2 * raid1.c : Multiple Devices driver for Linux
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 * RAID-1 management functions.
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
34 #include <linux/delay.h>
35 #include <linux/blkdev.h>
36 #include <linux/seq_file.h>
43 #define PRINTK(x...) printk(x)
49 * Number of guaranteed r1bios in case of extreme VM load:
51 #define NR_RAID1_BIOS 256
54 static void unplug_slaves(mddev_t *mddev);
56 static void allow_barrier(conf_t *conf);
57 static void lower_barrier(conf_t *conf);
59 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
61 struct pool_info *pi = data;
63 int size = offsetof(r1bio_t, bios[pi->raid_disks]);
65 /* allocate a r1bio with room for raid_disks entries in the bios array */
66 r1_bio = kzalloc(size, gfp_flags);
67 if (!r1_bio && pi->mddev)
68 unplug_slaves(pi->mddev);
73 static void r1bio_pool_free(void *r1_bio, void *data)
78 #define RESYNC_BLOCK_SIZE (64*1024)
79 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
80 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
81 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
82 #define RESYNC_WINDOW (2048*1024)
84 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
86 struct pool_info *pi = data;
92 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
94 unplug_slaves(pi->mddev);
99 * Allocate bios : 1 for reading, n-1 for writing
101 for (j = pi->raid_disks ; j-- ; ) {
102 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
105 r1_bio->bios[j] = bio;
108 * Allocate RESYNC_PAGES data pages and attach them to
110 * If this is a user-requested check/repair, allocate
111 * RESYNC_PAGES for each bio.
113 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
118 bio = r1_bio->bios[j];
119 for (i = 0; i < RESYNC_PAGES; i++) {
120 page = alloc_page(gfp_flags);
124 bio->bi_io_vec[i].bv_page = page;
128 /* If not user-requests, copy the page pointers to all bios */
129 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
130 for (i=0; i<RESYNC_PAGES ; i++)
131 for (j=1; j<pi->raid_disks; j++)
132 r1_bio->bios[j]->bi_io_vec[i].bv_page =
133 r1_bio->bios[0]->bi_io_vec[i].bv_page;
136 r1_bio->master_bio = NULL;
141 for (j=0 ; j < pi->raid_disks; j++)
142 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
143 put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
146 while ( ++j < pi->raid_disks )
147 bio_put(r1_bio->bios[j]);
148 r1bio_pool_free(r1_bio, data);
152 static void r1buf_pool_free(void *__r1_bio, void *data)
154 struct pool_info *pi = data;
156 r1bio_t *r1bio = __r1_bio;
158 for (i = 0; i < RESYNC_PAGES; i++)
159 for (j = pi->raid_disks; j-- ;) {
161 r1bio->bios[j]->bi_io_vec[i].bv_page !=
162 r1bio->bios[0]->bi_io_vec[i].bv_page)
163 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
165 for (i=0 ; i < pi->raid_disks; i++)
166 bio_put(r1bio->bios[i]);
168 r1bio_pool_free(r1bio, data);
171 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
175 for (i = 0; i < conf->raid_disks; i++) {
176 struct bio **bio = r1_bio->bios + i;
177 if (*bio && *bio != IO_BLOCKED)
183 static void free_r1bio(r1bio_t *r1_bio)
185 conf_t *conf = r1_bio->mddev->private;
188 * Wake up any possible resync thread that waits for the device
193 put_all_bios(conf, r1_bio);
194 mempool_free(r1_bio, conf->r1bio_pool);
197 static void put_buf(r1bio_t *r1_bio)
199 conf_t *conf = r1_bio->mddev->private;
202 for (i=0; i<conf->raid_disks; i++) {
203 struct bio *bio = r1_bio->bios[i];
205 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
208 mempool_free(r1_bio, conf->r1buf_pool);
213 static void reschedule_retry(r1bio_t *r1_bio)
216 mddev_t *mddev = r1_bio->mddev;
217 conf_t *conf = mddev->private;
219 spin_lock_irqsave(&conf->device_lock, flags);
220 list_add(&r1_bio->retry_list, &conf->retry_list);
222 spin_unlock_irqrestore(&conf->device_lock, flags);
224 wake_up(&conf->wait_barrier);
225 md_wakeup_thread(mddev->thread);
229 * raid_end_bio_io() is called when we have finished servicing a mirrored
230 * operation and are ready to return a success/failure code to the buffer
233 static void raid_end_bio_io(r1bio_t *r1_bio)
235 struct bio *bio = r1_bio->master_bio;
237 /* if nobody has done the final endio yet, do it now */
238 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
239 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
240 (bio_data_dir(bio) == WRITE) ? "write" : "read",
241 (unsigned long long) bio->bi_sector,
242 (unsigned long long) bio->bi_sector +
243 (bio->bi_size >> 9) - 1);
246 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
252 * Update disk head position estimator based on IRQ completion info.
254 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
256 conf_t *conf = r1_bio->mddev->private;
258 conf->mirrors[disk].head_position =
259 r1_bio->sector + (r1_bio->sectors);
262 static void raid1_end_read_request(struct bio *bio, int error)
264 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
265 r1bio_t *r1_bio = bio->bi_private;
267 conf_t *conf = r1_bio->mddev->private;
269 mirror = r1_bio->read_disk;
271 * this branch is our 'one mirror IO has finished' event handler:
273 update_head_pos(mirror, r1_bio);
276 set_bit(R1BIO_Uptodate, &r1_bio->state);
278 /* If all other devices have failed, we want to return
279 * the error upwards rather than fail the last device.
280 * Here we redefine "uptodate" to mean "Don't want to retry"
283 spin_lock_irqsave(&conf->device_lock, flags);
284 if (r1_bio->mddev->degraded == conf->raid_disks ||
285 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
286 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
288 spin_unlock_irqrestore(&conf->device_lock, flags);
292 raid_end_bio_io(r1_bio);
297 char b[BDEVNAME_SIZE];
298 if (printk_ratelimit())
299 printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
300 bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
301 reschedule_retry(r1_bio);
304 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
307 static void raid1_end_write_request(struct bio *bio, int error)
309 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
310 r1bio_t *r1_bio = bio->bi_private;
311 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
312 conf_t *conf = r1_bio->mddev->private;
313 struct bio *to_put = NULL;
316 for (mirror = 0; mirror < conf->raid_disks; mirror++)
317 if (r1_bio->bios[mirror] == bio)
320 if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
321 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
322 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
323 r1_bio->mddev->barriers_work = 0;
324 /* Don't rdev_dec_pending in this branch - keep it for the retry */
327 * this branch is our 'one mirror IO has finished' event handler:
329 r1_bio->bios[mirror] = NULL;
332 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
333 /* an I/O failed, we can't clear the bitmap */
334 set_bit(R1BIO_Degraded, &r1_bio->state);
337 * Set R1BIO_Uptodate in our master bio, so that
338 * we will return a good error code for to the higher
339 * levels even if IO on some other mirrored buffer fails.
341 * The 'master' represents the composite IO operation to
342 * user-side. So if something waits for IO, then it will
343 * wait for the 'master' bio.
345 set_bit(R1BIO_Uptodate, &r1_bio->state);
347 update_head_pos(mirror, r1_bio);
350 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
351 atomic_dec(&r1_bio->behind_remaining);
353 /* In behind mode, we ACK the master bio once the I/O has safely
354 * reached all non-writemostly disks. Setting the Returned bit
355 * ensures that this gets done only once -- we don't ever want to
356 * return -EIO here, instead we'll wait */
358 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
359 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
360 /* Maybe we can return now */
361 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
362 struct bio *mbio = r1_bio->master_bio;
363 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
364 (unsigned long long) mbio->bi_sector,
365 (unsigned long long) mbio->bi_sector +
366 (mbio->bi_size >> 9) - 1);
371 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
375 * Let's see if all mirrored write operations have finished
378 if (atomic_dec_and_test(&r1_bio->remaining)) {
379 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
380 reschedule_retry(r1_bio);
382 /* it really is the end of this request */
383 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
384 /* free extra copy of the data pages */
385 int i = bio->bi_vcnt;
387 safe_put_page(bio->bi_io_vec[i].bv_page);
389 /* clear the bitmap if all writes complete successfully */
390 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
392 !test_bit(R1BIO_Degraded, &r1_bio->state),
394 md_write_end(r1_bio->mddev);
395 raid_end_bio_io(r1_bio);
405 * This routine returns the disk from which the requested read should
406 * be done. There is a per-array 'next expected sequential IO' sector
407 * number - if this matches on the next IO then we use the last disk.
408 * There is also a per-disk 'last know head position' sector that is
409 * maintained from IRQ contexts, both the normal and the resync IO
410 * completion handlers update this position correctly. If there is no
411 * perfect sequential match then we pick the disk whose head is closest.
413 * If there are 2 mirrors in the same 2 devices, performance degrades
414 * because position is mirror, not device based.
416 * The rdev for the device selected will have nr_pending incremented.
418 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
420 const unsigned long this_sector = r1_bio->sector;
421 int new_disk = conf->last_used, disk = new_disk;
423 const int sectors = r1_bio->sectors;
424 sector_t new_distance, current_distance;
429 * Check if we can balance. We can balance on the whole
430 * device if no resync is going on, or below the resync window.
431 * We take the first readable disk when above the resync window.
434 if (conf->mddev->recovery_cp < MaxSector &&
435 (this_sector + sectors >= conf->next_resync)) {
436 /* Choose the first operation device, for consistancy */
439 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
440 r1_bio->bios[new_disk] == IO_BLOCKED ||
441 !rdev || !test_bit(In_sync, &rdev->flags)
442 || test_bit(WriteMostly, &rdev->flags);
443 rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
445 if (rdev && test_bit(In_sync, &rdev->flags) &&
446 r1_bio->bios[new_disk] != IO_BLOCKED)
447 wonly_disk = new_disk;
449 if (new_disk == conf->raid_disks - 1) {
450 new_disk = wonly_disk;
458 /* make sure the disk is operational */
459 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
460 r1_bio->bios[new_disk] == IO_BLOCKED ||
461 !rdev || !test_bit(In_sync, &rdev->flags) ||
462 test_bit(WriteMostly, &rdev->flags);
463 rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
465 if (rdev && test_bit(In_sync, &rdev->flags) &&
466 r1_bio->bios[new_disk] != IO_BLOCKED)
467 wonly_disk = new_disk;
470 new_disk = conf->raid_disks;
472 if (new_disk == disk) {
473 new_disk = wonly_disk;
482 /* now disk == new_disk == starting point for search */
485 * Don't change to another disk for sequential reads:
487 if (conf->next_seq_sect == this_sector)
489 if (this_sector == conf->mirrors[new_disk].head_position)
492 current_distance = abs(this_sector - conf->mirrors[disk].head_position);
494 /* Find the disk whose head is closest */
498 disk = conf->raid_disks;
501 rdev = rcu_dereference(conf->mirrors[disk].rdev);
503 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
504 !test_bit(In_sync, &rdev->flags) ||
505 test_bit(WriteMostly, &rdev->flags))
508 if (!atomic_read(&rdev->nr_pending)) {
512 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
513 if (new_distance < current_distance) {
514 current_distance = new_distance;
517 } while (disk != conf->last_used);
523 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
526 atomic_inc(&rdev->nr_pending);
527 if (!test_bit(In_sync, &rdev->flags)) {
528 /* cannot risk returning a device that failed
529 * before we inc'ed nr_pending
531 rdev_dec_pending(rdev, conf->mddev);
534 conf->next_seq_sect = this_sector + sectors;
535 conf->last_used = new_disk;
542 static void unplug_slaves(mddev_t *mddev)
544 conf_t *conf = mddev->private;
548 for (i=0; i<mddev->raid_disks; i++) {
549 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
550 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
551 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
553 atomic_inc(&rdev->nr_pending);
558 rdev_dec_pending(rdev, mddev);
565 static void raid1_unplug(struct request_queue *q)
567 mddev_t *mddev = q->queuedata;
569 unplug_slaves(mddev);
570 md_wakeup_thread(mddev->thread);
573 static int raid1_congested(void *data, int bits)
575 mddev_t *mddev = data;
576 conf_t *conf = mddev->private;
579 if (mddev_congested(mddev, bits))
583 for (i = 0; i < mddev->raid_disks; i++) {
584 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
585 if (rdev && !test_bit(Faulty, &rdev->flags)) {
586 struct request_queue *q = bdev_get_queue(rdev->bdev);
588 /* Note the '|| 1' - when read_balance prefers
589 * non-congested targets, it can be removed
591 if ((bits & (1<<BDI_async_congested)) || 1)
592 ret |= bdi_congested(&q->backing_dev_info, bits);
594 ret &= bdi_congested(&q->backing_dev_info, bits);
602 static int flush_pending_writes(conf_t *conf)
604 /* Any writes that have been queued but are awaiting
605 * bitmap updates get flushed here.
606 * We return 1 if any requests were actually submitted.
610 spin_lock_irq(&conf->device_lock);
612 if (conf->pending_bio_list.head) {
614 bio = bio_list_get(&conf->pending_bio_list);
615 blk_remove_plug(conf->mddev->queue);
616 spin_unlock_irq(&conf->device_lock);
617 /* flush any pending bitmap writes to
618 * disk before proceeding w/ I/O */
619 bitmap_unplug(conf->mddev->bitmap);
621 while (bio) { /* submit pending writes */
622 struct bio *next = bio->bi_next;
624 generic_make_request(bio);
629 spin_unlock_irq(&conf->device_lock);
634 * Sometimes we need to suspend IO while we do something else,
635 * either some resync/recovery, or reconfigure the array.
636 * To do this we raise a 'barrier'.
637 * The 'barrier' is a counter that can be raised multiple times
638 * to count how many activities are happening which preclude
640 * We can only raise the barrier if there is no pending IO.
641 * i.e. if nr_pending == 0.
642 * We choose only to raise the barrier if no-one is waiting for the
643 * barrier to go down. This means that as soon as an IO request
644 * is ready, no other operations which require a barrier will start
645 * until the IO request has had a chance.
647 * So: regular IO calls 'wait_barrier'. When that returns there
648 * is no backgroup IO happening, It must arrange to call
649 * allow_barrier when it has finished its IO.
650 * backgroup IO calls must call raise_barrier. Once that returns
651 * there is no normal IO happeing. It must arrange to call
652 * lower_barrier when the particular background IO completes.
654 #define RESYNC_DEPTH 32
656 static void raise_barrier(conf_t *conf)
658 spin_lock_irq(&conf->resync_lock);
660 /* Wait until no block IO is waiting */
661 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
663 raid1_unplug(conf->mddev->queue));
665 /* block any new IO from starting */
668 /* No wait for all pending IO to complete */
669 wait_event_lock_irq(conf->wait_barrier,
670 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
672 raid1_unplug(conf->mddev->queue));
674 spin_unlock_irq(&conf->resync_lock);
677 static void lower_barrier(conf_t *conf)
680 BUG_ON(conf->barrier <= 0);
681 spin_lock_irqsave(&conf->resync_lock, flags);
683 spin_unlock_irqrestore(&conf->resync_lock, flags);
684 wake_up(&conf->wait_barrier);
687 static void wait_barrier(conf_t *conf)
689 spin_lock_irq(&conf->resync_lock);
692 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
694 raid1_unplug(conf->mddev->queue));
698 spin_unlock_irq(&conf->resync_lock);
701 static void allow_barrier(conf_t *conf)
704 spin_lock_irqsave(&conf->resync_lock, flags);
706 spin_unlock_irqrestore(&conf->resync_lock, flags);
707 wake_up(&conf->wait_barrier);
710 static void freeze_array(conf_t *conf)
712 /* stop syncio and normal IO and wait for everything to
714 * We increment barrier and nr_waiting, and then
715 * wait until nr_pending match nr_queued+1
716 * This is called in the context of one normal IO request
717 * that has failed. Thus any sync request that might be pending
718 * will be blocked by nr_pending, and we need to wait for
719 * pending IO requests to complete or be queued for re-try.
720 * Thus the number queued (nr_queued) plus this request (1)
721 * must match the number of pending IOs (nr_pending) before
724 spin_lock_irq(&conf->resync_lock);
727 wait_event_lock_irq(conf->wait_barrier,
728 conf->nr_pending == conf->nr_queued+1,
730 ({ flush_pending_writes(conf);
731 raid1_unplug(conf->mddev->queue); }));
732 spin_unlock_irq(&conf->resync_lock);
734 static void unfreeze_array(conf_t *conf)
736 /* reverse the effect of the freeze */
737 spin_lock_irq(&conf->resync_lock);
740 wake_up(&conf->wait_barrier);
741 spin_unlock_irq(&conf->resync_lock);
745 /* duplicate the data pages for behind I/O */
746 static struct page **alloc_behind_pages(struct bio *bio)
749 struct bio_vec *bvec;
750 struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
752 if (unlikely(!pages))
755 bio_for_each_segment(bvec, bio, i) {
756 pages[i] = alloc_page(GFP_NOIO);
757 if (unlikely(!pages[i]))
759 memcpy(kmap(pages[i]) + bvec->bv_offset,
760 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
762 kunmap(bvec->bv_page);
769 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
772 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
776 static int make_request(struct request_queue *q, struct bio * bio)
778 mddev_t *mddev = q->queuedata;
779 conf_t *conf = mddev->private;
780 mirror_info_t *mirror;
782 struct bio *read_bio;
783 int i, targets = 0, disks;
784 struct bitmap *bitmap;
787 struct page **behind_pages = NULL;
788 const int rw = bio_data_dir(bio);
789 const bool do_sync = bio_rw_flagged(bio, BIO_RW_SYNCIO);
791 mdk_rdev_t *blocked_rdev;
794 * Register the new request and wait if the reconstruction
795 * thread has put up a bar for new requests.
796 * Continue immediately if no resync is active currently.
797 * We test barriers_work *after* md_write_start as md_write_start
798 * may cause the first superblock write, and that will check out
802 md_write_start(mddev, bio); /* wait on superblock update early */
804 if (bio_data_dir(bio) == WRITE &&
805 bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
806 bio->bi_sector < mddev->suspend_hi) {
807 /* As the suspend_* range is controlled by
808 * userspace, we want an interruptible
813 flush_signals(current);
814 prepare_to_wait(&conf->wait_barrier,
815 &w, TASK_INTERRUPTIBLE);
816 if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
817 bio->bi_sector >= mddev->suspend_hi)
821 finish_wait(&conf->wait_barrier, &w);
823 if (unlikely(!mddev->barriers_work &&
824 bio_rw_flagged(bio, BIO_RW_BARRIER))) {
827 bio_endio(bio, -EOPNOTSUPP);
833 bitmap = mddev->bitmap;
836 * make_request() can abort the operation when READA is being
837 * used and no empty request is available.
840 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
842 r1_bio->master_bio = bio;
843 r1_bio->sectors = bio->bi_size >> 9;
845 r1_bio->mddev = mddev;
846 r1_bio->sector = bio->bi_sector;
850 * read balancing logic:
852 int rdisk = read_balance(conf, r1_bio);
855 /* couldn't find anywhere to read from */
856 raid_end_bio_io(r1_bio);
859 mirror = conf->mirrors + rdisk;
861 r1_bio->read_disk = rdisk;
863 read_bio = bio_clone(bio, GFP_NOIO);
865 r1_bio->bios[rdisk] = read_bio;
867 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
868 read_bio->bi_bdev = mirror->rdev->bdev;
869 read_bio->bi_end_io = raid1_end_read_request;
870 read_bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
871 read_bio->bi_private = r1_bio;
873 generic_make_request(read_bio);
880 /* first select target devices under spinlock and
881 * inc refcount on their rdev. Record them by setting
884 disks = conf->raid_disks;
886 { static int first=1;
887 if (first) printk("First Write sector %llu disks %d\n",
888 (unsigned long long)r1_bio->sector, disks);
895 for (i = 0; i < disks; i++) {
896 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
897 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
898 atomic_inc(&rdev->nr_pending);
902 if (rdev && !test_bit(Faulty, &rdev->flags)) {
903 atomic_inc(&rdev->nr_pending);
904 if (test_bit(Faulty, &rdev->flags)) {
905 rdev_dec_pending(rdev, mddev);
906 r1_bio->bios[i] = NULL;
908 r1_bio->bios[i] = bio;
912 r1_bio->bios[i] = NULL;
916 if (unlikely(blocked_rdev)) {
917 /* Wait for this device to become unblocked */
920 for (j = 0; j < i; j++)
922 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
925 md_wait_for_blocked_rdev(blocked_rdev, mddev);
930 BUG_ON(targets == 0); /* we never fail the last device */
932 if (targets < conf->raid_disks) {
933 /* array is degraded, we will not clear the bitmap
934 * on I/O completion (see raid1_end_write_request) */
935 set_bit(R1BIO_Degraded, &r1_bio->state);
938 /* do behind I/O ? */
940 (atomic_read(&bitmap->behind_writes)
941 < mddev->bitmap_info.max_write_behind) &&
942 (behind_pages = alloc_behind_pages(bio)) != NULL)
943 set_bit(R1BIO_BehindIO, &r1_bio->state);
945 atomic_set(&r1_bio->remaining, 0);
946 atomic_set(&r1_bio->behind_remaining, 0);
948 do_barriers = bio_rw_flagged(bio, BIO_RW_BARRIER);
950 set_bit(R1BIO_Barrier, &r1_bio->state);
953 for (i = 0; i < disks; i++) {
955 if (!r1_bio->bios[i])
958 mbio = bio_clone(bio, GFP_NOIO);
959 r1_bio->bios[i] = mbio;
961 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
962 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
963 mbio->bi_end_io = raid1_end_write_request;
964 mbio->bi_rw = WRITE | (do_barriers << BIO_RW_BARRIER) |
965 (do_sync << BIO_RW_SYNCIO);
966 mbio->bi_private = r1_bio;
969 struct bio_vec *bvec;
972 /* Yes, I really want the '__' version so that
973 * we clear any unused pointer in the io_vec, rather
974 * than leave them unchanged. This is important
975 * because when we come to free the pages, we won't
976 * know the originial bi_idx, so we just free
979 __bio_for_each_segment(bvec, mbio, j, 0)
980 bvec->bv_page = behind_pages[j];
981 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
982 atomic_inc(&r1_bio->behind_remaining);
985 atomic_inc(&r1_bio->remaining);
987 bio_list_add(&bl, mbio);
989 kfree(behind_pages); /* the behind pages are attached to the bios now */
991 bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
992 test_bit(R1BIO_BehindIO, &r1_bio->state));
993 spin_lock_irqsave(&conf->device_lock, flags);
994 bio_list_merge(&conf->pending_bio_list, &bl);
997 blk_plug_device(mddev->queue);
998 spin_unlock_irqrestore(&conf->device_lock, flags);
1000 /* In case raid1d snuck into freeze_array */
1001 wake_up(&conf->wait_barrier);
1004 md_wakeup_thread(mddev->thread);
1006 while ((bio = bio_list_pop(&bl)) != NULL)
1007 generic_make_request(bio);
1013 static void status(struct seq_file *seq, mddev_t *mddev)
1015 conf_t *conf = mddev->private;
1018 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1019 conf->raid_disks - mddev->degraded);
1021 for (i = 0; i < conf->raid_disks; i++) {
1022 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1023 seq_printf(seq, "%s",
1024 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1027 seq_printf(seq, "]");
1031 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1033 char b[BDEVNAME_SIZE];
1034 conf_t *conf = mddev->private;
1037 * If it is not operational, then we have already marked it as dead
1038 * else if it is the last working disks, ignore the error, let the
1039 * next level up know.
1040 * else mark the drive as failed
1042 if (test_bit(In_sync, &rdev->flags)
1043 && (conf->raid_disks - mddev->degraded) == 1) {
1045 * Don't fail the drive, act as though we were just a
1046 * normal single drive.
1047 * However don't try a recovery from this drive as
1048 * it is very likely to fail.
1050 mddev->recovery_disabled = 1;
1053 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1054 unsigned long flags;
1055 spin_lock_irqsave(&conf->device_lock, flags);
1057 set_bit(Faulty, &rdev->flags);
1058 spin_unlock_irqrestore(&conf->device_lock, flags);
1060 * if recovery is running, make sure it aborts.
1062 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1064 set_bit(Faulty, &rdev->flags);
1065 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1066 printk(KERN_ALERT "raid1: Disk failure on %s, disabling device.\n"
1067 "raid1: Operation continuing on %d devices.\n",
1068 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1071 static void print_conf(conf_t *conf)
1075 printk("RAID1 conf printout:\n");
1077 printk("(!conf)\n");
1080 printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1084 for (i = 0; i < conf->raid_disks; i++) {
1085 char b[BDEVNAME_SIZE];
1086 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1088 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1089 i, !test_bit(In_sync, &rdev->flags),
1090 !test_bit(Faulty, &rdev->flags),
1091 bdevname(rdev->bdev,b));
1096 static void close_sync(conf_t *conf)
1099 allow_barrier(conf);
1101 mempool_destroy(conf->r1buf_pool);
1102 conf->r1buf_pool = NULL;
1105 static int raid1_spare_active(mddev_t *mddev)
1108 conf_t *conf = mddev->private;
1111 * Find all failed disks within the RAID1 configuration
1112 * and mark them readable.
1113 * Called under mddev lock, so rcu protection not needed.
1115 for (i = 0; i < conf->raid_disks; i++) {
1116 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1118 && !test_bit(Faulty, &rdev->flags)
1119 && !test_and_set_bit(In_sync, &rdev->flags)) {
1120 unsigned long flags;
1121 spin_lock_irqsave(&conf->device_lock, flags);
1123 spin_unlock_irqrestore(&conf->device_lock, flags);
1132 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1134 conf_t *conf = mddev->private;
1139 int last = mddev->raid_disks - 1;
1141 if (rdev->raid_disk >= 0)
1142 first = last = rdev->raid_disk;
1144 for (mirror = first; mirror <= last; mirror++)
1145 if ( !(p=conf->mirrors+mirror)->rdev) {
1147 disk_stack_limits(mddev->gendisk, rdev->bdev,
1148 rdev->data_offset << 9);
1149 /* as we don't honour merge_bvec_fn, we must
1150 * never risk violating it, so limit
1151 * ->max_segments to one lying with a single
1152 * page, as a one page request is never in
1155 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1156 blk_queue_max_segments(mddev->queue, 1);
1157 blk_queue_segment_boundary(mddev->queue,
1158 PAGE_CACHE_SIZE - 1);
1161 p->head_position = 0;
1162 rdev->raid_disk = mirror;
1164 /* As all devices are equivalent, we don't need a full recovery
1165 * if this was recently any drive of the array
1167 if (rdev->saved_raid_disk < 0)
1169 rcu_assign_pointer(p->rdev, rdev);
1172 md_integrity_add_rdev(rdev, mddev);
1177 static int raid1_remove_disk(mddev_t *mddev, int number)
1179 conf_t *conf = mddev->private;
1182 mirror_info_t *p = conf->mirrors+ number;
1187 if (test_bit(In_sync, &rdev->flags) ||
1188 atomic_read(&rdev->nr_pending)) {
1192 /* Only remove non-faulty devices is recovery
1195 if (!test_bit(Faulty, &rdev->flags) &&
1196 mddev->degraded < conf->raid_disks) {
1202 if (atomic_read(&rdev->nr_pending)) {
1203 /* lost the race, try later */
1208 md_integrity_register(mddev);
1217 static void end_sync_read(struct bio *bio, int error)
1219 r1bio_t *r1_bio = bio->bi_private;
1222 for (i=r1_bio->mddev->raid_disks; i--; )
1223 if (r1_bio->bios[i] == bio)
1226 update_head_pos(i, r1_bio);
1228 * we have read a block, now it needs to be re-written,
1229 * or re-read if the read failed.
1230 * We don't do much here, just schedule handling by raid1d
1232 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1233 set_bit(R1BIO_Uptodate, &r1_bio->state);
1235 if (atomic_dec_and_test(&r1_bio->remaining))
1236 reschedule_retry(r1_bio);
1239 static void end_sync_write(struct bio *bio, int error)
1241 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1242 r1bio_t *r1_bio = bio->bi_private;
1243 mddev_t *mddev = r1_bio->mddev;
1244 conf_t *conf = mddev->private;
1248 for (i = 0; i < conf->raid_disks; i++)
1249 if (r1_bio->bios[i] == bio) {
1254 int sync_blocks = 0;
1255 sector_t s = r1_bio->sector;
1256 long sectors_to_go = r1_bio->sectors;
1257 /* make sure these bits doesn't get cleared. */
1259 bitmap_end_sync(mddev->bitmap, s,
1262 sectors_to_go -= sync_blocks;
1263 } while (sectors_to_go > 0);
1264 md_error(mddev, conf->mirrors[mirror].rdev);
1267 update_head_pos(mirror, r1_bio);
1269 if (atomic_dec_and_test(&r1_bio->remaining)) {
1270 sector_t s = r1_bio->sectors;
1272 md_done_sync(mddev, s, uptodate);
1276 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1278 conf_t *conf = mddev->private;
1280 int disks = conf->raid_disks;
1281 struct bio *bio, *wbio;
1283 bio = r1_bio->bios[r1_bio->read_disk];
1286 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1287 /* We have read all readable devices. If we haven't
1288 * got the block, then there is no hope left.
1289 * If we have, then we want to do a comparison
1290 * and skip the write if everything is the same.
1291 * If any blocks failed to read, then we need to
1292 * attempt an over-write
1295 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1296 for (i=0; i<mddev->raid_disks; i++)
1297 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1298 md_error(mddev, conf->mirrors[i].rdev);
1300 md_done_sync(mddev, r1_bio->sectors, 1);
1304 for (primary=0; primary<mddev->raid_disks; primary++)
1305 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1306 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1307 r1_bio->bios[primary]->bi_end_io = NULL;
1308 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1311 r1_bio->read_disk = primary;
1312 for (i=0; i<mddev->raid_disks; i++)
1313 if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1315 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1316 struct bio *pbio = r1_bio->bios[primary];
1317 struct bio *sbio = r1_bio->bios[i];
1319 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1320 for (j = vcnt; j-- ; ) {
1322 p = pbio->bi_io_vec[j].bv_page;
1323 s = sbio->bi_io_vec[j].bv_page;
1324 if (memcmp(page_address(p),
1332 mddev->resync_mismatches += r1_bio->sectors;
1333 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1334 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1335 sbio->bi_end_io = NULL;
1336 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1338 /* fixup the bio for reuse */
1340 sbio->bi_vcnt = vcnt;
1341 sbio->bi_size = r1_bio->sectors << 9;
1343 sbio->bi_phys_segments = 0;
1344 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1345 sbio->bi_flags |= 1 << BIO_UPTODATE;
1346 sbio->bi_next = NULL;
1347 sbio->bi_sector = r1_bio->sector +
1348 conf->mirrors[i].rdev->data_offset;
1349 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1350 size = sbio->bi_size;
1351 for (j = 0; j < vcnt ; j++) {
1353 bi = &sbio->bi_io_vec[j];
1355 if (size > PAGE_SIZE)
1356 bi->bv_len = PAGE_SIZE;
1360 memcpy(page_address(bi->bv_page),
1361 page_address(pbio->bi_io_vec[j].bv_page),
1368 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1369 /* ouch - failed to read all of that.
1370 * Try some synchronous reads of other devices to get
1371 * good data, much like with normal read errors. Only
1372 * read into the pages we already have so we don't
1373 * need to re-issue the read request.
1374 * We don't need to freeze the array, because being in an
1375 * active sync request, there is no normal IO, and
1376 * no overlapping syncs.
1378 sector_t sect = r1_bio->sector;
1379 int sectors = r1_bio->sectors;
1384 int d = r1_bio->read_disk;
1388 if (s > (PAGE_SIZE>>9))
1391 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1392 /* No rcu protection needed here devices
1393 * can only be removed when no resync is
1394 * active, and resync is currently active
1396 rdev = conf->mirrors[d].rdev;
1397 if (sync_page_io(rdev->bdev,
1398 sect + rdev->data_offset,
1400 bio->bi_io_vec[idx].bv_page,
1407 if (d == conf->raid_disks)
1409 } while (!success && d != r1_bio->read_disk);
1413 /* write it back and re-read */
1414 set_bit(R1BIO_Uptodate, &r1_bio->state);
1415 while (d != r1_bio->read_disk) {
1417 d = conf->raid_disks;
1419 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1421 rdev = conf->mirrors[d].rdev;
1422 atomic_add(s, &rdev->corrected_errors);
1423 if (sync_page_io(rdev->bdev,
1424 sect + rdev->data_offset,
1426 bio->bi_io_vec[idx].bv_page,
1428 md_error(mddev, rdev);
1431 while (d != r1_bio->read_disk) {
1433 d = conf->raid_disks;
1435 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1437 rdev = conf->mirrors[d].rdev;
1438 if (sync_page_io(rdev->bdev,
1439 sect + rdev->data_offset,
1441 bio->bi_io_vec[idx].bv_page,
1443 md_error(mddev, rdev);
1446 char b[BDEVNAME_SIZE];
1447 /* Cannot read from anywhere, array is toast */
1448 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1449 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1450 " for block %llu\n",
1451 bdevname(bio->bi_bdev,b),
1452 (unsigned long long)r1_bio->sector);
1453 md_done_sync(mddev, r1_bio->sectors, 0);
1466 atomic_set(&r1_bio->remaining, 1);
1467 for (i = 0; i < disks ; i++) {
1468 wbio = r1_bio->bios[i];
1469 if (wbio->bi_end_io == NULL ||
1470 (wbio->bi_end_io == end_sync_read &&
1471 (i == r1_bio->read_disk ||
1472 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1475 wbio->bi_rw = WRITE;
1476 wbio->bi_end_io = end_sync_write;
1477 atomic_inc(&r1_bio->remaining);
1478 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1480 generic_make_request(wbio);
1483 if (atomic_dec_and_test(&r1_bio->remaining)) {
1484 /* if we're here, all write(s) have completed, so clean up */
1485 md_done_sync(mddev, r1_bio->sectors, 1);
1491 * This is a kernel thread which:
1493 * 1. Retries failed read operations on working mirrors.
1494 * 2. Updates the raid superblock when problems encounter.
1495 * 3. Performs writes following reads for array syncronising.
1498 static void fix_read_error(conf_t *conf, int read_disk,
1499 sector_t sect, int sectors)
1501 mddev_t *mddev = conf->mddev;
1509 if (s > (PAGE_SIZE>>9))
1513 /* Note: no rcu protection needed here
1514 * as this is synchronous in the raid1d thread
1515 * which is the thread that might remove
1516 * a device. If raid1d ever becomes multi-threaded....
1518 rdev = conf->mirrors[d].rdev;
1520 test_bit(In_sync, &rdev->flags) &&
1521 sync_page_io(rdev->bdev,
1522 sect + rdev->data_offset,
1524 conf->tmppage, READ))
1528 if (d == conf->raid_disks)
1531 } while (!success && d != read_disk);
1534 /* Cannot read from anywhere -- bye bye array */
1535 md_error(mddev, conf->mirrors[read_disk].rdev);
1538 /* write it back and re-read */
1540 while (d != read_disk) {
1542 d = conf->raid_disks;
1544 rdev = conf->mirrors[d].rdev;
1546 test_bit(In_sync, &rdev->flags)) {
1547 if (sync_page_io(rdev->bdev,
1548 sect + rdev->data_offset,
1549 s<<9, conf->tmppage, WRITE)
1551 /* Well, this device is dead */
1552 md_error(mddev, rdev);
1556 while (d != read_disk) {
1557 char b[BDEVNAME_SIZE];
1559 d = conf->raid_disks;
1561 rdev = conf->mirrors[d].rdev;
1563 test_bit(In_sync, &rdev->flags)) {
1564 if (sync_page_io(rdev->bdev,
1565 sect + rdev->data_offset,
1566 s<<9, conf->tmppage, READ)
1568 /* Well, this device is dead */
1569 md_error(mddev, rdev);
1571 atomic_add(s, &rdev->corrected_errors);
1573 "raid1:%s: read error corrected "
1574 "(%d sectors at %llu on %s)\n",
1576 (unsigned long long)(sect +
1578 bdevname(rdev->bdev, b));
1587 static void raid1d(mddev_t *mddev)
1591 unsigned long flags;
1592 conf_t *conf = mddev->private;
1593 struct list_head *head = &conf->retry_list;
1597 md_check_recovery(mddev);
1600 char b[BDEVNAME_SIZE];
1602 unplug += flush_pending_writes(conf);
1604 spin_lock_irqsave(&conf->device_lock, flags);
1605 if (list_empty(head)) {
1606 spin_unlock_irqrestore(&conf->device_lock, flags);
1609 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1610 list_del(head->prev);
1612 spin_unlock_irqrestore(&conf->device_lock, flags);
1614 mddev = r1_bio->mddev;
1615 conf = mddev->private;
1616 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1617 sync_request_write(mddev, r1_bio);
1619 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1620 /* some requests in the r1bio were BIO_RW_BARRIER
1621 * requests which failed with -EOPNOTSUPP. Hohumm..
1622 * Better resubmit without the barrier.
1623 * We know which devices to resubmit for, because
1624 * all others have had their bios[] entry cleared.
1625 * We already have a nr_pending reference on these rdevs.
1628 const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1629 clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1630 clear_bit(R1BIO_Barrier, &r1_bio->state);
1631 for (i=0; i < conf->raid_disks; i++)
1632 if (r1_bio->bios[i])
1633 atomic_inc(&r1_bio->remaining);
1634 for (i=0; i < conf->raid_disks; i++)
1635 if (r1_bio->bios[i]) {
1636 struct bio_vec *bvec;
1639 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1640 /* copy pages from the failed bio, as
1641 * this might be a write-behind device */
1642 __bio_for_each_segment(bvec, bio, j, 0)
1643 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1644 bio_put(r1_bio->bios[i]);
1645 bio->bi_sector = r1_bio->sector +
1646 conf->mirrors[i].rdev->data_offset;
1647 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1648 bio->bi_end_io = raid1_end_write_request;
1649 bio->bi_rw = WRITE |
1650 (do_sync << BIO_RW_SYNCIO);
1651 bio->bi_private = r1_bio;
1652 r1_bio->bios[i] = bio;
1653 generic_make_request(bio);
1658 /* we got a read error. Maybe the drive is bad. Maybe just
1659 * the block and we can fix it.
1660 * We freeze all other IO, and try reading the block from
1661 * other devices. When we find one, we re-write
1662 * and check it that fixes the read error.
1663 * This is all done synchronously while the array is
1666 if (mddev->ro == 0) {
1668 fix_read_error(conf, r1_bio->read_disk,
1671 unfreeze_array(conf);
1674 conf->mirrors[r1_bio->read_disk].rdev);
1676 bio = r1_bio->bios[r1_bio->read_disk];
1677 if ((disk=read_balance(conf, r1_bio)) == -1) {
1678 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1679 " read error for block %llu\n",
1680 bdevname(bio->bi_bdev,b),
1681 (unsigned long long)r1_bio->sector);
1682 raid_end_bio_io(r1_bio);
1684 const bool do_sync = bio_rw_flagged(r1_bio->master_bio, BIO_RW_SYNCIO);
1685 r1_bio->bios[r1_bio->read_disk] =
1686 mddev->ro ? IO_BLOCKED : NULL;
1687 r1_bio->read_disk = disk;
1689 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1690 r1_bio->bios[r1_bio->read_disk] = bio;
1691 rdev = conf->mirrors[disk].rdev;
1692 if (printk_ratelimit())
1693 printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1694 " another mirror\n",
1695 bdevname(rdev->bdev,b),
1696 (unsigned long long)r1_bio->sector);
1697 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1698 bio->bi_bdev = rdev->bdev;
1699 bio->bi_end_io = raid1_end_read_request;
1700 bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
1701 bio->bi_private = r1_bio;
1703 generic_make_request(bio);
1709 unplug_slaves(mddev);
1713 static int init_resync(conf_t *conf)
1717 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1718 BUG_ON(conf->r1buf_pool);
1719 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1721 if (!conf->r1buf_pool)
1723 conf->next_resync = 0;
1728 * perform a "sync" on one "block"
1730 * We need to make sure that no normal I/O request - particularly write
1731 * requests - conflict with active sync requests.
1733 * This is achieved by tracking pending requests and a 'barrier' concept
1734 * that can be installed to exclude normal IO requests.
1737 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1739 conf_t *conf = mddev->private;
1742 sector_t max_sector, nr_sectors;
1746 int write_targets = 0, read_targets = 0;
1748 int still_degraded = 0;
1750 if (!conf->r1buf_pool)
1753 printk("sync start - bitmap %p\n", mddev->bitmap);
1755 if (init_resync(conf))
1759 max_sector = mddev->dev_sectors;
1760 if (sector_nr >= max_sector) {
1761 /* If we aborted, we need to abort the
1762 * sync on the 'current' bitmap chunk (there will
1763 * only be one in raid1 resync.
1764 * We can find the current addess in mddev->curr_resync
1766 if (mddev->curr_resync < max_sector) /* aborted */
1767 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1769 else /* completed sync */
1772 bitmap_close_sync(mddev->bitmap);
1777 if (mddev->bitmap == NULL &&
1778 mddev->recovery_cp == MaxSector &&
1779 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1780 conf->fullsync == 0) {
1782 return max_sector - sector_nr;
1784 /* before building a request, check if we can skip these blocks..
1785 * This call the bitmap_start_sync doesn't actually record anything
1787 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1788 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1789 /* We can skip this block, and probably several more */
1794 * If there is non-resync activity waiting for a turn,
1795 * and resync is going fast enough,
1796 * then let it though before starting on this new sync request.
1798 if (!go_faster && conf->nr_waiting)
1799 msleep_interruptible(1000);
1801 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1802 raise_barrier(conf);
1804 conf->next_resync = sector_nr;
1806 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1809 * If we get a correctably read error during resync or recovery,
1810 * we might want to read from a different device. So we
1811 * flag all drives that could conceivably be read from for READ,
1812 * and any others (which will be non-In_sync devices) for WRITE.
1813 * If a read fails, we try reading from something else for which READ
1817 r1_bio->mddev = mddev;
1818 r1_bio->sector = sector_nr;
1820 set_bit(R1BIO_IsSync, &r1_bio->state);
1822 for (i=0; i < conf->raid_disks; i++) {
1824 bio = r1_bio->bios[i];
1826 /* take from bio_init */
1827 bio->bi_next = NULL;
1828 bio->bi_flags |= 1 << BIO_UPTODATE;
1832 bio->bi_phys_segments = 0;
1834 bio->bi_end_io = NULL;
1835 bio->bi_private = NULL;
1837 rdev = rcu_dereference(conf->mirrors[i].rdev);
1839 test_bit(Faulty, &rdev->flags)) {
1842 } else if (!test_bit(In_sync, &rdev->flags)) {
1844 bio->bi_end_io = end_sync_write;
1847 /* may need to read from here */
1849 bio->bi_end_io = end_sync_read;
1850 if (test_bit(WriteMostly, &rdev->flags)) {
1859 atomic_inc(&rdev->nr_pending);
1860 bio->bi_sector = sector_nr + rdev->data_offset;
1861 bio->bi_bdev = rdev->bdev;
1862 bio->bi_private = r1_bio;
1867 r1_bio->read_disk = disk;
1869 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1870 /* extra read targets are also write targets */
1871 write_targets += read_targets-1;
1873 if (write_targets == 0 || read_targets == 0) {
1874 /* There is nowhere to write, so all non-sync
1875 * drives must be failed - so we are finished
1877 sector_t rv = max_sector - sector_nr;
1883 if (max_sector > mddev->resync_max)
1884 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1889 int len = PAGE_SIZE;
1890 if (sector_nr + (len>>9) > max_sector)
1891 len = (max_sector - sector_nr) << 9;
1894 if (sync_blocks == 0) {
1895 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1896 &sync_blocks, still_degraded) &&
1898 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1900 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1901 if (len > (sync_blocks<<9))
1902 len = sync_blocks<<9;
1905 for (i=0 ; i < conf->raid_disks; i++) {
1906 bio = r1_bio->bios[i];
1907 if (bio->bi_end_io) {
1908 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1909 if (bio_add_page(bio, page, len, 0) == 0) {
1911 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1914 bio = r1_bio->bios[i];
1915 if (bio->bi_end_io==NULL)
1917 /* remove last page from this bio */
1919 bio->bi_size -= len;
1920 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1926 nr_sectors += len>>9;
1927 sector_nr += len>>9;
1928 sync_blocks -= (len>>9);
1929 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1931 r1_bio->sectors = nr_sectors;
1933 /* For a user-requested sync, we read all readable devices and do a
1936 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1937 atomic_set(&r1_bio->remaining, read_targets);
1938 for (i=0; i<conf->raid_disks; i++) {
1939 bio = r1_bio->bios[i];
1940 if (bio->bi_end_io == end_sync_read) {
1941 md_sync_acct(bio->bi_bdev, nr_sectors);
1942 generic_make_request(bio);
1946 atomic_set(&r1_bio->remaining, 1);
1947 bio = r1_bio->bios[r1_bio->read_disk];
1948 md_sync_acct(bio->bi_bdev, nr_sectors);
1949 generic_make_request(bio);
1955 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1960 return mddev->dev_sectors;
1963 static conf_t *setup_conf(mddev_t *mddev)
1967 mirror_info_t *disk;
1971 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1975 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1980 conf->tmppage = alloc_page(GFP_KERNEL);
1984 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1985 if (!conf->poolinfo)
1987 conf->poolinfo->raid_disks = mddev->raid_disks;
1988 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1991 if (!conf->r1bio_pool)
1994 conf->poolinfo->mddev = mddev;
1996 spin_lock_init(&conf->device_lock);
1997 list_for_each_entry(rdev, &mddev->disks, same_set) {
1998 int disk_idx = rdev->raid_disk;
1999 if (disk_idx >= mddev->raid_disks
2002 disk = conf->mirrors + disk_idx;
2006 disk->head_position = 0;
2008 conf->raid_disks = mddev->raid_disks;
2009 conf->mddev = mddev;
2010 INIT_LIST_HEAD(&conf->retry_list);
2012 spin_lock_init(&conf->resync_lock);
2013 init_waitqueue_head(&conf->wait_barrier);
2015 bio_list_init(&conf->pending_bio_list);
2016 bio_list_init(&conf->flushing_bio_list);
2018 conf->last_used = -1;
2019 for (i = 0; i < conf->raid_disks; i++) {
2021 disk = conf->mirrors + i;
2024 !test_bit(In_sync, &disk->rdev->flags)) {
2025 disk->head_position = 0;
2028 } else if (conf->last_used < 0)
2030 * The first working device is used as a
2031 * starting point to read balancing.
2033 conf->last_used = i;
2037 if (conf->last_used < 0) {
2038 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
2043 conf->thread = md_register_thread(raid1d, mddev, NULL);
2044 if (!conf->thread) {
2046 "raid1: couldn't allocate thread for %s\n",
2055 if (conf->r1bio_pool)
2056 mempool_destroy(conf->r1bio_pool);
2057 kfree(conf->mirrors);
2058 safe_put_page(conf->tmppage);
2059 kfree(conf->poolinfo);
2062 return ERR_PTR(err);
2065 static int run(mddev_t *mddev)
2071 if (mddev->level != 1) {
2072 printk("raid1: %s: raid level not set to mirroring (%d)\n",
2073 mdname(mddev), mddev->level);
2076 if (mddev->reshape_position != MaxSector) {
2077 printk("raid1: %s: reshape_position set but not supported\n",
2082 * copy the already verified devices into our private RAID1
2083 * bookkeeping area. [whatever we allocate in run(),
2084 * should be freed in stop()]
2086 if (mddev->private == NULL)
2087 conf = setup_conf(mddev);
2089 conf = mddev->private;
2092 return PTR_ERR(conf);
2094 mddev->queue->queue_lock = &conf->device_lock;
2095 list_for_each_entry(rdev, &mddev->disks, same_set) {
2096 disk_stack_limits(mddev->gendisk, rdev->bdev,
2097 rdev->data_offset << 9);
2098 /* as we don't honour merge_bvec_fn, we must never risk
2099 * violating it, so limit ->max_segments to 1 lying within
2100 * a single page, as a one page request is never in violation.
2102 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2103 blk_queue_max_segments(mddev->queue, 1);
2104 blk_queue_segment_boundary(mddev->queue,
2105 PAGE_CACHE_SIZE - 1);
2109 mddev->degraded = 0;
2110 for (i=0; i < conf->raid_disks; i++)
2111 if (conf->mirrors[i].rdev == NULL ||
2112 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2113 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2116 if (conf->raid_disks - mddev->degraded == 1)
2117 mddev->recovery_cp = MaxSector;
2119 if (mddev->recovery_cp != MaxSector)
2120 printk(KERN_NOTICE "raid1: %s is not clean"
2121 " -- starting background reconstruction\n",
2124 "raid1: raid set %s active with %d out of %d mirrors\n",
2125 mdname(mddev), mddev->raid_disks - mddev->degraded,
2129 * Ok, everything is just fine now
2131 mddev->thread = conf->thread;
2132 conf->thread = NULL;
2133 mddev->private = conf;
2135 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2137 mddev->queue->unplug_fn = raid1_unplug;
2138 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2139 mddev->queue->backing_dev_info.congested_data = mddev;
2140 md_integrity_register(mddev);
2144 static int stop(mddev_t *mddev)
2146 conf_t *conf = mddev->private;
2147 struct bitmap *bitmap = mddev->bitmap;
2148 int behind_wait = 0;
2150 /* wait for behind writes to complete */
2151 while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2153 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2154 set_current_state(TASK_UNINTERRUPTIBLE);
2155 schedule_timeout(HZ); /* wait a second */
2156 /* need to kick something here to make sure I/O goes? */
2159 raise_barrier(conf);
2160 lower_barrier(conf);
2162 md_unregister_thread(mddev->thread);
2163 mddev->thread = NULL;
2164 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2165 if (conf->r1bio_pool)
2166 mempool_destroy(conf->r1bio_pool);
2167 kfree(conf->mirrors);
2168 kfree(conf->poolinfo);
2170 mddev->private = NULL;
2174 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2176 /* no resync is happening, and there is enough space
2177 * on all devices, so we can resize.
2178 * We need to make sure resync covers any new space.
2179 * If the array is shrinking we should possibly wait until
2180 * any io in the removed space completes, but it hardly seems
2183 md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2184 if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2186 set_capacity(mddev->gendisk, mddev->array_sectors);
2188 revalidate_disk(mddev->gendisk);
2189 if (sectors > mddev->dev_sectors &&
2190 mddev->recovery_cp == MaxSector) {
2191 mddev->recovery_cp = mddev->dev_sectors;
2192 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2194 mddev->dev_sectors = sectors;
2195 mddev->resync_max_sectors = sectors;
2199 static int raid1_reshape(mddev_t *mddev)
2202 * 1/ resize the r1bio_pool
2203 * 2/ resize conf->mirrors
2205 * We allocate a new r1bio_pool if we can.
2206 * Then raise a device barrier and wait until all IO stops.
2207 * Then resize conf->mirrors and swap in the new r1bio pool.
2209 * At the same time, we "pack" the devices so that all the missing
2210 * devices have the higher raid_disk numbers.
2212 mempool_t *newpool, *oldpool;
2213 struct pool_info *newpoolinfo;
2214 mirror_info_t *newmirrors;
2215 conf_t *conf = mddev->private;
2216 int cnt, raid_disks;
2217 unsigned long flags;
2220 /* Cannot change chunk_size, layout, or level */
2221 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2222 mddev->layout != mddev->new_layout ||
2223 mddev->level != mddev->new_level) {
2224 mddev->new_chunk_sectors = mddev->chunk_sectors;
2225 mddev->new_layout = mddev->layout;
2226 mddev->new_level = mddev->level;
2230 err = md_allow_write(mddev);
2234 raid_disks = mddev->raid_disks + mddev->delta_disks;
2236 if (raid_disks < conf->raid_disks) {
2238 for (d= 0; d < conf->raid_disks; d++)
2239 if (conf->mirrors[d].rdev)
2241 if (cnt > raid_disks)
2245 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2248 newpoolinfo->mddev = mddev;
2249 newpoolinfo->raid_disks = raid_disks;
2251 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2252 r1bio_pool_free, newpoolinfo);
2257 newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2260 mempool_destroy(newpool);
2264 raise_barrier(conf);
2266 /* ok, everything is stopped */
2267 oldpool = conf->r1bio_pool;
2268 conf->r1bio_pool = newpool;
2270 for (d = d2 = 0; d < conf->raid_disks; d++) {
2271 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2272 if (rdev && rdev->raid_disk != d2) {
2274 sprintf(nm, "rd%d", rdev->raid_disk);
2275 sysfs_remove_link(&mddev->kobj, nm);
2276 rdev->raid_disk = d2;
2277 sprintf(nm, "rd%d", rdev->raid_disk);
2278 sysfs_remove_link(&mddev->kobj, nm);
2279 if (sysfs_create_link(&mddev->kobj,
2282 "md/raid1: cannot register "
2287 newmirrors[d2++].rdev = rdev;
2289 kfree(conf->mirrors);
2290 conf->mirrors = newmirrors;
2291 kfree(conf->poolinfo);
2292 conf->poolinfo = newpoolinfo;
2294 spin_lock_irqsave(&conf->device_lock, flags);
2295 mddev->degraded += (raid_disks - conf->raid_disks);
2296 spin_unlock_irqrestore(&conf->device_lock, flags);
2297 conf->raid_disks = mddev->raid_disks = raid_disks;
2298 mddev->delta_disks = 0;
2300 conf->last_used = 0; /* just make sure it is in-range */
2301 lower_barrier(conf);
2303 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2304 md_wakeup_thread(mddev->thread);
2306 mempool_destroy(oldpool);
2310 static void raid1_quiesce(mddev_t *mddev, int state)
2312 conf_t *conf = mddev->private;
2315 case 2: /* wake for suspend */
2316 wake_up(&conf->wait_barrier);
2319 raise_barrier(conf);
2322 lower_barrier(conf);
2327 static void *raid1_takeover(mddev_t *mddev)
2329 /* raid1 can take over:
2330 * raid5 with 2 devices, any layout or chunk size
2332 if (mddev->level == 5 && mddev->raid_disks == 2) {
2334 mddev->new_level = 1;
2335 mddev->new_layout = 0;
2336 mddev->new_chunk_sectors = 0;
2337 conf = setup_conf(mddev);
2342 return ERR_PTR(-EINVAL);
2345 static struct mdk_personality raid1_personality =
2349 .owner = THIS_MODULE,
2350 .make_request = make_request,
2354 .error_handler = error,
2355 .hot_add_disk = raid1_add_disk,
2356 .hot_remove_disk= raid1_remove_disk,
2357 .spare_active = raid1_spare_active,
2358 .sync_request = sync_request,
2359 .resize = raid1_resize,
2361 .check_reshape = raid1_reshape,
2362 .quiesce = raid1_quiesce,
2363 .takeover = raid1_takeover,
2366 static int __init raid_init(void)
2368 return register_md_personality(&raid1_personality);
2371 static void raid_exit(void)
2373 unregister_md_personality(&raid1_personality);
2376 module_init(raid_init);
2377 module_exit(raid_exit);
2378 MODULE_LICENSE("GPL");
2379 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2380 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2381 MODULE_ALIAS("md-raid1");
2382 MODULE_ALIAS("md-level-1");