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/*
2 * raid10.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 2000-2004 Neil Brown
5 *
6 * RAID-10 support for md.
7 *
8 * Base on code in raid1.c. See raid1.c for further copyright information.
9 *
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20
21#include <linux/slab.h>
22#include <linux/delay.h>
23#include <linux/blkdev.h>
24#include <linux/module.h>
25#include <linux/seq_file.h>
26#include <linux/ratelimit.h>
27#include <linux/kthread.h>
28#include <trace/events/block.h>
29#include "md.h"
30#include "raid10.h"
31#include "raid0.h"
32#include "md-bitmap.h"
33
34/*
35 * RAID10 provides a combination of RAID0 and RAID1 functionality.
36 * The layout of data is defined by
37 * chunk_size
38 * raid_disks
39 * near_copies (stored in low byte of layout)
40 * far_copies (stored in second byte of layout)
41 * far_offset (stored in bit 16 of layout )
42 * use_far_sets (stored in bit 17 of layout )
43 * use_far_sets_bugfixed (stored in bit 18 of layout )
44 *
45 * The data to be stored is divided into chunks using chunksize. Each device
46 * is divided into far_copies sections. In each section, chunks are laid out
47 * in a style similar to raid0, but near_copies copies of each chunk is stored
48 * (each on a different drive). The starting device for each section is offset
49 * near_copies from the starting device of the previous section. Thus there
50 * are (near_copies * far_copies) of each chunk, and each is on a different
51 * drive. near_copies and far_copies must be at least one, and their product
52 * is at most raid_disks.
53 *
54 * If far_offset is true, then the far_copies are handled a bit differently.
55 * The copies are still in different stripes, but instead of being very far
56 * apart on disk, there are adjacent stripes.
57 *
58 * The far and offset algorithms are handled slightly differently if
59 * 'use_far_sets' is true. In this case, the array's devices are grouped into
60 * sets that are (near_copies * far_copies) in size. The far copied stripes
61 * are still shifted by 'near_copies' devices, but this shifting stays confined
62 * to the set rather than the entire array. This is done to improve the number
63 * of device combinations that can fail without causing the array to fail.
64 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
65 * on a device):
66 * A B C D A B C D E
67 * ... ...
68 * D A B C E A B C D
69 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
70 * [A B] [C D] [A B] [C D E]
71 * |...| |...| |...| | ... |
72 * [B A] [D C] [B A] [E C D]
73 */
74
75/*
76 * Number of guaranteed r10bios in case of extreme VM load:
77 */
78#define NR_RAID10_BIOS 256
79
80/* when we get a read error on a read-only array, we redirect to another
81 * device without failing the first device, or trying to over-write to
82 * correct the read error. To keep track of bad blocks on a per-bio
83 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
84 */
85#define IO_BLOCKED ((struct bio *)1)
86/* When we successfully write to a known bad-block, we need to remove the
87 * bad-block marking which must be done from process context. So we record
88 * the success by setting devs[n].bio to IO_MADE_GOOD
89 */
90#define IO_MADE_GOOD ((struct bio *)2)
91
92#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
93
94/* When there are this many requests queued to be written by
95 * the raid10 thread, we become 'congested' to provide back-pressure
96 * for writeback.
97 */
98static int max_queued_requests = 1024;
99
100static void allow_barrier(struct r10conf *conf);
101static void lower_barrier(struct r10conf *conf);
102static int _enough(struct r10conf *conf, int previous, int ignore);
103static int enough(struct r10conf *conf, int ignore);
104static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
105 int *skipped);
106static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
107static void end_reshape_write(struct bio *bio);
108static void end_reshape(struct r10conf *conf);
109
110#define raid10_log(md, fmt, args...) \
111 do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
112
113#include "raid1-10.c"
114
115/*
116 * for resync bio, r10bio pointer can be retrieved from the per-bio
117 * 'struct resync_pages'.
118 */
119static inline struct r10bio *get_resync_r10bio(struct bio *bio)
120{
121 return get_resync_pages(bio)->raid_bio;
122}
123
124static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
125{
126 struct r10conf *conf = data;
127 int size = offsetof(struct r10bio, devs[conf->copies]);
128
129 /* allocate a r10bio with room for raid_disks entries in the
130 * bios array */
131 return kzalloc(size, gfp_flags);
132}
133
134static void r10bio_pool_free(void *r10_bio, void *data)
135{
136 kfree(r10_bio);
137}
138
139#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
140/* amount of memory to reserve for resync requests */
141#define RESYNC_WINDOW (1024*1024)
142/* maximum number of concurrent requests, memory permitting */
143#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
144#define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
145#define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
146
147/*
148 * When performing a resync, we need to read and compare, so
149 * we need as many pages are there are copies.
150 * When performing a recovery, we need 2 bios, one for read,
151 * one for write (we recover only one drive per r10buf)
152 *
153 */
154static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
155{
156 struct r10conf *conf = data;
157 struct r10bio *r10_bio;
158 struct bio *bio;
159 int j;
160 int nalloc, nalloc_rp;
161 struct resync_pages *rps;
162
163 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
164 if (!r10_bio)
165 return NULL;
166
167 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
168 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
169 nalloc = conf->copies; /* resync */
170 else
171 nalloc = 2; /* recovery */
172
173 /* allocate once for all bios */
174 if (!conf->have_replacement)
175 nalloc_rp = nalloc;
176 else
177 nalloc_rp = nalloc * 2;
178 rps = kmalloc(sizeof(struct resync_pages) * nalloc_rp, gfp_flags);
179 if (!rps)
180 goto out_free_r10bio;
181
182 /*
183 * Allocate bios.
184 */
185 for (j = nalloc ; j-- ; ) {
186 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
187 if (!bio)
188 goto out_free_bio;
189 r10_bio->devs[j].bio = bio;
190 if (!conf->have_replacement)
191 continue;
192 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
193 if (!bio)
194 goto out_free_bio;
195 r10_bio->devs[j].repl_bio = bio;
196 }
197 /*
198 * Allocate RESYNC_PAGES data pages and attach them
199 * where needed.
200 */
201 for (j = 0; j < nalloc; j++) {
202 struct bio *rbio = r10_bio->devs[j].repl_bio;
203 struct resync_pages *rp, *rp_repl;
204
205 rp = &rps[j];
206 if (rbio)
207 rp_repl = &rps[nalloc + j];
208
209 bio = r10_bio->devs[j].bio;
210
211 if (!j || test_bit(MD_RECOVERY_SYNC,
212 &conf->mddev->recovery)) {
213 if (resync_alloc_pages(rp, gfp_flags))
214 goto out_free_pages;
215 } else {
216 memcpy(rp, &rps[0], sizeof(*rp));
217 resync_get_all_pages(rp);
218 }
219
220 rp->raid_bio = r10_bio;
221 bio->bi_private = rp;
222 if (rbio) {
223 memcpy(rp_repl, rp, sizeof(*rp));
224 rbio->bi_private = rp_repl;
225 }
226 }
227
228 return r10_bio;
229
230out_free_pages:
231 while (--j >= 0)
232 resync_free_pages(&rps[j * 2]);
233
234 j = 0;
235out_free_bio:
236 for ( ; j < nalloc; j++) {
237 if (r10_bio->devs[j].bio)
238 bio_put(r10_bio->devs[j].bio);
239 if (r10_bio->devs[j].repl_bio)
240 bio_put(r10_bio->devs[j].repl_bio);
241 }
242 kfree(rps);
243out_free_r10bio:
244 r10bio_pool_free(r10_bio, conf);
245 return NULL;
246}
247
248static void r10buf_pool_free(void *__r10_bio, void *data)
249{
250 struct r10conf *conf = data;
251 struct r10bio *r10bio = __r10_bio;
252 int j;
253 struct resync_pages *rp = NULL;
254
255 for (j = conf->copies; j--; ) {
256 struct bio *bio = r10bio->devs[j].bio;
257
258 rp = get_resync_pages(bio);
259 resync_free_pages(rp);
260 bio_put(bio);
261
262 bio = r10bio->devs[j].repl_bio;
263 if (bio)
264 bio_put(bio);
265 }
266
267 /* resync pages array stored in the 1st bio's .bi_private */
268 kfree(rp);
269
270 r10bio_pool_free(r10bio, conf);
271}
272
273static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
274{
275 int i;
276
277 for (i = 0; i < conf->copies; i++) {
278 struct bio **bio = & r10_bio->devs[i].bio;
279 if (!BIO_SPECIAL(*bio))
280 bio_put(*bio);
281 *bio = NULL;
282 bio = &r10_bio->devs[i].repl_bio;
283 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
284 bio_put(*bio);
285 *bio = NULL;
286 }
287}
288
289static void free_r10bio(struct r10bio *r10_bio)
290{
291 struct r10conf *conf = r10_bio->mddev->private;
292
293 put_all_bios(conf, r10_bio);
294 mempool_free(r10_bio, conf->r10bio_pool);
295}
296
297static void put_buf(struct r10bio *r10_bio)
298{
299 struct r10conf *conf = r10_bio->mddev->private;
300
301 mempool_free(r10_bio, conf->r10buf_pool);
302
303 lower_barrier(conf);
304}
305
306static void reschedule_retry(struct r10bio *r10_bio)
307{
308 unsigned long flags;
309 struct mddev *mddev = r10_bio->mddev;
310 struct r10conf *conf = mddev->private;
311
312 spin_lock_irqsave(&conf->device_lock, flags);
313 list_add(&r10_bio->retry_list, &conf->retry_list);
314 conf->nr_queued ++;
315 spin_unlock_irqrestore(&conf->device_lock, flags);
316
317 /* wake up frozen array... */
318 wake_up(&conf->wait_barrier);
319
320 md_wakeup_thread(mddev->thread);
321}
322
323/*
324 * raid_end_bio_io() is called when we have finished servicing a mirrored
325 * operation and are ready to return a success/failure code to the buffer
326 * cache layer.
327 */
328static void raid_end_bio_io(struct r10bio *r10_bio)
329{
330 struct bio *bio = r10_bio->master_bio;
331 struct r10conf *conf = r10_bio->mddev->private;
332
333 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
334 bio->bi_status = BLK_STS_IOERR;
335
336 bio_endio(bio);
337 /*
338 * Wake up any possible resync thread that waits for the device
339 * to go idle.
340 */
341 allow_barrier(conf);
342
343 free_r10bio(r10_bio);
344}
345
346/*
347 * Update disk head position estimator based on IRQ completion info.
348 */
349static inline void update_head_pos(int slot, struct r10bio *r10_bio)
350{
351 struct r10conf *conf = r10_bio->mddev->private;
352
353 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
354 r10_bio->devs[slot].addr + (r10_bio->sectors);
355}
356
357/*
358 * Find the disk number which triggered given bio
359 */
360static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
361 struct bio *bio, int *slotp, int *replp)
362{
363 int slot;
364 int repl = 0;
365
366 for (slot = 0; slot < conf->copies; slot++) {
367 if (r10_bio->devs[slot].bio == bio)
368 break;
369 if (r10_bio->devs[slot].repl_bio == bio) {
370 repl = 1;
371 break;
372 }
373 }
374
375 BUG_ON(slot == conf->copies);
376 update_head_pos(slot, r10_bio);
377
378 if (slotp)
379 *slotp = slot;
380 if (replp)
381 *replp = repl;
382 return r10_bio->devs[slot].devnum;
383}
384
385static void raid10_end_read_request(struct bio *bio)
386{
387 int uptodate = !bio->bi_status;
388 struct r10bio *r10_bio = bio->bi_private;
389 int slot;
390 struct md_rdev *rdev;
391 struct r10conf *conf = r10_bio->mddev->private;
392
393 slot = r10_bio->read_slot;
394 rdev = r10_bio->devs[slot].rdev;
395 /*
396 * this branch is our 'one mirror IO has finished' event handler:
397 */
398 update_head_pos(slot, r10_bio);
399
400 if (uptodate) {
401 /*
402 * Set R10BIO_Uptodate in our master bio, so that
403 * we will return a good error code to the higher
404 * levels even if IO on some other mirrored buffer fails.
405 *
406 * The 'master' represents the composite IO operation to
407 * user-side. So if something waits for IO, then it will
408 * wait for the 'master' bio.
409 */
410 set_bit(R10BIO_Uptodate, &r10_bio->state);
411 } else {
412 /* If all other devices that store this block have
413 * failed, we want to return the error upwards rather
414 * than fail the last device. Here we redefine
415 * "uptodate" to mean "Don't want to retry"
416 */
417 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
418 rdev->raid_disk))
419 uptodate = 1;
420 }
421 if (uptodate) {
422 raid_end_bio_io(r10_bio);
423 rdev_dec_pending(rdev, conf->mddev);
424 } else {
425 /*
426 * oops, read error - keep the refcount on the rdev
427 */
428 char b[BDEVNAME_SIZE];
429 pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
430 mdname(conf->mddev),
431 bdevname(rdev->bdev, b),
432 (unsigned long long)r10_bio->sector);
433 set_bit(R10BIO_ReadError, &r10_bio->state);
434 reschedule_retry(r10_bio);
435 }
436}
437
438static void close_write(struct r10bio *r10_bio)
439{
440 /* clear the bitmap if all writes complete successfully */
441 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
442 r10_bio->sectors,
443 !test_bit(R10BIO_Degraded, &r10_bio->state),
444 0);
445 md_write_end(r10_bio->mddev);
446}
447
448static void one_write_done(struct r10bio *r10_bio)
449{
450 if (atomic_dec_and_test(&r10_bio->remaining)) {
451 if (test_bit(R10BIO_WriteError, &r10_bio->state))
452 reschedule_retry(r10_bio);
453 else {
454 close_write(r10_bio);
455 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
456 reschedule_retry(r10_bio);
457 else
458 raid_end_bio_io(r10_bio);
459 }
460 }
461}
462
463static void raid10_end_write_request(struct bio *bio)
464{
465 struct r10bio *r10_bio = bio->bi_private;
466 int dev;
467 int dec_rdev = 1;
468 struct r10conf *conf = r10_bio->mddev->private;
469 int slot, repl;
470 struct md_rdev *rdev = NULL;
471 struct bio *to_put = NULL;
472 bool discard_error;
473
474 discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
475
476 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
477
478 if (repl)
479 rdev = conf->mirrors[dev].replacement;
480 if (!rdev) {
481 smp_rmb();
482 repl = 0;
483 rdev = conf->mirrors[dev].rdev;
484 }
485 /*
486 * this branch is our 'one mirror IO has finished' event handler:
487 */
488 if (bio->bi_status && !discard_error) {
489 if (repl)
490 /* Never record new bad blocks to replacement,
491 * just fail it.
492 */
493 md_error(rdev->mddev, rdev);
494 else {
495 set_bit(WriteErrorSeen, &rdev->flags);
496 if (!test_and_set_bit(WantReplacement, &rdev->flags))
497 set_bit(MD_RECOVERY_NEEDED,
498 &rdev->mddev->recovery);
499
500 dec_rdev = 0;
501 if (test_bit(FailFast, &rdev->flags) &&
502 (bio->bi_opf & MD_FAILFAST)) {
503 md_error(rdev->mddev, rdev);
504 if (!test_bit(Faulty, &rdev->flags))
505 /* This is the only remaining device,
506 * We need to retry the write without
507 * FailFast
508 */
509 set_bit(R10BIO_WriteError, &r10_bio->state);
510 else {
511 r10_bio->devs[slot].bio = NULL;
512 to_put = bio;
513 dec_rdev = 1;
514 }
515 } else
516 set_bit(R10BIO_WriteError, &r10_bio->state);
517 }
518 } else {
519 /*
520 * Set R10BIO_Uptodate in our master bio, so that
521 * we will return a good error code for to the higher
522 * levels even if IO on some other mirrored buffer fails.
523 *
524 * The 'master' represents the composite IO operation to
525 * user-side. So if something waits for IO, then it will
526 * wait for the 'master' bio.
527 */
528 sector_t first_bad;
529 int bad_sectors;
530
531 /*
532 * Do not set R10BIO_Uptodate if the current device is
533 * rebuilding or Faulty. This is because we cannot use
534 * such device for properly reading the data back (we could
535 * potentially use it, if the current write would have felt
536 * before rdev->recovery_offset, but for simplicity we don't
537 * check this here.
538 */
539 if (test_bit(In_sync, &rdev->flags) &&
540 !test_bit(Faulty, &rdev->flags))
541 set_bit(R10BIO_Uptodate, &r10_bio->state);
542
543 /* Maybe we can clear some bad blocks. */
544 if (is_badblock(rdev,
545 r10_bio->devs[slot].addr,
546 r10_bio->sectors,
547 &first_bad, &bad_sectors) && !discard_error) {
548 bio_put(bio);
549 if (repl)
550 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
551 else
552 r10_bio->devs[slot].bio = IO_MADE_GOOD;
553 dec_rdev = 0;
554 set_bit(R10BIO_MadeGood, &r10_bio->state);
555 }
556 }
557
558 /*
559 *
560 * Let's see if all mirrored write operations have finished
561 * already.
562 */
563 one_write_done(r10_bio);
564 if (dec_rdev)
565 rdev_dec_pending(rdev, conf->mddev);
566 if (to_put)
567 bio_put(to_put);
568}
569
570/*
571 * RAID10 layout manager
572 * As well as the chunksize and raid_disks count, there are two
573 * parameters: near_copies and far_copies.
574 * near_copies * far_copies must be <= raid_disks.
575 * Normally one of these will be 1.
576 * If both are 1, we get raid0.
577 * If near_copies == raid_disks, we get raid1.
578 *
579 * Chunks are laid out in raid0 style with near_copies copies of the
580 * first chunk, followed by near_copies copies of the next chunk and
581 * so on.
582 * If far_copies > 1, then after 1/far_copies of the array has been assigned
583 * as described above, we start again with a device offset of near_copies.
584 * So we effectively have another copy of the whole array further down all
585 * the drives, but with blocks on different drives.
586 * With this layout, and block is never stored twice on the one device.
587 *
588 * raid10_find_phys finds the sector offset of a given virtual sector
589 * on each device that it is on.
590 *
591 * raid10_find_virt does the reverse mapping, from a device and a
592 * sector offset to a virtual address
593 */
594
595static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
596{
597 int n,f;
598 sector_t sector;
599 sector_t chunk;
600 sector_t stripe;
601 int dev;
602 int slot = 0;
603 int last_far_set_start, last_far_set_size;
604
605 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
606 last_far_set_start *= geo->far_set_size;
607
608 last_far_set_size = geo->far_set_size;
609 last_far_set_size += (geo->raid_disks % geo->far_set_size);
610
611 /* now calculate first sector/dev */
612 chunk = r10bio->sector >> geo->chunk_shift;
613 sector = r10bio->sector & geo->chunk_mask;
614
615 chunk *= geo->near_copies;
616 stripe = chunk;
617 dev = sector_div(stripe, geo->raid_disks);
618 if (geo->far_offset)
619 stripe *= geo->far_copies;
620
621 sector += stripe << geo->chunk_shift;
622
623 /* and calculate all the others */
624 for (n = 0; n < geo->near_copies; n++) {
625 int d = dev;
626 int set;
627 sector_t s = sector;
628 r10bio->devs[slot].devnum = d;
629 r10bio->devs[slot].addr = s;
630 slot++;
631
632 for (f = 1; f < geo->far_copies; f++) {
633 set = d / geo->far_set_size;
634 d += geo->near_copies;
635
636 if ((geo->raid_disks % geo->far_set_size) &&
637 (d > last_far_set_start)) {
638 d -= last_far_set_start;
639 d %= last_far_set_size;
640 d += last_far_set_start;
641 } else {
642 d %= geo->far_set_size;
643 d += geo->far_set_size * set;
644 }
645 s += geo->stride;
646 r10bio->devs[slot].devnum = d;
647 r10bio->devs[slot].addr = s;
648 slot++;
649 }
650 dev++;
651 if (dev >= geo->raid_disks) {
652 dev = 0;
653 sector += (geo->chunk_mask + 1);
654 }
655 }
656}
657
658static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
659{
660 struct geom *geo = &conf->geo;
661
662 if (conf->reshape_progress != MaxSector &&
663 ((r10bio->sector >= conf->reshape_progress) !=
664 conf->mddev->reshape_backwards)) {
665 set_bit(R10BIO_Previous, &r10bio->state);
666 geo = &conf->prev;
667 } else
668 clear_bit(R10BIO_Previous, &r10bio->state);
669
670 __raid10_find_phys(geo, r10bio);
671}
672
673static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
674{
675 sector_t offset, chunk, vchunk;
676 /* Never use conf->prev as this is only called during resync
677 * or recovery, so reshape isn't happening
678 */
679 struct geom *geo = &conf->geo;
680 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
681 int far_set_size = geo->far_set_size;
682 int last_far_set_start;
683
684 if (geo->raid_disks % geo->far_set_size) {
685 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
686 last_far_set_start *= geo->far_set_size;
687
688 if (dev >= last_far_set_start) {
689 far_set_size = geo->far_set_size;
690 far_set_size += (geo->raid_disks % geo->far_set_size);
691 far_set_start = last_far_set_start;
692 }
693 }
694
695 offset = sector & geo->chunk_mask;
696 if (geo->far_offset) {
697 int fc;
698 chunk = sector >> geo->chunk_shift;
699 fc = sector_div(chunk, geo->far_copies);
700 dev -= fc * geo->near_copies;
701 if (dev < far_set_start)
702 dev += far_set_size;
703 } else {
704 while (sector >= geo->stride) {
705 sector -= geo->stride;
706 if (dev < (geo->near_copies + far_set_start))
707 dev += far_set_size - geo->near_copies;
708 else
709 dev -= geo->near_copies;
710 }
711 chunk = sector >> geo->chunk_shift;
712 }
713 vchunk = chunk * geo->raid_disks + dev;
714 sector_div(vchunk, geo->near_copies);
715 return (vchunk << geo->chunk_shift) + offset;
716}
717
718/*
719 * This routine returns the disk from which the requested read should
720 * be done. There is a per-array 'next expected sequential IO' sector
721 * number - if this matches on the next IO then we use the last disk.
722 * There is also a per-disk 'last know head position' sector that is
723 * maintained from IRQ contexts, both the normal and the resync IO
724 * completion handlers update this position correctly. If there is no
725 * perfect sequential match then we pick the disk whose head is closest.
726 *
727 * If there are 2 mirrors in the same 2 devices, performance degrades
728 * because position is mirror, not device based.
729 *
730 * The rdev for the device selected will have nr_pending incremented.
731 */
732
733/*
734 * FIXME: possibly should rethink readbalancing and do it differently
735 * depending on near_copies / far_copies geometry.
736 */
737static struct md_rdev *read_balance(struct r10conf *conf,
738 struct r10bio *r10_bio,
739 int *max_sectors)
740{
741 const sector_t this_sector = r10_bio->sector;
742 int disk, slot;
743 int sectors = r10_bio->sectors;
744 int best_good_sectors;
745 sector_t new_distance, best_dist;
746 struct md_rdev *best_rdev, *rdev = NULL;
747 int do_balance;
748 int best_slot;
749 struct geom *geo = &conf->geo;
750
751 raid10_find_phys(conf, r10_bio);
752 rcu_read_lock();
753 best_slot = -1;
754 best_rdev = NULL;
755 best_dist = MaxSector;
756 best_good_sectors = 0;
757 do_balance = 1;
758 clear_bit(R10BIO_FailFast, &r10_bio->state);
759 /*
760 * Check if we can balance. We can balance on the whole
761 * device if no resync is going on (recovery is ok), or below
762 * the resync window. We take the first readable disk when
763 * above the resync window.
764 */
765 if ((conf->mddev->recovery_cp < MaxSector
766 && (this_sector + sectors >= conf->next_resync)) ||
767 (mddev_is_clustered(conf->mddev) &&
768 md_cluster_ops->area_resyncing(conf->mddev, READ, this_sector,
769 this_sector + sectors)))
770 do_balance = 0;
771
772 for (slot = 0; slot < conf->copies ; slot++) {
773 sector_t first_bad;
774 int bad_sectors;
775 sector_t dev_sector;
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_slot = slot;
813 best_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 if (best_slot >= 0)
827 /* At least 2 disks to choose from so failfast is OK */
828 set_bit(R10BIO_FailFast, &r10_bio->state);
829 /* This optimisation is debatable, and completely destroys
830 * sequential read speed for 'far copies' arrays. So only
831 * keep it for 'near' arrays, and review those later.
832 */
833 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
834 new_distance = 0;
835
836 /* for far > 1 always use the lowest address */
837 else if (geo->far_copies > 1)
838 new_distance = r10_bio->devs[slot].addr;
839 else
840 new_distance = abs(r10_bio->devs[slot].addr -
841 conf->mirrors[disk].head_position);
842 if (new_distance < best_dist) {
843 best_dist = new_distance;
844 best_slot = slot;
845 best_rdev = rdev;
846 }
847 }
848 if (slot >= conf->copies) {
849 slot = best_slot;
850 rdev = best_rdev;
851 }
852
853 if (slot >= 0) {
854 atomic_inc(&rdev->nr_pending);
855 r10_bio->read_slot = slot;
856 } else
857 rdev = NULL;
858 rcu_read_unlock();
859 *max_sectors = best_good_sectors;
860
861 return rdev;
862}
863
864static int raid10_congested(struct mddev *mddev, int bits)
865{
866 struct r10conf *conf = mddev->private;
867 int i, ret = 0;
868
869 if ((bits & (1 << WB_async_congested)) &&
870 conf->pending_count >= max_queued_requests)
871 return 1;
872
873 rcu_read_lock();
874 for (i = 0;
875 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
876 && ret == 0;
877 i++) {
878 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
879 if (rdev && !test_bit(Faulty, &rdev->flags)) {
880 struct request_queue *q = bdev_get_queue(rdev->bdev);
881
882 ret |= bdi_congested(q->backing_dev_info, bits);
883 }
884 }
885 rcu_read_unlock();
886 return ret;
887}
888
889static void flush_pending_writes(struct r10conf *conf)
890{
891 /* Any writes that have been queued but are awaiting
892 * bitmap updates get flushed here.
893 */
894 spin_lock_irq(&conf->device_lock);
895
896 if (conf->pending_bio_list.head) {
897 struct blk_plug plug;
898 struct bio *bio;
899
900 bio = bio_list_get(&conf->pending_bio_list);
901 conf->pending_count = 0;
902 spin_unlock_irq(&conf->device_lock);
903
904 /*
905 * As this is called in a wait_event() loop (see freeze_array),
906 * current->state might be TASK_UNINTERRUPTIBLE which will
907 * cause a warning when we prepare to wait again. As it is
908 * rare that this path is taken, it is perfectly safe to force
909 * us to go around the wait_event() loop again, so the warning
910 * is a false-positive. Silence the warning by resetting
911 * thread state
912 */
913 __set_current_state(TASK_RUNNING);
914
915 blk_start_plug(&plug);
916 /* flush any pending bitmap writes to disk
917 * before proceeding w/ I/O */
918 bitmap_unplug(conf->mddev->bitmap);
919 wake_up(&conf->wait_barrier);
920
921 while (bio) { /* submit pending writes */
922 struct bio *next = bio->bi_next;
923 struct md_rdev *rdev = (void*)bio->bi_disk;
924 bio->bi_next = NULL;
925 bio_set_dev(bio, rdev->bdev);
926 if (test_bit(Faulty, &rdev->flags)) {
927 bio_io_error(bio);
928 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
929 !blk_queue_discard(bio->bi_disk->queue)))
930 /* Just ignore it */
931 bio_endio(bio);
932 else
933 generic_make_request(bio);
934 bio = next;
935 }
936 blk_finish_plug(&plug);
937 } else
938 spin_unlock_irq(&conf->device_lock);
939}
940
941/* Barriers....
942 * Sometimes we need to suspend IO while we do something else,
943 * either some resync/recovery, or reconfigure the array.
944 * To do this we raise a 'barrier'.
945 * The 'barrier' is a counter that can be raised multiple times
946 * to count how many activities are happening which preclude
947 * normal IO.
948 * We can only raise the barrier if there is no pending IO.
949 * i.e. if nr_pending == 0.
950 * We choose only to raise the barrier if no-one is waiting for the
951 * barrier to go down. This means that as soon as an IO request
952 * is ready, no other operations which require a barrier will start
953 * until the IO request has had a chance.
954 *
955 * So: regular IO calls 'wait_barrier'. When that returns there
956 * is no backgroup IO happening, It must arrange to call
957 * allow_barrier when it has finished its IO.
958 * backgroup IO calls must call raise_barrier. Once that returns
959 * there is no normal IO happeing. It must arrange to call
960 * lower_barrier when the particular background IO completes.
961 */
962
963static void raise_barrier(struct r10conf *conf, int force)
964{
965 BUG_ON(force && !conf->barrier);
966 spin_lock_irq(&conf->resync_lock);
967
968 /* Wait until no block IO is waiting (unless 'force') */
969 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
970 conf->resync_lock);
971
972 /* block any new IO from starting */
973 conf->barrier++;
974
975 /* Now wait for all pending IO to complete */
976 wait_event_lock_irq(conf->wait_barrier,
977 !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
978 conf->resync_lock);
979
980 spin_unlock_irq(&conf->resync_lock);
981}
982
983static void lower_barrier(struct r10conf *conf)
984{
985 unsigned long flags;
986 spin_lock_irqsave(&conf->resync_lock, flags);
987 conf->barrier--;
988 spin_unlock_irqrestore(&conf->resync_lock, flags);
989 wake_up(&conf->wait_barrier);
990}
991
992static void wait_barrier(struct r10conf *conf)
993{
994 spin_lock_irq(&conf->resync_lock);
995 if (conf->barrier) {
996 conf->nr_waiting++;
997 /* Wait for the barrier to drop.
998 * However if there are already pending
999 * requests (preventing the barrier from
1000 * rising completely), and the
1001 * pre-process bio queue isn't empty,
1002 * then don't wait, as we need to empty
1003 * that queue to get the nr_pending
1004 * count down.
1005 */
1006 raid10_log(conf->mddev, "wait barrier");
1007 wait_event_lock_irq(conf->wait_barrier,
1008 !conf->barrier ||
1009 (atomic_read(&conf->nr_pending) &&
1010 current->bio_list &&
1011 (!bio_list_empty(¤t->bio_list[0]) ||
1012 !bio_list_empty(¤t->bio_list[1]))),
1013 conf->resync_lock);
1014 conf->nr_waiting--;
1015 if (!conf->nr_waiting)
1016 wake_up(&conf->wait_barrier);
1017 }
1018 atomic_inc(&conf->nr_pending);
1019 spin_unlock_irq(&conf->resync_lock);
1020}
1021
1022static void allow_barrier(struct r10conf *conf)
1023{
1024 if ((atomic_dec_and_test(&conf->nr_pending)) ||
1025 (conf->array_freeze_pending))
1026 wake_up(&conf->wait_barrier);
1027}
1028
1029static void freeze_array(struct r10conf *conf, int extra)
1030{
1031 /* stop syncio and normal IO and wait for everything to
1032 * go quiet.
1033 * We increment barrier and nr_waiting, and then
1034 * wait until nr_pending match nr_queued+extra
1035 * This is called in the context of one normal IO request
1036 * that has failed. Thus any sync request that might be pending
1037 * will be blocked by nr_pending, and we need to wait for
1038 * pending IO requests to complete or be queued for re-try.
1039 * Thus the number queued (nr_queued) plus this request (extra)
1040 * must match the number of pending IOs (nr_pending) before
1041 * we continue.
1042 */
1043 spin_lock_irq(&conf->resync_lock);
1044 conf->array_freeze_pending++;
1045 conf->barrier++;
1046 conf->nr_waiting++;
1047 wait_event_lock_irq_cmd(conf->wait_barrier,
1048 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1049 conf->resync_lock,
1050 flush_pending_writes(conf));
1051
1052 conf->array_freeze_pending--;
1053 spin_unlock_irq(&conf->resync_lock);
1054}
1055
1056static void unfreeze_array(struct r10conf *conf)
1057{
1058 /* reverse the effect of the freeze */
1059 spin_lock_irq(&conf->resync_lock);
1060 conf->barrier--;
1061 conf->nr_waiting--;
1062 wake_up(&conf->wait_barrier);
1063 spin_unlock_irq(&conf->resync_lock);
1064}
1065
1066static sector_t choose_data_offset(struct r10bio *r10_bio,
1067 struct md_rdev *rdev)
1068{
1069 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1070 test_bit(R10BIO_Previous, &r10_bio->state))
1071 return rdev->data_offset;
1072 else
1073 return rdev->new_data_offset;
1074}
1075
1076struct raid10_plug_cb {
1077 struct blk_plug_cb cb;
1078 struct bio_list pending;
1079 int pending_cnt;
1080};
1081
1082static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1083{
1084 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1085 cb);
1086 struct mddev *mddev = plug->cb.data;
1087 struct r10conf *conf = mddev->private;
1088 struct bio *bio;
1089
1090 if (from_schedule || current->bio_list) {
1091 spin_lock_irq(&conf->device_lock);
1092 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1093 conf->pending_count += plug->pending_cnt;
1094 spin_unlock_irq(&conf->device_lock);
1095 wake_up(&conf->wait_barrier);
1096 md_wakeup_thread(mddev->thread);
1097 kfree(plug);
1098 return;
1099 }
1100
1101 /* we aren't scheduling, so we can do the write-out directly. */
1102 bio = bio_list_get(&plug->pending);
1103 bitmap_unplug(mddev->bitmap);
1104 wake_up(&conf->wait_barrier);
1105
1106 while (bio) { /* submit pending writes */
1107 struct bio *next = bio->bi_next;
1108 struct md_rdev *rdev = (void*)bio->bi_disk;
1109 bio->bi_next = NULL;
1110 bio_set_dev(bio, rdev->bdev);
1111 if (test_bit(Faulty, &rdev->flags)) {
1112 bio_io_error(bio);
1113 } else if (unlikely((bio_op(bio) == REQ_OP_DISCARD) &&
1114 !blk_queue_discard(bio->bi_disk->queue)))
1115 /* Just ignore it */
1116 bio_endio(bio);
1117 else
1118 generic_make_request(bio);
1119 bio = next;
1120 }
1121 kfree(plug);
1122}
1123
1124static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1125 struct r10bio *r10_bio)
1126{
1127 struct r10conf *conf = mddev->private;
1128 struct bio *read_bio;
1129 const int op = bio_op(bio);
1130 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1131 int max_sectors;
1132 sector_t sectors;
1133 struct md_rdev *rdev;
1134 char b[BDEVNAME_SIZE];
1135 int slot = r10_bio->read_slot;
1136 struct md_rdev *err_rdev = NULL;
1137 gfp_t gfp = GFP_NOIO;
1138
1139 if (r10_bio->devs[slot].rdev) {
1140 /*
1141 * This is an error retry, but we cannot
1142 * safely dereference the rdev in the r10_bio,
1143 * we must use the one in conf.
1144 * If it has already been disconnected (unlikely)
1145 * we lose the device name in error messages.
1146 */
1147 int disk;
1148 /*
1149 * As we are blocking raid10, it is a little safer to
1150 * use __GFP_HIGH.
1151 */
1152 gfp = GFP_NOIO | __GFP_HIGH;
1153
1154 rcu_read_lock();
1155 disk = r10_bio->devs[slot].devnum;
1156 err_rdev = rcu_dereference(conf->mirrors[disk].rdev);
1157 if (err_rdev)
1158 bdevname(err_rdev->bdev, b);
1159 else {
1160 strcpy(b, "???");
1161 /* This never gets dereferenced */
1162 err_rdev = r10_bio->devs[slot].rdev;
1163 }
1164 rcu_read_unlock();
1165 }
1166 /*
1167 * Register the new request and wait if the reconstruction
1168 * thread has put up a bar for new requests.
1169 * Continue immediately if no resync is active currently.
1170 */
1171 wait_barrier(conf);
1172
1173 sectors = r10_bio->sectors;
1174 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1175 bio->bi_iter.bi_sector < conf->reshape_progress &&
1176 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1177 /*
1178 * IO spans the reshape position. Need to wait for reshape to
1179 * pass
1180 */
1181 raid10_log(conf->mddev, "wait reshape");
1182 allow_barrier(conf);
1183 wait_event(conf->wait_barrier,
1184 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1185 conf->reshape_progress >= bio->bi_iter.bi_sector +
1186 sectors);
1187 wait_barrier(conf);
1188 }
1189
1190 rdev = read_balance(conf, r10_bio, &max_sectors);
1191 if (!rdev) {
1192 if (err_rdev) {
1193 pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1194 mdname(mddev), b,
1195 (unsigned long long)r10_bio->sector);
1196 }
1197 raid_end_bio_io(r10_bio);
1198 return;
1199 }
1200 if (err_rdev)
1201 pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
1202 mdname(mddev),
1203 bdevname(rdev->bdev, b),
1204 (unsigned long long)r10_bio->sector);
1205 if (max_sectors < bio_sectors(bio)) {
1206 struct bio *split = bio_split(bio, max_sectors,
1207 gfp, conf->bio_split);
1208 bio_chain(split, bio);
1209 generic_make_request(bio);
1210 bio = split;
1211 r10_bio->master_bio = bio;
1212 r10_bio->sectors = max_sectors;
1213 }
1214 slot = r10_bio->read_slot;
1215
1216 read_bio = bio_clone_fast(bio, gfp, mddev->bio_set);
1217
1218 r10_bio->devs[slot].bio = read_bio;
1219 r10_bio->devs[slot].rdev = rdev;
1220
1221 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1222 choose_data_offset(r10_bio, rdev);
1223 bio_set_dev(read_bio, rdev->bdev);
1224 read_bio->bi_end_io = raid10_end_read_request;
1225 bio_set_op_attrs(read_bio, op, do_sync);
1226 if (test_bit(FailFast, &rdev->flags) &&
1227 test_bit(R10BIO_FailFast, &r10_bio->state))
1228 read_bio->bi_opf |= MD_FAILFAST;
1229 read_bio->bi_private = r10_bio;
1230
1231 if (mddev->gendisk)
1232 trace_block_bio_remap(read_bio->bi_disk->queue,
1233 read_bio, disk_devt(mddev->gendisk),
1234 r10_bio->sector);
1235 generic_make_request(read_bio);
1236 return;
1237}
1238
1239static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1240 struct bio *bio, bool replacement,
1241 int n_copy)
1242{
1243 const int op = bio_op(bio);
1244 const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1245 const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1246 unsigned long flags;
1247 struct blk_plug_cb *cb;
1248 struct raid10_plug_cb *plug = NULL;
1249 struct r10conf *conf = mddev->private;
1250 struct md_rdev *rdev;
1251 int devnum = r10_bio->devs[n_copy].devnum;
1252 struct bio *mbio;
1253
1254 if (replacement) {
1255 rdev = conf->mirrors[devnum].replacement;
1256 if (rdev == NULL) {
1257 /* Replacement just got moved to main 'rdev' */
1258 smp_mb();
1259 rdev = conf->mirrors[devnum].rdev;
1260 }
1261 } else
1262 rdev = conf->mirrors[devnum].rdev;
1263
1264 mbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
1265 if (replacement)
1266 r10_bio->devs[n_copy].repl_bio = mbio;
1267 else
1268 r10_bio->devs[n_copy].bio = mbio;
1269
1270 mbio->bi_iter.bi_sector = (r10_bio->devs[n_copy].addr +
1271 choose_data_offset(r10_bio, rdev));
1272 bio_set_dev(mbio, rdev->bdev);
1273 mbio->bi_end_io = raid10_end_write_request;
1274 bio_set_op_attrs(mbio, op, do_sync | do_fua);
1275 if (!replacement && test_bit(FailFast,
1276 &conf->mirrors[devnum].rdev->flags)
1277 && enough(conf, devnum))
1278 mbio->bi_opf |= MD_FAILFAST;
1279 mbio->bi_private = r10_bio;
1280
1281 if (conf->mddev->gendisk)
1282 trace_block_bio_remap(mbio->bi_disk->queue,
1283 mbio, disk_devt(conf->mddev->gendisk),
1284 r10_bio->sector);
1285 /* flush_pending_writes() needs access to the rdev so...*/
1286 mbio->bi_disk = (void *)rdev;
1287
1288 atomic_inc(&r10_bio->remaining);
1289
1290 cb = blk_check_plugged(raid10_unplug, mddev, sizeof(*plug));
1291 if (cb)
1292 plug = container_of(cb, struct raid10_plug_cb, cb);
1293 else
1294 plug = NULL;
1295 if (plug) {
1296 bio_list_add(&plug->pending, mbio);
1297 plug->pending_cnt++;
1298 } else {
1299 spin_lock_irqsave(&conf->device_lock, flags);
1300 bio_list_add(&conf->pending_bio_list, mbio);
1301 conf->pending_count++;
1302 spin_unlock_irqrestore(&conf->device_lock, flags);
1303 md_wakeup_thread(mddev->thread);
1304 }
1305}
1306
1307static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1308 struct r10bio *r10_bio)
1309{
1310 struct r10conf *conf = mddev->private;
1311 int i;
1312 struct md_rdev *blocked_rdev;
1313 sector_t sectors;
1314 int max_sectors;
1315
1316 if ((mddev_is_clustered(mddev) &&
1317 md_cluster_ops->area_resyncing(mddev, WRITE,
1318 bio->bi_iter.bi_sector,
1319 bio_end_sector(bio)))) {
1320 DEFINE_WAIT(w);
1321 for (;;) {
1322 prepare_to_wait(&conf->wait_barrier,
1323 &w, TASK_IDLE);
1324 if (!md_cluster_ops->area_resyncing(mddev, WRITE,
1325 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1326 break;
1327 schedule();
1328 }
1329 finish_wait(&conf->wait_barrier, &w);
1330 }
1331
1332 /*
1333 * Register the new request and wait if the reconstruction
1334 * thread has put up a bar for new requests.
1335 * Continue immediately if no resync is active currently.
1336 */
1337 wait_barrier(conf);
1338
1339 sectors = r10_bio->sectors;
1340 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1341 bio->bi_iter.bi_sector < conf->reshape_progress &&
1342 bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1343 /*
1344 * IO spans the reshape position. Need to wait for reshape to
1345 * pass
1346 */
1347 raid10_log(conf->mddev, "wait reshape");
1348 allow_barrier(conf);
1349 wait_event(conf->wait_barrier,
1350 conf->reshape_progress <= bio->bi_iter.bi_sector ||
1351 conf->reshape_progress >= bio->bi_iter.bi_sector +
1352 sectors);
1353 wait_barrier(conf);
1354 }
1355
1356 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1357 (mddev->reshape_backwards
1358 ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1359 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1360 : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1361 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1362 /* Need to update reshape_position in metadata */
1363 mddev->reshape_position = conf->reshape_progress;
1364 set_mask_bits(&mddev->sb_flags, 0,
1365 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1366 md_wakeup_thread(mddev->thread);
1367 raid10_log(conf->mddev, "wait reshape metadata");
1368 wait_event(mddev->sb_wait,
1369 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1370
1371 conf->reshape_safe = mddev->reshape_position;
1372 }
1373
1374 if (conf->pending_count >= max_queued_requests) {
1375 md_wakeup_thread(mddev->thread);
1376 raid10_log(mddev, "wait queued");
1377 wait_event(conf->wait_barrier,
1378 conf->pending_count < max_queued_requests);
1379 }
1380 /* first select target devices under rcu_lock and
1381 * inc refcount on their rdev. Record them by setting
1382 * bios[x] to bio
1383 * If there are known/acknowledged bad blocks on any device
1384 * on which we have seen a write error, we want to avoid
1385 * writing to those blocks. This potentially requires several
1386 * writes to write around the bad blocks. Each set of writes
1387 * gets its own r10_bio with a set of bios attached.
1388 */
1389
1390 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1391 raid10_find_phys(conf, r10_bio);
1392retry_write:
1393 blocked_rdev = NULL;
1394 rcu_read_lock();
1395 max_sectors = r10_bio->sectors;
1396
1397 for (i = 0; i < conf->copies; i++) {
1398 int d = r10_bio->devs[i].devnum;
1399 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1400 struct md_rdev *rrdev = rcu_dereference(
1401 conf->mirrors[d].replacement);
1402 if (rdev == rrdev)
1403 rrdev = NULL;
1404 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1405 atomic_inc(&rdev->nr_pending);
1406 blocked_rdev = rdev;
1407 break;
1408 }
1409 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1410 atomic_inc(&rrdev->nr_pending);
1411 blocked_rdev = rrdev;
1412 break;
1413 }
1414 if (rdev && (test_bit(Faulty, &rdev->flags)))
1415 rdev = NULL;
1416 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1417 rrdev = NULL;
1418
1419 r10_bio->devs[i].bio = NULL;
1420 r10_bio->devs[i].repl_bio = NULL;
1421
1422 if (!rdev && !rrdev) {
1423 set_bit(R10BIO_Degraded, &r10_bio->state);
1424 continue;
1425 }
1426 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1427 sector_t first_bad;
1428 sector_t dev_sector = r10_bio->devs[i].addr;
1429 int bad_sectors;
1430 int is_bad;
1431
1432 is_bad = is_badblock(rdev, dev_sector, max_sectors,
1433 &first_bad, &bad_sectors);
1434 if (is_bad < 0) {
1435 /* Mustn't write here until the bad block
1436 * is acknowledged
1437 */
1438 atomic_inc(&rdev->nr_pending);
1439 set_bit(BlockedBadBlocks, &rdev->flags);
1440 blocked_rdev = rdev;
1441 break;
1442 }
1443 if (is_bad && first_bad <= dev_sector) {
1444 /* Cannot write here at all */
1445 bad_sectors -= (dev_sector - first_bad);
1446 if (bad_sectors < max_sectors)
1447 /* Mustn't write more than bad_sectors
1448 * to other devices yet
1449 */
1450 max_sectors = bad_sectors;
1451 /* We don't set R10BIO_Degraded as that
1452 * only applies if the disk is missing,
1453 * so it might be re-added, and we want to
1454 * know to recover this chunk.
1455 * In this case the device is here, and the
1456 * fact that this chunk is not in-sync is
1457 * recorded in the bad block log.
1458 */
1459 continue;
1460 }
1461 if (is_bad) {
1462 int good_sectors = first_bad - dev_sector;
1463 if (good_sectors < max_sectors)
1464 max_sectors = good_sectors;
1465 }
1466 }
1467 if (rdev) {
1468 r10_bio->devs[i].bio = bio;
1469 atomic_inc(&rdev->nr_pending);
1470 }
1471 if (rrdev) {
1472 r10_bio->devs[i].repl_bio = bio;
1473 atomic_inc(&rrdev->nr_pending);
1474 }
1475 }
1476 rcu_read_unlock();
1477
1478 if (unlikely(blocked_rdev)) {
1479 /* Have to wait for this device to get unblocked, then retry */
1480 int j;
1481 int d;
1482
1483 for (j = 0; j < i; j++) {
1484 if (r10_bio->devs[j].bio) {
1485 d = r10_bio->devs[j].devnum;
1486 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1487 }
1488 if (r10_bio->devs[j].repl_bio) {
1489 struct md_rdev *rdev;
1490 d = r10_bio->devs[j].devnum;
1491 rdev = conf->mirrors[d].replacement;
1492 if (!rdev) {
1493 /* Race with remove_disk */
1494 smp_mb();
1495 rdev = conf->mirrors[d].rdev;
1496 }
1497 rdev_dec_pending(rdev, mddev);
1498 }
1499 }
1500 allow_barrier(conf);
1501 raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1502 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1503 wait_barrier(conf);
1504 goto retry_write;
1505 }
1506
1507 if (max_sectors < r10_bio->sectors)
1508 r10_bio->sectors = max_sectors;
1509
1510 if (r10_bio->sectors < bio_sectors(bio)) {
1511 struct bio *split = bio_split(bio, r10_bio->sectors,
1512 GFP_NOIO, conf->bio_split);
1513 bio_chain(split, bio);
1514 generic_make_request(bio);
1515 bio = split;
1516 r10_bio->master_bio = bio;
1517 }
1518
1519 atomic_set(&r10_bio->remaining, 1);
1520 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1521
1522 for (i = 0; i < conf->copies; i++) {
1523 if (r10_bio->devs[i].bio)
1524 raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1525 if (r10_bio->devs[i].repl_bio)
1526 raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1527 }
1528 one_write_done(r10_bio);
1529}
1530
1531static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1532{
1533 struct r10conf *conf = mddev->private;
1534 struct r10bio *r10_bio;
1535
1536 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1537
1538 r10_bio->master_bio = bio;
1539 r10_bio->sectors = sectors;
1540
1541 r10_bio->mddev = mddev;
1542 r10_bio->sector = bio->bi_iter.bi_sector;
1543 r10_bio->state = 0;
1544 memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * conf->copies);
1545
1546 if (bio_data_dir(bio) == READ)
1547 raid10_read_request(mddev, bio, r10_bio);
1548 else
1549 raid10_write_request(mddev, bio, r10_bio);
1550}
1551
1552static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1553{
1554 struct r10conf *conf = mddev->private;
1555 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1556 int chunk_sects = chunk_mask + 1;
1557 int sectors = bio_sectors(bio);
1558
1559 if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1560 md_flush_request(mddev, bio);
1561 return true;
1562 }
1563
1564 if (!md_write_start(mddev, bio))
1565 return false;
1566
1567 /*
1568 * If this request crosses a chunk boundary, we need to split
1569 * it.
1570 */
1571 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1572 sectors > chunk_sects
1573 && (conf->geo.near_copies < conf->geo.raid_disks
1574 || conf->prev.near_copies <
1575 conf->prev.raid_disks)))
1576 sectors = chunk_sects -
1577 (bio->bi_iter.bi_sector &
1578 (chunk_sects - 1));
1579 __make_request(mddev, bio, sectors);
1580
1581 /* In case raid10d snuck in to freeze_array */
1582 wake_up(&conf->wait_barrier);
1583 return true;
1584}
1585
1586static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1587{
1588 struct r10conf *conf = mddev->private;
1589 int i;
1590
1591 if (conf->geo.near_copies < conf->geo.raid_disks)
1592 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1593 if (conf->geo.near_copies > 1)
1594 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1595 if (conf->geo.far_copies > 1) {
1596 if (conf->geo.far_offset)
1597 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1598 else
1599 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1600 if (conf->geo.far_set_size != conf->geo.raid_disks)
1601 seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1602 }
1603 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1604 conf->geo.raid_disks - mddev->degraded);
1605 rcu_read_lock();
1606 for (i = 0; i < conf->geo.raid_disks; i++) {
1607 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1608 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1609 }
1610 rcu_read_unlock();
1611 seq_printf(seq, "]");
1612}
1613
1614/* check if there are enough drives for
1615 * every block to appear on atleast one.
1616 * Don't consider the device numbered 'ignore'
1617 * as we might be about to remove it.
1618 */
1619static int _enough(struct r10conf *conf, int previous, int ignore)
1620{
1621 int first = 0;
1622 int has_enough = 0;
1623 int disks, ncopies;
1624 if (previous) {
1625 disks = conf->prev.raid_disks;
1626 ncopies = conf->prev.near_copies;
1627 } else {
1628 disks = conf->geo.raid_disks;
1629 ncopies = conf->geo.near_copies;
1630 }
1631
1632 rcu_read_lock();
1633 do {
1634 int n = conf->copies;
1635 int cnt = 0;
1636 int this = first;
1637 while (n--) {
1638 struct md_rdev *rdev;
1639 if (this != ignore &&
1640 (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1641 test_bit(In_sync, &rdev->flags))
1642 cnt++;
1643 this = (this+1) % disks;
1644 }
1645 if (cnt == 0)
1646 goto out;
1647 first = (first + ncopies) % disks;
1648 } while (first != 0);
1649 has_enough = 1;
1650out:
1651 rcu_read_unlock();
1652 return has_enough;
1653}
1654
1655static int enough(struct r10conf *conf, int ignore)
1656{
1657 /* when calling 'enough', both 'prev' and 'geo' must
1658 * be stable.
1659 * This is ensured if ->reconfig_mutex or ->device_lock
1660 * is held.
1661 */
1662 return _enough(conf, 0, ignore) &&
1663 _enough(conf, 1, ignore);
1664}
1665
1666static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1667{
1668 char b[BDEVNAME_SIZE];
1669 struct r10conf *conf = mddev->private;
1670 unsigned long flags;
1671
1672 /*
1673 * If it is not operational, then we have already marked it as dead
1674 * else if it is the last working disks, ignore the error, let the
1675 * next level up know.
1676 * else mark the drive as failed
1677 */
1678 spin_lock_irqsave(&conf->device_lock, flags);
1679 if (test_bit(In_sync, &rdev->flags)
1680 && !enough(conf, rdev->raid_disk)) {
1681 /*
1682 * Don't fail the drive, just return an IO error.
1683 */
1684 spin_unlock_irqrestore(&conf->device_lock, flags);
1685 return;
1686 }
1687 if (test_and_clear_bit(In_sync, &rdev->flags))
1688 mddev->degraded++;
1689 /*
1690 * If recovery is running, make sure it aborts.
1691 */
1692 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1693 set_bit(Blocked, &rdev->flags);
1694 set_bit(Faulty, &rdev->flags);
1695 set_mask_bits(&mddev->sb_flags, 0,
1696 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1697 spin_unlock_irqrestore(&conf->device_lock, flags);
1698 pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1699 "md/raid10:%s: Operation continuing on %d devices.\n",
1700 mdname(mddev), bdevname(rdev->bdev, b),
1701 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1702}
1703
1704static void print_conf(struct r10conf *conf)
1705{
1706 int i;
1707 struct md_rdev *rdev;
1708
1709 pr_debug("RAID10 conf printout:\n");
1710 if (!conf) {
1711 pr_debug("(!conf)\n");
1712 return;
1713 }
1714 pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1715 conf->geo.raid_disks);
1716
1717 /* This is only called with ->reconfix_mutex held, so
1718 * rcu protection of rdev is not needed */
1719 for (i = 0; i < conf->geo.raid_disks; i++) {
1720 char b[BDEVNAME_SIZE];
1721 rdev = conf->mirrors[i].rdev;
1722 if (rdev)
1723 pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1724 i, !test_bit(In_sync, &rdev->flags),
1725 !test_bit(Faulty, &rdev->flags),
1726 bdevname(rdev->bdev,b));
1727 }
1728}
1729
1730static void close_sync(struct r10conf *conf)
1731{
1732 wait_barrier(conf);
1733 allow_barrier(conf);
1734
1735 mempool_destroy(conf->r10buf_pool);
1736 conf->r10buf_pool = NULL;
1737}
1738
1739static int raid10_spare_active(struct mddev *mddev)
1740{
1741 int i;
1742 struct r10conf *conf = mddev->private;
1743 struct raid10_info *tmp;
1744 int count = 0;
1745 unsigned long flags;
1746
1747 /*
1748 * Find all non-in_sync disks within the RAID10 configuration
1749 * and mark them in_sync
1750 */
1751 for (i = 0; i < conf->geo.raid_disks; i++) {
1752 tmp = conf->mirrors + i;
1753 if (tmp->replacement
1754 && tmp->replacement->recovery_offset == MaxSector
1755 && !test_bit(Faulty, &tmp->replacement->flags)
1756 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1757 /* Replacement has just become active */
1758 if (!tmp->rdev
1759 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1760 count++;
1761 if (tmp->rdev) {
1762 /* Replaced device not technically faulty,
1763 * but we need to be sure it gets removed
1764 * and never re-added.
1765 */
1766 set_bit(Faulty, &tmp->rdev->flags);
1767 sysfs_notify_dirent_safe(
1768 tmp->rdev->sysfs_state);
1769 }
1770 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1771 } else if (tmp->rdev
1772 && tmp->rdev->recovery_offset == MaxSector
1773 && !test_bit(Faulty, &tmp->rdev->flags)
1774 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1775 count++;
1776 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1777 }
1778 }
1779 spin_lock_irqsave(&conf->device_lock, flags);
1780 mddev->degraded -= count;
1781 spin_unlock_irqrestore(&conf->device_lock, flags);
1782
1783 print_conf(conf);
1784 return count;
1785}
1786
1787static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1788{
1789 struct r10conf *conf = mddev->private;
1790 int err = -EEXIST;
1791 int mirror;
1792 int first = 0;
1793 int last = conf->geo.raid_disks - 1;
1794
1795 if (mddev->recovery_cp < MaxSector)
1796 /* only hot-add to in-sync arrays, as recovery is
1797 * very different from resync
1798 */
1799 return -EBUSY;
1800 if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1801 return -EINVAL;
1802
1803 if (md_integrity_add_rdev(rdev, mddev))
1804 return -ENXIO;
1805
1806 if (rdev->raid_disk >= 0)
1807 first = last = rdev->raid_disk;
1808
1809 if (rdev->saved_raid_disk >= first &&
1810 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1811 mirror = rdev->saved_raid_disk;
1812 else
1813 mirror = first;
1814 for ( ; mirror <= last ; mirror++) {
1815 struct raid10_info *p = &conf->mirrors[mirror];
1816 if (p->recovery_disabled == mddev->recovery_disabled)
1817 continue;
1818 if (p->rdev) {
1819 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1820 p->replacement != NULL)
1821 continue;
1822 clear_bit(In_sync, &rdev->flags);
1823 set_bit(Replacement, &rdev->flags);
1824 rdev->raid_disk = mirror;
1825 err = 0;
1826 if (mddev->gendisk)
1827 disk_stack_limits(mddev->gendisk, rdev->bdev,
1828 rdev->data_offset << 9);
1829 conf->fullsync = 1;
1830 rcu_assign_pointer(p->replacement, rdev);
1831 break;
1832 }
1833
1834 if (mddev->gendisk)
1835 disk_stack_limits(mddev->gendisk, rdev->bdev,
1836 rdev->data_offset << 9);
1837
1838 p->head_position = 0;
1839 p->recovery_disabled = mddev->recovery_disabled - 1;
1840 rdev->raid_disk = mirror;
1841 err = 0;
1842 if (rdev->saved_raid_disk != mirror)
1843 conf->fullsync = 1;
1844 rcu_assign_pointer(p->rdev, rdev);
1845 break;
1846 }
1847 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1848 blk_queue_flag_set(QUEUE_FLAG_DISCARD, mddev->queue);
1849
1850 print_conf(conf);
1851 return err;
1852}
1853
1854static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1855{
1856 struct r10conf *conf = mddev->private;
1857 int err = 0;
1858 int number = rdev->raid_disk;
1859 struct md_rdev **rdevp;
1860 struct raid10_info *p = conf->mirrors + number;
1861
1862 print_conf(conf);
1863 if (rdev == p->rdev)
1864 rdevp = &p->rdev;
1865 else if (rdev == p->replacement)
1866 rdevp = &p->replacement;
1867 else
1868 return 0;
1869
1870 if (test_bit(In_sync, &rdev->flags) ||
1871 atomic_read(&rdev->nr_pending)) {
1872 err = -EBUSY;
1873 goto abort;
1874 }
1875 /* Only remove non-faulty devices if recovery
1876 * is not possible.
1877 */
1878 if (!test_bit(Faulty, &rdev->flags) &&
1879 mddev->recovery_disabled != p->recovery_disabled &&
1880 (!p->replacement || p->replacement == rdev) &&
1881 number < conf->geo.raid_disks &&
1882 enough(conf, -1)) {
1883 err = -EBUSY;
1884 goto abort;
1885 }
1886 *rdevp = NULL;
1887 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1888 synchronize_rcu();
1889 if (atomic_read(&rdev->nr_pending)) {
1890 /* lost the race, try later */
1891 err = -EBUSY;
1892 *rdevp = rdev;
1893 goto abort;
1894 }
1895 }
1896 if (p->replacement) {
1897 /* We must have just cleared 'rdev' */
1898 p->rdev = p->replacement;
1899 clear_bit(Replacement, &p->replacement->flags);
1900 smp_mb(); /* Make sure other CPUs may see both as identical
1901 * but will never see neither -- if they are careful.
1902 */
1903 p->replacement = NULL;
1904 }
1905
1906 clear_bit(WantReplacement, &rdev->flags);
1907 err = md_integrity_register(mddev);
1908
1909abort:
1910
1911 print_conf(conf);
1912 return err;
1913}
1914
1915static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
1916{
1917 struct r10conf *conf = r10_bio->mddev->private;
1918
1919 if (!bio->bi_status)
1920 set_bit(R10BIO_Uptodate, &r10_bio->state);
1921 else
1922 /* The write handler will notice the lack of
1923 * R10BIO_Uptodate and record any errors etc
1924 */
1925 atomic_add(r10_bio->sectors,
1926 &conf->mirrors[d].rdev->corrected_errors);
1927
1928 /* for reconstruct, we always reschedule after a read.
1929 * for resync, only after all reads
1930 */
1931 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1932 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1933 atomic_dec_and_test(&r10_bio->remaining)) {
1934 /* we have read all the blocks,
1935 * do the comparison in process context in raid10d
1936 */
1937 reschedule_retry(r10_bio);
1938 }
1939}
1940
1941static void end_sync_read(struct bio *bio)
1942{
1943 struct r10bio *r10_bio = get_resync_r10bio(bio);
1944 struct r10conf *conf = r10_bio->mddev->private;
1945 int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1946
1947 __end_sync_read(r10_bio, bio, d);
1948}
1949
1950static void end_reshape_read(struct bio *bio)
1951{
1952 /* reshape read bio isn't allocated from r10buf_pool */
1953 struct r10bio *r10_bio = bio->bi_private;
1954
1955 __end_sync_read(r10_bio, bio, r10_bio->read_slot);
1956}
1957
1958static void end_sync_request(struct r10bio *r10_bio)
1959{
1960 struct mddev *mddev = r10_bio->mddev;
1961
1962 while (atomic_dec_and_test(&r10_bio->remaining)) {
1963 if (r10_bio->master_bio == NULL) {
1964 /* the primary of several recovery bios */
1965 sector_t s = r10_bio->sectors;
1966 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1967 test_bit(R10BIO_WriteError, &r10_bio->state))
1968 reschedule_retry(r10_bio);
1969 else
1970 put_buf(r10_bio);
1971 md_done_sync(mddev, s, 1);
1972 break;
1973 } else {
1974 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1975 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1976 test_bit(R10BIO_WriteError, &r10_bio->state))
1977 reschedule_retry(r10_bio);
1978 else
1979 put_buf(r10_bio);
1980 r10_bio = r10_bio2;
1981 }
1982 }
1983}
1984
1985static void end_sync_write(struct bio *bio)
1986{
1987 struct r10bio *r10_bio = get_resync_r10bio(bio);
1988 struct mddev *mddev = r10_bio->mddev;
1989 struct r10conf *conf = mddev->private;
1990 int d;
1991 sector_t first_bad;
1992 int bad_sectors;
1993 int slot;
1994 int repl;
1995 struct md_rdev *rdev = NULL;
1996
1997 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1998 if (repl)
1999 rdev = conf->mirrors[d].replacement;
2000 else
2001 rdev = conf->mirrors[d].rdev;
2002
2003 if (bio->bi_status) {
2004 if (repl)
2005 md_error(mddev, rdev);
2006 else {
2007 set_bit(WriteErrorSeen, &rdev->flags);
2008 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2009 set_bit(MD_RECOVERY_NEEDED,
2010 &rdev->mddev->recovery);
2011 set_bit(R10BIO_WriteError, &r10_bio->state);
2012 }
2013 } else if (is_badblock(rdev,
2014 r10_bio->devs[slot].addr,
2015 r10_bio->sectors,
2016 &first_bad, &bad_sectors))
2017 set_bit(R10BIO_MadeGood, &r10_bio->state);
2018
2019 rdev_dec_pending(rdev, mddev);
2020
2021 end_sync_request(r10_bio);
2022}
2023
2024/*
2025 * Note: sync and recover and handled very differently for raid10
2026 * This code is for resync.
2027 * For resync, we read through virtual addresses and read all blocks.
2028 * If there is any error, we schedule a write. The lowest numbered
2029 * drive is authoritative.
2030 * However requests come for physical address, so we need to map.
2031 * For every physical address there are raid_disks/copies virtual addresses,
2032 * which is always are least one, but is not necessarly an integer.
2033 * This means that a physical address can span multiple chunks, so we may
2034 * have to submit multiple io requests for a single sync request.
2035 */
2036/*
2037 * We check if all blocks are in-sync and only write to blocks that
2038 * aren't in sync
2039 */
2040static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2041{
2042 struct r10conf *conf = mddev->private;
2043 int i, first;
2044 struct bio *tbio, *fbio;
2045 int vcnt;
2046 struct page **tpages, **fpages;
2047
2048 atomic_set(&r10_bio->remaining, 1);
2049
2050 /* find the first device with a block */
2051 for (i=0; i<conf->copies; i++)
2052 if (!r10_bio->devs[i].bio->bi_status)
2053 break;
2054
2055 if (i == conf->copies)
2056 goto done;
2057
2058 first = i;
2059 fbio = r10_bio->devs[i].bio;
2060 fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2061 fbio->bi_iter.bi_idx = 0;
2062 fpages = get_resync_pages(fbio)->pages;
2063
2064 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2065 /* now find blocks with errors */
2066 for (i=0 ; i < conf->copies ; i++) {
2067 int j, d;
2068 struct md_rdev *rdev;
2069 struct resync_pages *rp;
2070
2071 tbio = r10_bio->devs[i].bio;
2072
2073 if (tbio->bi_end_io != end_sync_read)
2074 continue;
2075 if (i == first)
2076 continue;
2077
2078 tpages = get_resync_pages(tbio)->pages;
2079 d = r10_bio->devs[i].devnum;
2080 rdev = conf->mirrors[d].rdev;
2081 if (!r10_bio->devs[i].bio->bi_status) {
2082 /* We know that the bi_io_vec layout is the same for
2083 * both 'first' and 'i', so we just compare them.
2084 * All vec entries are PAGE_SIZE;
2085 */
2086 int sectors = r10_bio->sectors;
2087 for (j = 0; j < vcnt; j++) {
2088 int len = PAGE_SIZE;
2089 if (sectors < (len / 512))
2090 len = sectors * 512;
2091 if (memcmp(page_address(fpages[j]),
2092 page_address(tpages[j]),
2093 len))
2094 break;
2095 sectors -= len/512;
2096 }
2097 if (j == vcnt)
2098 continue;
2099 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2100 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2101 /* Don't fix anything. */
2102 continue;
2103 } else if (test_bit(FailFast, &rdev->flags)) {
2104 /* Just give up on this device */
2105 md_error(rdev->mddev, rdev);
2106 continue;
2107 }
2108 /* Ok, we need to write this bio, either to correct an
2109 * inconsistency or to correct an unreadable block.
2110 * First we need to fixup bv_offset, bv_len and
2111 * bi_vecs, as the read request might have corrupted these
2112 */
2113 rp = get_resync_pages(tbio);
2114 bio_reset(tbio);
2115
2116 md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2117
2118 rp->raid_bio = r10_bio;
2119 tbio->bi_private = rp;
2120 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2121 tbio->bi_end_io = end_sync_write;
2122 bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2123
2124 bio_copy_data(tbio, fbio);
2125
2126 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2127 atomic_inc(&r10_bio->remaining);
2128 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2129
2130 if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2131 tbio->bi_opf |= MD_FAILFAST;
2132 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2133 bio_set_dev(tbio, conf->mirrors[d].rdev->bdev);
2134 generic_make_request(tbio);
2135 }
2136
2137 /* Now write out to any replacement devices
2138 * that are active
2139 */
2140 for (i = 0; i < conf->copies; i++) {
2141 int d;
2142
2143 tbio = r10_bio->devs[i].repl_bio;
2144 if (!tbio || !tbio->bi_end_io)
2145 continue;
2146 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2147 && r10_bio->devs[i].bio != fbio)
2148 bio_copy_data(tbio, fbio);
2149 d = r10_bio->devs[i].devnum;
2150 atomic_inc(&r10_bio->remaining);
2151 md_sync_acct(conf->mirrors[d].replacement->bdev,
2152 bio_sectors(tbio));
2153 generic_make_request(tbio);
2154 }
2155
2156done:
2157 if (atomic_dec_and_test(&r10_bio->remaining)) {
2158 md_done_sync(mddev, r10_bio->sectors, 1);
2159 put_buf(r10_bio);
2160 }
2161}
2162
2163/*
2164 * Now for the recovery code.
2165 * Recovery happens across physical sectors.
2166 * We recover all non-is_sync drives by finding the virtual address of
2167 * each, and then choose a working drive that also has that virt address.
2168 * There is a separate r10_bio for each non-in_sync drive.
2169 * Only the first two slots are in use. The first for reading,
2170 * The second for writing.
2171 *
2172 */
2173static void fix_recovery_read_error(struct r10bio *r10_bio)
2174{
2175 /* We got a read error during recovery.
2176 * We repeat the read in smaller page-sized sections.
2177 * If a read succeeds, write it to the new device or record
2178 * a bad block if we cannot.
2179 * If a read fails, record a bad block on both old and
2180 * new devices.
2181 */
2182 struct mddev *mddev = r10_bio->mddev;
2183 struct r10conf *conf = mddev->private;
2184 struct bio *bio = r10_bio->devs[0].bio;
2185 sector_t sect = 0;
2186 int sectors = r10_bio->sectors;
2187 int idx = 0;
2188 int dr = r10_bio->devs[0].devnum;
2189 int dw = r10_bio->devs[1].devnum;
2190 struct page **pages = get_resync_pages(bio)->pages;
2191
2192 while (sectors) {
2193 int s = sectors;
2194 struct md_rdev *rdev;
2195 sector_t addr;
2196 int ok;
2197
2198 if (s > (PAGE_SIZE>>9))
2199 s = PAGE_SIZE >> 9;
2200
2201 rdev = conf->mirrors[dr].rdev;
2202 addr = r10_bio->devs[0].addr + sect,
2203 ok = sync_page_io(rdev,
2204 addr,
2205 s << 9,
2206 pages[idx],
2207 REQ_OP_READ, 0, false);
2208 if (ok) {
2209 rdev = conf->mirrors[dw].rdev;
2210 addr = r10_bio->devs[1].addr + sect;
2211 ok = sync_page_io(rdev,
2212 addr,
2213 s << 9,
2214 pages[idx],
2215 REQ_OP_WRITE, 0, false);
2216 if (!ok) {
2217 set_bit(WriteErrorSeen, &rdev->flags);
2218 if (!test_and_set_bit(WantReplacement,
2219 &rdev->flags))
2220 set_bit(MD_RECOVERY_NEEDED,
2221 &rdev->mddev->recovery);
2222 }
2223 }
2224 if (!ok) {
2225 /* We don't worry if we cannot set a bad block -
2226 * it really is bad so there is no loss in not
2227 * recording it yet
2228 */
2229 rdev_set_badblocks(rdev, addr, s, 0);
2230
2231 if (rdev != conf->mirrors[dw].rdev) {
2232 /* need bad block on destination too */
2233 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2234 addr = r10_bio->devs[1].addr + sect;
2235 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2236 if (!ok) {
2237 /* just abort the recovery */
2238 pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2239 mdname(mddev));
2240
2241 conf->mirrors[dw].recovery_disabled
2242 = mddev->recovery_disabled;
2243 set_bit(MD_RECOVERY_INTR,
2244 &mddev->recovery);
2245 break;
2246 }
2247 }
2248 }
2249
2250 sectors -= s;
2251 sect += s;
2252 idx++;
2253 }
2254}
2255
2256static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2257{
2258 struct r10conf *conf = mddev->private;
2259 int d;
2260 struct bio *wbio, *wbio2;
2261
2262 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2263 fix_recovery_read_error(r10_bio);
2264 end_sync_request(r10_bio);
2265 return;
2266 }
2267
2268 /*
2269 * share the pages with the first bio
2270 * and submit the write request
2271 */
2272 d = r10_bio->devs[1].devnum;
2273 wbio = r10_bio->devs[1].bio;
2274 wbio2 = r10_bio->devs[1].repl_bio;
2275 /* Need to test wbio2->bi_end_io before we call
2276 * generic_make_request as if the former is NULL,
2277 * the latter is free to free wbio2.
2278 */
2279 if (wbio2 && !wbio2->bi_end_io)
2280 wbio2 = NULL;
2281 if (wbio->bi_end_io) {
2282 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2283 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2284 generic_make_request(wbio);
2285 }
2286 if (wbio2) {
2287 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2288 md_sync_acct(conf->mirrors[d].replacement->bdev,
2289 bio_sectors(wbio2));
2290 generic_make_request(wbio2);
2291 }
2292}
2293
2294/*
2295 * Used by fix_read_error() to decay the per rdev read_errors.
2296 * We halve the read error count for every hour that has elapsed
2297 * since the last recorded read error.
2298 *
2299 */
2300static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2301{
2302 long cur_time_mon;
2303 unsigned long hours_since_last;
2304 unsigned int read_errors = atomic_read(&rdev->read_errors);
2305
2306 cur_time_mon = ktime_get_seconds();
2307
2308 if (rdev->last_read_error == 0) {
2309 /* first time we've seen a read error */
2310 rdev->last_read_error = cur_time_mon;
2311 return;
2312 }
2313
2314 hours_since_last = (long)(cur_time_mon -
2315 rdev->last_read_error) / 3600;
2316
2317 rdev->last_read_error = cur_time_mon;
2318
2319 /*
2320 * if hours_since_last is > the number of bits in read_errors
2321 * just set read errors to 0. We do this to avoid
2322 * overflowing the shift of read_errors by hours_since_last.
2323 */
2324 if (hours_since_last >= 8 * sizeof(read_errors))
2325 atomic_set(&rdev->read_errors, 0);
2326 else
2327 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2328}
2329
2330static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2331 int sectors, struct page *page, int rw)
2332{
2333 sector_t first_bad;
2334 int bad_sectors;
2335
2336 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2337 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2338 return -1;
2339 if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2340 /* success */
2341 return 1;
2342 if (rw == WRITE) {
2343 set_bit(WriteErrorSeen, &rdev->flags);
2344 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2345 set_bit(MD_RECOVERY_NEEDED,
2346 &rdev->mddev->recovery);
2347 }
2348 /* need to record an error - either for the block or the device */
2349 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2350 md_error(rdev->mddev, rdev);
2351 return 0;
2352}
2353
2354/*
2355 * This is a kernel thread which:
2356 *
2357 * 1. Retries failed read operations on working mirrors.
2358 * 2. Updates the raid superblock when problems encounter.
2359 * 3. Performs writes following reads for array synchronising.
2360 */
2361
2362static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2363{
2364 int sect = 0; /* Offset from r10_bio->sector */
2365 int sectors = r10_bio->sectors;
2366 struct md_rdev*rdev;
2367 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2368 int d = r10_bio->devs[r10_bio->read_slot].devnum;
2369
2370 /* still own a reference to this rdev, so it cannot
2371 * have been cleared recently.
2372 */
2373 rdev = conf->mirrors[d].rdev;
2374
2375 if (test_bit(Faulty, &rdev->flags))
2376 /* drive has already been failed, just ignore any
2377 more fix_read_error() attempts */
2378 return;
2379
2380 check_decay_read_errors(mddev, rdev);
2381 atomic_inc(&rdev->read_errors);
2382 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2383 char b[BDEVNAME_SIZE];
2384 bdevname(rdev->bdev, b);
2385
2386 pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2387 mdname(mddev), b,
2388 atomic_read(&rdev->read_errors), max_read_errors);
2389 pr_notice("md/raid10:%s: %s: Failing raid device\n",
2390 mdname(mddev), b);
2391 md_error(mddev, rdev);
2392 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2393 return;
2394 }
2395
2396 while(sectors) {
2397 int s = sectors;
2398 int sl = r10_bio->read_slot;
2399 int success = 0;
2400 int start;
2401
2402 if (s > (PAGE_SIZE>>9))
2403 s = PAGE_SIZE >> 9;
2404
2405 rcu_read_lock();
2406 do {
2407 sector_t first_bad;
2408 int bad_sectors;
2409
2410 d = r10_bio->devs[sl].devnum;
2411 rdev = rcu_dereference(conf->mirrors[d].rdev);
2412 if (rdev &&
2413 test_bit(In_sync, &rdev->flags) &&
2414 !test_bit(Faulty, &rdev->flags) &&
2415 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2416 &first_bad, &bad_sectors) == 0) {
2417 atomic_inc(&rdev->nr_pending);
2418 rcu_read_unlock();
2419 success = sync_page_io(rdev,
2420 r10_bio->devs[sl].addr +
2421 sect,
2422 s<<9,
2423 conf->tmppage,
2424 REQ_OP_READ, 0, false);
2425 rdev_dec_pending(rdev, mddev);
2426 rcu_read_lock();
2427 if (success)
2428 break;
2429 }
2430 sl++;
2431 if (sl == conf->copies)
2432 sl = 0;
2433 } while (!success && sl != r10_bio->read_slot);
2434 rcu_read_unlock();
2435
2436 if (!success) {
2437 /* Cannot read from anywhere, just mark the block
2438 * as bad on the first device to discourage future
2439 * reads.
2440 */
2441 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2442 rdev = conf->mirrors[dn].rdev;
2443
2444 if (!rdev_set_badblocks(
2445 rdev,
2446 r10_bio->devs[r10_bio->read_slot].addr
2447 + sect,
2448 s, 0)) {
2449 md_error(mddev, rdev);
2450 r10_bio->devs[r10_bio->read_slot].bio
2451 = IO_BLOCKED;
2452 }
2453 break;
2454 }
2455
2456 start = sl;
2457 /* write it back and re-read */
2458 rcu_read_lock();
2459 while (sl != r10_bio->read_slot) {
2460 char b[BDEVNAME_SIZE];
2461
2462 if (sl==0)
2463 sl = conf->copies;
2464 sl--;
2465 d = r10_bio->devs[sl].devnum;
2466 rdev = rcu_dereference(conf->mirrors[d].rdev);
2467 if (!rdev ||
2468 test_bit(Faulty, &rdev->flags) ||
2469 !test_bit(In_sync, &rdev->flags))
2470 continue;
2471
2472 atomic_inc(&rdev->nr_pending);
2473 rcu_read_unlock();
2474 if (r10_sync_page_io(rdev,
2475 r10_bio->devs[sl].addr +
2476 sect,
2477 s, conf->tmppage, WRITE)
2478 == 0) {
2479 /* Well, this device is dead */
2480 pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2481 mdname(mddev), s,
2482 (unsigned long long)(
2483 sect +
2484 choose_data_offset(r10_bio,
2485 rdev)),
2486 bdevname(rdev->bdev, b));
2487 pr_notice("md/raid10:%s: %s: failing drive\n",
2488 mdname(mddev),
2489 bdevname(rdev->bdev, b));
2490 }
2491 rdev_dec_pending(rdev, mddev);
2492 rcu_read_lock();
2493 }
2494 sl = start;
2495 while (sl != r10_bio->read_slot) {
2496 char b[BDEVNAME_SIZE];
2497
2498 if (sl==0)
2499 sl = conf->copies;
2500 sl--;
2501 d = r10_bio->devs[sl].devnum;
2502 rdev = rcu_dereference(conf->mirrors[d].rdev);
2503 if (!rdev ||
2504 test_bit(Faulty, &rdev->flags) ||
2505 !test_bit(In_sync, &rdev->flags))
2506 continue;
2507
2508 atomic_inc(&rdev->nr_pending);
2509 rcu_read_unlock();
2510 switch (r10_sync_page_io(rdev,
2511 r10_bio->devs[sl].addr +
2512 sect,
2513 s, conf->tmppage,
2514 READ)) {
2515 case 0:
2516 /* Well, this device is dead */
2517 pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2518 mdname(mddev), s,
2519 (unsigned long long)(
2520 sect +
2521 choose_data_offset(r10_bio, rdev)),
2522 bdevname(rdev->bdev, b));
2523 pr_notice("md/raid10:%s: %s: failing drive\n",
2524 mdname(mddev),
2525 bdevname(rdev->bdev, b));
2526 break;
2527 case 1:
2528 pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2529 mdname(mddev), s,
2530 (unsigned long long)(
2531 sect +
2532 choose_data_offset(r10_bio, rdev)),
2533 bdevname(rdev->bdev, b));
2534 atomic_add(s, &rdev->corrected_errors);
2535 }
2536
2537 rdev_dec_pending(rdev, mddev);
2538 rcu_read_lock();
2539 }
2540 rcu_read_unlock();
2541
2542 sectors -= s;
2543 sect += s;
2544 }
2545}
2546
2547static int narrow_write_error(struct r10bio *r10_bio, int i)
2548{
2549 struct bio *bio = r10_bio->master_bio;
2550 struct mddev *mddev = r10_bio->mddev;
2551 struct r10conf *conf = mddev->private;
2552 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2553 /* bio has the data to be written to slot 'i' where
2554 * we just recently had a write error.
2555 * We repeatedly clone the bio and trim down to one block,
2556 * then try the write. Where the write fails we record
2557 * a bad block.
2558 * It is conceivable that the bio doesn't exactly align with
2559 * blocks. We must handle this.
2560 *
2561 * We currently own a reference to the rdev.
2562 */
2563
2564 int block_sectors;
2565 sector_t sector;
2566 int sectors;
2567 int sect_to_write = r10_bio->sectors;
2568 int ok = 1;
2569
2570 if (rdev->badblocks.shift < 0)
2571 return 0;
2572
2573 block_sectors = roundup(1 << rdev->badblocks.shift,
2574 bdev_logical_block_size(rdev->bdev) >> 9);
2575 sector = r10_bio->sector;
2576 sectors = ((r10_bio->sector + block_sectors)
2577 & ~(sector_t)(block_sectors - 1))
2578 - sector;
2579
2580 while (sect_to_write) {
2581 struct bio *wbio;
2582 sector_t wsector;
2583 if (sectors > sect_to_write)
2584 sectors = sect_to_write;
2585 /* Write at 'sector' for 'sectors' */
2586 wbio = bio_clone_fast(bio, GFP_NOIO, mddev->bio_set);
2587 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2588 wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2589 wbio->bi_iter.bi_sector = wsector +
2590 choose_data_offset(r10_bio, rdev);
2591 bio_set_dev(wbio, rdev->bdev);
2592 bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2593
2594 if (submit_bio_wait(wbio) < 0)
2595 /* Failure! */
2596 ok = rdev_set_badblocks(rdev, wsector,
2597 sectors, 0)
2598 && ok;
2599
2600 bio_put(wbio);
2601 sect_to_write -= sectors;
2602 sector += sectors;
2603 sectors = block_sectors;
2604 }
2605 return ok;
2606}
2607
2608static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2609{
2610 int slot = r10_bio->read_slot;
2611 struct bio *bio;
2612 struct r10conf *conf = mddev->private;
2613 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2614
2615 /* we got a read error. Maybe the drive is bad. Maybe just
2616 * the block and we can fix it.
2617 * We freeze all other IO, and try reading the block from
2618 * other devices. When we find one, we re-write
2619 * and check it that fixes the read error.
2620 * This is all done synchronously while the array is
2621 * frozen.
2622 */
2623 bio = r10_bio->devs[slot].bio;
2624 bio_put(bio);
2625 r10_bio->devs[slot].bio = NULL;
2626
2627 if (mddev->ro)
2628 r10_bio->devs[slot].bio = IO_BLOCKED;
2629 else if (!test_bit(FailFast, &rdev->flags)) {
2630 freeze_array(conf, 1);
2631 fix_read_error(conf, mddev, r10_bio);
2632 unfreeze_array(conf);
2633 } else
2634 md_error(mddev, rdev);
2635
2636 rdev_dec_pending(rdev, mddev);
2637 allow_barrier(conf);
2638 r10_bio->state = 0;
2639 raid10_read_request(mddev, r10_bio->master_bio, r10_bio);
2640}
2641
2642static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2643{
2644 /* Some sort of write request has finished and it
2645 * succeeded in writing where we thought there was a
2646 * bad block. So forget the bad block.
2647 * Or possibly if failed and we need to record
2648 * a bad block.
2649 */
2650 int m;
2651 struct md_rdev *rdev;
2652
2653 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2654 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2655 for (m = 0; m < conf->copies; m++) {
2656 int dev = r10_bio->devs[m].devnum;
2657 rdev = conf->mirrors[dev].rdev;
2658 if (r10_bio->devs[m].bio == NULL ||
2659 r10_bio->devs[m].bio->bi_end_io == NULL)
2660 continue;
2661 if (!r10_bio->devs[m].bio->bi_status) {
2662 rdev_clear_badblocks(
2663 rdev,
2664 r10_bio->devs[m].addr,
2665 r10_bio->sectors, 0);
2666 } else {
2667 if (!rdev_set_badblocks(
2668 rdev,
2669 r10_bio->devs[m].addr,
2670 r10_bio->sectors, 0))
2671 md_error(conf->mddev, rdev);
2672 }
2673 rdev = conf->mirrors[dev].replacement;
2674 if (r10_bio->devs[m].repl_bio == NULL ||
2675 r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2676 continue;
2677
2678 if (!r10_bio->devs[m].repl_bio->bi_status) {
2679 rdev_clear_badblocks(
2680 rdev,
2681 r10_bio->devs[m].addr,
2682 r10_bio->sectors, 0);
2683 } else {
2684 if (!rdev_set_badblocks(
2685 rdev,
2686 r10_bio->devs[m].addr,
2687 r10_bio->sectors, 0))
2688 md_error(conf->mddev, rdev);
2689 }
2690 }
2691 put_buf(r10_bio);
2692 } else {
2693 bool fail = false;
2694 for (m = 0; m < conf->copies; m++) {
2695 int dev = r10_bio->devs[m].devnum;
2696 struct bio *bio = r10_bio->devs[m].bio;
2697 rdev = conf->mirrors[dev].rdev;
2698 if (bio == IO_MADE_GOOD) {
2699 rdev_clear_badblocks(
2700 rdev,
2701 r10_bio->devs[m].addr,
2702 r10_bio->sectors, 0);
2703 rdev_dec_pending(rdev, conf->mddev);
2704 } else if (bio != NULL && bio->bi_status) {
2705 fail = true;
2706 if (!narrow_write_error(r10_bio, m)) {
2707 md_error(conf->mddev, rdev);
2708 set_bit(R10BIO_Degraded,
2709 &r10_bio->state);
2710 }
2711 rdev_dec_pending(rdev, conf->mddev);
2712 }
2713 bio = r10_bio->devs[m].repl_bio;
2714 rdev = conf->mirrors[dev].replacement;
2715 if (rdev && bio == IO_MADE_GOOD) {
2716 rdev_clear_badblocks(
2717 rdev,
2718 r10_bio->devs[m].addr,
2719 r10_bio->sectors, 0);
2720 rdev_dec_pending(rdev, conf->mddev);
2721 }
2722 }
2723 if (fail) {
2724 spin_lock_irq(&conf->device_lock);
2725 list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2726 conf->nr_queued++;
2727 spin_unlock_irq(&conf->device_lock);
2728 /*
2729 * In case freeze_array() is waiting for condition
2730 * nr_pending == nr_queued + extra to be true.
2731 */
2732 wake_up(&conf->wait_barrier);
2733 md_wakeup_thread(conf->mddev->thread);
2734 } else {
2735 if (test_bit(R10BIO_WriteError,
2736 &r10_bio->state))
2737 close_write(r10_bio);
2738 raid_end_bio_io(r10_bio);
2739 }
2740 }
2741}
2742
2743static void raid10d(struct md_thread *thread)
2744{
2745 struct mddev *mddev = thread->mddev;
2746 struct r10bio *r10_bio;
2747 unsigned long flags;
2748 struct r10conf *conf = mddev->private;
2749 struct list_head *head = &conf->retry_list;
2750 struct blk_plug plug;
2751
2752 md_check_recovery(mddev);
2753
2754 if (!list_empty_careful(&conf->bio_end_io_list) &&
2755 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2756 LIST_HEAD(tmp);
2757 spin_lock_irqsave(&conf->device_lock, flags);
2758 if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2759 while (!list_empty(&conf->bio_end_io_list)) {
2760 list_move(conf->bio_end_io_list.prev, &tmp);
2761 conf->nr_queued--;
2762 }
2763 }
2764 spin_unlock_irqrestore(&conf->device_lock, flags);
2765 while (!list_empty(&tmp)) {
2766 r10_bio = list_first_entry(&tmp, struct r10bio,
2767 retry_list);
2768 list_del(&r10_bio->retry_list);
2769 if (mddev->degraded)
2770 set_bit(R10BIO_Degraded, &r10_bio->state);
2771
2772 if (test_bit(R10BIO_WriteError,
2773 &r10_bio->state))
2774 close_write(r10_bio);
2775 raid_end_bio_io(r10_bio);
2776 }
2777 }
2778
2779 blk_start_plug(&plug);
2780 for (;;) {
2781
2782 flush_pending_writes(conf);
2783
2784 spin_lock_irqsave(&conf->device_lock, flags);
2785 if (list_empty(head)) {
2786 spin_unlock_irqrestore(&conf->device_lock, flags);
2787 break;
2788 }
2789 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2790 list_del(head->prev);
2791 conf->nr_queued--;
2792 spin_unlock_irqrestore(&conf->device_lock, flags);
2793
2794 mddev = r10_bio->mddev;
2795 conf = mddev->private;
2796 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2797 test_bit(R10BIO_WriteError, &r10_bio->state))
2798 handle_write_completed(conf, r10_bio);
2799 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2800 reshape_request_write(mddev, r10_bio);
2801 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2802 sync_request_write(mddev, r10_bio);
2803 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2804 recovery_request_write(mddev, r10_bio);
2805 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2806 handle_read_error(mddev, r10_bio);
2807 else
2808 WARN_ON_ONCE(1);
2809
2810 cond_resched();
2811 if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2812 md_check_recovery(mddev);
2813 }
2814 blk_finish_plug(&plug);
2815}
2816
2817static int init_resync(struct r10conf *conf)
2818{
2819 int buffs;
2820 int i;
2821
2822 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2823 BUG_ON(conf->r10buf_pool);
2824 conf->have_replacement = 0;
2825 for (i = 0; i < conf->geo.raid_disks; i++)
2826 if (conf->mirrors[i].replacement)
2827 conf->have_replacement = 1;
2828 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2829 if (!conf->r10buf_pool)
2830 return -ENOMEM;
2831 conf->next_resync = 0;
2832 return 0;
2833}
2834
2835static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
2836{
2837 struct r10bio *r10bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2838 struct rsync_pages *rp;
2839 struct bio *bio;
2840 int nalloc;
2841 int i;
2842
2843 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
2844 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
2845 nalloc = conf->copies; /* resync */
2846 else
2847 nalloc = 2; /* recovery */
2848
2849 for (i = 0; i < nalloc; i++) {
2850 bio = r10bio->devs[i].bio;
2851 rp = bio->bi_private;
2852 bio_reset(bio);
2853 bio->bi_private = rp;
2854 bio = r10bio->devs[i].repl_bio;
2855 if (bio) {
2856 rp = bio->bi_private;
2857 bio_reset(bio);
2858 bio->bi_private = rp;
2859 }
2860 }
2861 return r10bio;
2862}
2863
2864/*
2865 * Set cluster_sync_high since we need other nodes to add the
2866 * range [cluster_sync_low, cluster_sync_high] to suspend list.
2867 */
2868static void raid10_set_cluster_sync_high(struct r10conf *conf)
2869{
2870 sector_t window_size;
2871 int extra_chunk, chunks;
2872
2873 /*
2874 * First, here we define "stripe" as a unit which across
2875 * all member devices one time, so we get chunks by use
2876 * raid_disks / near_copies. Otherwise, if near_copies is
2877 * close to raid_disks, then resync window could increases
2878 * linearly with the increase of raid_disks, which means
2879 * we will suspend a really large IO window while it is not
2880 * necessary. If raid_disks is not divisible by near_copies,
2881 * an extra chunk is needed to ensure the whole "stripe" is
2882 * covered.
2883 */
2884
2885 chunks = conf->geo.raid_disks / conf->geo.near_copies;
2886 if (conf->geo.raid_disks % conf->geo.near_copies == 0)
2887 extra_chunk = 0;
2888 else
2889 extra_chunk = 1;
2890 window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
2891
2892 /*
2893 * At least use a 32M window to align with raid1's resync window
2894 */
2895 window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
2896 CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
2897
2898 conf->cluster_sync_high = conf->cluster_sync_low + window_size;
2899}
2900
2901/*
2902 * perform a "sync" on one "block"
2903 *
2904 * We need to make sure that no normal I/O request - particularly write
2905 * requests - conflict with active sync requests.
2906 *
2907 * This is achieved by tracking pending requests and a 'barrier' concept
2908 * that can be installed to exclude normal IO requests.
2909 *
2910 * Resync and recovery are handled very differently.
2911 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2912 *
2913 * For resync, we iterate over virtual addresses, read all copies,
2914 * and update if there are differences. If only one copy is live,
2915 * skip it.
2916 * For recovery, we iterate over physical addresses, read a good
2917 * value for each non-in_sync drive, and over-write.
2918 *
2919 * So, for recovery we may have several outstanding complex requests for a
2920 * given address, one for each out-of-sync device. We model this by allocating
2921 * a number of r10_bio structures, one for each out-of-sync device.
2922 * As we setup these structures, we collect all bio's together into a list
2923 * which we then process collectively to add pages, and then process again
2924 * to pass to generic_make_request.
2925 *
2926 * The r10_bio structures are linked using a borrowed master_bio pointer.
2927 * This link is counted in ->remaining. When the r10_bio that points to NULL
2928 * has its remaining count decremented to 0, the whole complex operation
2929 * is complete.
2930 *
2931 */
2932
2933static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2934 int *skipped)
2935{
2936 struct r10conf *conf = mddev->private;
2937 struct r10bio *r10_bio;
2938 struct bio *biolist = NULL, *bio;
2939 sector_t max_sector, nr_sectors;
2940 int i;
2941 int max_sync;
2942 sector_t sync_blocks;
2943 sector_t sectors_skipped = 0;
2944 int chunks_skipped = 0;
2945 sector_t chunk_mask = conf->geo.chunk_mask;
2946 int page_idx = 0;
2947
2948 if (!conf->r10buf_pool)
2949 if (init_resync(conf))
2950 return 0;
2951
2952 /*
2953 * Allow skipping a full rebuild for incremental assembly
2954 * of a clean array, like RAID1 does.
2955 */
2956 if (mddev->bitmap == NULL &&
2957 mddev->recovery_cp == MaxSector &&
2958 mddev->reshape_position == MaxSector &&
2959 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2960 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2961 !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2962 conf->fullsync == 0) {
2963 *skipped = 1;
2964 return mddev->dev_sectors - sector_nr;
2965 }
2966
2967 skipped:
2968 max_sector = mddev->dev_sectors;
2969 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2970 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2971 max_sector = mddev->resync_max_sectors;
2972 if (sector_nr >= max_sector) {
2973 conf->cluster_sync_low = 0;
2974 conf->cluster_sync_high = 0;
2975
2976 /* If we aborted, we need to abort the
2977 * sync on the 'current' bitmap chucks (there can
2978 * be several when recovering multiple devices).
2979 * as we may have started syncing it but not finished.
2980 * We can find the current address in
2981 * mddev->curr_resync, but for recovery,
2982 * we need to convert that to several
2983 * virtual addresses.
2984 */
2985 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2986 end_reshape(conf);
2987 close_sync(conf);
2988 return 0;
2989 }
2990
2991 if (mddev->curr_resync < max_sector) { /* aborted */
2992 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2993 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2994 &sync_blocks, 1);
2995 else for (i = 0; i < conf->geo.raid_disks; i++) {
2996 sector_t sect =
2997 raid10_find_virt(conf, mddev->curr_resync, i);
2998 bitmap_end_sync(mddev->bitmap, sect,
2999 &sync_blocks, 1);
3000 }
3001 } else {
3002 /* completed sync */
3003 if ((!mddev->bitmap || conf->fullsync)
3004 && conf->have_replacement
3005 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3006 /* Completed a full sync so the replacements
3007 * are now fully recovered.
3008 */
3009 rcu_read_lock();
3010 for (i = 0; i < conf->geo.raid_disks; i++) {
3011 struct md_rdev *rdev =
3012 rcu_dereference(conf->mirrors[i].replacement);
3013 if (rdev)
3014 rdev->recovery_offset = MaxSector;
3015 }
3016 rcu_read_unlock();
3017 }
3018 conf->fullsync = 0;
3019 }
3020 bitmap_close_sync(mddev->bitmap);
3021 close_sync(conf);
3022 *skipped = 1;
3023 return sectors_skipped;
3024 }
3025
3026 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3027 return reshape_request(mddev, sector_nr, skipped);
3028
3029 if (chunks_skipped >= conf->geo.raid_disks) {
3030 /* if there has been nothing to do on any drive,
3031 * then there is nothing to do at all..
3032 */
3033 *skipped = 1;
3034 return (max_sector - sector_nr) + sectors_skipped;
3035 }
3036
3037 if (max_sector > mddev->resync_max)
3038 max_sector = mddev->resync_max; /* Don't do IO beyond here */
3039
3040 /* make sure whole request will fit in a chunk - if chunks
3041 * are meaningful
3042 */
3043 if (conf->geo.near_copies < conf->geo.raid_disks &&
3044 max_sector > (sector_nr | chunk_mask))
3045 max_sector = (sector_nr | chunk_mask) + 1;
3046
3047 /*
3048 * If there is non-resync activity waiting for a turn, then let it
3049 * though before starting on this new sync request.
3050 */
3051 if (conf->nr_waiting)
3052 schedule_timeout_uninterruptible(1);
3053
3054 /* Again, very different code for resync and recovery.
3055 * Both must result in an r10bio with a list of bios that
3056 * have bi_end_io, bi_sector, bi_disk set,
3057 * and bi_private set to the r10bio.
3058 * For recovery, we may actually create several r10bios
3059 * with 2 bios in each, that correspond to the bios in the main one.
3060 * In this case, the subordinate r10bios link back through a
3061 * borrowed master_bio pointer, and the counter in the master
3062 * includes a ref from each subordinate.
3063 */
3064 /* First, we decide what to do and set ->bi_end_io
3065 * To end_sync_read if we want to read, and
3066 * end_sync_write if we will want to write.
3067 */
3068
3069 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3070 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3071 /* recovery... the complicated one */
3072 int j;
3073 r10_bio = NULL;
3074
3075 for (i = 0 ; i < conf->geo.raid_disks; i++) {
3076 int still_degraded;
3077 struct r10bio *rb2;
3078 sector_t sect;
3079 int must_sync;
3080 int any_working;
3081 struct raid10_info *mirror = &conf->mirrors[i];
3082 struct md_rdev *mrdev, *mreplace;
3083
3084 rcu_read_lock();
3085 mrdev = rcu_dereference(mirror->rdev);
3086 mreplace = rcu_dereference(mirror->replacement);
3087
3088 if ((mrdev == NULL ||
3089 test_bit(Faulty, &mrdev->flags) ||
3090 test_bit(In_sync, &mrdev->flags)) &&
3091 (mreplace == NULL ||
3092 test_bit(Faulty, &mreplace->flags))) {
3093 rcu_read_unlock();
3094 continue;
3095 }
3096
3097 still_degraded = 0;
3098 /* want to reconstruct this device */
3099 rb2 = r10_bio;
3100 sect = raid10_find_virt(conf, sector_nr, i);
3101 if (sect >= mddev->resync_max_sectors) {
3102 /* last stripe is not complete - don't
3103 * try to recover this sector.
3104 */
3105 rcu_read_unlock();
3106 continue;
3107 }
3108 if (mreplace && test_bit(Faulty, &mreplace->flags))
3109 mreplace = NULL;
3110 /* Unless we are doing a full sync, or a replacement
3111 * we only need to recover the block if it is set in
3112 * the bitmap
3113 */
3114 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3115 &sync_blocks, 1);
3116 if (sync_blocks < max_sync)
3117 max_sync = sync_blocks;
3118 if (!must_sync &&
3119 mreplace == NULL &&
3120 !conf->fullsync) {
3121 /* yep, skip the sync_blocks here, but don't assume
3122 * that there will never be anything to do here
3123 */
3124 chunks_skipped = -1;
3125 rcu_read_unlock();
3126 continue;
3127 }
3128 atomic_inc(&mrdev->nr_pending);
3129 if (mreplace)
3130 atomic_inc(&mreplace->nr_pending);
3131 rcu_read_unlock();
3132
3133 r10_bio = raid10_alloc_init_r10buf(conf);
3134 r10_bio->state = 0;
3135 raise_barrier(conf, rb2 != NULL);
3136 atomic_set(&r10_bio->remaining, 0);
3137
3138 r10_bio->master_bio = (struct bio*)rb2;
3139 if (rb2)
3140 atomic_inc(&rb2->remaining);
3141 r10_bio->mddev = mddev;
3142 set_bit(R10BIO_IsRecover, &r10_bio->state);
3143 r10_bio->sector = sect;
3144
3145 raid10_find_phys(conf, r10_bio);
3146
3147 /* Need to check if the array will still be
3148 * degraded
3149 */
3150 rcu_read_lock();
3151 for (j = 0; j < conf->geo.raid_disks; j++) {
3152 struct md_rdev *rdev = rcu_dereference(
3153 conf->mirrors[j].rdev);
3154 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3155 still_degraded = 1;
3156 break;
3157 }
3158 }
3159
3160 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3161 &sync_blocks, still_degraded);
3162
3163 any_working = 0;
3164 for (j=0; j<conf->copies;j++) {
3165 int k;
3166 int d = r10_bio->devs[j].devnum;
3167 sector_t from_addr, to_addr;
3168 struct md_rdev *rdev =
3169 rcu_dereference(conf->mirrors[d].rdev);
3170 sector_t sector, first_bad;
3171 int bad_sectors;
3172 if (!rdev ||
3173 !test_bit(In_sync, &rdev->flags))
3174 continue;
3175 /* This is where we read from */
3176 any_working = 1;
3177 sector = r10_bio->devs[j].addr;
3178
3179 if (is_badblock(rdev, sector, max_sync,
3180 &first_bad, &bad_sectors)) {
3181 if (first_bad > sector)
3182 max_sync = first_bad - sector;
3183 else {
3184 bad_sectors -= (sector
3185 - first_bad);
3186 if (max_sync > bad_sectors)
3187 max_sync = bad_sectors;
3188 continue;
3189 }
3190 }
3191 bio = r10_bio->devs[0].bio;
3192 bio->bi_next = biolist;
3193 biolist = bio;
3194 bio->bi_end_io = end_sync_read;
3195 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3196 if (test_bit(FailFast, &rdev->flags))
3197 bio->bi_opf |= MD_FAILFAST;
3198 from_addr = r10_bio->devs[j].addr;
3199 bio->bi_iter.bi_sector = from_addr +
3200 rdev->data_offset;
3201 bio_set_dev(bio, rdev->bdev);
3202 atomic_inc(&rdev->nr_pending);
3203 /* and we write to 'i' (if not in_sync) */
3204
3205 for (k=0; k<conf->copies; k++)
3206 if (r10_bio->devs[k].devnum == i)
3207 break;
3208 BUG_ON(k == conf->copies);
3209 to_addr = r10_bio->devs[k].addr;
3210 r10_bio->devs[0].devnum = d;
3211 r10_bio->devs[0].addr = from_addr;
3212 r10_bio->devs[1].devnum = i;
3213 r10_bio->devs[1].addr = to_addr;
3214
3215 if (!test_bit(In_sync, &mrdev->flags)) {
3216 bio = r10_bio->devs[1].bio;
3217 bio->bi_next = biolist;
3218 biolist = bio;
3219 bio->bi_end_io = end_sync_write;
3220 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3221 bio->bi_iter.bi_sector = to_addr
3222 + mrdev->data_offset;
3223 bio_set_dev(bio, mrdev->bdev);
3224 atomic_inc(&r10_bio->remaining);
3225 } else
3226 r10_bio->devs[1].bio->bi_end_io = NULL;
3227
3228 /* and maybe write to replacement */
3229 bio = r10_bio->devs[1].repl_bio;
3230 if (bio)
3231 bio->bi_end_io = NULL;
3232 /* Note: if mreplace != NULL, then bio
3233 * cannot be NULL as r10buf_pool_alloc will
3234 * have allocated it.
3235 * So the second test here is pointless.
3236 * But it keeps semantic-checkers happy, and
3237 * this comment keeps human reviewers
3238 * happy.
3239 */
3240 if (mreplace == NULL || bio == NULL ||
3241 test_bit(Faulty, &mreplace->flags))
3242 break;
3243 bio->bi_next = biolist;
3244 biolist = bio;
3245 bio->bi_end_io = end_sync_write;
3246 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3247 bio->bi_iter.bi_sector = to_addr +
3248 mreplace->data_offset;
3249 bio_set_dev(bio, mreplace->bdev);
3250 atomic_inc(&r10_bio->remaining);
3251 break;
3252 }
3253 rcu_read_unlock();
3254 if (j == conf->copies) {
3255 /* Cannot recover, so abort the recovery or
3256 * record a bad block */
3257 if (any_working) {
3258 /* problem is that there are bad blocks
3259 * on other device(s)
3260 */
3261 int k;
3262 for (k = 0; k < conf->copies; k++)
3263 if (r10_bio->devs[k].devnum == i)
3264 break;
3265 if (!test_bit(In_sync,
3266 &mrdev->flags)
3267 && !rdev_set_badblocks(
3268 mrdev,
3269 r10_bio->devs[k].addr,
3270 max_sync, 0))
3271 any_working = 0;
3272 if (mreplace &&
3273 !rdev_set_badblocks(
3274 mreplace,
3275 r10_bio->devs[k].addr,
3276 max_sync, 0))
3277 any_working = 0;
3278 }
3279 if (!any_working) {
3280 if (!test_and_set_bit(MD_RECOVERY_INTR,
3281 &mddev->recovery))
3282 pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3283 mdname(mddev));
3284 mirror->recovery_disabled
3285 = mddev->recovery_disabled;
3286 }
3287 put_buf(r10_bio);
3288 if (rb2)
3289 atomic_dec(&rb2->remaining);
3290 r10_bio = rb2;
3291 rdev_dec_pending(mrdev, mddev);
3292 if (mreplace)
3293 rdev_dec_pending(mreplace, mddev);
3294 break;
3295 }
3296 rdev_dec_pending(mrdev, mddev);
3297 if (mreplace)
3298 rdev_dec_pending(mreplace, mddev);
3299 if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3300 /* Only want this if there is elsewhere to
3301 * read from. 'j' is currently the first
3302 * readable copy.
3303 */
3304 int targets = 1;
3305 for (; j < conf->copies; j++) {
3306 int d = r10_bio->devs[j].devnum;
3307 if (conf->mirrors[d].rdev &&
3308 test_bit(In_sync,
3309 &conf->mirrors[d].rdev->flags))
3310 targets++;
3311 }
3312 if (targets == 1)
3313 r10_bio->devs[0].bio->bi_opf
3314 &= ~MD_FAILFAST;
3315 }
3316 }
3317 if (biolist == NULL) {
3318 while (r10_bio) {
3319 struct r10bio *rb2 = r10_bio;
3320 r10_bio = (struct r10bio*) rb2->master_bio;
3321 rb2->master_bio = NULL;
3322 put_buf(rb2);
3323 }
3324 goto giveup;
3325 }
3326 } else {
3327 /* resync. Schedule a read for every block at this virt offset */
3328 int count = 0;
3329
3330 /*
3331 * Since curr_resync_completed could probably not update in
3332 * time, and we will set cluster_sync_low based on it.
3333 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3334 * safety reason, which ensures curr_resync_completed is
3335 * updated in bitmap_cond_end_sync.
3336 */
3337 bitmap_cond_end_sync(mddev->bitmap, sector_nr,
3338 mddev_is_clustered(mddev) &&
3339 (sector_nr + 2 * RESYNC_SECTORS >
3340 conf->cluster_sync_high));
3341
3342 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3343 &sync_blocks, mddev->degraded) &&
3344 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3345 &mddev->recovery)) {
3346 /* We can skip this block */
3347 *skipped = 1;
3348 return sync_blocks + sectors_skipped;
3349 }
3350 if (sync_blocks < max_sync)
3351 max_sync = sync_blocks;
3352 r10_bio = raid10_alloc_init_r10buf(conf);
3353 r10_bio->state = 0;
3354
3355 r10_bio->mddev = mddev;
3356 atomic_set(&r10_bio->remaining, 0);
3357 raise_barrier(conf, 0);
3358 conf->next_resync = sector_nr;
3359
3360 r10_bio->master_bio = NULL;
3361 r10_bio->sector = sector_nr;
3362 set_bit(R10BIO_IsSync, &r10_bio->state);
3363 raid10_find_phys(conf, r10_bio);
3364 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3365
3366 for (i = 0; i < conf->copies; i++) {
3367 int d = r10_bio->devs[i].devnum;
3368 sector_t first_bad, sector;
3369 int bad_sectors;
3370 struct md_rdev *rdev;
3371
3372 if (r10_bio->devs[i].repl_bio)
3373 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3374
3375 bio = r10_bio->devs[i].bio;
3376 bio->bi_status = BLK_STS_IOERR;
3377 rcu_read_lock();
3378 rdev = rcu_dereference(conf->mirrors[d].rdev);
3379 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3380 rcu_read_unlock();
3381 continue;
3382 }
3383 sector = r10_bio->devs[i].addr;
3384 if (is_badblock(rdev, sector, max_sync,
3385 &first_bad, &bad_sectors)) {
3386 if (first_bad > sector)
3387 max_sync = first_bad - sector;
3388 else {
3389 bad_sectors -= (sector - first_bad);
3390 if (max_sync > bad_sectors)
3391 max_sync = bad_sectors;
3392 rcu_read_unlock();
3393 continue;
3394 }
3395 }
3396 atomic_inc(&rdev->nr_pending);
3397 atomic_inc(&r10_bio->remaining);
3398 bio->bi_next = biolist;
3399 biolist = bio;
3400 bio->bi_end_io = end_sync_read;
3401 bio_set_op_attrs(bio, REQ_OP_READ, 0);
3402 if (test_bit(FailFast, &rdev->flags))
3403 bio->bi_opf |= MD_FAILFAST;
3404 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3405 bio_set_dev(bio, rdev->bdev);
3406 count++;
3407
3408 rdev = rcu_dereference(conf->mirrors[d].replacement);
3409 if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3410 rcu_read_unlock();
3411 continue;
3412 }
3413 atomic_inc(&rdev->nr_pending);
3414
3415 /* Need to set up for writing to the replacement */
3416 bio = r10_bio->devs[i].repl_bio;
3417 bio->bi_status = BLK_STS_IOERR;
3418
3419 sector = r10_bio->devs[i].addr;
3420 bio->bi_next = biolist;
3421 biolist = bio;
3422 bio->bi_end_io = end_sync_write;
3423 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3424 if (test_bit(FailFast, &rdev->flags))
3425 bio->bi_opf |= MD_FAILFAST;
3426 bio->bi_iter.bi_sector = sector + rdev->data_offset;
3427 bio_set_dev(bio, rdev->bdev);
3428 count++;
3429 rcu_read_unlock();
3430 }
3431
3432 if (count < 2) {
3433 for (i=0; i<conf->copies; i++) {
3434 int d = r10_bio->devs[i].devnum;
3435 if (r10_bio->devs[i].bio->bi_end_io)
3436 rdev_dec_pending(conf->mirrors[d].rdev,
3437 mddev);
3438 if (r10_bio->devs[i].repl_bio &&
3439 r10_bio->devs[i].repl_bio->bi_end_io)
3440 rdev_dec_pending(
3441 conf->mirrors[d].replacement,
3442 mddev);
3443 }
3444 put_buf(r10_bio);
3445 biolist = NULL;
3446 goto giveup;
3447 }
3448 }
3449
3450 nr_sectors = 0;
3451 if (sector_nr + max_sync < max_sector)
3452 max_sector = sector_nr + max_sync;
3453 do {
3454 struct page *page;
3455 int len = PAGE_SIZE;
3456 if (sector_nr + (len>>9) > max_sector)
3457 len = (max_sector - sector_nr) << 9;
3458 if (len == 0)
3459 break;
3460 for (bio= biolist ; bio ; bio=bio->bi_next) {
3461 struct resync_pages *rp = get_resync_pages(bio);
3462 page = resync_fetch_page(rp, page_idx);
3463 /*
3464 * won't fail because the vec table is big enough
3465 * to hold all these pages
3466 */
3467 bio_add_page(bio, page, len, 0);
3468 }
3469 nr_sectors += len>>9;
3470 sector_nr += len>>9;
3471 } while (++page_idx < RESYNC_PAGES);
3472 r10_bio->sectors = nr_sectors;
3473
3474 if (mddev_is_clustered(mddev) &&
3475 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3476 /* It is resync not recovery */
3477 if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3478 conf->cluster_sync_low = mddev->curr_resync_completed;
3479 raid10_set_cluster_sync_high(conf);
3480 /* Send resync message */
3481 md_cluster_ops->resync_info_update(mddev,
3482 conf->cluster_sync_low,
3483 conf->cluster_sync_high);
3484 }
3485 } else if (mddev_is_clustered(mddev)) {
3486 /* This is recovery not resync */
3487 sector_t sect_va1, sect_va2;
3488 bool broadcast_msg = false;
3489
3490 for (i = 0; i < conf->geo.raid_disks; i++) {
3491 /*
3492 * sector_nr is a device address for recovery, so we
3493 * need translate it to array address before compare
3494 * with cluster_sync_high.
3495 */
3496 sect_va1 = raid10_find_virt(conf, sector_nr, i);
3497
3498 if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3499 broadcast_msg = true;
3500 /*
3501 * curr_resync_completed is similar as
3502 * sector_nr, so make the translation too.
3503 */
3504 sect_va2 = raid10_find_virt(conf,
3505 mddev->curr_resync_completed, i);
3506
3507 if (conf->cluster_sync_low == 0 ||
3508 conf->cluster_sync_low > sect_va2)
3509 conf->cluster_sync_low = sect_va2;
3510 }
3511 }
3512 if (broadcast_msg) {
3513 raid10_set_cluster_sync_high(conf);
3514 md_cluster_ops->resync_info_update(mddev,
3515 conf->cluster_sync_low,
3516 conf->cluster_sync_high);
3517 }
3518 }
3519
3520 while (biolist) {
3521 bio = biolist;
3522 biolist = biolist->bi_next;
3523
3524 bio->bi_next = NULL;
3525 r10_bio = get_resync_r10bio(bio);
3526 r10_bio->sectors = nr_sectors;
3527
3528 if (bio->bi_end_io == end_sync_read) {
3529 md_sync_acct_bio(bio, nr_sectors);
3530 bio->bi_status = 0;
3531 generic_make_request(bio);
3532 }
3533 }
3534
3535 if (sectors_skipped)
3536 /* pretend they weren't skipped, it makes
3537 * no important difference in this case
3538 */
3539 md_done_sync(mddev, sectors_skipped, 1);
3540
3541 return sectors_skipped + nr_sectors;
3542 giveup:
3543 /* There is nowhere to write, so all non-sync
3544 * drives must be failed or in resync, all drives
3545 * have a bad block, so try the next chunk...
3546 */
3547 if (sector_nr + max_sync < max_sector)
3548 max_sector = sector_nr + max_sync;
3549
3550 sectors_skipped += (max_sector - sector_nr);
3551 chunks_skipped ++;
3552 sector_nr = max_sector;
3553 goto skipped;
3554}
3555
3556static sector_t
3557raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3558{
3559 sector_t size;
3560 struct r10conf *conf = mddev->private;
3561
3562 if (!raid_disks)
3563 raid_disks = min(conf->geo.raid_disks,
3564 conf->prev.raid_disks);
3565 if (!sectors)
3566 sectors = conf->dev_sectors;
3567
3568 size = sectors >> conf->geo.chunk_shift;
3569 sector_div(size, conf->geo.far_copies);
3570 size = size * raid_disks;
3571 sector_div(size, conf->geo.near_copies);
3572
3573 return size << conf->geo.chunk_shift;
3574}
3575
3576static void calc_sectors(struct r10conf *conf, sector_t size)
3577{
3578 /* Calculate the number of sectors-per-device that will
3579 * actually be used, and set conf->dev_sectors and
3580 * conf->stride
3581 */
3582
3583 size = size >> conf->geo.chunk_shift;
3584 sector_div(size, conf->geo.far_copies);
3585 size = size * conf->geo.raid_disks;
3586 sector_div(size, conf->geo.near_copies);
3587 /* 'size' is now the number of chunks in the array */
3588 /* calculate "used chunks per device" */
3589 size = size * conf->copies;
3590
3591 /* We need to round up when dividing by raid_disks to
3592 * get the stride size.
3593 */
3594 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3595
3596 conf->dev_sectors = size << conf->geo.chunk_shift;
3597
3598 if (conf->geo.far_offset)
3599 conf->geo.stride = 1 << conf->geo.chunk_shift;
3600 else {
3601 sector_div(size, conf->geo.far_copies);
3602 conf->geo.stride = size << conf->geo.chunk_shift;
3603 }
3604}
3605
3606enum geo_type {geo_new, geo_old, geo_start};
3607static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3608{
3609 int nc, fc, fo;
3610 int layout, chunk, disks;
3611 switch (new) {
3612 case geo_old:
3613 layout = mddev->layout;
3614 chunk = mddev->chunk_sectors;
3615 disks = mddev->raid_disks - mddev->delta_disks;
3616 break;
3617 case geo_new:
3618 layout = mddev->new_layout;
3619 chunk = mddev->new_chunk_sectors;
3620 disks = mddev->raid_disks;
3621 break;
3622 default: /* avoid 'may be unused' warnings */
3623 case geo_start: /* new when starting reshape - raid_disks not
3624 * updated yet. */
3625 layout = mddev->new_layout;
3626 chunk = mddev->new_chunk_sectors;
3627 disks = mddev->raid_disks + mddev->delta_disks;
3628 break;
3629 }
3630 if (layout >> 19)
3631 return -1;
3632 if (chunk < (PAGE_SIZE >> 9) ||
3633 !is_power_of_2(chunk))
3634 return -2;
3635 nc = layout & 255;
3636 fc = (layout >> 8) & 255;
3637 fo = layout & (1<<16);
3638 geo->raid_disks = disks;
3639 geo->near_copies = nc;
3640 geo->far_copies = fc;
3641 geo->far_offset = fo;
3642 switch (layout >> 17) {
3643 case 0: /* original layout. simple but not always optimal */
3644 geo->far_set_size = disks;
3645 break;
3646 case 1: /* "improved" layout which was buggy. Hopefully no-one is
3647 * actually using this, but leave code here just in case.*/
3648 geo->far_set_size = disks/fc;
3649 WARN(geo->far_set_size < fc,
3650 "This RAID10 layout does not provide data safety - please backup and create new array\n");
3651 break;
3652 case 2: /* "improved" layout fixed to match documentation */
3653 geo->far_set_size = fc * nc;
3654 break;
3655 default: /* Not a valid layout */
3656 return -1;
3657 }
3658 geo->chunk_mask = chunk - 1;
3659 geo->chunk_shift = ffz(~chunk);
3660 return nc*fc;
3661}
3662
3663static struct r10conf *setup_conf(struct mddev *mddev)
3664{
3665 struct r10conf *conf = NULL;
3666 int err = -EINVAL;
3667 struct geom geo;
3668 int copies;
3669
3670 copies = setup_geo(&geo, mddev, geo_new);
3671
3672 if (copies == -2) {
3673 pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3674 mdname(mddev), PAGE_SIZE);
3675 goto out;
3676 }
3677
3678 if (copies < 2 || copies > mddev->raid_disks) {
3679 pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3680 mdname(mddev), mddev->new_layout);
3681 goto out;
3682 }
3683
3684 err = -ENOMEM;
3685 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3686 if (!conf)
3687 goto out;
3688
3689 /* FIXME calc properly */
3690 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3691 max(0,-mddev->delta_disks)),
3692 GFP_KERNEL);
3693 if (!conf->mirrors)
3694 goto out;
3695
3696 conf->tmppage = alloc_page(GFP_KERNEL);
3697 if (!conf->tmppage)
3698 goto out;
3699
3700 conf->geo = geo;
3701 conf->copies = copies;
3702 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3703 r10bio_pool_free, conf);
3704 if (!conf->r10bio_pool)
3705 goto out;
3706
3707 conf->bio_split = bioset_create(BIO_POOL_SIZE, 0, 0);
3708 if (!conf->bio_split)
3709 goto out;
3710
3711 calc_sectors(conf, mddev->dev_sectors);
3712 if (mddev->reshape_position == MaxSector) {
3713 conf->prev = conf->geo;
3714 conf->reshape_progress = MaxSector;
3715 } else {
3716 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3717 err = -EINVAL;
3718 goto out;
3719 }
3720 conf->reshape_progress = mddev->reshape_position;
3721 if (conf->prev.far_offset)
3722 conf->prev.stride = 1 << conf->prev.chunk_shift;
3723 else
3724 /* far_copies must be 1 */
3725 conf->prev.stride = conf->dev_sectors;
3726 }
3727 conf->reshape_safe = conf->reshape_progress;
3728 spin_lock_init(&conf->device_lock);
3729 INIT_LIST_HEAD(&conf->retry_list);
3730 INIT_LIST_HEAD(&conf->bio_end_io_list);
3731
3732 spin_lock_init(&conf->resync_lock);
3733 init_waitqueue_head(&conf->wait_barrier);
3734 atomic_set(&conf->nr_pending, 0);
3735
3736 conf->thread = md_register_thread(raid10d, mddev, "raid10");
3737 if (!conf->thread)
3738 goto out;
3739
3740 conf->mddev = mddev;
3741 return conf;
3742
3743 out:
3744 if (conf) {
3745 mempool_destroy(conf->r10bio_pool);
3746 kfree(conf->mirrors);
3747 safe_put_page(conf->tmppage);
3748 if (conf->bio_split)
3749 bioset_free(conf->bio_split);
3750 kfree(conf);
3751 }
3752 return ERR_PTR(err);
3753}
3754
3755static int raid10_run(struct mddev *mddev)
3756{
3757 struct r10conf *conf;
3758 int i, disk_idx, chunk_size;
3759 struct raid10_info *disk;
3760 struct md_rdev *rdev;
3761 sector_t size;
3762 sector_t min_offset_diff = 0;
3763 int first = 1;
3764 bool discard_supported = false;
3765
3766 if (mddev_init_writes_pending(mddev) < 0)
3767 return -ENOMEM;
3768
3769 if (mddev->private == NULL) {
3770 conf = setup_conf(mddev);
3771 if (IS_ERR(conf))
3772 return PTR_ERR(conf);
3773 mddev->private = conf;
3774 }
3775 conf = mddev->private;
3776 if (!conf)
3777 goto out;
3778
3779 if (mddev_is_clustered(conf->mddev)) {
3780 int fc, fo;
3781
3782 fc = (mddev->layout >> 8) & 255;
3783 fo = mddev->layout & (1<<16);
3784 if (fc > 1 || fo > 0) {
3785 pr_err("only near layout is supported by clustered"
3786 " raid10\n");
3787 goto out_free_conf;
3788 }
3789 }
3790
3791 mddev->thread = conf->thread;
3792 conf->thread = NULL;
3793
3794 chunk_size = mddev->chunk_sectors << 9;
3795 if (mddev->queue) {
3796 blk_queue_max_discard_sectors(mddev->queue,
3797 mddev->chunk_sectors);
3798 blk_queue_max_write_same_sectors(mddev->queue, 0);
3799 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
3800 blk_queue_io_min(mddev->queue, chunk_size);
3801 if (conf->geo.raid_disks % conf->geo.near_copies)
3802 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3803 else
3804 blk_queue_io_opt(mddev->queue, chunk_size *
3805 (conf->geo.raid_disks / conf->geo.near_copies));
3806 }
3807
3808 rdev_for_each(rdev, mddev) {
3809 long long diff;
3810
3811 disk_idx = rdev->raid_disk;
3812 if (disk_idx < 0)
3813 continue;
3814 if (disk_idx >= conf->geo.raid_disks &&
3815 disk_idx >= conf->prev.raid_disks)
3816 continue;
3817 disk = conf->mirrors + disk_idx;
3818
3819 if (test_bit(Replacement, &rdev->flags)) {
3820 if (disk->replacement)
3821 goto out_free_conf;
3822 disk->replacement = rdev;
3823 } else {
3824 if (disk->rdev)
3825 goto out_free_conf;
3826 disk->rdev = rdev;
3827 }
3828 diff = (rdev->new_data_offset - rdev->data_offset);
3829 if (!mddev->reshape_backwards)
3830 diff = -diff;
3831 if (diff < 0)
3832 diff = 0;
3833 if (first || diff < min_offset_diff)
3834 min_offset_diff = diff;
3835
3836 if (mddev->gendisk)
3837 disk_stack_limits(mddev->gendisk, rdev->bdev,
3838 rdev->data_offset << 9);
3839
3840 disk->head_position = 0;
3841
3842 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3843 discard_supported = true;
3844 first = 0;
3845 }
3846
3847 if (mddev->queue) {
3848 if (discard_supported)
3849 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
3850 mddev->queue);
3851 else
3852 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
3853 mddev->queue);
3854 }
3855 /* need to check that every block has at least one working mirror */
3856 if (!enough(conf, -1)) {
3857 pr_err("md/raid10:%s: not enough operational mirrors.\n",
3858 mdname(mddev));
3859 goto out_free_conf;
3860 }
3861
3862 if (conf->reshape_progress != MaxSector) {
3863 /* must ensure that shape change is supported */
3864 if (conf->geo.far_copies != 1 &&
3865 conf->geo.far_offset == 0)
3866 goto out_free_conf;
3867 if (conf->prev.far_copies != 1 &&
3868 conf->prev.far_offset == 0)
3869 goto out_free_conf;
3870 }
3871
3872 mddev->degraded = 0;
3873 for (i = 0;
3874 i < conf->geo.raid_disks
3875 || i < conf->prev.raid_disks;
3876 i++) {
3877
3878 disk = conf->mirrors + i;
3879
3880 if (!disk->rdev && disk->replacement) {
3881 /* The replacement is all we have - use it */
3882 disk->rdev = disk->replacement;
3883 disk->replacement = NULL;
3884 clear_bit(Replacement, &disk->rdev->flags);
3885 }
3886
3887 if (!disk->rdev ||
3888 !test_bit(In_sync, &disk->rdev->flags)) {
3889 disk->head_position = 0;
3890 mddev->degraded++;
3891 if (disk->rdev &&
3892 disk->rdev->saved_raid_disk < 0)
3893 conf->fullsync = 1;
3894 }
3895 disk->recovery_disabled = mddev->recovery_disabled - 1;
3896 }
3897
3898 if (mddev->recovery_cp != MaxSector)
3899 pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3900 mdname(mddev));
3901 pr_info("md/raid10:%s: active with %d out of %d devices\n",
3902 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3903 conf->geo.raid_disks);
3904 /*
3905 * Ok, everything is just fine now
3906 */
3907 mddev->dev_sectors = conf->dev_sectors;
3908 size = raid10_size(mddev, 0, 0);
3909 md_set_array_sectors(mddev, size);
3910 mddev->resync_max_sectors = size;
3911 set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3912
3913 if (mddev->queue) {
3914 int stripe = conf->geo.raid_disks *
3915 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3916
3917 /* Calculate max read-ahead size.
3918 * We need to readahead at least twice a whole stripe....
3919 * maybe...
3920 */
3921 stripe /= conf->geo.near_copies;
3922 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
3923 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
3924 }
3925
3926 if (md_integrity_register(mddev))
3927 goto out_free_conf;
3928
3929 if (conf->reshape_progress != MaxSector) {
3930 unsigned long before_length, after_length;
3931
3932 before_length = ((1 << conf->prev.chunk_shift) *
3933 conf->prev.far_copies);
3934 after_length = ((1 << conf->geo.chunk_shift) *
3935 conf->geo.far_copies);
3936
3937 if (max(before_length, after_length) > min_offset_diff) {
3938 /* This cannot work */
3939 pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3940 goto out_free_conf;
3941 }
3942 conf->offset_diff = min_offset_diff;
3943
3944 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3945 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3946 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3947 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3948 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3949 "reshape");
3950 }
3951
3952 return 0;
3953
3954out_free_conf:
3955 md_unregister_thread(&mddev->thread);
3956 mempool_destroy(conf->r10bio_pool);
3957 safe_put_page(conf->tmppage);
3958 kfree(conf->mirrors);
3959 kfree(conf);
3960 mddev->private = NULL;
3961out:
3962 return -EIO;
3963}
3964
3965static void raid10_free(struct mddev *mddev, void *priv)
3966{
3967 struct r10conf *conf = priv;
3968
3969 mempool_destroy(conf->r10bio_pool);
3970 safe_put_page(conf->tmppage);
3971 kfree(conf->mirrors);
3972 kfree(conf->mirrors_old);
3973 kfree(conf->mirrors_new);
3974 if (conf->bio_split)
3975 bioset_free(conf->bio_split);
3976 kfree(conf);
3977}
3978
3979static void raid10_quiesce(struct mddev *mddev, int quiesce)
3980{
3981 struct r10conf *conf = mddev->private;
3982
3983 if (quiesce)
3984 raise_barrier(conf, 0);
3985 else
3986 lower_barrier(conf);
3987}
3988
3989static int raid10_resize(struct mddev *mddev, sector_t sectors)
3990{
3991 /* Resize of 'far' arrays is not supported.
3992 * For 'near' and 'offset' arrays we can set the
3993 * number of sectors used to be an appropriate multiple
3994 * of the chunk size.
3995 * For 'offset', this is far_copies*chunksize.
3996 * For 'near' the multiplier is the LCM of
3997 * near_copies and raid_disks.
3998 * So if far_copies > 1 && !far_offset, fail.
3999 * Else find LCM(raid_disks, near_copy)*far_copies and
4000 * multiply by chunk_size. Then round to this number.
4001 * This is mostly done by raid10_size()
4002 */
4003 struct r10conf *conf = mddev->private;
4004 sector_t oldsize, size;
4005
4006 if (mddev->reshape_position != MaxSector)
4007 return -EBUSY;
4008
4009 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4010 return -EINVAL;
4011
4012 oldsize = raid10_size(mddev, 0, 0);
4013 size = raid10_size(mddev, sectors, 0);
4014 if (mddev->external_size &&
4015 mddev->array_sectors > size)
4016 return -EINVAL;
4017 if (mddev->bitmap) {
4018 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
4019 if (ret)
4020 return ret;
4021 }
4022 md_set_array_sectors(mddev, size);
4023 if (sectors > mddev->dev_sectors &&
4024 mddev->recovery_cp > oldsize) {
4025 mddev->recovery_cp = oldsize;
4026 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4027 }
4028 calc_sectors(conf, sectors);
4029 mddev->dev_sectors = conf->dev_sectors;
4030 mddev->resync_max_sectors = size;
4031 return 0;
4032}
4033
4034static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4035{
4036 struct md_rdev *rdev;
4037 struct r10conf *conf;
4038
4039 if (mddev->degraded > 0) {
4040 pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4041 mdname(mddev));
4042 return ERR_PTR(-EINVAL);
4043 }
4044 sector_div(size, devs);
4045
4046 /* Set new parameters */
4047 mddev->new_level = 10;
4048 /* new layout: far_copies = 1, near_copies = 2 */
4049 mddev->new_layout = (1<<8) + 2;
4050 mddev->new_chunk_sectors = mddev->chunk_sectors;
4051 mddev->delta_disks = mddev->raid_disks;
4052 mddev->raid_disks *= 2;
4053 /* make sure it will be not marked as dirty */
4054 mddev->recovery_cp = MaxSector;
4055 mddev->dev_sectors = size;
4056
4057 conf = setup_conf(mddev);
4058 if (!IS_ERR(conf)) {
4059 rdev_for_each(rdev, mddev)
4060 if (rdev->raid_disk >= 0) {
4061 rdev->new_raid_disk = rdev->raid_disk * 2;
4062 rdev->sectors = size;
4063 }
4064 conf->barrier = 1;
4065 }
4066
4067 return conf;
4068}
4069
4070static void *raid10_takeover(struct mddev *mddev)
4071{
4072 struct r0conf *raid0_conf;
4073
4074 /* raid10 can take over:
4075 * raid0 - providing it has only two drives
4076 */
4077 if (mddev->level == 0) {
4078 /* for raid0 takeover only one zone is supported */
4079 raid0_conf = mddev->private;
4080 if (raid0_conf->nr_strip_zones > 1) {
4081 pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4082 mdname(mddev));
4083 return ERR_PTR(-EINVAL);
4084 }
4085 return raid10_takeover_raid0(mddev,
4086 raid0_conf->strip_zone->zone_end,
4087 raid0_conf->strip_zone->nb_dev);
4088 }
4089 return ERR_PTR(-EINVAL);
4090}
4091
4092static int raid10_check_reshape(struct mddev *mddev)
4093{
4094 /* Called when there is a request to change
4095 * - layout (to ->new_layout)
4096 * - chunk size (to ->new_chunk_sectors)
4097 * - raid_disks (by delta_disks)
4098 * or when trying to restart a reshape that was ongoing.
4099 *
4100 * We need to validate the request and possibly allocate
4101 * space if that might be an issue later.
4102 *
4103 * Currently we reject any reshape of a 'far' mode array,
4104 * allow chunk size to change if new is generally acceptable,
4105 * allow raid_disks to increase, and allow
4106 * a switch between 'near' mode and 'offset' mode.
4107 */
4108 struct r10conf *conf = mddev->private;
4109 struct geom geo;
4110
4111 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4112 return -EINVAL;
4113
4114 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4115 /* mustn't change number of copies */
4116 return -EINVAL;
4117 if (geo.far_copies > 1 && !geo.far_offset)
4118 /* Cannot switch to 'far' mode */
4119 return -EINVAL;
4120
4121 if (mddev->array_sectors & geo.chunk_mask)
4122 /* not factor of array size */
4123 return -EINVAL;
4124
4125 if (!enough(conf, -1))
4126 return -EINVAL;
4127
4128 kfree(conf->mirrors_new);
4129 conf->mirrors_new = NULL;
4130 if (mddev->delta_disks > 0) {
4131 /* allocate new 'mirrors' list */
4132 conf->mirrors_new = kzalloc(
4133 sizeof(struct raid10_info)
4134 *(mddev->raid_disks +
4135 mddev->delta_disks),
4136 GFP_KERNEL);
4137 if (!conf->mirrors_new)
4138 return -ENOMEM;
4139 }
4140 return 0;
4141}
4142
4143/*
4144 * Need to check if array has failed when deciding whether to:
4145 * - start an array
4146 * - remove non-faulty devices
4147 * - add a spare
4148 * - allow a reshape
4149 * This determination is simple when no reshape is happening.
4150 * However if there is a reshape, we need to carefully check
4151 * both the before and after sections.
4152 * This is because some failed devices may only affect one
4153 * of the two sections, and some non-in_sync devices may
4154 * be insync in the section most affected by failed devices.
4155 */
4156static int calc_degraded(struct r10conf *conf)
4157{
4158 int degraded, degraded2;
4159 int i;
4160
4161 rcu_read_lock();
4162 degraded = 0;
4163 /* 'prev' section first */
4164 for (i = 0; i < conf->prev.raid_disks; i++) {
4165 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4166 if (!rdev || test_bit(Faulty, &rdev->flags))
4167 degraded++;
4168 else if (!test_bit(In_sync, &rdev->flags))
4169 /* When we can reduce the number of devices in
4170 * an array, this might not contribute to
4171 * 'degraded'. It does now.
4172 */
4173 degraded++;
4174 }
4175 rcu_read_unlock();
4176 if (conf->geo.raid_disks == conf->prev.raid_disks)
4177 return degraded;
4178 rcu_read_lock();
4179 degraded2 = 0;
4180 for (i = 0; i < conf->geo.raid_disks; i++) {
4181 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4182 if (!rdev || test_bit(Faulty, &rdev->flags))
4183 degraded2++;
4184 else if (!test_bit(In_sync, &rdev->flags)) {
4185 /* If reshape is increasing the number of devices,
4186 * this section has already been recovered, so
4187 * it doesn't contribute to degraded.
4188 * else it does.
4189 */
4190 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4191 degraded2++;
4192 }
4193 }
4194 rcu_read_unlock();
4195 if (degraded2 > degraded)
4196 return degraded2;
4197 return degraded;
4198}
4199
4200static int raid10_start_reshape(struct mddev *mddev)
4201{
4202 /* A 'reshape' has been requested. This commits
4203 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4204 * This also checks if there are enough spares and adds them
4205 * to the array.
4206 * We currently require enough spares to make the final
4207 * array non-degraded. We also require that the difference
4208 * between old and new data_offset - on each device - is
4209 * enough that we never risk over-writing.
4210 */
4211
4212 unsigned long before_length, after_length;
4213 sector_t min_offset_diff = 0;
4214 int first = 1;
4215 struct geom new;
4216 struct r10conf *conf = mddev->private;
4217 struct md_rdev *rdev;
4218 int spares = 0;
4219 int ret;
4220
4221 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4222 return -EBUSY;
4223
4224 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4225 return -EINVAL;
4226
4227 before_length = ((1 << conf->prev.chunk_shift) *
4228 conf->prev.far_copies);
4229 after_length = ((1 << conf->geo.chunk_shift) *
4230 conf->geo.far_copies);
4231
4232 rdev_for_each(rdev, mddev) {
4233 if (!test_bit(In_sync, &rdev->flags)
4234 && !test_bit(Faulty, &rdev->flags))
4235 spares++;
4236 if (rdev->raid_disk >= 0) {
4237 long long diff = (rdev->new_data_offset
4238 - rdev->data_offset);
4239 if (!mddev->reshape_backwards)
4240 diff = -diff;
4241 if (diff < 0)
4242 diff = 0;
4243 if (first || diff < min_offset_diff)
4244 min_offset_diff = diff;
4245 first = 0;
4246 }
4247 }
4248
4249 if (max(before_length, after_length) > min_offset_diff)
4250 return -EINVAL;
4251
4252 if (spares < mddev->delta_disks)
4253 return -EINVAL;
4254
4255 conf->offset_diff = min_offset_diff;
4256 spin_lock_irq(&conf->device_lock);
4257 if (conf->mirrors_new) {
4258 memcpy(conf->mirrors_new, conf->mirrors,
4259 sizeof(struct raid10_info)*conf->prev.raid_disks);
4260 smp_mb();
4261 kfree(conf->mirrors_old);
4262 conf->mirrors_old = conf->mirrors;
4263 conf->mirrors = conf->mirrors_new;
4264 conf->mirrors_new = NULL;
4265 }
4266 setup_geo(&conf->geo, mddev, geo_start);
4267 smp_mb();
4268 if (mddev->reshape_backwards) {
4269 sector_t size = raid10_size(mddev, 0, 0);
4270 if (size < mddev->array_sectors) {
4271 spin_unlock_irq(&conf->device_lock);
4272 pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4273 mdname(mddev));
4274 return -EINVAL;
4275 }
4276 mddev->resync_max_sectors = size;
4277 conf->reshape_progress = size;
4278 } else
4279 conf->reshape_progress = 0;
4280 conf->reshape_safe = conf->reshape_progress;
4281 spin_unlock_irq(&conf->device_lock);
4282
4283 if (mddev->delta_disks && mddev->bitmap) {
4284 ret = bitmap_resize(mddev->bitmap,
4285 raid10_size(mddev, 0,
4286 conf->geo.raid_disks),
4287 0, 0);
4288 if (ret)
4289 goto abort;
4290 }
4291 if (mddev->delta_disks > 0) {
4292 rdev_for_each(rdev, mddev)
4293 if (rdev->raid_disk < 0 &&
4294 !test_bit(Faulty, &rdev->flags)) {
4295 if (raid10_add_disk(mddev, rdev) == 0) {
4296 if (rdev->raid_disk >=
4297 conf->prev.raid_disks)
4298 set_bit(In_sync, &rdev->flags);
4299 else
4300 rdev->recovery_offset = 0;
4301
4302 if (sysfs_link_rdev(mddev, rdev))
4303 /* Failure here is OK */;
4304 }
4305 } else if (rdev->raid_disk >= conf->prev.raid_disks
4306 && !test_bit(Faulty, &rdev->flags)) {
4307 /* This is a spare that was manually added */
4308 set_bit(In_sync, &rdev->flags);
4309 }
4310 }
4311 /* When a reshape changes the number of devices,
4312 * ->degraded is measured against the larger of the
4313 * pre and post numbers.
4314 */
4315 spin_lock_irq(&conf->device_lock);
4316 mddev->degraded = calc_degraded(conf);
4317 spin_unlock_irq(&conf->device_lock);
4318 mddev->raid_disks = conf->geo.raid_disks;
4319 mddev->reshape_position = conf->reshape_progress;
4320 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4321
4322 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4323 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4324 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4325 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4326 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4327
4328 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4329 "reshape");
4330 if (!mddev->sync_thread) {
4331 ret = -EAGAIN;
4332 goto abort;
4333 }
4334 conf->reshape_checkpoint = jiffies;
4335 md_wakeup_thread(mddev->sync_thread);
4336 md_new_event(mddev);
4337 return 0;
4338
4339abort:
4340 mddev->recovery = 0;
4341 spin_lock_irq(&conf->device_lock);
4342 conf->geo = conf->prev;
4343 mddev->raid_disks = conf->geo.raid_disks;
4344 rdev_for_each(rdev, mddev)
4345 rdev->new_data_offset = rdev->data_offset;
4346 smp_wmb();
4347 conf->reshape_progress = MaxSector;
4348 conf->reshape_safe = MaxSector;
4349 mddev->reshape_position = MaxSector;
4350 spin_unlock_irq(&conf->device_lock);
4351 return ret;
4352}
4353
4354/* Calculate the last device-address that could contain
4355 * any block from the chunk that includes the array-address 's'
4356 * and report the next address.
4357 * i.e. the address returned will be chunk-aligned and after
4358 * any data that is in the chunk containing 's'.
4359 */
4360static sector_t last_dev_address(sector_t s, struct geom *geo)
4361{
4362 s = (s | geo->chunk_mask) + 1;
4363 s >>= geo->chunk_shift;
4364 s *= geo->near_copies;
4365 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4366 s *= geo->far_copies;
4367 s <<= geo->chunk_shift;
4368 return s;
4369}
4370
4371/* Calculate the first device-address that could contain
4372 * any block from the chunk that includes the array-address 's'.
4373 * This too will be the start of a chunk
4374 */
4375static sector_t first_dev_address(sector_t s, struct geom *geo)
4376{
4377 s >>= geo->chunk_shift;
4378 s *= geo->near_copies;
4379 sector_div(s, geo->raid_disks);
4380 s *= geo->far_copies;
4381 s <<= geo->chunk_shift;
4382 return s;
4383}
4384
4385static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4386 int *skipped)
4387{
4388 /* We simply copy at most one chunk (smallest of old and new)
4389 * at a time, possibly less if that exceeds RESYNC_PAGES,
4390 * or we hit a bad block or something.
4391 * This might mean we pause for normal IO in the middle of
4392 * a chunk, but that is not a problem as mddev->reshape_position
4393 * can record any location.
4394 *
4395 * If we will want to write to a location that isn't
4396 * yet recorded as 'safe' (i.e. in metadata on disk) then
4397 * we need to flush all reshape requests and update the metadata.
4398 *
4399 * When reshaping forwards (e.g. to more devices), we interpret
4400 * 'safe' as the earliest block which might not have been copied
4401 * down yet. We divide this by previous stripe size and multiply
4402 * by previous stripe length to get lowest device offset that we
4403 * cannot write to yet.
4404 * We interpret 'sector_nr' as an address that we want to write to.
4405 * From this we use last_device_address() to find where we might
4406 * write to, and first_device_address on the 'safe' position.
4407 * If this 'next' write position is after the 'safe' position,
4408 * we must update the metadata to increase the 'safe' position.
4409 *
4410 * When reshaping backwards, we round in the opposite direction
4411 * and perform the reverse test: next write position must not be
4412 * less than current safe position.
4413 *
4414 * In all this the minimum difference in data offsets
4415 * (conf->offset_diff - always positive) allows a bit of slack,
4416 * so next can be after 'safe', but not by more than offset_diff
4417 *
4418 * We need to prepare all the bios here before we start any IO
4419 * to ensure the size we choose is acceptable to all devices.
4420 * The means one for each copy for write-out and an extra one for
4421 * read-in.
4422 * We store the read-in bio in ->master_bio and the others in
4423 * ->devs[x].bio and ->devs[x].repl_bio.
4424 */
4425 struct r10conf *conf = mddev->private;
4426 struct r10bio *r10_bio;
4427 sector_t next, safe, last;
4428 int max_sectors;
4429 int nr_sectors;
4430 int s;
4431 struct md_rdev *rdev;
4432 int need_flush = 0;
4433 struct bio *blist;
4434 struct bio *bio, *read_bio;
4435 int sectors_done = 0;
4436 struct page **pages;
4437
4438 if (sector_nr == 0) {
4439 /* If restarting in the middle, skip the initial sectors */
4440 if (mddev->reshape_backwards &&
4441 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4442 sector_nr = (raid10_size(mddev, 0, 0)
4443 - conf->reshape_progress);
4444 } else if (!mddev->reshape_backwards &&
4445 conf->reshape_progress > 0)
4446 sector_nr = conf->reshape_progress;
4447 if (sector_nr) {
4448 mddev->curr_resync_completed = sector_nr;
4449 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4450 *skipped = 1;
4451 return sector_nr;
4452 }
4453 }
4454
4455 /* We don't use sector_nr to track where we are up to
4456 * as that doesn't work well for ->reshape_backwards.
4457 * So just use ->reshape_progress.
4458 */
4459 if (mddev->reshape_backwards) {
4460 /* 'next' is the earliest device address that we might
4461 * write to for this chunk in the new layout
4462 */
4463 next = first_dev_address(conf->reshape_progress - 1,
4464 &conf->geo);
4465
4466 /* 'safe' is the last device address that we might read from
4467 * in the old layout after a restart
4468 */
4469 safe = last_dev_address(conf->reshape_safe - 1,
4470 &conf->prev);
4471
4472 if (next + conf->offset_diff < safe)
4473 need_flush = 1;
4474
4475 last = conf->reshape_progress - 1;
4476 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4477 & conf->prev.chunk_mask);
4478 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4479 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4480 } else {
4481 /* 'next' is after the last device address that we
4482 * might write to for this chunk in the new layout
4483 */
4484 next = last_dev_address(conf->reshape_progress, &conf->geo);
4485
4486 /* 'safe' is the earliest device address that we might
4487 * read from in the old layout after a restart
4488 */
4489 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4490
4491 /* Need to update metadata if 'next' might be beyond 'safe'
4492 * as that would possibly corrupt data
4493 */
4494 if (next > safe + conf->offset_diff)
4495 need_flush = 1;
4496
4497 sector_nr = conf->reshape_progress;
4498 last = sector_nr | (conf->geo.chunk_mask
4499 & conf->prev.chunk_mask);
4500
4501 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4502 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4503 }
4504
4505 if (need_flush ||
4506 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4507 /* Need to update reshape_position in metadata */
4508 wait_barrier(conf);
4509 mddev->reshape_position = conf->reshape_progress;
4510 if (mddev->reshape_backwards)
4511 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4512 - conf->reshape_progress;
4513 else
4514 mddev->curr_resync_completed = conf->reshape_progress;
4515 conf->reshape_checkpoint = jiffies;
4516 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4517 md_wakeup_thread(mddev->thread);
4518 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4519 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4520 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4521 allow_barrier(conf);
4522 return sectors_done;
4523 }
4524 conf->reshape_safe = mddev->reshape_position;
4525 allow_barrier(conf);
4526 }
4527
4528read_more:
4529 /* Now schedule reads for blocks from sector_nr to last */
4530 r10_bio = raid10_alloc_init_r10buf(conf);
4531 r10_bio->state = 0;
4532 raise_barrier(conf, sectors_done != 0);
4533 atomic_set(&r10_bio->remaining, 0);
4534 r10_bio->mddev = mddev;
4535 r10_bio->sector = sector_nr;
4536 set_bit(R10BIO_IsReshape, &r10_bio->state);
4537 r10_bio->sectors = last - sector_nr + 1;
4538 rdev = read_balance(conf, r10_bio, &max_sectors);
4539 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4540
4541 if (!rdev) {
4542 /* Cannot read from here, so need to record bad blocks
4543 * on all the target devices.
4544 */
4545 // FIXME
4546 mempool_free(r10_bio, conf->r10buf_pool);
4547 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4548 return sectors_done;
4549 }
4550
4551 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4552
4553 bio_set_dev(read_bio, rdev->bdev);
4554 read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4555 + rdev->data_offset);
4556 read_bio->bi_private = r10_bio;
4557 read_bio->bi_end_io = end_reshape_read;
4558 bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4559 read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4560 read_bio->bi_status = 0;
4561 read_bio->bi_vcnt = 0;
4562 read_bio->bi_iter.bi_size = 0;
4563 r10_bio->master_bio = read_bio;
4564 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4565
4566 /* Now find the locations in the new layout */
4567 __raid10_find_phys(&conf->geo, r10_bio);
4568
4569 blist = read_bio;
4570 read_bio->bi_next = NULL;
4571
4572 rcu_read_lock();
4573 for (s = 0; s < conf->copies*2; s++) {
4574 struct bio *b;
4575 int d = r10_bio->devs[s/2].devnum;
4576 struct md_rdev *rdev2;
4577 if (s&1) {
4578 rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4579 b = r10_bio->devs[s/2].repl_bio;
4580 } else {
4581 rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4582 b = r10_bio->devs[s/2].bio;
4583 }
4584 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4585 continue;
4586
4587 bio_set_dev(b, rdev2->bdev);
4588 b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4589 rdev2->new_data_offset;
4590 b->bi_end_io = end_reshape_write;
4591 bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4592 b->bi_next = blist;
4593 blist = b;
4594 }
4595
4596 /* Now add as many pages as possible to all of these bios. */
4597
4598 nr_sectors = 0;
4599 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4600 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4601 struct page *page = pages[s / (PAGE_SIZE >> 9)];
4602 int len = (max_sectors - s) << 9;
4603 if (len > PAGE_SIZE)
4604 len = PAGE_SIZE;
4605 for (bio = blist; bio ; bio = bio->bi_next) {
4606 /*
4607 * won't fail because the vec table is big enough
4608 * to hold all these pages
4609 */
4610 bio_add_page(bio, page, len, 0);
4611 }
4612 sector_nr += len >> 9;
4613 nr_sectors += len >> 9;
4614 }
4615 rcu_read_unlock();
4616 r10_bio->sectors = nr_sectors;
4617
4618 /* Now submit the read */
4619 md_sync_acct_bio(read_bio, r10_bio->sectors);
4620 atomic_inc(&r10_bio->remaining);
4621 read_bio->bi_next = NULL;
4622 generic_make_request(read_bio);
4623 sector_nr += nr_sectors;
4624 sectors_done += nr_sectors;
4625 if (sector_nr <= last)
4626 goto read_more;
4627
4628 /* Now that we have done the whole section we can
4629 * update reshape_progress
4630 */
4631 if (mddev->reshape_backwards)
4632 conf->reshape_progress -= sectors_done;
4633 else
4634 conf->reshape_progress += sectors_done;
4635
4636 return sectors_done;
4637}
4638
4639static void end_reshape_request(struct r10bio *r10_bio);
4640static int handle_reshape_read_error(struct mddev *mddev,
4641 struct r10bio *r10_bio);
4642static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4643{
4644 /* Reshape read completed. Hopefully we have a block
4645 * to write out.
4646 * If we got a read error then we do sync 1-page reads from
4647 * elsewhere until we find the data - or give up.
4648 */
4649 struct r10conf *conf = mddev->private;
4650 int s;
4651
4652 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4653 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4654 /* Reshape has been aborted */
4655 md_done_sync(mddev, r10_bio->sectors, 0);
4656 return;
4657 }
4658
4659 /* We definitely have the data in the pages, schedule the
4660 * writes.
4661 */
4662 atomic_set(&r10_bio->remaining, 1);
4663 for (s = 0; s < conf->copies*2; s++) {
4664 struct bio *b;
4665 int d = r10_bio->devs[s/2].devnum;
4666 struct md_rdev *rdev;
4667 rcu_read_lock();
4668 if (s&1) {
4669 rdev = rcu_dereference(conf->mirrors[d].replacement);
4670 b = r10_bio->devs[s/2].repl_bio;
4671 } else {
4672 rdev = rcu_dereference(conf->mirrors[d].rdev);
4673 b = r10_bio->devs[s/2].bio;
4674 }
4675 if (!rdev || test_bit(Faulty, &rdev->flags)) {
4676 rcu_read_unlock();
4677 continue;
4678 }
4679 atomic_inc(&rdev->nr_pending);
4680 rcu_read_unlock();
4681 md_sync_acct_bio(b, r10_bio->sectors);
4682 atomic_inc(&r10_bio->remaining);
4683 b->bi_next = NULL;
4684 generic_make_request(b);
4685 }
4686 end_reshape_request(r10_bio);
4687}
4688
4689static void end_reshape(struct r10conf *conf)
4690{
4691 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4692 return;
4693
4694 spin_lock_irq(&conf->device_lock);
4695 conf->prev = conf->geo;
4696 md_finish_reshape(conf->mddev);
4697 smp_wmb();
4698 conf->reshape_progress = MaxSector;
4699 conf->reshape_safe = MaxSector;
4700 spin_unlock_irq(&conf->device_lock);
4701
4702 /* read-ahead size must cover two whole stripes, which is
4703 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4704 */
4705 if (conf->mddev->queue) {
4706 int stripe = conf->geo.raid_disks *
4707 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4708 stripe /= conf->geo.near_copies;
4709 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
4710 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
4711 }
4712 conf->fullsync = 0;
4713}
4714
4715static int handle_reshape_read_error(struct mddev *mddev,
4716 struct r10bio *r10_bio)
4717{
4718 /* Use sync reads to get the blocks from somewhere else */
4719 int sectors = r10_bio->sectors;
4720 struct r10conf *conf = mddev->private;
4721 struct r10bio *r10b;
4722 int slot = 0;
4723 int idx = 0;
4724 struct page **pages;
4725
4726 r10b = kmalloc(sizeof(*r10b) +
4727 sizeof(struct r10dev) * conf->copies, GFP_NOIO);
4728 if (!r10b) {
4729 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4730 return -ENOMEM;
4731 }
4732
4733 /* reshape IOs share pages from .devs[0].bio */
4734 pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4735
4736 r10b->sector = r10_bio->sector;
4737 __raid10_find_phys(&conf->prev, r10b);
4738
4739 while (sectors) {
4740 int s = sectors;
4741 int success = 0;
4742 int first_slot = slot;
4743
4744 if (s > (PAGE_SIZE >> 9))
4745 s = PAGE_SIZE >> 9;
4746
4747 rcu_read_lock();
4748 while (!success) {
4749 int d = r10b->devs[slot].devnum;
4750 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4751 sector_t addr;
4752 if (rdev == NULL ||
4753 test_bit(Faulty, &rdev->flags) ||
4754 !test_bit(In_sync, &rdev->flags))
4755 goto failed;
4756
4757 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4758 atomic_inc(&rdev->nr_pending);
4759 rcu_read_unlock();
4760 success = sync_page_io(rdev,
4761 addr,
4762 s << 9,
4763 pages[idx],
4764 REQ_OP_READ, 0, false);
4765 rdev_dec_pending(rdev, mddev);
4766 rcu_read_lock();
4767 if (success)
4768 break;
4769 failed:
4770 slot++;
4771 if (slot >= conf->copies)
4772 slot = 0;
4773 if (slot == first_slot)
4774 break;
4775 }
4776 rcu_read_unlock();
4777 if (!success) {
4778 /* couldn't read this block, must give up */
4779 set_bit(MD_RECOVERY_INTR,
4780 &mddev->recovery);
4781 kfree(r10b);
4782 return -EIO;
4783 }
4784 sectors -= s;
4785 idx++;
4786 }
4787 kfree(r10b);
4788 return 0;
4789}
4790
4791static void end_reshape_write(struct bio *bio)
4792{
4793 struct r10bio *r10_bio = get_resync_r10bio(bio);
4794 struct mddev *mddev = r10_bio->mddev;
4795 struct r10conf *conf = mddev->private;
4796 int d;
4797 int slot;
4798 int repl;
4799 struct md_rdev *rdev = NULL;
4800
4801 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4802 if (repl)
4803 rdev = conf->mirrors[d].replacement;
4804 if (!rdev) {
4805 smp_mb();
4806 rdev = conf->mirrors[d].rdev;
4807 }
4808
4809 if (bio->bi_status) {
4810 /* FIXME should record badblock */
4811 md_error(mddev, rdev);
4812 }
4813
4814 rdev_dec_pending(rdev, mddev);
4815 end_reshape_request(r10_bio);
4816}
4817
4818static void end_reshape_request(struct r10bio *r10_bio)
4819{
4820 if (!atomic_dec_and_test(&r10_bio->remaining))
4821 return;
4822 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4823 bio_put(r10_bio->master_bio);
4824 put_buf(r10_bio);
4825}
4826
4827static void raid10_finish_reshape(struct mddev *mddev)
4828{
4829 struct r10conf *conf = mddev->private;
4830
4831 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4832 return;
4833
4834 if (mddev->delta_disks > 0) {
4835 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4836 mddev->recovery_cp = mddev->resync_max_sectors;
4837 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4838 }
4839 mddev->resync_max_sectors = mddev->array_sectors;
4840 } else {
4841 int d;
4842 rcu_read_lock();
4843 for (d = conf->geo.raid_disks ;
4844 d < conf->geo.raid_disks - mddev->delta_disks;
4845 d++) {
4846 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4847 if (rdev)
4848 clear_bit(In_sync, &rdev->flags);
4849 rdev = rcu_dereference(conf->mirrors[d].replacement);
4850 if (rdev)
4851 clear_bit(In_sync, &rdev->flags);
4852 }
4853 rcu_read_unlock();
4854 }
4855 mddev->layout = mddev->new_layout;
4856 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4857 mddev->reshape_position = MaxSector;
4858 mddev->delta_disks = 0;
4859 mddev->reshape_backwards = 0;
4860}
4861
4862static struct md_personality raid10_personality =
4863{
4864 .name = "raid10",
4865 .level = 10,
4866 .owner = THIS_MODULE,
4867 .make_request = raid10_make_request,
4868 .run = raid10_run,
4869 .free = raid10_free,
4870 .status = raid10_status,
4871 .error_handler = raid10_error,
4872 .hot_add_disk = raid10_add_disk,
4873 .hot_remove_disk= raid10_remove_disk,
4874 .spare_active = raid10_spare_active,
4875 .sync_request = raid10_sync_request,
4876 .quiesce = raid10_quiesce,
4877 .size = raid10_size,
4878 .resize = raid10_resize,
4879 .takeover = raid10_takeover,
4880 .check_reshape = raid10_check_reshape,
4881 .start_reshape = raid10_start_reshape,
4882 .finish_reshape = raid10_finish_reshape,
4883 .congested = raid10_congested,
4884};
4885
4886static int __init raid_init(void)
4887{
4888 return register_md_personality(&raid10_personality);
4889}
4890
4891static void raid_exit(void)
4892{
4893 unregister_md_personality(&raid10_personality);
4894}
4895
4896module_init(raid_init);
4897module_exit(raid_exit);
4898MODULE_LICENSE("GPL");
4899MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4900MODULE_ALIAS("md-personality-9"); /* RAID10 */
4901MODULE_ALIAS("md-raid10");
4902MODULE_ALIAS("md-level-10");
4903
4904module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);