blob: 017f58113c33604e381a5fcc7c75f01344f67841 [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
79#define RESYNC_BLOCK_SIZE (64*1024)
80//#define RESYNC_BLOCK_SIZE PAGE_SIZE
81#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
82#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
83#define RESYNC_WINDOW (2048*1024)
84
85/*
86 * When performing a resync, we need to read and compare, so
87 * we need as many pages are there are copies.
88 * When performing a recovery, we need 2 bios, one for read,
89 * one for write (we recover only one drive per r10buf)
90 *
91 */
Al Virodd0fc662005-10-07 07:46:04 +010092static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070093{
94 conf_t *conf = data;
95 struct page *page;
96 r10bio_t *r10_bio;
97 struct bio *bio;
98 int i, j;
99 int nalloc;
100
101 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
102 if (!r10_bio) {
103 unplug_slaves(conf->mddev);
104 return NULL;
105 }
106
107 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
108 nalloc = conf->copies; /* resync */
109 else
110 nalloc = 2; /* recovery */
111
112 /*
113 * Allocate bios.
114 */
115 for (j = nalloc ; j-- ; ) {
116 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
117 if (!bio)
118 goto out_free_bio;
119 r10_bio->devs[j].bio = bio;
120 }
121 /*
122 * Allocate RESYNC_PAGES data pages and attach them
123 * where needed.
124 */
125 for (j = 0 ; j < nalloc; j++) {
126 bio = r10_bio->devs[j].bio;
127 for (i = 0; i < RESYNC_PAGES; i++) {
128 page = alloc_page(gfp_flags);
129 if (unlikely(!page))
130 goto out_free_pages;
131
132 bio->bi_io_vec[i].bv_page = page;
133 }
134 }
135
136 return r10_bio;
137
138out_free_pages:
139 for ( ; i > 0 ; i--)
NeilBrown1345b1d2006-01-06 00:20:40 -0800140 safe_put_page(bio->bi_io_vec[i-1].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141 while (j--)
142 for (i = 0; i < RESYNC_PAGES ; i++)
NeilBrown1345b1d2006-01-06 00:20:40 -0800143 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 j = -1;
145out_free_bio:
146 while ( ++j < nalloc )
147 bio_put(r10_bio->devs[j].bio);
148 r10bio_pool_free(r10_bio, conf);
149 return NULL;
150}
151
152static void r10buf_pool_free(void *__r10_bio, void *data)
153{
154 int i;
155 conf_t *conf = data;
156 r10bio_t *r10bio = __r10_bio;
157 int j;
158
159 for (j=0; j < conf->copies; j++) {
160 struct bio *bio = r10bio->devs[j].bio;
161 if (bio) {
162 for (i = 0; i < RESYNC_PAGES; i++) {
NeilBrown1345b1d2006-01-06 00:20:40 -0800163 safe_put_page(bio->bi_io_vec[i].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164 bio->bi_io_vec[i].bv_page = NULL;
165 }
166 bio_put(bio);
167 }
168 }
169 r10bio_pool_free(r10bio, conf);
170}
171
172static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
173{
174 int i;
175
176 for (i = 0; i < conf->copies; i++) {
177 struct bio **bio = & r10_bio->devs[i].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -0800178 if (*bio && *bio != IO_BLOCKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179 bio_put(*bio);
180 *bio = NULL;
181 }
182}
183
Arjan van de Ven858119e2006-01-14 13:20:43 -0800184static void free_r10bio(r10bio_t *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186 conf_t *conf = mddev_to_conf(r10_bio->mddev);
187
188 /*
189 * Wake up any possible resync thread that waits for the device
190 * to go idle.
191 */
NeilBrown0a27ec92006-01-06 00:20:13 -0800192 allow_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193
194 put_all_bios(conf, r10_bio);
195 mempool_free(r10_bio, conf->r10bio_pool);
196}
197
Arjan van de Ven858119e2006-01-14 13:20:43 -0800198static void put_buf(r10bio_t *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199{
200 conf_t *conf = mddev_to_conf(r10_bio->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201
202 mempool_free(r10_bio, conf->r10buf_pool);
203
NeilBrown0a27ec92006-01-06 00:20:13 -0800204 lower_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205}
206
207static void reschedule_retry(r10bio_t *r10_bio)
208{
209 unsigned long flags;
210 mddev_t *mddev = r10_bio->mddev;
211 conf_t *conf = mddev_to_conf(mddev);
212
213 spin_lock_irqsave(&conf->device_lock, flags);
214 list_add(&r10_bio->retry_list, &conf->retry_list);
NeilBrown4443ae12006-01-06 00:20:28 -0800215 conf->nr_queued ++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216 spin_unlock_irqrestore(&conf->device_lock, flags);
217
218 md_wakeup_thread(mddev->thread);
219}
220
221/*
222 * raid_end_bio_io() is called when we have finished servicing a mirrored
223 * operation and are ready to return a success/failure code to the buffer
224 * cache layer.
225 */
226static void raid_end_bio_io(r10bio_t *r10_bio)
227{
228 struct bio *bio = r10_bio->master_bio;
229
NeilBrown6712ecf2007-09-27 12:47:43 +0200230 bio_endio(bio,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
232 free_r10bio(r10_bio);
233}
234
235/*
236 * Update disk head position estimator based on IRQ completion info.
237 */
238static inline void update_head_pos(int slot, r10bio_t *r10_bio)
239{
240 conf_t *conf = mddev_to_conf(r10_bio->mddev);
241
242 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
243 r10_bio->devs[slot].addr + (r10_bio->sectors);
244}
245
NeilBrown6712ecf2007-09-27 12:47:43 +0200246static void raid10_end_read_request(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247{
248 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
249 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
250 int slot, dev;
251 conf_t *conf = mddev_to_conf(r10_bio->mddev);
252
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253
254 slot = r10_bio->read_slot;
255 dev = r10_bio->devs[slot].devnum;
256 /*
257 * this branch is our 'one mirror IO has finished' event handler:
258 */
NeilBrown4443ae12006-01-06 00:20:28 -0800259 update_head_pos(slot, r10_bio);
260
261 if (uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262 /*
263 * Set R10BIO_Uptodate in our master bio, so that
264 * we will return a good error code to the higher
265 * levels even if IO on some other mirrored buffer fails.
266 *
267 * The 'master' represents the composite IO operation to
268 * user-side. So if something waits for IO, then it will
269 * wait for the 'master' bio.
270 */
271 set_bit(R10BIO_Uptodate, &r10_bio->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 raid_end_bio_io(r10_bio);
NeilBrown4443ae12006-01-06 00:20:28 -0800273 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274 /*
275 * oops, read error:
276 */
277 char b[BDEVNAME_SIZE];
278 if (printk_ratelimit())
279 printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
280 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
281 reschedule_retry(r10_bio);
282 }
283
284 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285}
286
NeilBrown6712ecf2007-09-27 12:47:43 +0200287static void raid10_end_write_request(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288{
289 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
290 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
291 int slot, dev;
292 conf_t *conf = mddev_to_conf(r10_bio->mddev);
293
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294 for (slot = 0; slot < conf->copies; slot++)
295 if (r10_bio->devs[slot].bio == bio)
296 break;
297 dev = r10_bio->devs[slot].devnum;
298
299 /*
300 * this branch is our 'one mirror IO has finished' event handler:
301 */
NeilBrown6cce3b22006-01-06 00:20:16 -0800302 if (!uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
NeilBrown6cce3b22006-01-06 00:20:16 -0800304 /* an I/O failed, we can't clear the bitmap */
305 set_bit(R10BIO_Degraded, &r10_bio->state);
306 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 /*
308 * Set R10BIO_Uptodate in our master bio, so that
309 * we will return a good error code for to the higher
310 * levels even if IO on some other mirrored buffer fails.
311 *
312 * The 'master' represents the composite IO operation to
313 * user-side. So if something waits for IO, then it will
314 * wait for the 'master' bio.
315 */
316 set_bit(R10BIO_Uptodate, &r10_bio->state);
317
318 update_head_pos(slot, r10_bio);
319
320 /*
321 *
322 * Let's see if all mirrored write operations have finished
323 * already.
324 */
325 if (atomic_dec_and_test(&r10_bio->remaining)) {
NeilBrown6cce3b22006-01-06 00:20:16 -0800326 /* clear the bitmap if all writes complete successfully */
327 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
328 r10_bio->sectors,
329 !test_bit(R10BIO_Degraded, &r10_bio->state),
330 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331 md_write_end(r10_bio->mddev);
332 raid_end_bio_io(r10_bio);
333 }
334
335 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336}
337
338
339/*
340 * RAID10 layout manager
341 * Aswell as the chunksize and raid_disks count, there are two
342 * parameters: near_copies and far_copies.
343 * near_copies * far_copies must be <= raid_disks.
344 * Normally one of these will be 1.
345 * If both are 1, we get raid0.
346 * If near_copies == raid_disks, we get raid1.
347 *
348 * Chunks are layed out in raid0 style with near_copies copies of the
349 * first chunk, followed by near_copies copies of the next chunk and
350 * so on.
351 * If far_copies > 1, then after 1/far_copies of the array has been assigned
352 * as described above, we start again with a device offset of near_copies.
353 * So we effectively have another copy of the whole array further down all
354 * the drives, but with blocks on different drives.
355 * With this layout, and block is never stored twice on the one device.
356 *
357 * raid10_find_phys finds the sector offset of a given virtual sector
NeilBrownc93983b2006-06-26 00:27:41 -0700358 * on each device that it is on.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 *
360 * raid10_find_virt does the reverse mapping, from a device and a
361 * sector offset to a virtual address
362 */
363
364static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
365{
366 int n,f;
367 sector_t sector;
368 sector_t chunk;
369 sector_t stripe;
370 int dev;
371
372 int slot = 0;
373
374 /* now calculate first sector/dev */
375 chunk = r10bio->sector >> conf->chunk_shift;
376 sector = r10bio->sector & conf->chunk_mask;
377
378 chunk *= conf->near_copies;
379 stripe = chunk;
380 dev = sector_div(stripe, conf->raid_disks);
NeilBrownc93983b2006-06-26 00:27:41 -0700381 if (conf->far_offset)
382 stripe *= conf->far_copies;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383
384 sector += stripe << conf->chunk_shift;
385
386 /* and calculate all the others */
387 for (n=0; n < conf->near_copies; n++) {
388 int d = dev;
389 sector_t s = sector;
390 r10bio->devs[slot].addr = sector;
391 r10bio->devs[slot].devnum = d;
392 slot++;
393
394 for (f = 1; f < conf->far_copies; f++) {
395 d += conf->near_copies;
396 if (d >= conf->raid_disks)
397 d -= conf->raid_disks;
398 s += conf->stride;
399 r10bio->devs[slot].devnum = d;
400 r10bio->devs[slot].addr = s;
401 slot++;
402 }
403 dev++;
404 if (dev >= conf->raid_disks) {
405 dev = 0;
406 sector += (conf->chunk_mask + 1);
407 }
408 }
409 BUG_ON(slot != conf->copies);
410}
411
412static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
413{
414 sector_t offset, chunk, vchunk;
415
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 offset = sector & conf->chunk_mask;
NeilBrownc93983b2006-06-26 00:27:41 -0700417 if (conf->far_offset) {
418 int fc;
419 chunk = sector >> conf->chunk_shift;
420 fc = sector_div(chunk, conf->far_copies);
421 dev -= fc * conf->near_copies;
422 if (dev < 0)
423 dev += conf->raid_disks;
424 } else {
NeilBrown64a742b2007-02-28 20:11:18 -0800425 while (sector >= conf->stride) {
NeilBrownc93983b2006-06-26 00:27:41 -0700426 sector -= conf->stride;
427 if (dev < conf->near_copies)
428 dev += conf->raid_disks - conf->near_copies;
429 else
430 dev -= conf->near_copies;
431 }
432 chunk = sector >> conf->chunk_shift;
433 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434 vchunk = chunk * conf->raid_disks + dev;
435 sector_div(vchunk, conf->near_copies);
436 return (vchunk << conf->chunk_shift) + offset;
437}
438
439/**
440 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
441 * @q: request queue
442 * @bio: the buffer head that's been built up so far
443 * @biovec: the request that could be merged to it.
444 *
445 * Return amount of bytes we can accept at this offset
446 * If near_copies == raid_disk, there are no striping issues,
447 * but in that case, the function isn't called at all.
448 */
Jens Axboe165125e2007-07-24 09:28:11 +0200449static int raid10_mergeable_bvec(struct request_queue *q, struct bio *bio,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 struct bio_vec *bio_vec)
451{
452 mddev_t *mddev = q->queuedata;
453 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
454 int max;
455 unsigned int chunk_sectors = mddev->chunk_size >> 9;
456 unsigned int bio_sectors = bio->bi_size >> 9;
457
458 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
459 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
460 if (max <= bio_vec->bv_len && bio_sectors == 0)
461 return bio_vec->bv_len;
462 else
463 return max;
464}
465
466/*
467 * This routine returns the disk from which the requested read should
468 * be done. There is a per-array 'next expected sequential IO' sector
469 * number - if this matches on the next IO then we use the last disk.
470 * There is also a per-disk 'last know head position' sector that is
471 * maintained from IRQ contexts, both the normal and the resync IO
472 * completion handlers update this position correctly. If there is no
473 * perfect sequential match then we pick the disk whose head is closest.
474 *
475 * If there are 2 mirrors in the same 2 devices, performance degrades
476 * because position is mirror, not device based.
477 *
478 * The rdev for the device selected will have nr_pending incremented.
479 */
480
481/*
482 * FIXME: possibly should rethink readbalancing and do it differently
483 * depending on near_copies / far_copies geometry.
484 */
485static int read_balance(conf_t *conf, r10bio_t *r10_bio)
486{
487 const unsigned long this_sector = r10_bio->sector;
488 int disk, slot, nslot;
489 const int sectors = r10_bio->sectors;
490 sector_t new_distance, current_distance;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800491 mdk_rdev_t *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492
493 raid10_find_phys(conf, r10_bio);
494 rcu_read_lock();
495 /*
496 * Check if we can balance. We can balance on the whole
NeilBrown6cce3b22006-01-06 00:20:16 -0800497 * device if no resync is going on (recovery is ok), or below
498 * the resync window. We take the first readable disk when
499 * above the resync window.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 */
501 if (conf->mddev->recovery_cp < MaxSector
502 && (this_sector + sectors >= conf->next_resync)) {
503 /* make sure that disk is operational */
504 slot = 0;
505 disk = r10_bio->devs[slot].devnum;
506
Suzanne Woodd6065f72005-11-08 21:39:27 -0800507 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800508 r10_bio->devs[slot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800509 !test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 slot++;
511 if (slot == conf->copies) {
512 slot = 0;
513 disk = -1;
514 break;
515 }
516 disk = r10_bio->devs[slot].devnum;
517 }
518 goto rb_out;
519 }
520
521
522 /* make sure the disk is operational */
523 slot = 0;
524 disk = r10_bio->devs[slot].devnum;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800525 while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800526 r10_bio->devs[slot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800527 !test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 slot ++;
529 if (slot == conf->copies) {
530 disk = -1;
531 goto rb_out;
532 }
533 disk = r10_bio->devs[slot].devnum;
534 }
535
536
NeilBrown3ec67ac2005-09-09 16:23:40 -0700537 current_distance = abs(r10_bio->devs[slot].addr -
538 conf->mirrors[disk].head_position);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539
540 /* Find the disk whose head is closest */
541
542 for (nslot = slot; nslot < conf->copies; nslot++) {
543 int ndisk = r10_bio->devs[nslot].devnum;
544
545
Suzanne Woodd6065f72005-11-08 21:39:27 -0800546 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800547 r10_bio->devs[nslot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800548 !test_bit(In_sync, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 continue;
550
NeilBrown22dfdf52005-11-28 13:44:09 -0800551 /* This optimisation is debatable, and completely destroys
552 * sequential read speed for 'far copies' arrays. So only
553 * keep it for 'near' arrays, and review those later.
554 */
555 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556 disk = ndisk;
557 slot = nslot;
558 break;
559 }
560 new_distance = abs(r10_bio->devs[nslot].addr -
561 conf->mirrors[ndisk].head_position);
562 if (new_distance < current_distance) {
563 current_distance = new_distance;
564 disk = ndisk;
565 slot = nslot;
566 }
567 }
568
569rb_out:
570 r10_bio->read_slot = slot;
571/* conf->next_seq_sect = this_sector + sectors;*/
572
Suzanne Woodd6065f72005-11-08 21:39:27 -0800573 if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
NeilBrown29fc7e32006-02-03 03:03:41 -0800575 else
576 disk = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 rcu_read_unlock();
578
579 return disk;
580}
581
582static void unplug_slaves(mddev_t *mddev)
583{
584 conf_t *conf = mddev_to_conf(mddev);
585 int i;
586
587 rcu_read_lock();
588 for (i=0; i<mddev->raid_disks; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800589 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -0800590 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200591 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592
593 atomic_inc(&rdev->nr_pending);
594 rcu_read_unlock();
595
Alan D. Brunelle2ad8b1e2007-11-07 14:26:56 -0500596 blk_unplug(r_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597
598 rdev_dec_pending(rdev, mddev);
599 rcu_read_lock();
600 }
601 }
602 rcu_read_unlock();
603}
604
Jens Axboe165125e2007-07-24 09:28:11 +0200605static void raid10_unplug(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606{
NeilBrown6cce3b22006-01-06 00:20:16 -0800607 mddev_t *mddev = q->queuedata;
608
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 unplug_slaves(q->queuedata);
NeilBrown6cce3b22006-01-06 00:20:16 -0800610 md_wakeup_thread(mddev->thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611}
612
NeilBrown0d129222006-10-03 01:15:54 -0700613static int raid10_congested(void *data, int bits)
614{
615 mddev_t *mddev = data;
616 conf_t *conf = mddev_to_conf(mddev);
617 int i, ret = 0;
618
619 rcu_read_lock();
620 for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
621 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
622 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200623 struct request_queue *q = bdev_get_queue(rdev->bdev);
NeilBrown0d129222006-10-03 01:15:54 -0700624
625 ret |= bdi_congested(&q->backing_dev_info, bits);
626 }
627 }
628 rcu_read_unlock();
629 return ret;
630}
631
632
NeilBrown0a27ec92006-01-06 00:20:13 -0800633/* Barriers....
634 * Sometimes we need to suspend IO while we do something else,
635 * either some resync/recovery, or reconfigure the array.
636 * To do this we raise a 'barrier'.
637 * The 'barrier' is a counter that can be raised multiple times
638 * to count how many activities are happening which preclude
639 * normal IO.
640 * We can only raise the barrier if there is no pending IO.
641 * i.e. if nr_pending == 0.
642 * We choose only to raise the barrier if no-one is waiting for the
643 * barrier to go down. This means that as soon as an IO request
644 * is ready, no other operations which require a barrier will start
645 * until the IO request has had a chance.
646 *
647 * So: regular IO calls 'wait_barrier'. When that returns there
648 * is no backgroup IO happening, It must arrange to call
649 * allow_barrier when it has finished its IO.
650 * backgroup IO calls must call raise_barrier. Once that returns
651 * there is no normal IO happeing. It must arrange to call
652 * lower_barrier when the particular background IO completes.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653 */
654#define RESYNC_DEPTH 32
655
NeilBrown6cce3b22006-01-06 00:20:16 -0800656static void raise_barrier(conf_t *conf, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657{
NeilBrown6cce3b22006-01-06 00:20:16 -0800658 BUG_ON(force && !conf->barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 spin_lock_irq(&conf->resync_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660
NeilBrown6cce3b22006-01-06 00:20:16 -0800661 /* Wait until no block IO is waiting (unless 'force') */
662 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
NeilBrown0a27ec92006-01-06 00:20:13 -0800663 conf->resync_lock,
664 raid10_unplug(conf->mddev->queue));
665
666 /* block any new IO from starting */
667 conf->barrier++;
668
669 /* No wait for all pending IO to complete */
670 wait_event_lock_irq(conf->wait_barrier,
671 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
672 conf->resync_lock,
673 raid10_unplug(conf->mddev->queue));
674
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675 spin_unlock_irq(&conf->resync_lock);
676}
677
NeilBrown0a27ec92006-01-06 00:20:13 -0800678static void lower_barrier(conf_t *conf)
679{
680 unsigned long flags;
681 spin_lock_irqsave(&conf->resync_lock, flags);
682 conf->barrier--;
683 spin_unlock_irqrestore(&conf->resync_lock, flags);
684 wake_up(&conf->wait_barrier);
685}
686
687static void wait_barrier(conf_t *conf)
688{
689 spin_lock_irq(&conf->resync_lock);
690 if (conf->barrier) {
691 conf->nr_waiting++;
692 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
693 conf->resync_lock,
694 raid10_unplug(conf->mddev->queue));
695 conf->nr_waiting--;
696 }
697 conf->nr_pending++;
698 spin_unlock_irq(&conf->resync_lock);
699}
700
701static void allow_barrier(conf_t *conf)
702{
703 unsigned long flags;
704 spin_lock_irqsave(&conf->resync_lock, flags);
705 conf->nr_pending--;
706 spin_unlock_irqrestore(&conf->resync_lock, flags);
707 wake_up(&conf->wait_barrier);
708}
709
NeilBrown4443ae12006-01-06 00:20:28 -0800710static void freeze_array(conf_t *conf)
711{
712 /* stop syncio and normal IO and wait for everything to
NeilBrownf1885932006-01-06 00:20:42 -0800713 * go quiet.
NeilBrown4443ae12006-01-06 00:20:28 -0800714 * We increment barrier and nr_waiting, and then
715 * wait until barrier+nr_pending match nr_queued+2
716 */
717 spin_lock_irq(&conf->resync_lock);
718 conf->barrier++;
719 conf->nr_waiting++;
720 wait_event_lock_irq(conf->wait_barrier,
721 conf->barrier+conf->nr_pending == conf->nr_queued+2,
722 conf->resync_lock,
723 raid10_unplug(conf->mddev->queue));
724 spin_unlock_irq(&conf->resync_lock);
725}
726
727static void unfreeze_array(conf_t *conf)
728{
729 /* reverse the effect of the freeze */
730 spin_lock_irq(&conf->resync_lock);
731 conf->barrier--;
732 conf->nr_waiting--;
733 wake_up(&conf->wait_barrier);
734 spin_unlock_irq(&conf->resync_lock);
735}
736
Jens Axboe165125e2007-07-24 09:28:11 +0200737static int make_request(struct request_queue *q, struct bio * bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738{
739 mddev_t *mddev = q->queuedata;
740 conf_t *conf = mddev_to_conf(mddev);
741 mirror_info_t *mirror;
742 r10bio_t *r10_bio;
743 struct bio *read_bio;
744 int i;
745 int chunk_sects = conf->chunk_mask + 1;
Jens Axboea3623572005-11-01 09:26:16 +0100746 const int rw = bio_data_dir(bio);
Lars Ellenberge3881a62007-01-10 23:15:37 -0800747 const int do_sync = bio_sync(bio);
NeilBrown6cce3b22006-01-06 00:20:16 -0800748 struct bio_list bl;
749 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750
NeilBrowne5dcdd82005-09-09 16:23:41 -0700751 if (unlikely(bio_barrier(bio))) {
NeilBrown6712ecf2007-09-27 12:47:43 +0200752 bio_endio(bio, -EOPNOTSUPP);
NeilBrowne5dcdd82005-09-09 16:23:41 -0700753 return 0;
754 }
755
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 /* If this request crosses a chunk boundary, we need to
757 * split it. This will only happen for 1 PAGE (or less) requests.
758 */
759 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
760 > chunk_sects &&
761 conf->near_copies < conf->raid_disks)) {
762 struct bio_pair *bp;
763 /* Sanity check -- queue functions should prevent this happening */
764 if (bio->bi_vcnt != 1 ||
765 bio->bi_idx != 0)
766 goto bad_map;
767 /* This is a one page bio that upper layers
768 * refuse to split for us, so we need to split it.
769 */
770 bp = bio_split(bio, bio_split_pool,
771 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
772 if (make_request(q, &bp->bio1))
773 generic_make_request(&bp->bio1);
774 if (make_request(q, &bp->bio2))
775 generic_make_request(&bp->bio2);
776
777 bio_pair_release(bp);
778 return 0;
779 bad_map:
780 printk("raid10_make_request bug: can't convert block across chunks"
781 " or bigger than %dk %llu %d\n", chunk_sects/2,
782 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
783
NeilBrown6712ecf2007-09-27 12:47:43 +0200784 bio_io_error(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 return 0;
786 }
787
NeilBrown3d310eb2005-06-21 17:17:26 -0700788 md_write_start(mddev, bio);
NeilBrown06d91a52005-06-21 17:17:12 -0700789
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790 /*
791 * Register the new request and wait if the reconstruction
792 * thread has put up a bar for new requests.
793 * Continue immediately if no resync is active currently.
794 */
NeilBrown0a27ec92006-01-06 00:20:13 -0800795 wait_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796
Jens Axboea3623572005-11-01 09:26:16 +0100797 disk_stat_inc(mddev->gendisk, ios[rw]);
798 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799
800 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
801
802 r10_bio->master_bio = bio;
803 r10_bio->sectors = bio->bi_size >> 9;
804
805 r10_bio->mddev = mddev;
806 r10_bio->sector = bio->bi_sector;
NeilBrown6cce3b22006-01-06 00:20:16 -0800807 r10_bio->state = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808
Jens Axboea3623572005-11-01 09:26:16 +0100809 if (rw == READ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810 /*
811 * read balancing logic:
812 */
813 int disk = read_balance(conf, r10_bio);
814 int slot = r10_bio->read_slot;
815 if (disk < 0) {
816 raid_end_bio_io(r10_bio);
817 return 0;
818 }
819 mirror = conf->mirrors + disk;
820
821 read_bio = bio_clone(bio, GFP_NOIO);
822
823 r10_bio->devs[slot].bio = read_bio;
824
825 read_bio->bi_sector = r10_bio->devs[slot].addr +
826 mirror->rdev->data_offset;
827 read_bio->bi_bdev = mirror->rdev->bdev;
828 read_bio->bi_end_io = raid10_end_read_request;
Lars Ellenberge3881a62007-01-10 23:15:37 -0800829 read_bio->bi_rw = READ | do_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 read_bio->bi_private = r10_bio;
831
832 generic_make_request(read_bio);
833 return 0;
834 }
835
836 /*
837 * WRITE:
838 */
839 /* first select target devices under spinlock and
840 * inc refcount on their rdev. Record them by setting
841 * bios[x] to bio
842 */
843 raid10_find_phys(conf, r10_bio);
844 rcu_read_lock();
845 for (i = 0; i < conf->copies; i++) {
846 int d = r10_bio->devs[i].devnum;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800847 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
848 if (rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -0800849 !test_bit(Faulty, &rdev->flags)) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800850 atomic_inc(&rdev->nr_pending);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 r10_bio->devs[i].bio = bio;
NeilBrown6cce3b22006-01-06 00:20:16 -0800852 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 r10_bio->devs[i].bio = NULL;
NeilBrown6cce3b22006-01-06 00:20:16 -0800854 set_bit(R10BIO_Degraded, &r10_bio->state);
855 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 }
857 rcu_read_unlock();
858
NeilBrown6cce3b22006-01-06 00:20:16 -0800859 atomic_set(&r10_bio->remaining, 0);
NeilBrown06d91a52005-06-21 17:17:12 -0700860
NeilBrown6cce3b22006-01-06 00:20:16 -0800861 bio_list_init(&bl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 for (i = 0; i < conf->copies; i++) {
863 struct bio *mbio;
864 int d = r10_bio->devs[i].devnum;
865 if (!r10_bio->devs[i].bio)
866 continue;
867
868 mbio = bio_clone(bio, GFP_NOIO);
869 r10_bio->devs[i].bio = mbio;
870
871 mbio->bi_sector = r10_bio->devs[i].addr+
872 conf->mirrors[d].rdev->data_offset;
873 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
874 mbio->bi_end_io = raid10_end_write_request;
Lars Ellenberge3881a62007-01-10 23:15:37 -0800875 mbio->bi_rw = WRITE | do_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 mbio->bi_private = r10_bio;
877
878 atomic_inc(&r10_bio->remaining);
NeilBrown6cce3b22006-01-06 00:20:16 -0800879 bio_list_add(&bl, mbio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880 }
881
Arne Redlichf6f953a2007-07-31 00:37:57 -0700882 if (unlikely(!atomic_read(&r10_bio->remaining))) {
883 /* the array is dead */
884 md_write_end(mddev);
885 raid_end_bio_io(r10_bio);
886 return 0;
887 }
888
NeilBrown6cce3b22006-01-06 00:20:16 -0800889 bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
890 spin_lock_irqsave(&conf->device_lock, flags);
891 bio_list_merge(&conf->pending_bio_list, &bl);
892 blk_plug_device(mddev->queue);
893 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894
Lars Ellenberge3881a62007-01-10 23:15:37 -0800895 if (do_sync)
896 md_wakeup_thread(mddev->thread);
897
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898 return 0;
899}
900
901static void status(struct seq_file *seq, mddev_t *mddev)
902{
903 conf_t *conf = mddev_to_conf(mddev);
904 int i;
905
906 if (conf->near_copies < conf->raid_disks)
907 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
908 if (conf->near_copies > 1)
909 seq_printf(seq, " %d near-copies", conf->near_copies);
NeilBrownc93983b2006-06-26 00:27:41 -0700910 if (conf->far_copies > 1) {
911 if (conf->far_offset)
912 seq_printf(seq, " %d offset-copies", conf->far_copies);
913 else
914 seq_printf(seq, " %d far-copies", conf->far_copies);
915 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
NeilBrown76186dd2006-10-03 01:15:48 -0700917 conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918 for (i = 0; i < conf->raid_disks; i++)
919 seq_printf(seq, "%s",
920 conf->mirrors[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -0800921 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 seq_printf(seq, "]");
923}
924
925static void error(mddev_t *mddev, mdk_rdev_t *rdev)
926{
927 char b[BDEVNAME_SIZE];
928 conf_t *conf = mddev_to_conf(mddev);
929
930 /*
931 * If it is not operational, then we have already marked it as dead
932 * else if it is the last working disks, ignore the error, let the
933 * next level up know.
934 * else mark the drive as failed
935 */
NeilBrownb2d444d2005-11-08 21:39:31 -0800936 if (test_bit(In_sync, &rdev->flags)
NeilBrown76186dd2006-10-03 01:15:48 -0700937 && conf->raid_disks-mddev->degraded == 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 /*
939 * Don't fail the drive, just return an IO error.
940 * The test should really be more sophisticated than
941 * "working_disks == 1", but it isn't critical, and
942 * can wait until we do more sophisticated "is the drive
943 * really dead" tests...
944 */
945 return;
NeilBrownc04be0a2006-10-03 01:15:53 -0700946 if (test_and_clear_bit(In_sync, &rdev->flags)) {
947 unsigned long flags;
948 spin_lock_irqsave(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 mddev->degraded++;
NeilBrownc04be0a2006-10-03 01:15:53 -0700950 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951 /*
952 * if recovery is running, make sure it aborts.
953 */
954 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
955 }
NeilBrownb2d444d2005-11-08 21:39:31 -0800956 set_bit(Faulty, &rdev->flags);
NeilBrown850b2b42006-10-03 01:15:46 -0700957 set_bit(MD_CHANGE_DEVS, &mddev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 printk(KERN_ALERT "raid10: Disk failure on %s, disabling device. \n"
959 " Operation continuing on %d devices\n",
NeilBrown76186dd2006-10-03 01:15:48 -0700960 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961}
962
963static void print_conf(conf_t *conf)
964{
965 int i;
966 mirror_info_t *tmp;
967
968 printk("RAID10 conf printout:\n");
969 if (!conf) {
970 printk("(!conf)\n");
971 return;
972 }
NeilBrown76186dd2006-10-03 01:15:48 -0700973 printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 conf->raid_disks);
975
976 for (i = 0; i < conf->raid_disks; i++) {
977 char b[BDEVNAME_SIZE];
978 tmp = conf->mirrors + i;
979 if (tmp->rdev)
980 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
NeilBrownb2d444d2005-11-08 21:39:31 -0800981 i, !test_bit(In_sync, &tmp->rdev->flags),
982 !test_bit(Faulty, &tmp->rdev->flags),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983 bdevname(tmp->rdev->bdev,b));
984 }
985}
986
987static void close_sync(conf_t *conf)
988{
NeilBrown0a27ec92006-01-06 00:20:13 -0800989 wait_barrier(conf);
990 allow_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991
992 mempool_destroy(conf->r10buf_pool);
993 conf->r10buf_pool = NULL;
994}
995
NeilBrown6d508242005-09-09 16:24:03 -0700996/* check if there are enough drives for
997 * every block to appear on atleast one
998 */
999static int enough(conf_t *conf)
1000{
1001 int first = 0;
1002
1003 do {
1004 int n = conf->copies;
1005 int cnt = 0;
1006 while (n--) {
1007 if (conf->mirrors[first].rdev)
1008 cnt++;
1009 first = (first+1) % conf->raid_disks;
1010 }
1011 if (cnt == 0)
1012 return 0;
1013 } while (first != 0);
1014 return 1;
1015}
1016
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017static int raid10_spare_active(mddev_t *mddev)
1018{
1019 int i;
1020 conf_t *conf = mddev->private;
1021 mirror_info_t *tmp;
1022
1023 /*
1024 * Find all non-in_sync disks within the RAID10 configuration
1025 * and mark them in_sync
1026 */
1027 for (i = 0; i < conf->raid_disks; i++) {
1028 tmp = conf->mirrors + i;
1029 if (tmp->rdev
NeilBrownb2d444d2005-11-08 21:39:31 -08001030 && !test_bit(Faulty, &tmp->rdev->flags)
NeilBrownc04be0a2006-10-03 01:15:53 -07001031 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1032 unsigned long flags;
1033 spin_lock_irqsave(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034 mddev->degraded--;
NeilBrownc04be0a2006-10-03 01:15:53 -07001035 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 }
1037 }
1038
1039 print_conf(conf);
1040 return 0;
1041}
1042
1043
1044static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1045{
1046 conf_t *conf = mddev->private;
1047 int found = 0;
1048 int mirror;
1049 mirror_info_t *p;
1050
1051 if (mddev->recovery_cp < MaxSector)
1052 /* only hot-add to in-sync arrays, as recovery is
1053 * very different from resync
1054 */
1055 return 0;
NeilBrown6d508242005-09-09 16:24:03 -07001056 if (!enough(conf))
1057 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001058
NeilBrown6cce3b22006-01-06 00:20:16 -08001059 if (rdev->saved_raid_disk >= 0 &&
1060 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1061 mirror = rdev->saved_raid_disk;
1062 else
1063 mirror = 0;
1064 for ( ; mirror < mddev->raid_disks; mirror++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065 if ( !(p=conf->mirrors+mirror)->rdev) {
1066
1067 blk_queue_stack_limits(mddev->queue,
1068 rdev->bdev->bd_disk->queue);
1069 /* as we don't honour merge_bvec_fn, we must never risk
1070 * violating it, so limit ->max_sector to one PAGE, as
1071 * a one page request is never in violation.
1072 */
1073 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1074 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1075 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1076
1077 p->head_position = 0;
1078 rdev->raid_disk = mirror;
1079 found = 1;
NeilBrown6cce3b22006-01-06 00:20:16 -08001080 if (rdev->saved_raid_disk != mirror)
1081 conf->fullsync = 1;
Suzanne Woodd6065f72005-11-08 21:39:27 -08001082 rcu_assign_pointer(p->rdev, rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083 break;
1084 }
1085
1086 print_conf(conf);
1087 return found;
1088}
1089
1090static int raid10_remove_disk(mddev_t *mddev, int number)
1091{
1092 conf_t *conf = mddev->private;
1093 int err = 0;
1094 mdk_rdev_t *rdev;
1095 mirror_info_t *p = conf->mirrors+ number;
1096
1097 print_conf(conf);
1098 rdev = p->rdev;
1099 if (rdev) {
NeilBrownb2d444d2005-11-08 21:39:31 -08001100 if (test_bit(In_sync, &rdev->flags) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 atomic_read(&rdev->nr_pending)) {
1102 err = -EBUSY;
1103 goto abort;
1104 }
1105 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07001106 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 if (atomic_read(&rdev->nr_pending)) {
1108 /* lost the race, try later */
1109 err = -EBUSY;
1110 p->rdev = rdev;
1111 }
1112 }
1113abort:
1114
1115 print_conf(conf);
1116 return err;
1117}
1118
1119
NeilBrown6712ecf2007-09-27 12:47:43 +02001120static void end_sync_read(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1123 conf_t *conf = mddev_to_conf(r10_bio->mddev);
1124 int i,d;
1125
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 for (i=0; i<conf->copies; i++)
1127 if (r10_bio->devs[i].bio == bio)
1128 break;
Eric Sesterhennb6385482006-04-02 13:34:29 +02001129 BUG_ON(i == conf->copies);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 update_head_pos(i, r10_bio);
1131 d = r10_bio->devs[i].devnum;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001132
1133 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1134 set_bit(R10BIO_Uptodate, &r10_bio->state);
NeilBrown4dbcdc72006-01-06 00:20:52 -08001135 else {
1136 atomic_add(r10_bio->sectors,
1137 &conf->mirrors[d].rdev->corrected_errors);
1138 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1139 md_error(r10_bio->mddev,
1140 conf->mirrors[d].rdev);
1141 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142
1143 /* for reconstruct, we always reschedule after a read.
1144 * for resync, only after all reads
1145 */
1146 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1147 atomic_dec_and_test(&r10_bio->remaining)) {
1148 /* we have read all the blocks,
1149 * do the comparison in process context in raid10d
1150 */
1151 reschedule_retry(r10_bio);
1152 }
1153 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154}
1155
NeilBrown6712ecf2007-09-27 12:47:43 +02001156static void end_sync_write(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157{
1158 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1159 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1160 mddev_t *mddev = r10_bio->mddev;
1161 conf_t *conf = mddev_to_conf(mddev);
1162 int i,d;
1163
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164 for (i = 0; i < conf->copies; i++)
1165 if (r10_bio->devs[i].bio == bio)
1166 break;
1167 d = r10_bio->devs[i].devnum;
1168
1169 if (!uptodate)
1170 md_error(mddev, conf->mirrors[d].rdev);
1171 update_head_pos(i, r10_bio);
1172
1173 while (atomic_dec_and_test(&r10_bio->remaining)) {
1174 if (r10_bio->master_bio == NULL) {
1175 /* the primary of several recovery bios */
1176 md_done_sync(mddev, r10_bio->sectors, 1);
1177 put_buf(r10_bio);
1178 break;
1179 } else {
1180 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1181 put_buf(r10_bio);
1182 r10_bio = r10_bio2;
1183 }
1184 }
1185 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186}
1187
1188/*
1189 * Note: sync and recover and handled very differently for raid10
1190 * This code is for resync.
1191 * For resync, we read through virtual addresses and read all blocks.
1192 * If there is any error, we schedule a write. The lowest numbered
1193 * drive is authoritative.
1194 * However requests come for physical address, so we need to map.
1195 * For every physical address there are raid_disks/copies virtual addresses,
1196 * which is always are least one, but is not necessarly an integer.
1197 * This means that a physical address can span multiple chunks, so we may
1198 * have to submit multiple io requests for a single sync request.
1199 */
1200/*
1201 * We check if all blocks are in-sync and only write to blocks that
1202 * aren't in sync
1203 */
1204static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1205{
1206 conf_t *conf = mddev_to_conf(mddev);
1207 int i, first;
1208 struct bio *tbio, *fbio;
1209
1210 atomic_set(&r10_bio->remaining, 1);
1211
1212 /* find the first device with a block */
1213 for (i=0; i<conf->copies; i++)
1214 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1215 break;
1216
1217 if (i == conf->copies)
1218 goto done;
1219
1220 first = i;
1221 fbio = r10_bio->devs[i].bio;
1222
1223 /* now find blocks with errors */
NeilBrown0eb3ff12006-01-06 00:20:29 -08001224 for (i=0 ; i < conf->copies ; i++) {
1225 int j, d;
1226 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228 tbio = r10_bio->devs[i].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001229
1230 if (tbio->bi_end_io != end_sync_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231 continue;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001232 if (i == first)
1233 continue;
1234 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1235 /* We know that the bi_io_vec layout is the same for
1236 * both 'first' and 'i', so we just compare them.
1237 * All vec entries are PAGE_SIZE;
1238 */
1239 for (j = 0; j < vcnt; j++)
1240 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1241 page_address(tbio->bi_io_vec[j].bv_page),
1242 PAGE_SIZE))
1243 break;
1244 if (j == vcnt)
1245 continue;
1246 mddev->resync_mismatches += r10_bio->sectors;
1247 }
NeilBrown18f08812006-01-06 00:20:25 -08001248 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1249 /* Don't fix anything. */
1250 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 /* Ok, we need to write this bio
1252 * First we need to fixup bv_offset, bv_len and
1253 * bi_vecs, as the read request might have corrupted these
1254 */
1255 tbio->bi_vcnt = vcnt;
1256 tbio->bi_size = r10_bio->sectors << 9;
1257 tbio->bi_idx = 0;
1258 tbio->bi_phys_segments = 0;
1259 tbio->bi_hw_segments = 0;
1260 tbio->bi_hw_front_size = 0;
1261 tbio->bi_hw_back_size = 0;
1262 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1263 tbio->bi_flags |= 1 << BIO_UPTODATE;
1264 tbio->bi_next = NULL;
1265 tbio->bi_rw = WRITE;
1266 tbio->bi_private = r10_bio;
1267 tbio->bi_sector = r10_bio->devs[i].addr;
1268
1269 for (j=0; j < vcnt ; j++) {
1270 tbio->bi_io_vec[j].bv_offset = 0;
1271 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1272
1273 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1274 page_address(fbio->bi_io_vec[j].bv_page),
1275 PAGE_SIZE);
1276 }
1277 tbio->bi_end_io = end_sync_write;
1278
1279 d = r10_bio->devs[i].devnum;
1280 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1281 atomic_inc(&r10_bio->remaining);
1282 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1283
1284 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1285 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1286 generic_make_request(tbio);
1287 }
1288
1289done:
1290 if (atomic_dec_and_test(&r10_bio->remaining)) {
1291 md_done_sync(mddev, r10_bio->sectors, 1);
1292 put_buf(r10_bio);
1293 }
1294}
1295
1296/*
1297 * Now for the recovery code.
1298 * Recovery happens across physical sectors.
1299 * We recover all non-is_sync drives by finding the virtual address of
1300 * each, and then choose a working drive that also has that virt address.
1301 * There is a separate r10_bio for each non-in_sync drive.
1302 * Only the first two slots are in use. The first for reading,
1303 * The second for writing.
1304 *
1305 */
1306
1307static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1308{
1309 conf_t *conf = mddev_to_conf(mddev);
1310 int i, d;
1311 struct bio *bio, *wbio;
1312
1313
1314 /* move the pages across to the second bio
1315 * and submit the write request
1316 */
1317 bio = r10_bio->devs[0].bio;
1318 wbio = r10_bio->devs[1].bio;
1319 for (i=0; i < wbio->bi_vcnt; i++) {
1320 struct page *p = bio->bi_io_vec[i].bv_page;
1321 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1322 wbio->bi_io_vec[i].bv_page = p;
1323 }
1324 d = r10_bio->devs[1].devnum;
1325
1326 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1327 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
NeilBrown0eb3ff12006-01-06 00:20:29 -08001328 if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1329 generic_make_request(wbio);
1330 else
NeilBrown6712ecf2007-09-27 12:47:43 +02001331 bio_endio(wbio, -EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332}
1333
1334
1335/*
1336 * This is a kernel thread which:
1337 *
1338 * 1. Retries failed read operations on working mirrors.
1339 * 2. Updates the raid superblock when problems encounter.
NeilBrown6814d532006-10-03 01:15:45 -07001340 * 3. Performs writes following reads for array synchronising.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341 */
1342
NeilBrown6814d532006-10-03 01:15:45 -07001343static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1344{
1345 int sect = 0; /* Offset from r10_bio->sector */
1346 int sectors = r10_bio->sectors;
1347 mdk_rdev_t*rdev;
1348 while(sectors) {
1349 int s = sectors;
1350 int sl = r10_bio->read_slot;
1351 int success = 0;
1352 int start;
1353
1354 if (s > (PAGE_SIZE>>9))
1355 s = PAGE_SIZE >> 9;
1356
1357 rcu_read_lock();
1358 do {
1359 int d = r10_bio->devs[sl].devnum;
1360 rdev = rcu_dereference(conf->mirrors[d].rdev);
1361 if (rdev &&
1362 test_bit(In_sync, &rdev->flags)) {
1363 atomic_inc(&rdev->nr_pending);
1364 rcu_read_unlock();
1365 success = sync_page_io(rdev->bdev,
1366 r10_bio->devs[sl].addr +
1367 sect + rdev->data_offset,
1368 s<<9,
1369 conf->tmppage, READ);
1370 rdev_dec_pending(rdev, mddev);
1371 rcu_read_lock();
1372 if (success)
1373 break;
1374 }
1375 sl++;
1376 if (sl == conf->copies)
1377 sl = 0;
1378 } while (!success && sl != r10_bio->read_slot);
1379 rcu_read_unlock();
1380
1381 if (!success) {
1382 /* Cannot read from anywhere -- bye bye array */
1383 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1384 md_error(mddev, conf->mirrors[dn].rdev);
1385 break;
1386 }
1387
1388 start = sl;
1389 /* write it back and re-read */
1390 rcu_read_lock();
1391 while (sl != r10_bio->read_slot) {
1392 int d;
1393 if (sl==0)
1394 sl = conf->copies;
1395 sl--;
1396 d = r10_bio->devs[sl].devnum;
1397 rdev = rcu_dereference(conf->mirrors[d].rdev);
1398 if (rdev &&
1399 test_bit(In_sync, &rdev->flags)) {
1400 atomic_inc(&rdev->nr_pending);
1401 rcu_read_unlock();
1402 atomic_add(s, &rdev->corrected_errors);
1403 if (sync_page_io(rdev->bdev,
1404 r10_bio->devs[sl].addr +
1405 sect + rdev->data_offset,
1406 s<<9, conf->tmppage, WRITE)
1407 == 0)
1408 /* Well, this device is dead */
1409 md_error(mddev, rdev);
1410 rdev_dec_pending(rdev, mddev);
1411 rcu_read_lock();
1412 }
1413 }
1414 sl = start;
1415 while (sl != r10_bio->read_slot) {
1416 int d;
1417 if (sl==0)
1418 sl = conf->copies;
1419 sl--;
1420 d = r10_bio->devs[sl].devnum;
1421 rdev = rcu_dereference(conf->mirrors[d].rdev);
1422 if (rdev &&
1423 test_bit(In_sync, &rdev->flags)) {
1424 char b[BDEVNAME_SIZE];
1425 atomic_inc(&rdev->nr_pending);
1426 rcu_read_unlock();
1427 if (sync_page_io(rdev->bdev,
1428 r10_bio->devs[sl].addr +
1429 sect + rdev->data_offset,
1430 s<<9, conf->tmppage, READ) == 0)
1431 /* Well, this device is dead */
1432 md_error(mddev, rdev);
1433 else
1434 printk(KERN_INFO
1435 "raid10:%s: read error corrected"
1436 " (%d sectors at %llu on %s)\n",
1437 mdname(mddev), s,
Randy Dunlap969b7552006-10-28 10:38:32 -07001438 (unsigned long long)(sect+
1439 rdev->data_offset),
NeilBrown6814d532006-10-03 01:15:45 -07001440 bdevname(rdev->bdev, b));
1441
1442 rdev_dec_pending(rdev, mddev);
1443 rcu_read_lock();
1444 }
1445 }
1446 rcu_read_unlock();
1447
1448 sectors -= s;
1449 sect += s;
1450 }
1451}
1452
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453static void raid10d(mddev_t *mddev)
1454{
1455 r10bio_t *r10_bio;
1456 struct bio *bio;
1457 unsigned long flags;
1458 conf_t *conf = mddev_to_conf(mddev);
1459 struct list_head *head = &conf->retry_list;
1460 int unplug=0;
1461 mdk_rdev_t *rdev;
1462
1463 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464
1465 for (;;) {
1466 char b[BDEVNAME_SIZE];
1467 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrown6cce3b22006-01-06 00:20:16 -08001468
1469 if (conf->pending_bio_list.head) {
1470 bio = bio_list_get(&conf->pending_bio_list);
1471 blk_remove_plug(mddev->queue);
1472 spin_unlock_irqrestore(&conf->device_lock, flags);
1473 /* flush any pending bitmap writes to disk before proceeding w/ I/O */
NeilBrown4ad13662007-07-17 04:06:13 -07001474 bitmap_unplug(mddev->bitmap);
NeilBrown6cce3b22006-01-06 00:20:16 -08001475
1476 while (bio) { /* submit pending writes */
1477 struct bio *next = bio->bi_next;
1478 bio->bi_next = NULL;
1479 generic_make_request(bio);
1480 bio = next;
1481 }
1482 unplug = 1;
1483
1484 continue;
1485 }
1486
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 if (list_empty(head))
1488 break;
1489 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1490 list_del(head->prev);
NeilBrown4443ae12006-01-06 00:20:28 -08001491 conf->nr_queued--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 spin_unlock_irqrestore(&conf->device_lock, flags);
1493
1494 mddev = r10_bio->mddev;
1495 conf = mddev_to_conf(mddev);
1496 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1497 sync_request_write(mddev, r10_bio);
1498 unplug = 1;
1499 } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1500 recovery_request_write(mddev, r10_bio);
1501 unplug = 1;
1502 } else {
1503 int mirror;
NeilBrown4443ae12006-01-06 00:20:28 -08001504 /* we got a read error. Maybe the drive is bad. Maybe just
1505 * the block and we can fix it.
1506 * We freeze all other IO, and try reading the block from
1507 * other devices. When we find one, we re-write
1508 * and check it that fixes the read error.
1509 * This is all done synchronously while the array is
1510 * frozen.
1511 */
NeilBrown6814d532006-10-03 01:15:45 -07001512 if (mddev->ro == 0) {
1513 freeze_array(conf);
1514 fix_read_error(conf, mddev, r10_bio);
1515 unfreeze_array(conf);
NeilBrown4443ae12006-01-06 00:20:28 -08001516 }
1517
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 bio = r10_bio->devs[r10_bio->read_slot].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001519 r10_bio->devs[r10_bio->read_slot].bio =
1520 mddev->ro ? IO_BLOCKED : NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 mirror = read_balance(conf, r10_bio);
1522 if (mirror == -1) {
1523 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1524 " read error for block %llu\n",
1525 bdevname(bio->bi_bdev,b),
1526 (unsigned long long)r10_bio->sector);
1527 raid_end_bio_io(r10_bio);
Maik Hampel14e71342007-07-31 00:37:57 -07001528 bio_put(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001529 } else {
Lars Ellenberge3881a62007-01-10 23:15:37 -08001530 const int do_sync = bio_sync(r10_bio->master_bio);
Maik Hampel14e71342007-07-31 00:37:57 -07001531 bio_put(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532 rdev = conf->mirrors[mirror].rdev;
1533 if (printk_ratelimit())
1534 printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1535 " another mirror\n",
1536 bdevname(rdev->bdev,b),
1537 (unsigned long long)r10_bio->sector);
1538 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1539 r10_bio->devs[r10_bio->read_slot].bio = bio;
1540 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1541 + rdev->data_offset;
1542 bio->bi_bdev = rdev->bdev;
Lars Ellenberge3881a62007-01-10 23:15:37 -08001543 bio->bi_rw = READ | do_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544 bio->bi_private = r10_bio;
1545 bio->bi_end_io = raid10_end_read_request;
1546 unplug = 1;
1547 generic_make_request(bio);
1548 }
1549 }
1550 }
1551 spin_unlock_irqrestore(&conf->device_lock, flags);
1552 if (unplug)
1553 unplug_slaves(mddev);
1554}
1555
1556
1557static int init_resync(conf_t *conf)
1558{
1559 int buffs;
1560
1561 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
Eric Sesterhennb6385482006-04-02 13:34:29 +02001562 BUG_ON(conf->r10buf_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1564 if (!conf->r10buf_pool)
1565 return -ENOMEM;
1566 conf->next_resync = 0;
1567 return 0;
1568}
1569
1570/*
1571 * perform a "sync" on one "block"
1572 *
1573 * We need to make sure that no normal I/O request - particularly write
1574 * requests - conflict with active sync requests.
1575 *
1576 * This is achieved by tracking pending requests and a 'barrier' concept
1577 * that can be installed to exclude normal IO requests.
1578 *
1579 * Resync and recovery are handled very differently.
1580 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1581 *
1582 * For resync, we iterate over virtual addresses, read all copies,
1583 * and update if there are differences. If only one copy is live,
1584 * skip it.
1585 * For recovery, we iterate over physical addresses, read a good
1586 * value for each non-in_sync drive, and over-write.
1587 *
1588 * So, for recovery we may have several outstanding complex requests for a
1589 * given address, one for each out-of-sync device. We model this by allocating
1590 * a number of r10_bio structures, one for each out-of-sync device.
1591 * As we setup these structures, we collect all bio's together into a list
1592 * which we then process collectively to add pages, and then process again
1593 * to pass to generic_make_request.
1594 *
1595 * The r10_bio structures are linked using a borrowed master_bio pointer.
1596 * This link is counted in ->remaining. When the r10_bio that points to NULL
1597 * has its remaining count decremented to 0, the whole complex operation
1598 * is complete.
1599 *
1600 */
1601
NeilBrown57afd892005-06-21 17:17:13 -07001602static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603{
1604 conf_t *conf = mddev_to_conf(mddev);
1605 r10bio_t *r10_bio;
1606 struct bio *biolist = NULL, *bio;
1607 sector_t max_sector, nr_sectors;
1608 int disk;
1609 int i;
NeilBrown6cce3b22006-01-06 00:20:16 -08001610 int max_sync;
1611 int sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612
1613 sector_t sectors_skipped = 0;
1614 int chunks_skipped = 0;
1615
1616 if (!conf->r10buf_pool)
1617 if (init_resync(conf))
NeilBrown57afd892005-06-21 17:17:13 -07001618 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619
1620 skipped:
1621 max_sector = mddev->size << 1;
1622 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1623 max_sector = mddev->resync_max_sectors;
1624 if (sector_nr >= max_sector) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001625 /* If we aborted, we need to abort the
1626 * sync on the 'current' bitmap chucks (there can
1627 * be several when recovering multiple devices).
1628 * as we may have started syncing it but not finished.
1629 * We can find the current address in
1630 * mddev->curr_resync, but for recovery,
1631 * we need to convert that to several
1632 * virtual addresses.
1633 */
1634 if (mddev->curr_resync < max_sector) { /* aborted */
1635 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1636 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1637 &sync_blocks, 1);
1638 else for (i=0; i<conf->raid_disks; i++) {
1639 sector_t sect =
1640 raid10_find_virt(conf, mddev->curr_resync, i);
1641 bitmap_end_sync(mddev->bitmap, sect,
1642 &sync_blocks, 1);
1643 }
1644 } else /* completed sync */
1645 conf->fullsync = 0;
1646
1647 bitmap_close_sync(mddev->bitmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 close_sync(conf);
NeilBrown57afd892005-06-21 17:17:13 -07001649 *skipped = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650 return sectors_skipped;
1651 }
1652 if (chunks_skipped >= conf->raid_disks) {
1653 /* if there has been nothing to do on any drive,
1654 * then there is nothing to do at all..
1655 */
NeilBrown57afd892005-06-21 17:17:13 -07001656 *skipped = 1;
1657 return (max_sector - sector_nr) + sectors_skipped;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658 }
1659
NeilBrownc6207272008-02-06 01:39:52 -08001660 if (max_sector > mddev->resync_max)
1661 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1662
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 /* make sure whole request will fit in a chunk - if chunks
1664 * are meaningful
1665 */
1666 if (conf->near_copies < conf->raid_disks &&
1667 max_sector > (sector_nr | conf->chunk_mask))
1668 max_sector = (sector_nr | conf->chunk_mask) + 1;
1669 /*
1670 * If there is non-resync activity waiting for us then
1671 * put in a delay to throttle resync.
1672 */
NeilBrown0a27ec92006-01-06 00:20:13 -08001673 if (!go_faster && conf->nr_waiting)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674 msleep_interruptible(1000);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675
NeilBrownb47490c2008-02-06 01:39:50 -08001676 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1677
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678 /* Again, very different code for resync and recovery.
1679 * Both must result in an r10bio with a list of bios that
1680 * have bi_end_io, bi_sector, bi_bdev set,
1681 * and bi_private set to the r10bio.
1682 * For recovery, we may actually create several r10bios
1683 * with 2 bios in each, that correspond to the bios in the main one.
1684 * In this case, the subordinate r10bios link back through a
1685 * borrowed master_bio pointer, and the counter in the master
1686 * includes a ref from each subordinate.
1687 */
1688 /* First, we decide what to do and set ->bi_end_io
1689 * To end_sync_read if we want to read, and
1690 * end_sync_write if we will want to write.
1691 */
1692
NeilBrown6cce3b22006-01-06 00:20:16 -08001693 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1695 /* recovery... the complicated one */
1696 int i, j, k;
1697 r10_bio = NULL;
1698
1699 for (i=0 ; i<conf->raid_disks; i++)
1700 if (conf->mirrors[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001701 !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001702 int still_degraded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 /* want to reconstruct this device */
1704 r10bio_t *rb2 = r10_bio;
NeilBrown6cce3b22006-01-06 00:20:16 -08001705 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1706 int must_sync;
1707 /* Unless we are doing a full sync, we only need
1708 * to recover the block if it is set in the bitmap
1709 */
1710 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1711 &sync_blocks, 1);
1712 if (sync_blocks < max_sync)
1713 max_sync = sync_blocks;
1714 if (!must_sync &&
1715 !conf->fullsync) {
1716 /* yep, skip the sync_blocks here, but don't assume
1717 * that there will never be anything to do here
1718 */
1719 chunks_skipped = -1;
1720 continue;
1721 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722
1723 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
NeilBrown6cce3b22006-01-06 00:20:16 -08001724 raise_barrier(conf, rb2 != NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 atomic_set(&r10_bio->remaining, 0);
1726
1727 r10_bio->master_bio = (struct bio*)rb2;
1728 if (rb2)
1729 atomic_inc(&rb2->remaining);
1730 r10_bio->mddev = mddev;
1731 set_bit(R10BIO_IsRecover, &r10_bio->state);
NeilBrown6cce3b22006-01-06 00:20:16 -08001732 r10_bio->sector = sect;
1733
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 raid10_find_phys(conf, r10_bio);
NeilBrown6cce3b22006-01-06 00:20:16 -08001735 /* Need to check if this section will still be
1736 * degraded
1737 */
1738 for (j=0; j<conf->copies;j++) {
1739 int d = r10_bio->devs[j].devnum;
1740 if (conf->mirrors[d].rdev == NULL ||
NeilBrowna24a8dd2006-01-06 00:20:35 -08001741 test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001742 still_degraded = 1;
NeilBrowna24a8dd2006-01-06 00:20:35 -08001743 break;
1744 }
NeilBrown6cce3b22006-01-06 00:20:16 -08001745 }
1746 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1747 &sync_blocks, still_degraded);
1748
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749 for (j=0; j<conf->copies;j++) {
1750 int d = r10_bio->devs[j].devnum;
1751 if (conf->mirrors[d].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001752 test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753 /* This is where we read from */
1754 bio = r10_bio->devs[0].bio;
1755 bio->bi_next = biolist;
1756 biolist = bio;
1757 bio->bi_private = r10_bio;
1758 bio->bi_end_io = end_sync_read;
NeilBrown802ba062006-12-13 00:34:13 -08001759 bio->bi_rw = READ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760 bio->bi_sector = r10_bio->devs[j].addr +
1761 conf->mirrors[d].rdev->data_offset;
1762 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1763 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1764 atomic_inc(&r10_bio->remaining);
1765 /* and we write to 'i' */
1766
1767 for (k=0; k<conf->copies; k++)
1768 if (r10_bio->devs[k].devnum == i)
1769 break;
NeilBrown64a742b2007-02-28 20:11:18 -08001770 BUG_ON(k == conf->copies);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771 bio = r10_bio->devs[1].bio;
1772 bio->bi_next = biolist;
1773 biolist = bio;
1774 bio->bi_private = r10_bio;
1775 bio->bi_end_io = end_sync_write;
NeilBrown802ba062006-12-13 00:34:13 -08001776 bio->bi_rw = WRITE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 bio->bi_sector = r10_bio->devs[k].addr +
1778 conf->mirrors[i].rdev->data_offset;
1779 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1780
1781 r10_bio->devs[0].devnum = d;
1782 r10_bio->devs[1].devnum = i;
1783
1784 break;
1785 }
1786 }
1787 if (j == conf->copies) {
NeilBrown87fc7672005-09-09 16:24:04 -07001788 /* Cannot recover, so abort the recovery */
1789 put_buf(r10_bio);
1790 r10_bio = rb2;
1791 if (!test_and_set_bit(MD_RECOVERY_ERR, &mddev->recovery))
1792 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1793 mdname(mddev));
1794 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795 }
1796 }
1797 if (biolist == NULL) {
1798 while (r10_bio) {
1799 r10bio_t *rb2 = r10_bio;
1800 r10_bio = (r10bio_t*) rb2->master_bio;
1801 rb2->master_bio = NULL;
1802 put_buf(rb2);
1803 }
1804 goto giveup;
1805 }
1806 } else {
1807 /* resync. Schedule a read for every block at this virt offset */
1808 int count = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08001809
1810 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1811 &sync_blocks, mddev->degraded) &&
1812 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1813 /* We can skip this block */
1814 *skipped = 1;
1815 return sync_blocks + sectors_skipped;
1816 }
1817 if (sync_blocks < max_sync)
1818 max_sync = sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001819 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1820
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821 r10_bio->mddev = mddev;
1822 atomic_set(&r10_bio->remaining, 0);
NeilBrown6cce3b22006-01-06 00:20:16 -08001823 raise_barrier(conf, 0);
1824 conf->next_resync = sector_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825
1826 r10_bio->master_bio = NULL;
1827 r10_bio->sector = sector_nr;
1828 set_bit(R10BIO_IsSync, &r10_bio->state);
1829 raid10_find_phys(conf, r10_bio);
1830 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1831
1832 for (i=0; i<conf->copies; i++) {
1833 int d = r10_bio->devs[i].devnum;
1834 bio = r10_bio->devs[i].bio;
1835 bio->bi_end_io = NULL;
NeilBrownaf03b8e2007-06-16 10:16:06 -07001836 clear_bit(BIO_UPTODATE, &bio->bi_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001837 if (conf->mirrors[d].rdev == NULL ||
NeilBrownb2d444d2005-11-08 21:39:31 -08001838 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839 continue;
1840 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1841 atomic_inc(&r10_bio->remaining);
1842 bio->bi_next = biolist;
1843 biolist = bio;
1844 bio->bi_private = r10_bio;
1845 bio->bi_end_io = end_sync_read;
NeilBrown802ba062006-12-13 00:34:13 -08001846 bio->bi_rw = READ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 bio->bi_sector = r10_bio->devs[i].addr +
1848 conf->mirrors[d].rdev->data_offset;
1849 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1850 count++;
1851 }
1852
1853 if (count < 2) {
1854 for (i=0; i<conf->copies; i++) {
1855 int d = r10_bio->devs[i].devnum;
1856 if (r10_bio->devs[i].bio->bi_end_io)
1857 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1858 }
1859 put_buf(r10_bio);
1860 biolist = NULL;
1861 goto giveup;
1862 }
1863 }
1864
1865 for (bio = biolist; bio ; bio=bio->bi_next) {
1866
1867 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1868 if (bio->bi_end_io)
1869 bio->bi_flags |= 1 << BIO_UPTODATE;
1870 bio->bi_vcnt = 0;
1871 bio->bi_idx = 0;
1872 bio->bi_phys_segments = 0;
1873 bio->bi_hw_segments = 0;
1874 bio->bi_size = 0;
1875 }
1876
1877 nr_sectors = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08001878 if (sector_nr + max_sync < max_sector)
1879 max_sector = sector_nr + max_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880 do {
1881 struct page *page;
1882 int len = PAGE_SIZE;
1883 disk = 0;
1884 if (sector_nr + (len>>9) > max_sector)
1885 len = (max_sector - sector_nr) << 9;
1886 if (len == 0)
1887 break;
1888 for (bio= biolist ; bio ; bio=bio->bi_next) {
1889 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1890 if (bio_add_page(bio, page, len, 0) == 0) {
1891 /* stop here */
1892 struct bio *bio2;
1893 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1894 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1895 /* remove last page from this bio */
1896 bio2->bi_vcnt--;
1897 bio2->bi_size -= len;
1898 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1899 }
1900 goto bio_full;
1901 }
1902 disk = i;
1903 }
1904 nr_sectors += len>>9;
1905 sector_nr += len>>9;
1906 } while (biolist->bi_vcnt < RESYNC_PAGES);
1907 bio_full:
1908 r10_bio->sectors = nr_sectors;
1909
1910 while (biolist) {
1911 bio = biolist;
1912 biolist = biolist->bi_next;
1913
1914 bio->bi_next = NULL;
1915 r10_bio = bio->bi_private;
1916 r10_bio->sectors = nr_sectors;
1917
1918 if (bio->bi_end_io == end_sync_read) {
1919 md_sync_acct(bio->bi_bdev, nr_sectors);
1920 generic_make_request(bio);
1921 }
1922 }
1923
NeilBrown57afd892005-06-21 17:17:13 -07001924 if (sectors_skipped)
1925 /* pretend they weren't skipped, it makes
1926 * no important difference in this case
1927 */
1928 md_done_sync(mddev, sectors_skipped, 1);
1929
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 return sectors_skipped + nr_sectors;
1931 giveup:
1932 /* There is nowhere to write, so all non-sync
1933 * drives must be failed, so try the next chunk...
1934 */
1935 {
NeilBrown57afd892005-06-21 17:17:13 -07001936 sector_t sec = max_sector - sector_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937 sectors_skipped += sec;
1938 chunks_skipped ++;
1939 sector_nr = max_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 goto skipped;
1941 }
1942}
1943
1944static int run(mddev_t *mddev)
1945{
1946 conf_t *conf;
1947 int i, disk_idx;
1948 mirror_info_t *disk;
1949 mdk_rdev_t *rdev;
1950 struct list_head *tmp;
NeilBrownc93983b2006-06-26 00:27:41 -07001951 int nc, fc, fo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952 sector_t stride, size;
1953
NeilBrown2604b702006-01-06 00:20:36 -08001954 if (mddev->chunk_size == 0) {
1955 printk(KERN_ERR "md/raid10: non-zero chunk size required.\n");
1956 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957 }
NeilBrown2604b702006-01-06 00:20:36 -08001958
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 nc = mddev->layout & 255;
1960 fc = (mddev->layout >> 8) & 255;
NeilBrownc93983b2006-06-26 00:27:41 -07001961 fo = mddev->layout & (1<<16);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
NeilBrownc93983b2006-06-26 00:27:41 -07001963 (mddev->layout >> 17)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
1965 mdname(mddev), mddev->layout);
1966 goto out;
1967 }
1968 /*
1969 * copy the already verified devices into our private RAID10
1970 * bookkeeping area. [whatever we allocate in run(),
1971 * should be freed in stop()]
1972 */
NeilBrown4443ae12006-01-06 00:20:28 -08001973 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974 mddev->private = conf;
1975 if (!conf) {
1976 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1977 mdname(mddev));
1978 goto out;
1979 }
NeilBrown4443ae12006-01-06 00:20:28 -08001980 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 GFP_KERNEL);
1982 if (!conf->mirrors) {
1983 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
1984 mdname(mddev));
1985 goto out_free_conf;
1986 }
NeilBrown4443ae12006-01-06 00:20:28 -08001987
1988 conf->tmppage = alloc_page(GFP_KERNEL);
1989 if (!conf->tmppage)
1990 goto out_free_conf;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991
NeilBrown64a742b2007-02-28 20:11:18 -08001992 conf->mddev = mddev;
1993 conf->raid_disks = mddev->raid_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 conf->near_copies = nc;
1995 conf->far_copies = fc;
1996 conf->copies = nc*fc;
NeilBrownc93983b2006-06-26 00:27:41 -07001997 conf->far_offset = fo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998 conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
1999 conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
NeilBrown64a742b2007-02-28 20:11:18 -08002000 size = mddev->size >> (conf->chunk_shift-1);
2001 sector_div(size, fc);
2002 size = size * conf->raid_disks;
2003 sector_div(size, nc);
2004 /* 'size' is now the number of chunks in the array */
2005 /* calculate "used chunks per device" in 'stride' */
2006 stride = size * conf->copies;
NeilBrownaf03b8e2007-06-16 10:16:06 -07002007
2008 /* We need to round up when dividing by raid_disks to
2009 * get the stride size.
2010 */
2011 stride += conf->raid_disks - 1;
NeilBrown64a742b2007-02-28 20:11:18 -08002012 sector_div(stride, conf->raid_disks);
2013 mddev->size = stride << (conf->chunk_shift-1);
2014
NeilBrownc93983b2006-06-26 00:27:41 -07002015 if (fo)
NeilBrown64a742b2007-02-28 20:11:18 -08002016 stride = 1;
2017 else
NeilBrownc93983b2006-06-26 00:27:41 -07002018 sector_div(stride, fc);
NeilBrown64a742b2007-02-28 20:11:18 -08002019 conf->stride = stride << conf->chunk_shift;
2020
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2022 r10bio_pool_free, conf);
2023 if (!conf->r10bio_pool) {
2024 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2025 mdname(mddev));
2026 goto out_free_conf;
2027 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028
NeilBrownd089c6a2008-02-06 01:39:59 -08002029 rdev_for_each(rdev, tmp, mddev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 disk_idx = rdev->raid_disk;
2031 if (disk_idx >= mddev->raid_disks
2032 || disk_idx < 0)
2033 continue;
2034 disk = conf->mirrors + disk_idx;
2035
2036 disk->rdev = rdev;
2037
2038 blk_queue_stack_limits(mddev->queue,
2039 rdev->bdev->bd_disk->queue);
2040 /* as we don't honour merge_bvec_fn, we must never risk
2041 * violating it, so limit ->max_sector to one PAGE, as
2042 * a one page request is never in violation.
2043 */
2044 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2045 mddev->queue->max_sectors > (PAGE_SIZE>>9))
2046 mddev->queue->max_sectors = (PAGE_SIZE>>9);
2047
2048 disk->head_position = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050 spin_lock_init(&conf->device_lock);
2051 INIT_LIST_HEAD(&conf->retry_list);
2052
2053 spin_lock_init(&conf->resync_lock);
NeilBrown0a27ec92006-01-06 00:20:13 -08002054 init_waitqueue_head(&conf->wait_barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055
NeilBrown6d508242005-09-09 16:24:03 -07002056 /* need to check that every block has at least one working mirror */
2057 if (!enough(conf)) {
2058 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2059 mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060 goto out_free_conf;
2061 }
2062
2063 mddev->degraded = 0;
2064 for (i = 0; i < conf->raid_disks; i++) {
2065
2066 disk = conf->mirrors + i;
2067
NeilBrown5fd6c1d2006-06-26 00:27:40 -07002068 if (!disk->rdev ||
NeilBrown2e333e82006-10-21 10:24:07 -07002069 !test_bit(In_sync, &disk->rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070 disk->head_position = 0;
2071 mddev->degraded++;
2072 }
2073 }
2074
2075
2076 mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2077 if (!mddev->thread) {
2078 printk(KERN_ERR
2079 "raid10: couldn't allocate thread for %s\n",
2080 mdname(mddev));
2081 goto out_free_conf;
2082 }
2083
2084 printk(KERN_INFO
2085 "raid10: raid set %s active with %d out of %d devices\n",
2086 mdname(mddev), mddev->raid_disks - mddev->degraded,
2087 mddev->raid_disks);
2088 /*
2089 * Ok, everything is just fine now
2090 */
NeilBrown64a742b2007-02-28 20:11:18 -08002091 mddev->array_size = size << (conf->chunk_shift-1);
2092 mddev->resync_max_sectors = size << conf->chunk_shift;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093
NeilBrown7a5febe2005-05-16 21:53:16 -07002094 mddev->queue->unplug_fn = raid10_unplug;
NeilBrown0d129222006-10-03 01:15:54 -07002095 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2096 mddev->queue->backing_dev_info.congested_data = mddev;
NeilBrown7a5febe2005-05-16 21:53:16 -07002097
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 /* Calculate max read-ahead size.
2099 * We need to readahead at least twice a whole stripe....
2100 * maybe...
2101 */
2102 {
NeilBrown8932c2e2006-06-26 00:27:36 -07002103 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104 stripe /= conf->near_copies;
2105 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2106 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2107 }
2108
2109 if (conf->near_copies < mddev->raid_disks)
2110 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2111 return 0;
2112
2113out_free_conf:
2114 if (conf->r10bio_pool)
2115 mempool_destroy(conf->r10bio_pool);
NeilBrown1345b1d2006-01-06 00:20:40 -08002116 safe_put_page(conf->tmppage);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07002117 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118 kfree(conf);
2119 mddev->private = NULL;
2120out:
2121 return -EIO;
2122}
2123
2124static int stop(mddev_t *mddev)
2125{
2126 conf_t *conf = mddev_to_conf(mddev);
2127
2128 md_unregister_thread(mddev->thread);
2129 mddev->thread = NULL;
2130 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2131 if (conf->r10bio_pool)
2132 mempool_destroy(conf->r10bio_pool);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07002133 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134 kfree(conf);
2135 mddev->private = NULL;
2136 return 0;
2137}
2138
NeilBrown6cce3b22006-01-06 00:20:16 -08002139static void raid10_quiesce(mddev_t *mddev, int state)
2140{
2141 conf_t *conf = mddev_to_conf(mddev);
2142
2143 switch(state) {
2144 case 1:
2145 raise_barrier(conf, 0);
2146 break;
2147 case 0:
2148 lower_barrier(conf);
2149 break;
2150 }
2151 if (mddev->thread) {
2152 if (mddev->bitmap)
2153 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2154 else
2155 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2156 md_wakeup_thread(mddev->thread);
2157 }
2158}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159
NeilBrown2604b702006-01-06 00:20:36 -08002160static struct mdk_personality raid10_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161{
2162 .name = "raid10",
NeilBrown2604b702006-01-06 00:20:36 -08002163 .level = 10,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 .owner = THIS_MODULE,
2165 .make_request = make_request,
2166 .run = run,
2167 .stop = stop,
2168 .status = status,
2169 .error_handler = error,
2170 .hot_add_disk = raid10_add_disk,
2171 .hot_remove_disk= raid10_remove_disk,
2172 .spare_active = raid10_spare_active,
2173 .sync_request = sync_request,
NeilBrown6cce3b22006-01-06 00:20:16 -08002174 .quiesce = raid10_quiesce,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175};
2176
2177static int __init raid_init(void)
2178{
NeilBrown2604b702006-01-06 00:20:36 -08002179 return register_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180}
2181
2182static void raid_exit(void)
2183{
NeilBrown2604b702006-01-06 00:20:36 -08002184 unregister_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185}
2186
2187module_init(raid_init);
2188module_exit(raid_exit);
2189MODULE_LICENSE("GPL");
2190MODULE_ALIAS("md-personality-9"); /* RAID10 */
NeilBrownd9d166c2006-01-06 00:20:51 -08002191MODULE_ALIAS("md-raid10");
NeilBrown2604b702006-01-06 00:20:36 -08002192MODULE_ALIAS("md-level-10");