blob: 62ebb1bc72be24cc32ecc775c5bd71603d44e5c6 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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 futher 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/raid/raid10.h>
22
23/*
24 * RAID10 provides a combination of RAID0 and RAID1 functionality.
25 * The layout of data is defined by
26 * chunk_size
27 * raid_disks
28 * near_copies (stored in low byte of layout)
29 * far_copies (stored in second byte of layout)
30 *
31 * The data to be stored is divided into chunks using chunksize.
32 * Each device is divided into far_copies sections.
33 * In each section, chunks are laid out in a style similar to raid0, but
34 * near_copies copies of each chunk is stored (each on a different drive).
35 * The starting device for each section is offset near_copies from the starting
36 * device of the previous section.
37 * Thus there are (near_copies*far_copies) of each chunk, and each is on a different
38 * drive.
39 * near_copies and far_copies must be at least one, and their product is at most
40 * raid_disks.
41 */
42
43/*
44 * Number of guaranteed r10bios in case of extreme VM load:
45 */
46#define NR_RAID10_BIOS 256
47
48static void unplug_slaves(mddev_t *mddev);
49
50static void * r10bio_pool_alloc(unsigned int __nocast gfp_flags, void *data)
51{
52 conf_t *conf = data;
53 r10bio_t *r10_bio;
54 int size = offsetof(struct r10bio_s, devs[conf->copies]);
55
56 /* allocate a r10bio with room for raid_disks entries in the bios array */
57 r10_bio = kmalloc(size, gfp_flags);
58 if (r10_bio)
59 memset(r10_bio, 0, size);
60 else
61 unplug_slaves(conf->mddev);
62
63 return r10_bio;
64}
65
66static void r10bio_pool_free(void *r10_bio, void *data)
67{
68 kfree(r10_bio);
69}
70
71#define RESYNC_BLOCK_SIZE (64*1024)
72//#define RESYNC_BLOCK_SIZE PAGE_SIZE
73#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
74#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
75#define RESYNC_WINDOW (2048*1024)
76
77/*
78 * When performing a resync, we need to read and compare, so
79 * we need as many pages are there are copies.
80 * When performing a recovery, we need 2 bios, one for read,
81 * one for write (we recover only one drive per r10buf)
82 *
83 */
84static void * r10buf_pool_alloc(unsigned int __nocast gfp_flags, void *data)
85{
86 conf_t *conf = data;
87 struct page *page;
88 r10bio_t *r10_bio;
89 struct bio *bio;
90 int i, j;
91 int nalloc;
92
93 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
94 if (!r10_bio) {
95 unplug_slaves(conf->mddev);
96 return NULL;
97 }
98
99 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
100 nalloc = conf->copies; /* resync */
101 else
102 nalloc = 2; /* recovery */
103
104 /*
105 * Allocate bios.
106 */
107 for (j = nalloc ; j-- ; ) {
108 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
109 if (!bio)
110 goto out_free_bio;
111 r10_bio->devs[j].bio = bio;
112 }
113 /*
114 * Allocate RESYNC_PAGES data pages and attach them
115 * where needed.
116 */
117 for (j = 0 ; j < nalloc; j++) {
118 bio = r10_bio->devs[j].bio;
119 for (i = 0; i < RESYNC_PAGES; i++) {
120 page = alloc_page(gfp_flags);
121 if (unlikely(!page))
122 goto out_free_pages;
123
124 bio->bi_io_vec[i].bv_page = page;
125 }
126 }
127
128 return r10_bio;
129
130out_free_pages:
131 for ( ; i > 0 ; i--)
132 __free_page(bio->bi_io_vec[i-1].bv_page);
133 while (j--)
134 for (i = 0; i < RESYNC_PAGES ; i++)
135 __free_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
136 j = -1;
137out_free_bio:
138 while ( ++j < nalloc )
139 bio_put(r10_bio->devs[j].bio);
140 r10bio_pool_free(r10_bio, conf);
141 return NULL;
142}
143
144static void r10buf_pool_free(void *__r10_bio, void *data)
145{
146 int i;
147 conf_t *conf = data;
148 r10bio_t *r10bio = __r10_bio;
149 int j;
150
151 for (j=0; j < conf->copies; j++) {
152 struct bio *bio = r10bio->devs[j].bio;
153 if (bio) {
154 for (i = 0; i < RESYNC_PAGES; i++) {
155 __free_page(bio->bi_io_vec[i].bv_page);
156 bio->bi_io_vec[i].bv_page = NULL;
157 }
158 bio_put(bio);
159 }
160 }
161 r10bio_pool_free(r10bio, conf);
162}
163
164static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
165{
166 int i;
167
168 for (i = 0; i < conf->copies; i++) {
169 struct bio **bio = & r10_bio->devs[i].bio;
170 if (*bio)
171 bio_put(*bio);
172 *bio = NULL;
173 }
174}
175
176static inline void free_r10bio(r10bio_t *r10_bio)
177{
178 unsigned long flags;
179
180 conf_t *conf = mddev_to_conf(r10_bio->mddev);
181
182 /*
183 * Wake up any possible resync thread that waits for the device
184 * to go idle.
185 */
186 spin_lock_irqsave(&conf->resync_lock, flags);
187 if (!--conf->nr_pending) {
188 wake_up(&conf->wait_idle);
189 wake_up(&conf->wait_resume);
190 }
191 spin_unlock_irqrestore(&conf->resync_lock, flags);
192
193 put_all_bios(conf, r10_bio);
194 mempool_free(r10_bio, conf->r10bio_pool);
195}
196
197static inline void put_buf(r10bio_t *r10_bio)
198{
199 conf_t *conf = mddev_to_conf(r10_bio->mddev);
200 unsigned long flags;
201
202 mempool_free(r10_bio, conf->r10buf_pool);
203
204 spin_lock_irqsave(&conf->resync_lock, flags);
205 if (!conf->barrier)
206 BUG();
207 --conf->barrier;
208 wake_up(&conf->wait_resume);
209 wake_up(&conf->wait_idle);
210
211 if (!--conf->nr_pending) {
212 wake_up(&conf->wait_idle);
213 wake_up(&conf->wait_resume);
214 }
215 spin_unlock_irqrestore(&conf->resync_lock, flags);
216}
217
218static void reschedule_retry(r10bio_t *r10_bio)
219{
220 unsigned long flags;
221 mddev_t *mddev = r10_bio->mddev;
222 conf_t *conf = mddev_to_conf(mddev);
223
224 spin_lock_irqsave(&conf->device_lock, flags);
225 list_add(&r10_bio->retry_list, &conf->retry_list);
226 spin_unlock_irqrestore(&conf->device_lock, flags);
227
228 md_wakeup_thread(mddev->thread);
229}
230
231/*
232 * raid_end_bio_io() is called when we have finished servicing a mirrored
233 * operation and are ready to return a success/failure code to the buffer
234 * cache layer.
235 */
236static void raid_end_bio_io(r10bio_t *r10_bio)
237{
238 struct bio *bio = r10_bio->master_bio;
239
240 bio_endio(bio, bio->bi_size,
241 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
242 free_r10bio(r10_bio);
243}
244
245/*
246 * Update disk head position estimator based on IRQ completion info.
247 */
248static inline void update_head_pos(int slot, r10bio_t *r10_bio)
249{
250 conf_t *conf = mddev_to_conf(r10_bio->mddev);
251
252 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
253 r10_bio->devs[slot].addr + (r10_bio->sectors);
254}
255
256static int raid10_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
257{
258 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
259 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
260 int slot, dev;
261 conf_t *conf = mddev_to_conf(r10_bio->mddev);
262
263 if (bio->bi_size)
264 return 1;
265
266 slot = r10_bio->read_slot;
267 dev = r10_bio->devs[slot].devnum;
268 /*
269 * this branch is our 'one mirror IO has finished' event handler:
270 */
271 if (!uptodate)
272 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
273 else
274 /*
275 * Set R10BIO_Uptodate in our master bio, so that
276 * we will return a good error code to the higher
277 * levels even if IO on some other mirrored buffer fails.
278 *
279 * The 'master' represents the composite IO operation to
280 * user-side. So if something waits for IO, then it will
281 * wait for the 'master' bio.
282 */
283 set_bit(R10BIO_Uptodate, &r10_bio->state);
284
285 update_head_pos(slot, r10_bio);
286
287 /*
288 * we have only one bio on the read side
289 */
290 if (uptodate)
291 raid_end_bio_io(r10_bio);
292 else {
293 /*
294 * oops, read error:
295 */
296 char b[BDEVNAME_SIZE];
297 if (printk_ratelimit())
298 printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
299 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
300 reschedule_retry(r10_bio);
301 }
302
303 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
304 return 0;
305}
306
307static int raid10_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
308{
309 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
310 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
311 int slot, dev;
312 conf_t *conf = mddev_to_conf(r10_bio->mddev);
313
314 if (bio->bi_size)
315 return 1;
316
317 for (slot = 0; slot < conf->copies; slot++)
318 if (r10_bio->devs[slot].bio == bio)
319 break;
320 dev = r10_bio->devs[slot].devnum;
321
322 /*
323 * this branch is our 'one mirror IO has finished' event handler:
324 */
325 if (!uptodate)
326 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
327 else
328 /*
329 * Set R10BIO_Uptodate in our master bio, so that
330 * we will return a good error code for to the higher
331 * levels even if IO on some other mirrored buffer fails.
332 *
333 * The 'master' represents the composite IO operation to
334 * user-side. So if something waits for IO, then it will
335 * wait for the 'master' bio.
336 */
337 set_bit(R10BIO_Uptodate, &r10_bio->state);
338
339 update_head_pos(slot, r10_bio);
340
341 /*
342 *
343 * Let's see if all mirrored write operations have finished
344 * already.
345 */
346 if (atomic_dec_and_test(&r10_bio->remaining)) {
347 md_write_end(r10_bio->mddev);
348 raid_end_bio_io(r10_bio);
349 }
350
351 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
352 return 0;
353}
354
355
356/*
357 * RAID10 layout manager
358 * Aswell as the chunksize and raid_disks count, there are two
359 * parameters: near_copies and far_copies.
360 * near_copies * far_copies must be <= raid_disks.
361 * Normally one of these will be 1.
362 * If both are 1, we get raid0.
363 * If near_copies == raid_disks, we get raid1.
364 *
365 * Chunks are layed out in raid0 style with near_copies copies of the
366 * first chunk, followed by near_copies copies of the next chunk and
367 * so on.
368 * If far_copies > 1, then after 1/far_copies of the array has been assigned
369 * as described above, we start again with a device offset of near_copies.
370 * So we effectively have another copy of the whole array further down all
371 * the drives, but with blocks on different drives.
372 * With this layout, and block is never stored twice on the one device.
373 *
374 * raid10_find_phys finds the sector offset of a given virtual sector
375 * on each device that it is on. If a block isn't on a device,
376 * that entry in the array is set to MaxSector.
377 *
378 * raid10_find_virt does the reverse mapping, from a device and a
379 * sector offset to a virtual address
380 */
381
382static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
383{
384 int n,f;
385 sector_t sector;
386 sector_t chunk;
387 sector_t stripe;
388 int dev;
389
390 int slot = 0;
391
392 /* now calculate first sector/dev */
393 chunk = r10bio->sector >> conf->chunk_shift;
394 sector = r10bio->sector & conf->chunk_mask;
395
396 chunk *= conf->near_copies;
397 stripe = chunk;
398 dev = sector_div(stripe, conf->raid_disks);
399
400 sector += stripe << conf->chunk_shift;
401
402 /* and calculate all the others */
403 for (n=0; n < conf->near_copies; n++) {
404 int d = dev;
405 sector_t s = sector;
406 r10bio->devs[slot].addr = sector;
407 r10bio->devs[slot].devnum = d;
408 slot++;
409
410 for (f = 1; f < conf->far_copies; f++) {
411 d += conf->near_copies;
412 if (d >= conf->raid_disks)
413 d -= conf->raid_disks;
414 s += conf->stride;
415 r10bio->devs[slot].devnum = d;
416 r10bio->devs[slot].addr = s;
417 slot++;
418 }
419 dev++;
420 if (dev >= conf->raid_disks) {
421 dev = 0;
422 sector += (conf->chunk_mask + 1);
423 }
424 }
425 BUG_ON(slot != conf->copies);
426}
427
428static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
429{
430 sector_t offset, chunk, vchunk;
431
432 while (sector > conf->stride) {
433 sector -= conf->stride;
434 if (dev < conf->near_copies)
435 dev += conf->raid_disks - conf->near_copies;
436 else
437 dev -= conf->near_copies;
438 }
439
440 offset = sector & conf->chunk_mask;
441 chunk = sector >> conf->chunk_shift;
442 vchunk = chunk * conf->raid_disks + dev;
443 sector_div(vchunk, conf->near_copies);
444 return (vchunk << conf->chunk_shift) + offset;
445}
446
447/**
448 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
449 * @q: request queue
450 * @bio: the buffer head that's been built up so far
451 * @biovec: the request that could be merged to it.
452 *
453 * Return amount of bytes we can accept at this offset
454 * If near_copies == raid_disk, there are no striping issues,
455 * but in that case, the function isn't called at all.
456 */
457static int raid10_mergeable_bvec(request_queue_t *q, struct bio *bio,
458 struct bio_vec *bio_vec)
459{
460 mddev_t *mddev = q->queuedata;
461 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
462 int max;
463 unsigned int chunk_sectors = mddev->chunk_size >> 9;
464 unsigned int bio_sectors = bio->bi_size >> 9;
465
466 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
467 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
468 if (max <= bio_vec->bv_len && bio_sectors == 0)
469 return bio_vec->bv_len;
470 else
471 return max;
472}
473
474/*
475 * This routine returns the disk from which the requested read should
476 * be done. There is a per-array 'next expected sequential IO' sector
477 * number - if this matches on the next IO then we use the last disk.
478 * There is also a per-disk 'last know head position' sector that is
479 * maintained from IRQ contexts, both the normal and the resync IO
480 * completion handlers update this position correctly. If there is no
481 * perfect sequential match then we pick the disk whose head is closest.
482 *
483 * If there are 2 mirrors in the same 2 devices, performance degrades
484 * because position is mirror, not device based.
485 *
486 * The rdev for the device selected will have nr_pending incremented.
487 */
488
489/*
490 * FIXME: possibly should rethink readbalancing and do it differently
491 * depending on near_copies / far_copies geometry.
492 */
493static int read_balance(conf_t *conf, r10bio_t *r10_bio)
494{
495 const unsigned long this_sector = r10_bio->sector;
496 int disk, slot, nslot;
497 const int sectors = r10_bio->sectors;
498 sector_t new_distance, current_distance;
499
500 raid10_find_phys(conf, r10_bio);
501 rcu_read_lock();
502 /*
503 * Check if we can balance. We can balance on the whole
504 * device if no resync is going on, or below the resync window.
505 * We take the first readable disk when above the resync window.
506 */
507 if (conf->mddev->recovery_cp < MaxSector
508 && (this_sector + sectors >= conf->next_resync)) {
509 /* make sure that disk is operational */
510 slot = 0;
511 disk = r10_bio->devs[slot].devnum;
512
513 while (!conf->mirrors[disk].rdev ||
514 !conf->mirrors[disk].rdev->in_sync) {
515 slot++;
516 if (slot == conf->copies) {
517 slot = 0;
518 disk = -1;
519 break;
520 }
521 disk = r10_bio->devs[slot].devnum;
522 }
523 goto rb_out;
524 }
525
526
527 /* make sure the disk is operational */
528 slot = 0;
529 disk = r10_bio->devs[slot].devnum;
530 while (!conf->mirrors[disk].rdev ||
531 !conf->mirrors[disk].rdev->in_sync) {
532 slot ++;
533 if (slot == conf->copies) {
534 disk = -1;
535 goto rb_out;
536 }
537 disk = r10_bio->devs[slot].devnum;
538 }
539
540
541 current_distance = abs(this_sector - conf->mirrors[disk].head_position);
542
543 /* Find the disk whose head is closest */
544
545 for (nslot = slot; nslot < conf->copies; nslot++) {
546 int ndisk = r10_bio->devs[nslot].devnum;
547
548
549 if (!conf->mirrors[ndisk].rdev ||
550 !conf->mirrors[ndisk].rdev->in_sync)
551 continue;
552
553 if (!atomic_read(&conf->mirrors[ndisk].rdev->nr_pending)) {
554 disk = ndisk;
555 slot = nslot;
556 break;
557 }
558 new_distance = abs(r10_bio->devs[nslot].addr -
559 conf->mirrors[ndisk].head_position);
560 if (new_distance < current_distance) {
561 current_distance = new_distance;
562 disk = ndisk;
563 slot = nslot;
564 }
565 }
566
567rb_out:
568 r10_bio->read_slot = slot;
569/* conf->next_seq_sect = this_sector + sectors;*/
570
571 if (disk >= 0 && conf->mirrors[disk].rdev)
572 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
573 rcu_read_unlock();
574
575 return disk;
576}
577
578static void unplug_slaves(mddev_t *mddev)
579{
580 conf_t *conf = mddev_to_conf(mddev);
581 int i;
582
583 rcu_read_lock();
584 for (i=0; i<mddev->raid_disks; i++) {
585 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
586 if (rdev && !rdev->faulty && atomic_read(&rdev->nr_pending)) {
587 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
588
589 atomic_inc(&rdev->nr_pending);
590 rcu_read_unlock();
591
592 if (r_queue->unplug_fn)
593 r_queue->unplug_fn(r_queue);
594
595 rdev_dec_pending(rdev, mddev);
596 rcu_read_lock();
597 }
598 }
599 rcu_read_unlock();
600}
601
602static void raid10_unplug(request_queue_t *q)
603{
604 unplug_slaves(q->queuedata);
605}
606
607static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
608 sector_t *error_sector)
609{
610 mddev_t *mddev = q->queuedata;
611 conf_t *conf = mddev_to_conf(mddev);
612 int i, ret = 0;
613
614 rcu_read_lock();
615 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
616 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
617 if (rdev && !rdev->faulty) {
618 struct block_device *bdev = rdev->bdev;
619 request_queue_t *r_queue = bdev_get_queue(bdev);
620
621 if (!r_queue->issue_flush_fn)
622 ret = -EOPNOTSUPP;
623 else {
624 atomic_inc(&rdev->nr_pending);
625 rcu_read_unlock();
626 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
627 error_sector);
628 rdev_dec_pending(rdev, mddev);
629 rcu_read_lock();
630 }
631 }
632 }
633 rcu_read_unlock();
634 return ret;
635}
636
637/*
638 * Throttle resync depth, so that we can both get proper overlapping of
639 * requests, but are still able to handle normal requests quickly.
640 */
641#define RESYNC_DEPTH 32
642
643static void device_barrier(conf_t *conf, sector_t sect)
644{
645 spin_lock_irq(&conf->resync_lock);
646 wait_event_lock_irq(conf->wait_idle, !waitqueue_active(&conf->wait_resume),
647 conf->resync_lock, unplug_slaves(conf->mddev));
648
649 if (!conf->barrier++) {
650 wait_event_lock_irq(conf->wait_idle, !conf->nr_pending,
651 conf->resync_lock, unplug_slaves(conf->mddev));
652 if (conf->nr_pending)
653 BUG();
654 }
655 wait_event_lock_irq(conf->wait_resume, conf->barrier < RESYNC_DEPTH,
656 conf->resync_lock, unplug_slaves(conf->mddev));
657 conf->next_resync = sect;
658 spin_unlock_irq(&conf->resync_lock);
659}
660
661static int make_request(request_queue_t *q, struct bio * bio)
662{
663 mddev_t *mddev = q->queuedata;
664 conf_t *conf = mddev_to_conf(mddev);
665 mirror_info_t *mirror;
666 r10bio_t *r10_bio;
667 struct bio *read_bio;
668 int i;
669 int chunk_sects = conf->chunk_mask + 1;
670
671 /* If this request crosses a chunk boundary, we need to
672 * split it. This will only happen for 1 PAGE (or less) requests.
673 */
674 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
675 > chunk_sects &&
676 conf->near_copies < conf->raid_disks)) {
677 struct bio_pair *bp;
678 /* Sanity check -- queue functions should prevent this happening */
679 if (bio->bi_vcnt != 1 ||
680 bio->bi_idx != 0)
681 goto bad_map;
682 /* This is a one page bio that upper layers
683 * refuse to split for us, so we need to split it.
684 */
685 bp = bio_split(bio, bio_split_pool,
686 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
687 if (make_request(q, &bp->bio1))
688 generic_make_request(&bp->bio1);
689 if (make_request(q, &bp->bio2))
690 generic_make_request(&bp->bio2);
691
692 bio_pair_release(bp);
693 return 0;
694 bad_map:
695 printk("raid10_make_request bug: can't convert block across chunks"
696 " or bigger than %dk %llu %d\n", chunk_sects/2,
697 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
698
699 bio_io_error(bio, bio->bi_size);
700 return 0;
701 }
702
NeilBrown3d310eb2005-06-21 17:17:26 -0700703 md_write_start(mddev, bio);
NeilBrown06d91a52005-06-21 17:17:12 -0700704
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705 /*
706 * Register the new request and wait if the reconstruction
707 * thread has put up a bar for new requests.
708 * Continue immediately if no resync is active currently.
709 */
710 spin_lock_irq(&conf->resync_lock);
711 wait_event_lock_irq(conf->wait_resume, !conf->barrier, conf->resync_lock, );
712 conf->nr_pending++;
713 spin_unlock_irq(&conf->resync_lock);
714
715 if (bio_data_dir(bio)==WRITE) {
716 disk_stat_inc(mddev->gendisk, writes);
717 disk_stat_add(mddev->gendisk, write_sectors, bio_sectors(bio));
718 } else {
719 disk_stat_inc(mddev->gendisk, reads);
720 disk_stat_add(mddev->gendisk, read_sectors, bio_sectors(bio));
721 }
722
723 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
724
725 r10_bio->master_bio = bio;
726 r10_bio->sectors = bio->bi_size >> 9;
727
728 r10_bio->mddev = mddev;
729 r10_bio->sector = bio->bi_sector;
730
731 if (bio_data_dir(bio) == READ) {
732 /*
733 * read balancing logic:
734 */
735 int disk = read_balance(conf, r10_bio);
736 int slot = r10_bio->read_slot;
737 if (disk < 0) {
738 raid_end_bio_io(r10_bio);
739 return 0;
740 }
741 mirror = conf->mirrors + disk;
742
743 read_bio = bio_clone(bio, GFP_NOIO);
744
745 r10_bio->devs[slot].bio = read_bio;
746
747 read_bio->bi_sector = r10_bio->devs[slot].addr +
748 mirror->rdev->data_offset;
749 read_bio->bi_bdev = mirror->rdev->bdev;
750 read_bio->bi_end_io = raid10_end_read_request;
751 read_bio->bi_rw = READ;
752 read_bio->bi_private = r10_bio;
753
754 generic_make_request(read_bio);
755 return 0;
756 }
757
758 /*
759 * WRITE:
760 */
761 /* first select target devices under spinlock and
762 * inc refcount on their rdev. Record them by setting
763 * bios[x] to bio
764 */
765 raid10_find_phys(conf, r10_bio);
766 rcu_read_lock();
767 for (i = 0; i < conf->copies; i++) {
768 int d = r10_bio->devs[i].devnum;
769 if (conf->mirrors[d].rdev &&
770 !conf->mirrors[d].rdev->faulty) {
771 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
772 r10_bio->devs[i].bio = bio;
773 } else
774 r10_bio->devs[i].bio = NULL;
775 }
776 rcu_read_unlock();
777
778 atomic_set(&r10_bio->remaining, 1);
NeilBrown06d91a52005-06-21 17:17:12 -0700779
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780 for (i = 0; i < conf->copies; i++) {
781 struct bio *mbio;
782 int d = r10_bio->devs[i].devnum;
783 if (!r10_bio->devs[i].bio)
784 continue;
785
786 mbio = bio_clone(bio, GFP_NOIO);
787 r10_bio->devs[i].bio = mbio;
788
789 mbio->bi_sector = r10_bio->devs[i].addr+
790 conf->mirrors[d].rdev->data_offset;
791 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
792 mbio->bi_end_io = raid10_end_write_request;
793 mbio->bi_rw = WRITE;
794 mbio->bi_private = r10_bio;
795
796 atomic_inc(&r10_bio->remaining);
797 generic_make_request(mbio);
798 }
799
800 if (atomic_dec_and_test(&r10_bio->remaining)) {
801 md_write_end(mddev);
802 raid_end_bio_io(r10_bio);
803 }
804
805 return 0;
806}
807
808static void status(struct seq_file *seq, mddev_t *mddev)
809{
810 conf_t *conf = mddev_to_conf(mddev);
811 int i;
812
813 if (conf->near_copies < conf->raid_disks)
814 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
815 if (conf->near_copies > 1)
816 seq_printf(seq, " %d near-copies", conf->near_copies);
817 if (conf->far_copies > 1)
818 seq_printf(seq, " %d far-copies", conf->far_copies);
819
820 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
821 conf->working_disks);
822 for (i = 0; i < conf->raid_disks; i++)
823 seq_printf(seq, "%s",
824 conf->mirrors[i].rdev &&
825 conf->mirrors[i].rdev->in_sync ? "U" : "_");
826 seq_printf(seq, "]");
827}
828
829static void error(mddev_t *mddev, mdk_rdev_t *rdev)
830{
831 char b[BDEVNAME_SIZE];
832 conf_t *conf = mddev_to_conf(mddev);
833
834 /*
835 * If it is not operational, then we have already marked it as dead
836 * else if it is the last working disks, ignore the error, let the
837 * next level up know.
838 * else mark the drive as failed
839 */
840 if (rdev->in_sync
841 && conf->working_disks == 1)
842 /*
843 * Don't fail the drive, just return an IO error.
844 * The test should really be more sophisticated than
845 * "working_disks == 1", but it isn't critical, and
846 * can wait until we do more sophisticated "is the drive
847 * really dead" tests...
848 */
849 return;
850 if (rdev->in_sync) {
851 mddev->degraded++;
852 conf->working_disks--;
853 /*
854 * if recovery is running, make sure it aborts.
855 */
856 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
857 }
858 rdev->in_sync = 0;
859 rdev->faulty = 1;
860 mddev->sb_dirty = 1;
861 printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
862 " Operation continuing on %d devices\n",
863 bdevname(rdev->bdev,b), conf->working_disks);
864}
865
866static void print_conf(conf_t *conf)
867{
868 int i;
869 mirror_info_t *tmp;
870
871 printk("RAID10 conf printout:\n");
872 if (!conf) {
873 printk("(!conf)\n");
874 return;
875 }
876 printk(" --- wd:%d rd:%d\n", conf->working_disks,
877 conf->raid_disks);
878
879 for (i = 0; i < conf->raid_disks; i++) {
880 char b[BDEVNAME_SIZE];
881 tmp = conf->mirrors + i;
882 if (tmp->rdev)
883 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
884 i, !tmp->rdev->in_sync, !tmp->rdev->faulty,
885 bdevname(tmp->rdev->bdev,b));
886 }
887}
888
889static void close_sync(conf_t *conf)
890{
891 spin_lock_irq(&conf->resync_lock);
892 wait_event_lock_irq(conf->wait_resume, !conf->barrier,
893 conf->resync_lock, unplug_slaves(conf->mddev));
894 spin_unlock_irq(&conf->resync_lock);
895
896 if (conf->barrier) BUG();
897 if (waitqueue_active(&conf->wait_idle)) BUG();
898
899 mempool_destroy(conf->r10buf_pool);
900 conf->r10buf_pool = NULL;
901}
902
903static int raid10_spare_active(mddev_t *mddev)
904{
905 int i;
906 conf_t *conf = mddev->private;
907 mirror_info_t *tmp;
908
909 /*
910 * Find all non-in_sync disks within the RAID10 configuration
911 * and mark them in_sync
912 */
913 for (i = 0; i < conf->raid_disks; i++) {
914 tmp = conf->mirrors + i;
915 if (tmp->rdev
916 && !tmp->rdev->faulty
917 && !tmp->rdev->in_sync) {
918 conf->working_disks++;
919 mddev->degraded--;
920 tmp->rdev->in_sync = 1;
921 }
922 }
923
924 print_conf(conf);
925 return 0;
926}
927
928
929static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
930{
931 conf_t *conf = mddev->private;
932 int found = 0;
933 int mirror;
934 mirror_info_t *p;
935
936 if (mddev->recovery_cp < MaxSector)
937 /* only hot-add to in-sync arrays, as recovery is
938 * very different from resync
939 */
940 return 0;
941
942 for (mirror=0; mirror < mddev->raid_disks; mirror++)
943 if ( !(p=conf->mirrors+mirror)->rdev) {
944
945 blk_queue_stack_limits(mddev->queue,
946 rdev->bdev->bd_disk->queue);
947 /* as we don't honour merge_bvec_fn, we must never risk
948 * violating it, so limit ->max_sector to one PAGE, as
949 * a one page request is never in violation.
950 */
951 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
952 mddev->queue->max_sectors > (PAGE_SIZE>>9))
953 mddev->queue->max_sectors = (PAGE_SIZE>>9);
954
955 p->head_position = 0;
956 rdev->raid_disk = mirror;
957 found = 1;
958 p->rdev = rdev;
959 break;
960 }
961
962 print_conf(conf);
963 return found;
964}
965
966static int raid10_remove_disk(mddev_t *mddev, int number)
967{
968 conf_t *conf = mddev->private;
969 int err = 0;
970 mdk_rdev_t *rdev;
971 mirror_info_t *p = conf->mirrors+ number;
972
973 print_conf(conf);
974 rdev = p->rdev;
975 if (rdev) {
976 if (rdev->in_sync ||
977 atomic_read(&rdev->nr_pending)) {
978 err = -EBUSY;
979 goto abort;
980 }
981 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -0700982 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983 if (atomic_read(&rdev->nr_pending)) {
984 /* lost the race, try later */
985 err = -EBUSY;
986 p->rdev = rdev;
987 }
988 }
989abort:
990
991 print_conf(conf);
992 return err;
993}
994
995
996static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
997{
998 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
999 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1000 conf_t *conf = mddev_to_conf(r10_bio->mddev);
1001 int i,d;
1002
1003 if (bio->bi_size)
1004 return 1;
1005
1006 for (i=0; i<conf->copies; i++)
1007 if (r10_bio->devs[i].bio == bio)
1008 break;
1009 if (i == conf->copies)
1010 BUG();
1011 update_head_pos(i, r10_bio);
1012 d = r10_bio->devs[i].devnum;
1013 if (!uptodate)
1014 md_error(r10_bio->mddev,
1015 conf->mirrors[d].rdev);
1016
1017 /* for reconstruct, we always reschedule after a read.
1018 * for resync, only after all reads
1019 */
1020 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1021 atomic_dec_and_test(&r10_bio->remaining)) {
1022 /* we have read all the blocks,
1023 * do the comparison in process context in raid10d
1024 */
1025 reschedule_retry(r10_bio);
1026 }
1027 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1028 return 0;
1029}
1030
1031static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1032{
1033 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1034 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1035 mddev_t *mddev = r10_bio->mddev;
1036 conf_t *conf = mddev_to_conf(mddev);
1037 int i,d;
1038
1039 if (bio->bi_size)
1040 return 1;
1041
1042 for (i = 0; i < conf->copies; i++)
1043 if (r10_bio->devs[i].bio == bio)
1044 break;
1045 d = r10_bio->devs[i].devnum;
1046
1047 if (!uptodate)
1048 md_error(mddev, conf->mirrors[d].rdev);
1049 update_head_pos(i, r10_bio);
1050
1051 while (atomic_dec_and_test(&r10_bio->remaining)) {
1052 if (r10_bio->master_bio == NULL) {
1053 /* the primary of several recovery bios */
1054 md_done_sync(mddev, r10_bio->sectors, 1);
1055 put_buf(r10_bio);
1056 break;
1057 } else {
1058 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1059 put_buf(r10_bio);
1060 r10_bio = r10_bio2;
1061 }
1062 }
1063 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1064 return 0;
1065}
1066
1067/*
1068 * Note: sync and recover and handled very differently for raid10
1069 * This code is for resync.
1070 * For resync, we read through virtual addresses and read all blocks.
1071 * If there is any error, we schedule a write. The lowest numbered
1072 * drive is authoritative.
1073 * However requests come for physical address, so we need to map.
1074 * For every physical address there are raid_disks/copies virtual addresses,
1075 * which is always are least one, but is not necessarly an integer.
1076 * This means that a physical address can span multiple chunks, so we may
1077 * have to submit multiple io requests for a single sync request.
1078 */
1079/*
1080 * We check if all blocks are in-sync and only write to blocks that
1081 * aren't in sync
1082 */
1083static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1084{
1085 conf_t *conf = mddev_to_conf(mddev);
1086 int i, first;
1087 struct bio *tbio, *fbio;
1088
1089 atomic_set(&r10_bio->remaining, 1);
1090
1091 /* find the first device with a block */
1092 for (i=0; i<conf->copies; i++)
1093 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1094 break;
1095
1096 if (i == conf->copies)
1097 goto done;
1098
1099 first = i;
1100 fbio = r10_bio->devs[i].bio;
1101
1102 /* now find blocks with errors */
1103 for (i=first+1 ; i < conf->copies ; i++) {
1104 int vcnt, j, d;
1105
1106 if (!test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1107 continue;
1108 /* We know that the bi_io_vec layout is the same for
1109 * both 'first' and 'i', so we just compare them.
1110 * All vec entries are PAGE_SIZE;
1111 */
1112 tbio = r10_bio->devs[i].bio;
1113 vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1114 for (j = 0; j < vcnt; j++)
1115 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1116 page_address(tbio->bi_io_vec[j].bv_page),
1117 PAGE_SIZE))
1118 break;
1119 if (j == vcnt)
1120 continue;
1121 /* Ok, we need to write this bio
1122 * First we need to fixup bv_offset, bv_len and
1123 * bi_vecs, as the read request might have corrupted these
1124 */
1125 tbio->bi_vcnt = vcnt;
1126 tbio->bi_size = r10_bio->sectors << 9;
1127 tbio->bi_idx = 0;
1128 tbio->bi_phys_segments = 0;
1129 tbio->bi_hw_segments = 0;
1130 tbio->bi_hw_front_size = 0;
1131 tbio->bi_hw_back_size = 0;
1132 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1133 tbio->bi_flags |= 1 << BIO_UPTODATE;
1134 tbio->bi_next = NULL;
1135 tbio->bi_rw = WRITE;
1136 tbio->bi_private = r10_bio;
1137 tbio->bi_sector = r10_bio->devs[i].addr;
1138
1139 for (j=0; j < vcnt ; j++) {
1140 tbio->bi_io_vec[j].bv_offset = 0;
1141 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1142
1143 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1144 page_address(fbio->bi_io_vec[j].bv_page),
1145 PAGE_SIZE);
1146 }
1147 tbio->bi_end_io = end_sync_write;
1148
1149 d = r10_bio->devs[i].devnum;
1150 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1151 atomic_inc(&r10_bio->remaining);
1152 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1153
1154 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1155 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1156 generic_make_request(tbio);
1157 }
1158
1159done:
1160 if (atomic_dec_and_test(&r10_bio->remaining)) {
1161 md_done_sync(mddev, r10_bio->sectors, 1);
1162 put_buf(r10_bio);
1163 }
1164}
1165
1166/*
1167 * Now for the recovery code.
1168 * Recovery happens across physical sectors.
1169 * We recover all non-is_sync drives by finding the virtual address of
1170 * each, and then choose a working drive that also has that virt address.
1171 * There is a separate r10_bio for each non-in_sync drive.
1172 * Only the first two slots are in use. The first for reading,
1173 * The second for writing.
1174 *
1175 */
1176
1177static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1178{
1179 conf_t *conf = mddev_to_conf(mddev);
1180 int i, d;
1181 struct bio *bio, *wbio;
1182
1183
1184 /* move the pages across to the second bio
1185 * and submit the write request
1186 */
1187 bio = r10_bio->devs[0].bio;
1188 wbio = r10_bio->devs[1].bio;
1189 for (i=0; i < wbio->bi_vcnt; i++) {
1190 struct page *p = bio->bi_io_vec[i].bv_page;
1191 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1192 wbio->bi_io_vec[i].bv_page = p;
1193 }
1194 d = r10_bio->devs[1].devnum;
1195
1196 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1197 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1198 generic_make_request(wbio);
1199}
1200
1201
1202/*
1203 * This is a kernel thread which:
1204 *
1205 * 1. Retries failed read operations on working mirrors.
1206 * 2. Updates the raid superblock when problems encounter.
1207 * 3. Performs writes following reads for array syncronising.
1208 */
1209
1210static void raid10d(mddev_t *mddev)
1211{
1212 r10bio_t *r10_bio;
1213 struct bio *bio;
1214 unsigned long flags;
1215 conf_t *conf = mddev_to_conf(mddev);
1216 struct list_head *head = &conf->retry_list;
1217 int unplug=0;
1218 mdk_rdev_t *rdev;
1219
1220 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221
1222 for (;;) {
1223 char b[BDEVNAME_SIZE];
1224 spin_lock_irqsave(&conf->device_lock, flags);
1225 if (list_empty(head))
1226 break;
1227 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1228 list_del(head->prev);
1229 spin_unlock_irqrestore(&conf->device_lock, flags);
1230
1231 mddev = r10_bio->mddev;
1232 conf = mddev_to_conf(mddev);
1233 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1234 sync_request_write(mddev, r10_bio);
1235 unplug = 1;
1236 } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1237 recovery_request_write(mddev, r10_bio);
1238 unplug = 1;
1239 } else {
1240 int mirror;
1241 bio = r10_bio->devs[r10_bio->read_slot].bio;
1242 r10_bio->devs[r10_bio->read_slot].bio = NULL;
1243 bio_put(bio);
1244 mirror = read_balance(conf, r10_bio);
1245 if (mirror == -1) {
1246 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1247 " read error for block %llu\n",
1248 bdevname(bio->bi_bdev,b),
1249 (unsigned long long)r10_bio->sector);
1250 raid_end_bio_io(r10_bio);
1251 } else {
1252 rdev = conf->mirrors[mirror].rdev;
1253 if (printk_ratelimit())
1254 printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1255 " another mirror\n",
1256 bdevname(rdev->bdev,b),
1257 (unsigned long long)r10_bio->sector);
1258 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1259 r10_bio->devs[r10_bio->read_slot].bio = bio;
1260 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1261 + rdev->data_offset;
1262 bio->bi_bdev = rdev->bdev;
1263 bio->bi_rw = READ;
1264 bio->bi_private = r10_bio;
1265 bio->bi_end_io = raid10_end_read_request;
1266 unplug = 1;
1267 generic_make_request(bio);
1268 }
1269 }
1270 }
1271 spin_unlock_irqrestore(&conf->device_lock, flags);
1272 if (unplug)
1273 unplug_slaves(mddev);
1274}
1275
1276
1277static int init_resync(conf_t *conf)
1278{
1279 int buffs;
1280
1281 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1282 if (conf->r10buf_pool)
1283 BUG();
1284 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1285 if (!conf->r10buf_pool)
1286 return -ENOMEM;
1287 conf->next_resync = 0;
1288 return 0;
1289}
1290
1291/*
1292 * perform a "sync" on one "block"
1293 *
1294 * We need to make sure that no normal I/O request - particularly write
1295 * requests - conflict with active sync requests.
1296 *
1297 * This is achieved by tracking pending requests and a 'barrier' concept
1298 * that can be installed to exclude normal IO requests.
1299 *
1300 * Resync and recovery are handled very differently.
1301 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1302 *
1303 * For resync, we iterate over virtual addresses, read all copies,
1304 * and update if there are differences. If only one copy is live,
1305 * skip it.
1306 * For recovery, we iterate over physical addresses, read a good
1307 * value for each non-in_sync drive, and over-write.
1308 *
1309 * So, for recovery we may have several outstanding complex requests for a
1310 * given address, one for each out-of-sync device. We model this by allocating
1311 * a number of r10_bio structures, one for each out-of-sync device.
1312 * As we setup these structures, we collect all bio's together into a list
1313 * which we then process collectively to add pages, and then process again
1314 * to pass to generic_make_request.
1315 *
1316 * The r10_bio structures are linked using a borrowed master_bio pointer.
1317 * This link is counted in ->remaining. When the r10_bio that points to NULL
1318 * has its remaining count decremented to 0, the whole complex operation
1319 * is complete.
1320 *
1321 */
1322
NeilBrown57afd892005-06-21 17:17:13 -07001323static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324{
1325 conf_t *conf = mddev_to_conf(mddev);
1326 r10bio_t *r10_bio;
1327 struct bio *biolist = NULL, *bio;
1328 sector_t max_sector, nr_sectors;
1329 int disk;
1330 int i;
1331
1332 sector_t sectors_skipped = 0;
1333 int chunks_skipped = 0;
1334
1335 if (!conf->r10buf_pool)
1336 if (init_resync(conf))
NeilBrown57afd892005-06-21 17:17:13 -07001337 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338
1339 skipped:
1340 max_sector = mddev->size << 1;
1341 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1342 max_sector = mddev->resync_max_sectors;
1343 if (sector_nr >= max_sector) {
1344 close_sync(conf);
NeilBrown57afd892005-06-21 17:17:13 -07001345 *skipped = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 return sectors_skipped;
1347 }
1348 if (chunks_skipped >= conf->raid_disks) {
1349 /* if there has been nothing to do on any drive,
1350 * then there is nothing to do at all..
1351 */
NeilBrown57afd892005-06-21 17:17:13 -07001352 *skipped = 1;
1353 return (max_sector - sector_nr) + sectors_skipped;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 }
1355
1356 /* make sure whole request will fit in a chunk - if chunks
1357 * are meaningful
1358 */
1359 if (conf->near_copies < conf->raid_disks &&
1360 max_sector > (sector_nr | conf->chunk_mask))
1361 max_sector = (sector_nr | conf->chunk_mask) + 1;
1362 /*
1363 * If there is non-resync activity waiting for us then
1364 * put in a delay to throttle resync.
1365 */
1366 if (!go_faster && waitqueue_active(&conf->wait_resume))
1367 msleep_interruptible(1000);
1368 device_barrier(conf, sector_nr + RESYNC_SECTORS);
1369
1370 /* Again, very different code for resync and recovery.
1371 * Both must result in an r10bio with a list of bios that
1372 * have bi_end_io, bi_sector, bi_bdev set,
1373 * and bi_private set to the r10bio.
1374 * For recovery, we may actually create several r10bios
1375 * with 2 bios in each, that correspond to the bios in the main one.
1376 * In this case, the subordinate r10bios link back through a
1377 * borrowed master_bio pointer, and the counter in the master
1378 * includes a ref from each subordinate.
1379 */
1380 /* First, we decide what to do and set ->bi_end_io
1381 * To end_sync_read if we want to read, and
1382 * end_sync_write if we will want to write.
1383 */
1384
1385 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1386 /* recovery... the complicated one */
1387 int i, j, k;
1388 r10_bio = NULL;
1389
1390 for (i=0 ; i<conf->raid_disks; i++)
1391 if (conf->mirrors[i].rdev &&
1392 !conf->mirrors[i].rdev->in_sync) {
1393 /* want to reconstruct this device */
1394 r10bio_t *rb2 = r10_bio;
1395
1396 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1397 spin_lock_irq(&conf->resync_lock);
1398 conf->nr_pending++;
1399 if (rb2) conf->barrier++;
1400 spin_unlock_irq(&conf->resync_lock);
1401 atomic_set(&r10_bio->remaining, 0);
1402
1403 r10_bio->master_bio = (struct bio*)rb2;
1404 if (rb2)
1405 atomic_inc(&rb2->remaining);
1406 r10_bio->mddev = mddev;
1407 set_bit(R10BIO_IsRecover, &r10_bio->state);
1408 r10_bio->sector = raid10_find_virt(conf, sector_nr, i);
1409 raid10_find_phys(conf, r10_bio);
1410 for (j=0; j<conf->copies;j++) {
1411 int d = r10_bio->devs[j].devnum;
1412 if (conf->mirrors[d].rdev &&
1413 conf->mirrors[d].rdev->in_sync) {
1414 /* This is where we read from */
1415 bio = r10_bio->devs[0].bio;
1416 bio->bi_next = biolist;
1417 biolist = bio;
1418 bio->bi_private = r10_bio;
1419 bio->bi_end_io = end_sync_read;
1420 bio->bi_rw = 0;
1421 bio->bi_sector = r10_bio->devs[j].addr +
1422 conf->mirrors[d].rdev->data_offset;
1423 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1424 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1425 atomic_inc(&r10_bio->remaining);
1426 /* and we write to 'i' */
1427
1428 for (k=0; k<conf->copies; k++)
1429 if (r10_bio->devs[k].devnum == i)
1430 break;
1431 bio = r10_bio->devs[1].bio;
1432 bio->bi_next = biolist;
1433 biolist = bio;
1434 bio->bi_private = r10_bio;
1435 bio->bi_end_io = end_sync_write;
1436 bio->bi_rw = 1;
1437 bio->bi_sector = r10_bio->devs[k].addr +
1438 conf->mirrors[i].rdev->data_offset;
1439 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1440
1441 r10_bio->devs[0].devnum = d;
1442 r10_bio->devs[1].devnum = i;
1443
1444 break;
1445 }
1446 }
1447 if (j == conf->copies) {
1448 BUG();
1449 }
1450 }
1451 if (biolist == NULL) {
1452 while (r10_bio) {
1453 r10bio_t *rb2 = r10_bio;
1454 r10_bio = (r10bio_t*) rb2->master_bio;
1455 rb2->master_bio = NULL;
1456 put_buf(rb2);
1457 }
1458 goto giveup;
1459 }
1460 } else {
1461 /* resync. Schedule a read for every block at this virt offset */
1462 int count = 0;
1463 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1464
1465 spin_lock_irq(&conf->resync_lock);
1466 conf->nr_pending++;
1467 spin_unlock_irq(&conf->resync_lock);
1468
1469 r10_bio->mddev = mddev;
1470 atomic_set(&r10_bio->remaining, 0);
1471
1472 r10_bio->master_bio = NULL;
1473 r10_bio->sector = sector_nr;
1474 set_bit(R10BIO_IsSync, &r10_bio->state);
1475 raid10_find_phys(conf, r10_bio);
1476 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1477
1478 for (i=0; i<conf->copies; i++) {
1479 int d = r10_bio->devs[i].devnum;
1480 bio = r10_bio->devs[i].bio;
1481 bio->bi_end_io = NULL;
1482 if (conf->mirrors[d].rdev == NULL ||
1483 conf->mirrors[d].rdev->faulty)
1484 continue;
1485 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1486 atomic_inc(&r10_bio->remaining);
1487 bio->bi_next = biolist;
1488 biolist = bio;
1489 bio->bi_private = r10_bio;
1490 bio->bi_end_io = end_sync_read;
1491 bio->bi_rw = 0;
1492 bio->bi_sector = r10_bio->devs[i].addr +
1493 conf->mirrors[d].rdev->data_offset;
1494 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1495 count++;
1496 }
1497
1498 if (count < 2) {
1499 for (i=0; i<conf->copies; i++) {
1500 int d = r10_bio->devs[i].devnum;
1501 if (r10_bio->devs[i].bio->bi_end_io)
1502 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1503 }
1504 put_buf(r10_bio);
1505 biolist = NULL;
1506 goto giveup;
1507 }
1508 }
1509
1510 for (bio = biolist; bio ; bio=bio->bi_next) {
1511
1512 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1513 if (bio->bi_end_io)
1514 bio->bi_flags |= 1 << BIO_UPTODATE;
1515 bio->bi_vcnt = 0;
1516 bio->bi_idx = 0;
1517 bio->bi_phys_segments = 0;
1518 bio->bi_hw_segments = 0;
1519 bio->bi_size = 0;
1520 }
1521
1522 nr_sectors = 0;
1523 do {
1524 struct page *page;
1525 int len = PAGE_SIZE;
1526 disk = 0;
1527 if (sector_nr + (len>>9) > max_sector)
1528 len = (max_sector - sector_nr) << 9;
1529 if (len == 0)
1530 break;
1531 for (bio= biolist ; bio ; bio=bio->bi_next) {
1532 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1533 if (bio_add_page(bio, page, len, 0) == 0) {
1534 /* stop here */
1535 struct bio *bio2;
1536 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1537 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1538 /* remove last page from this bio */
1539 bio2->bi_vcnt--;
1540 bio2->bi_size -= len;
1541 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1542 }
1543 goto bio_full;
1544 }
1545 disk = i;
1546 }
1547 nr_sectors += len>>9;
1548 sector_nr += len>>9;
1549 } while (biolist->bi_vcnt < RESYNC_PAGES);
1550 bio_full:
1551 r10_bio->sectors = nr_sectors;
1552
1553 while (biolist) {
1554 bio = biolist;
1555 biolist = biolist->bi_next;
1556
1557 bio->bi_next = NULL;
1558 r10_bio = bio->bi_private;
1559 r10_bio->sectors = nr_sectors;
1560
1561 if (bio->bi_end_io == end_sync_read) {
1562 md_sync_acct(bio->bi_bdev, nr_sectors);
1563 generic_make_request(bio);
1564 }
1565 }
1566
NeilBrown57afd892005-06-21 17:17:13 -07001567 if (sectors_skipped)
1568 /* pretend they weren't skipped, it makes
1569 * no important difference in this case
1570 */
1571 md_done_sync(mddev, sectors_skipped, 1);
1572
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573 return sectors_skipped + nr_sectors;
1574 giveup:
1575 /* There is nowhere to write, so all non-sync
1576 * drives must be failed, so try the next chunk...
1577 */
1578 {
NeilBrown57afd892005-06-21 17:17:13 -07001579 sector_t sec = max_sector - sector_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 sectors_skipped += sec;
1581 chunks_skipped ++;
1582 sector_nr = max_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 goto skipped;
1584 }
1585}
1586
1587static int run(mddev_t *mddev)
1588{
1589 conf_t *conf;
1590 int i, disk_idx;
1591 mirror_info_t *disk;
1592 mdk_rdev_t *rdev;
1593 struct list_head *tmp;
1594 int nc, fc;
1595 sector_t stride, size;
1596
1597 if (mddev->level != 10) {
1598 printk(KERN_ERR "raid10: %s: raid level not set correctly... (%d)\n",
1599 mdname(mddev), mddev->level);
1600 goto out;
1601 }
1602 nc = mddev->layout & 255;
1603 fc = (mddev->layout >> 8) & 255;
1604 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1605 (mddev->layout >> 16)) {
1606 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1607 mdname(mddev), mddev->layout);
1608 goto out;
1609 }
1610 /*
1611 * copy the already verified devices into our private RAID10
1612 * bookkeeping area. [whatever we allocate in run(),
1613 * should be freed in stop()]
1614 */
1615 conf = kmalloc(sizeof(conf_t), GFP_KERNEL);
1616 mddev->private = conf;
1617 if (!conf) {
1618 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1619 mdname(mddev));
1620 goto out;
1621 }
1622 memset(conf, 0, sizeof(*conf));
1623 conf->mirrors = kmalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1624 GFP_KERNEL);
1625 if (!conf->mirrors) {
1626 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1627 mdname(mddev));
1628 goto out_free_conf;
1629 }
1630 memset(conf->mirrors, 0, sizeof(struct mirror_info)*mddev->raid_disks);
1631
1632 conf->near_copies = nc;
1633 conf->far_copies = fc;
1634 conf->copies = nc*fc;
1635 conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
1636 conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
1637 stride = mddev->size >> (conf->chunk_shift-1);
1638 sector_div(stride, fc);
1639 conf->stride = stride << conf->chunk_shift;
1640
1641 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
1642 r10bio_pool_free, conf);
1643 if (!conf->r10bio_pool) {
1644 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1645 mdname(mddev));
1646 goto out_free_conf;
1647 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648
1649 ITERATE_RDEV(mddev, rdev, tmp) {
1650 disk_idx = rdev->raid_disk;
1651 if (disk_idx >= mddev->raid_disks
1652 || disk_idx < 0)
1653 continue;
1654 disk = conf->mirrors + disk_idx;
1655
1656 disk->rdev = rdev;
1657
1658 blk_queue_stack_limits(mddev->queue,
1659 rdev->bdev->bd_disk->queue);
1660 /* as we don't honour merge_bvec_fn, we must never risk
1661 * violating it, so limit ->max_sector to one PAGE, as
1662 * a one page request is never in violation.
1663 */
1664 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1665 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1666 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1667
1668 disk->head_position = 0;
1669 if (!rdev->faulty && rdev->in_sync)
1670 conf->working_disks++;
1671 }
1672 conf->raid_disks = mddev->raid_disks;
1673 conf->mddev = mddev;
1674 spin_lock_init(&conf->device_lock);
1675 INIT_LIST_HEAD(&conf->retry_list);
1676
1677 spin_lock_init(&conf->resync_lock);
1678 init_waitqueue_head(&conf->wait_idle);
1679 init_waitqueue_head(&conf->wait_resume);
1680
1681 if (!conf->working_disks) {
1682 printk(KERN_ERR "raid10: no operational mirrors for %s\n",
1683 mdname(mddev));
1684 goto out_free_conf;
1685 }
1686
1687 mddev->degraded = 0;
1688 for (i = 0; i < conf->raid_disks; i++) {
1689
1690 disk = conf->mirrors + i;
1691
1692 if (!disk->rdev) {
1693 disk->head_position = 0;
1694 mddev->degraded++;
1695 }
1696 }
1697
1698
1699 mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
1700 if (!mddev->thread) {
1701 printk(KERN_ERR
1702 "raid10: couldn't allocate thread for %s\n",
1703 mdname(mddev));
1704 goto out_free_conf;
1705 }
1706
1707 printk(KERN_INFO
1708 "raid10: raid set %s active with %d out of %d devices\n",
1709 mdname(mddev), mddev->raid_disks - mddev->degraded,
1710 mddev->raid_disks);
1711 /*
1712 * Ok, everything is just fine now
1713 */
1714 size = conf->stride * conf->raid_disks;
1715 sector_div(size, conf->near_copies);
1716 mddev->array_size = size/2;
1717 mddev->resync_max_sectors = size;
1718
NeilBrown7a5febe2005-05-16 21:53:16 -07001719 mddev->queue->unplug_fn = raid10_unplug;
1720 mddev->queue->issue_flush_fn = raid10_issue_flush;
1721
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722 /* Calculate max read-ahead size.
1723 * We need to readahead at least twice a whole stripe....
1724 * maybe...
1725 */
1726 {
1727 int stripe = conf->raid_disks * mddev->chunk_size / PAGE_CACHE_SIZE;
1728 stripe /= conf->near_copies;
1729 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
1730 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
1731 }
1732
1733 if (conf->near_copies < mddev->raid_disks)
1734 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
1735 return 0;
1736
1737out_free_conf:
1738 if (conf->r10bio_pool)
1739 mempool_destroy(conf->r10bio_pool);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07001740 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 kfree(conf);
1742 mddev->private = NULL;
1743out:
1744 return -EIO;
1745}
1746
1747static int stop(mddev_t *mddev)
1748{
1749 conf_t *conf = mddev_to_conf(mddev);
1750
1751 md_unregister_thread(mddev->thread);
1752 mddev->thread = NULL;
1753 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
1754 if (conf->r10bio_pool)
1755 mempool_destroy(conf->r10bio_pool);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07001756 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757 kfree(conf);
1758 mddev->private = NULL;
1759 return 0;
1760}
1761
1762
1763static mdk_personality_t raid10_personality =
1764{
1765 .name = "raid10",
1766 .owner = THIS_MODULE,
1767 .make_request = make_request,
1768 .run = run,
1769 .stop = stop,
1770 .status = status,
1771 .error_handler = error,
1772 .hot_add_disk = raid10_add_disk,
1773 .hot_remove_disk= raid10_remove_disk,
1774 .spare_active = raid10_spare_active,
1775 .sync_request = sync_request,
1776};
1777
1778static int __init raid_init(void)
1779{
1780 return register_md_personality(RAID10, &raid10_personality);
1781}
1782
1783static void raid_exit(void)
1784{
1785 unregister_md_personality(RAID10);
1786}
1787
1788module_init(raid_init);
1789module_exit(raid_exit);
1790MODULE_LICENSE("GPL");
1791MODULE_ALIAS("md-personality-9"); /* RAID10 */