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 NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
67 #define HASH_MASK (NR_HASH - 1)
69 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
71 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
72 * order without overlap. There may be several bio's per stripe+device, and
73 * a bio could span several devices.
74 * When walking this list for a particular stripe+device, we must never proceed
75 * beyond a bio that extends past this device, as the next bio might no longer
77 * This macro is used to determine the 'next' bio in the list, given the sector
78 * of the current stripe+device
80 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
82 * The following can be used to debug the driver
84 #define RAID5_PARANOIA 1
85 #if RAID5_PARANOIA && defined(CONFIG_SMP)
86 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
88 # define CHECK_DEVLOCK()
96 #if !RAID6_USE_EMPTY_ZERO_PAGE
97 /* In .bss so it's zeroed */
98 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
101 static inline int raid6_next_disk(int disk, int raid_disks)
104 return (disk < raid_disks) ? disk : 0;
107 static void return_io(struct bio *return_bi)
109 struct bio *bi = return_bi;
111 int bytes = bi->bi_size;
113 return_bi = bi->bi_next;
116 bi->bi_end_io(bi, bytes,
117 test_bit(BIO_UPTODATE, &bi->bi_flags)
123 static void print_raid5_conf (raid5_conf_t *conf);
125 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
127 if (atomic_dec_and_test(&sh->count)) {
128 BUG_ON(!list_empty(&sh->lru));
129 BUG_ON(atomic_read(&conf->active_stripes)==0);
130 if (test_bit(STRIPE_HANDLE, &sh->state)) {
131 if (test_bit(STRIPE_DELAYED, &sh->state)) {
132 list_add_tail(&sh->lru, &conf->delayed_list);
133 blk_plug_device(conf->mddev->queue);
134 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
135 sh->bm_seq - conf->seq_write > 0) {
136 list_add_tail(&sh->lru, &conf->bitmap_list);
137 blk_plug_device(conf->mddev->queue);
139 clear_bit(STRIPE_BIT_DELAY, &sh->state);
140 list_add_tail(&sh->lru, &conf->handle_list);
142 md_wakeup_thread(conf->mddev->thread);
144 BUG_ON(sh->ops.pending);
145 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
146 atomic_dec(&conf->preread_active_stripes);
147 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
148 md_wakeup_thread(conf->mddev->thread);
150 atomic_dec(&conf->active_stripes);
151 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
152 list_add_tail(&sh->lru, &conf->inactive_list);
153 wake_up(&conf->wait_for_stripe);
154 if (conf->retry_read_aligned)
155 md_wakeup_thread(conf->mddev->thread);
160 static void release_stripe(struct stripe_head *sh)
162 raid5_conf_t *conf = sh->raid_conf;
165 spin_lock_irqsave(&conf->device_lock, flags);
166 __release_stripe(conf, sh);
167 spin_unlock_irqrestore(&conf->device_lock, flags);
170 static inline void remove_hash(struct stripe_head *sh)
172 pr_debug("remove_hash(), stripe %llu\n",
173 (unsigned long long)sh->sector);
175 hlist_del_init(&sh->hash);
178 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
180 struct hlist_head *hp = stripe_hash(conf, sh->sector);
182 pr_debug("insert_hash(), stripe %llu\n",
183 (unsigned long long)sh->sector);
186 hlist_add_head(&sh->hash, hp);
190 /* find an idle stripe, make sure it is unhashed, and return it. */
191 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
193 struct stripe_head *sh = NULL;
194 struct list_head *first;
197 if (list_empty(&conf->inactive_list))
199 first = conf->inactive_list.next;
200 sh = list_entry(first, struct stripe_head, lru);
201 list_del_init(first);
203 atomic_inc(&conf->active_stripes);
208 static void shrink_buffers(struct stripe_head *sh, int num)
213 for (i=0; i<num ; i++) {
217 sh->dev[i].page = NULL;
222 static int grow_buffers(struct stripe_head *sh, int num)
226 for (i=0; i<num; i++) {
229 if (!(page = alloc_page(GFP_KERNEL))) {
232 sh->dev[i].page = page;
237 static void raid5_build_block (struct stripe_head *sh, int i);
239 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
241 raid5_conf_t *conf = sh->raid_conf;
244 BUG_ON(atomic_read(&sh->count) != 0);
245 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
246 BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete);
249 pr_debug("init_stripe called, stripe %llu\n",
250 (unsigned long long)sh->sector);
260 for (i = sh->disks; i--; ) {
261 struct r5dev *dev = &sh->dev[i];
263 if (dev->toread || dev->read || dev->towrite || dev->written ||
264 test_bit(R5_LOCKED, &dev->flags)) {
265 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
266 (unsigned long long)sh->sector, i, dev->toread,
267 dev->read, dev->towrite, dev->written,
268 test_bit(R5_LOCKED, &dev->flags));
272 raid5_build_block(sh, i);
274 insert_hash(conf, sh);
277 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
279 struct stripe_head *sh;
280 struct hlist_node *hn;
283 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
284 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
285 if (sh->sector == sector && sh->disks == disks)
287 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
291 static void unplug_slaves(mddev_t *mddev);
292 static void raid5_unplug_device(request_queue_t *q);
294 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
295 int pd_idx, int noblock)
297 struct stripe_head *sh;
299 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
301 spin_lock_irq(&conf->device_lock);
304 wait_event_lock_irq(conf->wait_for_stripe,
306 conf->device_lock, /* nothing */);
307 sh = __find_stripe(conf, sector, disks);
309 if (!conf->inactive_blocked)
310 sh = get_free_stripe(conf);
311 if (noblock && sh == NULL)
314 conf->inactive_blocked = 1;
315 wait_event_lock_irq(conf->wait_for_stripe,
316 !list_empty(&conf->inactive_list) &&
317 (atomic_read(&conf->active_stripes)
318 < (conf->max_nr_stripes *3/4)
319 || !conf->inactive_blocked),
321 raid5_unplug_device(conf->mddev->queue)
323 conf->inactive_blocked = 0;
325 init_stripe(sh, sector, pd_idx, disks);
327 if (atomic_read(&sh->count)) {
328 BUG_ON(!list_empty(&sh->lru));
330 if (!test_bit(STRIPE_HANDLE, &sh->state))
331 atomic_inc(&conf->active_stripes);
332 if (list_empty(&sh->lru) &&
333 !test_bit(STRIPE_EXPANDING, &sh->state))
335 list_del_init(&sh->lru);
338 } while (sh == NULL);
341 atomic_inc(&sh->count);
343 spin_unlock_irq(&conf->device_lock);
347 /* test_and_ack_op() ensures that we only dequeue an operation once */
348 #define test_and_ack_op(op, pend) \
350 if (test_bit(op, &sh->ops.pending) && \
351 !test_bit(op, &sh->ops.complete)) { \
352 if (test_and_set_bit(op, &sh->ops.ack)) \
353 clear_bit(op, &pend); \
357 clear_bit(op, &pend); \
360 /* find new work to run, do not resubmit work that is already
363 static unsigned long get_stripe_work(struct stripe_head *sh)
365 unsigned long pending;
368 pending = sh->ops.pending;
370 test_and_ack_op(STRIPE_OP_BIOFILL, pending);
371 test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending);
372 test_and_ack_op(STRIPE_OP_PREXOR, pending);
373 test_and_ack_op(STRIPE_OP_BIODRAIN, pending);
374 test_and_ack_op(STRIPE_OP_POSTXOR, pending);
375 test_and_ack_op(STRIPE_OP_CHECK, pending);
376 if (test_and_clear_bit(STRIPE_OP_IO, &sh->ops.pending))
379 sh->ops.count -= ack;
380 BUG_ON(sh->ops.count < 0);
386 raid5_end_read_request(struct bio *bi, unsigned int bytes_done, int error);
388 raid5_end_write_request (struct bio *bi, unsigned int bytes_done, int error);
390 static void ops_run_io(struct stripe_head *sh)
392 raid5_conf_t *conf = sh->raid_conf;
393 int i, disks = sh->disks;
397 for (i = disks; i--; ) {
401 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
403 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
408 bi = &sh->dev[i].req;
412 bi->bi_end_io = raid5_end_write_request;
414 bi->bi_end_io = raid5_end_read_request;
417 rdev = rcu_dereference(conf->disks[i].rdev);
418 if (rdev && test_bit(Faulty, &rdev->flags))
421 atomic_inc(&rdev->nr_pending);
425 if (test_bit(STRIPE_SYNCING, &sh->state) ||
426 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
427 test_bit(STRIPE_EXPAND_READY, &sh->state))
428 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
430 bi->bi_bdev = rdev->bdev;
431 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
432 __FUNCTION__, (unsigned long long)sh->sector,
434 atomic_inc(&sh->count);
435 bi->bi_sector = sh->sector + rdev->data_offset;
436 bi->bi_flags = 1 << BIO_UPTODATE;
440 bi->bi_io_vec = &sh->dev[i].vec;
441 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
442 bi->bi_io_vec[0].bv_offset = 0;
443 bi->bi_size = STRIPE_SIZE;
446 test_bit(R5_ReWrite, &sh->dev[i].flags))
447 atomic_add(STRIPE_SECTORS,
448 &rdev->corrected_errors);
449 generic_make_request(bi);
452 set_bit(STRIPE_DEGRADED, &sh->state);
453 pr_debug("skip op %ld on disc %d for sector %llu\n",
454 bi->bi_rw, i, (unsigned long long)sh->sector);
455 clear_bit(R5_LOCKED, &sh->dev[i].flags);
456 set_bit(STRIPE_HANDLE, &sh->state);
461 static struct dma_async_tx_descriptor *
462 async_copy_data(int frombio, struct bio *bio, struct page *page,
463 sector_t sector, struct dma_async_tx_descriptor *tx)
466 struct page *bio_page;
470 if (bio->bi_sector >= sector)
471 page_offset = (signed)(bio->bi_sector - sector) * 512;
473 page_offset = (signed)(sector - bio->bi_sector) * -512;
474 bio_for_each_segment(bvl, bio, i) {
475 int len = bio_iovec_idx(bio, i)->bv_len;
479 if (page_offset < 0) {
480 b_offset = -page_offset;
481 page_offset += b_offset;
485 if (len > 0 && page_offset + len > STRIPE_SIZE)
486 clen = STRIPE_SIZE - page_offset;
491 b_offset += bio_iovec_idx(bio, i)->bv_offset;
492 bio_page = bio_iovec_idx(bio, i)->bv_page;
494 tx = async_memcpy(page, bio_page, page_offset,
496 ASYNC_TX_DEP_ACK | ASYNC_TX_KMAP_SRC,
499 tx = async_memcpy(bio_page, page, b_offset,
501 ASYNC_TX_DEP_ACK | ASYNC_TX_KMAP_DST,
504 if (clen < len) /* hit end of page */
512 static void ops_complete_biofill(void *stripe_head_ref)
514 struct stripe_head *sh = stripe_head_ref;
515 struct bio *return_bi = NULL;
516 raid5_conf_t *conf = sh->raid_conf;
517 int i, more_to_read = 0;
519 pr_debug("%s: stripe %llu\n", __FUNCTION__,
520 (unsigned long long)sh->sector);
522 /* clear completed biofills */
523 for (i = sh->disks; i--; ) {
524 struct r5dev *dev = &sh->dev[i];
525 /* check if this stripe has new incoming reads */
529 /* acknowledge completion of a biofill operation */
530 /* and check if we need to reply to a read request
532 if (test_bit(R5_Wantfill, &dev->flags) && !dev->toread) {
533 struct bio *rbi, *rbi2;
534 clear_bit(R5_Wantfill, &dev->flags);
536 /* The access to dev->read is outside of the
537 * spin_lock_irq(&conf->device_lock), but is protected
538 * by the STRIPE_OP_BIOFILL pending bit
543 while (rbi && rbi->bi_sector <
544 dev->sector + STRIPE_SECTORS) {
545 rbi2 = r5_next_bio(rbi, dev->sector);
546 spin_lock_irq(&conf->device_lock);
547 if (--rbi->bi_phys_segments == 0) {
548 rbi->bi_next = return_bi;
551 spin_unlock_irq(&conf->device_lock);
556 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
557 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
559 return_io(return_bi);
562 set_bit(STRIPE_HANDLE, &sh->state);
566 static void ops_run_biofill(struct stripe_head *sh)
568 struct dma_async_tx_descriptor *tx = NULL;
569 raid5_conf_t *conf = sh->raid_conf;
572 pr_debug("%s: stripe %llu\n", __FUNCTION__,
573 (unsigned long long)sh->sector);
575 for (i = sh->disks; i--; ) {
576 struct r5dev *dev = &sh->dev[i];
577 if (test_bit(R5_Wantfill, &dev->flags)) {
579 spin_lock_irq(&conf->device_lock);
580 dev->read = rbi = dev->toread;
582 spin_unlock_irq(&conf->device_lock);
583 while (rbi && rbi->bi_sector <
584 dev->sector + STRIPE_SECTORS) {
585 tx = async_copy_data(0, rbi, dev->page,
587 rbi = r5_next_bio(rbi, dev->sector);
592 atomic_inc(&sh->count);
593 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
594 ops_complete_biofill, sh);
597 static void ops_complete_compute5(void *stripe_head_ref)
599 struct stripe_head *sh = stripe_head_ref;
600 int target = sh->ops.target;
601 struct r5dev *tgt = &sh->dev[target];
603 pr_debug("%s: stripe %llu\n", __FUNCTION__,
604 (unsigned long long)sh->sector);
606 set_bit(R5_UPTODATE, &tgt->flags);
607 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
608 clear_bit(R5_Wantcompute, &tgt->flags);
609 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
610 set_bit(STRIPE_HANDLE, &sh->state);
614 static struct dma_async_tx_descriptor *
615 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
617 /* kernel stack size limits the total number of disks */
618 int disks = sh->disks;
619 struct page *xor_srcs[disks];
620 int target = sh->ops.target;
621 struct r5dev *tgt = &sh->dev[target];
622 struct page *xor_dest = tgt->page;
624 struct dma_async_tx_descriptor *tx;
627 pr_debug("%s: stripe %llu block: %d\n",
628 __FUNCTION__, (unsigned long long)sh->sector, target);
629 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
631 for (i = disks; i--; )
633 xor_srcs[count++] = sh->dev[i].page;
635 atomic_inc(&sh->count);
637 if (unlikely(count == 1))
638 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
639 0, NULL, ops_complete_compute5, sh);
641 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
642 ASYNC_TX_XOR_ZERO_DST, NULL,
643 ops_complete_compute5, sh);
645 /* ack now if postxor is not set to be run */
646 if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
652 static void ops_complete_prexor(void *stripe_head_ref)
654 struct stripe_head *sh = stripe_head_ref;
656 pr_debug("%s: stripe %llu\n", __FUNCTION__,
657 (unsigned long long)sh->sector);
659 set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
662 static struct dma_async_tx_descriptor *
663 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
665 /* kernel stack size limits the total number of disks */
666 int disks = sh->disks;
667 struct page *xor_srcs[disks];
668 int count = 0, pd_idx = sh->pd_idx, i;
670 /* existing parity data subtracted */
671 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
673 pr_debug("%s: stripe %llu\n", __FUNCTION__,
674 (unsigned long long)sh->sector);
676 for (i = disks; i--; ) {
677 struct r5dev *dev = &sh->dev[i];
678 /* Only process blocks that are known to be uptodate */
679 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
680 xor_srcs[count++] = dev->page;
683 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
684 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
685 ops_complete_prexor, sh);
690 static struct dma_async_tx_descriptor *
691 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
693 int disks = sh->disks;
694 int pd_idx = sh->pd_idx, i;
696 /* check if prexor is active which means only process blocks
697 * that are part of a read-modify-write (Wantprexor)
699 int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
701 pr_debug("%s: stripe %llu\n", __FUNCTION__,
702 (unsigned long long)sh->sector);
704 for (i = disks; i--; ) {
705 struct r5dev *dev = &sh->dev[i];
710 if (prexor) { /* rmw */
712 test_bit(R5_Wantprexor, &dev->flags))
715 if (i != pd_idx && dev->towrite &&
716 test_bit(R5_LOCKED, &dev->flags))
723 spin_lock(&sh->lock);
724 chosen = dev->towrite;
726 BUG_ON(dev->written);
727 wbi = dev->written = chosen;
728 spin_unlock(&sh->lock);
730 while (wbi && wbi->bi_sector <
731 dev->sector + STRIPE_SECTORS) {
732 tx = async_copy_data(1, wbi, dev->page,
734 wbi = r5_next_bio(wbi, dev->sector);
742 static void ops_complete_postxor(void *stripe_head_ref)
744 struct stripe_head *sh = stripe_head_ref;
746 pr_debug("%s: stripe %llu\n", __FUNCTION__,
747 (unsigned long long)sh->sector);
749 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
750 set_bit(STRIPE_HANDLE, &sh->state);
754 static void ops_complete_write(void *stripe_head_ref)
756 struct stripe_head *sh = stripe_head_ref;
757 int disks = sh->disks, i, pd_idx = sh->pd_idx;
759 pr_debug("%s: stripe %llu\n", __FUNCTION__,
760 (unsigned long long)sh->sector);
762 for (i = disks; i--; ) {
763 struct r5dev *dev = &sh->dev[i];
764 if (dev->written || i == pd_idx)
765 set_bit(R5_UPTODATE, &dev->flags);
768 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
769 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
771 set_bit(STRIPE_HANDLE, &sh->state);
776 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
778 /* kernel stack size limits the total number of disks */
779 int disks = sh->disks;
780 struct page *xor_srcs[disks];
782 int count = 0, pd_idx = sh->pd_idx, i;
783 struct page *xor_dest;
784 int prexor = test_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
786 dma_async_tx_callback callback;
788 pr_debug("%s: stripe %llu\n", __FUNCTION__,
789 (unsigned long long)sh->sector);
791 /* check if prexor is active which means only process blocks
792 * that are part of a read-modify-write (written)
795 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
796 for (i = disks; i--; ) {
797 struct r5dev *dev = &sh->dev[i];
799 xor_srcs[count++] = dev->page;
802 xor_dest = sh->dev[pd_idx].page;
803 for (i = disks; i--; ) {
804 struct r5dev *dev = &sh->dev[i];
806 xor_srcs[count++] = dev->page;
810 /* check whether this postxor is part of a write */
811 callback = test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending) ?
812 ops_complete_write : ops_complete_postxor;
814 /* 1/ if we prexor'd then the dest is reused as a source
815 * 2/ if we did not prexor then we are redoing the parity
816 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
817 * for the synchronous xor case
819 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
820 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
822 atomic_inc(&sh->count);
824 if (unlikely(count == 1)) {
825 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
826 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
827 flags, tx, callback, sh);
829 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
830 flags, tx, callback, sh);
833 static void ops_complete_check(void *stripe_head_ref)
835 struct stripe_head *sh = stripe_head_ref;
836 int pd_idx = sh->pd_idx;
838 pr_debug("%s: stripe %llu\n", __FUNCTION__,
839 (unsigned long long)sh->sector);
841 if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
842 sh->ops.zero_sum_result == 0)
843 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
845 set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
846 set_bit(STRIPE_HANDLE, &sh->state);
850 static void ops_run_check(struct stripe_head *sh)
852 /* kernel stack size limits the total number of disks */
853 int disks = sh->disks;
854 struct page *xor_srcs[disks];
855 struct dma_async_tx_descriptor *tx;
857 int count = 0, pd_idx = sh->pd_idx, i;
858 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
860 pr_debug("%s: stripe %llu\n", __FUNCTION__,
861 (unsigned long long)sh->sector);
863 for (i = disks; i--; ) {
864 struct r5dev *dev = &sh->dev[i];
866 xor_srcs[count++] = dev->page;
869 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
870 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
873 set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
875 clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
877 atomic_inc(&sh->count);
878 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
879 ops_complete_check, sh);
882 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
884 int overlap_clear = 0, i, disks = sh->disks;
885 struct dma_async_tx_descriptor *tx = NULL;
887 if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
892 if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
893 tx = ops_run_compute5(sh, pending);
895 if (test_bit(STRIPE_OP_PREXOR, &pending))
896 tx = ops_run_prexor(sh, tx);
898 if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
899 tx = ops_run_biodrain(sh, tx);
903 if (test_bit(STRIPE_OP_POSTXOR, &pending))
904 ops_run_postxor(sh, tx);
906 if (test_bit(STRIPE_OP_CHECK, &pending))
909 if (test_bit(STRIPE_OP_IO, &pending))
913 for (i = disks; i--; ) {
914 struct r5dev *dev = &sh->dev[i];
915 if (test_and_clear_bit(R5_Overlap, &dev->flags))
916 wake_up(&sh->raid_conf->wait_for_overlap);
920 static int grow_one_stripe(raid5_conf_t *conf)
922 struct stripe_head *sh;
923 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
926 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
927 sh->raid_conf = conf;
928 spin_lock_init(&sh->lock);
930 if (grow_buffers(sh, conf->raid_disks)) {
931 shrink_buffers(sh, conf->raid_disks);
932 kmem_cache_free(conf->slab_cache, sh);
935 sh->disks = conf->raid_disks;
936 /* we just created an active stripe so... */
937 atomic_set(&sh->count, 1);
938 atomic_inc(&conf->active_stripes);
939 INIT_LIST_HEAD(&sh->lru);
944 static int grow_stripes(raid5_conf_t *conf, int num)
946 struct kmem_cache *sc;
947 int devs = conf->raid_disks;
949 sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
950 sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
951 conf->active_name = 0;
952 sc = kmem_cache_create(conf->cache_name[conf->active_name],
953 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
957 conf->slab_cache = sc;
958 conf->pool_size = devs;
960 if (!grow_one_stripe(conf))
965 #ifdef CONFIG_MD_RAID5_RESHAPE
966 static int resize_stripes(raid5_conf_t *conf, int newsize)
968 /* Make all the stripes able to hold 'newsize' devices.
969 * New slots in each stripe get 'page' set to a new page.
971 * This happens in stages:
972 * 1/ create a new kmem_cache and allocate the required number of
974 * 2/ gather all the old stripe_heads and tranfer the pages across
975 * to the new stripe_heads. This will have the side effect of
976 * freezing the array as once all stripe_heads have been collected,
977 * no IO will be possible. Old stripe heads are freed once their
978 * pages have been transferred over, and the old kmem_cache is
979 * freed when all stripes are done.
980 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
981 * we simple return a failre status - no need to clean anything up.
982 * 4/ allocate new pages for the new slots in the new stripe_heads.
983 * If this fails, we don't bother trying the shrink the
984 * stripe_heads down again, we just leave them as they are.
985 * As each stripe_head is processed the new one is released into
988 * Once step2 is started, we cannot afford to wait for a write,
989 * so we use GFP_NOIO allocations.
991 struct stripe_head *osh, *nsh;
992 LIST_HEAD(newstripes);
993 struct disk_info *ndisks;
995 struct kmem_cache *sc;
998 if (newsize <= conf->pool_size)
999 return 0; /* never bother to shrink */
1001 md_allow_write(conf->mddev);
1004 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1005 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1010 for (i = conf->max_nr_stripes; i; i--) {
1011 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1015 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1017 nsh->raid_conf = conf;
1018 spin_lock_init(&nsh->lock);
1020 list_add(&nsh->lru, &newstripes);
1023 /* didn't get enough, give up */
1024 while (!list_empty(&newstripes)) {
1025 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1026 list_del(&nsh->lru);
1027 kmem_cache_free(sc, nsh);
1029 kmem_cache_destroy(sc);
1032 /* Step 2 - Must use GFP_NOIO now.
1033 * OK, we have enough stripes, start collecting inactive
1034 * stripes and copying them over
1036 list_for_each_entry(nsh, &newstripes, lru) {
1037 spin_lock_irq(&conf->device_lock);
1038 wait_event_lock_irq(conf->wait_for_stripe,
1039 !list_empty(&conf->inactive_list),
1041 unplug_slaves(conf->mddev)
1043 osh = get_free_stripe(conf);
1044 spin_unlock_irq(&conf->device_lock);
1045 atomic_set(&nsh->count, 1);
1046 for(i=0; i<conf->pool_size; i++)
1047 nsh->dev[i].page = osh->dev[i].page;
1048 for( ; i<newsize; i++)
1049 nsh->dev[i].page = NULL;
1050 kmem_cache_free(conf->slab_cache, osh);
1052 kmem_cache_destroy(conf->slab_cache);
1055 * At this point, we are holding all the stripes so the array
1056 * is completely stalled, so now is a good time to resize
1059 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1061 for (i=0; i<conf->raid_disks; i++)
1062 ndisks[i] = conf->disks[i];
1064 conf->disks = ndisks;
1068 /* Step 4, return new stripes to service */
1069 while(!list_empty(&newstripes)) {
1070 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1071 list_del_init(&nsh->lru);
1072 for (i=conf->raid_disks; i < newsize; i++)
1073 if (nsh->dev[i].page == NULL) {
1074 struct page *p = alloc_page(GFP_NOIO);
1075 nsh->dev[i].page = p;
1079 release_stripe(nsh);
1081 /* critical section pass, GFP_NOIO no longer needed */
1083 conf->slab_cache = sc;
1084 conf->active_name = 1-conf->active_name;
1085 conf->pool_size = newsize;
1090 static int drop_one_stripe(raid5_conf_t *conf)
1092 struct stripe_head *sh;
1094 spin_lock_irq(&conf->device_lock);
1095 sh = get_free_stripe(conf);
1096 spin_unlock_irq(&conf->device_lock);
1099 BUG_ON(atomic_read(&sh->count));
1100 shrink_buffers(sh, conf->pool_size);
1101 kmem_cache_free(conf->slab_cache, sh);
1102 atomic_dec(&conf->active_stripes);
1106 static void shrink_stripes(raid5_conf_t *conf)
1108 while (drop_one_stripe(conf))
1111 if (conf->slab_cache)
1112 kmem_cache_destroy(conf->slab_cache);
1113 conf->slab_cache = NULL;
1116 static int raid5_end_read_request(struct bio * bi, unsigned int bytes_done,
1119 struct stripe_head *sh = bi->bi_private;
1120 raid5_conf_t *conf = sh->raid_conf;
1121 int disks = sh->disks, i;
1122 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1123 char b[BDEVNAME_SIZE];
1129 for (i=0 ; i<disks; i++)
1130 if (bi == &sh->dev[i].req)
1133 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1134 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1142 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1143 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1144 rdev = conf->disks[i].rdev;
1145 printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1146 mdname(conf->mddev), STRIPE_SECTORS,
1147 (unsigned long long)sh->sector + rdev->data_offset,
1148 bdevname(rdev->bdev, b));
1149 clear_bit(R5_ReadError, &sh->dev[i].flags);
1150 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1152 if (atomic_read(&conf->disks[i].rdev->read_errors))
1153 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1155 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1157 rdev = conf->disks[i].rdev;
1159 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1160 atomic_inc(&rdev->read_errors);
1161 if (conf->mddev->degraded)
1162 printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1163 mdname(conf->mddev),
1164 (unsigned long long)sh->sector + rdev->data_offset,
1166 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1168 printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1169 mdname(conf->mddev),
1170 (unsigned long long)sh->sector + rdev->data_offset,
1172 else if (atomic_read(&rdev->read_errors)
1173 > conf->max_nr_stripes)
1175 "raid5:%s: Too many read errors, failing device %s.\n",
1176 mdname(conf->mddev), bdn);
1180 set_bit(R5_ReadError, &sh->dev[i].flags);
1182 clear_bit(R5_ReadError, &sh->dev[i].flags);
1183 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1184 md_error(conf->mddev, rdev);
1187 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1188 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1189 set_bit(STRIPE_HANDLE, &sh->state);
1194 static int raid5_end_write_request (struct bio *bi, unsigned int bytes_done,
1197 struct stripe_head *sh = bi->bi_private;
1198 raid5_conf_t *conf = sh->raid_conf;
1199 int disks = sh->disks, i;
1200 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1205 for (i=0 ; i<disks; i++)
1206 if (bi == &sh->dev[i].req)
1209 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1210 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1218 md_error(conf->mddev, conf->disks[i].rdev);
1220 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1222 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1223 set_bit(STRIPE_HANDLE, &sh->state);
1229 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1231 static void raid5_build_block (struct stripe_head *sh, int i)
1233 struct r5dev *dev = &sh->dev[i];
1235 bio_init(&dev->req);
1236 dev->req.bi_io_vec = &dev->vec;
1238 dev->req.bi_max_vecs++;
1239 dev->vec.bv_page = dev->page;
1240 dev->vec.bv_len = STRIPE_SIZE;
1241 dev->vec.bv_offset = 0;
1243 dev->req.bi_sector = sh->sector;
1244 dev->req.bi_private = sh;
1247 dev->sector = compute_blocknr(sh, i);
1250 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1252 char b[BDEVNAME_SIZE];
1253 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1254 pr_debug("raid5: error called\n");
1256 if (!test_bit(Faulty, &rdev->flags)) {
1257 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1258 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1259 unsigned long flags;
1260 spin_lock_irqsave(&conf->device_lock, flags);
1262 spin_unlock_irqrestore(&conf->device_lock, flags);
1264 * if recovery was running, make sure it aborts.
1266 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1268 set_bit(Faulty, &rdev->flags);
1270 "raid5: Disk failure on %s, disabling device."
1271 " Operation continuing on %d devices\n",
1272 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1277 * Input: a 'big' sector number,
1278 * Output: index of the data and parity disk, and the sector # in them.
1280 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1281 unsigned int data_disks, unsigned int * dd_idx,
1282 unsigned int * pd_idx, raid5_conf_t *conf)
1285 unsigned long chunk_number;
1286 unsigned int chunk_offset;
1287 sector_t new_sector;
1288 int sectors_per_chunk = conf->chunk_size >> 9;
1290 /* First compute the information on this sector */
1293 * Compute the chunk number and the sector offset inside the chunk
1295 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1296 chunk_number = r_sector;
1297 BUG_ON(r_sector != chunk_number);
1300 * Compute the stripe number
1302 stripe = chunk_number / data_disks;
1305 * Compute the data disk and parity disk indexes inside the stripe
1307 *dd_idx = chunk_number % data_disks;
1310 * Select the parity disk based on the user selected algorithm.
1312 switch(conf->level) {
1314 *pd_idx = data_disks;
1317 switch (conf->algorithm) {
1318 case ALGORITHM_LEFT_ASYMMETRIC:
1319 *pd_idx = data_disks - stripe % raid_disks;
1320 if (*dd_idx >= *pd_idx)
1323 case ALGORITHM_RIGHT_ASYMMETRIC:
1324 *pd_idx = stripe % raid_disks;
1325 if (*dd_idx >= *pd_idx)
1328 case ALGORITHM_LEFT_SYMMETRIC:
1329 *pd_idx = data_disks - stripe % raid_disks;
1330 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1332 case ALGORITHM_RIGHT_SYMMETRIC:
1333 *pd_idx = stripe % raid_disks;
1334 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1337 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1343 /**** FIX THIS ****/
1344 switch (conf->algorithm) {
1345 case ALGORITHM_LEFT_ASYMMETRIC:
1346 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1347 if (*pd_idx == raid_disks-1)
1348 (*dd_idx)++; /* Q D D D P */
1349 else if (*dd_idx >= *pd_idx)
1350 (*dd_idx) += 2; /* D D P Q D */
1352 case ALGORITHM_RIGHT_ASYMMETRIC:
1353 *pd_idx = stripe % raid_disks;
1354 if (*pd_idx == raid_disks-1)
1355 (*dd_idx)++; /* Q D D D P */
1356 else if (*dd_idx >= *pd_idx)
1357 (*dd_idx) += 2; /* D D P Q D */
1359 case ALGORITHM_LEFT_SYMMETRIC:
1360 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1361 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1363 case ALGORITHM_RIGHT_SYMMETRIC:
1364 *pd_idx = stripe % raid_disks;
1365 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1368 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1375 * Finally, compute the new sector number
1377 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1382 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1384 raid5_conf_t *conf = sh->raid_conf;
1385 int raid_disks = sh->disks;
1386 int data_disks = raid_disks - conf->max_degraded;
1387 sector_t new_sector = sh->sector, check;
1388 int sectors_per_chunk = conf->chunk_size >> 9;
1391 int chunk_number, dummy1, dummy2, dd_idx = i;
1395 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1396 stripe = new_sector;
1397 BUG_ON(new_sector != stripe);
1399 if (i == sh->pd_idx)
1401 switch(conf->level) {
1404 switch (conf->algorithm) {
1405 case ALGORITHM_LEFT_ASYMMETRIC:
1406 case ALGORITHM_RIGHT_ASYMMETRIC:
1410 case ALGORITHM_LEFT_SYMMETRIC:
1411 case ALGORITHM_RIGHT_SYMMETRIC:
1414 i -= (sh->pd_idx + 1);
1417 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1422 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1423 return 0; /* It is the Q disk */
1424 switch (conf->algorithm) {
1425 case ALGORITHM_LEFT_ASYMMETRIC:
1426 case ALGORITHM_RIGHT_ASYMMETRIC:
1427 if (sh->pd_idx == raid_disks-1)
1428 i--; /* Q D D D P */
1429 else if (i > sh->pd_idx)
1430 i -= 2; /* D D P Q D */
1432 case ALGORITHM_LEFT_SYMMETRIC:
1433 case ALGORITHM_RIGHT_SYMMETRIC:
1434 if (sh->pd_idx == raid_disks-1)
1435 i--; /* Q D D D P */
1440 i -= (sh->pd_idx + 2);
1444 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1450 chunk_number = stripe * data_disks + i;
1451 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1453 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1454 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1455 printk(KERN_ERR "compute_blocknr: map not correct\n");
1464 * Copy data between a page in the stripe cache, and one or more bion
1465 * The page could align with the middle of the bio, or there could be
1466 * several bion, each with several bio_vecs, which cover part of the page
1467 * Multiple bion are linked together on bi_next. There may be extras
1468 * at the end of this list. We ignore them.
1470 static void copy_data(int frombio, struct bio *bio,
1474 char *pa = page_address(page);
1475 struct bio_vec *bvl;
1479 if (bio->bi_sector >= sector)
1480 page_offset = (signed)(bio->bi_sector - sector) * 512;
1482 page_offset = (signed)(sector - bio->bi_sector) * -512;
1483 bio_for_each_segment(bvl, bio, i) {
1484 int len = bio_iovec_idx(bio,i)->bv_len;
1488 if (page_offset < 0) {
1489 b_offset = -page_offset;
1490 page_offset += b_offset;
1494 if (len > 0 && page_offset + len > STRIPE_SIZE)
1495 clen = STRIPE_SIZE - page_offset;
1499 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1501 memcpy(pa+page_offset, ba+b_offset, clen);
1503 memcpy(ba+b_offset, pa+page_offset, clen);
1504 __bio_kunmap_atomic(ba, KM_USER0);
1506 if (clen < len) /* hit end of page */
1512 #define check_xor() do { \
1513 if (count == MAX_XOR_BLOCKS) { \
1514 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1520 static void compute_block(struct stripe_head *sh, int dd_idx)
1522 int i, count, disks = sh->disks;
1523 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1525 pr_debug("compute_block, stripe %llu, idx %d\n",
1526 (unsigned long long)sh->sector, dd_idx);
1528 dest = page_address(sh->dev[dd_idx].page);
1529 memset(dest, 0, STRIPE_SIZE);
1531 for (i = disks ; i--; ) {
1534 p = page_address(sh->dev[i].page);
1535 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1538 printk(KERN_ERR "compute_block() %d, stripe %llu, %d"
1539 " not present\n", dd_idx,
1540 (unsigned long long)sh->sector, i);
1545 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1546 set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1549 static void compute_parity5(struct stripe_head *sh, int method)
1551 raid5_conf_t *conf = sh->raid_conf;
1552 int i, pd_idx = sh->pd_idx, disks = sh->disks, count;
1553 void *ptr[MAX_XOR_BLOCKS], *dest;
1556 pr_debug("compute_parity5, stripe %llu, method %d\n",
1557 (unsigned long long)sh->sector, method);
1560 dest = page_address(sh->dev[pd_idx].page);
1562 case READ_MODIFY_WRITE:
1563 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags));
1564 for (i=disks ; i-- ;) {
1567 if (sh->dev[i].towrite &&
1568 test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1569 ptr[count++] = page_address(sh->dev[i].page);
1570 chosen = sh->dev[i].towrite;
1571 sh->dev[i].towrite = NULL;
1573 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1574 wake_up(&conf->wait_for_overlap);
1576 BUG_ON(sh->dev[i].written);
1577 sh->dev[i].written = chosen;
1582 case RECONSTRUCT_WRITE:
1583 memset(dest, 0, STRIPE_SIZE);
1584 for (i= disks; i-- ;)
1585 if (i!=pd_idx && sh->dev[i].towrite) {
1586 chosen = sh->dev[i].towrite;
1587 sh->dev[i].towrite = NULL;
1589 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1590 wake_up(&conf->wait_for_overlap);
1592 BUG_ON(sh->dev[i].written);
1593 sh->dev[i].written = chosen;
1600 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1604 for (i = disks; i--;)
1605 if (sh->dev[i].written) {
1606 sector_t sector = sh->dev[i].sector;
1607 struct bio *wbi = sh->dev[i].written;
1608 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1609 copy_data(1, wbi, sh->dev[i].page, sector);
1610 wbi = r5_next_bio(wbi, sector);
1613 set_bit(R5_LOCKED, &sh->dev[i].flags);
1614 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1618 case RECONSTRUCT_WRITE:
1622 ptr[count++] = page_address(sh->dev[i].page);
1626 case READ_MODIFY_WRITE:
1627 for (i = disks; i--;)
1628 if (sh->dev[i].written) {
1629 ptr[count++] = page_address(sh->dev[i].page);
1634 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1636 if (method != CHECK_PARITY) {
1637 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1638 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1640 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1643 static void compute_parity6(struct stripe_head *sh, int method)
1645 raid6_conf_t *conf = sh->raid_conf;
1646 int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1648 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1651 qd_idx = raid6_next_disk(pd_idx, disks);
1652 d0_idx = raid6_next_disk(qd_idx, disks);
1654 pr_debug("compute_parity, stripe %llu, method %d\n",
1655 (unsigned long long)sh->sector, method);
1658 case READ_MODIFY_WRITE:
1659 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1660 case RECONSTRUCT_WRITE:
1661 for (i= disks; i-- ;)
1662 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1663 chosen = sh->dev[i].towrite;
1664 sh->dev[i].towrite = NULL;
1666 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1667 wake_up(&conf->wait_for_overlap);
1669 BUG_ON(sh->dev[i].written);
1670 sh->dev[i].written = chosen;
1674 BUG(); /* Not implemented yet */
1677 for (i = disks; i--;)
1678 if (sh->dev[i].written) {
1679 sector_t sector = sh->dev[i].sector;
1680 struct bio *wbi = sh->dev[i].written;
1681 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1682 copy_data(1, wbi, sh->dev[i].page, sector);
1683 wbi = r5_next_bio(wbi, sector);
1686 set_bit(R5_LOCKED, &sh->dev[i].flags);
1687 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1691 // case RECONSTRUCT_WRITE:
1692 // case CHECK_PARITY:
1693 // case UPDATE_PARITY:
1694 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1695 /* FIX: Is this ordering of drives even remotely optimal? */
1699 ptrs[count++] = page_address(sh->dev[i].page);
1700 if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1701 printk("block %d/%d not uptodate on parity calc\n", i,count);
1702 i = raid6_next_disk(i, disks);
1703 } while ( i != d0_idx );
1707 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1710 case RECONSTRUCT_WRITE:
1711 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1712 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1713 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1714 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1717 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1718 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1724 /* Compute one missing block */
1725 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1727 int i, count, disks = sh->disks;
1728 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1729 int pd_idx = sh->pd_idx;
1730 int qd_idx = raid6_next_disk(pd_idx, disks);
1732 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1733 (unsigned long long)sh->sector, dd_idx);
1735 if ( dd_idx == qd_idx ) {
1736 /* We're actually computing the Q drive */
1737 compute_parity6(sh, UPDATE_PARITY);
1739 dest = page_address(sh->dev[dd_idx].page);
1740 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1742 for (i = disks ; i--; ) {
1743 if (i == dd_idx || i == qd_idx)
1745 p = page_address(sh->dev[i].page);
1746 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1749 printk("compute_block() %d, stripe %llu, %d"
1750 " not present\n", dd_idx,
1751 (unsigned long long)sh->sector, i);
1756 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1757 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1758 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1762 /* Compute two missing blocks */
1763 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1765 int i, count, disks = sh->disks;
1766 int pd_idx = sh->pd_idx;
1767 int qd_idx = raid6_next_disk(pd_idx, disks);
1768 int d0_idx = raid6_next_disk(qd_idx, disks);
1771 /* faila and failb are disk numbers relative to d0_idx */
1772 /* pd_idx become disks-2 and qd_idx become disks-1 */
1773 faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1774 failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1776 BUG_ON(faila == failb);
1777 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1779 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1780 (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1782 if ( failb == disks-1 ) {
1783 /* Q disk is one of the missing disks */
1784 if ( faila == disks-2 ) {
1785 /* Missing P+Q, just recompute */
1786 compute_parity6(sh, UPDATE_PARITY);
1789 /* We're missing D+Q; recompute D from P */
1790 compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1791 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1796 /* We're missing D+P or D+D; build pointer table */
1798 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1804 ptrs[count++] = page_address(sh->dev[i].page);
1805 i = raid6_next_disk(i, disks);
1806 if (i != dd_idx1 && i != dd_idx2 &&
1807 !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1808 printk("compute_2 with missing block %d/%d\n", count, i);
1809 } while ( i != d0_idx );
1811 if ( failb == disks-2 ) {
1812 /* We're missing D+P. */
1813 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1815 /* We're missing D+D. */
1816 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1819 /* Both the above update both missing blocks */
1820 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1821 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1826 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1828 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1832 /* if we are not expanding this is a proper write request, and
1833 * there will be bios with new data to be drained into the
1837 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1841 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1844 for (i = disks; i--; ) {
1845 struct r5dev *dev = &sh->dev[i];
1848 set_bit(R5_LOCKED, &dev->flags);
1850 clear_bit(R5_UPTODATE, &dev->flags);
1855 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1856 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1858 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1859 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1860 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1864 for (i = disks; i--; ) {
1865 struct r5dev *dev = &sh->dev[i];
1869 /* For a read-modify write there may be blocks that are
1870 * locked for reading while others are ready to be
1871 * written so we distinguish these blocks by the
1875 (test_bit(R5_UPTODATE, &dev->flags) ||
1876 test_bit(R5_Wantcompute, &dev->flags))) {
1877 set_bit(R5_Wantprexor, &dev->flags);
1878 set_bit(R5_LOCKED, &dev->flags);
1879 clear_bit(R5_UPTODATE, &dev->flags);
1885 /* keep the parity disk locked while asynchronous operations
1888 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1889 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1892 pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1893 __FUNCTION__, (unsigned long long)sh->sector,
1894 locked, sh->ops.pending);
1900 * Each stripe/dev can have one or more bion attached.
1901 * toread/towrite point to the first in a chain.
1902 * The bi_next chain must be in order.
1904 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1907 raid5_conf_t *conf = sh->raid_conf;
1910 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1911 (unsigned long long)bi->bi_sector,
1912 (unsigned long long)sh->sector);
1915 spin_lock(&sh->lock);
1916 spin_lock_irq(&conf->device_lock);
1918 bip = &sh->dev[dd_idx].towrite;
1919 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1922 bip = &sh->dev[dd_idx].toread;
1923 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1924 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1926 bip = & (*bip)->bi_next;
1928 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1931 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1935 bi->bi_phys_segments ++;
1936 spin_unlock_irq(&conf->device_lock);
1937 spin_unlock(&sh->lock);
1939 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1940 (unsigned long long)bi->bi_sector,
1941 (unsigned long long)sh->sector, dd_idx);
1943 if (conf->mddev->bitmap && firstwrite) {
1944 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1946 sh->bm_seq = conf->seq_flush+1;
1947 set_bit(STRIPE_BIT_DELAY, &sh->state);
1951 /* check if page is covered */
1952 sector_t sector = sh->dev[dd_idx].sector;
1953 for (bi=sh->dev[dd_idx].towrite;
1954 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1955 bi && bi->bi_sector <= sector;
1956 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1957 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1958 sector = bi->bi_sector + (bi->bi_size>>9);
1960 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1961 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1966 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1967 spin_unlock_irq(&conf->device_lock);
1968 spin_unlock(&sh->lock);
1972 static void end_reshape(raid5_conf_t *conf);
1974 static int page_is_zero(struct page *p)
1976 char *a = page_address(p);
1977 return ((*(u32*)a) == 0 &&
1978 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1981 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1983 int sectors_per_chunk = conf->chunk_size >> 9;
1985 int chunk_offset = sector_div(stripe, sectors_per_chunk);
1987 raid5_compute_sector(stripe * (disks - conf->max_degraded)
1988 *sectors_per_chunk + chunk_offset,
1989 disks, disks - conf->max_degraded,
1990 &dd_idx, &pd_idx, conf);
1995 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1996 struct stripe_head_state *s, int disks,
1997 struct bio **return_bi)
2000 for (i = disks; i--; ) {
2004 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2007 rdev = rcu_dereference(conf->disks[i].rdev);
2008 if (rdev && test_bit(In_sync, &rdev->flags))
2009 /* multiple read failures in one stripe */
2010 md_error(conf->mddev, rdev);
2013 spin_lock_irq(&conf->device_lock);
2014 /* fail all writes first */
2015 bi = sh->dev[i].towrite;
2016 sh->dev[i].towrite = NULL;
2022 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2023 wake_up(&conf->wait_for_overlap);
2025 while (bi && bi->bi_sector <
2026 sh->dev[i].sector + STRIPE_SECTORS) {
2027 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2028 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2029 if (--bi->bi_phys_segments == 0) {
2030 md_write_end(conf->mddev);
2031 bi->bi_next = *return_bi;
2036 /* and fail all 'written' */
2037 bi = sh->dev[i].written;
2038 sh->dev[i].written = NULL;
2039 if (bi) bitmap_end = 1;
2040 while (bi && bi->bi_sector <
2041 sh->dev[i].sector + STRIPE_SECTORS) {
2042 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2043 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2044 if (--bi->bi_phys_segments == 0) {
2045 md_write_end(conf->mddev);
2046 bi->bi_next = *return_bi;
2052 /* fail any reads if this device is non-operational and
2053 * the data has not reached the cache yet.
2055 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2056 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2057 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2058 bi = sh->dev[i].toread;
2059 sh->dev[i].toread = NULL;
2060 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2061 wake_up(&conf->wait_for_overlap);
2062 if (bi) s->to_read--;
2063 while (bi && bi->bi_sector <
2064 sh->dev[i].sector + STRIPE_SECTORS) {
2065 struct bio *nextbi =
2066 r5_next_bio(bi, sh->dev[i].sector);
2067 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2068 if (--bi->bi_phys_segments == 0) {
2069 bi->bi_next = *return_bi;
2075 spin_unlock_irq(&conf->device_lock);
2077 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2078 STRIPE_SECTORS, 0, 0);
2083 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
2086 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
2087 struct stripe_head_state *s, int disk_idx, int disks)
2089 struct r5dev *dev = &sh->dev[disk_idx];
2090 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2092 /* don't schedule compute operations or reads on the parity block while
2093 * a check is in flight
2095 if ((disk_idx == sh->pd_idx) &&
2096 test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2099 /* is the data in this block needed, and can we get it? */
2100 if (!test_bit(R5_LOCKED, &dev->flags) &&
2101 !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
2102 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2103 s->syncing || s->expanding || (s->failed &&
2104 (failed_dev->toread || (failed_dev->towrite &&
2105 !test_bit(R5_OVERWRITE, &failed_dev->flags)
2107 /* 1/ We would like to get this block, possibly by computing it,
2108 * but we might not be able to.
2110 * 2/ Since parity check operations potentially make the parity
2111 * block !uptodate it will need to be refreshed before any
2112 * compute operations on data disks are scheduled.
2114 * 3/ We hold off parity block re-reads until check operations
2117 if ((s->uptodate == disks - 1) &&
2118 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2119 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2120 set_bit(R5_Wantcompute, &dev->flags);
2121 sh->ops.target = disk_idx;
2124 /* Careful: from this point on 'uptodate' is in the eye
2125 * of raid5_run_ops which services 'compute' operations
2126 * before writes. R5_Wantcompute flags a block that will
2127 * be R5_UPTODATE by the time it is needed for a
2128 * subsequent operation.
2131 return 0; /* uptodate + compute == disks */
2132 } else if ((s->uptodate < disks - 1) &&
2133 test_bit(R5_Insync, &dev->flags)) {
2134 /* Note: we hold off compute operations while checks are
2135 * in flight, but we still prefer 'compute' over 'read'
2136 * hence we only read if (uptodate < * disks-1)
2138 set_bit(R5_LOCKED, &dev->flags);
2139 set_bit(R5_Wantread, &dev->flags);
2140 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2143 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2151 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2152 struct stripe_head_state *s, int disks)
2156 /* Clear completed compute operations. Parity recovery
2157 * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2158 * later on in this routine
2160 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2161 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2162 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2163 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2164 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2167 /* look for blocks to read/compute, skip this if a compute
2168 * is already in flight, or if the stripe contents are in the
2169 * midst of changing due to a write
2171 if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2172 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2173 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2174 for (i = disks; i--; )
2175 if (__handle_issuing_new_read_requests5(
2176 sh, s, i, disks) == 0)
2179 set_bit(STRIPE_HANDLE, &sh->state);
2182 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2183 struct stripe_head_state *s, struct r6_state *r6s,
2187 for (i = disks; i--; ) {
2188 struct r5dev *dev = &sh->dev[i];
2189 if (!test_bit(R5_LOCKED, &dev->flags) &&
2190 !test_bit(R5_UPTODATE, &dev->flags) &&
2191 (dev->toread || (dev->towrite &&
2192 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2193 s->syncing || s->expanding ||
2195 (sh->dev[r6s->failed_num[0]].toread ||
2198 (sh->dev[r6s->failed_num[1]].toread ||
2200 /* we would like to get this block, possibly
2201 * by computing it, but we might not be able to
2203 if (s->uptodate == disks-1) {
2204 pr_debug("Computing stripe %llu block %d\n",
2205 (unsigned long long)sh->sector, i);
2206 compute_block_1(sh, i, 0);
2208 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2209 /* Computing 2-failure is *very* expensive; only
2210 * do it if failed >= 2
2213 for (other = disks; other--; ) {
2216 if (!test_bit(R5_UPTODATE,
2217 &sh->dev[other].flags))
2221 pr_debug("Computing stripe %llu blocks %d,%d\n",
2222 (unsigned long long)sh->sector,
2224 compute_block_2(sh, i, other);
2226 } else if (test_bit(R5_Insync, &dev->flags)) {
2227 set_bit(R5_LOCKED, &dev->flags);
2228 set_bit(R5_Wantread, &dev->flags);
2230 pr_debug("Reading block %d (sync=%d)\n",
2235 set_bit(STRIPE_HANDLE, &sh->state);
2239 /* handle_completed_write_requests
2240 * any written block on an uptodate or failed drive can be returned.
2241 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2242 * never LOCKED, so we don't need to test 'failed' directly.
2244 static void handle_completed_write_requests(raid5_conf_t *conf,
2245 struct stripe_head *sh, int disks, struct bio **return_bi)
2250 for (i = disks; i--; )
2251 if (sh->dev[i].written) {
2253 if (!test_bit(R5_LOCKED, &dev->flags) &&
2254 test_bit(R5_UPTODATE, &dev->flags)) {
2255 /* We can return any write requests */
2256 struct bio *wbi, *wbi2;
2258 pr_debug("Return write for disc %d\n", i);
2259 spin_lock_irq(&conf->device_lock);
2261 dev->written = NULL;
2262 while (wbi && wbi->bi_sector <
2263 dev->sector + STRIPE_SECTORS) {
2264 wbi2 = r5_next_bio(wbi, dev->sector);
2265 if (--wbi->bi_phys_segments == 0) {
2266 md_write_end(conf->mddev);
2267 wbi->bi_next = *return_bi;
2272 if (dev->towrite == NULL)
2274 spin_unlock_irq(&conf->device_lock);
2276 bitmap_endwrite(conf->mddev->bitmap,
2279 !test_bit(STRIPE_DEGRADED, &sh->state),
2285 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2286 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2288 int rmw = 0, rcw = 0, i;
2289 for (i = disks; i--; ) {
2290 /* would I have to read this buffer for read_modify_write */
2291 struct r5dev *dev = &sh->dev[i];
2292 if ((dev->towrite || i == sh->pd_idx) &&
2293 !test_bit(R5_LOCKED, &dev->flags) &&
2294 !(test_bit(R5_UPTODATE, &dev->flags) ||
2295 test_bit(R5_Wantcompute, &dev->flags))) {
2296 if (test_bit(R5_Insync, &dev->flags))
2299 rmw += 2*disks; /* cannot read it */
2301 /* Would I have to read this buffer for reconstruct_write */
2302 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2303 !test_bit(R5_LOCKED, &dev->flags) &&
2304 !(test_bit(R5_UPTODATE, &dev->flags) ||
2305 test_bit(R5_Wantcompute, &dev->flags))) {
2306 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2311 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2312 (unsigned long long)sh->sector, rmw, rcw);
2313 set_bit(STRIPE_HANDLE, &sh->state);
2314 if (rmw < rcw && rmw > 0)
2315 /* prefer read-modify-write, but need to get some data */
2316 for (i = disks; i--; ) {
2317 struct r5dev *dev = &sh->dev[i];
2318 if ((dev->towrite || i == sh->pd_idx) &&
2319 !test_bit(R5_LOCKED, &dev->flags) &&
2320 !(test_bit(R5_UPTODATE, &dev->flags) ||
2321 test_bit(R5_Wantcompute, &dev->flags)) &&
2322 test_bit(R5_Insync, &dev->flags)) {
2324 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2325 pr_debug("Read_old block "
2326 "%d for r-m-w\n", i);
2327 set_bit(R5_LOCKED, &dev->flags);
2328 set_bit(R5_Wantread, &dev->flags);
2331 set_bit(STRIPE_DELAYED, &sh->state);
2332 set_bit(STRIPE_HANDLE, &sh->state);
2336 if (rcw <= rmw && rcw > 0)
2337 /* want reconstruct write, but need to get some data */
2338 for (i = disks; i--; ) {
2339 struct r5dev *dev = &sh->dev[i];
2340 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2342 !test_bit(R5_LOCKED, &dev->flags) &&
2343 !(test_bit(R5_UPTODATE, &dev->flags) ||
2344 test_bit(R5_Wantcompute, &dev->flags)) &&
2345 test_bit(R5_Insync, &dev->flags)) {
2347 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2348 pr_debug("Read_old block "
2349 "%d for Reconstruct\n", i);
2350 set_bit(R5_LOCKED, &dev->flags);
2351 set_bit(R5_Wantread, &dev->flags);
2354 set_bit(STRIPE_DELAYED, &sh->state);
2355 set_bit(STRIPE_HANDLE, &sh->state);
2359 /* now if nothing is locked, and if we have enough data,
2360 * we can start a write request
2362 /* since handle_stripe can be called at any time we need to handle the
2363 * case where a compute block operation has been submitted and then a
2364 * subsequent call wants to start a write request. raid5_run_ops only
2365 * handles the case where compute block and postxor are requested
2366 * simultaneously. If this is not the case then new writes need to be
2367 * held off until the compute completes.
2369 if ((s->req_compute ||
2370 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2371 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2372 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2373 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2376 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2377 struct stripe_head *sh, struct stripe_head_state *s,
2378 struct r6_state *r6s, int disks)
2380 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2381 int qd_idx = r6s->qd_idx;
2382 for (i = disks; i--; ) {
2383 struct r5dev *dev = &sh->dev[i];
2384 /* Would I have to read this buffer for reconstruct_write */
2385 if (!test_bit(R5_OVERWRITE, &dev->flags)
2386 && i != pd_idx && i != qd_idx
2387 && (!test_bit(R5_LOCKED, &dev->flags)
2389 !test_bit(R5_UPTODATE, &dev->flags)) {
2390 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2392 pr_debug("raid6: must_compute: "
2393 "disk %d flags=%#lx\n", i, dev->flags);
2398 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2399 (unsigned long long)sh->sector, rcw, must_compute);
2400 set_bit(STRIPE_HANDLE, &sh->state);
2403 /* want reconstruct write, but need to get some data */
2404 for (i = disks; i--; ) {
2405 struct r5dev *dev = &sh->dev[i];
2406 if (!test_bit(R5_OVERWRITE, &dev->flags)
2407 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2408 && !test_bit(R5_LOCKED, &dev->flags) &&
2409 !test_bit(R5_UPTODATE, &dev->flags) &&
2410 test_bit(R5_Insync, &dev->flags)) {
2412 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2413 pr_debug("Read_old stripe %llu "
2414 "block %d for Reconstruct\n",
2415 (unsigned long long)sh->sector, i);
2416 set_bit(R5_LOCKED, &dev->flags);
2417 set_bit(R5_Wantread, &dev->flags);
2420 pr_debug("Request delayed stripe %llu "
2421 "block %d for Reconstruct\n",
2422 (unsigned long long)sh->sector, i);
2423 set_bit(STRIPE_DELAYED, &sh->state);
2424 set_bit(STRIPE_HANDLE, &sh->state);
2428 /* now if nothing is locked, and if we have enough data, we can start a
2431 if (s->locked == 0 && rcw == 0 &&
2432 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2433 if (must_compute > 0) {
2434 /* We have failed blocks and need to compute them */
2435 switch (s->failed) {
2439 compute_block_1(sh, r6s->failed_num[0], 0);
2442 compute_block_2(sh, r6s->failed_num[0],
2443 r6s->failed_num[1]);
2445 default: /* This request should have been failed? */
2450 pr_debug("Computing parity for stripe %llu\n",
2451 (unsigned long long)sh->sector);
2452 compute_parity6(sh, RECONSTRUCT_WRITE);
2453 /* now every locked buffer is ready to be written */
2454 for (i = disks; i--; )
2455 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2456 pr_debug("Writing stripe %llu block %d\n",
2457 (unsigned long long)sh->sector, i);
2459 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2461 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2462 set_bit(STRIPE_INSYNC, &sh->state);
2464 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2465 atomic_dec(&conf->preread_active_stripes);
2466 if (atomic_read(&conf->preread_active_stripes) <
2468 md_wakeup_thread(conf->mddev->thread);
2473 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2474 struct stripe_head_state *s, int disks)
2476 set_bit(STRIPE_HANDLE, &sh->state);
2477 /* Take one of the following actions:
2478 * 1/ start a check parity operation if (uptodate == disks)
2479 * 2/ finish a check parity operation and act on the result
2480 * 3/ skip to the writeback section if we previously
2481 * initiated a recovery operation
2483 if (s->failed == 0 &&
2484 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2485 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2486 BUG_ON(s->uptodate != disks);
2487 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2491 test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2492 clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2493 clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2495 if (sh->ops.zero_sum_result == 0)
2496 /* parity is correct (on disc,
2497 * not in buffer any more)
2499 set_bit(STRIPE_INSYNC, &sh->state);
2501 conf->mddev->resync_mismatches +=
2504 MD_RECOVERY_CHECK, &conf->mddev->recovery))
2505 /* don't try to repair!! */
2506 set_bit(STRIPE_INSYNC, &sh->state);
2508 set_bit(STRIPE_OP_COMPUTE_BLK,
2510 set_bit(STRIPE_OP_MOD_REPAIR_PD,
2512 set_bit(R5_Wantcompute,
2513 &sh->dev[sh->pd_idx].flags);
2514 sh->ops.target = sh->pd_idx;
2522 /* check if we can clear a parity disk reconstruct */
2523 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2524 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2526 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2527 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2528 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2529 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2532 /* Wait for check parity and compute block operations to complete
2535 if (!test_bit(STRIPE_INSYNC, &sh->state) &&
2536 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2537 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2539 /* either failed parity check, or recovery is happening */
2541 s->failed_num = sh->pd_idx;
2542 dev = &sh->dev[s->failed_num];
2543 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2544 BUG_ON(s->uptodate != disks);
2546 set_bit(R5_LOCKED, &dev->flags);
2547 set_bit(R5_Wantwrite, &dev->flags);
2548 clear_bit(STRIPE_DEGRADED, &sh->state);
2550 set_bit(STRIPE_INSYNC, &sh->state);
2555 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2556 struct stripe_head_state *s,
2557 struct r6_state *r6s, struct page *tmp_page,
2560 int update_p = 0, update_q = 0;
2562 int pd_idx = sh->pd_idx;
2563 int qd_idx = r6s->qd_idx;
2565 set_bit(STRIPE_HANDLE, &sh->state);
2567 BUG_ON(s->failed > 2);
2568 BUG_ON(s->uptodate < disks);
2569 /* Want to check and possibly repair P and Q.
2570 * However there could be one 'failed' device, in which
2571 * case we can only check one of them, possibly using the
2572 * other to generate missing data
2575 /* If !tmp_page, we cannot do the calculations,
2576 * but as we have set STRIPE_HANDLE, we will soon be called
2577 * by stripe_handle with a tmp_page - just wait until then.
2580 if (s->failed == r6s->q_failed) {
2581 /* The only possible failed device holds 'Q', so it
2582 * makes sense to check P (If anything else were failed,
2583 * we would have used P to recreate it).
2585 compute_block_1(sh, pd_idx, 1);
2586 if (!page_is_zero(sh->dev[pd_idx].page)) {
2587 compute_block_1(sh, pd_idx, 0);
2591 if (!r6s->q_failed && s->failed < 2) {
2592 /* q is not failed, and we didn't use it to generate
2593 * anything, so it makes sense to check it
2595 memcpy(page_address(tmp_page),
2596 page_address(sh->dev[qd_idx].page),
2598 compute_parity6(sh, UPDATE_PARITY);
2599 if (memcmp(page_address(tmp_page),
2600 page_address(sh->dev[qd_idx].page),
2601 STRIPE_SIZE) != 0) {
2602 clear_bit(STRIPE_INSYNC, &sh->state);
2606 if (update_p || update_q) {
2607 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2608 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2609 /* don't try to repair!! */
2610 update_p = update_q = 0;
2613 /* now write out any block on a failed drive,
2614 * or P or Q if they need it
2617 if (s->failed == 2) {
2618 dev = &sh->dev[r6s->failed_num[1]];
2620 set_bit(R5_LOCKED, &dev->flags);
2621 set_bit(R5_Wantwrite, &dev->flags);
2623 if (s->failed >= 1) {
2624 dev = &sh->dev[r6s->failed_num[0]];
2626 set_bit(R5_LOCKED, &dev->flags);
2627 set_bit(R5_Wantwrite, &dev->flags);
2631 dev = &sh->dev[pd_idx];
2633 set_bit(R5_LOCKED, &dev->flags);
2634 set_bit(R5_Wantwrite, &dev->flags);
2637 dev = &sh->dev[qd_idx];
2639 set_bit(R5_LOCKED, &dev->flags);
2640 set_bit(R5_Wantwrite, &dev->flags);
2642 clear_bit(STRIPE_DEGRADED, &sh->state);
2644 set_bit(STRIPE_INSYNC, &sh->state);
2648 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2649 struct r6_state *r6s)
2653 /* We have read all the blocks in this stripe and now we need to
2654 * copy some of them into a target stripe for expand.
2656 struct dma_async_tx_descriptor *tx = NULL;
2657 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2658 for (i = 0; i < sh->disks; i++)
2659 if (i != sh->pd_idx && (r6s && i != r6s->qd_idx)) {
2660 int dd_idx, pd_idx, j;
2661 struct stripe_head *sh2;
2663 sector_t bn = compute_blocknr(sh, i);
2664 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2666 conf->max_degraded, &dd_idx,
2668 sh2 = get_active_stripe(conf, s, conf->raid_disks,
2671 /* so far only the early blocks of this stripe
2672 * have been requested. When later blocks
2673 * get requested, we will try again
2676 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2677 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2678 /* must have already done this block */
2679 release_stripe(sh2);
2683 /* place all the copies on one channel */
2684 tx = async_memcpy(sh2->dev[dd_idx].page,
2685 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2686 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2688 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2689 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2690 for (j = 0; j < conf->raid_disks; j++)
2691 if (j != sh2->pd_idx &&
2692 (r6s && j != r6s->qd_idx) &&
2693 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2695 if (j == conf->raid_disks) {
2696 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2697 set_bit(STRIPE_HANDLE, &sh2->state);
2699 release_stripe(sh2);
2701 /* done submitting copies, wait for them to complete */
2702 if (i + 1 >= sh->disks) {
2704 dma_wait_for_async_tx(tx);
2710 * handle_stripe - do things to a stripe.
2712 * We lock the stripe and then examine the state of various bits
2713 * to see what needs to be done.
2715 * return some read request which now have data
2716 * return some write requests which are safely on disc
2717 * schedule a read on some buffers
2718 * schedule a write of some buffers
2719 * return confirmation of parity correctness
2721 * buffers are taken off read_list or write_list, and bh_cache buffers
2722 * get BH_Lock set before the stripe lock is released.
2726 static void handle_stripe5(struct stripe_head *sh)
2728 raid5_conf_t *conf = sh->raid_conf;
2729 int disks = sh->disks, i;
2730 struct bio *return_bi = NULL;
2731 struct stripe_head_state s;
2733 unsigned long pending = 0;
2735 memset(&s, 0, sizeof(s));
2736 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2737 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2738 atomic_read(&sh->count), sh->pd_idx,
2739 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2741 spin_lock(&sh->lock);
2742 clear_bit(STRIPE_HANDLE, &sh->state);
2743 clear_bit(STRIPE_DELAYED, &sh->state);
2745 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2746 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2747 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2748 /* Now to look around and see what can be done */
2751 for (i=disks; i--; ) {
2753 struct r5dev *dev = &sh->dev[i];
2754 clear_bit(R5_Insync, &dev->flags);
2756 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2757 "written %p\n", i, dev->flags, dev->toread, dev->read,
2758 dev->towrite, dev->written);
2760 /* maybe we can request a biofill operation
2762 * new wantfill requests are only permitted while
2763 * STRIPE_OP_BIOFILL is clear
2765 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2766 !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2767 set_bit(R5_Wantfill, &dev->flags);
2769 /* now count some things */
2770 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2771 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2772 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2774 if (test_bit(R5_Wantfill, &dev->flags))
2776 else if (dev->toread)
2780 if (!test_bit(R5_OVERWRITE, &dev->flags))
2785 rdev = rcu_dereference(conf->disks[i].rdev);
2786 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2787 /* The ReadError flag will just be confusing now */
2788 clear_bit(R5_ReadError, &dev->flags);
2789 clear_bit(R5_ReWrite, &dev->flags);
2791 if (!rdev || !test_bit(In_sync, &rdev->flags)
2792 || test_bit(R5_ReadError, &dev->flags)) {
2796 set_bit(R5_Insync, &dev->flags);
2800 if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2803 pr_debug("locked=%d uptodate=%d to_read=%d"
2804 " to_write=%d failed=%d failed_num=%d\n",
2805 s.locked, s.uptodate, s.to_read, s.to_write,
2806 s.failed, s.failed_num);
2807 /* check if the array has lost two devices and, if so, some requests might
2810 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2811 handle_requests_to_failed_array(conf, sh, &s, disks,
2813 if (s.failed > 1 && s.syncing) {
2814 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2815 clear_bit(STRIPE_SYNCING, &sh->state);
2819 /* might be able to return some write requests if the parity block
2820 * is safe, or on a failed drive
2822 dev = &sh->dev[sh->pd_idx];
2824 ((test_bit(R5_Insync, &dev->flags) &&
2825 !test_bit(R5_LOCKED, &dev->flags) &&
2826 test_bit(R5_UPTODATE, &dev->flags)) ||
2827 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2828 handle_completed_write_requests(conf, sh, disks, &return_bi);
2830 /* Now we might consider reading some blocks, either to check/generate
2831 * parity, or to satisfy requests
2832 * or to load a block that is being partially written.
2834 if (s.to_read || s.non_overwrite ||
2835 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2836 test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2837 handle_issuing_new_read_requests5(sh, &s, disks);
2839 /* Now we check to see if any write operations have recently
2843 /* leave prexor set until postxor is done, allows us to distinguish
2844 * a rmw from a rcw during biodrain
2846 if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2847 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2849 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2850 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2851 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2853 for (i = disks; i--; )
2854 clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2857 /* if only POSTXOR is set then this is an 'expand' postxor */
2858 if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2859 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2861 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2862 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2863 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2865 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2866 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2867 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2869 /* All the 'written' buffers and the parity block are ready to
2870 * be written back to disk
2872 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2873 for (i = disks; i--; ) {
2875 if (test_bit(R5_LOCKED, &dev->flags) &&
2876 (i == sh->pd_idx || dev->written)) {
2877 pr_debug("Writing block %d\n", i);
2878 set_bit(R5_Wantwrite, &dev->flags);
2879 if (!test_and_set_bit(
2880 STRIPE_OP_IO, &sh->ops.pending))
2882 if (!test_bit(R5_Insync, &dev->flags) ||
2883 (i == sh->pd_idx && s.failed == 0))
2884 set_bit(STRIPE_INSYNC, &sh->state);
2887 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2888 atomic_dec(&conf->preread_active_stripes);
2889 if (atomic_read(&conf->preread_active_stripes) <
2891 md_wakeup_thread(conf->mddev->thread);
2895 /* Now to consider new write requests and what else, if anything
2896 * should be read. We do not handle new writes when:
2897 * 1/ A 'write' operation (copy+xor) is already in flight.
2898 * 2/ A 'check' operation is in flight, as it may clobber the parity
2901 if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2902 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2903 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2905 /* maybe we need to check and possibly fix the parity for this stripe
2906 * Any reads will already have been scheduled, so we just see if enough
2907 * data is available. The parity check is held off while parity
2908 * dependent operations are in flight.
2910 if ((s.syncing && s.locked == 0 &&
2911 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2912 !test_bit(STRIPE_INSYNC, &sh->state)) ||
2913 test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2914 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2915 handle_parity_checks5(conf, sh, &s, disks);
2917 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2918 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2919 clear_bit(STRIPE_SYNCING, &sh->state);
2922 /* If the failed drive is just a ReadError, then we might need to progress
2923 * the repair/check process
2925 if (s.failed == 1 && !conf->mddev->ro &&
2926 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2927 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2928 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2930 dev = &sh->dev[s.failed_num];
2931 if (!test_bit(R5_ReWrite, &dev->flags)) {
2932 set_bit(R5_Wantwrite, &dev->flags);
2933 set_bit(R5_ReWrite, &dev->flags);
2934 set_bit(R5_LOCKED, &dev->flags);
2937 /* let's read it back */
2938 set_bit(R5_Wantread, &dev->flags);
2939 set_bit(R5_LOCKED, &dev->flags);
2944 /* Finish postxor operations initiated by the expansion
2947 if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2948 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2950 clear_bit(STRIPE_EXPANDING, &sh->state);
2952 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2953 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2954 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2956 for (i = conf->raid_disks; i--; ) {
2957 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2958 if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2963 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2964 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2965 /* Need to write out all blocks after computing parity */
2966 sh->disks = conf->raid_disks;
2967 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2969 s.locked += handle_write_operations5(sh, 0, 1);
2970 } else if (s.expanded &&
2971 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2972 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2973 atomic_dec(&conf->reshape_stripes);
2974 wake_up(&conf->wait_for_overlap);
2975 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2978 if (s.expanding && s.locked == 0)
2979 handle_stripe_expansion(conf, sh, NULL);
2982 pending = get_stripe_work(sh);
2984 spin_unlock(&sh->lock);
2987 raid5_run_ops(sh, pending);
2989 return_io(return_bi);
2991 for (i=disks; i-- ;) {
2995 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
2997 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3002 bi = &sh->dev[i].req;
3006 bi->bi_end_io = raid5_end_write_request;
3008 bi->bi_end_io = raid5_end_read_request;
3011 rdev = rcu_dereference(conf->disks[i].rdev);
3012 if (rdev && test_bit(Faulty, &rdev->flags))
3015 atomic_inc(&rdev->nr_pending);
3019 if (s.syncing || s.expanding || s.expanded)
3020 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3022 bi->bi_bdev = rdev->bdev;
3023 pr_debug("for %llu schedule op %ld on disc %d\n",
3024 (unsigned long long)sh->sector, bi->bi_rw, i);
3025 atomic_inc(&sh->count);
3026 bi->bi_sector = sh->sector + rdev->data_offset;
3027 bi->bi_flags = 1 << BIO_UPTODATE;
3029 bi->bi_max_vecs = 1;
3031 bi->bi_io_vec = &sh->dev[i].vec;
3032 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3033 bi->bi_io_vec[0].bv_offset = 0;
3034 bi->bi_size = STRIPE_SIZE;
3037 test_bit(R5_ReWrite, &sh->dev[i].flags))
3038 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3039 generic_make_request(bi);
3042 set_bit(STRIPE_DEGRADED, &sh->state);
3043 pr_debug("skip op %ld on disc %d for sector %llu\n",
3044 bi->bi_rw, i, (unsigned long long)sh->sector);
3045 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3046 set_bit(STRIPE_HANDLE, &sh->state);
3051 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
3053 raid6_conf_t *conf = sh->raid_conf;
3054 int disks = sh->disks;
3055 struct bio *return_bi = NULL;
3056 int i, pd_idx = sh->pd_idx;
3057 struct stripe_head_state s;
3058 struct r6_state r6s;
3059 struct r5dev *dev, *pdev, *qdev;
3061 r6s.qd_idx = raid6_next_disk(pd_idx, disks);
3062 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3063 "pd_idx=%d, qd_idx=%d\n",
3064 (unsigned long long)sh->sector, sh->state,
3065 atomic_read(&sh->count), pd_idx, r6s.qd_idx);
3066 memset(&s, 0, sizeof(s));
3068 spin_lock(&sh->lock);
3069 clear_bit(STRIPE_HANDLE, &sh->state);
3070 clear_bit(STRIPE_DELAYED, &sh->state);
3072 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3073 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3074 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3075 /* Now to look around and see what can be done */
3078 for (i=disks; i--; ) {
3081 clear_bit(R5_Insync, &dev->flags);
3083 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3084 i, dev->flags, dev->toread, dev->towrite, dev->written);
3085 /* maybe we can reply to a read */
3086 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3087 struct bio *rbi, *rbi2;
3088 pr_debug("Return read for disc %d\n", i);
3089 spin_lock_irq(&conf->device_lock);
3092 if (test_and_clear_bit(R5_Overlap, &dev->flags))
3093 wake_up(&conf->wait_for_overlap);
3094 spin_unlock_irq(&conf->device_lock);
3095 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3096 copy_data(0, rbi, dev->page, dev->sector);
3097 rbi2 = r5_next_bio(rbi, dev->sector);
3098 spin_lock_irq(&conf->device_lock);
3099 if (--rbi->bi_phys_segments == 0) {
3100 rbi->bi_next = return_bi;
3103 spin_unlock_irq(&conf->device_lock);
3108 /* now count some things */
3109 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3110 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3117 if (!test_bit(R5_OVERWRITE, &dev->flags))
3122 rdev = rcu_dereference(conf->disks[i].rdev);
3123 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3124 /* The ReadError flag will just be confusing now */
3125 clear_bit(R5_ReadError, &dev->flags);
3126 clear_bit(R5_ReWrite, &dev->flags);
3128 if (!rdev || !test_bit(In_sync, &rdev->flags)
3129 || test_bit(R5_ReadError, &dev->flags)) {
3131 r6s.failed_num[s.failed] = i;
3134 set_bit(R5_Insync, &dev->flags);
3137 pr_debug("locked=%d uptodate=%d to_read=%d"
3138 " to_write=%d failed=%d failed_num=%d,%d\n",
3139 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3140 r6s.failed_num[0], r6s.failed_num[1]);
3141 /* check if the array has lost >2 devices and, if so, some requests
3142 * might need to be failed
3144 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3145 handle_requests_to_failed_array(conf, sh, &s, disks,
3147 if (s.failed > 2 && s.syncing) {
3148 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3149 clear_bit(STRIPE_SYNCING, &sh->state);
3154 * might be able to return some write requests if the parity blocks
3155 * are safe, or on a failed drive
3157 pdev = &sh->dev[pd_idx];
3158 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3159 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3160 qdev = &sh->dev[r6s.qd_idx];
3161 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3162 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3165 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3166 && !test_bit(R5_LOCKED, &pdev->flags)
3167 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3168 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3169 && !test_bit(R5_LOCKED, &qdev->flags)
3170 && test_bit(R5_UPTODATE, &qdev->flags)))))
3171 handle_completed_write_requests(conf, sh, disks, &return_bi);
3173 /* Now we might consider reading some blocks, either to check/generate
3174 * parity, or to satisfy requests
3175 * or to load a block that is being partially written.
3177 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3178 (s.syncing && (s.uptodate < disks)) || s.expanding)
3179 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3181 /* now to consider writing and what else, if anything should be read */
3183 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3185 /* maybe we need to check and possibly fix the parity for this stripe
3186 * Any reads will already have been scheduled, so we just see if enough
3189 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3190 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3192 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3193 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3194 clear_bit(STRIPE_SYNCING, &sh->state);
3197 /* If the failed drives are just a ReadError, then we might need
3198 * to progress the repair/check process
3200 if (s.failed <= 2 && !conf->mddev->ro)
3201 for (i = 0; i < s.failed; i++) {
3202 dev = &sh->dev[r6s.failed_num[i]];
3203 if (test_bit(R5_ReadError, &dev->flags)
3204 && !test_bit(R5_LOCKED, &dev->flags)
3205 && test_bit(R5_UPTODATE, &dev->flags)
3207 if (!test_bit(R5_ReWrite, &dev->flags)) {
3208 set_bit(R5_Wantwrite, &dev->flags);
3209 set_bit(R5_ReWrite, &dev->flags);
3210 set_bit(R5_LOCKED, &dev->flags);
3212 /* let's read it back */
3213 set_bit(R5_Wantread, &dev->flags);
3214 set_bit(R5_LOCKED, &dev->flags);
3219 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3220 /* Need to write out all blocks after computing P&Q */
3221 sh->disks = conf->raid_disks;
3222 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3224 compute_parity6(sh, RECONSTRUCT_WRITE);
3225 for (i = conf->raid_disks ; i-- ; ) {
3226 set_bit(R5_LOCKED, &sh->dev[i].flags);
3228 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3230 clear_bit(STRIPE_EXPANDING, &sh->state);
3231 } else if (s.expanded) {
3232 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3233 atomic_dec(&conf->reshape_stripes);
3234 wake_up(&conf->wait_for_overlap);
3235 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3238 if (s.expanding && s.locked == 0)
3239 handle_stripe_expansion(conf, sh, &r6s);
3241 spin_unlock(&sh->lock);
3243 return_io(return_bi);
3245 for (i=disks; i-- ;) {
3249 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3251 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3256 bi = &sh->dev[i].req;
3260 bi->bi_end_io = raid5_end_write_request;
3262 bi->bi_end_io = raid5_end_read_request;
3265 rdev = rcu_dereference(conf->disks[i].rdev);
3266 if (rdev && test_bit(Faulty, &rdev->flags))
3269 atomic_inc(&rdev->nr_pending);
3273 if (s.syncing || s.expanding || s.expanded)
3274 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3276 bi->bi_bdev = rdev->bdev;
3277 pr_debug("for %llu schedule op %ld on disc %d\n",
3278 (unsigned long long)sh->sector, bi->bi_rw, i);
3279 atomic_inc(&sh->count);
3280 bi->bi_sector = sh->sector + rdev->data_offset;
3281 bi->bi_flags = 1 << BIO_UPTODATE;
3283 bi->bi_max_vecs = 1;
3285 bi->bi_io_vec = &sh->dev[i].vec;
3286 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3287 bi->bi_io_vec[0].bv_offset = 0;
3288 bi->bi_size = STRIPE_SIZE;
3291 test_bit(R5_ReWrite, &sh->dev[i].flags))
3292 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3293 generic_make_request(bi);
3296 set_bit(STRIPE_DEGRADED, &sh->state);
3297 pr_debug("skip op %ld on disc %d for sector %llu\n",
3298 bi->bi_rw, i, (unsigned long long)sh->sector);
3299 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3300 set_bit(STRIPE_HANDLE, &sh->state);
3305 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3307 if (sh->raid_conf->level == 6)
3308 handle_stripe6(sh, tmp_page);
3315 static void raid5_activate_delayed(raid5_conf_t *conf)
3317 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3318 while (!list_empty(&conf->delayed_list)) {
3319 struct list_head *l = conf->delayed_list.next;
3320 struct stripe_head *sh;
3321 sh = list_entry(l, struct stripe_head, lru);
3323 clear_bit(STRIPE_DELAYED, &sh->state);
3324 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3325 atomic_inc(&conf->preread_active_stripes);
3326 list_add_tail(&sh->lru, &conf->handle_list);
3331 static void activate_bit_delay(raid5_conf_t *conf)
3333 /* device_lock is held */
3334 struct list_head head;
3335 list_add(&head, &conf->bitmap_list);
3336 list_del_init(&conf->bitmap_list);
3337 while (!list_empty(&head)) {
3338 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3339 list_del_init(&sh->lru);
3340 atomic_inc(&sh->count);
3341 __release_stripe(conf, sh);
3345 static void unplug_slaves(mddev_t *mddev)
3347 raid5_conf_t *conf = mddev_to_conf(mddev);
3351 for (i=0; i<mddev->raid_disks; i++) {
3352 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3353 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3354 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
3356 atomic_inc(&rdev->nr_pending);
3359 if (r_queue->unplug_fn)
3360 r_queue->unplug_fn(r_queue);
3362 rdev_dec_pending(rdev, mddev);
3369 static void raid5_unplug_device(request_queue_t *q)
3371 mddev_t *mddev = q->queuedata;
3372 raid5_conf_t *conf = mddev_to_conf(mddev);
3373 unsigned long flags;
3375 spin_lock_irqsave(&conf->device_lock, flags);
3377 if (blk_remove_plug(q)) {
3379 raid5_activate_delayed(conf);
3381 md_wakeup_thread(mddev->thread);
3383 spin_unlock_irqrestore(&conf->device_lock, flags);
3385 unplug_slaves(mddev);
3388 static int raid5_issue_flush(request_queue_t *q, struct gendisk *disk,
3389 sector_t *error_sector)
3391 mddev_t *mddev = q->queuedata;
3392 raid5_conf_t *conf = mddev_to_conf(mddev);
3396 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
3397 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3398 if (rdev && !test_bit(Faulty, &rdev->flags)) {
3399 struct block_device *bdev = rdev->bdev;
3400 request_queue_t *r_queue = bdev_get_queue(bdev);
3402 if (!r_queue->issue_flush_fn)
3405 atomic_inc(&rdev->nr_pending);
3407 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
3409 rdev_dec_pending(rdev, mddev);
3418 static int raid5_congested(void *data, int bits)
3420 mddev_t *mddev = data;
3421 raid5_conf_t *conf = mddev_to_conf(mddev);
3423 /* No difference between reads and writes. Just check
3424 * how busy the stripe_cache is
3426 if (conf->inactive_blocked)
3430 if (list_empty_careful(&conf->inactive_list))
3436 /* We want read requests to align with chunks where possible,
3437 * but write requests don't need to.
3439 static int raid5_mergeable_bvec(request_queue_t *q, struct bio *bio, struct bio_vec *biovec)
3441 mddev_t *mddev = q->queuedata;
3442 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3444 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3445 unsigned int bio_sectors = bio->bi_size >> 9;
3447 if (bio_data_dir(bio) == WRITE)
3448 return biovec->bv_len; /* always allow writes to be mergeable */
3450 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3451 if (max < 0) max = 0;
3452 if (max <= biovec->bv_len && bio_sectors == 0)
3453 return biovec->bv_len;
3459 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3461 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3462 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3463 unsigned int bio_sectors = bio->bi_size >> 9;
3465 return chunk_sectors >=
3466 ((sector & (chunk_sectors - 1)) + bio_sectors);
3470 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3471 * later sampled by raid5d.
3473 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3475 unsigned long flags;
3477 spin_lock_irqsave(&conf->device_lock, flags);
3479 bi->bi_next = conf->retry_read_aligned_list;
3480 conf->retry_read_aligned_list = bi;
3482 spin_unlock_irqrestore(&conf->device_lock, flags);
3483 md_wakeup_thread(conf->mddev->thread);
3487 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3491 bi = conf->retry_read_aligned;
3493 conf->retry_read_aligned = NULL;
3496 bi = conf->retry_read_aligned_list;
3498 conf->retry_read_aligned_list = bi->bi_next;
3500 bi->bi_phys_segments = 1; /* biased count of active stripes */
3501 bi->bi_hw_segments = 0; /* count of processed stripes */
3509 * The "raid5_align_endio" should check if the read succeeded and if it
3510 * did, call bio_endio on the original bio (having bio_put the new bio
3512 * If the read failed..
3514 static int raid5_align_endio(struct bio *bi, unsigned int bytes, int error)
3516 struct bio* raid_bi = bi->bi_private;
3519 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3526 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3527 conf = mddev_to_conf(mddev);
3528 rdev = (void*)raid_bi->bi_next;
3529 raid_bi->bi_next = NULL;
3531 rdev_dec_pending(rdev, conf->mddev);
3533 if (!error && uptodate) {
3534 bio_endio(raid_bi, bytes, 0);
3535 if (atomic_dec_and_test(&conf->active_aligned_reads))
3536 wake_up(&conf->wait_for_stripe);
3541 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3543 add_bio_to_retry(raid_bi, conf);
3547 static int bio_fits_rdev(struct bio *bi)
3549 request_queue_t *q = bdev_get_queue(bi->bi_bdev);
3551 if ((bi->bi_size>>9) > q->max_sectors)
3553 blk_recount_segments(q, bi);
3554 if (bi->bi_phys_segments > q->max_phys_segments ||
3555 bi->bi_hw_segments > q->max_hw_segments)
3558 if (q->merge_bvec_fn)
3559 /* it's too hard to apply the merge_bvec_fn at this stage,
3568 static int chunk_aligned_read(request_queue_t *q, struct bio * raid_bio)
3570 mddev_t *mddev = q->queuedata;
3571 raid5_conf_t *conf = mddev_to_conf(mddev);
3572 const unsigned int raid_disks = conf->raid_disks;
3573 const unsigned int data_disks = raid_disks - conf->max_degraded;
3574 unsigned int dd_idx, pd_idx;
3575 struct bio* align_bi;
3578 if (!in_chunk_boundary(mddev, raid_bio)) {
3579 pr_debug("chunk_aligned_read : non aligned\n");
3583 * use bio_clone to make a copy of the bio
3585 align_bi = bio_clone(raid_bio, GFP_NOIO);
3589 * set bi_end_io to a new function, and set bi_private to the
3592 align_bi->bi_end_io = raid5_align_endio;
3593 align_bi->bi_private = raid_bio;
3597 align_bi->bi_sector = raid5_compute_sector(raid_bio->bi_sector,
3605 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3606 if (rdev && test_bit(In_sync, &rdev->flags)) {
3607 atomic_inc(&rdev->nr_pending);
3609 raid_bio->bi_next = (void*)rdev;
3610 align_bi->bi_bdev = rdev->bdev;
3611 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3612 align_bi->bi_sector += rdev->data_offset;
3614 if (!bio_fits_rdev(align_bi)) {
3615 /* too big in some way */
3617 rdev_dec_pending(rdev, mddev);
3621 spin_lock_irq(&conf->device_lock);
3622 wait_event_lock_irq(conf->wait_for_stripe,
3624 conf->device_lock, /* nothing */);
3625 atomic_inc(&conf->active_aligned_reads);
3626 spin_unlock_irq(&conf->device_lock);
3628 generic_make_request(align_bi);
3638 static int make_request(request_queue_t *q, struct bio * bi)
3640 mddev_t *mddev = q->queuedata;
3641 raid5_conf_t *conf = mddev_to_conf(mddev);
3642 unsigned int dd_idx, pd_idx;
3643 sector_t new_sector;
3644 sector_t logical_sector, last_sector;
3645 struct stripe_head *sh;
3646 const int rw = bio_data_dir(bi);
3649 if (unlikely(bio_barrier(bi))) {
3650 bio_endio(bi, bi->bi_size, -EOPNOTSUPP);
3654 md_write_start(mddev, bi);
3656 disk_stat_inc(mddev->gendisk, ios[rw]);
3657 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3660 mddev->reshape_position == MaxSector &&
3661 chunk_aligned_read(q,bi))
3664 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3665 last_sector = bi->bi_sector + (bi->bi_size>>9);
3667 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3669 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3671 int disks, data_disks;
3674 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3675 if (likely(conf->expand_progress == MaxSector))
3676 disks = conf->raid_disks;
3678 /* spinlock is needed as expand_progress may be
3679 * 64bit on a 32bit platform, and so it might be
3680 * possible to see a half-updated value
3681 * Ofcourse expand_progress could change after
3682 * the lock is dropped, so once we get a reference
3683 * to the stripe that we think it is, we will have
3686 spin_lock_irq(&conf->device_lock);
3687 disks = conf->raid_disks;
3688 if (logical_sector >= conf->expand_progress)
3689 disks = conf->previous_raid_disks;
3691 if (logical_sector >= conf->expand_lo) {
3692 spin_unlock_irq(&conf->device_lock);
3697 spin_unlock_irq(&conf->device_lock);
3699 data_disks = disks - conf->max_degraded;
3701 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3702 &dd_idx, &pd_idx, conf);
3703 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3704 (unsigned long long)new_sector,
3705 (unsigned long long)logical_sector);
3707 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3709 if (unlikely(conf->expand_progress != MaxSector)) {
3710 /* expansion might have moved on while waiting for a
3711 * stripe, so we must do the range check again.
3712 * Expansion could still move past after this
3713 * test, but as we are holding a reference to
3714 * 'sh', we know that if that happens,
3715 * STRIPE_EXPANDING will get set and the expansion
3716 * won't proceed until we finish with the stripe.
3719 spin_lock_irq(&conf->device_lock);
3720 if (logical_sector < conf->expand_progress &&
3721 disks == conf->previous_raid_disks)
3722 /* mismatch, need to try again */
3724 spin_unlock_irq(&conf->device_lock);
3730 /* FIXME what if we get a false positive because these
3731 * are being updated.
3733 if (logical_sector >= mddev->suspend_lo &&
3734 logical_sector < mddev->suspend_hi) {
3740 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3741 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3742 /* Stripe is busy expanding or
3743 * add failed due to overlap. Flush everything
3746 raid5_unplug_device(mddev->queue);
3751 finish_wait(&conf->wait_for_overlap, &w);
3752 handle_stripe(sh, NULL);
3755 /* cannot get stripe for read-ahead, just give-up */
3756 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3757 finish_wait(&conf->wait_for_overlap, &w);
3762 spin_lock_irq(&conf->device_lock);
3763 remaining = --bi->bi_phys_segments;
3764 spin_unlock_irq(&conf->device_lock);
3765 if (remaining == 0) {
3766 int bytes = bi->bi_size;
3769 md_write_end(mddev);
3771 bi->bi_end_io(bi, bytes,
3772 test_bit(BIO_UPTODATE, &bi->bi_flags)
3778 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3780 /* reshaping is quite different to recovery/resync so it is
3781 * handled quite separately ... here.
3783 * On each call to sync_request, we gather one chunk worth of
3784 * destination stripes and flag them as expanding.
3785 * Then we find all the source stripes and request reads.
3786 * As the reads complete, handle_stripe will copy the data
3787 * into the destination stripe and release that stripe.
3789 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3790 struct stripe_head *sh;
3792 sector_t first_sector, last_sector;
3793 int raid_disks = conf->previous_raid_disks;
3794 int data_disks = raid_disks - conf->max_degraded;
3795 int new_data_disks = conf->raid_disks - conf->max_degraded;
3798 sector_t writepos, safepos, gap;
3800 if (sector_nr == 0 &&
3801 conf->expand_progress != 0) {
3802 /* restarting in the middle, skip the initial sectors */
3803 sector_nr = conf->expand_progress;
3804 sector_div(sector_nr, new_data_disks);
3809 /* we update the metadata when there is more than 3Meg
3810 * in the block range (that is rather arbitrary, should
3811 * probably be time based) or when the data about to be
3812 * copied would over-write the source of the data at
3813 * the front of the range.
3814 * i.e. one new_stripe forward from expand_progress new_maps
3815 * to after where expand_lo old_maps to
3817 writepos = conf->expand_progress +
3818 conf->chunk_size/512*(new_data_disks);
3819 sector_div(writepos, new_data_disks);
3820 safepos = conf->expand_lo;
3821 sector_div(safepos, data_disks);
3822 gap = conf->expand_progress - conf->expand_lo;
3824 if (writepos >= safepos ||
3825 gap > (new_data_disks)*3000*2 /*3Meg*/) {
3826 /* Cannot proceed until we've updated the superblock... */
3827 wait_event(conf->wait_for_overlap,
3828 atomic_read(&conf->reshape_stripes)==0);
3829 mddev->reshape_position = conf->expand_progress;
3830 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3831 md_wakeup_thread(mddev->thread);
3832 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3833 kthread_should_stop());
3834 spin_lock_irq(&conf->device_lock);
3835 conf->expand_lo = mddev->reshape_position;
3836 spin_unlock_irq(&conf->device_lock);
3837 wake_up(&conf->wait_for_overlap);
3840 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3843 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3844 sh = get_active_stripe(conf, sector_nr+i,
3845 conf->raid_disks, pd_idx, 0);
3846 set_bit(STRIPE_EXPANDING, &sh->state);
3847 atomic_inc(&conf->reshape_stripes);
3848 /* If any of this stripe is beyond the end of the old
3849 * array, then we need to zero those blocks
3851 for (j=sh->disks; j--;) {
3853 if (j == sh->pd_idx)
3855 if (conf->level == 6 &&
3856 j == raid6_next_disk(sh->pd_idx, sh->disks))
3858 s = compute_blocknr(sh, j);
3859 if (s < (mddev->array_size<<1)) {
3863 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3864 set_bit(R5_Expanded, &sh->dev[j].flags);
3865 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3868 set_bit(STRIPE_EXPAND_READY, &sh->state);
3869 set_bit(STRIPE_HANDLE, &sh->state);
3873 spin_lock_irq(&conf->device_lock);
3874 conf->expand_progress = (sector_nr + i) * new_data_disks;
3875 spin_unlock_irq(&conf->device_lock);
3876 /* Ok, those stripe are ready. We can start scheduling
3877 * reads on the source stripes.
3878 * The source stripes are determined by mapping the first and last
3879 * block on the destination stripes.
3882 raid5_compute_sector(sector_nr*(new_data_disks),
3883 raid_disks, data_disks,
3884 &dd_idx, &pd_idx, conf);
3886 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3887 *(new_data_disks) -1,
3888 raid_disks, data_disks,
3889 &dd_idx, &pd_idx, conf);
3890 if (last_sector >= (mddev->size<<1))
3891 last_sector = (mddev->size<<1)-1;
3892 while (first_sector <= last_sector) {
3893 pd_idx = stripe_to_pdidx(first_sector, conf,
3894 conf->previous_raid_disks);
3895 sh = get_active_stripe(conf, first_sector,
3896 conf->previous_raid_disks, pd_idx, 0);
3897 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3898 set_bit(STRIPE_HANDLE, &sh->state);
3900 first_sector += STRIPE_SECTORS;
3902 return conf->chunk_size>>9;
3905 /* FIXME go_faster isn't used */
3906 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3908 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3909 struct stripe_head *sh;
3911 int raid_disks = conf->raid_disks;
3912 sector_t max_sector = mddev->size << 1;
3914 int still_degraded = 0;
3917 if (sector_nr >= max_sector) {
3918 /* just being told to finish up .. nothing much to do */
3919 unplug_slaves(mddev);
3920 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3925 if (mddev->curr_resync < max_sector) /* aborted */
3926 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3928 else /* completed sync */
3930 bitmap_close_sync(mddev->bitmap);
3935 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3936 return reshape_request(mddev, sector_nr, skipped);
3938 /* if there is too many failed drives and we are trying
3939 * to resync, then assert that we are finished, because there is
3940 * nothing we can do.
3942 if (mddev->degraded >= conf->max_degraded &&
3943 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3944 sector_t rv = (mddev->size << 1) - sector_nr;
3948 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3949 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3950 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3951 /* we can skip this block, and probably more */
3952 sync_blocks /= STRIPE_SECTORS;
3954 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3957 pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3958 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3960 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3961 /* make sure we don't swamp the stripe cache if someone else
3962 * is trying to get access
3964 schedule_timeout_uninterruptible(1);
3966 /* Need to check if array will still be degraded after recovery/resync
3967 * We don't need to check the 'failed' flag as when that gets set,
3970 for (i=0; i<mddev->raid_disks; i++)
3971 if (conf->disks[i].rdev == NULL)
3974 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3976 spin_lock(&sh->lock);
3977 set_bit(STRIPE_SYNCING, &sh->state);
3978 clear_bit(STRIPE_INSYNC, &sh->state);
3979 spin_unlock(&sh->lock);
3981 handle_stripe(sh, NULL);
3984 return STRIPE_SECTORS;
3987 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3989 /* We may not be able to submit a whole bio at once as there
3990 * may not be enough stripe_heads available.
3991 * We cannot pre-allocate enough stripe_heads as we may need
3992 * more than exist in the cache (if we allow ever large chunks).
3993 * So we do one stripe head at a time and record in
3994 * ->bi_hw_segments how many have been done.
3996 * We *know* that this entire raid_bio is in one chunk, so
3997 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3999 struct stripe_head *sh;
4001 sector_t sector, logical_sector, last_sector;
4006 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4007 sector = raid5_compute_sector( logical_sector,
4009 conf->raid_disks - conf->max_degraded,
4013 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4015 for (; logical_sector < last_sector;
4016 logical_sector += STRIPE_SECTORS,
4017 sector += STRIPE_SECTORS,
4020 if (scnt < raid_bio->bi_hw_segments)
4021 /* already done this stripe */
4024 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
4027 /* failed to get a stripe - must wait */
4028 raid_bio->bi_hw_segments = scnt;
4029 conf->retry_read_aligned = raid_bio;
4033 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4034 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4036 raid_bio->bi_hw_segments = scnt;
4037 conf->retry_read_aligned = raid_bio;
4041 handle_stripe(sh, NULL);
4045 spin_lock_irq(&conf->device_lock);
4046 remaining = --raid_bio->bi_phys_segments;
4047 spin_unlock_irq(&conf->device_lock);
4048 if (remaining == 0) {
4049 int bytes = raid_bio->bi_size;
4051 raid_bio->bi_size = 0;
4052 raid_bio->bi_end_io(raid_bio, bytes,
4053 test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
4056 if (atomic_dec_and_test(&conf->active_aligned_reads))
4057 wake_up(&conf->wait_for_stripe);
4064 * This is our raid5 kernel thread.
4066 * We scan the hash table for stripes which can be handled now.
4067 * During the scan, completed stripes are saved for us by the interrupt
4068 * handler, so that they will not have to wait for our next wakeup.
4070 static void raid5d (mddev_t *mddev)
4072 struct stripe_head *sh;
4073 raid5_conf_t *conf = mddev_to_conf(mddev);
4076 pr_debug("+++ raid5d active\n");
4078 md_check_recovery(mddev);
4081 spin_lock_irq(&conf->device_lock);
4083 struct list_head *first;
4086 if (conf->seq_flush != conf->seq_write) {
4087 int seq = conf->seq_flush;
4088 spin_unlock_irq(&conf->device_lock);
4089 bitmap_unplug(mddev->bitmap);
4090 spin_lock_irq(&conf->device_lock);
4091 conf->seq_write = seq;
4092 activate_bit_delay(conf);
4095 if (list_empty(&conf->handle_list) &&
4096 atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD &&
4097 !blk_queue_plugged(mddev->queue) &&
4098 !list_empty(&conf->delayed_list))
4099 raid5_activate_delayed(conf);
4101 while ((bio = remove_bio_from_retry(conf))) {
4103 spin_unlock_irq(&conf->device_lock);
4104 ok = retry_aligned_read(conf, bio);
4105 spin_lock_irq(&conf->device_lock);
4111 if (list_empty(&conf->handle_list)) {
4112 async_tx_issue_pending_all();
4116 first = conf->handle_list.next;
4117 sh = list_entry(first, struct stripe_head, lru);
4119 list_del_init(first);
4120 atomic_inc(&sh->count);
4121 BUG_ON(atomic_read(&sh->count)!= 1);
4122 spin_unlock_irq(&conf->device_lock);
4125 handle_stripe(sh, conf->spare_page);
4128 spin_lock_irq(&conf->device_lock);
4130 pr_debug("%d stripes handled\n", handled);
4132 spin_unlock_irq(&conf->device_lock);
4134 unplug_slaves(mddev);
4136 pr_debug("--- raid5d inactive\n");
4140 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4142 raid5_conf_t *conf = mddev_to_conf(mddev);
4144 return sprintf(page, "%d\n", conf->max_nr_stripes);
4150 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4152 raid5_conf_t *conf = mddev_to_conf(mddev);
4155 if (len >= PAGE_SIZE)
4160 new = simple_strtoul(page, &end, 10);
4161 if (!*page || (*end && *end != '\n') )
4163 if (new <= 16 || new > 32768)
4165 while (new < conf->max_nr_stripes) {
4166 if (drop_one_stripe(conf))
4167 conf->max_nr_stripes--;
4171 md_allow_write(mddev);
4172 while (new > conf->max_nr_stripes) {
4173 if (grow_one_stripe(conf))
4174 conf->max_nr_stripes++;
4180 static struct md_sysfs_entry
4181 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4182 raid5_show_stripe_cache_size,
4183 raid5_store_stripe_cache_size);
4186 stripe_cache_active_show(mddev_t *mddev, char *page)
4188 raid5_conf_t *conf = mddev_to_conf(mddev);
4190 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4195 static struct md_sysfs_entry
4196 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4198 static struct attribute *raid5_attrs[] = {
4199 &raid5_stripecache_size.attr,
4200 &raid5_stripecache_active.attr,
4203 static struct attribute_group raid5_attrs_group = {
4205 .attrs = raid5_attrs,
4208 static int run(mddev_t *mddev)
4211 int raid_disk, memory;
4213 struct disk_info *disk;
4214 struct list_head *tmp;
4215 int working_disks = 0;
4217 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4218 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4219 mdname(mddev), mddev->level);
4223 if (mddev->reshape_position != MaxSector) {
4224 /* Check that we can continue the reshape.
4225 * Currently only disks can change, it must
4226 * increase, and we must be past the point where
4227 * a stripe over-writes itself
4229 sector_t here_new, here_old;
4231 int max_degraded = (mddev->level == 5 ? 1 : 2);
4233 if (mddev->new_level != mddev->level ||
4234 mddev->new_layout != mddev->layout ||
4235 mddev->new_chunk != mddev->chunk_size) {
4236 printk(KERN_ERR "raid5: %s: unsupported reshape "
4237 "required - aborting.\n",
4241 if (mddev->delta_disks <= 0) {
4242 printk(KERN_ERR "raid5: %s: unsupported reshape "
4243 "(reduce disks) required - aborting.\n",
4247 old_disks = mddev->raid_disks - mddev->delta_disks;
4248 /* reshape_position must be on a new-stripe boundary, and one
4249 * further up in new geometry must map after here in old
4252 here_new = mddev->reshape_position;
4253 if (sector_div(here_new, (mddev->chunk_size>>9)*
4254 (mddev->raid_disks - max_degraded))) {
4255 printk(KERN_ERR "raid5: reshape_position not "
4256 "on a stripe boundary\n");
4259 /* here_new is the stripe we will write to */
4260 here_old = mddev->reshape_position;
4261 sector_div(here_old, (mddev->chunk_size>>9)*
4262 (old_disks-max_degraded));
4263 /* here_old is the first stripe that we might need to read
4265 if (here_new >= here_old) {
4266 /* Reading from the same stripe as writing to - bad */
4267 printk(KERN_ERR "raid5: reshape_position too early for "
4268 "auto-recovery - aborting.\n");
4271 printk(KERN_INFO "raid5: reshape will continue\n");
4272 /* OK, we should be able to continue; */
4276 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4277 if ((conf = mddev->private) == NULL)
4279 if (mddev->reshape_position == MaxSector) {
4280 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4282 conf->raid_disks = mddev->raid_disks;
4283 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4286 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4291 conf->mddev = mddev;
4293 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4296 if (mddev->level == 6) {
4297 conf->spare_page = alloc_page(GFP_KERNEL);
4298 if (!conf->spare_page)
4301 spin_lock_init(&conf->device_lock);
4302 init_waitqueue_head(&conf->wait_for_stripe);
4303 init_waitqueue_head(&conf->wait_for_overlap);
4304 INIT_LIST_HEAD(&conf->handle_list);
4305 INIT_LIST_HEAD(&conf->delayed_list);
4306 INIT_LIST_HEAD(&conf->bitmap_list);
4307 INIT_LIST_HEAD(&conf->inactive_list);
4308 atomic_set(&conf->active_stripes, 0);
4309 atomic_set(&conf->preread_active_stripes, 0);
4310 atomic_set(&conf->active_aligned_reads, 0);
4312 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4314 ITERATE_RDEV(mddev,rdev,tmp) {
4315 raid_disk = rdev->raid_disk;
4316 if (raid_disk >= conf->raid_disks
4319 disk = conf->disks + raid_disk;
4323 if (test_bit(In_sync, &rdev->flags)) {
4324 char b[BDEVNAME_SIZE];
4325 printk(KERN_INFO "raid5: device %s operational as raid"
4326 " disk %d\n", bdevname(rdev->bdev,b),
4333 * 0 for a fully functional array, 1 or 2 for a degraded array.
4335 mddev->degraded = conf->raid_disks - working_disks;
4336 conf->mddev = mddev;
4337 conf->chunk_size = mddev->chunk_size;
4338 conf->level = mddev->level;
4339 if (conf->level == 6)
4340 conf->max_degraded = 2;
4342 conf->max_degraded = 1;
4343 conf->algorithm = mddev->layout;
4344 conf->max_nr_stripes = NR_STRIPES;
4345 conf->expand_progress = mddev->reshape_position;
4347 /* device size must be a multiple of chunk size */
4348 mddev->size &= ~(mddev->chunk_size/1024 -1);
4349 mddev->resync_max_sectors = mddev->size << 1;
4351 if (conf->level == 6 && conf->raid_disks < 4) {
4352 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4353 mdname(mddev), conf->raid_disks);
4356 if (!conf->chunk_size || conf->chunk_size % 4) {
4357 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4358 conf->chunk_size, mdname(mddev));
4361 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4363 "raid5: unsupported parity algorithm %d for %s\n",
4364 conf->algorithm, mdname(mddev));
4367 if (mddev->degraded > conf->max_degraded) {
4368 printk(KERN_ERR "raid5: not enough operational devices for %s"
4369 " (%d/%d failed)\n",
4370 mdname(mddev), mddev->degraded, conf->raid_disks);
4374 if (mddev->degraded > 0 &&
4375 mddev->recovery_cp != MaxSector) {
4376 if (mddev->ok_start_degraded)
4378 "raid5: starting dirty degraded array: %s"
4379 "- data corruption possible.\n",
4383 "raid5: cannot start dirty degraded array for %s\n",
4390 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4391 if (!mddev->thread) {
4393 "raid5: couldn't allocate thread for %s\n",
4398 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4399 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4400 if (grow_stripes(conf, conf->max_nr_stripes)) {
4402 "raid5: couldn't allocate %dkB for buffers\n", memory);
4403 shrink_stripes(conf);
4404 md_unregister_thread(mddev->thread);
4407 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4408 memory, mdname(mddev));
4410 if (mddev->degraded == 0)
4411 printk("raid5: raid level %d set %s active with %d out of %d"
4412 " devices, algorithm %d\n", conf->level, mdname(mddev),
4413 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4416 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4417 " out of %d devices, algorithm %d\n", conf->level,
4418 mdname(mddev), mddev->raid_disks - mddev->degraded,
4419 mddev->raid_disks, conf->algorithm);
4421 print_raid5_conf(conf);
4423 if (conf->expand_progress != MaxSector) {
4424 printk("...ok start reshape thread\n");
4425 conf->expand_lo = conf->expand_progress;
4426 atomic_set(&conf->reshape_stripes, 0);
4427 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4428 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4429 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4430 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4431 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4435 /* read-ahead size must cover two whole stripes, which is
4436 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4439 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4440 int stripe = data_disks *
4441 (mddev->chunk_size / PAGE_SIZE);
4442 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4443 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4446 /* Ok, everything is just fine now */
4447 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4449 "raid5: failed to create sysfs attributes for %s\n",
4452 mddev->queue->unplug_fn = raid5_unplug_device;
4453 mddev->queue->issue_flush_fn = raid5_issue_flush;
4454 mddev->queue->backing_dev_info.congested_data = mddev;
4455 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4457 mddev->array_size = mddev->size * (conf->previous_raid_disks -
4458 conf->max_degraded);
4460 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4465 print_raid5_conf(conf);
4466 safe_put_page(conf->spare_page);
4468 kfree(conf->stripe_hashtbl);
4471 mddev->private = NULL;
4472 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4478 static int stop(mddev_t *mddev)
4480 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4482 md_unregister_thread(mddev->thread);
4483 mddev->thread = NULL;
4484 shrink_stripes(conf);
4485 kfree(conf->stripe_hashtbl);
4486 mddev->queue->backing_dev_info.congested_fn = NULL;
4487 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4488 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4491 mddev->private = NULL;
4496 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4500 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4501 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4502 seq_printf(seq, "sh %llu, count %d.\n",
4503 (unsigned long long)sh->sector, atomic_read(&sh->count));
4504 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4505 for (i = 0; i < sh->disks; i++) {
4506 seq_printf(seq, "(cache%d: %p %ld) ",
4507 i, sh->dev[i].page, sh->dev[i].flags);
4509 seq_printf(seq, "\n");
4512 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4514 struct stripe_head *sh;
4515 struct hlist_node *hn;
4518 spin_lock_irq(&conf->device_lock);
4519 for (i = 0; i < NR_HASH; i++) {
4520 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4521 if (sh->raid_conf != conf)
4526 spin_unlock_irq(&conf->device_lock);
4530 static void status (struct seq_file *seq, mddev_t *mddev)
4532 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4535 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4536 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4537 for (i = 0; i < conf->raid_disks; i++)
4538 seq_printf (seq, "%s",
4539 conf->disks[i].rdev &&
4540 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4541 seq_printf (seq, "]");
4543 seq_printf (seq, "\n");
4544 printall(seq, conf);
4548 static void print_raid5_conf (raid5_conf_t *conf)
4551 struct disk_info *tmp;
4553 printk("RAID5 conf printout:\n");
4555 printk("(conf==NULL)\n");
4558 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4559 conf->raid_disks - conf->mddev->degraded);
4561 for (i = 0; i < conf->raid_disks; i++) {
4562 char b[BDEVNAME_SIZE];
4563 tmp = conf->disks + i;
4565 printk(" disk %d, o:%d, dev:%s\n",
4566 i, !test_bit(Faulty, &tmp->rdev->flags),
4567 bdevname(tmp->rdev->bdev,b));
4571 static int raid5_spare_active(mddev_t *mddev)
4574 raid5_conf_t *conf = mddev->private;
4575 struct disk_info *tmp;
4577 for (i = 0; i < conf->raid_disks; i++) {
4578 tmp = conf->disks + i;
4580 && !test_bit(Faulty, &tmp->rdev->flags)
4581 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4582 unsigned long flags;
4583 spin_lock_irqsave(&conf->device_lock, flags);
4585 spin_unlock_irqrestore(&conf->device_lock, flags);
4588 print_raid5_conf(conf);
4592 static int raid5_remove_disk(mddev_t *mddev, int number)
4594 raid5_conf_t *conf = mddev->private;
4597 struct disk_info *p = conf->disks + number;
4599 print_raid5_conf(conf);
4602 if (test_bit(In_sync, &rdev->flags) ||
4603 atomic_read(&rdev->nr_pending)) {
4609 if (atomic_read(&rdev->nr_pending)) {
4610 /* lost the race, try later */
4617 print_raid5_conf(conf);
4621 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4623 raid5_conf_t *conf = mddev->private;
4626 struct disk_info *p;
4628 if (mddev->degraded > conf->max_degraded)
4629 /* no point adding a device */
4633 * find the disk ... but prefer rdev->saved_raid_disk
4636 if (rdev->saved_raid_disk >= 0 &&
4637 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4638 disk = rdev->saved_raid_disk;
4641 for ( ; disk < conf->raid_disks; disk++)
4642 if ((p=conf->disks + disk)->rdev == NULL) {
4643 clear_bit(In_sync, &rdev->flags);
4644 rdev->raid_disk = disk;
4646 if (rdev->saved_raid_disk != disk)
4648 rcu_assign_pointer(p->rdev, rdev);
4651 print_raid5_conf(conf);
4655 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4657 /* no resync is happening, and there is enough space
4658 * on all devices, so we can resize.
4659 * We need to make sure resync covers any new space.
4660 * If the array is shrinking we should possibly wait until
4661 * any io in the removed space completes, but it hardly seems
4664 raid5_conf_t *conf = mddev_to_conf(mddev);
4666 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4667 mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4668 set_capacity(mddev->gendisk, mddev->array_size << 1);
4670 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) {
4671 mddev->recovery_cp = mddev->size << 1;
4672 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4674 mddev->size = sectors /2;
4675 mddev->resync_max_sectors = sectors;
4679 #ifdef CONFIG_MD_RAID5_RESHAPE
4680 static int raid5_check_reshape(mddev_t *mddev)
4682 raid5_conf_t *conf = mddev_to_conf(mddev);
4685 if (mddev->delta_disks < 0 ||
4686 mddev->new_level != mddev->level)
4687 return -EINVAL; /* Cannot shrink array or change level yet */
4688 if (mddev->delta_disks == 0)
4689 return 0; /* nothing to do */
4691 /* Can only proceed if there are plenty of stripe_heads.
4692 * We need a minimum of one full stripe,, and for sensible progress
4693 * it is best to have about 4 times that.
4694 * If we require 4 times, then the default 256 4K stripe_heads will
4695 * allow for chunk sizes up to 256K, which is probably OK.
4696 * If the chunk size is greater, user-space should request more
4697 * stripe_heads first.
4699 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4700 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4701 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4702 (mddev->chunk_size / STRIPE_SIZE)*4);
4706 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4710 if (mddev->degraded > conf->max_degraded)
4712 /* looks like we might be able to manage this */
4716 static int raid5_start_reshape(mddev_t *mddev)
4718 raid5_conf_t *conf = mddev_to_conf(mddev);
4720 struct list_head *rtmp;
4722 int added_devices = 0;
4723 unsigned long flags;
4725 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4728 ITERATE_RDEV(mddev, rdev, rtmp)
4729 if (rdev->raid_disk < 0 &&
4730 !test_bit(Faulty, &rdev->flags))
4733 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4734 /* Not enough devices even to make a degraded array
4739 atomic_set(&conf->reshape_stripes, 0);
4740 spin_lock_irq(&conf->device_lock);
4741 conf->previous_raid_disks = conf->raid_disks;
4742 conf->raid_disks += mddev->delta_disks;
4743 conf->expand_progress = 0;
4744 conf->expand_lo = 0;
4745 spin_unlock_irq(&conf->device_lock);
4747 /* Add some new drives, as many as will fit.
4748 * We know there are enough to make the newly sized array work.
4750 ITERATE_RDEV(mddev, rdev, rtmp)
4751 if (rdev->raid_disk < 0 &&
4752 !test_bit(Faulty, &rdev->flags)) {
4753 if (raid5_add_disk(mddev, rdev)) {
4755 set_bit(In_sync, &rdev->flags);
4757 rdev->recovery_offset = 0;
4758 sprintf(nm, "rd%d", rdev->raid_disk);
4759 if (sysfs_create_link(&mddev->kobj,
4762 "raid5: failed to create "
4763 " link %s for %s\n",
4769 spin_lock_irqsave(&conf->device_lock, flags);
4770 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4771 spin_unlock_irqrestore(&conf->device_lock, flags);
4772 mddev->raid_disks = conf->raid_disks;
4773 mddev->reshape_position = 0;
4774 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4776 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4777 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4778 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4779 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4780 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4782 if (!mddev->sync_thread) {
4783 mddev->recovery = 0;
4784 spin_lock_irq(&conf->device_lock);
4785 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4786 conf->expand_progress = MaxSector;
4787 spin_unlock_irq(&conf->device_lock);
4790 md_wakeup_thread(mddev->sync_thread);
4791 md_new_event(mddev);
4796 static void end_reshape(raid5_conf_t *conf)
4798 struct block_device *bdev;
4800 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4801 conf->mddev->array_size = conf->mddev->size *
4802 (conf->raid_disks - conf->max_degraded);
4803 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4804 conf->mddev->changed = 1;
4806 bdev = bdget_disk(conf->mddev->gendisk, 0);
4808 mutex_lock(&bdev->bd_inode->i_mutex);
4809 i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4810 mutex_unlock(&bdev->bd_inode->i_mutex);
4813 spin_lock_irq(&conf->device_lock);
4814 conf->expand_progress = MaxSector;
4815 spin_unlock_irq(&conf->device_lock);
4816 conf->mddev->reshape_position = MaxSector;
4818 /* read-ahead size must cover two whole stripes, which is
4819 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4822 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4823 int stripe = data_disks *
4824 (conf->mddev->chunk_size / PAGE_SIZE);
4825 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4826 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4831 static void raid5_quiesce(mddev_t *mddev, int state)
4833 raid5_conf_t *conf = mddev_to_conf(mddev);
4836 case 2: /* resume for a suspend */
4837 wake_up(&conf->wait_for_overlap);
4840 case 1: /* stop all writes */
4841 spin_lock_irq(&conf->device_lock);
4843 wait_event_lock_irq(conf->wait_for_stripe,
4844 atomic_read(&conf->active_stripes) == 0 &&
4845 atomic_read(&conf->active_aligned_reads) == 0,
4846 conf->device_lock, /* nothing */);
4847 spin_unlock_irq(&conf->device_lock);
4850 case 0: /* re-enable writes */
4851 spin_lock_irq(&conf->device_lock);
4853 wake_up(&conf->wait_for_stripe);
4854 wake_up(&conf->wait_for_overlap);
4855 spin_unlock_irq(&conf->device_lock);
4860 static struct mdk_personality raid6_personality =
4864 .owner = THIS_MODULE,
4865 .make_request = make_request,
4869 .error_handler = error,
4870 .hot_add_disk = raid5_add_disk,
4871 .hot_remove_disk= raid5_remove_disk,
4872 .spare_active = raid5_spare_active,
4873 .sync_request = sync_request,
4874 .resize = raid5_resize,
4875 #ifdef CONFIG_MD_RAID5_RESHAPE
4876 .check_reshape = raid5_check_reshape,
4877 .start_reshape = raid5_start_reshape,
4879 .quiesce = raid5_quiesce,
4881 static struct mdk_personality raid5_personality =
4885 .owner = THIS_MODULE,
4886 .make_request = make_request,
4890 .error_handler = error,
4891 .hot_add_disk = raid5_add_disk,
4892 .hot_remove_disk= raid5_remove_disk,
4893 .spare_active = raid5_spare_active,
4894 .sync_request = sync_request,
4895 .resize = raid5_resize,
4896 #ifdef CONFIG_MD_RAID5_RESHAPE
4897 .check_reshape = raid5_check_reshape,
4898 .start_reshape = raid5_start_reshape,
4900 .quiesce = raid5_quiesce,
4903 static struct mdk_personality raid4_personality =
4907 .owner = THIS_MODULE,
4908 .make_request = make_request,
4912 .error_handler = error,
4913 .hot_add_disk = raid5_add_disk,
4914 .hot_remove_disk= raid5_remove_disk,
4915 .spare_active = raid5_spare_active,
4916 .sync_request = sync_request,
4917 .resize = raid5_resize,
4918 #ifdef CONFIG_MD_RAID5_RESHAPE
4919 .check_reshape = raid5_check_reshape,
4920 .start_reshape = raid5_start_reshape,
4922 .quiesce = raid5_quiesce,
4925 static int __init raid5_init(void)
4929 e = raid6_select_algo();
4932 register_md_personality(&raid6_personality);
4933 register_md_personality(&raid5_personality);
4934 register_md_personality(&raid4_personality);
4938 static void raid5_exit(void)
4940 unregister_md_personality(&raid6_personality);
4941 unregister_md_personality(&raid5_personality);
4942 unregister_md_personality(&raid4_personality);
4945 module_init(raid5_init);
4946 module_exit(raid5_exit);
4947 MODULE_LICENSE("GPL");
4948 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4949 MODULE_ALIAS("md-raid5");
4950 MODULE_ALIAS("md-raid4");
4951 MODULE_ALIAS("md-level-5");
4952 MODULE_ALIAS("md-level-4");
4953 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4954 MODULE_ALIAS("md-raid6");
4955 MODULE_ALIAS("md-level-6");
4957 /* This used to be two separate modules, they were: */
4958 MODULE_ALIAS("raid5");
4959 MODULE_ALIAS("raid6");