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)
139 if (sh->qd_idx == sh->disks - 1)
142 return sh->qd_idx + 1;
144 static inline int raid6_next_disk(int disk, int raid_disks)
147 return (disk < raid_disks) ? disk : 0;
150 /* When walking through the disks in a raid5, starting at raid6_d0,
151 * We need to map each disk to a 'slot', where the data disks are slot
152 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
153 * is raid_disks-1. This help does that mapping.
155 static int raid6_idx_to_slot(int idx, struct stripe_head *sh, int *count)
158 if (idx == sh->pd_idx)
159 return sh->disks - 2;
160 if (idx == sh->qd_idx)
161 return sh->disks - 1;
166 static void return_io(struct bio *return_bi)
168 struct bio *bi = return_bi;
171 return_bi = bi->bi_next;
179 static void print_raid5_conf (raid5_conf_t *conf);
181 static int stripe_operations_active(struct stripe_head *sh)
183 return sh->check_state || sh->reconstruct_state ||
184 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
185 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
188 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
190 if (atomic_dec_and_test(&sh->count)) {
191 BUG_ON(!list_empty(&sh->lru));
192 BUG_ON(atomic_read(&conf->active_stripes)==0);
193 if (test_bit(STRIPE_HANDLE, &sh->state)) {
194 if (test_bit(STRIPE_DELAYED, &sh->state)) {
195 list_add_tail(&sh->lru, &conf->delayed_list);
196 blk_plug_device(conf->mddev->queue);
197 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
198 sh->bm_seq - conf->seq_write > 0) {
199 list_add_tail(&sh->lru, &conf->bitmap_list);
200 blk_plug_device(conf->mddev->queue);
202 clear_bit(STRIPE_BIT_DELAY, &sh->state);
203 list_add_tail(&sh->lru, &conf->handle_list);
205 md_wakeup_thread(conf->mddev->thread);
207 BUG_ON(stripe_operations_active(sh));
208 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
209 atomic_dec(&conf->preread_active_stripes);
210 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
211 md_wakeup_thread(conf->mddev->thread);
213 atomic_dec(&conf->active_stripes);
214 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
215 list_add_tail(&sh->lru, &conf->inactive_list);
216 wake_up(&conf->wait_for_stripe);
217 if (conf->retry_read_aligned)
218 md_wakeup_thread(conf->mddev->thread);
224 static void release_stripe(struct stripe_head *sh)
226 raid5_conf_t *conf = sh->raid_conf;
229 spin_lock_irqsave(&conf->device_lock, flags);
230 __release_stripe(conf, sh);
231 spin_unlock_irqrestore(&conf->device_lock, flags);
234 static inline void remove_hash(struct stripe_head *sh)
236 pr_debug("remove_hash(), stripe %llu\n",
237 (unsigned long long)sh->sector);
239 hlist_del_init(&sh->hash);
242 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
244 struct hlist_head *hp = stripe_hash(conf, sh->sector);
246 pr_debug("insert_hash(), stripe %llu\n",
247 (unsigned long long)sh->sector);
250 hlist_add_head(&sh->hash, hp);
254 /* find an idle stripe, make sure it is unhashed, and return it. */
255 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
257 struct stripe_head *sh = NULL;
258 struct list_head *first;
261 if (list_empty(&conf->inactive_list))
263 first = conf->inactive_list.next;
264 sh = list_entry(first, struct stripe_head, lru);
265 list_del_init(first);
267 atomic_inc(&conf->active_stripes);
272 static void shrink_buffers(struct stripe_head *sh, int num)
277 for (i=0; i<num ; i++) {
281 sh->dev[i].page = NULL;
286 static int grow_buffers(struct stripe_head *sh, int num)
290 for (i=0; i<num; i++) {
293 if (!(page = alloc_page(GFP_KERNEL))) {
296 sh->dev[i].page = page;
301 static void raid5_build_block(struct stripe_head *sh, int i);
302 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
303 struct stripe_head *sh);
305 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
307 raid5_conf_t *conf = sh->raid_conf;
310 BUG_ON(atomic_read(&sh->count) != 0);
311 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
312 BUG_ON(stripe_operations_active(sh));
315 pr_debug("init_stripe called, stripe %llu\n",
316 (unsigned long long)sh->sector);
320 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
322 stripe_set_idx(sector, conf, previous, sh);
326 for (i = sh->disks; i--; ) {
327 struct r5dev *dev = &sh->dev[i];
329 if (dev->toread || dev->read || dev->towrite || dev->written ||
330 test_bit(R5_LOCKED, &dev->flags)) {
331 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
332 (unsigned long long)sh->sector, i, dev->toread,
333 dev->read, dev->towrite, dev->written,
334 test_bit(R5_LOCKED, &dev->flags));
338 raid5_build_block(sh, i);
340 insert_hash(conf, sh);
343 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
345 struct stripe_head *sh;
346 struct hlist_node *hn;
349 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
350 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
351 if (sh->sector == sector && sh->disks == disks)
353 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
357 static void unplug_slaves(mddev_t *mddev);
358 static void raid5_unplug_device(struct request_queue *q);
360 static struct stripe_head *
361 get_active_stripe(raid5_conf_t *conf, sector_t sector,
362 int previous, int noblock)
364 struct stripe_head *sh;
365 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
367 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
369 spin_lock_irq(&conf->device_lock);
372 wait_event_lock_irq(conf->wait_for_stripe,
374 conf->device_lock, /* nothing */);
375 sh = __find_stripe(conf, sector, disks);
377 if (!conf->inactive_blocked)
378 sh = get_free_stripe(conf);
379 if (noblock && sh == NULL)
382 conf->inactive_blocked = 1;
383 wait_event_lock_irq(conf->wait_for_stripe,
384 !list_empty(&conf->inactive_list) &&
385 (atomic_read(&conf->active_stripes)
386 < (conf->max_nr_stripes *3/4)
387 || !conf->inactive_blocked),
389 raid5_unplug_device(conf->mddev->queue)
391 conf->inactive_blocked = 0;
393 init_stripe(sh, sector, previous);
395 if (atomic_read(&sh->count)) {
396 BUG_ON(!list_empty(&sh->lru));
398 if (!test_bit(STRIPE_HANDLE, &sh->state))
399 atomic_inc(&conf->active_stripes);
400 if (list_empty(&sh->lru) &&
401 !test_bit(STRIPE_EXPANDING, &sh->state))
403 list_del_init(&sh->lru);
406 } while (sh == NULL);
409 atomic_inc(&sh->count);
411 spin_unlock_irq(&conf->device_lock);
416 raid5_end_read_request(struct bio *bi, int error);
418 raid5_end_write_request(struct bio *bi, int error);
420 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
422 raid5_conf_t *conf = sh->raid_conf;
423 int i, disks = sh->disks;
427 for (i = disks; i--; ) {
431 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
433 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
438 bi = &sh->dev[i].req;
442 bi->bi_end_io = raid5_end_write_request;
444 bi->bi_end_io = raid5_end_read_request;
447 rdev = rcu_dereference(conf->disks[i].rdev);
448 if (rdev && test_bit(Faulty, &rdev->flags))
451 atomic_inc(&rdev->nr_pending);
455 if (s->syncing || s->expanding || s->expanded)
456 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
458 set_bit(STRIPE_IO_STARTED, &sh->state);
460 bi->bi_bdev = rdev->bdev;
461 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
462 __func__, (unsigned long long)sh->sector,
464 atomic_inc(&sh->count);
465 bi->bi_sector = sh->sector + rdev->data_offset;
466 bi->bi_flags = 1 << BIO_UPTODATE;
470 bi->bi_io_vec = &sh->dev[i].vec;
471 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
472 bi->bi_io_vec[0].bv_offset = 0;
473 bi->bi_size = STRIPE_SIZE;
476 test_bit(R5_ReWrite, &sh->dev[i].flags))
477 atomic_add(STRIPE_SECTORS,
478 &rdev->corrected_errors);
479 generic_make_request(bi);
482 set_bit(STRIPE_DEGRADED, &sh->state);
483 pr_debug("skip op %ld on disc %d for sector %llu\n",
484 bi->bi_rw, i, (unsigned long long)sh->sector);
485 clear_bit(R5_LOCKED, &sh->dev[i].flags);
486 set_bit(STRIPE_HANDLE, &sh->state);
491 static struct dma_async_tx_descriptor *
492 async_copy_data(int frombio, struct bio *bio, struct page *page,
493 sector_t sector, struct dma_async_tx_descriptor *tx)
496 struct page *bio_page;
500 if (bio->bi_sector >= sector)
501 page_offset = (signed)(bio->bi_sector - sector) * 512;
503 page_offset = (signed)(sector - bio->bi_sector) * -512;
504 bio_for_each_segment(bvl, bio, i) {
505 int len = bio_iovec_idx(bio, i)->bv_len;
509 if (page_offset < 0) {
510 b_offset = -page_offset;
511 page_offset += b_offset;
515 if (len > 0 && page_offset + len > STRIPE_SIZE)
516 clen = STRIPE_SIZE - page_offset;
521 b_offset += bio_iovec_idx(bio, i)->bv_offset;
522 bio_page = bio_iovec_idx(bio, i)->bv_page;
524 tx = async_memcpy(page, bio_page, page_offset,
529 tx = async_memcpy(bio_page, page, b_offset,
534 if (clen < len) /* hit end of page */
542 static void ops_complete_biofill(void *stripe_head_ref)
544 struct stripe_head *sh = stripe_head_ref;
545 struct bio *return_bi = NULL;
546 raid5_conf_t *conf = sh->raid_conf;
549 pr_debug("%s: stripe %llu\n", __func__,
550 (unsigned long long)sh->sector);
552 /* clear completed biofills */
553 spin_lock_irq(&conf->device_lock);
554 for (i = sh->disks; i--; ) {
555 struct r5dev *dev = &sh->dev[i];
557 /* acknowledge completion of a biofill operation */
558 /* and check if we need to reply to a read request,
559 * new R5_Wantfill requests are held off until
560 * !STRIPE_BIOFILL_RUN
562 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
563 struct bio *rbi, *rbi2;
568 while (rbi && rbi->bi_sector <
569 dev->sector + STRIPE_SECTORS) {
570 rbi2 = r5_next_bio(rbi, dev->sector);
571 if (!raid5_dec_bi_phys_segments(rbi)) {
572 rbi->bi_next = return_bi;
579 spin_unlock_irq(&conf->device_lock);
580 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
582 return_io(return_bi);
584 set_bit(STRIPE_HANDLE, &sh->state);
588 static void ops_run_biofill(struct stripe_head *sh)
590 struct dma_async_tx_descriptor *tx = NULL;
591 raid5_conf_t *conf = sh->raid_conf;
594 pr_debug("%s: stripe %llu\n", __func__,
595 (unsigned long long)sh->sector);
597 for (i = sh->disks; i--; ) {
598 struct r5dev *dev = &sh->dev[i];
599 if (test_bit(R5_Wantfill, &dev->flags)) {
601 spin_lock_irq(&conf->device_lock);
602 dev->read = rbi = dev->toread;
604 spin_unlock_irq(&conf->device_lock);
605 while (rbi && rbi->bi_sector <
606 dev->sector + STRIPE_SECTORS) {
607 tx = async_copy_data(0, rbi, dev->page,
609 rbi = r5_next_bio(rbi, dev->sector);
614 atomic_inc(&sh->count);
615 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
616 ops_complete_biofill, sh);
619 static void ops_complete_compute5(void *stripe_head_ref)
621 struct stripe_head *sh = stripe_head_ref;
622 int target = sh->ops.target;
623 struct r5dev *tgt = &sh->dev[target];
625 pr_debug("%s: stripe %llu\n", __func__,
626 (unsigned long long)sh->sector);
628 set_bit(R5_UPTODATE, &tgt->flags);
629 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
630 clear_bit(R5_Wantcompute, &tgt->flags);
631 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
632 if (sh->check_state == check_state_compute_run)
633 sh->check_state = check_state_compute_result;
634 set_bit(STRIPE_HANDLE, &sh->state);
638 static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
640 /* kernel stack size limits the total number of disks */
641 int disks = sh->disks;
642 struct page *xor_srcs[disks];
643 int target = sh->ops.target;
644 struct r5dev *tgt = &sh->dev[target];
645 struct page *xor_dest = tgt->page;
647 struct dma_async_tx_descriptor *tx;
650 pr_debug("%s: stripe %llu block: %d\n",
651 __func__, (unsigned long long)sh->sector, target);
652 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
654 for (i = disks; i--; )
656 xor_srcs[count++] = sh->dev[i].page;
658 atomic_inc(&sh->count);
660 if (unlikely(count == 1))
661 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
662 0, NULL, ops_complete_compute5, sh);
664 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
665 ASYNC_TX_XOR_ZERO_DST, NULL,
666 ops_complete_compute5, sh);
671 static void ops_complete_prexor(void *stripe_head_ref)
673 struct stripe_head *sh = stripe_head_ref;
675 pr_debug("%s: stripe %llu\n", __func__,
676 (unsigned long long)sh->sector);
679 static struct dma_async_tx_descriptor *
680 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
682 /* kernel stack size limits the total number of disks */
683 int disks = sh->disks;
684 struct page *xor_srcs[disks];
685 int count = 0, pd_idx = sh->pd_idx, i;
687 /* existing parity data subtracted */
688 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
690 pr_debug("%s: stripe %llu\n", __func__,
691 (unsigned long long)sh->sector);
693 for (i = disks; i--; ) {
694 struct r5dev *dev = &sh->dev[i];
695 /* Only process blocks that are known to be uptodate */
696 if (test_bit(R5_Wantdrain, &dev->flags))
697 xor_srcs[count++] = dev->page;
700 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
701 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
702 ops_complete_prexor, sh);
707 static struct dma_async_tx_descriptor *
708 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
710 int disks = sh->disks;
713 pr_debug("%s: stripe %llu\n", __func__,
714 (unsigned long long)sh->sector);
716 for (i = disks; i--; ) {
717 struct r5dev *dev = &sh->dev[i];
720 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
723 spin_lock(&sh->lock);
724 chosen = dev->towrite;
726 BUG_ON(dev->written);
727 wbi = dev->written = chosen;
728 spin_unlock(&sh->lock);
730 while (wbi && wbi->bi_sector <
731 dev->sector + STRIPE_SECTORS) {
732 tx = async_copy_data(1, wbi, dev->page,
734 wbi = r5_next_bio(wbi, dev->sector);
742 static void ops_complete_postxor(void *stripe_head_ref)
744 struct stripe_head *sh = stripe_head_ref;
745 int disks = sh->disks, i, pd_idx = sh->pd_idx;
747 pr_debug("%s: stripe %llu\n", __func__,
748 (unsigned long long)sh->sector);
750 for (i = disks; i--; ) {
751 struct r5dev *dev = &sh->dev[i];
752 if (dev->written || i == pd_idx)
753 set_bit(R5_UPTODATE, &dev->flags);
756 if (sh->reconstruct_state == reconstruct_state_drain_run)
757 sh->reconstruct_state = reconstruct_state_drain_result;
758 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
759 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
761 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
762 sh->reconstruct_state = reconstruct_state_result;
765 set_bit(STRIPE_HANDLE, &sh->state);
770 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
772 /* kernel stack size limits the total number of disks */
773 int disks = sh->disks;
774 struct page *xor_srcs[disks];
776 int count = 0, pd_idx = sh->pd_idx, i;
777 struct page *xor_dest;
781 pr_debug("%s: stripe %llu\n", __func__,
782 (unsigned long long)sh->sector);
784 /* check if prexor is active which means only process blocks
785 * that are part of a read-modify-write (written)
787 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
789 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
790 for (i = disks; i--; ) {
791 struct r5dev *dev = &sh->dev[i];
793 xor_srcs[count++] = dev->page;
796 xor_dest = sh->dev[pd_idx].page;
797 for (i = disks; i--; ) {
798 struct r5dev *dev = &sh->dev[i];
800 xor_srcs[count++] = dev->page;
804 /* 1/ if we prexor'd then the dest is reused as a source
805 * 2/ if we did not prexor then we are redoing the parity
806 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
807 * for the synchronous xor case
809 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
810 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
812 atomic_inc(&sh->count);
814 if (unlikely(count == 1)) {
815 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
816 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
817 flags, tx, ops_complete_postxor, sh);
819 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
820 flags, tx, ops_complete_postxor, sh);
823 static void ops_complete_check(void *stripe_head_ref)
825 struct stripe_head *sh = stripe_head_ref;
827 pr_debug("%s: stripe %llu\n", __func__,
828 (unsigned long long)sh->sector);
830 sh->check_state = check_state_check_result;
831 set_bit(STRIPE_HANDLE, &sh->state);
835 static void ops_run_check(struct stripe_head *sh)
837 /* kernel stack size limits the total number of disks */
838 int disks = sh->disks;
839 struct page *xor_srcs[disks];
840 struct dma_async_tx_descriptor *tx;
842 int count = 0, pd_idx = sh->pd_idx, i;
843 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
845 pr_debug("%s: stripe %llu\n", __func__,
846 (unsigned long long)sh->sector);
848 for (i = disks; i--; ) {
849 struct r5dev *dev = &sh->dev[i];
851 xor_srcs[count++] = dev->page;
854 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
855 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
857 atomic_inc(&sh->count);
858 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
859 ops_complete_check, sh);
862 static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
864 int overlap_clear = 0, i, disks = sh->disks;
865 struct dma_async_tx_descriptor *tx = NULL;
867 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
872 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
873 tx = ops_run_compute5(sh);
874 /* terminate the chain if postxor is not set to be run */
875 if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
879 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
880 tx = ops_run_prexor(sh, tx);
882 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
883 tx = ops_run_biodrain(sh, tx);
887 if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
888 ops_run_postxor(sh, tx);
890 if (test_bit(STRIPE_OP_CHECK, &ops_request))
894 for (i = disks; i--; ) {
895 struct r5dev *dev = &sh->dev[i];
896 if (test_and_clear_bit(R5_Overlap, &dev->flags))
897 wake_up(&sh->raid_conf->wait_for_overlap);
901 static int grow_one_stripe(raid5_conf_t *conf)
903 struct stripe_head *sh;
904 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
907 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
908 sh->raid_conf = conf;
909 spin_lock_init(&sh->lock);
911 if (grow_buffers(sh, conf->raid_disks)) {
912 shrink_buffers(sh, conf->raid_disks);
913 kmem_cache_free(conf->slab_cache, sh);
916 sh->disks = conf->raid_disks;
917 /* we just created an active stripe so... */
918 atomic_set(&sh->count, 1);
919 atomic_inc(&conf->active_stripes);
920 INIT_LIST_HEAD(&sh->lru);
925 static int grow_stripes(raid5_conf_t *conf, int num)
927 struct kmem_cache *sc;
928 int devs = conf->raid_disks;
930 sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
931 sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
932 conf->active_name = 0;
933 sc = kmem_cache_create(conf->cache_name[conf->active_name],
934 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
938 conf->slab_cache = sc;
939 conf->pool_size = devs;
941 if (!grow_one_stripe(conf))
946 #ifdef CONFIG_MD_RAID5_RESHAPE
947 static int resize_stripes(raid5_conf_t *conf, int newsize)
949 /* Make all the stripes able to hold 'newsize' devices.
950 * New slots in each stripe get 'page' set to a new page.
952 * This happens in stages:
953 * 1/ create a new kmem_cache and allocate the required number of
955 * 2/ gather all the old stripe_heads and tranfer the pages across
956 * to the new stripe_heads. This will have the side effect of
957 * freezing the array as once all stripe_heads have been collected,
958 * no IO will be possible. Old stripe heads are freed once their
959 * pages have been transferred over, and the old kmem_cache is
960 * freed when all stripes are done.
961 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
962 * we simple return a failre status - no need to clean anything up.
963 * 4/ allocate new pages for the new slots in the new stripe_heads.
964 * If this fails, we don't bother trying the shrink the
965 * stripe_heads down again, we just leave them as they are.
966 * As each stripe_head is processed the new one is released into
969 * Once step2 is started, we cannot afford to wait for a write,
970 * so we use GFP_NOIO allocations.
972 struct stripe_head *osh, *nsh;
973 LIST_HEAD(newstripes);
974 struct disk_info *ndisks;
976 struct kmem_cache *sc;
979 if (newsize <= conf->pool_size)
980 return 0; /* never bother to shrink */
982 err = md_allow_write(conf->mddev);
987 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
988 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
993 for (i = conf->max_nr_stripes; i; i--) {
994 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
998 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1000 nsh->raid_conf = conf;
1001 spin_lock_init(&nsh->lock);
1003 list_add(&nsh->lru, &newstripes);
1006 /* didn't get enough, give up */
1007 while (!list_empty(&newstripes)) {
1008 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1009 list_del(&nsh->lru);
1010 kmem_cache_free(sc, nsh);
1012 kmem_cache_destroy(sc);
1015 /* Step 2 - Must use GFP_NOIO now.
1016 * OK, we have enough stripes, start collecting inactive
1017 * stripes and copying them over
1019 list_for_each_entry(nsh, &newstripes, lru) {
1020 spin_lock_irq(&conf->device_lock);
1021 wait_event_lock_irq(conf->wait_for_stripe,
1022 !list_empty(&conf->inactive_list),
1024 unplug_slaves(conf->mddev)
1026 osh = get_free_stripe(conf);
1027 spin_unlock_irq(&conf->device_lock);
1028 atomic_set(&nsh->count, 1);
1029 for(i=0; i<conf->pool_size; i++)
1030 nsh->dev[i].page = osh->dev[i].page;
1031 for( ; i<newsize; i++)
1032 nsh->dev[i].page = NULL;
1033 kmem_cache_free(conf->slab_cache, osh);
1035 kmem_cache_destroy(conf->slab_cache);
1038 * At this point, we are holding all the stripes so the array
1039 * is completely stalled, so now is a good time to resize
1042 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1044 for (i=0; i<conf->raid_disks; i++)
1045 ndisks[i] = conf->disks[i];
1047 conf->disks = ndisks;
1051 /* Step 4, return new stripes to service */
1052 while(!list_empty(&newstripes)) {
1053 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1054 list_del_init(&nsh->lru);
1055 for (i=conf->raid_disks; i < newsize; i++)
1056 if (nsh->dev[i].page == NULL) {
1057 struct page *p = alloc_page(GFP_NOIO);
1058 nsh->dev[i].page = p;
1062 release_stripe(nsh);
1064 /* critical section pass, GFP_NOIO no longer needed */
1066 conf->slab_cache = sc;
1067 conf->active_name = 1-conf->active_name;
1068 conf->pool_size = newsize;
1073 static int drop_one_stripe(raid5_conf_t *conf)
1075 struct stripe_head *sh;
1077 spin_lock_irq(&conf->device_lock);
1078 sh = get_free_stripe(conf);
1079 spin_unlock_irq(&conf->device_lock);
1082 BUG_ON(atomic_read(&sh->count));
1083 shrink_buffers(sh, conf->pool_size);
1084 kmem_cache_free(conf->slab_cache, sh);
1085 atomic_dec(&conf->active_stripes);
1089 static void shrink_stripes(raid5_conf_t *conf)
1091 while (drop_one_stripe(conf))
1094 if (conf->slab_cache)
1095 kmem_cache_destroy(conf->slab_cache);
1096 conf->slab_cache = NULL;
1099 static void raid5_end_read_request(struct bio * bi, int error)
1101 struct stripe_head *sh = bi->bi_private;
1102 raid5_conf_t *conf = sh->raid_conf;
1103 int disks = sh->disks, i;
1104 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1105 char b[BDEVNAME_SIZE];
1109 for (i=0 ; i<disks; i++)
1110 if (bi == &sh->dev[i].req)
1113 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1114 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1122 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1123 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1124 rdev = conf->disks[i].rdev;
1125 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1126 " (%lu sectors at %llu on %s)\n",
1127 mdname(conf->mddev), STRIPE_SECTORS,
1128 (unsigned long long)(sh->sector
1129 + rdev->data_offset),
1130 bdevname(rdev->bdev, b));
1131 clear_bit(R5_ReadError, &sh->dev[i].flags);
1132 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1134 if (atomic_read(&conf->disks[i].rdev->read_errors))
1135 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1137 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1139 rdev = conf->disks[i].rdev;
1141 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1142 atomic_inc(&rdev->read_errors);
1143 if (conf->mddev->degraded)
1144 printk_rl(KERN_WARNING
1145 "raid5:%s: read error not correctable "
1146 "(sector %llu on %s).\n",
1147 mdname(conf->mddev),
1148 (unsigned long long)(sh->sector
1149 + rdev->data_offset),
1151 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1153 printk_rl(KERN_WARNING
1154 "raid5:%s: read error NOT corrected!! "
1155 "(sector %llu on %s).\n",
1156 mdname(conf->mddev),
1157 (unsigned long long)(sh->sector
1158 + rdev->data_offset),
1160 else if (atomic_read(&rdev->read_errors)
1161 > conf->max_nr_stripes)
1163 "raid5:%s: Too many read errors, failing device %s.\n",
1164 mdname(conf->mddev), bdn);
1168 set_bit(R5_ReadError, &sh->dev[i].flags);
1170 clear_bit(R5_ReadError, &sh->dev[i].flags);
1171 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1172 md_error(conf->mddev, rdev);
1175 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1176 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1177 set_bit(STRIPE_HANDLE, &sh->state);
1181 static void raid5_end_write_request(struct bio *bi, int error)
1183 struct stripe_head *sh = bi->bi_private;
1184 raid5_conf_t *conf = sh->raid_conf;
1185 int disks = sh->disks, i;
1186 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1188 for (i=0 ; i<disks; i++)
1189 if (bi == &sh->dev[i].req)
1192 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1193 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1201 md_error(conf->mddev, conf->disks[i].rdev);
1203 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1205 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1206 set_bit(STRIPE_HANDLE, &sh->state);
1211 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1213 static void raid5_build_block(struct stripe_head *sh, int i)
1215 struct r5dev *dev = &sh->dev[i];
1217 bio_init(&dev->req);
1218 dev->req.bi_io_vec = &dev->vec;
1220 dev->req.bi_max_vecs++;
1221 dev->vec.bv_page = dev->page;
1222 dev->vec.bv_len = STRIPE_SIZE;
1223 dev->vec.bv_offset = 0;
1225 dev->req.bi_sector = sh->sector;
1226 dev->req.bi_private = sh;
1229 dev->sector = compute_blocknr(sh, i);
1232 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1234 char b[BDEVNAME_SIZE];
1235 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1236 pr_debug("raid5: error called\n");
1238 if (!test_bit(Faulty, &rdev->flags)) {
1239 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1240 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1241 unsigned long flags;
1242 spin_lock_irqsave(&conf->device_lock, flags);
1244 spin_unlock_irqrestore(&conf->device_lock, flags);
1246 * if recovery was running, make sure it aborts.
1248 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1250 set_bit(Faulty, &rdev->flags);
1252 "raid5: Disk failure on %s, disabling device.\n"
1253 "raid5: Operation continuing on %d devices.\n",
1254 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1259 * Input: a 'big' sector number,
1260 * Output: index of the data and parity disk, and the sector # in them.
1262 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1263 int previous, int *dd_idx,
1264 struct stripe_head *sh)
1267 unsigned long chunk_number;
1268 unsigned int chunk_offset;
1270 sector_t new_sector;
1271 int sectors_per_chunk = conf->chunk_size >> 9;
1272 int raid_disks = previous ? conf->previous_raid_disks
1274 int data_disks = raid_disks - conf->max_degraded;
1276 /* First compute the information on this sector */
1279 * Compute the chunk number and the sector offset inside the chunk
1281 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1282 chunk_number = r_sector;
1283 BUG_ON(r_sector != chunk_number);
1286 * Compute the stripe number
1288 stripe = chunk_number / data_disks;
1291 * Compute the data disk and parity disk indexes inside the stripe
1293 *dd_idx = chunk_number % data_disks;
1296 * Select the parity disk based on the user selected algorithm.
1298 pd_idx = qd_idx = ~0;
1299 switch(conf->level) {
1301 pd_idx = data_disks;
1304 switch (conf->algorithm) {
1305 case ALGORITHM_LEFT_ASYMMETRIC:
1306 pd_idx = data_disks - stripe % raid_disks;
1307 if (*dd_idx >= pd_idx)
1310 case ALGORITHM_RIGHT_ASYMMETRIC:
1311 pd_idx = stripe % raid_disks;
1312 if (*dd_idx >= pd_idx)
1315 case ALGORITHM_LEFT_SYMMETRIC:
1316 pd_idx = data_disks - stripe % raid_disks;
1317 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1319 case ALGORITHM_RIGHT_SYMMETRIC:
1320 pd_idx = stripe % raid_disks;
1321 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1323 case ALGORITHM_PARITY_0:
1327 case ALGORITHM_PARITY_N:
1328 pd_idx = data_disks;
1331 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1338 switch (conf->algorithm) {
1339 case ALGORITHM_LEFT_ASYMMETRIC:
1340 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1341 qd_idx = pd_idx + 1;
1342 if (pd_idx == raid_disks-1) {
1343 (*dd_idx)++; /* Q D D D P */
1345 } else if (*dd_idx >= pd_idx)
1346 (*dd_idx) += 2; /* D D P Q D */
1348 case ALGORITHM_RIGHT_ASYMMETRIC:
1349 pd_idx = stripe % raid_disks;
1350 qd_idx = pd_idx + 1;
1351 if (pd_idx == raid_disks-1) {
1352 (*dd_idx)++; /* Q D D D P */
1354 } else if (*dd_idx >= pd_idx)
1355 (*dd_idx) += 2; /* D D P Q D */
1357 case ALGORITHM_LEFT_SYMMETRIC:
1358 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1359 qd_idx = (pd_idx + 1) % raid_disks;
1360 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1362 case ALGORITHM_RIGHT_SYMMETRIC:
1363 pd_idx = stripe % raid_disks;
1364 qd_idx = (pd_idx + 1) % raid_disks;
1365 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1368 case ALGORITHM_PARITY_0:
1373 case ALGORITHM_PARITY_N:
1374 pd_idx = data_disks;
1375 qd_idx = data_disks + 1;
1378 case ALGORITHM_ROTATING_ZERO_RESTART:
1379 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1380 * of blocks for computing Q is different.
1382 pd_idx = stripe % raid_disks;
1383 qd_idx = pd_idx + 1;
1384 if (pd_idx == raid_disks-1) {
1385 (*dd_idx)++; /* Q D D D P */
1387 } else if (*dd_idx >= pd_idx)
1388 (*dd_idx) += 2; /* D D P Q D */
1391 case ALGORITHM_ROTATING_N_RESTART:
1392 /* Same a left_asymmetric, by first stripe is
1393 * D D D P Q rather than
1396 pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1397 qd_idx = pd_idx + 1;
1398 if (pd_idx == raid_disks-1) {
1399 (*dd_idx)++; /* Q D D D P */
1401 } else if (*dd_idx >= pd_idx)
1402 (*dd_idx) += 2; /* D D P Q D */
1405 case ALGORITHM_ROTATING_N_CONTINUE:
1406 /* Same as left_symmetric but Q is before P */
1407 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1408 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1409 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1412 case ALGORITHM_LEFT_ASYMMETRIC_6:
1413 /* RAID5 left_asymmetric, with Q on last device */
1414 pd_idx = data_disks - stripe % (raid_disks-1);
1415 if (*dd_idx >= pd_idx)
1417 qd_idx = raid_disks - 1;
1420 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1421 pd_idx = stripe % (raid_disks-1);
1422 if (*dd_idx >= pd_idx)
1424 qd_idx = raid_disks - 1;
1427 case ALGORITHM_LEFT_SYMMETRIC_6:
1428 pd_idx = data_disks - stripe % (raid_disks-1);
1429 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1430 qd_idx = raid_disks - 1;
1433 case ALGORITHM_RIGHT_SYMMETRIC_6:
1434 pd_idx = stripe % (raid_disks-1);
1435 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1436 qd_idx = raid_disks - 1;
1439 case ALGORITHM_PARITY_0_6:
1442 qd_idx = raid_disks - 1;
1447 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1455 sh->pd_idx = pd_idx;
1456 sh->qd_idx = qd_idx;
1459 * Finally, compute the new sector number
1461 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1466 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1468 raid5_conf_t *conf = sh->raid_conf;
1469 int raid_disks = sh->disks;
1470 int data_disks = raid_disks - conf->max_degraded;
1471 sector_t new_sector = sh->sector, check;
1472 int sectors_per_chunk = conf->chunk_size >> 9;
1475 int chunk_number, dummy1, dd_idx = i;
1477 struct stripe_head sh2;
1480 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1481 stripe = new_sector;
1482 BUG_ON(new_sector != stripe);
1484 if (i == sh->pd_idx)
1486 switch(conf->level) {
1489 switch (conf->algorithm) {
1490 case ALGORITHM_LEFT_ASYMMETRIC:
1491 case ALGORITHM_RIGHT_ASYMMETRIC:
1495 case ALGORITHM_LEFT_SYMMETRIC:
1496 case ALGORITHM_RIGHT_SYMMETRIC:
1499 i -= (sh->pd_idx + 1);
1501 case ALGORITHM_PARITY_0:
1504 case ALGORITHM_PARITY_N:
1507 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1513 if (i == sh->qd_idx)
1514 return 0; /* It is the Q disk */
1515 switch (conf->algorithm) {
1516 case ALGORITHM_LEFT_ASYMMETRIC:
1517 case ALGORITHM_RIGHT_ASYMMETRIC:
1518 case ALGORITHM_ROTATING_ZERO_RESTART:
1519 case ALGORITHM_ROTATING_N_RESTART:
1520 if (sh->pd_idx == raid_disks-1)
1521 i--; /* Q D D D P */
1522 else if (i > sh->pd_idx)
1523 i -= 2; /* D D P Q D */
1525 case ALGORITHM_LEFT_SYMMETRIC:
1526 case ALGORITHM_RIGHT_SYMMETRIC:
1527 if (sh->pd_idx == raid_disks-1)
1528 i--; /* Q D D D P */
1533 i -= (sh->pd_idx + 2);
1536 case ALGORITHM_PARITY_0:
1539 case ALGORITHM_PARITY_N:
1541 case ALGORITHM_ROTATING_N_CONTINUE:
1542 if (sh->pd_idx == 0)
1543 i--; /* P D D D Q */
1544 else if (i > sh->pd_idx)
1545 i -= 2; /* D D Q P D */
1547 case ALGORITHM_LEFT_ASYMMETRIC_6:
1548 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1552 case ALGORITHM_LEFT_SYMMETRIC_6:
1553 case ALGORITHM_RIGHT_SYMMETRIC_6:
1555 i += data_disks + 1;
1556 i -= (sh->pd_idx + 1);
1558 case ALGORITHM_PARITY_0_6:
1562 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1569 chunk_number = stripe * data_disks + i;
1570 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1572 check = raid5_compute_sector(conf, r_sector,
1573 (raid_disks != conf->raid_disks),
1575 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1576 || sh2.qd_idx != sh->qd_idx) {
1577 printk(KERN_ERR "compute_blocknr: map not correct\n");
1586 * Copy data between a page in the stripe cache, and one or more bion
1587 * The page could align with the middle of the bio, or there could be
1588 * several bion, each with several bio_vecs, which cover part of the page
1589 * Multiple bion are linked together on bi_next. There may be extras
1590 * at the end of this list. We ignore them.
1592 static void copy_data(int frombio, struct bio *bio,
1596 char *pa = page_address(page);
1597 struct bio_vec *bvl;
1601 if (bio->bi_sector >= sector)
1602 page_offset = (signed)(bio->bi_sector - sector) * 512;
1604 page_offset = (signed)(sector - bio->bi_sector) * -512;
1605 bio_for_each_segment(bvl, bio, i) {
1606 int len = bio_iovec_idx(bio,i)->bv_len;
1610 if (page_offset < 0) {
1611 b_offset = -page_offset;
1612 page_offset += b_offset;
1616 if (len > 0 && page_offset + len > STRIPE_SIZE)
1617 clen = STRIPE_SIZE - page_offset;
1621 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1623 memcpy(pa+page_offset, ba+b_offset, clen);
1625 memcpy(ba+b_offset, pa+page_offset, clen);
1626 __bio_kunmap_atomic(ba, KM_USER0);
1628 if (clen < len) /* hit end of page */
1634 #define check_xor() do { \
1635 if (count == MAX_XOR_BLOCKS) { \
1636 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1641 static void compute_parity6(struct stripe_head *sh, int method)
1643 raid5_conf_t *conf = sh->raid_conf;
1644 int i, pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1646 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1649 pd_idx = sh->pd_idx;
1650 qd_idx = sh->qd_idx;
1651 d0_idx = raid6_d0(sh);
1653 pr_debug("compute_parity, stripe %llu, method %d\n",
1654 (unsigned long long)sh->sector, method);
1657 case READ_MODIFY_WRITE:
1658 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1659 case RECONSTRUCT_WRITE:
1660 for (i= disks; i-- ;)
1661 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1662 chosen = sh->dev[i].towrite;
1663 sh->dev[i].towrite = NULL;
1665 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1666 wake_up(&conf->wait_for_overlap);
1668 BUG_ON(sh->dev[i].written);
1669 sh->dev[i].written = chosen;
1673 BUG(); /* Not implemented yet */
1676 for (i = disks; i--;)
1677 if (sh->dev[i].written) {
1678 sector_t sector = sh->dev[i].sector;
1679 struct bio *wbi = sh->dev[i].written;
1680 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1681 copy_data(1, wbi, sh->dev[i].page, sector);
1682 wbi = r5_next_bio(wbi, sector);
1685 set_bit(R5_LOCKED, &sh->dev[i].flags);
1686 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1689 /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1690 /* FIX: Is this ordering of drives even remotely optimal? */
1694 int slot = raid6_idx_to_slot(i, sh, &count);
1695 ptrs[slot] = page_address(sh->dev[i].page);
1696 if (slot < sh->disks - 2 &&
1697 !test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1698 printk(KERN_ERR "block %d/%d not uptodate "
1699 "on parity calc\n", i, count);
1702 i = raid6_next_disk(i, disks);
1703 } while (i != d0_idx);
1704 BUG_ON(count+2 != disks);
1706 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1709 case RECONSTRUCT_WRITE:
1710 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1711 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1712 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1713 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1716 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1717 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1723 /* Compute one missing block */
1724 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1726 int i, count, disks = sh->disks;
1727 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1728 int qd_idx = sh->qd_idx;
1730 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1731 (unsigned long long)sh->sector, dd_idx);
1733 if ( dd_idx == qd_idx ) {
1734 /* We're actually computing the Q drive */
1735 compute_parity6(sh, UPDATE_PARITY);
1737 dest = page_address(sh->dev[dd_idx].page);
1738 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1740 for (i = disks ; i--; ) {
1741 if (i == dd_idx || i == qd_idx)
1743 p = page_address(sh->dev[i].page);
1744 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1747 printk("compute_block() %d, stripe %llu, %d"
1748 " not present\n", dd_idx,
1749 (unsigned long long)sh->sector, i);
1754 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1755 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1756 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1760 /* Compute two missing blocks */
1761 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1763 int i, count, disks = sh->disks;
1764 int d0_idx = raid6_d0(sh);
1765 int faila = -1, failb = -1;
1766 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1773 slot = raid6_idx_to_slot(i, sh, &count);
1774 ptrs[slot] = page_address(sh->dev[i].page);
1779 i = raid6_next_disk(i, disks);
1780 } while (i != d0_idx);
1781 BUG_ON(count+2 != disks);
1783 BUG_ON(faila == failb);
1784 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1786 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1787 (unsigned long long)sh->sector, dd_idx1, dd_idx2,
1790 if ( failb == disks-1 ) {
1791 /* Q disk is one of the missing disks */
1792 if ( faila == disks-2 ) {
1793 /* Missing P+Q, just recompute */
1794 compute_parity6(sh, UPDATE_PARITY);
1797 /* We're missing D+Q; recompute D from P */
1798 compute_block_1(sh, ((dd_idx1 == sh->qd_idx) ?
1801 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1806 /* We're missing D+P or D+D; */
1807 if (failb == disks-2) {
1808 /* We're missing D+P. */
1809 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1811 /* We're missing D+D. */
1812 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1815 /* Both the above update both missing blocks */
1816 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1817 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1821 schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
1822 int rcw, int expand)
1824 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1827 /* if we are not expanding this is a proper write request, and
1828 * there will be bios with new data to be drained into the
1832 sh->reconstruct_state = reconstruct_state_drain_run;
1833 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1835 sh->reconstruct_state = reconstruct_state_run;
1837 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1839 for (i = disks; i--; ) {
1840 struct r5dev *dev = &sh->dev[i];
1843 set_bit(R5_LOCKED, &dev->flags);
1844 set_bit(R5_Wantdrain, &dev->flags);
1846 clear_bit(R5_UPTODATE, &dev->flags);
1850 if (s->locked + 1 == disks)
1851 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1852 atomic_inc(&sh->raid_conf->pending_full_writes);
1854 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1855 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1857 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1858 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1859 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1860 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1862 for (i = disks; i--; ) {
1863 struct r5dev *dev = &sh->dev[i];
1868 (test_bit(R5_UPTODATE, &dev->flags) ||
1869 test_bit(R5_Wantcompute, &dev->flags))) {
1870 set_bit(R5_Wantdrain, &dev->flags);
1871 set_bit(R5_LOCKED, &dev->flags);
1872 clear_bit(R5_UPTODATE, &dev->flags);
1878 /* keep the parity disk locked while asynchronous operations
1881 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1882 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1885 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1886 __func__, (unsigned long long)sh->sector,
1887 s->locked, s->ops_request);
1891 * Each stripe/dev can have one or more bion attached.
1892 * toread/towrite point to the first in a chain.
1893 * The bi_next chain must be in order.
1895 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1898 raid5_conf_t *conf = sh->raid_conf;
1901 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1902 (unsigned long long)bi->bi_sector,
1903 (unsigned long long)sh->sector);
1906 spin_lock(&sh->lock);
1907 spin_lock_irq(&conf->device_lock);
1909 bip = &sh->dev[dd_idx].towrite;
1910 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1913 bip = &sh->dev[dd_idx].toread;
1914 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1915 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1917 bip = & (*bip)->bi_next;
1919 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1922 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1926 bi->bi_phys_segments++;
1927 spin_unlock_irq(&conf->device_lock);
1928 spin_unlock(&sh->lock);
1930 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1931 (unsigned long long)bi->bi_sector,
1932 (unsigned long long)sh->sector, dd_idx);
1934 if (conf->mddev->bitmap && firstwrite) {
1935 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1937 sh->bm_seq = conf->seq_flush+1;
1938 set_bit(STRIPE_BIT_DELAY, &sh->state);
1942 /* check if page is covered */
1943 sector_t sector = sh->dev[dd_idx].sector;
1944 for (bi=sh->dev[dd_idx].towrite;
1945 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1946 bi && bi->bi_sector <= sector;
1947 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1948 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1949 sector = bi->bi_sector + (bi->bi_size>>9);
1951 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1952 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1957 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1958 spin_unlock_irq(&conf->device_lock);
1959 spin_unlock(&sh->lock);
1963 static void end_reshape(raid5_conf_t *conf);
1965 static int page_is_zero(struct page *p)
1967 char *a = page_address(p);
1968 return ((*(u32*)a) == 0 &&
1969 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1972 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
1973 struct stripe_head *sh)
1975 int sectors_per_chunk = conf->chunk_size >> 9;
1977 int chunk_offset = sector_div(stripe, sectors_per_chunk);
1978 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
1980 raid5_compute_sector(conf,
1981 stripe * (disks - conf->max_degraded)
1982 *sectors_per_chunk + chunk_offset,
1988 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
1989 struct stripe_head_state *s, int disks,
1990 struct bio **return_bi)
1993 for (i = disks; i--; ) {
1997 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2000 rdev = rcu_dereference(conf->disks[i].rdev);
2001 if (rdev && test_bit(In_sync, &rdev->flags))
2002 /* multiple read failures in one stripe */
2003 md_error(conf->mddev, rdev);
2006 spin_lock_irq(&conf->device_lock);
2007 /* fail all writes first */
2008 bi = sh->dev[i].towrite;
2009 sh->dev[i].towrite = NULL;
2015 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2016 wake_up(&conf->wait_for_overlap);
2018 while (bi && bi->bi_sector <
2019 sh->dev[i].sector + STRIPE_SECTORS) {
2020 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2021 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2022 if (!raid5_dec_bi_phys_segments(bi)) {
2023 md_write_end(conf->mddev);
2024 bi->bi_next = *return_bi;
2029 /* and fail all 'written' */
2030 bi = sh->dev[i].written;
2031 sh->dev[i].written = NULL;
2032 if (bi) bitmap_end = 1;
2033 while (bi && bi->bi_sector <
2034 sh->dev[i].sector + STRIPE_SECTORS) {
2035 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2036 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2037 if (!raid5_dec_bi_phys_segments(bi)) {
2038 md_write_end(conf->mddev);
2039 bi->bi_next = *return_bi;
2045 /* fail any reads if this device is non-operational and
2046 * the data has not reached the cache yet.
2048 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2049 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2050 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2051 bi = sh->dev[i].toread;
2052 sh->dev[i].toread = NULL;
2053 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2054 wake_up(&conf->wait_for_overlap);
2055 if (bi) s->to_read--;
2056 while (bi && bi->bi_sector <
2057 sh->dev[i].sector + STRIPE_SECTORS) {
2058 struct bio *nextbi =
2059 r5_next_bio(bi, sh->dev[i].sector);
2060 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2061 if (!raid5_dec_bi_phys_segments(bi)) {
2062 bi->bi_next = *return_bi;
2068 spin_unlock_irq(&conf->device_lock);
2070 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2071 STRIPE_SECTORS, 0, 0);
2074 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2075 if (atomic_dec_and_test(&conf->pending_full_writes))
2076 md_wakeup_thread(conf->mddev->thread);
2079 /* fetch_block5 - checks the given member device to see if its data needs
2080 * to be read or computed to satisfy a request.
2082 * Returns 1 when no more member devices need to be checked, otherwise returns
2083 * 0 to tell the loop in handle_stripe_fill5 to continue
2085 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2086 int disk_idx, int disks)
2088 struct r5dev *dev = &sh->dev[disk_idx];
2089 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2091 /* is the data in this block needed, and can we get it? */
2092 if (!test_bit(R5_LOCKED, &dev->flags) &&
2093 !test_bit(R5_UPTODATE, &dev->flags) &&
2095 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2096 s->syncing || s->expanding ||
2098 (failed_dev->toread ||
2099 (failed_dev->towrite &&
2100 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2101 /* We would like to get this block, possibly by computing it,
2102 * otherwise read it if the backing disk is insync
2104 if ((s->uptodate == disks - 1) &&
2105 (s->failed && disk_idx == s->failed_num)) {
2106 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2107 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2108 set_bit(R5_Wantcompute, &dev->flags);
2109 sh->ops.target = disk_idx;
2111 /* Careful: from this point on 'uptodate' is in the eye
2112 * of raid5_run_ops which services 'compute' operations
2113 * before writes. R5_Wantcompute flags a block that will
2114 * be R5_UPTODATE by the time it is needed for a
2115 * subsequent operation.
2118 return 1; /* uptodate + compute == disks */
2119 } else if (test_bit(R5_Insync, &dev->flags)) {
2120 set_bit(R5_LOCKED, &dev->flags);
2121 set_bit(R5_Wantread, &dev->flags);
2123 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2132 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2134 static void handle_stripe_fill5(struct stripe_head *sh,
2135 struct stripe_head_state *s, int disks)
2139 /* look for blocks to read/compute, skip this if a compute
2140 * is already in flight, or if the stripe contents are in the
2141 * midst of changing due to a write
2143 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2144 !sh->reconstruct_state)
2145 for (i = disks; i--; )
2146 if (fetch_block5(sh, s, i, disks))
2148 set_bit(STRIPE_HANDLE, &sh->state);
2151 static void handle_stripe_fill6(struct stripe_head *sh,
2152 struct stripe_head_state *s, struct r6_state *r6s,
2156 for (i = disks; i--; ) {
2157 struct r5dev *dev = &sh->dev[i];
2158 if (!test_bit(R5_LOCKED, &dev->flags) &&
2159 !test_bit(R5_UPTODATE, &dev->flags) &&
2160 (dev->toread || (dev->towrite &&
2161 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2162 s->syncing || s->expanding ||
2164 (sh->dev[r6s->failed_num[0]].toread ||
2167 (sh->dev[r6s->failed_num[1]].toread ||
2169 /* we would like to get this block, possibly
2170 * by computing it, but we might not be able to
2172 if ((s->uptodate == disks - 1) &&
2173 (s->failed && (i == r6s->failed_num[0] ||
2174 i == r6s->failed_num[1]))) {
2175 pr_debug("Computing stripe %llu block %d\n",
2176 (unsigned long long)sh->sector, i);
2177 compute_block_1(sh, i, 0);
2179 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2180 /* Computing 2-failure is *very* expensive; only
2181 * do it if failed >= 2
2184 for (other = disks; other--; ) {
2187 if (!test_bit(R5_UPTODATE,
2188 &sh->dev[other].flags))
2192 pr_debug("Computing stripe %llu blocks %d,%d\n",
2193 (unsigned long long)sh->sector,
2195 compute_block_2(sh, i, other);
2197 } else if (test_bit(R5_Insync, &dev->flags)) {
2198 set_bit(R5_LOCKED, &dev->flags);
2199 set_bit(R5_Wantread, &dev->flags);
2201 pr_debug("Reading block %d (sync=%d)\n",
2206 set_bit(STRIPE_HANDLE, &sh->state);
2210 /* handle_stripe_clean_event
2211 * any written block on an uptodate or failed drive can be returned.
2212 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2213 * never LOCKED, so we don't need to test 'failed' directly.
2215 static void handle_stripe_clean_event(raid5_conf_t *conf,
2216 struct stripe_head *sh, int disks, struct bio **return_bi)
2221 for (i = disks; i--; )
2222 if (sh->dev[i].written) {
2224 if (!test_bit(R5_LOCKED, &dev->flags) &&
2225 test_bit(R5_UPTODATE, &dev->flags)) {
2226 /* We can return any write requests */
2227 struct bio *wbi, *wbi2;
2229 pr_debug("Return write for disc %d\n", i);
2230 spin_lock_irq(&conf->device_lock);
2232 dev->written = NULL;
2233 while (wbi && wbi->bi_sector <
2234 dev->sector + STRIPE_SECTORS) {
2235 wbi2 = r5_next_bio(wbi, dev->sector);
2236 if (!raid5_dec_bi_phys_segments(wbi)) {
2237 md_write_end(conf->mddev);
2238 wbi->bi_next = *return_bi;
2243 if (dev->towrite == NULL)
2245 spin_unlock_irq(&conf->device_lock);
2247 bitmap_endwrite(conf->mddev->bitmap,
2250 !test_bit(STRIPE_DEGRADED, &sh->state),
2255 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2256 if (atomic_dec_and_test(&conf->pending_full_writes))
2257 md_wakeup_thread(conf->mddev->thread);
2260 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2261 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2263 int rmw = 0, rcw = 0, i;
2264 for (i = disks; i--; ) {
2265 /* would I have to read this buffer for read_modify_write */
2266 struct r5dev *dev = &sh->dev[i];
2267 if ((dev->towrite || i == sh->pd_idx) &&
2268 !test_bit(R5_LOCKED, &dev->flags) &&
2269 !(test_bit(R5_UPTODATE, &dev->flags) ||
2270 test_bit(R5_Wantcompute, &dev->flags))) {
2271 if (test_bit(R5_Insync, &dev->flags))
2274 rmw += 2*disks; /* cannot read it */
2276 /* Would I have to read this buffer for reconstruct_write */
2277 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2278 !test_bit(R5_LOCKED, &dev->flags) &&
2279 !(test_bit(R5_UPTODATE, &dev->flags) ||
2280 test_bit(R5_Wantcompute, &dev->flags))) {
2281 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2286 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2287 (unsigned long long)sh->sector, rmw, rcw);
2288 set_bit(STRIPE_HANDLE, &sh->state);
2289 if (rmw < rcw && rmw > 0)
2290 /* prefer read-modify-write, but need to get some data */
2291 for (i = disks; i--; ) {
2292 struct r5dev *dev = &sh->dev[i];
2293 if ((dev->towrite || i == sh->pd_idx) &&
2294 !test_bit(R5_LOCKED, &dev->flags) &&
2295 !(test_bit(R5_UPTODATE, &dev->flags) ||
2296 test_bit(R5_Wantcompute, &dev->flags)) &&
2297 test_bit(R5_Insync, &dev->flags)) {
2299 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2300 pr_debug("Read_old block "
2301 "%d for r-m-w\n", i);
2302 set_bit(R5_LOCKED, &dev->flags);
2303 set_bit(R5_Wantread, &dev->flags);
2306 set_bit(STRIPE_DELAYED, &sh->state);
2307 set_bit(STRIPE_HANDLE, &sh->state);
2311 if (rcw <= rmw && rcw > 0)
2312 /* want reconstruct write, but need to get some data */
2313 for (i = disks; i--; ) {
2314 struct r5dev *dev = &sh->dev[i];
2315 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2317 !test_bit(R5_LOCKED, &dev->flags) &&
2318 !(test_bit(R5_UPTODATE, &dev->flags) ||
2319 test_bit(R5_Wantcompute, &dev->flags)) &&
2320 test_bit(R5_Insync, &dev->flags)) {
2322 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2323 pr_debug("Read_old block "
2324 "%d for Reconstruct\n", i);
2325 set_bit(R5_LOCKED, &dev->flags);
2326 set_bit(R5_Wantread, &dev->flags);
2329 set_bit(STRIPE_DELAYED, &sh->state);
2330 set_bit(STRIPE_HANDLE, &sh->state);
2334 /* now if nothing is locked, and if we have enough data,
2335 * we can start a write request
2337 /* since handle_stripe can be called at any time we need to handle the
2338 * case where a compute block operation has been submitted and then a
2339 * subsequent call wants to start a write request. raid5_run_ops only
2340 * handles the case where compute block and postxor are requested
2341 * simultaneously. If this is not the case then new writes need to be
2342 * held off until the compute completes.
2344 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2345 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2346 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2347 schedule_reconstruction5(sh, s, rcw == 0, 0);
2350 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2351 struct stripe_head *sh, struct stripe_head_state *s,
2352 struct r6_state *r6s, int disks)
2354 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2355 int qd_idx = r6s->qd_idx;
2356 for (i = disks; i--; ) {
2357 struct r5dev *dev = &sh->dev[i];
2358 /* Would I have to read this buffer for reconstruct_write */
2359 if (!test_bit(R5_OVERWRITE, &dev->flags)
2360 && i != pd_idx && i != qd_idx
2361 && (!test_bit(R5_LOCKED, &dev->flags)
2363 !test_bit(R5_UPTODATE, &dev->flags)) {
2364 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2366 pr_debug("raid6: must_compute: "
2367 "disk %d flags=%#lx\n", i, dev->flags);
2372 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2373 (unsigned long long)sh->sector, rcw, must_compute);
2374 set_bit(STRIPE_HANDLE, &sh->state);
2377 /* want reconstruct write, but need to get some data */
2378 for (i = disks; i--; ) {
2379 struct r5dev *dev = &sh->dev[i];
2380 if (!test_bit(R5_OVERWRITE, &dev->flags)
2381 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2382 && !test_bit(R5_LOCKED, &dev->flags) &&
2383 !test_bit(R5_UPTODATE, &dev->flags) &&
2384 test_bit(R5_Insync, &dev->flags)) {
2386 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2387 pr_debug("Read_old stripe %llu "
2388 "block %d for Reconstruct\n",
2389 (unsigned long long)sh->sector, i);
2390 set_bit(R5_LOCKED, &dev->flags);
2391 set_bit(R5_Wantread, &dev->flags);
2394 pr_debug("Request delayed stripe %llu "
2395 "block %d for Reconstruct\n",
2396 (unsigned long long)sh->sector, i);
2397 set_bit(STRIPE_DELAYED, &sh->state);
2398 set_bit(STRIPE_HANDLE, &sh->state);
2402 /* now if nothing is locked, and if we have enough data, we can start a
2405 if (s->locked == 0 && rcw == 0 &&
2406 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2407 if (must_compute > 0) {
2408 /* We have failed blocks and need to compute them */
2409 switch (s->failed) {
2413 compute_block_1(sh, r6s->failed_num[0], 0);
2416 compute_block_2(sh, r6s->failed_num[0],
2417 r6s->failed_num[1]);
2419 default: /* This request should have been failed? */
2424 pr_debug("Computing parity for stripe %llu\n",
2425 (unsigned long long)sh->sector);
2426 compute_parity6(sh, RECONSTRUCT_WRITE);
2427 /* now every locked buffer is ready to be written */
2428 for (i = disks; i--; )
2429 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2430 pr_debug("Writing stripe %llu block %d\n",
2431 (unsigned long long)sh->sector, i);
2433 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2435 if (s->locked == disks)
2436 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2437 atomic_inc(&conf->pending_full_writes);
2438 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2439 set_bit(STRIPE_INSYNC, &sh->state);
2441 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2442 atomic_dec(&conf->preread_active_stripes);
2443 if (atomic_read(&conf->preread_active_stripes) <
2445 md_wakeup_thread(conf->mddev->thread);
2450 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2451 struct stripe_head_state *s, int disks)
2453 struct r5dev *dev = NULL;
2455 set_bit(STRIPE_HANDLE, &sh->state);
2457 switch (sh->check_state) {
2458 case check_state_idle:
2459 /* start a new check operation if there are no failures */
2460 if (s->failed == 0) {
2461 BUG_ON(s->uptodate != disks);
2462 sh->check_state = check_state_run;
2463 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2464 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2468 dev = &sh->dev[s->failed_num];
2470 case check_state_compute_result:
2471 sh->check_state = check_state_idle;
2473 dev = &sh->dev[sh->pd_idx];
2475 /* check that a write has not made the stripe insync */
2476 if (test_bit(STRIPE_INSYNC, &sh->state))
2479 /* either failed parity check, or recovery is happening */
2480 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2481 BUG_ON(s->uptodate != disks);
2483 set_bit(R5_LOCKED, &dev->flags);
2485 set_bit(R5_Wantwrite, &dev->flags);
2487 clear_bit(STRIPE_DEGRADED, &sh->state);
2488 set_bit(STRIPE_INSYNC, &sh->state);
2490 case check_state_run:
2491 break; /* we will be called again upon completion */
2492 case check_state_check_result:
2493 sh->check_state = check_state_idle;
2495 /* if a failure occurred during the check operation, leave
2496 * STRIPE_INSYNC not set and let the stripe be handled again
2501 /* handle a successful check operation, if parity is correct
2502 * we are done. Otherwise update the mismatch count and repair
2503 * parity if !MD_RECOVERY_CHECK
2505 if (sh->ops.zero_sum_result == 0)
2506 /* parity is correct (on disc,
2507 * not in buffer any more)
2509 set_bit(STRIPE_INSYNC, &sh->state);
2511 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2512 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2513 /* don't try to repair!! */
2514 set_bit(STRIPE_INSYNC, &sh->state);
2516 sh->check_state = check_state_compute_run;
2517 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2518 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2519 set_bit(R5_Wantcompute,
2520 &sh->dev[sh->pd_idx].flags);
2521 sh->ops.target = sh->pd_idx;
2526 case check_state_compute_run:
2529 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2530 __func__, sh->check_state,
2531 (unsigned long long) sh->sector);
2537 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2538 struct stripe_head_state *s,
2539 struct r6_state *r6s, struct page *tmp_page,
2542 int update_p = 0, update_q = 0;
2544 int pd_idx = sh->pd_idx;
2545 int qd_idx = r6s->qd_idx;
2547 set_bit(STRIPE_HANDLE, &sh->state);
2549 BUG_ON(s->failed > 2);
2550 BUG_ON(s->uptodate < disks);
2551 /* Want to check and possibly repair P and Q.
2552 * However there could be one 'failed' device, in which
2553 * case we can only check one of them, possibly using the
2554 * other to generate missing data
2557 /* If !tmp_page, we cannot do the calculations,
2558 * but as we have set STRIPE_HANDLE, we will soon be called
2559 * by stripe_handle with a tmp_page - just wait until then.
2562 if (s->failed == r6s->q_failed) {
2563 /* The only possible failed device holds 'Q', so it
2564 * makes sense to check P (If anything else were failed,
2565 * we would have used P to recreate it).
2567 compute_block_1(sh, pd_idx, 1);
2568 if (!page_is_zero(sh->dev[pd_idx].page)) {
2569 compute_block_1(sh, pd_idx, 0);
2573 if (!r6s->q_failed && s->failed < 2) {
2574 /* q is not failed, and we didn't use it to generate
2575 * anything, so it makes sense to check it
2577 memcpy(page_address(tmp_page),
2578 page_address(sh->dev[qd_idx].page),
2580 compute_parity6(sh, UPDATE_PARITY);
2581 if (memcmp(page_address(tmp_page),
2582 page_address(sh->dev[qd_idx].page),
2583 STRIPE_SIZE) != 0) {
2584 clear_bit(STRIPE_INSYNC, &sh->state);
2588 if (update_p || update_q) {
2589 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2590 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2591 /* don't try to repair!! */
2592 update_p = update_q = 0;
2595 /* now write out any block on a failed drive,
2596 * or P or Q if they need it
2599 if (s->failed == 2) {
2600 dev = &sh->dev[r6s->failed_num[1]];
2602 set_bit(R5_LOCKED, &dev->flags);
2603 set_bit(R5_Wantwrite, &dev->flags);
2605 if (s->failed >= 1) {
2606 dev = &sh->dev[r6s->failed_num[0]];
2608 set_bit(R5_LOCKED, &dev->flags);
2609 set_bit(R5_Wantwrite, &dev->flags);
2613 dev = &sh->dev[pd_idx];
2615 set_bit(R5_LOCKED, &dev->flags);
2616 set_bit(R5_Wantwrite, &dev->flags);
2619 dev = &sh->dev[qd_idx];
2621 set_bit(R5_LOCKED, &dev->flags);
2622 set_bit(R5_Wantwrite, &dev->flags);
2624 clear_bit(STRIPE_DEGRADED, &sh->state);
2626 set_bit(STRIPE_INSYNC, &sh->state);
2630 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2631 struct r6_state *r6s)
2635 /* We have read all the blocks in this stripe and now we need to
2636 * copy some of them into a target stripe for expand.
2638 struct dma_async_tx_descriptor *tx = NULL;
2639 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2640 for (i = 0; i < sh->disks; i++)
2641 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2643 struct stripe_head *sh2;
2645 sector_t bn = compute_blocknr(sh, i);
2646 sector_t s = raid5_compute_sector(conf, bn, 0,
2648 sh2 = get_active_stripe(conf, s, 0, 1);
2650 /* so far only the early blocks of this stripe
2651 * have been requested. When later blocks
2652 * get requested, we will try again
2655 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2656 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2657 /* must have already done this block */
2658 release_stripe(sh2);
2662 /* place all the copies on one channel */
2663 tx = async_memcpy(sh2->dev[dd_idx].page,
2664 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2665 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2667 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2668 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2669 for (j = 0; j < conf->raid_disks; j++)
2670 if (j != sh2->pd_idx &&
2671 (!r6s || j != sh2->qd_idx) &&
2672 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2674 if (j == conf->raid_disks) {
2675 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2676 set_bit(STRIPE_HANDLE, &sh2->state);
2678 release_stripe(sh2);
2681 /* done submitting copies, wait for them to complete */
2684 dma_wait_for_async_tx(tx);
2690 * handle_stripe - do things to a stripe.
2692 * We lock the stripe and then examine the state of various bits
2693 * to see what needs to be done.
2695 * return some read request which now have data
2696 * return some write requests which are safely on disc
2697 * schedule a read on some buffers
2698 * schedule a write of some buffers
2699 * return confirmation of parity correctness
2701 * buffers are taken off read_list or write_list, and bh_cache buffers
2702 * get BH_Lock set before the stripe lock is released.
2706 static bool handle_stripe5(struct stripe_head *sh)
2708 raid5_conf_t *conf = sh->raid_conf;
2709 int disks = sh->disks, i;
2710 struct bio *return_bi = NULL;
2711 struct stripe_head_state s;
2713 mdk_rdev_t *blocked_rdev = NULL;
2716 memset(&s, 0, sizeof(s));
2717 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2718 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2719 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2720 sh->reconstruct_state);
2722 spin_lock(&sh->lock);
2723 clear_bit(STRIPE_HANDLE, &sh->state);
2724 clear_bit(STRIPE_DELAYED, &sh->state);
2726 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2727 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2728 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2730 /* Now to look around and see what can be done */
2732 for (i=disks; i--; ) {
2734 struct r5dev *dev = &sh->dev[i];
2735 clear_bit(R5_Insync, &dev->flags);
2737 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2738 "written %p\n", i, dev->flags, dev->toread, dev->read,
2739 dev->towrite, dev->written);
2741 /* maybe we can request a biofill operation
2743 * new wantfill requests are only permitted while
2744 * ops_complete_biofill is guaranteed to be inactive
2746 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2747 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2748 set_bit(R5_Wantfill, &dev->flags);
2750 /* now count some things */
2751 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2752 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2753 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2755 if (test_bit(R5_Wantfill, &dev->flags))
2757 else if (dev->toread)
2761 if (!test_bit(R5_OVERWRITE, &dev->flags))
2766 rdev = rcu_dereference(conf->disks[i].rdev);
2767 if (blocked_rdev == NULL &&
2768 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2769 blocked_rdev = rdev;
2770 atomic_inc(&rdev->nr_pending);
2772 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2773 /* The ReadError flag will just be confusing now */
2774 clear_bit(R5_ReadError, &dev->flags);
2775 clear_bit(R5_ReWrite, &dev->flags);
2777 if (!rdev || !test_bit(In_sync, &rdev->flags)
2778 || test_bit(R5_ReadError, &dev->flags)) {
2782 set_bit(R5_Insync, &dev->flags);
2786 if (unlikely(blocked_rdev)) {
2787 if (s.syncing || s.expanding || s.expanded ||
2788 s.to_write || s.written) {
2789 set_bit(STRIPE_HANDLE, &sh->state);
2792 /* There is nothing for the blocked_rdev to block */
2793 rdev_dec_pending(blocked_rdev, conf->mddev);
2794 blocked_rdev = NULL;
2797 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2798 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2799 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2802 pr_debug("locked=%d uptodate=%d to_read=%d"
2803 " to_write=%d failed=%d failed_num=%d\n",
2804 s.locked, s.uptodate, s.to_read, s.to_write,
2805 s.failed, s.failed_num);
2806 /* check if the array has lost two devices and, if so, some requests might
2809 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2810 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2811 if (s.failed > 1 && s.syncing) {
2812 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2813 clear_bit(STRIPE_SYNCING, &sh->state);
2817 /* might be able to return some write requests if the parity block
2818 * is safe, or on a failed drive
2820 dev = &sh->dev[sh->pd_idx];
2822 ((test_bit(R5_Insync, &dev->flags) &&
2823 !test_bit(R5_LOCKED, &dev->flags) &&
2824 test_bit(R5_UPTODATE, &dev->flags)) ||
2825 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2826 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2828 /* Now we might consider reading some blocks, either to check/generate
2829 * parity, or to satisfy requests
2830 * or to load a block that is being partially written.
2832 if (s.to_read || s.non_overwrite ||
2833 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2834 handle_stripe_fill5(sh, &s, disks);
2836 /* Now we check to see if any write operations have recently
2840 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
2842 if (sh->reconstruct_state == reconstruct_state_drain_result ||
2843 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
2844 sh->reconstruct_state = reconstruct_state_idle;
2846 /* All the 'written' buffers and the parity block are ready to
2847 * be written back to disk
2849 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2850 for (i = disks; i--; ) {
2852 if (test_bit(R5_LOCKED, &dev->flags) &&
2853 (i == sh->pd_idx || dev->written)) {
2854 pr_debug("Writing block %d\n", i);
2855 set_bit(R5_Wantwrite, &dev->flags);
2858 if (!test_bit(R5_Insync, &dev->flags) ||
2859 (i == sh->pd_idx && s.failed == 0))
2860 set_bit(STRIPE_INSYNC, &sh->state);
2863 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2864 atomic_dec(&conf->preread_active_stripes);
2865 if (atomic_read(&conf->preread_active_stripes) <
2867 md_wakeup_thread(conf->mddev->thread);
2871 /* Now to consider new write requests and what else, if anything
2872 * should be read. We do not handle new writes when:
2873 * 1/ A 'write' operation (copy+xor) is already in flight.
2874 * 2/ A 'check' operation is in flight, as it may clobber the parity
2877 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
2878 handle_stripe_dirtying5(conf, sh, &s, disks);
2880 /* maybe we need to check and possibly fix the parity for this stripe
2881 * Any reads will already have been scheduled, so we just see if enough
2882 * data is available. The parity check is held off while parity
2883 * dependent operations are in flight.
2885 if (sh->check_state ||
2886 (s.syncing && s.locked == 0 &&
2887 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2888 !test_bit(STRIPE_INSYNC, &sh->state)))
2889 handle_parity_checks5(conf, sh, &s, disks);
2891 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2892 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2893 clear_bit(STRIPE_SYNCING, &sh->state);
2896 /* If the failed drive is just a ReadError, then we might need to progress
2897 * the repair/check process
2899 if (s.failed == 1 && !conf->mddev->ro &&
2900 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2901 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2902 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2904 dev = &sh->dev[s.failed_num];
2905 if (!test_bit(R5_ReWrite, &dev->flags)) {
2906 set_bit(R5_Wantwrite, &dev->flags);
2907 set_bit(R5_ReWrite, &dev->flags);
2908 set_bit(R5_LOCKED, &dev->flags);
2911 /* let's read it back */
2912 set_bit(R5_Wantread, &dev->flags);
2913 set_bit(R5_LOCKED, &dev->flags);
2918 /* Finish reconstruct operations initiated by the expansion process */
2919 if (sh->reconstruct_state == reconstruct_state_result) {
2920 sh->reconstruct_state = reconstruct_state_idle;
2921 clear_bit(STRIPE_EXPANDING, &sh->state);
2922 for (i = conf->raid_disks; i--; ) {
2923 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2924 set_bit(R5_LOCKED, &sh->dev[i].flags);
2929 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2930 !sh->reconstruct_state) {
2931 /* Need to write out all blocks after computing parity */
2932 sh->disks = conf->raid_disks;
2933 stripe_set_idx(sh->sector, conf, 0, sh);
2934 schedule_reconstruction5(sh, &s, 1, 1);
2935 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
2936 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2937 atomic_dec(&conf->reshape_stripes);
2938 wake_up(&conf->wait_for_overlap);
2939 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2942 if (s.expanding && s.locked == 0 &&
2943 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2944 handle_stripe_expansion(conf, sh, NULL);
2947 spin_unlock(&sh->lock);
2949 /* wait for this device to become unblocked */
2950 if (unlikely(blocked_rdev))
2951 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2954 raid5_run_ops(sh, s.ops_request);
2958 return_io(return_bi);
2960 return blocked_rdev == NULL;
2963 static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2965 raid5_conf_t *conf = sh->raid_conf;
2966 int disks = sh->disks;
2967 struct bio *return_bi = NULL;
2968 int i, pd_idx = sh->pd_idx;
2969 struct stripe_head_state s;
2970 struct r6_state r6s;
2971 struct r5dev *dev, *pdev, *qdev;
2972 mdk_rdev_t *blocked_rdev = NULL;
2974 r6s.qd_idx = sh->qd_idx;
2975 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2976 "pd_idx=%d, qd_idx=%d\n",
2977 (unsigned long long)sh->sector, sh->state,
2978 atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2979 memset(&s, 0, sizeof(s));
2981 spin_lock(&sh->lock);
2982 clear_bit(STRIPE_HANDLE, &sh->state);
2983 clear_bit(STRIPE_DELAYED, &sh->state);
2985 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2986 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2987 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2988 /* Now to look around and see what can be done */
2991 for (i=disks; i--; ) {
2994 clear_bit(R5_Insync, &dev->flags);
2996 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2997 i, dev->flags, dev->toread, dev->towrite, dev->written);
2998 /* maybe we can reply to a read */
2999 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3000 struct bio *rbi, *rbi2;
3001 pr_debug("Return read for disc %d\n", i);
3002 spin_lock_irq(&conf->device_lock);
3005 if (test_and_clear_bit(R5_Overlap, &dev->flags))
3006 wake_up(&conf->wait_for_overlap);
3007 spin_unlock_irq(&conf->device_lock);
3008 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3009 copy_data(0, rbi, dev->page, dev->sector);
3010 rbi2 = r5_next_bio(rbi, dev->sector);
3011 spin_lock_irq(&conf->device_lock);
3012 if (!raid5_dec_bi_phys_segments(rbi)) {
3013 rbi->bi_next = return_bi;
3016 spin_unlock_irq(&conf->device_lock);
3021 /* now count some things */
3022 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3023 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3030 if (!test_bit(R5_OVERWRITE, &dev->flags))
3035 rdev = rcu_dereference(conf->disks[i].rdev);
3036 if (blocked_rdev == NULL &&
3037 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3038 blocked_rdev = rdev;
3039 atomic_inc(&rdev->nr_pending);
3041 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3042 /* The ReadError flag will just be confusing now */
3043 clear_bit(R5_ReadError, &dev->flags);
3044 clear_bit(R5_ReWrite, &dev->flags);
3046 if (!rdev || !test_bit(In_sync, &rdev->flags)
3047 || test_bit(R5_ReadError, &dev->flags)) {
3049 r6s.failed_num[s.failed] = i;
3052 set_bit(R5_Insync, &dev->flags);
3056 if (unlikely(blocked_rdev)) {
3057 if (s.syncing || s.expanding || s.expanded ||
3058 s.to_write || s.written) {
3059 set_bit(STRIPE_HANDLE, &sh->state);
3062 /* There is nothing for the blocked_rdev to block */
3063 rdev_dec_pending(blocked_rdev, conf->mddev);
3064 blocked_rdev = NULL;
3067 pr_debug("locked=%d uptodate=%d to_read=%d"
3068 " to_write=%d failed=%d failed_num=%d,%d\n",
3069 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3070 r6s.failed_num[0], r6s.failed_num[1]);
3071 /* check if the array has lost >2 devices and, if so, some requests
3072 * might need to be failed
3074 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3075 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3076 if (s.failed > 2 && s.syncing) {
3077 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3078 clear_bit(STRIPE_SYNCING, &sh->state);
3083 * might be able to return some write requests if the parity blocks
3084 * are safe, or on a failed drive
3086 pdev = &sh->dev[pd_idx];
3087 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3088 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3089 qdev = &sh->dev[r6s.qd_idx];
3090 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3091 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3094 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3095 && !test_bit(R5_LOCKED, &pdev->flags)
3096 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3097 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3098 && !test_bit(R5_LOCKED, &qdev->flags)
3099 && test_bit(R5_UPTODATE, &qdev->flags)))))
3100 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3102 /* Now we might consider reading some blocks, either to check/generate
3103 * parity, or to satisfy requests
3104 * or to load a block that is being partially written.
3106 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3107 (s.syncing && (s.uptodate < disks)) || s.expanding)
3108 handle_stripe_fill6(sh, &s, &r6s, disks);
3110 /* now to consider writing and what else, if anything should be read */
3112 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3114 /* maybe we need to check and possibly fix the parity for this stripe
3115 * Any reads will already have been scheduled, so we just see if enough
3118 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3119 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3121 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3122 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3123 clear_bit(STRIPE_SYNCING, &sh->state);
3126 /* If the failed drives are just a ReadError, then we might need
3127 * to progress the repair/check process
3129 if (s.failed <= 2 && !conf->mddev->ro)
3130 for (i = 0; i < s.failed; i++) {
3131 dev = &sh->dev[r6s.failed_num[i]];
3132 if (test_bit(R5_ReadError, &dev->flags)
3133 && !test_bit(R5_LOCKED, &dev->flags)
3134 && test_bit(R5_UPTODATE, &dev->flags)
3136 if (!test_bit(R5_ReWrite, &dev->flags)) {
3137 set_bit(R5_Wantwrite, &dev->flags);
3138 set_bit(R5_ReWrite, &dev->flags);
3139 set_bit(R5_LOCKED, &dev->flags);
3141 /* let's read it back */
3142 set_bit(R5_Wantread, &dev->flags);
3143 set_bit(R5_LOCKED, &dev->flags);
3148 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3149 /* Need to write out all blocks after computing P&Q */
3150 sh->disks = conf->raid_disks;
3151 stripe_set_idx(sh->sector, conf, 0, sh);
3152 compute_parity6(sh, RECONSTRUCT_WRITE);
3153 for (i = conf->raid_disks ; i-- ; ) {
3154 set_bit(R5_LOCKED, &sh->dev[i].flags);
3156 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3158 clear_bit(STRIPE_EXPANDING, &sh->state);
3159 } else if (s.expanded) {
3160 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3161 atomic_dec(&conf->reshape_stripes);
3162 wake_up(&conf->wait_for_overlap);
3163 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3166 if (s.expanding && s.locked == 0 &&
3167 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3168 handle_stripe_expansion(conf, sh, &r6s);
3171 spin_unlock(&sh->lock);
3173 /* wait for this device to become unblocked */
3174 if (unlikely(blocked_rdev))
3175 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3179 return_io(return_bi);
3181 return blocked_rdev == NULL;
3184 /* returns true if the stripe was handled */
3185 static bool handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3187 if (sh->raid_conf->level == 6)
3188 return handle_stripe6(sh, tmp_page);
3190 return handle_stripe5(sh);
3195 static void raid5_activate_delayed(raid5_conf_t *conf)
3197 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3198 while (!list_empty(&conf->delayed_list)) {
3199 struct list_head *l = conf->delayed_list.next;
3200 struct stripe_head *sh;
3201 sh = list_entry(l, struct stripe_head, lru);
3203 clear_bit(STRIPE_DELAYED, &sh->state);
3204 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3205 atomic_inc(&conf->preread_active_stripes);
3206 list_add_tail(&sh->lru, &conf->hold_list);
3209 blk_plug_device(conf->mddev->queue);
3212 static void activate_bit_delay(raid5_conf_t *conf)
3214 /* device_lock is held */
3215 struct list_head head;
3216 list_add(&head, &conf->bitmap_list);
3217 list_del_init(&conf->bitmap_list);
3218 while (!list_empty(&head)) {
3219 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3220 list_del_init(&sh->lru);
3221 atomic_inc(&sh->count);
3222 __release_stripe(conf, sh);
3226 static void unplug_slaves(mddev_t *mddev)
3228 raid5_conf_t *conf = mddev_to_conf(mddev);
3232 for (i=0; i<mddev->raid_disks; i++) {
3233 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3234 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3235 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3237 atomic_inc(&rdev->nr_pending);
3240 blk_unplug(r_queue);
3242 rdev_dec_pending(rdev, mddev);
3249 static void raid5_unplug_device(struct request_queue *q)
3251 mddev_t *mddev = q->queuedata;
3252 raid5_conf_t *conf = mddev_to_conf(mddev);
3253 unsigned long flags;
3255 spin_lock_irqsave(&conf->device_lock, flags);
3257 if (blk_remove_plug(q)) {
3259 raid5_activate_delayed(conf);
3261 md_wakeup_thread(mddev->thread);
3263 spin_unlock_irqrestore(&conf->device_lock, flags);
3265 unplug_slaves(mddev);
3268 static int raid5_congested(void *data, int bits)
3270 mddev_t *mddev = data;
3271 raid5_conf_t *conf = mddev_to_conf(mddev);
3273 /* No difference between reads and writes. Just check
3274 * how busy the stripe_cache is
3276 if (conf->inactive_blocked)
3280 if (list_empty_careful(&conf->inactive_list))
3286 /* We want read requests to align with chunks where possible,
3287 * but write requests don't need to.
3289 static int raid5_mergeable_bvec(struct request_queue *q,
3290 struct bvec_merge_data *bvm,
3291 struct bio_vec *biovec)
3293 mddev_t *mddev = q->queuedata;
3294 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3296 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3297 unsigned int bio_sectors = bvm->bi_size >> 9;
3299 if ((bvm->bi_rw & 1) == WRITE)
3300 return biovec->bv_len; /* always allow writes to be mergeable */
3302 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3303 if (max < 0) max = 0;
3304 if (max <= biovec->bv_len && bio_sectors == 0)
3305 return biovec->bv_len;
3311 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3313 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3314 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3315 unsigned int bio_sectors = bio->bi_size >> 9;
3317 return chunk_sectors >=
3318 ((sector & (chunk_sectors - 1)) + bio_sectors);
3322 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3323 * later sampled by raid5d.
3325 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3327 unsigned long flags;
3329 spin_lock_irqsave(&conf->device_lock, flags);
3331 bi->bi_next = conf->retry_read_aligned_list;
3332 conf->retry_read_aligned_list = bi;
3334 spin_unlock_irqrestore(&conf->device_lock, flags);
3335 md_wakeup_thread(conf->mddev->thread);
3339 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3343 bi = conf->retry_read_aligned;
3345 conf->retry_read_aligned = NULL;
3348 bi = conf->retry_read_aligned_list;
3350 conf->retry_read_aligned_list = bi->bi_next;
3353 * this sets the active strip count to 1 and the processed
3354 * strip count to zero (upper 8 bits)
3356 bi->bi_phys_segments = 1; /* biased count of active stripes */
3364 * The "raid5_align_endio" should check if the read succeeded and if it
3365 * did, call bio_endio on the original bio (having bio_put the new bio
3367 * If the read failed..
3369 static void raid5_align_endio(struct bio *bi, int error)
3371 struct bio* raid_bi = bi->bi_private;
3374 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3379 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3380 conf = mddev_to_conf(mddev);
3381 rdev = (void*)raid_bi->bi_next;
3382 raid_bi->bi_next = NULL;
3384 rdev_dec_pending(rdev, conf->mddev);
3386 if (!error && uptodate) {
3387 bio_endio(raid_bi, 0);
3388 if (atomic_dec_and_test(&conf->active_aligned_reads))
3389 wake_up(&conf->wait_for_stripe);
3394 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3396 add_bio_to_retry(raid_bi, conf);
3399 static int bio_fits_rdev(struct bio *bi)
3401 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3403 if ((bi->bi_size>>9) > q->max_sectors)
3405 blk_recount_segments(q, bi);
3406 if (bi->bi_phys_segments > q->max_phys_segments)
3409 if (q->merge_bvec_fn)
3410 /* it's too hard to apply the merge_bvec_fn at this stage,
3419 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3421 mddev_t *mddev = q->queuedata;
3422 raid5_conf_t *conf = mddev_to_conf(mddev);
3423 unsigned int dd_idx;
3424 struct bio* align_bi;
3427 if (!in_chunk_boundary(mddev, raid_bio)) {
3428 pr_debug("chunk_aligned_read : non aligned\n");
3432 * use bio_clone to make a copy of the bio
3434 align_bi = bio_clone(raid_bio, GFP_NOIO);
3438 * set bi_end_io to a new function, and set bi_private to the
3441 align_bi->bi_end_io = raid5_align_endio;
3442 align_bi->bi_private = raid_bio;
3446 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3451 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3452 if (rdev && test_bit(In_sync, &rdev->flags)) {
3453 atomic_inc(&rdev->nr_pending);
3455 raid_bio->bi_next = (void*)rdev;
3456 align_bi->bi_bdev = rdev->bdev;
3457 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3458 align_bi->bi_sector += rdev->data_offset;
3460 if (!bio_fits_rdev(align_bi)) {
3461 /* too big in some way */
3463 rdev_dec_pending(rdev, mddev);
3467 spin_lock_irq(&conf->device_lock);
3468 wait_event_lock_irq(conf->wait_for_stripe,
3470 conf->device_lock, /* nothing */);
3471 atomic_inc(&conf->active_aligned_reads);
3472 spin_unlock_irq(&conf->device_lock);
3474 generic_make_request(align_bi);
3483 /* __get_priority_stripe - get the next stripe to process
3485 * Full stripe writes are allowed to pass preread active stripes up until
3486 * the bypass_threshold is exceeded. In general the bypass_count
3487 * increments when the handle_list is handled before the hold_list; however, it
3488 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3489 * stripe with in flight i/o. The bypass_count will be reset when the
3490 * head of the hold_list has changed, i.e. the head was promoted to the
3493 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3495 struct stripe_head *sh;
3497 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3499 list_empty(&conf->handle_list) ? "empty" : "busy",
3500 list_empty(&conf->hold_list) ? "empty" : "busy",
3501 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3503 if (!list_empty(&conf->handle_list)) {
3504 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3506 if (list_empty(&conf->hold_list))
3507 conf->bypass_count = 0;
3508 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3509 if (conf->hold_list.next == conf->last_hold)
3510 conf->bypass_count++;
3512 conf->last_hold = conf->hold_list.next;
3513 conf->bypass_count -= conf->bypass_threshold;
3514 if (conf->bypass_count < 0)
3515 conf->bypass_count = 0;
3518 } else if (!list_empty(&conf->hold_list) &&
3519 ((conf->bypass_threshold &&
3520 conf->bypass_count > conf->bypass_threshold) ||
3521 atomic_read(&conf->pending_full_writes) == 0)) {
3522 sh = list_entry(conf->hold_list.next,
3524 conf->bypass_count -= conf->bypass_threshold;
3525 if (conf->bypass_count < 0)
3526 conf->bypass_count = 0;
3530 list_del_init(&sh->lru);
3531 atomic_inc(&sh->count);
3532 BUG_ON(atomic_read(&sh->count) != 1);
3536 static int make_request(struct request_queue *q, struct bio * bi)
3538 mddev_t *mddev = q->queuedata;
3539 raid5_conf_t *conf = mddev_to_conf(mddev);
3541 sector_t new_sector;
3542 sector_t logical_sector, last_sector;
3543 struct stripe_head *sh;
3544 const int rw = bio_data_dir(bi);
3547 if (unlikely(bio_barrier(bi))) {
3548 bio_endio(bi, -EOPNOTSUPP);
3552 md_write_start(mddev, bi);
3554 cpu = part_stat_lock();
3555 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3556 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3561 mddev->reshape_position == MaxSector &&
3562 chunk_aligned_read(q,bi))
3565 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3566 last_sector = bi->bi_sector + (bi->bi_size>>9);
3568 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3570 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3572 int disks, data_disks;
3577 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3578 if (likely(conf->expand_progress == MaxSector))
3579 disks = conf->raid_disks;
3581 /* spinlock is needed as expand_progress may be
3582 * 64bit on a 32bit platform, and so it might be
3583 * possible to see a half-updated value
3584 * Ofcourse expand_progress could change after
3585 * the lock is dropped, so once we get a reference
3586 * to the stripe that we think it is, we will have
3589 spin_lock_irq(&conf->device_lock);
3590 disks = conf->raid_disks;
3591 if (logical_sector >= conf->expand_progress) {
3592 disks = conf->previous_raid_disks;
3595 if (logical_sector >= conf->expand_lo) {
3596 spin_unlock_irq(&conf->device_lock);
3601 spin_unlock_irq(&conf->device_lock);
3603 data_disks = disks - conf->max_degraded;
3605 new_sector = raid5_compute_sector(conf, logical_sector,
3608 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3609 (unsigned long long)new_sector,
3610 (unsigned long long)logical_sector);
3612 sh = get_active_stripe(conf, new_sector, previous,
3613 (bi->bi_rw&RWA_MASK));
3615 if (unlikely(conf->expand_progress != MaxSector)) {
3616 /* expansion might have moved on while waiting for a
3617 * stripe, so we must do the range check again.
3618 * Expansion could still move past after this
3619 * test, but as we are holding a reference to
3620 * 'sh', we know that if that happens,
3621 * STRIPE_EXPANDING will get set and the expansion
3622 * won't proceed until we finish with the stripe.
3625 spin_lock_irq(&conf->device_lock);
3626 if (logical_sector < conf->expand_progress &&
3627 disks == conf->previous_raid_disks)
3628 /* mismatch, need to try again */
3630 spin_unlock_irq(&conf->device_lock);
3636 /* FIXME what if we get a false positive because these
3637 * are being updated.
3639 if (logical_sector >= mddev->suspend_lo &&
3640 logical_sector < mddev->suspend_hi) {
3646 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3647 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3648 /* Stripe is busy expanding or
3649 * add failed due to overlap. Flush everything
3652 raid5_unplug_device(mddev->queue);
3657 finish_wait(&conf->wait_for_overlap, &w);
3658 set_bit(STRIPE_HANDLE, &sh->state);
3659 clear_bit(STRIPE_DELAYED, &sh->state);
3662 /* cannot get stripe for read-ahead, just give-up */
3663 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3664 finish_wait(&conf->wait_for_overlap, &w);
3669 spin_lock_irq(&conf->device_lock);
3670 remaining = raid5_dec_bi_phys_segments(bi);
3671 spin_unlock_irq(&conf->device_lock);
3672 if (remaining == 0) {
3675 md_write_end(mddev);
3682 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3684 /* reshaping is quite different to recovery/resync so it is
3685 * handled quite separately ... here.
3687 * On each call to sync_request, we gather one chunk worth of
3688 * destination stripes and flag them as expanding.
3689 * Then we find all the source stripes and request reads.
3690 * As the reads complete, handle_stripe will copy the data
3691 * into the destination stripe and release that stripe.
3693 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3694 struct stripe_head *sh;
3695 sector_t first_sector, last_sector;
3696 int raid_disks = conf->previous_raid_disks;
3697 int data_disks = raid_disks - conf->max_degraded;
3698 int new_data_disks = conf->raid_disks - conf->max_degraded;
3701 sector_t writepos, safepos, gap;
3703 if (sector_nr == 0 &&
3704 conf->expand_progress != 0) {
3705 /* restarting in the middle, skip the initial sectors */
3706 sector_nr = conf->expand_progress;
3707 sector_div(sector_nr, new_data_disks);
3712 /* we update the metadata when there is more than 3Meg
3713 * in the block range (that is rather arbitrary, should
3714 * probably be time based) or when the data about to be
3715 * copied would over-write the source of the data at
3716 * the front of the range.
3717 * i.e. one new_stripe forward from expand_progress new_maps
3718 * to after where expand_lo old_maps to
3720 writepos = conf->expand_progress +
3721 conf->chunk_size/512*(new_data_disks);
3722 sector_div(writepos, new_data_disks);
3723 safepos = conf->expand_lo;
3724 sector_div(safepos, data_disks);
3725 gap = conf->expand_progress - conf->expand_lo;
3727 if (writepos >= safepos ||
3728 gap > (new_data_disks)*3000*2 /*3Meg*/) {
3729 /* Cannot proceed until we've updated the superblock... */
3730 wait_event(conf->wait_for_overlap,
3731 atomic_read(&conf->reshape_stripes)==0);
3732 mddev->reshape_position = conf->expand_progress;
3733 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3734 md_wakeup_thread(mddev->thread);
3735 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3736 kthread_should_stop());
3737 spin_lock_irq(&conf->device_lock);
3738 conf->expand_lo = mddev->reshape_position;
3739 spin_unlock_irq(&conf->device_lock);
3740 wake_up(&conf->wait_for_overlap);
3743 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3746 sh = get_active_stripe(conf, sector_nr+i, 0, 0);
3747 set_bit(STRIPE_EXPANDING, &sh->state);
3748 atomic_inc(&conf->reshape_stripes);
3749 /* If any of this stripe is beyond the end of the old
3750 * array, then we need to zero those blocks
3752 for (j=sh->disks; j--;) {
3754 if (j == sh->pd_idx)
3756 if (conf->level == 6 &&
3759 s = compute_blocknr(sh, j);
3760 if (s < mddev->array_sectors) {
3764 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3765 set_bit(R5_Expanded, &sh->dev[j].flags);
3766 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3769 set_bit(STRIPE_EXPAND_READY, &sh->state);
3770 set_bit(STRIPE_HANDLE, &sh->state);
3774 spin_lock_irq(&conf->device_lock);
3775 conf->expand_progress = (sector_nr + i) * new_data_disks;
3776 spin_unlock_irq(&conf->device_lock);
3777 /* Ok, those stripe are ready. We can start scheduling
3778 * reads on the source stripes.
3779 * The source stripes are determined by mapping the first and last
3780 * block on the destination stripes.
3783 raid5_compute_sector(conf, sector_nr*(new_data_disks),
3786 raid5_compute_sector(conf, ((sector_nr+conf->chunk_size/512)
3787 *(new_data_disks) - 1),
3789 if (last_sector >= mddev->dev_sectors)
3790 last_sector = mddev->dev_sectors - 1;
3791 while (first_sector <= last_sector) {
3792 sh = get_active_stripe(conf, first_sector, 1, 0);
3793 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3794 set_bit(STRIPE_HANDLE, &sh->state);
3796 first_sector += STRIPE_SECTORS;
3798 /* If this takes us to the resync_max point where we have to pause,
3799 * then we need to write out the superblock.
3801 sector_nr += conf->chunk_size>>9;
3802 if (sector_nr >= mddev->resync_max) {
3803 /* Cannot proceed until we've updated the superblock... */
3804 wait_event(conf->wait_for_overlap,
3805 atomic_read(&conf->reshape_stripes) == 0);
3806 mddev->reshape_position = conf->expand_progress;
3807 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3808 md_wakeup_thread(mddev->thread);
3809 wait_event(mddev->sb_wait,
3810 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3811 || kthread_should_stop());
3812 spin_lock_irq(&conf->device_lock);
3813 conf->expand_lo = mddev->reshape_position;
3814 spin_unlock_irq(&conf->device_lock);
3815 wake_up(&conf->wait_for_overlap);
3817 return conf->chunk_size>>9;
3820 /* FIXME go_faster isn't used */
3821 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3823 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3824 struct stripe_head *sh;
3825 sector_t max_sector = mddev->dev_sectors;
3827 int still_degraded = 0;
3830 if (sector_nr >= max_sector) {
3831 /* just being told to finish up .. nothing much to do */
3832 unplug_slaves(mddev);
3833 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3838 if (mddev->curr_resync < max_sector) /* aborted */
3839 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3841 else /* completed sync */
3843 bitmap_close_sync(mddev->bitmap);
3848 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3849 return reshape_request(mddev, sector_nr, skipped);
3851 /* No need to check resync_max as we never do more than one
3852 * stripe, and as resync_max will always be on a chunk boundary,
3853 * if the check in md_do_sync didn't fire, there is no chance
3854 * of overstepping resync_max here
3857 /* if there is too many failed drives and we are trying
3858 * to resync, then assert that we are finished, because there is
3859 * nothing we can do.
3861 if (mddev->degraded >= conf->max_degraded &&
3862 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3863 sector_t rv = mddev->dev_sectors - sector_nr;
3867 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3868 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3869 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3870 /* we can skip this block, and probably more */
3871 sync_blocks /= STRIPE_SECTORS;
3873 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3877 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3879 sh = get_active_stripe(conf, sector_nr, 0, 1);
3881 sh = get_active_stripe(conf, sector_nr, 0, 0);
3882 /* make sure we don't swamp the stripe cache if someone else
3883 * is trying to get access
3885 schedule_timeout_uninterruptible(1);
3887 /* Need to check if array will still be degraded after recovery/resync
3888 * We don't need to check the 'failed' flag as when that gets set,
3891 for (i=0; i<mddev->raid_disks; i++)
3892 if (conf->disks[i].rdev == NULL)
3895 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3897 spin_lock(&sh->lock);
3898 set_bit(STRIPE_SYNCING, &sh->state);
3899 clear_bit(STRIPE_INSYNC, &sh->state);
3900 spin_unlock(&sh->lock);
3902 /* wait for any blocked device to be handled */
3903 while(unlikely(!handle_stripe(sh, NULL)))
3907 return STRIPE_SECTORS;
3910 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3912 /* We may not be able to submit a whole bio at once as there
3913 * may not be enough stripe_heads available.
3914 * We cannot pre-allocate enough stripe_heads as we may need
3915 * more than exist in the cache (if we allow ever large chunks).
3916 * So we do one stripe head at a time and record in
3917 * ->bi_hw_segments how many have been done.
3919 * We *know* that this entire raid_bio is in one chunk, so
3920 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3922 struct stripe_head *sh;
3924 sector_t sector, logical_sector, last_sector;
3929 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3930 sector = raid5_compute_sector(conf, logical_sector,
3932 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3934 for (; logical_sector < last_sector;
3935 logical_sector += STRIPE_SECTORS,
3936 sector += STRIPE_SECTORS,
3939 if (scnt < raid5_bi_hw_segments(raid_bio))
3940 /* already done this stripe */
3943 sh = get_active_stripe(conf, sector, 0, 1);
3946 /* failed to get a stripe - must wait */
3947 raid5_set_bi_hw_segments(raid_bio, scnt);
3948 conf->retry_read_aligned = raid_bio;
3952 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3953 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3955 raid5_set_bi_hw_segments(raid_bio, scnt);
3956 conf->retry_read_aligned = raid_bio;
3960 handle_stripe(sh, NULL);
3964 spin_lock_irq(&conf->device_lock);
3965 remaining = raid5_dec_bi_phys_segments(raid_bio);
3966 spin_unlock_irq(&conf->device_lock);
3968 bio_endio(raid_bio, 0);
3969 if (atomic_dec_and_test(&conf->active_aligned_reads))
3970 wake_up(&conf->wait_for_stripe);
3977 * This is our raid5 kernel thread.
3979 * We scan the hash table for stripes which can be handled now.
3980 * During the scan, completed stripes are saved for us by the interrupt
3981 * handler, so that they will not have to wait for our next wakeup.
3983 static void raid5d(mddev_t *mddev)
3985 struct stripe_head *sh;
3986 raid5_conf_t *conf = mddev_to_conf(mddev);
3989 pr_debug("+++ raid5d active\n");
3991 md_check_recovery(mddev);
3994 spin_lock_irq(&conf->device_lock);
3998 if (conf->seq_flush != conf->seq_write) {
3999 int seq = conf->seq_flush;
4000 spin_unlock_irq(&conf->device_lock);
4001 bitmap_unplug(mddev->bitmap);
4002 spin_lock_irq(&conf->device_lock);
4003 conf->seq_write = seq;
4004 activate_bit_delay(conf);
4007 while ((bio = remove_bio_from_retry(conf))) {
4009 spin_unlock_irq(&conf->device_lock);
4010 ok = retry_aligned_read(conf, bio);
4011 spin_lock_irq(&conf->device_lock);
4017 sh = __get_priority_stripe(conf);
4021 spin_unlock_irq(&conf->device_lock);
4024 handle_stripe(sh, conf->spare_page);
4027 spin_lock_irq(&conf->device_lock);
4029 pr_debug("%d stripes handled\n", handled);
4031 spin_unlock_irq(&conf->device_lock);
4033 async_tx_issue_pending_all();
4034 unplug_slaves(mddev);
4036 pr_debug("--- raid5d inactive\n");
4040 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4042 raid5_conf_t *conf = mddev_to_conf(mddev);
4044 return sprintf(page, "%d\n", conf->max_nr_stripes);
4050 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4052 raid5_conf_t *conf = mddev_to_conf(mddev);
4056 if (len >= PAGE_SIZE)
4061 if (strict_strtoul(page, 10, &new))
4063 if (new <= 16 || new > 32768)
4065 while (new < conf->max_nr_stripes) {
4066 if (drop_one_stripe(conf))
4067 conf->max_nr_stripes--;
4071 err = md_allow_write(mddev);
4074 while (new > conf->max_nr_stripes) {
4075 if (grow_one_stripe(conf))
4076 conf->max_nr_stripes++;
4082 static struct md_sysfs_entry
4083 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4084 raid5_show_stripe_cache_size,
4085 raid5_store_stripe_cache_size);
4088 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4090 raid5_conf_t *conf = mddev_to_conf(mddev);
4092 return sprintf(page, "%d\n", conf->bypass_threshold);
4098 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4100 raid5_conf_t *conf = mddev_to_conf(mddev);
4102 if (len >= PAGE_SIZE)
4107 if (strict_strtoul(page, 10, &new))
4109 if (new > conf->max_nr_stripes)
4111 conf->bypass_threshold = new;
4115 static struct md_sysfs_entry
4116 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4118 raid5_show_preread_threshold,
4119 raid5_store_preread_threshold);
4122 stripe_cache_active_show(mddev_t *mddev, char *page)
4124 raid5_conf_t *conf = mddev_to_conf(mddev);
4126 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4131 static struct md_sysfs_entry
4132 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4134 static struct attribute *raid5_attrs[] = {
4135 &raid5_stripecache_size.attr,
4136 &raid5_stripecache_active.attr,
4137 &raid5_preread_bypass_threshold.attr,
4140 static struct attribute_group raid5_attrs_group = {
4142 .attrs = raid5_attrs,
4145 static int run(mddev_t *mddev)
4148 int raid_disk, memory;
4150 struct disk_info *disk;
4151 int working_disks = 0;
4153 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4154 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4155 mdname(mddev), mddev->level);
4158 if ((mddev->level == 5 && !algorithm_valid_raid5(mddev->layout)) ||
4159 (mddev->level == 6 && !algorithm_valid_raid6(mddev->layout))) {
4160 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4161 mdname(mddev), mddev->layout);
4165 if (mddev->chunk_size < PAGE_SIZE) {
4166 printk(KERN_ERR "md/raid5: chunk_size must be at least "
4167 "PAGE_SIZE but %d < %ld\n",
4168 mddev->chunk_size, PAGE_SIZE);
4172 if (mddev->reshape_position != MaxSector) {
4173 /* Check that we can continue the reshape.
4174 * Currently only disks can change, it must
4175 * increase, and we must be past the point where
4176 * a stripe over-writes itself
4178 sector_t here_new, here_old;
4180 int max_degraded = (mddev->level == 5 ? 1 : 2);
4182 if (mddev->new_level != mddev->level ||
4183 mddev->new_layout != mddev->layout ||
4184 mddev->new_chunk != mddev->chunk_size) {
4185 printk(KERN_ERR "raid5: %s: unsupported reshape "
4186 "required - aborting.\n",
4190 if (mddev->delta_disks <= 0) {
4191 printk(KERN_ERR "raid5: %s: unsupported reshape "
4192 "(reduce disks) required - aborting.\n",
4196 old_disks = mddev->raid_disks - mddev->delta_disks;
4197 /* reshape_position must be on a new-stripe boundary, and one
4198 * further up in new geometry must map after here in old
4201 here_new = mddev->reshape_position;
4202 if (sector_div(here_new, (mddev->chunk_size>>9)*
4203 (mddev->raid_disks - max_degraded))) {
4204 printk(KERN_ERR "raid5: reshape_position not "
4205 "on a stripe boundary\n");
4208 /* here_new is the stripe we will write to */
4209 here_old = mddev->reshape_position;
4210 sector_div(here_old, (mddev->chunk_size>>9)*
4211 (old_disks-max_degraded));
4212 /* here_old is the first stripe that we might need to read
4214 if (here_new >= here_old) {
4215 /* Reading from the same stripe as writing to - bad */
4216 printk(KERN_ERR "raid5: reshape_position too early for "
4217 "auto-recovery - aborting.\n");
4220 printk(KERN_INFO "raid5: reshape will continue\n");
4221 /* OK, we should be able to continue; */
4225 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4226 if ((conf = mddev->private) == NULL)
4228 if (mddev->reshape_position == MaxSector) {
4229 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4231 conf->raid_disks = mddev->raid_disks;
4232 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4235 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4240 conf->mddev = mddev;
4242 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4245 if (mddev->level == 6) {
4246 conf->spare_page = alloc_page(GFP_KERNEL);
4247 if (!conf->spare_page)
4250 spin_lock_init(&conf->device_lock);
4251 mddev->queue->queue_lock = &conf->device_lock;
4252 init_waitqueue_head(&conf->wait_for_stripe);
4253 init_waitqueue_head(&conf->wait_for_overlap);
4254 INIT_LIST_HEAD(&conf->handle_list);
4255 INIT_LIST_HEAD(&conf->hold_list);
4256 INIT_LIST_HEAD(&conf->delayed_list);
4257 INIT_LIST_HEAD(&conf->bitmap_list);
4258 INIT_LIST_HEAD(&conf->inactive_list);
4259 atomic_set(&conf->active_stripes, 0);
4260 atomic_set(&conf->preread_active_stripes, 0);
4261 atomic_set(&conf->active_aligned_reads, 0);
4262 conf->bypass_threshold = BYPASS_THRESHOLD;
4264 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4266 list_for_each_entry(rdev, &mddev->disks, same_set) {
4267 raid_disk = rdev->raid_disk;
4268 if (raid_disk >= conf->raid_disks
4271 disk = conf->disks + raid_disk;
4275 if (test_bit(In_sync, &rdev->flags)) {
4276 char b[BDEVNAME_SIZE];
4277 printk(KERN_INFO "raid5: device %s operational as raid"
4278 " disk %d\n", bdevname(rdev->bdev,b),
4282 /* Cannot rely on bitmap to complete recovery */
4287 * 0 for a fully functional array, 1 or 2 for a degraded array.
4289 mddev->degraded = conf->raid_disks - working_disks;
4290 conf->mddev = mddev;
4291 conf->chunk_size = mddev->chunk_size;
4292 conf->level = mddev->level;
4293 if (conf->level == 6)
4294 conf->max_degraded = 2;
4296 conf->max_degraded = 1;
4297 conf->algorithm = mddev->layout;
4298 conf->max_nr_stripes = NR_STRIPES;
4299 conf->expand_progress = mddev->reshape_position;
4301 /* device size must be a multiple of chunk size */
4302 mddev->dev_sectors &= ~(mddev->chunk_size / 512 - 1);
4303 mddev->resync_max_sectors = mddev->dev_sectors;
4305 if (conf->level == 6 && conf->raid_disks < 4) {
4306 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4307 mdname(mddev), conf->raid_disks);
4310 if (!conf->chunk_size || conf->chunk_size % 4) {
4311 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4312 conf->chunk_size, mdname(mddev));
4315 if (mddev->degraded > conf->max_degraded) {
4316 printk(KERN_ERR "raid5: not enough operational devices for %s"
4317 " (%d/%d failed)\n",
4318 mdname(mddev), mddev->degraded, conf->raid_disks);
4322 if (mddev->degraded > 0 &&
4323 mddev->recovery_cp != MaxSector) {
4324 if (mddev->ok_start_degraded)
4326 "raid5: starting dirty degraded array: %s"
4327 "- data corruption possible.\n",
4331 "raid5: cannot start dirty degraded array for %s\n",
4338 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4339 if (!mddev->thread) {
4341 "raid5: couldn't allocate thread for %s\n",
4346 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4347 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4348 if (grow_stripes(conf, conf->max_nr_stripes)) {
4350 "raid5: couldn't allocate %dkB for buffers\n", memory);
4351 shrink_stripes(conf);
4352 md_unregister_thread(mddev->thread);
4355 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4356 memory, mdname(mddev));
4358 if (mddev->degraded == 0)
4359 printk("raid5: raid level %d set %s active with %d out of %d"
4360 " devices, algorithm %d\n", conf->level, mdname(mddev),
4361 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4364 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4365 " out of %d devices, algorithm %d\n", conf->level,
4366 mdname(mddev), mddev->raid_disks - mddev->degraded,
4367 mddev->raid_disks, conf->algorithm);
4369 print_raid5_conf(conf);
4371 if (conf->expand_progress != MaxSector) {
4372 printk("...ok start reshape thread\n");
4373 conf->expand_lo = conf->expand_progress;
4374 atomic_set(&conf->reshape_stripes, 0);
4375 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4376 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4377 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4378 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4379 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4383 /* read-ahead size must cover two whole stripes, which is
4384 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4387 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4388 int stripe = data_disks *
4389 (mddev->chunk_size / PAGE_SIZE);
4390 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4391 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4394 /* Ok, everything is just fine now */
4395 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4397 "raid5: failed to create sysfs attributes for %s\n",
4400 mddev->queue->unplug_fn = raid5_unplug_device;
4401 mddev->queue->backing_dev_info.congested_data = mddev;
4402 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4404 mddev->array_sectors = mddev->dev_sectors *
4405 (conf->previous_raid_disks - conf->max_degraded);
4407 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4412 print_raid5_conf(conf);
4413 safe_put_page(conf->spare_page);
4415 kfree(conf->stripe_hashtbl);
4418 mddev->private = NULL;
4419 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4425 static int stop(mddev_t *mddev)
4427 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4429 md_unregister_thread(mddev->thread);
4430 mddev->thread = NULL;
4431 shrink_stripes(conf);
4432 kfree(conf->stripe_hashtbl);
4433 mddev->queue->backing_dev_info.congested_fn = NULL;
4434 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4435 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4438 mddev->private = NULL;
4443 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4447 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4448 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4449 seq_printf(seq, "sh %llu, count %d.\n",
4450 (unsigned long long)sh->sector, atomic_read(&sh->count));
4451 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4452 for (i = 0; i < sh->disks; i++) {
4453 seq_printf(seq, "(cache%d: %p %ld) ",
4454 i, sh->dev[i].page, sh->dev[i].flags);
4456 seq_printf(seq, "\n");
4459 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4461 struct stripe_head *sh;
4462 struct hlist_node *hn;
4465 spin_lock_irq(&conf->device_lock);
4466 for (i = 0; i < NR_HASH; i++) {
4467 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4468 if (sh->raid_conf != conf)
4473 spin_unlock_irq(&conf->device_lock);
4477 static void status(struct seq_file *seq, mddev_t *mddev)
4479 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4482 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4483 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4484 for (i = 0; i < conf->raid_disks; i++)
4485 seq_printf (seq, "%s",
4486 conf->disks[i].rdev &&
4487 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4488 seq_printf (seq, "]");
4490 seq_printf (seq, "\n");
4491 printall(seq, conf);
4495 static void print_raid5_conf (raid5_conf_t *conf)
4498 struct disk_info *tmp;
4500 printk("RAID5 conf printout:\n");
4502 printk("(conf==NULL)\n");
4505 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4506 conf->raid_disks - conf->mddev->degraded);
4508 for (i = 0; i < conf->raid_disks; i++) {
4509 char b[BDEVNAME_SIZE];
4510 tmp = conf->disks + i;
4512 printk(" disk %d, o:%d, dev:%s\n",
4513 i, !test_bit(Faulty, &tmp->rdev->flags),
4514 bdevname(tmp->rdev->bdev,b));
4518 static int raid5_spare_active(mddev_t *mddev)
4521 raid5_conf_t *conf = mddev->private;
4522 struct disk_info *tmp;
4524 for (i = 0; i < conf->raid_disks; i++) {
4525 tmp = conf->disks + i;
4527 && !test_bit(Faulty, &tmp->rdev->flags)
4528 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4529 unsigned long flags;
4530 spin_lock_irqsave(&conf->device_lock, flags);
4532 spin_unlock_irqrestore(&conf->device_lock, flags);
4535 print_raid5_conf(conf);
4539 static int raid5_remove_disk(mddev_t *mddev, int number)
4541 raid5_conf_t *conf = mddev->private;
4544 struct disk_info *p = conf->disks + number;
4546 print_raid5_conf(conf);
4549 if (test_bit(In_sync, &rdev->flags) ||
4550 atomic_read(&rdev->nr_pending)) {
4554 /* Only remove non-faulty devices if recovery
4557 if (!test_bit(Faulty, &rdev->flags) &&
4558 mddev->degraded <= conf->max_degraded) {
4564 if (atomic_read(&rdev->nr_pending)) {
4565 /* lost the race, try later */
4572 print_raid5_conf(conf);
4576 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4578 raid5_conf_t *conf = mddev->private;
4581 struct disk_info *p;
4583 int last = conf->raid_disks - 1;
4585 if (mddev->degraded > conf->max_degraded)
4586 /* no point adding a device */
4589 if (rdev->raid_disk >= 0)
4590 first = last = rdev->raid_disk;
4593 * find the disk ... but prefer rdev->saved_raid_disk
4596 if (rdev->saved_raid_disk >= 0 &&
4597 rdev->saved_raid_disk >= first &&
4598 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4599 disk = rdev->saved_raid_disk;
4602 for ( ; disk <= last ; disk++)
4603 if ((p=conf->disks + disk)->rdev == NULL) {
4604 clear_bit(In_sync, &rdev->flags);
4605 rdev->raid_disk = disk;
4607 if (rdev->saved_raid_disk != disk)
4609 rcu_assign_pointer(p->rdev, rdev);
4612 print_raid5_conf(conf);
4616 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4618 /* no resync is happening, and there is enough space
4619 * on all devices, so we can resize.
4620 * We need to make sure resync covers any new space.
4621 * If the array is shrinking we should possibly wait until
4622 * any io in the removed space completes, but it hardly seems
4625 raid5_conf_t *conf = mddev_to_conf(mddev);
4627 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4628 mddev->array_sectors = sectors * (mddev->raid_disks
4629 - conf->max_degraded);
4630 set_capacity(mddev->gendisk, mddev->array_sectors);
4632 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
4633 mddev->recovery_cp = mddev->dev_sectors;
4634 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4636 mddev->dev_sectors = sectors;
4637 mddev->resync_max_sectors = sectors;
4641 #ifdef CONFIG_MD_RAID5_RESHAPE
4642 static int raid5_check_reshape(mddev_t *mddev)
4644 raid5_conf_t *conf = mddev_to_conf(mddev);
4647 if (mddev->delta_disks < 0 ||
4648 mddev->new_level != mddev->level)
4649 return -EINVAL; /* Cannot shrink array or change level yet */
4650 if (mddev->delta_disks == 0)
4651 return 0; /* nothing to do */
4653 /* Cannot grow a bitmap yet */
4656 /* Can only proceed if there are plenty of stripe_heads.
4657 * We need a minimum of one full stripe,, and for sensible progress
4658 * it is best to have about 4 times that.
4659 * If we require 4 times, then the default 256 4K stripe_heads will
4660 * allow for chunk sizes up to 256K, which is probably OK.
4661 * If the chunk size is greater, user-space should request more
4662 * stripe_heads first.
4664 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4665 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4666 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4667 (mddev->chunk_size / STRIPE_SIZE)*4);
4671 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4675 if (mddev->degraded > conf->max_degraded)
4677 /* looks like we might be able to manage this */
4681 static int raid5_start_reshape(mddev_t *mddev)
4683 raid5_conf_t *conf = mddev_to_conf(mddev);
4686 int added_devices = 0;
4687 unsigned long flags;
4689 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4692 list_for_each_entry(rdev, &mddev->disks, same_set)
4693 if (rdev->raid_disk < 0 &&
4694 !test_bit(Faulty, &rdev->flags))
4697 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4698 /* Not enough devices even to make a degraded array
4703 atomic_set(&conf->reshape_stripes, 0);
4704 spin_lock_irq(&conf->device_lock);
4705 conf->previous_raid_disks = conf->raid_disks;
4706 conf->raid_disks += mddev->delta_disks;
4707 conf->expand_progress = 0;
4708 conf->expand_lo = 0;
4709 spin_unlock_irq(&conf->device_lock);
4711 /* Add some new drives, as many as will fit.
4712 * We know there are enough to make the newly sized array work.
4714 list_for_each_entry(rdev, &mddev->disks, same_set)
4715 if (rdev->raid_disk < 0 &&
4716 !test_bit(Faulty, &rdev->flags)) {
4717 if (raid5_add_disk(mddev, rdev) == 0) {
4719 set_bit(In_sync, &rdev->flags);
4721 rdev->recovery_offset = 0;
4722 sprintf(nm, "rd%d", rdev->raid_disk);
4723 if (sysfs_create_link(&mddev->kobj,
4726 "raid5: failed to create "
4727 " link %s for %s\n",
4733 spin_lock_irqsave(&conf->device_lock, flags);
4734 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4735 spin_unlock_irqrestore(&conf->device_lock, flags);
4736 mddev->raid_disks = conf->raid_disks;
4737 mddev->reshape_position = 0;
4738 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4740 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4741 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4742 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4743 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4744 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4746 if (!mddev->sync_thread) {
4747 mddev->recovery = 0;
4748 spin_lock_irq(&conf->device_lock);
4749 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4750 conf->expand_progress = MaxSector;
4751 spin_unlock_irq(&conf->device_lock);
4754 md_wakeup_thread(mddev->sync_thread);
4755 md_new_event(mddev);
4760 static void end_reshape(raid5_conf_t *conf)
4762 struct block_device *bdev;
4764 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4765 conf->mddev->array_sectors = conf->mddev->dev_sectors *
4766 (conf->raid_disks - conf->max_degraded);
4767 set_capacity(conf->mddev->gendisk, conf->mddev->array_sectors);
4768 conf->mddev->changed = 1;
4770 bdev = bdget_disk(conf->mddev->gendisk, 0);
4772 mutex_lock(&bdev->bd_inode->i_mutex);
4773 i_size_write(bdev->bd_inode,
4774 (loff_t)conf->mddev->array_sectors << 9);
4775 mutex_unlock(&bdev->bd_inode->i_mutex);
4778 spin_lock_irq(&conf->device_lock);
4779 conf->expand_progress = MaxSector;
4780 spin_unlock_irq(&conf->device_lock);
4781 conf->mddev->reshape_position = MaxSector;
4783 /* read-ahead size must cover two whole stripes, which is
4784 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4787 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4788 int stripe = data_disks *
4789 (conf->mddev->chunk_size / PAGE_SIZE);
4790 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4791 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4796 static void raid5_quiesce(mddev_t *mddev, int state)
4798 raid5_conf_t *conf = mddev_to_conf(mddev);
4801 case 2: /* resume for a suspend */
4802 wake_up(&conf->wait_for_overlap);
4805 case 1: /* stop all writes */
4806 spin_lock_irq(&conf->device_lock);
4808 wait_event_lock_irq(conf->wait_for_stripe,
4809 atomic_read(&conf->active_stripes) == 0 &&
4810 atomic_read(&conf->active_aligned_reads) == 0,
4811 conf->device_lock, /* nothing */);
4812 spin_unlock_irq(&conf->device_lock);
4815 case 0: /* re-enable writes */
4816 spin_lock_irq(&conf->device_lock);
4818 wake_up(&conf->wait_for_stripe);
4819 wake_up(&conf->wait_for_overlap);
4820 spin_unlock_irq(&conf->device_lock);
4825 static struct mdk_personality raid6_personality =
4829 .owner = THIS_MODULE,
4830 .make_request = make_request,
4834 .error_handler = error,
4835 .hot_add_disk = raid5_add_disk,
4836 .hot_remove_disk= raid5_remove_disk,
4837 .spare_active = raid5_spare_active,
4838 .sync_request = sync_request,
4839 .resize = raid5_resize,
4840 #ifdef CONFIG_MD_RAID5_RESHAPE
4841 .check_reshape = raid5_check_reshape,
4842 .start_reshape = raid5_start_reshape,
4844 .quiesce = raid5_quiesce,
4846 static struct mdk_personality raid5_personality =
4850 .owner = THIS_MODULE,
4851 .make_request = make_request,
4855 .error_handler = error,
4856 .hot_add_disk = raid5_add_disk,
4857 .hot_remove_disk= raid5_remove_disk,
4858 .spare_active = raid5_spare_active,
4859 .sync_request = sync_request,
4860 .resize = raid5_resize,
4861 #ifdef CONFIG_MD_RAID5_RESHAPE
4862 .check_reshape = raid5_check_reshape,
4863 .start_reshape = raid5_start_reshape,
4865 .quiesce = raid5_quiesce,
4868 static struct mdk_personality raid4_personality =
4872 .owner = THIS_MODULE,
4873 .make_request = make_request,
4877 .error_handler = error,
4878 .hot_add_disk = raid5_add_disk,
4879 .hot_remove_disk= raid5_remove_disk,
4880 .spare_active = raid5_spare_active,
4881 .sync_request = sync_request,
4882 .resize = raid5_resize,
4883 #ifdef CONFIG_MD_RAID5_RESHAPE
4884 .check_reshape = raid5_check_reshape,
4885 .start_reshape = raid5_start_reshape,
4887 .quiesce = raid5_quiesce,
4890 static int __init raid5_init(void)
4894 e = raid6_select_algo();
4897 register_md_personality(&raid6_personality);
4898 register_md_personality(&raid5_personality);
4899 register_md_personality(&raid4_personality);
4903 static void raid5_exit(void)
4905 unregister_md_personality(&raid6_personality);
4906 unregister_md_personality(&raid5_personality);
4907 unregister_md_personality(&raid4_personality);
4910 module_init(raid5_init);
4911 module_exit(raid5_exit);
4912 MODULE_LICENSE("GPL");
4913 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4914 MODULE_ALIAS("md-raid5");
4915 MODULE_ALIAS("md-raid4");
4916 MODULE_ALIAS("md-level-5");
4917 MODULE_ALIAS("md-level-4");
4918 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4919 MODULE_ALIAS("md-raid6");
4920 MODULE_ALIAS("md-level-6");
4922 /* This used to be two separate modules, they were: */
4923 MODULE_ALIAS("raid5");
4924 MODULE_ALIAS("raid6");