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