blob: 201dc7168a5febe69572ede5dcd09797650dba1f [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
NeilBrown6cce3b22006-01-06 00:20:16 -080021#include "dm-bio-list.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070022#include <linux/raid/raid10.h>
NeilBrown6cce3b22006-01-06 00:20:16 -080023#include <linux/raid/bitmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024
25/*
26 * RAID10 provides a combination of RAID0 and RAID1 functionality.
27 * The layout of data is defined by
28 * chunk_size
29 * raid_disks
30 * near_copies (stored in low byte of layout)
31 * far_copies (stored in second byte of layout)
32 *
33 * The data to be stored is divided into chunks using chunksize.
34 * Each device is divided into far_copies sections.
35 * In each section, chunks are laid out in a style similar to raid0, but
36 * near_copies copies of each chunk is stored (each on a different drive).
37 * The starting device for each section is offset near_copies from the starting
38 * device of the previous section.
39 * Thus there are (near_copies*far_copies) of each chunk, and each is on a different
40 * drive.
41 * near_copies and far_copies must be at least one, and their product is at most
42 * raid_disks.
43 */
44
45/*
46 * Number of guaranteed r10bios in case of extreme VM load:
47 */
48#define NR_RAID10_BIOS 256
49
50static void unplug_slaves(mddev_t *mddev);
51
NeilBrown0a27ec92006-01-06 00:20:13 -080052static void allow_barrier(conf_t *conf);
53static void lower_barrier(conf_t *conf);
54
Al Virodd0fc662005-10-07 07:46:04 +010055static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070056{
57 conf_t *conf = data;
58 r10bio_t *r10_bio;
59 int size = offsetof(struct r10bio_s, devs[conf->copies]);
60
61 /* allocate a r10bio with room for raid_disks entries in the bios array */
NeilBrown9ffae0c2006-01-06 00:20:32 -080062 r10_bio = kzalloc(size, gfp_flags);
63 if (!r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -070064 unplug_slaves(conf->mddev);
65
66 return r10_bio;
67}
68
69static void r10bio_pool_free(void *r10_bio, void *data)
70{
71 kfree(r10_bio);
72}
73
74#define RESYNC_BLOCK_SIZE (64*1024)
75//#define RESYNC_BLOCK_SIZE PAGE_SIZE
76#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
77#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
78#define RESYNC_WINDOW (2048*1024)
79
80/*
81 * When performing a resync, we need to read and compare, so
82 * we need as many pages are there are copies.
83 * When performing a recovery, we need 2 bios, one for read,
84 * one for write (we recover only one drive per r10buf)
85 *
86 */
Al Virodd0fc662005-10-07 07:46:04 +010087static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070088{
89 conf_t *conf = data;
90 struct page *page;
91 r10bio_t *r10_bio;
92 struct bio *bio;
93 int i, j;
94 int nalloc;
95
96 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
97 if (!r10_bio) {
98 unplug_slaves(conf->mddev);
99 return NULL;
100 }
101
102 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
103 nalloc = conf->copies; /* resync */
104 else
105 nalloc = 2; /* recovery */
106
107 /*
108 * Allocate bios.
109 */
110 for (j = nalloc ; j-- ; ) {
111 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
112 if (!bio)
113 goto out_free_bio;
114 r10_bio->devs[j].bio = bio;
115 }
116 /*
117 * Allocate RESYNC_PAGES data pages and attach them
118 * where needed.
119 */
120 for (j = 0 ; j < nalloc; j++) {
121 bio = r10_bio->devs[j].bio;
122 for (i = 0; i < RESYNC_PAGES; i++) {
123 page = alloc_page(gfp_flags);
124 if (unlikely(!page))
125 goto out_free_pages;
126
127 bio->bi_io_vec[i].bv_page = page;
128 }
129 }
130
131 return r10_bio;
132
133out_free_pages:
134 for ( ; i > 0 ; i--)
NeilBrown1345b1d2006-01-06 00:20:40 -0800135 safe_put_page(bio->bi_io_vec[i-1].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136 while (j--)
137 for (i = 0; i < RESYNC_PAGES ; i++)
NeilBrown1345b1d2006-01-06 00:20:40 -0800138 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139 j = -1;
140out_free_bio:
141 while ( ++j < nalloc )
142 bio_put(r10_bio->devs[j].bio);
143 r10bio_pool_free(r10_bio, conf);
144 return NULL;
145}
146
147static void r10buf_pool_free(void *__r10_bio, void *data)
148{
149 int i;
150 conf_t *conf = data;
151 r10bio_t *r10bio = __r10_bio;
152 int j;
153
154 for (j=0; j < conf->copies; j++) {
155 struct bio *bio = r10bio->devs[j].bio;
156 if (bio) {
157 for (i = 0; i < RESYNC_PAGES; i++) {
NeilBrown1345b1d2006-01-06 00:20:40 -0800158 safe_put_page(bio->bi_io_vec[i].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159 bio->bi_io_vec[i].bv_page = NULL;
160 }
161 bio_put(bio);
162 }
163 }
164 r10bio_pool_free(r10bio, conf);
165}
166
167static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
168{
169 int i;
170
171 for (i = 0; i < conf->copies; i++) {
172 struct bio **bio = & r10_bio->devs[i].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -0800173 if (*bio && *bio != IO_BLOCKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174 bio_put(*bio);
175 *bio = NULL;
176 }
177}
178
179static inline void free_r10bio(r10bio_t *r10_bio)
180{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181 conf_t *conf = mddev_to_conf(r10_bio->mddev);
182
183 /*
184 * Wake up any possible resync thread that waits for the device
185 * to go idle.
186 */
NeilBrown0a27ec92006-01-06 00:20:13 -0800187 allow_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188
189 put_all_bios(conf, r10_bio);
190 mempool_free(r10_bio, conf->r10bio_pool);
191}
192
193static inline void put_buf(r10bio_t *r10_bio)
194{
195 conf_t *conf = mddev_to_conf(r10_bio->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196
197 mempool_free(r10_bio, conf->r10buf_pool);
198
NeilBrown0a27ec92006-01-06 00:20:13 -0800199 lower_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200}
201
202static void reschedule_retry(r10bio_t *r10_bio)
203{
204 unsigned long flags;
205 mddev_t *mddev = r10_bio->mddev;
206 conf_t *conf = mddev_to_conf(mddev);
207
208 spin_lock_irqsave(&conf->device_lock, flags);
209 list_add(&r10_bio->retry_list, &conf->retry_list);
NeilBrown4443ae12006-01-06 00:20:28 -0800210 conf->nr_queued ++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211 spin_unlock_irqrestore(&conf->device_lock, flags);
212
213 md_wakeup_thread(mddev->thread);
214}
215
216/*
217 * raid_end_bio_io() is called when we have finished servicing a mirrored
218 * operation and are ready to return a success/failure code to the buffer
219 * cache layer.
220 */
221static void raid_end_bio_io(r10bio_t *r10_bio)
222{
223 struct bio *bio = r10_bio->master_bio;
224
225 bio_endio(bio, bio->bi_size,
226 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
227 free_r10bio(r10_bio);
228}
229
230/*
231 * Update disk head position estimator based on IRQ completion info.
232 */
233static inline void update_head_pos(int slot, r10bio_t *r10_bio)
234{
235 conf_t *conf = mddev_to_conf(r10_bio->mddev);
236
237 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
238 r10_bio->devs[slot].addr + (r10_bio->sectors);
239}
240
241static int raid10_end_read_request(struct bio *bio, unsigned int bytes_done, int error)
242{
243 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
244 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
245 int slot, dev;
246 conf_t *conf = mddev_to_conf(r10_bio->mddev);
247
248 if (bio->bi_size)
249 return 1;
250
251 slot = r10_bio->read_slot;
252 dev = r10_bio->devs[slot].devnum;
253 /*
254 * this branch is our 'one mirror IO has finished' event handler:
255 */
NeilBrown4443ae12006-01-06 00:20:28 -0800256 update_head_pos(slot, r10_bio);
257
258 if (uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259 /*
260 * Set R10BIO_Uptodate in our master bio, so that
261 * we will return a good error code to the higher
262 * levels even if IO on some other mirrored buffer fails.
263 *
264 * The 'master' represents the composite IO operation to
265 * user-side. So if something waits for IO, then it will
266 * wait for the 'master' bio.
267 */
268 set_bit(R10BIO_Uptodate, &r10_bio->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700269 raid_end_bio_io(r10_bio);
NeilBrown4443ae12006-01-06 00:20:28 -0800270 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271 /*
272 * oops, read error:
273 */
274 char b[BDEVNAME_SIZE];
275 if (printk_ratelimit())
276 printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
277 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
278 reschedule_retry(r10_bio);
279 }
280
281 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
282 return 0;
283}
284
285static int raid10_end_write_request(struct bio *bio, unsigned int bytes_done, int error)
286{
287 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
288 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
289 int slot, dev;
290 conf_t *conf = mddev_to_conf(r10_bio->mddev);
291
292 if (bio->bi_size)
293 return 1;
294
295 for (slot = 0; slot < conf->copies; slot++)
296 if (r10_bio->devs[slot].bio == bio)
297 break;
298 dev = r10_bio->devs[slot].devnum;
299
300 /*
301 * this branch is our 'one mirror IO has finished' event handler:
302 */
NeilBrown6cce3b22006-01-06 00:20:16 -0800303 if (!uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
NeilBrown6cce3b22006-01-06 00:20:16 -0800305 /* an I/O failed, we can't clear the bitmap */
306 set_bit(R10BIO_Degraded, &r10_bio->state);
307 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308 /*
309 * Set R10BIO_Uptodate in our master bio, so that
310 * we will return a good error code for to the higher
311 * levels even if IO on some other mirrored buffer fails.
312 *
313 * The 'master' represents the composite IO operation to
314 * user-side. So if something waits for IO, then it will
315 * wait for the 'master' bio.
316 */
317 set_bit(R10BIO_Uptodate, &r10_bio->state);
318
319 update_head_pos(slot, r10_bio);
320
321 /*
322 *
323 * Let's see if all mirrored write operations have finished
324 * already.
325 */
326 if (atomic_dec_and_test(&r10_bio->remaining)) {
NeilBrown6cce3b22006-01-06 00:20:16 -0800327 /* clear the bitmap if all writes complete successfully */
328 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
329 r10_bio->sectors,
330 !test_bit(R10BIO_Degraded, &r10_bio->state),
331 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332 md_write_end(r10_bio->mddev);
333 raid_end_bio_io(r10_bio);
334 }
335
336 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
337 return 0;
338}
339
340
341/*
342 * RAID10 layout manager
343 * Aswell as the chunksize and raid_disks count, there are two
344 * parameters: near_copies and far_copies.
345 * near_copies * far_copies must be <= raid_disks.
346 * Normally one of these will be 1.
347 * If both are 1, we get raid0.
348 * If near_copies == raid_disks, we get raid1.
349 *
350 * Chunks are layed out in raid0 style with near_copies copies of the
351 * first chunk, followed by near_copies copies of the next chunk and
352 * so on.
353 * If far_copies > 1, then after 1/far_copies of the array has been assigned
354 * as described above, we start again with a device offset of near_copies.
355 * So we effectively have another copy of the whole array further down all
356 * the drives, but with blocks on different drives.
357 * With this layout, and block is never stored twice on the one device.
358 *
359 * raid10_find_phys finds the sector offset of a given virtual sector
360 * on each device that it is on. If a block isn't on a device,
361 * that entry in the array is set to MaxSector.
362 *
363 * raid10_find_virt does the reverse mapping, from a device and a
364 * sector offset to a virtual address
365 */
366
367static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
368{
369 int n,f;
370 sector_t sector;
371 sector_t chunk;
372 sector_t stripe;
373 int dev;
374
375 int slot = 0;
376
377 /* now calculate first sector/dev */
378 chunk = r10bio->sector >> conf->chunk_shift;
379 sector = r10bio->sector & conf->chunk_mask;
380
381 chunk *= conf->near_copies;
382 stripe = chunk;
383 dev = sector_div(stripe, conf->raid_disks);
384
385 sector += stripe << conf->chunk_shift;
386
387 /* and calculate all the others */
388 for (n=0; n < conf->near_copies; n++) {
389 int d = dev;
390 sector_t s = sector;
391 r10bio->devs[slot].addr = sector;
392 r10bio->devs[slot].devnum = d;
393 slot++;
394
395 for (f = 1; f < conf->far_copies; f++) {
396 d += conf->near_copies;
397 if (d >= conf->raid_disks)
398 d -= conf->raid_disks;
399 s += conf->stride;
400 r10bio->devs[slot].devnum = d;
401 r10bio->devs[slot].addr = s;
402 slot++;
403 }
404 dev++;
405 if (dev >= conf->raid_disks) {
406 dev = 0;
407 sector += (conf->chunk_mask + 1);
408 }
409 }
410 BUG_ON(slot != conf->copies);
411}
412
413static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
414{
415 sector_t offset, chunk, vchunk;
416
417 while (sector > conf->stride) {
418 sector -= conf->stride;
419 if (dev < conf->near_copies)
420 dev += conf->raid_disks - conf->near_copies;
421 else
422 dev -= conf->near_copies;
423 }
424
425 offset = sector & conf->chunk_mask;
426 chunk = sector >> conf->chunk_shift;
427 vchunk = chunk * conf->raid_disks + dev;
428 sector_div(vchunk, conf->near_copies);
429 return (vchunk << conf->chunk_shift) + offset;
430}
431
432/**
433 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
434 * @q: request queue
435 * @bio: the buffer head that's been built up so far
436 * @biovec: the request that could be merged to it.
437 *
438 * Return amount of bytes we can accept at this offset
439 * If near_copies == raid_disk, there are no striping issues,
440 * but in that case, the function isn't called at all.
441 */
442static int raid10_mergeable_bvec(request_queue_t *q, struct bio *bio,
443 struct bio_vec *bio_vec)
444{
445 mddev_t *mddev = q->queuedata;
446 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
447 int max;
448 unsigned int chunk_sectors = mddev->chunk_size >> 9;
449 unsigned int bio_sectors = bio->bi_size >> 9;
450
451 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
452 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
453 if (max <= bio_vec->bv_len && bio_sectors == 0)
454 return bio_vec->bv_len;
455 else
456 return max;
457}
458
459/*
460 * This routine returns the disk from which the requested read should
461 * be done. There is a per-array 'next expected sequential IO' sector
462 * number - if this matches on the next IO then we use the last disk.
463 * There is also a per-disk 'last know head position' sector that is
464 * maintained from IRQ contexts, both the normal and the resync IO
465 * completion handlers update this position correctly. If there is no
466 * perfect sequential match then we pick the disk whose head is closest.
467 *
468 * If there are 2 mirrors in the same 2 devices, performance degrades
469 * because position is mirror, not device based.
470 *
471 * The rdev for the device selected will have nr_pending incremented.
472 */
473
474/*
475 * FIXME: possibly should rethink readbalancing and do it differently
476 * depending on near_copies / far_copies geometry.
477 */
478static int read_balance(conf_t *conf, r10bio_t *r10_bio)
479{
480 const unsigned long this_sector = r10_bio->sector;
481 int disk, slot, nslot;
482 const int sectors = r10_bio->sectors;
483 sector_t new_distance, current_distance;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800484 mdk_rdev_t *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485
486 raid10_find_phys(conf, r10_bio);
487 rcu_read_lock();
488 /*
489 * Check if we can balance. We can balance on the whole
NeilBrown6cce3b22006-01-06 00:20:16 -0800490 * device if no resync is going on (recovery is ok), or below
491 * the resync window. We take the first readable disk when
492 * above the resync window.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 */
494 if (conf->mddev->recovery_cp < MaxSector
495 && (this_sector + sectors >= conf->next_resync)) {
496 /* make sure that disk is operational */
497 slot = 0;
498 disk = r10_bio->devs[slot].devnum;
499
Suzanne Woodd6065f72005-11-08 21:39:27 -0800500 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800501 r10_bio->devs[slot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800502 !test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 slot++;
504 if (slot == conf->copies) {
505 slot = 0;
506 disk = -1;
507 break;
508 }
509 disk = r10_bio->devs[slot].devnum;
510 }
511 goto rb_out;
512 }
513
514
515 /* make sure the disk is operational */
516 slot = 0;
517 disk = r10_bio->devs[slot].devnum;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800518 while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800519 r10_bio->devs[slot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800520 !test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 slot ++;
522 if (slot == conf->copies) {
523 disk = -1;
524 goto rb_out;
525 }
526 disk = r10_bio->devs[slot].devnum;
527 }
528
529
NeilBrown3ec67ac2005-09-09 16:23:40 -0700530 current_distance = abs(r10_bio->devs[slot].addr -
531 conf->mirrors[disk].head_position);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532
533 /* Find the disk whose head is closest */
534
535 for (nslot = slot; nslot < conf->copies; nslot++) {
536 int ndisk = r10_bio->devs[nslot].devnum;
537
538
Suzanne Woodd6065f72005-11-08 21:39:27 -0800539 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800540 r10_bio->devs[nslot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800541 !test_bit(In_sync, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542 continue;
543
NeilBrown22dfdf52005-11-28 13:44:09 -0800544 /* This optimisation is debatable, and completely destroys
545 * sequential read speed for 'far copies' arrays. So only
546 * keep it for 'near' arrays, and review those later.
547 */
548 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 disk = ndisk;
550 slot = nslot;
551 break;
552 }
553 new_distance = abs(r10_bio->devs[nslot].addr -
554 conf->mirrors[ndisk].head_position);
555 if (new_distance < current_distance) {
556 current_distance = new_distance;
557 disk = ndisk;
558 slot = nslot;
559 }
560 }
561
562rb_out:
563 r10_bio->read_slot = slot;
564/* conf->next_seq_sect = this_sector + sectors;*/
565
Suzanne Woodd6065f72005-11-08 21:39:27 -0800566 if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
568 rcu_read_unlock();
569
570 return disk;
571}
572
573static void unplug_slaves(mddev_t *mddev)
574{
575 conf_t *conf = mddev_to_conf(mddev);
576 int i;
577
578 rcu_read_lock();
579 for (i=0; i<mddev->raid_disks; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800580 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -0800581 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
583
584 atomic_inc(&rdev->nr_pending);
585 rcu_read_unlock();
586
587 if (r_queue->unplug_fn)
588 r_queue->unplug_fn(r_queue);
589
590 rdev_dec_pending(rdev, mddev);
591 rcu_read_lock();
592 }
593 }
594 rcu_read_unlock();
595}
596
597static void raid10_unplug(request_queue_t *q)
598{
NeilBrown6cce3b22006-01-06 00:20:16 -0800599 mddev_t *mddev = q->queuedata;
600
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601 unplug_slaves(q->queuedata);
NeilBrown6cce3b22006-01-06 00:20:16 -0800602 md_wakeup_thread(mddev->thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603}
604
605static int raid10_issue_flush(request_queue_t *q, struct gendisk *disk,
606 sector_t *error_sector)
607{
608 mddev_t *mddev = q->queuedata;
609 conf_t *conf = mddev_to_conf(mddev);
610 int i, ret = 0;
611
612 rcu_read_lock();
613 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800614 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -0800615 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616 struct block_device *bdev = rdev->bdev;
617 request_queue_t *r_queue = bdev_get_queue(bdev);
618
619 if (!r_queue->issue_flush_fn)
620 ret = -EOPNOTSUPP;
621 else {
622 atomic_inc(&rdev->nr_pending);
623 rcu_read_unlock();
624 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
625 error_sector);
626 rdev_dec_pending(rdev, mddev);
627 rcu_read_lock();
628 }
629 }
630 }
631 rcu_read_unlock();
632 return ret;
633}
634
NeilBrown0a27ec92006-01-06 00:20:13 -0800635/* Barriers....
636 * Sometimes we need to suspend IO while we do something else,
637 * either some resync/recovery, or reconfigure the array.
638 * To do this we raise a 'barrier'.
639 * The 'barrier' is a counter that can be raised multiple times
640 * to count how many activities are happening which preclude
641 * normal IO.
642 * We can only raise the barrier if there is no pending IO.
643 * i.e. if nr_pending == 0.
644 * We choose only to raise the barrier if no-one is waiting for the
645 * barrier to go down. This means that as soon as an IO request
646 * is ready, no other operations which require a barrier will start
647 * until the IO request has had a chance.
648 *
649 * So: regular IO calls 'wait_barrier'. When that returns there
650 * is no backgroup IO happening, It must arrange to call
651 * allow_barrier when it has finished its IO.
652 * backgroup IO calls must call raise_barrier. Once that returns
653 * there is no normal IO happeing. It must arrange to call
654 * lower_barrier when the particular background IO completes.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 */
656#define RESYNC_DEPTH 32
657
NeilBrown6cce3b22006-01-06 00:20:16 -0800658static void raise_barrier(conf_t *conf, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659{
NeilBrown6cce3b22006-01-06 00:20:16 -0800660 BUG_ON(force && !conf->barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700661 spin_lock_irq(&conf->resync_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
NeilBrown6cce3b22006-01-06 00:20:16 -0800663 /* Wait until no block IO is waiting (unless 'force') */
664 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
NeilBrown0a27ec92006-01-06 00:20:13 -0800665 conf->resync_lock,
666 raid10_unplug(conf->mddev->queue));
667
668 /* block any new IO from starting */
669 conf->barrier++;
670
671 /* No wait for all pending IO to complete */
672 wait_event_lock_irq(conf->wait_barrier,
673 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
674 conf->resync_lock,
675 raid10_unplug(conf->mddev->queue));
676
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 spin_unlock_irq(&conf->resync_lock);
678}
679
NeilBrown0a27ec92006-01-06 00:20:13 -0800680static void lower_barrier(conf_t *conf)
681{
682 unsigned long flags;
683 spin_lock_irqsave(&conf->resync_lock, flags);
684 conf->barrier--;
685 spin_unlock_irqrestore(&conf->resync_lock, flags);
686 wake_up(&conf->wait_barrier);
687}
688
689static void wait_barrier(conf_t *conf)
690{
691 spin_lock_irq(&conf->resync_lock);
692 if (conf->barrier) {
693 conf->nr_waiting++;
694 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
695 conf->resync_lock,
696 raid10_unplug(conf->mddev->queue));
697 conf->nr_waiting--;
698 }
699 conf->nr_pending++;
700 spin_unlock_irq(&conf->resync_lock);
701}
702
703static void allow_barrier(conf_t *conf)
704{
705 unsigned long flags;
706 spin_lock_irqsave(&conf->resync_lock, flags);
707 conf->nr_pending--;
708 spin_unlock_irqrestore(&conf->resync_lock, flags);
709 wake_up(&conf->wait_barrier);
710}
711
NeilBrown4443ae12006-01-06 00:20:28 -0800712static void freeze_array(conf_t *conf)
713{
714 /* stop syncio and normal IO and wait for everything to
NeilBrownf1885932006-01-06 00:20:42 -0800715 * go quiet.
NeilBrown4443ae12006-01-06 00:20:28 -0800716 * We increment barrier and nr_waiting, and then
717 * wait until barrier+nr_pending match nr_queued+2
718 */
719 spin_lock_irq(&conf->resync_lock);
720 conf->barrier++;
721 conf->nr_waiting++;
722 wait_event_lock_irq(conf->wait_barrier,
723 conf->barrier+conf->nr_pending == conf->nr_queued+2,
724 conf->resync_lock,
725 raid10_unplug(conf->mddev->queue));
726 spin_unlock_irq(&conf->resync_lock);
727}
728
729static void unfreeze_array(conf_t *conf)
730{
731 /* reverse the effect of the freeze */
732 spin_lock_irq(&conf->resync_lock);
733 conf->barrier--;
734 conf->nr_waiting--;
735 wake_up(&conf->wait_barrier);
736 spin_unlock_irq(&conf->resync_lock);
737}
738
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739static int make_request(request_queue_t *q, struct bio * bio)
740{
741 mddev_t *mddev = q->queuedata;
742 conf_t *conf = mddev_to_conf(mddev);
743 mirror_info_t *mirror;
744 r10bio_t *r10_bio;
745 struct bio *read_bio;
746 int i;
747 int chunk_sects = conf->chunk_mask + 1;
Jens Axboea3623572005-11-01 09:26:16 +0100748 const int rw = bio_data_dir(bio);
NeilBrown6cce3b22006-01-06 00:20:16 -0800749 struct bio_list bl;
750 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751
NeilBrowne5dcdd82005-09-09 16:23:41 -0700752 if (unlikely(bio_barrier(bio))) {
753 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
754 return 0;
755 }
756
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 /* If this request crosses a chunk boundary, we need to
758 * split it. This will only happen for 1 PAGE (or less) requests.
759 */
760 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
761 > chunk_sects &&
762 conf->near_copies < conf->raid_disks)) {
763 struct bio_pair *bp;
764 /* Sanity check -- queue functions should prevent this happening */
765 if (bio->bi_vcnt != 1 ||
766 bio->bi_idx != 0)
767 goto bad_map;
768 /* This is a one page bio that upper layers
769 * refuse to split for us, so we need to split it.
770 */
771 bp = bio_split(bio, bio_split_pool,
772 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
773 if (make_request(q, &bp->bio1))
774 generic_make_request(&bp->bio1);
775 if (make_request(q, &bp->bio2))
776 generic_make_request(&bp->bio2);
777
778 bio_pair_release(bp);
779 return 0;
780 bad_map:
781 printk("raid10_make_request bug: can't convert block across chunks"
782 " or bigger than %dk %llu %d\n", chunk_sects/2,
783 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
784
785 bio_io_error(bio, bio->bi_size);
786 return 0;
787 }
788
NeilBrown3d310eb2005-06-21 17:17:26 -0700789 md_write_start(mddev, bio);
NeilBrown06d91a52005-06-21 17:17:12 -0700790
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 /*
792 * Register the new request and wait if the reconstruction
793 * thread has put up a bar for new requests.
794 * Continue immediately if no resync is active currently.
795 */
NeilBrown0a27ec92006-01-06 00:20:13 -0800796 wait_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797
Jens Axboea3623572005-11-01 09:26:16 +0100798 disk_stat_inc(mddev->gendisk, ios[rw]);
799 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800
801 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
802
803 r10_bio->master_bio = bio;
804 r10_bio->sectors = bio->bi_size >> 9;
805
806 r10_bio->mddev = mddev;
807 r10_bio->sector = bio->bi_sector;
NeilBrown6cce3b22006-01-06 00:20:16 -0800808 r10_bio->state = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809
Jens Axboea3623572005-11-01 09:26:16 +0100810 if (rw == READ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 /*
812 * read balancing logic:
813 */
814 int disk = read_balance(conf, r10_bio);
815 int slot = r10_bio->read_slot;
816 if (disk < 0) {
817 raid_end_bio_io(r10_bio);
818 return 0;
819 }
820 mirror = conf->mirrors + disk;
821
822 read_bio = bio_clone(bio, GFP_NOIO);
823
824 r10_bio->devs[slot].bio = read_bio;
825
826 read_bio->bi_sector = r10_bio->devs[slot].addr +
827 mirror->rdev->data_offset;
828 read_bio->bi_bdev = mirror->rdev->bdev;
829 read_bio->bi_end_io = raid10_end_read_request;
830 read_bio->bi_rw = READ;
831 read_bio->bi_private = r10_bio;
832
833 generic_make_request(read_bio);
834 return 0;
835 }
836
837 /*
838 * WRITE:
839 */
840 /* first select target devices under spinlock and
841 * inc refcount on their rdev. Record them by setting
842 * bios[x] to bio
843 */
844 raid10_find_phys(conf, r10_bio);
845 rcu_read_lock();
846 for (i = 0; i < conf->copies; i++) {
847 int d = r10_bio->devs[i].devnum;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800848 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
849 if (rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -0800850 !test_bit(Faulty, &rdev->flags)) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800851 atomic_inc(&rdev->nr_pending);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 r10_bio->devs[i].bio = bio;
NeilBrown6cce3b22006-01-06 00:20:16 -0800853 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 r10_bio->devs[i].bio = NULL;
NeilBrown6cce3b22006-01-06 00:20:16 -0800855 set_bit(R10BIO_Degraded, &r10_bio->state);
856 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857 }
858 rcu_read_unlock();
859
NeilBrown6cce3b22006-01-06 00:20:16 -0800860 atomic_set(&r10_bio->remaining, 0);
NeilBrown06d91a52005-06-21 17:17:12 -0700861
NeilBrown6cce3b22006-01-06 00:20:16 -0800862 bio_list_init(&bl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 for (i = 0; i < conf->copies; i++) {
864 struct bio *mbio;
865 int d = r10_bio->devs[i].devnum;
866 if (!r10_bio->devs[i].bio)
867 continue;
868
869 mbio = bio_clone(bio, GFP_NOIO);
870 r10_bio->devs[i].bio = mbio;
871
872 mbio->bi_sector = r10_bio->devs[i].addr+
873 conf->mirrors[d].rdev->data_offset;
874 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
875 mbio->bi_end_io = raid10_end_write_request;
876 mbio->bi_rw = WRITE;
877 mbio->bi_private = r10_bio;
878
879 atomic_inc(&r10_bio->remaining);
NeilBrown6cce3b22006-01-06 00:20:16 -0800880 bio_list_add(&bl, mbio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 }
882
NeilBrown6cce3b22006-01-06 00:20:16 -0800883 bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
884 spin_lock_irqsave(&conf->device_lock, flags);
885 bio_list_merge(&conf->pending_bio_list, &bl);
886 blk_plug_device(mddev->queue);
887 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888
889 return 0;
890}
891
892static void status(struct seq_file *seq, mddev_t *mddev)
893{
894 conf_t *conf = mddev_to_conf(mddev);
895 int i;
896
897 if (conf->near_copies < conf->raid_disks)
898 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
899 if (conf->near_copies > 1)
900 seq_printf(seq, " %d near-copies", conf->near_copies);
901 if (conf->far_copies > 1)
902 seq_printf(seq, " %d far-copies", conf->far_copies);
903
904 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
905 conf->working_disks);
906 for (i = 0; i < conf->raid_disks; i++)
907 seq_printf(seq, "%s",
908 conf->mirrors[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -0800909 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 seq_printf(seq, "]");
911}
912
913static void error(mddev_t *mddev, mdk_rdev_t *rdev)
914{
915 char b[BDEVNAME_SIZE];
916 conf_t *conf = mddev_to_conf(mddev);
917
918 /*
919 * If it is not operational, then we have already marked it as dead
920 * else if it is the last working disks, ignore the error, let the
921 * next level up know.
922 * else mark the drive as failed
923 */
NeilBrownb2d444d2005-11-08 21:39:31 -0800924 if (test_bit(In_sync, &rdev->flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925 && conf->working_disks == 1)
926 /*
927 * Don't fail the drive, just return an IO error.
928 * The test should really be more sophisticated than
929 * "working_disks == 1", but it isn't critical, and
930 * can wait until we do more sophisticated "is the drive
931 * really dead" tests...
932 */
933 return;
NeilBrownb2d444d2005-11-08 21:39:31 -0800934 if (test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935 mddev->degraded++;
936 conf->working_disks--;
937 /*
938 * if recovery is running, make sure it aborts.
939 */
940 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
941 }
NeilBrownb2d444d2005-11-08 21:39:31 -0800942 clear_bit(In_sync, &rdev->flags);
943 set_bit(Faulty, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944 mddev->sb_dirty = 1;
945 printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
946 " Operation continuing on %d devices\n",
947 bdevname(rdev->bdev,b), conf->working_disks);
948}
949
950static void print_conf(conf_t *conf)
951{
952 int i;
953 mirror_info_t *tmp;
954
955 printk("RAID10 conf printout:\n");
956 if (!conf) {
957 printk("(!conf)\n");
958 return;
959 }
960 printk(" --- wd:%d rd:%d\n", conf->working_disks,
961 conf->raid_disks);
962
963 for (i = 0; i < conf->raid_disks; i++) {
964 char b[BDEVNAME_SIZE];
965 tmp = conf->mirrors + i;
966 if (tmp->rdev)
967 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
NeilBrownb2d444d2005-11-08 21:39:31 -0800968 i, !test_bit(In_sync, &tmp->rdev->flags),
969 !test_bit(Faulty, &tmp->rdev->flags),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970 bdevname(tmp->rdev->bdev,b));
971 }
972}
973
974static void close_sync(conf_t *conf)
975{
NeilBrown0a27ec92006-01-06 00:20:13 -0800976 wait_barrier(conf);
977 allow_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978
979 mempool_destroy(conf->r10buf_pool);
980 conf->r10buf_pool = NULL;
981}
982
NeilBrown6d508242005-09-09 16:24:03 -0700983/* check if there are enough drives for
984 * every block to appear on atleast one
985 */
986static int enough(conf_t *conf)
987{
988 int first = 0;
989
990 do {
991 int n = conf->copies;
992 int cnt = 0;
993 while (n--) {
994 if (conf->mirrors[first].rdev)
995 cnt++;
996 first = (first+1) % conf->raid_disks;
997 }
998 if (cnt == 0)
999 return 0;
1000 } while (first != 0);
1001 return 1;
1002}
1003
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004static int raid10_spare_active(mddev_t *mddev)
1005{
1006 int i;
1007 conf_t *conf = mddev->private;
1008 mirror_info_t *tmp;
1009
1010 /*
1011 * Find all non-in_sync disks within the RAID10 configuration
1012 * and mark them in_sync
1013 */
1014 for (i = 0; i < conf->raid_disks; i++) {
1015 tmp = conf->mirrors + i;
1016 if (tmp->rdev
NeilBrownb2d444d2005-11-08 21:39:31 -08001017 && !test_bit(Faulty, &tmp->rdev->flags)
1018 && !test_bit(In_sync, &tmp->rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019 conf->working_disks++;
1020 mddev->degraded--;
NeilBrownb2d444d2005-11-08 21:39:31 -08001021 set_bit(In_sync, &tmp->rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 }
1023 }
1024
1025 print_conf(conf);
1026 return 0;
1027}
1028
1029
1030static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1031{
1032 conf_t *conf = mddev->private;
1033 int found = 0;
1034 int mirror;
1035 mirror_info_t *p;
1036
1037 if (mddev->recovery_cp < MaxSector)
1038 /* only hot-add to in-sync arrays, as recovery is
1039 * very different from resync
1040 */
1041 return 0;
NeilBrown6d508242005-09-09 16:24:03 -07001042 if (!enough(conf))
1043 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044
NeilBrown6cce3b22006-01-06 00:20:16 -08001045 if (rdev->saved_raid_disk >= 0 &&
1046 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1047 mirror = rdev->saved_raid_disk;
1048 else
1049 mirror = 0;
1050 for ( ; mirror < mddev->raid_disks; mirror++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051 if ( !(p=conf->mirrors+mirror)->rdev) {
1052
1053 blk_queue_stack_limits(mddev->queue,
1054 rdev->bdev->bd_disk->queue);
1055 /* as we don't honour merge_bvec_fn, we must never risk
1056 * violating it, so limit ->max_sector to one PAGE, as
1057 * a one page request is never in violation.
1058 */
1059 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1060 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1061 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1062
1063 p->head_position = 0;
1064 rdev->raid_disk = mirror;
1065 found = 1;
NeilBrown6cce3b22006-01-06 00:20:16 -08001066 if (rdev->saved_raid_disk != mirror)
1067 conf->fullsync = 1;
Suzanne Woodd6065f72005-11-08 21:39:27 -08001068 rcu_assign_pointer(p->rdev, rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 break;
1070 }
1071
1072 print_conf(conf);
1073 return found;
1074}
1075
1076static int raid10_remove_disk(mddev_t *mddev, int number)
1077{
1078 conf_t *conf = mddev->private;
1079 int err = 0;
1080 mdk_rdev_t *rdev;
1081 mirror_info_t *p = conf->mirrors+ number;
1082
1083 print_conf(conf);
1084 rdev = p->rdev;
1085 if (rdev) {
NeilBrownb2d444d2005-11-08 21:39:31 -08001086 if (test_bit(In_sync, &rdev->flags) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 atomic_read(&rdev->nr_pending)) {
1088 err = -EBUSY;
1089 goto abort;
1090 }
1091 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07001092 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 if (atomic_read(&rdev->nr_pending)) {
1094 /* lost the race, try later */
1095 err = -EBUSY;
1096 p->rdev = rdev;
1097 }
1098 }
1099abort:
1100
1101 print_conf(conf);
1102 return err;
1103}
1104
1105
1106static int end_sync_read(struct bio *bio, unsigned int bytes_done, int error)
1107{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1109 conf_t *conf = mddev_to_conf(r10_bio->mddev);
1110 int i,d;
1111
1112 if (bio->bi_size)
1113 return 1;
1114
1115 for (i=0; i<conf->copies; i++)
1116 if (r10_bio->devs[i].bio == bio)
1117 break;
1118 if (i == conf->copies)
1119 BUG();
1120 update_head_pos(i, r10_bio);
1121 d = r10_bio->devs[i].devnum;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001122
1123 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1124 set_bit(R10BIO_Uptodate, &r10_bio->state);
1125 else if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 md_error(r10_bio->mddev,
1127 conf->mirrors[d].rdev);
1128
1129 /* for reconstruct, we always reschedule after a read.
1130 * for resync, only after all reads
1131 */
1132 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1133 atomic_dec_and_test(&r10_bio->remaining)) {
1134 /* we have read all the blocks,
1135 * do the comparison in process context in raid10d
1136 */
1137 reschedule_retry(r10_bio);
1138 }
1139 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1140 return 0;
1141}
1142
1143static int end_sync_write(struct bio *bio, unsigned int bytes_done, int error)
1144{
1145 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1146 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1147 mddev_t *mddev = r10_bio->mddev;
1148 conf_t *conf = mddev_to_conf(mddev);
1149 int i,d;
1150
1151 if (bio->bi_size)
1152 return 1;
1153
1154 for (i = 0; i < conf->copies; i++)
1155 if (r10_bio->devs[i].bio == bio)
1156 break;
1157 d = r10_bio->devs[i].devnum;
1158
1159 if (!uptodate)
1160 md_error(mddev, conf->mirrors[d].rdev);
1161 update_head_pos(i, r10_bio);
1162
1163 while (atomic_dec_and_test(&r10_bio->remaining)) {
1164 if (r10_bio->master_bio == NULL) {
1165 /* the primary of several recovery bios */
1166 md_done_sync(mddev, r10_bio->sectors, 1);
1167 put_buf(r10_bio);
1168 break;
1169 } else {
1170 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1171 put_buf(r10_bio);
1172 r10_bio = r10_bio2;
1173 }
1174 }
1175 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1176 return 0;
1177}
1178
1179/*
1180 * Note: sync and recover and handled very differently for raid10
1181 * This code is for resync.
1182 * For resync, we read through virtual addresses and read all blocks.
1183 * If there is any error, we schedule a write. The lowest numbered
1184 * drive is authoritative.
1185 * However requests come for physical address, so we need to map.
1186 * For every physical address there are raid_disks/copies virtual addresses,
1187 * which is always are least one, but is not necessarly an integer.
1188 * This means that a physical address can span multiple chunks, so we may
1189 * have to submit multiple io requests for a single sync request.
1190 */
1191/*
1192 * We check if all blocks are in-sync and only write to blocks that
1193 * aren't in sync
1194 */
1195static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1196{
1197 conf_t *conf = mddev_to_conf(mddev);
1198 int i, first;
1199 struct bio *tbio, *fbio;
1200
1201 atomic_set(&r10_bio->remaining, 1);
1202
1203 /* find the first device with a block */
1204 for (i=0; i<conf->copies; i++)
1205 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1206 break;
1207
1208 if (i == conf->copies)
1209 goto done;
1210
1211 first = i;
1212 fbio = r10_bio->devs[i].bio;
1213
1214 /* now find blocks with errors */
NeilBrown0eb3ff12006-01-06 00:20:29 -08001215 for (i=0 ; i < conf->copies ; i++) {
1216 int j, d;
1217 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 tbio = r10_bio->devs[i].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001220
1221 if (tbio->bi_end_io != end_sync_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222 continue;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001223 if (i == first)
1224 continue;
1225 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1226 /* We know that the bi_io_vec layout is the same for
1227 * both 'first' and 'i', so we just compare them.
1228 * All vec entries are PAGE_SIZE;
1229 */
1230 for (j = 0; j < vcnt; j++)
1231 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1232 page_address(tbio->bi_io_vec[j].bv_page),
1233 PAGE_SIZE))
1234 break;
1235 if (j == vcnt)
1236 continue;
1237 mddev->resync_mismatches += r10_bio->sectors;
1238 }
NeilBrown18f08812006-01-06 00:20:25 -08001239 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1240 /* Don't fix anything. */
1241 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242 /* Ok, we need to write this bio
1243 * First we need to fixup bv_offset, bv_len and
1244 * bi_vecs, as the read request might have corrupted these
1245 */
1246 tbio->bi_vcnt = vcnt;
1247 tbio->bi_size = r10_bio->sectors << 9;
1248 tbio->bi_idx = 0;
1249 tbio->bi_phys_segments = 0;
1250 tbio->bi_hw_segments = 0;
1251 tbio->bi_hw_front_size = 0;
1252 tbio->bi_hw_back_size = 0;
1253 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1254 tbio->bi_flags |= 1 << BIO_UPTODATE;
1255 tbio->bi_next = NULL;
1256 tbio->bi_rw = WRITE;
1257 tbio->bi_private = r10_bio;
1258 tbio->bi_sector = r10_bio->devs[i].addr;
1259
1260 for (j=0; j < vcnt ; j++) {
1261 tbio->bi_io_vec[j].bv_offset = 0;
1262 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1263
1264 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1265 page_address(fbio->bi_io_vec[j].bv_page),
1266 PAGE_SIZE);
1267 }
1268 tbio->bi_end_io = end_sync_write;
1269
1270 d = r10_bio->devs[i].devnum;
1271 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1272 atomic_inc(&r10_bio->remaining);
1273 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1274
1275 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1276 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1277 generic_make_request(tbio);
1278 }
1279
1280done:
1281 if (atomic_dec_and_test(&r10_bio->remaining)) {
1282 md_done_sync(mddev, r10_bio->sectors, 1);
1283 put_buf(r10_bio);
1284 }
1285}
1286
1287/*
1288 * Now for the recovery code.
1289 * Recovery happens across physical sectors.
1290 * We recover all non-is_sync drives by finding the virtual address of
1291 * each, and then choose a working drive that also has that virt address.
1292 * There is a separate r10_bio for each non-in_sync drive.
1293 * Only the first two slots are in use. The first for reading,
1294 * The second for writing.
1295 *
1296 */
1297
1298static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1299{
1300 conf_t *conf = mddev_to_conf(mddev);
1301 int i, d;
1302 struct bio *bio, *wbio;
1303
1304
1305 /* move the pages across to the second bio
1306 * and submit the write request
1307 */
1308 bio = r10_bio->devs[0].bio;
1309 wbio = r10_bio->devs[1].bio;
1310 for (i=0; i < wbio->bi_vcnt; i++) {
1311 struct page *p = bio->bi_io_vec[i].bv_page;
1312 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1313 wbio->bi_io_vec[i].bv_page = p;
1314 }
1315 d = r10_bio->devs[1].devnum;
1316
1317 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1318 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
NeilBrown0eb3ff12006-01-06 00:20:29 -08001319 if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1320 generic_make_request(wbio);
1321 else
1322 bio_endio(wbio, wbio->bi_size, -EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323}
1324
1325
1326/*
1327 * This is a kernel thread which:
1328 *
1329 * 1. Retries failed read operations on working mirrors.
1330 * 2. Updates the raid superblock when problems encounter.
1331 * 3. Performs writes following reads for array syncronising.
1332 */
1333
1334static void raid10d(mddev_t *mddev)
1335{
1336 r10bio_t *r10_bio;
1337 struct bio *bio;
1338 unsigned long flags;
1339 conf_t *conf = mddev_to_conf(mddev);
1340 struct list_head *head = &conf->retry_list;
1341 int unplug=0;
1342 mdk_rdev_t *rdev;
1343
1344 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345
1346 for (;;) {
1347 char b[BDEVNAME_SIZE];
1348 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrown6cce3b22006-01-06 00:20:16 -08001349
1350 if (conf->pending_bio_list.head) {
1351 bio = bio_list_get(&conf->pending_bio_list);
1352 blk_remove_plug(mddev->queue);
1353 spin_unlock_irqrestore(&conf->device_lock, flags);
1354 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
1355 if (bitmap_unplug(mddev->bitmap) != 0)
1356 printk("%s: bitmap file write failed!\n", mdname(mddev));
1357
1358 while (bio) { /* submit pending writes */
1359 struct bio *next = bio->bi_next;
1360 bio->bi_next = NULL;
1361 generic_make_request(bio);
1362 bio = next;
1363 }
1364 unplug = 1;
1365
1366 continue;
1367 }
1368
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 if (list_empty(head))
1370 break;
1371 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1372 list_del(head->prev);
NeilBrown4443ae12006-01-06 00:20:28 -08001373 conf->nr_queued--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 spin_unlock_irqrestore(&conf->device_lock, flags);
1375
1376 mddev = r10_bio->mddev;
1377 conf = mddev_to_conf(mddev);
1378 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1379 sync_request_write(mddev, r10_bio);
1380 unplug = 1;
1381 } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1382 recovery_request_write(mddev, r10_bio);
1383 unplug = 1;
1384 } else {
1385 int mirror;
NeilBrown4443ae12006-01-06 00:20:28 -08001386 /* we got a read error. Maybe the drive is bad. Maybe just
1387 * the block and we can fix it.
1388 * We freeze all other IO, and try reading the block from
1389 * other devices. When we find one, we re-write
1390 * and check it that fixes the read error.
1391 * This is all done synchronously while the array is
1392 * frozen.
1393 */
1394 int sect = 0; /* Offset from r10_bio->sector */
1395 int sectors = r10_bio->sectors;
1396 freeze_array(conf);
1397 if (mddev->ro == 0) while(sectors) {
1398 int s = sectors;
1399 int sl = r10_bio->read_slot;
1400 int success = 0;
1401
1402 if (s > (PAGE_SIZE>>9))
1403 s = PAGE_SIZE >> 9;
1404
1405 do {
1406 int d = r10_bio->devs[sl].devnum;
1407 rdev = conf->mirrors[d].rdev;
1408 if (rdev &&
1409 test_bit(In_sync, &rdev->flags) &&
1410 sync_page_io(rdev->bdev,
1411 r10_bio->devs[sl].addr +
1412 sect + rdev->data_offset,
1413 s<<9,
1414 conf->tmppage, READ))
1415 success = 1;
1416 else {
1417 sl++;
1418 if (sl == conf->copies)
1419 sl = 0;
1420 }
1421 } while (!success && sl != r10_bio->read_slot);
1422
1423 if (success) {
NeilBrown097426f2006-01-06 00:20:37 -08001424 int start = sl;
NeilBrown4443ae12006-01-06 00:20:28 -08001425 /* write it back and re-read */
1426 while (sl != r10_bio->read_slot) {
1427 int d;
1428 if (sl==0)
1429 sl = conf->copies;
1430 sl--;
1431 d = r10_bio->devs[sl].devnum;
1432 rdev = conf->mirrors[d].rdev;
1433 if (rdev &&
1434 test_bit(In_sync, &rdev->flags)) {
1435 if (sync_page_io(rdev->bdev,
1436 r10_bio->devs[sl].addr +
1437 sect + rdev->data_offset,
NeilBrown097426f2006-01-06 00:20:37 -08001438 s<<9, conf->tmppage, WRITE) == 0)
NeilBrown4443ae12006-01-06 00:20:28 -08001439 /* Well, this device is dead */
1440 md_error(mddev, rdev);
NeilBrown097426f2006-01-06 00:20:37 -08001441 }
1442 }
1443 sl = start;
1444 while (sl != r10_bio->read_slot) {
1445 int d;
1446 if (sl==0)
1447 sl = conf->copies;
1448 sl--;
1449 d = r10_bio->devs[sl].devnum;
1450 rdev = conf->mirrors[d].rdev;
1451 if (rdev &&
1452 test_bit(In_sync, &rdev->flags)) {
1453 if (sync_page_io(rdev->bdev,
1454 r10_bio->devs[sl].addr +
1455 sect + rdev->data_offset,
1456 s<<9, conf->tmppage, READ) == 0)
1457 /* Well, this device is dead */
1458 md_error(mddev, rdev);
NeilBrown4443ae12006-01-06 00:20:28 -08001459 }
1460 }
1461 } else {
1462 /* Cannot read from anywhere -- bye bye array */
1463 md_error(mddev, conf->mirrors[r10_bio->devs[r10_bio->read_slot].devnum].rdev);
1464 break;
1465 }
1466 sectors -= s;
1467 sect += s;
1468 }
1469
1470 unfreeze_array(conf);
1471
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 bio = r10_bio->devs[r10_bio->read_slot].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001473 r10_bio->devs[r10_bio->read_slot].bio =
1474 mddev->ro ? IO_BLOCKED : NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 bio_put(bio);
1476 mirror = read_balance(conf, r10_bio);
1477 if (mirror == -1) {
1478 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1479 " read error for block %llu\n",
1480 bdevname(bio->bi_bdev,b),
1481 (unsigned long long)r10_bio->sector);
1482 raid_end_bio_io(r10_bio);
1483 } else {
1484 rdev = conf->mirrors[mirror].rdev;
1485 if (printk_ratelimit())
1486 printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1487 " another mirror\n",
1488 bdevname(rdev->bdev,b),
1489 (unsigned long long)r10_bio->sector);
1490 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1491 r10_bio->devs[r10_bio->read_slot].bio = bio;
1492 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1493 + rdev->data_offset;
1494 bio->bi_bdev = rdev->bdev;
1495 bio->bi_rw = READ;
1496 bio->bi_private = r10_bio;
1497 bio->bi_end_io = raid10_end_read_request;
1498 unplug = 1;
1499 generic_make_request(bio);
1500 }
1501 }
1502 }
1503 spin_unlock_irqrestore(&conf->device_lock, flags);
1504 if (unplug)
1505 unplug_slaves(mddev);
1506}
1507
1508
1509static int init_resync(conf_t *conf)
1510{
1511 int buffs;
1512
1513 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1514 if (conf->r10buf_pool)
1515 BUG();
1516 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1517 if (!conf->r10buf_pool)
1518 return -ENOMEM;
1519 conf->next_resync = 0;
1520 return 0;
1521}
1522
1523/*
1524 * perform a "sync" on one "block"
1525 *
1526 * We need to make sure that no normal I/O request - particularly write
1527 * requests - conflict with active sync requests.
1528 *
1529 * This is achieved by tracking pending requests and a 'barrier' concept
1530 * that can be installed to exclude normal IO requests.
1531 *
1532 * Resync and recovery are handled very differently.
1533 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1534 *
1535 * For resync, we iterate over virtual addresses, read all copies,
1536 * and update if there are differences. If only one copy is live,
1537 * skip it.
1538 * For recovery, we iterate over physical addresses, read a good
1539 * value for each non-in_sync drive, and over-write.
1540 *
1541 * So, for recovery we may have several outstanding complex requests for a
1542 * given address, one for each out-of-sync device. We model this by allocating
1543 * a number of r10_bio structures, one for each out-of-sync device.
1544 * As we setup these structures, we collect all bio's together into a list
1545 * which we then process collectively to add pages, and then process again
1546 * to pass to generic_make_request.
1547 *
1548 * The r10_bio structures are linked using a borrowed master_bio pointer.
1549 * This link is counted in ->remaining. When the r10_bio that points to NULL
1550 * has its remaining count decremented to 0, the whole complex operation
1551 * is complete.
1552 *
1553 */
1554
NeilBrown57afd892005-06-21 17:17:13 -07001555static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556{
1557 conf_t *conf = mddev_to_conf(mddev);
1558 r10bio_t *r10_bio;
1559 struct bio *biolist = NULL, *bio;
1560 sector_t max_sector, nr_sectors;
1561 int disk;
1562 int i;
NeilBrown6cce3b22006-01-06 00:20:16 -08001563 int max_sync;
1564 int sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565
1566 sector_t sectors_skipped = 0;
1567 int chunks_skipped = 0;
1568
1569 if (!conf->r10buf_pool)
1570 if (init_resync(conf))
NeilBrown57afd892005-06-21 17:17:13 -07001571 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572
1573 skipped:
1574 max_sector = mddev->size << 1;
1575 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1576 max_sector = mddev->resync_max_sectors;
1577 if (sector_nr >= max_sector) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001578 /* If we aborted, we need to abort the
1579 * sync on the 'current' bitmap chucks (there can
1580 * be several when recovering multiple devices).
1581 * as we may have started syncing it but not finished.
1582 * We can find the current address in
1583 * mddev->curr_resync, but for recovery,
1584 * we need to convert that to several
1585 * virtual addresses.
1586 */
1587 if (mddev->curr_resync < max_sector) { /* aborted */
1588 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1589 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1590 &sync_blocks, 1);
1591 else for (i=0; i<conf->raid_disks; i++) {
1592 sector_t sect =
1593 raid10_find_virt(conf, mddev->curr_resync, i);
1594 bitmap_end_sync(mddev->bitmap, sect,
1595 &sync_blocks, 1);
1596 }
1597 } else /* completed sync */
1598 conf->fullsync = 0;
1599
1600 bitmap_close_sync(mddev->bitmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601 close_sync(conf);
NeilBrown57afd892005-06-21 17:17:13 -07001602 *skipped = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 return sectors_skipped;
1604 }
1605 if (chunks_skipped >= conf->raid_disks) {
1606 /* if there has been nothing to do on any drive,
1607 * then there is nothing to do at all..
1608 */
NeilBrown57afd892005-06-21 17:17:13 -07001609 *skipped = 1;
1610 return (max_sector - sector_nr) + sectors_skipped;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611 }
1612
1613 /* make sure whole request will fit in a chunk - if chunks
1614 * are meaningful
1615 */
1616 if (conf->near_copies < conf->raid_disks &&
1617 max_sector > (sector_nr | conf->chunk_mask))
1618 max_sector = (sector_nr | conf->chunk_mask) + 1;
1619 /*
1620 * If there is non-resync activity waiting for us then
1621 * put in a delay to throttle resync.
1622 */
NeilBrown0a27ec92006-01-06 00:20:13 -08001623 if (!go_faster && conf->nr_waiting)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624 msleep_interruptible(1000);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625
1626 /* Again, very different code for resync and recovery.
1627 * Both must result in an r10bio with a list of bios that
1628 * have bi_end_io, bi_sector, bi_bdev set,
1629 * and bi_private set to the r10bio.
1630 * For recovery, we may actually create several r10bios
1631 * with 2 bios in each, that correspond to the bios in the main one.
1632 * In this case, the subordinate r10bios link back through a
1633 * borrowed master_bio pointer, and the counter in the master
1634 * includes a ref from each subordinate.
1635 */
1636 /* First, we decide what to do and set ->bi_end_io
1637 * To end_sync_read if we want to read, and
1638 * end_sync_write if we will want to write.
1639 */
1640
NeilBrown6cce3b22006-01-06 00:20:16 -08001641 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1643 /* recovery... the complicated one */
1644 int i, j, k;
1645 r10_bio = NULL;
1646
1647 for (i=0 ; i<conf->raid_disks; i++)
1648 if (conf->mirrors[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001649 !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001650 int still_degraded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651 /* want to reconstruct this device */
1652 r10bio_t *rb2 = r10_bio;
NeilBrown6cce3b22006-01-06 00:20:16 -08001653 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1654 int must_sync;
1655 /* Unless we are doing a full sync, we only need
1656 * to recover the block if it is set in the bitmap
1657 */
1658 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1659 &sync_blocks, 1);
1660 if (sync_blocks < max_sync)
1661 max_sync = sync_blocks;
1662 if (!must_sync &&
1663 !conf->fullsync) {
1664 /* yep, skip the sync_blocks here, but don't assume
1665 * that there will never be anything to do here
1666 */
1667 chunks_skipped = -1;
1668 continue;
1669 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670
1671 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
NeilBrown6cce3b22006-01-06 00:20:16 -08001672 raise_barrier(conf, rb2 != NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 atomic_set(&r10_bio->remaining, 0);
1674
1675 r10_bio->master_bio = (struct bio*)rb2;
1676 if (rb2)
1677 atomic_inc(&rb2->remaining);
1678 r10_bio->mddev = mddev;
1679 set_bit(R10BIO_IsRecover, &r10_bio->state);
NeilBrown6cce3b22006-01-06 00:20:16 -08001680 r10_bio->sector = sect;
1681
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 raid10_find_phys(conf, r10_bio);
NeilBrown6cce3b22006-01-06 00:20:16 -08001683 /* Need to check if this section will still be
1684 * degraded
1685 */
1686 for (j=0; j<conf->copies;j++) {
1687 int d = r10_bio->devs[j].devnum;
1688 if (conf->mirrors[d].rdev == NULL ||
NeilBrowna24a8dd2006-01-06 00:20:35 -08001689 test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001690 still_degraded = 1;
NeilBrowna24a8dd2006-01-06 00:20:35 -08001691 break;
1692 }
NeilBrown6cce3b22006-01-06 00:20:16 -08001693 }
1694 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1695 &sync_blocks, still_degraded);
1696
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 for (j=0; j<conf->copies;j++) {
1698 int d = r10_bio->devs[j].devnum;
1699 if (conf->mirrors[d].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001700 test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 /* This is where we read from */
1702 bio = r10_bio->devs[0].bio;
1703 bio->bi_next = biolist;
1704 biolist = bio;
1705 bio->bi_private = r10_bio;
1706 bio->bi_end_io = end_sync_read;
1707 bio->bi_rw = 0;
1708 bio->bi_sector = r10_bio->devs[j].addr +
1709 conf->mirrors[d].rdev->data_offset;
1710 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1711 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1712 atomic_inc(&r10_bio->remaining);
1713 /* and we write to 'i' */
1714
1715 for (k=0; k<conf->copies; k++)
1716 if (r10_bio->devs[k].devnum == i)
1717 break;
1718 bio = r10_bio->devs[1].bio;
1719 bio->bi_next = biolist;
1720 biolist = bio;
1721 bio->bi_private = r10_bio;
1722 bio->bi_end_io = end_sync_write;
1723 bio->bi_rw = 1;
1724 bio->bi_sector = r10_bio->devs[k].addr +
1725 conf->mirrors[i].rdev->data_offset;
1726 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1727
1728 r10_bio->devs[0].devnum = d;
1729 r10_bio->devs[1].devnum = i;
1730
1731 break;
1732 }
1733 }
1734 if (j == conf->copies) {
NeilBrown87fc7672005-09-09 16:24:04 -07001735 /* Cannot recover, so abort the recovery */
1736 put_buf(r10_bio);
1737 r10_bio = rb2;
1738 if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
1739 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1740 mdname(mddev));
1741 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 }
1743 }
1744 if (biolist == NULL) {
1745 while (r10_bio) {
1746 r10bio_t *rb2 = r10_bio;
1747 r10_bio = (r10bio_t*) rb2->master_bio;
1748 rb2->master_bio = NULL;
1749 put_buf(rb2);
1750 }
1751 goto giveup;
1752 }
1753 } else {
1754 /* resync. Schedule a read for every block at this virt offset */
1755 int count = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08001756
1757 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1758 &sync_blocks, mddev->degraded) &&
1759 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1760 /* We can skip this block */
1761 *skipped = 1;
1762 return sync_blocks + sectors_skipped;
1763 }
1764 if (sync_blocks < max_sync)
1765 max_sync = sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1767
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 r10_bio->mddev = mddev;
1769 atomic_set(&r10_bio->remaining, 0);
NeilBrown6cce3b22006-01-06 00:20:16 -08001770 raise_barrier(conf, 0);
1771 conf->next_resync = sector_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772
1773 r10_bio->master_bio = NULL;
1774 r10_bio->sector = sector_nr;
1775 set_bit(R10BIO_IsSync, &r10_bio->state);
1776 raid10_find_phys(conf, r10_bio);
1777 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1778
1779 for (i=0; i<conf->copies; i++) {
1780 int d = r10_bio->devs[i].devnum;
1781 bio = r10_bio->devs[i].bio;
1782 bio->bi_end_io = NULL;
1783 if (conf->mirrors[d].rdev == NULL ||
NeilBrownb2d444d2005-11-08 21:39:31 -08001784 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 continue;
1786 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1787 atomic_inc(&r10_bio->remaining);
1788 bio->bi_next = biolist;
1789 biolist = bio;
1790 bio->bi_private = r10_bio;
1791 bio->bi_end_io = end_sync_read;
1792 bio->bi_rw = 0;
1793 bio->bi_sector = r10_bio->devs[i].addr +
1794 conf->mirrors[d].rdev->data_offset;
1795 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1796 count++;
1797 }
1798
1799 if (count < 2) {
1800 for (i=0; i<conf->copies; i++) {
1801 int d = r10_bio->devs[i].devnum;
1802 if (r10_bio->devs[i].bio->bi_end_io)
1803 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1804 }
1805 put_buf(r10_bio);
1806 biolist = NULL;
1807 goto giveup;
1808 }
1809 }
1810
1811 for (bio = biolist; bio ; bio=bio->bi_next) {
1812
1813 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1814 if (bio->bi_end_io)
1815 bio->bi_flags |= 1 << BIO_UPTODATE;
1816 bio->bi_vcnt = 0;
1817 bio->bi_idx = 0;
1818 bio->bi_phys_segments = 0;
1819 bio->bi_hw_segments = 0;
1820 bio->bi_size = 0;
1821 }
1822
1823 nr_sectors = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08001824 if (sector_nr + max_sync < max_sector)
1825 max_sector = sector_nr + max_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826 do {
1827 struct page *page;
1828 int len = PAGE_SIZE;
1829 disk = 0;
1830 if (sector_nr + (len>>9) > max_sector)
1831 len = (max_sector - sector_nr) << 9;
1832 if (len == 0)
1833 break;
1834 for (bio= biolist ; bio ; bio=bio->bi_next) {
1835 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1836 if (bio_add_page(bio, page, len, 0) == 0) {
1837 /* stop here */
1838 struct bio *bio2;
1839 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1840 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1841 /* remove last page from this bio */
1842 bio2->bi_vcnt--;
1843 bio2->bi_size -= len;
1844 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1845 }
1846 goto bio_full;
1847 }
1848 disk = i;
1849 }
1850 nr_sectors += len>>9;
1851 sector_nr += len>>9;
1852 } while (biolist->bi_vcnt < RESYNC_PAGES);
1853 bio_full:
1854 r10_bio->sectors = nr_sectors;
1855
1856 while (biolist) {
1857 bio = biolist;
1858 biolist = biolist->bi_next;
1859
1860 bio->bi_next = NULL;
1861 r10_bio = bio->bi_private;
1862 r10_bio->sectors = nr_sectors;
1863
1864 if (bio->bi_end_io == end_sync_read) {
1865 md_sync_acct(bio->bi_bdev, nr_sectors);
1866 generic_make_request(bio);
1867 }
1868 }
1869
NeilBrown57afd892005-06-21 17:17:13 -07001870 if (sectors_skipped)
1871 /* pretend they weren't skipped, it makes
1872 * no important difference in this case
1873 */
1874 md_done_sync(mddev, sectors_skipped, 1);
1875
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 return sectors_skipped + nr_sectors;
1877 giveup:
1878 /* There is nowhere to write, so all non-sync
1879 * drives must be failed, so try the next chunk...
1880 */
1881 {
NeilBrown57afd892005-06-21 17:17:13 -07001882 sector_t sec = max_sector - sector_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883 sectors_skipped += sec;
1884 chunks_skipped ++;
1885 sector_nr = max_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 goto skipped;
1887 }
1888}
1889
1890static int run(mddev_t *mddev)
1891{
1892 conf_t *conf;
1893 int i, disk_idx;
1894 mirror_info_t *disk;
1895 mdk_rdev_t *rdev;
1896 struct list_head *tmp;
1897 int nc, fc;
1898 sector_t stride, size;
1899
NeilBrown2604b702006-01-06 00:20:36 -08001900 if (mddev->chunk_size == 0) {
1901 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
1902 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903 }
NeilBrown2604b702006-01-06 00:20:36 -08001904
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905 nc = mddev->layout & 255;
1906 fc = (mddev->layout >> 8) & 255;
1907 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
1908 (mddev->layout >> 16)) {
1909 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1910 mdname(mddev), mddev->layout);
1911 goto out;
1912 }
1913 /*
1914 * copy the already verified devices into our private RAID10
1915 * bookkeeping area. [whatever we allocate in run(),
1916 * should be freed in stop()]
1917 */
NeilBrown4443ae12006-01-06 00:20:28 -08001918 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 mddev->private = conf;
1920 if (!conf) {
1921 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1922 mdname(mddev));
1923 goto out;
1924 }
NeilBrown4443ae12006-01-06 00:20:28 -08001925 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 GFP_KERNEL);
1927 if (!conf->mirrors) {
1928 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1929 mdname(mddev));
1930 goto out_free_conf;
1931 }
NeilBrown4443ae12006-01-06 00:20:28 -08001932
1933 conf->tmppage = alloc_page(GFP_KERNEL);
1934 if (!conf->tmppage)
1935 goto out_free_conf;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936
1937 conf->near_copies = nc;
1938 conf->far_copies = fc;
1939 conf->copies = nc*fc;
1940 conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
1941 conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
1942 stride = mddev->size >> (conf->chunk_shift-1);
1943 sector_div(stride, fc);
1944 conf->stride = stride << conf->chunk_shift;
1945
1946 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
1947 r10bio_pool_free, conf);
1948 if (!conf->r10bio_pool) {
1949 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1950 mdname(mddev));
1951 goto out_free_conf;
1952 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953
1954 ITERATE_RDEV(mddev, rdev, tmp) {
1955 disk_idx = rdev->raid_disk;
1956 if (disk_idx >= mddev->raid_disks
1957 || disk_idx < 0)
1958 continue;
1959 disk = conf->mirrors + disk_idx;
1960
1961 disk->rdev = rdev;
1962
1963 blk_queue_stack_limits(mddev->queue,
1964 rdev->bdev->bd_disk->queue);
1965 /* as we don't honour merge_bvec_fn, we must never risk
1966 * violating it, so limit ->max_sector to one PAGE, as
1967 * a one page request is never in violation.
1968 */
1969 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1970 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1971 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1972
1973 disk->head_position = 0;
NeilBrownb2d444d2005-11-08 21:39:31 -08001974 if (!test_bit(Faulty, &rdev->flags) && test_bit(In_sync, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975 conf->working_disks++;
1976 }
1977 conf->raid_disks = mddev->raid_disks;
1978 conf->mddev = mddev;
1979 spin_lock_init(&conf->device_lock);
1980 INIT_LIST_HEAD(&conf->retry_list);
1981
1982 spin_lock_init(&conf->resync_lock);
NeilBrown0a27ec92006-01-06 00:20:13 -08001983 init_waitqueue_head(&conf->wait_barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984
NeilBrown6d508242005-09-09 16:24:03 -07001985 /* need to check that every block has at least one working mirror */
1986 if (!enough(conf)) {
1987 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
1988 mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 goto out_free_conf;
1990 }
1991
1992 mddev->degraded = 0;
1993 for (i = 0; i < conf->raid_disks; i++) {
1994
1995 disk = conf->mirrors + i;
1996
1997 if (!disk->rdev) {
1998 disk->head_position = 0;
1999 mddev->degraded++;
2000 }
2001 }
2002
2003
2004 mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2005 if (!mddev->thread) {
2006 printk(KERN_ERR
2007 "raid10: couldn't allocate thread for %s\n",
2008 mdname(mddev));
2009 goto out_free_conf;
2010 }
2011
2012 printk(KERN_INFO
2013 "raid10: raid set %s active with %d out of %d devices\n",
2014 mdname(mddev), mddev->raid_disks - mddev->degraded,
2015 mddev->raid_disks);
2016 /*
2017 * Ok, everything is just fine now
2018 */
2019 size = conf->stride * conf->raid_disks;
2020 sector_div(size, conf->near_copies);
2021 mddev->array_size = size/2;
2022 mddev->resync_max_sectors = size;
2023
NeilBrown7a5febe2005-05-16 21:53:16 -07002024 mddev->queue->unplug_fn = raid10_unplug;
2025 mddev->queue->issue_flush_fn = raid10_issue_flush;
2026
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 /* Calculate max read-ahead size.
2028 * We need to readahead at least twice a whole stripe....
2029 * maybe...
2030 */
2031 {
NeilBrown2d1f3b52006-01-06 00:20:31 -08002032 int stripe = conf->raid_disks * mddev->chunk_size / PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 stripe /= conf->near_copies;
2034 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2035 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2036 }
2037
2038 if (conf->near_copies < mddev->raid_disks)
2039 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2040 return 0;
2041
2042out_free_conf:
2043 if (conf->r10bio_pool)
2044 mempool_destroy(conf->r10bio_pool);
NeilBrown1345b1d2006-01-06 00:20:40 -08002045 safe_put_page(conf->tmppage);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07002046 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047 kfree(conf);
2048 mddev->private = NULL;
2049out:
2050 return -EIO;
2051}
2052
2053static int stop(mddev_t *mddev)
2054{
2055 conf_t *conf = mddev_to_conf(mddev);
2056
2057 md_unregister_thread(mddev->thread);
2058 mddev->thread = NULL;
2059 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2060 if (conf->r10bio_pool)
2061 mempool_destroy(conf->r10bio_pool);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07002062 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002063 kfree(conf);
2064 mddev->private = NULL;
2065 return 0;
2066}
2067
NeilBrown6cce3b22006-01-06 00:20:16 -08002068static void raid10_quiesce(mddev_t *mddev, int state)
2069{
2070 conf_t *conf = mddev_to_conf(mddev);
2071
2072 switch(state) {
2073 case 1:
2074 raise_barrier(conf, 0);
2075 break;
2076 case 0:
2077 lower_barrier(conf);
2078 break;
2079 }
2080 if (mddev->thread) {
2081 if (mddev->bitmap)
2082 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2083 else
2084 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2085 md_wakeup_thread(mddev->thread);
2086 }
2087}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088
NeilBrown2604b702006-01-06 00:20:36 -08002089static struct mdk_personality raid10_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090{
2091 .name = "raid10",
NeilBrown2604b702006-01-06 00:20:36 -08002092 .level = 10,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093 .owner = THIS_MODULE,
2094 .make_request = make_request,
2095 .run = run,
2096 .stop = stop,
2097 .status = status,
2098 .error_handler = error,
2099 .hot_add_disk = raid10_add_disk,
2100 .hot_remove_disk= raid10_remove_disk,
2101 .spare_active = raid10_spare_active,
2102 .sync_request = sync_request,
NeilBrown6cce3b22006-01-06 00:20:16 -08002103 .quiesce = raid10_quiesce,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104};
2105
2106static int __init raid_init(void)
2107{
NeilBrown2604b702006-01-06 00:20:36 -08002108 return register_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109}
2110
2111static void raid_exit(void)
2112{
NeilBrown2604b702006-01-06 00:20:36 -08002113 unregister_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114}
2115
2116module_init(raid_init);
2117module_exit(raid_exit);
2118MODULE_LICENSE("GPL");
2119MODULE_ALIAS("md-personality-9"); /* RAID10 */
NeilBrownd9d166c2006-01-06 00:20:51 -08002120MODULE_ALIAS("md-raid10");
NeilBrown2604b702006-01-06 00:20:36 -08002121MODULE_ALIAS("md-level-10");