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/seq_file.h>
25#include <linux/ratelimit.h>
26#include "md.h"
27#include "raid10.h"
28#include "raid0.h"
29#include "bitmap.h"
30
31/*
32 * RAID10 provides a combination of RAID0 and RAID1 functionality.
33 * The layout of data is defined by
34 * chunk_size
35 * raid_disks
36 * near_copies (stored in low byte of layout)
37 * far_copies (stored in second byte of layout)
38 * far_offset (stored in bit 16 of layout )
39 *
40 * The data to be stored is divided into chunks using chunksize.
41 * Each device is divided into far_copies sections.
42 * In each section, chunks are laid out in a style similar to raid0, but
43 * near_copies copies of each chunk is stored (each on a different drive).
44 * The starting device for each section is offset near_copies from the starting
45 * device of the previous section.
46 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
47 * drive.
48 * near_copies and far_copies must be at least one, and their product is at most
49 * raid_disks.
50 *
51 * If far_offset is true, then the far_copies are handled a bit differently.
52 * The copies are still in different stripes, but instead of be very far apart
53 * on disk, there are adjacent stripes.
54 */
55
56/*
57 * Number of guaranteed r10bios in case of extreme VM load:
58 */
59#define NR_RAID10_BIOS 256
60
61static void allow_barrier(conf_t *conf);
62static void lower_barrier(conf_t *conf);
63
64static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
65{
66 conf_t *conf = data;
67 int size = offsetof(struct r10bio_s, devs[conf->copies]);
68
69 /* allocate a r10bio with room for raid_disks entries in the bios array */
70 return kzalloc(size, gfp_flags);
71}
72
73static void r10bio_pool_free(void *r10_bio, void *data)
74{
75 kfree(r10_bio);
76}
77
78/* Maximum size of each resync request */
79#define RESYNC_BLOCK_SIZE (64*1024)
80#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
81/* amount of memory to reserve for resync requests */
82#define RESYNC_WINDOW (1024*1024)
83/* maximum number of concurrent requests, memory permitting */
84#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
85
86/*
87 * When performing a resync, we need to read and compare, so
88 * we need as many pages are there are copies.
89 * When performing a recovery, we need 2 bios, one for read,
90 * one for write (we recover only one drive per r10buf)
91 *
92 */
93static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
94{
95 conf_t *conf = data;
96 struct page *page;
97 r10bio_t *r10_bio;
98 struct bio *bio;
99 int i, j;
100 int nalloc;
101
102 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
103 if (!r10_bio)
104 return NULL;
105
106 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
107 nalloc = conf->copies; /* resync */
108 else
109 nalloc = 2; /* recovery */
110
111 /*
112 * Allocate bios.
113 */
114 for (j = nalloc ; j-- ; ) {
115 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
116 if (!bio)
117 goto out_free_bio;
118 r10_bio->devs[j].bio = bio;
119 }
120 /*
121 * Allocate RESYNC_PAGES data pages and attach them
122 * where needed.
123 */
124 for (j = 0 ; j < nalloc; j++) {
125 bio = r10_bio->devs[j].bio;
126 for (i = 0; i < RESYNC_PAGES; i++) {
127 if (j == 1 && !test_bit(MD_RECOVERY_SYNC,
128 &conf->mddev->recovery)) {
129 /* we can share bv_page's during recovery */
130 struct bio *rbio = r10_bio->devs[0].bio;
131 page = rbio->bi_io_vec[i].bv_page;
132 get_page(page);
133 } else
134 page = alloc_page(gfp_flags);
135 if (unlikely(!page))
136 goto out_free_pages;
137
138 bio->bi_io_vec[i].bv_page = page;
139 }
140 }
141
142 return r10_bio;
143
144out_free_pages:
145 for ( ; i > 0 ; i--)
146 safe_put_page(bio->bi_io_vec[i-1].bv_page);
147 while (j--)
148 for (i = 0; i < RESYNC_PAGES ; i++)
149 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
150 j = -1;
151out_free_bio:
152 while ( ++j < nalloc )
153 bio_put(r10_bio->devs[j].bio);
154 r10bio_pool_free(r10_bio, conf);
155 return NULL;
156}
157
158static void r10buf_pool_free(void *__r10_bio, void *data)
159{
160 int i;
161 conf_t *conf = data;
162 r10bio_t *r10bio = __r10_bio;
163 int j;
164
165 for (j=0; j < conf->copies; j++) {
166 struct bio *bio = r10bio->devs[j].bio;
167 if (bio) {
168 for (i = 0; i < RESYNC_PAGES; i++) {
169 safe_put_page(bio->bi_io_vec[i].bv_page);
170 bio->bi_io_vec[i].bv_page = NULL;
171 }
172 bio_put(bio);
173 }
174 }
175 r10bio_pool_free(r10bio, conf);
176}
177
178static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
179{
180 int i;
181
182 for (i = 0; i < conf->copies; i++) {
183 struct bio **bio = & r10_bio->devs[i].bio;
184 if (!BIO_SPECIAL(*bio))
185 bio_put(*bio);
186 *bio = NULL;
187 }
188}
189
190static void free_r10bio(r10bio_t *r10_bio)
191{
192 conf_t *conf = r10_bio->mddev->private;
193
194 put_all_bios(conf, r10_bio);
195 mempool_free(r10_bio, conf->r10bio_pool);
196}
197
198static void put_buf(r10bio_t *r10_bio)
199{
200 conf_t *conf = r10_bio->mddev->private;
201
202 mempool_free(r10_bio, conf->r10buf_pool);
203
204 lower_barrier(conf);
205}
206
207static void reschedule_retry(r10bio_t *r10_bio)
208{
209 unsigned long flags;
210 mddev_t *mddev = r10_bio->mddev;
211 conf_t *conf = mddev->private;
212
213 spin_lock_irqsave(&conf->device_lock, flags);
214 list_add(&r10_bio->retry_list, &conf->retry_list);
215 conf->nr_queued ++;
216 spin_unlock_irqrestore(&conf->device_lock, flags);
217
218 /* wake up frozen array... */
219 wake_up(&conf->wait_barrier);
220
221 md_wakeup_thread(mddev->thread);
222}
223
224/*
225 * raid_end_bio_io() is called when we have finished servicing a mirrored
226 * operation and are ready to return a success/failure code to the buffer
227 * cache layer.
228 */
229static void raid_end_bio_io(r10bio_t *r10_bio)
230{
231 struct bio *bio = r10_bio->master_bio;
232 int done;
233 conf_t *conf = r10_bio->mddev->private;
234
235 if (bio->bi_phys_segments) {
236 unsigned long flags;
237 spin_lock_irqsave(&conf->device_lock, flags);
238 bio->bi_phys_segments--;
239 done = (bio->bi_phys_segments == 0);
240 spin_unlock_irqrestore(&conf->device_lock, flags);
241 } else
242 done = 1;
243 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
244 clear_bit(BIO_UPTODATE, &bio->bi_flags);
245 if (done) {
246 bio_endio(bio, 0);
247 /*
248 * Wake up any possible resync thread that waits for the device
249 * to go idle.
250 */
251 allow_barrier(conf);
252 }
253 free_r10bio(r10_bio);
254}
255
256/*
257 * Update disk head position estimator based on IRQ completion info.
258 */
259static inline void update_head_pos(int slot, r10bio_t *r10_bio)
260{
261 conf_t *conf = r10_bio->mddev->private;
262
263 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
264 r10_bio->devs[slot].addr + (r10_bio->sectors);
265}
266
267/*
268 * Find the disk number which triggered given bio
269 */
270static int find_bio_disk(conf_t *conf, r10bio_t *r10_bio,
271 struct bio *bio, int *slotp)
272{
273 int slot;
274
275 for (slot = 0; slot < conf->copies; slot++)
276 if (r10_bio->devs[slot].bio == bio)
277 break;
278
279 BUG_ON(slot == conf->copies);
280 update_head_pos(slot, r10_bio);
281
282 if (slotp)
283 *slotp = slot;
284 return r10_bio->devs[slot].devnum;
285}
286
287static void raid10_end_read_request(struct bio *bio, int error)
288{
289 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
290 r10bio_t *r10_bio = bio->bi_private;
291 int slot, dev;
292 conf_t *conf = r10_bio->mddev->private;
293
294
295 slot = r10_bio->read_slot;
296 dev = r10_bio->devs[slot].devnum;
297 /*
298 * this branch is our 'one mirror IO has finished' event handler:
299 */
300 update_head_pos(slot, r10_bio);
301
302 if (uptodate) {
303 /*
304 * Set R10BIO_Uptodate in our master bio, so that
305 * we will return a good error code to the higher
306 * levels even if IO on some other mirrored buffer fails.
307 *
308 * The 'master' represents the composite IO operation to
309 * user-side. So if something waits for IO, then it will
310 * wait for the 'master' bio.
311 */
312 set_bit(R10BIO_Uptodate, &r10_bio->state);
313 raid_end_bio_io(r10_bio);
314 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
315 } else {
316 /*
317 * oops, read error - keep the refcount on the rdev
318 */
319 char b[BDEVNAME_SIZE];
320 printk_ratelimited(KERN_ERR
321 "md/raid10:%s: %s: rescheduling sector %llu\n",
322 mdname(conf->mddev),
323 bdevname(conf->mirrors[dev].rdev->bdev, b),
324 (unsigned long long)r10_bio->sector);
325 set_bit(R10BIO_ReadError, &r10_bio->state);
326 reschedule_retry(r10_bio);
327 }
328}
329
330static void close_write(r10bio_t *r10_bio)
331{
332 /* clear the bitmap if all writes complete successfully */
333 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
334 r10_bio->sectors,
335 !test_bit(R10BIO_Degraded, &r10_bio->state),
336 0);
337 md_write_end(r10_bio->mddev);
338}
339
340static void one_write_done(r10bio_t *r10_bio)
341{
342 if (atomic_dec_and_test(&r10_bio->remaining)) {
343 if (test_bit(R10BIO_WriteError, &r10_bio->state))
344 reschedule_retry(r10_bio);
345 else {
346 close_write(r10_bio);
347 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
348 reschedule_retry(r10_bio);
349 else
350 raid_end_bio_io(r10_bio);
351 }
352 }
353}
354
355static void raid10_end_write_request(struct bio *bio, int error)
356{
357 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
358 r10bio_t *r10_bio = bio->bi_private;
359 int dev;
360 int dec_rdev = 1;
361 conf_t *conf = r10_bio->mddev->private;
362 int slot;
363
364 dev = find_bio_disk(conf, r10_bio, bio, &slot);
365
366 /*
367 * this branch is our 'one mirror IO has finished' event handler:
368 */
369 if (!uptodate) {
370 set_bit(WriteErrorSeen, &conf->mirrors[dev].rdev->flags);
371 set_bit(R10BIO_WriteError, &r10_bio->state);
372 dec_rdev = 0;
373 } else {
374 /*
375 * Set R10BIO_Uptodate in our master bio, so that
376 * we will return a good error code for to the higher
377 * levels even if IO on some other mirrored buffer fails.
378 *
379 * The 'master' represents the composite IO operation to
380 * user-side. So if something waits for IO, then it will
381 * wait for the 'master' bio.
382 */
383 sector_t first_bad;
384 int bad_sectors;
385
386 set_bit(R10BIO_Uptodate, &r10_bio->state);
387
388 /* Maybe we can clear some bad blocks. */
389 if (is_badblock(conf->mirrors[dev].rdev,
390 r10_bio->devs[slot].addr,
391 r10_bio->sectors,
392 &first_bad, &bad_sectors)) {
393 bio_put(bio);
394 r10_bio->devs[slot].bio = IO_MADE_GOOD;
395 dec_rdev = 0;
396 set_bit(R10BIO_MadeGood, &r10_bio->state);
397 }
398 }
399
400 /*
401 *
402 * Let's see if all mirrored write operations have finished
403 * already.
404 */
405 one_write_done(r10_bio);
406 if (dec_rdev)
407 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
408}
409
410
411/*
412 * RAID10 layout manager
413 * As well as the chunksize and raid_disks count, there are two
414 * parameters: near_copies and far_copies.
415 * near_copies * far_copies must be <= raid_disks.
416 * Normally one of these will be 1.
417 * If both are 1, we get raid0.
418 * If near_copies == raid_disks, we get raid1.
419 *
420 * Chunks are laid out in raid0 style with near_copies copies of the
421 * first chunk, followed by near_copies copies of the next chunk and
422 * so on.
423 * If far_copies > 1, then after 1/far_copies of the array has been assigned
424 * as described above, we start again with a device offset of near_copies.
425 * So we effectively have another copy of the whole array further down all
426 * the drives, but with blocks on different drives.
427 * With this layout, and block is never stored twice on the one device.
428 *
429 * raid10_find_phys finds the sector offset of a given virtual sector
430 * on each device that it is on.
431 *
432 * raid10_find_virt does the reverse mapping, from a device and a
433 * sector offset to a virtual address
434 */
435
436static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
437{
438 int n,f;
439 sector_t sector;
440 sector_t chunk;
441 sector_t stripe;
442 int dev;
443
444 int slot = 0;
445
446 /* now calculate first sector/dev */
447 chunk = r10bio->sector >> conf->chunk_shift;
448 sector = r10bio->sector & conf->chunk_mask;
449
450 chunk *= conf->near_copies;
451 stripe = chunk;
452 dev = sector_div(stripe, conf->raid_disks);
453 if (conf->far_offset)
454 stripe *= conf->far_copies;
455
456 sector += stripe << conf->chunk_shift;
457
458 /* and calculate all the others */
459 for (n=0; n < conf->near_copies; n++) {
460 int d = dev;
461 sector_t s = sector;
462 r10bio->devs[slot].addr = sector;
463 r10bio->devs[slot].devnum = d;
464 slot++;
465
466 for (f = 1; f < conf->far_copies; f++) {
467 d += conf->near_copies;
468 if (d >= conf->raid_disks)
469 d -= conf->raid_disks;
470 s += conf->stride;
471 r10bio->devs[slot].devnum = d;
472 r10bio->devs[slot].addr = s;
473 slot++;
474 }
475 dev++;
476 if (dev >= conf->raid_disks) {
477 dev = 0;
478 sector += (conf->chunk_mask + 1);
479 }
480 }
481 BUG_ON(slot != conf->copies);
482}
483
484static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
485{
486 sector_t offset, chunk, vchunk;
487
488 offset = sector & conf->chunk_mask;
489 if (conf->far_offset) {
490 int fc;
491 chunk = sector >> conf->chunk_shift;
492 fc = sector_div(chunk, conf->far_copies);
493 dev -= fc * conf->near_copies;
494 if (dev < 0)
495 dev += conf->raid_disks;
496 } else {
497 while (sector >= conf->stride) {
498 sector -= conf->stride;
499 if (dev < conf->near_copies)
500 dev += conf->raid_disks - conf->near_copies;
501 else
502 dev -= conf->near_copies;
503 }
504 chunk = sector >> conf->chunk_shift;
505 }
506 vchunk = chunk * conf->raid_disks + dev;
507 sector_div(vchunk, conf->near_copies);
508 return (vchunk << conf->chunk_shift) + offset;
509}
510
511/**
512 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
513 * @q: request queue
514 * @bvm: properties of new bio
515 * @biovec: the request that could be merged to it.
516 *
517 * Return amount of bytes we can accept at this offset
518 * If near_copies == raid_disk, there are no striping issues,
519 * but in that case, the function isn't called at all.
520 */
521static int raid10_mergeable_bvec(struct request_queue *q,
522 struct bvec_merge_data *bvm,
523 struct bio_vec *biovec)
524{
525 mddev_t *mddev = q->queuedata;
526 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
527 int max;
528 unsigned int chunk_sectors = mddev->chunk_sectors;
529 unsigned int bio_sectors = bvm->bi_size >> 9;
530
531 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
532 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
533 if (max <= biovec->bv_len && bio_sectors == 0)
534 return biovec->bv_len;
535 else
536 return max;
537}
538
539/*
540 * This routine returns the disk from which the requested read should
541 * be done. There is a per-array 'next expected sequential IO' sector
542 * number - if this matches on the next IO then we use the last disk.
543 * There is also a per-disk 'last know head position' sector that is
544 * maintained from IRQ contexts, both the normal and the resync IO
545 * completion handlers update this position correctly. If there is no
546 * perfect sequential match then we pick the disk whose head is closest.
547 *
548 * If there are 2 mirrors in the same 2 devices, performance degrades
549 * because position is mirror, not device based.
550 *
551 * The rdev for the device selected will have nr_pending incremented.
552 */
553
554/*
555 * FIXME: possibly should rethink readbalancing and do it differently
556 * depending on near_copies / far_copies geometry.
557 */
558static int read_balance(conf_t *conf, r10bio_t *r10_bio, int *max_sectors)
559{
560 const sector_t this_sector = r10_bio->sector;
561 int disk, slot;
562 int sectors = r10_bio->sectors;
563 int best_good_sectors;
564 sector_t new_distance, best_dist;
565 mdk_rdev_t *rdev;
566 int do_balance;
567 int best_slot;
568
569 raid10_find_phys(conf, r10_bio);
570 rcu_read_lock();
571retry:
572 sectors = r10_bio->sectors;
573 best_slot = -1;
574 best_dist = MaxSector;
575 best_good_sectors = 0;
576 do_balance = 1;
577 /*
578 * Check if we can balance. We can balance on the whole
579 * device if no resync is going on (recovery is ok), or below
580 * the resync window. We take the first readable disk when
581 * above the resync window.
582 */
583 if (conf->mddev->recovery_cp < MaxSector
584 && (this_sector + sectors >= conf->next_resync))
585 do_balance = 0;
586
587 for (slot = 0; slot < conf->copies ; slot++) {
588 sector_t first_bad;
589 int bad_sectors;
590 sector_t dev_sector;
591
592 if (r10_bio->devs[slot].bio == IO_BLOCKED)
593 continue;
594 disk = r10_bio->devs[slot].devnum;
595 rdev = rcu_dereference(conf->mirrors[disk].rdev);
596 if (rdev == NULL)
597 continue;
598 if (!test_bit(In_sync, &rdev->flags))
599 continue;
600
601 dev_sector = r10_bio->devs[slot].addr;
602 if (is_badblock(rdev, dev_sector, sectors,
603 &first_bad, &bad_sectors)) {
604 if (best_dist < MaxSector)
605 /* Already have a better slot */
606 continue;
607 if (first_bad <= dev_sector) {
608 /* Cannot read here. If this is the
609 * 'primary' device, then we must not read
610 * beyond 'bad_sectors' from another device.
611 */
612 bad_sectors -= (dev_sector - first_bad);
613 if (!do_balance && sectors > bad_sectors)
614 sectors = bad_sectors;
615 if (best_good_sectors > sectors)
616 best_good_sectors = sectors;
617 } else {
618 sector_t good_sectors =
619 first_bad - dev_sector;
620 if (good_sectors > best_good_sectors) {
621 best_good_sectors = good_sectors;
622 best_slot = slot;
623 }
624 if (!do_balance)
625 /* Must read from here */
626 break;
627 }
628 continue;
629 } else
630 best_good_sectors = sectors;
631
632 if (!do_balance)
633 break;
634
635 /* This optimisation is debatable, and completely destroys
636 * sequential read speed for 'far copies' arrays. So only
637 * keep it for 'near' arrays, and review those later.
638 */
639 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending))
640 break;
641
642 /* for far > 1 always use the lowest address */
643 if (conf->far_copies > 1)
644 new_distance = r10_bio->devs[slot].addr;
645 else
646 new_distance = abs(r10_bio->devs[slot].addr -
647 conf->mirrors[disk].head_position);
648 if (new_distance < best_dist) {
649 best_dist = new_distance;
650 best_slot = slot;
651 }
652 }
653 if (slot == conf->copies)
654 slot = best_slot;
655
656 if (slot >= 0) {
657 disk = r10_bio->devs[slot].devnum;
658 rdev = rcu_dereference(conf->mirrors[disk].rdev);
659 if (!rdev)
660 goto retry;
661 atomic_inc(&rdev->nr_pending);
662 if (test_bit(Faulty, &rdev->flags)) {
663 /* Cannot risk returning a device that failed
664 * before we inc'ed nr_pending
665 */
666 rdev_dec_pending(rdev, conf->mddev);
667 goto retry;
668 }
669 r10_bio->read_slot = slot;
670 } else
671 disk = -1;
672 rcu_read_unlock();
673 *max_sectors = best_good_sectors;
674
675 return disk;
676}
677
678static int raid10_congested(void *data, int bits)
679{
680 mddev_t *mddev = data;
681 conf_t *conf = mddev->private;
682 int i, ret = 0;
683
684 if (mddev_congested(mddev, bits))
685 return 1;
686 rcu_read_lock();
687 for (i = 0; i < conf->raid_disks && ret == 0; i++) {
688 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
689 if (rdev && !test_bit(Faulty, &rdev->flags)) {
690 struct request_queue *q = bdev_get_queue(rdev->bdev);
691
692 ret |= bdi_congested(&q->backing_dev_info, bits);
693 }
694 }
695 rcu_read_unlock();
696 return ret;
697}
698
699static void flush_pending_writes(conf_t *conf)
700{
701 /* Any writes that have been queued but are awaiting
702 * bitmap updates get flushed here.
703 */
704 spin_lock_irq(&conf->device_lock);
705
706 if (conf->pending_bio_list.head) {
707 struct bio *bio;
708 bio = bio_list_get(&conf->pending_bio_list);
709 spin_unlock_irq(&conf->device_lock);
710 /* flush any pending bitmap writes to disk
711 * before proceeding w/ I/O */
712 bitmap_unplug(conf->mddev->bitmap);
713
714 while (bio) { /* submit pending writes */
715 struct bio *next = bio->bi_next;
716 bio->bi_next = NULL;
717 generic_make_request(bio);
718 bio = next;
719 }
720 } else
721 spin_unlock_irq(&conf->device_lock);
722}
723
724/* Barriers....
725 * Sometimes we need to suspend IO while we do something else,
726 * either some resync/recovery, or reconfigure the array.
727 * To do this we raise a 'barrier'.
728 * The 'barrier' is a counter that can be raised multiple times
729 * to count how many activities are happening which preclude
730 * normal IO.
731 * We can only raise the barrier if there is no pending IO.
732 * i.e. if nr_pending == 0.
733 * We choose only to raise the barrier if no-one is waiting for the
734 * barrier to go down. This means that as soon as an IO request
735 * is ready, no other operations which require a barrier will start
736 * until the IO request has had a chance.
737 *
738 * So: regular IO calls 'wait_barrier'. When that returns there
739 * is no backgroup IO happening, It must arrange to call
740 * allow_barrier when it has finished its IO.
741 * backgroup IO calls must call raise_barrier. Once that returns
742 * there is no normal IO happeing. It must arrange to call
743 * lower_barrier when the particular background IO completes.
744 */
745
746static void raise_barrier(conf_t *conf, int force)
747{
748 BUG_ON(force && !conf->barrier);
749 spin_lock_irq(&conf->resync_lock);
750
751 /* Wait until no block IO is waiting (unless 'force') */
752 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
753 conf->resync_lock, );
754
755 /* block any new IO from starting */
756 conf->barrier++;
757
758 /* Now wait for all pending IO to complete */
759 wait_event_lock_irq(conf->wait_barrier,
760 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
761 conf->resync_lock, );
762
763 spin_unlock_irq(&conf->resync_lock);
764}
765
766static void lower_barrier(conf_t *conf)
767{
768 unsigned long flags;
769 spin_lock_irqsave(&conf->resync_lock, flags);
770 conf->barrier--;
771 spin_unlock_irqrestore(&conf->resync_lock, flags);
772 wake_up(&conf->wait_barrier);
773}
774
775static void wait_barrier(conf_t *conf)
776{
777 spin_lock_irq(&conf->resync_lock);
778 if (conf->barrier) {
779 conf->nr_waiting++;
780 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
781 conf->resync_lock,
782 );
783 conf->nr_waiting--;
784 }
785 conf->nr_pending++;
786 spin_unlock_irq(&conf->resync_lock);
787}
788
789static void allow_barrier(conf_t *conf)
790{
791 unsigned long flags;
792 spin_lock_irqsave(&conf->resync_lock, flags);
793 conf->nr_pending--;
794 spin_unlock_irqrestore(&conf->resync_lock, flags);
795 wake_up(&conf->wait_barrier);
796}
797
798static void freeze_array(conf_t *conf)
799{
800 /* stop syncio and normal IO and wait for everything to
801 * go quiet.
802 * We increment barrier and nr_waiting, and then
803 * wait until nr_pending match nr_queued+1
804 * This is called in the context of one normal IO request
805 * that has failed. Thus any sync request that might be pending
806 * will be blocked by nr_pending, and we need to wait for
807 * pending IO requests to complete or be queued for re-try.
808 * Thus the number queued (nr_queued) plus this request (1)
809 * must match the number of pending IOs (nr_pending) before
810 * we continue.
811 */
812 spin_lock_irq(&conf->resync_lock);
813 conf->barrier++;
814 conf->nr_waiting++;
815 wait_event_lock_irq(conf->wait_barrier,
816 conf->nr_pending == conf->nr_queued+1,
817 conf->resync_lock,
818 flush_pending_writes(conf));
819
820 spin_unlock_irq(&conf->resync_lock);
821}
822
823static void unfreeze_array(conf_t *conf)
824{
825 /* reverse the effect of the freeze */
826 spin_lock_irq(&conf->resync_lock);
827 conf->barrier--;
828 conf->nr_waiting--;
829 wake_up(&conf->wait_barrier);
830 spin_unlock_irq(&conf->resync_lock);
831}
832
833static int make_request(mddev_t *mddev, struct bio * bio)
834{
835 conf_t *conf = mddev->private;
836 mirror_info_t *mirror;
837 r10bio_t *r10_bio;
838 struct bio *read_bio;
839 int i;
840 int chunk_sects = conf->chunk_mask + 1;
841 const int rw = bio_data_dir(bio);
842 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
843 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
844 unsigned long flags;
845 mdk_rdev_t *blocked_rdev;
846 int plugged;
847 int sectors_handled;
848 int max_sectors;
849
850 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
851 md_flush_request(mddev, bio);
852 return 0;
853 }
854
855 /* If this request crosses a chunk boundary, we need to
856 * split it. This will only happen for 1 PAGE (or less) requests.
857 */
858 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
859 > chunk_sects &&
860 conf->near_copies < conf->raid_disks)) {
861 struct bio_pair *bp;
862 /* Sanity check -- queue functions should prevent this happening */
863 if (bio->bi_vcnt != 1 ||
864 bio->bi_idx != 0)
865 goto bad_map;
866 /* This is a one page bio that upper layers
867 * refuse to split for us, so we need to split it.
868 */
869 bp = bio_split(bio,
870 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
871
872 /* Each of these 'make_request' calls will call 'wait_barrier'.
873 * If the first succeeds but the second blocks due to the resync
874 * thread raising the barrier, we will deadlock because the
875 * IO to the underlying device will be queued in generic_make_request
876 * and will never complete, so will never reduce nr_pending.
877 * So increment nr_waiting here so no new raise_barriers will
878 * succeed, and so the second wait_barrier cannot block.
879 */
880 spin_lock_irq(&conf->resync_lock);
881 conf->nr_waiting++;
882 spin_unlock_irq(&conf->resync_lock);
883
884 if (make_request(mddev, &bp->bio1))
885 generic_make_request(&bp->bio1);
886 if (make_request(mddev, &bp->bio2))
887 generic_make_request(&bp->bio2);
888
889 spin_lock_irq(&conf->resync_lock);
890 conf->nr_waiting--;
891 wake_up(&conf->wait_barrier);
892 spin_unlock_irq(&conf->resync_lock);
893
894 bio_pair_release(bp);
895 return 0;
896 bad_map:
897 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
898 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
899 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
900
901 bio_io_error(bio);
902 return 0;
903 }
904
905 md_write_start(mddev, bio);
906
907 /*
908 * Register the new request and wait if the reconstruction
909 * thread has put up a bar for new requests.
910 * Continue immediately if no resync is active currently.
911 */
912 wait_barrier(conf);
913
914 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
915
916 r10_bio->master_bio = bio;
917 r10_bio->sectors = bio->bi_size >> 9;
918
919 r10_bio->mddev = mddev;
920 r10_bio->sector = bio->bi_sector;
921 r10_bio->state = 0;
922
923 /* We might need to issue multiple reads to different
924 * devices if there are bad blocks around, so we keep
925 * track of the number of reads in bio->bi_phys_segments.
926 * If this is 0, there is only one r10_bio and no locking
927 * will be needed when the request completes. If it is
928 * non-zero, then it is the number of not-completed requests.
929 */
930 bio->bi_phys_segments = 0;
931 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
932
933 if (rw == READ) {
934 /*
935 * read balancing logic:
936 */
937 int disk;
938 int slot;
939
940read_again:
941 disk = read_balance(conf, r10_bio, &max_sectors);
942 slot = r10_bio->read_slot;
943 if (disk < 0) {
944 raid_end_bio_io(r10_bio);
945 return 0;
946 }
947 mirror = conf->mirrors + disk;
948
949 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
950 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
951 max_sectors);
952
953 r10_bio->devs[slot].bio = read_bio;
954
955 read_bio->bi_sector = r10_bio->devs[slot].addr +
956 mirror->rdev->data_offset;
957 read_bio->bi_bdev = mirror->rdev->bdev;
958 read_bio->bi_end_io = raid10_end_read_request;
959 read_bio->bi_rw = READ | do_sync;
960 read_bio->bi_private = r10_bio;
961
962 if (max_sectors < r10_bio->sectors) {
963 /* Could not read all from this device, so we will
964 * need another r10_bio.
965 */
966 sectors_handled = (r10_bio->sectors + max_sectors
967 - bio->bi_sector);
968 r10_bio->sectors = max_sectors;
969 spin_lock_irq(&conf->device_lock);
970 if (bio->bi_phys_segments == 0)
971 bio->bi_phys_segments = 2;
972 else
973 bio->bi_phys_segments++;
974 spin_unlock(&conf->device_lock);
975 /* Cannot call generic_make_request directly
976 * as that will be queued in __generic_make_request
977 * and subsequent mempool_alloc might block
978 * waiting for it. so hand bio over to raid10d.
979 */
980 reschedule_retry(r10_bio);
981
982 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
983
984 r10_bio->master_bio = bio;
985 r10_bio->sectors = ((bio->bi_size >> 9)
986 - sectors_handled);
987 r10_bio->state = 0;
988 r10_bio->mddev = mddev;
989 r10_bio->sector = bio->bi_sector + sectors_handled;
990 goto read_again;
991 } else
992 generic_make_request(read_bio);
993 return 0;
994 }
995
996 /*
997 * WRITE:
998 */
999 /* first select target devices under rcu_lock and
1000 * inc refcount on their rdev. Record them by setting
1001 * bios[x] to bio
1002 * If there are known/acknowledged bad blocks on any device
1003 * on which we have seen a write error, we want to avoid
1004 * writing to those blocks. This potentially requires several
1005 * writes to write around the bad blocks. Each set of writes
1006 * gets its own r10_bio with a set of bios attached. The number
1007 * of r10_bios is recored in bio->bi_phys_segments just as with
1008 * the read case.
1009 */
1010 plugged = mddev_check_plugged(mddev);
1011
1012 raid10_find_phys(conf, r10_bio);
1013retry_write:
1014 blocked_rdev = NULL;
1015 rcu_read_lock();
1016 max_sectors = r10_bio->sectors;
1017
1018 for (i = 0; i < conf->copies; i++) {
1019 int d = r10_bio->devs[i].devnum;
1020 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
1021 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1022 atomic_inc(&rdev->nr_pending);
1023 blocked_rdev = rdev;
1024 break;
1025 }
1026 r10_bio->devs[i].bio = NULL;
1027 if (!rdev || test_bit(Faulty, &rdev->flags)) {
1028 set_bit(R10BIO_Degraded, &r10_bio->state);
1029 continue;
1030 }
1031 if (test_bit(WriteErrorSeen, &rdev->flags)) {
1032 sector_t first_bad;
1033 sector_t dev_sector = r10_bio->devs[i].addr;
1034 int bad_sectors;
1035 int is_bad;
1036
1037 is_bad = is_badblock(rdev, dev_sector,
1038 max_sectors,
1039 &first_bad, &bad_sectors);
1040 if (is_bad < 0) {
1041 /* Mustn't write here until the bad block
1042 * is acknowledged
1043 */
1044 atomic_inc(&rdev->nr_pending);
1045 set_bit(BlockedBadBlocks, &rdev->flags);
1046 blocked_rdev = rdev;
1047 break;
1048 }
1049 if (is_bad && first_bad <= dev_sector) {
1050 /* Cannot write here at all */
1051 bad_sectors -= (dev_sector - first_bad);
1052 if (bad_sectors < max_sectors)
1053 /* Mustn't write more than bad_sectors
1054 * to other devices yet
1055 */
1056 max_sectors = bad_sectors;
1057 /* We don't set R10BIO_Degraded as that
1058 * only applies if the disk is missing,
1059 * so it might be re-added, and we want to
1060 * know to recover this chunk.
1061 * In this case the device is here, and the
1062 * fact that this chunk is not in-sync is
1063 * recorded in the bad block log.
1064 */
1065 continue;
1066 }
1067 if (is_bad) {
1068 int good_sectors = first_bad - dev_sector;
1069 if (good_sectors < max_sectors)
1070 max_sectors = good_sectors;
1071 }
1072 }
1073 r10_bio->devs[i].bio = bio;
1074 atomic_inc(&rdev->nr_pending);
1075 }
1076 rcu_read_unlock();
1077
1078 if (unlikely(blocked_rdev)) {
1079 /* Have to wait for this device to get unblocked, then retry */
1080 int j;
1081 int d;
1082
1083 for (j = 0; j < i; j++)
1084 if (r10_bio->devs[j].bio) {
1085 d = r10_bio->devs[j].devnum;
1086 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1087 }
1088 allow_barrier(conf);
1089 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1090 wait_barrier(conf);
1091 goto retry_write;
1092 }
1093
1094 if (max_sectors < r10_bio->sectors) {
1095 /* We are splitting this into multiple parts, so
1096 * we need to prepare for allocating another r10_bio.
1097 */
1098 r10_bio->sectors = max_sectors;
1099 spin_lock_irq(&conf->device_lock);
1100 if (bio->bi_phys_segments == 0)
1101 bio->bi_phys_segments = 2;
1102 else
1103 bio->bi_phys_segments++;
1104 spin_unlock_irq(&conf->device_lock);
1105 }
1106 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1107
1108 atomic_set(&r10_bio->remaining, 1);
1109 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1110
1111 for (i = 0; i < conf->copies; i++) {
1112 struct bio *mbio;
1113 int d = r10_bio->devs[i].devnum;
1114 if (!r10_bio->devs[i].bio)
1115 continue;
1116
1117 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1118 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1119 max_sectors);
1120 r10_bio->devs[i].bio = mbio;
1121
1122 mbio->bi_sector = (r10_bio->devs[i].addr+
1123 conf->mirrors[d].rdev->data_offset);
1124 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1125 mbio->bi_end_io = raid10_end_write_request;
1126 mbio->bi_rw = WRITE | do_sync | do_fua;
1127 mbio->bi_private = r10_bio;
1128
1129 atomic_inc(&r10_bio->remaining);
1130 spin_lock_irqsave(&conf->device_lock, flags);
1131 bio_list_add(&conf->pending_bio_list, mbio);
1132 spin_unlock_irqrestore(&conf->device_lock, flags);
1133 }
1134
1135 /* Don't remove the bias on 'remaining' (one_write_done) until
1136 * after checking if we need to go around again.
1137 */
1138
1139 if (sectors_handled < (bio->bi_size >> 9)) {
1140 one_write_done(r10_bio);
1141 /* We need another r10_bio. It has already been counted
1142 * in bio->bi_phys_segments.
1143 */
1144 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1145
1146 r10_bio->master_bio = bio;
1147 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1148
1149 r10_bio->mddev = mddev;
1150 r10_bio->sector = bio->bi_sector + sectors_handled;
1151 r10_bio->state = 0;
1152 goto retry_write;
1153 }
1154 one_write_done(r10_bio);
1155
1156 /* In case raid10d snuck in to freeze_array */
1157 wake_up(&conf->wait_barrier);
1158
1159 if (do_sync || !mddev->bitmap || !plugged)
1160 md_wakeup_thread(mddev->thread);
1161 return 0;
1162}
1163
1164static void status(struct seq_file *seq, mddev_t *mddev)
1165{
1166 conf_t *conf = mddev->private;
1167 int i;
1168
1169 if (conf->near_copies < conf->raid_disks)
1170 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1171 if (conf->near_copies > 1)
1172 seq_printf(seq, " %d near-copies", conf->near_copies);
1173 if (conf->far_copies > 1) {
1174 if (conf->far_offset)
1175 seq_printf(seq, " %d offset-copies", conf->far_copies);
1176 else
1177 seq_printf(seq, " %d far-copies", conf->far_copies);
1178 }
1179 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1180 conf->raid_disks - mddev->degraded);
1181 for (i = 0; i < conf->raid_disks; i++)
1182 seq_printf(seq, "%s",
1183 conf->mirrors[i].rdev &&
1184 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1185 seq_printf(seq, "]");
1186}
1187
1188/* check if there are enough drives for
1189 * every block to appear on atleast one.
1190 * Don't consider the device numbered 'ignore'
1191 * as we might be about to remove it.
1192 */
1193static int enough(conf_t *conf, int ignore)
1194{
1195 int first = 0;
1196
1197 do {
1198 int n = conf->copies;
1199 int cnt = 0;
1200 while (n--) {
1201 if (conf->mirrors[first].rdev &&
1202 first != ignore)
1203 cnt++;
1204 first = (first+1) % conf->raid_disks;
1205 }
1206 if (cnt == 0)
1207 return 0;
1208 } while (first != 0);
1209 return 1;
1210}
1211
1212static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1213{
1214 char b[BDEVNAME_SIZE];
1215 conf_t *conf = mddev->private;
1216
1217 /*
1218 * If it is not operational, then we have already marked it as dead
1219 * else if it is the last working disks, ignore the error, let the
1220 * next level up know.
1221 * else mark the drive as failed
1222 */
1223 if (test_bit(In_sync, &rdev->flags)
1224 && !enough(conf, rdev->raid_disk))
1225 /*
1226 * Don't fail the drive, just return an IO error.
1227 */
1228 return;
1229 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1230 unsigned long flags;
1231 spin_lock_irqsave(&conf->device_lock, flags);
1232 mddev->degraded++;
1233 spin_unlock_irqrestore(&conf->device_lock, flags);
1234 /*
1235 * if recovery is running, make sure it aborts.
1236 */
1237 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1238 }
1239 set_bit(Blocked, &rdev->flags);
1240 set_bit(Faulty, &rdev->flags);
1241 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1242 printk(KERN_ALERT
1243 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1244 "md/raid10:%s: Operation continuing on %d devices.\n",
1245 mdname(mddev), bdevname(rdev->bdev, b),
1246 mdname(mddev), conf->raid_disks - mddev->degraded);
1247}
1248
1249static void print_conf(conf_t *conf)
1250{
1251 int i;
1252 mirror_info_t *tmp;
1253
1254 printk(KERN_DEBUG "RAID10 conf printout:\n");
1255 if (!conf) {
1256 printk(KERN_DEBUG "(!conf)\n");
1257 return;
1258 }
1259 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1260 conf->raid_disks);
1261
1262 for (i = 0; i < conf->raid_disks; i++) {
1263 char b[BDEVNAME_SIZE];
1264 tmp = conf->mirrors + i;
1265 if (tmp->rdev)
1266 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1267 i, !test_bit(In_sync, &tmp->rdev->flags),
1268 !test_bit(Faulty, &tmp->rdev->flags),
1269 bdevname(tmp->rdev->bdev,b));
1270 }
1271}
1272
1273static void close_sync(conf_t *conf)
1274{
1275 wait_barrier(conf);
1276 allow_barrier(conf);
1277
1278 mempool_destroy(conf->r10buf_pool);
1279 conf->r10buf_pool = NULL;
1280}
1281
1282static int raid10_spare_active(mddev_t *mddev)
1283{
1284 int i;
1285 conf_t *conf = mddev->private;
1286 mirror_info_t *tmp;
1287 int count = 0;
1288 unsigned long flags;
1289
1290 /*
1291 * Find all non-in_sync disks within the RAID10 configuration
1292 * and mark them in_sync
1293 */
1294 for (i = 0; i < conf->raid_disks; i++) {
1295 tmp = conf->mirrors + i;
1296 if (tmp->rdev
1297 && !test_bit(Faulty, &tmp->rdev->flags)
1298 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1299 count++;
1300 sysfs_notify_dirent(tmp->rdev->sysfs_state);
1301 }
1302 }
1303 spin_lock_irqsave(&conf->device_lock, flags);
1304 mddev->degraded -= count;
1305 spin_unlock_irqrestore(&conf->device_lock, flags);
1306
1307 print_conf(conf);
1308 return count;
1309}
1310
1311
1312static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1313{
1314 conf_t *conf = mddev->private;
1315 int err = -EEXIST;
1316 int mirror;
1317 int first = 0;
1318 int last = conf->raid_disks - 1;
1319
1320 if (mddev->recovery_cp < MaxSector)
1321 /* only hot-add to in-sync arrays, as recovery is
1322 * very different from resync
1323 */
1324 return -EBUSY;
1325 if (!enough(conf, -1))
1326 return -EINVAL;
1327
1328 if (rdev->raid_disk >= 0)
1329 first = last = rdev->raid_disk;
1330
1331 if (rdev->saved_raid_disk >= first &&
1332 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1333 mirror = rdev->saved_raid_disk;
1334 else
1335 mirror = first;
1336 for ( ; mirror <= last ; mirror++) {
1337 mirror_info_t *p = &conf->mirrors[mirror];
1338 if (p->recovery_disabled == mddev->recovery_disabled)
1339 continue;
1340 if (!p->rdev)
1341 continue;
1342
1343 disk_stack_limits(mddev->gendisk, rdev->bdev,
1344 rdev->data_offset << 9);
1345 /* as we don't honour merge_bvec_fn, we must
1346 * never risk violating it, so limit
1347 * ->max_segments to one lying with a single
1348 * page, as a one page request is never in
1349 * violation.
1350 */
1351 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1352 blk_queue_max_segments(mddev->queue, 1);
1353 blk_queue_segment_boundary(mddev->queue,
1354 PAGE_CACHE_SIZE - 1);
1355 }
1356
1357 p->head_position = 0;
1358 rdev->raid_disk = mirror;
1359 err = 0;
1360 if (rdev->saved_raid_disk != mirror)
1361 conf->fullsync = 1;
1362 rcu_assign_pointer(p->rdev, rdev);
1363 break;
1364 }
1365
1366 md_integrity_add_rdev(rdev, mddev);
1367 print_conf(conf);
1368 return err;
1369}
1370
1371static int raid10_remove_disk(mddev_t *mddev, int number)
1372{
1373 conf_t *conf = mddev->private;
1374 int err = 0;
1375 mdk_rdev_t *rdev;
1376 mirror_info_t *p = conf->mirrors+ number;
1377
1378 print_conf(conf);
1379 rdev = p->rdev;
1380 if (rdev) {
1381 if (test_bit(In_sync, &rdev->flags) ||
1382 atomic_read(&rdev->nr_pending)) {
1383 err = -EBUSY;
1384 goto abort;
1385 }
1386 /* Only remove faulty devices in recovery
1387 * is not possible.
1388 */
1389 if (!test_bit(Faulty, &rdev->flags) &&
1390 mddev->recovery_disabled != p->recovery_disabled &&
1391 enough(conf, -1)) {
1392 err = -EBUSY;
1393 goto abort;
1394 }
1395 p->rdev = NULL;
1396 synchronize_rcu();
1397 if (atomic_read(&rdev->nr_pending)) {
1398 /* lost the race, try later */
1399 err = -EBUSY;
1400 p->rdev = rdev;
1401 goto abort;
1402 }
1403 err = md_integrity_register(mddev);
1404 }
1405abort:
1406
1407 print_conf(conf);
1408 return err;
1409}
1410
1411
1412static void end_sync_read(struct bio *bio, int error)
1413{
1414 r10bio_t *r10_bio = bio->bi_private;
1415 conf_t *conf = r10_bio->mddev->private;
1416 int d;
1417
1418 d = find_bio_disk(conf, r10_bio, bio, NULL);
1419
1420 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1421 set_bit(R10BIO_Uptodate, &r10_bio->state);
1422 else
1423 /* The write handler will notice the lack of
1424 * R10BIO_Uptodate and record any errors etc
1425 */
1426 atomic_add(r10_bio->sectors,
1427 &conf->mirrors[d].rdev->corrected_errors);
1428
1429 /* for reconstruct, we always reschedule after a read.
1430 * for resync, only after all reads
1431 */
1432 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1433 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1434 atomic_dec_and_test(&r10_bio->remaining)) {
1435 /* we have read all the blocks,
1436 * do the comparison in process context in raid10d
1437 */
1438 reschedule_retry(r10_bio);
1439 }
1440}
1441
1442static void end_sync_request(r10bio_t *r10_bio)
1443{
1444 mddev_t *mddev = r10_bio->mddev;
1445
1446 while (atomic_dec_and_test(&r10_bio->remaining)) {
1447 if (r10_bio->master_bio == NULL) {
1448 /* the primary of several recovery bios */
1449 sector_t s = r10_bio->sectors;
1450 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1451 test_bit(R10BIO_WriteError, &r10_bio->state))
1452 reschedule_retry(r10_bio);
1453 else
1454 put_buf(r10_bio);
1455 md_done_sync(mddev, s, 1);
1456 break;
1457 } else {
1458 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1459 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1460 test_bit(R10BIO_WriteError, &r10_bio->state))
1461 reschedule_retry(r10_bio);
1462 else
1463 put_buf(r10_bio);
1464 r10_bio = r10_bio2;
1465 }
1466 }
1467}
1468
1469static void end_sync_write(struct bio *bio, int error)
1470{
1471 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1472 r10bio_t *r10_bio = bio->bi_private;
1473 mddev_t *mddev = r10_bio->mddev;
1474 conf_t *conf = mddev->private;
1475 int d;
1476 sector_t first_bad;
1477 int bad_sectors;
1478 int slot;
1479
1480 d = find_bio_disk(conf, r10_bio, bio, &slot);
1481
1482 if (!uptodate) {
1483 set_bit(WriteErrorSeen, &conf->mirrors[d].rdev->flags);
1484 set_bit(R10BIO_WriteError, &r10_bio->state);
1485 } else if (is_badblock(conf->mirrors[d].rdev,
1486 r10_bio->devs[slot].addr,
1487 r10_bio->sectors,
1488 &first_bad, &bad_sectors))
1489 set_bit(R10BIO_MadeGood, &r10_bio->state);
1490
1491 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1492
1493 end_sync_request(r10_bio);
1494}
1495
1496/*
1497 * Note: sync and recover and handled very differently for raid10
1498 * This code is for resync.
1499 * For resync, we read through virtual addresses and read all blocks.
1500 * If there is any error, we schedule a write. The lowest numbered
1501 * drive is authoritative.
1502 * However requests come for physical address, so we need to map.
1503 * For every physical address there are raid_disks/copies virtual addresses,
1504 * which is always are least one, but is not necessarly an integer.
1505 * This means that a physical address can span multiple chunks, so we may
1506 * have to submit multiple io requests for a single sync request.
1507 */
1508/*
1509 * We check if all blocks are in-sync and only write to blocks that
1510 * aren't in sync
1511 */
1512static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1513{
1514 conf_t *conf = mddev->private;
1515 int i, first;
1516 struct bio *tbio, *fbio;
1517
1518 atomic_set(&r10_bio->remaining, 1);
1519
1520 /* find the first device with a block */
1521 for (i=0; i<conf->copies; i++)
1522 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1523 break;
1524
1525 if (i == conf->copies)
1526 goto done;
1527
1528 first = i;
1529 fbio = r10_bio->devs[i].bio;
1530
1531 /* now find blocks with errors */
1532 for (i=0 ; i < conf->copies ; i++) {
1533 int j, d;
1534 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1535
1536 tbio = r10_bio->devs[i].bio;
1537
1538 if (tbio->bi_end_io != end_sync_read)
1539 continue;
1540 if (i == first)
1541 continue;
1542 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1543 /* We know that the bi_io_vec layout is the same for
1544 * both 'first' and 'i', so we just compare them.
1545 * All vec entries are PAGE_SIZE;
1546 */
1547 for (j = 0; j < vcnt; j++)
1548 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1549 page_address(tbio->bi_io_vec[j].bv_page),
1550 PAGE_SIZE))
1551 break;
1552 if (j == vcnt)
1553 continue;
1554 mddev->resync_mismatches += r10_bio->sectors;
1555 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1556 /* Don't fix anything. */
1557 continue;
1558 }
1559 /* Ok, we need to write this bio, either to correct an
1560 * inconsistency or to correct an unreadable block.
1561 * First we need to fixup bv_offset, bv_len and
1562 * bi_vecs, as the read request might have corrupted these
1563 */
1564 tbio->bi_vcnt = vcnt;
1565 tbio->bi_size = r10_bio->sectors << 9;
1566 tbio->bi_idx = 0;
1567 tbio->bi_phys_segments = 0;
1568 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1569 tbio->bi_flags |= 1 << BIO_UPTODATE;
1570 tbio->bi_next = NULL;
1571 tbio->bi_rw = WRITE;
1572 tbio->bi_private = r10_bio;
1573 tbio->bi_sector = r10_bio->devs[i].addr;
1574
1575 for (j=0; j < vcnt ; j++) {
1576 tbio->bi_io_vec[j].bv_offset = 0;
1577 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1578
1579 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1580 page_address(fbio->bi_io_vec[j].bv_page),
1581 PAGE_SIZE);
1582 }
1583 tbio->bi_end_io = end_sync_write;
1584
1585 d = r10_bio->devs[i].devnum;
1586 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1587 atomic_inc(&r10_bio->remaining);
1588 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1589
1590 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1591 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1592 generic_make_request(tbio);
1593 }
1594
1595done:
1596 if (atomic_dec_and_test(&r10_bio->remaining)) {
1597 md_done_sync(mddev, r10_bio->sectors, 1);
1598 put_buf(r10_bio);
1599 }
1600}
1601
1602/*
1603 * Now for the recovery code.
1604 * Recovery happens across physical sectors.
1605 * We recover all non-is_sync drives by finding the virtual address of
1606 * each, and then choose a working drive that also has that virt address.
1607 * There is a separate r10_bio for each non-in_sync drive.
1608 * Only the first two slots are in use. The first for reading,
1609 * The second for writing.
1610 *
1611 */
1612static void fix_recovery_read_error(r10bio_t *r10_bio)
1613{
1614 /* We got a read error during recovery.
1615 * We repeat the read in smaller page-sized sections.
1616 * If a read succeeds, write it to the new device or record
1617 * a bad block if we cannot.
1618 * If a read fails, record a bad block on both old and
1619 * new devices.
1620 */
1621 mddev_t *mddev = r10_bio->mddev;
1622 conf_t *conf = mddev->private;
1623 struct bio *bio = r10_bio->devs[0].bio;
1624 sector_t sect = 0;
1625 int sectors = r10_bio->sectors;
1626 int idx = 0;
1627 int dr = r10_bio->devs[0].devnum;
1628 int dw = r10_bio->devs[1].devnum;
1629
1630 while (sectors) {
1631 int s = sectors;
1632 mdk_rdev_t *rdev;
1633 sector_t addr;
1634 int ok;
1635
1636 if (s > (PAGE_SIZE>>9))
1637 s = PAGE_SIZE >> 9;
1638
1639 rdev = conf->mirrors[dr].rdev;
1640 addr = r10_bio->devs[0].addr + sect,
1641 ok = sync_page_io(rdev,
1642 addr,
1643 s << 9,
1644 bio->bi_io_vec[idx].bv_page,
1645 READ, false);
1646 if (ok) {
1647 rdev = conf->mirrors[dw].rdev;
1648 addr = r10_bio->devs[1].addr + sect;
1649 ok = sync_page_io(rdev,
1650 addr,
1651 s << 9,
1652 bio->bi_io_vec[idx].bv_page,
1653 WRITE, false);
1654 if (!ok)
1655 set_bit(WriteErrorSeen, &rdev->flags);
1656 }
1657 if (!ok) {
1658 /* We don't worry if we cannot set a bad block -
1659 * it really is bad so there is no loss in not
1660 * recording it yet
1661 */
1662 rdev_set_badblocks(rdev, addr, s, 0);
1663
1664 if (rdev != conf->mirrors[dw].rdev) {
1665 /* need bad block on destination too */
1666 mdk_rdev_t *rdev2 = conf->mirrors[dw].rdev;
1667 addr = r10_bio->devs[1].addr + sect;
1668 ok = rdev_set_badblocks(rdev2, addr, s, 0);
1669 if (!ok) {
1670 /* just abort the recovery */
1671 printk(KERN_NOTICE
1672 "md/raid10:%s: recovery aborted"
1673 " due to read error\n",
1674 mdname(mddev));
1675
1676 conf->mirrors[dw].recovery_disabled
1677 = mddev->recovery_disabled;
1678 set_bit(MD_RECOVERY_INTR,
1679 &mddev->recovery);
1680 break;
1681 }
1682 }
1683 }
1684
1685 sectors -= s;
1686 sect += s;
1687 idx++;
1688 }
1689}
1690
1691static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1692{
1693 conf_t *conf = mddev->private;
1694 int d;
1695 struct bio *wbio;
1696
1697 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
1698 fix_recovery_read_error(r10_bio);
1699 end_sync_request(r10_bio);
1700 return;
1701 }
1702
1703 /*
1704 * share the pages with the first bio
1705 * and submit the write request
1706 */
1707 wbio = r10_bio->devs[1].bio;
1708 d = r10_bio->devs[1].devnum;
1709
1710 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1711 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1712 generic_make_request(wbio);
1713}
1714
1715
1716/*
1717 * Used by fix_read_error() to decay the per rdev read_errors.
1718 * We halve the read error count for every hour that has elapsed
1719 * since the last recorded read error.
1720 *
1721 */
1722static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1723{
1724 struct timespec cur_time_mon;
1725 unsigned long hours_since_last;
1726 unsigned int read_errors = atomic_read(&rdev->read_errors);
1727
1728 ktime_get_ts(&cur_time_mon);
1729
1730 if (rdev->last_read_error.tv_sec == 0 &&
1731 rdev->last_read_error.tv_nsec == 0) {
1732 /* first time we've seen a read error */
1733 rdev->last_read_error = cur_time_mon;
1734 return;
1735 }
1736
1737 hours_since_last = (cur_time_mon.tv_sec -
1738 rdev->last_read_error.tv_sec) / 3600;
1739
1740 rdev->last_read_error = cur_time_mon;
1741
1742 /*
1743 * if hours_since_last is > the number of bits in read_errors
1744 * just set read errors to 0. We do this to avoid
1745 * overflowing the shift of read_errors by hours_since_last.
1746 */
1747 if (hours_since_last >= 8 * sizeof(read_errors))
1748 atomic_set(&rdev->read_errors, 0);
1749 else
1750 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1751}
1752
1753static int r10_sync_page_io(mdk_rdev_t *rdev, sector_t sector,
1754 int sectors, struct page *page, int rw)
1755{
1756 sector_t first_bad;
1757 int bad_sectors;
1758
1759 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
1760 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
1761 return -1;
1762 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1763 /* success */
1764 return 1;
1765 if (rw == WRITE)
1766 set_bit(WriteErrorSeen, &rdev->flags);
1767 /* need to record an error - either for the block or the device */
1768 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1769 md_error(rdev->mddev, rdev);
1770 return 0;
1771}
1772
1773/*
1774 * This is a kernel thread which:
1775 *
1776 * 1. Retries failed read operations on working mirrors.
1777 * 2. Updates the raid superblock when problems encounter.
1778 * 3. Performs writes following reads for array synchronising.
1779 */
1780
1781static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1782{
1783 int sect = 0; /* Offset from r10_bio->sector */
1784 int sectors = r10_bio->sectors;
1785 mdk_rdev_t*rdev;
1786 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
1787 int d = r10_bio->devs[r10_bio->read_slot].devnum;
1788
1789 /* still own a reference to this rdev, so it cannot
1790 * have been cleared recently.
1791 */
1792 rdev = conf->mirrors[d].rdev;
1793
1794 if (test_bit(Faulty, &rdev->flags))
1795 /* drive has already been failed, just ignore any
1796 more fix_read_error() attempts */
1797 return;
1798
1799 check_decay_read_errors(mddev, rdev);
1800 atomic_inc(&rdev->read_errors);
1801 if (atomic_read(&rdev->read_errors) > max_read_errors) {
1802 char b[BDEVNAME_SIZE];
1803 bdevname(rdev->bdev, b);
1804
1805 printk(KERN_NOTICE
1806 "md/raid10:%s: %s: Raid device exceeded "
1807 "read_error threshold [cur %d:max %d]\n",
1808 mdname(mddev), b,
1809 atomic_read(&rdev->read_errors), max_read_errors);
1810 printk(KERN_NOTICE
1811 "md/raid10:%s: %s: Failing raid device\n",
1812 mdname(mddev), b);
1813 md_error(mddev, conf->mirrors[d].rdev);
1814 return;
1815 }
1816
1817 while(sectors) {
1818 int s = sectors;
1819 int sl = r10_bio->read_slot;
1820 int success = 0;
1821 int start;
1822
1823 if (s > (PAGE_SIZE>>9))
1824 s = PAGE_SIZE >> 9;
1825
1826 rcu_read_lock();
1827 do {
1828 sector_t first_bad;
1829 int bad_sectors;
1830
1831 d = r10_bio->devs[sl].devnum;
1832 rdev = rcu_dereference(conf->mirrors[d].rdev);
1833 if (rdev &&
1834 test_bit(In_sync, &rdev->flags) &&
1835 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
1836 &first_bad, &bad_sectors) == 0) {
1837 atomic_inc(&rdev->nr_pending);
1838 rcu_read_unlock();
1839 success = sync_page_io(rdev,
1840 r10_bio->devs[sl].addr +
1841 sect,
1842 s<<9,
1843 conf->tmppage, READ, false);
1844 rdev_dec_pending(rdev, mddev);
1845 rcu_read_lock();
1846 if (success)
1847 break;
1848 }
1849 sl++;
1850 if (sl == conf->copies)
1851 sl = 0;
1852 } while (!success && sl != r10_bio->read_slot);
1853 rcu_read_unlock();
1854
1855 if (!success) {
1856 /* Cannot read from anywhere, just mark the block
1857 * as bad on the first device to discourage future
1858 * reads.
1859 */
1860 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1861 rdev = conf->mirrors[dn].rdev;
1862
1863 if (!rdev_set_badblocks(
1864 rdev,
1865 r10_bio->devs[r10_bio->read_slot].addr
1866 + sect,
1867 s, 0))
1868 md_error(mddev, rdev);
1869 break;
1870 }
1871
1872 start = sl;
1873 /* write it back and re-read */
1874 rcu_read_lock();
1875 while (sl != r10_bio->read_slot) {
1876 char b[BDEVNAME_SIZE];
1877
1878 if (sl==0)
1879 sl = conf->copies;
1880 sl--;
1881 d = r10_bio->devs[sl].devnum;
1882 rdev = rcu_dereference(conf->mirrors[d].rdev);
1883 if (!rdev ||
1884 !test_bit(In_sync, &rdev->flags))
1885 continue;
1886
1887 atomic_inc(&rdev->nr_pending);
1888 rcu_read_unlock();
1889 if (r10_sync_page_io(rdev,
1890 r10_bio->devs[sl].addr +
1891 sect,
1892 s<<9, conf->tmppage, WRITE)
1893 == 0) {
1894 /* Well, this device is dead */
1895 printk(KERN_NOTICE
1896 "md/raid10:%s: read correction "
1897 "write failed"
1898 " (%d sectors at %llu on %s)\n",
1899 mdname(mddev), s,
1900 (unsigned long long)(
1901 sect + rdev->data_offset),
1902 bdevname(rdev->bdev, b));
1903 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1904 "drive\n",
1905 mdname(mddev),
1906 bdevname(rdev->bdev, b));
1907 }
1908 rdev_dec_pending(rdev, mddev);
1909 rcu_read_lock();
1910 }
1911 sl = start;
1912 while (sl != r10_bio->read_slot) {
1913 char b[BDEVNAME_SIZE];
1914
1915 if (sl==0)
1916 sl = conf->copies;
1917 sl--;
1918 d = r10_bio->devs[sl].devnum;
1919 rdev = rcu_dereference(conf->mirrors[d].rdev);
1920 if (!rdev ||
1921 !test_bit(In_sync, &rdev->flags))
1922 continue;
1923
1924 atomic_inc(&rdev->nr_pending);
1925 rcu_read_unlock();
1926 switch (r10_sync_page_io(rdev,
1927 r10_bio->devs[sl].addr +
1928 sect,
1929 s<<9, conf->tmppage,
1930 READ)) {
1931 case 0:
1932 /* Well, this device is dead */
1933 printk(KERN_NOTICE
1934 "md/raid10:%s: unable to read back "
1935 "corrected sectors"
1936 " (%d sectors at %llu on %s)\n",
1937 mdname(mddev), s,
1938 (unsigned long long)(
1939 sect + rdev->data_offset),
1940 bdevname(rdev->bdev, b));
1941 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
1942 "drive\n",
1943 mdname(mddev),
1944 bdevname(rdev->bdev, b));
1945 break;
1946 case 1:
1947 printk(KERN_INFO
1948 "md/raid10:%s: read error corrected"
1949 " (%d sectors at %llu on %s)\n",
1950 mdname(mddev), s,
1951 (unsigned long long)(
1952 sect + rdev->data_offset),
1953 bdevname(rdev->bdev, b));
1954 atomic_add(s, &rdev->corrected_errors);
1955 }
1956
1957 rdev_dec_pending(rdev, mddev);
1958 rcu_read_lock();
1959 }
1960 rcu_read_unlock();
1961
1962 sectors -= s;
1963 sect += s;
1964 }
1965}
1966
1967static void bi_complete(struct bio *bio, int error)
1968{
1969 complete((struct completion *)bio->bi_private);
1970}
1971
1972static int submit_bio_wait(int rw, struct bio *bio)
1973{
1974 struct completion event;
1975 rw |= REQ_SYNC;
1976
1977 init_completion(&event);
1978 bio->bi_private = &event;
1979 bio->bi_end_io = bi_complete;
1980 submit_bio(rw, bio);
1981 wait_for_completion(&event);
1982
1983 return test_bit(BIO_UPTODATE, &bio->bi_flags);
1984}
1985
1986static int narrow_write_error(r10bio_t *r10_bio, int i)
1987{
1988 struct bio *bio = r10_bio->master_bio;
1989 mddev_t *mddev = r10_bio->mddev;
1990 conf_t *conf = mddev->private;
1991 mdk_rdev_t *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
1992 /* bio has the data to be written to slot 'i' where
1993 * we just recently had a write error.
1994 * We repeatedly clone the bio and trim down to one block,
1995 * then try the write. Where the write fails we record
1996 * a bad block.
1997 * It is conceivable that the bio doesn't exactly align with
1998 * blocks. We must handle this.
1999 *
2000 * We currently own a reference to the rdev.
2001 */
2002
2003 int block_sectors;
2004 sector_t sector;
2005 int sectors;
2006 int sect_to_write = r10_bio->sectors;
2007 int ok = 1;
2008
2009 if (rdev->badblocks.shift < 0)
2010 return 0;
2011
2012 block_sectors = 1 << rdev->badblocks.shift;
2013 sector = r10_bio->sector;
2014 sectors = ((r10_bio->sector + block_sectors)
2015 & ~(sector_t)(block_sectors - 1))
2016 - sector;
2017
2018 while (sect_to_write) {
2019 struct bio *wbio;
2020 if (sectors > sect_to_write)
2021 sectors = sect_to_write;
2022 /* Write at 'sector' for 'sectors' */
2023 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2024 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2025 wbio->bi_sector = (r10_bio->devs[i].addr+
2026 rdev->data_offset+
2027 (sector - r10_bio->sector));
2028 wbio->bi_bdev = rdev->bdev;
2029 if (submit_bio_wait(WRITE, wbio) == 0)
2030 /* Failure! */
2031 ok = rdev_set_badblocks(rdev, sector,
2032 sectors, 0)
2033 && ok;
2034
2035 bio_put(wbio);
2036 sect_to_write -= sectors;
2037 sector += sectors;
2038 sectors = block_sectors;
2039 }
2040 return ok;
2041}
2042
2043static void handle_read_error(mddev_t *mddev, r10bio_t *r10_bio)
2044{
2045 int slot = r10_bio->read_slot;
2046 int mirror = r10_bio->devs[slot].devnum;
2047 struct bio *bio;
2048 conf_t *conf = mddev->private;
2049 mdk_rdev_t *rdev;
2050 char b[BDEVNAME_SIZE];
2051 unsigned long do_sync;
2052 int max_sectors;
2053
2054 /* we got a read error. Maybe the drive is bad. Maybe just
2055 * the block and we can fix it.
2056 * We freeze all other IO, and try reading the block from
2057 * other devices. When we find one, we re-write
2058 * and check it that fixes the read error.
2059 * This is all done synchronously while the array is
2060 * frozen.
2061 */
2062 if (mddev->ro == 0) {
2063 freeze_array(conf);
2064 fix_read_error(conf, mddev, r10_bio);
2065 unfreeze_array(conf);
2066 }
2067 rdev_dec_pending(conf->mirrors[mirror].rdev, mddev);
2068
2069 bio = r10_bio->devs[slot].bio;
2070 bdevname(bio->bi_bdev, b);
2071 r10_bio->devs[slot].bio =
2072 mddev->ro ? IO_BLOCKED : NULL;
2073read_more:
2074 mirror = read_balance(conf, r10_bio, &max_sectors);
2075 if (mirror == -1) {
2076 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2077 " read error for block %llu\n",
2078 mdname(mddev), b,
2079 (unsigned long long)r10_bio->sector);
2080 raid_end_bio_io(r10_bio);
2081 bio_put(bio);
2082 return;
2083 }
2084
2085 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2086 if (bio)
2087 bio_put(bio);
2088 slot = r10_bio->read_slot;
2089 rdev = conf->mirrors[mirror].rdev;
2090 printk_ratelimited(
2091 KERN_ERR
2092 "md/raid10:%s: %s: redirecting"
2093 "sector %llu to another mirror\n",
2094 mdname(mddev),
2095 bdevname(rdev->bdev, b),
2096 (unsigned long long)r10_bio->sector);
2097 bio = bio_clone_mddev(r10_bio->master_bio,
2098 GFP_NOIO, mddev);
2099 md_trim_bio(bio,
2100 r10_bio->sector - bio->bi_sector,
2101 max_sectors);
2102 r10_bio->devs[slot].bio = bio;
2103 bio->bi_sector = r10_bio->devs[slot].addr
2104 + rdev->data_offset;
2105 bio->bi_bdev = rdev->bdev;
2106 bio->bi_rw = READ | do_sync;
2107 bio->bi_private = r10_bio;
2108 bio->bi_end_io = raid10_end_read_request;
2109 if (max_sectors < r10_bio->sectors) {
2110 /* Drat - have to split this up more */
2111 struct bio *mbio = r10_bio->master_bio;
2112 int sectors_handled =
2113 r10_bio->sector + max_sectors
2114 - mbio->bi_sector;
2115 r10_bio->sectors = max_sectors;
2116 spin_lock_irq(&conf->device_lock);
2117 if (mbio->bi_phys_segments == 0)
2118 mbio->bi_phys_segments = 2;
2119 else
2120 mbio->bi_phys_segments++;
2121 spin_unlock_irq(&conf->device_lock);
2122 generic_make_request(bio);
2123 bio = NULL;
2124
2125 r10_bio = mempool_alloc(conf->r10bio_pool,
2126 GFP_NOIO);
2127 r10_bio->master_bio = mbio;
2128 r10_bio->sectors = (mbio->bi_size >> 9)
2129 - sectors_handled;
2130 r10_bio->state = 0;
2131 set_bit(R10BIO_ReadError,
2132 &r10_bio->state);
2133 r10_bio->mddev = mddev;
2134 r10_bio->sector = mbio->bi_sector
2135 + sectors_handled;
2136
2137 goto read_more;
2138 } else
2139 generic_make_request(bio);
2140}
2141
2142static void handle_write_completed(conf_t *conf, r10bio_t *r10_bio)
2143{
2144 /* Some sort of write request has finished and it
2145 * succeeded in writing where we thought there was a
2146 * bad block. So forget the bad block.
2147 * Or possibly if failed and we need to record
2148 * a bad block.
2149 */
2150 int m;
2151 mdk_rdev_t *rdev;
2152
2153 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2154 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2155 for (m = 0; m < conf->copies; m++) {
2156 int dev = r10_bio->devs[m].devnum;
2157 rdev = conf->mirrors[dev].rdev;
2158 if (r10_bio->devs[m].bio == NULL)
2159 continue;
2160 if (test_bit(BIO_UPTODATE,
2161 &r10_bio->devs[m].bio->bi_flags)) {
2162 rdev_clear_badblocks(
2163 rdev,
2164 r10_bio->devs[m].addr,
2165 r10_bio->sectors);
2166 } else {
2167 if (!rdev_set_badblocks(
2168 rdev,
2169 r10_bio->devs[m].addr,
2170 r10_bio->sectors, 0))
2171 md_error(conf->mddev, rdev);
2172 }
2173 }
2174 put_buf(r10_bio);
2175 } else {
2176 for (m = 0; m < conf->copies; m++) {
2177 int dev = r10_bio->devs[m].devnum;
2178 struct bio *bio = r10_bio->devs[m].bio;
2179 rdev = conf->mirrors[dev].rdev;
2180 if (bio == IO_MADE_GOOD) {
2181 rdev_clear_badblocks(
2182 rdev,
2183 r10_bio->devs[m].addr,
2184 r10_bio->sectors);
2185 rdev_dec_pending(rdev, conf->mddev);
2186 } else if (bio != NULL &&
2187 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2188 if (!narrow_write_error(r10_bio, m)) {
2189 md_error(conf->mddev, rdev);
2190 set_bit(R10BIO_Degraded,
2191 &r10_bio->state);
2192 }
2193 rdev_dec_pending(rdev, conf->mddev);
2194 }
2195 }
2196 if (test_bit(R10BIO_WriteError,
2197 &r10_bio->state))
2198 close_write(r10_bio);
2199 raid_end_bio_io(r10_bio);
2200 }
2201}
2202
2203static void raid10d(mddev_t *mddev)
2204{
2205 r10bio_t *r10_bio;
2206 unsigned long flags;
2207 conf_t *conf = mddev->private;
2208 struct list_head *head = &conf->retry_list;
2209 struct blk_plug plug;
2210
2211 md_check_recovery(mddev);
2212
2213 blk_start_plug(&plug);
2214 for (;;) {
2215
2216 flush_pending_writes(conf);
2217
2218 spin_lock_irqsave(&conf->device_lock, flags);
2219 if (list_empty(head)) {
2220 spin_unlock_irqrestore(&conf->device_lock, flags);
2221 break;
2222 }
2223 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
2224 list_del(head->prev);
2225 conf->nr_queued--;
2226 spin_unlock_irqrestore(&conf->device_lock, flags);
2227
2228 mddev = r10_bio->mddev;
2229 conf = mddev->private;
2230 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2231 test_bit(R10BIO_WriteError, &r10_bio->state))
2232 handle_write_completed(conf, r10_bio);
2233 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2234 sync_request_write(mddev, r10_bio);
2235 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2236 recovery_request_write(mddev, r10_bio);
2237 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2238 handle_read_error(mddev, r10_bio);
2239 else {
2240 /* just a partial read to be scheduled from a
2241 * separate context
2242 */
2243 int slot = r10_bio->read_slot;
2244 generic_make_request(r10_bio->devs[slot].bio);
2245 }
2246
2247 cond_resched();
2248 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2249 md_check_recovery(mddev);
2250 }
2251 blk_finish_plug(&plug);
2252}
2253
2254
2255static int init_resync(conf_t *conf)
2256{
2257 int buffs;
2258
2259 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2260 BUG_ON(conf->r10buf_pool);
2261 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2262 if (!conf->r10buf_pool)
2263 return -ENOMEM;
2264 conf->next_resync = 0;
2265 return 0;
2266}
2267
2268/*
2269 * perform a "sync" on one "block"
2270 *
2271 * We need to make sure that no normal I/O request - particularly write
2272 * requests - conflict with active sync requests.
2273 *
2274 * This is achieved by tracking pending requests and a 'barrier' concept
2275 * that can be installed to exclude normal IO requests.
2276 *
2277 * Resync and recovery are handled very differently.
2278 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2279 *
2280 * For resync, we iterate over virtual addresses, read all copies,
2281 * and update if there are differences. If only one copy is live,
2282 * skip it.
2283 * For recovery, we iterate over physical addresses, read a good
2284 * value for each non-in_sync drive, and over-write.
2285 *
2286 * So, for recovery we may have several outstanding complex requests for a
2287 * given address, one for each out-of-sync device. We model this by allocating
2288 * a number of r10_bio structures, one for each out-of-sync device.
2289 * As we setup these structures, we collect all bio's together into a list
2290 * which we then process collectively to add pages, and then process again
2291 * to pass to generic_make_request.
2292 *
2293 * The r10_bio structures are linked using a borrowed master_bio pointer.
2294 * This link is counted in ->remaining. When the r10_bio that points to NULL
2295 * has its remaining count decremented to 0, the whole complex operation
2296 * is complete.
2297 *
2298 */
2299
2300static sector_t sync_request(mddev_t *mddev, sector_t sector_nr,
2301 int *skipped, int go_faster)
2302{
2303 conf_t *conf = mddev->private;
2304 r10bio_t *r10_bio;
2305 struct bio *biolist = NULL, *bio;
2306 sector_t max_sector, nr_sectors;
2307 int i;
2308 int max_sync;
2309 sector_t sync_blocks;
2310 sector_t sectors_skipped = 0;
2311 int chunks_skipped = 0;
2312
2313 if (!conf->r10buf_pool)
2314 if (init_resync(conf))
2315 return 0;
2316
2317 skipped:
2318 max_sector = mddev->dev_sectors;
2319 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2320 max_sector = mddev->resync_max_sectors;
2321 if (sector_nr >= max_sector) {
2322 /* If we aborted, we need to abort the
2323 * sync on the 'current' bitmap chucks (there can
2324 * be several when recovering multiple devices).
2325 * as we may have started syncing it but not finished.
2326 * We can find the current address in
2327 * mddev->curr_resync, but for recovery,
2328 * we need to convert that to several
2329 * virtual addresses.
2330 */
2331 if (mddev->curr_resync < max_sector) { /* aborted */
2332 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2333 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2334 &sync_blocks, 1);
2335 else for (i=0; i<conf->raid_disks; i++) {
2336 sector_t sect =
2337 raid10_find_virt(conf, mddev->curr_resync, i);
2338 bitmap_end_sync(mddev->bitmap, sect,
2339 &sync_blocks, 1);
2340 }
2341 } else /* completed sync */
2342 conf->fullsync = 0;
2343
2344 bitmap_close_sync(mddev->bitmap);
2345 close_sync(conf);
2346 *skipped = 1;
2347 return sectors_skipped;
2348 }
2349 if (chunks_skipped >= conf->raid_disks) {
2350 /* if there has been nothing to do on any drive,
2351 * then there is nothing to do at all..
2352 */
2353 *skipped = 1;
2354 return (max_sector - sector_nr) + sectors_skipped;
2355 }
2356
2357 if (max_sector > mddev->resync_max)
2358 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2359
2360 /* make sure whole request will fit in a chunk - if chunks
2361 * are meaningful
2362 */
2363 if (conf->near_copies < conf->raid_disks &&
2364 max_sector > (sector_nr | conf->chunk_mask))
2365 max_sector = (sector_nr | conf->chunk_mask) + 1;
2366 /*
2367 * If there is non-resync activity waiting for us then
2368 * put in a delay to throttle resync.
2369 */
2370 if (!go_faster && conf->nr_waiting)
2371 msleep_interruptible(1000);
2372
2373 /* Again, very different code for resync and recovery.
2374 * Both must result in an r10bio with a list of bios that
2375 * have bi_end_io, bi_sector, bi_bdev set,
2376 * and bi_private set to the r10bio.
2377 * For recovery, we may actually create several r10bios
2378 * with 2 bios in each, that correspond to the bios in the main one.
2379 * In this case, the subordinate r10bios link back through a
2380 * borrowed master_bio pointer, and the counter in the master
2381 * includes a ref from each subordinate.
2382 */
2383 /* First, we decide what to do and set ->bi_end_io
2384 * To end_sync_read if we want to read, and
2385 * end_sync_write if we will want to write.
2386 */
2387
2388 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2389 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2390 /* recovery... the complicated one */
2391 int j;
2392 r10_bio = NULL;
2393
2394 for (i=0 ; i<conf->raid_disks; i++) {
2395 int still_degraded;
2396 r10bio_t *rb2;
2397 sector_t sect;
2398 int must_sync;
2399 int any_working;
2400
2401 if (conf->mirrors[i].rdev == NULL ||
2402 test_bit(In_sync, &conf->mirrors[i].rdev->flags))
2403 continue;
2404
2405 still_degraded = 0;
2406 /* want to reconstruct this device */
2407 rb2 = r10_bio;
2408 sect = raid10_find_virt(conf, sector_nr, i);
2409 /* Unless we are doing a full sync, we only need
2410 * to recover the block if it is set in the bitmap
2411 */
2412 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2413 &sync_blocks, 1);
2414 if (sync_blocks < max_sync)
2415 max_sync = sync_blocks;
2416 if (!must_sync &&
2417 !conf->fullsync) {
2418 /* yep, skip the sync_blocks here, but don't assume
2419 * that there will never be anything to do here
2420 */
2421 chunks_skipped = -1;
2422 continue;
2423 }
2424
2425 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2426 raise_barrier(conf, rb2 != NULL);
2427 atomic_set(&r10_bio->remaining, 0);
2428
2429 r10_bio->master_bio = (struct bio*)rb2;
2430 if (rb2)
2431 atomic_inc(&rb2->remaining);
2432 r10_bio->mddev = mddev;
2433 set_bit(R10BIO_IsRecover, &r10_bio->state);
2434 r10_bio->sector = sect;
2435
2436 raid10_find_phys(conf, r10_bio);
2437
2438 /* Need to check if the array will still be
2439 * degraded
2440 */
2441 for (j=0; j<conf->raid_disks; j++)
2442 if (conf->mirrors[j].rdev == NULL ||
2443 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
2444 still_degraded = 1;
2445 break;
2446 }
2447
2448 must_sync = bitmap_start_sync(mddev->bitmap, sect,
2449 &sync_blocks, still_degraded);
2450
2451 any_working = 0;
2452 for (j=0; j<conf->copies;j++) {
2453 int k;
2454 int d = r10_bio->devs[j].devnum;
2455 sector_t from_addr, to_addr;
2456 mdk_rdev_t *rdev;
2457 sector_t sector, first_bad;
2458 int bad_sectors;
2459 if (!conf->mirrors[d].rdev ||
2460 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
2461 continue;
2462 /* This is where we read from */
2463 any_working = 1;
2464 rdev = conf->mirrors[d].rdev;
2465 sector = r10_bio->devs[j].addr;
2466
2467 if (is_badblock(rdev, sector, max_sync,
2468 &first_bad, &bad_sectors)) {
2469 if (first_bad > sector)
2470 max_sync = first_bad - sector;
2471 else {
2472 bad_sectors -= (sector
2473 - first_bad);
2474 if (max_sync > bad_sectors)
2475 max_sync = bad_sectors;
2476 continue;
2477 }
2478 }
2479 bio = r10_bio->devs[0].bio;
2480 bio->bi_next = biolist;
2481 biolist = bio;
2482 bio->bi_private = r10_bio;
2483 bio->bi_end_io = end_sync_read;
2484 bio->bi_rw = READ;
2485 from_addr = r10_bio->devs[j].addr;
2486 bio->bi_sector = from_addr +
2487 conf->mirrors[d].rdev->data_offset;
2488 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2489 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2490 atomic_inc(&r10_bio->remaining);
2491 /* and we write to 'i' */
2492
2493 for (k=0; k<conf->copies; k++)
2494 if (r10_bio->devs[k].devnum == i)
2495 break;
2496 BUG_ON(k == conf->copies);
2497 bio = r10_bio->devs[1].bio;
2498 bio->bi_next = biolist;
2499 biolist = bio;
2500 bio->bi_private = r10_bio;
2501 bio->bi_end_io = end_sync_write;
2502 bio->bi_rw = WRITE;
2503 to_addr = r10_bio->devs[k].addr;
2504 bio->bi_sector = to_addr +
2505 conf->mirrors[i].rdev->data_offset;
2506 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
2507
2508 r10_bio->devs[0].devnum = d;
2509 r10_bio->devs[0].addr = from_addr;
2510 r10_bio->devs[1].devnum = i;
2511 r10_bio->devs[1].addr = to_addr;
2512
2513 break;
2514 }
2515 if (j == conf->copies) {
2516 /* Cannot recover, so abort the recovery or
2517 * record a bad block */
2518 put_buf(r10_bio);
2519 if (rb2)
2520 atomic_dec(&rb2->remaining);
2521 r10_bio = rb2;
2522 if (any_working) {
2523 /* problem is that there are bad blocks
2524 * on other device(s)
2525 */
2526 int k;
2527 for (k = 0; k < conf->copies; k++)
2528 if (r10_bio->devs[k].devnum == i)
2529 break;
2530 if (!rdev_set_badblocks(
2531 conf->mirrors[i].rdev,
2532 r10_bio->devs[k].addr,
2533 max_sync, 0))
2534 any_working = 0;
2535 }
2536 if (!any_working) {
2537 if (!test_and_set_bit(MD_RECOVERY_INTR,
2538 &mddev->recovery))
2539 printk(KERN_INFO "md/raid10:%s: insufficient "
2540 "working devices for recovery.\n",
2541 mdname(mddev));
2542 conf->mirrors[i].recovery_disabled
2543 = mddev->recovery_disabled;
2544 }
2545 break;
2546 }
2547 }
2548 if (biolist == NULL) {
2549 while (r10_bio) {
2550 r10bio_t *rb2 = r10_bio;
2551 r10_bio = (r10bio_t*) rb2->master_bio;
2552 rb2->master_bio = NULL;
2553 put_buf(rb2);
2554 }
2555 goto giveup;
2556 }
2557 } else {
2558 /* resync. Schedule a read for every block at this virt offset */
2559 int count = 0;
2560
2561 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2562
2563 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2564 &sync_blocks, mddev->degraded) &&
2565 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
2566 &mddev->recovery)) {
2567 /* We can skip this block */
2568 *skipped = 1;
2569 return sync_blocks + sectors_skipped;
2570 }
2571 if (sync_blocks < max_sync)
2572 max_sync = sync_blocks;
2573 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2574
2575 r10_bio->mddev = mddev;
2576 atomic_set(&r10_bio->remaining, 0);
2577 raise_barrier(conf, 0);
2578 conf->next_resync = sector_nr;
2579
2580 r10_bio->master_bio = NULL;
2581 r10_bio->sector = sector_nr;
2582 set_bit(R10BIO_IsSync, &r10_bio->state);
2583 raid10_find_phys(conf, r10_bio);
2584 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2585
2586 for (i=0; i<conf->copies; i++) {
2587 int d = r10_bio->devs[i].devnum;
2588 sector_t first_bad, sector;
2589 int bad_sectors;
2590
2591 bio = r10_bio->devs[i].bio;
2592 bio->bi_end_io = NULL;
2593 clear_bit(BIO_UPTODATE, &bio->bi_flags);
2594 if (conf->mirrors[d].rdev == NULL ||
2595 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
2596 continue;
2597 sector = r10_bio->devs[i].addr;
2598 if (is_badblock(conf->mirrors[d].rdev,
2599 sector, max_sync,
2600 &first_bad, &bad_sectors)) {
2601 if (first_bad > sector)
2602 max_sync = first_bad - sector;
2603 else {
2604 bad_sectors -= (sector - first_bad);
2605 if (max_sync > bad_sectors)
2606 max_sync = max_sync;
2607 continue;
2608 }
2609 }
2610 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2611 atomic_inc(&r10_bio->remaining);
2612 bio->bi_next = biolist;
2613 biolist = bio;
2614 bio->bi_private = r10_bio;
2615 bio->bi_end_io = end_sync_read;
2616 bio->bi_rw = READ;
2617 bio->bi_sector = sector +
2618 conf->mirrors[d].rdev->data_offset;
2619 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2620 count++;
2621 }
2622
2623 if (count < 2) {
2624 for (i=0; i<conf->copies; i++) {
2625 int d = r10_bio->devs[i].devnum;
2626 if (r10_bio->devs[i].bio->bi_end_io)
2627 rdev_dec_pending(conf->mirrors[d].rdev,
2628 mddev);
2629 }
2630 put_buf(r10_bio);
2631 biolist = NULL;
2632 goto giveup;
2633 }
2634 }
2635
2636 for (bio = biolist; bio ; bio=bio->bi_next) {
2637
2638 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2639 if (bio->bi_end_io)
2640 bio->bi_flags |= 1 << BIO_UPTODATE;
2641 bio->bi_vcnt = 0;
2642 bio->bi_idx = 0;
2643 bio->bi_phys_segments = 0;
2644 bio->bi_size = 0;
2645 }
2646
2647 nr_sectors = 0;
2648 if (sector_nr + max_sync < max_sector)
2649 max_sector = sector_nr + max_sync;
2650 do {
2651 struct page *page;
2652 int len = PAGE_SIZE;
2653 if (sector_nr + (len>>9) > max_sector)
2654 len = (max_sector - sector_nr) << 9;
2655 if (len == 0)
2656 break;
2657 for (bio= biolist ; bio ; bio=bio->bi_next) {
2658 struct bio *bio2;
2659 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2660 if (bio_add_page(bio, page, len, 0))
2661 continue;
2662
2663 /* stop here */
2664 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2665 for (bio2 = biolist;
2666 bio2 && bio2 != bio;
2667 bio2 = bio2->bi_next) {
2668 /* remove last page from this bio */
2669 bio2->bi_vcnt--;
2670 bio2->bi_size -= len;
2671 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2672 }
2673 goto bio_full;
2674 }
2675 nr_sectors += len>>9;
2676 sector_nr += len>>9;
2677 } while (biolist->bi_vcnt < RESYNC_PAGES);
2678 bio_full:
2679 r10_bio->sectors = nr_sectors;
2680
2681 while (biolist) {
2682 bio = biolist;
2683 biolist = biolist->bi_next;
2684
2685 bio->bi_next = NULL;
2686 r10_bio = bio->bi_private;
2687 r10_bio->sectors = nr_sectors;
2688
2689 if (bio->bi_end_io == end_sync_read) {
2690 md_sync_acct(bio->bi_bdev, nr_sectors);
2691 generic_make_request(bio);
2692 }
2693 }
2694
2695 if (sectors_skipped)
2696 /* pretend they weren't skipped, it makes
2697 * no important difference in this case
2698 */
2699 md_done_sync(mddev, sectors_skipped, 1);
2700
2701 return sectors_skipped + nr_sectors;
2702 giveup:
2703 /* There is nowhere to write, so all non-sync
2704 * drives must be failed or in resync, all drives
2705 * have a bad block, so try the next chunk...
2706 */
2707 if (sector_nr + max_sync < max_sector)
2708 max_sector = sector_nr + max_sync;
2709
2710 sectors_skipped += (max_sector - sector_nr);
2711 chunks_skipped ++;
2712 sector_nr = max_sector;
2713 goto skipped;
2714}
2715
2716static sector_t
2717raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2718{
2719 sector_t size;
2720 conf_t *conf = mddev->private;
2721
2722 if (!raid_disks)
2723 raid_disks = conf->raid_disks;
2724 if (!sectors)
2725 sectors = conf->dev_sectors;
2726
2727 size = sectors >> conf->chunk_shift;
2728 sector_div(size, conf->far_copies);
2729 size = size * raid_disks;
2730 sector_div(size, conf->near_copies);
2731
2732 return size << conf->chunk_shift;
2733}
2734
2735
2736static conf_t *setup_conf(mddev_t *mddev)
2737{
2738 conf_t *conf = NULL;
2739 int nc, fc, fo;
2740 sector_t stride, size;
2741 int err = -EINVAL;
2742
2743 if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2744 !is_power_of_2(mddev->new_chunk_sectors)) {
2745 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2746 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2747 mdname(mddev), PAGE_SIZE);
2748 goto out;
2749 }
2750
2751 nc = mddev->new_layout & 255;
2752 fc = (mddev->new_layout >> 8) & 255;
2753 fo = mddev->new_layout & (1<<16);
2754
2755 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2756 (mddev->new_layout >> 17)) {
2757 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
2758 mdname(mddev), mddev->new_layout);
2759 goto out;
2760 }
2761
2762 err = -ENOMEM;
2763 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2764 if (!conf)
2765 goto out;
2766
2767 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2768 GFP_KERNEL);
2769 if (!conf->mirrors)
2770 goto out;
2771
2772 conf->tmppage = alloc_page(GFP_KERNEL);
2773 if (!conf->tmppage)
2774 goto out;
2775
2776
2777 conf->raid_disks = mddev->raid_disks;
2778 conf->near_copies = nc;
2779 conf->far_copies = fc;
2780 conf->copies = nc*fc;
2781 conf->far_offset = fo;
2782 conf->chunk_mask = mddev->new_chunk_sectors - 1;
2783 conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2784
2785 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2786 r10bio_pool_free, conf);
2787 if (!conf->r10bio_pool)
2788 goto out;
2789
2790 size = mddev->dev_sectors >> conf->chunk_shift;
2791 sector_div(size, fc);
2792 size = size * conf->raid_disks;
2793 sector_div(size, nc);
2794 /* 'size' is now the number of chunks in the array */
2795 /* calculate "used chunks per device" in 'stride' */
2796 stride = size * conf->copies;
2797
2798 /* We need to round up when dividing by raid_disks to
2799 * get the stride size.
2800 */
2801 stride += conf->raid_disks - 1;
2802 sector_div(stride, conf->raid_disks);
2803
2804 conf->dev_sectors = stride << conf->chunk_shift;
2805
2806 if (fo)
2807 stride = 1;
2808 else
2809 sector_div(stride, fc);
2810 conf->stride = stride << conf->chunk_shift;
2811
2812
2813 spin_lock_init(&conf->device_lock);
2814 INIT_LIST_HEAD(&conf->retry_list);
2815
2816 spin_lock_init(&conf->resync_lock);
2817 init_waitqueue_head(&conf->wait_barrier);
2818
2819 conf->thread = md_register_thread(raid10d, mddev, NULL);
2820 if (!conf->thread)
2821 goto out;
2822
2823 conf->mddev = mddev;
2824 return conf;
2825
2826 out:
2827 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
2828 mdname(mddev));
2829 if (conf) {
2830 if (conf->r10bio_pool)
2831 mempool_destroy(conf->r10bio_pool);
2832 kfree(conf->mirrors);
2833 safe_put_page(conf->tmppage);
2834 kfree(conf);
2835 }
2836 return ERR_PTR(err);
2837}
2838
2839static int run(mddev_t *mddev)
2840{
2841 conf_t *conf;
2842 int i, disk_idx, chunk_size;
2843 mirror_info_t *disk;
2844 mdk_rdev_t *rdev;
2845 sector_t size;
2846
2847 /*
2848 * copy the already verified devices into our private RAID10
2849 * bookkeeping area. [whatever we allocate in run(),
2850 * should be freed in stop()]
2851 */
2852
2853 if (mddev->private == NULL) {
2854 conf = setup_conf(mddev);
2855 if (IS_ERR(conf))
2856 return PTR_ERR(conf);
2857 mddev->private = conf;
2858 }
2859 conf = mddev->private;
2860 if (!conf)
2861 goto out;
2862
2863 mddev->thread = conf->thread;
2864 conf->thread = NULL;
2865
2866 chunk_size = mddev->chunk_sectors << 9;
2867 blk_queue_io_min(mddev->queue, chunk_size);
2868 if (conf->raid_disks % conf->near_copies)
2869 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2870 else
2871 blk_queue_io_opt(mddev->queue, chunk_size *
2872 (conf->raid_disks / conf->near_copies));
2873
2874 list_for_each_entry(rdev, &mddev->disks, same_set) {
2875
2876 disk_idx = rdev->raid_disk;
2877 if (disk_idx >= conf->raid_disks
2878 || disk_idx < 0)
2879 continue;
2880 disk = conf->mirrors + disk_idx;
2881
2882 disk->rdev = rdev;
2883 disk_stack_limits(mddev->gendisk, rdev->bdev,
2884 rdev->data_offset << 9);
2885 /* as we don't honour merge_bvec_fn, we must never risk
2886 * violating it, so limit max_segments to 1 lying
2887 * within a single page.
2888 */
2889 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2890 blk_queue_max_segments(mddev->queue, 1);
2891 blk_queue_segment_boundary(mddev->queue,
2892 PAGE_CACHE_SIZE - 1);
2893 }
2894
2895 disk->head_position = 0;
2896 }
2897 /* need to check that every block has at least one working mirror */
2898 if (!enough(conf, -1)) {
2899 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
2900 mdname(mddev));
2901 goto out_free_conf;
2902 }
2903
2904 mddev->degraded = 0;
2905 for (i = 0; i < conf->raid_disks; i++) {
2906
2907 disk = conf->mirrors + i;
2908
2909 if (!disk->rdev ||
2910 !test_bit(In_sync, &disk->rdev->flags)) {
2911 disk->head_position = 0;
2912 mddev->degraded++;
2913 if (disk->rdev)
2914 conf->fullsync = 1;
2915 }
2916 }
2917
2918 if (mddev->recovery_cp != MaxSector)
2919 printk(KERN_NOTICE "md/raid10:%s: not clean"
2920 " -- starting background reconstruction\n",
2921 mdname(mddev));
2922 printk(KERN_INFO
2923 "md/raid10:%s: active with %d out of %d devices\n",
2924 mdname(mddev), conf->raid_disks - mddev->degraded,
2925 conf->raid_disks);
2926 /*
2927 * Ok, everything is just fine now
2928 */
2929 mddev->dev_sectors = conf->dev_sectors;
2930 size = raid10_size(mddev, 0, 0);
2931 md_set_array_sectors(mddev, size);
2932 mddev->resync_max_sectors = size;
2933
2934 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2935 mddev->queue->backing_dev_info.congested_data = mddev;
2936
2937 /* Calculate max read-ahead size.
2938 * We need to readahead at least twice a whole stripe....
2939 * maybe...
2940 */
2941 {
2942 int stripe = conf->raid_disks *
2943 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2944 stripe /= conf->near_copies;
2945 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2946 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2947 }
2948
2949 if (conf->near_copies < conf->raid_disks)
2950 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2951
2952 if (md_integrity_register(mddev))
2953 goto out_free_conf;
2954
2955 return 0;
2956
2957out_free_conf:
2958 md_unregister_thread(&mddev->thread);
2959 if (conf->r10bio_pool)
2960 mempool_destroy(conf->r10bio_pool);
2961 safe_put_page(conf->tmppage);
2962 kfree(conf->mirrors);
2963 kfree(conf);
2964 mddev->private = NULL;
2965out:
2966 return -EIO;
2967}
2968
2969static int stop(mddev_t *mddev)
2970{
2971 conf_t *conf = mddev->private;
2972
2973 raise_barrier(conf, 0);
2974 lower_barrier(conf);
2975
2976 md_unregister_thread(&mddev->thread);
2977 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2978 if (conf->r10bio_pool)
2979 mempool_destroy(conf->r10bio_pool);
2980 kfree(conf->mirrors);
2981 kfree(conf);
2982 mddev->private = NULL;
2983 return 0;
2984}
2985
2986static void raid10_quiesce(mddev_t *mddev, int state)
2987{
2988 conf_t *conf = mddev->private;
2989
2990 switch(state) {
2991 case 1:
2992 raise_barrier(conf, 0);
2993 break;
2994 case 0:
2995 lower_barrier(conf);
2996 break;
2997 }
2998}
2999
3000static void *raid10_takeover_raid0(mddev_t *mddev)
3001{
3002 mdk_rdev_t *rdev;
3003 conf_t *conf;
3004
3005 if (mddev->degraded > 0) {
3006 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3007 mdname(mddev));
3008 return ERR_PTR(-EINVAL);
3009 }
3010
3011 /* Set new parameters */
3012 mddev->new_level = 10;
3013 /* new layout: far_copies = 1, near_copies = 2 */
3014 mddev->new_layout = (1<<8) + 2;
3015 mddev->new_chunk_sectors = mddev->chunk_sectors;
3016 mddev->delta_disks = mddev->raid_disks;
3017 mddev->raid_disks *= 2;
3018 /* make sure it will be not marked as dirty */
3019 mddev->recovery_cp = MaxSector;
3020
3021 conf = setup_conf(mddev);
3022 if (!IS_ERR(conf)) {
3023 list_for_each_entry(rdev, &mddev->disks, same_set)
3024 if (rdev->raid_disk >= 0)
3025 rdev->new_raid_disk = rdev->raid_disk * 2;
3026 conf->barrier = 1;
3027 }
3028
3029 return conf;
3030}
3031
3032static void *raid10_takeover(mddev_t *mddev)
3033{
3034 struct raid0_private_data *raid0_priv;
3035
3036 /* raid10 can take over:
3037 * raid0 - providing it has only two drives
3038 */
3039 if (mddev->level == 0) {
3040 /* for raid0 takeover only one zone is supported */
3041 raid0_priv = mddev->private;
3042 if (raid0_priv->nr_strip_zones > 1) {
3043 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3044 " with more than one zone.\n",
3045 mdname(mddev));
3046 return ERR_PTR(-EINVAL);
3047 }
3048 return raid10_takeover_raid0(mddev);
3049 }
3050 return ERR_PTR(-EINVAL);
3051}
3052
3053static struct mdk_personality raid10_personality =
3054{
3055 .name = "raid10",
3056 .level = 10,
3057 .owner = THIS_MODULE,
3058 .make_request = make_request,
3059 .run = run,
3060 .stop = stop,
3061 .status = status,
3062 .error_handler = error,
3063 .hot_add_disk = raid10_add_disk,
3064 .hot_remove_disk= raid10_remove_disk,
3065 .spare_active = raid10_spare_active,
3066 .sync_request = sync_request,
3067 .quiesce = raid10_quiesce,
3068 .size = raid10_size,
3069 .takeover = raid10_takeover,
3070};
3071
3072static int __init raid_init(void)
3073{
3074 return register_md_personality(&raid10_personality);
3075}
3076
3077static void raid_exit(void)
3078{
3079 unregister_md_personality(&raid10_personality);
3080}
3081
3082module_init(raid_init);
3083module_exit(raid_exit);
3084MODULE_LICENSE("GPL");
3085MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
3086MODULE_ALIAS("md-personality-9"); /* RAID10 */
3087MODULE_ALIAS("md-raid10");
3088MODULE_ALIAS("md-level-10");
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#define NULL_CMD
83#define cmd_before(conf, cmd) \
84 do { \
85 write_sequnlock_irq(&(conf)->resync_lock); \
86 cmd; \
87 } while (0)
88#define cmd_after(conf) write_seqlock_irq(&(conf)->resync_lock)
89
90#define wait_event_barrier_cmd(conf, cond, cmd) \
91 wait_event_cmd((conf)->wait_barrier, cond, cmd_before(conf, cmd), \
92 cmd_after(conf))
93
94#define wait_event_barrier(conf, cond) \
95 wait_event_barrier_cmd(conf, cond, NULL_CMD)
96
97/*
98 * for resync bio, r10bio pointer can be retrieved from the per-bio
99 * 'struct resync_pages'.
100 */
101static inline struct r10bio *get_resync_r10bio(struct bio *bio)
102{
103 return get_resync_pages(bio)->raid_bio;
104}
105
106static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
107{
108 struct r10conf *conf = data;
109 int size = offsetof(struct r10bio, devs[conf->geo.raid_disks]);
110
111 /* allocate a r10bio with room for raid_disks entries in the
112 * bios array */
113 return kzalloc(size, gfp_flags);
114}
115
116#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
117/* amount of memory to reserve for resync requests */
118#define RESYNC_WINDOW (1024*1024)
119/* maximum number of concurrent requests, memory permitting */
120#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
121#define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
122#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
123
124/*
125 * When performing a resync, we need to read and compare, so
126 * we need as many pages are there are copies.
127 * When performing a recovery, we need 2 bios, one for read,
128 * one for write (we recover only one drive per r10buf)
129 *
130 */
131static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
132{
133 struct r10conf *conf = data;
134 struct r10bio *r10_bio;
135 struct bio *bio;
136 int j;
137 int nalloc, nalloc_rp;
138 struct resync_pages *rps;
139
140 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
141 if (!r10_bio)
142 return NULL;
143
144 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
145 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
146 nalloc = conf->copies; /* resync */
147 else
148 nalloc = 2; /* recovery */
149
150 /* allocate once for all bios */
151 if (!conf->have_replacement)
152 nalloc_rp = nalloc;
153 else
154 nalloc_rp = nalloc * 2;
155 rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
156 if (!rps)
157 goto out_free_r10bio;
158
159 /*
160 * Allocate bios.
161 */
162 for (j = nalloc ; j-- ; ) {
163 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
164 if (!bio)
165 goto out_free_bio;
166 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
167 r10_bio->devs[j].bio = bio;
168 if (!conf->have_replacement)
169 continue;
170 bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
171 if (!bio)
172 goto out_free_bio;
173 bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
174 r10_bio->devs[j].repl_bio = bio;
175 }
176 /*
177 * Allocate RESYNC_PAGES data pages and attach them
178 * where needed.
179 */
180 for (j = 0; j < nalloc; j++) {
181 struct bio *rbio = r10_bio->devs[j].repl_bio;
182 struct resync_pages *rp, *rp_repl;
183
184 rp = &rps[j];
185 if (rbio)
186 rp_repl = &rps[nalloc + j];
187
188 bio = r10_bio->devs[j].bio;
189
190 if (!j || test_bit(MD_RECOVERY_SYNC,
191 &conf->mddev->recovery)) {
192 if (resync_alloc_pages(rp, gfp_flags))
193 goto out_free_pages;
194 } else {
195 memcpy(rp, &rps[0], sizeof(*rp));
196 resync_get_all_pages(rp);
197 }
198
199 rp->raid_bio = r10_bio;
200 bio->bi_private = rp;
201 if (rbio) {
202 memcpy(rp_repl, rp, sizeof(*rp));
203 rbio->bi_private = rp_repl;
204 }
205 }
206
207 return r10_bio;
208
209out_free_pages:
210 while (--j >= 0)
211 resync_free_pages(&rps[j]);
212
213 j = 0;
214out_free_bio:
215 for ( ; j < nalloc; j++) {
216 if (r10_bio->devs[j].bio)
217 bio_uninit(r10_bio->devs[j].bio);
218 kfree(r10_bio->devs[j].bio);
219 if (r10_bio->devs[j].repl_bio)
220 bio_uninit(r10_bio->devs[j].repl_bio);
221 kfree(r10_bio->devs[j].repl_bio);
222 }
223 kfree(rps);
224out_free_r10bio:
225 rbio_pool_free(r10_bio, conf);
226 return NULL;
227}
228
229static void r10buf_pool_free(void *__r10_bio, void *data)
230{
231 struct r10conf *conf = data;
232 struct r10bio *r10bio = __r10_bio;
233 int j;
234 struct resync_pages *rp = NULL;
235
236 for (j = conf->copies; j--; ) {
237 struct bio *bio = r10bio->devs[j].bio;
238
239 if (bio) {
240 rp = get_resync_pages(bio);
241 resync_free_pages(rp);
242 bio_uninit(bio);
243 kfree(bio);
244 }
245
246 bio = r10bio->devs[j].repl_bio;
247 if (bio) {
248 bio_uninit(bio);
249 kfree(bio);
250 }
251 }
252
253 /* resync pages array stored in the 1st bio's .bi_private */
254 kfree(rp);
255
256 rbio_pool_free(r10bio, conf);
257}
258
259static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
260{
261 int i;
262
263 for (i = 0; i < conf->geo.raid_disks; i++) {
264 struct bio **bio = & r10_bio->devs[i].bio;
265 if (!BIO_SPECIAL(*bio))
266 bio_put(*bio);
267 *bio = NULL;
268 bio = &r10_bio->devs[i].repl_bio;
269 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
270 bio_put(*bio);
271 *bio = NULL;
272 }
273}
274
275static void free_r10bio(struct r10bio *r10_bio)
276{
277 struct r10conf *conf = r10_bio->mddev->private;
278
279 put_all_bios(conf, r10_bio);
280 mempool_free(r10_bio, &conf->r10bio_pool);
281}
282
283static void put_buf(struct r10bio *r10_bio)
284{
285 struct r10conf *conf = r10_bio->mddev->private;
286
287 mempool_free(r10_bio, &conf->r10buf_pool);
288
289 lower_barrier(conf);
290}
291
292static void wake_up_barrier(struct r10conf *conf)
293{
294 if (wq_has_sleeper(&conf->wait_barrier))
295 wake_up(&conf->wait_barrier);
296}
297
298static void reschedule_retry(struct r10bio *r10_bio)
299{
300 unsigned long flags;
301 struct mddev *mddev = r10_bio->mddev;
302 struct r10conf *conf = mddev->private;
303
304 spin_lock_irqsave(&conf->device_lock, flags);
305 list_add(&r10_bio->retry_list, &conf->retry_list);
306 conf->nr_queued ++;
307 spin_unlock_irqrestore(&conf->device_lock, flags);
308
309 /* wake up frozen array... */
310 wake_up(&conf->wait_barrier);
311
312 md_wakeup_thread(mddev->thread);
313}
314
315/*
316 * raid_end_bio_io() is called when we have finished servicing a mirrored
317 * operation and are ready to return a success/failure code to the buffer
318 * cache layer.
319 */
320static void raid_end_bio_io(struct r10bio *r10_bio)
321{
322 struct bio *bio = r10_bio->master_bio;
323 struct r10conf *conf = r10_bio->mddev->private;
324
325 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
326 bio->bi_status = BLK_STS_IOERR;
327
328 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
329 bio_end_io_acct(bio, r10_bio->start_time);
330 bio_endio(bio);
331 /*
332 * Wake up any possible resync thread that waits for the device
333 * to go idle.
334 */
335 allow_barrier(conf);
336
337 free_r10bio(r10_bio);
338}
339
340/*
341 * Update disk head position estimator based on IRQ completion info.
342 */
343static inline void update_head_pos(int slot, struct r10bio *r10_bio)
344{
345 struct r10conf *conf = r10_bio->mddev->private;
346
347 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
348 r10_bio->devs[slot].addr + (r10_bio->sectors);
349}
350
351/*
352 * Find the disk number which triggered given bio
353 */
354static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
355 struct bio *bio, int *slotp, int *replp)
356{
357 int slot;
358 int repl = 0;
359
360 for (slot = 0; slot < conf->geo.raid_disks; slot++) {
361 if (r10_bio->devs[slot].bio == bio)
362 break;
363 if (r10_bio->devs[slot].repl_bio == bio) {
364 repl = 1;
365 break;
366 }
367 }
368
369 update_head_pos(slot, r10_bio);
370
371 if (slotp)
372 *slotp = slot;
373 if (replp)
374 *replp = repl;
375 return r10_bio->devs[slot].devnum;
376}
377
378static void raid10_end_read_request(struct bio *bio)
379{
380 int uptodate = !bio->bi_status;
381 struct r10bio *r10_bio = bio->bi_private;
382 int slot;
383 struct md_rdev *rdev;
384 struct r10conf *conf = r10_bio->mddev->private;
385
386 slot = r10_bio->read_slot;
387 rdev = r10_bio->devs[slot].rdev;
388 /*
389 * this branch is our 'one mirror IO has finished' event handler:
390 */
391 update_head_pos(slot, r10_bio);
392
393 if (uptodate) {
394 /*
395 * Set R10BIO_Uptodate in our master bio, so that
396 * we will return a good error code to the higher
397 * levels even if IO on some other mirrored buffer fails.
398 *
399 * The 'master' represents the composite IO operation to
400 * user-side. So if something waits for IO, then it will
401 * wait for the 'master' bio.
402 */
403 set_bit(R10BIO_Uptodate, &r10_bio->state);
404 } else {
405 /* If all other devices that store this block have
406 * failed, we want to return the error upwards rather
407 * than fail the last device. Here we redefine
408 * "uptodate" to mean "Don't want to retry"
409 */
410 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
411 rdev->raid_disk))
412 uptodate = 1;
413 }
414 if (uptodate) {
415 raid_end_bio_io(r10_bio);
416 rdev_dec_pending(rdev, conf->mddev);
417 } else {
418 /*
419 * oops, read error - keep the refcount on the rdev
420 */
421 pr_err_ratelimited("md/raid10:%s: %pg: rescheduling sector %llu\n",
422 mdname(conf->mddev),
423 rdev->bdev,
424 (unsigned long long)r10_bio->sector);
425 set_bit(R10BIO_ReadError, &r10_bio->state);
426 reschedule_retry(r10_bio);
427 }
428}
429
430static void close_write(struct r10bio *r10_bio)
431{
432 /* clear the bitmap if all writes complete successfully */
433 md_bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
434 r10_bio->sectors,
435 !test_bit(R10BIO_Degraded, &r10_bio->state),
436 0);
437 md_write_end(r10_bio->mddev);
438}
439
440static void one_write_done(struct r10bio *r10_bio)
441{
442 if (atomic_dec_and_test(&r10_bio->remaining)) {
443 if (test_bit(R10BIO_WriteError, &r10_bio->state))
444 reschedule_retry(r10_bio);
445 else {
446 close_write(r10_bio);
447 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
448 reschedule_retry(r10_bio);
449 else
450 raid_end_bio_io(r10_bio);
451 }
452 }
453}
454
455static void raid10_end_write_request(struct bio *bio)
456{
457 struct r10bio *r10_bio = bio->bi_private;
458 int dev;
459 int dec_rdev = 1;
460 struct r10conf *conf = r10_bio->mddev->private;
461 int slot, repl;
462 struct md_rdev *rdev = NULL;
463 struct bio *to_put = NULL;
464 bool discard_error;
465
466 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
467
468 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
469
470 if (repl)
471 rdev = conf->mirrors[dev].replacement;
472 if (!rdev) {
473 smp_rmb();
474 repl = 0;
475 rdev = conf->mirrors[dev].rdev;
476 }
477 /*
478 * this branch is our 'one mirror IO has finished' event handler:
479 */
480 if (bio->bi_status && !discard_error) {
481 if (repl)
482 /* Never record new bad blocks to replacement,
483 * just fail it.
484 */
485 md_error(rdev->mddev, rdev);
486 else {
487 set_bit(WriteErrorSeen, &rdev->flags);
488 if (!test_and_set_bit(WantReplacement, &rdev->flags))
489 set_bit(MD_RECOVERY_NEEDED,
490 &rdev->mddev->recovery);
491
492 dec_rdev = 0;
493 if (test_bit(FailFast, &rdev->flags) &&
494 (bio->bi_opf & MD_FAILFAST)) {
495 md_error(rdev->mddev, rdev);
496 }
497
498 /*
499 * When the device is faulty, it is not necessary to
500 * handle write error.
501 */
502 if (!test_bit(Faulty, &rdev->flags))
503 set_bit(R10BIO_WriteError, &r10_bio->state);
504 else {
505 /* Fail the request */
506 set_bit(R10BIO_Degraded, &r10_bio->state);
507 r10_bio->devs[slot].bio = NULL;
508 to_put = bio;
509 dec_rdev = 1;
510 }
511 }
512 } else {
513 /*
514 * Set R10BIO_Uptodate in our master bio, so that
515 * we will return a good error code for to the higher
516 * levels even if IO on some other mirrored buffer fails.
517 *
518 * The 'master' represents the composite IO operation to
519 * user-side. So if something waits for IO, then it will
520 * wait for the 'master' bio.
521 */
522 sector_t first_bad;
523 int bad_sectors;
524
525 /*
526 * Do not set R10BIO_Uptodate if the current device is
527 * rebuilding or Faulty. This is because we cannot use
528 * such device for properly reading the data back (we could
529 * potentially use it, if the current write would have felt
530 * before rdev->recovery_offset, but for simplicity we don't
531 * check this here.
532 */
533 if (test_bit(In_sync, &rdev->flags) &&
534 !test_bit(Faulty, &rdev->flags))
535 set_bit(R10BIO_Uptodate, &r10_bio->state);
536
537 /* Maybe we can clear some bad blocks. */
538 if (is_badblock(rdev,
539 r10_bio->devs[slot].addr,
540 r10_bio->sectors,
541 &first_bad, &bad_sectors) && !discard_error) {
542 bio_put(bio);
543 if (repl)
544 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
545 else
546 r10_bio->devs[slot].bio = IO_MADE_GOOD;
547 dec_rdev = 0;
548 set_bit(R10BIO_MadeGood, &r10_bio->state);
549 }
550 }
551
552 /*
553 *
554 * Let's see if all mirrored write operations have finished
555 * already.
556 */
557 one_write_done(r10_bio);
558 if (dec_rdev)
559 rdev_dec_pending(rdev, conf->mddev);
560 if (to_put)
561 bio_put(to_put);
562}
563
564/*
565 * RAID10 layout manager
566 * As well as the chunksize and raid_disks count, there are two
567 * parameters: near_copies and far_copies.
568 * near_copies * far_copies must be <= raid_disks.
569 * Normally one of these will be 1.
570 * If both are 1, we get raid0.
571 * If near_copies == raid_disks, we get raid1.
572 *
573 * Chunks are laid out in raid0 style with near_copies copies of the
574 * first chunk, followed by near_copies copies of the next chunk and
575 * so on.
576 * If far_copies > 1, then after 1/far_copies of the array has been assigned
577 * as described above, we start again with a device offset of near_copies.
578 * So we effectively have another copy of the whole array further down all
579 * the drives, but with blocks on different drives.
580 * With this layout, and block is never stored twice on the one device.
581 *
582 * raid10_find_phys finds the sector offset of a given virtual sector
583 * on each device that it is on.
584 *
585 * raid10_find_virt does the reverse mapping, from a device and a
586 * sector offset to a virtual address
587 */
588
589static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
590{
591 int n,f;
592 sector_t sector;
593 sector_t chunk;
594 sector_t stripe;
595 int dev;
596 int slot = 0;
597 int last_far_set_start, last_far_set_size;
598
599 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
600 last_far_set_start *= geo->far_set_size;
601
602 last_far_set_size = geo->far_set_size;
603 last_far_set_size += (geo->raid_disks % geo->far_set_size);
604
605 /* now calculate first sector/dev */
606 chunk = r10bio->sector >> geo->chunk_shift;
607 sector = r10bio->sector & geo->chunk_mask;
608
609 chunk *= geo->near_copies;
610 stripe = chunk;
611 dev = sector_div(stripe, geo->raid_disks);
612 if (geo->far_offset)
613 stripe *= geo->far_copies;
614
615 sector += stripe << geo->chunk_shift;
616
617 /* and calculate all the others */
618 for (n = 0; n < geo->near_copies; n++) {
619 int d = dev;
620 int set;
621 sector_t s = sector;
622 r10bio->devs[slot].devnum = d;
623 r10bio->devs[slot].addr = s;
624 slot++;
625
626 for (f = 1; f < geo->far_copies; f++) {
627 set = d / geo->far_set_size;
628 d += geo->near_copies;
629
630 if ((geo->raid_disks % geo->far_set_size) &&
631 (d > last_far_set_start)) {
632 d -= last_far_set_start;
633 d %= last_far_set_size;
634 d += last_far_set_start;
635 } else {
636 d %= geo->far_set_size;
637 d += geo->far_set_size * set;
638 }
639 s += geo->stride;
640 r10bio->devs[slot].devnum = d;
641 r10bio->devs[slot].addr = s;
642 slot++;
643 }
644 dev++;
645 if (dev >= geo->raid_disks) {
646 dev = 0;
647 sector += (geo->chunk_mask + 1);
648 }
649 }
650}
651
652static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
653{
654 struct geom *geo = &conf->geo;
655
656 if (conf->reshape_progress != MaxSector &&
657 ((r10bio->sector >= conf->reshape_progress) !=
658 conf->mddev->reshape_backwards)) {
659 set_bit(R10BIO_Previous, &r10bio->state);
660 geo = &conf->prev;
661 } else
662 clear_bit(R10BIO_Previous, &r10bio->state);
663
664 __raid10_find_phys(geo, r10bio);
665}
666
667static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
668{
669 sector_t offset, chunk, vchunk;
670 /* Never use conf->prev as this is only called during resync
671 * or recovery, so reshape isn't happening
672 */
673 struct geom *geo = &conf->geo;
674 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
675 int far_set_size = geo->far_set_size;
676 int last_far_set_start;
677
678 if (geo->raid_disks % geo->far_set_size) {
679 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
680 last_far_set_start *= geo->far_set_size;
681
682 if (dev >= last_far_set_start) {
683 far_set_size = geo->far_set_size;
684 far_set_size += (geo->raid_disks % geo->far_set_size);
685 far_set_start = last_far_set_start;
686 }
687 }
688
689 offset = sector & geo->chunk_mask;
690 if (geo->far_offset) {
691 int fc;
692 chunk = sector >> geo->chunk_shift;
693 fc = sector_div(chunk, geo->far_copies);
694 dev -= fc * geo->near_copies;
695 if (dev < far_set_start)
696 dev += far_set_size;
697 } else {
698 while (sector >= geo->stride) {
699 sector -= geo->stride;
700 if (dev < (geo->near_copies + far_set_start))
701 dev += far_set_size - geo->near_copies;
702 else
703 dev -= geo->near_copies;
704 }
705 chunk = sector >> geo->chunk_shift;
706 }
707 vchunk = chunk * geo->raid_disks + dev;
708 sector_div(vchunk, geo->near_copies);
709 return (vchunk << geo->chunk_shift) + offset;
710}
711
712/*
713 * This routine returns the disk from which the requested read should
714 * be done. There is a per-array 'next expected sequential IO' sector
715 * number - if this matches on the next IO then we use the last disk.
716 * There is also a per-disk 'last know head position' sector that is
717 * maintained from IRQ contexts, both the normal and the resync IO
718 * completion handlers update this position correctly. If there is no
719 * perfect sequential match then we pick the disk whose head is closest.
720 *
721 * If there are 2 mirrors in the same 2 devices, performance degrades
722 * because position is mirror, not device based.
723 *
724 * The rdev for the device selected will have nr_pending incremented.
725 */
726
727/*
728 * FIXME: possibly should rethink readbalancing and do it differently
729 * depending on near_copies / far_copies geometry.
730 */
731static struct md_rdev *read_balance(struct r10conf *conf,
732 struct r10bio *r10_bio,
733 int *max_sectors)
734{
735 const sector_t this_sector = r10_bio->sector;
736 int disk, slot;
737 int sectors = r10_bio->sectors;
738 int best_good_sectors;
739 sector_t new_distance, best_dist;
740 struct md_rdev *best_dist_rdev, *best_pending_rdev, *rdev = NULL;
741 int do_balance;
742 int best_dist_slot, best_pending_slot;
743 bool has_nonrot_disk = false;
744 unsigned int min_pending;
745 struct geom *geo = &conf->geo;
746
747 raid10_find_phys(conf, r10_bio);
748 rcu_read_lock();
749 best_dist_slot = -1;
750 min_pending = UINT_MAX;
751 best_dist_rdev = NULL;
752 best_pending_rdev = NULL;
753 best_dist = MaxSector;
754 best_good_sectors = 0;
755 do_balance = 1;
756 clear_bit(R10BIO_FailFast, &r10_bio->state);
757 /*
758 * Check if we can balance. We can balance on the whole
759 * device if no resync is going on (recovery is ok), or below
760 * the resync window. We take the first readable disk when
761 * above the resync window.
762 */
763 if ((conf->mddev->recovery_cp < MaxSector
764 && (this_sector + sectors >= conf->next_resync)) ||
765 (mddev_is_clustered(conf->mddev) &&
766 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
767 this_sector + sectors)))
768 do_balance = 0;
769
770 for (slot = 0; slot < conf->copies ; slot++) {
771 sector_t first_bad;
772 int bad_sectors;
773 sector_t dev_sector;
774 unsigned int pending;
775 bool nonrot;
776
777 if (r10_bio->devs[slot].bio == IO_BLOCKED)
778 continue;
779 disk = r10_bio->devs[slot].devnum;
780 rdev = rcu_dereference(conf->mirrors[disk].replacement);
781 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
782 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
783 rdev = rcu_dereference(conf->mirrors[disk].rdev);
784 if (rdev == NULL ||
785 test_bit(Faulty, &rdev->flags))
786 continue;
787 if (!test_bit(In_sync, &rdev->flags) &&
788 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
789 continue;
790
791 dev_sector = r10_bio->devs[slot].addr;
792 if (is_badblock(rdev, dev_sector, sectors,
793 &first_bad, &bad_sectors)) {
794 if (best_dist < MaxSector)
795 /* Already have a better slot */
796 continue;
797 if (first_bad <= dev_sector) {
798 /* Cannot read here. If this is the
799 * 'primary' device, then we must not read
800 * beyond 'bad_sectors' from another device.
801 */
802 bad_sectors -= (dev_sector - first_bad);
803 if (!do_balance && sectors > bad_sectors)
804 sectors = bad_sectors;
805 if (best_good_sectors > sectors)
806 best_good_sectors = sectors;
807 } else {
808 sector_t good_sectors =
809 first_bad - dev_sector;
810 if (good_sectors > best_good_sectors) {
811 best_good_sectors = good_sectors;
812 best_dist_slot = slot;
813 best_dist_rdev = rdev;
814 }
815 if (!do_balance)
816 /* Must read from here */
817 break;
818 }
819 continue;
820 } else
821 best_good_sectors = sectors;
822
823 if (!do_balance)
824 break;
825
826 nonrot = bdev_nonrot(rdev->bdev);
827 has_nonrot_disk |= nonrot;
828 pending = atomic_read(&rdev->nr_pending);
829 if (min_pending > pending && nonrot) {
830 min_pending = pending;
831 best_pending_slot = slot;
832 best_pending_rdev = rdev;
833 }
834
835 if (best_dist_slot >= 0)
836 /* At least 2 disks to choose from so failfast is OK */
837 set_bit(R10BIO_FailFast, &r10_bio->state);
838 /* This optimisation is debatable, and completely destroys
839 * sequential read speed for 'far copies' arrays. So only
840 * keep it for 'near' arrays, and review those later.
841 */
842 if (geo->near_copies > 1 && !pending)
843 new_distance = 0;
844
845 /* for far > 1 always use the lowest address */
846 else if (geo->far_copies > 1)
847 new_distance = r10_bio->devs[slot].addr;
848 else
849 new_distance = abs(r10_bio->devs[slot].addr -
850 conf->mirrors[disk].head_position);
851
852 if (new_distance < best_dist) {
853 best_dist = new_distance;
854 best_dist_slot = slot;
855 best_dist_rdev = rdev;
856 }
857 }
858 if (slot >= conf->copies) {
859 if (has_nonrot_disk) {
860 slot = best_pending_slot;
861 rdev = best_pending_rdev;
862 } else {
863 slot = best_dist_slot;
864 rdev = best_dist_rdev;
865 }
866 }
867
868 if (slot >= 0) {
869 atomic_inc(&rdev->nr_pending);
870 r10_bio->read_slot = slot;
871 } else
872 rdev = NULL;
873 rcu_read_unlock();
874 *max_sectors = best_good_sectors;
875
876 return rdev;
877}
878
879static void flush_pending_writes(struct r10conf *conf)
880{
881 /* Any writes that have been queued but are awaiting
882 * bitmap updates get flushed here.
883 */
884 spin_lock_irq(&conf->device_lock);
885
886 if (conf->pending_bio_list.head) {
887 struct blk_plug plug;
888 struct bio *bio;
889
890 bio = bio_list_get(&conf->pending_bio_list);
891 spin_unlock_irq(&conf->device_lock);
892
893 /*
894 * As this is called in a wait_event() loop (see freeze_array),
895 * current->state might be TASK_UNINTERRUPTIBLE which will
896 * cause a warning when we prepare to wait again. As it is
897 * rare that this path is taken, it is perfectly safe to force
898 * us to go around the wait_event() loop again, so the warning
899 * is a false-positive. Silence the warning by resetting
900 * thread state
901 */
902 __set_current_state(TASK_RUNNING);
903
904 blk_start_plug(&plug);
905 /* flush any pending bitmap writes to disk
906 * before proceeding w/ I/O */
907 md_bitmap_unplug(conf->mddev->bitmap);
908 wake_up(&conf->wait_barrier);
909
910 while (bio) { /* submit pending writes */
911 struct bio *next = bio->bi_next;
912 struct md_rdev *rdev = (void*)bio->bi_bdev;
913 bio->bi_next = NULL;
914 bio_set_dev(bio, rdev->bdev);
915 if (test_bit(Faulty, &rdev->flags)) {
916 bio_io_error(bio);
917 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
918 !bdev_max_discard_sectors(bio->bi_bdev)))
919 /* Just ignore it */
920 bio_endio(bio);
921 else
922 submit_bio_noacct(bio);
923 bio = next;
924 }
925 blk_finish_plug(&plug);
926 } else
927 spin_unlock_irq(&conf->device_lock);
928}
929
930/* Barriers....
931 * Sometimes we need to suspend IO while we do something else,
932 * either some resync/recovery, or reconfigure the array.
933 * To do this we raise a 'barrier'.
934 * The 'barrier' is a counter that can be raised multiple times
935 * to count how many activities are happening which preclude
936 * normal IO.
937 * We can only raise the barrier if there is no pending IO.
938 * i.e. if nr_pending == 0.
939 * We choose only to raise the barrier if no-one is waiting for the
940 * barrier to go down. This means that as soon as an IO request
941 * is ready, no other operations which require a barrier will start
942 * until the IO request has had a chance.
943 *
944 * So: regular IO calls 'wait_barrier'. When that returns there
945 * is no backgroup IO happening, It must arrange to call
946 * allow_barrier when it has finished its IO.
947 * backgroup IO calls must call raise_barrier. Once that returns
948 * there is no normal IO happeing. It must arrange to call
949 * lower_barrier when the particular background IO completes.
950 */
951
952static void raise_barrier(struct r10conf *conf, int force)
953{
954 write_seqlock_irq(&conf->resync_lock);
955 BUG_ON(force && !conf->barrier);
956
957 /* Wait until no block IO is waiting (unless 'force') */
958 wait_event_barrier(conf, force || !conf->nr_waiting);
959
960 /* block any new IO from starting */
961 WRITE_ONCE(conf->barrier, conf->barrier + 1);
962
963 /* Now wait for all pending IO to complete */
964 wait_event_barrier(conf, !atomic_read(&conf->nr_pending) &&
965 conf->barrier < RESYNC_DEPTH);
966
967 write_sequnlock_irq(&conf->resync_lock);
968}
969
970static void lower_barrier(struct r10conf *conf)
971{
972 unsigned long flags;
973
974 write_seqlock_irqsave(&conf->resync_lock, flags);
975 WRITE_ONCE(conf->barrier, conf->barrier - 1);
976 write_sequnlock_irqrestore(&conf->resync_lock, flags);
977 wake_up(&conf->wait_barrier);
978}
979
980static bool stop_waiting_barrier(struct r10conf *conf)
981{
982 struct bio_list *bio_list = current->bio_list;
983
984 /* barrier is dropped */
985 if (!conf->barrier)
986 return true;
987
988 /*
989 * If there are already pending requests (preventing the barrier from
990 * rising completely), and the pre-process bio queue isn't empty, then
991 * don't wait, as we need to empty that queue to get the nr_pending
992 * count down.
993 */
994 if (atomic_read(&conf->nr_pending) && bio_list &&
995 (!bio_list_empty(&bio_list[0]) || !bio_list_empty(&bio_list[1])))
996 return true;
997
998 /* move on if recovery thread is blocked by us */
999 if (conf->mddev->thread->tsk == current &&
1000 test_bit(MD_RECOVERY_RUNNING, &conf->mddev->recovery) &&
1001 conf->nr_queued > 0)
1002 return true;
1003
1004 return false;
1005}
1006
1007static bool wait_barrier_nolock(struct r10conf *conf)
1008{
1009 unsigned int seq = read_seqbegin(&conf->resync_lock);
1010
1011 if (READ_ONCE(conf->barrier))
1012 return false;
1013
1014 atomic_inc(&conf->nr_pending);
1015 if (!read_seqretry(&conf->resync_lock, seq))
1016 return true;
1017
1018 if (atomic_dec_and_test(&conf->nr_pending))
1019 wake_up_barrier(conf);
1020
1021 return false;
1022}
1023
1024static bool wait_barrier(struct r10conf *conf, bool nowait)
1025{
1026 bool ret = true;
1027
1028 if (wait_barrier_nolock(conf))
1029 return true;
1030
1031 write_seqlock_irq(&conf->resync_lock);
1032 if (conf->barrier) {
1033 /* Return false when nowait flag is set */
1034 if (nowait) {
1035 ret = false;
1036 } else {
1037 conf->nr_waiting++;
1038 raid10_log(conf->mddev, "wait barrier");
1039 wait_event_barrier(conf, stop_waiting_barrier(conf));
1040 conf->nr_waiting--;
1041 }
1042 if (!conf->nr_waiting)
1043 wake_up(&conf->wait_barrier);
1044 }
1045 /* Only increment nr_pending when we wait */
1046 if (ret)
1047 atomic_inc(&conf->nr_pending);
1048 write_sequnlock_irq(&conf->resync_lock);
1049 return ret;
1050}
1051
1052static void allow_barrier(struct r10conf *conf)
1053{
1054 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1055 (conf->array_freeze_pending))
1056 wake_up_barrier(conf);
1057}
1058
1059static void freeze_array(struct r10conf *conf, int extra)
1060{
1061 /* stop syncio and normal IO and wait for everything to
1062 * go quiet.
1063 * We increment barrier and nr_waiting, and then
1064 * wait until nr_pending match nr_queued+extra
1065 * This is called in the context of one normal IO request
1066 * that has failed. Thus any sync request that might be pending
1067 * will be blocked by nr_pending, and we need to wait for
1068 * pending IO requests to complete or be queued for re-try.
1069 * Thus the number queued (nr_queued) plus this request (extra)
1070 * must match the number of pending IOs (nr_pending) before
1071 * we continue.
1072 */
1073 write_seqlock_irq(&conf->resync_lock);
1074 conf->array_freeze_pending++;
1075 WRITE_ONCE(conf->barrier, conf->barrier + 1);
1076 conf->nr_waiting++;
1077 wait_event_barrier_cmd(conf, atomic_read(&conf->nr_pending) ==
1078 conf->nr_queued + extra, flush_pending_writes(conf));
1079 conf->array_freeze_pending--;
1080 write_sequnlock_irq(&conf->resync_lock);
1081}
1082
1083static void unfreeze_array(struct r10conf *conf)
1084{
1085 /* reverse the effect of the freeze */
1086 write_seqlock_irq(&conf->resync_lock);
1087 WRITE_ONCE(conf->barrier, conf->barrier - 1);
1088 conf->nr_waiting--;
1089 wake_up(&conf->wait_barrier);
1090 write_sequnlock_irq(&conf->resync_lock);
1091}
1092
1093static sector_t choose_data_offset(struct r10bio *r10_bio,
1094 struct md_rdev *rdev)
1095{
1096 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1097 test_bit(R10BIO_Previous, &r10_bio->state))
1098 return rdev->data_offset;
1099 else
1100 return rdev->new_data_offset;
1101}
1102
1103static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1104{
1105 struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, cb);
1106 struct mddev *mddev = plug->cb.data;
1107 struct r10conf *conf = mddev->private;
1108 struct bio *bio;
1109
1110 if (from_schedule || current->bio_list) {
1111 spin_lock_irq(&conf->device_lock);
1112 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1113 spin_unlock_irq(&conf->device_lock);
1114 wake_up(&conf->wait_barrier);
1115 md_wakeup_thread(mddev->thread);
1116 kfree(plug);
1117 return;
1118 }
1119
1120 /* we aren't scheduling, so we can do the write-out directly. */
1121 bio = bio_list_get(&plug->pending);
1122 md_bitmap_unplug(mddev->bitmap);
1123 wake_up(&conf->wait_barrier);
1124
1125 while (bio) { /* submit pending writes */
1126 struct bio *next = bio->bi_next;
1127 struct md_rdev *rdev = (void*)bio->bi_bdev;
1128 bio->bi_next = NULL;
1129 bio_set_dev(bio, rdev->bdev);
1130 if (test_bit(Faulty, &rdev->flags)) {
1131 bio_io_error(bio);
1132 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1133 !bdev_max_discard_sectors(bio->bi_bdev)))
1134 /* Just ignore it */
1135 bio_endio(bio);
1136 else
1137 submit_bio_noacct(bio);
1138 bio = next;
1139 }
1140 kfree(plug);
1141}
1142
1143/*
1144 * 1. Register the new request and wait if the reconstruction thread has put
1145 * up a bar for new requests. Continue immediately if no resync is active
1146 * currently.
1147 * 2. If IO spans the reshape position. Need to wait for reshape to pass.
1148 */
1149static bool regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1150 struct bio *bio, sector_t sectors)
1151{
1152 /* Bail out if REQ_NOWAIT is set for the bio */
1153 if (!wait_barrier(conf, bio->bi_opf & REQ_NOWAIT)) {
1154 bio_wouldblock_error(bio);
1155 return false;
1156 }
1157 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1158 bio->bi_iter.bi_sector < conf->reshape_progress &&
1159 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1160 allow_barrier(conf);
1161 if (bio->bi_opf & REQ_NOWAIT) {
1162 bio_wouldblock_error(bio);
1163 return false;
1164 }
1165 raid10_log(conf->mddev, "wait reshape");
1166 wait_event(conf->wait_barrier,
1167 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1168 conf->reshape_progress >= bio->bi_iter.bi_sector +
1169 sectors);
1170 wait_barrier(conf, false);
1171 }
1172 return true;
1173}
1174
1175static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1176 struct r10bio *r10_bio)
1177{
1178 struct r10conf *conf = mddev->private;
1179 struct bio *read_bio;
1180 const enum req_op op = bio_op(bio);
1181 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1182 int max_sectors;
1183 struct md_rdev *rdev;
1184 char b[BDEVNAME_SIZE];
1185 int slot = r10_bio->read_slot;
1186 struct md_rdev *err_rdev = NULL;
1187 gfp_t gfp = GFP_NOIO;
1188
1189 if (slot >= 0 && r10_bio->devs[slot].rdev) {
1190 /*
1191 * This is an error retry, but we cannot
1192 * safely dereference the rdev in the r10_bio,
1193 * we must use the one in conf.
1194 * If it has already been disconnected (unlikely)
1195 * we lose the device name in error messages.
1196 */
1197 int disk;
1198 /*
1199 * As we are blocking raid10, it is a little safer to
1200 * use __GFP_HIGH.
1201 */
1202 gfp = GFP_NOIO | __GFP_HIGH;
1203
1204 rcu_read_lock();
1205 disk = r10_bio->devs[slot].devnum;
1206 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1207 if (err_rdev)
1208 snprintf(b, sizeof(b), "%pg", err_rdev->bdev);
1209 else {
1210 strcpy(b, "???");
1211 /* This never gets dereferenced */
1212 err_rdev = r10_bio->devs[slot].rdev;
1213 }
1214 rcu_read_unlock();
1215 }
1216
1217 if (!regular_request_wait(mddev, conf, bio, r10_bio->sectors))
1218 return;
1219 rdev = read_balance(conf, r10_bio, &max_sectors);
1220 if (!rdev) {
1221 if (err_rdev) {
1222 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1223 mdname(mddev), b,
1224 (unsigned long long)r10_bio->sector);
1225 }
1226 raid_end_bio_io(r10_bio);
1227 return;
1228 }
1229 if (err_rdev)
1230 pr_err_ratelimited("md/raid10:%s: %pg: redirecting sector %llu to another mirror\n",
1231 mdname(mddev),
1232 rdev->bdev,
1233 (unsigned long long)r10_bio->sector);
1234 if (max_sectors < bio_sectors(bio)) {
1235 struct bio *split = bio_split(bio, max_sectors,
1236 gfp, &conf->bio_split);
1237 bio_chain(split, bio);
1238 allow_barrier(conf);
1239 submit_bio_noacct(bio);
1240 wait_barrier(conf, false);
1241 bio = split;
1242 r10_bio->master_bio = bio;
1243 r10_bio->sectors = max_sectors;
1244 }
1245 slot = r10_bio->read_slot;
1246
1247 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1248 r10_bio->start_time = bio_start_io_acct(bio);
1249 read_bio = bio_alloc_clone(rdev->bdev, bio, gfp, &mddev->bio_set);
1250
1251 r10_bio->devs[slot].bio = read_bio;
1252 r10_bio->devs[slot].rdev = rdev;
1253
1254 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1255 choose_data_offset(r10_bio, rdev);
1256 read_bio->bi_end_io = raid10_end_read_request;
1257 read_bio->bi_opf = op | do_sync;
1258 if (test_bit(FailFast, &rdev->flags) &&
1259 test_bit(R10BIO_FailFast, &r10_bio->state))
1260 read_bio->bi_opf |= MD_FAILFAST;
1261 read_bio->bi_private = r10_bio;
1262
1263 if (mddev->gendisk)
1264 trace_block_bio_remap(read_bio, disk_devt(mddev->gendisk),
1265 r10_bio->sector);
1266 submit_bio_noacct(read_bio);
1267 return;
1268}
1269
1270static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1271 struct bio *bio, bool replacement,
1272 int n_copy)
1273{
1274 const enum req_op op = bio_op(bio);
1275 const blk_opf_t do_sync = bio->bi_opf & REQ_SYNC;
1276 const blk_opf_t do_fua = bio->bi_opf & REQ_FUA;
1277 unsigned long flags;
1278 struct blk_plug_cb *cb;
1279 struct raid1_plug_cb *plug = NULL;
1280 struct r10conf *conf = mddev->private;
1281 struct md_rdev *rdev;
1282 int devnum = r10_bio->devs[n_copy].devnum;
1283 struct bio *mbio;
1284
1285 if (replacement) {
1286 rdev = conf->mirrors[devnum].replacement;
1287 if (rdev == NULL) {
1288 /* Replacement just got moved to main 'rdev' */
1289 smp_mb();
1290 rdev = conf->mirrors[devnum].rdev;
1291 }
1292 } else
1293 rdev = conf->mirrors[devnum].rdev;
1294
1295 mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO, &mddev->bio_set);
1296 if (replacement)
1297 r10_bio->devs[n_copy].repl_bio = mbio;
1298 else
1299 r10_bio->devs[n_copy].bio = mbio;
1300
1301 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1302 choose_data_offset(r10_bio, rdev));
1303 mbio->bi_end_io = raid10_end_write_request;
1304 mbio->bi_opf = op | do_sync | do_fua;
1305 if (!replacement && test_bit(FailFast,
1306 &conf->mirrors[devnum].rdev->flags)
1307 && enough(conf, devnum))
1308 mbio->bi_opf |= MD_FAILFAST;
1309 mbio->bi_private = r10_bio;
1310
1311 if (conf->mddev->gendisk)
1312 trace_block_bio_remap(mbio, disk_devt(conf->mddev->gendisk),
1313 r10_bio->sector);
1314 /* flush_pending_writes() needs access to the rdev so...*/
1315 mbio->bi_bdev = (void *)rdev;
1316
1317 atomic_inc(&r10_bio->remaining);
1318
1319 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1320 if (cb)
1321 plug = container_of(cb, struct raid1_plug_cb, cb);
1322 else
1323 plug = NULL;
1324 if (plug) {
1325 bio_list_add(&plug->pending, mbio);
1326 } else {
1327 spin_lock_irqsave(&conf->device_lock, flags);
1328 bio_list_add(&conf->pending_bio_list, mbio);
1329 spin_unlock_irqrestore(&conf->device_lock, flags);
1330 md_wakeup_thread(mddev->thread);
1331 }
1332}
1333
1334static void wait_blocked_dev(struct mddev *mddev, struct r10bio *r10_bio)
1335{
1336 int i;
1337 struct r10conf *conf = mddev->private;
1338 struct md_rdev *blocked_rdev;
1339
1340retry_wait:
1341 blocked_rdev = NULL;
1342 rcu_read_lock();
1343 for (i = 0; i < conf->copies; i++) {
1344 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1345 struct md_rdev *rrdev = rcu_dereference(
1346 conf->mirrors[i].replacement);
1347 if (rdev == rrdev)
1348 rrdev = NULL;
1349 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1350 atomic_inc(&rdev->nr_pending);
1351 blocked_rdev = rdev;
1352 break;
1353 }
1354 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1355 atomic_inc(&rrdev->nr_pending);
1356 blocked_rdev = rrdev;
1357 break;
1358 }
1359
1360 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1361 sector_t first_bad;
1362 sector_t dev_sector = r10_bio->devs[i].addr;
1363 int bad_sectors;
1364 int is_bad;
1365
1366 /*
1367 * Discard request doesn't care the write result
1368 * so it doesn't need to wait blocked disk here.
1369 */
1370 if (!r10_bio->sectors)
1371 continue;
1372
1373 is_bad = is_badblock(rdev, dev_sector, r10_bio->sectors,
1374 &first_bad, &bad_sectors);
1375 if (is_bad < 0) {
1376 /*
1377 * Mustn't write here until the bad block
1378 * is acknowledged
1379 */
1380 atomic_inc(&rdev->nr_pending);
1381 set_bit(BlockedBadBlocks, &rdev->flags);
1382 blocked_rdev = rdev;
1383 break;
1384 }
1385 }
1386 }
1387 rcu_read_unlock();
1388
1389 if (unlikely(blocked_rdev)) {
1390 /* Have to wait for this device to get unblocked, then retry */
1391 allow_barrier(conf);
1392 raid10_log(conf->mddev, "%s wait rdev %d blocked",
1393 __func__, blocked_rdev->raid_disk);
1394 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1395 wait_barrier(conf, false);
1396 goto retry_wait;
1397 }
1398}
1399
1400static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1401 struct r10bio *r10_bio)
1402{
1403 struct r10conf *conf = mddev->private;
1404 int i;
1405 sector_t sectors;
1406 int max_sectors;
1407
1408 if ((mddev_is_clustered(mddev) &&
1409 md_cluster_ops->area_resyncing(mddev, WRITE,
1410 bio->bi_iter.bi_sector,
1411 bio_end_sector(bio)))) {
1412 DEFINE_WAIT(w);
1413 /* Bail out if REQ_NOWAIT is set for the bio */
1414 if (bio->bi_opf & REQ_NOWAIT) {
1415 bio_wouldblock_error(bio);
1416 return;
1417 }
1418 for (;;) {
1419 prepare_to_wait(&conf->wait_barrier,
1420 &w, TASK_IDLE);
1421 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1422 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1423 break;
1424 schedule();
1425 }
1426 finish_wait(&conf->wait_barrier, &w);
1427 }
1428
1429 sectors = r10_bio->sectors;
1430 if (!regular_request_wait(mddev, conf, bio, sectors))
1431 return;
1432 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1433 (mddev->reshape_backwards
1434 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1435 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1436 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1437 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1438 /* Need to update reshape_position in metadata */
1439 mddev->reshape_position = conf->reshape_progress;
1440 set_mask_bits(&mddev->sb_flags, 0,
1441 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1442 md_wakeup_thread(mddev->thread);
1443 if (bio->bi_opf & REQ_NOWAIT) {
1444 allow_barrier(conf);
1445 bio_wouldblock_error(bio);
1446 return;
1447 }
1448 raid10_log(conf->mddev, "wait reshape metadata");
1449 wait_event(mddev->sb_wait,
1450 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1451
1452 conf->reshape_safe = mddev->reshape_position;
1453 }
1454
1455 /* first select target devices under rcu_lock and
1456 * inc refcount on their rdev. Record them by setting
1457 * bios[x] to bio
1458 * If there are known/acknowledged bad blocks on any device
1459 * on which we have seen a write error, we want to avoid
1460 * writing to those blocks. This potentially requires several
1461 * writes to write around the bad blocks. Each set of writes
1462 * gets its own r10_bio with a set of bios attached.
1463 */
1464
1465 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1466 raid10_find_phys(conf, r10_bio);
1467
1468 wait_blocked_dev(mddev, r10_bio);
1469
1470 rcu_read_lock();
1471 max_sectors = r10_bio->sectors;
1472
1473 for (i = 0; i < conf->copies; i++) {
1474 int d = r10_bio->devs[i].devnum;
1475 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1476 struct md_rdev *rrdev = rcu_dereference(
1477 conf->mirrors[d].replacement);
1478 if (rdev == rrdev)
1479 rrdev = NULL;
1480 if (rdev && (test_bit(Faulty, &rdev->flags)))
1481 rdev = NULL;
1482 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1483 rrdev = NULL;
1484
1485 r10_bio->devs[i].bio = NULL;
1486 r10_bio->devs[i].repl_bio = NULL;
1487
1488 if (!rdev && !rrdev) {
1489 set_bit(R10BIO_Degraded, &r10_bio->state);
1490 continue;
1491 }
1492 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1493 sector_t first_bad;
1494 sector_t dev_sector = r10_bio->devs[i].addr;
1495 int bad_sectors;
1496 int is_bad;
1497
1498 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1499 &first_bad, &bad_sectors);
1500 if (is_bad && first_bad <= dev_sector) {
1501 /* Cannot write here at all */
1502 bad_sectors -= (dev_sector - first_bad);
1503 if (bad_sectors < max_sectors)
1504 /* Mustn't write more than bad_sectors
1505 * to other devices yet
1506 */
1507 max_sectors = bad_sectors;
1508 /* We don't set R10BIO_Degraded as that
1509 * only applies if the disk is missing,
1510 * so it might be re-added, and we want to
1511 * know to recover this chunk.
1512 * In this case the device is here, and the
1513 * fact that this chunk is not in-sync is
1514 * recorded in the bad block log.
1515 */
1516 continue;
1517 }
1518 if (is_bad) {
1519 int good_sectors = first_bad - dev_sector;
1520 if (good_sectors < max_sectors)
1521 max_sectors = good_sectors;
1522 }
1523 }
1524 if (rdev) {
1525 r10_bio->devs[i].bio = bio;
1526 atomic_inc(&rdev->nr_pending);
1527 }
1528 if (rrdev) {
1529 r10_bio->devs[i].repl_bio = bio;
1530 atomic_inc(&rrdev->nr_pending);
1531 }
1532 }
1533 rcu_read_unlock();
1534
1535 if (max_sectors < r10_bio->sectors)
1536 r10_bio->sectors = max_sectors;
1537
1538 if (r10_bio->sectors < bio_sectors(bio)) {
1539 struct bio *split = bio_split(bio, r10_bio->sectors,
1540 GFP_NOIO, &conf->bio_split);
1541 bio_chain(split, bio);
1542 allow_barrier(conf);
1543 submit_bio_noacct(bio);
1544 wait_barrier(conf, false);
1545 bio = split;
1546 r10_bio->master_bio = bio;
1547 }
1548
1549 if (blk_queue_io_stat(bio->bi_bdev->bd_disk->queue))
1550 r10_bio->start_time = bio_start_io_acct(bio);
1551 atomic_set(&r10_bio->remaining, 1);
1552 md_bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1553
1554 for (i = 0; i < conf->copies; i++) {
1555 if (r10_bio->devs[i].bio)
1556 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1557 if (r10_bio->devs[i].repl_bio)
1558 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1559 }
1560 one_write_done(r10_bio);
1561}
1562
1563static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1564{
1565 struct r10conf *conf = mddev->private;
1566 struct r10bio *r10_bio;
1567
1568 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1569
1570 r10_bio->master_bio = bio;
1571 r10_bio->sectors = sectors;
1572
1573 r10_bio->mddev = mddev;
1574 r10_bio->sector = bio->bi_iter.bi_sector;
1575 r10_bio->state = 0;
1576 r10_bio->read_slot = -1;
1577 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) *
1578 conf->geo.raid_disks);
1579
1580 if (bio_data_dir(bio) == READ)
1581 raid10_read_request(mddev, bio, r10_bio);
1582 else
1583 raid10_write_request(mddev, bio, r10_bio);
1584}
1585
1586static void raid_end_discard_bio(struct r10bio *r10bio)
1587{
1588 struct r10conf *conf = r10bio->mddev->private;
1589 struct r10bio *first_r10bio;
1590
1591 while (atomic_dec_and_test(&r10bio->remaining)) {
1592
1593 allow_barrier(conf);
1594
1595 if (!test_bit(R10BIO_Discard, &r10bio->state)) {
1596 first_r10bio = (struct r10bio *)r10bio->master_bio;
1597 free_r10bio(r10bio);
1598 r10bio = first_r10bio;
1599 } else {
1600 md_write_end(r10bio->mddev);
1601 bio_endio(r10bio->master_bio);
1602 free_r10bio(r10bio);
1603 break;
1604 }
1605 }
1606}
1607
1608static void raid10_end_discard_request(struct bio *bio)
1609{
1610 struct r10bio *r10_bio = bio->bi_private;
1611 struct r10conf *conf = r10_bio->mddev->private;
1612 struct md_rdev *rdev = NULL;
1613 int dev;
1614 int slot, repl;
1615
1616 /*
1617 * We don't care the return value of discard bio
1618 */
1619 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
1620 set_bit(R10BIO_Uptodate, &r10_bio->state);
1621
1622 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1623 if (repl)
1624 rdev = conf->mirrors[dev].replacement;
1625 if (!rdev) {
1626 /*
1627 * raid10_remove_disk uses smp_mb to make sure rdev is set to
1628 * replacement before setting replacement to NULL. It can read
1629 * rdev first without barrier protect even replacment is NULL
1630 */
1631 smp_rmb();
1632 rdev = conf->mirrors[dev].rdev;
1633 }
1634
1635 raid_end_discard_bio(r10_bio);
1636 rdev_dec_pending(rdev, conf->mddev);
1637}
1638
1639/*
1640 * There are some limitations to handle discard bio
1641 * 1st, the discard size is bigger than stripe_size*2.
1642 * 2st, if the discard bio spans reshape progress, we use the old way to
1643 * handle discard bio
1644 */
1645static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
1646{
1647 struct r10conf *conf = mddev->private;
1648 struct geom *geo = &conf->geo;
1649 int far_copies = geo->far_copies;
1650 bool first_copy = true;
1651 struct r10bio *r10_bio, *first_r10bio;
1652 struct bio *split;
1653 int disk;
1654 sector_t chunk;
1655 unsigned int stripe_size;
1656 unsigned int stripe_data_disks;
1657 sector_t split_size;
1658 sector_t bio_start, bio_end;
1659 sector_t first_stripe_index, last_stripe_index;
1660 sector_t start_disk_offset;
1661 unsigned int start_disk_index;
1662 sector_t end_disk_offset;
1663 unsigned int end_disk_index;
1664 unsigned int remainder;
1665
1666 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1667 return -EAGAIN;
1668
1669 if (WARN_ON_ONCE(bio->bi_opf & REQ_NOWAIT)) {
1670 bio_wouldblock_error(bio);
1671 return 0;
1672 }
1673 wait_barrier(conf, false);
1674
1675 /*
1676 * Check reshape again to avoid reshape happens after checking
1677 * MD_RECOVERY_RESHAPE and before wait_barrier
1678 */
1679 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1680 goto out;
1681
1682 if (geo->near_copies)
1683 stripe_data_disks = geo->raid_disks / geo->near_copies +
1684 geo->raid_disks % geo->near_copies;
1685 else
1686 stripe_data_disks = geo->raid_disks;
1687
1688 stripe_size = stripe_data_disks << geo->chunk_shift;
1689
1690 bio_start = bio->bi_iter.bi_sector;
1691 bio_end = bio_end_sector(bio);
1692
1693 /*
1694 * Maybe one discard bio is smaller than strip size or across one
1695 * stripe and discard region is larger than one stripe size. For far
1696 * offset layout, if the discard region is not aligned with stripe
1697 * size, there is hole when we submit discard bio to member disk.
1698 * For simplicity, we only handle discard bio which discard region
1699 * is bigger than stripe_size * 2
1700 */
1701 if (bio_sectors(bio) < stripe_size*2)
1702 goto out;
1703
1704 /*
1705 * Keep bio aligned with strip size.
1706 */
1707 div_u64_rem(bio_start, stripe_size, &remainder);
1708 if (remainder) {
1709 split_size = stripe_size - remainder;
1710 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1711 bio_chain(split, bio);
1712 allow_barrier(conf);
1713 /* Resend the fist split part */
1714 submit_bio_noacct(split);
1715 wait_barrier(conf, false);
1716 }
1717 div_u64_rem(bio_end, stripe_size, &remainder);
1718 if (remainder) {
1719 split_size = bio_sectors(bio) - remainder;
1720 split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1721 bio_chain(split, bio);
1722 allow_barrier(conf);
1723 /* Resend the second split part */
1724 submit_bio_noacct(bio);
1725 bio = split;
1726 wait_barrier(conf, false);
1727 }
1728
1729 bio_start = bio->bi_iter.bi_sector;
1730 bio_end = bio_end_sector(bio);
1731
1732 /*
1733 * Raid10 uses chunk as the unit to store data. It's similar like raid0.
1734 * One stripe contains the chunks from all member disk (one chunk from
1735 * one disk at the same HBA address). For layout detail, see 'man md 4'
1736 */
1737 chunk = bio_start >> geo->chunk_shift;
1738 chunk *= geo->near_copies;
1739 first_stripe_index = chunk;
1740 start_disk_index = sector_div(first_stripe_index, geo->raid_disks);
1741 if (geo->far_offset)
1742 first_stripe_index *= geo->far_copies;
1743 start_disk_offset = (bio_start & geo->chunk_mask) +
1744 (first_stripe_index << geo->chunk_shift);
1745
1746 chunk = bio_end >> geo->chunk_shift;
1747 chunk *= geo->near_copies;
1748 last_stripe_index = chunk;
1749 end_disk_index = sector_div(last_stripe_index, geo->raid_disks);
1750 if (geo->far_offset)
1751 last_stripe_index *= geo->far_copies;
1752 end_disk_offset = (bio_end & geo->chunk_mask) +
1753 (last_stripe_index << geo->chunk_shift);
1754
1755retry_discard:
1756 r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1757 r10_bio->mddev = mddev;
1758 r10_bio->state = 0;
1759 r10_bio->sectors = 0;
1760 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * geo->raid_disks);
1761 wait_blocked_dev(mddev, r10_bio);
1762
1763 /*
1764 * For far layout it needs more than one r10bio to cover all regions.
1765 * Inspired by raid10_sync_request, we can use the first r10bio->master_bio
1766 * to record the discard bio. Other r10bio->master_bio record the first
1767 * r10bio. The first r10bio only release after all other r10bios finish.
1768 * The discard bio returns only first r10bio finishes
1769 */
1770 if (first_copy) {
1771 r10_bio->master_bio = bio;
1772 set_bit(R10BIO_Discard, &r10_bio->state);
1773 first_copy = false;
1774 first_r10bio = r10_bio;
1775 } else
1776 r10_bio->master_bio = (struct bio *)first_r10bio;
1777
1778 /*
1779 * first select target devices under rcu_lock and
1780 * inc refcount on their rdev. Record them by setting
1781 * bios[x] to bio
1782 */
1783 rcu_read_lock();
1784 for (disk = 0; disk < geo->raid_disks; disk++) {
1785 struct md_rdev *rdev = rcu_dereference(conf->mirrors[disk].rdev);
1786 struct md_rdev *rrdev = rcu_dereference(
1787 conf->mirrors[disk].replacement);
1788
1789 r10_bio->devs[disk].bio = NULL;
1790 r10_bio->devs[disk].repl_bio = NULL;
1791
1792 if (rdev && (test_bit(Faulty, &rdev->flags)))
1793 rdev = NULL;
1794 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1795 rrdev = NULL;
1796 if (!rdev && !rrdev)
1797 continue;
1798
1799 if (rdev) {
1800 r10_bio->devs[disk].bio = bio;
1801 atomic_inc(&rdev->nr_pending);
1802 }
1803 if (rrdev) {
1804 r10_bio->devs[disk].repl_bio = bio;
1805 atomic_inc(&rrdev->nr_pending);
1806 }
1807 }
1808 rcu_read_unlock();
1809
1810 atomic_set(&r10_bio->remaining, 1);
1811 for (disk = 0; disk < geo->raid_disks; disk++) {
1812 sector_t dev_start, dev_end;
1813 struct bio *mbio, *rbio = NULL;
1814
1815 /*
1816 * Now start to calculate the start and end address for each disk.
1817 * The space between dev_start and dev_end is the discard region.
1818 *
1819 * For dev_start, it needs to consider three conditions:
1820 * 1st, the disk is before start_disk, you can imagine the disk in
1821 * the next stripe. So the dev_start is the start address of next
1822 * stripe.
1823 * 2st, the disk is after start_disk, it means the disk is at the
1824 * same stripe of first disk
1825 * 3st, the first disk itself, we can use start_disk_offset directly
1826 */
1827 if (disk < start_disk_index)
1828 dev_start = (first_stripe_index + 1) * mddev->chunk_sectors;
1829 else if (disk > start_disk_index)
1830 dev_start = first_stripe_index * mddev->chunk_sectors;
1831 else
1832 dev_start = start_disk_offset;
1833
1834 if (disk < end_disk_index)
1835 dev_end = (last_stripe_index + 1) * mddev->chunk_sectors;
1836 else if (disk > end_disk_index)
1837 dev_end = last_stripe_index * mddev->chunk_sectors;
1838 else
1839 dev_end = end_disk_offset;
1840
1841 /*
1842 * It only handles discard bio which size is >= stripe size, so
1843 * dev_end > dev_start all the time.
1844 * It doesn't need to use rcu lock to get rdev here. We already
1845 * add rdev->nr_pending in the first loop.
1846 */
1847 if (r10_bio->devs[disk].bio) {
1848 struct md_rdev *rdev = conf->mirrors[disk].rdev;
1849 mbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1850 &mddev->bio_set);
1851 mbio->bi_end_io = raid10_end_discard_request;
1852 mbio->bi_private = r10_bio;
1853 r10_bio->devs[disk].bio = mbio;
1854 r10_bio->devs[disk].devnum = disk;
1855 atomic_inc(&r10_bio->remaining);
1856 md_submit_discard_bio(mddev, rdev, mbio,
1857 dev_start + choose_data_offset(r10_bio, rdev),
1858 dev_end - dev_start);
1859 bio_endio(mbio);
1860 }
1861 if (r10_bio->devs[disk].repl_bio) {
1862 struct md_rdev *rrdev = conf->mirrors[disk].replacement;
1863 rbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1864 &mddev->bio_set);
1865 rbio->bi_end_io = raid10_end_discard_request;
1866 rbio->bi_private = r10_bio;
1867 r10_bio->devs[disk].repl_bio = rbio;
1868 r10_bio->devs[disk].devnum = disk;
1869 atomic_inc(&r10_bio->remaining);
1870 md_submit_discard_bio(mddev, rrdev, rbio,
1871 dev_start + choose_data_offset(r10_bio, rrdev),
1872 dev_end - dev_start);
1873 bio_endio(rbio);
1874 }
1875 }
1876
1877 if (!geo->far_offset && --far_copies) {
1878 first_stripe_index += geo->stride >> geo->chunk_shift;
1879 start_disk_offset += geo->stride;
1880 last_stripe_index += geo->stride >> geo->chunk_shift;
1881 end_disk_offset += geo->stride;
1882 atomic_inc(&first_r10bio->remaining);
1883 raid_end_discard_bio(r10_bio);
1884 wait_barrier(conf, false);
1885 goto retry_discard;
1886 }
1887
1888 raid_end_discard_bio(r10_bio);
1889
1890 return 0;
1891out:
1892 allow_barrier(conf);
1893 return -EAGAIN;
1894}
1895
1896static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1897{
1898 struct r10conf *conf = mddev->private;
1899 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1900 int chunk_sects = chunk_mask + 1;
1901 int sectors = bio_sectors(bio);
1902
1903 if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1904 && md_flush_request(mddev, bio))
1905 return true;
1906
1907 if (!md_write_start(mddev, bio))
1908 return false;
1909
1910 if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
1911 if (!raid10_handle_discard(mddev, bio))
1912 return true;
1913
1914 /*
1915 * If this request crosses a chunk boundary, we need to split
1916 * it.
1917 */
1918 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1919 sectors > chunk_sects
1920 && (conf->geo.near_copies < conf->geo.raid_disks
1921 || conf->prev.near_copies <
1922 conf->prev.raid_disks)))
1923 sectors = chunk_sects -
1924 (bio->bi_iter.bi_sector &
1925 (chunk_sects - 1));
1926 __make_request(mddev, bio, sectors);
1927
1928 /* In case raid10d snuck in to freeze_array */
1929 wake_up_barrier(conf);
1930 return true;
1931}
1932
1933static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1934{
1935 struct r10conf *conf = mddev->private;
1936 int i;
1937
1938 if (conf->geo.near_copies < conf->geo.raid_disks)
1939 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1940 if (conf->geo.near_copies > 1)
1941 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1942 if (conf->geo.far_copies > 1) {
1943 if (conf->geo.far_offset)
1944 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1945 else
1946 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1947 if (conf->geo.far_set_size != conf->geo.raid_disks)
1948 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1949 }
1950 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1951 conf->geo.raid_disks - mddev->degraded);
1952 rcu_read_lock();
1953 for (i = 0; i < conf->geo.raid_disks; i++) {
1954 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1955 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1956 }
1957 rcu_read_unlock();
1958 seq_printf(seq, "]");
1959}
1960
1961/* check if there are enough drives for
1962 * every block to appear on atleast one.
1963 * Don't consider the device numbered 'ignore'
1964 * as we might be about to remove it.
1965 */
1966static int _enough(struct r10conf *conf, int previous, int ignore)
1967{
1968 int first = 0;
1969 int has_enough = 0;
1970 int disks, ncopies;
1971 if (previous) {
1972 disks = conf->prev.raid_disks;
1973 ncopies = conf->prev.near_copies;
1974 } else {
1975 disks = conf->geo.raid_disks;
1976 ncopies = conf->geo.near_copies;
1977 }
1978
1979 rcu_read_lock();
1980 do {
1981 int n = conf->copies;
1982 int cnt = 0;
1983 int this = first;
1984 while (n--) {
1985 struct md_rdev *rdev;
1986 if (this != ignore &&
1987 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1988 test_bit(In_sync, &rdev->flags))
1989 cnt++;
1990 this = (this+1) % disks;
1991 }
1992 if (cnt == 0)
1993 goto out;
1994 first = (first + ncopies) % disks;
1995 } while (first != 0);
1996 has_enough = 1;
1997out:
1998 rcu_read_unlock();
1999 return has_enough;
2000}
2001
2002static int enough(struct r10conf *conf, int ignore)
2003{
2004 /* when calling 'enough', both 'prev' and 'geo' must
2005 * be stable.
2006 * This is ensured if ->reconfig_mutex or ->device_lock
2007 * is held.
2008 */
2009 return _enough(conf, 0, ignore) &&
2010 _enough(conf, 1, ignore);
2011}
2012
2013/**
2014 * raid10_error() - RAID10 error handler.
2015 * @mddev: affected md device.
2016 * @rdev: member device to fail.
2017 *
2018 * The routine acknowledges &rdev failure and determines new @mddev state.
2019 * If it failed, then:
2020 * - &MD_BROKEN flag is set in &mddev->flags.
2021 * Otherwise, it must be degraded:
2022 * - recovery is interrupted.
2023 * - &mddev->degraded is bumped.
2024 *
2025 * @rdev is marked as &Faulty excluding case when array is failed and
2026 * &mddev->fail_last_dev is off.
2027 */
2028static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
2029{
2030 struct r10conf *conf = mddev->private;
2031 unsigned long flags;
2032
2033 spin_lock_irqsave(&conf->device_lock, flags);
2034
2035 if (test_bit(In_sync, &rdev->flags) && !enough(conf, rdev->raid_disk)) {
2036 set_bit(MD_BROKEN, &mddev->flags);
2037
2038 if (!mddev->fail_last_dev) {
2039 spin_unlock_irqrestore(&conf->device_lock, flags);
2040 return;
2041 }
2042 }
2043 if (test_and_clear_bit(In_sync, &rdev->flags))
2044 mddev->degraded++;
2045
2046 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2047 set_bit(Blocked, &rdev->flags);
2048 set_bit(Faulty, &rdev->flags);
2049 set_mask_bits(&mddev->sb_flags, 0,
2050 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2051 spin_unlock_irqrestore(&conf->device_lock, flags);
2052 pr_crit("md/raid10:%s: Disk failure on %pg, disabling device.\n"
2053 "md/raid10:%s: Operation continuing on %d devices.\n",
2054 mdname(mddev), rdev->bdev,
2055 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
2056}
2057
2058static void print_conf(struct r10conf *conf)
2059{
2060 int i;
2061 struct md_rdev *rdev;
2062
2063 pr_debug("RAID10 conf printout:\n");
2064 if (!conf) {
2065 pr_debug("(!conf)\n");
2066 return;
2067 }
2068 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
2069 conf->geo.raid_disks);
2070
2071 /* This is only called with ->reconfix_mutex held, so
2072 * rcu protection of rdev is not needed */
2073 for (i = 0; i < conf->geo.raid_disks; i++) {
2074 rdev = conf->mirrors[i].rdev;
2075 if (rdev)
2076 pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
2077 i, !test_bit(In_sync, &rdev->flags),
2078 !test_bit(Faulty, &rdev->flags),
2079 rdev->bdev);
2080 }
2081}
2082
2083static void close_sync(struct r10conf *conf)
2084{
2085 wait_barrier(conf, false);
2086 allow_barrier(conf);
2087
2088 mempool_exit(&conf->r10buf_pool);
2089}
2090
2091static int raid10_spare_active(struct mddev *mddev)
2092{
2093 int i;
2094 struct r10conf *conf = mddev->private;
2095 struct raid10_info *tmp;
2096 int count = 0;
2097 unsigned long flags;
2098
2099 /*
2100 * Find all non-in_sync disks within the RAID10 configuration
2101 * and mark them in_sync
2102 */
2103 for (i = 0; i < conf->geo.raid_disks; i++) {
2104 tmp = conf->mirrors + i;
2105 if (tmp->replacement
2106 && tmp->replacement->recovery_offset == MaxSector
2107 && !test_bit(Faulty, &tmp->replacement->flags)
2108 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
2109 /* Replacement has just become active */
2110 if (!tmp->rdev
2111 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
2112 count++;
2113 if (tmp->rdev) {
2114 /* Replaced device not technically faulty,
2115 * but we need to be sure it gets removed
2116 * and never re-added.
2117 */
2118 set_bit(Faulty, &tmp->rdev->flags);
2119 sysfs_notify_dirent_safe(
2120 tmp->rdev->sysfs_state);
2121 }
2122 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
2123 } else if (tmp->rdev
2124 && tmp->rdev->recovery_offset == MaxSector
2125 && !test_bit(Faulty, &tmp->rdev->flags)
2126 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
2127 count++;
2128 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
2129 }
2130 }
2131 spin_lock_irqsave(&conf->device_lock, flags);
2132 mddev->degraded -= count;
2133 spin_unlock_irqrestore(&conf->device_lock, flags);
2134
2135 print_conf(conf);
2136 return count;
2137}
2138
2139static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
2140{
2141 struct r10conf *conf = mddev->private;
2142 int err = -EEXIST;
2143 int mirror;
2144 int first = 0;
2145 int last = conf->geo.raid_disks - 1;
2146
2147 if (mddev->recovery_cp < MaxSector)
2148 /* only hot-add to in-sync arrays, as recovery is
2149 * very different from resync
2150 */
2151 return -EBUSY;
2152 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
2153 return -EINVAL;
2154
2155 if (md_integrity_add_rdev(rdev, mddev))
2156 return -ENXIO;
2157
2158 if (rdev->raid_disk >= 0)
2159 first = last = rdev->raid_disk;
2160
2161 if (rdev->saved_raid_disk >= first &&
2162 rdev->saved_raid_disk < conf->geo.raid_disks &&
2163 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
2164 mirror = rdev->saved_raid_disk;
2165 else
2166 mirror = first;
2167 for ( ; mirror <= last ; mirror++) {
2168 struct raid10_info *p = &conf->mirrors[mirror];
2169 if (p->recovery_disabled == mddev->recovery_disabled)
2170 continue;
2171 if (p->rdev) {
2172 if (!test_bit(WantReplacement, &p->rdev->flags) ||
2173 p->replacement != NULL)
2174 continue;
2175 clear_bit(In_sync, &rdev->flags);
2176 set_bit(Replacement, &rdev->flags);
2177 rdev->raid_disk = mirror;
2178 err = 0;
2179 if (mddev->gendisk)
2180 disk_stack_limits(mddev->gendisk, rdev->bdev,
2181 rdev->data_offset << 9);
2182 conf->fullsync = 1;
2183 rcu_assign_pointer(p->replacement, rdev);
2184 break;
2185 }
2186
2187 if (mddev->gendisk)
2188 disk_stack_limits(mddev->gendisk, rdev->bdev,
2189 rdev->data_offset << 9);
2190
2191 p->head_position = 0;
2192 p->recovery_disabled = mddev->recovery_disabled - 1;
2193 rdev->raid_disk = mirror;
2194 err = 0;
2195 if (rdev->saved_raid_disk != mirror)
2196 conf->fullsync = 1;
2197 rcu_assign_pointer(p->rdev, rdev);
2198 break;
2199 }
2200
2201 print_conf(conf);
2202 return err;
2203}
2204
2205static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
2206{
2207 struct r10conf *conf = mddev->private;
2208 int err = 0;
2209 int number = rdev->raid_disk;
2210 struct md_rdev **rdevp;
2211 struct raid10_info *p;
2212
2213 print_conf(conf);
2214 if (unlikely(number >= mddev->raid_disks))
2215 return 0;
2216 p = conf->mirrors + number;
2217 if (rdev == p->rdev)
2218 rdevp = &p->rdev;
2219 else if (rdev == p->replacement)
2220 rdevp = &p->replacement;
2221 else
2222 return 0;
2223
2224 if (test_bit(In_sync, &rdev->flags) ||
2225 atomic_read(&rdev->nr_pending)) {
2226 err = -EBUSY;
2227 goto abort;
2228 }
2229 /* Only remove non-faulty devices if recovery
2230 * is not possible.
2231 */
2232 if (!test_bit(Faulty, &rdev->flags) &&
2233 mddev->recovery_disabled != p->recovery_disabled &&
2234 (!p->replacement || p->replacement == rdev) &&
2235 number < conf->geo.raid_disks &&
2236 enough(conf, -1)) {
2237 err = -EBUSY;
2238 goto abort;
2239 }
2240 *rdevp = NULL;
2241 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
2242 synchronize_rcu();
2243 if (atomic_read(&rdev->nr_pending)) {
2244 /* lost the race, try later */
2245 err = -EBUSY;
2246 *rdevp = rdev;
2247 goto abort;
2248 }
2249 }
2250 if (p->replacement) {
2251 /* We must have just cleared 'rdev' */
2252 p->rdev = p->replacement;
2253 clear_bit(Replacement, &p->replacement->flags);
2254 smp_mb(); /* Make sure other CPUs may see both as identical
2255 * but will never see neither -- if they are careful.
2256 */
2257 p->replacement = NULL;
2258 }
2259
2260 clear_bit(WantReplacement, &rdev->flags);
2261 err = md_integrity_register(mddev);
2262
2263abort:
2264
2265 print_conf(conf);
2266 return err;
2267}
2268
2269static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
2270{
2271 struct r10conf *conf = r10_bio->mddev->private;
2272
2273 if (!bio->bi_status)
2274 set_bit(R10BIO_Uptodate, &r10_bio->state);
2275 else
2276 /* The write handler will notice the lack of
2277 * R10BIO_Uptodate and record any errors etc
2278 */
2279 atomic_add(r10_bio->sectors,
2280 &conf->mirrors[d].rdev->corrected_errors);
2281
2282 /* for reconstruct, we always reschedule after a read.
2283 * for resync, only after all reads
2284 */
2285 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
2286 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
2287 atomic_dec_and_test(&r10_bio->remaining)) {
2288 /* we have read all the blocks,
2289 * do the comparison in process context in raid10d
2290 */
2291 reschedule_retry(r10_bio);
2292 }
2293}
2294
2295static void end_sync_read(struct bio *bio)
2296{
2297 struct r10bio *r10_bio = get_resync_r10bio(bio);
2298 struct r10conf *conf = r10_bio->mddev->private;
2299 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
2300
2301 __end_sync_read(r10_bio, bio, d);
2302}
2303
2304static void end_reshape_read(struct bio *bio)
2305{
2306 /* reshape read bio isn't allocated from r10buf_pool */
2307 struct r10bio *r10_bio = bio->bi_private;
2308
2309 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
2310}
2311
2312static void end_sync_request(struct r10bio *r10_bio)
2313{
2314 struct mddev *mddev = r10_bio->mddev;
2315
2316 while (atomic_dec_and_test(&r10_bio->remaining)) {
2317 if (r10_bio->master_bio == NULL) {
2318 /* the primary of several recovery bios */
2319 sector_t s = r10_bio->sectors;
2320 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2321 test_bit(R10BIO_WriteError, &r10_bio->state))
2322 reschedule_retry(r10_bio);
2323 else
2324 put_buf(r10_bio);
2325 md_done_sync(mddev, s, 1);
2326 break;
2327 } else {
2328 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
2329 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2330 test_bit(R10BIO_WriteError, &r10_bio->state))
2331 reschedule_retry(r10_bio);
2332 else
2333 put_buf(r10_bio);
2334 r10_bio = r10_bio2;
2335 }
2336 }
2337}
2338
2339static void end_sync_write(struct bio *bio)
2340{
2341 struct r10bio *r10_bio = get_resync_r10bio(bio);
2342 struct mddev *mddev = r10_bio->mddev;
2343 struct r10conf *conf = mddev->private;
2344 int d;
2345 sector_t first_bad;
2346 int bad_sectors;
2347 int slot;
2348 int repl;
2349 struct md_rdev *rdev = NULL;
2350
2351 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2352 if (repl)
2353 rdev = conf->mirrors[d].replacement;
2354 else
2355 rdev = conf->mirrors[d].rdev;
2356
2357 if (bio->bi_status) {
2358 if (repl)
2359 md_error(mddev, rdev);
2360 else {
2361 set_bit(WriteErrorSeen, &rdev->flags);
2362 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2363 set_bit(MD_RECOVERY_NEEDED,
2364 &rdev->mddev->recovery);
2365 set_bit(R10BIO_WriteError, &r10_bio->state);
2366 }
2367 } else if (is_badblock(rdev,
2368 r10_bio->devs[slot].addr,
2369 r10_bio->sectors,
2370 &first_bad, &bad_sectors))
2371 set_bit(R10BIO_MadeGood, &r10_bio->state);
2372
2373 rdev_dec_pending(rdev, mddev);
2374
2375 end_sync_request(r10_bio);
2376}
2377
2378/*
2379 * Note: sync and recover and handled very differently for raid10
2380 * This code is for resync.
2381 * For resync, we read through virtual addresses and read all blocks.
2382 * If there is any error, we schedule a write. The lowest numbered
2383 * drive is authoritative.
2384 * However requests come for physical address, so we need to map.
2385 * For every physical address there are raid_disks/copies virtual addresses,
2386 * which is always are least one, but is not necessarly an integer.
2387 * This means that a physical address can span multiple chunks, so we may
2388 * have to submit multiple io requests for a single sync request.
2389 */
2390/*
2391 * We check if all blocks are in-sync and only write to blocks that
2392 * aren't in sync
2393 */
2394static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2395{
2396 struct r10conf *conf = mddev->private;
2397 int i, first;
2398 struct bio *tbio, *fbio;
2399 int vcnt;
2400 struct page **tpages, **fpages;
2401
2402 atomic_set(&r10_bio->remaining, 1);
2403
2404 /* find the first device with a block */
2405 for (i=0; i<conf->copies; i++)
2406 if (!r10_bio->devs[i].bio->bi_status)
2407 break;
2408
2409 if (i == conf->copies)
2410 goto done;
2411
2412 first = i;
2413 fbio = r10_bio->devs[i].bio;
2414 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2415 fbio->bi_iter.bi_idx = 0;
2416 fpages = get_resync_pages(fbio)->pages;
2417
2418 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2419 /* now find blocks with errors */
2420 for (i=0 ; i < conf->copies ; i++) {
2421 int j, d;
2422 struct md_rdev *rdev;
2423 struct resync_pages *rp;
2424
2425 tbio = r10_bio->devs[i].bio;
2426
2427 if (tbio->bi_end_io != end_sync_read)
2428 continue;
2429 if (i == first)
2430 continue;
2431
2432 tpages = get_resync_pages(tbio)->pages;
2433 d = r10_bio->devs[i].devnum;
2434 rdev = conf->mirrors[d].rdev;
2435 if (!r10_bio->devs[i].bio->bi_status) {
2436 /* We know that the bi_io_vec layout is the same for
2437 * both 'first' and 'i', so we just compare them.
2438 * All vec entries are PAGE_SIZE;
2439 */
2440 int sectors = r10_bio->sectors;
2441 for (j = 0; j < vcnt; j++) {
2442 int len = PAGE_SIZE;
2443 if (sectors < (len / 512))
2444 len = sectors * 512;
2445 if (memcmp(page_address(fpages[j]),
2446 page_address(tpages[j]),
2447 len))
2448 break;
2449 sectors -= len/512;
2450 }
2451 if (j == vcnt)
2452 continue;
2453 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2454 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2455 /* Don't fix anything. */
2456 continue;
2457 } else if (test_bit(FailFast, &rdev->flags)) {
2458 /* Just give up on this device */
2459 md_error(rdev->mddev, rdev);
2460 continue;
2461 }
2462 /* Ok, we need to write this bio, either to correct an
2463 * inconsistency or to correct an unreadable block.
2464 * First we need to fixup bv_offset, bv_len and
2465 * bi_vecs, as the read request might have corrupted these
2466 */
2467 rp = get_resync_pages(tbio);
2468 bio_reset(tbio, conf->mirrors[d].rdev->bdev, REQ_OP_WRITE);
2469
2470 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2471
2472 rp->raid_bio = r10_bio;
2473 tbio->bi_private = rp;
2474 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2475 tbio->bi_end_io = end_sync_write;
2476
2477 bio_copy_data(tbio, fbio);
2478
2479 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2480 atomic_inc(&r10_bio->remaining);
2481 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2482
2483 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2484 tbio->bi_opf |= MD_FAILFAST;
2485 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2486 submit_bio_noacct(tbio);
2487 }
2488
2489 /* Now write out to any replacement devices
2490 * that are active
2491 */
2492 for (i = 0; i < conf->copies; i++) {
2493 int d;
2494
2495 tbio = r10_bio->devs[i].repl_bio;
2496 if (!tbio || !tbio->bi_end_io)
2497 continue;
2498 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2499 && r10_bio->devs[i].bio != fbio)
2500 bio_copy_data(tbio, fbio);
2501 d = r10_bio->devs[i].devnum;
2502 atomic_inc(&r10_bio->remaining);
2503 md_sync_acct(conf->mirrors[d].replacement->bdev,
2504 bio_sectors(tbio));
2505 submit_bio_noacct(tbio);
2506 }
2507
2508done:
2509 if (atomic_dec_and_test(&r10_bio->remaining)) {
2510 md_done_sync(mddev, r10_bio->sectors, 1);
2511 put_buf(r10_bio);
2512 }
2513}
2514
2515/*
2516 * Now for the recovery code.
2517 * Recovery happens across physical sectors.
2518 * We recover all non-is_sync drives by finding the virtual address of
2519 * each, and then choose a working drive that also has that virt address.
2520 * There is a separate r10_bio for each non-in_sync drive.
2521 * Only the first two slots are in use. The first for reading,
2522 * The second for writing.
2523 *
2524 */
2525static void fix_recovery_read_error(struct r10bio *r10_bio)
2526{
2527 /* We got a read error during recovery.
2528 * We repeat the read in smaller page-sized sections.
2529 * If a read succeeds, write it to the new device or record
2530 * a bad block if we cannot.
2531 * If a read fails, record a bad block on both old and
2532 * new devices.
2533 */
2534 struct mddev *mddev = r10_bio->mddev;
2535 struct r10conf *conf = mddev->private;
2536 struct bio *bio = r10_bio->devs[0].bio;
2537 sector_t sect = 0;
2538 int sectors = r10_bio->sectors;
2539 int idx = 0;
2540 int dr = r10_bio->devs[0].devnum;
2541 int dw = r10_bio->devs[1].devnum;
2542 struct page **pages = get_resync_pages(bio)->pages;
2543
2544 while (sectors) {
2545 int s = sectors;
2546 struct md_rdev *rdev;
2547 sector_t addr;
2548 int ok;
2549
2550 if (s > (PAGE_SIZE>>9))
2551 s = PAGE_SIZE >> 9;
2552
2553 rdev = conf->mirrors[dr].rdev;
2554 addr = r10_bio->devs[0].addr + sect,
2555 ok = sync_page_io(rdev,
2556 addr,
2557 s << 9,
2558 pages[idx],
2559 REQ_OP_READ, false);
2560 if (ok) {
2561 rdev = conf->mirrors[dw].rdev;
2562 addr = r10_bio->devs[1].addr + sect;
2563 ok = sync_page_io(rdev,
2564 addr,
2565 s << 9,
2566 pages[idx],
2567 REQ_OP_WRITE, false);
2568 if (!ok) {
2569 set_bit(WriteErrorSeen, &rdev->flags);
2570 if (!test_and_set_bit(WantReplacement,
2571 &rdev->flags))
2572 set_bit(MD_RECOVERY_NEEDED,
2573 &rdev->mddev->recovery);
2574 }
2575 }
2576 if (!ok) {
2577 /* We don't worry if we cannot set a bad block -
2578 * it really is bad so there is no loss in not
2579 * recording it yet
2580 */
2581 rdev_set_badblocks(rdev, addr, s, 0);
2582
2583 if (rdev != conf->mirrors[dw].rdev) {
2584 /* need bad block on destination too */
2585 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2586 addr = r10_bio->devs[1].addr + sect;
2587 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2588 if (!ok) {
2589 /* just abort the recovery */
2590 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2591 mdname(mddev));
2592
2593 conf->mirrors[dw].recovery_disabled
2594 = mddev->recovery_disabled;
2595 set_bit(MD_RECOVERY_INTR,
2596 &mddev->recovery);
2597 break;
2598 }
2599 }
2600 }
2601
2602 sectors -= s;
2603 sect += s;
2604 idx++;
2605 }
2606}
2607
2608static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2609{
2610 struct r10conf *conf = mddev->private;
2611 int d;
2612 struct bio *wbio, *wbio2;
2613
2614 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2615 fix_recovery_read_error(r10_bio);
2616 end_sync_request(r10_bio);
2617 return;
2618 }
2619
2620 /*
2621 * share the pages with the first bio
2622 * and submit the write request
2623 */
2624 d = r10_bio->devs[1].devnum;
2625 wbio = r10_bio->devs[1].bio;
2626 wbio2 = r10_bio->devs[1].repl_bio;
2627 /* Need to test wbio2->bi_end_io before we call
2628 * submit_bio_noacct as if the former is NULL,
2629 * the latter is free to free wbio2.
2630 */
2631 if (wbio2 && !wbio2->bi_end_io)
2632 wbio2 = NULL;
2633 if (wbio->bi_end_io) {
2634 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2635 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2636 submit_bio_noacct(wbio);
2637 }
2638 if (wbio2) {
2639 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2640 md_sync_acct(conf->mirrors[d].replacement->bdev,
2641 bio_sectors(wbio2));
2642 submit_bio_noacct(wbio2);
2643 }
2644}
2645
2646/*
2647 * Used by fix_read_error() to decay the per rdev read_errors.
2648 * We halve the read error count for every hour that has elapsed
2649 * since the last recorded read error.
2650 *
2651 */
2652static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2653{
2654 long cur_time_mon;
2655 unsigned long hours_since_last;
2656 unsigned int read_errors = atomic_read(&rdev->read_errors);
2657
2658 cur_time_mon = ktime_get_seconds();
2659
2660 if (rdev->last_read_error == 0) {
2661 /* first time we've seen a read error */
2662 rdev->last_read_error = cur_time_mon;
2663 return;
2664 }
2665
2666 hours_since_last = (long)(cur_time_mon -
2667 rdev->last_read_error) / 3600;
2668
2669 rdev->last_read_error = cur_time_mon;
2670
2671 /*
2672 * if hours_since_last is > the number of bits in read_errors
2673 * just set read errors to 0. We do this to avoid
2674 * overflowing the shift of read_errors by hours_since_last.
2675 */
2676 if (hours_since_last >= 8 * sizeof(read_errors))
2677 atomic_set(&rdev->read_errors, 0);
2678 else
2679 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2680}
2681
2682static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2683 int sectors, struct page *page, enum req_op op)
2684{
2685 sector_t first_bad;
2686 int bad_sectors;
2687
2688 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2689 && (op == REQ_OP_READ || test_bit(WriteErrorSeen, &rdev->flags)))
2690 return -1;
2691 if (sync_page_io(rdev, sector, sectors << 9, page, op, false))
2692 /* success */
2693 return 1;
2694 if (op == REQ_OP_WRITE) {
2695 set_bit(WriteErrorSeen, &rdev->flags);
2696 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2697 set_bit(MD_RECOVERY_NEEDED,
2698 &rdev->mddev->recovery);
2699 }
2700 /* need to record an error - either for the block or the device */
2701 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2702 md_error(rdev->mddev, rdev);
2703 return 0;
2704}
2705
2706/*
2707 * This is a kernel thread which:
2708 *
2709 * 1. Retries failed read operations on working mirrors.
2710 * 2. Updates the raid superblock when problems encounter.
2711 * 3. Performs writes following reads for array synchronising.
2712 */
2713
2714static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2715{
2716 int sect = 0; /* Offset from r10_bio->sector */
2717 int sectors = r10_bio->sectors;
2718 struct md_rdev *rdev;
2719 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2720 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2721
2722 /* still own a reference to this rdev, so it cannot
2723 * have been cleared recently.
2724 */
2725 rdev = conf->mirrors[d].rdev;
2726
2727 if (test_bit(Faulty, &rdev->flags))
2728 /* drive has already been failed, just ignore any
2729 more fix_read_error() attempts */
2730 return;
2731
2732 check_decay_read_errors(mddev, rdev);
2733 atomic_inc(&rdev->read_errors);
2734 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2735 pr_notice("md/raid10:%s: %pg: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2736 mdname(mddev), rdev->bdev,
2737 atomic_read(&rdev->read_errors), max_read_errors);
2738 pr_notice("md/raid10:%s: %pg: Failing raid device\n",
2739 mdname(mddev), rdev->bdev);
2740 md_error(mddev, rdev);
2741 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2742 return;
2743 }
2744
2745 while(sectors) {
2746 int s = sectors;
2747 int sl = r10_bio->read_slot;
2748 int success = 0;
2749 int start;
2750
2751 if (s > (PAGE_SIZE>>9))
2752 s = PAGE_SIZE >> 9;
2753
2754 rcu_read_lock();
2755 do {
2756 sector_t first_bad;
2757 int bad_sectors;
2758
2759 d = r10_bio->devs[sl].devnum;
2760 rdev = rcu_dereference(conf->mirrors[d].rdev);
2761 if (rdev &&
2762 test_bit(In_sync, &rdev->flags) &&
2763 !test_bit(Faulty, &rdev->flags) &&
2764 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2765 &first_bad, &bad_sectors) == 0) {
2766 atomic_inc(&rdev->nr_pending);
2767 rcu_read_unlock();
2768 success = sync_page_io(rdev,
2769 r10_bio->devs[sl].addr +
2770 sect,
2771 s<<9,
2772 conf->tmppage,
2773 REQ_OP_READ, false);
2774 rdev_dec_pending(rdev, mddev);
2775 rcu_read_lock();
2776 if (success)
2777 break;
2778 }
2779 sl++;
2780 if (sl == conf->copies)
2781 sl = 0;
2782 } while (!success && sl != r10_bio->read_slot);
2783 rcu_read_unlock();
2784
2785 if (!success) {
2786 /* Cannot read from anywhere, just mark the block
2787 * as bad on the first device to discourage future
2788 * reads.
2789 */
2790 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2791 rdev = conf->mirrors[dn].rdev;
2792
2793 if (!rdev_set_badblocks(
2794 rdev,
2795 r10_bio->devs[r10_bio->read_slot].addr
2796 + sect,
2797 s, 0)) {
2798 md_error(mddev, rdev);
2799 r10_bio->devs[r10_bio->read_slot].bio
2800 = IO_BLOCKED;
2801 }
2802 break;
2803 }
2804
2805 start = sl;
2806 /* write it back and re-read */
2807 rcu_read_lock();
2808 while (sl != r10_bio->read_slot) {
2809 if (sl==0)
2810 sl = conf->copies;
2811 sl--;
2812 d = r10_bio->devs[sl].devnum;
2813 rdev = rcu_dereference(conf->mirrors[d].rdev);
2814 if (!rdev ||
2815 test_bit(Faulty, &rdev->flags) ||
2816 !test_bit(In_sync, &rdev->flags))
2817 continue;
2818
2819 atomic_inc(&rdev->nr_pending);
2820 rcu_read_unlock();
2821 if (r10_sync_page_io(rdev,
2822 r10_bio->devs[sl].addr +
2823 sect,
2824 s, conf->tmppage, REQ_OP_WRITE)
2825 == 0) {
2826 /* Well, this device is dead */
2827 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %pg)\n",
2828 mdname(mddev), s,
2829 (unsigned long long)(
2830 sect +
2831 choose_data_offset(r10_bio,
2832 rdev)),
2833 rdev->bdev);
2834 pr_notice("md/raid10:%s: %pg: failing drive\n",
2835 mdname(mddev),
2836 rdev->bdev);
2837 }
2838 rdev_dec_pending(rdev, mddev);
2839 rcu_read_lock();
2840 }
2841 sl = start;
2842 while (sl != r10_bio->read_slot) {
2843 if (sl==0)
2844 sl = conf->copies;
2845 sl--;
2846 d = r10_bio->devs[sl].devnum;
2847 rdev = rcu_dereference(conf->mirrors[d].rdev);
2848 if (!rdev ||
2849 test_bit(Faulty, &rdev->flags) ||
2850 !test_bit(In_sync, &rdev->flags))
2851 continue;
2852
2853 atomic_inc(&rdev->nr_pending);
2854 rcu_read_unlock();
2855 switch (r10_sync_page_io(rdev,
2856 r10_bio->devs[sl].addr +
2857 sect,
2858 s, conf->tmppage, REQ_OP_READ)) {
2859 case 0:
2860 /* Well, this device is dead */
2861 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %pg)\n",
2862 mdname(mddev), s,
2863 (unsigned long long)(
2864 sect +
2865 choose_data_offset(r10_bio, rdev)),
2866 rdev->bdev);
2867 pr_notice("md/raid10:%s: %pg: failing drive\n",
2868 mdname(mddev),
2869 rdev->bdev);
2870 break;
2871 case 1:
2872 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %pg)\n",
2873 mdname(mddev), s,
2874 (unsigned long long)(
2875 sect +
2876 choose_data_offset(r10_bio, rdev)),
2877 rdev->bdev);
2878 atomic_add(s, &rdev->corrected_errors);
2879 }
2880
2881 rdev_dec_pending(rdev, mddev);
2882 rcu_read_lock();
2883 }
2884 rcu_read_unlock();
2885
2886 sectors -= s;
2887 sect += s;
2888 }
2889}
2890
2891static int narrow_write_error(struct r10bio *r10_bio, int i)
2892{
2893 struct bio *bio = r10_bio->master_bio;
2894 struct mddev *mddev = r10_bio->mddev;
2895 struct r10conf *conf = mddev->private;
2896 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2897 /* bio has the data to be written to slot 'i' where
2898 * we just recently had a write error.
2899 * We repeatedly clone the bio and trim down to one block,
2900 * then try the write. Where the write fails we record
2901 * a bad block.
2902 * It is conceivable that the bio doesn't exactly align with
2903 * blocks. We must handle this.
2904 *
2905 * We currently own a reference to the rdev.
2906 */
2907
2908 int block_sectors;
2909 sector_t sector;
2910 int sectors;
2911 int sect_to_write = r10_bio->sectors;
2912 int ok = 1;
2913
2914 if (rdev->badblocks.shift < 0)
2915 return 0;
2916
2917 block_sectors = roundup(1 << rdev->badblocks.shift,
2918 bdev_logical_block_size(rdev->bdev) >> 9);
2919 sector = r10_bio->sector;
2920 sectors = ((r10_bio->sector + block_sectors)
2921 & ~(sector_t)(block_sectors - 1))
2922 - sector;
2923
2924 while (sect_to_write) {
2925 struct bio *wbio;
2926 sector_t wsector;
2927 if (sectors > sect_to_write)
2928 sectors = sect_to_write;
2929 /* Write at 'sector' for 'sectors' */
2930 wbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
2931 &mddev->bio_set);
2932 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2933 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2934 wbio->bi_iter.bi_sector = wsector +
2935 choose_data_offset(r10_bio, rdev);
2936 wbio->bi_opf = REQ_OP_WRITE;
2937
2938 if (submit_bio_wait(wbio) < 0)
2939 /* Failure! */
2940 ok = rdev_set_badblocks(rdev, wsector,
2941 sectors, 0)
2942 && ok;
2943
2944 bio_put(wbio);
2945 sect_to_write -= sectors;
2946 sector += sectors;
2947 sectors = block_sectors;
2948 }
2949 return ok;
2950}
2951
2952static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2953{
2954 int slot = r10_bio->read_slot;
2955 struct bio *bio;
2956 struct r10conf *conf = mddev->private;
2957 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2958
2959 /* we got a read error. Maybe the drive is bad. Maybe just
2960 * the block and we can fix it.
2961 * We freeze all other IO, and try reading the block from
2962 * other devices. When we find one, we re-write
2963 * and check it that fixes the read error.
2964 * This is all done synchronously while the array is
2965 * frozen.
2966 */
2967 bio = r10_bio->devs[slot].bio;
2968 bio_put(bio);
2969 r10_bio->devs[slot].bio = NULL;
2970
2971 if (mddev->ro)
2972 r10_bio->devs[slot].bio = IO_BLOCKED;
2973 else if (!test_bit(FailFast, &rdev->flags)) {
2974 freeze_array(conf, 1);
2975 fix_read_error(conf, mddev, r10_bio);
2976 unfreeze_array(conf);
2977 } else
2978 md_error(mddev, rdev);
2979
2980 rdev_dec_pending(rdev, mddev);
2981 allow_barrier(conf);
2982 r10_bio->state = 0;
2983 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2984}
2985
2986static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2987{
2988 /* Some sort of write request has finished and it
2989 * succeeded in writing where we thought there was a
2990 * bad block. So forget the bad block.
2991 * Or possibly if failed and we need to record
2992 * a bad block.
2993 */
2994 int m;
2995 struct md_rdev *rdev;
2996
2997 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2998 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2999 for (m = 0; m < conf->copies; m++) {
3000 int dev = r10_bio->devs[m].devnum;
3001 rdev = conf->mirrors[dev].rdev;
3002 if (r10_bio->devs[m].bio == NULL ||
3003 r10_bio->devs[m].bio->bi_end_io == NULL)
3004 continue;
3005 if (!r10_bio->devs[m].bio->bi_status) {
3006 rdev_clear_badblocks(
3007 rdev,
3008 r10_bio->devs[m].addr,
3009 r10_bio->sectors, 0);
3010 } else {
3011 if (!rdev_set_badblocks(
3012 rdev,
3013 r10_bio->devs[m].addr,
3014 r10_bio->sectors, 0))
3015 md_error(conf->mddev, rdev);
3016 }
3017 rdev = conf->mirrors[dev].replacement;
3018 if (r10_bio->devs[m].repl_bio == NULL ||
3019 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
3020 continue;
3021
3022 if (!r10_bio->devs[m].repl_bio->bi_status) {
3023 rdev_clear_badblocks(
3024 rdev,
3025 r10_bio->devs[m].addr,
3026 r10_bio->sectors, 0);
3027 } else {
3028 if (!rdev_set_badblocks(
3029 rdev,
3030 r10_bio->devs[m].addr,
3031 r10_bio->sectors, 0))
3032 md_error(conf->mddev, rdev);
3033 }
3034 }
3035 put_buf(r10_bio);
3036 } else {
3037 bool fail = false;
3038 for (m = 0; m < conf->copies; m++) {
3039 int dev = r10_bio->devs[m].devnum;
3040 struct bio *bio = r10_bio->devs[m].bio;
3041 rdev = conf->mirrors[dev].rdev;
3042 if (bio == IO_MADE_GOOD) {
3043 rdev_clear_badblocks(
3044 rdev,
3045 r10_bio->devs[m].addr,
3046 r10_bio->sectors, 0);
3047 rdev_dec_pending(rdev, conf->mddev);
3048 } else if (bio != NULL && bio->bi_status) {
3049 fail = true;
3050 if (!narrow_write_error(r10_bio, m)) {
3051 md_error(conf->mddev, rdev);
3052 set_bit(R10BIO_Degraded,
3053 &r10_bio->state);
3054 }
3055 rdev_dec_pending(rdev, conf->mddev);
3056 }
3057 bio = r10_bio->devs[m].repl_bio;
3058 rdev = conf->mirrors[dev].replacement;
3059 if (rdev && bio == IO_MADE_GOOD) {
3060 rdev_clear_badblocks(
3061 rdev,
3062 r10_bio->devs[m].addr,
3063 r10_bio->sectors, 0);
3064 rdev_dec_pending(rdev, conf->mddev);
3065 }
3066 }
3067 if (fail) {
3068 spin_lock_irq(&conf->device_lock);
3069 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
3070 conf->nr_queued++;
3071 spin_unlock_irq(&conf->device_lock);
3072 /*
3073 * In case freeze_array() is waiting for condition
3074 * nr_pending == nr_queued + extra to be true.
3075 */
3076 wake_up(&conf->wait_barrier);
3077 md_wakeup_thread(conf->mddev->thread);
3078 } else {
3079 if (test_bit(R10BIO_WriteError,
3080 &r10_bio->state))
3081 close_write(r10_bio);
3082 raid_end_bio_io(r10_bio);
3083 }
3084 }
3085}
3086
3087static void raid10d(struct md_thread *thread)
3088{
3089 struct mddev *mddev = thread->mddev;
3090 struct r10bio *r10_bio;
3091 unsigned long flags;
3092 struct r10conf *conf = mddev->private;
3093 struct list_head *head = &conf->retry_list;
3094 struct blk_plug plug;
3095
3096 md_check_recovery(mddev);
3097
3098 if (!list_empty_careful(&conf->bio_end_io_list) &&
3099 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3100 LIST_HEAD(tmp);
3101 spin_lock_irqsave(&conf->device_lock, flags);
3102 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
3103 while (!list_empty(&conf->bio_end_io_list)) {
3104 list_move(conf->bio_end_io_list.prev, &tmp);
3105 conf->nr_queued--;
3106 }
3107 }
3108 spin_unlock_irqrestore(&conf->device_lock, flags);
3109 while (!list_empty(&tmp)) {
3110 r10_bio = list_first_entry(&tmp, struct r10bio,
3111 retry_list);
3112 list_del(&r10_bio->retry_list);
3113 if (mddev->degraded)
3114 set_bit(R10BIO_Degraded, &r10_bio->state);
3115
3116 if (test_bit(R10BIO_WriteError,
3117 &r10_bio->state))
3118 close_write(r10_bio);
3119 raid_end_bio_io(r10_bio);
3120 }
3121 }
3122
3123 blk_start_plug(&plug);
3124 for (;;) {
3125
3126 flush_pending_writes(conf);
3127
3128 spin_lock_irqsave(&conf->device_lock, flags);
3129 if (list_empty(head)) {
3130 spin_unlock_irqrestore(&conf->device_lock, flags);
3131 break;
3132 }
3133 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
3134 list_del(head->prev);
3135 conf->nr_queued--;
3136 spin_unlock_irqrestore(&conf->device_lock, flags);
3137
3138 mddev = r10_bio->mddev;
3139 conf = mddev->private;
3140 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
3141 test_bit(R10BIO_WriteError, &r10_bio->state))
3142 handle_write_completed(conf, r10_bio);
3143 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
3144 reshape_request_write(mddev, r10_bio);
3145 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
3146 sync_request_write(mddev, r10_bio);
3147 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
3148 recovery_request_write(mddev, r10_bio);
3149 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
3150 handle_read_error(mddev, r10_bio);
3151 else
3152 WARN_ON_ONCE(1);
3153
3154 cond_resched();
3155 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
3156 md_check_recovery(mddev);
3157 }
3158 blk_finish_plug(&plug);
3159}
3160
3161static int init_resync(struct r10conf *conf)
3162{
3163 int ret, buffs, i;
3164
3165 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
3166 BUG_ON(mempool_initialized(&conf->r10buf_pool));
3167 conf->have_replacement = 0;
3168 for (i = 0; i < conf->geo.raid_disks; i++)
3169 if (conf->mirrors[i].replacement)
3170 conf->have_replacement = 1;
3171 ret = mempool_init(&conf->r10buf_pool, buffs,
3172 r10buf_pool_alloc, r10buf_pool_free, conf);
3173 if (ret)
3174 return ret;
3175 conf->next_resync = 0;
3176 return 0;
3177}
3178
3179static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
3180{
3181 struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
3182 struct rsync_pages *rp;
3183 struct bio *bio;
3184 int nalloc;
3185 int i;
3186
3187 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
3188 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
3189 nalloc = conf->copies; /* resync */
3190 else
3191 nalloc = 2; /* recovery */
3192
3193 for (i = 0; i < nalloc; i++) {
3194 bio = r10bio->devs[i].bio;
3195 rp = bio->bi_private;
3196 bio_reset(bio, NULL, 0);
3197 bio->bi_private = rp;
3198 bio = r10bio->devs[i].repl_bio;
3199 if (bio) {
3200 rp = bio->bi_private;
3201 bio_reset(bio, NULL, 0);
3202 bio->bi_private = rp;
3203 }
3204 }
3205 return r10bio;
3206}
3207
3208/*
3209 * Set cluster_sync_high since we need other nodes to add the
3210 * range [cluster_sync_low, cluster_sync_high] to suspend list.
3211 */
3212static void raid10_set_cluster_sync_high(struct r10conf *conf)
3213{
3214 sector_t window_size;
3215 int extra_chunk, chunks;
3216
3217 /*
3218 * First, here we define "stripe" as a unit which across
3219 * all member devices one time, so we get chunks by use
3220 * raid_disks / near_copies. Otherwise, if near_copies is
3221 * close to raid_disks, then resync window could increases
3222 * linearly with the increase of raid_disks, which means
3223 * we will suspend a really large IO window while it is not
3224 * necessary. If raid_disks is not divisible by near_copies,
3225 * an extra chunk is needed to ensure the whole "stripe" is
3226 * covered.
3227 */
3228
3229 chunks = conf->geo.raid_disks / conf->geo.near_copies;
3230 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
3231 extra_chunk = 0;
3232 else
3233 extra_chunk = 1;
3234 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
3235
3236 /*
3237 * At least use a 32M window to align with raid1's resync window
3238 */
3239 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
3240 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
3241
3242 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
3243}
3244
3245/*
3246 * perform a "sync" on one "block"
3247 *
3248 * We need to make sure that no normal I/O request - particularly write
3249 * requests - conflict with active sync requests.
3250 *
3251 * This is achieved by tracking pending requests and a 'barrier' concept
3252 * that can be installed to exclude normal IO requests.
3253 *
3254 * Resync and recovery are handled very differently.
3255 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
3256 *
3257 * For resync, we iterate over virtual addresses, read all copies,
3258 * and update if there are differences. If only one copy is live,
3259 * skip it.
3260 * For recovery, we iterate over physical addresses, read a good
3261 * value for each non-in_sync drive, and over-write.
3262 *
3263 * So, for recovery we may have several outstanding complex requests for a
3264 * given address, one for each out-of-sync device. We model this by allocating
3265 * a number of r10_bio structures, one for each out-of-sync device.
3266 * As we setup these structures, we collect all bio's together into a list
3267 * which we then process collectively to add pages, and then process again
3268 * to pass to submit_bio_noacct.
3269 *
3270 * The r10_bio structures are linked using a borrowed master_bio pointer.
3271 * This link is counted in ->remaining. When the r10_bio that points to NULL
3272 * has its remaining count decremented to 0, the whole complex operation
3273 * is complete.
3274 *
3275 */
3276
3277static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
3278 int *skipped)
3279{
3280 struct r10conf *conf = mddev->private;
3281 struct r10bio *r10_bio;
3282 struct bio *biolist = NULL, *bio;
3283 sector_t max_sector, nr_sectors;
3284 int i;
3285 int max_sync;
3286 sector_t sync_blocks;
3287 sector_t sectors_skipped = 0;
3288 int chunks_skipped = 0;
3289 sector_t chunk_mask = conf->geo.chunk_mask;
3290 int page_idx = 0;
3291
3292 if (!mempool_initialized(&conf->r10buf_pool))
3293 if (init_resync(conf))
3294 return 0;
3295
3296 /*
3297 * Allow skipping a full rebuild for incremental assembly
3298 * of a clean array, like RAID1 does.
3299 */
3300 if (mddev->bitmap == NULL &&
3301 mddev->recovery_cp == MaxSector &&
3302 mddev->reshape_position == MaxSector &&
3303 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
3304 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3305 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
3306 conf->fullsync == 0) {
3307 *skipped = 1;
3308 return mddev->dev_sectors - sector_nr;
3309 }
3310
3311 skipped:
3312 max_sector = mddev->dev_sectors;
3313 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
3314 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3315 max_sector = mddev->resync_max_sectors;
3316 if (sector_nr >= max_sector) {
3317 conf->cluster_sync_low = 0;
3318 conf->cluster_sync_high = 0;
3319
3320 /* If we aborted, we need to abort the
3321 * sync on the 'current' bitmap chucks (there can
3322 * be several when recovering multiple devices).
3323 * as we may have started syncing it but not finished.
3324 * We can find the current address in
3325 * mddev->curr_resync, but for recovery,
3326 * we need to convert that to several
3327 * virtual addresses.
3328 */
3329 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3330 end_reshape(conf);
3331 close_sync(conf);
3332 return 0;
3333 }
3334
3335 if (mddev->curr_resync < max_sector) { /* aborted */
3336 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
3337 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3338 &sync_blocks, 1);
3339 else for (i = 0; i < conf->geo.raid_disks; i++) {
3340 sector_t sect =
3341 raid10_find_virt(conf, mddev->curr_resync, i);
3342 md_bitmap_end_sync(mddev->bitmap, sect,
3343 &sync_blocks, 1);
3344 }
3345 } else {
3346 /* completed sync */
3347 if ((!mddev->bitmap || conf->fullsync)
3348 && conf->have_replacement
3349 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3350 /* Completed a full sync so the replacements
3351 * are now fully recovered.
3352 */
3353 rcu_read_lock();
3354 for (i = 0; i < conf->geo.raid_disks; i++) {
3355 struct md_rdev *rdev =
3356 rcu_dereference(conf->mirrors[i].replacement);
3357 if (rdev)
3358 rdev->recovery_offset = MaxSector;
3359 }
3360 rcu_read_unlock();
3361 }
3362 conf->fullsync = 0;
3363 }
3364 md_bitmap_close_sync(mddev->bitmap);
3365 close_sync(conf);
3366 *skipped = 1;
3367 return sectors_skipped;
3368 }
3369
3370 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3371 return reshape_request(mddev, sector_nr, skipped);
3372
3373 if (chunks_skipped >= conf->geo.raid_disks) {
3374 /* if there has been nothing to do on any drive,
3375 * then there is nothing to do at all..
3376 */
3377 *skipped = 1;
3378 return (max_sector - sector_nr) + sectors_skipped;
3379 }
3380
3381 if (max_sector > mddev->resync_max)
3382 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3383
3384 /* make sure whole request will fit in a chunk - if chunks
3385 * are meaningful
3386 */
3387 if (conf->geo.near_copies < conf->geo.raid_disks &&
3388 max_sector > (sector_nr | chunk_mask))
3389 max_sector = (sector_nr | chunk_mask) + 1;
3390
3391 /*
3392 * If there is non-resync activity waiting for a turn, then let it
3393 * though before starting on this new sync request.
3394 */
3395 if (conf->nr_waiting)
3396 schedule_timeout_uninterruptible(1);
3397
3398 /* Again, very different code for resync and recovery.
3399 * Both must result in an r10bio with a list of bios that
3400 * have bi_end_io, bi_sector, bi_bdev set,
3401 * and bi_private set to the r10bio.
3402 * For recovery, we may actually create several r10bios
3403 * with 2 bios in each, that correspond to the bios in the main one.
3404 * In this case, the subordinate r10bios link back through a
3405 * borrowed master_bio pointer, and the counter in the master
3406 * includes a ref from each subordinate.
3407 */
3408 /* First, we decide what to do and set ->bi_end_io
3409 * To end_sync_read if we want to read, and
3410 * end_sync_write if we will want to write.
3411 */
3412
3413 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3414 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3415 /* recovery... the complicated one */
3416 int j;
3417 r10_bio = NULL;
3418
3419 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3420 int still_degraded;
3421 struct r10bio *rb2;
3422 sector_t sect;
3423 int must_sync;
3424 int any_working;
3425 int need_recover = 0;
3426 int need_replace = 0;
3427 struct raid10_info *mirror = &conf->mirrors[i];
3428 struct md_rdev *mrdev, *mreplace;
3429
3430 rcu_read_lock();
3431 mrdev = rcu_dereference(mirror->rdev);
3432 mreplace = rcu_dereference(mirror->replacement);
3433
3434 if (mrdev != NULL &&
3435 !test_bit(Faulty, &mrdev->flags) &&
3436 !test_bit(In_sync, &mrdev->flags))
3437 need_recover = 1;
3438 if (mreplace != NULL &&
3439 !test_bit(Faulty, &mreplace->flags))
3440 need_replace = 1;
3441
3442 if (!need_recover && !need_replace) {
3443 rcu_read_unlock();
3444 continue;
3445 }
3446
3447 still_degraded = 0;
3448 /* want to reconstruct this device */
3449 rb2 = r10_bio;
3450 sect = raid10_find_virt(conf, sector_nr, i);
3451 if (sect >= mddev->resync_max_sectors) {
3452 /* last stripe is not complete - don't
3453 * try to recover this sector.
3454 */
3455 rcu_read_unlock();
3456 continue;
3457 }
3458 if (mreplace && test_bit(Faulty, &mreplace->flags))
3459 mreplace = NULL;
3460 /* Unless we are doing a full sync, or a replacement
3461 * we only need to recover the block if it is set in
3462 * the bitmap
3463 */
3464 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3465 &sync_blocks, 1);
3466 if (sync_blocks < max_sync)
3467 max_sync = sync_blocks;
3468 if (!must_sync &&
3469 mreplace == NULL &&
3470 !conf->fullsync) {
3471 /* yep, skip the sync_blocks here, but don't assume
3472 * that there will never be anything to do here
3473 */
3474 chunks_skipped = -1;
3475 rcu_read_unlock();
3476 continue;
3477 }
3478 atomic_inc(&mrdev->nr_pending);
3479 if (mreplace)
3480 atomic_inc(&mreplace->nr_pending);
3481 rcu_read_unlock();
3482
3483 r10_bio = raid10_alloc_init_r10buf(conf);
3484 r10_bio->state = 0;
3485 raise_barrier(conf, rb2 != NULL);
3486 atomic_set(&r10_bio->remaining, 0);
3487
3488 r10_bio->master_bio = (struct bio*)rb2;
3489 if (rb2)
3490 atomic_inc(&rb2->remaining);
3491 r10_bio->mddev = mddev;
3492 set_bit(R10BIO_IsRecover, &r10_bio->state);
3493 r10_bio->sector = sect;
3494
3495 raid10_find_phys(conf, r10_bio);
3496
3497 /* Need to check if the array will still be
3498 * degraded
3499 */
3500 rcu_read_lock();
3501 for (j = 0; j < conf->geo.raid_disks; j++) {
3502 struct md_rdev *rdev = rcu_dereference(
3503 conf->mirrors[j].rdev);
3504 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3505 still_degraded = 1;
3506 break;
3507 }
3508 }
3509
3510 must_sync = md_bitmap_start_sync(mddev->bitmap, sect,
3511 &sync_blocks, still_degraded);
3512
3513 any_working = 0;
3514 for (j=0; j<conf->copies;j++) {
3515 int k;
3516 int d = r10_bio->devs[j].devnum;
3517 sector_t from_addr, to_addr;
3518 struct md_rdev *rdev =
3519 rcu_dereference(conf->mirrors[d].rdev);
3520 sector_t sector, first_bad;
3521 int bad_sectors;
3522 if (!rdev ||
3523 !test_bit(In_sync, &rdev->flags))
3524 continue;
3525 /* This is where we read from */
3526 any_working = 1;
3527 sector = r10_bio->devs[j].addr;
3528
3529 if (is_badblock(rdev, sector, max_sync,
3530 &first_bad, &bad_sectors)) {
3531 if (first_bad > sector)
3532 max_sync = first_bad - sector;
3533 else {
3534 bad_sectors -= (sector
3535 - first_bad);
3536 if (max_sync > bad_sectors)
3537 max_sync = bad_sectors;
3538 continue;
3539 }
3540 }
3541 bio = r10_bio->devs[0].bio;
3542 bio->bi_next = biolist;
3543 biolist = bio;
3544 bio->bi_end_io = end_sync_read;
3545 bio->bi_opf = REQ_OP_READ;
3546 if (test_bit(FailFast, &rdev->flags))
3547 bio->bi_opf |= MD_FAILFAST;
3548 from_addr = r10_bio->devs[j].addr;
3549 bio->bi_iter.bi_sector = from_addr +
3550 rdev->data_offset;
3551 bio_set_dev(bio, rdev->bdev);
3552 atomic_inc(&rdev->nr_pending);
3553 /* and we write to 'i' (if not in_sync) */
3554
3555 for (k=0; k<conf->copies; k++)
3556 if (r10_bio->devs[k].devnum == i)
3557 break;
3558 BUG_ON(k == conf->copies);
3559 to_addr = r10_bio->devs[k].addr;
3560 r10_bio->devs[0].devnum = d;
3561 r10_bio->devs[0].addr = from_addr;
3562 r10_bio->devs[1].devnum = i;
3563 r10_bio->devs[1].addr = to_addr;
3564
3565 if (need_recover) {
3566 bio = r10_bio->devs[1].bio;
3567 bio->bi_next = biolist;
3568 biolist = bio;
3569 bio->bi_end_io = end_sync_write;
3570 bio->bi_opf = REQ_OP_WRITE;
3571 bio->bi_iter.bi_sector = to_addr
3572 + mrdev->data_offset;
3573 bio_set_dev(bio, mrdev->bdev);
3574 atomic_inc(&r10_bio->remaining);
3575 } else
3576 r10_bio->devs[1].bio->bi_end_io = NULL;
3577
3578 /* and maybe write to replacement */
3579 bio = r10_bio->devs[1].repl_bio;
3580 if (bio)
3581 bio->bi_end_io = NULL;
3582 /* Note: if need_replace, then bio
3583 * cannot be NULL as r10buf_pool_alloc will
3584 * have allocated it.
3585 */
3586 if (!need_replace)
3587 break;
3588 bio->bi_next = biolist;
3589 biolist = bio;
3590 bio->bi_end_io = end_sync_write;
3591 bio->bi_opf = REQ_OP_WRITE;
3592 bio->bi_iter.bi_sector = to_addr +
3593 mreplace->data_offset;
3594 bio_set_dev(bio, mreplace->bdev);
3595 atomic_inc(&r10_bio->remaining);
3596 break;
3597 }
3598 rcu_read_unlock();
3599 if (j == conf->copies) {
3600 /* Cannot recover, so abort the recovery or
3601 * record a bad block */
3602 if (any_working) {
3603 /* problem is that there are bad blocks
3604 * on other device(s)
3605 */
3606 int k;
3607 for (k = 0; k < conf->copies; k++)
3608 if (r10_bio->devs[k].devnum == i)
3609 break;
3610 if (!test_bit(In_sync,
3611 &mrdev->flags)
3612 && !rdev_set_badblocks(
3613 mrdev,
3614 r10_bio->devs[k].addr,
3615 max_sync, 0))
3616 any_working = 0;
3617 if (mreplace &&
3618 !rdev_set_badblocks(
3619 mreplace,
3620 r10_bio->devs[k].addr,
3621 max_sync, 0))
3622 any_working = 0;
3623 }
3624 if (!any_working) {
3625 if (!test_and_set_bit(MD_RECOVERY_INTR,
3626 &mddev->recovery))
3627 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3628 mdname(mddev));
3629 mirror->recovery_disabled
3630 = mddev->recovery_disabled;
3631 }
3632 put_buf(r10_bio);
3633 if (rb2)
3634 atomic_dec(&rb2->remaining);
3635 r10_bio = rb2;
3636 rdev_dec_pending(mrdev, mddev);
3637 if (mreplace)
3638 rdev_dec_pending(mreplace, mddev);
3639 break;
3640 }
3641 rdev_dec_pending(mrdev, mddev);
3642 if (mreplace)
3643 rdev_dec_pending(mreplace, mddev);
3644 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3645 /* Only want this if there is elsewhere to
3646 * read from. 'j' is currently the first
3647 * readable copy.
3648 */
3649 int targets = 1;
3650 for (; j < conf->copies; j++) {
3651 int d = r10_bio->devs[j].devnum;
3652 if (conf->mirrors[d].rdev &&
3653 test_bit(In_sync,
3654 &conf->mirrors[d].rdev->flags))
3655 targets++;
3656 }
3657 if (targets == 1)
3658 r10_bio->devs[0].bio->bi_opf
3659 &= ~MD_FAILFAST;
3660 }
3661 }
3662 if (biolist == NULL) {
3663 while (r10_bio) {
3664 struct r10bio *rb2 = r10_bio;
3665 r10_bio = (struct r10bio*) rb2->master_bio;
3666 rb2->master_bio = NULL;
3667 put_buf(rb2);
3668 }
3669 goto giveup;
3670 }
3671 } else {
3672 /* resync. Schedule a read for every block at this virt offset */
3673 int count = 0;
3674
3675 /*
3676 * Since curr_resync_completed could probably not update in
3677 * time, and we will set cluster_sync_low based on it.
3678 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3679 * safety reason, which ensures curr_resync_completed is
3680 * updated in bitmap_cond_end_sync.
3681 */
3682 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3683 mddev_is_clustered(mddev) &&
3684 (sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3685
3686 if (!md_bitmap_start_sync(mddev->bitmap, sector_nr,
3687 &sync_blocks, mddev->degraded) &&
3688 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3689 &mddev->recovery)) {
3690 /* We can skip this block */
3691 *skipped = 1;
3692 return sync_blocks + sectors_skipped;
3693 }
3694 if (sync_blocks < max_sync)
3695 max_sync = sync_blocks;
3696 r10_bio = raid10_alloc_init_r10buf(conf);
3697 r10_bio->state = 0;
3698
3699 r10_bio->mddev = mddev;
3700 atomic_set(&r10_bio->remaining, 0);
3701 raise_barrier(conf, 0);
3702 conf->next_resync = sector_nr;
3703
3704 r10_bio->master_bio = NULL;
3705 r10_bio->sector = sector_nr;
3706 set_bit(R10BIO_IsSync, &r10_bio->state);
3707 raid10_find_phys(conf, r10_bio);
3708 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3709
3710 for (i = 0; i < conf->copies; i++) {
3711 int d = r10_bio->devs[i].devnum;
3712 sector_t first_bad, sector;
3713 int bad_sectors;
3714 struct md_rdev *rdev;
3715
3716 if (r10_bio->devs[i].repl_bio)
3717 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3718
3719 bio = r10_bio->devs[i].bio;
3720 bio->bi_status = BLK_STS_IOERR;
3721 rcu_read_lock();
3722 rdev = rcu_dereference(conf->mirrors[d].rdev);
3723 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3724 rcu_read_unlock();
3725 continue;
3726 }
3727 sector = r10_bio->devs[i].addr;
3728 if (is_badblock(rdev, sector, max_sync,
3729 &first_bad, &bad_sectors)) {
3730 if (first_bad > sector)
3731 max_sync = first_bad - sector;
3732 else {
3733 bad_sectors -= (sector - first_bad);
3734 if (max_sync > bad_sectors)
3735 max_sync = bad_sectors;
3736 rcu_read_unlock();
3737 continue;
3738 }
3739 }
3740 atomic_inc(&rdev->nr_pending);
3741 atomic_inc(&r10_bio->remaining);
3742 bio->bi_next = biolist;
3743 biolist = bio;
3744 bio->bi_end_io = end_sync_read;
3745 bio->bi_opf = REQ_OP_READ;
3746 if (test_bit(FailFast, &rdev->flags))
3747 bio->bi_opf |= MD_FAILFAST;
3748 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3749 bio_set_dev(bio, rdev->bdev);
3750 count++;
3751
3752 rdev = rcu_dereference(conf->mirrors[d].replacement);
3753 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3754 rcu_read_unlock();
3755 continue;
3756 }
3757 atomic_inc(&rdev->nr_pending);
3758
3759 /* Need to set up for writing to the replacement */
3760 bio = r10_bio->devs[i].repl_bio;
3761 bio->bi_status = BLK_STS_IOERR;
3762
3763 sector = r10_bio->devs[i].addr;
3764 bio->bi_next = biolist;
3765 biolist = bio;
3766 bio->bi_end_io = end_sync_write;
3767 bio->bi_opf = REQ_OP_WRITE;
3768 if (test_bit(FailFast, &rdev->flags))
3769 bio->bi_opf |= MD_FAILFAST;
3770 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3771 bio_set_dev(bio, rdev->bdev);
3772 count++;
3773 rcu_read_unlock();
3774 }
3775
3776 if (count < 2) {
3777 for (i=0; i<conf->copies; i++) {
3778 int d = r10_bio->devs[i].devnum;
3779 if (r10_bio->devs[i].bio->bi_end_io)
3780 rdev_dec_pending(conf->mirrors[d].rdev,
3781 mddev);
3782 if (r10_bio->devs[i].repl_bio &&
3783 r10_bio->devs[i].repl_bio->bi_end_io)
3784 rdev_dec_pending(
3785 conf->mirrors[d].replacement,
3786 mddev);
3787 }
3788 put_buf(r10_bio);
3789 biolist = NULL;
3790 goto giveup;
3791 }
3792 }
3793
3794 nr_sectors = 0;
3795 if (sector_nr + max_sync < max_sector)
3796 max_sector = sector_nr + max_sync;
3797 do {
3798 struct page *page;
3799 int len = PAGE_SIZE;
3800 if (sector_nr + (len>>9) > max_sector)
3801 len = (max_sector - sector_nr) << 9;
3802 if (len == 0)
3803 break;
3804 for (bio= biolist ; bio ; bio=bio->bi_next) {
3805 struct resync_pages *rp = get_resync_pages(bio);
3806 page = resync_fetch_page(rp, page_idx);
3807 /*
3808 * won't fail because the vec table is big enough
3809 * to hold all these pages
3810 */
3811 bio_add_page(bio, page, len, 0);
3812 }
3813 nr_sectors += len>>9;
3814 sector_nr += len>>9;
3815 } while (++page_idx < RESYNC_PAGES);
3816 r10_bio->sectors = nr_sectors;
3817
3818 if (mddev_is_clustered(mddev) &&
3819 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3820 /* It is resync not recovery */
3821 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3822 conf->cluster_sync_low = mddev->curr_resync_completed;
3823 raid10_set_cluster_sync_high(conf);
3824 /* Send resync message */
3825 md_cluster_ops->resync_info_update(mddev,
3826 conf->cluster_sync_low,
3827 conf->cluster_sync_high);
3828 }
3829 } else if (mddev_is_clustered(mddev)) {
3830 /* This is recovery not resync */
3831 sector_t sect_va1, sect_va2;
3832 bool broadcast_msg = false;
3833
3834 for (i = 0; i < conf->geo.raid_disks; i++) {
3835 /*
3836 * sector_nr is a device address for recovery, so we
3837 * need translate it to array address before compare
3838 * with cluster_sync_high.
3839 */
3840 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3841
3842 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3843 broadcast_msg = true;
3844 /*
3845 * curr_resync_completed is similar as
3846 * sector_nr, so make the translation too.
3847 */
3848 sect_va2 = raid10_find_virt(conf,
3849 mddev->curr_resync_completed, i);
3850
3851 if (conf->cluster_sync_low == 0 ||
3852 conf->cluster_sync_low > sect_va2)
3853 conf->cluster_sync_low = sect_va2;
3854 }
3855 }
3856 if (broadcast_msg) {
3857 raid10_set_cluster_sync_high(conf);
3858 md_cluster_ops->resync_info_update(mddev,
3859 conf->cluster_sync_low,
3860 conf->cluster_sync_high);
3861 }
3862 }
3863
3864 while (biolist) {
3865 bio = biolist;
3866 biolist = biolist->bi_next;
3867
3868 bio->bi_next = NULL;
3869 r10_bio = get_resync_r10bio(bio);
3870 r10_bio->sectors = nr_sectors;
3871
3872 if (bio->bi_end_io == end_sync_read) {
3873 md_sync_acct_bio(bio, nr_sectors);
3874 bio->bi_status = 0;
3875 submit_bio_noacct(bio);
3876 }
3877 }
3878
3879 if (sectors_skipped)
3880 /* pretend they weren't skipped, it makes
3881 * no important difference in this case
3882 */
3883 md_done_sync(mddev, sectors_skipped, 1);
3884
3885 return sectors_skipped + nr_sectors;
3886 giveup:
3887 /* There is nowhere to write, so all non-sync
3888 * drives must be failed or in resync, all drives
3889 * have a bad block, so try the next chunk...
3890 */
3891 if (sector_nr + max_sync < max_sector)
3892 max_sector = sector_nr + max_sync;
3893
3894 sectors_skipped += (max_sector - sector_nr);
3895 chunks_skipped ++;
3896 sector_nr = max_sector;
3897 goto skipped;
3898}
3899
3900static sector_t
3901raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3902{
3903 sector_t size;
3904 struct r10conf *conf = mddev->private;
3905
3906 if (!raid_disks)
3907 raid_disks = min(conf->geo.raid_disks,
3908 conf->prev.raid_disks);
3909 if (!sectors)
3910 sectors = conf->dev_sectors;
3911
3912 size = sectors >> conf->geo.chunk_shift;
3913 sector_div(size, conf->geo.far_copies);
3914 size = size * raid_disks;
3915 sector_div(size, conf->geo.near_copies);
3916
3917 return size << conf->geo.chunk_shift;
3918}
3919
3920static void calc_sectors(struct r10conf *conf, sector_t size)
3921{
3922 /* Calculate the number of sectors-per-device that will
3923 * actually be used, and set conf->dev_sectors and
3924 * conf->stride
3925 */
3926
3927 size = size >> conf->geo.chunk_shift;
3928 sector_div(size, conf->geo.far_copies);
3929 size = size * conf->geo.raid_disks;
3930 sector_div(size, conf->geo.near_copies);
3931 /* 'size' is now the number of chunks in the array */
3932 /* calculate "used chunks per device" */
3933 size = size * conf->copies;
3934
3935 /* We need to round up when dividing by raid_disks to
3936 * get the stride size.
3937 */
3938 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3939
3940 conf->dev_sectors = size << conf->geo.chunk_shift;
3941
3942 if (conf->geo.far_offset)
3943 conf->geo.stride = 1 << conf->geo.chunk_shift;
3944 else {
3945 sector_div(size, conf->geo.far_copies);
3946 conf->geo.stride = size << conf->geo.chunk_shift;
3947 }
3948}
3949
3950enum geo_type {geo_new, geo_old, geo_start};
3951static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3952{
3953 int nc, fc, fo;
3954 int layout, chunk, disks;
3955 switch (new) {
3956 case geo_old:
3957 layout = mddev->layout;
3958 chunk = mddev->chunk_sectors;
3959 disks = mddev->raid_disks - mddev->delta_disks;
3960 break;
3961 case geo_new:
3962 layout = mddev->new_layout;
3963 chunk = mddev->new_chunk_sectors;
3964 disks = mddev->raid_disks;
3965 break;
3966 default: /* avoid 'may be unused' warnings */
3967 case geo_start: /* new when starting reshape - raid_disks not
3968 * updated yet. */
3969 layout = mddev->new_layout;
3970 chunk = mddev->new_chunk_sectors;
3971 disks = mddev->raid_disks + mddev->delta_disks;
3972 break;
3973 }
3974 if (layout >> 19)
3975 return -1;
3976 if (chunk < (PAGE_SIZE >> 9) ||
3977 !is_power_of_2(chunk))
3978 return -2;
3979 nc = layout & 255;
3980 fc = (layout >> 8) & 255;
3981 fo = layout & (1<<16);
3982 geo->raid_disks = disks;
3983 geo->near_copies = nc;
3984 geo->far_copies = fc;
3985 geo->far_offset = fo;
3986 switch (layout >> 17) {
3987 case 0: /* original layout. simple but not always optimal */
3988 geo->far_set_size = disks;
3989 break;
3990 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3991 * actually using this, but leave code here just in case.*/
3992 geo->far_set_size = disks/fc;
3993 WARN(geo->far_set_size < fc,
3994 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3995 break;
3996 case 2: /* "improved" layout fixed to match documentation */
3997 geo->far_set_size = fc * nc;
3998 break;
3999 default: /* Not a valid layout */
4000 return -1;
4001 }
4002 geo->chunk_mask = chunk - 1;
4003 geo->chunk_shift = ffz(~chunk);
4004 return nc*fc;
4005}
4006
4007static struct r10conf *setup_conf(struct mddev *mddev)
4008{
4009 struct r10conf *conf = NULL;
4010 int err = -EINVAL;
4011 struct geom geo;
4012 int copies;
4013
4014 copies = setup_geo(&geo, mddev, geo_new);
4015
4016 if (copies == -2) {
4017 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
4018 mdname(mddev), PAGE_SIZE);
4019 goto out;
4020 }
4021
4022 if (copies < 2 || copies > mddev->raid_disks) {
4023 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
4024 mdname(mddev), mddev->new_layout);
4025 goto out;
4026 }
4027
4028 err = -ENOMEM;
4029 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
4030 if (!conf)
4031 goto out;
4032
4033 /* FIXME calc properly */
4034 conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
4035 sizeof(struct raid10_info),
4036 GFP_KERNEL);
4037 if (!conf->mirrors)
4038 goto out;
4039
4040 conf->tmppage = alloc_page(GFP_KERNEL);
4041 if (!conf->tmppage)
4042 goto out;
4043
4044 conf->geo = geo;
4045 conf->copies = copies;
4046 err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
4047 rbio_pool_free, conf);
4048 if (err)
4049 goto out;
4050
4051 err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
4052 if (err)
4053 goto out;
4054
4055 calc_sectors(conf, mddev->dev_sectors);
4056 if (mddev->reshape_position == MaxSector) {
4057 conf->prev = conf->geo;
4058 conf->reshape_progress = MaxSector;
4059 } else {
4060 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
4061 err = -EINVAL;
4062 goto out;
4063 }
4064 conf->reshape_progress = mddev->reshape_position;
4065 if (conf->prev.far_offset)
4066 conf->prev.stride = 1 << conf->prev.chunk_shift;
4067 else
4068 /* far_copies must be 1 */
4069 conf->prev.stride = conf->dev_sectors;
4070 }
4071 conf->reshape_safe = conf->reshape_progress;
4072 spin_lock_init(&conf->device_lock);
4073 INIT_LIST_HEAD(&conf->retry_list);
4074 INIT_LIST_HEAD(&conf->bio_end_io_list);
4075
4076 seqlock_init(&conf->resync_lock);
4077 init_waitqueue_head(&conf->wait_barrier);
4078 atomic_set(&conf->nr_pending, 0);
4079
4080 err = -ENOMEM;
4081 conf->thread = md_register_thread(raid10d, mddev, "raid10");
4082 if (!conf->thread)
4083 goto out;
4084
4085 conf->mddev = mddev;
4086 return conf;
4087
4088 out:
4089 if (conf) {
4090 mempool_exit(&conf->r10bio_pool);
4091 kfree(conf->mirrors);
4092 safe_put_page(conf->tmppage);
4093 bioset_exit(&conf->bio_split);
4094 kfree(conf);
4095 }
4096 return ERR_PTR(err);
4097}
4098
4099static void raid10_set_io_opt(struct r10conf *conf)
4100{
4101 int raid_disks = conf->geo.raid_disks;
4102
4103 if (!(conf->geo.raid_disks % conf->geo.near_copies))
4104 raid_disks /= conf->geo.near_copies;
4105 blk_queue_io_opt(conf->mddev->queue, (conf->mddev->chunk_sectors << 9) *
4106 raid_disks);
4107}
4108
4109static int raid10_run(struct mddev *mddev)
4110{
4111 struct r10conf *conf;
4112 int i, disk_idx;
4113 struct raid10_info *disk;
4114 struct md_rdev *rdev;
4115 sector_t size;
4116 sector_t min_offset_diff = 0;
4117 int first = 1;
4118
4119 if (mddev_init_writes_pending(mddev) < 0)
4120 return -ENOMEM;
4121
4122 if (mddev->private == NULL) {
4123 conf = setup_conf(mddev);
4124 if (IS_ERR(conf))
4125 return PTR_ERR(conf);
4126 mddev->private = conf;
4127 }
4128 conf = mddev->private;
4129 if (!conf)
4130 goto out;
4131
4132 if (mddev_is_clustered(conf->mddev)) {
4133 int fc, fo;
4134
4135 fc = (mddev->layout >> 8) & 255;
4136 fo = mddev->layout & (1<<16);
4137 if (fc > 1 || fo > 0) {
4138 pr_err("only near layout is supported by clustered"
4139 " raid10\n");
4140 goto out_free_conf;
4141 }
4142 }
4143
4144 mddev->thread = conf->thread;
4145 conf->thread = NULL;
4146
4147 if (mddev->queue) {
4148 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
4149 blk_queue_io_min(mddev->queue, mddev->chunk_sectors << 9);
4150 raid10_set_io_opt(conf);
4151 }
4152
4153 rdev_for_each(rdev, mddev) {
4154 long long diff;
4155
4156 disk_idx = rdev->raid_disk;
4157 if (disk_idx < 0)
4158 continue;
4159 if (disk_idx >= conf->geo.raid_disks &&
4160 disk_idx >= conf->prev.raid_disks)
4161 continue;
4162 disk = conf->mirrors + disk_idx;
4163
4164 if (test_bit(Replacement, &rdev->flags)) {
4165 if (disk->replacement)
4166 goto out_free_conf;
4167 disk->replacement = rdev;
4168 } else {
4169 if (disk->rdev)
4170 goto out_free_conf;
4171 disk->rdev = rdev;
4172 }
4173 diff = (rdev->new_data_offset - rdev->data_offset);
4174 if (!mddev->reshape_backwards)
4175 diff = -diff;
4176 if (diff < 0)
4177 diff = 0;
4178 if (first || diff < min_offset_diff)
4179 min_offset_diff = diff;
4180
4181 if (mddev->gendisk)
4182 disk_stack_limits(mddev->gendisk, rdev->bdev,
4183 rdev->data_offset << 9);
4184
4185 disk->head_position = 0;
4186 first = 0;
4187 }
4188
4189 /* need to check that every block has at least one working mirror */
4190 if (!enough(conf, -1)) {
4191 pr_err("md/raid10:%s: not enough operational mirrors.\n",
4192 mdname(mddev));
4193 goto out_free_conf;
4194 }
4195
4196 if (conf->reshape_progress != MaxSector) {
4197 /* must ensure that shape change is supported */
4198 if (conf->geo.far_copies != 1 &&
4199 conf->geo.far_offset == 0)
4200 goto out_free_conf;
4201 if (conf->prev.far_copies != 1 &&
4202 conf->prev.far_offset == 0)
4203 goto out_free_conf;
4204 }
4205
4206 mddev->degraded = 0;
4207 for (i = 0;
4208 i < conf->geo.raid_disks
4209 || i < conf->prev.raid_disks;
4210 i++) {
4211
4212 disk = conf->mirrors + i;
4213
4214 if (!disk->rdev && disk->replacement) {
4215 /* The replacement is all we have - use it */
4216 disk->rdev = disk->replacement;
4217 disk->replacement = NULL;
4218 clear_bit(Replacement, &disk->rdev->flags);
4219 }
4220
4221 if (!disk->rdev ||
4222 !test_bit(In_sync, &disk->rdev->flags)) {
4223 disk->head_position = 0;
4224 mddev->degraded++;
4225 if (disk->rdev &&
4226 disk->rdev->saved_raid_disk < 0)
4227 conf->fullsync = 1;
4228 }
4229
4230 if (disk->replacement &&
4231 !test_bit(In_sync, &disk->replacement->flags) &&
4232 disk->replacement->saved_raid_disk < 0) {
4233 conf->fullsync = 1;
4234 }
4235
4236 disk->recovery_disabled = mddev->recovery_disabled - 1;
4237 }
4238
4239 if (mddev->recovery_cp != MaxSector)
4240 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
4241 mdname(mddev));
4242 pr_info("md/raid10:%s: active with %d out of %d devices\n",
4243 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
4244 conf->geo.raid_disks);
4245 /*
4246 * Ok, everything is just fine now
4247 */
4248 mddev->dev_sectors = conf->dev_sectors;
4249 size = raid10_size(mddev, 0, 0);
4250 md_set_array_sectors(mddev, size);
4251 mddev->resync_max_sectors = size;
4252 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
4253
4254 if (md_integrity_register(mddev))
4255 goto out_free_conf;
4256
4257 if (conf->reshape_progress != MaxSector) {
4258 unsigned long before_length, after_length;
4259
4260 before_length = ((1 << conf->prev.chunk_shift) *
4261 conf->prev.far_copies);
4262 after_length = ((1 << conf->geo.chunk_shift) *
4263 conf->geo.far_copies);
4264
4265 if (max(before_length, after_length) > min_offset_diff) {
4266 /* This cannot work */
4267 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
4268 goto out_free_conf;
4269 }
4270 conf->offset_diff = min_offset_diff;
4271
4272 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4273 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4274 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4275 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4276 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4277 "reshape");
4278 if (!mddev->sync_thread)
4279 goto out_free_conf;
4280 }
4281
4282 return 0;
4283
4284out_free_conf:
4285 md_unregister_thread(&mddev->thread);
4286 mempool_exit(&conf->r10bio_pool);
4287 safe_put_page(conf->tmppage);
4288 kfree(conf->mirrors);
4289 kfree(conf);
4290 mddev->private = NULL;
4291out:
4292 return -EIO;
4293}
4294
4295static void raid10_free(struct mddev *mddev, void *priv)
4296{
4297 struct r10conf *conf = priv;
4298
4299 mempool_exit(&conf->r10bio_pool);
4300 safe_put_page(conf->tmppage);
4301 kfree(conf->mirrors);
4302 kfree(conf->mirrors_old);
4303 kfree(conf->mirrors_new);
4304 bioset_exit(&conf->bio_split);
4305 kfree(conf);
4306}
4307
4308static void raid10_quiesce(struct mddev *mddev, int quiesce)
4309{
4310 struct r10conf *conf = mddev->private;
4311
4312 if (quiesce)
4313 raise_barrier(conf, 0);
4314 else
4315 lower_barrier(conf);
4316}
4317
4318static int raid10_resize(struct mddev *mddev, sector_t sectors)
4319{
4320 /* Resize of 'far' arrays is not supported.
4321 * For 'near' and 'offset' arrays we can set the
4322 * number of sectors used to be an appropriate multiple
4323 * of the chunk size.
4324 * For 'offset', this is far_copies*chunksize.
4325 * For 'near' the multiplier is the LCM of
4326 * near_copies and raid_disks.
4327 * So if far_copies > 1 && !far_offset, fail.
4328 * Else find LCM(raid_disks, near_copy)*far_copies and
4329 * multiply by chunk_size. Then round to this number.
4330 * This is mostly done by raid10_size()
4331 */
4332 struct r10conf *conf = mddev->private;
4333 sector_t oldsize, size;
4334
4335 if (mddev->reshape_position != MaxSector)
4336 return -EBUSY;
4337
4338 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4339 return -EINVAL;
4340
4341 oldsize = raid10_size(mddev, 0, 0);
4342 size = raid10_size(mddev, sectors, 0);
4343 if (mddev->external_size &&
4344 mddev->array_sectors > size)
4345 return -EINVAL;
4346 if (mddev->bitmap) {
4347 int ret = md_bitmap_resize(mddev->bitmap, size, 0, 0);
4348 if (ret)
4349 return ret;
4350 }
4351 md_set_array_sectors(mddev, size);
4352 if (sectors > mddev->dev_sectors &&
4353 mddev->recovery_cp > oldsize) {
4354 mddev->recovery_cp = oldsize;
4355 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4356 }
4357 calc_sectors(conf, sectors);
4358 mddev->dev_sectors = conf->dev_sectors;
4359 mddev->resync_max_sectors = size;
4360 return 0;
4361}
4362
4363static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4364{
4365 struct md_rdev *rdev;
4366 struct r10conf *conf;
4367
4368 if (mddev->degraded > 0) {
4369 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4370 mdname(mddev));
4371 return ERR_PTR(-EINVAL);
4372 }
4373 sector_div(size, devs);
4374
4375 /* Set new parameters */
4376 mddev->new_level = 10;
4377 /* new layout: far_copies = 1, near_copies = 2 */
4378 mddev->new_layout = (1<<8) + 2;
4379 mddev->new_chunk_sectors = mddev->chunk_sectors;
4380 mddev->delta_disks = mddev->raid_disks;
4381 mddev->raid_disks *= 2;
4382 /* make sure it will be not marked as dirty */
4383 mddev->recovery_cp = MaxSector;
4384 mddev->dev_sectors = size;
4385
4386 conf = setup_conf(mddev);
4387 if (!IS_ERR(conf)) {
4388 rdev_for_each(rdev, mddev)
4389 if (rdev->raid_disk >= 0) {
4390 rdev->new_raid_disk = rdev->raid_disk * 2;
4391 rdev->sectors = size;
4392 }
4393 WRITE_ONCE(conf->barrier, 1);
4394 }
4395
4396 return conf;
4397}
4398
4399static void *raid10_takeover(struct mddev *mddev)
4400{
4401 struct r0conf *raid0_conf;
4402
4403 /* raid10 can take over:
4404 * raid0 - providing it has only two drives
4405 */
4406 if (mddev->level == 0) {
4407 /* for raid0 takeover only one zone is supported */
4408 raid0_conf = mddev->private;
4409 if (raid0_conf->nr_strip_zones > 1) {
4410 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4411 mdname(mddev));
4412 return ERR_PTR(-EINVAL);
4413 }
4414 return raid10_takeover_raid0(mddev,
4415 raid0_conf->strip_zone->zone_end,
4416 raid0_conf->strip_zone->nb_dev);
4417 }
4418 return ERR_PTR(-EINVAL);
4419}
4420
4421static int raid10_check_reshape(struct mddev *mddev)
4422{
4423 /* Called when there is a request to change
4424 * - layout (to ->new_layout)
4425 * - chunk size (to ->new_chunk_sectors)
4426 * - raid_disks (by delta_disks)
4427 * or when trying to restart a reshape that was ongoing.
4428 *
4429 * We need to validate the request and possibly allocate
4430 * space if that might be an issue later.
4431 *
4432 * Currently we reject any reshape of a 'far' mode array,
4433 * allow chunk size to change if new is generally acceptable,
4434 * allow raid_disks to increase, and allow
4435 * a switch between 'near' mode and 'offset' mode.
4436 */
4437 struct r10conf *conf = mddev->private;
4438 struct geom geo;
4439
4440 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4441 return -EINVAL;
4442
4443 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4444 /* mustn't change number of copies */
4445 return -EINVAL;
4446 if (geo.far_copies > 1 && !geo.far_offset)
4447 /* Cannot switch to 'far' mode */
4448 return -EINVAL;
4449
4450 if (mddev->array_sectors & geo.chunk_mask)
4451 /* not factor of array size */
4452 return -EINVAL;
4453
4454 if (!enough(conf, -1))
4455 return -EINVAL;
4456
4457 kfree(conf->mirrors_new);
4458 conf->mirrors_new = NULL;
4459 if (mddev->delta_disks > 0) {
4460 /* allocate new 'mirrors' list */
4461 conf->mirrors_new =
4462 kcalloc(mddev->raid_disks + mddev->delta_disks,
4463 sizeof(struct raid10_info),
4464 GFP_KERNEL);
4465 if (!conf->mirrors_new)
4466 return -ENOMEM;
4467 }
4468 return 0;
4469}
4470
4471/*
4472 * Need to check if array has failed when deciding whether to:
4473 * - start an array
4474 * - remove non-faulty devices
4475 * - add a spare
4476 * - allow a reshape
4477 * This determination is simple when no reshape is happening.
4478 * However if there is a reshape, we need to carefully check
4479 * both the before and after sections.
4480 * This is because some failed devices may only affect one
4481 * of the two sections, and some non-in_sync devices may
4482 * be insync in the section most affected by failed devices.
4483 */
4484static int calc_degraded(struct r10conf *conf)
4485{
4486 int degraded, degraded2;
4487 int i;
4488
4489 rcu_read_lock();
4490 degraded = 0;
4491 /* 'prev' section first */
4492 for (i = 0; i < conf->prev.raid_disks; i++) {
4493 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4494 if (!rdev || test_bit(Faulty, &rdev->flags))
4495 degraded++;
4496 else if (!test_bit(In_sync, &rdev->flags))
4497 /* When we can reduce the number of devices in
4498 * an array, this might not contribute to
4499 * 'degraded'. It does now.
4500 */
4501 degraded++;
4502 }
4503 rcu_read_unlock();
4504 if (conf->geo.raid_disks == conf->prev.raid_disks)
4505 return degraded;
4506 rcu_read_lock();
4507 degraded2 = 0;
4508 for (i = 0; i < conf->geo.raid_disks; i++) {
4509 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4510 if (!rdev || test_bit(Faulty, &rdev->flags))
4511 degraded2++;
4512 else if (!test_bit(In_sync, &rdev->flags)) {
4513 /* If reshape is increasing the number of devices,
4514 * this section has already been recovered, so
4515 * it doesn't contribute to degraded.
4516 * else it does.
4517 */
4518 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4519 degraded2++;
4520 }
4521 }
4522 rcu_read_unlock();
4523 if (degraded2 > degraded)
4524 return degraded2;
4525 return degraded;
4526}
4527
4528static int raid10_start_reshape(struct mddev *mddev)
4529{
4530 /* A 'reshape' has been requested. This commits
4531 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4532 * This also checks if there are enough spares and adds them
4533 * to the array.
4534 * We currently require enough spares to make the final
4535 * array non-degraded. We also require that the difference
4536 * between old and new data_offset - on each device - is
4537 * enough that we never risk over-writing.
4538 */
4539
4540 unsigned long before_length, after_length;
4541 sector_t min_offset_diff = 0;
4542 int first = 1;
4543 struct geom new;
4544 struct r10conf *conf = mddev->private;
4545 struct md_rdev *rdev;
4546 int spares = 0;
4547 int ret;
4548
4549 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4550 return -EBUSY;
4551
4552 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4553 return -EINVAL;
4554
4555 before_length = ((1 << conf->prev.chunk_shift) *
4556 conf->prev.far_copies);
4557 after_length = ((1 << conf->geo.chunk_shift) *
4558 conf->geo.far_copies);
4559
4560 rdev_for_each(rdev, mddev) {
4561 if (!test_bit(In_sync, &rdev->flags)
4562 && !test_bit(Faulty, &rdev->flags))
4563 spares++;
4564 if (rdev->raid_disk >= 0) {
4565 long long diff = (rdev->new_data_offset
4566 - rdev->data_offset);
4567 if (!mddev->reshape_backwards)
4568 diff = -diff;
4569 if (diff < 0)
4570 diff = 0;
4571 if (first || diff < min_offset_diff)
4572 min_offset_diff = diff;
4573 first = 0;
4574 }
4575 }
4576
4577 if (max(before_length, after_length) > min_offset_diff)
4578 return -EINVAL;
4579
4580 if (spares < mddev->delta_disks)
4581 return -EINVAL;
4582
4583 conf->offset_diff = min_offset_diff;
4584 spin_lock_irq(&conf->device_lock);
4585 if (conf->mirrors_new) {
4586 memcpy(conf->mirrors_new, conf->mirrors,
4587 sizeof(struct raid10_info)*conf->prev.raid_disks);
4588 smp_mb();
4589 kfree(conf->mirrors_old);
4590 conf->mirrors_old = conf->mirrors;
4591 conf->mirrors = conf->mirrors_new;
4592 conf->mirrors_new = NULL;
4593 }
4594 setup_geo(&conf->geo, mddev, geo_start);
4595 smp_mb();
4596 if (mddev->reshape_backwards) {
4597 sector_t size = raid10_size(mddev, 0, 0);
4598 if (size < mddev->array_sectors) {
4599 spin_unlock_irq(&conf->device_lock);
4600 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4601 mdname(mddev));
4602 return -EINVAL;
4603 }
4604 mddev->resync_max_sectors = size;
4605 conf->reshape_progress = size;
4606 } else
4607 conf->reshape_progress = 0;
4608 conf->reshape_safe = conf->reshape_progress;
4609 spin_unlock_irq(&conf->device_lock);
4610
4611 if (mddev->delta_disks && mddev->bitmap) {
4612 struct mdp_superblock_1 *sb = NULL;
4613 sector_t oldsize, newsize;
4614
4615 oldsize = raid10_size(mddev, 0, 0);
4616 newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4617
4618 if (!mddev_is_clustered(mddev)) {
4619 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4620 if (ret)
4621 goto abort;
4622 else
4623 goto out;
4624 }
4625
4626 rdev_for_each(rdev, mddev) {
4627 if (rdev->raid_disk > -1 &&
4628 !test_bit(Faulty, &rdev->flags))
4629 sb = page_address(rdev->sb_page);
4630 }
4631
4632 /*
4633 * some node is already performing reshape, and no need to
4634 * call md_bitmap_resize again since it should be called when
4635 * receiving BITMAP_RESIZE msg
4636 */
4637 if ((sb && (le32_to_cpu(sb->feature_map) &
4638 MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4639 goto out;
4640
4641 ret = md_bitmap_resize(mddev->bitmap, newsize, 0, 0);
4642 if (ret)
4643 goto abort;
4644
4645 ret = md_cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4646 if (ret) {
4647 md_bitmap_resize(mddev->bitmap, oldsize, 0, 0);
4648 goto abort;
4649 }
4650 }
4651out:
4652 if (mddev->delta_disks > 0) {
4653 rdev_for_each(rdev, mddev)
4654 if (rdev->raid_disk < 0 &&
4655 !test_bit(Faulty, &rdev->flags)) {
4656 if (raid10_add_disk(mddev, rdev) == 0) {
4657 if (rdev->raid_disk >=
4658 conf->prev.raid_disks)
4659 set_bit(In_sync, &rdev->flags);
4660 else
4661 rdev->recovery_offset = 0;
4662
4663 /* Failure here is OK */
4664 sysfs_link_rdev(mddev, rdev);
4665 }
4666 } else if (rdev->raid_disk >= conf->prev.raid_disks
4667 && !test_bit(Faulty, &rdev->flags)) {
4668 /* This is a spare that was manually added */
4669 set_bit(In_sync, &rdev->flags);
4670 }
4671 }
4672 /* When a reshape changes the number of devices,
4673 * ->degraded is measured against the larger of the
4674 * pre and post numbers.
4675 */
4676 spin_lock_irq(&conf->device_lock);
4677 mddev->degraded = calc_degraded(conf);
4678 spin_unlock_irq(&conf->device_lock);
4679 mddev->raid_disks = conf->geo.raid_disks;
4680 mddev->reshape_position = conf->reshape_progress;
4681 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4682
4683 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4684 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4685 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4686 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4687 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4688
4689 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4690 "reshape");
4691 if (!mddev->sync_thread) {
4692 ret = -EAGAIN;
4693 goto abort;
4694 }
4695 conf->reshape_checkpoint = jiffies;
4696 md_wakeup_thread(mddev->sync_thread);
4697 md_new_event();
4698 return 0;
4699
4700abort:
4701 mddev->recovery = 0;
4702 spin_lock_irq(&conf->device_lock);
4703 conf->geo = conf->prev;
4704 mddev->raid_disks = conf->geo.raid_disks;
4705 rdev_for_each(rdev, mddev)
4706 rdev->new_data_offset = rdev->data_offset;
4707 smp_wmb();
4708 conf->reshape_progress = MaxSector;
4709 conf->reshape_safe = MaxSector;
4710 mddev->reshape_position = MaxSector;
4711 spin_unlock_irq(&conf->device_lock);
4712 return ret;
4713}
4714
4715/* Calculate the last device-address that could contain
4716 * any block from the chunk that includes the array-address 's'
4717 * and report the next address.
4718 * i.e. the address returned will be chunk-aligned and after
4719 * any data that is in the chunk containing 's'.
4720 */
4721static sector_t last_dev_address(sector_t s, struct geom *geo)
4722{
4723 s = (s | geo->chunk_mask) + 1;
4724 s >>= geo->chunk_shift;
4725 s *= geo->near_copies;
4726 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4727 s *= geo->far_copies;
4728 s <<= geo->chunk_shift;
4729 return s;
4730}
4731
4732/* Calculate the first device-address that could contain
4733 * any block from the chunk that includes the array-address 's'.
4734 * This too will be the start of a chunk
4735 */
4736static sector_t first_dev_address(sector_t s, struct geom *geo)
4737{
4738 s >>= geo->chunk_shift;
4739 s *= geo->near_copies;
4740 sector_div(s, geo->raid_disks);
4741 s *= geo->far_copies;
4742 s <<= geo->chunk_shift;
4743 return s;
4744}
4745
4746static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4747 int *skipped)
4748{
4749 /* We simply copy at most one chunk (smallest of old and new)
4750 * at a time, possibly less if that exceeds RESYNC_PAGES,
4751 * or we hit a bad block or something.
4752 * This might mean we pause for normal IO in the middle of
4753 * a chunk, but that is not a problem as mddev->reshape_position
4754 * can record any location.
4755 *
4756 * If we will want to write to a location that isn't
4757 * yet recorded as 'safe' (i.e. in metadata on disk) then
4758 * we need to flush all reshape requests and update the metadata.
4759 *
4760 * When reshaping forwards (e.g. to more devices), we interpret
4761 * 'safe' as the earliest block which might not have been copied
4762 * down yet. We divide this by previous stripe size and multiply
4763 * by previous stripe length to get lowest device offset that we
4764 * cannot write to yet.
4765 * We interpret 'sector_nr' as an address that we want to write to.
4766 * From this we use last_device_address() to find where we might
4767 * write to, and first_device_address on the 'safe' position.
4768 * If this 'next' write position is after the 'safe' position,
4769 * we must update the metadata to increase the 'safe' position.
4770 *
4771 * When reshaping backwards, we round in the opposite direction
4772 * and perform the reverse test: next write position must not be
4773 * less than current safe position.
4774 *
4775 * In all this the minimum difference in data offsets
4776 * (conf->offset_diff - always positive) allows a bit of slack,
4777 * so next can be after 'safe', but not by more than offset_diff
4778 *
4779 * We need to prepare all the bios here before we start any IO
4780 * to ensure the size we choose is acceptable to all devices.
4781 * The means one for each copy for write-out and an extra one for
4782 * read-in.
4783 * We store the read-in bio in ->master_bio and the others in
4784 * ->devs[x].bio and ->devs[x].repl_bio.
4785 */
4786 struct r10conf *conf = mddev->private;
4787 struct r10bio *r10_bio;
4788 sector_t next, safe, last;
4789 int max_sectors;
4790 int nr_sectors;
4791 int s;
4792 struct md_rdev *rdev;
4793 int need_flush = 0;
4794 struct bio *blist;
4795 struct bio *bio, *read_bio;
4796 int sectors_done = 0;
4797 struct page **pages;
4798
4799 if (sector_nr == 0) {
4800 /* If restarting in the middle, skip the initial sectors */
4801 if (mddev->reshape_backwards &&
4802 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4803 sector_nr = (raid10_size(mddev, 0, 0)
4804 - conf->reshape_progress);
4805 } else if (!mddev->reshape_backwards &&
4806 conf->reshape_progress > 0)
4807 sector_nr = conf->reshape_progress;
4808 if (sector_nr) {
4809 mddev->curr_resync_completed = sector_nr;
4810 sysfs_notify_dirent_safe(mddev->sysfs_completed);
4811 *skipped = 1;
4812 return sector_nr;
4813 }
4814 }
4815
4816 /* We don't use sector_nr to track where we are up to
4817 * as that doesn't work well for ->reshape_backwards.
4818 * So just use ->reshape_progress.
4819 */
4820 if (mddev->reshape_backwards) {
4821 /* 'next' is the earliest device address that we might
4822 * write to for this chunk in the new layout
4823 */
4824 next = first_dev_address(conf->reshape_progress - 1,
4825 &conf->geo);
4826
4827 /* 'safe' is the last device address that we might read from
4828 * in the old layout after a restart
4829 */
4830 safe = last_dev_address(conf->reshape_safe - 1,
4831 &conf->prev);
4832
4833 if (next + conf->offset_diff < safe)
4834 need_flush = 1;
4835
4836 last = conf->reshape_progress - 1;
4837 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4838 & conf->prev.chunk_mask);
4839 if (sector_nr + RESYNC_SECTORS < last)
4840 sector_nr = last + 1 - RESYNC_SECTORS;
4841 } else {
4842 /* 'next' is after the last device address that we
4843 * might write to for this chunk in the new layout
4844 */
4845 next = last_dev_address(conf->reshape_progress, &conf->geo);
4846
4847 /* 'safe' is the earliest device address that we might
4848 * read from in the old layout after a restart
4849 */
4850 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4851
4852 /* Need to update metadata if 'next' might be beyond 'safe'
4853 * as that would possibly corrupt data
4854 */
4855 if (next > safe + conf->offset_diff)
4856 need_flush = 1;
4857
4858 sector_nr = conf->reshape_progress;
4859 last = sector_nr | (conf->geo.chunk_mask
4860 & conf->prev.chunk_mask);
4861
4862 if (sector_nr + RESYNC_SECTORS <= last)
4863 last = sector_nr + RESYNC_SECTORS - 1;
4864 }
4865
4866 if (need_flush ||
4867 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4868 /* Need to update reshape_position in metadata */
4869 wait_barrier(conf, false);
4870 mddev->reshape_position = conf->reshape_progress;
4871 if (mddev->reshape_backwards)
4872 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4873 - conf->reshape_progress;
4874 else
4875 mddev->curr_resync_completed = conf->reshape_progress;
4876 conf->reshape_checkpoint = jiffies;
4877 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4878 md_wakeup_thread(mddev->thread);
4879 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4880 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4881 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4882 allow_barrier(conf);
4883 return sectors_done;
4884 }
4885 conf->reshape_safe = mddev->reshape_position;
4886 allow_barrier(conf);
4887 }
4888
4889 raise_barrier(conf, 0);
4890read_more:
4891 /* Now schedule reads for blocks from sector_nr to last */
4892 r10_bio = raid10_alloc_init_r10buf(conf);
4893 r10_bio->state = 0;
4894 raise_barrier(conf, 1);
4895 atomic_set(&r10_bio->remaining, 0);
4896 r10_bio->mddev = mddev;
4897 r10_bio->sector = sector_nr;
4898 set_bit(R10BIO_IsReshape, &r10_bio->state);
4899 r10_bio->sectors = last - sector_nr + 1;
4900 rdev = read_balance(conf, r10_bio, &max_sectors);
4901 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4902
4903 if (!rdev) {
4904 /* Cannot read from here, so need to record bad blocks
4905 * on all the target devices.
4906 */
4907 // FIXME
4908 mempool_free(r10_bio, &conf->r10buf_pool);
4909 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4910 return sectors_done;
4911 }
4912
4913 read_bio = bio_alloc_bioset(rdev->bdev, RESYNC_PAGES, REQ_OP_READ,
4914 GFP_KERNEL, &mddev->bio_set);
4915 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4916 + rdev->data_offset);
4917 read_bio->bi_private = r10_bio;
4918 read_bio->bi_end_io = end_reshape_read;
4919 r10_bio->master_bio = read_bio;
4920 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4921
4922 /*
4923 * Broadcast RESYNC message to other nodes, so all nodes would not
4924 * write to the region to avoid conflict.
4925 */
4926 if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4927 struct mdp_superblock_1 *sb = NULL;
4928 int sb_reshape_pos = 0;
4929
4930 conf->cluster_sync_low = sector_nr;
4931 conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4932 sb = page_address(rdev->sb_page);
4933 if (sb) {
4934 sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4935 /*
4936 * Set cluster_sync_low again if next address for array
4937 * reshape is less than cluster_sync_low. Since we can't
4938 * update cluster_sync_low until it has finished reshape.
4939 */
4940 if (sb_reshape_pos < conf->cluster_sync_low)
4941 conf->cluster_sync_low = sb_reshape_pos;
4942 }
4943
4944 md_cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4945 conf->cluster_sync_high);
4946 }
4947
4948 /* Now find the locations in the new layout */
4949 __raid10_find_phys(&conf->geo, r10_bio);
4950
4951 blist = read_bio;
4952 read_bio->bi_next = NULL;
4953
4954 rcu_read_lock();
4955 for (s = 0; s < conf->copies*2; s++) {
4956 struct bio *b;
4957 int d = r10_bio->devs[s/2].devnum;
4958 struct md_rdev *rdev2;
4959 if (s&1) {
4960 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4961 b = r10_bio->devs[s/2].repl_bio;
4962 } else {
4963 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4964 b = r10_bio->devs[s/2].bio;
4965 }
4966 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4967 continue;
4968
4969 bio_set_dev(b, rdev2->bdev);
4970 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4971 rdev2->new_data_offset;
4972 b->bi_end_io = end_reshape_write;
4973 b->bi_opf = REQ_OP_WRITE;
4974 b->bi_next = blist;
4975 blist = b;
4976 }
4977
4978 /* Now add as many pages as possible to all of these bios. */
4979
4980 nr_sectors = 0;
4981 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4982 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4983 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4984 int len = (max_sectors - s) << 9;
4985 if (len > PAGE_SIZE)
4986 len = PAGE_SIZE;
4987 for (bio = blist; bio ; bio = bio->bi_next) {
4988 /*
4989 * won't fail because the vec table is big enough
4990 * to hold all these pages
4991 */
4992 bio_add_page(bio, page, len, 0);
4993 }
4994 sector_nr += len >> 9;
4995 nr_sectors += len >> 9;
4996 }
4997 rcu_read_unlock();
4998 r10_bio->sectors = nr_sectors;
4999
5000 /* Now submit the read */
5001 md_sync_acct_bio(read_bio, r10_bio->sectors);
5002 atomic_inc(&r10_bio->remaining);
5003 read_bio->bi_next = NULL;
5004 submit_bio_noacct(read_bio);
5005 sectors_done += nr_sectors;
5006 if (sector_nr <= last)
5007 goto read_more;
5008
5009 lower_barrier(conf);
5010
5011 /* Now that we have done the whole section we can
5012 * update reshape_progress
5013 */
5014 if (mddev->reshape_backwards)
5015 conf->reshape_progress -= sectors_done;
5016 else
5017 conf->reshape_progress += sectors_done;
5018
5019 return sectors_done;
5020}
5021
5022static void end_reshape_request(struct r10bio *r10_bio);
5023static int handle_reshape_read_error(struct mddev *mddev,
5024 struct r10bio *r10_bio);
5025static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
5026{
5027 /* Reshape read completed. Hopefully we have a block
5028 * to write out.
5029 * If we got a read error then we do sync 1-page reads from
5030 * elsewhere until we find the data - or give up.
5031 */
5032 struct r10conf *conf = mddev->private;
5033 int s;
5034
5035 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
5036 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
5037 /* Reshape has been aborted */
5038 md_done_sync(mddev, r10_bio->sectors, 0);
5039 return;
5040 }
5041
5042 /* We definitely have the data in the pages, schedule the
5043 * writes.
5044 */
5045 atomic_set(&r10_bio->remaining, 1);
5046 for (s = 0; s < conf->copies*2; s++) {
5047 struct bio *b;
5048 int d = r10_bio->devs[s/2].devnum;
5049 struct md_rdev *rdev;
5050 rcu_read_lock();
5051 if (s&1) {
5052 rdev = rcu_dereference(conf->mirrors[d].replacement);
5053 b = r10_bio->devs[s/2].repl_bio;
5054 } else {
5055 rdev = rcu_dereference(conf->mirrors[d].rdev);
5056 b = r10_bio->devs[s/2].bio;
5057 }
5058 if (!rdev || test_bit(Faulty, &rdev->flags)) {
5059 rcu_read_unlock();
5060 continue;
5061 }
5062 atomic_inc(&rdev->nr_pending);
5063 rcu_read_unlock();
5064 md_sync_acct_bio(b, r10_bio->sectors);
5065 atomic_inc(&r10_bio->remaining);
5066 b->bi_next = NULL;
5067 submit_bio_noacct(b);
5068 }
5069 end_reshape_request(r10_bio);
5070}
5071
5072static void end_reshape(struct r10conf *conf)
5073{
5074 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
5075 return;
5076
5077 spin_lock_irq(&conf->device_lock);
5078 conf->prev = conf->geo;
5079 md_finish_reshape(conf->mddev);
5080 smp_wmb();
5081 conf->reshape_progress = MaxSector;
5082 conf->reshape_safe = MaxSector;
5083 spin_unlock_irq(&conf->device_lock);
5084
5085 if (conf->mddev->queue)
5086 raid10_set_io_opt(conf);
5087 conf->fullsync = 0;
5088}
5089
5090static void raid10_update_reshape_pos(struct mddev *mddev)
5091{
5092 struct r10conf *conf = mddev->private;
5093 sector_t lo, hi;
5094
5095 md_cluster_ops->resync_info_get(mddev, &lo, &hi);
5096 if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
5097 || mddev->reshape_position == MaxSector)
5098 conf->reshape_progress = mddev->reshape_position;
5099 else
5100 WARN_ON_ONCE(1);
5101}
5102
5103static int handle_reshape_read_error(struct mddev *mddev,
5104 struct r10bio *r10_bio)
5105{
5106 /* Use sync reads to get the blocks from somewhere else */
5107 int sectors = r10_bio->sectors;
5108 struct r10conf *conf = mddev->private;
5109 struct r10bio *r10b;
5110 int slot = 0;
5111 int idx = 0;
5112 struct page **pages;
5113
5114 r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
5115 if (!r10b) {
5116 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
5117 return -ENOMEM;
5118 }
5119
5120 /* reshape IOs share pages from .devs[0].bio */
5121 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
5122
5123 r10b->sector = r10_bio->sector;
5124 __raid10_find_phys(&conf->prev, r10b);
5125
5126 while (sectors) {
5127 int s = sectors;
5128 int success = 0;
5129 int first_slot = slot;
5130
5131 if (s > (PAGE_SIZE >> 9))
5132 s = PAGE_SIZE >> 9;
5133
5134 rcu_read_lock();
5135 while (!success) {
5136 int d = r10b->devs[slot].devnum;
5137 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
5138 sector_t addr;
5139 if (rdev == NULL ||
5140 test_bit(Faulty, &rdev->flags) ||
5141 !test_bit(In_sync, &rdev->flags))
5142 goto failed;
5143
5144 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
5145 atomic_inc(&rdev->nr_pending);
5146 rcu_read_unlock();
5147 success = sync_page_io(rdev,
5148 addr,
5149 s << 9,
5150 pages[idx],
5151 REQ_OP_READ, false);
5152 rdev_dec_pending(rdev, mddev);
5153 rcu_read_lock();
5154 if (success)
5155 break;
5156 failed:
5157 slot++;
5158 if (slot >= conf->copies)
5159 slot = 0;
5160 if (slot == first_slot)
5161 break;
5162 }
5163 rcu_read_unlock();
5164 if (!success) {
5165 /* couldn't read this block, must give up */
5166 set_bit(MD_RECOVERY_INTR,
5167 &mddev->recovery);
5168 kfree(r10b);
5169 return -EIO;
5170 }
5171 sectors -= s;
5172 idx++;
5173 }
5174 kfree(r10b);
5175 return 0;
5176}
5177
5178static void end_reshape_write(struct bio *bio)
5179{
5180 struct r10bio *r10_bio = get_resync_r10bio(bio);
5181 struct mddev *mddev = r10_bio->mddev;
5182 struct r10conf *conf = mddev->private;
5183 int d;
5184 int slot;
5185 int repl;
5186 struct md_rdev *rdev = NULL;
5187
5188 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
5189 if (repl)
5190 rdev = conf->mirrors[d].replacement;
5191 if (!rdev) {
5192 smp_mb();
5193 rdev = conf->mirrors[d].rdev;
5194 }
5195
5196 if (bio->bi_status) {
5197 /* FIXME should record badblock */
5198 md_error(mddev, rdev);
5199 }
5200
5201 rdev_dec_pending(rdev, mddev);
5202 end_reshape_request(r10_bio);
5203}
5204
5205static void end_reshape_request(struct r10bio *r10_bio)
5206{
5207 if (!atomic_dec_and_test(&r10_bio->remaining))
5208 return;
5209 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
5210 bio_put(r10_bio->master_bio);
5211 put_buf(r10_bio);
5212}
5213
5214static void raid10_finish_reshape(struct mddev *mddev)
5215{
5216 struct r10conf *conf = mddev->private;
5217
5218 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5219 return;
5220
5221 if (mddev->delta_disks > 0) {
5222 if (mddev->recovery_cp > mddev->resync_max_sectors) {
5223 mddev->recovery_cp = mddev->resync_max_sectors;
5224 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5225 }
5226 mddev->resync_max_sectors = mddev->array_sectors;
5227 } else {
5228 int d;
5229 rcu_read_lock();
5230 for (d = conf->geo.raid_disks ;
5231 d < conf->geo.raid_disks - mddev->delta_disks;
5232 d++) {
5233 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
5234 if (rdev)
5235 clear_bit(In_sync, &rdev->flags);
5236 rdev = rcu_dereference(conf->mirrors[d].replacement);
5237 if (rdev)
5238 clear_bit(In_sync, &rdev->flags);
5239 }
5240 rcu_read_unlock();
5241 }
5242 mddev->layout = mddev->new_layout;
5243 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
5244 mddev->reshape_position = MaxSector;
5245 mddev->delta_disks = 0;
5246 mddev->reshape_backwards = 0;
5247}
5248
5249static struct md_personality raid10_personality =
5250{
5251 .name = "raid10",
5252 .level = 10,
5253 .owner = THIS_MODULE,
5254 .make_request = raid10_make_request,
5255 .run = raid10_run,
5256 .free = raid10_free,
5257 .status = raid10_status,
5258 .error_handler = raid10_error,
5259 .hot_add_disk = raid10_add_disk,
5260 .hot_remove_disk= raid10_remove_disk,
5261 .spare_active = raid10_spare_active,
5262 .sync_request = raid10_sync_request,
5263 .quiesce = raid10_quiesce,
5264 .size = raid10_size,
5265 .resize = raid10_resize,
5266 .takeover = raid10_takeover,
5267 .check_reshape = raid10_check_reshape,
5268 .start_reshape = raid10_start_reshape,
5269 .finish_reshape = raid10_finish_reshape,
5270 .update_reshape_pos = raid10_update_reshape_pos,
5271};
5272
5273static int __init raid_init(void)
5274{
5275 return register_md_personality(&raid10_personality);
5276}
5277
5278static void raid_exit(void)
5279{
5280 unregister_md_personality(&raid10_personality);
5281}
5282
5283module_init(raid_init);
5284module_exit(raid_exit);
5285MODULE_LICENSE("GPL");
5286MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
5287MODULE_ALIAS("md-personality-9"); /* RAID10 */
5288MODULE_ALIAS("md-raid10");
5289MODULE_ALIAS("md-level-10");