Loading...
1/*
2 * raid1.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5 *
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7 *
8 * RAID-1 management functions.
9 *
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11 *
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14 *
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
17 *
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
20 *
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
23 *
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)
27 * any later version.
28 *
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.
32 */
33
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>
39#include "md.h"
40#include "raid1.h"
41#include "bitmap.h"
42
43#define DEBUG 0
44#define PRINTK(x...) do { if (DEBUG) printk(x); } while (0)
45
46/*
47 * Number of guaranteed r1bios in case of extreme VM load:
48 */
49#define NR_RAID1_BIOS 256
50
51
52static void allow_barrier(conf_t *conf);
53static void lower_barrier(conf_t *conf);
54
55static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
56{
57 struct pool_info *pi = data;
58 int size = offsetof(r1bio_t, bios[pi->raid_disks]);
59
60 /* allocate a r1bio with room for raid_disks entries in the bios array */
61 return kzalloc(size, gfp_flags);
62}
63
64static void r1bio_pool_free(void *r1_bio, void *data)
65{
66 kfree(r1_bio);
67}
68
69#define RESYNC_BLOCK_SIZE (64*1024)
70//#define RESYNC_BLOCK_SIZE PAGE_SIZE
71#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
72#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
73#define RESYNC_WINDOW (2048*1024)
74
75static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
76{
77 struct pool_info *pi = data;
78 struct page *page;
79 r1bio_t *r1_bio;
80 struct bio *bio;
81 int i, j;
82
83 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
84 if (!r1_bio)
85 return NULL;
86
87 /*
88 * Allocate bios : 1 for reading, n-1 for writing
89 */
90 for (j = pi->raid_disks ; j-- ; ) {
91 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
92 if (!bio)
93 goto out_free_bio;
94 r1_bio->bios[j] = bio;
95 }
96 /*
97 * Allocate RESYNC_PAGES data pages and attach them to
98 * the first bio.
99 * If this is a user-requested check/repair, allocate
100 * RESYNC_PAGES for each bio.
101 */
102 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
103 j = pi->raid_disks;
104 else
105 j = 1;
106 while(j--) {
107 bio = r1_bio->bios[j];
108 for (i = 0; i < RESYNC_PAGES; i++) {
109 page = alloc_page(gfp_flags);
110 if (unlikely(!page))
111 goto out_free_pages;
112
113 bio->bi_io_vec[i].bv_page = page;
114 bio->bi_vcnt = i+1;
115 }
116 }
117 /* If not user-requests, copy the page pointers to all bios */
118 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
119 for (i=0; i<RESYNC_PAGES ; i++)
120 for (j=1; j<pi->raid_disks; j++)
121 r1_bio->bios[j]->bi_io_vec[i].bv_page =
122 r1_bio->bios[0]->bi_io_vec[i].bv_page;
123 }
124
125 r1_bio->master_bio = NULL;
126
127 return r1_bio;
128
129out_free_pages:
130 for (j=0 ; j < pi->raid_disks; j++)
131 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
132 put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
133 j = -1;
134out_free_bio:
135 while ( ++j < pi->raid_disks )
136 bio_put(r1_bio->bios[j]);
137 r1bio_pool_free(r1_bio, data);
138 return NULL;
139}
140
141static void r1buf_pool_free(void *__r1_bio, void *data)
142{
143 struct pool_info *pi = data;
144 int i,j;
145 r1bio_t *r1bio = __r1_bio;
146
147 for (i = 0; i < RESYNC_PAGES; i++)
148 for (j = pi->raid_disks; j-- ;) {
149 if (j == 0 ||
150 r1bio->bios[j]->bi_io_vec[i].bv_page !=
151 r1bio->bios[0]->bi_io_vec[i].bv_page)
152 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
153 }
154 for (i=0 ; i < pi->raid_disks; i++)
155 bio_put(r1bio->bios[i]);
156
157 r1bio_pool_free(r1bio, data);
158}
159
160static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
161{
162 int i;
163
164 for (i = 0; i < conf->raid_disks; i++) {
165 struct bio **bio = r1_bio->bios + i;
166 if (!BIO_SPECIAL(*bio))
167 bio_put(*bio);
168 *bio = NULL;
169 }
170}
171
172static void free_r1bio(r1bio_t *r1_bio)
173{
174 conf_t *conf = r1_bio->mddev->private;
175
176 put_all_bios(conf, r1_bio);
177 mempool_free(r1_bio, conf->r1bio_pool);
178}
179
180static void put_buf(r1bio_t *r1_bio)
181{
182 conf_t *conf = r1_bio->mddev->private;
183 int i;
184
185 for (i=0; i<conf->raid_disks; i++) {
186 struct bio *bio = r1_bio->bios[i];
187 if (bio->bi_end_io)
188 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
189 }
190
191 mempool_free(r1_bio, conf->r1buf_pool);
192
193 lower_barrier(conf);
194}
195
196static void reschedule_retry(r1bio_t *r1_bio)
197{
198 unsigned long flags;
199 mddev_t *mddev = r1_bio->mddev;
200 conf_t *conf = mddev->private;
201
202 spin_lock_irqsave(&conf->device_lock, flags);
203 list_add(&r1_bio->retry_list, &conf->retry_list);
204 conf->nr_queued ++;
205 spin_unlock_irqrestore(&conf->device_lock, flags);
206
207 wake_up(&conf->wait_barrier);
208 md_wakeup_thread(mddev->thread);
209}
210
211/*
212 * raid_end_bio_io() is called when we have finished servicing a mirrored
213 * operation and are ready to return a success/failure code to the buffer
214 * cache layer.
215 */
216static void call_bio_endio(r1bio_t *r1_bio)
217{
218 struct bio *bio = r1_bio->master_bio;
219 int done;
220 conf_t *conf = r1_bio->mddev->private;
221
222 if (bio->bi_phys_segments) {
223 unsigned long flags;
224 spin_lock_irqsave(&conf->device_lock, flags);
225 bio->bi_phys_segments--;
226 done = (bio->bi_phys_segments == 0);
227 spin_unlock_irqrestore(&conf->device_lock, flags);
228 } else
229 done = 1;
230
231 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
232 clear_bit(BIO_UPTODATE, &bio->bi_flags);
233 if (done) {
234 bio_endio(bio, 0);
235 /*
236 * Wake up any possible resync thread that waits for the device
237 * to go idle.
238 */
239 allow_barrier(conf);
240 }
241}
242
243static void raid_end_bio_io(r1bio_t *r1_bio)
244{
245 struct bio *bio = r1_bio->master_bio;
246
247 /* if nobody has done the final endio yet, do it now */
248 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
249 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
250 (bio_data_dir(bio) == WRITE) ? "write" : "read",
251 (unsigned long long) bio->bi_sector,
252 (unsigned long long) bio->bi_sector +
253 (bio->bi_size >> 9) - 1);
254
255 call_bio_endio(r1_bio);
256 }
257 free_r1bio(r1_bio);
258}
259
260/*
261 * Update disk head position estimator based on IRQ completion info.
262 */
263static inline void update_head_pos(int disk, r1bio_t *r1_bio)
264{
265 conf_t *conf = r1_bio->mddev->private;
266
267 conf->mirrors[disk].head_position =
268 r1_bio->sector + (r1_bio->sectors);
269}
270
271static void raid1_end_read_request(struct bio *bio, int error)
272{
273 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
274 r1bio_t *r1_bio = bio->bi_private;
275 int mirror;
276 conf_t *conf = r1_bio->mddev->private;
277
278 mirror = r1_bio->read_disk;
279 /*
280 * this branch is our 'one mirror IO has finished' event handler:
281 */
282 update_head_pos(mirror, r1_bio);
283
284 if (uptodate)
285 set_bit(R1BIO_Uptodate, &r1_bio->state);
286 else {
287 /* If all other devices have failed, we want to return
288 * the error upwards rather than fail the last device.
289 * Here we redefine "uptodate" to mean "Don't want to retry"
290 */
291 unsigned long flags;
292 spin_lock_irqsave(&conf->device_lock, flags);
293 if (r1_bio->mddev->degraded == conf->raid_disks ||
294 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
295 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
296 uptodate = 1;
297 spin_unlock_irqrestore(&conf->device_lock, flags);
298 }
299
300 if (uptodate)
301 raid_end_bio_io(r1_bio);
302 else {
303 /*
304 * oops, read error:
305 */
306 char b[BDEVNAME_SIZE];
307 printk_ratelimited(
308 KERN_ERR "md/raid1:%s: %s: "
309 "rescheduling sector %llu\n",
310 mdname(conf->mddev),
311 bdevname(conf->mirrors[mirror].rdev->bdev,
312 b),
313 (unsigned long long)r1_bio->sector);
314 set_bit(R1BIO_ReadError, &r1_bio->state);
315 reschedule_retry(r1_bio);
316 }
317
318 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
319}
320
321static void close_write(r1bio_t *r1_bio)
322{
323 /* it really is the end of this request */
324 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
325 /* free extra copy of the data pages */
326 int i = r1_bio->behind_page_count;
327 while (i--)
328 safe_put_page(r1_bio->behind_bvecs[i].bv_page);
329 kfree(r1_bio->behind_bvecs);
330 r1_bio->behind_bvecs = NULL;
331 }
332 /* clear the bitmap if all writes complete successfully */
333 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
334 r1_bio->sectors,
335 !test_bit(R1BIO_Degraded, &r1_bio->state),
336 test_bit(R1BIO_BehindIO, &r1_bio->state));
337 md_write_end(r1_bio->mddev);
338}
339
340static void r1_bio_write_done(r1bio_t *r1_bio)
341{
342 if (!atomic_dec_and_test(&r1_bio->remaining))
343 return;
344
345 if (test_bit(R1BIO_WriteError, &r1_bio->state))
346 reschedule_retry(r1_bio);
347 else {
348 close_write(r1_bio);
349 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
350 reschedule_retry(r1_bio);
351 else
352 raid_end_bio_io(r1_bio);
353 }
354}
355
356static void raid1_end_write_request(struct bio *bio, int error)
357{
358 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
359 r1bio_t *r1_bio = bio->bi_private;
360 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
361 conf_t *conf = r1_bio->mddev->private;
362 struct bio *to_put = NULL;
363
364
365 for (mirror = 0; mirror < conf->raid_disks; mirror++)
366 if (r1_bio->bios[mirror] == bio)
367 break;
368
369 /*
370 * 'one mirror IO has finished' event handler:
371 */
372 if (!uptodate) {
373 set_bit(WriteErrorSeen,
374 &conf->mirrors[mirror].rdev->flags);
375 set_bit(R1BIO_WriteError, &r1_bio->state);
376 } else {
377 /*
378 * Set R1BIO_Uptodate in our master bio, so that we
379 * will return a good error code for to the higher
380 * levels even if IO on some other mirrored buffer
381 * fails.
382 *
383 * The 'master' represents the composite IO operation
384 * to user-side. So if something waits for IO, then it
385 * will wait for the 'master' bio.
386 */
387 sector_t first_bad;
388 int bad_sectors;
389
390 r1_bio->bios[mirror] = NULL;
391 to_put = bio;
392 set_bit(R1BIO_Uptodate, &r1_bio->state);
393
394 /* Maybe we can clear some bad blocks. */
395 if (is_badblock(conf->mirrors[mirror].rdev,
396 r1_bio->sector, r1_bio->sectors,
397 &first_bad, &bad_sectors)) {
398 r1_bio->bios[mirror] = IO_MADE_GOOD;
399 set_bit(R1BIO_MadeGood, &r1_bio->state);
400 }
401 }
402
403 update_head_pos(mirror, r1_bio);
404
405 if (behind) {
406 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
407 atomic_dec(&r1_bio->behind_remaining);
408
409 /*
410 * In behind mode, we ACK the master bio once the I/O
411 * has safely reached all non-writemostly
412 * disks. Setting the Returned bit ensures that this
413 * gets done only once -- we don't ever want to return
414 * -EIO here, instead we'll wait
415 */
416 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
417 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
418 /* Maybe we can return now */
419 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
420 struct bio *mbio = r1_bio->master_bio;
421 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
422 (unsigned long long) mbio->bi_sector,
423 (unsigned long long) mbio->bi_sector +
424 (mbio->bi_size >> 9) - 1);
425 call_bio_endio(r1_bio);
426 }
427 }
428 }
429 if (r1_bio->bios[mirror] == NULL)
430 rdev_dec_pending(conf->mirrors[mirror].rdev,
431 conf->mddev);
432
433 /*
434 * Let's see if all mirrored write operations have finished
435 * already.
436 */
437 r1_bio_write_done(r1_bio);
438
439 if (to_put)
440 bio_put(to_put);
441}
442
443
444/*
445 * This routine returns the disk from which the requested read should
446 * be done. There is a per-array 'next expected sequential IO' sector
447 * number - if this matches on the next IO then we use the last disk.
448 * There is also a per-disk 'last know head position' sector that is
449 * maintained from IRQ contexts, both the normal and the resync IO
450 * completion handlers update this position correctly. If there is no
451 * perfect sequential match then we pick the disk whose head is closest.
452 *
453 * If there are 2 mirrors in the same 2 devices, performance degrades
454 * because position is mirror, not device based.
455 *
456 * The rdev for the device selected will have nr_pending incremented.
457 */
458static int read_balance(conf_t *conf, r1bio_t *r1_bio, int *max_sectors)
459{
460 const sector_t this_sector = r1_bio->sector;
461 int sectors;
462 int best_good_sectors;
463 int start_disk;
464 int best_disk;
465 int i;
466 sector_t best_dist;
467 mdk_rdev_t *rdev;
468 int choose_first;
469
470 rcu_read_lock();
471 /*
472 * Check if we can balance. We can balance on the whole
473 * device if no resync is going on, or below the resync window.
474 * We take the first readable disk when above the resync window.
475 */
476 retry:
477 sectors = r1_bio->sectors;
478 best_disk = -1;
479 best_dist = MaxSector;
480 best_good_sectors = 0;
481
482 if (conf->mddev->recovery_cp < MaxSector &&
483 (this_sector + sectors >= conf->next_resync)) {
484 choose_first = 1;
485 start_disk = 0;
486 } else {
487 choose_first = 0;
488 start_disk = conf->last_used;
489 }
490
491 for (i = 0 ; i < conf->raid_disks ; i++) {
492 sector_t dist;
493 sector_t first_bad;
494 int bad_sectors;
495
496 int disk = start_disk + i;
497 if (disk >= conf->raid_disks)
498 disk -= conf->raid_disks;
499
500 rdev = rcu_dereference(conf->mirrors[disk].rdev);
501 if (r1_bio->bios[disk] == IO_BLOCKED
502 || rdev == NULL
503 || test_bit(Faulty, &rdev->flags))
504 continue;
505 if (!test_bit(In_sync, &rdev->flags) &&
506 rdev->recovery_offset < this_sector + sectors)
507 continue;
508 if (test_bit(WriteMostly, &rdev->flags)) {
509 /* Don't balance among write-mostly, just
510 * use the first as a last resort */
511 if (best_disk < 0)
512 best_disk = disk;
513 continue;
514 }
515 /* This is a reasonable device to use. It might
516 * even be best.
517 */
518 if (is_badblock(rdev, this_sector, sectors,
519 &first_bad, &bad_sectors)) {
520 if (best_dist < MaxSector)
521 /* already have a better device */
522 continue;
523 if (first_bad <= this_sector) {
524 /* cannot read here. If this is the 'primary'
525 * device, then we must not read beyond
526 * bad_sectors from another device..
527 */
528 bad_sectors -= (this_sector - first_bad);
529 if (choose_first && sectors > bad_sectors)
530 sectors = bad_sectors;
531 if (best_good_sectors > sectors)
532 best_good_sectors = sectors;
533
534 } else {
535 sector_t good_sectors = first_bad - this_sector;
536 if (good_sectors > best_good_sectors) {
537 best_good_sectors = good_sectors;
538 best_disk = disk;
539 }
540 if (choose_first)
541 break;
542 }
543 continue;
544 } else
545 best_good_sectors = sectors;
546
547 dist = abs(this_sector - conf->mirrors[disk].head_position);
548 if (choose_first
549 /* Don't change to another disk for sequential reads */
550 || conf->next_seq_sect == this_sector
551 || dist == 0
552 /* If device is idle, use it */
553 || atomic_read(&rdev->nr_pending) == 0) {
554 best_disk = disk;
555 break;
556 }
557 if (dist < best_dist) {
558 best_dist = dist;
559 best_disk = disk;
560 }
561 }
562
563 if (best_disk >= 0) {
564 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
565 if (!rdev)
566 goto retry;
567 atomic_inc(&rdev->nr_pending);
568 if (test_bit(Faulty, &rdev->flags)) {
569 /* cannot risk returning a device that failed
570 * before we inc'ed nr_pending
571 */
572 rdev_dec_pending(rdev, conf->mddev);
573 goto retry;
574 }
575 sectors = best_good_sectors;
576 conf->next_seq_sect = this_sector + sectors;
577 conf->last_used = best_disk;
578 }
579 rcu_read_unlock();
580 *max_sectors = sectors;
581
582 return best_disk;
583}
584
585int md_raid1_congested(mddev_t *mddev, int bits)
586{
587 conf_t *conf = mddev->private;
588 int i, ret = 0;
589
590 rcu_read_lock();
591 for (i = 0; i < mddev->raid_disks; i++) {
592 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
593 if (rdev && !test_bit(Faulty, &rdev->flags)) {
594 struct request_queue *q = bdev_get_queue(rdev->bdev);
595
596 BUG_ON(!q);
597
598 /* Note the '|| 1' - when read_balance prefers
599 * non-congested targets, it can be removed
600 */
601 if ((bits & (1<<BDI_async_congested)) || 1)
602 ret |= bdi_congested(&q->backing_dev_info, bits);
603 else
604 ret &= bdi_congested(&q->backing_dev_info, bits);
605 }
606 }
607 rcu_read_unlock();
608 return ret;
609}
610EXPORT_SYMBOL_GPL(md_raid1_congested);
611
612static int raid1_congested(void *data, int bits)
613{
614 mddev_t *mddev = data;
615
616 return mddev_congested(mddev, bits) ||
617 md_raid1_congested(mddev, bits);
618}
619
620static void flush_pending_writes(conf_t *conf)
621{
622 /* Any writes that have been queued but are awaiting
623 * bitmap updates get flushed here.
624 */
625 spin_lock_irq(&conf->device_lock);
626
627 if (conf->pending_bio_list.head) {
628 struct bio *bio;
629 bio = bio_list_get(&conf->pending_bio_list);
630 spin_unlock_irq(&conf->device_lock);
631 /* flush any pending bitmap writes to
632 * disk before proceeding w/ I/O */
633 bitmap_unplug(conf->mddev->bitmap);
634
635 while (bio) { /* submit pending writes */
636 struct bio *next = bio->bi_next;
637 bio->bi_next = NULL;
638 generic_make_request(bio);
639 bio = next;
640 }
641 } else
642 spin_unlock_irq(&conf->device_lock);
643}
644
645/* Barriers....
646 * Sometimes we need to suspend IO while we do something else,
647 * either some resync/recovery, or reconfigure the array.
648 * To do this we raise a 'barrier'.
649 * The 'barrier' is a counter that can be raised multiple times
650 * to count how many activities are happening which preclude
651 * normal IO.
652 * We can only raise the barrier if there is no pending IO.
653 * i.e. if nr_pending == 0.
654 * We choose only to raise the barrier if no-one is waiting for the
655 * barrier to go down. This means that as soon as an IO request
656 * is ready, no other operations which require a barrier will start
657 * until the IO request has had a chance.
658 *
659 * So: regular IO calls 'wait_barrier'. When that returns there
660 * is no backgroup IO happening, It must arrange to call
661 * allow_barrier when it has finished its IO.
662 * backgroup IO calls must call raise_barrier. Once that returns
663 * there is no normal IO happeing. It must arrange to call
664 * lower_barrier when the particular background IO completes.
665 */
666#define RESYNC_DEPTH 32
667
668static void raise_barrier(conf_t *conf)
669{
670 spin_lock_irq(&conf->resync_lock);
671
672 /* Wait until no block IO is waiting */
673 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
674 conf->resync_lock, );
675
676 /* block any new IO from starting */
677 conf->barrier++;
678
679 /* Now wait for all pending IO to complete */
680 wait_event_lock_irq(conf->wait_barrier,
681 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
682 conf->resync_lock, );
683
684 spin_unlock_irq(&conf->resync_lock);
685}
686
687static void lower_barrier(conf_t *conf)
688{
689 unsigned long flags;
690 BUG_ON(conf->barrier <= 0);
691 spin_lock_irqsave(&conf->resync_lock, flags);
692 conf->barrier--;
693 spin_unlock_irqrestore(&conf->resync_lock, flags);
694 wake_up(&conf->wait_barrier);
695}
696
697static void wait_barrier(conf_t *conf)
698{
699 spin_lock_irq(&conf->resync_lock);
700 if (conf->barrier) {
701 conf->nr_waiting++;
702 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
703 conf->resync_lock,
704 );
705 conf->nr_waiting--;
706 }
707 conf->nr_pending++;
708 spin_unlock_irq(&conf->resync_lock);
709}
710
711static void allow_barrier(conf_t *conf)
712{
713 unsigned long flags;
714 spin_lock_irqsave(&conf->resync_lock, flags);
715 conf->nr_pending--;
716 spin_unlock_irqrestore(&conf->resync_lock, flags);
717 wake_up(&conf->wait_barrier);
718}
719
720static void freeze_array(conf_t *conf)
721{
722 /* stop syncio and normal IO and wait for everything to
723 * go quite.
724 * We increment barrier and nr_waiting, and then
725 * wait until nr_pending match nr_queued+1
726 * This is called in the context of one normal IO request
727 * that has failed. Thus any sync request that might be pending
728 * will be blocked by nr_pending, and we need to wait for
729 * pending IO requests to complete or be queued for re-try.
730 * Thus the number queued (nr_queued) plus this request (1)
731 * must match the number of pending IOs (nr_pending) before
732 * we continue.
733 */
734 spin_lock_irq(&conf->resync_lock);
735 conf->barrier++;
736 conf->nr_waiting++;
737 wait_event_lock_irq(conf->wait_barrier,
738 conf->nr_pending == conf->nr_queued+1,
739 conf->resync_lock,
740 flush_pending_writes(conf));
741 spin_unlock_irq(&conf->resync_lock);
742}
743static void unfreeze_array(conf_t *conf)
744{
745 /* reverse the effect of the freeze */
746 spin_lock_irq(&conf->resync_lock);
747 conf->barrier--;
748 conf->nr_waiting--;
749 wake_up(&conf->wait_barrier);
750 spin_unlock_irq(&conf->resync_lock);
751}
752
753
754/* duplicate the data pages for behind I/O
755 */
756static void alloc_behind_pages(struct bio *bio, r1bio_t *r1_bio)
757{
758 int i;
759 struct bio_vec *bvec;
760 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
761 GFP_NOIO);
762 if (unlikely(!bvecs))
763 return;
764
765 bio_for_each_segment(bvec, bio, i) {
766 bvecs[i] = *bvec;
767 bvecs[i].bv_page = alloc_page(GFP_NOIO);
768 if (unlikely(!bvecs[i].bv_page))
769 goto do_sync_io;
770 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
771 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
772 kunmap(bvecs[i].bv_page);
773 kunmap(bvec->bv_page);
774 }
775 r1_bio->behind_bvecs = bvecs;
776 r1_bio->behind_page_count = bio->bi_vcnt;
777 set_bit(R1BIO_BehindIO, &r1_bio->state);
778 return;
779
780do_sync_io:
781 for (i = 0; i < bio->bi_vcnt; i++)
782 if (bvecs[i].bv_page)
783 put_page(bvecs[i].bv_page);
784 kfree(bvecs);
785 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
786}
787
788static int make_request(mddev_t *mddev, struct bio * bio)
789{
790 conf_t *conf = mddev->private;
791 mirror_info_t *mirror;
792 r1bio_t *r1_bio;
793 struct bio *read_bio;
794 int i, disks;
795 struct bitmap *bitmap;
796 unsigned long flags;
797 const int rw = bio_data_dir(bio);
798 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
799 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
800 mdk_rdev_t *blocked_rdev;
801 int plugged;
802 int first_clone;
803 int sectors_handled;
804 int max_sectors;
805
806 /*
807 * Register the new request and wait if the reconstruction
808 * thread has put up a bar for new requests.
809 * Continue immediately if no resync is active currently.
810 */
811
812 md_write_start(mddev, bio); /* wait on superblock update early */
813
814 if (bio_data_dir(bio) == WRITE &&
815 bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
816 bio->bi_sector < mddev->suspend_hi) {
817 /* As the suspend_* range is controlled by
818 * userspace, we want an interruptible
819 * wait.
820 */
821 DEFINE_WAIT(w);
822 for (;;) {
823 flush_signals(current);
824 prepare_to_wait(&conf->wait_barrier,
825 &w, TASK_INTERRUPTIBLE);
826 if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
827 bio->bi_sector >= mddev->suspend_hi)
828 break;
829 schedule();
830 }
831 finish_wait(&conf->wait_barrier, &w);
832 }
833
834 wait_barrier(conf);
835
836 bitmap = mddev->bitmap;
837
838 /*
839 * make_request() can abort the operation when READA is being
840 * used and no empty request is available.
841 *
842 */
843 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
844
845 r1_bio->master_bio = bio;
846 r1_bio->sectors = bio->bi_size >> 9;
847 r1_bio->state = 0;
848 r1_bio->mddev = mddev;
849 r1_bio->sector = bio->bi_sector;
850
851 /* We might need to issue multiple reads to different
852 * devices if there are bad blocks around, so we keep
853 * track of the number of reads in bio->bi_phys_segments.
854 * If this is 0, there is only one r1_bio and no locking
855 * will be needed when requests complete. If it is
856 * non-zero, then it is the number of not-completed requests.
857 */
858 bio->bi_phys_segments = 0;
859 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
860
861 if (rw == READ) {
862 /*
863 * read balancing logic:
864 */
865 int rdisk;
866
867read_again:
868 rdisk = read_balance(conf, r1_bio, &max_sectors);
869
870 if (rdisk < 0) {
871 /* couldn't find anywhere to read from */
872 raid_end_bio_io(r1_bio);
873 return 0;
874 }
875 mirror = conf->mirrors + rdisk;
876
877 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
878 bitmap) {
879 /* Reading from a write-mostly device must
880 * take care not to over-take any writes
881 * that are 'behind'
882 */
883 wait_event(bitmap->behind_wait,
884 atomic_read(&bitmap->behind_writes) == 0);
885 }
886 r1_bio->read_disk = rdisk;
887
888 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
889 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
890 max_sectors);
891
892 r1_bio->bios[rdisk] = read_bio;
893
894 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
895 read_bio->bi_bdev = mirror->rdev->bdev;
896 read_bio->bi_end_io = raid1_end_read_request;
897 read_bio->bi_rw = READ | do_sync;
898 read_bio->bi_private = r1_bio;
899
900 if (max_sectors < r1_bio->sectors) {
901 /* could not read all from this device, so we will
902 * need another r1_bio.
903 */
904
905 sectors_handled = (r1_bio->sector + max_sectors
906 - bio->bi_sector);
907 r1_bio->sectors = max_sectors;
908 spin_lock_irq(&conf->device_lock);
909 if (bio->bi_phys_segments == 0)
910 bio->bi_phys_segments = 2;
911 else
912 bio->bi_phys_segments++;
913 spin_unlock_irq(&conf->device_lock);
914 /* Cannot call generic_make_request directly
915 * as that will be queued in __make_request
916 * and subsequent mempool_alloc might block waiting
917 * for it. So hand bio over to raid1d.
918 */
919 reschedule_retry(r1_bio);
920
921 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
922
923 r1_bio->master_bio = bio;
924 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
925 r1_bio->state = 0;
926 r1_bio->mddev = mddev;
927 r1_bio->sector = bio->bi_sector + sectors_handled;
928 goto read_again;
929 } else
930 generic_make_request(read_bio);
931 return 0;
932 }
933
934 /*
935 * WRITE:
936 */
937 /* first select target devices under rcu_lock and
938 * inc refcount on their rdev. Record them by setting
939 * bios[x] to bio
940 * If there are known/acknowledged bad blocks on any device on
941 * which we have seen a write error, we want to avoid writing those
942 * blocks.
943 * This potentially requires several writes to write around
944 * the bad blocks. Each set of writes gets it's own r1bio
945 * with a set of bios attached.
946 */
947 plugged = mddev_check_plugged(mddev);
948
949 disks = conf->raid_disks;
950 retry_write:
951 blocked_rdev = NULL;
952 rcu_read_lock();
953 max_sectors = r1_bio->sectors;
954 for (i = 0; i < disks; i++) {
955 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
956 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
957 atomic_inc(&rdev->nr_pending);
958 blocked_rdev = rdev;
959 break;
960 }
961 r1_bio->bios[i] = NULL;
962 if (!rdev || test_bit(Faulty, &rdev->flags)) {
963 set_bit(R1BIO_Degraded, &r1_bio->state);
964 continue;
965 }
966
967 atomic_inc(&rdev->nr_pending);
968 if (test_bit(WriteErrorSeen, &rdev->flags)) {
969 sector_t first_bad;
970 int bad_sectors;
971 int is_bad;
972
973 is_bad = is_badblock(rdev, r1_bio->sector,
974 max_sectors,
975 &first_bad, &bad_sectors);
976 if (is_bad < 0) {
977 /* mustn't write here until the bad block is
978 * acknowledged*/
979 set_bit(BlockedBadBlocks, &rdev->flags);
980 blocked_rdev = rdev;
981 break;
982 }
983 if (is_bad && first_bad <= r1_bio->sector) {
984 /* Cannot write here at all */
985 bad_sectors -= (r1_bio->sector - first_bad);
986 if (bad_sectors < max_sectors)
987 /* mustn't write more than bad_sectors
988 * to other devices yet
989 */
990 max_sectors = bad_sectors;
991 rdev_dec_pending(rdev, mddev);
992 /* We don't set R1BIO_Degraded as that
993 * only applies if the disk is
994 * missing, so it might be re-added,
995 * and we want to know to recover this
996 * chunk.
997 * In this case the device is here,
998 * and the fact that this chunk is not
999 * in-sync is recorded in the bad
1000 * block log
1001 */
1002 continue;
1003 }
1004 if (is_bad) {
1005 int good_sectors = first_bad - r1_bio->sector;
1006 if (good_sectors < max_sectors)
1007 max_sectors = good_sectors;
1008 }
1009 }
1010 r1_bio->bios[i] = bio;
1011 }
1012 rcu_read_unlock();
1013
1014 if (unlikely(blocked_rdev)) {
1015 /* Wait for this device to become unblocked */
1016 int j;
1017
1018 for (j = 0; j < i; j++)
1019 if (r1_bio->bios[j])
1020 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1021 r1_bio->state = 0;
1022 allow_barrier(conf);
1023 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1024 wait_barrier(conf);
1025 goto retry_write;
1026 }
1027
1028 if (max_sectors < r1_bio->sectors) {
1029 /* We are splitting this write into multiple parts, so
1030 * we need to prepare for allocating another r1_bio.
1031 */
1032 r1_bio->sectors = max_sectors;
1033 spin_lock_irq(&conf->device_lock);
1034 if (bio->bi_phys_segments == 0)
1035 bio->bi_phys_segments = 2;
1036 else
1037 bio->bi_phys_segments++;
1038 spin_unlock_irq(&conf->device_lock);
1039 }
1040 sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1041
1042 atomic_set(&r1_bio->remaining, 1);
1043 atomic_set(&r1_bio->behind_remaining, 0);
1044
1045 first_clone = 1;
1046 for (i = 0; i < disks; i++) {
1047 struct bio *mbio;
1048 if (!r1_bio->bios[i])
1049 continue;
1050
1051 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1052 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1053
1054 if (first_clone) {
1055 /* do behind I/O ?
1056 * Not if there are too many, or cannot
1057 * allocate memory, or a reader on WriteMostly
1058 * is waiting for behind writes to flush */
1059 if (bitmap &&
1060 (atomic_read(&bitmap->behind_writes)
1061 < mddev->bitmap_info.max_write_behind) &&
1062 !waitqueue_active(&bitmap->behind_wait))
1063 alloc_behind_pages(mbio, r1_bio);
1064
1065 bitmap_startwrite(bitmap, r1_bio->sector,
1066 r1_bio->sectors,
1067 test_bit(R1BIO_BehindIO,
1068 &r1_bio->state));
1069 first_clone = 0;
1070 }
1071 if (r1_bio->behind_bvecs) {
1072 struct bio_vec *bvec;
1073 int j;
1074
1075 /* Yes, I really want the '__' version so that
1076 * we clear any unused pointer in the io_vec, rather
1077 * than leave them unchanged. This is important
1078 * because when we come to free the pages, we won't
1079 * know the original bi_idx, so we just free
1080 * them all
1081 */
1082 __bio_for_each_segment(bvec, mbio, j, 0)
1083 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1084 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1085 atomic_inc(&r1_bio->behind_remaining);
1086 }
1087
1088 r1_bio->bios[i] = mbio;
1089
1090 mbio->bi_sector = (r1_bio->sector +
1091 conf->mirrors[i].rdev->data_offset);
1092 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1093 mbio->bi_end_io = raid1_end_write_request;
1094 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
1095 mbio->bi_private = r1_bio;
1096
1097 atomic_inc(&r1_bio->remaining);
1098 spin_lock_irqsave(&conf->device_lock, flags);
1099 bio_list_add(&conf->pending_bio_list, mbio);
1100 spin_unlock_irqrestore(&conf->device_lock, flags);
1101 }
1102 /* Mustn't call r1_bio_write_done before this next test,
1103 * as it could result in the bio being freed.
1104 */
1105 if (sectors_handled < (bio->bi_size >> 9)) {
1106 r1_bio_write_done(r1_bio);
1107 /* We need another r1_bio. It has already been counted
1108 * in bio->bi_phys_segments
1109 */
1110 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1111 r1_bio->master_bio = bio;
1112 r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1113 r1_bio->state = 0;
1114 r1_bio->mddev = mddev;
1115 r1_bio->sector = bio->bi_sector + sectors_handled;
1116 goto retry_write;
1117 }
1118
1119 r1_bio_write_done(r1_bio);
1120
1121 /* In case raid1d snuck in to freeze_array */
1122 wake_up(&conf->wait_barrier);
1123
1124 if (do_sync || !bitmap || !plugged)
1125 md_wakeup_thread(mddev->thread);
1126
1127 return 0;
1128}
1129
1130static void status(struct seq_file *seq, mddev_t *mddev)
1131{
1132 conf_t *conf = mddev->private;
1133 int i;
1134
1135 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1136 conf->raid_disks - mddev->degraded);
1137 rcu_read_lock();
1138 for (i = 0; i < conf->raid_disks; i++) {
1139 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1140 seq_printf(seq, "%s",
1141 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1142 }
1143 rcu_read_unlock();
1144 seq_printf(seq, "]");
1145}
1146
1147
1148static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1149{
1150 char b[BDEVNAME_SIZE];
1151 conf_t *conf = mddev->private;
1152
1153 /*
1154 * If it is not operational, then we have already marked it as dead
1155 * else if it is the last working disks, ignore the error, let the
1156 * next level up know.
1157 * else mark the drive as failed
1158 */
1159 if (test_bit(In_sync, &rdev->flags)
1160 && (conf->raid_disks - mddev->degraded) == 1) {
1161 /*
1162 * Don't fail the drive, act as though we were just a
1163 * normal single drive.
1164 * However don't try a recovery from this drive as
1165 * it is very likely to fail.
1166 */
1167 conf->recovery_disabled = mddev->recovery_disabled;
1168 return;
1169 }
1170 set_bit(Blocked, &rdev->flags);
1171 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1172 unsigned long flags;
1173 spin_lock_irqsave(&conf->device_lock, flags);
1174 mddev->degraded++;
1175 set_bit(Faulty, &rdev->flags);
1176 spin_unlock_irqrestore(&conf->device_lock, flags);
1177 /*
1178 * if recovery is running, make sure it aborts.
1179 */
1180 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1181 } else
1182 set_bit(Faulty, &rdev->flags);
1183 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1184 printk(KERN_ALERT
1185 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1186 "md/raid1:%s: Operation continuing on %d devices.\n",
1187 mdname(mddev), bdevname(rdev->bdev, b),
1188 mdname(mddev), conf->raid_disks - mddev->degraded);
1189}
1190
1191static void print_conf(conf_t *conf)
1192{
1193 int i;
1194
1195 printk(KERN_DEBUG "RAID1 conf printout:\n");
1196 if (!conf) {
1197 printk(KERN_DEBUG "(!conf)\n");
1198 return;
1199 }
1200 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1201 conf->raid_disks);
1202
1203 rcu_read_lock();
1204 for (i = 0; i < conf->raid_disks; i++) {
1205 char b[BDEVNAME_SIZE];
1206 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1207 if (rdev)
1208 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1209 i, !test_bit(In_sync, &rdev->flags),
1210 !test_bit(Faulty, &rdev->flags),
1211 bdevname(rdev->bdev,b));
1212 }
1213 rcu_read_unlock();
1214}
1215
1216static void close_sync(conf_t *conf)
1217{
1218 wait_barrier(conf);
1219 allow_barrier(conf);
1220
1221 mempool_destroy(conf->r1buf_pool);
1222 conf->r1buf_pool = NULL;
1223}
1224
1225static int raid1_spare_active(mddev_t *mddev)
1226{
1227 int i;
1228 conf_t *conf = mddev->private;
1229 int count = 0;
1230 unsigned long flags;
1231
1232 /*
1233 * Find all failed disks within the RAID1 configuration
1234 * and mark them readable.
1235 * Called under mddev lock, so rcu protection not needed.
1236 */
1237 for (i = 0; i < conf->raid_disks; i++) {
1238 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1239 if (rdev
1240 && !test_bit(Faulty, &rdev->flags)
1241 && !test_and_set_bit(In_sync, &rdev->flags)) {
1242 count++;
1243 sysfs_notify_dirent_safe(rdev->sysfs_state);
1244 }
1245 }
1246 spin_lock_irqsave(&conf->device_lock, flags);
1247 mddev->degraded -= count;
1248 spin_unlock_irqrestore(&conf->device_lock, flags);
1249
1250 print_conf(conf);
1251 return count;
1252}
1253
1254
1255static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1256{
1257 conf_t *conf = mddev->private;
1258 int err = -EEXIST;
1259 int mirror = 0;
1260 mirror_info_t *p;
1261 int first = 0;
1262 int last = mddev->raid_disks - 1;
1263
1264 if (mddev->recovery_disabled == conf->recovery_disabled)
1265 return -EBUSY;
1266
1267 if (rdev->raid_disk >= 0)
1268 first = last = rdev->raid_disk;
1269
1270 for (mirror = first; mirror <= last; mirror++)
1271 if ( !(p=conf->mirrors+mirror)->rdev) {
1272
1273 disk_stack_limits(mddev->gendisk, rdev->bdev,
1274 rdev->data_offset << 9);
1275 /* as we don't honour merge_bvec_fn, we must
1276 * never risk violating it, so limit
1277 * ->max_segments to one lying with a single
1278 * page, as a one page request is never in
1279 * violation.
1280 */
1281 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1282 blk_queue_max_segments(mddev->queue, 1);
1283 blk_queue_segment_boundary(mddev->queue,
1284 PAGE_CACHE_SIZE - 1);
1285 }
1286
1287 p->head_position = 0;
1288 rdev->raid_disk = mirror;
1289 err = 0;
1290 /* As all devices are equivalent, we don't need a full recovery
1291 * if this was recently any drive of the array
1292 */
1293 if (rdev->saved_raid_disk < 0)
1294 conf->fullsync = 1;
1295 rcu_assign_pointer(p->rdev, rdev);
1296 break;
1297 }
1298 md_integrity_add_rdev(rdev, mddev);
1299 print_conf(conf);
1300 return err;
1301}
1302
1303static int raid1_remove_disk(mddev_t *mddev, int number)
1304{
1305 conf_t *conf = mddev->private;
1306 int err = 0;
1307 mdk_rdev_t *rdev;
1308 mirror_info_t *p = conf->mirrors+ number;
1309
1310 print_conf(conf);
1311 rdev = p->rdev;
1312 if (rdev) {
1313 if (test_bit(In_sync, &rdev->flags) ||
1314 atomic_read(&rdev->nr_pending)) {
1315 err = -EBUSY;
1316 goto abort;
1317 }
1318 /* Only remove non-faulty devices if recovery
1319 * is not possible.
1320 */
1321 if (!test_bit(Faulty, &rdev->flags) &&
1322 mddev->recovery_disabled != conf->recovery_disabled &&
1323 mddev->degraded < conf->raid_disks) {
1324 err = -EBUSY;
1325 goto abort;
1326 }
1327 p->rdev = NULL;
1328 synchronize_rcu();
1329 if (atomic_read(&rdev->nr_pending)) {
1330 /* lost the race, try later */
1331 err = -EBUSY;
1332 p->rdev = rdev;
1333 goto abort;
1334 }
1335 err = md_integrity_register(mddev);
1336 }
1337abort:
1338
1339 print_conf(conf);
1340 return err;
1341}
1342
1343
1344static void end_sync_read(struct bio *bio, int error)
1345{
1346 r1bio_t *r1_bio = bio->bi_private;
1347 int i;
1348
1349 for (i=r1_bio->mddev->raid_disks; i--; )
1350 if (r1_bio->bios[i] == bio)
1351 break;
1352 BUG_ON(i < 0);
1353 update_head_pos(i, r1_bio);
1354 /*
1355 * we have read a block, now it needs to be re-written,
1356 * or re-read if the read failed.
1357 * We don't do much here, just schedule handling by raid1d
1358 */
1359 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1360 set_bit(R1BIO_Uptodate, &r1_bio->state);
1361
1362 if (atomic_dec_and_test(&r1_bio->remaining))
1363 reschedule_retry(r1_bio);
1364}
1365
1366static void end_sync_write(struct bio *bio, int error)
1367{
1368 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1369 r1bio_t *r1_bio = bio->bi_private;
1370 mddev_t *mddev = r1_bio->mddev;
1371 conf_t *conf = mddev->private;
1372 int i;
1373 int mirror=0;
1374 sector_t first_bad;
1375 int bad_sectors;
1376
1377 for (i = 0; i < conf->raid_disks; i++)
1378 if (r1_bio->bios[i] == bio) {
1379 mirror = i;
1380 break;
1381 }
1382 if (!uptodate) {
1383 sector_t sync_blocks = 0;
1384 sector_t s = r1_bio->sector;
1385 long sectors_to_go = r1_bio->sectors;
1386 /* make sure these bits doesn't get cleared. */
1387 do {
1388 bitmap_end_sync(mddev->bitmap, s,
1389 &sync_blocks, 1);
1390 s += sync_blocks;
1391 sectors_to_go -= sync_blocks;
1392 } while (sectors_to_go > 0);
1393 set_bit(WriteErrorSeen,
1394 &conf->mirrors[mirror].rdev->flags);
1395 set_bit(R1BIO_WriteError, &r1_bio->state);
1396 } else if (is_badblock(conf->mirrors[mirror].rdev,
1397 r1_bio->sector,
1398 r1_bio->sectors,
1399 &first_bad, &bad_sectors) &&
1400 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1401 r1_bio->sector,
1402 r1_bio->sectors,
1403 &first_bad, &bad_sectors)
1404 )
1405 set_bit(R1BIO_MadeGood, &r1_bio->state);
1406
1407 update_head_pos(mirror, r1_bio);
1408
1409 if (atomic_dec_and_test(&r1_bio->remaining)) {
1410 int s = r1_bio->sectors;
1411 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1412 test_bit(R1BIO_WriteError, &r1_bio->state))
1413 reschedule_retry(r1_bio);
1414 else {
1415 put_buf(r1_bio);
1416 md_done_sync(mddev, s, uptodate);
1417 }
1418 }
1419}
1420
1421static int r1_sync_page_io(mdk_rdev_t *rdev, sector_t sector,
1422 int sectors, struct page *page, int rw)
1423{
1424 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1425 /* success */
1426 return 1;
1427 if (rw == WRITE)
1428 set_bit(WriteErrorSeen, &rdev->flags);
1429 /* need to record an error - either for the block or the device */
1430 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1431 md_error(rdev->mddev, rdev);
1432 return 0;
1433}
1434
1435static int fix_sync_read_error(r1bio_t *r1_bio)
1436{
1437 /* Try some synchronous reads of other devices to get
1438 * good data, much like with normal read errors. Only
1439 * read into the pages we already have so we don't
1440 * need to re-issue the read request.
1441 * We don't need to freeze the array, because being in an
1442 * active sync request, there is no normal IO, and
1443 * no overlapping syncs.
1444 * We don't need to check is_badblock() again as we
1445 * made sure that anything with a bad block in range
1446 * will have bi_end_io clear.
1447 */
1448 mddev_t *mddev = r1_bio->mddev;
1449 conf_t *conf = mddev->private;
1450 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1451 sector_t sect = r1_bio->sector;
1452 int sectors = r1_bio->sectors;
1453 int idx = 0;
1454
1455 while(sectors) {
1456 int s = sectors;
1457 int d = r1_bio->read_disk;
1458 int success = 0;
1459 mdk_rdev_t *rdev;
1460 int start;
1461
1462 if (s > (PAGE_SIZE>>9))
1463 s = PAGE_SIZE >> 9;
1464 do {
1465 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1466 /* No rcu protection needed here devices
1467 * can only be removed when no resync is
1468 * active, and resync is currently active
1469 */
1470 rdev = conf->mirrors[d].rdev;
1471 if (sync_page_io(rdev, sect, s<<9,
1472 bio->bi_io_vec[idx].bv_page,
1473 READ, false)) {
1474 success = 1;
1475 break;
1476 }
1477 }
1478 d++;
1479 if (d == conf->raid_disks)
1480 d = 0;
1481 } while (!success && d != r1_bio->read_disk);
1482
1483 if (!success) {
1484 char b[BDEVNAME_SIZE];
1485 int abort = 0;
1486 /* Cannot read from anywhere, this block is lost.
1487 * Record a bad block on each device. If that doesn't
1488 * work just disable and interrupt the recovery.
1489 * Don't fail devices as that won't really help.
1490 */
1491 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1492 " for block %llu\n",
1493 mdname(mddev),
1494 bdevname(bio->bi_bdev, b),
1495 (unsigned long long)r1_bio->sector);
1496 for (d = 0; d < conf->raid_disks; d++) {
1497 rdev = conf->mirrors[d].rdev;
1498 if (!rdev || test_bit(Faulty, &rdev->flags))
1499 continue;
1500 if (!rdev_set_badblocks(rdev, sect, s, 0))
1501 abort = 1;
1502 }
1503 if (abort) {
1504 mddev->recovery_disabled = 1;
1505 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1506 md_done_sync(mddev, r1_bio->sectors, 0);
1507 put_buf(r1_bio);
1508 return 0;
1509 }
1510 /* Try next page */
1511 sectors -= s;
1512 sect += s;
1513 idx++;
1514 continue;
1515 }
1516
1517 start = d;
1518 /* write it back and re-read */
1519 while (d != r1_bio->read_disk) {
1520 if (d == 0)
1521 d = conf->raid_disks;
1522 d--;
1523 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1524 continue;
1525 rdev = conf->mirrors[d].rdev;
1526 if (r1_sync_page_io(rdev, sect, s,
1527 bio->bi_io_vec[idx].bv_page,
1528 WRITE) == 0) {
1529 r1_bio->bios[d]->bi_end_io = NULL;
1530 rdev_dec_pending(rdev, mddev);
1531 }
1532 }
1533 d = start;
1534 while (d != r1_bio->read_disk) {
1535 if (d == 0)
1536 d = conf->raid_disks;
1537 d--;
1538 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1539 continue;
1540 rdev = conf->mirrors[d].rdev;
1541 if (r1_sync_page_io(rdev, sect, s,
1542 bio->bi_io_vec[idx].bv_page,
1543 READ) != 0)
1544 atomic_add(s, &rdev->corrected_errors);
1545 }
1546 sectors -= s;
1547 sect += s;
1548 idx ++;
1549 }
1550 set_bit(R1BIO_Uptodate, &r1_bio->state);
1551 set_bit(BIO_UPTODATE, &bio->bi_flags);
1552 return 1;
1553}
1554
1555static int process_checks(r1bio_t *r1_bio)
1556{
1557 /* We have read all readable devices. If we haven't
1558 * got the block, then there is no hope left.
1559 * If we have, then we want to do a comparison
1560 * and skip the write if everything is the same.
1561 * If any blocks failed to read, then we need to
1562 * attempt an over-write
1563 */
1564 mddev_t *mddev = r1_bio->mddev;
1565 conf_t *conf = mddev->private;
1566 int primary;
1567 int i;
1568
1569 for (primary = 0; primary < conf->raid_disks; primary++)
1570 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1571 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1572 r1_bio->bios[primary]->bi_end_io = NULL;
1573 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1574 break;
1575 }
1576 r1_bio->read_disk = primary;
1577 for (i = 0; i < conf->raid_disks; i++) {
1578 int j;
1579 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1580 struct bio *pbio = r1_bio->bios[primary];
1581 struct bio *sbio = r1_bio->bios[i];
1582 int size;
1583
1584 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1585 continue;
1586
1587 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1588 for (j = vcnt; j-- ; ) {
1589 struct page *p, *s;
1590 p = pbio->bi_io_vec[j].bv_page;
1591 s = sbio->bi_io_vec[j].bv_page;
1592 if (memcmp(page_address(p),
1593 page_address(s),
1594 PAGE_SIZE))
1595 break;
1596 }
1597 } else
1598 j = 0;
1599 if (j >= 0)
1600 mddev->resync_mismatches += r1_bio->sectors;
1601 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1602 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1603 /* No need to write to this device. */
1604 sbio->bi_end_io = NULL;
1605 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1606 continue;
1607 }
1608 /* fixup the bio for reuse */
1609 sbio->bi_vcnt = vcnt;
1610 sbio->bi_size = r1_bio->sectors << 9;
1611 sbio->bi_idx = 0;
1612 sbio->bi_phys_segments = 0;
1613 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1614 sbio->bi_flags |= 1 << BIO_UPTODATE;
1615 sbio->bi_next = NULL;
1616 sbio->bi_sector = r1_bio->sector +
1617 conf->mirrors[i].rdev->data_offset;
1618 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1619 size = sbio->bi_size;
1620 for (j = 0; j < vcnt ; j++) {
1621 struct bio_vec *bi;
1622 bi = &sbio->bi_io_vec[j];
1623 bi->bv_offset = 0;
1624 if (size > PAGE_SIZE)
1625 bi->bv_len = PAGE_SIZE;
1626 else
1627 bi->bv_len = size;
1628 size -= PAGE_SIZE;
1629 memcpy(page_address(bi->bv_page),
1630 page_address(pbio->bi_io_vec[j].bv_page),
1631 PAGE_SIZE);
1632 }
1633 }
1634 return 0;
1635}
1636
1637static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1638{
1639 conf_t *conf = mddev->private;
1640 int i;
1641 int disks = conf->raid_disks;
1642 struct bio *bio, *wbio;
1643
1644 bio = r1_bio->bios[r1_bio->read_disk];
1645
1646 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1647 /* ouch - failed to read all of that. */
1648 if (!fix_sync_read_error(r1_bio))
1649 return;
1650
1651 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1652 if (process_checks(r1_bio) < 0)
1653 return;
1654 /*
1655 * schedule writes
1656 */
1657 atomic_set(&r1_bio->remaining, 1);
1658 for (i = 0; i < disks ; i++) {
1659 wbio = r1_bio->bios[i];
1660 if (wbio->bi_end_io == NULL ||
1661 (wbio->bi_end_io == end_sync_read &&
1662 (i == r1_bio->read_disk ||
1663 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1664 continue;
1665
1666 wbio->bi_rw = WRITE;
1667 wbio->bi_end_io = end_sync_write;
1668 atomic_inc(&r1_bio->remaining);
1669 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1670
1671 generic_make_request(wbio);
1672 }
1673
1674 if (atomic_dec_and_test(&r1_bio->remaining)) {
1675 /* if we're here, all write(s) have completed, so clean up */
1676 md_done_sync(mddev, r1_bio->sectors, 1);
1677 put_buf(r1_bio);
1678 }
1679}
1680
1681/*
1682 * This is a kernel thread which:
1683 *
1684 * 1. Retries failed read operations on working mirrors.
1685 * 2. Updates the raid superblock when problems encounter.
1686 * 3. Performs writes following reads for array synchronising.
1687 */
1688
1689static void fix_read_error(conf_t *conf, int read_disk,
1690 sector_t sect, int sectors)
1691{
1692 mddev_t *mddev = conf->mddev;
1693 while(sectors) {
1694 int s = sectors;
1695 int d = read_disk;
1696 int success = 0;
1697 int start;
1698 mdk_rdev_t *rdev;
1699
1700 if (s > (PAGE_SIZE>>9))
1701 s = PAGE_SIZE >> 9;
1702
1703 do {
1704 /* Note: no rcu protection needed here
1705 * as this is synchronous in the raid1d thread
1706 * which is the thread that might remove
1707 * a device. If raid1d ever becomes multi-threaded....
1708 */
1709 sector_t first_bad;
1710 int bad_sectors;
1711
1712 rdev = conf->mirrors[d].rdev;
1713 if (rdev &&
1714 test_bit(In_sync, &rdev->flags) &&
1715 is_badblock(rdev, sect, s,
1716 &first_bad, &bad_sectors) == 0 &&
1717 sync_page_io(rdev, sect, s<<9,
1718 conf->tmppage, READ, false))
1719 success = 1;
1720 else {
1721 d++;
1722 if (d == conf->raid_disks)
1723 d = 0;
1724 }
1725 } while (!success && d != read_disk);
1726
1727 if (!success) {
1728 /* Cannot read from anywhere - mark it bad */
1729 mdk_rdev_t *rdev = conf->mirrors[read_disk].rdev;
1730 if (!rdev_set_badblocks(rdev, sect, s, 0))
1731 md_error(mddev, rdev);
1732 break;
1733 }
1734 /* write it back and re-read */
1735 start = d;
1736 while (d != read_disk) {
1737 if (d==0)
1738 d = conf->raid_disks;
1739 d--;
1740 rdev = conf->mirrors[d].rdev;
1741 if (rdev &&
1742 test_bit(In_sync, &rdev->flags))
1743 r1_sync_page_io(rdev, sect, s,
1744 conf->tmppage, WRITE);
1745 }
1746 d = start;
1747 while (d != read_disk) {
1748 char b[BDEVNAME_SIZE];
1749 if (d==0)
1750 d = conf->raid_disks;
1751 d--;
1752 rdev = conf->mirrors[d].rdev;
1753 if (rdev &&
1754 test_bit(In_sync, &rdev->flags)) {
1755 if (r1_sync_page_io(rdev, sect, s,
1756 conf->tmppage, READ)) {
1757 atomic_add(s, &rdev->corrected_errors);
1758 printk(KERN_INFO
1759 "md/raid1:%s: read error corrected "
1760 "(%d sectors at %llu on %s)\n",
1761 mdname(mddev), s,
1762 (unsigned long long)(sect +
1763 rdev->data_offset),
1764 bdevname(rdev->bdev, b));
1765 }
1766 }
1767 }
1768 sectors -= s;
1769 sect += s;
1770 }
1771}
1772
1773static void bi_complete(struct bio *bio, int error)
1774{
1775 complete((struct completion *)bio->bi_private);
1776}
1777
1778static int submit_bio_wait(int rw, struct bio *bio)
1779{
1780 struct completion event;
1781 rw |= REQ_SYNC;
1782
1783 init_completion(&event);
1784 bio->bi_private = &event;
1785 bio->bi_end_io = bi_complete;
1786 submit_bio(rw, bio);
1787 wait_for_completion(&event);
1788
1789 return test_bit(BIO_UPTODATE, &bio->bi_flags);
1790}
1791
1792static int narrow_write_error(r1bio_t *r1_bio, int i)
1793{
1794 mddev_t *mddev = r1_bio->mddev;
1795 conf_t *conf = mddev->private;
1796 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1797 int vcnt, idx;
1798 struct bio_vec *vec;
1799
1800 /* bio has the data to be written to device 'i' where
1801 * we just recently had a write error.
1802 * We repeatedly clone the bio and trim down to one block,
1803 * then try the write. Where the write fails we record
1804 * a bad block.
1805 * It is conceivable that the bio doesn't exactly align with
1806 * blocks. We must handle this somehow.
1807 *
1808 * We currently own a reference on the rdev.
1809 */
1810
1811 int block_sectors;
1812 sector_t sector;
1813 int sectors;
1814 int sect_to_write = r1_bio->sectors;
1815 int ok = 1;
1816
1817 if (rdev->badblocks.shift < 0)
1818 return 0;
1819
1820 block_sectors = 1 << rdev->badblocks.shift;
1821 sector = r1_bio->sector;
1822 sectors = ((sector + block_sectors)
1823 & ~(sector_t)(block_sectors - 1))
1824 - sector;
1825
1826 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
1827 vcnt = r1_bio->behind_page_count;
1828 vec = r1_bio->behind_bvecs;
1829 idx = 0;
1830 while (vec[idx].bv_page == NULL)
1831 idx++;
1832 } else {
1833 vcnt = r1_bio->master_bio->bi_vcnt;
1834 vec = r1_bio->master_bio->bi_io_vec;
1835 idx = r1_bio->master_bio->bi_idx;
1836 }
1837 while (sect_to_write) {
1838 struct bio *wbio;
1839 if (sectors > sect_to_write)
1840 sectors = sect_to_write;
1841 /* Write at 'sector' for 'sectors'*/
1842
1843 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
1844 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
1845 wbio->bi_sector = r1_bio->sector;
1846 wbio->bi_rw = WRITE;
1847 wbio->bi_vcnt = vcnt;
1848 wbio->bi_size = r1_bio->sectors << 9;
1849 wbio->bi_idx = idx;
1850
1851 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
1852 wbio->bi_sector += rdev->data_offset;
1853 wbio->bi_bdev = rdev->bdev;
1854 if (submit_bio_wait(WRITE, wbio) == 0)
1855 /* failure! */
1856 ok = rdev_set_badblocks(rdev, sector,
1857 sectors, 0)
1858 && ok;
1859
1860 bio_put(wbio);
1861 sect_to_write -= sectors;
1862 sector += sectors;
1863 sectors = block_sectors;
1864 }
1865 return ok;
1866}
1867
1868static void handle_sync_write_finished(conf_t *conf, r1bio_t *r1_bio)
1869{
1870 int m;
1871 int s = r1_bio->sectors;
1872 for (m = 0; m < conf->raid_disks ; m++) {
1873 mdk_rdev_t *rdev = conf->mirrors[m].rdev;
1874 struct bio *bio = r1_bio->bios[m];
1875 if (bio->bi_end_io == NULL)
1876 continue;
1877 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1878 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
1879 rdev_clear_badblocks(rdev, r1_bio->sector, s);
1880 }
1881 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
1882 test_bit(R1BIO_WriteError, &r1_bio->state)) {
1883 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
1884 md_error(conf->mddev, rdev);
1885 }
1886 }
1887 put_buf(r1_bio);
1888 md_done_sync(conf->mddev, s, 1);
1889}
1890
1891static void handle_write_finished(conf_t *conf, r1bio_t *r1_bio)
1892{
1893 int m;
1894 for (m = 0; m < conf->raid_disks ; m++)
1895 if (r1_bio->bios[m] == IO_MADE_GOOD) {
1896 mdk_rdev_t *rdev = conf->mirrors[m].rdev;
1897 rdev_clear_badblocks(rdev,
1898 r1_bio->sector,
1899 r1_bio->sectors);
1900 rdev_dec_pending(rdev, conf->mddev);
1901 } else if (r1_bio->bios[m] != NULL) {
1902 /* This drive got a write error. We need to
1903 * narrow down and record precise write
1904 * errors.
1905 */
1906 if (!narrow_write_error(r1_bio, m)) {
1907 md_error(conf->mddev,
1908 conf->mirrors[m].rdev);
1909 /* an I/O failed, we can't clear the bitmap */
1910 set_bit(R1BIO_Degraded, &r1_bio->state);
1911 }
1912 rdev_dec_pending(conf->mirrors[m].rdev,
1913 conf->mddev);
1914 }
1915 if (test_bit(R1BIO_WriteError, &r1_bio->state))
1916 close_write(r1_bio);
1917 raid_end_bio_io(r1_bio);
1918}
1919
1920static void handle_read_error(conf_t *conf, r1bio_t *r1_bio)
1921{
1922 int disk;
1923 int max_sectors;
1924 mddev_t *mddev = conf->mddev;
1925 struct bio *bio;
1926 char b[BDEVNAME_SIZE];
1927 mdk_rdev_t *rdev;
1928
1929 clear_bit(R1BIO_ReadError, &r1_bio->state);
1930 /* we got a read error. Maybe the drive is bad. Maybe just
1931 * the block and we can fix it.
1932 * We freeze all other IO, and try reading the block from
1933 * other devices. When we find one, we re-write
1934 * and check it that fixes the read error.
1935 * This is all done synchronously while the array is
1936 * frozen
1937 */
1938 if (mddev->ro == 0) {
1939 freeze_array(conf);
1940 fix_read_error(conf, r1_bio->read_disk,
1941 r1_bio->sector, r1_bio->sectors);
1942 unfreeze_array(conf);
1943 } else
1944 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1945
1946 bio = r1_bio->bios[r1_bio->read_disk];
1947 bdevname(bio->bi_bdev, b);
1948read_more:
1949 disk = read_balance(conf, r1_bio, &max_sectors);
1950 if (disk == -1) {
1951 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1952 " read error for block %llu\n",
1953 mdname(mddev), b, (unsigned long long)r1_bio->sector);
1954 raid_end_bio_io(r1_bio);
1955 } else {
1956 const unsigned long do_sync
1957 = r1_bio->master_bio->bi_rw & REQ_SYNC;
1958 if (bio) {
1959 r1_bio->bios[r1_bio->read_disk] =
1960 mddev->ro ? IO_BLOCKED : NULL;
1961 bio_put(bio);
1962 }
1963 r1_bio->read_disk = disk;
1964 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
1965 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
1966 r1_bio->bios[r1_bio->read_disk] = bio;
1967 rdev = conf->mirrors[disk].rdev;
1968 printk_ratelimited(KERN_ERR
1969 "md/raid1:%s: redirecting sector %llu"
1970 " to other mirror: %s\n",
1971 mdname(mddev),
1972 (unsigned long long)r1_bio->sector,
1973 bdevname(rdev->bdev, b));
1974 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1975 bio->bi_bdev = rdev->bdev;
1976 bio->bi_end_io = raid1_end_read_request;
1977 bio->bi_rw = READ | do_sync;
1978 bio->bi_private = r1_bio;
1979 if (max_sectors < r1_bio->sectors) {
1980 /* Drat - have to split this up more */
1981 struct bio *mbio = r1_bio->master_bio;
1982 int sectors_handled = (r1_bio->sector + max_sectors
1983 - mbio->bi_sector);
1984 r1_bio->sectors = max_sectors;
1985 spin_lock_irq(&conf->device_lock);
1986 if (mbio->bi_phys_segments == 0)
1987 mbio->bi_phys_segments = 2;
1988 else
1989 mbio->bi_phys_segments++;
1990 spin_unlock_irq(&conf->device_lock);
1991 generic_make_request(bio);
1992 bio = NULL;
1993
1994 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1995
1996 r1_bio->master_bio = mbio;
1997 r1_bio->sectors = (mbio->bi_size >> 9)
1998 - sectors_handled;
1999 r1_bio->state = 0;
2000 set_bit(R1BIO_ReadError, &r1_bio->state);
2001 r1_bio->mddev = mddev;
2002 r1_bio->sector = mbio->bi_sector + sectors_handled;
2003
2004 goto read_more;
2005 } else
2006 generic_make_request(bio);
2007 }
2008}
2009
2010static void raid1d(mddev_t *mddev)
2011{
2012 r1bio_t *r1_bio;
2013 unsigned long flags;
2014 conf_t *conf = mddev->private;
2015 struct list_head *head = &conf->retry_list;
2016 struct blk_plug plug;
2017
2018 md_check_recovery(mddev);
2019
2020 blk_start_plug(&plug);
2021 for (;;) {
2022
2023 if (atomic_read(&mddev->plug_cnt) == 0)
2024 flush_pending_writes(conf);
2025
2026 spin_lock_irqsave(&conf->device_lock, flags);
2027 if (list_empty(head)) {
2028 spin_unlock_irqrestore(&conf->device_lock, flags);
2029 break;
2030 }
2031 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
2032 list_del(head->prev);
2033 conf->nr_queued--;
2034 spin_unlock_irqrestore(&conf->device_lock, flags);
2035
2036 mddev = r1_bio->mddev;
2037 conf = mddev->private;
2038 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2039 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2040 test_bit(R1BIO_WriteError, &r1_bio->state))
2041 handle_sync_write_finished(conf, r1_bio);
2042 else
2043 sync_request_write(mddev, r1_bio);
2044 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2045 test_bit(R1BIO_WriteError, &r1_bio->state))
2046 handle_write_finished(conf, r1_bio);
2047 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2048 handle_read_error(conf, r1_bio);
2049 else
2050 /* just a partial read to be scheduled from separate
2051 * context
2052 */
2053 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2054
2055 cond_resched();
2056 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2057 md_check_recovery(mddev);
2058 }
2059 blk_finish_plug(&plug);
2060}
2061
2062
2063static int init_resync(conf_t *conf)
2064{
2065 int buffs;
2066
2067 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2068 BUG_ON(conf->r1buf_pool);
2069 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2070 conf->poolinfo);
2071 if (!conf->r1buf_pool)
2072 return -ENOMEM;
2073 conf->next_resync = 0;
2074 return 0;
2075}
2076
2077/*
2078 * perform a "sync" on one "block"
2079 *
2080 * We need to make sure that no normal I/O request - particularly write
2081 * requests - conflict with active sync requests.
2082 *
2083 * This is achieved by tracking pending requests and a 'barrier' concept
2084 * that can be installed to exclude normal IO requests.
2085 */
2086
2087static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
2088{
2089 conf_t *conf = mddev->private;
2090 r1bio_t *r1_bio;
2091 struct bio *bio;
2092 sector_t max_sector, nr_sectors;
2093 int disk = -1;
2094 int i;
2095 int wonly = -1;
2096 int write_targets = 0, read_targets = 0;
2097 sector_t sync_blocks;
2098 int still_degraded = 0;
2099 int good_sectors = RESYNC_SECTORS;
2100 int min_bad = 0; /* number of sectors that are bad in all devices */
2101
2102 if (!conf->r1buf_pool)
2103 if (init_resync(conf))
2104 return 0;
2105
2106 max_sector = mddev->dev_sectors;
2107 if (sector_nr >= max_sector) {
2108 /* If we aborted, we need to abort the
2109 * sync on the 'current' bitmap chunk (there will
2110 * only be one in raid1 resync.
2111 * We can find the current addess in mddev->curr_resync
2112 */
2113 if (mddev->curr_resync < max_sector) /* aborted */
2114 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2115 &sync_blocks, 1);
2116 else /* completed sync */
2117 conf->fullsync = 0;
2118
2119 bitmap_close_sync(mddev->bitmap);
2120 close_sync(conf);
2121 return 0;
2122 }
2123
2124 if (mddev->bitmap == NULL &&
2125 mddev->recovery_cp == MaxSector &&
2126 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2127 conf->fullsync == 0) {
2128 *skipped = 1;
2129 return max_sector - sector_nr;
2130 }
2131 /* before building a request, check if we can skip these blocks..
2132 * This call the bitmap_start_sync doesn't actually record anything
2133 */
2134 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2135 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2136 /* We can skip this block, and probably several more */
2137 *skipped = 1;
2138 return sync_blocks;
2139 }
2140 /*
2141 * If there is non-resync activity waiting for a turn,
2142 * and resync is going fast enough,
2143 * then let it though before starting on this new sync request.
2144 */
2145 if (!go_faster && conf->nr_waiting)
2146 msleep_interruptible(1000);
2147
2148 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2149 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2150 raise_barrier(conf);
2151
2152 conf->next_resync = sector_nr;
2153
2154 rcu_read_lock();
2155 /*
2156 * If we get a correctably read error during resync or recovery,
2157 * we might want to read from a different device. So we
2158 * flag all drives that could conceivably be read from for READ,
2159 * and any others (which will be non-In_sync devices) for WRITE.
2160 * If a read fails, we try reading from something else for which READ
2161 * is OK.
2162 */
2163
2164 r1_bio->mddev = mddev;
2165 r1_bio->sector = sector_nr;
2166 r1_bio->state = 0;
2167 set_bit(R1BIO_IsSync, &r1_bio->state);
2168
2169 for (i=0; i < conf->raid_disks; i++) {
2170 mdk_rdev_t *rdev;
2171 bio = r1_bio->bios[i];
2172
2173 /* take from bio_init */
2174 bio->bi_next = NULL;
2175 bio->bi_flags &= ~(BIO_POOL_MASK-1);
2176 bio->bi_flags |= 1 << BIO_UPTODATE;
2177 bio->bi_comp_cpu = -1;
2178 bio->bi_rw = READ;
2179 bio->bi_vcnt = 0;
2180 bio->bi_idx = 0;
2181 bio->bi_phys_segments = 0;
2182 bio->bi_size = 0;
2183 bio->bi_end_io = NULL;
2184 bio->bi_private = NULL;
2185
2186 rdev = rcu_dereference(conf->mirrors[i].rdev);
2187 if (rdev == NULL ||
2188 test_bit(Faulty, &rdev->flags)) {
2189 still_degraded = 1;
2190 } else if (!test_bit(In_sync, &rdev->flags)) {
2191 bio->bi_rw = WRITE;
2192 bio->bi_end_io = end_sync_write;
2193 write_targets ++;
2194 } else {
2195 /* may need to read from here */
2196 sector_t first_bad = MaxSector;
2197 int bad_sectors;
2198
2199 if (is_badblock(rdev, sector_nr, good_sectors,
2200 &first_bad, &bad_sectors)) {
2201 if (first_bad > sector_nr)
2202 good_sectors = first_bad - sector_nr;
2203 else {
2204 bad_sectors -= (sector_nr - first_bad);
2205 if (min_bad == 0 ||
2206 min_bad > bad_sectors)
2207 min_bad = bad_sectors;
2208 }
2209 }
2210 if (sector_nr < first_bad) {
2211 if (test_bit(WriteMostly, &rdev->flags)) {
2212 if (wonly < 0)
2213 wonly = i;
2214 } else {
2215 if (disk < 0)
2216 disk = i;
2217 }
2218 bio->bi_rw = READ;
2219 bio->bi_end_io = end_sync_read;
2220 read_targets++;
2221 }
2222 }
2223 if (bio->bi_end_io) {
2224 atomic_inc(&rdev->nr_pending);
2225 bio->bi_sector = sector_nr + rdev->data_offset;
2226 bio->bi_bdev = rdev->bdev;
2227 bio->bi_private = r1_bio;
2228 }
2229 }
2230 rcu_read_unlock();
2231 if (disk < 0)
2232 disk = wonly;
2233 r1_bio->read_disk = disk;
2234
2235 if (read_targets == 0 && min_bad > 0) {
2236 /* These sectors are bad on all InSync devices, so we
2237 * need to mark them bad on all write targets
2238 */
2239 int ok = 1;
2240 for (i = 0 ; i < conf->raid_disks ; i++)
2241 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2242 mdk_rdev_t *rdev =
2243 rcu_dereference(conf->mirrors[i].rdev);
2244 ok = rdev_set_badblocks(rdev, sector_nr,
2245 min_bad, 0
2246 ) && ok;
2247 }
2248 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2249 *skipped = 1;
2250 put_buf(r1_bio);
2251
2252 if (!ok) {
2253 /* Cannot record the badblocks, so need to
2254 * abort the resync.
2255 * If there are multiple read targets, could just
2256 * fail the really bad ones ???
2257 */
2258 conf->recovery_disabled = mddev->recovery_disabled;
2259 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2260 return 0;
2261 } else
2262 return min_bad;
2263
2264 }
2265 if (min_bad > 0 && min_bad < good_sectors) {
2266 /* only resync enough to reach the next bad->good
2267 * transition */
2268 good_sectors = min_bad;
2269 }
2270
2271 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2272 /* extra read targets are also write targets */
2273 write_targets += read_targets-1;
2274
2275 if (write_targets == 0 || read_targets == 0) {
2276 /* There is nowhere to write, so all non-sync
2277 * drives must be failed - so we are finished
2278 */
2279 sector_t rv = max_sector - sector_nr;
2280 *skipped = 1;
2281 put_buf(r1_bio);
2282 return rv;
2283 }
2284
2285 if (max_sector > mddev->resync_max)
2286 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2287 if (max_sector > sector_nr + good_sectors)
2288 max_sector = sector_nr + good_sectors;
2289 nr_sectors = 0;
2290 sync_blocks = 0;
2291 do {
2292 struct page *page;
2293 int len = PAGE_SIZE;
2294 if (sector_nr + (len>>9) > max_sector)
2295 len = (max_sector - sector_nr) << 9;
2296 if (len == 0)
2297 break;
2298 if (sync_blocks == 0) {
2299 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2300 &sync_blocks, still_degraded) &&
2301 !conf->fullsync &&
2302 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2303 break;
2304 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2305 if ((len >> 9) > sync_blocks)
2306 len = sync_blocks<<9;
2307 }
2308
2309 for (i=0 ; i < conf->raid_disks; i++) {
2310 bio = r1_bio->bios[i];
2311 if (bio->bi_end_io) {
2312 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2313 if (bio_add_page(bio, page, len, 0) == 0) {
2314 /* stop here */
2315 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2316 while (i > 0) {
2317 i--;
2318 bio = r1_bio->bios[i];
2319 if (bio->bi_end_io==NULL)
2320 continue;
2321 /* remove last page from this bio */
2322 bio->bi_vcnt--;
2323 bio->bi_size -= len;
2324 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2325 }
2326 goto bio_full;
2327 }
2328 }
2329 }
2330 nr_sectors += len>>9;
2331 sector_nr += len>>9;
2332 sync_blocks -= (len>>9);
2333 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2334 bio_full:
2335 r1_bio->sectors = nr_sectors;
2336
2337 /* For a user-requested sync, we read all readable devices and do a
2338 * compare
2339 */
2340 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2341 atomic_set(&r1_bio->remaining, read_targets);
2342 for (i=0; i<conf->raid_disks; i++) {
2343 bio = r1_bio->bios[i];
2344 if (bio->bi_end_io == end_sync_read) {
2345 md_sync_acct(bio->bi_bdev, nr_sectors);
2346 generic_make_request(bio);
2347 }
2348 }
2349 } else {
2350 atomic_set(&r1_bio->remaining, 1);
2351 bio = r1_bio->bios[r1_bio->read_disk];
2352 md_sync_acct(bio->bi_bdev, nr_sectors);
2353 generic_make_request(bio);
2354
2355 }
2356 return nr_sectors;
2357}
2358
2359static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2360{
2361 if (sectors)
2362 return sectors;
2363
2364 return mddev->dev_sectors;
2365}
2366
2367static conf_t *setup_conf(mddev_t *mddev)
2368{
2369 conf_t *conf;
2370 int i;
2371 mirror_info_t *disk;
2372 mdk_rdev_t *rdev;
2373 int err = -ENOMEM;
2374
2375 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2376 if (!conf)
2377 goto abort;
2378
2379 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2380 GFP_KERNEL);
2381 if (!conf->mirrors)
2382 goto abort;
2383
2384 conf->tmppage = alloc_page(GFP_KERNEL);
2385 if (!conf->tmppage)
2386 goto abort;
2387
2388 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2389 if (!conf->poolinfo)
2390 goto abort;
2391 conf->poolinfo->raid_disks = mddev->raid_disks;
2392 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2393 r1bio_pool_free,
2394 conf->poolinfo);
2395 if (!conf->r1bio_pool)
2396 goto abort;
2397
2398 conf->poolinfo->mddev = mddev;
2399
2400 spin_lock_init(&conf->device_lock);
2401 list_for_each_entry(rdev, &mddev->disks, same_set) {
2402 int disk_idx = rdev->raid_disk;
2403 if (disk_idx >= mddev->raid_disks
2404 || disk_idx < 0)
2405 continue;
2406 disk = conf->mirrors + disk_idx;
2407
2408 disk->rdev = rdev;
2409
2410 disk->head_position = 0;
2411 }
2412 conf->raid_disks = mddev->raid_disks;
2413 conf->mddev = mddev;
2414 INIT_LIST_HEAD(&conf->retry_list);
2415
2416 spin_lock_init(&conf->resync_lock);
2417 init_waitqueue_head(&conf->wait_barrier);
2418
2419 bio_list_init(&conf->pending_bio_list);
2420
2421 conf->last_used = -1;
2422 for (i = 0; i < conf->raid_disks; i++) {
2423
2424 disk = conf->mirrors + i;
2425
2426 if (!disk->rdev ||
2427 !test_bit(In_sync, &disk->rdev->flags)) {
2428 disk->head_position = 0;
2429 if (disk->rdev)
2430 conf->fullsync = 1;
2431 } else if (conf->last_used < 0)
2432 /*
2433 * The first working device is used as a
2434 * starting point to read balancing.
2435 */
2436 conf->last_used = i;
2437 }
2438
2439 err = -EIO;
2440 if (conf->last_used < 0) {
2441 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2442 mdname(mddev));
2443 goto abort;
2444 }
2445 err = -ENOMEM;
2446 conf->thread = md_register_thread(raid1d, mddev, NULL);
2447 if (!conf->thread) {
2448 printk(KERN_ERR
2449 "md/raid1:%s: couldn't allocate thread\n",
2450 mdname(mddev));
2451 goto abort;
2452 }
2453
2454 return conf;
2455
2456 abort:
2457 if (conf) {
2458 if (conf->r1bio_pool)
2459 mempool_destroy(conf->r1bio_pool);
2460 kfree(conf->mirrors);
2461 safe_put_page(conf->tmppage);
2462 kfree(conf->poolinfo);
2463 kfree(conf);
2464 }
2465 return ERR_PTR(err);
2466}
2467
2468static int run(mddev_t *mddev)
2469{
2470 conf_t *conf;
2471 int i;
2472 mdk_rdev_t *rdev;
2473
2474 if (mddev->level != 1) {
2475 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2476 mdname(mddev), mddev->level);
2477 return -EIO;
2478 }
2479 if (mddev->reshape_position != MaxSector) {
2480 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2481 mdname(mddev));
2482 return -EIO;
2483 }
2484 /*
2485 * copy the already verified devices into our private RAID1
2486 * bookkeeping area. [whatever we allocate in run(),
2487 * should be freed in stop()]
2488 */
2489 if (mddev->private == NULL)
2490 conf = setup_conf(mddev);
2491 else
2492 conf = mddev->private;
2493
2494 if (IS_ERR(conf))
2495 return PTR_ERR(conf);
2496
2497 list_for_each_entry(rdev, &mddev->disks, same_set) {
2498 if (!mddev->gendisk)
2499 continue;
2500 disk_stack_limits(mddev->gendisk, rdev->bdev,
2501 rdev->data_offset << 9);
2502 /* as we don't honour merge_bvec_fn, we must never risk
2503 * violating it, so limit ->max_segments to 1 lying within
2504 * a single page, as a one page request is never in violation.
2505 */
2506 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2507 blk_queue_max_segments(mddev->queue, 1);
2508 blk_queue_segment_boundary(mddev->queue,
2509 PAGE_CACHE_SIZE - 1);
2510 }
2511 }
2512
2513 mddev->degraded = 0;
2514 for (i=0; i < conf->raid_disks; i++)
2515 if (conf->mirrors[i].rdev == NULL ||
2516 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2517 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2518 mddev->degraded++;
2519
2520 if (conf->raid_disks - mddev->degraded == 1)
2521 mddev->recovery_cp = MaxSector;
2522
2523 if (mddev->recovery_cp != MaxSector)
2524 printk(KERN_NOTICE "md/raid1:%s: not clean"
2525 " -- starting background reconstruction\n",
2526 mdname(mddev));
2527 printk(KERN_INFO
2528 "md/raid1:%s: active with %d out of %d mirrors\n",
2529 mdname(mddev), mddev->raid_disks - mddev->degraded,
2530 mddev->raid_disks);
2531
2532 /*
2533 * Ok, everything is just fine now
2534 */
2535 mddev->thread = conf->thread;
2536 conf->thread = NULL;
2537 mddev->private = conf;
2538
2539 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2540
2541 if (mddev->queue) {
2542 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2543 mddev->queue->backing_dev_info.congested_data = mddev;
2544 }
2545 return md_integrity_register(mddev);
2546}
2547
2548static int stop(mddev_t *mddev)
2549{
2550 conf_t *conf = mddev->private;
2551 struct bitmap *bitmap = mddev->bitmap;
2552
2553 /* wait for behind writes to complete */
2554 if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2555 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2556 mdname(mddev));
2557 /* need to kick something here to make sure I/O goes? */
2558 wait_event(bitmap->behind_wait,
2559 atomic_read(&bitmap->behind_writes) == 0);
2560 }
2561
2562 raise_barrier(conf);
2563 lower_barrier(conf);
2564
2565 md_unregister_thread(&mddev->thread);
2566 if (conf->r1bio_pool)
2567 mempool_destroy(conf->r1bio_pool);
2568 kfree(conf->mirrors);
2569 kfree(conf->poolinfo);
2570 kfree(conf);
2571 mddev->private = NULL;
2572 return 0;
2573}
2574
2575static int raid1_resize(mddev_t *mddev, sector_t sectors)
2576{
2577 /* no resync is happening, and there is enough space
2578 * on all devices, so we can resize.
2579 * We need to make sure resync covers any new space.
2580 * If the array is shrinking we should possibly wait until
2581 * any io in the removed space completes, but it hardly seems
2582 * worth it.
2583 */
2584 md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2585 if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2586 return -EINVAL;
2587 set_capacity(mddev->gendisk, mddev->array_sectors);
2588 revalidate_disk(mddev->gendisk);
2589 if (sectors > mddev->dev_sectors &&
2590 mddev->recovery_cp > mddev->dev_sectors) {
2591 mddev->recovery_cp = mddev->dev_sectors;
2592 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2593 }
2594 mddev->dev_sectors = sectors;
2595 mddev->resync_max_sectors = sectors;
2596 return 0;
2597}
2598
2599static int raid1_reshape(mddev_t *mddev)
2600{
2601 /* We need to:
2602 * 1/ resize the r1bio_pool
2603 * 2/ resize conf->mirrors
2604 *
2605 * We allocate a new r1bio_pool if we can.
2606 * Then raise a device barrier and wait until all IO stops.
2607 * Then resize conf->mirrors and swap in the new r1bio pool.
2608 *
2609 * At the same time, we "pack" the devices so that all the missing
2610 * devices have the higher raid_disk numbers.
2611 */
2612 mempool_t *newpool, *oldpool;
2613 struct pool_info *newpoolinfo;
2614 mirror_info_t *newmirrors;
2615 conf_t *conf = mddev->private;
2616 int cnt, raid_disks;
2617 unsigned long flags;
2618 int d, d2, err;
2619
2620 /* Cannot change chunk_size, layout, or level */
2621 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2622 mddev->layout != mddev->new_layout ||
2623 mddev->level != mddev->new_level) {
2624 mddev->new_chunk_sectors = mddev->chunk_sectors;
2625 mddev->new_layout = mddev->layout;
2626 mddev->new_level = mddev->level;
2627 return -EINVAL;
2628 }
2629
2630 err = md_allow_write(mddev);
2631 if (err)
2632 return err;
2633
2634 raid_disks = mddev->raid_disks + mddev->delta_disks;
2635
2636 if (raid_disks < conf->raid_disks) {
2637 cnt=0;
2638 for (d= 0; d < conf->raid_disks; d++)
2639 if (conf->mirrors[d].rdev)
2640 cnt++;
2641 if (cnt > raid_disks)
2642 return -EBUSY;
2643 }
2644
2645 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2646 if (!newpoolinfo)
2647 return -ENOMEM;
2648 newpoolinfo->mddev = mddev;
2649 newpoolinfo->raid_disks = raid_disks;
2650
2651 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2652 r1bio_pool_free, newpoolinfo);
2653 if (!newpool) {
2654 kfree(newpoolinfo);
2655 return -ENOMEM;
2656 }
2657 newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2658 if (!newmirrors) {
2659 kfree(newpoolinfo);
2660 mempool_destroy(newpool);
2661 return -ENOMEM;
2662 }
2663
2664 raise_barrier(conf);
2665
2666 /* ok, everything is stopped */
2667 oldpool = conf->r1bio_pool;
2668 conf->r1bio_pool = newpool;
2669
2670 for (d = d2 = 0; d < conf->raid_disks; d++) {
2671 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2672 if (rdev && rdev->raid_disk != d2) {
2673 sysfs_unlink_rdev(mddev, rdev);
2674 rdev->raid_disk = d2;
2675 sysfs_unlink_rdev(mddev, rdev);
2676 if (sysfs_link_rdev(mddev, rdev))
2677 printk(KERN_WARNING
2678 "md/raid1:%s: cannot register rd%d\n",
2679 mdname(mddev), rdev->raid_disk);
2680 }
2681 if (rdev)
2682 newmirrors[d2++].rdev = rdev;
2683 }
2684 kfree(conf->mirrors);
2685 conf->mirrors = newmirrors;
2686 kfree(conf->poolinfo);
2687 conf->poolinfo = newpoolinfo;
2688
2689 spin_lock_irqsave(&conf->device_lock, flags);
2690 mddev->degraded += (raid_disks - conf->raid_disks);
2691 spin_unlock_irqrestore(&conf->device_lock, flags);
2692 conf->raid_disks = mddev->raid_disks = raid_disks;
2693 mddev->delta_disks = 0;
2694
2695 conf->last_used = 0; /* just make sure it is in-range */
2696 lower_barrier(conf);
2697
2698 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2699 md_wakeup_thread(mddev->thread);
2700
2701 mempool_destroy(oldpool);
2702 return 0;
2703}
2704
2705static void raid1_quiesce(mddev_t *mddev, int state)
2706{
2707 conf_t *conf = mddev->private;
2708
2709 switch(state) {
2710 case 2: /* wake for suspend */
2711 wake_up(&conf->wait_barrier);
2712 break;
2713 case 1:
2714 raise_barrier(conf);
2715 break;
2716 case 0:
2717 lower_barrier(conf);
2718 break;
2719 }
2720}
2721
2722static void *raid1_takeover(mddev_t *mddev)
2723{
2724 /* raid1 can take over:
2725 * raid5 with 2 devices, any layout or chunk size
2726 */
2727 if (mddev->level == 5 && mddev->raid_disks == 2) {
2728 conf_t *conf;
2729 mddev->new_level = 1;
2730 mddev->new_layout = 0;
2731 mddev->new_chunk_sectors = 0;
2732 conf = setup_conf(mddev);
2733 if (!IS_ERR(conf))
2734 conf->barrier = 1;
2735 return conf;
2736 }
2737 return ERR_PTR(-EINVAL);
2738}
2739
2740static struct mdk_personality raid1_personality =
2741{
2742 .name = "raid1",
2743 .level = 1,
2744 .owner = THIS_MODULE,
2745 .make_request = make_request,
2746 .run = run,
2747 .stop = stop,
2748 .status = status,
2749 .error_handler = error,
2750 .hot_add_disk = raid1_add_disk,
2751 .hot_remove_disk= raid1_remove_disk,
2752 .spare_active = raid1_spare_active,
2753 .sync_request = sync_request,
2754 .resize = raid1_resize,
2755 .size = raid1_size,
2756 .check_reshape = raid1_reshape,
2757 .quiesce = raid1_quiesce,
2758 .takeover = raid1_takeover,
2759};
2760
2761static int __init raid_init(void)
2762{
2763 return register_md_personality(&raid1_personality);
2764}
2765
2766static void raid_exit(void)
2767{
2768 unregister_md_personality(&raid1_personality);
2769}
2770
2771module_init(raid_init);
2772module_exit(raid_exit);
2773MODULE_LICENSE("GPL");
2774MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2775MODULE_ALIAS("md-personality-3"); /* RAID1 */
2776MODULE_ALIAS("md-raid1");
2777MODULE_ALIAS("md-level-1");
1/*
2 * raid1.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5 *
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7 *
8 * RAID-1 management functions.
9 *
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11 *
12 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14 *
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
17 *
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
20 *
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
23 *
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)
27 * any later version.
28 *
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.
32 */
33
34#include <linux/slab.h>
35#include <linux/delay.h>
36#include <linux/blkdev.h>
37#include <linux/module.h>
38#include <linux/seq_file.h>
39#include <linux/ratelimit.h>
40
41#include <trace/events/block.h>
42
43#include "md.h"
44#include "raid1.h"
45#include "md-bitmap.h"
46
47#define UNSUPPORTED_MDDEV_FLAGS \
48 ((1L << MD_HAS_JOURNAL) | \
49 (1L << MD_JOURNAL_CLEAN) | \
50 (1L << MD_HAS_PPL) | \
51 (1L << MD_HAS_MULTIPLE_PPLS))
52
53/*
54 * Number of guaranteed r1bios in case of extreme VM load:
55 */
56#define NR_RAID1_BIOS 256
57
58/* when we get a read error on a read-only array, we redirect to another
59 * device without failing the first device, or trying to over-write to
60 * correct the read error. To keep track of bad blocks on a per-bio
61 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
62 */
63#define IO_BLOCKED ((struct bio *)1)
64/* When we successfully write to a known bad-block, we need to remove the
65 * bad-block marking which must be done from process context. So we record
66 * the success by setting devs[n].bio to IO_MADE_GOOD
67 */
68#define IO_MADE_GOOD ((struct bio *)2)
69
70#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
71
72/* When there are this many requests queue to be written by
73 * the raid1 thread, we become 'congested' to provide back-pressure
74 * for writeback.
75 */
76static int max_queued_requests = 1024;
77
78static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
79static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
80
81#define raid1_log(md, fmt, args...) \
82 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid1 " fmt, ##args); } while (0)
83
84#include "raid1-10.c"
85
86/*
87 * for resync bio, r1bio pointer can be retrieved from the per-bio
88 * 'struct resync_pages'.
89 */
90static inline struct r1bio *get_resync_r1bio(struct bio *bio)
91{
92 return get_resync_pages(bio)->raid_bio;
93}
94
95static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
96{
97 struct pool_info *pi = data;
98 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
99
100 /* allocate a r1bio with room for raid_disks entries in the bios array */
101 return kzalloc(size, gfp_flags);
102}
103
104static void r1bio_pool_free(void *r1_bio, void *data)
105{
106 kfree(r1_bio);
107}
108
109#define RESYNC_DEPTH 32
110#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
111#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
112#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
113#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
114#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
115
116static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
117{
118 struct pool_info *pi = data;
119 struct r1bio *r1_bio;
120 struct bio *bio;
121 int need_pages;
122 int j;
123 struct resync_pages *rps;
124
125 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
126 if (!r1_bio)
127 return NULL;
128
129 rps = kmalloc(sizeof(struct resync_pages) * pi->raid_disks,
130 gfp_flags);
131 if (!rps)
132 goto out_free_r1bio;
133
134 /*
135 * Allocate bios : 1 for reading, n-1 for writing
136 */
137 for (j = pi->raid_disks ; j-- ; ) {
138 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
139 if (!bio)
140 goto out_free_bio;
141 r1_bio->bios[j] = bio;
142 }
143 /*
144 * Allocate RESYNC_PAGES data pages and attach them to
145 * the first bio.
146 * If this is a user-requested check/repair, allocate
147 * RESYNC_PAGES for each bio.
148 */
149 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
150 need_pages = pi->raid_disks;
151 else
152 need_pages = 1;
153 for (j = 0; j < pi->raid_disks; j++) {
154 struct resync_pages *rp = &rps[j];
155
156 bio = r1_bio->bios[j];
157
158 if (j < need_pages) {
159 if (resync_alloc_pages(rp, gfp_flags))
160 goto out_free_pages;
161 } else {
162 memcpy(rp, &rps[0], sizeof(*rp));
163 resync_get_all_pages(rp);
164 }
165
166 rp->raid_bio = r1_bio;
167 bio->bi_private = rp;
168 }
169
170 r1_bio->master_bio = NULL;
171
172 return r1_bio;
173
174out_free_pages:
175 while (--j >= 0)
176 resync_free_pages(&rps[j]);
177
178out_free_bio:
179 while (++j < pi->raid_disks)
180 bio_put(r1_bio->bios[j]);
181 kfree(rps);
182
183out_free_r1bio:
184 r1bio_pool_free(r1_bio, data);
185 return NULL;
186}
187
188static void r1buf_pool_free(void *__r1_bio, void *data)
189{
190 struct pool_info *pi = data;
191 int i;
192 struct r1bio *r1bio = __r1_bio;
193 struct resync_pages *rp = NULL;
194
195 for (i = pi->raid_disks; i--; ) {
196 rp = get_resync_pages(r1bio->bios[i]);
197 resync_free_pages(rp);
198 bio_put(r1bio->bios[i]);
199 }
200
201 /* resync pages array stored in the 1st bio's .bi_private */
202 kfree(rp);
203
204 r1bio_pool_free(r1bio, data);
205}
206
207static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
208{
209 int i;
210
211 for (i = 0; i < conf->raid_disks * 2; i++) {
212 struct bio **bio = r1_bio->bios + i;
213 if (!BIO_SPECIAL(*bio))
214 bio_put(*bio);
215 *bio = NULL;
216 }
217}
218
219static void free_r1bio(struct r1bio *r1_bio)
220{
221 struct r1conf *conf = r1_bio->mddev->private;
222
223 put_all_bios(conf, r1_bio);
224 mempool_free(r1_bio, conf->r1bio_pool);
225}
226
227static void put_buf(struct r1bio *r1_bio)
228{
229 struct r1conf *conf = r1_bio->mddev->private;
230 sector_t sect = r1_bio->sector;
231 int i;
232
233 for (i = 0; i < conf->raid_disks * 2; i++) {
234 struct bio *bio = r1_bio->bios[i];
235 if (bio->bi_end_io)
236 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
237 }
238
239 mempool_free(r1_bio, conf->r1buf_pool);
240
241 lower_barrier(conf, sect);
242}
243
244static void reschedule_retry(struct r1bio *r1_bio)
245{
246 unsigned long flags;
247 struct mddev *mddev = r1_bio->mddev;
248 struct r1conf *conf = mddev->private;
249 int idx;
250
251 idx = sector_to_idx(r1_bio->sector);
252 spin_lock_irqsave(&conf->device_lock, flags);
253 list_add(&r1_bio->retry_list, &conf->retry_list);
254 atomic_inc(&conf->nr_queued[idx]);
255 spin_unlock_irqrestore(&conf->device_lock, flags);
256
257 wake_up(&conf->wait_barrier);
258 md_wakeup_thread(mddev->thread);
259}
260
261/*
262 * raid_end_bio_io() is called when we have finished servicing a mirrored
263 * operation and are ready to return a success/failure code to the buffer
264 * cache layer.
265 */
266static void call_bio_endio(struct r1bio *r1_bio)
267{
268 struct bio *bio = r1_bio->master_bio;
269 struct r1conf *conf = r1_bio->mddev->private;
270
271 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
272 bio->bi_status = BLK_STS_IOERR;
273
274 bio_endio(bio);
275 /*
276 * Wake up any possible resync thread that waits for the device
277 * to go idle.
278 */
279 allow_barrier(conf, r1_bio->sector);
280}
281
282static void raid_end_bio_io(struct r1bio *r1_bio)
283{
284 struct bio *bio = r1_bio->master_bio;
285
286 /* if nobody has done the final endio yet, do it now */
287 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
288 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
289 (bio_data_dir(bio) == WRITE) ? "write" : "read",
290 (unsigned long long) bio->bi_iter.bi_sector,
291 (unsigned long long) bio_end_sector(bio) - 1);
292
293 call_bio_endio(r1_bio);
294 }
295 free_r1bio(r1_bio);
296}
297
298/*
299 * Update disk head position estimator based on IRQ completion info.
300 */
301static inline void update_head_pos(int disk, struct r1bio *r1_bio)
302{
303 struct r1conf *conf = r1_bio->mddev->private;
304
305 conf->mirrors[disk].head_position =
306 r1_bio->sector + (r1_bio->sectors);
307}
308
309/*
310 * Find the disk number which triggered given bio
311 */
312static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
313{
314 int mirror;
315 struct r1conf *conf = r1_bio->mddev->private;
316 int raid_disks = conf->raid_disks;
317
318 for (mirror = 0; mirror < raid_disks * 2; mirror++)
319 if (r1_bio->bios[mirror] == bio)
320 break;
321
322 BUG_ON(mirror == raid_disks * 2);
323 update_head_pos(mirror, r1_bio);
324
325 return mirror;
326}
327
328static void raid1_end_read_request(struct bio *bio)
329{
330 int uptodate = !bio->bi_status;
331 struct r1bio *r1_bio = bio->bi_private;
332 struct r1conf *conf = r1_bio->mddev->private;
333 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
334
335 /*
336 * this branch is our 'one mirror IO has finished' event handler:
337 */
338 update_head_pos(r1_bio->read_disk, r1_bio);
339
340 if (uptodate)
341 set_bit(R1BIO_Uptodate, &r1_bio->state);
342 else if (test_bit(FailFast, &rdev->flags) &&
343 test_bit(R1BIO_FailFast, &r1_bio->state))
344 /* This was a fail-fast read so we definitely
345 * want to retry */
346 ;
347 else {
348 /* If all other devices have failed, we want to return
349 * the error upwards rather than fail the last device.
350 * Here we redefine "uptodate" to mean "Don't want to retry"
351 */
352 unsigned long flags;
353 spin_lock_irqsave(&conf->device_lock, flags);
354 if (r1_bio->mddev->degraded == conf->raid_disks ||
355 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
356 test_bit(In_sync, &rdev->flags)))
357 uptodate = 1;
358 spin_unlock_irqrestore(&conf->device_lock, flags);
359 }
360
361 if (uptodate) {
362 raid_end_bio_io(r1_bio);
363 rdev_dec_pending(rdev, conf->mddev);
364 } else {
365 /*
366 * oops, read error:
367 */
368 char b[BDEVNAME_SIZE];
369 pr_err_ratelimited("md/raid1:%s: %s: rescheduling sector %llu\n",
370 mdname(conf->mddev),
371 bdevname(rdev->bdev, b),
372 (unsigned long long)r1_bio->sector);
373 set_bit(R1BIO_ReadError, &r1_bio->state);
374 reschedule_retry(r1_bio);
375 /* don't drop the reference on read_disk yet */
376 }
377}
378
379static void close_write(struct r1bio *r1_bio)
380{
381 /* it really is the end of this request */
382 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
383 bio_free_pages(r1_bio->behind_master_bio);
384 bio_put(r1_bio->behind_master_bio);
385 r1_bio->behind_master_bio = NULL;
386 }
387 /* clear the bitmap if all writes complete successfully */
388 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
389 r1_bio->sectors,
390 !test_bit(R1BIO_Degraded, &r1_bio->state),
391 test_bit(R1BIO_BehindIO, &r1_bio->state));
392 md_write_end(r1_bio->mddev);
393}
394
395static void r1_bio_write_done(struct r1bio *r1_bio)
396{
397 if (!atomic_dec_and_test(&r1_bio->remaining))
398 return;
399
400 if (test_bit(R1BIO_WriteError, &r1_bio->state))
401 reschedule_retry(r1_bio);
402 else {
403 close_write(r1_bio);
404 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
405 reschedule_retry(r1_bio);
406 else
407 raid_end_bio_io(r1_bio);
408 }
409}
410
411static void raid1_end_write_request(struct bio *bio)
412{
413 struct r1bio *r1_bio = bio->bi_private;
414 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
415 struct r1conf *conf = r1_bio->mddev->private;
416 struct bio *to_put = NULL;
417 int mirror = find_bio_disk(r1_bio, bio);
418 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
419 bool discard_error;
420
421 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
422
423 /*
424 * 'one mirror IO has finished' event handler:
425 */
426 if (bio->bi_status && !discard_error) {
427 set_bit(WriteErrorSeen, &rdev->flags);
428 if (!test_and_set_bit(WantReplacement, &rdev->flags))
429 set_bit(MD_RECOVERY_NEEDED, &
430 conf->mddev->recovery);
431
432 if (test_bit(FailFast, &rdev->flags) &&
433 (bio->bi_opf & MD_FAILFAST) &&
434 /* We never try FailFast to WriteMostly devices */
435 !test_bit(WriteMostly, &rdev->flags)) {
436 md_error(r1_bio->mddev, rdev);
437 if (!test_bit(Faulty, &rdev->flags))
438 /* This is the only remaining device,
439 * We need to retry the write without
440 * FailFast
441 */
442 set_bit(R1BIO_WriteError, &r1_bio->state);
443 else {
444 /* Finished with this branch */
445 r1_bio->bios[mirror] = NULL;
446 to_put = bio;
447 }
448 } else
449 set_bit(R1BIO_WriteError, &r1_bio->state);
450 } else {
451 /*
452 * Set R1BIO_Uptodate in our master bio, so that we
453 * will return a good error code for to the higher
454 * levels even if IO on some other mirrored buffer
455 * fails.
456 *
457 * The 'master' represents the composite IO operation
458 * to user-side. So if something waits for IO, then it
459 * will wait for the 'master' bio.
460 */
461 sector_t first_bad;
462 int bad_sectors;
463
464 r1_bio->bios[mirror] = NULL;
465 to_put = bio;
466 /*
467 * Do not set R1BIO_Uptodate if the current device is
468 * rebuilding or Faulty. This is because we cannot use
469 * such device for properly reading the data back (we could
470 * potentially use it, if the current write would have felt
471 * before rdev->recovery_offset, but for simplicity we don't
472 * check this here.
473 */
474 if (test_bit(In_sync, &rdev->flags) &&
475 !test_bit(Faulty, &rdev->flags))
476 set_bit(R1BIO_Uptodate, &r1_bio->state);
477
478 /* Maybe we can clear some bad blocks. */
479 if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
480 &first_bad, &bad_sectors) && !discard_error) {
481 r1_bio->bios[mirror] = IO_MADE_GOOD;
482 set_bit(R1BIO_MadeGood, &r1_bio->state);
483 }
484 }
485
486 if (behind) {
487 if (test_bit(WriteMostly, &rdev->flags))
488 atomic_dec(&r1_bio->behind_remaining);
489
490 /*
491 * In behind mode, we ACK the master bio once the I/O
492 * has safely reached all non-writemostly
493 * disks. Setting the Returned bit ensures that this
494 * gets done only once -- we don't ever want to return
495 * -EIO here, instead we'll wait
496 */
497 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
498 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
499 /* Maybe we can return now */
500 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
501 struct bio *mbio = r1_bio->master_bio;
502 pr_debug("raid1: behind end write sectors"
503 " %llu-%llu\n",
504 (unsigned long long) mbio->bi_iter.bi_sector,
505 (unsigned long long) bio_end_sector(mbio) - 1);
506 call_bio_endio(r1_bio);
507 }
508 }
509 }
510 if (r1_bio->bios[mirror] == NULL)
511 rdev_dec_pending(rdev, conf->mddev);
512
513 /*
514 * Let's see if all mirrored write operations have finished
515 * already.
516 */
517 r1_bio_write_done(r1_bio);
518
519 if (to_put)
520 bio_put(to_put);
521}
522
523static sector_t align_to_barrier_unit_end(sector_t start_sector,
524 sector_t sectors)
525{
526 sector_t len;
527
528 WARN_ON(sectors == 0);
529 /*
530 * len is the number of sectors from start_sector to end of the
531 * barrier unit which start_sector belongs to.
532 */
533 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
534 start_sector;
535
536 if (len > sectors)
537 len = sectors;
538
539 return len;
540}
541
542/*
543 * This routine returns the disk from which the requested read should
544 * be done. There is a per-array 'next expected sequential IO' sector
545 * number - if this matches on the next IO then we use the last disk.
546 * There is also a per-disk 'last know head position' sector that is
547 * maintained from IRQ contexts, both the normal and the resync IO
548 * completion handlers update this position correctly. If there is no
549 * perfect sequential match then we pick the disk whose head is closest.
550 *
551 * If there are 2 mirrors in the same 2 devices, performance degrades
552 * because position is mirror, not device based.
553 *
554 * The rdev for the device selected will have nr_pending incremented.
555 */
556static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
557{
558 const sector_t this_sector = r1_bio->sector;
559 int sectors;
560 int best_good_sectors;
561 int best_disk, best_dist_disk, best_pending_disk;
562 int has_nonrot_disk;
563 int disk;
564 sector_t best_dist;
565 unsigned int min_pending;
566 struct md_rdev *rdev;
567 int choose_first;
568 int choose_next_idle;
569
570 rcu_read_lock();
571 /*
572 * Check if we can balance. We can balance on the whole
573 * device if no resync is going on, or below the resync window.
574 * We take the first readable disk when above the resync window.
575 */
576 retry:
577 sectors = r1_bio->sectors;
578 best_disk = -1;
579 best_dist_disk = -1;
580 best_dist = MaxSector;
581 best_pending_disk = -1;
582 min_pending = UINT_MAX;
583 best_good_sectors = 0;
584 has_nonrot_disk = 0;
585 choose_next_idle = 0;
586 clear_bit(R1BIO_FailFast, &r1_bio->state);
587
588 if ((conf->mddev->recovery_cp < this_sector + sectors) ||
589 (mddev_is_clustered(conf->mddev) &&
590 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
591 this_sector + sectors)))
592 choose_first = 1;
593 else
594 choose_first = 0;
595
596 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
597 sector_t dist;
598 sector_t first_bad;
599 int bad_sectors;
600 unsigned int pending;
601 bool nonrot;
602
603 rdev = rcu_dereference(conf->mirrors[disk].rdev);
604 if (r1_bio->bios[disk] == IO_BLOCKED
605 || rdev == NULL
606 || test_bit(Faulty, &rdev->flags))
607 continue;
608 if (!test_bit(In_sync, &rdev->flags) &&
609 rdev->recovery_offset < this_sector + sectors)
610 continue;
611 if (test_bit(WriteMostly, &rdev->flags)) {
612 /* Don't balance among write-mostly, just
613 * use the first as a last resort */
614 if (best_dist_disk < 0) {
615 if (is_badblock(rdev, this_sector, sectors,
616 &first_bad, &bad_sectors)) {
617 if (first_bad <= this_sector)
618 /* Cannot use this */
619 continue;
620 best_good_sectors = first_bad - this_sector;
621 } else
622 best_good_sectors = sectors;
623 best_dist_disk = disk;
624 best_pending_disk = disk;
625 }
626 continue;
627 }
628 /* This is a reasonable device to use. It might
629 * even be best.
630 */
631 if (is_badblock(rdev, this_sector, sectors,
632 &first_bad, &bad_sectors)) {
633 if (best_dist < MaxSector)
634 /* already have a better device */
635 continue;
636 if (first_bad <= this_sector) {
637 /* cannot read here. If this is the 'primary'
638 * device, then we must not read beyond
639 * bad_sectors from another device..
640 */
641 bad_sectors -= (this_sector - first_bad);
642 if (choose_first && sectors > bad_sectors)
643 sectors = bad_sectors;
644 if (best_good_sectors > sectors)
645 best_good_sectors = sectors;
646
647 } else {
648 sector_t good_sectors = first_bad - this_sector;
649 if (good_sectors > best_good_sectors) {
650 best_good_sectors = good_sectors;
651 best_disk = disk;
652 }
653 if (choose_first)
654 break;
655 }
656 continue;
657 } else {
658 if ((sectors > best_good_sectors) && (best_disk >= 0))
659 best_disk = -1;
660 best_good_sectors = sectors;
661 }
662
663 if (best_disk >= 0)
664 /* At least two disks to choose from so failfast is OK */
665 set_bit(R1BIO_FailFast, &r1_bio->state);
666
667 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
668 has_nonrot_disk |= nonrot;
669 pending = atomic_read(&rdev->nr_pending);
670 dist = abs(this_sector - conf->mirrors[disk].head_position);
671 if (choose_first) {
672 best_disk = disk;
673 break;
674 }
675 /* Don't change to another disk for sequential reads */
676 if (conf->mirrors[disk].next_seq_sect == this_sector
677 || dist == 0) {
678 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
679 struct raid1_info *mirror = &conf->mirrors[disk];
680
681 best_disk = disk;
682 /*
683 * If buffered sequential IO size exceeds optimal
684 * iosize, check if there is idle disk. If yes, choose
685 * the idle disk. read_balance could already choose an
686 * idle disk before noticing it's a sequential IO in
687 * this disk. This doesn't matter because this disk
688 * will idle, next time it will be utilized after the
689 * first disk has IO size exceeds optimal iosize. In
690 * this way, iosize of the first disk will be optimal
691 * iosize at least. iosize of the second disk might be
692 * small, but not a big deal since when the second disk
693 * starts IO, the first disk is likely still busy.
694 */
695 if (nonrot && opt_iosize > 0 &&
696 mirror->seq_start != MaxSector &&
697 mirror->next_seq_sect > opt_iosize &&
698 mirror->next_seq_sect - opt_iosize >=
699 mirror->seq_start) {
700 choose_next_idle = 1;
701 continue;
702 }
703 break;
704 }
705
706 if (choose_next_idle)
707 continue;
708
709 if (min_pending > pending) {
710 min_pending = pending;
711 best_pending_disk = disk;
712 }
713
714 if (dist < best_dist) {
715 best_dist = dist;
716 best_dist_disk = disk;
717 }
718 }
719
720 /*
721 * If all disks are rotational, choose the closest disk. If any disk is
722 * non-rotational, choose the disk with less pending request even the
723 * disk is rotational, which might/might not be optimal for raids with
724 * mixed ratation/non-rotational disks depending on workload.
725 */
726 if (best_disk == -1) {
727 if (has_nonrot_disk || min_pending == 0)
728 best_disk = best_pending_disk;
729 else
730 best_disk = best_dist_disk;
731 }
732
733 if (best_disk >= 0) {
734 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
735 if (!rdev)
736 goto retry;
737 atomic_inc(&rdev->nr_pending);
738 sectors = best_good_sectors;
739
740 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
741 conf->mirrors[best_disk].seq_start = this_sector;
742
743 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
744 }
745 rcu_read_unlock();
746 *max_sectors = sectors;
747
748 return best_disk;
749}
750
751static int raid1_congested(struct mddev *mddev, int bits)
752{
753 struct r1conf *conf = mddev->private;
754 int i, ret = 0;
755
756 if ((bits & (1 << WB_async_congested)) &&
757 conf->pending_count >= max_queued_requests)
758 return 1;
759
760 rcu_read_lock();
761 for (i = 0; i < conf->raid_disks * 2; i++) {
762 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
763 if (rdev && !test_bit(Faulty, &rdev->flags)) {
764 struct request_queue *q = bdev_get_queue(rdev->bdev);
765
766 BUG_ON(!q);
767
768 /* Note the '|| 1' - when read_balance prefers
769 * non-congested targets, it can be removed
770 */
771 if ((bits & (1 << WB_async_congested)) || 1)
772 ret |= bdi_congested(q->backing_dev_info, bits);
773 else
774 ret &= bdi_congested(q->backing_dev_info, bits);
775 }
776 }
777 rcu_read_unlock();
778 return ret;
779}
780
781static void flush_bio_list(struct r1conf *conf, struct bio *bio)
782{
783 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
784 bitmap_unplug(conf->mddev->bitmap);
785 wake_up(&conf->wait_barrier);
786
787 while (bio) { /* submit pending writes */
788 struct bio *next = bio->bi_next;
789 struct md_rdev *rdev = (void *)bio->bi_disk;
790 bio->bi_next = NULL;
791 bio_set_dev(bio, rdev->bdev);
792 if (test_bit(Faulty, &rdev->flags)) {
793 bio_io_error(bio);
794 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
795 !blk_queue_discard(bio->bi_disk->queue)))
796 /* Just ignore it */
797 bio_endio(bio);
798 else
799 generic_make_request(bio);
800 bio = next;
801 }
802}
803
804static void flush_pending_writes(struct r1conf *conf)
805{
806 /* Any writes that have been queued but are awaiting
807 * bitmap updates get flushed here.
808 */
809 spin_lock_irq(&conf->device_lock);
810
811 if (conf->pending_bio_list.head) {
812 struct blk_plug plug;
813 struct bio *bio;
814
815 bio = bio_list_get(&conf->pending_bio_list);
816 conf->pending_count = 0;
817 spin_unlock_irq(&conf->device_lock);
818
819 /*
820 * As this is called in a wait_event() loop (see freeze_array),
821 * current->state might be TASK_UNINTERRUPTIBLE which will
822 * cause a warning when we prepare to wait again. As it is
823 * rare that this path is taken, it is perfectly safe to force
824 * us to go around the wait_event() loop again, so the warning
825 * is a false-positive. Silence the warning by resetting
826 * thread state
827 */
828 __set_current_state(TASK_RUNNING);
829 blk_start_plug(&plug);
830 flush_bio_list(conf, bio);
831 blk_finish_plug(&plug);
832 } else
833 spin_unlock_irq(&conf->device_lock);
834}
835
836/* Barriers....
837 * Sometimes we need to suspend IO while we do something else,
838 * either some resync/recovery, or reconfigure the array.
839 * To do this we raise a 'barrier'.
840 * The 'barrier' is a counter that can be raised multiple times
841 * to count how many activities are happening which preclude
842 * normal IO.
843 * We can only raise the barrier if there is no pending IO.
844 * i.e. if nr_pending == 0.
845 * We choose only to raise the barrier if no-one is waiting for the
846 * barrier to go down. This means that as soon as an IO request
847 * is ready, no other operations which require a barrier will start
848 * until the IO request has had a chance.
849 *
850 * So: regular IO calls 'wait_barrier'. When that returns there
851 * is no backgroup IO happening, It must arrange to call
852 * allow_barrier when it has finished its IO.
853 * backgroup IO calls must call raise_barrier. Once that returns
854 * there is no normal IO happeing. It must arrange to call
855 * lower_barrier when the particular background IO completes.
856 */
857static sector_t raise_barrier(struct r1conf *conf, sector_t sector_nr)
858{
859 int idx = sector_to_idx(sector_nr);
860
861 spin_lock_irq(&conf->resync_lock);
862
863 /* Wait until no block IO is waiting */
864 wait_event_lock_irq(conf->wait_barrier,
865 !atomic_read(&conf->nr_waiting[idx]),
866 conf->resync_lock);
867
868 /* block any new IO from starting */
869 atomic_inc(&conf->barrier[idx]);
870 /*
871 * In raise_barrier() we firstly increase conf->barrier[idx] then
872 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
873 * increase conf->nr_pending[idx] then check conf->barrier[idx].
874 * A memory barrier here to make sure conf->nr_pending[idx] won't
875 * be fetched before conf->barrier[idx] is increased. Otherwise
876 * there will be a race between raise_barrier() and _wait_barrier().
877 */
878 smp_mb__after_atomic();
879
880 /* For these conditions we must wait:
881 * A: while the array is in frozen state
882 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
883 * existing in corresponding I/O barrier bucket.
884 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
885 * max resync count which allowed on current I/O barrier bucket.
886 */
887 wait_event_lock_irq(conf->wait_barrier,
888 (!conf->array_frozen &&
889 !atomic_read(&conf->nr_pending[idx]) &&
890 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
891 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
892 conf->resync_lock);
893
894 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
895 atomic_dec(&conf->barrier[idx]);
896 spin_unlock_irq(&conf->resync_lock);
897 wake_up(&conf->wait_barrier);
898 return -EINTR;
899 }
900
901 atomic_inc(&conf->nr_sync_pending);
902 spin_unlock_irq(&conf->resync_lock);
903
904 return 0;
905}
906
907static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
908{
909 int idx = sector_to_idx(sector_nr);
910
911 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
912
913 atomic_dec(&conf->barrier[idx]);
914 atomic_dec(&conf->nr_sync_pending);
915 wake_up(&conf->wait_barrier);
916}
917
918static void _wait_barrier(struct r1conf *conf, int idx)
919{
920 /*
921 * We need to increase conf->nr_pending[idx] very early here,
922 * then raise_barrier() can be blocked when it waits for
923 * conf->nr_pending[idx] to be 0. Then we can avoid holding
924 * conf->resync_lock when there is no barrier raised in same
925 * barrier unit bucket. Also if the array is frozen, I/O
926 * should be blocked until array is unfrozen.
927 */
928 atomic_inc(&conf->nr_pending[idx]);
929 /*
930 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
931 * check conf->barrier[idx]. In raise_barrier() we firstly increase
932 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
933 * barrier is necessary here to make sure conf->barrier[idx] won't be
934 * fetched before conf->nr_pending[idx] is increased. Otherwise there
935 * will be a race between _wait_barrier() and raise_barrier().
936 */
937 smp_mb__after_atomic();
938
939 /*
940 * Don't worry about checking two atomic_t variables at same time
941 * here. If during we check conf->barrier[idx], the array is
942 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
943 * 0, it is safe to return and make the I/O continue. Because the
944 * array is frozen, all I/O returned here will eventually complete
945 * or be queued, no race will happen. See code comment in
946 * frozen_array().
947 */
948 if (!READ_ONCE(conf->array_frozen) &&
949 !atomic_read(&conf->barrier[idx]))
950 return;
951
952 /*
953 * After holding conf->resync_lock, conf->nr_pending[idx]
954 * should be decreased before waiting for barrier to drop.
955 * Otherwise, we may encounter a race condition because
956 * raise_barrer() might be waiting for conf->nr_pending[idx]
957 * to be 0 at same time.
958 */
959 spin_lock_irq(&conf->resync_lock);
960 atomic_inc(&conf->nr_waiting[idx]);
961 atomic_dec(&conf->nr_pending[idx]);
962 /*
963 * In case freeze_array() is waiting for
964 * get_unqueued_pending() == extra
965 */
966 wake_up(&conf->wait_barrier);
967 /* Wait for the barrier in same barrier unit bucket to drop. */
968 wait_event_lock_irq(conf->wait_barrier,
969 !conf->array_frozen &&
970 !atomic_read(&conf->barrier[idx]),
971 conf->resync_lock);
972 atomic_inc(&conf->nr_pending[idx]);
973 atomic_dec(&conf->nr_waiting[idx]);
974 spin_unlock_irq(&conf->resync_lock);
975}
976
977static void wait_read_barrier(struct r1conf *conf, sector_t sector_nr)
978{
979 int idx = sector_to_idx(sector_nr);
980
981 /*
982 * Very similar to _wait_barrier(). The difference is, for read
983 * I/O we don't need wait for sync I/O, but if the whole array
984 * is frozen, the read I/O still has to wait until the array is
985 * unfrozen. Since there is no ordering requirement with
986 * conf->barrier[idx] here, memory barrier is unnecessary as well.
987 */
988 atomic_inc(&conf->nr_pending[idx]);
989
990 if (!READ_ONCE(conf->array_frozen))
991 return;
992
993 spin_lock_irq(&conf->resync_lock);
994 atomic_inc(&conf->nr_waiting[idx]);
995 atomic_dec(&conf->nr_pending[idx]);
996 /*
997 * In case freeze_array() is waiting for
998 * get_unqueued_pending() == extra
999 */
1000 wake_up(&conf->wait_barrier);
1001 /* Wait for array to be unfrozen */
1002 wait_event_lock_irq(conf->wait_barrier,
1003 !conf->array_frozen,
1004 conf->resync_lock);
1005 atomic_inc(&conf->nr_pending[idx]);
1006 atomic_dec(&conf->nr_waiting[idx]);
1007 spin_unlock_irq(&conf->resync_lock);
1008}
1009
1010static void wait_barrier(struct r1conf *conf, sector_t sector_nr)
1011{
1012 int idx = sector_to_idx(sector_nr);
1013
1014 _wait_barrier(conf, idx);
1015}
1016
1017static void _allow_barrier(struct r1conf *conf, int idx)
1018{
1019 atomic_dec(&conf->nr_pending[idx]);
1020 wake_up(&conf->wait_barrier);
1021}
1022
1023static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1024{
1025 int idx = sector_to_idx(sector_nr);
1026
1027 _allow_barrier(conf, idx);
1028}
1029
1030/* conf->resync_lock should be held */
1031static int get_unqueued_pending(struct r1conf *conf)
1032{
1033 int idx, ret;
1034
1035 ret = atomic_read(&conf->nr_sync_pending);
1036 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1037 ret += atomic_read(&conf->nr_pending[idx]) -
1038 atomic_read(&conf->nr_queued[idx]);
1039
1040 return ret;
1041}
1042
1043static void freeze_array(struct r1conf *conf, int extra)
1044{
1045 /* Stop sync I/O and normal I/O and wait for everything to
1046 * go quiet.
1047 * This is called in two situations:
1048 * 1) management command handlers (reshape, remove disk, quiesce).
1049 * 2) one normal I/O request failed.
1050
1051 * After array_frozen is set to 1, new sync IO will be blocked at
1052 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1053 * or wait_read_barrier(). The flying I/Os will either complete or be
1054 * queued. When everything goes quite, there are only queued I/Os left.
1055
1056 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1057 * barrier bucket index which this I/O request hits. When all sync and
1058 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1059 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1060 * in handle_read_error(), we may call freeze_array() before trying to
1061 * fix the read error. In this case, the error read I/O is not queued,
1062 * so get_unqueued_pending() == 1.
1063 *
1064 * Therefore before this function returns, we need to wait until
1065 * get_unqueued_pendings(conf) gets equal to extra. For
1066 * normal I/O context, extra is 1, in rested situations extra is 0.
1067 */
1068 spin_lock_irq(&conf->resync_lock);
1069 conf->array_frozen = 1;
1070 raid1_log(conf->mddev, "wait freeze");
1071 wait_event_lock_irq_cmd(
1072 conf->wait_barrier,
1073 get_unqueued_pending(conf) == extra,
1074 conf->resync_lock,
1075 flush_pending_writes(conf));
1076 spin_unlock_irq(&conf->resync_lock);
1077}
1078static void unfreeze_array(struct r1conf *conf)
1079{
1080 /* reverse the effect of the freeze */
1081 spin_lock_irq(&conf->resync_lock);
1082 conf->array_frozen = 0;
1083 spin_unlock_irq(&conf->resync_lock);
1084 wake_up(&conf->wait_barrier);
1085}
1086
1087static void alloc_behind_master_bio(struct r1bio *r1_bio,
1088 struct bio *bio)
1089{
1090 int size = bio->bi_iter.bi_size;
1091 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1092 int i = 0;
1093 struct bio *behind_bio = NULL;
1094
1095 behind_bio = bio_alloc_mddev(GFP_NOIO, vcnt, r1_bio->mddev);
1096 if (!behind_bio)
1097 return;
1098
1099 /* discard op, we don't support writezero/writesame yet */
1100 if (!bio_has_data(bio)) {
1101 behind_bio->bi_iter.bi_size = size;
1102 goto skip_copy;
1103 }
1104
1105 behind_bio->bi_write_hint = bio->bi_write_hint;
1106
1107 while (i < vcnt && size) {
1108 struct page *page;
1109 int len = min_t(int, PAGE_SIZE, size);
1110
1111 page = alloc_page(GFP_NOIO);
1112 if (unlikely(!page))
1113 goto free_pages;
1114
1115 bio_add_page(behind_bio, page, len, 0);
1116
1117 size -= len;
1118 i++;
1119 }
1120
1121 bio_copy_data(behind_bio, bio);
1122skip_copy:
1123 r1_bio->behind_master_bio = behind_bio;
1124 set_bit(R1BIO_BehindIO, &r1_bio->state);
1125
1126 return;
1127
1128free_pages:
1129 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1130 bio->bi_iter.bi_size);
1131 bio_free_pages(behind_bio);
1132 bio_put(behind_bio);
1133}
1134
1135struct raid1_plug_cb {
1136 struct blk_plug_cb cb;
1137 struct bio_list pending;
1138 int pending_cnt;
1139};
1140
1141static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1142{
1143 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1144 cb);
1145 struct mddev *mddev = plug->cb.data;
1146 struct r1conf *conf = mddev->private;
1147 struct bio *bio;
1148
1149 if (from_schedule || current->bio_list) {
1150 spin_lock_irq(&conf->device_lock);
1151 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1152 conf->pending_count += plug->pending_cnt;
1153 spin_unlock_irq(&conf->device_lock);
1154 wake_up(&conf->wait_barrier);
1155 md_wakeup_thread(mddev->thread);
1156 kfree(plug);
1157 return;
1158 }
1159
1160 /* we aren't scheduling, so we can do the write-out directly. */
1161 bio = bio_list_get(&plug->pending);
1162 flush_bio_list(conf, bio);
1163 kfree(plug);
1164}
1165
1166static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1167{
1168 r1_bio->master_bio = bio;
1169 r1_bio->sectors = bio_sectors(bio);
1170 r1_bio->state = 0;
1171 r1_bio->mddev = mddev;
1172 r1_bio->sector = bio->bi_iter.bi_sector;
1173}
1174
1175static inline struct r1bio *
1176alloc_r1bio(struct mddev *mddev, struct bio *bio)
1177{
1178 struct r1conf *conf = mddev->private;
1179 struct r1bio *r1_bio;
1180
1181 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1182 /* Ensure no bio records IO_BLOCKED */
1183 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1184 init_r1bio(r1_bio, mddev, bio);
1185 return r1_bio;
1186}
1187
1188static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1189 int max_read_sectors, struct r1bio *r1_bio)
1190{
1191 struct r1conf *conf = mddev->private;
1192 struct raid1_info *mirror;
1193 struct bio *read_bio;
1194 struct bitmap *bitmap = mddev->bitmap;
1195 const int op = bio_op(bio);
1196 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1197 int max_sectors;
1198 int rdisk;
1199 bool print_msg = !!r1_bio;
1200 char b[BDEVNAME_SIZE];
1201
1202 /*
1203 * If r1_bio is set, we are blocking the raid1d thread
1204 * so there is a tiny risk of deadlock. So ask for
1205 * emergency memory if needed.
1206 */
1207 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1208
1209 if (print_msg) {
1210 /* Need to get the block device name carefully */
1211 struct md_rdev *rdev;
1212 rcu_read_lock();
1213 rdev = rcu_dereference(conf->mirrors[r1_bio->read_disk].rdev);
1214 if (rdev)
1215 bdevname(rdev->bdev, b);
1216 else
1217 strcpy(b, "???");
1218 rcu_read_unlock();
1219 }
1220
1221 /*
1222 * Still need barrier for READ in case that whole
1223 * array is frozen.
1224 */
1225 wait_read_barrier(conf, bio->bi_iter.bi_sector);
1226
1227 if (!r1_bio)
1228 r1_bio = alloc_r1bio(mddev, bio);
1229 else
1230 init_r1bio(r1_bio, mddev, bio);
1231 r1_bio->sectors = max_read_sectors;
1232
1233 /*
1234 * make_request() can abort the operation when read-ahead is being
1235 * used and no empty request is available.
1236 */
1237 rdisk = read_balance(conf, r1_bio, &max_sectors);
1238
1239 if (rdisk < 0) {
1240 /* couldn't find anywhere to read from */
1241 if (print_msg) {
1242 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1243 mdname(mddev),
1244 b,
1245 (unsigned long long)r1_bio->sector);
1246 }
1247 raid_end_bio_io(r1_bio);
1248 return;
1249 }
1250 mirror = conf->mirrors + rdisk;
1251
1252 if (print_msg)
1253 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %s\n",
1254 mdname(mddev),
1255 (unsigned long long)r1_bio->sector,
1256 bdevname(mirror->rdev->bdev, b));
1257
1258 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1259 bitmap) {
1260 /*
1261 * Reading from a write-mostly device must take care not to
1262 * over-take any writes that are 'behind'
1263 */
1264 raid1_log(mddev, "wait behind writes");
1265 wait_event(bitmap->behind_wait,
1266 atomic_read(&bitmap->behind_writes) == 0);
1267 }
1268
1269 if (max_sectors < bio_sectors(bio)) {
1270 struct bio *split = bio_split(bio, max_sectors,
1271 gfp, conf->bio_split);
1272 bio_chain(split, bio);
1273 generic_make_request(bio);
1274 bio = split;
1275 r1_bio->master_bio = bio;
1276 r1_bio->sectors = max_sectors;
1277 }
1278
1279 r1_bio->read_disk = rdisk;
1280
1281 read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1282
1283 r1_bio->bios[rdisk] = read_bio;
1284
1285 read_bio->bi_iter.bi_sector = r1_bio->sector +
1286 mirror->rdev->data_offset;
1287 bio_set_dev(read_bio, mirror->rdev->bdev);
1288 read_bio->bi_end_io = raid1_end_read_request;
1289 bio_set_op_attrs(read_bio, op, do_sync);
1290 if (test_bit(FailFast, &mirror->rdev->flags) &&
1291 test_bit(R1BIO_FailFast, &r1_bio->state))
1292 read_bio->bi_opf |= MD_FAILFAST;
1293 read_bio->bi_private = r1_bio;
1294
1295 if (mddev->gendisk)
1296 trace_block_bio_remap(read_bio->bi_disk->queue, read_bio,
1297 disk_devt(mddev->gendisk), r1_bio->sector);
1298
1299 generic_make_request(read_bio);
1300}
1301
1302static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1303 int max_write_sectors)
1304{
1305 struct r1conf *conf = mddev->private;
1306 struct r1bio *r1_bio;
1307 int i, disks;
1308 struct bitmap *bitmap = mddev->bitmap;
1309 unsigned long flags;
1310 struct md_rdev *blocked_rdev;
1311 struct blk_plug_cb *cb;
1312 struct raid1_plug_cb *plug = NULL;
1313 int first_clone;
1314 int max_sectors;
1315
1316 if (mddev_is_clustered(mddev) &&
1317 md_cluster_ops->area_resyncing(mddev, WRITE,
1318 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1319
1320 DEFINE_WAIT(w);
1321 for (;;) {
1322 prepare_to_wait(&conf->wait_barrier,
1323 &w, TASK_IDLE);
1324 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1325 bio->bi_iter.bi_sector,
1326 bio_end_sector(bio)))
1327 break;
1328 schedule();
1329 }
1330 finish_wait(&conf->wait_barrier, &w);
1331 }
1332
1333 /*
1334 * Register the new request and wait if the reconstruction
1335 * thread has put up a bar for new requests.
1336 * Continue immediately if no resync is active currently.
1337 */
1338 wait_barrier(conf, bio->bi_iter.bi_sector);
1339
1340 r1_bio = alloc_r1bio(mddev, bio);
1341 r1_bio->sectors = max_write_sectors;
1342
1343 if (conf->pending_count >= max_queued_requests) {
1344 md_wakeup_thread(mddev->thread);
1345 raid1_log(mddev, "wait queued");
1346 wait_event(conf->wait_barrier,
1347 conf->pending_count < max_queued_requests);
1348 }
1349 /* first select target devices under rcu_lock and
1350 * inc refcount on their rdev. Record them by setting
1351 * bios[x] to bio
1352 * If there are known/acknowledged bad blocks on any device on
1353 * which we have seen a write error, we want to avoid writing those
1354 * blocks.
1355 * This potentially requires several writes to write around
1356 * the bad blocks. Each set of writes gets it's own r1bio
1357 * with a set of bios attached.
1358 */
1359
1360 disks = conf->raid_disks * 2;
1361 retry_write:
1362 blocked_rdev = NULL;
1363 rcu_read_lock();
1364 max_sectors = r1_bio->sectors;
1365 for (i = 0; i < disks; i++) {
1366 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1367 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1368 atomic_inc(&rdev->nr_pending);
1369 blocked_rdev = rdev;
1370 break;
1371 }
1372 r1_bio->bios[i] = NULL;
1373 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1374 if (i < conf->raid_disks)
1375 set_bit(R1BIO_Degraded, &r1_bio->state);
1376 continue;
1377 }
1378
1379 atomic_inc(&rdev->nr_pending);
1380 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1381 sector_t first_bad;
1382 int bad_sectors;
1383 int is_bad;
1384
1385 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1386 &first_bad, &bad_sectors);
1387 if (is_bad < 0) {
1388 /* mustn't write here until the bad block is
1389 * acknowledged*/
1390 set_bit(BlockedBadBlocks, &rdev->flags);
1391 blocked_rdev = rdev;
1392 break;
1393 }
1394 if (is_bad && first_bad <= r1_bio->sector) {
1395 /* Cannot write here at all */
1396 bad_sectors -= (r1_bio->sector - first_bad);
1397 if (bad_sectors < max_sectors)
1398 /* mustn't write more than bad_sectors
1399 * to other devices yet
1400 */
1401 max_sectors = bad_sectors;
1402 rdev_dec_pending(rdev, mddev);
1403 /* We don't set R1BIO_Degraded as that
1404 * only applies if the disk is
1405 * missing, so it might be re-added,
1406 * and we want to know to recover this
1407 * chunk.
1408 * In this case the device is here,
1409 * and the fact that this chunk is not
1410 * in-sync is recorded in the bad
1411 * block log
1412 */
1413 continue;
1414 }
1415 if (is_bad) {
1416 int good_sectors = first_bad - r1_bio->sector;
1417 if (good_sectors < max_sectors)
1418 max_sectors = good_sectors;
1419 }
1420 }
1421 r1_bio->bios[i] = bio;
1422 }
1423 rcu_read_unlock();
1424
1425 if (unlikely(blocked_rdev)) {
1426 /* Wait for this device to become unblocked */
1427 int j;
1428
1429 for (j = 0; j < i; j++)
1430 if (r1_bio->bios[j])
1431 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1432 r1_bio->state = 0;
1433 allow_barrier(conf, bio->bi_iter.bi_sector);
1434 raid1_log(mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1435 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1436 wait_barrier(conf, bio->bi_iter.bi_sector);
1437 goto retry_write;
1438 }
1439
1440 if (max_sectors < bio_sectors(bio)) {
1441 struct bio *split = bio_split(bio, max_sectors,
1442 GFP_NOIO, conf->bio_split);
1443 bio_chain(split, bio);
1444 generic_make_request(bio);
1445 bio = split;
1446 r1_bio->master_bio = bio;
1447 r1_bio->sectors = max_sectors;
1448 }
1449
1450 atomic_set(&r1_bio->remaining, 1);
1451 atomic_set(&r1_bio->behind_remaining, 0);
1452
1453 first_clone = 1;
1454
1455 for (i = 0; i < disks; i++) {
1456 struct bio *mbio = NULL;
1457 if (!r1_bio->bios[i])
1458 continue;
1459
1460
1461 if (first_clone) {
1462 /* do behind I/O ?
1463 * Not if there are too many, or cannot
1464 * allocate memory, or a reader on WriteMostly
1465 * is waiting for behind writes to flush */
1466 if (bitmap &&
1467 (atomic_read(&bitmap->behind_writes)
1468 < mddev->bitmap_info.max_write_behind) &&
1469 !waitqueue_active(&bitmap->behind_wait)) {
1470 alloc_behind_master_bio(r1_bio, bio);
1471 }
1472
1473 bitmap_startwrite(bitmap, r1_bio->sector,
1474 r1_bio->sectors,
1475 test_bit(R1BIO_BehindIO,
1476 &r1_bio->state));
1477 first_clone = 0;
1478 }
1479
1480 if (r1_bio->behind_master_bio)
1481 mbio = bio_clone_fast(r1_bio->behind_master_bio,
1482 GFP_NOIO, mddev->bio_set);
1483 else
1484 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1485
1486 if (r1_bio->behind_master_bio) {
1487 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1488 atomic_inc(&r1_bio->behind_remaining);
1489 }
1490
1491 r1_bio->bios[i] = mbio;
1492
1493 mbio->bi_iter.bi_sector = (r1_bio->sector +
1494 conf->mirrors[i].rdev->data_offset);
1495 bio_set_dev(mbio, conf->mirrors[i].rdev->bdev);
1496 mbio->bi_end_io = raid1_end_write_request;
1497 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1498 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags) &&
1499 !test_bit(WriteMostly, &conf->mirrors[i].rdev->flags) &&
1500 conf->raid_disks - mddev->degraded > 1)
1501 mbio->bi_opf |= MD_FAILFAST;
1502 mbio->bi_private = r1_bio;
1503
1504 atomic_inc(&r1_bio->remaining);
1505
1506 if (mddev->gendisk)
1507 trace_block_bio_remap(mbio->bi_disk->queue,
1508 mbio, disk_devt(mddev->gendisk),
1509 r1_bio->sector);
1510 /* flush_pending_writes() needs access to the rdev so...*/
1511 mbio->bi_disk = (void *)conf->mirrors[i].rdev;
1512
1513 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1514 if (cb)
1515 plug = container_of(cb, struct raid1_plug_cb, cb);
1516 else
1517 plug = NULL;
1518 if (plug) {
1519 bio_list_add(&plug->pending, mbio);
1520 plug->pending_cnt++;
1521 } else {
1522 spin_lock_irqsave(&conf->device_lock, flags);
1523 bio_list_add(&conf->pending_bio_list, mbio);
1524 conf->pending_count++;
1525 spin_unlock_irqrestore(&conf->device_lock, flags);
1526 md_wakeup_thread(mddev->thread);
1527 }
1528 }
1529
1530 r1_bio_write_done(r1_bio);
1531
1532 /* In case raid1d snuck in to freeze_array */
1533 wake_up(&conf->wait_barrier);
1534}
1535
1536static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1537{
1538 sector_t sectors;
1539
1540 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1541 md_flush_request(mddev, bio);
1542 return true;
1543 }
1544
1545 /*
1546 * There is a limit to the maximum size, but
1547 * the read/write handler might find a lower limit
1548 * due to bad blocks. To avoid multiple splits,
1549 * we pass the maximum number of sectors down
1550 * and let the lower level perform the split.
1551 */
1552 sectors = align_to_barrier_unit_end(
1553 bio->bi_iter.bi_sector, bio_sectors(bio));
1554
1555 if (bio_data_dir(bio) == READ)
1556 raid1_read_request(mddev, bio, sectors, NULL);
1557 else {
1558 if (!md_write_start(mddev,bio))
1559 return false;
1560 raid1_write_request(mddev, bio, sectors);
1561 }
1562 return true;
1563}
1564
1565static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1566{
1567 struct r1conf *conf = mddev->private;
1568 int i;
1569
1570 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1571 conf->raid_disks - mddev->degraded);
1572 rcu_read_lock();
1573 for (i = 0; i < conf->raid_disks; i++) {
1574 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1575 seq_printf(seq, "%s",
1576 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1577 }
1578 rcu_read_unlock();
1579 seq_printf(seq, "]");
1580}
1581
1582static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1583{
1584 char b[BDEVNAME_SIZE];
1585 struct r1conf *conf = mddev->private;
1586 unsigned long flags;
1587
1588 /*
1589 * If it is not operational, then we have already marked it as dead
1590 * else if it is the last working disks, ignore the error, let the
1591 * next level up know.
1592 * else mark the drive as failed
1593 */
1594 spin_lock_irqsave(&conf->device_lock, flags);
1595 if (test_bit(In_sync, &rdev->flags)
1596 && (conf->raid_disks - mddev->degraded) == 1) {
1597 /*
1598 * Don't fail the drive, act as though we were just a
1599 * normal single drive.
1600 * However don't try a recovery from this drive as
1601 * it is very likely to fail.
1602 */
1603 conf->recovery_disabled = mddev->recovery_disabled;
1604 spin_unlock_irqrestore(&conf->device_lock, flags);
1605 return;
1606 }
1607 set_bit(Blocked, &rdev->flags);
1608 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1609 mddev->degraded++;
1610 set_bit(Faulty, &rdev->flags);
1611 } else
1612 set_bit(Faulty, &rdev->flags);
1613 spin_unlock_irqrestore(&conf->device_lock, flags);
1614 /*
1615 * if recovery is running, make sure it aborts.
1616 */
1617 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1618 set_mask_bits(&mddev->sb_flags, 0,
1619 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1620 pr_crit("md/raid1:%s: Disk failure on %s, disabling device.\n"
1621 "md/raid1:%s: Operation continuing on %d devices.\n",
1622 mdname(mddev), bdevname(rdev->bdev, b),
1623 mdname(mddev), conf->raid_disks - mddev->degraded);
1624}
1625
1626static void print_conf(struct r1conf *conf)
1627{
1628 int i;
1629
1630 pr_debug("RAID1 conf printout:\n");
1631 if (!conf) {
1632 pr_debug("(!conf)\n");
1633 return;
1634 }
1635 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1636 conf->raid_disks);
1637
1638 rcu_read_lock();
1639 for (i = 0; i < conf->raid_disks; i++) {
1640 char b[BDEVNAME_SIZE];
1641 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1642 if (rdev)
1643 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1644 i, !test_bit(In_sync, &rdev->flags),
1645 !test_bit(Faulty, &rdev->flags),
1646 bdevname(rdev->bdev,b));
1647 }
1648 rcu_read_unlock();
1649}
1650
1651static void close_sync(struct r1conf *conf)
1652{
1653 int idx;
1654
1655 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1656 _wait_barrier(conf, idx);
1657 _allow_barrier(conf, idx);
1658 }
1659
1660 mempool_destroy(conf->r1buf_pool);
1661 conf->r1buf_pool = NULL;
1662}
1663
1664static int raid1_spare_active(struct mddev *mddev)
1665{
1666 int i;
1667 struct r1conf *conf = mddev->private;
1668 int count = 0;
1669 unsigned long flags;
1670
1671 /*
1672 * Find all failed disks within the RAID1 configuration
1673 * and mark them readable.
1674 * Called under mddev lock, so rcu protection not needed.
1675 * device_lock used to avoid races with raid1_end_read_request
1676 * which expects 'In_sync' flags and ->degraded to be consistent.
1677 */
1678 spin_lock_irqsave(&conf->device_lock, flags);
1679 for (i = 0; i < conf->raid_disks; i++) {
1680 struct md_rdev *rdev = conf->mirrors[i].rdev;
1681 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1682 if (repl
1683 && !test_bit(Candidate, &repl->flags)
1684 && repl->recovery_offset == MaxSector
1685 && !test_bit(Faulty, &repl->flags)
1686 && !test_and_set_bit(In_sync, &repl->flags)) {
1687 /* replacement has just become active */
1688 if (!rdev ||
1689 !test_and_clear_bit(In_sync, &rdev->flags))
1690 count++;
1691 if (rdev) {
1692 /* Replaced device not technically
1693 * faulty, but we need to be sure
1694 * it gets removed and never re-added
1695 */
1696 set_bit(Faulty, &rdev->flags);
1697 sysfs_notify_dirent_safe(
1698 rdev->sysfs_state);
1699 }
1700 }
1701 if (rdev
1702 && rdev->recovery_offset == MaxSector
1703 && !test_bit(Faulty, &rdev->flags)
1704 && !test_and_set_bit(In_sync, &rdev->flags)) {
1705 count++;
1706 sysfs_notify_dirent_safe(rdev->sysfs_state);
1707 }
1708 }
1709 mddev->degraded -= count;
1710 spin_unlock_irqrestore(&conf->device_lock, flags);
1711
1712 print_conf(conf);
1713 return count;
1714}
1715
1716static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1717{
1718 struct r1conf *conf = mddev->private;
1719 int err = -EEXIST;
1720 int mirror = 0;
1721 struct raid1_info *p;
1722 int first = 0;
1723 int last = conf->raid_disks - 1;
1724
1725 if (mddev->recovery_disabled == conf->recovery_disabled)
1726 return -EBUSY;
1727
1728 if (md_integrity_add_rdev(rdev, mddev))
1729 return -ENXIO;
1730
1731 if (rdev->raid_disk >= 0)
1732 first = last = rdev->raid_disk;
1733
1734 /*
1735 * find the disk ... but prefer rdev->saved_raid_disk
1736 * if possible.
1737 */
1738 if (rdev->saved_raid_disk >= 0 &&
1739 rdev->saved_raid_disk >= first &&
1740 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1741 first = last = rdev->saved_raid_disk;
1742
1743 for (mirror = first; mirror <= last; mirror++) {
1744 p = conf->mirrors+mirror;
1745 if (!p->rdev) {
1746
1747 if (mddev->gendisk)
1748 disk_stack_limits(mddev->gendisk, rdev->bdev,
1749 rdev->data_offset << 9);
1750
1751 p->head_position = 0;
1752 rdev->raid_disk = mirror;
1753 err = 0;
1754 /* As all devices are equivalent, we don't need a full recovery
1755 * if this was recently any drive of the array
1756 */
1757 if (rdev->saved_raid_disk < 0)
1758 conf->fullsync = 1;
1759 rcu_assign_pointer(p->rdev, rdev);
1760 break;
1761 }
1762 if (test_bit(WantReplacement, &p->rdev->flags) &&
1763 p[conf->raid_disks].rdev == NULL) {
1764 /* Add this device as a replacement */
1765 clear_bit(In_sync, &rdev->flags);
1766 set_bit(Replacement, &rdev->flags);
1767 rdev->raid_disk = mirror;
1768 err = 0;
1769 conf->fullsync = 1;
1770 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1771 break;
1772 }
1773 }
1774 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1775 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1776 print_conf(conf);
1777 return err;
1778}
1779
1780static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1781{
1782 struct r1conf *conf = mddev->private;
1783 int err = 0;
1784 int number = rdev->raid_disk;
1785 struct raid1_info *p = conf->mirrors + number;
1786
1787 if (rdev != p->rdev)
1788 p = conf->mirrors + conf->raid_disks + number;
1789
1790 print_conf(conf);
1791 if (rdev == p->rdev) {
1792 if (test_bit(In_sync, &rdev->flags) ||
1793 atomic_read(&rdev->nr_pending)) {
1794 err = -EBUSY;
1795 goto abort;
1796 }
1797 /* Only remove non-faulty devices if recovery
1798 * is not possible.
1799 */
1800 if (!test_bit(Faulty, &rdev->flags) &&
1801 mddev->recovery_disabled != conf->recovery_disabled &&
1802 mddev->degraded < conf->raid_disks) {
1803 err = -EBUSY;
1804 goto abort;
1805 }
1806 p->rdev = NULL;
1807 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1808 synchronize_rcu();
1809 if (atomic_read(&rdev->nr_pending)) {
1810 /* lost the race, try later */
1811 err = -EBUSY;
1812 p->rdev = rdev;
1813 goto abort;
1814 }
1815 }
1816 if (conf->mirrors[conf->raid_disks + number].rdev) {
1817 /* We just removed a device that is being replaced.
1818 * Move down the replacement. We drain all IO before
1819 * doing this to avoid confusion.
1820 */
1821 struct md_rdev *repl =
1822 conf->mirrors[conf->raid_disks + number].rdev;
1823 freeze_array(conf, 0);
1824 if (atomic_read(&repl->nr_pending)) {
1825 /* It means that some queued IO of retry_list
1826 * hold repl. Thus, we cannot set replacement
1827 * as NULL, avoiding rdev NULL pointer
1828 * dereference in sync_request_write and
1829 * handle_write_finished.
1830 */
1831 err = -EBUSY;
1832 unfreeze_array(conf);
1833 goto abort;
1834 }
1835 clear_bit(Replacement, &repl->flags);
1836 p->rdev = repl;
1837 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1838 unfreeze_array(conf);
1839 }
1840
1841 clear_bit(WantReplacement, &rdev->flags);
1842 err = md_integrity_register(mddev);
1843 }
1844abort:
1845
1846 print_conf(conf);
1847 return err;
1848}
1849
1850static void end_sync_read(struct bio *bio)
1851{
1852 struct r1bio *r1_bio = get_resync_r1bio(bio);
1853
1854 update_head_pos(r1_bio->read_disk, r1_bio);
1855
1856 /*
1857 * we have read a block, now it needs to be re-written,
1858 * or re-read if the read failed.
1859 * We don't do much here, just schedule handling by raid1d
1860 */
1861 if (!bio->bi_status)
1862 set_bit(R1BIO_Uptodate, &r1_bio->state);
1863
1864 if (atomic_dec_and_test(&r1_bio->remaining))
1865 reschedule_retry(r1_bio);
1866}
1867
1868static void end_sync_write(struct bio *bio)
1869{
1870 int uptodate = !bio->bi_status;
1871 struct r1bio *r1_bio = get_resync_r1bio(bio);
1872 struct mddev *mddev = r1_bio->mddev;
1873 struct r1conf *conf = mddev->private;
1874 sector_t first_bad;
1875 int bad_sectors;
1876 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
1877
1878 if (!uptodate) {
1879 sector_t sync_blocks = 0;
1880 sector_t s = r1_bio->sector;
1881 long sectors_to_go = r1_bio->sectors;
1882 /* make sure these bits doesn't get cleared. */
1883 do {
1884 bitmap_end_sync(mddev->bitmap, s,
1885 &sync_blocks, 1);
1886 s += sync_blocks;
1887 sectors_to_go -= sync_blocks;
1888 } while (sectors_to_go > 0);
1889 set_bit(WriteErrorSeen, &rdev->flags);
1890 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1891 set_bit(MD_RECOVERY_NEEDED, &
1892 mddev->recovery);
1893 set_bit(R1BIO_WriteError, &r1_bio->state);
1894 } else if (is_badblock(rdev, r1_bio->sector, r1_bio->sectors,
1895 &first_bad, &bad_sectors) &&
1896 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1897 r1_bio->sector,
1898 r1_bio->sectors,
1899 &first_bad, &bad_sectors)
1900 )
1901 set_bit(R1BIO_MadeGood, &r1_bio->state);
1902
1903 if (atomic_dec_and_test(&r1_bio->remaining)) {
1904 int s = r1_bio->sectors;
1905 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1906 test_bit(R1BIO_WriteError, &r1_bio->state))
1907 reschedule_retry(r1_bio);
1908 else {
1909 put_buf(r1_bio);
1910 md_done_sync(mddev, s, uptodate);
1911 }
1912 }
1913}
1914
1915static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1916 int sectors, struct page *page, int rw)
1917{
1918 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
1919 /* success */
1920 return 1;
1921 if (rw == WRITE) {
1922 set_bit(WriteErrorSeen, &rdev->flags);
1923 if (!test_and_set_bit(WantReplacement,
1924 &rdev->flags))
1925 set_bit(MD_RECOVERY_NEEDED, &
1926 rdev->mddev->recovery);
1927 }
1928 /* need to record an error - either for the block or the device */
1929 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1930 md_error(rdev->mddev, rdev);
1931 return 0;
1932}
1933
1934static int fix_sync_read_error(struct r1bio *r1_bio)
1935{
1936 /* Try some synchronous reads of other devices to get
1937 * good data, much like with normal read errors. Only
1938 * read into the pages we already have so we don't
1939 * need to re-issue the read request.
1940 * We don't need to freeze the array, because being in an
1941 * active sync request, there is no normal IO, and
1942 * no overlapping syncs.
1943 * We don't need to check is_badblock() again as we
1944 * made sure that anything with a bad block in range
1945 * will have bi_end_io clear.
1946 */
1947 struct mddev *mddev = r1_bio->mddev;
1948 struct r1conf *conf = mddev->private;
1949 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1950 struct page **pages = get_resync_pages(bio)->pages;
1951 sector_t sect = r1_bio->sector;
1952 int sectors = r1_bio->sectors;
1953 int idx = 0;
1954 struct md_rdev *rdev;
1955
1956 rdev = conf->mirrors[r1_bio->read_disk].rdev;
1957 if (test_bit(FailFast, &rdev->flags)) {
1958 /* Don't try recovering from here - just fail it
1959 * ... unless it is the last working device of course */
1960 md_error(mddev, rdev);
1961 if (test_bit(Faulty, &rdev->flags))
1962 /* Don't try to read from here, but make sure
1963 * put_buf does it's thing
1964 */
1965 bio->bi_end_io = end_sync_write;
1966 }
1967
1968 while(sectors) {
1969 int s = sectors;
1970 int d = r1_bio->read_disk;
1971 int success = 0;
1972 int start;
1973
1974 if (s > (PAGE_SIZE>>9))
1975 s = PAGE_SIZE >> 9;
1976 do {
1977 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1978 /* No rcu protection needed here devices
1979 * can only be removed when no resync is
1980 * active, and resync is currently active
1981 */
1982 rdev = conf->mirrors[d].rdev;
1983 if (sync_page_io(rdev, sect, s<<9,
1984 pages[idx],
1985 REQ_OP_READ, 0, false)) {
1986 success = 1;
1987 break;
1988 }
1989 }
1990 d++;
1991 if (d == conf->raid_disks * 2)
1992 d = 0;
1993 } while (!success && d != r1_bio->read_disk);
1994
1995 if (!success) {
1996 char b[BDEVNAME_SIZE];
1997 int abort = 0;
1998 /* Cannot read from anywhere, this block is lost.
1999 * Record a bad block on each device. If that doesn't
2000 * work just disable and interrupt the recovery.
2001 * Don't fail devices as that won't really help.
2002 */
2003 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
2004 mdname(mddev), bio_devname(bio, b),
2005 (unsigned long long)r1_bio->sector);
2006 for (d = 0; d < conf->raid_disks * 2; d++) {
2007 rdev = conf->mirrors[d].rdev;
2008 if (!rdev || test_bit(Faulty, &rdev->flags))
2009 continue;
2010 if (!rdev_set_badblocks(rdev, sect, s, 0))
2011 abort = 1;
2012 }
2013 if (abort) {
2014 conf->recovery_disabled =
2015 mddev->recovery_disabled;
2016 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2017 md_done_sync(mddev, r1_bio->sectors, 0);
2018 put_buf(r1_bio);
2019 return 0;
2020 }
2021 /* Try next page */
2022 sectors -= s;
2023 sect += s;
2024 idx++;
2025 continue;
2026 }
2027
2028 start = d;
2029 /* write it back and re-read */
2030 while (d != r1_bio->read_disk) {
2031 if (d == 0)
2032 d = conf->raid_disks * 2;
2033 d--;
2034 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2035 continue;
2036 rdev = conf->mirrors[d].rdev;
2037 if (r1_sync_page_io(rdev, sect, s,
2038 pages[idx],
2039 WRITE) == 0) {
2040 r1_bio->bios[d]->bi_end_io = NULL;
2041 rdev_dec_pending(rdev, mddev);
2042 }
2043 }
2044 d = start;
2045 while (d != r1_bio->read_disk) {
2046 if (d == 0)
2047 d = conf->raid_disks * 2;
2048 d--;
2049 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2050 continue;
2051 rdev = conf->mirrors[d].rdev;
2052 if (r1_sync_page_io(rdev, sect, s,
2053 pages[idx],
2054 READ) != 0)
2055 atomic_add(s, &rdev->corrected_errors);
2056 }
2057 sectors -= s;
2058 sect += s;
2059 idx ++;
2060 }
2061 set_bit(R1BIO_Uptodate, &r1_bio->state);
2062 bio->bi_status = 0;
2063 return 1;
2064}
2065
2066static void process_checks(struct r1bio *r1_bio)
2067{
2068 /* We have read all readable devices. If we haven't
2069 * got the block, then there is no hope left.
2070 * If we have, then we want to do a comparison
2071 * and skip the write if everything is the same.
2072 * If any blocks failed to read, then we need to
2073 * attempt an over-write
2074 */
2075 struct mddev *mddev = r1_bio->mddev;
2076 struct r1conf *conf = mddev->private;
2077 int primary;
2078 int i;
2079 int vcnt;
2080
2081 /* Fix variable parts of all bios */
2082 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2083 for (i = 0; i < conf->raid_disks * 2; i++) {
2084 blk_status_t status;
2085 struct bio *b = r1_bio->bios[i];
2086 struct resync_pages *rp = get_resync_pages(b);
2087 if (b->bi_end_io != end_sync_read)
2088 continue;
2089 /* fixup the bio for reuse, but preserve errno */
2090 status = b->bi_status;
2091 bio_reset(b);
2092 b->bi_status = status;
2093 b->bi_iter.bi_sector = r1_bio->sector +
2094 conf->mirrors[i].rdev->data_offset;
2095 bio_set_dev(b, conf->mirrors[i].rdev->bdev);
2096 b->bi_end_io = end_sync_read;
2097 rp->raid_bio = r1_bio;
2098 b->bi_private = rp;
2099
2100 /* initialize bvec table again */
2101 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2102 }
2103 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2104 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2105 !r1_bio->bios[primary]->bi_status) {
2106 r1_bio->bios[primary]->bi_end_io = NULL;
2107 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2108 break;
2109 }
2110 r1_bio->read_disk = primary;
2111 for (i = 0; i < conf->raid_disks * 2; i++) {
2112 int j;
2113 struct bio *pbio = r1_bio->bios[primary];
2114 struct bio *sbio = r1_bio->bios[i];
2115 blk_status_t status = sbio->bi_status;
2116 struct page **ppages = get_resync_pages(pbio)->pages;
2117 struct page **spages = get_resync_pages(sbio)->pages;
2118 struct bio_vec *bi;
2119 int page_len[RESYNC_PAGES] = { 0 };
2120
2121 if (sbio->bi_end_io != end_sync_read)
2122 continue;
2123 /* Now we can 'fixup' the error value */
2124 sbio->bi_status = 0;
2125
2126 bio_for_each_segment_all(bi, sbio, j)
2127 page_len[j] = bi->bv_len;
2128
2129 if (!status) {
2130 for (j = vcnt; j-- ; ) {
2131 if (memcmp(page_address(ppages[j]),
2132 page_address(spages[j]),
2133 page_len[j]))
2134 break;
2135 }
2136 } else
2137 j = 0;
2138 if (j >= 0)
2139 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2140 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2141 && !status)) {
2142 /* No need to write to this device. */
2143 sbio->bi_end_io = NULL;
2144 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2145 continue;
2146 }
2147
2148 bio_copy_data(sbio, pbio);
2149 }
2150}
2151
2152static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2153{
2154 struct r1conf *conf = mddev->private;
2155 int i;
2156 int disks = conf->raid_disks * 2;
2157 struct bio *wbio;
2158
2159 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2160 /* ouch - failed to read all of that. */
2161 if (!fix_sync_read_error(r1_bio))
2162 return;
2163
2164 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2165 process_checks(r1_bio);
2166
2167 /*
2168 * schedule writes
2169 */
2170 atomic_set(&r1_bio->remaining, 1);
2171 for (i = 0; i < disks ; i++) {
2172 wbio = r1_bio->bios[i];
2173 if (wbio->bi_end_io == NULL ||
2174 (wbio->bi_end_io == end_sync_read &&
2175 (i == r1_bio->read_disk ||
2176 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2177 continue;
2178 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2179 continue;
2180
2181 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2182 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2183 wbio->bi_opf |= MD_FAILFAST;
2184
2185 wbio->bi_end_io = end_sync_write;
2186 atomic_inc(&r1_bio->remaining);
2187 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2188
2189 generic_make_request(wbio);
2190 }
2191
2192 if (atomic_dec_and_test(&r1_bio->remaining)) {
2193 /* if we're here, all write(s) have completed, so clean up */
2194 int s = r1_bio->sectors;
2195 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2196 test_bit(R1BIO_WriteError, &r1_bio->state))
2197 reschedule_retry(r1_bio);
2198 else {
2199 put_buf(r1_bio);
2200 md_done_sync(mddev, s, 1);
2201 }
2202 }
2203}
2204
2205/*
2206 * This is a kernel thread which:
2207 *
2208 * 1. Retries failed read operations on working mirrors.
2209 * 2. Updates the raid superblock when problems encounter.
2210 * 3. Performs writes following reads for array synchronising.
2211 */
2212
2213static void fix_read_error(struct r1conf *conf, int read_disk,
2214 sector_t sect, int sectors)
2215{
2216 struct mddev *mddev = conf->mddev;
2217 while(sectors) {
2218 int s = sectors;
2219 int d = read_disk;
2220 int success = 0;
2221 int start;
2222 struct md_rdev *rdev;
2223
2224 if (s > (PAGE_SIZE>>9))
2225 s = PAGE_SIZE >> 9;
2226
2227 do {
2228 sector_t first_bad;
2229 int bad_sectors;
2230
2231 rcu_read_lock();
2232 rdev = rcu_dereference(conf->mirrors[d].rdev);
2233 if (rdev &&
2234 (test_bit(In_sync, &rdev->flags) ||
2235 (!test_bit(Faulty, &rdev->flags) &&
2236 rdev->recovery_offset >= sect + s)) &&
2237 is_badblock(rdev, sect, s,
2238 &first_bad, &bad_sectors) == 0) {
2239 atomic_inc(&rdev->nr_pending);
2240 rcu_read_unlock();
2241 if (sync_page_io(rdev, sect, s<<9,
2242 conf->tmppage, REQ_OP_READ, 0, false))
2243 success = 1;
2244 rdev_dec_pending(rdev, mddev);
2245 if (success)
2246 break;
2247 } else
2248 rcu_read_unlock();
2249 d++;
2250 if (d == conf->raid_disks * 2)
2251 d = 0;
2252 } while (!success && d != read_disk);
2253
2254 if (!success) {
2255 /* Cannot read from anywhere - mark it bad */
2256 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2257 if (!rdev_set_badblocks(rdev, sect, s, 0))
2258 md_error(mddev, rdev);
2259 break;
2260 }
2261 /* write it back and re-read */
2262 start = d;
2263 while (d != read_disk) {
2264 if (d==0)
2265 d = conf->raid_disks * 2;
2266 d--;
2267 rcu_read_lock();
2268 rdev = rcu_dereference(conf->mirrors[d].rdev);
2269 if (rdev &&
2270 !test_bit(Faulty, &rdev->flags)) {
2271 atomic_inc(&rdev->nr_pending);
2272 rcu_read_unlock();
2273 r1_sync_page_io(rdev, sect, s,
2274 conf->tmppage, WRITE);
2275 rdev_dec_pending(rdev, mddev);
2276 } else
2277 rcu_read_unlock();
2278 }
2279 d = start;
2280 while (d != read_disk) {
2281 char b[BDEVNAME_SIZE];
2282 if (d==0)
2283 d = conf->raid_disks * 2;
2284 d--;
2285 rcu_read_lock();
2286 rdev = rcu_dereference(conf->mirrors[d].rdev);
2287 if (rdev &&
2288 !test_bit(Faulty, &rdev->flags)) {
2289 atomic_inc(&rdev->nr_pending);
2290 rcu_read_unlock();
2291 if (r1_sync_page_io(rdev, sect, s,
2292 conf->tmppage, READ)) {
2293 atomic_add(s, &rdev->corrected_errors);
2294 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %s)\n",
2295 mdname(mddev), s,
2296 (unsigned long long)(sect +
2297 rdev->data_offset),
2298 bdevname(rdev->bdev, b));
2299 }
2300 rdev_dec_pending(rdev, mddev);
2301 } else
2302 rcu_read_unlock();
2303 }
2304 sectors -= s;
2305 sect += s;
2306 }
2307}
2308
2309static int narrow_write_error(struct r1bio *r1_bio, int i)
2310{
2311 struct mddev *mddev = r1_bio->mddev;
2312 struct r1conf *conf = mddev->private;
2313 struct md_rdev *rdev = conf->mirrors[i].rdev;
2314
2315 /* bio has the data to be written to device 'i' where
2316 * we just recently had a write error.
2317 * We repeatedly clone the bio and trim down to one block,
2318 * then try the write. Where the write fails we record
2319 * a bad block.
2320 * It is conceivable that the bio doesn't exactly align with
2321 * blocks. We must handle this somehow.
2322 *
2323 * We currently own a reference on the rdev.
2324 */
2325
2326 int block_sectors;
2327 sector_t sector;
2328 int sectors;
2329 int sect_to_write = r1_bio->sectors;
2330 int ok = 1;
2331
2332 if (rdev->badblocks.shift < 0)
2333 return 0;
2334
2335 block_sectors = roundup(1 << rdev->badblocks.shift,
2336 bdev_logical_block_size(rdev->bdev) >> 9);
2337 sector = r1_bio->sector;
2338 sectors = ((sector + block_sectors)
2339 & ~(sector_t)(block_sectors - 1))
2340 - sector;
2341
2342 while (sect_to_write) {
2343 struct bio *wbio;
2344 if (sectors > sect_to_write)
2345 sectors = sect_to_write;
2346 /* Write at 'sector' for 'sectors'*/
2347
2348 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2349 wbio = bio_clone_fast(r1_bio->behind_master_bio,
2350 GFP_NOIO,
2351 mddev->bio_set);
2352 } else {
2353 wbio = bio_clone_fast(r1_bio->master_bio, GFP_NOIO,
2354 mddev->bio_set);
2355 }
2356
2357 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2358 wbio->bi_iter.bi_sector = r1_bio->sector;
2359 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2360
2361 bio_trim(wbio, sector - r1_bio->sector, sectors);
2362 wbio->bi_iter.bi_sector += rdev->data_offset;
2363 bio_set_dev(wbio, rdev->bdev);
2364
2365 if (submit_bio_wait(wbio) < 0)
2366 /* failure! */
2367 ok = rdev_set_badblocks(rdev, sector,
2368 sectors, 0)
2369 && ok;
2370
2371 bio_put(wbio);
2372 sect_to_write -= sectors;
2373 sector += sectors;
2374 sectors = block_sectors;
2375 }
2376 return ok;
2377}
2378
2379static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2380{
2381 int m;
2382 int s = r1_bio->sectors;
2383 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2384 struct md_rdev *rdev = conf->mirrors[m].rdev;
2385 struct bio *bio = r1_bio->bios[m];
2386 if (bio->bi_end_io == NULL)
2387 continue;
2388 if (!bio->bi_status &&
2389 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2390 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2391 }
2392 if (bio->bi_status &&
2393 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2394 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2395 md_error(conf->mddev, rdev);
2396 }
2397 }
2398 put_buf(r1_bio);
2399 md_done_sync(conf->mddev, s, 1);
2400}
2401
2402static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2403{
2404 int m, idx;
2405 bool fail = false;
2406
2407 for (m = 0; m < conf->raid_disks * 2 ; m++)
2408 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2409 struct md_rdev *rdev = conf->mirrors[m].rdev;
2410 rdev_clear_badblocks(rdev,
2411 r1_bio->sector,
2412 r1_bio->sectors, 0);
2413 rdev_dec_pending(rdev, conf->mddev);
2414 } else if (r1_bio->bios[m] != NULL) {
2415 /* This drive got a write error. We need to
2416 * narrow down and record precise write
2417 * errors.
2418 */
2419 fail = true;
2420 if (!narrow_write_error(r1_bio, m)) {
2421 md_error(conf->mddev,
2422 conf->mirrors[m].rdev);
2423 /* an I/O failed, we can't clear the bitmap */
2424 set_bit(R1BIO_Degraded, &r1_bio->state);
2425 }
2426 rdev_dec_pending(conf->mirrors[m].rdev,
2427 conf->mddev);
2428 }
2429 if (fail) {
2430 spin_lock_irq(&conf->device_lock);
2431 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2432 idx = sector_to_idx(r1_bio->sector);
2433 atomic_inc(&conf->nr_queued[idx]);
2434 spin_unlock_irq(&conf->device_lock);
2435 /*
2436 * In case freeze_array() is waiting for condition
2437 * get_unqueued_pending() == extra to be true.
2438 */
2439 wake_up(&conf->wait_barrier);
2440 md_wakeup_thread(conf->mddev->thread);
2441 } else {
2442 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2443 close_write(r1_bio);
2444 raid_end_bio_io(r1_bio);
2445 }
2446}
2447
2448static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2449{
2450 struct mddev *mddev = conf->mddev;
2451 struct bio *bio;
2452 struct md_rdev *rdev;
2453 sector_t bio_sector;
2454
2455 clear_bit(R1BIO_ReadError, &r1_bio->state);
2456 /* we got a read error. Maybe the drive is bad. Maybe just
2457 * the block and we can fix it.
2458 * We freeze all other IO, and try reading the block from
2459 * other devices. When we find one, we re-write
2460 * and check it that fixes the read error.
2461 * This is all done synchronously while the array is
2462 * frozen
2463 */
2464
2465 bio = r1_bio->bios[r1_bio->read_disk];
2466 bio_sector = conf->mirrors[r1_bio->read_disk].rdev->data_offset + r1_bio->sector;
2467 bio_put(bio);
2468 r1_bio->bios[r1_bio->read_disk] = NULL;
2469
2470 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2471 if (mddev->ro == 0
2472 && !test_bit(FailFast, &rdev->flags)) {
2473 freeze_array(conf, 1);
2474 fix_read_error(conf, r1_bio->read_disk,
2475 r1_bio->sector, r1_bio->sectors);
2476 unfreeze_array(conf);
2477 } else {
2478 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2479 }
2480
2481 rdev_dec_pending(rdev, conf->mddev);
2482 allow_barrier(conf, r1_bio->sector);
2483 bio = r1_bio->master_bio;
2484
2485 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2486 r1_bio->state = 0;
2487 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2488}
2489
2490static void raid1d(struct md_thread *thread)
2491{
2492 struct mddev *mddev = thread->mddev;
2493 struct r1bio *r1_bio;
2494 unsigned long flags;
2495 struct r1conf *conf = mddev->private;
2496 struct list_head *head = &conf->retry_list;
2497 struct blk_plug plug;
2498 int idx;
2499
2500 md_check_recovery(mddev);
2501
2502 if (!list_empty_careful(&conf->bio_end_io_list) &&
2503 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2504 LIST_HEAD(tmp);
2505 spin_lock_irqsave(&conf->device_lock, flags);
2506 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2507 list_splice_init(&conf->bio_end_io_list, &tmp);
2508 spin_unlock_irqrestore(&conf->device_lock, flags);
2509 while (!list_empty(&tmp)) {
2510 r1_bio = list_first_entry(&tmp, struct r1bio,
2511 retry_list);
2512 list_del(&r1_bio->retry_list);
2513 idx = sector_to_idx(r1_bio->sector);
2514 atomic_dec(&conf->nr_queued[idx]);
2515 if (mddev->degraded)
2516 set_bit(R1BIO_Degraded, &r1_bio->state);
2517 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2518 close_write(r1_bio);
2519 raid_end_bio_io(r1_bio);
2520 }
2521 }
2522
2523 blk_start_plug(&plug);
2524 for (;;) {
2525
2526 flush_pending_writes(conf);
2527
2528 spin_lock_irqsave(&conf->device_lock, flags);
2529 if (list_empty(head)) {
2530 spin_unlock_irqrestore(&conf->device_lock, flags);
2531 break;
2532 }
2533 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2534 list_del(head->prev);
2535 idx = sector_to_idx(r1_bio->sector);
2536 atomic_dec(&conf->nr_queued[idx]);
2537 spin_unlock_irqrestore(&conf->device_lock, flags);
2538
2539 mddev = r1_bio->mddev;
2540 conf = mddev->private;
2541 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2542 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2543 test_bit(R1BIO_WriteError, &r1_bio->state))
2544 handle_sync_write_finished(conf, r1_bio);
2545 else
2546 sync_request_write(mddev, r1_bio);
2547 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2548 test_bit(R1BIO_WriteError, &r1_bio->state))
2549 handle_write_finished(conf, r1_bio);
2550 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2551 handle_read_error(conf, r1_bio);
2552 else
2553 WARN_ON_ONCE(1);
2554
2555 cond_resched();
2556 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2557 md_check_recovery(mddev);
2558 }
2559 blk_finish_plug(&plug);
2560}
2561
2562static int init_resync(struct r1conf *conf)
2563{
2564 int buffs;
2565
2566 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2567 BUG_ON(conf->r1buf_pool);
2568 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2569 conf->poolinfo);
2570 if (!conf->r1buf_pool)
2571 return -ENOMEM;
2572 return 0;
2573}
2574
2575static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2576{
2577 struct r1bio *r1bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2578 struct resync_pages *rps;
2579 struct bio *bio;
2580 int i;
2581
2582 for (i = conf->poolinfo->raid_disks; i--; ) {
2583 bio = r1bio->bios[i];
2584 rps = bio->bi_private;
2585 bio_reset(bio);
2586 bio->bi_private = rps;
2587 }
2588 r1bio->master_bio = NULL;
2589 return r1bio;
2590}
2591
2592/*
2593 * perform a "sync" on one "block"
2594 *
2595 * We need to make sure that no normal I/O request - particularly write
2596 * requests - conflict with active sync requests.
2597 *
2598 * This is achieved by tracking pending requests and a 'barrier' concept
2599 * that can be installed to exclude normal IO requests.
2600 */
2601
2602static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2603 int *skipped)
2604{
2605 struct r1conf *conf = mddev->private;
2606 struct r1bio *r1_bio;
2607 struct bio *bio;
2608 sector_t max_sector, nr_sectors;
2609 int disk = -1;
2610 int i;
2611 int wonly = -1;
2612 int write_targets = 0, read_targets = 0;
2613 sector_t sync_blocks;
2614 int still_degraded = 0;
2615 int good_sectors = RESYNC_SECTORS;
2616 int min_bad = 0; /* number of sectors that are bad in all devices */
2617 int idx = sector_to_idx(sector_nr);
2618 int page_idx = 0;
2619
2620 if (!conf->r1buf_pool)
2621 if (init_resync(conf))
2622 return 0;
2623
2624 max_sector = mddev->dev_sectors;
2625 if (sector_nr >= max_sector) {
2626 /* If we aborted, we need to abort the
2627 * sync on the 'current' bitmap chunk (there will
2628 * only be one in raid1 resync.
2629 * We can find the current addess in mddev->curr_resync
2630 */
2631 if (mddev->curr_resync < max_sector) /* aborted */
2632 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2633 &sync_blocks, 1);
2634 else /* completed sync */
2635 conf->fullsync = 0;
2636
2637 bitmap_close_sync(mddev->bitmap);
2638 close_sync(conf);
2639
2640 if (mddev_is_clustered(mddev)) {
2641 conf->cluster_sync_low = 0;
2642 conf->cluster_sync_high = 0;
2643 }
2644 return 0;
2645 }
2646
2647 if (mddev->bitmap == NULL &&
2648 mddev->recovery_cp == MaxSector &&
2649 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2650 conf->fullsync == 0) {
2651 *skipped = 1;
2652 return max_sector - sector_nr;
2653 }
2654 /* before building a request, check if we can skip these blocks..
2655 * This call the bitmap_start_sync doesn't actually record anything
2656 */
2657 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2658 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2659 /* We can skip this block, and probably several more */
2660 *skipped = 1;
2661 return sync_blocks;
2662 }
2663
2664 /*
2665 * If there is non-resync activity waiting for a turn, then let it
2666 * though before starting on this new sync request.
2667 */
2668 if (atomic_read(&conf->nr_waiting[idx]))
2669 schedule_timeout_uninterruptible(1);
2670
2671 /* we are incrementing sector_nr below. To be safe, we check against
2672 * sector_nr + two times RESYNC_SECTORS
2673 */
2674
2675 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2676 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2677
2678
2679 if (raise_barrier(conf, sector_nr))
2680 return 0;
2681
2682 r1_bio = raid1_alloc_init_r1buf(conf);
2683
2684 rcu_read_lock();
2685 /*
2686 * If we get a correctably read error during resync or recovery,
2687 * we might want to read from a different device. So we
2688 * flag all drives that could conceivably be read from for READ,
2689 * and any others (which will be non-In_sync devices) for WRITE.
2690 * If a read fails, we try reading from something else for which READ
2691 * is OK.
2692 */
2693
2694 r1_bio->mddev = mddev;
2695 r1_bio->sector = sector_nr;
2696 r1_bio->state = 0;
2697 set_bit(R1BIO_IsSync, &r1_bio->state);
2698 /* make sure good_sectors won't go across barrier unit boundary */
2699 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2700
2701 for (i = 0; i < conf->raid_disks * 2; i++) {
2702 struct md_rdev *rdev;
2703 bio = r1_bio->bios[i];
2704
2705 rdev = rcu_dereference(conf->mirrors[i].rdev);
2706 if (rdev == NULL ||
2707 test_bit(Faulty, &rdev->flags)) {
2708 if (i < conf->raid_disks)
2709 still_degraded = 1;
2710 } else if (!test_bit(In_sync, &rdev->flags)) {
2711 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2712 bio->bi_end_io = end_sync_write;
2713 write_targets ++;
2714 } else {
2715 /* may need to read from here */
2716 sector_t first_bad = MaxSector;
2717 int bad_sectors;
2718
2719 if (is_badblock(rdev, sector_nr, good_sectors,
2720 &first_bad, &bad_sectors)) {
2721 if (first_bad > sector_nr)
2722 good_sectors = first_bad - sector_nr;
2723 else {
2724 bad_sectors -= (sector_nr - first_bad);
2725 if (min_bad == 0 ||
2726 min_bad > bad_sectors)
2727 min_bad = bad_sectors;
2728 }
2729 }
2730 if (sector_nr < first_bad) {
2731 if (test_bit(WriteMostly, &rdev->flags)) {
2732 if (wonly < 0)
2733 wonly = i;
2734 } else {
2735 if (disk < 0)
2736 disk = i;
2737 }
2738 bio_set_op_attrs(bio, REQ_OP_READ, 0);
2739 bio->bi_end_io = end_sync_read;
2740 read_targets++;
2741 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2742 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2743 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2744 /*
2745 * The device is suitable for reading (InSync),
2746 * but has bad block(s) here. Let's try to correct them,
2747 * if we are doing resync or repair. Otherwise, leave
2748 * this device alone for this sync request.
2749 */
2750 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
2751 bio->bi_end_io = end_sync_write;
2752 write_targets++;
2753 }
2754 }
2755 if (bio->bi_end_io) {
2756 atomic_inc(&rdev->nr_pending);
2757 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2758 bio_set_dev(bio, rdev->bdev);
2759 if (test_bit(FailFast, &rdev->flags))
2760 bio->bi_opf |= MD_FAILFAST;
2761 }
2762 }
2763 rcu_read_unlock();
2764 if (disk < 0)
2765 disk = wonly;
2766 r1_bio->read_disk = disk;
2767
2768 if (read_targets == 0 && min_bad > 0) {
2769 /* These sectors are bad on all InSync devices, so we
2770 * need to mark them bad on all write targets
2771 */
2772 int ok = 1;
2773 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2774 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2775 struct md_rdev *rdev = conf->mirrors[i].rdev;
2776 ok = rdev_set_badblocks(rdev, sector_nr,
2777 min_bad, 0
2778 ) && ok;
2779 }
2780 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2781 *skipped = 1;
2782 put_buf(r1_bio);
2783
2784 if (!ok) {
2785 /* Cannot record the badblocks, so need to
2786 * abort the resync.
2787 * If there are multiple read targets, could just
2788 * fail the really bad ones ???
2789 */
2790 conf->recovery_disabled = mddev->recovery_disabled;
2791 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2792 return 0;
2793 } else
2794 return min_bad;
2795
2796 }
2797 if (min_bad > 0 && min_bad < good_sectors) {
2798 /* only resync enough to reach the next bad->good
2799 * transition */
2800 good_sectors = min_bad;
2801 }
2802
2803 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2804 /* extra read targets are also write targets */
2805 write_targets += read_targets-1;
2806
2807 if (write_targets == 0 || read_targets == 0) {
2808 /* There is nowhere to write, so all non-sync
2809 * drives must be failed - so we are finished
2810 */
2811 sector_t rv;
2812 if (min_bad > 0)
2813 max_sector = sector_nr + min_bad;
2814 rv = max_sector - sector_nr;
2815 *skipped = 1;
2816 put_buf(r1_bio);
2817 return rv;
2818 }
2819
2820 if (max_sector > mddev->resync_max)
2821 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2822 if (max_sector > sector_nr + good_sectors)
2823 max_sector = sector_nr + good_sectors;
2824 nr_sectors = 0;
2825 sync_blocks = 0;
2826 do {
2827 struct page *page;
2828 int len = PAGE_SIZE;
2829 if (sector_nr + (len>>9) > max_sector)
2830 len = (max_sector - sector_nr) << 9;
2831 if (len == 0)
2832 break;
2833 if (sync_blocks == 0) {
2834 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2835 &sync_blocks, still_degraded) &&
2836 !conf->fullsync &&
2837 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2838 break;
2839 if ((len >> 9) > sync_blocks)
2840 len = sync_blocks<<9;
2841 }
2842
2843 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2844 struct resync_pages *rp;
2845
2846 bio = r1_bio->bios[i];
2847 rp = get_resync_pages(bio);
2848 if (bio->bi_end_io) {
2849 page = resync_fetch_page(rp, page_idx);
2850
2851 /*
2852 * won't fail because the vec table is big
2853 * enough to hold all these pages
2854 */
2855 bio_add_page(bio, page, len, 0);
2856 }
2857 }
2858 nr_sectors += len>>9;
2859 sector_nr += len>>9;
2860 sync_blocks -= (len>>9);
2861 } while (++page_idx < RESYNC_PAGES);
2862
2863 r1_bio->sectors = nr_sectors;
2864
2865 if (mddev_is_clustered(mddev) &&
2866 conf->cluster_sync_high < sector_nr + nr_sectors) {
2867 conf->cluster_sync_low = mddev->curr_resync_completed;
2868 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
2869 /* Send resync message */
2870 md_cluster_ops->resync_info_update(mddev,
2871 conf->cluster_sync_low,
2872 conf->cluster_sync_high);
2873 }
2874
2875 /* For a user-requested sync, we read all readable devices and do a
2876 * compare
2877 */
2878 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2879 atomic_set(&r1_bio->remaining, read_targets);
2880 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2881 bio = r1_bio->bios[i];
2882 if (bio->bi_end_io == end_sync_read) {
2883 read_targets--;
2884 md_sync_acct_bio(bio, nr_sectors);
2885 if (read_targets == 1)
2886 bio->bi_opf &= ~MD_FAILFAST;
2887 generic_make_request(bio);
2888 }
2889 }
2890 } else {
2891 atomic_set(&r1_bio->remaining, 1);
2892 bio = r1_bio->bios[r1_bio->read_disk];
2893 md_sync_acct_bio(bio, nr_sectors);
2894 if (read_targets == 1)
2895 bio->bi_opf &= ~MD_FAILFAST;
2896 generic_make_request(bio);
2897
2898 }
2899 return nr_sectors;
2900}
2901
2902static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2903{
2904 if (sectors)
2905 return sectors;
2906
2907 return mddev->dev_sectors;
2908}
2909
2910static struct r1conf *setup_conf(struct mddev *mddev)
2911{
2912 struct r1conf *conf;
2913 int i;
2914 struct raid1_info *disk;
2915 struct md_rdev *rdev;
2916 int err = -ENOMEM;
2917
2918 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2919 if (!conf)
2920 goto abort;
2921
2922 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
2923 sizeof(atomic_t), GFP_KERNEL);
2924 if (!conf->nr_pending)
2925 goto abort;
2926
2927 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
2928 sizeof(atomic_t), GFP_KERNEL);
2929 if (!conf->nr_waiting)
2930 goto abort;
2931
2932 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
2933 sizeof(atomic_t), GFP_KERNEL);
2934 if (!conf->nr_queued)
2935 goto abort;
2936
2937 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
2938 sizeof(atomic_t), GFP_KERNEL);
2939 if (!conf->barrier)
2940 goto abort;
2941
2942 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2943 * mddev->raid_disks * 2,
2944 GFP_KERNEL);
2945 if (!conf->mirrors)
2946 goto abort;
2947
2948 conf->tmppage = alloc_page(GFP_KERNEL);
2949 if (!conf->tmppage)
2950 goto abort;
2951
2952 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2953 if (!conf->poolinfo)
2954 goto abort;
2955 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2956 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2957 r1bio_pool_free,
2958 conf->poolinfo);
2959 if (!conf->r1bio_pool)
2960 goto abort;
2961
2962 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
2963 if (!conf->bio_split)
2964 goto abort;
2965
2966 conf->poolinfo->mddev = mddev;
2967
2968 err = -EINVAL;
2969 spin_lock_init(&conf->device_lock);
2970 rdev_for_each(rdev, mddev) {
2971 int disk_idx = rdev->raid_disk;
2972 if (disk_idx >= mddev->raid_disks
2973 || disk_idx < 0)
2974 continue;
2975 if (test_bit(Replacement, &rdev->flags))
2976 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2977 else
2978 disk = conf->mirrors + disk_idx;
2979
2980 if (disk->rdev)
2981 goto abort;
2982 disk->rdev = rdev;
2983 disk->head_position = 0;
2984 disk->seq_start = MaxSector;
2985 }
2986 conf->raid_disks = mddev->raid_disks;
2987 conf->mddev = mddev;
2988 INIT_LIST_HEAD(&conf->retry_list);
2989 INIT_LIST_HEAD(&conf->bio_end_io_list);
2990
2991 spin_lock_init(&conf->resync_lock);
2992 init_waitqueue_head(&conf->wait_barrier);
2993
2994 bio_list_init(&conf->pending_bio_list);
2995 conf->pending_count = 0;
2996 conf->recovery_disabled = mddev->recovery_disabled - 1;
2997
2998 err = -EIO;
2999 for (i = 0; i < conf->raid_disks * 2; i++) {
3000
3001 disk = conf->mirrors + i;
3002
3003 if (i < conf->raid_disks &&
3004 disk[conf->raid_disks].rdev) {
3005 /* This slot has a replacement. */
3006 if (!disk->rdev) {
3007 /* No original, just make the replacement
3008 * a recovering spare
3009 */
3010 disk->rdev =
3011 disk[conf->raid_disks].rdev;
3012 disk[conf->raid_disks].rdev = NULL;
3013 } else if (!test_bit(In_sync, &disk->rdev->flags))
3014 /* Original is not in_sync - bad */
3015 goto abort;
3016 }
3017
3018 if (!disk->rdev ||
3019 !test_bit(In_sync, &disk->rdev->flags)) {
3020 disk->head_position = 0;
3021 if (disk->rdev &&
3022 (disk->rdev->saved_raid_disk < 0))
3023 conf->fullsync = 1;
3024 }
3025 }
3026
3027 err = -ENOMEM;
3028 conf->thread = md_register_thread(raid1d, mddev, "raid1");
3029 if (!conf->thread)
3030 goto abort;
3031
3032 return conf;
3033
3034 abort:
3035 if (conf) {
3036 mempool_destroy(conf->r1bio_pool);
3037 kfree(conf->mirrors);
3038 safe_put_page(conf->tmppage);
3039 kfree(conf->poolinfo);
3040 kfree(conf->nr_pending);
3041 kfree(conf->nr_waiting);
3042 kfree(conf->nr_queued);
3043 kfree(conf->barrier);
3044 if (conf->bio_split)
3045 bioset_free(conf->bio_split);
3046 kfree(conf);
3047 }
3048 return ERR_PTR(err);
3049}
3050
3051static void raid1_free(struct mddev *mddev, void *priv);
3052static int raid1_run(struct mddev *mddev)
3053{
3054 struct r1conf *conf;
3055 int i;
3056 struct md_rdev *rdev;
3057 int ret;
3058 bool discard_supported = false;
3059
3060 if (mddev->level != 1) {
3061 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3062 mdname(mddev), mddev->level);
3063 return -EIO;
3064 }
3065 if (mddev->reshape_position != MaxSector) {
3066 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3067 mdname(mddev));
3068 return -EIO;
3069 }
3070 if (mddev_init_writes_pending(mddev) < 0)
3071 return -ENOMEM;
3072 /*
3073 * copy the already verified devices into our private RAID1
3074 * bookkeeping area. [whatever we allocate in run(),
3075 * should be freed in raid1_free()]
3076 */
3077 if (mddev->private == NULL)
3078 conf = setup_conf(mddev);
3079 else
3080 conf = mddev->private;
3081
3082 if (IS_ERR(conf))
3083 return PTR_ERR(conf);
3084
3085 if (mddev->queue) {
3086 blk_queue_max_write_same_sectors(mddev->queue, 0);
3087 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3088 }
3089
3090 rdev_for_each(rdev, mddev) {
3091 if (!mddev->gendisk)
3092 continue;
3093 disk_stack_limits(mddev->gendisk, rdev->bdev,
3094 rdev->data_offset << 9);
3095 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3096 discard_supported = true;
3097 }
3098
3099 mddev->degraded = 0;
3100 for (i=0; i < conf->raid_disks; i++)
3101 if (conf->mirrors[i].rdev == NULL ||
3102 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3103 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3104 mddev->degraded++;
3105
3106 if (conf->raid_disks - mddev->degraded == 1)
3107 mddev->recovery_cp = MaxSector;
3108
3109 if (mddev->recovery_cp != MaxSector)
3110 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3111 mdname(mddev));
3112 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3113 mdname(mddev), mddev->raid_disks - mddev->degraded,
3114 mddev->raid_disks);
3115
3116 /*
3117 * Ok, everything is just fine now
3118 */
3119 mddev->thread = conf->thread;
3120 conf->thread = NULL;
3121 mddev->private = conf;
3122 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3123
3124 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3125
3126 if (mddev->queue) {
3127 if (discard_supported)
3128 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3129 mddev->queue);
3130 else
3131 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3132 mddev->queue);
3133 }
3134
3135 ret = md_integrity_register(mddev);
3136 if (ret) {
3137 md_unregister_thread(&mddev->thread);
3138 raid1_free(mddev, conf);
3139 }
3140 return ret;
3141}
3142
3143static void raid1_free(struct mddev *mddev, void *priv)
3144{
3145 struct r1conf *conf = priv;
3146
3147 mempool_destroy(conf->r1bio_pool);
3148 kfree(conf->mirrors);
3149 safe_put_page(conf->tmppage);
3150 kfree(conf->poolinfo);
3151 kfree(conf->nr_pending);
3152 kfree(conf->nr_waiting);
3153 kfree(conf->nr_queued);
3154 kfree(conf->barrier);
3155 if (conf->bio_split)
3156 bioset_free(conf->bio_split);
3157 kfree(conf);
3158}
3159
3160static int raid1_resize(struct mddev *mddev, sector_t sectors)
3161{
3162 /* no resync is happening, and there is enough space
3163 * on all devices, so we can resize.
3164 * We need to make sure resync covers any new space.
3165 * If the array is shrinking we should possibly wait until
3166 * any io in the removed space completes, but it hardly seems
3167 * worth it.
3168 */
3169 sector_t newsize = raid1_size(mddev, sectors, 0);
3170 if (mddev->external_size &&
3171 mddev->array_sectors > newsize)
3172 return -EINVAL;
3173 if (mddev->bitmap) {
3174 int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
3175 if (ret)
3176 return ret;
3177 }
3178 md_set_array_sectors(mddev, newsize);
3179 if (sectors > mddev->dev_sectors &&
3180 mddev->recovery_cp > mddev->dev_sectors) {
3181 mddev->recovery_cp = mddev->dev_sectors;
3182 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3183 }
3184 mddev->dev_sectors = sectors;
3185 mddev->resync_max_sectors = sectors;
3186 return 0;
3187}
3188
3189static int raid1_reshape(struct mddev *mddev)
3190{
3191 /* We need to:
3192 * 1/ resize the r1bio_pool
3193 * 2/ resize conf->mirrors
3194 *
3195 * We allocate a new r1bio_pool if we can.
3196 * Then raise a device barrier and wait until all IO stops.
3197 * Then resize conf->mirrors and swap in the new r1bio pool.
3198 *
3199 * At the same time, we "pack" the devices so that all the missing
3200 * devices have the higher raid_disk numbers.
3201 */
3202 mempool_t *newpool, *oldpool;
3203 struct pool_info *newpoolinfo;
3204 struct raid1_info *newmirrors;
3205 struct r1conf *conf = mddev->private;
3206 int cnt, raid_disks;
3207 unsigned long flags;
3208 int d, d2;
3209
3210 /* Cannot change chunk_size, layout, or level */
3211 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3212 mddev->layout != mddev->new_layout ||
3213 mddev->level != mddev->new_level) {
3214 mddev->new_chunk_sectors = mddev->chunk_sectors;
3215 mddev->new_layout = mddev->layout;
3216 mddev->new_level = mddev->level;
3217 return -EINVAL;
3218 }
3219
3220 if (!mddev_is_clustered(mddev))
3221 md_allow_write(mddev);
3222
3223 raid_disks = mddev->raid_disks + mddev->delta_disks;
3224
3225 if (raid_disks < conf->raid_disks) {
3226 cnt=0;
3227 for (d= 0; d < conf->raid_disks; d++)
3228 if (conf->mirrors[d].rdev)
3229 cnt++;
3230 if (cnt > raid_disks)
3231 return -EBUSY;
3232 }
3233
3234 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3235 if (!newpoolinfo)
3236 return -ENOMEM;
3237 newpoolinfo->mddev = mddev;
3238 newpoolinfo->raid_disks = raid_disks * 2;
3239
3240 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
3241 r1bio_pool_free, newpoolinfo);
3242 if (!newpool) {
3243 kfree(newpoolinfo);
3244 return -ENOMEM;
3245 }
3246 newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3247 GFP_KERNEL);
3248 if (!newmirrors) {
3249 kfree(newpoolinfo);
3250 mempool_destroy(newpool);
3251 return -ENOMEM;
3252 }
3253
3254 freeze_array(conf, 0);
3255
3256 /* ok, everything is stopped */
3257 oldpool = conf->r1bio_pool;
3258 conf->r1bio_pool = newpool;
3259
3260 for (d = d2 = 0; d < conf->raid_disks; d++) {
3261 struct md_rdev *rdev = conf->mirrors[d].rdev;
3262 if (rdev && rdev->raid_disk != d2) {
3263 sysfs_unlink_rdev(mddev, rdev);
3264 rdev->raid_disk = d2;
3265 sysfs_unlink_rdev(mddev, rdev);
3266 if (sysfs_link_rdev(mddev, rdev))
3267 pr_warn("md/raid1:%s: cannot register rd%d\n",
3268 mdname(mddev), rdev->raid_disk);
3269 }
3270 if (rdev)
3271 newmirrors[d2++].rdev = rdev;
3272 }
3273 kfree(conf->mirrors);
3274 conf->mirrors = newmirrors;
3275 kfree(conf->poolinfo);
3276 conf->poolinfo = newpoolinfo;
3277
3278 spin_lock_irqsave(&conf->device_lock, flags);
3279 mddev->degraded += (raid_disks - conf->raid_disks);
3280 spin_unlock_irqrestore(&conf->device_lock, flags);
3281 conf->raid_disks = mddev->raid_disks = raid_disks;
3282 mddev->delta_disks = 0;
3283
3284 unfreeze_array(conf);
3285
3286 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3287 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3288 md_wakeup_thread(mddev->thread);
3289
3290 mempool_destroy(oldpool);
3291 return 0;
3292}
3293
3294static void raid1_quiesce(struct mddev *mddev, int quiesce)
3295{
3296 struct r1conf *conf = mddev->private;
3297
3298 if (quiesce)
3299 freeze_array(conf, 0);
3300 else
3301 unfreeze_array(conf);
3302}
3303
3304static void *raid1_takeover(struct mddev *mddev)
3305{
3306 /* raid1 can take over:
3307 * raid5 with 2 devices, any layout or chunk size
3308 */
3309 if (mddev->level == 5 && mddev->raid_disks == 2) {
3310 struct r1conf *conf;
3311 mddev->new_level = 1;
3312 mddev->new_layout = 0;
3313 mddev->new_chunk_sectors = 0;
3314 conf = setup_conf(mddev);
3315 if (!IS_ERR(conf)) {
3316 /* Array must appear to be quiesced */
3317 conf->array_frozen = 1;
3318 mddev_clear_unsupported_flags(mddev,
3319 UNSUPPORTED_MDDEV_FLAGS);
3320 }
3321 return conf;
3322 }
3323 return ERR_PTR(-EINVAL);
3324}
3325
3326static struct md_personality raid1_personality =
3327{
3328 .name = "raid1",
3329 .level = 1,
3330 .owner = THIS_MODULE,
3331 .make_request = raid1_make_request,
3332 .run = raid1_run,
3333 .free = raid1_free,
3334 .status = raid1_status,
3335 .error_handler = raid1_error,
3336 .hot_add_disk = raid1_add_disk,
3337 .hot_remove_disk= raid1_remove_disk,
3338 .spare_active = raid1_spare_active,
3339 .sync_request = raid1_sync_request,
3340 .resize = raid1_resize,
3341 .size = raid1_size,
3342 .check_reshape = raid1_reshape,
3343 .quiesce = raid1_quiesce,
3344 .takeover = raid1_takeover,
3345 .congested = raid1_congested,
3346};
3347
3348static int __init raid_init(void)
3349{
3350 return register_md_personality(&raid1_personality);
3351}
3352
3353static void raid_exit(void)
3354{
3355 unregister_md_personality(&raid1_personality);
3356}
3357
3358module_init(raid_init);
3359module_exit(raid_exit);
3360MODULE_LICENSE("GPL");
3361MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3362MODULE_ALIAS("md-personality-3"); /* RAID1 */
3363MODULE_ALIAS("md-raid1");
3364MODULE_ALIAS("md-level-1");
3365
3366module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);