Loading...
1/*
2 * raid10.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 2000-2004 Neil Brown
5 *
6 * RAID-10 support for md.
7 *
8 * Base on code in raid1.c. See raid1.c for further copyright information.
9 *
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20
21#include <linux/slab.h>
22#include <linux/delay.h>
23#include <linux/blkdev.h>
24#include <linux/module.h>
25#include <linux/seq_file.h>
26#include <linux/ratelimit.h>
27#include <linux/kthread.h>
28#include "md.h"
29#include "raid10.h"
30#include "raid0.h"
31#include "bitmap.h"
32
33/*
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
36 * chunk_size
37 * raid_disks
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
40 * far_offset (stored in bit 16 of layout )
41 *
42 * The data to be stored is divided into chunks using chunksize.
43 * Each device is divided into far_copies sections.
44 * In each section, chunks are laid out in a style similar to raid0, but
45 * near_copies copies of each chunk is stored (each on a different drive).
46 * The starting device for each section is offset near_copies from the starting
47 * device of the previous section.
48 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
49 * drive.
50 * near_copies and far_copies must be at least one, and their product is at most
51 * raid_disks.
52 *
53 * If far_offset is true, then the far_copies are handled a bit differently.
54 * The copies are still in different stripes, but instead of be very far apart
55 * on disk, there are adjacent stripes.
56 */
57
58/*
59 * Number of guaranteed r10bios in case of extreme VM load:
60 */
61#define NR_RAID10_BIOS 256
62
63/* When there are this many requests queue to be written by
64 * the raid10 thread, we become 'congested' to provide back-pressure
65 * for writeback.
66 */
67static int max_queued_requests = 1024;
68
69static void allow_barrier(struct r10conf *conf);
70static void lower_barrier(struct r10conf *conf);
71static int enough(struct r10conf *conf, int ignore);
72static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
73 int *skipped);
74static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
75static void end_reshape_write(struct bio *bio, int error);
76static void end_reshape(struct r10conf *conf);
77
78static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
79{
80 struct r10conf *conf = data;
81 int size = offsetof(struct r10bio, devs[conf->copies]);
82
83 /* allocate a r10bio with room for raid_disks entries in the
84 * bios array */
85 return kzalloc(size, gfp_flags);
86}
87
88static void r10bio_pool_free(void *r10_bio, void *data)
89{
90 kfree(r10_bio);
91}
92
93/* Maximum size of each resync request */
94#define RESYNC_BLOCK_SIZE (64*1024)
95#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
96/* amount of memory to reserve for resync requests */
97#define RESYNC_WINDOW (1024*1024)
98/* maximum number of concurrent requests, memory permitting */
99#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
100
101/*
102 * When performing a resync, we need to read and compare, so
103 * we need as many pages are there are copies.
104 * When performing a recovery, we need 2 bios, one for read,
105 * one for write (we recover only one drive per r10buf)
106 *
107 */
108static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
109{
110 struct r10conf *conf = data;
111 struct page *page;
112 struct r10bio *r10_bio;
113 struct bio *bio;
114 int i, j;
115 int nalloc;
116
117 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
118 if (!r10_bio)
119 return NULL;
120
121 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
122 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
123 nalloc = conf->copies; /* resync */
124 else
125 nalloc = 2; /* recovery */
126
127 /*
128 * Allocate bios.
129 */
130 for (j = nalloc ; j-- ; ) {
131 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
132 if (!bio)
133 goto out_free_bio;
134 r10_bio->devs[j].bio = bio;
135 if (!conf->have_replacement)
136 continue;
137 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
138 if (!bio)
139 goto out_free_bio;
140 r10_bio->devs[j].repl_bio = bio;
141 }
142 /*
143 * Allocate RESYNC_PAGES data pages and attach them
144 * where needed.
145 */
146 for (j = 0 ; j < nalloc; j++) {
147 struct bio *rbio = r10_bio->devs[j].repl_bio;
148 bio = r10_bio->devs[j].bio;
149 for (i = 0; i < RESYNC_PAGES; i++) {
150 if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
151 &conf->mddev->recovery)) {
152 /* we can share bv_page's during recovery
153 * and reshape */
154 struct bio *rbio = r10_bio->devs[0].bio;
155 page = rbio->bi_io_vec[i].bv_page;
156 get_page(page);
157 } else
158 page = alloc_page(gfp_flags);
159 if (unlikely(!page))
160 goto out_free_pages;
161
162 bio->bi_io_vec[i].bv_page = page;
163 if (rbio)
164 rbio->bi_io_vec[i].bv_page = page;
165 }
166 }
167
168 return r10_bio;
169
170out_free_pages:
171 for ( ; i > 0 ; i--)
172 safe_put_page(bio->bi_io_vec[i-1].bv_page);
173 while (j--)
174 for (i = 0; i < RESYNC_PAGES ; i++)
175 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
176 j = 0;
177out_free_bio:
178 for ( ; j < nalloc; j++) {
179 if (r10_bio->devs[j].bio)
180 bio_put(r10_bio->devs[j].bio);
181 if (r10_bio->devs[j].repl_bio)
182 bio_put(r10_bio->devs[j].repl_bio);
183 }
184 r10bio_pool_free(r10_bio, conf);
185 return NULL;
186}
187
188static void r10buf_pool_free(void *__r10_bio, void *data)
189{
190 int i;
191 struct r10conf *conf = data;
192 struct r10bio *r10bio = __r10_bio;
193 int j;
194
195 for (j=0; j < conf->copies; j++) {
196 struct bio *bio = r10bio->devs[j].bio;
197 if (bio) {
198 for (i = 0; i < RESYNC_PAGES; i++) {
199 safe_put_page(bio->bi_io_vec[i].bv_page);
200 bio->bi_io_vec[i].bv_page = NULL;
201 }
202 bio_put(bio);
203 }
204 bio = r10bio->devs[j].repl_bio;
205 if (bio)
206 bio_put(bio);
207 }
208 r10bio_pool_free(r10bio, conf);
209}
210
211static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
212{
213 int i;
214
215 for (i = 0; i < conf->copies; i++) {
216 struct bio **bio = & r10_bio->devs[i].bio;
217 if (!BIO_SPECIAL(*bio))
218 bio_put(*bio);
219 *bio = NULL;
220 bio = &r10_bio->devs[i].repl_bio;
221 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
222 bio_put(*bio);
223 *bio = NULL;
224 }
225}
226
227static void free_r10bio(struct r10bio *r10_bio)
228{
229 struct r10conf *conf = r10_bio->mddev->private;
230
231 put_all_bios(conf, r10_bio);
232 mempool_free(r10_bio, conf->r10bio_pool);
233}
234
235static void put_buf(struct r10bio *r10_bio)
236{
237 struct r10conf *conf = r10_bio->mddev->private;
238
239 mempool_free(r10_bio, conf->r10buf_pool);
240
241 lower_barrier(conf);
242}
243
244static void reschedule_retry(struct r10bio *r10_bio)
245{
246 unsigned long flags;
247 struct mddev *mddev = r10_bio->mddev;
248 struct r10conf *conf = mddev->private;
249
250 spin_lock_irqsave(&conf->device_lock, flags);
251 list_add(&r10_bio->retry_list, &conf->retry_list);
252 conf->nr_queued ++;
253 spin_unlock_irqrestore(&conf->device_lock, flags);
254
255 /* wake up frozen array... */
256 wake_up(&conf->wait_barrier);
257
258 md_wakeup_thread(mddev->thread);
259}
260
261/*
262 * raid_end_bio_io() is called when we have finished servicing a mirrored
263 * operation and are ready to return a success/failure code to the buffer
264 * cache layer.
265 */
266static void raid_end_bio_io(struct r10bio *r10_bio)
267{
268 struct bio *bio = r10_bio->master_bio;
269 int done;
270 struct r10conf *conf = r10_bio->mddev->private;
271
272 if (bio->bi_phys_segments) {
273 unsigned long flags;
274 spin_lock_irqsave(&conf->device_lock, flags);
275 bio->bi_phys_segments--;
276 done = (bio->bi_phys_segments == 0);
277 spin_unlock_irqrestore(&conf->device_lock, flags);
278 } else
279 done = 1;
280 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
281 clear_bit(BIO_UPTODATE, &bio->bi_flags);
282 if (done) {
283 bio_endio(bio, 0);
284 /*
285 * Wake up any possible resync thread that waits for the device
286 * to go idle.
287 */
288 allow_barrier(conf);
289 }
290 free_r10bio(r10_bio);
291}
292
293/*
294 * Update disk head position estimator based on IRQ completion info.
295 */
296static inline void update_head_pos(int slot, struct r10bio *r10_bio)
297{
298 struct r10conf *conf = r10_bio->mddev->private;
299
300 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
301 r10_bio->devs[slot].addr + (r10_bio->sectors);
302}
303
304/*
305 * Find the disk number which triggered given bio
306 */
307static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
308 struct bio *bio, int *slotp, int *replp)
309{
310 int slot;
311 int repl = 0;
312
313 for (slot = 0; slot < conf->copies; slot++) {
314 if (r10_bio->devs[slot].bio == bio)
315 break;
316 if (r10_bio->devs[slot].repl_bio == bio) {
317 repl = 1;
318 break;
319 }
320 }
321
322 BUG_ON(slot == conf->copies);
323 update_head_pos(slot, r10_bio);
324
325 if (slotp)
326 *slotp = slot;
327 if (replp)
328 *replp = repl;
329 return r10_bio->devs[slot].devnum;
330}
331
332static void raid10_end_read_request(struct bio *bio, int error)
333{
334 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
335 struct r10bio *r10_bio = bio->bi_private;
336 int slot, dev;
337 struct md_rdev *rdev;
338 struct r10conf *conf = r10_bio->mddev->private;
339
340
341 slot = r10_bio->read_slot;
342 dev = r10_bio->devs[slot].devnum;
343 rdev = r10_bio->devs[slot].rdev;
344 /*
345 * this branch is our 'one mirror IO has finished' event handler:
346 */
347 update_head_pos(slot, r10_bio);
348
349 if (uptodate) {
350 /*
351 * Set R10BIO_Uptodate in our master bio, so that
352 * we will return a good error code to the higher
353 * levels even if IO on some other mirrored buffer fails.
354 *
355 * The 'master' represents the composite IO operation to
356 * user-side. So if something waits for IO, then it will
357 * wait for the 'master' bio.
358 */
359 set_bit(R10BIO_Uptodate, &r10_bio->state);
360 } else {
361 /* If all other devices that store this block have
362 * failed, we want to return the error upwards rather
363 * than fail the last device. Here we redefine
364 * "uptodate" to mean "Don't want to retry"
365 */
366 unsigned long flags;
367 spin_lock_irqsave(&conf->device_lock, flags);
368 if (!enough(conf, rdev->raid_disk))
369 uptodate = 1;
370 spin_unlock_irqrestore(&conf->device_lock, flags);
371 }
372 if (uptodate) {
373 raid_end_bio_io(r10_bio);
374 rdev_dec_pending(rdev, conf->mddev);
375 } else {
376 /*
377 * oops, read error - keep the refcount on the rdev
378 */
379 char b[BDEVNAME_SIZE];
380 printk_ratelimited(KERN_ERR
381 "md/raid10:%s: %s: rescheduling sector %llu\n",
382 mdname(conf->mddev),
383 bdevname(rdev->bdev, b),
384 (unsigned long long)r10_bio->sector);
385 set_bit(R10BIO_ReadError, &r10_bio->state);
386 reschedule_retry(r10_bio);
387 }
388}
389
390static void close_write(struct r10bio *r10_bio)
391{
392 /* clear the bitmap if all writes complete successfully */
393 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
394 r10_bio->sectors,
395 !test_bit(R10BIO_Degraded, &r10_bio->state),
396 0);
397 md_write_end(r10_bio->mddev);
398}
399
400static void one_write_done(struct r10bio *r10_bio)
401{
402 if (atomic_dec_and_test(&r10_bio->remaining)) {
403 if (test_bit(R10BIO_WriteError, &r10_bio->state))
404 reschedule_retry(r10_bio);
405 else {
406 close_write(r10_bio);
407 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
408 reschedule_retry(r10_bio);
409 else
410 raid_end_bio_io(r10_bio);
411 }
412 }
413}
414
415static void raid10_end_write_request(struct bio *bio, int error)
416{
417 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
418 struct r10bio *r10_bio = bio->bi_private;
419 int dev;
420 int dec_rdev = 1;
421 struct r10conf *conf = r10_bio->mddev->private;
422 int slot, repl;
423 struct md_rdev *rdev = NULL;
424
425 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
426
427 if (repl)
428 rdev = conf->mirrors[dev].replacement;
429 if (!rdev) {
430 smp_rmb();
431 repl = 0;
432 rdev = conf->mirrors[dev].rdev;
433 }
434 /*
435 * this branch is our 'one mirror IO has finished' event handler:
436 */
437 if (!uptodate) {
438 if (repl)
439 /* Never record new bad blocks to replacement,
440 * just fail it.
441 */
442 md_error(rdev->mddev, rdev);
443 else {
444 set_bit(WriteErrorSeen, &rdev->flags);
445 if (!test_and_set_bit(WantReplacement, &rdev->flags))
446 set_bit(MD_RECOVERY_NEEDED,
447 &rdev->mddev->recovery);
448 set_bit(R10BIO_WriteError, &r10_bio->state);
449 dec_rdev = 0;
450 }
451 } else {
452 /*
453 * Set R10BIO_Uptodate in our master bio, so that
454 * we will return a good error code for to the higher
455 * levels even if IO on some other mirrored buffer fails.
456 *
457 * The 'master' represents the composite IO operation to
458 * user-side. So if something waits for IO, then it will
459 * wait for the 'master' bio.
460 */
461 sector_t first_bad;
462 int bad_sectors;
463
464 set_bit(R10BIO_Uptodate, &r10_bio->state);
465
466 /* Maybe we can clear some bad blocks. */
467 if (is_badblock(rdev,
468 r10_bio->devs[slot].addr,
469 r10_bio->sectors,
470 &first_bad, &bad_sectors)) {
471 bio_put(bio);
472 if (repl)
473 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
474 else
475 r10_bio->devs[slot].bio = IO_MADE_GOOD;
476 dec_rdev = 0;
477 set_bit(R10BIO_MadeGood, &r10_bio->state);
478 }
479 }
480
481 /*
482 *
483 * Let's see if all mirrored write operations have finished
484 * already.
485 */
486 one_write_done(r10_bio);
487 if (dec_rdev)
488 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
489}
490
491/*
492 * RAID10 layout manager
493 * As well as the chunksize and raid_disks count, there are two
494 * parameters: near_copies and far_copies.
495 * near_copies * far_copies must be <= raid_disks.
496 * Normally one of these will be 1.
497 * If both are 1, we get raid0.
498 * If near_copies == raid_disks, we get raid1.
499 *
500 * Chunks are laid out in raid0 style with near_copies copies of the
501 * first chunk, followed by near_copies copies of the next chunk and
502 * so on.
503 * If far_copies > 1, then after 1/far_copies of the array has been assigned
504 * as described above, we start again with a device offset of near_copies.
505 * So we effectively have another copy of the whole array further down all
506 * the drives, but with blocks on different drives.
507 * With this layout, and block is never stored twice on the one device.
508 *
509 * raid10_find_phys finds the sector offset of a given virtual sector
510 * on each device that it is on.
511 *
512 * raid10_find_virt does the reverse mapping, from a device and a
513 * sector offset to a virtual address
514 */
515
516static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
517{
518 int n,f;
519 sector_t sector;
520 sector_t chunk;
521 sector_t stripe;
522 int dev;
523 int slot = 0;
524
525 /* now calculate first sector/dev */
526 chunk = r10bio->sector >> geo->chunk_shift;
527 sector = r10bio->sector & geo->chunk_mask;
528
529 chunk *= geo->near_copies;
530 stripe = chunk;
531 dev = sector_div(stripe, geo->raid_disks);
532 if (geo->far_offset)
533 stripe *= geo->far_copies;
534
535 sector += stripe << geo->chunk_shift;
536
537 /* and calculate all the others */
538 for (n = 0; n < geo->near_copies; n++) {
539 int d = dev;
540 sector_t s = sector;
541 r10bio->devs[slot].addr = sector;
542 r10bio->devs[slot].devnum = d;
543 slot++;
544
545 for (f = 1; f < geo->far_copies; f++) {
546 d += geo->near_copies;
547 if (d >= geo->raid_disks)
548 d -= geo->raid_disks;
549 s += geo->stride;
550 r10bio->devs[slot].devnum = d;
551 r10bio->devs[slot].addr = s;
552 slot++;
553 }
554 dev++;
555 if (dev >= geo->raid_disks) {
556 dev = 0;
557 sector += (geo->chunk_mask + 1);
558 }
559 }
560}
561
562static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
563{
564 struct geom *geo = &conf->geo;
565
566 if (conf->reshape_progress != MaxSector &&
567 ((r10bio->sector >= conf->reshape_progress) !=
568 conf->mddev->reshape_backwards)) {
569 set_bit(R10BIO_Previous, &r10bio->state);
570 geo = &conf->prev;
571 } else
572 clear_bit(R10BIO_Previous, &r10bio->state);
573
574 __raid10_find_phys(geo, r10bio);
575}
576
577static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
578{
579 sector_t offset, chunk, vchunk;
580 /* Never use conf->prev as this is only called during resync
581 * or recovery, so reshape isn't happening
582 */
583 struct geom *geo = &conf->geo;
584
585 offset = sector & geo->chunk_mask;
586 if (geo->far_offset) {
587 int fc;
588 chunk = sector >> geo->chunk_shift;
589 fc = sector_div(chunk, geo->far_copies);
590 dev -= fc * geo->near_copies;
591 if (dev < 0)
592 dev += geo->raid_disks;
593 } else {
594 while (sector >= geo->stride) {
595 sector -= geo->stride;
596 if (dev < geo->near_copies)
597 dev += geo->raid_disks - geo->near_copies;
598 else
599 dev -= geo->near_copies;
600 }
601 chunk = sector >> geo->chunk_shift;
602 }
603 vchunk = chunk * geo->raid_disks + dev;
604 sector_div(vchunk, geo->near_copies);
605 return (vchunk << geo->chunk_shift) + offset;
606}
607
608/**
609 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
610 * @q: request queue
611 * @bvm: properties of new bio
612 * @biovec: the request that could be merged to it.
613 *
614 * Return amount of bytes we can accept at this offset
615 * This requires checking for end-of-chunk if near_copies != raid_disks,
616 * and for subordinate merge_bvec_fns if merge_check_needed.
617 */
618static int raid10_mergeable_bvec(struct request_queue *q,
619 struct bvec_merge_data *bvm,
620 struct bio_vec *biovec)
621{
622 struct mddev *mddev = q->queuedata;
623 struct r10conf *conf = mddev->private;
624 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
625 int max;
626 unsigned int chunk_sectors;
627 unsigned int bio_sectors = bvm->bi_size >> 9;
628 struct geom *geo = &conf->geo;
629
630 chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
631 if (conf->reshape_progress != MaxSector &&
632 ((sector >= conf->reshape_progress) !=
633 conf->mddev->reshape_backwards))
634 geo = &conf->prev;
635
636 if (geo->near_copies < geo->raid_disks) {
637 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
638 + bio_sectors)) << 9;
639 if (max < 0)
640 /* bio_add cannot handle a negative return */
641 max = 0;
642 if (max <= biovec->bv_len && bio_sectors == 0)
643 return biovec->bv_len;
644 } else
645 max = biovec->bv_len;
646
647 if (mddev->merge_check_needed) {
648 struct {
649 struct r10bio r10_bio;
650 struct r10dev devs[conf->copies];
651 } on_stack;
652 struct r10bio *r10_bio = &on_stack.r10_bio;
653 int s;
654 if (conf->reshape_progress != MaxSector) {
655 /* Cannot give any guidance during reshape */
656 if (max <= biovec->bv_len && bio_sectors == 0)
657 return biovec->bv_len;
658 return 0;
659 }
660 r10_bio->sector = sector;
661 raid10_find_phys(conf, r10_bio);
662 rcu_read_lock();
663 for (s = 0; s < conf->copies; s++) {
664 int disk = r10_bio->devs[s].devnum;
665 struct md_rdev *rdev = rcu_dereference(
666 conf->mirrors[disk].rdev);
667 if (rdev && !test_bit(Faulty, &rdev->flags)) {
668 struct request_queue *q =
669 bdev_get_queue(rdev->bdev);
670 if (q->merge_bvec_fn) {
671 bvm->bi_sector = r10_bio->devs[s].addr
672 + rdev->data_offset;
673 bvm->bi_bdev = rdev->bdev;
674 max = min(max, q->merge_bvec_fn(
675 q, bvm, biovec));
676 }
677 }
678 rdev = rcu_dereference(conf->mirrors[disk].replacement);
679 if (rdev && !test_bit(Faulty, &rdev->flags)) {
680 struct request_queue *q =
681 bdev_get_queue(rdev->bdev);
682 if (q->merge_bvec_fn) {
683 bvm->bi_sector = r10_bio->devs[s].addr
684 + rdev->data_offset;
685 bvm->bi_bdev = rdev->bdev;
686 max = min(max, q->merge_bvec_fn(
687 q, bvm, biovec));
688 }
689 }
690 }
691 rcu_read_unlock();
692 }
693 return max;
694}
695
696/*
697 * This routine returns the disk from which the requested read should
698 * be done. There is a per-array 'next expected sequential IO' sector
699 * number - if this matches on the next IO then we use the last disk.
700 * There is also a per-disk 'last know head position' sector that is
701 * maintained from IRQ contexts, both the normal and the resync IO
702 * completion handlers update this position correctly. If there is no
703 * perfect sequential match then we pick the disk whose head is closest.
704 *
705 * If there are 2 mirrors in the same 2 devices, performance degrades
706 * because position is mirror, not device based.
707 *
708 * The rdev for the device selected will have nr_pending incremented.
709 */
710
711/*
712 * FIXME: possibly should rethink readbalancing and do it differently
713 * depending on near_copies / far_copies geometry.
714 */
715static struct md_rdev *read_balance(struct r10conf *conf,
716 struct r10bio *r10_bio,
717 int *max_sectors)
718{
719 const sector_t this_sector = r10_bio->sector;
720 int disk, slot;
721 int sectors = r10_bio->sectors;
722 int best_good_sectors;
723 sector_t new_distance, best_dist;
724 struct md_rdev *rdev, *best_rdev;
725 int do_balance;
726 int best_slot;
727 struct geom *geo = &conf->geo;
728
729 raid10_find_phys(conf, r10_bio);
730 rcu_read_lock();
731retry:
732 sectors = r10_bio->sectors;
733 best_slot = -1;
734 best_rdev = NULL;
735 best_dist = MaxSector;
736 best_good_sectors = 0;
737 do_balance = 1;
738 /*
739 * Check if we can balance. We can balance on the whole
740 * device if no resync is going on (recovery is ok), or below
741 * the resync window. We take the first readable disk when
742 * above the resync window.
743 */
744 if (conf->mddev->recovery_cp < MaxSector
745 && (this_sector + sectors >= conf->next_resync))
746 do_balance = 0;
747
748 for (slot = 0; slot < conf->copies ; slot++) {
749 sector_t first_bad;
750 int bad_sectors;
751 sector_t dev_sector;
752
753 if (r10_bio->devs[slot].bio == IO_BLOCKED)
754 continue;
755 disk = r10_bio->devs[slot].devnum;
756 rdev = rcu_dereference(conf->mirrors[disk].replacement);
757 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
758 test_bit(Unmerged, &rdev->flags) ||
759 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
760 rdev = rcu_dereference(conf->mirrors[disk].rdev);
761 if (rdev == NULL ||
762 test_bit(Faulty, &rdev->flags) ||
763 test_bit(Unmerged, &rdev->flags))
764 continue;
765 if (!test_bit(In_sync, &rdev->flags) &&
766 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
767 continue;
768
769 dev_sector = r10_bio->devs[slot].addr;
770 if (is_badblock(rdev, dev_sector, sectors,
771 &first_bad, &bad_sectors)) {
772 if (best_dist < MaxSector)
773 /* Already have a better slot */
774 continue;
775 if (first_bad <= dev_sector) {
776 /* Cannot read here. If this is the
777 * 'primary' device, then we must not read
778 * beyond 'bad_sectors' from another device.
779 */
780 bad_sectors -= (dev_sector - first_bad);
781 if (!do_balance && sectors > bad_sectors)
782 sectors = bad_sectors;
783 if (best_good_sectors > sectors)
784 best_good_sectors = sectors;
785 } else {
786 sector_t good_sectors =
787 first_bad - dev_sector;
788 if (good_sectors > best_good_sectors) {
789 best_good_sectors = good_sectors;
790 best_slot = slot;
791 best_rdev = rdev;
792 }
793 if (!do_balance)
794 /* Must read from here */
795 break;
796 }
797 continue;
798 } else
799 best_good_sectors = sectors;
800
801 if (!do_balance)
802 break;
803
804 /* This optimisation is debatable, and completely destroys
805 * sequential read speed for 'far copies' arrays. So only
806 * keep it for 'near' arrays, and review those later.
807 */
808 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
809 break;
810
811 /* for far > 1 always use the lowest address */
812 if (geo->far_copies > 1)
813 new_distance = r10_bio->devs[slot].addr;
814 else
815 new_distance = abs(r10_bio->devs[slot].addr -
816 conf->mirrors[disk].head_position);
817 if (new_distance < best_dist) {
818 best_dist = new_distance;
819 best_slot = slot;
820 best_rdev = rdev;
821 }
822 }
823 if (slot >= conf->copies) {
824 slot = best_slot;
825 rdev = best_rdev;
826 }
827
828 if (slot >= 0) {
829 atomic_inc(&rdev->nr_pending);
830 if (test_bit(Faulty, &rdev->flags)) {
831 /* Cannot risk returning a device that failed
832 * before we inc'ed nr_pending
833 */
834 rdev_dec_pending(rdev, conf->mddev);
835 goto retry;
836 }
837 r10_bio->read_slot = slot;
838 } else
839 rdev = NULL;
840 rcu_read_unlock();
841 *max_sectors = best_good_sectors;
842
843 return rdev;
844}
845
846static int raid10_congested(void *data, int bits)
847{
848 struct mddev *mddev = data;
849 struct r10conf *conf = mddev->private;
850 int i, ret = 0;
851
852 if ((bits & (1 << BDI_async_congested)) &&
853 conf->pending_count >= max_queued_requests)
854 return 1;
855
856 if (mddev_congested(mddev, bits))
857 return 1;
858 rcu_read_lock();
859 for (i = 0;
860 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
861 && ret == 0;
862 i++) {
863 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
864 if (rdev && !test_bit(Faulty, &rdev->flags)) {
865 struct request_queue *q = bdev_get_queue(rdev->bdev);
866
867 ret |= bdi_congested(&q->backing_dev_info, bits);
868 }
869 }
870 rcu_read_unlock();
871 return ret;
872}
873
874static void flush_pending_writes(struct r10conf *conf)
875{
876 /* Any writes that have been queued but are awaiting
877 * bitmap updates get flushed here.
878 */
879 spin_lock_irq(&conf->device_lock);
880
881 if (conf->pending_bio_list.head) {
882 struct bio *bio;
883 bio = bio_list_get(&conf->pending_bio_list);
884 conf->pending_count = 0;
885 spin_unlock_irq(&conf->device_lock);
886 /* flush any pending bitmap writes to disk
887 * before proceeding w/ I/O */
888 bitmap_unplug(conf->mddev->bitmap);
889 wake_up(&conf->wait_barrier);
890
891 while (bio) { /* submit pending writes */
892 struct bio *next = bio->bi_next;
893 bio->bi_next = NULL;
894 generic_make_request(bio);
895 bio = next;
896 }
897 } else
898 spin_unlock_irq(&conf->device_lock);
899}
900
901/* Barriers....
902 * Sometimes we need to suspend IO while we do something else,
903 * either some resync/recovery, or reconfigure the array.
904 * To do this we raise a 'barrier'.
905 * The 'barrier' is a counter that can be raised multiple times
906 * to count how many activities are happening which preclude
907 * normal IO.
908 * We can only raise the barrier if there is no pending IO.
909 * i.e. if nr_pending == 0.
910 * We choose only to raise the barrier if no-one is waiting for the
911 * barrier to go down. This means that as soon as an IO request
912 * is ready, no other operations which require a barrier will start
913 * until the IO request has had a chance.
914 *
915 * So: regular IO calls 'wait_barrier'. When that returns there
916 * is no backgroup IO happening, It must arrange to call
917 * allow_barrier when it has finished its IO.
918 * backgroup IO calls must call raise_barrier. Once that returns
919 * there is no normal IO happeing. It must arrange to call
920 * lower_barrier when the particular background IO completes.
921 */
922
923static void raise_barrier(struct r10conf *conf, int force)
924{
925 BUG_ON(force && !conf->barrier);
926 spin_lock_irq(&conf->resync_lock);
927
928 /* Wait until no block IO is waiting (unless 'force') */
929 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
930 conf->resync_lock, );
931
932 /* block any new IO from starting */
933 conf->barrier++;
934
935 /* Now wait for all pending IO to complete */
936 wait_event_lock_irq(conf->wait_barrier,
937 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
938 conf->resync_lock, );
939
940 spin_unlock_irq(&conf->resync_lock);
941}
942
943static void lower_barrier(struct r10conf *conf)
944{
945 unsigned long flags;
946 spin_lock_irqsave(&conf->resync_lock, flags);
947 conf->barrier--;
948 spin_unlock_irqrestore(&conf->resync_lock, flags);
949 wake_up(&conf->wait_barrier);
950}
951
952static void wait_barrier(struct r10conf *conf)
953{
954 spin_lock_irq(&conf->resync_lock);
955 if (conf->barrier) {
956 conf->nr_waiting++;
957 /* Wait for the barrier to drop.
958 * However if there are already pending
959 * requests (preventing the barrier from
960 * rising completely), and the
961 * pre-process bio queue isn't empty,
962 * then don't wait, as we need to empty
963 * that queue to get the nr_pending
964 * count down.
965 */
966 wait_event_lock_irq(conf->wait_barrier,
967 !conf->barrier ||
968 (conf->nr_pending &&
969 current->bio_list &&
970 !bio_list_empty(current->bio_list)),
971 conf->resync_lock,
972 );
973 conf->nr_waiting--;
974 }
975 conf->nr_pending++;
976 spin_unlock_irq(&conf->resync_lock);
977}
978
979static void allow_barrier(struct r10conf *conf)
980{
981 unsigned long flags;
982 spin_lock_irqsave(&conf->resync_lock, flags);
983 conf->nr_pending--;
984 spin_unlock_irqrestore(&conf->resync_lock, flags);
985 wake_up(&conf->wait_barrier);
986}
987
988static void freeze_array(struct r10conf *conf)
989{
990 /* stop syncio and normal IO and wait for everything to
991 * go quiet.
992 * We increment barrier and nr_waiting, and then
993 * wait until nr_pending match nr_queued+1
994 * This is called in the context of one normal IO request
995 * that has failed. Thus any sync request that might be pending
996 * will be blocked by nr_pending, and we need to wait for
997 * pending IO requests to complete or be queued for re-try.
998 * Thus the number queued (nr_queued) plus this request (1)
999 * must match the number of pending IOs (nr_pending) before
1000 * we continue.
1001 */
1002 spin_lock_irq(&conf->resync_lock);
1003 conf->barrier++;
1004 conf->nr_waiting++;
1005 wait_event_lock_irq(conf->wait_barrier,
1006 conf->nr_pending == conf->nr_queued+1,
1007 conf->resync_lock,
1008 flush_pending_writes(conf));
1009
1010 spin_unlock_irq(&conf->resync_lock);
1011}
1012
1013static void unfreeze_array(struct r10conf *conf)
1014{
1015 /* reverse the effect of the freeze */
1016 spin_lock_irq(&conf->resync_lock);
1017 conf->barrier--;
1018 conf->nr_waiting--;
1019 wake_up(&conf->wait_barrier);
1020 spin_unlock_irq(&conf->resync_lock);
1021}
1022
1023static sector_t choose_data_offset(struct r10bio *r10_bio,
1024 struct md_rdev *rdev)
1025{
1026 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1027 test_bit(R10BIO_Previous, &r10_bio->state))
1028 return rdev->data_offset;
1029 else
1030 return rdev->new_data_offset;
1031}
1032
1033static void make_request(struct mddev *mddev, struct bio * bio)
1034{
1035 struct r10conf *conf = mddev->private;
1036 struct r10bio *r10_bio;
1037 struct bio *read_bio;
1038 int i;
1039 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1040 int chunk_sects = chunk_mask + 1;
1041 const int rw = bio_data_dir(bio);
1042 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1043 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1044 unsigned long flags;
1045 struct md_rdev *blocked_rdev;
1046 int sectors_handled;
1047 int max_sectors;
1048 int sectors;
1049
1050 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1051 md_flush_request(mddev, bio);
1052 return;
1053 }
1054
1055 /* If this request crosses a chunk boundary, we need to
1056 * split it. This will only happen for 1 PAGE (or less) requests.
1057 */
1058 if (unlikely((bio->bi_sector & chunk_mask) + (bio->bi_size >> 9)
1059 > chunk_sects
1060 && (conf->geo.near_copies < conf->geo.raid_disks
1061 || conf->prev.near_copies < conf->prev.raid_disks))) {
1062 struct bio_pair *bp;
1063 /* Sanity check -- queue functions should prevent this happening */
1064 if (bio->bi_vcnt != 1 ||
1065 bio->bi_idx != 0)
1066 goto bad_map;
1067 /* This is a one page bio that upper layers
1068 * refuse to split for us, so we need to split it.
1069 */
1070 bp = bio_split(bio,
1071 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1072
1073 /* Each of these 'make_request' calls will call 'wait_barrier'.
1074 * If the first succeeds but the second blocks due to the resync
1075 * thread raising the barrier, we will deadlock because the
1076 * IO to the underlying device will be queued in generic_make_request
1077 * and will never complete, so will never reduce nr_pending.
1078 * So increment nr_waiting here so no new raise_barriers will
1079 * succeed, and so the second wait_barrier cannot block.
1080 */
1081 spin_lock_irq(&conf->resync_lock);
1082 conf->nr_waiting++;
1083 spin_unlock_irq(&conf->resync_lock);
1084
1085 make_request(mddev, &bp->bio1);
1086 make_request(mddev, &bp->bio2);
1087
1088 spin_lock_irq(&conf->resync_lock);
1089 conf->nr_waiting--;
1090 wake_up(&conf->wait_barrier);
1091 spin_unlock_irq(&conf->resync_lock);
1092
1093 bio_pair_release(bp);
1094 return;
1095 bad_map:
1096 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1097 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1098 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
1099
1100 bio_io_error(bio);
1101 return;
1102 }
1103
1104 md_write_start(mddev, bio);
1105
1106 /*
1107 * Register the new request and wait if the reconstruction
1108 * thread has put up a bar for new requests.
1109 * Continue immediately if no resync is active currently.
1110 */
1111 wait_barrier(conf);
1112
1113 sectors = bio->bi_size >> 9;
1114 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1115 bio->bi_sector < conf->reshape_progress &&
1116 bio->bi_sector + sectors > conf->reshape_progress) {
1117 /* IO spans the reshape position. Need to wait for
1118 * reshape to pass
1119 */
1120 allow_barrier(conf);
1121 wait_event(conf->wait_barrier,
1122 conf->reshape_progress <= bio->bi_sector ||
1123 conf->reshape_progress >= bio->bi_sector + sectors);
1124 wait_barrier(conf);
1125 }
1126 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1127 bio_data_dir(bio) == WRITE &&
1128 (mddev->reshape_backwards
1129 ? (bio->bi_sector < conf->reshape_safe &&
1130 bio->bi_sector + sectors > conf->reshape_progress)
1131 : (bio->bi_sector + sectors > conf->reshape_safe &&
1132 bio->bi_sector < conf->reshape_progress))) {
1133 /* Need to update reshape_position in metadata */
1134 mddev->reshape_position = conf->reshape_progress;
1135 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1136 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1137 md_wakeup_thread(mddev->thread);
1138 wait_event(mddev->sb_wait,
1139 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1140
1141 conf->reshape_safe = mddev->reshape_position;
1142 }
1143
1144 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1145
1146 r10_bio->master_bio = bio;
1147 r10_bio->sectors = sectors;
1148
1149 r10_bio->mddev = mddev;
1150 r10_bio->sector = bio->bi_sector;
1151 r10_bio->state = 0;
1152
1153 /* We might need to issue multiple reads to different
1154 * devices if there are bad blocks around, so we keep
1155 * track of the number of reads in bio->bi_phys_segments.
1156 * If this is 0, there is only one r10_bio and no locking
1157 * will be needed when the request completes. If it is
1158 * non-zero, then it is the number of not-completed requests.
1159 */
1160 bio->bi_phys_segments = 0;
1161 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1162
1163 if (rw == READ) {
1164 /*
1165 * read balancing logic:
1166 */
1167 struct md_rdev *rdev;
1168 int slot;
1169
1170read_again:
1171 rdev = read_balance(conf, r10_bio, &max_sectors);
1172 if (!rdev) {
1173 raid_end_bio_io(r10_bio);
1174 return;
1175 }
1176 slot = r10_bio->read_slot;
1177
1178 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1179 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1180 max_sectors);
1181
1182 r10_bio->devs[slot].bio = read_bio;
1183 r10_bio->devs[slot].rdev = rdev;
1184
1185 read_bio->bi_sector = r10_bio->devs[slot].addr +
1186 choose_data_offset(r10_bio, rdev);
1187 read_bio->bi_bdev = rdev->bdev;
1188 read_bio->bi_end_io = raid10_end_read_request;
1189 read_bio->bi_rw = READ | do_sync;
1190 read_bio->bi_private = r10_bio;
1191
1192 if (max_sectors < r10_bio->sectors) {
1193 /* Could not read all from this device, so we will
1194 * need another r10_bio.
1195 */
1196 sectors_handled = (r10_bio->sectors + max_sectors
1197 - bio->bi_sector);
1198 r10_bio->sectors = max_sectors;
1199 spin_lock_irq(&conf->device_lock);
1200 if (bio->bi_phys_segments == 0)
1201 bio->bi_phys_segments = 2;
1202 else
1203 bio->bi_phys_segments++;
1204 spin_unlock(&conf->device_lock);
1205 /* Cannot call generic_make_request directly
1206 * as that will be queued in __generic_make_request
1207 * and subsequent mempool_alloc might block
1208 * waiting for it. so hand bio over to raid10d.
1209 */
1210 reschedule_retry(r10_bio);
1211
1212 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1213
1214 r10_bio->master_bio = bio;
1215 r10_bio->sectors = ((bio->bi_size >> 9)
1216 - sectors_handled);
1217 r10_bio->state = 0;
1218 r10_bio->mddev = mddev;
1219 r10_bio->sector = bio->bi_sector + sectors_handled;
1220 goto read_again;
1221 } else
1222 generic_make_request(read_bio);
1223 return;
1224 }
1225
1226 /*
1227 * WRITE:
1228 */
1229 if (conf->pending_count >= max_queued_requests) {
1230 md_wakeup_thread(mddev->thread);
1231 wait_event(conf->wait_barrier,
1232 conf->pending_count < max_queued_requests);
1233 }
1234 /* first select target devices under rcu_lock and
1235 * inc refcount on their rdev. Record them by setting
1236 * bios[x] to bio
1237 * If there are known/acknowledged bad blocks on any device
1238 * on which we have seen a write error, we want to avoid
1239 * writing to those blocks. This potentially requires several
1240 * writes to write around the bad blocks. Each set of writes
1241 * gets its own r10_bio with a set of bios attached. The number
1242 * of r10_bios is recored in bio->bi_phys_segments just as with
1243 * the read case.
1244 */
1245
1246 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1247 raid10_find_phys(conf, r10_bio);
1248retry_write:
1249 blocked_rdev = NULL;
1250 rcu_read_lock();
1251 max_sectors = r10_bio->sectors;
1252
1253 for (i = 0; i < conf->copies; i++) {
1254 int d = r10_bio->devs[i].devnum;
1255 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1256 struct md_rdev *rrdev = rcu_dereference(
1257 conf->mirrors[d].replacement);
1258 if (rdev == rrdev)
1259 rrdev = NULL;
1260 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1261 atomic_inc(&rdev->nr_pending);
1262 blocked_rdev = rdev;
1263 break;
1264 }
1265 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1266 atomic_inc(&rrdev->nr_pending);
1267 blocked_rdev = rrdev;
1268 break;
1269 }
1270 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1271 || test_bit(Unmerged, &rrdev->flags)))
1272 rrdev = NULL;
1273
1274 r10_bio->devs[i].bio = NULL;
1275 r10_bio->devs[i].repl_bio = NULL;
1276 if (!rdev || test_bit(Faulty, &rdev->flags) ||
1277 test_bit(Unmerged, &rdev->flags)) {
1278 set_bit(R10BIO_Degraded, &r10_bio->state);
1279 continue;
1280 }
1281 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1282 sector_t first_bad;
1283 sector_t dev_sector = r10_bio->devs[i].addr;
1284 int bad_sectors;
1285 int is_bad;
1286
1287 is_bad = is_badblock(rdev, dev_sector,
1288 max_sectors,
1289 &first_bad, &bad_sectors);
1290 if (is_bad < 0) {
1291 /* Mustn't write here until the bad block
1292 * is acknowledged
1293 */
1294 atomic_inc(&rdev->nr_pending);
1295 set_bit(BlockedBadBlocks, &rdev->flags);
1296 blocked_rdev = rdev;
1297 break;
1298 }
1299 if (is_bad && first_bad <= dev_sector) {
1300 /* Cannot write here at all */
1301 bad_sectors -= (dev_sector - first_bad);
1302 if (bad_sectors < max_sectors)
1303 /* Mustn't write more than bad_sectors
1304 * to other devices yet
1305 */
1306 max_sectors = bad_sectors;
1307 /* We don't set R10BIO_Degraded as that
1308 * only applies if the disk is missing,
1309 * so it might be re-added, and we want to
1310 * know to recover this chunk.
1311 * In this case the device is here, and the
1312 * fact that this chunk is not in-sync is
1313 * recorded in the bad block log.
1314 */
1315 continue;
1316 }
1317 if (is_bad) {
1318 int good_sectors = first_bad - dev_sector;
1319 if (good_sectors < max_sectors)
1320 max_sectors = good_sectors;
1321 }
1322 }
1323 r10_bio->devs[i].bio = bio;
1324 atomic_inc(&rdev->nr_pending);
1325 if (rrdev) {
1326 r10_bio->devs[i].repl_bio = bio;
1327 atomic_inc(&rrdev->nr_pending);
1328 }
1329 }
1330 rcu_read_unlock();
1331
1332 if (unlikely(blocked_rdev)) {
1333 /* Have to wait for this device to get unblocked, then retry */
1334 int j;
1335 int d;
1336
1337 for (j = 0; j < i; j++) {
1338 if (r10_bio->devs[j].bio) {
1339 d = r10_bio->devs[j].devnum;
1340 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1341 }
1342 if (r10_bio->devs[j].repl_bio) {
1343 struct md_rdev *rdev;
1344 d = r10_bio->devs[j].devnum;
1345 rdev = conf->mirrors[d].replacement;
1346 if (!rdev) {
1347 /* Race with remove_disk */
1348 smp_mb();
1349 rdev = conf->mirrors[d].rdev;
1350 }
1351 rdev_dec_pending(rdev, mddev);
1352 }
1353 }
1354 allow_barrier(conf);
1355 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1356 wait_barrier(conf);
1357 goto retry_write;
1358 }
1359
1360 if (max_sectors < r10_bio->sectors) {
1361 /* We are splitting this into multiple parts, so
1362 * we need to prepare for allocating another r10_bio.
1363 */
1364 r10_bio->sectors = max_sectors;
1365 spin_lock_irq(&conf->device_lock);
1366 if (bio->bi_phys_segments == 0)
1367 bio->bi_phys_segments = 2;
1368 else
1369 bio->bi_phys_segments++;
1370 spin_unlock_irq(&conf->device_lock);
1371 }
1372 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1373
1374 atomic_set(&r10_bio->remaining, 1);
1375 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1376
1377 for (i = 0; i < conf->copies; i++) {
1378 struct bio *mbio;
1379 int d = r10_bio->devs[i].devnum;
1380 if (!r10_bio->devs[i].bio)
1381 continue;
1382
1383 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1384 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1385 max_sectors);
1386 r10_bio->devs[i].bio = mbio;
1387
1388 mbio->bi_sector = (r10_bio->devs[i].addr+
1389 choose_data_offset(r10_bio,
1390 conf->mirrors[d].rdev));
1391 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1392 mbio->bi_end_io = raid10_end_write_request;
1393 mbio->bi_rw = WRITE | do_sync | do_fua;
1394 mbio->bi_private = r10_bio;
1395
1396 atomic_inc(&r10_bio->remaining);
1397 spin_lock_irqsave(&conf->device_lock, flags);
1398 bio_list_add(&conf->pending_bio_list, mbio);
1399 conf->pending_count++;
1400 spin_unlock_irqrestore(&conf->device_lock, flags);
1401 if (!mddev_check_plugged(mddev))
1402 md_wakeup_thread(mddev->thread);
1403
1404 if (!r10_bio->devs[i].repl_bio)
1405 continue;
1406
1407 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1408 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1409 max_sectors);
1410 r10_bio->devs[i].repl_bio = mbio;
1411
1412 /* We are actively writing to the original device
1413 * so it cannot disappear, so the replacement cannot
1414 * become NULL here
1415 */
1416 mbio->bi_sector = (r10_bio->devs[i].addr +
1417 choose_data_offset(
1418 r10_bio,
1419 conf->mirrors[d].replacement));
1420 mbio->bi_bdev = conf->mirrors[d].replacement->bdev;
1421 mbio->bi_end_io = raid10_end_write_request;
1422 mbio->bi_rw = WRITE | do_sync | do_fua;
1423 mbio->bi_private = r10_bio;
1424
1425 atomic_inc(&r10_bio->remaining);
1426 spin_lock_irqsave(&conf->device_lock, flags);
1427 bio_list_add(&conf->pending_bio_list, mbio);
1428 conf->pending_count++;
1429 spin_unlock_irqrestore(&conf->device_lock, flags);
1430 if (!mddev_check_plugged(mddev))
1431 md_wakeup_thread(mddev->thread);
1432 }
1433
1434 /* Don't remove the bias on 'remaining' (one_write_done) until
1435 * after checking if we need to go around again.
1436 */
1437
1438 if (sectors_handled < (bio->bi_size >> 9)) {
1439 one_write_done(r10_bio);
1440 /* We need another r10_bio. It has already been counted
1441 * in bio->bi_phys_segments.
1442 */
1443 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1444
1445 r10_bio->master_bio = bio;
1446 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1447
1448 r10_bio->mddev = mddev;
1449 r10_bio->sector = bio->bi_sector + sectors_handled;
1450 r10_bio->state = 0;
1451 goto retry_write;
1452 }
1453 one_write_done(r10_bio);
1454
1455 /* In case raid10d snuck in to freeze_array */
1456 wake_up(&conf->wait_barrier);
1457}
1458
1459static void status(struct seq_file *seq, struct mddev *mddev)
1460{
1461 struct r10conf *conf = mddev->private;
1462 int i;
1463
1464 if (conf->geo.near_copies < conf->geo.raid_disks)
1465 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1466 if (conf->geo.near_copies > 1)
1467 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1468 if (conf->geo.far_copies > 1) {
1469 if (conf->geo.far_offset)
1470 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1471 else
1472 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1473 }
1474 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1475 conf->geo.raid_disks - mddev->degraded);
1476 for (i = 0; i < conf->geo.raid_disks; i++)
1477 seq_printf(seq, "%s",
1478 conf->mirrors[i].rdev &&
1479 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1480 seq_printf(seq, "]");
1481}
1482
1483/* check if there are enough drives for
1484 * every block to appear on atleast one.
1485 * Don't consider the device numbered 'ignore'
1486 * as we might be about to remove it.
1487 */
1488static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
1489{
1490 int first = 0;
1491
1492 do {
1493 int n = conf->copies;
1494 int cnt = 0;
1495 int this = first;
1496 while (n--) {
1497 if (conf->mirrors[this].rdev &&
1498 this != ignore)
1499 cnt++;
1500 this = (this+1) % geo->raid_disks;
1501 }
1502 if (cnt == 0)
1503 return 0;
1504 first = (first + geo->near_copies) % geo->raid_disks;
1505 } while (first != 0);
1506 return 1;
1507}
1508
1509static int enough(struct r10conf *conf, int ignore)
1510{
1511 return _enough(conf, &conf->geo, ignore) &&
1512 _enough(conf, &conf->prev, ignore);
1513}
1514
1515static void error(struct mddev *mddev, struct md_rdev *rdev)
1516{
1517 char b[BDEVNAME_SIZE];
1518 struct r10conf *conf = mddev->private;
1519
1520 /*
1521 * If it is not operational, then we have already marked it as dead
1522 * else if it is the last working disks, ignore the error, let the
1523 * next level up know.
1524 * else mark the drive as failed
1525 */
1526 if (test_bit(In_sync, &rdev->flags)
1527 && !enough(conf, rdev->raid_disk))
1528 /*
1529 * Don't fail the drive, just return an IO error.
1530 */
1531 return;
1532 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1533 unsigned long flags;
1534 spin_lock_irqsave(&conf->device_lock, flags);
1535 mddev->degraded++;
1536 spin_unlock_irqrestore(&conf->device_lock, flags);
1537 /*
1538 * if recovery is running, make sure it aborts.
1539 */
1540 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1541 }
1542 set_bit(Blocked, &rdev->flags);
1543 set_bit(Faulty, &rdev->flags);
1544 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1545 printk(KERN_ALERT
1546 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1547 "md/raid10:%s: Operation continuing on %d devices.\n",
1548 mdname(mddev), bdevname(rdev->bdev, b),
1549 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1550}
1551
1552static void print_conf(struct r10conf *conf)
1553{
1554 int i;
1555 struct mirror_info *tmp;
1556
1557 printk(KERN_DEBUG "RAID10 conf printout:\n");
1558 if (!conf) {
1559 printk(KERN_DEBUG "(!conf)\n");
1560 return;
1561 }
1562 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1563 conf->geo.raid_disks);
1564
1565 for (i = 0; i < conf->geo.raid_disks; i++) {
1566 char b[BDEVNAME_SIZE];
1567 tmp = conf->mirrors + i;
1568 if (tmp->rdev)
1569 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1570 i, !test_bit(In_sync, &tmp->rdev->flags),
1571 !test_bit(Faulty, &tmp->rdev->flags),
1572 bdevname(tmp->rdev->bdev,b));
1573 }
1574}
1575
1576static void close_sync(struct r10conf *conf)
1577{
1578 wait_barrier(conf);
1579 allow_barrier(conf);
1580
1581 mempool_destroy(conf->r10buf_pool);
1582 conf->r10buf_pool = NULL;
1583}
1584
1585static int raid10_spare_active(struct mddev *mddev)
1586{
1587 int i;
1588 struct r10conf *conf = mddev->private;
1589 struct mirror_info *tmp;
1590 int count = 0;
1591 unsigned long flags;
1592
1593 /*
1594 * Find all non-in_sync disks within the RAID10 configuration
1595 * and mark them in_sync
1596 */
1597 for (i = 0; i < conf->geo.raid_disks; i++) {
1598 tmp = conf->mirrors + i;
1599 if (tmp->replacement
1600 && tmp->replacement->recovery_offset == MaxSector
1601 && !test_bit(Faulty, &tmp->replacement->flags)
1602 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1603 /* Replacement has just become active */
1604 if (!tmp->rdev
1605 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1606 count++;
1607 if (tmp->rdev) {
1608 /* Replaced device not technically faulty,
1609 * but we need to be sure it gets removed
1610 * and never re-added.
1611 */
1612 set_bit(Faulty, &tmp->rdev->flags);
1613 sysfs_notify_dirent_safe(
1614 tmp->rdev->sysfs_state);
1615 }
1616 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1617 } else if (tmp->rdev
1618 && !test_bit(Faulty, &tmp->rdev->flags)
1619 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1620 count++;
1621 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1622 }
1623 }
1624 spin_lock_irqsave(&conf->device_lock, flags);
1625 mddev->degraded -= count;
1626 spin_unlock_irqrestore(&conf->device_lock, flags);
1627
1628 print_conf(conf);
1629 return count;
1630}
1631
1632
1633static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1634{
1635 struct r10conf *conf = mddev->private;
1636 int err = -EEXIST;
1637 int mirror;
1638 int first = 0;
1639 int last = conf->geo.raid_disks - 1;
1640 struct request_queue *q = bdev_get_queue(rdev->bdev);
1641
1642 if (mddev->recovery_cp < MaxSector)
1643 /* only hot-add to in-sync arrays, as recovery is
1644 * very different from resync
1645 */
1646 return -EBUSY;
1647 if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
1648 return -EINVAL;
1649
1650 if (rdev->raid_disk >= 0)
1651 first = last = rdev->raid_disk;
1652
1653 if (q->merge_bvec_fn) {
1654 set_bit(Unmerged, &rdev->flags);
1655 mddev->merge_check_needed = 1;
1656 }
1657
1658 if (rdev->saved_raid_disk >= first &&
1659 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1660 mirror = rdev->saved_raid_disk;
1661 else
1662 mirror = first;
1663 for ( ; mirror <= last ; mirror++) {
1664 struct mirror_info *p = &conf->mirrors[mirror];
1665 if (p->recovery_disabled == mddev->recovery_disabled)
1666 continue;
1667 if (p->rdev) {
1668 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1669 p->replacement != NULL)
1670 continue;
1671 clear_bit(In_sync, &rdev->flags);
1672 set_bit(Replacement, &rdev->flags);
1673 rdev->raid_disk = mirror;
1674 err = 0;
1675 disk_stack_limits(mddev->gendisk, rdev->bdev,
1676 rdev->data_offset << 9);
1677 conf->fullsync = 1;
1678 rcu_assign_pointer(p->replacement, rdev);
1679 break;
1680 }
1681
1682 disk_stack_limits(mddev->gendisk, rdev->bdev,
1683 rdev->data_offset << 9);
1684
1685 p->head_position = 0;
1686 p->recovery_disabled = mddev->recovery_disabled - 1;
1687 rdev->raid_disk = mirror;
1688 err = 0;
1689 if (rdev->saved_raid_disk != mirror)
1690 conf->fullsync = 1;
1691 rcu_assign_pointer(p->rdev, rdev);
1692 break;
1693 }
1694 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1695 /* Some requests might not have seen this new
1696 * merge_bvec_fn. We must wait for them to complete
1697 * before merging the device fully.
1698 * First we make sure any code which has tested
1699 * our function has submitted the request, then
1700 * we wait for all outstanding requests to complete.
1701 */
1702 synchronize_sched();
1703 raise_barrier(conf, 0);
1704 lower_barrier(conf);
1705 clear_bit(Unmerged, &rdev->flags);
1706 }
1707 md_integrity_add_rdev(rdev, mddev);
1708 print_conf(conf);
1709 return err;
1710}
1711
1712static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1713{
1714 struct r10conf *conf = mddev->private;
1715 int err = 0;
1716 int number = rdev->raid_disk;
1717 struct md_rdev **rdevp;
1718 struct mirror_info *p = conf->mirrors + number;
1719
1720 print_conf(conf);
1721 if (rdev == p->rdev)
1722 rdevp = &p->rdev;
1723 else if (rdev == p->replacement)
1724 rdevp = &p->replacement;
1725 else
1726 return 0;
1727
1728 if (test_bit(In_sync, &rdev->flags) ||
1729 atomic_read(&rdev->nr_pending)) {
1730 err = -EBUSY;
1731 goto abort;
1732 }
1733 /* Only remove faulty devices if recovery
1734 * is not possible.
1735 */
1736 if (!test_bit(Faulty, &rdev->flags) &&
1737 mddev->recovery_disabled != p->recovery_disabled &&
1738 (!p->replacement || p->replacement == rdev) &&
1739 number < conf->geo.raid_disks &&
1740 enough(conf, -1)) {
1741 err = -EBUSY;
1742 goto abort;
1743 }
1744 *rdevp = NULL;
1745 synchronize_rcu();
1746 if (atomic_read(&rdev->nr_pending)) {
1747 /* lost the race, try later */
1748 err = -EBUSY;
1749 *rdevp = rdev;
1750 goto abort;
1751 } else if (p->replacement) {
1752 /* We must have just cleared 'rdev' */
1753 p->rdev = p->replacement;
1754 clear_bit(Replacement, &p->replacement->flags);
1755 smp_mb(); /* Make sure other CPUs may see both as identical
1756 * but will never see neither -- if they are careful.
1757 */
1758 p->replacement = NULL;
1759 clear_bit(WantReplacement, &rdev->flags);
1760 } else
1761 /* We might have just remove the Replacement as faulty
1762 * Clear the flag just in case
1763 */
1764 clear_bit(WantReplacement, &rdev->flags);
1765
1766 err = md_integrity_register(mddev);
1767
1768abort:
1769
1770 print_conf(conf);
1771 return err;
1772}
1773
1774
1775static void end_sync_read(struct bio *bio, int error)
1776{
1777 struct r10bio *r10_bio = bio->bi_private;
1778 struct r10conf *conf = r10_bio->mddev->private;
1779 int d;
1780
1781 if (bio == r10_bio->master_bio) {
1782 /* this is a reshape read */
1783 d = r10_bio->read_slot; /* really the read dev */
1784 } else
1785 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1786
1787 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1788 set_bit(R10BIO_Uptodate, &r10_bio->state);
1789 else
1790 /* The write handler will notice the lack of
1791 * R10BIO_Uptodate and record any errors etc
1792 */
1793 atomic_add(r10_bio->sectors,
1794 &conf->mirrors[d].rdev->corrected_errors);
1795
1796 /* for reconstruct, we always reschedule after a read.
1797 * for resync, only after all reads
1798 */
1799 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1800 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1801 atomic_dec_and_test(&r10_bio->remaining)) {
1802 /* we have read all the blocks,
1803 * do the comparison in process context in raid10d
1804 */
1805 reschedule_retry(r10_bio);
1806 }
1807}
1808
1809static void end_sync_request(struct r10bio *r10_bio)
1810{
1811 struct mddev *mddev = r10_bio->mddev;
1812
1813 while (atomic_dec_and_test(&r10_bio->remaining)) {
1814 if (r10_bio->master_bio == NULL) {
1815 /* the primary of several recovery bios */
1816 sector_t s = r10_bio->sectors;
1817 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1818 test_bit(R10BIO_WriteError, &r10_bio->state))
1819 reschedule_retry(r10_bio);
1820 else
1821 put_buf(r10_bio);
1822 md_done_sync(mddev, s, 1);
1823 break;
1824 } else {
1825 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1826 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1827 test_bit(R10BIO_WriteError, &r10_bio->state))
1828 reschedule_retry(r10_bio);
1829 else
1830 put_buf(r10_bio);
1831 r10_bio = r10_bio2;
1832 }
1833 }
1834}
1835
1836static void end_sync_write(struct bio *bio, int error)
1837{
1838 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1839 struct r10bio *r10_bio = bio->bi_private;
1840 struct mddev *mddev = r10_bio->mddev;
1841 struct r10conf *conf = mddev->private;
1842 int d;
1843 sector_t first_bad;
1844 int bad_sectors;
1845 int slot;
1846 int repl;
1847 struct md_rdev *rdev = NULL;
1848
1849 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1850 if (repl)
1851 rdev = conf->mirrors[d].replacement;
1852 else
1853 rdev = conf->mirrors[d].rdev;
1854
1855 if (!uptodate) {
1856 if (repl)
1857 md_error(mddev, rdev);
1858 else {
1859 set_bit(WriteErrorSeen, &rdev->flags);
1860 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1861 set_bit(MD_RECOVERY_NEEDED,
1862 &rdev->mddev->recovery);
1863 set_bit(R10BIO_WriteError, &r10_bio->state);
1864 }
1865 } else if (is_badblock(rdev,
1866 r10_bio->devs[slot].addr,
1867 r10_bio->sectors,
1868 &first_bad, &bad_sectors))
1869 set_bit(R10BIO_MadeGood, &r10_bio->state);
1870
1871 rdev_dec_pending(rdev, mddev);
1872
1873 end_sync_request(r10_bio);
1874}
1875
1876/*
1877 * Note: sync and recover and handled very differently for raid10
1878 * This code is for resync.
1879 * For resync, we read through virtual addresses and read all blocks.
1880 * If there is any error, we schedule a write. The lowest numbered
1881 * drive is authoritative.
1882 * However requests come for physical address, so we need to map.
1883 * For every physical address there are raid_disks/copies virtual addresses,
1884 * which is always are least one, but is not necessarly an integer.
1885 * This means that a physical address can span multiple chunks, so we may
1886 * have to submit multiple io requests for a single sync request.
1887 */
1888/*
1889 * We check if all blocks are in-sync and only write to blocks that
1890 * aren't in sync
1891 */
1892static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1893{
1894 struct r10conf *conf = mddev->private;
1895 int i, first;
1896 struct bio *tbio, *fbio;
1897 int vcnt;
1898
1899 atomic_set(&r10_bio->remaining, 1);
1900
1901 /* find the first device with a block */
1902 for (i=0; i<conf->copies; i++)
1903 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1904 break;
1905
1906 if (i == conf->copies)
1907 goto done;
1908
1909 first = i;
1910 fbio = r10_bio->devs[i].bio;
1911
1912 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1913 /* now find blocks with errors */
1914 for (i=0 ; i < conf->copies ; i++) {
1915 int j, d;
1916
1917 tbio = r10_bio->devs[i].bio;
1918
1919 if (tbio->bi_end_io != end_sync_read)
1920 continue;
1921 if (i == first)
1922 continue;
1923 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1924 /* We know that the bi_io_vec layout is the same for
1925 * both 'first' and 'i', so we just compare them.
1926 * All vec entries are PAGE_SIZE;
1927 */
1928 for (j = 0; j < vcnt; j++)
1929 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1930 page_address(tbio->bi_io_vec[j].bv_page),
1931 fbio->bi_io_vec[j].bv_len))
1932 break;
1933 if (j == vcnt)
1934 continue;
1935 mddev->resync_mismatches += r10_bio->sectors;
1936 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1937 /* Don't fix anything. */
1938 continue;
1939 }
1940 /* Ok, we need to write this bio, either to correct an
1941 * inconsistency or to correct an unreadable block.
1942 * First we need to fixup bv_offset, bv_len and
1943 * bi_vecs, as the read request might have corrupted these
1944 */
1945 tbio->bi_vcnt = vcnt;
1946 tbio->bi_size = r10_bio->sectors << 9;
1947 tbio->bi_idx = 0;
1948 tbio->bi_phys_segments = 0;
1949 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1950 tbio->bi_flags |= 1 << BIO_UPTODATE;
1951 tbio->bi_next = NULL;
1952 tbio->bi_rw = WRITE;
1953 tbio->bi_private = r10_bio;
1954 tbio->bi_sector = r10_bio->devs[i].addr;
1955
1956 for (j=0; j < vcnt ; j++) {
1957 tbio->bi_io_vec[j].bv_offset = 0;
1958 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1959
1960 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1961 page_address(fbio->bi_io_vec[j].bv_page),
1962 PAGE_SIZE);
1963 }
1964 tbio->bi_end_io = end_sync_write;
1965
1966 d = r10_bio->devs[i].devnum;
1967 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1968 atomic_inc(&r10_bio->remaining);
1969 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1970
1971 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1972 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1973 generic_make_request(tbio);
1974 }
1975
1976 /* Now write out to any replacement devices
1977 * that are active
1978 */
1979 for (i = 0; i < conf->copies; i++) {
1980 int j, d;
1981
1982 tbio = r10_bio->devs[i].repl_bio;
1983 if (!tbio || !tbio->bi_end_io)
1984 continue;
1985 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
1986 && r10_bio->devs[i].bio != fbio)
1987 for (j = 0; j < vcnt; j++)
1988 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1989 page_address(fbio->bi_io_vec[j].bv_page),
1990 PAGE_SIZE);
1991 d = r10_bio->devs[i].devnum;
1992 atomic_inc(&r10_bio->remaining);
1993 md_sync_acct(conf->mirrors[d].replacement->bdev,
1994 tbio->bi_size >> 9);
1995 generic_make_request(tbio);
1996 }
1997
1998done:
1999 if (atomic_dec_and_test(&r10_bio->remaining)) {
2000 md_done_sync(mddev, r10_bio->sectors, 1);
2001 put_buf(r10_bio);
2002 }
2003}
2004
2005/*
2006 * Now for the recovery code.
2007 * Recovery happens across physical sectors.
2008 * We recover all non-is_sync drives by finding the virtual address of
2009 * each, and then choose a working drive that also has that virt address.
2010 * There is a separate r10_bio for each non-in_sync drive.
2011 * Only the first two slots are in use. The first for reading,
2012 * The second for writing.
2013 *
2014 */
2015static void fix_recovery_read_error(struct r10bio *r10_bio)
2016{
2017 /* We got a read error during recovery.
2018 * We repeat the read in smaller page-sized sections.
2019 * If a read succeeds, write it to the new device or record
2020 * a bad block if we cannot.
2021 * If a read fails, record a bad block on both old and
2022 * new devices.
2023 */
2024 struct mddev *mddev = r10_bio->mddev;
2025 struct r10conf *conf = mddev->private;
2026 struct bio *bio = r10_bio->devs[0].bio;
2027 sector_t sect = 0;
2028 int sectors = r10_bio->sectors;
2029 int idx = 0;
2030 int dr = r10_bio->devs[0].devnum;
2031 int dw = r10_bio->devs[1].devnum;
2032
2033 while (sectors) {
2034 int s = sectors;
2035 struct md_rdev *rdev;
2036 sector_t addr;
2037 int ok;
2038
2039 if (s > (PAGE_SIZE>>9))
2040 s = PAGE_SIZE >> 9;
2041
2042 rdev = conf->mirrors[dr].rdev;
2043 addr = r10_bio->devs[0].addr + sect,
2044 ok = sync_page_io(rdev,
2045 addr,
2046 s << 9,
2047 bio->bi_io_vec[idx].bv_page,
2048 READ, false);
2049 if (ok) {
2050 rdev = conf->mirrors[dw].rdev;
2051 addr = r10_bio->devs[1].addr + sect;
2052 ok = sync_page_io(rdev,
2053 addr,
2054 s << 9,
2055 bio->bi_io_vec[idx].bv_page,
2056 WRITE, false);
2057 if (!ok) {
2058 set_bit(WriteErrorSeen, &rdev->flags);
2059 if (!test_and_set_bit(WantReplacement,
2060 &rdev->flags))
2061 set_bit(MD_RECOVERY_NEEDED,
2062 &rdev->mddev->recovery);
2063 }
2064 }
2065 if (!ok) {
2066 /* We don't worry if we cannot set a bad block -
2067 * it really is bad so there is no loss in not
2068 * recording it yet
2069 */
2070 rdev_set_badblocks(rdev, addr, s, 0);
2071
2072 if (rdev != conf->mirrors[dw].rdev) {
2073 /* need bad block on destination too */
2074 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2075 addr = r10_bio->devs[1].addr + sect;
2076 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2077 if (!ok) {
2078 /* just abort the recovery */
2079 printk(KERN_NOTICE
2080 "md/raid10:%s: recovery aborted"
2081 " due to read error\n",
2082 mdname(mddev));
2083
2084 conf->mirrors[dw].recovery_disabled
2085 = mddev->recovery_disabled;
2086 set_bit(MD_RECOVERY_INTR,
2087 &mddev->recovery);
2088 break;
2089 }
2090 }
2091 }
2092
2093 sectors -= s;
2094 sect += s;
2095 idx++;
2096 }
2097}
2098
2099static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2100{
2101 struct r10conf *conf = mddev->private;
2102 int d;
2103 struct bio *wbio, *wbio2;
2104
2105 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2106 fix_recovery_read_error(r10_bio);
2107 end_sync_request(r10_bio);
2108 return;
2109 }
2110
2111 /*
2112 * share the pages with the first bio
2113 * and submit the write request
2114 */
2115 d = r10_bio->devs[1].devnum;
2116 wbio = r10_bio->devs[1].bio;
2117 wbio2 = r10_bio->devs[1].repl_bio;
2118 if (wbio->bi_end_io) {
2119 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2120 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
2121 generic_make_request(wbio);
2122 }
2123 if (wbio2 && wbio2->bi_end_io) {
2124 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2125 md_sync_acct(conf->mirrors[d].replacement->bdev,
2126 wbio2->bi_size >> 9);
2127 generic_make_request(wbio2);
2128 }
2129}
2130
2131
2132/*
2133 * Used by fix_read_error() to decay the per rdev read_errors.
2134 * We halve the read error count for every hour that has elapsed
2135 * since the last recorded read error.
2136 *
2137 */
2138static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2139{
2140 struct timespec cur_time_mon;
2141 unsigned long hours_since_last;
2142 unsigned int read_errors = atomic_read(&rdev->read_errors);
2143
2144 ktime_get_ts(&cur_time_mon);
2145
2146 if (rdev->last_read_error.tv_sec == 0 &&
2147 rdev->last_read_error.tv_nsec == 0) {
2148 /* first time we've seen a read error */
2149 rdev->last_read_error = cur_time_mon;
2150 return;
2151 }
2152
2153 hours_since_last = (cur_time_mon.tv_sec -
2154 rdev->last_read_error.tv_sec) / 3600;
2155
2156 rdev->last_read_error = cur_time_mon;
2157
2158 /*
2159 * if hours_since_last is > the number of bits in read_errors
2160 * just set read errors to 0. We do this to avoid
2161 * overflowing the shift of read_errors by hours_since_last.
2162 */
2163 if (hours_since_last >= 8 * sizeof(read_errors))
2164 atomic_set(&rdev->read_errors, 0);
2165 else
2166 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2167}
2168
2169static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2170 int sectors, struct page *page, int rw)
2171{
2172 sector_t first_bad;
2173 int bad_sectors;
2174
2175 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2176 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2177 return -1;
2178 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2179 /* success */
2180 return 1;
2181 if (rw == WRITE) {
2182 set_bit(WriteErrorSeen, &rdev->flags);
2183 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2184 set_bit(MD_RECOVERY_NEEDED,
2185 &rdev->mddev->recovery);
2186 }
2187 /* need to record an error - either for the block or the device */
2188 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2189 md_error(rdev->mddev, rdev);
2190 return 0;
2191}
2192
2193/*
2194 * This is a kernel thread which:
2195 *
2196 * 1. Retries failed read operations on working mirrors.
2197 * 2. Updates the raid superblock when problems encounter.
2198 * 3. Performs writes following reads for array synchronising.
2199 */
2200
2201static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2202{
2203 int sect = 0; /* Offset from r10_bio->sector */
2204 int sectors = r10_bio->sectors;
2205 struct md_rdev*rdev;
2206 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2207 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2208
2209 /* still own a reference to this rdev, so it cannot
2210 * have been cleared recently.
2211 */
2212 rdev = conf->mirrors[d].rdev;
2213
2214 if (test_bit(Faulty, &rdev->flags))
2215 /* drive has already been failed, just ignore any
2216 more fix_read_error() attempts */
2217 return;
2218
2219 check_decay_read_errors(mddev, rdev);
2220 atomic_inc(&rdev->read_errors);
2221 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2222 char b[BDEVNAME_SIZE];
2223 bdevname(rdev->bdev, b);
2224
2225 printk(KERN_NOTICE
2226 "md/raid10:%s: %s: Raid device exceeded "
2227 "read_error threshold [cur %d:max %d]\n",
2228 mdname(mddev), b,
2229 atomic_read(&rdev->read_errors), max_read_errors);
2230 printk(KERN_NOTICE
2231 "md/raid10:%s: %s: Failing raid device\n",
2232 mdname(mddev), b);
2233 md_error(mddev, conf->mirrors[d].rdev);
2234 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2235 return;
2236 }
2237
2238 while(sectors) {
2239 int s = sectors;
2240 int sl = r10_bio->read_slot;
2241 int success = 0;
2242 int start;
2243
2244 if (s > (PAGE_SIZE>>9))
2245 s = PAGE_SIZE >> 9;
2246
2247 rcu_read_lock();
2248 do {
2249 sector_t first_bad;
2250 int bad_sectors;
2251
2252 d = r10_bio->devs[sl].devnum;
2253 rdev = rcu_dereference(conf->mirrors[d].rdev);
2254 if (rdev &&
2255 !test_bit(Unmerged, &rdev->flags) &&
2256 test_bit(In_sync, &rdev->flags) &&
2257 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2258 &first_bad, &bad_sectors) == 0) {
2259 atomic_inc(&rdev->nr_pending);
2260 rcu_read_unlock();
2261 success = sync_page_io(rdev,
2262 r10_bio->devs[sl].addr +
2263 sect,
2264 s<<9,
2265 conf->tmppage, READ, false);
2266 rdev_dec_pending(rdev, mddev);
2267 rcu_read_lock();
2268 if (success)
2269 break;
2270 }
2271 sl++;
2272 if (sl == conf->copies)
2273 sl = 0;
2274 } while (!success && sl != r10_bio->read_slot);
2275 rcu_read_unlock();
2276
2277 if (!success) {
2278 /* Cannot read from anywhere, just mark the block
2279 * as bad on the first device to discourage future
2280 * reads.
2281 */
2282 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2283 rdev = conf->mirrors[dn].rdev;
2284
2285 if (!rdev_set_badblocks(
2286 rdev,
2287 r10_bio->devs[r10_bio->read_slot].addr
2288 + sect,
2289 s, 0)) {
2290 md_error(mddev, rdev);
2291 r10_bio->devs[r10_bio->read_slot].bio
2292 = IO_BLOCKED;
2293 }
2294 break;
2295 }
2296
2297 start = sl;
2298 /* write it back and re-read */
2299 rcu_read_lock();
2300 while (sl != r10_bio->read_slot) {
2301 char b[BDEVNAME_SIZE];
2302
2303 if (sl==0)
2304 sl = conf->copies;
2305 sl--;
2306 d = r10_bio->devs[sl].devnum;
2307 rdev = rcu_dereference(conf->mirrors[d].rdev);
2308 if (!rdev ||
2309 test_bit(Unmerged, &rdev->flags) ||
2310 !test_bit(In_sync, &rdev->flags))
2311 continue;
2312
2313 atomic_inc(&rdev->nr_pending);
2314 rcu_read_unlock();
2315 if (r10_sync_page_io(rdev,
2316 r10_bio->devs[sl].addr +
2317 sect,
2318 s, conf->tmppage, WRITE)
2319 == 0) {
2320 /* Well, this device is dead */
2321 printk(KERN_NOTICE
2322 "md/raid10:%s: read correction "
2323 "write failed"
2324 " (%d sectors at %llu on %s)\n",
2325 mdname(mddev), s,
2326 (unsigned long long)(
2327 sect +
2328 choose_data_offset(r10_bio,
2329 rdev)),
2330 bdevname(rdev->bdev, b));
2331 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2332 "drive\n",
2333 mdname(mddev),
2334 bdevname(rdev->bdev, b));
2335 }
2336 rdev_dec_pending(rdev, mddev);
2337 rcu_read_lock();
2338 }
2339 sl = start;
2340 while (sl != r10_bio->read_slot) {
2341 char b[BDEVNAME_SIZE];
2342
2343 if (sl==0)
2344 sl = conf->copies;
2345 sl--;
2346 d = r10_bio->devs[sl].devnum;
2347 rdev = rcu_dereference(conf->mirrors[d].rdev);
2348 if (!rdev ||
2349 !test_bit(In_sync, &rdev->flags))
2350 continue;
2351
2352 atomic_inc(&rdev->nr_pending);
2353 rcu_read_unlock();
2354 switch (r10_sync_page_io(rdev,
2355 r10_bio->devs[sl].addr +
2356 sect,
2357 s, conf->tmppage,
2358 READ)) {
2359 case 0:
2360 /* Well, this device is dead */
2361 printk(KERN_NOTICE
2362 "md/raid10:%s: unable to read back "
2363 "corrected sectors"
2364 " (%d sectors at %llu on %s)\n",
2365 mdname(mddev), s,
2366 (unsigned long long)(
2367 sect +
2368 choose_data_offset(r10_bio, rdev)),
2369 bdevname(rdev->bdev, b));
2370 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2371 "drive\n",
2372 mdname(mddev),
2373 bdevname(rdev->bdev, b));
2374 break;
2375 case 1:
2376 printk(KERN_INFO
2377 "md/raid10:%s: read error corrected"
2378 " (%d sectors at %llu on %s)\n",
2379 mdname(mddev), s,
2380 (unsigned long long)(
2381 sect +
2382 choose_data_offset(r10_bio, rdev)),
2383 bdevname(rdev->bdev, b));
2384 atomic_add(s, &rdev->corrected_errors);
2385 }
2386
2387 rdev_dec_pending(rdev, mddev);
2388 rcu_read_lock();
2389 }
2390 rcu_read_unlock();
2391
2392 sectors -= s;
2393 sect += s;
2394 }
2395}
2396
2397static void bi_complete(struct bio *bio, int error)
2398{
2399 complete((struct completion *)bio->bi_private);
2400}
2401
2402static int submit_bio_wait(int rw, struct bio *bio)
2403{
2404 struct completion event;
2405 rw |= REQ_SYNC;
2406
2407 init_completion(&event);
2408 bio->bi_private = &event;
2409 bio->bi_end_io = bi_complete;
2410 submit_bio(rw, bio);
2411 wait_for_completion(&event);
2412
2413 return test_bit(BIO_UPTODATE, &bio->bi_flags);
2414}
2415
2416static int narrow_write_error(struct r10bio *r10_bio, int i)
2417{
2418 struct bio *bio = r10_bio->master_bio;
2419 struct mddev *mddev = r10_bio->mddev;
2420 struct r10conf *conf = mddev->private;
2421 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2422 /* bio has the data to be written to slot 'i' where
2423 * we just recently had a write error.
2424 * We repeatedly clone the bio and trim down to one block,
2425 * then try the write. Where the write fails we record
2426 * a bad block.
2427 * It is conceivable that the bio doesn't exactly align with
2428 * blocks. We must handle this.
2429 *
2430 * We currently own a reference to the rdev.
2431 */
2432
2433 int block_sectors;
2434 sector_t sector;
2435 int sectors;
2436 int sect_to_write = r10_bio->sectors;
2437 int ok = 1;
2438
2439 if (rdev->badblocks.shift < 0)
2440 return 0;
2441
2442 block_sectors = 1 << rdev->badblocks.shift;
2443 sector = r10_bio->sector;
2444 sectors = ((r10_bio->sector + block_sectors)
2445 & ~(sector_t)(block_sectors - 1))
2446 - sector;
2447
2448 while (sect_to_write) {
2449 struct bio *wbio;
2450 if (sectors > sect_to_write)
2451 sectors = sect_to_write;
2452 /* Write at 'sector' for 'sectors' */
2453 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2454 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2455 wbio->bi_sector = (r10_bio->devs[i].addr+
2456 choose_data_offset(r10_bio, rdev) +
2457 (sector - r10_bio->sector));
2458 wbio->bi_bdev = rdev->bdev;
2459 if (submit_bio_wait(WRITE, wbio) == 0)
2460 /* Failure! */
2461 ok = rdev_set_badblocks(rdev, sector,
2462 sectors, 0)
2463 && ok;
2464
2465 bio_put(wbio);
2466 sect_to_write -= sectors;
2467 sector += sectors;
2468 sectors = block_sectors;
2469 }
2470 return ok;
2471}
2472
2473static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2474{
2475 int slot = r10_bio->read_slot;
2476 struct bio *bio;
2477 struct r10conf *conf = mddev->private;
2478 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2479 char b[BDEVNAME_SIZE];
2480 unsigned long do_sync;
2481 int max_sectors;
2482
2483 /* we got a read error. Maybe the drive is bad. Maybe just
2484 * the block and we can fix it.
2485 * We freeze all other IO, and try reading the block from
2486 * other devices. When we find one, we re-write
2487 * and check it that fixes the read error.
2488 * This is all done synchronously while the array is
2489 * frozen.
2490 */
2491 bio = r10_bio->devs[slot].bio;
2492 bdevname(bio->bi_bdev, b);
2493 bio_put(bio);
2494 r10_bio->devs[slot].bio = NULL;
2495
2496 if (mddev->ro == 0) {
2497 freeze_array(conf);
2498 fix_read_error(conf, mddev, r10_bio);
2499 unfreeze_array(conf);
2500 } else
2501 r10_bio->devs[slot].bio = IO_BLOCKED;
2502
2503 rdev_dec_pending(rdev, mddev);
2504
2505read_more:
2506 rdev = read_balance(conf, r10_bio, &max_sectors);
2507 if (rdev == NULL) {
2508 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2509 " read error for block %llu\n",
2510 mdname(mddev), b,
2511 (unsigned long long)r10_bio->sector);
2512 raid_end_bio_io(r10_bio);
2513 return;
2514 }
2515
2516 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2517 slot = r10_bio->read_slot;
2518 printk_ratelimited(
2519 KERN_ERR
2520 "md/raid10:%s: %s: redirecting "
2521 "sector %llu to another mirror\n",
2522 mdname(mddev),
2523 bdevname(rdev->bdev, b),
2524 (unsigned long long)r10_bio->sector);
2525 bio = bio_clone_mddev(r10_bio->master_bio,
2526 GFP_NOIO, mddev);
2527 md_trim_bio(bio,
2528 r10_bio->sector - bio->bi_sector,
2529 max_sectors);
2530 r10_bio->devs[slot].bio = bio;
2531 r10_bio->devs[slot].rdev = rdev;
2532 bio->bi_sector = r10_bio->devs[slot].addr
2533 + choose_data_offset(r10_bio, rdev);
2534 bio->bi_bdev = rdev->bdev;
2535 bio->bi_rw = READ | do_sync;
2536 bio->bi_private = r10_bio;
2537 bio->bi_end_io = raid10_end_read_request;
2538 if (max_sectors < r10_bio->sectors) {
2539 /* Drat - have to split this up more */
2540 struct bio *mbio = r10_bio->master_bio;
2541 int sectors_handled =
2542 r10_bio->sector + max_sectors
2543 - mbio->bi_sector;
2544 r10_bio->sectors = max_sectors;
2545 spin_lock_irq(&conf->device_lock);
2546 if (mbio->bi_phys_segments == 0)
2547 mbio->bi_phys_segments = 2;
2548 else
2549 mbio->bi_phys_segments++;
2550 spin_unlock_irq(&conf->device_lock);
2551 generic_make_request(bio);
2552
2553 r10_bio = mempool_alloc(conf->r10bio_pool,
2554 GFP_NOIO);
2555 r10_bio->master_bio = mbio;
2556 r10_bio->sectors = (mbio->bi_size >> 9)
2557 - sectors_handled;
2558 r10_bio->state = 0;
2559 set_bit(R10BIO_ReadError,
2560 &r10_bio->state);
2561 r10_bio->mddev = mddev;
2562 r10_bio->sector = mbio->bi_sector
2563 + sectors_handled;
2564
2565 goto read_more;
2566 } else
2567 generic_make_request(bio);
2568}
2569
2570static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2571{
2572 /* Some sort of write request has finished and it
2573 * succeeded in writing where we thought there was a
2574 * bad block. So forget the bad block.
2575 * Or possibly if failed and we need to record
2576 * a bad block.
2577 */
2578 int m;
2579 struct md_rdev *rdev;
2580
2581 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2582 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2583 for (m = 0; m < conf->copies; m++) {
2584 int dev = r10_bio->devs[m].devnum;
2585 rdev = conf->mirrors[dev].rdev;
2586 if (r10_bio->devs[m].bio == NULL)
2587 continue;
2588 if (test_bit(BIO_UPTODATE,
2589 &r10_bio->devs[m].bio->bi_flags)) {
2590 rdev_clear_badblocks(
2591 rdev,
2592 r10_bio->devs[m].addr,
2593 r10_bio->sectors, 0);
2594 } else {
2595 if (!rdev_set_badblocks(
2596 rdev,
2597 r10_bio->devs[m].addr,
2598 r10_bio->sectors, 0))
2599 md_error(conf->mddev, rdev);
2600 }
2601 rdev = conf->mirrors[dev].replacement;
2602 if (r10_bio->devs[m].repl_bio == NULL)
2603 continue;
2604 if (test_bit(BIO_UPTODATE,
2605 &r10_bio->devs[m].repl_bio->bi_flags)) {
2606 rdev_clear_badblocks(
2607 rdev,
2608 r10_bio->devs[m].addr,
2609 r10_bio->sectors, 0);
2610 } else {
2611 if (!rdev_set_badblocks(
2612 rdev,
2613 r10_bio->devs[m].addr,
2614 r10_bio->sectors, 0))
2615 md_error(conf->mddev, rdev);
2616 }
2617 }
2618 put_buf(r10_bio);
2619 } else {
2620 for (m = 0; m < conf->copies; m++) {
2621 int dev = r10_bio->devs[m].devnum;
2622 struct bio *bio = r10_bio->devs[m].bio;
2623 rdev = conf->mirrors[dev].rdev;
2624 if (bio == IO_MADE_GOOD) {
2625 rdev_clear_badblocks(
2626 rdev,
2627 r10_bio->devs[m].addr,
2628 r10_bio->sectors, 0);
2629 rdev_dec_pending(rdev, conf->mddev);
2630 } else if (bio != NULL &&
2631 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2632 if (!narrow_write_error(r10_bio, m)) {
2633 md_error(conf->mddev, rdev);
2634 set_bit(R10BIO_Degraded,
2635 &r10_bio->state);
2636 }
2637 rdev_dec_pending(rdev, conf->mddev);
2638 }
2639 bio = r10_bio->devs[m].repl_bio;
2640 rdev = conf->mirrors[dev].replacement;
2641 if (rdev && bio == IO_MADE_GOOD) {
2642 rdev_clear_badblocks(
2643 rdev,
2644 r10_bio->devs[m].addr,
2645 r10_bio->sectors, 0);
2646 rdev_dec_pending(rdev, conf->mddev);
2647 }
2648 }
2649 if (test_bit(R10BIO_WriteError,
2650 &r10_bio->state))
2651 close_write(r10_bio);
2652 raid_end_bio_io(r10_bio);
2653 }
2654}
2655
2656static void raid10d(struct mddev *mddev)
2657{
2658 struct r10bio *r10_bio;
2659 unsigned long flags;
2660 struct r10conf *conf = mddev->private;
2661 struct list_head *head = &conf->retry_list;
2662 struct blk_plug plug;
2663
2664 md_check_recovery(mddev);
2665
2666 blk_start_plug(&plug);
2667 for (;;) {
2668
2669 if (atomic_read(&mddev->plug_cnt) == 0)
2670 flush_pending_writes(conf);
2671
2672 spin_lock_irqsave(&conf->device_lock, flags);
2673 if (list_empty(head)) {
2674 spin_unlock_irqrestore(&conf->device_lock, flags);
2675 break;
2676 }
2677 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2678 list_del(head->prev);
2679 conf->nr_queued--;
2680 spin_unlock_irqrestore(&conf->device_lock, flags);
2681
2682 mddev = r10_bio->mddev;
2683 conf = mddev->private;
2684 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2685 test_bit(R10BIO_WriteError, &r10_bio->state))
2686 handle_write_completed(conf, r10_bio);
2687 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2688 reshape_request_write(mddev, r10_bio);
2689 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2690 sync_request_write(mddev, r10_bio);
2691 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2692 recovery_request_write(mddev, r10_bio);
2693 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2694 handle_read_error(mddev, r10_bio);
2695 else {
2696 /* just a partial read to be scheduled from a
2697 * separate context
2698 */
2699 int slot = r10_bio->read_slot;
2700 generic_make_request(r10_bio->devs[slot].bio);
2701 }
2702
2703 cond_resched();
2704 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2705 md_check_recovery(mddev);
2706 }
2707 blk_finish_plug(&plug);
2708}
2709
2710
2711static int init_resync(struct r10conf *conf)
2712{
2713 int buffs;
2714 int i;
2715
2716 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2717 BUG_ON(conf->r10buf_pool);
2718 conf->have_replacement = 0;
2719 for (i = 0; i < conf->geo.raid_disks; i++)
2720 if (conf->mirrors[i].replacement)
2721 conf->have_replacement = 1;
2722 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2723 if (!conf->r10buf_pool)
2724 return -ENOMEM;
2725 conf->next_resync = 0;
2726 return 0;
2727}
2728
2729/*
2730 * perform a "sync" on one "block"
2731 *
2732 * We need to make sure that no normal I/O request - particularly write
2733 * requests - conflict with active sync requests.
2734 *
2735 * This is achieved by tracking pending requests and a 'barrier' concept
2736 * that can be installed to exclude normal IO requests.
2737 *
2738 * Resync and recovery are handled very differently.
2739 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2740 *
2741 * For resync, we iterate over virtual addresses, read all copies,
2742 * and update if there are differences. If only one copy is live,
2743 * skip it.
2744 * For recovery, we iterate over physical addresses, read a good
2745 * value for each non-in_sync drive, and over-write.
2746 *
2747 * So, for recovery we may have several outstanding complex requests for a
2748 * given address, one for each out-of-sync device. We model this by allocating
2749 * a number of r10_bio structures, one for each out-of-sync device.
2750 * As we setup these structures, we collect all bio's together into a list
2751 * which we then process collectively to add pages, and then process again
2752 * to pass to generic_make_request.
2753 *
2754 * The r10_bio structures are linked using a borrowed master_bio pointer.
2755 * This link is counted in ->remaining. When the r10_bio that points to NULL
2756 * has its remaining count decremented to 0, the whole complex operation
2757 * is complete.
2758 *
2759 */
2760
2761static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2762 int *skipped, int go_faster)
2763{
2764 struct r10conf *conf = mddev->private;
2765 struct r10bio *r10_bio;
2766 struct bio *biolist = NULL, *bio;
2767 sector_t max_sector, nr_sectors;
2768 int i;
2769 int max_sync;
2770 sector_t sync_blocks;
2771 sector_t sectors_skipped = 0;
2772 int chunks_skipped = 0;
2773 sector_t chunk_mask = conf->geo.chunk_mask;
2774
2775 if (!conf->r10buf_pool)
2776 if (init_resync(conf))
2777 return 0;
2778
2779 skipped:
2780 max_sector = mddev->dev_sectors;
2781 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2782 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2783 max_sector = mddev->resync_max_sectors;
2784 if (sector_nr >= max_sector) {
2785 /* If we aborted, we need to abort the
2786 * sync on the 'current' bitmap chucks (there can
2787 * be several when recovering multiple devices).
2788 * as we may have started syncing it but not finished.
2789 * We can find the current address in
2790 * mddev->curr_resync, but for recovery,
2791 * we need to convert that to several
2792 * virtual addresses.
2793 */
2794 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2795 end_reshape(conf);
2796 return 0;
2797 }
2798
2799 if (mddev->curr_resync < max_sector) { /* aborted */
2800 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2801 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2802 &sync_blocks, 1);
2803 else for (i = 0; i < conf->geo.raid_disks; i++) {
2804 sector_t sect =
2805 raid10_find_virt(conf, mddev->curr_resync, i);
2806 bitmap_end_sync(mddev->bitmap, sect,
2807 &sync_blocks, 1);
2808 }
2809 } else {
2810 /* completed sync */
2811 if ((!mddev->bitmap || conf->fullsync)
2812 && conf->have_replacement
2813 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2814 /* Completed a full sync so the replacements
2815 * are now fully recovered.
2816 */
2817 for (i = 0; i < conf->geo.raid_disks; i++)
2818 if (conf->mirrors[i].replacement)
2819 conf->mirrors[i].replacement
2820 ->recovery_offset
2821 = MaxSector;
2822 }
2823 conf->fullsync = 0;
2824 }
2825 bitmap_close_sync(mddev->bitmap);
2826 close_sync(conf);
2827 *skipped = 1;
2828 return sectors_skipped;
2829 }
2830
2831 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2832 return reshape_request(mddev, sector_nr, skipped);
2833
2834 if (chunks_skipped >= conf->geo.raid_disks) {
2835 /* if there has been nothing to do on any drive,
2836 * then there is nothing to do at all..
2837 */
2838 *skipped = 1;
2839 return (max_sector - sector_nr) + sectors_skipped;
2840 }
2841
2842 if (max_sector > mddev->resync_max)
2843 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2844
2845 /* make sure whole request will fit in a chunk - if chunks
2846 * are meaningful
2847 */
2848 if (conf->geo.near_copies < conf->geo.raid_disks &&
2849 max_sector > (sector_nr | chunk_mask))
2850 max_sector = (sector_nr | chunk_mask) + 1;
2851 /*
2852 * If there is non-resync activity waiting for us then
2853 * put in a delay to throttle resync.
2854 */
2855 if (!go_faster && conf->nr_waiting)
2856 msleep_interruptible(1000);
2857
2858 /* Again, very different code for resync and recovery.
2859 * Both must result in an r10bio with a list of bios that
2860 * have bi_end_io, bi_sector, bi_bdev set,
2861 * and bi_private set to the r10bio.
2862 * For recovery, we may actually create several r10bios
2863 * with 2 bios in each, that correspond to the bios in the main one.
2864 * In this case, the subordinate r10bios link back through a
2865 * borrowed master_bio pointer, and the counter in the master
2866 * includes a ref from each subordinate.
2867 */
2868 /* First, we decide what to do and set ->bi_end_io
2869 * To end_sync_read if we want to read, and
2870 * end_sync_write if we will want to write.
2871 */
2872
2873 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2874 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2875 /* recovery... the complicated one */
2876 int j;
2877 r10_bio = NULL;
2878
2879 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2880 int still_degraded;
2881 struct r10bio *rb2;
2882 sector_t sect;
2883 int must_sync;
2884 int any_working;
2885 struct mirror_info *mirror = &conf->mirrors[i];
2886
2887 if ((mirror->rdev == NULL ||
2888 test_bit(In_sync, &mirror->rdev->flags))
2889 &&
2890 (mirror->replacement == NULL ||
2891 test_bit(Faulty,
2892 &mirror->replacement->flags)))
2893 continue;
2894
2895 still_degraded = 0;
2896 /* want to reconstruct this device */
2897 rb2 = r10_bio;
2898 sect = raid10_find_virt(conf, sector_nr, i);
2899 if (sect >= mddev->resync_max_sectors) {
2900 /* last stripe is not complete - don't
2901 * try to recover this sector.
2902 */
2903 continue;
2904 }
2905 /* Unless we are doing a full sync, or a replacement
2906 * we only need to recover the block if it is set in
2907 * the bitmap
2908 */
2909 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2910 &sync_blocks, 1);
2911 if (sync_blocks < max_sync)
2912 max_sync = sync_blocks;
2913 if (!must_sync &&
2914 mirror->replacement == NULL &&
2915 !conf->fullsync) {
2916 /* yep, skip the sync_blocks here, but don't assume
2917 * that there will never be anything to do here
2918 */
2919 chunks_skipped = -1;
2920 continue;
2921 }
2922
2923 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2924 raise_barrier(conf, rb2 != NULL);
2925 atomic_set(&r10_bio->remaining, 0);
2926
2927 r10_bio->master_bio = (struct bio*)rb2;
2928 if (rb2)
2929 atomic_inc(&rb2->remaining);
2930 r10_bio->mddev = mddev;
2931 set_bit(R10BIO_IsRecover, &r10_bio->state);
2932 r10_bio->sector = sect;
2933
2934 raid10_find_phys(conf, r10_bio);
2935
2936 /* Need to check if the array will still be
2937 * degraded
2938 */
2939 for (j = 0; j < conf->geo.raid_disks; j++)
2940 if (conf->mirrors[j].rdev == NULL ||
2941 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2942 still_degraded = 1;
2943 break;
2944 }
2945
2946 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2947 &sync_blocks, still_degraded);
2948
2949 any_working = 0;
2950 for (j=0; j<conf->copies;j++) {
2951 int k;
2952 int d = r10_bio->devs[j].devnum;
2953 sector_t from_addr, to_addr;
2954 struct md_rdev *rdev;
2955 sector_t sector, first_bad;
2956 int bad_sectors;
2957 if (!conf->mirrors[d].rdev ||
2958 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2959 continue;
2960 /* This is where we read from */
2961 any_working = 1;
2962 rdev = conf->mirrors[d].rdev;
2963 sector = r10_bio->devs[j].addr;
2964
2965 if (is_badblock(rdev, sector, max_sync,
2966 &first_bad, &bad_sectors)) {
2967 if (first_bad > sector)
2968 max_sync = first_bad - sector;
2969 else {
2970 bad_sectors -= (sector
2971 - first_bad);
2972 if (max_sync > bad_sectors)
2973 max_sync = bad_sectors;
2974 continue;
2975 }
2976 }
2977 bio = r10_bio->devs[0].bio;
2978 bio->bi_next = biolist;
2979 biolist = bio;
2980 bio->bi_private = r10_bio;
2981 bio->bi_end_io = end_sync_read;
2982 bio->bi_rw = READ;
2983 from_addr = r10_bio->devs[j].addr;
2984 bio->bi_sector = from_addr + rdev->data_offset;
2985 bio->bi_bdev = rdev->bdev;
2986 atomic_inc(&rdev->nr_pending);
2987 /* and we write to 'i' (if not in_sync) */
2988
2989 for (k=0; k<conf->copies; k++)
2990 if (r10_bio->devs[k].devnum == i)
2991 break;
2992 BUG_ON(k == conf->copies);
2993 to_addr = r10_bio->devs[k].addr;
2994 r10_bio->devs[0].devnum = d;
2995 r10_bio->devs[0].addr = from_addr;
2996 r10_bio->devs[1].devnum = i;
2997 r10_bio->devs[1].addr = to_addr;
2998
2999 rdev = mirror->rdev;
3000 if (!test_bit(In_sync, &rdev->flags)) {
3001 bio = r10_bio->devs[1].bio;
3002 bio->bi_next = biolist;
3003 biolist = bio;
3004 bio->bi_private = r10_bio;
3005 bio->bi_end_io = end_sync_write;
3006 bio->bi_rw = WRITE;
3007 bio->bi_sector = to_addr
3008 + rdev->data_offset;
3009 bio->bi_bdev = rdev->bdev;
3010 atomic_inc(&r10_bio->remaining);
3011 } else
3012 r10_bio->devs[1].bio->bi_end_io = NULL;
3013
3014 /* and maybe write to replacement */
3015 bio = r10_bio->devs[1].repl_bio;
3016 if (bio)
3017 bio->bi_end_io = NULL;
3018 rdev = mirror->replacement;
3019 /* Note: if rdev != NULL, then bio
3020 * cannot be NULL as r10buf_pool_alloc will
3021 * have allocated it.
3022 * So the second test here is pointless.
3023 * But it keeps semantic-checkers happy, and
3024 * this comment keeps human reviewers
3025 * happy.
3026 */
3027 if (rdev == NULL || bio == NULL ||
3028 test_bit(Faulty, &rdev->flags))
3029 break;
3030 bio->bi_next = biolist;
3031 biolist = bio;
3032 bio->bi_private = r10_bio;
3033 bio->bi_end_io = end_sync_write;
3034 bio->bi_rw = WRITE;
3035 bio->bi_sector = to_addr + rdev->data_offset;
3036 bio->bi_bdev = rdev->bdev;
3037 atomic_inc(&r10_bio->remaining);
3038 break;
3039 }
3040 if (j == conf->copies) {
3041 /* Cannot recover, so abort the recovery or
3042 * record a bad block */
3043 put_buf(r10_bio);
3044 if (rb2)
3045 atomic_dec(&rb2->remaining);
3046 r10_bio = rb2;
3047 if (any_working) {
3048 /* problem is that there are bad blocks
3049 * on other device(s)
3050 */
3051 int k;
3052 for (k = 0; k < conf->copies; k++)
3053 if (r10_bio->devs[k].devnum == i)
3054 break;
3055 if (!test_bit(In_sync,
3056 &mirror->rdev->flags)
3057 && !rdev_set_badblocks(
3058 mirror->rdev,
3059 r10_bio->devs[k].addr,
3060 max_sync, 0))
3061 any_working = 0;
3062 if (mirror->replacement &&
3063 !rdev_set_badblocks(
3064 mirror->replacement,
3065 r10_bio->devs[k].addr,
3066 max_sync, 0))
3067 any_working = 0;
3068 }
3069 if (!any_working) {
3070 if (!test_and_set_bit(MD_RECOVERY_INTR,
3071 &mddev->recovery))
3072 printk(KERN_INFO "md/raid10:%s: insufficient "
3073 "working devices for recovery.\n",
3074 mdname(mddev));
3075 mirror->recovery_disabled
3076 = mddev->recovery_disabled;
3077 }
3078 break;
3079 }
3080 }
3081 if (biolist == NULL) {
3082 while (r10_bio) {
3083 struct r10bio *rb2 = r10_bio;
3084 r10_bio = (struct r10bio*) rb2->master_bio;
3085 rb2->master_bio = NULL;
3086 put_buf(rb2);
3087 }
3088 goto giveup;
3089 }
3090 } else {
3091 /* resync. Schedule a read for every block at this virt offset */
3092 int count = 0;
3093
3094 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3095
3096 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3097 &sync_blocks, mddev->degraded) &&
3098 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3099 &mddev->recovery)) {
3100 /* We can skip this block */
3101 *skipped = 1;
3102 return sync_blocks + sectors_skipped;
3103 }
3104 if (sync_blocks < max_sync)
3105 max_sync = sync_blocks;
3106 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3107
3108 r10_bio->mddev = mddev;
3109 atomic_set(&r10_bio->remaining, 0);
3110 raise_barrier(conf, 0);
3111 conf->next_resync = sector_nr;
3112
3113 r10_bio->master_bio = NULL;
3114 r10_bio->sector = sector_nr;
3115 set_bit(R10BIO_IsSync, &r10_bio->state);
3116 raid10_find_phys(conf, r10_bio);
3117 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3118
3119 for (i = 0; i < conf->copies; i++) {
3120 int d = r10_bio->devs[i].devnum;
3121 sector_t first_bad, sector;
3122 int bad_sectors;
3123
3124 if (r10_bio->devs[i].repl_bio)
3125 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3126
3127 bio = r10_bio->devs[i].bio;
3128 bio->bi_end_io = NULL;
3129 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3130 if (conf->mirrors[d].rdev == NULL ||
3131 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3132 continue;
3133 sector = r10_bio->devs[i].addr;
3134 if (is_badblock(conf->mirrors[d].rdev,
3135 sector, max_sync,
3136 &first_bad, &bad_sectors)) {
3137 if (first_bad > sector)
3138 max_sync = first_bad - sector;
3139 else {
3140 bad_sectors -= (sector - first_bad);
3141 if (max_sync > bad_sectors)
3142 max_sync = max_sync;
3143 continue;
3144 }
3145 }
3146 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3147 atomic_inc(&r10_bio->remaining);
3148 bio->bi_next = biolist;
3149 biolist = bio;
3150 bio->bi_private = r10_bio;
3151 bio->bi_end_io = end_sync_read;
3152 bio->bi_rw = READ;
3153 bio->bi_sector = sector +
3154 conf->mirrors[d].rdev->data_offset;
3155 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3156 count++;
3157
3158 if (conf->mirrors[d].replacement == NULL ||
3159 test_bit(Faulty,
3160 &conf->mirrors[d].replacement->flags))
3161 continue;
3162
3163 /* Need to set up for writing to the replacement */
3164 bio = r10_bio->devs[i].repl_bio;
3165 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3166
3167 sector = r10_bio->devs[i].addr;
3168 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3169 bio->bi_next = biolist;
3170 biolist = bio;
3171 bio->bi_private = r10_bio;
3172 bio->bi_end_io = end_sync_write;
3173 bio->bi_rw = WRITE;
3174 bio->bi_sector = sector +
3175 conf->mirrors[d].replacement->data_offset;
3176 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3177 count++;
3178 }
3179
3180 if (count < 2) {
3181 for (i=0; i<conf->copies; i++) {
3182 int d = r10_bio->devs[i].devnum;
3183 if (r10_bio->devs[i].bio->bi_end_io)
3184 rdev_dec_pending(conf->mirrors[d].rdev,
3185 mddev);
3186 if (r10_bio->devs[i].repl_bio &&
3187 r10_bio->devs[i].repl_bio->bi_end_io)
3188 rdev_dec_pending(
3189 conf->mirrors[d].replacement,
3190 mddev);
3191 }
3192 put_buf(r10_bio);
3193 biolist = NULL;
3194 goto giveup;
3195 }
3196 }
3197
3198 for (bio = biolist; bio ; bio=bio->bi_next) {
3199
3200 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
3201 if (bio->bi_end_io)
3202 bio->bi_flags |= 1 << BIO_UPTODATE;
3203 bio->bi_vcnt = 0;
3204 bio->bi_idx = 0;
3205 bio->bi_phys_segments = 0;
3206 bio->bi_size = 0;
3207 }
3208
3209 nr_sectors = 0;
3210 if (sector_nr + max_sync < max_sector)
3211 max_sector = sector_nr + max_sync;
3212 do {
3213 struct page *page;
3214 int len = PAGE_SIZE;
3215 if (sector_nr + (len>>9) > max_sector)
3216 len = (max_sector - sector_nr) << 9;
3217 if (len == 0)
3218 break;
3219 for (bio= biolist ; bio ; bio=bio->bi_next) {
3220 struct bio *bio2;
3221 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3222 if (bio_add_page(bio, page, len, 0))
3223 continue;
3224
3225 /* stop here */
3226 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3227 for (bio2 = biolist;
3228 bio2 && bio2 != bio;
3229 bio2 = bio2->bi_next) {
3230 /* remove last page from this bio */
3231 bio2->bi_vcnt--;
3232 bio2->bi_size -= len;
3233 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3234 }
3235 goto bio_full;
3236 }
3237 nr_sectors += len>>9;
3238 sector_nr += len>>9;
3239 } while (biolist->bi_vcnt < RESYNC_PAGES);
3240 bio_full:
3241 r10_bio->sectors = nr_sectors;
3242
3243 while (biolist) {
3244 bio = biolist;
3245 biolist = biolist->bi_next;
3246
3247 bio->bi_next = NULL;
3248 r10_bio = bio->bi_private;
3249 r10_bio->sectors = nr_sectors;
3250
3251 if (bio->bi_end_io == end_sync_read) {
3252 md_sync_acct(bio->bi_bdev, nr_sectors);
3253 generic_make_request(bio);
3254 }
3255 }
3256
3257 if (sectors_skipped)
3258 /* pretend they weren't skipped, it makes
3259 * no important difference in this case
3260 */
3261 md_done_sync(mddev, sectors_skipped, 1);
3262
3263 return sectors_skipped + nr_sectors;
3264 giveup:
3265 /* There is nowhere to write, so all non-sync
3266 * drives must be failed or in resync, all drives
3267 * have a bad block, so try the next chunk...
3268 */
3269 if (sector_nr + max_sync < max_sector)
3270 max_sector = sector_nr + max_sync;
3271
3272 sectors_skipped += (max_sector - sector_nr);
3273 chunks_skipped ++;
3274 sector_nr = max_sector;
3275 goto skipped;
3276}
3277
3278static sector_t
3279raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3280{
3281 sector_t size;
3282 struct r10conf *conf = mddev->private;
3283
3284 if (!raid_disks)
3285 raid_disks = min(conf->geo.raid_disks,
3286 conf->prev.raid_disks);
3287 if (!sectors)
3288 sectors = conf->dev_sectors;
3289
3290 size = sectors >> conf->geo.chunk_shift;
3291 sector_div(size, conf->geo.far_copies);
3292 size = size * raid_disks;
3293 sector_div(size, conf->geo.near_copies);
3294
3295 return size << conf->geo.chunk_shift;
3296}
3297
3298static void calc_sectors(struct r10conf *conf, sector_t size)
3299{
3300 /* Calculate the number of sectors-per-device that will
3301 * actually be used, and set conf->dev_sectors and
3302 * conf->stride
3303 */
3304
3305 size = size >> conf->geo.chunk_shift;
3306 sector_div(size, conf->geo.far_copies);
3307 size = size * conf->geo.raid_disks;
3308 sector_div(size, conf->geo.near_copies);
3309 /* 'size' is now the number of chunks in the array */
3310 /* calculate "used chunks per device" */
3311 size = size * conf->copies;
3312
3313 /* We need to round up when dividing by raid_disks to
3314 * get the stride size.
3315 */
3316 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3317
3318 conf->dev_sectors = size << conf->geo.chunk_shift;
3319
3320 if (conf->geo.far_offset)
3321 conf->geo.stride = 1 << conf->geo.chunk_shift;
3322 else {
3323 sector_div(size, conf->geo.far_copies);
3324 conf->geo.stride = size << conf->geo.chunk_shift;
3325 }
3326}
3327
3328enum geo_type {geo_new, geo_old, geo_start};
3329static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3330{
3331 int nc, fc, fo;
3332 int layout, chunk, disks;
3333 switch (new) {
3334 case geo_old:
3335 layout = mddev->layout;
3336 chunk = mddev->chunk_sectors;
3337 disks = mddev->raid_disks - mddev->delta_disks;
3338 break;
3339 case geo_new:
3340 layout = mddev->new_layout;
3341 chunk = mddev->new_chunk_sectors;
3342 disks = mddev->raid_disks;
3343 break;
3344 default: /* avoid 'may be unused' warnings */
3345 case geo_start: /* new when starting reshape - raid_disks not
3346 * updated yet. */
3347 layout = mddev->new_layout;
3348 chunk = mddev->new_chunk_sectors;
3349 disks = mddev->raid_disks + mddev->delta_disks;
3350 break;
3351 }
3352 if (layout >> 17)
3353 return -1;
3354 if (chunk < (PAGE_SIZE >> 9) ||
3355 !is_power_of_2(chunk))
3356 return -2;
3357 nc = layout & 255;
3358 fc = (layout >> 8) & 255;
3359 fo = layout & (1<<16);
3360 geo->raid_disks = disks;
3361 geo->near_copies = nc;
3362 geo->far_copies = fc;
3363 geo->far_offset = fo;
3364 geo->chunk_mask = chunk - 1;
3365 geo->chunk_shift = ffz(~chunk);
3366 return nc*fc;
3367}
3368
3369static struct r10conf *setup_conf(struct mddev *mddev)
3370{
3371 struct r10conf *conf = NULL;
3372 int err = -EINVAL;
3373 struct geom geo;
3374 int copies;
3375
3376 copies = setup_geo(&geo, mddev, geo_new);
3377
3378 if (copies == -2) {
3379 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3380 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3381 mdname(mddev), PAGE_SIZE);
3382 goto out;
3383 }
3384
3385 if (copies < 2 || copies > mddev->raid_disks) {
3386 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3387 mdname(mddev), mddev->new_layout);
3388 goto out;
3389 }
3390
3391 err = -ENOMEM;
3392 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3393 if (!conf)
3394 goto out;
3395
3396 /* FIXME calc properly */
3397 conf->mirrors = kzalloc(sizeof(struct mirror_info)*(mddev->raid_disks +
3398 max(0,mddev->delta_disks)),
3399 GFP_KERNEL);
3400 if (!conf->mirrors)
3401 goto out;
3402
3403 conf->tmppage = alloc_page(GFP_KERNEL);
3404 if (!conf->tmppage)
3405 goto out;
3406
3407 conf->geo = geo;
3408 conf->copies = copies;
3409 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3410 r10bio_pool_free, conf);
3411 if (!conf->r10bio_pool)
3412 goto out;
3413
3414 calc_sectors(conf, mddev->dev_sectors);
3415 if (mddev->reshape_position == MaxSector) {
3416 conf->prev = conf->geo;
3417 conf->reshape_progress = MaxSector;
3418 } else {
3419 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3420 err = -EINVAL;
3421 goto out;
3422 }
3423 conf->reshape_progress = mddev->reshape_position;
3424 if (conf->prev.far_offset)
3425 conf->prev.stride = 1 << conf->prev.chunk_shift;
3426 else
3427 /* far_copies must be 1 */
3428 conf->prev.stride = conf->dev_sectors;
3429 }
3430 spin_lock_init(&conf->device_lock);
3431 INIT_LIST_HEAD(&conf->retry_list);
3432
3433 spin_lock_init(&conf->resync_lock);
3434 init_waitqueue_head(&conf->wait_barrier);
3435
3436 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3437 if (!conf->thread)
3438 goto out;
3439
3440 conf->mddev = mddev;
3441 return conf;
3442
3443 out:
3444 if (err == -ENOMEM)
3445 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3446 mdname(mddev));
3447 if (conf) {
3448 if (conf->r10bio_pool)
3449 mempool_destroy(conf->r10bio_pool);
3450 kfree(conf->mirrors);
3451 safe_put_page(conf->tmppage);
3452 kfree(conf);
3453 }
3454 return ERR_PTR(err);
3455}
3456
3457static int run(struct mddev *mddev)
3458{
3459 struct r10conf *conf;
3460 int i, disk_idx, chunk_size;
3461 struct mirror_info *disk;
3462 struct md_rdev *rdev;
3463 sector_t size;
3464 sector_t min_offset_diff = 0;
3465 int first = 1;
3466
3467 if (mddev->private == NULL) {
3468 conf = setup_conf(mddev);
3469 if (IS_ERR(conf))
3470 return PTR_ERR(conf);
3471 mddev->private = conf;
3472 }
3473 conf = mddev->private;
3474 if (!conf)
3475 goto out;
3476
3477 mddev->thread = conf->thread;
3478 conf->thread = NULL;
3479
3480 chunk_size = mddev->chunk_sectors << 9;
3481 blk_queue_io_min(mddev->queue, chunk_size);
3482 if (conf->geo.raid_disks % conf->geo.near_copies)
3483 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3484 else
3485 blk_queue_io_opt(mddev->queue, chunk_size *
3486 (conf->geo.raid_disks / conf->geo.near_copies));
3487
3488 rdev_for_each(rdev, mddev) {
3489 long long diff;
3490 struct request_queue *q;
3491
3492 disk_idx = rdev->raid_disk;
3493 if (disk_idx < 0)
3494 continue;
3495 if (disk_idx >= conf->geo.raid_disks &&
3496 disk_idx >= conf->prev.raid_disks)
3497 continue;
3498 disk = conf->mirrors + disk_idx;
3499
3500 if (test_bit(Replacement, &rdev->flags)) {
3501 if (disk->replacement)
3502 goto out_free_conf;
3503 disk->replacement = rdev;
3504 } else {
3505 if (disk->rdev)
3506 goto out_free_conf;
3507 disk->rdev = rdev;
3508 }
3509 q = bdev_get_queue(rdev->bdev);
3510 if (q->merge_bvec_fn)
3511 mddev->merge_check_needed = 1;
3512 diff = (rdev->new_data_offset - rdev->data_offset);
3513 if (!mddev->reshape_backwards)
3514 diff = -diff;
3515 if (diff < 0)
3516 diff = 0;
3517 if (first || diff < min_offset_diff)
3518 min_offset_diff = diff;
3519
3520 disk_stack_limits(mddev->gendisk, rdev->bdev,
3521 rdev->data_offset << 9);
3522
3523 disk->head_position = 0;
3524 }
3525
3526 /* need to check that every block has at least one working mirror */
3527 if (!enough(conf, -1)) {
3528 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3529 mdname(mddev));
3530 goto out_free_conf;
3531 }
3532
3533 if (conf->reshape_progress != MaxSector) {
3534 /* must ensure that shape change is supported */
3535 if (conf->geo.far_copies != 1 &&
3536 conf->geo.far_offset == 0)
3537 goto out_free_conf;
3538 if (conf->prev.far_copies != 1 &&
3539 conf->geo.far_offset == 0)
3540 goto out_free_conf;
3541 }
3542
3543 mddev->degraded = 0;
3544 for (i = 0;
3545 i < conf->geo.raid_disks
3546 || i < conf->prev.raid_disks;
3547 i++) {
3548
3549 disk = conf->mirrors + i;
3550
3551 if (!disk->rdev && disk->replacement) {
3552 /* The replacement is all we have - use it */
3553 disk->rdev = disk->replacement;
3554 disk->replacement = NULL;
3555 clear_bit(Replacement, &disk->rdev->flags);
3556 }
3557
3558 if (!disk->rdev ||
3559 !test_bit(In_sync, &disk->rdev->flags)) {
3560 disk->head_position = 0;
3561 mddev->degraded++;
3562 if (disk->rdev)
3563 conf->fullsync = 1;
3564 }
3565 disk->recovery_disabled = mddev->recovery_disabled - 1;
3566 }
3567
3568 if (mddev->recovery_cp != MaxSector)
3569 printk(KERN_NOTICE "md/raid10:%s: not clean"
3570 " -- starting background reconstruction\n",
3571 mdname(mddev));
3572 printk(KERN_INFO
3573 "md/raid10:%s: active with %d out of %d devices\n",
3574 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3575 conf->geo.raid_disks);
3576 /*
3577 * Ok, everything is just fine now
3578 */
3579 mddev->dev_sectors = conf->dev_sectors;
3580 size = raid10_size(mddev, 0, 0);
3581 md_set_array_sectors(mddev, size);
3582 mddev->resync_max_sectors = size;
3583
3584 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3585 mddev->queue->backing_dev_info.congested_data = mddev;
3586
3587 /* Calculate max read-ahead size.
3588 * We need to readahead at least twice a whole stripe....
3589 * maybe...
3590 */
3591 {
3592 int stripe = conf->geo.raid_disks *
3593 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3594 stripe /= conf->geo.near_copies;
3595 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3596 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3597 }
3598
3599 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3600
3601 if (md_integrity_register(mddev))
3602 goto out_free_conf;
3603
3604 if (conf->reshape_progress != MaxSector) {
3605 unsigned long before_length, after_length;
3606
3607 before_length = ((1 << conf->prev.chunk_shift) *
3608 conf->prev.far_copies);
3609 after_length = ((1 << conf->geo.chunk_shift) *
3610 conf->geo.far_copies);
3611
3612 if (max(before_length, after_length) > min_offset_diff) {
3613 /* This cannot work */
3614 printk("md/raid10: offset difference not enough to continue reshape\n");
3615 goto out_free_conf;
3616 }
3617 conf->offset_diff = min_offset_diff;
3618
3619 conf->reshape_safe = conf->reshape_progress;
3620 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3621 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3622 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3623 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3624 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3625 "reshape");
3626 }
3627
3628 return 0;
3629
3630out_free_conf:
3631 md_unregister_thread(&mddev->thread);
3632 if (conf->r10bio_pool)
3633 mempool_destroy(conf->r10bio_pool);
3634 safe_put_page(conf->tmppage);
3635 kfree(conf->mirrors);
3636 kfree(conf);
3637 mddev->private = NULL;
3638out:
3639 return -EIO;
3640}
3641
3642static int stop(struct mddev *mddev)
3643{
3644 struct r10conf *conf = mddev->private;
3645
3646 raise_barrier(conf, 0);
3647 lower_barrier(conf);
3648
3649 md_unregister_thread(&mddev->thread);
3650 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
3651 if (conf->r10bio_pool)
3652 mempool_destroy(conf->r10bio_pool);
3653 kfree(conf->mirrors);
3654 kfree(conf);
3655 mddev->private = NULL;
3656 return 0;
3657}
3658
3659static void raid10_quiesce(struct mddev *mddev, int state)
3660{
3661 struct r10conf *conf = mddev->private;
3662
3663 switch(state) {
3664 case 1:
3665 raise_barrier(conf, 0);
3666 break;
3667 case 0:
3668 lower_barrier(conf);
3669 break;
3670 }
3671}
3672
3673static int raid10_resize(struct mddev *mddev, sector_t sectors)
3674{
3675 /* Resize of 'far' arrays is not supported.
3676 * For 'near' and 'offset' arrays we can set the
3677 * number of sectors used to be an appropriate multiple
3678 * of the chunk size.
3679 * For 'offset', this is far_copies*chunksize.
3680 * For 'near' the multiplier is the LCM of
3681 * near_copies and raid_disks.
3682 * So if far_copies > 1 && !far_offset, fail.
3683 * Else find LCM(raid_disks, near_copy)*far_copies and
3684 * multiply by chunk_size. Then round to this number.
3685 * This is mostly done by raid10_size()
3686 */
3687 struct r10conf *conf = mddev->private;
3688 sector_t oldsize, size;
3689
3690 if (mddev->reshape_position != MaxSector)
3691 return -EBUSY;
3692
3693 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3694 return -EINVAL;
3695
3696 oldsize = raid10_size(mddev, 0, 0);
3697 size = raid10_size(mddev, sectors, 0);
3698 if (mddev->external_size &&
3699 mddev->array_sectors > size)
3700 return -EINVAL;
3701 if (mddev->bitmap) {
3702 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3703 if (ret)
3704 return ret;
3705 }
3706 md_set_array_sectors(mddev, size);
3707 set_capacity(mddev->gendisk, mddev->array_sectors);
3708 revalidate_disk(mddev->gendisk);
3709 if (sectors > mddev->dev_sectors &&
3710 mddev->recovery_cp > oldsize) {
3711 mddev->recovery_cp = oldsize;
3712 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3713 }
3714 calc_sectors(conf, sectors);
3715 mddev->dev_sectors = conf->dev_sectors;
3716 mddev->resync_max_sectors = size;
3717 return 0;
3718}
3719
3720static void *raid10_takeover_raid0(struct mddev *mddev)
3721{
3722 struct md_rdev *rdev;
3723 struct r10conf *conf;
3724
3725 if (mddev->degraded > 0) {
3726 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3727 mdname(mddev));
3728 return ERR_PTR(-EINVAL);
3729 }
3730
3731 /* Set new parameters */
3732 mddev->new_level = 10;
3733 /* new layout: far_copies = 1, near_copies = 2 */
3734 mddev->new_layout = (1<<8) + 2;
3735 mddev->new_chunk_sectors = mddev->chunk_sectors;
3736 mddev->delta_disks = mddev->raid_disks;
3737 mddev->raid_disks *= 2;
3738 /* make sure it will be not marked as dirty */
3739 mddev->recovery_cp = MaxSector;
3740
3741 conf = setup_conf(mddev);
3742 if (!IS_ERR(conf)) {
3743 rdev_for_each(rdev, mddev)
3744 if (rdev->raid_disk >= 0)
3745 rdev->new_raid_disk = rdev->raid_disk * 2;
3746 conf->barrier = 1;
3747 }
3748
3749 return conf;
3750}
3751
3752static void *raid10_takeover(struct mddev *mddev)
3753{
3754 struct r0conf *raid0_conf;
3755
3756 /* raid10 can take over:
3757 * raid0 - providing it has only two drives
3758 */
3759 if (mddev->level == 0) {
3760 /* for raid0 takeover only one zone is supported */
3761 raid0_conf = mddev->private;
3762 if (raid0_conf->nr_strip_zones > 1) {
3763 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3764 " with more than one zone.\n",
3765 mdname(mddev));
3766 return ERR_PTR(-EINVAL);
3767 }
3768 return raid10_takeover_raid0(mddev);
3769 }
3770 return ERR_PTR(-EINVAL);
3771}
3772
3773static int raid10_check_reshape(struct mddev *mddev)
3774{
3775 /* Called when there is a request to change
3776 * - layout (to ->new_layout)
3777 * - chunk size (to ->new_chunk_sectors)
3778 * - raid_disks (by delta_disks)
3779 * or when trying to restart a reshape that was ongoing.
3780 *
3781 * We need to validate the request and possibly allocate
3782 * space if that might be an issue later.
3783 *
3784 * Currently we reject any reshape of a 'far' mode array,
3785 * allow chunk size to change if new is generally acceptable,
3786 * allow raid_disks to increase, and allow
3787 * a switch between 'near' mode and 'offset' mode.
3788 */
3789 struct r10conf *conf = mddev->private;
3790 struct geom geo;
3791
3792 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3793 return -EINVAL;
3794
3795 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3796 /* mustn't change number of copies */
3797 return -EINVAL;
3798 if (geo.far_copies > 1 && !geo.far_offset)
3799 /* Cannot switch to 'far' mode */
3800 return -EINVAL;
3801
3802 if (mddev->array_sectors & geo.chunk_mask)
3803 /* not factor of array size */
3804 return -EINVAL;
3805
3806 if (!enough(conf, -1))
3807 return -EINVAL;
3808
3809 kfree(conf->mirrors_new);
3810 conf->mirrors_new = NULL;
3811 if (mddev->delta_disks > 0) {
3812 /* allocate new 'mirrors' list */
3813 conf->mirrors_new = kzalloc(
3814 sizeof(struct mirror_info)
3815 *(mddev->raid_disks +
3816 mddev->delta_disks),
3817 GFP_KERNEL);
3818 if (!conf->mirrors_new)
3819 return -ENOMEM;
3820 }
3821 return 0;
3822}
3823
3824/*
3825 * Need to check if array has failed when deciding whether to:
3826 * - start an array
3827 * - remove non-faulty devices
3828 * - add a spare
3829 * - allow a reshape
3830 * This determination is simple when no reshape is happening.
3831 * However if there is a reshape, we need to carefully check
3832 * both the before and after sections.
3833 * This is because some failed devices may only affect one
3834 * of the two sections, and some non-in_sync devices may
3835 * be insync in the section most affected by failed devices.
3836 */
3837static int calc_degraded(struct r10conf *conf)
3838{
3839 int degraded, degraded2;
3840 int i;
3841
3842 rcu_read_lock();
3843 degraded = 0;
3844 /* 'prev' section first */
3845 for (i = 0; i < conf->prev.raid_disks; i++) {
3846 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3847 if (!rdev || test_bit(Faulty, &rdev->flags))
3848 degraded++;
3849 else if (!test_bit(In_sync, &rdev->flags))
3850 /* When we can reduce the number of devices in
3851 * an array, this might not contribute to
3852 * 'degraded'. It does now.
3853 */
3854 degraded++;
3855 }
3856 rcu_read_unlock();
3857 if (conf->geo.raid_disks == conf->prev.raid_disks)
3858 return degraded;
3859 rcu_read_lock();
3860 degraded2 = 0;
3861 for (i = 0; i < conf->geo.raid_disks; i++) {
3862 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3863 if (!rdev || test_bit(Faulty, &rdev->flags))
3864 degraded2++;
3865 else if (!test_bit(In_sync, &rdev->flags)) {
3866 /* If reshape is increasing the number of devices,
3867 * this section has already been recovered, so
3868 * it doesn't contribute to degraded.
3869 * else it does.
3870 */
3871 if (conf->geo.raid_disks <= conf->prev.raid_disks)
3872 degraded2++;
3873 }
3874 }
3875 rcu_read_unlock();
3876 if (degraded2 > degraded)
3877 return degraded2;
3878 return degraded;
3879}
3880
3881static int raid10_start_reshape(struct mddev *mddev)
3882{
3883 /* A 'reshape' has been requested. This commits
3884 * the various 'new' fields and sets MD_RECOVER_RESHAPE
3885 * This also checks if there are enough spares and adds them
3886 * to the array.
3887 * We currently require enough spares to make the final
3888 * array non-degraded. We also require that the difference
3889 * between old and new data_offset - on each device - is
3890 * enough that we never risk over-writing.
3891 */
3892
3893 unsigned long before_length, after_length;
3894 sector_t min_offset_diff = 0;
3895 int first = 1;
3896 struct geom new;
3897 struct r10conf *conf = mddev->private;
3898 struct md_rdev *rdev;
3899 int spares = 0;
3900 int ret;
3901
3902 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
3903 return -EBUSY;
3904
3905 if (setup_geo(&new, mddev, geo_start) != conf->copies)
3906 return -EINVAL;
3907
3908 before_length = ((1 << conf->prev.chunk_shift) *
3909 conf->prev.far_copies);
3910 after_length = ((1 << conf->geo.chunk_shift) *
3911 conf->geo.far_copies);
3912
3913 rdev_for_each(rdev, mddev) {
3914 if (!test_bit(In_sync, &rdev->flags)
3915 && !test_bit(Faulty, &rdev->flags))
3916 spares++;
3917 if (rdev->raid_disk >= 0) {
3918 long long diff = (rdev->new_data_offset
3919 - rdev->data_offset);
3920 if (!mddev->reshape_backwards)
3921 diff = -diff;
3922 if (diff < 0)
3923 diff = 0;
3924 if (first || diff < min_offset_diff)
3925 min_offset_diff = diff;
3926 }
3927 }
3928
3929 if (max(before_length, after_length) > min_offset_diff)
3930 return -EINVAL;
3931
3932 if (spares < mddev->delta_disks)
3933 return -EINVAL;
3934
3935 conf->offset_diff = min_offset_diff;
3936 spin_lock_irq(&conf->device_lock);
3937 if (conf->mirrors_new) {
3938 memcpy(conf->mirrors_new, conf->mirrors,
3939 sizeof(struct mirror_info)*conf->prev.raid_disks);
3940 smp_mb();
3941 kfree(conf->mirrors_old); /* FIXME and elsewhere */
3942 conf->mirrors_old = conf->mirrors;
3943 conf->mirrors = conf->mirrors_new;
3944 conf->mirrors_new = NULL;
3945 }
3946 setup_geo(&conf->geo, mddev, geo_start);
3947 smp_mb();
3948 if (mddev->reshape_backwards) {
3949 sector_t size = raid10_size(mddev, 0, 0);
3950 if (size < mddev->array_sectors) {
3951 spin_unlock_irq(&conf->device_lock);
3952 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
3953 mdname(mddev));
3954 return -EINVAL;
3955 }
3956 mddev->resync_max_sectors = size;
3957 conf->reshape_progress = size;
3958 } else
3959 conf->reshape_progress = 0;
3960 spin_unlock_irq(&conf->device_lock);
3961
3962 if (mddev->delta_disks && mddev->bitmap) {
3963 ret = bitmap_resize(mddev->bitmap,
3964 raid10_size(mddev, 0,
3965 conf->geo.raid_disks),
3966 0, 0);
3967 if (ret)
3968 goto abort;
3969 }
3970 if (mddev->delta_disks > 0) {
3971 rdev_for_each(rdev, mddev)
3972 if (rdev->raid_disk < 0 &&
3973 !test_bit(Faulty, &rdev->flags)) {
3974 if (raid10_add_disk(mddev, rdev) == 0) {
3975 if (rdev->raid_disk >=
3976 conf->prev.raid_disks)
3977 set_bit(In_sync, &rdev->flags);
3978 else
3979 rdev->recovery_offset = 0;
3980
3981 if (sysfs_link_rdev(mddev, rdev))
3982 /* Failure here is OK */;
3983 }
3984 } else if (rdev->raid_disk >= conf->prev.raid_disks
3985 && !test_bit(Faulty, &rdev->flags)) {
3986 /* This is a spare that was manually added */
3987 set_bit(In_sync, &rdev->flags);
3988 }
3989 }
3990 /* When a reshape changes the number of devices,
3991 * ->degraded is measured against the larger of the
3992 * pre and post numbers.
3993 */
3994 spin_lock_irq(&conf->device_lock);
3995 mddev->degraded = calc_degraded(conf);
3996 spin_unlock_irq(&conf->device_lock);
3997 mddev->raid_disks = conf->geo.raid_disks;
3998 mddev->reshape_position = conf->reshape_progress;
3999 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4000
4001 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4002 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4003 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4004 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4005
4006 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4007 "reshape");
4008 if (!mddev->sync_thread) {
4009 ret = -EAGAIN;
4010 goto abort;
4011 }
4012 conf->reshape_checkpoint = jiffies;
4013 md_wakeup_thread(mddev->sync_thread);
4014 md_new_event(mddev);
4015 return 0;
4016
4017abort:
4018 mddev->recovery = 0;
4019 spin_lock_irq(&conf->device_lock);
4020 conf->geo = conf->prev;
4021 mddev->raid_disks = conf->geo.raid_disks;
4022 rdev_for_each(rdev, mddev)
4023 rdev->new_data_offset = rdev->data_offset;
4024 smp_wmb();
4025 conf->reshape_progress = MaxSector;
4026 mddev->reshape_position = MaxSector;
4027 spin_unlock_irq(&conf->device_lock);
4028 return ret;
4029}
4030
4031/* Calculate the last device-address that could contain
4032 * any block from the chunk that includes the array-address 's'
4033 * and report the next address.
4034 * i.e. the address returned will be chunk-aligned and after
4035 * any data that is in the chunk containing 's'.
4036 */
4037static sector_t last_dev_address(sector_t s, struct geom *geo)
4038{
4039 s = (s | geo->chunk_mask) + 1;
4040 s >>= geo->chunk_shift;
4041 s *= geo->near_copies;
4042 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4043 s *= geo->far_copies;
4044 s <<= geo->chunk_shift;
4045 return s;
4046}
4047
4048/* Calculate the first device-address that could contain
4049 * any block from the chunk that includes the array-address 's'.
4050 * This too will be the start of a chunk
4051 */
4052static sector_t first_dev_address(sector_t s, struct geom *geo)
4053{
4054 s >>= geo->chunk_shift;
4055 s *= geo->near_copies;
4056 sector_div(s, geo->raid_disks);
4057 s *= geo->far_copies;
4058 s <<= geo->chunk_shift;
4059 return s;
4060}
4061
4062static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4063 int *skipped)
4064{
4065 /* We simply copy at most one chunk (smallest of old and new)
4066 * at a time, possibly less if that exceeds RESYNC_PAGES,
4067 * or we hit a bad block or something.
4068 * This might mean we pause for normal IO in the middle of
4069 * a chunk, but that is not a problem was mddev->reshape_position
4070 * can record any location.
4071 *
4072 * If we will want to write to a location that isn't
4073 * yet recorded as 'safe' (i.e. in metadata on disk) then
4074 * we need to flush all reshape requests and update the metadata.
4075 *
4076 * When reshaping forwards (e.g. to more devices), we interpret
4077 * 'safe' as the earliest block which might not have been copied
4078 * down yet. We divide this by previous stripe size and multiply
4079 * by previous stripe length to get lowest device offset that we
4080 * cannot write to yet.
4081 * We interpret 'sector_nr' as an address that we want to write to.
4082 * From this we use last_device_address() to find where we might
4083 * write to, and first_device_address on the 'safe' position.
4084 * If this 'next' write position is after the 'safe' position,
4085 * we must update the metadata to increase the 'safe' position.
4086 *
4087 * When reshaping backwards, we round in the opposite direction
4088 * and perform the reverse test: next write position must not be
4089 * less than current safe position.
4090 *
4091 * In all this the minimum difference in data offsets
4092 * (conf->offset_diff - always positive) allows a bit of slack,
4093 * so next can be after 'safe', but not by more than offset_disk
4094 *
4095 * We need to prepare all the bios here before we start any IO
4096 * to ensure the size we choose is acceptable to all devices.
4097 * The means one for each copy for write-out and an extra one for
4098 * read-in.
4099 * We store the read-in bio in ->master_bio and the others in
4100 * ->devs[x].bio and ->devs[x].repl_bio.
4101 */
4102 struct r10conf *conf = mddev->private;
4103 struct r10bio *r10_bio;
4104 sector_t next, safe, last;
4105 int max_sectors;
4106 int nr_sectors;
4107 int s;
4108 struct md_rdev *rdev;
4109 int need_flush = 0;
4110 struct bio *blist;
4111 struct bio *bio, *read_bio;
4112 int sectors_done = 0;
4113
4114 if (sector_nr == 0) {
4115 /* If restarting in the middle, skip the initial sectors */
4116 if (mddev->reshape_backwards &&
4117 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4118 sector_nr = (raid10_size(mddev, 0, 0)
4119 - conf->reshape_progress);
4120 } else if (!mddev->reshape_backwards &&
4121 conf->reshape_progress > 0)
4122 sector_nr = conf->reshape_progress;
4123 if (sector_nr) {
4124 mddev->curr_resync_completed = sector_nr;
4125 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4126 *skipped = 1;
4127 return sector_nr;
4128 }
4129 }
4130
4131 /* We don't use sector_nr to track where we are up to
4132 * as that doesn't work well for ->reshape_backwards.
4133 * So just use ->reshape_progress.
4134 */
4135 if (mddev->reshape_backwards) {
4136 /* 'next' is the earliest device address that we might
4137 * write to for this chunk in the new layout
4138 */
4139 next = first_dev_address(conf->reshape_progress - 1,
4140 &conf->geo);
4141
4142 /* 'safe' is the last device address that we might read from
4143 * in the old layout after a restart
4144 */
4145 safe = last_dev_address(conf->reshape_safe - 1,
4146 &conf->prev);
4147
4148 if (next + conf->offset_diff < safe)
4149 need_flush = 1;
4150
4151 last = conf->reshape_progress - 1;
4152 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4153 & conf->prev.chunk_mask);
4154 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4155 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4156 } else {
4157 /* 'next' is after the last device address that we
4158 * might write to for this chunk in the new layout
4159 */
4160 next = last_dev_address(conf->reshape_progress, &conf->geo);
4161
4162 /* 'safe' is the earliest device address that we might
4163 * read from in the old layout after a restart
4164 */
4165 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4166
4167 /* Need to update metadata if 'next' might be beyond 'safe'
4168 * as that would possibly corrupt data
4169 */
4170 if (next > safe + conf->offset_diff)
4171 need_flush = 1;
4172
4173 sector_nr = conf->reshape_progress;
4174 last = sector_nr | (conf->geo.chunk_mask
4175 & conf->prev.chunk_mask);
4176
4177 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4178 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4179 }
4180
4181 if (need_flush ||
4182 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4183 /* Need to update reshape_position in metadata */
4184 wait_barrier(conf);
4185 mddev->reshape_position = conf->reshape_progress;
4186 if (mddev->reshape_backwards)
4187 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4188 - conf->reshape_progress;
4189 else
4190 mddev->curr_resync_completed = conf->reshape_progress;
4191 conf->reshape_checkpoint = jiffies;
4192 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4193 md_wakeup_thread(mddev->thread);
4194 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4195 kthread_should_stop());
4196 conf->reshape_safe = mddev->reshape_position;
4197 allow_barrier(conf);
4198 }
4199
4200read_more:
4201 /* Now schedule reads for blocks from sector_nr to last */
4202 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4203 raise_barrier(conf, sectors_done != 0);
4204 atomic_set(&r10_bio->remaining, 0);
4205 r10_bio->mddev = mddev;
4206 r10_bio->sector = sector_nr;
4207 set_bit(R10BIO_IsReshape, &r10_bio->state);
4208 r10_bio->sectors = last - sector_nr + 1;
4209 rdev = read_balance(conf, r10_bio, &max_sectors);
4210 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4211
4212 if (!rdev) {
4213 /* Cannot read from here, so need to record bad blocks
4214 * on all the target devices.
4215 */
4216 // FIXME
4217 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4218 return sectors_done;
4219 }
4220
4221 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4222
4223 read_bio->bi_bdev = rdev->bdev;
4224 read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4225 + rdev->data_offset);
4226 read_bio->bi_private = r10_bio;
4227 read_bio->bi_end_io = end_sync_read;
4228 read_bio->bi_rw = READ;
4229 read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4230 read_bio->bi_flags |= 1 << BIO_UPTODATE;
4231 read_bio->bi_vcnt = 0;
4232 read_bio->bi_idx = 0;
4233 read_bio->bi_size = 0;
4234 r10_bio->master_bio = read_bio;
4235 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4236
4237 /* Now find the locations in the new layout */
4238 __raid10_find_phys(&conf->geo, r10_bio);
4239
4240 blist = read_bio;
4241 read_bio->bi_next = NULL;
4242
4243 for (s = 0; s < conf->copies*2; s++) {
4244 struct bio *b;
4245 int d = r10_bio->devs[s/2].devnum;
4246 struct md_rdev *rdev2;
4247 if (s&1) {
4248 rdev2 = conf->mirrors[d].replacement;
4249 b = r10_bio->devs[s/2].repl_bio;
4250 } else {
4251 rdev2 = conf->mirrors[d].rdev;
4252 b = r10_bio->devs[s/2].bio;
4253 }
4254 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4255 continue;
4256 b->bi_bdev = rdev2->bdev;
4257 b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4258 b->bi_private = r10_bio;
4259 b->bi_end_io = end_reshape_write;
4260 b->bi_rw = WRITE;
4261 b->bi_flags &= ~(BIO_POOL_MASK - 1);
4262 b->bi_flags |= 1 << BIO_UPTODATE;
4263 b->bi_next = blist;
4264 b->bi_vcnt = 0;
4265 b->bi_idx = 0;
4266 b->bi_size = 0;
4267 blist = b;
4268 }
4269
4270 /* Now add as many pages as possible to all of these bios. */
4271
4272 nr_sectors = 0;
4273 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4274 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4275 int len = (max_sectors - s) << 9;
4276 if (len > PAGE_SIZE)
4277 len = PAGE_SIZE;
4278 for (bio = blist; bio ; bio = bio->bi_next) {
4279 struct bio *bio2;
4280 if (bio_add_page(bio, page, len, 0))
4281 continue;
4282
4283 /* Didn't fit, must stop */
4284 for (bio2 = blist;
4285 bio2 && bio2 != bio;
4286 bio2 = bio2->bi_next) {
4287 /* Remove last page from this bio */
4288 bio2->bi_vcnt--;
4289 bio2->bi_size -= len;
4290 bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4291 }
4292 goto bio_full;
4293 }
4294 sector_nr += len >> 9;
4295 nr_sectors += len >> 9;
4296 }
4297bio_full:
4298 r10_bio->sectors = nr_sectors;
4299
4300 /* Now submit the read */
4301 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4302 atomic_inc(&r10_bio->remaining);
4303 read_bio->bi_next = NULL;
4304 generic_make_request(read_bio);
4305 sector_nr += nr_sectors;
4306 sectors_done += nr_sectors;
4307 if (sector_nr <= last)
4308 goto read_more;
4309
4310 /* Now that we have done the whole section we can
4311 * update reshape_progress
4312 */
4313 if (mddev->reshape_backwards)
4314 conf->reshape_progress -= sectors_done;
4315 else
4316 conf->reshape_progress += sectors_done;
4317
4318 return sectors_done;
4319}
4320
4321static void end_reshape_request(struct r10bio *r10_bio);
4322static int handle_reshape_read_error(struct mddev *mddev,
4323 struct r10bio *r10_bio);
4324static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4325{
4326 /* Reshape read completed. Hopefully we have a block
4327 * to write out.
4328 * If we got a read error then we do sync 1-page reads from
4329 * elsewhere until we find the data - or give up.
4330 */
4331 struct r10conf *conf = mddev->private;
4332 int s;
4333
4334 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4335 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4336 /* Reshape has been aborted */
4337 md_done_sync(mddev, r10_bio->sectors, 0);
4338 return;
4339 }
4340
4341 /* We definitely have the data in the pages, schedule the
4342 * writes.
4343 */
4344 atomic_set(&r10_bio->remaining, 1);
4345 for (s = 0; s < conf->copies*2; s++) {
4346 struct bio *b;
4347 int d = r10_bio->devs[s/2].devnum;
4348 struct md_rdev *rdev;
4349 if (s&1) {
4350 rdev = conf->mirrors[d].replacement;
4351 b = r10_bio->devs[s/2].repl_bio;
4352 } else {
4353 rdev = conf->mirrors[d].rdev;
4354 b = r10_bio->devs[s/2].bio;
4355 }
4356 if (!rdev || test_bit(Faulty, &rdev->flags))
4357 continue;
4358 atomic_inc(&rdev->nr_pending);
4359 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4360 atomic_inc(&r10_bio->remaining);
4361 b->bi_next = NULL;
4362 generic_make_request(b);
4363 }
4364 end_reshape_request(r10_bio);
4365}
4366
4367static void end_reshape(struct r10conf *conf)
4368{
4369 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4370 return;
4371
4372 spin_lock_irq(&conf->device_lock);
4373 conf->prev = conf->geo;
4374 md_finish_reshape(conf->mddev);
4375 smp_wmb();
4376 conf->reshape_progress = MaxSector;
4377 spin_unlock_irq(&conf->device_lock);
4378
4379 /* read-ahead size must cover two whole stripes, which is
4380 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4381 */
4382 if (conf->mddev->queue) {
4383 int stripe = conf->geo.raid_disks *
4384 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4385 stripe /= conf->geo.near_copies;
4386 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4387 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4388 }
4389 conf->fullsync = 0;
4390}
4391
4392
4393static int handle_reshape_read_error(struct mddev *mddev,
4394 struct r10bio *r10_bio)
4395{
4396 /* Use sync reads to get the blocks from somewhere else */
4397 int sectors = r10_bio->sectors;
4398 struct r10conf *conf = mddev->private;
4399 struct {
4400 struct r10bio r10_bio;
4401 struct r10dev devs[conf->copies];
4402 } on_stack;
4403 struct r10bio *r10b = &on_stack.r10_bio;
4404 int slot = 0;
4405 int idx = 0;
4406 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4407
4408 r10b->sector = r10_bio->sector;
4409 __raid10_find_phys(&conf->prev, r10b);
4410
4411 while (sectors) {
4412 int s = sectors;
4413 int success = 0;
4414 int first_slot = slot;
4415
4416 if (s > (PAGE_SIZE >> 9))
4417 s = PAGE_SIZE >> 9;
4418
4419 while (!success) {
4420 int d = r10b->devs[slot].devnum;
4421 struct md_rdev *rdev = conf->mirrors[d].rdev;
4422 sector_t addr;
4423 if (rdev == NULL ||
4424 test_bit(Faulty, &rdev->flags) ||
4425 !test_bit(In_sync, &rdev->flags))
4426 goto failed;
4427
4428 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4429 success = sync_page_io(rdev,
4430 addr,
4431 s << 9,
4432 bvec[idx].bv_page,
4433 READ, false);
4434 if (success)
4435 break;
4436 failed:
4437 slot++;
4438 if (slot >= conf->copies)
4439 slot = 0;
4440 if (slot == first_slot)
4441 break;
4442 }
4443 if (!success) {
4444 /* couldn't read this block, must give up */
4445 set_bit(MD_RECOVERY_INTR,
4446 &mddev->recovery);
4447 return -EIO;
4448 }
4449 sectors -= s;
4450 idx++;
4451 }
4452 return 0;
4453}
4454
4455static void end_reshape_write(struct bio *bio, int error)
4456{
4457 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4458 struct r10bio *r10_bio = bio->bi_private;
4459 struct mddev *mddev = r10_bio->mddev;
4460 struct r10conf *conf = mddev->private;
4461 int d;
4462 int slot;
4463 int repl;
4464 struct md_rdev *rdev = NULL;
4465
4466 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4467 if (repl)
4468 rdev = conf->mirrors[d].replacement;
4469 if (!rdev) {
4470 smp_mb();
4471 rdev = conf->mirrors[d].rdev;
4472 }
4473
4474 if (!uptodate) {
4475 /* FIXME should record badblock */
4476 md_error(mddev, rdev);
4477 }
4478
4479 rdev_dec_pending(rdev, mddev);
4480 end_reshape_request(r10_bio);
4481}
4482
4483static void end_reshape_request(struct r10bio *r10_bio)
4484{
4485 if (!atomic_dec_and_test(&r10_bio->remaining))
4486 return;
4487 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4488 bio_put(r10_bio->master_bio);
4489 put_buf(r10_bio);
4490}
4491
4492static void raid10_finish_reshape(struct mddev *mddev)
4493{
4494 struct r10conf *conf = mddev->private;
4495
4496 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4497 return;
4498
4499 if (mddev->delta_disks > 0) {
4500 sector_t size = raid10_size(mddev, 0, 0);
4501 md_set_array_sectors(mddev, size);
4502 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4503 mddev->recovery_cp = mddev->resync_max_sectors;
4504 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4505 }
4506 mddev->resync_max_sectors = size;
4507 set_capacity(mddev->gendisk, mddev->array_sectors);
4508 revalidate_disk(mddev->gendisk);
4509 } else {
4510 int d;
4511 for (d = conf->geo.raid_disks ;
4512 d < conf->geo.raid_disks - mddev->delta_disks;
4513 d++) {
4514 struct md_rdev *rdev = conf->mirrors[d].rdev;
4515 if (rdev)
4516 clear_bit(In_sync, &rdev->flags);
4517 rdev = conf->mirrors[d].replacement;
4518 if (rdev)
4519 clear_bit(In_sync, &rdev->flags);
4520 }
4521 }
4522 mddev->layout = mddev->new_layout;
4523 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4524 mddev->reshape_position = MaxSector;
4525 mddev->delta_disks = 0;
4526 mddev->reshape_backwards = 0;
4527}
4528
4529static struct md_personality raid10_personality =
4530{
4531 .name = "raid10",
4532 .level = 10,
4533 .owner = THIS_MODULE,
4534 .make_request = make_request,
4535 .run = run,
4536 .stop = stop,
4537 .status = status,
4538 .error_handler = error,
4539 .hot_add_disk = raid10_add_disk,
4540 .hot_remove_disk= raid10_remove_disk,
4541 .spare_active = raid10_spare_active,
4542 .sync_request = sync_request,
4543 .quiesce = raid10_quiesce,
4544 .size = raid10_size,
4545 .resize = raid10_resize,
4546 .takeover = raid10_takeover,
4547 .check_reshape = raid10_check_reshape,
4548 .start_reshape = raid10_start_reshape,
4549 .finish_reshape = raid10_finish_reshape,
4550};
4551
4552static int __init raid_init(void)
4553{
4554 return register_md_personality(&raid10_personality);
4555}
4556
4557static void raid_exit(void)
4558{
4559 unregister_md_personality(&raid10_personality);
4560}
4561
4562module_init(raid_init);
4563module_exit(raid_exit);
4564MODULE_LICENSE("GPL");
4565MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4566MODULE_ALIAS("md-personality-9"); /* RAID10 */
4567MODULE_ALIAS("md-raid10");
4568MODULE_ALIAS("md-level-10");
4569
4570module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
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
67static void allow_barrier(struct r10conf *conf);
68static void lower_barrier(struct r10conf *conf);
69static int _enough(struct r10conf *conf, int previous, int ignore);
70static int enough(struct r10conf *conf, int ignore);
71static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
72 int *skipped);
73static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
74static void end_reshape_write(struct bio *bio);
75static 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 */
86static inline struct r10bio *get_resync_r10bio(struct bio *bio)
87{
88 return get_resync_pages(bio)->raid_bio;
89}
90
91static 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 */
116static 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
192out_free_pages:
193 while (--j >= 0)
194 resync_free_pages(&rps[j]);
195
196 j = 0;
197out_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);
205out_free_r10bio:
206 rbio_pool_free(r10_bio, conf);
207 return NULL;
208}
209
210static 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
237static 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
253static 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
261static 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
270static 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 */
292static 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 */
313static 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 */
324static 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
349static 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
402static 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
412static 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
427static 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 * For failfast, this is the only remaining device,
474 * We need to retry the write without FailFast.
475 */
476 if (!test_bit(Faulty, &rdev->flags))
477 set_bit(R10BIO_WriteError, &r10_bio->state);
478 else {
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
561static 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
624static 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
639static 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 */
703static 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
851static 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
925static 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
945static 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
954static 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
992static 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
999static 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
1026static 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
1036static 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
1046struct raid10_plug_cb {
1047 struct blk_plug_cb cb;
1048 struct bio_list pending;
1049 int pending_cnt;
1050};
1051
1052static 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 */
1100static 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
1117static 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 (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
1211static 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
1279static 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);
1342retry_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
1483static 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 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1497
1498 if (bio_data_dir(bio) == READ)
1499 raid10_read_request(mddev, bio, r10_bio);
1500 else
1501 raid10_write_request(mddev, bio, r10_bio);
1502}
1503
1504static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1505{
1506 struct r10conf *conf = mddev->private;
1507 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1508 int chunk_sects = chunk_mask + 1;
1509 int sectors = bio_sectors(bio);
1510
1511 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1512 && md_flush_request(mddev, bio))
1513 return true;
1514
1515 if (!md_write_start(mddev, bio))
1516 return false;
1517
1518 /*
1519 * If this request crosses a chunk boundary, we need to split
1520 * it.
1521 */
1522 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1523 sectors > chunk_sects
1524 && (conf->geo.near_copies < conf->geo.raid_disks
1525 || conf->prev.near_copies <
1526 conf->prev.raid_disks)))
1527 sectors = chunk_sects -
1528 (bio->bi_iter.bi_sector &
1529 (chunk_sects - 1));
1530 __make_request(mddev, bio, sectors);
1531
1532 /* In case raid10d snuck in to freeze_array */
1533 wake_up(&conf->wait_barrier);
1534 return true;
1535}
1536
1537static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1538{
1539 struct r10conf *conf = mddev->private;
1540 int i;
1541
1542 if (conf->geo.near_copies < conf->geo.raid_disks)
1543 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1544 if (conf->geo.near_copies > 1)
1545 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1546 if (conf->geo.far_copies > 1) {
1547 if (conf->geo.far_offset)
1548 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1549 else
1550 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1551 if (conf->geo.far_set_size != conf->geo.raid_disks)
1552 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1553 }
1554 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1555 conf->geo.raid_disks - mddev->degraded);
1556 rcu_read_lock();
1557 for (i = 0; i < conf->geo.raid_disks; i++) {
1558 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1559 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1560 }
1561 rcu_read_unlock();
1562 seq_printf(seq, "]");
1563}
1564
1565/* check if there are enough drives for
1566 * every block to appear on atleast one.
1567 * Don't consider the device numbered 'ignore'
1568 * as we might be about to remove it.
1569 */
1570static int _enough(struct r10conf *conf, int previous, int ignore)
1571{
1572 int first = 0;
1573 int has_enough = 0;
1574 int disks, ncopies;
1575 if (previous) {
1576 disks = conf->prev.raid_disks;
1577 ncopies = conf->prev.near_copies;
1578 } else {
1579 disks = conf->geo.raid_disks;
1580 ncopies = conf->geo.near_copies;
1581 }
1582
1583 rcu_read_lock();
1584 do {
1585 int n = conf->copies;
1586 int cnt = 0;
1587 int this = first;
1588 while (n--) {
1589 struct md_rdev *rdev;
1590 if (this != ignore &&
1591 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1592 test_bit(In_sync, &rdev->flags))
1593 cnt++;
1594 this = (this+1) % disks;
1595 }
1596 if (cnt == 0)
1597 goto out;
1598 first = (first + ncopies) % disks;
1599 } while (first != 0);
1600 has_enough = 1;
1601out:
1602 rcu_read_unlock();
1603 return has_enough;
1604}
1605
1606static int enough(struct r10conf *conf, int ignore)
1607{
1608 /* when calling 'enough', both 'prev' and 'geo' must
1609 * be stable.
1610 * This is ensured if ->reconfig_mutex or ->device_lock
1611 * is held.
1612 */
1613 return _enough(conf, 0, ignore) &&
1614 _enough(conf, 1, ignore);
1615}
1616
1617static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1618{
1619 char b[BDEVNAME_SIZE];
1620 struct r10conf *conf = mddev->private;
1621 unsigned long flags;
1622
1623 /*
1624 * If it is not operational, then we have already marked it as dead
1625 * else if it is the last working disks with "fail_last_dev == false",
1626 * ignore the error, let the next level up know.
1627 * else mark the drive as failed
1628 */
1629 spin_lock_irqsave(&conf->device_lock, flags);
1630 if (test_bit(In_sync, &rdev->flags) && !mddev->fail_last_dev
1631 && !enough(conf, rdev->raid_disk)) {
1632 /*
1633 * Don't fail the drive, just return an IO error.
1634 */
1635 spin_unlock_irqrestore(&conf->device_lock, flags);
1636 return;
1637 }
1638 if (test_and_clear_bit(In_sync, &rdev->flags))
1639 mddev->degraded++;
1640 /*
1641 * If recovery is running, make sure it aborts.
1642 */
1643 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1644 set_bit(Blocked, &rdev->flags);
1645 set_bit(Faulty, &rdev->flags);
1646 set_mask_bits(&mddev->sb_flags, 0,
1647 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1648 spin_unlock_irqrestore(&conf->device_lock, flags);
1649 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1650 "md/raid10:%s: Operation continuing on %d devices.\n",
1651 mdname(mddev), bdevname(rdev->bdev, b),
1652 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1653}
1654
1655static void print_conf(struct r10conf *conf)
1656{
1657 int i;
1658 struct md_rdev *rdev;
1659
1660 pr_debug("RAID10 conf printout:\n");
1661 if (!conf) {
1662 pr_debug("(!conf)\n");
1663 return;
1664 }
1665 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1666 conf->geo.raid_disks);
1667
1668 /* This is only called with ->reconfix_mutex held, so
1669 * rcu protection of rdev is not needed */
1670 for (i = 0; i < conf->geo.raid_disks; i++) {
1671 char b[BDEVNAME_SIZE];
1672 rdev = conf->mirrors[i].rdev;
1673 if (rdev)
1674 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1675 i, !test_bit(In_sync, &rdev->flags),
1676 !test_bit(Faulty, &rdev->flags),
1677 bdevname(rdev->bdev,b));
1678 }
1679}
1680
1681static void close_sync(struct r10conf *conf)
1682{
1683 wait_barrier(conf);
1684 allow_barrier(conf);
1685
1686 mempool_exit(&conf->r10buf_pool);
1687}
1688
1689static int raid10_spare_active(struct mddev *mddev)
1690{
1691 int i;
1692 struct r10conf *conf = mddev->private;
1693 struct raid10_info *tmp;
1694 int count = 0;
1695 unsigned long flags;
1696
1697 /*
1698 * Find all non-in_sync disks within the RAID10 configuration
1699 * and mark them in_sync
1700 */
1701 for (i = 0; i < conf->geo.raid_disks; i++) {
1702 tmp = conf->mirrors + i;
1703 if (tmp->replacement
1704 && tmp->replacement->recovery_offset == MaxSector
1705 && !test_bit(Faulty, &tmp->replacement->flags)
1706 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1707 /* Replacement has just become active */
1708 if (!tmp->rdev
1709 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1710 count++;
1711 if (tmp->rdev) {
1712 /* Replaced device not technically faulty,
1713 * but we need to be sure it gets removed
1714 * and never re-added.
1715 */
1716 set_bit(Faulty, &tmp->rdev->flags);
1717 sysfs_notify_dirent_safe(
1718 tmp->rdev->sysfs_state);
1719 }
1720 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1721 } else if (tmp->rdev
1722 && tmp->rdev->recovery_offset == MaxSector
1723 && !test_bit(Faulty, &tmp->rdev->flags)
1724 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1725 count++;
1726 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1727 }
1728 }
1729 spin_lock_irqsave(&conf->device_lock, flags);
1730 mddev->degraded -= count;
1731 spin_unlock_irqrestore(&conf->device_lock, flags);
1732
1733 print_conf(conf);
1734 return count;
1735}
1736
1737static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1738{
1739 struct r10conf *conf = mddev->private;
1740 int err = -EEXIST;
1741 int mirror;
1742 int first = 0;
1743 int last = conf->geo.raid_disks - 1;
1744
1745 if (mddev->recovery_cp < MaxSector)
1746 /* only hot-add to in-sync arrays, as recovery is
1747 * very different from resync
1748 */
1749 return -EBUSY;
1750 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1751 return -EINVAL;
1752
1753 if (md_integrity_add_rdev(rdev, mddev))
1754 return -ENXIO;
1755
1756 if (rdev->raid_disk >= 0)
1757 first = last = rdev->raid_disk;
1758
1759 if (rdev->saved_raid_disk >= first &&
1760 rdev->saved_raid_disk < conf->geo.raid_disks &&
1761 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1762 mirror = rdev->saved_raid_disk;
1763 else
1764 mirror = first;
1765 for ( ; mirror <= last ; mirror++) {
1766 struct raid10_info *p = &conf->mirrors[mirror];
1767 if (p->recovery_disabled == mddev->recovery_disabled)
1768 continue;
1769 if (p->rdev) {
1770 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1771 p->replacement != NULL)
1772 continue;
1773 clear_bit(In_sync, &rdev->flags);
1774 set_bit(Replacement, &rdev->flags);
1775 rdev->raid_disk = mirror;
1776 err = 0;
1777 if (mddev->gendisk)
1778 disk_stack_limits(mddev->gendisk, rdev->bdev,
1779 rdev->data_offset << 9);
1780 conf->fullsync = 1;
1781 rcu_assign_pointer(p->replacement, rdev);
1782 break;
1783 }
1784
1785 if (mddev->gendisk)
1786 disk_stack_limits(mddev->gendisk, rdev->bdev,
1787 rdev->data_offset << 9);
1788
1789 p->head_position = 0;
1790 p->recovery_disabled = mddev->recovery_disabled - 1;
1791 rdev->raid_disk = mirror;
1792 err = 0;
1793 if (rdev->saved_raid_disk != mirror)
1794 conf->fullsync = 1;
1795 rcu_assign_pointer(p->rdev, rdev);
1796 break;
1797 }
1798 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1799 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1800
1801 print_conf(conf);
1802 return err;
1803}
1804
1805static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1806{
1807 struct r10conf *conf = mddev->private;
1808 int err = 0;
1809 int number = rdev->raid_disk;
1810 struct md_rdev **rdevp;
1811 struct raid10_info *p = conf->mirrors + number;
1812
1813 print_conf(conf);
1814 if (rdev == p->rdev)
1815 rdevp = &p->rdev;
1816 else if (rdev == p->replacement)
1817 rdevp = &p->replacement;
1818 else
1819 return 0;
1820
1821 if (test_bit(In_sync, &rdev->flags) ||
1822 atomic_read(&rdev->nr_pending)) {
1823 err = -EBUSY;
1824 goto abort;
1825 }
1826 /* Only remove non-faulty devices if recovery
1827 * is not possible.
1828 */
1829 if (!test_bit(Faulty, &rdev->flags) &&
1830 mddev->recovery_disabled != p->recovery_disabled &&
1831 (!p->replacement || p->replacement == rdev) &&
1832 number < conf->geo.raid_disks &&
1833 enough(conf, -1)) {
1834 err = -EBUSY;
1835 goto abort;
1836 }
1837 *rdevp = NULL;
1838 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1839 synchronize_rcu();
1840 if (atomic_read(&rdev->nr_pending)) {
1841 /* lost the race, try later */
1842 err = -EBUSY;
1843 *rdevp = rdev;
1844 goto abort;
1845 }
1846 }
1847 if (p->replacement) {
1848 /* We must have just cleared 'rdev' */
1849 p->rdev = p->replacement;
1850 clear_bit(Replacement, &p->replacement->flags);
1851 smp_mb(); /* Make sure other CPUs may see both as identical
1852 * but will never see neither -- if they are careful.
1853 */
1854 p->replacement = NULL;
1855 }
1856
1857 clear_bit(WantReplacement, &rdev->flags);
1858 err = md_integrity_register(mddev);
1859
1860abort:
1861
1862 print_conf(conf);
1863 return err;
1864}
1865
1866static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1867{
1868 struct r10conf *conf = r10_bio->mddev->private;
1869
1870 if (!bio->bi_status)
1871 set_bit(R10BIO_Uptodate, &r10_bio->state);
1872 else
1873 /* The write handler will notice the lack of
1874 * R10BIO_Uptodate and record any errors etc
1875 */
1876 atomic_add(r10_bio->sectors,
1877 &conf->mirrors[d].rdev->corrected_errors);
1878
1879 /* for reconstruct, we always reschedule after a read.
1880 * for resync, only after all reads
1881 */
1882 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1883 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1884 atomic_dec_and_test(&r10_bio->remaining)) {
1885 /* we have read all the blocks,
1886 * do the comparison in process context in raid10d
1887 */
1888 reschedule_retry(r10_bio);
1889 }
1890}
1891
1892static void end_sync_read(struct bio *bio)
1893{
1894 struct r10bio *r10_bio = get_resync_r10bio(bio);
1895 struct r10conf *conf = r10_bio->mddev->private;
1896 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1897
1898 __end_sync_read(r10_bio, bio, d);
1899}
1900
1901static void end_reshape_read(struct bio *bio)
1902{
1903 /* reshape read bio isn't allocated from r10buf_pool */
1904 struct r10bio *r10_bio = bio->bi_private;
1905
1906 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1907}
1908
1909static void end_sync_request(struct r10bio *r10_bio)
1910{
1911 struct mddev *mddev = r10_bio->mddev;
1912
1913 while (atomic_dec_and_test(&r10_bio->remaining)) {
1914 if (r10_bio->master_bio == NULL) {
1915 /* the primary of several recovery bios */
1916 sector_t s = r10_bio->sectors;
1917 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1918 test_bit(R10BIO_WriteError, &r10_bio->state))
1919 reschedule_retry(r10_bio);
1920 else
1921 put_buf(r10_bio);
1922 md_done_sync(mddev, s, 1);
1923 break;
1924 } else {
1925 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1926 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1927 test_bit(R10BIO_WriteError, &r10_bio->state))
1928 reschedule_retry(r10_bio);
1929 else
1930 put_buf(r10_bio);
1931 r10_bio = r10_bio2;
1932 }
1933 }
1934}
1935
1936static void end_sync_write(struct bio *bio)
1937{
1938 struct r10bio *r10_bio = get_resync_r10bio(bio);
1939 struct mddev *mddev = r10_bio->mddev;
1940 struct r10conf *conf = mddev->private;
1941 int d;
1942 sector_t first_bad;
1943 int bad_sectors;
1944 int slot;
1945 int repl;
1946 struct md_rdev *rdev = NULL;
1947
1948 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1949 if (repl)
1950 rdev = conf->mirrors[d].replacement;
1951 else
1952 rdev = conf->mirrors[d].rdev;
1953
1954 if (bio->bi_status) {
1955 if (repl)
1956 md_error(mddev, rdev);
1957 else {
1958 set_bit(WriteErrorSeen, &rdev->flags);
1959 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1960 set_bit(MD_RECOVERY_NEEDED,
1961 &rdev->mddev->recovery);
1962 set_bit(R10BIO_WriteError, &r10_bio->state);
1963 }
1964 } else if (is_badblock(rdev,
1965 r10_bio->devs[slot].addr,
1966 r10_bio->sectors,
1967 &first_bad, &bad_sectors))
1968 set_bit(R10BIO_MadeGood, &r10_bio->state);
1969
1970 rdev_dec_pending(rdev, mddev);
1971
1972 end_sync_request(r10_bio);
1973}
1974
1975/*
1976 * Note: sync and recover and handled very differently for raid10
1977 * This code is for resync.
1978 * For resync, we read through virtual addresses and read all blocks.
1979 * If there is any error, we schedule a write. The lowest numbered
1980 * drive is authoritative.
1981 * However requests come for physical address, so we need to map.
1982 * For every physical address there are raid_disks/copies virtual addresses,
1983 * which is always are least one, but is not necessarly an integer.
1984 * This means that a physical address can span multiple chunks, so we may
1985 * have to submit multiple io requests for a single sync request.
1986 */
1987/*
1988 * We check if all blocks are in-sync and only write to blocks that
1989 * aren't in sync
1990 */
1991static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1992{
1993 struct r10conf *conf = mddev->private;
1994 int i, first;
1995 struct bio *tbio, *fbio;
1996 int vcnt;
1997 struct page **tpages, **fpages;
1998
1999 atomic_set(&r10_bio->remaining, 1);
2000
2001 /* find the first device with a block */
2002 for (i=0; i<conf->copies; i++)
2003 if (!r10_bio->devs[i].bio->bi_status)
2004 break;
2005
2006 if (i == conf->copies)
2007 goto done;
2008
2009 first = i;
2010 fbio = r10_bio->devs[i].bio;
2011 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2012 fbio->bi_iter.bi_idx = 0;
2013 fpages = get_resync_pages(fbio)->pages;
2014
2015 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2016 /* now find blocks with errors */
2017 for (i=0 ; i < conf->copies ; i++) {
2018 int j, d;
2019 struct md_rdev *rdev;
2020 struct resync_pages *rp;
2021
2022 tbio = r10_bio->devs[i].bio;
2023
2024 if (tbio->bi_end_io != end_sync_read)
2025 continue;
2026 if (i == first)
2027 continue;
2028
2029 tpages = get_resync_pages(tbio)->pages;
2030 d = r10_bio->devs[i].devnum;
2031 rdev = conf->mirrors[d].rdev;
2032 if (!r10_bio->devs[i].bio->bi_status) {
2033 /* We know that the bi_io_vec layout is the same for
2034 * both 'first' and 'i', so we just compare them.
2035 * All vec entries are PAGE_SIZE;
2036 */
2037 int sectors = r10_bio->sectors;
2038 for (j = 0; j < vcnt; j++) {
2039 int len = PAGE_SIZE;
2040 if (sectors < (len / 512))
2041 len = sectors * 512;
2042 if (memcmp(page_address(fpages[j]),
2043 page_address(tpages[j]),
2044 len))
2045 break;
2046 sectors -= len/512;
2047 }
2048 if (j == vcnt)
2049 continue;
2050 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2051 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2052 /* Don't fix anything. */
2053 continue;
2054 } else if (test_bit(FailFast, &rdev->flags)) {
2055 /* Just give up on this device */
2056 md_error(rdev->mddev, rdev);
2057 continue;
2058 }
2059 /* Ok, we need to write this bio, either to correct an
2060 * inconsistency or to correct an unreadable block.
2061 * First we need to fixup bv_offset, bv_len and
2062 * bi_vecs, as the read request might have corrupted these
2063 */
2064 rp = get_resync_pages(tbio);
2065 bio_reset(tbio);
2066
2067 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2068
2069 rp->raid_bio = r10_bio;
2070 tbio->bi_private = rp;
2071 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2072 tbio->bi_end_io = end_sync_write;
2073 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2074
2075 bio_copy_data(tbio, fbio);
2076
2077 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2078 atomic_inc(&r10_bio->remaining);
2079 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2080
2081 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2082 tbio->bi_opf |= MD_FAILFAST;
2083 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2084 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2085 submit_bio_noacct(tbio);
2086 }
2087
2088 /* Now write out to any replacement devices
2089 * that are active
2090 */
2091 for (i = 0; i < conf->copies; i++) {
2092 int d;
2093
2094 tbio = r10_bio->devs[i].repl_bio;
2095 if (!tbio || !tbio->bi_end_io)
2096 continue;
2097 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2098 && r10_bio->devs[i].bio != fbio)
2099 bio_copy_data(tbio, fbio);
2100 d = r10_bio->devs[i].devnum;
2101 atomic_inc(&r10_bio->remaining);
2102 md_sync_acct(conf->mirrors[d].replacement->bdev,
2103 bio_sectors(tbio));
2104 submit_bio_noacct(tbio);
2105 }
2106
2107done:
2108 if (atomic_dec_and_test(&r10_bio->remaining)) {
2109 md_done_sync(mddev, r10_bio->sectors, 1);
2110 put_buf(r10_bio);
2111 }
2112}
2113
2114/*
2115 * Now for the recovery code.
2116 * Recovery happens across physical sectors.
2117 * We recover all non-is_sync drives by finding the virtual address of
2118 * each, and then choose a working drive that also has that virt address.
2119 * There is a separate r10_bio for each non-in_sync drive.
2120 * Only the first two slots are in use. The first for reading,
2121 * The second for writing.
2122 *
2123 */
2124static void fix_recovery_read_error(struct r10bio *r10_bio)
2125{
2126 /* We got a read error during recovery.
2127 * We repeat the read in smaller page-sized sections.
2128 * If a read succeeds, write it to the new device or record
2129 * a bad block if we cannot.
2130 * If a read fails, record a bad block on both old and
2131 * new devices.
2132 */
2133 struct mddev *mddev = r10_bio->mddev;
2134 struct r10conf *conf = mddev->private;
2135 struct bio *bio = r10_bio->devs[0].bio;
2136 sector_t sect = 0;
2137 int sectors = r10_bio->sectors;
2138 int idx = 0;
2139 int dr = r10_bio->devs[0].devnum;
2140 int dw = r10_bio->devs[1].devnum;
2141 struct page **pages = get_resync_pages(bio)->pages;
2142
2143 while (sectors) {
2144 int s = sectors;
2145 struct md_rdev *rdev;
2146 sector_t addr;
2147 int ok;
2148
2149 if (s > (PAGE_SIZE>>9))
2150 s = PAGE_SIZE >> 9;
2151
2152 rdev = conf->mirrors[dr].rdev;
2153 addr = r10_bio->devs[0].addr + sect,
2154 ok = sync_page_io(rdev,
2155 addr,
2156 s << 9,
2157 pages[idx],
2158 REQ_OP_READ, 0, false);
2159 if (ok) {
2160 rdev = conf->mirrors[dw].rdev;
2161 addr = r10_bio->devs[1].addr + sect;
2162 ok = sync_page_io(rdev,
2163 addr,
2164 s << 9,
2165 pages[idx],
2166 REQ_OP_WRITE, 0, false);
2167 if (!ok) {
2168 set_bit(WriteErrorSeen, &rdev->flags);
2169 if (!test_and_set_bit(WantReplacement,
2170 &rdev->flags))
2171 set_bit(MD_RECOVERY_NEEDED,
2172 &rdev->mddev->recovery);
2173 }
2174 }
2175 if (!ok) {
2176 /* We don't worry if we cannot set a bad block -
2177 * it really is bad so there is no loss in not
2178 * recording it yet
2179 */
2180 rdev_set_badblocks(rdev, addr, s, 0);
2181
2182 if (rdev != conf->mirrors[dw].rdev) {
2183 /* need bad block on destination too */
2184 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2185 addr = r10_bio->devs[1].addr + sect;
2186 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2187 if (!ok) {
2188 /* just abort the recovery */
2189 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2190 mdname(mddev));
2191
2192 conf->mirrors[dw].recovery_disabled
2193 = mddev->recovery_disabled;
2194 set_bit(MD_RECOVERY_INTR,
2195 &mddev->recovery);
2196 break;
2197 }
2198 }
2199 }
2200
2201 sectors -= s;
2202 sect += s;
2203 idx++;
2204 }
2205}
2206
2207static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2208{
2209 struct r10conf *conf = mddev->private;
2210 int d;
2211 struct bio *wbio, *wbio2;
2212
2213 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2214 fix_recovery_read_error(r10_bio);
2215 end_sync_request(r10_bio);
2216 return;
2217 }
2218
2219 /*
2220 * share the pages with the first bio
2221 * and submit the write request
2222 */
2223 d = r10_bio->devs[1].devnum;
2224 wbio = r10_bio->devs[1].bio;
2225 wbio2 = r10_bio->devs[1].repl_bio;
2226 /* Need to test wbio2->bi_end_io before we call
2227 * submit_bio_noacct as if the former is NULL,
2228 * the latter is free to free wbio2.
2229 */
2230 if (wbio2 && !wbio2->bi_end_io)
2231 wbio2 = NULL;
2232 if (wbio->bi_end_io) {
2233 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2234 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2235 submit_bio_noacct(wbio);
2236 }
2237 if (wbio2) {
2238 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2239 md_sync_acct(conf->mirrors[d].replacement->bdev,
2240 bio_sectors(wbio2));
2241 submit_bio_noacct(wbio2);
2242 }
2243}
2244
2245/*
2246 * Used by fix_read_error() to decay the per rdev read_errors.
2247 * We halve the read error count for every hour that has elapsed
2248 * since the last recorded read error.
2249 *
2250 */
2251static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2252{
2253 long cur_time_mon;
2254 unsigned long hours_since_last;
2255 unsigned int read_errors = atomic_read(&rdev->read_errors);
2256
2257 cur_time_mon = ktime_get_seconds();
2258
2259 if (rdev->last_read_error == 0) {
2260 /* first time we've seen a read error */
2261 rdev->last_read_error = cur_time_mon;
2262 return;
2263 }
2264
2265 hours_since_last = (long)(cur_time_mon -
2266 rdev->last_read_error) / 3600;
2267
2268 rdev->last_read_error = cur_time_mon;
2269
2270 /*
2271 * if hours_since_last is > the number of bits in read_errors
2272 * just set read errors to 0. We do this to avoid
2273 * overflowing the shift of read_errors by hours_since_last.
2274 */
2275 if (hours_since_last >= 8 * sizeof(read_errors))
2276 atomic_set(&rdev->read_errors, 0);
2277 else
2278 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2279}
2280
2281static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2282 int sectors, struct page *page, int rw)
2283{
2284 sector_t first_bad;
2285 int bad_sectors;
2286
2287 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2288 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2289 return -1;
2290 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2291 /* success */
2292 return 1;
2293 if (rw == WRITE) {
2294 set_bit(WriteErrorSeen, &rdev->flags);
2295 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2296 set_bit(MD_RECOVERY_NEEDED,
2297 &rdev->mddev->recovery);
2298 }
2299 /* need to record an error - either for the block or the device */
2300 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2301 md_error(rdev->mddev, rdev);
2302 return 0;
2303}
2304
2305/*
2306 * This is a kernel thread which:
2307 *
2308 * 1. Retries failed read operations on working mirrors.
2309 * 2. Updates the raid superblock when problems encounter.
2310 * 3. Performs writes following reads for array synchronising.
2311 */
2312
2313static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2314{
2315 int sect = 0; /* Offset from r10_bio->sector */
2316 int sectors = r10_bio->sectors;
2317 struct md_rdev *rdev;
2318 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2319 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2320
2321 /* still own a reference to this rdev, so it cannot
2322 * have been cleared recently.
2323 */
2324 rdev = conf->mirrors[d].rdev;
2325
2326 if (test_bit(Faulty, &rdev->flags))
2327 /* drive has already been failed, just ignore any
2328 more fix_read_error() attempts */
2329 return;
2330
2331 check_decay_read_errors(mddev, rdev);
2332 atomic_inc(&rdev->read_errors);
2333 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2334 char b[BDEVNAME_SIZE];
2335 bdevname(rdev->bdev, b);
2336
2337 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2338 mdname(mddev), b,
2339 atomic_read(&rdev->read_errors), max_read_errors);
2340 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2341 mdname(mddev), b);
2342 md_error(mddev, rdev);
2343 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2344 return;
2345 }
2346
2347 while(sectors) {
2348 int s = sectors;
2349 int sl = r10_bio->read_slot;
2350 int success = 0;
2351 int start;
2352
2353 if (s > (PAGE_SIZE>>9))
2354 s = PAGE_SIZE >> 9;
2355
2356 rcu_read_lock();
2357 do {
2358 sector_t first_bad;
2359 int bad_sectors;
2360
2361 d = r10_bio->devs[sl].devnum;
2362 rdev = rcu_dereference(conf->mirrors[d].rdev);
2363 if (rdev &&
2364 test_bit(In_sync, &rdev->flags) &&
2365 !test_bit(Faulty, &rdev->flags) &&
2366 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2367 &first_bad, &bad_sectors) == 0) {
2368 atomic_inc(&rdev->nr_pending);
2369 rcu_read_unlock();
2370 success = sync_page_io(rdev,
2371 r10_bio->devs[sl].addr +
2372 sect,
2373 s<<9,
2374 conf->tmppage,
2375 REQ_OP_READ, 0, false);
2376 rdev_dec_pending(rdev, mddev);
2377 rcu_read_lock();
2378 if (success)
2379 break;
2380 }
2381 sl++;
2382 if (sl == conf->copies)
2383 sl = 0;
2384 } while (!success && sl != r10_bio->read_slot);
2385 rcu_read_unlock();
2386
2387 if (!success) {
2388 /* Cannot read from anywhere, just mark the block
2389 * as bad on the first device to discourage future
2390 * reads.
2391 */
2392 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2393 rdev = conf->mirrors[dn].rdev;
2394
2395 if (!rdev_set_badblocks(
2396 rdev,
2397 r10_bio->devs[r10_bio->read_slot].addr
2398 + sect,
2399 s, 0)) {
2400 md_error(mddev, rdev);
2401 r10_bio->devs[r10_bio->read_slot].bio
2402 = IO_BLOCKED;
2403 }
2404 break;
2405 }
2406
2407 start = sl;
2408 /* write it back and re-read */
2409 rcu_read_lock();
2410 while (sl != r10_bio->read_slot) {
2411 char b[BDEVNAME_SIZE];
2412
2413 if (sl==0)
2414 sl = conf->copies;
2415 sl--;
2416 d = r10_bio->devs[sl].devnum;
2417 rdev = rcu_dereference(conf->mirrors[d].rdev);
2418 if (!rdev ||
2419 test_bit(Faulty, &rdev->flags) ||
2420 !test_bit(In_sync, &rdev->flags))
2421 continue;
2422
2423 atomic_inc(&rdev->nr_pending);
2424 rcu_read_unlock();
2425 if (r10_sync_page_io(rdev,
2426 r10_bio->devs[sl].addr +
2427 sect,
2428 s, conf->tmppage, WRITE)
2429 == 0) {
2430 /* Well, this device is dead */
2431 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2432 mdname(mddev), s,
2433 (unsigned long long)(
2434 sect +
2435 choose_data_offset(r10_bio,
2436 rdev)),
2437 bdevname(rdev->bdev, b));
2438 pr_notice("md/raid10:%s: %s: failing drive\n",
2439 mdname(mddev),
2440 bdevname(rdev->bdev, b));
2441 }
2442 rdev_dec_pending(rdev, mddev);
2443 rcu_read_lock();
2444 }
2445 sl = start;
2446 while (sl != r10_bio->read_slot) {
2447 char b[BDEVNAME_SIZE];
2448
2449 if (sl==0)
2450 sl = conf->copies;
2451 sl--;
2452 d = r10_bio->devs[sl].devnum;
2453 rdev = rcu_dereference(conf->mirrors[d].rdev);
2454 if (!rdev ||
2455 test_bit(Faulty, &rdev->flags) ||
2456 !test_bit(In_sync, &rdev->flags))
2457 continue;
2458
2459 atomic_inc(&rdev->nr_pending);
2460 rcu_read_unlock();
2461 switch (r10_sync_page_io(rdev,
2462 r10_bio->devs[sl].addr +
2463 sect,
2464 s, conf->tmppage,
2465 READ)) {
2466 case 0:
2467 /* Well, this device is dead */
2468 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2469 mdname(mddev), s,
2470 (unsigned long long)(
2471 sect +
2472 choose_data_offset(r10_bio, rdev)),
2473 bdevname(rdev->bdev, b));
2474 pr_notice("md/raid10:%s: %s: failing drive\n",
2475 mdname(mddev),
2476 bdevname(rdev->bdev, b));
2477 break;
2478 case 1:
2479 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2480 mdname(mddev), s,
2481 (unsigned long long)(
2482 sect +
2483 choose_data_offset(r10_bio, rdev)),
2484 bdevname(rdev->bdev, b));
2485 atomic_add(s, &rdev->corrected_errors);
2486 }
2487
2488 rdev_dec_pending(rdev, mddev);
2489 rcu_read_lock();
2490 }
2491 rcu_read_unlock();
2492
2493 sectors -= s;
2494 sect += s;
2495 }
2496}
2497
2498static int narrow_write_error(struct r10bio *r10_bio, int i)
2499{
2500 struct bio *bio = r10_bio->master_bio;
2501 struct mddev *mddev = r10_bio->mddev;
2502 struct r10conf *conf = mddev->private;
2503 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2504 /* bio has the data to be written to slot 'i' where
2505 * we just recently had a write error.
2506 * We repeatedly clone the bio and trim down to one block,
2507 * then try the write. Where the write fails we record
2508 * a bad block.
2509 * It is conceivable that the bio doesn't exactly align with
2510 * blocks. We must handle this.
2511 *
2512 * We currently own a reference to the rdev.
2513 */
2514
2515 int block_sectors;
2516 sector_t sector;
2517 int sectors;
2518 int sect_to_write = r10_bio->sectors;
2519 int ok = 1;
2520
2521 if (rdev->badblocks.shift < 0)
2522 return 0;
2523
2524 block_sectors = roundup(1 << rdev->badblocks.shift,
2525 bdev_logical_block_size(rdev->bdev) >> 9);
2526 sector = r10_bio->sector;
2527 sectors = ((r10_bio->sector + block_sectors)
2528 & ~(sector_t)(block_sectors - 1))
2529 - sector;
2530
2531 while (sect_to_write) {
2532 struct bio *wbio;
2533 sector_t wsector;
2534 if (sectors > sect_to_write)
2535 sectors = sect_to_write;
2536 /* Write at 'sector' for 'sectors' */
2537 wbio = bio_clone_fast(bio, GFP_NOIO, &mddev->bio_set);
2538 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2539 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2540 wbio->bi_iter.bi_sector = wsector +
2541 choose_data_offset(r10_bio, rdev);
2542 bio_set_dev(wbio, rdev->bdev);
2543 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2544
2545 if (submit_bio_wait(wbio) < 0)
2546 /* Failure! */
2547 ok = rdev_set_badblocks(rdev, wsector,
2548 sectors, 0)
2549 && ok;
2550
2551 bio_put(wbio);
2552 sect_to_write -= sectors;
2553 sector += sectors;
2554 sectors = block_sectors;
2555 }
2556 return ok;
2557}
2558
2559static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2560{
2561 int slot = r10_bio->read_slot;
2562 struct bio *bio;
2563 struct r10conf *conf = mddev->private;
2564 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2565
2566 /* we got a read error. Maybe the drive is bad. Maybe just
2567 * the block and we can fix it.
2568 * We freeze all other IO, and try reading the block from
2569 * other devices. When we find one, we re-write
2570 * and check it that fixes the read error.
2571 * This is all done synchronously while the array is
2572 * frozen.
2573 */
2574 bio = r10_bio->devs[slot].bio;
2575 bio_put(bio);
2576 r10_bio->devs[slot].bio = NULL;
2577
2578 if (mddev->ro)
2579 r10_bio->devs[slot].bio = IO_BLOCKED;
2580 else if (!test_bit(FailFast, &rdev->flags)) {
2581 freeze_array(conf, 1);
2582 fix_read_error(conf, mddev, r10_bio);
2583 unfreeze_array(conf);
2584 } else
2585 md_error(mddev, rdev);
2586
2587 rdev_dec_pending(rdev, mddev);
2588 allow_barrier(conf);
2589 r10_bio->state = 0;
2590 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2591}
2592
2593static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2594{
2595 /* Some sort of write request has finished and it
2596 * succeeded in writing where we thought there was a
2597 * bad block. So forget the bad block.
2598 * Or possibly if failed and we need to record
2599 * a bad block.
2600 */
2601 int m;
2602 struct md_rdev *rdev;
2603
2604 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2605 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2606 for (m = 0; m < conf->copies; m++) {
2607 int dev = r10_bio->devs[m].devnum;
2608 rdev = conf->mirrors[dev].rdev;
2609 if (r10_bio->devs[m].bio == NULL ||
2610 r10_bio->devs[m].bio->bi_end_io == NULL)
2611 continue;
2612 if (!r10_bio->devs[m].bio->bi_status) {
2613 rdev_clear_badblocks(
2614 rdev,
2615 r10_bio->devs[m].addr,
2616 r10_bio->sectors, 0);
2617 } else {
2618 if (!rdev_set_badblocks(
2619 rdev,
2620 r10_bio->devs[m].addr,
2621 r10_bio->sectors, 0))
2622 md_error(conf->mddev, rdev);
2623 }
2624 rdev = conf->mirrors[dev].replacement;
2625 if (r10_bio->devs[m].repl_bio == NULL ||
2626 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2627 continue;
2628
2629 if (!r10_bio->devs[m].repl_bio->bi_status) {
2630 rdev_clear_badblocks(
2631 rdev,
2632 r10_bio->devs[m].addr,
2633 r10_bio->sectors, 0);
2634 } else {
2635 if (!rdev_set_badblocks(
2636 rdev,
2637 r10_bio->devs[m].addr,
2638 r10_bio->sectors, 0))
2639 md_error(conf->mddev, rdev);
2640 }
2641 }
2642 put_buf(r10_bio);
2643 } else {
2644 bool fail = false;
2645 for (m = 0; m < conf->copies; m++) {
2646 int dev = r10_bio->devs[m].devnum;
2647 struct bio *bio = r10_bio->devs[m].bio;
2648 rdev = conf->mirrors[dev].rdev;
2649 if (bio == IO_MADE_GOOD) {
2650 rdev_clear_badblocks(
2651 rdev,
2652 r10_bio->devs[m].addr,
2653 r10_bio->sectors, 0);
2654 rdev_dec_pending(rdev, conf->mddev);
2655 } else if (bio != NULL && bio->bi_status) {
2656 fail = true;
2657 if (!narrow_write_error(r10_bio, m)) {
2658 md_error(conf->mddev, rdev);
2659 set_bit(R10BIO_Degraded,
2660 &r10_bio->state);
2661 }
2662 rdev_dec_pending(rdev, conf->mddev);
2663 }
2664 bio = r10_bio->devs[m].repl_bio;
2665 rdev = conf->mirrors[dev].replacement;
2666 if (rdev && bio == IO_MADE_GOOD) {
2667 rdev_clear_badblocks(
2668 rdev,
2669 r10_bio->devs[m].addr,
2670 r10_bio->sectors, 0);
2671 rdev_dec_pending(rdev, conf->mddev);
2672 }
2673 }
2674 if (fail) {
2675 spin_lock_irq(&conf->device_lock);
2676 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2677 conf->nr_queued++;
2678 spin_unlock_irq(&conf->device_lock);
2679 /*
2680 * In case freeze_array() is waiting for condition
2681 * nr_pending == nr_queued + extra to be true.
2682 */
2683 wake_up(&conf->wait_barrier);
2684 md_wakeup_thread(conf->mddev->thread);
2685 } else {
2686 if (test_bit(R10BIO_WriteError,
2687 &r10_bio->state))
2688 close_write(r10_bio);
2689 raid_end_bio_io(r10_bio);
2690 }
2691 }
2692}
2693
2694static void raid10d(struct md_thread *thread)
2695{
2696 struct mddev *mddev = thread->mddev;
2697 struct r10bio *r10_bio;
2698 unsigned long flags;
2699 struct r10conf *conf = mddev->private;
2700 struct list_head *head = &conf->retry_list;
2701 struct blk_plug plug;
2702
2703 md_check_recovery(mddev);
2704
2705 if (!list_empty_careful(&conf->bio_end_io_list) &&
2706 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2707 LIST_HEAD(tmp);
2708 spin_lock_irqsave(&conf->device_lock, flags);
2709 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2710 while (!list_empty(&conf->bio_end_io_list)) {
2711 list_move(conf->bio_end_io_list.prev, &tmp);
2712 conf->nr_queued--;
2713 }
2714 }
2715 spin_unlock_irqrestore(&conf->device_lock, flags);
2716 while (!list_empty(&tmp)) {
2717 r10_bio = list_first_entry(&tmp, struct r10bio,
2718 retry_list);
2719 list_del(&r10_bio->retry_list);
2720 if (mddev->degraded)
2721 set_bit(R10BIO_Degraded, &r10_bio->state);
2722
2723 if (test_bit(R10BIO_WriteError,
2724 &r10_bio->state))
2725 close_write(r10_bio);
2726 raid_end_bio_io(r10_bio);
2727 }
2728 }
2729
2730 blk_start_plug(&plug);
2731 for (;;) {
2732
2733 flush_pending_writes(conf);
2734
2735 spin_lock_irqsave(&conf->device_lock, flags);
2736 if (list_empty(head)) {
2737 spin_unlock_irqrestore(&conf->device_lock, flags);
2738 break;
2739 }
2740 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2741 list_del(head->prev);
2742 conf->nr_queued--;
2743 spin_unlock_irqrestore(&conf->device_lock, flags);
2744
2745 mddev = r10_bio->mddev;
2746 conf = mddev->private;
2747 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2748 test_bit(R10BIO_WriteError, &r10_bio->state))
2749 handle_write_completed(conf, r10_bio);
2750 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2751 reshape_request_write(mddev, r10_bio);
2752 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2753 sync_request_write(mddev, r10_bio);
2754 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2755 recovery_request_write(mddev, r10_bio);
2756 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2757 handle_read_error(mddev, r10_bio);
2758 else
2759 WARN_ON_ONCE(1);
2760
2761 cond_resched();
2762 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2763 md_check_recovery(mddev);
2764 }
2765 blk_finish_plug(&plug);
2766}
2767
2768static int init_resync(struct r10conf *conf)
2769{
2770 int ret, buffs, i;
2771
2772 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2773 BUG_ON(mempool_initialized(&conf->r10buf_pool));
2774 conf->have_replacement = 0;
2775 for (i = 0; i < conf->geo.raid_disks; i++)
2776 if (conf->mirrors[i].replacement)
2777 conf->have_replacement = 1;
2778 ret = mempool_init(&conf->r10buf_pool, buffs,
2779 r10buf_pool_alloc, r10buf_pool_free, conf);
2780 if (ret)
2781 return ret;
2782 conf->next_resync = 0;
2783 return 0;
2784}
2785
2786static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2787{
2788 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
2789 struct rsync_pages *rp;
2790 struct bio *bio;
2791 int nalloc;
2792 int i;
2793
2794 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2795 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2796 nalloc = conf->copies; /* resync */
2797 else
2798 nalloc = 2; /* recovery */
2799
2800 for (i = 0; i < nalloc; i++) {
2801 bio = r10bio->devs[i].bio;
2802 rp = bio->bi_private;
2803 bio_reset(bio);
2804 bio->bi_private = rp;
2805 bio = r10bio->devs[i].repl_bio;
2806 if (bio) {
2807 rp = bio->bi_private;
2808 bio_reset(bio);
2809 bio->bi_private = rp;
2810 }
2811 }
2812 return r10bio;
2813}
2814
2815/*
2816 * Set cluster_sync_high since we need other nodes to add the
2817 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2818 */
2819static void raid10_set_cluster_sync_high(struct r10conf *conf)
2820{
2821 sector_t window_size;
2822 int extra_chunk, chunks;
2823
2824 /*
2825 * First, here we define "stripe" as a unit which across
2826 * all member devices one time, so we get chunks by use
2827 * raid_disks / near_copies. Otherwise, if near_copies is
2828 * close to raid_disks, then resync window could increases
2829 * linearly with the increase of raid_disks, which means
2830 * we will suspend a really large IO window while it is not
2831 * necessary. If raid_disks is not divisible by near_copies,
2832 * an extra chunk is needed to ensure the whole "stripe" is
2833 * covered.
2834 */
2835
2836 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2837 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2838 extra_chunk = 0;
2839 else
2840 extra_chunk = 1;
2841 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2842
2843 /*
2844 * At least use a 32M window to align with raid1's resync window
2845 */
2846 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2847 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2848
2849 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2850}
2851
2852/*
2853 * perform a "sync" on one "block"
2854 *
2855 * We need to make sure that no normal I/O request - particularly write
2856 * requests - conflict with active sync requests.
2857 *
2858 * This is achieved by tracking pending requests and a 'barrier' concept
2859 * that can be installed to exclude normal IO requests.
2860 *
2861 * Resync and recovery are handled very differently.
2862 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2863 *
2864 * For resync, we iterate over virtual addresses, read all copies,
2865 * and update if there are differences. If only one copy is live,
2866 * skip it.
2867 * For recovery, we iterate over physical addresses, read a good
2868 * value for each non-in_sync drive, and over-write.
2869 *
2870 * So, for recovery we may have several outstanding complex requests for a
2871 * given address, one for each out-of-sync device. We model this by allocating
2872 * a number of r10_bio structures, one for each out-of-sync device.
2873 * As we setup these structures, we collect all bio's together into a list
2874 * which we then process collectively to add pages, and then process again
2875 * to pass to submit_bio_noacct.
2876 *
2877 * The r10_bio structures are linked using a borrowed master_bio pointer.
2878 * This link is counted in ->remaining. When the r10_bio that points to NULL
2879 * has its remaining count decremented to 0, the whole complex operation
2880 * is complete.
2881 *
2882 */
2883
2884static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2885 int *skipped)
2886{
2887 struct r10conf *conf = mddev->private;
2888 struct r10bio *r10_bio;
2889 struct bio *biolist = NULL, *bio;
2890 sector_t max_sector, nr_sectors;
2891 int i;
2892 int max_sync;
2893 sector_t sync_blocks;
2894 sector_t sectors_skipped = 0;
2895 int chunks_skipped = 0;
2896 sector_t chunk_mask = conf->geo.chunk_mask;
2897 int page_idx = 0;
2898
2899 if (!mempool_initialized(&conf->r10buf_pool))
2900 if (init_resync(conf))
2901 return 0;
2902
2903 /*
2904 * Allow skipping a full rebuild for incremental assembly
2905 * of a clean array, like RAID1 does.
2906 */
2907 if (mddev->bitmap == NULL &&
2908 mddev->recovery_cp == MaxSector &&
2909 mddev->reshape_position == MaxSector &&
2910 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2911 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2912 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2913 conf->fullsync == 0) {
2914 *skipped = 1;
2915 return mddev->dev_sectors - sector_nr;
2916 }
2917
2918 skipped:
2919 max_sector = mddev->dev_sectors;
2920 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2921 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2922 max_sector = mddev->resync_max_sectors;
2923 if (sector_nr >= max_sector) {
2924 conf->cluster_sync_low = 0;
2925 conf->cluster_sync_high = 0;
2926
2927 /* If we aborted, we need to abort the
2928 * sync on the 'current' bitmap chucks (there can
2929 * be several when recovering multiple devices).
2930 * as we may have started syncing it but not finished.
2931 * We can find the current address in
2932 * mddev->curr_resync, but for recovery,
2933 * we need to convert that to several
2934 * virtual addresses.
2935 */
2936 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2937 end_reshape(conf);
2938 close_sync(conf);
2939 return 0;
2940 }
2941
2942 if (mddev->curr_resync < max_sector) { /* aborted */
2943 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2944 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2945 &sync_blocks, 1);
2946 else for (i = 0; i < conf->geo.raid_disks; i++) {
2947 sector_t sect =
2948 raid10_find_virt(conf, mddev->curr_resync, i);
2949 md_bitmap_end_sync(mddev->bitmap, sect,
2950 &sync_blocks, 1);
2951 }
2952 } else {
2953 /* completed sync */
2954 if ((!mddev->bitmap || conf->fullsync)
2955 && conf->have_replacement
2956 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2957 /* Completed a full sync so the replacements
2958 * are now fully recovered.
2959 */
2960 rcu_read_lock();
2961 for (i = 0; i < conf->geo.raid_disks; i++) {
2962 struct md_rdev *rdev =
2963 rcu_dereference(conf->mirrors[i].replacement);
2964 if (rdev)
2965 rdev->recovery_offset = MaxSector;
2966 }
2967 rcu_read_unlock();
2968 }
2969 conf->fullsync = 0;
2970 }
2971 md_bitmap_close_sync(mddev->bitmap);
2972 close_sync(conf);
2973 *skipped = 1;
2974 return sectors_skipped;
2975 }
2976
2977 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2978 return reshape_request(mddev, sector_nr, skipped);
2979
2980 if (chunks_skipped >= conf->geo.raid_disks) {
2981 /* if there has been nothing to do on any drive,
2982 * then there is nothing to do at all..
2983 */
2984 *skipped = 1;
2985 return (max_sector - sector_nr) + sectors_skipped;
2986 }
2987
2988 if (max_sector > mddev->resync_max)
2989 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2990
2991 /* make sure whole request will fit in a chunk - if chunks
2992 * are meaningful
2993 */
2994 if (conf->geo.near_copies < conf->geo.raid_disks &&
2995 max_sector > (sector_nr | chunk_mask))
2996 max_sector = (sector_nr | chunk_mask) + 1;
2997
2998 /*
2999 * If there is non-resync activity waiting for a turn, then let it
3000 * though before starting on this new sync request.
3001 */
3002 if (conf->nr_waiting)
3003 schedule_timeout_uninterruptible(1);
3004
3005 /* Again, very different code for resync and recovery.
3006 * Both must result in an r10bio with a list of bios that
3007 * have bi_end_io, bi_sector, bi_disk set,
3008 * and bi_private set to the r10bio.
3009 * For recovery, we may actually create several r10bios
3010 * with 2 bios in each, that correspond to the bios in the main one.
3011 * In this case, the subordinate r10bios link back through a
3012 * borrowed master_bio pointer, and the counter in the master
3013 * includes a ref from each subordinate.
3014 */
3015 /* First, we decide what to do and set ->bi_end_io
3016 * To end_sync_read if we want to read, and
3017 * end_sync_write if we will want to write.
3018 */
3019
3020 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3021 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3022 /* recovery... the complicated one */
3023 int j;
3024 r10_bio = NULL;
3025
3026 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3027 int still_degraded;
3028 struct r10bio *rb2;
3029 sector_t sect;
3030 int must_sync;
3031 int any_working;
3032 int need_recover = 0;
3033 int need_replace = 0;
3034 struct raid10_info *mirror = &conf->mirrors[i];
3035 struct md_rdev *mrdev, *mreplace;
3036
3037 rcu_read_lock();
3038 mrdev = rcu_dereference(mirror->rdev);
3039 mreplace = rcu_dereference(mirror->replacement);
3040
3041 if (mrdev != NULL &&
3042 !test_bit(Faulty, &mrdev->flags) &&
3043 !test_bit(In_sync, &mrdev->flags))
3044 need_recover = 1;
3045 if (mreplace != NULL &&
3046 !test_bit(Faulty, &mreplace->flags))
3047 need_replace = 1;
3048
3049 if (!need_recover && !need_replace) {
3050 rcu_read_unlock();
3051 continue;
3052 }
3053
3054 still_degraded = 0;
3055 /* want to reconstruct this device */
3056 rb2 = r10_bio;
3057 sect = raid10_find_virt(conf, sector_nr, i);
3058 if (sect >= mddev->resync_max_sectors) {
3059 /* last stripe is not complete - don't
3060 * try to recover this sector.
3061 */
3062 rcu_read_unlock();
3063 continue;
3064 }
3065 if (mreplace && test_bit(Faulty, &mreplace->flags))
3066 mreplace = NULL;
3067 /* Unless we are doing a full sync, or a replacement
3068 * we only need to recover the block if it is set in
3069 * the bitmap
3070 */
3071 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3072 &sync_blocks, 1);
3073 if (sync_blocks < max_sync)
3074 max_sync = sync_blocks;
3075 if (!must_sync &&
3076 mreplace == NULL &&
3077 !conf->fullsync) {
3078 /* yep, skip the sync_blocks here, but don't assume
3079 * that there will never be anything to do here
3080 */
3081 chunks_skipped = -1;
3082 rcu_read_unlock();
3083 continue;
3084 }
3085 atomic_inc(&mrdev->nr_pending);
3086 if (mreplace)
3087 atomic_inc(&mreplace->nr_pending);
3088 rcu_read_unlock();
3089
3090 r10_bio = raid10_alloc_init_r10buf(conf);
3091 r10_bio->state = 0;
3092 raise_barrier(conf, rb2 != NULL);
3093 atomic_set(&r10_bio->remaining, 0);
3094
3095 r10_bio->master_bio = (struct bio*)rb2;
3096 if (rb2)
3097 atomic_inc(&rb2->remaining);
3098 r10_bio->mddev = mddev;
3099 set_bit(R10BIO_IsRecover, &r10_bio->state);
3100 r10_bio->sector = sect;
3101
3102 raid10_find_phys(conf, r10_bio);
3103
3104 /* Need to check if the array will still be
3105 * degraded
3106 */
3107 rcu_read_lock();
3108 for (j = 0; j < conf->geo.raid_disks; j++) {
3109 struct md_rdev *rdev = rcu_dereference(
3110 conf->mirrors[j].rdev);
3111 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3112 still_degraded = 1;
3113 break;
3114 }
3115 }
3116
3117 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3118 &sync_blocks, still_degraded);
3119
3120 any_working = 0;
3121 for (j=0; j<conf->copies;j++) {
3122 int k;
3123 int d = r10_bio->devs[j].devnum;
3124 sector_t from_addr, to_addr;
3125 struct md_rdev *rdev =
3126 rcu_dereference(conf->mirrors[d].rdev);
3127 sector_t sector, first_bad;
3128 int bad_sectors;
3129 if (!rdev ||
3130 !test_bit(In_sync, &rdev->flags))
3131 continue;
3132 /* This is where we read from */
3133 any_working = 1;
3134 sector = r10_bio->devs[j].addr;
3135
3136 if (is_badblock(rdev, sector, max_sync,
3137 &first_bad, &bad_sectors)) {
3138 if (first_bad > sector)
3139 max_sync = first_bad - sector;
3140 else {
3141 bad_sectors -= (sector
3142 - first_bad);
3143 if (max_sync > bad_sectors)
3144 max_sync = bad_sectors;
3145 continue;
3146 }
3147 }
3148 bio = r10_bio->devs[0].bio;
3149 bio->bi_next = biolist;
3150 biolist = bio;
3151 bio->bi_end_io = end_sync_read;
3152 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3153 if (test_bit(FailFast, &rdev->flags))
3154 bio->bi_opf |= MD_FAILFAST;
3155 from_addr = r10_bio->devs[j].addr;
3156 bio->bi_iter.bi_sector = from_addr +
3157 rdev->data_offset;
3158 bio_set_dev(bio, rdev->bdev);
3159 atomic_inc(&rdev->nr_pending);
3160 /* and we write to 'i' (if not in_sync) */
3161
3162 for (k=0; k<conf->copies; k++)
3163 if (r10_bio->devs[k].devnum == i)
3164 break;
3165 BUG_ON(k == conf->copies);
3166 to_addr = r10_bio->devs[k].addr;
3167 r10_bio->devs[0].devnum = d;
3168 r10_bio->devs[0].addr = from_addr;
3169 r10_bio->devs[1].devnum = i;
3170 r10_bio->devs[1].addr = to_addr;
3171
3172 if (need_recover) {
3173 bio = r10_bio->devs[1].bio;
3174 bio->bi_next = biolist;
3175 biolist = bio;
3176 bio->bi_end_io = end_sync_write;
3177 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3178 bio->bi_iter.bi_sector = to_addr
3179 + mrdev->data_offset;
3180 bio_set_dev(bio, mrdev->bdev);
3181 atomic_inc(&r10_bio->remaining);
3182 } else
3183 r10_bio->devs[1].bio->bi_end_io = NULL;
3184
3185 /* and maybe write to replacement */
3186 bio = r10_bio->devs[1].repl_bio;
3187 if (bio)
3188 bio->bi_end_io = NULL;
3189 /* Note: if need_replace, then bio
3190 * cannot be NULL as r10buf_pool_alloc will
3191 * have allocated it.
3192 */
3193 if (!need_replace)
3194 break;
3195 bio->bi_next = biolist;
3196 biolist = bio;
3197 bio->bi_end_io = end_sync_write;
3198 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3199 bio->bi_iter.bi_sector = to_addr +
3200 mreplace->data_offset;
3201 bio_set_dev(bio, mreplace->bdev);
3202 atomic_inc(&r10_bio->remaining);
3203 break;
3204 }
3205 rcu_read_unlock();
3206 if (j == conf->copies) {
3207 /* Cannot recover, so abort the recovery or
3208 * record a bad block */
3209 if (any_working) {
3210 /* problem is that there are bad blocks
3211 * on other device(s)
3212 */
3213 int k;
3214 for (k = 0; k < conf->copies; k++)
3215 if (r10_bio->devs[k].devnum == i)
3216 break;
3217 if (!test_bit(In_sync,
3218 &mrdev->flags)
3219 && !rdev_set_badblocks(
3220 mrdev,
3221 r10_bio->devs[k].addr,
3222 max_sync, 0))
3223 any_working = 0;
3224 if (mreplace &&
3225 !rdev_set_badblocks(
3226 mreplace,
3227 r10_bio->devs[k].addr,
3228 max_sync, 0))
3229 any_working = 0;
3230 }
3231 if (!any_working) {
3232 if (!test_and_set_bit(MD_RECOVERY_INTR,
3233 &mddev->recovery))
3234 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3235 mdname(mddev));
3236 mirror->recovery_disabled
3237 = mddev->recovery_disabled;
3238 }
3239 put_buf(r10_bio);
3240 if (rb2)
3241 atomic_dec(&rb2->remaining);
3242 r10_bio = rb2;
3243 rdev_dec_pending(mrdev, mddev);
3244 if (mreplace)
3245 rdev_dec_pending(mreplace, mddev);
3246 break;
3247 }
3248 rdev_dec_pending(mrdev, mddev);
3249 if (mreplace)
3250 rdev_dec_pending(mreplace, mddev);
3251 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3252 /* Only want this if there is elsewhere to
3253 * read from. 'j' is currently the first
3254 * readable copy.
3255 */
3256 int targets = 1;
3257 for (; j < conf->copies; j++) {
3258 int d = r10_bio->devs[j].devnum;
3259 if (conf->mirrors[d].rdev &&
3260 test_bit(In_sync,
3261 &conf->mirrors[d].rdev->flags))
3262 targets++;
3263 }
3264 if (targets == 1)
3265 r10_bio->devs[0].bio->bi_opf
3266 &= ~MD_FAILFAST;
3267 }
3268 }
3269 if (biolist == NULL) {
3270 while (r10_bio) {
3271 struct r10bio *rb2 = r10_bio;
3272 r10_bio = (struct r10bio*) rb2->master_bio;
3273 rb2->master_bio = NULL;
3274 put_buf(rb2);
3275 }
3276 goto giveup;
3277 }
3278 } else {
3279 /* resync. Schedule a read for every block at this virt offset */
3280 int count = 0;
3281
3282 /*
3283 * Since curr_resync_completed could probably not update in
3284 * time, and we will set cluster_sync_low based on it.
3285 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3286 * safety reason, which ensures curr_resync_completed is
3287 * updated in bitmap_cond_end_sync.
3288 */
3289 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3290 mddev_is_clustered(mddev) &&
3291 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3292
3293 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3294 &sync_blocks, mddev->degraded) &&
3295 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3296 &mddev->recovery)) {
3297 /* We can skip this block */
3298 *skipped = 1;
3299 return sync_blocks + sectors_skipped;
3300 }
3301 if (sync_blocks < max_sync)
3302 max_sync = sync_blocks;
3303 r10_bio = raid10_alloc_init_r10buf(conf);
3304 r10_bio->state = 0;
3305
3306 r10_bio->mddev = mddev;
3307 atomic_set(&r10_bio->remaining, 0);
3308 raise_barrier(conf, 0);
3309 conf->next_resync = sector_nr;
3310
3311 r10_bio->master_bio = NULL;
3312 r10_bio->sector = sector_nr;
3313 set_bit(R10BIO_IsSync, &r10_bio->state);
3314 raid10_find_phys(conf, r10_bio);
3315 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3316
3317 for (i = 0; i < conf->copies; i++) {
3318 int d = r10_bio->devs[i].devnum;
3319 sector_t first_bad, sector;
3320 int bad_sectors;
3321 struct md_rdev *rdev;
3322
3323 if (r10_bio->devs[i].repl_bio)
3324 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3325
3326 bio = r10_bio->devs[i].bio;
3327 bio->bi_status = BLK_STS_IOERR;
3328 rcu_read_lock();
3329 rdev = rcu_dereference(conf->mirrors[d].rdev);
3330 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3331 rcu_read_unlock();
3332 continue;
3333 }
3334 sector = r10_bio->devs[i].addr;
3335 if (is_badblock(rdev, sector, max_sync,
3336 &first_bad, &bad_sectors)) {
3337 if (first_bad > sector)
3338 max_sync = first_bad - sector;
3339 else {
3340 bad_sectors -= (sector - first_bad);
3341 if (max_sync > bad_sectors)
3342 max_sync = bad_sectors;
3343 rcu_read_unlock();
3344 continue;
3345 }
3346 }
3347 atomic_inc(&rdev->nr_pending);
3348 atomic_inc(&r10_bio->remaining);
3349 bio->bi_next = biolist;
3350 biolist = bio;
3351 bio->bi_end_io = end_sync_read;
3352 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3353 if (test_bit(FailFast, &rdev->flags))
3354 bio->bi_opf |= MD_FAILFAST;
3355 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3356 bio_set_dev(bio, rdev->bdev);
3357 count++;
3358
3359 rdev = rcu_dereference(conf->mirrors[d].replacement);
3360 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3361 rcu_read_unlock();
3362 continue;
3363 }
3364 atomic_inc(&rdev->nr_pending);
3365
3366 /* Need to set up for writing to the replacement */
3367 bio = r10_bio->devs[i].repl_bio;
3368 bio->bi_status = BLK_STS_IOERR;
3369
3370 sector = r10_bio->devs[i].addr;
3371 bio->bi_next = biolist;
3372 biolist = bio;
3373 bio->bi_end_io = end_sync_write;
3374 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3375 if (test_bit(FailFast, &rdev->flags))
3376 bio->bi_opf |= MD_FAILFAST;
3377 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3378 bio_set_dev(bio, rdev->bdev);
3379 count++;
3380 rcu_read_unlock();
3381 }
3382
3383 if (count < 2) {
3384 for (i=0; i<conf->copies; i++) {
3385 int d = r10_bio->devs[i].devnum;
3386 if (r10_bio->devs[i].bio->bi_end_io)
3387 rdev_dec_pending(conf->mirrors[d].rdev,
3388 mddev);
3389 if (r10_bio->devs[i].repl_bio &&
3390 r10_bio->devs[i].repl_bio->bi_end_io)
3391 rdev_dec_pending(
3392 conf->mirrors[d].replacement,
3393 mddev);
3394 }
3395 put_buf(r10_bio);
3396 biolist = NULL;
3397 goto giveup;
3398 }
3399 }
3400
3401 nr_sectors = 0;
3402 if (sector_nr + max_sync < max_sector)
3403 max_sector = sector_nr + max_sync;
3404 do {
3405 struct page *page;
3406 int len = PAGE_SIZE;
3407 if (sector_nr + (len>>9) > max_sector)
3408 len = (max_sector - sector_nr) << 9;
3409 if (len == 0)
3410 break;
3411 for (bio= biolist ; bio ; bio=bio->bi_next) {
3412 struct resync_pages *rp = get_resync_pages(bio);
3413 page = resync_fetch_page(rp, page_idx);
3414 /*
3415 * won't fail because the vec table is big enough
3416 * to hold all these pages
3417 */
3418 bio_add_page(bio, page, len, 0);
3419 }
3420 nr_sectors += len>>9;
3421 sector_nr += len>>9;
3422 } while (++page_idx < RESYNC_PAGES);
3423 r10_bio->sectors = nr_sectors;
3424
3425 if (mddev_is_clustered(mddev) &&
3426 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3427 /* It is resync not recovery */
3428 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3429 conf->cluster_sync_low = mddev->curr_resync_completed;
3430 raid10_set_cluster_sync_high(conf);
3431 /* Send resync message */
3432 md_cluster_ops->resync_info_update(mddev,
3433 conf->cluster_sync_low,
3434 conf->cluster_sync_high);
3435 }
3436 } else if (mddev_is_clustered(mddev)) {
3437 /* This is recovery not resync */
3438 sector_t sect_va1, sect_va2;
3439 bool broadcast_msg = false;
3440
3441 for (i = 0; i < conf->geo.raid_disks; i++) {
3442 /*
3443 * sector_nr is a device address for recovery, so we
3444 * need translate it to array address before compare
3445 * with cluster_sync_high.
3446 */
3447 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3448
3449 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3450 broadcast_msg = true;
3451 /*
3452 * curr_resync_completed is similar as
3453 * sector_nr, so make the translation too.
3454 */
3455 sect_va2 = raid10_find_virt(conf,
3456 mddev->curr_resync_completed, i);
3457
3458 if (conf->cluster_sync_low == 0 ||
3459 conf->cluster_sync_low > sect_va2)
3460 conf->cluster_sync_low = sect_va2;
3461 }
3462 }
3463 if (broadcast_msg) {
3464 raid10_set_cluster_sync_high(conf);
3465 md_cluster_ops->resync_info_update(mddev,
3466 conf->cluster_sync_low,
3467 conf->cluster_sync_high);
3468 }
3469 }
3470
3471 while (biolist) {
3472 bio = biolist;
3473 biolist = biolist->bi_next;
3474
3475 bio->bi_next = NULL;
3476 r10_bio = get_resync_r10bio(bio);
3477 r10_bio->sectors = nr_sectors;
3478
3479 if (bio->bi_end_io == end_sync_read) {
3480 md_sync_acct_bio(bio, nr_sectors);
3481 bio->bi_status = 0;
3482 submit_bio_noacct(bio);
3483 }
3484 }
3485
3486 if (sectors_skipped)
3487 /* pretend they weren't skipped, it makes
3488 * no important difference in this case
3489 */
3490 md_done_sync(mddev, sectors_skipped, 1);
3491
3492 return sectors_skipped + nr_sectors;
3493 giveup:
3494 /* There is nowhere to write, so all non-sync
3495 * drives must be failed or in resync, all drives
3496 * have a bad block, so try the next chunk...
3497 */
3498 if (sector_nr + max_sync < max_sector)
3499 max_sector = sector_nr + max_sync;
3500
3501 sectors_skipped += (max_sector - sector_nr);
3502 chunks_skipped ++;
3503 sector_nr = max_sector;
3504 goto skipped;
3505}
3506
3507static sector_t
3508raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3509{
3510 sector_t size;
3511 struct r10conf *conf = mddev->private;
3512
3513 if (!raid_disks)
3514 raid_disks = min(conf->geo.raid_disks,
3515 conf->prev.raid_disks);
3516 if (!sectors)
3517 sectors = conf->dev_sectors;
3518
3519 size = sectors >> conf->geo.chunk_shift;
3520 sector_div(size, conf->geo.far_copies);
3521 size = size * raid_disks;
3522 sector_div(size, conf->geo.near_copies);
3523
3524 return size << conf->geo.chunk_shift;
3525}
3526
3527static void calc_sectors(struct r10conf *conf, sector_t size)
3528{
3529 /* Calculate the number of sectors-per-device that will
3530 * actually be used, and set conf->dev_sectors and
3531 * conf->stride
3532 */
3533
3534 size = size >> conf->geo.chunk_shift;
3535 sector_div(size, conf->geo.far_copies);
3536 size = size * conf->geo.raid_disks;
3537 sector_div(size, conf->geo.near_copies);
3538 /* 'size' is now the number of chunks in the array */
3539 /* calculate "used chunks per device" */
3540 size = size * conf->copies;
3541
3542 /* We need to round up when dividing by raid_disks to
3543 * get the stride size.
3544 */
3545 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3546
3547 conf->dev_sectors = size << conf->geo.chunk_shift;
3548
3549 if (conf->geo.far_offset)
3550 conf->geo.stride = 1 << conf->geo.chunk_shift;
3551 else {
3552 sector_div(size, conf->geo.far_copies);
3553 conf->geo.stride = size << conf->geo.chunk_shift;
3554 }
3555}
3556
3557enum geo_type {geo_new, geo_old, geo_start};
3558static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3559{
3560 int nc, fc, fo;
3561 int layout, chunk, disks;
3562 switch (new) {
3563 case geo_old:
3564 layout = mddev->layout;
3565 chunk = mddev->chunk_sectors;
3566 disks = mddev->raid_disks - mddev->delta_disks;
3567 break;
3568 case geo_new:
3569 layout = mddev->new_layout;
3570 chunk = mddev->new_chunk_sectors;
3571 disks = mddev->raid_disks;
3572 break;
3573 default: /* avoid 'may be unused' warnings */
3574 case geo_start: /* new when starting reshape - raid_disks not
3575 * updated yet. */
3576 layout = mddev->new_layout;
3577 chunk = mddev->new_chunk_sectors;
3578 disks = mddev->raid_disks + mddev->delta_disks;
3579 break;
3580 }
3581 if (layout >> 19)
3582 return -1;
3583 if (chunk < (PAGE_SIZE >> 9) ||
3584 !is_power_of_2(chunk))
3585 return -2;
3586 nc = layout & 255;
3587 fc = (layout >> 8) & 255;
3588 fo = layout & (1<<16);
3589 geo->raid_disks = disks;
3590 geo->near_copies = nc;
3591 geo->far_copies = fc;
3592 geo->far_offset = fo;
3593 switch (layout >> 17) {
3594 case 0: /* original layout. simple but not always optimal */
3595 geo->far_set_size = disks;
3596 break;
3597 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3598 * actually using this, but leave code here just in case.*/
3599 geo->far_set_size = disks/fc;
3600 WARN(geo->far_set_size < fc,
3601 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3602 break;
3603 case 2: /* "improved" layout fixed to match documentation */
3604 geo->far_set_size = fc * nc;
3605 break;
3606 default: /* Not a valid layout */
3607 return -1;
3608 }
3609 geo->chunk_mask = chunk - 1;
3610 geo->chunk_shift = ffz(~chunk);
3611 return nc*fc;
3612}
3613
3614static struct r10conf *setup_conf(struct mddev *mddev)
3615{
3616 struct r10conf *conf = NULL;
3617 int err = -EINVAL;
3618 struct geom geo;
3619 int copies;
3620
3621 copies = setup_geo(&geo, mddev, geo_new);
3622
3623 if (copies == -2) {
3624 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3625 mdname(mddev), PAGE_SIZE);
3626 goto out;
3627 }
3628
3629 if (copies < 2 || copies > mddev->raid_disks) {
3630 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3631 mdname(mddev), mddev->new_layout);
3632 goto out;
3633 }
3634
3635 err = -ENOMEM;
3636 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3637 if (!conf)
3638 goto out;
3639
3640 /* FIXME calc properly */
3641 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3642 sizeof(struct raid10_info),
3643 GFP_KERNEL);
3644 if (!conf->mirrors)
3645 goto out;
3646
3647 conf->tmppage = alloc_page(GFP_KERNEL);
3648 if (!conf->tmppage)
3649 goto out;
3650
3651 conf->geo = geo;
3652 conf->copies = copies;
3653 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3654 rbio_pool_free, conf);
3655 if (err)
3656 goto out;
3657
3658 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3659 if (err)
3660 goto out;
3661
3662 calc_sectors(conf, mddev->dev_sectors);
3663 if (mddev->reshape_position == MaxSector) {
3664 conf->prev = conf->geo;
3665 conf->reshape_progress = MaxSector;
3666 } else {
3667 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3668 err = -EINVAL;
3669 goto out;
3670 }
3671 conf->reshape_progress = mddev->reshape_position;
3672 if (conf->prev.far_offset)
3673 conf->prev.stride = 1 << conf->prev.chunk_shift;
3674 else
3675 /* far_copies must be 1 */
3676 conf->prev.stride = conf->dev_sectors;
3677 }
3678 conf->reshape_safe = conf->reshape_progress;
3679 spin_lock_init(&conf->device_lock);
3680 INIT_LIST_HEAD(&conf->retry_list);
3681 INIT_LIST_HEAD(&conf->bio_end_io_list);
3682
3683 spin_lock_init(&conf->resync_lock);
3684 init_waitqueue_head(&conf->wait_barrier);
3685 atomic_set(&conf->nr_pending, 0);
3686
3687 err = -ENOMEM;
3688 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3689 if (!conf->thread)
3690 goto out;
3691
3692 conf->mddev = mddev;
3693 return conf;
3694
3695 out:
3696 if (conf) {
3697 mempool_exit(&conf->r10bio_pool);
3698 kfree(conf->mirrors);
3699 safe_put_page(conf->tmppage);
3700 bioset_exit(&conf->bio_split);
3701 kfree(conf);
3702 }
3703 return ERR_PTR(err);
3704}
3705
3706static int raid10_run(struct mddev *mddev)
3707{
3708 struct r10conf *conf;
3709 int i, disk_idx, chunk_size;
3710 struct raid10_info *disk;
3711 struct md_rdev *rdev;
3712 sector_t size;
3713 sector_t min_offset_diff = 0;
3714 int first = 1;
3715 bool discard_supported = false;
3716
3717 if (mddev_init_writes_pending(mddev) < 0)
3718 return -ENOMEM;
3719
3720 if (mddev->private == NULL) {
3721 conf = setup_conf(mddev);
3722 if (IS_ERR(conf))
3723 return PTR_ERR(conf);
3724 mddev->private = conf;
3725 }
3726 conf = mddev->private;
3727 if (!conf)
3728 goto out;
3729
3730 if (mddev_is_clustered(conf->mddev)) {
3731 int fc, fo;
3732
3733 fc = (mddev->layout >> 8) & 255;
3734 fo = mddev->layout & (1<<16);
3735 if (fc > 1 || fo > 0) {
3736 pr_err("only near layout is supported by clustered"
3737 " raid10\n");
3738 goto out_free_conf;
3739 }
3740 }
3741
3742 mddev->thread = conf->thread;
3743 conf->thread = NULL;
3744
3745 chunk_size = mddev->chunk_sectors << 9;
3746 if (mddev->queue) {
3747 blk_queue_max_discard_sectors(mddev->queue,
3748 mddev->chunk_sectors);
3749 blk_queue_max_write_same_sectors(mddev->queue, 0);
3750 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3751 blk_queue_io_min(mddev->queue, chunk_size);
3752 if (conf->geo.raid_disks % conf->geo.near_copies)
3753 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3754 else
3755 blk_queue_io_opt(mddev->queue, chunk_size *
3756 (conf->geo.raid_disks / conf->geo.near_copies));
3757 }
3758
3759 rdev_for_each(rdev, mddev) {
3760 long long diff;
3761
3762 disk_idx = rdev->raid_disk;
3763 if (disk_idx < 0)
3764 continue;
3765 if (disk_idx >= conf->geo.raid_disks &&
3766 disk_idx >= conf->prev.raid_disks)
3767 continue;
3768 disk = conf->mirrors + disk_idx;
3769
3770 if (test_bit(Replacement, &rdev->flags)) {
3771 if (disk->replacement)
3772 goto out_free_conf;
3773 disk->replacement = rdev;
3774 } else {
3775 if (disk->rdev)
3776 goto out_free_conf;
3777 disk->rdev = rdev;
3778 }
3779 diff = (rdev->new_data_offset - rdev->data_offset);
3780 if (!mddev->reshape_backwards)
3781 diff = -diff;
3782 if (diff < 0)
3783 diff = 0;
3784 if (first || diff < min_offset_diff)
3785 min_offset_diff = diff;
3786
3787 if (mddev->gendisk)
3788 disk_stack_limits(mddev->gendisk, rdev->bdev,
3789 rdev->data_offset << 9);
3790
3791 disk->head_position = 0;
3792
3793 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3794 discard_supported = true;
3795 first = 0;
3796 }
3797
3798 if (mddev->queue) {
3799 if (discard_supported)
3800 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3801 mddev->queue);
3802 else
3803 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3804 mddev->queue);
3805 }
3806 /* need to check that every block has at least one working mirror */
3807 if (!enough(conf, -1)) {
3808 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3809 mdname(mddev));
3810 goto out_free_conf;
3811 }
3812
3813 if (conf->reshape_progress != MaxSector) {
3814 /* must ensure that shape change is supported */
3815 if (conf->geo.far_copies != 1 &&
3816 conf->geo.far_offset == 0)
3817 goto out_free_conf;
3818 if (conf->prev.far_copies != 1 &&
3819 conf->prev.far_offset == 0)
3820 goto out_free_conf;
3821 }
3822
3823 mddev->degraded = 0;
3824 for (i = 0;
3825 i < conf->geo.raid_disks
3826 || i < conf->prev.raid_disks;
3827 i++) {
3828
3829 disk = conf->mirrors + i;
3830
3831 if (!disk->rdev && disk->replacement) {
3832 /* The replacement is all we have - use it */
3833 disk->rdev = disk->replacement;
3834 disk->replacement = NULL;
3835 clear_bit(Replacement, &disk->rdev->flags);
3836 }
3837
3838 if (!disk->rdev ||
3839 !test_bit(In_sync, &disk->rdev->flags)) {
3840 disk->head_position = 0;
3841 mddev->degraded++;
3842 if (disk->rdev &&
3843 disk->rdev->saved_raid_disk < 0)
3844 conf->fullsync = 1;
3845 }
3846
3847 if (disk->replacement &&
3848 !test_bit(In_sync, &disk->replacement->flags) &&
3849 disk->replacement->saved_raid_disk < 0) {
3850 conf->fullsync = 1;
3851 }
3852
3853 disk->recovery_disabled = mddev->recovery_disabled - 1;
3854 }
3855
3856 if (mddev->recovery_cp != MaxSector)
3857 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3858 mdname(mddev));
3859 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3860 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3861 conf->geo.raid_disks);
3862 /*
3863 * Ok, everything is just fine now
3864 */
3865 mddev->dev_sectors = conf->dev_sectors;
3866 size = raid10_size(mddev, 0, 0);
3867 md_set_array_sectors(mddev, size);
3868 mddev->resync_max_sectors = size;
3869 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3870
3871 if (mddev->queue) {
3872 int stripe = conf->geo.raid_disks *
3873 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3874
3875 /* Calculate max read-ahead size.
3876 * We need to readahead at least twice a whole stripe....
3877 * maybe...
3878 */
3879 stripe /= conf->geo.near_copies;
3880 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3881 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3882 }
3883
3884 if (md_integrity_register(mddev))
3885 goto out_free_conf;
3886
3887 if (conf->reshape_progress != MaxSector) {
3888 unsigned long before_length, after_length;
3889
3890 before_length = ((1 << conf->prev.chunk_shift) *
3891 conf->prev.far_copies);
3892 after_length = ((1 << conf->geo.chunk_shift) *
3893 conf->geo.far_copies);
3894
3895 if (max(before_length, after_length) > min_offset_diff) {
3896 /* This cannot work */
3897 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3898 goto out_free_conf;
3899 }
3900 conf->offset_diff = min_offset_diff;
3901
3902 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3903 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3904 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3905 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3906 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3907 "reshape");
3908 if (!mddev->sync_thread)
3909 goto out_free_conf;
3910 }
3911
3912 return 0;
3913
3914out_free_conf:
3915 md_unregister_thread(&mddev->thread);
3916 mempool_exit(&conf->r10bio_pool);
3917 safe_put_page(conf->tmppage);
3918 kfree(conf->mirrors);
3919 kfree(conf);
3920 mddev->private = NULL;
3921out:
3922 return -EIO;
3923}
3924
3925static void raid10_free(struct mddev *mddev, void *priv)
3926{
3927 struct r10conf *conf = priv;
3928
3929 mempool_exit(&conf->r10bio_pool);
3930 safe_put_page(conf->tmppage);
3931 kfree(conf->mirrors);
3932 kfree(conf->mirrors_old);
3933 kfree(conf->mirrors_new);
3934 bioset_exit(&conf->bio_split);
3935 kfree(conf);
3936}
3937
3938static void raid10_quiesce(struct mddev *mddev, int quiesce)
3939{
3940 struct r10conf *conf = mddev->private;
3941
3942 if (quiesce)
3943 raise_barrier(conf, 0);
3944 else
3945 lower_barrier(conf);
3946}
3947
3948static int raid10_resize(struct mddev *mddev, sector_t sectors)
3949{
3950 /* Resize of 'far' arrays is not supported.
3951 * For 'near' and 'offset' arrays we can set the
3952 * number of sectors used to be an appropriate multiple
3953 * of the chunk size.
3954 * For 'offset', this is far_copies*chunksize.
3955 * For 'near' the multiplier is the LCM of
3956 * near_copies and raid_disks.
3957 * So if far_copies > 1 && !far_offset, fail.
3958 * Else find LCM(raid_disks, near_copy)*far_copies and
3959 * multiply by chunk_size. Then round to this number.
3960 * This is mostly done by raid10_size()
3961 */
3962 struct r10conf *conf = mddev->private;
3963 sector_t oldsize, size;
3964
3965 if (mddev->reshape_position != MaxSector)
3966 return -EBUSY;
3967
3968 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3969 return -EINVAL;
3970
3971 oldsize = raid10_size(mddev, 0, 0);
3972 size = raid10_size(mddev, sectors, 0);
3973 if (mddev->external_size &&
3974 mddev->array_sectors > size)
3975 return -EINVAL;
3976 if (mddev->bitmap) {
3977 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
3978 if (ret)
3979 return ret;
3980 }
3981 md_set_array_sectors(mddev, size);
3982 if (sectors > mddev->dev_sectors &&
3983 mddev->recovery_cp > oldsize) {
3984 mddev->recovery_cp = oldsize;
3985 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3986 }
3987 calc_sectors(conf, sectors);
3988 mddev->dev_sectors = conf->dev_sectors;
3989 mddev->resync_max_sectors = size;
3990 return 0;
3991}
3992
3993static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3994{
3995 struct md_rdev *rdev;
3996 struct r10conf *conf;
3997
3998 if (mddev->degraded > 0) {
3999 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4000 mdname(mddev));
4001 return ERR_PTR(-EINVAL);
4002 }
4003 sector_div(size, devs);
4004
4005 /* Set new parameters */
4006 mddev->new_level = 10;
4007 /* new layout: far_copies = 1, near_copies = 2 */
4008 mddev->new_layout = (1<<8) + 2;
4009 mddev->new_chunk_sectors = mddev->chunk_sectors;
4010 mddev->delta_disks = mddev->raid_disks;
4011 mddev->raid_disks *= 2;
4012 /* make sure it will be not marked as dirty */
4013 mddev->recovery_cp = MaxSector;
4014 mddev->dev_sectors = size;
4015
4016 conf = setup_conf(mddev);
4017 if (!IS_ERR(conf)) {
4018 rdev_for_each(rdev, mddev)
4019 if (rdev->raid_disk >= 0) {
4020 rdev->new_raid_disk = rdev->raid_disk * 2;
4021 rdev->sectors = size;
4022 }
4023 conf->barrier = 1;
4024 }
4025
4026 return conf;
4027}
4028
4029static void *raid10_takeover(struct mddev *mddev)
4030{
4031 struct r0conf *raid0_conf;
4032
4033 /* raid10 can take over:
4034 * raid0 - providing it has only two drives
4035 */
4036 if (mddev->level == 0) {
4037 /* for raid0 takeover only one zone is supported */
4038 raid0_conf = mddev->private;
4039 if (raid0_conf->nr_strip_zones > 1) {
4040 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4041 mdname(mddev));
4042 return ERR_PTR(-EINVAL);
4043 }
4044 return raid10_takeover_raid0(mddev,
4045 raid0_conf->strip_zone->zone_end,
4046 raid0_conf->strip_zone->nb_dev);
4047 }
4048 return ERR_PTR(-EINVAL);
4049}
4050
4051static int raid10_check_reshape(struct mddev *mddev)
4052{
4053 /* Called when there is a request to change
4054 * - layout (to ->new_layout)
4055 * - chunk size (to ->new_chunk_sectors)
4056 * - raid_disks (by delta_disks)
4057 * or when trying to restart a reshape that was ongoing.
4058 *
4059 * We need to validate the request and possibly allocate
4060 * space if that might be an issue later.
4061 *
4062 * Currently we reject any reshape of a 'far' mode array,
4063 * allow chunk size to change if new is generally acceptable,
4064 * allow raid_disks to increase, and allow
4065 * a switch between 'near' mode and 'offset' mode.
4066 */
4067 struct r10conf *conf = mddev->private;
4068 struct geom geo;
4069
4070 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4071 return -EINVAL;
4072
4073 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4074 /* mustn't change number of copies */
4075 return -EINVAL;
4076 if (geo.far_copies > 1 && !geo.far_offset)
4077 /* Cannot switch to 'far' mode */
4078 return -EINVAL;
4079
4080 if (mddev->array_sectors & geo.chunk_mask)
4081 /* not factor of array size */
4082 return -EINVAL;
4083
4084 if (!enough(conf, -1))
4085 return -EINVAL;
4086
4087 kfree(conf->mirrors_new);
4088 conf->mirrors_new = NULL;
4089 if (mddev->delta_disks > 0) {
4090 /* allocate new 'mirrors' list */
4091 conf->mirrors_new =
4092 kcalloc(mddev->raid_disks + mddev->delta_disks,
4093 sizeof(struct raid10_info),
4094 GFP_KERNEL);
4095 if (!conf->mirrors_new)
4096 return -ENOMEM;
4097 }
4098 return 0;
4099}
4100
4101/*
4102 * Need to check if array has failed when deciding whether to:
4103 * - start an array
4104 * - remove non-faulty devices
4105 * - add a spare
4106 * - allow a reshape
4107 * This determination is simple when no reshape is happening.
4108 * However if there is a reshape, we need to carefully check
4109 * both the before and after sections.
4110 * This is because some failed devices may only affect one
4111 * of the two sections, and some non-in_sync devices may
4112 * be insync in the section most affected by failed devices.
4113 */
4114static int calc_degraded(struct r10conf *conf)
4115{
4116 int degraded, degraded2;
4117 int i;
4118
4119 rcu_read_lock();
4120 degraded = 0;
4121 /* 'prev' section first */
4122 for (i = 0; i < conf->prev.raid_disks; i++) {
4123 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4124 if (!rdev || test_bit(Faulty, &rdev->flags))
4125 degraded++;
4126 else if (!test_bit(In_sync, &rdev->flags))
4127 /* When we can reduce the number of devices in
4128 * an array, this might not contribute to
4129 * 'degraded'. It does now.
4130 */
4131 degraded++;
4132 }
4133 rcu_read_unlock();
4134 if (conf->geo.raid_disks == conf->prev.raid_disks)
4135 return degraded;
4136 rcu_read_lock();
4137 degraded2 = 0;
4138 for (i = 0; i < conf->geo.raid_disks; i++) {
4139 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4140 if (!rdev || test_bit(Faulty, &rdev->flags))
4141 degraded2++;
4142 else if (!test_bit(In_sync, &rdev->flags)) {
4143 /* If reshape is increasing the number of devices,
4144 * this section has already been recovered, so
4145 * it doesn't contribute to degraded.
4146 * else it does.
4147 */
4148 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4149 degraded2++;
4150 }
4151 }
4152 rcu_read_unlock();
4153 if (degraded2 > degraded)
4154 return degraded2;
4155 return degraded;
4156}
4157
4158static int raid10_start_reshape(struct mddev *mddev)
4159{
4160 /* A 'reshape' has been requested. This commits
4161 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4162 * This also checks if there are enough spares and adds them
4163 * to the array.
4164 * We currently require enough spares to make the final
4165 * array non-degraded. We also require that the difference
4166 * between old and new data_offset - on each device - is
4167 * enough that we never risk over-writing.
4168 */
4169
4170 unsigned long before_length, after_length;
4171 sector_t min_offset_diff = 0;
4172 int first = 1;
4173 struct geom new;
4174 struct r10conf *conf = mddev->private;
4175 struct md_rdev *rdev;
4176 int spares = 0;
4177 int ret;
4178
4179 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4180 return -EBUSY;
4181
4182 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4183 return -EINVAL;
4184
4185 before_length = ((1 << conf->prev.chunk_shift) *
4186 conf->prev.far_copies);
4187 after_length = ((1 << conf->geo.chunk_shift) *
4188 conf->geo.far_copies);
4189
4190 rdev_for_each(rdev, mddev) {
4191 if (!test_bit(In_sync, &rdev->flags)
4192 && !test_bit(Faulty, &rdev->flags))
4193 spares++;
4194 if (rdev->raid_disk >= 0) {
4195 long long diff = (rdev->new_data_offset
4196 - rdev->data_offset);
4197 if (!mddev->reshape_backwards)
4198 diff = -diff;
4199 if (diff < 0)
4200 diff = 0;
4201 if (first || diff < min_offset_diff)
4202 min_offset_diff = diff;
4203 first = 0;
4204 }
4205 }
4206
4207 if (max(before_length, after_length) > min_offset_diff)
4208 return -EINVAL;
4209
4210 if (spares < mddev->delta_disks)
4211 return -EINVAL;
4212
4213 conf->offset_diff = min_offset_diff;
4214 spin_lock_irq(&conf->device_lock);
4215 if (conf->mirrors_new) {
4216 memcpy(conf->mirrors_new, conf->mirrors,
4217 sizeof(struct raid10_info)*conf->prev.raid_disks);
4218 smp_mb();
4219 kfree(conf->mirrors_old);
4220 conf->mirrors_old = conf->mirrors;
4221 conf->mirrors = conf->mirrors_new;
4222 conf->mirrors_new = NULL;
4223 }
4224 setup_geo(&conf->geo, mddev, geo_start);
4225 smp_mb();
4226 if (mddev->reshape_backwards) {
4227 sector_t size = raid10_size(mddev, 0, 0);
4228 if (size < mddev->array_sectors) {
4229 spin_unlock_irq(&conf->device_lock);
4230 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4231 mdname(mddev));
4232 return -EINVAL;
4233 }
4234 mddev->resync_max_sectors = size;
4235 conf->reshape_progress = size;
4236 } else
4237 conf->reshape_progress = 0;
4238 conf->reshape_safe = conf->reshape_progress;
4239 spin_unlock_irq(&conf->device_lock);
4240
4241 if (mddev->delta_disks && mddev->bitmap) {
4242 struct mdp_superblock_1 *sb = NULL;
4243 sector_t oldsize, newsize;
4244
4245 oldsize = raid10_size(mddev, 0, 0);
4246 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4247
4248 if (!mddev_is_clustered(mddev)) {
4249 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4250 if (ret)
4251 goto abort;
4252 else
4253 goto out;
4254 }
4255
4256 rdev_for_each(rdev, mddev) {
4257 if (rdev->raid_disk > -1 &&
4258 !test_bit(Faulty, &rdev->flags))
4259 sb = page_address(rdev->sb_page);
4260 }
4261
4262 /*
4263 * some node is already performing reshape, and no need to
4264 * call md_bitmap_resize again since it should be called when
4265 * receiving BITMAP_RESIZE msg
4266 */
4267 if ((sb && (le32_to_cpu(sb->feature_map) &
4268 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4269 goto out;
4270
4271 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4272 if (ret)
4273 goto abort;
4274
4275 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4276 if (ret) {
4277 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4278 goto abort;
4279 }
4280 }
4281out:
4282 if (mddev->delta_disks > 0) {
4283 rdev_for_each(rdev, mddev)
4284 if (rdev->raid_disk < 0 &&
4285 !test_bit(Faulty, &rdev->flags)) {
4286 if (raid10_add_disk(mddev, rdev) == 0) {
4287 if (rdev->raid_disk >=
4288 conf->prev.raid_disks)
4289 set_bit(In_sync, &rdev->flags);
4290 else
4291 rdev->recovery_offset = 0;
4292
4293 /* Failure here is OK */
4294 sysfs_link_rdev(mddev, rdev);
4295 }
4296 } else if (rdev->raid_disk >= conf->prev.raid_disks
4297 && !test_bit(Faulty, &rdev->flags)) {
4298 /* This is a spare that was manually added */
4299 set_bit(In_sync, &rdev->flags);
4300 }
4301 }
4302 /* When a reshape changes the number of devices,
4303 * ->degraded is measured against the larger of the
4304 * pre and post numbers.
4305 */
4306 spin_lock_irq(&conf->device_lock);
4307 mddev->degraded = calc_degraded(conf);
4308 spin_unlock_irq(&conf->device_lock);
4309 mddev->raid_disks = conf->geo.raid_disks;
4310 mddev->reshape_position = conf->reshape_progress;
4311 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4312
4313 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4314 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4315 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4316 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4317 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4318
4319 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4320 "reshape");
4321 if (!mddev->sync_thread) {
4322 ret = -EAGAIN;
4323 goto abort;
4324 }
4325 conf->reshape_checkpoint = jiffies;
4326 md_wakeup_thread(mddev->sync_thread);
4327 md_new_event(mddev);
4328 return 0;
4329
4330abort:
4331 mddev->recovery = 0;
4332 spin_lock_irq(&conf->device_lock);
4333 conf->geo = conf->prev;
4334 mddev->raid_disks = conf->geo.raid_disks;
4335 rdev_for_each(rdev, mddev)
4336 rdev->new_data_offset = rdev->data_offset;
4337 smp_wmb();
4338 conf->reshape_progress = MaxSector;
4339 conf->reshape_safe = MaxSector;
4340 mddev->reshape_position = MaxSector;
4341 spin_unlock_irq(&conf->device_lock);
4342 return ret;
4343}
4344
4345/* Calculate the last device-address that could contain
4346 * any block from the chunk that includes the array-address 's'
4347 * and report the next address.
4348 * i.e. the address returned will be chunk-aligned and after
4349 * any data that is in the chunk containing 's'.
4350 */
4351static sector_t last_dev_address(sector_t s, struct geom *geo)
4352{
4353 s = (s | geo->chunk_mask) + 1;
4354 s >>= geo->chunk_shift;
4355 s *= geo->near_copies;
4356 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4357 s *= geo->far_copies;
4358 s <<= geo->chunk_shift;
4359 return s;
4360}
4361
4362/* Calculate the first device-address that could contain
4363 * any block from the chunk that includes the array-address 's'.
4364 * This too will be the start of a chunk
4365 */
4366static sector_t first_dev_address(sector_t s, struct geom *geo)
4367{
4368 s >>= geo->chunk_shift;
4369 s *= geo->near_copies;
4370 sector_div(s, geo->raid_disks);
4371 s *= geo->far_copies;
4372 s <<= geo->chunk_shift;
4373 return s;
4374}
4375
4376static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4377 int *skipped)
4378{
4379 /* We simply copy at most one chunk (smallest of old and new)
4380 * at a time, possibly less if that exceeds RESYNC_PAGES,
4381 * or we hit a bad block or something.
4382 * This might mean we pause for normal IO in the middle of
4383 * a chunk, but that is not a problem as mddev->reshape_position
4384 * can record any location.
4385 *
4386 * If we will want to write to a location that isn't
4387 * yet recorded as 'safe' (i.e. in metadata on disk) then
4388 * we need to flush all reshape requests and update the metadata.
4389 *
4390 * When reshaping forwards (e.g. to more devices), we interpret
4391 * 'safe' as the earliest block which might not have been copied
4392 * down yet. We divide this by previous stripe size and multiply
4393 * by previous stripe length to get lowest device offset that we
4394 * cannot write to yet.
4395 * We interpret 'sector_nr' as an address that we want to write to.
4396 * From this we use last_device_address() to find where we might
4397 * write to, and first_device_address on the 'safe' position.
4398 * If this 'next' write position is after the 'safe' position,
4399 * we must update the metadata to increase the 'safe' position.
4400 *
4401 * When reshaping backwards, we round in the opposite direction
4402 * and perform the reverse test: next write position must not be
4403 * less than current safe position.
4404 *
4405 * In all this the minimum difference in data offsets
4406 * (conf->offset_diff - always positive) allows a bit of slack,
4407 * so next can be after 'safe', but not by more than offset_diff
4408 *
4409 * We need to prepare all the bios here before we start any IO
4410 * to ensure the size we choose is acceptable to all devices.
4411 * The means one for each copy for write-out and an extra one for
4412 * read-in.
4413 * We store the read-in bio in ->master_bio and the others in
4414 * ->devs[x].bio and ->devs[x].repl_bio.
4415 */
4416 struct r10conf *conf = mddev->private;
4417 struct r10bio *r10_bio;
4418 sector_t next, safe, last;
4419 int max_sectors;
4420 int nr_sectors;
4421 int s;
4422 struct md_rdev *rdev;
4423 int need_flush = 0;
4424 struct bio *blist;
4425 struct bio *bio, *read_bio;
4426 int sectors_done = 0;
4427 struct page **pages;
4428
4429 if (sector_nr == 0) {
4430 /* If restarting in the middle, skip the initial sectors */
4431 if (mddev->reshape_backwards &&
4432 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4433 sector_nr = (raid10_size(mddev, 0, 0)
4434 - conf->reshape_progress);
4435 } else if (!mddev->reshape_backwards &&
4436 conf->reshape_progress > 0)
4437 sector_nr = conf->reshape_progress;
4438 if (sector_nr) {
4439 mddev->curr_resync_completed = sector_nr;
4440 sysfs_notify_dirent_safe(mddev->sysfs_completed);
4441 *skipped = 1;
4442 return sector_nr;
4443 }
4444 }
4445
4446 /* We don't use sector_nr to track where we are up to
4447 * as that doesn't work well for ->reshape_backwards.
4448 * So just use ->reshape_progress.
4449 */
4450 if (mddev->reshape_backwards) {
4451 /* 'next' is the earliest device address that we might
4452 * write to for this chunk in the new layout
4453 */
4454 next = first_dev_address(conf->reshape_progress - 1,
4455 &conf->geo);
4456
4457 /* 'safe' is the last device address that we might read from
4458 * in the old layout after a restart
4459 */
4460 safe = last_dev_address(conf->reshape_safe - 1,
4461 &conf->prev);
4462
4463 if (next + conf->offset_diff < safe)
4464 need_flush = 1;
4465
4466 last = conf->reshape_progress - 1;
4467 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4468 & conf->prev.chunk_mask);
4469 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4470 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4471 } else {
4472 /* 'next' is after the last device address that we
4473 * might write to for this chunk in the new layout
4474 */
4475 next = last_dev_address(conf->reshape_progress, &conf->geo);
4476
4477 /* 'safe' is the earliest device address that we might
4478 * read from in the old layout after a restart
4479 */
4480 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4481
4482 /* Need to update metadata if 'next' might be beyond 'safe'
4483 * as that would possibly corrupt data
4484 */
4485 if (next > safe + conf->offset_diff)
4486 need_flush = 1;
4487
4488 sector_nr = conf->reshape_progress;
4489 last = sector_nr | (conf->geo.chunk_mask
4490 & conf->prev.chunk_mask);
4491
4492 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4493 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4494 }
4495
4496 if (need_flush ||
4497 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4498 /* Need to update reshape_position in metadata */
4499 wait_barrier(conf);
4500 mddev->reshape_position = conf->reshape_progress;
4501 if (mddev->reshape_backwards)
4502 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4503 - conf->reshape_progress;
4504 else
4505 mddev->curr_resync_completed = conf->reshape_progress;
4506 conf->reshape_checkpoint = jiffies;
4507 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4508 md_wakeup_thread(mddev->thread);
4509 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4510 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4511 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4512 allow_barrier(conf);
4513 return sectors_done;
4514 }
4515 conf->reshape_safe = mddev->reshape_position;
4516 allow_barrier(conf);
4517 }
4518
4519 raise_barrier(conf, 0);
4520read_more:
4521 /* Now schedule reads for blocks from sector_nr to last */
4522 r10_bio = raid10_alloc_init_r10buf(conf);
4523 r10_bio->state = 0;
4524 raise_barrier(conf, 1);
4525 atomic_set(&r10_bio->remaining, 0);
4526 r10_bio->mddev = mddev;
4527 r10_bio->sector = sector_nr;
4528 set_bit(R10BIO_IsReshape, &r10_bio->state);
4529 r10_bio->sectors = last - sector_nr + 1;
4530 rdev = read_balance(conf, r10_bio, &max_sectors);
4531 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4532
4533 if (!rdev) {
4534 /* Cannot read from here, so need to record bad blocks
4535 * on all the target devices.
4536 */
4537 // FIXME
4538 mempool_free(r10_bio, &conf->r10buf_pool);
4539 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4540 return sectors_done;
4541 }
4542
4543 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4544
4545 bio_set_dev(read_bio, rdev->bdev);
4546 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4547 + rdev->data_offset);
4548 read_bio->bi_private = r10_bio;
4549 read_bio->bi_end_io = end_reshape_read;
4550 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4551 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4552 read_bio->bi_status = 0;
4553 read_bio->bi_vcnt = 0;
4554 read_bio->bi_iter.bi_size = 0;
4555 r10_bio->master_bio = read_bio;
4556 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4557
4558 /*
4559 * Broadcast RESYNC message to other nodes, so all nodes would not
4560 * write to the region to avoid conflict.
4561 */
4562 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4563 struct mdp_superblock_1 *sb = NULL;
4564 int sb_reshape_pos = 0;
4565
4566 conf->cluster_sync_low = sector_nr;
4567 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4568 sb = page_address(rdev->sb_page);
4569 if (sb) {
4570 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4571 /*
4572 * Set cluster_sync_low again if next address for array
4573 * reshape is less than cluster_sync_low. Since we can't
4574 * update cluster_sync_low until it has finished reshape.
4575 */
4576 if (sb_reshape_pos < conf->cluster_sync_low)
4577 conf->cluster_sync_low = sb_reshape_pos;
4578 }
4579
4580 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4581 conf->cluster_sync_high);
4582 }
4583
4584 /* Now find the locations in the new layout */
4585 __raid10_find_phys(&conf->geo, r10_bio);
4586
4587 blist = read_bio;
4588 read_bio->bi_next = NULL;
4589
4590 rcu_read_lock();
4591 for (s = 0; s < conf->copies*2; s++) {
4592 struct bio *b;
4593 int d = r10_bio->devs[s/2].devnum;
4594 struct md_rdev *rdev2;
4595 if (s&1) {
4596 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4597 b = r10_bio->devs[s/2].repl_bio;
4598 } else {
4599 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4600 b = r10_bio->devs[s/2].bio;
4601 }
4602 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4603 continue;
4604
4605 bio_set_dev(b, rdev2->bdev);
4606 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4607 rdev2->new_data_offset;
4608 b->bi_end_io = end_reshape_write;
4609 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4610 b->bi_next = blist;
4611 blist = b;
4612 }
4613
4614 /* Now add as many pages as possible to all of these bios. */
4615
4616 nr_sectors = 0;
4617 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4618 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4619 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4620 int len = (max_sectors - s) << 9;
4621 if (len > PAGE_SIZE)
4622 len = PAGE_SIZE;
4623 for (bio = blist; bio ; bio = bio->bi_next) {
4624 /*
4625 * won't fail because the vec table is big enough
4626 * to hold all these pages
4627 */
4628 bio_add_page(bio, page, len, 0);
4629 }
4630 sector_nr += len >> 9;
4631 nr_sectors += len >> 9;
4632 }
4633 rcu_read_unlock();
4634 r10_bio->sectors = nr_sectors;
4635
4636 /* Now submit the read */
4637 md_sync_acct_bio(read_bio, r10_bio->sectors);
4638 atomic_inc(&r10_bio->remaining);
4639 read_bio->bi_next = NULL;
4640 submit_bio_noacct(read_bio);
4641 sectors_done += nr_sectors;
4642 if (sector_nr <= last)
4643 goto read_more;
4644
4645 lower_barrier(conf);
4646
4647 /* Now that we have done the whole section we can
4648 * update reshape_progress
4649 */
4650 if (mddev->reshape_backwards)
4651 conf->reshape_progress -= sectors_done;
4652 else
4653 conf->reshape_progress += sectors_done;
4654
4655 return sectors_done;
4656}
4657
4658static void end_reshape_request(struct r10bio *r10_bio);
4659static int handle_reshape_read_error(struct mddev *mddev,
4660 struct r10bio *r10_bio);
4661static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4662{
4663 /* Reshape read completed. Hopefully we have a block
4664 * to write out.
4665 * If we got a read error then we do sync 1-page reads from
4666 * elsewhere until we find the data - or give up.
4667 */
4668 struct r10conf *conf = mddev->private;
4669 int s;
4670
4671 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4672 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4673 /* Reshape has been aborted */
4674 md_done_sync(mddev, r10_bio->sectors, 0);
4675 return;
4676 }
4677
4678 /* We definitely have the data in the pages, schedule the
4679 * writes.
4680 */
4681 atomic_set(&r10_bio->remaining, 1);
4682 for (s = 0; s < conf->copies*2; s++) {
4683 struct bio *b;
4684 int d = r10_bio->devs[s/2].devnum;
4685 struct md_rdev *rdev;
4686 rcu_read_lock();
4687 if (s&1) {
4688 rdev = rcu_dereference(conf->mirrors[d].replacement);
4689 b = r10_bio->devs[s/2].repl_bio;
4690 } else {
4691 rdev = rcu_dereference(conf->mirrors[d].rdev);
4692 b = r10_bio->devs[s/2].bio;
4693 }
4694 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4695 rcu_read_unlock();
4696 continue;
4697 }
4698 atomic_inc(&rdev->nr_pending);
4699 rcu_read_unlock();
4700 md_sync_acct_bio(b, r10_bio->sectors);
4701 atomic_inc(&r10_bio->remaining);
4702 b->bi_next = NULL;
4703 submit_bio_noacct(b);
4704 }
4705 end_reshape_request(r10_bio);
4706}
4707
4708static void end_reshape(struct r10conf *conf)
4709{
4710 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4711 return;
4712
4713 spin_lock_irq(&conf->device_lock);
4714 conf->prev = conf->geo;
4715 md_finish_reshape(conf->mddev);
4716 smp_wmb();
4717 conf->reshape_progress = MaxSector;
4718 conf->reshape_safe = MaxSector;
4719 spin_unlock_irq(&conf->device_lock);
4720
4721 /* read-ahead size must cover two whole stripes, which is
4722 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4723 */
4724 if (conf->mddev->queue) {
4725 int stripe = conf->geo.raid_disks *
4726 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4727 stripe /= conf->geo.near_copies;
4728 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4729 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4730 }
4731 conf->fullsync = 0;
4732}
4733
4734static void raid10_update_reshape_pos(struct mddev *mddev)
4735{
4736 struct r10conf *conf = mddev->private;
4737 sector_t lo, hi;
4738
4739 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
4740 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4741 || mddev->reshape_position == MaxSector)
4742 conf->reshape_progress = mddev->reshape_position;
4743 else
4744 WARN_ON_ONCE(1);
4745}
4746
4747static int handle_reshape_read_error(struct mddev *mddev,
4748 struct r10bio *r10_bio)
4749{
4750 /* Use sync reads to get the blocks from somewhere else */
4751 int sectors = r10_bio->sectors;
4752 struct r10conf *conf = mddev->private;
4753 struct r10bio *r10b;
4754 int slot = 0;
4755 int idx = 0;
4756 struct page **pages;
4757
4758 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
4759 if (!r10b) {
4760 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4761 return -ENOMEM;
4762 }
4763
4764 /* reshape IOs share pages from .devs[0].bio */
4765 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4766
4767 r10b->sector = r10_bio->sector;
4768 __raid10_find_phys(&conf->prev, r10b);
4769
4770 while (sectors) {
4771 int s = sectors;
4772 int success = 0;
4773 int first_slot = slot;
4774
4775 if (s > (PAGE_SIZE >> 9))
4776 s = PAGE_SIZE >> 9;
4777
4778 rcu_read_lock();
4779 while (!success) {
4780 int d = r10b->devs[slot].devnum;
4781 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4782 sector_t addr;
4783 if (rdev == NULL ||
4784 test_bit(Faulty, &rdev->flags) ||
4785 !test_bit(In_sync, &rdev->flags))
4786 goto failed;
4787
4788 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4789 atomic_inc(&rdev->nr_pending);
4790 rcu_read_unlock();
4791 success = sync_page_io(rdev,
4792 addr,
4793 s << 9,
4794 pages[idx],
4795 REQ_OP_READ, 0, false);
4796 rdev_dec_pending(rdev, mddev);
4797 rcu_read_lock();
4798 if (success)
4799 break;
4800 failed:
4801 slot++;
4802 if (slot >= conf->copies)
4803 slot = 0;
4804 if (slot == first_slot)
4805 break;
4806 }
4807 rcu_read_unlock();
4808 if (!success) {
4809 /* couldn't read this block, must give up */
4810 set_bit(MD_RECOVERY_INTR,
4811 &mddev->recovery);
4812 kfree(r10b);
4813 return -EIO;
4814 }
4815 sectors -= s;
4816 idx++;
4817 }
4818 kfree(r10b);
4819 return 0;
4820}
4821
4822static void end_reshape_write(struct bio *bio)
4823{
4824 struct r10bio *r10_bio = get_resync_r10bio(bio);
4825 struct mddev *mddev = r10_bio->mddev;
4826 struct r10conf *conf = mddev->private;
4827 int d;
4828 int slot;
4829 int repl;
4830 struct md_rdev *rdev = NULL;
4831
4832 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4833 if (repl)
4834 rdev = conf->mirrors[d].replacement;
4835 if (!rdev) {
4836 smp_mb();
4837 rdev = conf->mirrors[d].rdev;
4838 }
4839
4840 if (bio->bi_status) {
4841 /* FIXME should record badblock */
4842 md_error(mddev, rdev);
4843 }
4844
4845 rdev_dec_pending(rdev, mddev);
4846 end_reshape_request(r10_bio);
4847}
4848
4849static void end_reshape_request(struct r10bio *r10_bio)
4850{
4851 if (!atomic_dec_and_test(&r10_bio->remaining))
4852 return;
4853 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4854 bio_put(r10_bio->master_bio);
4855 put_buf(r10_bio);
4856}
4857
4858static void raid10_finish_reshape(struct mddev *mddev)
4859{
4860 struct r10conf *conf = mddev->private;
4861
4862 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4863 return;
4864
4865 if (mddev->delta_disks > 0) {
4866 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4867 mddev->recovery_cp = mddev->resync_max_sectors;
4868 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4869 }
4870 mddev->resync_max_sectors = mddev->array_sectors;
4871 } else {
4872 int d;
4873 rcu_read_lock();
4874 for (d = conf->geo.raid_disks ;
4875 d < conf->geo.raid_disks - mddev->delta_disks;
4876 d++) {
4877 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4878 if (rdev)
4879 clear_bit(In_sync, &rdev->flags);
4880 rdev = rcu_dereference(conf->mirrors[d].replacement);
4881 if (rdev)
4882 clear_bit(In_sync, &rdev->flags);
4883 }
4884 rcu_read_unlock();
4885 }
4886 mddev->layout = mddev->new_layout;
4887 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4888 mddev->reshape_position = MaxSector;
4889 mddev->delta_disks = 0;
4890 mddev->reshape_backwards = 0;
4891}
4892
4893static struct md_personality raid10_personality =
4894{
4895 .name = "raid10",
4896 .level = 10,
4897 .owner = THIS_MODULE,
4898 .make_request = raid10_make_request,
4899 .run = raid10_run,
4900 .free = raid10_free,
4901 .status = raid10_status,
4902 .error_handler = raid10_error,
4903 .hot_add_disk = raid10_add_disk,
4904 .hot_remove_disk= raid10_remove_disk,
4905 .spare_active = raid10_spare_active,
4906 .sync_request = raid10_sync_request,
4907 .quiesce = raid10_quiesce,
4908 .size = raid10_size,
4909 .resize = raid10_resize,
4910 .takeover = raid10_takeover,
4911 .check_reshape = raid10_check_reshape,
4912 .start_reshape = raid10_start_reshape,
4913 .finish_reshape = raid10_finish_reshape,
4914 .update_reshape_pos = raid10_update_reshape_pos,
4915};
4916
4917static int __init raid_init(void)
4918{
4919 return register_md_personality(&raid10_personality);
4920}
4921
4922static void raid_exit(void)
4923{
4924 unregister_md_personality(&raid10_personality);
4925}
4926
4927module_init(raid_init);
4928module_exit(raid_exit);
4929MODULE_LICENSE("GPL");
4930MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4931MODULE_ALIAS("md-personality-9"); /* RAID10 */
4932MODULE_ALIAS("md-raid10");
4933MODULE_ALIAS("md-level-10");
4934
4935module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);