1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * raid10.c : Multiple Devices driver for Linux
4 *
5 * Copyright (C) 2000-2004 Neil Brown
6 *
7 * RAID-10 support for md.
8 *
9 * Base on code in raid1.c. See raid1.c for further copyright information.
10 */
11
12 #include <linux/slab.h>
13 #include <linux/delay.h>
14 #include <linux/blkdev.h>
15 #include <linux/module.h>
16 #include <linux/seq_file.h>
17 #include <linux/ratelimit.h>
18 #include <linux/kthread.h>
19 #include <linux/raid/md_p.h>
20 #include <trace/events/block.h>
21 #include "md.h"
22 #include "raid10.h"
23 #include "raid0.h"
24 #include "md-bitmap.h"
25
26 /*
27 * RAID10 provides a combination of RAID0 and RAID1 functionality.
28 * The layout of data is defined by
29 * chunk_size
30 * raid_disks
31 * near_copies (stored in low byte of layout)
32 * far_copies (stored in second byte of layout)
33 * far_offset (stored in bit 16 of layout )
34 * use_far_sets (stored in bit 17 of layout )
35 * use_far_sets_bugfixed (stored in bit 18 of layout )
36 *
37 * The data to be stored is divided into chunks using chunksize. Each device
38 * is divided into far_copies sections. In each section, chunks are laid out
39 * in a style similar to raid0, but near_copies copies of each chunk is stored
40 * (each on a different drive). The starting device for each section is offset
41 * near_copies from the starting device of the previous section. Thus there
42 * are (near_copies * far_copies) of each chunk, and each is on a different
43 * drive. near_copies and far_copies must be at least one, and their product
44 * is at most raid_disks.
45 *
46 * If far_offset is true, then the far_copies are handled a bit differently.
47 * The copies are still in different stripes, but instead of being very far
48 * apart on disk, there are adjacent stripes.
49 *
50 * The far and offset algorithms are handled slightly differently if
51 * 'use_far_sets' is true. In this case, the array's devices are grouped into
52 * sets that are (near_copies * far_copies) in size. The far copied stripes
53 * are still shifted by 'near_copies' devices, but this shifting stays confined
54 * to the set rather than the entire array. This is done to improve the number
55 * of device combinations that can fail without causing the array to fail.
56 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
57 * on a device):
58 * A B C D A B C D E
59 * ... ...
60 * D A B C E A B C D
61 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
62 * [A B] [C D] [A B] [C D E]
63 * |...| |...| |...| | ... |
64 * [B A] [D C] [B A] [E C D]
65 */
66
67 static void allow_barrier(struct r10conf *conf);
68 static void lower_barrier(struct r10conf *conf);
69 static int _enough(struct r10conf *conf, int previous, int ignore);
70 static int enough(struct r10conf *conf, int ignore);
71 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
72 int *skipped);
73 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
74 static void end_reshape_write(struct bio *bio);
75 static void end_reshape(struct r10conf *conf);
76
77 #define raid10_log(md, fmt, args...) \
78 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
79
80 #include "raid1-10.c"
81
82 /*
83 * for resync bio, r10bio pointer can be retrieved from the per-bio
84 * 'struct resync_pages'.
85 */
get_resync_r10bio(struct bio *bio)86 static inline struct r10bio *get_resync_r10bio(struct bio *bio)
87 {
88 return get_resync_pages(bio)->raid_bio;
89 }
90
r10bio_pool_alloc(gfp_t gfp_flags, void *data)91 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
92 {
93 struct r10conf *conf = data;
94 int size = offsetof(struct r10bio, devs[conf->copies]);
95
96 /* allocate a r10bio with room for raid_disks entries in the
97 * bios array */
98 return kzalloc(size, gfp_flags);
99 }
100
101 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
102 /* amount of memory to reserve for resync requests */
103 #define RESYNC_WINDOW (1024*1024)
104 /* maximum number of concurrent requests, memory permitting */
105 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
106 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
107 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
108
109 /*
110 * When performing a resync, we need to read and compare, so
111 * we need as many pages are there are copies.
112 * When performing a recovery, we need 2 bios, one for read,
113 * one for write (we recover only one drive per r10buf)
114 *
115 */
r10buf_pool_alloc(gfp_t gfp_flags, void *data)116 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
117 {
118 struct r10conf *conf = data;
119 struct r10bio *r10_bio;
120 struct bio *bio;
121 int j;
122 int nalloc, nalloc_rp;
123 struct resync_pages *rps;
124
125 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
126 if (!r10_bio)
127 return NULL;
128
129 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
130 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
131 nalloc = conf->copies; /* resync */
132 else
133 nalloc = 2; /* recovery */
134
135 /* allocate once for all bios */
136 if (!conf->have_replacement)
137 nalloc_rp = nalloc;
138 else
139 nalloc_rp = nalloc * 2;
140 rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
141 if (!rps)
142 goto out_free_r10bio;
143
144 /*
145 * Allocate bios.
146 */
147 for (j = nalloc ; j-- ; ) {
148 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
149 if (!bio)
150 goto out_free_bio;
151 r10_bio->devs[j].bio = bio;
152 if (!conf->have_replacement)
153 continue;
154 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
155 if (!bio)
156 goto out_free_bio;
157 r10_bio->devs[j].repl_bio = bio;
158 }
159 /*
160 * Allocate RESYNC_PAGES data pages and attach them
161 * where needed.
162 */
163 for (j = 0; j < nalloc; j++) {
164 struct bio *rbio = r10_bio->devs[j].repl_bio;
165 struct resync_pages *rp, *rp_repl;
166
167 rp = &rps[j];
168 if (rbio)
169 rp_repl = &rps[nalloc + j];
170
171 bio = r10_bio->devs[j].bio;
172
173 if (!j || test_bit(MD_RECOVERY_SYNC,
174 &conf->mddev->recovery)) {
175 if (resync_alloc_pages(rp, gfp_flags))
176 goto out_free_pages;
177 } else {
178 memcpy(rp, &rps[0], sizeof(*rp));
179 resync_get_all_pages(rp);
180 }
181
182 rp->raid_bio = r10_bio;
183 bio->bi_private = rp;
184 if (rbio) {
185 memcpy(rp_repl, rp, sizeof(*rp));
186 rbio->bi_private = rp_repl;
187 }
188 }
189
190 return r10_bio;
191
192 out_free_pages:
193 while (--j >= 0)
194 resync_free_pages(&rps[j]);
195
196 j = 0;
197 out_free_bio:
198 for ( ; j < nalloc; j++) {
199 if (r10_bio->devs[j].bio)
200 bio_put(r10_bio->devs[j].bio);
201 if (r10_bio->devs[j].repl_bio)
202 bio_put(r10_bio->devs[j].repl_bio);
203 }
204 kfree(rps);
205 out_free_r10bio:
206 rbio_pool_free(r10_bio, conf);
207 return NULL;
208 }
209
r10buf_pool_free(void *__r10_bio, void *data)210 static void r10buf_pool_free(void *__r10_bio, void *data)
211 {
212 struct r10conf *conf = data;
213 struct r10bio *r10bio = __r10_bio;
214 int j;
215 struct resync_pages *rp = NULL;
216
217 for (j = conf->copies; j--; ) {
218 struct bio *bio = r10bio->devs[j].bio;
219
220 if (bio) {
221 rp = get_resync_pages(bio);
222 resync_free_pages(rp);
223 bio_put(bio);
224 }
225
226 bio = r10bio->devs[j].repl_bio;
227 if (bio)
228 bio_put(bio);
229 }
230
231 /* resync pages array stored in the 1st bio's .bi_private */
232 kfree(rp);
233
234 rbio_pool_free(r10bio, conf);
235 }
236
put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)237 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
238 {
239 int i;
240
241 for (i = 0; i < conf->copies; i++) {
242 struct bio **bio = & r10_bio->devs[i].bio;
243 if (!BIO_SPECIAL(*bio))
244 bio_put(*bio);
245 *bio = NULL;
246 bio = &r10_bio->devs[i].repl_bio;
247 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
248 bio_put(*bio);
249 *bio = NULL;
250 }
251 }
252
free_r10bio(struct r10bio *r10_bio)253 static void free_r10bio(struct r10bio *r10_bio)
254 {
255 struct r10conf *conf = r10_bio->mddev->private;
256
257 put_all_bios(conf, r10_bio);
258 mempool_free(r10_bio, &conf->r10bio_pool);
259 }
260
put_buf(struct r10bio *r10_bio)261 static void put_buf(struct r10bio *r10_bio)
262 {
263 struct r10conf *conf = r10_bio->mddev->private;
264
265 mempool_free(r10_bio, &conf->r10buf_pool);
266
267 lower_barrier(conf);
268 }
269
reschedule_retry(struct r10bio *r10_bio)270 static void reschedule_retry(struct r10bio *r10_bio)
271 {
272 unsigned long flags;
273 struct mddev *mddev = r10_bio->mddev;
274 struct r10conf *conf = mddev->private;
275
276 spin_lock_irqsave(&conf->device_lock, flags);
277 list_add(&r10_bio->retry_list, &conf->retry_list);
278 conf->nr_queued ++;
279 spin_unlock_irqrestore(&conf->device_lock, flags);
280
281 /* wake up frozen array... */
282 wake_up(&conf->wait_barrier);
283
284 md_wakeup_thread(mddev->thread);
285 }
286
287 /*
288 * raid_end_bio_io() is called when we have finished servicing a mirrored
289 * operation and are ready to return a success/failure code to the buffer
290 * cache layer.
291 */
raid_end_bio_io(struct r10bio *r10_bio)292 static void raid_end_bio_io(struct r10bio *r10_bio)
293 {
294 struct bio *bio = r10_bio->master_bio;
295 struct r10conf *conf = r10_bio->mddev->private;
296
297 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
298 bio->bi_status = BLK_STS_IOERR;
299
300 bio_endio(bio);
301 /*
302 * Wake up any possible resync thread that waits for the device
303 * to go idle.
304 */
305 allow_barrier(conf);
306
307 free_r10bio(r10_bio);
308 }
309
310 /*
311 * Update disk head position estimator based on IRQ completion info.
312 */
update_head_pos(int slot, struct r10bio *r10_bio)313 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
314 {
315 struct r10conf *conf = r10_bio->mddev->private;
316
317 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
318 r10_bio->devs[slot].addr + (r10_bio->sectors);
319 }
320
321 /*
322 * Find the disk number which triggered given bio
323 */
find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio, struct bio *bio, int *slotp, int *replp)324 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
325 struct bio *bio, int *slotp, int *replp)
326 {
327 int slot;
328 int repl = 0;
329
330 for (slot = 0; slot < conf->copies; slot++) {
331 if (r10_bio->devs[slot].bio == bio)
332 break;
333 if (r10_bio->devs[slot].repl_bio == bio) {
334 repl = 1;
335 break;
336 }
337 }
338
339 BUG_ON(slot == conf->copies);
340 update_head_pos(slot, r10_bio);
341
342 if (slotp)
343 *slotp = slot;
344 if (replp)
345 *replp = repl;
346 return r10_bio->devs[slot].devnum;
347 }
348
raid10_end_read_request(struct bio *bio)349 static void raid10_end_read_request(struct bio *bio)
350 {
351 int uptodate = !bio->bi_status;
352 struct r10bio *r10_bio = bio->bi_private;
353 int slot;
354 struct md_rdev *rdev;
355 struct r10conf *conf = r10_bio->mddev->private;
356
357 slot = r10_bio->read_slot;
358 rdev = r10_bio->devs[slot].rdev;
359 /*
360 * this branch is our 'one mirror IO has finished' event handler:
361 */
362 update_head_pos(slot, r10_bio);
363
364 if (uptodate) {
365 /*
366 * Set R10BIO_Uptodate in our master bio, so that
367 * we will return a good error code to the higher
368 * levels even if IO on some other mirrored buffer fails.
369 *
370 * The 'master' represents the composite IO operation to
371 * user-side. So if something waits for IO, then it will
372 * wait for the 'master' bio.
373 */
374 set_bit(R10BIO_Uptodate, &r10_bio->state);
375 } else {
376 /* If all other devices that store this block have
377 * failed, we want to return the error upwards rather
378 * than fail the last device. Here we redefine
379 * "uptodate" to mean "Don't want to retry"
380 */
381 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
382 rdev->raid_disk))
383 uptodate = 1;
384 }
385 if (uptodate) {
386 raid_end_bio_io(r10_bio);
387 rdev_dec_pending(rdev, conf->mddev);
388 } else {
389 /*
390 * oops, read error - keep the refcount on the rdev
391 */
392 char b[BDEVNAME_SIZE];
393 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
394 mdname(conf->mddev),
395 bdevname(rdev->bdev, b),
396 (unsigned long long)r10_bio->sector);
397 set_bit(R10BIO_ReadError, &r10_bio->state);
398 reschedule_retry(r10_bio);
399 }
400 }
401
close_write(struct r10bio *r10_bio)402 static void close_write(struct r10bio *r10_bio)
403 {
404 /* clear the bitmap if all writes complete successfully */
405 md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
406 r10_bio->sectors,
407 !test_bit(R10BIO_Degraded, &r10_bio->state),
408 0);
409 md_write_end(r10_bio->mddev);
410 }
411
one_write_done(struct r10bio *r10_bio)412 static void one_write_done(struct r10bio *r10_bio)
413 {
414 if (atomic_dec_and_test(&r10_bio->remaining)) {
415 if (test_bit(R10BIO_WriteError, &r10_bio->state))
416 reschedule_retry(r10_bio);
417 else {
418 close_write(r10_bio);
419 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
420 reschedule_retry(r10_bio);
421 else
422 raid_end_bio_io(r10_bio);
423 }
424 }
425 }
426
raid10_end_write_request(struct bio *bio)427 static void raid10_end_write_request(struct bio *bio)
428 {
429 struct r10bio *r10_bio = bio->bi_private;
430 int dev;
431 int dec_rdev = 1;
432 struct r10conf *conf = r10_bio->mddev->private;
433 int slot, repl;
434 struct md_rdev *rdev = NULL;
435 struct bio *to_put = NULL;
436 bool discard_error;
437
438 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
439
440 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
441
442 if (repl)
443 rdev = conf->mirrors[dev].replacement;
444 if (!rdev) {
445 smp_rmb();
446 repl = 0;
447 rdev = conf->mirrors[dev].rdev;
448 }
449 /*
450 * this branch is our 'one mirror IO has finished' event handler:
451 */
452 if (bio->bi_status && !discard_error) {
453 if (repl)
454 /* Never record new bad blocks to replacement,
455 * just fail it.
456 */
457 md_error(rdev->mddev, rdev);
458 else {
459 set_bit(WriteErrorSeen, &rdev->flags);
460 if (!test_and_set_bit(WantReplacement, &rdev->flags))
461 set_bit(MD_RECOVERY_NEEDED,
462 &rdev->mddev->recovery);
463
464 dec_rdev = 0;
465 if (test_bit(FailFast, &rdev->flags) &&
466 (bio->bi_opf & MD_FAILFAST)) {
467 md_error(rdev->mddev, rdev);
468 }
469
470 /*
471 * When the device is faulty, it is not necessary to
472 * handle write error.
473 */
474 if (!test_bit(Faulty, &rdev->flags))
475 set_bit(R10BIO_WriteError, &r10_bio->state);
476 else {
477 /* Fail the request */
478 set_bit(R10BIO_Degraded, &r10_bio->state);
479 r10_bio->devs[slot].bio = NULL;
480 to_put = bio;
481 dec_rdev = 1;
482 }
483 }
484 } else {
485 /*
486 * Set R10BIO_Uptodate in our master bio, so that
487 * we will return a good error code for to the higher
488 * levels even if IO on some other mirrored buffer fails.
489 *
490 * The 'master' represents the composite IO operation to
491 * user-side. So if something waits for IO, then it will
492 * wait for the 'master' bio.
493 */
494 sector_t first_bad;
495 int bad_sectors;
496
497 /*
498 * Do not set R10BIO_Uptodate if the current device is
499 * rebuilding or Faulty. This is because we cannot use
500 * such device for properly reading the data back (we could
501 * potentially use it, if the current write would have felt
502 * before rdev->recovery_offset, but for simplicity we don't
503 * check this here.
504 */
505 if (test_bit(In_sync, &rdev->flags) &&
506 !test_bit(Faulty, &rdev->flags))
507 set_bit(R10BIO_Uptodate, &r10_bio->state);
508
509 /* Maybe we can clear some bad blocks. */
510 if (is_badblock(rdev,
511 r10_bio->devs[slot].addr,
512 r10_bio->sectors,
513 &first_bad, &bad_sectors) && !discard_error) {
514 bio_put(bio);
515 if (repl)
516 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
517 else
518 r10_bio->devs[slot].bio = IO_MADE_GOOD;
519 dec_rdev = 0;
520 set_bit(R10BIO_MadeGood, &r10_bio->state);
521 }
522 }
523
524 /*
525 *
526 * Let's see if all mirrored write operations have finished
527 * already.
528 */
529 one_write_done(r10_bio);
530 if (dec_rdev)
531 rdev_dec_pending(rdev, conf->mddev);
532 if (to_put)
533 bio_put(to_put);
534 }
535
536 /*
537 * RAID10 layout manager
538 * As well as the chunksize and raid_disks count, there are two
539 * parameters: near_copies and far_copies.
540 * near_copies * far_copies must be <= raid_disks.
541 * Normally one of these will be 1.
542 * If both are 1, we get raid0.
543 * If near_copies == raid_disks, we get raid1.
544 *
545 * Chunks are laid out in raid0 style with near_copies copies of the
546 * first chunk, followed by near_copies copies of the next chunk and
547 * so on.
548 * If far_copies > 1, then after 1/far_copies of the array has been assigned
549 * as described above, we start again with a device offset of near_copies.
550 * So we effectively have another copy of the whole array further down all
551 * the drives, but with blocks on different drives.
552 * With this layout, and block is never stored twice on the one device.
553 *
554 * raid10_find_phys finds the sector offset of a given virtual sector
555 * on each device that it is on.
556 *
557 * raid10_find_virt does the reverse mapping, from a device and a
558 * sector offset to a virtual address
559 */
560
__raid10_find_phys(struct geom *geo, struct r10bio *r10bio)561 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
562 {
563 int n,f;
564 sector_t sector;
565 sector_t chunk;
566 sector_t stripe;
567 int dev;
568 int slot = 0;
569 int last_far_set_start, last_far_set_size;
570
571 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
572 last_far_set_start *= geo->far_set_size;
573
574 last_far_set_size = geo->far_set_size;
575 last_far_set_size += (geo->raid_disks % geo->far_set_size);
576
577 /* now calculate first sector/dev */
578 chunk = r10bio->sector >> geo->chunk_shift;
579 sector = r10bio->sector & geo->chunk_mask;
580
581 chunk *= geo->near_copies;
582 stripe = chunk;
583 dev = sector_div(stripe, geo->raid_disks);
584 if (geo->far_offset)
585 stripe *= geo->far_copies;
586
587 sector += stripe << geo->chunk_shift;
588
589 /* and calculate all the others */
590 for (n = 0; n < geo->near_copies; n++) {
591 int d = dev;
592 int set;
593 sector_t s = sector;
594 r10bio->devs[slot].devnum = d;
595 r10bio->devs[slot].addr = s;
596 slot++;
597
598 for (f = 1; f < geo->far_copies; f++) {
599 set = d / geo->far_set_size;
600 d += geo->near_copies;
601
602 if ((geo->raid_disks % geo->far_set_size) &&
603 (d > last_far_set_start)) {
604 d -= last_far_set_start;
605 d %= last_far_set_size;
606 d += last_far_set_start;
607 } else {
608 d %= geo->far_set_size;
609 d += geo->far_set_size * set;
610 }
611 s += geo->stride;
612 r10bio->devs[slot].devnum = d;
613 r10bio->devs[slot].addr = s;
614 slot++;
615 }
616 dev++;
617 if (dev >= geo->raid_disks) {
618 dev = 0;
619 sector += (geo->chunk_mask + 1);
620 }
621 }
622 }
623
raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)624 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
625 {
626 struct geom *geo = &conf->geo;
627
628 if (conf->reshape_progress != MaxSector &&
629 ((r10bio->sector >= conf->reshape_progress) !=
630 conf->mddev->reshape_backwards)) {
631 set_bit(R10BIO_Previous, &r10bio->state);
632 geo = &conf->prev;
633 } else
634 clear_bit(R10BIO_Previous, &r10bio->state);
635
636 __raid10_find_phys(geo, r10bio);
637 }
638
raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)639 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
640 {
641 sector_t offset, chunk, vchunk;
642 /* Never use conf->prev as this is only called during resync
643 * or recovery, so reshape isn't happening
644 */
645 struct geom *geo = &conf->geo;
646 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
647 int far_set_size = geo->far_set_size;
648 int last_far_set_start;
649
650 if (geo->raid_disks % geo->far_set_size) {
651 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
652 last_far_set_start *= geo->far_set_size;
653
654 if (dev >= last_far_set_start) {
655 far_set_size = geo->far_set_size;
656 far_set_size += (geo->raid_disks % geo->far_set_size);
657 far_set_start = last_far_set_start;
658 }
659 }
660
661 offset = sector & geo->chunk_mask;
662 if (geo->far_offset) {
663 int fc;
664 chunk = sector >> geo->chunk_shift;
665 fc = sector_div(chunk, geo->far_copies);
666 dev -= fc * geo->near_copies;
667 if (dev < far_set_start)
668 dev += far_set_size;
669 } else {
670 while (sector >= geo->stride) {
671 sector -= geo->stride;
672 if (dev < (geo->near_copies + far_set_start))
673 dev += far_set_size - geo->near_copies;
674 else
675 dev -= geo->near_copies;
676 }
677 chunk = sector >> geo->chunk_shift;
678 }
679 vchunk = chunk * geo->raid_disks + dev;
680 sector_div(vchunk, geo->near_copies);
681 return (vchunk << geo->chunk_shift) + offset;
682 }
683
684 /*
685 * This routine returns the disk from which the requested read should
686 * be done. There is a per-array 'next expected sequential IO' sector
687 * number - if this matches on the next IO then we use the last disk.
688 * There is also a per-disk 'last know head position' sector that is
689 * maintained from IRQ contexts, both the normal and the resync IO
690 * completion handlers update this position correctly. If there is no
691 * perfect sequential match then we pick the disk whose head is closest.
692 *
693 * If there are 2 mirrors in the same 2 devices, performance degrades
694 * because position is mirror, not device based.
695 *
696 * The rdev for the device selected will have nr_pending incremented.
697 */
698
699 /*
700 * FIXME: possibly should rethink readbalancing and do it differently
701 * depending on near_copies / far_copies geometry.
702 */
read_balance(struct r10conf *conf, struct r10bio *r10_bio, int *max_sectors)703 static struct md_rdev *read_balance(struct r10conf *conf,
704 struct r10bio *r10_bio,
705 int *max_sectors)
706 {
707 const sector_t this_sector = r10_bio->sector;
708 int disk, slot;
709 int sectors = r10_bio->sectors;
710 int best_good_sectors;
711 sector_t new_distance, best_dist;
712 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
713 int do_balance;
714 int best_dist_slot, best_pending_slot;
715 bool has_nonrot_disk = false;
716 unsigned int min_pending;
717 struct geom *geo = &conf->geo;
718
719 raid10_find_phys(conf, r10_bio);
720 rcu_read_lock();
721 best_dist_slot = -1;
722 min_pending = UINT_MAX;
723 best_dist_rdev = NULL;
724 best_pending_rdev = NULL;
725 best_dist = MaxSector;
726 best_good_sectors = 0;
727 do_balance = 1;
728 clear_bit(R10BIO_FailFast, &r10_bio->state);
729 /*
730 * Check if we can balance. We can balance on the whole
731 * device if no resync is going on (recovery is ok), or below
732 * the resync window. We take the first readable disk when
733 * above the resync window.
734 */
735 if ((conf->mddev->recovery_cp < MaxSector
736 && (this_sector + sectors >= conf->next_resync)) ||
737 (mddev_is_clustered(conf->mddev) &&
738 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
739 this_sector + sectors)))
740 do_balance = 0;
741
742 for (slot = 0; slot < conf->copies ; slot++) {
743 sector_t first_bad;
744 int bad_sectors;
745 sector_t dev_sector;
746 unsigned int pending;
747 bool nonrot;
748
749 if (r10_bio->devs[slot].bio == IO_BLOCKED)
750 continue;
751 disk = r10_bio->devs[slot].devnum;
752 rdev = rcu_dereference(conf->mirrors[disk].replacement);
753 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
754 r10_bio->devs[slot].addr + sectors >
755 rdev->recovery_offset) {
756 /*
757 * Read replacement first to prevent reading both rdev
758 * and replacement as NULL during replacement replace
759 * rdev.
760 */
761 smp_mb();
762 rdev = rcu_dereference(conf->mirrors[disk].rdev);
763 }
764 if (rdev == NULL ||
765 test_bit(Faulty, &rdev->flags))
766 continue;
767 if (!test_bit(In_sync, &rdev->flags) &&
768 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
769 continue;
770
771 dev_sector = r10_bio->devs[slot].addr;
772 if (is_badblock(rdev, dev_sector, sectors,
773 &first_bad, &bad_sectors)) {
774 if (best_dist < MaxSector)
775 /* Already have a better slot */
776 continue;
777 if (first_bad <= dev_sector) {
778 /* Cannot read here. If this is the
779 * 'primary' device, then we must not read
780 * beyond 'bad_sectors' from another device.
781 */
782 bad_sectors -= (dev_sector - first_bad);
783 if (!do_balance && sectors > bad_sectors)
784 sectors = bad_sectors;
785 if (best_good_sectors > sectors)
786 best_good_sectors = sectors;
787 } else {
788 sector_t good_sectors =
789 first_bad - dev_sector;
790 if (good_sectors > best_good_sectors) {
791 best_good_sectors = good_sectors;
792 best_dist_slot = slot;
793 best_dist_rdev = rdev;
794 }
795 if (!do_balance)
796 /* Must read from here */
797 break;
798 }
799 continue;
800 } else
801 best_good_sectors = sectors;
802
803 if (!do_balance)
804 break;
805
806 nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
807 has_nonrot_disk |= nonrot;
808 pending = atomic_read(&rdev->nr_pending);
809 if (min_pending > pending && nonrot) {
810 min_pending = pending;
811 best_pending_slot = slot;
812 best_pending_rdev = rdev;
813 }
814
815 if (best_dist_slot >= 0)
816 /* At least 2 disks to choose from so failfast is OK */
817 set_bit(R10BIO_FailFast, &r10_bio->state);
818 /* This optimisation is debatable, and completely destroys
819 * sequential read speed for 'far copies' arrays. So only
820 * keep it for 'near' arrays, and review those later.
821 */
822 if (geo->near_copies > 1 && !pending)
823 new_distance = 0;
824
825 /* for far > 1 always use the lowest address */
826 else if (geo->far_copies > 1)
827 new_distance = r10_bio->devs[slot].addr;
828 else
829 new_distance = abs(r10_bio->devs[slot].addr -
830 conf->mirrors[disk].head_position);
831
832 if (new_distance < best_dist) {
833 best_dist = new_distance;
834 best_dist_slot = slot;
835 best_dist_rdev = rdev;
836 }
837 }
838 if (slot >= conf->copies) {
839 if (has_nonrot_disk) {
840 slot = best_pending_slot;
841 rdev = best_pending_rdev;
842 } else {
843 slot = best_dist_slot;
844 rdev = best_dist_rdev;
845 }
846 }
847
848 if (slot >= 0) {
849 atomic_inc(&rdev->nr_pending);
850 r10_bio->read_slot = slot;
851 } else
852 rdev = NULL;
853 rcu_read_unlock();
854 *max_sectors = best_good_sectors;
855
856 return rdev;
857 }
858
flush_pending_writes(struct r10conf *conf)859 static void flush_pending_writes(struct r10conf *conf)
860 {
861 /* Any writes that have been queued but are awaiting
862 * bitmap updates get flushed here.
863 */
864 spin_lock_irq(&conf->device_lock);
865
866 if (conf->pending_bio_list.head) {
867 struct blk_plug plug;
868 struct bio *bio;
869
870 bio = bio_list_get(&conf->pending_bio_list);
871 conf->pending_count = 0;
872 spin_unlock_irq(&conf->device_lock);
873
874 /*
875 * As this is called in a wait_event() loop (see freeze_array),
876 * current->state might be TASK_UNINTERRUPTIBLE which will
877 * cause a warning when we prepare to wait again. As it is
878 * rare that this path is taken, it is perfectly safe to force
879 * us to go around the wait_event() loop again, so the warning
880 * is a false-positive. Silence the warning by resetting
881 * thread state
882 */
883 __set_current_state(TASK_RUNNING);
884
885 blk_start_plug(&plug);
886 /* flush any pending bitmap writes to disk
887 * before proceeding w/ I/O */
888 md_bitmap_unplug(conf->mddev->bitmap);
889 wake_up(&conf->wait_barrier);
890
891 while (bio) { /* submit pending writes */
892 struct bio *next = bio->bi_next;
893 struct md_rdev *rdev = (void*)bio->bi_disk;
894 bio->bi_next = NULL;
895 bio_set_dev(bio, rdev->bdev);
896 if (test_bit(Faulty, &rdev->flags)) {
897 bio_io_error(bio);
898 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
899 !blk_queue_discard(bio->bi_disk->queue)))
900 /* Just ignore it */
901 bio_endio(bio);
902 else
903 submit_bio_noacct(bio);
904 bio = next;
905 cond_resched();
906 }
907 blk_finish_plug(&plug);
908 } else
909 spin_unlock_irq(&conf->device_lock);
910 }
911
912 /* Barriers....
913 * Sometimes we need to suspend IO while we do something else,
914 * either some resync/recovery, or reconfigure the array.
915 * To do this we raise a 'barrier'.
916 * The 'barrier' is a counter that can be raised multiple times
917 * to count how many activities are happening which preclude
918 * normal IO.
919 * We can only raise the barrier if there is no pending IO.
920 * i.e. if nr_pending == 0.
921 * We choose only to raise the barrier if no-one is waiting for the
922 * barrier to go down. This means that as soon as an IO request
923 * is ready, no other operations which require a barrier will start
924 * until the IO request has had a chance.
925 *
926 * So: regular IO calls 'wait_barrier'. When that returns there
927 * is no backgroup IO happening, It must arrange to call
928 * allow_barrier when it has finished its IO.
929 * backgroup IO calls must call raise_barrier. Once that returns
930 * there is no normal IO happeing. It must arrange to call
931 * lower_barrier when the particular background IO completes.
932 */
933
raise_barrier(struct r10conf *conf, int force)934 static void raise_barrier(struct r10conf *conf, int force)
935 {
936 BUG_ON(force && !conf->barrier);
937 spin_lock_irq(&conf->resync_lock);
938
939 /* Wait until no block IO is waiting (unless 'force') */
940 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
941 conf->resync_lock);
942
943 /* block any new IO from starting */
944 conf->barrier++;
945
946 /* Now wait for all pending IO to complete */
947 wait_event_lock_irq(conf->wait_barrier,
948 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
949 conf->resync_lock);
950
951 spin_unlock_irq(&conf->resync_lock);
952 }
953
lower_barrier(struct r10conf *conf)954 static void lower_barrier(struct r10conf *conf)
955 {
956 unsigned long flags;
957 spin_lock_irqsave(&conf->resync_lock, flags);
958 conf->barrier--;
959 spin_unlock_irqrestore(&conf->resync_lock, flags);
960 wake_up(&conf->wait_barrier);
961 }
962
wait_barrier(struct r10conf *conf)963 static void wait_barrier(struct r10conf *conf)
964 {
965 spin_lock_irq(&conf->resync_lock);
966 if (conf->barrier) {
967 struct bio_list *bio_list = current->bio_list;
968 conf->nr_waiting++;
969 /* Wait for the barrier to drop.
970 * However if there are already pending
971 * requests (preventing the barrier from
972 * rising completely), and the
973 * pre-process bio queue isn't empty,
974 * then don't wait, as we need to empty
975 * that queue to get the nr_pending
976 * count down.
977 */
978 raid10_log(conf->mddev, "wait barrier");
979 wait_event_lock_irq(conf->wait_barrier,
980 !conf->barrier ||
981 (atomic_read(&conf->nr_pending) &&
982 bio_list &&
983 (!bio_list_empty(&bio_list[0]) ||
984 !bio_list_empty(&bio_list[1]))) ||
985 /* move on if recovery thread is
986 * blocked by us
987 */
988 (conf->mddev->thread->tsk == current &&
989 test_bit(MD_RECOVERY_RUNNING,
990 &conf->mddev->recovery) &&
991 conf->nr_queued > 0),
992 conf->resync_lock);
993 conf->nr_waiting--;
994 if (!conf->nr_waiting)
995 wake_up(&conf->wait_barrier);
996 }
997 atomic_inc(&conf->nr_pending);
998 spin_unlock_irq(&conf->resync_lock);
999 }
1000
allow_barrier(struct r10conf *conf)1001 static void allow_barrier(struct r10conf *conf)
1002 {
1003 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1004 (conf->array_freeze_pending))
1005 wake_up(&conf->wait_barrier);
1006 }
1007
freeze_array(struct r10conf *conf, int extra)1008 static void freeze_array(struct r10conf *conf, int extra)
1009 {
1010 /* stop syncio and normal IO and wait for everything to
1011 * go quiet.
1012 * We increment barrier and nr_waiting, and then
1013 * wait until nr_pending match nr_queued+extra
1014 * This is called in the context of one normal IO request
1015 * that has failed. Thus any sync request that might be pending
1016 * will be blocked by nr_pending, and we need to wait for
1017 * pending IO requests to complete or be queued for re-try.
1018 * Thus the number queued (nr_queued) plus this request (extra)
1019 * must match the number of pending IOs (nr_pending) before
1020 * we continue.
1021 */
1022 spin_lock_irq(&conf->resync_lock);
1023 conf->array_freeze_pending++;
1024 conf->barrier++;
1025 conf->nr_waiting++;
1026 wait_event_lock_irq_cmd(conf->wait_barrier,
1027 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1028 conf->resync_lock,
1029 flush_pending_writes(conf));
1030
1031 conf->array_freeze_pending--;
1032 spin_unlock_irq(&conf->resync_lock);
1033 }
1034
unfreeze_array(struct r10conf *conf)1035 static void unfreeze_array(struct r10conf *conf)
1036 {
1037 /* reverse the effect of the freeze */
1038 spin_lock_irq(&conf->resync_lock);
1039 conf->barrier--;
1040 conf->nr_waiting--;
1041 wake_up(&conf->wait_barrier);
1042 spin_unlock_irq(&conf->resync_lock);
1043 }
1044
choose_data_offset(struct r10bio *r10_bio, struct md_rdev *rdev)1045 static sector_t choose_data_offset(struct r10bio *r10_bio,
1046 struct md_rdev *rdev)
1047 {
1048 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1049 test_bit(R10BIO_Previous, &r10_bio->state))
1050 return rdev->data_offset;
1051 else
1052 return rdev->new_data_offset;
1053 }
1054
1055 struct raid10_plug_cb {
1056 struct blk_plug_cb cb;
1057 struct bio_list pending;
1058 int pending_cnt;
1059 };
1060
raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)1061 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1062 {
1063 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1064 cb);
1065 struct mddev *mddev = plug->cb.data;
1066 struct r10conf *conf = mddev->private;
1067 struct bio *bio;
1068
1069 if (from_schedule || current->bio_list) {
1070 spin_lock_irq(&conf->device_lock);
1071 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1072 conf->pending_count += plug->pending_cnt;
1073 spin_unlock_irq(&conf->device_lock);
1074 wake_up(&conf->wait_barrier);
1075 md_wakeup_thread(mddev->thread);
1076 kfree(plug);
1077 return;
1078 }
1079
1080 /* we aren't scheduling, so we can do the write-out directly. */
1081 bio = bio_list_get(&plug->pending);
1082 md_bitmap_unplug(mddev->bitmap);
1083 wake_up(&conf->wait_barrier);
1084
1085 while (bio) { /* submit pending writes */
1086 struct bio *next = bio->bi_next;
1087 struct md_rdev *rdev = (void*)bio->bi_disk;
1088 bio->bi_next = NULL;
1089 bio_set_dev(bio, rdev->bdev);
1090 if (test_bit(Faulty, &rdev->flags)) {
1091 bio_io_error(bio);
1092 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1093 !blk_queue_discard(bio->bi_disk->queue)))
1094 /* Just ignore it */
1095 bio_endio(bio);
1096 else
1097 submit_bio_noacct(bio);
1098 bio = next;
1099 cond_resched();
1100 }
1101 kfree(plug);
1102 }
1103
1104 /*
1105 * 1. Register the new request and wait if the reconstruction thread has put
1106 * up a bar for new requests. Continue immediately if no resync is active
1107 * currently.
1108 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1109 */
regular_request_wait(struct mddev *mddev, struct r10conf *conf, struct bio *bio, sector_t sectors)1110 static void regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1111 struct bio *bio, sector_t sectors)
1112 {
1113 wait_barrier(conf);
1114 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1115 bio->bi_iter.bi_sector < conf->reshape_progress &&
1116 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1117 raid10_log(conf->mddev, "wait reshape");
1118 allow_barrier(conf);
1119 wait_event(conf->wait_barrier,
1120 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1121 conf->reshape_progress >= bio->bi_iter.bi_sector +
1122 sectors);
1123 wait_barrier(conf);
1124 }
1125 }
1126
raid10_read_request(struct mddev *mddev, struct bio *bio, struct r10bio *r10_bio)1127 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1128 struct r10bio *r10_bio)
1129 {
1130 struct r10conf *conf = mddev->private;
1131 struct bio *read_bio;
1132 const int op = bio_op(bio);
1133 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1134 int max_sectors;
1135 struct md_rdev *rdev;
1136 char b[BDEVNAME_SIZE];
1137 int slot = r10_bio->read_slot;
1138 struct md_rdev *err_rdev = NULL;
1139 gfp_t gfp = GFP_NOIO;
1140
1141 if (slot >= 0 && r10_bio->devs[slot].rdev) {
1142 /*
1143 * This is an error retry, but we cannot
1144 * safely dereference the rdev in the r10_bio,
1145 * we must use the one in conf.
1146 * If it has already been disconnected (unlikely)
1147 * we lose the device name in error messages.
1148 */
1149 int disk;
1150 /*
1151 * As we are blocking raid10, it is a little safer to
1152 * use __GFP_HIGH.
1153 */
1154 gfp = GFP_NOIO | __GFP_HIGH;
1155
1156 rcu_read_lock();
1157 disk = r10_bio->devs[slot].devnum;
1158 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1159 if (err_rdev)
1160 bdevname(err_rdev->bdev, b);
1161 else {
1162 strcpy(b, "???");
1163 /* This never gets dereferenced */
1164 err_rdev = r10_bio->devs[slot].rdev;
1165 }
1166 rcu_read_unlock();
1167 }
1168
1169 regular_request_wait(mddev, conf, bio, r10_bio->sectors);
1170 rdev = read_balance(conf, r10_bio, &max_sectors);
1171 if (!rdev) {
1172 if (err_rdev) {
1173 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1174 mdname(mddev), b,
1175 (unsigned long long)r10_bio->sector);
1176 }
1177 raid_end_bio_io(r10_bio);
1178 return;
1179 }
1180 if (err_rdev)
1181 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1182 mdname(mddev),
1183 bdevname(rdev->bdev, b),
1184 (unsigned long long)r10_bio->sector);
1185 if (max_sectors < bio_sectors(bio)) {
1186 struct bio *split = bio_split(bio, max_sectors,
1187 gfp, &conf->bio_split);
1188 bio_chain(split, bio);
1189 allow_barrier(conf);
1190 submit_bio_noacct(bio);
1191 wait_barrier(conf);
1192 bio = split;
1193 r10_bio->master_bio = bio;
1194 r10_bio->sectors = max_sectors;
1195 }
1196 slot = r10_bio->read_slot;
1197
1198 read_bio = bio_clone_fast(bio, gfp, &mddev->bio_set);
1199
1200 r10_bio->devs[slot].bio = read_bio;
1201 r10_bio->devs[slot].rdev = rdev;
1202
1203 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1204 choose_data_offset(r10_bio, rdev);
1205 bio_set_dev(read_bio, rdev->bdev);
1206 read_bio->bi_end_io = raid10_end_read_request;
1207 bio_set_op_attrs(read_bio, op, do_sync);
1208 if (test_bit(FailFast, &rdev->flags) &&
1209 test_bit(R10BIO_FailFast, &r10_bio->state))
1210 read_bio->bi_opf |= MD_FAILFAST;
1211 read_bio->bi_private = r10_bio;
1212
1213 if (mddev->gendisk)
1214 trace_block_bio_remap(read_bio->bi_disk->queue,
1215 read_bio, disk_devt(mddev->gendisk),
1216 r10_bio->sector);
1217 submit_bio_noacct(read_bio);
1218 return;
1219 }
1220
raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio, struct bio *bio, bool replacement, int n_copy)1221 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1222 struct bio *bio, bool replacement,
1223 int n_copy)
1224 {
1225 const int op = bio_op(bio);
1226 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1227 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1228 unsigned long flags;
1229 struct blk_plug_cb *cb;
1230 struct raid10_plug_cb *plug = NULL;
1231 struct r10conf *conf = mddev->private;
1232 struct md_rdev *rdev;
1233 int devnum = r10_bio->devs[n_copy].devnum;
1234 struct bio *mbio;
1235
1236 if (replacement) {
1237 rdev = conf->mirrors[devnum].replacement;
1238 if (rdev == NULL) {
1239 /* Replacement just got moved to main 'rdev' */
1240 smp_mb();
1241 rdev = conf->mirrors[devnum].rdev;
1242 }
1243 } else
1244 rdev = conf->mirrors[devnum].rdev;
1245
1246 mbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
1247 if (replacement)
1248 r10_bio->devs[n_copy].repl_bio = mbio;
1249 else
1250 r10_bio->devs[n_copy].bio = mbio;
1251
1252 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1253 choose_data_offset(r10_bio, rdev));
1254 bio_set_dev(mbio, rdev->bdev);
1255 mbio->bi_end_io = raid10_end_write_request;
1256 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1257 if (!replacement && test_bit(FailFast,
1258 &conf->mirrors[devnum].rdev->flags)
1259 && enough(conf, devnum))
1260 mbio->bi_opf |= MD_FAILFAST;
1261 mbio->bi_private = r10_bio;
1262
1263 if (conf->mddev->gendisk)
1264 trace_block_bio_remap(mbio->bi_disk->queue,
1265 mbio, disk_devt(conf->mddev->gendisk),
1266 r10_bio->sector);
1267 /* flush_pending_writes() needs access to the rdev so...*/
1268 mbio->bi_disk = (void *)rdev;
1269
1270 atomic_inc(&r10_bio->remaining);
1271
1272 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1273 if (cb)
1274 plug = container_of(cb, struct raid10_plug_cb, cb);
1275 else
1276 plug = NULL;
1277 if (plug) {
1278 bio_list_add(&plug->pending, mbio);
1279 plug->pending_cnt++;
1280 } else {
1281 spin_lock_irqsave(&conf->device_lock, flags);
1282 bio_list_add(&conf->pending_bio_list, mbio);
1283 conf->pending_count++;
1284 spin_unlock_irqrestore(&conf->device_lock, flags);
1285 md_wakeup_thread(mddev->thread);
1286 }
1287 }
1288
raid10_write_request(struct mddev *mddev, struct bio *bio, struct r10bio *r10_bio)1289 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1290 struct r10bio *r10_bio)
1291 {
1292 struct r10conf *conf = mddev->private;
1293 int i;
1294 struct md_rdev *blocked_rdev;
1295 sector_t sectors;
1296 int max_sectors;
1297
1298 if ((mddev_is_clustered(mddev) &&
1299 md_cluster_ops->area_resyncing(mddev, WRITE,
1300 bio->bi_iter.bi_sector,
1301 bio_end_sector(bio)))) {
1302 DEFINE_WAIT(w);
1303 for (;;) {
1304 prepare_to_wait(&conf->wait_barrier,
1305 &w, TASK_IDLE);
1306 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1307 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1308 break;
1309 schedule();
1310 }
1311 finish_wait(&conf->wait_barrier, &w);
1312 }
1313
1314 sectors = r10_bio->sectors;
1315 regular_request_wait(mddev, conf, bio, sectors);
1316 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1317 (mddev->reshape_backwards
1318 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1319 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1320 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1321 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1322 /* Need to update reshape_position in metadata */
1323 mddev->reshape_position = conf->reshape_progress;
1324 set_mask_bits(&mddev->sb_flags, 0,
1325 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1326 md_wakeup_thread(mddev->thread);
1327 raid10_log(conf->mddev, "wait reshape metadata");
1328 wait_event(mddev->sb_wait,
1329 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1330
1331 conf->reshape_safe = mddev->reshape_position;
1332 }
1333
1334 if (conf->pending_count >= max_queued_requests) {
1335 md_wakeup_thread(mddev->thread);
1336 raid10_log(mddev, "wait queued");
1337 wait_event(conf->wait_barrier,
1338 conf->pending_count < max_queued_requests);
1339 }
1340 /* first select target devices under rcu_lock and
1341 * inc refcount on their rdev. Record them by setting
1342 * bios[x] to bio
1343 * If there are known/acknowledged bad blocks on any device
1344 * on which we have seen a write error, we want to avoid
1345 * writing to those blocks. This potentially requires several
1346 * writes to write around the bad blocks. Each set of writes
1347 * gets its own r10_bio with a set of bios attached.
1348 */
1349
1350 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1351 raid10_find_phys(conf, r10_bio);
1352 retry_write:
1353 blocked_rdev = NULL;
1354 rcu_read_lock();
1355 max_sectors = r10_bio->sectors;
1356
1357 for (i = 0; i < conf->copies; i++) {
1358 int d = r10_bio->devs[i].devnum;
1359 struct md_rdev *rdev, *rrdev;
1360
1361 rrdev = rcu_dereference(conf->mirrors[d].replacement);
1362 /*
1363 * Read replacement first to prevent reading both rdev and
1364 * replacement as NULL during replacement replace rdev.
1365 */
1366 smp_mb();
1367 rdev = rcu_dereference(conf->mirrors[d].rdev);
1368 if (rdev == rrdev)
1369 rrdev = NULL;
1370 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1371 atomic_inc(&rdev->nr_pending);
1372 blocked_rdev = rdev;
1373 break;
1374 }
1375 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1376 atomic_inc(&rrdev->nr_pending);
1377 blocked_rdev = rrdev;
1378 break;
1379 }
1380 if (rdev && (test_bit(Faulty, &rdev->flags)))
1381 rdev = NULL;
1382 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1383 rrdev = NULL;
1384
1385 r10_bio->devs[i].bio = NULL;
1386 r10_bio->devs[i].repl_bio = NULL;
1387
1388 if (!rdev && !rrdev) {
1389 set_bit(R10BIO_Degraded, &r10_bio->state);
1390 continue;
1391 }
1392 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1393 sector_t first_bad;
1394 sector_t dev_sector = r10_bio->devs[i].addr;
1395 int bad_sectors;
1396 int is_bad;
1397
1398 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1399 &first_bad, &bad_sectors);
1400 if (is_bad < 0) {
1401 /* Mustn't write here until the bad block
1402 * is acknowledged
1403 */
1404 atomic_inc(&rdev->nr_pending);
1405 set_bit(BlockedBadBlocks, &rdev->flags);
1406 blocked_rdev = rdev;
1407 break;
1408 }
1409 if (is_bad && first_bad <= dev_sector) {
1410 /* Cannot write here at all */
1411 bad_sectors -= (dev_sector - first_bad);
1412 if (bad_sectors < max_sectors)
1413 /* Mustn't write more than bad_sectors
1414 * to other devices yet
1415 */
1416 max_sectors = bad_sectors;
1417 /* We don't set R10BIO_Degraded as that
1418 * only applies if the disk is missing,
1419 * so it might be re-added, and we want to
1420 * know to recover this chunk.
1421 * In this case the device is here, and the
1422 * fact that this chunk is not in-sync is
1423 * recorded in the bad block log.
1424 */
1425 continue;
1426 }
1427 if (is_bad) {
1428 int good_sectors = first_bad - dev_sector;
1429 if (good_sectors < max_sectors)
1430 max_sectors = good_sectors;
1431 }
1432 }
1433 if (rdev) {
1434 r10_bio->devs[i].bio = bio;
1435 atomic_inc(&rdev->nr_pending);
1436 }
1437 if (rrdev) {
1438 r10_bio->devs[i].repl_bio = bio;
1439 atomic_inc(&rrdev->nr_pending);
1440 }
1441 }
1442 rcu_read_unlock();
1443
1444 if (unlikely(blocked_rdev)) {
1445 /* Have to wait for this device to get unblocked, then retry */
1446 int j;
1447 int d;
1448
1449 for (j = 0; j < i; j++) {
1450 if (r10_bio->devs[j].bio) {
1451 d = r10_bio->devs[j].devnum;
1452 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1453 }
1454 if (r10_bio->devs[j].repl_bio) {
1455 struct md_rdev *rdev;
1456 d = r10_bio->devs[j].devnum;
1457 rdev = conf->mirrors[d].replacement;
1458 if (!rdev) {
1459 /* Race with remove_disk */
1460 smp_mb();
1461 rdev = conf->mirrors[d].rdev;
1462 }
1463 rdev_dec_pending(rdev, mddev);
1464 }
1465 }
1466 allow_barrier(conf);
1467 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1468 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1469 wait_barrier(conf);
1470 goto retry_write;
1471 }
1472
1473 if (max_sectors < r10_bio->sectors)
1474 r10_bio->sectors = max_sectors;
1475
1476 if (r10_bio->sectors < bio_sectors(bio)) {
1477 struct bio *split = bio_split(bio, r10_bio->sectors,
1478 GFP_NOIO, &conf->bio_split);
1479 bio_chain(split, bio);
1480 allow_barrier(conf);
1481 submit_bio_noacct(bio);
1482 wait_barrier(conf);
1483 bio = split;
1484 r10_bio->master_bio = bio;
1485 }
1486
1487 atomic_set(&r10_bio->remaining, 1);
1488 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1489
1490 for (i = 0; i < conf->copies; i++) {
1491 if (r10_bio->devs[i].bio)
1492 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1493 if (r10_bio->devs[i].repl_bio)
1494 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1495 }
1496 one_write_done(r10_bio);
1497 }
1498
__make_request(struct mddev *mddev, struct bio *bio, int sectors)1499 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1500 {
1501 struct r10conf *conf = mddev->private;
1502 struct r10bio *r10_bio;
1503
1504 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1505
1506 r10_bio->master_bio = bio;
1507 r10_bio->sectors = sectors;
1508
1509 r10_bio->mddev = mddev;
1510 r10_bio->sector = bio->bi_iter.bi_sector;
1511 r10_bio->state = 0;
1512 r10_bio->read_slot = -1;
1513 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1514
1515 if (bio_data_dir(bio) == READ)
1516 raid10_read_request(mddev, bio, r10_bio);
1517 else
1518 raid10_write_request(mddev, bio, r10_bio);
1519 }
1520
raid10_make_request(struct mddev *mddev, struct bio *bio)1521 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1522 {
1523 struct r10conf *conf = mddev->private;
1524 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1525 int chunk_sects = chunk_mask + 1;
1526 int sectors = bio_sectors(bio);
1527
1528 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1529 && md_flush_request(mddev, bio))
1530 return true;
1531
1532 if (!md_write_start(mddev, bio))
1533 return false;
1534
1535 /*
1536 * If this request crosses a chunk boundary, we need to split
1537 * it.
1538 */
1539 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1540 sectors > chunk_sects
1541 && (conf->geo.near_copies < conf->geo.raid_disks
1542 || conf->prev.near_copies <
1543 conf->prev.raid_disks)))
1544 sectors = chunk_sects -
1545 (bio->bi_iter.bi_sector &
1546 (chunk_sects - 1));
1547 __make_request(mddev, bio, sectors);
1548
1549 /* In case raid10d snuck in to freeze_array */
1550 wake_up(&conf->wait_barrier);
1551 return true;
1552 }
1553
raid10_status(struct seq_file *seq, struct mddev *mddev)1554 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1555 {
1556 struct r10conf *conf = mddev->private;
1557 int i;
1558
1559 if (conf->geo.near_copies < conf->geo.raid_disks)
1560 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1561 if (conf->geo.near_copies > 1)
1562 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1563 if (conf->geo.far_copies > 1) {
1564 if (conf->geo.far_offset)
1565 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1566 else
1567 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1568 if (conf->geo.far_set_size != conf->geo.raid_disks)
1569 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1570 }
1571 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1572 conf->geo.raid_disks - mddev->degraded);
1573 rcu_read_lock();
1574 for (i = 0; i < conf->geo.raid_disks; i++) {
1575 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1576 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1577 }
1578 rcu_read_unlock();
1579 seq_printf(seq, "]");
1580 }
1581
1582 /* check if there are enough drives for
1583 * every block to appear on atleast one.
1584 * Don't consider the device numbered 'ignore'
1585 * as we might be about to remove it.
1586 */
_enough(struct r10conf *conf, int previous, int ignore)1587 static int _enough(struct r10conf *conf, int previous, int ignore)
1588 {
1589 int first = 0;
1590 int has_enough = 0;
1591 int disks, ncopies;
1592 if (previous) {
1593 disks = conf->prev.raid_disks;
1594 ncopies = conf->prev.near_copies;
1595 } else {
1596 disks = conf->geo.raid_disks;
1597 ncopies = conf->geo.near_copies;
1598 }
1599
1600 rcu_read_lock();
1601 do {
1602 int n = conf->copies;
1603 int cnt = 0;
1604 int this = first;
1605 while (n--) {
1606 struct md_rdev *rdev;
1607 if (this != ignore &&
1608 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1609 test_bit(In_sync, &rdev->flags))
1610 cnt++;
1611 this = (this+1) % disks;
1612 }
1613 if (cnt == 0)
1614 goto out;
1615 first = (first + ncopies) % disks;
1616 } while (first != 0);
1617 has_enough = 1;
1618 out:
1619 rcu_read_unlock();
1620 return has_enough;
1621 }
1622
enough(struct r10conf *conf, int ignore)1623 static int enough(struct r10conf *conf, int ignore)
1624 {
1625 /* when calling 'enough', both 'prev' and 'geo' must
1626 * be stable.
1627 * This is ensured if ->reconfig_mutex or ->device_lock
1628 * is held.
1629 */
1630 return _enough(conf, 0, ignore) &&
1631 _enough(conf, 1, ignore);
1632 }
1633
raid10_error(struct mddev *mddev, struct md_rdev *rdev)1634 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1635 {
1636 char b[BDEVNAME_SIZE];
1637 struct r10conf *conf = mddev->private;
1638 unsigned long flags;
1639
1640 /*
1641 * If it is not operational, then we have already marked it as dead
1642 * else if it is the last working disks with "fail_last_dev == false",
1643 * ignore the error, let the next level up know.
1644 * else mark the drive as failed
1645 */
1646 spin_lock_irqsave(&conf->device_lock, flags);
1647 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1648 && !enough(conf, rdev->raid_disk)) {
1649 /*
1650 * Don't fail the drive, just return an IO error.
1651 */
1652 spin_unlock_irqrestore(&conf->device_lock, flags);
1653 return;
1654 }
1655 if (test_and_clear_bit(In_sync, &rdev->flags))
1656 mddev->degraded++;
1657 /*
1658 * If recovery is running, make sure it aborts.
1659 */
1660 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1661 set_bit(Blocked, &rdev->flags);
1662 set_bit(Faulty, &rdev->flags);
1663 set_mask_bits(&mddev->sb_flags, 0,
1664 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1665 spin_unlock_irqrestore(&conf->device_lock, flags);
1666 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1667 "md/raid10:%s: Operation continuing on %d devices.\n",
1668 mdname(mddev), bdevname(rdev->bdev, b),
1669 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1670 }
1671
print_conf(struct r10conf *conf)1672 static void print_conf(struct r10conf *conf)
1673 {
1674 int i;
1675 struct md_rdev *rdev;
1676
1677 pr_debug("RAID10 conf printout:\n");
1678 if (!conf) {
1679 pr_debug("(!conf)\n");
1680 return;
1681 }
1682 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1683 conf->geo.raid_disks);
1684
1685 /* This is only called with ->reconfix_mutex held, so
1686 * rcu protection of rdev is not needed */
1687 for (i = 0; i < conf->geo.raid_disks; i++) {
1688 char b[BDEVNAME_SIZE];
1689 rdev = conf->mirrors[i].rdev;
1690 if (rdev)
1691 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1692 i, !test_bit(In_sync, &rdev->flags),
1693 !test_bit(Faulty, &rdev->flags),
1694 bdevname(rdev->bdev,b));
1695 }
1696 }
1697
close_sync(struct r10conf *conf)1698 static void close_sync(struct r10conf *conf)
1699 {
1700 wait_barrier(conf);
1701 allow_barrier(conf);
1702
1703 mempool_exit(&conf->r10buf_pool);
1704 }
1705
raid10_spare_active(struct mddev *mddev)1706 static int raid10_spare_active(struct mddev *mddev)
1707 {
1708 int i;
1709 struct r10conf *conf = mddev->private;
1710 struct raid10_info *tmp;
1711 int count = 0;
1712 unsigned long flags;
1713
1714 /*
1715 * Find all non-in_sync disks within the RAID10 configuration
1716 * and mark them in_sync
1717 */
1718 for (i = 0; i < conf->geo.raid_disks; i++) {
1719 tmp = conf->mirrors + i;
1720 if (tmp->replacement
1721 && tmp->replacement->recovery_offset == MaxSector
1722 && !test_bit(Faulty, &tmp->replacement->flags)
1723 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1724 /* Replacement has just become active */
1725 if (!tmp->rdev
1726 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1727 count++;
1728 if (tmp->rdev) {
1729 /* Replaced device not technically faulty,
1730 * but we need to be sure it gets removed
1731 * and never re-added.
1732 */
1733 set_bit(Faulty, &tmp->rdev->flags);
1734 sysfs_notify_dirent_safe(
1735 tmp->rdev->sysfs_state);
1736 }
1737 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1738 } else if (tmp->rdev
1739 && tmp->rdev->recovery_offset == MaxSector
1740 && !test_bit(Faulty, &tmp->rdev->flags)
1741 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1742 count++;
1743 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1744 }
1745 }
1746 spin_lock_irqsave(&conf->device_lock, flags);
1747 mddev->degraded -= count;
1748 spin_unlock_irqrestore(&conf->device_lock, flags);
1749
1750 print_conf(conf);
1751 return count;
1752 }
1753
raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)1754 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1755 {
1756 struct r10conf *conf = mddev->private;
1757 int err = -EEXIST;
1758 int mirror;
1759 int first = 0;
1760 int last = conf->geo.raid_disks - 1;
1761
1762 if (mddev->recovery_cp < MaxSector)
1763 /* only hot-add to in-sync arrays, as recovery is
1764 * very different from resync
1765 */
1766 return -EBUSY;
1767 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1768 return -EINVAL;
1769
1770 if (md_integrity_add_rdev(rdev, mddev))
1771 return -ENXIO;
1772
1773 if (rdev->raid_disk >= 0)
1774 first = last = rdev->raid_disk;
1775
1776 if (rdev->saved_raid_disk >= first &&
1777 rdev->saved_raid_disk < conf->geo.raid_disks &&
1778 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1779 mirror = rdev->saved_raid_disk;
1780 else
1781 mirror = first;
1782 for ( ; mirror <= last ; mirror++) {
1783 struct raid10_info *p = &conf->mirrors[mirror];
1784 if (p->recovery_disabled == mddev->recovery_disabled)
1785 continue;
1786 if (p->rdev) {
1787 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1788 p->replacement != NULL)
1789 continue;
1790 clear_bit(In_sync, &rdev->flags);
1791 set_bit(Replacement, &rdev->flags);
1792 rdev->raid_disk = mirror;
1793 err = 0;
1794 if (mddev->gendisk)
1795 disk_stack_limits(mddev->gendisk, rdev->bdev,
1796 rdev->data_offset << 9);
1797 conf->fullsync = 1;
1798 rcu_assign_pointer(p->replacement, rdev);
1799 break;
1800 }
1801
1802 if (mddev->gendisk)
1803 disk_stack_limits(mddev->gendisk, rdev->bdev,
1804 rdev->data_offset << 9);
1805
1806 p->head_position = 0;
1807 p->recovery_disabled = mddev->recovery_disabled - 1;
1808 rdev->raid_disk = mirror;
1809 err = 0;
1810 if (rdev->saved_raid_disk != mirror)
1811 conf->fullsync = 1;
1812 rcu_assign_pointer(p->rdev, rdev);
1813 break;
1814 }
1815 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1816 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1817
1818 print_conf(conf);
1819 return err;
1820 }
1821
raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)1822 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1823 {
1824 struct r10conf *conf = mddev->private;
1825 int err = 0;
1826 int number = rdev->raid_disk;
1827 struct md_rdev **rdevp;
1828 struct raid10_info *p;
1829
1830 print_conf(conf);
1831 if (unlikely(number >= mddev->raid_disks))
1832 return 0;
1833 p = conf->mirrors + number;
1834 if (rdev == p->rdev)
1835 rdevp = &p->rdev;
1836 else if (rdev == p->replacement)
1837 rdevp = &p->replacement;
1838 else
1839 return 0;
1840
1841 if (test_bit(In_sync, &rdev->flags) ||
1842 atomic_read(&rdev->nr_pending)) {
1843 err = -EBUSY;
1844 goto abort;
1845 }
1846 /* Only remove non-faulty devices if recovery
1847 * is not possible.
1848 */
1849 if (!test_bit(Faulty, &rdev->flags) &&
1850 mddev->recovery_disabled != p->recovery_disabled &&
1851 (!p->replacement || p->replacement == rdev) &&
1852 number < conf->geo.raid_disks &&
1853 enough(conf, -1)) {
1854 err = -EBUSY;
1855 goto abort;
1856 }
1857 *rdevp = NULL;
1858 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1859 synchronize_rcu();
1860 if (atomic_read(&rdev->nr_pending)) {
1861 /* lost the race, try later */
1862 err = -EBUSY;
1863 *rdevp = rdev;
1864 goto abort;
1865 }
1866 }
1867 if (p->replacement) {
1868 /* We must have just cleared 'rdev' */
1869 p->rdev = p->replacement;
1870 clear_bit(Replacement, &p->replacement->flags);
1871 smp_mb(); /* Make sure other CPUs may see both as identical
1872 * but will never see neither -- if they are careful.
1873 */
1874 p->replacement = NULL;
1875 }
1876
1877 clear_bit(WantReplacement, &rdev->flags);
1878 err = md_integrity_register(mddev);
1879
1880 abort:
1881
1882 print_conf(conf);
1883 return err;
1884 }
1885
__end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)1886 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1887 {
1888 struct r10conf *conf = r10_bio->mddev->private;
1889
1890 if (!bio->bi_status)
1891 set_bit(R10BIO_Uptodate, &r10_bio->state);
1892 else
1893 /* The write handler will notice the lack of
1894 * R10BIO_Uptodate and record any errors etc
1895 */
1896 atomic_add(r10_bio->sectors,
1897 &conf->mirrors[d].rdev->corrected_errors);
1898
1899 /* for reconstruct, we always reschedule after a read.
1900 * for resync, only after all reads
1901 */
1902 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1903 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1904 atomic_dec_and_test(&r10_bio->remaining)) {
1905 /* we have read all the blocks,
1906 * do the comparison in process context in raid10d
1907 */
1908 reschedule_retry(r10_bio);
1909 }
1910 }
1911
end_sync_read(struct bio *bio)1912 static void end_sync_read(struct bio *bio)
1913 {
1914 struct r10bio *r10_bio = get_resync_r10bio(bio);
1915 struct r10conf *conf = r10_bio->mddev->private;
1916 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1917
1918 __end_sync_read(r10_bio, bio, d);
1919 }
1920
end_reshape_read(struct bio *bio)1921 static void end_reshape_read(struct bio *bio)
1922 {
1923 /* reshape read bio isn't allocated from r10buf_pool */
1924 struct r10bio *r10_bio = bio->bi_private;
1925
1926 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1927 }
1928
end_sync_request(struct r10bio *r10_bio)1929 static void end_sync_request(struct r10bio *r10_bio)
1930 {
1931 struct mddev *mddev = r10_bio->mddev;
1932
1933 while (atomic_dec_and_test(&r10_bio->remaining)) {
1934 if (r10_bio->master_bio == NULL) {
1935 /* the primary of several recovery bios */
1936 sector_t s = r10_bio->sectors;
1937 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1938 test_bit(R10BIO_WriteError, &r10_bio->state))
1939 reschedule_retry(r10_bio);
1940 else
1941 put_buf(r10_bio);
1942 md_done_sync(mddev, s, 1);
1943 break;
1944 } else {
1945 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1946 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1947 test_bit(R10BIO_WriteError, &r10_bio->state))
1948 reschedule_retry(r10_bio);
1949 else
1950 put_buf(r10_bio);
1951 r10_bio = r10_bio2;
1952 }
1953 }
1954 }
1955
end_sync_write(struct bio *bio)1956 static void end_sync_write(struct bio *bio)
1957 {
1958 struct r10bio *r10_bio = get_resync_r10bio(bio);
1959 struct mddev *mddev = r10_bio->mddev;
1960 struct r10conf *conf = mddev->private;
1961 int d;
1962 sector_t first_bad;
1963 int bad_sectors;
1964 int slot;
1965 int repl;
1966 struct md_rdev *rdev = NULL;
1967
1968 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1969 if (repl)
1970 rdev = conf->mirrors[d].replacement;
1971 else
1972 rdev = conf->mirrors[d].rdev;
1973
1974 if (bio->bi_status) {
1975 if (repl)
1976 md_error(mddev, rdev);
1977 else {
1978 set_bit(WriteErrorSeen, &rdev->flags);
1979 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1980 set_bit(MD_RECOVERY_NEEDED,
1981 &rdev->mddev->recovery);
1982 set_bit(R10BIO_WriteError, &r10_bio->state);
1983 }
1984 } else if (is_badblock(rdev,
1985 r10_bio->devs[slot].addr,
1986 r10_bio->sectors,
1987 &first_bad, &bad_sectors))
1988 set_bit(R10BIO_MadeGood, &r10_bio->state);
1989
1990 rdev_dec_pending(rdev, mddev);
1991
1992 end_sync_request(r10_bio);
1993 }
1994
1995 /*
1996 * Note: sync and recover and handled very differently for raid10
1997 * This code is for resync.
1998 * For resync, we read through virtual addresses and read all blocks.
1999 * If there is any error, we schedule a write. The lowest numbered
2000 * drive is authoritative.
2001 * However requests come for physical address, so we need to map.
2002 * For every physical address there are raid_disks/copies virtual addresses,
2003 * which is always are least one, but is not necessarly an integer.
2004 * This means that a physical address can span multiple chunks, so we may
2005 * have to submit multiple io requests for a single sync request.
2006 */
2007 /*
2008 * We check if all blocks are in-sync and only write to blocks that
2009 * aren't in sync
2010 */
sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)2011 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2012 {
2013 struct r10conf *conf = mddev->private;
2014 int i, first;
2015 struct bio *tbio, *fbio;
2016 int vcnt;
2017 struct page **tpages, **fpages;
2018
2019 atomic_set(&r10_bio->remaining, 1);
2020
2021 /* find the first device with a block */
2022 for (i=0; i<conf->copies; i++)
2023 if (!r10_bio->devs[i].bio->bi_status)
2024 break;
2025
2026 if (i == conf->copies)
2027 goto done;
2028
2029 first = i;
2030 fbio = r10_bio->devs[i].bio;
2031 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2032 fbio->bi_iter.bi_idx = 0;
2033 fpages = get_resync_pages(fbio)->pages;
2034
2035 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2036 /* now find blocks with errors */
2037 for (i=0 ; i < conf->copies ; i++) {
2038 int j, d;
2039 struct md_rdev *rdev;
2040 struct resync_pages *rp;
2041
2042 tbio = r10_bio->devs[i].bio;
2043
2044 if (tbio->bi_end_io != end_sync_read)
2045 continue;
2046 if (i == first)
2047 continue;
2048
2049 tpages = get_resync_pages(tbio)->pages;
2050 d = r10_bio->devs[i].devnum;
2051 rdev = conf->mirrors[d].rdev;
2052 if (!r10_bio->devs[i].bio->bi_status) {
2053 /* We know that the bi_io_vec layout is the same for
2054 * both 'first' and 'i', so we just compare them.
2055 * All vec entries are PAGE_SIZE;
2056 */
2057 int sectors = r10_bio->sectors;
2058 for (j = 0; j < vcnt; j++) {
2059 int len = PAGE_SIZE;
2060 if (sectors < (len / 512))
2061 len = sectors * 512;
2062 if (memcmp(page_address(fpages[j]),
2063 page_address(tpages[j]),
2064 len))
2065 break;
2066 sectors -= len/512;
2067 }
2068 if (j == vcnt)
2069 continue;
2070 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2071 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2072 /* Don't fix anything. */
2073 continue;
2074 } else if (test_bit(FailFast, &rdev->flags)) {
2075 /* Just give up on this device */
2076 md_error(rdev->mddev, rdev);
2077 continue;
2078 }
2079 /* Ok, we need to write this bio, either to correct an
2080 * inconsistency or to correct an unreadable block.
2081 * First we need to fixup bv_offset, bv_len and
2082 * bi_vecs, as the read request might have corrupted these
2083 */
2084 rp = get_resync_pages(tbio);
2085 bio_reset(tbio);
2086
2087 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2088
2089 rp->raid_bio = r10_bio;
2090 tbio->bi_private = rp;
2091 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2092 tbio->bi_end_io = end_sync_write;
2093 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2094
2095 bio_copy_data(tbio, fbio);
2096
2097 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2098 atomic_inc(&r10_bio->remaining);
2099 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2100
2101 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2102 tbio->bi_opf |= MD_FAILFAST;
2103 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2104 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2105 submit_bio_noacct(tbio);
2106 }
2107
2108 /* Now write out to any replacement devices
2109 * that are active
2110 */
2111 for (i = 0; i < conf->copies; i++) {
2112 int d;
2113
2114 tbio = r10_bio->devs[i].repl_bio;
2115 if (!tbio || !tbio->bi_end_io)
2116 continue;
2117 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2118 && r10_bio->devs[i].bio != fbio)
2119 bio_copy_data(tbio, fbio);
2120 d = r10_bio->devs[i].devnum;
2121 atomic_inc(&r10_bio->remaining);
2122 md_sync_acct(conf->mirrors[d].replacement->bdev,
2123 bio_sectors(tbio));
2124 submit_bio_noacct(tbio);
2125 }
2126
2127 done:
2128 if (atomic_dec_and_test(&r10_bio->remaining)) {
2129 md_done_sync(mddev, r10_bio->sectors, 1);
2130 put_buf(r10_bio);
2131 }
2132 }
2133
2134 /*
2135 * Now for the recovery code.
2136 * Recovery happens across physical sectors.
2137 * We recover all non-is_sync drives by finding the virtual address of
2138 * each, and then choose a working drive that also has that virt address.
2139 * There is a separate r10_bio for each non-in_sync drive.
2140 * Only the first two slots are in use. The first for reading,
2141 * The second for writing.
2142 *
2143 */
fix_recovery_read_error(struct r10bio *r10_bio)2144 static void fix_recovery_read_error(struct r10bio *r10_bio)
2145 {
2146 /* We got a read error during recovery.
2147 * We repeat the read in smaller page-sized sections.
2148 * If a read succeeds, write it to the new device or record
2149 * a bad block if we cannot.
2150 * If a read fails, record a bad block on both old and
2151 * new devices.
2152 */
2153 struct mddev *mddev = r10_bio->mddev;
2154 struct r10conf *conf = mddev->private;
2155 struct bio *bio = r10_bio->devs[0].bio;
2156 sector_t sect = 0;
2157 int sectors = r10_bio->sectors;
2158 int idx = 0;
2159 int dr = r10_bio->devs[0].devnum;
2160 int dw = r10_bio->devs[1].devnum;
2161 struct page **pages = get_resync_pages(bio)->pages;
2162
2163 while (sectors) {
2164 int s = sectors;
2165 struct md_rdev *rdev;
2166 sector_t addr;
2167 int ok;
2168
2169 if (s > (PAGE_SIZE>>9))
2170 s = PAGE_SIZE >> 9;
2171
2172 rdev = conf->mirrors[dr].rdev;
2173 addr = r10_bio->devs[0].addr + sect,
2174 ok = sync_page_io(rdev,
2175 addr,
2176 s << 9,
2177 pages[idx],
2178 REQ_OP_READ, 0, false);
2179 if (ok) {
2180 rdev = conf->mirrors[dw].rdev;
2181 addr = r10_bio->devs[1].addr + sect;
2182 ok = sync_page_io(rdev,
2183 addr,
2184 s << 9,
2185 pages[idx],
2186 REQ_OP_WRITE, 0, false);
2187 if (!ok) {
2188 set_bit(WriteErrorSeen, &rdev->flags);
2189 if (!test_and_set_bit(WantReplacement,
2190 &rdev->flags))
2191 set_bit(MD_RECOVERY_NEEDED,
2192 &rdev->mddev->recovery);
2193 }
2194 }
2195 if (!ok) {
2196 /* We don't worry if we cannot set a bad block -
2197 * it really is bad so there is no loss in not
2198 * recording it yet
2199 */
2200 rdev_set_badblocks(rdev, addr, s, 0);
2201
2202 if (rdev != conf->mirrors[dw].rdev) {
2203 /* need bad block on destination too */
2204 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2205 addr = r10_bio->devs[1].addr + sect;
2206 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2207 if (!ok) {
2208 /* just abort the recovery */
2209 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2210 mdname(mddev));
2211
2212 conf->mirrors[dw].recovery_disabled
2213 = mddev->recovery_disabled;
2214 set_bit(MD_RECOVERY_INTR,
2215 &mddev->recovery);
2216 break;
2217 }
2218 }
2219 }
2220
2221 sectors -= s;
2222 sect += s;
2223 idx++;
2224 }
2225 }
2226
recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)2227 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2228 {
2229 struct r10conf *conf = mddev->private;
2230 int d;
2231 struct bio *wbio = r10_bio->devs[1].bio;
2232 struct bio *wbio2 = r10_bio->devs[1].repl_bio;
2233
2234 /* Need to test wbio2->bi_end_io before we call
2235 * submit_bio_noacct as if the former is NULL,
2236 * the latter is free to free wbio2.
2237 */
2238 if (wbio2 && !wbio2->bi_end_io)
2239 wbio2 = NULL;
2240
2241 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2242 fix_recovery_read_error(r10_bio);
2243 if (wbio->bi_end_io)
2244 end_sync_request(r10_bio);
2245 if (wbio2)
2246 end_sync_request(r10_bio);
2247 return;
2248 }
2249
2250 /*
2251 * share the pages with the first bio
2252 * and submit the write request
2253 */
2254 d = r10_bio->devs[1].devnum;
2255 if (wbio->bi_end_io) {
2256 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2257 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2258 submit_bio_noacct(wbio);
2259 }
2260 if (wbio2) {
2261 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2262 md_sync_acct(conf->mirrors[d].replacement->bdev,
2263 bio_sectors(wbio2));
2264 submit_bio_noacct(wbio2);
2265 }
2266 }
2267
2268 /*
2269 * Used by fix_read_error() to decay the per rdev read_errors.
2270 * We halve the read error count for every hour that has elapsed
2271 * since the last recorded read error.
2272 *
2273 */
check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)2274 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2275 {
2276 long cur_time_mon;
2277 unsigned long hours_since_last;
2278 unsigned int read_errors = atomic_read(&rdev->read_errors);
2279
2280 cur_time_mon = ktime_get_seconds();
2281
2282 if (rdev->last_read_error == 0) {
2283 /* first time we've seen a read error */
2284 rdev->last_read_error = cur_time_mon;
2285 return;
2286 }
2287
2288 hours_since_last = (long)(cur_time_mon -
2289 rdev->last_read_error) / 3600;
2290
2291 rdev->last_read_error = cur_time_mon;
2292
2293 /*
2294 * if hours_since_last is > the number of bits in read_errors
2295 * just set read errors to 0. We do this to avoid
2296 * overflowing the shift of read_errors by hours_since_last.
2297 */
2298 if (hours_since_last >= 8 * sizeof(read_errors))
2299 atomic_set(&rdev->read_errors, 0);
2300 else
2301 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2302 }
2303
r10_sync_page_io(struct md_rdev *rdev, sector_t sector, int sectors, struct page *page, int rw)2304 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2305 int sectors, struct page *page, int rw)
2306 {
2307 sector_t first_bad;
2308 int bad_sectors;
2309
2310 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2311 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2312 return -1;
2313 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2314 /* success */
2315 return 1;
2316 if (rw == WRITE) {
2317 set_bit(WriteErrorSeen, &rdev->flags);
2318 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2319 set_bit(MD_RECOVERY_NEEDED,
2320 &rdev->mddev->recovery);
2321 }
2322 /* need to record an error - either for the block or the device */
2323 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2324 md_error(rdev->mddev, rdev);
2325 return 0;
2326 }
2327
2328 /*
2329 * This is a kernel thread which:
2330 *
2331 * 1. Retries failed read operations on working mirrors.
2332 * 2. Updates the raid superblock when problems encounter.
2333 * 3. Performs writes following reads for array synchronising.
2334 */
2335
fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)2336 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2337 {
2338 int sect = 0; /* Offset from r10_bio->sector */
2339 int sectors = r10_bio->sectors;
2340 struct md_rdev *rdev;
2341 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2342 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2343
2344 /* still own a reference to this rdev, so it cannot
2345 * have been cleared recently.
2346 */
2347 rdev = conf->mirrors[d].rdev;
2348
2349 if (test_bit(Faulty, &rdev->flags))
2350 /* drive has already been failed, just ignore any
2351 more fix_read_error() attempts */
2352 return;
2353
2354 check_decay_read_errors(mddev, rdev);
2355 atomic_inc(&rdev->read_errors);
2356 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2357 char b[BDEVNAME_SIZE];
2358 bdevname(rdev->bdev, b);
2359
2360 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2361 mdname(mddev), b,
2362 atomic_read(&rdev->read_errors), max_read_errors);
2363 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2364 mdname(mddev), b);
2365 md_error(mddev, rdev);
2366 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2367 return;
2368 }
2369
2370 while(sectors) {
2371 int s = sectors;
2372 int sl = r10_bio->read_slot;
2373 int success = 0;
2374 int start;
2375
2376 if (s > (PAGE_SIZE>>9))
2377 s = PAGE_SIZE >> 9;
2378
2379 rcu_read_lock();
2380 do {
2381 sector_t first_bad;
2382 int bad_sectors;
2383
2384 d = r10_bio->devs[sl].devnum;
2385 rdev = rcu_dereference(conf->mirrors[d].rdev);
2386 if (rdev &&
2387 test_bit(In_sync, &rdev->flags) &&
2388 !test_bit(Faulty, &rdev->flags) &&
2389 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2390 &first_bad, &bad_sectors) == 0) {
2391 atomic_inc(&rdev->nr_pending);
2392 rcu_read_unlock();
2393 success = sync_page_io(rdev,
2394 r10_bio->devs[sl].addr +
2395 sect,
2396 s<<9,
2397 conf->tmppage,
2398 REQ_OP_READ, 0, false);
2399 rdev_dec_pending(rdev, mddev);
2400 rcu_read_lock();
2401 if (success)
2402 break;
2403 }
2404 sl++;
2405 if (sl == conf->copies)
2406 sl = 0;
2407 } while (!success && sl != r10_bio->read_slot);
2408 rcu_read_unlock();
2409
2410 if (!success) {
2411 /* Cannot read from anywhere, just mark the block
2412 * as bad on the first device to discourage future
2413 * reads.
2414 */
2415 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2416 rdev = conf->mirrors[dn].rdev;
2417
2418 if (!rdev_set_badblocks(
2419 rdev,
2420 r10_bio->devs[r10_bio->read_slot].addr
2421 + sect,
2422 s, 0)) {
2423 md_error(mddev, rdev);
2424 r10_bio->devs[r10_bio->read_slot].bio
2425 = IO_BLOCKED;
2426 }
2427 break;
2428 }
2429
2430 start = sl;
2431 /* write it back and re-read */
2432 rcu_read_lock();
2433 while (sl != r10_bio->read_slot) {
2434 char b[BDEVNAME_SIZE];
2435
2436 if (sl==0)
2437 sl = conf->copies;
2438 sl--;
2439 d = r10_bio->devs[sl].devnum;
2440 rdev = rcu_dereference(conf->mirrors[d].rdev);
2441 if (!rdev ||
2442 test_bit(Faulty, &rdev->flags) ||
2443 !test_bit(In_sync, &rdev->flags))
2444 continue;
2445
2446 atomic_inc(&rdev->nr_pending);
2447 rcu_read_unlock();
2448 if (r10_sync_page_io(rdev,
2449 r10_bio->devs[sl].addr +
2450 sect,
2451 s, conf->tmppage, WRITE)
2452 == 0) {
2453 /* Well, this device is dead */
2454 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2455 mdname(mddev), s,
2456 (unsigned long long)(
2457 sect +
2458 choose_data_offset(r10_bio,
2459 rdev)),
2460 bdevname(rdev->bdev, b));
2461 pr_notice("md/raid10:%s: %s: failing drive\n",
2462 mdname(mddev),
2463 bdevname(rdev->bdev, b));
2464 }
2465 rdev_dec_pending(rdev, mddev);
2466 rcu_read_lock();
2467 }
2468 sl = start;
2469 while (sl != r10_bio->read_slot) {
2470 char b[BDEVNAME_SIZE];
2471
2472 if (sl==0)
2473 sl = conf->copies;
2474 sl--;
2475 d = r10_bio->devs[sl].devnum;
2476 rdev = rcu_dereference(conf->mirrors[d].rdev);
2477 if (!rdev ||
2478 test_bit(Faulty, &rdev->flags) ||
2479 !test_bit(In_sync, &rdev->flags))
2480 continue;
2481
2482 atomic_inc(&rdev->nr_pending);
2483 rcu_read_unlock();
2484 switch (r10_sync_page_io(rdev,
2485 r10_bio->devs[sl].addr +
2486 sect,
2487 s, conf->tmppage,
2488 READ)) {
2489 case 0:
2490 /* Well, this device is dead */
2491 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2492 mdname(mddev), s,
2493 (unsigned long long)(
2494 sect +
2495 choose_data_offset(r10_bio, rdev)),
2496 bdevname(rdev->bdev, b));
2497 pr_notice("md/raid10:%s: %s: failing drive\n",
2498 mdname(mddev),
2499 bdevname(rdev->bdev, b));
2500 break;
2501 case 1:
2502 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2503 mdname(mddev), s,
2504 (unsigned long long)(
2505 sect +
2506 choose_data_offset(r10_bio, rdev)),
2507 bdevname(rdev->bdev, b));
2508 atomic_add(s, &rdev->corrected_errors);
2509 }
2510
2511 rdev_dec_pending(rdev, mddev);
2512 rcu_read_lock();
2513 }
2514 rcu_read_unlock();
2515
2516 sectors -= s;
2517 sect += s;
2518 }
2519 }
2520
narrow_write_error(struct r10bio *r10_bio, int i)2521 static int narrow_write_error(struct r10bio *r10_bio, int i)
2522 {
2523 struct bio *bio = r10_bio->master_bio;
2524 struct mddev *mddev = r10_bio->mddev;
2525 struct r10conf *conf = mddev->private;
2526 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2527 /* bio has the data to be written to slot 'i' where
2528 * we just recently had a write error.
2529 * We repeatedly clone the bio and trim down to one block,
2530 * then try the write. Where the write fails we record
2531 * a bad block.
2532 * It is conceivable that the bio doesn't exactly align with
2533 * blocks. We must handle this.
2534 *
2535 * We currently own a reference to the rdev.
2536 */
2537
2538 int block_sectors;
2539 sector_t sector;
2540 int sectors;
2541 int sect_to_write = r10_bio->sectors;
2542 int ok = 1;
2543
2544 if (rdev->badblocks.shift < 0)
2545 return 0;
2546
2547 block_sectors = roundup(1 << rdev->badblocks.shift,
2548 bdev_logical_block_size(rdev->bdev) >> 9);
2549 sector = r10_bio->sector;
2550 sectors = ((r10_bio->sector + block_sectors)
2551 & ~(sector_t)(block_sectors - 1))
2552 - sector;
2553
2554 while (sect_to_write) {
2555 struct bio *wbio;
2556 sector_t wsector;
2557 if (sectors > sect_to_write)
2558 sectors = sect_to_write;
2559 /* Write at 'sector' for 'sectors' */
2560 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2561 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2562 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2563 wbio->bi_iter.bi_sector = wsector +
2564 choose_data_offset(r10_bio, rdev);
2565 bio_set_dev(wbio, rdev->bdev);
2566 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2567
2568 if (submit_bio_wait(wbio) < 0)
2569 /* Failure! */
2570 ok = rdev_set_badblocks(rdev, wsector,
2571 sectors, 0)
2572 && ok;
2573
2574 bio_put(wbio);
2575 sect_to_write -= sectors;
2576 sector += sectors;
2577 sectors = block_sectors;
2578 }
2579 return ok;
2580 }
2581
handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)2582 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2583 {
2584 int slot = r10_bio->read_slot;
2585 struct bio *bio;
2586 struct r10conf *conf = mddev->private;
2587 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2588
2589 /* we got a read error. Maybe the drive is bad. Maybe just
2590 * the block and we can fix it.
2591 * We freeze all other IO, and try reading the block from
2592 * other devices. When we find one, we re-write
2593 * and check it that fixes the read error.
2594 * This is all done synchronously while the array is
2595 * frozen.
2596 */
2597 bio = r10_bio->devs[slot].bio;
2598 bio_put(bio);
2599 r10_bio->devs[slot].bio = NULL;
2600
2601 if (mddev->ro)
2602 r10_bio->devs[slot].bio = IO_BLOCKED;
2603 else if (!test_bit(FailFast, &rdev->flags)) {
2604 freeze_array(conf, 1);
2605 fix_read_error(conf, mddev, r10_bio);
2606 unfreeze_array(conf);
2607 } else
2608 md_error(mddev, rdev);
2609
2610 rdev_dec_pending(rdev, mddev);
2611 allow_barrier(conf);
2612 r10_bio->state = 0;
2613 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2614 }
2615
handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)2616 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2617 {
2618 /* Some sort of write request has finished and it
2619 * succeeded in writing where we thought there was a
2620 * bad block. So forget the bad block.
2621 * Or possibly if failed and we need to record
2622 * a bad block.
2623 */
2624 int m;
2625 struct md_rdev *rdev;
2626
2627 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2628 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2629 for (m = 0; m < conf->copies; m++) {
2630 int dev = r10_bio->devs[m].devnum;
2631 rdev = conf->mirrors[dev].rdev;
2632 if (r10_bio->devs[m].bio == NULL ||
2633 r10_bio->devs[m].bio->bi_end_io == NULL)
2634 continue;
2635 if (!r10_bio->devs[m].bio->bi_status) {
2636 rdev_clear_badblocks(
2637 rdev,
2638 r10_bio->devs[m].addr,
2639 r10_bio->sectors, 0);
2640 } else {
2641 if (!rdev_set_badblocks(
2642 rdev,
2643 r10_bio->devs[m].addr,
2644 r10_bio->sectors, 0))
2645 md_error(conf->mddev, rdev);
2646 }
2647 rdev = conf->mirrors[dev].replacement;
2648 if (r10_bio->devs[m].repl_bio == NULL ||
2649 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2650 continue;
2651
2652 if (!r10_bio->devs[m].repl_bio->bi_status) {
2653 rdev_clear_badblocks(
2654 rdev,
2655 r10_bio->devs[m].addr,
2656 r10_bio->sectors, 0);
2657 } else {
2658 if (!rdev_set_badblocks(
2659 rdev,
2660 r10_bio->devs[m].addr,
2661 r10_bio->sectors, 0))
2662 md_error(conf->mddev, rdev);
2663 }
2664 }
2665 put_buf(r10_bio);
2666 } else {
2667 bool fail = false;
2668 for (m = 0; m < conf->copies; m++) {
2669 int dev = r10_bio->devs[m].devnum;
2670 struct bio *bio = r10_bio->devs[m].bio;
2671 rdev = conf->mirrors[dev].rdev;
2672 if (bio == IO_MADE_GOOD) {
2673 rdev_clear_badblocks(
2674 rdev,
2675 r10_bio->devs[m].addr,
2676 r10_bio->sectors, 0);
2677 rdev_dec_pending(rdev, conf->mddev);
2678 } else if (bio != NULL && bio->bi_status) {
2679 fail = true;
2680 if (!narrow_write_error(r10_bio, m)) {
2681 md_error(conf->mddev, rdev);
2682 set_bit(R10BIO_Degraded,
2683 &r10_bio->state);
2684 }
2685 rdev_dec_pending(rdev, conf->mddev);
2686 }
2687 bio = r10_bio->devs[m].repl_bio;
2688 rdev = conf->mirrors[dev].replacement;
2689 if (rdev && bio == IO_MADE_GOOD) {
2690 rdev_clear_badblocks(
2691 rdev,
2692 r10_bio->devs[m].addr,
2693 r10_bio->sectors, 0);
2694 rdev_dec_pending(rdev, conf->mddev);
2695 }
2696 }
2697 if (fail) {
2698 spin_lock_irq(&conf->device_lock);
2699 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2700 conf->nr_queued++;
2701 spin_unlock_irq(&conf->device_lock);
2702 /*
2703 * In case freeze_array() is waiting for condition
2704 * nr_pending == nr_queued + extra to be true.
2705 */
2706 wake_up(&conf->wait_barrier);
2707 md_wakeup_thread(conf->mddev->thread);
2708 } else {
2709 if (test_bit(R10BIO_WriteError,
2710 &r10_bio->state))
2711 close_write(r10_bio);
2712 raid_end_bio_io(r10_bio);
2713 }
2714 }
2715 }
2716
raid10d(struct md_thread *thread)2717 static void raid10d(struct md_thread *thread)
2718 {
2719 struct mddev *mddev = thread->mddev;
2720 struct r10bio *r10_bio;
2721 unsigned long flags;
2722 struct r10conf *conf = mddev->private;
2723 struct list_head *head = &conf->retry_list;
2724 struct blk_plug plug;
2725
2726 md_check_recovery(mddev);
2727
2728 if (!list_empty_careful(&conf->bio_end_io_list) &&
2729 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2730 LIST_HEAD(tmp);
2731 spin_lock_irqsave(&conf->device_lock, flags);
2732 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2733 while (!list_empty(&conf->bio_end_io_list)) {
2734 list_move(conf->bio_end_io_list.prev, &tmp);
2735 conf->nr_queued--;
2736 }
2737 }
2738 spin_unlock_irqrestore(&conf->device_lock, flags);
2739 while (!list_empty(&tmp)) {
2740 r10_bio = list_first_entry(&tmp, struct r10bio,
2741 retry_list);
2742 list_del(&r10_bio->retry_list);
2743 if (mddev->degraded)
2744 set_bit(R10BIO_Degraded, &r10_bio->state);
2745
2746 if (test_bit(R10BIO_WriteError,
2747 &r10_bio->state))
2748 close_write(r10_bio);
2749 raid_end_bio_io(r10_bio);
2750 }
2751 }
2752
2753 blk_start_plug(&plug);
2754 for (;;) {
2755
2756 flush_pending_writes(conf);
2757
2758 spin_lock_irqsave(&conf->device_lock, flags);
2759 if (list_empty(head)) {
2760 spin_unlock_irqrestore(&conf->device_lock, flags);
2761 break;
2762 }
2763 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2764 list_del(head->prev);
2765 conf->nr_queued--;
2766 spin_unlock_irqrestore(&conf->device_lock, flags);
2767
2768 mddev = r10_bio->mddev;
2769 conf = mddev->private;
2770 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2771 test_bit(R10BIO_WriteError, &r10_bio->state))
2772 handle_write_completed(conf, r10_bio);
2773 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2774 reshape_request_write(mddev, r10_bio);
2775 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2776 sync_request_write(mddev, r10_bio);
2777 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2778 recovery_request_write(mddev, r10_bio);
2779 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2780 handle_read_error(mddev, r10_bio);
2781 else
2782 WARN_ON_ONCE(1);
2783
2784 cond_resched();
2785 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2786 md_check_recovery(mddev);
2787 }
2788 blk_finish_plug(&plug);
2789 }
2790
init_resync(struct r10conf *conf)2791 static int init_resync(struct r10conf *conf)
2792 {
2793 int ret, buffs, i;
2794
2795 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2796 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2797 conf->have_replacement = 0;
2798 for (i = 0; i < conf->geo.raid_disks; i++)
2799 if (conf->mirrors[i].replacement)
2800 conf->have_replacement = 1;
2801 ret = mempool_init(&conf->r10buf_pool, buffs,
2802 r10buf_pool_alloc, r10buf_pool_free, conf);
2803 if (ret)
2804 return ret;
2805 conf->next_resync = 0;
2806 return 0;
2807 }
2808
raid10_alloc_init_r10buf(struct r10conf *conf)2809 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2810 {
2811 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2812 struct rsync_pages *rp;
2813 struct bio *bio;
2814 int nalloc;
2815 int i;
2816
2817 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2818 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2819 nalloc = conf->copies; /* resync */
2820 else
2821 nalloc = 2; /* recovery */
2822
2823 for (i = 0; i < nalloc; i++) {
2824 bio = r10bio->devs[i].bio;
2825 rp = bio->bi_private;
2826 bio_reset(bio);
2827 bio->bi_private = rp;
2828 bio = r10bio->devs[i].repl_bio;
2829 if (bio) {
2830 rp = bio->bi_private;
2831 bio_reset(bio);
2832 bio->bi_private = rp;
2833 }
2834 }
2835 return r10bio;
2836 }
2837
2838 /*
2839 * Set cluster_sync_high since we need other nodes to add the
2840 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2841 */
raid10_set_cluster_sync_high(struct r10conf *conf)2842 static void raid10_set_cluster_sync_high(struct r10conf *conf)
2843 {
2844 sector_t window_size;
2845 int extra_chunk, chunks;
2846
2847 /*
2848 * First, here we define "stripe" as a unit which across
2849 * all member devices one time, so we get chunks by use
2850 * raid_disks / near_copies. Otherwise, if near_copies is
2851 * close to raid_disks, then resync window could increases
2852 * linearly with the increase of raid_disks, which means
2853 * we will suspend a really large IO window while it is not
2854 * necessary. If raid_disks is not divisible by near_copies,
2855 * an extra chunk is needed to ensure the whole "stripe" is
2856 * covered.
2857 */
2858
2859 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2860 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2861 extra_chunk = 0;
2862 else
2863 extra_chunk = 1;
2864 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2865
2866 /*
2867 * At least use a 32M window to align with raid1's resync window
2868 */
2869 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2870 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2871
2872 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2873 }
2874
2875 /*
2876 * perform a "sync" on one "block"
2877 *
2878 * We need to make sure that no normal I/O request - particularly write
2879 * requests - conflict with active sync requests.
2880 *
2881 * This is achieved by tracking pending requests and a 'barrier' concept
2882 * that can be installed to exclude normal IO requests.
2883 *
2884 * Resync and recovery are handled very differently.
2885 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2886 *
2887 * For resync, we iterate over virtual addresses, read all copies,
2888 * and update if there are differences. If only one copy is live,
2889 * skip it.
2890 * For recovery, we iterate over physical addresses, read a good
2891 * value for each non-in_sync drive, and over-write.
2892 *
2893 * So, for recovery we may have several outstanding complex requests for a
2894 * given address, one for each out-of-sync device. We model this by allocating
2895 * a number of r10_bio structures, one for each out-of-sync device.
2896 * As we setup these structures, we collect all bio's together into a list
2897 * which we then process collectively to add pages, and then process again
2898 * to pass to submit_bio_noacct.
2899 *
2900 * The r10_bio structures are linked using a borrowed master_bio pointer.
2901 * This link is counted in ->remaining. When the r10_bio that points to NULL
2902 * has its remaining count decremented to 0, the whole complex operation
2903 * is complete.
2904 *
2905 */
2906
raid10_sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)2907 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2908 int *skipped)
2909 {
2910 struct r10conf *conf = mddev->private;
2911 struct r10bio *r10_bio;
2912 struct bio *biolist = NULL, *bio;
2913 sector_t max_sector, nr_sectors;
2914 int i;
2915 int max_sync;
2916 sector_t sync_blocks;
2917 sector_t sectors_skipped = 0;
2918 int chunks_skipped = 0;
2919 sector_t chunk_mask = conf->geo.chunk_mask;
2920 int page_idx = 0;
2921
2922 /*
2923 * Allow skipping a full rebuild for incremental assembly
2924 * of a clean array, like RAID1 does.
2925 */
2926 if (mddev->bitmap == NULL &&
2927 mddev->recovery_cp == MaxSector &&
2928 mddev->reshape_position == MaxSector &&
2929 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2930 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2931 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2932 conf->fullsync == 0) {
2933 *skipped = 1;
2934 return mddev->dev_sectors - sector_nr;
2935 }
2936
2937 if (!mempool_initialized(&conf->r10buf_pool))
2938 if (init_resync(conf))
2939 return 0;
2940
2941 skipped:
2942 max_sector = mddev->dev_sectors;
2943 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2944 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2945 max_sector = mddev->resync_max_sectors;
2946 if (sector_nr >= max_sector) {
2947 conf->cluster_sync_low = 0;
2948 conf->cluster_sync_high = 0;
2949
2950 /* If we aborted, we need to abort the
2951 * sync on the 'current' bitmap chucks (there can
2952 * be several when recovering multiple devices).
2953 * as we may have started syncing it but not finished.
2954 * We can find the current address in
2955 * mddev->curr_resync, but for recovery,
2956 * we need to convert that to several
2957 * virtual addresses.
2958 */
2959 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2960 end_reshape(conf);
2961 close_sync(conf);
2962 return 0;
2963 }
2964
2965 if (mddev->curr_resync < max_sector) { /* aborted */
2966 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2967 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2968 &sync_blocks, 1);
2969 else for (i = 0; i < conf->geo.raid_disks; i++) {
2970 sector_t sect =
2971 raid10_find_virt(conf, mddev->curr_resync, i);
2972 md_bitmap_end_sync(mddev->bitmap, sect,
2973 &sync_blocks, 1);
2974 }
2975 } else {
2976 /* completed sync */
2977 if ((!mddev->bitmap || conf->fullsync)
2978 && conf->have_replacement
2979 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2980 /* Completed a full sync so the replacements
2981 * are now fully recovered.
2982 */
2983 rcu_read_lock();
2984 for (i = 0; i < conf->geo.raid_disks; i++) {
2985 struct md_rdev *rdev =
2986 rcu_dereference(conf->mirrors[i].replacement);
2987 if (rdev)
2988 rdev->recovery_offset = MaxSector;
2989 }
2990 rcu_read_unlock();
2991 }
2992 conf->fullsync = 0;
2993 }
2994 md_bitmap_close_sync(mddev->bitmap);
2995 close_sync(conf);
2996 *skipped = 1;
2997 return sectors_skipped;
2998 }
2999
3000 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3001 return reshape_request(mddev, sector_nr, skipped);
3002
3003 if (chunks_skipped >= conf->geo.raid_disks) {
3004 /* if there has been nothing to do on any drive,
3005 * then there is nothing to do at all..
3006 */
3007 *skipped = 1;
3008 return (max_sector - sector_nr) + sectors_skipped;
3009 }
3010
3011 if (max_sector > mddev->resync_max)
3012 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3013
3014 /* make sure whole request will fit in a chunk - if chunks
3015 * are meaningful
3016 */
3017 if (conf->geo.near_copies < conf->geo.raid_disks &&
3018 max_sector > (sector_nr | chunk_mask))
3019 max_sector = (sector_nr | chunk_mask) + 1;
3020
3021 /*
3022 * If there is non-resync activity waiting for a turn, then let it
3023 * though before starting on this new sync request.
3024 */
3025 if (conf->nr_waiting)
3026 schedule_timeout_uninterruptible(1);
3027
3028 /* Again, very different code for resync and recovery.
3029 * Both must result in an r10bio with a list of bios that
3030 * have bi_end_io, bi_sector, bi_disk set,
3031 * and bi_private set to the r10bio.
3032 * For recovery, we may actually create several r10bios
3033 * with 2 bios in each, that correspond to the bios in the main one.
3034 * In this case, the subordinate r10bios link back through a
3035 * borrowed master_bio pointer, and the counter in the master
3036 * includes a ref from each subordinate.
3037 */
3038 /* First, we decide what to do and set ->bi_end_io
3039 * To end_sync_read if we want to read, and
3040 * end_sync_write if we will want to write.
3041 */
3042
3043 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3044 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3045 /* recovery... the complicated one */
3046 int j;
3047 r10_bio = NULL;
3048
3049 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3050 int still_degraded;
3051 struct r10bio *rb2;
3052 sector_t sect;
3053 int must_sync;
3054 int any_working;
3055 int need_recover = 0;
3056 struct raid10_info *mirror = &conf->mirrors[i];
3057 struct md_rdev *mrdev, *mreplace;
3058
3059 rcu_read_lock();
3060 mrdev = rcu_dereference(mirror->rdev);
3061 mreplace = rcu_dereference(mirror->replacement);
3062
3063 if (mrdev != NULL &&
3064 !test_bit(Faulty, &mrdev->flags) &&
3065 !test_bit(In_sync, &mrdev->flags))
3066 need_recover = 1;
3067 if (mreplace && test_bit(Faulty, &mreplace->flags))
3068 mreplace = NULL;
3069
3070 if (!need_recover && !mreplace) {
3071 rcu_read_unlock();
3072 continue;
3073 }
3074
3075 still_degraded = 0;
3076 /* want to reconstruct this device */
3077 rb2 = r10_bio;
3078 sect = raid10_find_virt(conf, sector_nr, i);
3079 if (sect >= mddev->resync_max_sectors) {
3080 /* last stripe is not complete - don't
3081 * try to recover this sector.
3082 */
3083 rcu_read_unlock();
3084 continue;
3085 }
3086 /* Unless we are doing a full sync, or a replacement
3087 * we only need to recover the block if it is set in
3088 * the bitmap
3089 */
3090 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3091 &sync_blocks, 1);
3092 if (sync_blocks < max_sync)
3093 max_sync = sync_blocks;
3094 if (!must_sync &&
3095 mreplace == NULL &&
3096 !conf->fullsync) {
3097 /* yep, skip the sync_blocks here, but don't assume
3098 * that there will never be anything to do here
3099 */
3100 chunks_skipped = -1;
3101 rcu_read_unlock();
3102 continue;
3103 }
3104 atomic_inc(&mrdev->nr_pending);
3105 if (mreplace)
3106 atomic_inc(&mreplace->nr_pending);
3107 rcu_read_unlock();
3108
3109 r10_bio = raid10_alloc_init_r10buf(conf);
3110 r10_bio->state = 0;
3111 raise_barrier(conf, rb2 != NULL);
3112 atomic_set(&r10_bio->remaining, 0);
3113
3114 r10_bio->master_bio = (struct bio*)rb2;
3115 if (rb2)
3116 atomic_inc(&rb2->remaining);
3117 r10_bio->mddev = mddev;
3118 set_bit(R10BIO_IsRecover, &r10_bio->state);
3119 r10_bio->sector = sect;
3120
3121 raid10_find_phys(conf, r10_bio);
3122
3123 /* Need to check if the array will still be
3124 * degraded
3125 */
3126 rcu_read_lock();
3127 for (j = 0; j < conf->geo.raid_disks; j++) {
3128 struct md_rdev *rdev = rcu_dereference(
3129 conf->mirrors[j].rdev);
3130 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3131 still_degraded = 1;
3132 break;
3133 }
3134 }
3135
3136 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3137 &sync_blocks, still_degraded);
3138
3139 any_working = 0;
3140 for (j=0; j<conf->copies;j++) {
3141 int k;
3142 int d = r10_bio->devs[j].devnum;
3143 sector_t from_addr, to_addr;
3144 struct md_rdev *rdev =
3145 rcu_dereference(conf->mirrors[d].rdev);
3146 sector_t sector, first_bad;
3147 int bad_sectors;
3148 if (!rdev ||
3149 !test_bit(In_sync, &rdev->flags))
3150 continue;
3151 /* This is where we read from */
3152 any_working = 1;
3153 sector = r10_bio->devs[j].addr;
3154
3155 if (is_badblock(rdev, sector, max_sync,
3156 &first_bad, &bad_sectors)) {
3157 if (first_bad > sector)
3158 max_sync = first_bad - sector;
3159 else {
3160 bad_sectors -= (sector
3161 - first_bad);
3162 if (max_sync > bad_sectors)
3163 max_sync = bad_sectors;
3164 continue;
3165 }
3166 }
3167 bio = r10_bio->devs[0].bio;
3168 bio->bi_next = biolist;
3169 biolist = bio;
3170 bio->bi_end_io = end_sync_read;
3171 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3172 if (test_bit(FailFast, &rdev->flags))
3173 bio->bi_opf |= MD_FAILFAST;
3174 from_addr = r10_bio->devs[j].addr;
3175 bio->bi_iter.bi_sector = from_addr +
3176 rdev->data_offset;
3177 bio_set_dev(bio, rdev->bdev);
3178 atomic_inc(&rdev->nr_pending);
3179 /* and we write to 'i' (if not in_sync) */
3180
3181 for (k=0; k<conf->copies; k++)
3182 if (r10_bio->devs[k].devnum == i)
3183 break;
3184 BUG_ON(k == conf->copies);
3185 to_addr = r10_bio->devs[k].addr;
3186 r10_bio->devs[0].devnum = d;
3187 r10_bio->devs[0].addr = from_addr;
3188 r10_bio->devs[1].devnum = i;
3189 r10_bio->devs[1].addr = to_addr;
3190
3191 if (need_recover) {
3192 bio = r10_bio->devs[1].bio;
3193 bio->bi_next = biolist;
3194 biolist = bio;
3195 bio->bi_end_io = end_sync_write;
3196 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3197 bio->bi_iter.bi_sector = to_addr
3198 + mrdev->data_offset;
3199 bio_set_dev(bio, mrdev->bdev);
3200 atomic_inc(&r10_bio->remaining);
3201 } else
3202 r10_bio->devs[1].bio->bi_end_io = NULL;
3203
3204 /* and maybe write to replacement */
3205 bio = r10_bio->devs[1].repl_bio;
3206 if (bio)
3207 bio->bi_end_io = NULL;
3208 /* Note: if replace is not NULL, then bio
3209 * cannot be NULL as r10buf_pool_alloc will
3210 * have allocated it.
3211 */
3212 if (!mreplace)
3213 break;
3214 bio->bi_next = biolist;
3215 biolist = bio;
3216 bio->bi_end_io = end_sync_write;
3217 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3218 bio->bi_iter.bi_sector = to_addr +
3219 mreplace->data_offset;
3220 bio_set_dev(bio, mreplace->bdev);
3221 atomic_inc(&r10_bio->remaining);
3222 break;
3223 }
3224 rcu_read_unlock();
3225 if (j == conf->copies) {
3226 /* Cannot recover, so abort the recovery or
3227 * record a bad block */
3228 if (any_working) {
3229 /* problem is that there are bad blocks
3230 * on other device(s)
3231 */
3232 int k;
3233 for (k = 0; k < conf->copies; k++)
3234 if (r10_bio->devs[k].devnum == i)
3235 break;
3236 if (!test_bit(In_sync,
3237 &mrdev->flags)
3238 && !rdev_set_badblocks(
3239 mrdev,
3240 r10_bio->devs[k].addr,
3241 max_sync, 0))
3242 any_working = 0;
3243 if (mreplace &&
3244 !rdev_set_badblocks(
3245 mreplace,
3246 r10_bio->devs[k].addr,
3247 max_sync, 0))
3248 any_working = 0;
3249 }
3250 if (!any_working) {
3251 if (!test_and_set_bit(MD_RECOVERY_INTR,
3252 &mddev->recovery))
3253 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3254 mdname(mddev));
3255 mirror->recovery_disabled
3256 = mddev->recovery_disabled;
3257 }
3258 put_buf(r10_bio);
3259 if (rb2)
3260 atomic_dec(&rb2->remaining);
3261 r10_bio = rb2;
3262 rdev_dec_pending(mrdev, mddev);
3263 if (mreplace)
3264 rdev_dec_pending(mreplace, mddev);
3265 break;
3266 }
3267 rdev_dec_pending(mrdev, mddev);
3268 if (mreplace)
3269 rdev_dec_pending(mreplace, mddev);
3270 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3271 /* Only want this if there is elsewhere to
3272 * read from. 'j' is currently the first
3273 * readable copy.
3274 */
3275 int targets = 1;
3276 for (; j < conf->copies; j++) {
3277 int d = r10_bio->devs[j].devnum;
3278 if (conf->mirrors[d].rdev &&
3279 test_bit(In_sync,
3280 &conf->mirrors[d].rdev->flags))
3281 targets++;
3282 }
3283 if (targets == 1)
3284 r10_bio->devs[0].bio->bi_opf
3285 &= ~MD_FAILFAST;
3286 }
3287 }
3288 if (biolist == NULL) {
3289 while (r10_bio) {
3290 struct r10bio *rb2 = r10_bio;
3291 r10_bio = (struct r10bio*) rb2->master_bio;
3292 rb2->master_bio = NULL;
3293 put_buf(rb2);
3294 }
3295 goto giveup;
3296 }
3297 } else {
3298 /* resync. Schedule a read for every block at this virt offset */
3299 int count = 0;
3300
3301 /*
3302 * Since curr_resync_completed could probably not update in
3303 * time, and we will set cluster_sync_low based on it.
3304 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3305 * safety reason, which ensures curr_resync_completed is
3306 * updated in bitmap_cond_end_sync.
3307 */
3308 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3309 mddev_is_clustered(mddev) &&
3310 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3311
3312 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3313 &sync_blocks, mddev->degraded) &&
3314 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3315 &mddev->recovery)) {
3316 /* We can skip this block */
3317 *skipped = 1;
3318 return sync_blocks + sectors_skipped;
3319 }
3320 if (sync_blocks < max_sync)
3321 max_sync = sync_blocks;
3322 r10_bio = raid10_alloc_init_r10buf(conf);
3323 r10_bio->state = 0;
3324
3325 r10_bio->mddev = mddev;
3326 atomic_set(&r10_bio->remaining, 0);
3327 raise_barrier(conf, 0);
3328 conf->next_resync = sector_nr;
3329
3330 r10_bio->master_bio = NULL;
3331 r10_bio->sector = sector_nr;
3332 set_bit(R10BIO_IsSync, &r10_bio->state);
3333 raid10_find_phys(conf, r10_bio);
3334 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3335
3336 for (i = 0; i < conf->copies; i++) {
3337 int d = r10_bio->devs[i].devnum;
3338 sector_t first_bad, sector;
3339 int bad_sectors;
3340 struct md_rdev *rdev;
3341
3342 if (r10_bio->devs[i].repl_bio)
3343 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3344
3345 bio = r10_bio->devs[i].bio;
3346 bio->bi_status = BLK_STS_IOERR;
3347 rcu_read_lock();
3348 rdev = rcu_dereference(conf->mirrors[d].rdev);
3349 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3350 rcu_read_unlock();
3351 continue;
3352 }
3353 sector = r10_bio->devs[i].addr;
3354 if (is_badblock(rdev, sector, max_sync,
3355 &first_bad, &bad_sectors)) {
3356 if (first_bad > sector)
3357 max_sync = first_bad - sector;
3358 else {
3359 bad_sectors -= (sector - first_bad);
3360 if (max_sync > bad_sectors)
3361 max_sync = bad_sectors;
3362 rcu_read_unlock();
3363 continue;
3364 }
3365 }
3366 atomic_inc(&rdev->nr_pending);
3367 atomic_inc(&r10_bio->remaining);
3368 bio->bi_next = biolist;
3369 biolist = bio;
3370 bio->bi_end_io = end_sync_read;
3371 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3372 if (test_bit(FailFast, &rdev->flags))
3373 bio->bi_opf |= MD_FAILFAST;
3374 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3375 bio_set_dev(bio, rdev->bdev);
3376 count++;
3377
3378 rdev = rcu_dereference(conf->mirrors[d].replacement);
3379 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3380 rcu_read_unlock();
3381 continue;
3382 }
3383 atomic_inc(&rdev->nr_pending);
3384
3385 /* Need to set up for writing to the replacement */
3386 bio = r10_bio->devs[i].repl_bio;
3387 bio->bi_status = BLK_STS_IOERR;
3388
3389 sector = r10_bio->devs[i].addr;
3390 bio->bi_next = biolist;
3391 biolist = bio;
3392 bio->bi_end_io = end_sync_write;
3393 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3394 if (test_bit(FailFast, &rdev->flags))
3395 bio->bi_opf |= MD_FAILFAST;
3396 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3397 bio_set_dev(bio, rdev->bdev);
3398 count++;
3399 rcu_read_unlock();
3400 }
3401
3402 if (count < 2) {
3403 for (i=0; i<conf->copies; i++) {
3404 int d = r10_bio->devs[i].devnum;
3405 if (r10_bio->devs[i].bio->bi_end_io)
3406 rdev_dec_pending(conf->mirrors[d].rdev,
3407 mddev);
3408 if (r10_bio->devs[i].repl_bio &&
3409 r10_bio->devs[i].repl_bio->bi_end_io)
3410 rdev_dec_pending(
3411 conf->mirrors[d].replacement,
3412 mddev);
3413 }
3414 put_buf(r10_bio);
3415 biolist = NULL;
3416 goto giveup;
3417 }
3418 }
3419
3420 nr_sectors = 0;
3421 if (sector_nr + max_sync < max_sector)
3422 max_sector = sector_nr + max_sync;
3423 do {
3424 struct page *page;
3425 int len = PAGE_SIZE;
3426 if (sector_nr + (len>>9) > max_sector)
3427 len = (max_sector - sector_nr) << 9;
3428 if (len == 0)
3429 break;
3430 for (bio= biolist ; bio ; bio=bio->bi_next) {
3431 struct resync_pages *rp = get_resync_pages(bio);
3432 page = resync_fetch_page(rp, page_idx);
3433 /*
3434 * won't fail because the vec table is big enough
3435 * to hold all these pages
3436 */
3437 bio_add_page(bio, page, len, 0);
3438 }
3439 nr_sectors += len>>9;
3440 sector_nr += len>>9;
3441 } while (++page_idx < RESYNC_PAGES);
3442 r10_bio->sectors = nr_sectors;
3443
3444 if (mddev_is_clustered(mddev) &&
3445 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3446 /* It is resync not recovery */
3447 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3448 conf->cluster_sync_low = mddev->curr_resync_completed;
3449 raid10_set_cluster_sync_high(conf);
3450 /* Send resync message */
3451 md_cluster_ops->resync_info_update(mddev,
3452 conf->cluster_sync_low,
3453 conf->cluster_sync_high);
3454 }
3455 } else if (mddev_is_clustered(mddev)) {
3456 /* This is recovery not resync */
3457 sector_t sect_va1, sect_va2;
3458 bool broadcast_msg = false;
3459
3460 for (i = 0; i < conf->geo.raid_disks; i++) {
3461 /*
3462 * sector_nr is a device address for recovery, so we
3463 * need translate it to array address before compare
3464 * with cluster_sync_high.
3465 */
3466 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3467
3468 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3469 broadcast_msg = true;
3470 /*
3471 * curr_resync_completed is similar as
3472 * sector_nr, so make the translation too.
3473 */
3474 sect_va2 = raid10_find_virt(conf,
3475 mddev->curr_resync_completed, i);
3476
3477 if (conf->cluster_sync_low == 0 ||
3478 conf->cluster_sync_low > sect_va2)
3479 conf->cluster_sync_low = sect_va2;
3480 }
3481 }
3482 if (broadcast_msg) {
3483 raid10_set_cluster_sync_high(conf);
3484 md_cluster_ops->resync_info_update(mddev,
3485 conf->cluster_sync_low,
3486 conf->cluster_sync_high);
3487 }
3488 }
3489
3490 while (biolist) {
3491 bio = biolist;
3492 biolist = biolist->bi_next;
3493
3494 bio->bi_next = NULL;
3495 r10_bio = get_resync_r10bio(bio);
3496 r10_bio->sectors = nr_sectors;
3497
3498 if (bio->bi_end_io == end_sync_read) {
3499 md_sync_acct_bio(bio, nr_sectors);
3500 bio->bi_status = 0;
3501 submit_bio_noacct(bio);
3502 }
3503 }
3504
3505 if (sectors_skipped)
3506 /* pretend they weren't skipped, it makes
3507 * no important difference in this case
3508 */
3509 md_done_sync(mddev, sectors_skipped, 1);
3510
3511 return sectors_skipped + nr_sectors;
3512 giveup:
3513 /* There is nowhere to write, so all non-sync
3514 * drives must be failed or in resync, all drives
3515 * have a bad block, so try the next chunk...
3516 */
3517 if (sector_nr + max_sync < max_sector)
3518 max_sector = sector_nr + max_sync;
3519
3520 sectors_skipped += (max_sector - sector_nr);
3521 chunks_skipped ++;
3522 sector_nr = max_sector;
3523 goto skipped;
3524 }
3525
3526 static sector_t
raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)3527 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3528 {
3529 sector_t size;
3530 struct r10conf *conf = mddev->private;
3531
3532 if (!raid_disks)
3533 raid_disks = min(conf->geo.raid_disks,
3534 conf->prev.raid_disks);
3535 if (!sectors)
3536 sectors = conf->dev_sectors;
3537
3538 size = sectors >> conf->geo.chunk_shift;
3539 sector_div(size, conf->geo.far_copies);
3540 size = size * raid_disks;
3541 sector_div(size, conf->geo.near_copies);
3542
3543 return size << conf->geo.chunk_shift;
3544 }
3545
calc_sectors(struct r10conf *conf, sector_t size)3546 static void calc_sectors(struct r10conf *conf, sector_t size)
3547 {
3548 /* Calculate the number of sectors-per-device that will
3549 * actually be used, and set conf->dev_sectors and
3550 * conf->stride
3551 */
3552
3553 size = size >> conf->geo.chunk_shift;
3554 sector_div(size, conf->geo.far_copies);
3555 size = size * conf->geo.raid_disks;
3556 sector_div(size, conf->geo.near_copies);
3557 /* 'size' is now the number of chunks in the array */
3558 /* calculate "used chunks per device" */
3559 size = size * conf->copies;
3560
3561 /* We need to round up when dividing by raid_disks to
3562 * get the stride size.
3563 */
3564 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3565
3566 conf->dev_sectors = size << conf->geo.chunk_shift;
3567
3568 if (conf->geo.far_offset)
3569 conf->geo.stride = 1 << conf->geo.chunk_shift;
3570 else {
3571 sector_div(size, conf->geo.far_copies);
3572 conf->geo.stride = size << conf->geo.chunk_shift;
3573 }
3574 }
3575
3576 enum geo_type {geo_new, geo_old, geo_start};
setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)3577 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3578 {
3579 int nc, fc, fo;
3580 int layout, chunk, disks;
3581 switch (new) {
3582 case geo_old:
3583 layout = mddev->layout;
3584 chunk = mddev->chunk_sectors;
3585 disks = mddev->raid_disks - mddev->delta_disks;
3586 break;
3587 case geo_new:
3588 layout = mddev->new_layout;
3589 chunk = mddev->new_chunk_sectors;
3590 disks = mddev->raid_disks;
3591 break;
3592 default: /* avoid 'may be unused' warnings */
3593 case geo_start: /* new when starting reshape - raid_disks not
3594 * updated yet. */
3595 layout = mddev->new_layout;
3596 chunk = mddev->new_chunk_sectors;
3597 disks = mddev->raid_disks + mddev->delta_disks;
3598 break;
3599 }
3600 if (layout >> 19)
3601 return -1;
3602 if (chunk < (PAGE_SIZE >> 9) ||
3603 !is_power_of_2(chunk))
3604 return -2;
3605 nc = layout & 255;
3606 fc = (layout >> 8) & 255;
3607 fo = layout & (1<<16);
3608 geo->raid_disks = disks;
3609 geo->near_copies = nc;
3610 geo->far_copies = fc;
3611 geo->far_offset = fo;
3612 switch (layout >> 17) {
3613 case 0: /* original layout. simple but not always optimal */
3614 geo->far_set_size = disks;
3615 break;
3616 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3617 * actually using this, but leave code here just in case.*/
3618 geo->far_set_size = disks/fc;
3619 WARN(geo->far_set_size < fc,
3620 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3621 break;
3622 case 2: /* "improved" layout fixed to match documentation */
3623 geo->far_set_size = fc * nc;
3624 break;
3625 default: /* Not a valid layout */
3626 return -1;
3627 }
3628 geo->chunk_mask = chunk - 1;
3629 geo->chunk_shift = ffz(~chunk);
3630 return nc*fc;
3631 }
3632
raid10_free_conf(struct r10conf *conf)3633 static void raid10_free_conf(struct r10conf *conf)
3634 {
3635 if (!conf)
3636 return;
3637
3638 mempool_exit(&conf->r10bio_pool);
3639 kfree(conf->mirrors);
3640 kfree(conf->mirrors_old);
3641 kfree(conf->mirrors_new);
3642 safe_put_page(conf->tmppage);
3643 bioset_exit(&conf->bio_split);
3644 kfree(conf);
3645 }
3646
setup_conf(struct mddev *mddev)3647 static struct r10conf *setup_conf(struct mddev *mddev)
3648 {
3649 struct r10conf *conf = NULL;
3650 int err = -EINVAL;
3651 struct geom geo;
3652 int copies;
3653
3654 copies = setup_geo(&geo, mddev, geo_new);
3655
3656 if (copies == -2) {
3657 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3658 mdname(mddev), PAGE_SIZE);
3659 goto out;
3660 }
3661
3662 if (copies < 2 || copies > mddev->raid_disks) {
3663 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3664 mdname(mddev), mddev->new_layout);
3665 goto out;
3666 }
3667
3668 err = -ENOMEM;
3669 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3670 if (!conf)
3671 goto out;
3672
3673 /* FIXME calc properly */
3674 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3675 sizeof(struct raid10_info),
3676 GFP_KERNEL);
3677 if (!conf->mirrors)
3678 goto out;
3679
3680 conf->tmppage = alloc_page(GFP_KERNEL);
3681 if (!conf->tmppage)
3682 goto out;
3683
3684 conf->geo = geo;
3685 conf->copies = copies;
3686 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3687 rbio_pool_free, conf);
3688 if (err)
3689 goto out;
3690
3691 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3692 if (err)
3693 goto out;
3694
3695 calc_sectors(conf, mddev->dev_sectors);
3696 if (mddev->reshape_position == MaxSector) {
3697 conf->prev = conf->geo;
3698 conf->reshape_progress = MaxSector;
3699 } else {
3700 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3701 err = -EINVAL;
3702 goto out;
3703 }
3704 conf->reshape_progress = mddev->reshape_position;
3705 if (conf->prev.far_offset)
3706 conf->prev.stride = 1 << conf->prev.chunk_shift;
3707 else
3708 /* far_copies must be 1 */
3709 conf->prev.stride = conf->dev_sectors;
3710 }
3711 conf->reshape_safe = conf->reshape_progress;
3712 spin_lock_init(&conf->device_lock);
3713 INIT_LIST_HEAD(&conf->retry_list);
3714 INIT_LIST_HEAD(&conf->bio_end_io_list);
3715
3716 spin_lock_init(&conf->resync_lock);
3717 init_waitqueue_head(&conf->wait_barrier);
3718 atomic_set(&conf->nr_pending, 0);
3719
3720 err = -ENOMEM;
3721 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3722 if (!conf->thread)
3723 goto out;
3724
3725 conf->mddev = mddev;
3726 return conf;
3727
3728 out:
3729 raid10_free_conf(conf);
3730 return ERR_PTR(err);
3731 }
3732
raid10_set_io_opt(struct r10conf *conf)3733 static void raid10_set_io_opt(struct r10conf *conf)
3734 {
3735 int raid_disks = conf->geo.raid_disks;
3736
3737 if (!(conf->geo.raid_disks % conf->geo.near_copies))
3738 raid_disks /= conf->geo.near_copies;
3739 blk_queue_io_opt(conf->mddev->queue, (conf->mddev->chunk_sectors << 9) *
3740 raid_disks);
3741 }
3742
raid10_run(struct mddev *mddev)3743 static int raid10_run(struct mddev *mddev)
3744 {
3745 struct r10conf *conf;
3746 int i, disk_idx;
3747 struct raid10_info *disk;
3748 struct md_rdev *rdev;
3749 sector_t size;
3750 sector_t min_offset_diff = 0;
3751 int first = 1;
3752 bool discard_supported = false;
3753
3754 if (mddev_init_writes_pending(mddev) < 0)
3755 return -ENOMEM;
3756
3757 if (mddev->private == NULL) {
3758 conf = setup_conf(mddev);
3759 if (IS_ERR(conf))
3760 return PTR_ERR(conf);
3761 mddev->private = conf;
3762 }
3763 conf = mddev->private;
3764 if (!conf)
3765 goto out;
3766
3767 mddev->thread = conf->thread;
3768 conf->thread = NULL;
3769
3770 if (mddev_is_clustered(conf->mddev)) {
3771 int fc, fo;
3772
3773 fc = (mddev->layout >> 8) & 255;
3774 fo = mddev->layout & (1<<16);
3775 if (fc > 1 || fo > 0) {
3776 pr_err("only near layout is supported by clustered"
3777 " raid10\n");
3778 goto out_free_conf;
3779 }
3780 }
3781
3782 if (mddev->queue) {
3783 blk_queue_max_discard_sectors(mddev->queue,
3784 mddev->chunk_sectors);
3785 blk_queue_max_write_same_sectors(mddev->queue, 0);
3786 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3787 blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9);
3788 raid10_set_io_opt(conf);
3789 }
3790
3791 rdev_for_each(rdev, mddev) {
3792 long long diff;
3793
3794 disk_idx = rdev->raid_disk;
3795 if (disk_idx < 0)
3796 continue;
3797 if (disk_idx >= conf->geo.raid_disks &&
3798 disk_idx >= conf->prev.raid_disks)
3799 continue;
3800 disk = conf->mirrors + disk_idx;
3801
3802 if (test_bit(Replacement, &rdev->flags)) {
3803 if (disk->replacement)
3804 goto out_free_conf;
3805 disk->replacement = rdev;
3806 } else {
3807 if (disk->rdev)
3808 goto out_free_conf;
3809 disk->rdev = rdev;
3810 }
3811 diff = (rdev->new_data_offset - rdev->data_offset);
3812 if (!mddev->reshape_backwards)
3813 diff = -diff;
3814 if (diff < 0)
3815 diff = 0;
3816 if (first || diff < min_offset_diff)
3817 min_offset_diff = diff;
3818
3819 if (mddev->gendisk)
3820 disk_stack_limits(mddev->gendisk, rdev->bdev,
3821 rdev->data_offset << 9);
3822
3823 disk->head_position = 0;
3824
3825 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3826 discard_supported = true;
3827 first = 0;
3828 }
3829
3830 if (mddev->queue) {
3831 if (discard_supported)
3832 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3833 mddev->queue);
3834 else
3835 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3836 mddev->queue);
3837 }
3838 /* need to check that every block has at least one working mirror */
3839 if (!enough(conf, -1)) {
3840 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3841 mdname(mddev));
3842 goto out_free_conf;
3843 }
3844
3845 if (conf->reshape_progress != MaxSector) {
3846 /* must ensure that shape change is supported */
3847 if (conf->geo.far_copies != 1 &&
3848 conf->geo.far_offset == 0)
3849 goto out_free_conf;
3850 if (conf->prev.far_copies != 1 &&
3851 conf->prev.far_offset == 0)
3852 goto out_free_conf;
3853 }
3854
3855 mddev->degraded = 0;
3856 for (i = 0;
3857 i < conf->geo.raid_disks
3858 || i < conf->prev.raid_disks;
3859 i++) {
3860
3861 disk = conf->mirrors + i;
3862
3863 if (!disk->rdev && disk->replacement) {
3864 /* The replacement is all we have - use it */
3865 disk->rdev = disk->replacement;
3866 disk->replacement = NULL;
3867 clear_bit(Replacement, &disk->rdev->flags);
3868 }
3869
3870 if (!disk->rdev ||
3871 !test_bit(In_sync, &disk->rdev->flags)) {
3872 disk->head_position = 0;
3873 mddev->degraded++;
3874 if (disk->rdev &&
3875 disk->rdev->saved_raid_disk < 0)
3876 conf->fullsync = 1;
3877 }
3878
3879 if (disk->replacement &&
3880 !test_bit(In_sync, &disk->replacement->flags) &&
3881 disk->replacement->saved_raid_disk < 0) {
3882 conf->fullsync = 1;
3883 }
3884
3885 disk->recovery_disabled = mddev->recovery_disabled - 1;
3886 }
3887
3888 if (mddev->recovery_cp != MaxSector)
3889 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3890 mdname(mddev));
3891 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3892 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3893 conf->geo.raid_disks);
3894 /*
3895 * Ok, everything is just fine now
3896 */
3897 mddev->dev_sectors = conf->dev_sectors;
3898 size = raid10_size(mddev, 0, 0);
3899 md_set_array_sectors(mddev, size);
3900 mddev->resync_max_sectors = size;
3901 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3902
3903 if (md_integrity_register(mddev))
3904 goto out_free_conf;
3905
3906 if (conf->reshape_progress != MaxSector) {
3907 unsigned long before_length, after_length;
3908
3909 before_length = ((1 << conf->prev.chunk_shift) *
3910 conf->prev.far_copies);
3911 after_length = ((1 << conf->geo.chunk_shift) *
3912 conf->geo.far_copies);
3913
3914 if (max(before_length, after_length) > min_offset_diff) {
3915 /* This cannot work */
3916 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3917 goto out_free_conf;
3918 }
3919 conf->offset_diff = min_offset_diff;
3920
3921 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3922 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3923 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3924 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3925 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3926 "reshape");
3927 if (!mddev->sync_thread)
3928 goto out_free_conf;
3929 }
3930
3931 return 0;
3932
3933 out_free_conf:
3934 md_unregister_thread(&mddev->thread);
3935 raid10_free_conf(conf);
3936 mddev->private = NULL;
3937 out:
3938 return -EIO;
3939 }
3940
raid10_free(struct mddev *mddev, void *priv)3941 static void raid10_free(struct mddev *mddev, void *priv)
3942 {
3943 raid10_free_conf(priv);
3944 }
3945
raid10_quiesce(struct mddev *mddev, int quiesce)3946 static void raid10_quiesce(struct mddev *mddev, int quiesce)
3947 {
3948 struct r10conf *conf = mddev->private;
3949
3950 if (quiesce)
3951 raise_barrier(conf, 0);
3952 else
3953 lower_barrier(conf);
3954 }
3955
raid10_resize(struct mddev *mddev, sector_t sectors)3956 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3957 {
3958 /* Resize of 'far' arrays is not supported.
3959 * For 'near' and 'offset' arrays we can set the
3960 * number of sectors used to be an appropriate multiple
3961 * of the chunk size.
3962 * For 'offset', this is far_copies*chunksize.
3963 * For 'near' the multiplier is the LCM of
3964 * near_copies and raid_disks.
3965 * So if far_copies > 1 && !far_offset, fail.
3966 * Else find LCM(raid_disks, near_copy)*far_copies and
3967 * multiply by chunk_size. Then round to this number.
3968 * This is mostly done by raid10_size()
3969 */
3970 struct r10conf *conf = mddev->private;
3971 sector_t oldsize, size;
3972
3973 if (mddev->reshape_position != MaxSector)
3974 return -EBUSY;
3975
3976 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3977 return -EINVAL;
3978
3979 oldsize = raid10_size(mddev, 0, 0);
3980 size = raid10_size(mddev, sectors, 0);
3981 if (mddev->external_size &&
3982 mddev->array_sectors > size)
3983 return -EINVAL;
3984 if (mddev->bitmap) {
3985 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
3986 if (ret)
3987 return ret;
3988 }
3989 md_set_array_sectors(mddev, size);
3990 if (sectors > mddev->dev_sectors &&
3991 mddev->recovery_cp > oldsize) {
3992 mddev->recovery_cp = oldsize;
3993 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3994 }
3995 calc_sectors(conf, sectors);
3996 mddev->dev_sectors = conf->dev_sectors;
3997 mddev->resync_max_sectors = size;
3998 return 0;
3999 }
4000
raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)4001 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4002 {
4003 struct md_rdev *rdev;
4004 struct r10conf *conf;
4005
4006 if (mddev->degraded > 0) {
4007 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4008 mdname(mddev));
4009 return ERR_PTR(-EINVAL);
4010 }
4011 sector_div(size, devs);
4012
4013 /* Set new parameters */
4014 mddev->new_level = 10;
4015 /* new layout: far_copies = 1, near_copies = 2 */
4016 mddev->new_layout = (1<<8) + 2;
4017 mddev->new_chunk_sectors = mddev->chunk_sectors;
4018 mddev->delta_disks = mddev->raid_disks;
4019 mddev->raid_disks *= 2;
4020 /* make sure it will be not marked as dirty */
4021 mddev->recovery_cp = MaxSector;
4022 mddev->dev_sectors = size;
4023
4024 conf = setup_conf(mddev);
4025 if (!IS_ERR(conf)) {
4026 rdev_for_each(rdev, mddev)
4027 if (rdev->raid_disk >= 0) {
4028 rdev->new_raid_disk = rdev->raid_disk * 2;
4029 rdev->sectors = size;
4030 }
4031 conf->barrier = 1;
4032 }
4033
4034 return conf;
4035 }
4036
raid10_takeover(struct mddev *mddev)4037 static void *raid10_takeover(struct mddev *mddev)
4038 {
4039 struct r0conf *raid0_conf;
4040
4041 /* raid10 can take over:
4042 * raid0 - providing it has only two drives
4043 */
4044 if (mddev->level == 0) {
4045 /* for raid0 takeover only one zone is supported */
4046 raid0_conf = mddev->private;
4047 if (raid0_conf->nr_strip_zones > 1) {
4048 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4049 mdname(mddev));
4050 return ERR_PTR(-EINVAL);
4051 }
4052 return raid10_takeover_raid0(mddev,
4053 raid0_conf->strip_zone->zone_end,
4054 raid0_conf->strip_zone->nb_dev);
4055 }
4056 return ERR_PTR(-EINVAL);
4057 }
4058
raid10_check_reshape(struct mddev *mddev)4059 static int raid10_check_reshape(struct mddev *mddev)
4060 {
4061 /* Called when there is a request to change
4062 * - layout (to ->new_layout)
4063 * - chunk size (to ->new_chunk_sectors)
4064 * - raid_disks (by delta_disks)
4065 * or when trying to restart a reshape that was ongoing.
4066 *
4067 * We need to validate the request and possibly allocate
4068 * space if that might be an issue later.
4069 *
4070 * Currently we reject any reshape of a 'far' mode array,
4071 * allow chunk size to change if new is generally acceptable,
4072 * allow raid_disks to increase, and allow
4073 * a switch between 'near' mode and 'offset' mode.
4074 */
4075 struct r10conf *conf = mddev->private;
4076 struct geom geo;
4077
4078 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4079 return -EINVAL;
4080
4081 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4082 /* mustn't change number of copies */
4083 return -EINVAL;
4084 if (geo.far_copies > 1 && !geo.far_offset)
4085 /* Cannot switch to 'far' mode */
4086 return -EINVAL;
4087
4088 if (mddev->array_sectors & geo.chunk_mask)
4089 /* not factor of array size */
4090 return -EINVAL;
4091
4092 if (!enough(conf, -1))
4093 return -EINVAL;
4094
4095 kfree(conf->mirrors_new);
4096 conf->mirrors_new = NULL;
4097 if (mddev->delta_disks > 0) {
4098 /* allocate new 'mirrors' list */
4099 conf->mirrors_new =
4100 kcalloc(mddev->raid_disks + mddev->delta_disks,
4101 sizeof(struct raid10_info),
4102 GFP_KERNEL);
4103 if (!conf->mirrors_new)
4104 return -ENOMEM;
4105 }
4106 return 0;
4107 }
4108
4109 /*
4110 * Need to check if array has failed when deciding whether to:
4111 * - start an array
4112 * - remove non-faulty devices
4113 * - add a spare
4114 * - allow a reshape
4115 * This determination is simple when no reshape is happening.
4116 * However if there is a reshape, we need to carefully check
4117 * both the before and after sections.
4118 * This is because some failed devices may only affect one
4119 * of the two sections, and some non-in_sync devices may
4120 * be insync in the section most affected by failed devices.
4121 */
calc_degraded(struct r10conf *conf)4122 static int calc_degraded(struct r10conf *conf)
4123 {
4124 int degraded, degraded2;
4125 int i;
4126
4127 rcu_read_lock();
4128 degraded = 0;
4129 /* 'prev' section first */
4130 for (i = 0; i < conf->prev.raid_disks; i++) {
4131 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4132 if (!rdev || test_bit(Faulty, &rdev->flags))
4133 degraded++;
4134 else if (!test_bit(In_sync, &rdev->flags))
4135 /* When we can reduce the number of devices in
4136 * an array, this might not contribute to
4137 * 'degraded'. It does now.
4138 */
4139 degraded++;
4140 }
4141 rcu_read_unlock();
4142 if (conf->geo.raid_disks == conf->prev.raid_disks)
4143 return degraded;
4144 rcu_read_lock();
4145 degraded2 = 0;
4146 for (i = 0; i < conf->geo.raid_disks; i++) {
4147 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4148 if (!rdev || test_bit(Faulty, &rdev->flags))
4149 degraded2++;
4150 else if (!test_bit(In_sync, &rdev->flags)) {
4151 /* If reshape is increasing the number of devices,
4152 * this section has already been recovered, so
4153 * it doesn't contribute to degraded.
4154 * else it does.
4155 */
4156 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4157 degraded2++;
4158 }
4159 }
4160 rcu_read_unlock();
4161 if (degraded2 > degraded)
4162 return degraded2;
4163 return degraded;
4164 }
4165
raid10_start_reshape(struct mddev *mddev)4166 static int raid10_start_reshape(struct mddev *mddev)
4167 {
4168 /* A 'reshape' has been requested. This commits
4169 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4170 * This also checks if there are enough spares and adds them
4171 * to the array.
4172 * We currently require enough spares to make the final
4173 * array non-degraded. We also require that the difference
4174 * between old and new data_offset - on each device - is
4175 * enough that we never risk over-writing.
4176 */
4177
4178 unsigned long before_length, after_length;
4179 sector_t min_offset_diff = 0;
4180 int first = 1;
4181 struct geom new;
4182 struct r10conf *conf = mddev->private;
4183 struct md_rdev *rdev;
4184 int spares = 0;
4185 int ret;
4186
4187 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4188 return -EBUSY;
4189
4190 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4191 return -EINVAL;
4192
4193 before_length = ((1 << conf->prev.chunk_shift) *
4194 conf->prev.far_copies);
4195 after_length = ((1 << conf->geo.chunk_shift) *
4196 conf->geo.far_copies);
4197
4198 rdev_for_each(rdev, mddev) {
4199 if (!test_bit(In_sync, &rdev->flags)
4200 && !test_bit(Faulty, &rdev->flags))
4201 spares++;
4202 if (rdev->raid_disk >= 0) {
4203 long long diff = (rdev->new_data_offset
4204 - rdev->data_offset);
4205 if (!mddev->reshape_backwards)
4206 diff = -diff;
4207 if (diff < 0)
4208 diff = 0;
4209 if (first || diff < min_offset_diff)
4210 min_offset_diff = diff;
4211 first = 0;
4212 }
4213 }
4214
4215 if (max(before_length, after_length) > min_offset_diff)
4216 return -EINVAL;
4217
4218 if (spares < mddev->delta_disks)
4219 return -EINVAL;
4220
4221 conf->offset_diff = min_offset_diff;
4222 spin_lock_irq(&conf->device_lock);
4223 if (conf->mirrors_new) {
4224 memcpy(conf->mirrors_new, conf->mirrors,
4225 sizeof(struct raid10_info)*conf->prev.raid_disks);
4226 smp_mb();
4227 kfree(conf->mirrors_old);
4228 conf->mirrors_old = conf->mirrors;
4229 conf->mirrors = conf->mirrors_new;
4230 conf->mirrors_new = NULL;
4231 }
4232 setup_geo(&conf->geo, mddev, geo_start);
4233 smp_mb();
4234 if (mddev->reshape_backwards) {
4235 sector_t size = raid10_size(mddev, 0, 0);
4236 if (size < mddev->array_sectors) {
4237 spin_unlock_irq(&conf->device_lock);
4238 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4239 mdname(mddev));
4240 return -EINVAL;
4241 }
4242 mddev->resync_max_sectors = size;
4243 conf->reshape_progress = size;
4244 } else
4245 conf->reshape_progress = 0;
4246 conf->reshape_safe = conf->reshape_progress;
4247 spin_unlock_irq(&conf->device_lock);
4248
4249 if (mddev->delta_disks && mddev->bitmap) {
4250 struct mdp_superblock_1 *sb = NULL;
4251 sector_t oldsize, newsize;
4252
4253 oldsize = raid10_size(mddev, 0, 0);
4254 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4255
4256 if (!mddev_is_clustered(mddev)) {
4257 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4258 if (ret)
4259 goto abort;
4260 else
4261 goto out;
4262 }
4263
4264 rdev_for_each(rdev, mddev) {
4265 if (rdev->raid_disk > -1 &&
4266 !test_bit(Faulty, &rdev->flags))
4267 sb = page_address(rdev->sb_page);
4268 }
4269
4270 /*
4271 * some node is already performing reshape, and no need to
4272 * call md_bitmap_resize again since it should be called when
4273 * receiving BITMAP_RESIZE msg
4274 */
4275 if ((sb && (le32_to_cpu(sb->feature_map) &
4276 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4277 goto out;
4278
4279 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4280 if (ret)
4281 goto abort;
4282
4283 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4284 if (ret) {
4285 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4286 goto abort;
4287 }
4288 }
4289 out:
4290 if (mddev->delta_disks > 0) {
4291 rdev_for_each(rdev, mddev)
4292 if (rdev->raid_disk < 0 &&
4293 !test_bit(Faulty, &rdev->flags)) {
4294 if (raid10_add_disk(mddev, rdev) == 0) {
4295 if (rdev->raid_disk >=
4296 conf->prev.raid_disks)
4297 set_bit(In_sync, &rdev->flags);
4298 else
4299 rdev->recovery_offset = 0;
4300
4301 /* Failure here is OK */
4302 sysfs_link_rdev(mddev, rdev);
4303 }
4304 } else if (rdev->raid_disk >= conf->prev.raid_disks
4305 && !test_bit(Faulty, &rdev->flags)) {
4306 /* This is a spare that was manually added */
4307 set_bit(In_sync, &rdev->flags);
4308 }
4309 }
4310 /* When a reshape changes the number of devices,
4311 * ->degraded is measured against the larger of the
4312 * pre and post numbers.
4313 */
4314 spin_lock_irq(&conf->device_lock);
4315 mddev->degraded = calc_degraded(conf);
4316 spin_unlock_irq(&conf->device_lock);
4317 mddev->raid_disks = conf->geo.raid_disks;
4318 mddev->reshape_position = conf->reshape_progress;
4319 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4320
4321 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4322 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4323 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4324 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4325 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4326
4327 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4328 "reshape");
4329 if (!mddev->sync_thread) {
4330 ret = -EAGAIN;
4331 goto abort;
4332 }
4333 conf->reshape_checkpoint = jiffies;
4334 md_wakeup_thread(mddev->sync_thread);
4335 md_new_event(mddev);
4336 return 0;
4337
4338 abort:
4339 mddev->recovery = 0;
4340 spin_lock_irq(&conf->device_lock);
4341 conf->geo = conf->prev;
4342 mddev->raid_disks = conf->geo.raid_disks;
4343 rdev_for_each(rdev, mddev)
4344 rdev->new_data_offset = rdev->data_offset;
4345 smp_wmb();
4346 conf->reshape_progress = MaxSector;
4347 conf->reshape_safe = MaxSector;
4348 mddev->reshape_position = MaxSector;
4349 spin_unlock_irq(&conf->device_lock);
4350 return ret;
4351 }
4352
4353 /* Calculate the last device-address that could contain
4354 * any block from the chunk that includes the array-address 's'
4355 * and report the next address.
4356 * i.e. the address returned will be chunk-aligned and after
4357 * any data that is in the chunk containing 's'.
4358 */
last_dev_address(sector_t s, struct geom *geo)4359 static sector_t last_dev_address(sector_t s, struct geom *geo)
4360 {
4361 s = (s | geo->chunk_mask) + 1;
4362 s >>= geo->chunk_shift;
4363 s *= geo->near_copies;
4364 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4365 s *= geo->far_copies;
4366 s <<= geo->chunk_shift;
4367 return s;
4368 }
4369
4370 /* Calculate the first device-address that could contain
4371 * any block from the chunk that includes the array-address 's'.
4372 * This too will be the start of a chunk
4373 */
first_dev_address(sector_t s, struct geom *geo)4374 static sector_t first_dev_address(sector_t s, struct geom *geo)
4375 {
4376 s >>= geo->chunk_shift;
4377 s *= geo->near_copies;
4378 sector_div(s, geo->raid_disks);
4379 s *= geo->far_copies;
4380 s <<= geo->chunk_shift;
4381 return s;
4382 }
4383
reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)4384 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4385 int *skipped)
4386 {
4387 /* We simply copy at most one chunk (smallest of old and new)
4388 * at a time, possibly less if that exceeds RESYNC_PAGES,
4389 * or we hit a bad block or something.
4390 * This might mean we pause for normal IO in the middle of
4391 * a chunk, but that is not a problem as mddev->reshape_position
4392 * can record any location.
4393 *
4394 * If we will want to write to a location that isn't
4395 * yet recorded as 'safe' (i.e. in metadata on disk) then
4396 * we need to flush all reshape requests and update the metadata.
4397 *
4398 * When reshaping forwards (e.g. to more devices), we interpret
4399 * 'safe' as the earliest block which might not have been copied
4400 * down yet. We divide this by previous stripe size and multiply
4401 * by previous stripe length to get lowest device offset that we
4402 * cannot write to yet.
4403 * We interpret 'sector_nr' as an address that we want to write to.
4404 * From this we use last_device_address() to find where we might
4405 * write to, and first_device_address on the 'safe' position.
4406 * If this 'next' write position is after the 'safe' position,
4407 * we must update the metadata to increase the 'safe' position.
4408 *
4409 * When reshaping backwards, we round in the opposite direction
4410 * and perform the reverse test: next write position must not be
4411 * less than current safe position.
4412 *
4413 * In all this the minimum difference in data offsets
4414 * (conf->offset_diff - always positive) allows a bit of slack,
4415 * so next can be after 'safe', but not by more than offset_diff
4416 *
4417 * We need to prepare all the bios here before we start any IO
4418 * to ensure the size we choose is acceptable to all devices.
4419 * The means one for each copy for write-out and an extra one for
4420 * read-in.
4421 * We store the read-in bio in ->master_bio and the others in
4422 * ->devs[x].bio and ->devs[x].repl_bio.
4423 */
4424 struct r10conf *conf = mddev->private;
4425 struct r10bio *r10_bio;
4426 sector_t next, safe, last;
4427 int max_sectors;
4428 int nr_sectors;
4429 int s;
4430 struct md_rdev *rdev;
4431 int need_flush = 0;
4432 struct bio *blist;
4433 struct bio *bio, *read_bio;
4434 int sectors_done = 0;
4435 struct page **pages;
4436
4437 if (sector_nr == 0) {
4438 /* If restarting in the middle, skip the initial sectors */
4439 if (mddev->reshape_backwards &&
4440 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4441 sector_nr = (raid10_size(mddev, 0, 0)
4442 - conf->reshape_progress);
4443 } else if (!mddev->reshape_backwards &&
4444 conf->reshape_progress > 0)
4445 sector_nr = conf->reshape_progress;
4446 if (sector_nr) {
4447 mddev->curr_resync_completed = sector_nr;
4448 sysfs_notify_dirent_safe(mddev->sysfs_completed);
4449 *skipped = 1;
4450 return sector_nr;
4451 }
4452 }
4453
4454 /* We don't use sector_nr to track where we are up to
4455 * as that doesn't work well for ->reshape_backwards.
4456 * So just use ->reshape_progress.
4457 */
4458 if (mddev->reshape_backwards) {
4459 /* 'next' is the earliest device address that we might
4460 * write to for this chunk in the new layout
4461 */
4462 next = first_dev_address(conf->reshape_progress - 1,
4463 &conf->geo);
4464
4465 /* 'safe' is the last device address that we might read from
4466 * in the old layout after a restart
4467 */
4468 safe = last_dev_address(conf->reshape_safe - 1,
4469 &conf->prev);
4470
4471 if (next + conf->offset_diff < safe)
4472 need_flush = 1;
4473
4474 last = conf->reshape_progress - 1;
4475 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4476 & conf->prev.chunk_mask);
4477 if (sector_nr + RESYNC_SECTORS < last)
4478 sector_nr = last + 1 - RESYNC_SECTORS;
4479 } else {
4480 /* 'next' is after the last device address that we
4481 * might write to for this chunk in the new layout
4482 */
4483 next = last_dev_address(conf->reshape_progress, &conf->geo);
4484
4485 /* 'safe' is the earliest device address that we might
4486 * read from in the old layout after a restart
4487 */
4488 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4489
4490 /* Need to update metadata if 'next' might be beyond 'safe'
4491 * as that would possibly corrupt data
4492 */
4493 if (next > safe + conf->offset_diff)
4494 need_flush = 1;
4495
4496 sector_nr = conf->reshape_progress;
4497 last = sector_nr | (conf->geo.chunk_mask
4498 & conf->prev.chunk_mask);
4499
4500 if (sector_nr + RESYNC_SECTORS <= last)
4501 last = sector_nr + RESYNC_SECTORS - 1;
4502 }
4503
4504 if (need_flush ||
4505 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4506 /* Need to update reshape_position in metadata */
4507 wait_barrier(conf);
4508 mddev->reshape_position = conf->reshape_progress;
4509 if (mddev->reshape_backwards)
4510 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4511 - conf->reshape_progress;
4512 else
4513 mddev->curr_resync_completed = conf->reshape_progress;
4514 conf->reshape_checkpoint = jiffies;
4515 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4516 md_wakeup_thread(mddev->thread);
4517 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4518 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4519 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4520 allow_barrier(conf);
4521 return sectors_done;
4522 }
4523 conf->reshape_safe = mddev->reshape_position;
4524 allow_barrier(conf);
4525 }
4526
4527 raise_barrier(conf, 0);
4528 read_more:
4529 /* Now schedule reads for blocks from sector_nr to last */
4530 r10_bio = raid10_alloc_init_r10buf(conf);
4531 r10_bio->state = 0;
4532 raise_barrier(conf, 1);
4533 atomic_set(&r10_bio->remaining, 0);
4534 r10_bio->mddev = mddev;
4535 r10_bio->sector = sector_nr;
4536 set_bit(R10BIO_IsReshape, &r10_bio->state);
4537 r10_bio->sectors = last - sector_nr + 1;
4538 rdev = read_balance(conf, r10_bio, &max_sectors);
4539 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4540
4541 if (!rdev) {
4542 /* Cannot read from here, so need to record bad blocks
4543 * on all the target devices.
4544 */
4545 // FIXME
4546 mempool_free(r10_bio, &conf->r10buf_pool);
4547 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4548 return sectors_done;
4549 }
4550
4551 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4552
4553 bio_set_dev(read_bio, rdev->bdev);
4554 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4555 + rdev->data_offset);
4556 read_bio->bi_private = r10_bio;
4557 read_bio->bi_end_io = end_reshape_read;
4558 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4559 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4560 read_bio->bi_status = 0;
4561 read_bio->bi_vcnt = 0;
4562 read_bio->bi_iter.bi_size = 0;
4563 r10_bio->master_bio = read_bio;
4564 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4565
4566 /*
4567 * Broadcast RESYNC message to other nodes, so all nodes would not
4568 * write to the region to avoid conflict.
4569 */
4570 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4571 struct mdp_superblock_1 *sb = NULL;
4572 int sb_reshape_pos = 0;
4573
4574 conf->cluster_sync_low = sector_nr;
4575 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4576 sb = page_address(rdev->sb_page);
4577 if (sb) {
4578 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4579 /*
4580 * Set cluster_sync_low again if next address for array
4581 * reshape is less than cluster_sync_low. Since we can't
4582 * update cluster_sync_low until it has finished reshape.
4583 */
4584 if (sb_reshape_pos < conf->cluster_sync_low)
4585 conf->cluster_sync_low = sb_reshape_pos;
4586 }
4587
4588 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4589 conf->cluster_sync_high);
4590 }
4591
4592 /* Now find the locations in the new layout */
4593 __raid10_find_phys(&conf->geo, r10_bio);
4594
4595 blist = read_bio;
4596 read_bio->bi_next = NULL;
4597
4598 rcu_read_lock();
4599 for (s = 0; s < conf->copies*2; s++) {
4600 struct bio *b;
4601 int d = r10_bio->devs[s/2].devnum;
4602 struct md_rdev *rdev2;
4603 if (s&1) {
4604 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4605 b = r10_bio->devs[s/2].repl_bio;
4606 } else {
4607 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4608 b = r10_bio->devs[s/2].bio;
4609 }
4610 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4611 continue;
4612
4613 bio_set_dev(b, rdev2->bdev);
4614 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4615 rdev2->new_data_offset;
4616 b->bi_end_io = end_reshape_write;
4617 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4618 b->bi_next = blist;
4619 blist = b;
4620 }
4621
4622 /* Now add as many pages as possible to all of these bios. */
4623
4624 nr_sectors = 0;
4625 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4626 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4627 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4628 int len = (max_sectors - s) << 9;
4629 if (len > PAGE_SIZE)
4630 len = PAGE_SIZE;
4631 for (bio = blist; bio ; bio = bio->bi_next) {
4632 /*
4633 * won't fail because the vec table is big enough
4634 * to hold all these pages
4635 */
4636 bio_add_page(bio, page, len, 0);
4637 }
4638 sector_nr += len >> 9;
4639 nr_sectors += len >> 9;
4640 }
4641 rcu_read_unlock();
4642 r10_bio->sectors = nr_sectors;
4643
4644 /* Now submit the read */
4645 md_sync_acct_bio(read_bio, r10_bio->sectors);
4646 atomic_inc(&r10_bio->remaining);
4647 read_bio->bi_next = NULL;
4648 submit_bio_noacct(read_bio);
4649 sectors_done += nr_sectors;
4650 if (sector_nr <= last)
4651 goto read_more;
4652
4653 lower_barrier(conf);
4654
4655 /* Now that we have done the whole section we can
4656 * update reshape_progress
4657 */
4658 if (mddev->reshape_backwards)
4659 conf->reshape_progress -= sectors_done;
4660 else
4661 conf->reshape_progress += sectors_done;
4662
4663 return sectors_done;
4664 }
4665
4666 static void end_reshape_request(struct r10bio *r10_bio);
4667 static int handle_reshape_read_error(struct mddev *mddev,
4668 struct r10bio *r10_bio);
reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)4669 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4670 {
4671 /* Reshape read completed. Hopefully we have a block
4672 * to write out.
4673 * If we got a read error then we do sync 1-page reads from
4674 * elsewhere until we find the data - or give up.
4675 */
4676 struct r10conf *conf = mddev->private;
4677 int s;
4678
4679 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4680 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4681 /* Reshape has been aborted */
4682 md_done_sync(mddev, r10_bio->sectors, 0);
4683 return;
4684 }
4685
4686 /* We definitely have the data in the pages, schedule the
4687 * writes.
4688 */
4689 atomic_set(&r10_bio->remaining, 1);
4690 for (s = 0; s < conf->copies*2; s++) {
4691 struct bio *b;
4692 int d = r10_bio->devs[s/2].devnum;
4693 struct md_rdev *rdev;
4694 rcu_read_lock();
4695 if (s&1) {
4696 rdev = rcu_dereference(conf->mirrors[d].replacement);
4697 b = r10_bio->devs[s/2].repl_bio;
4698 } else {
4699 rdev = rcu_dereference(conf->mirrors[d].rdev);
4700 b = r10_bio->devs[s/2].bio;
4701 }
4702 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4703 rcu_read_unlock();
4704 continue;
4705 }
4706 atomic_inc(&rdev->nr_pending);
4707 rcu_read_unlock();
4708 md_sync_acct_bio(b, r10_bio->sectors);
4709 atomic_inc(&r10_bio->remaining);
4710 b->bi_next = NULL;
4711 submit_bio_noacct(b);
4712 }
4713 end_reshape_request(r10_bio);
4714 }
4715
end_reshape(struct r10conf *conf)4716 static void end_reshape(struct r10conf *conf)
4717 {
4718 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4719 return;
4720
4721 spin_lock_irq(&conf->device_lock);
4722 conf->prev = conf->geo;
4723 md_finish_reshape(conf->mddev);
4724 smp_wmb();
4725 conf->reshape_progress = MaxSector;
4726 conf->reshape_safe = MaxSector;
4727 spin_unlock_irq(&conf->device_lock);
4728
4729 if (conf->mddev->queue)
4730 raid10_set_io_opt(conf);
4731 conf->fullsync = 0;
4732 }
4733
raid10_update_reshape_pos(struct mddev *mddev)4734 static void raid10_update_reshape_pos(struct mddev *mddev)
4735 {
4736 struct r10conf *conf = mddev->private;
4737 sector_t lo, hi;
4738
4739 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
4740 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4741 || mddev->reshape_position == MaxSector)
4742 conf->reshape_progress = mddev->reshape_position;
4743 else
4744 WARN_ON_ONCE(1);
4745 }
4746
handle_reshape_read_error(struct mddev *mddev, struct r10bio *r10_bio)4747 static int handle_reshape_read_error(struct mddev *mddev,
4748 struct r10bio *r10_bio)
4749 {
4750 /* Use sync reads to get the blocks from somewhere else */
4751 int sectors = r10_bio->sectors;
4752 struct r10conf *conf = mddev->private;
4753 struct r10bio *r10b;
4754 int slot = 0;
4755 int idx = 0;
4756 struct page **pages;
4757
4758 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
4759 if (!r10b) {
4760 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4761 return -ENOMEM;
4762 }
4763
4764 /* reshape IOs share pages from .devs[0].bio */
4765 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4766
4767 r10b->sector = r10_bio->sector;
4768 __raid10_find_phys(&conf->prev, r10b);
4769
4770 while (sectors) {
4771 int s = sectors;
4772 int success = 0;
4773 int first_slot = slot;
4774
4775 if (s > (PAGE_SIZE >> 9))
4776 s = PAGE_SIZE >> 9;
4777
4778 rcu_read_lock();
4779 while (!success) {
4780 int d = r10b->devs[slot].devnum;
4781 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4782 sector_t addr;
4783 if (rdev == NULL ||
4784 test_bit(Faulty, &rdev->flags) ||
4785 !test_bit(In_sync, &rdev->flags))
4786 goto failed;
4787
4788 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4789 atomic_inc(&rdev->nr_pending);
4790 rcu_read_unlock();
4791 success = sync_page_io(rdev,
4792 addr,
4793 s << 9,
4794 pages[idx],
4795 REQ_OP_READ, 0, false);
4796 rdev_dec_pending(rdev, mddev);
4797 rcu_read_lock();
4798 if (success)
4799 break;
4800 failed:
4801 slot++;
4802 if (slot >= conf->copies)
4803 slot = 0;
4804 if (slot == first_slot)
4805 break;
4806 }
4807 rcu_read_unlock();
4808 if (!success) {
4809 /* couldn't read this block, must give up */
4810 set_bit(MD_RECOVERY_INTR,
4811 &mddev->recovery);
4812 kfree(r10b);
4813 return -EIO;
4814 }
4815 sectors -= s;
4816 idx++;
4817 }
4818 kfree(r10b);
4819 return 0;
4820 }
4821
end_reshape_write(struct bio *bio)4822 static void end_reshape_write(struct bio *bio)
4823 {
4824 struct r10bio *r10_bio = get_resync_r10bio(bio);
4825 struct mddev *mddev = r10_bio->mddev;
4826 struct r10conf *conf = mddev->private;
4827 int d;
4828 int slot;
4829 int repl;
4830 struct md_rdev *rdev = NULL;
4831
4832 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4833 if (repl)
4834 rdev = conf->mirrors[d].replacement;
4835 if (!rdev) {
4836 smp_mb();
4837 rdev = conf->mirrors[d].rdev;
4838 }
4839
4840 if (bio->bi_status) {
4841 /* FIXME should record badblock */
4842 md_error(mddev, rdev);
4843 }
4844
4845 rdev_dec_pending(rdev, mddev);
4846 end_reshape_request(r10_bio);
4847 }
4848
end_reshape_request(struct r10bio *r10_bio)4849 static void end_reshape_request(struct r10bio *r10_bio)
4850 {
4851 if (!atomic_dec_and_test(&r10_bio->remaining))
4852 return;
4853 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4854 bio_put(r10_bio->master_bio);
4855 put_buf(r10_bio);
4856 }
4857
raid10_finish_reshape(struct mddev *mddev)4858 static void raid10_finish_reshape(struct mddev *mddev)
4859 {
4860 struct r10conf *conf = mddev->private;
4861
4862 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4863 return;
4864
4865 if (mddev->delta_disks > 0) {
4866 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4867 mddev->recovery_cp = mddev->resync_max_sectors;
4868 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4869 }
4870 mddev->resync_max_sectors = mddev->array_sectors;
4871 } else {
4872 int d;
4873 rcu_read_lock();
4874 for (d = conf->geo.raid_disks ;
4875 d < conf->geo.raid_disks - mddev->delta_disks;
4876 d++) {
4877 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4878 if (rdev)
4879 clear_bit(In_sync, &rdev->flags);
4880 rdev = rcu_dereference(conf->mirrors[d].replacement);
4881 if (rdev)
4882 clear_bit(In_sync, &rdev->flags);
4883 }
4884 rcu_read_unlock();
4885 }
4886 mddev->layout = mddev->new_layout;
4887 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4888 mddev->reshape_position = MaxSector;
4889 mddev->delta_disks = 0;
4890 mddev->reshape_backwards = 0;
4891 }
4892
4893 static struct md_personality raid10_personality =
4894 {
4895 .name = "raid10",
4896 .level = 10,
4897 .owner = THIS_MODULE,
4898 .make_request = raid10_make_request,
4899 .run = raid10_run,
4900 .free = raid10_free,
4901 .status = raid10_status,
4902 .error_handler = raid10_error,
4903 .hot_add_disk = raid10_add_disk,
4904 .hot_remove_disk= raid10_remove_disk,
4905 .spare_active = raid10_spare_active,
4906 .sync_request = raid10_sync_request,
4907 .quiesce = raid10_quiesce,
4908 .size = raid10_size,
4909 .resize = raid10_resize,
4910 .takeover = raid10_takeover,
4911 .check_reshape = raid10_check_reshape,
4912 .start_reshape = raid10_start_reshape,
4913 .finish_reshape = raid10_finish_reshape,
4914 .update_reshape_pos = raid10_update_reshape_pos,
4915 };
4916
raid_init(void)4917 static int __init raid_init(void)
4918 {
4919 return register_md_personality(&raid10_personality);
4920 }
4921
raid_exit(void)4922 static void raid_exit(void)
4923 {
4924 unregister_md_personality(&raid10_personality);
4925 }
4926
4927 module_init(raid_init);
4928 module_exit(raid_exit);
4929 MODULE_LICENSE("GPL");
4930 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4931 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4932 MODULE_ALIAS("md-raid10");
4933 MODULE_ALIAS("md-level-10");
4934
4935 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
4936