Merge tag 'v3.10.55' into update
[GitHub/mt8127/android_kernel_alcatel_ttab.git] / drivers / md / raid1.c
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 Ƙstergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14 *
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
17 *
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
20 *
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
23 *
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
27 * any later version.
28 *
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32 */
33
34 #include <linux/slab.h>
35 #include <linux/delay.h>
36 #include <linux/blkdev.h>
37 #include <linux/module.h>
38 #include <linux/seq_file.h>
39 #include <linux/ratelimit.h>
40 #include "md.h"
41 #include "raid1.h"
42 #include "bitmap.h"
43
44 /*
45 * Number of guaranteed r1bios in case of extreme VM load:
46 */
47 #define NR_RAID1_BIOS 256
48
49 /* when we get a read error on a read-only array, we redirect to another
50 * device without failing the first device, or trying to over-write to
51 * correct the read error. To keep track of bad blocks on a per-bio
52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
53 */
54 #define IO_BLOCKED ((struct bio *)1)
55 /* When we successfully write to a known bad-block, we need to remove the
56 * bad-block marking which must be done from process context. So we record
57 * the success by setting devs[n].bio to IO_MADE_GOOD
58 */
59 #define IO_MADE_GOOD ((struct bio *)2)
60
61 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
62
63 /* When there are this many requests queue to be written by
64 * the raid1 thread, we become 'congested' to provide back-pressure
65 * for writeback.
66 */
67 static int max_queued_requests = 1024;
68
69 static void allow_barrier(struct r1conf *conf);
70 static void lower_barrier(struct r1conf *conf);
71
72 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
73 {
74 struct pool_info *pi = data;
75 int size = offsetof(struct r1bio, bios[pi->raid_disks]);
76
77 /* allocate a r1bio with room for raid_disks entries in the bios array */
78 return kzalloc(size, gfp_flags);
79 }
80
81 static void r1bio_pool_free(void *r1_bio, void *data)
82 {
83 kfree(r1_bio);
84 }
85
86 #define RESYNC_BLOCK_SIZE (64*1024)
87 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
88 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
89 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
90 #define RESYNC_WINDOW (2048*1024)
91
92 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
93 {
94 struct pool_info *pi = data;
95 struct r1bio *r1_bio;
96 struct bio *bio;
97 int need_pages;
98 int i, j;
99
100 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
101 if (!r1_bio)
102 return NULL;
103
104 /*
105 * Allocate bios : 1 for reading, n-1 for writing
106 */
107 for (j = pi->raid_disks ; j-- ; ) {
108 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
109 if (!bio)
110 goto out_free_bio;
111 r1_bio->bios[j] = bio;
112 }
113 /*
114 * Allocate RESYNC_PAGES data pages and attach them to
115 * the first bio.
116 * If this is a user-requested check/repair, allocate
117 * RESYNC_PAGES for each bio.
118 */
119 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
120 need_pages = pi->raid_disks;
121 else
122 need_pages = 1;
123 for (j = 0; j < need_pages; j++) {
124 bio = r1_bio->bios[j];
125 bio->bi_vcnt = RESYNC_PAGES;
126
127 if (bio_alloc_pages(bio, gfp_flags))
128 goto out_free_pages;
129 }
130 /* If not user-requests, copy the page pointers to all bios */
131 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
132 for (i=0; i<RESYNC_PAGES ; i++)
133 for (j=1; j<pi->raid_disks; j++)
134 r1_bio->bios[j]->bi_io_vec[i].bv_page =
135 r1_bio->bios[0]->bi_io_vec[i].bv_page;
136 }
137
138 r1_bio->master_bio = NULL;
139
140 return r1_bio;
141
142 out_free_pages:
143 while (--j >= 0) {
144 struct bio_vec *bv;
145
146 bio_for_each_segment_all(bv, r1_bio->bios[j], i)
147 __free_page(bv->bv_page);
148 }
149
150 out_free_bio:
151 while (++j < pi->raid_disks)
152 bio_put(r1_bio->bios[j]);
153 r1bio_pool_free(r1_bio, data);
154 return NULL;
155 }
156
157 static void r1buf_pool_free(void *__r1_bio, void *data)
158 {
159 struct pool_info *pi = data;
160 int i,j;
161 struct r1bio *r1bio = __r1_bio;
162
163 for (i = 0; i < RESYNC_PAGES; i++)
164 for (j = pi->raid_disks; j-- ;) {
165 if (j == 0 ||
166 r1bio->bios[j]->bi_io_vec[i].bv_page !=
167 r1bio->bios[0]->bi_io_vec[i].bv_page)
168 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
169 }
170 for (i=0 ; i < pi->raid_disks; i++)
171 bio_put(r1bio->bios[i]);
172
173 r1bio_pool_free(r1bio, data);
174 }
175
176 static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
177 {
178 int i;
179
180 for (i = 0; i < conf->raid_disks * 2; i++) {
181 struct bio **bio = r1_bio->bios + i;
182 if (!BIO_SPECIAL(*bio))
183 bio_put(*bio);
184 *bio = NULL;
185 }
186 }
187
188 static void free_r1bio(struct r1bio *r1_bio)
189 {
190 struct r1conf *conf = r1_bio->mddev->private;
191
192 put_all_bios(conf, r1_bio);
193 mempool_free(r1_bio, conf->r1bio_pool);
194 }
195
196 static void put_buf(struct r1bio *r1_bio)
197 {
198 struct r1conf *conf = r1_bio->mddev->private;
199 int i;
200
201 for (i = 0; i < conf->raid_disks * 2; i++) {
202 struct bio *bio = r1_bio->bios[i];
203 if (bio->bi_end_io)
204 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
205 }
206
207 mempool_free(r1_bio, conf->r1buf_pool);
208
209 lower_barrier(conf);
210 }
211
212 static void reschedule_retry(struct r1bio *r1_bio)
213 {
214 unsigned long flags;
215 struct mddev *mddev = r1_bio->mddev;
216 struct r1conf *conf = mddev->private;
217
218 spin_lock_irqsave(&conf->device_lock, flags);
219 list_add(&r1_bio->retry_list, &conf->retry_list);
220 conf->nr_queued ++;
221 spin_unlock_irqrestore(&conf->device_lock, flags);
222
223 wake_up(&conf->wait_barrier);
224 md_wakeup_thread(mddev->thread);
225 }
226
227 /*
228 * raid_end_bio_io() is called when we have finished servicing a mirrored
229 * operation and are ready to return a success/failure code to the buffer
230 * cache layer.
231 */
232 static void call_bio_endio(struct r1bio *r1_bio)
233 {
234 struct bio *bio = r1_bio->master_bio;
235 int done;
236 struct r1conf *conf = r1_bio->mddev->private;
237
238 if (bio->bi_phys_segments) {
239 unsigned long flags;
240 spin_lock_irqsave(&conf->device_lock, flags);
241 bio->bi_phys_segments--;
242 done = (bio->bi_phys_segments == 0);
243 spin_unlock_irqrestore(&conf->device_lock, flags);
244 } else
245 done = 1;
246
247 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
248 clear_bit(BIO_UPTODATE, &bio->bi_flags);
249 if (done) {
250 bio_endio(bio, 0);
251 /*
252 * Wake up any possible resync thread that waits for the device
253 * to go idle.
254 */
255 allow_barrier(conf);
256 }
257 }
258
259 static void raid_end_bio_io(struct r1bio *r1_bio)
260 {
261 struct bio *bio = r1_bio->master_bio;
262
263 /* if nobody has done the final endio yet, do it now */
264 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
265 pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
266 (bio_data_dir(bio) == WRITE) ? "write" : "read",
267 (unsigned long long) bio->bi_sector,
268 (unsigned long long) bio->bi_sector +
269 bio_sectors(bio) - 1);
270
271 call_bio_endio(r1_bio);
272 }
273 free_r1bio(r1_bio);
274 }
275
276 /*
277 * Update disk head position estimator based on IRQ completion info.
278 */
279 static inline void update_head_pos(int disk, struct r1bio *r1_bio)
280 {
281 struct r1conf *conf = r1_bio->mddev->private;
282
283 conf->mirrors[disk].head_position =
284 r1_bio->sector + (r1_bio->sectors);
285 }
286
287 /*
288 * Find the disk number which triggered given bio
289 */
290 static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
291 {
292 int mirror;
293 struct r1conf *conf = r1_bio->mddev->private;
294 int raid_disks = conf->raid_disks;
295
296 for (mirror = 0; mirror < raid_disks * 2; mirror++)
297 if (r1_bio->bios[mirror] == bio)
298 break;
299
300 BUG_ON(mirror == raid_disks * 2);
301 update_head_pos(mirror, r1_bio);
302
303 return mirror;
304 }
305
306 static void raid1_end_read_request(struct bio *bio, int error)
307 {
308 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
309 struct r1bio *r1_bio = bio->bi_private;
310 int mirror;
311 struct r1conf *conf = r1_bio->mddev->private;
312
313 mirror = r1_bio->read_disk;
314 /*
315 * this branch is our 'one mirror IO has finished' event handler:
316 */
317 update_head_pos(mirror, r1_bio);
318
319 if (uptodate)
320 set_bit(R1BIO_Uptodate, &r1_bio->state);
321 else {
322 /* If all other devices have failed, we want to return
323 * the error upwards rather than fail the last device.
324 * Here we redefine "uptodate" to mean "Don't want to retry"
325 */
326 unsigned long flags;
327 spin_lock_irqsave(&conf->device_lock, flags);
328 if (r1_bio->mddev->degraded == conf->raid_disks ||
329 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
330 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
331 uptodate = 1;
332 spin_unlock_irqrestore(&conf->device_lock, flags);
333 }
334
335 if (uptodate) {
336 raid_end_bio_io(r1_bio);
337 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
338 } else {
339 /*
340 * oops, read error:
341 */
342 char b[BDEVNAME_SIZE];
343 printk_ratelimited(
344 KERN_ERR "md/raid1:%s: %s: "
345 "rescheduling sector %llu\n",
346 mdname(conf->mddev),
347 bdevname(conf->mirrors[mirror].rdev->bdev,
348 b),
349 (unsigned long long)r1_bio->sector);
350 set_bit(R1BIO_ReadError, &r1_bio->state);
351 reschedule_retry(r1_bio);
352 /* don't drop the reference on read_disk yet */
353 }
354 }
355
356 static void close_write(struct r1bio *r1_bio)
357 {
358 /* it really is the end of this request */
359 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
360 /* free extra copy of the data pages */
361 int i = r1_bio->behind_page_count;
362 while (i--)
363 safe_put_page(r1_bio->behind_bvecs[i].bv_page);
364 kfree(r1_bio->behind_bvecs);
365 r1_bio->behind_bvecs = NULL;
366 }
367 /* clear the bitmap if all writes complete successfully */
368 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
369 r1_bio->sectors,
370 !test_bit(R1BIO_Degraded, &r1_bio->state),
371 test_bit(R1BIO_BehindIO, &r1_bio->state));
372 md_write_end(r1_bio->mddev);
373 }
374
375 static void r1_bio_write_done(struct r1bio *r1_bio)
376 {
377 if (!atomic_dec_and_test(&r1_bio->remaining))
378 return;
379
380 if (test_bit(R1BIO_WriteError, &r1_bio->state))
381 reschedule_retry(r1_bio);
382 else {
383 close_write(r1_bio);
384 if (test_bit(R1BIO_MadeGood, &r1_bio->state))
385 reschedule_retry(r1_bio);
386 else
387 raid_end_bio_io(r1_bio);
388 }
389 }
390
391 static void raid1_end_write_request(struct bio *bio, int error)
392 {
393 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
394 struct r1bio *r1_bio = bio->bi_private;
395 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
396 struct r1conf *conf = r1_bio->mddev->private;
397 struct bio *to_put = NULL;
398
399 mirror = find_bio_disk(r1_bio, bio);
400
401 /*
402 * 'one mirror IO has finished' event handler:
403 */
404 if (!uptodate) {
405 set_bit(WriteErrorSeen,
406 &conf->mirrors[mirror].rdev->flags);
407 if (!test_and_set_bit(WantReplacement,
408 &conf->mirrors[mirror].rdev->flags))
409 set_bit(MD_RECOVERY_NEEDED, &
410 conf->mddev->recovery);
411
412 set_bit(R1BIO_WriteError, &r1_bio->state);
413 } else {
414 /*
415 * Set R1BIO_Uptodate in our master bio, so that we
416 * will return a good error code for to the higher
417 * levels even if IO on some other mirrored buffer
418 * fails.
419 *
420 * The 'master' represents the composite IO operation
421 * to user-side. So if something waits for IO, then it
422 * will wait for the 'master' bio.
423 */
424 sector_t first_bad;
425 int bad_sectors;
426
427 r1_bio->bios[mirror] = NULL;
428 to_put = bio;
429 /*
430 * Do not set R1BIO_Uptodate if the current device is
431 * rebuilding or Faulty. This is because we cannot use
432 * such device for properly reading the data back (we could
433 * potentially use it, if the current write would have felt
434 * before rdev->recovery_offset, but for simplicity we don't
435 * check this here.
436 */
437 if (test_bit(In_sync, &conf->mirrors[mirror].rdev->flags) &&
438 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags))
439 set_bit(R1BIO_Uptodate, &r1_bio->state);
440
441 /* Maybe we can clear some bad blocks. */
442 if (is_badblock(conf->mirrors[mirror].rdev,
443 r1_bio->sector, r1_bio->sectors,
444 &first_bad, &bad_sectors)) {
445 r1_bio->bios[mirror] = IO_MADE_GOOD;
446 set_bit(R1BIO_MadeGood, &r1_bio->state);
447 }
448 }
449
450 if (behind) {
451 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
452 atomic_dec(&r1_bio->behind_remaining);
453
454 /*
455 * In behind mode, we ACK the master bio once the I/O
456 * has safely reached all non-writemostly
457 * disks. Setting the Returned bit ensures that this
458 * gets done only once -- we don't ever want to return
459 * -EIO here, instead we'll wait
460 */
461 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
462 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
463 /* Maybe we can return now */
464 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
465 struct bio *mbio = r1_bio->master_bio;
466 pr_debug("raid1: behind end write sectors"
467 " %llu-%llu\n",
468 (unsigned long long) mbio->bi_sector,
469 (unsigned long long) mbio->bi_sector +
470 bio_sectors(mbio) - 1);
471 call_bio_endio(r1_bio);
472 }
473 }
474 }
475 if (r1_bio->bios[mirror] == NULL)
476 rdev_dec_pending(conf->mirrors[mirror].rdev,
477 conf->mddev);
478
479 /*
480 * Let's see if all mirrored write operations have finished
481 * already.
482 */
483 r1_bio_write_done(r1_bio);
484
485 if (to_put)
486 bio_put(to_put);
487 }
488
489
490 /*
491 * This routine returns the disk from which the requested read should
492 * be done. There is a per-array 'next expected sequential IO' sector
493 * number - if this matches on the next IO then we use the last disk.
494 * There is also a per-disk 'last know head position' sector that is
495 * maintained from IRQ contexts, both the normal and the resync IO
496 * completion handlers update this position correctly. If there is no
497 * perfect sequential match then we pick the disk whose head is closest.
498 *
499 * If there are 2 mirrors in the same 2 devices, performance degrades
500 * because position is mirror, not device based.
501 *
502 * The rdev for the device selected will have nr_pending incremented.
503 */
504 static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
505 {
506 const sector_t this_sector = r1_bio->sector;
507 int sectors;
508 int best_good_sectors;
509 int best_disk, best_dist_disk, best_pending_disk;
510 int has_nonrot_disk;
511 int disk;
512 sector_t best_dist;
513 unsigned int min_pending;
514 struct md_rdev *rdev;
515 int choose_first;
516 int choose_next_idle;
517
518 rcu_read_lock();
519 /*
520 * Check if we can balance. We can balance on the whole
521 * device if no resync is going on, or below the resync window.
522 * We take the first readable disk when above the resync window.
523 */
524 retry:
525 sectors = r1_bio->sectors;
526 best_disk = -1;
527 best_dist_disk = -1;
528 best_dist = MaxSector;
529 best_pending_disk = -1;
530 min_pending = UINT_MAX;
531 best_good_sectors = 0;
532 has_nonrot_disk = 0;
533 choose_next_idle = 0;
534
535 if (conf->mddev->recovery_cp < MaxSector &&
536 (this_sector + sectors >= conf->next_resync))
537 choose_first = 1;
538 else
539 choose_first = 0;
540
541 for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
542 sector_t dist;
543 sector_t first_bad;
544 int bad_sectors;
545 unsigned int pending;
546 bool nonrot;
547
548 rdev = rcu_dereference(conf->mirrors[disk].rdev);
549 if (r1_bio->bios[disk] == IO_BLOCKED
550 || rdev == NULL
551 || test_bit(Unmerged, &rdev->flags)
552 || test_bit(Faulty, &rdev->flags))
553 continue;
554 if (!test_bit(In_sync, &rdev->flags) &&
555 rdev->recovery_offset < this_sector + sectors)
556 continue;
557 if (test_bit(WriteMostly, &rdev->flags)) {
558 /* Don't balance among write-mostly, just
559 * use the first as a last resort */
560 if (best_disk < 0) {
561 if (is_badblock(rdev, this_sector, sectors,
562 &first_bad, &bad_sectors)) {
563 if (first_bad < this_sector)
564 /* Cannot use this */
565 continue;
566 best_good_sectors = first_bad - this_sector;
567 } else
568 best_good_sectors = sectors;
569 best_disk = disk;
570 }
571 continue;
572 }
573 /* This is a reasonable device to use. It might
574 * even be best.
575 */
576 if (is_badblock(rdev, this_sector, sectors,
577 &first_bad, &bad_sectors)) {
578 if (best_dist < MaxSector)
579 /* already have a better device */
580 continue;
581 if (first_bad <= this_sector) {
582 /* cannot read here. If this is the 'primary'
583 * device, then we must not read beyond
584 * bad_sectors from another device..
585 */
586 bad_sectors -= (this_sector - first_bad);
587 if (choose_first && sectors > bad_sectors)
588 sectors = bad_sectors;
589 if (best_good_sectors > sectors)
590 best_good_sectors = sectors;
591
592 } else {
593 sector_t good_sectors = first_bad - this_sector;
594 if (good_sectors > best_good_sectors) {
595 best_good_sectors = good_sectors;
596 best_disk = disk;
597 }
598 if (choose_first)
599 break;
600 }
601 continue;
602 } else
603 best_good_sectors = sectors;
604
605 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
606 has_nonrot_disk |= nonrot;
607 pending = atomic_read(&rdev->nr_pending);
608 dist = abs(this_sector - conf->mirrors[disk].head_position);
609 if (choose_first) {
610 best_disk = disk;
611 break;
612 }
613 /* Don't change to another disk for sequential reads */
614 if (conf->mirrors[disk].next_seq_sect == this_sector
615 || dist == 0) {
616 int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
617 struct raid1_info *mirror = &conf->mirrors[disk];
618
619 best_disk = disk;
620 /*
621 * If buffered sequential IO size exceeds optimal
622 * iosize, check if there is idle disk. If yes, choose
623 * the idle disk. read_balance could already choose an
624 * idle disk before noticing it's a sequential IO in
625 * this disk. This doesn't matter because this disk
626 * will idle, next time it will be utilized after the
627 * first disk has IO size exceeds optimal iosize. In
628 * this way, iosize of the first disk will be optimal
629 * iosize at least. iosize of the second disk might be
630 * small, but not a big deal since when the second disk
631 * starts IO, the first disk is likely still busy.
632 */
633 if (nonrot && opt_iosize > 0 &&
634 mirror->seq_start != MaxSector &&
635 mirror->next_seq_sect > opt_iosize &&
636 mirror->next_seq_sect - opt_iosize >=
637 mirror->seq_start) {
638 choose_next_idle = 1;
639 continue;
640 }
641 break;
642 }
643 /* If device is idle, use it */
644 if (pending == 0) {
645 best_disk = disk;
646 break;
647 }
648
649 if (choose_next_idle)
650 continue;
651
652 if (min_pending > pending) {
653 min_pending = pending;
654 best_pending_disk = disk;
655 }
656
657 if (dist < best_dist) {
658 best_dist = dist;
659 best_dist_disk = disk;
660 }
661 }
662
663 /*
664 * If all disks are rotational, choose the closest disk. If any disk is
665 * non-rotational, choose the disk with less pending request even the
666 * disk is rotational, which might/might not be optimal for raids with
667 * mixed ratation/non-rotational disks depending on workload.
668 */
669 if (best_disk == -1) {
670 if (has_nonrot_disk)
671 best_disk = best_pending_disk;
672 else
673 best_disk = best_dist_disk;
674 }
675
676 if (best_disk >= 0) {
677 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
678 if (!rdev)
679 goto retry;
680 atomic_inc(&rdev->nr_pending);
681 if (test_bit(Faulty, &rdev->flags)) {
682 /* cannot risk returning a device that failed
683 * before we inc'ed nr_pending
684 */
685 rdev_dec_pending(rdev, conf->mddev);
686 goto retry;
687 }
688 sectors = best_good_sectors;
689
690 if (conf->mirrors[best_disk].next_seq_sect != this_sector)
691 conf->mirrors[best_disk].seq_start = this_sector;
692
693 conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
694 }
695 rcu_read_unlock();
696 *max_sectors = sectors;
697
698 return best_disk;
699 }
700
701 static int raid1_mergeable_bvec(struct request_queue *q,
702 struct bvec_merge_data *bvm,
703 struct bio_vec *biovec)
704 {
705 struct mddev *mddev = q->queuedata;
706 struct r1conf *conf = mddev->private;
707 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
708 int max = biovec->bv_len;
709
710 if (mddev->merge_check_needed) {
711 int disk;
712 rcu_read_lock();
713 for (disk = 0; disk < conf->raid_disks * 2; disk++) {
714 struct md_rdev *rdev = rcu_dereference(
715 conf->mirrors[disk].rdev);
716 if (rdev && !test_bit(Faulty, &rdev->flags)) {
717 struct request_queue *q =
718 bdev_get_queue(rdev->bdev);
719 if (q->merge_bvec_fn) {
720 bvm->bi_sector = sector +
721 rdev->data_offset;
722 bvm->bi_bdev = rdev->bdev;
723 max = min(max, q->merge_bvec_fn(
724 q, bvm, biovec));
725 }
726 }
727 }
728 rcu_read_unlock();
729 }
730 return max;
731
732 }
733
734 int md_raid1_congested(struct mddev *mddev, int bits)
735 {
736 struct r1conf *conf = mddev->private;
737 int i, ret = 0;
738
739 if ((bits & (1 << BDI_async_congested)) &&
740 conf->pending_count >= max_queued_requests)
741 return 1;
742
743 rcu_read_lock();
744 for (i = 0; i < conf->raid_disks * 2; i++) {
745 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
746 if (rdev && !test_bit(Faulty, &rdev->flags)) {
747 struct request_queue *q = bdev_get_queue(rdev->bdev);
748
749 BUG_ON(!q);
750
751 /* Note the '|| 1' - when read_balance prefers
752 * non-congested targets, it can be removed
753 */
754 if ((bits & (1<<BDI_async_congested)) || 1)
755 ret |= bdi_congested(&q->backing_dev_info, bits);
756 else
757 ret &= bdi_congested(&q->backing_dev_info, bits);
758 }
759 }
760 rcu_read_unlock();
761 return ret;
762 }
763 EXPORT_SYMBOL_GPL(md_raid1_congested);
764
765 static int raid1_congested(void *data, int bits)
766 {
767 struct mddev *mddev = data;
768
769 return mddev_congested(mddev, bits) ||
770 md_raid1_congested(mddev, bits);
771 }
772
773 static void flush_pending_writes(struct r1conf *conf)
774 {
775 /* Any writes that have been queued but are awaiting
776 * bitmap updates get flushed here.
777 */
778 spin_lock_irq(&conf->device_lock);
779
780 if (conf->pending_bio_list.head) {
781 struct bio *bio;
782 bio = bio_list_get(&conf->pending_bio_list);
783 conf->pending_count = 0;
784 spin_unlock_irq(&conf->device_lock);
785 /* flush any pending bitmap writes to
786 * disk before proceeding w/ I/O */
787 bitmap_unplug(conf->mddev->bitmap);
788 wake_up(&conf->wait_barrier);
789
790 while (bio) { /* submit pending writes */
791 struct bio *next = bio->bi_next;
792 bio->bi_next = NULL;
793 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
794 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
795 /* Just ignore it */
796 bio_endio(bio, 0);
797 else
798 generic_make_request(bio);
799 bio = next;
800 }
801 } else
802 spin_unlock_irq(&conf->device_lock);
803 }
804
805 /* Barriers....
806 * Sometimes we need to suspend IO while we do something else,
807 * either some resync/recovery, or reconfigure the array.
808 * To do this we raise a 'barrier'.
809 * The 'barrier' is a counter that can be raised multiple times
810 * to count how many activities are happening which preclude
811 * normal IO.
812 * We can only raise the barrier if there is no pending IO.
813 * i.e. if nr_pending == 0.
814 * We choose only to raise the barrier if no-one is waiting for the
815 * barrier to go down. This means that as soon as an IO request
816 * is ready, no other operations which require a barrier will start
817 * until the IO request has had a chance.
818 *
819 * So: regular IO calls 'wait_barrier'. When that returns there
820 * is no backgroup IO happening, It must arrange to call
821 * allow_barrier when it has finished its IO.
822 * backgroup IO calls must call raise_barrier. Once that returns
823 * there is no normal IO happeing. It must arrange to call
824 * lower_barrier when the particular background IO completes.
825 */
826 #define RESYNC_DEPTH 32
827
828 static void raise_barrier(struct r1conf *conf)
829 {
830 spin_lock_irq(&conf->resync_lock);
831
832 /* Wait until no block IO is waiting */
833 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
834 conf->resync_lock);
835
836 /* block any new IO from starting */
837 conf->barrier++;
838
839 /* Now wait for all pending IO to complete */
840 wait_event_lock_irq(conf->wait_barrier,
841 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
842 conf->resync_lock);
843
844 spin_unlock_irq(&conf->resync_lock);
845 }
846
847 static void lower_barrier(struct r1conf *conf)
848 {
849 unsigned long flags;
850 BUG_ON(conf->barrier <= 0);
851 spin_lock_irqsave(&conf->resync_lock, flags);
852 conf->barrier--;
853 spin_unlock_irqrestore(&conf->resync_lock, flags);
854 wake_up(&conf->wait_barrier);
855 }
856
857 static void wait_barrier(struct r1conf *conf)
858 {
859 spin_lock_irq(&conf->resync_lock);
860 if (conf->barrier) {
861 conf->nr_waiting++;
862 /* Wait for the barrier to drop.
863 * However if there are already pending
864 * requests (preventing the barrier from
865 * rising completely), and the
866 * pre-process bio queue isn't empty,
867 * then don't wait, as we need to empty
868 * that queue to get the nr_pending
869 * count down.
870 */
871 wait_event_lock_irq(conf->wait_barrier,
872 !conf->barrier ||
873 (conf->nr_pending &&
874 current->bio_list &&
875 !bio_list_empty(current->bio_list)),
876 conf->resync_lock);
877 conf->nr_waiting--;
878 }
879 conf->nr_pending++;
880 spin_unlock_irq(&conf->resync_lock);
881 }
882
883 static void allow_barrier(struct r1conf *conf)
884 {
885 unsigned long flags;
886 spin_lock_irqsave(&conf->resync_lock, flags);
887 conf->nr_pending--;
888 spin_unlock_irqrestore(&conf->resync_lock, flags);
889 wake_up(&conf->wait_barrier);
890 }
891
892 static void freeze_array(struct r1conf *conf, int extra)
893 {
894 /* stop syncio and normal IO and wait for everything to
895 * go quite.
896 * We increment barrier and nr_waiting, and then
897 * wait until nr_pending match nr_queued+extra
898 * This is called in the context of one normal IO request
899 * that has failed. Thus any sync request that might be pending
900 * will be blocked by nr_pending, and we need to wait for
901 * pending IO requests to complete or be queued for re-try.
902 * Thus the number queued (nr_queued) plus this request (extra)
903 * must match the number of pending IOs (nr_pending) before
904 * we continue.
905 */
906 spin_lock_irq(&conf->resync_lock);
907 conf->barrier++;
908 conf->nr_waiting++;
909 wait_event_lock_irq_cmd(conf->wait_barrier,
910 conf->nr_pending == conf->nr_queued+extra,
911 conf->resync_lock,
912 flush_pending_writes(conf));
913 spin_unlock_irq(&conf->resync_lock);
914 }
915 static void unfreeze_array(struct r1conf *conf)
916 {
917 /* reverse the effect of the freeze */
918 spin_lock_irq(&conf->resync_lock);
919 conf->barrier--;
920 conf->nr_waiting--;
921 wake_up(&conf->wait_barrier);
922 spin_unlock_irq(&conf->resync_lock);
923 }
924
925
926 /* duplicate the data pages for behind I/O
927 */
928 static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
929 {
930 int i;
931 struct bio_vec *bvec;
932 struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
933 GFP_NOIO);
934 if (unlikely(!bvecs))
935 return;
936
937 bio_for_each_segment_all(bvec, bio, i) {
938 bvecs[i] = *bvec;
939 bvecs[i].bv_page = alloc_page(GFP_NOIO);
940 if (unlikely(!bvecs[i].bv_page))
941 goto do_sync_io;
942 memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
943 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
944 kunmap(bvecs[i].bv_page);
945 kunmap(bvec->bv_page);
946 }
947 r1_bio->behind_bvecs = bvecs;
948 r1_bio->behind_page_count = bio->bi_vcnt;
949 set_bit(R1BIO_BehindIO, &r1_bio->state);
950 return;
951
952 do_sync_io:
953 for (i = 0; i < bio->bi_vcnt; i++)
954 if (bvecs[i].bv_page)
955 put_page(bvecs[i].bv_page);
956 kfree(bvecs);
957 pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
958 }
959
960 struct raid1_plug_cb {
961 struct blk_plug_cb cb;
962 struct bio_list pending;
963 int pending_cnt;
964 };
965
966 static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
967 {
968 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
969 cb);
970 struct mddev *mddev = plug->cb.data;
971 struct r1conf *conf = mddev->private;
972 struct bio *bio;
973
974 if (from_schedule || current->bio_list) {
975 spin_lock_irq(&conf->device_lock);
976 bio_list_merge(&conf->pending_bio_list, &plug->pending);
977 conf->pending_count += plug->pending_cnt;
978 spin_unlock_irq(&conf->device_lock);
979 wake_up(&conf->wait_barrier);
980 md_wakeup_thread(mddev->thread);
981 kfree(plug);
982 return;
983 }
984
985 /* we aren't scheduling, so we can do the write-out directly. */
986 bio = bio_list_get(&plug->pending);
987 bitmap_unplug(mddev->bitmap);
988 wake_up(&conf->wait_barrier);
989
990 while (bio) { /* submit pending writes */
991 struct bio *next = bio->bi_next;
992 bio->bi_next = NULL;
993 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
994 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
995 /* Just ignore it */
996 bio_endio(bio, 0);
997 else
998 generic_make_request(bio);
999 bio = next;
1000 }
1001 kfree(plug);
1002 }
1003
1004 static void make_request(struct mddev *mddev, struct bio * bio)
1005 {
1006 struct r1conf *conf = mddev->private;
1007 struct raid1_info *mirror;
1008 struct r1bio *r1_bio;
1009 struct bio *read_bio;
1010 int i, disks;
1011 struct bitmap *bitmap;
1012 unsigned long flags;
1013 const int rw = bio_data_dir(bio);
1014 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1015 const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1016 const unsigned long do_discard = (bio->bi_rw
1017 & (REQ_DISCARD | REQ_SECURE));
1018 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1019 struct md_rdev *blocked_rdev;
1020 struct blk_plug_cb *cb;
1021 struct raid1_plug_cb *plug = NULL;
1022 int first_clone;
1023 int sectors_handled;
1024 int max_sectors;
1025
1026 /*
1027 * Register the new request and wait if the reconstruction
1028 * thread has put up a bar for new requests.
1029 * Continue immediately if no resync is active currently.
1030 */
1031
1032 md_write_start(mddev, bio); /* wait on superblock update early */
1033
1034 if (bio_data_dir(bio) == WRITE &&
1035 bio_end_sector(bio) > mddev->suspend_lo &&
1036 bio->bi_sector < mddev->suspend_hi) {
1037 /* As the suspend_* range is controlled by
1038 * userspace, we want an interruptible
1039 * wait.
1040 */
1041 DEFINE_WAIT(w);
1042 for (;;) {
1043 flush_signals(current);
1044 prepare_to_wait(&conf->wait_barrier,
1045 &w, TASK_INTERRUPTIBLE);
1046 if (bio_end_sector(bio) <= mddev->suspend_lo ||
1047 bio->bi_sector >= mddev->suspend_hi)
1048 break;
1049 schedule();
1050 }
1051 finish_wait(&conf->wait_barrier, &w);
1052 }
1053
1054 wait_barrier(conf);
1055
1056 bitmap = mddev->bitmap;
1057
1058 /*
1059 * make_request() can abort the operation when READA is being
1060 * used and no empty request is available.
1061 *
1062 */
1063 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1064
1065 r1_bio->master_bio = bio;
1066 r1_bio->sectors = bio_sectors(bio);
1067 r1_bio->state = 0;
1068 r1_bio->mddev = mddev;
1069 r1_bio->sector = bio->bi_sector;
1070
1071 /* We might need to issue multiple reads to different
1072 * devices if there are bad blocks around, so we keep
1073 * track of the number of reads in bio->bi_phys_segments.
1074 * If this is 0, there is only one r1_bio and no locking
1075 * will be needed when requests complete. If it is
1076 * non-zero, then it is the number of not-completed requests.
1077 */
1078 bio->bi_phys_segments = 0;
1079 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1080
1081 if (rw == READ) {
1082 /*
1083 * read balancing logic:
1084 */
1085 int rdisk;
1086
1087 read_again:
1088 rdisk = read_balance(conf, r1_bio, &max_sectors);
1089
1090 if (rdisk < 0) {
1091 /* couldn't find anywhere to read from */
1092 raid_end_bio_io(r1_bio);
1093 return;
1094 }
1095 mirror = conf->mirrors + rdisk;
1096
1097 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1098 bitmap) {
1099 /* Reading from a write-mostly device must
1100 * take care not to over-take any writes
1101 * that are 'behind'
1102 */
1103 wait_event(bitmap->behind_wait,
1104 atomic_read(&bitmap->behind_writes) == 0);
1105 }
1106 r1_bio->read_disk = rdisk;
1107
1108 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1109 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
1110 max_sectors);
1111
1112 r1_bio->bios[rdisk] = read_bio;
1113
1114 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
1115 read_bio->bi_bdev = mirror->rdev->bdev;
1116 read_bio->bi_end_io = raid1_end_read_request;
1117 read_bio->bi_rw = READ | do_sync;
1118 read_bio->bi_private = r1_bio;
1119
1120 if (max_sectors < r1_bio->sectors) {
1121 /* could not read all from this device, so we will
1122 * need another r1_bio.
1123 */
1124
1125 sectors_handled = (r1_bio->sector + max_sectors
1126 - bio->bi_sector);
1127 r1_bio->sectors = max_sectors;
1128 spin_lock_irq(&conf->device_lock);
1129 if (bio->bi_phys_segments == 0)
1130 bio->bi_phys_segments = 2;
1131 else
1132 bio->bi_phys_segments++;
1133 spin_unlock_irq(&conf->device_lock);
1134 /* Cannot call generic_make_request directly
1135 * as that will be queued in __make_request
1136 * and subsequent mempool_alloc might block waiting
1137 * for it. So hand bio over to raid1d.
1138 */
1139 reschedule_retry(r1_bio);
1140
1141 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1142
1143 r1_bio->master_bio = bio;
1144 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1145 r1_bio->state = 0;
1146 r1_bio->mddev = mddev;
1147 r1_bio->sector = bio->bi_sector + sectors_handled;
1148 goto read_again;
1149 } else
1150 generic_make_request(read_bio);
1151 return;
1152 }
1153
1154 /*
1155 * WRITE:
1156 */
1157 if (conf->pending_count >= max_queued_requests) {
1158 md_wakeup_thread(mddev->thread);
1159 wait_event(conf->wait_barrier,
1160 conf->pending_count < max_queued_requests);
1161 }
1162 /* first select target devices under rcu_lock and
1163 * inc refcount on their rdev. Record them by setting
1164 * bios[x] to bio
1165 * If there are known/acknowledged bad blocks on any device on
1166 * which we have seen a write error, we want to avoid writing those
1167 * blocks.
1168 * This potentially requires several writes to write around
1169 * the bad blocks. Each set of writes gets it's own r1bio
1170 * with a set of bios attached.
1171 */
1172
1173 disks = conf->raid_disks * 2;
1174 retry_write:
1175 blocked_rdev = NULL;
1176 rcu_read_lock();
1177 max_sectors = r1_bio->sectors;
1178 for (i = 0; i < disks; i++) {
1179 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1180 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1181 atomic_inc(&rdev->nr_pending);
1182 blocked_rdev = rdev;
1183 break;
1184 }
1185 r1_bio->bios[i] = NULL;
1186 if (!rdev || test_bit(Faulty, &rdev->flags)
1187 || test_bit(Unmerged, &rdev->flags)) {
1188 if (i < conf->raid_disks)
1189 set_bit(R1BIO_Degraded, &r1_bio->state);
1190 continue;
1191 }
1192
1193 atomic_inc(&rdev->nr_pending);
1194 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1195 sector_t first_bad;
1196 int bad_sectors;
1197 int is_bad;
1198
1199 is_bad = is_badblock(rdev, r1_bio->sector,
1200 max_sectors,
1201 &first_bad, &bad_sectors);
1202 if (is_bad < 0) {
1203 /* mustn't write here until the bad block is
1204 * acknowledged*/
1205 set_bit(BlockedBadBlocks, &rdev->flags);
1206 blocked_rdev = rdev;
1207 break;
1208 }
1209 if (is_bad && first_bad <= r1_bio->sector) {
1210 /* Cannot write here at all */
1211 bad_sectors -= (r1_bio->sector - first_bad);
1212 if (bad_sectors < max_sectors)
1213 /* mustn't write more than bad_sectors
1214 * to other devices yet
1215 */
1216 max_sectors = bad_sectors;
1217 rdev_dec_pending(rdev, mddev);
1218 /* We don't set R1BIO_Degraded as that
1219 * only applies if the disk is
1220 * missing, so it might be re-added,
1221 * and we want to know to recover this
1222 * chunk.
1223 * In this case the device is here,
1224 * and the fact that this chunk is not
1225 * in-sync is recorded in the bad
1226 * block log
1227 */
1228 continue;
1229 }
1230 if (is_bad) {
1231 int good_sectors = first_bad - r1_bio->sector;
1232 if (good_sectors < max_sectors)
1233 max_sectors = good_sectors;
1234 }
1235 }
1236 r1_bio->bios[i] = bio;
1237 }
1238 rcu_read_unlock();
1239
1240 if (unlikely(blocked_rdev)) {
1241 /* Wait for this device to become unblocked */
1242 int j;
1243
1244 for (j = 0; j < i; j++)
1245 if (r1_bio->bios[j])
1246 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1247 r1_bio->state = 0;
1248 allow_barrier(conf);
1249 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1250 wait_barrier(conf);
1251 goto retry_write;
1252 }
1253
1254 if (max_sectors < r1_bio->sectors) {
1255 /* We are splitting this write into multiple parts, so
1256 * we need to prepare for allocating another r1_bio.
1257 */
1258 r1_bio->sectors = max_sectors;
1259 spin_lock_irq(&conf->device_lock);
1260 if (bio->bi_phys_segments == 0)
1261 bio->bi_phys_segments = 2;
1262 else
1263 bio->bi_phys_segments++;
1264 spin_unlock_irq(&conf->device_lock);
1265 }
1266 sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1267
1268 atomic_set(&r1_bio->remaining, 1);
1269 atomic_set(&r1_bio->behind_remaining, 0);
1270
1271 first_clone = 1;
1272 for (i = 0; i < disks; i++) {
1273 struct bio *mbio;
1274 if (!r1_bio->bios[i])
1275 continue;
1276
1277 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1278 md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1279
1280 if (first_clone) {
1281 /* do behind I/O ?
1282 * Not if there are too many, or cannot
1283 * allocate memory, or a reader on WriteMostly
1284 * is waiting for behind writes to flush */
1285 if (bitmap &&
1286 (atomic_read(&bitmap->behind_writes)
1287 < mddev->bitmap_info.max_write_behind) &&
1288 !waitqueue_active(&bitmap->behind_wait))
1289 alloc_behind_pages(mbio, r1_bio);
1290
1291 bitmap_startwrite(bitmap, r1_bio->sector,
1292 r1_bio->sectors,
1293 test_bit(R1BIO_BehindIO,
1294 &r1_bio->state));
1295 first_clone = 0;
1296 }
1297 if (r1_bio->behind_bvecs) {
1298 struct bio_vec *bvec;
1299 int j;
1300
1301 /*
1302 * We trimmed the bio, so _all is legit
1303 */
1304 bio_for_each_segment_all(bvec, mbio, j)
1305 bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1306 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1307 atomic_inc(&r1_bio->behind_remaining);
1308 }
1309
1310 r1_bio->bios[i] = mbio;
1311
1312 mbio->bi_sector = (r1_bio->sector +
1313 conf->mirrors[i].rdev->data_offset);
1314 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1315 mbio->bi_end_io = raid1_end_write_request;
1316 mbio->bi_rw =
1317 WRITE | do_flush_fua | do_sync | do_discard | do_same;
1318 mbio->bi_private = r1_bio;
1319
1320 atomic_inc(&r1_bio->remaining);
1321
1322 cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1323 if (cb)
1324 plug = container_of(cb, struct raid1_plug_cb, cb);
1325 else
1326 plug = NULL;
1327 spin_lock_irqsave(&conf->device_lock, flags);
1328 if (plug) {
1329 bio_list_add(&plug->pending, mbio);
1330 plug->pending_cnt++;
1331 } else {
1332 bio_list_add(&conf->pending_bio_list, mbio);
1333 conf->pending_count++;
1334 }
1335 spin_unlock_irqrestore(&conf->device_lock, flags);
1336 if (!plug)
1337 md_wakeup_thread(mddev->thread);
1338 }
1339 /* Mustn't call r1_bio_write_done before this next test,
1340 * as it could result in the bio being freed.
1341 */
1342 if (sectors_handled < bio_sectors(bio)) {
1343 r1_bio_write_done(r1_bio);
1344 /* We need another r1_bio. It has already been counted
1345 * in bio->bi_phys_segments
1346 */
1347 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1348 r1_bio->master_bio = bio;
1349 r1_bio->sectors = bio_sectors(bio) - sectors_handled;
1350 r1_bio->state = 0;
1351 r1_bio->mddev = mddev;
1352 r1_bio->sector = bio->bi_sector + sectors_handled;
1353 goto retry_write;
1354 }
1355
1356 r1_bio_write_done(r1_bio);
1357
1358 /* In case raid1d snuck in to freeze_array */
1359 wake_up(&conf->wait_barrier);
1360 }
1361
1362 static void status(struct seq_file *seq, struct mddev *mddev)
1363 {
1364 struct r1conf *conf = mddev->private;
1365 int i;
1366
1367 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1368 conf->raid_disks - mddev->degraded);
1369 rcu_read_lock();
1370 for (i = 0; i < conf->raid_disks; i++) {
1371 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1372 seq_printf(seq, "%s",
1373 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1374 }
1375 rcu_read_unlock();
1376 seq_printf(seq, "]");
1377 }
1378
1379
1380 static void error(struct mddev *mddev, struct md_rdev *rdev)
1381 {
1382 char b[BDEVNAME_SIZE];
1383 struct r1conf *conf = mddev->private;
1384
1385 /*
1386 * If it is not operational, then we have already marked it as dead
1387 * else if it is the last working disks, ignore the error, let the
1388 * next level up know.
1389 * else mark the drive as failed
1390 */
1391 if (test_bit(In_sync, &rdev->flags)
1392 && (conf->raid_disks - mddev->degraded) == 1) {
1393 /*
1394 * Don't fail the drive, act as though we were just a
1395 * normal single drive.
1396 * However don't try a recovery from this drive as
1397 * it is very likely to fail.
1398 */
1399 conf->recovery_disabled = mddev->recovery_disabled;
1400 return;
1401 }
1402 set_bit(Blocked, &rdev->flags);
1403 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1404 unsigned long flags;
1405 spin_lock_irqsave(&conf->device_lock, flags);
1406 mddev->degraded++;
1407 set_bit(Faulty, &rdev->flags);
1408 spin_unlock_irqrestore(&conf->device_lock, flags);
1409 } else
1410 set_bit(Faulty, &rdev->flags);
1411 /*
1412 * if recovery is running, make sure it aborts.
1413 */
1414 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1415 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1416 printk(KERN_ALERT
1417 "md/raid1:%s: Disk failure on %s, disabling device.\n"
1418 "md/raid1:%s: Operation continuing on %d devices.\n",
1419 mdname(mddev), bdevname(rdev->bdev, b),
1420 mdname(mddev), conf->raid_disks - mddev->degraded);
1421 }
1422
1423 static void print_conf(struct r1conf *conf)
1424 {
1425 int i;
1426
1427 printk(KERN_DEBUG "RAID1 conf printout:\n");
1428 if (!conf) {
1429 printk(KERN_DEBUG "(!conf)\n");
1430 return;
1431 }
1432 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1433 conf->raid_disks);
1434
1435 rcu_read_lock();
1436 for (i = 0; i < conf->raid_disks; i++) {
1437 char b[BDEVNAME_SIZE];
1438 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1439 if (rdev)
1440 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1441 i, !test_bit(In_sync, &rdev->flags),
1442 !test_bit(Faulty, &rdev->flags),
1443 bdevname(rdev->bdev,b));
1444 }
1445 rcu_read_unlock();
1446 }
1447
1448 static void close_sync(struct r1conf *conf)
1449 {
1450 wait_barrier(conf);
1451 allow_barrier(conf);
1452
1453 mempool_destroy(conf->r1buf_pool);
1454 conf->r1buf_pool = NULL;
1455 }
1456
1457 static int raid1_spare_active(struct mddev *mddev)
1458 {
1459 int i;
1460 struct r1conf *conf = mddev->private;
1461 int count = 0;
1462 unsigned long flags;
1463
1464 /*
1465 * Find all failed disks within the RAID1 configuration
1466 * and mark them readable.
1467 * Called under mddev lock, so rcu protection not needed.
1468 */
1469 for (i = 0; i < conf->raid_disks; i++) {
1470 struct md_rdev *rdev = conf->mirrors[i].rdev;
1471 struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1472 if (repl
1473 && repl->recovery_offset == MaxSector
1474 && !test_bit(Faulty, &repl->flags)
1475 && !test_and_set_bit(In_sync, &repl->flags)) {
1476 /* replacement has just become active */
1477 if (!rdev ||
1478 !test_and_clear_bit(In_sync, &rdev->flags))
1479 count++;
1480 if (rdev) {
1481 /* Replaced device not technically
1482 * faulty, but we need to be sure
1483 * it gets removed and never re-added
1484 */
1485 set_bit(Faulty, &rdev->flags);
1486 sysfs_notify_dirent_safe(
1487 rdev->sysfs_state);
1488 }
1489 }
1490 if (rdev
1491 && rdev->recovery_offset == MaxSector
1492 && !test_bit(Faulty, &rdev->flags)
1493 && !test_and_set_bit(In_sync, &rdev->flags)) {
1494 count++;
1495 sysfs_notify_dirent_safe(rdev->sysfs_state);
1496 }
1497 }
1498 spin_lock_irqsave(&conf->device_lock, flags);
1499 mddev->degraded -= count;
1500 spin_unlock_irqrestore(&conf->device_lock, flags);
1501
1502 print_conf(conf);
1503 return count;
1504 }
1505
1506
1507 static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1508 {
1509 struct r1conf *conf = mddev->private;
1510 int err = -EEXIST;
1511 int mirror = 0;
1512 struct raid1_info *p;
1513 int first = 0;
1514 int last = conf->raid_disks - 1;
1515 struct request_queue *q = bdev_get_queue(rdev->bdev);
1516
1517 if (mddev->recovery_disabled == conf->recovery_disabled)
1518 return -EBUSY;
1519
1520 if (rdev->raid_disk >= 0)
1521 first = last = rdev->raid_disk;
1522
1523 if (q->merge_bvec_fn) {
1524 set_bit(Unmerged, &rdev->flags);
1525 mddev->merge_check_needed = 1;
1526 }
1527
1528 for (mirror = first; mirror <= last; mirror++) {
1529 p = conf->mirrors+mirror;
1530 if (!p->rdev) {
1531
1532 disk_stack_limits(mddev->gendisk, rdev->bdev,
1533 rdev->data_offset << 9);
1534
1535 p->head_position = 0;
1536 rdev->raid_disk = mirror;
1537 err = 0;
1538 /* As all devices are equivalent, we don't need a full recovery
1539 * if this was recently any drive of the array
1540 */
1541 if (rdev->saved_raid_disk < 0)
1542 conf->fullsync = 1;
1543 rcu_assign_pointer(p->rdev, rdev);
1544 break;
1545 }
1546 if (test_bit(WantReplacement, &p->rdev->flags) &&
1547 p[conf->raid_disks].rdev == NULL) {
1548 /* Add this device as a replacement */
1549 clear_bit(In_sync, &rdev->flags);
1550 set_bit(Replacement, &rdev->flags);
1551 rdev->raid_disk = mirror;
1552 err = 0;
1553 conf->fullsync = 1;
1554 rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1555 break;
1556 }
1557 }
1558 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1559 /* Some requests might not have seen this new
1560 * merge_bvec_fn. We must wait for them to complete
1561 * before merging the device fully.
1562 * First we make sure any code which has tested
1563 * our function has submitted the request, then
1564 * we wait for all outstanding requests to complete.
1565 */
1566 synchronize_sched();
1567 freeze_array(conf, 0);
1568 unfreeze_array(conf);
1569 clear_bit(Unmerged, &rdev->flags);
1570 }
1571 md_integrity_add_rdev(rdev, mddev);
1572 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
1573 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1574 print_conf(conf);
1575 return err;
1576 }
1577
1578 static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1579 {
1580 struct r1conf *conf = mddev->private;
1581 int err = 0;
1582 int number = rdev->raid_disk;
1583 struct raid1_info *p = conf->mirrors + number;
1584
1585 if (rdev != p->rdev)
1586 p = conf->mirrors + conf->raid_disks + number;
1587
1588 print_conf(conf);
1589 if (rdev == p->rdev) {
1590 if (test_bit(In_sync, &rdev->flags) ||
1591 atomic_read(&rdev->nr_pending)) {
1592 err = -EBUSY;
1593 goto abort;
1594 }
1595 /* Only remove non-faulty devices if recovery
1596 * is not possible.
1597 */
1598 if (!test_bit(Faulty, &rdev->flags) &&
1599 mddev->recovery_disabled != conf->recovery_disabled &&
1600 mddev->degraded < conf->raid_disks) {
1601 err = -EBUSY;
1602 goto abort;
1603 }
1604 p->rdev = NULL;
1605 synchronize_rcu();
1606 if (atomic_read(&rdev->nr_pending)) {
1607 /* lost the race, try later */
1608 err = -EBUSY;
1609 p->rdev = rdev;
1610 goto abort;
1611 } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1612 /* We just removed a device that is being replaced.
1613 * Move down the replacement. We drain all IO before
1614 * doing this to avoid confusion.
1615 */
1616 struct md_rdev *repl =
1617 conf->mirrors[conf->raid_disks + number].rdev;
1618 freeze_array(conf, 0);
1619 clear_bit(Replacement, &repl->flags);
1620 p->rdev = repl;
1621 conf->mirrors[conf->raid_disks + number].rdev = NULL;
1622 unfreeze_array(conf);
1623 clear_bit(WantReplacement, &rdev->flags);
1624 } else
1625 clear_bit(WantReplacement, &rdev->flags);
1626 err = md_integrity_register(mddev);
1627 }
1628 abort:
1629
1630 print_conf(conf);
1631 return err;
1632 }
1633
1634
1635 static void end_sync_read(struct bio *bio, int error)
1636 {
1637 struct r1bio *r1_bio = bio->bi_private;
1638
1639 update_head_pos(r1_bio->read_disk, r1_bio);
1640
1641 /*
1642 * we have read a block, now it needs to be re-written,
1643 * or re-read if the read failed.
1644 * We don't do much here, just schedule handling by raid1d
1645 */
1646 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1647 set_bit(R1BIO_Uptodate, &r1_bio->state);
1648
1649 if (atomic_dec_and_test(&r1_bio->remaining))
1650 reschedule_retry(r1_bio);
1651 }
1652
1653 static void end_sync_write(struct bio *bio, int error)
1654 {
1655 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1656 struct r1bio *r1_bio = bio->bi_private;
1657 struct mddev *mddev = r1_bio->mddev;
1658 struct r1conf *conf = mddev->private;
1659 int mirror=0;
1660 sector_t first_bad;
1661 int bad_sectors;
1662
1663 mirror = find_bio_disk(r1_bio, bio);
1664
1665 if (!uptodate) {
1666 sector_t sync_blocks = 0;
1667 sector_t s = r1_bio->sector;
1668 long sectors_to_go = r1_bio->sectors;
1669 /* make sure these bits doesn't get cleared. */
1670 do {
1671 bitmap_end_sync(mddev->bitmap, s,
1672 &sync_blocks, 1);
1673 s += sync_blocks;
1674 sectors_to_go -= sync_blocks;
1675 } while (sectors_to_go > 0);
1676 set_bit(WriteErrorSeen,
1677 &conf->mirrors[mirror].rdev->flags);
1678 if (!test_and_set_bit(WantReplacement,
1679 &conf->mirrors[mirror].rdev->flags))
1680 set_bit(MD_RECOVERY_NEEDED, &
1681 mddev->recovery);
1682 set_bit(R1BIO_WriteError, &r1_bio->state);
1683 } else if (is_badblock(conf->mirrors[mirror].rdev,
1684 r1_bio->sector,
1685 r1_bio->sectors,
1686 &first_bad, &bad_sectors) &&
1687 !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1688 r1_bio->sector,
1689 r1_bio->sectors,
1690 &first_bad, &bad_sectors)
1691 )
1692 set_bit(R1BIO_MadeGood, &r1_bio->state);
1693
1694 if (atomic_dec_and_test(&r1_bio->remaining)) {
1695 int s = r1_bio->sectors;
1696 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1697 test_bit(R1BIO_WriteError, &r1_bio->state))
1698 reschedule_retry(r1_bio);
1699 else {
1700 put_buf(r1_bio);
1701 md_done_sync(mddev, s, uptodate);
1702 }
1703 }
1704 }
1705
1706 static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1707 int sectors, struct page *page, int rw)
1708 {
1709 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1710 /* success */
1711 return 1;
1712 if (rw == WRITE) {
1713 set_bit(WriteErrorSeen, &rdev->flags);
1714 if (!test_and_set_bit(WantReplacement,
1715 &rdev->flags))
1716 set_bit(MD_RECOVERY_NEEDED, &
1717 rdev->mddev->recovery);
1718 }
1719 /* need to record an error - either for the block or the device */
1720 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1721 md_error(rdev->mddev, rdev);
1722 return 0;
1723 }
1724
1725 static int fix_sync_read_error(struct r1bio *r1_bio)
1726 {
1727 /* Try some synchronous reads of other devices to get
1728 * good data, much like with normal read errors. Only
1729 * read into the pages we already have so we don't
1730 * need to re-issue the read request.
1731 * We don't need to freeze the array, because being in an
1732 * active sync request, there is no normal IO, and
1733 * no overlapping syncs.
1734 * We don't need to check is_badblock() again as we
1735 * made sure that anything with a bad block in range
1736 * will have bi_end_io clear.
1737 */
1738 struct mddev *mddev = r1_bio->mddev;
1739 struct r1conf *conf = mddev->private;
1740 struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1741 sector_t sect = r1_bio->sector;
1742 int sectors = r1_bio->sectors;
1743 int idx = 0;
1744
1745 while(sectors) {
1746 int s = sectors;
1747 int d = r1_bio->read_disk;
1748 int success = 0;
1749 struct md_rdev *rdev;
1750 int start;
1751
1752 if (s > (PAGE_SIZE>>9))
1753 s = PAGE_SIZE >> 9;
1754 do {
1755 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1756 /* No rcu protection needed here devices
1757 * can only be removed when no resync is
1758 * active, and resync is currently active
1759 */
1760 rdev = conf->mirrors[d].rdev;
1761 if (sync_page_io(rdev, sect, s<<9,
1762 bio->bi_io_vec[idx].bv_page,
1763 READ, false)) {
1764 success = 1;
1765 break;
1766 }
1767 }
1768 d++;
1769 if (d == conf->raid_disks * 2)
1770 d = 0;
1771 } while (!success && d != r1_bio->read_disk);
1772
1773 if (!success) {
1774 char b[BDEVNAME_SIZE];
1775 int abort = 0;
1776 /* Cannot read from anywhere, this block is lost.
1777 * Record a bad block on each device. If that doesn't
1778 * work just disable and interrupt the recovery.
1779 * Don't fail devices as that won't really help.
1780 */
1781 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1782 " for block %llu\n",
1783 mdname(mddev),
1784 bdevname(bio->bi_bdev, b),
1785 (unsigned long long)r1_bio->sector);
1786 for (d = 0; d < conf->raid_disks * 2; d++) {
1787 rdev = conf->mirrors[d].rdev;
1788 if (!rdev || test_bit(Faulty, &rdev->flags))
1789 continue;
1790 if (!rdev_set_badblocks(rdev, sect, s, 0))
1791 abort = 1;
1792 }
1793 if (abort) {
1794 conf->recovery_disabled =
1795 mddev->recovery_disabled;
1796 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1797 md_done_sync(mddev, r1_bio->sectors, 0);
1798 put_buf(r1_bio);
1799 return 0;
1800 }
1801 /* Try next page */
1802 sectors -= s;
1803 sect += s;
1804 idx++;
1805 continue;
1806 }
1807
1808 start = d;
1809 /* write it back and re-read */
1810 while (d != r1_bio->read_disk) {
1811 if (d == 0)
1812 d = conf->raid_disks * 2;
1813 d--;
1814 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1815 continue;
1816 rdev = conf->mirrors[d].rdev;
1817 if (r1_sync_page_io(rdev, sect, s,
1818 bio->bi_io_vec[idx].bv_page,
1819 WRITE) == 0) {
1820 r1_bio->bios[d]->bi_end_io = NULL;
1821 rdev_dec_pending(rdev, mddev);
1822 }
1823 }
1824 d = start;
1825 while (d != r1_bio->read_disk) {
1826 if (d == 0)
1827 d = conf->raid_disks * 2;
1828 d--;
1829 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1830 continue;
1831 rdev = conf->mirrors[d].rdev;
1832 if (r1_sync_page_io(rdev, sect, s,
1833 bio->bi_io_vec[idx].bv_page,
1834 READ) != 0)
1835 atomic_add(s, &rdev->corrected_errors);
1836 }
1837 sectors -= s;
1838 sect += s;
1839 idx ++;
1840 }
1841 set_bit(R1BIO_Uptodate, &r1_bio->state);
1842 set_bit(BIO_UPTODATE, &bio->bi_flags);
1843 return 1;
1844 }
1845
1846 static int process_checks(struct r1bio *r1_bio)
1847 {
1848 /* We have read all readable devices. If we haven't
1849 * got the block, then there is no hope left.
1850 * If we have, then we want to do a comparison
1851 * and skip the write if everything is the same.
1852 * If any blocks failed to read, then we need to
1853 * attempt an over-write
1854 */
1855 struct mddev *mddev = r1_bio->mddev;
1856 struct r1conf *conf = mddev->private;
1857 int primary;
1858 int i;
1859 int vcnt;
1860
1861 /* Fix variable parts of all bios */
1862 vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1863 for (i = 0; i < conf->raid_disks * 2; i++) {
1864 int j;
1865 int size;
1866 int uptodate;
1867 struct bio *b = r1_bio->bios[i];
1868 if (b->bi_end_io != end_sync_read)
1869 continue;
1870 /* fixup the bio for reuse, but preserve BIO_UPTODATE */
1871 uptodate = test_bit(BIO_UPTODATE, &b->bi_flags);
1872 bio_reset(b);
1873 if (!uptodate)
1874 clear_bit(BIO_UPTODATE, &b->bi_flags);
1875 b->bi_vcnt = vcnt;
1876 b->bi_size = r1_bio->sectors << 9;
1877 b->bi_sector = r1_bio->sector +
1878 conf->mirrors[i].rdev->data_offset;
1879 b->bi_bdev = conf->mirrors[i].rdev->bdev;
1880 b->bi_end_io = end_sync_read;
1881 b->bi_private = r1_bio;
1882
1883 size = b->bi_size;
1884 for (j = 0; j < vcnt ; j++) {
1885 struct bio_vec *bi;
1886 bi = &b->bi_io_vec[j];
1887 bi->bv_offset = 0;
1888 if (size > PAGE_SIZE)
1889 bi->bv_len = PAGE_SIZE;
1890 else
1891 bi->bv_len = size;
1892 size -= PAGE_SIZE;
1893 }
1894 }
1895 for (primary = 0; primary < conf->raid_disks * 2; primary++)
1896 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1897 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1898 r1_bio->bios[primary]->bi_end_io = NULL;
1899 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1900 break;
1901 }
1902 r1_bio->read_disk = primary;
1903 for (i = 0; i < conf->raid_disks * 2; i++) {
1904 int j;
1905 struct bio *pbio = r1_bio->bios[primary];
1906 struct bio *sbio = r1_bio->bios[i];
1907 int uptodate = test_bit(BIO_UPTODATE, &sbio->bi_flags);
1908
1909 if (sbio->bi_end_io != end_sync_read)
1910 continue;
1911 /* Now we can 'fixup' the BIO_UPTODATE flag */
1912 set_bit(BIO_UPTODATE, &sbio->bi_flags);
1913
1914 if (uptodate) {
1915 for (j = vcnt; j-- ; ) {
1916 struct page *p, *s;
1917 p = pbio->bi_io_vec[j].bv_page;
1918 s = sbio->bi_io_vec[j].bv_page;
1919 if (memcmp(page_address(p),
1920 page_address(s),
1921 sbio->bi_io_vec[j].bv_len))
1922 break;
1923 }
1924 } else
1925 j = 0;
1926 if (j >= 0)
1927 atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1928 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1929 && uptodate)) {
1930 /* No need to write to this device. */
1931 sbio->bi_end_io = NULL;
1932 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1933 continue;
1934 }
1935
1936 bio_copy_data(sbio, pbio);
1937 }
1938 return 0;
1939 }
1940
1941 static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1942 {
1943 struct r1conf *conf = mddev->private;
1944 int i;
1945 int disks = conf->raid_disks * 2;
1946 struct bio *bio, *wbio;
1947
1948 bio = r1_bio->bios[r1_bio->read_disk];
1949
1950 if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1951 /* ouch - failed to read all of that. */
1952 if (!fix_sync_read_error(r1_bio))
1953 return;
1954
1955 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1956 if (process_checks(r1_bio) < 0)
1957 return;
1958 /*
1959 * schedule writes
1960 */
1961 atomic_set(&r1_bio->remaining, 1);
1962 for (i = 0; i < disks ; i++) {
1963 wbio = r1_bio->bios[i];
1964 if (wbio->bi_end_io == NULL ||
1965 (wbio->bi_end_io == end_sync_read &&
1966 (i == r1_bio->read_disk ||
1967 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1968 continue;
1969
1970 wbio->bi_rw = WRITE;
1971 wbio->bi_end_io = end_sync_write;
1972 atomic_inc(&r1_bio->remaining);
1973 md_sync_acct(conf->mirrors[i].rdev->bdev, bio_sectors(wbio));
1974
1975 generic_make_request(wbio);
1976 }
1977
1978 if (atomic_dec_and_test(&r1_bio->remaining)) {
1979 /* if we're here, all write(s) have completed, so clean up */
1980 int s = r1_bio->sectors;
1981 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1982 test_bit(R1BIO_WriteError, &r1_bio->state))
1983 reschedule_retry(r1_bio);
1984 else {
1985 put_buf(r1_bio);
1986 md_done_sync(mddev, s, 1);
1987 }
1988 }
1989 }
1990
1991 /*
1992 * This is a kernel thread which:
1993 *
1994 * 1. Retries failed read operations on working mirrors.
1995 * 2. Updates the raid superblock when problems encounter.
1996 * 3. Performs writes following reads for array synchronising.
1997 */
1998
1999 static void fix_read_error(struct r1conf *conf, int read_disk,
2000 sector_t sect, int sectors)
2001 {
2002 struct mddev *mddev = conf->mddev;
2003 while(sectors) {
2004 int s = sectors;
2005 int d = read_disk;
2006 int success = 0;
2007 int start;
2008 struct md_rdev *rdev;
2009
2010 if (s > (PAGE_SIZE>>9))
2011 s = PAGE_SIZE >> 9;
2012
2013 do {
2014 /* Note: no rcu protection needed here
2015 * as this is synchronous in the raid1d thread
2016 * which is the thread that might remove
2017 * a device. If raid1d ever becomes multi-threaded....
2018 */
2019 sector_t first_bad;
2020 int bad_sectors;
2021
2022 rdev = conf->mirrors[d].rdev;
2023 if (rdev &&
2024 (test_bit(In_sync, &rdev->flags) ||
2025 (!test_bit(Faulty, &rdev->flags) &&
2026 rdev->recovery_offset >= sect + s)) &&
2027 is_badblock(rdev, sect, s,
2028 &first_bad, &bad_sectors) == 0 &&
2029 sync_page_io(rdev, sect, s<<9,
2030 conf->tmppage, READ, false))
2031 success = 1;
2032 else {
2033 d++;
2034 if (d == conf->raid_disks * 2)
2035 d = 0;
2036 }
2037 } while (!success && d != read_disk);
2038
2039 if (!success) {
2040 /* Cannot read from anywhere - mark it bad */
2041 struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2042 if (!rdev_set_badblocks(rdev, sect, s, 0))
2043 md_error(mddev, rdev);
2044 break;
2045 }
2046 /* write it back and re-read */
2047 start = d;
2048 while (d != read_disk) {
2049 if (d==0)
2050 d = conf->raid_disks * 2;
2051 d--;
2052 rdev = conf->mirrors[d].rdev;
2053 if (rdev &&
2054 test_bit(In_sync, &rdev->flags))
2055 r1_sync_page_io(rdev, sect, s,
2056 conf->tmppage, WRITE);
2057 }
2058 d = start;
2059 while (d != read_disk) {
2060 char b[BDEVNAME_SIZE];
2061 if (d==0)
2062 d = conf->raid_disks * 2;
2063 d--;
2064 rdev = conf->mirrors[d].rdev;
2065 if (rdev &&
2066 test_bit(In_sync, &rdev->flags)) {
2067 if (r1_sync_page_io(rdev, sect, s,
2068 conf->tmppage, READ)) {
2069 atomic_add(s, &rdev->corrected_errors);
2070 printk(KERN_INFO
2071 "md/raid1:%s: read error corrected "
2072 "(%d sectors at %llu on %s)\n",
2073 mdname(mddev), s,
2074 (unsigned long long)(sect +
2075 rdev->data_offset),
2076 bdevname(rdev->bdev, b));
2077 }
2078 }
2079 }
2080 sectors -= s;
2081 sect += s;
2082 }
2083 }
2084
2085 static int narrow_write_error(struct r1bio *r1_bio, int i)
2086 {
2087 struct mddev *mddev = r1_bio->mddev;
2088 struct r1conf *conf = mddev->private;
2089 struct md_rdev *rdev = conf->mirrors[i].rdev;
2090
2091 /* bio has the data to be written to device 'i' where
2092 * we just recently had a write error.
2093 * We repeatedly clone the bio and trim down to one block,
2094 * then try the write. Where the write fails we record
2095 * a bad block.
2096 * It is conceivable that the bio doesn't exactly align with
2097 * blocks. We must handle this somehow.
2098 *
2099 * We currently own a reference on the rdev.
2100 */
2101
2102 int block_sectors;
2103 sector_t sector;
2104 int sectors;
2105 int sect_to_write = r1_bio->sectors;
2106 int ok = 1;
2107
2108 if (rdev->badblocks.shift < 0)
2109 return 0;
2110
2111 block_sectors = 1 << rdev->badblocks.shift;
2112 sector = r1_bio->sector;
2113 sectors = ((sector + block_sectors)
2114 & ~(sector_t)(block_sectors - 1))
2115 - sector;
2116
2117 while (sect_to_write) {
2118 struct bio *wbio;
2119 if (sectors > sect_to_write)
2120 sectors = sect_to_write;
2121 /* Write at 'sector' for 'sectors'*/
2122
2123 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2124 unsigned vcnt = r1_bio->behind_page_count;
2125 struct bio_vec *vec = r1_bio->behind_bvecs;
2126
2127 while (!vec->bv_page) {
2128 vec++;
2129 vcnt--;
2130 }
2131
2132 wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2133 memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2134
2135 wbio->bi_vcnt = vcnt;
2136 } else {
2137 wbio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2138 }
2139
2140 wbio->bi_rw = WRITE;
2141 wbio->bi_sector = r1_bio->sector;
2142 wbio->bi_size = r1_bio->sectors << 9;
2143
2144 md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2145 wbio->bi_sector += rdev->data_offset;
2146 wbio->bi_bdev = rdev->bdev;
2147 if (submit_bio_wait(WRITE, wbio) == 0)
2148 /* failure! */
2149 ok = rdev_set_badblocks(rdev, sector,
2150 sectors, 0)
2151 && ok;
2152
2153 bio_put(wbio);
2154 sect_to_write -= sectors;
2155 sector += sectors;
2156 sectors = block_sectors;
2157 }
2158 return ok;
2159 }
2160
2161 static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2162 {
2163 int m;
2164 int s = r1_bio->sectors;
2165 for (m = 0; m < conf->raid_disks * 2 ; m++) {
2166 struct md_rdev *rdev = conf->mirrors[m].rdev;
2167 struct bio *bio = r1_bio->bios[m];
2168 if (bio->bi_end_io == NULL)
2169 continue;
2170 if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2171 test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2172 rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2173 }
2174 if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2175 test_bit(R1BIO_WriteError, &r1_bio->state)) {
2176 if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2177 md_error(conf->mddev, rdev);
2178 }
2179 }
2180 put_buf(r1_bio);
2181 md_done_sync(conf->mddev, s, 1);
2182 }
2183
2184 static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2185 {
2186 int m;
2187 for (m = 0; m < conf->raid_disks * 2 ; m++)
2188 if (r1_bio->bios[m] == IO_MADE_GOOD) {
2189 struct md_rdev *rdev = conf->mirrors[m].rdev;
2190 rdev_clear_badblocks(rdev,
2191 r1_bio->sector,
2192 r1_bio->sectors, 0);
2193 rdev_dec_pending(rdev, conf->mddev);
2194 } else if (r1_bio->bios[m] != NULL) {
2195 /* This drive got a write error. We need to
2196 * narrow down and record precise write
2197 * errors.
2198 */
2199 if (!narrow_write_error(r1_bio, m)) {
2200 md_error(conf->mddev,
2201 conf->mirrors[m].rdev);
2202 /* an I/O failed, we can't clear the bitmap */
2203 set_bit(R1BIO_Degraded, &r1_bio->state);
2204 }
2205 rdev_dec_pending(conf->mirrors[m].rdev,
2206 conf->mddev);
2207 }
2208 if (test_bit(R1BIO_WriteError, &r1_bio->state))
2209 close_write(r1_bio);
2210 raid_end_bio_io(r1_bio);
2211 }
2212
2213 static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2214 {
2215 int disk;
2216 int max_sectors;
2217 struct mddev *mddev = conf->mddev;
2218 struct bio *bio;
2219 char b[BDEVNAME_SIZE];
2220 struct md_rdev *rdev;
2221
2222 clear_bit(R1BIO_ReadError, &r1_bio->state);
2223 /* we got a read error. Maybe the drive is bad. Maybe just
2224 * the block and we can fix it.
2225 * We freeze all other IO, and try reading the block from
2226 * other devices. When we find one, we re-write
2227 * and check it that fixes the read error.
2228 * This is all done synchronously while the array is
2229 * frozen
2230 */
2231 if (mddev->ro == 0) {
2232 freeze_array(conf, 1);
2233 fix_read_error(conf, r1_bio->read_disk,
2234 r1_bio->sector, r1_bio->sectors);
2235 unfreeze_array(conf);
2236 } else
2237 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2238 rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2239
2240 bio = r1_bio->bios[r1_bio->read_disk];
2241 bdevname(bio->bi_bdev, b);
2242 read_more:
2243 disk = read_balance(conf, r1_bio, &max_sectors);
2244 if (disk == -1) {
2245 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2246 " read error for block %llu\n",
2247 mdname(mddev), b, (unsigned long long)r1_bio->sector);
2248 raid_end_bio_io(r1_bio);
2249 } else {
2250 const unsigned long do_sync
2251 = r1_bio->master_bio->bi_rw & REQ_SYNC;
2252 if (bio) {
2253 r1_bio->bios[r1_bio->read_disk] =
2254 mddev->ro ? IO_BLOCKED : NULL;
2255 bio_put(bio);
2256 }
2257 r1_bio->read_disk = disk;
2258 bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2259 md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2260 r1_bio->bios[r1_bio->read_disk] = bio;
2261 rdev = conf->mirrors[disk].rdev;
2262 printk_ratelimited(KERN_ERR
2263 "md/raid1:%s: redirecting sector %llu"
2264 " to other mirror: %s\n",
2265 mdname(mddev),
2266 (unsigned long long)r1_bio->sector,
2267 bdevname(rdev->bdev, b));
2268 bio->bi_sector = r1_bio->sector + rdev->data_offset;
2269 bio->bi_bdev = rdev->bdev;
2270 bio->bi_end_io = raid1_end_read_request;
2271 bio->bi_rw = READ | do_sync;
2272 bio->bi_private = r1_bio;
2273 if (max_sectors < r1_bio->sectors) {
2274 /* Drat - have to split this up more */
2275 struct bio *mbio = r1_bio->master_bio;
2276 int sectors_handled = (r1_bio->sector + max_sectors
2277 - mbio->bi_sector);
2278 r1_bio->sectors = max_sectors;
2279 spin_lock_irq(&conf->device_lock);
2280 if (mbio->bi_phys_segments == 0)
2281 mbio->bi_phys_segments = 2;
2282 else
2283 mbio->bi_phys_segments++;
2284 spin_unlock_irq(&conf->device_lock);
2285 generic_make_request(bio);
2286 bio = NULL;
2287
2288 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2289
2290 r1_bio->master_bio = mbio;
2291 r1_bio->sectors = bio_sectors(mbio) - sectors_handled;
2292 r1_bio->state = 0;
2293 set_bit(R1BIO_ReadError, &r1_bio->state);
2294 r1_bio->mddev = mddev;
2295 r1_bio->sector = mbio->bi_sector + sectors_handled;
2296
2297 goto read_more;
2298 } else
2299 generic_make_request(bio);
2300 }
2301 }
2302
2303 static void raid1d(struct md_thread *thread)
2304 {
2305 struct mddev *mddev = thread->mddev;
2306 struct r1bio *r1_bio;
2307 unsigned long flags;
2308 struct r1conf *conf = mddev->private;
2309 struct list_head *head = &conf->retry_list;
2310 struct blk_plug plug;
2311
2312 md_check_recovery(mddev);
2313
2314 blk_start_plug(&plug);
2315 for (;;) {
2316
2317 flush_pending_writes(conf);
2318
2319 spin_lock_irqsave(&conf->device_lock, flags);
2320 if (list_empty(head)) {
2321 spin_unlock_irqrestore(&conf->device_lock, flags);
2322 break;
2323 }
2324 r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2325 list_del(head->prev);
2326 conf->nr_queued--;
2327 spin_unlock_irqrestore(&conf->device_lock, flags);
2328
2329 mddev = r1_bio->mddev;
2330 conf = mddev->private;
2331 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2332 if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2333 test_bit(R1BIO_WriteError, &r1_bio->state))
2334 handle_sync_write_finished(conf, r1_bio);
2335 else
2336 sync_request_write(mddev, r1_bio);
2337 } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2338 test_bit(R1BIO_WriteError, &r1_bio->state))
2339 handle_write_finished(conf, r1_bio);
2340 else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2341 handle_read_error(conf, r1_bio);
2342 else
2343 /* just a partial read to be scheduled from separate
2344 * context
2345 */
2346 generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2347
2348 cond_resched();
2349 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2350 md_check_recovery(mddev);
2351 }
2352 blk_finish_plug(&plug);
2353 }
2354
2355
2356 static int init_resync(struct r1conf *conf)
2357 {
2358 int buffs;
2359
2360 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2361 BUG_ON(conf->r1buf_pool);
2362 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2363 conf->poolinfo);
2364 if (!conf->r1buf_pool)
2365 return -ENOMEM;
2366 conf->next_resync = 0;
2367 return 0;
2368 }
2369
2370 /*
2371 * perform a "sync" on one "block"
2372 *
2373 * We need to make sure that no normal I/O request - particularly write
2374 * requests - conflict with active sync requests.
2375 *
2376 * This is achieved by tracking pending requests and a 'barrier' concept
2377 * that can be installed to exclude normal IO requests.
2378 */
2379
2380 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2381 {
2382 struct r1conf *conf = mddev->private;
2383 struct r1bio *r1_bio;
2384 struct bio *bio;
2385 sector_t max_sector, nr_sectors;
2386 int disk = -1;
2387 int i;
2388 int wonly = -1;
2389 int write_targets = 0, read_targets = 0;
2390 sector_t sync_blocks;
2391 int still_degraded = 0;
2392 int good_sectors = RESYNC_SECTORS;
2393 int min_bad = 0; /* number of sectors that are bad in all devices */
2394
2395 if (!conf->r1buf_pool)
2396 if (init_resync(conf))
2397 return 0;
2398
2399 max_sector = mddev->dev_sectors;
2400 if (sector_nr >= max_sector) {
2401 /* If we aborted, we need to abort the
2402 * sync on the 'current' bitmap chunk (there will
2403 * only be one in raid1 resync.
2404 * We can find the current addess in mddev->curr_resync
2405 */
2406 if (mddev->curr_resync < max_sector) /* aborted */
2407 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2408 &sync_blocks, 1);
2409 else /* completed sync */
2410 conf->fullsync = 0;
2411
2412 bitmap_close_sync(mddev->bitmap);
2413 close_sync(conf);
2414 return 0;
2415 }
2416
2417 if (mddev->bitmap == NULL &&
2418 mddev->recovery_cp == MaxSector &&
2419 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2420 conf->fullsync == 0) {
2421 *skipped = 1;
2422 return max_sector - sector_nr;
2423 }
2424 /* before building a request, check if we can skip these blocks..
2425 * This call the bitmap_start_sync doesn't actually record anything
2426 */
2427 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2428 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2429 /* We can skip this block, and probably several more */
2430 *skipped = 1;
2431 return sync_blocks;
2432 }
2433 /*
2434 * If there is non-resync activity waiting for a turn,
2435 * and resync is going fast enough,
2436 * then let it though before starting on this new sync request.
2437 */
2438 if (!go_faster && conf->nr_waiting)
2439 msleep_interruptible(1000);
2440
2441 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2442 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2443 raise_barrier(conf);
2444
2445 conf->next_resync = sector_nr;
2446
2447 rcu_read_lock();
2448 /*
2449 * If we get a correctably read error during resync or recovery,
2450 * we might want to read from a different device. So we
2451 * flag all drives that could conceivably be read from for READ,
2452 * and any others (which will be non-In_sync devices) for WRITE.
2453 * If a read fails, we try reading from something else for which READ
2454 * is OK.
2455 */
2456
2457 r1_bio->mddev = mddev;
2458 r1_bio->sector = sector_nr;
2459 r1_bio->state = 0;
2460 set_bit(R1BIO_IsSync, &r1_bio->state);
2461
2462 for (i = 0; i < conf->raid_disks * 2; i++) {
2463 struct md_rdev *rdev;
2464 bio = r1_bio->bios[i];
2465 bio_reset(bio);
2466
2467 rdev = rcu_dereference(conf->mirrors[i].rdev);
2468 if (rdev == NULL ||
2469 test_bit(Faulty, &rdev->flags)) {
2470 if (i < conf->raid_disks)
2471 still_degraded = 1;
2472 } else if (!test_bit(In_sync, &rdev->flags)) {
2473 bio->bi_rw = WRITE;
2474 bio->bi_end_io = end_sync_write;
2475 write_targets ++;
2476 } else {
2477 /* may need to read from here */
2478 sector_t first_bad = MaxSector;
2479 int bad_sectors;
2480
2481 if (is_badblock(rdev, sector_nr, good_sectors,
2482 &first_bad, &bad_sectors)) {
2483 if (first_bad > sector_nr)
2484 good_sectors = first_bad - sector_nr;
2485 else {
2486 bad_sectors -= (sector_nr - first_bad);
2487 if (min_bad == 0 ||
2488 min_bad > bad_sectors)
2489 min_bad = bad_sectors;
2490 }
2491 }
2492 if (sector_nr < first_bad) {
2493 if (test_bit(WriteMostly, &rdev->flags)) {
2494 if (wonly < 0)
2495 wonly = i;
2496 } else {
2497 if (disk < 0)
2498 disk = i;
2499 }
2500 bio->bi_rw = READ;
2501 bio->bi_end_io = end_sync_read;
2502 read_targets++;
2503 } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2504 test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2505 !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2506 /*
2507 * The device is suitable for reading (InSync),
2508 * but has bad block(s) here. Let's try to correct them,
2509 * if we are doing resync or repair. Otherwise, leave
2510 * this device alone for this sync request.
2511 */
2512 bio->bi_rw = WRITE;
2513 bio->bi_end_io = end_sync_write;
2514 write_targets++;
2515 }
2516 }
2517 if (bio->bi_end_io) {
2518 atomic_inc(&rdev->nr_pending);
2519 bio->bi_sector = sector_nr + rdev->data_offset;
2520 bio->bi_bdev = rdev->bdev;
2521 bio->bi_private = r1_bio;
2522 }
2523 }
2524 rcu_read_unlock();
2525 if (disk < 0)
2526 disk = wonly;
2527 r1_bio->read_disk = disk;
2528
2529 if (read_targets == 0 && min_bad > 0) {
2530 /* These sectors are bad on all InSync devices, so we
2531 * need to mark them bad on all write targets
2532 */
2533 int ok = 1;
2534 for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2535 if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2536 struct md_rdev *rdev = conf->mirrors[i].rdev;
2537 ok = rdev_set_badblocks(rdev, sector_nr,
2538 min_bad, 0
2539 ) && ok;
2540 }
2541 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2542 *skipped = 1;
2543 put_buf(r1_bio);
2544
2545 if (!ok) {
2546 /* Cannot record the badblocks, so need to
2547 * abort the resync.
2548 * If there are multiple read targets, could just
2549 * fail the really bad ones ???
2550 */
2551 conf->recovery_disabled = mddev->recovery_disabled;
2552 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2553 return 0;
2554 } else
2555 return min_bad;
2556
2557 }
2558 if (min_bad > 0 && min_bad < good_sectors) {
2559 /* only resync enough to reach the next bad->good
2560 * transition */
2561 good_sectors = min_bad;
2562 }
2563
2564 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2565 /* extra read targets are also write targets */
2566 write_targets += read_targets-1;
2567
2568 if (write_targets == 0 || read_targets == 0) {
2569 /* There is nowhere to write, so all non-sync
2570 * drives must be failed - so we are finished
2571 */
2572 sector_t rv;
2573 if (min_bad > 0)
2574 max_sector = sector_nr + min_bad;
2575 rv = max_sector - sector_nr;
2576 *skipped = 1;
2577 put_buf(r1_bio);
2578 return rv;
2579 }
2580
2581 if (max_sector > mddev->resync_max)
2582 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2583 if (max_sector > sector_nr + good_sectors)
2584 max_sector = sector_nr + good_sectors;
2585 nr_sectors = 0;
2586 sync_blocks = 0;
2587 do {
2588 struct page *page;
2589 int len = PAGE_SIZE;
2590 if (sector_nr + (len>>9) > max_sector)
2591 len = (max_sector - sector_nr) << 9;
2592 if (len == 0)
2593 break;
2594 if (sync_blocks == 0) {
2595 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2596 &sync_blocks, still_degraded) &&
2597 !conf->fullsync &&
2598 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2599 break;
2600 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2601 if ((len >> 9) > sync_blocks)
2602 len = sync_blocks<<9;
2603 }
2604
2605 for (i = 0 ; i < conf->raid_disks * 2; i++) {
2606 bio = r1_bio->bios[i];
2607 if (bio->bi_end_io) {
2608 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2609 if (bio_add_page(bio, page, len, 0) == 0) {
2610 /* stop here */
2611 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2612 while (i > 0) {
2613 i--;
2614 bio = r1_bio->bios[i];
2615 if (bio->bi_end_io==NULL)
2616 continue;
2617 /* remove last page from this bio */
2618 bio->bi_vcnt--;
2619 bio->bi_size -= len;
2620 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2621 }
2622 goto bio_full;
2623 }
2624 }
2625 }
2626 nr_sectors += len>>9;
2627 sector_nr += len>>9;
2628 sync_blocks -= (len>>9);
2629 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2630 bio_full:
2631 r1_bio->sectors = nr_sectors;
2632
2633 /* For a user-requested sync, we read all readable devices and do a
2634 * compare
2635 */
2636 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2637 atomic_set(&r1_bio->remaining, read_targets);
2638 for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2639 bio = r1_bio->bios[i];
2640 if (bio->bi_end_io == end_sync_read) {
2641 read_targets--;
2642 md_sync_acct(bio->bi_bdev, nr_sectors);
2643 generic_make_request(bio);
2644 }
2645 }
2646 } else {
2647 atomic_set(&r1_bio->remaining, 1);
2648 bio = r1_bio->bios[r1_bio->read_disk];
2649 md_sync_acct(bio->bi_bdev, nr_sectors);
2650 generic_make_request(bio);
2651
2652 }
2653 return nr_sectors;
2654 }
2655
2656 static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2657 {
2658 if (sectors)
2659 return sectors;
2660
2661 return mddev->dev_sectors;
2662 }
2663
2664 static struct r1conf *setup_conf(struct mddev *mddev)
2665 {
2666 struct r1conf *conf;
2667 int i;
2668 struct raid1_info *disk;
2669 struct md_rdev *rdev;
2670 int err = -ENOMEM;
2671
2672 conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2673 if (!conf)
2674 goto abort;
2675
2676 conf->mirrors = kzalloc(sizeof(struct raid1_info)
2677 * mddev->raid_disks * 2,
2678 GFP_KERNEL);
2679 if (!conf->mirrors)
2680 goto abort;
2681
2682 conf->tmppage = alloc_page(GFP_KERNEL);
2683 if (!conf->tmppage)
2684 goto abort;
2685
2686 conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2687 if (!conf->poolinfo)
2688 goto abort;
2689 conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2690 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2691 r1bio_pool_free,
2692 conf->poolinfo);
2693 if (!conf->r1bio_pool)
2694 goto abort;
2695
2696 conf->poolinfo->mddev = mddev;
2697
2698 err = -EINVAL;
2699 spin_lock_init(&conf->device_lock);
2700 rdev_for_each(rdev, mddev) {
2701 struct request_queue *q;
2702 int disk_idx = rdev->raid_disk;
2703 if (disk_idx >= mddev->raid_disks
2704 || disk_idx < 0)
2705 continue;
2706 if (test_bit(Replacement, &rdev->flags))
2707 disk = conf->mirrors + mddev->raid_disks + disk_idx;
2708 else
2709 disk = conf->mirrors + disk_idx;
2710
2711 if (disk->rdev)
2712 goto abort;
2713 disk->rdev = rdev;
2714 q = bdev_get_queue(rdev->bdev);
2715 if (q->merge_bvec_fn)