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/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
54 #include <linux/raid/bitmap.h>
55 #include <linux/async_tx.h>
61 #define NR_STRIPES 256
62 #define STRIPE_SIZE PAGE_SIZE
63 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD 1
66 #define BYPASS_THRESHOLD 1
67 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
68 #define HASH_MASK (NR_HASH - 1)
70 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
72 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
73 * order without overlap. There may be several bio's per stripe+device, and
74 * a bio could span several devices.
75 * When walking this list for a particular stripe+device, we must never proceed
76 * beyond a bio that extends past this device, as the next bio might no longer
78 * This macro is used to determine the 'next' bio in the list, given the sector
79 * of the current stripe+device
81 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
83 * The following can be used to debug the driver
85 #define RAID5_PARANOIA 1
86 #if RAID5_PARANOIA && defined(CONFIG_SMP)
87 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
89 # define CHECK_DEVLOCK()
97 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
99 #if !RAID6_USE_EMPTY_ZERO_PAGE
100 /* In .bss so it's zeroed */
101 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
104 static inline int raid6_next_disk(int disk, int raid_disks)
107 return (disk < raid_disks) ? disk : 0;
110 static void return_io(struct bio *return_bi)
112 struct bio *bi = return_bi;
115 return_bi = bi->bi_next;
123 static void print_raid5_conf (raid5_conf_t *conf);
125 static int stripe_operations_active(struct stripe_head *sh)
127 return sh->check_state || sh->reconstruct_state ||
128 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
129 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
132 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
134 if (atomic_dec_and_test(&sh->count)) {
135 BUG_ON(!list_empty(&sh->lru));
136 BUG_ON(atomic_read(&conf->active_stripes)==0);
137 if (test_bit(STRIPE_HANDLE, &sh->state)) {
138 if (test_bit(STRIPE_DELAYED, &sh->state)) {
139 list_add_tail(&sh->lru, &conf->delayed_list);
140 blk_plug_device(conf->mddev->queue);
141 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
142 sh->bm_seq - conf->seq_write > 0) {
143 list_add_tail(&sh->lru, &conf->bitmap_list);
144 blk_plug_device(conf->mddev->queue);
146 clear_bit(STRIPE_BIT_DELAY, &sh->state);
147 list_add_tail(&sh->lru, &conf->handle_list);
149 md_wakeup_thread(conf->mddev->thread);
151 BUG_ON(stripe_operations_active(sh));
152 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
153 atomic_dec(&conf->preread_active_stripes);
154 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
155 md_wakeup_thread(conf->mddev->thread);
157 atomic_dec(&conf->active_stripes);
158 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
159 list_add_tail(&sh->lru, &conf->inactive_list);
160 wake_up(&conf->wait_for_stripe);
161 if (conf->retry_read_aligned)
162 md_wakeup_thread(conf->mddev->thread);
167 static void release_stripe(struct stripe_head *sh)
169 raid5_conf_t *conf = sh->raid_conf;
172 spin_lock_irqsave(&conf->device_lock, flags);
173 __release_stripe(conf, sh);
174 spin_unlock_irqrestore(&conf->device_lock, flags);
177 static inline void remove_hash(struct stripe_head *sh)
179 pr_debug("remove_hash(), stripe %llu\n",
180 (unsigned long long)sh->sector);
182 hlist_del_init(&sh->hash);
185 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
187 struct hlist_head *hp = stripe_hash(conf, sh->sector);
189 pr_debug("insert_hash(), stripe %llu\n",
190 (unsigned long long)sh->sector);
193 hlist_add_head(&sh->hash, hp);
197 /* find an idle stripe, make sure it is unhashed, and return it. */
198 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
200 struct stripe_head *sh = NULL;
201 struct list_head *first;
204 if (list_empty(&conf->inactive_list))
206 first = conf->inactive_list.next;
207 sh = list_entry(first, struct stripe_head, lru);
208 list_del_init(first);
210 atomic_inc(&conf->active_stripes);
215 static void shrink_buffers(struct stripe_head *sh, int num)
220 for (i=0; i<num ; i++) {
224 sh->dev[i].page = NULL;
229 static int grow_buffers(struct stripe_head *sh, int num)
233 for (i=0; i<num; i++) {
236 if (!(page = alloc_page(GFP_KERNEL))) {
239 sh->dev[i].page = page;
244 static void raid5_build_block (struct stripe_head *sh, int i);
246 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
248 raid5_conf_t *conf = sh->raid_conf;
251 BUG_ON(atomic_read(&sh->count) != 0);
252 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
253 BUG_ON(stripe_operations_active(sh));
256 pr_debug("init_stripe called, stripe %llu\n",
257 (unsigned long long)sh->sector);
267 for (i = sh->disks; i--; ) {
268 struct r5dev *dev = &sh->dev[i];
270 if (dev->toread || dev->read || dev->towrite || dev->written ||
271 test_bit(R5_LOCKED, &dev->flags)) {
272 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
273 (unsigned long long)sh->sector, i, dev->toread,
274 dev->read, dev->towrite, dev->written,
275 test_bit(R5_LOCKED, &dev->flags));
279 raid5_build_block(sh, i);
281 insert_hash(conf, sh);
284 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
286 struct stripe_head *sh;
287 struct hlist_node *hn;
290 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
291 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
292 if (sh->sector == sector && sh->disks == disks)
294 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
298 static void unplug_slaves(mddev_t *mddev);
299 static void raid5_unplug_device(struct request_queue *q);
301 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
302 int pd_idx, int noblock)
304 struct stripe_head *sh;
306 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
308 spin_lock_irq(&conf->device_lock);
311 wait_event_lock_irq(conf->wait_for_stripe,
313 conf->device_lock, /* nothing */);
314 sh = __find_stripe(conf, sector, disks);
316 if (!conf->inactive_blocked)
317 sh = get_free_stripe(conf);
318 if (noblock && sh == NULL)
321 conf->inactive_blocked = 1;
322 wait_event_lock_irq(conf->wait_for_stripe,
323 !list_empty(&conf->inactive_list) &&
324 (atomic_read(&conf->active_stripes)
325 < (conf->max_nr_stripes *3/4)
326 || !conf->inactive_blocked),
328 raid5_unplug_device(conf->mddev->queue)
330 conf->inactive_blocked = 0;
332 init_stripe(sh, sector, pd_idx, disks);
334 if (atomic_read(&sh->count)) {
335 BUG_ON(!list_empty(&sh->lru));
337 if (!test_bit(STRIPE_HANDLE, &sh->state))
338 atomic_inc(&conf->active_stripes);
339 if (list_empty(&sh->lru) &&
340 !test_bit(STRIPE_EXPANDING, &sh->state))
342 list_del_init(&sh->lru);
345 } while (sh == NULL);
348 atomic_inc(&sh->count);
350 spin_unlock_irq(&conf->device_lock);
355 raid5_end_read_request(struct bio *bi, int error);
357 raid5_end_write_request(struct bio *bi, int error);
359 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
361 raid5_conf_t *conf = sh->raid_conf;
362 int i, disks = sh->disks;
366 for (i = disks; i--; ) {
370 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
372 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
377 bi = &sh->dev[i].req;
381 bi->bi_end_io = raid5_end_write_request;
383 bi->bi_end_io = raid5_end_read_request;
386 rdev = rcu_dereference(conf->disks[i].rdev);
387 if (rdev && test_bit(Faulty, &rdev->flags))
390 atomic_inc(&rdev->nr_pending);
394 if (s->syncing || s->expanding || s->expanded)
395 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
397 set_bit(STRIPE_IO_STARTED, &sh->state);
399 bi->bi_bdev = rdev->bdev;
400 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
401 __func__, (unsigned long long)sh->sector,
403 atomic_inc(&sh->count);
404 bi->bi_sector = sh->sector + rdev->data_offset;
405 bi->bi_flags = 1 << BIO_UPTODATE;
409 bi->bi_io_vec = &sh->dev[i].vec;
410 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
411 bi->bi_io_vec[0].bv_offset = 0;
412 bi->bi_size = STRIPE_SIZE;
415 test_bit(R5_ReWrite, &sh->dev[i].flags))
416 atomic_add(STRIPE_SECTORS,
417 &rdev->corrected_errors);
418 generic_make_request(bi);
421 set_bit(STRIPE_DEGRADED, &sh->state);
422 pr_debug("skip op %ld on disc %d for sector %llu\n",
423 bi->bi_rw, i, (unsigned long long)sh->sector);
424 clear_bit(R5_LOCKED, &sh->dev[i].flags);
425 set_bit(STRIPE_HANDLE, &sh->state);
430 static struct dma_async_tx_descriptor *
431 async_copy_data(int frombio, struct bio *bio, struct page *page,
432 sector_t sector, struct dma_async_tx_descriptor *tx)
435 struct page *bio_page;
439 if (bio->bi_sector >= sector)
440 page_offset = (signed)(bio->bi_sector - sector) * 512;
442 page_offset = (signed)(sector - bio->bi_sector) * -512;
443 bio_for_each_segment(bvl, bio, i) {
444 int len = bio_iovec_idx(bio, i)->bv_len;
448 if (page_offset < 0) {
449 b_offset = -page_offset;
450 page_offset += b_offset;
454 if (len > 0 && page_offset + len > STRIPE_SIZE)
455 clen = STRIPE_SIZE - page_offset;
460 b_offset += bio_iovec_idx(bio, i)->bv_offset;
461 bio_page = bio_iovec_idx(bio, i)->bv_page;
463 tx = async_memcpy(page, bio_page, page_offset,
468 tx = async_memcpy(bio_page, page, b_offset,
473 if (clen < len) /* hit end of page */
481 static void ops_complete_biofill(void *stripe_head_ref)
483 struct stripe_head *sh = stripe_head_ref;
484 struct bio *return_bi = NULL;
485 raid5_conf_t *conf = sh->raid_conf;
488 pr_debug("%s: stripe %llu\n", __func__,
489 (unsigned long long)sh->sector);
491 /* clear completed biofills */
492 spin_lock_irq(&conf->device_lock);
493 for (i = sh->disks; i--; ) {
494 struct r5dev *dev = &sh->dev[i];
496 /* acknowledge completion of a biofill operation */
497 /* and check if we need to reply to a read request,
498 * new R5_Wantfill requests are held off until
499 * !STRIPE_BIOFILL_RUN
501 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
502 struct bio *rbi, *rbi2;
507 while (rbi && rbi->bi_sector <
508 dev->sector + STRIPE_SECTORS) {
509 rbi2 = r5_next_bio(rbi, dev->sector);
510 if (--rbi->bi_phys_segments == 0) {
511 rbi->bi_next = return_bi;
518 spin_unlock_irq(&conf->device_lock);
519 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
521 return_io(return_bi);
523 set_bit(STRIPE_HANDLE, &sh->state);
527 static void ops_run_biofill(struct stripe_head *sh)
529 struct dma_async_tx_descriptor *tx = NULL;
530 raid5_conf_t *conf = sh->raid_conf;
533 pr_debug("%s: stripe %llu\n", __func__,
534 (unsigned long long)sh->sector);
536 for (i = sh->disks; i--; ) {
537 struct r5dev *dev = &sh->dev[i];
538 if (test_bit(R5_Wantfill, &dev->flags)) {
540 spin_lock_irq(&conf->device_lock);
541 dev->read = rbi = dev->toread;
543 spin_unlock_irq(&conf->device_lock);
544 while (rbi && rbi->bi_sector <
545 dev->sector + STRIPE_SECTORS) {
546 tx = async_copy_data(0, rbi, dev->page,
548 rbi = r5_next_bio(rbi, dev->sector);
553 atomic_inc(&sh->count);
554 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
555 ops_complete_biofill, sh);
558 static void ops_complete_compute5(void *stripe_head_ref)
560 struct stripe_head *sh = stripe_head_ref;
561 int target = sh->ops.target;
562 struct r5dev *tgt = &sh->dev[target];
564 pr_debug("%s: stripe %llu\n", __func__,
565 (unsigned long long)sh->sector);
567 set_bit(R5_UPTODATE, &tgt->flags);
568 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
569 clear_bit(R5_Wantcompute, &tgt->flags);
570 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
571 if (sh->check_state == check_state_compute_run)
572 sh->check_state = check_state_compute_result;
573 set_bit(STRIPE_HANDLE, &sh->state);
577 static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
579 /* kernel stack size limits the total number of disks */
580 int disks = sh->disks;
581 struct page *xor_srcs[disks];
582 int target = sh->ops.target;
583 struct r5dev *tgt = &sh->dev[target];
584 struct page *xor_dest = tgt->page;
586 struct dma_async_tx_descriptor *tx;
589 pr_debug("%s: stripe %llu block: %d\n",
590 __func__, (unsigned long long)sh->sector, target);
591 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
593 for (i = disks; i--; )
595 xor_srcs[count++] = sh->dev[i].page;
597 atomic_inc(&sh->count);
599 if (unlikely(count == 1))
600 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
601 0, NULL, ops_complete_compute5, sh);
603 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
604 ASYNC_TX_XOR_ZERO_DST, NULL,
605 ops_complete_compute5, sh);
610 static void ops_complete_prexor(void *stripe_head_ref)
612 struct stripe_head *sh = stripe_head_ref;
614 pr_debug("%s: stripe %llu\n", __func__,
615 (unsigned long long)sh->sector);
618 static struct dma_async_tx_descriptor *
619 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
621 /* kernel stack size limits the total number of disks */
622 int disks = sh->disks;
623 struct page *xor_srcs[disks];
624 int count = 0, pd_idx = sh->pd_idx, i;
626 /* existing parity data subtracted */
627 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
629 pr_debug("%s: stripe %llu\n", __func__,
630 (unsigned long long)sh->sector);
632 for (i = disks; i--; ) {
633 struct r5dev *dev = &sh->dev[i];
634 /* Only process blocks that are known to be uptodate */
635 if (test_bit(R5_Wantdrain, &dev->flags))
636 xor_srcs[count++] = dev->page;
639 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
640 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
641 ops_complete_prexor, sh);
646 static struct dma_async_tx_descriptor *
647 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
649 int disks = sh->disks;
652 pr_debug("%s: stripe %llu\n", __func__,
653 (unsigned long long)sh->sector);
655 for (i = disks; i--; ) {
656 struct r5dev *dev = &sh->dev[i];
659 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
662 spin_lock(&sh->lock);
663 chosen = dev->towrite;
665 BUG_ON(dev->written);
666 wbi = dev->written = chosen;
667 spin_unlock(&sh->lock);
669 while (wbi && wbi->bi_sector <
670 dev->sector + STRIPE_SECTORS) {
671 tx = async_copy_data(1, wbi, dev->page,
673 wbi = r5_next_bio(wbi, dev->sector);
681 static void ops_complete_postxor(void *stripe_head_ref)
683 struct stripe_head *sh = stripe_head_ref;
684 int disks = sh->disks, i, pd_idx = sh->pd_idx;
686 pr_debug("%s: stripe %llu\n", __func__,
687 (unsigned long long)sh->sector);
689 for (i = disks; i--; ) {
690 struct r5dev *dev = &sh->dev[i];
691 if (dev->written || i == pd_idx)
692 set_bit(R5_UPTODATE, &dev->flags);
695 if (sh->reconstruct_state == reconstruct_state_drain_run)
696 sh->reconstruct_state = reconstruct_state_drain_result;
697 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
698 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
700 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
701 sh->reconstruct_state = reconstruct_state_result;
704 set_bit(STRIPE_HANDLE, &sh->state);
709 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
711 /* kernel stack size limits the total number of disks */
712 int disks = sh->disks;
713 struct page *xor_srcs[disks];
715 int count = 0, pd_idx = sh->pd_idx, i;
716 struct page *xor_dest;
720 pr_debug("%s: stripe %llu\n", __func__,
721 (unsigned long long)sh->sector);
723 /* check if prexor is active which means only process blocks
724 * that are part of a read-modify-write (written)
726 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
728 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
729 for (i = disks; i--; ) {
730 struct r5dev *dev = &sh->dev[i];
732 xor_srcs[count++] = dev->page;
735 xor_dest = sh->dev[pd_idx].page;
736 for (i = disks; i--; ) {
737 struct r5dev *dev = &sh->dev[i];
739 xor_srcs[count++] = dev->page;
743 /* 1/ if we prexor'd then the dest is reused as a source
744 * 2/ if we did not prexor then we are redoing the parity
745 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
746 * for the synchronous xor case
748 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
749 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
751 atomic_inc(&sh->count);
753 if (unlikely(count == 1)) {
754 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
755 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
756 flags, tx, ops_complete_postxor, sh);
758 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
759 flags, tx, ops_complete_postxor, sh);
762 static void ops_complete_check(void *stripe_head_ref)
764 struct stripe_head *sh = stripe_head_ref;
766 pr_debug("%s: stripe %llu\n", __func__,
767 (unsigned long long)sh->sector);
769 sh->check_state = check_state_check_result;
770 set_bit(STRIPE_HANDLE, &sh->state);
774 static void ops_run_check(struct stripe_head *sh)
776 /* kernel stack size limits the total number of disks */
777 int disks = sh->disks;
778 struct page *xor_srcs[disks];
779 struct dma_async_tx_descriptor *tx;
781 int count = 0, pd_idx = sh->pd_idx, i;
782 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
784 pr_debug("%s: stripe %llu\n", __func__,
785 (unsigned long long)sh->sector);
787 for (i = disks; i--; ) {
788 struct r5dev *dev = &sh->dev[i];
790 xor_srcs[count++] = dev->page;
793 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
794 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
796 atomic_inc(&sh->count);
797 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
798 ops_complete_check, sh);
801 static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
803 int overlap_clear = 0, i, disks = sh->disks;
804 struct dma_async_tx_descriptor *tx = NULL;
806 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
811 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
812 tx = ops_run_compute5(sh);
813 /* terminate the chain if postxor is not set to be run */
814 if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
818 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
819 tx = ops_run_prexor(sh, tx);
821 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
822 tx = ops_run_biodrain(sh, tx);
826 if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
827 ops_run_postxor(sh, tx);
829 if (test_bit(STRIPE_OP_CHECK, &ops_request))
833 for (i = disks; i--; ) {
834 struct r5dev *dev = &sh->dev[i];
835 if (test_and_clear_bit(R5_Overlap, &dev->flags))
836 wake_up(&sh->raid_conf->wait_for_overlap);
840 static int grow_one_stripe(raid5_conf_t *conf)
842 struct stripe_head *sh;
843 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
846 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
847 sh->raid_conf = conf;
848 spin_lock_init(&sh->lock);
850 if (grow_buffers(sh, conf->raid_disks)) {
851 shrink_buffers(sh, conf->raid_disks);
852 kmem_cache_free(conf->slab_cache, sh);
855 sh->disks = conf->raid_disks;
856 /* we just created an active stripe so... */
857 atomic_set(&sh->count, 1);
858 atomic_inc(&conf->active_stripes);
859 INIT_LIST_HEAD(&sh->lru);
864 static int grow_stripes(raid5_conf_t *conf, int num)
866 struct kmem_cache *sc;
867 int devs = conf->raid_disks;
869 sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
870 sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
871 conf->active_name = 0;
872 sc = kmem_cache_create(conf->cache_name[conf->active_name],
873 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
877 conf->slab_cache = sc;
878 conf->pool_size = devs;
880 if (!grow_one_stripe(conf))
885 #ifdef CONFIG_MD_RAID5_RESHAPE
886 static int resize_stripes(raid5_conf_t *conf, int newsize)
888 /* Make all the stripes able to hold 'newsize' devices.
889 * New slots in each stripe get 'page' set to a new page.
891 * This happens in stages:
892 * 1/ create a new kmem_cache and allocate the required number of
894 * 2/ gather all the old stripe_heads and tranfer the pages across
895 * to the new stripe_heads. This will have the side effect of
896 * freezing the array as once all stripe_heads have been collected,
897 * no IO will be possible. Old stripe heads are freed once their
898 * pages have been transferred over, and the old kmem_cache is
899 * freed when all stripes are done.
900 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
901 * we simple return a failre status - no need to clean anything up.
902 * 4/ allocate new pages for the new slots in the new stripe_heads.
903 * If this fails, we don't bother trying the shrink the
904 * stripe_heads down again, we just leave them as they are.
905 * As each stripe_head is processed the new one is released into
908 * Once step2 is started, we cannot afford to wait for a write,
909 * so we use GFP_NOIO allocations.
911 struct stripe_head *osh, *nsh;
912 LIST_HEAD(newstripes);
913 struct disk_info *ndisks;
915 struct kmem_cache *sc;
918 if (newsize <= conf->pool_size)
919 return 0; /* never bother to shrink */
921 md_allow_write(conf->mddev);
924 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
925 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
930 for (i = conf->max_nr_stripes; i; i--) {
931 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
935 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
937 nsh->raid_conf = conf;
938 spin_lock_init(&nsh->lock);
940 list_add(&nsh->lru, &newstripes);
943 /* didn't get enough, give up */
944 while (!list_empty(&newstripes)) {
945 nsh = list_entry(newstripes.next, struct stripe_head, lru);
947 kmem_cache_free(sc, nsh);
949 kmem_cache_destroy(sc);
952 /* Step 2 - Must use GFP_NOIO now.
953 * OK, we have enough stripes, start collecting inactive
954 * stripes and copying them over
956 list_for_each_entry(nsh, &newstripes, lru) {
957 spin_lock_irq(&conf->device_lock);
958 wait_event_lock_irq(conf->wait_for_stripe,
959 !list_empty(&conf->inactive_list),
961 unplug_slaves(conf->mddev)
963 osh = get_free_stripe(conf);
964 spin_unlock_irq(&conf->device_lock);
965 atomic_set(&nsh->count, 1);
966 for(i=0; i<conf->pool_size; i++)
967 nsh->dev[i].page = osh->dev[i].page;
968 for( ; i<newsize; i++)
969 nsh->dev[i].page = NULL;
970 kmem_cache_free(conf->slab_cache, osh);
972 kmem_cache_destroy(conf->slab_cache);
975 * At this point, we are holding all the stripes so the array
976 * is completely stalled, so now is a good time to resize
979 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
981 for (i=0; i<conf->raid_disks; i++)
982 ndisks[i] = conf->disks[i];
984 conf->disks = ndisks;
988 /* Step 4, return new stripes to service */
989 while(!list_empty(&newstripes)) {
990 nsh = list_entry(newstripes.next, struct stripe_head, lru);
991 list_del_init(&nsh->lru);
992 for (i=conf->raid_disks; i < newsize; i++)
993 if (nsh->dev[i].page == NULL) {
994 struct page *p = alloc_page(GFP_NOIO);
995 nsh->dev[i].page = p;
1001 /* critical section pass, GFP_NOIO no longer needed */
1003 conf->slab_cache = sc;
1004 conf->active_name = 1-conf->active_name;
1005 conf->pool_size = newsize;
1010 static int drop_one_stripe(raid5_conf_t *conf)
1012 struct stripe_head *sh;
1014 spin_lock_irq(&conf->device_lock);
1015 sh = get_free_stripe(conf);
1016 spin_unlock_irq(&conf->device_lock);
1019 BUG_ON(atomic_read(&sh->count));
1020 shrink_buffers(sh, conf->pool_size);
1021 kmem_cache_free(conf->slab_cache, sh);
1022 atomic_dec(&conf->active_stripes);
1026 static void shrink_stripes(raid5_conf_t *conf)
1028 while (drop_one_stripe(conf))
1031 if (conf->slab_cache)
1032 kmem_cache_destroy(conf->slab_cache);
1033 conf->slab_cache = NULL;
1036 static void raid5_end_read_request(struct bio * bi, int error)
1038 struct stripe_head *sh = bi->bi_private;
1039 raid5_conf_t *conf = sh->raid_conf;
1040 int disks = sh->disks, i;
1041 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1042 char b[BDEVNAME_SIZE];
1046 for (i=0 ; i<disks; i++)
1047 if (bi == &sh->dev[i].req)
1050 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1051 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1059 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1060 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1061 rdev = conf->disks[i].rdev;
1062 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1063 " (%lu sectors at %llu on %s)\n",
1064 mdname(conf->mddev), STRIPE_SECTORS,
1065 (unsigned long long)(sh->sector
1066 + rdev->data_offset),
1067 bdevname(rdev->bdev, b));
1068 clear_bit(R5_ReadError, &sh->dev[i].flags);
1069 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1071 if (atomic_read(&conf->disks[i].rdev->read_errors))
1072 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1074 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1076 rdev = conf->disks[i].rdev;
1078 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1079 atomic_inc(&rdev->read_errors);
1080 if (conf->mddev->degraded)
1081 printk_rl(KERN_WARNING
1082 "raid5:%s: read error not correctable "
1083 "(sector %llu on %s).\n",
1084 mdname(conf->mddev),
1085 (unsigned long long)(sh->sector
1086 + rdev->data_offset),
1088 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1090 printk_rl(KERN_WARNING
1091 "raid5:%s: read error NOT corrected!! "
1092 "(sector %llu on %s).\n",
1093 mdname(conf->mddev),
1094 (unsigned long long)(sh->sector
1095 + rdev->data_offset),
1097 else if (atomic_read(&rdev->read_errors)
1098 > conf->max_nr_stripes)
1100 "raid5:%s: Too many read errors, failing device %s.\n",
1101 mdname(conf->mddev), bdn);
1105 set_bit(R5_ReadError, &sh->dev[i].flags);
1107 clear_bit(R5_ReadError, &sh->dev[i].flags);
1108 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1109 md_error(conf->mddev, rdev);
1112 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1113 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1114 set_bit(STRIPE_HANDLE, &sh->state);
1118 static void raid5_end_write_request (struct bio *bi, int error)
1120 struct stripe_head *sh = bi->bi_private;
1121 raid5_conf_t *conf = sh->raid_conf;
1122 int disks = sh->disks, i;
1123 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1125 for (i=0 ; i<disks; i++)
1126 if (bi == &sh->dev[i].req)
1129 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1130 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1138 md_error(conf->mddev, conf->disks[i].rdev);
1140 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1142 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1143 set_bit(STRIPE_HANDLE, &sh->state);
1148 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1150 static void raid5_build_block (struct stripe_head *sh, int i)
1152 struct r5dev *dev = &sh->dev[i];
1154 bio_init(&dev->req);
1155 dev->req.bi_io_vec = &dev->vec;
1157 dev->req.bi_max_vecs++;
1158 dev->vec.bv_page = dev->page;
1159 dev->vec.bv_len = STRIPE_SIZE;
1160 dev->vec.bv_offset = 0;
1162 dev->req.bi_sector = sh->sector;
1163 dev->req.bi_private = sh;
1166 dev->sector = compute_blocknr(sh, i);
1169 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1171 char b[BDEVNAME_SIZE];
1172 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1173 pr_debug("raid5: error called\n");
1175 if (!test_bit(Faulty, &rdev->flags)) {
1176 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1177 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1178 unsigned long flags;
1179 spin_lock_irqsave(&conf->device_lock, flags);
1181 spin_unlock_irqrestore(&conf->device_lock, flags);
1183 * if recovery was running, make sure it aborts.
1185 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1187 set_bit(Faulty, &rdev->flags);
1189 "raid5: Disk failure on %s, disabling device.\n"
1190 "raid5: Operation continuing on %d devices.\n",
1191 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1196 * Input: a 'big' sector number,
1197 * Output: index of the data and parity disk, and the sector # in them.
1199 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1200 unsigned int data_disks, unsigned int * dd_idx,
1201 unsigned int * pd_idx, raid5_conf_t *conf)
1204 unsigned long chunk_number;
1205 unsigned int chunk_offset;
1206 sector_t new_sector;
1207 int sectors_per_chunk = conf->chunk_size >> 9;
1209 /* First compute the information on this sector */
1212 * Compute the chunk number and the sector offset inside the chunk
1214 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1215 chunk_number = r_sector;
1216 BUG_ON(r_sector != chunk_number);
1219 * Compute the stripe number
1221 stripe = chunk_number / data_disks;
1224 * Compute the data disk and parity disk indexes inside the stripe
1226 *dd_idx = chunk_number % data_disks;
1229 * Select the parity disk based on the user selected algorithm.
1231 switch(conf->level) {
1233 *pd_idx = data_disks;
1236 switch (conf->algorithm) {
1237 case ALGORITHM_LEFT_ASYMMETRIC:
1238 *pd_idx = data_disks - stripe % raid_disks;
1239 if (*dd_idx >= *pd_idx)
1242 case ALGORITHM_RIGHT_ASYMMETRIC:
1243 *pd_idx = stripe % raid_disks;
1244 if (*dd_idx >= *pd_idx)
1247 case ALGORITHM_LEFT_SYMMETRIC:
1248 *pd_idx = data_disks - stripe % raid_disks;
1249 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1251 case ALGORITHM_RIGHT_SYMMETRIC:
1252 *pd_idx = stripe % raid_disks;
1253 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1256 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1262 /**** FIX THIS ****/
1263 switch (conf->algorithm) {
1264 case ALGORITHM_LEFT_ASYMMETRIC:
1265 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1266 if (*pd_idx == raid_disks-1)
1267 (*dd_idx)++; /* Q D D D P */
1268 else if (*dd_idx >= *pd_idx)
1269 (*dd_idx) += 2; /* D D P Q D */
1271 case ALGORITHM_RIGHT_ASYMMETRIC:
1272 *pd_idx = stripe % raid_disks;
1273 if (*pd_idx == raid_disks-1)
1274 (*dd_idx)++; /* Q D D D P */
1275 else if (*dd_idx >= *pd_idx)
1276 (*dd_idx) += 2; /* D D P Q D */
1278 case ALGORITHM_LEFT_SYMMETRIC:
1279 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1280 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1282 case ALGORITHM_RIGHT_SYMMETRIC:
1283 *pd_idx = stripe % raid_disks;
1284 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1287 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1294 * Finally, compute the new sector number
1296 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1301 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1303 raid5_conf_t *conf = sh->raid_conf;
1304 int raid_disks = sh->disks;
1305 int data_disks = raid_disks - conf->max_degraded;
1306 sector_t new_sector = sh->sector, check;
1307 int sectors_per_chunk = conf->chunk_size >> 9;
1310 int chunk_number, dummy1, dummy2, dd_idx = i;
1314 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1315 stripe = new_sector;
1316 BUG_ON(new_sector != stripe);
1318 if (i == sh->pd_idx)
1320 switch(conf->level) {
1323 switch (conf->algorithm) {
1324 case ALGORITHM_LEFT_ASYMMETRIC:
1325 case ALGORITHM_RIGHT_ASYMMETRIC:
1329 case ALGORITHM_LEFT_SYMMETRIC:
1330 case ALGORITHM_RIGHT_SYMMETRIC:
1333 i -= (sh->pd_idx + 1);
1336 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1341 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1342 return 0; /* It is the Q disk */
1343 switch (conf->algorithm) {
1344 case ALGORITHM_LEFT_ASYMMETRIC:
1345 case ALGORITHM_RIGHT_ASYMMETRIC:
1346 if (sh->pd_idx == raid_disks-1)
1347 i--; /* Q D D D P */
1348 else if (i > sh->pd_idx)
1349 i -= 2; /* D D P Q D */
1351 case ALGORITHM_LEFT_SYMMETRIC:
1352 case ALGORITHM_RIGHT_SYMMETRIC:
1353 if (sh->pd_idx == raid_disks-1)
1354 i--; /* Q D D D P */
1359 i -= (sh->pd_idx + 2);
1363 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1369 chunk_number = stripe * data_disks + i;
1370 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1372 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1373 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1374 printk(KERN_ERR "compute_blocknr: map not correct\n");
1383 * Copy data between a page in the stripe cache, and one or more bion
1384 * The page could align with the middle of the bio, or there could be
1385 * several bion, each with several bio_vecs, which cover part of the page
1386 * Multiple bion are linked together on bi_next. There may be extras
1387 * at the end of this list. We ignore them.
1389 static void copy_data(int frombio, struct bio *bio,
1393 char *pa = page_address(page);
1394 struct bio_vec *bvl;
1398 if (bio->bi_sector >= sector)
1399 page_offset = (signed)(bio->bi_sector - sector) * 512;
1401 page_offset = (signed)(sector - bio->bi_sector) * -512;
1402 bio_for_each_segment(bvl, bio, i) {
1403 int len = bio_iovec_idx(bio,i)->bv_len;
1407 if (page_offset < 0) {
1408 b_offset = -page_offset;
1409 page_offset += b_offset;
1413 if (len > 0 && page_offset + len > STRIPE_SIZE)
1414 clen = STRIPE_SIZE - page_offset;
1418 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1420 memcpy(pa+page_offset, ba+b_offset, clen);
1422 memcpy(ba+b_offset, pa+page_offset, clen);
1423 __bio_kunmap_atomic(ba, KM_USER0);
1425 if (clen < len) /* hit end of page */
1431 #define check_xor() do { \
1432 if (count == MAX_XOR_BLOCKS) { \
1433 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1438 static void compute_parity6(struct stripe_head *sh, int method)
1440 raid6_conf_t *conf = sh->raid_conf;
1441 int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1443 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1446 qd_idx = raid6_next_disk(pd_idx, disks);
1447 d0_idx = raid6_next_disk(qd_idx, disks);
1449 pr_debug("compute_parity, stripe %llu, method %d\n",
1450 (unsigned long long)sh->sector, method);
1453 case READ_MODIFY_WRITE:
1454 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1455 case RECONSTRUCT_WRITE:
1456 for (i= disks; i-- ;)
1457 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1458 chosen = sh->dev[i].towrite;
1459 sh->dev[i].towrite = NULL;
1461 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1462 wake_up(&conf->wait_for_overlap);
1464 BUG_ON(sh->dev[i].written);
1465 sh->dev[i].written = chosen;
1469 BUG(); /* Not implemented yet */
1472 for (i = disks; i--;)
1473 if (sh->dev[i].written) {
1474 sector_t sector = sh->dev[i].sector;
1475 struct bio *wbi = sh->dev[i].written;
1476 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1477 copy_data(1, wbi, sh->dev[i].page, sector);
1478 wbi = r5_next_bio(wbi, sector);
1481 set_bit(R5_LOCKED, &sh->dev[i].flags);
1482 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1486 // case RECONSTRUCT_WRITE:
1487 // case CHECK_PARITY:
1488 // case UPDATE_PARITY:
1489 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1490 /* FIX: Is this ordering of drives even remotely optimal? */
1494 ptrs[count++] = page_address(sh->dev[i].page);
1495 if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1496 printk("block %d/%d not uptodate on parity calc\n", i,count);
1497 i = raid6_next_disk(i, disks);
1498 } while ( i != d0_idx );
1502 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1505 case RECONSTRUCT_WRITE:
1506 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1507 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1508 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1509 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1512 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1513 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1519 /* Compute one missing block */
1520 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1522 int i, count, disks = sh->disks;
1523 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1524 int pd_idx = sh->pd_idx;
1525 int qd_idx = raid6_next_disk(pd_idx, disks);
1527 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1528 (unsigned long long)sh->sector, dd_idx);
1530 if ( dd_idx == qd_idx ) {
1531 /* We're actually computing the Q drive */
1532 compute_parity6(sh, UPDATE_PARITY);
1534 dest = page_address(sh->dev[dd_idx].page);
1535 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1537 for (i = disks ; i--; ) {
1538 if (i == dd_idx || i == qd_idx)
1540 p = page_address(sh->dev[i].page);
1541 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1544 printk("compute_block() %d, stripe %llu, %d"
1545 " not present\n", dd_idx,
1546 (unsigned long long)sh->sector, i);
1551 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1552 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1553 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1557 /* Compute two missing blocks */
1558 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1560 int i, count, disks = sh->disks;
1561 int pd_idx = sh->pd_idx;
1562 int qd_idx = raid6_next_disk(pd_idx, disks);
1563 int d0_idx = raid6_next_disk(qd_idx, disks);
1566 /* faila and failb are disk numbers relative to d0_idx */
1567 /* pd_idx become disks-2 and qd_idx become disks-1 */
1568 faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1569 failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1571 BUG_ON(faila == failb);
1572 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1574 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1575 (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1577 if ( failb == disks-1 ) {
1578 /* Q disk is one of the missing disks */
1579 if ( faila == disks-2 ) {
1580 /* Missing P+Q, just recompute */
1581 compute_parity6(sh, UPDATE_PARITY);
1584 /* We're missing D+Q; recompute D from P */
1585 compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1586 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1591 /* We're missing D+P or D+D; build pointer table */
1593 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1599 ptrs[count++] = page_address(sh->dev[i].page);
1600 i = raid6_next_disk(i, disks);
1601 if (i != dd_idx1 && i != dd_idx2 &&
1602 !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1603 printk("compute_2 with missing block %d/%d\n", count, i);
1604 } while ( i != d0_idx );
1606 if ( failb == disks-2 ) {
1607 /* We're missing D+P. */
1608 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1610 /* We're missing D+D. */
1611 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1614 /* Both the above update both missing blocks */
1615 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1616 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1621 schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
1622 int rcw, int expand)
1624 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1627 /* if we are not expanding this is a proper write request, and
1628 * there will be bios with new data to be drained into the
1632 sh->reconstruct_state = reconstruct_state_drain_run;
1633 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1635 sh->reconstruct_state = reconstruct_state_run;
1637 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1639 for (i = disks; i--; ) {
1640 struct r5dev *dev = &sh->dev[i];
1643 set_bit(R5_LOCKED, &dev->flags);
1644 set_bit(R5_Wantdrain, &dev->flags);
1646 clear_bit(R5_UPTODATE, &dev->flags);
1650 if (s->locked + 1 == disks)
1651 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1652 atomic_inc(&sh->raid_conf->pending_full_writes);
1654 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1655 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1657 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1658 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1659 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1660 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1662 for (i = disks; i--; ) {
1663 struct r5dev *dev = &sh->dev[i];
1668 (test_bit(R5_UPTODATE, &dev->flags) ||
1669 test_bit(R5_Wantcompute, &dev->flags))) {
1670 set_bit(R5_Wantdrain, &dev->flags);
1671 set_bit(R5_LOCKED, &dev->flags);
1672 clear_bit(R5_UPTODATE, &dev->flags);
1678 /* keep the parity disk locked while asynchronous operations
1681 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1682 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1685 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1686 __func__, (unsigned long long)sh->sector,
1687 s->locked, s->ops_request);
1691 * Each stripe/dev can have one or more bion attached.
1692 * toread/towrite point to the first in a chain.
1693 * The bi_next chain must be in order.
1695 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1698 raid5_conf_t *conf = sh->raid_conf;
1701 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1702 (unsigned long long)bi->bi_sector,
1703 (unsigned long long)sh->sector);
1706 spin_lock(&sh->lock);
1707 spin_lock_irq(&conf->device_lock);
1709 bip = &sh->dev[dd_idx].towrite;
1710 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1713 bip = &sh->dev[dd_idx].toread;
1714 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1715 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1717 bip = & (*bip)->bi_next;
1719 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1722 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1726 bi->bi_phys_segments ++;
1727 spin_unlock_irq(&conf->device_lock);
1728 spin_unlock(&sh->lock);
1730 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1731 (unsigned long long)bi->bi_sector,
1732 (unsigned long long)sh->sector, dd_idx);
1734 if (conf->mddev->bitmap && firstwrite) {
1735 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1737 sh->bm_seq = conf->seq_flush+1;
1738 set_bit(STRIPE_BIT_DELAY, &sh->state);
1742 /* check if page is covered */
1743 sector_t sector = sh->dev[dd_idx].sector;
1744 for (bi=sh->dev[dd_idx].towrite;
1745 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1746 bi && bi->bi_sector <= sector;
1747 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1748 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1749 sector = bi->bi_sector + (bi->bi_size>>9);
1751 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1752 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1757 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1758 spin_unlock_irq(&conf->device_lock);
1759 spin_unlock(&sh->lock);
1763 static void end_reshape(raid5_conf_t *conf);
1765 static int page_is_zero(struct page *p)
1767 char *a = page_address(p);
1768 return ((*(u32*)a) == 0 &&
1769 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1772 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1774 int sectors_per_chunk = conf->chunk_size >> 9;
1776 int chunk_offset = sector_div(stripe, sectors_per_chunk);
1778 raid5_compute_sector(stripe * (disks - conf->max_degraded)
1779 *sectors_per_chunk + chunk_offset,
1780 disks, disks - conf->max_degraded,
1781 &dd_idx, &pd_idx, conf);
1786 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
1787 struct stripe_head_state *s, int disks,
1788 struct bio **return_bi)
1791 for (i = disks; i--; ) {
1795 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1798 rdev = rcu_dereference(conf->disks[i].rdev);
1799 if (rdev && test_bit(In_sync, &rdev->flags))
1800 /* multiple read failures in one stripe */
1801 md_error(conf->mddev, rdev);
1804 spin_lock_irq(&conf->device_lock);
1805 /* fail all writes first */
1806 bi = sh->dev[i].towrite;
1807 sh->dev[i].towrite = NULL;
1813 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1814 wake_up(&conf->wait_for_overlap);
1816 while (bi && bi->bi_sector <
1817 sh->dev[i].sector + STRIPE_SECTORS) {
1818 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1819 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1820 if (--bi->bi_phys_segments == 0) {
1821 md_write_end(conf->mddev);
1822 bi->bi_next = *return_bi;
1827 /* and fail all 'written' */
1828 bi = sh->dev[i].written;
1829 sh->dev[i].written = NULL;
1830 if (bi) bitmap_end = 1;
1831 while (bi && bi->bi_sector <
1832 sh->dev[i].sector + STRIPE_SECTORS) {
1833 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1834 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1835 if (--bi->bi_phys_segments == 0) {
1836 md_write_end(conf->mddev);
1837 bi->bi_next = *return_bi;
1843 /* fail any reads if this device is non-operational and
1844 * the data has not reached the cache yet.
1846 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1847 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1848 test_bit(R5_ReadError, &sh->dev[i].flags))) {
1849 bi = sh->dev[i].toread;
1850 sh->dev[i].toread = NULL;
1851 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1852 wake_up(&conf->wait_for_overlap);
1853 if (bi) s->to_read--;
1854 while (bi && bi->bi_sector <
1855 sh->dev[i].sector + STRIPE_SECTORS) {
1856 struct bio *nextbi =
1857 r5_next_bio(bi, sh->dev[i].sector);
1858 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1859 if (--bi->bi_phys_segments == 0) {
1860 bi->bi_next = *return_bi;
1866 spin_unlock_irq(&conf->device_lock);
1868 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1869 STRIPE_SECTORS, 0, 0);
1872 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
1873 if (atomic_dec_and_test(&conf->pending_full_writes))
1874 md_wakeup_thread(conf->mddev->thread);
1877 /* fetch_block5 - checks the given member device to see if its data needs
1878 * to be read or computed to satisfy a request.
1880 * Returns 1 when no more member devices need to be checked, otherwise returns
1881 * 0 to tell the loop in handle_stripe_fill5 to continue
1883 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
1884 int disk_idx, int disks)
1886 struct r5dev *dev = &sh->dev[disk_idx];
1887 struct r5dev *failed_dev = &sh->dev[s->failed_num];
1889 /* is the data in this block needed, and can we get it? */
1890 if (!test_bit(R5_LOCKED, &dev->flags) &&
1891 !test_bit(R5_UPTODATE, &dev->flags) &&
1893 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1894 s->syncing || s->expanding ||
1896 (failed_dev->toread ||
1897 (failed_dev->towrite &&
1898 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
1899 /* We would like to get this block, possibly by computing it,
1900 * otherwise read it if the backing disk is insync
1902 if ((s->uptodate == disks - 1) &&
1903 (s->failed && disk_idx == s->failed_num)) {
1904 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
1905 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
1906 set_bit(R5_Wantcompute, &dev->flags);
1907 sh->ops.target = disk_idx;
1909 /* Careful: from this point on 'uptodate' is in the eye
1910 * of raid5_run_ops which services 'compute' operations
1911 * before writes. R5_Wantcompute flags a block that will
1912 * be R5_UPTODATE by the time it is needed for a
1913 * subsequent operation.
1916 return 1; /* uptodate + compute == disks */
1917 } else if (test_bit(R5_Insync, &dev->flags)) {
1918 set_bit(R5_LOCKED, &dev->flags);
1919 set_bit(R5_Wantread, &dev->flags);
1921 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
1930 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
1932 static void handle_stripe_fill5(struct stripe_head *sh,
1933 struct stripe_head_state *s, int disks)
1937 /* look for blocks to read/compute, skip this if a compute
1938 * is already in flight, or if the stripe contents are in the
1939 * midst of changing due to a write
1941 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
1942 !sh->reconstruct_state)
1943 for (i = disks; i--; )
1944 if (fetch_block5(sh, s, i, disks))
1946 set_bit(STRIPE_HANDLE, &sh->state);
1949 static void handle_stripe_fill6(struct stripe_head *sh,
1950 struct stripe_head_state *s, struct r6_state *r6s,
1954 for (i = disks; i--; ) {
1955 struct r5dev *dev = &sh->dev[i];
1956 if (!test_bit(R5_LOCKED, &dev->flags) &&
1957 !test_bit(R5_UPTODATE, &dev->flags) &&
1958 (dev->toread || (dev->towrite &&
1959 !test_bit(R5_OVERWRITE, &dev->flags)) ||
1960 s->syncing || s->expanding ||
1962 (sh->dev[r6s->failed_num[0]].toread ||
1965 (sh->dev[r6s->failed_num[1]].toread ||
1967 /* we would like to get this block, possibly
1968 * by computing it, but we might not be able to
1970 if ((s->uptodate == disks - 1) &&
1971 (s->failed && (i == r6s->failed_num[0] ||
1972 i == r6s->failed_num[1]))) {
1973 pr_debug("Computing stripe %llu block %d\n",
1974 (unsigned long long)sh->sector, i);
1975 compute_block_1(sh, i, 0);
1977 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
1978 /* Computing 2-failure is *very* expensive; only
1979 * do it if failed >= 2
1982 for (other = disks; other--; ) {
1985 if (!test_bit(R5_UPTODATE,
1986 &sh->dev[other].flags))
1990 pr_debug("Computing stripe %llu blocks %d,%d\n",
1991 (unsigned long long)sh->sector,
1993 compute_block_2(sh, i, other);
1995 } else if (test_bit(R5_Insync, &dev->flags)) {
1996 set_bit(R5_LOCKED, &dev->flags);
1997 set_bit(R5_Wantread, &dev->flags);
1999 pr_debug("Reading block %d (sync=%d)\n",
2004 set_bit(STRIPE_HANDLE, &sh->state);
2008 /* handle_stripe_clean_event
2009 * any written block on an uptodate or failed drive can be returned.
2010 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2011 * never LOCKED, so we don't need to test 'failed' directly.
2013 static void handle_stripe_clean_event(raid5_conf_t *conf,
2014 struct stripe_head *sh, int disks, struct bio **return_bi)
2019 for (i = disks; i--; )
2020 if (sh->dev[i].written) {
2022 if (!test_bit(R5_LOCKED, &dev->flags) &&
2023 test_bit(R5_UPTODATE, &dev->flags)) {
2024 /* We can return any write requests */
2025 struct bio *wbi, *wbi2;
2027 pr_debug("Return write for disc %d\n", i);
2028 spin_lock_irq(&conf->device_lock);
2030 dev->written = NULL;
2031 while (wbi && wbi->bi_sector <
2032 dev->sector + STRIPE_SECTORS) {
2033 wbi2 = r5_next_bio(wbi, dev->sector);
2034 if (--wbi->bi_phys_segments == 0) {
2035 md_write_end(conf->mddev);
2036 wbi->bi_next = *return_bi;
2041 if (dev->towrite == NULL)
2043 spin_unlock_irq(&conf->device_lock);
2045 bitmap_endwrite(conf->mddev->bitmap,
2048 !test_bit(STRIPE_DEGRADED, &sh->state),
2053 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2054 if (atomic_dec_and_test(&conf->pending_full_writes))
2055 md_wakeup_thread(conf->mddev->thread);
2058 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2059 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2061 int rmw = 0, rcw = 0, i;
2062 for (i = disks; i--; ) {
2063 /* would I have to read this buffer for read_modify_write */
2064 struct r5dev *dev = &sh->dev[i];
2065 if ((dev->towrite || i == sh->pd_idx) &&
2066 !test_bit(R5_LOCKED, &dev->flags) &&
2067 !(test_bit(R5_UPTODATE, &dev->flags) ||
2068 test_bit(R5_Wantcompute, &dev->flags))) {
2069 if (test_bit(R5_Insync, &dev->flags))
2072 rmw += 2*disks; /* cannot read it */
2074 /* Would I have to read this buffer for reconstruct_write */
2075 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2076 !test_bit(R5_LOCKED, &dev->flags) &&
2077 !(test_bit(R5_UPTODATE, &dev->flags) ||
2078 test_bit(R5_Wantcompute, &dev->flags))) {
2079 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2084 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2085 (unsigned long long)sh->sector, rmw, rcw);
2086 set_bit(STRIPE_HANDLE, &sh->state);
2087 if (rmw < rcw && rmw > 0)
2088 /* prefer read-modify-write, but need to get some data */
2089 for (i = disks; i--; ) {
2090 struct r5dev *dev = &sh->dev[i];
2091 if ((dev->towrite || i == sh->pd_idx) &&
2092 !test_bit(R5_LOCKED, &dev->flags) &&
2093 !(test_bit(R5_UPTODATE, &dev->flags) ||
2094 test_bit(R5_Wantcompute, &dev->flags)) &&
2095 test_bit(R5_Insync, &dev->flags)) {
2097 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2098 pr_debug("Read_old block "
2099 "%d for r-m-w\n", i);
2100 set_bit(R5_LOCKED, &dev->flags);
2101 set_bit(R5_Wantread, &dev->flags);
2104 set_bit(STRIPE_DELAYED, &sh->state);
2105 set_bit(STRIPE_HANDLE, &sh->state);
2109 if (rcw <= rmw && rcw > 0)
2110 /* want reconstruct write, but need to get some data */
2111 for (i = disks; i--; ) {
2112 struct r5dev *dev = &sh->dev[i];
2113 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2115 !test_bit(R5_LOCKED, &dev->flags) &&
2116 !(test_bit(R5_UPTODATE, &dev->flags) ||
2117 test_bit(R5_Wantcompute, &dev->flags)) &&
2118 test_bit(R5_Insync, &dev->flags)) {
2120 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2121 pr_debug("Read_old block "
2122 "%d for Reconstruct\n", i);
2123 set_bit(R5_LOCKED, &dev->flags);
2124 set_bit(R5_Wantread, &dev->flags);
2127 set_bit(STRIPE_DELAYED, &sh->state);
2128 set_bit(STRIPE_HANDLE, &sh->state);
2132 /* now if nothing is locked, and if we have enough data,
2133 * we can start a write request
2135 /* since handle_stripe can be called at any time we need to handle the
2136 * case where a compute block operation has been submitted and then a
2137 * subsequent call wants to start a write request. raid5_run_ops only
2138 * handles the case where compute block and postxor are requested
2139 * simultaneously. If this is not the case then new writes need to be
2140 * held off until the compute completes.
2142 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2143 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2144 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2145 schedule_reconstruction5(sh, s, rcw == 0, 0);
2148 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2149 struct stripe_head *sh, struct stripe_head_state *s,
2150 struct r6_state *r6s, int disks)
2152 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2153 int qd_idx = r6s->qd_idx;
2154 for (i = disks; i--; ) {
2155 struct r5dev *dev = &sh->dev[i];
2156 /* Would I have to read this buffer for reconstruct_write */
2157 if (!test_bit(R5_OVERWRITE, &dev->flags)
2158 && i != pd_idx && i != qd_idx
2159 && (!test_bit(R5_LOCKED, &dev->flags)
2161 !test_bit(R5_UPTODATE, &dev->flags)) {
2162 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2164 pr_debug("raid6: must_compute: "
2165 "disk %d flags=%#lx\n", i, dev->flags);
2170 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2171 (unsigned long long)sh->sector, rcw, must_compute);
2172 set_bit(STRIPE_HANDLE, &sh->state);
2175 /* want reconstruct write, but need to get some data */
2176 for (i = disks; i--; ) {
2177 struct r5dev *dev = &sh->dev[i];
2178 if (!test_bit(R5_OVERWRITE, &dev->flags)
2179 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2180 && !test_bit(R5_LOCKED, &dev->flags) &&
2181 !test_bit(R5_UPTODATE, &dev->flags) &&
2182 test_bit(R5_Insync, &dev->flags)) {
2184 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2185 pr_debug("Read_old stripe %llu "
2186 "block %d for Reconstruct\n",
2187 (unsigned long long)sh->sector, i);
2188 set_bit(R5_LOCKED, &dev->flags);
2189 set_bit(R5_Wantread, &dev->flags);
2192 pr_debug("Request delayed stripe %llu "
2193 "block %d for Reconstruct\n",
2194 (unsigned long long)sh->sector, i);
2195 set_bit(STRIPE_DELAYED, &sh->state);
2196 set_bit(STRIPE_HANDLE, &sh->state);
2200 /* now if nothing is locked, and if we have enough data, we can start a
2203 if (s->locked == 0 && rcw == 0 &&
2204 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2205 if (must_compute > 0) {
2206 /* We have failed blocks and need to compute them */
2207 switch (s->failed) {
2211 compute_block_1(sh, r6s->failed_num[0], 0);
2214 compute_block_2(sh, r6s->failed_num[0],
2215 r6s->failed_num[1]);
2217 default: /* This request should have been failed? */
2222 pr_debug("Computing parity for stripe %llu\n",
2223 (unsigned long long)sh->sector);
2224 compute_parity6(sh, RECONSTRUCT_WRITE);
2225 /* now every locked buffer is ready to be written */
2226 for (i = disks; i--; )
2227 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2228 pr_debug("Writing stripe %llu block %d\n",
2229 (unsigned long long)sh->sector, i);
2231 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2233 if (s->locked == disks)
2234 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2235 atomic_inc(&conf->pending_full_writes);
2236 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2237 set_bit(STRIPE_INSYNC, &sh->state);
2239 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2240 atomic_dec(&conf->preread_active_stripes);
2241 if (atomic_read(&conf->preread_active_stripes) <
2243 md_wakeup_thread(conf->mddev->thread);
2248 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2249 struct stripe_head_state *s, int disks)
2251 struct r5dev *dev = NULL;
2253 set_bit(STRIPE_HANDLE, &sh->state);
2255 switch (sh->check_state) {
2256 case check_state_idle:
2257 /* start a new check operation if there are no failures */
2258 if (s->failed == 0) {
2259 BUG_ON(s->uptodate != disks);
2260 sh->check_state = check_state_run;
2261 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2262 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2266 dev = &sh->dev[s->failed_num];
2268 case check_state_compute_result:
2269 sh->check_state = check_state_idle;
2271 dev = &sh->dev[sh->pd_idx];
2273 /* check that a write has not made the stripe insync */
2274 if (test_bit(STRIPE_INSYNC, &sh->state))
2277 /* either failed parity check, or recovery is happening */
2278 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2279 BUG_ON(s->uptodate != disks);
2281 set_bit(R5_LOCKED, &dev->flags);
2283 set_bit(R5_Wantwrite, &dev->flags);
2285 clear_bit(STRIPE_DEGRADED, &sh->state);
2286 set_bit(STRIPE_INSYNC, &sh->state);
2288 case check_state_run:
2289 break; /* we will be called again upon completion */
2290 case check_state_check_result:
2291 sh->check_state = check_state_idle;
2293 /* if a failure occurred during the check operation, leave
2294 * STRIPE_INSYNC not set and let the stripe be handled again
2299 /* handle a successful check operation, if parity is correct
2300 * we are done. Otherwise update the mismatch count and repair
2301 * parity if !MD_RECOVERY_CHECK
2303 if (sh->ops.zero_sum_result == 0)
2304 /* parity is correct (on disc,
2305 * not in buffer any more)
2307 set_bit(STRIPE_INSYNC, &sh->state);
2309 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2310 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2311 /* don't try to repair!! */
2312 set_bit(STRIPE_INSYNC, &sh->state);
2314 sh->check_state = check_state_compute_run;
2315 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2316 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2317 set_bit(R5_Wantcompute,
2318 &sh->dev[sh->pd_idx].flags);
2319 sh->ops.target = sh->pd_idx;
2324 case check_state_compute_run:
2327 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2328 __func__, sh->check_state,
2329 (unsigned long long) sh->sector);
2335 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2336 struct stripe_head_state *s,
2337 struct r6_state *r6s, struct page *tmp_page,
2340 int update_p = 0, update_q = 0;
2342 int pd_idx = sh->pd_idx;
2343 int qd_idx = r6s->qd_idx;
2345 set_bit(STRIPE_HANDLE, &sh->state);
2347 BUG_ON(s->failed > 2);
2348 BUG_ON(s->uptodate < disks);
2349 /* Want to check and possibly repair P and Q.
2350 * However there could be one 'failed' device, in which
2351 * case we can only check one of them, possibly using the
2352 * other to generate missing data
2355 /* If !tmp_page, we cannot do the calculations,
2356 * but as we have set STRIPE_HANDLE, we will soon be called
2357 * by stripe_handle with a tmp_page - just wait until then.
2360 if (s->failed == r6s->q_failed) {
2361 /* The only possible failed device holds 'Q', so it
2362 * makes sense to check P (If anything else were failed,
2363 * we would have used P to recreate it).
2365 compute_block_1(sh, pd_idx, 1);
2366 if (!page_is_zero(sh->dev[pd_idx].page)) {
2367 compute_block_1(sh, pd_idx, 0);
2371 if (!r6s->q_failed && s->failed < 2) {
2372 /* q is not failed, and we didn't use it to generate
2373 * anything, so it makes sense to check it
2375 memcpy(page_address(tmp_page),
2376 page_address(sh->dev[qd_idx].page),
2378 compute_parity6(sh, UPDATE_PARITY);
2379 if (memcmp(page_address(tmp_page),
2380 page_address(sh->dev[qd_idx].page),
2381 STRIPE_SIZE) != 0) {
2382 clear_bit(STRIPE_INSYNC, &sh->state);
2386 if (update_p || update_q) {
2387 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2388 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2389 /* don't try to repair!! */
2390 update_p = update_q = 0;
2393 /* now write out any block on a failed drive,
2394 * or P or Q if they need it
2397 if (s->failed == 2) {
2398 dev = &sh->dev[r6s->failed_num[1]];
2400 set_bit(R5_LOCKED, &dev->flags);
2401 set_bit(R5_Wantwrite, &dev->flags);
2403 if (s->failed >= 1) {
2404 dev = &sh->dev[r6s->failed_num[0]];
2406 set_bit(R5_LOCKED, &dev->flags);
2407 set_bit(R5_Wantwrite, &dev->flags);
2411 dev = &sh->dev[pd_idx];
2413 set_bit(R5_LOCKED, &dev->flags);
2414 set_bit(R5_Wantwrite, &dev->flags);
2417 dev = &sh->dev[qd_idx];
2419 set_bit(R5_LOCKED, &dev->flags);
2420 set_bit(R5_Wantwrite, &dev->flags);
2422 clear_bit(STRIPE_DEGRADED, &sh->state);
2424 set_bit(STRIPE_INSYNC, &sh->state);
2428 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2429 struct r6_state *r6s)
2433 /* We have read all the blocks in this stripe and now we need to
2434 * copy some of them into a target stripe for expand.
2436 struct dma_async_tx_descriptor *tx = NULL;
2437 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2438 for (i = 0; i < sh->disks; i++)
2439 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2440 int dd_idx, pd_idx, j;
2441 struct stripe_head *sh2;
2443 sector_t bn = compute_blocknr(sh, i);
2444 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2446 conf->max_degraded, &dd_idx,
2448 sh2 = get_active_stripe(conf, s, conf->raid_disks,
2451 /* so far only the early blocks of this stripe
2452 * have been requested. When later blocks
2453 * get requested, we will try again
2456 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2457 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2458 /* must have already done this block */
2459 release_stripe(sh2);
2463 /* place all the copies on one channel */
2464 tx = async_memcpy(sh2->dev[dd_idx].page,
2465 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2466 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2468 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2469 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2470 for (j = 0; j < conf->raid_disks; j++)
2471 if (j != sh2->pd_idx &&
2472 (!r6s || j != raid6_next_disk(sh2->pd_idx,
2474 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2476 if (j == conf->raid_disks) {
2477 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2478 set_bit(STRIPE_HANDLE, &sh2->state);
2480 release_stripe(sh2);
2483 /* done submitting copies, wait for them to complete */
2486 dma_wait_for_async_tx(tx);
2492 * handle_stripe - do things to a stripe.
2494 * We lock the stripe and then examine the state of various bits
2495 * to see what needs to be done.
2497 * return some read request which now have data
2498 * return some write requests which are safely on disc
2499 * schedule a read on some buffers
2500 * schedule a write of some buffers
2501 * return confirmation of parity correctness
2503 * buffers are taken off read_list or write_list, and bh_cache buffers
2504 * get BH_Lock set before the stripe lock is released.
2508 static void handle_stripe5(struct stripe_head *sh)
2510 raid5_conf_t *conf = sh->raid_conf;
2511 int disks = sh->disks, i;
2512 struct bio *return_bi = NULL;
2513 struct stripe_head_state s;
2515 mdk_rdev_t *blocked_rdev = NULL;
2518 memset(&s, 0, sizeof(s));
2519 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2520 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2521 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2522 sh->reconstruct_state);
2524 spin_lock(&sh->lock);
2525 clear_bit(STRIPE_HANDLE, &sh->state);
2526 clear_bit(STRIPE_DELAYED, &sh->state);
2528 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2529 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2530 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2532 /* Now to look around and see what can be done */
2534 for (i=disks; i--; ) {
2536 struct r5dev *dev = &sh->dev[i];
2537 clear_bit(R5_Insync, &dev->flags);
2539 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2540 "written %p\n", i, dev->flags, dev->toread, dev->read,
2541 dev->towrite, dev->written);
2543 /* maybe we can request a biofill operation
2545 * new wantfill requests are only permitted while
2546 * ops_complete_biofill is guaranteed to be inactive
2548 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2549 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2550 set_bit(R5_Wantfill, &dev->flags);
2552 /* now count some things */
2553 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2554 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2555 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2557 if (test_bit(R5_Wantfill, &dev->flags))
2559 else if (dev->toread)
2563 if (!test_bit(R5_OVERWRITE, &dev->flags))
2568 rdev = rcu_dereference(conf->disks[i].rdev);
2569 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2570 blocked_rdev = rdev;
2571 atomic_inc(&rdev->nr_pending);
2574 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2575 /* The ReadError flag will just be confusing now */
2576 clear_bit(R5_ReadError, &dev->flags);
2577 clear_bit(R5_ReWrite, &dev->flags);
2579 if (!rdev || !test_bit(In_sync, &rdev->flags)
2580 || test_bit(R5_ReadError, &dev->flags)) {
2584 set_bit(R5_Insync, &dev->flags);
2588 if (unlikely(blocked_rdev)) {
2589 set_bit(STRIPE_HANDLE, &sh->state);
2593 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2594 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2595 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2598 pr_debug("locked=%d uptodate=%d to_read=%d"
2599 " to_write=%d failed=%d failed_num=%d\n",
2600 s.locked, s.uptodate, s.to_read, s.to_write,
2601 s.failed, s.failed_num);
2602 /* check if the array has lost two devices and, if so, some requests might
2605 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2606 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2607 if (s.failed > 1 && s.syncing) {
2608 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2609 clear_bit(STRIPE_SYNCING, &sh->state);
2613 /* might be able to return some write requests if the parity block
2614 * is safe, or on a failed drive
2616 dev = &sh->dev[sh->pd_idx];
2618 ((test_bit(R5_Insync, &dev->flags) &&
2619 !test_bit(R5_LOCKED, &dev->flags) &&
2620 test_bit(R5_UPTODATE, &dev->flags)) ||
2621 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2622 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2624 /* Now we might consider reading some blocks, either to check/generate
2625 * parity, or to satisfy requests
2626 * or to load a block that is being partially written.
2628 if (s.to_read || s.non_overwrite ||
2629 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2630 handle_stripe_fill5(sh, &s, disks);
2632 /* Now we check to see if any write operations have recently
2636 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
2638 if (sh->reconstruct_state == reconstruct_state_drain_result ||
2639 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
2640 sh->reconstruct_state = reconstruct_state_idle;
2642 /* All the 'written' buffers and the parity block are ready to
2643 * be written back to disk
2645 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2646 for (i = disks; i--; ) {
2648 if (test_bit(R5_LOCKED, &dev->flags) &&
2649 (i == sh->pd_idx || dev->written)) {
2650 pr_debug("Writing block %d\n", i);
2651 set_bit(R5_Wantwrite, &dev->flags);
2654 if (!test_bit(R5_Insync, &dev->flags) ||
2655 (i == sh->pd_idx && s.failed == 0))
2656 set_bit(STRIPE_INSYNC, &sh->state);
2659 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2660 atomic_dec(&conf->preread_active_stripes);
2661 if (atomic_read(&conf->preread_active_stripes) <
2663 md_wakeup_thread(conf->mddev->thread);
2667 /* Now to consider new write requests and what else, if anything
2668 * should be read. We do not handle new writes when:
2669 * 1/ A 'write' operation (copy+xor) is already in flight.
2670 * 2/ A 'check' operation is in flight, as it may clobber the parity
2673 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
2674 handle_stripe_dirtying5(conf, sh, &s, disks);
2676 /* maybe we need to check and possibly fix the parity for this stripe
2677 * Any reads will already have been scheduled, so we just see if enough
2678 * data is available. The parity check is held off while parity
2679 * dependent operations are in flight.
2681 if (sh->check_state ||
2682 (s.syncing && s.locked == 0 &&
2683 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2684 !test_bit(STRIPE_INSYNC, &sh->state)))
2685 handle_parity_checks5(conf, sh, &s, disks);
2687 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2688 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2689 clear_bit(STRIPE_SYNCING, &sh->state);
2692 /* If the failed drive is just a ReadError, then we might need to progress
2693 * the repair/check process
2695 if (s.failed == 1 && !conf->mddev->ro &&
2696 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2697 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2698 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2700 dev = &sh->dev[s.failed_num];
2701 if (!test_bit(R5_ReWrite, &dev->flags)) {
2702 set_bit(R5_Wantwrite, &dev->flags);
2703 set_bit(R5_ReWrite, &dev->flags);
2704 set_bit(R5_LOCKED, &dev->flags);
2707 /* let's read it back */
2708 set_bit(R5_Wantread, &dev->flags);
2709 set_bit(R5_LOCKED, &dev->flags);
2714 /* Finish reconstruct operations initiated by the expansion process */
2715 if (sh->reconstruct_state == reconstruct_state_result) {
2716 sh->reconstruct_state = reconstruct_state_idle;
2717 clear_bit(STRIPE_EXPANDING, &sh->state);
2718 for (i = conf->raid_disks; i--; )
2719 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2720 set_bit(R5_LOCKED, &dev->flags);
2724 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2725 !sh->reconstruct_state) {
2726 /* Need to write out all blocks after computing parity */
2727 sh->disks = conf->raid_disks;
2728 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2730 schedule_reconstruction5(sh, &s, 1, 1);
2731 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
2732 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2733 atomic_dec(&conf->reshape_stripes);
2734 wake_up(&conf->wait_for_overlap);
2735 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2738 if (s.expanding && s.locked == 0 &&
2739 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2740 handle_stripe_expansion(conf, sh, NULL);
2743 spin_unlock(&sh->lock);
2745 /* wait for this device to become unblocked */
2746 if (unlikely(blocked_rdev))
2747 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2750 raid5_run_ops(sh, s.ops_request);
2754 return_io(return_bi);
2757 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2759 raid6_conf_t *conf = sh->raid_conf;
2760 int disks = sh->disks;
2761 struct bio *return_bi = NULL;
2762 int i, pd_idx = sh->pd_idx;
2763 struct stripe_head_state s;
2764 struct r6_state r6s;
2765 struct r5dev *dev, *pdev, *qdev;
2766 mdk_rdev_t *blocked_rdev = NULL;
2768 r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2769 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2770 "pd_idx=%d, qd_idx=%d\n",
2771 (unsigned long long)sh->sector, sh->state,
2772 atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2773 memset(&s, 0, sizeof(s));
2775 spin_lock(&sh->lock);
2776 clear_bit(STRIPE_HANDLE, &sh->state);
2777 clear_bit(STRIPE_DELAYED, &sh->state);
2779 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2780 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2781 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2782 /* Now to look around and see what can be done */
2785 for (i=disks; i--; ) {
2788 clear_bit(R5_Insync, &dev->flags);
2790 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2791 i, dev->flags, dev->toread, dev->towrite, dev->written);
2792 /* maybe we can reply to a read */
2793 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2794 struct bio *rbi, *rbi2;
2795 pr_debug("Return read for disc %d\n", i);
2796 spin_lock_irq(&conf->device_lock);
2799 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2800 wake_up(&conf->wait_for_overlap);
2801 spin_unlock_irq(&conf->device_lock);
2802 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2803 copy_data(0, rbi, dev->page, dev->sector);
2804 rbi2 = r5_next_bio(rbi, dev->sector);
2805 spin_lock_irq(&conf->device_lock);
2806 if (--rbi->bi_phys_segments == 0) {
2807 rbi->bi_next = return_bi;
2810 spin_unlock_irq(&conf->device_lock);
2815 /* now count some things */
2816 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2817 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2824 if (!test_bit(R5_OVERWRITE, &dev->flags))
2829 rdev = rcu_dereference(conf->disks[i].rdev);
2830 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2831 blocked_rdev = rdev;
2832 atomic_inc(&rdev->nr_pending);
2835 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2836 /* The ReadError flag will just be confusing now */
2837 clear_bit(R5_ReadError, &dev->flags);
2838 clear_bit(R5_ReWrite, &dev->flags);
2840 if (!rdev || !test_bit(In_sync, &rdev->flags)
2841 || test_bit(R5_ReadError, &dev->flags)) {
2843 r6s.failed_num[s.failed] = i;
2846 set_bit(R5_Insync, &dev->flags);
2850 if (unlikely(blocked_rdev)) {
2851 set_bit(STRIPE_HANDLE, &sh->state);
2854 pr_debug("locked=%d uptodate=%d to_read=%d"
2855 " to_write=%d failed=%d failed_num=%d,%d\n",
2856 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
2857 r6s.failed_num[0], r6s.failed_num[1]);
2858 /* check if the array has lost >2 devices and, if so, some requests
2859 * might need to be failed
2861 if (s.failed > 2 && s.to_read+s.to_write+s.written)
2862 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2863 if (s.failed > 2 && s.syncing) {
2864 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2865 clear_bit(STRIPE_SYNCING, &sh->state);
2870 * might be able to return some write requests if the parity blocks
2871 * are safe, or on a failed drive
2873 pdev = &sh->dev[pd_idx];
2874 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
2875 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
2876 qdev = &sh->dev[r6s.qd_idx];
2877 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
2878 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
2881 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
2882 && !test_bit(R5_LOCKED, &pdev->flags)
2883 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
2884 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
2885 && !test_bit(R5_LOCKED, &qdev->flags)
2886 && test_bit(R5_UPTODATE, &qdev->flags)))))
2887 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2889 /* Now we might consider reading some blocks, either to check/generate
2890 * parity, or to satisfy requests
2891 * or to load a block that is being partially written.
2893 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
2894 (s.syncing && (s.uptodate < disks)) || s.expanding)
2895 handle_stripe_fill6(sh, &s, &r6s, disks);
2897 /* now to consider writing and what else, if anything should be read */
2899 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
2901 /* maybe we need to check and possibly fix the parity for this stripe
2902 * Any reads will already have been scheduled, so we just see if enough
2905 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
2906 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
2908 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2909 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2910 clear_bit(STRIPE_SYNCING, &sh->state);
2913 /* If the failed drives are just a ReadError, then we might need
2914 * to progress the repair/check process
2916 if (s.failed <= 2 && !conf->mddev->ro)
2917 for (i = 0; i < s.failed; i++) {
2918 dev = &sh->dev[r6s.failed_num[i]];
2919 if (test_bit(R5_ReadError, &dev->flags)
2920 && !test_bit(R5_LOCKED, &dev->flags)
2921 && test_bit(R5_UPTODATE, &dev->flags)
2923 if (!test_bit(R5_ReWrite, &dev->flags)) {
2924 set_bit(R5_Wantwrite, &dev->flags);
2925 set_bit(R5_ReWrite, &dev->flags);
2926 set_bit(R5_LOCKED, &dev->flags);
2928 /* let's read it back */
2929 set_bit(R5_Wantread, &dev->flags);
2930 set_bit(R5_LOCKED, &dev->flags);
2935 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
2936 /* Need to write out all blocks after computing P&Q */
2937 sh->disks = conf->raid_disks;
2938 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2940 compute_parity6(sh, RECONSTRUCT_WRITE);
2941 for (i = conf->raid_disks ; i-- ; ) {
2942 set_bit(R5_LOCKED, &sh->dev[i].flags);
2944 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2946 clear_bit(STRIPE_EXPANDING, &sh->state);
2947 } else if (s.expanded) {
2948 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2949 atomic_dec(&conf->reshape_stripes);
2950 wake_up(&conf->wait_for_overlap);
2951 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2954 if (s.expanding && s.locked == 0 &&
2955 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2956 handle_stripe_expansion(conf, sh, &r6s);
2959 spin_unlock(&sh->lock);
2961 /* wait for this device to become unblocked */
2962 if (unlikely(blocked_rdev))
2963 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2967 return_io(return_bi);
2970 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
2972 if (sh->raid_conf->level == 6)
2973 handle_stripe6(sh, tmp_page);
2980 static void raid5_activate_delayed(raid5_conf_t *conf)
2982 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
2983 while (!list_empty(&conf->delayed_list)) {
2984 struct list_head *l = conf->delayed_list.next;
2985 struct stripe_head *sh;
2986 sh = list_entry(l, struct stripe_head, lru);
2988 clear_bit(STRIPE_DELAYED, &sh->state);
2989 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
2990 atomic_inc(&conf->preread_active_stripes);
2991 list_add_tail(&sh->lru, &conf->hold_list);
2994 blk_plug_device(conf->mddev->queue);
2997 static void activate_bit_delay(raid5_conf_t *conf)
2999 /* device_lock is held */
3000 struct list_head head;
3001 list_add(&head, &conf->bitmap_list);
3002 list_del_init(&conf->bitmap_list);
3003 while (!list_empty(&head)) {
3004 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3005 list_del_init(&sh->lru);
3006 atomic_inc(&sh->count);
3007 __release_stripe(conf, sh);
3011 static void unplug_slaves(mddev_t *mddev)
3013 raid5_conf_t *conf = mddev_to_conf(mddev);
3017 for (i=0; i<mddev->raid_disks; i++) {
3018 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3019 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3020 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3022 atomic_inc(&rdev->nr_pending);
3025 blk_unplug(r_queue);
3027 rdev_dec_pending(rdev, mddev);
3034 static void raid5_unplug_device(struct request_queue *q)
3036 mddev_t *mddev = q->queuedata;
3037 raid5_conf_t *conf = mddev_to_conf(mddev);
3038 unsigned long flags;
3040 spin_lock_irqsave(&conf->device_lock, flags);
3042 if (blk_remove_plug(q)) {
3044 raid5_activate_delayed(conf);
3046 md_wakeup_thread(mddev->thread);
3048 spin_unlock_irqrestore(&conf->device_lock, flags);
3050 unplug_slaves(mddev);
3053 static int raid5_congested(void *data, int bits)
3055 mddev_t *mddev = data;
3056 raid5_conf_t *conf = mddev_to_conf(mddev);
3058 /* No difference between reads and writes. Just check
3059 * how busy the stripe_cache is
3061 if (conf->inactive_blocked)
3065 if (list_empty_careful(&conf->inactive_list))
3071 /* We want read requests to align with chunks where possible,
3072 * but write requests don't need to.
3074 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3076 mddev_t *mddev = q->queuedata;
3077 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3079 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3080 unsigned int bio_sectors = bio->bi_size >> 9;
3082 if (bio_data_dir(bio) == WRITE)
3083 return biovec->bv_len; /* always allow writes to be mergeable */
3085 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3086 if (max < 0) max = 0;
3087 if (max <= biovec->bv_len && bio_sectors == 0)
3088 return biovec->bv_len;
3094 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3096 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3097 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3098 unsigned int bio_sectors = bio->bi_size >> 9;
3100 return chunk_sectors >=
3101 ((sector & (chunk_sectors - 1)) + bio_sectors);
3105 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3106 * later sampled by raid5d.
3108 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3110 unsigned long flags;
3112 spin_lock_irqsave(&conf->device_lock, flags);
3114 bi->bi_next = conf->retry_read_aligned_list;
3115 conf->retry_read_aligned_list = bi;
3117 spin_unlock_irqrestore(&conf->device_lock, flags);
3118 md_wakeup_thread(conf->mddev->thread);
3122 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3126 bi = conf->retry_read_aligned;
3128 conf->retry_read_aligned = NULL;
3131 bi = conf->retry_read_aligned_list;
3133 conf->retry_read_aligned_list = bi->bi_next;
3135 bi->bi_phys_segments = 1; /* biased count of active stripes */
3136 bi->bi_hw_segments = 0; /* count of processed stripes */
3144 * The "raid5_align_endio" should check if the read succeeded and if it
3145 * did, call bio_endio on the original bio (having bio_put the new bio
3147 * If the read failed..
3149 static void raid5_align_endio(struct bio *bi, int error)
3151 struct bio* raid_bi = bi->bi_private;
3154 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3159 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3160 conf = mddev_to_conf(mddev);
3161 rdev = (void*)raid_bi->bi_next;
3162 raid_bi->bi_next = NULL;
3164 rdev_dec_pending(rdev, conf->mddev);
3166 if (!error && uptodate) {
3167 bio_endio(raid_bi, 0);
3168 if (atomic_dec_and_test(&conf->active_aligned_reads))
3169 wake_up(&conf->wait_for_stripe);
3174 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3176 add_bio_to_retry(raid_bi, conf);
3179 static int bio_fits_rdev(struct bio *bi)
3181 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3183 if ((bi->bi_size>>9) > q->max_sectors)
3185 blk_recount_segments(q, bi);
3186 if (bi->bi_phys_segments > q->max_phys_segments ||
3187 bi->bi_hw_segments > q->max_hw_segments)
3190 if (q->merge_bvec_fn)
3191 /* it's too hard to apply the merge_bvec_fn at this stage,
3200 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3202 mddev_t *mddev = q->queuedata;
3203 raid5_conf_t *conf = mddev_to_conf(mddev);
3204 const unsigned int raid_disks = conf->raid_disks;
3205 const unsigned int data_disks = raid_disks - conf->max_degraded;
3206 unsigned int dd_idx, pd_idx;
3207 struct bio* align_bi;
3210 if (!in_chunk_boundary(mddev, raid_bio)) {
3211 pr_debug("chunk_aligned_read : non aligned\n");
3215 * use bio_clone to make a copy of the bio
3217 align_bi = bio_clone(raid_bio, GFP_NOIO);
3221 * set bi_end_io to a new function, and set bi_private to the
3224 align_bi->bi_end_io = raid5_align_endio;
3225 align_bi->bi_private = raid_bio;
3229 align_bi->bi_sector = raid5_compute_sector(raid_bio->bi_sector,
3237 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3238 if (rdev && test_bit(In_sync, &rdev->flags)) {
3239 atomic_inc(&rdev->nr_pending);
3241 raid_bio->bi_next = (void*)rdev;
3242 align_bi->bi_bdev = rdev->bdev;
3243 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3244 align_bi->bi_sector += rdev->data_offset;
3246 if (!bio_fits_rdev(align_bi)) {
3247 /* too big in some way */
3249 rdev_dec_pending(rdev, mddev);
3253 spin_lock_irq(&conf->device_lock);
3254 wait_event_lock_irq(conf->wait_for_stripe,
3256 conf->device_lock, /* nothing */);
3257 atomic_inc(&conf->active_aligned_reads);
3258 spin_unlock_irq(&conf->device_lock);
3260 generic_make_request(align_bi);
3269 /* __get_priority_stripe - get the next stripe to process
3271 * Full stripe writes are allowed to pass preread active stripes up until
3272 * the bypass_threshold is exceeded. In general the bypass_count
3273 * increments when the handle_list is handled before the hold_list; however, it
3274 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3275 * stripe with in flight i/o. The bypass_count will be reset when the
3276 * head of the hold_list has changed, i.e. the head was promoted to the
3279 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3281 struct stripe_head *sh;
3283 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3285 list_empty(&conf->handle_list) ? "empty" : "busy",
3286 list_empty(&conf->hold_list) ? "empty" : "busy",
3287 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3289 if (!list_empty(&conf->handle_list)) {
3290 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3292 if (list_empty(&conf->hold_list))
3293 conf->bypass_count = 0;
3294 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3295 if (conf->hold_list.next == conf->last_hold)
3296 conf->bypass_count++;
3298 conf->last_hold = conf->hold_list.next;
3299 conf->bypass_count -= conf->bypass_threshold;
3300 if (conf->bypass_count < 0)
3301 conf->bypass_count = 0;
3304 } else if (!list_empty(&conf->hold_list) &&
3305 ((conf->bypass_threshold &&
3306 conf->bypass_count > conf->bypass_threshold) ||
3307 atomic_read(&conf->pending_full_writes) == 0)) {
3308 sh = list_entry(conf->hold_list.next,
3310 conf->bypass_count -= conf->bypass_threshold;
3311 if (conf->bypass_count < 0)
3312 conf->bypass_count = 0;
3316 list_del_init(&sh->lru);
3317 atomic_inc(&sh->count);
3318 BUG_ON(atomic_read(&sh->count) != 1);
3322 static int make_request(struct request_queue *q, struct bio * bi)
3324 mddev_t *mddev = q->queuedata;
3325 raid5_conf_t *conf = mddev_to_conf(mddev);
3326 unsigned int dd_idx, pd_idx;
3327 sector_t new_sector;
3328 sector_t logical_sector, last_sector;
3329 struct stripe_head *sh;
3330 const int rw = bio_data_dir(bi);
3333 if (unlikely(bio_barrier(bi))) {
3334 bio_endio(bi, -EOPNOTSUPP);
3338 md_write_start(mddev, bi);
3340 disk_stat_inc(mddev->gendisk, ios[rw]);
3341 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3344 mddev->reshape_position == MaxSector &&
3345 chunk_aligned_read(q,bi))
3348 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3349 last_sector = bi->bi_sector + (bi->bi_size>>9);
3351 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3353 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3355 int disks, data_disks;
3358 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3359 if (likely(conf->expand_progress == MaxSector))
3360 disks = conf->raid_disks;
3362 /* spinlock is needed as expand_progress may be
3363 * 64bit on a 32bit platform, and so it might be
3364 * possible to see a half-updated value
3365 * Ofcourse expand_progress could change after
3366 * the lock is dropped, so once we get a reference
3367 * to the stripe that we think it is, we will have
3370 spin_lock_irq(&conf->device_lock);
3371 disks = conf->raid_disks;
3372 if (logical_sector >= conf->expand_progress)
3373 disks = conf->previous_raid_disks;
3375 if (logical_sector >= conf->expand_lo) {
3376 spin_unlock_irq(&conf->device_lock);
3381 spin_unlock_irq(&conf->device_lock);
3383 data_disks = disks - conf->max_degraded;
3385 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3386 &dd_idx, &pd_idx, conf);
3387 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3388 (unsigned long long)new_sector,
3389 (unsigned long long)logical_sector);
3391 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3393 if (unlikely(conf->expand_progress != MaxSector)) {
3394 /* expansion might have moved on while waiting for a
3395 * stripe, so we must do the range check again.
3396 * Expansion could still move past after this
3397 * test, but as we are holding a reference to
3398 * 'sh', we know that if that happens,
3399 * STRIPE_EXPANDING will get set and the expansion
3400 * won't proceed until we finish with the stripe.
3403 spin_lock_irq(&conf->device_lock);
3404 if (logical_sector < conf->expand_progress &&
3405 disks == conf->previous_raid_disks)
3406 /* mismatch, need to try again */
3408 spin_unlock_irq(&conf->device_lock);
3414 /* FIXME what if we get a false positive because these
3415 * are being updated.
3417 if (logical_sector >= mddev->suspend_lo &&
3418 logical_sector < mddev->suspend_hi) {
3424 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3425 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3426 /* Stripe is busy expanding or
3427 * add failed due to overlap. Flush everything
3430 raid5_unplug_device(mddev->queue);
3435 finish_wait(&conf->wait_for_overlap, &w);
3436 set_bit(STRIPE_HANDLE, &sh->state);
3437 clear_bit(STRIPE_DELAYED, &sh->state);
3440 /* cannot get stripe for read-ahead, just give-up */
3441 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3442 finish_wait(&conf->wait_for_overlap, &w);
3447 spin_lock_irq(&conf->device_lock);
3448 remaining = --bi->bi_phys_segments;
3449 spin_unlock_irq(&conf->device_lock);
3450 if (remaining == 0) {
3453 md_write_end(mddev);
3460 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3462 /* reshaping is quite different to recovery/resync so it is
3463 * handled quite separately ... here.
3465 * On each call to sync_request, we gather one chunk worth of
3466 * destination stripes and flag them as expanding.
3467 * Then we find all the source stripes and request reads.
3468 * As the reads complete, handle_stripe will copy the data
3469 * into the destination stripe and release that stripe.
3471 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3472 struct stripe_head *sh;
3474 sector_t first_sector, last_sector;
3475 int raid_disks = conf->previous_raid_disks;
3476 int data_disks = raid_disks - conf->max_degraded;
3477 int new_data_disks = conf->raid_disks - conf->max_degraded;
3480 sector_t writepos, safepos, gap;
3482 if (sector_nr == 0 &&
3483 conf->expand_progress != 0) {
3484 /* restarting in the middle, skip the initial sectors */
3485 sector_nr = conf->expand_progress;
3486 sector_div(sector_nr, new_data_disks);
3491 /* we update the metadata when there is more than 3Meg
3492 * in the block range (that is rather arbitrary, should
3493 * probably be time based) or when the data about to be
3494 * copied would over-write the source of the data at
3495 * the front of the range.
3496 * i.e. one new_stripe forward from expand_progress new_maps
3497 * to after where expand_lo old_maps to
3499 writepos = conf->expand_progress +
3500 conf->chunk_size/512*(new_data_disks);
3501 sector_div(writepos, new_data_disks);
3502 safepos = conf->expand_lo;
3503 sector_div(safepos, data_disks);
3504 gap = conf->expand_progress - conf->expand_lo;
3506 if (writepos >= safepos ||
3507 gap > (new_data_disks)*3000*2 /*3Meg*/) {
3508 /* Cannot proceed until we've updated the superblock... */
3509 wait_event(conf->wait_for_overlap,
3510 atomic_read(&conf->reshape_stripes)==0);
3511 mddev->reshape_position = conf->expand_progress;
3512 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3513 md_wakeup_thread(mddev->thread);
3514 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3515 kthread_should_stop());
3516 spin_lock_irq(&conf->device_lock);
3517 conf->expand_lo = mddev->reshape_position;
3518 spin_unlock_irq(&conf->device_lock);
3519 wake_up(&conf->wait_for_overlap);
3522 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3525 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3526 sh = get_active_stripe(conf, sector_nr+i,
3527 conf->raid_disks, pd_idx, 0);
3528 set_bit(STRIPE_EXPANDING, &sh->state);
3529 atomic_inc(&conf->reshape_stripes);
3530 /* If any of this stripe is beyond the end of the old
3531 * array, then we need to zero those blocks
3533 for (j=sh->disks; j--;) {
3535 if (j == sh->pd_idx)
3537 if (conf->level == 6 &&
3538 j == raid6_next_disk(sh->pd_idx, sh->disks))
3540 s = compute_blocknr(sh, j);
3541 if (s < (mddev->array_size<<1)) {
3545 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3546 set_bit(R5_Expanded, &sh->dev[j].flags);
3547 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3550 set_bit(STRIPE_EXPAND_READY, &sh->state);
3551 set_bit(STRIPE_HANDLE, &sh->state);
3555 spin_lock_irq(&conf->device_lock);
3556 conf->expand_progress = (sector_nr + i) * new_data_disks;
3557 spin_unlock_irq(&conf->device_lock);
3558 /* Ok, those stripe are ready. We can start scheduling
3559 * reads on the source stripes.
3560 * The source stripes are determined by mapping the first and last
3561 * block on the destination stripes.
3564 raid5_compute_sector(sector_nr*(new_data_disks),
3565 raid_disks, data_disks,
3566 &dd_idx, &pd_idx, conf);
3568 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3569 *(new_data_disks) -1,
3570 raid_disks, data_disks,
3571 &dd_idx, &pd_idx, conf);
3572 if (last_sector >= (mddev->size<<1))
3573 last_sector = (mddev->size<<1)-1;
3574 while (first_sector <= last_sector) {
3575 pd_idx = stripe_to_pdidx(first_sector, conf,
3576 conf->previous_raid_disks);
3577 sh = get_active_stripe(conf, first_sector,
3578 conf->previous_raid_disks, pd_idx, 0);
3579 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3580 set_bit(STRIPE_HANDLE, &sh->state);
3582 first_sector += STRIPE_SECTORS;
3584 /* If this takes us to the resync_max point where we have to pause,
3585 * then we need to write out the superblock.
3587 sector_nr += conf->chunk_size>>9;
3588 if (sector_nr >= mddev->resync_max) {
3589 /* Cannot proceed until we've updated the superblock... */
3590 wait_event(conf->wait_for_overlap,
3591 atomic_read(&conf->reshape_stripes) == 0);
3592 mddev->reshape_position = conf->expand_progress;
3593 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3594 md_wakeup_thread(mddev->thread);
3595 wait_event(mddev->sb_wait,
3596 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3597 || kthread_should_stop());
3598 spin_lock_irq(&conf->device_lock);
3599 conf->expand_lo = mddev->reshape_position;
3600 spin_unlock_irq(&conf->device_lock);
3601 wake_up(&conf->wait_for_overlap);
3603 return conf->chunk_size>>9;
3606 /* FIXME go_faster isn't used */
3607 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3609 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3610 struct stripe_head *sh;
3612 int raid_disks = conf->raid_disks;
3613 sector_t max_sector = mddev->size << 1;
3615 int still_degraded = 0;
3618 if (sector_nr >= max_sector) {
3619 /* just being told to finish up .. nothing much to do */
3620 unplug_slaves(mddev);
3621 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3626 if (mddev->curr_resync < max_sector) /* aborted */
3627 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3629 else /* completed sync */
3631 bitmap_close_sync(mddev->bitmap);
3636 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3637 return reshape_request(mddev, sector_nr, skipped);
3639 /* No need to check resync_max as we never do more than one
3640 * stripe, and as resync_max will always be on a chunk boundary,
3641 * if the check in md_do_sync didn't fire, there is no chance
3642 * of overstepping resync_max here
3645 /* if there is too many failed drives and we are trying
3646 * to resync, then assert that we are finished, because there is
3647 * nothing we can do.
3649 if (mddev->degraded >= conf->max_degraded &&
3650 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3651 sector_t rv = (mddev->size << 1) - sector_nr;
3655 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3656 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3657 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3658 /* we can skip this block, and probably more */
3659 sync_blocks /= STRIPE_SECTORS;
3661 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3665 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3667 pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3668 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3670 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3671 /* make sure we don't swamp the stripe cache if someone else
3672 * is trying to get access
3674 schedule_timeout_uninterruptible(1);
3676 /* Need to check if array will still be degraded after recovery/resync
3677 * We don't need to check the 'failed' flag as when that gets set,
3680 for (i=0; i<mddev->raid_disks; i++)
3681 if (conf->disks[i].rdev == NULL)
3684 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3686 spin_lock(&sh->lock);
3687 set_bit(STRIPE_SYNCING, &sh->state);
3688 clear_bit(STRIPE_INSYNC, &sh->state);
3689 spin_unlock(&sh->lock);
3691 handle_stripe(sh, NULL);
3694 return STRIPE_SECTORS;
3697 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3699 /* We may not be able to submit a whole bio at once as there
3700 * may not be enough stripe_heads available.
3701 * We cannot pre-allocate enough stripe_heads as we may need
3702 * more than exist in the cache (if we allow ever large chunks).
3703 * So we do one stripe head at a time and record in
3704 * ->bi_hw_segments how many have been done.
3706 * We *know* that this entire raid_bio is in one chunk, so
3707 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3709 struct stripe_head *sh;
3711 sector_t sector, logical_sector, last_sector;
3716 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3717 sector = raid5_compute_sector( logical_sector,
3719 conf->raid_disks - conf->max_degraded,
3723 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3725 for (; logical_sector < last_sector;
3726 logical_sector += STRIPE_SECTORS,
3727 sector += STRIPE_SECTORS,
3730 if (scnt < raid_bio->bi_hw_segments)
3731 /* already done this stripe */
3734 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3737 /* failed to get a stripe - must wait */
3738 raid_bio->bi_hw_segments = scnt;
3739 conf->retry_read_aligned = raid_bio;
3743 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3744 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3746 raid_bio->bi_hw_segments = scnt;
3747 conf->retry_read_aligned = raid_bio;
3751 handle_stripe(sh, NULL);
3755 spin_lock_irq(&conf->device_lock);
3756 remaining = --raid_bio->bi_phys_segments;
3757 spin_unlock_irq(&conf->device_lock);
3759 bio_endio(raid_bio, 0);
3760 if (atomic_dec_and_test(&conf->active_aligned_reads))
3761 wake_up(&conf->wait_for_stripe);
3768 * This is our raid5 kernel thread.
3770 * We scan the hash table for stripes which can be handled now.
3771 * During the scan, completed stripes are saved for us by the interrupt
3772 * handler, so that they will not have to wait for our next wakeup.
3774 static void raid5d(mddev_t *mddev)
3776 struct stripe_head *sh;
3777 raid5_conf_t *conf = mddev_to_conf(mddev);
3780 pr_debug("+++ raid5d active\n");
3782 md_check_recovery(mddev);
3785 spin_lock_irq(&conf->device_lock);
3789 if (conf->seq_flush != conf->seq_write) {
3790 int seq = conf->seq_flush;
3791 spin_unlock_irq(&conf->device_lock);
3792 bitmap_unplug(mddev->bitmap);
3793 spin_lock_irq(&conf->device_lock);
3794 conf->seq_write = seq;
3795 activate_bit_delay(conf);
3798 while ((bio = remove_bio_from_retry(conf))) {
3800 spin_unlock_irq(&conf->device_lock);
3801 ok = retry_aligned_read(conf, bio);
3802 spin_lock_irq(&conf->device_lock);
3808 sh = __get_priority_stripe(conf);
3811 async_tx_issue_pending_all();
3814 spin_unlock_irq(&conf->device_lock);
3817 handle_stripe(sh, conf->spare_page);
3820 spin_lock_irq(&conf->device_lock);
3822 pr_debug("%d stripes handled\n", handled);
3824 spin_unlock_irq(&conf->device_lock);
3826 unplug_slaves(mddev);
3828 pr_debug("--- raid5d inactive\n");
3832 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3834 raid5_conf_t *conf = mddev_to_conf(mddev);
3836 return sprintf(page, "%d\n", conf->max_nr_stripes);
3842 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3844 raid5_conf_t *conf = mddev_to_conf(mddev);
3846 if (len >= PAGE_SIZE)
3851 if (strict_strtoul(page, 10, &new))
3853 if (new <= 16 || new > 32768)
3855 while (new < conf->max_nr_stripes) {
3856 if (drop_one_stripe(conf))
3857 conf->max_nr_stripes--;
3861 md_allow_write(mddev);
3862 while (new > conf->max_nr_stripes) {
3863 if (grow_one_stripe(conf))
3864 conf->max_nr_stripes++;
3870 static struct md_sysfs_entry
3871 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
3872 raid5_show_stripe_cache_size,
3873 raid5_store_stripe_cache_size);
3876 raid5_show_preread_threshold(mddev_t *mddev, char *page)
3878 raid5_conf_t *conf = mddev_to_conf(mddev);
3880 return sprintf(page, "%d\n", conf->bypass_threshold);
3886 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
3888 raid5_conf_t *conf = mddev_to_conf(mddev);
3890 if (len >= PAGE_SIZE)
3895 if (strict_strtoul(page, 10, &new))
3897 if (new > conf->max_nr_stripes)
3899 conf->bypass_threshold = new;
3903 static struct md_sysfs_entry
3904 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
3906 raid5_show_preread_threshold,
3907 raid5_store_preread_threshold);
3910 stripe_cache_active_show(mddev_t *mddev, char *page)
3912 raid5_conf_t *conf = mddev_to_conf(mddev);
3914 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
3919 static struct md_sysfs_entry
3920 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
3922 static struct attribute *raid5_attrs[] = {
3923 &raid5_stripecache_size.attr,
3924 &raid5_stripecache_active.attr,
3925 &raid5_preread_bypass_threshold.attr,
3928 static struct attribute_group raid5_attrs_group = {
3930 .attrs = raid5_attrs,
3933 static int run(mddev_t *mddev)
3936 int raid_disk, memory;
3938 struct disk_info *disk;
3939 struct list_head *tmp;
3940 int working_disks = 0;
3942 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
3943 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
3944 mdname(mddev), mddev->level);
3948 if (mddev->reshape_position != MaxSector) {
3949 /* Check that we can continue the reshape.
3950 * Currently only disks can change, it must
3951 * increase, and we must be past the point where
3952 * a stripe over-writes itself
3954 sector_t here_new, here_old;
3956 int max_degraded = (mddev->level == 5 ? 1 : 2);
3958 if (mddev->new_level != mddev->level ||
3959 mddev->new_layout != mddev->layout ||
3960 mddev->new_chunk != mddev->chunk_size) {
3961 printk(KERN_ERR "raid5: %s: unsupported reshape "
3962 "required - aborting.\n",
3966 if (mddev->delta_disks <= 0) {
3967 printk(KERN_ERR "raid5: %s: unsupported reshape "
3968 "(reduce disks) required - aborting.\n",
3972 old_disks = mddev->raid_disks - mddev->delta_disks;
3973 /* reshape_position must be on a new-stripe boundary, and one
3974 * further up in new geometry must map after here in old
3977 here_new = mddev->reshape_position;
3978 if (sector_div(here_new, (mddev->chunk_size>>9)*
3979 (mddev->raid_disks - max_degraded))) {
3980 printk(KERN_ERR "raid5: reshape_position not "
3981 "on a stripe boundary\n");
3984 /* here_new is the stripe we will write to */
3985 here_old = mddev->reshape_position;
3986 sector_div(here_old, (mddev->chunk_size>>9)*
3987 (old_disks-max_degraded));
3988 /* here_old is the first stripe that we might need to read
3990 if (here_new >= here_old) {
3991 /* Reading from the same stripe as writing to - bad */
3992 printk(KERN_ERR "raid5: reshape_position too early for "
3993 "auto-recovery - aborting.\n");
3996 printk(KERN_INFO "raid5: reshape will continue\n");
3997 /* OK, we should be able to continue; */
4001 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4002 if ((conf = mddev->private) == NULL)
4004 if (mddev->reshape_position == MaxSector) {
4005 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4007 conf->raid_disks = mddev->raid_disks;
4008 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4011 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4016 conf->mddev = mddev;
4018 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4021 if (mddev->level == 6) {
4022 conf->spare_page = alloc_page(GFP_KERNEL);
4023 if (!conf->spare_page)
4026 spin_lock_init(&conf->device_lock);
4027 mddev->queue->queue_lock = &conf->device_lock;
4028 init_waitqueue_head(&conf->wait_for_stripe);
4029 init_waitqueue_head(&conf->wait_for_overlap);
4030 INIT_LIST_HEAD(&conf->handle_list);
4031 INIT_LIST_HEAD(&conf->hold_list);
4032 INIT_LIST_HEAD(&conf->delayed_list);
4033 INIT_LIST_HEAD(&conf->bitmap_list);
4034 INIT_LIST_HEAD(&conf->inactive_list);
4035 atomic_set(&conf->active_stripes, 0);
4036 atomic_set(&conf->preread_active_stripes, 0);
4037 atomic_set(&conf->active_aligned_reads, 0);
4038 conf->bypass_threshold = BYPASS_THRESHOLD;
4040 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4042 rdev_for_each(rdev, tmp, mddev) {
4043 raid_disk = rdev->raid_disk;
4044 if (raid_disk >= conf->raid_disks
4047 disk = conf->disks + raid_disk;
4051 if (test_bit(In_sync, &rdev->flags)) {
4052 char b[BDEVNAME_SIZE];
4053 printk(KERN_INFO "raid5: device %s operational as raid"
4054 " disk %d\n", bdevname(rdev->bdev,b),
4058 /* Cannot rely on bitmap to complete recovery */
4063 * 0 for a fully functional array, 1 or 2 for a degraded array.
4065 mddev->degraded = conf->raid_disks - working_disks;
4066 conf->mddev = mddev;
4067 conf->chunk_size = mddev->chunk_size;
4068 conf->level = mddev->level;
4069 if (conf->level == 6)
4070 conf->max_degraded = 2;
4072 conf->max_degraded = 1;
4073 conf->algorithm = mddev->layout;
4074 conf->max_nr_stripes = NR_STRIPES;
4075 conf->expand_progress = mddev->reshape_position;
4077 /* device size must be a multiple of chunk size */
4078 mddev->size &= ~(mddev->chunk_size/1024 -1);
4079 mddev->resync_max_sectors = mddev->size << 1;
4081 if (conf->level == 6 && conf->raid_disks < 4) {
4082 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4083 mdname(mddev), conf->raid_disks);
4086 if (!conf->chunk_size || conf->chunk_size % 4) {
4087 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4088 conf->chunk_size, mdname(mddev));
4091 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4093 "raid5: unsupported parity algorithm %d for %s\n",
4094 conf->algorithm, mdname(mddev));
4097 if (mddev->degraded > conf->max_degraded) {
4098 printk(KERN_ERR "raid5: not enough operational devices for %s"
4099 " (%d/%d failed)\n",
4100 mdname(mddev), mddev->degraded, conf->raid_disks);
4104 if (mddev->degraded > 0 &&
4105 mddev->recovery_cp != MaxSector) {
4106 if (mddev->ok_start_degraded)
4108 "raid5: starting dirty degraded array: %s"
4109 "- data corruption possible.\n",
4113 "raid5: cannot start dirty degraded array for %s\n",
4120 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4121 if (!mddev->thread) {
4123 "raid5: couldn't allocate thread for %s\n",
4128 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4129 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4130 if (grow_stripes(conf, conf->max_nr_stripes)) {
4132 "raid5: couldn't allocate %dkB for buffers\n", memory);
4133 shrink_stripes(conf);
4134 md_unregister_thread(mddev->thread);
4137 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4138 memory, mdname(mddev));
4140 if (mddev->degraded == 0)
4141 printk("raid5: raid level %d set %s active with %d out of %d"
4142 " devices, algorithm %d\n", conf->level, mdname(mddev),
4143 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4146 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4147 " out of %d devices, algorithm %d\n", conf->level,
4148 mdname(mddev), mddev->raid_disks - mddev->degraded,
4149 mddev->raid_disks, conf->algorithm);
4151 print_raid5_conf(conf);
4153 if (conf->expand_progress != MaxSector) {
4154 printk("...ok start reshape thread\n");
4155 conf->expand_lo = conf->expand_progress;
4156 atomic_set(&conf->reshape_stripes, 0);
4157 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4158 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4159 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4160 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4161 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4165 /* read-ahead size must cover two whole stripes, which is
4166 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4169 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4170 int stripe = data_disks *
4171 (mddev->chunk_size / PAGE_SIZE);
4172 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4173 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4176 /* Ok, everything is just fine now */
4177 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4179 "raid5: failed to create sysfs attributes for %s\n",
4182 mddev->queue->unplug_fn = raid5_unplug_device;
4183 mddev->queue->backing_dev_info.congested_data = mddev;
4184 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4186 mddev->array_size = mddev->size * (conf->previous_raid_disks -
4187 conf->max_degraded);
4189 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4194 print_raid5_conf(conf);
4195 safe_put_page(conf->spare_page);
4197 kfree(conf->stripe_hashtbl);
4200 mddev->private = NULL;
4201 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4207 static int stop(mddev_t *mddev)
4209 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4211 md_unregister_thread(mddev->thread);
4212 mddev->thread = NULL;
4213 shrink_stripes(conf);
4214 kfree(conf->stripe_hashtbl);
4215 mddev->queue->backing_dev_info.congested_fn = NULL;
4216 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4217 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4220 mddev->private = NULL;
4225 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4229 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4230 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4231 seq_printf(seq, "sh %llu, count %d.\n",
4232 (unsigned long long)sh->sector, atomic_read(&sh->count));
4233 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4234 for (i = 0; i < sh->disks; i++) {
4235 seq_printf(seq, "(cache%d: %p %ld) ",
4236 i, sh->dev[i].page, sh->dev[i].flags);
4238 seq_printf(seq, "\n");
4241 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4243 struct stripe_head *sh;
4244 struct hlist_node *hn;
4247 spin_lock_irq(&conf->device_lock);
4248 for (i = 0; i < NR_HASH; i++) {
4249 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4250 if (sh->raid_conf != conf)
4255 spin_unlock_irq(&conf->device_lock);
4259 static void status (struct seq_file *seq, mddev_t *mddev)
4261 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4264 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4265 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4266 for (i = 0; i < conf->raid_disks; i++)
4267 seq_printf (seq, "%s",
4268 conf->disks[i].rdev &&
4269 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4270 seq_printf (seq, "]");
4272 seq_printf (seq, "\n");
4273 printall(seq, conf);
4277 static void print_raid5_conf (raid5_conf_t *conf)
4280 struct disk_info *tmp;
4282 printk("RAID5 conf printout:\n");
4284 printk("(conf==NULL)\n");
4287 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4288 conf->raid_disks - conf->mddev->degraded);
4290 for (i = 0; i < conf->raid_disks; i++) {
4291 char b[BDEVNAME_SIZE];
4292 tmp = conf->disks + i;
4294 printk(" disk %d, o:%d, dev:%s\n",
4295 i, !test_bit(Faulty, &tmp->rdev->flags),
4296 bdevname(tmp->rdev->bdev,b));
4300 static int raid5_spare_active(mddev_t *mddev)
4303 raid5_conf_t *conf = mddev->private;
4304 struct disk_info *tmp;
4306 for (i = 0; i < conf->raid_disks; i++) {
4307 tmp = conf->disks + i;
4309 && !test_bit(Faulty, &tmp->rdev->flags)
4310 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4311 unsigned long flags;
4312 spin_lock_irqsave(&conf->device_lock, flags);
4314 spin_unlock_irqrestore(&conf->device_lock, flags);
4317 print_raid5_conf(conf);
4321 static int raid5_remove_disk(mddev_t *mddev, int number)
4323 raid5_conf_t *conf = mddev->private;
4326 struct disk_info *p = conf->disks + number;
4328 print_raid5_conf(conf);
4331 if (test_bit(In_sync, &rdev->flags) ||
4332 atomic_read(&rdev->nr_pending)) {
4336 /* Only remove non-faulty devices if recovery
4339 if (!test_bit(Faulty, &rdev->flags) &&
4340 mddev->degraded <= conf->max_degraded) {
4346 if (atomic_read(&rdev->nr_pending)) {
4347 /* lost the race, try later */
4354 print_raid5_conf(conf);
4358 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4360 raid5_conf_t *conf = mddev->private;
4363 struct disk_info *p;
4365 int last = conf->raid_disks - 1;
4367 if (mddev->degraded > conf->max_degraded)
4368 /* no point adding a device */
4371 if (rdev->raid_disk >= 0)
4372 first = last = rdev->raid_disk;
4375 * find the disk ... but prefer rdev->saved_raid_disk
4378 if (rdev->saved_raid_disk >= 0 &&
4379 rdev->saved_raid_disk >= first &&
4380 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4381 disk = rdev->saved_raid_disk;
4384 for ( ; disk <= last ; disk++)
4385 if ((p=conf->disks + disk)->rdev == NULL) {
4386 clear_bit(In_sync, &rdev->flags);
4387 rdev->raid_disk = disk;
4389 if (rdev->saved_raid_disk != disk)
4391 rcu_assign_pointer(p->rdev, rdev);
4394 print_raid5_conf(conf);
4398 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4400 /* no resync is happening, and there is enough space
4401 * on all devices, so we can resize.
4402 * We need to make sure resync covers any new space.
4403 * If the array is shrinking we should possibly wait until
4404 * any io in the removed space completes, but it hardly seems
4407 raid5_conf_t *conf = mddev_to_conf(mddev);
4409 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4410 mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4411 set_capacity(mddev->gendisk, mddev->array_size << 1);
4413 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) {
4414 mddev->recovery_cp = mddev->size << 1;
4415 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4417 mddev->size = sectors /2;
4418 mddev->resync_max_sectors = sectors;
4422 #ifdef CONFIG_MD_RAID5_RESHAPE
4423 static int raid5_check_reshape(mddev_t *mddev)
4425 raid5_conf_t *conf = mddev_to_conf(mddev);
4428 if (mddev->delta_disks < 0 ||
4429 mddev->new_level != mddev->level)
4430 return -EINVAL; /* Cannot shrink array or change level yet */
4431 if (mddev->delta_disks == 0)
4432 return 0; /* nothing to do */
4434 /* Can only proceed if there are plenty of stripe_heads.
4435 * We need a minimum of one full stripe,, and for sensible progress
4436 * it is best to have about 4 times that.
4437 * If we require 4 times, then the default 256 4K stripe_heads will
4438 * allow for chunk sizes up to 256K, which is probably OK.
4439 * If the chunk size is greater, user-space should request more
4440 * stripe_heads first.
4442 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4443 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4444 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4445 (mddev->chunk_size / STRIPE_SIZE)*4);
4449 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4453 if (mddev->degraded > conf->max_degraded)
4455 /* looks like we might be able to manage this */
4459 static int raid5_start_reshape(mddev_t *mddev)
4461 raid5_conf_t *conf = mddev_to_conf(mddev);
4463 struct list_head *rtmp;
4465 int added_devices = 0;
4466 unsigned long flags;
4468 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4471 rdev_for_each(rdev, rtmp, mddev)
4472 if (rdev->raid_disk < 0 &&
4473 !test_bit(Faulty, &rdev->flags))
4476 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4477 /* Not enough devices even to make a degraded array
4482 atomic_set(&conf->reshape_stripes, 0);
4483 spin_lock_irq(&conf->device_lock);
4484 conf->previous_raid_disks = conf->raid_disks;
4485 conf->raid_disks += mddev->delta_disks;
4486 conf->expand_progress = 0;
4487 conf->expand_lo = 0;
4488 spin_unlock_irq(&conf->device_lock);
4490 /* Add some new drives, as many as will fit.
4491 * We know there are enough to make the newly sized array work.
4493 rdev_for_each(rdev, rtmp, mddev)
4494 if (rdev->raid_disk < 0 &&
4495 !test_bit(Faulty, &rdev->flags)) {
4496 if (raid5_add_disk(mddev, rdev) == 0) {
4498 set_bit(In_sync, &rdev->flags);
4500 rdev->recovery_offset = 0;
4501 sprintf(nm, "rd%d", rdev->raid_disk);
4502 if (sysfs_create_link(&mddev->kobj,
4505 "raid5: failed to create "
4506 " link %s for %s\n",
4512 spin_lock_irqsave(&conf->device_lock, flags);
4513 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4514 spin_unlock_irqrestore(&conf->device_lock, flags);
4515 mddev->raid_disks = conf->raid_disks;
4516 mddev->reshape_position = 0;
4517 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4519 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4520 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4521 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4522 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4523 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4525 if (!mddev->sync_thread) {
4526 mddev->recovery = 0;
4527 spin_lock_irq(&conf->device_lock);
4528 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4529 conf->expand_progress = MaxSector;
4530 spin_unlock_irq(&conf->device_lock);
4533 md_wakeup_thread(mddev->sync_thread);
4534 md_new_event(mddev);
4539 static void end_reshape(raid5_conf_t *conf)
4541 struct block_device *bdev;
4543 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4544 conf->mddev->array_size = conf->mddev->size *
4545 (conf->raid_disks - conf->max_degraded);
4546 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4547 conf->mddev->changed = 1;
4549 bdev = bdget_disk(conf->mddev->gendisk, 0);
4551 mutex_lock(&bdev->bd_inode->i_mutex);
4552 i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4553 mutex_unlock(&bdev->bd_inode->i_mutex);
4556 spin_lock_irq(&conf->device_lock);
4557 conf->expand_progress = MaxSector;
4558 spin_unlock_irq(&conf->device_lock);
4559 conf->mddev->reshape_position = MaxSector;
4561 /* read-ahead size must cover two whole stripes, which is
4562 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4565 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4566 int stripe = data_disks *
4567 (conf->mddev->chunk_size / PAGE_SIZE);
4568 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4569 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4574 static void raid5_quiesce(mddev_t *mddev, int state)
4576 raid5_conf_t *conf = mddev_to_conf(mddev);
4579 case 2: /* resume for a suspend */
4580 wake_up(&conf->wait_for_overlap);
4583 case 1: /* stop all writes */
4584 spin_lock_irq(&conf->device_lock);
4586 wait_event_lock_irq(conf->wait_for_stripe,
4587 atomic_read(&conf->active_stripes) == 0 &&
4588 atomic_read(&conf->active_aligned_reads) == 0,
4589 conf->device_lock, /* nothing */);
4590 spin_unlock_irq(&conf->device_lock);
4593 case 0: /* re-enable writes */
4594 spin_lock_irq(&conf->device_lock);
4596 wake_up(&conf->wait_for_stripe);
4597 wake_up(&conf->wait_for_overlap);
4598 spin_unlock_irq(&conf->device_lock);
4603 static struct mdk_personality raid6_personality =
4607 .owner = THIS_MODULE,
4608 .make_request = make_request,
4612 .error_handler = error,
4613 .hot_add_disk = raid5_add_disk,
4614 .hot_remove_disk= raid5_remove_disk,
4615 .spare_active = raid5_spare_active,
4616 .sync_request = sync_request,
4617 .resize = raid5_resize,
4618 #ifdef CONFIG_MD_RAID5_RESHAPE
4619 .check_reshape = raid5_check_reshape,
4620 .start_reshape = raid5_start_reshape,
4622 .quiesce = raid5_quiesce,
4624 static struct mdk_personality raid5_personality =
4628 .owner = THIS_MODULE,
4629 .make_request = make_request,
4633 .error_handler = error,
4634 .hot_add_disk = raid5_add_disk,
4635 .hot_remove_disk= raid5_remove_disk,
4636 .spare_active = raid5_spare_active,
4637 .sync_request = sync_request,
4638 .resize = raid5_resize,
4639 #ifdef CONFIG_MD_RAID5_RESHAPE
4640 .check_reshape = raid5_check_reshape,
4641 .start_reshape = raid5_start_reshape,
4643 .quiesce = raid5_quiesce,
4646 static struct mdk_personality raid4_personality =
4650 .owner = THIS_MODULE,
4651 .make_request = make_request,
4655 .error_handler = error,
4656 .hot_add_disk = raid5_add_disk,
4657 .hot_remove_disk= raid5_remove_disk,
4658 .spare_active = raid5_spare_active,
4659 .sync_request = sync_request,
4660 .resize = raid5_resize,
4661 #ifdef CONFIG_MD_RAID5_RESHAPE
4662 .check_reshape = raid5_check_reshape,
4663 .start_reshape = raid5_start_reshape,
4665 .quiesce = raid5_quiesce,
4668 static int __init raid5_init(void)
4672 e = raid6_select_algo();
4675 register_md_personality(&raid6_personality);
4676 register_md_personality(&raid5_personality);
4677 register_md_personality(&raid4_personality);
4681 static void raid5_exit(void)
4683 unregister_md_personality(&raid6_personality);
4684 unregister_md_personality(&raid5_personality);
4685 unregister_md_personality(&raid4_personality);
4688 module_init(raid5_init);
4689 module_exit(raid5_exit);
4690 MODULE_LICENSE("GPL");
4691 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4692 MODULE_ALIAS("md-raid5");
4693 MODULE_ALIAS("md-raid4");
4694 MODULE_ALIAS("md-level-5");
4695 MODULE_ALIAS("md-level-4");
4696 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4697 MODULE_ALIAS("md-raid6");
4698 MODULE_ALIAS("md-level-6");
4700 /* This used to be two separate modules, they were: */
4701 MODULE_ALIAS("raid5");
4702 MODULE_ALIAS("raid6");