ocfs2: Add metaecc for ocfs2_refcount_block.
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / drivers / md / raid5.c
blob37835538b58ef8aa4e07595274b1ea5ec27e7c81
1 /*
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)
14 * any later version.
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.
22 * BITMAP UNPLUGGING:
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
26 * explanation.
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
32 * new additions.
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
39 * batch.
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
43 * miss any bits.
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>
51 #include "md.h"
52 #include "raid5.h"
53 #include "bitmap.h"
56 * Stripe cache
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
75 * be valid.
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)
86 #else
87 # define CHECK_DEVLOCK()
88 #endif
90 #ifdef DEBUG
91 #define inline
92 #define __inline__
93 #endif
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);
121 --val;
122 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
123 return val;
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)
134 if (sh->ddf_layout)
135 /* ddf always start from first device */
136 return 0;
137 /* md starts just after Q block */
138 if (sh->qd_idx == sh->disks - 1)
139 return 0;
140 else
141 return sh->qd_idx + 1;
143 static inline int raid6_next_disk(int disk, int raid_disks)
145 disk++;
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)
157 int slot;
159 if (idx == sh->pd_idx)
160 return syndrome_disks;
161 if (idx == sh->qd_idx)
162 return syndrome_disks + 1;
163 slot = (*count)++;
164 return slot;
167 static void return_io(struct bio *return_bi)
169 struct bio *bi = return_bi;
170 while (bi) {
172 return_bi = bi->bi_next;
173 bi->bi_next = NULL;
174 bi->bi_size = 0;
175 bio_endio(bi, 0);
176 bi = return_bi;
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);
202 } else {
203 clear_bit(STRIPE_BIT_DELAY, &sh->state);
204 list_add_tail(&sh->lru, &conf->handle_list);
206 md_wakeup_thread(conf->mddev->thread);
207 } else {
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;
228 unsigned long flags;
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);
250 CHECK_DEVLOCK();
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;
261 CHECK_DEVLOCK();
262 if (list_empty(&conf->inactive_list))
263 goto out;
264 first = conf->inactive_list.next;
265 sh = list_entry(first, struct stripe_head, lru);
266 list_del_init(first);
267 remove_hash(sh);
268 atomic_inc(&conf->active_stripes);
269 out:
270 return sh;
273 static void shrink_buffers(struct stripe_head *sh, int num)
275 struct page *p;
276 int i;
278 for (i=0; i<num ; i++) {
279 p = sh->dev[i].page;
280 if (!p)
281 continue;
282 sh->dev[i].page = NULL;
283 put_page(p);
287 static int grow_buffers(struct stripe_head *sh, int num)
289 int i;
291 for (i=0; i<num; i++) {
292 struct page *page;
294 if (!(page = alloc_page(GFP_KERNEL))) {
295 return 1;
297 sh->dev[i].page = page;
299 return 0;
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;
309 int i;
311 BUG_ON(atomic_read(&sh->count) != 0);
312 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
313 BUG_ON(stripe_operations_active(sh));
315 CHECK_DEVLOCK();
316 pr_debug("init_stripe called, stripe %llu\n",
317 (unsigned long long)sh->sector);
319 remove_hash(sh);
321 sh->generation = conf->generation - previous;
322 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
323 sh->sector = sector;
324 stripe_set_idx(sector, conf, previous, sh);
325 sh->state = 0;
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));
337 BUG();
339 dev->flags = 0;
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,
346 short generation)
348 struct stripe_head *sh;
349 struct hlist_node *hn;
351 CHECK_DEVLOCK();
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)
355 return sh;
356 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
357 return NULL;
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, int noquiesce)
367 struct stripe_head *sh;
369 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
371 spin_lock_irq(&conf->device_lock);
373 do {
374 wait_event_lock_irq(conf->wait_for_stripe,
375 conf->quiesce == 0 || noquiesce,
376 conf->device_lock, /* nothing */);
377 sh = __find_stripe(conf, sector, conf->generation - previous);
378 if (!sh) {
379 if (!conf->inactive_blocked)
380 sh = get_free_stripe(conf);
381 if (noblock && sh == NULL)
382 break;
383 if (!sh) {
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),
390 conf->device_lock,
391 raid5_unplug_device(conf->mddev->queue)
393 conf->inactive_blocked = 0;
394 } else
395 init_stripe(sh, sector, previous);
396 } else {
397 if (atomic_read(&sh->count)) {
398 BUG_ON(!list_empty(&sh->lru)
399 && !test_bit(STRIPE_EXPANDING, &sh->state));
400 } else {
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))
405 BUG();
406 list_del_init(&sh->lru);
409 } while (sh == NULL);
411 if (sh)
412 atomic_inc(&sh->count);
414 spin_unlock_irq(&conf->device_lock);
415 return sh;
418 static void
419 raid5_end_read_request(struct bio *bi, int error);
420 static void
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;
428 might_sleep();
430 for (i = disks; i--; ) {
431 int rw;
432 struct bio *bi;
433 mdk_rdev_t *rdev;
434 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
435 rw = WRITE;
436 else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
437 rw = READ;
438 else
439 continue;
441 bi = &sh->dev[i].req;
443 bi->bi_rw = rw;
444 if (rw == WRITE)
445 bi->bi_end_io = raid5_end_write_request;
446 else
447 bi->bi_end_io = raid5_end_read_request;
449 rcu_read_lock();
450 rdev = rcu_dereference(conf->disks[i].rdev);
451 if (rdev && test_bit(Faulty, &rdev->flags))
452 rdev = NULL;
453 if (rdev)
454 atomic_inc(&rdev->nr_pending);
455 rcu_read_unlock();
457 if (rdev) {
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,
466 bi->bi_rw, i);
467 atomic_inc(&sh->count);
468 bi->bi_sector = sh->sector + rdev->data_offset;
469 bi->bi_flags = 1 << BIO_UPTODATE;
470 bi->bi_vcnt = 1;
471 bi->bi_max_vecs = 1;
472 bi->bi_idx = 0;
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;
477 bi->bi_next = NULL;
478 if (rw == WRITE &&
479 test_bit(R5_ReWrite, &sh->dev[i].flags))
480 atomic_add(STRIPE_SECTORS,
481 &rdev->corrected_errors);
482 generic_make_request(bi);
483 } else {
484 if (rw == WRITE)
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)
498 struct bio_vec *bvl;
499 struct page *bio_page;
500 int i;
501 int page_offset;
503 if (bio->bi_sector >= sector)
504 page_offset = (signed)(bio->bi_sector - sector) * 512;
505 else
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;
509 int clen;
510 int b_offset = 0;
512 if (page_offset < 0) {
513 b_offset = -page_offset;
514 page_offset += b_offset;
515 len -= b_offset;
518 if (len > 0 && page_offset + len > STRIPE_SIZE)
519 clen = STRIPE_SIZE - page_offset;
520 else
521 clen = len;
523 if (clen > 0) {
524 b_offset += bio_iovec_idx(bio, i)->bv_offset;
525 bio_page = bio_iovec_idx(bio, i)->bv_page;
526 if (frombio)
527 tx = async_memcpy(page, bio_page, page_offset,
528 b_offset, clen,
529 ASYNC_TX_DEP_ACK,
530 tx, NULL, NULL);
531 else
532 tx = async_memcpy(bio_page, page, b_offset,
533 page_offset, clen,
534 ASYNC_TX_DEP_ACK,
535 tx, NULL, NULL);
537 if (clen < len) /* hit end of page */
538 break;
539 page_offset += len;
542 return tx;
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;
550 int i;
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;
568 BUG_ON(!dev->read);
569 rbi = dev->read;
570 dev->read = NULL;
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;
576 return_bi = rbi;
578 rbi = rbi2;
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);
588 release_stripe(sh);
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;
595 int i;
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)) {
603 struct bio *rbi;
604 spin_lock_irq(&conf->device_lock);
605 dev->read = rbi = dev->toread;
606 dev->toread = NULL;
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,
611 dev->sector, tx);
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);
638 release_stripe(sh);
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;
649 int count = 0;
650 struct dma_async_tx_descriptor *tx;
651 int i;
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--; )
658 if (i != target)
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);
666 else
667 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
668 ASYNC_TX_XOR_ZERO_DST, NULL,
669 ops_complete_compute5, sh);
671 return tx;
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);
707 return tx;
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;
714 int i;
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];
721 struct bio *chosen;
723 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
724 struct bio *wbi;
726 spin_lock(&sh->lock);
727 chosen = dev->towrite;
728 dev->towrite = NULL;
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,
736 dev->sector, tx);
737 wbi = r5_next_bio(wbi, dev->sector);
742 return tx;
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;
763 else {
764 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
765 sh->reconstruct_state = reconstruct_state_result;
768 set_bit(STRIPE_HANDLE, &sh->state);
769 release_stripe(sh);
772 static void
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;
781 int prexor = 0;
782 unsigned long flags;
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) {
791 prexor = 1;
792 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
793 for (i = disks; i--; ) {
794 struct r5dev *dev = &sh->dev[i];
795 if (dev->written)
796 xor_srcs[count++] = dev->page;
798 } else {
799 xor_dest = sh->dev[pd_idx].page;
800 for (i = disks; i--; ) {
801 struct r5dev *dev = &sh->dev[i];
802 if (i != pd_idx)
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);
821 } else
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);
835 release_stripe(sh);
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];
853 if (i != pd_idx)
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)) {
871 ops_run_biofill(sh);
872 overlap_clear++;
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))
879 async_tx_ack(tx);
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);
887 overlap_clear++;
890 if (test_bit(STRIPE_OP_POSTXOR, &ops_request))
891 ops_run_postxor(sh, tx);
893 if (test_bit(STRIPE_OP_CHECK, &ops_request))
894 ops_run_check(sh);
896 if (overlap_clear)
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);
908 if (!sh)
909 return 0;
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);
917 return 0;
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);
924 release_stripe(sh);
925 return 1;
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),
940 0, 0, NULL);
941 if (!sc)
942 return 1;
943 conf->slab_cache = sc;
944 conf->pool_size = devs;
945 while (num--)
946 if (!grow_one_stripe(conf))
947 return 1;
948 return 0;
951 static int resize_stripes(raid5_conf_t *conf, int newsize)
953 /* Make all the stripes able to hold 'newsize' devices.
954 * New slots in each stripe get 'page' set to a new page.
956 * This happens in stages:
957 * 1/ create a new kmem_cache and allocate the required number of
958 * stripe_heads.
959 * 2/ gather all the old stripe_heads and tranfer the pages across
960 * to the new stripe_heads. This will have the side effect of
961 * freezing the array as once all stripe_heads have been collected,
962 * no IO will be possible. Old stripe heads are freed once their
963 * pages have been transferred over, and the old kmem_cache is
964 * freed when all stripes are done.
965 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
966 * we simple return a failre status - no need to clean anything up.
967 * 4/ allocate new pages for the new slots in the new stripe_heads.
968 * If this fails, we don't bother trying the shrink the
969 * stripe_heads down again, we just leave them as they are.
970 * As each stripe_head is processed the new one is released into
971 * active service.
973 * Once step2 is started, we cannot afford to wait for a write,
974 * so we use GFP_NOIO allocations.
976 struct stripe_head *osh, *nsh;
977 LIST_HEAD(newstripes);
978 struct disk_info *ndisks;
979 int err;
980 struct kmem_cache *sc;
981 int i;
983 if (newsize <= conf->pool_size)
984 return 0; /* never bother to shrink */
986 err = md_allow_write(conf->mddev);
987 if (err)
988 return err;
990 /* Step 1 */
991 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
992 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
993 0, 0, NULL);
994 if (!sc)
995 return -ENOMEM;
997 for (i = conf->max_nr_stripes; i; i--) {
998 nsh = kmem_cache_alloc(sc, GFP_KERNEL);
999 if (!nsh)
1000 break;
1002 memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1004 nsh->raid_conf = conf;
1005 spin_lock_init(&nsh->lock);
1007 list_add(&nsh->lru, &newstripes);
1009 if (i) {
1010 /* didn't get enough, give up */
1011 while (!list_empty(&newstripes)) {
1012 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1013 list_del(&nsh->lru);
1014 kmem_cache_free(sc, nsh);
1016 kmem_cache_destroy(sc);
1017 return -ENOMEM;
1019 /* Step 2 - Must use GFP_NOIO now.
1020 * OK, we have enough stripes, start collecting inactive
1021 * stripes and copying them over
1023 list_for_each_entry(nsh, &newstripes, lru) {
1024 spin_lock_irq(&conf->device_lock);
1025 wait_event_lock_irq(conf->wait_for_stripe,
1026 !list_empty(&conf->inactive_list),
1027 conf->device_lock,
1028 unplug_slaves(conf->mddev)
1030 osh = get_free_stripe(conf);
1031 spin_unlock_irq(&conf->device_lock);
1032 atomic_set(&nsh->count, 1);
1033 for(i=0; i<conf->pool_size; i++)
1034 nsh->dev[i].page = osh->dev[i].page;
1035 for( ; i<newsize; i++)
1036 nsh->dev[i].page = NULL;
1037 kmem_cache_free(conf->slab_cache, osh);
1039 kmem_cache_destroy(conf->slab_cache);
1041 /* Step 3.
1042 * At this point, we are holding all the stripes so the array
1043 * is completely stalled, so now is a good time to resize
1044 * conf->disks.
1046 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1047 if (ndisks) {
1048 for (i=0; i<conf->raid_disks; i++)
1049 ndisks[i] = conf->disks[i];
1050 kfree(conf->disks);
1051 conf->disks = ndisks;
1052 } else
1053 err = -ENOMEM;
1055 /* Step 4, return new stripes to service */
1056 while(!list_empty(&newstripes)) {
1057 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1058 list_del_init(&nsh->lru);
1059 for (i=conf->raid_disks; i < newsize; i++)
1060 if (nsh->dev[i].page == NULL) {
1061 struct page *p = alloc_page(GFP_NOIO);
1062 nsh->dev[i].page = p;
1063 if (!p)
1064 err = -ENOMEM;
1066 release_stripe(nsh);
1068 /* critical section pass, GFP_NOIO no longer needed */
1070 conf->slab_cache = sc;
1071 conf->active_name = 1-conf->active_name;
1072 conf->pool_size = newsize;
1073 return err;
1076 static int drop_one_stripe(raid5_conf_t *conf)
1078 struct stripe_head *sh;
1080 spin_lock_irq(&conf->device_lock);
1081 sh = get_free_stripe(conf);
1082 spin_unlock_irq(&conf->device_lock);
1083 if (!sh)
1084 return 0;
1085 BUG_ON(atomic_read(&sh->count));
1086 shrink_buffers(sh, conf->pool_size);
1087 kmem_cache_free(conf->slab_cache, sh);
1088 atomic_dec(&conf->active_stripes);
1089 return 1;
1092 static void shrink_stripes(raid5_conf_t *conf)
1094 while (drop_one_stripe(conf))
1097 if (conf->slab_cache)
1098 kmem_cache_destroy(conf->slab_cache);
1099 conf->slab_cache = NULL;
1102 static void raid5_end_read_request(struct bio * bi, int error)
1104 struct stripe_head *sh = bi->bi_private;
1105 raid5_conf_t *conf = sh->raid_conf;
1106 int disks = sh->disks, i;
1107 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1108 char b[BDEVNAME_SIZE];
1109 mdk_rdev_t *rdev;
1112 for (i=0 ; i<disks; i++)
1113 if (bi == &sh->dev[i].req)
1114 break;
1116 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1117 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1118 uptodate);
1119 if (i == disks) {
1120 BUG();
1121 return;
1124 if (uptodate) {
1125 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1126 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1127 rdev = conf->disks[i].rdev;
1128 printk_rl(KERN_INFO "raid5:%s: read error corrected"
1129 " (%lu sectors at %llu on %s)\n",
1130 mdname(conf->mddev), STRIPE_SECTORS,
1131 (unsigned long long)(sh->sector
1132 + rdev->data_offset),
1133 bdevname(rdev->bdev, b));
1134 clear_bit(R5_ReadError, &sh->dev[i].flags);
1135 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1137 if (atomic_read(&conf->disks[i].rdev->read_errors))
1138 atomic_set(&conf->disks[i].rdev->read_errors, 0);
1139 } else {
1140 const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1141 int retry = 0;
1142 rdev = conf->disks[i].rdev;
1144 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1145 atomic_inc(&rdev->read_errors);
1146 if (conf->mddev->degraded)
1147 printk_rl(KERN_WARNING
1148 "raid5:%s: read error not correctable "
1149 "(sector %llu on %s).\n",
1150 mdname(conf->mddev),
1151 (unsigned long long)(sh->sector
1152 + rdev->data_offset),
1153 bdn);
1154 else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1155 /* Oh, no!!! */
1156 printk_rl(KERN_WARNING
1157 "raid5:%s: read error NOT corrected!! "
1158 "(sector %llu on %s).\n",
1159 mdname(conf->mddev),
1160 (unsigned long long)(sh->sector
1161 + rdev->data_offset),
1162 bdn);
1163 else if (atomic_read(&rdev->read_errors)
1164 > conf->max_nr_stripes)
1165 printk(KERN_WARNING
1166 "raid5:%s: Too many read errors, failing device %s.\n",
1167 mdname(conf->mddev), bdn);
1168 else
1169 retry = 1;
1170 if (retry)
1171 set_bit(R5_ReadError, &sh->dev[i].flags);
1172 else {
1173 clear_bit(R5_ReadError, &sh->dev[i].flags);
1174 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1175 md_error(conf->mddev, rdev);
1178 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1179 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1180 set_bit(STRIPE_HANDLE, &sh->state);
1181 release_stripe(sh);
1184 static void raid5_end_write_request(struct bio *bi, int error)
1186 struct stripe_head *sh = bi->bi_private;
1187 raid5_conf_t *conf = sh->raid_conf;
1188 int disks = sh->disks, i;
1189 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1191 for (i=0 ; i<disks; i++)
1192 if (bi == &sh->dev[i].req)
1193 break;
1195 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1196 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1197 uptodate);
1198 if (i == disks) {
1199 BUG();
1200 return;
1203 if (!uptodate)
1204 md_error(conf->mddev, conf->disks[i].rdev);
1206 rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1208 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1209 set_bit(STRIPE_HANDLE, &sh->state);
1210 release_stripe(sh);
1214 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1216 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1218 struct r5dev *dev = &sh->dev[i];
1220 bio_init(&dev->req);
1221 dev->req.bi_io_vec = &dev->vec;
1222 dev->req.bi_vcnt++;
1223 dev->req.bi_max_vecs++;
1224 dev->vec.bv_page = dev->page;
1225 dev->vec.bv_len = STRIPE_SIZE;
1226 dev->vec.bv_offset = 0;
1228 dev->req.bi_sector = sh->sector;
1229 dev->req.bi_private = sh;
1231 dev->flags = 0;
1232 dev->sector = compute_blocknr(sh, i, previous);
1235 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1237 char b[BDEVNAME_SIZE];
1238 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1239 pr_debug("raid5: error called\n");
1241 if (!test_bit(Faulty, &rdev->flags)) {
1242 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1243 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1244 unsigned long flags;
1245 spin_lock_irqsave(&conf->device_lock, flags);
1246 mddev->degraded++;
1247 spin_unlock_irqrestore(&conf->device_lock, flags);
1249 * if recovery was running, make sure it aborts.
1251 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1253 set_bit(Faulty, &rdev->flags);
1254 printk(KERN_ALERT
1255 "raid5: Disk failure on %s, disabling device.\n"
1256 "raid5: Operation continuing on %d devices.\n",
1257 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1262 * Input: a 'big' sector number,
1263 * Output: index of the data and parity disk, and the sector # in them.
1265 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1266 int previous, int *dd_idx,
1267 struct stripe_head *sh)
1269 long stripe;
1270 unsigned long chunk_number;
1271 unsigned int chunk_offset;
1272 int pd_idx, qd_idx;
1273 int ddf_layout = 0;
1274 sector_t new_sector;
1275 int algorithm = previous ? conf->prev_algo
1276 : conf->algorithm;
1277 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1278 : conf->chunk_sectors;
1279 int raid_disks = previous ? conf->previous_raid_disks
1280 : conf->raid_disks;
1281 int data_disks = raid_disks - conf->max_degraded;
1283 /* First compute the information on this sector */
1286 * Compute the chunk number and the sector offset inside the chunk
1288 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1289 chunk_number = r_sector;
1290 BUG_ON(r_sector != chunk_number);
1293 * Compute the stripe number
1295 stripe = chunk_number / data_disks;
1298 * Compute the data disk and parity disk indexes inside the stripe
1300 *dd_idx = chunk_number % data_disks;
1303 * Select the parity disk based on the user selected algorithm.
1305 pd_idx = qd_idx = ~0;
1306 switch(conf->level) {
1307 case 4:
1308 pd_idx = data_disks;
1309 break;
1310 case 5:
1311 switch (algorithm) {
1312 case ALGORITHM_LEFT_ASYMMETRIC:
1313 pd_idx = data_disks - stripe % raid_disks;
1314 if (*dd_idx >= pd_idx)
1315 (*dd_idx)++;
1316 break;
1317 case ALGORITHM_RIGHT_ASYMMETRIC:
1318 pd_idx = stripe % raid_disks;
1319 if (*dd_idx >= pd_idx)
1320 (*dd_idx)++;
1321 break;
1322 case ALGORITHM_LEFT_SYMMETRIC:
1323 pd_idx = data_disks - stripe % raid_disks;
1324 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1325 break;
1326 case ALGORITHM_RIGHT_SYMMETRIC:
1327 pd_idx = stripe % raid_disks;
1328 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1329 break;
1330 case ALGORITHM_PARITY_0:
1331 pd_idx = 0;
1332 (*dd_idx)++;
1333 break;
1334 case ALGORITHM_PARITY_N:
1335 pd_idx = data_disks;
1336 break;
1337 default:
1338 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1339 algorithm);
1340 BUG();
1342 break;
1343 case 6:
1345 switch (algorithm) {
1346 case ALGORITHM_LEFT_ASYMMETRIC:
1347 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1348 qd_idx = pd_idx + 1;
1349 if (pd_idx == raid_disks-1) {
1350 (*dd_idx)++; /* Q D D D P */
1351 qd_idx = 0;
1352 } else if (*dd_idx >= pd_idx)
1353 (*dd_idx) += 2; /* D D P Q D */
1354 break;
1355 case ALGORITHM_RIGHT_ASYMMETRIC:
1356 pd_idx = stripe % raid_disks;
1357 qd_idx = pd_idx + 1;
1358 if (pd_idx == raid_disks-1) {
1359 (*dd_idx)++; /* Q D D D P */
1360 qd_idx = 0;
1361 } else if (*dd_idx >= pd_idx)
1362 (*dd_idx) += 2; /* D D P Q D */
1363 break;
1364 case ALGORITHM_LEFT_SYMMETRIC:
1365 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1366 qd_idx = (pd_idx + 1) % raid_disks;
1367 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1368 break;
1369 case ALGORITHM_RIGHT_SYMMETRIC:
1370 pd_idx = stripe % raid_disks;
1371 qd_idx = (pd_idx + 1) % raid_disks;
1372 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1373 break;
1375 case ALGORITHM_PARITY_0:
1376 pd_idx = 0;
1377 qd_idx = 1;
1378 (*dd_idx) += 2;
1379 break;
1380 case ALGORITHM_PARITY_N:
1381 pd_idx = data_disks;
1382 qd_idx = data_disks + 1;
1383 break;
1385 case ALGORITHM_ROTATING_ZERO_RESTART:
1386 /* Exactly the same as RIGHT_ASYMMETRIC, but or
1387 * of blocks for computing Q is different.
1389 pd_idx = stripe % raid_disks;
1390 qd_idx = pd_idx + 1;
1391 if (pd_idx == raid_disks-1) {
1392 (*dd_idx)++; /* Q D D D P */
1393 qd_idx = 0;
1394 } else if (*dd_idx >= pd_idx)
1395 (*dd_idx) += 2; /* D D P Q D */
1396 ddf_layout = 1;
1397 break;
1399 case ALGORITHM_ROTATING_N_RESTART:
1400 /* Same a left_asymmetric, by first stripe is
1401 * D D D P Q rather than
1402 * Q D D D P
1404 pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1405 qd_idx = pd_idx + 1;
1406 if (pd_idx == raid_disks-1) {
1407 (*dd_idx)++; /* Q D D D P */
1408 qd_idx = 0;
1409 } else if (*dd_idx >= pd_idx)
1410 (*dd_idx) += 2; /* D D P Q D */
1411 ddf_layout = 1;
1412 break;
1414 case ALGORITHM_ROTATING_N_CONTINUE:
1415 /* Same as left_symmetric but Q is before P */
1416 pd_idx = raid_disks - 1 - (stripe % raid_disks);
1417 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1418 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1419 ddf_layout = 1;
1420 break;
1422 case ALGORITHM_LEFT_ASYMMETRIC_6:
1423 /* RAID5 left_asymmetric, with Q on last device */
1424 pd_idx = data_disks - stripe % (raid_disks-1);
1425 if (*dd_idx >= pd_idx)
1426 (*dd_idx)++;
1427 qd_idx = raid_disks - 1;
1428 break;
1430 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1431 pd_idx = stripe % (raid_disks-1);
1432 if (*dd_idx >= pd_idx)
1433 (*dd_idx)++;
1434 qd_idx = raid_disks - 1;
1435 break;
1437 case ALGORITHM_LEFT_SYMMETRIC_6:
1438 pd_idx = data_disks - stripe % (raid_disks-1);
1439 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1440 qd_idx = raid_disks - 1;
1441 break;
1443 case ALGORITHM_RIGHT_SYMMETRIC_6:
1444 pd_idx = stripe % (raid_disks-1);
1445 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1446 qd_idx = raid_disks - 1;
1447 break;
1449 case ALGORITHM_PARITY_0_6:
1450 pd_idx = 0;
1451 (*dd_idx)++;
1452 qd_idx = raid_disks - 1;
1453 break;
1456 default:
1457 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1458 algorithm);
1459 BUG();
1461 break;
1464 if (sh) {
1465 sh->pd_idx = pd_idx;
1466 sh->qd_idx = qd_idx;
1467 sh->ddf_layout = ddf_layout;
1470 * Finally, compute the new sector number
1472 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1473 return new_sector;
1477 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1479 raid5_conf_t *conf = sh->raid_conf;
1480 int raid_disks = sh->disks;
1481 int data_disks = raid_disks - conf->max_degraded;
1482 sector_t new_sector = sh->sector, check;
1483 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1484 : conf->chunk_sectors;
1485 int algorithm = previous ? conf->prev_algo
1486 : conf->algorithm;
1487 sector_t stripe;
1488 int chunk_offset;
1489 int chunk_number, dummy1, dd_idx = i;
1490 sector_t r_sector;
1491 struct stripe_head sh2;
1494 chunk_offset = sector_div(new_sector, sectors_per_chunk);
1495 stripe = new_sector;
1496 BUG_ON(new_sector != stripe);
1498 if (i == sh->pd_idx)
1499 return 0;
1500 switch(conf->level) {
1501 case 4: break;
1502 case 5:
1503 switch (algorithm) {
1504 case ALGORITHM_LEFT_ASYMMETRIC:
1505 case ALGORITHM_RIGHT_ASYMMETRIC:
1506 if (i > sh->pd_idx)
1507 i--;
1508 break;
1509 case ALGORITHM_LEFT_SYMMETRIC:
1510 case ALGORITHM_RIGHT_SYMMETRIC:
1511 if (i < sh->pd_idx)
1512 i += raid_disks;
1513 i -= (sh->pd_idx + 1);
1514 break;
1515 case ALGORITHM_PARITY_0:
1516 i -= 1;
1517 break;
1518 case ALGORITHM_PARITY_N:
1519 break;
1520 default:
1521 printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1522 algorithm);
1523 BUG();
1525 break;
1526 case 6:
1527 if (i == sh->qd_idx)
1528 return 0; /* It is the Q disk */
1529 switch (algorithm) {
1530 case ALGORITHM_LEFT_ASYMMETRIC:
1531 case ALGORITHM_RIGHT_ASYMMETRIC:
1532 case ALGORITHM_ROTATING_ZERO_RESTART:
1533 case ALGORITHM_ROTATING_N_RESTART:
1534 if (sh->pd_idx == raid_disks-1)
1535 i--; /* Q D D D P */
1536 else if (i > sh->pd_idx)
1537 i -= 2; /* D D P Q D */
1538 break;
1539 case ALGORITHM_LEFT_SYMMETRIC:
1540 case ALGORITHM_RIGHT_SYMMETRIC:
1541 if (sh->pd_idx == raid_disks-1)
1542 i--; /* Q D D D P */
1543 else {
1544 /* D D P Q D */
1545 if (i < sh->pd_idx)
1546 i += raid_disks;
1547 i -= (sh->pd_idx + 2);
1549 break;
1550 case ALGORITHM_PARITY_0:
1551 i -= 2;
1552 break;
1553 case ALGORITHM_PARITY_N:
1554 break;
1555 case ALGORITHM_ROTATING_N_CONTINUE:
1556 if (sh->pd_idx == 0)
1557 i--; /* P D D D Q */
1558 else if (i > sh->pd_idx)
1559 i -= 2; /* D D Q P D */
1560 break;
1561 case ALGORITHM_LEFT_ASYMMETRIC_6:
1562 case ALGORITHM_RIGHT_ASYMMETRIC_6:
1563 if (i > sh->pd_idx)
1564 i--;
1565 break;
1566 case ALGORITHM_LEFT_SYMMETRIC_6:
1567 case ALGORITHM_RIGHT_SYMMETRIC_6:
1568 if (i < sh->pd_idx)
1569 i += data_disks + 1;
1570 i -= (sh->pd_idx + 1);
1571 break;
1572 case ALGORITHM_PARITY_0_6:
1573 i -= 1;
1574 break;
1575 default:
1576 printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1577 algorithm);
1578 BUG();
1580 break;
1583 chunk_number = stripe * data_disks + i;
1584 r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1586 check = raid5_compute_sector(conf, r_sector,
1587 previous, &dummy1, &sh2);
1588 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1589 || sh2.qd_idx != sh->qd_idx) {
1590 printk(KERN_ERR "compute_blocknr: map not correct\n");
1591 return 0;
1593 return r_sector;
1599 * Copy data between a page in the stripe cache, and one or more bion
1600 * The page could align with the middle of the bio, or there could be
1601 * several bion, each with several bio_vecs, which cover part of the page
1602 * Multiple bion are linked together on bi_next. There may be extras
1603 * at the end of this list. We ignore them.
1605 static void copy_data(int frombio, struct bio *bio,
1606 struct page *page,
1607 sector_t sector)
1609 char *pa = page_address(page);
1610 struct bio_vec *bvl;
1611 int i;
1612 int page_offset;
1614 if (bio->bi_sector >= sector)
1615 page_offset = (signed)(bio->bi_sector - sector) * 512;
1616 else
1617 page_offset = (signed)(sector - bio->bi_sector) * -512;
1618 bio_for_each_segment(bvl, bio, i) {
1619 int len = bio_iovec_idx(bio,i)->bv_len;
1620 int clen;
1621 int b_offset = 0;
1623 if (page_offset < 0) {
1624 b_offset = -page_offset;
1625 page_offset += b_offset;
1626 len -= b_offset;
1629 if (len > 0 && page_offset + len > STRIPE_SIZE)
1630 clen = STRIPE_SIZE - page_offset;
1631 else clen = len;
1633 if (clen > 0) {
1634 char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1635 if (frombio)
1636 memcpy(pa+page_offset, ba+b_offset, clen);
1637 else
1638 memcpy(ba+b_offset, pa+page_offset, clen);
1639 __bio_kunmap_atomic(ba, KM_USER0);
1641 if (clen < len) /* hit end of page */
1642 break;
1643 page_offset += len;
1647 #define check_xor() do { \
1648 if (count == MAX_XOR_BLOCKS) { \
1649 xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1650 count = 0; \
1652 } while(0)
1654 static void compute_parity6(struct stripe_head *sh, int method)
1656 raid5_conf_t *conf = sh->raid_conf;
1657 int i, pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1658 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1659 struct bio *chosen;
1660 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1661 void *ptrs[syndrome_disks+2];
1663 pd_idx = sh->pd_idx;
1664 qd_idx = sh->qd_idx;
1665 d0_idx = raid6_d0(sh);
1667 pr_debug("compute_parity, stripe %llu, method %d\n",
1668 (unsigned long long)sh->sector, method);
1670 switch(method) {
1671 case READ_MODIFY_WRITE:
1672 BUG(); /* READ_MODIFY_WRITE N/A for RAID-6 */
1673 case RECONSTRUCT_WRITE:
1674 for (i= disks; i-- ;)
1675 if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1676 chosen = sh->dev[i].towrite;
1677 sh->dev[i].towrite = NULL;
1679 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1680 wake_up(&conf->wait_for_overlap);
1682 BUG_ON(sh->dev[i].written);
1683 sh->dev[i].written = chosen;
1685 break;
1686 case CHECK_PARITY:
1687 BUG(); /* Not implemented yet */
1690 for (i = disks; i--;)
1691 if (sh->dev[i].written) {
1692 sector_t sector = sh->dev[i].sector;
1693 struct bio *wbi = sh->dev[i].written;
1694 while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1695 copy_data(1, wbi, sh->dev[i].page, sector);
1696 wbi = r5_next_bio(wbi, sector);
1699 set_bit(R5_LOCKED, &sh->dev[i].flags);
1700 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1703 /* Note that unlike RAID-5, the ordering of the disks matters greatly.*/
1705 for (i = 0; i < disks; i++)
1706 ptrs[i] = (void *)raid6_empty_zero_page;
1708 count = 0;
1709 i = d0_idx;
1710 do {
1711 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1713 ptrs[slot] = page_address(sh->dev[i].page);
1714 if (slot < syndrome_disks &&
1715 !test_bit(R5_UPTODATE, &sh->dev[i].flags)) {
1716 printk(KERN_ERR "block %d/%d not uptodate "
1717 "on parity calc\n", i, count);
1718 BUG();
1721 i = raid6_next_disk(i, disks);
1722 } while (i != d0_idx);
1723 BUG_ON(count != syndrome_disks);
1725 raid6_call.gen_syndrome(syndrome_disks+2, STRIPE_SIZE, ptrs);
1727 switch(method) {
1728 case RECONSTRUCT_WRITE:
1729 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1730 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1731 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1732 set_bit(R5_LOCKED, &sh->dev[qd_idx].flags);
1733 break;
1734 case UPDATE_PARITY:
1735 set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1736 set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1737 break;
1742 /* Compute one missing block */
1743 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1745 int i, count, disks = sh->disks;
1746 void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1747 int qd_idx = sh->qd_idx;
1749 pr_debug("compute_block_1, stripe %llu, idx %d\n",
1750 (unsigned long long)sh->sector, dd_idx);
1752 if ( dd_idx == qd_idx ) {
1753 /* We're actually computing the Q drive */
1754 compute_parity6(sh, UPDATE_PARITY);
1755 } else {
1756 dest = page_address(sh->dev[dd_idx].page);
1757 if (!nozero) memset(dest, 0, STRIPE_SIZE);
1758 count = 0;
1759 for (i = disks ; i--; ) {
1760 if (i == dd_idx || i == qd_idx)
1761 continue;
1762 p = page_address(sh->dev[i].page);
1763 if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1764 ptr[count++] = p;
1765 else
1766 printk("compute_block() %d, stripe %llu, %d"
1767 " not present\n", dd_idx,
1768 (unsigned long long)sh->sector, i);
1770 check_xor();
1772 if (count)
1773 xor_blocks(count, STRIPE_SIZE, dest, ptr);
1774 if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1775 else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1779 /* Compute two missing blocks */
1780 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1782 int i, count, disks = sh->disks;
1783 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1784 int d0_idx = raid6_d0(sh);
1785 int faila = -1, failb = -1;
1786 /**** FIX THIS: This could be very bad if disks is close to 256 ****/
1787 void *ptrs[syndrome_disks+2];
1789 for (i = 0; i < disks ; i++)
1790 ptrs[i] = (void *)raid6_empty_zero_page;
1791 count = 0;
1792 i = d0_idx;
1793 do {
1794 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1796 ptrs[slot] = page_address(sh->dev[i].page);
1798 if (i == dd_idx1)
1799 faila = slot;
1800 if (i == dd_idx2)
1801 failb = slot;
1802 i = raid6_next_disk(i, disks);
1803 } while (i != d0_idx);
1804 BUG_ON(count != syndrome_disks);
1806 BUG_ON(faila == failb);
1807 if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1809 pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1810 (unsigned long long)sh->sector, dd_idx1, dd_idx2,
1811 faila, failb);
1813 if (failb == syndrome_disks+1) {
1814 /* Q disk is one of the missing disks */
1815 if (faila == syndrome_disks) {
1816 /* Missing P+Q, just recompute */
1817 compute_parity6(sh, UPDATE_PARITY);
1818 return;
1819 } else {
1820 /* We're missing D+Q; recompute D from P */
1821 compute_block_1(sh, ((dd_idx1 == sh->qd_idx) ?
1822 dd_idx2 : dd_idx1),
1824 compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1825 return;
1829 /* We're missing D+P or D+D; */
1830 if (failb == syndrome_disks) {
1831 /* We're missing D+P. */
1832 raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE, faila, ptrs);
1833 } else {
1834 /* We're missing D+D. */
1835 raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE, faila, failb,
1836 ptrs);
1839 /* Both the above update both missing blocks */
1840 set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1841 set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1844 static void
1845 schedule_reconstruction5(struct stripe_head *sh, struct stripe_head_state *s,
1846 int rcw, int expand)
1848 int i, pd_idx = sh->pd_idx, disks = sh->disks;
1850 if (rcw) {
1851 /* if we are not expanding this is a proper write request, and
1852 * there will be bios with new data to be drained into the
1853 * stripe cache
1855 if (!expand) {
1856 sh->reconstruct_state = reconstruct_state_drain_run;
1857 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1858 } else
1859 sh->reconstruct_state = reconstruct_state_run;
1861 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1863 for (i = disks; i--; ) {
1864 struct r5dev *dev = &sh->dev[i];
1866 if (dev->towrite) {
1867 set_bit(R5_LOCKED, &dev->flags);
1868 set_bit(R5_Wantdrain, &dev->flags);
1869 if (!expand)
1870 clear_bit(R5_UPTODATE, &dev->flags);
1871 s->locked++;
1874 if (s->locked + 1 == disks)
1875 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1876 atomic_inc(&sh->raid_conf->pending_full_writes);
1877 } else {
1878 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1879 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1881 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1882 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1883 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1884 set_bit(STRIPE_OP_POSTXOR, &s->ops_request);
1886 for (i = disks; i--; ) {
1887 struct r5dev *dev = &sh->dev[i];
1888 if (i == pd_idx)
1889 continue;
1891 if (dev->towrite &&
1892 (test_bit(R5_UPTODATE, &dev->flags) ||
1893 test_bit(R5_Wantcompute, &dev->flags))) {
1894 set_bit(R5_Wantdrain, &dev->flags);
1895 set_bit(R5_LOCKED, &dev->flags);
1896 clear_bit(R5_UPTODATE, &dev->flags);
1897 s->locked++;
1902 /* keep the parity disk locked while asynchronous operations
1903 * are in flight
1905 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1906 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1907 s->locked++;
1909 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
1910 __func__, (unsigned long long)sh->sector,
1911 s->locked, s->ops_request);
1915 * Each stripe/dev can have one or more bion attached.
1916 * toread/towrite point to the first in a chain.
1917 * The bi_next chain must be in order.
1919 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1921 struct bio **bip;
1922 raid5_conf_t *conf = sh->raid_conf;
1923 int firstwrite=0;
1925 pr_debug("adding bh b#%llu to stripe s#%llu\n",
1926 (unsigned long long)bi->bi_sector,
1927 (unsigned long long)sh->sector);
1930 spin_lock(&sh->lock);
1931 spin_lock_irq(&conf->device_lock);
1932 if (forwrite) {
1933 bip = &sh->dev[dd_idx].towrite;
1934 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1935 firstwrite = 1;
1936 } else
1937 bip = &sh->dev[dd_idx].toread;
1938 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1939 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1940 goto overlap;
1941 bip = & (*bip)->bi_next;
1943 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1944 goto overlap;
1946 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1947 if (*bip)
1948 bi->bi_next = *bip;
1949 *bip = bi;
1950 bi->bi_phys_segments++;
1951 spin_unlock_irq(&conf->device_lock);
1952 spin_unlock(&sh->lock);
1954 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1955 (unsigned long long)bi->bi_sector,
1956 (unsigned long long)sh->sector, dd_idx);
1958 if (conf->mddev->bitmap && firstwrite) {
1959 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1960 STRIPE_SECTORS, 0);
1961 sh->bm_seq = conf->seq_flush+1;
1962 set_bit(STRIPE_BIT_DELAY, &sh->state);
1965 if (forwrite) {
1966 /* check if page is covered */
1967 sector_t sector = sh->dev[dd_idx].sector;
1968 for (bi=sh->dev[dd_idx].towrite;
1969 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1970 bi && bi->bi_sector <= sector;
1971 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1972 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1973 sector = bi->bi_sector + (bi->bi_size>>9);
1975 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1976 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1978 return 1;
1980 overlap:
1981 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1982 spin_unlock_irq(&conf->device_lock);
1983 spin_unlock(&sh->lock);
1984 return 0;
1987 static void end_reshape(raid5_conf_t *conf);
1989 static int page_is_zero(struct page *p)
1991 char *a = page_address(p);
1992 return ((*(u32*)a) == 0 &&
1993 memcmp(a, a+4, STRIPE_SIZE-4)==0);
1996 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
1997 struct stripe_head *sh)
1999 int sectors_per_chunk =
2000 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2001 int dd_idx;
2002 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2003 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2005 raid5_compute_sector(conf,
2006 stripe * (disks - conf->max_degraded)
2007 *sectors_per_chunk + chunk_offset,
2008 previous,
2009 &dd_idx, sh);
2012 static void
2013 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2014 struct stripe_head_state *s, int disks,
2015 struct bio **return_bi)
2017 int i;
2018 for (i = disks; i--; ) {
2019 struct bio *bi;
2020 int bitmap_end = 0;
2022 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2023 mdk_rdev_t *rdev;
2024 rcu_read_lock();
2025 rdev = rcu_dereference(conf->disks[i].rdev);
2026 if (rdev && test_bit(In_sync, &rdev->flags))
2027 /* multiple read failures in one stripe */
2028 md_error(conf->mddev, rdev);
2029 rcu_read_unlock();
2031 spin_lock_irq(&conf->device_lock);
2032 /* fail all writes first */
2033 bi = sh->dev[i].towrite;
2034 sh->dev[i].towrite = NULL;
2035 if (bi) {
2036 s->to_write--;
2037 bitmap_end = 1;
2040 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2041 wake_up(&conf->wait_for_overlap);
2043 while (bi && bi->bi_sector <
2044 sh->dev[i].sector + STRIPE_SECTORS) {
2045 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2046 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2047 if (!raid5_dec_bi_phys_segments(bi)) {
2048 md_write_end(conf->mddev);
2049 bi->bi_next = *return_bi;
2050 *return_bi = bi;
2052 bi = nextbi;
2054 /* and fail all 'written' */
2055 bi = sh->dev[i].written;
2056 sh->dev[i].written = NULL;
2057 if (bi) bitmap_end = 1;
2058 while (bi && bi->bi_sector <
2059 sh->dev[i].sector + STRIPE_SECTORS) {
2060 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2061 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2062 if (!raid5_dec_bi_phys_segments(bi)) {
2063 md_write_end(conf->mddev);
2064 bi->bi_next = *return_bi;
2065 *return_bi = bi;
2067 bi = bi2;
2070 /* fail any reads if this device is non-operational and
2071 * the data has not reached the cache yet.
2073 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2074 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2075 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2076 bi = sh->dev[i].toread;
2077 sh->dev[i].toread = NULL;
2078 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2079 wake_up(&conf->wait_for_overlap);
2080 if (bi) s->to_read--;
2081 while (bi && bi->bi_sector <
2082 sh->dev[i].sector + STRIPE_SECTORS) {
2083 struct bio *nextbi =
2084 r5_next_bio(bi, sh->dev[i].sector);
2085 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2086 if (!raid5_dec_bi_phys_segments(bi)) {
2087 bi->bi_next = *return_bi;
2088 *return_bi = bi;
2090 bi = nextbi;
2093 spin_unlock_irq(&conf->device_lock);
2094 if (bitmap_end)
2095 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2096 STRIPE_SECTORS, 0, 0);
2099 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2100 if (atomic_dec_and_test(&conf->pending_full_writes))
2101 md_wakeup_thread(conf->mddev->thread);
2104 /* fetch_block5 - checks the given member device to see if its data needs
2105 * to be read or computed to satisfy a request.
2107 * Returns 1 when no more member devices need to be checked, otherwise returns
2108 * 0 to tell the loop in handle_stripe_fill5 to continue
2110 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2111 int disk_idx, int disks)
2113 struct r5dev *dev = &sh->dev[disk_idx];
2114 struct r5dev *failed_dev = &sh->dev[s->failed_num];
2116 /* is the data in this block needed, and can we get it? */
2117 if (!test_bit(R5_LOCKED, &dev->flags) &&
2118 !test_bit(R5_UPTODATE, &dev->flags) &&
2119 (dev->toread ||
2120 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2121 s->syncing || s->expanding ||
2122 (s->failed &&
2123 (failed_dev->toread ||
2124 (failed_dev->towrite &&
2125 !test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2126 /* We would like to get this block, possibly by computing it,
2127 * otherwise read it if the backing disk is insync
2129 if ((s->uptodate == disks - 1) &&
2130 (s->failed && disk_idx == s->failed_num)) {
2131 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2132 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2133 set_bit(R5_Wantcompute, &dev->flags);
2134 sh->ops.target = disk_idx;
2135 s->req_compute = 1;
2136 /* Careful: from this point on 'uptodate' is in the eye
2137 * of raid5_run_ops which services 'compute' operations
2138 * before writes. R5_Wantcompute flags a block that will
2139 * be R5_UPTODATE by the time it is needed for a
2140 * subsequent operation.
2142 s->uptodate++;
2143 return 1; /* uptodate + compute == disks */
2144 } else if (test_bit(R5_Insync, &dev->flags)) {
2145 set_bit(R5_LOCKED, &dev->flags);
2146 set_bit(R5_Wantread, &dev->flags);
2147 s->locked++;
2148 pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2149 s->syncing);
2153 return 0;
2157 * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2159 static void handle_stripe_fill5(struct stripe_head *sh,
2160 struct stripe_head_state *s, int disks)
2162 int i;
2164 /* look for blocks to read/compute, skip this if a compute
2165 * is already in flight, or if the stripe contents are in the
2166 * midst of changing due to a write
2168 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2169 !sh->reconstruct_state)
2170 for (i = disks; i--; )
2171 if (fetch_block5(sh, s, i, disks))
2172 break;
2173 set_bit(STRIPE_HANDLE, &sh->state);
2176 static void handle_stripe_fill6(struct stripe_head *sh,
2177 struct stripe_head_state *s, struct r6_state *r6s,
2178 int disks)
2180 int i;
2181 for (i = disks; i--; ) {
2182 struct r5dev *dev = &sh->dev[i];
2183 if (!test_bit(R5_LOCKED, &dev->flags) &&
2184 !test_bit(R5_UPTODATE, &dev->flags) &&
2185 (dev->toread || (dev->towrite &&
2186 !test_bit(R5_OVERWRITE, &dev->flags)) ||
2187 s->syncing || s->expanding ||
2188 (s->failed >= 1 &&
2189 (sh->dev[r6s->failed_num[0]].toread ||
2190 s->to_write)) ||
2191 (s->failed >= 2 &&
2192 (sh->dev[r6s->failed_num[1]].toread ||
2193 s->to_write)))) {
2194 /* we would like to get this block, possibly
2195 * by computing it, but we might not be able to
2197 if ((s->uptodate == disks - 1) &&
2198 (s->failed && (i == r6s->failed_num[0] ||
2199 i == r6s->failed_num[1]))) {
2200 pr_debug("Computing stripe %llu block %d\n",
2201 (unsigned long long)sh->sector, i);
2202 compute_block_1(sh, i, 0);
2203 s->uptodate++;
2204 } else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2205 /* Computing 2-failure is *very* expensive; only
2206 * do it if failed >= 2
2208 int other;
2209 for (other = disks; other--; ) {
2210 if (other == i)
2211 continue;
2212 if (!test_bit(R5_UPTODATE,
2213 &sh->dev[other].flags))
2214 break;
2216 BUG_ON(other < 0);
2217 pr_debug("Computing stripe %llu blocks %d,%d\n",
2218 (unsigned long long)sh->sector,
2219 i, other);
2220 compute_block_2(sh, i, other);
2221 s->uptodate += 2;
2222 } else if (test_bit(R5_Insync, &dev->flags)) {
2223 set_bit(R5_LOCKED, &dev->flags);
2224 set_bit(R5_Wantread, &dev->flags);
2225 s->locked++;
2226 pr_debug("Reading block %d (sync=%d)\n",
2227 i, s->syncing);
2231 set_bit(STRIPE_HANDLE, &sh->state);
2235 /* handle_stripe_clean_event
2236 * any written block on an uptodate or failed drive can be returned.
2237 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2238 * never LOCKED, so we don't need to test 'failed' directly.
2240 static void handle_stripe_clean_event(raid5_conf_t *conf,
2241 struct stripe_head *sh, int disks, struct bio **return_bi)
2243 int i;
2244 struct r5dev *dev;
2246 for (i = disks; i--; )
2247 if (sh->dev[i].written) {
2248 dev = &sh->dev[i];
2249 if (!test_bit(R5_LOCKED, &dev->flags) &&
2250 test_bit(R5_UPTODATE, &dev->flags)) {
2251 /* We can return any write requests */
2252 struct bio *wbi, *wbi2;
2253 int bitmap_end = 0;
2254 pr_debug("Return write for disc %d\n", i);
2255 spin_lock_irq(&conf->device_lock);
2256 wbi = dev->written;
2257 dev->written = NULL;
2258 while (wbi && wbi->bi_sector <
2259 dev->sector + STRIPE_SECTORS) {
2260 wbi2 = r5_next_bio(wbi, dev->sector);
2261 if (!raid5_dec_bi_phys_segments(wbi)) {
2262 md_write_end(conf->mddev);
2263 wbi->bi_next = *return_bi;
2264 *return_bi = wbi;
2266 wbi = wbi2;
2268 if (dev->towrite == NULL)
2269 bitmap_end = 1;
2270 spin_unlock_irq(&conf->device_lock);
2271 if (bitmap_end)
2272 bitmap_endwrite(conf->mddev->bitmap,
2273 sh->sector,
2274 STRIPE_SECTORS,
2275 !test_bit(STRIPE_DEGRADED, &sh->state),
2280 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2281 if (atomic_dec_and_test(&conf->pending_full_writes))
2282 md_wakeup_thread(conf->mddev->thread);
2285 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2286 struct stripe_head *sh, struct stripe_head_state *s, int disks)
2288 int rmw = 0, rcw = 0, i;
2289 for (i = disks; i--; ) {
2290 /* would I have to read this buffer for read_modify_write */
2291 struct r5dev *dev = &sh->dev[i];
2292 if ((dev->towrite || i == sh->pd_idx) &&
2293 !test_bit(R5_LOCKED, &dev->flags) &&
2294 !(test_bit(R5_UPTODATE, &dev->flags) ||
2295 test_bit(R5_Wantcompute, &dev->flags))) {
2296 if (test_bit(R5_Insync, &dev->flags))
2297 rmw++;
2298 else
2299 rmw += 2*disks; /* cannot read it */
2301 /* Would I have to read this buffer for reconstruct_write */
2302 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2303 !test_bit(R5_LOCKED, &dev->flags) &&
2304 !(test_bit(R5_UPTODATE, &dev->flags) ||
2305 test_bit(R5_Wantcompute, &dev->flags))) {
2306 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2307 else
2308 rcw += 2*disks;
2311 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2312 (unsigned long long)sh->sector, rmw, rcw);
2313 set_bit(STRIPE_HANDLE, &sh->state);
2314 if (rmw < rcw && rmw > 0)
2315 /* prefer read-modify-write, but need to get some data */
2316 for (i = disks; i--; ) {
2317 struct r5dev *dev = &sh->dev[i];
2318 if ((dev->towrite || i == sh->pd_idx) &&
2319 !test_bit(R5_LOCKED, &dev->flags) &&
2320 !(test_bit(R5_UPTODATE, &dev->flags) ||
2321 test_bit(R5_Wantcompute, &dev->flags)) &&
2322 test_bit(R5_Insync, &dev->flags)) {
2323 if (
2324 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2325 pr_debug("Read_old block "
2326 "%d for r-m-w\n", i);
2327 set_bit(R5_LOCKED, &dev->flags);
2328 set_bit(R5_Wantread, &dev->flags);
2329 s->locked++;
2330 } else {
2331 set_bit(STRIPE_DELAYED, &sh->state);
2332 set_bit(STRIPE_HANDLE, &sh->state);
2336 if (rcw <= rmw && rcw > 0)
2337 /* want reconstruct write, but need to get some data */
2338 for (i = disks; i--; ) {
2339 struct r5dev *dev = &sh->dev[i];
2340 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2341 i != sh->pd_idx &&
2342 !test_bit(R5_LOCKED, &dev->flags) &&
2343 !(test_bit(R5_UPTODATE, &dev->flags) ||
2344 test_bit(R5_Wantcompute, &dev->flags)) &&
2345 test_bit(R5_Insync, &dev->flags)) {
2346 if (
2347 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2348 pr_debug("Read_old block "
2349 "%d for Reconstruct\n", i);
2350 set_bit(R5_LOCKED, &dev->flags);
2351 set_bit(R5_Wantread, &dev->flags);
2352 s->locked++;
2353 } else {
2354 set_bit(STRIPE_DELAYED, &sh->state);
2355 set_bit(STRIPE_HANDLE, &sh->state);
2359 /* now if nothing is locked, and if we have enough data,
2360 * we can start a write request
2362 /* since handle_stripe can be called at any time we need to handle the
2363 * case where a compute block operation has been submitted and then a
2364 * subsequent call wants to start a write request. raid5_run_ops only
2365 * handles the case where compute block and postxor are requested
2366 * simultaneously. If this is not the case then new writes need to be
2367 * held off until the compute completes.
2369 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2370 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2371 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2372 schedule_reconstruction5(sh, s, rcw == 0, 0);
2375 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2376 struct stripe_head *sh, struct stripe_head_state *s,
2377 struct r6_state *r6s, int disks)
2379 int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2380 int qd_idx = sh->qd_idx;
2381 for (i = disks; i--; ) {
2382 struct r5dev *dev = &sh->dev[i];
2383 /* Would I have to read this buffer for reconstruct_write */
2384 if (!test_bit(R5_OVERWRITE, &dev->flags)
2385 && i != pd_idx && i != qd_idx
2386 && (!test_bit(R5_LOCKED, &dev->flags)
2387 ) &&
2388 !test_bit(R5_UPTODATE, &dev->flags)) {
2389 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2390 else {
2391 pr_debug("raid6: must_compute: "
2392 "disk %d flags=%#lx\n", i, dev->flags);
2393 must_compute++;
2397 pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2398 (unsigned long long)sh->sector, rcw, must_compute);
2399 set_bit(STRIPE_HANDLE, &sh->state);
2401 if (rcw > 0)
2402 /* want reconstruct write, but need to get some data */
2403 for (i = disks; i--; ) {
2404 struct r5dev *dev = &sh->dev[i];
2405 if (!test_bit(R5_OVERWRITE, &dev->flags)
2406 && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2407 && !test_bit(R5_LOCKED, &dev->flags) &&
2408 !test_bit(R5_UPTODATE, &dev->flags) &&
2409 test_bit(R5_Insync, &dev->flags)) {
2410 if (
2411 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2412 pr_debug("Read_old stripe %llu "
2413 "block %d for Reconstruct\n",
2414 (unsigned long long)sh->sector, i);
2415 set_bit(R5_LOCKED, &dev->flags);
2416 set_bit(R5_Wantread, &dev->flags);
2417 s->locked++;
2418 } else {
2419 pr_debug("Request delayed stripe %llu "
2420 "block %d for Reconstruct\n",
2421 (unsigned long long)sh->sector, i);
2422 set_bit(STRIPE_DELAYED, &sh->state);
2423 set_bit(STRIPE_HANDLE, &sh->state);
2427 /* now if nothing is locked, and if we have enough data, we can start a
2428 * write request
2430 if (s->locked == 0 && rcw == 0 &&
2431 !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2432 if (must_compute > 0) {
2433 /* We have failed blocks and need to compute them */
2434 switch (s->failed) {
2435 case 0:
2436 BUG();
2437 case 1:
2438 compute_block_1(sh, r6s->failed_num[0], 0);
2439 break;
2440 case 2:
2441 compute_block_2(sh, r6s->failed_num[0],
2442 r6s->failed_num[1]);
2443 break;
2444 default: /* This request should have been failed? */
2445 BUG();
2449 pr_debug("Computing parity for stripe %llu\n",
2450 (unsigned long long)sh->sector);
2451 compute_parity6(sh, RECONSTRUCT_WRITE);
2452 /* now every locked buffer is ready to be written */
2453 for (i = disks; i--; )
2454 if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2455 pr_debug("Writing stripe %llu block %d\n",
2456 (unsigned long long)sh->sector, i);
2457 s->locked++;
2458 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2460 if (s->locked == disks)
2461 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2462 atomic_inc(&conf->pending_full_writes);
2463 /* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2464 set_bit(STRIPE_INSYNC, &sh->state);
2466 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2467 atomic_dec(&conf->preread_active_stripes);
2468 if (atomic_read(&conf->preread_active_stripes) <
2469 IO_THRESHOLD)
2470 md_wakeup_thread(conf->mddev->thread);
2475 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2476 struct stripe_head_state *s, int disks)
2478 struct r5dev *dev = NULL;
2480 set_bit(STRIPE_HANDLE, &sh->state);
2482 switch (sh->check_state) {
2483 case check_state_idle:
2484 /* start a new check operation if there are no failures */
2485 if (s->failed == 0) {
2486 BUG_ON(s->uptodate != disks);
2487 sh->check_state = check_state_run;
2488 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2489 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2490 s->uptodate--;
2491 break;
2493 dev = &sh->dev[s->failed_num];
2494 /* fall through */
2495 case check_state_compute_result:
2496 sh->check_state = check_state_idle;
2497 if (!dev)
2498 dev = &sh->dev[sh->pd_idx];
2500 /* check that a write has not made the stripe insync */
2501 if (test_bit(STRIPE_INSYNC, &sh->state))
2502 break;
2504 /* either failed parity check, or recovery is happening */
2505 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2506 BUG_ON(s->uptodate != disks);
2508 set_bit(R5_LOCKED, &dev->flags);
2509 s->locked++;
2510 set_bit(R5_Wantwrite, &dev->flags);
2512 clear_bit(STRIPE_DEGRADED, &sh->state);
2513 set_bit(STRIPE_INSYNC, &sh->state);
2514 break;
2515 case check_state_run:
2516 break; /* we will be called again upon completion */
2517 case check_state_check_result:
2518 sh->check_state = check_state_idle;
2520 /* if a failure occurred during the check operation, leave
2521 * STRIPE_INSYNC not set and let the stripe be handled again
2523 if (s->failed)
2524 break;
2526 /* handle a successful check operation, if parity is correct
2527 * we are done. Otherwise update the mismatch count and repair
2528 * parity if !MD_RECOVERY_CHECK
2530 if (sh->ops.zero_sum_result == 0)
2531 /* parity is correct (on disc,
2532 * not in buffer any more)
2534 set_bit(STRIPE_INSYNC, &sh->state);
2535 else {
2536 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2537 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2538 /* don't try to repair!! */
2539 set_bit(STRIPE_INSYNC, &sh->state);
2540 else {
2541 sh->check_state = check_state_compute_run;
2542 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2543 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2544 set_bit(R5_Wantcompute,
2545 &sh->dev[sh->pd_idx].flags);
2546 sh->ops.target = sh->pd_idx;
2547 s->uptodate++;
2550 break;
2551 case check_state_compute_run:
2552 break;
2553 default:
2554 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2555 __func__, sh->check_state,
2556 (unsigned long long) sh->sector);
2557 BUG();
2562 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2563 struct stripe_head_state *s,
2564 struct r6_state *r6s, struct page *tmp_page,
2565 int disks)
2567 int update_p = 0, update_q = 0;
2568 struct r5dev *dev;
2569 int pd_idx = sh->pd_idx;
2570 int qd_idx = sh->qd_idx;
2572 set_bit(STRIPE_HANDLE, &sh->state);
2574 BUG_ON(s->failed > 2);
2575 BUG_ON(s->uptodate < disks);
2576 /* Want to check and possibly repair P and Q.
2577 * However there could be one 'failed' device, in which
2578 * case we can only check one of them, possibly using the
2579 * other to generate missing data
2582 /* If !tmp_page, we cannot do the calculations,
2583 * but as we have set STRIPE_HANDLE, we will soon be called
2584 * by stripe_handle with a tmp_page - just wait until then.
2586 if (tmp_page) {
2587 if (s->failed == r6s->q_failed) {
2588 /* The only possible failed device holds 'Q', so it
2589 * makes sense to check P (If anything else were failed,
2590 * we would have used P to recreate it).
2592 compute_block_1(sh, pd_idx, 1);
2593 if (!page_is_zero(sh->dev[pd_idx].page)) {
2594 compute_block_1(sh, pd_idx, 0);
2595 update_p = 1;
2598 if (!r6s->q_failed && s->failed < 2) {
2599 /* q is not failed, and we didn't use it to generate
2600 * anything, so it makes sense to check it
2602 memcpy(page_address(tmp_page),
2603 page_address(sh->dev[qd_idx].page),
2604 STRIPE_SIZE);
2605 compute_parity6(sh, UPDATE_PARITY);
2606 if (memcmp(page_address(tmp_page),
2607 page_address(sh->dev[qd_idx].page),
2608 STRIPE_SIZE) != 0) {
2609 clear_bit(STRIPE_INSYNC, &sh->state);
2610 update_q = 1;
2613 if (update_p || update_q) {
2614 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2615 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2616 /* don't try to repair!! */
2617 update_p = update_q = 0;
2620 /* now write out any block on a failed drive,
2621 * or P or Q if they need it
2624 if (s->failed == 2) {
2625 dev = &sh->dev[r6s->failed_num[1]];
2626 s->locked++;
2627 set_bit(R5_LOCKED, &dev->flags);
2628 set_bit(R5_Wantwrite, &dev->flags);
2630 if (s->failed >= 1) {
2631 dev = &sh->dev[r6s->failed_num[0]];
2632 s->locked++;
2633 set_bit(R5_LOCKED, &dev->flags);
2634 set_bit(R5_Wantwrite, &dev->flags);
2637 if (update_p) {
2638 dev = &sh->dev[pd_idx];
2639 s->locked++;
2640 set_bit(R5_LOCKED, &dev->flags);
2641 set_bit(R5_Wantwrite, &dev->flags);
2643 if (update_q) {
2644 dev = &sh->dev[qd_idx];
2645 s->locked++;
2646 set_bit(R5_LOCKED, &dev->flags);
2647 set_bit(R5_Wantwrite, &dev->flags);
2649 clear_bit(STRIPE_DEGRADED, &sh->state);
2651 set_bit(STRIPE_INSYNC, &sh->state);
2655 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2656 struct r6_state *r6s)
2658 int i;
2660 /* We have read all the blocks in this stripe and now we need to
2661 * copy some of them into a target stripe for expand.
2663 struct dma_async_tx_descriptor *tx = NULL;
2664 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2665 for (i = 0; i < sh->disks; i++)
2666 if (i != sh->pd_idx && i != sh->qd_idx) {
2667 int dd_idx, j;
2668 struct stripe_head *sh2;
2670 sector_t bn = compute_blocknr(sh, i, 1);
2671 sector_t s = raid5_compute_sector(conf, bn, 0,
2672 &dd_idx, NULL);
2673 sh2 = get_active_stripe(conf, s, 0, 1, 1);
2674 if (sh2 == NULL)
2675 /* so far only the early blocks of this stripe
2676 * have been requested. When later blocks
2677 * get requested, we will try again
2679 continue;
2680 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2681 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2682 /* must have already done this block */
2683 release_stripe(sh2);
2684 continue;
2687 /* place all the copies on one channel */
2688 tx = async_memcpy(sh2->dev[dd_idx].page,
2689 sh->dev[i].page, 0, 0, STRIPE_SIZE,
2690 ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2692 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2693 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2694 for (j = 0; j < conf->raid_disks; j++)
2695 if (j != sh2->pd_idx &&
2696 (!r6s || j != sh2->qd_idx) &&
2697 !test_bit(R5_Expanded, &sh2->dev[j].flags))
2698 break;
2699 if (j == conf->raid_disks) {
2700 set_bit(STRIPE_EXPAND_READY, &sh2->state);
2701 set_bit(STRIPE_HANDLE, &sh2->state);
2703 release_stripe(sh2);
2706 /* done submitting copies, wait for them to complete */
2707 if (tx) {
2708 async_tx_ack(tx);
2709 dma_wait_for_async_tx(tx);
2715 * handle_stripe - do things to a stripe.
2717 * We lock the stripe and then examine the state of various bits
2718 * to see what needs to be done.
2719 * Possible results:
2720 * return some read request which now have data
2721 * return some write requests which are safely on disc
2722 * schedule a read on some buffers
2723 * schedule a write of some buffers
2724 * return confirmation of parity correctness
2726 * buffers are taken off read_list or write_list, and bh_cache buffers
2727 * get BH_Lock set before the stripe lock is released.
2731 static bool handle_stripe5(struct stripe_head *sh)
2733 raid5_conf_t *conf = sh->raid_conf;
2734 int disks = sh->disks, i;
2735 struct bio *return_bi = NULL;
2736 struct stripe_head_state s;
2737 struct r5dev *dev;
2738 mdk_rdev_t *blocked_rdev = NULL;
2739 int prexor;
2741 memset(&s, 0, sizeof(s));
2742 pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2743 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2744 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2745 sh->reconstruct_state);
2747 spin_lock(&sh->lock);
2748 clear_bit(STRIPE_HANDLE, &sh->state);
2749 clear_bit(STRIPE_DELAYED, &sh->state);
2751 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2752 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2753 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2755 /* Now to look around and see what can be done */
2756 rcu_read_lock();
2757 for (i=disks; i--; ) {
2758 mdk_rdev_t *rdev;
2759 struct r5dev *dev = &sh->dev[i];
2760 clear_bit(R5_Insync, &dev->flags);
2762 pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2763 "written %p\n", i, dev->flags, dev->toread, dev->read,
2764 dev->towrite, dev->written);
2766 /* maybe we can request a biofill operation
2768 * new wantfill requests are only permitted while
2769 * ops_complete_biofill is guaranteed to be inactive
2771 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2772 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2773 set_bit(R5_Wantfill, &dev->flags);
2775 /* now count some things */
2776 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2777 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2778 if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2780 if (test_bit(R5_Wantfill, &dev->flags))
2781 s.to_fill++;
2782 else if (dev->toread)
2783 s.to_read++;
2784 if (dev->towrite) {
2785 s.to_write++;
2786 if (!test_bit(R5_OVERWRITE, &dev->flags))
2787 s.non_overwrite++;
2789 if (dev->written)
2790 s.written++;
2791 rdev = rcu_dereference(conf->disks[i].rdev);
2792 if (blocked_rdev == NULL &&
2793 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2794 blocked_rdev = rdev;
2795 atomic_inc(&rdev->nr_pending);
2797 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2798 /* The ReadError flag will just be confusing now */
2799 clear_bit(R5_ReadError, &dev->flags);
2800 clear_bit(R5_ReWrite, &dev->flags);
2802 if (!rdev || !test_bit(In_sync, &rdev->flags)
2803 || test_bit(R5_ReadError, &dev->flags)) {
2804 s.failed++;
2805 s.failed_num = i;
2806 } else
2807 set_bit(R5_Insync, &dev->flags);
2809 rcu_read_unlock();
2811 if (unlikely(blocked_rdev)) {
2812 if (s.syncing || s.expanding || s.expanded ||
2813 s.to_write || s.written) {
2814 set_bit(STRIPE_HANDLE, &sh->state);
2815 goto unlock;
2817 /* There is nothing for the blocked_rdev to block */
2818 rdev_dec_pending(blocked_rdev, conf->mddev);
2819 blocked_rdev = NULL;
2822 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2823 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2824 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2827 pr_debug("locked=%d uptodate=%d to_read=%d"
2828 " to_write=%d failed=%d failed_num=%d\n",
2829 s.locked, s.uptodate, s.to_read, s.to_write,
2830 s.failed, s.failed_num);
2831 /* check if the array has lost two devices and, if so, some requests might
2832 * need to be failed
2834 if (s.failed > 1 && s.to_read+s.to_write+s.written)
2835 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
2836 if (s.failed > 1 && s.syncing) {
2837 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2838 clear_bit(STRIPE_SYNCING, &sh->state);
2839 s.syncing = 0;
2842 /* might be able to return some write requests if the parity block
2843 * is safe, or on a failed drive
2845 dev = &sh->dev[sh->pd_idx];
2846 if ( s.written &&
2847 ((test_bit(R5_Insync, &dev->flags) &&
2848 !test_bit(R5_LOCKED, &dev->flags) &&
2849 test_bit(R5_UPTODATE, &dev->flags)) ||
2850 (s.failed == 1 && s.failed_num == sh->pd_idx)))
2851 handle_stripe_clean_event(conf, sh, disks, &return_bi);
2853 /* Now we might consider reading some blocks, either to check/generate
2854 * parity, or to satisfy requests
2855 * or to load a block that is being partially written.
2857 if (s.to_read || s.non_overwrite ||
2858 (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
2859 handle_stripe_fill5(sh, &s, disks);
2861 /* Now we check to see if any write operations have recently
2862 * completed
2864 prexor = 0;
2865 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
2866 prexor = 1;
2867 if (sh->reconstruct_state == reconstruct_state_drain_result ||
2868 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
2869 sh->reconstruct_state = reconstruct_state_idle;
2871 /* All the 'written' buffers and the parity block are ready to
2872 * be written back to disk
2874 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2875 for (i = disks; i--; ) {
2876 dev = &sh->dev[i];
2877 if (test_bit(R5_LOCKED, &dev->flags) &&
2878 (i == sh->pd_idx || dev->written)) {
2879 pr_debug("Writing block %d\n", i);
2880 set_bit(R5_Wantwrite, &dev->flags);
2881 if (prexor)
2882 continue;
2883 if (!test_bit(R5_Insync, &dev->flags) ||
2884 (i == sh->pd_idx && s.failed == 0))
2885 set_bit(STRIPE_INSYNC, &sh->state);
2888 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2889 atomic_dec(&conf->preread_active_stripes);
2890 if (atomic_read(&conf->preread_active_stripes) <
2891 IO_THRESHOLD)
2892 md_wakeup_thread(conf->mddev->thread);
2896 /* Now to consider new write requests and what else, if anything
2897 * should be read. We do not handle new writes when:
2898 * 1/ A 'write' operation (copy+xor) is already in flight.
2899 * 2/ A 'check' operation is in flight, as it may clobber the parity
2900 * block.
2902 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
2903 handle_stripe_dirtying5(conf, sh, &s, disks);
2905 /* maybe we need to check and possibly fix the parity for this stripe
2906 * Any reads will already have been scheduled, so we just see if enough
2907 * data is available. The parity check is held off while parity
2908 * dependent operations are in flight.
2910 if (sh->check_state ||
2911 (s.syncing && s.locked == 0 &&
2912 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
2913 !test_bit(STRIPE_INSYNC, &sh->state)))
2914 handle_parity_checks5(conf, sh, &s, disks);
2916 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2917 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2918 clear_bit(STRIPE_SYNCING, &sh->state);
2921 /* If the failed drive is just a ReadError, then we might need to progress
2922 * the repair/check process
2924 if (s.failed == 1 && !conf->mddev->ro &&
2925 test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2926 && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2927 && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2929 dev = &sh->dev[s.failed_num];
2930 if (!test_bit(R5_ReWrite, &dev->flags)) {
2931 set_bit(R5_Wantwrite, &dev->flags);
2932 set_bit(R5_ReWrite, &dev->flags);
2933 set_bit(R5_LOCKED, &dev->flags);
2934 s.locked++;
2935 } else {
2936 /* let's read it back */
2937 set_bit(R5_Wantread, &dev->flags);
2938 set_bit(R5_LOCKED, &dev->flags);
2939 s.locked++;
2943 /* Finish reconstruct operations initiated by the expansion process */
2944 if (sh->reconstruct_state == reconstruct_state_result) {
2945 struct stripe_head *sh2
2946 = get_active_stripe(conf, sh->sector, 1, 1, 1);
2947 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
2948 /* sh cannot be written until sh2 has been read.
2949 * so arrange for sh to be delayed a little
2951 set_bit(STRIPE_DELAYED, &sh->state);
2952 set_bit(STRIPE_HANDLE, &sh->state);
2953 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
2954 &sh2->state))
2955 atomic_inc(&conf->preread_active_stripes);
2956 release_stripe(sh2);
2957 goto unlock;
2959 if (sh2)
2960 release_stripe(sh2);
2962 sh->reconstruct_state = reconstruct_state_idle;
2963 clear_bit(STRIPE_EXPANDING, &sh->state);
2964 for (i = conf->raid_disks; i--; ) {
2965 set_bit(R5_Wantwrite, &sh->dev[i].flags);
2966 set_bit(R5_LOCKED, &sh->dev[i].flags);
2967 s.locked++;
2971 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2972 !sh->reconstruct_state) {
2973 /* Need to write out all blocks after computing parity */
2974 sh->disks = conf->raid_disks;
2975 stripe_set_idx(sh->sector, conf, 0, sh);
2976 schedule_reconstruction5(sh, &s, 1, 1);
2977 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
2978 clear_bit(STRIPE_EXPAND_READY, &sh->state);
2979 atomic_dec(&conf->reshape_stripes);
2980 wake_up(&conf->wait_for_overlap);
2981 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2984 if (s.expanding && s.locked == 0 &&
2985 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
2986 handle_stripe_expansion(conf, sh, NULL);
2988 unlock:
2989 spin_unlock(&sh->lock);
2991 /* wait for this device to become unblocked */
2992 if (unlikely(blocked_rdev))
2993 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2995 if (s.ops_request)
2996 raid5_run_ops(sh, s.ops_request);
2998 ops_run_io(sh, &s);
3000 return_io(return_bi);
3002 return blocked_rdev == NULL;
3005 static bool handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
3007 raid5_conf_t *conf = sh->raid_conf;
3008 int disks = sh->disks;
3009 struct bio *return_bi = NULL;
3010 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3011 struct stripe_head_state s;
3012 struct r6_state r6s;
3013 struct r5dev *dev, *pdev, *qdev;
3014 mdk_rdev_t *blocked_rdev = NULL;
3016 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3017 "pd_idx=%d, qd_idx=%d\n",
3018 (unsigned long long)sh->sector, sh->state,
3019 atomic_read(&sh->count), pd_idx, qd_idx);
3020 memset(&s, 0, sizeof(s));
3022 spin_lock(&sh->lock);
3023 clear_bit(STRIPE_HANDLE, &sh->state);
3024 clear_bit(STRIPE_DELAYED, &sh->state);
3026 s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3027 s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3028 s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3029 /* Now to look around and see what can be done */
3031 rcu_read_lock();
3032 for (i=disks; i--; ) {
3033 mdk_rdev_t *rdev;
3034 dev = &sh->dev[i];
3035 clear_bit(R5_Insync, &dev->flags);
3037 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3038 i, dev->flags, dev->toread, dev->towrite, dev->written);
3039 /* maybe we can reply to a read */
3040 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
3041 struct bio *rbi, *rbi2;
3042 pr_debug("Return read for disc %d\n", i);
3043 spin_lock_irq(&conf->device_lock);
3044 rbi = dev->toread;
3045 dev->toread = NULL;
3046 if (test_and_clear_bit(R5_Overlap, &dev->flags))
3047 wake_up(&conf->wait_for_overlap);
3048 spin_unlock_irq(&conf->device_lock);
3049 while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
3050 copy_data(0, rbi, dev->page, dev->sector);
3051 rbi2 = r5_next_bio(rbi, dev->sector);
3052 spin_lock_irq(&conf->device_lock);
3053 if (!raid5_dec_bi_phys_segments(rbi)) {
3054 rbi->bi_next = return_bi;
3055 return_bi = rbi;
3057 spin_unlock_irq(&conf->device_lock);
3058 rbi = rbi2;
3062 /* now count some things */
3063 if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3064 if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3067 if (dev->toread)
3068 s.to_read++;
3069 if (dev->towrite) {
3070 s.to_write++;
3071 if (!test_bit(R5_OVERWRITE, &dev->flags))
3072 s.non_overwrite++;
3074 if (dev->written)
3075 s.written++;
3076 rdev = rcu_dereference(conf->disks[i].rdev);
3077 if (blocked_rdev == NULL &&
3078 rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3079 blocked_rdev = rdev;
3080 atomic_inc(&rdev->nr_pending);
3082 if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3083 /* The ReadError flag will just be confusing now */
3084 clear_bit(R5_ReadError, &dev->flags);
3085 clear_bit(R5_ReWrite, &dev->flags);
3087 if (!rdev || !test_bit(In_sync, &rdev->flags)
3088 || test_bit(R5_ReadError, &dev->flags)) {
3089 if (s.failed < 2)
3090 r6s.failed_num[s.failed] = i;
3091 s.failed++;
3092 } else
3093 set_bit(R5_Insync, &dev->flags);
3095 rcu_read_unlock();
3097 if (unlikely(blocked_rdev)) {
3098 if (s.syncing || s.expanding || s.expanded ||
3099 s.to_write || s.written) {
3100 set_bit(STRIPE_HANDLE, &sh->state);
3101 goto unlock;
3103 /* There is nothing for the blocked_rdev to block */
3104 rdev_dec_pending(blocked_rdev, conf->mddev);
3105 blocked_rdev = NULL;
3108 pr_debug("locked=%d uptodate=%d to_read=%d"
3109 " to_write=%d failed=%d failed_num=%d,%d\n",
3110 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3111 r6s.failed_num[0], r6s.failed_num[1]);
3112 /* check if the array has lost >2 devices and, if so, some requests
3113 * might need to be failed
3115 if (s.failed > 2 && s.to_read+s.to_write+s.written)
3116 handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3117 if (s.failed > 2 && s.syncing) {
3118 md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3119 clear_bit(STRIPE_SYNCING, &sh->state);
3120 s.syncing = 0;
3124 * might be able to return some write requests if the parity blocks
3125 * are safe, or on a failed drive
3127 pdev = &sh->dev[pd_idx];
3128 r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3129 || (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3130 qdev = &sh->dev[qd_idx];
3131 r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3132 || (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3134 if ( s.written &&
3135 ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3136 && !test_bit(R5_LOCKED, &pdev->flags)
3137 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3138 ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3139 && !test_bit(R5_LOCKED, &qdev->flags)
3140 && test_bit(R5_UPTODATE, &qdev->flags)))))
3141 handle_stripe_clean_event(conf, sh, disks, &return_bi);
3143 /* Now we might consider reading some blocks, either to check/generate
3144 * parity, or to satisfy requests
3145 * or to load a block that is being partially written.
3147 if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3148 (s.syncing && (s.uptodate < disks)) || s.expanding)
3149 handle_stripe_fill6(sh, &s, &r6s, disks);
3151 /* now to consider writing and what else, if anything should be read */
3152 if (s.to_write)
3153 handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3155 /* maybe we need to check and possibly fix the parity for this stripe
3156 * Any reads will already have been scheduled, so we just see if enough
3157 * data is available
3159 if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3160 handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3162 if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3163 md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3164 clear_bit(STRIPE_SYNCING, &sh->state);
3167 /* If the failed drives are just a ReadError, then we might need
3168 * to progress the repair/check process
3170 if (s.failed <= 2 && !conf->mddev->ro)
3171 for (i = 0; i < s.failed; i++) {
3172 dev = &sh->dev[r6s.failed_num[i]];
3173 if (test_bit(R5_ReadError, &dev->flags)
3174 && !test_bit(R5_LOCKED, &dev->flags)
3175 && test_bit(R5_UPTODATE, &dev->flags)
3177 if (!test_bit(R5_ReWrite, &dev->flags)) {
3178 set_bit(R5_Wantwrite, &dev->flags);
3179 set_bit(R5_ReWrite, &dev->flags);
3180 set_bit(R5_LOCKED, &dev->flags);
3181 } else {
3182 /* let's read it back */
3183 set_bit(R5_Wantread, &dev->flags);
3184 set_bit(R5_LOCKED, &dev->flags);
3189 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3190 struct stripe_head *sh2
3191 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3192 if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3193 /* sh cannot be written until sh2 has been read.
3194 * so arrange for sh to be delayed a little
3196 set_bit(STRIPE_DELAYED, &sh->state);
3197 set_bit(STRIPE_HANDLE, &sh->state);
3198 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3199 &sh2->state))
3200 atomic_inc(&conf->preread_active_stripes);
3201 release_stripe(sh2);
3202 goto unlock;
3204 if (sh2)
3205 release_stripe(sh2);
3207 /* Need to write out all blocks after computing P&Q */
3208 sh->disks = conf->raid_disks;
3209 stripe_set_idx(sh->sector, conf, 0, sh);
3210 compute_parity6(sh, RECONSTRUCT_WRITE);
3211 for (i = conf->raid_disks ; i-- ; ) {
3212 set_bit(R5_LOCKED, &sh->dev[i].flags);
3213 s.locked++;
3214 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3216 clear_bit(STRIPE_EXPANDING, &sh->state);
3217 } else if (s.expanded) {
3218 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3219 atomic_dec(&conf->reshape_stripes);
3220 wake_up(&conf->wait_for_overlap);
3221 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3224 if (s.expanding && s.locked == 0 &&
3225 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3226 handle_stripe_expansion(conf, sh, &r6s);
3228 unlock:
3229 spin_unlock(&sh->lock);
3231 /* wait for this device to become unblocked */
3232 if (unlikely(blocked_rdev))
3233 md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3235 ops_run_io(sh, &s);
3237 return_io(return_bi);
3239 return blocked_rdev == NULL;
3242 /* returns true if the stripe was handled */
3243 static bool handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3245 if (sh->raid_conf->level == 6)
3246 return handle_stripe6(sh, tmp_page);
3247 else
3248 return handle_stripe5(sh);
3253 static void raid5_activate_delayed(raid5_conf_t *conf)
3255 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3256 while (!list_empty(&conf->delayed_list)) {
3257 struct list_head *l = conf->delayed_list.next;
3258 struct stripe_head *sh;
3259 sh = list_entry(l, struct stripe_head, lru);
3260 list_del_init(l);
3261 clear_bit(STRIPE_DELAYED, &sh->state);
3262 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3263 atomic_inc(&conf->preread_active_stripes);
3264 list_add_tail(&sh->lru, &conf->hold_list);
3266 } else
3267 blk_plug_device(conf->mddev->queue);
3270 static void activate_bit_delay(raid5_conf_t *conf)
3272 /* device_lock is held */
3273 struct list_head head;
3274 list_add(&head, &conf->bitmap_list);
3275 list_del_init(&conf->bitmap_list);
3276 while (!list_empty(&head)) {
3277 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3278 list_del_init(&sh->lru);
3279 atomic_inc(&sh->count);
3280 __release_stripe(conf, sh);
3284 static void unplug_slaves(mddev_t *mddev)
3286 raid5_conf_t *conf = mddev->private;
3287 int i;
3289 rcu_read_lock();
3290 for (i = 0; i < conf->raid_disks; i++) {
3291 mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3292 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3293 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3295 atomic_inc(&rdev->nr_pending);
3296 rcu_read_unlock();
3298 blk_unplug(r_queue);
3300 rdev_dec_pending(rdev, mddev);
3301 rcu_read_lock();
3304 rcu_read_unlock();
3307 static void raid5_unplug_device(struct request_queue *q)
3309 mddev_t *mddev = q->queuedata;
3310 raid5_conf_t *conf = mddev->private;
3311 unsigned long flags;
3313 spin_lock_irqsave(&conf->device_lock, flags);
3315 if (blk_remove_plug(q)) {
3316 conf->seq_flush++;
3317 raid5_activate_delayed(conf);
3319 md_wakeup_thread(mddev->thread);
3321 spin_unlock_irqrestore(&conf->device_lock, flags);
3323 unplug_slaves(mddev);
3326 static int raid5_congested(void *data, int bits)
3328 mddev_t *mddev = data;
3329 raid5_conf_t *conf = mddev->private;
3331 /* No difference between reads and writes. Just check
3332 * how busy the stripe_cache is
3334 if (conf->inactive_blocked)
3335 return 1;
3336 if (conf->quiesce)
3337 return 1;
3338 if (list_empty_careful(&conf->inactive_list))
3339 return 1;
3341 return 0;
3344 /* We want read requests to align with chunks where possible,
3345 * but write requests don't need to.
3347 static int raid5_mergeable_bvec(struct request_queue *q,
3348 struct bvec_merge_data *bvm,
3349 struct bio_vec *biovec)
3351 mddev_t *mddev = q->queuedata;
3352 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3353 int max;
3354 unsigned int chunk_sectors = mddev->chunk_sectors;
3355 unsigned int bio_sectors = bvm->bi_size >> 9;
3357 if ((bvm->bi_rw & 1) == WRITE)
3358 return biovec->bv_len; /* always allow writes to be mergeable */
3360 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3361 chunk_sectors = mddev->new_chunk_sectors;
3362 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3363 if (max < 0) max = 0;
3364 if (max <= biovec->bv_len && bio_sectors == 0)
3365 return biovec->bv_len;
3366 else
3367 return max;
3371 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3373 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3374 unsigned int chunk_sectors = mddev->chunk_sectors;
3375 unsigned int bio_sectors = bio->bi_size >> 9;
3377 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3378 chunk_sectors = mddev->new_chunk_sectors;
3379 return chunk_sectors >=
3380 ((sector & (chunk_sectors - 1)) + bio_sectors);
3384 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3385 * later sampled by raid5d.
3387 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3389 unsigned long flags;
3391 spin_lock_irqsave(&conf->device_lock, flags);
3393 bi->bi_next = conf->retry_read_aligned_list;
3394 conf->retry_read_aligned_list = bi;
3396 spin_unlock_irqrestore(&conf->device_lock, flags);
3397 md_wakeup_thread(conf->mddev->thread);
3401 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3403 struct bio *bi;
3405 bi = conf->retry_read_aligned;
3406 if (bi) {
3407 conf->retry_read_aligned = NULL;
3408 return bi;
3410 bi = conf->retry_read_aligned_list;
3411 if(bi) {
3412 conf->retry_read_aligned_list = bi->bi_next;
3413 bi->bi_next = NULL;
3415 * this sets the active strip count to 1 and the processed
3416 * strip count to zero (upper 8 bits)
3418 bi->bi_phys_segments = 1; /* biased count of active stripes */
3421 return bi;
3426 * The "raid5_align_endio" should check if the read succeeded and if it
3427 * did, call bio_endio on the original bio (having bio_put the new bio
3428 * first).
3429 * If the read failed..
3431 static void raid5_align_endio(struct bio *bi, int error)
3433 struct bio* raid_bi = bi->bi_private;
3434 mddev_t *mddev;
3435 raid5_conf_t *conf;
3436 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3437 mdk_rdev_t *rdev;
3439 bio_put(bi);
3441 mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3442 conf = mddev->private;
3443 rdev = (void*)raid_bi->bi_next;
3444 raid_bi->bi_next = NULL;
3446 rdev_dec_pending(rdev, conf->mddev);
3448 if (!error && uptodate) {
3449 bio_endio(raid_bi, 0);
3450 if (atomic_dec_and_test(&conf->active_aligned_reads))
3451 wake_up(&conf->wait_for_stripe);
3452 return;
3456 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3458 add_bio_to_retry(raid_bi, conf);
3461 static int bio_fits_rdev(struct bio *bi)
3463 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3465 if ((bi->bi_size>>9) > queue_max_sectors(q))
3466 return 0;
3467 blk_recount_segments(q, bi);
3468 if (bi->bi_phys_segments > queue_max_phys_segments(q))
3469 return 0;
3471 if (q->merge_bvec_fn)
3472 /* it's too hard to apply the merge_bvec_fn at this stage,
3473 * just just give up
3475 return 0;
3477 return 1;
3481 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3483 mddev_t *mddev = q->queuedata;
3484 raid5_conf_t *conf = mddev->private;
3485 unsigned int dd_idx;
3486 struct bio* align_bi;
3487 mdk_rdev_t *rdev;
3489 if (!in_chunk_boundary(mddev, raid_bio)) {
3490 pr_debug("chunk_aligned_read : non aligned\n");
3491 return 0;
3494 * use bio_clone to make a copy of the bio
3496 align_bi = bio_clone(raid_bio, GFP_NOIO);
3497 if (!align_bi)
3498 return 0;
3500 * set bi_end_io to a new function, and set bi_private to the
3501 * original bio.
3503 align_bi->bi_end_io = raid5_align_endio;
3504 align_bi->bi_private = raid_bio;
3506 * compute position
3508 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3510 &dd_idx, NULL);
3512 rcu_read_lock();
3513 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3514 if (rdev && test_bit(In_sync, &rdev->flags)) {
3515 atomic_inc(&rdev->nr_pending);
3516 rcu_read_unlock();
3517 raid_bio->bi_next = (void*)rdev;
3518 align_bi->bi_bdev = rdev->bdev;
3519 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3520 align_bi->bi_sector += rdev->data_offset;
3522 if (!bio_fits_rdev(align_bi)) {
3523 /* too big in some way */
3524 bio_put(align_bi);
3525 rdev_dec_pending(rdev, mddev);
3526 return 0;
3529 spin_lock_irq(&conf->device_lock);
3530 wait_event_lock_irq(conf->wait_for_stripe,
3531 conf->quiesce == 0,
3532 conf->device_lock, /* nothing */);
3533 atomic_inc(&conf->active_aligned_reads);
3534 spin_unlock_irq(&conf->device_lock);
3536 generic_make_request(align_bi);
3537 return 1;
3538 } else {
3539 rcu_read_unlock();
3540 bio_put(align_bi);
3541 return 0;
3545 /* __get_priority_stripe - get the next stripe to process
3547 * Full stripe writes are allowed to pass preread active stripes up until
3548 * the bypass_threshold is exceeded. In general the bypass_count
3549 * increments when the handle_list is handled before the hold_list; however, it
3550 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3551 * stripe with in flight i/o. The bypass_count will be reset when the
3552 * head of the hold_list has changed, i.e. the head was promoted to the
3553 * handle_list.
3555 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3557 struct stripe_head *sh;
3559 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3560 __func__,
3561 list_empty(&conf->handle_list) ? "empty" : "busy",
3562 list_empty(&conf->hold_list) ? "empty" : "busy",
3563 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3565 if (!list_empty(&conf->handle_list)) {
3566 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3568 if (list_empty(&conf->hold_list))
3569 conf->bypass_count = 0;
3570 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3571 if (conf->hold_list.next == conf->last_hold)
3572 conf->bypass_count++;
3573 else {
3574 conf->last_hold = conf->hold_list.next;
3575 conf->bypass_count -= conf->bypass_threshold;
3576 if (conf->bypass_count < 0)
3577 conf->bypass_count = 0;
3580 } else if (!list_empty(&conf->hold_list) &&
3581 ((conf->bypass_threshold &&
3582 conf->bypass_count > conf->bypass_threshold) ||
3583 atomic_read(&conf->pending_full_writes) == 0)) {
3584 sh = list_entry(conf->hold_list.next,
3585 typeof(*sh), lru);
3586 conf->bypass_count -= conf->bypass_threshold;
3587 if (conf->bypass_count < 0)
3588 conf->bypass_count = 0;
3589 } else
3590 return NULL;
3592 list_del_init(&sh->lru);
3593 atomic_inc(&sh->count);
3594 BUG_ON(atomic_read(&sh->count) != 1);
3595 return sh;
3598 static int make_request(struct request_queue *q, struct bio * bi)
3600 mddev_t *mddev = q->queuedata;
3601 raid5_conf_t *conf = mddev->private;
3602 int dd_idx;
3603 sector_t new_sector;
3604 sector_t logical_sector, last_sector;
3605 struct stripe_head *sh;
3606 const int rw = bio_data_dir(bi);
3607 int cpu, remaining;
3609 if (unlikely(bio_barrier(bi))) {
3610 bio_endio(bi, -EOPNOTSUPP);
3611 return 0;
3614 md_write_start(mddev, bi);
3616 cpu = part_stat_lock();
3617 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3618 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3619 bio_sectors(bi));
3620 part_stat_unlock();
3622 if (rw == READ &&
3623 mddev->reshape_position == MaxSector &&
3624 chunk_aligned_read(q,bi))
3625 return 0;
3627 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3628 last_sector = bi->bi_sector + (bi->bi_size>>9);
3629 bi->bi_next = NULL;
3630 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
3632 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3633 DEFINE_WAIT(w);
3634 int disks, data_disks;
3635 int previous;
3637 retry:
3638 previous = 0;
3639 disks = conf->raid_disks;
3640 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3641 if (unlikely(conf->reshape_progress != MaxSector)) {
3642 /* spinlock is needed as reshape_progress may be
3643 * 64bit on a 32bit platform, and so it might be
3644 * possible to see a half-updated value
3645 * Ofcourse reshape_progress could change after
3646 * the lock is dropped, so once we get a reference
3647 * to the stripe that we think it is, we will have
3648 * to check again.
3650 spin_lock_irq(&conf->device_lock);
3651 if (mddev->delta_disks < 0
3652 ? logical_sector < conf->reshape_progress
3653 : logical_sector >= conf->reshape_progress) {
3654 disks = conf->previous_raid_disks;
3655 previous = 1;
3656 } else {
3657 if (mddev->delta_disks < 0
3658 ? logical_sector < conf->reshape_safe
3659 : logical_sector >= conf->reshape_safe) {
3660 spin_unlock_irq(&conf->device_lock);
3661 schedule();
3662 goto retry;
3665 spin_unlock_irq(&conf->device_lock);
3667 data_disks = disks - conf->max_degraded;
3669 new_sector = raid5_compute_sector(conf, logical_sector,
3670 previous,
3671 &dd_idx, NULL);
3672 pr_debug("raid5: make_request, sector %llu logical %llu\n",
3673 (unsigned long long)new_sector,
3674 (unsigned long long)logical_sector);
3676 sh = get_active_stripe(conf, new_sector, previous,
3677 (bi->bi_rw&RWA_MASK), 0);
3678 if (sh) {
3679 if (unlikely(previous)) {
3680 /* expansion might have moved on while waiting for a
3681 * stripe, so we must do the range check again.
3682 * Expansion could still move past after this
3683 * test, but as we are holding a reference to
3684 * 'sh', we know that if that happens,
3685 * STRIPE_EXPANDING will get set and the expansion
3686 * won't proceed until we finish with the stripe.
3688 int must_retry = 0;
3689 spin_lock_irq(&conf->device_lock);
3690 if (mddev->delta_disks < 0
3691 ? logical_sector >= conf->reshape_progress
3692 : logical_sector < conf->reshape_progress)
3693 /* mismatch, need to try again */
3694 must_retry = 1;
3695 spin_unlock_irq(&conf->device_lock);
3696 if (must_retry) {
3697 release_stripe(sh);
3698 schedule();
3699 goto retry;
3703 if (bio_data_dir(bi) == WRITE &&
3704 logical_sector >= mddev->suspend_lo &&
3705 logical_sector < mddev->suspend_hi) {
3706 release_stripe(sh);
3707 /* As the suspend_* range is controlled by
3708 * userspace, we want an interruptible
3709 * wait.
3711 flush_signals(current);
3712 prepare_to_wait(&conf->wait_for_overlap,
3713 &w, TASK_INTERRUPTIBLE);
3714 if (logical_sector >= mddev->suspend_lo &&
3715 logical_sector < mddev->suspend_hi)
3716 schedule();
3717 goto retry;
3720 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3721 !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3722 /* Stripe is busy expanding or
3723 * add failed due to overlap. Flush everything
3724 * and wait a while
3726 raid5_unplug_device(mddev->queue);
3727 release_stripe(sh);
3728 schedule();
3729 goto retry;
3731 finish_wait(&conf->wait_for_overlap, &w);
3732 set_bit(STRIPE_HANDLE, &sh->state);
3733 clear_bit(STRIPE_DELAYED, &sh->state);
3734 release_stripe(sh);
3735 } else {
3736 /* cannot get stripe for read-ahead, just give-up */
3737 clear_bit(BIO_UPTODATE, &bi->bi_flags);
3738 finish_wait(&conf->wait_for_overlap, &w);
3739 break;
3743 spin_lock_irq(&conf->device_lock);
3744 remaining = raid5_dec_bi_phys_segments(bi);
3745 spin_unlock_irq(&conf->device_lock);
3746 if (remaining == 0) {
3748 if ( rw == WRITE )
3749 md_write_end(mddev);
3751 bio_endio(bi, 0);
3753 return 0;
3756 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3758 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3760 /* reshaping is quite different to recovery/resync so it is
3761 * handled quite separately ... here.
3763 * On each call to sync_request, we gather one chunk worth of
3764 * destination stripes and flag them as expanding.
3765 * Then we find all the source stripes and request reads.
3766 * As the reads complete, handle_stripe will copy the data
3767 * into the destination stripe and release that stripe.
3769 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3770 struct stripe_head *sh;
3771 sector_t first_sector, last_sector;
3772 int raid_disks = conf->previous_raid_disks;
3773 int data_disks = raid_disks - conf->max_degraded;
3774 int new_data_disks = conf->raid_disks - conf->max_degraded;
3775 int i;
3776 int dd_idx;
3777 sector_t writepos, readpos, safepos;
3778 sector_t stripe_addr;
3779 int reshape_sectors;
3780 struct list_head stripes;
3782 if (sector_nr == 0) {
3783 /* If restarting in the middle, skip the initial sectors */
3784 if (mddev->delta_disks < 0 &&
3785 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3786 sector_nr = raid5_size(mddev, 0, 0)
3787 - conf->reshape_progress;
3788 } else if (mddev->delta_disks > 0 &&
3789 conf->reshape_progress > 0)
3790 sector_nr = conf->reshape_progress;
3791 sector_div(sector_nr, new_data_disks);
3792 if (sector_nr) {
3793 *skipped = 1;
3794 return sector_nr;
3798 /* We need to process a full chunk at a time.
3799 * If old and new chunk sizes differ, we need to process the
3800 * largest of these
3802 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3803 reshape_sectors = mddev->new_chunk_sectors;
3804 else
3805 reshape_sectors = mddev->chunk_sectors;
3807 /* we update the metadata when there is more than 3Meg
3808 * in the block range (that is rather arbitrary, should
3809 * probably be time based) or when the data about to be
3810 * copied would over-write the source of the data at
3811 * the front of the range.
3812 * i.e. one new_stripe along from reshape_progress new_maps
3813 * to after where reshape_safe old_maps to
3815 writepos = conf->reshape_progress;
3816 sector_div(writepos, new_data_disks);
3817 readpos = conf->reshape_progress;
3818 sector_div(readpos, data_disks);
3819 safepos = conf->reshape_safe;
3820 sector_div(safepos, data_disks);
3821 if (mddev->delta_disks < 0) {
3822 writepos -= min_t(sector_t, reshape_sectors, writepos);
3823 readpos += reshape_sectors;
3824 safepos += reshape_sectors;
3825 } else {
3826 writepos += reshape_sectors;
3827 readpos -= min_t(sector_t, reshape_sectors, readpos);
3828 safepos -= min_t(sector_t, reshape_sectors, safepos);
3831 /* 'writepos' is the most advanced device address we might write.
3832 * 'readpos' is the least advanced device address we might read.
3833 * 'safepos' is the least address recorded in the metadata as having
3834 * been reshaped.
3835 * If 'readpos' is behind 'writepos', then there is no way that we can
3836 * ensure safety in the face of a crash - that must be done by userspace
3837 * making a backup of the data. So in that case there is no particular
3838 * rush to update metadata.
3839 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3840 * update the metadata to advance 'safepos' to match 'readpos' so that
3841 * we can be safe in the event of a crash.
3842 * So we insist on updating metadata if safepos is behind writepos and
3843 * readpos is beyond writepos.
3844 * In any case, update the metadata every 10 seconds.
3845 * Maybe that number should be configurable, but I'm not sure it is
3846 * worth it.... maybe it could be a multiple of safemode_delay???
3848 if ((mddev->delta_disks < 0
3849 ? (safepos > writepos && readpos < writepos)
3850 : (safepos < writepos && readpos > writepos)) ||
3851 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3852 /* Cannot proceed until we've updated the superblock... */
3853 wait_event(conf->wait_for_overlap,
3854 atomic_read(&conf->reshape_stripes)==0);
3855 mddev->reshape_position = conf->reshape_progress;
3856 mddev->curr_resync_completed = mddev->curr_resync;
3857 conf->reshape_checkpoint = jiffies;
3858 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3859 md_wakeup_thread(mddev->thread);
3860 wait_event(mddev->sb_wait, mddev->flags == 0 ||
3861 kthread_should_stop());
3862 spin_lock_irq(&conf->device_lock);
3863 conf->reshape_safe = mddev->reshape_position;
3864 spin_unlock_irq(&conf->device_lock);
3865 wake_up(&conf->wait_for_overlap);
3866 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3869 if (mddev->delta_disks < 0) {
3870 BUG_ON(conf->reshape_progress == 0);
3871 stripe_addr = writepos;
3872 BUG_ON((mddev->dev_sectors &
3873 ~((sector_t)reshape_sectors - 1))
3874 - reshape_sectors - stripe_addr
3875 != sector_nr);
3876 } else {
3877 BUG_ON(writepos != sector_nr + reshape_sectors);
3878 stripe_addr = sector_nr;
3880 INIT_LIST_HEAD(&stripes);
3881 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3882 int j;
3883 int skipped = 0;
3884 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
3885 set_bit(STRIPE_EXPANDING, &sh->state);
3886 atomic_inc(&conf->reshape_stripes);
3887 /* If any of this stripe is beyond the end of the old
3888 * array, then we need to zero those blocks
3890 for (j=sh->disks; j--;) {
3891 sector_t s;
3892 if (j == sh->pd_idx)
3893 continue;
3894 if (conf->level == 6 &&
3895 j == sh->qd_idx)
3896 continue;
3897 s = compute_blocknr(sh, j, 0);
3898 if (s < raid5_size(mddev, 0, 0)) {
3899 skipped = 1;
3900 continue;
3902 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3903 set_bit(R5_Expanded, &sh->dev[j].flags);
3904 set_bit(R5_UPTODATE, &sh->dev[j].flags);
3906 if (!skipped) {
3907 set_bit(STRIPE_EXPAND_READY, &sh->state);
3908 set_bit(STRIPE_HANDLE, &sh->state);
3910 list_add(&sh->lru, &stripes);
3912 spin_lock_irq(&conf->device_lock);
3913 if (mddev->delta_disks < 0)
3914 conf->reshape_progress -= reshape_sectors * new_data_disks;
3915 else
3916 conf->reshape_progress += reshape_sectors * new_data_disks;
3917 spin_unlock_irq(&conf->device_lock);
3918 /* Ok, those stripe are ready. We can start scheduling
3919 * reads on the source stripes.
3920 * The source stripes are determined by mapping the first and last
3921 * block on the destination stripes.
3923 first_sector =
3924 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
3925 1, &dd_idx, NULL);
3926 last_sector =
3927 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
3928 * new_data_disks - 1),
3929 1, &dd_idx, NULL);
3930 if (last_sector >= mddev->dev_sectors)
3931 last_sector = mddev->dev_sectors - 1;
3932 while (first_sector <= last_sector) {
3933 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
3934 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3935 set_bit(STRIPE_HANDLE, &sh->state);
3936 release_stripe(sh);
3937 first_sector += STRIPE_SECTORS;
3939 /* Now that the sources are clearly marked, we can release
3940 * the destination stripes
3942 while (!list_empty(&stripes)) {
3943 sh = list_entry(stripes.next, struct stripe_head, lru);
3944 list_del_init(&sh->lru);
3945 release_stripe(sh);
3947 /* If this takes us to the resync_max point where we have to pause,
3948 * then we need to write out the superblock.
3950 sector_nr += reshape_sectors;
3951 if ((sector_nr - mddev->curr_resync_completed) * 2
3952 >= mddev->resync_max - mddev->curr_resync_completed) {
3953 /* Cannot proceed until we've updated the superblock... */
3954 wait_event(conf->wait_for_overlap,
3955 atomic_read(&conf->reshape_stripes) == 0);
3956 mddev->reshape_position = conf->reshape_progress;
3957 mddev->curr_resync_completed = mddev->curr_resync + reshape_sectors;
3958 conf->reshape_checkpoint = jiffies;
3959 set_bit(MD_CHANGE_DEVS, &mddev->flags);
3960 md_wakeup_thread(mddev->thread);
3961 wait_event(mddev->sb_wait,
3962 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3963 || kthread_should_stop());
3964 spin_lock_irq(&conf->device_lock);
3965 conf->reshape_safe = mddev->reshape_position;
3966 spin_unlock_irq(&conf->device_lock);
3967 wake_up(&conf->wait_for_overlap);
3968 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3970 return reshape_sectors;
3973 /* FIXME go_faster isn't used */
3974 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3976 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3977 struct stripe_head *sh;
3978 sector_t max_sector = mddev->dev_sectors;
3979 int sync_blocks;
3980 int still_degraded = 0;
3981 int i;
3983 if (sector_nr >= max_sector) {
3984 /* just being told to finish up .. nothing much to do */
3985 unplug_slaves(mddev);
3987 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3988 end_reshape(conf);
3989 return 0;
3992 if (mddev->curr_resync < max_sector) /* aborted */
3993 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3994 &sync_blocks, 1);
3995 else /* completed sync */
3996 conf->fullsync = 0;
3997 bitmap_close_sync(mddev->bitmap);
3999 return 0;
4002 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4003 return reshape_request(mddev, sector_nr, skipped);
4005 /* No need to check resync_max as we never do more than one
4006 * stripe, and as resync_max will always be on a chunk boundary,
4007 * if the check in md_do_sync didn't fire, there is no chance
4008 * of overstepping resync_max here
4011 /* if there is too many failed drives and we are trying
4012 * to resync, then assert that we are finished, because there is
4013 * nothing we can do.
4015 if (mddev->degraded >= conf->max_degraded &&
4016 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4017 sector_t rv = mddev->dev_sectors - sector_nr;
4018 *skipped = 1;
4019 return rv;
4021 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4022 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4023 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4024 /* we can skip this block, and probably more */
4025 sync_blocks /= STRIPE_SECTORS;
4026 *skipped = 1;
4027 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4031 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4033 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4034 if (sh == NULL) {
4035 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4036 /* make sure we don't swamp the stripe cache if someone else
4037 * is trying to get access
4039 schedule_timeout_uninterruptible(1);
4041 /* Need to check if array will still be degraded after recovery/resync
4042 * We don't need to check the 'failed' flag as when that gets set,
4043 * recovery aborts.
4045 for (i = 0; i < conf->raid_disks; i++)
4046 if (conf->disks[i].rdev == NULL)
4047 still_degraded = 1;
4049 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4051 spin_lock(&sh->lock);
4052 set_bit(STRIPE_SYNCING, &sh->state);
4053 clear_bit(STRIPE_INSYNC, &sh->state);
4054 spin_unlock(&sh->lock);
4056 /* wait for any blocked device to be handled */
4057 while(unlikely(!handle_stripe(sh, NULL)))
4059 release_stripe(sh);
4061 return STRIPE_SECTORS;
4064 static int retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4066 /* We may not be able to submit a whole bio at once as there
4067 * may not be enough stripe_heads available.
4068 * We cannot pre-allocate enough stripe_heads as we may need
4069 * more than exist in the cache (if we allow ever large chunks).
4070 * So we do one stripe head at a time and record in
4071 * ->bi_hw_segments how many have been done.
4073 * We *know* that this entire raid_bio is in one chunk, so
4074 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4076 struct stripe_head *sh;
4077 int dd_idx;
4078 sector_t sector, logical_sector, last_sector;
4079 int scnt = 0;
4080 int remaining;
4081 int handled = 0;
4083 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4084 sector = raid5_compute_sector(conf, logical_sector,
4085 0, &dd_idx, NULL);
4086 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4088 for (; logical_sector < last_sector;
4089 logical_sector += STRIPE_SECTORS,
4090 sector += STRIPE_SECTORS,
4091 scnt++) {
4093 if (scnt < raid5_bi_hw_segments(raid_bio))
4094 /* already done this stripe */
4095 continue;
4097 sh = get_active_stripe(conf, sector, 0, 1, 0);
4099 if (!sh) {
4100 /* failed to get a stripe - must wait */
4101 raid5_set_bi_hw_segments(raid_bio, scnt);
4102 conf->retry_read_aligned = raid_bio;
4103 return handled;
4106 set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4107 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4108 release_stripe(sh);
4109 raid5_set_bi_hw_segments(raid_bio, scnt);
4110 conf->retry_read_aligned = raid_bio;
4111 return handled;
4114 handle_stripe(sh, NULL);
4115 release_stripe(sh);
4116 handled++;
4118 spin_lock_irq(&conf->device_lock);
4119 remaining = raid5_dec_bi_phys_segments(raid_bio);
4120 spin_unlock_irq(&conf->device_lock);
4121 if (remaining == 0)
4122 bio_endio(raid_bio, 0);
4123 if (atomic_dec_and_test(&conf->active_aligned_reads))
4124 wake_up(&conf->wait_for_stripe);
4125 return handled;
4131 * This is our raid5 kernel thread.
4133 * We scan the hash table for stripes which can be handled now.
4134 * During the scan, completed stripes are saved for us by the interrupt
4135 * handler, so that they will not have to wait for our next wakeup.
4137 static void raid5d(mddev_t *mddev)
4139 struct stripe_head *sh;
4140 raid5_conf_t *conf = mddev->private;
4141 int handled;
4143 pr_debug("+++ raid5d active\n");
4145 md_check_recovery(mddev);
4147 handled = 0;
4148 spin_lock_irq(&conf->device_lock);
4149 while (1) {
4150 struct bio *bio;
4152 if (conf->seq_flush != conf->seq_write) {
4153 int seq = conf->seq_flush;
4154 spin_unlock_irq(&conf->device_lock);
4155 bitmap_unplug(mddev->bitmap);
4156 spin_lock_irq(&conf->device_lock);
4157 conf->seq_write = seq;
4158 activate_bit_delay(conf);
4161 while ((bio = remove_bio_from_retry(conf))) {
4162 int ok;
4163 spin_unlock_irq(&conf->device_lock);
4164 ok = retry_aligned_read(conf, bio);
4165 spin_lock_irq(&conf->device_lock);
4166 if (!ok)
4167 break;
4168 handled++;
4171 sh = __get_priority_stripe(conf);
4173 if (!sh)
4174 break;
4175 spin_unlock_irq(&conf->device_lock);
4177 handled++;
4178 handle_stripe(sh, conf->spare_page);
4179 release_stripe(sh);
4181 spin_lock_irq(&conf->device_lock);
4183 pr_debug("%d stripes handled\n", handled);
4185 spin_unlock_irq(&conf->device_lock);
4187 async_tx_issue_pending_all();
4188 unplug_slaves(mddev);
4190 pr_debug("--- raid5d inactive\n");
4193 static ssize_t
4194 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4196 raid5_conf_t *conf = mddev->private;
4197 if (conf)
4198 return sprintf(page, "%d\n", conf->max_nr_stripes);
4199 else
4200 return 0;
4203 static ssize_t
4204 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4206 raid5_conf_t *conf = mddev->private;
4207 unsigned long new;
4208 int err;
4210 if (len >= PAGE_SIZE)
4211 return -EINVAL;
4212 if (!conf)
4213 return -ENODEV;
4215 if (strict_strtoul(page, 10, &new))
4216 return -EINVAL;
4217 if (new <= 16 || new > 32768)
4218 return -EINVAL;
4219 while (new < conf->max_nr_stripes) {
4220 if (drop_one_stripe(conf))
4221 conf->max_nr_stripes--;
4222 else
4223 break;
4225 err = md_allow_write(mddev);
4226 if (err)
4227 return err;
4228 while (new > conf->max_nr_stripes) {
4229 if (grow_one_stripe(conf))
4230 conf->max_nr_stripes++;
4231 else break;
4233 return len;
4236 static struct md_sysfs_entry
4237 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4238 raid5_show_stripe_cache_size,
4239 raid5_store_stripe_cache_size);
4241 static ssize_t
4242 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4244 raid5_conf_t *conf = mddev->private;
4245 if (conf)
4246 return sprintf(page, "%d\n", conf->bypass_threshold);
4247 else
4248 return 0;
4251 static ssize_t
4252 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4254 raid5_conf_t *conf = mddev->private;
4255 unsigned long new;
4256 if (len >= PAGE_SIZE)
4257 return -EINVAL;
4258 if (!conf)
4259 return -ENODEV;
4261 if (strict_strtoul(page, 10, &new))
4262 return -EINVAL;
4263 if (new > conf->max_nr_stripes)
4264 return -EINVAL;
4265 conf->bypass_threshold = new;
4266 return len;
4269 static struct md_sysfs_entry
4270 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4271 S_IRUGO | S_IWUSR,
4272 raid5_show_preread_threshold,
4273 raid5_store_preread_threshold);
4275 static ssize_t
4276 stripe_cache_active_show(mddev_t *mddev, char *page)
4278 raid5_conf_t *conf = mddev->private;
4279 if (conf)
4280 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4281 else
4282 return 0;
4285 static struct md_sysfs_entry
4286 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4288 static struct attribute *raid5_attrs[] = {
4289 &raid5_stripecache_size.attr,
4290 &raid5_stripecache_active.attr,
4291 &raid5_preread_bypass_threshold.attr,
4292 NULL,
4294 static struct attribute_group raid5_attrs_group = {
4295 .name = NULL,
4296 .attrs = raid5_attrs,
4299 static sector_t
4300 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4302 raid5_conf_t *conf = mddev->private;
4304 if (!sectors)
4305 sectors = mddev->dev_sectors;
4306 if (!raid_disks) {
4307 /* size is defined by the smallest of previous and new size */
4308 if (conf->raid_disks < conf->previous_raid_disks)
4309 raid_disks = conf->raid_disks;
4310 else
4311 raid_disks = conf->previous_raid_disks;
4314 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4315 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4316 return sectors * (raid_disks - conf->max_degraded);
4319 static raid5_conf_t *setup_conf(mddev_t *mddev)
4321 raid5_conf_t *conf;
4322 int raid_disk, memory;
4323 mdk_rdev_t *rdev;
4324 struct disk_info *disk;
4326 if (mddev->new_level != 5
4327 && mddev->new_level != 4
4328 && mddev->new_level != 6) {
4329 printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4330 mdname(mddev), mddev->new_level);
4331 return ERR_PTR(-EIO);
4333 if ((mddev->new_level == 5
4334 && !algorithm_valid_raid5(mddev->new_layout)) ||
4335 (mddev->new_level == 6
4336 && !algorithm_valid_raid6(mddev->new_layout))) {
4337 printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4338 mdname(mddev), mddev->new_layout);
4339 return ERR_PTR(-EIO);
4341 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4342 printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4343 mdname(mddev), mddev->raid_disks);
4344 return ERR_PTR(-EINVAL);
4347 if (!mddev->new_chunk_sectors ||
4348 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4349 !is_power_of_2(mddev->new_chunk_sectors)) {
4350 printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4351 mddev->new_chunk_sectors << 9, mdname(mddev));
4352 return ERR_PTR(-EINVAL);
4355 conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4356 if (conf == NULL)
4357 goto abort;
4359 conf->raid_disks = mddev->raid_disks;
4360 if (mddev->reshape_position == MaxSector)
4361 conf->previous_raid_disks = mddev->raid_disks;
4362 else
4363 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4365 conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4366 GFP_KERNEL);
4367 if (!conf->disks)
4368 goto abort;
4370 conf->mddev = mddev;
4372 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4373 goto abort;
4375 if (mddev->new_level == 6) {
4376 conf->spare_page = alloc_page(GFP_KERNEL);
4377 if (!conf->spare_page)
4378 goto abort;
4380 spin_lock_init(&conf->device_lock);
4381 init_waitqueue_head(&conf->wait_for_stripe);
4382 init_waitqueue_head(&conf->wait_for_overlap);
4383 INIT_LIST_HEAD(&conf->handle_list);
4384 INIT_LIST_HEAD(&conf->hold_list);
4385 INIT_LIST_HEAD(&conf->delayed_list);
4386 INIT_LIST_HEAD(&conf->bitmap_list);
4387 INIT_LIST_HEAD(&conf->inactive_list);
4388 atomic_set(&conf->active_stripes, 0);
4389 atomic_set(&conf->preread_active_stripes, 0);
4390 atomic_set(&conf->active_aligned_reads, 0);
4391 conf->bypass_threshold = BYPASS_THRESHOLD;
4393 pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4395 list_for_each_entry(rdev, &mddev->disks, same_set) {
4396 raid_disk = rdev->raid_disk;
4397 if (raid_disk >= conf->raid_disks
4398 || raid_disk < 0)
4399 continue;
4400 disk = conf->disks + raid_disk;
4402 disk->rdev = rdev;
4404 if (test_bit(In_sync, &rdev->flags)) {
4405 char b[BDEVNAME_SIZE];
4406 printk(KERN_INFO "raid5: device %s operational as raid"
4407 " disk %d\n", bdevname(rdev->bdev,b),
4408 raid_disk);
4409 } else
4410 /* Cannot rely on bitmap to complete recovery */
4411 conf->fullsync = 1;
4414 conf->chunk_sectors = mddev->new_chunk_sectors;
4415 conf->level = mddev->new_level;
4416 if (conf->level == 6)
4417 conf->max_degraded = 2;
4418 else
4419 conf->max_degraded = 1;
4420 conf->algorithm = mddev->new_layout;
4421 conf->max_nr_stripes = NR_STRIPES;
4422 conf->reshape_progress = mddev->reshape_position;
4423 if (conf->reshape_progress != MaxSector) {
4424 conf->prev_chunk_sectors = mddev->chunk_sectors;
4425 conf->prev_algo = mddev->layout;
4428 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4429 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4430 if (grow_stripes(conf, conf->max_nr_stripes)) {
4431 printk(KERN_ERR
4432 "raid5: couldn't allocate %dkB for buffers\n", memory);
4433 goto abort;
4434 } else
4435 printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4436 memory, mdname(mddev));
4438 conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4439 if (!conf->thread) {
4440 printk(KERN_ERR
4441 "raid5: couldn't allocate thread for %s\n",
4442 mdname(mddev));
4443 goto abort;
4446 return conf;
4448 abort:
4449 if (conf) {
4450 shrink_stripes(conf);
4451 safe_put_page(conf->spare_page);
4452 kfree(conf->disks);
4453 kfree(conf->stripe_hashtbl);
4454 kfree(conf);
4455 return ERR_PTR(-EIO);
4456 } else
4457 return ERR_PTR(-ENOMEM);
4460 static int run(mddev_t *mddev)
4462 raid5_conf_t *conf;
4463 int working_disks = 0, chunk_size;
4464 mdk_rdev_t *rdev;
4466 if (mddev->recovery_cp != MaxSector)
4467 printk(KERN_NOTICE "raid5: %s is not clean"
4468 " -- starting background reconstruction\n",
4469 mdname(mddev));
4470 if (mddev->reshape_position != MaxSector) {
4471 /* Check that we can continue the reshape.
4472 * Currently only disks can change, it must
4473 * increase, and we must be past the point where
4474 * a stripe over-writes itself
4476 sector_t here_new, here_old;
4477 int old_disks;
4478 int max_degraded = (mddev->level == 6 ? 2 : 1);
4480 if (mddev->new_level != mddev->level) {
4481 printk(KERN_ERR "raid5: %s: unsupported reshape "
4482 "required - aborting.\n",
4483 mdname(mddev));
4484 return -EINVAL;
4486 old_disks = mddev->raid_disks - mddev->delta_disks;
4487 /* reshape_position must be on a new-stripe boundary, and one
4488 * further up in new geometry must map after here in old
4489 * geometry.
4491 here_new = mddev->reshape_position;
4492 if (sector_div(here_new, mddev->new_chunk_sectors *
4493 (mddev->raid_disks - max_degraded))) {
4494 printk(KERN_ERR "raid5: reshape_position not "
4495 "on a stripe boundary\n");
4496 return -EINVAL;
4498 /* here_new is the stripe we will write to */
4499 here_old = mddev->reshape_position;
4500 sector_div(here_old, mddev->chunk_sectors *
4501 (old_disks-max_degraded));
4502 /* here_old is the first stripe that we might need to read
4503 * from */
4504 if (here_new >= here_old) {
4505 /* Reading from the same stripe as writing to - bad */
4506 printk(KERN_ERR "raid5: reshape_position too early for "
4507 "auto-recovery - aborting.\n");
4508 return -EINVAL;
4510 printk(KERN_INFO "raid5: reshape will continue\n");
4511 /* OK, we should be able to continue; */
4512 } else {
4513 BUG_ON(mddev->level != mddev->new_level);
4514 BUG_ON(mddev->layout != mddev->new_layout);
4515 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4516 BUG_ON(mddev->delta_disks != 0);
4519 if (mddev->private == NULL)
4520 conf = setup_conf(mddev);
4521 else
4522 conf = mddev->private;
4524 if (IS_ERR(conf))
4525 return PTR_ERR(conf);
4527 mddev->thread = conf->thread;
4528 conf->thread = NULL;
4529 mddev->private = conf;
4532 * 0 for a fully functional array, 1 or 2 for a degraded array.
4534 list_for_each_entry(rdev, &mddev->disks, same_set)
4535 if (rdev->raid_disk >= 0 &&
4536 test_bit(In_sync, &rdev->flags))
4537 working_disks++;
4539 mddev->degraded = conf->raid_disks - working_disks;
4541 if (mddev->degraded > conf->max_degraded) {
4542 printk(KERN_ERR "raid5: not enough operational devices for %s"
4543 " (%d/%d failed)\n",
4544 mdname(mddev), mddev->degraded, conf->raid_disks);
4545 goto abort;
4548 /* device size must be a multiple of chunk size */
4549 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4550 mddev->resync_max_sectors = mddev->dev_sectors;
4552 if (mddev->degraded > 0 &&
4553 mddev->recovery_cp != MaxSector) {
4554 if (mddev->ok_start_degraded)
4555 printk(KERN_WARNING
4556 "raid5: starting dirty degraded array: %s"
4557 "- data corruption possible.\n",
4558 mdname(mddev));
4559 else {
4560 printk(KERN_ERR
4561 "raid5: cannot start dirty degraded array for %s\n",
4562 mdname(mddev));
4563 goto abort;
4567 if (mddev->degraded == 0)
4568 printk("raid5: raid level %d set %s active with %d out of %d"
4569 " devices, algorithm %d\n", conf->level, mdname(mddev),
4570 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4571 mddev->new_layout);
4572 else
4573 printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4574 " out of %d devices, algorithm %d\n", conf->level,
4575 mdname(mddev), mddev->raid_disks - mddev->degraded,
4576 mddev->raid_disks, mddev->new_layout);
4578 print_raid5_conf(conf);
4580 if (conf->reshape_progress != MaxSector) {
4581 printk("...ok start reshape thread\n");
4582 conf->reshape_safe = conf->reshape_progress;
4583 atomic_set(&conf->reshape_stripes, 0);
4584 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4585 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4586 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4587 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4588 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4589 "%s_reshape");
4592 /* read-ahead size must cover two whole stripes, which is
4593 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4596 int data_disks = conf->previous_raid_disks - conf->max_degraded;
4597 int stripe = data_disks *
4598 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
4599 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4600 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4603 /* Ok, everything is just fine now */
4604 if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4605 printk(KERN_WARNING
4606 "raid5: failed to create sysfs attributes for %s\n",
4607 mdname(mddev));
4609 mddev->queue->queue_lock = &conf->device_lock;
4611 mddev->queue->unplug_fn = raid5_unplug_device;
4612 mddev->queue->backing_dev_info.congested_data = mddev;
4613 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4615 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4617 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4618 chunk_size = mddev->chunk_sectors << 9;
4619 blk_queue_io_min(mddev->queue, chunk_size);
4620 blk_queue_io_opt(mddev->queue, chunk_size *
4621 (conf->raid_disks - conf->max_degraded));
4623 list_for_each_entry(rdev, &mddev->disks, same_set)
4624 disk_stack_limits(mddev->gendisk, rdev->bdev,
4625 rdev->data_offset << 9);
4627 return 0;
4628 abort:
4629 md_unregister_thread(mddev->thread);
4630 mddev->thread = NULL;
4631 if (conf) {
4632 shrink_stripes(conf);
4633 print_raid5_conf(conf);
4634 safe_put_page(conf->spare_page);
4635 kfree(conf->disks);
4636 kfree(conf->stripe_hashtbl);
4637 kfree(conf);
4639 mddev->private = NULL;
4640 printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4641 return -EIO;
4646 static int stop(mddev_t *mddev)
4648 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4650 md_unregister_thread(mddev->thread);
4651 mddev->thread = NULL;
4652 shrink_stripes(conf);
4653 kfree(conf->stripe_hashtbl);
4654 mddev->queue->backing_dev_info.congested_fn = NULL;
4655 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4656 sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4657 kfree(conf->disks);
4658 kfree(conf);
4659 mddev->private = NULL;
4660 return 0;
4663 #ifdef DEBUG
4664 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
4666 int i;
4668 seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4669 (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4670 seq_printf(seq, "sh %llu, count %d.\n",
4671 (unsigned long long)sh->sector, atomic_read(&sh->count));
4672 seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4673 for (i = 0; i < sh->disks; i++) {
4674 seq_printf(seq, "(cache%d: %p %ld) ",
4675 i, sh->dev[i].page, sh->dev[i].flags);
4677 seq_printf(seq, "\n");
4680 static void printall(struct seq_file *seq, raid5_conf_t *conf)
4682 struct stripe_head *sh;
4683 struct hlist_node *hn;
4684 int i;
4686 spin_lock_irq(&conf->device_lock);
4687 for (i = 0; i < NR_HASH; i++) {
4688 hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4689 if (sh->raid_conf != conf)
4690 continue;
4691 print_sh(seq, sh);
4694 spin_unlock_irq(&conf->device_lock);
4696 #endif
4698 static void status(struct seq_file *seq, mddev_t *mddev)
4700 raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4701 int i;
4703 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
4704 mddev->chunk_sectors / 2, mddev->layout);
4705 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4706 for (i = 0; i < conf->raid_disks; i++)
4707 seq_printf (seq, "%s",
4708 conf->disks[i].rdev &&
4709 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4710 seq_printf (seq, "]");
4711 #ifdef DEBUG
4712 seq_printf (seq, "\n");
4713 printall(seq, conf);
4714 #endif
4717 static void print_raid5_conf (raid5_conf_t *conf)
4719 int i;
4720 struct disk_info *tmp;
4722 printk("RAID5 conf printout:\n");
4723 if (!conf) {
4724 printk("(conf==NULL)\n");
4725 return;
4727 printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4728 conf->raid_disks - conf->mddev->degraded);
4730 for (i = 0; i < conf->raid_disks; i++) {
4731 char b[BDEVNAME_SIZE];
4732 tmp = conf->disks + i;
4733 if (tmp->rdev)
4734 printk(" disk %d, o:%d, dev:%s\n",
4735 i, !test_bit(Faulty, &tmp->rdev->flags),
4736 bdevname(tmp->rdev->bdev,b));
4740 static int raid5_spare_active(mddev_t *mddev)
4742 int i;
4743 raid5_conf_t *conf = mddev->private;
4744 struct disk_info *tmp;
4746 for (i = 0; i < conf->raid_disks; i++) {
4747 tmp = conf->disks + i;
4748 if (tmp->rdev
4749 && !test_bit(Faulty, &tmp->rdev->flags)
4750 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4751 unsigned long flags;
4752 spin_lock_irqsave(&conf->device_lock, flags);
4753 mddev->degraded--;
4754 spin_unlock_irqrestore(&conf->device_lock, flags);
4757 print_raid5_conf(conf);
4758 return 0;
4761 static int raid5_remove_disk(mddev_t *mddev, int number)
4763 raid5_conf_t *conf = mddev->private;
4764 int err = 0;
4765 mdk_rdev_t *rdev;
4766 struct disk_info *p = conf->disks + number;
4768 print_raid5_conf(conf);
4769 rdev = p->rdev;
4770 if (rdev) {
4771 if (number >= conf->raid_disks &&
4772 conf->reshape_progress == MaxSector)
4773 clear_bit(In_sync, &rdev->flags);
4775 if (test_bit(In_sync, &rdev->flags) ||
4776 atomic_read(&rdev->nr_pending)) {
4777 err = -EBUSY;
4778 goto abort;
4780 /* Only remove non-faulty devices if recovery
4781 * isn't possible.
4783 if (!test_bit(Faulty, &rdev->flags) &&
4784 mddev->degraded <= conf->max_degraded &&
4785 number < conf->raid_disks) {
4786 err = -EBUSY;
4787 goto abort;
4789 p->rdev = NULL;
4790 synchronize_rcu();
4791 if (atomic_read(&rdev->nr_pending)) {
4792 /* lost the race, try later */
4793 err = -EBUSY;
4794 p->rdev = rdev;
4797 abort:
4799 print_raid5_conf(conf);
4800 return err;
4803 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4805 raid5_conf_t *conf = mddev->private;
4806 int err = -EEXIST;
4807 int disk;
4808 struct disk_info *p;
4809 int first = 0;
4810 int last = conf->raid_disks - 1;
4812 if (mddev->degraded > conf->max_degraded)
4813 /* no point adding a device */
4814 return -EINVAL;
4816 if (rdev->raid_disk >= 0)
4817 first = last = rdev->raid_disk;
4820 * find the disk ... but prefer rdev->saved_raid_disk
4821 * if possible.
4823 if (rdev->saved_raid_disk >= 0 &&
4824 rdev->saved_raid_disk >= first &&
4825 conf->disks[rdev->saved_raid_disk].rdev == NULL)
4826 disk = rdev->saved_raid_disk;
4827 else
4828 disk = first;
4829 for ( ; disk <= last ; disk++)
4830 if ((p=conf->disks + disk)->rdev == NULL) {
4831 clear_bit(In_sync, &rdev->flags);
4832 rdev->raid_disk = disk;
4833 err = 0;
4834 if (rdev->saved_raid_disk != disk)
4835 conf->fullsync = 1;
4836 rcu_assign_pointer(p->rdev, rdev);
4837 break;
4839 print_raid5_conf(conf);
4840 return err;
4843 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4845 /* no resync is happening, and there is enough space
4846 * on all devices, so we can resize.
4847 * We need to make sure resync covers any new space.
4848 * If the array is shrinking we should possibly wait until
4849 * any io in the removed space completes, but it hardly seems
4850 * worth it.
4852 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4853 md_set_array_sectors(mddev, raid5_size(mddev, sectors,
4854 mddev->raid_disks));
4855 if (mddev->array_sectors >
4856 raid5_size(mddev, sectors, mddev->raid_disks))
4857 return -EINVAL;
4858 set_capacity(mddev->gendisk, mddev->array_sectors);
4859 mddev->changed = 1;
4860 if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
4861 mddev->recovery_cp = mddev->dev_sectors;
4862 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4864 mddev->dev_sectors = sectors;
4865 mddev->resync_max_sectors = sectors;
4866 return 0;
4869 static int check_stripe_cache(mddev_t *mddev)
4871 /* Can only proceed if there are plenty of stripe_heads.
4872 * We need a minimum of one full stripe,, and for sensible progress
4873 * it is best to have about 4 times that.
4874 * If we require 4 times, then the default 256 4K stripe_heads will
4875 * allow for chunk sizes up to 256K, which is probably OK.
4876 * If the chunk size is greater, user-space should request more
4877 * stripe_heads first.
4879 raid5_conf_t *conf = mddev->private;
4880 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
4881 > conf->max_nr_stripes ||
4882 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
4883 > conf->max_nr_stripes) {
4884 printk(KERN_WARNING "raid5: reshape: not enough stripes. Needed %lu\n",
4885 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
4886 / STRIPE_SIZE)*4);
4887 return 0;
4889 return 1;
4892 static int check_reshape(mddev_t *mddev)
4894 raid5_conf_t *conf = mddev->private;
4896 if (mddev->delta_disks == 0 &&
4897 mddev->new_layout == mddev->layout &&
4898 mddev->new_chunk_sectors == mddev->chunk_sectors)
4899 return 0; /* nothing to do */
4900 if (mddev->bitmap)
4901 /* Cannot grow a bitmap yet */
4902 return -EBUSY;
4903 if (mddev->degraded > conf->max_degraded)
4904 return -EINVAL;
4905 if (mddev->delta_disks < 0) {
4906 /* We might be able to shrink, but the devices must
4907 * be made bigger first.
4908 * For raid6, 4 is the minimum size.
4909 * Otherwise 2 is the minimum
4911 int min = 2;
4912 if (mddev->level == 6)
4913 min = 4;
4914 if (mddev->raid_disks + mddev->delta_disks < min)
4915 return -EINVAL;
4918 if (!check_stripe_cache(mddev))
4919 return -ENOSPC;
4921 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4924 static int raid5_start_reshape(mddev_t *mddev)
4926 raid5_conf_t *conf = mddev->private;
4927 mdk_rdev_t *rdev;
4928 int spares = 0;
4929 int added_devices = 0;
4930 unsigned long flags;
4932 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4933 return -EBUSY;
4935 if (!check_stripe_cache(mddev))
4936 return -ENOSPC;
4938 list_for_each_entry(rdev, &mddev->disks, same_set)
4939 if (rdev->raid_disk < 0 &&
4940 !test_bit(Faulty, &rdev->flags))
4941 spares++;
4943 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4944 /* Not enough devices even to make a degraded array
4945 * of that size
4947 return -EINVAL;
4949 /* Refuse to reduce size of the array. Any reductions in
4950 * array size must be through explicit setting of array_size
4951 * attribute.
4953 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
4954 < mddev->array_sectors) {
4955 printk(KERN_ERR "md: %s: array size must be reduced "
4956 "before number of disks\n", mdname(mddev));
4957 return -EINVAL;
4960 atomic_set(&conf->reshape_stripes, 0);
4961 spin_lock_irq(&conf->device_lock);
4962 conf->previous_raid_disks = conf->raid_disks;
4963 conf->raid_disks += mddev->delta_disks;
4964 conf->prev_chunk_sectors = conf->chunk_sectors;
4965 conf->chunk_sectors = mddev->new_chunk_sectors;
4966 conf->prev_algo = conf->algorithm;
4967 conf->algorithm = mddev->new_layout;
4968 if (mddev->delta_disks < 0)
4969 conf->reshape_progress = raid5_size(mddev, 0, 0);
4970 else
4971 conf->reshape_progress = 0;
4972 conf->reshape_safe = conf->reshape_progress;
4973 conf->generation++;
4974 spin_unlock_irq(&conf->device_lock);
4976 /* Add some new drives, as many as will fit.
4977 * We know there are enough to make the newly sized array work.
4979 list_for_each_entry(rdev, &mddev->disks, same_set)
4980 if (rdev->raid_disk < 0 &&
4981 !test_bit(Faulty, &rdev->flags)) {
4982 if (raid5_add_disk(mddev, rdev) == 0) {
4983 char nm[20];
4984 set_bit(In_sync, &rdev->flags);
4985 added_devices++;
4986 rdev->recovery_offset = 0;
4987 sprintf(nm, "rd%d", rdev->raid_disk);
4988 if (sysfs_create_link(&mddev->kobj,
4989 &rdev->kobj, nm))
4990 printk(KERN_WARNING
4991 "raid5: failed to create "
4992 " link %s for %s\n",
4993 nm, mdname(mddev));
4994 } else
4995 break;
4998 if (mddev->delta_disks > 0) {
4999 spin_lock_irqsave(&conf->device_lock, flags);
5000 mddev->degraded = (conf->raid_disks - conf->previous_raid_disks)
5001 - added_devices;
5002 spin_unlock_irqrestore(&conf->device_lock, flags);
5004 mddev->raid_disks = conf->raid_disks;
5005 mddev->reshape_position = 0;
5006 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5008 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5009 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5010 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5011 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5012 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5013 "%s_reshape");
5014 if (!mddev->sync_thread) {
5015 mddev->recovery = 0;
5016 spin_lock_irq(&conf->device_lock);
5017 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5018 conf->reshape_progress = MaxSector;
5019 spin_unlock_irq(&conf->device_lock);
5020 return -EAGAIN;
5022 conf->reshape_checkpoint = jiffies;
5023 md_wakeup_thread(mddev->sync_thread);
5024 md_new_event(mddev);
5025 return 0;
5028 /* This is called from the reshape thread and should make any
5029 * changes needed in 'conf'
5031 static void end_reshape(raid5_conf_t *conf)
5034 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5036 spin_lock_irq(&conf->device_lock);
5037 conf->previous_raid_disks = conf->raid_disks;
5038 conf->reshape_progress = MaxSector;
5039 spin_unlock_irq(&conf->device_lock);
5040 wake_up(&conf->wait_for_overlap);
5042 /* read-ahead size must cover two whole stripes, which is
5043 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5046 int data_disks = conf->raid_disks - conf->max_degraded;
5047 int stripe = data_disks * ((conf->chunk_sectors << 9)
5048 / PAGE_SIZE);
5049 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5050 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5055 /* This is called from the raid5d thread with mddev_lock held.
5056 * It makes config changes to the device.
5058 static void raid5_finish_reshape(mddev_t *mddev)
5060 struct block_device *bdev;
5061 raid5_conf_t *conf = mddev->private;
5063 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5065 if (mddev->delta_disks > 0) {
5066 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5067 set_capacity(mddev->gendisk, mddev->array_sectors);
5068 mddev->changed = 1;
5070 bdev = bdget_disk(mddev->gendisk, 0);
5071 if (bdev) {
5072 mutex_lock(&bdev->bd_inode->i_mutex);
5073 i_size_write(bdev->bd_inode,
5074 (loff_t)mddev->array_sectors << 9);
5075 mutex_unlock(&bdev->bd_inode->i_mutex);
5076 bdput(bdev);
5078 } else {
5079 int d;
5080 mddev->degraded = conf->raid_disks;
5081 for (d = 0; d < conf->raid_disks ; d++)
5082 if (conf->disks[d].rdev &&
5083 test_bit(In_sync,
5084 &conf->disks[d].rdev->flags))
5085 mddev->degraded--;
5086 for (d = conf->raid_disks ;
5087 d < conf->raid_disks - mddev->delta_disks;
5088 d++)
5089 raid5_remove_disk(mddev, d);
5091 mddev->layout = conf->algorithm;
5092 mddev->chunk_sectors = conf->chunk_sectors;
5093 mddev->reshape_position = MaxSector;
5094 mddev->delta_disks = 0;
5098 static void raid5_quiesce(mddev_t *mddev, int state)
5100 raid5_conf_t *conf = mddev->private;
5102 switch(state) {
5103 case 2: /* resume for a suspend */
5104 wake_up(&conf->wait_for_overlap);
5105 break;
5107 case 1: /* stop all writes */
5108 spin_lock_irq(&conf->device_lock);
5109 conf->quiesce = 1;
5110 wait_event_lock_irq(conf->wait_for_stripe,
5111 atomic_read(&conf->active_stripes) == 0 &&
5112 atomic_read(&conf->active_aligned_reads) == 0,
5113 conf->device_lock, /* nothing */);
5114 spin_unlock_irq(&conf->device_lock);
5115 break;
5117 case 0: /* re-enable writes */
5118 spin_lock_irq(&conf->device_lock);
5119 conf->quiesce = 0;
5120 wake_up(&conf->wait_for_stripe);
5121 wake_up(&conf->wait_for_overlap);
5122 spin_unlock_irq(&conf->device_lock);
5123 break;
5128 static void *raid5_takeover_raid1(mddev_t *mddev)
5130 int chunksect;
5132 if (mddev->raid_disks != 2 ||
5133 mddev->degraded > 1)
5134 return ERR_PTR(-EINVAL);
5136 /* Should check if there are write-behind devices? */
5138 chunksect = 64*2; /* 64K by default */
5140 /* The array must be an exact multiple of chunksize */
5141 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5142 chunksect >>= 1;
5144 if ((chunksect<<9) < STRIPE_SIZE)
5145 /* array size does not allow a suitable chunk size */
5146 return ERR_PTR(-EINVAL);
5148 mddev->new_level = 5;
5149 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5150 mddev->new_chunk_sectors = chunksect;
5152 return setup_conf(mddev);
5155 static void *raid5_takeover_raid6(mddev_t *mddev)
5157 int new_layout;
5159 switch (mddev->layout) {
5160 case ALGORITHM_LEFT_ASYMMETRIC_6:
5161 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5162 break;
5163 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5164 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5165 break;
5166 case ALGORITHM_LEFT_SYMMETRIC_6:
5167 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5168 break;
5169 case ALGORITHM_RIGHT_SYMMETRIC_6:
5170 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5171 break;
5172 case ALGORITHM_PARITY_0_6:
5173 new_layout = ALGORITHM_PARITY_0;
5174 break;
5175 case ALGORITHM_PARITY_N:
5176 new_layout = ALGORITHM_PARITY_N;
5177 break;
5178 default:
5179 return ERR_PTR(-EINVAL);
5181 mddev->new_level = 5;
5182 mddev->new_layout = new_layout;
5183 mddev->delta_disks = -1;
5184 mddev->raid_disks -= 1;
5185 return setup_conf(mddev);
5189 static int raid5_check_reshape(mddev_t *mddev)
5191 /* For a 2-drive array, the layout and chunk size can be changed
5192 * immediately as not restriping is needed.
5193 * For larger arrays we record the new value - after validation
5194 * to be used by a reshape pass.
5196 raid5_conf_t *conf = mddev->private;
5197 int new_chunk = mddev->new_chunk_sectors;
5199 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5200 return -EINVAL;
5201 if (new_chunk > 0) {
5202 if (!is_power_of_2(new_chunk))
5203 return -EINVAL;
5204 if (new_chunk < (PAGE_SIZE>>9))
5205 return -EINVAL;
5206 if (mddev->array_sectors & (new_chunk-1))
5207 /* not factor of array size */
5208 return -EINVAL;
5211 /* They look valid */
5213 if (mddev->raid_disks == 2) {
5214 /* can make the change immediately */
5215 if (mddev->new_layout >= 0) {
5216 conf->algorithm = mddev->new_layout;
5217 mddev->layout = mddev->new_layout;
5219 if (new_chunk > 0) {
5220 conf->chunk_sectors = new_chunk ;
5221 mddev->chunk_sectors = new_chunk;
5223 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5224 md_wakeup_thread(mddev->thread);
5226 return check_reshape(mddev);
5229 static int raid6_check_reshape(mddev_t *mddev)
5231 int new_chunk = mddev->new_chunk_sectors;
5233 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5234 return -EINVAL;
5235 if (new_chunk > 0) {
5236 if (!is_power_of_2(new_chunk))
5237 return -EINVAL;
5238 if (new_chunk < (PAGE_SIZE >> 9))
5239 return -EINVAL;
5240 if (mddev->array_sectors & (new_chunk-1))
5241 /* not factor of array size */
5242 return -EINVAL;
5245 /* They look valid */
5246 return check_reshape(mddev);
5249 static void *raid5_takeover(mddev_t *mddev)
5251 /* raid5 can take over:
5252 * raid0 - if all devices are the same - make it a raid4 layout
5253 * raid1 - if there are two drives. We need to know the chunk size
5254 * raid4 - trivial - just use a raid4 layout.
5255 * raid6 - Providing it is a *_6 layout
5258 if (mddev->level == 1)
5259 return raid5_takeover_raid1(mddev);
5260 if (mddev->level == 4) {
5261 mddev->new_layout = ALGORITHM_PARITY_N;
5262 mddev->new_level = 5;
5263 return setup_conf(mddev);
5265 if (mddev->level == 6)
5266 return raid5_takeover_raid6(mddev);
5268 return ERR_PTR(-EINVAL);
5272 static struct mdk_personality raid5_personality;
5274 static void *raid6_takeover(mddev_t *mddev)
5276 /* Currently can only take over a raid5. We map the
5277 * personality to an equivalent raid6 personality
5278 * with the Q block at the end.
5280 int new_layout;
5282 if (mddev->pers != &raid5_personality)
5283 return ERR_PTR(-EINVAL);
5284 if (mddev->degraded > 1)
5285 return ERR_PTR(-EINVAL);
5286 if (mddev->raid_disks > 253)
5287 return ERR_PTR(-EINVAL);
5288 if (mddev->raid_disks < 3)
5289 return ERR_PTR(-EINVAL);
5291 switch (mddev->layout) {
5292 case ALGORITHM_LEFT_ASYMMETRIC:
5293 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5294 break;
5295 case ALGORITHM_RIGHT_ASYMMETRIC:
5296 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5297 break;
5298 case ALGORITHM_LEFT_SYMMETRIC:
5299 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5300 break;
5301 case ALGORITHM_RIGHT_SYMMETRIC:
5302 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5303 break;
5304 case ALGORITHM_PARITY_0:
5305 new_layout = ALGORITHM_PARITY_0_6;
5306 break;
5307 case ALGORITHM_PARITY_N:
5308 new_layout = ALGORITHM_PARITY_N;
5309 break;
5310 default:
5311 return ERR_PTR(-EINVAL);
5313 mddev->new_level = 6;
5314 mddev->new_layout = new_layout;
5315 mddev->delta_disks = 1;
5316 mddev->raid_disks += 1;
5317 return setup_conf(mddev);
5321 static struct mdk_personality raid6_personality =
5323 .name = "raid6",
5324 .level = 6,
5325 .owner = THIS_MODULE,
5326 .make_request = make_request,
5327 .run = run,
5328 .stop = stop,
5329 .status = status,
5330 .error_handler = error,
5331 .hot_add_disk = raid5_add_disk,
5332 .hot_remove_disk= raid5_remove_disk,
5333 .spare_active = raid5_spare_active,
5334 .sync_request = sync_request,
5335 .resize = raid5_resize,
5336 .size = raid5_size,
5337 .check_reshape = raid6_check_reshape,
5338 .start_reshape = raid5_start_reshape,
5339 .finish_reshape = raid5_finish_reshape,
5340 .quiesce = raid5_quiesce,
5341 .takeover = raid6_takeover,
5343 static struct mdk_personality raid5_personality =
5345 .name = "raid5",
5346 .level = 5,
5347 .owner = THIS_MODULE,
5348 .make_request = make_request,
5349 .run = run,
5350 .stop = stop,
5351 .status = status,
5352 .error_handler = error,
5353 .hot_add_disk = raid5_add_disk,
5354 .hot_remove_disk= raid5_remove_disk,
5355 .spare_active = raid5_spare_active,
5356 .sync_request = sync_request,
5357 .resize = raid5_resize,
5358 .size = raid5_size,
5359 .check_reshape = raid5_check_reshape,
5360 .start_reshape = raid5_start_reshape,
5361 .finish_reshape = raid5_finish_reshape,
5362 .quiesce = raid5_quiesce,
5363 .takeover = raid5_takeover,
5366 static struct mdk_personality raid4_personality =
5368 .name = "raid4",
5369 .level = 4,
5370 .owner = THIS_MODULE,
5371 .make_request = make_request,
5372 .run = run,
5373 .stop = stop,
5374 .status = status,
5375 .error_handler = error,
5376 .hot_add_disk = raid5_add_disk,
5377 .hot_remove_disk= raid5_remove_disk,
5378 .spare_active = raid5_spare_active,
5379 .sync_request = sync_request,
5380 .resize = raid5_resize,
5381 .size = raid5_size,
5382 .check_reshape = raid5_check_reshape,
5383 .start_reshape = raid5_start_reshape,
5384 .finish_reshape = raid5_finish_reshape,
5385 .quiesce = raid5_quiesce,
5388 static int __init raid5_init(void)
5390 register_md_personality(&raid6_personality);
5391 register_md_personality(&raid5_personality);
5392 register_md_personality(&raid4_personality);
5393 return 0;
5396 static void raid5_exit(void)
5398 unregister_md_personality(&raid6_personality);
5399 unregister_md_personality(&raid5_personality);
5400 unregister_md_personality(&raid4_personality);
5403 module_init(raid5_init);
5404 module_exit(raid5_exit);
5405 MODULE_LICENSE("GPL");
5406 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5407 MODULE_ALIAS("md-raid5");
5408 MODULE_ALIAS("md-raid4");
5409 MODULE_ALIAS("md-level-5");
5410 MODULE_ALIAS("md-level-4");
5411 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5412 MODULE_ALIAS("md-raid6");
5413 MODULE_ALIAS("md-level-6");
5415 /* This used to be two separate modules, they were: */
5416 MODULE_ALIAS("raid5");
5417 MODULE_ALIAS("raid6");