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, struct stripe_head_state *s)
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 (s->syncing || s->expanding || s->expanded)
429 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
431 set_bit(STRIPE_IO_STARTED, &sh->state);
433 bi->bi_bdev = rdev->bdev;
434 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
435 __func__, (unsigned long long)sh->sector,
437 atomic_inc(&sh->count);
438 bi->bi_sector = sh->sector + rdev->data_offset;
439 bi->bi_flags = 1 << BIO_UPTODATE;
443 bi->bi_io_vec = &sh->dev[i].vec;
444 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
445 bi->bi_io_vec[0].bv_offset = 0;
446 bi->bi_size = STRIPE_SIZE;
449 test_bit(R5_ReWrite, &sh->dev[i].flags))
450 atomic_add(STRIPE_SECTORS,
451 &rdev->corrected_errors);
452 generic_make_request(bi);
455 set_bit(STRIPE_DEGRADED, &sh->state);
456 pr_debug("skip op %ld on disc %d for sector %llu\n",
457 bi->bi_rw, i, (unsigned long long)sh->sector);
458 clear_bit(R5_LOCKED, &sh->dev[i].flags);
459 set_bit(STRIPE_HANDLE, &sh->state);
464 static struct dma_async_tx_descriptor *
465 async_copy_data(int frombio, struct bio *bio, struct page *page,
466 sector_t sector, struct dma_async_tx_descriptor *tx)
469 struct page *bio_page;
473 if (bio->bi_sector >= sector)
474 page_offset = (signed)(bio->bi_sector - sector) * 512;
476 page_offset = (signed)(sector - bio->bi_sector) * -512;
477 bio_for_each_segment(bvl, bio, i) {
478 int len = bio_iovec_idx(bio, i)->bv_len;
482 if (page_offset < 0) {
483 b_offset = -page_offset;
484 page_offset += b_offset;
488 if (len > 0 && page_offset + len > STRIPE_SIZE)
489 clen = STRIPE_SIZE - page_offset;
494 b_offset += bio_iovec_idx(bio, i)->bv_offset;
495 bio_page = bio_iovec_idx(bio, i)->bv_page;
497 tx = async_memcpy(page, bio_page, page_offset,
502 tx = async_memcpy(bio_page, page, b_offset,
507 if (clen < len) /* hit end of page */
515 static void ops_complete_biofill(void *stripe_head_ref)
517 struct stripe_head *sh = stripe_head_ref;
518 struct bio *return_bi = NULL;
519 raid5_conf_t *conf = sh->raid_conf;
522 pr_debug("%s: stripe %llu\n", __func__,
523 (unsigned long long)sh->sector);
525 /* clear completed biofills */
526 for (i = sh->disks; i--; ) {
527 struct r5dev *dev = &sh->dev[i];
529 /* acknowledge completion of a biofill operation */
530 /* and check if we need to reply to a read request,
531 * new R5_Wantfill requests are held off until
532 * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
534 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
535 struct bio *rbi, *rbi2;
537 /* The access to dev->read is outside of the
538 * spin_lock_irq(&conf->device_lock), but is protected
539 * by the STRIPE_OP_BIOFILL pending bit
544 while (rbi && rbi->bi_sector <
545 dev->sector + STRIPE_SECTORS) {
546 rbi2 = r5_next_bio(rbi, dev->sector);
547 spin_lock_irq(&conf->device_lock);
548 if (--rbi->bi_phys_segments == 0) {
549 rbi->bi_next = return_bi;
552 spin_unlock_irq(&conf->device_lock);
557 set_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
559 return_io(return_bi);
561 set_bit(STRIPE_HANDLE, &sh->state);
565 static void ops_run_biofill(struct stripe_head *sh)
567 struct dma_async_tx_descriptor *tx = NULL;
568 raid5_conf_t *conf = sh->raid_conf;
571 pr_debug("%s: stripe %llu\n", __func__,
572 (unsigned long long)sh->sector);
574 for (i = sh->disks; i--; ) {
575 struct r5dev *dev = &sh->dev[i];
576 if (test_bit(R5_Wantfill, &dev->flags)) {
578 spin_lock_irq(&conf->device_lock);
579 dev->read = rbi = dev->toread;
581 spin_unlock_irq(&conf->device_lock);
582 while (rbi && rbi->bi_sector <
583 dev->sector + STRIPE_SECTORS) {
584 tx = async_copy_data(0, rbi, dev->page,
586 rbi = r5_next_bio(rbi, dev->sector);
591 atomic_inc(&sh->count);
592 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
593 ops_complete_biofill, sh);
596 static void ops_complete_compute5(void *stripe_head_ref)
598 struct stripe_head *sh = stripe_head_ref;
599 int target = sh->ops.target;
600 struct r5dev *tgt = &sh->dev[target];
602 pr_debug("%s: stripe %llu\n", __func__,
603 (unsigned long long)sh->sector);
605 set_bit(R5_UPTODATE, &tgt->flags);
606 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
607 clear_bit(R5_Wantcompute, &tgt->flags);
608 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
609 if (sh->check_state == check_state_compute_run)
610 sh->check_state = check_state_compute_result;
612 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
613 set_bit(STRIPE_HANDLE, &sh->state);
617 static struct dma_async_tx_descriptor *
618 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
620 /* kernel stack size limits the total number of disks */
621 int disks = sh->disks;
622 struct page *xor_srcs[disks];
623 int target = sh->ops.target;
624 struct r5dev *tgt = &sh->dev[target];
625 struct page *xor_dest = tgt->page;
627 struct dma_async_tx_descriptor *tx;
630 pr_debug("%s: stripe %llu block: %d\n",
631 __func__, (unsigned long long)sh->sector, target);
632 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
634 for (i = disks; i--; )
636 xor_srcs[count++] = sh->dev[i].page;
638 atomic_inc(&sh->count);
640 if (unlikely(count == 1))
641 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
642 0, NULL, ops_complete_compute5, sh);
644 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
645 ASYNC_TX_XOR_ZERO_DST, NULL,
646 ops_complete_compute5, sh);
648 /* ack now if postxor is not set to be run */
649 if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
655 static void ops_complete_prexor(void *stripe_head_ref)
657 struct stripe_head *sh = stripe_head_ref;
659 pr_debug("%s: stripe %llu\n", __func__,
660 (unsigned long long)sh->sector);
662 set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
665 static struct dma_async_tx_descriptor *
666 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
668 /* kernel stack size limits the total number of disks */
669 int disks = sh->disks;
670 struct page *xor_srcs[disks];
671 int count = 0, pd_idx = sh->pd_idx, i;
673 /* existing parity data subtracted */
674 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
676 pr_debug("%s: stripe %llu\n", __func__,
677 (unsigned long long)sh->sector);
679 for (i = disks; i--; ) {
680 struct r5dev *dev = &sh->dev[i];
681 /* Only process blocks that are known to be uptodate */
682 if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
683 xor_srcs[count++] = dev->page;
686 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
687 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
688 ops_complete_prexor, sh);
693 static struct dma_async_tx_descriptor *
694 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
695 unsigned long pending)
697 int disks = sh->disks;
698 int pd_idx = sh->pd_idx, i;
700 /* check if prexor is active which means only process blocks
701 * that are part of a read-modify-write (Wantprexor)
703 int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
705 pr_debug("%s: stripe %llu\n", __func__,
706 (unsigned long long)sh->sector);
708 for (i = disks; i--; ) {
709 struct r5dev *dev = &sh->dev[i];
714 if (prexor) { /* rmw */
716 test_bit(R5_Wantprexor, &dev->flags))
719 if (i != pd_idx && dev->towrite &&
720 test_bit(R5_LOCKED, &dev->flags))
727 spin_lock(&sh->lock);
728 chosen = dev->towrite;
730 BUG_ON(dev->written);
731 wbi = dev->written = chosen;
732 spin_unlock(&sh->lock);
734 while (wbi && wbi->bi_sector <
735 dev->sector + STRIPE_SECTORS) {
736 tx = async_copy_data(1, wbi, dev->page,
738 wbi = r5_next_bio(wbi, dev->sector);
746 static void ops_complete_postxor(void *stripe_head_ref)
748 struct stripe_head *sh = stripe_head_ref;
750 pr_debug("%s: stripe %llu\n", __func__,
751 (unsigned long long)sh->sector);
753 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
754 set_bit(STRIPE_HANDLE, &sh->state);
758 static void ops_complete_write(void *stripe_head_ref)
760 struct stripe_head *sh = stripe_head_ref;
761 int disks = sh->disks, i, pd_idx = sh->pd_idx;
763 pr_debug("%s: stripe %llu\n", __func__,
764 (unsigned long long)sh->sector);
766 for (i = disks; i--; ) {
767 struct r5dev *dev = &sh->dev[i];
768 if (dev->written || i == pd_idx)
769 set_bit(R5_UPTODATE, &dev->flags);
772 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
773 set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
775 set_bit(STRIPE_HANDLE, &sh->state);
780 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
781 unsigned long pending)
783 /* kernel stack size limits the total number of disks */
784 int disks = sh->disks;
785 struct page *xor_srcs[disks];
787 int count = 0, pd_idx = sh->pd_idx, i;
788 struct page *xor_dest;
789 int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
791 dma_async_tx_callback callback;
793 pr_debug("%s: stripe %llu\n", __func__,
794 (unsigned long long)sh->sector);
796 /* check if prexor is active which means only process blocks
797 * that are part of a read-modify-write (written)
800 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
801 for (i = disks; i--; ) {
802 struct r5dev *dev = &sh->dev[i];
804 xor_srcs[count++] = dev->page;
807 xor_dest = sh->dev[pd_idx].page;
808 for (i = disks; i--; ) {
809 struct r5dev *dev = &sh->dev[i];
811 xor_srcs[count++] = dev->page;
815 /* check whether this postxor is part of a write */
816 callback = test_bit(STRIPE_OP_BIODRAIN, &pending) ?
817 ops_complete_write : ops_complete_postxor;
819 /* 1/ if we prexor'd then the dest is reused as a source
820 * 2/ if we did not prexor then we are redoing the parity
821 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
822 * for the synchronous xor case
824 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
825 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
827 atomic_inc(&sh->count);
829 if (unlikely(count == 1)) {
830 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
831 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
832 flags, tx, callback, sh);
834 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
835 flags, tx, callback, sh);
838 static void ops_complete_check(void *stripe_head_ref)
840 struct stripe_head *sh = stripe_head_ref;
842 pr_debug("%s: stripe %llu\n", __func__,
843 (unsigned long long)sh->sector);
845 sh->check_state = check_state_check_result;
846 set_bit(STRIPE_HANDLE, &sh->state);
850 static void ops_run_check(struct stripe_head *sh)
852 /* kernel stack size limits the total number of disks */
853 int disks = sh->disks;
854 struct page *xor_srcs[disks];
855 struct dma_async_tx_descriptor *tx;
857 int count = 0, pd_idx = sh->pd_idx, i;
858 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
860 pr_debug("%s: stripe %llu\n", __func__,
861 (unsigned long long)sh->sector);
863 for (i = disks; i--; ) {
864 struct r5dev *dev = &sh->dev[i];
866 xor_srcs[count++] = dev->page;
869 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
870 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
872 atomic_inc(&sh->count);
873 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
874 ops_complete_check, sh);
877 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending,
878 unsigned long ops_request)
880 int overlap_clear = 0, i, disks = sh->disks;
881 struct dma_async_tx_descriptor *tx = NULL;
883 if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
888 if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending) ||
889 test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request))
890 tx = ops_run_compute5(sh, pending);
892 if (test_bit(STRIPE_OP_PREXOR, &pending))
893 tx = ops_run_prexor(sh, tx);
895 if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
896 tx = ops_run_biodrain(sh, tx, pending);
900 if (test_bit(STRIPE_OP_POSTXOR, &pending))
901 ops_run_postxor(sh, tx, pending);
903 if (test_bit(STRIPE_OP_CHECK, &ops_request))
907 for (i = disks; i--; ) {
908 struct r5dev *dev = &sh->dev[i];
909 if (test_and_clear_bit(R5_Overlap, &dev->flags))
910 wake_up(&sh->raid_conf->wait_for_overlap);
914 static int grow_one_stripe(raid5_conf_t *conf)
916 struct stripe_head *sh;
917 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
920 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
921 sh->raid_conf = conf;
922 spin_lock_init(&sh->lock);
924 if (grow_buffers(sh, conf->raid_disks)) {
925 shrink_buffers(sh, conf->raid_disks);
926 kmem_cache_free(conf->slab_cache, sh);
929 sh->disks = conf->raid_disks;
930 /* we just created an active stripe so... */
931 atomic_set(&sh->count, 1);
932 atomic_inc(&conf->active_stripes);
933 INIT_LIST_HEAD(&sh->lru);
938 static int grow_stripes(raid5_conf_t *conf, int num)
940 struct kmem_cache *sc;
941 int devs = conf->raid_disks;
943 sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
944 sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
945 conf->active_name = 0;
946 sc = kmem_cache_create(conf->cache_name[conf->active_name],
947 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
951 conf->slab_cache = sc;
952 conf->pool_size = devs;
954 if (!grow_one_stripe(conf))
959 #ifdef CONFIG_MD_RAID5_RESHAPE
960 static int resize_stripes(raid5_conf_t *conf, int newsize)
962 /* Make all the stripes able to hold 'newsize' devices.
963 * New slots in each stripe get 'page' set to a new page.
965 * This happens in stages:
966 * 1/ create a new kmem_cache and allocate the required number of
968 * 2/ gather all the old stripe_heads and tranfer the pages across
969 * to the new stripe_heads. This will have the side effect of
970 * freezing the array as once all stripe_heads have been collected,
971 * no IO will be possible. Old stripe heads are freed once their
972 * pages have been transferred over, and the old kmem_cache is
973 * freed when all stripes are done.
974 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
975 * we simple return a failre status - no need to clean anything up.
976 * 4/ allocate new pages for the new slots in the new stripe_heads.
977 * If this fails, we don't bother trying the shrink the
978 * stripe_heads down again, we just leave them as they are.
979 * As each stripe_head is processed the new one is released into
982 * Once step2 is started, we cannot afford to wait for a write,
983 * so we use GFP_NOIO allocations.
985 struct stripe_head *osh, *nsh;
986 LIST_HEAD(newstripes);
987 struct disk_info *ndisks;
989 struct kmem_cache *sc;
992 if (newsize <= conf->pool_size)
993 return 0; /* never bother to shrink */
995 md_allow_write(conf->mddev);
998 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
999 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1004 for (i = conf->max_nr_stripes; i; i--) {
1005 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1009 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1011 nsh->raid_conf = conf;
1012 spin_lock_init(&nsh->lock);
1014 list_add(&nsh->lru, &newstripes);
1017 /* didn't get enough, give up */
1018 while (!list_empty(&newstripes)) {
1019 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1020 list_del(&nsh->lru);
1021 kmem_cache_free(sc, nsh);
1023 kmem_cache_destroy(sc);
1026 /* Step 2 - Must use GFP_NOIO now.
1027 * OK, we have enough stripes, start collecting inactive
1028 * stripes and copying them over
1030 list_for_each_entry(nsh, &newstripes, lru) {
1031 spin_lock_irq(&conf->device_lock);
1032 wait_event_lock_irq(conf->wait_for_stripe,
1033 !list_empty(&conf->inactive_list),
1035 unplug_slaves(conf->mddev)
1037 osh = get_free_stripe(conf);
1038 spin_unlock_irq(&conf->device_lock);
1039 atomic_set(&nsh->count, 1);
1040 for(i=0; i<conf->pool_size; i++)
1041 nsh->dev[i].page = osh->dev[i].page;
1042 for( ; i<newsize; i++)
1043 nsh->dev[i].page = NULL;
1044 kmem_cache_free(conf->slab_cache, osh);
1046 kmem_cache_destroy(conf->slab_cache);
1049 * At this point, we are holding all the stripes so the array
1050 * is completely stalled, so now is a good time to resize
1053 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1055 for (i=0; i<conf->raid_disks; i++)
1056 ndisks[i] = conf->disks[i];
1058 conf->disks = ndisks;
1062 /* Step 4, return new stripes to service */
1063 while(!list_empty(&newstripes)) {
1064 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1065 list_del_init(&nsh->lru);
1066 for (i=conf->raid_disks; i < newsize; i++)
1067 if (nsh->dev[i].page == NULL) {
1068 struct page *p = alloc_page(GFP_NOIO);
1069 nsh->dev[i].page = p;
1073 release_stripe(nsh);
1075 /* critical section pass, GFP_NOIO no longer needed */
1077 conf->slab_cache = sc;
1078 conf->active_name = 1-conf->active_name;
1079 conf->pool_size = newsize;
1084 static int drop_one_stripe(raid5_conf_t *conf)
1086 struct stripe_head *sh;
1088 spin_lock_irq(&conf->device_lock);
1089 sh = get_free_stripe(conf);
1090 spin_unlock_irq(&conf->device_lock);
1093 BUG_ON(atomic_read(&sh->count));
1094 shrink_buffers(sh, conf->pool_size);
1095 kmem_cache_free(conf->slab_cache, sh);
1096 atomic_dec(&conf->active_stripes);
1100 static void shrink_stripes(raid5_conf_t *conf)
1102 while (drop_one_stripe(conf))
1105 if (conf->slab_cache)
1106 kmem_cache_destroy(conf->slab_cache);
1107 conf->slab_cache = NULL;
1110 static void raid5_end_read_request(struct bio * bi, int error)
1112 struct stripe_head *sh = bi->bi_private;
1113 raid5_conf_t *conf = sh->raid_conf;
1114 int disks = sh->disks, i;
1115 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1116 char b[BDEVNAME_SIZE];
1120 for (i=0 ; i<disks; i++)
1121 if (bi == &sh->dev[i].req)
1124 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1125 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1133 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1134 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1135 rdev = conf->disks[i].rdev;
1136 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1137 " (%lu sectors at %llu on %s)\n",
1138 mdname(conf->mddev), STRIPE_SECTORS,
1139 (unsigned long long)(sh->sector
1140 + rdev->data_offset),
1141 bdevname(rdev->bdev, b));
1142 clear_bit(R5_ReadError, &sh->dev[i].flags);
1143 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1145 if (atomic_read(&conf->disks[i].rdev->read_errors))
1146 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1148 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1150 rdev = conf->disks[i].rdev;
1152 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1153 atomic_inc(&rdev->read_errors);
1154 if (conf->mddev->degraded)
1155 printk_rl(KERN_WARNING
1156 "raid5:%s: read error not correctable "
1157 "(sector %llu on %s).\n",
1158 mdname(conf->mddev),
1159 (unsigned long long)(sh->sector
1160 + rdev->data_offset),
1162 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1164 printk_rl(KERN_WARNING
1165 "raid5:%s: read error NOT corrected!! "
1166 "(sector %llu on %s).\n",
1167 mdname(conf->mddev),
1168 (unsigned long long)(sh->sector
1169 + rdev->data_offset),
1171 else if (atomic_read(&rdev->read_errors)
1172 > conf->max_nr_stripes)
1174 "raid5:%s: Too many read errors, failing device %s.\n",
1175 mdname(conf->mddev), bdn);
1179 set_bit(R5_ReadError, &sh->dev[i].flags);
1181 clear_bit(R5_ReadError, &sh->dev[i].flags);
1182 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1183 md_error(conf->mddev, rdev);
1186 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1187 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1188 set_bit(STRIPE_HANDLE, &sh->state);
1192 static void raid5_end_write_request (struct bio *bi, int error)
1194 struct stripe_head *sh = bi->bi_private;
1195 raid5_conf_t *conf = sh->raid_conf;
1196 int disks = sh->disks, i;
1197 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1199 for (i=0 ; i<disks; i++)
1200 if (bi == &sh->dev[i].req)
1203 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1204 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1212 md_error(conf->mddev, conf->disks[i].rdev);
1214 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1216 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1217 set_bit(STRIPE_HANDLE, &sh->state);
1222 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1224 static void raid5_build_block (struct stripe_head *sh, int i)
1226 struct r5dev *dev = &sh->dev[i];
1228 bio_init(&dev->req);
1229 dev->req.bi_io_vec = &dev->vec;
1231 dev->req.bi_max_vecs++;
1232 dev->vec.bv_page = dev->page;
1233 dev->vec.bv_len = STRIPE_SIZE;
1234 dev->vec.bv_offset = 0;
1236 dev->req.bi_sector = sh->sector;
1237 dev->req.bi_private = sh;
1240 dev->sector = compute_blocknr(sh, i);
1243 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1245 char b[BDEVNAME_SIZE];
1246 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1247 pr_debug("raid5: error called\n");
1249 if (!test_bit(Faulty, &rdev->flags)) {
1250 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1251 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1252 unsigned long flags;
1253 spin_lock_irqsave(&conf->device_lock, flags);
1255 spin_unlock_irqrestore(&conf->device_lock, flags);
1257 * if recovery was running, make sure it aborts.
1259 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1261 set_bit(Faulty, &rdev->flags);
1263 "raid5: Disk failure on %s, disabling device.\n"
1264 "raid5: Operation continuing on %d devices.\n",
1265 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1270 * Input: a 'big' sector number,
1271 * Output: index of the data and parity disk, and the sector # in them.
1273 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1274 unsigned int data_disks, unsigned int * dd_idx,
1275 unsigned int * pd_idx, raid5_conf_t *conf)
1278 unsigned long chunk_number;
1279 unsigned int chunk_offset;
1280 sector_t new_sector;
1281 int sectors_per_chunk = conf->chunk_size >> 9;
1283 /* First compute the information on this sector */
1286 * Compute the chunk number and the sector offset inside the chunk
1288 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1289 chunk_number = r_sector;
1290 BUG_ON(r_sector != chunk_number);
1293 * Compute the stripe number
1295 stripe = chunk_number / data_disks;
1298 * Compute the data disk and parity disk indexes inside the stripe
1300 *dd_idx = chunk_number % data_disks;
1303 * Select the parity disk based on the user selected algorithm.
1305 switch(conf->level) {
1307 *pd_idx = data_disks;
1310 switch (conf->algorithm) {
1311 case ALGORITHM_LEFT_ASYMMETRIC:
1312 *pd_idx = data_disks - stripe % raid_disks;
1313 if (*dd_idx >= *pd_idx)
1316 case ALGORITHM_RIGHT_ASYMMETRIC:
1317 *pd_idx = stripe % raid_disks;
1318 if (*dd_idx >= *pd_idx)
1321 case ALGORITHM_LEFT_SYMMETRIC:
1322 *pd_idx = data_disks - stripe % raid_disks;
1323 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1325 case ALGORITHM_RIGHT_SYMMETRIC:
1326 *pd_idx = stripe % raid_disks;
1327 *dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1330 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1336 /**** FIX THIS ****/
1337 switch (conf->algorithm) {
1338 case ALGORITHM_LEFT_ASYMMETRIC:
1339 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1340 if (*pd_idx == raid_disks-1)
1341 (*dd_idx)++; /* Q D D D P */
1342 else if (*dd_idx >= *pd_idx)
1343 (*dd_idx) += 2; /* D D P Q D */
1345 case ALGORITHM_RIGHT_ASYMMETRIC:
1346 *pd_idx = stripe % raid_disks;
1347 if (*pd_idx == raid_disks-1)
1348 (*dd_idx)++; /* Q D D D P */
1349 else if (*dd_idx >= *pd_idx)
1350 (*dd_idx) += 2; /* D D P Q D */
1352 case ALGORITHM_LEFT_SYMMETRIC:
1353 *pd_idx = raid_disks - 1 - (stripe % raid_disks);
1354 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1356 case ALGORITHM_RIGHT_SYMMETRIC:
1357 *pd_idx = stripe % raid_disks;
1358 *dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1361 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1368 * Finally, compute the new sector number
1370 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1375 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1377 raid5_conf_t *conf = sh->raid_conf;
1378 int raid_disks = sh->disks;
1379 int data_disks = raid_disks - conf->max_degraded;
1380 sector_t new_sector = sh->sector, check;
1381 int sectors_per_chunk = conf->chunk_size >> 9;
1384 int chunk_number, dummy1, dummy2, dd_idx = i;
1388 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1389 stripe = new_sector;
1390 BUG_ON(new_sector != stripe);
1392 if (i == sh->pd_idx)
1394 switch(conf->level) {
1397 switch (conf->algorithm) {
1398 case ALGORITHM_LEFT_ASYMMETRIC:
1399 case ALGORITHM_RIGHT_ASYMMETRIC:
1403 case ALGORITHM_LEFT_SYMMETRIC:
1404 case ALGORITHM_RIGHT_SYMMETRIC:
1407 i -= (sh->pd_idx + 1);
1410 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1415 if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1416 return 0; /* It is the Q disk */
1417 switch (conf->algorithm) {
1418 case ALGORITHM_LEFT_ASYMMETRIC:
1419 case ALGORITHM_RIGHT_ASYMMETRIC:
1420 if (sh->pd_idx == raid_disks-1)
1421 i--; /* Q D D D P */
1422 else if (i > sh->pd_idx)
1423 i -= 2; /* D D P Q D */
1425 case ALGORITHM_LEFT_SYMMETRIC:
1426 case ALGORITHM_RIGHT_SYMMETRIC:
1427 if (sh->pd_idx == raid_disks-1)
1428 i--; /* Q D D D P */
1433 i -= (sh->pd_idx + 2);
1437 printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1443 chunk_number = stripe * data_disks + i;
1444 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1446 check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1447 if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1448 printk(KERN_ERR "compute_blocknr: map not correct\n");
1457 * Copy data between a page in the stripe cache, and one or more bion
1458 * The page could align with the middle of the bio, or there could be
1459 * several bion, each with several bio_vecs, which cover part of the page
1460 * Multiple bion are linked together on bi_next. There may be extras
1461 * at the end of this list. We ignore them.
1463 static void copy_data(int frombio, struct bio *bio,
1467 char *pa = page_address(page);
1468 struct bio_vec *bvl;
1472 if (bio->bi_sector >= sector)
1473 page_offset = (signed)(bio->bi_sector - sector) * 512;
1475 page_offset = (signed)(sector - bio->bi_sector) * -512;
1476 bio_for_each_segment(bvl, bio, i) {
1477 int len = bio_iovec_idx(bio,i)->bv_len;
1481 if (page_offset < 0) {
1482 b_offset = -page_offset;
1483 page_offset += b_offset;
1487 if (len > 0 && page_offset + len > STRIPE_SIZE)
1488 clen = STRIPE_SIZE - page_offset;
1492 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1494 memcpy(pa+page_offset, ba+b_offset, clen);
1496 memcpy(ba+b_offset, pa+page_offset, clen);
1497 __bio_kunmap_atomic(ba, KM_USER0);
1499 if (clen < len) /* hit end of page */
1505 #define check_xor() do { \
1506 if (count == MAX_XOR_BLOCKS) { \
1507 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1512 static void compute_parity6(struct stripe_head *sh, int method)
1514 raid6_conf_t *conf = sh->raid_conf;
1515 int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1517 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1520 qd_idx = raid6_next_disk(pd_idx, disks);
1521 d0_idx = raid6_next_disk(qd_idx, disks);
1523 pr_debug("compute_parity, stripe %llu, method %d\n",
1524 (unsigned long long)sh->sector, method);
1527 case READ_MODIFY_WRITE:
1528 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1529 case RECONSTRUCT_WRITE:
1530 for (i= disks; i-- ;)
1531 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1532 chosen = sh->dev[i].towrite;
1533 sh->dev[i].towrite = NULL;
1535 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1536 wake_up(&conf->wait_for_overlap);
1538 BUG_ON(sh->dev[i].written);
1539 sh->dev[i].written = chosen;
1543 BUG(); /* Not implemented yet */
1546 for (i = disks; i--;)
1547 if (sh->dev[i].written) {
1548 sector_t sector = sh->dev[i].sector;
1549 struct bio *wbi = sh->dev[i].written;
1550 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1551 copy_data(1, wbi, sh->dev[i].page, sector);
1552 wbi = r5_next_bio(wbi, sector);
1555 set_bit(R5_LOCKED, &sh->dev[i].flags);
1556 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1560 // case RECONSTRUCT_WRITE:
1561 // case CHECK_PARITY:
1562 // case UPDATE_PARITY:
1563 /* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1564 /* FIX: Is this ordering of drives even remotely optimal? */
1568 ptrs[count++] = page_address(sh->dev[i].page);
1569 if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1570 printk("block %d/%d not uptodate on parity calc\n", i,count);
1571 i = raid6_next_disk(i, disks);
1572 } while ( i != d0_idx );
1576 raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1579 case RECONSTRUCT_WRITE:
1580 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1581 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1582 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1583 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1586 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1587 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1593 /* Compute one missing block */
1594 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1596 int i, count, disks = sh->disks;
1597 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1598 int pd_idx = sh->pd_idx;
1599 int qd_idx = raid6_next_disk(pd_idx, disks);
1601 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1602 (unsigned long long)sh->sector, dd_idx);
1604 if ( dd_idx == qd_idx ) {
1605 /* We're actually computing the Q drive */
1606 compute_parity6(sh, UPDATE_PARITY);
1608 dest = page_address(sh->dev[dd_idx].page);
1609 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1611 for (i = disks ; i--; ) {
1612 if (i == dd_idx || i == qd_idx)
1614 p = page_address(sh->dev[i].page);
1615 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1618 printk("compute_block() %d, stripe %llu, %d"
1619 " not present\n", dd_idx,
1620 (unsigned long long)sh->sector, i);
1625 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1626 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1627 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1631 /* Compute two missing blocks */
1632 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1634 int i, count, disks = sh->disks;
1635 int pd_idx = sh->pd_idx;
1636 int qd_idx = raid6_next_disk(pd_idx, disks);
1637 int d0_idx = raid6_next_disk(qd_idx, disks);
1640 /* faila and failb are disk numbers relative to d0_idx */
1641 /* pd_idx become disks-2 and qd_idx become disks-1 */
1642 faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1643 failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1645 BUG_ON(faila == failb);
1646 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1648 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1649 (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1651 if ( failb == disks-1 ) {
1652 /* Q disk is one of the missing disks */
1653 if ( faila == disks-2 ) {
1654 /* Missing P+Q, just recompute */
1655 compute_parity6(sh, UPDATE_PARITY);
1658 /* We're missing D+Q; recompute D from P */
1659 compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1660 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1665 /* We're missing D+P or D+D; build pointer table */
1667 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1673 ptrs[count++] = page_address(sh->dev[i].page);
1674 i = raid6_next_disk(i, disks);
1675 if (i != dd_idx1 && i != dd_idx2 &&
1676 !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1677 printk("compute_2 with missing block %d/%d\n", count, i);
1678 } while ( i != d0_idx );
1680 if ( failb == disks-2 ) {
1681 /* We're missing D+P. */
1682 raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1684 /* We're missing D+D. */
1685 raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1688 /* Both the above update both missing blocks */
1689 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1690 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1695 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1697 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1701 /* if we are not expanding this is a proper write request, and
1702 * there will be bios with new data to be drained into the
1706 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1710 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1713 for (i = disks; i--; ) {
1714 struct r5dev *dev = &sh->dev[i];
1717 set_bit(R5_LOCKED, &dev->flags);
1719 clear_bit(R5_UPTODATE, &dev->flags);
1723 if (locked + 1 == disks)
1724 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1725 atomic_inc(&sh->raid_conf->pending_full_writes);
1727 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1728 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1730 set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1731 set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1732 set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1736 for (i = disks; i--; ) {
1737 struct r5dev *dev = &sh->dev[i];
1741 /* For a read-modify write there may be blocks that are
1742 * locked for reading while others are ready to be
1743 * written so we distinguish these blocks by the
1747 (test_bit(R5_UPTODATE, &dev->flags) ||
1748 test_bit(R5_Wantcompute, &dev->flags))) {
1749 set_bit(R5_Wantprexor, &dev->flags);
1750 set_bit(R5_LOCKED, &dev->flags);
1751 clear_bit(R5_UPTODATE, &dev->flags);
1757 /* keep the parity disk locked while asynchronous operations
1760 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1761 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1764 pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1765 __func__, (unsigned long long)sh->sector,
1766 locked, sh->ops.pending);
1772 * Each stripe/dev can have one or more bion attached.
1773 * toread/towrite point to the first in a chain.
1774 * The bi_next chain must be in order.
1776 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1779 raid5_conf_t *conf = sh->raid_conf;
1782 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1783 (unsigned long long)bi->bi_sector,
1784 (unsigned long long)sh->sector);
1787 spin_lock(&sh->lock);
1788 spin_lock_irq(&conf->device_lock);
1790 bip = &sh->dev[dd_idx].towrite;
1791 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1794 bip = &sh->dev[dd_idx].toread;
1795 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1796 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1798 bip = & (*bip)->bi_next;
1800 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1803 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1807 bi->bi_phys_segments ++;
1808 spin_unlock_irq(&conf->device_lock);
1809 spin_unlock(&sh->lock);
1811 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1812 (unsigned long long)bi->bi_sector,
1813 (unsigned long long)sh->sector, dd_idx);
1815 if (conf->mddev->bitmap && firstwrite) {
1816 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1818 sh->bm_seq = conf->seq_flush+1;
1819 set_bit(STRIPE_BIT_DELAY, &sh->state);
1823 /* check if page is covered */
1824 sector_t sector = sh->dev[dd_idx].sector;
1825 for (bi=sh->dev[dd_idx].towrite;
1826 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1827 bi && bi->bi_sector <= sector;
1828 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1829 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1830 sector = bi->bi_sector + (bi->bi_size>>9);
1832 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1833 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1838 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1839 spin_unlock_irq(&conf->device_lock);
1840 spin_unlock(&sh->lock);
1844 static void end_reshape(raid5_conf_t *conf);
1846 static int page_is_zero(struct page *p)
1848 char *a = page_address(p);
1849 return ((*(u32*)a) == 0 &&
1850 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1853 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1855 int sectors_per_chunk = conf->chunk_size >> 9;
1857 int chunk_offset = sector_div(stripe, sectors_per_chunk);
1859 raid5_compute_sector(stripe * (disks - conf->max_degraded)
1860 *sectors_per_chunk + chunk_offset,
1861 disks, disks - conf->max_degraded,
1862 &dd_idx, &pd_idx, conf);
1867 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1868 struct stripe_head_state *s, int disks,
1869 struct bio **return_bi)
1872 for (i = disks; i--; ) {
1876 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1879 rdev = rcu_dereference(conf->disks[i].rdev);
1880 if (rdev && test_bit(In_sync, &rdev->flags))
1881 /* multiple read failures in one stripe */
1882 md_error(conf->mddev, rdev);
1885 spin_lock_irq(&conf->device_lock);
1886 /* fail all writes first */
1887 bi = sh->dev[i].towrite;
1888 sh->dev[i].towrite = NULL;
1894 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1895 wake_up(&conf->wait_for_overlap);
1897 while (bi && bi->bi_sector <
1898 sh->dev[i].sector + STRIPE_SECTORS) {
1899 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1900 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1901 if (--bi->bi_phys_segments == 0) {
1902 md_write_end(conf->mddev);
1903 bi->bi_next = *return_bi;
1908 /* and fail all 'written' */
1909 bi = sh->dev[i].written;
1910 sh->dev[i].written = NULL;
1911 if (bi) bitmap_end = 1;
1912 while (bi && bi->bi_sector <
1913 sh->dev[i].sector + STRIPE_SECTORS) {
1914 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1915 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1916 if (--bi->bi_phys_segments == 0) {
1917 md_write_end(conf->mddev);
1918 bi->bi_next = *return_bi;
1924 /* fail any reads if this device is non-operational and
1925 * the data has not reached the cache yet.
1927 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1928 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1929 test_bit(R5_ReadError, &sh->dev[i].flags))) {
1930 bi = sh->dev[i].toread;
1931 sh->dev[i].toread = NULL;
1932 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1933 wake_up(&conf->wait_for_overlap);
1934 if (bi) s->to_read--;
1935 while (bi && bi->bi_sector <
1936 sh->dev[i].sector + STRIPE_SECTORS) {
1937 struct bio *nextbi =
1938 r5_next_bio(bi, sh->dev[i].sector);
1939 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1940 if (--bi->bi_phys_segments == 0) {
1941 bi->bi_next = *return_bi;
1947 spin_unlock_irq(&conf->device_lock);
1949 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1950 STRIPE_SECTORS, 0, 0);
1953 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
1954 if (atomic_dec_and_test(&conf->pending_full_writes))
1955 md_wakeup_thread(conf->mddev->thread);
1958 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1961 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1962 struct stripe_head_state *s, int disk_idx, int disks)
1964 struct r5dev *dev = &sh->dev[disk_idx];
1965 struct r5dev *failed_dev = &sh->dev[s->failed_num];
1967 /* don't schedule compute operations or reads on the parity block while
1968 * a check is in flight
1970 if (disk_idx == sh->pd_idx && sh->check_state)
1973 /* is the data in this block needed, and can we get it? */
1974 if (!test_bit(R5_LOCKED, &dev->flags) &&
1975 !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1976 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1977 s->syncing || s->expanding || (s->failed &&
1978 (failed_dev->toread || (failed_dev->towrite &&
1979 !test_bit(R5_OVERWRITE, &failed_dev->flags)
1981 /* 1/ We would like to get this block, possibly by computing it,
1982 * but we might not be able to.
1984 * 2/ Since parity check operations potentially make the parity
1985 * block !uptodate it will need to be refreshed before any
1986 * compute operations on data disks are scheduled.
1988 * 3/ We hold off parity block re-reads until check operations
1991 if ((s->uptodate == disks - 1) && !sh->check_state &&
1992 (s->failed && disk_idx == s->failed_num)) {
1993 set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1994 set_bit(R5_Wantcompute, &dev->flags);
1995 sh->ops.target = disk_idx;
1998 /* Careful: from this point on 'uptodate' is in the eye
1999 * of raid5_run_ops which services 'compute' operations
2000 * before writes. R5_Wantcompute flags a block that will
2001 * be R5_UPTODATE by the time it is needed for a
2002 * subsequent operation.
2005 return 0; /* uptodate + compute == disks */
2006 } else if ((s->uptodate < disks - 1) &&
2007 test_bit(R5_Insync, &dev->flags)) {
2008 /* Note: we hold off compute operations while checks are
2009 * in flight, but we still prefer 'compute' over 'read'
2010 * hence we only read if (uptodate < * disks-1)
2012 set_bit(R5_LOCKED, &dev->flags);
2013 set_bit(R5_Wantread, &dev->flags);
2015 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2023 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2024 struct stripe_head_state *s, int disks)
2028 /* Clear completed compute operations */
2029 if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete)) {
2030 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2031 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2032 clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2035 /* look for blocks to read/compute, skip this if a compute
2036 * is already in flight, or if the stripe contents are in the
2037 * midst of changing due to a write
2039 if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2040 !test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2041 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2042 for (i = disks; i--; )
2043 if (__handle_issuing_new_read_requests5(
2044 sh, s, i, disks) == 0)
2047 set_bit(STRIPE_HANDLE, &sh->state);
2050 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2051 struct stripe_head_state *s, struct r6_state *r6s,
2055 for (i = disks; i--; ) {
2056 struct r5dev *dev = &sh->dev[i];
2057 if (!test_bit(R5_LOCKED, &dev->flags) &&
2058 !test_bit(R5_UPTODATE, &dev->flags) &&
2059 (dev->toread || (dev->towrite &&
2060 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2061 s->syncing || s->expanding ||
2063 (sh->dev[r6s->failed_num[0]].toread ||
2066 (sh->dev[r6s->failed_num[1]].toread ||
2068 /* we would like to get this block, possibly
2069 * by computing it, but we might not be able to
2071 if ((s->uptodate == disks - 1) &&
2072 (s->failed && (i == r6s->failed_num[0] ||
2073 i == r6s->failed_num[1]))) {
2074 pr_debug("Computing stripe %llu block %d\n",
2075 (unsigned long long)sh->sector, i);
2076 compute_block_1(sh, i, 0);
2078 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2079 /* Computing 2-failure is *very* expensive; only
2080 * do it if failed >= 2
2083 for (other = disks; other--; ) {
2086 if (!test_bit(R5_UPTODATE,
2087 &sh->dev[other].flags))
2091 pr_debug("Computing stripe %llu blocks %d,%d\n",
2092 (unsigned long long)sh->sector,
2094 compute_block_2(sh, i, other);
2096 } else if (test_bit(R5_Insync, &dev->flags)) {
2097 set_bit(R5_LOCKED, &dev->flags);
2098 set_bit(R5_Wantread, &dev->flags);
2100 pr_debug("Reading block %d (sync=%d)\n",
2105 set_bit(STRIPE_HANDLE, &sh->state);
2109 /* handle_completed_write_requests
2110 * any written block on an uptodate or failed drive can be returned.
2111 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2112 * never LOCKED, so we don't need to test 'failed' directly.
2114 static void handle_completed_write_requests(raid5_conf_t *conf,
2115 struct stripe_head *sh, int disks, struct bio **return_bi)
2120 for (i = disks; i--; )
2121 if (sh->dev[i].written) {
2123 if (!test_bit(R5_LOCKED, &dev->flags) &&
2124 test_bit(R5_UPTODATE, &dev->flags)) {
2125 /* We can return any write requests */
2126 struct bio *wbi, *wbi2;
2128 pr_debug("Return write for disc %d\n", i);
2129 spin_lock_irq(&conf->device_lock);
2131 dev->written = NULL;
2132 while (wbi && wbi->bi_sector <
2133 dev->sector + STRIPE_SECTORS) {
2134 wbi2 = r5_next_bio(wbi, dev->sector);
2135 if (--wbi->bi_phys_segments == 0) {
2136 md_write_end(conf->mddev);
2137 wbi->bi_next = *return_bi;
2142 if (dev->towrite == NULL)
2144 spin_unlock_irq(&conf->device_lock);
2146 bitmap_endwrite(conf->mddev->bitmap,
2149 !test_bit(STRIPE_DEGRADED, &sh->state),
2154 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2155 if (atomic_dec_and_test(&conf->pending_full_writes))
2156 md_wakeup_thread(conf->mddev->thread);
2159 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2160 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2162 int rmw = 0, rcw = 0, i;
2163 for (i = disks; i--; ) {
2164 /* would I have to read this buffer for read_modify_write */
2165 struct r5dev *dev = &sh->dev[i];
2166 if ((dev->towrite || i == sh->pd_idx) &&
2167 !test_bit(R5_LOCKED, &dev->flags) &&
2168 !(test_bit(R5_UPTODATE, &dev->flags) ||
2169 test_bit(R5_Wantcompute, &dev->flags))) {
2170 if (test_bit(R5_Insync, &dev->flags))
2173 rmw += 2*disks; /* cannot read it */
2175 /* Would I have to read this buffer for reconstruct_write */
2176 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2177 !test_bit(R5_LOCKED, &dev->flags) &&
2178 !(test_bit(R5_UPTODATE, &dev->flags) ||
2179 test_bit(R5_Wantcompute, &dev->flags))) {
2180 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2185 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2186 (unsigned long long)sh->sector, rmw, rcw);
2187 set_bit(STRIPE_HANDLE, &sh->state);
2188 if (rmw < rcw && rmw > 0)
2189 /* prefer read-modify-write, but need to get some data */
2190 for (i = disks; i--; ) {
2191 struct r5dev *dev = &sh->dev[i];
2192 if ((dev->towrite || i == sh->pd_idx) &&
2193 !test_bit(R5_LOCKED, &dev->flags) &&
2194 !(test_bit(R5_UPTODATE, &dev->flags) ||
2195 test_bit(R5_Wantcompute, &dev->flags)) &&
2196 test_bit(R5_Insync, &dev->flags)) {
2198 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2199 pr_debug("Read_old block "
2200 "%d for r-m-w\n", i);
2201 set_bit(R5_LOCKED, &dev->flags);
2202 set_bit(R5_Wantread, &dev->flags);
2205 set_bit(STRIPE_DELAYED, &sh->state);
2206 set_bit(STRIPE_HANDLE, &sh->state);
2210 if (rcw <= rmw && rcw > 0)
2211 /* want reconstruct write, but need to get some data */
2212 for (i = disks; i--; ) {
2213 struct r5dev *dev = &sh->dev[i];
2214 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2216 !test_bit(R5_LOCKED, &dev->flags) &&
2217 !(test_bit(R5_UPTODATE, &dev->flags) ||
2218 test_bit(R5_Wantcompute, &dev->flags)) &&
2219 test_bit(R5_Insync, &dev->flags)) {
2221 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2222 pr_debug("Read_old block "
2223 "%d for Reconstruct\n", i);
2224 set_bit(R5_LOCKED, &dev->flags);
2225 set_bit(R5_Wantread, &dev->flags);
2228 set_bit(STRIPE_DELAYED, &sh->state);
2229 set_bit(STRIPE_HANDLE, &sh->state);
2233 /* now if nothing is locked, and if we have enough data,
2234 * we can start a write request
2236 /* since handle_stripe can be called at any time we need to handle the
2237 * case where a compute block operation has been submitted and then a
2238 * subsequent call wants to start a write request. raid5_run_ops only
2239 * handles the case where compute block and postxor are requested
2240 * simultaneously. If this is not the case then new writes need to be
2241 * held off until the compute completes.
2243 if ((s->req_compute ||
2244 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2245 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2246 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2247 s->locked += handle_write_operations5(sh, rcw == 0, 0);
2250 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2251 struct stripe_head *sh, struct stripe_head_state *s,
2252 struct r6_state *r6s, int disks)
2254 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2255 int qd_idx = r6s->qd_idx;
2256 for (i = disks; i--; ) {
2257 struct r5dev *dev = &sh->dev[i];
2258 /* Would I have to read this buffer for reconstruct_write */
2259 if (!test_bit(R5_OVERWRITE, &dev->flags)
2260 && i != pd_idx && i != qd_idx
2261 && (!test_bit(R5_LOCKED, &dev->flags)
2263 !test_bit(R5_UPTODATE, &dev->flags)) {
2264 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2266 pr_debug("raid6: must_compute: "
2267 "disk %d flags=%#lx\n", i, dev->flags);
2272 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2273 (unsigned long long)sh->sector, rcw, must_compute);
2274 set_bit(STRIPE_HANDLE, &sh->state);
2277 /* want reconstruct write, but need to get some data */
2278 for (i = disks; i--; ) {
2279 struct r5dev *dev = &sh->dev[i];
2280 if (!test_bit(R5_OVERWRITE, &dev->flags)
2281 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2282 && !test_bit(R5_LOCKED, &dev->flags) &&
2283 !test_bit(R5_UPTODATE, &dev->flags) &&
2284 test_bit(R5_Insync, &dev->flags)) {
2286 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2287 pr_debug("Read_old stripe %llu "
2288 "block %d for Reconstruct\n",
2289 (unsigned long long)sh->sector, i);
2290 set_bit(R5_LOCKED, &dev->flags);
2291 set_bit(R5_Wantread, &dev->flags);
2294 pr_debug("Request delayed stripe %llu "
2295 "block %d for Reconstruct\n",
2296 (unsigned long long)sh->sector, i);
2297 set_bit(STRIPE_DELAYED, &sh->state);
2298 set_bit(STRIPE_HANDLE, &sh->state);
2302 /* now if nothing is locked, and if we have enough data, we can start a
2305 if (s->locked == 0 && rcw == 0 &&
2306 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2307 if (must_compute > 0) {
2308 /* We have failed blocks and need to compute them */
2309 switch (s->failed) {
2313 compute_block_1(sh, r6s->failed_num[0], 0);
2316 compute_block_2(sh, r6s->failed_num[0],
2317 r6s->failed_num[1]);
2319 default: /* This request should have been failed? */
2324 pr_debug("Computing parity for stripe %llu\n",
2325 (unsigned long long)sh->sector);
2326 compute_parity6(sh, RECONSTRUCT_WRITE);
2327 /* now every locked buffer is ready to be written */
2328 for (i = disks; i--; )
2329 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2330 pr_debug("Writing stripe %llu block %d\n",
2331 (unsigned long long)sh->sector, i);
2333 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2335 if (s->locked == disks)
2336 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2337 atomic_inc(&conf->pending_full_writes);
2338 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2339 set_bit(STRIPE_INSYNC, &sh->state);
2341 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2342 atomic_dec(&conf->preread_active_stripes);
2343 if (atomic_read(&conf->preread_active_stripes) <
2345 md_wakeup_thread(conf->mddev->thread);
2350 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2351 struct stripe_head_state *s, int disks)
2353 struct r5dev *dev = NULL;
2355 set_bit(STRIPE_HANDLE, &sh->state);
2357 switch (sh->check_state) {
2358 case check_state_idle:
2359 /* start a new check operation if there are no failures */
2360 if (s->failed == 0) {
2361 BUG_ON(s->uptodate != disks);
2362 sh->check_state = check_state_run;
2363 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2364 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2368 dev = &sh->dev[s->failed_num];
2370 case check_state_compute_result:
2371 sh->check_state = check_state_idle;
2373 dev = &sh->dev[sh->pd_idx];
2375 /* check that a write has not made the stripe insync */
2376 if (test_bit(STRIPE_INSYNC, &sh->state))
2379 /* either failed parity check, or recovery is happening */
2380 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2381 BUG_ON(s->uptodate != disks);
2383 set_bit(R5_LOCKED, &dev->flags);
2385 set_bit(R5_Wantwrite, &dev->flags);
2387 clear_bit(STRIPE_DEGRADED, &sh->state);
2388 set_bit(STRIPE_INSYNC, &sh->state);
2390 case check_state_run:
2391 break; /* we will be called again upon completion */
2392 case check_state_check_result:
2393 sh->check_state = check_state_idle;
2395 /* if a failure occurred during the check operation, leave
2396 * STRIPE_INSYNC not set and let the stripe be handled again
2401 /* handle a successful check operation, if parity is correct
2402 * we are done. Otherwise update the mismatch count and repair
2403 * parity if !MD_RECOVERY_CHECK
2405 if (sh->ops.zero_sum_result == 0)
2406 /* parity is correct (on disc,
2407 * not in buffer any more)
2409 set_bit(STRIPE_INSYNC, &sh->state);
2411 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2412 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2413 /* don't try to repair!! */
2414 set_bit(STRIPE_INSYNC, &sh->state);
2416 sh->check_state = check_state_compute_run;
2417 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2418 set_bit(R5_Wantcompute,
2419 &sh->dev[sh->pd_idx].flags);
2420 sh->ops.target = sh->pd_idx;
2425 case check_state_compute_run:
2428 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2429 __func__, sh->check_state,
2430 (unsigned long long) sh->sector);
2436 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2437 struct stripe_head_state *s,
2438 struct r6_state *r6s, struct page *tmp_page,
2441 int update_p = 0, update_q = 0;
2443 int pd_idx = sh->pd_idx;
2444 int qd_idx = r6s->qd_idx;
2446 set_bit(STRIPE_HANDLE, &sh->state);
2448 BUG_ON(s->failed > 2);
2449 BUG_ON(s->uptodate < disks);
2450 /* Want to check and possibly repair P and Q.
2451 * However there could be one 'failed' device, in which
2452 * case we can only check one of them, possibly using the
2453 * other to generate missing data
2456 /* If !tmp_page, we cannot do the calculations,
2457 * but as we have set STRIPE_HANDLE, we will soon be called
2458 * by stripe_handle with a tmp_page - just wait until then.
2461 if (s->failed == r6s->q_failed) {
2462 /* The only possible failed device holds 'Q', so it
2463 * makes sense to check P (If anything else were failed,
2464 * we would have used P to recreate it).
2466 compute_block_1(sh, pd_idx, 1);
2467 if (!page_is_zero(sh->dev[pd_idx].page)) {
2468 compute_block_1(sh, pd_idx, 0);
2472 if (!r6s->q_failed && s->failed < 2) {
2473 /* q is not failed, and we didn't use it to generate
2474 * anything, so it makes sense to check it
2476 memcpy(page_address(tmp_page),
2477 page_address(sh->dev[qd_idx].page),
2479 compute_parity6(sh, UPDATE_PARITY);
2480 if (memcmp(page_address(tmp_page),
2481 page_address(sh->dev[qd_idx].page),
2482 STRIPE_SIZE) != 0) {
2483 clear_bit(STRIPE_INSYNC, &sh->state);
2487 if (update_p || update_q) {
2488 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2489 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2490 /* don't try to repair!! */
2491 update_p = update_q = 0;
2494 /* now write out any block on a failed drive,
2495 * or P or Q if they need it
2498 if (s->failed == 2) {
2499 dev = &sh->dev[r6s->failed_num[1]];
2501 set_bit(R5_LOCKED, &dev->flags);
2502 set_bit(R5_Wantwrite, &dev->flags);
2504 if (s->failed >= 1) {
2505 dev = &sh->dev[r6s->failed_num[0]];
2507 set_bit(R5_LOCKED, &dev->flags);
2508 set_bit(R5_Wantwrite, &dev->flags);
2512 dev = &sh->dev[pd_idx];
2514 set_bit(R5_LOCKED, &dev->flags);
2515 set_bit(R5_Wantwrite, &dev->flags);
2518 dev = &sh->dev[qd_idx];
2520 set_bit(R5_LOCKED, &dev->flags);
2521 set_bit(R5_Wantwrite, &dev->flags);
2523 clear_bit(STRIPE_DEGRADED, &sh->state);
2525 set_bit(STRIPE_INSYNC, &sh->state);
2529 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2530 struct r6_state *r6s)
2534 /* We have read all the blocks in this stripe and now we need to
2535 * copy some of them into a target stripe for expand.
2537 struct dma_async_tx_descriptor *tx = NULL;
2538 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2539 for (i = 0; i < sh->disks; i++)
2540 if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2541 int dd_idx, pd_idx, j;
2542 struct stripe_head *sh2;
2544 sector_t bn = compute_blocknr(sh, i);
2545 sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2547 conf->max_degraded, &dd_idx,
2549 sh2 = get_active_stripe(conf, s, conf->raid_disks,
2552 /* so far only the early blocks of this stripe
2553 * have been requested. When later blocks
2554 * get requested, we will try again
2557 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2558 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2559 /* must have already done this block */
2560 release_stripe(sh2);
2564 /* place all the copies on one channel */
2565 tx = async_memcpy(sh2->dev[dd_idx].page,
2566 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2567 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2569 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2570 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2571 for (j = 0; j < conf->raid_disks; j++)
2572 if (j != sh2->pd_idx &&
2573 (!r6s || j != raid6_next_disk(sh2->pd_idx,
2575 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2577 if (j == conf->raid_disks) {
2578 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2579 set_bit(STRIPE_HANDLE, &sh2->state);
2581 release_stripe(sh2);
2584 /* done submitting copies, wait for them to complete */
2587 dma_wait_for_async_tx(tx);
2593 * handle_stripe - do things to a stripe.
2595 * We lock the stripe and then examine the state of various bits
2596 * to see what needs to be done.
2598 * return some read request which now have data
2599 * return some write requests which are safely on disc
2600 * schedule a read on some buffers
2601 * schedule a write of some buffers
2602 * return confirmation of parity correctness
2604 * buffers are taken off read_list or write_list, and bh_cache buffers
2605 * get BH_Lock set before the stripe lock is released.
2609 static void handle_stripe5(struct stripe_head *sh)
2611 raid5_conf_t *conf = sh->raid_conf;
2612 int disks = sh->disks, i;
2613 struct bio *return_bi = NULL;
2614 struct stripe_head_state s;
2616 unsigned long pending = 0;
2617 mdk_rdev_t *blocked_rdev = NULL;
2620 memset(&s, 0, sizeof(s));
2621 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2622 "ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2623 atomic_read(&sh->count), sh->pd_idx,
2624 sh->ops.pending, sh->ops.ack, sh->ops.complete);
2626 spin_lock(&sh->lock);
2627 clear_bit(STRIPE_HANDLE, &sh->state);
2628 clear_bit(STRIPE_DELAYED, &sh->state);
2630 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2631 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2632 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2633 /* Now to look around and see what can be done */
2635 /* clean-up completed biofill operations */
2636 if (test_bit(STRIPE_OP_BIOFILL, &sh->ops.complete)) {
2637 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
2638 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
2639 clear_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
2643 for (i=disks; i--; ) {
2645 struct r5dev *dev = &sh->dev[i];
2646 clear_bit(R5_Insync, &dev->flags);
2648 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2649 "written %p\n", i, dev->flags, dev->toread, dev->read,
2650 dev->towrite, dev->written);
2652 /* maybe we can request a biofill operation
2654 * new wantfill requests are only permitted while
2655 * STRIPE_OP_BIOFILL is clear
2657 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2658 !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2659 set_bit(R5_Wantfill, &dev->flags);
2661 /* now count some things */
2662 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2663 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2664 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2666 if (test_bit(R5_Wantfill, &dev->flags))
2668 else if (dev->toread)
2672 if (!test_bit(R5_OVERWRITE, &dev->flags))
2677 rdev = rcu_dereference(conf->disks[i].rdev);
2678 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2679 blocked_rdev = rdev;
2680 atomic_inc(&rdev->nr_pending);
2683 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2684 /* The ReadError flag will just be confusing now */
2685 clear_bit(R5_ReadError, &dev->flags);
2686 clear_bit(R5_ReWrite, &dev->flags);
2688 if (!rdev || !test_bit(In_sync, &rdev->flags)
2689 || test_bit(R5_ReadError, &dev->flags)) {
2693 set_bit(R5_Insync, &dev->flags);
2697 if (unlikely(blocked_rdev)) {
2698 set_bit(STRIPE_HANDLE, &sh->state);
2702 if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2705 pr_debug("locked=%d uptodate=%d to_read=%d"
2706 " to_write=%d failed=%d failed_num=%d\n",
2707 s.locked, s.uptodate, s.to_read, s.to_write,
2708 s.failed, s.failed_num);
2709 /* check if the array has lost two devices and, if so, some requests might
2712 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2713 handle_requests_to_failed_array(conf, sh, &s, disks,
2715 if (s.failed > 1 && s.syncing) {
2716 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2717 clear_bit(STRIPE_SYNCING, &sh->state);
2721 /* might be able to return some write requests if the parity block
2722 * is safe, or on a failed drive
2724 dev = &sh->dev[sh->pd_idx];
2726 ((test_bit(R5_Insync, &dev->flags) &&
2727 !test_bit(R5_LOCKED, &dev->flags) &&
2728 test_bit(R5_UPTODATE, &dev->flags)) ||
2729 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2730 handle_completed_write_requests(conf, sh, disks, &return_bi);
2732 /* Now we might consider reading some blocks, either to check/generate
2733 * parity, or to satisfy requests
2734 * or to load a block that is being partially written.
2736 if (s.to_read || s.non_overwrite ||
2737 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2738 test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2739 handle_issuing_new_read_requests5(sh, &s, disks);
2741 /* Now we check to see if any write operations have recently
2745 /* leave prexor set until postxor is done, allows us to distinguish
2746 * a rmw from a rcw during biodrain
2749 if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2750 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2753 clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2754 clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2755 clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2757 for (i = disks; i--; )
2758 clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2761 /* if only POSTXOR is set then this is an 'expand' postxor */
2762 if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2763 test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2765 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2766 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2767 clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2769 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2770 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2771 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2773 /* All the 'written' buffers and the parity block are ready to
2774 * be written back to disk
2776 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2777 for (i = disks; i--; ) {
2779 if (test_bit(R5_LOCKED, &dev->flags) &&
2780 (i == sh->pd_idx || dev->written)) {
2781 pr_debug("Writing block %d\n", i);
2782 set_bit(R5_Wantwrite, &dev->flags);
2785 if (!test_bit(R5_Insync, &dev->flags) ||
2786 (i == sh->pd_idx && s.failed == 0))
2787 set_bit(STRIPE_INSYNC, &sh->state);
2790 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2791 atomic_dec(&conf->preread_active_stripes);
2792 if (atomic_read(&conf->preread_active_stripes) <
2794 md_wakeup_thread(conf->mddev->thread);
2798 /* Now to consider new write requests and what else, if anything
2799 * should be read. We do not handle new writes when:
2800 * 1/ A 'write' operation (copy+xor) is already in flight.
2801 * 2/ A 'check' operation is in flight, as it may clobber the parity
2804 if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2806 handle_issuing_new_write_requests5(conf, sh, &s, disks);
2808 /* maybe we need to check and possibly fix the parity for this stripe
2809 * Any reads will already have been scheduled, so we just see if enough
2810 * data is available. The parity check is held off while parity
2811 * dependent operations are in flight.
2813 if (sh->check_state ||
2814 (s.syncing && s.locked == 0 &&
2815 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2816 !test_bit(STRIPE_INSYNC, &sh->state)))
2817 handle_parity_checks5(conf, sh, &s, disks);
2819 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2820 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2821 clear_bit(STRIPE_SYNCING, &sh->state);
2824 /* If the failed drive is just a ReadError, then we might need to progress
2825 * the repair/check process
2827 if (s.failed == 1 && !conf->mddev->ro &&
2828 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2829 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2830 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2832 dev = &sh->dev[s.failed_num];
2833 if (!test_bit(R5_ReWrite, &dev->flags)) {
2834 set_bit(R5_Wantwrite, &dev->flags);
2835 set_bit(R5_ReWrite, &dev->flags);
2836 set_bit(R5_LOCKED, &dev->flags);
2839 /* let's read it back */
2840 set_bit(R5_Wantread, &dev->flags);
2841 set_bit(R5_LOCKED, &dev->flags);
2846 /* Finish postxor operations initiated by the expansion
2849 if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2850 !test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2852 clear_bit(STRIPE_EXPANDING, &sh->state);
2854 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2855 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2856 clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2858 for (i = conf->raid_disks; i--; )
2859 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2860 set_bit(R5_LOCKED, &dev->flags);
2864 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2865 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2866 /* Need to write out all blocks after computing parity */
2867 sh->disks = conf->raid_disks;
2868 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2870 s.locked += handle_write_operations5(sh, 1, 1);
2871 } else if (s.expanded &&
2873 !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2874 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2875 atomic_dec(&conf->reshape_stripes);
2876 wake_up(&conf->wait_for_overlap);
2877 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2880 if (s.expanding && s.locked == 0 &&
2881 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2882 handle_stripe_expansion(conf, sh, NULL);
2885 pending = get_stripe_work(sh);
2888 spin_unlock(&sh->lock);
2890 /* wait for this device to become unblocked */
2891 if (unlikely(blocked_rdev))
2892 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2894 if (pending || s.ops_request)
2895 raid5_run_ops(sh, pending, s.ops_request);
2899 return_io(return_bi);
2902 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2904 raid6_conf_t *conf = sh->raid_conf;
2905 int disks = sh->disks;
2906 struct bio *return_bi = NULL;
2907 int i, pd_idx = sh->pd_idx;
2908 struct stripe_head_state s;
2909 struct r6_state r6s;
2910 struct r5dev *dev, *pdev, *qdev;
2911 mdk_rdev_t *blocked_rdev = NULL;
2913 r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2914 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2915 "pd_idx=%d, qd_idx=%d\n",
2916 (unsigned long long)sh->sector, sh->state,
2917 atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2918 memset(&s, 0, sizeof(s));
2920 spin_lock(&sh->lock);
2921 clear_bit(STRIPE_HANDLE, &sh->state);
2922 clear_bit(STRIPE_DELAYED, &sh->state);
2924 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2925 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2926 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2927 /* Now to look around and see what can be done */
2930 for (i=disks; i--; ) {
2933 clear_bit(R5_Insync, &dev->flags);
2935 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2936 i, dev->flags, dev->toread, dev->towrite, dev->written);
2937 /* maybe we can reply to a read */
2938 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2939 struct bio *rbi, *rbi2;
2940 pr_debug("Return read for disc %d\n", i);
2941 spin_lock_irq(&conf->device_lock);
2944 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2945 wake_up(&conf->wait_for_overlap);
2946 spin_unlock_irq(&conf->device_lock);
2947 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2948 copy_data(0, rbi, dev->page, dev->sector);
2949 rbi2 = r5_next_bio(rbi, dev->sector);
2950 spin_lock_irq(&conf->device_lock);
2951 if (--rbi->bi_phys_segments == 0) {
2952 rbi->bi_next = return_bi;
2955 spin_unlock_irq(&conf->device_lock);
2960 /* now count some things */
2961 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2962 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2969 if (!test_bit(R5_OVERWRITE, &dev->flags))
2974 rdev = rcu_dereference(conf->disks[i].rdev);
2975 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2976 blocked_rdev = rdev;
2977 atomic_inc(&rdev->nr_pending);
2980 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2981 /* The ReadError flag will just be confusing now */
2982 clear_bit(R5_ReadError, &dev->flags);
2983 clear_bit(R5_ReWrite, &dev->flags);
2985 if (!rdev || !test_bit(In_sync, &rdev->flags)
2986 || test_bit(R5_ReadError, &dev->flags)) {
2988 r6s.failed_num[s.failed] = i;
2991 set_bit(R5_Insync, &dev->flags);
2995 if (unlikely(blocked_rdev)) {
2996 set_bit(STRIPE_HANDLE, &sh->state);
2999 pr_debug("locked=%d uptodate=%d to_read=%d"
3000 " to_write=%d failed=%d failed_num=%d,%d\n",
3001 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3002 r6s.failed_num[0], r6s.failed_num[1]);
3003 /* check if the array has lost >2 devices and, if so, some requests
3004 * might need to be failed
3006 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3007 handle_requests_to_failed_array(conf, sh, &s, disks,
3009 if (s.failed > 2 && s.syncing) {
3010 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3011 clear_bit(STRIPE_SYNCING, &sh->state);
3016 * might be able to return some write requests if the parity blocks
3017 * are safe, or on a failed drive
3019 pdev = &sh->dev[pd_idx];
3020 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3021 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3022 qdev = &sh->dev[r6s.qd_idx];
3023 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3024 || (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3027 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3028 && !test_bit(R5_LOCKED, &pdev->flags)
3029 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3030 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3031 && !test_bit(R5_LOCKED, &qdev->flags)
3032 && test_bit(R5_UPTODATE, &qdev->flags)))))
3033 handle_completed_write_requests(conf, sh, disks, &return_bi);
3035 /* Now we might consider reading some blocks, either to check/generate
3036 * parity, or to satisfy requests
3037 * or to load a block that is being partially written.
3039 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3040 (s.syncing && (s.uptodate < disks)) || s.expanding)
3041 handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3043 /* now to consider writing and what else, if anything should be read */
3045 handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3047 /* maybe we need to check and possibly fix the parity for this stripe
3048 * Any reads will already have been scheduled, so we just see if enough
3051 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3052 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3054 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3055 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3056 clear_bit(STRIPE_SYNCING, &sh->state);
3059 /* If the failed drives are just a ReadError, then we might need
3060 * to progress the repair/check process
3062 if (s.failed <= 2 && !conf->mddev->ro)
3063 for (i = 0; i < s.failed; i++) {
3064 dev = &sh->dev[r6s.failed_num[i]];
3065 if (test_bit(R5_ReadError, &dev->flags)
3066 && !test_bit(R5_LOCKED, &dev->flags)
3067 && test_bit(R5_UPTODATE, &dev->flags)
3069 if (!test_bit(R5_ReWrite, &dev->flags)) {
3070 set_bit(R5_Wantwrite, &dev->flags);
3071 set_bit(R5_ReWrite, &dev->flags);
3072 set_bit(R5_LOCKED, &dev->flags);
3074 /* let's read it back */
3075 set_bit(R5_Wantread, &dev->flags);
3076 set_bit(R5_LOCKED, &dev->flags);
3081 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3082 /* Need to write out all blocks after computing P&Q */
3083 sh->disks = conf->raid_disks;
3084 sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3086 compute_parity6(sh, RECONSTRUCT_WRITE);
3087 for (i = conf->raid_disks ; i-- ; ) {
3088 set_bit(R5_LOCKED, &sh->dev[i].flags);
3090 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3092 clear_bit(STRIPE_EXPANDING, &sh->state);
3093 } else if (s.expanded) {
3094 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3095 atomic_dec(&conf->reshape_stripes);
3096 wake_up(&conf->wait_for_overlap);
3097 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3100 if (s.expanding && s.locked == 0 &&
3101 !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
3102 handle_stripe_expansion(conf, sh, &r6s);
3105 spin_unlock(&sh->lock);
3107 /* wait for this device to become unblocked */
3108 if (unlikely(blocked_rdev))
3109 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3113 return_io(return_bi);
3116 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3118 if (sh->raid_conf->level == 6)
3119 handle_stripe6(sh, tmp_page);
3126 static void raid5_activate_delayed(raid5_conf_t *conf)
3128 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3129 while (!list_empty(&conf->delayed_list)) {
3130 struct list_head *l = conf->delayed_list.next;
3131 struct stripe_head *sh;
3132 sh = list_entry(l, struct stripe_head, lru);
3134 clear_bit(STRIPE_DELAYED, &sh->state);
3135 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3136 atomic_inc(&conf->preread_active_stripes);
3137 list_add_tail(&sh->lru, &conf->hold_list);
3140 blk_plug_device(conf->mddev->queue);
3143 static void activate_bit_delay(raid5_conf_t *conf)
3145 /* device_lock is held */
3146 struct list_head head;
3147 list_add(&head, &conf->bitmap_list);
3148 list_del_init(&conf->bitmap_list);
3149 while (!list_empty(&head)) {
3150 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3151 list_del_init(&sh->lru);
3152 atomic_inc(&sh->count);
3153 __release_stripe(conf, sh);
3157 static void unplug_slaves(mddev_t *mddev)
3159 raid5_conf_t *conf = mddev_to_conf(mddev);
3163 for (i=0; i<mddev->raid_disks; i++) {
3164 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3165 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3166 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3168 atomic_inc(&rdev->nr_pending);
3171 blk_unplug(r_queue);
3173 rdev_dec_pending(rdev, mddev);
3180 static void raid5_unplug_device(struct request_queue *q)
3182 mddev_t *mddev = q->queuedata;
3183 raid5_conf_t *conf = mddev_to_conf(mddev);
3184 unsigned long flags;
3186 spin_lock_irqsave(&conf->device_lock, flags);
3188 if (blk_remove_plug(q)) {
3190 raid5_activate_delayed(conf);
3192 md_wakeup_thread(mddev->thread);
3194 spin_unlock_irqrestore(&conf->device_lock, flags);
3196 unplug_slaves(mddev);
3199 static int raid5_congested(void *data, int bits)
3201 mddev_t *mddev = data;
3202 raid5_conf_t *conf = mddev_to_conf(mddev);
3204 /* No difference between reads and writes. Just check
3205 * how busy the stripe_cache is
3207 if (conf->inactive_blocked)
3211 if (list_empty_careful(&conf->inactive_list))
3217 /* We want read requests to align with chunks where possible,
3218 * but write requests don't need to.
3220 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3222 mddev_t *mddev = q->queuedata;
3223 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3225 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3226 unsigned int bio_sectors = bio->bi_size >> 9;
3228 if (bio_data_dir(bio) == WRITE)
3229 return biovec->bv_len; /* always allow writes to be mergeable */
3231 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3232 if (max < 0) max = 0;
3233 if (max <= biovec->bv_len && bio_sectors == 0)
3234 return biovec->bv_len;
3240 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3242 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3243 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3244 unsigned int bio_sectors = bio->bi_size >> 9;
3246 return chunk_sectors >=
3247 ((sector & (chunk_sectors - 1)) + bio_sectors);
3251 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3252 * later sampled by raid5d.
3254 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3256 unsigned long flags;
3258 spin_lock_irqsave(&conf->device_lock, flags);
3260 bi->bi_next = conf->retry_read_aligned_list;
3261 conf->retry_read_aligned_list = bi;
3263 spin_unlock_irqrestore(&conf->device_lock, flags);
3264 md_wakeup_thread(conf->mddev->thread);
3268 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3272 bi = conf->retry_read_aligned;
3274 conf->retry_read_aligned = NULL;
3277 bi = conf->retry_read_aligned_list;
3279 conf->retry_read_aligned_list = bi->bi_next;
3281 bi->bi_phys_segments = 1; /* biased count of active stripes */
3282 bi->bi_hw_segments = 0; /* count of processed stripes */
3290 * The "raid5_align_endio" should check if the read succeeded and if it
3291 * did, call bio_endio on the original bio (having bio_put the new bio
3293 * If the read failed..
3295 static void raid5_align_endio(struct bio *bi, int error)
3297 struct bio* raid_bi = bi->bi_private;
3300 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3305 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3306 conf = mddev_to_conf(mddev);
3307 rdev = (void*)raid_bi->bi_next;
3308 raid_bi->bi_next = NULL;
3310 rdev_dec_pending(rdev, conf->mddev);
3312 if (!error && uptodate) {
3313 bio_endio(raid_bi, 0);
3314 if (atomic_dec_and_test(&conf->active_aligned_reads))
3315 wake_up(&conf->wait_for_stripe);
3320 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3322 add_bio_to_retry(raid_bi, conf);
3325 static int bio_fits_rdev(struct bio *bi)
3327 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3329 if ((bi->bi_size>>9) > q->max_sectors)
3331 blk_recount_segments(q, bi);
3332 if (bi->bi_phys_segments > q->max_phys_segments ||
3333 bi->bi_hw_segments > q->max_hw_segments)
3336 if (q->merge_bvec_fn)
3337 /* it's too hard to apply the merge_bvec_fn at this stage,
3346 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3348 mddev_t *mddev = q->queuedata;
3349 raid5_conf_t *conf = mddev_to_conf(mddev);
3350 const unsigned int raid_disks = conf->raid_disks;
3351 const unsigned int data_disks = raid_disks - conf->max_degraded;
3352 unsigned int dd_idx, pd_idx;
3353 struct bio* align_bi;
3356 if (!in_chunk_boundary(mddev, raid_bio)) {
3357 pr_debug("chunk_aligned_read : non aligned\n");
3361 * use bio_clone to make a copy of the bio
3363 align_bi = bio_clone(raid_bio, GFP_NOIO);
3367 * set bi_end_io to a new function, and set bi_private to the
3370 align_bi->bi_end_io = raid5_align_endio;
3371 align_bi->bi_private = raid_bio;
3375 align_bi->bi_sector = raid5_compute_sector(raid_bio->bi_sector,
3383 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3384 if (rdev && test_bit(In_sync, &rdev->flags)) {
3385 atomic_inc(&rdev->nr_pending);
3387 raid_bio->bi_next = (void*)rdev;
3388 align_bi->bi_bdev = rdev->bdev;
3389 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3390 align_bi->bi_sector += rdev->data_offset;
3392 if (!bio_fits_rdev(align_bi)) {
3393 /* too big in some way */
3395 rdev_dec_pending(rdev, mddev);
3399 spin_lock_irq(&conf->device_lock);
3400 wait_event_lock_irq(conf->wait_for_stripe,
3402 conf->device_lock, /* nothing */);
3403 atomic_inc(&conf->active_aligned_reads);
3404 spin_unlock_irq(&conf->device_lock);
3406 generic_make_request(align_bi);
3415 /* __get_priority_stripe - get the next stripe to process
3417 * Full stripe writes are allowed to pass preread active stripes up until
3418 * the bypass_threshold is exceeded. In general the bypass_count
3419 * increments when the handle_list is handled before the hold_list; however, it
3420 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3421 * stripe with in flight i/o. The bypass_count will be reset when the
3422 * head of the hold_list has changed, i.e. the head was promoted to the
3425 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3427 struct stripe_head *sh;
3429 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3431 list_empty(&conf->handle_list) ? "empty" : "busy",
3432 list_empty(&conf->hold_list) ? "empty" : "busy",
3433 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3435 if (!list_empty(&conf->handle_list)) {
3436 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3438 if (list_empty(&conf->hold_list))
3439 conf->bypass_count = 0;
3440 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3441 if (conf->hold_list.next == conf->last_hold)
3442 conf->bypass_count++;
3444 conf->last_hold = conf->hold_list.next;
3445 conf->bypass_count -= conf->bypass_threshold;
3446 if (conf->bypass_count < 0)
3447 conf->bypass_count = 0;
3450 } else if (!list_empty(&conf->hold_list) &&
3451 ((conf->bypass_threshold &&
3452 conf->bypass_count > conf->bypass_threshold) ||
3453 atomic_read(&conf->pending_full_writes) == 0)) {
3454 sh = list_entry(conf->hold_list.next,
3456 conf->bypass_count -= conf->bypass_threshold;
3457 if (conf->bypass_count < 0)
3458 conf->bypass_count = 0;
3462 list_del_init(&sh->lru);
3463 atomic_inc(&sh->count);
3464 BUG_ON(atomic_read(&sh->count) != 1);
3468 static int make_request(struct request_queue *q, struct bio * bi)
3470 mddev_t *mddev = q->queuedata;
3471 raid5_conf_t *conf = mddev_to_conf(mddev);
3472 unsigned int dd_idx, pd_idx;
3473 sector_t new_sector;
3474 sector_t logical_sector, last_sector;
3475 struct stripe_head *sh;
3476 const int rw = bio_data_dir(bi);
3479 if (unlikely(bio_barrier(bi))) {
3480 bio_endio(bi, -EOPNOTSUPP);
3484 md_write_start(mddev, bi);
3486 disk_stat_inc(mddev->gendisk, ios[rw]);
3487 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3490 mddev->reshape_position == MaxSector &&
3491 chunk_aligned_read(q,bi))
3494 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3495 last_sector = bi->bi_sector + (bi->bi_size>>9);
3497 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3499 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3501 int disks, data_disks;
3504 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3505 if (likely(conf->expand_progress == MaxSector))
3506 disks = conf->raid_disks;
3508 /* spinlock is needed as expand_progress may be
3509 * 64bit on a 32bit platform, and so it might be
3510 * possible to see a half-updated value
3511 * Ofcourse expand_progress could change after
3512 * the lock is dropped, so once we get a reference
3513 * to the stripe that we think it is, we will have
3516 spin_lock_irq(&conf->device_lock);
3517 disks = conf->raid_disks;
3518 if (logical_sector >= conf->expand_progress)
3519 disks = conf->previous_raid_disks;
3521 if (logical_sector >= conf->expand_lo) {
3522 spin_unlock_irq(&conf->device_lock);
3527 spin_unlock_irq(&conf->device_lock);
3529 data_disks = disks - conf->max_degraded;
3531 new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3532 &dd_idx, &pd_idx, conf);
3533 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3534 (unsigned long long)new_sector,
3535 (unsigned long long)logical_sector);
3537 sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3539 if (unlikely(conf->expand_progress != MaxSector)) {
3540 /* expansion might have moved on while waiting for a
3541 * stripe, so we must do the range check again.
3542 * Expansion could still move past after this
3543 * test, but as we are holding a reference to
3544 * 'sh', we know that if that happens,
3545 * STRIPE_EXPANDING will get set and the expansion
3546 * won't proceed until we finish with the stripe.
3549 spin_lock_irq(&conf->device_lock);
3550 if (logical_sector < conf->expand_progress &&
3551 disks == conf->previous_raid_disks)
3552 /* mismatch, need to try again */
3554 spin_unlock_irq(&conf->device_lock);
3560 /* FIXME what if we get a false positive because these
3561 * are being updated.
3563 if (logical_sector >= mddev->suspend_lo &&
3564 logical_sector < mddev->suspend_hi) {
3570 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3571 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3572 /* Stripe is busy expanding or
3573 * add failed due to overlap. Flush everything
3576 raid5_unplug_device(mddev->queue);
3581 finish_wait(&conf->wait_for_overlap, &w);
3582 set_bit(STRIPE_HANDLE, &sh->state);
3583 clear_bit(STRIPE_DELAYED, &sh->state);
3586 /* cannot get stripe for read-ahead, just give-up */
3587 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3588 finish_wait(&conf->wait_for_overlap, &w);
3593 spin_lock_irq(&conf->device_lock);
3594 remaining = --bi->bi_phys_segments;
3595 spin_unlock_irq(&conf->device_lock);
3596 if (remaining == 0) {
3599 md_write_end(mddev);
3606 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3608 /* reshaping is quite different to recovery/resync so it is
3609 * handled quite separately ... here.
3611 * On each call to sync_request, we gather one chunk worth of
3612 * destination stripes and flag them as expanding.
3613 * Then we find all the source stripes and request reads.
3614 * As the reads complete, handle_stripe will copy the data
3615 * into the destination stripe and release that stripe.
3617 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3618 struct stripe_head *sh;
3620 sector_t first_sector, last_sector;
3621 int raid_disks = conf->previous_raid_disks;
3622 int data_disks = raid_disks - conf->max_degraded;
3623 int new_data_disks = conf->raid_disks - conf->max_degraded;
3626 sector_t writepos, safepos, gap;
3628 if (sector_nr == 0 &&
3629 conf->expand_progress != 0) {
3630 /* restarting in the middle, skip the initial sectors */
3631 sector_nr = conf->expand_progress;
3632 sector_div(sector_nr, new_data_disks);
3637 /* we update the metadata when there is more than 3Meg
3638 * in the block range (that is rather arbitrary, should
3639 * probably be time based) or when the data about to be
3640 * copied would over-write the source of the data at
3641 * the front of the range.
3642 * i.e. one new_stripe forward from expand_progress new_maps
3643 * to after where expand_lo old_maps to
3645 writepos = conf->expand_progress +
3646 conf->chunk_size/512*(new_data_disks);
3647 sector_div(writepos, new_data_disks);
3648 safepos = conf->expand_lo;
3649 sector_div(safepos, data_disks);
3650 gap = conf->expand_progress - conf->expand_lo;
3652 if (writepos >= safepos ||
3653 gap > (new_data_disks)*3000*2 /*3Meg*/) {
3654 /* Cannot proceed until we've updated the superblock... */
3655 wait_event(conf->wait_for_overlap,
3656 atomic_read(&conf->reshape_stripes)==0);
3657 mddev->reshape_position = conf->expand_progress;
3658 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3659 md_wakeup_thread(mddev->thread);
3660 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3661 kthread_should_stop());
3662 spin_lock_irq(&conf->device_lock);
3663 conf->expand_lo = mddev->reshape_position;
3664 spin_unlock_irq(&conf->device_lock);
3665 wake_up(&conf->wait_for_overlap);
3668 for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3671 pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3672 sh = get_active_stripe(conf, sector_nr+i,
3673 conf->raid_disks, pd_idx, 0);
3674 set_bit(STRIPE_EXPANDING, &sh->state);
3675 atomic_inc(&conf->reshape_stripes);
3676 /* If any of this stripe is beyond the end of the old
3677 * array, then we need to zero those blocks
3679 for (j=sh->disks; j--;) {
3681 if (j == sh->pd_idx)
3683 if (conf->level == 6 &&
3684 j == raid6_next_disk(sh->pd_idx, sh->disks))
3686 s = compute_blocknr(sh, j);
3687 if (s < (mddev->array_size<<1)) {
3691 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3692 set_bit(R5_Expanded, &sh->dev[j].flags);
3693 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3696 set_bit(STRIPE_EXPAND_READY, &sh->state);
3697 set_bit(STRIPE_HANDLE, &sh->state);
3701 spin_lock_irq(&conf->device_lock);
3702 conf->expand_progress = (sector_nr + i) * new_data_disks;
3703 spin_unlock_irq(&conf->device_lock);
3704 /* Ok, those stripe are ready. We can start scheduling
3705 * reads on the source stripes.
3706 * The source stripes are determined by mapping the first and last
3707 * block on the destination stripes.
3710 raid5_compute_sector(sector_nr*(new_data_disks),
3711 raid_disks, data_disks,
3712 &dd_idx, &pd_idx, conf);
3714 raid5_compute_sector((sector_nr+conf->chunk_size/512)
3715 *(new_data_disks) -1,
3716 raid_disks, data_disks,
3717 &dd_idx, &pd_idx, conf);
3718 if (last_sector >= (mddev->size<<1))
3719 last_sector = (mddev->size<<1)-1;
3720 while (first_sector <= last_sector) {
3721 pd_idx = stripe_to_pdidx(first_sector, conf,
3722 conf->previous_raid_disks);
3723 sh = get_active_stripe(conf, first_sector,
3724 conf->previous_raid_disks, pd_idx, 0);
3725 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3726 set_bit(STRIPE_HANDLE, &sh->state);
3728 first_sector += STRIPE_SECTORS;
3730 /* If this takes us to the resync_max point where we have to pause,
3731 * then we need to write out the superblock.
3733 sector_nr += conf->chunk_size>>9;
3734 if (sector_nr >= mddev->resync_max) {
3735 /* Cannot proceed until we've updated the superblock... */
3736 wait_event(conf->wait_for_overlap,
3737 atomic_read(&conf->reshape_stripes) == 0);
3738 mddev->reshape_position = conf->expand_progress;
3739 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3740 md_wakeup_thread(mddev->thread);
3741 wait_event(mddev->sb_wait,
3742 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3743 || kthread_should_stop());
3744 spin_lock_irq(&conf->device_lock);
3745 conf->expand_lo = mddev->reshape_position;
3746 spin_unlock_irq(&conf->device_lock);
3747 wake_up(&conf->wait_for_overlap);
3749 return conf->chunk_size>>9;
3752 /* FIXME go_faster isn't used */
3753 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3755 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3756 struct stripe_head *sh;
3758 int raid_disks = conf->raid_disks;
3759 sector_t max_sector = mddev->size << 1;
3761 int still_degraded = 0;
3764 if (sector_nr >= max_sector) {
3765 /* just being told to finish up .. nothing much to do */
3766 unplug_slaves(mddev);
3767 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3772 if (mddev->curr_resync < max_sector) /* aborted */
3773 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3775 else /* completed sync */
3777 bitmap_close_sync(mddev->bitmap);
3782 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3783 return reshape_request(mddev, sector_nr, skipped);
3785 /* No need to check resync_max as we never do more than one
3786 * stripe, and as resync_max will always be on a chunk boundary,
3787 * if the check in md_do_sync didn't fire, there is no chance
3788 * of overstepping resync_max here
3791 /* if there is too many failed drives and we are trying
3792 * to resync, then assert that we are finished, because there is
3793 * nothing we can do.
3795 if (mddev->degraded >= conf->max_degraded &&
3796 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3797 sector_t rv = (mddev->size << 1) - sector_nr;
3801 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3802 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3803 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3804 /* we can skip this block, and probably more */
3805 sync_blocks /= STRIPE_SECTORS;
3807 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3811 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3813 pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3814 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3816 sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3817 /* make sure we don't swamp the stripe cache if someone else
3818 * is trying to get access
3820 schedule_timeout_uninterruptible(1);
3822 /* Need to check if array will still be degraded after recovery/resync
3823 * We don't need to check the 'failed' flag as when that gets set,
3826 for (i=0; i<mddev->raid_disks; i++)
3827 if (conf->disks[i].rdev == NULL)
3830 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3832 spin_lock(&sh->lock);
3833 set_bit(STRIPE_SYNCING, &sh->state);
3834 clear_bit(STRIPE_INSYNC, &sh->state);
3835 spin_unlock(&sh->lock);
3837 handle_stripe(sh, NULL);
3840 return STRIPE_SECTORS;
3843 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3845 /* We may not be able to submit a whole bio at once as there
3846 * may not be enough stripe_heads available.
3847 * We cannot pre-allocate enough stripe_heads as we may need
3848 * more than exist in the cache (if we allow ever large chunks).
3849 * So we do one stripe head at a time and record in
3850 * ->bi_hw_segments how many have been done.
3852 * We *know* that this entire raid_bio is in one chunk, so
3853 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3855 struct stripe_head *sh;
3857 sector_t sector, logical_sector, last_sector;
3862 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3863 sector = raid5_compute_sector( logical_sector,
3865 conf->raid_disks - conf->max_degraded,
3869 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3871 for (; logical_sector < last_sector;
3872 logical_sector += STRIPE_SECTORS,
3873 sector += STRIPE_SECTORS,
3876 if (scnt < raid_bio->bi_hw_segments)
3877 /* already done this stripe */
3880 sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3883 /* failed to get a stripe - must wait */
3884 raid_bio->bi_hw_segments = scnt;
3885 conf->retry_read_aligned = raid_bio;
3889 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3890 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3892 raid_bio->bi_hw_segments = scnt;
3893 conf->retry_read_aligned = raid_bio;
3897 handle_stripe(sh, NULL);
3901 spin_lock_irq(&conf->device_lock);
3902 remaining = --raid_bio->bi_phys_segments;
3903 spin_unlock_irq(&conf->device_lock);
3905 bio_endio(raid_bio, 0);
3906 if (atomic_dec_and_test(&conf->active_aligned_reads))
3907 wake_up(&conf->wait_for_stripe);
3914 * This is our raid5 kernel thread.
3916 * We scan the hash table for stripes which can be handled now.
3917 * During the scan, completed stripes are saved for us by the interrupt
3918 * handler, so that they will not have to wait for our next wakeup.
3920 static void raid5d(mddev_t *mddev)
3922 struct stripe_head *sh;
3923 raid5_conf_t *conf = mddev_to_conf(mddev);
3926 pr_debug("+++ raid5d active\n");
3928 md_check_recovery(mddev);
3931 spin_lock_irq(&conf->device_lock);
3935 if (conf->seq_flush != conf->seq_write) {
3936 int seq = conf->seq_flush;
3937 spin_unlock_irq(&conf->device_lock);
3938 bitmap_unplug(mddev->bitmap);
3939 spin_lock_irq(&conf->device_lock);
3940 conf->seq_write = seq;
3941 activate_bit_delay(conf);
3944 while ((bio = remove_bio_from_retry(conf))) {
3946 spin_unlock_irq(&conf->device_lock);
3947 ok = retry_aligned_read(conf, bio);
3948 spin_lock_irq(&conf->device_lock);
3954 sh = __get_priority_stripe(conf);
3957 async_tx_issue_pending_all();
3960 spin_unlock_irq(&conf->device_lock);
3963 handle_stripe(sh, conf->spare_page);
3966 spin_lock_irq(&conf->device_lock);
3968 pr_debug("%d stripes handled\n", handled);
3970 spin_unlock_irq(&conf->device_lock);
3972 unplug_slaves(mddev);
3974 pr_debug("--- raid5d inactive\n");
3978 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
3980 raid5_conf_t *conf = mddev_to_conf(mddev);
3982 return sprintf(page, "%d\n", conf->max_nr_stripes);
3988 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
3990 raid5_conf_t *conf = mddev_to_conf(mddev);
3992 if (len >= PAGE_SIZE)
3997 if (strict_strtoul(page, 10, &new))
3999 if (new <= 16 || new > 32768)
4001 while (new < conf->max_nr_stripes) {
4002 if (drop_one_stripe(conf))
4003 conf->max_nr_stripes--;
4007 md_allow_write(mddev);
4008 while (new > conf->max_nr_stripes) {
4009 if (grow_one_stripe(conf))
4010 conf->max_nr_stripes++;
4016 static struct md_sysfs_entry
4017 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4018 raid5_show_stripe_cache_size,
4019 raid5_store_stripe_cache_size);
4022 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4024 raid5_conf_t *conf = mddev_to_conf(mddev);
4026 return sprintf(page, "%d\n", conf->bypass_threshold);
4032 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4034 raid5_conf_t *conf = mddev_to_conf(mddev);
4036 if (len >= PAGE_SIZE)
4041 if (strict_strtoul(page, 10, &new))
4043 if (new > conf->max_nr_stripes)
4045 conf->bypass_threshold = new;
4049 static struct md_sysfs_entry
4050 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4052 raid5_show_preread_threshold,
4053 raid5_store_preread_threshold);
4056 stripe_cache_active_show(mddev_t *mddev, char *page)
4058 raid5_conf_t *conf = mddev_to_conf(mddev);
4060 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4065 static struct md_sysfs_entry
4066 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4068 static struct attribute *raid5_attrs[] = {
4069 &raid5_stripecache_size.attr,
4070 &raid5_stripecache_active.attr,
4071 &raid5_preread_bypass_threshold.attr,
4074 static struct attribute_group raid5_attrs_group = {
4076 .attrs = raid5_attrs,
4079 static int run(mddev_t *mddev)
4082 int raid_disk, memory;
4084 struct disk_info *disk;
4085 struct list_head *tmp;
4086 int working_disks = 0;
4088 if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4089 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4090 mdname(mddev), mddev->level);
4094 if (mddev->reshape_position != MaxSector) {
4095 /* Check that we can continue the reshape.
4096 * Currently only disks can change, it must
4097 * increase, and we must be past the point where
4098 * a stripe over-writes itself
4100 sector_t here_new, here_old;
4102 int max_degraded = (mddev->level == 5 ? 1 : 2);
4104 if (mddev->new_level != mddev->level ||
4105 mddev->new_layout != mddev->layout ||
4106 mddev->new_chunk != mddev->chunk_size) {
4107 printk(KERN_ERR "raid5: %s: unsupported reshape "
4108 "required - aborting.\n",
4112 if (mddev->delta_disks <= 0) {
4113 printk(KERN_ERR "raid5: %s: unsupported reshape "
4114 "(reduce disks) required - aborting.\n",
4118 old_disks = mddev->raid_disks - mddev->delta_disks;
4119 /* reshape_position must be on a new-stripe boundary, and one
4120 * further up in new geometry must map after here in old
4123 here_new = mddev->reshape_position;
4124 if (sector_div(here_new, (mddev->chunk_size>>9)*
4125 (mddev->raid_disks - max_degraded))) {
4126 printk(KERN_ERR "raid5: reshape_position not "
4127 "on a stripe boundary\n");
4130 /* here_new is the stripe we will write to */
4131 here_old = mddev->reshape_position;
4132 sector_div(here_old, (mddev->chunk_size>>9)*
4133 (old_disks-max_degraded));
4134 /* here_old is the first stripe that we might need to read
4136 if (here_new >= here_old) {
4137 /* Reading from the same stripe as writing to - bad */
4138 printk(KERN_ERR "raid5: reshape_position too early for "
4139 "auto-recovery - aborting.\n");
4142 printk(KERN_INFO "raid5: reshape will continue\n");
4143 /* OK, we should be able to continue; */
4147 mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4148 if ((conf = mddev->private) == NULL)
4150 if (mddev->reshape_position == MaxSector) {
4151 conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4153 conf->raid_disks = mddev->raid_disks;
4154 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4157 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4162 conf->mddev = mddev;
4164 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4167 if (mddev->level == 6) {
4168 conf->spare_page = alloc_page(GFP_KERNEL);
4169 if (!conf->spare_page)
4172 spin_lock_init(&conf->device_lock);
4173 mddev->queue->queue_lock = &conf->device_lock;
4174 init_waitqueue_head(&conf->wait_for_stripe);
4175 init_waitqueue_head(&conf->wait_for_overlap);
4176 INIT_LIST_HEAD(&conf->handle_list);
4177 INIT_LIST_HEAD(&conf->hold_list);
4178 INIT_LIST_HEAD(&conf->delayed_list);
4179 INIT_LIST_HEAD(&conf->bitmap_list);
4180 INIT_LIST_HEAD(&conf->inactive_list);
4181 atomic_set(&conf->active_stripes, 0);
4182 atomic_set(&conf->preread_active_stripes, 0);
4183 atomic_set(&conf->active_aligned_reads, 0);
4184 conf->bypass_threshold = BYPASS_THRESHOLD;
4186 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4188 rdev_for_each(rdev, tmp, mddev) {
4189 raid_disk = rdev->raid_disk;
4190 if (raid_disk >= conf->raid_disks
4193 disk = conf->disks + raid_disk;
4197 if (test_bit(In_sync, &rdev->flags)) {
4198 char b[BDEVNAME_SIZE];
4199 printk(KERN_INFO "raid5: device %s operational as raid"
4200 " disk %d\n", bdevname(rdev->bdev,b),
4204 /* Cannot rely on bitmap to complete recovery */
4209 * 0 for a fully functional array, 1 or 2 for a degraded array.
4211 mddev->degraded = conf->raid_disks - working_disks;
4212 conf->mddev = mddev;
4213 conf->chunk_size = mddev->chunk_size;
4214 conf->level = mddev->level;
4215 if (conf->level == 6)
4216 conf->max_degraded = 2;
4218 conf->max_degraded = 1;
4219 conf->algorithm = mddev->layout;
4220 conf->max_nr_stripes = NR_STRIPES;
4221 conf->expand_progress = mddev->reshape_position;
4223 /* device size must be a multiple of chunk size */
4224 mddev->size &= ~(mddev->chunk_size/1024 -1);
4225 mddev->resync_max_sectors = mddev->size << 1;
4227 if (conf->level == 6 && conf->raid_disks < 4) {
4228 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4229 mdname(mddev), conf->raid_disks);
4232 if (!conf->chunk_size || conf->chunk_size % 4) {
4233 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4234 conf->chunk_size, mdname(mddev));
4237 if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4239 "raid5: unsupported parity algorithm %d for %s\n",
4240 conf->algorithm, mdname(mddev));
4243 if (mddev->degraded > conf->max_degraded) {
4244 printk(KERN_ERR "raid5: not enough operational devices for %s"
4245 " (%d/%d failed)\n",
4246 mdname(mddev), mddev->degraded, conf->raid_disks);
4250 if (mddev->degraded > 0 &&
4251 mddev->recovery_cp != MaxSector) {
4252 if (mddev->ok_start_degraded)
4254 "raid5: starting dirty degraded array: %s"
4255 "- data corruption possible.\n",
4259 "raid5: cannot start dirty degraded array for %s\n",
4266 mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4267 if (!mddev->thread) {
4269 "raid5: couldn't allocate thread for %s\n",
4274 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4275 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4276 if (grow_stripes(conf, conf->max_nr_stripes)) {
4278 "raid5: couldn't allocate %dkB for buffers\n", memory);
4279 shrink_stripes(conf);
4280 md_unregister_thread(mddev->thread);
4283 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4284 memory, mdname(mddev));
4286 if (mddev->degraded == 0)
4287 printk("raid5: raid level %d set %s active with %d out of %d"
4288 " devices, algorithm %d\n", conf->level, mdname(mddev),
4289 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4292 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4293 " out of %d devices, algorithm %d\n", conf->level,
4294 mdname(mddev), mddev->raid_disks - mddev->degraded,
4295 mddev->raid_disks, conf->algorithm);
4297 print_raid5_conf(conf);
4299 if (conf->expand_progress != MaxSector) {
4300 printk("...ok start reshape thread\n");
4301 conf->expand_lo = conf->expand_progress;
4302 atomic_set(&conf->reshape_stripes, 0);
4303 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4304 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4305 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4306 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4307 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4311 /* read-ahead size must cover two whole stripes, which is
4312 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4315 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4316 int stripe = data_disks *
4317 (mddev->chunk_size / PAGE_SIZE);
4318 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4319 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4322 /* Ok, everything is just fine now */
4323 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4325 "raid5: failed to create sysfs attributes for %s\n",
4328 mddev->queue->unplug_fn = raid5_unplug_device;
4329 mddev->queue->backing_dev_info.congested_data = mddev;
4330 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4332 mddev->array_size = mddev->size * (conf->previous_raid_disks -
4333 conf->max_degraded);
4335 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4340 print_raid5_conf(conf);
4341 safe_put_page(conf->spare_page);
4343 kfree(conf->stripe_hashtbl);
4346 mddev->private = NULL;
4347 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4353 static int stop(mddev_t *mddev)
4355 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4357 md_unregister_thread(mddev->thread);
4358 mddev->thread = NULL;
4359 shrink_stripes(conf);
4360 kfree(conf->stripe_hashtbl);
4361 mddev->queue->backing_dev_info.congested_fn = NULL;
4362 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4363 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4366 mddev->private = NULL;
4371 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4375 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4376 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4377 seq_printf(seq, "sh %llu, count %d.\n",
4378 (unsigned long long)sh->sector, atomic_read(&sh->count));
4379 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4380 for (i = 0; i < sh->disks; i++) {
4381 seq_printf(seq, "(cache%d: %p %ld) ",
4382 i, sh->dev[i].page, sh->dev[i].flags);
4384 seq_printf(seq, "\n");
4387 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4389 struct stripe_head *sh;
4390 struct hlist_node *hn;
4393 spin_lock_irq(&conf->device_lock);
4394 for (i = 0; i < NR_HASH; i++) {
4395 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4396 if (sh->raid_conf != conf)
4401 spin_unlock_irq(&conf->device_lock);
4405 static void status (struct seq_file *seq, mddev_t *mddev)
4407 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4410 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4411 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4412 for (i = 0; i < conf->raid_disks; i++)
4413 seq_printf (seq, "%s",
4414 conf->disks[i].rdev &&
4415 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4416 seq_printf (seq, "]");
4418 seq_printf (seq, "\n");
4419 printall(seq, conf);
4423 static void print_raid5_conf (raid5_conf_t *conf)
4426 struct disk_info *tmp;
4428 printk("RAID5 conf printout:\n");
4430 printk("(conf==NULL)\n");
4433 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4434 conf->raid_disks - conf->mddev->degraded);
4436 for (i = 0; i < conf->raid_disks; i++) {
4437 char b[BDEVNAME_SIZE];
4438 tmp = conf->disks + i;
4440 printk(" disk %d, o:%d, dev:%s\n",
4441 i, !test_bit(Faulty, &tmp->rdev->flags),
4442 bdevname(tmp->rdev->bdev,b));
4446 static int raid5_spare_active(mddev_t *mddev)
4449 raid5_conf_t *conf = mddev->private;
4450 struct disk_info *tmp;
4452 for (i = 0; i < conf->raid_disks; i++) {
4453 tmp = conf->disks + i;
4455 && !test_bit(Faulty, &tmp->rdev->flags)
4456 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4457 unsigned long flags;
4458 spin_lock_irqsave(&conf->device_lock, flags);
4460 spin_unlock_irqrestore(&conf->device_lock, flags);
4463 print_raid5_conf(conf);
4467 static int raid5_remove_disk(mddev_t *mddev, int number)
4469 raid5_conf_t *conf = mddev->private;
4472 struct disk_info *p = conf->disks + number;
4474 print_raid5_conf(conf);
4477 if (test_bit(In_sync, &rdev->flags) ||
4478 atomic_read(&rdev->nr_pending)) {
4482 /* Only remove non-faulty devices if recovery
4485 if (!test_bit(Faulty, &rdev->flags) &&
4486 mddev->degraded <= conf->max_degraded) {
4492 if (atomic_read(&rdev->nr_pending)) {
4493 /* lost the race, try later */
4500 print_raid5_conf(conf);
4504 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4506 raid5_conf_t *conf = mddev->private;
4509 struct disk_info *p;
4511 int last = conf->raid_disks - 1;
4513 if (mddev->degraded > conf->max_degraded)
4514 /* no point adding a device */
4517 if (rdev->raid_disk >= 0)
4518 first = last = rdev->raid_disk;
4521 * find the disk ... but prefer rdev->saved_raid_disk
4524 if (rdev->saved_raid_disk >= 0 &&
4525 rdev->saved_raid_disk >= first &&
4526 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4527 disk = rdev->saved_raid_disk;
4530 for ( ; disk <= last ; disk++)
4531 if ((p=conf->disks + disk)->rdev == NULL) {
4532 clear_bit(In_sync, &rdev->flags);
4533 rdev->raid_disk = disk;
4535 if (rdev->saved_raid_disk != disk)
4537 rcu_assign_pointer(p->rdev, rdev);
4540 print_raid5_conf(conf);
4544 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4546 /* no resync is happening, and there is enough space
4547 * on all devices, so we can resize.
4548 * We need to make sure resync covers any new space.
4549 * If the array is shrinking we should possibly wait until
4550 * any io in the removed space completes, but it hardly seems
4553 raid5_conf_t *conf = mddev_to_conf(mddev);
4555 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4556 mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4557 set_capacity(mddev->gendisk, mddev->array_size << 1);
4559 if (sectors/2 > mddev->size && mddev->recovery_cp == MaxSector) {
4560 mddev->recovery_cp = mddev->size << 1;
4561 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4563 mddev->size = sectors /2;
4564 mddev->resync_max_sectors = sectors;
4568 #ifdef CONFIG_MD_RAID5_RESHAPE
4569 static int raid5_check_reshape(mddev_t *mddev)
4571 raid5_conf_t *conf = mddev_to_conf(mddev);
4574 if (mddev->delta_disks < 0 ||
4575 mddev->new_level != mddev->level)
4576 return -EINVAL; /* Cannot shrink array or change level yet */
4577 if (mddev->delta_disks == 0)
4578 return 0; /* nothing to do */
4580 /* Can only proceed if there are plenty of stripe_heads.
4581 * We need a minimum of one full stripe,, and for sensible progress
4582 * it is best to have about 4 times that.
4583 * If we require 4 times, then the default 256 4K stripe_heads will
4584 * allow for chunk sizes up to 256K, which is probably OK.
4585 * If the chunk size is greater, user-space should request more
4586 * stripe_heads first.
4588 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4589 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4590 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4591 (mddev->chunk_size / STRIPE_SIZE)*4);
4595 err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4599 if (mddev->degraded > conf->max_degraded)
4601 /* looks like we might be able to manage this */
4605 static int raid5_start_reshape(mddev_t *mddev)
4607 raid5_conf_t *conf = mddev_to_conf(mddev);
4609 struct list_head *rtmp;
4611 int added_devices = 0;
4612 unsigned long flags;
4614 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4617 rdev_for_each(rdev, rtmp, mddev)
4618 if (rdev->raid_disk < 0 &&
4619 !test_bit(Faulty, &rdev->flags))
4622 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4623 /* Not enough devices even to make a degraded array
4628 atomic_set(&conf->reshape_stripes, 0);
4629 spin_lock_irq(&conf->device_lock);
4630 conf->previous_raid_disks = conf->raid_disks;
4631 conf->raid_disks += mddev->delta_disks;
4632 conf->expand_progress = 0;
4633 conf->expand_lo = 0;
4634 spin_unlock_irq(&conf->device_lock);
4636 /* Add some new drives, as many as will fit.
4637 * We know there are enough to make the newly sized array work.
4639 rdev_for_each(rdev, rtmp, mddev)
4640 if (rdev->raid_disk < 0 &&
4641 !test_bit(Faulty, &rdev->flags)) {
4642 if (raid5_add_disk(mddev, rdev) == 0) {
4644 set_bit(In_sync, &rdev->flags);
4646 rdev->recovery_offset = 0;
4647 sprintf(nm, "rd%d", rdev->raid_disk);
4648 if (sysfs_create_link(&mddev->kobj,
4651 "raid5: failed to create "
4652 " link %s for %s\n",
4658 spin_lock_irqsave(&conf->device_lock, flags);
4659 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4660 spin_unlock_irqrestore(&conf->device_lock, flags);
4661 mddev->raid_disks = conf->raid_disks;
4662 mddev->reshape_position = 0;
4663 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4665 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4666 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4667 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4668 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4669 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4671 if (!mddev->sync_thread) {
4672 mddev->recovery = 0;
4673 spin_lock_irq(&conf->device_lock);
4674 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4675 conf->expand_progress = MaxSector;
4676 spin_unlock_irq(&conf->device_lock);
4679 md_wakeup_thread(mddev->sync_thread);
4680 md_new_event(mddev);
4685 static void end_reshape(raid5_conf_t *conf)
4687 struct block_device *bdev;
4689 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4690 conf->mddev->array_size = conf->mddev->size *
4691 (conf->raid_disks - conf->max_degraded);
4692 set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4693 conf->mddev->changed = 1;
4695 bdev = bdget_disk(conf->mddev->gendisk, 0);
4697 mutex_lock(&bdev->bd_inode->i_mutex);
4698 i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4699 mutex_unlock(&bdev->bd_inode->i_mutex);
4702 spin_lock_irq(&conf->device_lock);
4703 conf->expand_progress = MaxSector;
4704 spin_unlock_irq(&conf->device_lock);
4705 conf->mddev->reshape_position = MaxSector;
4707 /* read-ahead size must cover two whole stripes, which is
4708 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4711 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4712 int stripe = data_disks *
4713 (conf->mddev->chunk_size / PAGE_SIZE);
4714 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4715 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4720 static void raid5_quiesce(mddev_t *mddev, int state)
4722 raid5_conf_t *conf = mddev_to_conf(mddev);
4725 case 2: /* resume for a suspend */
4726 wake_up(&conf->wait_for_overlap);
4729 case 1: /* stop all writes */
4730 spin_lock_irq(&conf->device_lock);
4732 wait_event_lock_irq(conf->wait_for_stripe,
4733 atomic_read(&conf->active_stripes) == 0 &&
4734 atomic_read(&conf->active_aligned_reads) == 0,
4735 conf->device_lock, /* nothing */);
4736 spin_unlock_irq(&conf->device_lock);
4739 case 0: /* re-enable writes */
4740 spin_lock_irq(&conf->device_lock);
4742 wake_up(&conf->wait_for_stripe);
4743 wake_up(&conf->wait_for_overlap);
4744 spin_unlock_irq(&conf->device_lock);
4749 static struct mdk_personality raid6_personality =
4753 .owner = THIS_MODULE,
4754 .make_request = make_request,
4758 .error_handler = error,
4759 .hot_add_disk = raid5_add_disk,
4760 .hot_remove_disk= raid5_remove_disk,
4761 .spare_active = raid5_spare_active,
4762 .sync_request = sync_request,
4763 .resize = raid5_resize,
4764 #ifdef CONFIG_MD_RAID5_RESHAPE
4765 .check_reshape = raid5_check_reshape,
4766 .start_reshape = raid5_start_reshape,
4768 .quiesce = raid5_quiesce,
4770 static struct mdk_personality raid5_personality =
4774 .owner = THIS_MODULE,
4775 .make_request = make_request,
4779 .error_handler = error,
4780 .hot_add_disk = raid5_add_disk,
4781 .hot_remove_disk= raid5_remove_disk,
4782 .spare_active = raid5_spare_active,
4783 .sync_request = sync_request,
4784 .resize = raid5_resize,
4785 #ifdef CONFIG_MD_RAID5_RESHAPE
4786 .check_reshape = raid5_check_reshape,
4787 .start_reshape = raid5_start_reshape,
4789 .quiesce = raid5_quiesce,
4792 static struct mdk_personality raid4_personality =
4796 .owner = THIS_MODULE,
4797 .make_request = make_request,
4801 .error_handler = error,
4802 .hot_add_disk = raid5_add_disk,
4803 .hot_remove_disk= raid5_remove_disk,
4804 .spare_active = raid5_spare_active,
4805 .sync_request = sync_request,
4806 .resize = raid5_resize,
4807 #ifdef CONFIG_MD_RAID5_RESHAPE
4808 .check_reshape = raid5_check_reshape,
4809 .start_reshape = raid5_start_reshape,
4811 .quiesce = raid5_quiesce,
4814 static int __init raid5_init(void)
4818 e = raid6_select_algo();
4821 register_md_personality(&raid6_personality);
4822 register_md_personality(&raid5_personality);
4823 register_md_personality(&raid4_personality);
4827 static void raid5_exit(void)
4829 unregister_md_personality(&raid6_personality);
4830 unregister_md_personality(&raid5_personality);
4831 unregister_md_personality(&raid4_personality);
4834 module_init(raid5_init);
4835 module_exit(raid5_exit);
4836 MODULE_LICENSE("GPL");
4837 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4838 MODULE_ALIAS("md-raid5");
4839 MODULE_ALIAS("md-raid4");
4840 MODULE_ALIAS("md-level-5");
4841 MODULE_ALIAS("md-level-4");
4842 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4843 MODULE_ALIAS("md-raid6");
4844 MODULE_ALIAS("md-level-6");
4846 /* This used to be two separate modules, they were: */
4847 MODULE_ALIAS("raid5");
4848 MODULE_ALIAS("raid6");