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// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * raid1.c : Multiple Devices driver for Linux
4 *
5 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
6 *
7 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
8 *
9 * RAID-1 management functions.
10 *
11 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
12 *
13 * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
14 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
15 *
16 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
17 * bitmapped intelligence in resync:
18 *
19 * - bitmap marked during normal i/o
20 * - bitmap used to skip nondirty blocks during sync
21 *
22 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
23 * - persistent bitmap code
24 */
25
26#include <linux/slab.h>
27#include <linux/delay.h>
28#include <linux/blkdev.h>
29#include <linux/module.h>
30#include <linux/seq_file.h>
31#include <linux/ratelimit.h>
32#include <linux/interval_tree_generic.h>
33
34#include <trace/events/block.h>
35
36#include "md.h"
37#include "raid1.h"
38#include "md-bitmap.h"
39
40#define UNSUPPORTED_MDDEV_FLAGS \
41 ((1L << MD_HAS_JOURNAL) | \
42 (1L << MD_JOURNAL_CLEAN) | \
43 (1L << MD_HAS_PPL) | \
44 (1L << MD_HAS_MULTIPLE_PPLS))
45
46static void allow_barrier(struct r1conf *conf, sector_t sector_nr);
47static void lower_barrier(struct r1conf *conf, sector_t sector_nr);
48
49#define RAID_1_10_NAME "raid1"
50#include "raid1-10.c"
51
52#define START(node) ((node)->start)
53#define LAST(node) ((node)->last)
54INTERVAL_TREE_DEFINE(struct serial_info, node, sector_t, _subtree_last,
55 START, LAST, static inline, raid1_rb);
56
57static int check_and_add_serial(struct md_rdev *rdev, struct r1bio *r1_bio,
58 struct serial_info *si, int idx)
59{
60 unsigned long flags;
61 int ret = 0;
62 sector_t lo = r1_bio->sector;
63 sector_t hi = lo + r1_bio->sectors;
64 struct serial_in_rdev *serial = &rdev->serial[idx];
65
66 spin_lock_irqsave(&serial->serial_lock, flags);
67 /* collision happened */
68 if (raid1_rb_iter_first(&serial->serial_rb, lo, hi))
69 ret = -EBUSY;
70 else {
71 si->start = lo;
72 si->last = hi;
73 raid1_rb_insert(si, &serial->serial_rb);
74 }
75 spin_unlock_irqrestore(&serial->serial_lock, flags);
76
77 return ret;
78}
79
80static void wait_for_serialization(struct md_rdev *rdev, struct r1bio *r1_bio)
81{
82 struct mddev *mddev = rdev->mddev;
83 struct serial_info *si;
84 int idx = sector_to_idx(r1_bio->sector);
85 struct serial_in_rdev *serial = &rdev->serial[idx];
86
87 if (WARN_ON(!mddev->serial_info_pool))
88 return;
89 si = mempool_alloc(mddev->serial_info_pool, GFP_NOIO);
90 wait_event(serial->serial_io_wait,
91 check_and_add_serial(rdev, r1_bio, si, idx) == 0);
92}
93
94static void remove_serial(struct md_rdev *rdev, sector_t lo, sector_t hi)
95{
96 struct serial_info *si;
97 unsigned long flags;
98 int found = 0;
99 struct mddev *mddev = rdev->mddev;
100 int idx = sector_to_idx(lo);
101 struct serial_in_rdev *serial = &rdev->serial[idx];
102
103 spin_lock_irqsave(&serial->serial_lock, flags);
104 for (si = raid1_rb_iter_first(&serial->serial_rb, lo, hi);
105 si; si = raid1_rb_iter_next(si, lo, hi)) {
106 if (si->start == lo && si->last == hi) {
107 raid1_rb_remove(si, &serial->serial_rb);
108 mempool_free(si, mddev->serial_info_pool);
109 found = 1;
110 break;
111 }
112 }
113 if (!found)
114 WARN(1, "The write IO is not recorded for serialization\n");
115 spin_unlock_irqrestore(&serial->serial_lock, flags);
116 wake_up(&serial->serial_io_wait);
117}
118
119/*
120 * for resync bio, r1bio pointer can be retrieved from the per-bio
121 * 'struct resync_pages'.
122 */
123static inline struct r1bio *get_resync_r1bio(struct bio *bio)
124{
125 return get_resync_pages(bio)->raid_bio;
126}
127
128static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
129{
130 struct pool_info *pi = data;
131 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
132
133 /* allocate a r1bio with room for raid_disks entries in the bios array */
134 return kzalloc(size, gfp_flags);
135}
136
137#define RESYNC_DEPTH 32
138#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
139#define RESYNC_WINDOW (RESYNC_BLOCK_SIZE * RESYNC_DEPTH)
140#define RESYNC_WINDOW_SECTORS (RESYNC_WINDOW >> 9)
141#define CLUSTER_RESYNC_WINDOW (16 * RESYNC_WINDOW)
142#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
143
144static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
145{
146 struct pool_info *pi = data;
147 struct r1bio *r1_bio;
148 struct bio *bio;
149 int need_pages;
150 int j;
151 struct resync_pages *rps;
152
153 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
154 if (!r1_bio)
155 return NULL;
156
157 rps = kmalloc_array(pi->raid_disks, sizeof(struct resync_pages),
158 gfp_flags);
159 if (!rps)
160 goto out_free_r1bio;
161
162 /*
163 * Allocate bios : 1 for reading, n-1 for writing
164 */
165 for (j = pi->raid_disks ; j-- ; ) {
166 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
167 if (!bio)
168 goto out_free_bio;
169 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
170 r1_bio->bios[j] = bio;
171 }
172 /*
173 * Allocate RESYNC_PAGES data pages and attach them to
174 * the first bio.
175 * If this is a user-requested check/repair, allocate
176 * RESYNC_PAGES for each bio.
177 */
178 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
179 need_pages = pi->raid_disks;
180 else
181 need_pages = 1;
182 for (j = 0; j < pi->raid_disks; j++) {
183 struct resync_pages *rp = &rps[j];
184
185 bio = r1_bio->bios[j];
186
187 if (j < need_pages) {
188 if (resync_alloc_pages(rp, gfp_flags))
189 goto out_free_pages;
190 } else {
191 memcpy(rp, &rps[0], sizeof(*rp));
192 resync_get_all_pages(rp);
193 }
194
195 rp->raid_bio = r1_bio;
196 bio->bi_private = rp;
197 }
198
199 r1_bio->master_bio = NULL;
200
201 return r1_bio;
202
203out_free_pages:
204 while (--j >= 0)
205 resync_free_pages(&rps[j]);
206
207out_free_bio:
208 while (++j < pi->raid_disks) {
209 bio_uninit(r1_bio->bios[j]);
210 kfree(r1_bio->bios[j]);
211 }
212 kfree(rps);
213
214out_free_r1bio:
215 rbio_pool_free(r1_bio, data);
216 return NULL;
217}
218
219static void r1buf_pool_free(void *__r1_bio, void *data)
220{
221 struct pool_info *pi = data;
222 int i;
223 struct r1bio *r1bio = __r1_bio;
224 struct resync_pages *rp = NULL;
225
226 for (i = pi->raid_disks; i--; ) {
227 rp = get_resync_pages(r1bio->bios[i]);
228 resync_free_pages(rp);
229 bio_uninit(r1bio->bios[i]);
230 kfree(r1bio->bios[i]);
231 }
232
233 /* resync pages array stored in the 1st bio's .bi_private */
234 kfree(rp);
235
236 rbio_pool_free(r1bio, data);
237}
238
239static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
240{
241 int i;
242
243 for (i = 0; i < conf->raid_disks * 2; i++) {
244 struct bio **bio = r1_bio->bios + i;
245 if (!BIO_SPECIAL(*bio))
246 bio_put(*bio);
247 *bio = NULL;
248 }
249}
250
251static void free_r1bio(struct r1bio *r1_bio)
252{
253 struct r1conf *conf = r1_bio->mddev->private;
254
255 put_all_bios(conf, r1_bio);
256 mempool_free(r1_bio, &conf->r1bio_pool);
257}
258
259static void put_buf(struct r1bio *r1_bio)
260{
261 struct r1conf *conf = r1_bio->mddev->private;
262 sector_t sect = r1_bio->sector;
263 int i;
264
265 for (i = 0; i < conf->raid_disks * 2; i++) {
266 struct bio *bio = r1_bio->bios[i];
267 if (bio->bi_end_io)
268 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
269 }
270
271 mempool_free(r1_bio, &conf->r1buf_pool);
272
273 lower_barrier(conf, sect);
274}
275
276static void reschedule_retry(struct r1bio *r1_bio)
277{
278 unsigned long flags;
279 struct mddev *mddev = r1_bio->mddev;
280 struct r1conf *conf = mddev->private;
281 int idx;
282
283 idx = sector_to_idx(r1_bio->sector);
284 spin_lock_irqsave(&conf->device_lock, flags);
285 list_add(&r1_bio->retry_list, &conf->retry_list);
286 atomic_inc(&conf->nr_queued[idx]);
287 spin_unlock_irqrestore(&conf->device_lock, flags);
288
289 wake_up(&conf->wait_barrier);
290 md_wakeup_thread(mddev->thread);
291}
292
293/*
294 * raid_end_bio_io() is called when we have finished servicing a mirrored
295 * operation and are ready to return a success/failure code to the buffer
296 * cache layer.
297 */
298static void call_bio_endio(struct r1bio *r1_bio)
299{
300 struct bio *bio = r1_bio->master_bio;
301
302 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
303 bio->bi_status = BLK_STS_IOERR;
304
305 bio_endio(bio);
306}
307
308static void raid_end_bio_io(struct r1bio *r1_bio)
309{
310 struct bio *bio = r1_bio->master_bio;
311 struct r1conf *conf = r1_bio->mddev->private;
312 sector_t sector = r1_bio->sector;
313
314 /* if nobody has done the final endio yet, do it now */
315 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
316 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
317 (bio_data_dir(bio) == WRITE) ? "write" : "read",
318 (unsigned long long) bio->bi_iter.bi_sector,
319 (unsigned long long) bio_end_sector(bio) - 1);
320
321 call_bio_endio(r1_bio);
322 }
323
324 free_r1bio(r1_bio);
325 /*
326 * Wake up any possible resync thread that waits for the device
327 * to go idle. All I/Os, even write-behind writes, are done.
328 */
329 allow_barrier(conf, sector);
330}
331
332/*
333 * Update disk head position estimator based on IRQ completion info.
334 */
335static inline void update_head_pos(int disk, struct r1bio *r1_bio)
336{
337 struct r1conf *conf = r1_bio->mddev->private;
338
339 conf->mirrors[disk].head_position =
340 r1_bio->sector + (r1_bio->sectors);
341}
342
343/*
344 * Find the disk number which triggered given bio
345 */
346static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
347{
348 int mirror;
349 struct r1conf *conf = r1_bio->mddev->private;
350 int raid_disks = conf->raid_disks;
351
352 for (mirror = 0; mirror < raid_disks * 2; mirror++)
353 if (r1_bio->bios[mirror] == bio)
354 break;
355
356 BUG_ON(mirror == raid_disks * 2);
357 update_head_pos(mirror, r1_bio);
358
359 return mirror;
360}
361
362static void raid1_end_read_request(struct bio *bio)
363{
364 int uptodate = !bio->bi_status;
365 struct r1bio *r1_bio = bio->bi_private;
366 struct r1conf *conf = r1_bio->mddev->private;
367 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
368
369 /*
370 * this branch is our 'one mirror IO has finished' event handler:
371 */
372 update_head_pos(r1_bio->read_disk, r1_bio);
373
374 if (uptodate)
375 set_bit(R1BIO_Uptodate, &r1_bio->state);
376 else if (test_bit(FailFast, &rdev->flags) &&
377 test_bit(R1BIO_FailFast, &r1_bio->state))
378 /* This was a fail-fast read so we definitely
379 * want to retry */
380 ;
381 else {
382 /* If all other devices have failed, we want to return
383 * the error upwards rather than fail the last device.
384 * Here we redefine "uptodate" to mean "Don't want to retry"
385 */
386 unsigned long flags;
387 spin_lock_irqsave(&conf->device_lock, flags);
388 if (r1_bio->mddev->degraded == conf->raid_disks ||
389 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
390 test_bit(In_sync, &rdev->flags)))
391 uptodate = 1;
392 spin_unlock_irqrestore(&conf->device_lock, flags);
393 }
394
395 if (uptodate) {
396 raid_end_bio_io(r1_bio);
397 rdev_dec_pending(rdev, conf->mddev);
398 } else {
399 /*
400 * oops, read error:
401 */
402 pr_err_ratelimited("md/raid1:%s: %pg: rescheduling sector %llu\n",
403 mdname(conf->mddev),
404 rdev->bdev,
405 (unsigned long long)r1_bio->sector);
406 set_bit(R1BIO_ReadError, &r1_bio->state);
407 reschedule_retry(r1_bio);
408 /* don't drop the reference on read_disk yet */
409 }
410}
411
412static void close_write(struct r1bio *r1_bio)
413{
414 /* it really is the end of this request */
415 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
416 bio_free_pages(r1_bio->behind_master_bio);
417 bio_put(r1_bio->behind_master_bio);
418 r1_bio->behind_master_bio = NULL;
419 }
420 /* clear the bitmap if all writes complete successfully */
421 md_bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
422 r1_bio->sectors,
423 !test_bit(R1BIO_Degraded, &r1_bio->state),
424 test_bit(R1BIO_BehindIO, &r1_bio->state));
425 md_write_end(r1_bio->mddev);
426}
427
428static void r1_bio_write_done(struct r1bio *r1_bio)
429{
430 if (!atomic_dec_and_test(&r1_bio->remaining))
431 return;
432
433 if (test_bit(R1BIO_WriteError, &r1_bio->state))
434 reschedule_retry(r1_bio);
435 else {
436 close_write(r1_bio);
437 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
438 reschedule_retry(r1_bio);
439 else
440 raid_end_bio_io(r1_bio);
441 }
442}
443
444static void raid1_end_write_request(struct bio *bio)
445{
446 struct r1bio *r1_bio = bio->bi_private;
447 int behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
448 struct r1conf *conf = r1_bio->mddev->private;
449 struct bio *to_put = NULL;
450 int mirror = find_bio_disk(r1_bio, bio);
451 struct md_rdev *rdev = conf->mirrors[mirror].rdev;
452 bool discard_error;
453 sector_t lo = r1_bio->sector;
454 sector_t hi = r1_bio->sector + r1_bio->sectors;
455
456 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
457
458 /*
459 * 'one mirror IO has finished' event handler:
460 */
461 if (bio->bi_status && !discard_error) {
462 set_bit(WriteErrorSeen, &rdev->flags);
463 if (!test_and_set_bit(WantReplacement, &rdev->flags))
464 set_bit(MD_RECOVERY_NEEDED, &
465 conf->mddev->recovery);
466
467 if (test_bit(FailFast, &rdev->flags) &&
468 (bio->bi_opf & MD_FAILFAST) &&
469 /* We never try FailFast to WriteMostly devices */
470 !test_bit(WriteMostly, &rdev->flags)) {
471 md_error(r1_bio->mddev, rdev);
472 }
473
474 /*
475 * When the device is faulty, it is not necessary to
476 * handle write error.
477 */
478 if (!test_bit(Faulty, &rdev->flags))
479 set_bit(R1BIO_WriteError, &r1_bio->state);
480 else {
481 /* Fail the request */
482 set_bit(R1BIO_Degraded, &r1_bio->state);
483 /* Finished with this branch */
484 r1_bio->bios[mirror] = NULL;
485 to_put = bio;
486 }
487 } else {
488 /*
489 * Set R1BIO_Uptodate in our master bio, so that we
490 * will return a good error code for to the higher
491 * levels even if IO on some other mirrored buffer
492 * fails.
493 *
494 * The 'master' represents the composite IO operation
495 * to user-side. So if something waits for IO, then it
496 * will wait for the 'master' bio.
497 */
498 r1_bio->bios[mirror] = NULL;
499 to_put = bio;
500 /*
501 * Do not set R1BIO_Uptodate if the current device is
502 * rebuilding or Faulty. This is because we cannot use
503 * such device for properly reading the data back (we could
504 * potentially use it, if the current write would have felt
505 * before rdev->recovery_offset, but for simplicity we don't
506 * check this here.
507 */
508 if (test_bit(In_sync, &rdev->flags) &&
509 !test_bit(Faulty, &rdev->flags))
510 set_bit(R1BIO_Uptodate, &r1_bio->state);
511
512 /* Maybe we can clear some bad blocks. */
513 if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors) &&
514 !discard_error) {
515 r1_bio->bios[mirror] = IO_MADE_GOOD;
516 set_bit(R1BIO_MadeGood, &r1_bio->state);
517 }
518 }
519
520 if (behind) {
521 if (test_bit(CollisionCheck, &rdev->flags))
522 remove_serial(rdev, lo, hi);
523 if (test_bit(WriteMostly, &rdev->flags))
524 atomic_dec(&r1_bio->behind_remaining);
525
526 /*
527 * In behind mode, we ACK the master bio once the I/O
528 * has safely reached all non-writemostly
529 * disks. Setting the Returned bit ensures that this
530 * gets done only once -- we don't ever want to return
531 * -EIO here, instead we'll wait
532 */
533 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
534 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
535 /* Maybe we can return now */
536 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
537 struct bio *mbio = r1_bio->master_bio;
538 pr_debug("raid1: behind end write sectors"
539 " %llu-%llu\n",
540 (unsigned long long) mbio->bi_iter.bi_sector,
541 (unsigned long long) bio_end_sector(mbio) - 1);
542 call_bio_endio(r1_bio);
543 }
544 }
545 } else if (rdev->mddev->serialize_policy)
546 remove_serial(rdev, lo, hi);
547 if (r1_bio->bios[mirror] == NULL)
548 rdev_dec_pending(rdev, conf->mddev);
549
550 /*
551 * Let's see if all mirrored write operations have finished
552 * already.
553 */
554 r1_bio_write_done(r1_bio);
555
556 if (to_put)
557 bio_put(to_put);
558}
559
560static sector_t align_to_barrier_unit_end(sector_t start_sector,
561 sector_t sectors)
562{
563 sector_t len;
564
565 WARN_ON(sectors == 0);
566 /*
567 * len is the number of sectors from start_sector to end of the
568 * barrier unit which start_sector belongs to.
569 */
570 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
571 start_sector;
572
573 if (len > sectors)
574 len = sectors;
575
576 return len;
577}
578
579static void update_read_sectors(struct r1conf *conf, int disk,
580 sector_t this_sector, int len)
581{
582 struct raid1_info *info = &conf->mirrors[disk];
583
584 atomic_inc(&info->rdev->nr_pending);
585 if (info->next_seq_sect != this_sector)
586 info->seq_start = this_sector;
587 info->next_seq_sect = this_sector + len;
588}
589
590static int choose_first_rdev(struct r1conf *conf, struct r1bio *r1_bio,
591 int *max_sectors)
592{
593 sector_t this_sector = r1_bio->sector;
594 int len = r1_bio->sectors;
595 int disk;
596
597 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
598 struct md_rdev *rdev;
599 int read_len;
600
601 if (r1_bio->bios[disk] == IO_BLOCKED)
602 continue;
603
604 rdev = conf->mirrors[disk].rdev;
605 if (!rdev || test_bit(Faulty, &rdev->flags))
606 continue;
607
608 /* choose the first disk even if it has some bad blocks. */
609 read_len = raid1_check_read_range(rdev, this_sector, &len);
610 if (read_len > 0) {
611 update_read_sectors(conf, disk, this_sector, read_len);
612 *max_sectors = read_len;
613 return disk;
614 }
615 }
616
617 return -1;
618}
619
620static int choose_bb_rdev(struct r1conf *conf, struct r1bio *r1_bio,
621 int *max_sectors)
622{
623 sector_t this_sector = r1_bio->sector;
624 int best_disk = -1;
625 int best_len = 0;
626 int disk;
627
628 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
629 struct md_rdev *rdev;
630 int len;
631 int read_len;
632
633 if (r1_bio->bios[disk] == IO_BLOCKED)
634 continue;
635
636 rdev = conf->mirrors[disk].rdev;
637 if (!rdev || test_bit(Faulty, &rdev->flags) ||
638 test_bit(WriteMostly, &rdev->flags))
639 continue;
640
641 /* keep track of the disk with the most readable sectors. */
642 len = r1_bio->sectors;
643 read_len = raid1_check_read_range(rdev, this_sector, &len);
644 if (read_len > best_len) {
645 best_disk = disk;
646 best_len = read_len;
647 }
648 }
649
650 if (best_disk != -1) {
651 *max_sectors = best_len;
652 update_read_sectors(conf, best_disk, this_sector, best_len);
653 }
654
655 return best_disk;
656}
657
658static int choose_slow_rdev(struct r1conf *conf, struct r1bio *r1_bio,
659 int *max_sectors)
660{
661 sector_t this_sector = r1_bio->sector;
662 int bb_disk = -1;
663 int bb_read_len = 0;
664 int disk;
665
666 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
667 struct md_rdev *rdev;
668 int len;
669 int read_len;
670
671 if (r1_bio->bios[disk] == IO_BLOCKED)
672 continue;
673
674 rdev = conf->mirrors[disk].rdev;
675 if (!rdev || test_bit(Faulty, &rdev->flags) ||
676 !test_bit(WriteMostly, &rdev->flags))
677 continue;
678
679 /* there are no bad blocks, we can use this disk */
680 len = r1_bio->sectors;
681 read_len = raid1_check_read_range(rdev, this_sector, &len);
682 if (read_len == r1_bio->sectors) {
683 update_read_sectors(conf, disk, this_sector, read_len);
684 return disk;
685 }
686
687 /*
688 * there are partial bad blocks, choose the rdev with largest
689 * read length.
690 */
691 if (read_len > bb_read_len) {
692 bb_disk = disk;
693 bb_read_len = read_len;
694 }
695 }
696
697 if (bb_disk != -1) {
698 *max_sectors = bb_read_len;
699 update_read_sectors(conf, bb_disk, this_sector, bb_read_len);
700 }
701
702 return bb_disk;
703}
704
705static bool is_sequential(struct r1conf *conf, int disk, struct r1bio *r1_bio)
706{
707 /* TODO: address issues with this check and concurrency. */
708 return conf->mirrors[disk].next_seq_sect == r1_bio->sector ||
709 conf->mirrors[disk].head_position == r1_bio->sector;
710}
711
712/*
713 * If buffered sequential IO size exceeds optimal iosize, check if there is idle
714 * disk. If yes, choose the idle disk.
715 */
716static bool should_choose_next(struct r1conf *conf, int disk)
717{
718 struct raid1_info *mirror = &conf->mirrors[disk];
719 int opt_iosize;
720
721 if (!test_bit(Nonrot, &mirror->rdev->flags))
722 return false;
723
724 opt_iosize = bdev_io_opt(mirror->rdev->bdev) >> 9;
725 return opt_iosize > 0 && mirror->seq_start != MaxSector &&
726 mirror->next_seq_sect > opt_iosize &&
727 mirror->next_seq_sect - opt_iosize >= mirror->seq_start;
728}
729
730static bool rdev_readable(struct md_rdev *rdev, struct r1bio *r1_bio)
731{
732 if (!rdev || test_bit(Faulty, &rdev->flags))
733 return false;
734
735 /* still in recovery */
736 if (!test_bit(In_sync, &rdev->flags) &&
737 rdev->recovery_offset < r1_bio->sector + r1_bio->sectors)
738 return false;
739
740 /* don't read from slow disk unless have to */
741 if (test_bit(WriteMostly, &rdev->flags))
742 return false;
743
744 /* don't split IO for bad blocks unless have to */
745 if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors))
746 return false;
747
748 return true;
749}
750
751struct read_balance_ctl {
752 sector_t closest_dist;
753 int closest_dist_disk;
754 int min_pending;
755 int min_pending_disk;
756 int sequential_disk;
757 int readable_disks;
758};
759
760static int choose_best_rdev(struct r1conf *conf, struct r1bio *r1_bio)
761{
762 int disk;
763 struct read_balance_ctl ctl = {
764 .closest_dist_disk = -1,
765 .closest_dist = MaxSector,
766 .min_pending_disk = -1,
767 .min_pending = UINT_MAX,
768 .sequential_disk = -1,
769 };
770
771 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
772 struct md_rdev *rdev;
773 sector_t dist;
774 unsigned int pending;
775
776 if (r1_bio->bios[disk] == IO_BLOCKED)
777 continue;
778
779 rdev = conf->mirrors[disk].rdev;
780 if (!rdev_readable(rdev, r1_bio))
781 continue;
782
783 /* At least two disks to choose from so failfast is OK */
784 if (ctl.readable_disks++ == 1)
785 set_bit(R1BIO_FailFast, &r1_bio->state);
786
787 pending = atomic_read(&rdev->nr_pending);
788 dist = abs(r1_bio->sector - conf->mirrors[disk].head_position);
789
790 /* Don't change to another disk for sequential reads */
791 if (is_sequential(conf, disk, r1_bio)) {
792 if (!should_choose_next(conf, disk))
793 return disk;
794
795 /*
796 * Add 'pending' to avoid choosing this disk if
797 * there is other idle disk.
798 */
799 pending++;
800 /*
801 * If there is no other idle disk, this disk
802 * will be chosen.
803 */
804 ctl.sequential_disk = disk;
805 }
806
807 if (ctl.min_pending > pending) {
808 ctl.min_pending = pending;
809 ctl.min_pending_disk = disk;
810 }
811
812 if (ctl.closest_dist > dist) {
813 ctl.closest_dist = dist;
814 ctl.closest_dist_disk = disk;
815 }
816 }
817
818 /*
819 * sequential IO size exceeds optimal iosize, however, there is no other
820 * idle disk, so choose the sequential disk.
821 */
822 if (ctl.sequential_disk != -1 && ctl.min_pending != 0)
823 return ctl.sequential_disk;
824
825 /*
826 * If all disks are rotational, choose the closest disk. If any disk is
827 * non-rotational, choose the disk with less pending request even the
828 * disk is rotational, which might/might not be optimal for raids with
829 * mixed ratation/non-rotational disks depending on workload.
830 */
831 if (ctl.min_pending_disk != -1 &&
832 (READ_ONCE(conf->nonrot_disks) || ctl.min_pending == 0))
833 return ctl.min_pending_disk;
834 else
835 return ctl.closest_dist_disk;
836}
837
838/*
839 * This routine returns the disk from which the requested read should be done.
840 *
841 * 1) If resync is in progress, find the first usable disk and use it even if it
842 * has some bad blocks.
843 *
844 * 2) Now that there is no resync, loop through all disks and skipping slow
845 * disks and disks with bad blocks for now. Only pay attention to key disk
846 * choice.
847 *
848 * 3) If we've made it this far, now look for disks with bad blocks and choose
849 * the one with most number of sectors.
850 *
851 * 4) If we are all the way at the end, we have no choice but to use a disk even
852 * if it is write mostly.
853 *
854 * The rdev for the device selected will have nr_pending incremented.
855 */
856static int read_balance(struct r1conf *conf, struct r1bio *r1_bio,
857 int *max_sectors)
858{
859 int disk;
860
861 clear_bit(R1BIO_FailFast, &r1_bio->state);
862
863 if (raid1_should_read_first(conf->mddev, r1_bio->sector,
864 r1_bio->sectors))
865 return choose_first_rdev(conf, r1_bio, max_sectors);
866
867 disk = choose_best_rdev(conf, r1_bio);
868 if (disk >= 0) {
869 *max_sectors = r1_bio->sectors;
870 update_read_sectors(conf, disk, r1_bio->sector,
871 r1_bio->sectors);
872 return disk;
873 }
874
875 /*
876 * If we are here it means we didn't find a perfectly good disk so
877 * now spend a bit more time trying to find one with the most good
878 * sectors.
879 */
880 disk = choose_bb_rdev(conf, r1_bio, max_sectors);
881 if (disk >= 0)
882 return disk;
883
884 return choose_slow_rdev(conf, r1_bio, max_sectors);
885}
886
887static void wake_up_barrier(struct r1conf *conf)
888{
889 if (wq_has_sleeper(&conf->wait_barrier))
890 wake_up(&conf->wait_barrier);
891}
892
893static void flush_bio_list(struct r1conf *conf, struct bio *bio)
894{
895 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
896 raid1_prepare_flush_writes(conf->mddev->bitmap);
897 wake_up_barrier(conf);
898
899 while (bio) { /* submit pending writes */
900 struct bio *next = bio->bi_next;
901
902 raid1_submit_write(bio);
903 bio = next;
904 cond_resched();
905 }
906}
907
908static void flush_pending_writes(struct r1conf *conf)
909{
910 /* Any writes that have been queued but are awaiting
911 * bitmap updates get flushed here.
912 */
913 spin_lock_irq(&conf->device_lock);
914
915 if (conf->pending_bio_list.head) {
916 struct blk_plug plug;
917 struct bio *bio;
918
919 bio = bio_list_get(&conf->pending_bio_list);
920 spin_unlock_irq(&conf->device_lock);
921
922 /*
923 * As this is called in a wait_event() loop (see freeze_array),
924 * current->state might be TASK_UNINTERRUPTIBLE which will
925 * cause a warning when we prepare to wait again. As it is
926 * rare that this path is taken, it is perfectly safe to force
927 * us to go around the wait_event() loop again, so the warning
928 * is a false-positive. Silence the warning by resetting
929 * thread state
930 */
931 __set_current_state(TASK_RUNNING);
932 blk_start_plug(&plug);
933 flush_bio_list(conf, bio);
934 blk_finish_plug(&plug);
935 } else
936 spin_unlock_irq(&conf->device_lock);
937}
938
939/* Barriers....
940 * Sometimes we need to suspend IO while we do something else,
941 * either some resync/recovery, or reconfigure the array.
942 * To do this we raise a 'barrier'.
943 * The 'barrier' is a counter that can be raised multiple times
944 * to count how many activities are happening which preclude
945 * normal IO.
946 * We can only raise the barrier if there is no pending IO.
947 * i.e. if nr_pending == 0.
948 * We choose only to raise the barrier if no-one is waiting for the
949 * barrier to go down. This means that as soon as an IO request
950 * is ready, no other operations which require a barrier will start
951 * until the IO request has had a chance.
952 *
953 * So: regular IO calls 'wait_barrier'. When that returns there
954 * is no backgroup IO happening, It must arrange to call
955 * allow_barrier when it has finished its IO.
956 * backgroup IO calls must call raise_barrier. Once that returns
957 * there is no normal IO happeing. It must arrange to call
958 * lower_barrier when the particular background IO completes.
959 *
960 * If resync/recovery is interrupted, returns -EINTR;
961 * Otherwise, returns 0.
962 */
963static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
964{
965 int idx = sector_to_idx(sector_nr);
966
967 spin_lock_irq(&conf->resync_lock);
968
969 /* Wait until no block IO is waiting */
970 wait_event_lock_irq(conf->wait_barrier,
971 !atomic_read(&conf->nr_waiting[idx]),
972 conf->resync_lock);
973
974 /* block any new IO from starting */
975 atomic_inc(&conf->barrier[idx]);
976 /*
977 * In raise_barrier() we firstly increase conf->barrier[idx] then
978 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
979 * increase conf->nr_pending[idx] then check conf->barrier[idx].
980 * A memory barrier here to make sure conf->nr_pending[idx] won't
981 * be fetched before conf->barrier[idx] is increased. Otherwise
982 * there will be a race between raise_barrier() and _wait_barrier().
983 */
984 smp_mb__after_atomic();
985
986 /* For these conditions we must wait:
987 * A: while the array is in frozen state
988 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
989 * existing in corresponding I/O barrier bucket.
990 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
991 * max resync count which allowed on current I/O barrier bucket.
992 */
993 wait_event_lock_irq(conf->wait_barrier,
994 (!conf->array_frozen &&
995 !atomic_read(&conf->nr_pending[idx]) &&
996 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
997 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
998 conf->resync_lock);
999
1000 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
1001 atomic_dec(&conf->barrier[idx]);
1002 spin_unlock_irq(&conf->resync_lock);
1003 wake_up(&conf->wait_barrier);
1004 return -EINTR;
1005 }
1006
1007 atomic_inc(&conf->nr_sync_pending);
1008 spin_unlock_irq(&conf->resync_lock);
1009
1010 return 0;
1011}
1012
1013static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
1014{
1015 int idx = sector_to_idx(sector_nr);
1016
1017 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
1018
1019 atomic_dec(&conf->barrier[idx]);
1020 atomic_dec(&conf->nr_sync_pending);
1021 wake_up(&conf->wait_barrier);
1022}
1023
1024static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
1025{
1026 bool ret = true;
1027
1028 /*
1029 * We need to increase conf->nr_pending[idx] very early here,
1030 * then raise_barrier() can be blocked when it waits for
1031 * conf->nr_pending[idx] to be 0. Then we can avoid holding
1032 * conf->resync_lock when there is no barrier raised in same
1033 * barrier unit bucket. Also if the array is frozen, I/O
1034 * should be blocked until array is unfrozen.
1035 */
1036 atomic_inc(&conf->nr_pending[idx]);
1037 /*
1038 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
1039 * check conf->barrier[idx]. In raise_barrier() we firstly increase
1040 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
1041 * barrier is necessary here to make sure conf->barrier[idx] won't be
1042 * fetched before conf->nr_pending[idx] is increased. Otherwise there
1043 * will be a race between _wait_barrier() and raise_barrier().
1044 */
1045 smp_mb__after_atomic();
1046
1047 /*
1048 * Don't worry about checking two atomic_t variables at same time
1049 * here. If during we check conf->barrier[idx], the array is
1050 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
1051 * 0, it is safe to return and make the I/O continue. Because the
1052 * array is frozen, all I/O returned here will eventually complete
1053 * or be queued, no race will happen. See code comment in
1054 * frozen_array().
1055 */
1056 if (!READ_ONCE(conf->array_frozen) &&
1057 !atomic_read(&conf->barrier[idx]))
1058 return ret;
1059
1060 /*
1061 * After holding conf->resync_lock, conf->nr_pending[idx]
1062 * should be decreased before waiting for barrier to drop.
1063 * Otherwise, we may encounter a race condition because
1064 * raise_barrer() might be waiting for conf->nr_pending[idx]
1065 * to be 0 at same time.
1066 */
1067 spin_lock_irq(&conf->resync_lock);
1068 atomic_inc(&conf->nr_waiting[idx]);
1069 atomic_dec(&conf->nr_pending[idx]);
1070 /*
1071 * In case freeze_array() is waiting for
1072 * get_unqueued_pending() == extra
1073 */
1074 wake_up_barrier(conf);
1075 /* Wait for the barrier in same barrier unit bucket to drop. */
1076
1077 /* Return false when nowait flag is set */
1078 if (nowait) {
1079 ret = false;
1080 } else {
1081 wait_event_lock_irq(conf->wait_barrier,
1082 !conf->array_frozen &&
1083 !atomic_read(&conf->barrier[idx]),
1084 conf->resync_lock);
1085 atomic_inc(&conf->nr_pending[idx]);
1086 }
1087
1088 atomic_dec(&conf->nr_waiting[idx]);
1089 spin_unlock_irq(&conf->resync_lock);
1090 return ret;
1091}
1092
1093static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1094{
1095 int idx = sector_to_idx(sector_nr);
1096 bool ret = true;
1097
1098 /*
1099 * Very similar to _wait_barrier(). The difference is, for read
1100 * I/O we don't need wait for sync I/O, but if the whole array
1101 * is frozen, the read I/O still has to wait until the array is
1102 * unfrozen. Since there is no ordering requirement with
1103 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1104 */
1105 atomic_inc(&conf->nr_pending[idx]);
1106
1107 if (!READ_ONCE(conf->array_frozen))
1108 return ret;
1109
1110 spin_lock_irq(&conf->resync_lock);
1111 atomic_inc(&conf->nr_waiting[idx]);
1112 atomic_dec(&conf->nr_pending[idx]);
1113 /*
1114 * In case freeze_array() is waiting for
1115 * get_unqueued_pending() == extra
1116 */
1117 wake_up_barrier(conf);
1118 /* Wait for array to be unfrozen */
1119
1120 /* Return false when nowait flag is set */
1121 if (nowait) {
1122 /* Return false when nowait flag is set */
1123 ret = false;
1124 } else {
1125 wait_event_lock_irq(conf->wait_barrier,
1126 !conf->array_frozen,
1127 conf->resync_lock);
1128 atomic_inc(&conf->nr_pending[idx]);
1129 }
1130
1131 atomic_dec(&conf->nr_waiting[idx]);
1132 spin_unlock_irq(&conf->resync_lock);
1133 return ret;
1134}
1135
1136static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1137{
1138 int idx = sector_to_idx(sector_nr);
1139
1140 return _wait_barrier(conf, idx, nowait);
1141}
1142
1143static void _allow_barrier(struct r1conf *conf, int idx)
1144{
1145 atomic_dec(&conf->nr_pending[idx]);
1146 wake_up_barrier(conf);
1147}
1148
1149static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1150{
1151 int idx = sector_to_idx(sector_nr);
1152
1153 _allow_barrier(conf, idx);
1154}
1155
1156/* conf->resync_lock should be held */
1157static int get_unqueued_pending(struct r1conf *conf)
1158{
1159 int idx, ret;
1160
1161 ret = atomic_read(&conf->nr_sync_pending);
1162 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1163 ret += atomic_read(&conf->nr_pending[idx]) -
1164 atomic_read(&conf->nr_queued[idx]);
1165
1166 return ret;
1167}
1168
1169static void freeze_array(struct r1conf *conf, int extra)
1170{
1171 /* Stop sync I/O and normal I/O and wait for everything to
1172 * go quiet.
1173 * This is called in two situations:
1174 * 1) management command handlers (reshape, remove disk, quiesce).
1175 * 2) one normal I/O request failed.
1176
1177 * After array_frozen is set to 1, new sync IO will be blocked at
1178 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1179 * or wait_read_barrier(). The flying I/Os will either complete or be
1180 * queued. When everything goes quite, there are only queued I/Os left.
1181
1182 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1183 * barrier bucket index which this I/O request hits. When all sync and
1184 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1185 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1186 * in handle_read_error(), we may call freeze_array() before trying to
1187 * fix the read error. In this case, the error read I/O is not queued,
1188 * so get_unqueued_pending() == 1.
1189 *
1190 * Therefore before this function returns, we need to wait until
1191 * get_unqueued_pendings(conf) gets equal to extra. For
1192 * normal I/O context, extra is 1, in rested situations extra is 0.
1193 */
1194 spin_lock_irq(&conf->resync_lock);
1195 conf->array_frozen = 1;
1196 mddev_add_trace_msg(conf->mddev, "raid1 wait freeze");
1197 wait_event_lock_irq_cmd(
1198 conf->wait_barrier,
1199 get_unqueued_pending(conf) == extra,
1200 conf->resync_lock,
1201 flush_pending_writes(conf));
1202 spin_unlock_irq(&conf->resync_lock);
1203}
1204static void unfreeze_array(struct r1conf *conf)
1205{
1206 /* reverse the effect of the freeze */
1207 spin_lock_irq(&conf->resync_lock);
1208 conf->array_frozen = 0;
1209 spin_unlock_irq(&conf->resync_lock);
1210 wake_up(&conf->wait_barrier);
1211}
1212
1213static void alloc_behind_master_bio(struct r1bio *r1_bio,
1214 struct bio *bio)
1215{
1216 int size = bio->bi_iter.bi_size;
1217 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1218 int i = 0;
1219 struct bio *behind_bio = NULL;
1220
1221 behind_bio = bio_alloc_bioset(NULL, vcnt, 0, GFP_NOIO,
1222 &r1_bio->mddev->bio_set);
1223
1224 /* discard op, we don't support writezero/writesame yet */
1225 if (!bio_has_data(bio)) {
1226 behind_bio->bi_iter.bi_size = size;
1227 goto skip_copy;
1228 }
1229
1230 while (i < vcnt && size) {
1231 struct page *page;
1232 int len = min_t(int, PAGE_SIZE, size);
1233
1234 page = alloc_page(GFP_NOIO);
1235 if (unlikely(!page))
1236 goto free_pages;
1237
1238 if (!bio_add_page(behind_bio, page, len, 0)) {
1239 put_page(page);
1240 goto free_pages;
1241 }
1242
1243 size -= len;
1244 i++;
1245 }
1246
1247 bio_copy_data(behind_bio, bio);
1248skip_copy:
1249 r1_bio->behind_master_bio = behind_bio;
1250 set_bit(R1BIO_BehindIO, &r1_bio->state);
1251
1252 return;
1253
1254free_pages:
1255 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1256 bio->bi_iter.bi_size);
1257 bio_free_pages(behind_bio);
1258 bio_put(behind_bio);
1259}
1260
1261static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1262{
1263 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1264 cb);
1265 struct mddev *mddev = plug->cb.data;
1266 struct r1conf *conf = mddev->private;
1267 struct bio *bio;
1268
1269 if (from_schedule) {
1270 spin_lock_irq(&conf->device_lock);
1271 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1272 spin_unlock_irq(&conf->device_lock);
1273 wake_up_barrier(conf);
1274 md_wakeup_thread(mddev->thread);
1275 kfree(plug);
1276 return;
1277 }
1278
1279 /* we aren't scheduling, so we can do the write-out directly. */
1280 bio = bio_list_get(&plug->pending);
1281 flush_bio_list(conf, bio);
1282 kfree(plug);
1283}
1284
1285static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1286{
1287 r1_bio->master_bio = bio;
1288 r1_bio->sectors = bio_sectors(bio);
1289 r1_bio->state = 0;
1290 r1_bio->mddev = mddev;
1291 r1_bio->sector = bio->bi_iter.bi_sector;
1292}
1293
1294static inline struct r1bio *
1295alloc_r1bio(struct mddev *mddev, struct bio *bio)
1296{
1297 struct r1conf *conf = mddev->private;
1298 struct r1bio *r1_bio;
1299
1300 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1301 /* Ensure no bio records IO_BLOCKED */
1302 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1303 init_r1bio(r1_bio, mddev, bio);
1304 return r1_bio;
1305}
1306
1307static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1308 int max_read_sectors, struct r1bio *r1_bio)
1309{
1310 struct r1conf *conf = mddev->private;
1311 struct raid1_info *mirror;
1312 struct bio *read_bio;
1313 struct bitmap *bitmap = mddev->bitmap;
1314 const enum req_op op = bio_op(bio);
1315 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1316 int max_sectors;
1317 int rdisk;
1318 bool r1bio_existed = !!r1_bio;
1319 char b[BDEVNAME_SIZE];
1320
1321 /*
1322 * If r1_bio is set, we are blocking the raid1d thread
1323 * so there is a tiny risk of deadlock. So ask for
1324 * emergency memory if needed.
1325 */
1326 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1327
1328 if (r1bio_existed) {
1329 /* Need to get the block device name carefully */
1330 struct md_rdev *rdev = conf->mirrors[r1_bio->read_disk].rdev;
1331
1332 if (rdev)
1333 snprintf(b, sizeof(b), "%pg", rdev->bdev);
1334 else
1335 strcpy(b, "???");
1336 }
1337
1338 /*
1339 * Still need barrier for READ in case that whole
1340 * array is frozen.
1341 */
1342 if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1343 bio->bi_opf & REQ_NOWAIT)) {
1344 bio_wouldblock_error(bio);
1345 return;
1346 }
1347
1348 if (!r1_bio)
1349 r1_bio = alloc_r1bio(mddev, bio);
1350 else
1351 init_r1bio(r1_bio, mddev, bio);
1352 r1_bio->sectors = max_read_sectors;
1353
1354 /*
1355 * make_request() can abort the operation when read-ahead is being
1356 * used and no empty request is available.
1357 */
1358 rdisk = read_balance(conf, r1_bio, &max_sectors);
1359
1360 if (rdisk < 0) {
1361 /* couldn't find anywhere to read from */
1362 if (r1bio_existed) {
1363 pr_crit_ratelimited("md/raid1:%s: %s: unrecoverable I/O read error for block %llu\n",
1364 mdname(mddev),
1365 b,
1366 (unsigned long long)r1_bio->sector);
1367 }
1368 raid_end_bio_io(r1_bio);
1369 return;
1370 }
1371 mirror = conf->mirrors + rdisk;
1372
1373 if (r1bio_existed)
1374 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %pg\n",
1375 mdname(mddev),
1376 (unsigned long long)r1_bio->sector,
1377 mirror->rdev->bdev);
1378
1379 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1380 bitmap) {
1381 /*
1382 * Reading from a write-mostly device must take care not to
1383 * over-take any writes that are 'behind'
1384 */
1385 mddev_add_trace_msg(mddev, "raid1 wait behind writes");
1386 wait_event(bitmap->behind_wait,
1387 atomic_read(&bitmap->behind_writes) == 0);
1388 }
1389
1390 if (max_sectors < bio_sectors(bio)) {
1391 struct bio *split = bio_split(bio, max_sectors,
1392 gfp, &conf->bio_split);
1393 bio_chain(split, bio);
1394 submit_bio_noacct(bio);
1395 bio = split;
1396 r1_bio->master_bio = bio;
1397 r1_bio->sectors = max_sectors;
1398 }
1399
1400 r1_bio->read_disk = rdisk;
1401 if (!r1bio_existed) {
1402 md_account_bio(mddev, &bio);
1403 r1_bio->master_bio = bio;
1404 }
1405 read_bio = bio_alloc_clone(mirror->rdev->bdev, bio, gfp,
1406 &mddev->bio_set);
1407
1408 r1_bio->bios[rdisk] = read_bio;
1409
1410 read_bio->bi_iter.bi_sector = r1_bio->sector +
1411 mirror->rdev->data_offset;
1412 read_bio->bi_end_io = raid1_end_read_request;
1413 read_bio->bi_opf = op | do_sync;
1414 if (test_bit(FailFast, &mirror->rdev->flags) &&
1415 test_bit(R1BIO_FailFast, &r1_bio->state))
1416 read_bio->bi_opf |= MD_FAILFAST;
1417 read_bio->bi_private = r1_bio;
1418 mddev_trace_remap(mddev, read_bio, r1_bio->sector);
1419 submit_bio_noacct(read_bio);
1420}
1421
1422static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1423 int max_write_sectors)
1424{
1425 struct r1conf *conf = mddev->private;
1426 struct r1bio *r1_bio;
1427 int i, disks;
1428 struct bitmap *bitmap = mddev->bitmap;
1429 unsigned long flags;
1430 struct md_rdev *blocked_rdev;
1431 int first_clone;
1432 int max_sectors;
1433 bool write_behind = false;
1434 bool is_discard = (bio_op(bio) == REQ_OP_DISCARD);
1435
1436 if (mddev_is_clustered(mddev) &&
1437 md_cluster_ops->area_resyncing(mddev, WRITE,
1438 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1439
1440 DEFINE_WAIT(w);
1441 if (bio->bi_opf & REQ_NOWAIT) {
1442 bio_wouldblock_error(bio);
1443 return;
1444 }
1445 for (;;) {
1446 prepare_to_wait(&conf->wait_barrier,
1447 &w, TASK_IDLE);
1448 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1449 bio->bi_iter.bi_sector,
1450 bio_end_sector(bio)))
1451 break;
1452 schedule();
1453 }
1454 finish_wait(&conf->wait_barrier, &w);
1455 }
1456
1457 /*
1458 * Register the new request and wait if the reconstruction
1459 * thread has put up a bar for new requests.
1460 * Continue immediately if no resync is active currently.
1461 */
1462 if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1463 bio->bi_opf & REQ_NOWAIT)) {
1464 bio_wouldblock_error(bio);
1465 return;
1466 }
1467
1468 retry_write:
1469 r1_bio = alloc_r1bio(mddev, bio);
1470 r1_bio->sectors = max_write_sectors;
1471
1472 /* first select target devices under rcu_lock and
1473 * inc refcount on their rdev. Record them by setting
1474 * bios[x] to bio
1475 * If there are known/acknowledged bad blocks on any device on
1476 * which we have seen a write error, we want to avoid writing those
1477 * blocks.
1478 * This potentially requires several writes to write around
1479 * the bad blocks. Each set of writes gets it's own r1bio
1480 * with a set of bios attached.
1481 */
1482
1483 disks = conf->raid_disks * 2;
1484 blocked_rdev = NULL;
1485 max_sectors = r1_bio->sectors;
1486 for (i = 0; i < disks; i++) {
1487 struct md_rdev *rdev = conf->mirrors[i].rdev;
1488
1489 /*
1490 * The write-behind io is only attempted on drives marked as
1491 * write-mostly, which means we could allocate write behind
1492 * bio later.
1493 */
1494 if (!is_discard && rdev && test_bit(WriteMostly, &rdev->flags))
1495 write_behind = true;
1496
1497 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1498 atomic_inc(&rdev->nr_pending);
1499 blocked_rdev = rdev;
1500 break;
1501 }
1502 r1_bio->bios[i] = NULL;
1503 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1504 if (i < conf->raid_disks)
1505 set_bit(R1BIO_Degraded, &r1_bio->state);
1506 continue;
1507 }
1508
1509 atomic_inc(&rdev->nr_pending);
1510 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1511 sector_t first_bad;
1512 int bad_sectors;
1513 int is_bad;
1514
1515 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1516 &first_bad, &bad_sectors);
1517 if (is_bad < 0) {
1518 /* mustn't write here until the bad block is
1519 * acknowledged*/
1520 set_bit(BlockedBadBlocks, &rdev->flags);
1521 blocked_rdev = rdev;
1522 break;
1523 }
1524 if (is_bad && first_bad <= r1_bio->sector) {
1525 /* Cannot write here at all */
1526 bad_sectors -= (r1_bio->sector - first_bad);
1527 if (bad_sectors < max_sectors)
1528 /* mustn't write more than bad_sectors
1529 * to other devices yet
1530 */
1531 max_sectors = bad_sectors;
1532 rdev_dec_pending(rdev, mddev);
1533 /* We don't set R1BIO_Degraded as that
1534 * only applies if the disk is
1535 * missing, so it might be re-added,
1536 * and we want to know to recover this
1537 * chunk.
1538 * In this case the device is here,
1539 * and the fact that this chunk is not
1540 * in-sync is recorded in the bad
1541 * block log
1542 */
1543 continue;
1544 }
1545 if (is_bad) {
1546 int good_sectors = first_bad - r1_bio->sector;
1547 if (good_sectors < max_sectors)
1548 max_sectors = good_sectors;
1549 }
1550 }
1551 r1_bio->bios[i] = bio;
1552 }
1553
1554 if (unlikely(blocked_rdev)) {
1555 /* Wait for this device to become unblocked */
1556 int j;
1557
1558 for (j = 0; j < i; j++)
1559 if (r1_bio->bios[j])
1560 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1561 mempool_free(r1_bio, &conf->r1bio_pool);
1562 allow_barrier(conf, bio->bi_iter.bi_sector);
1563
1564 if (bio->bi_opf & REQ_NOWAIT) {
1565 bio_wouldblock_error(bio);
1566 return;
1567 }
1568 mddev_add_trace_msg(mddev, "raid1 wait rdev %d blocked",
1569 blocked_rdev->raid_disk);
1570 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1571 wait_barrier(conf, bio->bi_iter.bi_sector, false);
1572 goto retry_write;
1573 }
1574
1575 /*
1576 * When using a bitmap, we may call alloc_behind_master_bio below.
1577 * alloc_behind_master_bio allocates a copy of the data payload a page
1578 * at a time and thus needs a new bio that can fit the whole payload
1579 * this bio in page sized chunks.
1580 */
1581 if (write_behind && bitmap)
1582 max_sectors = min_t(int, max_sectors,
1583 BIO_MAX_VECS * (PAGE_SIZE >> 9));
1584 if (max_sectors < bio_sectors(bio)) {
1585 struct bio *split = bio_split(bio, max_sectors,
1586 GFP_NOIO, &conf->bio_split);
1587 bio_chain(split, bio);
1588 submit_bio_noacct(bio);
1589 bio = split;
1590 r1_bio->master_bio = bio;
1591 r1_bio->sectors = max_sectors;
1592 }
1593
1594 md_account_bio(mddev, &bio);
1595 r1_bio->master_bio = bio;
1596 atomic_set(&r1_bio->remaining, 1);
1597 atomic_set(&r1_bio->behind_remaining, 0);
1598
1599 first_clone = 1;
1600
1601 for (i = 0; i < disks; i++) {
1602 struct bio *mbio = NULL;
1603 struct md_rdev *rdev = conf->mirrors[i].rdev;
1604 if (!r1_bio->bios[i])
1605 continue;
1606
1607 if (first_clone) {
1608 /* do behind I/O ?
1609 * Not if there are too many, or cannot
1610 * allocate memory, or a reader on WriteMostly
1611 * is waiting for behind writes to flush */
1612 if (bitmap && write_behind &&
1613 (atomic_read(&bitmap->behind_writes)
1614 < mddev->bitmap_info.max_write_behind) &&
1615 !waitqueue_active(&bitmap->behind_wait)) {
1616 alloc_behind_master_bio(r1_bio, bio);
1617 }
1618
1619 md_bitmap_startwrite(bitmap, r1_bio->sector, r1_bio->sectors,
1620 test_bit(R1BIO_BehindIO, &r1_bio->state));
1621 first_clone = 0;
1622 }
1623
1624 if (r1_bio->behind_master_bio) {
1625 mbio = bio_alloc_clone(rdev->bdev,
1626 r1_bio->behind_master_bio,
1627 GFP_NOIO, &mddev->bio_set);
1628 if (test_bit(CollisionCheck, &rdev->flags))
1629 wait_for_serialization(rdev, r1_bio);
1630 if (test_bit(WriteMostly, &rdev->flags))
1631 atomic_inc(&r1_bio->behind_remaining);
1632 } else {
1633 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
1634 &mddev->bio_set);
1635
1636 if (mddev->serialize_policy)
1637 wait_for_serialization(rdev, r1_bio);
1638 }
1639
1640 r1_bio->bios[i] = mbio;
1641
1642 mbio->bi_iter.bi_sector = (r1_bio->sector + rdev->data_offset);
1643 mbio->bi_end_io = raid1_end_write_request;
1644 mbio->bi_opf = bio_op(bio) | (bio->bi_opf & (REQ_SYNC | REQ_FUA));
1645 if (test_bit(FailFast, &rdev->flags) &&
1646 !test_bit(WriteMostly, &rdev->flags) &&
1647 conf->raid_disks - mddev->degraded > 1)
1648 mbio->bi_opf |= MD_FAILFAST;
1649 mbio->bi_private = r1_bio;
1650
1651 atomic_inc(&r1_bio->remaining);
1652 mddev_trace_remap(mddev, mbio, r1_bio->sector);
1653 /* flush_pending_writes() needs access to the rdev so...*/
1654 mbio->bi_bdev = (void *)rdev;
1655 if (!raid1_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) {
1656 spin_lock_irqsave(&conf->device_lock, flags);
1657 bio_list_add(&conf->pending_bio_list, mbio);
1658 spin_unlock_irqrestore(&conf->device_lock, flags);
1659 md_wakeup_thread(mddev->thread);
1660 }
1661 }
1662
1663 r1_bio_write_done(r1_bio);
1664
1665 /* In case raid1d snuck in to freeze_array */
1666 wake_up_barrier(conf);
1667}
1668
1669static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1670{
1671 sector_t sectors;
1672
1673 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1674 && md_flush_request(mddev, bio))
1675 return true;
1676
1677 /*
1678 * There is a limit to the maximum size, but
1679 * the read/write handler might find a lower limit
1680 * due to bad blocks. To avoid multiple splits,
1681 * we pass the maximum number of sectors down
1682 * and let the lower level perform the split.
1683 */
1684 sectors = align_to_barrier_unit_end(
1685 bio->bi_iter.bi_sector, bio_sectors(bio));
1686
1687 if (bio_data_dir(bio) == READ)
1688 raid1_read_request(mddev, bio, sectors, NULL);
1689 else {
1690 if (!md_write_start(mddev,bio))
1691 return false;
1692 raid1_write_request(mddev, bio, sectors);
1693 }
1694 return true;
1695}
1696
1697static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1698{
1699 struct r1conf *conf = mddev->private;
1700 int i;
1701
1702 lockdep_assert_held(&mddev->lock);
1703
1704 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1705 conf->raid_disks - mddev->degraded);
1706 for (i = 0; i < conf->raid_disks; i++) {
1707 struct md_rdev *rdev = READ_ONCE(conf->mirrors[i].rdev);
1708
1709 seq_printf(seq, "%s",
1710 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1711 }
1712 seq_printf(seq, "]");
1713}
1714
1715/**
1716 * raid1_error() - RAID1 error handler.
1717 * @mddev: affected md device.
1718 * @rdev: member device to fail.
1719 *
1720 * The routine acknowledges &rdev failure and determines new @mddev state.
1721 * If it failed, then:
1722 * - &MD_BROKEN flag is set in &mddev->flags.
1723 * - recovery is disabled.
1724 * Otherwise, it must be degraded:
1725 * - recovery is interrupted.
1726 * - &mddev->degraded is bumped.
1727 *
1728 * @rdev is marked as &Faulty excluding case when array is failed and
1729 * &mddev->fail_last_dev is off.
1730 */
1731static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1732{
1733 struct r1conf *conf = mddev->private;
1734 unsigned long flags;
1735
1736 spin_lock_irqsave(&conf->device_lock, flags);
1737
1738 if (test_bit(In_sync, &rdev->flags) &&
1739 (conf->raid_disks - mddev->degraded) == 1) {
1740 set_bit(MD_BROKEN, &mddev->flags);
1741
1742 if (!mddev->fail_last_dev) {
1743 conf->recovery_disabled = mddev->recovery_disabled;
1744 spin_unlock_irqrestore(&conf->device_lock, flags);
1745 return;
1746 }
1747 }
1748 set_bit(Blocked, &rdev->flags);
1749 if (test_and_clear_bit(In_sync, &rdev->flags))
1750 mddev->degraded++;
1751 set_bit(Faulty, &rdev->flags);
1752 spin_unlock_irqrestore(&conf->device_lock, flags);
1753 /*
1754 * if recovery is running, make sure it aborts.
1755 */
1756 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1757 set_mask_bits(&mddev->sb_flags, 0,
1758 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1759 pr_crit("md/raid1:%s: Disk failure on %pg, disabling device.\n"
1760 "md/raid1:%s: Operation continuing on %d devices.\n",
1761 mdname(mddev), rdev->bdev,
1762 mdname(mddev), conf->raid_disks - mddev->degraded);
1763}
1764
1765static void print_conf(struct r1conf *conf)
1766{
1767 int i;
1768
1769 pr_debug("RAID1 conf printout:\n");
1770 if (!conf) {
1771 pr_debug("(!conf)\n");
1772 return;
1773 }
1774 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1775 conf->raid_disks);
1776
1777 lockdep_assert_held(&conf->mddev->reconfig_mutex);
1778 for (i = 0; i < conf->raid_disks; i++) {
1779 struct md_rdev *rdev = conf->mirrors[i].rdev;
1780 if (rdev)
1781 pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
1782 i, !test_bit(In_sync, &rdev->flags),
1783 !test_bit(Faulty, &rdev->flags),
1784 rdev->bdev);
1785 }
1786}
1787
1788static void close_sync(struct r1conf *conf)
1789{
1790 int idx;
1791
1792 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1793 _wait_barrier(conf, idx, false);
1794 _allow_barrier(conf, idx);
1795 }
1796
1797 mempool_exit(&conf->r1buf_pool);
1798}
1799
1800static int raid1_spare_active(struct mddev *mddev)
1801{
1802 int i;
1803 struct r1conf *conf = mddev->private;
1804 int count = 0;
1805 unsigned long flags;
1806
1807 /*
1808 * Find all failed disks within the RAID1 configuration
1809 * and mark them readable.
1810 * Called under mddev lock, so rcu protection not needed.
1811 * device_lock used to avoid races with raid1_end_read_request
1812 * which expects 'In_sync' flags and ->degraded to be consistent.
1813 */
1814 spin_lock_irqsave(&conf->device_lock, flags);
1815 for (i = 0; i < conf->raid_disks; i++) {
1816 struct md_rdev *rdev = conf->mirrors[i].rdev;
1817 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1818 if (repl
1819 && !test_bit(Candidate, &repl->flags)
1820 && repl->recovery_offset == MaxSector
1821 && !test_bit(Faulty, &repl->flags)
1822 && !test_and_set_bit(In_sync, &repl->flags)) {
1823 /* replacement has just become active */
1824 if (!rdev ||
1825 !test_and_clear_bit(In_sync, &rdev->flags))
1826 count++;
1827 if (rdev) {
1828 /* Replaced device not technically
1829 * faulty, but we need to be sure
1830 * it gets removed and never re-added
1831 */
1832 set_bit(Faulty, &rdev->flags);
1833 sysfs_notify_dirent_safe(
1834 rdev->sysfs_state);
1835 }
1836 }
1837 if (rdev
1838 && rdev->recovery_offset == MaxSector
1839 && !test_bit(Faulty, &rdev->flags)
1840 && !test_and_set_bit(In_sync, &rdev->flags)) {
1841 count++;
1842 sysfs_notify_dirent_safe(rdev->sysfs_state);
1843 }
1844 }
1845 mddev->degraded -= count;
1846 spin_unlock_irqrestore(&conf->device_lock, flags);
1847
1848 print_conf(conf);
1849 return count;
1850}
1851
1852static bool raid1_add_conf(struct r1conf *conf, struct md_rdev *rdev, int disk,
1853 bool replacement)
1854{
1855 struct raid1_info *info = conf->mirrors + disk;
1856
1857 if (replacement)
1858 info += conf->raid_disks;
1859
1860 if (info->rdev)
1861 return false;
1862
1863 if (bdev_nonrot(rdev->bdev)) {
1864 set_bit(Nonrot, &rdev->flags);
1865 WRITE_ONCE(conf->nonrot_disks, conf->nonrot_disks + 1);
1866 }
1867
1868 rdev->raid_disk = disk;
1869 info->head_position = 0;
1870 info->seq_start = MaxSector;
1871 WRITE_ONCE(info->rdev, rdev);
1872
1873 return true;
1874}
1875
1876static bool raid1_remove_conf(struct r1conf *conf, int disk)
1877{
1878 struct raid1_info *info = conf->mirrors + disk;
1879 struct md_rdev *rdev = info->rdev;
1880
1881 if (!rdev || test_bit(In_sync, &rdev->flags) ||
1882 atomic_read(&rdev->nr_pending))
1883 return false;
1884
1885 /* Only remove non-faulty devices if recovery is not possible. */
1886 if (!test_bit(Faulty, &rdev->flags) &&
1887 rdev->mddev->recovery_disabled != conf->recovery_disabled &&
1888 rdev->mddev->degraded < conf->raid_disks)
1889 return false;
1890
1891 if (test_and_clear_bit(Nonrot, &rdev->flags))
1892 WRITE_ONCE(conf->nonrot_disks, conf->nonrot_disks - 1);
1893
1894 WRITE_ONCE(info->rdev, NULL);
1895 return true;
1896}
1897
1898static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1899{
1900 struct r1conf *conf = mddev->private;
1901 int err = -EEXIST;
1902 int mirror = 0, repl_slot = -1;
1903 struct raid1_info *p;
1904 int first = 0;
1905 int last = conf->raid_disks - 1;
1906
1907 if (mddev->recovery_disabled == conf->recovery_disabled)
1908 return -EBUSY;
1909
1910 if (md_integrity_add_rdev(rdev, mddev))
1911 return -ENXIO;
1912
1913 if (rdev->raid_disk >= 0)
1914 first = last = rdev->raid_disk;
1915
1916 /*
1917 * find the disk ... but prefer rdev->saved_raid_disk
1918 * if possible.
1919 */
1920 if (rdev->saved_raid_disk >= 0 &&
1921 rdev->saved_raid_disk >= first &&
1922 rdev->saved_raid_disk < conf->raid_disks &&
1923 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1924 first = last = rdev->saved_raid_disk;
1925
1926 for (mirror = first; mirror <= last; mirror++) {
1927 p = conf->mirrors + mirror;
1928 if (!p->rdev) {
1929 err = mddev_stack_new_rdev(mddev, rdev);
1930 if (err)
1931 return err;
1932
1933 raid1_add_conf(conf, rdev, mirror, false);
1934 /* As all devices are equivalent, we don't need a full recovery
1935 * if this was recently any drive of the array
1936 */
1937 if (rdev->saved_raid_disk < 0)
1938 conf->fullsync = 1;
1939 break;
1940 }
1941 if (test_bit(WantReplacement, &p->rdev->flags) &&
1942 p[conf->raid_disks].rdev == NULL && repl_slot < 0)
1943 repl_slot = mirror;
1944 }
1945
1946 if (err && repl_slot >= 0) {
1947 /* Add this device as a replacement */
1948 clear_bit(In_sync, &rdev->flags);
1949 set_bit(Replacement, &rdev->flags);
1950 raid1_add_conf(conf, rdev, repl_slot, true);
1951 err = 0;
1952 conf->fullsync = 1;
1953 }
1954
1955 print_conf(conf);
1956 return err;
1957}
1958
1959static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1960{
1961 struct r1conf *conf = mddev->private;
1962 int err = 0;
1963 int number = rdev->raid_disk;
1964 struct raid1_info *p = conf->mirrors + number;
1965
1966 if (unlikely(number >= conf->raid_disks))
1967 goto abort;
1968
1969 if (rdev != p->rdev) {
1970 number += conf->raid_disks;
1971 p = conf->mirrors + number;
1972 }
1973
1974 print_conf(conf);
1975 if (rdev == p->rdev) {
1976 if (!raid1_remove_conf(conf, number)) {
1977 err = -EBUSY;
1978 goto abort;
1979 }
1980
1981 if (number < conf->raid_disks &&
1982 conf->mirrors[conf->raid_disks + number].rdev) {
1983 /* We just removed a device that is being replaced.
1984 * Move down the replacement. We drain all IO before
1985 * doing this to avoid confusion.
1986 */
1987 struct md_rdev *repl =
1988 conf->mirrors[conf->raid_disks + number].rdev;
1989 freeze_array(conf, 0);
1990 if (atomic_read(&repl->nr_pending)) {
1991 /* It means that some queued IO of retry_list
1992 * hold repl. Thus, we cannot set replacement
1993 * as NULL, avoiding rdev NULL pointer
1994 * dereference in sync_request_write and
1995 * handle_write_finished.
1996 */
1997 err = -EBUSY;
1998 unfreeze_array(conf);
1999 goto abort;
2000 }
2001 clear_bit(Replacement, &repl->flags);
2002 WRITE_ONCE(p->rdev, repl);
2003 conf->mirrors[conf->raid_disks + number].rdev = NULL;
2004 unfreeze_array(conf);
2005 }
2006
2007 clear_bit(WantReplacement, &rdev->flags);
2008 err = md_integrity_register(mddev);
2009 }
2010abort:
2011
2012 print_conf(conf);
2013 return err;
2014}
2015
2016static void end_sync_read(struct bio *bio)
2017{
2018 struct r1bio *r1_bio = get_resync_r1bio(bio);
2019
2020 update_head_pos(r1_bio->read_disk, r1_bio);
2021
2022 /*
2023 * we have read a block, now it needs to be re-written,
2024 * or re-read if the read failed.
2025 * We don't do much here, just schedule handling by raid1d
2026 */
2027 if (!bio->bi_status)
2028 set_bit(R1BIO_Uptodate, &r1_bio->state);
2029
2030 if (atomic_dec_and_test(&r1_bio->remaining))
2031 reschedule_retry(r1_bio);
2032}
2033
2034static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
2035{
2036 sector_t sync_blocks = 0;
2037 sector_t s = r1_bio->sector;
2038 long sectors_to_go = r1_bio->sectors;
2039
2040 /* make sure these bits don't get cleared. */
2041 do {
2042 md_bitmap_end_sync(mddev->bitmap, s, &sync_blocks, 1);
2043 s += sync_blocks;
2044 sectors_to_go -= sync_blocks;
2045 } while (sectors_to_go > 0);
2046}
2047
2048static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
2049{
2050 if (atomic_dec_and_test(&r1_bio->remaining)) {
2051 struct mddev *mddev = r1_bio->mddev;
2052 int s = r1_bio->sectors;
2053
2054 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2055 test_bit(R1BIO_WriteError, &r1_bio->state))
2056 reschedule_retry(r1_bio);
2057 else {
2058 put_buf(r1_bio);
2059 md_done_sync(mddev, s, uptodate);
2060 }
2061 }
2062}
2063
2064static void end_sync_write(struct bio *bio)
2065{
2066 int uptodate = !bio->bi_status;
2067 struct r1bio *r1_bio = get_resync_r1bio(bio);
2068 struct mddev *mddev = r1_bio->mddev;
2069 struct r1conf *conf = mddev->private;
2070 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
2071
2072 if (!uptodate) {
2073 abort_sync_write(mddev, r1_bio);
2074 set_bit(WriteErrorSeen, &rdev->flags);
2075 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2076 set_bit(MD_RECOVERY_NEEDED, &
2077 mddev->recovery);
2078 set_bit(R1BIO_WriteError, &r1_bio->state);
2079 } else if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors) &&
2080 !rdev_has_badblock(conf->mirrors[r1_bio->read_disk].rdev,
2081 r1_bio->sector, r1_bio->sectors)) {
2082 set_bit(R1BIO_MadeGood, &r1_bio->state);
2083 }
2084
2085 put_sync_write_buf(r1_bio, uptodate);
2086}
2087
2088static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
2089 int sectors, struct page *page, blk_opf_t rw)
2090{
2091 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2092 /* success */
2093 return 1;
2094 if (rw == REQ_OP_WRITE) {
2095 set_bit(WriteErrorSeen, &rdev->flags);
2096 if (!test_and_set_bit(WantReplacement,
2097 &rdev->flags))
2098 set_bit(MD_RECOVERY_NEEDED, &
2099 rdev->mddev->recovery);
2100 }
2101 /* need to record an error - either for the block or the device */
2102 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2103 md_error(rdev->mddev, rdev);
2104 return 0;
2105}
2106
2107static int fix_sync_read_error(struct r1bio *r1_bio)
2108{
2109 /* Try some synchronous reads of other devices to get
2110 * good data, much like with normal read errors. Only
2111 * read into the pages we already have so we don't
2112 * need to re-issue the read request.
2113 * We don't need to freeze the array, because being in an
2114 * active sync request, there is no normal IO, and
2115 * no overlapping syncs.
2116 * We don't need to check is_badblock() again as we
2117 * made sure that anything with a bad block in range
2118 * will have bi_end_io clear.
2119 */
2120 struct mddev *mddev = r1_bio->mddev;
2121 struct r1conf *conf = mddev->private;
2122 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2123 struct page **pages = get_resync_pages(bio)->pages;
2124 sector_t sect = r1_bio->sector;
2125 int sectors = r1_bio->sectors;
2126 int idx = 0;
2127 struct md_rdev *rdev;
2128
2129 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2130 if (test_bit(FailFast, &rdev->flags)) {
2131 /* Don't try recovering from here - just fail it
2132 * ... unless it is the last working device of course */
2133 md_error(mddev, rdev);
2134 if (test_bit(Faulty, &rdev->flags))
2135 /* Don't try to read from here, but make sure
2136 * put_buf does it's thing
2137 */
2138 bio->bi_end_io = end_sync_write;
2139 }
2140
2141 while(sectors) {
2142 int s = sectors;
2143 int d = r1_bio->read_disk;
2144 int success = 0;
2145 int start;
2146
2147 if (s > (PAGE_SIZE>>9))
2148 s = PAGE_SIZE >> 9;
2149 do {
2150 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2151 /* No rcu protection needed here devices
2152 * can only be removed when no resync is
2153 * active, and resync is currently active
2154 */
2155 rdev = conf->mirrors[d].rdev;
2156 if (sync_page_io(rdev, sect, s<<9,
2157 pages[idx],
2158 REQ_OP_READ, false)) {
2159 success = 1;
2160 break;
2161 }
2162 }
2163 d++;
2164 if (d == conf->raid_disks * 2)
2165 d = 0;
2166 } while (!success && d != r1_bio->read_disk);
2167
2168 if (!success) {
2169 int abort = 0;
2170 /* Cannot read from anywhere, this block is lost.
2171 * Record a bad block on each device. If that doesn't
2172 * work just disable and interrupt the recovery.
2173 * Don't fail devices as that won't really help.
2174 */
2175 pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
2176 mdname(mddev), bio->bi_bdev,
2177 (unsigned long long)r1_bio->sector);
2178 for (d = 0; d < conf->raid_disks * 2; d++) {
2179 rdev = conf->mirrors[d].rdev;
2180 if (!rdev || test_bit(Faulty, &rdev->flags))
2181 continue;
2182 if (!rdev_set_badblocks(rdev, sect, s, 0))
2183 abort = 1;
2184 }
2185 if (abort) {
2186 conf->recovery_disabled =
2187 mddev->recovery_disabled;
2188 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2189 md_done_sync(mddev, r1_bio->sectors, 0);
2190 put_buf(r1_bio);
2191 return 0;
2192 }
2193 /* Try next page */
2194 sectors -= s;
2195 sect += s;
2196 idx++;
2197 continue;
2198 }
2199
2200 start = d;
2201 /* write it back and re-read */
2202 while (d != r1_bio->read_disk) {
2203 if (d == 0)
2204 d = conf->raid_disks * 2;
2205 d--;
2206 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2207 continue;
2208 rdev = conf->mirrors[d].rdev;
2209 if (r1_sync_page_io(rdev, sect, s,
2210 pages[idx],
2211 REQ_OP_WRITE) == 0) {
2212 r1_bio->bios[d]->bi_end_io = NULL;
2213 rdev_dec_pending(rdev, mddev);
2214 }
2215 }
2216 d = start;
2217 while (d != r1_bio->read_disk) {
2218 if (d == 0)
2219 d = conf->raid_disks * 2;
2220 d--;
2221 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2222 continue;
2223 rdev = conf->mirrors[d].rdev;
2224 if (r1_sync_page_io(rdev, sect, s,
2225 pages[idx],
2226 REQ_OP_READ) != 0)
2227 atomic_add(s, &rdev->corrected_errors);
2228 }
2229 sectors -= s;
2230 sect += s;
2231 idx ++;
2232 }
2233 set_bit(R1BIO_Uptodate, &r1_bio->state);
2234 bio->bi_status = 0;
2235 return 1;
2236}
2237
2238static void process_checks(struct r1bio *r1_bio)
2239{
2240 /* We have read all readable devices. If we haven't
2241 * got the block, then there is no hope left.
2242 * If we have, then we want to do a comparison
2243 * and skip the write if everything is the same.
2244 * If any blocks failed to read, then we need to
2245 * attempt an over-write
2246 */
2247 struct mddev *mddev = r1_bio->mddev;
2248 struct r1conf *conf = mddev->private;
2249 int primary;
2250 int i;
2251 int vcnt;
2252
2253 /* Fix variable parts of all bios */
2254 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2255 for (i = 0; i < conf->raid_disks * 2; i++) {
2256 blk_status_t status;
2257 struct bio *b = r1_bio->bios[i];
2258 struct resync_pages *rp = get_resync_pages(b);
2259 if (b->bi_end_io != end_sync_read)
2260 continue;
2261 /* fixup the bio for reuse, but preserve errno */
2262 status = b->bi_status;
2263 bio_reset(b, conf->mirrors[i].rdev->bdev, REQ_OP_READ);
2264 b->bi_status = status;
2265 b->bi_iter.bi_sector = r1_bio->sector +
2266 conf->mirrors[i].rdev->data_offset;
2267 b->bi_end_io = end_sync_read;
2268 rp->raid_bio = r1_bio;
2269 b->bi_private = rp;
2270
2271 /* initialize bvec table again */
2272 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2273 }
2274 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2275 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2276 !r1_bio->bios[primary]->bi_status) {
2277 r1_bio->bios[primary]->bi_end_io = NULL;
2278 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2279 break;
2280 }
2281 r1_bio->read_disk = primary;
2282 for (i = 0; i < conf->raid_disks * 2; i++) {
2283 int j = 0;
2284 struct bio *pbio = r1_bio->bios[primary];
2285 struct bio *sbio = r1_bio->bios[i];
2286 blk_status_t status = sbio->bi_status;
2287 struct page **ppages = get_resync_pages(pbio)->pages;
2288 struct page **spages = get_resync_pages(sbio)->pages;
2289 struct bio_vec *bi;
2290 int page_len[RESYNC_PAGES] = { 0 };
2291 struct bvec_iter_all iter_all;
2292
2293 if (sbio->bi_end_io != end_sync_read)
2294 continue;
2295 /* Now we can 'fixup' the error value */
2296 sbio->bi_status = 0;
2297
2298 bio_for_each_segment_all(bi, sbio, iter_all)
2299 page_len[j++] = bi->bv_len;
2300
2301 if (!status) {
2302 for (j = vcnt; j-- ; ) {
2303 if (memcmp(page_address(ppages[j]),
2304 page_address(spages[j]),
2305 page_len[j]))
2306 break;
2307 }
2308 } else
2309 j = 0;
2310 if (j >= 0)
2311 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2312 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2313 && !status)) {
2314 /* No need to write to this device. */
2315 sbio->bi_end_io = NULL;
2316 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2317 continue;
2318 }
2319
2320 bio_copy_data(sbio, pbio);
2321 }
2322}
2323
2324static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2325{
2326 struct r1conf *conf = mddev->private;
2327 int i;
2328 int disks = conf->raid_disks * 2;
2329 struct bio *wbio;
2330
2331 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2332 /* ouch - failed to read all of that. */
2333 if (!fix_sync_read_error(r1_bio))
2334 return;
2335
2336 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2337 process_checks(r1_bio);
2338
2339 /*
2340 * schedule writes
2341 */
2342 atomic_set(&r1_bio->remaining, 1);
2343 for (i = 0; i < disks ; i++) {
2344 wbio = r1_bio->bios[i];
2345 if (wbio->bi_end_io == NULL ||
2346 (wbio->bi_end_io == end_sync_read &&
2347 (i == r1_bio->read_disk ||
2348 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2349 continue;
2350 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2351 abort_sync_write(mddev, r1_bio);
2352 continue;
2353 }
2354
2355 wbio->bi_opf = REQ_OP_WRITE;
2356 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2357 wbio->bi_opf |= MD_FAILFAST;
2358
2359 wbio->bi_end_io = end_sync_write;
2360 atomic_inc(&r1_bio->remaining);
2361 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2362
2363 submit_bio_noacct(wbio);
2364 }
2365
2366 put_sync_write_buf(r1_bio, 1);
2367}
2368
2369/*
2370 * This is a kernel thread which:
2371 *
2372 * 1. Retries failed read operations on working mirrors.
2373 * 2. Updates the raid superblock when problems encounter.
2374 * 3. Performs writes following reads for array synchronising.
2375 */
2376
2377static void fix_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2378{
2379 sector_t sect = r1_bio->sector;
2380 int sectors = r1_bio->sectors;
2381 int read_disk = r1_bio->read_disk;
2382 struct mddev *mddev = conf->mddev;
2383 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2384
2385 if (exceed_read_errors(mddev, rdev)) {
2386 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2387 return;
2388 }
2389
2390 while(sectors) {
2391 int s = sectors;
2392 int d = read_disk;
2393 int success = 0;
2394 int start;
2395
2396 if (s > (PAGE_SIZE>>9))
2397 s = PAGE_SIZE >> 9;
2398
2399 do {
2400 rdev = conf->mirrors[d].rdev;
2401 if (rdev &&
2402 (test_bit(In_sync, &rdev->flags) ||
2403 (!test_bit(Faulty, &rdev->flags) &&
2404 rdev->recovery_offset >= sect + s)) &&
2405 rdev_has_badblock(rdev, sect, s) == 0) {
2406 atomic_inc(&rdev->nr_pending);
2407 if (sync_page_io(rdev, sect, s<<9,
2408 conf->tmppage, REQ_OP_READ, false))
2409 success = 1;
2410 rdev_dec_pending(rdev, mddev);
2411 if (success)
2412 break;
2413 }
2414
2415 d++;
2416 if (d == conf->raid_disks * 2)
2417 d = 0;
2418 } while (d != read_disk);
2419
2420 if (!success) {
2421 /* Cannot read from anywhere - mark it bad */
2422 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2423 if (!rdev_set_badblocks(rdev, sect, s, 0))
2424 md_error(mddev, rdev);
2425 break;
2426 }
2427 /* write it back and re-read */
2428 start = d;
2429 while (d != read_disk) {
2430 if (d==0)
2431 d = conf->raid_disks * 2;
2432 d--;
2433 rdev = conf->mirrors[d].rdev;
2434 if (rdev &&
2435 !test_bit(Faulty, &rdev->flags)) {
2436 atomic_inc(&rdev->nr_pending);
2437 r1_sync_page_io(rdev, sect, s,
2438 conf->tmppage, REQ_OP_WRITE);
2439 rdev_dec_pending(rdev, mddev);
2440 }
2441 }
2442 d = start;
2443 while (d != read_disk) {
2444 if (d==0)
2445 d = conf->raid_disks * 2;
2446 d--;
2447 rdev = conf->mirrors[d].rdev;
2448 if (rdev &&
2449 !test_bit(Faulty, &rdev->flags)) {
2450 atomic_inc(&rdev->nr_pending);
2451 if (r1_sync_page_io(rdev, sect, s,
2452 conf->tmppage, REQ_OP_READ)) {
2453 atomic_add(s, &rdev->corrected_errors);
2454 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %pg)\n",
2455 mdname(mddev), s,
2456 (unsigned long long)(sect +
2457 rdev->data_offset),
2458 rdev->bdev);
2459 }
2460 rdev_dec_pending(rdev, mddev);
2461 }
2462 }
2463 sectors -= s;
2464 sect += s;
2465 }
2466}
2467
2468static int narrow_write_error(struct r1bio *r1_bio, int i)
2469{
2470 struct mddev *mddev = r1_bio->mddev;
2471 struct r1conf *conf = mddev->private;
2472 struct md_rdev *rdev = conf->mirrors[i].rdev;
2473
2474 /* bio has the data to be written to device 'i' where
2475 * we just recently had a write error.
2476 * We repeatedly clone the bio and trim down to one block,
2477 * then try the write. Where the write fails we record
2478 * a bad block.
2479 * It is conceivable that the bio doesn't exactly align with
2480 * blocks. We must handle this somehow.
2481 *
2482 * We currently own a reference on the rdev.
2483 */
2484
2485 int block_sectors;
2486 sector_t sector;
2487 int sectors;
2488 int sect_to_write = r1_bio->sectors;
2489 int ok = 1;
2490
2491 if (rdev->badblocks.shift < 0)
2492 return 0;
2493
2494 block_sectors = roundup(1 << rdev->badblocks.shift,
2495 bdev_logical_block_size(rdev->bdev) >> 9);
2496 sector = r1_bio->sector;
2497 sectors = ((sector + block_sectors)
2498 & ~(sector_t)(block_sectors - 1))
2499 - sector;
2500
2501 while (sect_to_write) {
2502 struct bio *wbio;
2503 if (sectors > sect_to_write)
2504 sectors = sect_to_write;
2505 /* Write at 'sector' for 'sectors'*/
2506
2507 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2508 wbio = bio_alloc_clone(rdev->bdev,
2509 r1_bio->behind_master_bio,
2510 GFP_NOIO, &mddev->bio_set);
2511 } else {
2512 wbio = bio_alloc_clone(rdev->bdev, r1_bio->master_bio,
2513 GFP_NOIO, &mddev->bio_set);
2514 }
2515
2516 wbio->bi_opf = REQ_OP_WRITE;
2517 wbio->bi_iter.bi_sector = r1_bio->sector;
2518 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2519
2520 bio_trim(wbio, sector - r1_bio->sector, sectors);
2521 wbio->bi_iter.bi_sector += rdev->data_offset;
2522
2523 if (submit_bio_wait(wbio) < 0)
2524 /* failure! */
2525 ok = rdev_set_badblocks(rdev, sector,
2526 sectors, 0)
2527 && ok;
2528
2529 bio_put(wbio);
2530 sect_to_write -= sectors;
2531 sector += sectors;
2532 sectors = block_sectors;
2533 }
2534 return ok;
2535}
2536
2537static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2538{
2539 int m;
2540 int s = r1_bio->sectors;
2541 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2542 struct md_rdev *rdev = conf->mirrors[m].rdev;
2543 struct bio *bio = r1_bio->bios[m];
2544 if (bio->bi_end_io == NULL)
2545 continue;
2546 if (!bio->bi_status &&
2547 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2548 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2549 }
2550 if (bio->bi_status &&
2551 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2552 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2553 md_error(conf->mddev, rdev);
2554 }
2555 }
2556 put_buf(r1_bio);
2557 md_done_sync(conf->mddev, s, 1);
2558}
2559
2560static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2561{
2562 int m, idx;
2563 bool fail = false;
2564
2565 for (m = 0; m < conf->raid_disks * 2 ; m++)
2566 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2567 struct md_rdev *rdev = conf->mirrors[m].rdev;
2568 rdev_clear_badblocks(rdev,
2569 r1_bio->sector,
2570 r1_bio->sectors, 0);
2571 rdev_dec_pending(rdev, conf->mddev);
2572 } else if (r1_bio->bios[m] != NULL) {
2573 /* This drive got a write error. We need to
2574 * narrow down and record precise write
2575 * errors.
2576 */
2577 fail = true;
2578 if (!narrow_write_error(r1_bio, m)) {
2579 md_error(conf->mddev,
2580 conf->mirrors[m].rdev);
2581 /* an I/O failed, we can't clear the bitmap */
2582 set_bit(R1BIO_Degraded, &r1_bio->state);
2583 }
2584 rdev_dec_pending(conf->mirrors[m].rdev,
2585 conf->mddev);
2586 }
2587 if (fail) {
2588 spin_lock_irq(&conf->device_lock);
2589 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2590 idx = sector_to_idx(r1_bio->sector);
2591 atomic_inc(&conf->nr_queued[idx]);
2592 spin_unlock_irq(&conf->device_lock);
2593 /*
2594 * In case freeze_array() is waiting for condition
2595 * get_unqueued_pending() == extra to be true.
2596 */
2597 wake_up(&conf->wait_barrier);
2598 md_wakeup_thread(conf->mddev->thread);
2599 } else {
2600 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2601 close_write(r1_bio);
2602 raid_end_bio_io(r1_bio);
2603 }
2604}
2605
2606static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2607{
2608 struct mddev *mddev = conf->mddev;
2609 struct bio *bio;
2610 struct md_rdev *rdev;
2611 sector_t sector;
2612
2613 clear_bit(R1BIO_ReadError, &r1_bio->state);
2614 /* we got a read error. Maybe the drive is bad. Maybe just
2615 * the block and we can fix it.
2616 * We freeze all other IO, and try reading the block from
2617 * other devices. When we find one, we re-write
2618 * and check it that fixes the read error.
2619 * This is all done synchronously while the array is
2620 * frozen
2621 */
2622
2623 bio = r1_bio->bios[r1_bio->read_disk];
2624 bio_put(bio);
2625 r1_bio->bios[r1_bio->read_disk] = NULL;
2626
2627 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2628 if (mddev->ro == 0
2629 && !test_bit(FailFast, &rdev->flags)) {
2630 freeze_array(conf, 1);
2631 fix_read_error(conf, r1_bio);
2632 unfreeze_array(conf);
2633 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2634 md_error(mddev, rdev);
2635 } else {
2636 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2637 }
2638
2639 rdev_dec_pending(rdev, conf->mddev);
2640 sector = r1_bio->sector;
2641 bio = r1_bio->master_bio;
2642
2643 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2644 r1_bio->state = 0;
2645 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2646 allow_barrier(conf, sector);
2647}
2648
2649static void raid1d(struct md_thread *thread)
2650{
2651 struct mddev *mddev = thread->mddev;
2652 struct r1bio *r1_bio;
2653 unsigned long flags;
2654 struct r1conf *conf = mddev->private;
2655 struct list_head *head = &conf->retry_list;
2656 struct blk_plug plug;
2657 int idx;
2658
2659 md_check_recovery(mddev);
2660
2661 if (!list_empty_careful(&conf->bio_end_io_list) &&
2662 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2663 LIST_HEAD(tmp);
2664 spin_lock_irqsave(&conf->device_lock, flags);
2665 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2666 list_splice_init(&conf->bio_end_io_list, &tmp);
2667 spin_unlock_irqrestore(&conf->device_lock, flags);
2668 while (!list_empty(&tmp)) {
2669 r1_bio = list_first_entry(&tmp, struct r1bio,
2670 retry_list);
2671 list_del(&r1_bio->retry_list);
2672 idx = sector_to_idx(r1_bio->sector);
2673 atomic_dec(&conf->nr_queued[idx]);
2674 if (mddev->degraded)
2675 set_bit(R1BIO_Degraded, &r1_bio->state);
2676 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2677 close_write(r1_bio);
2678 raid_end_bio_io(r1_bio);
2679 }
2680 }
2681
2682 blk_start_plug(&plug);
2683 for (;;) {
2684
2685 flush_pending_writes(conf);
2686
2687 spin_lock_irqsave(&conf->device_lock, flags);
2688 if (list_empty(head)) {
2689 spin_unlock_irqrestore(&conf->device_lock, flags);
2690 break;
2691 }
2692 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2693 list_del(head->prev);
2694 idx = sector_to_idx(r1_bio->sector);
2695 atomic_dec(&conf->nr_queued[idx]);
2696 spin_unlock_irqrestore(&conf->device_lock, flags);
2697
2698 mddev = r1_bio->mddev;
2699 conf = mddev->private;
2700 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2701 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2702 test_bit(R1BIO_WriteError, &r1_bio->state))
2703 handle_sync_write_finished(conf, r1_bio);
2704 else
2705 sync_request_write(mddev, r1_bio);
2706 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2707 test_bit(R1BIO_WriteError, &r1_bio->state))
2708 handle_write_finished(conf, r1_bio);
2709 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2710 handle_read_error(conf, r1_bio);
2711 else
2712 WARN_ON_ONCE(1);
2713
2714 cond_resched();
2715 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2716 md_check_recovery(mddev);
2717 }
2718 blk_finish_plug(&plug);
2719}
2720
2721static int init_resync(struct r1conf *conf)
2722{
2723 int buffs;
2724
2725 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2726 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2727
2728 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2729 r1buf_pool_free, conf->poolinfo);
2730}
2731
2732static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2733{
2734 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2735 struct resync_pages *rps;
2736 struct bio *bio;
2737 int i;
2738
2739 for (i = conf->poolinfo->raid_disks; i--; ) {
2740 bio = r1bio->bios[i];
2741 rps = bio->bi_private;
2742 bio_reset(bio, NULL, 0);
2743 bio->bi_private = rps;
2744 }
2745 r1bio->master_bio = NULL;
2746 return r1bio;
2747}
2748
2749/*
2750 * perform a "sync" on one "block"
2751 *
2752 * We need to make sure that no normal I/O request - particularly write
2753 * requests - conflict with active sync requests.
2754 *
2755 * This is achieved by tracking pending requests and a 'barrier' concept
2756 * that can be installed to exclude normal IO requests.
2757 */
2758
2759static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2760 int *skipped)
2761{
2762 struct r1conf *conf = mddev->private;
2763 struct r1bio *r1_bio;
2764 struct bio *bio;
2765 sector_t max_sector, nr_sectors;
2766 int disk = -1;
2767 int i;
2768 int wonly = -1;
2769 int write_targets = 0, read_targets = 0;
2770 sector_t sync_blocks;
2771 int still_degraded = 0;
2772 int good_sectors = RESYNC_SECTORS;
2773 int min_bad = 0; /* number of sectors that are bad in all devices */
2774 int idx = sector_to_idx(sector_nr);
2775 int page_idx = 0;
2776
2777 if (!mempool_initialized(&conf->r1buf_pool))
2778 if (init_resync(conf))
2779 return 0;
2780
2781 max_sector = mddev->dev_sectors;
2782 if (sector_nr >= max_sector) {
2783 /* If we aborted, we need to abort the
2784 * sync on the 'current' bitmap chunk (there will
2785 * only be one in raid1 resync.
2786 * We can find the current addess in mddev->curr_resync
2787 */
2788 if (mddev->curr_resync < max_sector) /* aborted */
2789 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2790 &sync_blocks, 1);
2791 else /* completed sync */
2792 conf->fullsync = 0;
2793
2794 md_bitmap_close_sync(mddev->bitmap);
2795 close_sync(conf);
2796
2797 if (mddev_is_clustered(mddev)) {
2798 conf->cluster_sync_low = 0;
2799 conf->cluster_sync_high = 0;
2800 }
2801 return 0;
2802 }
2803
2804 if (mddev->bitmap == NULL &&
2805 mddev->recovery_cp == MaxSector &&
2806 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2807 conf->fullsync == 0) {
2808 *skipped = 1;
2809 return max_sector - sector_nr;
2810 }
2811 /* before building a request, check if we can skip these blocks..
2812 * This call the bitmap_start_sync doesn't actually record anything
2813 */
2814 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2815 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2816 /* We can skip this block, and probably several more */
2817 *skipped = 1;
2818 return sync_blocks;
2819 }
2820
2821 /*
2822 * If there is non-resync activity waiting for a turn, then let it
2823 * though before starting on this new sync request.
2824 */
2825 if (atomic_read(&conf->nr_waiting[idx]))
2826 schedule_timeout_uninterruptible(1);
2827
2828 /* we are incrementing sector_nr below. To be safe, we check against
2829 * sector_nr + two times RESYNC_SECTORS
2830 */
2831
2832 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
2833 mddev_is_clustered(mddev) && (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2834
2835
2836 if (raise_barrier(conf, sector_nr))
2837 return 0;
2838
2839 r1_bio = raid1_alloc_init_r1buf(conf);
2840
2841 /*
2842 * If we get a correctably read error during resync or recovery,
2843 * we might want to read from a different device. So we
2844 * flag all drives that could conceivably be read from for READ,
2845 * and any others (which will be non-In_sync devices) for WRITE.
2846 * If a read fails, we try reading from something else for which READ
2847 * is OK.
2848 */
2849
2850 r1_bio->mddev = mddev;
2851 r1_bio->sector = sector_nr;
2852 r1_bio->state = 0;
2853 set_bit(R1BIO_IsSync, &r1_bio->state);
2854 /* make sure good_sectors won't go across barrier unit boundary */
2855 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2856
2857 for (i = 0; i < conf->raid_disks * 2; i++) {
2858 struct md_rdev *rdev;
2859 bio = r1_bio->bios[i];
2860
2861 rdev = conf->mirrors[i].rdev;
2862 if (rdev == NULL ||
2863 test_bit(Faulty, &rdev->flags)) {
2864 if (i < conf->raid_disks)
2865 still_degraded = 1;
2866 } else if (!test_bit(In_sync, &rdev->flags)) {
2867 bio->bi_opf = REQ_OP_WRITE;
2868 bio->bi_end_io = end_sync_write;
2869 write_targets ++;
2870 } else {
2871 /* may need to read from here */
2872 sector_t first_bad = MaxSector;
2873 int bad_sectors;
2874
2875 if (is_badblock(rdev, sector_nr, good_sectors,
2876 &first_bad, &bad_sectors)) {
2877 if (first_bad > sector_nr)
2878 good_sectors = first_bad - sector_nr;
2879 else {
2880 bad_sectors -= (sector_nr - first_bad);
2881 if (min_bad == 0 ||
2882 min_bad > bad_sectors)
2883 min_bad = bad_sectors;
2884 }
2885 }
2886 if (sector_nr < first_bad) {
2887 if (test_bit(WriteMostly, &rdev->flags)) {
2888 if (wonly < 0)
2889 wonly = i;
2890 } else {
2891 if (disk < 0)
2892 disk = i;
2893 }
2894 bio->bi_opf = REQ_OP_READ;
2895 bio->bi_end_io = end_sync_read;
2896 read_targets++;
2897 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2898 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2899 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2900 /*
2901 * The device is suitable for reading (InSync),
2902 * but has bad block(s) here. Let's try to correct them,
2903 * if we are doing resync or repair. Otherwise, leave
2904 * this device alone for this sync request.
2905 */
2906 bio->bi_opf = REQ_OP_WRITE;
2907 bio->bi_end_io = end_sync_write;
2908 write_targets++;
2909 }
2910 }
2911 if (rdev && bio->bi_end_io) {
2912 atomic_inc(&rdev->nr_pending);
2913 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2914 bio_set_dev(bio, rdev->bdev);
2915 if (test_bit(FailFast, &rdev->flags))
2916 bio->bi_opf |= MD_FAILFAST;
2917 }
2918 }
2919 if (disk < 0)
2920 disk = wonly;
2921 r1_bio->read_disk = disk;
2922
2923 if (read_targets == 0 && min_bad > 0) {
2924 /* These sectors are bad on all InSync devices, so we
2925 * need to mark them bad on all write targets
2926 */
2927 int ok = 1;
2928 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2929 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2930 struct md_rdev *rdev = conf->mirrors[i].rdev;
2931 ok = rdev_set_badblocks(rdev, sector_nr,
2932 min_bad, 0
2933 ) && ok;
2934 }
2935 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2936 *skipped = 1;
2937 put_buf(r1_bio);
2938
2939 if (!ok) {
2940 /* Cannot record the badblocks, so need to
2941 * abort the resync.
2942 * If there are multiple read targets, could just
2943 * fail the really bad ones ???
2944 */
2945 conf->recovery_disabled = mddev->recovery_disabled;
2946 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2947 return 0;
2948 } else
2949 return min_bad;
2950
2951 }
2952 if (min_bad > 0 && min_bad < good_sectors) {
2953 /* only resync enough to reach the next bad->good
2954 * transition */
2955 good_sectors = min_bad;
2956 }
2957
2958 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2959 /* extra read targets are also write targets */
2960 write_targets += read_targets-1;
2961
2962 if (write_targets == 0 || read_targets == 0) {
2963 /* There is nowhere to write, so all non-sync
2964 * drives must be failed - so we are finished
2965 */
2966 sector_t rv;
2967 if (min_bad > 0)
2968 max_sector = sector_nr + min_bad;
2969 rv = max_sector - sector_nr;
2970 *skipped = 1;
2971 put_buf(r1_bio);
2972 return rv;
2973 }
2974
2975 if (max_sector > mddev->resync_max)
2976 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2977 if (max_sector > sector_nr + good_sectors)
2978 max_sector = sector_nr + good_sectors;
2979 nr_sectors = 0;
2980 sync_blocks = 0;
2981 do {
2982 struct page *page;
2983 int len = PAGE_SIZE;
2984 if (sector_nr + (len>>9) > max_sector)
2985 len = (max_sector - sector_nr) << 9;
2986 if (len == 0)
2987 break;
2988 if (sync_blocks == 0) {
2989 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
2990 &sync_blocks, still_degraded) &&
2991 !conf->fullsync &&
2992 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2993 break;
2994 if ((len >> 9) > sync_blocks)
2995 len = sync_blocks<<9;
2996 }
2997
2998 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2999 struct resync_pages *rp;
3000
3001 bio = r1_bio->bios[i];
3002 rp = get_resync_pages(bio);
3003 if (bio->bi_end_io) {
3004 page = resync_fetch_page(rp, page_idx);
3005
3006 /*
3007 * won't fail because the vec table is big
3008 * enough to hold all these pages
3009 */
3010 __bio_add_page(bio, page, len, 0);
3011 }
3012 }
3013 nr_sectors += len>>9;
3014 sector_nr += len>>9;
3015 sync_blocks -= (len>>9);
3016 } while (++page_idx < RESYNC_PAGES);
3017
3018 r1_bio->sectors = nr_sectors;
3019
3020 if (mddev_is_clustered(mddev) &&
3021 conf->cluster_sync_high < sector_nr + nr_sectors) {
3022 conf->cluster_sync_low = mddev->curr_resync_completed;
3023 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
3024 /* Send resync message */
3025 md_cluster_ops->resync_info_update(mddev,
3026 conf->cluster_sync_low,
3027 conf->cluster_sync_high);
3028 }
3029
3030 /* For a user-requested sync, we read all readable devices and do a
3031 * compare
3032 */
3033 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
3034 atomic_set(&r1_bio->remaining, read_targets);
3035 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
3036 bio = r1_bio->bios[i];
3037 if (bio->bi_end_io == end_sync_read) {
3038 read_targets--;
3039 md_sync_acct_bio(bio, nr_sectors);
3040 if (read_targets == 1)
3041 bio->bi_opf &= ~MD_FAILFAST;
3042 submit_bio_noacct(bio);
3043 }
3044 }
3045 } else {
3046 atomic_set(&r1_bio->remaining, 1);
3047 bio = r1_bio->bios[r1_bio->read_disk];
3048 md_sync_acct_bio(bio, nr_sectors);
3049 if (read_targets == 1)
3050 bio->bi_opf &= ~MD_FAILFAST;
3051 submit_bio_noacct(bio);
3052 }
3053 return nr_sectors;
3054}
3055
3056static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3057{
3058 if (sectors)
3059 return sectors;
3060
3061 return mddev->dev_sectors;
3062}
3063
3064static struct r1conf *setup_conf(struct mddev *mddev)
3065{
3066 struct r1conf *conf;
3067 int i;
3068 struct raid1_info *disk;
3069 struct md_rdev *rdev;
3070 int err = -ENOMEM;
3071
3072 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
3073 if (!conf)
3074 goto abort;
3075
3076 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
3077 sizeof(atomic_t), GFP_KERNEL);
3078 if (!conf->nr_pending)
3079 goto abort;
3080
3081 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
3082 sizeof(atomic_t), GFP_KERNEL);
3083 if (!conf->nr_waiting)
3084 goto abort;
3085
3086 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
3087 sizeof(atomic_t), GFP_KERNEL);
3088 if (!conf->nr_queued)
3089 goto abort;
3090
3091 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
3092 sizeof(atomic_t), GFP_KERNEL);
3093 if (!conf->barrier)
3094 goto abort;
3095
3096 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3097 mddev->raid_disks, 2),
3098 GFP_KERNEL);
3099 if (!conf->mirrors)
3100 goto abort;
3101
3102 conf->tmppage = alloc_page(GFP_KERNEL);
3103 if (!conf->tmppage)
3104 goto abort;
3105
3106 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
3107 if (!conf->poolinfo)
3108 goto abort;
3109 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
3110 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
3111 rbio_pool_free, conf->poolinfo);
3112 if (err)
3113 goto abort;
3114
3115 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3116 if (err)
3117 goto abort;
3118
3119 conf->poolinfo->mddev = mddev;
3120
3121 err = -EINVAL;
3122 spin_lock_init(&conf->device_lock);
3123 conf->raid_disks = mddev->raid_disks;
3124 rdev_for_each(rdev, mddev) {
3125 int disk_idx = rdev->raid_disk;
3126
3127 if (disk_idx >= conf->raid_disks || disk_idx < 0)
3128 continue;
3129
3130 if (!raid1_add_conf(conf, rdev, disk_idx,
3131 test_bit(Replacement, &rdev->flags)))
3132 goto abort;
3133 }
3134 conf->mddev = mddev;
3135 INIT_LIST_HEAD(&conf->retry_list);
3136 INIT_LIST_HEAD(&conf->bio_end_io_list);
3137
3138 spin_lock_init(&conf->resync_lock);
3139 init_waitqueue_head(&conf->wait_barrier);
3140
3141 bio_list_init(&conf->pending_bio_list);
3142 conf->recovery_disabled = mddev->recovery_disabled - 1;
3143
3144 err = -EIO;
3145 for (i = 0; i < conf->raid_disks * 2; i++) {
3146
3147 disk = conf->mirrors + i;
3148
3149 if (i < conf->raid_disks &&
3150 disk[conf->raid_disks].rdev) {
3151 /* This slot has a replacement. */
3152 if (!disk->rdev) {
3153 /* No original, just make the replacement
3154 * a recovering spare
3155 */
3156 disk->rdev =
3157 disk[conf->raid_disks].rdev;
3158 disk[conf->raid_disks].rdev = NULL;
3159 } else if (!test_bit(In_sync, &disk->rdev->flags))
3160 /* Original is not in_sync - bad */
3161 goto abort;
3162 }
3163
3164 if (!disk->rdev ||
3165 !test_bit(In_sync, &disk->rdev->flags)) {
3166 disk->head_position = 0;
3167 if (disk->rdev &&
3168 (disk->rdev->saved_raid_disk < 0))
3169 conf->fullsync = 1;
3170 }
3171 }
3172
3173 err = -ENOMEM;
3174 rcu_assign_pointer(conf->thread,
3175 md_register_thread(raid1d, mddev, "raid1"));
3176 if (!conf->thread)
3177 goto abort;
3178
3179 return conf;
3180
3181 abort:
3182 if (conf) {
3183 mempool_exit(&conf->r1bio_pool);
3184 kfree(conf->mirrors);
3185 safe_put_page(conf->tmppage);
3186 kfree(conf->poolinfo);
3187 kfree(conf->nr_pending);
3188 kfree(conf->nr_waiting);
3189 kfree(conf->nr_queued);
3190 kfree(conf->barrier);
3191 bioset_exit(&conf->bio_split);
3192 kfree(conf);
3193 }
3194 return ERR_PTR(err);
3195}
3196
3197static int raid1_set_limits(struct mddev *mddev)
3198{
3199 struct queue_limits lim;
3200
3201 blk_set_stacking_limits(&lim);
3202 lim.max_write_zeroes_sectors = 0;
3203 mddev_stack_rdev_limits(mddev, &lim);
3204 return queue_limits_set(mddev->gendisk->queue, &lim);
3205}
3206
3207static void raid1_free(struct mddev *mddev, void *priv);
3208static int raid1_run(struct mddev *mddev)
3209{
3210 struct r1conf *conf;
3211 int i;
3212 int ret;
3213
3214 if (mddev->level != 1) {
3215 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3216 mdname(mddev), mddev->level);
3217 return -EIO;
3218 }
3219 if (mddev->reshape_position != MaxSector) {
3220 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3221 mdname(mddev));
3222 return -EIO;
3223 }
3224
3225 /*
3226 * copy the already verified devices into our private RAID1
3227 * bookkeeping area. [whatever we allocate in run(),
3228 * should be freed in raid1_free()]
3229 */
3230 if (mddev->private == NULL)
3231 conf = setup_conf(mddev);
3232 else
3233 conf = mddev->private;
3234
3235 if (IS_ERR(conf))
3236 return PTR_ERR(conf);
3237
3238 if (!mddev_is_dm(mddev)) {
3239 ret = raid1_set_limits(mddev);
3240 if (ret)
3241 goto abort;
3242 }
3243
3244 mddev->degraded = 0;
3245 for (i = 0; i < conf->raid_disks; i++)
3246 if (conf->mirrors[i].rdev == NULL ||
3247 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3248 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3249 mddev->degraded++;
3250 /*
3251 * RAID1 needs at least one disk in active
3252 */
3253 if (conf->raid_disks - mddev->degraded < 1) {
3254 md_unregister_thread(mddev, &conf->thread);
3255 ret = -EINVAL;
3256 goto abort;
3257 }
3258
3259 if (conf->raid_disks - mddev->degraded == 1)
3260 mddev->recovery_cp = MaxSector;
3261
3262 if (mddev->recovery_cp != MaxSector)
3263 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3264 mdname(mddev));
3265 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3266 mdname(mddev), mddev->raid_disks - mddev->degraded,
3267 mddev->raid_disks);
3268
3269 /*
3270 * Ok, everything is just fine now
3271 */
3272 rcu_assign_pointer(mddev->thread, conf->thread);
3273 rcu_assign_pointer(conf->thread, NULL);
3274 mddev->private = conf;
3275 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3276
3277 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3278
3279 ret = md_integrity_register(mddev);
3280 if (ret) {
3281 md_unregister_thread(mddev, &mddev->thread);
3282 goto abort;
3283 }
3284 return 0;
3285
3286abort:
3287 raid1_free(mddev, conf);
3288 return ret;
3289}
3290
3291static void raid1_free(struct mddev *mddev, void *priv)
3292{
3293 struct r1conf *conf = priv;
3294
3295 mempool_exit(&conf->r1bio_pool);
3296 kfree(conf->mirrors);
3297 safe_put_page(conf->tmppage);
3298 kfree(conf->poolinfo);
3299 kfree(conf->nr_pending);
3300 kfree(conf->nr_waiting);
3301 kfree(conf->nr_queued);
3302 kfree(conf->barrier);
3303 bioset_exit(&conf->bio_split);
3304 kfree(conf);
3305}
3306
3307static int raid1_resize(struct mddev *mddev, sector_t sectors)
3308{
3309 /* no resync is happening, and there is enough space
3310 * on all devices, so we can resize.
3311 * We need to make sure resync covers any new space.
3312 * If the array is shrinking we should possibly wait until
3313 * any io in the removed space completes, but it hardly seems
3314 * worth it.
3315 */
3316 sector_t newsize = raid1_size(mddev, sectors, 0);
3317 if (mddev->external_size &&
3318 mddev->array_sectors > newsize)
3319 return -EINVAL;
3320 if (mddev->bitmap) {
3321 int ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
3322 if (ret)
3323 return ret;
3324 }
3325 md_set_array_sectors(mddev, newsize);
3326 if (sectors > mddev->dev_sectors &&
3327 mddev->recovery_cp > mddev->dev_sectors) {
3328 mddev->recovery_cp = mddev->dev_sectors;
3329 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3330 }
3331 mddev->dev_sectors = sectors;
3332 mddev->resync_max_sectors = sectors;
3333 return 0;
3334}
3335
3336static int raid1_reshape(struct mddev *mddev)
3337{
3338 /* We need to:
3339 * 1/ resize the r1bio_pool
3340 * 2/ resize conf->mirrors
3341 *
3342 * We allocate a new r1bio_pool if we can.
3343 * Then raise a device barrier and wait until all IO stops.
3344 * Then resize conf->mirrors and swap in the new r1bio pool.
3345 *
3346 * At the same time, we "pack" the devices so that all the missing
3347 * devices have the higher raid_disk numbers.
3348 */
3349 mempool_t newpool, oldpool;
3350 struct pool_info *newpoolinfo;
3351 struct raid1_info *newmirrors;
3352 struct r1conf *conf = mddev->private;
3353 int cnt, raid_disks;
3354 unsigned long flags;
3355 int d, d2;
3356 int ret;
3357
3358 memset(&newpool, 0, sizeof(newpool));
3359 memset(&oldpool, 0, sizeof(oldpool));
3360
3361 /* Cannot change chunk_size, layout, or level */
3362 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3363 mddev->layout != mddev->new_layout ||
3364 mddev->level != mddev->new_level) {
3365 mddev->new_chunk_sectors = mddev->chunk_sectors;
3366 mddev->new_layout = mddev->layout;
3367 mddev->new_level = mddev->level;
3368 return -EINVAL;
3369 }
3370
3371 if (!mddev_is_clustered(mddev))
3372 md_allow_write(mddev);
3373
3374 raid_disks = mddev->raid_disks + mddev->delta_disks;
3375
3376 if (raid_disks < conf->raid_disks) {
3377 cnt=0;
3378 for (d= 0; d < conf->raid_disks; d++)
3379 if (conf->mirrors[d].rdev)
3380 cnt++;
3381 if (cnt > raid_disks)
3382 return -EBUSY;
3383 }
3384
3385 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3386 if (!newpoolinfo)
3387 return -ENOMEM;
3388 newpoolinfo->mddev = mddev;
3389 newpoolinfo->raid_disks = raid_disks * 2;
3390
3391 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3392 rbio_pool_free, newpoolinfo);
3393 if (ret) {
3394 kfree(newpoolinfo);
3395 return ret;
3396 }
3397 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3398 raid_disks, 2),
3399 GFP_KERNEL);
3400 if (!newmirrors) {
3401 kfree(newpoolinfo);
3402 mempool_exit(&newpool);
3403 return -ENOMEM;
3404 }
3405
3406 freeze_array(conf, 0);
3407
3408 /* ok, everything is stopped */
3409 oldpool = conf->r1bio_pool;
3410 conf->r1bio_pool = newpool;
3411
3412 for (d = d2 = 0; d < conf->raid_disks; d++) {
3413 struct md_rdev *rdev = conf->mirrors[d].rdev;
3414 if (rdev && rdev->raid_disk != d2) {
3415 sysfs_unlink_rdev(mddev, rdev);
3416 rdev->raid_disk = d2;
3417 sysfs_unlink_rdev(mddev, rdev);
3418 if (sysfs_link_rdev(mddev, rdev))
3419 pr_warn("md/raid1:%s: cannot register rd%d\n",
3420 mdname(mddev), rdev->raid_disk);
3421 }
3422 if (rdev)
3423 newmirrors[d2++].rdev = rdev;
3424 }
3425 kfree(conf->mirrors);
3426 conf->mirrors = newmirrors;
3427 kfree(conf->poolinfo);
3428 conf->poolinfo = newpoolinfo;
3429
3430 spin_lock_irqsave(&conf->device_lock, flags);
3431 mddev->degraded += (raid_disks - conf->raid_disks);
3432 spin_unlock_irqrestore(&conf->device_lock, flags);
3433 conf->raid_disks = mddev->raid_disks = raid_disks;
3434 mddev->delta_disks = 0;
3435
3436 unfreeze_array(conf);
3437
3438 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3439 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3440 md_wakeup_thread(mddev->thread);
3441
3442 mempool_exit(&oldpool);
3443 return 0;
3444}
3445
3446static void raid1_quiesce(struct mddev *mddev, int quiesce)
3447{
3448 struct r1conf *conf = mddev->private;
3449
3450 if (quiesce)
3451 freeze_array(conf, 0);
3452 else
3453 unfreeze_array(conf);
3454}
3455
3456static void *raid1_takeover(struct mddev *mddev)
3457{
3458 /* raid1 can take over:
3459 * raid5 with 2 devices, any layout or chunk size
3460 */
3461 if (mddev->level == 5 && mddev->raid_disks == 2) {
3462 struct r1conf *conf;
3463 mddev->new_level = 1;
3464 mddev->new_layout = 0;
3465 mddev->new_chunk_sectors = 0;
3466 conf = setup_conf(mddev);
3467 if (!IS_ERR(conf)) {
3468 /* Array must appear to be quiesced */
3469 conf->array_frozen = 1;
3470 mddev_clear_unsupported_flags(mddev,
3471 UNSUPPORTED_MDDEV_FLAGS);
3472 }
3473 return conf;
3474 }
3475 return ERR_PTR(-EINVAL);
3476}
3477
3478static struct md_personality raid1_personality =
3479{
3480 .name = "raid1",
3481 .level = 1,
3482 .owner = THIS_MODULE,
3483 .make_request = raid1_make_request,
3484 .run = raid1_run,
3485 .free = raid1_free,
3486 .status = raid1_status,
3487 .error_handler = raid1_error,
3488 .hot_add_disk = raid1_add_disk,
3489 .hot_remove_disk= raid1_remove_disk,
3490 .spare_active = raid1_spare_active,
3491 .sync_request = raid1_sync_request,
3492 .resize = raid1_resize,
3493 .size = raid1_size,
3494 .check_reshape = raid1_reshape,
3495 .quiesce = raid1_quiesce,
3496 .takeover = raid1_takeover,
3497};
3498
3499static int __init raid_init(void)
3500{
3501 return register_md_personality(&raid1_personality);
3502}
3503
3504static void raid_exit(void)
3505{
3506 unregister_md_personality(&raid1_personality);
3507}
3508
3509module_init(raid_init);
3510module_exit(raid_exit);
3511MODULE_LICENSE("GPL");
3512MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3513MODULE_ALIAS("md-personality-3"); /* RAID1 */
3514MODULE_ALIAS("md-raid1");
3515MODULE_ALIAS("md-level-1");