Loading...
1/*
2 * raid10.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 2000-2004 Neil Brown
5 *
6 * RAID-10 support for md.
7 *
8 * Base on code in raid1.c. See raid1.c for further copyright information.
9 *
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20
21#include <linux/slab.h>
22#include <linux/delay.h>
23#include <linux/blkdev.h>
24#include <linux/module.h>
25#include <linux/seq_file.h>
26#include <linux/ratelimit.h>
27#include <linux/kthread.h>
28#include "md.h"
29#include "raid10.h"
30#include "raid0.h"
31#include "bitmap.h"
32
33/*
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
36 * chunk_size
37 * raid_disks
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
41 * use_far_sets (stored in bit 17 of layout )
42 * use_far_sets_bugfixed (stored in bit 18 of layout )
43 *
44 * The data to be stored is divided into chunks using chunksize. Each device
45 * is divided into far_copies sections. In each section, chunks are laid out
46 * in a style similar to raid0, but near_copies copies of each chunk is stored
47 * (each on a different drive). The starting device for each section is offset
48 * near_copies from the starting device of the previous section. Thus there
49 * are (near_copies * far_copies) of each chunk, and each is on a different
50 * drive. near_copies and far_copies must be at least one, and their product
51 * is at most raid_disks.
52 *
53 * If far_offset is true, then the far_copies are handled a bit differently.
54 * The copies are still in different stripes, but instead of being very far
55 * apart on disk, there are adjacent stripes.
56 *
57 * The far and offset algorithms are handled slightly differently if
58 * 'use_far_sets' is true. In this case, the array's devices are grouped into
59 * sets that are (near_copies * far_copies) in size. The far copied stripes
60 * are still shifted by 'near_copies' devices, but this shifting stays confined
61 * to the set rather than the entire array. This is done to improve the number
62 * of device combinations that can fail without causing the array to fail.
63 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
64 * on a device):
65 * A B C D A B C D E
66 * ... ...
67 * D A B C E A B C D
68 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
69 * [A B] [C D] [A B] [C D E]
70 * |...| |...| |...| | ... |
71 * [B A] [D C] [B A] [E C D]
72 */
73
74/*
75 * Number of guaranteed r10bios in case of extreme VM load:
76 */
77#define NR_RAID10_BIOS 256
78
79/* when we get a read error on a read-only array, we redirect to another
80 * device without failing the first device, or trying to over-write to
81 * correct the read error. To keep track of bad blocks on a per-bio
82 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
83 */
84#define IO_BLOCKED ((struct bio *)1)
85/* When we successfully write to a known bad-block, we need to remove the
86 * bad-block marking which must be done from process context. So we record
87 * the success by setting devs[n].bio to IO_MADE_GOOD
88 */
89#define IO_MADE_GOOD ((struct bio *)2)
90
91#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
92
93/* When there are this many requests queued to be written by
94 * the raid10 thread, we become 'congested' to provide back-pressure
95 * for writeback.
96 */
97static int max_queued_requests = 1024;
98
99static void allow_barrier(struct r10conf *conf);
100static void lower_barrier(struct r10conf *conf);
101static int _enough(struct r10conf *conf, int previous, int ignore);
102static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
103 int *skipped);
104static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
105static void end_reshape_write(struct bio *bio);
106static void end_reshape(struct r10conf *conf);
107
108static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
109{
110 struct r10conf *conf = data;
111 int size = offsetof(struct r10bio, devs[conf->copies]);
112
113 /* allocate a r10bio with room for raid_disks entries in the
114 * bios array */
115 return kzalloc(size, gfp_flags);
116}
117
118static void r10bio_pool_free(void *r10_bio, void *data)
119{
120 kfree(r10_bio);
121}
122
123/* Maximum size of each resync request */
124#define RESYNC_BLOCK_SIZE (64*1024)
125#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
126/* amount of memory to reserve for resync requests */
127#define RESYNC_WINDOW (1024*1024)
128/* maximum number of concurrent requests, memory permitting */
129#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
130
131/*
132 * When performing a resync, we need to read and compare, so
133 * we need as many pages are there are copies.
134 * When performing a recovery, we need 2 bios, one for read,
135 * one for write (we recover only one drive per r10buf)
136 *
137 */
138static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
139{
140 struct r10conf *conf = data;
141 struct page *page;
142 struct r10bio *r10_bio;
143 struct bio *bio;
144 int i, j;
145 int nalloc;
146
147 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
148 if (!r10_bio)
149 return NULL;
150
151 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
152 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
153 nalloc = conf->copies; /* resync */
154 else
155 nalloc = 2; /* recovery */
156
157 /*
158 * Allocate bios.
159 */
160 for (j = nalloc ; j-- ; ) {
161 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
162 if (!bio)
163 goto out_free_bio;
164 r10_bio->devs[j].bio = bio;
165 if (!conf->have_replacement)
166 continue;
167 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
168 if (!bio)
169 goto out_free_bio;
170 r10_bio->devs[j].repl_bio = bio;
171 }
172 /*
173 * Allocate RESYNC_PAGES data pages and attach them
174 * where needed.
175 */
176 for (j = 0 ; j < nalloc; j++) {
177 struct bio *rbio = r10_bio->devs[j].repl_bio;
178 bio = r10_bio->devs[j].bio;
179 for (i = 0; i < RESYNC_PAGES; i++) {
180 if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
181 &conf->mddev->recovery)) {
182 /* we can share bv_page's during recovery
183 * and reshape */
184 struct bio *rbio = r10_bio->devs[0].bio;
185 page = rbio->bi_io_vec[i].bv_page;
186 get_page(page);
187 } else
188 page = alloc_page(gfp_flags);
189 if (unlikely(!page))
190 goto out_free_pages;
191
192 bio->bi_io_vec[i].bv_page = page;
193 if (rbio)
194 rbio->bi_io_vec[i].bv_page = page;
195 }
196 }
197
198 return r10_bio;
199
200out_free_pages:
201 for ( ; i > 0 ; i--)
202 safe_put_page(bio->bi_io_vec[i-1].bv_page);
203 while (j--)
204 for (i = 0; i < RESYNC_PAGES ; i++)
205 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
206 j = 0;
207out_free_bio:
208 for ( ; j < nalloc; j++) {
209 if (r10_bio->devs[j].bio)
210 bio_put(r10_bio->devs[j].bio);
211 if (r10_bio->devs[j].repl_bio)
212 bio_put(r10_bio->devs[j].repl_bio);
213 }
214 r10bio_pool_free(r10_bio, conf);
215 return NULL;
216}
217
218static void r10buf_pool_free(void *__r10_bio, void *data)
219{
220 int i;
221 struct r10conf *conf = data;
222 struct r10bio *r10bio = __r10_bio;
223 int j;
224
225 for (j=0; j < conf->copies; j++) {
226 struct bio *bio = r10bio->devs[j].bio;
227 if (bio) {
228 for (i = 0; i < RESYNC_PAGES; i++) {
229 safe_put_page(bio->bi_io_vec[i].bv_page);
230 bio->bi_io_vec[i].bv_page = NULL;
231 }
232 bio_put(bio);
233 }
234 bio = r10bio->devs[j].repl_bio;
235 if (bio)
236 bio_put(bio);
237 }
238 r10bio_pool_free(r10bio, conf);
239}
240
241static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
242{
243 int i;
244
245 for (i = 0; i < conf->copies; i++) {
246 struct bio **bio = & r10_bio->devs[i].bio;
247 if (!BIO_SPECIAL(*bio))
248 bio_put(*bio);
249 *bio = NULL;
250 bio = &r10_bio->devs[i].repl_bio;
251 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
252 bio_put(*bio);
253 *bio = NULL;
254 }
255}
256
257static void free_r10bio(struct r10bio *r10_bio)
258{
259 struct r10conf *conf = r10_bio->mddev->private;
260
261 put_all_bios(conf, r10_bio);
262 mempool_free(r10_bio, conf->r10bio_pool);
263}
264
265static void put_buf(struct r10bio *r10_bio)
266{
267 struct r10conf *conf = r10_bio->mddev->private;
268
269 mempool_free(r10_bio, conf->r10buf_pool);
270
271 lower_barrier(conf);
272}
273
274static void reschedule_retry(struct r10bio *r10_bio)
275{
276 unsigned long flags;
277 struct mddev *mddev = r10_bio->mddev;
278 struct r10conf *conf = mddev->private;
279
280 spin_lock_irqsave(&conf->device_lock, flags);
281 list_add(&r10_bio->retry_list, &conf->retry_list);
282 conf->nr_queued ++;
283 spin_unlock_irqrestore(&conf->device_lock, flags);
284
285 /* wake up frozen array... */
286 wake_up(&conf->wait_barrier);
287
288 md_wakeup_thread(mddev->thread);
289}
290
291/*
292 * raid_end_bio_io() is called when we have finished servicing a mirrored
293 * operation and are ready to return a success/failure code to the buffer
294 * cache layer.
295 */
296static void raid_end_bio_io(struct r10bio *r10_bio)
297{
298 struct bio *bio = r10_bio->master_bio;
299 int done;
300 struct r10conf *conf = r10_bio->mddev->private;
301
302 if (bio->bi_phys_segments) {
303 unsigned long flags;
304 spin_lock_irqsave(&conf->device_lock, flags);
305 bio->bi_phys_segments--;
306 done = (bio->bi_phys_segments == 0);
307 spin_unlock_irqrestore(&conf->device_lock, flags);
308 } else
309 done = 1;
310 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
311 bio->bi_error = -EIO;
312 if (done) {
313 bio_endio(bio);
314 /*
315 * Wake up any possible resync thread that waits for the device
316 * to go idle.
317 */
318 allow_barrier(conf);
319 }
320 free_r10bio(r10_bio);
321}
322
323/*
324 * Update disk head position estimator based on IRQ completion info.
325 */
326static inline void update_head_pos(int slot, struct r10bio *r10_bio)
327{
328 struct r10conf *conf = r10_bio->mddev->private;
329
330 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
331 r10_bio->devs[slot].addr + (r10_bio->sectors);
332}
333
334/*
335 * Find the disk number which triggered given bio
336 */
337static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
338 struct bio *bio, int *slotp, int *replp)
339{
340 int slot;
341 int repl = 0;
342
343 for (slot = 0; slot < conf->copies; slot++) {
344 if (r10_bio->devs[slot].bio == bio)
345 break;
346 if (r10_bio->devs[slot].repl_bio == bio) {
347 repl = 1;
348 break;
349 }
350 }
351
352 BUG_ON(slot == conf->copies);
353 update_head_pos(slot, r10_bio);
354
355 if (slotp)
356 *slotp = slot;
357 if (replp)
358 *replp = repl;
359 return r10_bio->devs[slot].devnum;
360}
361
362static void raid10_end_read_request(struct bio *bio)
363{
364 int uptodate = !bio->bi_error;
365 struct r10bio *r10_bio = bio->bi_private;
366 int slot, dev;
367 struct md_rdev *rdev;
368 struct r10conf *conf = r10_bio->mddev->private;
369
370 slot = r10_bio->read_slot;
371 dev = r10_bio->devs[slot].devnum;
372 rdev = r10_bio->devs[slot].rdev;
373 /*
374 * this branch is our 'one mirror IO has finished' event handler:
375 */
376 update_head_pos(slot, r10_bio);
377
378 if (uptodate) {
379 /*
380 * Set R10BIO_Uptodate in our master bio, so that
381 * we will return a good error code to the higher
382 * levels even if IO on some other mirrored buffer fails.
383 *
384 * The 'master' represents the composite IO operation to
385 * user-side. So if something waits for IO, then it will
386 * wait for the 'master' bio.
387 */
388 set_bit(R10BIO_Uptodate, &r10_bio->state);
389 } else {
390 /* If all other devices that store this block have
391 * failed, we want to return the error upwards rather
392 * than fail the last device. Here we redefine
393 * "uptodate" to mean "Don't want to retry"
394 */
395 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
396 rdev->raid_disk))
397 uptodate = 1;
398 }
399 if (uptodate) {
400 raid_end_bio_io(r10_bio);
401 rdev_dec_pending(rdev, conf->mddev);
402 } else {
403 /*
404 * oops, read error - keep the refcount on the rdev
405 */
406 char b[BDEVNAME_SIZE];
407 printk_ratelimited(KERN_ERR
408 "md/raid10:%s: %s: rescheduling sector %llu\n",
409 mdname(conf->mddev),
410 bdevname(rdev->bdev, b),
411 (unsigned long long)r10_bio->sector);
412 set_bit(R10BIO_ReadError, &r10_bio->state);
413 reschedule_retry(r10_bio);
414 }
415}
416
417static void close_write(struct r10bio *r10_bio)
418{
419 /* clear the bitmap if all writes complete successfully */
420 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
421 r10_bio->sectors,
422 !test_bit(R10BIO_Degraded, &r10_bio->state),
423 0);
424 md_write_end(r10_bio->mddev);
425}
426
427static void one_write_done(struct r10bio *r10_bio)
428{
429 if (atomic_dec_and_test(&r10_bio->remaining)) {
430 if (test_bit(R10BIO_WriteError, &r10_bio->state))
431 reschedule_retry(r10_bio);
432 else {
433 close_write(r10_bio);
434 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
435 reschedule_retry(r10_bio);
436 else
437 raid_end_bio_io(r10_bio);
438 }
439 }
440}
441
442static void raid10_end_write_request(struct bio *bio)
443{
444 struct r10bio *r10_bio = bio->bi_private;
445 int dev;
446 int dec_rdev = 1;
447 struct r10conf *conf = r10_bio->mddev->private;
448 int slot, repl;
449 struct md_rdev *rdev = NULL;
450
451 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
452
453 if (repl)
454 rdev = conf->mirrors[dev].replacement;
455 if (!rdev) {
456 smp_rmb();
457 repl = 0;
458 rdev = conf->mirrors[dev].rdev;
459 }
460 /*
461 * this branch is our 'one mirror IO has finished' event handler:
462 */
463 if (bio->bi_error) {
464 if (repl)
465 /* Never record new bad blocks to replacement,
466 * just fail it.
467 */
468 md_error(rdev->mddev, rdev);
469 else {
470 set_bit(WriteErrorSeen, &rdev->flags);
471 if (!test_and_set_bit(WantReplacement, &rdev->flags))
472 set_bit(MD_RECOVERY_NEEDED,
473 &rdev->mddev->recovery);
474 set_bit(R10BIO_WriteError, &r10_bio->state);
475 dec_rdev = 0;
476 }
477 } else {
478 /*
479 * Set R10BIO_Uptodate in our master bio, so that
480 * we will return a good error code for to the higher
481 * levels even if IO on some other mirrored buffer fails.
482 *
483 * The 'master' represents the composite IO operation to
484 * user-side. So if something waits for IO, then it will
485 * wait for the 'master' bio.
486 */
487 sector_t first_bad;
488 int bad_sectors;
489
490 /*
491 * Do not set R10BIO_Uptodate if the current device is
492 * rebuilding or Faulty. This is because we cannot use
493 * such device for properly reading the data back (we could
494 * potentially use it, if the current write would have felt
495 * before rdev->recovery_offset, but for simplicity we don't
496 * check this here.
497 */
498 if (test_bit(In_sync, &rdev->flags) &&
499 !test_bit(Faulty, &rdev->flags))
500 set_bit(R10BIO_Uptodate, &r10_bio->state);
501
502 /* Maybe we can clear some bad blocks. */
503 if (is_badblock(rdev,
504 r10_bio->devs[slot].addr,
505 r10_bio->sectors,
506 &first_bad, &bad_sectors)) {
507 bio_put(bio);
508 if (repl)
509 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
510 else
511 r10_bio->devs[slot].bio = IO_MADE_GOOD;
512 dec_rdev = 0;
513 set_bit(R10BIO_MadeGood, &r10_bio->state);
514 }
515 }
516
517 /*
518 *
519 * Let's see if all mirrored write operations have finished
520 * already.
521 */
522 one_write_done(r10_bio);
523 if (dec_rdev)
524 rdev_dec_pending(rdev, conf->mddev);
525}
526
527/*
528 * RAID10 layout manager
529 * As well as the chunksize and raid_disks count, there are two
530 * parameters: near_copies and far_copies.
531 * near_copies * far_copies must be <= raid_disks.
532 * Normally one of these will be 1.
533 * If both are 1, we get raid0.
534 * If near_copies == raid_disks, we get raid1.
535 *
536 * Chunks are laid out in raid0 style with near_copies copies of the
537 * first chunk, followed by near_copies copies of the next chunk and
538 * so on.
539 * If far_copies > 1, then after 1/far_copies of the array has been assigned
540 * as described above, we start again with a device offset of near_copies.
541 * So we effectively have another copy of the whole array further down all
542 * the drives, but with blocks on different drives.
543 * With this layout, and block is never stored twice on the one device.
544 *
545 * raid10_find_phys finds the sector offset of a given virtual sector
546 * on each device that it is on.
547 *
548 * raid10_find_virt does the reverse mapping, from a device and a
549 * sector offset to a virtual address
550 */
551
552static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
553{
554 int n,f;
555 sector_t sector;
556 sector_t chunk;
557 sector_t stripe;
558 int dev;
559 int slot = 0;
560 int last_far_set_start, last_far_set_size;
561
562 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
563 last_far_set_start *= geo->far_set_size;
564
565 last_far_set_size = geo->far_set_size;
566 last_far_set_size += (geo->raid_disks % geo->far_set_size);
567
568 /* now calculate first sector/dev */
569 chunk = r10bio->sector >> geo->chunk_shift;
570 sector = r10bio->sector & geo->chunk_mask;
571
572 chunk *= geo->near_copies;
573 stripe = chunk;
574 dev = sector_div(stripe, geo->raid_disks);
575 if (geo->far_offset)
576 stripe *= geo->far_copies;
577
578 sector += stripe << geo->chunk_shift;
579
580 /* and calculate all the others */
581 for (n = 0; n < geo->near_copies; n++) {
582 int d = dev;
583 int set;
584 sector_t s = sector;
585 r10bio->devs[slot].devnum = d;
586 r10bio->devs[slot].addr = s;
587 slot++;
588
589 for (f = 1; f < geo->far_copies; f++) {
590 set = d / geo->far_set_size;
591 d += geo->near_copies;
592
593 if ((geo->raid_disks % geo->far_set_size) &&
594 (d > last_far_set_start)) {
595 d -= last_far_set_start;
596 d %= last_far_set_size;
597 d += last_far_set_start;
598 } else {
599 d %= geo->far_set_size;
600 d += geo->far_set_size * set;
601 }
602 s += geo->stride;
603 r10bio->devs[slot].devnum = d;
604 r10bio->devs[slot].addr = s;
605 slot++;
606 }
607 dev++;
608 if (dev >= geo->raid_disks) {
609 dev = 0;
610 sector += (geo->chunk_mask + 1);
611 }
612 }
613}
614
615static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
616{
617 struct geom *geo = &conf->geo;
618
619 if (conf->reshape_progress != MaxSector &&
620 ((r10bio->sector >= conf->reshape_progress) !=
621 conf->mddev->reshape_backwards)) {
622 set_bit(R10BIO_Previous, &r10bio->state);
623 geo = &conf->prev;
624 } else
625 clear_bit(R10BIO_Previous, &r10bio->state);
626
627 __raid10_find_phys(geo, r10bio);
628}
629
630static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
631{
632 sector_t offset, chunk, vchunk;
633 /* Never use conf->prev as this is only called during resync
634 * or recovery, so reshape isn't happening
635 */
636 struct geom *geo = &conf->geo;
637 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
638 int far_set_size = geo->far_set_size;
639 int last_far_set_start;
640
641 if (geo->raid_disks % geo->far_set_size) {
642 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
643 last_far_set_start *= geo->far_set_size;
644
645 if (dev >= last_far_set_start) {
646 far_set_size = geo->far_set_size;
647 far_set_size += (geo->raid_disks % geo->far_set_size);
648 far_set_start = last_far_set_start;
649 }
650 }
651
652 offset = sector & geo->chunk_mask;
653 if (geo->far_offset) {
654 int fc;
655 chunk = sector >> geo->chunk_shift;
656 fc = sector_div(chunk, geo->far_copies);
657 dev -= fc * geo->near_copies;
658 if (dev < far_set_start)
659 dev += far_set_size;
660 } else {
661 while (sector >= geo->stride) {
662 sector -= geo->stride;
663 if (dev < (geo->near_copies + far_set_start))
664 dev += far_set_size - geo->near_copies;
665 else
666 dev -= geo->near_copies;
667 }
668 chunk = sector >> geo->chunk_shift;
669 }
670 vchunk = chunk * geo->raid_disks + dev;
671 sector_div(vchunk, geo->near_copies);
672 return (vchunk << geo->chunk_shift) + offset;
673}
674
675/*
676 * This routine returns the disk from which the requested read should
677 * be done. There is a per-array 'next expected sequential IO' sector
678 * number - if this matches on the next IO then we use the last disk.
679 * There is also a per-disk 'last know head position' sector that is
680 * maintained from IRQ contexts, both the normal and the resync IO
681 * completion handlers update this position correctly. If there is no
682 * perfect sequential match then we pick the disk whose head is closest.
683 *
684 * If there are 2 mirrors in the same 2 devices, performance degrades
685 * because position is mirror, not device based.
686 *
687 * The rdev for the device selected will have nr_pending incremented.
688 */
689
690/*
691 * FIXME: possibly should rethink readbalancing and do it differently
692 * depending on near_copies / far_copies geometry.
693 */
694static struct md_rdev *read_balance(struct r10conf *conf,
695 struct r10bio *r10_bio,
696 int *max_sectors)
697{
698 const sector_t this_sector = r10_bio->sector;
699 int disk, slot;
700 int sectors = r10_bio->sectors;
701 int best_good_sectors;
702 sector_t new_distance, best_dist;
703 struct md_rdev *best_rdev, *rdev = NULL;
704 int do_balance;
705 int best_slot;
706 struct geom *geo = &conf->geo;
707
708 raid10_find_phys(conf, r10_bio);
709 rcu_read_lock();
710retry:
711 sectors = r10_bio->sectors;
712 best_slot = -1;
713 best_rdev = NULL;
714 best_dist = MaxSector;
715 best_good_sectors = 0;
716 do_balance = 1;
717 /*
718 * Check if we can balance. We can balance on the whole
719 * device if no resync is going on (recovery is ok), or below
720 * the resync window. We take the first readable disk when
721 * above the resync window.
722 */
723 if (conf->mddev->recovery_cp < MaxSector
724 && (this_sector + sectors >= conf->next_resync))
725 do_balance = 0;
726
727 for (slot = 0; slot < conf->copies ; slot++) {
728 sector_t first_bad;
729 int bad_sectors;
730 sector_t dev_sector;
731
732 if (r10_bio->devs[slot].bio == IO_BLOCKED)
733 continue;
734 disk = r10_bio->devs[slot].devnum;
735 rdev = rcu_dereference(conf->mirrors[disk].replacement);
736 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
737 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
738 rdev = rcu_dereference(conf->mirrors[disk].rdev);
739 if (rdev == NULL ||
740 test_bit(Faulty, &rdev->flags))
741 continue;
742 if (!test_bit(In_sync, &rdev->flags) &&
743 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
744 continue;
745
746 dev_sector = r10_bio->devs[slot].addr;
747 if (is_badblock(rdev, dev_sector, sectors,
748 &first_bad, &bad_sectors)) {
749 if (best_dist < MaxSector)
750 /* Already have a better slot */
751 continue;
752 if (first_bad <= dev_sector) {
753 /* Cannot read here. If this is the
754 * 'primary' device, then we must not read
755 * beyond 'bad_sectors' from another device.
756 */
757 bad_sectors -= (dev_sector - first_bad);
758 if (!do_balance && sectors > bad_sectors)
759 sectors = bad_sectors;
760 if (best_good_sectors > sectors)
761 best_good_sectors = sectors;
762 } else {
763 sector_t good_sectors =
764 first_bad - dev_sector;
765 if (good_sectors > best_good_sectors) {
766 best_good_sectors = good_sectors;
767 best_slot = slot;
768 best_rdev = rdev;
769 }
770 if (!do_balance)
771 /* Must read from here */
772 break;
773 }
774 continue;
775 } else
776 best_good_sectors = sectors;
777
778 if (!do_balance)
779 break;
780
781 /* This optimisation is debatable, and completely destroys
782 * sequential read speed for 'far copies' arrays. So only
783 * keep it for 'near' arrays, and review those later.
784 */
785 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
786 break;
787
788 /* for far > 1 always use the lowest address */
789 if (geo->far_copies > 1)
790 new_distance = r10_bio->devs[slot].addr;
791 else
792 new_distance = abs(r10_bio->devs[slot].addr -
793 conf->mirrors[disk].head_position);
794 if (new_distance < best_dist) {
795 best_dist = new_distance;
796 best_slot = slot;
797 best_rdev = rdev;
798 }
799 }
800 if (slot >= conf->copies) {
801 slot = best_slot;
802 rdev = best_rdev;
803 }
804
805 if (slot >= 0) {
806 atomic_inc(&rdev->nr_pending);
807 if (test_bit(Faulty, &rdev->flags)) {
808 /* Cannot risk returning a device that failed
809 * before we inc'ed nr_pending
810 */
811 rdev_dec_pending(rdev, conf->mddev);
812 goto retry;
813 }
814 r10_bio->read_slot = slot;
815 } else
816 rdev = NULL;
817 rcu_read_unlock();
818 *max_sectors = best_good_sectors;
819
820 return rdev;
821}
822
823static int raid10_congested(struct mddev *mddev, int bits)
824{
825 struct r10conf *conf = mddev->private;
826 int i, ret = 0;
827
828 if ((bits & (1 << WB_async_congested)) &&
829 conf->pending_count >= max_queued_requests)
830 return 1;
831
832 rcu_read_lock();
833 for (i = 0;
834 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
835 && ret == 0;
836 i++) {
837 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
838 if (rdev && !test_bit(Faulty, &rdev->flags)) {
839 struct request_queue *q = bdev_get_queue(rdev->bdev);
840
841 ret |= bdi_congested(&q->backing_dev_info, bits);
842 }
843 }
844 rcu_read_unlock();
845 return ret;
846}
847
848static void flush_pending_writes(struct r10conf *conf)
849{
850 /* Any writes that have been queued but are awaiting
851 * bitmap updates get flushed here.
852 */
853 spin_lock_irq(&conf->device_lock);
854
855 if (conf->pending_bio_list.head) {
856 struct bio *bio;
857 bio = bio_list_get(&conf->pending_bio_list);
858 conf->pending_count = 0;
859 spin_unlock_irq(&conf->device_lock);
860 /* flush any pending bitmap writes to disk
861 * before proceeding w/ I/O */
862 bitmap_unplug(conf->mddev->bitmap);
863 wake_up(&conf->wait_barrier);
864
865 while (bio) { /* submit pending writes */
866 struct bio *next = bio->bi_next;
867 bio->bi_next = NULL;
868 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
869 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
870 /* Just ignore it */
871 bio_endio(bio);
872 else
873 generic_make_request(bio);
874 bio = next;
875 }
876 } else
877 spin_unlock_irq(&conf->device_lock);
878}
879
880/* Barriers....
881 * Sometimes we need to suspend IO while we do something else,
882 * either some resync/recovery, or reconfigure the array.
883 * To do this we raise a 'barrier'.
884 * The 'barrier' is a counter that can be raised multiple times
885 * to count how many activities are happening which preclude
886 * normal IO.
887 * We can only raise the barrier if there is no pending IO.
888 * i.e. if nr_pending == 0.
889 * We choose only to raise the barrier if no-one is waiting for the
890 * barrier to go down. This means that as soon as an IO request
891 * is ready, no other operations which require a barrier will start
892 * until the IO request has had a chance.
893 *
894 * So: regular IO calls 'wait_barrier'. When that returns there
895 * is no backgroup IO happening, It must arrange to call
896 * allow_barrier when it has finished its IO.
897 * backgroup IO calls must call raise_barrier. Once that returns
898 * there is no normal IO happeing. It must arrange to call
899 * lower_barrier when the particular background IO completes.
900 */
901
902static void raise_barrier(struct r10conf *conf, int force)
903{
904 BUG_ON(force && !conf->barrier);
905 spin_lock_irq(&conf->resync_lock);
906
907 /* Wait until no block IO is waiting (unless 'force') */
908 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
909 conf->resync_lock);
910
911 /* block any new IO from starting */
912 conf->barrier++;
913
914 /* Now wait for all pending IO to complete */
915 wait_event_lock_irq(conf->wait_barrier,
916 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
917 conf->resync_lock);
918
919 spin_unlock_irq(&conf->resync_lock);
920}
921
922static void lower_barrier(struct r10conf *conf)
923{
924 unsigned long flags;
925 spin_lock_irqsave(&conf->resync_lock, flags);
926 conf->barrier--;
927 spin_unlock_irqrestore(&conf->resync_lock, flags);
928 wake_up(&conf->wait_barrier);
929}
930
931static void wait_barrier(struct r10conf *conf)
932{
933 spin_lock_irq(&conf->resync_lock);
934 if (conf->barrier) {
935 conf->nr_waiting++;
936 /* Wait for the barrier to drop.
937 * However if there are already pending
938 * requests (preventing the barrier from
939 * rising completely), and the
940 * pre-process bio queue isn't empty,
941 * then don't wait, as we need to empty
942 * that queue to get the nr_pending
943 * count down.
944 */
945 wait_event_lock_irq(conf->wait_barrier,
946 !conf->barrier ||
947 (conf->nr_pending &&
948 current->bio_list &&
949 !bio_list_empty(current->bio_list)),
950 conf->resync_lock);
951 conf->nr_waiting--;
952 }
953 conf->nr_pending++;
954 spin_unlock_irq(&conf->resync_lock);
955}
956
957static void allow_barrier(struct r10conf *conf)
958{
959 unsigned long flags;
960 spin_lock_irqsave(&conf->resync_lock, flags);
961 conf->nr_pending--;
962 spin_unlock_irqrestore(&conf->resync_lock, flags);
963 wake_up(&conf->wait_barrier);
964}
965
966static void freeze_array(struct r10conf *conf, int extra)
967{
968 /* stop syncio and normal IO and wait for everything to
969 * go quiet.
970 * We increment barrier and nr_waiting, and then
971 * wait until nr_pending match nr_queued+extra
972 * This is called in the context of one normal IO request
973 * that has failed. Thus any sync request that might be pending
974 * will be blocked by nr_pending, and we need to wait for
975 * pending IO requests to complete or be queued for re-try.
976 * Thus the number queued (nr_queued) plus this request (extra)
977 * must match the number of pending IOs (nr_pending) before
978 * we continue.
979 */
980 spin_lock_irq(&conf->resync_lock);
981 conf->barrier++;
982 conf->nr_waiting++;
983 wait_event_lock_irq_cmd(conf->wait_barrier,
984 conf->nr_pending == conf->nr_queued+extra,
985 conf->resync_lock,
986 flush_pending_writes(conf));
987
988 spin_unlock_irq(&conf->resync_lock);
989}
990
991static void unfreeze_array(struct r10conf *conf)
992{
993 /* reverse the effect of the freeze */
994 spin_lock_irq(&conf->resync_lock);
995 conf->barrier--;
996 conf->nr_waiting--;
997 wake_up(&conf->wait_barrier);
998 spin_unlock_irq(&conf->resync_lock);
999}
1000
1001static sector_t choose_data_offset(struct r10bio *r10_bio,
1002 struct md_rdev *rdev)
1003{
1004 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1005 test_bit(R10BIO_Previous, &r10_bio->state))
1006 return rdev->data_offset;
1007 else
1008 return rdev->new_data_offset;
1009}
1010
1011struct raid10_plug_cb {
1012 struct blk_plug_cb cb;
1013 struct bio_list pending;
1014 int pending_cnt;
1015};
1016
1017static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1018{
1019 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1020 cb);
1021 struct mddev *mddev = plug->cb.data;
1022 struct r10conf *conf = mddev->private;
1023 struct bio *bio;
1024
1025 if (from_schedule || current->bio_list) {
1026 spin_lock_irq(&conf->device_lock);
1027 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1028 conf->pending_count += plug->pending_cnt;
1029 spin_unlock_irq(&conf->device_lock);
1030 wake_up(&conf->wait_barrier);
1031 md_wakeup_thread(mddev->thread);
1032 kfree(plug);
1033 return;
1034 }
1035
1036 /* we aren't scheduling, so we can do the write-out directly. */
1037 bio = bio_list_get(&plug->pending);
1038 bitmap_unplug(mddev->bitmap);
1039 wake_up(&conf->wait_barrier);
1040
1041 while (bio) { /* submit pending writes */
1042 struct bio *next = bio->bi_next;
1043 bio->bi_next = NULL;
1044 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1045 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1046 /* Just ignore it */
1047 bio_endio(bio);
1048 else
1049 generic_make_request(bio);
1050 bio = next;
1051 }
1052 kfree(plug);
1053}
1054
1055static void __make_request(struct mddev *mddev, struct bio *bio)
1056{
1057 struct r10conf *conf = mddev->private;
1058 struct r10bio *r10_bio;
1059 struct bio *read_bio;
1060 int i;
1061 const int rw = bio_data_dir(bio);
1062 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1063 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1064 const unsigned long do_discard = (bio->bi_rw
1065 & (REQ_DISCARD | REQ_SECURE));
1066 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1067 unsigned long flags;
1068 struct md_rdev *blocked_rdev;
1069 struct blk_plug_cb *cb;
1070 struct raid10_plug_cb *plug = NULL;
1071 int sectors_handled;
1072 int max_sectors;
1073 int sectors;
1074
1075 /*
1076 * Register the new request and wait if the reconstruction
1077 * thread has put up a bar for new requests.
1078 * Continue immediately if no resync is active currently.
1079 */
1080 wait_barrier(conf);
1081
1082 sectors = bio_sectors(bio);
1083 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1084 bio->bi_iter.bi_sector < conf->reshape_progress &&
1085 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1086 /* IO spans the reshape position. Need to wait for
1087 * reshape to pass
1088 */
1089 allow_barrier(conf);
1090 wait_event(conf->wait_barrier,
1091 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1092 conf->reshape_progress >= bio->bi_iter.bi_sector +
1093 sectors);
1094 wait_barrier(conf);
1095 }
1096 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1097 bio_data_dir(bio) == WRITE &&
1098 (mddev->reshape_backwards
1099 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1100 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1101 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1102 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1103 /* Need to update reshape_position in metadata */
1104 mddev->reshape_position = conf->reshape_progress;
1105 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1106 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1107 md_wakeup_thread(mddev->thread);
1108 wait_event(mddev->sb_wait,
1109 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1110
1111 conf->reshape_safe = mddev->reshape_position;
1112 }
1113
1114 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1115
1116 r10_bio->master_bio = bio;
1117 r10_bio->sectors = sectors;
1118
1119 r10_bio->mddev = mddev;
1120 r10_bio->sector = bio->bi_iter.bi_sector;
1121 r10_bio->state = 0;
1122
1123 /* We might need to issue multiple reads to different
1124 * devices if there are bad blocks around, so we keep
1125 * track of the number of reads in bio->bi_phys_segments.
1126 * If this is 0, there is only one r10_bio and no locking
1127 * will be needed when the request completes. If it is
1128 * non-zero, then it is the number of not-completed requests.
1129 */
1130 bio->bi_phys_segments = 0;
1131 bio_clear_flag(bio, BIO_SEG_VALID);
1132
1133 if (rw == READ) {
1134 /*
1135 * read balancing logic:
1136 */
1137 struct md_rdev *rdev;
1138 int slot;
1139
1140read_again:
1141 rdev = read_balance(conf, r10_bio, &max_sectors);
1142 if (!rdev) {
1143 raid_end_bio_io(r10_bio);
1144 return;
1145 }
1146 slot = r10_bio->read_slot;
1147
1148 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1149 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1150 max_sectors);
1151
1152 r10_bio->devs[slot].bio = read_bio;
1153 r10_bio->devs[slot].rdev = rdev;
1154
1155 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1156 choose_data_offset(r10_bio, rdev);
1157 read_bio->bi_bdev = rdev->bdev;
1158 read_bio->bi_end_io = raid10_end_read_request;
1159 read_bio->bi_rw = READ | do_sync;
1160 read_bio->bi_private = r10_bio;
1161
1162 if (max_sectors < r10_bio->sectors) {
1163 /* Could not read all from this device, so we will
1164 * need another r10_bio.
1165 */
1166 sectors_handled = (r10_bio->sector + max_sectors
1167 - bio->bi_iter.bi_sector);
1168 r10_bio->sectors = max_sectors;
1169 spin_lock_irq(&conf->device_lock);
1170 if (bio->bi_phys_segments == 0)
1171 bio->bi_phys_segments = 2;
1172 else
1173 bio->bi_phys_segments++;
1174 spin_unlock_irq(&conf->device_lock);
1175 /* Cannot call generic_make_request directly
1176 * as that will be queued in __generic_make_request
1177 * and subsequent mempool_alloc might block
1178 * waiting for it. so hand bio over to raid10d.
1179 */
1180 reschedule_retry(r10_bio);
1181
1182 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1183
1184 r10_bio->master_bio = bio;
1185 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1186 r10_bio->state = 0;
1187 r10_bio->mddev = mddev;
1188 r10_bio->sector = bio->bi_iter.bi_sector +
1189 sectors_handled;
1190 goto read_again;
1191 } else
1192 generic_make_request(read_bio);
1193 return;
1194 }
1195
1196 /*
1197 * WRITE:
1198 */
1199 if (conf->pending_count >= max_queued_requests) {
1200 md_wakeup_thread(mddev->thread);
1201 wait_event(conf->wait_barrier,
1202 conf->pending_count < max_queued_requests);
1203 }
1204 /* first select target devices under rcu_lock and
1205 * inc refcount on their rdev. Record them by setting
1206 * bios[x] to bio
1207 * If there are known/acknowledged bad blocks on any device
1208 * on which we have seen a write error, we want to avoid
1209 * writing to those blocks. This potentially requires several
1210 * writes to write around the bad blocks. Each set of writes
1211 * gets its own r10_bio with a set of bios attached. The number
1212 * of r10_bios is recored in bio->bi_phys_segments just as with
1213 * the read case.
1214 */
1215
1216 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1217 raid10_find_phys(conf, r10_bio);
1218retry_write:
1219 blocked_rdev = NULL;
1220 rcu_read_lock();
1221 max_sectors = r10_bio->sectors;
1222
1223 for (i = 0; i < conf->copies; i++) {
1224 int d = r10_bio->devs[i].devnum;
1225 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1226 struct md_rdev *rrdev = rcu_dereference(
1227 conf->mirrors[d].replacement);
1228 if (rdev == rrdev)
1229 rrdev = NULL;
1230 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1231 atomic_inc(&rdev->nr_pending);
1232 blocked_rdev = rdev;
1233 break;
1234 }
1235 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1236 atomic_inc(&rrdev->nr_pending);
1237 blocked_rdev = rrdev;
1238 break;
1239 }
1240 if (rdev && (test_bit(Faulty, &rdev->flags)))
1241 rdev = NULL;
1242 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1243 rrdev = NULL;
1244
1245 r10_bio->devs[i].bio = NULL;
1246 r10_bio->devs[i].repl_bio = NULL;
1247
1248 if (!rdev && !rrdev) {
1249 set_bit(R10BIO_Degraded, &r10_bio->state);
1250 continue;
1251 }
1252 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1253 sector_t first_bad;
1254 sector_t dev_sector = r10_bio->devs[i].addr;
1255 int bad_sectors;
1256 int is_bad;
1257
1258 is_bad = is_badblock(rdev, dev_sector,
1259 max_sectors,
1260 &first_bad, &bad_sectors);
1261 if (is_bad < 0) {
1262 /* Mustn't write here until the bad block
1263 * is acknowledged
1264 */
1265 atomic_inc(&rdev->nr_pending);
1266 set_bit(BlockedBadBlocks, &rdev->flags);
1267 blocked_rdev = rdev;
1268 break;
1269 }
1270 if (is_bad && first_bad <= dev_sector) {
1271 /* Cannot write here at all */
1272 bad_sectors -= (dev_sector - first_bad);
1273 if (bad_sectors < max_sectors)
1274 /* Mustn't write more than bad_sectors
1275 * to other devices yet
1276 */
1277 max_sectors = bad_sectors;
1278 /* We don't set R10BIO_Degraded as that
1279 * only applies if the disk is missing,
1280 * so it might be re-added, and we want to
1281 * know to recover this chunk.
1282 * In this case the device is here, and the
1283 * fact that this chunk is not in-sync is
1284 * recorded in the bad block log.
1285 */
1286 continue;
1287 }
1288 if (is_bad) {
1289 int good_sectors = first_bad - dev_sector;
1290 if (good_sectors < max_sectors)
1291 max_sectors = good_sectors;
1292 }
1293 }
1294 if (rdev) {
1295 r10_bio->devs[i].bio = bio;
1296 atomic_inc(&rdev->nr_pending);
1297 }
1298 if (rrdev) {
1299 r10_bio->devs[i].repl_bio = bio;
1300 atomic_inc(&rrdev->nr_pending);
1301 }
1302 }
1303 rcu_read_unlock();
1304
1305 if (unlikely(blocked_rdev)) {
1306 /* Have to wait for this device to get unblocked, then retry */
1307 int j;
1308 int d;
1309
1310 for (j = 0; j < i; j++) {
1311 if (r10_bio->devs[j].bio) {
1312 d = r10_bio->devs[j].devnum;
1313 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1314 }
1315 if (r10_bio->devs[j].repl_bio) {
1316 struct md_rdev *rdev;
1317 d = r10_bio->devs[j].devnum;
1318 rdev = conf->mirrors[d].replacement;
1319 if (!rdev) {
1320 /* Race with remove_disk */
1321 smp_mb();
1322 rdev = conf->mirrors[d].rdev;
1323 }
1324 rdev_dec_pending(rdev, mddev);
1325 }
1326 }
1327 allow_barrier(conf);
1328 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1329 wait_barrier(conf);
1330 goto retry_write;
1331 }
1332
1333 if (max_sectors < r10_bio->sectors) {
1334 /* We are splitting this into multiple parts, so
1335 * we need to prepare for allocating another r10_bio.
1336 */
1337 r10_bio->sectors = max_sectors;
1338 spin_lock_irq(&conf->device_lock);
1339 if (bio->bi_phys_segments == 0)
1340 bio->bi_phys_segments = 2;
1341 else
1342 bio->bi_phys_segments++;
1343 spin_unlock_irq(&conf->device_lock);
1344 }
1345 sectors_handled = r10_bio->sector + max_sectors -
1346 bio->bi_iter.bi_sector;
1347
1348 atomic_set(&r10_bio->remaining, 1);
1349 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1350
1351 for (i = 0; i < conf->copies; i++) {
1352 struct bio *mbio;
1353 int d = r10_bio->devs[i].devnum;
1354 if (r10_bio->devs[i].bio) {
1355 struct md_rdev *rdev = conf->mirrors[d].rdev;
1356 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1357 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1358 max_sectors);
1359 r10_bio->devs[i].bio = mbio;
1360
1361 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1362 choose_data_offset(r10_bio,
1363 rdev));
1364 mbio->bi_bdev = rdev->bdev;
1365 mbio->bi_end_io = raid10_end_write_request;
1366 mbio->bi_rw =
1367 WRITE | do_sync | do_fua | do_discard | do_same;
1368 mbio->bi_private = r10_bio;
1369
1370 atomic_inc(&r10_bio->remaining);
1371
1372 cb = blk_check_plugged(raid10_unplug, mddev,
1373 sizeof(*plug));
1374 if (cb)
1375 plug = container_of(cb, struct raid10_plug_cb,
1376 cb);
1377 else
1378 plug = NULL;
1379 spin_lock_irqsave(&conf->device_lock, flags);
1380 if (plug) {
1381 bio_list_add(&plug->pending, mbio);
1382 plug->pending_cnt++;
1383 } else {
1384 bio_list_add(&conf->pending_bio_list, mbio);
1385 conf->pending_count++;
1386 }
1387 spin_unlock_irqrestore(&conf->device_lock, flags);
1388 if (!plug)
1389 md_wakeup_thread(mddev->thread);
1390 }
1391
1392 if (r10_bio->devs[i].repl_bio) {
1393 struct md_rdev *rdev = conf->mirrors[d].replacement;
1394 if (rdev == NULL) {
1395 /* Replacement just got moved to main 'rdev' */
1396 smp_mb();
1397 rdev = conf->mirrors[d].rdev;
1398 }
1399 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1400 bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1401 max_sectors);
1402 r10_bio->devs[i].repl_bio = mbio;
1403
1404 mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1405 choose_data_offset(
1406 r10_bio, rdev));
1407 mbio->bi_bdev = rdev->bdev;
1408 mbio->bi_end_io = raid10_end_write_request;
1409 mbio->bi_rw =
1410 WRITE | do_sync | do_fua | do_discard | do_same;
1411 mbio->bi_private = r10_bio;
1412
1413 atomic_inc(&r10_bio->remaining);
1414 spin_lock_irqsave(&conf->device_lock, flags);
1415 bio_list_add(&conf->pending_bio_list, mbio);
1416 conf->pending_count++;
1417 spin_unlock_irqrestore(&conf->device_lock, flags);
1418 if (!mddev_check_plugged(mddev))
1419 md_wakeup_thread(mddev->thread);
1420 }
1421 }
1422
1423 /* Don't remove the bias on 'remaining' (one_write_done) until
1424 * after checking if we need to go around again.
1425 */
1426
1427 if (sectors_handled < bio_sectors(bio)) {
1428 one_write_done(r10_bio);
1429 /* We need another r10_bio. It has already been counted
1430 * in bio->bi_phys_segments.
1431 */
1432 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1433
1434 r10_bio->master_bio = bio;
1435 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1436
1437 r10_bio->mddev = mddev;
1438 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1439 r10_bio->state = 0;
1440 goto retry_write;
1441 }
1442 one_write_done(r10_bio);
1443}
1444
1445static void raid10_make_request(struct mddev *mddev, struct bio *bio)
1446{
1447 struct r10conf *conf = mddev->private;
1448 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1449 int chunk_sects = chunk_mask + 1;
1450
1451 struct bio *split;
1452
1453 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1454 md_flush_request(mddev, bio);
1455 return;
1456 }
1457
1458 md_write_start(mddev, bio);
1459
1460 do {
1461
1462 /*
1463 * If this request crosses a chunk boundary, we need to split
1464 * it.
1465 */
1466 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1467 bio_sectors(bio) > chunk_sects
1468 && (conf->geo.near_copies < conf->geo.raid_disks
1469 || conf->prev.near_copies <
1470 conf->prev.raid_disks))) {
1471 split = bio_split(bio, chunk_sects -
1472 (bio->bi_iter.bi_sector &
1473 (chunk_sects - 1)),
1474 GFP_NOIO, fs_bio_set);
1475 bio_chain(split, bio);
1476 } else {
1477 split = bio;
1478 }
1479
1480 __make_request(mddev, split);
1481 } while (split != bio);
1482
1483 /* In case raid10d snuck in to freeze_array */
1484 wake_up(&conf->wait_barrier);
1485}
1486
1487static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1488{
1489 struct r10conf *conf = mddev->private;
1490 int i;
1491
1492 if (conf->geo.near_copies < conf->geo.raid_disks)
1493 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1494 if (conf->geo.near_copies > 1)
1495 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1496 if (conf->geo.far_copies > 1) {
1497 if (conf->geo.far_offset)
1498 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1499 else
1500 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1501 if (conf->geo.far_set_size != conf->geo.raid_disks)
1502 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1503 }
1504 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1505 conf->geo.raid_disks - mddev->degraded);
1506 for (i = 0; i < conf->geo.raid_disks; i++)
1507 seq_printf(seq, "%s",
1508 conf->mirrors[i].rdev &&
1509 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1510 seq_printf(seq, "]");
1511}
1512
1513/* check if there are enough drives for
1514 * every block to appear on atleast one.
1515 * Don't consider the device numbered 'ignore'
1516 * as we might be about to remove it.
1517 */
1518static int _enough(struct r10conf *conf, int previous, int ignore)
1519{
1520 int first = 0;
1521 int has_enough = 0;
1522 int disks, ncopies;
1523 if (previous) {
1524 disks = conf->prev.raid_disks;
1525 ncopies = conf->prev.near_copies;
1526 } else {
1527 disks = conf->geo.raid_disks;
1528 ncopies = conf->geo.near_copies;
1529 }
1530
1531 rcu_read_lock();
1532 do {
1533 int n = conf->copies;
1534 int cnt = 0;
1535 int this = first;
1536 while (n--) {
1537 struct md_rdev *rdev;
1538 if (this != ignore &&
1539 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1540 test_bit(In_sync, &rdev->flags))
1541 cnt++;
1542 this = (this+1) % disks;
1543 }
1544 if (cnt == 0)
1545 goto out;
1546 first = (first + ncopies) % disks;
1547 } while (first != 0);
1548 has_enough = 1;
1549out:
1550 rcu_read_unlock();
1551 return has_enough;
1552}
1553
1554static int enough(struct r10conf *conf, int ignore)
1555{
1556 /* when calling 'enough', both 'prev' and 'geo' must
1557 * be stable.
1558 * This is ensured if ->reconfig_mutex or ->device_lock
1559 * is held.
1560 */
1561 return _enough(conf, 0, ignore) &&
1562 _enough(conf, 1, ignore);
1563}
1564
1565static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1566{
1567 char b[BDEVNAME_SIZE];
1568 struct r10conf *conf = mddev->private;
1569 unsigned long flags;
1570
1571 /*
1572 * If it is not operational, then we have already marked it as dead
1573 * else if it is the last working disks, ignore the error, let the
1574 * next level up know.
1575 * else mark the drive as failed
1576 */
1577 spin_lock_irqsave(&conf->device_lock, flags);
1578 if (test_bit(In_sync, &rdev->flags)
1579 && !enough(conf, rdev->raid_disk)) {
1580 /*
1581 * Don't fail the drive, just return an IO error.
1582 */
1583 spin_unlock_irqrestore(&conf->device_lock, flags);
1584 return;
1585 }
1586 if (test_and_clear_bit(In_sync, &rdev->flags))
1587 mddev->degraded++;
1588 /*
1589 * If recovery is running, make sure it aborts.
1590 */
1591 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1592 set_bit(Blocked, &rdev->flags);
1593 set_bit(Faulty, &rdev->flags);
1594 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1595 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1596 spin_unlock_irqrestore(&conf->device_lock, flags);
1597 printk(KERN_ALERT
1598 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1599 "md/raid10:%s: Operation continuing on %d devices.\n",
1600 mdname(mddev), bdevname(rdev->bdev, b),
1601 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1602}
1603
1604static void print_conf(struct r10conf *conf)
1605{
1606 int i;
1607 struct raid10_info *tmp;
1608
1609 printk(KERN_DEBUG "RAID10 conf printout:\n");
1610 if (!conf) {
1611 printk(KERN_DEBUG "(!conf)\n");
1612 return;
1613 }
1614 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1615 conf->geo.raid_disks);
1616
1617 for (i = 0; i < conf->geo.raid_disks; i++) {
1618 char b[BDEVNAME_SIZE];
1619 tmp = conf->mirrors + i;
1620 if (tmp->rdev)
1621 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1622 i, !test_bit(In_sync, &tmp->rdev->flags),
1623 !test_bit(Faulty, &tmp->rdev->flags),
1624 bdevname(tmp->rdev->bdev,b));
1625 }
1626}
1627
1628static void close_sync(struct r10conf *conf)
1629{
1630 wait_barrier(conf);
1631 allow_barrier(conf);
1632
1633 mempool_destroy(conf->r10buf_pool);
1634 conf->r10buf_pool = NULL;
1635}
1636
1637static int raid10_spare_active(struct mddev *mddev)
1638{
1639 int i;
1640 struct r10conf *conf = mddev->private;
1641 struct raid10_info *tmp;
1642 int count = 0;
1643 unsigned long flags;
1644
1645 /*
1646 * Find all non-in_sync disks within the RAID10 configuration
1647 * and mark them in_sync
1648 */
1649 for (i = 0; i < conf->geo.raid_disks; i++) {
1650 tmp = conf->mirrors + i;
1651 if (tmp->replacement
1652 && tmp->replacement->recovery_offset == MaxSector
1653 && !test_bit(Faulty, &tmp->replacement->flags)
1654 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1655 /* Replacement has just become active */
1656 if (!tmp->rdev
1657 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1658 count++;
1659 if (tmp->rdev) {
1660 /* Replaced device not technically faulty,
1661 * but we need to be sure it gets removed
1662 * and never re-added.
1663 */
1664 set_bit(Faulty, &tmp->rdev->flags);
1665 sysfs_notify_dirent_safe(
1666 tmp->rdev->sysfs_state);
1667 }
1668 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1669 } else if (tmp->rdev
1670 && tmp->rdev->recovery_offset == MaxSector
1671 && !test_bit(Faulty, &tmp->rdev->flags)
1672 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1673 count++;
1674 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1675 }
1676 }
1677 spin_lock_irqsave(&conf->device_lock, flags);
1678 mddev->degraded -= count;
1679 spin_unlock_irqrestore(&conf->device_lock, flags);
1680
1681 print_conf(conf);
1682 return count;
1683}
1684
1685static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1686{
1687 struct r10conf *conf = mddev->private;
1688 int err = -EEXIST;
1689 int mirror;
1690 int first = 0;
1691 int last = conf->geo.raid_disks - 1;
1692
1693 if (mddev->recovery_cp < MaxSector)
1694 /* only hot-add to in-sync arrays, as recovery is
1695 * very different from resync
1696 */
1697 return -EBUSY;
1698 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1699 return -EINVAL;
1700
1701 if (md_integrity_add_rdev(rdev, mddev))
1702 return -ENXIO;
1703
1704 if (rdev->raid_disk >= 0)
1705 first = last = rdev->raid_disk;
1706
1707 if (rdev->saved_raid_disk >= first &&
1708 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1709 mirror = rdev->saved_raid_disk;
1710 else
1711 mirror = first;
1712 for ( ; mirror <= last ; mirror++) {
1713 struct raid10_info *p = &conf->mirrors[mirror];
1714 if (p->recovery_disabled == mddev->recovery_disabled)
1715 continue;
1716 if (p->rdev) {
1717 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1718 p->replacement != NULL)
1719 continue;
1720 clear_bit(In_sync, &rdev->flags);
1721 set_bit(Replacement, &rdev->flags);
1722 rdev->raid_disk = mirror;
1723 err = 0;
1724 if (mddev->gendisk)
1725 disk_stack_limits(mddev->gendisk, rdev->bdev,
1726 rdev->data_offset << 9);
1727 conf->fullsync = 1;
1728 rcu_assign_pointer(p->replacement, rdev);
1729 break;
1730 }
1731
1732 if (mddev->gendisk)
1733 disk_stack_limits(mddev->gendisk, rdev->bdev,
1734 rdev->data_offset << 9);
1735
1736 p->head_position = 0;
1737 p->recovery_disabled = mddev->recovery_disabled - 1;
1738 rdev->raid_disk = mirror;
1739 err = 0;
1740 if (rdev->saved_raid_disk != mirror)
1741 conf->fullsync = 1;
1742 rcu_assign_pointer(p->rdev, rdev);
1743 break;
1744 }
1745 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1746 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1747
1748 print_conf(conf);
1749 return err;
1750}
1751
1752static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1753{
1754 struct r10conf *conf = mddev->private;
1755 int err = 0;
1756 int number = rdev->raid_disk;
1757 struct md_rdev **rdevp;
1758 struct raid10_info *p = conf->mirrors + number;
1759
1760 print_conf(conf);
1761 if (rdev == p->rdev)
1762 rdevp = &p->rdev;
1763 else if (rdev == p->replacement)
1764 rdevp = &p->replacement;
1765 else
1766 return 0;
1767
1768 if (test_bit(In_sync, &rdev->flags) ||
1769 atomic_read(&rdev->nr_pending)) {
1770 err = -EBUSY;
1771 goto abort;
1772 }
1773 /* Only remove faulty devices if recovery
1774 * is not possible.
1775 */
1776 if (!test_bit(Faulty, &rdev->flags) &&
1777 mddev->recovery_disabled != p->recovery_disabled &&
1778 (!p->replacement || p->replacement == rdev) &&
1779 number < conf->geo.raid_disks &&
1780 enough(conf, -1)) {
1781 err = -EBUSY;
1782 goto abort;
1783 }
1784 *rdevp = NULL;
1785 synchronize_rcu();
1786 if (atomic_read(&rdev->nr_pending)) {
1787 /* lost the race, try later */
1788 err = -EBUSY;
1789 *rdevp = rdev;
1790 goto abort;
1791 } else if (p->replacement) {
1792 /* We must have just cleared 'rdev' */
1793 p->rdev = p->replacement;
1794 clear_bit(Replacement, &p->replacement->flags);
1795 smp_mb(); /* Make sure other CPUs may see both as identical
1796 * but will never see neither -- if they are careful.
1797 */
1798 p->replacement = NULL;
1799 clear_bit(WantReplacement, &rdev->flags);
1800 } else
1801 /* We might have just remove the Replacement as faulty
1802 * Clear the flag just in case
1803 */
1804 clear_bit(WantReplacement, &rdev->flags);
1805
1806 err = md_integrity_register(mddev);
1807
1808abort:
1809
1810 print_conf(conf);
1811 return err;
1812}
1813
1814static void end_sync_read(struct bio *bio)
1815{
1816 struct r10bio *r10_bio = bio->bi_private;
1817 struct r10conf *conf = r10_bio->mddev->private;
1818 int d;
1819
1820 if (bio == r10_bio->master_bio) {
1821 /* this is a reshape read */
1822 d = r10_bio->read_slot; /* really the read dev */
1823 } else
1824 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1825
1826 if (!bio->bi_error)
1827 set_bit(R10BIO_Uptodate, &r10_bio->state);
1828 else
1829 /* The write handler will notice the lack of
1830 * R10BIO_Uptodate and record any errors etc
1831 */
1832 atomic_add(r10_bio->sectors,
1833 &conf->mirrors[d].rdev->corrected_errors);
1834
1835 /* for reconstruct, we always reschedule after a read.
1836 * for resync, only after all reads
1837 */
1838 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1839 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1840 atomic_dec_and_test(&r10_bio->remaining)) {
1841 /* we have read all the blocks,
1842 * do the comparison in process context in raid10d
1843 */
1844 reschedule_retry(r10_bio);
1845 }
1846}
1847
1848static void end_sync_request(struct r10bio *r10_bio)
1849{
1850 struct mddev *mddev = r10_bio->mddev;
1851
1852 while (atomic_dec_and_test(&r10_bio->remaining)) {
1853 if (r10_bio->master_bio == NULL) {
1854 /* the primary of several recovery bios */
1855 sector_t s = r10_bio->sectors;
1856 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1857 test_bit(R10BIO_WriteError, &r10_bio->state))
1858 reschedule_retry(r10_bio);
1859 else
1860 put_buf(r10_bio);
1861 md_done_sync(mddev, s, 1);
1862 break;
1863 } else {
1864 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1865 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1866 test_bit(R10BIO_WriteError, &r10_bio->state))
1867 reschedule_retry(r10_bio);
1868 else
1869 put_buf(r10_bio);
1870 r10_bio = r10_bio2;
1871 }
1872 }
1873}
1874
1875static void end_sync_write(struct bio *bio)
1876{
1877 struct r10bio *r10_bio = bio->bi_private;
1878 struct mddev *mddev = r10_bio->mddev;
1879 struct r10conf *conf = mddev->private;
1880 int d;
1881 sector_t first_bad;
1882 int bad_sectors;
1883 int slot;
1884 int repl;
1885 struct md_rdev *rdev = NULL;
1886
1887 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1888 if (repl)
1889 rdev = conf->mirrors[d].replacement;
1890 else
1891 rdev = conf->mirrors[d].rdev;
1892
1893 if (bio->bi_error) {
1894 if (repl)
1895 md_error(mddev, rdev);
1896 else {
1897 set_bit(WriteErrorSeen, &rdev->flags);
1898 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1899 set_bit(MD_RECOVERY_NEEDED,
1900 &rdev->mddev->recovery);
1901 set_bit(R10BIO_WriteError, &r10_bio->state);
1902 }
1903 } else if (is_badblock(rdev,
1904 r10_bio->devs[slot].addr,
1905 r10_bio->sectors,
1906 &first_bad, &bad_sectors))
1907 set_bit(R10BIO_MadeGood, &r10_bio->state);
1908
1909 rdev_dec_pending(rdev, mddev);
1910
1911 end_sync_request(r10_bio);
1912}
1913
1914/*
1915 * Note: sync and recover and handled very differently for raid10
1916 * This code is for resync.
1917 * For resync, we read through virtual addresses and read all blocks.
1918 * If there is any error, we schedule a write. The lowest numbered
1919 * drive is authoritative.
1920 * However requests come for physical address, so we need to map.
1921 * For every physical address there are raid_disks/copies virtual addresses,
1922 * which is always are least one, but is not necessarly an integer.
1923 * This means that a physical address can span multiple chunks, so we may
1924 * have to submit multiple io requests for a single sync request.
1925 */
1926/*
1927 * We check if all blocks are in-sync and only write to blocks that
1928 * aren't in sync
1929 */
1930static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1931{
1932 struct r10conf *conf = mddev->private;
1933 int i, first;
1934 struct bio *tbio, *fbio;
1935 int vcnt;
1936
1937 atomic_set(&r10_bio->remaining, 1);
1938
1939 /* find the first device with a block */
1940 for (i=0; i<conf->copies; i++)
1941 if (!r10_bio->devs[i].bio->bi_error)
1942 break;
1943
1944 if (i == conf->copies)
1945 goto done;
1946
1947 first = i;
1948 fbio = r10_bio->devs[i].bio;
1949 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
1950 fbio->bi_iter.bi_idx = 0;
1951
1952 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1953 /* now find blocks with errors */
1954 for (i=0 ; i < conf->copies ; i++) {
1955 int j, d;
1956
1957 tbio = r10_bio->devs[i].bio;
1958
1959 if (tbio->bi_end_io != end_sync_read)
1960 continue;
1961 if (i == first)
1962 continue;
1963 if (!r10_bio->devs[i].bio->bi_error) {
1964 /* We know that the bi_io_vec layout is the same for
1965 * both 'first' and 'i', so we just compare them.
1966 * All vec entries are PAGE_SIZE;
1967 */
1968 int sectors = r10_bio->sectors;
1969 for (j = 0; j < vcnt; j++) {
1970 int len = PAGE_SIZE;
1971 if (sectors < (len / 512))
1972 len = sectors * 512;
1973 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1974 page_address(tbio->bi_io_vec[j].bv_page),
1975 len))
1976 break;
1977 sectors -= len/512;
1978 }
1979 if (j == vcnt)
1980 continue;
1981 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
1982 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1983 /* Don't fix anything. */
1984 continue;
1985 }
1986 /* Ok, we need to write this bio, either to correct an
1987 * inconsistency or to correct an unreadable block.
1988 * First we need to fixup bv_offset, bv_len and
1989 * bi_vecs, as the read request might have corrupted these
1990 */
1991 bio_reset(tbio);
1992
1993 tbio->bi_vcnt = vcnt;
1994 tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
1995 tbio->bi_rw = WRITE;
1996 tbio->bi_private = r10_bio;
1997 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
1998 tbio->bi_end_io = end_sync_write;
1999
2000 bio_copy_data(tbio, fbio);
2001
2002 d = r10_bio->devs[i].devnum;
2003 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2004 atomic_inc(&r10_bio->remaining);
2005 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2006
2007 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2008 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2009 generic_make_request(tbio);
2010 }
2011
2012 /* Now write out to any replacement devices
2013 * that are active
2014 */
2015 for (i = 0; i < conf->copies; i++) {
2016 int d;
2017
2018 tbio = r10_bio->devs[i].repl_bio;
2019 if (!tbio || !tbio->bi_end_io)
2020 continue;
2021 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2022 && r10_bio->devs[i].bio != fbio)
2023 bio_copy_data(tbio, fbio);
2024 d = r10_bio->devs[i].devnum;
2025 atomic_inc(&r10_bio->remaining);
2026 md_sync_acct(conf->mirrors[d].replacement->bdev,
2027 bio_sectors(tbio));
2028 generic_make_request(tbio);
2029 }
2030
2031done:
2032 if (atomic_dec_and_test(&r10_bio->remaining)) {
2033 md_done_sync(mddev, r10_bio->sectors, 1);
2034 put_buf(r10_bio);
2035 }
2036}
2037
2038/*
2039 * Now for the recovery code.
2040 * Recovery happens across physical sectors.
2041 * We recover all non-is_sync drives by finding the virtual address of
2042 * each, and then choose a working drive that also has that virt address.
2043 * There is a separate r10_bio for each non-in_sync drive.
2044 * Only the first two slots are in use. The first for reading,
2045 * The second for writing.
2046 *
2047 */
2048static void fix_recovery_read_error(struct r10bio *r10_bio)
2049{
2050 /* We got a read error during recovery.
2051 * We repeat the read in smaller page-sized sections.
2052 * If a read succeeds, write it to the new device or record
2053 * a bad block if we cannot.
2054 * If a read fails, record a bad block on both old and
2055 * new devices.
2056 */
2057 struct mddev *mddev = r10_bio->mddev;
2058 struct r10conf *conf = mddev->private;
2059 struct bio *bio = r10_bio->devs[0].bio;
2060 sector_t sect = 0;
2061 int sectors = r10_bio->sectors;
2062 int idx = 0;
2063 int dr = r10_bio->devs[0].devnum;
2064 int dw = r10_bio->devs[1].devnum;
2065
2066 while (sectors) {
2067 int s = sectors;
2068 struct md_rdev *rdev;
2069 sector_t addr;
2070 int ok;
2071
2072 if (s > (PAGE_SIZE>>9))
2073 s = PAGE_SIZE >> 9;
2074
2075 rdev = conf->mirrors[dr].rdev;
2076 addr = r10_bio->devs[0].addr + sect,
2077 ok = sync_page_io(rdev,
2078 addr,
2079 s << 9,
2080 bio->bi_io_vec[idx].bv_page,
2081 READ, false);
2082 if (ok) {
2083 rdev = conf->mirrors[dw].rdev;
2084 addr = r10_bio->devs[1].addr + sect;
2085 ok = sync_page_io(rdev,
2086 addr,
2087 s << 9,
2088 bio->bi_io_vec[idx].bv_page,
2089 WRITE, false);
2090 if (!ok) {
2091 set_bit(WriteErrorSeen, &rdev->flags);
2092 if (!test_and_set_bit(WantReplacement,
2093 &rdev->flags))
2094 set_bit(MD_RECOVERY_NEEDED,
2095 &rdev->mddev->recovery);
2096 }
2097 }
2098 if (!ok) {
2099 /* We don't worry if we cannot set a bad block -
2100 * it really is bad so there is no loss in not
2101 * recording it yet
2102 */
2103 rdev_set_badblocks(rdev, addr, s, 0);
2104
2105 if (rdev != conf->mirrors[dw].rdev) {
2106 /* need bad block on destination too */
2107 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2108 addr = r10_bio->devs[1].addr + sect;
2109 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2110 if (!ok) {
2111 /* just abort the recovery */
2112 printk(KERN_NOTICE
2113 "md/raid10:%s: recovery aborted"
2114 " due to read error\n",
2115 mdname(mddev));
2116
2117 conf->mirrors[dw].recovery_disabled
2118 = mddev->recovery_disabled;
2119 set_bit(MD_RECOVERY_INTR,
2120 &mddev->recovery);
2121 break;
2122 }
2123 }
2124 }
2125
2126 sectors -= s;
2127 sect += s;
2128 idx++;
2129 }
2130}
2131
2132static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2133{
2134 struct r10conf *conf = mddev->private;
2135 int d;
2136 struct bio *wbio, *wbio2;
2137
2138 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2139 fix_recovery_read_error(r10_bio);
2140 end_sync_request(r10_bio);
2141 return;
2142 }
2143
2144 /*
2145 * share the pages with the first bio
2146 * and submit the write request
2147 */
2148 d = r10_bio->devs[1].devnum;
2149 wbio = r10_bio->devs[1].bio;
2150 wbio2 = r10_bio->devs[1].repl_bio;
2151 /* Need to test wbio2->bi_end_io before we call
2152 * generic_make_request as if the former is NULL,
2153 * the latter is free to free wbio2.
2154 */
2155 if (wbio2 && !wbio2->bi_end_io)
2156 wbio2 = NULL;
2157 if (wbio->bi_end_io) {
2158 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2159 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2160 generic_make_request(wbio);
2161 }
2162 if (wbio2) {
2163 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2164 md_sync_acct(conf->mirrors[d].replacement->bdev,
2165 bio_sectors(wbio2));
2166 generic_make_request(wbio2);
2167 }
2168}
2169
2170/*
2171 * Used by fix_read_error() to decay the per rdev read_errors.
2172 * We halve the read error count for every hour that has elapsed
2173 * since the last recorded read error.
2174 *
2175 */
2176static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2177{
2178 struct timespec cur_time_mon;
2179 unsigned long hours_since_last;
2180 unsigned int read_errors = atomic_read(&rdev->read_errors);
2181
2182 ktime_get_ts(&cur_time_mon);
2183
2184 if (rdev->last_read_error.tv_sec == 0 &&
2185 rdev->last_read_error.tv_nsec == 0) {
2186 /* first time we've seen a read error */
2187 rdev->last_read_error = cur_time_mon;
2188 return;
2189 }
2190
2191 hours_since_last = (cur_time_mon.tv_sec -
2192 rdev->last_read_error.tv_sec) / 3600;
2193
2194 rdev->last_read_error = cur_time_mon;
2195
2196 /*
2197 * if hours_since_last is > the number of bits in read_errors
2198 * just set read errors to 0. We do this to avoid
2199 * overflowing the shift of read_errors by hours_since_last.
2200 */
2201 if (hours_since_last >= 8 * sizeof(read_errors))
2202 atomic_set(&rdev->read_errors, 0);
2203 else
2204 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2205}
2206
2207static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2208 int sectors, struct page *page, int rw)
2209{
2210 sector_t first_bad;
2211 int bad_sectors;
2212
2213 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2214 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2215 return -1;
2216 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2217 /* success */
2218 return 1;
2219 if (rw == WRITE) {
2220 set_bit(WriteErrorSeen, &rdev->flags);
2221 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2222 set_bit(MD_RECOVERY_NEEDED,
2223 &rdev->mddev->recovery);
2224 }
2225 /* need to record an error - either for the block or the device */
2226 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2227 md_error(rdev->mddev, rdev);
2228 return 0;
2229}
2230
2231/*
2232 * This is a kernel thread which:
2233 *
2234 * 1. Retries failed read operations on working mirrors.
2235 * 2. Updates the raid superblock when problems encounter.
2236 * 3. Performs writes following reads for array synchronising.
2237 */
2238
2239static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2240{
2241 int sect = 0; /* Offset from r10_bio->sector */
2242 int sectors = r10_bio->sectors;
2243 struct md_rdev*rdev;
2244 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2245 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2246
2247 /* still own a reference to this rdev, so it cannot
2248 * have been cleared recently.
2249 */
2250 rdev = conf->mirrors[d].rdev;
2251
2252 if (test_bit(Faulty, &rdev->flags))
2253 /* drive has already been failed, just ignore any
2254 more fix_read_error() attempts */
2255 return;
2256
2257 check_decay_read_errors(mddev, rdev);
2258 atomic_inc(&rdev->read_errors);
2259 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2260 char b[BDEVNAME_SIZE];
2261 bdevname(rdev->bdev, b);
2262
2263 printk(KERN_NOTICE
2264 "md/raid10:%s: %s: Raid device exceeded "
2265 "read_error threshold [cur %d:max %d]\n",
2266 mdname(mddev), b,
2267 atomic_read(&rdev->read_errors), max_read_errors);
2268 printk(KERN_NOTICE
2269 "md/raid10:%s: %s: Failing raid device\n",
2270 mdname(mddev), b);
2271 md_error(mddev, conf->mirrors[d].rdev);
2272 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2273 return;
2274 }
2275
2276 while(sectors) {
2277 int s = sectors;
2278 int sl = r10_bio->read_slot;
2279 int success = 0;
2280 int start;
2281
2282 if (s > (PAGE_SIZE>>9))
2283 s = PAGE_SIZE >> 9;
2284
2285 rcu_read_lock();
2286 do {
2287 sector_t first_bad;
2288 int bad_sectors;
2289
2290 d = r10_bio->devs[sl].devnum;
2291 rdev = rcu_dereference(conf->mirrors[d].rdev);
2292 if (rdev &&
2293 test_bit(In_sync, &rdev->flags) &&
2294 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2295 &first_bad, &bad_sectors) == 0) {
2296 atomic_inc(&rdev->nr_pending);
2297 rcu_read_unlock();
2298 success = sync_page_io(rdev,
2299 r10_bio->devs[sl].addr +
2300 sect,
2301 s<<9,
2302 conf->tmppage, READ, false);
2303 rdev_dec_pending(rdev, mddev);
2304 rcu_read_lock();
2305 if (success)
2306 break;
2307 }
2308 sl++;
2309 if (sl == conf->copies)
2310 sl = 0;
2311 } while (!success && sl != r10_bio->read_slot);
2312 rcu_read_unlock();
2313
2314 if (!success) {
2315 /* Cannot read from anywhere, just mark the block
2316 * as bad on the first device to discourage future
2317 * reads.
2318 */
2319 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2320 rdev = conf->mirrors[dn].rdev;
2321
2322 if (!rdev_set_badblocks(
2323 rdev,
2324 r10_bio->devs[r10_bio->read_slot].addr
2325 + sect,
2326 s, 0)) {
2327 md_error(mddev, rdev);
2328 r10_bio->devs[r10_bio->read_slot].bio
2329 = IO_BLOCKED;
2330 }
2331 break;
2332 }
2333
2334 start = sl;
2335 /* write it back and re-read */
2336 rcu_read_lock();
2337 while (sl != r10_bio->read_slot) {
2338 char b[BDEVNAME_SIZE];
2339
2340 if (sl==0)
2341 sl = conf->copies;
2342 sl--;
2343 d = r10_bio->devs[sl].devnum;
2344 rdev = rcu_dereference(conf->mirrors[d].rdev);
2345 if (!rdev ||
2346 !test_bit(In_sync, &rdev->flags))
2347 continue;
2348
2349 atomic_inc(&rdev->nr_pending);
2350 rcu_read_unlock();
2351 if (r10_sync_page_io(rdev,
2352 r10_bio->devs[sl].addr +
2353 sect,
2354 s, conf->tmppage, WRITE)
2355 == 0) {
2356 /* Well, this device is dead */
2357 printk(KERN_NOTICE
2358 "md/raid10:%s: read correction "
2359 "write failed"
2360 " (%d sectors at %llu on %s)\n",
2361 mdname(mddev), s,
2362 (unsigned long long)(
2363 sect +
2364 choose_data_offset(r10_bio,
2365 rdev)),
2366 bdevname(rdev->bdev, b));
2367 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2368 "drive\n",
2369 mdname(mddev),
2370 bdevname(rdev->bdev, b));
2371 }
2372 rdev_dec_pending(rdev, mddev);
2373 rcu_read_lock();
2374 }
2375 sl = start;
2376 while (sl != r10_bio->read_slot) {
2377 char b[BDEVNAME_SIZE];
2378
2379 if (sl==0)
2380 sl = conf->copies;
2381 sl--;
2382 d = r10_bio->devs[sl].devnum;
2383 rdev = rcu_dereference(conf->mirrors[d].rdev);
2384 if (!rdev ||
2385 !test_bit(In_sync, &rdev->flags))
2386 continue;
2387
2388 atomic_inc(&rdev->nr_pending);
2389 rcu_read_unlock();
2390 switch (r10_sync_page_io(rdev,
2391 r10_bio->devs[sl].addr +
2392 sect,
2393 s, conf->tmppage,
2394 READ)) {
2395 case 0:
2396 /* Well, this device is dead */
2397 printk(KERN_NOTICE
2398 "md/raid10:%s: unable to read back "
2399 "corrected sectors"
2400 " (%d sectors at %llu on %s)\n",
2401 mdname(mddev), s,
2402 (unsigned long long)(
2403 sect +
2404 choose_data_offset(r10_bio, rdev)),
2405 bdevname(rdev->bdev, b));
2406 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2407 "drive\n",
2408 mdname(mddev),
2409 bdevname(rdev->bdev, b));
2410 break;
2411 case 1:
2412 printk(KERN_INFO
2413 "md/raid10:%s: read error corrected"
2414 " (%d sectors at %llu on %s)\n",
2415 mdname(mddev), s,
2416 (unsigned long long)(
2417 sect +
2418 choose_data_offset(r10_bio, rdev)),
2419 bdevname(rdev->bdev, b));
2420 atomic_add(s, &rdev->corrected_errors);
2421 }
2422
2423 rdev_dec_pending(rdev, mddev);
2424 rcu_read_lock();
2425 }
2426 rcu_read_unlock();
2427
2428 sectors -= s;
2429 sect += s;
2430 }
2431}
2432
2433static int narrow_write_error(struct r10bio *r10_bio, int i)
2434{
2435 struct bio *bio = r10_bio->master_bio;
2436 struct mddev *mddev = r10_bio->mddev;
2437 struct r10conf *conf = mddev->private;
2438 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2439 /* bio has the data to be written to slot 'i' where
2440 * we just recently had a write error.
2441 * We repeatedly clone the bio and trim down to one block,
2442 * then try the write. Where the write fails we record
2443 * a bad block.
2444 * It is conceivable that the bio doesn't exactly align with
2445 * blocks. We must handle this.
2446 *
2447 * We currently own a reference to the rdev.
2448 */
2449
2450 int block_sectors;
2451 sector_t sector;
2452 int sectors;
2453 int sect_to_write = r10_bio->sectors;
2454 int ok = 1;
2455
2456 if (rdev->badblocks.shift < 0)
2457 return 0;
2458
2459 block_sectors = roundup(1 << rdev->badblocks.shift,
2460 bdev_logical_block_size(rdev->bdev) >> 9);
2461 sector = r10_bio->sector;
2462 sectors = ((r10_bio->sector + block_sectors)
2463 & ~(sector_t)(block_sectors - 1))
2464 - sector;
2465
2466 while (sect_to_write) {
2467 struct bio *wbio;
2468 if (sectors > sect_to_write)
2469 sectors = sect_to_write;
2470 /* Write at 'sector' for 'sectors' */
2471 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2472 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2473 wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2474 choose_data_offset(r10_bio, rdev) +
2475 (sector - r10_bio->sector));
2476 wbio->bi_bdev = rdev->bdev;
2477 if (submit_bio_wait(WRITE, wbio) < 0)
2478 /* Failure! */
2479 ok = rdev_set_badblocks(rdev, sector,
2480 sectors, 0)
2481 && ok;
2482
2483 bio_put(wbio);
2484 sect_to_write -= sectors;
2485 sector += sectors;
2486 sectors = block_sectors;
2487 }
2488 return ok;
2489}
2490
2491static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2492{
2493 int slot = r10_bio->read_slot;
2494 struct bio *bio;
2495 struct r10conf *conf = mddev->private;
2496 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2497 char b[BDEVNAME_SIZE];
2498 unsigned long do_sync;
2499 int max_sectors;
2500
2501 /* we got a read error. Maybe the drive is bad. Maybe just
2502 * the block and we can fix it.
2503 * We freeze all other IO, and try reading the block from
2504 * other devices. When we find one, we re-write
2505 * and check it that fixes the read error.
2506 * This is all done synchronously while the array is
2507 * frozen.
2508 */
2509 bio = r10_bio->devs[slot].bio;
2510 bdevname(bio->bi_bdev, b);
2511 bio_put(bio);
2512 r10_bio->devs[slot].bio = NULL;
2513
2514 if (mddev->ro == 0) {
2515 freeze_array(conf, 1);
2516 fix_read_error(conf, mddev, r10_bio);
2517 unfreeze_array(conf);
2518 } else
2519 r10_bio->devs[slot].bio = IO_BLOCKED;
2520
2521 rdev_dec_pending(rdev, mddev);
2522
2523read_more:
2524 rdev = read_balance(conf, r10_bio, &max_sectors);
2525 if (rdev == NULL) {
2526 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2527 " read error for block %llu\n",
2528 mdname(mddev), b,
2529 (unsigned long long)r10_bio->sector);
2530 raid_end_bio_io(r10_bio);
2531 return;
2532 }
2533
2534 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2535 slot = r10_bio->read_slot;
2536 printk_ratelimited(
2537 KERN_ERR
2538 "md/raid10:%s: %s: redirecting "
2539 "sector %llu to another mirror\n",
2540 mdname(mddev),
2541 bdevname(rdev->bdev, b),
2542 (unsigned long long)r10_bio->sector);
2543 bio = bio_clone_mddev(r10_bio->master_bio,
2544 GFP_NOIO, mddev);
2545 bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2546 r10_bio->devs[slot].bio = bio;
2547 r10_bio->devs[slot].rdev = rdev;
2548 bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2549 + choose_data_offset(r10_bio, rdev);
2550 bio->bi_bdev = rdev->bdev;
2551 bio->bi_rw = READ | do_sync;
2552 bio->bi_private = r10_bio;
2553 bio->bi_end_io = raid10_end_read_request;
2554 if (max_sectors < r10_bio->sectors) {
2555 /* Drat - have to split this up more */
2556 struct bio *mbio = r10_bio->master_bio;
2557 int sectors_handled =
2558 r10_bio->sector + max_sectors
2559 - mbio->bi_iter.bi_sector;
2560 r10_bio->sectors = max_sectors;
2561 spin_lock_irq(&conf->device_lock);
2562 if (mbio->bi_phys_segments == 0)
2563 mbio->bi_phys_segments = 2;
2564 else
2565 mbio->bi_phys_segments++;
2566 spin_unlock_irq(&conf->device_lock);
2567 generic_make_request(bio);
2568
2569 r10_bio = mempool_alloc(conf->r10bio_pool,
2570 GFP_NOIO);
2571 r10_bio->master_bio = mbio;
2572 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2573 r10_bio->state = 0;
2574 set_bit(R10BIO_ReadError,
2575 &r10_bio->state);
2576 r10_bio->mddev = mddev;
2577 r10_bio->sector = mbio->bi_iter.bi_sector
2578 + sectors_handled;
2579
2580 goto read_more;
2581 } else
2582 generic_make_request(bio);
2583}
2584
2585static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2586{
2587 /* Some sort of write request has finished and it
2588 * succeeded in writing where we thought there was a
2589 * bad block. So forget the bad block.
2590 * Or possibly if failed and we need to record
2591 * a bad block.
2592 */
2593 int m;
2594 struct md_rdev *rdev;
2595
2596 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2597 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2598 for (m = 0; m < conf->copies; m++) {
2599 int dev = r10_bio->devs[m].devnum;
2600 rdev = conf->mirrors[dev].rdev;
2601 if (r10_bio->devs[m].bio == NULL)
2602 continue;
2603 if (!r10_bio->devs[m].bio->bi_error) {
2604 rdev_clear_badblocks(
2605 rdev,
2606 r10_bio->devs[m].addr,
2607 r10_bio->sectors, 0);
2608 } else {
2609 if (!rdev_set_badblocks(
2610 rdev,
2611 r10_bio->devs[m].addr,
2612 r10_bio->sectors, 0))
2613 md_error(conf->mddev, rdev);
2614 }
2615 rdev = conf->mirrors[dev].replacement;
2616 if (r10_bio->devs[m].repl_bio == NULL)
2617 continue;
2618
2619 if (!r10_bio->devs[m].repl_bio->bi_error) {
2620 rdev_clear_badblocks(
2621 rdev,
2622 r10_bio->devs[m].addr,
2623 r10_bio->sectors, 0);
2624 } else {
2625 if (!rdev_set_badblocks(
2626 rdev,
2627 r10_bio->devs[m].addr,
2628 r10_bio->sectors, 0))
2629 md_error(conf->mddev, rdev);
2630 }
2631 }
2632 put_buf(r10_bio);
2633 } else {
2634 bool fail = false;
2635 for (m = 0; m < conf->copies; m++) {
2636 int dev = r10_bio->devs[m].devnum;
2637 struct bio *bio = r10_bio->devs[m].bio;
2638 rdev = conf->mirrors[dev].rdev;
2639 if (bio == IO_MADE_GOOD) {
2640 rdev_clear_badblocks(
2641 rdev,
2642 r10_bio->devs[m].addr,
2643 r10_bio->sectors, 0);
2644 rdev_dec_pending(rdev, conf->mddev);
2645 } else if (bio != NULL && bio->bi_error) {
2646 fail = true;
2647 if (!narrow_write_error(r10_bio, m)) {
2648 md_error(conf->mddev, rdev);
2649 set_bit(R10BIO_Degraded,
2650 &r10_bio->state);
2651 }
2652 rdev_dec_pending(rdev, conf->mddev);
2653 }
2654 bio = r10_bio->devs[m].repl_bio;
2655 rdev = conf->mirrors[dev].replacement;
2656 if (rdev && bio == IO_MADE_GOOD) {
2657 rdev_clear_badblocks(
2658 rdev,
2659 r10_bio->devs[m].addr,
2660 r10_bio->sectors, 0);
2661 rdev_dec_pending(rdev, conf->mddev);
2662 }
2663 }
2664 if (fail) {
2665 spin_lock_irq(&conf->device_lock);
2666 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2667 conf->nr_queued++;
2668 spin_unlock_irq(&conf->device_lock);
2669 md_wakeup_thread(conf->mddev->thread);
2670 } else {
2671 if (test_bit(R10BIO_WriteError,
2672 &r10_bio->state))
2673 close_write(r10_bio);
2674 raid_end_bio_io(r10_bio);
2675 }
2676 }
2677}
2678
2679static void raid10d(struct md_thread *thread)
2680{
2681 struct mddev *mddev = thread->mddev;
2682 struct r10bio *r10_bio;
2683 unsigned long flags;
2684 struct r10conf *conf = mddev->private;
2685 struct list_head *head = &conf->retry_list;
2686 struct blk_plug plug;
2687
2688 md_check_recovery(mddev);
2689
2690 if (!list_empty_careful(&conf->bio_end_io_list) &&
2691 !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2692 LIST_HEAD(tmp);
2693 spin_lock_irqsave(&conf->device_lock, flags);
2694 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2695 while (!list_empty(&conf->bio_end_io_list)) {
2696 list_move(conf->bio_end_io_list.prev, &tmp);
2697 conf->nr_queued--;
2698 }
2699 }
2700 spin_unlock_irqrestore(&conf->device_lock, flags);
2701 while (!list_empty(&tmp)) {
2702 r10_bio = list_first_entry(&tmp, struct r10bio,
2703 retry_list);
2704 list_del(&r10_bio->retry_list);
2705 if (mddev->degraded)
2706 set_bit(R10BIO_Degraded, &r10_bio->state);
2707
2708 if (test_bit(R10BIO_WriteError,
2709 &r10_bio->state))
2710 close_write(r10_bio);
2711 raid_end_bio_io(r10_bio);
2712 }
2713 }
2714
2715 blk_start_plug(&plug);
2716 for (;;) {
2717
2718 flush_pending_writes(conf);
2719
2720 spin_lock_irqsave(&conf->device_lock, flags);
2721 if (list_empty(head)) {
2722 spin_unlock_irqrestore(&conf->device_lock, flags);
2723 break;
2724 }
2725 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2726 list_del(head->prev);
2727 conf->nr_queued--;
2728 spin_unlock_irqrestore(&conf->device_lock, flags);
2729
2730 mddev = r10_bio->mddev;
2731 conf = mddev->private;
2732 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2733 test_bit(R10BIO_WriteError, &r10_bio->state))
2734 handle_write_completed(conf, r10_bio);
2735 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2736 reshape_request_write(mddev, r10_bio);
2737 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2738 sync_request_write(mddev, r10_bio);
2739 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2740 recovery_request_write(mddev, r10_bio);
2741 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2742 handle_read_error(mddev, r10_bio);
2743 else {
2744 /* just a partial read to be scheduled from a
2745 * separate context
2746 */
2747 int slot = r10_bio->read_slot;
2748 generic_make_request(r10_bio->devs[slot].bio);
2749 }
2750
2751 cond_resched();
2752 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2753 md_check_recovery(mddev);
2754 }
2755 blk_finish_plug(&plug);
2756}
2757
2758static int init_resync(struct r10conf *conf)
2759{
2760 int buffs;
2761 int i;
2762
2763 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2764 BUG_ON(conf->r10buf_pool);
2765 conf->have_replacement = 0;
2766 for (i = 0; i < conf->geo.raid_disks; i++)
2767 if (conf->mirrors[i].replacement)
2768 conf->have_replacement = 1;
2769 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2770 if (!conf->r10buf_pool)
2771 return -ENOMEM;
2772 conf->next_resync = 0;
2773 return 0;
2774}
2775
2776/*
2777 * perform a "sync" on one "block"
2778 *
2779 * We need to make sure that no normal I/O request - particularly write
2780 * requests - conflict with active sync requests.
2781 *
2782 * This is achieved by tracking pending requests and a 'barrier' concept
2783 * that can be installed to exclude normal IO requests.
2784 *
2785 * Resync and recovery are handled very differently.
2786 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2787 *
2788 * For resync, we iterate over virtual addresses, read all copies,
2789 * and update if there are differences. If only one copy is live,
2790 * skip it.
2791 * For recovery, we iterate over physical addresses, read a good
2792 * value for each non-in_sync drive, and over-write.
2793 *
2794 * So, for recovery we may have several outstanding complex requests for a
2795 * given address, one for each out-of-sync device. We model this by allocating
2796 * a number of r10_bio structures, one for each out-of-sync device.
2797 * As we setup these structures, we collect all bio's together into a list
2798 * which we then process collectively to add pages, and then process again
2799 * to pass to generic_make_request.
2800 *
2801 * The r10_bio structures are linked using a borrowed master_bio pointer.
2802 * This link is counted in ->remaining. When the r10_bio that points to NULL
2803 * has its remaining count decremented to 0, the whole complex operation
2804 * is complete.
2805 *
2806 */
2807
2808static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2809 int *skipped)
2810{
2811 struct r10conf *conf = mddev->private;
2812 struct r10bio *r10_bio;
2813 struct bio *biolist = NULL, *bio;
2814 sector_t max_sector, nr_sectors;
2815 int i;
2816 int max_sync;
2817 sector_t sync_blocks;
2818 sector_t sectors_skipped = 0;
2819 int chunks_skipped = 0;
2820 sector_t chunk_mask = conf->geo.chunk_mask;
2821
2822 if (!conf->r10buf_pool)
2823 if (init_resync(conf))
2824 return 0;
2825
2826 /*
2827 * Allow skipping a full rebuild for incremental assembly
2828 * of a clean array, like RAID1 does.
2829 */
2830 if (mddev->bitmap == NULL &&
2831 mddev->recovery_cp == MaxSector &&
2832 mddev->reshape_position == MaxSector &&
2833 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2834 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2835 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2836 conf->fullsync == 0) {
2837 *skipped = 1;
2838 return mddev->dev_sectors - sector_nr;
2839 }
2840
2841 skipped:
2842 max_sector = mddev->dev_sectors;
2843 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2844 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2845 max_sector = mddev->resync_max_sectors;
2846 if (sector_nr >= max_sector) {
2847 /* If we aborted, we need to abort the
2848 * sync on the 'current' bitmap chucks (there can
2849 * be several when recovering multiple devices).
2850 * as we may have started syncing it but not finished.
2851 * We can find the current address in
2852 * mddev->curr_resync, but for recovery,
2853 * we need to convert that to several
2854 * virtual addresses.
2855 */
2856 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2857 end_reshape(conf);
2858 close_sync(conf);
2859 return 0;
2860 }
2861
2862 if (mddev->curr_resync < max_sector) { /* aborted */
2863 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2864 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2865 &sync_blocks, 1);
2866 else for (i = 0; i < conf->geo.raid_disks; i++) {
2867 sector_t sect =
2868 raid10_find_virt(conf, mddev->curr_resync, i);
2869 bitmap_end_sync(mddev->bitmap, sect,
2870 &sync_blocks, 1);
2871 }
2872 } else {
2873 /* completed sync */
2874 if ((!mddev->bitmap || conf->fullsync)
2875 && conf->have_replacement
2876 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2877 /* Completed a full sync so the replacements
2878 * are now fully recovered.
2879 */
2880 for (i = 0; i < conf->geo.raid_disks; i++)
2881 if (conf->mirrors[i].replacement)
2882 conf->mirrors[i].replacement
2883 ->recovery_offset
2884 = MaxSector;
2885 }
2886 conf->fullsync = 0;
2887 }
2888 bitmap_close_sync(mddev->bitmap);
2889 close_sync(conf);
2890 *skipped = 1;
2891 return sectors_skipped;
2892 }
2893
2894 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2895 return reshape_request(mddev, sector_nr, skipped);
2896
2897 if (chunks_skipped >= conf->geo.raid_disks) {
2898 /* if there has been nothing to do on any drive,
2899 * then there is nothing to do at all..
2900 */
2901 *skipped = 1;
2902 return (max_sector - sector_nr) + sectors_skipped;
2903 }
2904
2905 if (max_sector > mddev->resync_max)
2906 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2907
2908 /* make sure whole request will fit in a chunk - if chunks
2909 * are meaningful
2910 */
2911 if (conf->geo.near_copies < conf->geo.raid_disks &&
2912 max_sector > (sector_nr | chunk_mask))
2913 max_sector = (sector_nr | chunk_mask) + 1;
2914
2915 /* Again, very different code for resync and recovery.
2916 * Both must result in an r10bio with a list of bios that
2917 * have bi_end_io, bi_sector, bi_bdev set,
2918 * and bi_private set to the r10bio.
2919 * For recovery, we may actually create several r10bios
2920 * with 2 bios in each, that correspond to the bios in the main one.
2921 * In this case, the subordinate r10bios link back through a
2922 * borrowed master_bio pointer, and the counter in the master
2923 * includes a ref from each subordinate.
2924 */
2925 /* First, we decide what to do and set ->bi_end_io
2926 * To end_sync_read if we want to read, and
2927 * end_sync_write if we will want to write.
2928 */
2929
2930 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2931 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2932 /* recovery... the complicated one */
2933 int j;
2934 r10_bio = NULL;
2935
2936 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2937 int still_degraded;
2938 struct r10bio *rb2;
2939 sector_t sect;
2940 int must_sync;
2941 int any_working;
2942 struct raid10_info *mirror = &conf->mirrors[i];
2943
2944 if ((mirror->rdev == NULL ||
2945 test_bit(In_sync, &mirror->rdev->flags))
2946 &&
2947 (mirror->replacement == NULL ||
2948 test_bit(Faulty,
2949 &mirror->replacement->flags)))
2950 continue;
2951
2952 still_degraded = 0;
2953 /* want to reconstruct this device */
2954 rb2 = r10_bio;
2955 sect = raid10_find_virt(conf, sector_nr, i);
2956 if (sect >= mddev->resync_max_sectors) {
2957 /* last stripe is not complete - don't
2958 * try to recover this sector.
2959 */
2960 continue;
2961 }
2962 /* Unless we are doing a full sync, or a replacement
2963 * we only need to recover the block if it is set in
2964 * the bitmap
2965 */
2966 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2967 &sync_blocks, 1);
2968 if (sync_blocks < max_sync)
2969 max_sync = sync_blocks;
2970 if (!must_sync &&
2971 mirror->replacement == NULL &&
2972 !conf->fullsync) {
2973 /* yep, skip the sync_blocks here, but don't assume
2974 * that there will never be anything to do here
2975 */
2976 chunks_skipped = -1;
2977 continue;
2978 }
2979
2980 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2981 r10_bio->state = 0;
2982 raise_barrier(conf, rb2 != NULL);
2983 atomic_set(&r10_bio->remaining, 0);
2984
2985 r10_bio->master_bio = (struct bio*)rb2;
2986 if (rb2)
2987 atomic_inc(&rb2->remaining);
2988 r10_bio->mddev = mddev;
2989 set_bit(R10BIO_IsRecover, &r10_bio->state);
2990 r10_bio->sector = sect;
2991
2992 raid10_find_phys(conf, r10_bio);
2993
2994 /* Need to check if the array will still be
2995 * degraded
2996 */
2997 for (j = 0; j < conf->geo.raid_disks; j++)
2998 if (conf->mirrors[j].rdev == NULL ||
2999 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3000 still_degraded = 1;
3001 break;
3002 }
3003
3004 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3005 &sync_blocks, still_degraded);
3006
3007 any_working = 0;
3008 for (j=0; j<conf->copies;j++) {
3009 int k;
3010 int d = r10_bio->devs[j].devnum;
3011 sector_t from_addr, to_addr;
3012 struct md_rdev *rdev;
3013 sector_t sector, first_bad;
3014 int bad_sectors;
3015 if (!conf->mirrors[d].rdev ||
3016 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3017 continue;
3018 /* This is where we read from */
3019 any_working = 1;
3020 rdev = conf->mirrors[d].rdev;
3021 sector = r10_bio->devs[j].addr;
3022
3023 if (is_badblock(rdev, sector, max_sync,
3024 &first_bad, &bad_sectors)) {
3025 if (first_bad > sector)
3026 max_sync = first_bad - sector;
3027 else {
3028 bad_sectors -= (sector
3029 - first_bad);
3030 if (max_sync > bad_sectors)
3031 max_sync = bad_sectors;
3032 continue;
3033 }
3034 }
3035 bio = r10_bio->devs[0].bio;
3036 bio_reset(bio);
3037 bio->bi_next = biolist;
3038 biolist = bio;
3039 bio->bi_private = r10_bio;
3040 bio->bi_end_io = end_sync_read;
3041 bio->bi_rw = READ;
3042 from_addr = r10_bio->devs[j].addr;
3043 bio->bi_iter.bi_sector = from_addr +
3044 rdev->data_offset;
3045 bio->bi_bdev = rdev->bdev;
3046 atomic_inc(&rdev->nr_pending);
3047 /* and we write to 'i' (if not in_sync) */
3048
3049 for (k=0; k<conf->copies; k++)
3050 if (r10_bio->devs[k].devnum == i)
3051 break;
3052 BUG_ON(k == conf->copies);
3053 to_addr = r10_bio->devs[k].addr;
3054 r10_bio->devs[0].devnum = d;
3055 r10_bio->devs[0].addr = from_addr;
3056 r10_bio->devs[1].devnum = i;
3057 r10_bio->devs[1].addr = to_addr;
3058
3059 rdev = mirror->rdev;
3060 if (!test_bit(In_sync, &rdev->flags)) {
3061 bio = r10_bio->devs[1].bio;
3062 bio_reset(bio);
3063 bio->bi_next = biolist;
3064 biolist = bio;
3065 bio->bi_private = r10_bio;
3066 bio->bi_end_io = end_sync_write;
3067 bio->bi_rw = WRITE;
3068 bio->bi_iter.bi_sector = to_addr
3069 + rdev->data_offset;
3070 bio->bi_bdev = rdev->bdev;
3071 atomic_inc(&r10_bio->remaining);
3072 } else
3073 r10_bio->devs[1].bio->bi_end_io = NULL;
3074
3075 /* and maybe write to replacement */
3076 bio = r10_bio->devs[1].repl_bio;
3077 if (bio)
3078 bio->bi_end_io = NULL;
3079 rdev = mirror->replacement;
3080 /* Note: if rdev != NULL, then bio
3081 * cannot be NULL as r10buf_pool_alloc will
3082 * have allocated it.
3083 * So the second test here is pointless.
3084 * But it keeps semantic-checkers happy, and
3085 * this comment keeps human reviewers
3086 * happy.
3087 */
3088 if (rdev == NULL || bio == NULL ||
3089 test_bit(Faulty, &rdev->flags))
3090 break;
3091 bio_reset(bio);
3092 bio->bi_next = biolist;
3093 biolist = bio;
3094 bio->bi_private = r10_bio;
3095 bio->bi_end_io = end_sync_write;
3096 bio->bi_rw = WRITE;
3097 bio->bi_iter.bi_sector = to_addr +
3098 rdev->data_offset;
3099 bio->bi_bdev = rdev->bdev;
3100 atomic_inc(&r10_bio->remaining);
3101 break;
3102 }
3103 if (j == conf->copies) {
3104 /* Cannot recover, so abort the recovery or
3105 * record a bad block */
3106 if (any_working) {
3107 /* problem is that there are bad blocks
3108 * on other device(s)
3109 */
3110 int k;
3111 for (k = 0; k < conf->copies; k++)
3112 if (r10_bio->devs[k].devnum == i)
3113 break;
3114 if (!test_bit(In_sync,
3115 &mirror->rdev->flags)
3116 && !rdev_set_badblocks(
3117 mirror->rdev,
3118 r10_bio->devs[k].addr,
3119 max_sync, 0))
3120 any_working = 0;
3121 if (mirror->replacement &&
3122 !rdev_set_badblocks(
3123 mirror->replacement,
3124 r10_bio->devs[k].addr,
3125 max_sync, 0))
3126 any_working = 0;
3127 }
3128 if (!any_working) {
3129 if (!test_and_set_bit(MD_RECOVERY_INTR,
3130 &mddev->recovery))
3131 printk(KERN_INFO "md/raid10:%s: insufficient "
3132 "working devices for recovery.\n",
3133 mdname(mddev));
3134 mirror->recovery_disabled
3135 = mddev->recovery_disabled;
3136 }
3137 put_buf(r10_bio);
3138 if (rb2)
3139 atomic_dec(&rb2->remaining);
3140 r10_bio = rb2;
3141 break;
3142 }
3143 }
3144 if (biolist == NULL) {
3145 while (r10_bio) {
3146 struct r10bio *rb2 = r10_bio;
3147 r10_bio = (struct r10bio*) rb2->master_bio;
3148 rb2->master_bio = NULL;
3149 put_buf(rb2);
3150 }
3151 goto giveup;
3152 }
3153 } else {
3154 /* resync. Schedule a read for every block at this virt offset */
3155 int count = 0;
3156
3157 bitmap_cond_end_sync(mddev->bitmap, sector_nr, 0);
3158
3159 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3160 &sync_blocks, mddev->degraded) &&
3161 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3162 &mddev->recovery)) {
3163 /* We can skip this block */
3164 *skipped = 1;
3165 return sync_blocks + sectors_skipped;
3166 }
3167 if (sync_blocks < max_sync)
3168 max_sync = sync_blocks;
3169 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3170 r10_bio->state = 0;
3171
3172 r10_bio->mddev = mddev;
3173 atomic_set(&r10_bio->remaining, 0);
3174 raise_barrier(conf, 0);
3175 conf->next_resync = sector_nr;
3176
3177 r10_bio->master_bio = NULL;
3178 r10_bio->sector = sector_nr;
3179 set_bit(R10BIO_IsSync, &r10_bio->state);
3180 raid10_find_phys(conf, r10_bio);
3181 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3182
3183 for (i = 0; i < conf->copies; i++) {
3184 int d = r10_bio->devs[i].devnum;
3185 sector_t first_bad, sector;
3186 int bad_sectors;
3187
3188 if (r10_bio->devs[i].repl_bio)
3189 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3190
3191 bio = r10_bio->devs[i].bio;
3192 bio_reset(bio);
3193 bio->bi_error = -EIO;
3194 if (conf->mirrors[d].rdev == NULL ||
3195 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3196 continue;
3197 sector = r10_bio->devs[i].addr;
3198 if (is_badblock(conf->mirrors[d].rdev,
3199 sector, max_sync,
3200 &first_bad, &bad_sectors)) {
3201 if (first_bad > sector)
3202 max_sync = first_bad - sector;
3203 else {
3204 bad_sectors -= (sector - first_bad);
3205 if (max_sync > bad_sectors)
3206 max_sync = bad_sectors;
3207 continue;
3208 }
3209 }
3210 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3211 atomic_inc(&r10_bio->remaining);
3212 bio->bi_next = biolist;
3213 biolist = bio;
3214 bio->bi_private = r10_bio;
3215 bio->bi_end_io = end_sync_read;
3216 bio->bi_rw = READ;
3217 bio->bi_iter.bi_sector = sector +
3218 conf->mirrors[d].rdev->data_offset;
3219 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3220 count++;
3221
3222 if (conf->mirrors[d].replacement == NULL ||
3223 test_bit(Faulty,
3224 &conf->mirrors[d].replacement->flags))
3225 continue;
3226
3227 /* Need to set up for writing to the replacement */
3228 bio = r10_bio->devs[i].repl_bio;
3229 bio_reset(bio);
3230 bio->bi_error = -EIO;
3231
3232 sector = r10_bio->devs[i].addr;
3233 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3234 bio->bi_next = biolist;
3235 biolist = bio;
3236 bio->bi_private = r10_bio;
3237 bio->bi_end_io = end_sync_write;
3238 bio->bi_rw = WRITE;
3239 bio->bi_iter.bi_sector = sector +
3240 conf->mirrors[d].replacement->data_offset;
3241 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3242 count++;
3243 }
3244
3245 if (count < 2) {
3246 for (i=0; i<conf->copies; i++) {
3247 int d = r10_bio->devs[i].devnum;
3248 if (r10_bio->devs[i].bio->bi_end_io)
3249 rdev_dec_pending(conf->mirrors[d].rdev,
3250 mddev);
3251 if (r10_bio->devs[i].repl_bio &&
3252 r10_bio->devs[i].repl_bio->bi_end_io)
3253 rdev_dec_pending(
3254 conf->mirrors[d].replacement,
3255 mddev);
3256 }
3257 put_buf(r10_bio);
3258 biolist = NULL;
3259 goto giveup;
3260 }
3261 }
3262
3263 nr_sectors = 0;
3264 if (sector_nr + max_sync < max_sector)
3265 max_sector = sector_nr + max_sync;
3266 do {
3267 struct page *page;
3268 int len = PAGE_SIZE;
3269 if (sector_nr + (len>>9) > max_sector)
3270 len = (max_sector - sector_nr) << 9;
3271 if (len == 0)
3272 break;
3273 for (bio= biolist ; bio ; bio=bio->bi_next) {
3274 struct bio *bio2;
3275 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3276 if (bio_add_page(bio, page, len, 0))
3277 continue;
3278
3279 /* stop here */
3280 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3281 for (bio2 = biolist;
3282 bio2 && bio2 != bio;
3283 bio2 = bio2->bi_next) {
3284 /* remove last page from this bio */
3285 bio2->bi_vcnt--;
3286 bio2->bi_iter.bi_size -= len;
3287 bio_clear_flag(bio2, BIO_SEG_VALID);
3288 }
3289 goto bio_full;
3290 }
3291 nr_sectors += len>>9;
3292 sector_nr += len>>9;
3293 } while (biolist->bi_vcnt < RESYNC_PAGES);
3294 bio_full:
3295 r10_bio->sectors = nr_sectors;
3296
3297 while (biolist) {
3298 bio = biolist;
3299 biolist = biolist->bi_next;
3300
3301 bio->bi_next = NULL;
3302 r10_bio = bio->bi_private;
3303 r10_bio->sectors = nr_sectors;
3304
3305 if (bio->bi_end_io == end_sync_read) {
3306 md_sync_acct(bio->bi_bdev, nr_sectors);
3307 bio->bi_error = 0;
3308 generic_make_request(bio);
3309 }
3310 }
3311
3312 if (sectors_skipped)
3313 /* pretend they weren't skipped, it makes
3314 * no important difference in this case
3315 */
3316 md_done_sync(mddev, sectors_skipped, 1);
3317
3318 return sectors_skipped + nr_sectors;
3319 giveup:
3320 /* There is nowhere to write, so all non-sync
3321 * drives must be failed or in resync, all drives
3322 * have a bad block, so try the next chunk...
3323 */
3324 if (sector_nr + max_sync < max_sector)
3325 max_sector = sector_nr + max_sync;
3326
3327 sectors_skipped += (max_sector - sector_nr);
3328 chunks_skipped ++;
3329 sector_nr = max_sector;
3330 goto skipped;
3331}
3332
3333static sector_t
3334raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3335{
3336 sector_t size;
3337 struct r10conf *conf = mddev->private;
3338
3339 if (!raid_disks)
3340 raid_disks = min(conf->geo.raid_disks,
3341 conf->prev.raid_disks);
3342 if (!sectors)
3343 sectors = conf->dev_sectors;
3344
3345 size = sectors >> conf->geo.chunk_shift;
3346 sector_div(size, conf->geo.far_copies);
3347 size = size * raid_disks;
3348 sector_div(size, conf->geo.near_copies);
3349
3350 return size << conf->geo.chunk_shift;
3351}
3352
3353static void calc_sectors(struct r10conf *conf, sector_t size)
3354{
3355 /* Calculate the number of sectors-per-device that will
3356 * actually be used, and set conf->dev_sectors and
3357 * conf->stride
3358 */
3359
3360 size = size >> conf->geo.chunk_shift;
3361 sector_div(size, conf->geo.far_copies);
3362 size = size * conf->geo.raid_disks;
3363 sector_div(size, conf->geo.near_copies);
3364 /* 'size' is now the number of chunks in the array */
3365 /* calculate "used chunks per device" */
3366 size = size * conf->copies;
3367
3368 /* We need to round up when dividing by raid_disks to
3369 * get the stride size.
3370 */
3371 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3372
3373 conf->dev_sectors = size << conf->geo.chunk_shift;
3374
3375 if (conf->geo.far_offset)
3376 conf->geo.stride = 1 << conf->geo.chunk_shift;
3377 else {
3378 sector_div(size, conf->geo.far_copies);
3379 conf->geo.stride = size << conf->geo.chunk_shift;
3380 }
3381}
3382
3383enum geo_type {geo_new, geo_old, geo_start};
3384static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3385{
3386 int nc, fc, fo;
3387 int layout, chunk, disks;
3388 switch (new) {
3389 case geo_old:
3390 layout = mddev->layout;
3391 chunk = mddev->chunk_sectors;
3392 disks = mddev->raid_disks - mddev->delta_disks;
3393 break;
3394 case geo_new:
3395 layout = mddev->new_layout;
3396 chunk = mddev->new_chunk_sectors;
3397 disks = mddev->raid_disks;
3398 break;
3399 default: /* avoid 'may be unused' warnings */
3400 case geo_start: /* new when starting reshape - raid_disks not
3401 * updated yet. */
3402 layout = mddev->new_layout;
3403 chunk = mddev->new_chunk_sectors;
3404 disks = mddev->raid_disks + mddev->delta_disks;
3405 break;
3406 }
3407 if (layout >> 19)
3408 return -1;
3409 if (chunk < (PAGE_SIZE >> 9) ||
3410 !is_power_of_2(chunk))
3411 return -2;
3412 nc = layout & 255;
3413 fc = (layout >> 8) & 255;
3414 fo = layout & (1<<16);
3415 geo->raid_disks = disks;
3416 geo->near_copies = nc;
3417 geo->far_copies = fc;
3418 geo->far_offset = fo;
3419 switch (layout >> 17) {
3420 case 0: /* original layout. simple but not always optimal */
3421 geo->far_set_size = disks;
3422 break;
3423 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3424 * actually using this, but leave code here just in case.*/
3425 geo->far_set_size = disks/fc;
3426 WARN(geo->far_set_size < fc,
3427 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3428 break;
3429 case 2: /* "improved" layout fixed to match documentation */
3430 geo->far_set_size = fc * nc;
3431 break;
3432 default: /* Not a valid layout */
3433 return -1;
3434 }
3435 geo->chunk_mask = chunk - 1;
3436 geo->chunk_shift = ffz(~chunk);
3437 return nc*fc;
3438}
3439
3440static struct r10conf *setup_conf(struct mddev *mddev)
3441{
3442 struct r10conf *conf = NULL;
3443 int err = -EINVAL;
3444 struct geom geo;
3445 int copies;
3446
3447 copies = setup_geo(&geo, mddev, geo_new);
3448
3449 if (copies == -2) {
3450 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3451 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3452 mdname(mddev), PAGE_SIZE);
3453 goto out;
3454 }
3455
3456 if (copies < 2 || copies > mddev->raid_disks) {
3457 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3458 mdname(mddev), mddev->new_layout);
3459 goto out;
3460 }
3461
3462 err = -ENOMEM;
3463 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3464 if (!conf)
3465 goto out;
3466
3467 /* FIXME calc properly */
3468 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3469 max(0,-mddev->delta_disks)),
3470 GFP_KERNEL);
3471 if (!conf->mirrors)
3472 goto out;
3473
3474 conf->tmppage = alloc_page(GFP_KERNEL);
3475 if (!conf->tmppage)
3476 goto out;
3477
3478 conf->geo = geo;
3479 conf->copies = copies;
3480 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3481 r10bio_pool_free, conf);
3482 if (!conf->r10bio_pool)
3483 goto out;
3484
3485 calc_sectors(conf, mddev->dev_sectors);
3486 if (mddev->reshape_position == MaxSector) {
3487 conf->prev = conf->geo;
3488 conf->reshape_progress = MaxSector;
3489 } else {
3490 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3491 err = -EINVAL;
3492 goto out;
3493 }
3494 conf->reshape_progress = mddev->reshape_position;
3495 if (conf->prev.far_offset)
3496 conf->prev.stride = 1 << conf->prev.chunk_shift;
3497 else
3498 /* far_copies must be 1 */
3499 conf->prev.stride = conf->dev_sectors;
3500 }
3501 conf->reshape_safe = conf->reshape_progress;
3502 spin_lock_init(&conf->device_lock);
3503 INIT_LIST_HEAD(&conf->retry_list);
3504 INIT_LIST_HEAD(&conf->bio_end_io_list);
3505
3506 spin_lock_init(&conf->resync_lock);
3507 init_waitqueue_head(&conf->wait_barrier);
3508
3509 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3510 if (!conf->thread)
3511 goto out;
3512
3513 conf->mddev = mddev;
3514 return conf;
3515
3516 out:
3517 if (err == -ENOMEM)
3518 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3519 mdname(mddev));
3520 if (conf) {
3521 mempool_destroy(conf->r10bio_pool);
3522 kfree(conf->mirrors);
3523 safe_put_page(conf->tmppage);
3524 kfree(conf);
3525 }
3526 return ERR_PTR(err);
3527}
3528
3529static int raid10_run(struct mddev *mddev)
3530{
3531 struct r10conf *conf;
3532 int i, disk_idx, chunk_size;
3533 struct raid10_info *disk;
3534 struct md_rdev *rdev;
3535 sector_t size;
3536 sector_t min_offset_diff = 0;
3537 int first = 1;
3538 bool discard_supported = false;
3539
3540 if (mddev->private == NULL) {
3541 conf = setup_conf(mddev);
3542 if (IS_ERR(conf))
3543 return PTR_ERR(conf);
3544 mddev->private = conf;
3545 }
3546 conf = mddev->private;
3547 if (!conf)
3548 goto out;
3549
3550 mddev->thread = conf->thread;
3551 conf->thread = NULL;
3552
3553 chunk_size = mddev->chunk_sectors << 9;
3554 if (mddev->queue) {
3555 blk_queue_max_discard_sectors(mddev->queue,
3556 mddev->chunk_sectors);
3557 blk_queue_max_write_same_sectors(mddev->queue, 0);
3558 blk_queue_io_min(mddev->queue, chunk_size);
3559 if (conf->geo.raid_disks % conf->geo.near_copies)
3560 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3561 else
3562 blk_queue_io_opt(mddev->queue, chunk_size *
3563 (conf->geo.raid_disks / conf->geo.near_copies));
3564 }
3565
3566 rdev_for_each(rdev, mddev) {
3567 long long diff;
3568 struct request_queue *q;
3569
3570 disk_idx = rdev->raid_disk;
3571 if (disk_idx < 0)
3572 continue;
3573 if (disk_idx >= conf->geo.raid_disks &&
3574 disk_idx >= conf->prev.raid_disks)
3575 continue;
3576 disk = conf->mirrors + disk_idx;
3577
3578 if (test_bit(Replacement, &rdev->flags)) {
3579 if (disk->replacement)
3580 goto out_free_conf;
3581 disk->replacement = rdev;
3582 } else {
3583 if (disk->rdev)
3584 goto out_free_conf;
3585 disk->rdev = rdev;
3586 }
3587 q = bdev_get_queue(rdev->bdev);
3588 diff = (rdev->new_data_offset - rdev->data_offset);
3589 if (!mddev->reshape_backwards)
3590 diff = -diff;
3591 if (diff < 0)
3592 diff = 0;
3593 if (first || diff < min_offset_diff)
3594 min_offset_diff = diff;
3595
3596 if (mddev->gendisk)
3597 disk_stack_limits(mddev->gendisk, rdev->bdev,
3598 rdev->data_offset << 9);
3599
3600 disk->head_position = 0;
3601
3602 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3603 discard_supported = true;
3604 }
3605
3606 if (mddev->queue) {
3607 if (discard_supported)
3608 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3609 mddev->queue);
3610 else
3611 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3612 mddev->queue);
3613 }
3614 /* need to check that every block has at least one working mirror */
3615 if (!enough(conf, -1)) {
3616 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3617 mdname(mddev));
3618 goto out_free_conf;
3619 }
3620
3621 if (conf->reshape_progress != MaxSector) {
3622 /* must ensure that shape change is supported */
3623 if (conf->geo.far_copies != 1 &&
3624 conf->geo.far_offset == 0)
3625 goto out_free_conf;
3626 if (conf->prev.far_copies != 1 &&
3627 conf->prev.far_offset == 0)
3628 goto out_free_conf;
3629 }
3630
3631 mddev->degraded = 0;
3632 for (i = 0;
3633 i < conf->geo.raid_disks
3634 || i < conf->prev.raid_disks;
3635 i++) {
3636
3637 disk = conf->mirrors + i;
3638
3639 if (!disk->rdev && disk->replacement) {
3640 /* The replacement is all we have - use it */
3641 disk->rdev = disk->replacement;
3642 disk->replacement = NULL;
3643 clear_bit(Replacement, &disk->rdev->flags);
3644 }
3645
3646 if (!disk->rdev ||
3647 !test_bit(In_sync, &disk->rdev->flags)) {
3648 disk->head_position = 0;
3649 mddev->degraded++;
3650 if (disk->rdev &&
3651 disk->rdev->saved_raid_disk < 0)
3652 conf->fullsync = 1;
3653 }
3654 disk->recovery_disabled = mddev->recovery_disabled - 1;
3655 }
3656
3657 if (mddev->recovery_cp != MaxSector)
3658 printk(KERN_NOTICE "md/raid10:%s: not clean"
3659 " -- starting background reconstruction\n",
3660 mdname(mddev));
3661 printk(KERN_INFO
3662 "md/raid10:%s: active with %d out of %d devices\n",
3663 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3664 conf->geo.raid_disks);
3665 /*
3666 * Ok, everything is just fine now
3667 */
3668 mddev->dev_sectors = conf->dev_sectors;
3669 size = raid10_size(mddev, 0, 0);
3670 md_set_array_sectors(mddev, size);
3671 mddev->resync_max_sectors = size;
3672
3673 if (mddev->queue) {
3674 int stripe = conf->geo.raid_disks *
3675 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3676
3677 /* Calculate max read-ahead size.
3678 * We need to readahead at least twice a whole stripe....
3679 * maybe...
3680 */
3681 stripe /= conf->geo.near_copies;
3682 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3683 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3684 }
3685
3686 if (md_integrity_register(mddev))
3687 goto out_free_conf;
3688
3689 if (conf->reshape_progress != MaxSector) {
3690 unsigned long before_length, after_length;
3691
3692 before_length = ((1 << conf->prev.chunk_shift) *
3693 conf->prev.far_copies);
3694 after_length = ((1 << conf->geo.chunk_shift) *
3695 conf->geo.far_copies);
3696
3697 if (max(before_length, after_length) > min_offset_diff) {
3698 /* This cannot work */
3699 printk("md/raid10: offset difference not enough to continue reshape\n");
3700 goto out_free_conf;
3701 }
3702 conf->offset_diff = min_offset_diff;
3703
3704 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3705 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3706 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3707 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3708 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3709 "reshape");
3710 }
3711
3712 return 0;
3713
3714out_free_conf:
3715 md_unregister_thread(&mddev->thread);
3716 mempool_destroy(conf->r10bio_pool);
3717 safe_put_page(conf->tmppage);
3718 kfree(conf->mirrors);
3719 kfree(conf);
3720 mddev->private = NULL;
3721out:
3722 return -EIO;
3723}
3724
3725static void raid10_free(struct mddev *mddev, void *priv)
3726{
3727 struct r10conf *conf = priv;
3728
3729 mempool_destroy(conf->r10bio_pool);
3730 safe_put_page(conf->tmppage);
3731 kfree(conf->mirrors);
3732 kfree(conf->mirrors_old);
3733 kfree(conf->mirrors_new);
3734 kfree(conf);
3735}
3736
3737static void raid10_quiesce(struct mddev *mddev, int state)
3738{
3739 struct r10conf *conf = mddev->private;
3740
3741 switch(state) {
3742 case 1:
3743 raise_barrier(conf, 0);
3744 break;
3745 case 0:
3746 lower_barrier(conf);
3747 break;
3748 }
3749}
3750
3751static int raid10_resize(struct mddev *mddev, sector_t sectors)
3752{
3753 /* Resize of 'far' arrays is not supported.
3754 * For 'near' and 'offset' arrays we can set the
3755 * number of sectors used to be an appropriate multiple
3756 * of the chunk size.
3757 * For 'offset', this is far_copies*chunksize.
3758 * For 'near' the multiplier is the LCM of
3759 * near_copies and raid_disks.
3760 * So if far_copies > 1 && !far_offset, fail.
3761 * Else find LCM(raid_disks, near_copy)*far_copies and
3762 * multiply by chunk_size. Then round to this number.
3763 * This is mostly done by raid10_size()
3764 */
3765 struct r10conf *conf = mddev->private;
3766 sector_t oldsize, size;
3767
3768 if (mddev->reshape_position != MaxSector)
3769 return -EBUSY;
3770
3771 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3772 return -EINVAL;
3773
3774 oldsize = raid10_size(mddev, 0, 0);
3775 size = raid10_size(mddev, sectors, 0);
3776 if (mddev->external_size &&
3777 mddev->array_sectors > size)
3778 return -EINVAL;
3779 if (mddev->bitmap) {
3780 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3781 if (ret)
3782 return ret;
3783 }
3784 md_set_array_sectors(mddev, size);
3785 set_capacity(mddev->gendisk, mddev->array_sectors);
3786 revalidate_disk(mddev->gendisk);
3787 if (sectors > mddev->dev_sectors &&
3788 mddev->recovery_cp > oldsize) {
3789 mddev->recovery_cp = oldsize;
3790 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3791 }
3792 calc_sectors(conf, sectors);
3793 mddev->dev_sectors = conf->dev_sectors;
3794 mddev->resync_max_sectors = size;
3795 return 0;
3796}
3797
3798static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3799{
3800 struct md_rdev *rdev;
3801 struct r10conf *conf;
3802
3803 if (mddev->degraded > 0) {
3804 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3805 mdname(mddev));
3806 return ERR_PTR(-EINVAL);
3807 }
3808 sector_div(size, devs);
3809
3810 /* Set new parameters */
3811 mddev->new_level = 10;
3812 /* new layout: far_copies = 1, near_copies = 2 */
3813 mddev->new_layout = (1<<8) + 2;
3814 mddev->new_chunk_sectors = mddev->chunk_sectors;
3815 mddev->delta_disks = mddev->raid_disks;
3816 mddev->raid_disks *= 2;
3817 /* make sure it will be not marked as dirty */
3818 mddev->recovery_cp = MaxSector;
3819 mddev->dev_sectors = size;
3820
3821 conf = setup_conf(mddev);
3822 if (!IS_ERR(conf)) {
3823 rdev_for_each(rdev, mddev)
3824 if (rdev->raid_disk >= 0) {
3825 rdev->new_raid_disk = rdev->raid_disk * 2;
3826 rdev->sectors = size;
3827 }
3828 conf->barrier = 1;
3829 }
3830
3831 return conf;
3832}
3833
3834static void *raid10_takeover(struct mddev *mddev)
3835{
3836 struct r0conf *raid0_conf;
3837
3838 /* raid10 can take over:
3839 * raid0 - providing it has only two drives
3840 */
3841 if (mddev->level == 0) {
3842 /* for raid0 takeover only one zone is supported */
3843 raid0_conf = mddev->private;
3844 if (raid0_conf->nr_strip_zones > 1) {
3845 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3846 " with more than one zone.\n",
3847 mdname(mddev));
3848 return ERR_PTR(-EINVAL);
3849 }
3850 return raid10_takeover_raid0(mddev,
3851 raid0_conf->strip_zone->zone_end,
3852 raid0_conf->strip_zone->nb_dev);
3853 }
3854 return ERR_PTR(-EINVAL);
3855}
3856
3857static int raid10_check_reshape(struct mddev *mddev)
3858{
3859 /* Called when there is a request to change
3860 * - layout (to ->new_layout)
3861 * - chunk size (to ->new_chunk_sectors)
3862 * - raid_disks (by delta_disks)
3863 * or when trying to restart a reshape that was ongoing.
3864 *
3865 * We need to validate the request and possibly allocate
3866 * space if that might be an issue later.
3867 *
3868 * Currently we reject any reshape of a 'far' mode array,
3869 * allow chunk size to change if new is generally acceptable,
3870 * allow raid_disks to increase, and allow
3871 * a switch between 'near' mode and 'offset' mode.
3872 */
3873 struct r10conf *conf = mddev->private;
3874 struct geom geo;
3875
3876 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3877 return -EINVAL;
3878
3879 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3880 /* mustn't change number of copies */
3881 return -EINVAL;
3882 if (geo.far_copies > 1 && !geo.far_offset)
3883 /* Cannot switch to 'far' mode */
3884 return -EINVAL;
3885
3886 if (mddev->array_sectors & geo.chunk_mask)
3887 /* not factor of array size */
3888 return -EINVAL;
3889
3890 if (!enough(conf, -1))
3891 return -EINVAL;
3892
3893 kfree(conf->mirrors_new);
3894 conf->mirrors_new = NULL;
3895 if (mddev->delta_disks > 0) {
3896 /* allocate new 'mirrors' list */
3897 conf->mirrors_new = kzalloc(
3898 sizeof(struct raid10_info)
3899 *(mddev->raid_disks +
3900 mddev->delta_disks),
3901 GFP_KERNEL);
3902 if (!conf->mirrors_new)
3903 return -ENOMEM;
3904 }
3905 return 0;
3906}
3907
3908/*
3909 * Need to check if array has failed when deciding whether to:
3910 * - start an array
3911 * - remove non-faulty devices
3912 * - add a spare
3913 * - allow a reshape
3914 * This determination is simple when no reshape is happening.
3915 * However if there is a reshape, we need to carefully check
3916 * both the before and after sections.
3917 * This is because some failed devices may only affect one
3918 * of the two sections, and some non-in_sync devices may
3919 * be insync in the section most affected by failed devices.
3920 */
3921static int calc_degraded(struct r10conf *conf)
3922{
3923 int degraded, degraded2;
3924 int i;
3925
3926 rcu_read_lock();
3927 degraded = 0;
3928 /* 'prev' section first */
3929 for (i = 0; i < conf->prev.raid_disks; i++) {
3930 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3931 if (!rdev || test_bit(Faulty, &rdev->flags))
3932 degraded++;
3933 else if (!test_bit(In_sync, &rdev->flags))
3934 /* When we can reduce the number of devices in
3935 * an array, this might not contribute to
3936 * 'degraded'. It does now.
3937 */
3938 degraded++;
3939 }
3940 rcu_read_unlock();
3941 if (conf->geo.raid_disks == conf->prev.raid_disks)
3942 return degraded;
3943 rcu_read_lock();
3944 degraded2 = 0;
3945 for (i = 0; i < conf->geo.raid_disks; i++) {
3946 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3947 if (!rdev || test_bit(Faulty, &rdev->flags))
3948 degraded2++;
3949 else if (!test_bit(In_sync, &rdev->flags)) {
3950 /* If reshape is increasing the number of devices,
3951 * this section has already been recovered, so
3952 * it doesn't contribute to degraded.
3953 * else it does.
3954 */
3955 if (conf->geo.raid_disks <= conf->prev.raid_disks)
3956 degraded2++;
3957 }
3958 }
3959 rcu_read_unlock();
3960 if (degraded2 > degraded)
3961 return degraded2;
3962 return degraded;
3963}
3964
3965static int raid10_start_reshape(struct mddev *mddev)
3966{
3967 /* A 'reshape' has been requested. This commits
3968 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3969 * This also checks if there are enough spares and adds them
3970 * to the array.
3971 * We currently require enough spares to make the final
3972 * array non-degraded. We also require that the difference
3973 * between old and new data_offset - on each device - is
3974 * enough that we never risk over-writing.
3975 */
3976
3977 unsigned long before_length, after_length;
3978 sector_t min_offset_diff = 0;
3979 int first = 1;
3980 struct geom new;
3981 struct r10conf *conf = mddev->private;
3982 struct md_rdev *rdev;
3983 int spares = 0;
3984 int ret;
3985
3986 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3987 return -EBUSY;
3988
3989 if (setup_geo(&new, mddev, geo_start) != conf->copies)
3990 return -EINVAL;
3991
3992 before_length = ((1 << conf->prev.chunk_shift) *
3993 conf->prev.far_copies);
3994 after_length = ((1 << conf->geo.chunk_shift) *
3995 conf->geo.far_copies);
3996
3997 rdev_for_each(rdev, mddev) {
3998 if (!test_bit(In_sync, &rdev->flags)
3999 && !test_bit(Faulty, &rdev->flags))
4000 spares++;
4001 if (rdev->raid_disk >= 0) {
4002 long long diff = (rdev->new_data_offset
4003 - rdev->data_offset);
4004 if (!mddev->reshape_backwards)
4005 diff = -diff;
4006 if (diff < 0)
4007 diff = 0;
4008 if (first || diff < min_offset_diff)
4009 min_offset_diff = diff;
4010 }
4011 }
4012
4013 if (max(before_length, after_length) > min_offset_diff)
4014 return -EINVAL;
4015
4016 if (spares < mddev->delta_disks)
4017 return -EINVAL;
4018
4019 conf->offset_diff = min_offset_diff;
4020 spin_lock_irq(&conf->device_lock);
4021 if (conf->mirrors_new) {
4022 memcpy(conf->mirrors_new, conf->mirrors,
4023 sizeof(struct raid10_info)*conf->prev.raid_disks);
4024 smp_mb();
4025 kfree(conf->mirrors_old);
4026 conf->mirrors_old = conf->mirrors;
4027 conf->mirrors = conf->mirrors_new;
4028 conf->mirrors_new = NULL;
4029 }
4030 setup_geo(&conf->geo, mddev, geo_start);
4031 smp_mb();
4032 if (mddev->reshape_backwards) {
4033 sector_t size = raid10_size(mddev, 0, 0);
4034 if (size < mddev->array_sectors) {
4035 spin_unlock_irq(&conf->device_lock);
4036 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4037 mdname(mddev));
4038 return -EINVAL;
4039 }
4040 mddev->resync_max_sectors = size;
4041 conf->reshape_progress = size;
4042 } else
4043 conf->reshape_progress = 0;
4044 conf->reshape_safe = conf->reshape_progress;
4045 spin_unlock_irq(&conf->device_lock);
4046
4047 if (mddev->delta_disks && mddev->bitmap) {
4048 ret = bitmap_resize(mddev->bitmap,
4049 raid10_size(mddev, 0,
4050 conf->geo.raid_disks),
4051 0, 0);
4052 if (ret)
4053 goto abort;
4054 }
4055 if (mddev->delta_disks > 0) {
4056 rdev_for_each(rdev, mddev)
4057 if (rdev->raid_disk < 0 &&
4058 !test_bit(Faulty, &rdev->flags)) {
4059 if (raid10_add_disk(mddev, rdev) == 0) {
4060 if (rdev->raid_disk >=
4061 conf->prev.raid_disks)
4062 set_bit(In_sync, &rdev->flags);
4063 else
4064 rdev->recovery_offset = 0;
4065
4066 if (sysfs_link_rdev(mddev, rdev))
4067 /* Failure here is OK */;
4068 }
4069 } else if (rdev->raid_disk >= conf->prev.raid_disks
4070 && !test_bit(Faulty, &rdev->flags)) {
4071 /* This is a spare that was manually added */
4072 set_bit(In_sync, &rdev->flags);
4073 }
4074 }
4075 /* When a reshape changes the number of devices,
4076 * ->degraded is measured against the larger of the
4077 * pre and post numbers.
4078 */
4079 spin_lock_irq(&conf->device_lock);
4080 mddev->degraded = calc_degraded(conf);
4081 spin_unlock_irq(&conf->device_lock);
4082 mddev->raid_disks = conf->geo.raid_disks;
4083 mddev->reshape_position = conf->reshape_progress;
4084 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4085
4086 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4087 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4088 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4089 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4090 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4091
4092 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4093 "reshape");
4094 if (!mddev->sync_thread) {
4095 ret = -EAGAIN;
4096 goto abort;
4097 }
4098 conf->reshape_checkpoint = jiffies;
4099 md_wakeup_thread(mddev->sync_thread);
4100 md_new_event(mddev);
4101 return 0;
4102
4103abort:
4104 mddev->recovery = 0;
4105 spin_lock_irq(&conf->device_lock);
4106 conf->geo = conf->prev;
4107 mddev->raid_disks = conf->geo.raid_disks;
4108 rdev_for_each(rdev, mddev)
4109 rdev->new_data_offset = rdev->data_offset;
4110 smp_wmb();
4111 conf->reshape_progress = MaxSector;
4112 conf->reshape_safe = MaxSector;
4113 mddev->reshape_position = MaxSector;
4114 spin_unlock_irq(&conf->device_lock);
4115 return ret;
4116}
4117
4118/* Calculate the last device-address that could contain
4119 * any block from the chunk that includes the array-address 's'
4120 * and report the next address.
4121 * i.e. the address returned will be chunk-aligned and after
4122 * any data that is in the chunk containing 's'.
4123 */
4124static sector_t last_dev_address(sector_t s, struct geom *geo)
4125{
4126 s = (s | geo->chunk_mask) + 1;
4127 s >>= geo->chunk_shift;
4128 s *= geo->near_copies;
4129 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4130 s *= geo->far_copies;
4131 s <<= geo->chunk_shift;
4132 return s;
4133}
4134
4135/* Calculate the first device-address that could contain
4136 * any block from the chunk that includes the array-address 's'.
4137 * This too will be the start of a chunk
4138 */
4139static sector_t first_dev_address(sector_t s, struct geom *geo)
4140{
4141 s >>= geo->chunk_shift;
4142 s *= geo->near_copies;
4143 sector_div(s, geo->raid_disks);
4144 s *= geo->far_copies;
4145 s <<= geo->chunk_shift;
4146 return s;
4147}
4148
4149static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4150 int *skipped)
4151{
4152 /* We simply copy at most one chunk (smallest of old and new)
4153 * at a time, possibly less if that exceeds RESYNC_PAGES,
4154 * or we hit a bad block or something.
4155 * This might mean we pause for normal IO in the middle of
4156 * a chunk, but that is not a problem as mddev->reshape_position
4157 * can record any location.
4158 *
4159 * If we will want to write to a location that isn't
4160 * yet recorded as 'safe' (i.e. in metadata on disk) then
4161 * we need to flush all reshape requests and update the metadata.
4162 *
4163 * When reshaping forwards (e.g. to more devices), we interpret
4164 * 'safe' as the earliest block which might not have been copied
4165 * down yet. We divide this by previous stripe size and multiply
4166 * by previous stripe length to get lowest device offset that we
4167 * cannot write to yet.
4168 * We interpret 'sector_nr' as an address that we want to write to.
4169 * From this we use last_device_address() to find where we might
4170 * write to, and first_device_address on the 'safe' position.
4171 * If this 'next' write position is after the 'safe' position,
4172 * we must update the metadata to increase the 'safe' position.
4173 *
4174 * When reshaping backwards, we round in the opposite direction
4175 * and perform the reverse test: next write position must not be
4176 * less than current safe position.
4177 *
4178 * In all this the minimum difference in data offsets
4179 * (conf->offset_diff - always positive) allows a bit of slack,
4180 * so next can be after 'safe', but not by more than offset_diff
4181 *
4182 * We need to prepare all the bios here before we start any IO
4183 * to ensure the size we choose is acceptable to all devices.
4184 * The means one for each copy for write-out and an extra one for
4185 * read-in.
4186 * We store the read-in bio in ->master_bio and the others in
4187 * ->devs[x].bio and ->devs[x].repl_bio.
4188 */
4189 struct r10conf *conf = mddev->private;
4190 struct r10bio *r10_bio;
4191 sector_t next, safe, last;
4192 int max_sectors;
4193 int nr_sectors;
4194 int s;
4195 struct md_rdev *rdev;
4196 int need_flush = 0;
4197 struct bio *blist;
4198 struct bio *bio, *read_bio;
4199 int sectors_done = 0;
4200
4201 if (sector_nr == 0) {
4202 /* If restarting in the middle, skip the initial sectors */
4203 if (mddev->reshape_backwards &&
4204 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4205 sector_nr = (raid10_size(mddev, 0, 0)
4206 - conf->reshape_progress);
4207 } else if (!mddev->reshape_backwards &&
4208 conf->reshape_progress > 0)
4209 sector_nr = conf->reshape_progress;
4210 if (sector_nr) {
4211 mddev->curr_resync_completed = sector_nr;
4212 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4213 *skipped = 1;
4214 return sector_nr;
4215 }
4216 }
4217
4218 /* We don't use sector_nr to track where we are up to
4219 * as that doesn't work well for ->reshape_backwards.
4220 * So just use ->reshape_progress.
4221 */
4222 if (mddev->reshape_backwards) {
4223 /* 'next' is the earliest device address that we might
4224 * write to for this chunk in the new layout
4225 */
4226 next = first_dev_address(conf->reshape_progress - 1,
4227 &conf->geo);
4228
4229 /* 'safe' is the last device address that we might read from
4230 * in the old layout after a restart
4231 */
4232 safe = last_dev_address(conf->reshape_safe - 1,
4233 &conf->prev);
4234
4235 if (next + conf->offset_diff < safe)
4236 need_flush = 1;
4237
4238 last = conf->reshape_progress - 1;
4239 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4240 & conf->prev.chunk_mask);
4241 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4242 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4243 } else {
4244 /* 'next' is after the last device address that we
4245 * might write to for this chunk in the new layout
4246 */
4247 next = last_dev_address(conf->reshape_progress, &conf->geo);
4248
4249 /* 'safe' is the earliest device address that we might
4250 * read from in the old layout after a restart
4251 */
4252 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4253
4254 /* Need to update metadata if 'next' might be beyond 'safe'
4255 * as that would possibly corrupt data
4256 */
4257 if (next > safe + conf->offset_diff)
4258 need_flush = 1;
4259
4260 sector_nr = conf->reshape_progress;
4261 last = sector_nr | (conf->geo.chunk_mask
4262 & conf->prev.chunk_mask);
4263
4264 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4265 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4266 }
4267
4268 if (need_flush ||
4269 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4270 /* Need to update reshape_position in metadata */
4271 wait_barrier(conf);
4272 mddev->reshape_position = conf->reshape_progress;
4273 if (mddev->reshape_backwards)
4274 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4275 - conf->reshape_progress;
4276 else
4277 mddev->curr_resync_completed = conf->reshape_progress;
4278 conf->reshape_checkpoint = jiffies;
4279 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4280 md_wakeup_thread(mddev->thread);
4281 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4282 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4283 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4284 allow_barrier(conf);
4285 return sectors_done;
4286 }
4287 conf->reshape_safe = mddev->reshape_position;
4288 allow_barrier(conf);
4289 }
4290
4291read_more:
4292 /* Now schedule reads for blocks from sector_nr to last */
4293 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4294 r10_bio->state = 0;
4295 raise_barrier(conf, sectors_done != 0);
4296 atomic_set(&r10_bio->remaining, 0);
4297 r10_bio->mddev = mddev;
4298 r10_bio->sector = sector_nr;
4299 set_bit(R10BIO_IsReshape, &r10_bio->state);
4300 r10_bio->sectors = last - sector_nr + 1;
4301 rdev = read_balance(conf, r10_bio, &max_sectors);
4302 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4303
4304 if (!rdev) {
4305 /* Cannot read from here, so need to record bad blocks
4306 * on all the target devices.
4307 */
4308 // FIXME
4309 mempool_free(r10_bio, conf->r10buf_pool);
4310 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4311 return sectors_done;
4312 }
4313
4314 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4315
4316 read_bio->bi_bdev = rdev->bdev;
4317 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4318 + rdev->data_offset);
4319 read_bio->bi_private = r10_bio;
4320 read_bio->bi_end_io = end_sync_read;
4321 read_bio->bi_rw = READ;
4322 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4323 read_bio->bi_error = 0;
4324 read_bio->bi_vcnt = 0;
4325 read_bio->bi_iter.bi_size = 0;
4326 r10_bio->master_bio = read_bio;
4327 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4328
4329 /* Now find the locations in the new layout */
4330 __raid10_find_phys(&conf->geo, r10_bio);
4331
4332 blist = read_bio;
4333 read_bio->bi_next = NULL;
4334
4335 for (s = 0; s < conf->copies*2; s++) {
4336 struct bio *b;
4337 int d = r10_bio->devs[s/2].devnum;
4338 struct md_rdev *rdev2;
4339 if (s&1) {
4340 rdev2 = conf->mirrors[d].replacement;
4341 b = r10_bio->devs[s/2].repl_bio;
4342 } else {
4343 rdev2 = conf->mirrors[d].rdev;
4344 b = r10_bio->devs[s/2].bio;
4345 }
4346 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4347 continue;
4348
4349 bio_reset(b);
4350 b->bi_bdev = rdev2->bdev;
4351 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4352 rdev2->new_data_offset;
4353 b->bi_private = r10_bio;
4354 b->bi_end_io = end_reshape_write;
4355 b->bi_rw = WRITE;
4356 b->bi_next = blist;
4357 blist = b;
4358 }
4359
4360 /* Now add as many pages as possible to all of these bios. */
4361
4362 nr_sectors = 0;
4363 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4364 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4365 int len = (max_sectors - s) << 9;
4366 if (len > PAGE_SIZE)
4367 len = PAGE_SIZE;
4368 for (bio = blist; bio ; bio = bio->bi_next) {
4369 struct bio *bio2;
4370 if (bio_add_page(bio, page, len, 0))
4371 continue;
4372
4373 /* Didn't fit, must stop */
4374 for (bio2 = blist;
4375 bio2 && bio2 != bio;
4376 bio2 = bio2->bi_next) {
4377 /* Remove last page from this bio */
4378 bio2->bi_vcnt--;
4379 bio2->bi_iter.bi_size -= len;
4380 bio_clear_flag(bio2, BIO_SEG_VALID);
4381 }
4382 goto bio_full;
4383 }
4384 sector_nr += len >> 9;
4385 nr_sectors += len >> 9;
4386 }
4387bio_full:
4388 r10_bio->sectors = nr_sectors;
4389
4390 /* Now submit the read */
4391 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4392 atomic_inc(&r10_bio->remaining);
4393 read_bio->bi_next = NULL;
4394 generic_make_request(read_bio);
4395 sector_nr += nr_sectors;
4396 sectors_done += nr_sectors;
4397 if (sector_nr <= last)
4398 goto read_more;
4399
4400 /* Now that we have done the whole section we can
4401 * update reshape_progress
4402 */
4403 if (mddev->reshape_backwards)
4404 conf->reshape_progress -= sectors_done;
4405 else
4406 conf->reshape_progress += sectors_done;
4407
4408 return sectors_done;
4409}
4410
4411static void end_reshape_request(struct r10bio *r10_bio);
4412static int handle_reshape_read_error(struct mddev *mddev,
4413 struct r10bio *r10_bio);
4414static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4415{
4416 /* Reshape read completed. Hopefully we have a block
4417 * to write out.
4418 * If we got a read error then we do sync 1-page reads from
4419 * elsewhere until we find the data - or give up.
4420 */
4421 struct r10conf *conf = mddev->private;
4422 int s;
4423
4424 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4425 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4426 /* Reshape has been aborted */
4427 md_done_sync(mddev, r10_bio->sectors, 0);
4428 return;
4429 }
4430
4431 /* We definitely have the data in the pages, schedule the
4432 * writes.
4433 */
4434 atomic_set(&r10_bio->remaining, 1);
4435 for (s = 0; s < conf->copies*2; s++) {
4436 struct bio *b;
4437 int d = r10_bio->devs[s/2].devnum;
4438 struct md_rdev *rdev;
4439 if (s&1) {
4440 rdev = conf->mirrors[d].replacement;
4441 b = r10_bio->devs[s/2].repl_bio;
4442 } else {
4443 rdev = conf->mirrors[d].rdev;
4444 b = r10_bio->devs[s/2].bio;
4445 }
4446 if (!rdev || test_bit(Faulty, &rdev->flags))
4447 continue;
4448 atomic_inc(&rdev->nr_pending);
4449 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4450 atomic_inc(&r10_bio->remaining);
4451 b->bi_next = NULL;
4452 generic_make_request(b);
4453 }
4454 end_reshape_request(r10_bio);
4455}
4456
4457static void end_reshape(struct r10conf *conf)
4458{
4459 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4460 return;
4461
4462 spin_lock_irq(&conf->device_lock);
4463 conf->prev = conf->geo;
4464 md_finish_reshape(conf->mddev);
4465 smp_wmb();
4466 conf->reshape_progress = MaxSector;
4467 conf->reshape_safe = MaxSector;
4468 spin_unlock_irq(&conf->device_lock);
4469
4470 /* read-ahead size must cover two whole stripes, which is
4471 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4472 */
4473 if (conf->mddev->queue) {
4474 int stripe = conf->geo.raid_disks *
4475 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4476 stripe /= conf->geo.near_copies;
4477 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4478 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4479 }
4480 conf->fullsync = 0;
4481}
4482
4483static int handle_reshape_read_error(struct mddev *mddev,
4484 struct r10bio *r10_bio)
4485{
4486 /* Use sync reads to get the blocks from somewhere else */
4487 int sectors = r10_bio->sectors;
4488 struct r10conf *conf = mddev->private;
4489 struct {
4490 struct r10bio r10_bio;
4491 struct r10dev devs[conf->copies];
4492 } on_stack;
4493 struct r10bio *r10b = &on_stack.r10_bio;
4494 int slot = 0;
4495 int idx = 0;
4496 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4497
4498 r10b->sector = r10_bio->sector;
4499 __raid10_find_phys(&conf->prev, r10b);
4500
4501 while (sectors) {
4502 int s = sectors;
4503 int success = 0;
4504 int first_slot = slot;
4505
4506 if (s > (PAGE_SIZE >> 9))
4507 s = PAGE_SIZE >> 9;
4508
4509 while (!success) {
4510 int d = r10b->devs[slot].devnum;
4511 struct md_rdev *rdev = conf->mirrors[d].rdev;
4512 sector_t addr;
4513 if (rdev == NULL ||
4514 test_bit(Faulty, &rdev->flags) ||
4515 !test_bit(In_sync, &rdev->flags))
4516 goto failed;
4517
4518 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4519 success = sync_page_io(rdev,
4520 addr,
4521 s << 9,
4522 bvec[idx].bv_page,
4523 READ, false);
4524 if (success)
4525 break;
4526 failed:
4527 slot++;
4528 if (slot >= conf->copies)
4529 slot = 0;
4530 if (slot == first_slot)
4531 break;
4532 }
4533 if (!success) {
4534 /* couldn't read this block, must give up */
4535 set_bit(MD_RECOVERY_INTR,
4536 &mddev->recovery);
4537 return -EIO;
4538 }
4539 sectors -= s;
4540 idx++;
4541 }
4542 return 0;
4543}
4544
4545static void end_reshape_write(struct bio *bio)
4546{
4547 struct r10bio *r10_bio = bio->bi_private;
4548 struct mddev *mddev = r10_bio->mddev;
4549 struct r10conf *conf = mddev->private;
4550 int d;
4551 int slot;
4552 int repl;
4553 struct md_rdev *rdev = NULL;
4554
4555 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4556 if (repl)
4557 rdev = conf->mirrors[d].replacement;
4558 if (!rdev) {
4559 smp_mb();
4560 rdev = conf->mirrors[d].rdev;
4561 }
4562
4563 if (bio->bi_error) {
4564 /* FIXME should record badblock */
4565 md_error(mddev, rdev);
4566 }
4567
4568 rdev_dec_pending(rdev, mddev);
4569 end_reshape_request(r10_bio);
4570}
4571
4572static void end_reshape_request(struct r10bio *r10_bio)
4573{
4574 if (!atomic_dec_and_test(&r10_bio->remaining))
4575 return;
4576 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4577 bio_put(r10_bio->master_bio);
4578 put_buf(r10_bio);
4579}
4580
4581static void raid10_finish_reshape(struct mddev *mddev)
4582{
4583 struct r10conf *conf = mddev->private;
4584
4585 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4586 return;
4587
4588 if (mddev->delta_disks > 0) {
4589 sector_t size = raid10_size(mddev, 0, 0);
4590 md_set_array_sectors(mddev, size);
4591 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4592 mddev->recovery_cp = mddev->resync_max_sectors;
4593 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4594 }
4595 mddev->resync_max_sectors = size;
4596 set_capacity(mddev->gendisk, mddev->array_sectors);
4597 revalidate_disk(mddev->gendisk);
4598 } else {
4599 int d;
4600 for (d = conf->geo.raid_disks ;
4601 d < conf->geo.raid_disks - mddev->delta_disks;
4602 d++) {
4603 struct md_rdev *rdev = conf->mirrors[d].rdev;
4604 if (rdev)
4605 clear_bit(In_sync, &rdev->flags);
4606 rdev = conf->mirrors[d].replacement;
4607 if (rdev)
4608 clear_bit(In_sync, &rdev->flags);
4609 }
4610 }
4611 mddev->layout = mddev->new_layout;
4612 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4613 mddev->reshape_position = MaxSector;
4614 mddev->delta_disks = 0;
4615 mddev->reshape_backwards = 0;
4616}
4617
4618static struct md_personality raid10_personality =
4619{
4620 .name = "raid10",
4621 .level = 10,
4622 .owner = THIS_MODULE,
4623 .make_request = raid10_make_request,
4624 .run = raid10_run,
4625 .free = raid10_free,
4626 .status = raid10_status,
4627 .error_handler = raid10_error,
4628 .hot_add_disk = raid10_add_disk,
4629 .hot_remove_disk= raid10_remove_disk,
4630 .spare_active = raid10_spare_active,
4631 .sync_request = raid10_sync_request,
4632 .quiesce = raid10_quiesce,
4633 .size = raid10_size,
4634 .resize = raid10_resize,
4635 .takeover = raid10_takeover,
4636 .check_reshape = raid10_check_reshape,
4637 .start_reshape = raid10_start_reshape,
4638 .finish_reshape = raid10_finish_reshape,
4639 .congested = raid10_congested,
4640};
4641
4642static int __init raid_init(void)
4643{
4644 return register_md_personality(&raid10_personality);
4645}
4646
4647static void raid_exit(void)
4648{
4649 unregister_md_personality(&raid10_personality);
4650}
4651
4652module_init(raid_init);
4653module_exit(raid_exit);
4654MODULE_LICENSE("GPL");
4655MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4656MODULE_ALIAS("md-personality-9"); /* RAID10 */
4657MODULE_ALIAS("md-raid10");
4658MODULE_ALIAS("md-level-10");
4659
4660module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
1/*
2 * raid10.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 2000-2004 Neil Brown
5 *
6 * RAID-10 support for md.
7 *
8 * Base on code in raid1.c. See raid1.c for further copyright information.
9 *
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20
21#include <linux/slab.h>
22#include <linux/delay.h>
23#include <linux/blkdev.h>
24#include <linux/seq_file.h>
25#include <linux/ratelimit.h>
26#include "md.h"
27#include "raid10.h"
28#include "raid0.h"
29#include "bitmap.h"
30
31/*
32 * RAID10 provides a combination of RAID0 and RAID1 functionality.
33 * The layout of data is defined by
34 * chunk_size
35 * raid_disks
36 * near_copies (stored in low byte of layout)
37 * far_copies (stored in second byte of layout)
38 * far_offset (stored in bit 16 of layout )
39 *
40 * The data to be stored is divided into chunks using chunksize.
41 * Each device is divided into far_copies sections.
42 * In each section, chunks are laid out in a style similar to raid0, but
43 * near_copies copies of each chunk is stored (each on a different drive).
44 * The starting device for each section is offset near_copies from the starting
45 * device of the previous section.
46 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
47 * drive.
48 * near_copies and far_copies must be at least one, and their product is at most
49 * raid_disks.
50 *
51 * If far_offset is true, then the far_copies are handled a bit differently.
52 * The copies are still in different stripes, but instead of be very far apart
53 * on disk, there are adjacent stripes.
54 */
55
56/*
57 * Number of guaranteed r10bios in case of extreme VM load:
58 */
59#define NR_RAID10_BIOS 256
60
61static void allow_barrier(conf_t *conf);
62static void lower_barrier(conf_t *conf);
63
64static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
65{
66 conf_t *conf = data;
67 int size = offsetof(struct r10bio_s, devs[conf->copies]);
68
69 /* allocate a r10bio with room for raid_disks entries in the bios array */
70 return kzalloc(size, gfp_flags);
71}
72
73static void r10bio_pool_free(void *r10_bio, void *data)
74{
75 kfree(r10_bio);
76}
77
78/* Maximum size of each resync request */
79#define RESYNC_BLOCK_SIZE (64*1024)
80#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
81/* amount of memory to reserve for resync requests */
82#define RESYNC_WINDOW (1024*1024)
83/* maximum number of concurrent requests, memory permitting */
84#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
85
86/*
87 * When performing a resync, we need to read and compare, so
88 * we need as many pages are there are copies.
89 * When performing a recovery, we need 2 bios, one for read,
90 * one for write (we recover only one drive per r10buf)
91 *
92 */
93static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
94{
95 conf_t *conf = data;
96 struct page *page;
97 r10bio_t *r10_bio;
98 struct bio *bio;
99 int i, j;
100 int nalloc;
101
102 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
103 if (!r10_bio)
104 return NULL;
105
106 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
107 nalloc = conf->copies; /* resync */
108 else
109 nalloc = 2; /* recovery */
110
111 /*
112 * Allocate bios.
113 */
114 for (j = nalloc ; j-- ; ) {
115 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
116 if (!bio)
117 goto out_free_bio;
118 r10_bio->devs[j].bio = bio;
119 }
120 /*
121 * Allocate RESYNC_PAGES data pages and attach them
122 * where needed.
123 */
124 for (j = 0 ; j < nalloc; j++) {
125 bio = r10_bio->devs[j].bio;
126 for (i = 0; i < RESYNC_PAGES; i++) {
127 if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
128 &conf->mddev->recovery)) {
129 /* we can share bv_page's during recovery */
130 struct bio *rbio = r10_bio->devs[0].bio;
131 page = rbio->bi_io_vec[i].bv_page;
132 get_page(page);
133 } else
134 page = alloc_page(gfp_flags);
135 if (unlikely(!page))
136 goto out_free_pages;
137
138 bio->bi_io_vec[i].bv_page = page;
139 }
140 }
141
142 return r10_bio;
143
144out_free_pages:
145 for ( ; i > 0 ; i--)
146 safe_put_page(bio->bi_io_vec[i-1].bv_page);
147 while (j--)
148 for (i = 0; i < RESYNC_PAGES ; i++)
149 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
150 j = -1;
151out_free_bio:
152 while ( ++j < nalloc )
153 bio_put(r10_bio->devs[j].bio);
154 r10bio_pool_free(r10_bio, conf);
155 return NULL;
156}
157
158static void r10buf_pool_free(void *__r10_bio, void *data)
159{
160 int i;
161 conf_t *conf = data;
162 r10bio_t *r10bio = __r10_bio;
163 int j;
164
165 for (j=0; j < conf->copies; j++) {
166 struct bio *bio = r10bio->devs[j].bio;
167 if (bio) {
168 for (i = 0; i < RESYNC_PAGES; i++) {
169 safe_put_page(bio->bi_io_vec[i].bv_page);
170 bio->bi_io_vec[i].bv_page = NULL;
171 }
172 bio_put(bio);
173 }
174 }
175 r10bio_pool_free(r10bio, conf);
176}
177
178static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
179{
180 int i;
181
182 for (i = 0; i < conf->copies; i++) {
183 struct bio **bio = & r10_bio->devs[i].bio;
184 if (!BIO_SPECIAL(*bio))
185 bio_put(*bio);
186 *bio = NULL;
187 }
188}
189
190static void free_r10bio(r10bio_t *r10_bio)
191{
192 conf_t *conf = r10_bio->mddev->private;
193
194 put_all_bios(conf, r10_bio);
195 mempool_free(r10_bio, conf->r10bio_pool);
196}
197
198static void put_buf(r10bio_t *r10_bio)
199{
200 conf_t *conf = r10_bio->mddev->private;
201
202 mempool_free(r10_bio, conf->r10buf_pool);
203
204 lower_barrier(conf);
205}
206
207static void reschedule_retry(r10bio_t *r10_bio)
208{
209 unsigned long flags;
210 mddev_t *mddev = r10_bio->mddev;
211 conf_t *conf = mddev->private;
212
213 spin_lock_irqsave(&conf->device_lock, flags);
214 list_add(&r10_bio->retry_list, &conf->retry_list);
215 conf->nr_queued ++;
216 spin_unlock_irqrestore(&conf->device_lock, flags);
217
218 /* wake up frozen array... */
219 wake_up(&conf->wait_barrier);
220
221 md_wakeup_thread(mddev->thread);
222}
223
224/*
225 * raid_end_bio_io() is called when we have finished servicing a mirrored
226 * operation and are ready to return a success/failure code to the buffer
227 * cache layer.
228 */
229static void raid_end_bio_io(r10bio_t *r10_bio)
230{
231 struct bio *bio = r10_bio->master_bio;
232 int done;
233 conf_t *conf = r10_bio->mddev->private;
234
235 if (bio->bi_phys_segments) {
236 unsigned long flags;
237 spin_lock_irqsave(&conf->device_lock, flags);
238 bio->bi_phys_segments--;
239 done = (bio->bi_phys_segments == 0);
240 spin_unlock_irqrestore(&conf->device_lock, flags);
241 } else
242 done = 1;
243 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
244 clear_bit(BIO_UPTODATE, &bio->bi_flags);
245 if (done) {
246 bio_endio(bio, 0);
247 /*
248 * Wake up any possible resync thread that waits for the device
249 * to go idle.
250 */
251 allow_barrier(conf);
252 }
253 free_r10bio(r10_bio);
254}
255
256/*
257 * Update disk head position estimator based on IRQ completion info.
258 */
259static inline void update_head_pos(int slot, r10bio_t *r10_bio)
260{
261 conf_t *conf = r10_bio->mddev->private;
262
263 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
264 r10_bio->devs[slot].addr + (r10_bio->sectors);
265}
266
267/*
268 * Find the disk number which triggered given bio
269 */
270static int find_bio_disk(conf_t *conf, r10bio_t *r10_bio,
271 struct bio *bio, int *slotp)
272{
273 int slot;
274
275 for (slot = 0; slot < conf->copies; slot++)
276 if (r10_bio->devs[slot].bio == bio)
277 break;
278
279 BUG_ON(slot == conf->copies);
280 update_head_pos(slot, r10_bio);
281
282 if (slotp)
283 *slotp = slot;
284 return r10_bio->devs[slot].devnum;
285}
286
287static void raid10_end_read_request(struct bio *bio, int error)
288{
289 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
290 r10bio_t *r10_bio = bio->bi_private;
291 int slot, dev;
292 conf_t *conf = r10_bio->mddev->private;
293
294
295 slot = r10_bio->read_slot;
296 dev = r10_bio->devs[slot].devnum;
297 /*
298 * this branch is our 'one mirror IO has finished' event handler:
299 */
300 update_head_pos(slot, r10_bio);
301
302 if (uptodate) {
303 /*
304 * Set R10BIO_Uptodate in our master bio, so that
305 * we will return a good error code to the higher
306 * levels even if IO on some other mirrored buffer fails.
307 *
308 * The 'master' represents the composite IO operation to
309 * user-side. So if something waits for IO, then it will
310 * wait for the 'master' bio.
311 */
312 set_bit(R10BIO_Uptodate, &r10_bio->state);
313 raid_end_bio_io(r10_bio);
314 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
315 } else {
316 /*
317 * oops, read error - keep the refcount on the rdev
318 */
319 char b[BDEVNAME_SIZE];
320 printk_ratelimited(KERN_ERR
321 "md/raid10:%s: %s: rescheduling sector %llu\n",
322 mdname(conf->mddev),
323 bdevname(conf->mirrors[dev].rdev->bdev, b),
324 (unsigned long long)r10_bio->sector);
325 set_bit(R10BIO_ReadError, &r10_bio->state);
326 reschedule_retry(r10_bio);
327 }
328}
329
330static void close_write(r10bio_t *r10_bio)
331{
332 /* clear the bitmap if all writes complete successfully */
333 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
334 r10_bio->sectors,
335 !test_bit(R10BIO_Degraded, &r10_bio->state),
336 0);
337 md_write_end(r10_bio->mddev);
338}
339
340static void one_write_done(r10bio_t *r10_bio)
341{
342 if (atomic_dec_and_test(&r10_bio->remaining)) {
343 if (test_bit(R10BIO_WriteError, &r10_bio->state))
344 reschedule_retry(r10_bio);
345 else {
346 close_write(r10_bio);
347 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
348 reschedule_retry(r10_bio);
349 else
350 raid_end_bio_io(r10_bio);
351 }
352 }
353}
354
355static void raid10_end_write_request(struct bio *bio, int error)
356{
357 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
358 r10bio_t *r10_bio = bio->bi_private;
359 int dev;
360 int dec_rdev = 1;
361 conf_t *conf = r10_bio->mddev->private;
362 int slot;
363
364 dev = find_bio_disk(conf, r10_bio, bio, &slot);
365
366 /*
367 * this branch is our 'one mirror IO has finished' event handler:
368 */
369 if (!uptodate) {
370 set_bit(WriteErrorSeen, &conf->mirrors[dev].rdev->flags);
371 set_bit(R10BIO_WriteError, &r10_bio->state);
372 dec_rdev = 0;
373 } else {
374 /*
375 * Set R10BIO_Uptodate in our master bio, so that
376 * we will return a good error code for to the higher
377 * levels even if IO on some other mirrored buffer fails.
378 *
379 * The 'master' represents the composite IO operation to
380 * user-side. So if something waits for IO, then it will
381 * wait for the 'master' bio.
382 */
383 sector_t first_bad;
384 int bad_sectors;
385
386 set_bit(R10BIO_Uptodate, &r10_bio->state);
387
388 /* Maybe we can clear some bad blocks. */
389 if (is_badblock(conf->mirrors[dev].rdev,
390 r10_bio->devs[slot].addr,
391 r10_bio->sectors,
392 &first_bad, &bad_sectors)) {
393 bio_put(bio);
394 r10_bio->devs[slot].bio = IO_MADE_GOOD;
395 dec_rdev = 0;
396 set_bit(R10BIO_MadeGood, &r10_bio->state);
397 }
398 }
399
400 /*
401 *
402 * Let's see if all mirrored write operations have finished
403 * already.
404 */
405 one_write_done(r10_bio);
406 if (dec_rdev)
407 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
408}
409
410
411/*
412 * RAID10 layout manager
413 * As well as the chunksize and raid_disks count, there are two
414 * parameters: near_copies and far_copies.
415 * near_copies * far_copies must be <= raid_disks.
416 * Normally one of these will be 1.
417 * If both are 1, we get raid0.
418 * If near_copies == raid_disks, we get raid1.
419 *
420 * Chunks are laid out in raid0 style with near_copies copies of the
421 * first chunk, followed by near_copies copies of the next chunk and
422 * so on.
423 * If far_copies > 1, then after 1/far_copies of the array has been assigned
424 * as described above, we start again with a device offset of near_copies.
425 * So we effectively have another copy of the whole array further down all
426 * the drives, but with blocks on different drives.
427 * With this layout, and block is never stored twice on the one device.
428 *
429 * raid10_find_phys finds the sector offset of a given virtual sector
430 * on each device that it is on.
431 *
432 * raid10_find_virt does the reverse mapping, from a device and a
433 * sector offset to a virtual address
434 */
435
436static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
437{
438 int n,f;
439 sector_t sector;
440 sector_t chunk;
441 sector_t stripe;
442 int dev;
443
444 int slot = 0;
445
446 /* now calculate first sector/dev */
447 chunk = r10bio->sector >> conf->chunk_shift;
448 sector = r10bio->sector & conf->chunk_mask;
449
450 chunk *= conf->near_copies;
451 stripe = chunk;
452 dev = sector_div(stripe, conf->raid_disks);
453 if (conf->far_offset)
454 stripe *= conf->far_copies;
455
456 sector += stripe << conf->chunk_shift;
457
458 /* and calculate all the others */
459 for (n=0; n < conf->near_copies; n++) {
460 int d = dev;
461 sector_t s = sector;
462 r10bio->devs[slot].addr = sector;
463 r10bio->devs[slot].devnum = d;
464 slot++;
465
466 for (f = 1; f < conf->far_copies; f++) {
467 d += conf->near_copies;
468 if (d >= conf->raid_disks)
469 d -= conf->raid_disks;
470 s += conf->stride;
471 r10bio->devs[slot].devnum = d;
472 r10bio->devs[slot].addr = s;
473 slot++;
474 }
475 dev++;
476 if (dev >= conf->raid_disks) {
477 dev = 0;
478 sector += (conf->chunk_mask + 1);
479 }
480 }
481 BUG_ON(slot != conf->copies);
482}
483
484static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
485{
486 sector_t offset, chunk, vchunk;
487
488 offset = sector & conf->chunk_mask;
489 if (conf->far_offset) {
490 int fc;
491 chunk = sector >> conf->chunk_shift;
492 fc = sector_div(chunk, conf->far_copies);
493 dev -= fc * conf->near_copies;
494 if (dev < 0)
495 dev += conf->raid_disks;
496 } else {
497 while (sector >= conf->stride) {
498 sector -= conf->stride;
499 if (dev < conf->near_copies)
500 dev += conf->raid_disks - conf->near_copies;
501 else
502 dev -= conf->near_copies;
503 }
504 chunk = sector >> conf->chunk_shift;
505 }
506 vchunk = chunk * conf->raid_disks + dev;
507 sector_div(vchunk, conf->near_copies);
508 return (vchunk << conf->chunk_shift) + offset;
509}
510
511/**
512 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
513 * @q: request queue
514 * @bvm: properties of new bio
515 * @biovec: the request that could be merged to it.
516 *
517 * Return amount of bytes we can accept at this offset
518 * If near_copies == raid_disk, there are no striping issues,
519 * but in that case, the function isn't called at all.
520 */
521static int raid10_mergeable_bvec(struct request_queue *q,
522 struct bvec_merge_data *bvm,
523 struct bio_vec *biovec)
524{
525 mddev_t *mddev = q->queuedata;
526 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
527 int max;
528 unsigned int chunk_sectors = mddev->chunk_sectors;
529 unsigned int bio_sectors = bvm->bi_size >> 9;
530
531 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
532 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
533 if (max <= biovec->bv_len && bio_sectors == 0)
534 return biovec->bv_len;
535 else
536 return max;
537}
538
539/*
540 * This routine returns the disk from which the requested read should
541 * be done. There is a per-array 'next expected sequential IO' sector
542 * number - if this matches on the next IO then we use the last disk.
543 * There is also a per-disk 'last know head position' sector that is
544 * maintained from IRQ contexts, both the normal and the resync IO
545 * completion handlers update this position correctly. If there is no
546 * perfect sequential match then we pick the disk whose head is closest.
547 *
548 * If there are 2 mirrors in the same 2 devices, performance degrades
549 * because position is mirror, not device based.
550 *
551 * The rdev for the device selected will have nr_pending incremented.
552 */
553
554/*
555 * FIXME: possibly should rethink readbalancing and do it differently
556 * depending on near_copies / far_copies geometry.
557 */
558static int read_balance(conf_t *conf, r10bio_t *r10_bio, int *max_sectors)
559{
560 const sector_t this_sector = r10_bio->sector;
561 int disk, slot;
562 int sectors = r10_bio->sectors;
563 int best_good_sectors;
564 sector_t new_distance, best_dist;
565 mdk_rdev_t *rdev;
566 int do_balance;
567 int best_slot;
568
569 raid10_find_phys(conf, r10_bio);
570 rcu_read_lock();
571retry:
572 sectors = r10_bio->sectors;
573 best_slot = -1;
574 best_dist = MaxSector;
575 best_good_sectors = 0;
576 do_balance = 1;
577 /*
578 * Check if we can balance. We can balance on the whole
579 * device if no resync is going on (recovery is ok), or below
580 * the resync window. We take the first readable disk when
581 * above the resync window.
582 */
583 if (conf->mddev->recovery_cp < MaxSector
584 && (this_sector + sectors >= conf->next_resync))
585 do_balance = 0;
586
587 for (slot = 0; slot < conf->copies ; slot++) {
588 sector_t first_bad;
589 int bad_sectors;
590 sector_t dev_sector;
591
592 if (r10_bio->devs[slot].bio == IO_BLOCKED)
593 continue;
594 disk = r10_bio->devs[slot].devnum;
595 rdev = rcu_dereference(conf->mirrors[disk].rdev);
596 if (rdev == NULL)
597 continue;
598 if (!test_bit(In_sync, &rdev->flags))
599 continue;
600
601 dev_sector = r10_bio->devs[slot].addr;
602 if (is_badblock(rdev, dev_sector, sectors,
603 &first_bad, &bad_sectors)) {
604 if (best_dist < MaxSector)
605 /* Already have a better slot */
606 continue;
607 if (first_bad <= dev_sector) {
608 /* Cannot read here. If this is the
609 * 'primary' device, then we must not read
610 * beyond 'bad_sectors' from another device.
611 */
612 bad_sectors -= (dev_sector - first_bad);
613 if (!do_balance && sectors > bad_sectors)
614 sectors = bad_sectors;
615 if (best_good_sectors > sectors)
616 best_good_sectors = sectors;
617 } else {
618 sector_t good_sectors =
619 first_bad - dev_sector;
620 if (good_sectors > best_good_sectors) {
621 best_good_sectors = good_sectors;
622 best_slot = slot;
623 }
624 if (!do_balance)
625 /* Must read from here */
626 break;
627 }
628 continue;
629 } else
630 best_good_sectors = sectors;
631
632 if (!do_balance)
633 break;
634
635 /* This optimisation is debatable, and completely destroys
636 * sequential read speed for 'far copies' arrays. So only
637 * keep it for 'near' arrays, and review those later.
638 */
639 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
640 break;
641
642 /* for far > 1 always use the lowest address */
643 if (conf->far_copies > 1)
644 new_distance = r10_bio->devs[slot].addr;
645 else
646 new_distance = abs(r10_bio->devs[slot].addr -
647 conf->mirrors[disk].head_position);
648 if (new_distance < best_dist) {
649 best_dist = new_distance;
650 best_slot = slot;
651 }
652 }
653 if (slot == conf->copies)
654 slot = best_slot;
655
656 if (slot >= 0) {
657 disk = r10_bio->devs[slot].devnum;
658 rdev = rcu_dereference(conf->mirrors[disk].rdev);
659 if (!rdev)
660 goto retry;
661 atomic_inc(&rdev->nr_pending);
662 if (test_bit(Faulty, &rdev->flags)) {
663 /* Cannot risk returning a device that failed
664 * before we inc'ed nr_pending
665 */
666 rdev_dec_pending(rdev, conf->mddev);
667 goto retry;
668 }
669 r10_bio->read_slot = slot;
670 } else
671 disk = -1;
672 rcu_read_unlock();
673 *max_sectors = best_good_sectors;
674
675 return disk;
676}
677
678static int raid10_congested(void *data, int bits)
679{
680 mddev_t *mddev = data;
681 conf_t *conf = mddev->private;
682 int i, ret = 0;
683
684 if (mddev_congested(mddev, bits))
685 return 1;
686 rcu_read_lock();
687 for (i = 0; i < conf->raid_disks && ret == 0; i++) {
688 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
689 if (rdev && !test_bit(Faulty, &rdev->flags)) {
690 struct request_queue *q = bdev_get_queue(rdev->bdev);
691
692 ret |= bdi_congested(&q->backing_dev_info, bits);
693 }
694 }
695 rcu_read_unlock();
696 return ret;
697}
698
699static void flush_pending_writes(conf_t *conf)
700{
701 /* Any writes that have been queued but are awaiting
702 * bitmap updates get flushed here.
703 */
704 spin_lock_irq(&conf->device_lock);
705
706 if (conf->pending_bio_list.head) {
707 struct bio *bio;
708 bio = bio_list_get(&conf->pending_bio_list);
709 spin_unlock_irq(&conf->device_lock);
710 /* flush any pending bitmap writes to disk
711 * before proceeding w/ I/O */
712 bitmap_unplug(conf->mddev->bitmap);
713
714 while (bio) { /* submit pending writes */
715 struct bio *next = bio->bi_next;
716 bio->bi_next = NULL;
717 generic_make_request(bio);
718 bio = next;
719 }
720 } else
721 spin_unlock_irq(&conf->device_lock);
722}
723
724/* Barriers....
725 * Sometimes we need to suspend IO while we do something else,
726 * either some resync/recovery, or reconfigure the array.
727 * To do this we raise a 'barrier'.
728 * The 'barrier' is a counter that can be raised multiple times
729 * to count how many activities are happening which preclude
730 * normal IO.
731 * We can only raise the barrier if there is no pending IO.
732 * i.e. if nr_pending == 0.
733 * We choose only to raise the barrier if no-one is waiting for the
734 * barrier to go down. This means that as soon as an IO request
735 * is ready, no other operations which require a barrier will start
736 * until the IO request has had a chance.
737 *
738 * So: regular IO calls 'wait_barrier'. When that returns there
739 * is no backgroup IO happening, It must arrange to call
740 * allow_barrier when it has finished its IO.
741 * backgroup IO calls must call raise_barrier. Once that returns
742 * there is no normal IO happeing. It must arrange to call
743 * lower_barrier when the particular background IO completes.
744 */
745
746static void raise_barrier(conf_t *conf, int force)
747{
748 BUG_ON(force && !conf->barrier);
749 spin_lock_irq(&conf->resync_lock);
750
751 /* Wait until no block IO is waiting (unless 'force') */
752 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
753 conf->resync_lock, );
754
755 /* block any new IO from starting */
756 conf->barrier++;
757
758 /* Now wait for all pending IO to complete */
759 wait_event_lock_irq(conf->wait_barrier,
760 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
761 conf->resync_lock, );
762
763 spin_unlock_irq(&conf->resync_lock);
764}
765
766static void lower_barrier(conf_t *conf)
767{
768 unsigned long flags;
769 spin_lock_irqsave(&conf->resync_lock, flags);
770 conf->barrier--;
771 spin_unlock_irqrestore(&conf->resync_lock, flags);
772 wake_up(&conf->wait_barrier);
773}
774
775static void wait_barrier(conf_t *conf)
776{
777 spin_lock_irq(&conf->resync_lock);
778 if (conf->barrier) {
779 conf->nr_waiting++;
780 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
781 conf->resync_lock,
782 );
783 conf->nr_waiting--;
784 }
785 conf->nr_pending++;
786 spin_unlock_irq(&conf->resync_lock);
787}
788
789static void allow_barrier(conf_t *conf)
790{
791 unsigned long flags;
792 spin_lock_irqsave(&conf->resync_lock, flags);
793 conf->nr_pending--;
794 spin_unlock_irqrestore(&conf->resync_lock, flags);
795 wake_up(&conf->wait_barrier);
796}
797
798static void freeze_array(conf_t *conf)
799{
800 /* stop syncio and normal IO and wait for everything to
801 * go quiet.
802 * We increment barrier and nr_waiting, and then
803 * wait until nr_pending match nr_queued+1
804 * This is called in the context of one normal IO request
805 * that has failed. Thus any sync request that might be pending
806 * will be blocked by nr_pending, and we need to wait for
807 * pending IO requests to complete or be queued for re-try.
808 * Thus the number queued (nr_queued) plus this request (1)
809 * must match the number of pending IOs (nr_pending) before
810 * we continue.
811 */
812 spin_lock_irq(&conf->resync_lock);
813 conf->barrier++;
814 conf->nr_waiting++;
815 wait_event_lock_irq(conf->wait_barrier,
816 conf->nr_pending == conf->nr_queued+1,
817 conf->resync_lock,
818 flush_pending_writes(conf));
819
820 spin_unlock_irq(&conf->resync_lock);
821}
822
823static void unfreeze_array(conf_t *conf)
824{
825 /* reverse the effect of the freeze */
826 spin_lock_irq(&conf->resync_lock);
827 conf->barrier--;
828 conf->nr_waiting--;
829 wake_up(&conf->wait_barrier);
830 spin_unlock_irq(&conf->resync_lock);
831}
832
833static int make_request(mddev_t *mddev, struct bio * bio)
834{
835 conf_t *conf = mddev->private;
836 mirror_info_t *mirror;
837 r10bio_t *r10_bio;
838 struct bio *read_bio;
839 int i;
840 int chunk_sects = conf->chunk_mask + 1;
841 const int rw = bio_data_dir(bio);
842 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
843 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
844 unsigned long flags;
845 mdk_rdev_t *blocked_rdev;
846 int plugged;
847 int sectors_handled;
848 int max_sectors;
849
850 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
851 md_flush_request(mddev, bio);
852 return 0;
853 }
854
855 /* If this request crosses a chunk boundary, we need to
856 * split it. This will only happen for 1 PAGE (or less) requests.
857 */
858 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
859 > chunk_sects &&
860 conf->near_copies < conf->raid_disks)) {
861 struct bio_pair *bp;
862 /* Sanity check -- queue functions should prevent this happening */
863 if (bio->bi_vcnt != 1 ||
864 bio->bi_idx != 0)
865 goto bad_map;
866 /* This is a one page bio that upper layers
867 * refuse to split for us, so we need to split it.
868 */
869 bp = bio_split(bio,
870 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
871
872 /* Each of these 'make_request' calls will call 'wait_barrier'.
873 * If the first succeeds but the second blocks due to the resync
874 * thread raising the barrier, we will deadlock because the
875 * IO to the underlying device will be queued in generic_make_request
876 * and will never complete, so will never reduce nr_pending.
877 * So increment nr_waiting here so no new raise_barriers will
878 * succeed, and so the second wait_barrier cannot block.
879 */
880 spin_lock_irq(&conf->resync_lock);
881 conf->nr_waiting++;
882 spin_unlock_irq(&conf->resync_lock);
883
884 if (make_request(mddev, &bp->bio1))
885 generic_make_request(&bp->bio1);
886 if (make_request(mddev, &bp->bio2))
887 generic_make_request(&bp->bio2);
888
889 spin_lock_irq(&conf->resync_lock);
890 conf->nr_waiting--;
891 wake_up(&conf->wait_barrier);
892 spin_unlock_irq(&conf->resync_lock);
893
894 bio_pair_release(bp);
895 return 0;
896 bad_map:
897 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
898 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
899 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
900
901 bio_io_error(bio);
902 return 0;
903 }
904
905 md_write_start(mddev, bio);
906
907 /*
908 * Register the new request and wait if the reconstruction
909 * thread has put up a bar for new requests.
910 * Continue immediately if no resync is active currently.
911 */
912 wait_barrier(conf);
913
914 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
915
916 r10_bio->master_bio = bio;
917 r10_bio->sectors = bio->bi_size >> 9;
918
919 r10_bio->mddev = mddev;
920 r10_bio->sector = bio->bi_sector;
921 r10_bio->state = 0;
922
923 /* We might need to issue multiple reads to different
924 * devices if there are bad blocks around, so we keep
925 * track of the number of reads in bio->bi_phys_segments.
926 * If this is 0, there is only one r10_bio and no locking
927 * will be needed when the request completes. If it is
928 * non-zero, then it is the number of not-completed requests.
929 */
930 bio->bi_phys_segments = 0;
931 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
932
933 if (rw == READ) {
934 /*
935 * read balancing logic:
936 */
937 int disk;
938 int slot;
939
940read_again:
941 disk = read_balance(conf, r10_bio, &max_sectors);
942 slot = r10_bio->read_slot;
943 if (disk < 0) {
944 raid_end_bio_io(r10_bio);
945 return 0;
946 }
947 mirror = conf->mirrors + disk;
948
949 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
950 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
951 max_sectors);
952
953 r10_bio->devs[slot].bio = read_bio;
954
955 read_bio->bi_sector = r10_bio->devs[slot].addr +
956 mirror->rdev->data_offset;
957 read_bio->bi_bdev = mirror->rdev->bdev;
958 read_bio->bi_end_io = raid10_end_read_request;
959 read_bio->bi_rw = READ | do_sync;
960 read_bio->bi_private = r10_bio;
961
962 if (max_sectors < r10_bio->sectors) {
963 /* Could not read all from this device, so we will
964 * need another r10_bio.
965 */
966 sectors_handled = (r10_bio->sectors + max_sectors
967 - bio->bi_sector);
968 r10_bio->sectors = max_sectors;
969 spin_lock_irq(&conf->device_lock);
970 if (bio->bi_phys_segments == 0)
971 bio->bi_phys_segments = 2;
972 else
973 bio->bi_phys_segments++;
974 spin_unlock(&conf->device_lock);
975 /* Cannot call generic_make_request directly
976 * as that will be queued in __generic_make_request
977 * and subsequent mempool_alloc might block
978 * waiting for it. so hand bio over to raid10d.
979 */
980 reschedule_retry(r10_bio);
981
982 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
983
984 r10_bio->master_bio = bio;
985 r10_bio->sectors = ((bio->bi_size >> 9)
986 - sectors_handled);
987 r10_bio->state = 0;
988 r10_bio->mddev = mddev;
989 r10_bio->sector = bio->bi_sector + sectors_handled;
990 goto read_again;
991 } else
992 generic_make_request(read_bio);
993 return 0;
994 }
995
996 /*
997 * WRITE:
998 */
999 /* first select target devices under rcu_lock and
1000 * inc refcount on their rdev. Record them by setting
1001 * bios[x] to bio
1002 * If there are known/acknowledged bad blocks on any device
1003 * on which we have seen a write error, we want to avoid
1004 * writing to those blocks. This potentially requires several
1005 * writes to write around the bad blocks. Each set of writes
1006 * gets its own r10_bio with a set of bios attached. The number
1007 * of r10_bios is recored in bio->bi_phys_segments just as with
1008 * the read case.
1009 */
1010 plugged = mddev_check_plugged(mddev);
1011
1012 raid10_find_phys(conf, r10_bio);
1013retry_write:
1014 blocked_rdev = NULL;
1015 rcu_read_lock();
1016 max_sectors = r10_bio->sectors;
1017
1018 for (i = 0; i < conf->copies; i++) {
1019 int d = r10_bio->devs[i].devnum;
1020 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
1021 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1022 atomic_inc(&rdev->nr_pending);
1023 blocked_rdev = rdev;
1024 break;
1025 }
1026 r10_bio->devs[i].bio = NULL;
1027 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1028 set_bit(R10BIO_Degraded, &r10_bio->state);
1029 continue;
1030 }
1031 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1032 sector_t first_bad;
1033 sector_t dev_sector = r10_bio->devs[i].addr;
1034 int bad_sectors;
1035 int is_bad;
1036
1037 is_bad = is_badblock(rdev, dev_sector,
1038 max_sectors,
1039 &first_bad, &bad_sectors);
1040 if (is_bad < 0) {
1041 /* Mustn't write here until the bad block
1042 * is acknowledged
1043 */
1044 atomic_inc(&rdev->nr_pending);
1045 set_bit(BlockedBadBlocks, &rdev->flags);
1046 blocked_rdev = rdev;
1047 break;
1048 }
1049 if (is_bad && first_bad <= dev_sector) {
1050 /* Cannot write here at all */
1051 bad_sectors -= (dev_sector - first_bad);
1052 if (bad_sectors < max_sectors)
1053 /* Mustn't write more than bad_sectors
1054 * to other devices yet
1055 */
1056 max_sectors = bad_sectors;
1057 /* We don't set R10BIO_Degraded as that
1058 * only applies if the disk is missing,
1059 * so it might be re-added, and we want to
1060 * know to recover this chunk.
1061 * In this case the device is here, and the
1062 * fact that this chunk is not in-sync is
1063 * recorded in the bad block log.
1064 */
1065 continue;
1066 }
1067 if (is_bad) {
1068 int good_sectors = first_bad - dev_sector;
1069 if (good_sectors < max_sectors)
1070 max_sectors = good_sectors;
1071 }
1072 }
1073 r10_bio->devs[i].bio = bio;
1074 atomic_inc(&rdev->nr_pending);
1075 }
1076 rcu_read_unlock();
1077
1078 if (unlikely(blocked_rdev)) {
1079 /* Have to wait for this device to get unblocked, then retry */
1080 int j;
1081 int d;
1082
1083 for (j = 0; j < i; j++)
1084 if (r10_bio->devs[j].bio) {
1085 d = r10_bio->devs[j].devnum;
1086 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1087 }
1088 allow_barrier(conf);
1089 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1090 wait_barrier(conf);
1091 goto retry_write;
1092 }
1093
1094 if (max_sectors < r10_bio->sectors) {
1095 /* We are splitting this into multiple parts, so
1096 * we need to prepare for allocating another r10_bio.
1097 */
1098 r10_bio->sectors = max_sectors;
1099 spin_lock_irq(&conf->device_lock);
1100 if (bio->bi_phys_segments == 0)
1101 bio->bi_phys_segments = 2;
1102 else
1103 bio->bi_phys_segments++;
1104 spin_unlock_irq(&conf->device_lock);
1105 }
1106 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1107
1108 atomic_set(&r10_bio->remaining, 1);
1109 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1110
1111 for (i = 0; i < conf->copies; i++) {
1112 struct bio *mbio;
1113 int d = r10_bio->devs[i].devnum;
1114 if (!r10_bio->devs[i].bio)
1115 continue;
1116
1117 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1118 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1119 max_sectors);
1120 r10_bio->devs[i].bio = mbio;
1121
1122 mbio->bi_sector = (r10_bio->devs[i].addr+
1123 conf->mirrors[d].rdev->data_offset);
1124 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1125 mbio->bi_end_io = raid10_end_write_request;
1126 mbio->bi_rw = WRITE | do_sync | do_fua;
1127 mbio->bi_private = r10_bio;
1128
1129 atomic_inc(&r10_bio->remaining);
1130 spin_lock_irqsave(&conf->device_lock, flags);
1131 bio_list_add(&conf->pending_bio_list, mbio);
1132 spin_unlock_irqrestore(&conf->device_lock, flags);
1133 }
1134
1135 /* Don't remove the bias on 'remaining' (one_write_done) until
1136 * after checking if we need to go around again.
1137 */
1138
1139 if (sectors_handled < (bio->bi_size >> 9)) {
1140 one_write_done(r10_bio);
1141 /* We need another r10_bio. It has already been counted
1142 * in bio->bi_phys_segments.
1143 */
1144 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1145
1146 r10_bio->master_bio = bio;
1147 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1148
1149 r10_bio->mddev = mddev;
1150 r10_bio->sector = bio->bi_sector + sectors_handled;
1151 r10_bio->state = 0;
1152 goto retry_write;
1153 }
1154 one_write_done(r10_bio);
1155
1156 /* In case raid10d snuck in to freeze_array */
1157 wake_up(&conf->wait_barrier);
1158
1159 if (do_sync || !mddev->bitmap || !plugged)
1160 md_wakeup_thread(mddev->thread);
1161 return 0;
1162}
1163
1164static void status(struct seq_file *seq, mddev_t *mddev)
1165{
1166 conf_t *conf = mddev->private;
1167 int i;
1168
1169 if (conf->near_copies < conf->raid_disks)
1170 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1171 if (conf->near_copies > 1)
1172 seq_printf(seq, " %d near-copies", conf->near_copies);
1173 if (conf->far_copies > 1) {
1174 if (conf->far_offset)
1175 seq_printf(seq, " %d offset-copies", conf->far_copies);
1176 else
1177 seq_printf(seq, " %d far-copies", conf->far_copies);
1178 }
1179 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1180 conf->raid_disks - mddev->degraded);
1181 for (i = 0; i < conf->raid_disks; i++)
1182 seq_printf(seq, "%s",
1183 conf->mirrors[i].rdev &&
1184 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1185 seq_printf(seq, "]");
1186}
1187
1188/* check if there are enough drives for
1189 * every block to appear on atleast one.
1190 * Don't consider the device numbered 'ignore'
1191 * as we might be about to remove it.
1192 */
1193static int enough(conf_t *conf, int ignore)
1194{
1195 int first = 0;
1196
1197 do {
1198 int n = conf->copies;
1199 int cnt = 0;
1200 while (n--) {
1201 if (conf->mirrors[first].rdev &&
1202 first != ignore)
1203 cnt++;
1204 first = (first+1) % conf->raid_disks;
1205 }
1206 if (cnt == 0)
1207 return 0;
1208 } while (first != 0);
1209 return 1;
1210}
1211
1212static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1213{
1214 char b[BDEVNAME_SIZE];
1215 conf_t *conf = mddev->private;
1216
1217 /*
1218 * If it is not operational, then we have already marked it as dead
1219 * else if it is the last working disks, ignore the error, let the
1220 * next level up know.
1221 * else mark the drive as failed
1222 */
1223 if (test_bit(In_sync, &rdev->flags)
1224 && !enough(conf, rdev->raid_disk))
1225 /*
1226 * Don't fail the drive, just return an IO error.
1227 */
1228 return;
1229 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1230 unsigned long flags;
1231 spin_lock_irqsave(&conf->device_lock, flags);
1232 mddev->degraded++;
1233 spin_unlock_irqrestore(&conf->device_lock, flags);
1234 /*
1235 * if recovery is running, make sure it aborts.
1236 */
1237 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1238 }
1239 set_bit(Blocked, &rdev->flags);
1240 set_bit(Faulty, &rdev->flags);
1241 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1242 printk(KERN_ALERT
1243 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1244 "md/raid10:%s: Operation continuing on %d devices.\n",
1245 mdname(mddev), bdevname(rdev->bdev, b),
1246 mdname(mddev), conf->raid_disks - mddev->degraded);
1247}
1248
1249static void print_conf(conf_t *conf)
1250{
1251 int i;
1252 mirror_info_t *tmp;
1253
1254 printk(KERN_DEBUG "RAID10 conf printout:\n");
1255 if (!conf) {
1256 printk(KERN_DEBUG "(!conf)\n");
1257 return;
1258 }
1259 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1260 conf->raid_disks);
1261
1262 for (i = 0; i < conf->raid_disks; i++) {
1263 char b[BDEVNAME_SIZE];
1264 tmp = conf->mirrors + i;
1265 if (tmp->rdev)
1266 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1267 i, !test_bit(In_sync, &tmp->rdev->flags),
1268 !test_bit(Faulty, &tmp->rdev->flags),
1269 bdevname(tmp->rdev->bdev,b));
1270 }
1271}
1272
1273static void close_sync(conf_t *conf)
1274{
1275 wait_barrier(conf);
1276 allow_barrier(conf);
1277
1278 mempool_destroy(conf->r10buf_pool);
1279 conf->r10buf_pool = NULL;
1280}
1281
1282static int raid10_spare_active(mddev_t *mddev)
1283{
1284 int i;
1285 conf_t *conf = mddev->private;
1286 mirror_info_t *tmp;
1287 int count = 0;
1288 unsigned long flags;
1289
1290 /*
1291 * Find all non-in_sync disks within the RAID10 configuration
1292 * and mark them in_sync
1293 */
1294 for (i = 0; i < conf->raid_disks; i++) {
1295 tmp = conf->mirrors + i;
1296 if (tmp->rdev
1297 && !test_bit(Faulty, &tmp->rdev->flags)
1298 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1299 count++;
1300 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1301 }
1302 }
1303 spin_lock_irqsave(&conf->device_lock, flags);
1304 mddev->degraded -= count;
1305 spin_unlock_irqrestore(&conf->device_lock, flags);
1306
1307 print_conf(conf);
1308 return count;
1309}
1310
1311
1312static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1313{
1314 conf_t *conf = mddev->private;
1315 int err = -EEXIST;
1316 int mirror;
1317 int first = 0;
1318 int last = conf->raid_disks - 1;
1319
1320 if (mddev->recovery_cp < MaxSector)
1321 /* only hot-add to in-sync arrays, as recovery is
1322 * very different from resync
1323 */
1324 return -EBUSY;
1325 if (!enough(conf, -1))
1326 return -EINVAL;
1327
1328 if (rdev->raid_disk >= 0)
1329 first = last = rdev->raid_disk;
1330
1331 if (rdev->saved_raid_disk >= first &&
1332 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1333 mirror = rdev->saved_raid_disk;
1334 else
1335 mirror = first;
1336 for ( ; mirror <= last ; mirror++) {
1337 mirror_info_t *p = &conf->mirrors[mirror];
1338 if (p->recovery_disabled == mddev->recovery_disabled)
1339 continue;
1340 if (!p->rdev)
1341 continue;
1342
1343 disk_stack_limits(mddev->gendisk, rdev->bdev,
1344 rdev->data_offset << 9);
1345 /* as we don't honour merge_bvec_fn, we must
1346 * never risk violating it, so limit
1347 * ->max_segments to one lying with a single
1348 * page, as a one page request is never in
1349 * violation.
1350 */
1351 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1352 blk_queue_max_segments(mddev->queue, 1);
1353 blk_queue_segment_boundary(mddev->queue,
1354 PAGE_CACHE_SIZE - 1);
1355 }
1356
1357 p->head_position = 0;
1358 rdev->raid_disk = mirror;
1359 err = 0;
1360 if (rdev->saved_raid_disk != mirror)
1361 conf->fullsync = 1;
1362 rcu_assign_pointer(p->rdev, rdev);
1363 break;
1364 }
1365
1366 md_integrity_add_rdev(rdev, mddev);
1367 print_conf(conf);
1368 return err;
1369}
1370
1371static int raid10_remove_disk(mddev_t *mddev, int number)
1372{
1373 conf_t *conf = mddev->private;
1374 int err = 0;
1375 mdk_rdev_t *rdev;
1376 mirror_info_t *p = conf->mirrors+ number;
1377
1378 print_conf(conf);
1379 rdev = p->rdev;
1380 if (rdev) {
1381 if (test_bit(In_sync, &rdev->flags) ||
1382 atomic_read(&rdev->nr_pending)) {
1383 err = -EBUSY;
1384 goto abort;
1385 }
1386 /* Only remove faulty devices in recovery
1387 * is not possible.
1388 */
1389 if (!test_bit(Faulty, &rdev->flags) &&
1390 mddev->recovery_disabled != p->recovery_disabled &&
1391 enough(conf, -1)) {
1392 err = -EBUSY;
1393 goto abort;
1394 }
1395 p->rdev = NULL;
1396 synchronize_rcu();
1397 if (atomic_read(&rdev->nr_pending)) {
1398 /* lost the race, try later */
1399 err = -EBUSY;
1400 p->rdev = rdev;
1401 goto abort;
1402 }
1403 err = md_integrity_register(mddev);
1404 }
1405abort:
1406
1407 print_conf(conf);
1408 return err;
1409}
1410
1411
1412static void end_sync_read(struct bio *bio, int error)
1413{
1414 r10bio_t *r10_bio = bio->bi_private;
1415 conf_t *conf = r10_bio->mddev->private;
1416 int d;
1417
1418 d = find_bio_disk(conf, r10_bio, bio, NULL);
1419
1420 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1421 set_bit(R10BIO_Uptodate, &r10_bio->state);
1422 else
1423 /* The write handler will notice the lack of
1424 * R10BIO_Uptodate and record any errors etc
1425 */
1426 atomic_add(r10_bio->sectors,
1427 &conf->mirrors[d].rdev->corrected_errors);
1428
1429 /* for reconstruct, we always reschedule after a read.
1430 * for resync, only after all reads
1431 */
1432 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1433 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1434 atomic_dec_and_test(&r10_bio->remaining)) {
1435 /* we have read all the blocks,
1436 * do the comparison in process context in raid10d
1437 */
1438 reschedule_retry(r10_bio);
1439 }
1440}
1441
1442static void end_sync_request(r10bio_t *r10_bio)
1443{
1444 mddev_t *mddev = r10_bio->mddev;
1445
1446 while (atomic_dec_and_test(&r10_bio->remaining)) {
1447 if (r10_bio->master_bio == NULL) {
1448 /* the primary of several recovery bios */
1449 sector_t s = r10_bio->sectors;
1450 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1451 test_bit(R10BIO_WriteError, &r10_bio->state))
1452 reschedule_retry(r10_bio);
1453 else
1454 put_buf(r10_bio);
1455 md_done_sync(mddev, s, 1);
1456 break;
1457 } else {
1458 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1459 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1460 test_bit(R10BIO_WriteError, &r10_bio->state))
1461 reschedule_retry(r10_bio);
1462 else
1463 put_buf(r10_bio);
1464 r10_bio = r10_bio2;
1465 }
1466 }
1467}
1468
1469static void end_sync_write(struct bio *bio, int error)
1470{
1471 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1472 r10bio_t *r10_bio = bio->bi_private;
1473 mddev_t *mddev = r10_bio->mddev;
1474 conf_t *conf = mddev->private;
1475 int d;
1476 sector_t first_bad;
1477 int bad_sectors;
1478 int slot;
1479
1480 d = find_bio_disk(conf, r10_bio, bio, &slot);
1481
1482 if (!uptodate) {
1483 set_bit(WriteErrorSeen, &conf->mirrors[d].rdev->flags);
1484 set_bit(R10BIO_WriteError, &r10_bio->state);
1485 } else if (is_badblock(conf->mirrors[d].rdev,
1486 r10_bio->devs[slot].addr,
1487 r10_bio->sectors,
1488 &first_bad, &bad_sectors))
1489 set_bit(R10BIO_MadeGood, &r10_bio->state);
1490
1491 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1492
1493 end_sync_request(r10_bio);
1494}
1495
1496/*
1497 * Note: sync and recover and handled very differently for raid10
1498 * This code is for resync.
1499 * For resync, we read through virtual addresses and read all blocks.
1500 * If there is any error, we schedule a write. The lowest numbered
1501 * drive is authoritative.
1502 * However requests come for physical address, so we need to map.
1503 * For every physical address there are raid_disks/copies virtual addresses,
1504 * which is always are least one, but is not necessarly an integer.
1505 * This means that a physical address can span multiple chunks, so we may
1506 * have to submit multiple io requests for a single sync request.
1507 */
1508/*
1509 * We check if all blocks are in-sync and only write to blocks that
1510 * aren't in sync
1511 */
1512static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1513{
1514 conf_t *conf = mddev->private;
1515 int i, first;
1516 struct bio *tbio, *fbio;
1517
1518 atomic_set(&r10_bio->remaining, 1);
1519
1520 /* find the first device with a block */
1521 for (i=0; i<conf->copies; i++)
1522 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1523 break;
1524
1525 if (i == conf->copies)
1526 goto done;
1527
1528 first = i;
1529 fbio = r10_bio->devs[i].bio;
1530
1531 /* now find blocks with errors */
1532 for (i=0 ; i < conf->copies ; i++) {
1533 int j, d;
1534 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1535
1536 tbio = r10_bio->devs[i].bio;
1537
1538 if (tbio->bi_end_io != end_sync_read)
1539 continue;
1540 if (i == first)
1541 continue;
1542 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1543 /* We know that the bi_io_vec layout is the same for
1544 * both 'first' and 'i', so we just compare them.
1545 * All vec entries are PAGE_SIZE;
1546 */
1547 for (j = 0; j < vcnt; j++)
1548 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1549 page_address(tbio->bi_io_vec[j].bv_page),
1550 PAGE_SIZE))
1551 break;
1552 if (j == vcnt)
1553 continue;
1554 mddev->resync_mismatches += r10_bio->sectors;
1555 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1556 /* Don't fix anything. */
1557 continue;
1558 }
1559 /* Ok, we need to write this bio, either to correct an
1560 * inconsistency or to correct an unreadable block.
1561 * First we need to fixup bv_offset, bv_len and
1562 * bi_vecs, as the read request might have corrupted these
1563 */
1564 tbio->bi_vcnt = vcnt;
1565 tbio->bi_size = r10_bio->sectors << 9;
1566 tbio->bi_idx = 0;
1567 tbio->bi_phys_segments = 0;
1568 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1569 tbio->bi_flags |= 1 << BIO_UPTODATE;
1570 tbio->bi_next = NULL;
1571 tbio->bi_rw = WRITE;
1572 tbio->bi_private = r10_bio;
1573 tbio->bi_sector = r10_bio->devs[i].addr;
1574
1575 for (j=0; j < vcnt ; j++) {
1576 tbio->bi_io_vec[j].bv_offset = 0;
1577 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1578
1579 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1580 page_address(fbio->bi_io_vec[j].bv_page),
1581 PAGE_SIZE);
1582 }
1583 tbio->bi_end_io = end_sync_write;
1584
1585 d = r10_bio->devs[i].devnum;
1586 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1587 atomic_inc(&r10_bio->remaining);
1588 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1589
1590 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1591 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1592 generic_make_request(tbio);
1593 }
1594
1595done:
1596 if (atomic_dec_and_test(&r10_bio->remaining)) {
1597 md_done_sync(mddev, r10_bio->sectors, 1);
1598 put_buf(r10_bio);
1599 }
1600}
1601
1602/*
1603 * Now for the recovery code.
1604 * Recovery happens across physical sectors.
1605 * We recover all non-is_sync drives by finding the virtual address of
1606 * each, and then choose a working drive that also has that virt address.
1607 * There is a separate r10_bio for each non-in_sync drive.
1608 * Only the first two slots are in use. The first for reading,
1609 * The second for writing.
1610 *
1611 */
1612static void fix_recovery_read_error(r10bio_t *r10_bio)
1613{
1614 /* We got a read error during recovery.
1615 * We repeat the read in smaller page-sized sections.
1616 * If a read succeeds, write it to the new device or record
1617 * a bad block if we cannot.
1618 * If a read fails, record a bad block on both old and
1619 * new devices.
1620 */
1621 mddev_t *mddev = r10_bio->mddev;
1622 conf_t *conf = mddev->private;
1623 struct bio *bio = r10_bio->devs[0].bio;
1624 sector_t sect = 0;
1625 int sectors = r10_bio->sectors;
1626 int idx = 0;
1627 int dr = r10_bio->devs[0].devnum;
1628 int dw = r10_bio->devs[1].devnum;
1629
1630 while (sectors) {
1631 int s = sectors;
1632 mdk_rdev_t *rdev;
1633 sector_t addr;
1634 int ok;
1635
1636 if (s > (PAGE_SIZE>>9))
1637 s = PAGE_SIZE >> 9;
1638
1639 rdev = conf->mirrors[dr].rdev;
1640 addr = r10_bio->devs[0].addr + sect,
1641 ok = sync_page_io(rdev,
1642 addr,
1643 s << 9,
1644 bio->bi_io_vec[idx].bv_page,
1645 READ, false);
1646 if (ok) {
1647 rdev = conf->mirrors[dw].rdev;
1648 addr = r10_bio->devs[1].addr + sect;
1649 ok = sync_page_io(rdev,
1650 addr,
1651 s << 9,
1652 bio->bi_io_vec[idx].bv_page,
1653 WRITE, false);
1654 if (!ok)
1655 set_bit(WriteErrorSeen, &rdev->flags);
1656 }
1657 if (!ok) {
1658 /* We don't worry if we cannot set a bad block -
1659 * it really is bad so there is no loss in not
1660 * recording it yet
1661 */
1662 rdev_set_badblocks(rdev, addr, s, 0);
1663
1664 if (rdev != conf->mirrors[dw].rdev) {
1665 /* need bad block on destination too */
1666 mdk_rdev_t *rdev2 = conf->mirrors[dw].rdev;
1667 addr = r10_bio->devs[1].addr + sect;
1668 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1669 if (!ok) {
1670 /* just abort the recovery */
1671 printk(KERN_NOTICE
1672 "md/raid10:%s: recovery aborted"
1673 " due to read error\n",
1674 mdname(mddev));
1675
1676 conf->mirrors[dw].recovery_disabled
1677 = mddev->recovery_disabled;
1678 set_bit(MD_RECOVERY_INTR,
1679 &mddev->recovery);
1680 break;
1681 }
1682 }
1683 }
1684
1685 sectors -= s;
1686 sect += s;
1687 idx++;
1688 }
1689}
1690
1691static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1692{
1693 conf_t *conf = mddev->private;
1694 int d;
1695 struct bio *wbio;
1696
1697 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1698 fix_recovery_read_error(r10_bio);
1699 end_sync_request(r10_bio);
1700 return;
1701 }
1702
1703 /*
1704 * share the pages with the first bio
1705 * and submit the write request
1706 */
1707 wbio = r10_bio->devs[1].bio;
1708 d = r10_bio->devs[1].devnum;
1709
1710 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1711 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1712 generic_make_request(wbio);
1713}
1714
1715
1716/*
1717 * Used by fix_read_error() to decay the per rdev read_errors.
1718 * We halve the read error count for every hour that has elapsed
1719 * since the last recorded read error.
1720 *
1721 */
1722static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1723{
1724 struct timespec cur_time_mon;
1725 unsigned long hours_since_last;
1726 unsigned int read_errors = atomic_read(&rdev->read_errors);
1727
1728 ktime_get_ts(&cur_time_mon);
1729
1730 if (rdev->last_read_error.tv_sec == 0 &&
1731 rdev->last_read_error.tv_nsec == 0) {
1732 /* first time we've seen a read error */
1733 rdev->last_read_error = cur_time_mon;
1734 return;
1735 }
1736
1737 hours_since_last = (cur_time_mon.tv_sec -
1738 rdev->last_read_error.tv_sec) / 3600;
1739
1740 rdev->last_read_error = cur_time_mon;
1741
1742 /*
1743 * if hours_since_last is > the number of bits in read_errors
1744 * just set read errors to 0. We do this to avoid
1745 * overflowing the shift of read_errors by hours_since_last.
1746 */
1747 if (hours_since_last >= 8 * sizeof(read_errors))
1748 atomic_set(&rdev->read_errors, 0);
1749 else
1750 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1751}
1752
1753static int r10_sync_page_io(mdk_rdev_t *rdev, sector_t sector,
1754 int sectors, struct page *page, int rw)
1755{
1756 sector_t first_bad;
1757 int bad_sectors;
1758
1759 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1760 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1761 return -1;
1762 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1763 /* success */
1764 return 1;
1765 if (rw == WRITE)
1766 set_bit(WriteErrorSeen, &rdev->flags);
1767 /* need to record an error - either for the block or the device */
1768 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1769 md_error(rdev->mddev, rdev);
1770 return 0;
1771}
1772
1773/*
1774 * This is a kernel thread which:
1775 *
1776 * 1. Retries failed read operations on working mirrors.
1777 * 2. Updates the raid superblock when problems encounter.
1778 * 3. Performs writes following reads for array synchronising.
1779 */
1780
1781static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1782{
1783 int sect = 0; /* Offset from r10_bio->sector */
1784 int sectors = r10_bio->sectors;
1785 mdk_rdev_t*rdev;
1786 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1787 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1788
1789 /* still own a reference to this rdev, so it cannot
1790 * have been cleared recently.
1791 */
1792 rdev = conf->mirrors[d].rdev;
1793
1794 if (test_bit(Faulty, &rdev->flags))
1795 /* drive has already been failed, just ignore any
1796 more fix_read_error() attempts */
1797 return;
1798
1799 check_decay_read_errors(mddev, rdev);
1800 atomic_inc(&rdev->read_errors);
1801 if (atomic_read(&rdev->read_errors) > max_read_errors) {
1802 char b[BDEVNAME_SIZE];
1803 bdevname(rdev->bdev, b);
1804
1805 printk(KERN_NOTICE
1806 "md/raid10:%s: %s: Raid device exceeded "
1807 "read_error threshold [cur %d:max %d]\n",
1808 mdname(mddev), b,
1809 atomic_read(&rdev->read_errors), max_read_errors);
1810 printk(KERN_NOTICE
1811 "md/raid10:%s: %s: Failing raid device\n",
1812 mdname(mddev), b);
1813 md_error(mddev, conf->mirrors[d].rdev);
1814 return;
1815 }
1816
1817 while(sectors) {
1818 int s = sectors;
1819 int sl = r10_bio->read_slot;
1820 int success = 0;
1821 int start;
1822
1823 if (s > (PAGE_SIZE>>9))
1824 s = PAGE_SIZE >> 9;
1825
1826 rcu_read_lock();
1827 do {
1828 sector_t first_bad;
1829 int bad_sectors;
1830
1831 d = r10_bio->devs[sl].devnum;
1832 rdev = rcu_dereference(conf->mirrors[d].rdev);
1833 if (rdev &&
1834 test_bit(In_sync, &rdev->flags) &&
1835 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
1836 &first_bad, &bad_sectors) == 0) {
1837 atomic_inc(&rdev->nr_pending);
1838 rcu_read_unlock();
1839 success = sync_page_io(rdev,
1840 r10_bio->devs[sl].addr +
1841 sect,
1842 s<<9,
1843 conf->tmppage, READ, false);
1844 rdev_dec_pending(rdev, mddev);
1845 rcu_read_lock();
1846 if (success)
1847 break;
1848 }
1849 sl++;
1850 if (sl == conf->copies)
1851 sl = 0;
1852 } while (!success && sl != r10_bio->read_slot);
1853 rcu_read_unlock();
1854
1855 if (!success) {
1856 /* Cannot read from anywhere, just mark the block
1857 * as bad on the first device to discourage future
1858 * reads.
1859 */
1860 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1861 rdev = conf->mirrors[dn].rdev;
1862
1863 if (!rdev_set_badblocks(
1864 rdev,
1865 r10_bio->devs[r10_bio->read_slot].addr
1866 + sect,
1867 s, 0))
1868 md_error(mddev, rdev);
1869 break;
1870 }
1871
1872 start = sl;
1873 /* write it back and re-read */
1874 rcu_read_lock();
1875 while (sl != r10_bio->read_slot) {
1876 char b[BDEVNAME_SIZE];
1877
1878 if (sl==0)
1879 sl = conf->copies;
1880 sl--;
1881 d = r10_bio->devs[sl].devnum;
1882 rdev = rcu_dereference(conf->mirrors[d].rdev);
1883 if (!rdev ||
1884 !test_bit(In_sync, &rdev->flags))
1885 continue;
1886
1887 atomic_inc(&rdev->nr_pending);
1888 rcu_read_unlock();
1889 if (r10_sync_page_io(rdev,
1890 r10_bio->devs[sl].addr +
1891 sect,
1892 s<<9, conf->tmppage, WRITE)
1893 == 0) {
1894 /* Well, this device is dead */
1895 printk(KERN_NOTICE
1896 "md/raid10:%s: read correction "
1897 "write failed"
1898 " (%d sectors at %llu on %s)\n",
1899 mdname(mddev), s,
1900 (unsigned long long)(
1901 sect + rdev->data_offset),
1902 bdevname(rdev->bdev, b));
1903 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1904 "drive\n",
1905 mdname(mddev),
1906 bdevname(rdev->bdev, b));
1907 }
1908 rdev_dec_pending(rdev, mddev);
1909 rcu_read_lock();
1910 }
1911 sl = start;
1912 while (sl != r10_bio->read_slot) {
1913 char b[BDEVNAME_SIZE];
1914
1915 if (sl==0)
1916 sl = conf->copies;
1917 sl--;
1918 d = r10_bio->devs[sl].devnum;
1919 rdev = rcu_dereference(conf->mirrors[d].rdev);
1920 if (!rdev ||
1921 !test_bit(In_sync, &rdev->flags))
1922 continue;
1923
1924 atomic_inc(&rdev->nr_pending);
1925 rcu_read_unlock();
1926 switch (r10_sync_page_io(rdev,
1927 r10_bio->devs[sl].addr +
1928 sect,
1929 s<<9, conf->tmppage,
1930 READ)) {
1931 case 0:
1932 /* Well, this device is dead */
1933 printk(KERN_NOTICE
1934 "md/raid10:%s: unable to read back "
1935 "corrected sectors"
1936 " (%d sectors at %llu on %s)\n",
1937 mdname(mddev), s,
1938 (unsigned long long)(
1939 sect + rdev->data_offset),
1940 bdevname(rdev->bdev, b));
1941 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1942 "drive\n",
1943 mdname(mddev),
1944 bdevname(rdev->bdev, b));
1945 break;
1946 case 1:
1947 printk(KERN_INFO
1948 "md/raid10:%s: read error corrected"
1949 " (%d sectors at %llu on %s)\n",
1950 mdname(mddev), s,
1951 (unsigned long long)(
1952 sect + rdev->data_offset),
1953 bdevname(rdev->bdev, b));
1954 atomic_add(s, &rdev->corrected_errors);
1955 }
1956
1957 rdev_dec_pending(rdev, mddev);
1958 rcu_read_lock();
1959 }
1960 rcu_read_unlock();
1961
1962 sectors -= s;
1963 sect += s;
1964 }
1965}
1966
1967static void bi_complete(struct bio *bio, int error)
1968{
1969 complete((struct completion *)bio->bi_private);
1970}
1971
1972static int submit_bio_wait(int rw, struct bio *bio)
1973{
1974 struct completion event;
1975 rw |= REQ_SYNC;
1976
1977 init_completion(&event);
1978 bio->bi_private = &event;
1979 bio->bi_end_io = bi_complete;
1980 submit_bio(rw, bio);
1981 wait_for_completion(&event);
1982
1983 return test_bit(BIO_UPTODATE, &bio->bi_flags);
1984}
1985
1986static int narrow_write_error(r10bio_t *r10_bio, int i)
1987{
1988 struct bio *bio = r10_bio->master_bio;
1989 mddev_t *mddev = r10_bio->mddev;
1990 conf_t *conf = mddev->private;
1991 mdk_rdev_t *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
1992 /* bio has the data to be written to slot 'i' where
1993 * we just recently had a write error.
1994 * We repeatedly clone the bio and trim down to one block,
1995 * then try the write. Where the write fails we record
1996 * a bad block.
1997 * It is conceivable that the bio doesn't exactly align with
1998 * blocks. We must handle this.
1999 *
2000 * We currently own a reference to the rdev.
2001 */
2002
2003 int block_sectors;
2004 sector_t sector;
2005 int sectors;
2006 int sect_to_write = r10_bio->sectors;
2007 int ok = 1;
2008
2009 if (rdev->badblocks.shift < 0)
2010 return 0;
2011
2012 block_sectors = 1 << rdev->badblocks.shift;
2013 sector = r10_bio->sector;
2014 sectors = ((r10_bio->sector + block_sectors)
2015 & ~(sector_t)(block_sectors - 1))
2016 - sector;
2017
2018 while (sect_to_write) {
2019 struct bio *wbio;
2020 if (sectors > sect_to_write)
2021 sectors = sect_to_write;
2022 /* Write at 'sector' for 'sectors' */
2023 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2024 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2025 wbio->bi_sector = (r10_bio->devs[i].addr+
2026 rdev->data_offset+
2027 (sector - r10_bio->sector));
2028 wbio->bi_bdev = rdev->bdev;
2029 if (submit_bio_wait(WRITE, wbio) == 0)
2030 /* Failure! */
2031 ok = rdev_set_badblocks(rdev, sector,
2032 sectors, 0)
2033 && ok;
2034
2035 bio_put(wbio);
2036 sect_to_write -= sectors;
2037 sector += sectors;
2038 sectors = block_sectors;
2039 }
2040 return ok;
2041}
2042
2043static void handle_read_error(mddev_t *mddev, r10bio_t *r10_bio)
2044{
2045 int slot = r10_bio->read_slot;
2046 int mirror = r10_bio->devs[slot].devnum;
2047 struct bio *bio;
2048 conf_t *conf = mddev->private;
2049 mdk_rdev_t *rdev;
2050 char b[BDEVNAME_SIZE];
2051 unsigned long do_sync;
2052 int max_sectors;
2053
2054 /* we got a read error. Maybe the drive is bad. Maybe just
2055 * the block and we can fix it.
2056 * We freeze all other IO, and try reading the block from
2057 * other devices. When we find one, we re-write
2058 * and check it that fixes the read error.
2059 * This is all done synchronously while the array is
2060 * frozen.
2061 */
2062 if (mddev->ro == 0) {
2063 freeze_array(conf);
2064 fix_read_error(conf, mddev, r10_bio);
2065 unfreeze_array(conf);
2066 }
2067 rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
2068
2069 bio = r10_bio->devs[slot].bio;
2070 bdevname(bio->bi_bdev, b);
2071 r10_bio->devs[slot].bio =
2072 mddev->ro ? IO_BLOCKED : NULL;
2073read_more:
2074 mirror = read_balance(conf, r10_bio, &max_sectors);
2075 if (mirror == -1) {
2076 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2077 " read error for block %llu\n",
2078 mdname(mddev), b,
2079 (unsigned long long)r10_bio->sector);
2080 raid_end_bio_io(r10_bio);
2081 bio_put(bio);
2082 return;
2083 }
2084
2085 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2086 if (bio)
2087 bio_put(bio);
2088 slot = r10_bio->read_slot;
2089 rdev = conf->mirrors[mirror].rdev;
2090 printk_ratelimited(
2091 KERN_ERR
2092 "md/raid10:%s: %s: redirecting"
2093 "sector %llu to another mirror\n",
2094 mdname(mddev),
2095 bdevname(rdev->bdev, b),
2096 (unsigned long long)r10_bio->sector);
2097 bio = bio_clone_mddev(r10_bio->master_bio,
2098 GFP_NOIO, mddev);
2099 md_trim_bio(bio,
2100 r10_bio->sector - bio->bi_sector,
2101 max_sectors);
2102 r10_bio->devs[slot].bio = bio;
2103 bio->bi_sector = r10_bio->devs[slot].addr
2104 + rdev->data_offset;
2105 bio->bi_bdev = rdev->bdev;
2106 bio->bi_rw = READ | do_sync;
2107 bio->bi_private = r10_bio;
2108 bio->bi_end_io = raid10_end_read_request;
2109 if (max_sectors < r10_bio->sectors) {
2110 /* Drat - have to split this up more */
2111 struct bio *mbio = r10_bio->master_bio;
2112 int sectors_handled =
2113 r10_bio->sector + max_sectors
2114 - mbio->bi_sector;
2115 r10_bio->sectors = max_sectors;
2116 spin_lock_irq(&conf->device_lock);
2117 if (mbio->bi_phys_segments == 0)
2118 mbio->bi_phys_segments = 2;
2119 else
2120 mbio->bi_phys_segments++;
2121 spin_unlock_irq(&conf->device_lock);
2122 generic_make_request(bio);
2123 bio = NULL;
2124
2125 r10_bio = mempool_alloc(conf->r10bio_pool,
2126 GFP_NOIO);
2127 r10_bio->master_bio = mbio;
2128 r10_bio->sectors = (mbio->bi_size >> 9)
2129 - sectors_handled;
2130 r10_bio->state = 0;
2131 set_bit(R10BIO_ReadError,
2132 &r10_bio->state);
2133 r10_bio->mddev = mddev;
2134 r10_bio->sector = mbio->bi_sector
2135 + sectors_handled;
2136
2137 goto read_more;
2138 } else
2139 generic_make_request(bio);
2140}
2141
2142static void handle_write_completed(conf_t *conf, r10bio_t *r10_bio)
2143{
2144 /* Some sort of write request has finished and it
2145 * succeeded in writing where we thought there was a
2146 * bad block. So forget the bad block.
2147 * Or possibly if failed and we need to record
2148 * a bad block.
2149 */
2150 int m;
2151 mdk_rdev_t *rdev;
2152
2153 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2154 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2155 for (m = 0; m < conf->copies; m++) {
2156 int dev = r10_bio->devs[m].devnum;
2157 rdev = conf->mirrors[dev].rdev;
2158 if (r10_bio->devs[m].bio == NULL)
2159 continue;
2160 if (test_bit(BIO_UPTODATE,
2161 &r10_bio->devs[m].bio->bi_flags)) {
2162 rdev_clear_badblocks(
2163 rdev,
2164 r10_bio->devs[m].addr,
2165 r10_bio->sectors);
2166 } else {
2167 if (!rdev_set_badblocks(
2168 rdev,
2169 r10_bio->devs[m].addr,
2170 r10_bio->sectors, 0))
2171 md_error(conf->mddev, rdev);
2172 }
2173 }
2174 put_buf(r10_bio);
2175 } else {
2176 for (m = 0; m < conf->copies; m++) {
2177 int dev = r10_bio->devs[m].devnum;
2178 struct bio *bio = r10_bio->devs[m].bio;
2179 rdev = conf->mirrors[dev].rdev;
2180 if (bio == IO_MADE_GOOD) {
2181 rdev_clear_badblocks(
2182 rdev,
2183 r10_bio->devs[m].addr,
2184 r10_bio->sectors);
2185 rdev_dec_pending(rdev, conf->mddev);
2186 } else if (bio != NULL &&
2187 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2188 if (!narrow_write_error(r10_bio, m)) {
2189 md_error(conf->mddev, rdev);
2190 set_bit(R10BIO_Degraded,
2191 &r10_bio->state);
2192 }
2193 rdev_dec_pending(rdev, conf->mddev);
2194 }
2195 }
2196 if (test_bit(R10BIO_WriteError,
2197 &r10_bio->state))
2198 close_write(r10_bio);
2199 raid_end_bio_io(r10_bio);
2200 }
2201}
2202
2203static void raid10d(mddev_t *mddev)
2204{
2205 r10bio_t *r10_bio;
2206 unsigned long flags;
2207 conf_t *conf = mddev->private;
2208 struct list_head *head = &conf->retry_list;
2209 struct blk_plug plug;
2210
2211 md_check_recovery(mddev);
2212
2213 blk_start_plug(&plug);
2214 for (;;) {
2215
2216 flush_pending_writes(conf);
2217
2218 spin_lock_irqsave(&conf->device_lock, flags);
2219 if (list_empty(head)) {
2220 spin_unlock_irqrestore(&conf->device_lock, flags);
2221 break;
2222 }
2223 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
2224 list_del(head->prev);
2225 conf->nr_queued--;
2226 spin_unlock_irqrestore(&conf->device_lock, flags);
2227
2228 mddev = r10_bio->mddev;
2229 conf = mddev->private;
2230 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2231 test_bit(R10BIO_WriteError, &r10_bio->state))
2232 handle_write_completed(conf, r10_bio);
2233 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2234 sync_request_write(mddev, r10_bio);
2235 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2236 recovery_request_write(mddev, r10_bio);
2237 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2238 handle_read_error(mddev, r10_bio);
2239 else {
2240 /* just a partial read to be scheduled from a
2241 * separate context
2242 */
2243 int slot = r10_bio->read_slot;
2244 generic_make_request(r10_bio->devs[slot].bio);
2245 }
2246
2247 cond_resched();
2248 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2249 md_check_recovery(mddev);
2250 }
2251 blk_finish_plug(&plug);
2252}
2253
2254
2255static int init_resync(conf_t *conf)
2256{
2257 int buffs;
2258
2259 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2260 BUG_ON(conf->r10buf_pool);
2261 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2262 if (!conf->r10buf_pool)
2263 return -ENOMEM;
2264 conf->next_resync = 0;
2265 return 0;
2266}
2267
2268/*
2269 * perform a "sync" on one "block"
2270 *
2271 * We need to make sure that no normal I/O request - particularly write
2272 * requests - conflict with active sync requests.
2273 *
2274 * This is achieved by tracking pending requests and a 'barrier' concept
2275 * that can be installed to exclude normal IO requests.
2276 *
2277 * Resync and recovery are handled very differently.
2278 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2279 *
2280 * For resync, we iterate over virtual addresses, read all copies,
2281 * and update if there are differences. If only one copy is live,
2282 * skip it.
2283 * For recovery, we iterate over physical addresses, read a good
2284 * value for each non-in_sync drive, and over-write.
2285 *
2286 * So, for recovery we may have several outstanding complex requests for a
2287 * given address, one for each out-of-sync device. We model this by allocating
2288 * a number of r10_bio structures, one for each out-of-sync device.
2289 * As we setup these structures, we collect all bio's together into a list
2290 * which we then process collectively to add pages, and then process again
2291 * to pass to generic_make_request.
2292 *
2293 * The r10_bio structures are linked using a borrowed master_bio pointer.
2294 * This link is counted in ->remaining. When the r10_bio that points to NULL
2295 * has its remaining count decremented to 0, the whole complex operation
2296 * is complete.
2297 *
2298 */
2299
2300static sector_t sync_request(mddev_t *mddev, sector_t sector_nr,
2301 int *skipped, int go_faster)
2302{
2303 conf_t *conf = mddev->private;
2304 r10bio_t *r10_bio;
2305 struct bio *biolist = NULL, *bio;
2306 sector_t max_sector, nr_sectors;
2307 int i;
2308 int max_sync;
2309 sector_t sync_blocks;
2310 sector_t sectors_skipped = 0;
2311 int chunks_skipped = 0;
2312
2313 if (!conf->r10buf_pool)
2314 if (init_resync(conf))
2315 return 0;
2316
2317 skipped:
2318 max_sector = mddev->dev_sectors;
2319 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2320 max_sector = mddev->resync_max_sectors;
2321 if (sector_nr >= max_sector) {
2322 /* If we aborted, we need to abort the
2323 * sync on the 'current' bitmap chucks (there can
2324 * be several when recovering multiple devices).
2325 * as we may have started syncing it but not finished.
2326 * We can find the current address in
2327 * mddev->curr_resync, but for recovery,
2328 * we need to convert that to several
2329 * virtual addresses.
2330 */
2331 if (mddev->curr_resync < max_sector) { /* aborted */
2332 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2333 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2334 &sync_blocks, 1);
2335 else for (i=0; i<conf->raid_disks; i++) {
2336 sector_t sect =
2337 raid10_find_virt(conf, mddev->curr_resync, i);
2338 bitmap_end_sync(mddev->bitmap, sect,
2339 &sync_blocks, 1);
2340 }
2341 } else /* completed sync */
2342 conf->fullsync = 0;
2343
2344 bitmap_close_sync(mddev->bitmap);
2345 close_sync(conf);
2346 *skipped = 1;
2347 return sectors_skipped;
2348 }
2349 if (chunks_skipped >= conf->raid_disks) {
2350 /* if there has been nothing to do on any drive,
2351 * then there is nothing to do at all..
2352 */
2353 *skipped = 1;
2354 return (max_sector - sector_nr) + sectors_skipped;
2355 }
2356
2357 if (max_sector > mddev->resync_max)
2358 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2359
2360 /* make sure whole request will fit in a chunk - if chunks
2361 * are meaningful
2362 */
2363 if (conf->near_copies < conf->raid_disks &&
2364 max_sector > (sector_nr | conf->chunk_mask))
2365 max_sector = (sector_nr | conf->chunk_mask) + 1;
2366 /*
2367 * If there is non-resync activity waiting for us then
2368 * put in a delay to throttle resync.
2369 */
2370 if (!go_faster && conf->nr_waiting)
2371 msleep_interruptible(1000);
2372
2373 /* Again, very different code for resync and recovery.
2374 * Both must result in an r10bio with a list of bios that
2375 * have bi_end_io, bi_sector, bi_bdev set,
2376 * and bi_private set to the r10bio.
2377 * For recovery, we may actually create several r10bios
2378 * with 2 bios in each, that correspond to the bios in the main one.
2379 * In this case, the subordinate r10bios link back through a
2380 * borrowed master_bio pointer, and the counter in the master
2381 * includes a ref from each subordinate.
2382 */
2383 /* First, we decide what to do and set ->bi_end_io
2384 * To end_sync_read if we want to read, and
2385 * end_sync_write if we will want to write.
2386 */
2387
2388 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2389 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2390 /* recovery... the complicated one */
2391 int j;
2392 r10_bio = NULL;
2393
2394 for (i=0 ; i<conf->raid_disks; i++) {
2395 int still_degraded;
2396 r10bio_t *rb2;
2397 sector_t sect;
2398 int must_sync;
2399 int any_working;
2400
2401 if (conf->mirrors[i].rdev == NULL ||
2402 test_bit(In_sync, &conf->mirrors[i].rdev->flags))
2403 continue;
2404
2405 still_degraded = 0;
2406 /* want to reconstruct this device */
2407 rb2 = r10_bio;
2408 sect = raid10_find_virt(conf, sector_nr, i);
2409 /* Unless we are doing a full sync, we only need
2410 * to recover the block if it is set in the bitmap
2411 */
2412 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2413 &sync_blocks, 1);
2414 if (sync_blocks < max_sync)
2415 max_sync = sync_blocks;
2416 if (!must_sync &&
2417 !conf->fullsync) {
2418 /* yep, skip the sync_blocks here, but don't assume
2419 * that there will never be anything to do here
2420 */
2421 chunks_skipped = -1;
2422 continue;
2423 }
2424
2425 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2426 raise_barrier(conf, rb2 != NULL);
2427 atomic_set(&r10_bio->remaining, 0);
2428
2429 r10_bio->master_bio = (struct bio*)rb2;
2430 if (rb2)
2431 atomic_inc(&rb2->remaining);
2432 r10_bio->mddev = mddev;
2433 set_bit(R10BIO_IsRecover, &r10_bio->state);
2434 r10_bio->sector = sect;
2435
2436 raid10_find_phys(conf, r10_bio);
2437
2438 /* Need to check if the array will still be
2439 * degraded
2440 */
2441 for (j=0; j<conf->raid_disks; j++)
2442 if (conf->mirrors[j].rdev == NULL ||
2443 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2444 still_degraded = 1;
2445 break;
2446 }
2447
2448 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2449 &sync_blocks, still_degraded);
2450
2451 any_working = 0;
2452 for (j=0; j<conf->copies;j++) {
2453 int k;
2454 int d = r10_bio->devs[j].devnum;
2455 sector_t from_addr, to_addr;
2456 mdk_rdev_t *rdev;
2457 sector_t sector, first_bad;
2458 int bad_sectors;
2459 if (!conf->mirrors[d].rdev ||
2460 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2461 continue;
2462 /* This is where we read from */
2463 any_working = 1;
2464 rdev = conf->mirrors[d].rdev;
2465 sector = r10_bio->devs[j].addr;
2466
2467 if (is_badblock(rdev, sector, max_sync,
2468 &first_bad, &bad_sectors)) {
2469 if (first_bad > sector)
2470 max_sync = first_bad - sector;
2471 else {
2472 bad_sectors -= (sector
2473 - first_bad);
2474 if (max_sync > bad_sectors)
2475 max_sync = bad_sectors;
2476 continue;
2477 }
2478 }
2479 bio = r10_bio->devs[0].bio;
2480 bio->bi_next = biolist;
2481 biolist = bio;
2482 bio->bi_private = r10_bio;
2483 bio->bi_end_io = end_sync_read;
2484 bio->bi_rw = READ;
2485 from_addr = r10_bio->devs[j].addr;
2486 bio->bi_sector = from_addr +
2487 conf->mirrors[d].rdev->data_offset;
2488 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2489 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2490 atomic_inc(&r10_bio->remaining);
2491 /* and we write to 'i' */
2492
2493 for (k=0; k<conf->copies; k++)
2494 if (r10_bio->devs[k].devnum == i)
2495 break;
2496 BUG_ON(k == conf->copies);
2497 bio = r10_bio->devs[1].bio;
2498 bio->bi_next = biolist;
2499 biolist = bio;
2500 bio->bi_private = r10_bio;
2501 bio->bi_end_io = end_sync_write;
2502 bio->bi_rw = WRITE;
2503 to_addr = r10_bio->devs[k].addr;
2504 bio->bi_sector = to_addr +
2505 conf->mirrors[i].rdev->data_offset;
2506 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
2507
2508 r10_bio->devs[0].devnum = d;
2509 r10_bio->devs[0].addr = from_addr;
2510 r10_bio->devs[1].devnum = i;
2511 r10_bio->devs[1].addr = to_addr;
2512
2513 break;
2514 }
2515 if (j == conf->copies) {
2516 /* Cannot recover, so abort the recovery or
2517 * record a bad block */
2518 put_buf(r10_bio);
2519 if (rb2)
2520 atomic_dec(&rb2->remaining);
2521 r10_bio = rb2;
2522 if (any_working) {
2523 /* problem is that there are bad blocks
2524 * on other device(s)
2525 */
2526 int k;
2527 for (k = 0; k < conf->copies; k++)
2528 if (r10_bio->devs[k].devnum == i)
2529 break;
2530 if (!rdev_set_badblocks(
2531 conf->mirrors[i].rdev,
2532 r10_bio->devs[k].addr,
2533 max_sync, 0))
2534 any_working = 0;
2535 }
2536 if (!any_working) {
2537 if (!test_and_set_bit(MD_RECOVERY_INTR,
2538 &mddev->recovery))
2539 printk(KERN_INFO "md/raid10:%s: insufficient "
2540 "working devices for recovery.\n",
2541 mdname(mddev));
2542 conf->mirrors[i].recovery_disabled
2543 = mddev->recovery_disabled;
2544 }
2545 break;
2546 }
2547 }
2548 if (biolist == NULL) {
2549 while (r10_bio) {
2550 r10bio_t *rb2 = r10_bio;
2551 r10_bio = (r10bio_t*) rb2->master_bio;
2552 rb2->master_bio = NULL;
2553 put_buf(rb2);
2554 }
2555 goto giveup;
2556 }
2557 } else {
2558 /* resync. Schedule a read for every block at this virt offset */
2559 int count = 0;
2560
2561 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2562
2563 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2564 &sync_blocks, mddev->degraded) &&
2565 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2566 &mddev->recovery)) {
2567 /* We can skip this block */
2568 *skipped = 1;
2569 return sync_blocks + sectors_skipped;
2570 }
2571 if (sync_blocks < max_sync)
2572 max_sync = sync_blocks;
2573 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2574
2575 r10_bio->mddev = mddev;
2576 atomic_set(&r10_bio->remaining, 0);
2577 raise_barrier(conf, 0);
2578 conf->next_resync = sector_nr;
2579
2580 r10_bio->master_bio = NULL;
2581 r10_bio->sector = sector_nr;
2582 set_bit(R10BIO_IsSync, &r10_bio->state);
2583 raid10_find_phys(conf, r10_bio);
2584 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2585
2586 for (i=0; i<conf->copies; i++) {
2587 int d = r10_bio->devs[i].devnum;
2588 sector_t first_bad, sector;
2589 int bad_sectors;
2590
2591 bio = r10_bio->devs[i].bio;
2592 bio->bi_end_io = NULL;
2593 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2594 if (conf->mirrors[d].rdev == NULL ||
2595 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2596 continue;
2597 sector = r10_bio->devs[i].addr;
2598 if (is_badblock(conf->mirrors[d].rdev,
2599 sector, max_sync,
2600 &first_bad, &bad_sectors)) {
2601 if (first_bad > sector)
2602 max_sync = first_bad - sector;
2603 else {
2604 bad_sectors -= (sector - first_bad);
2605 if (max_sync > bad_sectors)
2606 max_sync = max_sync;
2607 continue;
2608 }
2609 }
2610 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2611 atomic_inc(&r10_bio->remaining);
2612 bio->bi_next = biolist;
2613 biolist = bio;
2614 bio->bi_private = r10_bio;
2615 bio->bi_end_io = end_sync_read;
2616 bio->bi_rw = READ;
2617 bio->bi_sector = sector +
2618 conf->mirrors[d].rdev->data_offset;
2619 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2620 count++;
2621 }
2622
2623 if (count < 2) {
2624 for (i=0; i<conf->copies; i++) {
2625 int d = r10_bio->devs[i].devnum;
2626 if (r10_bio->devs[i].bio->bi_end_io)
2627 rdev_dec_pending(conf->mirrors[d].rdev,
2628 mddev);
2629 }
2630 put_buf(r10_bio);
2631 biolist = NULL;
2632 goto giveup;
2633 }
2634 }
2635
2636 for (bio = biolist; bio ; bio=bio->bi_next) {
2637
2638 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2639 if (bio->bi_end_io)
2640 bio->bi_flags |= 1 << BIO_UPTODATE;
2641 bio->bi_vcnt = 0;
2642 bio->bi_idx = 0;
2643 bio->bi_phys_segments = 0;
2644 bio->bi_size = 0;
2645 }
2646
2647 nr_sectors = 0;
2648 if (sector_nr + max_sync < max_sector)
2649 max_sector = sector_nr + max_sync;
2650 do {
2651 struct page *page;
2652 int len = PAGE_SIZE;
2653 if (sector_nr + (len>>9) > max_sector)
2654 len = (max_sector - sector_nr) << 9;
2655 if (len == 0)
2656 break;
2657 for (bio= biolist ; bio ; bio=bio->bi_next) {
2658 struct bio *bio2;
2659 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2660 if (bio_add_page(bio, page, len, 0))
2661 continue;
2662
2663 /* stop here */
2664 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2665 for (bio2 = biolist;
2666 bio2 && bio2 != bio;
2667 bio2 = bio2->bi_next) {
2668 /* remove last page from this bio */
2669 bio2->bi_vcnt--;
2670 bio2->bi_size -= len;
2671 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2672 }
2673 goto bio_full;
2674 }
2675 nr_sectors += len>>9;
2676 sector_nr += len>>9;
2677 } while (biolist->bi_vcnt < RESYNC_PAGES);
2678 bio_full:
2679 r10_bio->sectors = nr_sectors;
2680
2681 while (biolist) {
2682 bio = biolist;
2683 biolist = biolist->bi_next;
2684
2685 bio->bi_next = NULL;
2686 r10_bio = bio->bi_private;
2687 r10_bio->sectors = nr_sectors;
2688
2689 if (bio->bi_end_io == end_sync_read) {
2690 md_sync_acct(bio->bi_bdev, nr_sectors);
2691 generic_make_request(bio);
2692 }
2693 }
2694
2695 if (sectors_skipped)
2696 /* pretend they weren't skipped, it makes
2697 * no important difference in this case
2698 */
2699 md_done_sync(mddev, sectors_skipped, 1);
2700
2701 return sectors_skipped + nr_sectors;
2702 giveup:
2703 /* There is nowhere to write, so all non-sync
2704 * drives must be failed or in resync, all drives
2705 * have a bad block, so try the next chunk...
2706 */
2707 if (sector_nr + max_sync < max_sector)
2708 max_sector = sector_nr + max_sync;
2709
2710 sectors_skipped += (max_sector - sector_nr);
2711 chunks_skipped ++;
2712 sector_nr = max_sector;
2713 goto skipped;
2714}
2715
2716static sector_t
2717raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2718{
2719 sector_t size;
2720 conf_t *conf = mddev->private;
2721
2722 if (!raid_disks)
2723 raid_disks = conf->raid_disks;
2724 if (!sectors)
2725 sectors = conf->dev_sectors;
2726
2727 size = sectors >> conf->chunk_shift;
2728 sector_div(size, conf->far_copies);
2729 size = size * raid_disks;
2730 sector_div(size, conf->near_copies);
2731
2732 return size << conf->chunk_shift;
2733}
2734
2735
2736static conf_t *setup_conf(mddev_t *mddev)
2737{
2738 conf_t *conf = NULL;
2739 int nc, fc, fo;
2740 sector_t stride, size;
2741 int err = -EINVAL;
2742
2743 if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2744 !is_power_of_2(mddev->new_chunk_sectors)) {
2745 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2746 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2747 mdname(mddev), PAGE_SIZE);
2748 goto out;
2749 }
2750
2751 nc = mddev->new_layout & 255;
2752 fc = (mddev->new_layout >> 8) & 255;
2753 fo = mddev->new_layout & (1<<16);
2754
2755 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2756 (mddev->new_layout >> 17)) {
2757 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2758 mdname(mddev), mddev->new_layout);
2759 goto out;
2760 }
2761
2762 err = -ENOMEM;
2763 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2764 if (!conf)
2765 goto out;
2766
2767 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2768 GFP_KERNEL);
2769 if (!conf->mirrors)
2770 goto out;
2771
2772 conf->tmppage = alloc_page(GFP_KERNEL);
2773 if (!conf->tmppage)
2774 goto out;
2775
2776
2777 conf->raid_disks = mddev->raid_disks;
2778 conf->near_copies = nc;
2779 conf->far_copies = fc;
2780 conf->copies = nc*fc;
2781 conf->far_offset = fo;
2782 conf->chunk_mask = mddev->new_chunk_sectors - 1;
2783 conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2784
2785 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2786 r10bio_pool_free, conf);
2787 if (!conf->r10bio_pool)
2788 goto out;
2789
2790 size = mddev->dev_sectors >> conf->chunk_shift;
2791 sector_div(size, fc);
2792 size = size * conf->raid_disks;
2793 sector_div(size, nc);
2794 /* 'size' is now the number of chunks in the array */
2795 /* calculate "used chunks per device" in 'stride' */
2796 stride = size * conf->copies;
2797
2798 /* We need to round up when dividing by raid_disks to
2799 * get the stride size.
2800 */
2801 stride += conf->raid_disks - 1;
2802 sector_div(stride, conf->raid_disks);
2803
2804 conf->dev_sectors = stride << conf->chunk_shift;
2805
2806 if (fo)
2807 stride = 1;
2808 else
2809 sector_div(stride, fc);
2810 conf->stride = stride << conf->chunk_shift;
2811
2812
2813 spin_lock_init(&conf->device_lock);
2814 INIT_LIST_HEAD(&conf->retry_list);
2815
2816 spin_lock_init(&conf->resync_lock);
2817 init_waitqueue_head(&conf->wait_barrier);
2818
2819 conf->thread = md_register_thread(raid10d, mddev, NULL);
2820 if (!conf->thread)
2821 goto out;
2822
2823 conf->mddev = mddev;
2824 return conf;
2825
2826 out:
2827 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2828 mdname(mddev));
2829 if (conf) {
2830 if (conf->r10bio_pool)
2831 mempool_destroy(conf->r10bio_pool);
2832 kfree(conf->mirrors);
2833 safe_put_page(conf->tmppage);
2834 kfree(conf);
2835 }
2836 return ERR_PTR(err);
2837}
2838
2839static int run(mddev_t *mddev)
2840{
2841 conf_t *conf;
2842 int i, disk_idx, chunk_size;
2843 mirror_info_t *disk;
2844 mdk_rdev_t *rdev;
2845 sector_t size;
2846
2847 /*
2848 * copy the already verified devices into our private RAID10
2849 * bookkeeping area. [whatever we allocate in run(),
2850 * should be freed in stop()]
2851 */
2852
2853 if (mddev->private == NULL) {
2854 conf = setup_conf(mddev);
2855 if (IS_ERR(conf))
2856 return PTR_ERR(conf);
2857 mddev->private = conf;
2858 }
2859 conf = mddev->private;
2860 if (!conf)
2861 goto out;
2862
2863 mddev->thread = conf->thread;
2864 conf->thread = NULL;
2865
2866 chunk_size = mddev->chunk_sectors << 9;
2867 blk_queue_io_min(mddev->queue, chunk_size);
2868 if (conf->raid_disks % conf->near_copies)
2869 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2870 else
2871 blk_queue_io_opt(mddev->queue, chunk_size *
2872 (conf->raid_disks / conf->near_copies));
2873
2874 list_for_each_entry(rdev, &mddev->disks, same_set) {
2875
2876 disk_idx = rdev->raid_disk;
2877 if (disk_idx >= conf->raid_disks
2878 || disk_idx < 0)
2879 continue;
2880 disk = conf->mirrors + disk_idx;
2881
2882 disk->rdev = rdev;
2883 disk_stack_limits(mddev->gendisk, rdev->bdev,
2884 rdev->data_offset << 9);
2885 /* as we don't honour merge_bvec_fn, we must never risk
2886 * violating it, so limit max_segments to 1 lying
2887 * within a single page.
2888 */
2889 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2890 blk_queue_max_segments(mddev->queue, 1);
2891 blk_queue_segment_boundary(mddev->queue,
2892 PAGE_CACHE_SIZE - 1);
2893 }
2894
2895 disk->head_position = 0;
2896 }
2897 /* need to check that every block has at least one working mirror */
2898 if (!enough(conf, -1)) {
2899 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2900 mdname(mddev));
2901 goto out_free_conf;
2902 }
2903
2904 mddev->degraded = 0;
2905 for (i = 0; i < conf->raid_disks; i++) {
2906
2907 disk = conf->mirrors + i;
2908
2909 if (!disk->rdev ||
2910 !test_bit(In_sync, &disk->rdev->flags)) {
2911 disk->head_position = 0;
2912 mddev->degraded++;
2913 if (disk->rdev)
2914 conf->fullsync = 1;
2915 }
2916 }
2917
2918 if (mddev->recovery_cp != MaxSector)
2919 printk(KERN_NOTICE "md/raid10:%s: not clean"
2920 " -- starting background reconstruction\n",
2921 mdname(mddev));
2922 printk(KERN_INFO
2923 "md/raid10:%s: active with %d out of %d devices\n",
2924 mdname(mddev), conf->raid_disks - mddev->degraded,
2925 conf->raid_disks);
2926 /*
2927 * Ok, everything is just fine now
2928 */
2929 mddev->dev_sectors = conf->dev_sectors;
2930 size = raid10_size(mddev, 0, 0);
2931 md_set_array_sectors(mddev, size);
2932 mddev->resync_max_sectors = size;
2933
2934 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2935 mddev->queue->backing_dev_info.congested_data = mddev;
2936
2937 /* Calculate max read-ahead size.
2938 * We need to readahead at least twice a whole stripe....
2939 * maybe...
2940 */
2941 {
2942 int stripe = conf->raid_disks *
2943 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2944 stripe /= conf->near_copies;
2945 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2946 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2947 }
2948
2949 if (conf->near_copies < conf->raid_disks)
2950 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2951
2952 if (md_integrity_register(mddev))
2953 goto out_free_conf;
2954
2955 return 0;
2956
2957out_free_conf:
2958 md_unregister_thread(&mddev->thread);
2959 if (conf->r10bio_pool)
2960 mempool_destroy(conf->r10bio_pool);
2961 safe_put_page(conf->tmppage);
2962 kfree(conf->mirrors);
2963 kfree(conf);
2964 mddev->private = NULL;
2965out:
2966 return -EIO;
2967}
2968
2969static int stop(mddev_t *mddev)
2970{
2971 conf_t *conf = mddev->private;
2972
2973 raise_barrier(conf, 0);
2974 lower_barrier(conf);
2975
2976 md_unregister_thread(&mddev->thread);
2977 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2978 if (conf->r10bio_pool)
2979 mempool_destroy(conf->r10bio_pool);
2980 kfree(conf->mirrors);
2981 kfree(conf);
2982 mddev->private = NULL;
2983 return 0;
2984}
2985
2986static void raid10_quiesce(mddev_t *mddev, int state)
2987{
2988 conf_t *conf = mddev->private;
2989
2990 switch(state) {
2991 case 1:
2992 raise_barrier(conf, 0);
2993 break;
2994 case 0:
2995 lower_barrier(conf);
2996 break;
2997 }
2998}
2999
3000static void *raid10_takeover_raid0(mddev_t *mddev)
3001{
3002 mdk_rdev_t *rdev;
3003 conf_t *conf;
3004
3005 if (mddev->degraded > 0) {
3006 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3007 mdname(mddev));
3008 return ERR_PTR(-EINVAL);
3009 }
3010
3011 /* Set new parameters */
3012 mddev->new_level = 10;
3013 /* new layout: far_copies = 1, near_copies = 2 */
3014 mddev->new_layout = (1<<8) + 2;
3015 mddev->new_chunk_sectors = mddev->chunk_sectors;
3016 mddev->delta_disks = mddev->raid_disks;
3017 mddev->raid_disks *= 2;
3018 /* make sure it will be not marked as dirty */
3019 mddev->recovery_cp = MaxSector;
3020
3021 conf = setup_conf(mddev);
3022 if (!IS_ERR(conf)) {
3023 list_for_each_entry(rdev, &mddev->disks, same_set)
3024 if (rdev->raid_disk >= 0)
3025 rdev->new_raid_disk = rdev->raid_disk * 2;
3026 conf->barrier = 1;
3027 }
3028
3029 return conf;
3030}
3031
3032static void *raid10_takeover(mddev_t *mddev)
3033{
3034 struct raid0_private_data *raid0_priv;
3035
3036 /* raid10 can take over:
3037 * raid0 - providing it has only two drives
3038 */
3039 if (mddev->level == 0) {
3040 /* for raid0 takeover only one zone is supported */
3041 raid0_priv = mddev->private;
3042 if (raid0_priv->nr_strip_zones > 1) {
3043 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3044 " with more than one zone.\n",
3045 mdname(mddev));
3046 return ERR_PTR(-EINVAL);
3047 }
3048 return raid10_takeover_raid0(mddev);
3049 }
3050 return ERR_PTR(-EINVAL);
3051}
3052
3053static struct mdk_personality raid10_personality =
3054{
3055 .name = "raid10",
3056 .level = 10,
3057 .owner = THIS_MODULE,
3058 .make_request = make_request,
3059 .run = run,
3060 .stop = stop,
3061 .status = status,
3062 .error_handler = error,
3063 .hot_add_disk = raid10_add_disk,
3064 .hot_remove_disk= raid10_remove_disk,
3065 .spare_active = raid10_spare_active,
3066 .sync_request = sync_request,
3067 .quiesce = raid10_quiesce,
3068 .size = raid10_size,
3069 .takeover = raid10_takeover,
3070};
3071
3072static int __init raid_init(void)
3073{
3074 return register_md_personality(&raid10_personality);
3075}
3076
3077static void raid_exit(void)
3078{
3079 unregister_md_personality(&raid10_personality);
3080}
3081
3082module_init(raid_init);
3083module_exit(raid_exit);
3084MODULE_LICENSE("GPL");
3085MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3086MODULE_ALIAS("md-personality-9"); /* RAID10 */
3087MODULE_ALIAS("md-raid10");
3088MODULE_ALIAS("md-level-10");