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/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/seq_file.h>
44 #define PRINTK(x...) printk(x)
50 * Number of guaranteed r1bios in case of extreme VM load:
52 #define NR_RAID1_BIOS 256
55 static void unplug_slaves(mddev_t *mddev);
57 static void allow_barrier(conf_t *conf);
58 static void lower_barrier(conf_t *conf);
60 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
62 struct pool_info *pi = data;
64 int size = offsetof(r1bio_t, bios[pi->raid_disks]);
66 /* allocate a r1bio with room for raid_disks entries in the bios array */
67 r1_bio = kzalloc(size, gfp_flags);
68 if (!r1_bio && pi->mddev)
69 unplug_slaves(pi->mddev);
74 static void r1bio_pool_free(void *r1_bio, void *data)
79 #define RESYNC_BLOCK_SIZE (64*1024)
80 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
81 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
82 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
83 #define RESYNC_WINDOW (2048*1024)
85 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
87 struct pool_info *pi = data;
93 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
95 unplug_slaves(pi->mddev);
100 * Allocate bios : 1 for reading, n-1 for writing
102 for (j = pi->raid_disks ; j-- ; ) {
103 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
106 r1_bio->bios[j] = bio;
109 * Allocate RESYNC_PAGES data pages and attach them to
111 * If this is a user-requested check/repair, allocate
112 * RESYNC_PAGES for each bio.
114 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
119 bio = r1_bio->bios[j];
120 for (i = 0; i < RESYNC_PAGES; i++) {
121 page = alloc_page(gfp_flags);
125 bio->bi_io_vec[i].bv_page = page;
129 /* If not user-requests, copy the page pointers to all bios */
130 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
131 for (i=0; i<RESYNC_PAGES ; i++)
132 for (j=1; j<pi->raid_disks; j++)
133 r1_bio->bios[j]->bi_io_vec[i].bv_page =
134 r1_bio->bios[0]->bi_io_vec[i].bv_page;
137 r1_bio->master_bio = NULL;
142 for (j=0 ; j < pi->raid_disks; j++)
143 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
144 put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
147 while ( ++j < pi->raid_disks )
148 bio_put(r1_bio->bios[j]);
149 r1bio_pool_free(r1_bio, data);
153 static void r1buf_pool_free(void *__r1_bio, void *data)
155 struct pool_info *pi = data;
157 r1bio_t *r1bio = __r1_bio;
159 for (i = 0; i < RESYNC_PAGES; i++)
160 for (j = pi->raid_disks; j-- ;) {
162 r1bio->bios[j]->bi_io_vec[i].bv_page !=
163 r1bio->bios[0]->bi_io_vec[i].bv_page)
164 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
166 for (i=0 ; i < pi->raid_disks; i++)
167 bio_put(r1bio->bios[i]);
169 r1bio_pool_free(r1bio, data);
172 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
176 for (i = 0; i < conf->raid_disks; i++) {
177 struct bio **bio = r1_bio->bios + i;
178 if (*bio && *bio != IO_BLOCKED)
184 static void free_r1bio(r1bio_t *r1_bio)
186 conf_t *conf = r1_bio->mddev->private;
189 * Wake up any possible resync thread that waits for the device
194 put_all_bios(conf, r1_bio);
195 mempool_free(r1_bio, conf->r1bio_pool);
198 static void put_buf(r1bio_t *r1_bio)
200 conf_t *conf = r1_bio->mddev->private;
203 for (i=0; i<conf->raid_disks; i++) {
204 struct bio *bio = r1_bio->bios[i];
206 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
209 mempool_free(r1_bio, conf->r1buf_pool);
214 static void reschedule_retry(r1bio_t *r1_bio)
217 mddev_t *mddev = r1_bio->mddev;
218 conf_t *conf = mddev->private;
220 spin_lock_irqsave(&conf->device_lock, flags);
221 list_add(&r1_bio->retry_list, &conf->retry_list);
223 spin_unlock_irqrestore(&conf->device_lock, flags);
225 wake_up(&conf->wait_barrier);
226 md_wakeup_thread(mddev->thread);
230 * raid_end_bio_io() is called when we have finished servicing a mirrored
231 * operation and are ready to return a success/failure code to the buffer
234 static void raid_end_bio_io(r1bio_t *r1_bio)
236 struct bio *bio = r1_bio->master_bio;
238 /* if nobody has done the final endio yet, do it now */
239 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
240 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
241 (bio_data_dir(bio) == WRITE) ? "write" : "read",
242 (unsigned long long) bio->bi_sector,
243 (unsigned long long) bio->bi_sector +
244 (bio->bi_size >> 9) - 1);
247 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
253 * Update disk head position estimator based on IRQ completion info.
255 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
257 conf_t *conf = r1_bio->mddev->private;
259 conf->mirrors[disk].head_position =
260 r1_bio->sector + (r1_bio->sectors);
263 static void raid1_end_read_request(struct bio *bio, int error)
265 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
266 r1bio_t *r1_bio = bio->bi_private;
268 conf_t *conf = r1_bio->mddev->private;
270 mirror = r1_bio->read_disk;
272 * this branch is our 'one mirror IO has finished' event handler:
274 update_head_pos(mirror, r1_bio);
277 set_bit(R1BIO_Uptodate, &r1_bio->state);
279 /* If all other devices have failed, we want to return
280 * the error upwards rather than fail the last device.
281 * Here we redefine "uptodate" to mean "Don't want to retry"
284 spin_lock_irqsave(&conf->device_lock, flags);
285 if (r1_bio->mddev->degraded == conf->raid_disks ||
286 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
287 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
289 spin_unlock_irqrestore(&conf->device_lock, flags);
293 raid_end_bio_io(r1_bio);
298 char b[BDEVNAME_SIZE];
299 if (printk_ratelimit())
300 printk(KERN_ERR "md/raid1:%s: %s: rescheduling sector %llu\n",
302 bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
303 reschedule_retry(r1_bio);
306 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
309 static void r1_bio_write_done(r1bio_t *r1_bio, int vcnt, struct bio_vec *bv,
312 if (atomic_dec_and_test(&r1_bio->remaining))
314 /* it really is the end of this request */
315 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
316 /* free extra copy of the data pages */
319 safe_put_page(bv[i].bv_page);
321 /* clear the bitmap if all writes complete successfully */
322 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
324 !test_bit(R1BIO_Degraded, &r1_bio->state),
326 md_write_end(r1_bio->mddev);
327 raid_end_bio_io(r1_bio);
331 static void raid1_end_write_request(struct bio *bio, int error)
333 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
334 r1bio_t *r1_bio = bio->bi_private;
335 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
336 conf_t *conf = r1_bio->mddev->private;
337 struct bio *to_put = NULL;
340 for (mirror = 0; mirror < conf->raid_disks; mirror++)
341 if (r1_bio->bios[mirror] == bio)
345 * 'one mirror IO has finished' event handler:
347 r1_bio->bios[mirror] = NULL;
350 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
351 /* an I/O failed, we can't clear the bitmap */
352 set_bit(R1BIO_Degraded, &r1_bio->state);
355 * Set R1BIO_Uptodate in our master bio, so that we
356 * will return a good error code for to the higher
357 * levels even if IO on some other mirrored buffer
360 * The 'master' represents the composite IO operation
361 * to user-side. So if something waits for IO, then it
362 * will wait for the 'master' bio.
364 set_bit(R1BIO_Uptodate, &r1_bio->state);
366 update_head_pos(mirror, r1_bio);
369 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
370 atomic_dec(&r1_bio->behind_remaining);
373 * In behind mode, we ACK the master bio once the I/O
374 * has safely reached all non-writemostly
375 * disks. Setting the Returned bit ensures that this
376 * gets done only once -- we don't ever want to return
377 * -EIO here, instead we'll wait
379 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
380 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
381 /* Maybe we can return now */
382 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
383 struct bio *mbio = r1_bio->master_bio;
384 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
385 (unsigned long long) mbio->bi_sector,
386 (unsigned long long) mbio->bi_sector +
387 (mbio->bi_size >> 9) - 1);
392 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
395 * Let's see if all mirrored write operations have finished
398 r1_bio_write_done(r1_bio, bio->bi_vcnt, bio->bi_io_vec, behind);
406 * This routine returns the disk from which the requested read should
407 * be done. There is a per-array 'next expected sequential IO' sector
408 * number - if this matches on the next IO then we use the last disk.
409 * There is also a per-disk 'last know head position' sector that is
410 * maintained from IRQ contexts, both the normal and the resync IO
411 * completion handlers update this position correctly. If there is no
412 * perfect sequential match then we pick the disk whose head is closest.
414 * If there are 2 mirrors in the same 2 devices, performance degrades
415 * because position is mirror, not device based.
417 * The rdev for the device selected will have nr_pending incremented.
419 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
421 const sector_t this_sector = r1_bio->sector;
422 const int sectors = r1_bio->sectors;
426 sector_t new_distance, current_distance;
432 * Check if we can balance. We can balance on the whole
433 * device if no resync is going on, or below the resync window.
434 * We take the first readable disk when above the resync window.
437 if (conf->mddev->recovery_cp < MaxSector &&
438 (this_sector + sectors >= conf->next_resync)) {
443 start_disk = conf->last_used;
446 /* make sure the disk is operational */
447 for (i = 0 ; i < conf->raid_disks ; i++) {
448 int disk = start_disk + i;
449 if (disk >= conf->raid_disks)
450 disk -= conf->raid_disks;
452 rdev = rcu_dereference(conf->mirrors[disk].rdev);
453 if (r1_bio->bios[disk] == IO_BLOCKED
455 || !test_bit(In_sync, &rdev->flags))
459 if (!test_bit(WriteMostly, &rdev->flags))
463 if (new_disk < 0 || choose_first)
467 * Don't change to another disk for sequential reads:
469 if (conf->next_seq_sect == this_sector)
471 if (this_sector == conf->mirrors[new_disk].head_position)
474 current_distance = abs(this_sector
475 - conf->mirrors[new_disk].head_position);
477 /* look for a better disk - i.e. head is closer */
478 start_disk = new_disk;
479 for (i = 1; i < conf->raid_disks; i++) {
480 int disk = start_disk + 1;
481 if (disk >= conf->raid_disks)
482 disk -= conf->raid_disks;
484 rdev = rcu_dereference(conf->mirrors[disk].rdev);
485 if (r1_bio->bios[disk] == IO_BLOCKED
487 || !test_bit(In_sync, &rdev->flags)
488 || test_bit(WriteMostly, &rdev->flags))
491 if (!atomic_read(&rdev->nr_pending)) {
495 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
496 if (new_distance < current_distance) {
497 current_distance = new_distance;
504 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
507 atomic_inc(&rdev->nr_pending);
508 if (!test_bit(In_sync, &rdev->flags)) {
509 /* cannot risk returning a device that failed
510 * before we inc'ed nr_pending
512 rdev_dec_pending(rdev, conf->mddev);
515 conf->next_seq_sect = this_sector + sectors;
516 conf->last_used = new_disk;
523 static void unplug_slaves(mddev_t *mddev)
525 conf_t *conf = mddev->private;
529 for (i=0; i<mddev->raid_disks; i++) {
530 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
531 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
532 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
534 atomic_inc(&rdev->nr_pending);
539 rdev_dec_pending(rdev, mddev);
546 static void raid1_unplug(struct request_queue *q)
548 mddev_t *mddev = q->queuedata;
550 unplug_slaves(mddev);
551 md_wakeup_thread(mddev->thread);
554 static int raid1_congested(void *data, int bits)
556 mddev_t *mddev = data;
557 conf_t *conf = mddev->private;
560 if (mddev_congested(mddev, bits))
564 for (i = 0; i < mddev->raid_disks; i++) {
565 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
566 if (rdev && !test_bit(Faulty, &rdev->flags)) {
567 struct request_queue *q = bdev_get_queue(rdev->bdev);
569 /* Note the '|| 1' - when read_balance prefers
570 * non-congested targets, it can be removed
572 if ((bits & (1<<BDI_async_congested)) || 1)
573 ret |= bdi_congested(&q->backing_dev_info, bits);
575 ret &= bdi_congested(&q->backing_dev_info, bits);
583 static int flush_pending_writes(conf_t *conf)
585 /* Any writes that have been queued but are awaiting
586 * bitmap updates get flushed here.
587 * We return 1 if any requests were actually submitted.
591 spin_lock_irq(&conf->device_lock);
593 if (conf->pending_bio_list.head) {
595 bio = bio_list_get(&conf->pending_bio_list);
596 blk_remove_plug(conf->mddev->queue);
597 spin_unlock_irq(&conf->device_lock);
598 /* flush any pending bitmap writes to
599 * disk before proceeding w/ I/O */
600 bitmap_unplug(conf->mddev->bitmap);
602 while (bio) { /* submit pending writes */
603 struct bio *next = bio->bi_next;
605 generic_make_request(bio);
610 spin_unlock_irq(&conf->device_lock);
615 * Sometimes we need to suspend IO while we do something else,
616 * either some resync/recovery, or reconfigure the array.
617 * To do this we raise a 'barrier'.
618 * The 'barrier' is a counter that can be raised multiple times
619 * to count how many activities are happening which preclude
621 * We can only raise the barrier if there is no pending IO.
622 * i.e. if nr_pending == 0.
623 * We choose only to raise the barrier if no-one is waiting for the
624 * barrier to go down. This means that as soon as an IO request
625 * is ready, no other operations which require a barrier will start
626 * until the IO request has had a chance.
628 * So: regular IO calls 'wait_barrier'. When that returns there
629 * is no backgroup IO happening, It must arrange to call
630 * allow_barrier when it has finished its IO.
631 * backgroup IO calls must call raise_barrier. Once that returns
632 * there is no normal IO happeing. It must arrange to call
633 * lower_barrier when the particular background IO completes.
635 #define RESYNC_DEPTH 32
637 static void raise_barrier(conf_t *conf)
639 spin_lock_irq(&conf->resync_lock);
641 /* Wait until no block IO is waiting */
642 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
644 raid1_unplug(conf->mddev->queue));
646 /* block any new IO from starting */
649 /* Now wait for all pending IO to complete */
650 wait_event_lock_irq(conf->wait_barrier,
651 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
653 raid1_unplug(conf->mddev->queue));
655 spin_unlock_irq(&conf->resync_lock);
658 static void lower_barrier(conf_t *conf)
661 BUG_ON(conf->barrier <= 0);
662 spin_lock_irqsave(&conf->resync_lock, flags);
664 spin_unlock_irqrestore(&conf->resync_lock, flags);
665 wake_up(&conf->wait_barrier);
668 static void wait_barrier(conf_t *conf)
670 spin_lock_irq(&conf->resync_lock);
673 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
675 raid1_unplug(conf->mddev->queue));
679 spin_unlock_irq(&conf->resync_lock);
682 static void allow_barrier(conf_t *conf)
685 spin_lock_irqsave(&conf->resync_lock, flags);
687 spin_unlock_irqrestore(&conf->resync_lock, flags);
688 wake_up(&conf->wait_barrier);
691 static void freeze_array(conf_t *conf)
693 /* stop syncio and normal IO and wait for everything to
695 * We increment barrier and nr_waiting, and then
696 * wait until nr_pending match nr_queued+1
697 * This is called in the context of one normal IO request
698 * that has failed. Thus any sync request that might be pending
699 * will be blocked by nr_pending, and we need to wait for
700 * pending IO requests to complete or be queued for re-try.
701 * Thus the number queued (nr_queued) plus this request (1)
702 * must match the number of pending IOs (nr_pending) before
705 spin_lock_irq(&conf->resync_lock);
708 wait_event_lock_irq(conf->wait_barrier,
709 conf->nr_pending == conf->nr_queued+1,
711 ({ flush_pending_writes(conf);
712 raid1_unplug(conf->mddev->queue); }));
713 spin_unlock_irq(&conf->resync_lock);
715 static void unfreeze_array(conf_t *conf)
717 /* reverse the effect of the freeze */
718 spin_lock_irq(&conf->resync_lock);
721 wake_up(&conf->wait_barrier);
722 spin_unlock_irq(&conf->resync_lock);
726 /* duplicate the data pages for behind I/O
727 * We return a list of bio_vec rather than just page pointers
728 * as it makes freeing easier
730 static struct bio_vec *alloc_behind_pages(struct bio *bio)
733 struct bio_vec *bvec;
734 struct bio_vec *pages = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
736 if (unlikely(!pages))
739 bio_for_each_segment(bvec, bio, i) {
740 pages[i].bv_page = alloc_page(GFP_NOIO);
741 if (unlikely(!pages[i].bv_page))
743 memcpy(kmap(pages[i].bv_page) + bvec->bv_offset,
744 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
745 kunmap(pages[i].bv_page);
746 kunmap(bvec->bv_page);
753 for (i = 0; i < bio->bi_vcnt && pages[i].bv_page; i++)
754 put_page(pages[i].bv_page);
756 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
760 static int make_request(mddev_t *mddev, struct bio * bio)
762 conf_t *conf = mddev->private;
763 mirror_info_t *mirror;
765 struct bio *read_bio;
766 int i, targets = 0, disks;
767 struct bitmap *bitmap;
769 struct bio_vec *behind_pages = NULL;
770 const int rw = bio_data_dir(bio);
771 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
772 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
773 mdk_rdev_t *blocked_rdev;
776 * Register the new request and wait if the reconstruction
777 * thread has put up a bar for new requests.
778 * Continue immediately if no resync is active currently.
781 md_write_start(mddev, bio); /* wait on superblock update early */
783 if (bio_data_dir(bio) == WRITE &&
784 bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
785 bio->bi_sector < mddev->suspend_hi) {
786 /* As the suspend_* range is controlled by
787 * userspace, we want an interruptible
792 flush_signals(current);
793 prepare_to_wait(&conf->wait_barrier,
794 &w, TASK_INTERRUPTIBLE);
795 if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
796 bio->bi_sector >= mddev->suspend_hi)
800 finish_wait(&conf->wait_barrier, &w);
805 bitmap = mddev->bitmap;
808 * make_request() can abort the operation when READA is being
809 * used and no empty request is available.
812 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
814 r1_bio->master_bio = bio;
815 r1_bio->sectors = bio->bi_size >> 9;
817 r1_bio->mddev = mddev;
818 r1_bio->sector = bio->bi_sector;
822 * read balancing logic:
824 int rdisk = read_balance(conf, r1_bio);
827 /* couldn't find anywhere to read from */
828 raid_end_bio_io(r1_bio);
831 mirror = conf->mirrors + rdisk;
833 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
835 /* Reading from a write-mostly device must
836 * take care not to over-take any writes
839 wait_event(bitmap->behind_wait,
840 atomic_read(&bitmap->behind_writes) == 0);
842 r1_bio->read_disk = rdisk;
844 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
846 r1_bio->bios[rdisk] = read_bio;
848 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
849 read_bio->bi_bdev = mirror->rdev->bdev;
850 read_bio->bi_end_io = raid1_end_read_request;
851 read_bio->bi_rw = READ | do_sync;
852 read_bio->bi_private = r1_bio;
854 generic_make_request(read_bio);
861 /* first select target devices under spinlock and
862 * inc refcount on their rdev. Record them by setting
865 disks = conf->raid_disks;
869 for (i = 0; i < disks; i++) {
870 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
871 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
872 atomic_inc(&rdev->nr_pending);
876 if (rdev && !test_bit(Faulty, &rdev->flags)) {
877 atomic_inc(&rdev->nr_pending);
878 if (test_bit(Faulty, &rdev->flags)) {
879 rdev_dec_pending(rdev, mddev);
880 r1_bio->bios[i] = NULL;
882 r1_bio->bios[i] = bio;
886 r1_bio->bios[i] = NULL;
890 if (unlikely(blocked_rdev)) {
891 /* Wait for this device to become unblocked */
894 for (j = 0; j < i; j++)
896 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
899 md_wait_for_blocked_rdev(blocked_rdev, mddev);
904 BUG_ON(targets == 0); /* we never fail the last device */
906 if (targets < conf->raid_disks) {
907 /* array is degraded, we will not clear the bitmap
908 * on I/O completion (see raid1_end_write_request) */
909 set_bit(R1BIO_Degraded, &r1_bio->state);
913 * Not if there are too many, or cannot allocate memory,
914 * or a reader on WriteMostly is waiting for behind writes
917 (atomic_read(&bitmap->behind_writes)
918 < mddev->bitmap_info.max_write_behind) &&
919 !waitqueue_active(&bitmap->behind_wait) &&
920 (behind_pages = alloc_behind_pages(bio)) != NULL)
921 set_bit(R1BIO_BehindIO, &r1_bio->state);
923 atomic_set(&r1_bio->remaining, 1);
924 atomic_set(&r1_bio->behind_remaining, 0);
926 bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
927 test_bit(R1BIO_BehindIO, &r1_bio->state));
928 for (i = 0; i < disks; i++) {
930 if (!r1_bio->bios[i])
933 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
934 r1_bio->bios[i] = mbio;
936 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
937 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
938 mbio->bi_end_io = raid1_end_write_request;
939 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
940 mbio->bi_private = r1_bio;
943 struct bio_vec *bvec;
946 /* Yes, I really want the '__' version so that
947 * we clear any unused pointer in the io_vec, rather
948 * than leave them unchanged. This is important
949 * because when we come to free the pages, we won't
950 * know the original bi_idx, so we just free
953 __bio_for_each_segment(bvec, mbio, j, 0)
954 bvec->bv_page = behind_pages[j].bv_page;
955 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
956 atomic_inc(&r1_bio->behind_remaining);
959 atomic_inc(&r1_bio->remaining);
960 spin_lock_irqsave(&conf->device_lock, flags);
961 bio_list_add(&conf->pending_bio_list, mbio);
962 blk_plug_device(mddev->queue);
963 spin_unlock_irqrestore(&conf->device_lock, flags);
965 r1_bio_write_done(r1_bio, bio->bi_vcnt, behind_pages, behind_pages != NULL);
966 kfree(behind_pages); /* the behind pages are attached to the bios now */
968 /* In case raid1d snuck in to freeze_array */
969 wake_up(&conf->wait_barrier);
972 md_wakeup_thread(mddev->thread);
977 static void status(struct seq_file *seq, mddev_t *mddev)
979 conf_t *conf = mddev->private;
982 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
983 conf->raid_disks - mddev->degraded);
985 for (i = 0; i < conf->raid_disks; i++) {
986 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
987 seq_printf(seq, "%s",
988 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
991 seq_printf(seq, "]");
995 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
997 char b[BDEVNAME_SIZE];
998 conf_t *conf = mddev->private;
1001 * If it is not operational, then we have already marked it as dead
1002 * else if it is the last working disks, ignore the error, let the
1003 * next level up know.
1004 * else mark the drive as failed
1006 if (test_bit(In_sync, &rdev->flags)
1007 && (conf->raid_disks - mddev->degraded) == 1) {
1009 * Don't fail the drive, act as though we were just a
1010 * normal single drive.
1011 * However don't try a recovery from this drive as
1012 * it is very likely to fail.
1014 mddev->recovery_disabled = 1;
1017 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1018 unsigned long flags;
1019 spin_lock_irqsave(&conf->device_lock, flags);
1021 set_bit(Faulty, &rdev->flags);
1022 spin_unlock_irqrestore(&conf->device_lock, flags);
1024 * if recovery is running, make sure it aborts.
1026 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1028 set_bit(Faulty, &rdev->flags);
1029 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1030 printk(KERN_ALERT "md/raid1:%s: Disk failure on %s, disabling device.\n"
1031 KERN_ALERT "md/raid1:%s: Operation continuing on %d devices.\n",
1032 mdname(mddev), bdevname(rdev->bdev, b),
1033 mdname(mddev), conf->raid_disks - mddev->degraded);
1036 static void print_conf(conf_t *conf)
1040 printk(KERN_DEBUG "RAID1 conf printout:\n");
1042 printk(KERN_DEBUG "(!conf)\n");
1045 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1049 for (i = 0; i < conf->raid_disks; i++) {
1050 char b[BDEVNAME_SIZE];
1051 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1053 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1054 i, !test_bit(In_sync, &rdev->flags),
1055 !test_bit(Faulty, &rdev->flags),
1056 bdevname(rdev->bdev,b));
1061 static void close_sync(conf_t *conf)
1064 allow_barrier(conf);
1066 mempool_destroy(conf->r1buf_pool);
1067 conf->r1buf_pool = NULL;
1070 static int raid1_spare_active(mddev_t *mddev)
1073 conf_t *conf = mddev->private;
1075 unsigned long flags;
1078 * Find all failed disks within the RAID1 configuration
1079 * and mark them readable.
1080 * Called under mddev lock, so rcu protection not needed.
1082 for (i = 0; i < conf->raid_disks; i++) {
1083 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1085 && !test_bit(Faulty, &rdev->flags)
1086 && !test_and_set_bit(In_sync, &rdev->flags)) {
1088 sysfs_notify_dirent(rdev->sysfs_state);
1091 spin_lock_irqsave(&conf->device_lock, flags);
1092 mddev->degraded -= count;
1093 spin_unlock_irqrestore(&conf->device_lock, flags);
1100 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1102 conf_t *conf = mddev->private;
1107 int last = mddev->raid_disks - 1;
1109 if (rdev->raid_disk >= 0)
1110 first = last = rdev->raid_disk;
1112 for (mirror = first; mirror <= last; mirror++)
1113 if ( !(p=conf->mirrors+mirror)->rdev) {
1115 disk_stack_limits(mddev->gendisk, rdev->bdev,
1116 rdev->data_offset << 9);
1117 /* as we don't honour merge_bvec_fn, we must
1118 * never risk violating it, so limit
1119 * ->max_segments to one lying with a single
1120 * page, as a one page request is never in
1123 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1124 blk_queue_max_segments(mddev->queue, 1);
1125 blk_queue_segment_boundary(mddev->queue,
1126 PAGE_CACHE_SIZE - 1);
1129 p->head_position = 0;
1130 rdev->raid_disk = mirror;
1132 /* As all devices are equivalent, we don't need a full recovery
1133 * if this was recently any drive of the array
1135 if (rdev->saved_raid_disk < 0)
1137 rcu_assign_pointer(p->rdev, rdev);
1140 md_integrity_add_rdev(rdev, mddev);
1145 static int raid1_remove_disk(mddev_t *mddev, int number)
1147 conf_t *conf = mddev->private;
1150 mirror_info_t *p = conf->mirrors+ number;
1155 if (test_bit(In_sync, &rdev->flags) ||
1156 atomic_read(&rdev->nr_pending)) {
1160 /* Only remove non-faulty devices if recovery
1163 if (!test_bit(Faulty, &rdev->flags) &&
1164 !mddev->recovery_disabled &&
1165 mddev->degraded < conf->raid_disks) {
1171 if (atomic_read(&rdev->nr_pending)) {
1172 /* lost the race, try later */
1177 md_integrity_register(mddev);
1186 static void end_sync_read(struct bio *bio, int error)
1188 r1bio_t *r1_bio = bio->bi_private;
1191 for (i=r1_bio->mddev->raid_disks; i--; )
1192 if (r1_bio->bios[i] == bio)
1195 update_head_pos(i, r1_bio);
1197 * we have read a block, now it needs to be re-written,
1198 * or re-read if the read failed.
1199 * We don't do much here, just schedule handling by raid1d
1201 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1202 set_bit(R1BIO_Uptodate, &r1_bio->state);
1204 if (atomic_dec_and_test(&r1_bio->remaining))
1205 reschedule_retry(r1_bio);
1208 static void end_sync_write(struct bio *bio, int error)
1210 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1211 r1bio_t *r1_bio = bio->bi_private;
1212 mddev_t *mddev = r1_bio->mddev;
1213 conf_t *conf = mddev->private;
1217 for (i = 0; i < conf->raid_disks; i++)
1218 if (r1_bio->bios[i] == bio) {
1223 sector_t sync_blocks = 0;
1224 sector_t s = r1_bio->sector;
1225 long sectors_to_go = r1_bio->sectors;
1226 /* make sure these bits doesn't get cleared. */
1228 bitmap_end_sync(mddev->bitmap, s,
1231 sectors_to_go -= sync_blocks;
1232 } while (sectors_to_go > 0);
1233 md_error(mddev, conf->mirrors[mirror].rdev);
1236 update_head_pos(mirror, r1_bio);
1238 if (atomic_dec_and_test(&r1_bio->remaining)) {
1239 sector_t s = r1_bio->sectors;
1241 md_done_sync(mddev, s, uptodate);
1245 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1247 conf_t *conf = mddev->private;
1249 int disks = conf->raid_disks;
1250 struct bio *bio, *wbio;
1252 bio = r1_bio->bios[r1_bio->read_disk];
1255 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1256 /* We have read all readable devices. If we haven't
1257 * got the block, then there is no hope left.
1258 * If we have, then we want to do a comparison
1259 * and skip the write if everything is the same.
1260 * If any blocks failed to read, then we need to
1261 * attempt an over-write
1264 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1265 for (i=0; i<mddev->raid_disks; i++)
1266 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1267 md_error(mddev, conf->mirrors[i].rdev);
1269 md_done_sync(mddev, r1_bio->sectors, 1);
1273 for (primary=0; primary<mddev->raid_disks; primary++)
1274 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1275 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1276 r1_bio->bios[primary]->bi_end_io = NULL;
1277 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1280 r1_bio->read_disk = primary;
1281 for (i=0; i<mddev->raid_disks; i++)
1282 if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1284 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1285 struct bio *pbio = r1_bio->bios[primary];
1286 struct bio *sbio = r1_bio->bios[i];
1288 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1289 for (j = vcnt; j-- ; ) {
1291 p = pbio->bi_io_vec[j].bv_page;
1292 s = sbio->bi_io_vec[j].bv_page;
1293 if (memcmp(page_address(p),
1301 mddev->resync_mismatches += r1_bio->sectors;
1302 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1303 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1304 sbio->bi_end_io = NULL;
1305 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1307 /* fixup the bio for reuse */
1309 sbio->bi_vcnt = vcnt;
1310 sbio->bi_size = r1_bio->sectors << 9;
1312 sbio->bi_phys_segments = 0;
1313 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1314 sbio->bi_flags |= 1 << BIO_UPTODATE;
1315 sbio->bi_next = NULL;
1316 sbio->bi_sector = r1_bio->sector +
1317 conf->mirrors[i].rdev->data_offset;
1318 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1319 size = sbio->bi_size;
1320 for (j = 0; j < vcnt ; j++) {
1322 bi = &sbio->bi_io_vec[j];
1324 if (size > PAGE_SIZE)
1325 bi->bv_len = PAGE_SIZE;
1329 memcpy(page_address(bi->bv_page),
1330 page_address(pbio->bi_io_vec[j].bv_page),
1337 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1338 /* ouch - failed to read all of that.
1339 * Try some synchronous reads of other devices to get
1340 * good data, much like with normal read errors. Only
1341 * read into the pages we already have so we don't
1342 * need to re-issue the read request.
1343 * We don't need to freeze the array, because being in an
1344 * active sync request, there is no normal IO, and
1345 * no overlapping syncs.
1347 sector_t sect = r1_bio->sector;
1348 int sectors = r1_bio->sectors;
1353 int d = r1_bio->read_disk;
1357 if (s > (PAGE_SIZE>>9))
1360 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1361 /* No rcu protection needed here devices
1362 * can only be removed when no resync is
1363 * active, and resync is currently active
1365 rdev = conf->mirrors[d].rdev;
1366 if (sync_page_io(rdev,
1367 sect + rdev->data_offset,
1369 bio->bi_io_vec[idx].bv_page,
1376 if (d == conf->raid_disks)
1378 } while (!success && d != r1_bio->read_disk);
1382 /* write it back and re-read */
1383 set_bit(R1BIO_Uptodate, &r1_bio->state);
1384 while (d != r1_bio->read_disk) {
1386 d = conf->raid_disks;
1388 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1390 rdev = conf->mirrors[d].rdev;
1391 atomic_add(s, &rdev->corrected_errors);
1392 if (sync_page_io(rdev,
1393 sect + rdev->data_offset,
1395 bio->bi_io_vec[idx].bv_page,
1397 md_error(mddev, rdev);
1400 while (d != r1_bio->read_disk) {
1402 d = conf->raid_disks;
1404 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1406 rdev = conf->mirrors[d].rdev;
1407 if (sync_page_io(rdev,
1408 sect + rdev->data_offset,
1410 bio->bi_io_vec[idx].bv_page,
1412 md_error(mddev, rdev);
1415 char b[BDEVNAME_SIZE];
1416 /* Cannot read from anywhere, array is toast */
1417 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1418 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1419 " for block %llu\n",
1421 bdevname(bio->bi_bdev, b),
1422 (unsigned long long)r1_bio->sector);
1423 md_done_sync(mddev, r1_bio->sectors, 0);
1436 atomic_set(&r1_bio->remaining, 1);
1437 for (i = 0; i < disks ; i++) {
1438 wbio = r1_bio->bios[i];
1439 if (wbio->bi_end_io == NULL ||
1440 (wbio->bi_end_io == end_sync_read &&
1441 (i == r1_bio->read_disk ||
1442 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1445 wbio->bi_rw = WRITE;
1446 wbio->bi_end_io = end_sync_write;
1447 atomic_inc(&r1_bio->remaining);
1448 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1450 generic_make_request(wbio);
1453 if (atomic_dec_and_test(&r1_bio->remaining)) {
1454 /* if we're here, all write(s) have completed, so clean up */
1455 md_done_sync(mddev, r1_bio->sectors, 1);
1461 * This is a kernel thread which:
1463 * 1. Retries failed read operations on working mirrors.
1464 * 2. Updates the raid superblock when problems encounter.
1465 * 3. Performs writes following reads for array syncronising.
1468 static void fix_read_error(conf_t *conf, int read_disk,
1469 sector_t sect, int sectors)
1471 mddev_t *mddev = conf->mddev;
1479 if (s > (PAGE_SIZE>>9))
1483 /* Note: no rcu protection needed here
1484 * as this is synchronous in the raid1d thread
1485 * which is the thread that might remove
1486 * a device. If raid1d ever becomes multi-threaded....
1488 rdev = conf->mirrors[d].rdev;
1490 test_bit(In_sync, &rdev->flags) &&
1492 sect + rdev->data_offset,
1494 conf->tmppage, READ))
1498 if (d == conf->raid_disks)
1501 } while (!success && d != read_disk);
1504 /* Cannot read from anywhere -- bye bye array */
1505 md_error(mddev, conf->mirrors[read_disk].rdev);
1508 /* write it back and re-read */
1510 while (d != read_disk) {
1512 d = conf->raid_disks;
1514 rdev = conf->mirrors[d].rdev;
1516 test_bit(In_sync, &rdev->flags)) {
1517 if (sync_page_io(rdev,
1518 sect + rdev->data_offset,
1519 s<<9, conf->tmppage, WRITE)
1521 /* Well, this device is dead */
1522 md_error(mddev, rdev);
1526 while (d != read_disk) {
1527 char b[BDEVNAME_SIZE];
1529 d = conf->raid_disks;
1531 rdev = conf->mirrors[d].rdev;
1533 test_bit(In_sync, &rdev->flags)) {
1534 if (sync_page_io(rdev,
1535 sect + rdev->data_offset,
1536 s<<9, conf->tmppage, READ)
1538 /* Well, this device is dead */
1539 md_error(mddev, rdev);
1541 atomic_add(s, &rdev->corrected_errors);
1543 "md/raid1:%s: read error corrected "
1544 "(%d sectors at %llu on %s)\n",
1546 (unsigned long long)(sect +
1548 bdevname(rdev->bdev, b));
1557 static void raid1d(mddev_t *mddev)
1561 unsigned long flags;
1562 conf_t *conf = mddev->private;
1563 struct list_head *head = &conf->retry_list;
1567 md_check_recovery(mddev);
1570 char b[BDEVNAME_SIZE];
1572 unplug += flush_pending_writes(conf);
1574 spin_lock_irqsave(&conf->device_lock, flags);
1575 if (list_empty(head)) {
1576 spin_unlock_irqrestore(&conf->device_lock, flags);
1579 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1580 list_del(head->prev);
1582 spin_unlock_irqrestore(&conf->device_lock, flags);
1584 mddev = r1_bio->mddev;
1585 conf = mddev->private;
1586 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1587 sync_request_write(mddev, r1_bio);
1592 /* we got a read error. Maybe the drive is bad. Maybe just
1593 * the block and we can fix it.
1594 * We freeze all other IO, and try reading the block from
1595 * other devices. When we find one, we re-write
1596 * and check it that fixes the read error.
1597 * This is all done synchronously while the array is
1600 if (mddev->ro == 0) {
1602 fix_read_error(conf, r1_bio->read_disk,
1605 unfreeze_array(conf);
1608 conf->mirrors[r1_bio->read_disk].rdev);
1610 bio = r1_bio->bios[r1_bio->read_disk];
1611 if ((disk=read_balance(conf, r1_bio)) == -1) {
1612 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1613 " read error for block %llu\n",
1615 bdevname(bio->bi_bdev,b),
1616 (unsigned long long)r1_bio->sector);
1617 raid_end_bio_io(r1_bio);
1619 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1620 r1_bio->bios[r1_bio->read_disk] =
1621 mddev->ro ? IO_BLOCKED : NULL;
1622 r1_bio->read_disk = disk;
1624 bio = bio_clone_mddev(r1_bio->master_bio,
1626 r1_bio->bios[r1_bio->read_disk] = bio;
1627 rdev = conf->mirrors[disk].rdev;
1628 if (printk_ratelimit())
1629 printk(KERN_ERR "md/raid1:%s: redirecting sector %llu to"
1630 " other mirror: %s\n",
1632 (unsigned long long)r1_bio->sector,
1633 bdevname(rdev->bdev,b));
1634 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1635 bio->bi_bdev = rdev->bdev;
1636 bio->bi_end_io = raid1_end_read_request;
1637 bio->bi_rw = READ | do_sync;
1638 bio->bi_private = r1_bio;
1640 generic_make_request(bio);
1646 unplug_slaves(mddev);
1650 static int init_resync(conf_t *conf)
1654 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1655 BUG_ON(conf->r1buf_pool);
1656 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1658 if (!conf->r1buf_pool)
1660 conf->next_resync = 0;
1665 * perform a "sync" on one "block"
1667 * We need to make sure that no normal I/O request - particularly write
1668 * requests - conflict with active sync requests.
1670 * This is achieved by tracking pending requests and a 'barrier' concept
1671 * that can be installed to exclude normal IO requests.
1674 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1676 conf_t *conf = mddev->private;
1679 sector_t max_sector, nr_sectors;
1683 int write_targets = 0, read_targets = 0;
1684 sector_t sync_blocks;
1685 int still_degraded = 0;
1687 if (!conf->r1buf_pool)
1688 if (init_resync(conf))
1691 max_sector = mddev->dev_sectors;
1692 if (sector_nr >= max_sector) {
1693 /* If we aborted, we need to abort the
1694 * sync on the 'current' bitmap chunk (there will
1695 * only be one in raid1 resync.
1696 * We can find the current addess in mddev->curr_resync
1698 if (mddev->curr_resync < max_sector) /* aborted */
1699 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1701 else /* completed sync */
1704 bitmap_close_sync(mddev->bitmap);
1709 if (mddev->bitmap == NULL &&
1710 mddev->recovery_cp == MaxSector &&
1711 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1712 conf->fullsync == 0) {
1714 return max_sector - sector_nr;
1716 /* before building a request, check if we can skip these blocks..
1717 * This call the bitmap_start_sync doesn't actually record anything
1719 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1720 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1721 /* We can skip this block, and probably several more */
1726 * If there is non-resync activity waiting for a turn,
1727 * and resync is going fast enough,
1728 * then let it though before starting on this new sync request.
1730 if (!go_faster && conf->nr_waiting)
1731 msleep_interruptible(1000);
1733 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1734 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1735 raise_barrier(conf);
1737 conf->next_resync = sector_nr;
1741 * If we get a correctably read error during resync or recovery,
1742 * we might want to read from a different device. So we
1743 * flag all drives that could conceivably be read from for READ,
1744 * and any others (which will be non-In_sync devices) for WRITE.
1745 * If a read fails, we try reading from something else for which READ
1749 r1_bio->mddev = mddev;
1750 r1_bio->sector = sector_nr;
1752 set_bit(R1BIO_IsSync, &r1_bio->state);
1754 for (i=0; i < conf->raid_disks; i++) {
1756 bio = r1_bio->bios[i];
1758 /* take from bio_init */
1759 bio->bi_next = NULL;
1760 bio->bi_flags &= ~(BIO_POOL_MASK-1);
1761 bio->bi_flags |= 1 << BIO_UPTODATE;
1762 bio->bi_comp_cpu = -1;
1766 bio->bi_phys_segments = 0;
1768 bio->bi_end_io = NULL;
1769 bio->bi_private = NULL;
1771 rdev = rcu_dereference(conf->mirrors[i].rdev);
1773 test_bit(Faulty, &rdev->flags)) {
1776 } else if (!test_bit(In_sync, &rdev->flags)) {
1778 bio->bi_end_io = end_sync_write;
1781 /* may need to read from here */
1783 bio->bi_end_io = end_sync_read;
1784 if (test_bit(WriteMostly, &rdev->flags)) {
1793 atomic_inc(&rdev->nr_pending);
1794 bio->bi_sector = sector_nr + rdev->data_offset;
1795 bio->bi_bdev = rdev->bdev;
1796 bio->bi_private = r1_bio;
1801 r1_bio->read_disk = disk;
1803 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1804 /* extra read targets are also write targets */
1805 write_targets += read_targets-1;
1807 if (write_targets == 0 || read_targets == 0) {
1808 /* There is nowhere to write, so all non-sync
1809 * drives must be failed - so we are finished
1811 sector_t rv = max_sector - sector_nr;
1817 if (max_sector > mddev->resync_max)
1818 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1823 int len = PAGE_SIZE;
1824 if (sector_nr + (len>>9) > max_sector)
1825 len = (max_sector - sector_nr) << 9;
1828 if (sync_blocks == 0) {
1829 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1830 &sync_blocks, still_degraded) &&
1832 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1834 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1835 if ((len >> 9) > sync_blocks)
1836 len = sync_blocks<<9;
1839 for (i=0 ; i < conf->raid_disks; i++) {
1840 bio = r1_bio->bios[i];
1841 if (bio->bi_end_io) {
1842 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1843 if (bio_add_page(bio, page, len, 0) == 0) {
1845 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1848 bio = r1_bio->bios[i];
1849 if (bio->bi_end_io==NULL)
1851 /* remove last page from this bio */
1853 bio->bi_size -= len;
1854 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1860 nr_sectors += len>>9;
1861 sector_nr += len>>9;
1862 sync_blocks -= (len>>9);
1863 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1865 r1_bio->sectors = nr_sectors;
1867 /* For a user-requested sync, we read all readable devices and do a
1870 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1871 atomic_set(&r1_bio->remaining, read_targets);
1872 for (i=0; i<conf->raid_disks; i++) {
1873 bio = r1_bio->bios[i];
1874 if (bio->bi_end_io == end_sync_read) {
1875 md_sync_acct(bio->bi_bdev, nr_sectors);
1876 generic_make_request(bio);
1880 atomic_set(&r1_bio->remaining, 1);
1881 bio = r1_bio->bios[r1_bio->read_disk];
1882 md_sync_acct(bio->bi_bdev, nr_sectors);
1883 generic_make_request(bio);
1889 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1894 return mddev->dev_sectors;
1897 static conf_t *setup_conf(mddev_t *mddev)
1901 mirror_info_t *disk;
1905 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1909 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1914 conf->tmppage = alloc_page(GFP_KERNEL);
1918 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1919 if (!conf->poolinfo)
1921 conf->poolinfo->raid_disks = mddev->raid_disks;
1922 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1925 if (!conf->r1bio_pool)
1928 conf->poolinfo->mddev = mddev;
1930 spin_lock_init(&conf->device_lock);
1931 list_for_each_entry(rdev, &mddev->disks, same_set) {
1932 int disk_idx = rdev->raid_disk;
1933 if (disk_idx >= mddev->raid_disks
1936 disk = conf->mirrors + disk_idx;
1940 disk->head_position = 0;
1942 conf->raid_disks = mddev->raid_disks;
1943 conf->mddev = mddev;
1944 INIT_LIST_HEAD(&conf->retry_list);
1946 spin_lock_init(&conf->resync_lock);
1947 init_waitqueue_head(&conf->wait_barrier);
1949 bio_list_init(&conf->pending_bio_list);
1951 conf->last_used = -1;
1952 for (i = 0; i < conf->raid_disks; i++) {
1954 disk = conf->mirrors + i;
1957 !test_bit(In_sync, &disk->rdev->flags)) {
1958 disk->head_position = 0;
1961 } else if (conf->last_used < 0)
1963 * The first working device is used as a
1964 * starting point to read balancing.
1966 conf->last_used = i;
1970 if (conf->last_used < 0) {
1971 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
1976 conf->thread = md_register_thread(raid1d, mddev, NULL);
1977 if (!conf->thread) {
1979 "md/raid1:%s: couldn't allocate thread\n",
1988 if (conf->r1bio_pool)
1989 mempool_destroy(conf->r1bio_pool);
1990 kfree(conf->mirrors);
1991 safe_put_page(conf->tmppage);
1992 kfree(conf->poolinfo);
1995 return ERR_PTR(err);
1998 static int run(mddev_t *mddev)
2004 if (mddev->level != 1) {
2005 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2006 mdname(mddev), mddev->level);
2009 if (mddev->reshape_position != MaxSector) {
2010 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2015 * copy the already verified devices into our private RAID1
2016 * bookkeeping area. [whatever we allocate in run(),
2017 * should be freed in stop()]
2019 if (mddev->private == NULL)
2020 conf = setup_conf(mddev);
2022 conf = mddev->private;
2025 return PTR_ERR(conf);
2027 mddev->queue->queue_lock = &conf->device_lock;
2028 list_for_each_entry(rdev, &mddev->disks, same_set) {
2029 disk_stack_limits(mddev->gendisk, rdev->bdev,
2030 rdev->data_offset << 9);
2031 /* as we don't honour merge_bvec_fn, we must never risk
2032 * violating it, so limit ->max_segments to 1 lying within
2033 * a single page, as a one page request is never in violation.
2035 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2036 blk_queue_max_segments(mddev->queue, 1);
2037 blk_queue_segment_boundary(mddev->queue,
2038 PAGE_CACHE_SIZE - 1);
2042 mddev->degraded = 0;
2043 for (i=0; i < conf->raid_disks; i++)
2044 if (conf->mirrors[i].rdev == NULL ||
2045 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2046 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2049 if (conf->raid_disks - mddev->degraded == 1)
2050 mddev->recovery_cp = MaxSector;
2052 if (mddev->recovery_cp != MaxSector)
2053 printk(KERN_NOTICE "md/raid1:%s: not clean"
2054 " -- starting background reconstruction\n",
2057 "md/raid1:%s: active with %d out of %d mirrors\n",
2058 mdname(mddev), mddev->raid_disks - mddev->degraded,
2062 * Ok, everything is just fine now
2064 mddev->thread = conf->thread;
2065 conf->thread = NULL;
2066 mddev->private = conf;
2068 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2070 mddev->queue->unplug_fn = raid1_unplug;
2071 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2072 mddev->queue->backing_dev_info.congested_data = mddev;
2073 md_integrity_register(mddev);
2077 static int stop(mddev_t *mddev)
2079 conf_t *conf = mddev->private;
2080 struct bitmap *bitmap = mddev->bitmap;
2082 /* wait for behind writes to complete */
2083 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2084 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2086 /* need to kick something here to make sure I/O goes? */
2087 wait_event(bitmap->behind_wait,
2088 atomic_read(&bitmap->behind_writes) == 0);
2091 raise_barrier(conf);
2092 lower_barrier(conf);
2094 md_unregister_thread(mddev->thread);
2095 mddev->thread = NULL;
2096 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2097 if (conf->r1bio_pool)
2098 mempool_destroy(conf->r1bio_pool);
2099 kfree(conf->mirrors);
2100 kfree(conf->poolinfo);
2102 mddev->private = NULL;
2106 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2108 /* no resync is happening, and there is enough space
2109 * on all devices, so we can resize.
2110 * We need to make sure resync covers any new space.
2111 * If the array is shrinking we should possibly wait until
2112 * any io in the removed space completes, but it hardly seems
2115 md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2116 if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2118 set_capacity(mddev->gendisk, mddev->array_sectors);
2119 revalidate_disk(mddev->gendisk);
2120 if (sectors > mddev->dev_sectors &&
2121 mddev->recovery_cp == MaxSector) {
2122 mddev->recovery_cp = mddev->dev_sectors;
2123 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2125 mddev->dev_sectors = sectors;
2126 mddev->resync_max_sectors = sectors;
2130 static int raid1_reshape(mddev_t *mddev)
2133 * 1/ resize the r1bio_pool
2134 * 2/ resize conf->mirrors
2136 * We allocate a new r1bio_pool if we can.
2137 * Then raise a device barrier and wait until all IO stops.
2138 * Then resize conf->mirrors and swap in the new r1bio pool.
2140 * At the same time, we "pack" the devices so that all the missing
2141 * devices have the higher raid_disk numbers.
2143 mempool_t *newpool, *oldpool;
2144 struct pool_info *newpoolinfo;
2145 mirror_info_t *newmirrors;
2146 conf_t *conf = mddev->private;
2147 int cnt, raid_disks;
2148 unsigned long flags;
2151 /* Cannot change chunk_size, layout, or level */
2152 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2153 mddev->layout != mddev->new_layout ||
2154 mddev->level != mddev->new_level) {
2155 mddev->new_chunk_sectors = mddev->chunk_sectors;
2156 mddev->new_layout = mddev->layout;
2157 mddev->new_level = mddev->level;
2161 err = md_allow_write(mddev);
2165 raid_disks = mddev->raid_disks + mddev->delta_disks;
2167 if (raid_disks < conf->raid_disks) {
2169 for (d= 0; d < conf->raid_disks; d++)
2170 if (conf->mirrors[d].rdev)
2172 if (cnt > raid_disks)
2176 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2179 newpoolinfo->mddev = mddev;
2180 newpoolinfo->raid_disks = raid_disks;
2182 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2183 r1bio_pool_free, newpoolinfo);
2188 newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2191 mempool_destroy(newpool);
2195 raise_barrier(conf);
2197 /* ok, everything is stopped */
2198 oldpool = conf->r1bio_pool;
2199 conf->r1bio_pool = newpool;
2201 for (d = d2 = 0; d < conf->raid_disks; d++) {
2202 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2203 if (rdev && rdev->raid_disk != d2) {
2205 sprintf(nm, "rd%d", rdev->raid_disk);
2206 sysfs_remove_link(&mddev->kobj, nm);
2207 rdev->raid_disk = d2;
2208 sprintf(nm, "rd%d", rdev->raid_disk);
2209 sysfs_remove_link(&mddev->kobj, nm);
2210 if (sysfs_create_link(&mddev->kobj,
2213 "md/raid1:%s: cannot register "
2218 newmirrors[d2++].rdev = rdev;
2220 kfree(conf->mirrors);
2221 conf->mirrors = newmirrors;
2222 kfree(conf->poolinfo);
2223 conf->poolinfo = newpoolinfo;
2225 spin_lock_irqsave(&conf->device_lock, flags);
2226 mddev->degraded += (raid_disks - conf->raid_disks);
2227 spin_unlock_irqrestore(&conf->device_lock, flags);
2228 conf->raid_disks = mddev->raid_disks = raid_disks;
2229 mddev->delta_disks = 0;
2231 conf->last_used = 0; /* just make sure it is in-range */
2232 lower_barrier(conf);
2234 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2235 md_wakeup_thread(mddev->thread);
2237 mempool_destroy(oldpool);
2241 static void raid1_quiesce(mddev_t *mddev, int state)
2243 conf_t *conf = mddev->private;
2246 case 2: /* wake for suspend */
2247 wake_up(&conf->wait_barrier);
2250 raise_barrier(conf);
2253 lower_barrier(conf);
2258 static void *raid1_takeover(mddev_t *mddev)
2260 /* raid1 can take over:
2261 * raid5 with 2 devices, any layout or chunk size
2263 if (mddev->level == 5 && mddev->raid_disks == 2) {
2265 mddev->new_level = 1;
2266 mddev->new_layout = 0;
2267 mddev->new_chunk_sectors = 0;
2268 conf = setup_conf(mddev);
2273 return ERR_PTR(-EINVAL);
2276 static struct mdk_personality raid1_personality =
2280 .owner = THIS_MODULE,
2281 .make_request = make_request,
2285 .error_handler = error,
2286 .hot_add_disk = raid1_add_disk,
2287 .hot_remove_disk= raid1_remove_disk,
2288 .spare_active = raid1_spare_active,
2289 .sync_request = sync_request,
2290 .resize = raid1_resize,
2292 .check_reshape = raid1_reshape,
2293 .quiesce = raid1_quiesce,
2294 .takeover = raid1_takeover,
2297 static int __init raid_init(void)
2299 return register_md_personality(&raid1_personality);
2302 static void raid_exit(void)
2304 unregister_md_personality(&raid1_personality);
2307 module_init(raid_init);
2308 module_exit(raid_exit);
2309 MODULE_LICENSE("GPL");
2310 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2311 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2312 MODULE_ALIAS("md-raid1");
2313 MODULE_ALIAS("md-level-1");