2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
5 * Copyright (C) 2002, 2003 H. Peter Anvin
7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
30 * conf->bm_write is the number of the last batch successfully written.
31 * conf->bm_flush is the number of the last batch that was closed to
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35 * the number of the batch it will be in. This is bm_flush+1.
36 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/seq_file.h>
59 #define NR_STRIPES 256
60 #define STRIPE_SIZE PAGE_SIZE
61 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
62 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
63 #define IO_THRESHOLD 1
64 #define BYPASS_THRESHOLD 1
65 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
66 #define HASH_MASK (NR_HASH - 1)
68 #define stripe_hash(conf, sect) (&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
70 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
71 * order without overlap. There may be several bio's per stripe+device, and
72 * a bio could span several devices.
73 * When walking this list for a particular stripe+device, we must never proceed
74 * beyond a bio that extends past this device, as the next bio might no longer
76 * This macro is used to determine the 'next' bio in the list, given the sector
77 * of the current stripe+device
79 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
81 * The following can be used to debug the driver
83 #define RAID5_PARANOIA 1
84 #if RAID5_PARANOIA && defined(CONFIG_SMP)
85 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
87 # define CHECK_DEVLOCK()
95 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
98 * We maintain a biased count of active stripes in the bottom 16 bits of
99 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
101 static inline int raid5_bi_phys_segments(struct bio *bio)
103 return bio->bi_phys_segments & 0xffff;
106 static inline int raid5_bi_hw_segments(struct bio *bio)
108 return (bio->bi_phys_segments >> 16) & 0xffff;
111 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
113 --bio->bi_phys_segments;
114 return raid5_bi_phys_segments(bio);
117 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
119 unsigned short val = raid5_bi_hw_segments(bio);
122 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
126 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
128 bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
131 /* Find first data disk in a raid6 stripe */
132 static inline int raid6_d0(struct stripe_head *sh)
135 /* ddf always start from first device */
137 /* md starts just after Q block */
138 if (sh->qd_idx == sh->disks - 1)
141 return sh->qd_idx + 1;
143 static inline int raid6_next_disk(int disk, int raid_disks)
146 return (disk < raid_disks) ? disk : 0;
149 /* When walking through the disks in a raid5, starting at raid6_d0,
150 * We need to map each disk to a 'slot', where the data disks are slot
151 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
152 * is raid_disks-1. This help does that mapping.
154 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
155 int *count, int syndrome_disks)
159 if (idx == sh->pd_idx)
160 return syndrome_disks;
161 if (idx == sh->qd_idx)
162 return syndrome_disks + 1;
167 static void return_io(struct bio *return_bi)
169 struct bio *bi = return_bi;
172 return_bi = bi->bi_next;
180 static void print_raid5_conf (raid5_conf_t *conf);
182 static int stripe_operations_active(struct stripe_head *sh)
184 return sh->check_state || sh->reconstruct_state ||
185 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
186 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
189 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
191 if (atomic_dec_and_test(&sh->count)) {
192 BUG_ON(!list_empty(&sh->lru));
193 BUG_ON(atomic_read(&conf->active_stripes)==0);
194 if (test_bit(STRIPE_HANDLE, &sh->state)) {
195 if (test_bit(STRIPE_DELAYED, &sh->state)) {
196 list_add_tail(&sh->lru, &conf->delayed_list);
197 blk_plug_device(conf->mddev->queue);
198 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
199 sh->bm_seq - conf->seq_write > 0) {
200 list_add_tail(&sh->lru, &conf->bitmap_list);
201 blk_plug_device(conf->mddev->queue);
203 clear_bit(STRIPE_BIT_DELAY, &sh->state);
204 list_add_tail(&sh->lru, &conf->handle_list);
206 md_wakeup_thread(conf->mddev->thread);
208 BUG_ON(stripe_operations_active(sh));
209 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
210 atomic_dec(&conf->preread_active_stripes);
211 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
212 md_wakeup_thread(conf->mddev->thread);
214 atomic_dec(&conf->active_stripes);
215 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
216 list_add_tail(&sh->lru, &conf->inactive_list);
217 wake_up(&conf->wait_for_stripe);
218 if (conf->retry_read_aligned)
219 md_wakeup_thread(conf->mddev->thread);
225 static void release_stripe(struct stripe_head *sh)
227 raid5_conf_t *conf = sh->raid_conf;
230 spin_lock_irqsave(&conf->device_lock, flags);
231 __release_stripe(conf, sh);
232 spin_unlock_irqrestore(&conf->device_lock, flags);
235 static inline void remove_hash(struct stripe_head *sh)
237 pr_debug("remove_hash(), stripe %llu\n",
238 (unsigned long long)sh->sector);
240 hlist_del_init(&sh->hash);
243 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
245 struct hlist_head *hp = stripe_hash(conf, sh->sector);
247 pr_debug("insert_hash(), stripe %llu\n",
248 (unsigned long long)sh->sector);
251 hlist_add_head(&sh->hash, hp);
255 /* find an idle stripe, make sure it is unhashed, and return it. */
256 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
258 struct stripe_head *sh = NULL;
259 struct list_head *first;
262 if (list_empty(&conf->inactive_list))
264 first = conf->inactive_list.next;
265 sh = list_entry(first, struct stripe_head, lru);
266 list_del_init(first);
268 atomic_inc(&conf->active_stripes);
273 static void shrink_buffers(struct stripe_head *sh, int num)
278 for (i=0; i<num ; i++) {
282 sh->dev[i].page = NULL;
287 static int grow_buffers(struct stripe_head *sh, int num)
291 for (i=0; i<num; i++) {
294 if (!(page = alloc_page(GFP_KERNEL))) {
297 sh->dev[i].page = page;
302 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
303 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
304 struct stripe_head *sh);
306 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
308 raid5_conf_t *conf = sh->raid_conf;
311 BUG_ON(atomic_read(&sh->count) != 0);
312 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
313 BUG_ON(stripe_operations_active(sh));
316 pr_debug("init_stripe called, stripe %llu\n",
317 (unsigned long long)sh->sector);
321 sh->generation = conf->generation - previous;
322 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
324 stripe_set_idx(sector, conf, previous, sh);
328 for (i = sh->disks; i--; ) {
329 struct r5dev *dev = &sh->dev[i];
331 if (dev->toread || dev->read || dev->towrite || dev->written ||
332 test_bit(R5_LOCKED, &dev->flags)) {
333 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
334 (unsigned long long)sh->sector, i, dev->toread,
335 dev->read, dev->towrite, dev->written,
336 test_bit(R5_LOCKED, &dev->flags));
340 raid5_build_block(sh, i, previous);
342 insert_hash(conf, sh);
345 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
348 struct stripe_head *sh;
349 struct hlist_node *hn;
352 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
353 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
354 if (sh->sector == sector && sh->generation == generation)
356 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
360 static void unplug_slaves(mddev_t *mddev);
361 static void raid5_unplug_device(struct request_queue *q);
363 static struct stripe_head *
364 get_active_stripe(raid5_conf_t *conf, sector_t sector,
365 int previous, int noblock)
367 struct stripe_head *sh;
369 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
371 spin_lock_irq(&conf->device_lock);
374 wait_event_lock_irq(conf->wait_for_stripe,
376 conf->device_lock, /* nothing */);
377 sh = __find_stripe(conf, sector, conf->generation - previous);
379 if (!conf->inactive_blocked)
380 sh = get_free_stripe(conf);
381 if (noblock && sh == NULL)
384 conf->inactive_blocked = 1;
385 wait_event_lock_irq(conf->wait_for_stripe,
386 !list_empty(&conf->inactive_list) &&
387 (atomic_read(&conf->active_stripes)
388 < (conf->max_nr_stripes *3/4)
389 || !conf->inactive_blocked),
391 raid5_unplug_device(conf->mddev->queue)
393 conf->inactive_blocked = 0;
395 init_stripe(sh, sector, previous);
397 if (atomic_read(&sh->count)) {
398 BUG_ON(!list_empty(&sh->lru)
399 && !test_bit(STRIPE_EXPANDING, &sh->state));
401 if (!test_bit(STRIPE_HANDLE, &sh->state))
402 atomic_inc(&conf->active_stripes);
403 if (list_empty(&sh->lru) &&
404 !test_bit(STRIPE_EXPANDING, &sh->state))
406 list_del_init(&sh->lru);
409 } while (sh == NULL);
412 atomic_inc(&sh->count);
414 spin_unlock_irq(&conf->device_lock);
419 raid5_end_read_request(struct bio *bi, int error);
421 raid5_end_write_request(struct bio *bi, int error);
423 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
425 raid5_conf_t *conf = sh->raid_conf;
426 int i, disks = sh->disks;
430 for (i = disks; i--; ) {
434 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
436 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
441 bi = &sh->dev[i].req;
445 bi->bi_end_io = raid5_end_write_request;
447 bi->bi_end_io = raid5_end_read_request;
450 rdev = rcu_dereference(conf->disks[i].rdev);
451 if (rdev && test_bit(Faulty, &rdev->flags))
454 atomic_inc(&rdev->nr_pending);
458 if (s->syncing || s->expanding || s->expanded)
459 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
461 set_bit(STRIPE_IO_STARTED, &sh->state);
463 bi->bi_bdev = rdev->bdev;
464 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
465 __func__, (unsigned long long)sh->sector,
467 atomic_inc(&sh->count);
468 bi->bi_sector = sh->sector + rdev->data_offset;
469 bi->bi_flags = 1 << BIO_UPTODATE;
473 bi->bi_io_vec = &sh->dev[i].vec;
474 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
475 bi->bi_io_vec[0].bv_offset = 0;
476 bi->bi_size = STRIPE_SIZE;
479 test_bit(R5_ReWrite, &sh->dev[i].flags))
480 atomic_add(STRIPE_SECTORS,
481 &rdev->corrected_errors);
482 generic_make_request(bi);
485 set_bit(STRIPE_DEGRADED, &sh->state);
486 pr_debug("skip op %ld on disc %d for sector %llu\n",
487 bi->bi_rw, i, (unsigned long long)sh->sector);
488 clear_bit(R5_LOCKED, &sh->dev[i].flags);
489 set_bit(STRIPE_HANDLE, &sh->state);
494 static struct dma_async_tx_descriptor *
495 async_copy_data(int frombio, struct bio *bio, struct page *page,
496 sector_t sector, struct dma_async_tx_descriptor *tx)
499 struct page *bio_page;
503 if (bio->bi_sector >= sector)
504 page_offset = (signed)(bio->bi_sector - sector) * 512;
506 page_offset = (signed)(sector - bio->bi_sector) * -512;
507 bio_for_each_segment(bvl, bio, i) {
508 int len = bio_iovec_idx(bio, i)->bv_len;
512 if (page_offset < 0) {
513 b_offset = -page_offset;
514 page_offset += b_offset;
518 if (len > 0 && page_offset + len > STRIPE_SIZE)
519 clen = STRIPE_SIZE - page_offset;
524 b_offset += bio_iovec_idx(bio, i)->bv_offset;
525 bio_page = bio_iovec_idx(bio, i)->bv_page;
527 tx = async_memcpy(page, bio_page, page_offset,
532 tx = async_memcpy(bio_page, page, b_offset,
537 if (clen < len) /* hit end of page */
545 static void ops_complete_biofill(void *stripe_head_ref)
547 struct stripe_head *sh = stripe_head_ref;
548 struct bio *return_bi = NULL;
549 raid5_conf_t *conf = sh->raid_conf;
552 pr_debug("%s: stripe %llu\n", __func__,
553 (unsigned long long)sh->sector);
555 /* clear completed biofills */
556 spin_lock_irq(&conf->device_lock);
557 for (i = sh->disks; i--; ) {
558 struct r5dev *dev = &sh->dev[i];
560 /* acknowledge completion of a biofill operation */
561 /* and check if we need to reply to a read request,
562 * new R5_Wantfill requests are held off until
563 * !STRIPE_BIOFILL_RUN
565 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
566 struct bio *rbi, *rbi2;
571 while (rbi && rbi->bi_sector <
572 dev->sector + STRIPE_SECTORS) {
573 rbi2 = r5_next_bio(rbi, dev->sector);
574 if (!raid5_dec_bi_phys_segments(rbi)) {
575 rbi->bi_next = return_bi;
582 spin_unlock_irq(&conf->device_lock);
583 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
585 return_io(return_bi);
587 set_bit(STRIPE_HANDLE, &sh->state);
591 static void ops_run_biofill(struct stripe_head *sh)
593 struct dma_async_tx_descriptor *tx = NULL;
594 raid5_conf_t *conf = sh->raid_conf;
597 pr_debug("%s: stripe %llu\n", __func__,
598 (unsigned long long)sh->sector);
600 for (i = sh->disks; i--; ) {
601 struct r5dev *dev = &sh->dev[i];
602 if (test_bit(R5_Wantfill, &dev->flags)) {
604 spin_lock_irq(&conf->device_lock);
605 dev->read = rbi = dev->toread;
607 spin_unlock_irq(&conf->device_lock);
608 while (rbi && rbi->bi_sector <
609 dev->sector + STRIPE_SECTORS) {
610 tx = async_copy_data(0, rbi, dev->page,
612 rbi = r5_next_bio(rbi, dev->sector);
617 atomic_inc(&sh->count);
618 async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
619 ops_complete_biofill, sh);
622 static void ops_complete_compute5(void *stripe_head_ref)
624 struct stripe_head *sh = stripe_head_ref;
625 int target = sh->ops.target;
626 struct r5dev *tgt = &sh->dev[target];
628 pr_debug("%s: stripe %llu\n", __func__,
629 (unsigned long long)sh->sector);
631 set_bit(R5_UPTODATE, &tgt->flags);
632 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
633 clear_bit(R5_Wantcompute, &tgt->flags);
634 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
635 if (sh->check_state == check_state_compute_run)
636 sh->check_state = check_state_compute_result;
637 set_bit(STRIPE_HANDLE, &sh->state);
641 static struct dma_async_tx_descriptor *ops_run_compute5(struct stripe_head *sh)
643 /* kernel stack size limits the total number of disks */
644 int disks = sh->disks;
645 struct page *xor_srcs[disks];
646 int target = sh->ops.target;
647 struct r5dev *tgt = &sh->dev[target];
648 struct page *xor_dest = tgt->page;
650 struct dma_async_tx_descriptor *tx;
653 pr_debug("%s: stripe %llu block: %d\n",
654 __func__, (unsigned long long)sh->sector, target);
655 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
657 for (i = disks; i--; )
659 xor_srcs[count++] = sh->dev[i].page;
661 atomic_inc(&sh->count);
663 if (unlikely(count == 1))
664 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
665 0, NULL, ops_complete_compute5, sh);
667 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
668 ASYNC_TX_XOR_ZERO_DST, NULL,
669 ops_complete_compute5, sh);
674 static void ops_complete_prexor(void *stripe_head_ref)
676 struct stripe_head *sh = stripe_head_ref;
678 pr_debug("%s: stripe %llu\n", __func__,
679 (unsigned long long)sh->sector);
682 static struct dma_async_tx_descriptor *
683 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
685 /* kernel stack size limits the total number of disks */
686 int disks = sh->disks;
687 struct page *xor_srcs[disks];
688 int count = 0, pd_idx = sh->pd_idx, i;
690 /* existing parity data subtracted */
691 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
693 pr_debug("%s: stripe %llu\n", __func__,
694 (unsigned long long)sh->sector);
696 for (i = disks; i--; ) {
697 struct r5dev *dev = &sh->dev[i];
698 /* Only process blocks that are known to be uptodate */
699 if (test_bit(R5_Wantdrain, &dev->flags))
700 xor_srcs[count++] = dev->page;
703 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
704 ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
705 ops_complete_prexor, sh);
710 static struct dma_async_tx_descriptor *
711 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
713 int disks = sh->disks;
716 pr_debug("%s: stripe %llu\n", __func__,
717 (unsigned long long)sh->sector);
719 for (i = disks; i--; ) {
720 struct r5dev *dev = &sh->dev[i];
723 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
726 spin_lock(&sh->lock);
727 chosen = dev->towrite;
729 BUG_ON(dev->written);
730 wbi = dev->written = chosen;
731 spin_unlock(&sh->lock);
733 while (wbi && wbi->bi_sector <
734 dev->sector + STRIPE_SECTORS) {
735 tx = async_copy_data(1, wbi, dev->page,
737 wbi = r5_next_bio(wbi, dev->sector);
745 static void ops_complete_postxor(void *stripe_head_ref)
747 struct stripe_head *sh = stripe_head_ref;
748 int disks = sh->disks, i, pd_idx = sh->pd_idx;
750 pr_debug("%s: stripe %llu\n", __func__,
751 (unsigned long long)sh->sector);
753 for (i = disks; i--; ) {
754 struct r5dev *dev = &sh->dev[i];
755 if (dev->written || i == pd_idx)
756 set_bit(R5_UPTODATE, &dev->flags);
759 if (sh->reconstruct_state == reconstruct_state_drain_run)
760 sh->reconstruct_state = reconstruct_state_drain_result;
761 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
762 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
764 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
765 sh->reconstruct_state = reconstruct_state_result;
768 set_bit(STRIPE_HANDLE, &sh->state);
773 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
775 /* kernel stack size limits the total number of disks */
776 int disks = sh->disks;
777 struct page *xor_srcs[disks];
779 int count = 0, pd_idx = sh->pd_idx, i;
780 struct page *xor_dest;
784 pr_debug("%s: stripe %llu\n", __func__,
785 (unsigned long long)sh->sector);
787 /* check if prexor is active which means only process blocks
788 * that are part of a read-modify-write (written)
790 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
792 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
793 for (i = disks; i--; ) {
794 struct r5dev *dev = &sh->dev[i];
796 xor_srcs[count++] = dev->page;
799 xor_dest = sh->dev[pd_idx].page;
800 for (i = disks; i--; ) {
801 struct r5dev *dev = &sh->dev[i];
803 xor_srcs[count++] = dev->page;
807 /* 1/ if we prexor'd then the dest is reused as a source
808 * 2/ if we did not prexor then we are redoing the parity
809 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
810 * for the synchronous xor case
812 flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
813 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
815 atomic_inc(&sh->count);
817 if (unlikely(count == 1)) {
818 flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
819 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
820 flags, tx, ops_complete_postxor, sh);
822 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
823 flags, tx, ops_complete_postxor, sh);
826 static void ops_complete_check(void *stripe_head_ref)
828 struct stripe_head *sh = stripe_head_ref;
830 pr_debug("%s: stripe %llu\n", __func__,
831 (unsigned long long)sh->sector);
833 sh->check_state = check_state_check_result;
834 set_bit(STRIPE_HANDLE, &sh->state);
838 static void ops_run_check(struct stripe_head *sh)
840 /* kernel stack size limits the total number of disks */
841 int disks = sh->disks;
842 struct page *xor_srcs[disks];
843 struct dma_async_tx_descriptor *tx;
845 int count = 0, pd_idx = sh->pd_idx, i;
846 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
848 pr_debug("%s: stripe %llu\n", __func__,
849 (unsigned long long)sh->sector);
851 for (i = disks; i--; ) {
852 struct r5dev *dev = &sh->dev[i];
854 xor_srcs[count++] = dev->page;
857 tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
858 &sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
860 atomic_inc(&sh->count);
861 tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
862 ops_complete_check, sh);
865 static void raid5_run_ops(struct stripe_head *sh, unsigned long ops_request)
867 int overlap_clear = 0, i, disks = sh->disks;
868 struct dma_async_tx_descriptor *tx = NULL;
870 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
875 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
876 tx = ops_run_compute5(sh);
877 /* terminate the chain if postxor is not set to be run */
878 if (tx && !test_bit(STRIPE_OP_POSTXOR, &ops_request))
882 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
883 tx = ops_run_prexor(sh, tx);
885 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
886 tx = ops_run_biodrain(sh, tx);
890 if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
891 ops_run_postxor(sh, tx);
893 if (test_bit(STRIPE_OP_CHECK, &ops_request))
897 for (i = disks; i--; ) {
898 struct r5dev *dev = &sh->dev[i];
899 if (test_and_clear_bit(R5_Overlap, &dev->flags))
900 wake_up(&sh->raid_conf->wait_for_overlap);
904 static int grow_one_stripe(raid5_conf_t *conf)
906 struct stripe_head *sh;
907 sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
910 memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
911 sh->raid_conf = conf;
912 spin_lock_init(&sh->lock);
914 if (grow_buffers(sh, conf->raid_disks)) {
915 shrink_buffers(sh, conf->raid_disks);
916 kmem_cache_free(conf->slab_cache, sh);
919 sh->disks = conf->raid_disks;
920 /* we just created an active stripe so... */
921 atomic_set(&sh->count, 1);
922 atomic_inc(&conf->active_stripes);
923 INIT_LIST_HEAD(&sh->lru);
928 static int grow_stripes(raid5_conf_t *conf, int num)
930 struct kmem_cache *sc;
931 int devs = conf->raid_disks;
933 sprintf(conf->cache_name[0],
934 "raid%d-%s", conf->level, mdname(conf->mddev));
935 sprintf(conf->cache_name[1],
936 "raid%d-%s-alt", conf->level, mdname(conf->mddev));
937 conf->active_name = 0;
938 sc = kmem_cache_create(conf->cache_name[conf->active_name],
939 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
943 conf->slab_cache = sc;
944 conf->pool_size = devs;
946 if (!grow_one_stripe(conf))
951 #ifdef CONFIG_MD_RAID5_RESHAPE
952 static int resize_stripes(raid5_conf_t *conf, int newsize)
954 /* Make all the stripes able to hold 'newsize' devices.
955 * New slots in each stripe get 'page' set to a new page.
957 * This happens in stages:
958 * 1/ create a new kmem_cache and allocate the required number of
960 * 2/ gather all the old stripe_heads and tranfer the pages across
961 * to the new stripe_heads. This will have the side effect of
962 * freezing the array as once all stripe_heads have been collected,
963 * no IO will be possible. Old stripe heads are freed once their
964 * pages have been transferred over, and the old kmem_cache is
965 * freed when all stripes are done.
966 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
967 * we simple return a failre status - no need to clean anything up.
968 * 4/ allocate new pages for the new slots in the new stripe_heads.
969 * If this fails, we don't bother trying the shrink the
970 * stripe_heads down again, we just leave them as they are.
971 * As each stripe_head is processed the new one is released into
974 * Once step2 is started, we cannot afford to wait for a write,
975 * so we use GFP_NOIO allocations.
977 struct stripe_head *osh, *nsh;
978 LIST_HEAD(newstripes);
979 struct disk_info *ndisks;
981 struct kmem_cache *sc;
984 if (newsize <= conf->pool_size)
985 return 0; /* never bother to shrink */
987 err = md_allow_write(conf->mddev);
992 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
993 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
998 for (i = conf->max_nr_stripes; i; i--) {
999 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1003 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1005 nsh->raid_conf = conf;
1006 spin_lock_init(&nsh->lock);
1008 list_add(&nsh->lru, &newstripes);
1011 /* didn't get enough, give up */
1012 while (!list_empty(&newstripes)) {
1013 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1014 list_del(&nsh->lru);
1015 kmem_cache_free(sc, nsh);
1017 kmem_cache_destroy(sc);
1020 /* Step 2 - Must use GFP_NOIO now.
1021 * OK, we have enough stripes, start collecting inactive
1022 * stripes and copying them over
1024 list_for_each_entry(nsh, &newstripes, lru) {
1025 spin_lock_irq(&conf->device_lock);
1026 wait_event_lock_irq(conf->wait_for_stripe,
1027 !list_empty(&conf->inactive_list),
1029 unplug_slaves(conf->mddev)
1031 osh = get_free_stripe(conf);
1032 spin_unlock_irq(&conf->device_lock);
1033 atomic_set(&nsh->count, 1);
1034 for(i=0; i<conf->pool_size; i++)
1035 nsh->dev[i].page = osh->dev[i].page;
1036 for( ; i<newsize; i++)
1037 nsh->dev[i].page = NULL;
1038 kmem_cache_free(conf->slab_cache, osh);
1040 kmem_cache_destroy(conf->slab_cache);
1043 * At this point, we are holding all the stripes so the array
1044 * is completely stalled, so now is a good time to resize
1047 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1049 for (i=0; i<conf->raid_disks; i++)
1050 ndisks[i] = conf->disks[i];
1052 conf->disks = ndisks;
1056 /* Step 4, return new stripes to service */
1057 while(!list_empty(&newstripes)) {
1058 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1059 list_del_init(&nsh->lru);
1060 for (i=conf->raid_disks; i < newsize; i++)
1061 if (nsh->dev[i].page == NULL) {
1062 struct page *p = alloc_page(GFP_NOIO);
1063 nsh->dev[i].page = p;
1067 release_stripe(nsh);
1069 /* critical section pass, GFP_NOIO no longer needed */
1071 conf->slab_cache = sc;
1072 conf->active_name = 1-conf->active_name;
1073 conf->pool_size = newsize;
1078 static int drop_one_stripe(raid5_conf_t *conf)
1080 struct stripe_head *sh;
1082 spin_lock_irq(&conf->device_lock);
1083 sh = get_free_stripe(conf);
1084 spin_unlock_irq(&conf->device_lock);
1087 BUG_ON(atomic_read(&sh->count));
1088 shrink_buffers(sh, conf->pool_size);
1089 kmem_cache_free(conf->slab_cache, sh);
1090 atomic_dec(&conf->active_stripes);
1094 static void shrink_stripes(raid5_conf_t *conf)
1096 while (drop_one_stripe(conf))
1099 if (conf->slab_cache)
1100 kmem_cache_destroy(conf->slab_cache);
1101 conf->slab_cache = NULL;
1104 static void raid5_end_read_request(struct bio * bi, int error)
1106 struct stripe_head *sh = bi->bi_private;
1107 raid5_conf_t *conf = sh->raid_conf;
1108 int disks = sh->disks, i;
1109 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1110 char b[BDEVNAME_SIZE];
1114 for (i=0 ; i<disks; i++)
1115 if (bi == &sh->dev[i].req)
1118 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1119 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1127 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1128 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1129 rdev = conf->disks[i].rdev;
1130 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1131 " (%lu sectors at %llu on %s)\n",
1132 mdname(conf->mddev), STRIPE_SECTORS,
1133 (unsigned long long)(sh->sector
1134 + rdev->data_offset),
1135 bdevname(rdev->bdev, b));
1136 clear_bit(R5_ReadError, &sh->dev[i].flags);
1137 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1139 if (atomic_read(&conf->disks[i].rdev->read_errors))
1140 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1142 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1144 rdev = conf->disks[i].rdev;
1146 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1147 atomic_inc(&rdev->read_errors);
1148 if (conf->mddev->degraded)
1149 printk_rl(KERN_WARNING
1150 "raid5:%s: read error not correctable "
1151 "(sector %llu on %s).\n",
1152 mdname(conf->mddev),
1153 (unsigned long long)(sh->sector
1154 + rdev->data_offset),
1156 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1158 printk_rl(KERN_WARNING
1159 "raid5:%s: read error NOT corrected!! "
1160 "(sector %llu on %s).\n",
1161 mdname(conf->mddev),
1162 (unsigned long long)(sh->sector
1163 + rdev->data_offset),
1165 else if (atomic_read(&rdev->read_errors)
1166 > conf->max_nr_stripes)
1168 "raid5:%s: Too many read errors, failing device %s.\n",
1169 mdname(conf->mddev), bdn);
1173 set_bit(R5_ReadError, &sh->dev[i].flags);
1175 clear_bit(R5_ReadError, &sh->dev[i].flags);
1176 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1177 md_error(conf->mddev, rdev);
1180 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1181 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1182 set_bit(STRIPE_HANDLE, &sh->state);
1186 static void raid5_end_write_request(struct bio *bi, int error)
1188 struct stripe_head *sh = bi->bi_private;
1189 raid5_conf_t *conf = sh->raid_conf;
1190 int disks = sh->disks, i;
1191 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1193 for (i=0 ; i<disks; i++)
1194 if (bi == &sh->dev[i].req)
1197 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1198 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1206 md_error(conf->mddev, conf->disks[i].rdev);
1208 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1210 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1211 set_bit(STRIPE_HANDLE, &sh->state);
1216 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1218 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1220 struct r5dev *dev = &sh->dev[i];
1222 bio_init(&dev->req);
1223 dev->req.bi_io_vec = &dev->vec;
1225 dev->req.bi_max_vecs++;
1226 dev->vec.bv_page = dev->page;
1227 dev->vec.bv_len = STRIPE_SIZE;
1228 dev->vec.bv_offset = 0;
1230 dev->req.bi_sector = sh->sector;
1231 dev->req.bi_private = sh;
1234 dev->sector = compute_blocknr(sh, i, previous);
1237 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1239 char b[BDEVNAME_SIZE];
1240 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1241 pr_debug("raid5: error called\n");
1243 if (!test_bit(Faulty, &rdev->flags)) {
1244 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1245 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1246 unsigned long flags;
1247 spin_lock_irqsave(&conf->device_lock, flags);
1249 spin_unlock_irqrestore(&conf->device_lock, flags);
1251 * if recovery was running, make sure it aborts.
1253 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1255 set_bit(Faulty, &rdev->flags);
1257 "raid5: Disk failure on %s, disabling device.\n"
1258 "raid5: Operation continuing on %d devices.\n",
1259 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1264 * Input: a 'big' sector number,
1265 * Output: index of the data and parity disk, and the sector # in them.
1267 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1268 int previous, int *dd_idx,
1269 struct stripe_head *sh)
1272 unsigned long chunk_number;
1273 unsigned int chunk_offset;
1276 sector_t new_sector;
1277 int algorithm = previous ? conf->prev_algo
1279 int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1280 : (conf->chunk_size >> 9);
1281 int raid_disks = previous ? conf->previous_raid_disks
1283 int data_disks = raid_disks - conf->max_degraded;
1285 /* First compute the information on this sector */
1288 * Compute the chunk number and the sector offset inside the chunk
1290 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1291 chunk_number = r_sector;
1292 BUG_ON(r_sector != chunk_number);
1295 * Compute the stripe number
1297 stripe = chunk_number / data_disks;
1300 * Compute the data disk and parity disk indexes inside the stripe
1302 *dd_idx = chunk_number % data_disks;
1305 * Select the parity disk based on the user selected algorithm.
1307 pd_idx = qd_idx = ~0;
1308 switch(conf->level) {
1310 pd_idx = data_disks;
1313 switch (algorithm) {
1314 case ALGORITHM_LEFT_ASYMMETRIC:
1315 pd_idx = data_disks - stripe % raid_disks;
1316 if (*dd_idx >= pd_idx)
1319 case ALGORITHM_RIGHT_ASYMMETRIC:
1320 pd_idx = stripe % raid_disks;
1321 if (*dd_idx >= pd_idx)
1324 case ALGORITHM_LEFT_SYMMETRIC:
1325 pd_idx = data_disks - stripe % raid_disks;
1326 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1328 case ALGORITHM_RIGHT_SYMMETRIC:
1329 pd_idx = stripe % raid_disks;
1330 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1332 case ALGORITHM_PARITY_0:
1336 case ALGORITHM_PARITY_N:
1337 pd_idx = data_disks;
1340 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1347 switch (algorithm) {
1348 case ALGORITHM_LEFT_ASYMMETRIC:
1349 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1350 qd_idx = pd_idx + 1;
1351 if (pd_idx == raid_disks-1) {
1352 (*dd_idx)++; /* Q D D D P */
1354 } else if (*dd_idx >= pd_idx)
1355 (*dd_idx) += 2; /* D D P Q D */
1357 case ALGORITHM_RIGHT_ASYMMETRIC:
1358 pd_idx = stripe % raid_disks;
1359 qd_idx = pd_idx + 1;
1360 if (pd_idx == raid_disks-1) {
1361 (*dd_idx)++; /* Q D D D P */
1363 } else if (*dd_idx >= pd_idx)
1364 (*dd_idx) += 2; /* D D P Q D */
1366 case ALGORITHM_LEFT_SYMMETRIC:
1367 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1368 qd_idx = (pd_idx + 1) % raid_disks;
1369 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1371 case ALGORITHM_RIGHT_SYMMETRIC:
1372 pd_idx = stripe % raid_disks;
1373 qd_idx = (pd_idx + 1) % raid_disks;
1374 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1377 case ALGORITHM_PARITY_0:
1382 case ALGORITHM_PARITY_N:
1383 pd_idx = data_disks;
1384 qd_idx = data_disks + 1;
1387 case ALGORITHM_ROTATING_ZERO_RESTART:
1388 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1389 * of blocks for computing Q is different.
1391 pd_idx = stripe % raid_disks;
1392 qd_idx = pd_idx + 1;
1393 if (pd_idx == raid_disks-1) {
1394 (*dd_idx)++; /* Q D D D P */
1396 } else if (*dd_idx >= pd_idx)
1397 (*dd_idx) += 2; /* D D P Q D */
1401 case ALGORITHM_ROTATING_N_RESTART:
1402 /* Same a left_asymmetric, by first stripe is
1403 * D D D P Q rather than
1406 pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1407 qd_idx = pd_idx + 1;
1408 if (pd_idx == raid_disks-1) {
1409 (*dd_idx)++; /* Q D D D P */
1411 } else if (*dd_idx >= pd_idx)
1412 (*dd_idx) += 2; /* D D P Q D */
1416 case ALGORITHM_ROTATING_N_CONTINUE:
1417 /* Same as left_symmetric but Q is before P */
1418 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1419 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1420 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1424 case ALGORITHM_LEFT_ASYMMETRIC_6:
1425 /* RAID5 left_asymmetric, with Q on last device */
1426 pd_idx = data_disks - stripe % (raid_disks-1);
1427 if (*dd_idx >= pd_idx)
1429 qd_idx = raid_disks - 1;
1432 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1433 pd_idx = stripe % (raid_disks-1);
1434 if (*dd_idx >= pd_idx)
1436 qd_idx = raid_disks - 1;
1439 case ALGORITHM_LEFT_SYMMETRIC_6:
1440 pd_idx = data_disks - stripe % (raid_disks-1);
1441 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1442 qd_idx = raid_disks - 1;
1445 case ALGORITHM_RIGHT_SYMMETRIC_6:
1446 pd_idx = stripe % (raid_disks-1);
1447 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1448 qd_idx = raid_disks - 1;
1451 case ALGORITHM_PARITY_0_6:
1454 qd_idx = raid_disks - 1;
1459 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1467 sh->pd_idx = pd_idx;
1468 sh->qd_idx = qd_idx;
1469 sh->ddf_layout = ddf_layout;
1472 * Finally, compute the new sector number
1474 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1479 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1481 raid5_conf_t *conf = sh->raid_conf;
1482 int raid_disks = sh->disks;
1483 int data_disks = raid_disks - conf->max_degraded;
1484 sector_t new_sector = sh->sector, check;
1485 int sectors_per_chunk = previous ? (conf->prev_chunk >> 9)
1486 : (conf->chunk_size >> 9);
1487 int algorithm = previous ? conf->prev_algo
1491 int chunk_number, dummy1, dd_idx = i;
1493 struct stripe_head sh2;
1496 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1497 stripe = new_sector;
1498 BUG_ON(new_sector != stripe);
1500 if (i == sh->pd_idx)
1502 switch(conf->level) {
1505 switch (algorithm) {
1506 case ALGORITHM_LEFT_ASYMMETRIC:
1507 case ALGORITHM_RIGHT_ASYMMETRIC:
1511 case ALGORITHM_LEFT_SYMMETRIC:
1512 case ALGORITHM_RIGHT_SYMMETRIC:
1515 i -= (sh->pd_idx + 1);
1517 case ALGORITHM_PARITY_0:
1520 case ALGORITHM_PARITY_N:
1523 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1529 if (i == sh->qd_idx)
1530 return 0; /* It is the Q disk */
1531 switch (algorithm) {
1532 case ALGORITHM_LEFT_ASYMMETRIC:
1533 case ALGORITHM_RIGHT_ASYMMETRIC:
1534 case ALGORITHM_ROTATING_ZERO_RESTART:
1535 case ALGORITHM_ROTATING_N_RESTART:
1536 if (sh->pd_idx == raid_disks-1)
1537 i--; /* Q D D D P */
1538 else if (i > sh->pd_idx)
1539 i -= 2; /* D D P Q D */
1541 case ALGORITHM_LEFT_SYMMETRIC:
1542 case ALGORITHM_RIGHT_SYMMETRIC:
1543 if (sh->pd_idx == raid_disks-1)
1544 i--; /* Q D D D P */
1549 i -= (sh->pd_idx + 2);
1552 case ALGORITHM_PARITY_0:
1555 case ALGORITHM_PARITY_N:
1557 case ALGORITHM_ROTATING_N_CONTINUE:
1558 if (sh->pd_idx == 0)
1559 i--; /* P D D D Q */
1560 else if (i > sh->pd_idx)
1561 i -= 2; /* D D Q P D */
1563 case ALGORITHM_LEFT_ASYMMETRIC_6:
1564 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1568 case ALGORITHM_LEFT_SYMMETRIC_6:
1569 case ALGORITHM_RIGHT_SYMMETRIC_6:
1571 i += data_disks + 1;
1572 i -= (sh->pd_idx + 1);
1574 case ALGORITHM_PARITY_0_6:
1578 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1585 chunk_number = stripe * data_disks + i;
1586 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1588 check = raid5_compute_sector(conf, r_sector,
1589 previous, &dummy1, &sh2);
1590 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1591 || sh2.qd_idx != sh->qd_idx) {
1592 printk(KERN_ERR "compute_blocknr: map not correct\n");
1601 * Copy data between a page in the stripe cache, and one or more bion
1602 * The page could align with the middle of the bio, or there could be
1603 * several bion, each with several bio_vecs, which cover part of the page
1604 * Multiple bion are linked together on bi_next. There may be extras
1605 * at the end of this list. We ignore them.
1607 static void copy_data(int frombio, struct bio *bio,
1611 char *pa = page_address(page);
1612 struct bio_vec *bvl;
1616 if (bio->bi_sector >= sector)
1617 page_offset = (signed)(bio->bi_sector - sector) * 512;
1619 page_offset = (signed)(sector - bio->bi_sector) * -512;
1620 bio_for_each_segment(bvl, bio, i) {
1621 int len = bio_iovec_idx(bio,i)->bv_len;
1625 if (page_offset < 0) {
1626 b_offset = -page_offset;
1627 page_offset += b_offset;
1631 if (len > 0 && page_offset + len > STRIPE_SIZE)
1632 clen = STRIPE_SIZE - page_offset;
1636 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1638 memcpy(pa+page_offset, ba+b_offset, clen);
1640 memcpy(ba+b_offset, pa+page_offset, clen);
1641 __bio_kunmap_atomic(ba, KM_USER0);
1643 if (clen < len) /* hit end of page */
1649 #define check_xor() do { \
1650 if (count == MAX_XOR_BLOCKS) { \
1651 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1656 static void compute_parity6(struct stripe_head *sh, int method)
1658 raid5_conf_t *conf = sh->raid_conf;
1659 int i, pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1660 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1662 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1663 void *ptrs[syndrome_disks+2];
1665 pd_idx = sh->pd_idx;
1666 qd_idx = sh->qd_idx;
1667 d0_idx = raid6_d0(sh);
1669 pr_debug("compute_parity, stripe %llu, method %d\n",
1670 (unsigned long long)sh->sector, method);
1673 case READ_MODIFY_WRITE:
1674 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1675 case RECONSTRUCT_WRITE:
1676 for (i= disks; i-- ;)
1677 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1678 chosen = sh->dev[i].towrite;
1679 sh->dev[i].towrite = NULL;
1681 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1682 wake_up(&conf->wait_for_overlap);
1684 BUG_ON(sh->dev[i].written);
1685 sh->dev[i].written = chosen;
1689 BUG(); /* Not implemented yet */
1692 for (i = disks; i--;)
1693 if (sh->dev[i].written) {
1694 sector_t sector = sh->dev[i].sector;
1695 struct bio *wbi = sh->dev[i].written;
1696 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1697 copy_data(1, wbi, sh->dev[i].page, sector);
1698 wbi = r5_next_bio(wbi, sector);
1701 set_bit(R5_LOCKED, &sh->dev[i].flags);
1702 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1705 /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1707 for (i = 0; i < disks; i++)
1708 ptrs[i] = (void *)raid6_empty_zero_page;
1713 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1715 ptrs[slot] = page_address(sh->dev[i].page);
1716 if (slot < syndrome_disks &&
1717 !test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1718 printk(KERN_ERR "block %d/%d not uptodate "
1719 "on parity calc\n", i, count);
1723 i = raid6_next_disk(i, disks);
1724 } while (i != d0_idx);
1725 BUG_ON(count != syndrome_disks);
1727 raid6_call.gen_syndrome(syndrome_disks+2, STRIPE_SIZE, ptrs);
1730 case RECONSTRUCT_WRITE:
1731 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1732 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1733 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1734 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1737 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1738 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1744 /* Compute one missing block */
1745 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1747 int i, count, disks = sh->disks;
1748 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1749 int qd_idx = sh->qd_idx;
1751 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1752 (unsigned long long)sh->sector, dd_idx);
1754 if ( dd_idx == qd_idx ) {
1755 /* We're actually computing the Q drive */
1756 compute_parity6(sh, UPDATE_PARITY);
1758 dest = page_address(sh->dev[dd_idx].page);
1759 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1761 for (i = disks ; i--; ) {
1762 if (i == dd_idx || i == qd_idx)
1764 p = page_address(sh->dev[i].page);
1765 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1768 printk("compute_block() %d, stripe %llu, %d"
1769 " not present\n", dd_idx,
1770 (unsigned long long)sh->sector, i);
1775 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1776 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1777 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1781 /* Compute two missing blocks */
1782 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1784 int i, count, disks = sh->disks;
1785 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1786 int d0_idx = raid6_d0(sh);
1787 int faila = -1, failb = -1;
1788 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1789 void *ptrs[syndrome_disks+2];
1791 for (i = 0; i < disks ; i++)
1792 ptrs[i] = (void *)raid6_empty_zero_page;
1796 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1798 ptrs[slot] = page_address(sh->dev[i].page);
1804 i = raid6_next_disk(i, disks);
1805 } while (i != d0_idx);
1806 BUG_ON(count != syndrome_disks);
1808 BUG_ON(faila == failb);
1809 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1811 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1812 (unsigned long long)sh->sector, dd_idx1, dd_idx2,
1815 if (failb == syndrome_disks+1) {
1816 /* Q disk is one of the missing disks */
1817 if (faila == syndrome_disks) {
1818 /* Missing P+Q, just recompute */
1819 compute_parity6(sh, UPDATE_PARITY);
1822 /* We're missing D+Q; recompute D from P */
1823 compute_block_1(sh, ((dd_idx1 == sh->qd_idx) ?
1826 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1831 /* We're missing D+P or D+D; */
1832 if (failb == syndrome_disks) {
1833 /* We're missing D+P. */
1834 raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE, faila, ptrs);
1836 /* We're missing D+D. */
1837 raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE, faila, failb,
1841 /* Both the above update both missing blocks */
1842 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1843 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1847 schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
1848 int rcw, int expand)
1850 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1853 /* if we are not expanding this is a proper write request, and
1854 * there will be bios with new data to be drained into the
1858 sh->reconstruct_state = reconstruct_state_drain_run;
1859 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1861 sh->reconstruct_state = reconstruct_state_run;
1863 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1865 for (i = disks; i--; ) {
1866 struct r5dev *dev = &sh->dev[i];
1869 set_bit(R5_LOCKED, &dev->flags);
1870 set_bit(R5_Wantdrain, &dev->flags);
1872 clear_bit(R5_UPTODATE, &dev->flags);
1876 if (s->locked + 1 == disks)
1877 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1878 atomic_inc(&sh->raid_conf->pending_full_writes);
1880 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1881 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1883 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1884 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1885 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1886 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1888 for (i = disks; i--; ) {
1889 struct r5dev *dev = &sh->dev[i];
1894 (test_bit(R5_UPTODATE, &dev->flags) ||
1895 test_bit(R5_Wantcompute, &dev->flags))) {
1896 set_bit(R5_Wantdrain, &dev->flags);
1897 set_bit(R5_LOCKED, &dev->flags);
1898 clear_bit(R5_UPTODATE, &dev->flags);
1904 /* keep the parity disk locked while asynchronous operations
1907 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1908 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1911 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1912 __func__, (unsigned long long)sh->sector,
1913 s->locked, s->ops_request);
1917 * Each stripe/dev can have one or more bion attached.
1918 * toread/towrite point to the first in a chain.
1919 * The bi_next chain must be in order.
1921 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1924 raid5_conf_t *conf = sh->raid_conf;
1927 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1928 (unsigned long long)bi->bi_sector,
1929 (unsigned long long)sh->sector);
1932 spin_lock(&sh->lock);
1933 spin_lock_irq(&conf->device_lock);
1935 bip = &sh->dev[dd_idx].towrite;
1936 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1939 bip = &sh->dev[dd_idx].toread;
1940 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1941 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1943 bip = & (*bip)->bi_next;
1945 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1948 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1952 bi->bi_phys_segments++;
1953 spin_unlock_irq(&conf->device_lock);
1954 spin_unlock(&sh->lock);
1956 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1957 (unsigned long long)bi->bi_sector,
1958 (unsigned long long)sh->sector, dd_idx);
1960 if (conf->mddev->bitmap && firstwrite) {
1961 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1963 sh->bm_seq = conf->seq_flush+1;
1964 set_bit(STRIPE_BIT_DELAY, &sh->state);
1968 /* check if page is covered */
1969 sector_t sector = sh->dev[dd_idx].sector;
1970 for (bi=sh->dev[dd_idx].towrite;
1971 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1972 bi && bi->bi_sector <= sector;
1973 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1974 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1975 sector = bi->bi_sector + (bi->bi_size>>9);
1977 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1978 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1983 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1984 spin_unlock_irq(&conf->device_lock);
1985 spin_unlock(&sh->lock);
1989 static void end_reshape(raid5_conf_t *conf);
1991 static int page_is_zero(struct page *p)
1993 char *a = page_address(p);
1994 return ((*(u32*)a) == 0 &&
1995 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1998 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
1999 struct stripe_head *sh)
2001 int sectors_per_chunk =
2002 previous ? (conf->prev_chunk >> 9)
2003 : (conf->chunk_size >> 9);
2005 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2006 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2008 raid5_compute_sector(conf,
2009 stripe * (disks - conf->max_degraded)
2010 *sectors_per_chunk + chunk_offset,
2016 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2017 struct stripe_head_state *s, int disks,
2018 struct bio **return_bi)
2021 for (i = disks; i--; ) {
2025 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2028 rdev = rcu_dereference(conf->disks[i].rdev);
2029 if (rdev && test_bit(In_sync, &rdev->flags))
2030 /* multiple read failures in one stripe */
2031 md_error(conf->mddev, rdev);
2034 spin_lock_irq(&conf->device_lock);
2035 /* fail all writes first */
2036 bi = sh->dev[i].towrite;
2037 sh->dev[i].towrite = NULL;
2043 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2044 wake_up(&conf->wait_for_overlap);
2046 while (bi && bi->bi_sector <
2047 sh->dev[i].sector + STRIPE_SECTORS) {
2048 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2049 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2050 if (!raid5_dec_bi_phys_segments(bi)) {
2051 md_write_end(conf->mddev);
2052 bi->bi_next = *return_bi;
2057 /* and fail all 'written' */
2058 bi = sh->dev[i].written;
2059 sh->dev[i].written = NULL;
2060 if (bi) bitmap_end = 1;
2061 while (bi && bi->bi_sector <
2062 sh->dev[i].sector + STRIPE_SECTORS) {
2063 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2064 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2065 if (!raid5_dec_bi_phys_segments(bi)) {
2066 md_write_end(conf->mddev);
2067 bi->bi_next = *return_bi;
2073 /* fail any reads if this device is non-operational and
2074 * the data has not reached the cache yet.
2076 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2077 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2078 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2079 bi = sh->dev[i].toread;
2080 sh->dev[i].toread = NULL;
2081 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2082 wake_up(&conf->wait_for_overlap);
2083 if (bi) s->to_read--;
2084 while (bi && bi->bi_sector <
2085 sh->dev[i].sector + STRIPE_SECTORS) {
2086 struct bio *nextbi =
2087 r5_next_bio(bi, sh->dev[i].sector);
2088 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2089 if (!raid5_dec_bi_phys_segments(bi)) {
2090 bi->bi_next = *return_bi;
2096 spin_unlock_irq(&conf->device_lock);
2098 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2099 STRIPE_SECTORS, 0, 0);
2102 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2103 if (atomic_dec_and_test(&conf->pending_full_writes))
2104 md_wakeup_thread(conf->mddev->thread);
2107 /* fetch_block5 - checks the given member device to see if its data needs
2108 * to be read or computed to satisfy a request.
2110 * Returns 1 when no more member devices need to be checked, otherwise returns
2111 * 0 to tell the loop in handle_stripe_fill5 to continue
2113 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2114 int disk_idx, int disks)
2116 struct r5dev *dev = &sh->dev[disk_idx];
2117 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2119 /* is the data in this block needed, and can we get it? */
2120 if (!test_bit(R5_LOCKED, &dev->flags) &&
2121 !test_bit(R5_UPTODATE, &dev->flags) &&
2123 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2124 s->syncing || s->expanding ||
2126 (failed_dev->toread ||
2127 (failed_dev->towrite &&
2128 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2129 /* We would like to get this block, possibly by computing it,
2130 * otherwise read it if the backing disk is insync
2132 if ((s->uptodate == disks - 1) &&
2133 (s->failed && disk_idx == s->failed_num)) {
2134 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2135 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2136 set_bit(R5_Wantcompute, &dev->flags);
2137 sh->ops.target = disk_idx;
2139 /* Careful: from this point on 'uptodate' is in the eye
2140 * of raid5_run_ops which services 'compute' operations
2141 * before writes. R5_Wantcompute flags a block that will
2142 * be R5_UPTODATE by the time it is needed for a
2143 * subsequent operation.
2146 return 1; /* uptodate + compute == disks */
2147 } else if (test_bit(R5_Insync, &dev->flags)) {
2148 set_bit(R5_LOCKED, &dev->flags);
2149 set_bit(R5_Wantread, &dev->flags);
2151 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2160 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2162 static void handle_stripe_fill5(struct stripe_head *sh,
2163 struct stripe_head_state *s, int disks)
2167 /* look for blocks to read/compute, skip this if a compute
2168 * is already in flight, or if the stripe contents are in the
2169 * midst of changing due to a write
2171 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2172 !sh->reconstruct_state)
2173 for (i = disks; i--; )
2174 if (fetch_block5(sh, s, i, disks))
2176 set_bit(STRIPE_HANDLE, &sh->state);
2179 static void handle_stripe_fill6(struct stripe_head *sh,
2180 struct stripe_head_state *s, struct r6_state *r6s,
2184 for (i = disks; i--; ) {
2185 struct r5dev *dev = &sh->dev[i];
2186 if (!test_bit(R5_LOCKED, &dev->flags) &&
2187 !test_bit(R5_UPTODATE, &dev->flags) &&
2188 (dev->toread || (dev->towrite &&
2189 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2190 s->syncing || s->expanding ||
2192 (sh->dev[r6s->failed_num[0]].toread ||
2195 (sh->dev[r6s->failed_num[1]].toread ||
2197 /* we would like to get this block, possibly
2198 * by computing it, but we might not be able to
2200 if ((s->uptodate == disks - 1) &&
2201 (s->failed && (i == r6s->failed_num[0] ||
2202 i == r6s->failed_num[1]))) {
2203 pr_debug("Computing stripe %llu block %d\n",
2204 (unsigned long long)sh->sector, i);
2205 compute_block_1(sh, i, 0);
2207 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2208 /* Computing 2-failure is *very* expensive; only
2209 * do it if failed >= 2
2212 for (other = disks; other--; ) {
2215 if (!test_bit(R5_UPTODATE,
2216 &sh->dev[other].flags))
2220 pr_debug("Computing stripe %llu blocks %d,%d\n",
2221 (unsigned long long)sh->sector,
2223 compute_block_2(sh, i, other);
2225 } else if (test_bit(R5_Insync, &dev->flags)) {
2226 set_bit(R5_LOCKED, &dev->flags);
2227 set_bit(R5_Wantread, &dev->flags);
2229 pr_debug("Reading block %d (sync=%d)\n",
2234 set_bit(STRIPE_HANDLE, &sh->state);
2238 /* handle_stripe_clean_event
2239 * any written block on an uptodate or failed drive can be returned.
2240 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2241 * never LOCKED, so we don't need to test 'failed' directly.
2243 static void handle_stripe_clean_event(raid5_conf_t *conf,
2244 struct stripe_head *sh, int disks, struct bio **return_bi)
2249 for (i = disks; i--; )
2250 if (sh->dev[i].written) {
2252 if (!test_bit(R5_LOCKED, &dev->flags) &&
2253 test_bit(R5_UPTODATE, &dev->flags)) {
2254 /* We can return any write requests */
2255 struct bio *wbi, *wbi2;
2257 pr_debug("Return write for disc %d\n", i);
2258 spin_lock_irq(&conf->device_lock);
2260 dev->written = NULL;
2261 while (wbi && wbi->bi_sector <
2262 dev->sector + STRIPE_SECTORS) {
2263 wbi2 = r5_next_bio(wbi, dev->sector);
2264 if (!raid5_dec_bi_phys_segments(wbi)) {
2265 md_write_end(conf->mddev);
2266 wbi->bi_next = *return_bi;
2271 if (dev->towrite == NULL)
2273 spin_unlock_irq(&conf->device_lock);
2275 bitmap_endwrite(conf->mddev->bitmap,
2278 !test_bit(STRIPE_DEGRADED, &sh->state),
2283 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2284 if (atomic_dec_and_test(&conf->pending_full_writes))
2285 md_wakeup_thread(conf->mddev->thread);
2288 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2289 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2291 int rmw = 0, rcw = 0, i;
2292 for (i = disks; i--; ) {
2293 /* would I have to read this buffer for read_modify_write */
2294 struct r5dev *dev = &sh->dev[i];
2295 if ((dev->towrite || i == sh->pd_idx) &&
2296 !test_bit(R5_LOCKED, &dev->flags) &&
2297 !(test_bit(R5_UPTODATE, &dev->flags) ||
2298 test_bit(R5_Wantcompute, &dev->flags))) {
2299 if (test_bit(R5_Insync, &dev->flags))
2302 rmw += 2*disks; /* cannot read it */
2304 /* Would I have to read this buffer for reconstruct_write */
2305 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2306 !test_bit(R5_LOCKED, &dev->flags) &&
2307 !(test_bit(R5_UPTODATE, &dev->flags) ||
2308 test_bit(R5_Wantcompute, &dev->flags))) {
2309 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2314 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2315 (unsigned long long)sh->sector, rmw, rcw);
2316 set_bit(STRIPE_HANDLE, &sh->state);
2317 if (rmw < rcw && rmw > 0)
2318 /* prefer read-modify-write, but need to get some data */
2319 for (i = disks; i--; ) {
2320 struct r5dev *dev = &sh->dev[i];
2321 if ((dev->towrite || i == sh->pd_idx) &&
2322 !test_bit(R5_LOCKED, &dev->flags) &&
2323 !(test_bit(R5_UPTODATE, &dev->flags) ||
2324 test_bit(R5_Wantcompute, &dev->flags)) &&
2325 test_bit(R5_Insync, &dev->flags)) {
2327 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2328 pr_debug("Read_old block "
2329 "%d for r-m-w\n", i);
2330 set_bit(R5_LOCKED, &dev->flags);
2331 set_bit(R5_Wantread, &dev->flags);
2334 set_bit(STRIPE_DELAYED, &sh->state);
2335 set_bit(STRIPE_HANDLE, &sh->state);
2339 if (rcw <= rmw && rcw > 0)
2340 /* want reconstruct write, but need to get some data */
2341 for (i = disks; i--; ) {
2342 struct r5dev *dev = &sh->dev[i];
2343 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2345 !test_bit(R5_LOCKED, &dev->flags) &&
2346 !(test_bit(R5_UPTODATE, &dev->flags) ||
2347 test_bit(R5_Wantcompute, &dev->flags)) &&
2348 test_bit(R5_Insync, &dev->flags)) {
2350 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2351 pr_debug("Read_old block "
2352 "%d for Reconstruct\n", i);
2353 set_bit(R5_LOCKED, &dev->flags);
2354 set_bit(R5_Wantread, &dev->flags);
2357 set_bit(STRIPE_DELAYED, &sh->state);
2358 set_bit(STRIPE_HANDLE, &sh->state);
2362 /* now if nothing is locked, and if we have enough data,
2363 * we can start a write request
2365 /* since handle_stripe can be called at any time we need to handle the
2366 * case where a compute block operation has been submitted and then a
2367 * subsequent call wants to start a write request. raid5_run_ops only
2368 * handles the case where compute block and postxor are requested
2369 * simultaneously. If this is not the case then new writes need to be
2370 * held off until the compute completes.
2372 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2373 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2374 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2375 schedule_reconstruction5(sh, s, rcw == 0, 0);
2378 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2379 struct stripe_head *sh, struct stripe_head_state *s,
2380 struct r6_state *r6s, int disks)
2382 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2383 int qd_idx = sh->qd_idx;
2384 for (i = disks; i--; ) {
2385 struct r5dev *dev = &sh->dev[i];
2386 /* Would I have to read this buffer for reconstruct_write */
2387 if (!test_bit(R5_OVERWRITE, &dev->flags)
2388 && i != pd_idx && i != qd_idx
2389 && (!test_bit(R5_LOCKED, &dev->flags)
2391 !test_bit(R5_UPTODATE, &dev->flags)) {
2392 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2394 pr_debug("raid6: must_compute: "
2395 "disk %d flags=%#lx\n", i, dev->flags);
2400 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2401 (unsigned long long)sh->sector, rcw, must_compute);
2402 set_bit(STRIPE_HANDLE, &sh->state);
2405 /* want reconstruct write, but need to get some data */
2406 for (i = disks; i--; ) {
2407 struct r5dev *dev = &sh->dev[i];
2408 if (!test_bit(R5_OVERWRITE, &dev->flags)
2409 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2410 && !test_bit(R5_LOCKED, &dev->flags) &&
2411 !test_bit(R5_UPTODATE, &dev->flags) &&
2412 test_bit(R5_Insync, &dev->flags)) {
2414 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2415 pr_debug("Read_old stripe %llu "
2416 "block %d for Reconstruct\n",
2417 (unsigned long long)sh->sector, i);
2418 set_bit(R5_LOCKED, &dev->flags);
2419 set_bit(R5_Wantread, &dev->flags);
2422 pr_debug("Request delayed stripe %llu "
2423 "block %d for Reconstruct\n",
2424 (unsigned long long)sh->sector, i);
2425 set_bit(STRIPE_DELAYED, &sh->state);
2426 set_bit(STRIPE_HANDLE, &sh->state);
2430 /* now if nothing is locked, and if we have enough data, we can start a
2433 if (s->locked == 0 && rcw == 0 &&
2434 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2435 if (must_compute > 0) {
2436 /* We have failed blocks and need to compute them */
2437 switch (s->failed) {
2441 compute_block_1(sh, r6s->failed_num[0], 0);
2444 compute_block_2(sh, r6s->failed_num[0],
2445 r6s->failed_num[1]);
2447 default: /* This request should have been failed? */
2452 pr_debug("Computing parity for stripe %llu\n",
2453 (unsigned long long)sh->sector);
2454 compute_parity6(sh, RECONSTRUCT_WRITE);
2455 /* now every locked buffer is ready to be written */
2456 for (i = disks; i--; )
2457 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2458 pr_debug("Writing stripe %llu block %d\n",
2459 (unsigned long long)sh->sector, i);
2461 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2463 if (s->locked == disks)
2464 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2465 atomic_inc(&conf->pending_full_writes);
2466 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2467 set_bit(STRIPE_INSYNC, &sh->state);
2469 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2470 atomic_dec(&conf->preread_active_stripes);
2471 if (atomic_read(&conf->preread_active_stripes) <
2473 md_wakeup_thread(conf->mddev->thread);
2478 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2479 struct stripe_head_state *s, int disks)
2481 struct r5dev *dev = NULL;
2483 set_bit(STRIPE_HANDLE, &sh->state);
2485 switch (sh->check_state) {
2486 case check_state_idle:
2487 /* start a new check operation if there are no failures */
2488 if (s->failed == 0) {
2489 BUG_ON(s->uptodate != disks);
2490 sh->check_state = check_state_run;
2491 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2492 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2496 dev = &sh->dev[s->failed_num];
2498 case check_state_compute_result:
2499 sh->check_state = check_state_idle;
2501 dev = &sh->dev[sh->pd_idx];
2503 /* check that a write has not made the stripe insync */
2504 if (test_bit(STRIPE_INSYNC, &sh->state))
2507 /* either failed parity check, or recovery is happening */
2508 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2509 BUG_ON(s->uptodate != disks);
2511 set_bit(R5_LOCKED, &dev->flags);
2513 set_bit(R5_Wantwrite, &dev->flags);
2515 clear_bit(STRIPE_DEGRADED, &sh->state);
2516 set_bit(STRIPE_INSYNC, &sh->state);
2518 case check_state_run:
2519 break; /* we will be called again upon completion */
2520 case check_state_check_result:
2521 sh->check_state = check_state_idle;
2523 /* if a failure occurred during the check operation, leave
2524 * STRIPE_INSYNC not set and let the stripe be handled again
2529 /* handle a successful check operation, if parity is correct
2530 * we are done. Otherwise update the mismatch count and repair
2531 * parity if !MD_RECOVERY_CHECK
2533 if (sh->ops.zero_sum_result == 0)
2534 /* parity is correct (on disc,
2535 * not in buffer any more)
2537 set_bit(STRIPE_INSYNC, &sh->state);
2539 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2540 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2541 /* don't try to repair!! */
2542 set_bit(STRIPE_INSYNC, &sh->state);
2544 sh->check_state = check_state_compute_run;
2545 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2546 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2547 set_bit(R5_Wantcompute,
2548 &sh->dev[sh->pd_idx].flags);
2549 sh->ops.target = sh->pd_idx;
2554 case check_state_compute_run:
2557 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2558 __func__, sh->check_state,
2559 (unsigned long long) sh->sector);
2565 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2566 struct stripe_head_state *s,
2567 struct r6_state *r6s, struct page *tmp_page,
2570 int update_p = 0, update_q = 0;
2572 int pd_idx = sh->pd_idx;
2573 int qd_idx = sh->qd_idx;
2575 set_bit(STRIPE_HANDLE, &sh->state);
2577 BUG_ON(s->failed > 2);
2578 BUG_ON(s->uptodate < disks);
2579 /* Want to check and possibly repair P and Q.
2580 * However there could be one 'failed' device, in which
2581 * case we can only check one of them, possibly using the
2582 * other to generate missing data
2585 /* If !tmp_page, we cannot do the calculations,
2586 * but as we have set STRIPE_HANDLE, we will soon be called
2587 * by stripe_handle with a tmp_page - just wait until then.
2590 if (s->failed == r6s->q_failed) {
2591 /* The only possible failed device holds 'Q', so it
2592 * makes sense to check P (If anything else were failed,
2593 * we would have used P to recreate it).
2595 compute_block_1(sh, pd_idx, 1);
2596 if (!page_is_zero(sh->dev[pd_idx].page)) {
2597 compute_block_1(sh, pd_idx, 0);
2601 if (!r6s->q_failed && s->failed < 2) {
2602 /* q is not failed, and we didn't use it to generate
2603 * anything, so it makes sense to check it
2605 memcpy(page_address(tmp_page),
2606 page_address(sh->dev[qd_idx].page),
2608 compute_parity6(sh, UPDATE_PARITY);
2609 if (memcmp(page_address(tmp_page),
2610 page_address(sh->dev[qd_idx].page),
2611 STRIPE_SIZE) != 0) {
2612 clear_bit(STRIPE_INSYNC, &sh->state);
2616 if (update_p || update_q) {
2617 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2618 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2619 /* don't try to repair!! */
2620 update_p = update_q = 0;
2623 /* now write out any block on a failed drive,
2624 * or P or Q if they need it
2627 if (s->failed == 2) {
2628 dev = &sh->dev[r6s->failed_num[1]];
2630 set_bit(R5_LOCKED, &dev->flags);
2631 set_bit(R5_Wantwrite, &dev->flags);
2633 if (s->failed >= 1) {
2634 dev = &sh->dev[r6s->failed_num[0]];
2636 set_bit(R5_LOCKED, &dev->flags);
2637 set_bit(R5_Wantwrite, &dev->flags);
2641 dev = &sh->dev[pd_idx];
2643 set_bit(R5_LOCKED, &dev->flags);
2644 set_bit(R5_Wantwrite, &dev->flags);
2647 dev = &sh->dev[qd_idx];
2649 set_bit(R5_LOCKED, &dev->flags);
2650 set_bit(R5_Wantwrite, &dev->flags);
2652 clear_bit(STRIPE_DEGRADED, &sh->state);
2654 set_bit(STRIPE_INSYNC, &sh->state);
2658 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2659 struct r6_state *r6s)
2663 /* We have read all the blocks in this stripe and now we need to
2664 * copy some of them into a target stripe for expand.
2666 struct dma_async_tx_descriptor *tx = NULL;
2667 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2668 for (i = 0; i < sh->disks; i++)
2669 if (i != sh->pd_idx && i != sh->qd_idx) {
2671 struct stripe_head *sh2;
2673 sector_t bn = compute_blocknr(sh, i, 1);
2674 sector_t s = raid5_compute_sector(conf, bn, 0,
2676 sh2 = get_active_stripe(conf, s, 0, 1);
2678 /* so far only the early blocks of this stripe
2679 * have been requested. When later blocks
2680 * get requested, we will try again
2683 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2684 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2685 /* must have already done this block */
2686 release_stripe(sh2);
2690 /* place all the copies on one channel */
2691 tx = async_memcpy(sh2->dev[dd_idx].page,
2692 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2693 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2695 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2696 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2697 for (j = 0; j < conf->raid_disks; j++)
2698 if (j != sh2->pd_idx &&
2699 (!r6s || j != sh2->qd_idx) &&
2700 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2702 if (j == conf->raid_disks) {
2703 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2704 set_bit(STRIPE_HANDLE, &sh2->state);
2706 release_stripe(sh2);
2709 /* done submitting copies, wait for them to complete */
2712 dma_wait_for_async_tx(tx);
2718 * handle_stripe - do things to a stripe.
2720 * We lock the stripe and then examine the state of various bits
2721 * to see what needs to be done.
2723 * return some read request which now have data
2724 * return some write requests which are safely on disc
2725 * schedule a read on some buffers
2726 * schedule a write of some buffers
2727 * return confirmation of parity correctness
2729 * buffers are taken off read_list or write_list, and bh_cache buffers
2730 * get BH_Lock set before the stripe lock is released.
2734 static bool handle_stripe5(struct stripe_head *sh)
2736 raid5_conf_t *conf = sh->raid_conf;
2737 int disks = sh->disks, i;
2738 struct bio *return_bi = NULL;
2739 struct stripe_head_state s;
2741 mdk_rdev_t *blocked_rdev = NULL;
2744 memset(&s, 0, sizeof(s));
2745 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2746 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2747 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2748 sh->reconstruct_state);
2750 spin_lock(&sh->lock);
2751 clear_bit(STRIPE_HANDLE, &sh->state);
2752 clear_bit(STRIPE_DELAYED, &sh->state);
2754 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2755 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2756 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2758 /* Now to look around and see what can be done */
2760 for (i=disks; i--; ) {
2762 struct r5dev *dev = &sh->dev[i];
2763 clear_bit(R5_Insync, &dev->flags);
2765 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2766 "written %p\n", i, dev->flags, dev->toread, dev->read,
2767 dev->towrite, dev->written);
2769 /* maybe we can request a biofill operation
2771 * new wantfill requests are only permitted while
2772 * ops_complete_biofill is guaranteed to be inactive
2774 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2775 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2776 set_bit(R5_Wantfill, &dev->flags);
2778 /* now count some things */
2779 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2780 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2781 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2783 if (test_bit(R5_Wantfill, &dev->flags))
2785 else if (dev->toread)
2789 if (!test_bit(R5_OVERWRITE, &dev->flags))
2794 rdev = rcu_dereference(conf->disks[i].rdev);
2795 if (blocked_rdev == NULL &&
2796 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2797 blocked_rdev = rdev;
2798 atomic_inc(&rdev->nr_pending);
2800 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2801 /* The ReadError flag will just be confusing now */
2802 clear_bit(R5_ReadError, &dev->flags);
2803 clear_bit(R5_ReWrite, &dev->flags);
2805 if (!rdev || !test_bit(In_sync, &rdev->flags)
2806 || test_bit(R5_ReadError, &dev->flags)) {
2810 set_bit(R5_Insync, &dev->flags);
2814 if (unlikely(blocked_rdev)) {
2815 if (s.syncing || s.expanding || s.expanded ||
2816 s.to_write || s.written) {
2817 set_bit(STRIPE_HANDLE, &sh->state);
2820 /* There is nothing for the blocked_rdev to block */
2821 rdev_dec_pending(blocked_rdev, conf->mddev);
2822 blocked_rdev = NULL;
2825 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2826 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2827 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2830 pr_debug("locked=%d uptodate=%d to_read=%d"
2831 " to_write=%d failed=%d failed_num=%d\n",
2832 s.locked, s.uptodate, s.to_read, s.to_write,
2833 s.failed, s.failed_num);
2834 /* check if the array has lost two devices and, if so, some requests might
2837 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2838 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2839 if (s.failed > 1 && s.syncing) {
2840 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2841 clear_bit(STRIPE_SYNCING, &sh->state);
2845 /* might be able to return some write requests if the parity block
2846 * is safe, or on a failed drive
2848 dev = &sh->dev[sh->pd_idx];
2850 ((test_bit(R5_Insync, &dev->flags) &&
2851 !test_bit(R5_LOCKED, &dev->flags) &&
2852 test_bit(R5_UPTODATE, &dev->flags)) ||
2853 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2854 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2856 /* Now we might consider reading some blocks, either to check/generate
2857 * parity, or to satisfy requests
2858 * or to load a block that is being partially written.
2860 if (s.to_read || s.non_overwrite ||
2861 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2862 handle_stripe_fill5(sh, &s, disks);
2864 /* Now we check to see if any write operations have recently
2868 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
2870 if (sh->reconstruct_state == reconstruct_state_drain_result ||
2871 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
2872 sh->reconstruct_state = reconstruct_state_idle;
2874 /* All the 'written' buffers and the parity block are ready to
2875 * be written back to disk
2877 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2878 for (i = disks; i--; ) {
2880 if (test_bit(R5_LOCKED, &dev->flags) &&
2881 (i == sh->pd_idx || dev->written)) {
2882 pr_debug("Writing block %d\n", i);
2883 set_bit(R5_Wantwrite, &dev->flags);
2886 if (!test_bit(R5_Insync, &dev->flags) ||
2887 (i == sh->pd_idx && s.failed == 0))
2888 set_bit(STRIPE_INSYNC, &sh->state);
2891 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2892 atomic_dec(&conf->preread_active_stripes);
2893 if (atomic_read(&conf->preread_active_stripes) <
2895 md_wakeup_thread(conf->mddev->thread);
2899 /* Now to consider new write requests and what else, if anything
2900 * should be read. We do not handle new writes when:
2901 * 1/ A 'write' operation (copy+xor) is already in flight.
2902 * 2/ A 'check' operation is in flight, as it may clobber the parity
2905 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
2906 handle_stripe_dirtying5(conf, sh, &s, disks);
2908 /* maybe we need to check and possibly fix the parity for this stripe
2909 * Any reads will already have been scheduled, so we just see if enough
2910 * data is available. The parity check is held off while parity
2911 * dependent operations are in flight.
2913 if (sh->check_state ||
2914 (s.syncing && s.locked == 0 &&
2915 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2916 !test_bit(STRIPE_INSYNC, &sh->state)))
2917 handle_parity_checks5(conf, sh, &s, disks);
2919 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2920 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2921 clear_bit(STRIPE_SYNCING, &sh->state);
2924 /* If the failed drive is just a ReadError, then we might need to progress
2925 * the repair/check process
2927 if (s.failed == 1 && !conf->mddev->ro &&
2928 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2929 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2930 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2932 dev = &sh->dev[s.failed_num];
2933 if (!test_bit(R5_ReWrite, &dev->flags)) {
2934 set_bit(R5_Wantwrite, &dev->flags);
2935 set_bit(R5_ReWrite, &dev->flags);
2936 set_bit(R5_LOCKED, &dev->flags);
2939 /* let's read it back */
2940 set_bit(R5_Wantread, &dev->flags);
2941 set_bit(R5_LOCKED, &dev->flags);
2946 /* Finish reconstruct operations initiated by the expansion process */
2947 if (sh->reconstruct_state == reconstruct_state_result) {
2948 struct stripe_head *sh2
2949 = get_active_stripe(conf, sh->sector, 1, 1);
2950 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
2951 /* sh cannot be written until sh2 has been read.
2952 * so arrange for sh to be delayed a little
2954 set_bit(STRIPE_DELAYED, &sh->state);
2955 set_bit(STRIPE_HANDLE, &sh->state);
2956 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
2958 atomic_inc(&conf->preread_active_stripes);
2959 release_stripe(sh2);
2963 release_stripe(sh2);
2965 sh->reconstruct_state = reconstruct_state_idle;
2966 clear_bit(STRIPE_EXPANDING, &sh->state);
2967 for (i = conf->raid_disks; i--; ) {
2968 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2969 set_bit(R5_LOCKED, &sh->dev[i].flags);
2974 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2975 !sh->reconstruct_state) {
2976 /* Need to write out all blocks after computing parity */
2977 sh->disks = conf->raid_disks;
2978 stripe_set_idx(sh->sector, conf, 0, sh);
2979 schedule_reconstruction5(sh, &s, 1, 1);
2980 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
2981 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2982 atomic_dec(&conf->reshape_stripes);
2983 wake_up(&conf->wait_for_overlap);
2984 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2987 if (s.expanding && s.locked == 0 &&
2988 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2989 handle_stripe_expansion(conf, sh, NULL);
2992 spin_unlock(&sh->lock);
2994 /* wait for this device to become unblocked */
2995 if (unlikely(blocked_rdev))
2996 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2999 raid5_run_ops(sh, s.ops_request);
3003 return_io(return_bi);
3005 return blocked_rdev == NULL;
3008 static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
3010 raid5_conf_t *conf = sh->raid_conf;
3011 int disks = sh->disks;
3012 struct bio *return_bi = NULL;
3013 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3014 struct stripe_head_state s;
3015 struct r6_state r6s;
3016 struct r5dev *dev, *pdev, *qdev;
3017 mdk_rdev_t *blocked_rdev = NULL;
3019 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3020 "pd_idx=%d, qd_idx=%d\n",
3021 (unsigned long long)sh->sector, sh->state,
3022 atomic_read(&sh->count), pd_idx, qd_idx);
3023 memset(&s, 0, sizeof(s));
3025 spin_lock(&sh->lock);
3026 clear_bit(STRIPE_HANDLE, &sh->state);
3027 clear_bit(STRIPE_DELAYED, &sh->state);
3029 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3030 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3031 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3032 /* Now to look around and see what can be done */
3035 for (i=disks; i--; ) {
3038 clear_bit(R5_Insync, &dev->flags);
3040 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3041 i, dev->flags, dev->toread, dev->towrite, dev->written);
3042 /* maybe we can reply to a read */
3043 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3044 struct bio *rbi, *rbi2;
3045 pr_debug("Return read for disc %d\n", i);
3046 spin_lock_irq(&conf->device_lock);
3049 if (test_and_clear_bit(R5_Overlap, &dev->flags))
3050 wake_up(&conf->wait_for_overlap);
3051 spin_unlock_irq(&conf->device_lock);
3052 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3053 copy_data(0, rbi, dev->page, dev->sector);
3054 rbi2 = r5_next_bio(rbi, dev->sector);
3055 spin_lock_irq(&conf->device_lock);
3056 if (!raid5_dec_bi_phys_segments(rbi)) {
3057 rbi->bi_next = return_bi;
3060 spin_unlock_irq(&conf->device_lock);
3065 /* now count some things */
3066 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3067 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3074 if (!test_bit(R5_OVERWRITE, &dev->flags))
3079 rdev = rcu_dereference(conf->disks[i].rdev);
3080 if (blocked_rdev == NULL &&
3081 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3082 blocked_rdev = rdev;
3083 atomic_inc(&rdev->nr_pending);
3085 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3086 /* The ReadError flag will just be confusing now */
3087 clear_bit(R5_ReadError, &dev->flags);
3088 clear_bit(R5_ReWrite, &dev->flags);
3090 if (!rdev || !test_bit(In_sync, &rdev->flags)
3091 || test_bit(R5_ReadError, &dev->flags)) {
3093 r6s.failed_num[s.failed] = i;
3096 set_bit(R5_Insync, &dev->flags);
3100 if (unlikely(blocked_rdev)) {
3101 if (s.syncing || s.expanding || s.expanded ||
3102 s.to_write || s.written) {
3103 set_bit(STRIPE_HANDLE, &sh->state);
3106 /* There is nothing for the blocked_rdev to block */
3107 rdev_dec_pending(blocked_rdev, conf->mddev);
3108 blocked_rdev = NULL;
3111 pr_debug("locked=%d uptodate=%d to_read=%d"
3112 " to_write=%d failed=%d failed_num=%d,%d\n",
3113 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3114 r6s.failed_num[0], r6s.failed_num[1]);
3115 /* check if the array has lost >2 devices and, if so, some requests
3116 * might need to be failed
3118 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3119 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3120 if (s.failed > 2 && s.syncing) {
3121 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3122 clear_bit(STRIPE_SYNCING, &sh->state);
3127 * might be able to return some write requests if the parity blocks
3128 * are safe, or on a failed drive
3130 pdev = &sh->dev[pd_idx];
3131 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3132 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3133 qdev = &sh->dev[qd_idx];
3134 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3135 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3138 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3139 && !test_bit(R5_LOCKED, &pdev->flags)
3140 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3141 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3142 && !test_bit(R5_LOCKED, &qdev->flags)
3143 && test_bit(R5_UPTODATE, &qdev->flags)))))
3144 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3146 /* Now we might consider reading some blocks, either to check/generate
3147 * parity, or to satisfy requests
3148 * or to load a block that is being partially written.
3150 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3151 (s.syncing && (s.uptodate < disks)) || s.expanding)
3152 handle_stripe_fill6(sh, &s, &r6s, disks);
3154 /* now to consider writing and what else, if anything should be read */
3156 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3158 /* maybe we need to check and possibly fix the parity for this stripe
3159 * Any reads will already have been scheduled, so we just see if enough
3162 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3163 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3165 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3166 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3167 clear_bit(STRIPE_SYNCING, &sh->state);
3170 /* If the failed drives are just a ReadError, then we might need
3171 * to progress the repair/check process
3173 if (s.failed <= 2 && !conf->mddev->ro)
3174 for (i = 0; i < s.failed; i++) {
3175 dev = &sh->dev[r6s.failed_num[i]];
3176 if (test_bit(R5_ReadError, &dev->flags)
3177 && !test_bit(R5_LOCKED, &dev->flags)
3178 && test_bit(R5_UPTODATE, &dev->flags)
3180 if (!test_bit(R5_ReWrite, &dev->flags)) {
3181 set_bit(R5_Wantwrite, &dev->flags);
3182 set_bit(R5_ReWrite, &dev->flags);
3183 set_bit(R5_LOCKED, &dev->flags);
3185 /* let's read it back */
3186 set_bit(R5_Wantread, &dev->flags);
3187 set_bit(R5_LOCKED, &dev->flags);
3192 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3193 struct stripe_head *sh2
3194 = get_active_stripe(conf, sh->sector, 1, 1);
3195 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3196 /* sh cannot be written until sh2 has been read.
3197 * so arrange for sh to be delayed a little
3199 set_bit(STRIPE_DELAYED, &sh->state);
3200 set_bit(STRIPE_HANDLE, &sh->state);
3201 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3203 atomic_inc(&conf->preread_active_stripes);
3204 release_stripe(sh2);
3208 release_stripe(sh2);
3210 /* Need to write out all blocks after computing P&Q */
3211 sh->disks = conf->raid_disks;
3212 stripe_set_idx(sh->sector, conf, 0, sh);
3213 compute_parity6(sh, RECONSTRUCT_WRITE);
3214 for (i = conf->raid_disks ; i-- ; ) {
3215 set_bit(R5_LOCKED, &sh->dev[i].flags);
3217 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3219 clear_bit(STRIPE_EXPANDING, &sh->state);
3220 } else if (s.expanded) {
3221 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3222 atomic_dec(&conf->reshape_stripes);
3223 wake_up(&conf->wait_for_overlap);
3224 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3227 if (s.expanding && s.locked == 0 &&
3228 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3229 handle_stripe_expansion(conf, sh, &r6s);
3232 spin_unlock(&sh->lock);
3234 /* wait for this device to become unblocked */
3235 if (unlikely(blocked_rdev))
3236 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3240 return_io(return_bi);
3242 return blocked_rdev == NULL;
3245 /* returns true if the stripe was handled */
3246 static bool handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3248 if (sh->raid_conf->level == 6)
3249 return handle_stripe6(sh, tmp_page);
3251 return handle_stripe5(sh);
3256 static void raid5_activate_delayed(raid5_conf_t *conf)
3258 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3259 while (!list_empty(&conf->delayed_list)) {
3260 struct list_head *l = conf->delayed_list.next;
3261 struct stripe_head *sh;
3262 sh = list_entry(l, struct stripe_head, lru);
3264 clear_bit(STRIPE_DELAYED, &sh->state);
3265 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3266 atomic_inc(&conf->preread_active_stripes);
3267 list_add_tail(&sh->lru, &conf->hold_list);
3270 blk_plug_device(conf->mddev->queue);
3273 static void activate_bit_delay(raid5_conf_t *conf)
3275 /* device_lock is held */
3276 struct list_head head;
3277 list_add(&head, &conf->bitmap_list);
3278 list_del_init(&conf->bitmap_list);
3279 while (!list_empty(&head)) {
3280 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3281 list_del_init(&sh->lru);
3282 atomic_inc(&sh->count);
3283 __release_stripe(conf, sh);
3287 static void unplug_slaves(mddev_t *mddev)
3289 raid5_conf_t *conf = mddev_to_conf(mddev);
3293 for (i=0; i<mddev->raid_disks; i++) {
3294 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3295 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3296 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3298 atomic_inc(&rdev->nr_pending);
3301 blk_unplug(r_queue);
3303 rdev_dec_pending(rdev, mddev);
3310 static void raid5_unplug_device(struct request_queue *q)
3312 mddev_t *mddev = q->queuedata;
3313 raid5_conf_t *conf = mddev_to_conf(mddev);
3314 unsigned long flags;
3316 spin_lock_irqsave(&conf->device_lock, flags);
3318 if (blk_remove_plug(q)) {
3320 raid5_activate_delayed(conf);
3322 md_wakeup_thread(mddev->thread);
3324 spin_unlock_irqrestore(&conf->device_lock, flags);
3326 unplug_slaves(mddev);
3329 static int raid5_congested(void *data, int bits)
3331 mddev_t *mddev = data;
3332 raid5_conf_t *conf = mddev_to_conf(mddev);
3334 /* No difference between reads and writes. Just check
3335 * how busy the stripe_cache is
3337 if (conf->inactive_blocked)
3341 if (list_empty_careful(&conf->inactive_list))
3347 /* We want read requests to align with chunks where possible,
3348 * but write requests don't need to.
3350 static int raid5_mergeable_bvec(struct request_queue *q,
3351 struct bvec_merge_data *bvm,
3352 struct bio_vec *biovec)
3354 mddev_t *mddev = q->queuedata;
3355 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3357 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3358 unsigned int bio_sectors = bvm->bi_size >> 9;
3360 if ((bvm->bi_rw & 1) == WRITE)
3361 return biovec->bv_len; /* always allow writes to be mergeable */
3363 if (mddev->new_chunk < mddev->chunk_size)
3364 chunk_sectors = mddev->new_chunk >> 9;
3365 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3366 if (max < 0) max = 0;
3367 if (max <= biovec->bv_len && bio_sectors == 0)
3368 return biovec->bv_len;
3374 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3376 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3377 unsigned int chunk_sectors = mddev->chunk_size >> 9;
3378 unsigned int bio_sectors = bio->bi_size >> 9;
3380 if (mddev->new_chunk < mddev->chunk_size)
3381 chunk_sectors = mddev->new_chunk >> 9;
3382 return chunk_sectors >=
3383 ((sector & (chunk_sectors - 1)) + bio_sectors);
3387 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3388 * later sampled by raid5d.
3390 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3392 unsigned long flags;
3394 spin_lock_irqsave(&conf->device_lock, flags);
3396 bi->bi_next = conf->retry_read_aligned_list;
3397 conf->retry_read_aligned_list = bi;
3399 spin_unlock_irqrestore(&conf->device_lock, flags);
3400 md_wakeup_thread(conf->mddev->thread);
3404 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3408 bi = conf->retry_read_aligned;
3410 conf->retry_read_aligned = NULL;
3413 bi = conf->retry_read_aligned_list;
3415 conf->retry_read_aligned_list = bi->bi_next;
3418 * this sets the active strip count to 1 and the processed
3419 * strip count to zero (upper 8 bits)
3421 bi->bi_phys_segments = 1; /* biased count of active stripes */
3429 * The "raid5_align_endio" should check if the read succeeded and if it
3430 * did, call bio_endio on the original bio (having bio_put the new bio
3432 * If the read failed..
3434 static void raid5_align_endio(struct bio *bi, int error)
3436 struct bio* raid_bi = bi->bi_private;
3439 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3444 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3445 conf = mddev_to_conf(mddev);
3446 rdev = (void*)raid_bi->bi_next;
3447 raid_bi->bi_next = NULL;
3449 rdev_dec_pending(rdev, conf->mddev);
3451 if (!error && uptodate) {
3452 bio_endio(raid_bi, 0);
3453 if (atomic_dec_and_test(&conf->active_aligned_reads))
3454 wake_up(&conf->wait_for_stripe);
3459 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3461 add_bio_to_retry(raid_bi, conf);
3464 static int bio_fits_rdev(struct bio *bi)
3466 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3468 if ((bi->bi_size>>9) > q->max_sectors)
3470 blk_recount_segments(q, bi);
3471 if (bi->bi_phys_segments > q->max_phys_segments)
3474 if (q->merge_bvec_fn)
3475 /* it's too hard to apply the merge_bvec_fn at this stage,
3484 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3486 mddev_t *mddev = q->queuedata;
3487 raid5_conf_t *conf = mddev_to_conf(mddev);
3488 unsigned int dd_idx;
3489 struct bio* align_bi;
3492 if (!in_chunk_boundary(mddev, raid_bio)) {
3493 pr_debug("chunk_aligned_read : non aligned\n");
3497 * use bio_clone to make a copy of the bio
3499 align_bi = bio_clone(raid_bio, GFP_NOIO);
3503 * set bi_end_io to a new function, and set bi_private to the
3506 align_bi->bi_end_io = raid5_align_endio;
3507 align_bi->bi_private = raid_bio;
3511 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3516 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3517 if (rdev && test_bit(In_sync, &rdev->flags)) {
3518 atomic_inc(&rdev->nr_pending);
3520 raid_bio->bi_next = (void*)rdev;
3521 align_bi->bi_bdev = rdev->bdev;
3522 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3523 align_bi->bi_sector += rdev->data_offset;
3525 if (!bio_fits_rdev(align_bi)) {
3526 /* too big in some way */
3528 rdev_dec_pending(rdev, mddev);
3532 spin_lock_irq(&conf->device_lock);
3533 wait_event_lock_irq(conf->wait_for_stripe,
3535 conf->device_lock, /* nothing */);
3536 atomic_inc(&conf->active_aligned_reads);
3537 spin_unlock_irq(&conf->device_lock);
3539 generic_make_request(align_bi);
3548 /* __get_priority_stripe - get the next stripe to process
3550 * Full stripe writes are allowed to pass preread active stripes up until
3551 * the bypass_threshold is exceeded. In general the bypass_count
3552 * increments when the handle_list is handled before the hold_list; however, it
3553 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3554 * stripe with in flight i/o. The bypass_count will be reset when the
3555 * head of the hold_list has changed, i.e. the head was promoted to the
3558 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3560 struct stripe_head *sh;
3562 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3564 list_empty(&conf->handle_list) ? "empty" : "busy",
3565 list_empty(&conf->hold_list) ? "empty" : "busy",
3566 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3568 if (!list_empty(&conf->handle_list)) {
3569 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3571 if (list_empty(&conf->hold_list))
3572 conf->bypass_count = 0;
3573 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3574 if (conf->hold_list.next == conf->last_hold)
3575 conf->bypass_count++;
3577 conf->last_hold = conf->hold_list.next;
3578 conf->bypass_count -= conf->bypass_threshold;
3579 if (conf->bypass_count < 0)
3580 conf->bypass_count = 0;
3583 } else if (!list_empty(&conf->hold_list) &&
3584 ((conf->bypass_threshold &&
3585 conf->bypass_count > conf->bypass_threshold) ||
3586 atomic_read(&conf->pending_full_writes) == 0)) {
3587 sh = list_entry(conf->hold_list.next,
3589 conf->bypass_count -= conf->bypass_threshold;
3590 if (conf->bypass_count < 0)
3591 conf->bypass_count = 0;
3595 list_del_init(&sh->lru);
3596 atomic_inc(&sh->count);
3597 BUG_ON(atomic_read(&sh->count) != 1);
3601 static int make_request(struct request_queue *q, struct bio * bi)
3603 mddev_t *mddev = q->queuedata;
3604 raid5_conf_t *conf = mddev_to_conf(mddev);
3606 sector_t new_sector;
3607 sector_t logical_sector, last_sector;
3608 struct stripe_head *sh;
3609 const int rw = bio_data_dir(bi);
3612 if (unlikely(bio_barrier(bi))) {
3613 bio_endio(bi, -EOPNOTSUPP);
3617 md_write_start(mddev, bi);
3619 cpu = part_stat_lock();
3620 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3621 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3626 mddev->reshape_position == MaxSector &&
3627 chunk_aligned_read(q,bi))
3630 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3631 last_sector = bi->bi_sector + (bi->bi_size>>9);
3633 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3635 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3637 int disks, data_disks;
3642 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3643 if (likely(conf->reshape_progress == MaxSector))
3644 disks = conf->raid_disks;
3646 /* spinlock is needed as reshape_progress may be
3647 * 64bit on a 32bit platform, and so it might be
3648 * possible to see a half-updated value
3649 * Ofcourse reshape_progress could change after
3650 * the lock is dropped, so once we get a reference
3651 * to the stripe that we think it is, we will have
3654 spin_lock_irq(&conf->device_lock);
3655 disks = conf->raid_disks;
3656 if (mddev->delta_disks < 0
3657 ? logical_sector < conf->reshape_progress
3658 : logical_sector >= conf->reshape_progress) {
3659 disks = conf->previous_raid_disks;
3662 if (mddev->delta_disks < 0
3663 ? logical_sector < conf->reshape_safe
3664 : logical_sector >= conf->reshape_safe) {
3665 spin_unlock_irq(&conf->device_lock);
3670 spin_unlock_irq(&conf->device_lock);
3672 data_disks = disks - conf->max_degraded;
3674 new_sector = raid5_compute_sector(conf, logical_sector,
3677 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3678 (unsigned long long)new_sector,
3679 (unsigned long long)logical_sector);
3681 sh = get_active_stripe(conf, new_sector, previous,
3682 (bi->bi_rw&RWA_MASK));
3684 if (unlikely(conf->reshape_progress != MaxSector)) {
3685 /* expansion might have moved on while waiting for a
3686 * stripe, so we must do the range check again.
3687 * Expansion could still move past after this
3688 * test, but as we are holding a reference to
3689 * 'sh', we know that if that happens,
3690 * STRIPE_EXPANDING will get set and the expansion
3691 * won't proceed until we finish with the stripe.
3694 spin_lock_irq(&conf->device_lock);
3695 if ((mddev->delta_disks < 0
3696 ? logical_sector >= conf->reshape_progress
3697 : logical_sector < conf->reshape_progress)
3699 /* mismatch, need to try again */
3701 spin_unlock_irq(&conf->device_lock);
3707 /* FIXME what if we get a false positive because these
3708 * are being updated.
3710 if (logical_sector >= mddev->suspend_lo &&
3711 logical_sector < mddev->suspend_hi) {
3717 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3718 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3719 /* Stripe is busy expanding or
3720 * add failed due to overlap. Flush everything
3723 raid5_unplug_device(mddev->queue);
3728 finish_wait(&conf->wait_for_overlap, &w);
3729 set_bit(STRIPE_HANDLE, &sh->state);
3730 clear_bit(STRIPE_DELAYED, &sh->state);
3733 /* cannot get stripe for read-ahead, just give-up */
3734 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3735 finish_wait(&conf->wait_for_overlap, &w);
3740 spin_lock_irq(&conf->device_lock);
3741 remaining = raid5_dec_bi_phys_segments(bi);
3742 spin_unlock_irq(&conf->device_lock);
3743 if (remaining == 0) {
3746 md_write_end(mddev);
3753 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3755 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3757 /* reshaping is quite different to recovery/resync so it is
3758 * handled quite separately ... here.
3760 * On each call to sync_request, we gather one chunk worth of
3761 * destination stripes and flag them as expanding.
3762 * Then we find all the source stripes and request reads.
3763 * As the reads complete, handle_stripe will copy the data
3764 * into the destination stripe and release that stripe.
3766 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3767 struct stripe_head *sh;
3768 sector_t first_sector, last_sector;
3769 int raid_disks = conf->previous_raid_disks;
3770 int data_disks = raid_disks - conf->max_degraded;
3771 int new_data_disks = conf->raid_disks - conf->max_degraded;
3774 sector_t writepos, safepos, gap;
3775 sector_t stripe_addr;
3776 int reshape_sectors;
3777 struct list_head stripes;
3779 if (sector_nr == 0) {
3780 /* If restarting in the middle, skip the initial sectors */
3781 if (mddev->delta_disks < 0 &&
3782 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3783 sector_nr = raid5_size(mddev, 0, 0)
3784 - conf->reshape_progress;
3785 } else if (mddev->delta_disks > 0 &&
3786 conf->reshape_progress > 0)
3787 sector_nr = conf->reshape_progress;
3788 sector_div(sector_nr, new_data_disks);
3795 /* We need to process a full chunk at a time.
3796 * If old and new chunk sizes differ, we need to process the
3799 if (mddev->new_chunk > mddev->chunk_size)
3800 reshape_sectors = mddev->new_chunk / 512;
3802 reshape_sectors = mddev->chunk_size / 512;
3804 /* we update the metadata when there is more than 3Meg
3805 * in the block range (that is rather arbitrary, should
3806 * probably be time based) or when the data about to be
3807 * copied would over-write the source of the data at
3808 * the front of the range.
3809 * i.e. one new_stripe along from reshape_progress new_maps
3810 * to after where reshape_safe old_maps to
3812 writepos = conf->reshape_progress;
3813 sector_div(writepos, new_data_disks);
3814 safepos = conf->reshape_safe;
3815 sector_div(safepos, data_disks);
3816 if (mddev->delta_disks < 0) {
3817 writepos -= reshape_sectors;
3818 safepos += reshape_sectors;
3819 gap = conf->reshape_safe - conf->reshape_progress;
3821 writepos += reshape_sectors;
3822 safepos -= reshape_sectors;
3823 gap = conf->reshape_progress - conf->reshape_safe;
3826 if ((mddev->delta_disks < 0
3827 ? writepos < safepos
3828 : writepos > safepos) ||
3829 gap > (new_data_disks)*3000*2 /*3Meg*/) {
3830 /* Cannot proceed until we've updated the superblock... */
3831 wait_event(conf->wait_for_overlap,
3832 atomic_read(&conf->reshape_stripes)==0);
3833 mddev->reshape_position = conf->reshape_progress;
3834 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3835 md_wakeup_thread(mddev->thread);
3836 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3837 kthread_should_stop());
3838 spin_lock_irq(&conf->device_lock);
3839 conf->reshape_safe = mddev->reshape_position;
3840 spin_unlock_irq(&conf->device_lock);
3841 wake_up(&conf->wait_for_overlap);
3844 if (mddev->delta_disks < 0) {
3845 BUG_ON(conf->reshape_progress == 0);
3846 stripe_addr = writepos;
3847 BUG_ON((mddev->dev_sectors &
3848 ~((sector_t)reshape_sectors - 1))
3849 - reshape_sectors - stripe_addr
3852 BUG_ON(writepos != sector_nr + reshape_sectors);
3853 stripe_addr = sector_nr;
3855 INIT_LIST_HEAD(&stripes);
3856 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3859 sh = get_active_stripe(conf, stripe_addr+i, 0, 0);
3860 set_bit(STRIPE_EXPANDING, &sh->state);
3861 atomic_inc(&conf->reshape_stripes);
3862 /* If any of this stripe is beyond the end of the old
3863 * array, then we need to zero those blocks
3865 for (j=sh->disks; j--;) {
3867 if (j == sh->pd_idx)
3869 if (conf->level == 6 &&
3872 s = compute_blocknr(sh, j, 0);
3873 if (s < raid5_size(mddev, 0, 0)) {
3877 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3878 set_bit(R5_Expanded, &sh->dev[j].flags);
3879 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3882 set_bit(STRIPE_EXPAND_READY, &sh->state);
3883 set_bit(STRIPE_HANDLE, &sh->state);
3885 list_add(&sh->lru, &stripes);
3887 spin_lock_irq(&conf->device_lock);
3888 if (mddev->delta_disks < 0)
3889 conf->reshape_progress -= reshape_sectors * new_data_disks;
3891 conf->reshape_progress += reshape_sectors * new_data_disks;
3892 spin_unlock_irq(&conf->device_lock);
3893 /* Ok, those stripe are ready. We can start scheduling
3894 * reads on the source stripes.
3895 * The source stripes are determined by mapping the first and last
3896 * block on the destination stripes.
3899 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
3902 raid5_compute_sector(conf, ((stripe_addr+conf->chunk_size/512)
3903 *(new_data_disks) - 1),
3905 if (last_sector >= mddev->dev_sectors)
3906 last_sector = mddev->dev_sectors - 1;
3907 while (first_sector <= last_sector) {
3908 sh = get_active_stripe(conf, first_sector, 1, 0);
3909 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3910 set_bit(STRIPE_HANDLE, &sh->state);
3912 first_sector += STRIPE_SECTORS;
3914 /* Now that the sources are clearly marked, we can release
3915 * the destination stripes
3917 while (!list_empty(&stripes)) {
3918 sh = list_entry(stripes.next, struct stripe_head, lru);
3919 list_del_init(&sh->lru);
3922 /* If this takes us to the resync_max point where we have to pause,
3923 * then we need to write out the superblock.
3925 sector_nr += reshape_sectors;
3926 if (sector_nr >= mddev->resync_max) {
3927 /* Cannot proceed until we've updated the superblock... */
3928 wait_event(conf->wait_for_overlap,
3929 atomic_read(&conf->reshape_stripes) == 0);
3930 mddev->reshape_position = conf->reshape_progress;
3931 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3932 md_wakeup_thread(mddev->thread);
3933 wait_event(mddev->sb_wait,
3934 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3935 || kthread_should_stop());
3936 spin_lock_irq(&conf->device_lock);
3937 conf->reshape_safe = mddev->reshape_position;
3938 spin_unlock_irq(&conf->device_lock);
3939 wake_up(&conf->wait_for_overlap);
3941 return reshape_sectors;
3944 /* FIXME go_faster isn't used */
3945 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3947 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3948 struct stripe_head *sh;
3949 sector_t max_sector = mddev->dev_sectors;
3951 int still_degraded = 0;
3954 if (sector_nr >= max_sector) {
3955 /* just being told to finish up .. nothing much to do */
3956 unplug_slaves(mddev);
3958 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3963 if (mddev->curr_resync < max_sector) /* aborted */
3964 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3966 else /* completed sync */
3968 bitmap_close_sync(mddev->bitmap);
3973 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3974 return reshape_request(mddev, sector_nr, skipped);
3976 /* No need to check resync_max as we never do more than one
3977 * stripe, and as resync_max will always be on a chunk boundary,
3978 * if the check in md_do_sync didn't fire, there is no chance
3979 * of overstepping resync_max here
3982 /* if there is too many failed drives and we are trying
3983 * to resync, then assert that we are finished, because there is
3984 * nothing we can do.
3986 if (mddev->degraded >= conf->max_degraded &&
3987 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3988 sector_t rv = mddev->dev_sectors - sector_nr;
3992 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3993 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3994 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3995 /* we can skip this block, and probably more */
3996 sync_blocks /= STRIPE_SECTORS;
3998 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4002 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4004 sh = get_active_stripe(conf, sector_nr, 0, 1);
4006 sh = get_active_stripe(conf, sector_nr, 0, 0);
4007 /* make sure we don't swamp the stripe cache if someone else
4008 * is trying to get access
4010 schedule_timeout_uninterruptible(1);
4012 /* Need to check if array will still be degraded after recovery/resync
4013 * We don't need to check the 'failed' flag as when that gets set,
4016 for (i=0; i<mddev->raid_disks; i++)
4017 if (conf->disks[i].rdev == NULL)
4020 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4022 spin_lock(&sh->lock);
4023 set_bit(STRIPE_SYNCING, &sh->state);
4024 clear_bit(STRIPE_INSYNC, &sh->state);
4025 spin_unlock(&sh->lock);
4027 /* wait for any blocked device to be handled */
4028 while(unlikely(!handle_stripe(sh, NULL)))
4032 return STRIPE_SECTORS;
4035 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4037 /* We may not be able to submit a whole bio at once as there
4038 * may not be enough stripe_heads available.
4039 * We cannot pre-allocate enough stripe_heads as we may need
4040 * more than exist in the cache (if we allow ever large chunks).
4041 * So we do one stripe head at a time and record in
4042 * ->bi_hw_segments how many have been done.
4044 * We *know* that this entire raid_bio is in one chunk, so
4045 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4047 struct stripe_head *sh;
4049 sector_t sector, logical_sector, last_sector;
4054 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4055 sector = raid5_compute_sector(conf, logical_sector,
4057 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4059 for (; logical_sector < last_sector;
4060 logical_sector += STRIPE_SECTORS,
4061 sector += STRIPE_SECTORS,
4064 if (scnt < raid5_bi_hw_segments(raid_bio))
4065 /* already done this stripe */
4068 sh = get_active_stripe(conf, sector, 0, 1);
4071 /* failed to get a stripe - must wait */
4072 raid5_set_bi_hw_segments(raid_bio, scnt);
4073 conf->retry_read_aligned = raid_bio;
4077 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4078 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4080 raid5_set_bi_hw_segments(raid_bio, scnt);
4081 conf->retry_read_aligned = raid_bio;
4085 handle_stripe(sh, NULL);
4089 spin_lock_irq(&conf->device_lock);
4090 remaining = raid5_dec_bi_phys_segments(raid_bio);
4091 spin_unlock_irq(&conf->device_lock);
4093 bio_endio(raid_bio, 0);
4094 if (atomic_dec_and_test(&conf->active_aligned_reads))
4095 wake_up(&conf->wait_for_stripe);
4102 * This is our raid5 kernel thread.
4104 * We scan the hash table for stripes which can be handled now.
4105 * During the scan, completed stripes are saved for us by the interrupt
4106 * handler, so that they will not have to wait for our next wakeup.
4108 static void raid5d(mddev_t *mddev)
4110 struct stripe_head *sh;
4111 raid5_conf_t *conf = mddev_to_conf(mddev);
4114 pr_debug("+++ raid5d active\n");
4116 md_check_recovery(mddev);
4119 spin_lock_irq(&conf->device_lock);
4123 if (conf->seq_flush != conf->seq_write) {
4124 int seq = conf->seq_flush;
4125 spin_unlock_irq(&conf->device_lock);
4126 bitmap_unplug(mddev->bitmap);
4127 spin_lock_irq(&conf->device_lock);
4128 conf->seq_write = seq;
4129 activate_bit_delay(conf);
4132 while ((bio = remove_bio_from_retry(conf))) {
4134 spin_unlock_irq(&conf->device_lock);
4135 ok = retry_aligned_read(conf, bio);
4136 spin_lock_irq(&conf->device_lock);
4142 sh = __get_priority_stripe(conf);
4146 spin_unlock_irq(&conf->device_lock);
4149 handle_stripe(sh, conf->spare_page);
4152 spin_lock_irq(&conf->device_lock);
4154 pr_debug("%d stripes handled\n", handled);
4156 spin_unlock_irq(&conf->device_lock);
4158 async_tx_issue_pending_all();
4159 unplug_slaves(mddev);
4161 pr_debug("--- raid5d inactive\n");
4165 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4167 raid5_conf_t *conf = mddev_to_conf(mddev);
4169 return sprintf(page, "%d\n", conf->max_nr_stripes);
4175 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4177 raid5_conf_t *conf = mddev_to_conf(mddev);
4181 if (len >= PAGE_SIZE)
4186 if (strict_strtoul(page, 10, &new))
4188 if (new <= 16 || new > 32768)
4190 while (new < conf->max_nr_stripes) {
4191 if (drop_one_stripe(conf))
4192 conf->max_nr_stripes--;
4196 err = md_allow_write(mddev);
4199 while (new > conf->max_nr_stripes) {
4200 if (grow_one_stripe(conf))
4201 conf->max_nr_stripes++;
4207 static struct md_sysfs_entry
4208 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4209 raid5_show_stripe_cache_size,
4210 raid5_store_stripe_cache_size);
4213 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4215 raid5_conf_t *conf = mddev_to_conf(mddev);
4217 return sprintf(page, "%d\n", conf->bypass_threshold);
4223 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4225 raid5_conf_t *conf = mddev_to_conf(mddev);
4227 if (len >= PAGE_SIZE)
4232 if (strict_strtoul(page, 10, &new))
4234 if (new > conf->max_nr_stripes)
4236 conf->bypass_threshold = new;
4240 static struct md_sysfs_entry
4241 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4243 raid5_show_preread_threshold,
4244 raid5_store_preread_threshold);
4247 stripe_cache_active_show(mddev_t *mddev, char *page)
4249 raid5_conf_t *conf = mddev_to_conf(mddev);
4251 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4256 static struct md_sysfs_entry
4257 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4259 static struct attribute *raid5_attrs[] = {
4260 &raid5_stripecache_size.attr,
4261 &raid5_stripecache_active.attr,
4262 &raid5_preread_bypass_threshold.attr,
4265 static struct attribute_group raid5_attrs_group = {
4267 .attrs = raid5_attrs,
4271 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4273 raid5_conf_t *conf = mddev_to_conf(mddev);
4276 sectors = mddev->dev_sectors;
4278 /* size is defined by the smallest of previous and new size */
4279 if (conf->raid_disks < conf->previous_raid_disks)
4280 raid_disks = conf->raid_disks;
4282 raid_disks = conf->previous_raid_disks;
4285 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4286 sectors &= ~((sector_t)mddev->new_chunk/512 - 1);
4287 return sectors * (raid_disks - conf->max_degraded);
4290 static raid5_conf_t *setup_conf(mddev_t *mddev)
4293 int raid_disk, memory;
4295 struct disk_info *disk;
4297 if (mddev->new_level != 5
4298 && mddev->new_level != 4
4299 && mddev->new_level != 6) {
4300 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4301 mdname(mddev), mddev->new_level);
4302 return ERR_PTR(-EIO);
4304 if ((mddev->new_level == 5
4305 && !algorithm_valid_raid5(mddev->new_layout)) ||
4306 (mddev->new_level == 6
4307 && !algorithm_valid_raid6(mddev->new_layout))) {
4308 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4309 mdname(mddev), mddev->new_layout);
4310 return ERR_PTR(-EIO);
4312 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4313 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4314 mdname(mddev), mddev->raid_disks);
4315 return ERR_PTR(-EINVAL);
4318 if (!mddev->new_chunk || mddev->new_chunk % PAGE_SIZE) {
4319 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4320 mddev->new_chunk, mdname(mddev));
4321 return ERR_PTR(-EINVAL);
4324 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4328 conf->raid_disks = mddev->raid_disks;
4329 if (mddev->reshape_position == MaxSector)
4330 conf->previous_raid_disks = mddev->raid_disks;
4332 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4334 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4339 conf->mddev = mddev;
4341 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4344 if (mddev->new_level == 6) {
4345 conf->spare_page = alloc_page(GFP_KERNEL);
4346 if (!conf->spare_page)
4349 spin_lock_init(&conf->device_lock);
4350 init_waitqueue_head(&conf->wait_for_stripe);
4351 init_waitqueue_head(&conf->wait_for_overlap);
4352 INIT_LIST_HEAD(&conf->handle_list);
4353 INIT_LIST_HEAD(&conf->hold_list);
4354 INIT_LIST_HEAD(&conf->delayed_list);
4355 INIT_LIST_HEAD(&conf->bitmap_list);
4356 INIT_LIST_HEAD(&conf->inactive_list);
4357 atomic_set(&conf->active_stripes, 0);
4358 atomic_set(&conf->preread_active_stripes, 0);
4359 atomic_set(&conf->active_aligned_reads, 0);
4360 conf->bypass_threshold = BYPASS_THRESHOLD;
4362 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4364 list_for_each_entry(rdev, &mddev->disks, same_set) {
4365 raid_disk = rdev->raid_disk;
4366 if (raid_disk >= conf->raid_disks
4369 disk = conf->disks + raid_disk;
4373 if (test_bit(In_sync, &rdev->flags)) {
4374 char b[BDEVNAME_SIZE];
4375 printk(KERN_INFO "raid5: device %s operational as raid"
4376 " disk %d\n", bdevname(rdev->bdev,b),
4379 /* Cannot rely on bitmap to complete recovery */
4383 conf->chunk_size = mddev->new_chunk;
4384 conf->level = mddev->new_level;
4385 if (conf->level == 6)
4386 conf->max_degraded = 2;
4388 conf->max_degraded = 1;
4389 conf->algorithm = mddev->new_layout;
4390 conf->max_nr_stripes = NR_STRIPES;
4391 conf->reshape_progress = mddev->reshape_position;
4392 if (conf->reshape_progress != MaxSector) {
4393 conf->prev_chunk = mddev->chunk_size;
4394 conf->prev_algo = mddev->layout;
4397 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4398 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4399 if (grow_stripes(conf, conf->max_nr_stripes)) {
4401 "raid5: couldn't allocate %dkB for buffers\n", memory);
4404 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4405 memory, mdname(mddev));
4407 conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4408 if (!conf->thread) {
4410 "raid5: couldn't allocate thread for %s\n",
4419 shrink_stripes(conf);
4420 safe_put_page(conf->spare_page);
4422 kfree(conf->stripe_hashtbl);
4424 return ERR_PTR(-EIO);
4426 return ERR_PTR(-ENOMEM);
4429 static int run(mddev_t *mddev)
4432 int working_disks = 0;
4435 if (mddev->reshape_position != MaxSector) {
4436 /* Check that we can continue the reshape.
4437 * Currently only disks can change, it must
4438 * increase, and we must be past the point where
4439 * a stripe over-writes itself
4441 sector_t here_new, here_old;
4443 int max_degraded = (mddev->level == 6 ? 2 : 1);
4445 if (mddev->new_level != mddev->level) {
4446 printk(KERN_ERR "raid5: %s: unsupported reshape "
4447 "required - aborting.\n",
4451 old_disks = mddev->raid_disks - mddev->delta_disks;
4452 /* reshape_position must be on a new-stripe boundary, and one
4453 * further up in new geometry must map after here in old
4456 here_new = mddev->reshape_position;
4457 if (sector_div(here_new, (mddev->new_chunk>>9)*
4458 (mddev->raid_disks - max_degraded))) {
4459 printk(KERN_ERR "raid5: reshape_position not "
4460 "on a stripe boundary\n");
4463 /* here_new is the stripe we will write to */
4464 here_old = mddev->reshape_position;
4465 sector_div(here_old, (mddev->chunk_size>>9)*
4466 (old_disks-max_degraded));
4467 /* here_old is the first stripe that we might need to read
4469 if (here_new >= here_old) {
4470 /* Reading from the same stripe as writing to - bad */
4471 printk(KERN_ERR "raid5: reshape_position too early for "
4472 "auto-recovery - aborting.\n");
4475 printk(KERN_INFO "raid5: reshape will continue\n");
4476 /* OK, we should be able to continue; */
4478 BUG_ON(mddev->level != mddev->new_level);
4479 BUG_ON(mddev->layout != mddev->new_layout);
4480 BUG_ON(mddev->chunk_size != mddev->new_chunk);
4481 BUG_ON(mddev->delta_disks != 0);
4484 if (mddev->private == NULL)
4485 conf = setup_conf(mddev);
4487 conf = mddev->private;
4490 return PTR_ERR(conf);
4492 mddev->thread = conf->thread;
4493 conf->thread = NULL;
4494 mddev->private = conf;
4497 * 0 for a fully functional array, 1 or 2 for a degraded array.
4499 list_for_each_entry(rdev, &mddev->disks, same_set)
4500 if (rdev->raid_disk >= 0 &&
4501 test_bit(In_sync, &rdev->flags))
4504 mddev->degraded = conf->raid_disks - working_disks;
4506 if (mddev->degraded > conf->max_degraded) {
4507 printk(KERN_ERR "raid5: not enough operational devices for %s"
4508 " (%d/%d failed)\n",
4509 mdname(mddev), mddev->degraded, conf->raid_disks);
4513 /* device size must be a multiple of chunk size */
4514 mddev->dev_sectors &= ~(mddev->chunk_size / 512 - 1);
4515 mddev->resync_max_sectors = mddev->dev_sectors;
4517 if (mddev->degraded > 0 &&
4518 mddev->recovery_cp != MaxSector) {
4519 if (mddev->ok_start_degraded)
4521 "raid5: starting dirty degraded array: %s"
4522 "- data corruption possible.\n",
4526 "raid5: cannot start dirty degraded array for %s\n",
4532 if (mddev->degraded == 0)
4533 printk("raid5: raid level %d set %s active with %d out of %d"
4534 " devices, algorithm %d\n", conf->level, mdname(mddev),
4535 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4538 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4539 " out of %d devices, algorithm %d\n", conf->level,
4540 mdname(mddev), mddev->raid_disks - mddev->degraded,
4541 mddev->raid_disks, mddev->new_layout);
4543 print_raid5_conf(conf);
4545 if (conf->reshape_progress != MaxSector) {
4546 printk("...ok start reshape thread\n");
4547 conf->reshape_safe = conf->reshape_progress;
4548 atomic_set(&conf->reshape_stripes, 0);
4549 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4550 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4551 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4552 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4553 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4557 /* read-ahead size must cover two whole stripes, which is
4558 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4561 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4562 int stripe = data_disks *
4563 (mddev->chunk_size / PAGE_SIZE);
4564 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4565 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4568 /* Ok, everything is just fine now */
4569 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4571 "raid5: failed to create sysfs attributes for %s\n",
4574 mddev->queue->queue_lock = &conf->device_lock;
4576 mddev->queue->unplug_fn = raid5_unplug_device;
4577 mddev->queue->backing_dev_info.congested_data = mddev;
4578 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4580 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4582 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4586 md_unregister_thread(mddev->thread);
4587 mddev->thread = NULL;
4589 shrink_stripes(conf);
4590 print_raid5_conf(conf);
4591 safe_put_page(conf->spare_page);
4593 kfree(conf->stripe_hashtbl);
4596 mddev->private = NULL;
4597 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4603 static int stop(mddev_t *mddev)
4605 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4607 md_unregister_thread(mddev->thread);
4608 mddev->thread = NULL;
4609 shrink_stripes(conf);
4610 kfree(conf->stripe_hashtbl);
4611 mddev->queue->backing_dev_info.congested_fn = NULL;
4612 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4613 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4616 mddev->private = NULL;
4621 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4625 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4626 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4627 seq_printf(seq, "sh %llu, count %d.\n",
4628 (unsigned long long)sh->sector, atomic_read(&sh->count));
4629 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4630 for (i = 0; i < sh->disks; i++) {
4631 seq_printf(seq, "(cache%d: %p %ld) ",
4632 i, sh->dev[i].page, sh->dev[i].flags);
4634 seq_printf(seq, "\n");
4637 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4639 struct stripe_head *sh;
4640 struct hlist_node *hn;
4643 spin_lock_irq(&conf->device_lock);
4644 for (i = 0; i < NR_HASH; i++) {
4645 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4646 if (sh->raid_conf != conf)
4651 spin_unlock_irq(&conf->device_lock);
4655 static void status(struct seq_file *seq, mddev_t *mddev)
4657 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4660 seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4661 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4662 for (i = 0; i < conf->raid_disks; i++)
4663 seq_printf (seq, "%s",
4664 conf->disks[i].rdev &&
4665 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4666 seq_printf (seq, "]");
4668 seq_printf (seq, "\n");
4669 printall(seq, conf);
4673 static void print_raid5_conf (raid5_conf_t *conf)
4676 struct disk_info *tmp;
4678 printk("RAID5 conf printout:\n");
4680 printk("(conf==NULL)\n");
4683 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4684 conf->raid_disks - conf->mddev->degraded);
4686 for (i = 0; i < conf->raid_disks; i++) {
4687 char b[BDEVNAME_SIZE];
4688 tmp = conf->disks + i;
4690 printk(" disk %d, o:%d, dev:%s\n",
4691 i, !test_bit(Faulty, &tmp->rdev->flags),
4692 bdevname(tmp->rdev->bdev,b));
4696 static int raid5_spare_active(mddev_t *mddev)
4699 raid5_conf_t *conf = mddev->private;
4700 struct disk_info *tmp;
4702 for (i = 0; i < conf->raid_disks; i++) {
4703 tmp = conf->disks + i;
4705 && !test_bit(Faulty, &tmp->rdev->flags)
4706 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4707 unsigned long flags;
4708 spin_lock_irqsave(&conf->device_lock, flags);
4710 spin_unlock_irqrestore(&conf->device_lock, flags);
4713 print_raid5_conf(conf);
4717 static int raid5_remove_disk(mddev_t *mddev, int number)
4719 raid5_conf_t *conf = mddev->private;
4722 struct disk_info *p = conf->disks + number;
4724 print_raid5_conf(conf);
4727 if (number >= conf->raid_disks &&
4728 conf->reshape_progress == MaxSector)
4729 clear_bit(In_sync, &rdev->flags);
4731 if (test_bit(In_sync, &rdev->flags) ||
4732 atomic_read(&rdev->nr_pending)) {
4736 /* Only remove non-faulty devices if recovery
4739 if (!test_bit(Faulty, &rdev->flags) &&
4740 mddev->degraded <= conf->max_degraded &&
4741 number < conf->raid_disks) {
4747 if (atomic_read(&rdev->nr_pending)) {
4748 /* lost the race, try later */
4755 print_raid5_conf(conf);
4759 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4761 raid5_conf_t *conf = mddev->private;
4764 struct disk_info *p;
4766 int last = conf->raid_disks - 1;
4768 if (mddev->degraded > conf->max_degraded)
4769 /* no point adding a device */
4772 if (rdev->raid_disk >= 0)
4773 first = last = rdev->raid_disk;
4776 * find the disk ... but prefer rdev->saved_raid_disk
4779 if (rdev->saved_raid_disk >= 0 &&
4780 rdev->saved_raid_disk >= first &&
4781 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4782 disk = rdev->saved_raid_disk;
4785 for ( ; disk <= last ; disk++)
4786 if ((p=conf->disks + disk)->rdev == NULL) {
4787 clear_bit(In_sync, &rdev->flags);
4788 rdev->raid_disk = disk;
4790 if (rdev->saved_raid_disk != disk)
4792 rcu_assign_pointer(p->rdev, rdev);
4795 print_raid5_conf(conf);
4799 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4801 /* no resync is happening, and there is enough space
4802 * on all devices, so we can resize.
4803 * We need to make sure resync covers any new space.
4804 * If the array is shrinking we should possibly wait until
4805 * any io in the removed space completes, but it hardly seems
4808 sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4809 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
4810 mddev->raid_disks));
4811 if (mddev->array_sectors >
4812 raid5_size(mddev, sectors, mddev->raid_disks))
4814 set_capacity(mddev->gendisk, mddev->array_sectors);
4816 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
4817 mddev->recovery_cp = mddev->dev_sectors;
4818 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4820 mddev->dev_sectors = sectors;
4821 mddev->resync_max_sectors = sectors;
4825 #ifdef CONFIG_MD_RAID5_RESHAPE
4826 static int raid5_check_reshape(mddev_t *mddev)
4828 raid5_conf_t *conf = mddev_to_conf(mddev);
4830 if (mddev->delta_disks == 0 &&
4831 mddev->new_layout == mddev->layout &&
4832 mddev->new_chunk == mddev->chunk_size)
4833 return -EINVAL; /* nothing to do */
4835 /* Cannot grow a bitmap yet */
4837 if (mddev->degraded > conf->max_degraded)
4839 if (mddev->delta_disks < 0) {
4840 /* We might be able to shrink, but the devices must
4841 * be made bigger first.
4842 * For raid6, 4 is the minimum size.
4843 * Otherwise 2 is the minimum
4846 if (mddev->level == 6)
4848 if (mddev->raid_disks + mddev->delta_disks < min)
4852 /* Can only proceed if there are plenty of stripe_heads.
4853 * We need a minimum of one full stripe,, and for sensible progress
4854 * it is best to have about 4 times that.
4855 * If we require 4 times, then the default 256 4K stripe_heads will
4856 * allow for chunk sizes up to 256K, which is probably OK.
4857 * If the chunk size is greater, user-space should request more
4858 * stripe_heads first.
4860 if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4861 (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4862 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4863 (max(mddev->chunk_size, mddev->new_chunk)
4868 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4871 static int raid5_start_reshape(mddev_t *mddev)
4873 raid5_conf_t *conf = mddev_to_conf(mddev);
4876 int added_devices = 0;
4877 unsigned long flags;
4879 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4882 list_for_each_entry(rdev, &mddev->disks, same_set)
4883 if (rdev->raid_disk < 0 &&
4884 !test_bit(Faulty, &rdev->flags))
4887 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4888 /* Not enough devices even to make a degraded array
4893 /* Refuse to reduce size of the array. Any reductions in
4894 * array size must be through explicit setting of array_size
4897 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
4898 < mddev->array_sectors) {
4899 printk(KERN_ERR "md: %s: array size must be reduced "
4900 "before number of disks\n", mdname(mddev));
4904 atomic_set(&conf->reshape_stripes, 0);
4905 spin_lock_irq(&conf->device_lock);
4906 conf->previous_raid_disks = conf->raid_disks;
4907 conf->raid_disks += mddev->delta_disks;
4908 conf->prev_chunk = conf->chunk_size;
4909 conf->chunk_size = mddev->new_chunk;
4910 conf->prev_algo = conf->algorithm;
4911 conf->algorithm = mddev->new_layout;
4912 if (mddev->delta_disks < 0)
4913 conf->reshape_progress = raid5_size(mddev, 0, 0);
4915 conf->reshape_progress = 0;
4916 conf->reshape_safe = conf->reshape_progress;
4918 spin_unlock_irq(&conf->device_lock);
4920 /* Add some new drives, as many as will fit.
4921 * We know there are enough to make the newly sized array work.
4923 list_for_each_entry(rdev, &mddev->disks, same_set)
4924 if (rdev->raid_disk < 0 &&
4925 !test_bit(Faulty, &rdev->flags)) {
4926 if (raid5_add_disk(mddev, rdev) == 0) {
4928 set_bit(In_sync, &rdev->flags);
4930 rdev->recovery_offset = 0;
4931 sprintf(nm, "rd%d", rdev->raid_disk);
4932 if (sysfs_create_link(&mddev->kobj,
4935 "raid5: failed to create "
4936 " link %s for %s\n",
4942 if (mddev->delta_disks > 0) {
4943 spin_lock_irqsave(&conf->device_lock, flags);
4944 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks)
4946 spin_unlock_irqrestore(&conf->device_lock, flags);
4948 mddev->raid_disks = conf->raid_disks;
4949 mddev->reshape_position = 0;
4950 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4952 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4953 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4954 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4955 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4956 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4958 if (!mddev->sync_thread) {
4959 mddev->recovery = 0;
4960 spin_lock_irq(&conf->device_lock);
4961 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4962 conf->reshape_progress = MaxSector;
4963 spin_unlock_irq(&conf->device_lock);
4966 md_wakeup_thread(mddev->sync_thread);
4967 md_new_event(mddev);
4972 /* This is called from the reshape thread and should make any
4973 * changes needed in 'conf'
4975 static void end_reshape(raid5_conf_t *conf)
4978 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4980 spin_lock_irq(&conf->device_lock);
4981 conf->previous_raid_disks = conf->raid_disks;
4982 conf->reshape_progress = MaxSector;
4983 spin_unlock_irq(&conf->device_lock);
4985 /* read-ahead size must cover two whole stripes, which is
4986 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4989 int data_disks = conf->raid_disks - conf->max_degraded;
4990 int stripe = data_disks * (conf->chunk_size
4992 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4993 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4998 /* This is called from the raid5d thread with mddev_lock held.
4999 * It makes config changes to the device.
5001 static void raid5_finish_reshape(mddev_t *mddev)
5003 struct block_device *bdev;
5004 raid5_conf_t *conf = mddev_to_conf(mddev);
5006 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5008 if (mddev->delta_disks > 0) {
5009 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5010 set_capacity(mddev->gendisk, mddev->array_sectors);
5013 bdev = bdget_disk(mddev->gendisk, 0);
5015 mutex_lock(&bdev->bd_inode->i_mutex);
5016 i_size_write(bdev->bd_inode,
5017 (loff_t)mddev->array_sectors << 9);
5018 mutex_unlock(&bdev->bd_inode->i_mutex);
5023 mddev->degraded = conf->raid_disks;
5024 for (d = 0; d < conf->raid_disks ; d++)
5025 if (conf->disks[d].rdev &&
5027 &conf->disks[d].rdev->flags))
5029 for (d = conf->raid_disks ;
5030 d < conf->raid_disks - mddev->delta_disks;
5032 raid5_remove_disk(mddev, d);
5034 mddev->layout = conf->algorithm;
5035 mddev->chunk_size = conf->chunk_size;
5036 mddev->reshape_position = MaxSector;
5037 mddev->delta_disks = 0;
5041 static void raid5_quiesce(mddev_t *mddev, int state)
5043 raid5_conf_t *conf = mddev_to_conf(mddev);
5046 case 2: /* resume for a suspend */
5047 wake_up(&conf->wait_for_overlap);
5050 case 1: /* stop all writes */
5051 spin_lock_irq(&conf->device_lock);
5053 wait_event_lock_irq(conf->wait_for_stripe,
5054 atomic_read(&conf->active_stripes) == 0 &&
5055 atomic_read(&conf->active_aligned_reads) == 0,
5056 conf->device_lock, /* nothing */);
5057 spin_unlock_irq(&conf->device_lock);
5060 case 0: /* re-enable writes */
5061 spin_lock_irq(&conf->device_lock);
5063 wake_up(&conf->wait_for_stripe);
5064 wake_up(&conf->wait_for_overlap);
5065 spin_unlock_irq(&conf->device_lock);
5071 static void *raid5_takeover_raid1(mddev_t *mddev)
5075 if (mddev->raid_disks != 2 ||
5076 mddev->degraded > 1)
5077 return ERR_PTR(-EINVAL);
5079 /* Should check if there are write-behind devices? */
5081 chunksect = 64*2; /* 64K by default */
5083 /* The array must be an exact multiple of chunksize */
5084 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5087 if ((chunksect<<9) < STRIPE_SIZE)
5088 /* array size does not allow a suitable chunk size */
5089 return ERR_PTR(-EINVAL);
5091 mddev->new_level = 5;
5092 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5093 mddev->new_chunk = chunksect << 9;
5095 return setup_conf(mddev);
5098 static void *raid5_takeover_raid6(mddev_t *mddev)
5102 switch (mddev->layout) {
5103 case ALGORITHM_LEFT_ASYMMETRIC_6:
5104 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5106 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5107 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5109 case ALGORITHM_LEFT_SYMMETRIC_6:
5110 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5112 case ALGORITHM_RIGHT_SYMMETRIC_6:
5113 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5115 case ALGORITHM_PARITY_0_6:
5116 new_layout = ALGORITHM_PARITY_0;
5118 case ALGORITHM_PARITY_N:
5119 new_layout = ALGORITHM_PARITY_N;
5122 return ERR_PTR(-EINVAL);
5124 mddev->new_level = 5;
5125 mddev->new_layout = new_layout;
5126 mddev->delta_disks = -1;
5127 mddev->raid_disks -= 1;
5128 return setup_conf(mddev);
5132 static int raid5_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5134 /* For a 2-drive array, the layout and chunk size can be changed
5135 * immediately as not restriping is needed.
5136 * For larger arrays we record the new value - after validation
5137 * to be used by a reshape pass.
5139 raid5_conf_t *conf = mddev_to_conf(mddev);
5141 if (new_layout >= 0 && !algorithm_valid_raid5(new_layout))
5143 if (new_chunk > 0) {
5144 if (new_chunk & (new_chunk-1))
5145 /* not a power of 2 */
5147 if (new_chunk < PAGE_SIZE)
5149 if (mddev->array_sectors & ((new_chunk>>9)-1))
5150 /* not factor of array size */
5154 /* They look valid */
5156 if (mddev->raid_disks == 2) {
5158 if (new_layout >= 0) {
5159 conf->algorithm = new_layout;
5160 mddev->layout = mddev->new_layout = new_layout;
5162 if (new_chunk > 0) {
5163 conf->chunk_size = new_chunk;
5164 mddev->chunk_size = mddev->new_chunk = new_chunk;
5166 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5167 md_wakeup_thread(mddev->thread);
5169 if (new_layout >= 0)
5170 mddev->new_layout = new_layout;
5172 mddev->new_chunk = new_chunk;
5177 static int raid6_reconfig(mddev_t *mddev, int new_layout, int new_chunk)
5179 if (new_layout >= 0 && !algorithm_valid_raid6(new_layout))
5181 if (new_chunk > 0) {
5182 if (new_chunk & (new_chunk-1))
5183 /* not a power of 2 */
5185 if (new_chunk < PAGE_SIZE)
5187 if (mddev->array_sectors & ((new_chunk>>9)-1))
5188 /* not factor of array size */
5192 /* They look valid */
5194 if (new_layout >= 0)
5195 mddev->new_layout = new_layout;
5197 mddev->new_chunk = new_chunk;
5202 static void *raid5_takeover(mddev_t *mddev)
5204 /* raid5 can take over:
5205 * raid0 - if all devices are the same - make it a raid4 layout
5206 * raid1 - if there are two drives. We need to know the chunk size
5207 * raid4 - trivial - just use a raid4 layout.
5208 * raid6 - Providing it is a *_6 layout
5210 * For now, just do raid1
5213 if (mddev->level == 1)
5214 return raid5_takeover_raid1(mddev);
5215 if (mddev->level == 4) {
5216 mddev->new_layout = ALGORITHM_PARITY_N;
5217 mddev->new_level = 5;
5218 return setup_conf(mddev);
5220 if (mddev->level == 6)
5221 return raid5_takeover_raid6(mddev);
5223 return ERR_PTR(-EINVAL);
5227 static struct mdk_personality raid5_personality;
5229 static void *raid6_takeover(mddev_t *mddev)
5231 /* Currently can only take over a raid5. We map the
5232 * personality to an equivalent raid6 personality
5233 * with the Q block at the end.
5237 if (mddev->pers != &raid5_personality)
5238 return ERR_PTR(-EINVAL);
5239 if (mddev->degraded > 1)
5240 return ERR_PTR(-EINVAL);
5241 if (mddev->raid_disks > 253)
5242 return ERR_PTR(-EINVAL);
5243 if (mddev->raid_disks < 3)
5244 return ERR_PTR(-EINVAL);
5246 switch (mddev->layout) {
5247 case ALGORITHM_LEFT_ASYMMETRIC:
5248 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5250 case ALGORITHM_RIGHT_ASYMMETRIC:
5251 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5253 case ALGORITHM_LEFT_SYMMETRIC:
5254 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5256 case ALGORITHM_RIGHT_SYMMETRIC:
5257 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5259 case ALGORITHM_PARITY_0:
5260 new_layout = ALGORITHM_PARITY_0_6;
5262 case ALGORITHM_PARITY_N:
5263 new_layout = ALGORITHM_PARITY_N;
5266 return ERR_PTR(-EINVAL);
5268 mddev->new_level = 6;
5269 mddev->new_layout = new_layout;
5270 mddev->delta_disks = 1;
5271 mddev->raid_disks += 1;
5272 return setup_conf(mddev);
5276 static struct mdk_personality raid6_personality =
5280 .owner = THIS_MODULE,
5281 .make_request = make_request,
5285 .error_handler = error,
5286 .hot_add_disk = raid5_add_disk,
5287 .hot_remove_disk= raid5_remove_disk,
5288 .spare_active = raid5_spare_active,
5289 .sync_request = sync_request,
5290 .resize = raid5_resize,
5292 #ifdef CONFIG_MD_RAID5_RESHAPE
5293 .check_reshape = raid5_check_reshape,
5294 .start_reshape = raid5_start_reshape,
5295 .finish_reshape = raid5_finish_reshape,
5297 .quiesce = raid5_quiesce,
5298 .takeover = raid6_takeover,
5299 .reconfig = raid6_reconfig,
5301 static struct mdk_personality raid5_personality =
5305 .owner = THIS_MODULE,
5306 .make_request = make_request,
5310 .error_handler = error,
5311 .hot_add_disk = raid5_add_disk,
5312 .hot_remove_disk= raid5_remove_disk,
5313 .spare_active = raid5_spare_active,
5314 .sync_request = sync_request,
5315 .resize = raid5_resize,
5317 #ifdef CONFIG_MD_RAID5_RESHAPE
5318 .check_reshape = raid5_check_reshape,
5319 .start_reshape = raid5_start_reshape,
5320 .finish_reshape = raid5_finish_reshape,
5322 .quiesce = raid5_quiesce,
5323 .takeover = raid5_takeover,
5324 .reconfig = raid5_reconfig,
5327 static struct mdk_personality raid4_personality =
5331 .owner = THIS_MODULE,
5332 .make_request = make_request,
5336 .error_handler = error,
5337 .hot_add_disk = raid5_add_disk,
5338 .hot_remove_disk= raid5_remove_disk,
5339 .spare_active = raid5_spare_active,
5340 .sync_request = sync_request,
5341 .resize = raid5_resize,
5343 #ifdef CONFIG_MD_RAID5_RESHAPE
5344 .check_reshape = raid5_check_reshape,
5345 .start_reshape = raid5_start_reshape,
5346 .finish_reshape = raid5_finish_reshape,
5348 .quiesce = raid5_quiesce,
5351 static int __init raid5_init(void)
5353 register_md_personality(&raid6_personality);
5354 register_md_personality(&raid5_personality);
5355 register_md_personality(&raid4_personality);
5359 static void raid5_exit(void)
5361 unregister_md_personality(&raid6_personality);
5362 unregister_md_personality(&raid5_personality);
5363 unregister_md_personality(&raid4_personality);
5366 module_init(raid5_init);
5367 module_exit(raid5_exit);
5368 MODULE_LICENSE("GPL");
5369 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5370 MODULE_ALIAS("md-raid5");
5371 MODULE_ALIAS("md-raid4");
5372 MODULE_ALIAS("md-level-5");
5373 MODULE_ALIAS("md-level-4");
5374 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5375 MODULE_ALIAS("md-raid6");
5376 MODULE_ALIAS("md-level-6");
5378 /* This used to be two separate modules, they were: */
5379 MODULE_ALIAS("raid5");
5380 MODULE_ALIAS("raid6");