2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->bm_write is the number of the last batch successfully written.
31 * conf->bm_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is bm_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
54 #include <linux/raid/bitmap.h>
55 #include <linux/async_tx.h>
61 #define NR_STRIPES 256
62 #define STRIPE_SIZE PAGE_SIZE
63 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
65 #define IO_THRESHOLD 1
66 #define BYPASS_THRESHOLD 1
67 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
68 #define HASH_MASK (NR_HASH - 1)
70 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
72 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
73 * order without overlap. There may be several bio's per stripe+device, and
74 * a bio could span several devices.
75 * When walking this list for a particular stripe+device, we must never proceed
76 * beyond a bio that extends past this device, as the next bio might no longer
78 * This macro is used to determine the 'next' bio in the list, given the sector
79 * of the current stripe+device
81 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
83 * The following can be used to debug the driver
85 #define RAID5_PARANOIA 1
86 #if RAID5_PARANOIA && defined(CONFIG_SMP)
87 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
89 # define CHECK_DEVLOCK()
97 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
99 #if !RAID6_USE_EMPTY_ZERO_PAGE
100 /* In .bss so it's zeroed */
101 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
104 static inline int raid6_next_disk(int disk, int raid_disks)
107 return (disk < raid_disks) ? disk : 0;
110 static void return_io(struct bio *return_bi)
112 struct bio *bi = return_bi;
115 return_bi = bi->bi_next;
123 static void print_raid5_conf (raid5_conf_t *conf);
125 static 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(struct request_queue *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);
377 sh->ops.count -= ack;
378 if (unlikely(sh->ops.count < 0)) {
379 printk(KERN_ERR "pending: %#lx ops.pending: %#lx ops.ack: %#lx "
380 "ops.complete: %#lx\n", pending, sh->ops.pending,
381 sh->ops.ack, sh->ops.complete);
389 raid5_end_read_request(struct bio *bi, int error);
391 raid5_end_write_request(struct bio *bi, int error);
393 static void ops_run_io(struct stripe_head *sh)
395 raid5_conf_t *conf = sh->raid_conf;
396 int i, disks = sh->disks;
400 for (i = disks; i--; ) {
404 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
406 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
411 bi = &sh->dev[i].req;
415 bi->bi_end_io = raid5_end_write_request;
417 bi->bi_end_io = raid5_end_read_request;
420 rdev = rcu_dereference(conf->disks[i].rdev);
421 if (rdev && test_bit(Faulty, &rdev->flags))
424 atomic_inc(&rdev->nr_pending);
428 if (test_bit(STRIPE_SYNCING, &sh->state) ||
429 test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
430 test_bit(STRIPE_EXPAND_READY, &sh->state))
431 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
433 set_bit(STRIPE_IO_STARTED, &sh->state);
435 bi->bi_bdev = rdev->bdev;
436 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
437 __func__, (unsigned long long)sh->sector,
439 atomic_inc(&sh->count);
440 bi->bi_sector = sh->sector + rdev->data_offset;
441 bi->bi_flags = 1 << BIO_UPTODATE;
445 bi->bi_io_vec = &sh->dev[i].vec;
446 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
447 bi->bi_io_vec[0].bv_offset = 0;
448 bi->bi_size = STRIPE_SIZE;
451 test_bit(R5_ReWrite, &sh->dev[i].flags))
452 atomic_add(STRIPE_SECTORS,
453 &rdev->corrected_errors);
454 generic_make_request(bi);
457 set_bit(STRIPE_DEGRADED, &sh->state);
458 pr_debug("skip op %ld on disc %d for sector %llu\n",
459 bi->bi_rw, i, (unsigned long long)sh->sector);
460 clear_bit(R5_LOCKED, &sh->dev[i].flags);
461 set_bit(STRIPE_HANDLE, &sh->state);
466 static struct dma_async_tx_descriptor *
467 async_copy_data(int frombio, struct bio *bio, struct page *page,
468 sector_t sector, struct dma_async_tx_descriptor *tx)
471 struct page *bio_page;
475 if (bio->bi_sector >= sector)
476 page_offset = (signed)(bio->bi_sector - sector) * 512;
478 page_offset = (signed)(sector - bio->bi_sector) * -512;
479 bio_for_each_segment(bvl, bio, i) {
480 int len = bio_iovec_idx(bio, i)->bv_len;
484 if (page_offset < 0) {
485 b_offset = -page_offset;
486 page_offset += b_offset;
490 if (len > 0 && page_offset + len > STRIPE_SIZE)
491 clen = STRIPE_SIZE - page_offset;
496 b_offset += bio_iovec_idx(bio, i)->bv_offset;
497 bio_page = bio_iovec_idx(bio, i)->bv_page;
499 tx = async_memcpy(page, bio_page, page_offset,
504 tx = async_memcpy(bio_page, page, b_offset,
509 if (clen < len) /* hit end of page */
517 static void ops_complete_biofill(void *stripe_head_ref)
519 struct stripe_head *sh = stripe_head_ref;
520 struct bio *return_bi = NULL;
521 raid5_conf_t *conf = sh->raid_conf;
524 pr_debug("%s: stripe %llu\n", __func__,
525 (unsigned long long)sh->sector);
527 /* clear completed biofills */
528 for (i = sh->disks; i--; ) {
529 struct r5dev *dev = &sh->dev[i];
531 /* acknowledge completion of a biofill operation */
532 /* and check if we need to reply to a read request,
533 * new R5_Wantfill requests are held off until
534 * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
536 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
537 struct bio *rbi, *rbi2;
539 /* The access to dev->read is outside of the
540 * spin_lock_irq(&conf->device_lock), but is protected
541 * by the STRIPE_OP_BIOFILL pending bit
546 while (rbi && rbi->bi_sector <
547 dev->sector + STRIPE_SECTORS) {
548 rbi2 = r5_next_bio(rbi, dev->sector);
549 spin_lock_irq(&conf->device_lock);
550 if (--rbi->bi_phys_segments == 0) {
551 rbi->bi_next = return_bi;
554 spin_unlock_irq(&conf->device_lock);
559 set_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
561 return_io(return_bi);
563 set_bit(STRIPE_HANDLE, &sh->state);
567 static void ops_run_biofill(struct stripe_head *sh)
569 struct dma_async_tx_descriptor *tx = NULL;
570 raid5_conf_t *conf = sh->raid_conf;
573 pr_debug("%s: stripe %llu\n", __func__,
574 (unsigned long long)sh->sector);
576 for (i = sh->disks; i--; ) {
577 struct r5dev *dev = &sh->dev[i];
578 if (test_bit(R5_Wantfill, &dev->flags)) {
580 spin_lock_irq(&conf->device_lock);
581 dev->read = rbi = dev->toread;
583 spin_unlock_irq(&conf->device_lock);
584 while (rbi && rbi->bi_sector <
585 dev->sector + STRIPE_SECTORS) {
586 tx = async_copy_data(0, rbi, dev->page,
588 rbi = r5_next_bio(rbi, dev->sector);
593 atomic_inc(&sh->count);
594 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
595 ops_complete_biofill, sh);
598 static void ops_complete_compute5(void *stripe_head_ref)
600 struct stripe_head *sh = stripe_head_ref;
601 int target = sh->ops.target;
602 struct r5dev *tgt = &sh->dev[target];
604 pr_debug("%s: stripe %llu\n", __func__,
605 (unsigned long long)sh->sector);
607 set_bit(R5_UPTODATE, &tgt->flags);
608 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
609 clear_bit(R5_Wantcompute, &tgt->flags);
610 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
611 set_bit(STRIPE_HANDLE, &sh->state);
615 static struct dma_async_tx_descriptor *
616 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
618 /* kernel stack size limits the total number of disks */
619 int disks = sh->disks;
620 struct page *xor_srcs[disks];
621 int target = sh->ops.target;
622 struct r5dev *tgt = &sh->dev[target];
623 struct page *xor_dest = tgt->page;
625 struct dma_async_tx_descriptor *tx;
628 pr_debug("%s: stripe %llu block: %d\n",
629 __func__, (unsigned long long)sh->sector, target);
630 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
632 for (i = disks; i--; )
634 xor_srcs[count++] = sh->dev[i].page;
636 atomic_inc(&sh->count);
638 if (unlikely(count == 1))
639 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
640 0, NULL, ops_complete_compute5, sh);
642 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
643 ASYNC_TX_XOR_ZERO_DST, NULL,
644 ops_complete_compute5, sh);
646 /* ack now if postxor is not set to be run */
647 if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
653 static void ops_complete_prexor(void *stripe_head_ref)
655 struct stripe_head *sh = stripe_head_ref;
657 pr_debug("%s: stripe %llu\n", __func__,
658 (unsigned long long)sh->sector);
660 set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
663 static struct dma_async_tx_descriptor *
664 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
666 /* kernel stack size limits the total number of disks */
667 int disks = sh->disks;
668 struct page *xor_srcs[disks];
669 int count = 0, pd_idx = sh->pd_idx, i;
671 /* existing parity data subtracted */
672 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
674 pr_debug("%s: stripe %llu\n", __func__,
675 (unsigned long long)sh->sector);
677 for (i = disks; i--; ) {
678 struct r5dev *dev = &sh->dev[i];
679 /* Only process blocks that are known to be uptodate */
680 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
681 xor_srcs[count++] = dev->page;
684 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
685 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
686 ops_complete_prexor, sh);
691 static struct dma_async_tx_descriptor *
692 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
693 unsigned long pending)
695 int disks = sh->disks;
696 int pd_idx = sh->pd_idx, i;
698 /* check if prexor is active which means only process blocks
699 * that are part of a read-modify-write (Wantprexor)
701 int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
703 pr_debug("%s: stripe %llu\n", __func__,
704 (unsigned long long)sh->sector);
706 for (i = disks; i--; ) {
707 struct r5dev *dev = &sh->dev[i];
712 if (prexor) { /* rmw */
714 test_bit(R5_Wantprexor, &dev->flags))
717 if (i != pd_idx && dev->towrite &&
718 test_bit(R5_LOCKED, &dev->flags))
725 spin_lock(&sh->lock);
726 chosen = dev->towrite;
728 BUG_ON(dev->written);
729 wbi = dev->written = chosen;
730 spin_unlock(&sh->lock);
732 while (wbi && wbi->bi_sector <
733 dev->sector + STRIPE_SECTORS) {
734 tx = async_copy_data(1, wbi, dev->page,
736 wbi = r5_next_bio(wbi, dev->sector);
744 static void ops_complete_postxor(void *stripe_head_ref)
746 struct stripe_head *sh = stripe_head_ref;
748 pr_debug("%s: stripe %llu\n", __func__,
749 (unsigned long long)sh->sector);
751 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
752 set_bit(STRIPE_HANDLE, &sh->state);
756 static void ops_complete_write(void *stripe_head_ref)
758 struct stripe_head *sh = stripe_head_ref;
759 int disks = sh->disks, i, pd_idx = sh->pd_idx;
761 pr_debug("%s: stripe %llu\n", __func__,
762 (unsigned long long)sh->sector);
764 for (i = disks; i--; ) {
765 struct r5dev *dev = &sh->dev[i];
766 if (dev->written || i == pd_idx)
767 set_bit(R5_UPTODATE, &dev->flags);
770 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
771 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
773 set_bit(STRIPE_HANDLE, &sh->state);
778 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
779 unsigned long pending)
781 /* kernel stack size limits the total number of disks */
782 int disks = sh->disks;
783 struct page *xor_srcs[disks];
785 int count = 0, pd_idx = sh->pd_idx, i;
786 struct page *xor_dest;
787 int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
789 dma_async_tx_callback callback;
791 pr_debug("%s: stripe %llu\n", __func__,
792 (unsigned long long)sh->sector);
794 /* check if prexor is active which means only process blocks
795 * that are part of a read-modify-write (written)
798 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
799 for (i = disks; i--; ) {
800 struct r5dev *dev = &sh->dev[i];
802 xor_srcs[count++] = dev->page;
805 xor_dest = sh->dev[pd_idx].page;
806 for (i = disks; i--; ) {
807 struct r5dev *dev = &sh->dev[i];
809 xor_srcs[count++] = dev->page;
813 /* check whether this postxor is part of a write */
814 callback = test_bit(STRIPE_OP_BIODRAIN, &pending) ?
815 ops_complete_write : ops_complete_postxor;
817 /* 1/ if we prexor'd then the dest is reused as a source
818 * 2/ if we did not prexor then we are redoing the parity
819 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
820 * for the synchronous xor case
822 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
823 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
825 atomic_inc(&sh->count);
827 if (unlikely(count == 1)) {
828 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
829 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
830 flags, tx, callback, sh);
832 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
833 flags, tx, callback, sh);
836 static void ops_complete_check(void *stripe_head_ref)
838 struct stripe_head *sh = stripe_head_ref;
840 pr_debug("%s: stripe %llu\n", __func__,
841 (unsigned long long)sh->sector);
843 set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
844 set_bit(STRIPE_HANDLE, &sh->state);
848 static void ops_run_check(struct stripe_head *sh)
850 /* kernel stack size limits the total number of disks */
851 int disks = sh->disks;
852 struct page *xor_srcs[disks];
853 struct dma_async_tx_descriptor *tx;
855 int count = 0, pd_idx = sh->pd_idx, i;
856 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
858 pr_debug("%s: stripe %llu\n", __func__,
859 (unsigned long long)sh->sector);
861 for (i = disks; i--; ) {
862 struct r5dev *dev = &sh->dev[i];
864 xor_srcs[count++] = dev->page;
867 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
868 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
870 atomic_inc(&sh->count);
871 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
872 ops_complete_check, sh);
875 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
877 int overlap_clear = 0, i, disks = sh->disks;
878 struct dma_async_tx_descriptor *tx = NULL;
880 if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
885 if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
886 tx = ops_run_compute5(sh, pending);
888 if (test_bit(STRIPE_OP_PREXOR, &pending))
889 tx = ops_run_prexor(sh, tx);
891 if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
892 tx = ops_run_biodrain(sh, tx, pending);
896 if (test_bit(STRIPE_OP_POSTXOR, &pending))
897 ops_run_postxor(sh, tx, pending);
899 if (test_bit(STRIPE_OP_CHECK, &pending))
903 for (i = disks; i--; ) {
904 struct r5dev *dev = &sh->dev[i];
905 if (test_and_clear_bit(R5_Overlap, &dev->flags))
906 wake_up(&sh->raid_conf->wait_for_overlap);
910 static int grow_one_stripe(raid5_conf_t *conf)
912 struct stripe_head *sh;
913 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
916 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
917 sh->raid_conf = conf;
918 spin_lock_init(&sh->lock);
920 if (grow_buffers(sh, conf->raid_disks)) {
921 shrink_buffers(sh, conf->raid_disks);
922 kmem_cache_free(conf->slab_cache, sh);
925 sh->disks = conf->raid_disks;
926 /* we just created an active stripe so... */
927 atomic_set(&sh->count, 1);
928 atomic_inc(&conf->active_stripes);
929 INIT_LIST_HEAD(&sh->lru);
934 static int grow_stripes(raid5_conf_t *conf, int num)
936 struct kmem_cache *sc;
937 int devs = conf->raid_disks;
939 sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
940 sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
941 conf->active_name = 0;
942 sc = kmem_cache_create(conf->cache_name[conf->active_name],
943 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
947 conf->slab_cache = sc;
948 conf->pool_size = devs;
950 if (!grow_one_stripe(conf))
955 #ifdef CONFIG_MD_RAID5_RESHAPE
956 static int resize_stripes(raid5_conf_t *conf, int newsize)
958 /* Make all the stripes able to hold 'newsize' devices.
959 * New slots in each stripe get 'page' set to a new page.
961 * This happens in stages:
962 * 1/ create a new kmem_cache and allocate the required number of
964 * 2/ gather all the old stripe_heads and tranfer the pages across
965 * to the new stripe_heads. This will have the side effect of
966 * freezing the array as once all stripe_heads have been collected,
967 * no IO will be possible. Old stripe heads are freed once their
968 * pages have been transferred over, and the old kmem_cache is
969 * freed when all stripes are done.
970 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
971 * we simple return a failre status - no need to clean anything up.
972 * 4/ allocate new pages for the new slots in the new stripe_heads.
973 * If this fails, we don't bother trying the shrink the
974 * stripe_heads down again, we just leave them as they are.
975 * As each stripe_head is processed the new one is released into
978 * Once step2 is started, we cannot afford to wait for a write,
979 * so we use GFP_NOIO allocations.
981 struct stripe_head *osh, *nsh;
982 LIST_HEAD(newstripes);
983 struct disk_info *ndisks;
985 struct kmem_cache *sc;
988 if (newsize <= conf->pool_size)
989 return 0; /* never bother to shrink */
991 md_allow_write(conf->mddev);
994 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
995 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1000 for (i = conf->max_nr_stripes; i; i--) {
1001 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1005 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1007 nsh->raid_conf = conf;
1008 spin_lock_init(&nsh->lock);
1010 list_add(&nsh->lru, &newstripes);
1013 /* didn't get enough, give up */
1014 while (!list_empty(&newstripes)) {
1015 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1016 list_del(&nsh->lru);
1017 kmem_cache_free(sc, nsh);
1019 kmem_cache_destroy(sc);
1022 /* Step 2 - Must use GFP_NOIO now.
1023 * OK, we have enough stripes, start collecting inactive
1024 * stripes and copying them over
1026 list_for_each_entry(nsh, &newstripes, lru) {
1027 spin_lock_irq(&conf->device_lock);
1028 wait_event_lock_irq(conf->wait_for_stripe,
1029 !list_empty(&conf->inactive_list),
1031 unplug_slaves(conf->mddev)
1033 osh = get_free_stripe(conf);
1034 spin_unlock_irq(&conf->device_lock);
1035 atomic_set(&nsh->count, 1);
1036 for(i=0; i<conf->pool_size; i++)
1037 nsh->dev[i].page = osh->dev[i].page;
1038 for( ; i<newsize; i++)
1039 nsh->dev[i].page = NULL;
1040 kmem_cache_free(conf->slab_cache, osh);
1042 kmem_cache_destroy(conf->slab_cache);
1045 * At this point, we are holding all the stripes so the array
1046 * is completely stalled, so now is a good time to resize
1049 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1051 for (i=0; i<conf->raid_disks; i++)
1052 ndisks[i] = conf->disks[i];
1054 conf->disks = ndisks;
1058 /* Step 4, return new stripes to service */
1059 while(!list_empty(&newstripes)) {
1060 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1061 list_del_init(&nsh->lru);
1062 for (i=conf->raid_disks; i < newsize; i++)
1063 if (nsh->dev[i].page == NULL) {
1064 struct page *p = alloc_page(GFP_NOIO);
1065 nsh->dev[i].page = p;
1069 release_stripe(nsh);
1071 /* critical section pass, GFP_NOIO no longer needed */
1073 conf->slab_cache = sc;
1074 conf->active_name = 1-conf->active_name;
1075 conf->pool_size = newsize;
1080 static int drop_one_stripe(raid5_conf_t *conf)
1082 struct stripe_head *sh;
1084 spin_lock_irq(&conf->device_lock);
1085 sh = get_free_stripe(conf);
1086 spin_unlock_irq(&conf->device_lock);
1089 BUG_ON(atomic_read(&sh->count));
1090 shrink_buffers(sh, conf->pool_size);
1091 kmem_cache_free(conf->slab_cache, sh);
1092 atomic_dec(&conf->active_stripes);
1096 static void shrink_stripes(raid5_conf_t *conf)
1098 while (drop_one_stripe(conf))
1101 if (conf->slab_cache)
1102 kmem_cache_destroy(conf->slab_cache);
1103 conf->slab_cache = NULL;
1106 static void raid5_end_read_request(struct bio * bi, int error)
1108 struct stripe_head *sh = bi->bi_private;
1109 raid5_conf_t *conf = sh->raid_conf;
1110 int disks = sh->disks, i;
1111 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1112 char b[BDEVNAME_SIZE];
1116 for (i=0 ; i<disks; i++)
1117 if (bi == &sh->dev[i].req)
1120 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1121 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1129 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1130 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1131 rdev = conf->disks[i].rdev;
1132 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1133 " (%lu sectors at %llu on %s)\n",
1134 mdname(conf->mddev), STRIPE_SECTORS,
1135 (unsigned long long)(sh->sector
1136 + rdev->data_offset),
1137 bdevname(rdev->bdev, b));
1138 clear_bit(R5_ReadError, &sh->dev[i].flags);
1139 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1141 if (atomic_read(&conf->disks[i].rdev->read_errors))
1142 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1144 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1146 rdev = conf->disks[i].rdev;
1148 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1149 atomic_inc(&rdev->read_errors);
1150 if (conf->mddev->degraded)
1151 printk_rl(KERN_WARNING
1152 "raid5:%s: read error not correctable "
1153 "(sector %llu on %s).\n",
1154 mdname(conf->mddev),
1155 (unsigned long long)(sh->sector
1156 + rdev->data_offset),
1158 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1160 printk_rl(KERN_WARNING
1161 "raid5:%s: read error NOT corrected!! "
1162 "(sector %llu on %s).\n",
1163 mdname(conf->mddev),
1164 (unsigned long long)(sh->sector
1165 + rdev->data_offset),
1167 else if (atomic_read(&rdev->read_errors)
1168 > conf->max_nr_stripes)
1170 "raid5:%s: Too many read errors, failing device %s.\n",
1171 mdname(conf->mddev), bdn);
1175 set_bit(R5_ReadError, &sh->dev[i].flags);
1177 clear_bit(R5_ReadError, &sh->dev[i].flags);
1178 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1179 md_error(conf->mddev, rdev);
1182 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1183 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1184 set_bit(STRIPE_HANDLE, &sh->state);
1188 static void raid5_end_write_request (struct bio *bi, int error)
1190 struct stripe_head *sh = bi->bi_private;
1191 raid5_conf_t *conf = sh->raid_conf;
1192 int disks = sh->disks, i;
1193 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1195 for (i=0 ; i<disks; i++)
1196 if (bi == &sh->dev[i].req)
1199 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1200 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1208 md_error(conf->mddev, conf->disks[i].rdev);
1210 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1212 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1213 set_bit(STRIPE_HANDLE, &sh->state);
1218 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1220 static void raid5_build_block (struct stripe_head *sh, int i)
1222 struct r5dev *dev = &sh->dev[i];
1224 bio_init(&dev->req);
1225 dev->req.bi_io_vec = &dev->vec;
1227 dev->req.bi_max_vecs++;
1228 dev->vec.bv_page = dev->page;
1229 dev->vec.bv_len = STRIPE_SIZE;
1230 dev->vec.bv_offset = 0;
1232 dev->req.bi_sector = sh->sector;
1233 dev->req.bi_private = sh;
1236 dev->sector = compute_blocknr(sh, i);
1239 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1241 char b[BDEVNAME_SIZE];
1242 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1243 pr_debug("raid5: error called\n");
1245 if (!test_bit(Faulty, &rdev->flags)) {
1246 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1247 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1248 unsigned long flags;
1249 spin_lock_irqsave(&conf->device_lock, flags);
1251 spin_unlock_irqrestore(&conf->device_lock, flags);
1253 * if recovery was running, make sure it aborts.
1255 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1257 set_bit(Faulty, &rdev->flags);
1259 "raid5: Disk failure on %s, disabling device.\n"
1260 "raid5: Operation continuing on %d devices.\n",
1261 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1266 * Input: a 'big' sector number,
1267 * Output: index of the data and parity disk, and the sector # in them.
1269 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1270 unsigned int data_disks, unsigned int * dd_idx,
1271 unsigned int * pd_idx, raid5_conf_t *conf)
1274 unsigned long chunk_number;
1275 unsigned int chunk_offset;
1276 sector_t new_sector;
1277 int sectors_per_chunk = conf->chunk_size >> 9;
1279 /* First compute the information on this sector */
1282 * Compute the chunk number and the sector offset inside the chunk
1284 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1285 chunk_number = r_sector;
1286 BUG_ON(r_sector != chunk_number);
1289 * Compute the stripe number
1291 stripe = chunk_number / data_disks;
1294 * Compute the data disk and parity disk indexes inside the stripe
1296 *dd_idx = chunk_number % data_disks;
1299 * Select the parity disk based on the user selected algorithm.
1301 switch(conf->level) {
1303 *pd_idx = data_disks;
1306 switch (conf->algorithm) {
1307 case ALGORITHM_LEFT_ASYMMETRIC:
1308 *pd_idx = data_disks - stripe % raid_disks;
1309 if (*dd_idx >= *pd_idx)
1312 case ALGORITHM_RIGHT_ASYMMETRIC:
1313 *pd_idx = stripe % raid_disks;
1314 if (*dd_idx >= *pd_idx)
1317 case ALGORITHM_LEFT_SYMMETRIC:
1318 *pd_idx = data_disks - stripe % raid_disks;
1319 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1321 case ALGORITHM_RIGHT_SYMMETRIC:
1322 *pd_idx = stripe % raid_disks;
1323 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1326 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1332 /**** FIX THIS ****/
1333 switch (conf->algorithm) {
1334 case ALGORITHM_LEFT_ASYMMETRIC:
1335 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1336 if (*pd_idx == raid_disks-1)
1337 (*dd_idx)++; /* Q D D D P */
1338 else if (*dd_idx >= *pd_idx)
1339 (*dd_idx) += 2; /* D D P Q D */
1341 case ALGORITHM_RIGHT_ASYMMETRIC:
1342 *pd_idx = stripe % raid_disks;
1343 if (*pd_idx == raid_disks-1)
1344 (*dd_idx)++; /* Q D D D P */
1345 else if (*dd_idx >= *pd_idx)
1346 (*dd_idx) += 2; /* D D P Q D */
1348 case ALGORITHM_LEFT_SYMMETRIC:
1349 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1350 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1352 case ALGORITHM_RIGHT_SYMMETRIC:
1353 *pd_idx = stripe % raid_disks;
1354 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1357 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1364 * Finally, compute the new sector number
1366 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1371 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1373 raid5_conf_t *conf = sh->raid_conf;
1374 int raid_disks = sh->disks;
1375 int data_disks = raid_disks - conf->max_degraded;
1376 sector_t new_sector = sh->sector, check;
1377 int sectors_per_chunk = conf->chunk_size >> 9;
1380 int chunk_number, dummy1, dummy2, dd_idx = i;
1384 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1385 stripe = new_sector;
1386 BUG_ON(new_sector != stripe);
1388 if (i == sh->pd_idx)
1390 switch(conf->level) {
1393 switch (conf->algorithm) {
1394 case ALGORITHM_LEFT_ASYMMETRIC:
1395 case ALGORITHM_RIGHT_ASYMMETRIC:
1399 case ALGORITHM_LEFT_SYMMETRIC:
1400 case ALGORITHM_RIGHT_SYMMETRIC:
1403 i -= (sh->pd_idx + 1);
1406 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1411 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1412 return 0; /* It is the Q disk */
1413 switch (conf->algorithm) {
1414 case ALGORITHM_LEFT_ASYMMETRIC:
1415 case ALGORITHM_RIGHT_ASYMMETRIC:
1416 if (sh->pd_idx == raid_disks-1)
1417 i--; /* Q D D D P */
1418 else if (i > sh->pd_idx)
1419 i -= 2; /* D D P Q D */
1421 case ALGORITHM_LEFT_SYMMETRIC:
1422 case ALGORITHM_RIGHT_SYMMETRIC:
1423 if (sh->pd_idx == raid_disks-1)
1424 i--; /* Q D D D P */
1429 i -= (sh->pd_idx + 2);
1433 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1439 chunk_number = stripe * data_disks + i;
1440 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1442 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1443 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1444 printk(KERN_ERR "compute_blocknr: map not correct\n");
1453 * Copy data between a page in the stripe cache, and one or more bion
1454 * The page could align with the middle of the bio, or there could be
1455 * several bion, each with several bio_vecs, which cover part of the page
1456 * Multiple bion are linked together on bi_next. There may be extras
1457 * at the end of this list. We ignore them.
1459 static void copy_data(int frombio, struct bio *bio,
1463 char *pa = page_address(page);
1464 struct bio_vec *bvl;
1468 if (bio->bi_sector >= sector)
1469 page_offset = (signed)(bio->bi_sector - sector) * 512;
1471 page_offset = (signed)(sector - bio->bi_sector) * -512;
1472 bio_for_each_segment(bvl, bio, i) {
1473 int len = bio_iovec_idx(bio,i)->bv_len;
1477 if (page_offset < 0) {
1478 b_offset = -page_offset;
1479 page_offset += b_offset;
1483 if (len > 0 && page_offset + len > STRIPE_SIZE)
1484 clen = STRIPE_SIZE - page_offset;
1488 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1490 memcpy(pa+page_offset, ba+b_offset, clen);
1492 memcpy(ba+b_offset, pa+page_offset, clen);
1493 __bio_kunmap_atomic(ba, KM_USER0);
1495 if (clen < len) /* hit end of page */
1501 #define check_xor() do { \
1502 if (count == MAX_XOR_BLOCKS) { \
1503 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1508 static void compute_parity6(struct stripe_head *sh, int method)
1510 raid6_conf_t *conf = sh->raid_conf;
1511 int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1513 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1516 qd_idx = raid6_next_disk(pd_idx, disks);
1517 d0_idx = raid6_next_disk(qd_idx, disks);
1519 pr_debug("compute_parity, stripe %llu, method %d\n",
1520 (unsigned long long)sh->sector, method);
1523 case READ_MODIFY_WRITE:
1524 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1525 case RECONSTRUCT_WRITE:
1526 for (i= disks; i-- ;)
1527 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1528 chosen = sh->dev[i].towrite;
1529 sh->dev[i].towrite = NULL;
1531 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1532 wake_up(&conf->wait_for_overlap);
1534 BUG_ON(sh->dev[i].written);
1535 sh->dev[i].written = chosen;
1539 BUG(); /* Not implemented yet */
1542 for (i = disks; i--;)
1543 if (sh->dev[i].written) {
1544 sector_t sector = sh->dev[i].sector;
1545 struct bio *wbi = sh->dev[i].written;
1546 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1547 copy_data(1, wbi, sh->dev[i].page, sector);
1548 wbi = r5_next_bio(wbi, sector);
1551 set_bit(R5_LOCKED, &sh->dev[i].flags);
1552 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1556 // case RECONSTRUCT_WRITE:
1557 // case CHECK_PARITY:
1558 // case UPDATE_PARITY:
1559 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1560 /* FIX: Is this ordering of drives even remotely optimal? */
1564 ptrs[count++] = page_address(sh->dev[i].page);
1565 if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1566 printk("block %d/%d not uptodate on parity calc\n", i,count);
1567 i = raid6_next_disk(i, disks);
1568 } while ( i != d0_idx );
1572 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1575 case RECONSTRUCT_WRITE:
1576 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1577 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1578 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1579 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1582 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1583 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1589 /* Compute one missing block */
1590 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1592 int i, count, disks = sh->disks;
1593 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1594 int pd_idx = sh->pd_idx;
1595 int qd_idx = raid6_next_disk(pd_idx, disks);
1597 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1598 (unsigned long long)sh->sector, dd_idx);
1600 if ( dd_idx == qd_idx ) {
1601 /* We're actually computing the Q drive */
1602 compute_parity6(sh, UPDATE_PARITY);
1604 dest = page_address(sh->dev[dd_idx].page);
1605 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1607 for (i = disks ; i--; ) {
1608 if (i == dd_idx || i == qd_idx)
1610 p = page_address(sh->dev[i].page);
1611 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1614 printk("compute_block() %d, stripe %llu, %d"
1615 " not present\n", dd_idx,
1616 (unsigned long long)sh->sector, i);
1621 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1622 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1623 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1627 /* Compute two missing blocks */
1628 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1630 int i, count, disks = sh->disks;
1631 int pd_idx = sh->pd_idx;
1632 int qd_idx = raid6_next_disk(pd_idx, disks);
1633 int d0_idx = raid6_next_disk(qd_idx, disks);
1636 /* faila and failb are disk numbers relative to d0_idx */
1637 /* pd_idx become disks-2 and qd_idx become disks-1 */
1638 faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1639 failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1641 BUG_ON(faila == failb);
1642 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1644 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1645 (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1647 if ( failb == disks-1 ) {
1648 /* Q disk is one of the missing disks */
1649 if ( faila == disks-2 ) {
1650 /* Missing P+Q, just recompute */
1651 compute_parity6(sh, UPDATE_PARITY);
1654 /* We're missing D+Q; recompute D from P */
1655 compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1656 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1661 /* We're missing D+P or D+D; build pointer table */
1663 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1669 ptrs[count++] = page_address(sh->dev[i].page);
1670 i = raid6_next_disk(i, disks);
1671 if (i != dd_idx1 && i != dd_idx2 &&
1672 !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1673 printk("compute_2 with missing block %d/%d\n", count, i);
1674 } while ( i != d0_idx );
1676 if ( failb == disks-2 ) {
1677 /* We're missing D+P. */
1678 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1680 /* We're missing D+D. */
1681 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1684 /* Both the above update both missing blocks */
1685 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1686 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1691 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1693 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1697 /* if we are not expanding this is a proper write request, and
1698 * there will be bios with new data to be drained into the
1702 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1706 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1709 for (i = disks; i--; ) {
1710 struct r5dev *dev = &sh->dev[i];
1713 set_bit(R5_LOCKED, &dev->flags);
1715 clear_bit(R5_UPTODATE, &dev->flags);
1719 if (locked + 1 == disks)
1720 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1721 atomic_inc(&sh->raid_conf->pending_full_writes);
1723 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1724 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1726 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1727 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1728 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1732 for (i = disks; i--; ) {
1733 struct r5dev *dev = &sh->dev[i];
1737 /* For a read-modify write there may be blocks that are
1738 * locked for reading while others are ready to be
1739 * written so we distinguish these blocks by the
1743 (test_bit(R5_UPTODATE, &dev->flags) ||
1744 test_bit(R5_Wantcompute, &dev->flags))) {
1745 set_bit(R5_Wantprexor, &dev->flags);
1746 set_bit(R5_LOCKED, &dev->flags);
1747 clear_bit(R5_UPTODATE, &dev->flags);
1753 /* keep the parity disk locked while asynchronous operations
1756 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1757 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1760 pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1761 __func__, (unsigned long long)sh->sector,
1762 locked, sh->ops.pending);
1768 * Each stripe/dev can have one or more bion attached.
1769 * toread/towrite point to the first in a chain.
1770 * The bi_next chain must be in order.
1772 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1775 raid5_conf_t *conf = sh->raid_conf;
1778 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1779 (unsigned long long)bi->bi_sector,
1780 (unsigned long long)sh->sector);
1783 spin_lock(&sh->lock);
1784 spin_lock_irq(&conf->device_lock);
1786 bip = &sh->dev[dd_idx].towrite;
1787 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1790 bip = &sh->dev[dd_idx].toread;
1791 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1792 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1794 bip = & (*bip)->bi_next;
1796 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1799 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1803 bi->bi_phys_segments ++;
1804 spin_unlock_irq(&conf->device_lock);
1805 spin_unlock(&sh->lock);
1807 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1808 (unsigned long long)bi->bi_sector,
1809 (unsigned long long)sh->sector, dd_idx);
1811 if (conf->mddev->bitmap && firstwrite) {
1812 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1814 sh->bm_seq = conf->seq_flush+1;
1815 set_bit(STRIPE_BIT_DELAY, &sh->state);
1819 /* check if page is covered */
1820 sector_t sector = sh->dev[dd_idx].sector;
1821 for (bi=sh->dev[dd_idx].towrite;
1822 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1823 bi && bi->bi_sector <= sector;
1824 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1825 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1826 sector = bi->bi_sector + (bi->bi_size>>9);
1828 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1829 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1834 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1835 spin_unlock_irq(&conf->device_lock);
1836 spin_unlock(&sh->lock);
1840 static void end_reshape(raid5_conf_t *conf);
1842 static int page_is_zero(struct page *p)
1844 char *a = page_address(p);
1845 return ((*(u32*)a) == 0 &&
1846 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1849 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1851 int sectors_per_chunk = conf->chunk_size >> 9;
1853 int chunk_offset = sector_div(stripe, sectors_per_chunk);
1855 raid5_compute_sector(stripe * (disks - conf->max_degraded)
1856 *sectors_per_chunk + chunk_offset,
1857 disks, disks - conf->max_degraded,
1858 &dd_idx, &pd_idx, conf);
1863 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1864 struct stripe_head_state *s, int disks,
1865 struct bio **return_bi)
1868 for (i = disks; i--; ) {
1872 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1875 rdev = rcu_dereference(conf->disks[i].rdev);
1876 if (rdev && test_bit(In_sync, &rdev->flags))
1877 /* multiple read failures in one stripe */
1878 md_error(conf->mddev, rdev);
1881 spin_lock_irq(&conf->device_lock);
1882 /* fail all writes first */
1883 bi = sh->dev[i].towrite;
1884 sh->dev[i].towrite = NULL;
1890 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1891 wake_up(&conf->wait_for_overlap);
1893 while (bi && bi->bi_sector <
1894 sh->dev[i].sector + STRIPE_SECTORS) {
1895 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1896 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1897 if (--bi->bi_phys_segments == 0) {
1898 md_write_end(conf->mddev);
1899 bi->bi_next = *return_bi;
1904 /* and fail all 'written' */
1905 bi = sh->dev[i].written;
1906 sh->dev[i].written = NULL;
1907 if (bi) bitmap_end = 1;
1908 while (bi && bi->bi_sector <
1909 sh->dev[i].sector + STRIPE_SECTORS) {
1910 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1911 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1912 if (--bi->bi_phys_segments == 0) {
1913 md_write_end(conf->mddev);
1914 bi->bi_next = *return_bi;
1920 /* fail any reads if this device is non-operational and
1921 * the data has not reached the cache yet.
1923 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1924 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1925 test_bit(R5_ReadError, &sh->dev[i].flags))) {
1926 bi = sh->dev[i].toread;
1927 sh->dev[i].toread = NULL;
1928 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1929 wake_up(&conf->wait_for_overlap);
1930 if (bi) s->to_read--;
1931 while (bi && bi->bi_sector <
1932 sh->dev[i].sector + STRIPE_SECTORS) {
1933 struct bio *nextbi =
1934 r5_next_bio(bi, sh->dev[i].sector);
1935 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1936 if (--bi->bi_phys_segments == 0) {
1937 bi->bi_next = *return_bi;
1943 spin_unlock_irq(&conf->device_lock);
1945 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1946 STRIPE_SECTORS, 0, 0);
1949 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
1950 if (atomic_dec_and_test(&conf->pending_full_writes))
1951 md_wakeup_thread(conf->mddev->thread);
1954 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1957 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1958 struct stripe_head_state *s, int disk_idx, int disks)
1960 struct r5dev *dev = &sh->dev[disk_idx];
1961 struct r5dev *failed_dev = &sh->dev[s->failed_num];
1963 /* don't schedule compute operations or reads on the parity block while
1964 * a check is in flight
1966 if ((disk_idx == sh->pd_idx) &&
1967 test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1970 /* is the data in this block needed, and can we get it? */
1971 if (!test_bit(R5_LOCKED, &dev->flags) &&
1972 !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1973 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1974 s->syncing || s->expanding || (s->failed &&
1975 (failed_dev->toread || (failed_dev->towrite &&
1976 !test_bit(R5_OVERWRITE, &failed_dev->flags)
1978 /* 1/ We would like to get this block, possibly by computing it,
1979 * but we might not be able to.
1981 * 2/ Since parity check operations potentially make the parity
1982 * block !uptodate it will need to be refreshed before any
1983 * compute operations on data disks are scheduled.
1985 * 3/ We hold off parity block re-reads until check operations
1988 if ((s->uptodate == disks - 1) &&
1989 (s->failed && disk_idx == s->failed_num) &&
1990 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1991 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1992 set_bit(R5_Wantcompute, &dev->flags);
1993 sh->ops.target = disk_idx;
1996 /* Careful: from this point on 'uptodate' is in the eye
1997 * of raid5_run_ops which services 'compute' operations
1998 * before writes. R5_Wantcompute flags a block that will
1999 * be R5_UPTODATE by the time it is needed for a
2000 * subsequent operation.
2003 return 0; /* uptodate + compute == disks */
2004 } else if ((s->uptodate < disks - 1) &&
2005 test_bit(R5_Insync, &dev->flags)) {
2006 /* Note: we hold off compute operations while checks are
2007 * in flight, but we still prefer 'compute' over 'read'
2008 * hence we only read if (uptodate < * disks-1)
2010 set_bit(R5_LOCKED, &dev->flags);
2011 set_bit(R5_Wantread, &dev->flags);
2013 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2021 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2022 struct stripe_head_state *s, int disks)
2026 /* Clear completed compute operations. Parity recovery
2027 * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2028 * later on in this routine
2030 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2031 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2032 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2033 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2034 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2037 /* look for blocks to read/compute, skip this if a compute
2038 * is already in flight, or if the stripe contents are in the
2039 * midst of changing due to a write
2041 if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2042 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2043 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2044 for (i = disks; i--; )
2045 if (__handle_issuing_new_read_requests5(
2046 sh, s, i, disks) == 0)
2049 set_bit(STRIPE_HANDLE, &sh->state);
2052 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2053 struct stripe_head_state *s, struct r6_state *r6s,
2057 for (i = disks; i--; ) {
2058 struct r5dev *dev = &sh->dev[i];
2059 if (!test_bit(R5_LOCKED, &dev->flags) &&
2060 !test_bit(R5_UPTODATE, &dev->flags) &&
2061 (dev->toread || (dev->towrite &&
2062 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2063 s->syncing || s->expanding ||
2065 (sh->dev[r6s->failed_num[0]].toread ||
2068 (sh->dev[r6s->failed_num[1]].toread ||
2070 /* we would like to get this block, possibly
2071 * by computing it, but we might not be able to
2073 if ((s->uptodate == disks - 1) &&
2074 (s->failed && (i == r6s->failed_num[0] ||
2075 i == r6s->failed_num[1]))) {
2076 pr_debug("Computing stripe %llu block %d\n",
2077 (unsigned long long)sh->sector, i);
2078 compute_block_1(sh, i, 0);
2080 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2081 /* Computing 2-failure is *very* expensive; only
2082 * do it if failed >= 2
2085 for (other = disks; other--; ) {
2088 if (!test_bit(R5_UPTODATE,
2089 &sh->dev[other].flags))
2093 pr_debug("Computing stripe %llu blocks %d,%d\n",
2094 (unsigned long long)sh->sector,
2096 compute_block_2(sh, i, other);
2098 } else if (test_bit(R5_Insync, &dev->flags)) {
2099 set_bit(R5_LOCKED, &dev->flags);
2100 set_bit(R5_Wantread, &dev->flags);
2102 pr_debug("Reading block %d (sync=%d)\n",
2107 set_bit(STRIPE_HANDLE, &sh->state);
2111 /* handle_completed_write_requests
2112 * any written block on an uptodate or failed drive can be returned.
2113 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2114 * never LOCKED, so we don't need to test 'failed' directly.
2116 static void handle_completed_write_requests(raid5_conf_t *conf,
2117 struct stripe_head *sh, int disks, struct bio **return_bi)
2122 for (i = disks; i--; )
2123 if (sh->dev[i].written) {
2125 if (!test_bit(R5_LOCKED, &dev->flags) &&
2126 test_bit(R5_UPTODATE, &dev->flags)) {
2127 /* We can return any write requests */
2128 struct bio *wbi, *wbi2;
2130 pr_debug("Return write for disc %d\n", i);
2131 spin_lock_irq(&conf->device_lock);
2133 dev->written = NULL;
2134 while (wbi && wbi->bi_sector <
2135 dev->sector + STRIPE_SECTORS) {
2136 wbi2 = r5_next_bio(wbi, dev->sector);
2137 if (--wbi->bi_phys_segments == 0) {
2138 md_write_end(conf->mddev);
2139 wbi->bi_next = *return_bi;
2144 if (dev->towrite == NULL)
2146 spin_unlock_irq(&conf->device_lock);
2148 bitmap_endwrite(conf->mddev->bitmap,
2151 !test_bit(STRIPE_DEGRADED, &sh->state),
2156 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2157 if (atomic_dec_and_test(&conf->pending_full_writes))
2158 md_wakeup_thread(conf->mddev->thread);
2161 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2162 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2164 int rmw = 0, rcw = 0, i;
2165 for (i = disks; i--; ) {
2166 /* would I have to read this buffer for read_modify_write */
2167 struct r5dev *dev = &sh->dev[i];
2168 if ((dev->towrite || i == sh->pd_idx) &&
2169 !test_bit(R5_LOCKED, &dev->flags) &&
2170 !(test_bit(R5_UPTODATE, &dev->flags) ||
2171 test_bit(R5_Wantcompute, &dev->flags))) {
2172 if (test_bit(R5_Insync, &dev->flags))
2175 rmw += 2*disks; /* cannot read it */
2177 /* Would I have to read this buffer for reconstruct_write */
2178 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2179 !test_bit(R5_LOCKED, &dev->flags) &&
2180 !(test_bit(R5_UPTODATE, &dev->flags) ||
2181 test_bit(R5_Wantcompute, &dev->flags))) {
2182 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2187 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2188 (unsigned long long)sh->sector, rmw, rcw);
2189 set_bit(STRIPE_HANDLE, &sh->state);
2190 if (rmw < rcw && rmw > 0)
2191 /* prefer read-modify-write, but need to get some data */
2192 for (i = disks; i--; ) {
2193 struct r5dev *dev = &sh->dev[i];
2194 if ((dev->towrite || i == sh->pd_idx) &&
2195 !test_bit(R5_LOCKED, &dev->flags) &&
2196 !(test_bit(R5_UPTODATE, &dev->flags) ||
2197 test_bit(R5_Wantcompute, &dev->flags)) &&
2198 test_bit(R5_Insync, &dev->flags)) {
2200 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2201 pr_debug("Read_old block "
2202 "%d for r-m-w\n", i);
2203 set_bit(R5_LOCKED, &dev->flags);
2204 set_bit(R5_Wantread, &dev->flags);
2207 set_bit(STRIPE_DELAYED, &sh->state);
2208 set_bit(STRIPE_HANDLE, &sh->state);
2212 if (rcw <= rmw && rcw > 0)
2213 /* want reconstruct write, but need to get some data */
2214 for (i = disks; i--; ) {
2215 struct r5dev *dev = &sh->dev[i];
2216 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2218 !test_bit(R5_LOCKED, &dev->flags) &&
2219 !(test_bit(R5_UPTODATE, &dev->flags) ||
2220 test_bit(R5_Wantcompute, &dev->flags)) &&
2221 test_bit(R5_Insync, &dev->flags)) {
2223 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2224 pr_debug("Read_old block "
2225 "%d for Reconstruct\n", i);
2226 set_bit(R5_LOCKED, &dev->flags);
2227 set_bit(R5_Wantread, &dev->flags);
2230 set_bit(STRIPE_DELAYED, &sh->state);
2231 set_bit(STRIPE_HANDLE, &sh->state);
2235 /* now if nothing is locked, and if we have enough data,
2236 * we can start a write request
2238 /* since handle_stripe can be called at any time we need to handle the
2239 * case where a compute block operation has been submitted and then a
2240 * subsequent call wants to start a write request. raid5_run_ops only
2241 * handles the case where compute block and postxor are requested
2242 * simultaneously. If this is not the case then new writes need to be
2243 * held off until the compute completes.
2245 if ((s->req_compute ||
2246 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2247 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2248 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2249 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2252 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2253 struct stripe_head *sh, struct stripe_head_state *s,
2254 struct r6_state *r6s, int disks)
2256 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2257 int qd_idx = r6s->qd_idx;
2258 for (i = disks; i--; ) {
2259 struct r5dev *dev = &sh->dev[i];
2260 /* Would I have to read this buffer for reconstruct_write */
2261 if (!test_bit(R5_OVERWRITE, &dev->flags)
2262 && i != pd_idx && i != qd_idx
2263 && (!test_bit(R5_LOCKED, &dev->flags)
2265 !test_bit(R5_UPTODATE, &dev->flags)) {
2266 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2268 pr_debug("raid6: must_compute: "
2269 "disk %d flags=%#lx\n", i, dev->flags);
2274 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2275 (unsigned long long)sh->sector, rcw, must_compute);
2276 set_bit(STRIPE_HANDLE, &sh->state);
2279 /* want reconstruct write, but need to get some data */
2280 for (i = disks; i--; ) {
2281 struct r5dev *dev = &sh->dev[i];
2282 if (!test_bit(R5_OVERWRITE, &dev->flags)
2283 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2284 && !test_bit(R5_LOCKED, &dev->flags) &&
2285 !test_bit(R5_UPTODATE, &dev->flags) &&
2286 test_bit(R5_Insync, &dev->flags)) {
2288 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2289 pr_debug("Read_old stripe %llu "
2290 "block %d for Reconstruct\n",
2291 (unsigned long long)sh->sector, i);
2292 set_bit(R5_LOCKED, &dev->flags);
2293 set_bit(R5_Wantread, &dev->flags);
2296 pr_debug("Request delayed stripe %llu "
2297 "block %d for Reconstruct\n",
2298 (unsigned long long)sh->sector, i);
2299 set_bit(STRIPE_DELAYED, &sh->state);
2300 set_bit(STRIPE_HANDLE, &sh->state);
2304 /* now if nothing is locked, and if we have enough data, we can start a
2307 if (s->locked == 0 && rcw == 0 &&
2308 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2309 if (must_compute > 0) {
2310 /* We have failed blocks and need to compute them */
2311 switch (s->failed) {
2315 compute_block_1(sh, r6s->failed_num[0], 0);
2318 compute_block_2(sh, r6s->failed_num[0],
2319 r6s->failed_num[1]);
2321 default: /* This request should have been failed? */
2326 pr_debug("Computing parity for stripe %llu\n",
2327 (unsigned long long)sh->sector);
2328 compute_parity6(sh, RECONSTRUCT_WRITE);
2329 /* now every locked buffer is ready to be written */
2330 for (i = disks; i--; )
2331 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2332 pr_debug("Writing stripe %llu block %d\n",
2333 (unsigned long long)sh->sector, i);
2335 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2337 if (s->locked == disks)
2338 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2339 atomic_inc(&conf->pending_full_writes);
2340 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2341 set_bit(STRIPE_INSYNC, &sh->state);
2343 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2344 atomic_dec(&conf->preread_active_stripes);
2345 if (atomic_read(&conf->preread_active_stripes) <
2347 md_wakeup_thread(conf->mddev->thread);
2352 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2353 struct stripe_head_state *s, int disks)
2355 int canceled_check = 0;
2357 set_bit(STRIPE_HANDLE, &sh->state);
2359 /* complete a check operation */
2360 if (test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2361 clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2362 clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2363 if (s->failed == 0) {
2364 if (sh->ops.zero_sum_result == 0)
2365 /* parity is correct (on disc,
2366 * not in buffer any more)
2368 set_bit(STRIPE_INSYNC, &sh->state);
2370 conf->mddev->resync_mismatches +=
2373 MD_RECOVERY_CHECK, &conf->mddev->recovery))
2374 /* don't try to repair!! */
2375 set_bit(STRIPE_INSYNC, &sh->state);
2377 set_bit(STRIPE_OP_COMPUTE_BLK,
2379 set_bit(STRIPE_OP_MOD_REPAIR_PD,
2381 set_bit(R5_Wantcompute,
2382 &sh->dev[sh->pd_idx].flags);
2383 sh->ops.target = sh->pd_idx;
2389 canceled_check = 1; /* STRIPE_INSYNC is not set */
2392 /* start a new check operation if there are no failures, the stripe is
2393 * not insync, and a repair is not in flight
2395 if (s->failed == 0 &&
2396 !test_bit(STRIPE_INSYNC, &sh->state) &&
2397 !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2398 if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2399 BUG_ON(s->uptodate != disks);
2400 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2406 /* check if we can clear a parity disk reconstruct */
2407 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2408 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2410 clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2411 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2412 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2413 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2417 /* Wait for check parity and compute block operations to complete
2418 * before write-back. If a failure occurred while the check operation
2419 * was in flight we need to cycle this stripe through handle_stripe
2420 * since the parity block may not be uptodate
2422 if (!canceled_check && !test_bit(STRIPE_INSYNC, &sh->state) &&
2423 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2424 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2426 /* either failed parity check, or recovery is happening */
2428 s->failed_num = sh->pd_idx;
2429 dev = &sh->dev[s->failed_num];
2430 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2431 BUG_ON(s->uptodate != disks);
2433 set_bit(R5_LOCKED, &dev->flags);
2434 set_bit(R5_Wantwrite, &dev->flags);
2436 clear_bit(STRIPE_DEGRADED, &sh->state);
2438 set_bit(STRIPE_INSYNC, &sh->state);
2443 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2444 struct stripe_head_state *s,
2445 struct r6_state *r6s, struct page *tmp_page,
2448 int update_p = 0, update_q = 0;
2450 int pd_idx = sh->pd_idx;
2451 int qd_idx = r6s->qd_idx;
2453 set_bit(STRIPE_HANDLE, &sh->state);
2455 BUG_ON(s->failed > 2);
2456 BUG_ON(s->uptodate < disks);
2457 /* Want to check and possibly repair P and Q.
2458 * However there could be one 'failed' device, in which
2459 * case we can only check one of them, possibly using the
2460 * other to generate missing data
2463 /* If !tmp_page, we cannot do the calculations,
2464 * but as we have set STRIPE_HANDLE, we will soon be called
2465 * by stripe_handle with a tmp_page - just wait until then.
2468 if (s->failed == r6s->q_failed) {
2469 /* The only possible failed device holds 'Q', so it
2470 * makes sense to check P (If anything else were failed,
2471 * we would have used P to recreate it).
2473 compute_block_1(sh, pd_idx, 1);
2474 if (!page_is_zero(sh->dev[pd_idx].page)) {
2475 compute_block_1(sh, pd_idx, 0);
2479 if (!r6s->q_failed && s->failed < 2) {
2480 /* q is not failed, and we didn't use it to generate
2481 * anything, so it makes sense to check it
2483 memcpy(page_address(tmp_page),
2484 page_address(sh->dev[qd_idx].page),
2486 compute_parity6(sh, UPDATE_PARITY);
2487 if (memcmp(page_address(tmp_page),
2488 page_address(sh->dev[qd_idx].page),
2489 STRIPE_SIZE) != 0) {
2490 clear_bit(STRIPE_INSYNC, &sh->state);
2494 if (update_p || update_q) {
2495 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2496 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2497 /* don't try to repair!! */
2498 update_p = update_q = 0;
2501 /* now write out any block on a failed drive,
2502 * or P or Q if they need it
2505 if (s->failed == 2) {
2506 dev = &sh->dev[r6s->failed_num[1]];
2508 set_bit(R5_LOCKED, &dev->flags);
2509 set_bit(R5_Wantwrite, &dev->flags);
2511 if (s->failed >= 1) {
2512 dev = &sh->dev[r6s->failed_num[0]];
2514 set_bit(R5_LOCKED, &dev->flags);
2515 set_bit(R5_Wantwrite, &dev->flags);
2519 dev = &sh->dev[pd_idx];
2521 set_bit(R5_LOCKED, &dev->flags);
2522 set_bit(R5_Wantwrite, &dev->flags);
2525 dev = &sh->dev[qd_idx];
2527 set_bit(R5_LOCKED, &dev->flags);
2528 set_bit(R5_Wantwrite, &dev->flags);
2530 clear_bit(STRIPE_DEGRADED, &sh->state);
2532 set_bit(STRIPE_INSYNC, &sh->state);
2536 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2537 struct r6_state *r6s)
2541 /* We have read all the blocks in this stripe and now we need to
2542 * copy some of them into a target stripe for expand.
2544 struct dma_async_tx_descriptor *tx = NULL;
2545 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2546 for (i = 0; i < sh->disks; i++)
2547 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2548 int dd_idx, pd_idx, j;
2549 struct stripe_head *sh2;
2551 sector_t bn = compute_blocknr(sh, i);
2552 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2554 conf->max_degraded, &dd_idx,
2556 sh2 = get_active_stripe(conf, s, conf->raid_disks,
2559 /* so far only the early blocks of this stripe
2560 * have been requested. When later blocks
2561 * get requested, we will try again
2564 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2565 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2566 /* must have already done this block */
2567 release_stripe(sh2);
2571 /* place all the copies on one channel */
2572 tx = async_memcpy(sh2->dev[dd_idx].page,
2573 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2574 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2576 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2577 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2578 for (j = 0; j < conf->raid_disks; j++)
2579 if (j != sh2->pd_idx &&
2580 (!r6s || j != raid6_next_disk(sh2->pd_idx,
2582 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2584 if (j == conf->raid_disks) {
2585 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2586 set_bit(STRIPE_HANDLE, &sh2->state);
2588 release_stripe(sh2);
2591 /* done submitting copies, wait for them to complete */
2594 dma_wait_for_async_tx(tx);
2600 * handle_stripe - do things to a stripe.
2602 * We lock the stripe and then examine the state of various bits
2603 * to see what needs to be done.
2605 * return some read request which now have data
2606 * return some write requests which are safely on disc
2607 * schedule a read on some buffers
2608 * schedule a write of some buffers
2609 * return confirmation of parity correctness
2611 * buffers are taken off read_list or write_list, and bh_cache buffers
2612 * get BH_Lock set before the stripe lock is released.
2616 static void handle_stripe5(struct stripe_head *sh)
2618 raid5_conf_t *conf = sh->raid_conf;
2619 int disks = sh->disks, i;
2620 struct bio *return_bi = NULL;
2621 struct stripe_head_state s;
2623 unsigned long pending = 0;
2624 mdk_rdev_t *blocked_rdev = NULL;
2627 memset(&s, 0, sizeof(s));
2628 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2629 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2630 atomic_read(&sh->count), sh->pd_idx,
2631 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2633 spin_lock(&sh->lock);
2634 clear_bit(STRIPE_HANDLE, &sh->state);
2635 clear_bit(STRIPE_DELAYED, &sh->state);
2637 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2638 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2639 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2640 /* Now to look around and see what can be done */
2642 /* clean-up completed biofill operations */
2643 if (test_bit(STRIPE_OP_BIOFILL, &sh->ops.complete)) {
2644 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
2645 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
2646 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
2650 for (i=disks; i--; ) {
2652 struct r5dev *dev = &sh->dev[i];
2653 clear_bit(R5_Insync, &dev->flags);
2655 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2656 "written %p\n", i, dev->flags, dev->toread, dev->read,
2657 dev->towrite, dev->written);
2659 /* maybe we can request a biofill operation
2661 * new wantfill requests are only permitted while
2662 * STRIPE_OP_BIOFILL is clear
2664 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2665 !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2666 set_bit(R5_Wantfill, &dev->flags);
2668 /* now count some things */
2669 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2670 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2671 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2673 if (test_bit(R5_Wantfill, &dev->flags))
2675 else if (dev->toread)
2679 if (!test_bit(R5_OVERWRITE, &dev->flags))
2684 rdev = rcu_dereference(conf->disks[i].rdev);
2685 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2686 blocked_rdev = rdev;
2687 atomic_inc(&rdev->nr_pending);
2690 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2691 /* The ReadError flag will just be confusing now */
2692 clear_bit(R5_ReadError, &dev->flags);
2693 clear_bit(R5_ReWrite, &dev->flags);
2695 if (!rdev || !test_bit(In_sync, &rdev->flags)
2696 || test_bit(R5_ReadError, &dev->flags)) {
2700 set_bit(R5_Insync, &dev->flags);
2704 if (unlikely(blocked_rdev)) {
2705 set_bit(STRIPE_HANDLE, &sh->state);
2709 if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2712 pr_debug("locked=%d uptodate=%d to_read=%d"
2713 " to_write=%d failed=%d failed_num=%d\n",
2714 s.locked, s.uptodate, s.to_read, s.to_write,
2715 s.failed, s.failed_num);
2716 /* check if the array has lost two devices and, if so, some requests might
2719 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2720 handle_requests_to_failed_array(conf, sh, &s, disks,
2722 if (s.failed > 1 && s.syncing) {
2723 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2724 clear_bit(STRIPE_SYNCING, &sh->state);
2728 /* might be able to return some write requests if the parity block
2729 * is safe, or on a failed drive
2731 dev = &sh->dev[sh->pd_idx];
2733 ((test_bit(R5_Insync, &dev->flags) &&
2734 !test_bit(R5_LOCKED, &dev->flags) &&
2735 test_bit(R5_UPTODATE, &dev->flags)) ||
2736 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2737 handle_completed_write_requests(conf, sh, disks, &return_bi);
2739 /* Now we might consider reading some blocks, either to check/generate
2740 * parity, or to satisfy requests
2741 * or to load a block that is being partially written.
2743 if (s.to_read || s.non_overwrite ||
2744 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2745 test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2746 handle_issuing_new_read_requests5(sh, &s, disks);
2748 /* Now we check to see if any write operations have recently
2752 /* leave prexor set until postxor is done, allows us to distinguish
2753 * a rmw from a rcw during biodrain
2756 if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2757 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2760 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2761 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2762 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2764 for (i = disks; i--; )
2765 clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2768 /* if only POSTXOR is set then this is an 'expand' postxor */
2769 if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2770 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2772 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2773 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2774 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2776 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2777 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2778 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2780 /* All the 'written' buffers and the parity block are ready to
2781 * be written back to disk
2783 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2784 for (i = disks; i--; ) {
2786 if (test_bit(R5_LOCKED, &dev->flags) &&
2787 (i == sh->pd_idx || dev->written)) {
2788 pr_debug("Writing block %d\n", i);
2789 set_bit(R5_Wantwrite, &dev->flags);
2792 if (!test_bit(R5_Insync, &dev->flags) ||
2793 (i == sh->pd_idx && s.failed == 0))
2794 set_bit(STRIPE_INSYNC, &sh->state);
2797 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2798 atomic_dec(&conf->preread_active_stripes);
2799 if (atomic_read(&conf->preread_active_stripes) <
2801 md_wakeup_thread(conf->mddev->thread);
2805 /* Now to consider new write requests and what else, if anything
2806 * should be read. We do not handle new writes when:
2807 * 1/ A 'write' operation (copy+xor) is already in flight.
2808 * 2/ A 'check' operation is in flight, as it may clobber the parity
2811 if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2812 !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2813 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2815 /* maybe we need to check and possibly fix the parity for this stripe
2816 * Any reads will already have been scheduled, so we just see if enough
2817 * data is available. The parity check is held off while parity
2818 * dependent operations are in flight.
2820 if ((s.syncing && s.locked == 0 &&
2821 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2822 !test_bit(STRIPE_INSYNC, &sh->state)) ||
2823 test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2824 test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2825 handle_parity_checks5(conf, sh, &s, disks);
2827 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2828 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2829 clear_bit(STRIPE_SYNCING, &sh->state);
2832 /* If the failed drive is just a ReadError, then we might need to progress
2833 * the repair/check process
2835 if (s.failed == 1 && !conf->mddev->ro &&
2836 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2837 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2838 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2840 dev = &sh->dev[s.failed_num];
2841 if (!test_bit(R5_ReWrite, &dev->flags)) {
2842 set_bit(R5_Wantwrite, &dev->flags);
2843 set_bit(R5_ReWrite, &dev->flags);
2844 set_bit(R5_LOCKED, &dev->flags);
2847 /* let's read it back */
2848 set_bit(R5_Wantread, &dev->flags);
2849 set_bit(R5_LOCKED, &dev->flags);
2854 /* Finish postxor operations initiated by the expansion
2857 if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2858 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2860 clear_bit(STRIPE_EXPANDING, &sh->state);
2862 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2863 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2864 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2866 for (i = conf->raid_disks; i--; )
2867 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2868 set_bit(R5_LOCKED, &dev->flags);
2872 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2873 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2874 /* Need to write out all blocks after computing parity */
2875 sh->disks = conf->raid_disks;
2876 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2878 s.locked += handle_write_operations5(sh, 1, 1);
2879 } else if (s.expanded &&
2881 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2882 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2883 atomic_dec(&conf->reshape_stripes);
2884 wake_up(&conf->wait_for_overlap);
2885 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2888 if (s.expanding && s.locked == 0 &&
2889 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2890 handle_stripe_expansion(conf, sh, NULL);
2893 pending = get_stripe_work(sh);
2896 spin_unlock(&sh->lock);
2898 /* wait for this device to become unblocked */
2899 if (unlikely(blocked_rdev))
2900 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2903 raid5_run_ops(sh, pending);
2907 return_io(return_bi);
2911 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2913 raid6_conf_t *conf = sh->raid_conf;
2914 int disks = sh->disks;
2915 struct bio *return_bi = NULL;
2916 int i, pd_idx = sh->pd_idx;
2917 struct stripe_head_state s;
2918 struct r6_state r6s;
2919 struct r5dev *dev, *pdev, *qdev;
2920 mdk_rdev_t *blocked_rdev = NULL;
2922 r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2923 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2924 "pd_idx=%d, qd_idx=%d\n",
2925 (unsigned long long)sh->sector, sh->state,
2926 atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2927 memset(&s, 0, sizeof(s));
2929 spin_lock(&sh->lock);
2930 clear_bit(STRIPE_HANDLE, &sh->state);
2931 clear_bit(STRIPE_DELAYED, &sh->state);
2933 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2934 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2935 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2936 /* Now to look around and see what can be done */
2939 for (i=disks; i--; ) {
2942 clear_bit(R5_Insync, &dev->flags);
2944 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2945 i, dev->flags, dev->toread, dev->towrite, dev->written);
2946 /* maybe we can reply to a read */
2947 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2948 struct bio *rbi, *rbi2;
2949 pr_debug("Return read for disc %d\n", i);
2950 spin_lock_irq(&conf->device_lock);
2953 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2954 wake_up(&conf->wait_for_overlap);
2955 spin_unlock_irq(&conf->device_lock);
2956 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2957 copy_data(0, rbi, dev->page, dev->sector);
2958 rbi2 = r5_next_bio(rbi, dev->sector);
2959 spin_lock_irq(&conf->device_lock);
2960 if (--rbi->bi_phys_segments == 0) {
2961 rbi->bi_next = return_bi;
2964 spin_unlock_irq(&conf->device_lock);
2969 /* now count some things */
2970 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2971 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2978 if (!test_bit(R5_OVERWRITE, &dev->flags))
2983 rdev = rcu_dereference(conf->disks[i].rdev);
2984 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2985 blocked_rdev = rdev;
2986 atomic_inc(&rdev->nr_pending);
2989 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2990 /* The ReadError flag will just be confusing now */
2991 clear_bit(R5_ReadError, &dev->flags);
2992 clear_bit(R5_ReWrite, &dev->flags);
2994 if (!rdev || !test_bit(In_sync, &rdev->flags)
2995 || test_bit(R5_ReadError, &dev->flags)) {
2997 r6s.failed_num[s.failed] = i;
3000 set_bit(R5_Insync, &dev->flags);
3004 if (unlikely(blocked_rdev)) {
3005 set_bit(STRIPE_HANDLE, &sh->state);
3008 pr_debug("locked=%d uptodate=%d to_read=%d"
3009 " to_write=%d failed=%d failed_num=%d,%d\n",
3010 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3011 r6s.failed_num[0], r6s.failed_num[1]);
3012 /* check if the array has lost >2 devices and, if so, some requests
3013 * might need to be failed
3015 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3016 handle_requests_to_failed_array(conf, sh, &s, disks,
3018 if (s.failed > 2 && s.syncing) {
3019 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3020 clear_bit(STRIPE_SYNCING, &sh->state);
3025 * might be able to return some write requests if the parity blocks
3026 * are safe, or on a failed drive
3028 pdev = &sh->dev[pd_idx];
3029 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3030 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3031 qdev = &sh->dev[r6s.qd_idx];
3032 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3033 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3036 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3037 && !test_bit(R5_LOCKED, &pdev->flags)
3038 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3039 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3040 && !test_bit(R5_LOCKED, &qdev->flags)
3041 && test_bit(R5_UPTODATE, &qdev->flags)))))
3042 handle_completed_write_requests(conf, sh, disks, &return_bi);
3044 /* Now we might consider reading some blocks, either to check/generate
3045 * parity, or to satisfy requests
3046 * or to load a block that is being partially written.
3048 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3049 (s.syncing && (s.uptodate < disks)) || s.expanding)
3050 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3052 /* now to consider writing and what else, if anything should be read */
3054 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3056 /* maybe we need to check and possibly fix the parity for this stripe
3057 * Any reads will already have been scheduled, so we just see if enough
3060 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3061 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3063 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3064 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3065 clear_bit(STRIPE_SYNCING, &sh->state);
3068 /* If the failed drives are just a ReadError, then we might need
3069 * to progress the repair/check process
3071 if (s.failed <= 2 && !conf->mddev->ro)
3072 for (i = 0; i < s.failed; i++) {
3073 dev = &sh->dev[r6s.failed_num[i]];
3074 if (test_bit(R5_ReadError, &dev->flags)
3075 && !test_bit(R5_LOCKED, &dev->flags)
3076 && test_bit(R5_UPTODATE, &dev->flags)
3078 if (!test_bit(R5_ReWrite, &dev->flags)) {
3079 set_bit(R5_Wantwrite, &dev->flags);
3080 set_bit(R5_ReWrite, &dev->flags);
3081 set_bit(R5_LOCKED, &dev->flags);
3083 /* let's read it back */
3084 set_bit(R5_Wantread, &dev->flags);
3085 set_bit(R5_LOCKED, &dev->flags);
3090 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3091 /* Need to write out all blocks after computing P&Q */
3092 sh->disks = conf->raid_disks;
3093 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3095 compute_parity6(sh, RECONSTRUCT_WRITE);
3096 for (i = conf->raid_disks ; i-- ; ) {
3097 set_bit(R5_LOCKED, &sh->dev[i].flags);
3099 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3101 clear_bit(STRIPE_EXPANDING, &sh->state);
3102 } else if (s.expanded) {
3103 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3104 atomic_dec(&conf->reshape_stripes);
3105 wake_up(&conf->wait_for_overlap);
3106 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3109 if (s.expanding && s.locked == 0 &&
3110 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
3111 handle_stripe_expansion(conf, sh, &r6s);
3114 spin_unlock(&sh->lock);
3116 /* wait for this device to become unblocked */
3117 if (unlikely(blocked_rdev))
3118 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3120 return_io(return_bi);
3122 for (i=disks; i-- ;) {
3126 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3128 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3133 set_bit(STRIPE_IO_STARTED, &sh->state);
3135 bi = &sh->dev[i].req;
3139 bi->bi_end_io = raid5_end_write_request;
3141 bi->bi_end_io = raid5_end_read_request;
3144 rdev = rcu_dereference(conf->disks[i].rdev);
3145 if (rdev && test_bit(Faulty, &rdev->flags))
3148 atomic_inc(&rdev->nr_pending);
3152 if (s.syncing || s.expanding || s.expanded)
3153 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3155 bi->bi_bdev = rdev->bdev;
3156 pr_debug("for %llu schedule op %ld on disc %d\n",
3157 (unsigned long long)sh->sector, bi->bi_rw, i);
3158 atomic_inc(&sh->count);
3159 bi->bi_sector = sh->sector + rdev->data_offset;
3160 bi->bi_flags = 1 << BIO_UPTODATE;
3162 bi->bi_max_vecs = 1;
3164 bi->bi_io_vec = &sh->dev[i].vec;
3165 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3166 bi->bi_io_vec[0].bv_offset = 0;
3167 bi->bi_size = STRIPE_SIZE;
3170 test_bit(R5_ReWrite, &sh->dev[i].flags))
3171 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3172 generic_make_request(bi);
3175 set_bit(STRIPE_DEGRADED, &sh->state);
3176 pr_debug("skip op %ld on disc %d for sector %llu\n",
3177 bi->bi_rw, i, (unsigned long long)sh->sector);
3178 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3179 set_bit(STRIPE_HANDLE, &sh->state);
3184 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3186 if (sh->raid_conf->level == 6)
3187 handle_stripe6(sh, tmp_page);
3194 static void raid5_activate_delayed(raid5_conf_t *conf)
3196 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3197 while (!list_empty(&conf->delayed_list)) {
3198 struct list_head *l = conf->delayed_list.next;
3199 struct stripe_head *sh;
3200 sh = list_entry(l, struct stripe_head, lru);
3202 clear_bit(STRIPE_DELAYED, &sh->state);
3203 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3204 atomic_inc(&conf->preread_active_stripes);
3205 list_add_tail(&sh->lru, &conf->hold_list);
3208 blk_plug_device(conf->mddev->queue);
3211 static void activate_bit_delay(raid5_conf_t *conf)
3213 /* device_lock is held */
3214 struct list_head head;
3215 list_add(&head, &conf->bitmap_list);
3216 list_del_init(&conf->bitmap_list);
3217 while (!list_empty(&head)) {
3218 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3219 list_del_init(&sh->lru);
3220 atomic_inc(&sh->count);
3221 __release_stripe(conf, sh);
3225 static void unplug_slaves(mddev_t *mddev)
3227 raid5_conf_t *conf = mddev_to_conf(mddev);
3231 for (i=0; i<mddev->raid_disks; i++) {
3232 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3233 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3234 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3236 atomic_inc(&rdev->nr_pending);
3239 blk_unplug(r_queue);
3241 rdev_dec_pending(rdev, mddev);
3248 static void raid5_unplug_device(struct request_queue *q)
3250 mddev_t *mddev = q->queuedata;
3251 raid5_conf_t *conf = mddev_to_conf(mddev);
3252 unsigned long flags;
3254 spin_lock_irqsave(&conf->device_lock, flags);
3256 if (blk_remove_plug(q)) {
3258 raid5_activate_delayed(conf);
3260 md_wakeup_thread(mddev->thread);
3262 spin_unlock_irqrestore(&conf->device_lock, flags);
3264 unplug_slaves(mddev);
3267 static int raid5_congested(void *data, int bits)
3269 mddev_t *mddev = data;
3270 raid5_conf_t *conf = mddev_to_conf(mddev);
3272 /* No difference between reads and writes. Just check
3273 * how busy the stripe_cache is
3275 if (conf->inactive_blocked)
3279 if (list_empty_careful(&conf->inactive_list))
3285 /* We want read requests to align with chunks where possible,
3286 * but write requests don't need to.
3288 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3290 mddev_t *mddev = q->queuedata;
3291 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3293 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3294 unsigned int bio_sectors = bio->bi_size >> 9;
3296 if (bio_data_dir(bio) == WRITE)
3297 return biovec->bv_len; /* always allow writes to be mergeable */
3299 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3300 if (max < 0) max = 0;
3301 if (max <= biovec->bv_len && bio_sectors == 0)
3302 return biovec->bv_len;
3308 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3310 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3311 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3312 unsigned int bio_sectors = bio->bi_size >> 9;
3314 return chunk_sectors >=
3315 ((sector & (chunk_sectors - 1)) + bio_sectors);
3319 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3320 * later sampled by raid5d.
3322 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3324 unsigned long flags;
3326 spin_lock_irqsave(&conf->device_lock, flags);
3328 bi->bi_next = conf->retry_read_aligned_list;
3329 conf->retry_read_aligned_list = bi;
3331 spin_unlock_irqrestore(&conf->device_lock, flags);
3332 md_wakeup_thread(conf->mddev->thread);
3336 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3340 bi = conf->retry_read_aligned;
3342 conf->retry_read_aligned = NULL;
3345 bi = conf->retry_read_aligned_list;
3347 conf->retry_read_aligned_list = bi->bi_next;
3349 bi->bi_phys_segments = 1; /* biased count of active stripes */
3350 bi->bi_hw_segments = 0; /* count of processed stripes */
3358 * The "raid5_align_endio" should check if the read succeeded and if it
3359 * did, call bio_endio on the original bio (having bio_put the new bio
3361 * If the read failed..
3363 static void raid5_align_endio(struct bio *bi, int error)
3365 struct bio* raid_bi = bi->bi_private;
3368 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3373 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3374 conf = mddev_to_conf(mddev);
3375 rdev = (void*)raid_bi->bi_next;
3376 raid_bi->bi_next = NULL;
3378 rdev_dec_pending(rdev, conf->mddev);
3380 if (!error && uptodate) {
3381 bio_endio(raid_bi, 0);
3382 if (atomic_dec_and_test(&conf->active_aligned_reads))
3383 wake_up(&conf->wait_for_stripe);
3388 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3390 add_bio_to_retry(raid_bi, conf);
3393 static int bio_fits_rdev(struct bio *bi)
3395 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3397 if ((bi->bi_size>>9) > q->max_sectors)
3399 blk_recount_segments(q, bi);
3400 if (bi->bi_phys_segments > q->max_phys_segments ||
3401 bi->bi_hw_segments > q->max_hw_segments)
3404 if (q->merge_bvec_fn)
3405 /* it's too hard to apply the merge_bvec_fn at this stage,
3414 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3416 mddev_t *mddev = q->queuedata;
3417 raid5_conf_t *conf = mddev_to_conf(mddev);
3418 const unsigned int raid_disks = conf->raid_disks;
3419 const unsigned int data_disks = raid_disks - conf->max_degraded;
3420 unsigned int dd_idx, pd_idx;
3421 struct bio* align_bi;
3424 if (!in_chunk_boundary(mddev, raid_bio)) {
3425 pr_debug("chunk_aligned_read : non aligned\n");
3429 * use bio_clone to make a copy of the bio
3431 align_bi = bio_clone(raid_bio, GFP_NOIO);
3435 * set bi_end_io to a new function, and set bi_private to the
3438 align_bi->bi_end_io = raid5_align_endio;
3439 align_bi->bi_private = raid_bio;
3443 align_bi->bi_sector = raid5_compute_sector(raid_bio->bi_sector,
3451 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3452 if (rdev && test_bit(In_sync, &rdev->flags)) {
3453 atomic_inc(&rdev->nr_pending);
3455 raid_bio->bi_next = (void*)rdev;
3456 align_bi->bi_bdev = rdev->bdev;
3457 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3458 align_bi->bi_sector += rdev->data_offset;
3460 if (!bio_fits_rdev(align_bi)) {
3461 /* too big in some way */
3463 rdev_dec_pending(rdev, mddev);
3467 spin_lock_irq(&conf->device_lock);
3468 wait_event_lock_irq(conf->wait_for_stripe,
3470 conf->device_lock, /* nothing */);
3471 atomic_inc(&conf->active_aligned_reads);
3472 spin_unlock_irq(&conf->device_lock);
3474 generic_make_request(align_bi);
3483 /* __get_priority_stripe - get the next stripe to process
3485 * Full stripe writes are allowed to pass preread active stripes up until
3486 * the bypass_threshold is exceeded. In general the bypass_count
3487 * increments when the handle_list is handled before the hold_list; however, it
3488 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3489 * stripe with in flight i/o. The bypass_count will be reset when the
3490 * head of the hold_list has changed, i.e. the head was promoted to the
3493 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3495 struct stripe_head *sh;
3497 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3499 list_empty(&conf->handle_list) ? "empty" : "busy",
3500 list_empty(&conf->hold_list) ? "empty" : "busy",
3501 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3503 if (!list_empty(&conf->handle_list)) {
3504 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3506 if (list_empty(&conf->hold_list))
3507 conf->bypass_count = 0;
3508 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3509 if (conf->hold_list.next == conf->last_hold)
3510 conf->bypass_count++;
3512 conf->last_hold = conf->hold_list.next;
3513 conf->bypass_count -= conf->bypass_threshold;
3514 if (conf->bypass_count < 0)
3515 conf->bypass_count = 0;
3518 } else if (!list_empty(&conf->hold_list) &&
3519 ((conf->bypass_threshold &&
3520 conf->bypass_count > conf->bypass_threshold) ||
3521 atomic_read(&conf->pending_full_writes) == 0)) {
3522 sh = list_entry(conf->hold_list.next,
3524 conf->bypass_count -= conf->bypass_threshold;
3525 if (conf->bypass_count < 0)
3526 conf->bypass_count = 0;
3530 list_del_init(&sh->lru);
3531 atomic_inc(&sh->count);
3532 BUG_ON(atomic_read(&sh->count) != 1);
3536 static int make_request(struct request_queue *q, struct bio * bi)
3538 mddev_t *mddev = q->queuedata;
3539 raid5_conf_t *conf = mddev_to_conf(mddev);
3540 unsigned int dd_idx, pd_idx;
3541 sector_t new_sector;
3542 sector_t logical_sector, last_sector;
3543 struct stripe_head *sh;
3544 const int rw = bio_data_dir(bi);
3547 if (unlikely(bio_barrier(bi))) {
3548 bio_endio(bi, -EOPNOTSUPP);
3552 md_write_start(mddev, bi);
3554 disk_stat_inc(mddev->gendisk, ios[rw]);
3555 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3558 mddev->reshape_position == MaxSector &&
3559 chunk_aligned_read(q,bi))
3562 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3563 last_sector = bi->bi_sector + (bi->bi_size>>9);
3565 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3567 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3569 int disks, data_disks;
3572 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3573 if (likely(conf->expand_progress == MaxSector))
3574 disks = conf->raid_disks;
3576 /* spinlock is needed as expand_progress may be
3577 * 64bit on a 32bit platform, and so it might be
3578 * possible to see a half-updated value
3579 * Ofcourse expand_progress could change after
3580 * the lock is dropped, so once we get a reference
3581 * to the stripe that we think it is, we will have
3584 spin_lock_irq(&conf->device_lock);
3585 disks = conf->raid_disks;
3586 if (logical_sector >= conf->expand_progress)
3587 disks = conf->previous_raid_disks;
3589 if (logical_sector >= conf->expand_lo) {
3590 spin_unlock_irq(&conf->device_lock);
3595 spin_unlock_irq(&conf->device_lock);
3597 data_disks = disks - conf->max_degraded;
3599 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3600 &dd_idx, &pd_idx, conf);
3601 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3602 (unsigned long long)new_sector,
3603 (unsigned long long)logical_sector);
3605 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3607 if (unlikely(conf->expand_progress != MaxSector)) {
3608 /* expansion might have moved on while waiting for a
3609 * stripe, so we must do the range check again.
3610 * Expansion could still move past after this
3611 * test, but as we are holding a reference to
3612 * 'sh', we know that if that happens,
3613 * STRIPE_EXPANDING will get set and the expansion
3614 * won't proceed until we finish with the stripe.
3617 spin_lock_irq(&conf->device_lock);
3618 if (logical_sector < conf->expand_progress &&
3619 disks == conf->previous_raid_disks)
3620 /* mismatch, need to try again */
3622 spin_unlock_irq(&conf->device_lock);
3628 /* FIXME what if we get a false positive because these
3629 * are being updated.
3631 if (logical_sector >= mddev->suspend_lo &&
3632 logical_sector < mddev->suspend_hi) {
3638 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3639 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3640 /* Stripe is busy expanding or
3641 * add failed due to overlap. Flush everything
3644 raid5_unplug_device(mddev->queue);
3649 finish_wait(&conf->wait_for_overlap, &w);
3650 set_bit(STRIPE_HANDLE, &sh->state);
3651 clear_bit(STRIPE_DELAYED, &sh->state);
3654 /* cannot get stripe for read-ahead, just give-up */
3655 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3656 finish_wait(&conf->wait_for_overlap, &w);
3661 spin_lock_irq(&conf->device_lock);
3662 remaining = --bi->bi_phys_segments;
3663 spin_unlock_irq(&conf->device_lock);
3664 if (remaining == 0) {
3667 md_write_end(mddev);
3674 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3676 /* reshaping is quite different to recovery/resync so it is
3677 * handled quite separately ... here.
3679 * On each call to sync_request, we gather one chunk worth of
3680 * destination stripes and flag them as expanding.
3681 * Then we find all the source stripes and request reads.
3682 * As the reads complete, handle_stripe will copy the data
3683 * into the destination stripe and release that stripe.
3685 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3686 struct stripe_head *sh;
3688 sector_t first_sector, last_sector;
3689 int raid_disks = conf->previous_raid_disks;
3690 int data_disks = raid_disks - conf->max_degraded;
3691 int new_data_disks = conf->raid_disks - conf->max_degraded;
3694 sector_t writepos, safepos, gap;
3696 if (sector_nr == 0 &&
3697 conf->expand_progress != 0) {
3698 /* restarting in the middle, skip the initial sectors */
3699 sector_nr = conf->expand_progress;
3700 sector_div(sector_nr, new_data_disks);
3705 /* we update the metadata when there is more than 3Meg
3706 * in the block range (that is rather arbitrary, should
3707 * probably be time based) or when the data about to be
3708 * copied would over-write the source of the data at
3709 * the front of the range.
3710 * i.e. one new_stripe forward from expand_progress new_maps
3711 * to after where expand_lo old_maps to
3713 writepos = conf->expand_progress +
3714 conf->chunk_size/512*(new_data_disks);
3715 sector_div(writepos, new_data_disks);
3716 safepos = conf->expand_lo;
3717 sector_div(safepos, data_disks);
3718 gap = conf->expand_progress - conf->expand_lo;
3720 if (writepos >= safepos ||
3721 gap > (new_data_disks)*3000*2 /*3Meg*/) {
3722 /* Cannot proceed until we've updated the superblock... */
3723 wait_event(conf->wait_for_overlap,
3724 atomic_read(&conf->reshape_stripes)==0);
3725 mddev->reshape_position = conf->expand_progress;
3726 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3727 md_wakeup_thread(mddev->thread);
3728 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3729 kthread_should_stop());
3730 spin_lock_irq(&conf->device_lock);
3731 conf->expand_lo = mddev->reshape_position;
3732 spin_unlock_irq(&conf->device_lock);
3733 wake_up(&conf->wait_for_overlap);
3736 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3739 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3740 sh = get_active_stripe(conf, sector_nr+i,
3741 conf->raid_disks, pd_idx, 0);
3742 set_bit(STRIPE_EXPANDING, &sh->state);
3743 atomic_inc(&conf->reshape_stripes);
3744 /* If any of this stripe is beyond the end of the old
3745 * array, then we need to zero those blocks
3747 for (j=sh->disks; j--;) {
3749 if (j == sh->pd_idx)
3751 if (conf->level == 6 &&
3752 j == raid6_next_disk(sh->pd_idx, sh->disks))
3754 s = compute_blocknr(sh, j);
3755 if (s < (mddev->array_size<<1)) {
3759 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3760 set_bit(R5_Expanded, &sh->dev[j].flags);
3761 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3764 set_bit(STRIPE_EXPAND_READY, &sh->state);
3765 set_bit(STRIPE_HANDLE, &sh->state);
3769 spin_lock_irq(&conf->device_lock);
3770 conf->expand_progress = (sector_nr + i) * new_data_disks;
3771 spin_unlock_irq(&conf->device_lock);
3772 /* Ok, those stripe are ready. We can start scheduling
3773 * reads on the source stripes.
3774 * The source stripes are determined by mapping the first and last
3775 * block on the destination stripes.
3778 raid5_compute_sector(sector_nr*(new_data_disks),
3779 raid_disks, data_disks,
3780 &dd_idx, &pd_idx, conf);
3782 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3783 *(new_data_disks) -1,
3784 raid_disks, data_disks,
3785 &dd_idx, &pd_idx, conf);
3786 if (last_sector >= (mddev->size<<1))
3787 last_sector = (mddev->size<<1)-1;
3788 while (first_sector <= last_sector) {
3789 pd_idx = stripe_to_pdidx(first_sector, conf,
3790 conf->previous_raid_disks);
3791 sh = get_active_stripe(conf, first_sector,
3792 conf->previous_raid_disks, pd_idx, 0);
3793 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3794 set_bit(STRIPE_HANDLE, &sh->state);
3796 first_sector += STRIPE_SECTORS;
3798 /* If this takes us to the resync_max point where we have to pause,
3799 * then we need to write out the superblock.
3801 sector_nr += conf->chunk_size>>9;
3802 if (sector_nr >= mddev->resync_max) {
3803 /* Cannot proceed until we've updated the superblock... */
3804 wait_event(conf->wait_for_overlap,
3805 atomic_read(&conf->reshape_stripes) == 0);
3806 mddev->reshape_position = conf->expand_progress;
3807 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3808 md_wakeup_thread(mddev->thread);
3809 wait_event(mddev->sb_wait,
3810 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3811 || kthread_should_stop());
3812 spin_lock_irq(&conf->device_lock);
3813 conf->expand_lo = mddev->reshape_position;
3814 spin_unlock_irq(&conf->device_lock);
3815 wake_up(&conf->wait_for_overlap);
3817 return conf->chunk_size>>9;
3820 /* FIXME go_faster isn't used */
3821 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3823 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3824 struct stripe_head *sh;
3826 int raid_disks = conf->raid_disks;
3827 sector_t max_sector = mddev->size << 1;
3829 int still_degraded = 0;
3832 if (sector_nr >= max_sector) {
3833 /* just being told to finish up .. nothing much to do */
3834 unplug_slaves(mddev);
3835 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3840 if (mddev->curr_resync < max_sector) /* aborted */
3841 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3843 else /* completed sync */
3845 bitmap_close_sync(mddev->bitmap);
3850 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3851 return reshape_request(mddev, sector_nr, skipped);
3853 /* No need to check resync_max as we never do more than one
3854 * stripe, and as resync_max will always be on a chunk boundary,
3855 * if the check in md_do_sync didn't fire, there is no chance
3856 * of overstepping resync_max here
3859 /* if there is too many failed drives and we are trying
3860 * to resync, then assert that we are finished, because there is
3861 * nothing we can do.
3863 if (mddev->degraded >= conf->max_degraded &&
3864 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3865 sector_t rv = (mddev->size << 1) - sector_nr;
3869 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3870 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3871 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3872 /* we can skip this block, and probably more */
3873 sync_blocks /= STRIPE_SECTORS;
3875 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3879 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3881 pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3882 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3884 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3885 /* make sure we don't swamp the stripe cache if someone else
3886 * is trying to get access
3888 schedule_timeout_uninterruptible(1);
3890 /* Need to check if array will still be degraded after recovery/resync
3891 * We don't need to check the 'failed' flag as when that gets set,
3894 for (i=0; i<mddev->raid_disks; i++)
3895 if (conf->disks[i].rdev == NULL)
3898 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3900 spin_lock(&sh->lock);
3901 set_bit(STRIPE_SYNCING, &sh->state);
3902 clear_bit(STRIPE_INSYNC, &sh->state);
3903 spin_unlock(&sh->lock);
3905 handle_stripe(sh, NULL);
3908 return STRIPE_SECTORS;
3911 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3913 /* We may not be able to submit a whole bio at once as there
3914 * may not be enough stripe_heads available.
3915 * We cannot pre-allocate enough stripe_heads as we may need
3916 * more than exist in the cache (if we allow ever large chunks).
3917 * So we do one stripe head at a time and record in
3918 * ->bi_hw_segments how many have been done.
3920 * We *know* that this entire raid_bio is in one chunk, so
3921 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3923 struct stripe_head *sh;
3925 sector_t sector, logical_sector, last_sector;
3930 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3931 sector = raid5_compute_sector( logical_sector,
3933 conf->raid_disks - conf->max_degraded,
3937 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3939 for (; logical_sector < last_sector;
3940 logical_sector += STRIPE_SECTORS,
3941 sector += STRIPE_SECTORS,
3944 if (scnt < raid_bio->bi_hw_segments)
3945 /* already done this stripe */
3948 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3951 /* failed to get a stripe - must wait */
3952 raid_bio->bi_hw_segments = scnt;
3953 conf->retry_read_aligned = raid_bio;
3957 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3958 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3960 raid_bio->bi_hw_segments = scnt;
3961 conf->retry_read_aligned = raid_bio;
3965 handle_stripe(sh, NULL);
3969 spin_lock_irq(&conf->device_lock);
3970 remaining = --raid_bio->bi_phys_segments;
3971 spin_unlock_irq(&conf->device_lock);
3973 bio_endio(raid_bio, 0);
3974 if (atomic_dec_and_test(&conf->active_aligned_reads))
3975 wake_up(&conf->wait_for_stripe);
3982 * This is our raid5 kernel thread.
3984 * We scan the hash table for stripes which can be handled now.
3985 * During the scan, completed stripes are saved for us by the interrupt
3986 * handler, so that they will not have to wait for our next wakeup.
3988 static void raid5d(mddev_t *mddev)
3990 struct stripe_head *sh;
3991 raid5_conf_t *conf = mddev_to_conf(mddev);
3994 pr_debug("+++ raid5d active\n");
3996 md_check_recovery(mddev);
3999 spin_lock_irq(&conf->device_lock);
4003 if (conf->seq_flush != conf->seq_write) {
4004 int seq = conf->seq_flush;
4005 spin_unlock_irq(&conf->device_lock);
4006 bitmap_unplug(mddev->bitmap);
4007 spin_lock_irq(&conf->device_lock);
4008 conf->seq_write = seq;
4009 activate_bit_delay(conf);
4012 while ((bio = remove_bio_from_retry(conf))) {
4014 spin_unlock_irq(&conf->device_lock);
4015 ok = retry_aligned_read(conf, bio);
4016 spin_lock_irq(&conf->device_lock);
4022 sh = __get_priority_stripe(conf);
4025 async_tx_issue_pending_all();
4028 spin_unlock_irq(&conf->device_lock);
4031 handle_stripe(sh, conf->spare_page);
4034 spin_lock_irq(&conf->device_lock);
4036 pr_debug("%d stripes handled\n", handled);
4038 spin_unlock_irq(&conf->device_lock);
4040 unplug_slaves(mddev);
4042 pr_debug("--- raid5d inactive\n");
4046 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4048 raid5_conf_t *conf = mddev_to_conf(mddev);
4050 return sprintf(page, "%d\n", conf->max_nr_stripes);
4056 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4058 raid5_conf_t *conf = mddev_to_conf(mddev);
4060 if (len >= PAGE_SIZE)
4065 if (strict_strtoul(page, 10, &new))
4067 if (new <= 16 || new > 32768)
4069 while (new < conf->max_nr_stripes) {
4070 if (drop_one_stripe(conf))
4071 conf->max_nr_stripes--;
4075 md_allow_write(mddev);
4076 while (new > conf->max_nr_stripes) {
4077 if (grow_one_stripe(conf))
4078 conf->max_nr_stripes++;
4084 static struct md_sysfs_entry
4085 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4086 raid5_show_stripe_cache_size,
4087 raid5_store_stripe_cache_size);
4090 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4092 raid5_conf_t *conf = mddev_to_conf(mddev);
4094 return sprintf(page, "%d\n", conf->bypass_threshold);
4100 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4102 raid5_conf_t *conf = mddev_to_conf(mddev);
4104 if (len >= PAGE_SIZE)
4109 if (strict_strtoul(page, 10, &new))
4111 if (new > conf->max_nr_stripes)
4113 conf->bypass_threshold = new;
4117 static struct md_sysfs_entry
4118 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4120 raid5_show_preread_threshold,
4121 raid5_store_preread_threshold);
4124 stripe_cache_active_show(mddev_t *mddev, char *page)
4126 raid5_conf_t *conf = mddev_to_conf(mddev);
4128 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4133 static struct md_sysfs_entry
4134 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4136 static struct attribute *raid5_attrs[] = {
4137 &raid5_stripecache_size.attr,
4138 &raid5_stripecache_active.attr,
4139 &raid5_preread_bypass_threshold.attr,
4142 static struct attribute_group raid5_attrs_group = {
4144 .attrs = raid5_attrs,
4147 static int run(mddev_t *mddev)
4150 int raid_disk, memory;
4152 struct disk_info *disk;
4153 struct list_head *tmp;
4154 int working_disks = 0;
4156 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4157 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4158 mdname(mddev), mddev->level);
4162 if (mddev->reshape_position != MaxSector) {
4163 /* Check that we can continue the reshape.
4164 * Currently only disks can change, it must
4165 * increase, and we must be past the point where
4166 * a stripe over-writes itself
4168 sector_t here_new, here_old;
4170 int max_degraded = (mddev->level == 5 ? 1 : 2);
4172 if (mddev->new_level != mddev->level ||
4173 mddev->new_layout != mddev->layout ||
4174 mddev->new_chunk != mddev->chunk_size) {
4175 printk(KERN_ERR "raid5: %s: unsupported reshape "
4176 "required - aborting.\n",
4180 if (mddev->delta_disks <= 0) {
4181 printk(KERN_ERR "raid5: %s: unsupported reshape "
4182 "(reduce disks) required - aborting.\n",
4186 old_disks = mddev->raid_disks - mddev->delta_disks;
4187 /* reshape_position must be on a new-stripe boundary, and one
4188 * further up in new geometry must map after here in old
4191 here_new = mddev->reshape_position;
4192 if (sector_div(here_new, (mddev->chunk_size>>9)*
4193 (mddev->raid_disks - max_degraded))) {
4194 printk(KERN_ERR "raid5: reshape_position not "
4195 "on a stripe boundary\n");
4198 /* here_new is the stripe we will write to */
4199 here_old = mddev->reshape_position;
4200 sector_div(here_old, (mddev->chunk_size>>9)*
4201 (old_disks-max_degraded));
4202 /* here_old is the first stripe that we might need to read
4204 if (here_new >= here_old) {
4205 /* Reading from the same stripe as writing to - bad */
4206 printk(KERN_ERR "raid5: reshape_position too early for "
4207 "auto-recovery - aborting.\n");
4210 printk(KERN_INFO "raid5: reshape will continue\n");
4211 /* OK, we should be able to continue; */
4215 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4216 if ((conf = mddev->private) == NULL)
4218 if (mddev->reshape_position == MaxSector) {
4219 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4221 conf->raid_disks = mddev->raid_disks;
4222 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4225 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4230 conf->mddev = mddev;
4232 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4235 if (mddev->level == 6) {
4236 conf->spare_page = alloc_page(GFP_KERNEL);
4237 if (!conf->spare_page)
4240 spin_lock_init(&conf->device_lock);
4241 mddev->queue->queue_lock = &conf->device_lock;
4242 init_waitqueue_head(&conf->wait_for_stripe);
4243 init_waitqueue_head(&conf->wait_for_overlap);
4244 INIT_LIST_HEAD(&conf->handle_list);
4245 INIT_LIST_HEAD(&conf->hold_list);
4246 INIT_LIST_HEAD(&conf->delayed_list);
4247 INIT_LIST_HEAD(&conf->bitmap_list);
4248 INIT_LIST_HEAD(&conf->inactive_list);
4249 atomic_set(&conf->active_stripes, 0);
4250 atomic_set(&conf->preread_active_stripes, 0);
4251 atomic_set(&conf->active_aligned_reads, 0);
4252 conf->bypass_threshold = BYPASS_THRESHOLD;
4254 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4256 rdev_for_each(rdev, tmp, mddev) {
4257 raid_disk = rdev->raid_disk;
4258 if (raid_disk >= conf->raid_disks
4261 disk = conf->disks + raid_disk;
4265 if (test_bit(In_sync, &rdev->flags)) {
4266 char b[BDEVNAME_SIZE];
4267 printk(KERN_INFO "raid5: device %s operational as raid"
4268 " disk %d\n", bdevname(rdev->bdev,b),
4272 /* Cannot rely on bitmap to complete recovery */
4277 * 0 for a fully functional array, 1 or 2 for a degraded array.
4279 mddev->degraded = conf->raid_disks - working_disks;
4280 conf->mddev = mddev;
4281 conf->chunk_size = mddev->chunk_size;
4282 conf->level = mddev->level;
4283 if (conf->level == 6)
4284 conf->max_degraded = 2;
4286 conf->max_degraded = 1;
4287 conf->algorithm = mddev->layout;
4288 conf->max_nr_stripes = NR_STRIPES;
4289 conf->expand_progress = mddev->reshape_position;
4291 /* device size must be a multiple of chunk size */
4292 mddev->size &= ~(mddev->chunk_size/1024 -1);
4293 mddev->resync_max_sectors = mddev->size << 1;
4295 if (conf->level == 6 && conf->raid_disks < 4) {
4296 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4297 mdname(mddev), conf->raid_disks);
4300 if (!conf->chunk_size || conf->chunk_size % 4) {
4301 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4302 conf->chunk_size, mdname(mddev));
4305 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4307 "raid5: unsupported parity algorithm %d for %s\n",
4308 conf->algorithm, mdname(mddev));
4311 if (mddev->degraded > conf->max_degraded) {
4312 printk(KERN_ERR "raid5: not enough operational devices for %s"
4313 " (%d/%d failed)\n",
4314 mdname(mddev), mddev->degraded, conf->raid_disks);
4318 if (mddev->degraded > 0 &&
4319 mddev->recovery_cp != MaxSector) {
4320 if (mddev->ok_start_degraded)
4322 "raid5: starting dirty degraded array: %s"
4323 "- data corruption possible.\n",
4327 "raid5: cannot start dirty degraded array for %s\n",
4334 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4335 if (!mddev->thread) {
4337 "raid5: couldn't allocate thread for %s\n",
4342 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4343 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4344 if (grow_stripes(conf, conf->max_nr_stripes)) {
4346 "raid5: couldn't allocate %dkB for buffers\n", memory);
4347 shrink_stripes(conf);
4348 md_unregister_thread(mddev->thread);
4351 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4352 memory, mdname(mddev));
4354 if (mddev->degraded == 0)
4355 printk("raid5: raid level %d set %s active with %d out of %d"
4356 " devices, algorithm %d\n", conf->level, mdname(mddev),
4357 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4360 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4361 " out of %d devices, algorithm %d\n", conf->level,
4362 mdname(mddev), mddev->raid_disks - mddev->degraded,
4363 mddev->raid_disks, conf->algorithm);
4365 print_raid5_conf(conf);
4367 if (conf->expand_progress != MaxSector) {
4368 printk("...ok start reshape thread\n");
4369 conf->expand_lo = conf->expand_progress;
4370 atomic_set(&conf->reshape_stripes, 0);
4371 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4372 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4373 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4374 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4375 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4379 /* read-ahead size must cover two whole stripes, which is
4380 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4383 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4384 int stripe = data_disks *
4385 (mddev->chunk_size / PAGE_SIZE);
4386 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4387 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4390 /* Ok, everything is just fine now */
4391 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4393 "raid5: failed to create sysfs attributes for %s\n",
4396 mddev->queue->unplug_fn = raid5_unplug_device;
4397 mddev->queue->backing_dev_info.congested_data = mddev;
4398 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4400 mddev->array_size = mddev->size * (conf->previous_raid_disks -
4401 conf->max_degraded);
4403 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4408 print_raid5_conf(conf);
4409 safe_put_page(conf->spare_page);
4411 kfree(conf->stripe_hashtbl);
4414 mddev->private = NULL;
4415 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4421 static int stop(mddev_t *mddev)
4423 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4425 md_unregister_thread(mddev->thread);
4426 mddev->thread = NULL;
4427 shrink_stripes(conf);
4428 kfree(conf->stripe_hashtbl);
4429 mddev->queue->backing_dev_info.congested_fn = NULL;
4430 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4431 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4434 mddev->private = NULL;
4439 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4443 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4444 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4445 seq_printf(seq, "sh %llu, count %d.\n",
4446 (unsigned long long)sh->sector, atomic_read(&sh->count));
4447 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4448 for (i = 0; i < sh->disks; i++) {
4449 seq_printf(seq, "(cache%d: %p %ld) ",
4450 i, sh->dev[i].page, sh->dev[i].flags);
4452 seq_printf(seq, "\n");
4455 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4457 struct stripe_head *sh;
4458 struct hlist_node *hn;
4461 spin_lock_irq(&conf->device_lock);
4462 for (i = 0; i < NR_HASH; i++) {
4463 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4464 if (sh->raid_conf != conf)
4469 spin_unlock_irq(&conf->device_lock);
4473 static void status (struct seq_file *seq, mddev_t *mddev)
4475 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4478 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4479 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4480 for (i = 0; i < conf->raid_disks; i++)
4481 seq_printf (seq, "%s",
4482 conf->disks[i].rdev &&
4483 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4484 seq_printf (seq, "]");
4486 seq_printf (seq, "\n");
4487 printall(seq, conf);
4491 static void print_raid5_conf (raid5_conf_t *conf)
4494 struct disk_info *tmp;
4496 printk("RAID5 conf printout:\n");
4498 printk("(conf==NULL)\n");
4501 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4502 conf->raid_disks - conf->mddev->degraded);
4504 for (i = 0; i < conf->raid_disks; i++) {
4505 char b[BDEVNAME_SIZE];
4506 tmp = conf->disks + i;
4508 printk(" disk %d, o:%d, dev:%s\n",
4509 i, !test_bit(Faulty, &tmp->rdev->flags),
4510 bdevname(tmp->rdev->bdev,b));
4514 static int raid5_spare_active(mddev_t *mddev)
4517 raid5_conf_t *conf = mddev->private;
4518 struct disk_info *tmp;
4520 for (i = 0; i < conf->raid_disks; i++) {
4521 tmp = conf->disks + i;
4523 && !test_bit(Faulty, &tmp->rdev->flags)
4524 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4525 unsigned long flags;
4526 spin_lock_irqsave(&conf->device_lock, flags);
4528 spin_unlock_irqrestore(&conf->device_lock, flags);
4531 print_raid5_conf(conf);
4535 static int raid5_remove_disk(mddev_t *mddev, int number)
4537 raid5_conf_t *conf = mddev->private;
4540 struct disk_info *p = conf->disks + number;
4542 print_raid5_conf(conf);
4545 if (test_bit(In_sync, &rdev->flags) ||
4546 atomic_read(&rdev->nr_pending)) {
4550 /* Only remove non-faulty devices if recovery
4553 if (!test_bit(Faulty, &rdev->flags) &&
4554 mddev->degraded <= conf->max_degraded) {
4560 if (atomic_read(&rdev->nr_pending)) {
4561 /* lost the race, try later */
4568 print_raid5_conf(conf);
4572 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4574 raid5_conf_t *conf = mddev->private;
4577 struct disk_info *p;
4579 int last = conf->raid_disks - 1;
4581 if (mddev->degraded > conf->max_degraded)
4582 /* no point adding a device */
4585 if (rdev->raid_disk >= 0)
4586 first = last = rdev->raid_disk;
4589 * find the disk ... but prefer rdev->saved_raid_disk
4592 if (rdev->saved_raid_disk >= 0 &&
4593 rdev->saved_raid_disk >= first &&
4594 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4595 disk = rdev->saved_raid_disk;
4598 for ( ; disk <= last ; disk++)
4599 if ((p=conf->disks + disk)->rdev == NULL) {
4600 clear_bit(In_sync, &rdev->flags);
4601 rdev->raid_disk = disk;
4603 if (rdev->saved_raid_disk != disk)
4605 rcu_assign_pointer(p->rdev, rdev);
4608 print_raid5_conf(conf);
4612 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4614 /* no resync is happening, and there is enough space
4615 * on all devices, so we can resize.
4616 * We need to make sure resync covers any new space.
4617 * If the array is shrinking we should possibly wait until
4618 * any io in the removed space completes, but it hardly seems
4621 raid5_conf_t *conf = mddev_to_conf(mddev);
4623 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4624 mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4625 set_capacity(mddev->gendisk, mddev->array_size << 1);
4627 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) {
4628 mddev->recovery_cp = mddev->size << 1;
4629 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4631 mddev->size = sectors /2;
4632 mddev->resync_max_sectors = sectors;
4636 #ifdef CONFIG_MD_RAID5_RESHAPE
4637 static int raid5_check_reshape(mddev_t *mddev)
4639 raid5_conf_t *conf = mddev_to_conf(mddev);
4642 if (mddev->delta_disks < 0 ||
4643 mddev->new_level != mddev->level)
4644 return -EINVAL; /* Cannot shrink array or change level yet */
4645 if (mddev->delta_disks == 0)
4646 return 0; /* nothing to do */
4648 /* Can only proceed if there are plenty of stripe_heads.
4649 * We need a minimum of one full stripe,, and for sensible progress
4650 * it is best to have about 4 times that.
4651 * If we require 4 times, then the default 256 4K stripe_heads will
4652 * allow for chunk sizes up to 256K, which is probably OK.
4653 * If the chunk size is greater, user-space should request more
4654 * stripe_heads first.
4656 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4657 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4658 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4659 (mddev->chunk_size / STRIPE_SIZE)*4);
4663 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4667 if (mddev->degraded > conf->max_degraded)
4669 /* looks like we might be able to manage this */
4673 static int raid5_start_reshape(mddev_t *mddev)
4675 raid5_conf_t *conf = mddev_to_conf(mddev);
4677 struct list_head *rtmp;
4679 int added_devices = 0;
4680 unsigned long flags;
4682 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4685 rdev_for_each(rdev, rtmp, mddev)
4686 if (rdev->raid_disk < 0 &&
4687 !test_bit(Faulty, &rdev->flags))
4690 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4691 /* Not enough devices even to make a degraded array
4696 atomic_set(&conf->reshape_stripes, 0);
4697 spin_lock_irq(&conf->device_lock);
4698 conf->previous_raid_disks = conf->raid_disks;
4699 conf->raid_disks += mddev->delta_disks;
4700 conf->expand_progress = 0;
4701 conf->expand_lo = 0;
4702 spin_unlock_irq(&conf->device_lock);
4704 /* Add some new drives, as many as will fit.
4705 * We know there are enough to make the newly sized array work.
4707 rdev_for_each(rdev, rtmp, mddev)
4708 if (rdev->raid_disk < 0 &&
4709 !test_bit(Faulty, &rdev->flags)) {
4710 if (raid5_add_disk(mddev, rdev) == 0) {
4712 set_bit(In_sync, &rdev->flags);
4714 rdev->recovery_offset = 0;
4715 sprintf(nm, "rd%d", rdev->raid_disk);
4716 if (sysfs_create_link(&mddev->kobj,
4719 "raid5: failed to create "
4720 " link %s for %s\n",
4726 spin_lock_irqsave(&conf->device_lock, flags);
4727 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4728 spin_unlock_irqrestore(&conf->device_lock, flags);
4729 mddev->raid_disks = conf->raid_disks;
4730 mddev->reshape_position = 0;
4731 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4733 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4734 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4735 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4736 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4737 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4739 if (!mddev->sync_thread) {
4740 mddev->recovery = 0;
4741 spin_lock_irq(&conf->device_lock);
4742 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4743 conf->expand_progress = MaxSector;
4744 spin_unlock_irq(&conf->device_lock);
4747 md_wakeup_thread(mddev->sync_thread);
4748 md_new_event(mddev);
4753 static void end_reshape(raid5_conf_t *conf)
4755 struct block_device *bdev;
4757 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4758 conf->mddev->array_size = conf->mddev->size *
4759 (conf->raid_disks - conf->max_degraded);
4760 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4761 conf->mddev->changed = 1;
4763 bdev = bdget_disk(conf->mddev->gendisk, 0);
4765 mutex_lock(&bdev->bd_inode->i_mutex);
4766 i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4767 mutex_unlock(&bdev->bd_inode->i_mutex);
4770 spin_lock_irq(&conf->device_lock);
4771 conf->expand_progress = MaxSector;
4772 spin_unlock_irq(&conf->device_lock);
4773 conf->mddev->reshape_position = MaxSector;
4775 /* read-ahead size must cover two whole stripes, which is
4776 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4779 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4780 int stripe = data_disks *
4781 (conf->mddev->chunk_size / PAGE_SIZE);
4782 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4783 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4788 static void raid5_quiesce(mddev_t *mddev, int state)
4790 raid5_conf_t *conf = mddev_to_conf(mddev);
4793 case 2: /* resume for a suspend */
4794 wake_up(&conf->wait_for_overlap);
4797 case 1: /* stop all writes */
4798 spin_lock_irq(&conf->device_lock);
4800 wait_event_lock_irq(conf->wait_for_stripe,
4801 atomic_read(&conf->active_stripes) == 0 &&
4802 atomic_read(&conf->active_aligned_reads) == 0,
4803 conf->device_lock, /* nothing */);
4804 spin_unlock_irq(&conf->device_lock);
4807 case 0: /* re-enable writes */
4808 spin_lock_irq(&conf->device_lock);
4810 wake_up(&conf->wait_for_stripe);
4811 wake_up(&conf->wait_for_overlap);
4812 spin_unlock_irq(&conf->device_lock);
4817 static struct mdk_personality raid6_personality =
4821 .owner = THIS_MODULE,
4822 .make_request = make_request,
4826 .error_handler = error,
4827 .hot_add_disk = raid5_add_disk,
4828 .hot_remove_disk= raid5_remove_disk,
4829 .spare_active = raid5_spare_active,
4830 .sync_request = sync_request,
4831 .resize = raid5_resize,
4832 #ifdef CONFIG_MD_RAID5_RESHAPE
4833 .check_reshape = raid5_check_reshape,
4834 .start_reshape = raid5_start_reshape,
4836 .quiesce = raid5_quiesce,
4838 static struct mdk_personality raid5_personality =
4842 .owner = THIS_MODULE,
4843 .make_request = make_request,
4847 .error_handler = error,
4848 .hot_add_disk = raid5_add_disk,
4849 .hot_remove_disk= raid5_remove_disk,
4850 .spare_active = raid5_spare_active,
4851 .sync_request = sync_request,
4852 .resize = raid5_resize,
4853 #ifdef CONFIG_MD_RAID5_RESHAPE
4854 .check_reshape = raid5_check_reshape,
4855 .start_reshape = raid5_start_reshape,
4857 .quiesce = raid5_quiesce,
4860 static struct mdk_personality raid4_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,
4882 static int __init raid5_init(void)
4886 e = raid6_select_algo();
4889 register_md_personality(&raid6_personality);
4890 register_md_personality(&raid5_personality);
4891 register_md_personality(&raid4_personality);
4895 static void raid5_exit(void)
4897 unregister_md_personality(&raid6_personality);
4898 unregister_md_personality(&raid5_personality);
4899 unregister_md_personality(&raid4_personality);
4902 module_init(raid5_init);
4903 module_exit(raid5_exit);
4904 MODULE_LICENSE("GPL");
4905 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4906 MODULE_ALIAS("md-raid5");
4907 MODULE_ALIAS("md-raid4");
4908 MODULE_ALIAS("md-level-5");
4909 MODULE_ALIAS("md-level-4");
4910 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4911 MODULE_ALIAS("md-raid6");
4912 MODULE_ALIAS("md-level-6");
4914 /* This used to be two separate modules, they were: */
4915 MODULE_ALIAS("raid5");
4916 MODULE_ALIAS("raid6");