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 struct mddev *mddev = r1_bio->mddev;
415
416 /* it really is the end of this request */
417 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
418 bio_free_pages(r1_bio->behind_master_bio);
419 bio_put(r1_bio->behind_master_bio);
420 r1_bio->behind_master_bio = NULL;
421 }
422
423 if (test_bit(R1BIO_BehindIO, &r1_bio->state))
424 mddev->bitmap_ops->end_behind_write(mddev);
425 md_write_end(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 /* Finished with this branch */
482 r1_bio->bios[mirror] = NULL;
483 to_put = bio;
484 }
485 } else {
486 /*
487 * Set R1BIO_Uptodate in our master bio, so that we
488 * will return a good error code for to the higher
489 * levels even if IO on some other mirrored buffer
490 * fails.
491 *
492 * The 'master' represents the composite IO operation
493 * to user-side. So if something waits for IO, then it
494 * will wait for the 'master' bio.
495 */
496 r1_bio->bios[mirror] = NULL;
497 to_put = bio;
498 /*
499 * Do not set R1BIO_Uptodate if the current device is
500 * rebuilding or Faulty. This is because we cannot use
501 * such device for properly reading the data back (we could
502 * potentially use it, if the current write would have felt
503 * before rdev->recovery_offset, but for simplicity we don't
504 * check this here.
505 */
506 if (test_bit(In_sync, &rdev->flags) &&
507 !test_bit(Faulty, &rdev->flags))
508 set_bit(R1BIO_Uptodate, &r1_bio->state);
509
510 /* Maybe we can clear some bad blocks. */
511 if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors) &&
512 !discard_error) {
513 r1_bio->bios[mirror] = IO_MADE_GOOD;
514 set_bit(R1BIO_MadeGood, &r1_bio->state);
515 }
516 }
517
518 if (behind) {
519 if (test_bit(CollisionCheck, &rdev->flags))
520 remove_serial(rdev, lo, hi);
521 if (test_bit(WriteMostly, &rdev->flags))
522 atomic_dec(&r1_bio->behind_remaining);
523
524 /*
525 * In behind mode, we ACK the master bio once the I/O
526 * has safely reached all non-writemostly
527 * disks. Setting the Returned bit ensures that this
528 * gets done only once -- we don't ever want to return
529 * -EIO here, instead we'll wait
530 */
531 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
532 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
533 /* Maybe we can return now */
534 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
535 struct bio *mbio = r1_bio->master_bio;
536 pr_debug("raid1: behind end write sectors"
537 " %llu-%llu\n",
538 (unsigned long long) mbio->bi_iter.bi_sector,
539 (unsigned long long) bio_end_sector(mbio) - 1);
540 call_bio_endio(r1_bio);
541 }
542 }
543 } else if (rdev->mddev->serialize_policy)
544 remove_serial(rdev, lo, hi);
545 if (r1_bio->bios[mirror] == NULL)
546 rdev_dec_pending(rdev, conf->mddev);
547
548 /*
549 * Let's see if all mirrored write operations have finished
550 * already.
551 */
552 r1_bio_write_done(r1_bio);
553
554 if (to_put)
555 bio_put(to_put);
556}
557
558static sector_t align_to_barrier_unit_end(sector_t start_sector,
559 sector_t sectors)
560{
561 sector_t len;
562
563 WARN_ON(sectors == 0);
564 /*
565 * len is the number of sectors from start_sector to end of the
566 * barrier unit which start_sector belongs to.
567 */
568 len = round_up(start_sector + 1, BARRIER_UNIT_SECTOR_SIZE) -
569 start_sector;
570
571 if (len > sectors)
572 len = sectors;
573
574 return len;
575}
576
577static void update_read_sectors(struct r1conf *conf, int disk,
578 sector_t this_sector, int len)
579{
580 struct raid1_info *info = &conf->mirrors[disk];
581
582 atomic_inc(&info->rdev->nr_pending);
583 if (info->next_seq_sect != this_sector)
584 info->seq_start = this_sector;
585 info->next_seq_sect = this_sector + len;
586}
587
588static int choose_first_rdev(struct r1conf *conf, struct r1bio *r1_bio,
589 int *max_sectors)
590{
591 sector_t this_sector = r1_bio->sector;
592 int len = r1_bio->sectors;
593 int disk;
594
595 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
596 struct md_rdev *rdev;
597 int read_len;
598
599 if (r1_bio->bios[disk] == IO_BLOCKED)
600 continue;
601
602 rdev = conf->mirrors[disk].rdev;
603 if (!rdev || test_bit(Faulty, &rdev->flags))
604 continue;
605
606 /* choose the first disk even if it has some bad blocks. */
607 read_len = raid1_check_read_range(rdev, this_sector, &len);
608 if (read_len > 0) {
609 update_read_sectors(conf, disk, this_sector, read_len);
610 *max_sectors = read_len;
611 return disk;
612 }
613 }
614
615 return -1;
616}
617
618static bool rdev_in_recovery(struct md_rdev *rdev, struct r1bio *r1_bio)
619{
620 return !test_bit(In_sync, &rdev->flags) &&
621 rdev->recovery_offset < r1_bio->sector + r1_bio->sectors;
622}
623
624static int choose_bb_rdev(struct r1conf *conf, struct r1bio *r1_bio,
625 int *max_sectors)
626{
627 sector_t this_sector = r1_bio->sector;
628 int best_disk = -1;
629 int best_len = 0;
630 int disk;
631
632 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
633 struct md_rdev *rdev;
634 int len;
635 int read_len;
636
637 if (r1_bio->bios[disk] == IO_BLOCKED)
638 continue;
639
640 rdev = conf->mirrors[disk].rdev;
641 if (!rdev || test_bit(Faulty, &rdev->flags) ||
642 rdev_in_recovery(rdev, r1_bio) ||
643 test_bit(WriteMostly, &rdev->flags))
644 continue;
645
646 /* keep track of the disk with the most readable sectors. */
647 len = r1_bio->sectors;
648 read_len = raid1_check_read_range(rdev, this_sector, &len);
649 if (read_len > best_len) {
650 best_disk = disk;
651 best_len = read_len;
652 }
653 }
654
655 if (best_disk != -1) {
656 *max_sectors = best_len;
657 update_read_sectors(conf, best_disk, this_sector, best_len);
658 }
659
660 return best_disk;
661}
662
663static int choose_slow_rdev(struct r1conf *conf, struct r1bio *r1_bio,
664 int *max_sectors)
665{
666 sector_t this_sector = r1_bio->sector;
667 int bb_disk = -1;
668 int bb_read_len = 0;
669 int disk;
670
671 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
672 struct md_rdev *rdev;
673 int len;
674 int read_len;
675
676 if (r1_bio->bios[disk] == IO_BLOCKED)
677 continue;
678
679 rdev = conf->mirrors[disk].rdev;
680 if (!rdev || test_bit(Faulty, &rdev->flags) ||
681 !test_bit(WriteMostly, &rdev->flags) ||
682 rdev_in_recovery(rdev, r1_bio))
683 continue;
684
685 /* there are no bad blocks, we can use this disk */
686 len = r1_bio->sectors;
687 read_len = raid1_check_read_range(rdev, this_sector, &len);
688 if (read_len == r1_bio->sectors) {
689 *max_sectors = read_len;
690 update_read_sectors(conf, disk, this_sector, read_len);
691 return disk;
692 }
693
694 /*
695 * there are partial bad blocks, choose the rdev with largest
696 * read length.
697 */
698 if (read_len > bb_read_len) {
699 bb_disk = disk;
700 bb_read_len = read_len;
701 }
702 }
703
704 if (bb_disk != -1) {
705 *max_sectors = bb_read_len;
706 update_read_sectors(conf, bb_disk, this_sector, bb_read_len);
707 }
708
709 return bb_disk;
710}
711
712static bool is_sequential(struct r1conf *conf, int disk, struct r1bio *r1_bio)
713{
714 /* TODO: address issues with this check and concurrency. */
715 return conf->mirrors[disk].next_seq_sect == r1_bio->sector ||
716 conf->mirrors[disk].head_position == r1_bio->sector;
717}
718
719/*
720 * If buffered sequential IO size exceeds optimal iosize, check if there is idle
721 * disk. If yes, choose the idle disk.
722 */
723static bool should_choose_next(struct r1conf *conf, int disk)
724{
725 struct raid1_info *mirror = &conf->mirrors[disk];
726 int opt_iosize;
727
728 if (!test_bit(Nonrot, &mirror->rdev->flags))
729 return false;
730
731 opt_iosize = bdev_io_opt(mirror->rdev->bdev) >> 9;
732 return opt_iosize > 0 && mirror->seq_start != MaxSector &&
733 mirror->next_seq_sect > opt_iosize &&
734 mirror->next_seq_sect - opt_iosize >= mirror->seq_start;
735}
736
737static bool rdev_readable(struct md_rdev *rdev, struct r1bio *r1_bio)
738{
739 if (!rdev || test_bit(Faulty, &rdev->flags))
740 return false;
741
742 if (rdev_in_recovery(rdev, r1_bio))
743 return false;
744
745 /* don't read from slow disk unless have to */
746 if (test_bit(WriteMostly, &rdev->flags))
747 return false;
748
749 /* don't split IO for bad blocks unless have to */
750 if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors))
751 return false;
752
753 return true;
754}
755
756struct read_balance_ctl {
757 sector_t closest_dist;
758 int closest_dist_disk;
759 int min_pending;
760 int min_pending_disk;
761 int sequential_disk;
762 int readable_disks;
763};
764
765static int choose_best_rdev(struct r1conf *conf, struct r1bio *r1_bio)
766{
767 int disk;
768 struct read_balance_ctl ctl = {
769 .closest_dist_disk = -1,
770 .closest_dist = MaxSector,
771 .min_pending_disk = -1,
772 .min_pending = UINT_MAX,
773 .sequential_disk = -1,
774 };
775
776 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
777 struct md_rdev *rdev;
778 sector_t dist;
779 unsigned int pending;
780
781 if (r1_bio->bios[disk] == IO_BLOCKED)
782 continue;
783
784 rdev = conf->mirrors[disk].rdev;
785 if (!rdev_readable(rdev, r1_bio))
786 continue;
787
788 /* At least two disks to choose from so failfast is OK */
789 if (ctl.readable_disks++ == 1)
790 set_bit(R1BIO_FailFast, &r1_bio->state);
791
792 pending = atomic_read(&rdev->nr_pending);
793 dist = abs(r1_bio->sector - conf->mirrors[disk].head_position);
794
795 /* Don't change to another disk for sequential reads */
796 if (is_sequential(conf, disk, r1_bio)) {
797 if (!should_choose_next(conf, disk))
798 return disk;
799
800 /*
801 * Add 'pending' to avoid choosing this disk if
802 * there is other idle disk.
803 */
804 pending++;
805 /*
806 * If there is no other idle disk, this disk
807 * will be chosen.
808 */
809 ctl.sequential_disk = disk;
810 }
811
812 if (ctl.min_pending > pending) {
813 ctl.min_pending = pending;
814 ctl.min_pending_disk = disk;
815 }
816
817 if (ctl.closest_dist > dist) {
818 ctl.closest_dist = dist;
819 ctl.closest_dist_disk = disk;
820 }
821 }
822
823 /*
824 * sequential IO size exceeds optimal iosize, however, there is no other
825 * idle disk, so choose the sequential disk.
826 */
827 if (ctl.sequential_disk != -1 && ctl.min_pending != 0)
828 return ctl.sequential_disk;
829
830 /*
831 * If all disks are rotational, choose the closest disk. If any disk is
832 * non-rotational, choose the disk with less pending request even the
833 * disk is rotational, which might/might not be optimal for raids with
834 * mixed ratation/non-rotational disks depending on workload.
835 */
836 if (ctl.min_pending_disk != -1 &&
837 (READ_ONCE(conf->nonrot_disks) || ctl.min_pending == 0))
838 return ctl.min_pending_disk;
839 else
840 return ctl.closest_dist_disk;
841}
842
843/*
844 * This routine returns the disk from which the requested read should be done.
845 *
846 * 1) If resync is in progress, find the first usable disk and use it even if it
847 * has some bad blocks.
848 *
849 * 2) Now that there is no resync, loop through all disks and skipping slow
850 * disks and disks with bad blocks for now. Only pay attention to key disk
851 * choice.
852 *
853 * 3) If we've made it this far, now look for disks with bad blocks and choose
854 * the one with most number of sectors.
855 *
856 * 4) If we are all the way at the end, we have no choice but to use a disk even
857 * if it is write mostly.
858 *
859 * The rdev for the device selected will have nr_pending incremented.
860 */
861static int read_balance(struct r1conf *conf, struct r1bio *r1_bio,
862 int *max_sectors)
863{
864 int disk;
865
866 clear_bit(R1BIO_FailFast, &r1_bio->state);
867
868 if (raid1_should_read_first(conf->mddev, r1_bio->sector,
869 r1_bio->sectors))
870 return choose_first_rdev(conf, r1_bio, max_sectors);
871
872 disk = choose_best_rdev(conf, r1_bio);
873 if (disk >= 0) {
874 *max_sectors = r1_bio->sectors;
875 update_read_sectors(conf, disk, r1_bio->sector,
876 r1_bio->sectors);
877 return disk;
878 }
879
880 /*
881 * If we are here it means we didn't find a perfectly good disk so
882 * now spend a bit more time trying to find one with the most good
883 * sectors.
884 */
885 disk = choose_bb_rdev(conf, r1_bio, max_sectors);
886 if (disk >= 0)
887 return disk;
888
889 return choose_slow_rdev(conf, r1_bio, max_sectors);
890}
891
892static void wake_up_barrier(struct r1conf *conf)
893{
894 if (wq_has_sleeper(&conf->wait_barrier))
895 wake_up(&conf->wait_barrier);
896}
897
898static void flush_bio_list(struct r1conf *conf, struct bio *bio)
899{
900 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
901 raid1_prepare_flush_writes(conf->mddev);
902 wake_up_barrier(conf);
903
904 while (bio) { /* submit pending writes */
905 struct bio *next = bio->bi_next;
906
907 raid1_submit_write(bio);
908 bio = next;
909 cond_resched();
910 }
911}
912
913static void flush_pending_writes(struct r1conf *conf)
914{
915 /* Any writes that have been queued but are awaiting
916 * bitmap updates get flushed here.
917 */
918 spin_lock_irq(&conf->device_lock);
919
920 if (conf->pending_bio_list.head) {
921 struct blk_plug plug;
922 struct bio *bio;
923
924 bio = bio_list_get(&conf->pending_bio_list);
925 spin_unlock_irq(&conf->device_lock);
926
927 /*
928 * As this is called in a wait_event() loop (see freeze_array),
929 * current->state might be TASK_UNINTERRUPTIBLE which will
930 * cause a warning when we prepare to wait again. As it is
931 * rare that this path is taken, it is perfectly safe to force
932 * us to go around the wait_event() loop again, so the warning
933 * is a false-positive. Silence the warning by resetting
934 * thread state
935 */
936 __set_current_state(TASK_RUNNING);
937 blk_start_plug(&plug);
938 flush_bio_list(conf, bio);
939 blk_finish_plug(&plug);
940 } else
941 spin_unlock_irq(&conf->device_lock);
942}
943
944/* Barriers....
945 * Sometimes we need to suspend IO while we do something else,
946 * either some resync/recovery, or reconfigure the array.
947 * To do this we raise a 'barrier'.
948 * The 'barrier' is a counter that can be raised multiple times
949 * to count how many activities are happening which preclude
950 * normal IO.
951 * We can only raise the barrier if there is no pending IO.
952 * i.e. if nr_pending == 0.
953 * We choose only to raise the barrier if no-one is waiting for the
954 * barrier to go down. This means that as soon as an IO request
955 * is ready, no other operations which require a barrier will start
956 * until the IO request has had a chance.
957 *
958 * So: regular IO calls 'wait_barrier'. When that returns there
959 * is no backgroup IO happening, It must arrange to call
960 * allow_barrier when it has finished its IO.
961 * backgroup IO calls must call raise_barrier. Once that returns
962 * there is no normal IO happeing. It must arrange to call
963 * lower_barrier when the particular background IO completes.
964 *
965 * If resync/recovery is interrupted, returns -EINTR;
966 * Otherwise, returns 0.
967 */
968static int raise_barrier(struct r1conf *conf, sector_t sector_nr)
969{
970 int idx = sector_to_idx(sector_nr);
971
972 spin_lock_irq(&conf->resync_lock);
973
974 /* Wait until no block IO is waiting */
975 wait_event_lock_irq(conf->wait_barrier,
976 !atomic_read(&conf->nr_waiting[idx]),
977 conf->resync_lock);
978
979 /* block any new IO from starting */
980 atomic_inc(&conf->barrier[idx]);
981 /*
982 * In raise_barrier() we firstly increase conf->barrier[idx] then
983 * check conf->nr_pending[idx]. In _wait_barrier() we firstly
984 * increase conf->nr_pending[idx] then check conf->barrier[idx].
985 * A memory barrier here to make sure conf->nr_pending[idx] won't
986 * be fetched before conf->barrier[idx] is increased. Otherwise
987 * there will be a race between raise_barrier() and _wait_barrier().
988 */
989 smp_mb__after_atomic();
990
991 /* For these conditions we must wait:
992 * A: while the array is in frozen state
993 * B: while conf->nr_pending[idx] is not 0, meaning regular I/O
994 * existing in corresponding I/O barrier bucket.
995 * C: while conf->barrier[idx] >= RESYNC_DEPTH, meaning reaches
996 * max resync count which allowed on current I/O barrier bucket.
997 */
998 wait_event_lock_irq(conf->wait_barrier,
999 (!conf->array_frozen &&
1000 !atomic_read(&conf->nr_pending[idx]) &&
1001 atomic_read(&conf->barrier[idx]) < RESYNC_DEPTH) ||
1002 test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery),
1003 conf->resync_lock);
1004
1005 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
1006 atomic_dec(&conf->barrier[idx]);
1007 spin_unlock_irq(&conf->resync_lock);
1008 wake_up(&conf->wait_barrier);
1009 return -EINTR;
1010 }
1011
1012 atomic_inc(&conf->nr_sync_pending);
1013 spin_unlock_irq(&conf->resync_lock);
1014
1015 return 0;
1016}
1017
1018static void lower_barrier(struct r1conf *conf, sector_t sector_nr)
1019{
1020 int idx = sector_to_idx(sector_nr);
1021
1022 BUG_ON(atomic_read(&conf->barrier[idx]) <= 0);
1023
1024 atomic_dec(&conf->barrier[idx]);
1025 atomic_dec(&conf->nr_sync_pending);
1026 wake_up(&conf->wait_barrier);
1027}
1028
1029static bool _wait_barrier(struct r1conf *conf, int idx, bool nowait)
1030{
1031 bool ret = true;
1032
1033 /*
1034 * We need to increase conf->nr_pending[idx] very early here,
1035 * then raise_barrier() can be blocked when it waits for
1036 * conf->nr_pending[idx] to be 0. Then we can avoid holding
1037 * conf->resync_lock when there is no barrier raised in same
1038 * barrier unit bucket. Also if the array is frozen, I/O
1039 * should be blocked until array is unfrozen.
1040 */
1041 atomic_inc(&conf->nr_pending[idx]);
1042 /*
1043 * In _wait_barrier() we firstly increase conf->nr_pending[idx], then
1044 * check conf->barrier[idx]. In raise_barrier() we firstly increase
1045 * conf->barrier[idx], then check conf->nr_pending[idx]. A memory
1046 * barrier is necessary here to make sure conf->barrier[idx] won't be
1047 * fetched before conf->nr_pending[idx] is increased. Otherwise there
1048 * will be a race between _wait_barrier() and raise_barrier().
1049 */
1050 smp_mb__after_atomic();
1051
1052 /*
1053 * Don't worry about checking two atomic_t variables at same time
1054 * here. If during we check conf->barrier[idx], the array is
1055 * frozen (conf->array_frozen is 1), and chonf->barrier[idx] is
1056 * 0, it is safe to return and make the I/O continue. Because the
1057 * array is frozen, all I/O returned here will eventually complete
1058 * or be queued, no race will happen. See code comment in
1059 * frozen_array().
1060 */
1061 if (!READ_ONCE(conf->array_frozen) &&
1062 !atomic_read(&conf->barrier[idx]))
1063 return ret;
1064
1065 /*
1066 * After holding conf->resync_lock, conf->nr_pending[idx]
1067 * should be decreased before waiting for barrier to drop.
1068 * Otherwise, we may encounter a race condition because
1069 * raise_barrer() might be waiting for conf->nr_pending[idx]
1070 * to be 0 at same time.
1071 */
1072 spin_lock_irq(&conf->resync_lock);
1073 atomic_inc(&conf->nr_waiting[idx]);
1074 atomic_dec(&conf->nr_pending[idx]);
1075 /*
1076 * In case freeze_array() is waiting for
1077 * get_unqueued_pending() == extra
1078 */
1079 wake_up_barrier(conf);
1080 /* Wait for the barrier in same barrier unit bucket to drop. */
1081
1082 /* Return false when nowait flag is set */
1083 if (nowait) {
1084 ret = false;
1085 } else {
1086 wait_event_lock_irq(conf->wait_barrier,
1087 !conf->array_frozen &&
1088 !atomic_read(&conf->barrier[idx]),
1089 conf->resync_lock);
1090 atomic_inc(&conf->nr_pending[idx]);
1091 }
1092
1093 atomic_dec(&conf->nr_waiting[idx]);
1094 spin_unlock_irq(&conf->resync_lock);
1095 return ret;
1096}
1097
1098static bool wait_read_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1099{
1100 int idx = sector_to_idx(sector_nr);
1101 bool ret = true;
1102
1103 /*
1104 * Very similar to _wait_barrier(). The difference is, for read
1105 * I/O we don't need wait for sync I/O, but if the whole array
1106 * is frozen, the read I/O still has to wait until the array is
1107 * unfrozen. Since there is no ordering requirement with
1108 * conf->barrier[idx] here, memory barrier is unnecessary as well.
1109 */
1110 atomic_inc(&conf->nr_pending[idx]);
1111
1112 if (!READ_ONCE(conf->array_frozen))
1113 return ret;
1114
1115 spin_lock_irq(&conf->resync_lock);
1116 atomic_inc(&conf->nr_waiting[idx]);
1117 atomic_dec(&conf->nr_pending[idx]);
1118 /*
1119 * In case freeze_array() is waiting for
1120 * get_unqueued_pending() == extra
1121 */
1122 wake_up_barrier(conf);
1123 /* Wait for array to be unfrozen */
1124
1125 /* Return false when nowait flag is set */
1126 if (nowait) {
1127 /* Return false when nowait flag is set */
1128 ret = false;
1129 } else {
1130 wait_event_lock_irq(conf->wait_barrier,
1131 !conf->array_frozen,
1132 conf->resync_lock);
1133 atomic_inc(&conf->nr_pending[idx]);
1134 }
1135
1136 atomic_dec(&conf->nr_waiting[idx]);
1137 spin_unlock_irq(&conf->resync_lock);
1138 return ret;
1139}
1140
1141static bool wait_barrier(struct r1conf *conf, sector_t sector_nr, bool nowait)
1142{
1143 int idx = sector_to_idx(sector_nr);
1144
1145 return _wait_barrier(conf, idx, nowait);
1146}
1147
1148static void _allow_barrier(struct r1conf *conf, int idx)
1149{
1150 atomic_dec(&conf->nr_pending[idx]);
1151 wake_up_barrier(conf);
1152}
1153
1154static void allow_barrier(struct r1conf *conf, sector_t sector_nr)
1155{
1156 int idx = sector_to_idx(sector_nr);
1157
1158 _allow_barrier(conf, idx);
1159}
1160
1161/* conf->resync_lock should be held */
1162static int get_unqueued_pending(struct r1conf *conf)
1163{
1164 int idx, ret;
1165
1166 ret = atomic_read(&conf->nr_sync_pending);
1167 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++)
1168 ret += atomic_read(&conf->nr_pending[idx]) -
1169 atomic_read(&conf->nr_queued[idx]);
1170
1171 return ret;
1172}
1173
1174static void freeze_array(struct r1conf *conf, int extra)
1175{
1176 /* Stop sync I/O and normal I/O and wait for everything to
1177 * go quiet.
1178 * This is called in two situations:
1179 * 1) management command handlers (reshape, remove disk, quiesce).
1180 * 2) one normal I/O request failed.
1181
1182 * After array_frozen is set to 1, new sync IO will be blocked at
1183 * raise_barrier(), and new normal I/O will blocked at _wait_barrier()
1184 * or wait_read_barrier(). The flying I/Os will either complete or be
1185 * queued. When everything goes quite, there are only queued I/Os left.
1186
1187 * Every flying I/O contributes to a conf->nr_pending[idx], idx is the
1188 * barrier bucket index which this I/O request hits. When all sync and
1189 * normal I/O are queued, sum of all conf->nr_pending[] will match sum
1190 * of all conf->nr_queued[]. But normal I/O failure is an exception,
1191 * in handle_read_error(), we may call freeze_array() before trying to
1192 * fix the read error. In this case, the error read I/O is not queued,
1193 * so get_unqueued_pending() == 1.
1194 *
1195 * Therefore before this function returns, we need to wait until
1196 * get_unqueued_pendings(conf) gets equal to extra. For
1197 * normal I/O context, extra is 1, in rested situations extra is 0.
1198 */
1199 spin_lock_irq(&conf->resync_lock);
1200 conf->array_frozen = 1;
1201 mddev_add_trace_msg(conf->mddev, "raid1 wait freeze");
1202 wait_event_lock_irq_cmd(
1203 conf->wait_barrier,
1204 get_unqueued_pending(conf) == extra,
1205 conf->resync_lock,
1206 flush_pending_writes(conf));
1207 spin_unlock_irq(&conf->resync_lock);
1208}
1209static void unfreeze_array(struct r1conf *conf)
1210{
1211 /* reverse the effect of the freeze */
1212 spin_lock_irq(&conf->resync_lock);
1213 conf->array_frozen = 0;
1214 spin_unlock_irq(&conf->resync_lock);
1215 wake_up(&conf->wait_barrier);
1216}
1217
1218static void alloc_behind_master_bio(struct r1bio *r1_bio,
1219 struct bio *bio)
1220{
1221 int size = bio->bi_iter.bi_size;
1222 unsigned vcnt = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1223 int i = 0;
1224 struct bio *behind_bio = NULL;
1225
1226 behind_bio = bio_alloc_bioset(NULL, vcnt, 0, GFP_NOIO,
1227 &r1_bio->mddev->bio_set);
1228
1229 /* discard op, we don't support writezero/writesame yet */
1230 if (!bio_has_data(bio)) {
1231 behind_bio->bi_iter.bi_size = size;
1232 goto skip_copy;
1233 }
1234
1235 while (i < vcnt && size) {
1236 struct page *page;
1237 int len = min_t(int, PAGE_SIZE, size);
1238
1239 page = alloc_page(GFP_NOIO);
1240 if (unlikely(!page))
1241 goto free_pages;
1242
1243 if (!bio_add_page(behind_bio, page, len, 0)) {
1244 put_page(page);
1245 goto free_pages;
1246 }
1247
1248 size -= len;
1249 i++;
1250 }
1251
1252 bio_copy_data(behind_bio, bio);
1253skip_copy:
1254 r1_bio->behind_master_bio = behind_bio;
1255 set_bit(R1BIO_BehindIO, &r1_bio->state);
1256
1257 return;
1258
1259free_pages:
1260 pr_debug("%dB behind alloc failed, doing sync I/O\n",
1261 bio->bi_iter.bi_size);
1262 bio_free_pages(behind_bio);
1263 bio_put(behind_bio);
1264}
1265
1266static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
1267{
1268 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
1269 cb);
1270 struct mddev *mddev = plug->cb.data;
1271 struct r1conf *conf = mddev->private;
1272 struct bio *bio;
1273
1274 if (from_schedule) {
1275 spin_lock_irq(&conf->device_lock);
1276 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1277 spin_unlock_irq(&conf->device_lock);
1278 wake_up_barrier(conf);
1279 md_wakeup_thread(mddev->thread);
1280 kfree(plug);
1281 return;
1282 }
1283
1284 /* we aren't scheduling, so we can do the write-out directly. */
1285 bio = bio_list_get(&plug->pending);
1286 flush_bio_list(conf, bio);
1287 kfree(plug);
1288}
1289
1290static void init_r1bio(struct r1bio *r1_bio, struct mddev *mddev, struct bio *bio)
1291{
1292 r1_bio->master_bio = bio;
1293 r1_bio->sectors = bio_sectors(bio);
1294 r1_bio->state = 0;
1295 r1_bio->mddev = mddev;
1296 r1_bio->sector = bio->bi_iter.bi_sector;
1297}
1298
1299static inline struct r1bio *
1300alloc_r1bio(struct mddev *mddev, struct bio *bio)
1301{
1302 struct r1conf *conf = mddev->private;
1303 struct r1bio *r1_bio;
1304
1305 r1_bio = mempool_alloc(&conf->r1bio_pool, GFP_NOIO);
1306 /* Ensure no bio records IO_BLOCKED */
1307 memset(r1_bio->bios, 0, conf->raid_disks * sizeof(r1_bio->bios[0]));
1308 init_r1bio(r1_bio, mddev, bio);
1309 return r1_bio;
1310}
1311
1312static void raid1_read_request(struct mddev *mddev, struct bio *bio,
1313 int max_read_sectors, struct r1bio *r1_bio)
1314{
1315 struct r1conf *conf = mddev->private;
1316 struct raid1_info *mirror;
1317 struct bio *read_bio;
1318 const enum req_op op = bio_op(bio);
1319 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1320 int max_sectors;
1321 int rdisk, error;
1322 bool r1bio_existed = !!r1_bio;
1323
1324 /*
1325 * If r1_bio is set, we are blocking the raid1d thread
1326 * so there is a tiny risk of deadlock. So ask for
1327 * emergency memory if needed.
1328 */
1329 gfp_t gfp = r1_bio ? (GFP_NOIO | __GFP_HIGH) : GFP_NOIO;
1330
1331 /*
1332 * Still need barrier for READ in case that whole
1333 * array is frozen.
1334 */
1335 if (!wait_read_barrier(conf, bio->bi_iter.bi_sector,
1336 bio->bi_opf & REQ_NOWAIT)) {
1337 bio_wouldblock_error(bio);
1338 return;
1339 }
1340
1341 if (!r1_bio)
1342 r1_bio = alloc_r1bio(mddev, bio);
1343 else
1344 init_r1bio(r1_bio, mddev, bio);
1345 r1_bio->sectors = max_read_sectors;
1346
1347 /*
1348 * make_request() can abort the operation when read-ahead is being
1349 * used and no empty request is available.
1350 */
1351 rdisk = read_balance(conf, r1_bio, &max_sectors);
1352 if (rdisk < 0) {
1353 /* couldn't find anywhere to read from */
1354 if (r1bio_existed)
1355 pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
1356 mdname(mddev),
1357 conf->mirrors[r1_bio->read_disk].rdev->bdev,
1358 r1_bio->sector);
1359 raid_end_bio_io(r1_bio);
1360 return;
1361 }
1362 mirror = conf->mirrors + rdisk;
1363
1364 if (r1bio_existed)
1365 pr_info_ratelimited("md/raid1:%s: redirecting sector %llu to other mirror: %pg\n",
1366 mdname(mddev),
1367 (unsigned long long)r1_bio->sector,
1368 mirror->rdev->bdev);
1369
1370 if (test_bit(WriteMostly, &mirror->rdev->flags)) {
1371 /*
1372 * Reading from a write-mostly device must take care not to
1373 * over-take any writes that are 'behind'
1374 */
1375 mddev_add_trace_msg(mddev, "raid1 wait behind writes");
1376 mddev->bitmap_ops->wait_behind_writes(mddev);
1377 }
1378
1379 if (max_sectors < bio_sectors(bio)) {
1380 struct bio *split = bio_split(bio, max_sectors,
1381 gfp, &conf->bio_split);
1382
1383 if (IS_ERR(split)) {
1384 error = PTR_ERR(split);
1385 goto err_handle;
1386 }
1387 bio_chain(split, bio);
1388 submit_bio_noacct(bio);
1389 bio = split;
1390 r1_bio->master_bio = bio;
1391 r1_bio->sectors = max_sectors;
1392 }
1393
1394 r1_bio->read_disk = rdisk;
1395 if (!r1bio_existed) {
1396 md_account_bio(mddev, &bio);
1397 r1_bio->master_bio = bio;
1398 }
1399 read_bio = bio_alloc_clone(mirror->rdev->bdev, bio, gfp,
1400 &mddev->bio_set);
1401
1402 r1_bio->bios[rdisk] = read_bio;
1403
1404 read_bio->bi_iter.bi_sector = r1_bio->sector +
1405 mirror->rdev->data_offset;
1406 read_bio->bi_end_io = raid1_end_read_request;
1407 read_bio->bi_opf = op | do_sync;
1408 if (test_bit(FailFast, &mirror->rdev->flags) &&
1409 test_bit(R1BIO_FailFast, &r1_bio->state))
1410 read_bio->bi_opf |= MD_FAILFAST;
1411 read_bio->bi_private = r1_bio;
1412 mddev_trace_remap(mddev, read_bio, r1_bio->sector);
1413 submit_bio_noacct(read_bio);
1414 return;
1415
1416err_handle:
1417 atomic_dec(&mirror->rdev->nr_pending);
1418 bio->bi_status = errno_to_blk_status(error);
1419 set_bit(R1BIO_Uptodate, &r1_bio->state);
1420 raid_end_bio_io(r1_bio);
1421}
1422
1423static bool wait_blocked_rdev(struct mddev *mddev, struct bio *bio)
1424{
1425 struct r1conf *conf = mddev->private;
1426 int disks = conf->raid_disks * 2;
1427 int i;
1428
1429retry:
1430 for (i = 0; i < disks; i++) {
1431 struct md_rdev *rdev = conf->mirrors[i].rdev;
1432
1433 if (!rdev)
1434 continue;
1435
1436 /* don't write here until the bad block is acknowledged */
1437 if (test_bit(WriteErrorSeen, &rdev->flags) &&
1438 rdev_has_badblock(rdev, bio->bi_iter.bi_sector,
1439 bio_sectors(bio)) < 0)
1440 set_bit(BlockedBadBlocks, &rdev->flags);
1441
1442 if (rdev_blocked(rdev)) {
1443 if (bio->bi_opf & REQ_NOWAIT)
1444 return false;
1445
1446 mddev_add_trace_msg(rdev->mddev, "raid1 wait rdev %d blocked",
1447 rdev->raid_disk);
1448 atomic_inc(&rdev->nr_pending);
1449 md_wait_for_blocked_rdev(rdev, rdev->mddev);
1450 goto retry;
1451 }
1452 }
1453
1454 return true;
1455}
1456
1457static void raid1_write_request(struct mddev *mddev, struct bio *bio,
1458 int max_write_sectors)
1459{
1460 struct r1conf *conf = mddev->private;
1461 struct r1bio *r1_bio;
1462 int i, disks, k, error;
1463 unsigned long flags;
1464 int first_clone;
1465 int max_sectors;
1466 bool write_behind = false;
1467 bool is_discard = (bio_op(bio) == REQ_OP_DISCARD);
1468
1469 if (mddev_is_clustered(mddev) &&
1470 md_cluster_ops->area_resyncing(mddev, WRITE,
1471 bio->bi_iter.bi_sector, bio_end_sector(bio))) {
1472
1473 DEFINE_WAIT(w);
1474 if (bio->bi_opf & REQ_NOWAIT) {
1475 bio_wouldblock_error(bio);
1476 return;
1477 }
1478 for (;;) {
1479 prepare_to_wait(&conf->wait_barrier,
1480 &w, TASK_IDLE);
1481 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1482 bio->bi_iter.bi_sector,
1483 bio_end_sector(bio)))
1484 break;
1485 schedule();
1486 }
1487 finish_wait(&conf->wait_barrier, &w);
1488 }
1489
1490 /*
1491 * Register the new request and wait if the reconstruction
1492 * thread has put up a bar for new requests.
1493 * Continue immediately if no resync is active currently.
1494 */
1495 if (!wait_barrier(conf, bio->bi_iter.bi_sector,
1496 bio->bi_opf & REQ_NOWAIT)) {
1497 bio_wouldblock_error(bio);
1498 return;
1499 }
1500
1501 if (!wait_blocked_rdev(mddev, bio)) {
1502 bio_wouldblock_error(bio);
1503 return;
1504 }
1505
1506 r1_bio = alloc_r1bio(mddev, bio);
1507 r1_bio->sectors = max_write_sectors;
1508
1509 /* first select target devices under rcu_lock and
1510 * inc refcount on their rdev. Record them by setting
1511 * bios[x] to bio
1512 * If there are known/acknowledged bad blocks on any device on
1513 * which we have seen a write error, we want to avoid writing those
1514 * blocks.
1515 * This potentially requires several writes to write around
1516 * the bad blocks. Each set of writes gets it's own r1bio
1517 * with a set of bios attached.
1518 */
1519
1520 disks = conf->raid_disks * 2;
1521 max_sectors = r1_bio->sectors;
1522 for (i = 0; i < disks; i++) {
1523 struct md_rdev *rdev = conf->mirrors[i].rdev;
1524
1525 /*
1526 * The write-behind io is only attempted on drives marked as
1527 * write-mostly, which means we could allocate write behind
1528 * bio later.
1529 */
1530 if (!is_discard && rdev && test_bit(WriteMostly, &rdev->flags))
1531 write_behind = true;
1532
1533 r1_bio->bios[i] = NULL;
1534 if (!rdev || test_bit(Faulty, &rdev->flags))
1535 continue;
1536
1537 atomic_inc(&rdev->nr_pending);
1538 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1539 sector_t first_bad;
1540 int bad_sectors;
1541 int is_bad;
1542
1543 is_bad = is_badblock(rdev, r1_bio->sector, max_sectors,
1544 &first_bad, &bad_sectors);
1545 if (is_bad && first_bad <= r1_bio->sector) {
1546 /* Cannot write here at all */
1547 bad_sectors -= (r1_bio->sector - first_bad);
1548 if (bad_sectors < max_sectors)
1549 /* mustn't write more than bad_sectors
1550 * to other devices yet
1551 */
1552 max_sectors = bad_sectors;
1553 rdev_dec_pending(rdev, mddev);
1554 continue;
1555 }
1556 if (is_bad) {
1557 int good_sectors;
1558
1559 /*
1560 * We cannot atomically write this, so just
1561 * error in that case. It could be possible to
1562 * atomically write other mirrors, but the
1563 * complexity of supporting that is not worth
1564 * the benefit.
1565 */
1566 if (bio->bi_opf & REQ_ATOMIC) {
1567 error = -EIO;
1568 goto err_handle;
1569 }
1570
1571 good_sectors = first_bad - r1_bio->sector;
1572 if (good_sectors < max_sectors)
1573 max_sectors = good_sectors;
1574 }
1575 }
1576 r1_bio->bios[i] = bio;
1577 }
1578
1579 /*
1580 * When using a bitmap, we may call alloc_behind_master_bio below.
1581 * alloc_behind_master_bio allocates a copy of the data payload a page
1582 * at a time and thus needs a new bio that can fit the whole payload
1583 * this bio in page sized chunks.
1584 */
1585 if (write_behind && mddev->bitmap)
1586 max_sectors = min_t(int, max_sectors,
1587 BIO_MAX_VECS * (PAGE_SIZE >> 9));
1588 if (max_sectors < bio_sectors(bio)) {
1589 struct bio *split = bio_split(bio, max_sectors,
1590 GFP_NOIO, &conf->bio_split);
1591
1592 if (IS_ERR(split)) {
1593 error = PTR_ERR(split);
1594 goto err_handle;
1595 }
1596 bio_chain(split, bio);
1597 submit_bio_noacct(bio);
1598 bio = split;
1599 r1_bio->master_bio = bio;
1600 r1_bio->sectors = max_sectors;
1601 }
1602
1603 md_account_bio(mddev, &bio);
1604 r1_bio->master_bio = bio;
1605 atomic_set(&r1_bio->remaining, 1);
1606 atomic_set(&r1_bio->behind_remaining, 0);
1607
1608 first_clone = 1;
1609
1610 for (i = 0; i < disks; i++) {
1611 struct bio *mbio = NULL;
1612 struct md_rdev *rdev = conf->mirrors[i].rdev;
1613 if (!r1_bio->bios[i])
1614 continue;
1615
1616 if (first_clone) {
1617 unsigned long max_write_behind =
1618 mddev->bitmap_info.max_write_behind;
1619 struct md_bitmap_stats stats;
1620 int err;
1621
1622 /* do behind I/O ?
1623 * Not if there are too many, or cannot
1624 * allocate memory, or a reader on WriteMostly
1625 * is waiting for behind writes to flush */
1626 err = mddev->bitmap_ops->get_stats(mddev->bitmap, &stats);
1627 if (!err && write_behind && !stats.behind_wait &&
1628 stats.behind_writes < max_write_behind)
1629 alloc_behind_master_bio(r1_bio, bio);
1630
1631 if (test_bit(R1BIO_BehindIO, &r1_bio->state))
1632 mddev->bitmap_ops->start_behind_write(mddev);
1633 first_clone = 0;
1634 }
1635
1636 if (r1_bio->behind_master_bio) {
1637 mbio = bio_alloc_clone(rdev->bdev,
1638 r1_bio->behind_master_bio,
1639 GFP_NOIO, &mddev->bio_set);
1640 if (test_bit(CollisionCheck, &rdev->flags))
1641 wait_for_serialization(rdev, r1_bio);
1642 if (test_bit(WriteMostly, &rdev->flags))
1643 atomic_inc(&r1_bio->behind_remaining);
1644 } else {
1645 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
1646 &mddev->bio_set);
1647
1648 if (mddev->serialize_policy)
1649 wait_for_serialization(rdev, r1_bio);
1650 }
1651
1652 r1_bio->bios[i] = mbio;
1653
1654 mbio->bi_iter.bi_sector = (r1_bio->sector + rdev->data_offset);
1655 mbio->bi_end_io = raid1_end_write_request;
1656 mbio->bi_opf = bio_op(bio) |
1657 (bio->bi_opf & (REQ_SYNC | REQ_FUA | REQ_ATOMIC));
1658 if (test_bit(FailFast, &rdev->flags) &&
1659 !test_bit(WriteMostly, &rdev->flags) &&
1660 conf->raid_disks - mddev->degraded > 1)
1661 mbio->bi_opf |= MD_FAILFAST;
1662 mbio->bi_private = r1_bio;
1663
1664 atomic_inc(&r1_bio->remaining);
1665 mddev_trace_remap(mddev, mbio, r1_bio->sector);
1666 /* flush_pending_writes() needs access to the rdev so...*/
1667 mbio->bi_bdev = (void *)rdev;
1668 if (!raid1_add_bio_to_plug(mddev, mbio, raid1_unplug, disks)) {
1669 spin_lock_irqsave(&conf->device_lock, flags);
1670 bio_list_add(&conf->pending_bio_list, mbio);
1671 spin_unlock_irqrestore(&conf->device_lock, flags);
1672 md_wakeup_thread(mddev->thread);
1673 }
1674 }
1675
1676 r1_bio_write_done(r1_bio);
1677
1678 /* In case raid1d snuck in to freeze_array */
1679 wake_up_barrier(conf);
1680 return;
1681err_handle:
1682 for (k = 0; k < i; k++) {
1683 if (r1_bio->bios[k]) {
1684 rdev_dec_pending(conf->mirrors[k].rdev, mddev);
1685 r1_bio->bios[k] = NULL;
1686 }
1687 }
1688
1689 bio->bi_status = errno_to_blk_status(error);
1690 set_bit(R1BIO_Uptodate, &r1_bio->state);
1691 raid_end_bio_io(r1_bio);
1692}
1693
1694static bool raid1_make_request(struct mddev *mddev, struct bio *bio)
1695{
1696 sector_t sectors;
1697
1698 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1699 && md_flush_request(mddev, bio))
1700 return true;
1701
1702 /*
1703 * There is a limit to the maximum size, but
1704 * the read/write handler might find a lower limit
1705 * due to bad blocks. To avoid multiple splits,
1706 * we pass the maximum number of sectors down
1707 * and let the lower level perform the split.
1708 */
1709 sectors = align_to_barrier_unit_end(
1710 bio->bi_iter.bi_sector, bio_sectors(bio));
1711
1712 if (bio_data_dir(bio) == READ)
1713 raid1_read_request(mddev, bio, sectors, NULL);
1714 else {
1715 md_write_start(mddev,bio);
1716 raid1_write_request(mddev, bio, sectors);
1717 }
1718 return true;
1719}
1720
1721static void raid1_status(struct seq_file *seq, struct mddev *mddev)
1722{
1723 struct r1conf *conf = mddev->private;
1724 int i;
1725
1726 lockdep_assert_held(&mddev->lock);
1727
1728 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1729 conf->raid_disks - mddev->degraded);
1730 for (i = 0; i < conf->raid_disks; i++) {
1731 struct md_rdev *rdev = READ_ONCE(conf->mirrors[i].rdev);
1732
1733 seq_printf(seq, "%s",
1734 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1735 }
1736 seq_printf(seq, "]");
1737}
1738
1739/**
1740 * raid1_error() - RAID1 error handler.
1741 * @mddev: affected md device.
1742 * @rdev: member device to fail.
1743 *
1744 * The routine acknowledges &rdev failure and determines new @mddev state.
1745 * If it failed, then:
1746 * - &MD_BROKEN flag is set in &mddev->flags.
1747 * - recovery is disabled.
1748 * Otherwise, it must be degraded:
1749 * - recovery is interrupted.
1750 * - &mddev->degraded is bumped.
1751 *
1752 * @rdev is marked as &Faulty excluding case when array is failed and
1753 * &mddev->fail_last_dev is off.
1754 */
1755static void raid1_error(struct mddev *mddev, struct md_rdev *rdev)
1756{
1757 struct r1conf *conf = mddev->private;
1758 unsigned long flags;
1759
1760 spin_lock_irqsave(&conf->device_lock, flags);
1761
1762 if (test_bit(In_sync, &rdev->flags) &&
1763 (conf->raid_disks - mddev->degraded) == 1) {
1764 set_bit(MD_BROKEN, &mddev->flags);
1765
1766 if (!mddev->fail_last_dev) {
1767 conf->recovery_disabled = mddev->recovery_disabled;
1768 spin_unlock_irqrestore(&conf->device_lock, flags);
1769 return;
1770 }
1771 }
1772 set_bit(Blocked, &rdev->flags);
1773 if (test_and_clear_bit(In_sync, &rdev->flags))
1774 mddev->degraded++;
1775 set_bit(Faulty, &rdev->flags);
1776 spin_unlock_irqrestore(&conf->device_lock, flags);
1777 /*
1778 * if recovery is running, make sure it aborts.
1779 */
1780 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1781 set_mask_bits(&mddev->sb_flags, 0,
1782 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1783 pr_crit("md/raid1:%s: Disk failure on %pg, disabling device.\n"
1784 "md/raid1:%s: Operation continuing on %d devices.\n",
1785 mdname(mddev), rdev->bdev,
1786 mdname(mddev), conf->raid_disks - mddev->degraded);
1787}
1788
1789static void print_conf(struct r1conf *conf)
1790{
1791 int i;
1792
1793 pr_debug("RAID1 conf printout:\n");
1794 if (!conf) {
1795 pr_debug("(!conf)\n");
1796 return;
1797 }
1798 pr_debug(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1799 conf->raid_disks);
1800
1801 lockdep_assert_held(&conf->mddev->reconfig_mutex);
1802 for (i = 0; i < conf->raid_disks; i++) {
1803 struct md_rdev *rdev = conf->mirrors[i].rdev;
1804 if (rdev)
1805 pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
1806 i, !test_bit(In_sync, &rdev->flags),
1807 !test_bit(Faulty, &rdev->flags),
1808 rdev->bdev);
1809 }
1810}
1811
1812static void close_sync(struct r1conf *conf)
1813{
1814 int idx;
1815
1816 for (idx = 0; idx < BARRIER_BUCKETS_NR; idx++) {
1817 _wait_barrier(conf, idx, false);
1818 _allow_barrier(conf, idx);
1819 }
1820
1821 mempool_exit(&conf->r1buf_pool);
1822}
1823
1824static int raid1_spare_active(struct mddev *mddev)
1825{
1826 int i;
1827 struct r1conf *conf = mddev->private;
1828 int count = 0;
1829 unsigned long flags;
1830
1831 /*
1832 * Find all failed disks within the RAID1 configuration
1833 * and mark them readable.
1834 * Called under mddev lock, so rcu protection not needed.
1835 * device_lock used to avoid races with raid1_end_read_request
1836 * which expects 'In_sync' flags and ->degraded to be consistent.
1837 */
1838 spin_lock_irqsave(&conf->device_lock, flags);
1839 for (i = 0; i < conf->raid_disks; i++) {
1840 struct md_rdev *rdev = conf->mirrors[i].rdev;
1841 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1842 if (repl
1843 && !test_bit(Candidate, &repl->flags)
1844 && repl->recovery_offset == MaxSector
1845 && !test_bit(Faulty, &repl->flags)
1846 && !test_and_set_bit(In_sync, &repl->flags)) {
1847 /* replacement has just become active */
1848 if (!rdev ||
1849 !test_and_clear_bit(In_sync, &rdev->flags))
1850 count++;
1851 if (rdev) {
1852 /* Replaced device not technically
1853 * faulty, but we need to be sure
1854 * it gets removed and never re-added
1855 */
1856 set_bit(Faulty, &rdev->flags);
1857 sysfs_notify_dirent_safe(
1858 rdev->sysfs_state);
1859 }
1860 }
1861 if (rdev
1862 && rdev->recovery_offset == MaxSector
1863 && !test_bit(Faulty, &rdev->flags)
1864 && !test_and_set_bit(In_sync, &rdev->flags)) {
1865 count++;
1866 sysfs_notify_dirent_safe(rdev->sysfs_state);
1867 }
1868 }
1869 mddev->degraded -= count;
1870 spin_unlock_irqrestore(&conf->device_lock, flags);
1871
1872 print_conf(conf);
1873 return count;
1874}
1875
1876static bool raid1_add_conf(struct r1conf *conf, struct md_rdev *rdev, int disk,
1877 bool replacement)
1878{
1879 struct raid1_info *info = conf->mirrors + disk;
1880
1881 if (replacement)
1882 info += conf->raid_disks;
1883
1884 if (info->rdev)
1885 return false;
1886
1887 if (bdev_nonrot(rdev->bdev)) {
1888 set_bit(Nonrot, &rdev->flags);
1889 WRITE_ONCE(conf->nonrot_disks, conf->nonrot_disks + 1);
1890 }
1891
1892 rdev->raid_disk = disk;
1893 info->head_position = 0;
1894 info->seq_start = MaxSector;
1895 WRITE_ONCE(info->rdev, rdev);
1896
1897 return true;
1898}
1899
1900static bool raid1_remove_conf(struct r1conf *conf, int disk)
1901{
1902 struct raid1_info *info = conf->mirrors + disk;
1903 struct md_rdev *rdev = info->rdev;
1904
1905 if (!rdev || test_bit(In_sync, &rdev->flags) ||
1906 atomic_read(&rdev->nr_pending))
1907 return false;
1908
1909 /* Only remove non-faulty devices if recovery is not possible. */
1910 if (!test_bit(Faulty, &rdev->flags) &&
1911 rdev->mddev->recovery_disabled != conf->recovery_disabled &&
1912 rdev->mddev->degraded < conf->raid_disks)
1913 return false;
1914
1915 if (test_and_clear_bit(Nonrot, &rdev->flags))
1916 WRITE_ONCE(conf->nonrot_disks, conf->nonrot_disks - 1);
1917
1918 WRITE_ONCE(info->rdev, NULL);
1919 return true;
1920}
1921
1922static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1923{
1924 struct r1conf *conf = mddev->private;
1925 int err = -EEXIST;
1926 int mirror = 0, repl_slot = -1;
1927 struct raid1_info *p;
1928 int first = 0;
1929 int last = conf->raid_disks - 1;
1930
1931 if (mddev->recovery_disabled == conf->recovery_disabled)
1932 return -EBUSY;
1933
1934 if (rdev->raid_disk >= 0)
1935 first = last = rdev->raid_disk;
1936
1937 /*
1938 * find the disk ... but prefer rdev->saved_raid_disk
1939 * if possible.
1940 */
1941 if (rdev->saved_raid_disk >= 0 &&
1942 rdev->saved_raid_disk >= first &&
1943 rdev->saved_raid_disk < conf->raid_disks &&
1944 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1945 first = last = rdev->saved_raid_disk;
1946
1947 for (mirror = first; mirror <= last; mirror++) {
1948 p = conf->mirrors + mirror;
1949 if (!p->rdev) {
1950 err = mddev_stack_new_rdev(mddev, rdev);
1951 if (err)
1952 return err;
1953
1954 raid1_add_conf(conf, rdev, mirror, false);
1955 /* As all devices are equivalent, we don't need a full recovery
1956 * if this was recently any drive of the array
1957 */
1958 if (rdev->saved_raid_disk < 0)
1959 conf->fullsync = 1;
1960 break;
1961 }
1962 if (test_bit(WantReplacement, &p->rdev->flags) &&
1963 p[conf->raid_disks].rdev == NULL && repl_slot < 0)
1964 repl_slot = mirror;
1965 }
1966
1967 if (err && repl_slot >= 0) {
1968 /* Add this device as a replacement */
1969 clear_bit(In_sync, &rdev->flags);
1970 set_bit(Replacement, &rdev->flags);
1971 raid1_add_conf(conf, rdev, repl_slot, true);
1972 err = 0;
1973 conf->fullsync = 1;
1974 }
1975
1976 print_conf(conf);
1977 return err;
1978}
1979
1980static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1981{
1982 struct r1conf *conf = mddev->private;
1983 int err = 0;
1984 int number = rdev->raid_disk;
1985 struct raid1_info *p = conf->mirrors + number;
1986
1987 if (unlikely(number >= conf->raid_disks))
1988 goto abort;
1989
1990 if (rdev != p->rdev) {
1991 number += conf->raid_disks;
1992 p = conf->mirrors + number;
1993 }
1994
1995 print_conf(conf);
1996 if (rdev == p->rdev) {
1997 if (!raid1_remove_conf(conf, number)) {
1998 err = -EBUSY;
1999 goto abort;
2000 }
2001
2002 if (number < conf->raid_disks &&
2003 conf->mirrors[conf->raid_disks + number].rdev) {
2004 /* We just removed a device that is being replaced.
2005 * Move down the replacement. We drain all IO before
2006 * doing this to avoid confusion.
2007 */
2008 struct md_rdev *repl =
2009 conf->mirrors[conf->raid_disks + number].rdev;
2010 freeze_array(conf, 0);
2011 if (atomic_read(&repl->nr_pending)) {
2012 /* It means that some queued IO of retry_list
2013 * hold repl. Thus, we cannot set replacement
2014 * as NULL, avoiding rdev NULL pointer
2015 * dereference in sync_request_write and
2016 * handle_write_finished.
2017 */
2018 err = -EBUSY;
2019 unfreeze_array(conf);
2020 goto abort;
2021 }
2022 clear_bit(Replacement, &repl->flags);
2023 WRITE_ONCE(p->rdev, repl);
2024 conf->mirrors[conf->raid_disks + number].rdev = NULL;
2025 unfreeze_array(conf);
2026 }
2027
2028 clear_bit(WantReplacement, &rdev->flags);
2029 err = md_integrity_register(mddev);
2030 }
2031abort:
2032
2033 print_conf(conf);
2034 return err;
2035}
2036
2037static void end_sync_read(struct bio *bio)
2038{
2039 struct r1bio *r1_bio = get_resync_r1bio(bio);
2040
2041 update_head_pos(r1_bio->read_disk, r1_bio);
2042
2043 /*
2044 * we have read a block, now it needs to be re-written,
2045 * or re-read if the read failed.
2046 * We don't do much here, just schedule handling by raid1d
2047 */
2048 if (!bio->bi_status)
2049 set_bit(R1BIO_Uptodate, &r1_bio->state);
2050
2051 if (atomic_dec_and_test(&r1_bio->remaining))
2052 reschedule_retry(r1_bio);
2053}
2054
2055static void abort_sync_write(struct mddev *mddev, struct r1bio *r1_bio)
2056{
2057 sector_t sync_blocks = 0;
2058 sector_t s = r1_bio->sector;
2059 long sectors_to_go = r1_bio->sectors;
2060
2061 /* make sure these bits don't get cleared. */
2062 do {
2063 mddev->bitmap_ops->end_sync(mddev, s, &sync_blocks);
2064 s += sync_blocks;
2065 sectors_to_go -= sync_blocks;
2066 } while (sectors_to_go > 0);
2067}
2068
2069static void put_sync_write_buf(struct r1bio *r1_bio, int uptodate)
2070{
2071 if (atomic_dec_and_test(&r1_bio->remaining)) {
2072 struct mddev *mddev = r1_bio->mddev;
2073 int s = r1_bio->sectors;
2074
2075 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2076 test_bit(R1BIO_WriteError, &r1_bio->state))
2077 reschedule_retry(r1_bio);
2078 else {
2079 put_buf(r1_bio);
2080 md_done_sync(mddev, s, uptodate);
2081 }
2082 }
2083}
2084
2085static void end_sync_write(struct bio *bio)
2086{
2087 int uptodate = !bio->bi_status;
2088 struct r1bio *r1_bio = get_resync_r1bio(bio);
2089 struct mddev *mddev = r1_bio->mddev;
2090 struct r1conf *conf = mddev->private;
2091 struct md_rdev *rdev = conf->mirrors[find_bio_disk(r1_bio, bio)].rdev;
2092
2093 if (!uptodate) {
2094 abort_sync_write(mddev, r1_bio);
2095 set_bit(WriteErrorSeen, &rdev->flags);
2096 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2097 set_bit(MD_RECOVERY_NEEDED, &
2098 mddev->recovery);
2099 set_bit(R1BIO_WriteError, &r1_bio->state);
2100 } else if (rdev_has_badblock(rdev, r1_bio->sector, r1_bio->sectors) &&
2101 !rdev_has_badblock(conf->mirrors[r1_bio->read_disk].rdev,
2102 r1_bio->sector, r1_bio->sectors)) {
2103 set_bit(R1BIO_MadeGood, &r1_bio->state);
2104 }
2105
2106 put_sync_write_buf(r1_bio, uptodate);
2107}
2108
2109static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
2110 int sectors, struct page *page, blk_opf_t rw)
2111{
2112 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2113 /* success */
2114 return 1;
2115 if (rw == REQ_OP_WRITE) {
2116 set_bit(WriteErrorSeen, &rdev->flags);
2117 if (!test_and_set_bit(WantReplacement,
2118 &rdev->flags))
2119 set_bit(MD_RECOVERY_NEEDED, &
2120 rdev->mddev->recovery);
2121 }
2122 /* need to record an error - either for the block or the device */
2123 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2124 md_error(rdev->mddev, rdev);
2125 return 0;
2126}
2127
2128static int fix_sync_read_error(struct r1bio *r1_bio)
2129{
2130 /* Try some synchronous reads of other devices to get
2131 * good data, much like with normal read errors. Only
2132 * read into the pages we already have so we don't
2133 * need to re-issue the read request.
2134 * We don't need to freeze the array, because being in an
2135 * active sync request, there is no normal IO, and
2136 * no overlapping syncs.
2137 * We don't need to check is_badblock() again as we
2138 * made sure that anything with a bad block in range
2139 * will have bi_end_io clear.
2140 */
2141 struct mddev *mddev = r1_bio->mddev;
2142 struct r1conf *conf = mddev->private;
2143 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
2144 struct page **pages = get_resync_pages(bio)->pages;
2145 sector_t sect = r1_bio->sector;
2146 int sectors = r1_bio->sectors;
2147 int idx = 0;
2148 struct md_rdev *rdev;
2149
2150 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2151 if (test_bit(FailFast, &rdev->flags)) {
2152 /* Don't try recovering from here - just fail it
2153 * ... unless it is the last working device of course */
2154 md_error(mddev, rdev);
2155 if (test_bit(Faulty, &rdev->flags))
2156 /* Don't try to read from here, but make sure
2157 * put_buf does it's thing
2158 */
2159 bio->bi_end_io = end_sync_write;
2160 }
2161
2162 while(sectors) {
2163 int s = sectors;
2164 int d = r1_bio->read_disk;
2165 int success = 0;
2166 int start;
2167
2168 if (s > (PAGE_SIZE>>9))
2169 s = PAGE_SIZE >> 9;
2170 do {
2171 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
2172 /* No rcu protection needed here devices
2173 * can only be removed when no resync is
2174 * active, and resync is currently active
2175 */
2176 rdev = conf->mirrors[d].rdev;
2177 if (sync_page_io(rdev, sect, s<<9,
2178 pages[idx],
2179 REQ_OP_READ, false)) {
2180 success = 1;
2181 break;
2182 }
2183 }
2184 d++;
2185 if (d == conf->raid_disks * 2)
2186 d = 0;
2187 } while (!success && d != r1_bio->read_disk);
2188
2189 if (!success) {
2190 int abort = 0;
2191 /* Cannot read from anywhere, this block is lost.
2192 * Record a bad block on each device. If that doesn't
2193 * work just disable and interrupt the recovery.
2194 * Don't fail devices as that won't really help.
2195 */
2196 pr_crit_ratelimited("md/raid1:%s: %pg: unrecoverable I/O read error for block %llu\n",
2197 mdname(mddev), bio->bi_bdev,
2198 (unsigned long long)r1_bio->sector);
2199 for (d = 0; d < conf->raid_disks * 2; d++) {
2200 rdev = conf->mirrors[d].rdev;
2201 if (!rdev || test_bit(Faulty, &rdev->flags))
2202 continue;
2203 if (!rdev_set_badblocks(rdev, sect, s, 0))
2204 abort = 1;
2205 }
2206 if (abort) {
2207 conf->recovery_disabled =
2208 mddev->recovery_disabled;
2209 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2210 md_done_sync(mddev, r1_bio->sectors, 0);
2211 put_buf(r1_bio);
2212 return 0;
2213 }
2214 /* Try next page */
2215 sectors -= s;
2216 sect += s;
2217 idx++;
2218 continue;
2219 }
2220
2221 start = d;
2222 /* write it back and re-read */
2223 while (d != r1_bio->read_disk) {
2224 if (d == 0)
2225 d = conf->raid_disks * 2;
2226 d--;
2227 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2228 continue;
2229 rdev = conf->mirrors[d].rdev;
2230 if (r1_sync_page_io(rdev, sect, s,
2231 pages[idx],
2232 REQ_OP_WRITE) == 0) {
2233 r1_bio->bios[d]->bi_end_io = NULL;
2234 rdev_dec_pending(rdev, mddev);
2235 }
2236 }
2237 d = start;
2238 while (d != r1_bio->read_disk) {
2239 if (d == 0)
2240 d = conf->raid_disks * 2;
2241 d--;
2242 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
2243 continue;
2244 rdev = conf->mirrors[d].rdev;
2245 if (r1_sync_page_io(rdev, sect, s,
2246 pages[idx],
2247 REQ_OP_READ) != 0)
2248 atomic_add(s, &rdev->corrected_errors);
2249 }
2250 sectors -= s;
2251 sect += s;
2252 idx ++;
2253 }
2254 set_bit(R1BIO_Uptodate, &r1_bio->state);
2255 bio->bi_status = 0;
2256 return 1;
2257}
2258
2259static void process_checks(struct r1bio *r1_bio)
2260{
2261 /* We have read all readable devices. If we haven't
2262 * got the block, then there is no hope left.
2263 * If we have, then we want to do a comparison
2264 * and skip the write if everything is the same.
2265 * If any blocks failed to read, then we need to
2266 * attempt an over-write
2267 */
2268 struct mddev *mddev = r1_bio->mddev;
2269 struct r1conf *conf = mddev->private;
2270 int primary;
2271 int i;
2272 int vcnt;
2273
2274 /* Fix variable parts of all bios */
2275 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
2276 for (i = 0; i < conf->raid_disks * 2; i++) {
2277 blk_status_t status;
2278 struct bio *b = r1_bio->bios[i];
2279 struct resync_pages *rp = get_resync_pages(b);
2280 if (b->bi_end_io != end_sync_read)
2281 continue;
2282 /* fixup the bio for reuse, but preserve errno */
2283 status = b->bi_status;
2284 bio_reset(b, conf->mirrors[i].rdev->bdev, REQ_OP_READ);
2285 b->bi_status = status;
2286 b->bi_iter.bi_sector = r1_bio->sector +
2287 conf->mirrors[i].rdev->data_offset;
2288 b->bi_end_io = end_sync_read;
2289 rp->raid_bio = r1_bio;
2290 b->bi_private = rp;
2291
2292 /* initialize bvec table again */
2293 md_bio_reset_resync_pages(b, rp, r1_bio->sectors << 9);
2294 }
2295 for (primary = 0; primary < conf->raid_disks * 2; primary++)
2296 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
2297 !r1_bio->bios[primary]->bi_status) {
2298 r1_bio->bios[primary]->bi_end_io = NULL;
2299 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
2300 break;
2301 }
2302 r1_bio->read_disk = primary;
2303 for (i = 0; i < conf->raid_disks * 2; i++) {
2304 int j = 0;
2305 struct bio *pbio = r1_bio->bios[primary];
2306 struct bio *sbio = r1_bio->bios[i];
2307 blk_status_t status = sbio->bi_status;
2308 struct page **ppages = get_resync_pages(pbio)->pages;
2309 struct page **spages = get_resync_pages(sbio)->pages;
2310 struct bio_vec *bi;
2311 int page_len[RESYNC_PAGES] = { 0 };
2312 struct bvec_iter_all iter_all;
2313
2314 if (sbio->bi_end_io != end_sync_read)
2315 continue;
2316 /* Now we can 'fixup' the error value */
2317 sbio->bi_status = 0;
2318
2319 bio_for_each_segment_all(bi, sbio, iter_all)
2320 page_len[j++] = bi->bv_len;
2321
2322 if (!status) {
2323 for (j = vcnt; j-- ; ) {
2324 if (memcmp(page_address(ppages[j]),
2325 page_address(spages[j]),
2326 page_len[j]))
2327 break;
2328 }
2329 } else
2330 j = 0;
2331 if (j >= 0)
2332 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
2333 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
2334 && !status)) {
2335 /* No need to write to this device. */
2336 sbio->bi_end_io = NULL;
2337 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
2338 continue;
2339 }
2340
2341 bio_copy_data(sbio, pbio);
2342 }
2343}
2344
2345static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
2346{
2347 struct r1conf *conf = mddev->private;
2348 int i;
2349 int disks = conf->raid_disks * 2;
2350 struct bio *wbio;
2351
2352 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
2353 /* ouch - failed to read all of that. */
2354 if (!fix_sync_read_error(r1_bio))
2355 return;
2356
2357 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2358 process_checks(r1_bio);
2359
2360 /*
2361 * schedule writes
2362 */
2363 atomic_set(&r1_bio->remaining, 1);
2364 for (i = 0; i < disks ; i++) {
2365 wbio = r1_bio->bios[i];
2366 if (wbio->bi_end_io == NULL ||
2367 (wbio->bi_end_io == end_sync_read &&
2368 (i == r1_bio->read_disk ||
2369 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
2370 continue;
2371 if (test_bit(Faulty, &conf->mirrors[i].rdev->flags)) {
2372 abort_sync_write(mddev, r1_bio);
2373 continue;
2374 }
2375
2376 wbio->bi_opf = REQ_OP_WRITE;
2377 if (test_bit(FailFast, &conf->mirrors[i].rdev->flags))
2378 wbio->bi_opf |= MD_FAILFAST;
2379
2380 wbio->bi_end_io = end_sync_write;
2381 atomic_inc(&r1_bio->remaining);
2382 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
2383
2384 submit_bio_noacct(wbio);
2385 }
2386
2387 put_sync_write_buf(r1_bio, 1);
2388}
2389
2390/*
2391 * This is a kernel thread which:
2392 *
2393 * 1. Retries failed read operations on working mirrors.
2394 * 2. Updates the raid superblock when problems encounter.
2395 * 3. Performs writes following reads for array synchronising.
2396 */
2397
2398static void fix_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2399{
2400 sector_t sect = r1_bio->sector;
2401 int sectors = r1_bio->sectors;
2402 int read_disk = r1_bio->read_disk;
2403 struct mddev *mddev = conf->mddev;
2404 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2405
2406 if (exceed_read_errors(mddev, rdev)) {
2407 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2408 return;
2409 }
2410
2411 while(sectors) {
2412 int s = sectors;
2413 int d = read_disk;
2414 int success = 0;
2415 int start;
2416
2417 if (s > (PAGE_SIZE>>9))
2418 s = PAGE_SIZE >> 9;
2419
2420 do {
2421 rdev = conf->mirrors[d].rdev;
2422 if (rdev &&
2423 (test_bit(In_sync, &rdev->flags) ||
2424 (!test_bit(Faulty, &rdev->flags) &&
2425 rdev->recovery_offset >= sect + s)) &&
2426 rdev_has_badblock(rdev, sect, s) == 0) {
2427 atomic_inc(&rdev->nr_pending);
2428 if (sync_page_io(rdev, sect, s<<9,
2429 conf->tmppage, REQ_OP_READ, false))
2430 success = 1;
2431 rdev_dec_pending(rdev, mddev);
2432 if (success)
2433 break;
2434 }
2435
2436 d++;
2437 if (d == conf->raid_disks * 2)
2438 d = 0;
2439 } while (d != read_disk);
2440
2441 if (!success) {
2442 /* Cannot read from anywhere - mark it bad */
2443 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2444 if (!rdev_set_badblocks(rdev, sect, s, 0))
2445 md_error(mddev, rdev);
2446 break;
2447 }
2448 /* write it back and re-read */
2449 start = d;
2450 while (d != read_disk) {
2451 if (d==0)
2452 d = conf->raid_disks * 2;
2453 d--;
2454 rdev = conf->mirrors[d].rdev;
2455 if (rdev &&
2456 !test_bit(Faulty, &rdev->flags)) {
2457 atomic_inc(&rdev->nr_pending);
2458 r1_sync_page_io(rdev, sect, s,
2459 conf->tmppage, REQ_OP_WRITE);
2460 rdev_dec_pending(rdev, mddev);
2461 }
2462 }
2463 d = start;
2464 while (d != read_disk) {
2465 if (d==0)
2466 d = conf->raid_disks * 2;
2467 d--;
2468 rdev = conf->mirrors[d].rdev;
2469 if (rdev &&
2470 !test_bit(Faulty, &rdev->flags)) {
2471 atomic_inc(&rdev->nr_pending);
2472 if (r1_sync_page_io(rdev, sect, s,
2473 conf->tmppage, REQ_OP_READ)) {
2474 atomic_add(s, &rdev->corrected_errors);
2475 pr_info("md/raid1:%s: read error corrected (%d sectors at %llu on %pg)\n",
2476 mdname(mddev), s,
2477 (unsigned long long)(sect +
2478 rdev->data_offset),
2479 rdev->bdev);
2480 }
2481 rdev_dec_pending(rdev, mddev);
2482 }
2483 }
2484 sectors -= s;
2485 sect += s;
2486 }
2487}
2488
2489static int narrow_write_error(struct r1bio *r1_bio, int i)
2490{
2491 struct mddev *mddev = r1_bio->mddev;
2492 struct r1conf *conf = mddev->private;
2493 struct md_rdev *rdev = conf->mirrors[i].rdev;
2494
2495 /* bio has the data to be written to device 'i' where
2496 * we just recently had a write error.
2497 * We repeatedly clone the bio and trim down to one block,
2498 * then try the write. Where the write fails we record
2499 * a bad block.
2500 * It is conceivable that the bio doesn't exactly align with
2501 * blocks. We must handle this somehow.
2502 *
2503 * We currently own a reference on the rdev.
2504 */
2505
2506 int block_sectors;
2507 sector_t sector;
2508 int sectors;
2509 int sect_to_write = r1_bio->sectors;
2510 int ok = 1;
2511
2512 if (rdev->badblocks.shift < 0)
2513 return 0;
2514
2515 block_sectors = roundup(1 << rdev->badblocks.shift,
2516 bdev_logical_block_size(rdev->bdev) >> 9);
2517 sector = r1_bio->sector;
2518 sectors = ((sector + block_sectors)
2519 & ~(sector_t)(block_sectors - 1))
2520 - sector;
2521
2522 while (sect_to_write) {
2523 struct bio *wbio;
2524 if (sectors > sect_to_write)
2525 sectors = sect_to_write;
2526 /* Write at 'sector' for 'sectors'*/
2527
2528 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2529 wbio = bio_alloc_clone(rdev->bdev,
2530 r1_bio->behind_master_bio,
2531 GFP_NOIO, &mddev->bio_set);
2532 } else {
2533 wbio = bio_alloc_clone(rdev->bdev, r1_bio->master_bio,
2534 GFP_NOIO, &mddev->bio_set);
2535 }
2536
2537 wbio->bi_opf = REQ_OP_WRITE;
2538 wbio->bi_iter.bi_sector = r1_bio->sector;
2539 wbio->bi_iter.bi_size = r1_bio->sectors << 9;
2540
2541 bio_trim(wbio, sector - r1_bio->sector, sectors);
2542 wbio->bi_iter.bi_sector += rdev->data_offset;
2543
2544 if (submit_bio_wait(wbio) < 0)
2545 /* failure! */
2546 ok = rdev_set_badblocks(rdev, sector,
2547 sectors, 0)
2548 && ok;
2549
2550 bio_put(wbio);
2551 sect_to_write -= sectors;
2552 sector += sectors;
2553 sectors = block_sectors;
2554 }
2555 return ok;
2556}
2557
2558static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2559{
2560 int m;
2561 int s = r1_bio->sectors;
2562 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2563 struct md_rdev *rdev = conf->mirrors[m].rdev;
2564 struct bio *bio = r1_bio->bios[m];
2565 if (bio->bi_end_io == NULL)
2566 continue;
2567 if (!bio->bi_status &&
2568 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2569 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2570 }
2571 if (bio->bi_status &&
2572 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2573 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2574 md_error(conf->mddev, rdev);
2575 }
2576 }
2577 put_buf(r1_bio);
2578 md_done_sync(conf->mddev, s, 1);
2579}
2580
2581static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2582{
2583 int m, idx;
2584 bool fail = false;
2585
2586 for (m = 0; m < conf->raid_disks * 2 ; m++)
2587 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2588 struct md_rdev *rdev = conf->mirrors[m].rdev;
2589 rdev_clear_badblocks(rdev,
2590 r1_bio->sector,
2591 r1_bio->sectors, 0);
2592 rdev_dec_pending(rdev, conf->mddev);
2593 } else if (r1_bio->bios[m] != NULL) {
2594 /* This drive got a write error. We need to
2595 * narrow down and record precise write
2596 * errors.
2597 */
2598 fail = true;
2599 if (!narrow_write_error(r1_bio, m))
2600 md_error(conf->mddev,
2601 conf->mirrors[m].rdev);
2602 /* an I/O failed, we can't clear the bitmap */
2603 rdev_dec_pending(conf->mirrors[m].rdev,
2604 conf->mddev);
2605 }
2606 if (fail) {
2607 spin_lock_irq(&conf->device_lock);
2608 list_add(&r1_bio->retry_list, &conf->bio_end_io_list);
2609 idx = sector_to_idx(r1_bio->sector);
2610 atomic_inc(&conf->nr_queued[idx]);
2611 spin_unlock_irq(&conf->device_lock);
2612 /*
2613 * In case freeze_array() is waiting for condition
2614 * get_unqueued_pending() == extra to be true.
2615 */
2616 wake_up(&conf->wait_barrier);
2617 md_wakeup_thread(conf->mddev->thread);
2618 } else {
2619 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2620 close_write(r1_bio);
2621 raid_end_bio_io(r1_bio);
2622 }
2623}
2624
2625static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2626{
2627 struct mddev *mddev = conf->mddev;
2628 struct bio *bio;
2629 struct md_rdev *rdev;
2630 sector_t sector;
2631
2632 clear_bit(R1BIO_ReadError, &r1_bio->state);
2633 /* we got a read error. Maybe the drive is bad. Maybe just
2634 * the block and we can fix it.
2635 * We freeze all other IO, and try reading the block from
2636 * other devices. When we find one, we re-write
2637 * and check it that fixes the read error.
2638 * This is all done synchronously while the array is
2639 * frozen
2640 */
2641
2642 bio = r1_bio->bios[r1_bio->read_disk];
2643 bio_put(bio);
2644 r1_bio->bios[r1_bio->read_disk] = NULL;
2645
2646 rdev = conf->mirrors[r1_bio->read_disk].rdev;
2647 if (mddev->ro == 0
2648 && !test_bit(FailFast, &rdev->flags)) {
2649 freeze_array(conf, 1);
2650 fix_read_error(conf, r1_bio);
2651 unfreeze_array(conf);
2652 } else if (mddev->ro == 0 && test_bit(FailFast, &rdev->flags)) {
2653 md_error(mddev, rdev);
2654 } else {
2655 r1_bio->bios[r1_bio->read_disk] = IO_BLOCKED;
2656 }
2657
2658 rdev_dec_pending(rdev, conf->mddev);
2659 sector = r1_bio->sector;
2660 bio = r1_bio->master_bio;
2661
2662 /* Reuse the old r1_bio so that the IO_BLOCKED settings are preserved */
2663 r1_bio->state = 0;
2664 raid1_read_request(mddev, bio, r1_bio->sectors, r1_bio);
2665 allow_barrier(conf, sector);
2666}
2667
2668static void raid1d(struct md_thread *thread)
2669{
2670 struct mddev *mddev = thread->mddev;
2671 struct r1bio *r1_bio;
2672 unsigned long flags;
2673 struct r1conf *conf = mddev->private;
2674 struct list_head *head = &conf->retry_list;
2675 struct blk_plug plug;
2676 int idx;
2677
2678 md_check_recovery(mddev);
2679
2680 if (!list_empty_careful(&conf->bio_end_io_list) &&
2681 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2682 LIST_HEAD(tmp);
2683 spin_lock_irqsave(&conf->device_lock, flags);
2684 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags))
2685 list_splice_init(&conf->bio_end_io_list, &tmp);
2686 spin_unlock_irqrestore(&conf->device_lock, flags);
2687 while (!list_empty(&tmp)) {
2688 r1_bio = list_first_entry(&tmp, struct r1bio,
2689 retry_list);
2690 list_del(&r1_bio->retry_list);
2691 idx = sector_to_idx(r1_bio->sector);
2692 atomic_dec(&conf->nr_queued[idx]);
2693 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2694 close_write(r1_bio);
2695 raid_end_bio_io(r1_bio);
2696 }
2697 }
2698
2699 blk_start_plug(&plug);
2700 for (;;) {
2701
2702 flush_pending_writes(conf);
2703
2704 spin_lock_irqsave(&conf->device_lock, flags);
2705 if (list_empty(head)) {
2706 spin_unlock_irqrestore(&conf->device_lock, flags);
2707 break;
2708 }
2709 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2710 list_del(head->prev);
2711 idx = sector_to_idx(r1_bio->sector);
2712 atomic_dec(&conf->nr_queued[idx]);
2713 spin_unlock_irqrestore(&conf->device_lock, flags);
2714
2715 mddev = r1_bio->mddev;
2716 conf = mddev->private;
2717 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2718 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2719 test_bit(R1BIO_WriteError, &r1_bio->state))
2720 handle_sync_write_finished(conf, r1_bio);
2721 else
2722 sync_request_write(mddev, r1_bio);
2723 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2724 test_bit(R1BIO_WriteError, &r1_bio->state))
2725 handle_write_finished(conf, r1_bio);
2726 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2727 handle_read_error(conf, r1_bio);
2728 else
2729 WARN_ON_ONCE(1);
2730
2731 cond_resched();
2732 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2733 md_check_recovery(mddev);
2734 }
2735 blk_finish_plug(&plug);
2736}
2737
2738static int init_resync(struct r1conf *conf)
2739{
2740 int buffs;
2741
2742 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2743 BUG_ON(mempool_initialized(&conf->r1buf_pool));
2744
2745 return mempool_init(&conf->r1buf_pool, buffs, r1buf_pool_alloc,
2746 r1buf_pool_free, conf->poolinfo);
2747}
2748
2749static struct r1bio *raid1_alloc_init_r1buf(struct r1conf *conf)
2750{
2751 struct r1bio *r1bio = mempool_alloc(&conf->r1buf_pool, GFP_NOIO);
2752 struct resync_pages *rps;
2753 struct bio *bio;
2754 int i;
2755
2756 for (i = conf->poolinfo->raid_disks; i--; ) {
2757 bio = r1bio->bios[i];
2758 rps = bio->bi_private;
2759 bio_reset(bio, NULL, 0);
2760 bio->bi_private = rps;
2761 }
2762 r1bio->master_bio = NULL;
2763 return r1bio;
2764}
2765
2766/*
2767 * perform a "sync" on one "block"
2768 *
2769 * We need to make sure that no normal I/O request - particularly write
2770 * requests - conflict with active sync requests.
2771 *
2772 * This is achieved by tracking pending requests and a 'barrier' concept
2773 * that can be installed to exclude normal IO requests.
2774 */
2775
2776static sector_t raid1_sync_request(struct mddev *mddev, sector_t sector_nr,
2777 sector_t max_sector, int *skipped)
2778{
2779 struct r1conf *conf = mddev->private;
2780 struct r1bio *r1_bio;
2781 struct bio *bio;
2782 sector_t nr_sectors;
2783 int disk = -1;
2784 int i;
2785 int wonly = -1;
2786 int write_targets = 0, read_targets = 0;
2787 sector_t sync_blocks;
2788 bool still_degraded = false;
2789 int good_sectors = RESYNC_SECTORS;
2790 int min_bad = 0; /* number of sectors that are bad in all devices */
2791 int idx = sector_to_idx(sector_nr);
2792 int page_idx = 0;
2793
2794 if (!mempool_initialized(&conf->r1buf_pool))
2795 if (init_resync(conf))
2796 return 0;
2797
2798 if (sector_nr >= max_sector) {
2799 /* If we aborted, we need to abort the
2800 * sync on the 'current' bitmap chunk (there will
2801 * only be one in raid1 resync.
2802 * We can find the current addess in mddev->curr_resync
2803 */
2804 if (mddev->curr_resync < max_sector) /* aborted */
2805 mddev->bitmap_ops->end_sync(mddev, mddev->curr_resync,
2806 &sync_blocks);
2807 else /* completed sync */
2808 conf->fullsync = 0;
2809
2810 mddev->bitmap_ops->close_sync(mddev);
2811 close_sync(conf);
2812
2813 if (mddev_is_clustered(mddev)) {
2814 conf->cluster_sync_low = 0;
2815 conf->cluster_sync_high = 0;
2816 }
2817 return 0;
2818 }
2819
2820 if (mddev->bitmap == NULL &&
2821 mddev->recovery_cp == MaxSector &&
2822 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2823 conf->fullsync == 0) {
2824 *skipped = 1;
2825 return max_sector - sector_nr;
2826 }
2827 /* before building a request, check if we can skip these blocks..
2828 * This call the bitmap_start_sync doesn't actually record anything
2829 */
2830 if (!mddev->bitmap_ops->start_sync(mddev, sector_nr, &sync_blocks, true) &&
2831 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2832 /* We can skip this block, and probably several more */
2833 *skipped = 1;
2834 return sync_blocks;
2835 }
2836
2837 /*
2838 * If there is non-resync activity waiting for a turn, then let it
2839 * though before starting on this new sync request.
2840 */
2841 if (atomic_read(&conf->nr_waiting[idx]))
2842 schedule_timeout_uninterruptible(1);
2843
2844 /* we are incrementing sector_nr below. To be safe, we check against
2845 * sector_nr + two times RESYNC_SECTORS
2846 */
2847
2848 mddev->bitmap_ops->cond_end_sync(mddev, sector_nr,
2849 mddev_is_clustered(mddev) &&
2850 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
2851
2852 if (raise_barrier(conf, sector_nr))
2853 return 0;
2854
2855 r1_bio = raid1_alloc_init_r1buf(conf);
2856
2857 /*
2858 * If we get a correctably read error during resync or recovery,
2859 * we might want to read from a different device. So we
2860 * flag all drives that could conceivably be read from for READ,
2861 * and any others (which will be non-In_sync devices) for WRITE.
2862 * If a read fails, we try reading from something else for which READ
2863 * is OK.
2864 */
2865
2866 r1_bio->mddev = mddev;
2867 r1_bio->sector = sector_nr;
2868 r1_bio->state = 0;
2869 set_bit(R1BIO_IsSync, &r1_bio->state);
2870 /* make sure good_sectors won't go across barrier unit boundary */
2871 good_sectors = align_to_barrier_unit_end(sector_nr, good_sectors);
2872
2873 for (i = 0; i < conf->raid_disks * 2; i++) {
2874 struct md_rdev *rdev;
2875 bio = r1_bio->bios[i];
2876
2877 rdev = conf->mirrors[i].rdev;
2878 if (rdev == NULL ||
2879 test_bit(Faulty, &rdev->flags)) {
2880 if (i < conf->raid_disks)
2881 still_degraded = true;
2882 } else if (!test_bit(In_sync, &rdev->flags)) {
2883 bio->bi_opf = REQ_OP_WRITE;
2884 bio->bi_end_io = end_sync_write;
2885 write_targets ++;
2886 } else {
2887 /* may need to read from here */
2888 sector_t first_bad = MaxSector;
2889 int bad_sectors;
2890
2891 if (is_badblock(rdev, sector_nr, good_sectors,
2892 &first_bad, &bad_sectors)) {
2893 if (first_bad > sector_nr)
2894 good_sectors = first_bad - sector_nr;
2895 else {
2896 bad_sectors -= (sector_nr - first_bad);
2897 if (min_bad == 0 ||
2898 min_bad > bad_sectors)
2899 min_bad = bad_sectors;
2900 }
2901 }
2902 if (sector_nr < first_bad) {
2903 if (test_bit(WriteMostly, &rdev->flags)) {
2904 if (wonly < 0)
2905 wonly = i;
2906 } else {
2907 if (disk < 0)
2908 disk = i;
2909 }
2910 bio->bi_opf = REQ_OP_READ;
2911 bio->bi_end_io = end_sync_read;
2912 read_targets++;
2913 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2914 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2915 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2916 /*
2917 * The device is suitable for reading (InSync),
2918 * but has bad block(s) here. Let's try to correct them,
2919 * if we are doing resync or repair. Otherwise, leave
2920 * this device alone for this sync request.
2921 */
2922 bio->bi_opf = REQ_OP_WRITE;
2923 bio->bi_end_io = end_sync_write;
2924 write_targets++;
2925 }
2926 }
2927 if (rdev && bio->bi_end_io) {
2928 atomic_inc(&rdev->nr_pending);
2929 bio->bi_iter.bi_sector = sector_nr + rdev->data_offset;
2930 bio_set_dev(bio, rdev->bdev);
2931 if (test_bit(FailFast, &rdev->flags))
2932 bio->bi_opf |= MD_FAILFAST;
2933 }
2934 }
2935 if (disk < 0)
2936 disk = wonly;
2937 r1_bio->read_disk = disk;
2938
2939 if (read_targets == 0 && min_bad > 0) {
2940 /* These sectors are bad on all InSync devices, so we
2941 * need to mark them bad on all write targets
2942 */
2943 int ok = 1;
2944 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2945 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2946 struct md_rdev *rdev = conf->mirrors[i].rdev;
2947 ok = rdev_set_badblocks(rdev, sector_nr,
2948 min_bad, 0
2949 ) && ok;
2950 }
2951 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2952 *skipped = 1;
2953 put_buf(r1_bio);
2954
2955 if (!ok) {
2956 /* Cannot record the badblocks, so need to
2957 * abort the resync.
2958 * If there are multiple read targets, could just
2959 * fail the really bad ones ???
2960 */
2961 conf->recovery_disabled = mddev->recovery_disabled;
2962 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2963 return 0;
2964 } else
2965 return min_bad;
2966
2967 }
2968 if (min_bad > 0 && min_bad < good_sectors) {
2969 /* only resync enough to reach the next bad->good
2970 * transition */
2971 good_sectors = min_bad;
2972 }
2973
2974 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2975 /* extra read targets are also write targets */
2976 write_targets += read_targets-1;
2977
2978 if (write_targets == 0 || read_targets == 0) {
2979 /* There is nowhere to write, so all non-sync
2980 * drives must be failed - so we are finished
2981 */
2982 sector_t rv;
2983 if (min_bad > 0)
2984 max_sector = sector_nr + min_bad;
2985 rv = max_sector - sector_nr;
2986 *skipped = 1;
2987 put_buf(r1_bio);
2988 return rv;
2989 }
2990
2991 if (max_sector > mddev->resync_max)
2992 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2993 if (max_sector > sector_nr + good_sectors)
2994 max_sector = sector_nr + good_sectors;
2995 nr_sectors = 0;
2996 sync_blocks = 0;
2997 do {
2998 struct page *page;
2999 int len = PAGE_SIZE;
3000 if (sector_nr + (len>>9) > max_sector)
3001 len = (max_sector - sector_nr) << 9;
3002 if (len == 0)
3003 break;
3004 if (sync_blocks == 0) {
3005 if (!mddev->bitmap_ops->start_sync(mddev, sector_nr,
3006 &sync_blocks, still_degraded) &&
3007 !conf->fullsync &&
3008 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
3009 break;
3010 if ((len >> 9) > sync_blocks)
3011 len = sync_blocks<<9;
3012 }
3013
3014 for (i = 0 ; i < conf->raid_disks * 2; i++) {
3015 struct resync_pages *rp;
3016
3017 bio = r1_bio->bios[i];
3018 rp = get_resync_pages(bio);
3019 if (bio->bi_end_io) {
3020 page = resync_fetch_page(rp, page_idx);
3021
3022 /*
3023 * won't fail because the vec table is big
3024 * enough to hold all these pages
3025 */
3026 __bio_add_page(bio, page, len, 0);
3027 }
3028 }
3029 nr_sectors += len>>9;
3030 sector_nr += len>>9;
3031 sync_blocks -= (len>>9);
3032 } while (++page_idx < RESYNC_PAGES);
3033
3034 r1_bio->sectors = nr_sectors;
3035
3036 if (mddev_is_clustered(mddev) &&
3037 conf->cluster_sync_high < sector_nr + nr_sectors) {
3038 conf->cluster_sync_low = mddev->curr_resync_completed;
3039 conf->cluster_sync_high = conf->cluster_sync_low + CLUSTER_RESYNC_WINDOW_SECTORS;
3040 /* Send resync message */
3041 md_cluster_ops->resync_info_update(mddev,
3042 conf->cluster_sync_low,
3043 conf->cluster_sync_high);
3044 }
3045
3046 /* For a user-requested sync, we read all readable devices and do a
3047 * compare
3048 */
3049 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
3050 atomic_set(&r1_bio->remaining, read_targets);
3051 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
3052 bio = r1_bio->bios[i];
3053 if (bio->bi_end_io == end_sync_read) {
3054 read_targets--;
3055 md_sync_acct_bio(bio, nr_sectors);
3056 if (read_targets == 1)
3057 bio->bi_opf &= ~MD_FAILFAST;
3058 submit_bio_noacct(bio);
3059 }
3060 }
3061 } else {
3062 atomic_set(&r1_bio->remaining, 1);
3063 bio = r1_bio->bios[r1_bio->read_disk];
3064 md_sync_acct_bio(bio, nr_sectors);
3065 if (read_targets == 1)
3066 bio->bi_opf &= ~MD_FAILFAST;
3067 submit_bio_noacct(bio);
3068 }
3069 return nr_sectors;
3070}
3071
3072static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3073{
3074 if (sectors)
3075 return sectors;
3076
3077 return mddev->dev_sectors;
3078}
3079
3080static struct r1conf *setup_conf(struct mddev *mddev)
3081{
3082 struct r1conf *conf;
3083 int i;
3084 struct raid1_info *disk;
3085 struct md_rdev *rdev;
3086 int err = -ENOMEM;
3087
3088 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
3089 if (!conf)
3090 goto abort;
3091
3092 conf->nr_pending = kcalloc(BARRIER_BUCKETS_NR,
3093 sizeof(atomic_t), GFP_KERNEL);
3094 if (!conf->nr_pending)
3095 goto abort;
3096
3097 conf->nr_waiting = kcalloc(BARRIER_BUCKETS_NR,
3098 sizeof(atomic_t), GFP_KERNEL);
3099 if (!conf->nr_waiting)
3100 goto abort;
3101
3102 conf->nr_queued = kcalloc(BARRIER_BUCKETS_NR,
3103 sizeof(atomic_t), GFP_KERNEL);
3104 if (!conf->nr_queued)
3105 goto abort;
3106
3107 conf->barrier = kcalloc(BARRIER_BUCKETS_NR,
3108 sizeof(atomic_t), GFP_KERNEL);
3109 if (!conf->barrier)
3110 goto abort;
3111
3112 conf->mirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3113 mddev->raid_disks, 2),
3114 GFP_KERNEL);
3115 if (!conf->mirrors)
3116 goto abort;
3117
3118 conf->tmppage = alloc_page(GFP_KERNEL);
3119 if (!conf->tmppage)
3120 goto abort;
3121
3122 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
3123 if (!conf->poolinfo)
3124 goto abort;
3125 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
3126 err = mempool_init(&conf->r1bio_pool, NR_RAID_BIOS, r1bio_pool_alloc,
3127 rbio_pool_free, conf->poolinfo);
3128 if (err)
3129 goto abort;
3130
3131 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3132 if (err)
3133 goto abort;
3134
3135 conf->poolinfo->mddev = mddev;
3136
3137 err = -EINVAL;
3138 spin_lock_init(&conf->device_lock);
3139 conf->raid_disks = mddev->raid_disks;
3140 rdev_for_each(rdev, mddev) {
3141 int disk_idx = rdev->raid_disk;
3142
3143 if (disk_idx >= conf->raid_disks || disk_idx < 0)
3144 continue;
3145
3146 if (!raid1_add_conf(conf, rdev, disk_idx,
3147 test_bit(Replacement, &rdev->flags)))
3148 goto abort;
3149 }
3150 conf->mddev = mddev;
3151 INIT_LIST_HEAD(&conf->retry_list);
3152 INIT_LIST_HEAD(&conf->bio_end_io_list);
3153
3154 spin_lock_init(&conf->resync_lock);
3155 init_waitqueue_head(&conf->wait_barrier);
3156
3157 bio_list_init(&conf->pending_bio_list);
3158 conf->recovery_disabled = mddev->recovery_disabled - 1;
3159
3160 err = -EIO;
3161 for (i = 0; i < conf->raid_disks * 2; i++) {
3162
3163 disk = conf->mirrors + i;
3164
3165 if (i < conf->raid_disks &&
3166 disk[conf->raid_disks].rdev) {
3167 /* This slot has a replacement. */
3168 if (!disk->rdev) {
3169 /* No original, just make the replacement
3170 * a recovering spare
3171 */
3172 disk->rdev =
3173 disk[conf->raid_disks].rdev;
3174 disk[conf->raid_disks].rdev = NULL;
3175 } else if (!test_bit(In_sync, &disk->rdev->flags))
3176 /* Original is not in_sync - bad */
3177 goto abort;
3178 }
3179
3180 if (!disk->rdev ||
3181 !test_bit(In_sync, &disk->rdev->flags)) {
3182 disk->head_position = 0;
3183 if (disk->rdev &&
3184 (disk->rdev->saved_raid_disk < 0))
3185 conf->fullsync = 1;
3186 }
3187 }
3188
3189 err = -ENOMEM;
3190 rcu_assign_pointer(conf->thread,
3191 md_register_thread(raid1d, mddev, "raid1"));
3192 if (!conf->thread)
3193 goto abort;
3194
3195 return conf;
3196
3197 abort:
3198 if (conf) {
3199 mempool_exit(&conf->r1bio_pool);
3200 kfree(conf->mirrors);
3201 safe_put_page(conf->tmppage);
3202 kfree(conf->poolinfo);
3203 kfree(conf->nr_pending);
3204 kfree(conf->nr_waiting);
3205 kfree(conf->nr_queued);
3206 kfree(conf->barrier);
3207 bioset_exit(&conf->bio_split);
3208 kfree(conf);
3209 }
3210 return ERR_PTR(err);
3211}
3212
3213static int raid1_set_limits(struct mddev *mddev)
3214{
3215 struct queue_limits lim;
3216 int err;
3217
3218 md_init_stacking_limits(&lim);
3219 lim.max_write_zeroes_sectors = 0;
3220 lim.features |= BLK_FEAT_ATOMIC_WRITES_STACKED;
3221 err = mddev_stack_rdev_limits(mddev, &lim, MDDEV_STACK_INTEGRITY);
3222 if (err)
3223 return err;
3224 return queue_limits_set(mddev->gendisk->queue, &lim);
3225}
3226
3227static int raid1_run(struct mddev *mddev)
3228{
3229 struct r1conf *conf;
3230 int i;
3231 int ret;
3232
3233 if (mddev->level != 1) {
3234 pr_warn("md/raid1:%s: raid level not set to mirroring (%d)\n",
3235 mdname(mddev), mddev->level);
3236 return -EIO;
3237 }
3238 if (mddev->reshape_position != MaxSector) {
3239 pr_warn("md/raid1:%s: reshape_position set but not supported\n",
3240 mdname(mddev));
3241 return -EIO;
3242 }
3243
3244 /*
3245 * copy the already verified devices into our private RAID1
3246 * bookkeeping area. [whatever we allocate in run(),
3247 * should be freed in raid1_free()]
3248 */
3249 if (mddev->private == NULL)
3250 conf = setup_conf(mddev);
3251 else
3252 conf = mddev->private;
3253
3254 if (IS_ERR(conf))
3255 return PTR_ERR(conf);
3256
3257 if (!mddev_is_dm(mddev)) {
3258 ret = raid1_set_limits(mddev);
3259 if (ret)
3260 return ret;
3261 }
3262
3263 mddev->degraded = 0;
3264 for (i = 0; i < conf->raid_disks; i++)
3265 if (conf->mirrors[i].rdev == NULL ||
3266 !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
3267 test_bit(Faulty, &conf->mirrors[i].rdev->flags))
3268 mddev->degraded++;
3269 /*
3270 * RAID1 needs at least one disk in active
3271 */
3272 if (conf->raid_disks - mddev->degraded < 1) {
3273 md_unregister_thread(mddev, &conf->thread);
3274 return -EINVAL;
3275 }
3276
3277 if (conf->raid_disks - mddev->degraded == 1)
3278 mddev->recovery_cp = MaxSector;
3279
3280 if (mddev->recovery_cp != MaxSector)
3281 pr_info("md/raid1:%s: not clean -- starting background reconstruction\n",
3282 mdname(mddev));
3283 pr_info("md/raid1:%s: active with %d out of %d mirrors\n",
3284 mdname(mddev), mddev->raid_disks - mddev->degraded,
3285 mddev->raid_disks);
3286
3287 /*
3288 * Ok, everything is just fine now
3289 */
3290 rcu_assign_pointer(mddev->thread, conf->thread);
3291 rcu_assign_pointer(conf->thread, NULL);
3292 mddev->private = conf;
3293 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3294
3295 md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
3296
3297 ret = md_integrity_register(mddev);
3298 if (ret)
3299 md_unregister_thread(mddev, &mddev->thread);
3300 return ret;
3301}
3302
3303static void raid1_free(struct mddev *mddev, void *priv)
3304{
3305 struct r1conf *conf = priv;
3306
3307 mempool_exit(&conf->r1bio_pool);
3308 kfree(conf->mirrors);
3309 safe_put_page(conf->tmppage);
3310 kfree(conf->poolinfo);
3311 kfree(conf->nr_pending);
3312 kfree(conf->nr_waiting);
3313 kfree(conf->nr_queued);
3314 kfree(conf->barrier);
3315 bioset_exit(&conf->bio_split);
3316 kfree(conf);
3317}
3318
3319static int raid1_resize(struct mddev *mddev, sector_t sectors)
3320{
3321 /* no resync is happening, and there is enough space
3322 * on all devices, so we can resize.
3323 * We need to make sure resync covers any new space.
3324 * If the array is shrinking we should possibly wait until
3325 * any io in the removed space completes, but it hardly seems
3326 * worth it.
3327 */
3328 sector_t newsize = raid1_size(mddev, sectors, 0);
3329 int ret;
3330
3331 if (mddev->external_size &&
3332 mddev->array_sectors > newsize)
3333 return -EINVAL;
3334
3335 ret = mddev->bitmap_ops->resize(mddev, newsize, 0, false);
3336 if (ret)
3337 return ret;
3338
3339 md_set_array_sectors(mddev, newsize);
3340 if (sectors > mddev->dev_sectors &&
3341 mddev->recovery_cp > mddev->dev_sectors) {
3342 mddev->recovery_cp = mddev->dev_sectors;
3343 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3344 }
3345 mddev->dev_sectors = sectors;
3346 mddev->resync_max_sectors = sectors;
3347 return 0;
3348}
3349
3350static int raid1_reshape(struct mddev *mddev)
3351{
3352 /* We need to:
3353 * 1/ resize the r1bio_pool
3354 * 2/ resize conf->mirrors
3355 *
3356 * We allocate a new r1bio_pool if we can.
3357 * Then raise a device barrier and wait until all IO stops.
3358 * Then resize conf->mirrors and swap in the new r1bio pool.
3359 *
3360 * At the same time, we "pack" the devices so that all the missing
3361 * devices have the higher raid_disk numbers.
3362 */
3363 mempool_t newpool, oldpool;
3364 struct pool_info *newpoolinfo;
3365 struct raid1_info *newmirrors;
3366 struct r1conf *conf = mddev->private;
3367 int cnt, raid_disks;
3368 unsigned long flags;
3369 int d, d2;
3370 int ret;
3371
3372 memset(&newpool, 0, sizeof(newpool));
3373 memset(&oldpool, 0, sizeof(oldpool));
3374
3375 /* Cannot change chunk_size, layout, or level */
3376 if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
3377 mddev->layout != mddev->new_layout ||
3378 mddev->level != mddev->new_level) {
3379 mddev->new_chunk_sectors = mddev->chunk_sectors;
3380 mddev->new_layout = mddev->layout;
3381 mddev->new_level = mddev->level;
3382 return -EINVAL;
3383 }
3384
3385 if (!mddev_is_clustered(mddev))
3386 md_allow_write(mddev);
3387
3388 raid_disks = mddev->raid_disks + mddev->delta_disks;
3389
3390 if (raid_disks < conf->raid_disks) {
3391 cnt=0;
3392 for (d= 0; d < conf->raid_disks; d++)
3393 if (conf->mirrors[d].rdev)
3394 cnt++;
3395 if (cnt > raid_disks)
3396 return -EBUSY;
3397 }
3398
3399 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
3400 if (!newpoolinfo)
3401 return -ENOMEM;
3402 newpoolinfo->mddev = mddev;
3403 newpoolinfo->raid_disks = raid_disks * 2;
3404
3405 ret = mempool_init(&newpool, NR_RAID_BIOS, r1bio_pool_alloc,
3406 rbio_pool_free, newpoolinfo);
3407 if (ret) {
3408 kfree(newpoolinfo);
3409 return ret;
3410 }
3411 newmirrors = kzalloc(array3_size(sizeof(struct raid1_info),
3412 raid_disks, 2),
3413 GFP_KERNEL);
3414 if (!newmirrors) {
3415 kfree(newpoolinfo);
3416 mempool_exit(&newpool);
3417 return -ENOMEM;
3418 }
3419
3420 freeze_array(conf, 0);
3421
3422 /* ok, everything is stopped */
3423 oldpool = conf->r1bio_pool;
3424 conf->r1bio_pool = newpool;
3425
3426 for (d = d2 = 0; d < conf->raid_disks; d++) {
3427 struct md_rdev *rdev = conf->mirrors[d].rdev;
3428 if (rdev && rdev->raid_disk != d2) {
3429 sysfs_unlink_rdev(mddev, rdev);
3430 rdev->raid_disk = d2;
3431 sysfs_unlink_rdev(mddev, rdev);
3432 if (sysfs_link_rdev(mddev, rdev))
3433 pr_warn("md/raid1:%s: cannot register rd%d\n",
3434 mdname(mddev), rdev->raid_disk);
3435 }
3436 if (rdev)
3437 newmirrors[d2++].rdev = rdev;
3438 }
3439 kfree(conf->mirrors);
3440 conf->mirrors = newmirrors;
3441 kfree(conf->poolinfo);
3442 conf->poolinfo = newpoolinfo;
3443
3444 spin_lock_irqsave(&conf->device_lock, flags);
3445 mddev->degraded += (raid_disks - conf->raid_disks);
3446 spin_unlock_irqrestore(&conf->device_lock, flags);
3447 conf->raid_disks = mddev->raid_disks = raid_disks;
3448 mddev->delta_disks = 0;
3449
3450 unfreeze_array(conf);
3451
3452 set_bit(MD_RECOVERY_RECOVER, &mddev->recovery);
3453 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3454 md_wakeup_thread(mddev->thread);
3455
3456 mempool_exit(&oldpool);
3457 return 0;
3458}
3459
3460static void raid1_quiesce(struct mddev *mddev, int quiesce)
3461{
3462 struct r1conf *conf = mddev->private;
3463
3464 if (quiesce)
3465 freeze_array(conf, 0);
3466 else
3467 unfreeze_array(conf);
3468}
3469
3470static void *raid1_takeover(struct mddev *mddev)
3471{
3472 /* raid1 can take over:
3473 * raid5 with 2 devices, any layout or chunk size
3474 */
3475 if (mddev->level == 5 && mddev->raid_disks == 2) {
3476 struct r1conf *conf;
3477 mddev->new_level = 1;
3478 mddev->new_layout = 0;
3479 mddev->new_chunk_sectors = 0;
3480 conf = setup_conf(mddev);
3481 if (!IS_ERR(conf)) {
3482 /* Array must appear to be quiesced */
3483 conf->array_frozen = 1;
3484 mddev_clear_unsupported_flags(mddev,
3485 UNSUPPORTED_MDDEV_FLAGS);
3486 }
3487 return conf;
3488 }
3489 return ERR_PTR(-EINVAL);
3490}
3491
3492static struct md_personality raid1_personality =
3493{
3494 .name = "raid1",
3495 .level = 1,
3496 .owner = THIS_MODULE,
3497 .make_request = raid1_make_request,
3498 .run = raid1_run,
3499 .free = raid1_free,
3500 .status = raid1_status,
3501 .error_handler = raid1_error,
3502 .hot_add_disk = raid1_add_disk,
3503 .hot_remove_disk= raid1_remove_disk,
3504 .spare_active = raid1_spare_active,
3505 .sync_request = raid1_sync_request,
3506 .resize = raid1_resize,
3507 .size = raid1_size,
3508 .check_reshape = raid1_reshape,
3509 .quiesce = raid1_quiesce,
3510 .takeover = raid1_takeover,
3511};
3512
3513static int __init raid_init(void)
3514{
3515 return register_md_personality(&raid1_personality);
3516}
3517
3518static void raid_exit(void)
3519{
3520 unregister_md_personality(&raid1_personality);
3521}
3522
3523module_init(raid_init);
3524module_exit(raid_exit);
3525MODULE_LICENSE("GPL");
3526MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3527MODULE_ALIAS("md-personality-3"); /* RAID1 */
3528MODULE_ALIAS("md-raid1");
3529MODULE_ALIAS("md-level-1");