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>
38 #include <linux/ratelimit.h>
45 #define PRINTK(x...) printk(x)
51 * Number of guaranteed r1bios in case of extreme VM load:
53 #define NR_RAID1_BIOS 256
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;
62 int size = offsetof(r1bio_t, bios[pi->raid_disks]);
64 /* allocate a r1bio with room for raid_disks entries in the bios array */
65 return kzalloc(size, gfp_flags);
68 static void r1bio_pool_free(void *r1_bio, void *data)
73 #define RESYNC_BLOCK_SIZE (64*1024)
74 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
75 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
76 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
77 #define RESYNC_WINDOW (2048*1024)
79 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
81 struct pool_info *pi = data;
87 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
92 * Allocate bios : 1 for reading, n-1 for writing
94 for (j = pi->raid_disks ; j-- ; ) {
95 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
98 r1_bio->bios[j] = bio;
101 * Allocate RESYNC_PAGES data pages and attach them to
103 * If this is a user-requested check/repair, allocate
104 * RESYNC_PAGES for each bio.
106 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
111 bio = r1_bio->bios[j];
112 for (i = 0; i < RESYNC_PAGES; i++) {
113 page = alloc_page(gfp_flags);
117 bio->bi_io_vec[i].bv_page = page;
121 /* If not user-requests, copy the page pointers to all bios */
122 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
123 for (i=0; i<RESYNC_PAGES ; i++)
124 for (j=1; j<pi->raid_disks; j++)
125 r1_bio->bios[j]->bi_io_vec[i].bv_page =
126 r1_bio->bios[0]->bi_io_vec[i].bv_page;
129 r1_bio->master_bio = NULL;
134 for (j=0 ; j < pi->raid_disks; j++)
135 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
136 put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
139 while ( ++j < pi->raid_disks )
140 bio_put(r1_bio->bios[j]);
141 r1bio_pool_free(r1_bio, data);
145 static void r1buf_pool_free(void *__r1_bio, void *data)
147 struct pool_info *pi = data;
149 r1bio_t *r1bio = __r1_bio;
151 for (i = 0; i < RESYNC_PAGES; i++)
152 for (j = pi->raid_disks; j-- ;) {
154 r1bio->bios[j]->bi_io_vec[i].bv_page !=
155 r1bio->bios[0]->bi_io_vec[i].bv_page)
156 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
158 for (i=0 ; i < pi->raid_disks; i++)
159 bio_put(r1bio->bios[i]);
161 r1bio_pool_free(r1bio, data);
164 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
168 for (i = 0; i < conf->raid_disks; i++) {
169 struct bio **bio = r1_bio->bios + i;
170 if (*bio && *bio != IO_BLOCKED)
176 static void free_r1bio(r1bio_t *r1_bio)
178 conf_t *conf = r1_bio->mddev->private;
181 * Wake up any possible resync thread that waits for the device
186 put_all_bios(conf, r1_bio);
187 mempool_free(r1_bio, conf->r1bio_pool);
190 static void put_buf(r1bio_t *r1_bio)
192 conf_t *conf = r1_bio->mddev->private;
195 for (i=0; i<conf->raid_disks; i++) {
196 struct bio *bio = r1_bio->bios[i];
198 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
201 mempool_free(r1_bio, conf->r1buf_pool);
206 static void reschedule_retry(r1bio_t *r1_bio)
209 mddev_t *mddev = r1_bio->mddev;
210 conf_t *conf = mddev->private;
212 spin_lock_irqsave(&conf->device_lock, flags);
213 list_add(&r1_bio->retry_list, &conf->retry_list);
215 spin_unlock_irqrestore(&conf->device_lock, flags);
217 wake_up(&conf->wait_barrier);
218 md_wakeup_thread(mddev->thread);
222 * raid_end_bio_io() is called when we have finished servicing a mirrored
223 * operation and are ready to return a success/failure code to the buffer
226 static void raid_end_bio_io(r1bio_t *r1_bio)
228 struct bio *bio = r1_bio->master_bio;
230 /* if nobody has done the final endio yet, do it now */
231 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
232 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
233 (bio_data_dir(bio) == WRITE) ? "write" : "read",
234 (unsigned long long) bio->bi_sector,
235 (unsigned long long) bio->bi_sector +
236 (bio->bi_size >> 9) - 1);
239 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
245 * Update disk head position estimator based on IRQ completion info.
247 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
249 conf_t *conf = r1_bio->mddev->private;
251 conf->mirrors[disk].head_position =
252 r1_bio->sector + (r1_bio->sectors);
255 static void raid1_end_read_request(struct bio *bio, int error)
257 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
258 r1bio_t *r1_bio = bio->bi_private;
260 conf_t *conf = r1_bio->mddev->private;
262 mirror = r1_bio->read_disk;
264 * this branch is our 'one mirror IO has finished' event handler:
266 update_head_pos(mirror, r1_bio);
269 set_bit(R1BIO_Uptodate, &r1_bio->state);
271 /* If all other devices have failed, we want to return
272 * the error upwards rather than fail the last device.
273 * Here we redefine "uptodate" to mean "Don't want to retry"
276 spin_lock_irqsave(&conf->device_lock, flags);
277 if (r1_bio->mddev->degraded == conf->raid_disks ||
278 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
279 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
281 spin_unlock_irqrestore(&conf->device_lock, flags);
285 raid_end_bio_io(r1_bio);
290 char b[BDEVNAME_SIZE];
292 KERN_ERR "md/raid1:%s: %s: "
293 "rescheduling sector %llu\n",
295 bdevname(conf->mirrors[mirror].rdev->bdev,
297 (unsigned long long)r1_bio->sector);
298 reschedule_retry(r1_bio);
301 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
304 static void r1_bio_write_done(r1bio_t *r1_bio)
306 if (atomic_dec_and_test(&r1_bio->remaining))
308 /* it really is the end of this request */
309 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
310 /* free extra copy of the data pages */
311 int i = r1_bio->behind_page_count;
313 safe_put_page(r1_bio->behind_pages[i]);
314 kfree(r1_bio->behind_pages);
315 r1_bio->behind_pages = NULL;
317 /* clear the bitmap if all writes complete successfully */
318 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
320 !test_bit(R1BIO_Degraded, &r1_bio->state),
321 test_bit(R1BIO_BehindIO, &r1_bio->state));
322 md_write_end(r1_bio->mddev);
323 raid_end_bio_io(r1_bio);
327 static void raid1_end_write_request(struct bio *bio, int error)
329 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
330 r1bio_t *r1_bio = bio->bi_private;
331 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
332 conf_t *conf = r1_bio->mddev->private;
333 struct bio *to_put = NULL;
336 for (mirror = 0; mirror < conf->raid_disks; mirror++)
337 if (r1_bio->bios[mirror] == bio)
341 * 'one mirror IO has finished' event handler:
343 r1_bio->bios[mirror] = NULL;
346 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
347 /* an I/O failed, we can't clear the bitmap */
348 set_bit(R1BIO_Degraded, &r1_bio->state);
351 * Set R1BIO_Uptodate in our master bio, so that we
352 * will return a good error code for to the higher
353 * levels even if IO on some other mirrored buffer
356 * The 'master' represents the composite IO operation
357 * to user-side. So if something waits for IO, then it
358 * will wait for the 'master' bio.
360 set_bit(R1BIO_Uptodate, &r1_bio->state);
362 update_head_pos(mirror, r1_bio);
365 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
366 atomic_dec(&r1_bio->behind_remaining);
369 * In behind mode, we ACK the master bio once the I/O
370 * has safely reached all non-writemostly
371 * disks. Setting the Returned bit ensures that this
372 * gets done only once -- we don't ever want to return
373 * -EIO here, instead we'll wait
375 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
376 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
377 /* Maybe we can return now */
378 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
379 struct bio *mbio = r1_bio->master_bio;
380 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
381 (unsigned long long) mbio->bi_sector,
382 (unsigned long long) mbio->bi_sector +
383 (mbio->bi_size >> 9) - 1);
388 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
391 * Let's see if all mirrored write operations have finished
394 r1_bio_write_done(r1_bio);
402 * This routine returns the disk from which the requested read should
403 * be done. There is a per-array 'next expected sequential IO' sector
404 * number - if this matches on the next IO then we use the last disk.
405 * There is also a per-disk 'last know head position' sector that is
406 * maintained from IRQ contexts, both the normal and the resync IO
407 * completion handlers update this position correctly. If there is no
408 * perfect sequential match then we pick the disk whose head is closest.
410 * If there are 2 mirrors in the same 2 devices, performance degrades
411 * because position is mirror, not device based.
413 * The rdev for the device selected will have nr_pending incremented.
415 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
417 const sector_t this_sector = r1_bio->sector;
418 const int sectors = r1_bio->sectors;
428 * Check if we can balance. We can balance on the whole
429 * device if no resync is going on, or below the resync window.
430 * We take the first readable disk when above the resync window.
434 best_dist = MaxSector;
435 if (conf->mddev->recovery_cp < MaxSector &&
436 (this_sector + sectors >= conf->next_resync)) {
441 start_disk = conf->last_used;
444 for (i = 0 ; i < conf->raid_disks ; i++) {
446 int disk = start_disk + i;
447 if (disk >= conf->raid_disks)
448 disk -= conf->raid_disks;
450 rdev = rcu_dereference(conf->mirrors[disk].rdev);
451 if (r1_bio->bios[disk] == IO_BLOCKED
453 || test_bit(Faulty, &rdev->flags))
455 if (!test_bit(In_sync, &rdev->flags) &&
456 rdev->recovery_offset < this_sector + sectors)
458 if (test_bit(WriteMostly, &rdev->flags)) {
459 /* Don't balance among write-mostly, just
460 * use the first as a last resort */
465 /* This is a reasonable device to use. It might
468 dist = abs(this_sector - conf->mirrors[disk].head_position);
470 /* Don't change to another disk for sequential reads */
471 || conf->next_seq_sect == this_sector
473 /* If device is idle, use it */
474 || atomic_read(&rdev->nr_pending) == 0) {
478 if (dist < best_dist) {
484 if (best_disk >= 0) {
485 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
488 atomic_inc(&rdev->nr_pending);
489 if (test_bit(Faulty, &rdev->flags)) {
490 /* cannot risk returning a device that failed
491 * before we inc'ed nr_pending
493 rdev_dec_pending(rdev, conf->mddev);
496 conf->next_seq_sect = this_sector + sectors;
497 conf->last_used = best_disk;
504 int md_raid1_congested(mddev_t *mddev, int bits)
506 conf_t *conf = mddev->private;
510 for (i = 0; i < mddev->raid_disks; i++) {
511 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
512 if (rdev && !test_bit(Faulty, &rdev->flags)) {
513 struct request_queue *q = bdev_get_queue(rdev->bdev);
517 /* Note the '|| 1' - when read_balance prefers
518 * non-congested targets, it can be removed
520 if ((bits & (1<<BDI_async_congested)) || 1)
521 ret |= bdi_congested(&q->backing_dev_info, bits);
523 ret &= bdi_congested(&q->backing_dev_info, bits);
529 EXPORT_SYMBOL_GPL(md_raid1_congested);
531 static int raid1_congested(void *data, int bits)
533 mddev_t *mddev = data;
535 return mddev_congested(mddev, bits) ||
536 md_raid1_congested(mddev, bits);
539 static void flush_pending_writes(conf_t *conf)
541 /* Any writes that have been queued but are awaiting
542 * bitmap updates get flushed here.
544 spin_lock_irq(&conf->device_lock);
546 if (conf->pending_bio_list.head) {
548 bio = bio_list_get(&conf->pending_bio_list);
549 spin_unlock_irq(&conf->device_lock);
550 /* flush any pending bitmap writes to
551 * disk before proceeding w/ I/O */
552 bitmap_unplug(conf->mddev->bitmap);
554 while (bio) { /* submit pending writes */
555 struct bio *next = bio->bi_next;
557 generic_make_request(bio);
561 spin_unlock_irq(&conf->device_lock);
565 * Sometimes we need to suspend IO while we do something else,
566 * either some resync/recovery, or reconfigure the array.
567 * To do this we raise a 'barrier'.
568 * The 'barrier' is a counter that can be raised multiple times
569 * to count how many activities are happening which preclude
571 * We can only raise the barrier if there is no pending IO.
572 * i.e. if nr_pending == 0.
573 * We choose only to raise the barrier if no-one is waiting for the
574 * barrier to go down. This means that as soon as an IO request
575 * is ready, no other operations which require a barrier will start
576 * until the IO request has had a chance.
578 * So: regular IO calls 'wait_barrier'. When that returns there
579 * is no backgroup IO happening, It must arrange to call
580 * allow_barrier when it has finished its IO.
581 * backgroup IO calls must call raise_barrier. Once that returns
582 * there is no normal IO happeing. It must arrange to call
583 * lower_barrier when the particular background IO completes.
585 #define RESYNC_DEPTH 32
587 static void raise_barrier(conf_t *conf)
589 spin_lock_irq(&conf->resync_lock);
591 /* Wait until no block IO is waiting */
592 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
593 conf->resync_lock, );
595 /* block any new IO from starting */
598 /* Now wait for all pending IO to complete */
599 wait_event_lock_irq(conf->wait_barrier,
600 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
601 conf->resync_lock, );
603 spin_unlock_irq(&conf->resync_lock);
606 static void lower_barrier(conf_t *conf)
609 BUG_ON(conf->barrier <= 0);
610 spin_lock_irqsave(&conf->resync_lock, flags);
612 spin_unlock_irqrestore(&conf->resync_lock, flags);
613 wake_up(&conf->wait_barrier);
616 static void wait_barrier(conf_t *conf)
618 spin_lock_irq(&conf->resync_lock);
621 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
627 spin_unlock_irq(&conf->resync_lock);
630 static void allow_barrier(conf_t *conf)
633 spin_lock_irqsave(&conf->resync_lock, flags);
635 spin_unlock_irqrestore(&conf->resync_lock, flags);
636 wake_up(&conf->wait_barrier);
639 static void freeze_array(conf_t *conf)
641 /* stop syncio and normal IO and wait for everything to
643 * We increment barrier and nr_waiting, and then
644 * wait until nr_pending match nr_queued+1
645 * This is called in the context of one normal IO request
646 * that has failed. Thus any sync request that might be pending
647 * will be blocked by nr_pending, and we need to wait for
648 * pending IO requests to complete or be queued for re-try.
649 * Thus the number queued (nr_queued) plus this request (1)
650 * must match the number of pending IOs (nr_pending) before
653 spin_lock_irq(&conf->resync_lock);
656 wait_event_lock_irq(conf->wait_barrier,
657 conf->nr_pending == conf->nr_queued+1,
659 flush_pending_writes(conf));
660 spin_unlock_irq(&conf->resync_lock);
662 static void unfreeze_array(conf_t *conf)
664 /* reverse the effect of the freeze */
665 spin_lock_irq(&conf->resync_lock);
668 wake_up(&conf->wait_barrier);
669 spin_unlock_irq(&conf->resync_lock);
673 /* duplicate the data pages for behind I/O
675 static void alloc_behind_pages(struct bio *bio, r1bio_t *r1_bio)
678 struct bio_vec *bvec;
679 struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page*),
681 if (unlikely(!pages))
684 bio_for_each_segment(bvec, bio, i) {
685 pages[i] = alloc_page(GFP_NOIO);
686 if (unlikely(!pages[i]))
688 memcpy(kmap(pages[i]) + bvec->bv_offset,
689 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
691 kunmap(bvec->bv_page);
693 r1_bio->behind_pages = pages;
694 r1_bio->behind_page_count = bio->bi_vcnt;
695 set_bit(R1BIO_BehindIO, &r1_bio->state);
699 for (i = 0; i < bio->bi_vcnt; i++)
703 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
706 static int make_request(mddev_t *mddev, struct bio * bio)
708 conf_t *conf = mddev->private;
709 mirror_info_t *mirror;
711 struct bio *read_bio;
712 int i, targets = 0, disks;
713 struct bitmap *bitmap;
715 const int rw = bio_data_dir(bio);
716 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
717 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
718 mdk_rdev_t *blocked_rdev;
722 * Register the new request and wait if the reconstruction
723 * thread has put up a bar for new requests.
724 * Continue immediately if no resync is active currently.
727 md_write_start(mddev, bio); /* wait on superblock update early */
729 if (bio_data_dir(bio) == WRITE &&
730 bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
731 bio->bi_sector < mddev->suspend_hi) {
732 /* As the suspend_* range is controlled by
733 * userspace, we want an interruptible
738 flush_signals(current);
739 prepare_to_wait(&conf->wait_barrier,
740 &w, TASK_INTERRUPTIBLE);
741 if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
742 bio->bi_sector >= mddev->suspend_hi)
746 finish_wait(&conf->wait_barrier, &w);
751 bitmap = mddev->bitmap;
754 * make_request() can abort the operation when READA is being
755 * used and no empty request is available.
758 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
760 r1_bio->master_bio = bio;
761 r1_bio->sectors = bio->bi_size >> 9;
763 r1_bio->mddev = mddev;
764 r1_bio->sector = bio->bi_sector;
768 * read balancing logic:
770 int rdisk = read_balance(conf, r1_bio);
773 /* couldn't find anywhere to read from */
774 raid_end_bio_io(r1_bio);
777 mirror = conf->mirrors + rdisk;
779 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
781 /* Reading from a write-mostly device must
782 * take care not to over-take any writes
785 wait_event(bitmap->behind_wait,
786 atomic_read(&bitmap->behind_writes) == 0);
788 r1_bio->read_disk = rdisk;
790 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
792 r1_bio->bios[rdisk] = read_bio;
794 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
795 read_bio->bi_bdev = mirror->rdev->bdev;
796 read_bio->bi_end_io = raid1_end_read_request;
797 read_bio->bi_rw = READ | do_sync;
798 read_bio->bi_private = r1_bio;
800 generic_make_request(read_bio);
807 /* first select target devices under spinlock and
808 * inc refcount on their rdev. Record them by setting
811 plugged = mddev_check_plugged(mddev);
813 disks = conf->raid_disks;
817 for (i = 0; i < disks; i++) {
818 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
819 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
820 atomic_inc(&rdev->nr_pending);
824 if (rdev && !test_bit(Faulty, &rdev->flags)) {
825 atomic_inc(&rdev->nr_pending);
826 if (test_bit(Faulty, &rdev->flags)) {
827 rdev_dec_pending(rdev, mddev);
828 r1_bio->bios[i] = NULL;
830 r1_bio->bios[i] = bio;
834 r1_bio->bios[i] = NULL;
838 if (unlikely(blocked_rdev)) {
839 /* Wait for this device to become unblocked */
842 for (j = 0; j < i; j++)
844 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
847 md_wait_for_blocked_rdev(blocked_rdev, mddev);
852 BUG_ON(targets == 0); /* we never fail the last device */
854 if (targets < conf->raid_disks) {
855 /* array is degraded, we will not clear the bitmap
856 * on I/O completion (see raid1_end_write_request) */
857 set_bit(R1BIO_Degraded, &r1_bio->state);
861 * Not if there are too many, or cannot allocate memory,
862 * or a reader on WriteMostly is waiting for behind writes
865 (atomic_read(&bitmap->behind_writes)
866 < mddev->bitmap_info.max_write_behind) &&
867 !waitqueue_active(&bitmap->behind_wait))
868 alloc_behind_pages(bio, r1_bio);
870 atomic_set(&r1_bio->remaining, 1);
871 atomic_set(&r1_bio->behind_remaining, 0);
873 bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
874 test_bit(R1BIO_BehindIO, &r1_bio->state));
875 for (i = 0; i < disks; i++) {
877 if (!r1_bio->bios[i])
880 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
881 r1_bio->bios[i] = mbio;
883 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
884 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
885 mbio->bi_end_io = raid1_end_write_request;
886 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
887 mbio->bi_private = r1_bio;
889 if (r1_bio->behind_pages) {
890 struct bio_vec *bvec;
893 /* Yes, I really want the '__' version so that
894 * we clear any unused pointer in the io_vec, rather
895 * than leave them unchanged. This is important
896 * because when we come to free the pages, we won't
897 * know the original bi_idx, so we just free
900 __bio_for_each_segment(bvec, mbio, j, 0)
901 bvec->bv_page = r1_bio->behind_pages[j];
902 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
903 atomic_inc(&r1_bio->behind_remaining);
906 atomic_inc(&r1_bio->remaining);
907 spin_lock_irqsave(&conf->device_lock, flags);
908 bio_list_add(&conf->pending_bio_list, mbio);
909 spin_unlock_irqrestore(&conf->device_lock, flags);
911 r1_bio_write_done(r1_bio);
913 /* In case raid1d snuck in to freeze_array */
914 wake_up(&conf->wait_barrier);
916 if (do_sync || !bitmap || !plugged)
917 md_wakeup_thread(mddev->thread);
922 static void status(struct seq_file *seq, mddev_t *mddev)
924 conf_t *conf = mddev->private;
927 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
928 conf->raid_disks - mddev->degraded);
930 for (i = 0; i < conf->raid_disks; i++) {
931 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
932 seq_printf(seq, "%s",
933 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
936 seq_printf(seq, "]");
940 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
942 char b[BDEVNAME_SIZE];
943 conf_t *conf = mddev->private;
946 * If it is not operational, then we have already marked it as dead
947 * else if it is the last working disks, ignore the error, let the
948 * next level up know.
949 * else mark the drive as failed
951 if (test_bit(In_sync, &rdev->flags)
952 && (conf->raid_disks - mddev->degraded) == 1) {
954 * Don't fail the drive, act as though we were just a
955 * normal single drive.
956 * However don't try a recovery from this drive as
957 * it is very likely to fail.
959 conf->recovery_disabled = mddev->recovery_disabled;
962 if (test_and_clear_bit(In_sync, &rdev->flags)) {
964 spin_lock_irqsave(&conf->device_lock, flags);
966 set_bit(Faulty, &rdev->flags);
967 spin_unlock_irqrestore(&conf->device_lock, flags);
969 * if recovery is running, make sure it aborts.
971 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
973 set_bit(Faulty, &rdev->flags);
974 set_bit(MD_CHANGE_DEVS, &mddev->flags);
976 "md/raid1:%s: Disk failure on %s, disabling device.\n"
977 "md/raid1:%s: Operation continuing on %d devices.\n",
978 mdname(mddev), bdevname(rdev->bdev, b),
979 mdname(mddev), conf->raid_disks - mddev->degraded);
982 static void print_conf(conf_t *conf)
986 printk(KERN_DEBUG "RAID1 conf printout:\n");
988 printk(KERN_DEBUG "(!conf)\n");
991 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
995 for (i = 0; i < conf->raid_disks; i++) {
996 char b[BDEVNAME_SIZE];
997 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
999 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1000 i, !test_bit(In_sync, &rdev->flags),
1001 !test_bit(Faulty, &rdev->flags),
1002 bdevname(rdev->bdev,b));
1007 static void close_sync(conf_t *conf)
1010 allow_barrier(conf);
1012 mempool_destroy(conf->r1buf_pool);
1013 conf->r1buf_pool = NULL;
1016 static int raid1_spare_active(mddev_t *mddev)
1019 conf_t *conf = mddev->private;
1021 unsigned long flags;
1024 * Find all failed disks within the RAID1 configuration
1025 * and mark them readable.
1026 * Called under mddev lock, so rcu protection not needed.
1028 for (i = 0; i < conf->raid_disks; i++) {
1029 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1031 && !test_bit(Faulty, &rdev->flags)
1032 && !test_and_set_bit(In_sync, &rdev->flags)) {
1034 sysfs_notify_dirent_safe(rdev->sysfs_state);
1037 spin_lock_irqsave(&conf->device_lock, flags);
1038 mddev->degraded -= count;
1039 spin_unlock_irqrestore(&conf->device_lock, flags);
1046 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1048 conf_t *conf = mddev->private;
1053 int last = mddev->raid_disks - 1;
1055 if (mddev->recovery_disabled == conf->recovery_disabled)
1058 if (rdev->raid_disk >= 0)
1059 first = last = rdev->raid_disk;
1061 for (mirror = first; mirror <= last; mirror++)
1062 if ( !(p=conf->mirrors+mirror)->rdev) {
1064 disk_stack_limits(mddev->gendisk, rdev->bdev,
1065 rdev->data_offset << 9);
1066 /* as we don't honour merge_bvec_fn, we must
1067 * never risk violating it, so limit
1068 * ->max_segments to one lying with a single
1069 * page, as a one page request is never in
1072 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1073 blk_queue_max_segments(mddev->queue, 1);
1074 blk_queue_segment_boundary(mddev->queue,
1075 PAGE_CACHE_SIZE - 1);
1078 p->head_position = 0;
1079 rdev->raid_disk = mirror;
1081 /* As all devices are equivalent, we don't need a full recovery
1082 * if this was recently any drive of the array
1084 if (rdev->saved_raid_disk < 0)
1086 rcu_assign_pointer(p->rdev, rdev);
1089 md_integrity_add_rdev(rdev, mddev);
1094 static int raid1_remove_disk(mddev_t *mddev, int number)
1096 conf_t *conf = mddev->private;
1099 mirror_info_t *p = conf->mirrors+ number;
1104 if (test_bit(In_sync, &rdev->flags) ||
1105 atomic_read(&rdev->nr_pending)) {
1109 /* Only remove non-faulty devices if recovery
1112 if (!test_bit(Faulty, &rdev->flags) &&
1113 mddev->recovery_disabled != conf->recovery_disabled &&
1114 mddev->degraded < conf->raid_disks) {
1120 if (atomic_read(&rdev->nr_pending)) {
1121 /* lost the race, try later */
1126 err = md_integrity_register(mddev);
1135 static void end_sync_read(struct bio *bio, int error)
1137 r1bio_t *r1_bio = bio->bi_private;
1140 for (i=r1_bio->mddev->raid_disks; i--; )
1141 if (r1_bio->bios[i] == bio)
1144 update_head_pos(i, r1_bio);
1146 * we have read a block, now it needs to be re-written,
1147 * or re-read if the read failed.
1148 * We don't do much here, just schedule handling by raid1d
1150 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1151 set_bit(R1BIO_Uptodate, &r1_bio->state);
1153 if (atomic_dec_and_test(&r1_bio->remaining))
1154 reschedule_retry(r1_bio);
1157 static void end_sync_write(struct bio *bio, int error)
1159 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1160 r1bio_t *r1_bio = bio->bi_private;
1161 mddev_t *mddev = r1_bio->mddev;
1162 conf_t *conf = mddev->private;
1166 for (i = 0; i < conf->raid_disks; i++)
1167 if (r1_bio->bios[i] == bio) {
1172 sector_t sync_blocks = 0;
1173 sector_t s = r1_bio->sector;
1174 long sectors_to_go = r1_bio->sectors;
1175 /* make sure these bits doesn't get cleared. */
1177 bitmap_end_sync(mddev->bitmap, s,
1180 sectors_to_go -= sync_blocks;
1181 } while (sectors_to_go > 0);
1182 md_error(mddev, conf->mirrors[mirror].rdev);
1185 update_head_pos(mirror, r1_bio);
1187 if (atomic_dec_and_test(&r1_bio->remaining)) {
1188 sector_t s = r1_bio->sectors;
1190 md_done_sync(mddev, s, uptodate);
1194 static int fix_sync_read_error(r1bio_t *r1_bio)
1196 /* Try some synchronous reads of other devices to get
1197 * good data, much like with normal read errors. Only
1198 * read into the pages we already have so we don't
1199 * need to re-issue the read request.
1200 * We don't need to freeze the array, because being in an
1201 * active sync request, there is no normal IO, and
1202 * no overlapping syncs.
1204 mddev_t *mddev = r1_bio->mddev;
1205 conf_t *conf = mddev->private;
1206 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1207 sector_t sect = r1_bio->sector;
1208 int sectors = r1_bio->sectors;
1213 int d = r1_bio->read_disk;
1218 if (s > (PAGE_SIZE>>9))
1221 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1222 /* No rcu protection needed here devices
1223 * can only be removed when no resync is
1224 * active, and resync is currently active
1226 rdev = conf->mirrors[d].rdev;
1227 if (sync_page_io(rdev, sect, s<<9,
1228 bio->bi_io_vec[idx].bv_page,
1235 if (d == conf->raid_disks)
1237 } while (!success && d != r1_bio->read_disk);
1240 char b[BDEVNAME_SIZE];
1241 /* Cannot read from anywhere, array is toast */
1242 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1243 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1244 " for block %llu\n",
1246 bdevname(bio->bi_bdev, b),
1247 (unsigned long long)r1_bio->sector);
1248 md_done_sync(mddev, r1_bio->sectors, 0);
1254 /* write it back and re-read */
1255 while (d != r1_bio->read_disk) {
1257 d = conf->raid_disks;
1259 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1261 rdev = conf->mirrors[d].rdev;
1262 if (sync_page_io(rdev, sect, s<<9,
1263 bio->bi_io_vec[idx].bv_page,
1264 WRITE, false) == 0) {
1265 r1_bio->bios[d]->bi_end_io = NULL;
1266 rdev_dec_pending(rdev, mddev);
1267 md_error(mddev, rdev);
1271 while (d != r1_bio->read_disk) {
1273 d = conf->raid_disks;
1275 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1277 rdev = conf->mirrors[d].rdev;
1278 if (sync_page_io(rdev, sect, s<<9,
1279 bio->bi_io_vec[idx].bv_page,
1281 md_error(mddev, rdev);
1283 atomic_add(s, &rdev->corrected_errors);
1289 set_bit(R1BIO_Uptodate, &r1_bio->state);
1290 set_bit(BIO_UPTODATE, &bio->bi_flags);
1294 static int process_checks(r1bio_t *r1_bio)
1296 /* We have read all readable devices. If we haven't
1297 * got the block, then there is no hope left.
1298 * If we have, then we want to do a comparison
1299 * and skip the write if everything is the same.
1300 * If any blocks failed to read, then we need to
1301 * attempt an over-write
1303 mddev_t *mddev = r1_bio->mddev;
1304 conf_t *conf = mddev->private;
1308 for (primary = 0; primary < conf->raid_disks; primary++)
1309 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1310 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1311 r1_bio->bios[primary]->bi_end_io = NULL;
1312 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1315 r1_bio->read_disk = primary;
1316 for (i = 0; i < conf->raid_disks; i++) {
1318 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1319 struct bio *pbio = r1_bio->bios[primary];
1320 struct bio *sbio = r1_bio->bios[i];
1323 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1326 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1327 for (j = vcnt; j-- ; ) {
1329 p = pbio->bi_io_vec[j].bv_page;
1330 s = sbio->bi_io_vec[j].bv_page;
1331 if (memcmp(page_address(p),
1339 mddev->resync_mismatches += r1_bio->sectors;
1340 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1341 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1342 /* No need to write to this device. */
1343 sbio->bi_end_io = NULL;
1344 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1347 /* fixup the bio for reuse */
1348 sbio->bi_vcnt = vcnt;
1349 sbio->bi_size = r1_bio->sectors << 9;
1351 sbio->bi_phys_segments = 0;
1352 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1353 sbio->bi_flags |= 1 << BIO_UPTODATE;
1354 sbio->bi_next = NULL;
1355 sbio->bi_sector = r1_bio->sector +
1356 conf->mirrors[i].rdev->data_offset;
1357 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1358 size = sbio->bi_size;
1359 for (j = 0; j < vcnt ; j++) {
1361 bi = &sbio->bi_io_vec[j];
1363 if (size > PAGE_SIZE)
1364 bi->bv_len = PAGE_SIZE;
1368 memcpy(page_address(bi->bv_page),
1369 page_address(pbio->bi_io_vec[j].bv_page),
1376 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1378 conf_t *conf = mddev->private;
1380 int disks = conf->raid_disks;
1381 struct bio *bio, *wbio;
1383 bio = r1_bio->bios[r1_bio->read_disk];
1385 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1386 /* ouch - failed to read all of that. */
1387 if (!fix_sync_read_error(r1_bio))
1390 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1391 if (process_checks(r1_bio) < 0)
1396 atomic_set(&r1_bio->remaining, 1);
1397 for (i = 0; i < disks ; i++) {
1398 wbio = r1_bio->bios[i];
1399 if (wbio->bi_end_io == NULL ||
1400 (wbio->bi_end_io == end_sync_read &&
1401 (i == r1_bio->read_disk ||
1402 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1405 wbio->bi_rw = WRITE;
1406 wbio->bi_end_io = end_sync_write;
1407 atomic_inc(&r1_bio->remaining);
1408 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1410 generic_make_request(wbio);
1413 if (atomic_dec_and_test(&r1_bio->remaining)) {
1414 /* if we're here, all write(s) have completed, so clean up */
1415 md_done_sync(mddev, r1_bio->sectors, 1);
1421 * This is a kernel thread which:
1423 * 1. Retries failed read operations on working mirrors.
1424 * 2. Updates the raid superblock when problems encounter.
1425 * 3. Performs writes following reads for array syncronising.
1428 static void fix_read_error(conf_t *conf, int read_disk,
1429 sector_t sect, int sectors)
1431 mddev_t *mddev = conf->mddev;
1439 if (s > (PAGE_SIZE>>9))
1443 /* Note: no rcu protection needed here
1444 * as this is synchronous in the raid1d thread
1445 * which is the thread that might remove
1446 * a device. If raid1d ever becomes multi-threaded....
1448 rdev = conf->mirrors[d].rdev;
1450 test_bit(In_sync, &rdev->flags) &&
1451 sync_page_io(rdev, sect, s<<9,
1452 conf->tmppage, READ, false))
1456 if (d == conf->raid_disks)
1459 } while (!success && d != read_disk);
1462 /* Cannot read from anywhere -- bye bye array */
1463 md_error(mddev, conf->mirrors[read_disk].rdev);
1466 /* write it back and re-read */
1468 while (d != read_disk) {
1470 d = conf->raid_disks;
1472 rdev = conf->mirrors[d].rdev;
1474 test_bit(In_sync, &rdev->flags)) {
1475 if (sync_page_io(rdev, sect, s<<9,
1476 conf->tmppage, WRITE, false)
1478 /* Well, this device is dead */
1479 md_error(mddev, rdev);
1483 while (d != read_disk) {
1484 char b[BDEVNAME_SIZE];
1486 d = conf->raid_disks;
1488 rdev = conf->mirrors[d].rdev;
1490 test_bit(In_sync, &rdev->flags)) {
1491 if (sync_page_io(rdev, sect, s<<9,
1492 conf->tmppage, READ, false)
1494 /* Well, this device is dead */
1495 md_error(mddev, rdev);
1497 atomic_add(s, &rdev->corrected_errors);
1499 "md/raid1:%s: read error corrected "
1500 "(%d sectors at %llu on %s)\n",
1502 (unsigned long long)(sect +
1504 bdevname(rdev->bdev, b));
1513 static void raid1d(mddev_t *mddev)
1517 unsigned long flags;
1518 conf_t *conf = mddev->private;
1519 struct list_head *head = &conf->retry_list;
1521 struct blk_plug plug;
1523 md_check_recovery(mddev);
1525 blk_start_plug(&plug);
1527 char b[BDEVNAME_SIZE];
1529 if (atomic_read(&mddev->plug_cnt) == 0)
1530 flush_pending_writes(conf);
1532 spin_lock_irqsave(&conf->device_lock, flags);
1533 if (list_empty(head)) {
1534 spin_unlock_irqrestore(&conf->device_lock, flags);
1537 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1538 list_del(head->prev);
1540 spin_unlock_irqrestore(&conf->device_lock, flags);
1542 mddev = r1_bio->mddev;
1543 conf = mddev->private;
1544 if (test_bit(R1BIO_IsSync, &r1_bio->state))
1545 sync_request_write(mddev, r1_bio);
1549 /* we got a read error. Maybe the drive is bad. Maybe just
1550 * the block and we can fix it.
1551 * We freeze all other IO, and try reading the block from
1552 * other devices. When we find one, we re-write
1553 * and check it that fixes the read error.
1554 * This is all done synchronously while the array is
1557 if (mddev->ro == 0) {
1559 fix_read_error(conf, r1_bio->read_disk,
1562 unfreeze_array(conf);
1565 conf->mirrors[r1_bio->read_disk].rdev);
1567 bio = r1_bio->bios[r1_bio->read_disk];
1568 if ((disk=read_balance(conf, r1_bio)) == -1) {
1569 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1570 " read error for block %llu\n",
1572 bdevname(bio->bi_bdev,b),
1573 (unsigned long long)r1_bio->sector);
1574 raid_end_bio_io(r1_bio);
1576 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1577 r1_bio->bios[r1_bio->read_disk] =
1578 mddev->ro ? IO_BLOCKED : NULL;
1579 r1_bio->read_disk = disk;
1581 bio = bio_clone_mddev(r1_bio->master_bio,
1583 r1_bio->bios[r1_bio->read_disk] = bio;
1584 rdev = conf->mirrors[disk].rdev;
1587 "md/raid1:%s: redirecting sector %llu"
1588 " to other mirror: %s\n",
1590 (unsigned long long)r1_bio->sector,
1591 bdevname(rdev->bdev, b));
1592 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1593 bio->bi_bdev = rdev->bdev;
1594 bio->bi_end_io = raid1_end_read_request;
1595 bio->bi_rw = READ | do_sync;
1596 bio->bi_private = r1_bio;
1597 generic_make_request(bio);
1602 blk_finish_plug(&plug);
1606 static int init_resync(conf_t *conf)
1610 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1611 BUG_ON(conf->r1buf_pool);
1612 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1614 if (!conf->r1buf_pool)
1616 conf->next_resync = 0;
1621 * perform a "sync" on one "block"
1623 * We need to make sure that no normal I/O request - particularly write
1624 * requests - conflict with active sync requests.
1626 * This is achieved by tracking pending requests and a 'barrier' concept
1627 * that can be installed to exclude normal IO requests.
1630 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1632 conf_t *conf = mddev->private;
1635 sector_t max_sector, nr_sectors;
1639 int write_targets = 0, read_targets = 0;
1640 sector_t sync_blocks;
1641 int still_degraded = 0;
1643 if (!conf->r1buf_pool)
1644 if (init_resync(conf))
1647 max_sector = mddev->dev_sectors;
1648 if (sector_nr >= max_sector) {
1649 /* If we aborted, we need to abort the
1650 * sync on the 'current' bitmap chunk (there will
1651 * only be one in raid1 resync.
1652 * We can find the current addess in mddev->curr_resync
1654 if (mddev->curr_resync < max_sector) /* aborted */
1655 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1657 else /* completed sync */
1660 bitmap_close_sync(mddev->bitmap);
1665 if (mddev->bitmap == NULL &&
1666 mddev->recovery_cp == MaxSector &&
1667 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1668 conf->fullsync == 0) {
1670 return max_sector - sector_nr;
1672 /* before building a request, check if we can skip these blocks..
1673 * This call the bitmap_start_sync doesn't actually record anything
1675 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1676 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1677 /* We can skip this block, and probably several more */
1682 * If there is non-resync activity waiting for a turn,
1683 * and resync is going fast enough,
1684 * then let it though before starting on this new sync request.
1686 if (!go_faster && conf->nr_waiting)
1687 msleep_interruptible(1000);
1689 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1690 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1691 raise_barrier(conf);
1693 conf->next_resync = sector_nr;
1697 * If we get a correctably read error during resync or recovery,
1698 * we might want to read from a different device. So we
1699 * flag all drives that could conceivably be read from for READ,
1700 * and any others (which will be non-In_sync devices) for WRITE.
1701 * If a read fails, we try reading from something else for which READ
1705 r1_bio->mddev = mddev;
1706 r1_bio->sector = sector_nr;
1708 set_bit(R1BIO_IsSync, &r1_bio->state);
1710 for (i=0; i < conf->raid_disks; i++) {
1712 bio = r1_bio->bios[i];
1714 /* take from bio_init */
1715 bio->bi_next = NULL;
1716 bio->bi_flags &= ~(BIO_POOL_MASK-1);
1717 bio->bi_flags |= 1 << BIO_UPTODATE;
1718 bio->bi_comp_cpu = -1;
1722 bio->bi_phys_segments = 0;
1724 bio->bi_end_io = NULL;
1725 bio->bi_private = NULL;
1727 rdev = rcu_dereference(conf->mirrors[i].rdev);
1729 test_bit(Faulty, &rdev->flags)) {
1732 } else if (!test_bit(In_sync, &rdev->flags)) {
1734 bio->bi_end_io = end_sync_write;
1737 /* may need to read from here */
1739 bio->bi_end_io = end_sync_read;
1740 if (test_bit(WriteMostly, &rdev->flags)) {
1749 atomic_inc(&rdev->nr_pending);
1750 bio->bi_sector = sector_nr + rdev->data_offset;
1751 bio->bi_bdev = rdev->bdev;
1752 bio->bi_private = r1_bio;
1757 r1_bio->read_disk = disk;
1759 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1760 /* extra read targets are also write targets */
1761 write_targets += read_targets-1;
1763 if (write_targets == 0 || read_targets == 0) {
1764 /* There is nowhere to write, so all non-sync
1765 * drives must be failed - so we are finished
1767 sector_t rv = max_sector - sector_nr;
1773 if (max_sector > mddev->resync_max)
1774 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1779 int len = PAGE_SIZE;
1780 if (sector_nr + (len>>9) > max_sector)
1781 len = (max_sector - sector_nr) << 9;
1784 if (sync_blocks == 0) {
1785 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1786 &sync_blocks, still_degraded) &&
1788 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1790 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1791 if ((len >> 9) > sync_blocks)
1792 len = sync_blocks<<9;
1795 for (i=0 ; i < conf->raid_disks; i++) {
1796 bio = r1_bio->bios[i];
1797 if (bio->bi_end_io) {
1798 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1799 if (bio_add_page(bio, page, len, 0) == 0) {
1801 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1804 bio = r1_bio->bios[i];
1805 if (bio->bi_end_io==NULL)
1807 /* remove last page from this bio */
1809 bio->bi_size -= len;
1810 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1816 nr_sectors += len>>9;
1817 sector_nr += len>>9;
1818 sync_blocks -= (len>>9);
1819 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1821 r1_bio->sectors = nr_sectors;
1823 /* For a user-requested sync, we read all readable devices and do a
1826 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1827 atomic_set(&r1_bio->remaining, read_targets);
1828 for (i=0; i<conf->raid_disks; i++) {
1829 bio = r1_bio->bios[i];
1830 if (bio->bi_end_io == end_sync_read) {
1831 md_sync_acct(bio->bi_bdev, nr_sectors);
1832 generic_make_request(bio);
1836 atomic_set(&r1_bio->remaining, 1);
1837 bio = r1_bio->bios[r1_bio->read_disk];
1838 md_sync_acct(bio->bi_bdev, nr_sectors);
1839 generic_make_request(bio);
1845 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1850 return mddev->dev_sectors;
1853 static conf_t *setup_conf(mddev_t *mddev)
1857 mirror_info_t *disk;
1861 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1865 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1870 conf->tmppage = alloc_page(GFP_KERNEL);
1874 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1875 if (!conf->poolinfo)
1877 conf->poolinfo->raid_disks = mddev->raid_disks;
1878 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1881 if (!conf->r1bio_pool)
1884 conf->poolinfo->mddev = mddev;
1886 spin_lock_init(&conf->device_lock);
1887 list_for_each_entry(rdev, &mddev->disks, same_set) {
1888 int disk_idx = rdev->raid_disk;
1889 if (disk_idx >= mddev->raid_disks
1892 disk = conf->mirrors + disk_idx;
1896 disk->head_position = 0;
1898 conf->raid_disks = mddev->raid_disks;
1899 conf->mddev = mddev;
1900 INIT_LIST_HEAD(&conf->retry_list);
1902 spin_lock_init(&conf->resync_lock);
1903 init_waitqueue_head(&conf->wait_barrier);
1905 bio_list_init(&conf->pending_bio_list);
1907 conf->last_used = -1;
1908 for (i = 0; i < conf->raid_disks; i++) {
1910 disk = conf->mirrors + i;
1913 !test_bit(In_sync, &disk->rdev->flags)) {
1914 disk->head_position = 0;
1917 } else if (conf->last_used < 0)
1919 * The first working device is used as a
1920 * starting point to read balancing.
1922 conf->last_used = i;
1926 if (conf->last_used < 0) {
1927 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
1932 conf->thread = md_register_thread(raid1d, mddev, NULL);
1933 if (!conf->thread) {
1935 "md/raid1:%s: couldn't allocate thread\n",
1944 if (conf->r1bio_pool)
1945 mempool_destroy(conf->r1bio_pool);
1946 kfree(conf->mirrors);
1947 safe_put_page(conf->tmppage);
1948 kfree(conf->poolinfo);
1951 return ERR_PTR(err);
1954 static int run(mddev_t *mddev)
1960 if (mddev->level != 1) {
1961 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
1962 mdname(mddev), mddev->level);
1965 if (mddev->reshape_position != MaxSector) {
1966 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
1971 * copy the already verified devices into our private RAID1
1972 * bookkeeping area. [whatever we allocate in run(),
1973 * should be freed in stop()]
1975 if (mddev->private == NULL)
1976 conf = setup_conf(mddev);
1978 conf = mddev->private;
1981 return PTR_ERR(conf);
1983 list_for_each_entry(rdev, &mddev->disks, same_set) {
1984 if (!mddev->gendisk)
1986 disk_stack_limits(mddev->gendisk, rdev->bdev,
1987 rdev->data_offset << 9);
1988 /* as we don't honour merge_bvec_fn, we must never risk
1989 * violating it, so limit ->max_segments to 1 lying within
1990 * a single page, as a one page request is never in violation.
1992 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1993 blk_queue_max_segments(mddev->queue, 1);
1994 blk_queue_segment_boundary(mddev->queue,
1995 PAGE_CACHE_SIZE - 1);
1999 mddev->degraded = 0;
2000 for (i=0; i < conf->raid_disks; i++)
2001 if (conf->mirrors[i].rdev == NULL ||
2002 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2003 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2006 if (conf->raid_disks - mddev->degraded == 1)
2007 mddev->recovery_cp = MaxSector;
2009 if (mddev->recovery_cp != MaxSector)
2010 printk(KERN_NOTICE "md/raid1:%s: not clean"
2011 " -- starting background reconstruction\n",
2014 "md/raid1:%s: active with %d out of %d mirrors\n",
2015 mdname(mddev), mddev->raid_disks - mddev->degraded,
2019 * Ok, everything is just fine now
2021 mddev->thread = conf->thread;
2022 conf->thread = NULL;
2023 mddev->private = conf;
2025 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2028 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2029 mddev->queue->backing_dev_info.congested_data = mddev;
2031 return md_integrity_register(mddev);
2034 static int stop(mddev_t *mddev)
2036 conf_t *conf = mddev->private;
2037 struct bitmap *bitmap = mddev->bitmap;
2039 /* wait for behind writes to complete */
2040 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2041 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2043 /* need to kick something here to make sure I/O goes? */
2044 wait_event(bitmap->behind_wait,
2045 atomic_read(&bitmap->behind_writes) == 0);
2048 raise_barrier(conf);
2049 lower_barrier(conf);
2051 md_unregister_thread(mddev->thread);
2052 mddev->thread = NULL;
2053 if (conf->r1bio_pool)
2054 mempool_destroy(conf->r1bio_pool);
2055 kfree(conf->mirrors);
2056 kfree(conf->poolinfo);
2058 mddev->private = NULL;
2062 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2064 /* no resync is happening, and there is enough space
2065 * on all devices, so we can resize.
2066 * We need to make sure resync covers any new space.
2067 * If the array is shrinking we should possibly wait until
2068 * any io in the removed space completes, but it hardly seems
2071 md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2072 if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2074 set_capacity(mddev->gendisk, mddev->array_sectors);
2075 revalidate_disk(mddev->gendisk);
2076 if (sectors > mddev->dev_sectors &&
2077 mddev->recovery_cp > mddev->dev_sectors) {
2078 mddev->recovery_cp = mddev->dev_sectors;
2079 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2081 mddev->dev_sectors = sectors;
2082 mddev->resync_max_sectors = sectors;
2086 static int raid1_reshape(mddev_t *mddev)
2089 * 1/ resize the r1bio_pool
2090 * 2/ resize conf->mirrors
2092 * We allocate a new r1bio_pool if we can.
2093 * Then raise a device barrier and wait until all IO stops.
2094 * Then resize conf->mirrors and swap in the new r1bio pool.
2096 * At the same time, we "pack" the devices so that all the missing
2097 * devices have the higher raid_disk numbers.
2099 mempool_t *newpool, *oldpool;
2100 struct pool_info *newpoolinfo;
2101 mirror_info_t *newmirrors;
2102 conf_t *conf = mddev->private;
2103 int cnt, raid_disks;
2104 unsigned long flags;
2107 /* Cannot change chunk_size, layout, or level */
2108 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2109 mddev->layout != mddev->new_layout ||
2110 mddev->level != mddev->new_level) {
2111 mddev->new_chunk_sectors = mddev->chunk_sectors;
2112 mddev->new_layout = mddev->layout;
2113 mddev->new_level = mddev->level;
2117 err = md_allow_write(mddev);
2121 raid_disks = mddev->raid_disks + mddev->delta_disks;
2123 if (raid_disks < conf->raid_disks) {
2125 for (d= 0; d < conf->raid_disks; d++)
2126 if (conf->mirrors[d].rdev)
2128 if (cnt > raid_disks)
2132 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2135 newpoolinfo->mddev = mddev;
2136 newpoolinfo->raid_disks = raid_disks;
2138 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2139 r1bio_pool_free, newpoolinfo);
2144 newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2147 mempool_destroy(newpool);
2151 raise_barrier(conf);
2153 /* ok, everything is stopped */
2154 oldpool = conf->r1bio_pool;
2155 conf->r1bio_pool = newpool;
2157 for (d = d2 = 0; d < conf->raid_disks; d++) {
2158 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2159 if (rdev && rdev->raid_disk != d2) {
2160 sysfs_unlink_rdev(mddev, rdev);
2161 rdev->raid_disk = d2;
2162 sysfs_unlink_rdev(mddev, rdev);
2163 if (sysfs_link_rdev(mddev, rdev))
2165 "md/raid1:%s: cannot register rd%d\n",
2166 mdname(mddev), rdev->raid_disk);
2169 newmirrors[d2++].rdev = rdev;
2171 kfree(conf->mirrors);
2172 conf->mirrors = newmirrors;
2173 kfree(conf->poolinfo);
2174 conf->poolinfo = newpoolinfo;
2176 spin_lock_irqsave(&conf->device_lock, flags);
2177 mddev->degraded += (raid_disks - conf->raid_disks);
2178 spin_unlock_irqrestore(&conf->device_lock, flags);
2179 conf->raid_disks = mddev->raid_disks = raid_disks;
2180 mddev->delta_disks = 0;
2182 conf->last_used = 0; /* just make sure it is in-range */
2183 lower_barrier(conf);
2185 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2186 md_wakeup_thread(mddev->thread);
2188 mempool_destroy(oldpool);
2192 static void raid1_quiesce(mddev_t *mddev, int state)
2194 conf_t *conf = mddev->private;
2197 case 2: /* wake for suspend */
2198 wake_up(&conf->wait_barrier);
2201 raise_barrier(conf);
2204 lower_barrier(conf);
2209 static void *raid1_takeover(mddev_t *mddev)
2211 /* raid1 can take over:
2212 * raid5 with 2 devices, any layout or chunk size
2214 if (mddev->level == 5 && mddev->raid_disks == 2) {
2216 mddev->new_level = 1;
2217 mddev->new_layout = 0;
2218 mddev->new_chunk_sectors = 0;
2219 conf = setup_conf(mddev);
2224 return ERR_PTR(-EINVAL);
2227 static struct mdk_personality raid1_personality =
2231 .owner = THIS_MODULE,
2232 .make_request = make_request,
2236 .error_handler = error,
2237 .hot_add_disk = raid1_add_disk,
2238 .hot_remove_disk= raid1_remove_disk,
2239 .spare_active = raid1_spare_active,
2240 .sync_request = sync_request,
2241 .resize = raid1_resize,
2243 .check_reshape = raid1_reshape,
2244 .quiesce = raid1_quiesce,
2245 .takeover = raid1_takeover,
2248 static int __init raid_init(void)
2250 return register_md_personality(&raid1_personality);
2253 static void raid_exit(void)
2255 unregister_md_personality(&raid1_personality);
2258 module_init(raid_init);
2259 module_exit(raid_exit);
2260 MODULE_LICENSE("GPL");
2261 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2262 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2263 MODULE_ALIAS("md-raid1");
2264 MODULE_ALIAS("md-level-1");