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 allow_barrier(conf_t *conf);
56 static void lower_barrier(conf_t *conf);
58 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
60 struct pool_info *pi = data;
61 int size = offsetof(r1bio_t, bios[pi->raid_disks]);
63 /* allocate a r1bio with room for raid_disks entries in the bios array */
64 return kzalloc(size, gfp_flags);
67 static void r1bio_pool_free(void *r1_bio, void *data)
72 #define RESYNC_BLOCK_SIZE (64*1024)
73 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
74 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
75 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
76 #define RESYNC_WINDOW (2048*1024)
78 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
80 struct pool_info *pi = data;
86 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
91 * Allocate bios : 1 for reading, n-1 for writing
93 for (j = pi->raid_disks ; j-- ; ) {
94 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
97 r1_bio->bios[j] = bio;
100 * Allocate RESYNC_PAGES data pages and attach them to
102 * If this is a user-requested check/repair, allocate
103 * RESYNC_PAGES for each bio.
105 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
110 bio = r1_bio->bios[j];
111 for (i = 0; i < RESYNC_PAGES; i++) {
112 page = alloc_page(gfp_flags);
116 bio->bi_io_vec[i].bv_page = page;
120 /* If not user-requests, copy the page pointers to all bios */
121 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
122 for (i=0; i<RESYNC_PAGES ; i++)
123 for (j=1; j<pi->raid_disks; j++)
124 r1_bio->bios[j]->bi_io_vec[i].bv_page =
125 r1_bio->bios[0]->bi_io_vec[i].bv_page;
128 r1_bio->master_bio = NULL;
133 for (j=0 ; j < pi->raid_disks; j++)
134 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
135 put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
138 while ( ++j < pi->raid_disks )
139 bio_put(r1_bio->bios[j]);
140 r1bio_pool_free(r1_bio, data);
144 static void r1buf_pool_free(void *__r1_bio, void *data)
146 struct pool_info *pi = data;
148 r1bio_t *r1bio = __r1_bio;
150 for (i = 0; i < RESYNC_PAGES; i++)
151 for (j = pi->raid_disks; j-- ;) {
153 r1bio->bios[j]->bi_io_vec[i].bv_page !=
154 r1bio->bios[0]->bi_io_vec[i].bv_page)
155 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
157 for (i=0 ; i < pi->raid_disks; i++)
158 bio_put(r1bio->bios[i]);
160 r1bio_pool_free(r1bio, data);
163 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
167 for (i = 0; i < conf->raid_disks; i++) {
168 struct bio **bio = r1_bio->bios + i;
169 if (*bio && *bio != IO_BLOCKED)
175 static void free_r1bio(r1bio_t *r1_bio)
177 conf_t *conf = r1_bio->mddev->private;
180 * Wake up any possible resync thread that waits for the device
185 put_all_bios(conf, r1_bio);
186 mempool_free(r1_bio, conf->r1bio_pool);
189 static void put_buf(r1bio_t *r1_bio)
191 conf_t *conf = r1_bio->mddev->private;
194 for (i=0; i<conf->raid_disks; i++) {
195 struct bio *bio = r1_bio->bios[i];
197 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
200 mempool_free(r1_bio, conf->r1buf_pool);
205 static void reschedule_retry(r1bio_t *r1_bio)
208 mddev_t *mddev = r1_bio->mddev;
209 conf_t *conf = mddev->private;
211 spin_lock_irqsave(&conf->device_lock, flags);
212 list_add(&r1_bio->retry_list, &conf->retry_list);
214 spin_unlock_irqrestore(&conf->device_lock, flags);
216 wake_up(&conf->wait_barrier);
217 md_wakeup_thread(mddev->thread);
221 * raid_end_bio_io() is called when we have finished servicing a mirrored
222 * operation and are ready to return a success/failure code to the buffer
225 static void raid_end_bio_io(r1bio_t *r1_bio)
227 struct bio *bio = r1_bio->master_bio;
229 /* if nobody has done the final endio yet, do it now */
230 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
231 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
232 (bio_data_dir(bio) == WRITE) ? "write" : "read",
233 (unsigned long long) bio->bi_sector,
234 (unsigned long long) bio->bi_sector +
235 (bio->bi_size >> 9) - 1);
238 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
244 * Update disk head position estimator based on IRQ completion info.
246 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
248 conf_t *conf = r1_bio->mddev->private;
250 conf->mirrors[disk].head_position =
251 r1_bio->sector + (r1_bio->sectors);
254 static void raid1_end_read_request(struct bio *bio, int error)
256 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
257 r1bio_t *r1_bio = bio->bi_private;
259 conf_t *conf = r1_bio->mddev->private;
261 mirror = r1_bio->read_disk;
263 * this branch is our 'one mirror IO has finished' event handler:
265 update_head_pos(mirror, r1_bio);
268 set_bit(R1BIO_Uptodate, &r1_bio->state);
270 /* If all other devices have failed, we want to return
271 * the error upwards rather than fail the last device.
272 * Here we redefine "uptodate" to mean "Don't want to retry"
275 spin_lock_irqsave(&conf->device_lock, flags);
276 if (r1_bio->mddev->degraded == conf->raid_disks ||
277 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
278 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
280 spin_unlock_irqrestore(&conf->device_lock, flags);
284 raid_end_bio_io(r1_bio);
289 char b[BDEVNAME_SIZE];
290 if (printk_ratelimit())
291 printk(KERN_ERR "md/raid1:%s: %s: rescheduling sector %llu\n",
293 bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
294 reschedule_retry(r1_bio);
297 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
300 static void r1_bio_write_done(r1bio_t *r1_bio)
302 if (atomic_dec_and_test(&r1_bio->remaining))
304 /* it really is the end of this request */
305 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
306 /* free extra copy of the data pages */
307 int i = r1_bio->behind_page_count;
309 safe_put_page(r1_bio->behind_pages[i]);
310 kfree(r1_bio->behind_pages);
311 r1_bio->behind_pages = NULL;
313 /* clear the bitmap if all writes complete successfully */
314 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
316 !test_bit(R1BIO_Degraded, &r1_bio->state),
317 test_bit(R1BIO_BehindIO, &r1_bio->state));
318 md_write_end(r1_bio->mddev);
319 raid_end_bio_io(r1_bio);
323 static void raid1_end_write_request(struct bio *bio, int error)
325 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
326 r1bio_t *r1_bio = bio->bi_private;
327 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
328 conf_t *conf = r1_bio->mddev->private;
329 struct bio *to_put = NULL;
332 for (mirror = 0; mirror < conf->raid_disks; mirror++)
333 if (r1_bio->bios[mirror] == bio)
337 * 'one mirror IO has finished' event handler:
339 r1_bio->bios[mirror] = NULL;
342 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
343 /* an I/O failed, we can't clear the bitmap */
344 set_bit(R1BIO_Degraded, &r1_bio->state);
347 * Set R1BIO_Uptodate in our master bio, so that we
348 * will return a good error code for to the higher
349 * levels even if IO on some other mirrored buffer
352 * The 'master' represents the composite IO operation
353 * to user-side. So if something waits for IO, then it
354 * will wait for the 'master' bio.
356 set_bit(R1BIO_Uptodate, &r1_bio->state);
358 update_head_pos(mirror, r1_bio);
361 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
362 atomic_dec(&r1_bio->behind_remaining);
365 * In behind mode, we ACK the master bio once the I/O
366 * has safely reached all non-writemostly
367 * disks. Setting the Returned bit ensures that this
368 * gets done only once -- we don't ever want to return
369 * -EIO here, instead we'll wait
371 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
372 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
373 /* Maybe we can return now */
374 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
375 struct bio *mbio = r1_bio->master_bio;
376 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
377 (unsigned long long) mbio->bi_sector,
378 (unsigned long long) mbio->bi_sector +
379 (mbio->bi_size >> 9) - 1);
384 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
387 * Let's see if all mirrored write operations have finished
390 r1_bio_write_done(r1_bio);
398 * This routine returns the disk from which the requested read should
399 * be done. There is a per-array 'next expected sequential IO' sector
400 * number - if this matches on the next IO then we use the last disk.
401 * There is also a per-disk 'last know head position' sector that is
402 * maintained from IRQ contexts, both the normal and the resync IO
403 * completion handlers update this position correctly. If there is no
404 * perfect sequential match then we pick the disk whose head is closest.
406 * If there are 2 mirrors in the same 2 devices, performance degrades
407 * because position is mirror, not device based.
409 * The rdev for the device selected will have nr_pending incremented.
411 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
413 const sector_t this_sector = r1_bio->sector;
414 const int sectors = r1_bio->sectors;
424 * Check if we can balance. We can balance on the whole
425 * device if no resync is going on, or below the resync window.
426 * We take the first readable disk when above the resync window.
430 best_dist = MaxSector;
431 if (conf->mddev->recovery_cp < MaxSector &&
432 (this_sector + sectors >= conf->next_resync)) {
437 start_disk = conf->last_used;
440 for (i = 0 ; i < conf->raid_disks ; i++) {
442 int disk = start_disk + i;
443 if (disk >= conf->raid_disks)
444 disk -= conf->raid_disks;
446 rdev = rcu_dereference(conf->mirrors[disk].rdev);
447 if (r1_bio->bios[disk] == IO_BLOCKED
449 || test_bit(Faulty, &rdev->flags))
451 if (!test_bit(In_sync, &rdev->flags) &&
452 rdev->recovery_offset < this_sector + sectors)
454 if (test_bit(WriteMostly, &rdev->flags)) {
455 /* Don't balance among write-mostly, just
456 * use the first as a last resort */
461 /* This is a reasonable device to use. It might
464 dist = abs(this_sector - conf->mirrors[disk].head_position);
466 /* Don't change to another disk for sequential reads */
467 || conf->next_seq_sect == this_sector
469 /* If device is idle, use it */
470 || atomic_read(&rdev->nr_pending) == 0) {
474 if (dist < best_dist) {
480 if (best_disk >= 0) {
481 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
484 atomic_inc(&rdev->nr_pending);
485 if (test_bit(Faulty, &rdev->flags)) {
486 /* cannot risk returning a device that failed
487 * before we inc'ed nr_pending
489 rdev_dec_pending(rdev, conf->mddev);
492 conf->next_seq_sect = this_sector + sectors;
493 conf->last_used = best_disk;
500 int md_raid1_congested(mddev_t *mddev, int bits)
502 conf_t *conf = mddev->private;
506 for (i = 0; i < mddev->raid_disks; i++) {
507 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
508 if (rdev && !test_bit(Faulty, &rdev->flags)) {
509 struct request_queue *q = bdev_get_queue(rdev->bdev);
513 /* Note the '|| 1' - when read_balance prefers
514 * non-congested targets, it can be removed
516 if ((bits & (1<<BDI_async_congested)) || 1)
517 ret |= bdi_congested(&q->backing_dev_info, bits);
519 ret &= bdi_congested(&q->backing_dev_info, bits);
525 EXPORT_SYMBOL_GPL(md_raid1_congested);
527 static int raid1_congested(void *data, int bits)
529 mddev_t *mddev = data;
531 return mddev_congested(mddev, bits) ||
532 md_raid1_congested(mddev, bits);
535 static void flush_pending_writes(conf_t *conf)
537 /* Any writes that have been queued but are awaiting
538 * bitmap updates get flushed here.
540 spin_lock_irq(&conf->device_lock);
542 if (conf->pending_bio_list.head) {
544 bio = bio_list_get(&conf->pending_bio_list);
545 spin_unlock_irq(&conf->device_lock);
546 /* flush any pending bitmap writes to
547 * disk before proceeding w/ I/O */
548 bitmap_unplug(conf->mddev->bitmap);
550 while (bio) { /* submit pending writes */
551 struct bio *next = bio->bi_next;
553 generic_make_request(bio);
557 spin_unlock_irq(&conf->device_lock);
561 * Sometimes we need to suspend IO while we do something else,
562 * either some resync/recovery, or reconfigure the array.
563 * To do this we raise a 'barrier'.
564 * The 'barrier' is a counter that can be raised multiple times
565 * to count how many activities are happening which preclude
567 * We can only raise the barrier if there is no pending IO.
568 * i.e. if nr_pending == 0.
569 * We choose only to raise the barrier if no-one is waiting for the
570 * barrier to go down. This means that as soon as an IO request
571 * is ready, no other operations which require a barrier will start
572 * until the IO request has had a chance.
574 * So: regular IO calls 'wait_barrier'. When that returns there
575 * is no backgroup IO happening, It must arrange to call
576 * allow_barrier when it has finished its IO.
577 * backgroup IO calls must call raise_barrier. Once that returns
578 * there is no normal IO happeing. It must arrange to call
579 * lower_barrier when the particular background IO completes.
581 #define RESYNC_DEPTH 32
583 static void raise_barrier(conf_t *conf)
585 spin_lock_irq(&conf->resync_lock);
587 /* Wait until no block IO is waiting */
588 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
589 conf->resync_lock, );
591 /* block any new IO from starting */
594 /* Now wait for all pending IO to complete */
595 wait_event_lock_irq(conf->wait_barrier,
596 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
597 conf->resync_lock, );
599 spin_unlock_irq(&conf->resync_lock);
602 static void lower_barrier(conf_t *conf)
605 BUG_ON(conf->barrier <= 0);
606 spin_lock_irqsave(&conf->resync_lock, flags);
608 spin_unlock_irqrestore(&conf->resync_lock, flags);
609 wake_up(&conf->wait_barrier);
612 static void wait_barrier(conf_t *conf)
614 spin_lock_irq(&conf->resync_lock);
617 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
623 spin_unlock_irq(&conf->resync_lock);
626 static void allow_barrier(conf_t *conf)
629 spin_lock_irqsave(&conf->resync_lock, flags);
631 spin_unlock_irqrestore(&conf->resync_lock, flags);
632 wake_up(&conf->wait_barrier);
635 static void freeze_array(conf_t *conf)
637 /* stop syncio and normal IO and wait for everything to
639 * We increment barrier and nr_waiting, and then
640 * wait until nr_pending match nr_queued+1
641 * This is called in the context of one normal IO request
642 * that has failed. Thus any sync request that might be pending
643 * will be blocked by nr_pending, and we need to wait for
644 * pending IO requests to complete or be queued for re-try.
645 * Thus the number queued (nr_queued) plus this request (1)
646 * must match the number of pending IOs (nr_pending) before
649 spin_lock_irq(&conf->resync_lock);
652 wait_event_lock_irq(conf->wait_barrier,
653 conf->nr_pending == conf->nr_queued+1,
655 flush_pending_writes(conf));
656 spin_unlock_irq(&conf->resync_lock);
658 static void unfreeze_array(conf_t *conf)
660 /* reverse the effect of the freeze */
661 spin_lock_irq(&conf->resync_lock);
664 wake_up(&conf->wait_barrier);
665 spin_unlock_irq(&conf->resync_lock);
669 /* duplicate the data pages for behind I/O
671 static void alloc_behind_pages(struct bio *bio, r1bio_t *r1_bio)
674 struct bio_vec *bvec;
675 struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page*),
677 if (unlikely(!pages))
680 bio_for_each_segment(bvec, bio, i) {
681 pages[i] = alloc_page(GFP_NOIO);
682 if (unlikely(!pages[i]))
684 memcpy(kmap(pages[i]) + bvec->bv_offset,
685 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
687 kunmap(bvec->bv_page);
689 r1_bio->behind_pages = pages;
690 r1_bio->behind_page_count = bio->bi_vcnt;
691 set_bit(R1BIO_BehindIO, &r1_bio->state);
695 for (i = 0; i < bio->bi_vcnt; i++)
699 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
702 static int make_request(mddev_t *mddev, struct bio * bio)
704 conf_t *conf = mddev->private;
705 mirror_info_t *mirror;
707 struct bio *read_bio;
708 int i, targets = 0, disks;
709 struct bitmap *bitmap;
711 const int rw = bio_data_dir(bio);
712 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
713 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
714 mdk_rdev_t *blocked_rdev;
718 * Register the new request and wait if the reconstruction
719 * thread has put up a bar for new requests.
720 * Continue immediately if no resync is active currently.
723 md_write_start(mddev, bio); /* wait on superblock update early */
725 if (bio_data_dir(bio) == WRITE &&
726 bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
727 bio->bi_sector < mddev->suspend_hi) {
728 /* As the suspend_* range is controlled by
729 * userspace, we want an interruptible
734 flush_signals(current);
735 prepare_to_wait(&conf->wait_barrier,
736 &w, TASK_INTERRUPTIBLE);
737 if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
738 bio->bi_sector >= mddev->suspend_hi)
742 finish_wait(&conf->wait_barrier, &w);
747 bitmap = mddev->bitmap;
750 * make_request() can abort the operation when READA is being
751 * used and no empty request is available.
754 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
756 r1_bio->master_bio = bio;
757 r1_bio->sectors = bio->bi_size >> 9;
759 r1_bio->mddev = mddev;
760 r1_bio->sector = bio->bi_sector;
764 * read balancing logic:
766 int rdisk = read_balance(conf, r1_bio);
769 /* couldn't find anywhere to read from */
770 raid_end_bio_io(r1_bio);
773 mirror = conf->mirrors + rdisk;
775 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
777 /* Reading from a write-mostly device must
778 * take care not to over-take any writes
781 wait_event(bitmap->behind_wait,
782 atomic_read(&bitmap->behind_writes) == 0);
784 r1_bio->read_disk = rdisk;
786 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
788 r1_bio->bios[rdisk] = read_bio;
790 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
791 read_bio->bi_bdev = mirror->rdev->bdev;
792 read_bio->bi_end_io = raid1_end_read_request;
793 read_bio->bi_rw = READ | do_sync;
794 read_bio->bi_private = r1_bio;
796 generic_make_request(read_bio);
803 /* first select target devices under spinlock and
804 * inc refcount on their rdev. Record them by setting
807 plugged = mddev_check_plugged(mddev);
809 disks = conf->raid_disks;
813 for (i = 0; i < disks; i++) {
814 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
815 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
816 atomic_inc(&rdev->nr_pending);
820 if (rdev && !test_bit(Faulty, &rdev->flags)) {
821 atomic_inc(&rdev->nr_pending);
822 if (test_bit(Faulty, &rdev->flags)) {
823 rdev_dec_pending(rdev, mddev);
824 r1_bio->bios[i] = NULL;
826 r1_bio->bios[i] = bio;
830 r1_bio->bios[i] = NULL;
834 if (unlikely(blocked_rdev)) {
835 /* Wait for this device to become unblocked */
838 for (j = 0; j < i; j++)
840 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
843 md_wait_for_blocked_rdev(blocked_rdev, mddev);
848 BUG_ON(targets == 0); /* we never fail the last device */
850 if (targets < conf->raid_disks) {
851 /* array is degraded, we will not clear the bitmap
852 * on I/O completion (see raid1_end_write_request) */
853 set_bit(R1BIO_Degraded, &r1_bio->state);
857 * Not if there are too many, or cannot allocate memory,
858 * or a reader on WriteMostly is waiting for behind writes
861 (atomic_read(&bitmap->behind_writes)
862 < mddev->bitmap_info.max_write_behind) &&
863 !waitqueue_active(&bitmap->behind_wait))
864 alloc_behind_pages(bio, r1_bio);
866 atomic_set(&r1_bio->remaining, 1);
867 atomic_set(&r1_bio->behind_remaining, 0);
869 bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
870 test_bit(R1BIO_BehindIO, &r1_bio->state));
871 for (i = 0; i < disks; i++) {
873 if (!r1_bio->bios[i])
876 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
877 r1_bio->bios[i] = mbio;
879 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
880 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
881 mbio->bi_end_io = raid1_end_write_request;
882 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
883 mbio->bi_private = r1_bio;
885 if (r1_bio->behind_pages) {
886 struct bio_vec *bvec;
889 /* Yes, I really want the '__' version so that
890 * we clear any unused pointer in the io_vec, rather
891 * than leave them unchanged. This is important
892 * because when we come to free the pages, we won't
893 * know the original bi_idx, so we just free
896 __bio_for_each_segment(bvec, mbio, j, 0)
897 bvec->bv_page = r1_bio->behind_pages[j];
898 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
899 atomic_inc(&r1_bio->behind_remaining);
902 atomic_inc(&r1_bio->remaining);
903 spin_lock_irqsave(&conf->device_lock, flags);
904 bio_list_add(&conf->pending_bio_list, mbio);
905 spin_unlock_irqrestore(&conf->device_lock, flags);
907 r1_bio_write_done(r1_bio);
909 /* In case raid1d snuck in to freeze_array */
910 wake_up(&conf->wait_barrier);
912 if (do_sync || !bitmap || !plugged)
913 md_wakeup_thread(mddev->thread);
918 static void status(struct seq_file *seq, mddev_t *mddev)
920 conf_t *conf = mddev->private;
923 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
924 conf->raid_disks - mddev->degraded);
926 for (i = 0; i < conf->raid_disks; i++) {
927 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
928 seq_printf(seq, "%s",
929 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
932 seq_printf(seq, "]");
936 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
938 char b[BDEVNAME_SIZE];
939 conf_t *conf = mddev->private;
942 * If it is not operational, then we have already marked it as dead
943 * else if it is the last working disks, ignore the error, let the
944 * next level up know.
945 * else mark the drive as failed
947 if (test_bit(In_sync, &rdev->flags)
948 && (conf->raid_disks - mddev->degraded) == 1) {
950 * Don't fail the drive, act as though we were just a
951 * normal single drive.
952 * However don't try a recovery from this drive as
953 * it is very likely to fail.
955 mddev->recovery_disabled = 1;
958 if (test_and_clear_bit(In_sync, &rdev->flags)) {
960 spin_lock_irqsave(&conf->device_lock, flags);
962 set_bit(Faulty, &rdev->flags);
963 spin_unlock_irqrestore(&conf->device_lock, flags);
965 * if recovery is running, make sure it aborts.
967 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
969 set_bit(Faulty, &rdev->flags);
970 set_bit(MD_CHANGE_DEVS, &mddev->flags);
972 "md/raid1:%s: Disk failure on %s, disabling device.\n"
973 "md/raid1:%s: Operation continuing on %d devices.\n",
974 mdname(mddev), bdevname(rdev->bdev, b),
975 mdname(mddev), conf->raid_disks - mddev->degraded);
978 static void print_conf(conf_t *conf)
982 printk(KERN_DEBUG "RAID1 conf printout:\n");
984 printk(KERN_DEBUG "(!conf)\n");
987 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
991 for (i = 0; i < conf->raid_disks; i++) {
992 char b[BDEVNAME_SIZE];
993 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
995 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
996 i, !test_bit(In_sync, &rdev->flags),
997 !test_bit(Faulty, &rdev->flags),
998 bdevname(rdev->bdev,b));
1003 static void close_sync(conf_t *conf)
1006 allow_barrier(conf);
1008 mempool_destroy(conf->r1buf_pool);
1009 conf->r1buf_pool = NULL;
1012 static int raid1_spare_active(mddev_t *mddev)
1015 conf_t *conf = mddev->private;
1017 unsigned long flags;
1020 * Find all failed disks within the RAID1 configuration
1021 * and mark them readable.
1022 * Called under mddev lock, so rcu protection not needed.
1024 for (i = 0; i < conf->raid_disks; i++) {
1025 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1027 && !test_bit(Faulty, &rdev->flags)
1028 && !test_and_set_bit(In_sync, &rdev->flags)) {
1030 sysfs_notify_dirent(rdev->sysfs_state);
1033 spin_lock_irqsave(&conf->device_lock, flags);
1034 mddev->degraded -= count;
1035 spin_unlock_irqrestore(&conf->device_lock, flags);
1042 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1044 conf_t *conf = mddev->private;
1049 int last = mddev->raid_disks - 1;
1051 if (rdev->raid_disk >= 0)
1052 first = last = rdev->raid_disk;
1054 for (mirror = first; mirror <= last; mirror++)
1055 if ( !(p=conf->mirrors+mirror)->rdev) {
1057 disk_stack_limits(mddev->gendisk, rdev->bdev,
1058 rdev->data_offset << 9);
1059 /* as we don't honour merge_bvec_fn, we must
1060 * never risk violating it, so limit
1061 * ->max_segments to one lying with a single
1062 * page, as a one page request is never in
1065 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1066 blk_queue_max_segments(mddev->queue, 1);
1067 blk_queue_segment_boundary(mddev->queue,
1068 PAGE_CACHE_SIZE - 1);
1071 p->head_position = 0;
1072 rdev->raid_disk = mirror;
1074 /* As all devices are equivalent, we don't need a full recovery
1075 * if this was recently any drive of the array
1077 if (rdev->saved_raid_disk < 0)
1079 rcu_assign_pointer(p->rdev, rdev);
1082 md_integrity_add_rdev(rdev, mddev);
1087 static int raid1_remove_disk(mddev_t *mddev, int number)
1089 conf_t *conf = mddev->private;
1092 mirror_info_t *p = conf->mirrors+ number;
1097 if (test_bit(In_sync, &rdev->flags) ||
1098 atomic_read(&rdev->nr_pending)) {
1102 /* Only remove non-faulty devices if recovery
1105 if (!test_bit(Faulty, &rdev->flags) &&
1106 !mddev->recovery_disabled &&
1107 mddev->degraded < conf->raid_disks) {
1113 if (atomic_read(&rdev->nr_pending)) {
1114 /* lost the race, try later */
1119 err = md_integrity_register(mddev);
1128 static void end_sync_read(struct bio *bio, int error)
1130 r1bio_t *r1_bio = bio->bi_private;
1133 for (i=r1_bio->mddev->raid_disks; i--; )
1134 if (r1_bio->bios[i] == bio)
1137 update_head_pos(i, r1_bio);
1139 * we have read a block, now it needs to be re-written,
1140 * or re-read if the read failed.
1141 * We don't do much here, just schedule handling by raid1d
1143 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1144 set_bit(R1BIO_Uptodate, &r1_bio->state);
1146 if (atomic_dec_and_test(&r1_bio->remaining))
1147 reschedule_retry(r1_bio);
1150 static void end_sync_write(struct bio *bio, int error)
1152 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1153 r1bio_t *r1_bio = bio->bi_private;
1154 mddev_t *mddev = r1_bio->mddev;
1155 conf_t *conf = mddev->private;
1159 for (i = 0; i < conf->raid_disks; i++)
1160 if (r1_bio->bios[i] == bio) {
1165 sector_t sync_blocks = 0;
1166 sector_t s = r1_bio->sector;
1167 long sectors_to_go = r1_bio->sectors;
1168 /* make sure these bits doesn't get cleared. */
1170 bitmap_end_sync(mddev->bitmap, s,
1173 sectors_to_go -= sync_blocks;
1174 } while (sectors_to_go > 0);
1175 md_error(mddev, conf->mirrors[mirror].rdev);
1178 update_head_pos(mirror, r1_bio);
1180 if (atomic_dec_and_test(&r1_bio->remaining)) {
1181 sector_t s = r1_bio->sectors;
1183 md_done_sync(mddev, s, uptodate);
1187 static int fix_sync_read_error(r1bio_t *r1_bio)
1189 /* Try some synchronous reads of other devices to get
1190 * good data, much like with normal read errors. Only
1191 * read into the pages we already have so we don't
1192 * need to re-issue the read request.
1193 * We don't need to freeze the array, because being in an
1194 * active sync request, there is no normal IO, and
1195 * no overlapping syncs.
1197 mddev_t *mddev = r1_bio->mddev;
1198 conf_t *conf = mddev->private;
1199 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1200 sector_t sect = r1_bio->sector;
1201 int sectors = r1_bio->sectors;
1206 int d = r1_bio->read_disk;
1211 if (s > (PAGE_SIZE>>9))
1214 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1215 /* No rcu protection needed here devices
1216 * can only be removed when no resync is
1217 * active, and resync is currently active
1219 rdev = conf->mirrors[d].rdev;
1220 if (sync_page_io(rdev,
1223 bio->bi_io_vec[idx].bv_page,
1230 if (d == conf->raid_disks)
1232 } while (!success && d != r1_bio->read_disk);
1235 char b[BDEVNAME_SIZE];
1236 /* Cannot read from anywhere, array is toast */
1237 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1238 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1239 " for block %llu\n",
1241 bdevname(bio->bi_bdev, b),
1242 (unsigned long long)r1_bio->sector);
1243 md_done_sync(mddev, r1_bio->sectors, 0);
1249 /* write it back and re-read */
1250 while (d != r1_bio->read_disk) {
1252 d = conf->raid_disks;
1254 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1256 rdev = conf->mirrors[d].rdev;
1257 if (sync_page_io(rdev,
1260 bio->bi_io_vec[idx].bv_page,
1261 WRITE, false) == 0) {
1262 r1_bio->bios[d]->bi_end_io = NULL;
1263 rdev_dec_pending(rdev, mddev);
1264 md_error(mddev, rdev);
1266 atomic_add(s, &rdev->corrected_errors);
1269 while (d != r1_bio->read_disk) {
1271 d = conf->raid_disks;
1273 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1275 rdev = conf->mirrors[d].rdev;
1276 if (sync_page_io(rdev,
1279 bio->bi_io_vec[idx].bv_page,
1281 md_error(mddev, rdev);
1287 set_bit(R1BIO_Uptodate, &r1_bio->state);
1288 set_bit(BIO_UPTODATE, &bio->bi_flags);
1292 static int process_checks(r1bio_t *r1_bio)
1294 /* We have read all readable devices. If we haven't
1295 * got the block, then there is no hope left.
1296 * If we have, then we want to do a comparison
1297 * and skip the write if everything is the same.
1298 * If any blocks failed to read, then we need to
1299 * attempt an over-write
1301 mddev_t *mddev = r1_bio->mddev;
1302 conf_t *conf = mddev->private;
1306 for (primary = 0; primary < conf->raid_disks; primary++)
1307 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1308 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1309 r1_bio->bios[primary]->bi_end_io = NULL;
1310 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1313 r1_bio->read_disk = primary;
1314 for (i = 0; i < conf->raid_disks; i++) {
1316 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1317 struct bio *pbio = r1_bio->bios[primary];
1318 struct bio *sbio = r1_bio->bios[i];
1321 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1324 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1325 for (j = vcnt; j-- ; ) {
1327 p = pbio->bi_io_vec[j].bv_page;
1328 s = sbio->bi_io_vec[j].bv_page;
1329 if (memcmp(page_address(p),
1337 mddev->resync_mismatches += r1_bio->sectors;
1338 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1339 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1340 /* No need to write to this device. */
1341 sbio->bi_end_io = NULL;
1342 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1345 /* fixup the bio for reuse */
1346 sbio->bi_vcnt = vcnt;
1347 sbio->bi_size = r1_bio->sectors << 9;
1349 sbio->bi_phys_segments = 0;
1350 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1351 sbio->bi_flags |= 1 << BIO_UPTODATE;
1352 sbio->bi_next = NULL;
1353 sbio->bi_sector = r1_bio->sector +
1354 conf->mirrors[i].rdev->data_offset;
1355 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1356 size = sbio->bi_size;
1357 for (j = 0; j < vcnt ; j++) {
1359 bi = &sbio->bi_io_vec[j];
1361 if (size > PAGE_SIZE)
1362 bi->bv_len = PAGE_SIZE;
1366 memcpy(page_address(bi->bv_page),
1367 page_address(pbio->bi_io_vec[j].bv_page),
1374 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1376 conf_t *conf = mddev->private;
1378 int disks = conf->raid_disks;
1379 struct bio *bio, *wbio;
1381 bio = r1_bio->bios[r1_bio->read_disk];
1383 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1384 /* ouch - failed to read all of that. */
1385 if (!fix_sync_read_error(r1_bio))
1388 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1389 if (process_checks(r1_bio) < 0)
1394 atomic_set(&r1_bio->remaining, 1);
1395 for (i = 0; i < disks ; i++) {
1396 wbio = r1_bio->bios[i];
1397 if (wbio->bi_end_io == NULL ||
1398 (wbio->bi_end_io == end_sync_read &&
1399 (i == r1_bio->read_disk ||
1400 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1403 wbio->bi_rw = WRITE;
1404 wbio->bi_end_io = end_sync_write;
1405 atomic_inc(&r1_bio->remaining);
1406 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1408 generic_make_request(wbio);
1411 if (atomic_dec_and_test(&r1_bio->remaining)) {
1412 /* if we're here, all write(s) have completed, so clean up */
1413 md_done_sync(mddev, r1_bio->sectors, 1);
1419 * This is a kernel thread which:
1421 * 1. Retries failed read operations on working mirrors.
1422 * 2. Updates the raid superblock when problems encounter.
1423 * 3. Performs writes following reads for array syncronising.
1426 static void fix_read_error(conf_t *conf, int read_disk,
1427 sector_t sect, int sectors)
1429 mddev_t *mddev = conf->mddev;
1437 if (s > (PAGE_SIZE>>9))
1441 /* Note: no rcu protection needed here
1442 * as this is synchronous in the raid1d thread
1443 * which is the thread that might remove
1444 * a device. If raid1d ever becomes multi-threaded....
1446 rdev = conf->mirrors[d].rdev;
1448 test_bit(In_sync, &rdev->flags) &&
1449 sync_page_io(rdev, sect, s<<9,
1450 conf->tmppage, READ, false))
1454 if (d == conf->raid_disks)
1457 } while (!success && d != read_disk);
1460 /* Cannot read from anywhere -- bye bye array */
1461 md_error(mddev, conf->mirrors[read_disk].rdev);
1464 /* write it back and re-read */
1466 while (d != read_disk) {
1468 d = conf->raid_disks;
1470 rdev = conf->mirrors[d].rdev;
1472 test_bit(In_sync, &rdev->flags)) {
1473 if (sync_page_io(rdev, sect, s<<9,
1474 conf->tmppage, WRITE, false)
1476 /* Well, this device is dead */
1477 md_error(mddev, rdev);
1481 while (d != read_disk) {
1482 char b[BDEVNAME_SIZE];
1484 d = conf->raid_disks;
1486 rdev = conf->mirrors[d].rdev;
1488 test_bit(In_sync, &rdev->flags)) {
1489 if (sync_page_io(rdev, sect, s<<9,
1490 conf->tmppage, READ, false)
1492 /* Well, this device is dead */
1493 md_error(mddev, rdev);
1495 atomic_add(s, &rdev->corrected_errors);
1497 "md/raid1:%s: read error corrected "
1498 "(%d sectors at %llu on %s)\n",
1500 (unsigned long long)(sect +
1502 bdevname(rdev->bdev, b));
1511 static void raid1d(mddev_t *mddev)
1515 unsigned long flags;
1516 conf_t *conf = mddev->private;
1517 struct list_head *head = &conf->retry_list;
1519 struct blk_plug plug;
1521 md_check_recovery(mddev);
1523 blk_start_plug(&plug);
1525 char b[BDEVNAME_SIZE];
1527 if (atomic_read(&mddev->plug_cnt) == 0)
1528 flush_pending_writes(conf);
1530 spin_lock_irqsave(&conf->device_lock, flags);
1531 if (list_empty(head)) {
1532 spin_unlock_irqrestore(&conf->device_lock, flags);
1535 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1536 list_del(head->prev);
1538 spin_unlock_irqrestore(&conf->device_lock, flags);
1540 mddev = r1_bio->mddev;
1541 conf = mddev->private;
1542 if (test_bit(R1BIO_IsSync, &r1_bio->state))
1543 sync_request_write(mddev, r1_bio);
1547 /* we got a read error. Maybe the drive is bad. Maybe just
1548 * the block and we can fix it.
1549 * We freeze all other IO, and try reading the block from
1550 * other devices. When we find one, we re-write
1551 * and check it that fixes the read error.
1552 * This is all done synchronously while the array is
1555 if (mddev->ro == 0) {
1557 fix_read_error(conf, r1_bio->read_disk,
1560 unfreeze_array(conf);
1563 conf->mirrors[r1_bio->read_disk].rdev);
1565 bio = r1_bio->bios[r1_bio->read_disk];
1566 if ((disk=read_balance(conf, r1_bio)) == -1) {
1567 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1568 " read error for block %llu\n",
1570 bdevname(bio->bi_bdev,b),
1571 (unsigned long long)r1_bio->sector);
1572 raid_end_bio_io(r1_bio);
1574 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1575 r1_bio->bios[r1_bio->read_disk] =
1576 mddev->ro ? IO_BLOCKED : NULL;
1577 r1_bio->read_disk = disk;
1579 bio = bio_clone_mddev(r1_bio->master_bio,
1581 r1_bio->bios[r1_bio->read_disk] = bio;
1582 rdev = conf->mirrors[disk].rdev;
1583 if (printk_ratelimit())
1584 printk(KERN_ERR "md/raid1:%s: redirecting sector %llu to"
1585 " other mirror: %s\n",
1587 (unsigned long long)r1_bio->sector,
1588 bdevname(rdev->bdev,b));
1589 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1590 bio->bi_bdev = rdev->bdev;
1591 bio->bi_end_io = raid1_end_read_request;
1592 bio->bi_rw = READ | do_sync;
1593 bio->bi_private = r1_bio;
1594 generic_make_request(bio);
1599 blk_finish_plug(&plug);
1603 static int init_resync(conf_t *conf)
1607 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1608 BUG_ON(conf->r1buf_pool);
1609 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1611 if (!conf->r1buf_pool)
1613 conf->next_resync = 0;
1618 * perform a "sync" on one "block"
1620 * We need to make sure that no normal I/O request - particularly write
1621 * requests - conflict with active sync requests.
1623 * This is achieved by tracking pending requests and a 'barrier' concept
1624 * that can be installed to exclude normal IO requests.
1627 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1629 conf_t *conf = mddev->private;
1632 sector_t max_sector, nr_sectors;
1636 int write_targets = 0, read_targets = 0;
1637 sector_t sync_blocks;
1638 int still_degraded = 0;
1640 if (!conf->r1buf_pool)
1641 if (init_resync(conf))
1644 max_sector = mddev->dev_sectors;
1645 if (sector_nr >= max_sector) {
1646 /* If we aborted, we need to abort the
1647 * sync on the 'current' bitmap chunk (there will
1648 * only be one in raid1 resync.
1649 * We can find the current addess in mddev->curr_resync
1651 if (mddev->curr_resync < max_sector) /* aborted */
1652 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1654 else /* completed sync */
1657 bitmap_close_sync(mddev->bitmap);
1662 if (mddev->bitmap == NULL &&
1663 mddev->recovery_cp == MaxSector &&
1664 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1665 conf->fullsync == 0) {
1667 return max_sector - sector_nr;
1669 /* before building a request, check if we can skip these blocks..
1670 * This call the bitmap_start_sync doesn't actually record anything
1672 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1673 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1674 /* We can skip this block, and probably several more */
1679 * If there is non-resync activity waiting for a turn,
1680 * and resync is going fast enough,
1681 * then let it though before starting on this new sync request.
1683 if (!go_faster && conf->nr_waiting)
1684 msleep_interruptible(1000);
1686 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1687 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1688 raise_barrier(conf);
1690 conf->next_resync = sector_nr;
1694 * If we get a correctably read error during resync or recovery,
1695 * we might want to read from a different device. So we
1696 * flag all drives that could conceivably be read from for READ,
1697 * and any others (which will be non-In_sync devices) for WRITE.
1698 * If a read fails, we try reading from something else for which READ
1702 r1_bio->mddev = mddev;
1703 r1_bio->sector = sector_nr;
1705 set_bit(R1BIO_IsSync, &r1_bio->state);
1707 for (i=0; i < conf->raid_disks; i++) {
1709 bio = r1_bio->bios[i];
1711 /* take from bio_init */
1712 bio->bi_next = NULL;
1713 bio->bi_flags &= ~(BIO_POOL_MASK-1);
1714 bio->bi_flags |= 1 << BIO_UPTODATE;
1715 bio->bi_comp_cpu = -1;
1719 bio->bi_phys_segments = 0;
1721 bio->bi_end_io = NULL;
1722 bio->bi_private = NULL;
1724 rdev = rcu_dereference(conf->mirrors[i].rdev);
1726 test_bit(Faulty, &rdev->flags)) {
1729 } else if (!test_bit(In_sync, &rdev->flags)) {
1731 bio->bi_end_io = end_sync_write;
1734 /* may need to read from here */
1736 bio->bi_end_io = end_sync_read;
1737 if (test_bit(WriteMostly, &rdev->flags)) {
1746 atomic_inc(&rdev->nr_pending);
1747 bio->bi_sector = sector_nr + rdev->data_offset;
1748 bio->bi_bdev = rdev->bdev;
1749 bio->bi_private = r1_bio;
1754 r1_bio->read_disk = disk;
1756 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1757 /* extra read targets are also write targets */
1758 write_targets += read_targets-1;
1760 if (write_targets == 0 || read_targets == 0) {
1761 /* There is nowhere to write, so all non-sync
1762 * drives must be failed - so we are finished
1764 sector_t rv = max_sector - sector_nr;
1770 if (max_sector > mddev->resync_max)
1771 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1776 int len = PAGE_SIZE;
1777 if (sector_nr + (len>>9) > max_sector)
1778 len = (max_sector - sector_nr) << 9;
1781 if (sync_blocks == 0) {
1782 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1783 &sync_blocks, still_degraded) &&
1785 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1787 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1788 if ((len >> 9) > sync_blocks)
1789 len = sync_blocks<<9;
1792 for (i=0 ; i < conf->raid_disks; i++) {
1793 bio = r1_bio->bios[i];
1794 if (bio->bi_end_io) {
1795 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1796 if (bio_add_page(bio, page, len, 0) == 0) {
1798 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1801 bio = r1_bio->bios[i];
1802 if (bio->bi_end_io==NULL)
1804 /* remove last page from this bio */
1806 bio->bi_size -= len;
1807 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1813 nr_sectors += len>>9;
1814 sector_nr += len>>9;
1815 sync_blocks -= (len>>9);
1816 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1818 r1_bio->sectors = nr_sectors;
1820 /* For a user-requested sync, we read all readable devices and do a
1823 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1824 atomic_set(&r1_bio->remaining, read_targets);
1825 for (i=0; i<conf->raid_disks; i++) {
1826 bio = r1_bio->bios[i];
1827 if (bio->bi_end_io == end_sync_read) {
1828 md_sync_acct(bio->bi_bdev, nr_sectors);
1829 generic_make_request(bio);
1833 atomic_set(&r1_bio->remaining, 1);
1834 bio = r1_bio->bios[r1_bio->read_disk];
1835 md_sync_acct(bio->bi_bdev, nr_sectors);
1836 generic_make_request(bio);
1842 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1847 return mddev->dev_sectors;
1850 static conf_t *setup_conf(mddev_t *mddev)
1854 mirror_info_t *disk;
1858 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1862 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1867 conf->tmppage = alloc_page(GFP_KERNEL);
1871 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1872 if (!conf->poolinfo)
1874 conf->poolinfo->raid_disks = mddev->raid_disks;
1875 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1878 if (!conf->r1bio_pool)
1881 conf->poolinfo->mddev = mddev;
1883 spin_lock_init(&conf->device_lock);
1884 list_for_each_entry(rdev, &mddev->disks, same_set) {
1885 int disk_idx = rdev->raid_disk;
1886 if (disk_idx >= mddev->raid_disks
1889 disk = conf->mirrors + disk_idx;
1893 disk->head_position = 0;
1895 conf->raid_disks = mddev->raid_disks;
1896 conf->mddev = mddev;
1897 INIT_LIST_HEAD(&conf->retry_list);
1899 spin_lock_init(&conf->resync_lock);
1900 init_waitqueue_head(&conf->wait_barrier);
1902 bio_list_init(&conf->pending_bio_list);
1904 conf->last_used = -1;
1905 for (i = 0; i < conf->raid_disks; i++) {
1907 disk = conf->mirrors + i;
1910 !test_bit(In_sync, &disk->rdev->flags)) {
1911 disk->head_position = 0;
1914 } else if (conf->last_used < 0)
1916 * The first working device is used as a
1917 * starting point to read balancing.
1919 conf->last_used = i;
1923 if (conf->last_used < 0) {
1924 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
1929 conf->thread = md_register_thread(raid1d, mddev, NULL);
1930 if (!conf->thread) {
1932 "md/raid1:%s: couldn't allocate thread\n",
1941 if (conf->r1bio_pool)
1942 mempool_destroy(conf->r1bio_pool);
1943 kfree(conf->mirrors);
1944 safe_put_page(conf->tmppage);
1945 kfree(conf->poolinfo);
1948 return ERR_PTR(err);
1951 static int run(mddev_t *mddev)
1957 if (mddev->level != 1) {
1958 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
1959 mdname(mddev), mddev->level);
1962 if (mddev->reshape_position != MaxSector) {
1963 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
1968 * copy the already verified devices into our private RAID1
1969 * bookkeeping area. [whatever we allocate in run(),
1970 * should be freed in stop()]
1972 if (mddev->private == NULL)
1973 conf = setup_conf(mddev);
1975 conf = mddev->private;
1978 return PTR_ERR(conf);
1980 list_for_each_entry(rdev, &mddev->disks, same_set) {
1981 if (!mddev->gendisk)
1983 disk_stack_limits(mddev->gendisk, rdev->bdev,
1984 rdev->data_offset << 9);
1985 /* as we don't honour merge_bvec_fn, we must never risk
1986 * violating it, so limit ->max_segments to 1 lying within
1987 * a single page, as a one page request is never in violation.
1989 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1990 blk_queue_max_segments(mddev->queue, 1);
1991 blk_queue_segment_boundary(mddev->queue,
1992 PAGE_CACHE_SIZE - 1);
1996 mddev->degraded = 0;
1997 for (i=0; i < conf->raid_disks; i++)
1998 if (conf->mirrors[i].rdev == NULL ||
1999 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2000 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2003 if (conf->raid_disks - mddev->degraded == 1)
2004 mddev->recovery_cp = MaxSector;
2006 if (mddev->recovery_cp != MaxSector)
2007 printk(KERN_NOTICE "md/raid1:%s: not clean"
2008 " -- starting background reconstruction\n",
2011 "md/raid1:%s: active with %d out of %d mirrors\n",
2012 mdname(mddev), mddev->raid_disks - mddev->degraded,
2016 * Ok, everything is just fine now
2018 mddev->thread = conf->thread;
2019 conf->thread = NULL;
2020 mddev->private = conf;
2022 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2025 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2026 mddev->queue->backing_dev_info.congested_data = mddev;
2028 return md_integrity_register(mddev);
2031 static int stop(mddev_t *mddev)
2033 conf_t *conf = mddev->private;
2034 struct bitmap *bitmap = mddev->bitmap;
2036 /* wait for behind writes to complete */
2037 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2038 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2040 /* need to kick something here to make sure I/O goes? */
2041 wait_event(bitmap->behind_wait,
2042 atomic_read(&bitmap->behind_writes) == 0);
2045 raise_barrier(conf);
2046 lower_barrier(conf);
2048 md_unregister_thread(mddev->thread);
2049 mddev->thread = NULL;
2050 if (conf->r1bio_pool)
2051 mempool_destroy(conf->r1bio_pool);
2052 kfree(conf->mirrors);
2053 kfree(conf->poolinfo);
2055 mddev->private = NULL;
2059 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2061 /* no resync is happening, and there is enough space
2062 * on all devices, so we can resize.
2063 * We need to make sure resync covers any new space.
2064 * If the array is shrinking we should possibly wait until
2065 * any io in the removed space completes, but it hardly seems
2068 md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2069 if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2071 set_capacity(mddev->gendisk, mddev->array_sectors);
2072 revalidate_disk(mddev->gendisk);
2073 if (sectors > mddev->dev_sectors &&
2074 mddev->recovery_cp > mddev->dev_sectors) {
2075 mddev->recovery_cp = mddev->dev_sectors;
2076 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2078 mddev->dev_sectors = sectors;
2079 mddev->resync_max_sectors = sectors;
2083 static int raid1_reshape(mddev_t *mddev)
2086 * 1/ resize the r1bio_pool
2087 * 2/ resize conf->mirrors
2089 * We allocate a new r1bio_pool if we can.
2090 * Then raise a device barrier and wait until all IO stops.
2091 * Then resize conf->mirrors and swap in the new r1bio pool.
2093 * At the same time, we "pack" the devices so that all the missing
2094 * devices have the higher raid_disk numbers.
2096 mempool_t *newpool, *oldpool;
2097 struct pool_info *newpoolinfo;
2098 mirror_info_t *newmirrors;
2099 conf_t *conf = mddev->private;
2100 int cnt, raid_disks;
2101 unsigned long flags;
2104 /* Cannot change chunk_size, layout, or level */
2105 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2106 mddev->layout != mddev->new_layout ||
2107 mddev->level != mddev->new_level) {
2108 mddev->new_chunk_sectors = mddev->chunk_sectors;
2109 mddev->new_layout = mddev->layout;
2110 mddev->new_level = mddev->level;
2114 err = md_allow_write(mddev);
2118 raid_disks = mddev->raid_disks + mddev->delta_disks;
2120 if (raid_disks < conf->raid_disks) {
2122 for (d= 0; d < conf->raid_disks; d++)
2123 if (conf->mirrors[d].rdev)
2125 if (cnt > raid_disks)
2129 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2132 newpoolinfo->mddev = mddev;
2133 newpoolinfo->raid_disks = raid_disks;
2135 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2136 r1bio_pool_free, newpoolinfo);
2141 newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2144 mempool_destroy(newpool);
2148 raise_barrier(conf);
2150 /* ok, everything is stopped */
2151 oldpool = conf->r1bio_pool;
2152 conf->r1bio_pool = newpool;
2154 for (d = d2 = 0; d < conf->raid_disks; d++) {
2155 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2156 if (rdev && rdev->raid_disk != d2) {
2158 sprintf(nm, "rd%d", rdev->raid_disk);
2159 sysfs_remove_link(&mddev->kobj, nm);
2160 rdev->raid_disk = d2;
2161 sprintf(nm, "rd%d", rdev->raid_disk);
2162 sysfs_remove_link(&mddev->kobj, nm);
2163 if (sysfs_create_link(&mddev->kobj,
2166 "md/raid1:%s: cannot register "
2171 newmirrors[d2++].rdev = rdev;
2173 kfree(conf->mirrors);
2174 conf->mirrors = newmirrors;
2175 kfree(conf->poolinfo);
2176 conf->poolinfo = newpoolinfo;
2178 spin_lock_irqsave(&conf->device_lock, flags);
2179 mddev->degraded += (raid_disks - conf->raid_disks);
2180 spin_unlock_irqrestore(&conf->device_lock, flags);
2181 conf->raid_disks = mddev->raid_disks = raid_disks;
2182 mddev->delta_disks = 0;
2184 conf->last_used = 0; /* just make sure it is in-range */
2185 lower_barrier(conf);
2187 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2188 md_wakeup_thread(mddev->thread);
2190 mempool_destroy(oldpool);
2194 static void raid1_quiesce(mddev_t *mddev, int state)
2196 conf_t *conf = mddev->private;
2199 case 2: /* wake for suspend */
2200 wake_up(&conf->wait_barrier);
2203 raise_barrier(conf);
2206 lower_barrier(conf);
2211 static void *raid1_takeover(mddev_t *mddev)
2213 /* raid1 can take over:
2214 * raid5 with 2 devices, any layout or chunk size
2216 if (mddev->level == 5 && mddev->raid_disks == 2) {
2218 mddev->new_level = 1;
2219 mddev->new_layout = 0;
2220 mddev->new_chunk_sectors = 0;
2221 conf = setup_conf(mddev);
2226 return ERR_PTR(-EINVAL);
2229 static struct mdk_personality raid1_personality =
2233 .owner = THIS_MODULE,
2234 .make_request = make_request,
2238 .error_handler = error,
2239 .hot_add_disk = raid1_add_disk,
2240 .hot_remove_disk= raid1_remove_disk,
2241 .spare_active = raid1_spare_active,
2242 .sync_request = sync_request,
2243 .resize = raid1_resize,
2245 .check_reshape = raid1_reshape,
2246 .quiesce = raid1_quiesce,
2247 .takeover = raid1_takeover,
2250 static int __init raid_init(void)
2252 return register_md_personality(&raid1_personality);
2255 static void raid_exit(void)
2257 unregister_md_personality(&raid1_personality);
2260 module_init(raid_init);
2261 module_exit(raid_exit);
2262 MODULE_LICENSE("GPL");
2263 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2264 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2265 MODULE_ALIAS("md-raid1");
2266 MODULE_ALIAS("md-level-1");