blob: e34cd0e6247385f5827d7232ae4257c664f47c7b [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)
NeilBrownc93983b2006-06-26 00:27:41 -070032 * far_offset (stored in bit 16 of layout )
Linus Torvalds1da177e2005-04-16 15:20:36 -070033 *
34 * The data to be stored is divided into chunks using chunksize.
35 * Each device is divided into far_copies sections.
36 * In each section, chunks are laid out in a style similar to raid0, but
37 * near_copies copies of each chunk is stored (each on a different drive).
38 * The starting device for each section is offset near_copies from the starting
39 * device of the previous section.
NeilBrownc93983b2006-06-26 00:27:41 -070040 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
Linus Torvalds1da177e2005-04-16 15:20:36 -070041 * drive.
42 * near_copies and far_copies must be at least one, and their product is at most
43 * raid_disks.
NeilBrownc93983b2006-06-26 00:27:41 -070044 *
45 * If far_offset is true, then the far_copies are handled a bit differently.
46 * The copies are still in different stripes, but instead of be very far apart
47 * on disk, there are adjacent stripes.
Linus Torvalds1da177e2005-04-16 15:20:36 -070048 */
49
50/*
51 * Number of guaranteed r10bios in case of extreme VM load:
52 */
53#define NR_RAID10_BIOS 256
54
55static void unplug_slaves(mddev_t *mddev);
56
NeilBrown0a27ec92006-01-06 00:20:13 -080057static void allow_barrier(conf_t *conf);
58static void lower_barrier(conf_t *conf);
59
Al Virodd0fc662005-10-07 07:46:04 +010060static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070061{
62 conf_t *conf = data;
63 r10bio_t *r10_bio;
64 int size = offsetof(struct r10bio_s, devs[conf->copies]);
65
66 /* allocate a r10bio with room for raid_disks entries in the bios array */
NeilBrown9ffae0c2006-01-06 00:20:32 -080067 r10_bio = kzalloc(size, gfp_flags);
68 if (!r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -070069 unplug_slaves(conf->mddev);
70
71 return r10_bio;
72}
73
74static void r10bio_pool_free(void *r10_bio, void *data)
75{
76 kfree(r10_bio);
77}
78
NeilBrown0310fa22008-08-05 15:54:14 +100079/* Maximum size of each resync request */
Linus Torvalds1da177e2005-04-16 15:20:36 -070080#define RESYNC_BLOCK_SIZE (64*1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -070081#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
NeilBrown0310fa22008-08-05 15:54:14 +100082/* amount of memory to reserve for resync requests */
83#define RESYNC_WINDOW (1024*1024)
84/* maximum number of concurrent requests, memory permitting */
85#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -070086
87/*
88 * When performing a resync, we need to read and compare, so
89 * we need as many pages are there are copies.
90 * When performing a recovery, we need 2 bios, one for read,
91 * one for write (we recover only one drive per r10buf)
92 *
93 */
Al Virodd0fc662005-10-07 07:46:04 +010094static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070095{
96 conf_t *conf = data;
97 struct page *page;
98 r10bio_t *r10_bio;
99 struct bio *bio;
100 int i, j;
101 int nalloc;
102
103 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
104 if (!r10_bio) {
105 unplug_slaves(conf->mddev);
106 return NULL;
107 }
108
109 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
110 nalloc = conf->copies; /* resync */
111 else
112 nalloc = 2; /* recovery */
113
114 /*
115 * Allocate bios.
116 */
117 for (j = nalloc ; j-- ; ) {
118 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
119 if (!bio)
120 goto out_free_bio;
121 r10_bio->devs[j].bio = bio;
122 }
123 /*
124 * Allocate RESYNC_PAGES data pages and attach them
125 * where needed.
126 */
127 for (j = 0 ; j < nalloc; j++) {
128 bio = r10_bio->devs[j].bio;
129 for (i = 0; i < RESYNC_PAGES; i++) {
130 page = alloc_page(gfp_flags);
131 if (unlikely(!page))
132 goto out_free_pages;
133
134 bio->bi_io_vec[i].bv_page = page;
135 }
136 }
137
138 return r10_bio;
139
140out_free_pages:
141 for ( ; i > 0 ; i--)
NeilBrown1345b1d2006-01-06 00:20:40 -0800142 safe_put_page(bio->bi_io_vec[i-1].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143 while (j--)
144 for (i = 0; i < RESYNC_PAGES ; i++)
NeilBrown1345b1d2006-01-06 00:20:40 -0800145 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146 j = -1;
147out_free_bio:
148 while ( ++j < nalloc )
149 bio_put(r10_bio->devs[j].bio);
150 r10bio_pool_free(r10_bio, conf);
151 return NULL;
152}
153
154static void r10buf_pool_free(void *__r10_bio, void *data)
155{
156 int i;
157 conf_t *conf = data;
158 r10bio_t *r10bio = __r10_bio;
159 int j;
160
161 for (j=0; j < conf->copies; j++) {
162 struct bio *bio = r10bio->devs[j].bio;
163 if (bio) {
164 for (i = 0; i < RESYNC_PAGES; i++) {
NeilBrown1345b1d2006-01-06 00:20:40 -0800165 safe_put_page(bio->bi_io_vec[i].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 bio->bi_io_vec[i].bv_page = NULL;
167 }
168 bio_put(bio);
169 }
170 }
171 r10bio_pool_free(r10bio, conf);
172}
173
174static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
175{
176 int i;
177
178 for (i = 0; i < conf->copies; i++) {
179 struct bio **bio = & r10_bio->devs[i].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -0800180 if (*bio && *bio != IO_BLOCKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181 bio_put(*bio);
182 *bio = NULL;
183 }
184}
185
Arjan van de Ven858119e2006-01-14 13:20:43 -0800186static void free_r10bio(r10bio_t *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188 conf_t *conf = mddev_to_conf(r10_bio->mddev);
189
190 /*
191 * Wake up any possible resync thread that waits for the device
192 * to go idle.
193 */
NeilBrown0a27ec92006-01-06 00:20:13 -0800194 allow_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195
196 put_all_bios(conf, r10_bio);
197 mempool_free(r10_bio, conf->r10bio_pool);
198}
199
Arjan van de Ven858119e2006-01-14 13:20:43 -0800200static void put_buf(r10bio_t *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201{
202 conf_t *conf = mddev_to_conf(r10_bio->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203
204 mempool_free(r10_bio, conf->r10buf_pool);
205
NeilBrown0a27ec92006-01-06 00:20:13 -0800206 lower_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207}
208
209static void reschedule_retry(r10bio_t *r10_bio)
210{
211 unsigned long flags;
212 mddev_t *mddev = r10_bio->mddev;
213 conf_t *conf = mddev_to_conf(mddev);
214
215 spin_lock_irqsave(&conf->device_lock, flags);
216 list_add(&r10_bio->retry_list, &conf->retry_list);
NeilBrown4443ae12006-01-06 00:20:28 -0800217 conf->nr_queued ++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218 spin_unlock_irqrestore(&conf->device_lock, flags);
219
Arthur Jones388667b2008-07-25 12:03:38 -0700220 /* wake up frozen array... */
221 wake_up(&conf->wait_barrier);
222
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223 md_wakeup_thread(mddev->thread);
224}
225
226/*
227 * raid_end_bio_io() is called when we have finished servicing a mirrored
228 * operation and are ready to return a success/failure code to the buffer
229 * cache layer.
230 */
231static void raid_end_bio_io(r10bio_t *r10_bio)
232{
233 struct bio *bio = r10_bio->master_bio;
234
NeilBrown6712ecf2007-09-27 12:47:43 +0200235 bio_endio(bio,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
237 free_r10bio(r10_bio);
238}
239
240/*
241 * Update disk head position estimator based on IRQ completion info.
242 */
243static inline void update_head_pos(int slot, r10bio_t *r10_bio)
244{
245 conf_t *conf = mddev_to_conf(r10_bio->mddev);
246
247 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
248 r10_bio->devs[slot].addr + (r10_bio->sectors);
249}
250
NeilBrown6712ecf2007-09-27 12:47:43 +0200251static void raid10_end_read_request(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252{
253 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
254 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
255 int slot, dev;
256 conf_t *conf = mddev_to_conf(r10_bio->mddev);
257
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258
259 slot = r10_bio->read_slot;
260 dev = r10_bio->devs[slot].devnum;
261 /*
262 * this branch is our 'one mirror IO has finished' event handler:
263 */
NeilBrown4443ae12006-01-06 00:20:28 -0800264 update_head_pos(slot, r10_bio);
265
266 if (uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267 /*
268 * Set R10BIO_Uptodate in our master bio, so that
269 * we will return a good error code to the higher
270 * levels even if IO on some other mirrored buffer fails.
271 *
272 * The 'master' represents the composite IO operation to
273 * user-side. So if something waits for IO, then it will
274 * wait for the 'master' bio.
275 */
276 set_bit(R10BIO_Uptodate, &r10_bio->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277 raid_end_bio_io(r10_bio);
NeilBrown4443ae12006-01-06 00:20:28 -0800278 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279 /*
280 * oops, read error:
281 */
282 char b[BDEVNAME_SIZE];
283 if (printk_ratelimit())
284 printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
285 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
286 reschedule_retry(r10_bio);
287 }
288
289 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290}
291
NeilBrown6712ecf2007-09-27 12:47:43 +0200292static void raid10_end_write_request(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293{
294 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
295 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
296 int slot, dev;
297 conf_t *conf = mddev_to_conf(r10_bio->mddev);
298
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299 for (slot = 0; slot < conf->copies; slot++)
300 if (r10_bio->devs[slot].bio == bio)
301 break;
302 dev = r10_bio->devs[slot].devnum;
303
304 /*
305 * this branch is our 'one mirror IO has finished' event handler:
306 */
NeilBrown6cce3b22006-01-06 00:20:16 -0800307 if (!uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
NeilBrown6cce3b22006-01-06 00:20:16 -0800309 /* an I/O failed, we can't clear the bitmap */
310 set_bit(R10BIO_Degraded, &r10_bio->state);
311 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312 /*
313 * Set R10BIO_Uptodate in our master bio, so that
314 * we will return a good error code for to the higher
315 * levels even if IO on some other mirrored buffer fails.
316 *
317 * The 'master' represents the composite IO operation to
318 * user-side. So if something waits for IO, then it will
319 * wait for the 'master' bio.
320 */
321 set_bit(R10BIO_Uptodate, &r10_bio->state);
322
323 update_head_pos(slot, r10_bio);
324
325 /*
326 *
327 * Let's see if all mirrored write operations have finished
328 * already.
329 */
330 if (atomic_dec_and_test(&r10_bio->remaining)) {
NeilBrown6cce3b22006-01-06 00:20:16 -0800331 /* clear the bitmap if all writes complete successfully */
332 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
333 r10_bio->sectors,
334 !test_bit(R10BIO_Degraded, &r10_bio->state),
335 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336 md_write_end(r10_bio->mddev);
337 raid_end_bio_io(r10_bio);
338 }
339
340 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341}
342
343
344/*
345 * RAID10 layout manager
346 * Aswell as the chunksize and raid_disks count, there are two
347 * parameters: near_copies and far_copies.
348 * near_copies * far_copies must be <= raid_disks.
349 * Normally one of these will be 1.
350 * If both are 1, we get raid0.
351 * If near_copies == raid_disks, we get raid1.
352 *
353 * Chunks are layed out in raid0 style with near_copies copies of the
354 * first chunk, followed by near_copies copies of the next chunk and
355 * so on.
356 * If far_copies > 1, then after 1/far_copies of the array has been assigned
357 * as described above, we start again with a device offset of near_copies.
358 * So we effectively have another copy of the whole array further down all
359 * the drives, but with blocks on different drives.
360 * With this layout, and block is never stored twice on the one device.
361 *
362 * raid10_find_phys finds the sector offset of a given virtual sector
NeilBrownc93983b2006-06-26 00:27:41 -0700363 * on each device that it is on.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 *
365 * raid10_find_virt does the reverse mapping, from a device and a
366 * sector offset to a virtual address
367 */
368
369static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
370{
371 int n,f;
372 sector_t sector;
373 sector_t chunk;
374 sector_t stripe;
375 int dev;
376
377 int slot = 0;
378
379 /* now calculate first sector/dev */
380 chunk = r10bio->sector >> conf->chunk_shift;
381 sector = r10bio->sector & conf->chunk_mask;
382
383 chunk *= conf->near_copies;
384 stripe = chunk;
385 dev = sector_div(stripe, conf->raid_disks);
NeilBrownc93983b2006-06-26 00:27:41 -0700386 if (conf->far_offset)
387 stripe *= conf->far_copies;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388
389 sector += stripe << conf->chunk_shift;
390
391 /* and calculate all the others */
392 for (n=0; n < conf->near_copies; n++) {
393 int d = dev;
394 sector_t s = sector;
395 r10bio->devs[slot].addr = sector;
396 r10bio->devs[slot].devnum = d;
397 slot++;
398
399 for (f = 1; f < conf->far_copies; f++) {
400 d += conf->near_copies;
401 if (d >= conf->raid_disks)
402 d -= conf->raid_disks;
403 s += conf->stride;
404 r10bio->devs[slot].devnum = d;
405 r10bio->devs[slot].addr = s;
406 slot++;
407 }
408 dev++;
409 if (dev >= conf->raid_disks) {
410 dev = 0;
411 sector += (conf->chunk_mask + 1);
412 }
413 }
414 BUG_ON(slot != conf->copies);
415}
416
417static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
418{
419 sector_t offset, chunk, vchunk;
420
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421 offset = sector & conf->chunk_mask;
NeilBrownc93983b2006-06-26 00:27:41 -0700422 if (conf->far_offset) {
423 int fc;
424 chunk = sector >> conf->chunk_shift;
425 fc = sector_div(chunk, conf->far_copies);
426 dev -= fc * conf->near_copies;
427 if (dev < 0)
428 dev += conf->raid_disks;
429 } else {
NeilBrown64a742b2007-02-28 20:11:18 -0800430 while (sector >= conf->stride) {
NeilBrownc93983b2006-06-26 00:27:41 -0700431 sector -= conf->stride;
432 if (dev < conf->near_copies)
433 dev += conf->raid_disks - conf->near_copies;
434 else
435 dev -= conf->near_copies;
436 }
437 chunk = sector >> conf->chunk_shift;
438 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700439 vchunk = chunk * conf->raid_disks + dev;
440 sector_div(vchunk, conf->near_copies);
441 return (vchunk << conf->chunk_shift) + offset;
442}
443
444/**
445 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
446 * @q: request queue
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200447 * @bvm: properties of new bio
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 * @biovec: the request that could be merged to it.
449 *
450 * Return amount of bytes we can accept at this offset
451 * If near_copies == raid_disk, there are no striping issues,
452 * but in that case, the function isn't called at all.
453 */
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200454static int raid10_mergeable_bvec(struct request_queue *q,
455 struct bvec_merge_data *bvm,
456 struct bio_vec *biovec)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457{
458 mddev_t *mddev = q->queuedata;
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200459 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 int max;
461 unsigned int chunk_sectors = mddev->chunk_size >> 9;
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200462 unsigned int bio_sectors = bvm->bi_size >> 9;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463
464 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
465 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200466 if (max <= biovec->bv_len && bio_sectors == 0)
467 return biovec->bv_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 else
469 return max;
470}
471
472/*
473 * This routine returns the disk from which the requested read should
474 * be done. There is a per-array 'next expected sequential IO' sector
475 * number - if this matches on the next IO then we use the last disk.
476 * There is also a per-disk 'last know head position' sector that is
477 * maintained from IRQ contexts, both the normal and the resync IO
478 * completion handlers update this position correctly. If there is no
479 * perfect sequential match then we pick the disk whose head is closest.
480 *
481 * If there are 2 mirrors in the same 2 devices, performance degrades
482 * because position is mirror, not device based.
483 *
484 * The rdev for the device selected will have nr_pending incremented.
485 */
486
487/*
488 * FIXME: possibly should rethink readbalancing and do it differently
489 * depending on near_copies / far_copies geometry.
490 */
491static int read_balance(conf_t *conf, r10bio_t *r10_bio)
492{
493 const unsigned long this_sector = r10_bio->sector;
494 int disk, slot, nslot;
495 const int sectors = r10_bio->sectors;
496 sector_t new_distance, current_distance;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800497 mdk_rdev_t *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498
499 raid10_find_phys(conf, r10_bio);
500 rcu_read_lock();
501 /*
502 * Check if we can balance. We can balance on the whole
NeilBrown6cce3b22006-01-06 00:20:16 -0800503 * device if no resync is going on (recovery is ok), or below
504 * the resync window. We take the first readable disk when
505 * above the resync window.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 */
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
Suzanne Woodd6065f72005-11-08 21:39:27 -0800513 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800514 r10_bio->devs[slot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800515 !test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 slot++;
517 if (slot == conf->copies) {
518 slot = 0;
519 disk = -1;
520 break;
521 }
522 disk = r10_bio->devs[slot].devnum;
523 }
524 goto rb_out;
525 }
526
527
528 /* make sure the disk is operational */
529 slot = 0;
530 disk = r10_bio->devs[slot].devnum;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800531 while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800532 r10_bio->devs[slot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800533 !test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 slot ++;
535 if (slot == conf->copies) {
536 disk = -1;
537 goto rb_out;
538 }
539 disk = r10_bio->devs[slot].devnum;
540 }
541
542
NeilBrown3ec67ac2005-09-09 16:23:40 -0700543 current_distance = abs(r10_bio->devs[slot].addr -
544 conf->mirrors[disk].head_position);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545
Keld Simonsen8ed3a192008-03-04 14:29:34 -0800546 /* Find the disk whose head is closest,
547 * or - for far > 1 - find the closest to partition beginning */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548
549 for (nslot = slot; nslot < conf->copies; nslot++) {
550 int ndisk = r10_bio->devs[nslot].devnum;
551
552
Suzanne Woodd6065f72005-11-08 21:39:27 -0800553 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800554 r10_bio->devs[nslot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800555 !test_bit(In_sync, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556 continue;
557
NeilBrown22dfdf52005-11-28 13:44:09 -0800558 /* This optimisation is debatable, and completely destroys
559 * sequential read speed for 'far copies' arrays. So only
560 * keep it for 'near' arrays, and review those later.
561 */
562 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 disk = ndisk;
564 slot = nslot;
565 break;
566 }
Keld Simonsen8ed3a192008-03-04 14:29:34 -0800567
568 /* for far > 1 always use the lowest address */
569 if (conf->far_copies > 1)
570 new_distance = r10_bio->devs[nslot].addr;
571 else
572 new_distance = abs(r10_bio->devs[nslot].addr -
573 conf->mirrors[ndisk].head_position);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 if (new_distance < current_distance) {
575 current_distance = new_distance;
576 disk = ndisk;
577 slot = nslot;
578 }
579 }
580
581rb_out:
582 r10_bio->read_slot = slot;
583/* conf->next_seq_sect = this_sector + sectors;*/
584
Suzanne Woodd6065f72005-11-08 21:39:27 -0800585 if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
NeilBrown29fc7e32006-02-03 03:03:41 -0800587 else
588 disk = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589 rcu_read_unlock();
590
591 return disk;
592}
593
594static void unplug_slaves(mddev_t *mddev)
595{
596 conf_t *conf = mddev_to_conf(mddev);
597 int i;
598
599 rcu_read_lock();
600 for (i=0; i<mddev->raid_disks; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800601 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -0800602 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200603 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604
605 atomic_inc(&rdev->nr_pending);
606 rcu_read_unlock();
607
Alan D. Brunelle2ad8b1e2007-11-07 14:26:56 -0500608 blk_unplug(r_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609
610 rdev_dec_pending(rdev, mddev);
611 rcu_read_lock();
612 }
613 }
614 rcu_read_unlock();
615}
616
Jens Axboe165125e2007-07-24 09:28:11 +0200617static void raid10_unplug(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618{
NeilBrown6cce3b22006-01-06 00:20:16 -0800619 mddev_t *mddev = q->queuedata;
620
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 unplug_slaves(q->queuedata);
NeilBrown6cce3b22006-01-06 00:20:16 -0800622 md_wakeup_thread(mddev->thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623}
624
NeilBrown0d129222006-10-03 01:15:54 -0700625static int raid10_congested(void *data, int bits)
626{
627 mddev_t *mddev = data;
628 conf_t *conf = mddev_to_conf(mddev);
629 int i, ret = 0;
630
631 rcu_read_lock();
632 for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
633 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
634 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200635 struct request_queue *q = bdev_get_queue(rdev->bdev);
NeilBrown0d129222006-10-03 01:15:54 -0700636
637 ret |= bdi_congested(&q->backing_dev_info, bits);
638 }
639 }
640 rcu_read_unlock();
641 return ret;
642}
643
NeilBrowna35e63e2008-03-04 14:29:29 -0800644static int flush_pending_writes(conf_t *conf)
645{
646 /* Any writes that have been queued but are awaiting
647 * bitmap updates get flushed here.
648 * We return 1 if any requests were actually submitted.
649 */
650 int rv = 0;
NeilBrown0d129222006-10-03 01:15:54 -0700651
NeilBrowna35e63e2008-03-04 14:29:29 -0800652 spin_lock_irq(&conf->device_lock);
653
654 if (conf->pending_bio_list.head) {
655 struct bio *bio;
656 bio = bio_list_get(&conf->pending_bio_list);
657 blk_remove_plug(conf->mddev->queue);
658 spin_unlock_irq(&conf->device_lock);
659 /* flush any pending bitmap writes to disk
660 * before proceeding w/ I/O */
661 bitmap_unplug(conf->mddev->bitmap);
662
663 while (bio) { /* submit pending writes */
664 struct bio *next = bio->bi_next;
665 bio->bi_next = NULL;
666 generic_make_request(bio);
667 bio = next;
668 }
669 rv = 1;
670 } else
671 spin_unlock_irq(&conf->device_lock);
672 return rv;
673}
NeilBrown0a27ec92006-01-06 00:20:13 -0800674/* Barriers....
675 * Sometimes we need to suspend IO while we do something else,
676 * either some resync/recovery, or reconfigure the array.
677 * To do this we raise a 'barrier'.
678 * The 'barrier' is a counter that can be raised multiple times
679 * to count how many activities are happening which preclude
680 * normal IO.
681 * We can only raise the barrier if there is no pending IO.
682 * i.e. if nr_pending == 0.
683 * We choose only to raise the barrier if no-one is waiting for the
684 * barrier to go down. This means that as soon as an IO request
685 * is ready, no other operations which require a barrier will start
686 * until the IO request has had a chance.
687 *
688 * So: regular IO calls 'wait_barrier'. When that returns there
689 * is no backgroup IO happening, It must arrange to call
690 * allow_barrier when it has finished its IO.
691 * backgroup IO calls must call raise_barrier. Once that returns
692 * there is no normal IO happeing. It must arrange to call
693 * lower_barrier when the particular background IO completes.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695
NeilBrown6cce3b22006-01-06 00:20:16 -0800696static void raise_barrier(conf_t *conf, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697{
NeilBrown6cce3b22006-01-06 00:20:16 -0800698 BUG_ON(force && !conf->barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699 spin_lock_irq(&conf->resync_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700
NeilBrown6cce3b22006-01-06 00:20:16 -0800701 /* Wait until no block IO is waiting (unless 'force') */
702 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
NeilBrown0a27ec92006-01-06 00:20:13 -0800703 conf->resync_lock,
704 raid10_unplug(conf->mddev->queue));
705
706 /* block any new IO from starting */
707 conf->barrier++;
708
709 /* No wait for all pending IO to complete */
710 wait_event_lock_irq(conf->wait_barrier,
711 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
712 conf->resync_lock,
713 raid10_unplug(conf->mddev->queue));
714
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 spin_unlock_irq(&conf->resync_lock);
716}
717
NeilBrown0a27ec92006-01-06 00:20:13 -0800718static void lower_barrier(conf_t *conf)
719{
720 unsigned long flags;
721 spin_lock_irqsave(&conf->resync_lock, flags);
722 conf->barrier--;
723 spin_unlock_irqrestore(&conf->resync_lock, flags);
724 wake_up(&conf->wait_barrier);
725}
726
727static void wait_barrier(conf_t *conf)
728{
729 spin_lock_irq(&conf->resync_lock);
730 if (conf->barrier) {
731 conf->nr_waiting++;
732 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
733 conf->resync_lock,
734 raid10_unplug(conf->mddev->queue));
735 conf->nr_waiting--;
736 }
737 conf->nr_pending++;
738 spin_unlock_irq(&conf->resync_lock);
739}
740
741static void allow_barrier(conf_t *conf)
742{
743 unsigned long flags;
744 spin_lock_irqsave(&conf->resync_lock, flags);
745 conf->nr_pending--;
746 spin_unlock_irqrestore(&conf->resync_lock, flags);
747 wake_up(&conf->wait_barrier);
748}
749
NeilBrown4443ae12006-01-06 00:20:28 -0800750static void freeze_array(conf_t *conf)
751{
752 /* stop syncio and normal IO and wait for everything to
NeilBrownf1885932006-01-06 00:20:42 -0800753 * go quiet.
NeilBrown4443ae12006-01-06 00:20:28 -0800754 * We increment barrier and nr_waiting, and then
NeilBrown1c830532008-03-04 14:29:35 -0800755 * wait until nr_pending match nr_queued+1
756 * This is called in the context of one normal IO request
757 * that has failed. Thus any sync request that might be pending
758 * will be blocked by nr_pending, and we need to wait for
759 * pending IO requests to complete or be queued for re-try.
760 * Thus the number queued (nr_queued) plus this request (1)
761 * must match the number of pending IOs (nr_pending) before
762 * we continue.
NeilBrown4443ae12006-01-06 00:20:28 -0800763 */
764 spin_lock_irq(&conf->resync_lock);
765 conf->barrier++;
766 conf->nr_waiting++;
767 wait_event_lock_irq(conf->wait_barrier,
NeilBrown1c830532008-03-04 14:29:35 -0800768 conf->nr_pending == conf->nr_queued+1,
NeilBrown4443ae12006-01-06 00:20:28 -0800769 conf->resync_lock,
NeilBrowna35e63e2008-03-04 14:29:29 -0800770 ({ flush_pending_writes(conf);
771 raid10_unplug(conf->mddev->queue); }));
NeilBrown4443ae12006-01-06 00:20:28 -0800772 spin_unlock_irq(&conf->resync_lock);
773}
774
775static void unfreeze_array(conf_t *conf)
776{
777 /* reverse the effect of the freeze */
778 spin_lock_irq(&conf->resync_lock);
779 conf->barrier--;
780 conf->nr_waiting--;
781 wake_up(&conf->wait_barrier);
782 spin_unlock_irq(&conf->resync_lock);
783}
784
Jens Axboe165125e2007-07-24 09:28:11 +0200785static int make_request(struct request_queue *q, struct bio * bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786{
787 mddev_t *mddev = q->queuedata;
788 conf_t *conf = mddev_to_conf(mddev);
789 mirror_info_t *mirror;
790 r10bio_t *r10_bio;
791 struct bio *read_bio;
792 int i;
793 int chunk_sects = conf->chunk_mask + 1;
Jens Axboea3623572005-11-01 09:26:16 +0100794 const int rw = bio_data_dir(bio);
Lars Ellenberge3881a62007-01-10 23:15:37 -0800795 const int do_sync = bio_sync(bio);
NeilBrown6cce3b22006-01-06 00:20:16 -0800796 struct bio_list bl;
797 unsigned long flags;
Dan Williams6bfe0b42008-04-30 00:52:32 -0700798 mdk_rdev_t *blocked_rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799
NeilBrowne5dcdd82005-09-09 16:23:41 -0700800 if (unlikely(bio_barrier(bio))) {
NeilBrown6712ecf2007-09-27 12:47:43 +0200801 bio_endio(bio, -EOPNOTSUPP);
NeilBrowne5dcdd82005-09-09 16:23:41 -0700802 return 0;
803 }
804
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 /* If this request crosses a chunk boundary, we need to
806 * split it. This will only happen for 1 PAGE (or less) requests.
807 */
808 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
809 > chunk_sects &&
810 conf->near_copies < conf->raid_disks)) {
811 struct bio_pair *bp;
812 /* Sanity check -- queue functions should prevent this happening */
813 if (bio->bi_vcnt != 1 ||
814 bio->bi_idx != 0)
815 goto bad_map;
816 /* This is a one page bio that upper layers
817 * refuse to split for us, so we need to split it.
818 */
819 bp = bio_split(bio, bio_split_pool,
820 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
821 if (make_request(q, &bp->bio1))
822 generic_make_request(&bp->bio1);
823 if (make_request(q, &bp->bio2))
824 generic_make_request(&bp->bio2);
825
826 bio_pair_release(bp);
827 return 0;
828 bad_map:
829 printk("raid10_make_request bug: can't convert block across chunks"
830 " or bigger than %dk %llu %d\n", chunk_sects/2,
831 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
832
NeilBrown6712ecf2007-09-27 12:47:43 +0200833 bio_io_error(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 return 0;
835 }
836
NeilBrown3d310eb2005-06-21 17:17:26 -0700837 md_write_start(mddev, bio);
NeilBrown06d91a52005-06-21 17:17:12 -0700838
Linus Torvalds1da177e2005-04-16 15:20:36 -0700839 /*
840 * Register the new request and wait if the reconstruction
841 * thread has put up a bar for new requests.
842 * Continue immediately if no resync is active currently.
843 */
NeilBrown0a27ec92006-01-06 00:20:13 -0800844 wait_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845
Jens Axboea3623572005-11-01 09:26:16 +0100846 disk_stat_inc(mddev->gendisk, ios[rw]);
847 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848
849 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
850
851 r10_bio->master_bio = bio;
852 r10_bio->sectors = bio->bi_size >> 9;
853
854 r10_bio->mddev = mddev;
855 r10_bio->sector = bio->bi_sector;
NeilBrown6cce3b22006-01-06 00:20:16 -0800856 r10_bio->state = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857
Jens Axboea3623572005-11-01 09:26:16 +0100858 if (rw == READ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 /*
860 * read balancing logic:
861 */
862 int disk = read_balance(conf, r10_bio);
863 int slot = r10_bio->read_slot;
864 if (disk < 0) {
865 raid_end_bio_io(r10_bio);
866 return 0;
867 }
868 mirror = conf->mirrors + disk;
869
870 read_bio = bio_clone(bio, GFP_NOIO);
871
872 r10_bio->devs[slot].bio = read_bio;
873
874 read_bio->bi_sector = r10_bio->devs[slot].addr +
875 mirror->rdev->data_offset;
876 read_bio->bi_bdev = mirror->rdev->bdev;
877 read_bio->bi_end_io = raid10_end_read_request;
Lars Ellenberge3881a62007-01-10 23:15:37 -0800878 read_bio->bi_rw = READ | do_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 read_bio->bi_private = r10_bio;
880
881 generic_make_request(read_bio);
882 return 0;
883 }
884
885 /*
886 * WRITE:
887 */
Dan Williams6bfe0b42008-04-30 00:52:32 -0700888 /* first select target devices under rcu_lock and
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 * inc refcount on their rdev. Record them by setting
890 * bios[x] to bio
891 */
892 raid10_find_phys(conf, r10_bio);
Dan Williams6bfe0b42008-04-30 00:52:32 -0700893 retry_write:
Harvey Harrisoncb6969e2008-05-06 20:42:32 -0700894 blocked_rdev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 rcu_read_lock();
896 for (i = 0; i < conf->copies; i++) {
897 int d = r10_bio->devs[i].devnum;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800898 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
Dan Williams6bfe0b42008-04-30 00:52:32 -0700899 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
900 atomic_inc(&rdev->nr_pending);
901 blocked_rdev = rdev;
902 break;
903 }
904 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800905 atomic_inc(&rdev->nr_pending);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 r10_bio->devs[i].bio = bio;
NeilBrown6cce3b22006-01-06 00:20:16 -0800907 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908 r10_bio->devs[i].bio = NULL;
NeilBrown6cce3b22006-01-06 00:20:16 -0800909 set_bit(R10BIO_Degraded, &r10_bio->state);
910 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 }
912 rcu_read_unlock();
913
Dan Williams6bfe0b42008-04-30 00:52:32 -0700914 if (unlikely(blocked_rdev)) {
915 /* Have to wait for this device to get unblocked, then retry */
916 int j;
917 int d;
918
919 for (j = 0; j < i; j++)
920 if (r10_bio->devs[j].bio) {
921 d = r10_bio->devs[j].devnum;
922 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
923 }
924 allow_barrier(conf);
925 md_wait_for_blocked_rdev(blocked_rdev, mddev);
926 wait_barrier(conf);
927 goto retry_write;
928 }
929
NeilBrown6cce3b22006-01-06 00:20:16 -0800930 atomic_set(&r10_bio->remaining, 0);
NeilBrown06d91a52005-06-21 17:17:12 -0700931
NeilBrown6cce3b22006-01-06 00:20:16 -0800932 bio_list_init(&bl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933 for (i = 0; i < conf->copies; i++) {
934 struct bio *mbio;
935 int d = r10_bio->devs[i].devnum;
936 if (!r10_bio->devs[i].bio)
937 continue;
938
939 mbio = bio_clone(bio, GFP_NOIO);
940 r10_bio->devs[i].bio = mbio;
941
942 mbio->bi_sector = r10_bio->devs[i].addr+
943 conf->mirrors[d].rdev->data_offset;
944 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
945 mbio->bi_end_io = raid10_end_write_request;
Lars Ellenberge3881a62007-01-10 23:15:37 -0800946 mbio->bi_rw = WRITE | do_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 mbio->bi_private = r10_bio;
948
949 atomic_inc(&r10_bio->remaining);
NeilBrown6cce3b22006-01-06 00:20:16 -0800950 bio_list_add(&bl, mbio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951 }
952
Arne Redlichf6f953a2007-07-31 00:37:57 -0700953 if (unlikely(!atomic_read(&r10_bio->remaining))) {
954 /* the array is dead */
955 md_write_end(mddev);
956 raid_end_bio_io(r10_bio);
957 return 0;
958 }
959
NeilBrown6cce3b22006-01-06 00:20:16 -0800960 bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
961 spin_lock_irqsave(&conf->device_lock, flags);
962 bio_list_merge(&conf->pending_bio_list, &bl);
963 blk_plug_device(mddev->queue);
964 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965
NeilBrowna35e63e2008-03-04 14:29:29 -0800966 /* In case raid10d snuck in to freeze_array */
967 wake_up(&conf->wait_barrier);
968
Lars Ellenberge3881a62007-01-10 23:15:37 -0800969 if (do_sync)
970 md_wakeup_thread(mddev->thread);
971
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972 return 0;
973}
974
975static void status(struct seq_file *seq, mddev_t *mddev)
976{
977 conf_t *conf = mddev_to_conf(mddev);
978 int i;
979
980 if (conf->near_copies < conf->raid_disks)
981 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
982 if (conf->near_copies > 1)
983 seq_printf(seq, " %d near-copies", conf->near_copies);
NeilBrownc93983b2006-06-26 00:27:41 -0700984 if (conf->far_copies > 1) {
985 if (conf->far_offset)
986 seq_printf(seq, " %d offset-copies", conf->far_copies);
987 else
988 seq_printf(seq, " %d far-copies", conf->far_copies);
989 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
NeilBrown76186dd2006-10-03 01:15:48 -0700991 conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992 for (i = 0; i < conf->raid_disks; i++)
993 seq_printf(seq, "%s",
994 conf->mirrors[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -0800995 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 seq_printf(seq, "]");
997}
998
999static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1000{
1001 char b[BDEVNAME_SIZE];
1002 conf_t *conf = mddev_to_conf(mddev);
1003
1004 /*
1005 * If it is not operational, then we have already marked it as dead
1006 * else if it is the last working disks, ignore the error, let the
1007 * next level up know.
1008 * else mark the drive as failed
1009 */
NeilBrownb2d444d2005-11-08 21:39:31 -08001010 if (test_bit(In_sync, &rdev->flags)
NeilBrown76186dd2006-10-03 01:15:48 -07001011 && conf->raid_disks-mddev->degraded == 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 /*
1013 * Don't fail the drive, just return an IO error.
1014 * The test should really be more sophisticated than
1015 * "working_disks == 1", but it isn't critical, and
1016 * can wait until we do more sophisticated "is the drive
1017 * really dead" tests...
1018 */
1019 return;
NeilBrownc04be0a2006-10-03 01:15:53 -07001020 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1021 unsigned long flags;
1022 spin_lock_irqsave(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023 mddev->degraded++;
NeilBrownc04be0a2006-10-03 01:15:53 -07001024 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025 /*
1026 * if recovery is running, make sure it aborts.
1027 */
NeilBrowndfc70642008-05-23 13:04:39 -07001028 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029 }
NeilBrownb2d444d2005-11-08 21:39:31 -08001030 set_bit(Faulty, &rdev->flags);
NeilBrown850b2b42006-10-03 01:15:46 -07001031 set_bit(MD_CHANGE_DEVS, &mddev->flags);
Nick Andrewd7a420c2008-04-28 02:15:55 -07001032 printk(KERN_ALERT "raid10: Disk failure on %s, disabling device.\n"
1033 "raid10: Operation continuing on %d devices.\n",
NeilBrown76186dd2006-10-03 01:15:48 -07001034 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035}
1036
1037static void print_conf(conf_t *conf)
1038{
1039 int i;
1040 mirror_info_t *tmp;
1041
1042 printk("RAID10 conf printout:\n");
1043 if (!conf) {
1044 printk("(!conf)\n");
1045 return;
1046 }
NeilBrown76186dd2006-10-03 01:15:48 -07001047 printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 conf->raid_disks);
1049
1050 for (i = 0; i < conf->raid_disks; i++) {
1051 char b[BDEVNAME_SIZE];
1052 tmp = conf->mirrors + i;
1053 if (tmp->rdev)
1054 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
NeilBrownb2d444d2005-11-08 21:39:31 -08001055 i, !test_bit(In_sync, &tmp->rdev->flags),
1056 !test_bit(Faulty, &tmp->rdev->flags),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057 bdevname(tmp->rdev->bdev,b));
1058 }
1059}
1060
1061static void close_sync(conf_t *conf)
1062{
NeilBrown0a27ec92006-01-06 00:20:13 -08001063 wait_barrier(conf);
1064 allow_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065
1066 mempool_destroy(conf->r10buf_pool);
1067 conf->r10buf_pool = NULL;
1068}
1069
NeilBrown6d508242005-09-09 16:24:03 -07001070/* check if there are enough drives for
1071 * every block to appear on atleast one
1072 */
1073static int enough(conf_t *conf)
1074{
1075 int first = 0;
1076
1077 do {
1078 int n = conf->copies;
1079 int cnt = 0;
1080 while (n--) {
1081 if (conf->mirrors[first].rdev)
1082 cnt++;
1083 first = (first+1) % conf->raid_disks;
1084 }
1085 if (cnt == 0)
1086 return 0;
1087 } while (first != 0);
1088 return 1;
1089}
1090
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091static int raid10_spare_active(mddev_t *mddev)
1092{
1093 int i;
1094 conf_t *conf = mddev->private;
1095 mirror_info_t *tmp;
1096
1097 /*
1098 * Find all non-in_sync disks within the RAID10 configuration
1099 * and mark them in_sync
1100 */
1101 for (i = 0; i < conf->raid_disks; i++) {
1102 tmp = conf->mirrors + i;
1103 if (tmp->rdev
NeilBrownb2d444d2005-11-08 21:39:31 -08001104 && !test_bit(Faulty, &tmp->rdev->flags)
NeilBrownc04be0a2006-10-03 01:15:53 -07001105 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1106 unsigned long flags;
1107 spin_lock_irqsave(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 mddev->degraded--;
NeilBrownc04be0a2006-10-03 01:15:53 -07001109 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 }
1111 }
1112
1113 print_conf(conf);
1114 return 0;
1115}
1116
1117
1118static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1119{
1120 conf_t *conf = mddev->private;
Neil Brown199050e2008-06-28 08:31:33 +10001121 int err = -EEXIST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 int mirror;
1123 mirror_info_t *p;
Neil Brown6c2fce22008-06-28 08:31:31 +10001124 int first = 0;
1125 int last = mddev->raid_disks - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126
1127 if (mddev->recovery_cp < MaxSector)
1128 /* only hot-add to in-sync arrays, as recovery is
1129 * very different from resync
1130 */
Neil Brown199050e2008-06-28 08:31:33 +10001131 return -EBUSY;
NeilBrown6d508242005-09-09 16:24:03 -07001132 if (!enough(conf))
Neil Brown199050e2008-06-28 08:31:33 +10001133 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134
Neil Brown6c2fce22008-06-28 08:31:31 +10001135 if (rdev->raid_disk)
1136 first = last = rdev->raid_disk;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137
NeilBrown6cce3b22006-01-06 00:20:16 -08001138 if (rdev->saved_raid_disk >= 0 &&
Neil Brown6c2fce22008-06-28 08:31:31 +10001139 rdev->saved_raid_disk >= first &&
NeilBrown6cce3b22006-01-06 00:20:16 -08001140 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1141 mirror = rdev->saved_raid_disk;
1142 else
Neil Brown6c2fce22008-06-28 08:31:31 +10001143 mirror = first;
1144 for ( ; mirror <= last ; mirror++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145 if ( !(p=conf->mirrors+mirror)->rdev) {
1146
1147 blk_queue_stack_limits(mddev->queue,
1148 rdev->bdev->bd_disk->queue);
1149 /* as we don't honour merge_bvec_fn, we must never risk
1150 * violating it, so limit ->max_sector to one PAGE, as
1151 * a one page request is never in violation.
1152 */
1153 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1154 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1155 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1156
1157 p->head_position = 0;
1158 rdev->raid_disk = mirror;
Neil Brown199050e2008-06-28 08:31:33 +10001159 err = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08001160 if (rdev->saved_raid_disk != mirror)
1161 conf->fullsync = 1;
Suzanne Woodd6065f72005-11-08 21:39:27 -08001162 rcu_assign_pointer(p->rdev, rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001163 break;
1164 }
1165
1166 print_conf(conf);
Neil Brown199050e2008-06-28 08:31:33 +10001167 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168}
1169
1170static int raid10_remove_disk(mddev_t *mddev, int number)
1171{
1172 conf_t *conf = mddev->private;
1173 int err = 0;
1174 mdk_rdev_t *rdev;
1175 mirror_info_t *p = conf->mirrors+ number;
1176
1177 print_conf(conf);
1178 rdev = p->rdev;
1179 if (rdev) {
NeilBrownb2d444d2005-11-08 21:39:31 -08001180 if (test_bit(In_sync, &rdev->flags) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181 atomic_read(&rdev->nr_pending)) {
1182 err = -EBUSY;
1183 goto abort;
1184 }
NeilBrowndfc70642008-05-23 13:04:39 -07001185 /* Only remove faulty devices in recovery
1186 * is not possible.
1187 */
1188 if (!test_bit(Faulty, &rdev->flags) &&
1189 enough(conf)) {
1190 err = -EBUSY;
1191 goto abort;
1192 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07001194 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 if (atomic_read(&rdev->nr_pending)) {
1196 /* lost the race, try later */
1197 err = -EBUSY;
1198 p->rdev = rdev;
1199 }
1200 }
1201abort:
1202
1203 print_conf(conf);
1204 return err;
1205}
1206
1207
NeilBrown6712ecf2007-09-27 12:47:43 +02001208static void end_sync_read(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1211 conf_t *conf = mddev_to_conf(r10_bio->mddev);
1212 int i,d;
1213
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214 for (i=0; i<conf->copies; i++)
1215 if (r10_bio->devs[i].bio == bio)
1216 break;
Eric Sesterhennb6385482006-04-02 13:34:29 +02001217 BUG_ON(i == conf->copies);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 update_head_pos(i, r10_bio);
1219 d = r10_bio->devs[i].devnum;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001220
1221 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1222 set_bit(R10BIO_Uptodate, &r10_bio->state);
NeilBrown4dbcdc72006-01-06 00:20:52 -08001223 else {
1224 atomic_add(r10_bio->sectors,
1225 &conf->mirrors[d].rdev->corrected_errors);
1226 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1227 md_error(r10_bio->mddev,
1228 conf->mirrors[d].rdev);
1229 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230
1231 /* for reconstruct, we always reschedule after a read.
1232 * for resync, only after all reads
1233 */
1234 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1235 atomic_dec_and_test(&r10_bio->remaining)) {
1236 /* we have read all the blocks,
1237 * do the comparison in process context in raid10d
1238 */
1239 reschedule_retry(r10_bio);
1240 }
1241 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242}
1243
NeilBrown6712ecf2007-09-27 12:47:43 +02001244static void end_sync_write(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245{
1246 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1247 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1248 mddev_t *mddev = r10_bio->mddev;
1249 conf_t *conf = mddev_to_conf(mddev);
1250 int i,d;
1251
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 for (i = 0; i < conf->copies; i++)
1253 if (r10_bio->devs[i].bio == bio)
1254 break;
1255 d = r10_bio->devs[i].devnum;
1256
1257 if (!uptodate)
1258 md_error(mddev, conf->mirrors[d].rdev);
NeilBrowndfc70642008-05-23 13:04:39 -07001259
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260 update_head_pos(i, r10_bio);
1261
1262 while (atomic_dec_and_test(&r10_bio->remaining)) {
1263 if (r10_bio->master_bio == NULL) {
1264 /* the primary of several recovery bios */
1265 md_done_sync(mddev, r10_bio->sectors, 1);
1266 put_buf(r10_bio);
1267 break;
1268 } else {
1269 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1270 put_buf(r10_bio);
1271 r10_bio = r10_bio2;
1272 }
1273 }
1274 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275}
1276
1277/*
1278 * Note: sync and recover and handled very differently for raid10
1279 * This code is for resync.
1280 * For resync, we read through virtual addresses and read all blocks.
1281 * If there is any error, we schedule a write. The lowest numbered
1282 * drive is authoritative.
1283 * However requests come for physical address, so we need to map.
1284 * For every physical address there are raid_disks/copies virtual addresses,
1285 * which is always are least one, but is not necessarly an integer.
1286 * This means that a physical address can span multiple chunks, so we may
1287 * have to submit multiple io requests for a single sync request.
1288 */
1289/*
1290 * We check if all blocks are in-sync and only write to blocks that
1291 * aren't in sync
1292 */
1293static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1294{
1295 conf_t *conf = mddev_to_conf(mddev);
1296 int i, first;
1297 struct bio *tbio, *fbio;
1298
1299 atomic_set(&r10_bio->remaining, 1);
1300
1301 /* find the first device with a block */
1302 for (i=0; i<conf->copies; i++)
1303 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1304 break;
1305
1306 if (i == conf->copies)
1307 goto done;
1308
1309 first = i;
1310 fbio = r10_bio->devs[i].bio;
1311
1312 /* now find blocks with errors */
NeilBrown0eb3ff12006-01-06 00:20:29 -08001313 for (i=0 ; i < conf->copies ; i++) {
1314 int j, d;
1315 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317 tbio = r10_bio->devs[i].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001318
1319 if (tbio->bi_end_io != end_sync_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 continue;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001321 if (i == first)
1322 continue;
1323 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1324 /* We know that the bi_io_vec layout is the same for
1325 * both 'first' and 'i', so we just compare them.
1326 * All vec entries are PAGE_SIZE;
1327 */
1328 for (j = 0; j < vcnt; j++)
1329 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1330 page_address(tbio->bi_io_vec[j].bv_page),
1331 PAGE_SIZE))
1332 break;
1333 if (j == vcnt)
1334 continue;
1335 mddev->resync_mismatches += r10_bio->sectors;
1336 }
NeilBrown18f08812006-01-06 00:20:25 -08001337 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1338 /* Don't fix anything. */
1339 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 /* Ok, we need to write this bio
1341 * First we need to fixup bv_offset, bv_len and
1342 * bi_vecs, as the read request might have corrupted these
1343 */
1344 tbio->bi_vcnt = vcnt;
1345 tbio->bi_size = r10_bio->sectors << 9;
1346 tbio->bi_idx = 0;
1347 tbio->bi_phys_segments = 0;
1348 tbio->bi_hw_segments = 0;
1349 tbio->bi_hw_front_size = 0;
1350 tbio->bi_hw_back_size = 0;
1351 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1352 tbio->bi_flags |= 1 << BIO_UPTODATE;
1353 tbio->bi_next = NULL;
1354 tbio->bi_rw = WRITE;
1355 tbio->bi_private = r10_bio;
1356 tbio->bi_sector = r10_bio->devs[i].addr;
1357
1358 for (j=0; j < vcnt ; j++) {
1359 tbio->bi_io_vec[j].bv_offset = 0;
1360 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1361
1362 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1363 page_address(fbio->bi_io_vec[j].bv_page),
1364 PAGE_SIZE);
1365 }
1366 tbio->bi_end_io = end_sync_write;
1367
1368 d = r10_bio->devs[i].devnum;
1369 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1370 atomic_inc(&r10_bio->remaining);
1371 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1372
1373 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1374 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1375 generic_make_request(tbio);
1376 }
1377
1378done:
1379 if (atomic_dec_and_test(&r10_bio->remaining)) {
1380 md_done_sync(mddev, r10_bio->sectors, 1);
1381 put_buf(r10_bio);
1382 }
1383}
1384
1385/*
1386 * Now for the recovery code.
1387 * Recovery happens across physical sectors.
1388 * We recover all non-is_sync drives by finding the virtual address of
1389 * each, and then choose a working drive that also has that virt address.
1390 * There is a separate r10_bio for each non-in_sync drive.
1391 * Only the first two slots are in use. The first for reading,
1392 * The second for writing.
1393 *
1394 */
1395
1396static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1397{
1398 conf_t *conf = mddev_to_conf(mddev);
1399 int i, d;
1400 struct bio *bio, *wbio;
1401
1402
1403 /* move the pages across to the second bio
1404 * and submit the write request
1405 */
1406 bio = r10_bio->devs[0].bio;
1407 wbio = r10_bio->devs[1].bio;
1408 for (i=0; i < wbio->bi_vcnt; i++) {
1409 struct page *p = bio->bi_io_vec[i].bv_page;
1410 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1411 wbio->bi_io_vec[i].bv_page = p;
1412 }
1413 d = r10_bio->devs[1].devnum;
1414
1415 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1416 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
NeilBrown0eb3ff12006-01-06 00:20:29 -08001417 if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1418 generic_make_request(wbio);
1419 else
NeilBrown6712ecf2007-09-27 12:47:43 +02001420 bio_endio(wbio, -EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421}
1422
1423
1424/*
1425 * This is a kernel thread which:
1426 *
1427 * 1. Retries failed read operations on working mirrors.
1428 * 2. Updates the raid superblock when problems encounter.
NeilBrown6814d532006-10-03 01:15:45 -07001429 * 3. Performs writes following reads for array synchronising.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430 */
1431
NeilBrown6814d532006-10-03 01:15:45 -07001432static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1433{
1434 int sect = 0; /* Offset from r10_bio->sector */
1435 int sectors = r10_bio->sectors;
1436 mdk_rdev_t*rdev;
1437 while(sectors) {
1438 int s = sectors;
1439 int sl = r10_bio->read_slot;
1440 int success = 0;
1441 int start;
1442
1443 if (s > (PAGE_SIZE>>9))
1444 s = PAGE_SIZE >> 9;
1445
1446 rcu_read_lock();
1447 do {
1448 int d = r10_bio->devs[sl].devnum;
1449 rdev = rcu_dereference(conf->mirrors[d].rdev);
1450 if (rdev &&
1451 test_bit(In_sync, &rdev->flags)) {
1452 atomic_inc(&rdev->nr_pending);
1453 rcu_read_unlock();
1454 success = sync_page_io(rdev->bdev,
1455 r10_bio->devs[sl].addr +
1456 sect + rdev->data_offset,
1457 s<<9,
1458 conf->tmppage, READ);
1459 rdev_dec_pending(rdev, mddev);
1460 rcu_read_lock();
1461 if (success)
1462 break;
1463 }
1464 sl++;
1465 if (sl == conf->copies)
1466 sl = 0;
1467 } while (!success && sl != r10_bio->read_slot);
1468 rcu_read_unlock();
1469
1470 if (!success) {
1471 /* Cannot read from anywhere -- bye bye array */
1472 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1473 md_error(mddev, conf->mirrors[dn].rdev);
1474 break;
1475 }
1476
1477 start = sl;
1478 /* write it back and re-read */
1479 rcu_read_lock();
1480 while (sl != r10_bio->read_slot) {
1481 int d;
1482 if (sl==0)
1483 sl = conf->copies;
1484 sl--;
1485 d = r10_bio->devs[sl].devnum;
1486 rdev = rcu_dereference(conf->mirrors[d].rdev);
1487 if (rdev &&
1488 test_bit(In_sync, &rdev->flags)) {
1489 atomic_inc(&rdev->nr_pending);
1490 rcu_read_unlock();
1491 atomic_add(s, &rdev->corrected_errors);
1492 if (sync_page_io(rdev->bdev,
1493 r10_bio->devs[sl].addr +
1494 sect + rdev->data_offset,
1495 s<<9, conf->tmppage, WRITE)
1496 == 0)
1497 /* Well, this device is dead */
1498 md_error(mddev, rdev);
1499 rdev_dec_pending(rdev, mddev);
1500 rcu_read_lock();
1501 }
1502 }
1503 sl = start;
1504 while (sl != r10_bio->read_slot) {
1505 int d;
1506 if (sl==0)
1507 sl = conf->copies;
1508 sl--;
1509 d = r10_bio->devs[sl].devnum;
1510 rdev = rcu_dereference(conf->mirrors[d].rdev);
1511 if (rdev &&
1512 test_bit(In_sync, &rdev->flags)) {
1513 char b[BDEVNAME_SIZE];
1514 atomic_inc(&rdev->nr_pending);
1515 rcu_read_unlock();
1516 if (sync_page_io(rdev->bdev,
1517 r10_bio->devs[sl].addr +
1518 sect + rdev->data_offset,
1519 s<<9, conf->tmppage, READ) == 0)
1520 /* Well, this device is dead */
1521 md_error(mddev, rdev);
1522 else
1523 printk(KERN_INFO
1524 "raid10:%s: read error corrected"
1525 " (%d sectors at %llu on %s)\n",
1526 mdname(mddev), s,
Randy Dunlap969b7552006-10-28 10:38:32 -07001527 (unsigned long long)(sect+
1528 rdev->data_offset),
NeilBrown6814d532006-10-03 01:15:45 -07001529 bdevname(rdev->bdev, b));
1530
1531 rdev_dec_pending(rdev, mddev);
1532 rcu_read_lock();
1533 }
1534 }
1535 rcu_read_unlock();
1536
1537 sectors -= s;
1538 sect += s;
1539 }
1540}
1541
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542static void raid10d(mddev_t *mddev)
1543{
1544 r10bio_t *r10_bio;
1545 struct bio *bio;
1546 unsigned long flags;
1547 conf_t *conf = mddev_to_conf(mddev);
1548 struct list_head *head = &conf->retry_list;
1549 int unplug=0;
1550 mdk_rdev_t *rdev;
1551
1552 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553
1554 for (;;) {
1555 char b[BDEVNAME_SIZE];
NeilBrowna35e63e2008-03-04 14:29:29 -08001556
1557 unplug += flush_pending_writes(conf);
1558
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrowna35e63e2008-03-04 14:29:29 -08001560 if (list_empty(head)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001561 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 break;
NeilBrowna35e63e2008-03-04 14:29:29 -08001563 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1565 list_del(head->prev);
NeilBrown4443ae12006-01-06 00:20:28 -08001566 conf->nr_queued--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567 spin_unlock_irqrestore(&conf->device_lock, flags);
1568
1569 mddev = r10_bio->mddev;
1570 conf = mddev_to_conf(mddev);
1571 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1572 sync_request_write(mddev, r10_bio);
1573 unplug = 1;
1574 } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1575 recovery_request_write(mddev, r10_bio);
1576 unplug = 1;
1577 } else {
1578 int mirror;
NeilBrown4443ae12006-01-06 00:20:28 -08001579 /* we got a read error. Maybe the drive is bad. Maybe just
1580 * the block and we can fix it.
1581 * We freeze all other IO, and try reading the block from
1582 * other devices. When we find one, we re-write
1583 * and check it that fixes the read error.
1584 * This is all done synchronously while the array is
1585 * frozen.
1586 */
NeilBrown6814d532006-10-03 01:15:45 -07001587 if (mddev->ro == 0) {
1588 freeze_array(conf);
1589 fix_read_error(conf, mddev, r10_bio);
1590 unfreeze_array(conf);
NeilBrown4443ae12006-01-06 00:20:28 -08001591 }
1592
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 bio = r10_bio->devs[r10_bio->read_slot].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001594 r10_bio->devs[r10_bio->read_slot].bio =
1595 mddev->ro ? IO_BLOCKED : NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 mirror = read_balance(conf, r10_bio);
1597 if (mirror == -1) {
1598 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1599 " read error for block %llu\n",
1600 bdevname(bio->bi_bdev,b),
1601 (unsigned long long)r10_bio->sector);
1602 raid_end_bio_io(r10_bio);
Maik Hampel14e71342007-07-31 00:37:57 -07001603 bio_put(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604 } else {
Lars Ellenberge3881a62007-01-10 23:15:37 -08001605 const int do_sync = bio_sync(r10_bio->master_bio);
Maik Hampel14e71342007-07-31 00:37:57 -07001606 bio_put(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607 rdev = conf->mirrors[mirror].rdev;
1608 if (printk_ratelimit())
1609 printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1610 " another mirror\n",
1611 bdevname(rdev->bdev,b),
1612 (unsigned long long)r10_bio->sector);
1613 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1614 r10_bio->devs[r10_bio->read_slot].bio = bio;
1615 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1616 + rdev->data_offset;
1617 bio->bi_bdev = rdev->bdev;
Lars Ellenberge3881a62007-01-10 23:15:37 -08001618 bio->bi_rw = READ | do_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 bio->bi_private = r10_bio;
1620 bio->bi_end_io = raid10_end_read_request;
1621 unplug = 1;
1622 generic_make_request(bio);
1623 }
1624 }
1625 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 if (unplug)
1627 unplug_slaves(mddev);
1628}
1629
1630
1631static int init_resync(conf_t *conf)
1632{
1633 int buffs;
1634
1635 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
Eric Sesterhennb6385482006-04-02 13:34:29 +02001636 BUG_ON(conf->r10buf_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1638 if (!conf->r10buf_pool)
1639 return -ENOMEM;
1640 conf->next_resync = 0;
1641 return 0;
1642}
1643
1644/*
1645 * perform a "sync" on one "block"
1646 *
1647 * We need to make sure that no normal I/O request - particularly write
1648 * requests - conflict with active sync requests.
1649 *
1650 * This is achieved by tracking pending requests and a 'barrier' concept
1651 * that can be installed to exclude normal IO requests.
1652 *
1653 * Resync and recovery are handled very differently.
1654 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1655 *
1656 * For resync, we iterate over virtual addresses, read all copies,
1657 * and update if there are differences. If only one copy is live,
1658 * skip it.
1659 * For recovery, we iterate over physical addresses, read a good
1660 * value for each non-in_sync drive, and over-write.
1661 *
1662 * So, for recovery we may have several outstanding complex requests for a
1663 * given address, one for each out-of-sync device. We model this by allocating
1664 * a number of r10_bio structures, one for each out-of-sync device.
1665 * As we setup these structures, we collect all bio's together into a list
1666 * which we then process collectively to add pages, and then process again
1667 * to pass to generic_make_request.
1668 *
1669 * The r10_bio structures are linked using a borrowed master_bio pointer.
1670 * This link is counted in ->remaining. When the r10_bio that points to NULL
1671 * has its remaining count decremented to 0, the whole complex operation
1672 * is complete.
1673 *
1674 */
1675
NeilBrown57afd892005-06-21 17:17:13 -07001676static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677{
1678 conf_t *conf = mddev_to_conf(mddev);
1679 r10bio_t *r10_bio;
1680 struct bio *biolist = NULL, *bio;
1681 sector_t max_sector, nr_sectors;
1682 int disk;
1683 int i;
NeilBrown6cce3b22006-01-06 00:20:16 -08001684 int max_sync;
1685 int sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
1687 sector_t sectors_skipped = 0;
1688 int chunks_skipped = 0;
1689
1690 if (!conf->r10buf_pool)
1691 if (init_resync(conf))
NeilBrown57afd892005-06-21 17:17:13 -07001692 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693
1694 skipped:
1695 max_sector = mddev->size << 1;
1696 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1697 max_sector = mddev->resync_max_sectors;
1698 if (sector_nr >= max_sector) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001699 /* If we aborted, we need to abort the
1700 * sync on the 'current' bitmap chucks (there can
1701 * be several when recovering multiple devices).
1702 * as we may have started syncing it but not finished.
1703 * We can find the current address in
1704 * mddev->curr_resync, but for recovery,
1705 * we need to convert that to several
1706 * virtual addresses.
1707 */
1708 if (mddev->curr_resync < max_sector) { /* aborted */
1709 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1710 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1711 &sync_blocks, 1);
1712 else for (i=0; i<conf->raid_disks; i++) {
1713 sector_t sect =
1714 raid10_find_virt(conf, mddev->curr_resync, i);
1715 bitmap_end_sync(mddev->bitmap, sect,
1716 &sync_blocks, 1);
1717 }
1718 } else /* completed sync */
1719 conf->fullsync = 0;
1720
1721 bitmap_close_sync(mddev->bitmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722 close_sync(conf);
NeilBrown57afd892005-06-21 17:17:13 -07001723 *skipped = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 return sectors_skipped;
1725 }
1726 if (chunks_skipped >= conf->raid_disks) {
1727 /* if there has been nothing to do on any drive,
1728 * then there is nothing to do at all..
1729 */
NeilBrown57afd892005-06-21 17:17:13 -07001730 *skipped = 1;
1731 return (max_sector - sector_nr) + sectors_skipped;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 }
1733
NeilBrownc6207272008-02-06 01:39:52 -08001734 if (max_sector > mddev->resync_max)
1735 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1736
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 /* make sure whole request will fit in a chunk - if chunks
1738 * are meaningful
1739 */
1740 if (conf->near_copies < conf->raid_disks &&
1741 max_sector > (sector_nr | conf->chunk_mask))
1742 max_sector = (sector_nr | conf->chunk_mask) + 1;
1743 /*
1744 * If there is non-resync activity waiting for us then
1745 * put in a delay to throttle resync.
1746 */
NeilBrown0a27ec92006-01-06 00:20:13 -08001747 if (!go_faster && conf->nr_waiting)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748 msleep_interruptible(1000);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749
NeilBrownb47490c2008-02-06 01:39:50 -08001750 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1751
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752 /* Again, very different code for resync and recovery.
1753 * Both must result in an r10bio with a list of bios that
1754 * have bi_end_io, bi_sector, bi_bdev set,
1755 * and bi_private set to the r10bio.
1756 * For recovery, we may actually create several r10bios
1757 * with 2 bios in each, that correspond to the bios in the main one.
1758 * In this case, the subordinate r10bios link back through a
1759 * borrowed master_bio pointer, and the counter in the master
1760 * includes a ref from each subordinate.
1761 */
1762 /* First, we decide what to do and set ->bi_end_io
1763 * To end_sync_read if we want to read, and
1764 * end_sync_write if we will want to write.
1765 */
1766
NeilBrown6cce3b22006-01-06 00:20:16 -08001767 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1769 /* recovery... the complicated one */
1770 int i, j, k;
1771 r10_bio = NULL;
1772
1773 for (i=0 ; i<conf->raid_disks; i++)
1774 if (conf->mirrors[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001775 !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001776 int still_degraded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 /* want to reconstruct this device */
1778 r10bio_t *rb2 = r10_bio;
NeilBrown6cce3b22006-01-06 00:20:16 -08001779 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1780 int must_sync;
1781 /* Unless we are doing a full sync, we only need
1782 * to recover the block if it is set in the bitmap
1783 */
1784 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1785 &sync_blocks, 1);
1786 if (sync_blocks < max_sync)
1787 max_sync = sync_blocks;
1788 if (!must_sync &&
1789 !conf->fullsync) {
1790 /* yep, skip the sync_blocks here, but don't assume
1791 * that there will never be anything to do here
1792 */
1793 chunks_skipped = -1;
1794 continue;
1795 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796
1797 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
NeilBrown6cce3b22006-01-06 00:20:16 -08001798 raise_barrier(conf, rb2 != NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 atomic_set(&r10_bio->remaining, 0);
1800
1801 r10_bio->master_bio = (struct bio*)rb2;
1802 if (rb2)
1803 atomic_inc(&rb2->remaining);
1804 r10_bio->mddev = mddev;
1805 set_bit(R10BIO_IsRecover, &r10_bio->state);
NeilBrown6cce3b22006-01-06 00:20:16 -08001806 r10_bio->sector = sect;
1807
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808 raid10_find_phys(conf, r10_bio);
NeilBrown6cce3b22006-01-06 00:20:16 -08001809 /* Need to check if this section will still be
1810 * degraded
1811 */
1812 for (j=0; j<conf->copies;j++) {
1813 int d = r10_bio->devs[j].devnum;
1814 if (conf->mirrors[d].rdev == NULL ||
NeilBrowna24a8dd2006-01-06 00:20:35 -08001815 test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001816 still_degraded = 1;
NeilBrowna24a8dd2006-01-06 00:20:35 -08001817 break;
1818 }
NeilBrown6cce3b22006-01-06 00:20:16 -08001819 }
1820 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1821 &sync_blocks, still_degraded);
1822
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823 for (j=0; j<conf->copies;j++) {
1824 int d = r10_bio->devs[j].devnum;
1825 if (conf->mirrors[d].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001826 test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 /* This is where we read from */
1828 bio = r10_bio->devs[0].bio;
1829 bio->bi_next = biolist;
1830 biolist = bio;
1831 bio->bi_private = r10_bio;
1832 bio->bi_end_io = end_sync_read;
NeilBrown802ba062006-12-13 00:34:13 -08001833 bio->bi_rw = READ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834 bio->bi_sector = r10_bio->devs[j].addr +
1835 conf->mirrors[d].rdev->data_offset;
1836 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1837 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1838 atomic_inc(&r10_bio->remaining);
1839 /* and we write to 'i' */
1840
1841 for (k=0; k<conf->copies; k++)
1842 if (r10_bio->devs[k].devnum == i)
1843 break;
NeilBrown64a742b2007-02-28 20:11:18 -08001844 BUG_ON(k == conf->copies);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845 bio = r10_bio->devs[1].bio;
1846 bio->bi_next = biolist;
1847 biolist = bio;
1848 bio->bi_private = r10_bio;
1849 bio->bi_end_io = end_sync_write;
NeilBrown802ba062006-12-13 00:34:13 -08001850 bio->bi_rw = WRITE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851 bio->bi_sector = r10_bio->devs[k].addr +
1852 conf->mirrors[i].rdev->data_offset;
1853 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1854
1855 r10_bio->devs[0].devnum = d;
1856 r10_bio->devs[1].devnum = i;
1857
1858 break;
1859 }
1860 }
1861 if (j == conf->copies) {
NeilBrown87fc7672005-09-09 16:24:04 -07001862 /* Cannot recover, so abort the recovery */
1863 put_buf(r10_bio);
K.Tanakaa07e6ab2008-03-04 14:29:37 -08001864 if (rb2)
1865 atomic_dec(&rb2->remaining);
NeilBrown87fc7672005-09-09 16:24:04 -07001866 r10_bio = rb2;
NeilBrowndfc70642008-05-23 13:04:39 -07001867 if (!test_and_set_bit(MD_RECOVERY_INTR,
1868 &mddev->recovery))
NeilBrown87fc7672005-09-09 16:24:04 -07001869 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1870 mdname(mddev));
1871 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872 }
1873 }
1874 if (biolist == NULL) {
1875 while (r10_bio) {
1876 r10bio_t *rb2 = r10_bio;
1877 r10_bio = (r10bio_t*) rb2->master_bio;
1878 rb2->master_bio = NULL;
1879 put_buf(rb2);
1880 }
1881 goto giveup;
1882 }
1883 } else {
1884 /* resync. Schedule a read for every block at this virt offset */
1885 int count = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08001886
1887 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1888 &sync_blocks, mddev->degraded) &&
1889 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1890 /* We can skip this block */
1891 *skipped = 1;
1892 return sync_blocks + sectors_skipped;
1893 }
1894 if (sync_blocks < max_sync)
1895 max_sync = sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1897
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 r10_bio->mddev = mddev;
1899 atomic_set(&r10_bio->remaining, 0);
NeilBrown6cce3b22006-01-06 00:20:16 -08001900 raise_barrier(conf, 0);
1901 conf->next_resync = sector_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902
1903 r10_bio->master_bio = NULL;
1904 r10_bio->sector = sector_nr;
1905 set_bit(R10BIO_IsSync, &r10_bio->state);
1906 raid10_find_phys(conf, r10_bio);
1907 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1908
1909 for (i=0; i<conf->copies; i++) {
1910 int d = r10_bio->devs[i].devnum;
1911 bio = r10_bio->devs[i].bio;
1912 bio->bi_end_io = NULL;
NeilBrownaf03b8e2007-06-16 10:16:06 -07001913 clear_bit(BIO_UPTODATE, &bio->bi_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914 if (conf->mirrors[d].rdev == NULL ||
NeilBrownb2d444d2005-11-08 21:39:31 -08001915 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916 continue;
1917 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1918 atomic_inc(&r10_bio->remaining);
1919 bio->bi_next = biolist;
1920 biolist = bio;
1921 bio->bi_private = r10_bio;
1922 bio->bi_end_io = end_sync_read;
NeilBrown802ba062006-12-13 00:34:13 -08001923 bio->bi_rw = READ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924 bio->bi_sector = r10_bio->devs[i].addr +
1925 conf->mirrors[d].rdev->data_offset;
1926 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1927 count++;
1928 }
1929
1930 if (count < 2) {
1931 for (i=0; i<conf->copies; i++) {
1932 int d = r10_bio->devs[i].devnum;
1933 if (r10_bio->devs[i].bio->bi_end_io)
1934 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1935 }
1936 put_buf(r10_bio);
1937 biolist = NULL;
1938 goto giveup;
1939 }
1940 }
1941
1942 for (bio = biolist; bio ; bio=bio->bi_next) {
1943
1944 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1945 if (bio->bi_end_io)
1946 bio->bi_flags |= 1 << BIO_UPTODATE;
1947 bio->bi_vcnt = 0;
1948 bio->bi_idx = 0;
1949 bio->bi_phys_segments = 0;
1950 bio->bi_hw_segments = 0;
1951 bio->bi_size = 0;
1952 }
1953
1954 nr_sectors = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08001955 if (sector_nr + max_sync < max_sector)
1956 max_sector = sector_nr + max_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957 do {
1958 struct page *page;
1959 int len = PAGE_SIZE;
1960 disk = 0;
1961 if (sector_nr + (len>>9) > max_sector)
1962 len = (max_sector - sector_nr) << 9;
1963 if (len == 0)
1964 break;
1965 for (bio= biolist ; bio ; bio=bio->bi_next) {
1966 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1967 if (bio_add_page(bio, page, len, 0) == 0) {
1968 /* stop here */
1969 struct bio *bio2;
1970 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1971 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1972 /* remove last page from this bio */
1973 bio2->bi_vcnt--;
1974 bio2->bi_size -= len;
1975 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1976 }
1977 goto bio_full;
1978 }
1979 disk = i;
1980 }
1981 nr_sectors += len>>9;
1982 sector_nr += len>>9;
1983 } while (biolist->bi_vcnt < RESYNC_PAGES);
1984 bio_full:
1985 r10_bio->sectors = nr_sectors;
1986
1987 while (biolist) {
1988 bio = biolist;
1989 biolist = biolist->bi_next;
1990
1991 bio->bi_next = NULL;
1992 r10_bio = bio->bi_private;
1993 r10_bio->sectors = nr_sectors;
1994
1995 if (bio->bi_end_io == end_sync_read) {
1996 md_sync_acct(bio->bi_bdev, nr_sectors);
1997 generic_make_request(bio);
1998 }
1999 }
2000
NeilBrown57afd892005-06-21 17:17:13 -07002001 if (sectors_skipped)
2002 /* pretend they weren't skipped, it makes
2003 * no important difference in this case
2004 */
2005 md_done_sync(mddev, sectors_skipped, 1);
2006
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007 return sectors_skipped + nr_sectors;
2008 giveup:
2009 /* There is nowhere to write, so all non-sync
2010 * drives must be failed, so try the next chunk...
2011 */
2012 {
NeilBrown57afd892005-06-21 17:17:13 -07002013 sector_t sec = max_sector - sector_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014 sectors_skipped += sec;
2015 chunks_skipped ++;
2016 sector_nr = max_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002017 goto skipped;
2018 }
2019}
2020
2021static int run(mddev_t *mddev)
2022{
2023 conf_t *conf;
2024 int i, disk_idx;
2025 mirror_info_t *disk;
2026 mdk_rdev_t *rdev;
2027 struct list_head *tmp;
NeilBrownc93983b2006-06-26 00:27:41 -07002028 int nc, fc, fo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029 sector_t stride, size;
2030
NeilBrown2604b702006-01-06 00:20:36 -08002031 if (mddev->chunk_size == 0) {
2032 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
2033 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 }
NeilBrown2604b702006-01-06 00:20:36 -08002035
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 nc = mddev->layout & 255;
2037 fc = (mddev->layout >> 8) & 255;
NeilBrownc93983b2006-06-26 00:27:41 -07002038 fo = mddev->layout & (1<<16);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
NeilBrownc93983b2006-06-26 00:27:41 -07002040 (mddev->layout >> 17)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
2042 mdname(mddev), mddev->layout);
2043 goto out;
2044 }
2045 /*
2046 * copy the already verified devices into our private RAID10
2047 * bookkeeping area. [whatever we allocate in run(),
2048 * should be freed in stop()]
2049 */
NeilBrown4443ae12006-01-06 00:20:28 -08002050 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051 mddev->private = conf;
2052 if (!conf) {
2053 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2054 mdname(mddev));
2055 goto out;
2056 }
NeilBrown4443ae12006-01-06 00:20:28 -08002057 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058 GFP_KERNEL);
2059 if (!conf->mirrors) {
2060 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2061 mdname(mddev));
2062 goto out_free_conf;
2063 }
NeilBrown4443ae12006-01-06 00:20:28 -08002064
2065 conf->tmppage = alloc_page(GFP_KERNEL);
2066 if (!conf->tmppage)
2067 goto out_free_conf;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068
NeilBrown64a742b2007-02-28 20:11:18 -08002069 conf->mddev = mddev;
2070 conf->raid_disks = mddev->raid_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071 conf->near_copies = nc;
2072 conf->far_copies = fc;
2073 conf->copies = nc*fc;
NeilBrownc93983b2006-06-26 00:27:41 -07002074 conf->far_offset = fo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075 conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
2076 conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
NeilBrown64a742b2007-02-28 20:11:18 -08002077 size = mddev->size >> (conf->chunk_shift-1);
2078 sector_div(size, fc);
2079 size = size * conf->raid_disks;
2080 sector_div(size, nc);
2081 /* 'size' is now the number of chunks in the array */
2082 /* calculate "used chunks per device" in 'stride' */
2083 stride = size * conf->copies;
NeilBrownaf03b8e2007-06-16 10:16:06 -07002084
2085 /* We need to round up when dividing by raid_disks to
2086 * get the stride size.
2087 */
2088 stride += conf->raid_disks - 1;
NeilBrown64a742b2007-02-28 20:11:18 -08002089 sector_div(stride, conf->raid_disks);
2090 mddev->size = stride << (conf->chunk_shift-1);
2091
NeilBrownc93983b2006-06-26 00:27:41 -07002092 if (fo)
NeilBrown64a742b2007-02-28 20:11:18 -08002093 stride = 1;
2094 else
NeilBrownc93983b2006-06-26 00:27:41 -07002095 sector_div(stride, fc);
NeilBrown64a742b2007-02-28 20:11:18 -08002096 conf->stride = stride << conf->chunk_shift;
2097
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2099 r10bio_pool_free, conf);
2100 if (!conf->r10bio_pool) {
2101 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2102 mdname(mddev));
2103 goto out_free_conf;
2104 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105
Neil Browne7e72bf2008-05-14 16:05:54 -07002106 spin_lock_init(&conf->device_lock);
2107 mddev->queue->queue_lock = &conf->device_lock;
2108
NeilBrownd089c6a2008-02-06 01:39:59 -08002109 rdev_for_each(rdev, tmp, mddev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110 disk_idx = rdev->raid_disk;
2111 if (disk_idx >= mddev->raid_disks
2112 || disk_idx < 0)
2113 continue;
2114 disk = conf->mirrors + disk_idx;
2115
2116 disk->rdev = rdev;
2117
2118 blk_queue_stack_limits(mddev->queue,
2119 rdev->bdev->bd_disk->queue);
2120 /* as we don't honour merge_bvec_fn, we must never risk
2121 * violating it, so limit ->max_sector to one PAGE, as
2122 * a one page request is never in violation.
2123 */
2124 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2125 mddev->queue->max_sectors > (PAGE_SIZE>>9))
2126 mddev->queue->max_sectors = (PAGE_SIZE>>9);
2127
2128 disk->head_position = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130 INIT_LIST_HEAD(&conf->retry_list);
2131
2132 spin_lock_init(&conf->resync_lock);
NeilBrown0a27ec92006-01-06 00:20:13 -08002133 init_waitqueue_head(&conf->wait_barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134
NeilBrown6d508242005-09-09 16:24:03 -07002135 /* need to check that every block has at least one working mirror */
2136 if (!enough(conf)) {
2137 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2138 mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139 goto out_free_conf;
2140 }
2141
2142 mddev->degraded = 0;
2143 for (i = 0; i < conf->raid_disks; i++) {
2144
2145 disk = conf->mirrors + i;
2146
NeilBrown5fd6c1d2006-06-26 00:27:40 -07002147 if (!disk->rdev ||
NeilBrown2e333e82006-10-21 10:24:07 -07002148 !test_bit(In_sync, &disk->rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 disk->head_position = 0;
2150 mddev->degraded++;
Neil Brown8c2e8702008-06-28 08:30:52 +10002151 if (disk->rdev)
2152 conf->fullsync = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153 }
2154 }
2155
2156
2157 mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2158 if (!mddev->thread) {
2159 printk(KERN_ERR
2160 "raid10: couldn't allocate thread for %s\n",
2161 mdname(mddev));
2162 goto out_free_conf;
2163 }
2164
2165 printk(KERN_INFO
2166 "raid10: raid set %s active with %d out of %d devices\n",
2167 mdname(mddev), mddev->raid_disks - mddev->degraded,
2168 mddev->raid_disks);
2169 /*
2170 * Ok, everything is just fine now
2171 */
Andre Nollf233ea52008-07-21 17:05:22 +10002172 mddev->array_sectors = size << conf->chunk_shift;
NeilBrown64a742b2007-02-28 20:11:18 -08002173 mddev->resync_max_sectors = size << conf->chunk_shift;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174
NeilBrown7a5febe2005-05-16 21:53:16 -07002175 mddev->queue->unplug_fn = raid10_unplug;
NeilBrown0d129222006-10-03 01:15:54 -07002176 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2177 mddev->queue->backing_dev_info.congested_data = mddev;
NeilBrown7a5febe2005-05-16 21:53:16 -07002178
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179 /* Calculate max read-ahead size.
2180 * We need to readahead at least twice a whole stripe....
2181 * maybe...
2182 */
2183 {
NeilBrown8932c2e2006-06-26 00:27:36 -07002184 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 stripe /= conf->near_copies;
2186 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2187 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2188 }
2189
2190 if (conf->near_copies < mddev->raid_disks)
2191 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2192 return 0;
2193
2194out_free_conf:
2195 if (conf->r10bio_pool)
2196 mempool_destroy(conf->r10bio_pool);
NeilBrown1345b1d2006-01-06 00:20:40 -08002197 safe_put_page(conf->tmppage);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07002198 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199 kfree(conf);
2200 mddev->private = NULL;
2201out:
2202 return -EIO;
2203}
2204
2205static int stop(mddev_t *mddev)
2206{
2207 conf_t *conf = mddev_to_conf(mddev);
2208
2209 md_unregister_thread(mddev->thread);
2210 mddev->thread = NULL;
2211 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2212 if (conf->r10bio_pool)
2213 mempool_destroy(conf->r10bio_pool);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07002214 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 kfree(conf);
2216 mddev->private = NULL;
2217 return 0;
2218}
2219
NeilBrown6cce3b22006-01-06 00:20:16 -08002220static void raid10_quiesce(mddev_t *mddev, int state)
2221{
2222 conf_t *conf = mddev_to_conf(mddev);
2223
2224 switch(state) {
2225 case 1:
2226 raise_barrier(conf, 0);
2227 break;
2228 case 0:
2229 lower_barrier(conf);
2230 break;
2231 }
2232 if (mddev->thread) {
2233 if (mddev->bitmap)
2234 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2235 else
2236 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2237 md_wakeup_thread(mddev->thread);
2238 }
2239}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240
NeilBrown2604b702006-01-06 00:20:36 -08002241static struct mdk_personality raid10_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242{
2243 .name = "raid10",
NeilBrown2604b702006-01-06 00:20:36 -08002244 .level = 10,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245 .owner = THIS_MODULE,
2246 .make_request = make_request,
2247 .run = run,
2248 .stop = stop,
2249 .status = status,
2250 .error_handler = error,
2251 .hot_add_disk = raid10_add_disk,
2252 .hot_remove_disk= raid10_remove_disk,
2253 .spare_active = raid10_spare_active,
2254 .sync_request = sync_request,
NeilBrown6cce3b22006-01-06 00:20:16 -08002255 .quiesce = raid10_quiesce,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256};
2257
2258static int __init raid_init(void)
2259{
NeilBrown2604b702006-01-06 00:20:36 -08002260 return register_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261}
2262
2263static void raid_exit(void)
2264{
NeilBrown2604b702006-01-06 00:20:36 -08002265 unregister_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266}
2267
2268module_init(raid_init);
2269module_exit(raid_exit);
2270MODULE_LICENSE("GPL");
2271MODULE_ALIAS("md-personality-9"); /* RAID10 */
NeilBrownd9d166c2006-01-06 00:20:51 -08002272MODULE_ALIAS("md-raid10");
NeilBrown2604b702006-01-06 00:20:36 -08002273MODULE_ALIAS("md-level-10");