blob: d1d6891263469f5c84be025c004e411ed93f2171 [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
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090021#include <linux/slab.h>
Stephen Rothwell25570722008-10-15 09:09:21 +110022#include <linux/delay.h>
NeilBrownbff61972009-03-31 14:33:13 +110023#include <linux/blkdev.h>
NeilBrownbff61972009-03-31 14:33:13 +110024#include <linux/seq_file.h>
NeilBrown43b2e5d2009-03-31 14:33:13 +110025#include "md.h"
Christoph Hellwigef740c32009-03-31 14:27:03 +110026#include "raid10.h"
Trela, Maciejdab8b292010-03-08 16:02:45 +110027#include "raid0.h"
Christoph Hellwigef740c32009-03-31 14:27:03 +110028#include "bitmap.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070029
30/*
31 * RAID10 provides a combination of RAID0 and RAID1 functionality.
32 * The layout of data is defined by
33 * chunk_size
34 * raid_disks
35 * near_copies (stored in low byte of layout)
36 * far_copies (stored in second byte of layout)
NeilBrownc93983b2006-06-26 00:27:41 -070037 * far_offset (stored in bit 16 of layout )
Linus Torvalds1da177e2005-04-16 15:20:36 -070038 *
39 * The data to be stored is divided into chunks using chunksize.
40 * Each device is divided into far_copies sections.
41 * In each section, chunks are laid out in a style similar to raid0, but
42 * near_copies copies of each chunk is stored (each on a different drive).
43 * The starting device for each section is offset near_copies from the starting
44 * device of the previous section.
NeilBrownc93983b2006-06-26 00:27:41 -070045 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
Linus Torvalds1da177e2005-04-16 15:20:36 -070046 * drive.
47 * near_copies and far_copies must be at least one, and their product is at most
48 * raid_disks.
NeilBrownc93983b2006-06-26 00:27:41 -070049 *
50 * If far_offset is true, then the far_copies are handled a bit differently.
51 * The copies are still in different stripes, but instead of be very far apart
52 * on disk, there are adjacent stripes.
Linus Torvalds1da177e2005-04-16 15:20:36 -070053 */
54
55/*
56 * Number of guaranteed r10bios in case of extreme VM load:
57 */
58#define NR_RAID10_BIOS 256
59
60static void unplug_slaves(mddev_t *mddev);
61
NeilBrown0a27ec92006-01-06 00:20:13 -080062static void allow_barrier(conf_t *conf);
63static void lower_barrier(conf_t *conf);
64
Al Virodd0fc662005-10-07 07:46:04 +010065static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070066{
67 conf_t *conf = data;
68 r10bio_t *r10_bio;
69 int size = offsetof(struct r10bio_s, devs[conf->copies]);
70
71 /* allocate a r10bio with room for raid_disks entries in the bios array */
NeilBrown9ffae0c2006-01-06 00:20:32 -080072 r10_bio = kzalloc(size, gfp_flags);
NeilBrowned9bfdf2009-10-16 15:55:44 +110073 if (!r10_bio && conf->mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -070074 unplug_slaves(conf->mddev);
75
76 return r10_bio;
77}
78
79static void r10bio_pool_free(void *r10_bio, void *data)
80{
81 kfree(r10_bio);
82}
83
NeilBrown0310fa22008-08-05 15:54:14 +100084/* Maximum size of each resync request */
Linus Torvalds1da177e2005-04-16 15:20:36 -070085#define RESYNC_BLOCK_SIZE (64*1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -070086#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
NeilBrown0310fa22008-08-05 15:54:14 +100087/* amount of memory to reserve for resync requests */
88#define RESYNC_WINDOW (1024*1024)
89/* maximum number of concurrent requests, memory permitting */
90#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -070091
92/*
93 * When performing a resync, we need to read and compare, so
94 * we need as many pages are there are copies.
95 * When performing a recovery, we need 2 bios, one for read,
96 * one for write (we recover only one drive per r10buf)
97 *
98 */
Al Virodd0fc662005-10-07 07:46:04 +010099static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100{
101 conf_t *conf = data;
102 struct page *page;
103 r10bio_t *r10_bio;
104 struct bio *bio;
105 int i, j;
106 int nalloc;
107
108 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
109 if (!r10_bio) {
110 unplug_slaves(conf->mddev);
111 return NULL;
112 }
113
114 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
115 nalloc = conf->copies; /* resync */
116 else
117 nalloc = 2; /* recovery */
118
119 /*
120 * Allocate bios.
121 */
122 for (j = nalloc ; j-- ; ) {
123 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
124 if (!bio)
125 goto out_free_bio;
126 r10_bio->devs[j].bio = bio;
127 }
128 /*
129 * Allocate RESYNC_PAGES data pages and attach them
130 * where needed.
131 */
132 for (j = 0 ; j < nalloc; j++) {
133 bio = r10_bio->devs[j].bio;
134 for (i = 0; i < RESYNC_PAGES; i++) {
135 page = alloc_page(gfp_flags);
136 if (unlikely(!page))
137 goto out_free_pages;
138
139 bio->bi_io_vec[i].bv_page = page;
140 }
141 }
142
143 return r10_bio;
144
145out_free_pages:
146 for ( ; i > 0 ; i--)
NeilBrown1345b1d2006-01-06 00:20:40 -0800147 safe_put_page(bio->bi_io_vec[i-1].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148 while (j--)
149 for (i = 0; i < RESYNC_PAGES ; i++)
NeilBrown1345b1d2006-01-06 00:20:40 -0800150 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151 j = -1;
152out_free_bio:
153 while ( ++j < nalloc )
154 bio_put(r10_bio->devs[j].bio);
155 r10bio_pool_free(r10_bio, conf);
156 return NULL;
157}
158
159static void r10buf_pool_free(void *__r10_bio, void *data)
160{
161 int i;
162 conf_t *conf = data;
163 r10bio_t *r10bio = __r10_bio;
164 int j;
165
166 for (j=0; j < conf->copies; j++) {
167 struct bio *bio = r10bio->devs[j].bio;
168 if (bio) {
169 for (i = 0; i < RESYNC_PAGES; i++) {
NeilBrown1345b1d2006-01-06 00:20:40 -0800170 safe_put_page(bio->bi_io_vec[i].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 bio->bi_io_vec[i].bv_page = NULL;
172 }
173 bio_put(bio);
174 }
175 }
176 r10bio_pool_free(r10bio, conf);
177}
178
179static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
180{
181 int i;
182
183 for (i = 0; i < conf->copies; i++) {
184 struct bio **bio = & r10_bio->devs[i].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -0800185 if (*bio && *bio != IO_BLOCKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186 bio_put(*bio);
187 *bio = NULL;
188 }
189}
190
Arjan van de Ven858119e2006-01-14 13:20:43 -0800191static void free_r10bio(r10bio_t *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192{
NeilBrown070ec552009-06-16 16:54:21 +1000193 conf_t *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194
195 /*
196 * Wake up any possible resync thread that waits for the device
197 * to go idle.
198 */
NeilBrown0a27ec92006-01-06 00:20:13 -0800199 allow_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200
201 put_all_bios(conf, r10_bio);
202 mempool_free(r10_bio, conf->r10bio_pool);
203}
204
Arjan van de Ven858119e2006-01-14 13:20:43 -0800205static void put_buf(r10bio_t *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206{
NeilBrown070ec552009-06-16 16:54:21 +1000207 conf_t *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208
209 mempool_free(r10_bio, conf->r10buf_pool);
210
NeilBrown0a27ec92006-01-06 00:20:13 -0800211 lower_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212}
213
214static void reschedule_retry(r10bio_t *r10_bio)
215{
216 unsigned long flags;
217 mddev_t *mddev = r10_bio->mddev;
NeilBrown070ec552009-06-16 16:54:21 +1000218 conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219
220 spin_lock_irqsave(&conf->device_lock, flags);
221 list_add(&r10_bio->retry_list, &conf->retry_list);
NeilBrown4443ae12006-01-06 00:20:28 -0800222 conf->nr_queued ++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223 spin_unlock_irqrestore(&conf->device_lock, flags);
224
Arthur Jones388667b2008-07-25 12:03:38 -0700225 /* wake up frozen array... */
226 wake_up(&conf->wait_barrier);
227
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228 md_wakeup_thread(mddev->thread);
229}
230
231/*
232 * raid_end_bio_io() is called when we have finished servicing a mirrored
233 * operation and are ready to return a success/failure code to the buffer
234 * cache layer.
235 */
236static void raid_end_bio_io(r10bio_t *r10_bio)
237{
238 struct bio *bio = r10_bio->master_bio;
239
NeilBrown6712ecf2007-09-27 12:47:43 +0200240 bio_endio(bio,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
242 free_r10bio(r10_bio);
243}
244
245/*
246 * Update disk head position estimator based on IRQ completion info.
247 */
248static inline void update_head_pos(int slot, r10bio_t *r10_bio)
249{
NeilBrown070ec552009-06-16 16:54:21 +1000250 conf_t *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251
252 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
253 r10_bio->devs[slot].addr + (r10_bio->sectors);
254}
255
NeilBrown6712ecf2007-09-27 12:47:43 +0200256static void raid10_end_read_request(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257{
258 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
H Hartley Sweeten7b928132010-03-08 16:02:40 +1100259 r10bio_t *r10_bio = bio->bi_private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 int slot, dev;
NeilBrown070ec552009-06-16 16:54:21 +1000261 conf_t *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263
264 slot = r10_bio->read_slot;
265 dev = r10_bio->devs[slot].devnum;
266 /*
267 * this branch is our 'one mirror IO has finished' event handler:
268 */
NeilBrown4443ae12006-01-06 00:20:28 -0800269 update_head_pos(slot, r10_bio);
270
271 if (uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272 /*
273 * Set R10BIO_Uptodate in our master bio, so that
274 * we will return a good error code to the higher
275 * levels even if IO on some other mirrored buffer fails.
276 *
277 * The 'master' represents the composite IO operation to
278 * user-side. So if something waits for IO, then it will
279 * wait for the 'master' bio.
280 */
281 set_bit(R10BIO_Uptodate, &r10_bio->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 raid_end_bio_io(r10_bio);
NeilBrown4443ae12006-01-06 00:20:28 -0800283 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284 /*
285 * oops, read error:
286 */
287 char b[BDEVNAME_SIZE];
288 if (printk_ratelimit())
NeilBrown128595e2010-05-03 14:47:14 +1000289 printk(KERN_ERR "md/raid10:%s: %s: rescheduling sector %llu\n",
290 mdname(conf->mddev),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
292 reschedule_retry(r10_bio);
293 }
294
295 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296}
297
NeilBrown6712ecf2007-09-27 12:47:43 +0200298static void raid10_end_write_request(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299{
300 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
H Hartley Sweeten7b928132010-03-08 16:02:40 +1100301 r10bio_t *r10_bio = bio->bi_private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302 int slot, dev;
NeilBrown070ec552009-06-16 16:54:21 +1000303 conf_t *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305 for (slot = 0; slot < conf->copies; slot++)
306 if (r10_bio->devs[slot].bio == bio)
307 break;
308 dev = r10_bio->devs[slot].devnum;
309
310 /*
311 * this branch is our 'one mirror IO has finished' event handler:
312 */
NeilBrown6cce3b22006-01-06 00:20:16 -0800313 if (!uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
NeilBrown6cce3b22006-01-06 00:20:16 -0800315 /* an I/O failed, we can't clear the bitmap */
316 set_bit(R10BIO_Degraded, &r10_bio->state);
317 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 /*
319 * Set R10BIO_Uptodate in our master bio, so that
320 * we will return a good error code for to the higher
321 * levels even if IO on some other mirrored buffer fails.
322 *
323 * The 'master' represents the composite IO operation to
324 * user-side. So if something waits for IO, then it will
325 * wait for the 'master' bio.
326 */
327 set_bit(R10BIO_Uptodate, &r10_bio->state);
328
329 update_head_pos(slot, r10_bio);
330
331 /*
332 *
333 * Let's see if all mirrored write operations have finished
334 * already.
335 */
336 if (atomic_dec_and_test(&r10_bio->remaining)) {
NeilBrown6cce3b22006-01-06 00:20:16 -0800337 /* clear the bitmap if all writes complete successfully */
338 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
339 r10_bio->sectors,
340 !test_bit(R10BIO_Degraded, &r10_bio->state),
341 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 md_write_end(r10_bio->mddev);
343 raid_end_bio_io(r10_bio);
344 }
345
346 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347}
348
349
350/*
351 * RAID10 layout manager
352 * Aswell as the chunksize and raid_disks count, there are two
353 * parameters: near_copies and far_copies.
354 * near_copies * far_copies must be <= raid_disks.
355 * Normally one of these will be 1.
356 * If both are 1, we get raid0.
357 * If near_copies == raid_disks, we get raid1.
358 *
359 * Chunks are layed out in raid0 style with near_copies copies of the
360 * first chunk, followed by near_copies copies of the next chunk and
361 * so on.
362 * If far_copies > 1, then after 1/far_copies of the array has been assigned
363 * as described above, we start again with a device offset of near_copies.
364 * So we effectively have another copy of the whole array further down all
365 * the drives, but with blocks on different drives.
366 * With this layout, and block is never stored twice on the one device.
367 *
368 * raid10_find_phys finds the sector offset of a given virtual sector
NeilBrownc93983b2006-06-26 00:27:41 -0700369 * on each device that it is on.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370 *
371 * raid10_find_virt does the reverse mapping, from a device and a
372 * sector offset to a virtual address
373 */
374
375static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
376{
377 int n,f;
378 sector_t sector;
379 sector_t chunk;
380 sector_t stripe;
381 int dev;
382
383 int slot = 0;
384
385 /* now calculate first sector/dev */
386 chunk = r10bio->sector >> conf->chunk_shift;
387 sector = r10bio->sector & conf->chunk_mask;
388
389 chunk *= conf->near_copies;
390 stripe = chunk;
391 dev = sector_div(stripe, conf->raid_disks);
NeilBrownc93983b2006-06-26 00:27:41 -0700392 if (conf->far_offset)
393 stripe *= conf->far_copies;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394
395 sector += stripe << conf->chunk_shift;
396
397 /* and calculate all the others */
398 for (n=0; n < conf->near_copies; n++) {
399 int d = dev;
400 sector_t s = sector;
401 r10bio->devs[slot].addr = sector;
402 r10bio->devs[slot].devnum = d;
403 slot++;
404
405 for (f = 1; f < conf->far_copies; f++) {
406 d += conf->near_copies;
407 if (d >= conf->raid_disks)
408 d -= conf->raid_disks;
409 s += conf->stride;
410 r10bio->devs[slot].devnum = d;
411 r10bio->devs[slot].addr = s;
412 slot++;
413 }
414 dev++;
415 if (dev >= conf->raid_disks) {
416 dev = 0;
417 sector += (conf->chunk_mask + 1);
418 }
419 }
420 BUG_ON(slot != conf->copies);
421}
422
423static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
424{
425 sector_t offset, chunk, vchunk;
426
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427 offset = sector & conf->chunk_mask;
NeilBrownc93983b2006-06-26 00:27:41 -0700428 if (conf->far_offset) {
429 int fc;
430 chunk = sector >> conf->chunk_shift;
431 fc = sector_div(chunk, conf->far_copies);
432 dev -= fc * conf->near_copies;
433 if (dev < 0)
434 dev += conf->raid_disks;
435 } else {
NeilBrown64a742b2007-02-28 20:11:18 -0800436 while (sector >= conf->stride) {
NeilBrownc93983b2006-06-26 00:27:41 -0700437 sector -= conf->stride;
438 if (dev < conf->near_copies)
439 dev += conf->raid_disks - conf->near_copies;
440 else
441 dev -= conf->near_copies;
442 }
443 chunk = sector >> conf->chunk_shift;
444 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 vchunk = chunk * conf->raid_disks + dev;
446 sector_div(vchunk, conf->near_copies);
447 return (vchunk << conf->chunk_shift) + offset;
448}
449
450/**
451 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
452 * @q: request queue
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200453 * @bvm: properties of new bio
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 * @biovec: the request that could be merged to it.
455 *
456 * Return amount of bytes we can accept at this offset
457 * If near_copies == raid_disk, there are no striping issues,
458 * but in that case, the function isn't called at all.
459 */
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200460static int raid10_mergeable_bvec(struct request_queue *q,
461 struct bvec_merge_data *bvm,
462 struct bio_vec *biovec)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463{
464 mddev_t *mddev = q->queuedata;
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200465 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 int max;
Andre Noll9d8f0362009-06-18 08:45:01 +1000467 unsigned int chunk_sectors = mddev->chunk_sectors;
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200468 unsigned int bio_sectors = bvm->bi_size >> 9;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469
470 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
471 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200472 if (max <= biovec->bv_len && bio_sectors == 0)
473 return biovec->bv_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474 else
475 return max;
476}
477
478/*
479 * This routine returns the disk from which the requested read should
480 * be done. There is a per-array 'next expected sequential IO' sector
481 * number - if this matches on the next IO then we use the last disk.
482 * There is also a per-disk 'last know head position' sector that is
483 * maintained from IRQ contexts, both the normal and the resync IO
484 * completion handlers update this position correctly. If there is no
485 * perfect sequential match then we pick the disk whose head is closest.
486 *
487 * If there are 2 mirrors in the same 2 devices, performance degrades
488 * because position is mirror, not device based.
489 *
490 * The rdev for the device selected will have nr_pending incremented.
491 */
492
493/*
494 * FIXME: possibly should rethink readbalancing and do it differently
495 * depending on near_copies / far_copies geometry.
496 */
497static int read_balance(conf_t *conf, r10bio_t *r10_bio)
498{
NeilBrownaf3a2cd2010-05-08 08:20:17 +1000499 const sector_t this_sector = r10_bio->sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 int disk, slot, nslot;
501 const int sectors = r10_bio->sectors;
502 sector_t new_distance, current_distance;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800503 mdk_rdev_t *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504
505 raid10_find_phys(conf, r10_bio);
506 rcu_read_lock();
507 /*
508 * Check if we can balance. We can balance on the whole
NeilBrown6cce3b22006-01-06 00:20:16 -0800509 * device if no resync is going on (recovery is ok), or below
510 * the resync window. We take the first readable disk when
511 * above the resync window.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 */
513 if (conf->mddev->recovery_cp < MaxSector
514 && (this_sector + sectors >= conf->next_resync)) {
515 /* make sure that disk is operational */
516 slot = 0;
517 disk = r10_bio->devs[slot].devnum;
518
Suzanne Woodd6065f72005-11-08 21:39:27 -0800519 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800520 r10_bio->devs[slot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800521 !test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 slot++;
523 if (slot == conf->copies) {
524 slot = 0;
525 disk = -1;
526 break;
527 }
528 disk = r10_bio->devs[slot].devnum;
529 }
530 goto rb_out;
531 }
532
533
534 /* make sure the disk is operational */
535 slot = 0;
536 disk = r10_bio->devs[slot].devnum;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800537 while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800538 r10_bio->devs[slot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800539 !test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540 slot ++;
541 if (slot == conf->copies) {
542 disk = -1;
543 goto rb_out;
544 }
545 disk = r10_bio->devs[slot].devnum;
546 }
547
548
NeilBrown3ec67ac2005-09-09 16:23:40 -0700549 current_distance = abs(r10_bio->devs[slot].addr -
550 conf->mirrors[disk].head_position);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551
Keld Simonsen8ed3a192008-03-04 14:29:34 -0800552 /* Find the disk whose head is closest,
553 * or - for far > 1 - find the closest to partition beginning */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554
555 for (nslot = slot; nslot < conf->copies; nslot++) {
556 int ndisk = r10_bio->devs[nslot].devnum;
557
558
Suzanne Woodd6065f72005-11-08 21:39:27 -0800559 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800560 r10_bio->devs[nslot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800561 !test_bit(In_sync, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 continue;
563
NeilBrown22dfdf52005-11-28 13:44:09 -0800564 /* This optimisation is debatable, and completely destroys
565 * sequential read speed for 'far copies' arrays. So only
566 * keep it for 'near' arrays, and review those later.
567 */
568 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 disk = ndisk;
570 slot = nslot;
571 break;
572 }
Keld Simonsen8ed3a192008-03-04 14:29:34 -0800573
574 /* for far > 1 always use the lowest address */
575 if (conf->far_copies > 1)
576 new_distance = r10_bio->devs[nslot].addr;
577 else
578 new_distance = abs(r10_bio->devs[nslot].addr -
579 conf->mirrors[ndisk].head_position);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 if (new_distance < current_distance) {
581 current_distance = new_distance;
582 disk = ndisk;
583 slot = nslot;
584 }
585 }
586
587rb_out:
588 r10_bio->read_slot = slot;
589/* conf->next_seq_sect = this_sector + sectors;*/
590
Suzanne Woodd6065f72005-11-08 21:39:27 -0800591 if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
NeilBrown29fc7e32006-02-03 03:03:41 -0800593 else
594 disk = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595 rcu_read_unlock();
596
597 return disk;
598}
599
600static void unplug_slaves(mddev_t *mddev)
601{
NeilBrown070ec552009-06-16 16:54:21 +1000602 conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603 int i;
604
605 rcu_read_lock();
NeilBrown84707f32010-03-16 17:23:35 +1100606 for (i=0; i < conf->raid_disks; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800607 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -0800608 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200609 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610
611 atomic_inc(&rdev->nr_pending);
612 rcu_read_unlock();
613
Alan D. Brunelle2ad8b1e2007-11-07 14:26:56 -0500614 blk_unplug(r_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615
616 rdev_dec_pending(rdev, mddev);
617 rcu_read_lock();
618 }
619 }
620 rcu_read_unlock();
621}
622
Jens Axboe165125e2007-07-24 09:28:11 +0200623static void raid10_unplug(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624{
NeilBrown6cce3b22006-01-06 00:20:16 -0800625 mddev_t *mddev = q->queuedata;
626
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 unplug_slaves(q->queuedata);
NeilBrown6cce3b22006-01-06 00:20:16 -0800628 md_wakeup_thread(mddev->thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629}
630
NeilBrown0d129222006-10-03 01:15:54 -0700631static int raid10_congested(void *data, int bits)
632{
633 mddev_t *mddev = data;
NeilBrown070ec552009-06-16 16:54:21 +1000634 conf_t *conf = mddev->private;
NeilBrown0d129222006-10-03 01:15:54 -0700635 int i, ret = 0;
636
NeilBrown3fa841d2009-09-23 18:10:29 +1000637 if (mddev_congested(mddev, bits))
638 return 1;
NeilBrown0d129222006-10-03 01:15:54 -0700639 rcu_read_lock();
NeilBrown84707f32010-03-16 17:23:35 +1100640 for (i = 0; i < conf->raid_disks && ret == 0; i++) {
NeilBrown0d129222006-10-03 01:15:54 -0700641 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
642 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200643 struct request_queue *q = bdev_get_queue(rdev->bdev);
NeilBrown0d129222006-10-03 01:15:54 -0700644
645 ret |= bdi_congested(&q->backing_dev_info, bits);
646 }
647 }
648 rcu_read_unlock();
649 return ret;
650}
651
NeilBrowna35e63e2008-03-04 14:29:29 -0800652static int flush_pending_writes(conf_t *conf)
653{
654 /* Any writes that have been queued but are awaiting
655 * bitmap updates get flushed here.
656 * We return 1 if any requests were actually submitted.
657 */
658 int rv = 0;
NeilBrown0d129222006-10-03 01:15:54 -0700659
NeilBrowna35e63e2008-03-04 14:29:29 -0800660 spin_lock_irq(&conf->device_lock);
661
662 if (conf->pending_bio_list.head) {
663 struct bio *bio;
664 bio = bio_list_get(&conf->pending_bio_list);
665 blk_remove_plug(conf->mddev->queue);
666 spin_unlock_irq(&conf->device_lock);
667 /* flush any pending bitmap writes to disk
668 * before proceeding w/ I/O */
669 bitmap_unplug(conf->mddev->bitmap);
670
671 while (bio) { /* submit pending writes */
672 struct bio *next = bio->bi_next;
673 bio->bi_next = NULL;
674 generic_make_request(bio);
675 bio = next;
676 }
677 rv = 1;
678 } else
679 spin_unlock_irq(&conf->device_lock);
680 return rv;
681}
NeilBrown0a27ec92006-01-06 00:20:13 -0800682/* Barriers....
683 * Sometimes we need to suspend IO while we do something else,
684 * either some resync/recovery, or reconfigure the array.
685 * To do this we raise a 'barrier'.
686 * The 'barrier' is a counter that can be raised multiple times
687 * to count how many activities are happening which preclude
688 * normal IO.
689 * We can only raise the barrier if there is no pending IO.
690 * i.e. if nr_pending == 0.
691 * We choose only to raise the barrier if no-one is waiting for the
692 * barrier to go down. This means that as soon as an IO request
693 * is ready, no other operations which require a barrier will start
694 * until the IO request has had a chance.
695 *
696 * So: regular IO calls 'wait_barrier'. When that returns there
697 * is no backgroup IO happening, It must arrange to call
698 * allow_barrier when it has finished its IO.
699 * backgroup IO calls must call raise_barrier. Once that returns
700 * there is no normal IO happeing. It must arrange to call
701 * lower_barrier when the particular background IO completes.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700702 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703
NeilBrown6cce3b22006-01-06 00:20:16 -0800704static void raise_barrier(conf_t *conf, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705{
NeilBrown6cce3b22006-01-06 00:20:16 -0800706 BUG_ON(force && !conf->barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707 spin_lock_irq(&conf->resync_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708
NeilBrown6cce3b22006-01-06 00:20:16 -0800709 /* Wait until no block IO is waiting (unless 'force') */
710 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
NeilBrown0a27ec92006-01-06 00:20:13 -0800711 conf->resync_lock,
712 raid10_unplug(conf->mddev->queue));
713
714 /* block any new IO from starting */
715 conf->barrier++;
716
717 /* No wait for all pending IO to complete */
718 wait_event_lock_irq(conf->wait_barrier,
719 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
720 conf->resync_lock,
721 raid10_unplug(conf->mddev->queue));
722
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 spin_unlock_irq(&conf->resync_lock);
724}
725
NeilBrown0a27ec92006-01-06 00:20:13 -0800726static void lower_barrier(conf_t *conf)
727{
728 unsigned long flags;
729 spin_lock_irqsave(&conf->resync_lock, flags);
730 conf->barrier--;
731 spin_unlock_irqrestore(&conf->resync_lock, flags);
732 wake_up(&conf->wait_barrier);
733}
734
735static void wait_barrier(conf_t *conf)
736{
737 spin_lock_irq(&conf->resync_lock);
738 if (conf->barrier) {
739 conf->nr_waiting++;
740 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
741 conf->resync_lock,
742 raid10_unplug(conf->mddev->queue));
743 conf->nr_waiting--;
744 }
745 conf->nr_pending++;
746 spin_unlock_irq(&conf->resync_lock);
747}
748
749static void allow_barrier(conf_t *conf)
750{
751 unsigned long flags;
752 spin_lock_irqsave(&conf->resync_lock, flags);
753 conf->nr_pending--;
754 spin_unlock_irqrestore(&conf->resync_lock, flags);
755 wake_up(&conf->wait_barrier);
756}
757
NeilBrown4443ae12006-01-06 00:20:28 -0800758static void freeze_array(conf_t *conf)
759{
760 /* stop syncio and normal IO and wait for everything to
NeilBrownf1885932006-01-06 00:20:42 -0800761 * go quiet.
NeilBrown4443ae12006-01-06 00:20:28 -0800762 * We increment barrier and nr_waiting, and then
NeilBrown1c830532008-03-04 14:29:35 -0800763 * wait until nr_pending match nr_queued+1
764 * This is called in the context of one normal IO request
765 * that has failed. Thus any sync request that might be pending
766 * will be blocked by nr_pending, and we need to wait for
767 * pending IO requests to complete or be queued for re-try.
768 * Thus the number queued (nr_queued) plus this request (1)
769 * must match the number of pending IOs (nr_pending) before
770 * we continue.
NeilBrown4443ae12006-01-06 00:20:28 -0800771 */
772 spin_lock_irq(&conf->resync_lock);
773 conf->barrier++;
774 conf->nr_waiting++;
775 wait_event_lock_irq(conf->wait_barrier,
NeilBrown1c830532008-03-04 14:29:35 -0800776 conf->nr_pending == conf->nr_queued+1,
NeilBrown4443ae12006-01-06 00:20:28 -0800777 conf->resync_lock,
NeilBrowna35e63e2008-03-04 14:29:29 -0800778 ({ flush_pending_writes(conf);
779 raid10_unplug(conf->mddev->queue); }));
NeilBrown4443ae12006-01-06 00:20:28 -0800780 spin_unlock_irq(&conf->resync_lock);
781}
782
783static void unfreeze_array(conf_t *conf)
784{
785 /* reverse the effect of the freeze */
786 spin_lock_irq(&conf->resync_lock);
787 conf->barrier--;
788 conf->nr_waiting--;
789 wake_up(&conf->wait_barrier);
790 spin_unlock_irq(&conf->resync_lock);
791}
792
NeilBrown21a52c62010-04-01 15:02:13 +1100793static int make_request(mddev_t *mddev, struct bio * bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794{
NeilBrown070ec552009-06-16 16:54:21 +1000795 conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 mirror_info_t *mirror;
797 r10bio_t *r10_bio;
798 struct bio *read_bio;
799 int i;
800 int chunk_sects = conf->chunk_mask + 1;
Jens Axboea3623572005-11-01 09:26:16 +0100801 const int rw = bio_data_dir(bio);
Jens Axboe1f98a132009-09-11 14:32:04 +0200802 const bool do_sync = bio_rw_flagged(bio, BIO_RW_SYNCIO);
NeilBrown6cce3b22006-01-06 00:20:16 -0800803 struct bio_list bl;
804 unsigned long flags;
Dan Williams6bfe0b42008-04-30 00:52:32 -0700805 mdk_rdev_t *blocked_rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806
Jens Axboe1f98a132009-09-11 14:32:04 +0200807 if (unlikely(bio_rw_flagged(bio, BIO_RW_BARRIER))) {
NeilBrowna2826aa2009-12-14 12:49:49 +1100808 md_barrier_request(mddev, bio);
NeilBrowne5dcdd82005-09-09 16:23:41 -0700809 return 0;
810 }
811
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812 /* If this request crosses a chunk boundary, we need to
813 * split it. This will only happen for 1 PAGE (or less) requests.
814 */
815 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
816 > chunk_sects &&
817 conf->near_copies < conf->raid_disks)) {
818 struct bio_pair *bp;
819 /* Sanity check -- queue functions should prevent this happening */
820 if (bio->bi_vcnt != 1 ||
821 bio->bi_idx != 0)
822 goto bad_map;
823 /* This is a one page bio that upper layers
824 * refuse to split for us, so we need to split it.
825 */
Denis ChengRq6feef532008-10-09 08:57:05 +0200826 bp = bio_split(bio,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
NeilBrown51e9ac72010-08-07 21:17:00 +1000828
829 /* Each of these 'make_request' calls will call 'wait_barrier'.
830 * If the first succeeds but the second blocks due to the resync
831 * thread raising the barrier, we will deadlock because the
832 * IO to the underlying device will be queued in generic_make_request
833 * and will never complete, so will never reduce nr_pending.
834 * So increment nr_waiting here so no new raise_barriers will
835 * succeed, and so the second wait_barrier cannot block.
836 */
837 spin_lock_irq(&conf->resync_lock);
838 conf->nr_waiting++;
839 spin_unlock_irq(&conf->resync_lock);
840
NeilBrown21a52c62010-04-01 15:02:13 +1100841 if (make_request(mddev, &bp->bio1))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700842 generic_make_request(&bp->bio1);
NeilBrown21a52c62010-04-01 15:02:13 +1100843 if (make_request(mddev, &bp->bio2))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844 generic_make_request(&bp->bio2);
845
NeilBrown51e9ac72010-08-07 21:17:00 +1000846 spin_lock_irq(&conf->resync_lock);
847 conf->nr_waiting--;
848 wake_up(&conf->wait_barrier);
849 spin_unlock_irq(&conf->resync_lock);
850
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 bio_pair_release(bp);
852 return 0;
853 bad_map:
NeilBrown128595e2010-05-03 14:47:14 +1000854 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
855 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
857
NeilBrown6712ecf2007-09-27 12:47:43 +0200858 bio_io_error(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700859 return 0;
860 }
861
NeilBrown3d310eb2005-06-21 17:17:26 -0700862 md_write_start(mddev, bio);
NeilBrown06d91a52005-06-21 17:17:12 -0700863
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864 /*
865 * Register the new request and wait if the reconstruction
866 * thread has put up a bar for new requests.
867 * Continue immediately if no resync is active currently.
868 */
NeilBrown0a27ec92006-01-06 00:20:13 -0800869 wait_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
872
873 r10_bio->master_bio = bio;
874 r10_bio->sectors = bio->bi_size >> 9;
875
876 r10_bio->mddev = mddev;
877 r10_bio->sector = bio->bi_sector;
NeilBrown6cce3b22006-01-06 00:20:16 -0800878 r10_bio->state = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879
Jens Axboea3623572005-11-01 09:26:16 +0100880 if (rw == READ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 /*
882 * read balancing logic:
883 */
884 int disk = read_balance(conf, r10_bio);
885 int slot = r10_bio->read_slot;
886 if (disk < 0) {
887 raid_end_bio_io(r10_bio);
888 return 0;
889 }
890 mirror = conf->mirrors + disk;
891
892 read_bio = bio_clone(bio, GFP_NOIO);
893
894 r10_bio->devs[slot].bio = read_bio;
895
896 read_bio->bi_sector = r10_bio->devs[slot].addr +
897 mirror->rdev->data_offset;
898 read_bio->bi_bdev = mirror->rdev->bdev;
899 read_bio->bi_end_io = raid10_end_read_request;
Dmitry Monakhov1ef04fe2009-09-20 05:52:25 +0400900 read_bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901 read_bio->bi_private = r10_bio;
902
903 generic_make_request(read_bio);
904 return 0;
905 }
906
907 /*
908 * WRITE:
909 */
Dan Williams6bfe0b42008-04-30 00:52:32 -0700910 /* first select target devices under rcu_lock and
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 * inc refcount on their rdev. Record them by setting
912 * bios[x] to bio
913 */
914 raid10_find_phys(conf, r10_bio);
Dan Williams6bfe0b42008-04-30 00:52:32 -0700915 retry_write:
Harvey Harrisoncb6969e2008-05-06 20:42:32 -0700916 blocked_rdev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 rcu_read_lock();
918 for (i = 0; i < conf->copies; i++) {
919 int d = r10_bio->devs[i].devnum;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800920 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
Dan Williams6bfe0b42008-04-30 00:52:32 -0700921 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
922 atomic_inc(&rdev->nr_pending);
923 blocked_rdev = rdev;
924 break;
925 }
926 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800927 atomic_inc(&rdev->nr_pending);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 r10_bio->devs[i].bio = bio;
NeilBrown6cce3b22006-01-06 00:20:16 -0800929 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 r10_bio->devs[i].bio = NULL;
NeilBrown6cce3b22006-01-06 00:20:16 -0800931 set_bit(R10BIO_Degraded, &r10_bio->state);
932 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933 }
934 rcu_read_unlock();
935
Dan Williams6bfe0b42008-04-30 00:52:32 -0700936 if (unlikely(blocked_rdev)) {
937 /* Have to wait for this device to get unblocked, then retry */
938 int j;
939 int d;
940
941 for (j = 0; j < i; j++)
942 if (r10_bio->devs[j].bio) {
943 d = r10_bio->devs[j].devnum;
944 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
945 }
946 allow_barrier(conf);
947 md_wait_for_blocked_rdev(blocked_rdev, mddev);
948 wait_barrier(conf);
949 goto retry_write;
950 }
951
NeilBrown6cce3b22006-01-06 00:20:16 -0800952 atomic_set(&r10_bio->remaining, 0);
NeilBrown06d91a52005-06-21 17:17:12 -0700953
NeilBrown6cce3b22006-01-06 00:20:16 -0800954 bio_list_init(&bl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 for (i = 0; i < conf->copies; i++) {
956 struct bio *mbio;
957 int d = r10_bio->devs[i].devnum;
958 if (!r10_bio->devs[i].bio)
959 continue;
960
961 mbio = bio_clone(bio, GFP_NOIO);
962 r10_bio->devs[i].bio = mbio;
963
964 mbio->bi_sector = r10_bio->devs[i].addr+
965 conf->mirrors[d].rdev->data_offset;
966 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
967 mbio->bi_end_io = raid10_end_write_request;
Dmitry Monakhov1ef04fe2009-09-20 05:52:25 +0400968 mbio->bi_rw = WRITE | (do_sync << BIO_RW_SYNCIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 mbio->bi_private = r10_bio;
970
971 atomic_inc(&r10_bio->remaining);
NeilBrown6cce3b22006-01-06 00:20:16 -0800972 bio_list_add(&bl, mbio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 }
974
Arne Redlichf6f953a2007-07-31 00:37:57 -0700975 if (unlikely(!atomic_read(&r10_bio->remaining))) {
976 /* the array is dead */
977 md_write_end(mddev);
978 raid_end_bio_io(r10_bio);
979 return 0;
980 }
981
NeilBrown6cce3b22006-01-06 00:20:16 -0800982 bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
983 spin_lock_irqsave(&conf->device_lock, flags);
984 bio_list_merge(&conf->pending_bio_list, &bl);
985 blk_plug_device(mddev->queue);
986 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987
NeilBrowna35e63e2008-03-04 14:29:29 -0800988 /* In case raid10d snuck in to freeze_array */
989 wake_up(&conf->wait_barrier);
990
Lars Ellenberge3881a62007-01-10 23:15:37 -0800991 if (do_sync)
992 md_wakeup_thread(mddev->thread);
993
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994 return 0;
995}
996
997static void status(struct seq_file *seq, mddev_t *mddev)
998{
NeilBrown070ec552009-06-16 16:54:21 +1000999 conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 int i;
1001
1002 if (conf->near_copies < conf->raid_disks)
Andre Noll9d8f0362009-06-18 08:45:01 +10001003 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 if (conf->near_copies > 1)
1005 seq_printf(seq, " %d near-copies", conf->near_copies);
NeilBrownc93983b2006-06-26 00:27:41 -07001006 if (conf->far_copies > 1) {
1007 if (conf->far_offset)
1008 seq_printf(seq, " %d offset-copies", conf->far_copies);
1009 else
1010 seq_printf(seq, " %d far-copies", conf->far_copies);
1011 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
NeilBrown76186dd2006-10-03 01:15:48 -07001013 conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 for (i = 0; i < conf->raid_disks; i++)
1015 seq_printf(seq, "%s",
1016 conf->mirrors[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001017 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018 seq_printf(seq, "]");
1019}
1020
1021static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1022{
1023 char b[BDEVNAME_SIZE];
NeilBrown070ec552009-06-16 16:54:21 +10001024 conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001025
1026 /*
1027 * If it is not operational, then we have already marked it as dead
1028 * else if it is the last working disks, ignore the error, let the
1029 * next level up know.
1030 * else mark the drive as failed
1031 */
NeilBrownb2d444d2005-11-08 21:39:31 -08001032 if (test_bit(In_sync, &rdev->flags)
NeilBrown76186dd2006-10-03 01:15:48 -07001033 && conf->raid_disks-mddev->degraded == 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034 /*
1035 * Don't fail the drive, just return an IO error.
1036 * The test should really be more sophisticated than
1037 * "working_disks == 1", but it isn't critical, and
1038 * can wait until we do more sophisticated "is the drive
1039 * really dead" tests...
1040 */
1041 return;
NeilBrownc04be0a2006-10-03 01:15:53 -07001042 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1043 unsigned long flags;
1044 spin_lock_irqsave(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045 mddev->degraded++;
NeilBrownc04be0a2006-10-03 01:15:53 -07001046 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047 /*
1048 * if recovery is running, make sure it aborts.
1049 */
NeilBrowndfc70642008-05-23 13:04:39 -07001050 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001051 }
NeilBrownb2d444d2005-11-08 21:39:31 -08001052 set_bit(Faulty, &rdev->flags);
NeilBrown850b2b42006-10-03 01:15:46 -07001053 set_bit(MD_CHANGE_DEVS, &mddev->flags);
NeilBrown128595e2010-05-03 14:47:14 +10001054 printk(KERN_ALERT "md/raid10:%s: Disk failure on %s, disabling device.\n"
1055 KERN_ALERT "md/raid10:%s: Operation continuing on %d devices.\n",
1056 mdname(mddev), bdevname(rdev->bdev, b),
1057 mdname(mddev), conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001058}
1059
1060static void print_conf(conf_t *conf)
1061{
1062 int i;
1063 mirror_info_t *tmp;
1064
NeilBrown128595e2010-05-03 14:47:14 +10001065 printk(KERN_DEBUG "RAID10 conf printout:\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066 if (!conf) {
NeilBrown128595e2010-05-03 14:47:14 +10001067 printk(KERN_DEBUG "(!conf)\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068 return;
1069 }
NeilBrown128595e2010-05-03 14:47:14 +10001070 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071 conf->raid_disks);
1072
1073 for (i = 0; i < conf->raid_disks; i++) {
1074 char b[BDEVNAME_SIZE];
1075 tmp = conf->mirrors + i;
1076 if (tmp->rdev)
NeilBrown128595e2010-05-03 14:47:14 +10001077 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
NeilBrownb2d444d2005-11-08 21:39:31 -08001078 i, !test_bit(In_sync, &tmp->rdev->flags),
1079 !test_bit(Faulty, &tmp->rdev->flags),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 bdevname(tmp->rdev->bdev,b));
1081 }
1082}
1083
1084static void close_sync(conf_t *conf)
1085{
NeilBrown0a27ec92006-01-06 00:20:13 -08001086 wait_barrier(conf);
1087 allow_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088
1089 mempool_destroy(conf->r10buf_pool);
1090 conf->r10buf_pool = NULL;
1091}
1092
NeilBrown6d508242005-09-09 16:24:03 -07001093/* check if there are enough drives for
1094 * every block to appear on atleast one
1095 */
1096static int enough(conf_t *conf)
1097{
1098 int first = 0;
1099
1100 do {
1101 int n = conf->copies;
1102 int cnt = 0;
1103 while (n--) {
1104 if (conf->mirrors[first].rdev)
1105 cnt++;
1106 first = (first+1) % conf->raid_disks;
1107 }
1108 if (cnt == 0)
1109 return 0;
1110 } while (first != 0);
1111 return 1;
1112}
1113
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114static int raid10_spare_active(mddev_t *mddev)
1115{
1116 int i;
1117 conf_t *conf = mddev->private;
1118 mirror_info_t *tmp;
1119
1120 /*
1121 * Find all non-in_sync disks within the RAID10 configuration
1122 * and mark them in_sync
1123 */
1124 for (i = 0; i < conf->raid_disks; i++) {
1125 tmp = conf->mirrors + i;
1126 if (tmp->rdev
NeilBrownb2d444d2005-11-08 21:39:31 -08001127 && !test_bit(Faulty, &tmp->rdev->flags)
NeilBrownc04be0a2006-10-03 01:15:53 -07001128 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1129 unsigned long flags;
1130 spin_lock_irqsave(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 mddev->degraded--;
NeilBrownc04be0a2006-10-03 01:15:53 -07001132 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133 }
1134 }
1135
1136 print_conf(conf);
1137 return 0;
1138}
1139
1140
1141static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1142{
1143 conf_t *conf = mddev->private;
Neil Brown199050e2008-06-28 08:31:33 +10001144 int err = -EEXIST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001145 int mirror;
1146 mirror_info_t *p;
Neil Brown6c2fce22008-06-28 08:31:31 +10001147 int first = 0;
NeilBrown84707f32010-03-16 17:23:35 +11001148 int last = conf->raid_disks - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149
1150 if (mddev->recovery_cp < MaxSector)
1151 /* only hot-add to in-sync arrays, as recovery is
1152 * very different from resync
1153 */
Neil Brown199050e2008-06-28 08:31:33 +10001154 return -EBUSY;
NeilBrown6d508242005-09-09 16:24:03 -07001155 if (!enough(conf))
Neil Brown199050e2008-06-28 08:31:33 +10001156 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157
NeilBrowna53a6c82008-11-06 17:28:20 +11001158 if (rdev->raid_disk >= 0)
Neil Brown6c2fce22008-06-28 08:31:31 +10001159 first = last = rdev->raid_disk;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160
NeilBrown6cce3b22006-01-06 00:20:16 -08001161 if (rdev->saved_raid_disk >= 0 &&
Neil Brown6c2fce22008-06-28 08:31:31 +10001162 rdev->saved_raid_disk >= first &&
NeilBrown6cce3b22006-01-06 00:20:16 -08001163 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1164 mirror = rdev->saved_raid_disk;
1165 else
Neil Brown6c2fce22008-06-28 08:31:31 +10001166 mirror = first;
1167 for ( ; mirror <= last ; mirror++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 if ( !(p=conf->mirrors+mirror)->rdev) {
1169
Martin K. Petersen8f6c2e42009-07-01 11:13:45 +10001170 disk_stack_limits(mddev->gendisk, rdev->bdev,
1171 rdev->data_offset << 9);
NeilBrown627a2d32010-03-08 16:44:38 +11001172 /* as we don't honour merge_bvec_fn, we must
1173 * never risk violating it, so limit
1174 * ->max_segments to one lying with a single
1175 * page, as a one page request is never in
1176 * violation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177 */
NeilBrown627a2d32010-03-08 16:44:38 +11001178 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1179 blk_queue_max_segments(mddev->queue, 1);
1180 blk_queue_segment_boundary(mddev->queue,
1181 PAGE_CACHE_SIZE - 1);
1182 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183
1184 p->head_position = 0;
1185 rdev->raid_disk = mirror;
Neil Brown199050e2008-06-28 08:31:33 +10001186 err = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08001187 if (rdev->saved_raid_disk != mirror)
1188 conf->fullsync = 1;
Suzanne Woodd6065f72005-11-08 21:39:27 -08001189 rcu_assign_pointer(p->rdev, rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001190 break;
1191 }
1192
Andre Nollac5e7112009-08-03 10:59:47 +10001193 md_integrity_add_rdev(rdev, mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 print_conf(conf);
Neil Brown199050e2008-06-28 08:31:33 +10001195 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196}
1197
1198static int raid10_remove_disk(mddev_t *mddev, int number)
1199{
1200 conf_t *conf = mddev->private;
1201 int err = 0;
1202 mdk_rdev_t *rdev;
1203 mirror_info_t *p = conf->mirrors+ number;
1204
1205 print_conf(conf);
1206 rdev = p->rdev;
1207 if (rdev) {
NeilBrownb2d444d2005-11-08 21:39:31 -08001208 if (test_bit(In_sync, &rdev->flags) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209 atomic_read(&rdev->nr_pending)) {
1210 err = -EBUSY;
1211 goto abort;
1212 }
NeilBrowndfc70642008-05-23 13:04:39 -07001213 /* Only remove faulty devices in recovery
1214 * is not possible.
1215 */
1216 if (!test_bit(Faulty, &rdev->flags) &&
1217 enough(conf)) {
1218 err = -EBUSY;
1219 goto abort;
1220 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07001222 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 if (atomic_read(&rdev->nr_pending)) {
1224 /* lost the race, try later */
1225 err = -EBUSY;
1226 p->rdev = rdev;
Andre Nollac5e7112009-08-03 10:59:47 +10001227 goto abort;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228 }
Andre Nollac5e7112009-08-03 10:59:47 +10001229 md_integrity_register(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 }
1231abort:
1232
1233 print_conf(conf);
1234 return err;
1235}
1236
1237
NeilBrown6712ecf2007-09-27 12:47:43 +02001238static void end_sync_read(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239{
H Hartley Sweeten7b928132010-03-08 16:02:40 +11001240 r10bio_t *r10_bio = bio->bi_private;
NeilBrown070ec552009-06-16 16:54:21 +10001241 conf_t *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242 int i,d;
1243
Linus Torvalds1da177e2005-04-16 15:20:36 -07001244 for (i=0; i<conf->copies; i++)
1245 if (r10_bio->devs[i].bio == bio)
1246 break;
Eric Sesterhennb6385482006-04-02 13:34:29 +02001247 BUG_ON(i == conf->copies);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 update_head_pos(i, r10_bio);
1249 d = r10_bio->devs[i].devnum;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001250
1251 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1252 set_bit(R10BIO_Uptodate, &r10_bio->state);
NeilBrown4dbcdc72006-01-06 00:20:52 -08001253 else {
1254 atomic_add(r10_bio->sectors,
1255 &conf->mirrors[d].rdev->corrected_errors);
1256 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1257 md_error(r10_bio->mddev,
1258 conf->mirrors[d].rdev);
1259 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260
1261 /* for reconstruct, we always reschedule after a read.
1262 * for resync, only after all reads
1263 */
NeilBrown73d5c382009-02-25 13:18:47 +11001264 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1266 atomic_dec_and_test(&r10_bio->remaining)) {
1267 /* we have read all the blocks,
1268 * do the comparison in process context in raid10d
1269 */
1270 reschedule_retry(r10_bio);
1271 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272}
1273
NeilBrown6712ecf2007-09-27 12:47:43 +02001274static void end_sync_write(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275{
1276 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
H Hartley Sweeten7b928132010-03-08 16:02:40 +11001277 r10bio_t *r10_bio = bio->bi_private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 mddev_t *mddev = r10_bio->mddev;
NeilBrown070ec552009-06-16 16:54:21 +10001279 conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 int i,d;
1281
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 for (i = 0; i < conf->copies; i++)
1283 if (r10_bio->devs[i].bio == bio)
1284 break;
1285 d = r10_bio->devs[i].devnum;
1286
1287 if (!uptodate)
1288 md_error(mddev, conf->mirrors[d].rdev);
NeilBrowndfc70642008-05-23 13:04:39 -07001289
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290 update_head_pos(i, r10_bio);
1291
NeilBrown73d5c382009-02-25 13:18:47 +11001292 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 while (atomic_dec_and_test(&r10_bio->remaining)) {
1294 if (r10_bio->master_bio == NULL) {
1295 /* the primary of several recovery bios */
NeilBrown73d5c382009-02-25 13:18:47 +11001296 sector_t s = r10_bio->sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297 put_buf(r10_bio);
NeilBrown73d5c382009-02-25 13:18:47 +11001298 md_done_sync(mddev, s, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299 break;
1300 } else {
1301 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1302 put_buf(r10_bio);
1303 r10_bio = r10_bio2;
1304 }
1305 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306}
1307
1308/*
1309 * Note: sync and recover and handled very differently for raid10
1310 * This code is for resync.
1311 * For resync, we read through virtual addresses and read all blocks.
1312 * If there is any error, we schedule a write. The lowest numbered
1313 * drive is authoritative.
1314 * However requests come for physical address, so we need to map.
1315 * For every physical address there are raid_disks/copies virtual addresses,
1316 * which is always are least one, but is not necessarly an integer.
1317 * This means that a physical address can span multiple chunks, so we may
1318 * have to submit multiple io requests for a single sync request.
1319 */
1320/*
1321 * We check if all blocks are in-sync and only write to blocks that
1322 * aren't in sync
1323 */
1324static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1325{
NeilBrown070ec552009-06-16 16:54:21 +10001326 conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327 int i, first;
1328 struct bio *tbio, *fbio;
1329
1330 atomic_set(&r10_bio->remaining, 1);
1331
1332 /* find the first device with a block */
1333 for (i=0; i<conf->copies; i++)
1334 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1335 break;
1336
1337 if (i == conf->copies)
1338 goto done;
1339
1340 first = i;
1341 fbio = r10_bio->devs[i].bio;
1342
1343 /* now find blocks with errors */
NeilBrown0eb3ff12006-01-06 00:20:29 -08001344 for (i=0 ; i < conf->copies ; i++) {
1345 int j, d;
1346 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 tbio = r10_bio->devs[i].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001349
1350 if (tbio->bi_end_io != end_sync_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 continue;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001352 if (i == first)
1353 continue;
1354 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1355 /* We know that the bi_io_vec layout is the same for
1356 * both 'first' and 'i', so we just compare them.
1357 * All vec entries are PAGE_SIZE;
1358 */
1359 for (j = 0; j < vcnt; j++)
1360 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1361 page_address(tbio->bi_io_vec[j].bv_page),
1362 PAGE_SIZE))
1363 break;
1364 if (j == vcnt)
1365 continue;
1366 mddev->resync_mismatches += r10_bio->sectors;
1367 }
NeilBrown18f08812006-01-06 00:20:25 -08001368 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1369 /* Don't fix anything. */
1370 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 /* Ok, we need to write this bio
1372 * First we need to fixup bv_offset, bv_len and
1373 * bi_vecs, as the read request might have corrupted these
1374 */
1375 tbio->bi_vcnt = vcnt;
1376 tbio->bi_size = r10_bio->sectors << 9;
1377 tbio->bi_idx = 0;
1378 tbio->bi_phys_segments = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1380 tbio->bi_flags |= 1 << BIO_UPTODATE;
1381 tbio->bi_next = NULL;
1382 tbio->bi_rw = WRITE;
1383 tbio->bi_private = r10_bio;
1384 tbio->bi_sector = r10_bio->devs[i].addr;
1385
1386 for (j=0; j < vcnt ; j++) {
1387 tbio->bi_io_vec[j].bv_offset = 0;
1388 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1389
1390 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1391 page_address(fbio->bi_io_vec[j].bv_page),
1392 PAGE_SIZE);
1393 }
1394 tbio->bi_end_io = end_sync_write;
1395
1396 d = r10_bio->devs[i].devnum;
1397 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1398 atomic_inc(&r10_bio->remaining);
1399 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1400
1401 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1402 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1403 generic_make_request(tbio);
1404 }
1405
1406done:
1407 if (atomic_dec_and_test(&r10_bio->remaining)) {
1408 md_done_sync(mddev, r10_bio->sectors, 1);
1409 put_buf(r10_bio);
1410 }
1411}
1412
1413/*
1414 * Now for the recovery code.
1415 * Recovery happens across physical sectors.
1416 * We recover all non-is_sync drives by finding the virtual address of
1417 * each, and then choose a working drive that also has that virt address.
1418 * There is a separate r10_bio for each non-in_sync drive.
1419 * Only the first two slots are in use. The first for reading,
1420 * The second for writing.
1421 *
1422 */
1423
1424static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1425{
NeilBrown070ec552009-06-16 16:54:21 +10001426 conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 int i, d;
1428 struct bio *bio, *wbio;
1429
1430
1431 /* move the pages across to the second bio
1432 * and submit the write request
1433 */
1434 bio = r10_bio->devs[0].bio;
1435 wbio = r10_bio->devs[1].bio;
1436 for (i=0; i < wbio->bi_vcnt; i++) {
1437 struct page *p = bio->bi_io_vec[i].bv_page;
1438 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1439 wbio->bi_io_vec[i].bv_page = p;
1440 }
1441 d = r10_bio->devs[1].devnum;
1442
1443 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1444 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
NeilBrown0eb3ff12006-01-06 00:20:29 -08001445 if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1446 generic_make_request(wbio);
1447 else
NeilBrown6712ecf2007-09-27 12:47:43 +02001448 bio_endio(wbio, -EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449}
1450
1451
1452/*
Robert Becker1e509152009-12-14 12:49:58 +11001453 * Used by fix_read_error() to decay the per rdev read_errors.
1454 * We halve the read error count for every hour that has elapsed
1455 * since the last recorded read error.
1456 *
1457 */
1458static void check_decay_read_errors(mddev_t *mddev, mdk_rdev_t *rdev)
1459{
1460 struct timespec cur_time_mon;
1461 unsigned long hours_since_last;
1462 unsigned int read_errors = atomic_read(&rdev->read_errors);
1463
1464 ktime_get_ts(&cur_time_mon);
1465
1466 if (rdev->last_read_error.tv_sec == 0 &&
1467 rdev->last_read_error.tv_nsec == 0) {
1468 /* first time we've seen a read error */
1469 rdev->last_read_error = cur_time_mon;
1470 return;
1471 }
1472
1473 hours_since_last = (cur_time_mon.tv_sec -
1474 rdev->last_read_error.tv_sec) / 3600;
1475
1476 rdev->last_read_error = cur_time_mon;
1477
1478 /*
1479 * if hours_since_last is > the number of bits in read_errors
1480 * just set read errors to 0. We do this to avoid
1481 * overflowing the shift of read_errors by hours_since_last.
1482 */
1483 if (hours_since_last >= 8 * sizeof(read_errors))
1484 atomic_set(&rdev->read_errors, 0);
1485 else
1486 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
1487}
1488
1489/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 * This is a kernel thread which:
1491 *
1492 * 1. Retries failed read operations on working mirrors.
1493 * 2. Updates the raid superblock when problems encounter.
NeilBrown6814d532006-10-03 01:15:45 -07001494 * 3. Performs writes following reads for array synchronising.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 */
1496
NeilBrown6814d532006-10-03 01:15:45 -07001497static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1498{
1499 int sect = 0; /* Offset from r10_bio->sector */
1500 int sectors = r10_bio->sectors;
1501 mdk_rdev_t*rdev;
Robert Becker1e509152009-12-14 12:49:58 +11001502 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10001503 int d = r10_bio->devs[r10_bio->read_slot].devnum;
Robert Becker1e509152009-12-14 12:49:58 +11001504
1505 rcu_read_lock();
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10001506 rdev = rcu_dereference(conf->mirrors[d].rdev);
1507 if (rdev) { /* If rdev is not NULL */
Robert Becker1e509152009-12-14 12:49:58 +11001508 char b[BDEVNAME_SIZE];
1509 int cur_read_error_count = 0;
1510
Robert Becker1e509152009-12-14 12:49:58 +11001511 bdevname(rdev->bdev, b);
1512
1513 if (test_bit(Faulty, &rdev->flags)) {
1514 rcu_read_unlock();
1515 /* drive has already been failed, just ignore any
1516 more fix_read_error() attempts */
1517 return;
1518 }
1519
1520 check_decay_read_errors(mddev, rdev);
1521 atomic_inc(&rdev->read_errors);
1522 cur_read_error_count = atomic_read(&rdev->read_errors);
1523 if (cur_read_error_count > max_read_errors) {
1524 rcu_read_unlock();
1525 printk(KERN_NOTICE
NeilBrown128595e2010-05-03 14:47:14 +10001526 "md/raid10:%s: %s: Raid device exceeded "
Robert Becker1e509152009-12-14 12:49:58 +11001527 "read_error threshold "
1528 "[cur %d:max %d]\n",
NeilBrown128595e2010-05-03 14:47:14 +10001529 mdname(mddev),
Robert Becker1e509152009-12-14 12:49:58 +11001530 b, cur_read_error_count, max_read_errors);
1531 printk(KERN_NOTICE
NeilBrown128595e2010-05-03 14:47:14 +10001532 "md/raid10:%s: %s: Failing raid "
1533 "device\n", mdname(mddev), b);
Robert Becker1e509152009-12-14 12:49:58 +11001534 md_error(mddev, conf->mirrors[d].rdev);
1535 return;
1536 }
1537 }
1538 rcu_read_unlock();
1539
NeilBrown6814d532006-10-03 01:15:45 -07001540 while(sectors) {
1541 int s = sectors;
1542 int sl = r10_bio->read_slot;
1543 int success = 0;
1544 int start;
1545
1546 if (s > (PAGE_SIZE>>9))
1547 s = PAGE_SIZE >> 9;
1548
1549 rcu_read_lock();
1550 do {
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10001551 d = r10_bio->devs[sl].devnum;
NeilBrown6814d532006-10-03 01:15:45 -07001552 rdev = rcu_dereference(conf->mirrors[d].rdev);
1553 if (rdev &&
1554 test_bit(In_sync, &rdev->flags)) {
1555 atomic_inc(&rdev->nr_pending);
1556 rcu_read_unlock();
1557 success = sync_page_io(rdev->bdev,
1558 r10_bio->devs[sl].addr +
1559 sect + rdev->data_offset,
1560 s<<9,
1561 conf->tmppage, READ);
1562 rdev_dec_pending(rdev, mddev);
1563 rcu_read_lock();
1564 if (success)
1565 break;
1566 }
1567 sl++;
1568 if (sl == conf->copies)
1569 sl = 0;
1570 } while (!success && sl != r10_bio->read_slot);
1571 rcu_read_unlock();
1572
1573 if (!success) {
1574 /* Cannot read from anywhere -- bye bye array */
1575 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1576 md_error(mddev, conf->mirrors[dn].rdev);
1577 break;
1578 }
1579
1580 start = sl;
1581 /* write it back and re-read */
1582 rcu_read_lock();
1583 while (sl != r10_bio->read_slot) {
Robert Becker67b8dc42009-12-14 12:49:57 +11001584 char b[BDEVNAME_SIZE];
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10001585
NeilBrown6814d532006-10-03 01:15:45 -07001586 if (sl==0)
1587 sl = conf->copies;
1588 sl--;
1589 d = r10_bio->devs[sl].devnum;
1590 rdev = rcu_dereference(conf->mirrors[d].rdev);
1591 if (rdev &&
1592 test_bit(In_sync, &rdev->flags)) {
1593 atomic_inc(&rdev->nr_pending);
1594 rcu_read_unlock();
1595 atomic_add(s, &rdev->corrected_errors);
1596 if (sync_page_io(rdev->bdev,
1597 r10_bio->devs[sl].addr +
1598 sect + rdev->data_offset,
1599 s<<9, conf->tmppage, WRITE)
Robert Becker67b8dc42009-12-14 12:49:57 +11001600 == 0) {
NeilBrown6814d532006-10-03 01:15:45 -07001601 /* Well, this device is dead */
Robert Becker67b8dc42009-12-14 12:49:57 +11001602 printk(KERN_NOTICE
NeilBrown128595e2010-05-03 14:47:14 +10001603 "md/raid10:%s: read correction "
Robert Becker67b8dc42009-12-14 12:49:57 +11001604 "write failed"
1605 " (%d sectors at %llu on %s)\n",
1606 mdname(mddev), s,
1607 (unsigned long long)(sect+
1608 rdev->data_offset),
1609 bdevname(rdev->bdev, b));
NeilBrown128595e2010-05-03 14:47:14 +10001610 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
Robert Becker67b8dc42009-12-14 12:49:57 +11001611 "drive\n",
NeilBrown128595e2010-05-03 14:47:14 +10001612 mdname(mddev),
Robert Becker67b8dc42009-12-14 12:49:57 +11001613 bdevname(rdev->bdev, b));
NeilBrown6814d532006-10-03 01:15:45 -07001614 md_error(mddev, rdev);
Robert Becker67b8dc42009-12-14 12:49:57 +11001615 }
NeilBrown6814d532006-10-03 01:15:45 -07001616 rdev_dec_pending(rdev, mddev);
1617 rcu_read_lock();
1618 }
1619 }
1620 sl = start;
1621 while (sl != r10_bio->read_slot) {
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10001622
NeilBrown6814d532006-10-03 01:15:45 -07001623 if (sl==0)
1624 sl = conf->copies;
1625 sl--;
1626 d = r10_bio->devs[sl].devnum;
1627 rdev = rcu_dereference(conf->mirrors[d].rdev);
1628 if (rdev &&
1629 test_bit(In_sync, &rdev->flags)) {
1630 char b[BDEVNAME_SIZE];
1631 atomic_inc(&rdev->nr_pending);
1632 rcu_read_unlock();
1633 if (sync_page_io(rdev->bdev,
1634 r10_bio->devs[sl].addr +
1635 sect + rdev->data_offset,
Robert Becker67b8dc42009-12-14 12:49:57 +11001636 s<<9, conf->tmppage,
1637 READ) == 0) {
NeilBrown6814d532006-10-03 01:15:45 -07001638 /* Well, this device is dead */
Robert Becker67b8dc42009-12-14 12:49:57 +11001639 printk(KERN_NOTICE
NeilBrown128595e2010-05-03 14:47:14 +10001640 "md/raid10:%s: unable to read back "
Robert Becker67b8dc42009-12-14 12:49:57 +11001641 "corrected sectors"
1642 " (%d sectors at %llu on %s)\n",
1643 mdname(mddev), s,
1644 (unsigned long long)(sect+
1645 rdev->data_offset),
1646 bdevname(rdev->bdev, b));
NeilBrown128595e2010-05-03 14:47:14 +10001647 printk(KERN_NOTICE "md/raid10:%s: %s: failing drive\n",
1648 mdname(mddev),
Robert Becker67b8dc42009-12-14 12:49:57 +11001649 bdevname(rdev->bdev, b));
1650
NeilBrown6814d532006-10-03 01:15:45 -07001651 md_error(mddev, rdev);
Robert Becker67b8dc42009-12-14 12:49:57 +11001652 } else {
NeilBrown6814d532006-10-03 01:15:45 -07001653 printk(KERN_INFO
NeilBrown128595e2010-05-03 14:47:14 +10001654 "md/raid10:%s: read error corrected"
NeilBrown6814d532006-10-03 01:15:45 -07001655 " (%d sectors at %llu on %s)\n",
1656 mdname(mddev), s,
Randy Dunlap969b7552006-10-28 10:38:32 -07001657 (unsigned long long)(sect+
1658 rdev->data_offset),
NeilBrown6814d532006-10-03 01:15:45 -07001659 bdevname(rdev->bdev, b));
Robert Becker67b8dc42009-12-14 12:49:57 +11001660 }
NeilBrown6814d532006-10-03 01:15:45 -07001661
1662 rdev_dec_pending(rdev, mddev);
1663 rcu_read_lock();
1664 }
1665 }
1666 rcu_read_unlock();
1667
1668 sectors -= s;
1669 sect += s;
1670 }
1671}
1672
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673static void raid10d(mddev_t *mddev)
1674{
1675 r10bio_t *r10_bio;
1676 struct bio *bio;
1677 unsigned long flags;
NeilBrown070ec552009-06-16 16:54:21 +10001678 conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679 struct list_head *head = &conf->retry_list;
1680 int unplug=0;
1681 mdk_rdev_t *rdev;
1682
1683 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684
1685 for (;;) {
1686 char b[BDEVNAME_SIZE];
NeilBrowna35e63e2008-03-04 14:29:29 -08001687
1688 unplug += flush_pending_writes(conf);
1689
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrowna35e63e2008-03-04 14:29:29 -08001691 if (list_empty(head)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001692 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693 break;
NeilBrowna35e63e2008-03-04 14:29:29 -08001694 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1696 list_del(head->prev);
NeilBrown4443ae12006-01-06 00:20:28 -08001697 conf->nr_queued--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698 spin_unlock_irqrestore(&conf->device_lock, flags);
1699
1700 mddev = r10_bio->mddev;
NeilBrown070ec552009-06-16 16:54:21 +10001701 conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1703 sync_request_write(mddev, r10_bio);
1704 unplug = 1;
1705 } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1706 recovery_request_write(mddev, r10_bio);
1707 unplug = 1;
1708 } else {
1709 int mirror;
NeilBrown4443ae12006-01-06 00:20:28 -08001710 /* we got a read error. Maybe the drive is bad. Maybe just
1711 * the block and we can fix it.
1712 * We freeze all other IO, and try reading the block from
1713 * other devices. When we find one, we re-write
1714 * and check it that fixes the read error.
1715 * This is all done synchronously while the array is
1716 * frozen.
1717 */
NeilBrown6814d532006-10-03 01:15:45 -07001718 if (mddev->ro == 0) {
1719 freeze_array(conf);
1720 fix_read_error(conf, mddev, r10_bio);
1721 unfreeze_array(conf);
NeilBrown4443ae12006-01-06 00:20:28 -08001722 }
1723
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 bio = r10_bio->devs[r10_bio->read_slot].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001725 r10_bio->devs[r10_bio->read_slot].bio =
1726 mddev->ro ? IO_BLOCKED : NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 mirror = read_balance(conf, r10_bio);
1728 if (mirror == -1) {
NeilBrown128595e2010-05-03 14:47:14 +10001729 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730 " read error for block %llu\n",
NeilBrown128595e2010-05-03 14:47:14 +10001731 mdname(mddev),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 bdevname(bio->bi_bdev,b),
1733 (unsigned long long)r10_bio->sector);
1734 raid_end_bio_io(r10_bio);
Maik Hampel14e71342007-07-31 00:37:57 -07001735 bio_put(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736 } else {
Jens Axboe1f98a132009-09-11 14:32:04 +02001737 const bool do_sync = bio_rw_flagged(r10_bio->master_bio, BIO_RW_SYNCIO);
Maik Hampel14e71342007-07-31 00:37:57 -07001738 bio_put(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739 rdev = conf->mirrors[mirror].rdev;
1740 if (printk_ratelimit())
NeilBrown128595e2010-05-03 14:47:14 +10001741 printk(KERN_ERR "md/raid10:%s: %s: redirecting sector %llu to"
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742 " another mirror\n",
NeilBrown128595e2010-05-03 14:47:14 +10001743 mdname(mddev),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 bdevname(rdev->bdev,b),
1745 (unsigned long long)r10_bio->sector);
1746 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1747 r10_bio->devs[r10_bio->read_slot].bio = bio;
1748 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1749 + rdev->data_offset;
1750 bio->bi_bdev = rdev->bdev;
Dmitry Monakhov1ef04fe2009-09-20 05:52:25 +04001751 bio->bi_rw = READ | (do_sync << BIO_RW_SYNCIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752 bio->bi_private = r10_bio;
1753 bio->bi_end_io = raid10_end_read_request;
1754 unplug = 1;
1755 generic_make_request(bio);
1756 }
1757 }
NeilBrown1d9d5242009-10-16 15:55:32 +11001758 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760 if (unplug)
1761 unplug_slaves(mddev);
1762}
1763
1764
1765static int init_resync(conf_t *conf)
1766{
1767 int buffs;
1768
1769 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
Eric Sesterhennb6385482006-04-02 13:34:29 +02001770 BUG_ON(conf->r10buf_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1772 if (!conf->r10buf_pool)
1773 return -ENOMEM;
1774 conf->next_resync = 0;
1775 return 0;
1776}
1777
1778/*
1779 * perform a "sync" on one "block"
1780 *
1781 * We need to make sure that no normal I/O request - particularly write
1782 * requests - conflict with active sync requests.
1783 *
1784 * This is achieved by tracking pending requests and a 'barrier' concept
1785 * that can be installed to exclude normal IO requests.
1786 *
1787 * Resync and recovery are handled very differently.
1788 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1789 *
1790 * For resync, we iterate over virtual addresses, read all copies,
1791 * and update if there are differences. If only one copy is live,
1792 * skip it.
1793 * For recovery, we iterate over physical addresses, read a good
1794 * value for each non-in_sync drive, and over-write.
1795 *
1796 * So, for recovery we may have several outstanding complex requests for a
1797 * given address, one for each out-of-sync device. We model this by allocating
1798 * a number of r10_bio structures, one for each out-of-sync device.
1799 * As we setup these structures, we collect all bio's together into a list
1800 * which we then process collectively to add pages, and then process again
1801 * to pass to generic_make_request.
1802 *
1803 * The r10_bio structures are linked using a borrowed master_bio pointer.
1804 * This link is counted in ->remaining. When the r10_bio that points to NULL
1805 * has its remaining count decremented to 0, the whole complex operation
1806 * is complete.
1807 *
1808 */
1809
NeilBrown57afd892005-06-21 17:17:13 -07001810static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811{
NeilBrown070ec552009-06-16 16:54:21 +10001812 conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 r10bio_t *r10_bio;
1814 struct bio *biolist = NULL, *bio;
1815 sector_t max_sector, nr_sectors;
1816 int disk;
1817 int i;
NeilBrown6cce3b22006-01-06 00:20:16 -08001818 int max_sync;
1819 int sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820
1821 sector_t sectors_skipped = 0;
1822 int chunks_skipped = 0;
1823
1824 if (!conf->r10buf_pool)
1825 if (init_resync(conf))
NeilBrown57afd892005-06-21 17:17:13 -07001826 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827
1828 skipped:
Andre Noll58c0fed2009-03-31 14:33:13 +11001829 max_sector = mddev->dev_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1831 max_sector = mddev->resync_max_sectors;
1832 if (sector_nr >= max_sector) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001833 /* If we aborted, we need to abort the
1834 * sync on the 'current' bitmap chucks (there can
1835 * be several when recovering multiple devices).
1836 * as we may have started syncing it but not finished.
1837 * We can find the current address in
1838 * mddev->curr_resync, but for recovery,
1839 * we need to convert that to several
1840 * virtual addresses.
1841 */
1842 if (mddev->curr_resync < max_sector) { /* aborted */
1843 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1844 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1845 &sync_blocks, 1);
1846 else for (i=0; i<conf->raid_disks; i++) {
1847 sector_t sect =
1848 raid10_find_virt(conf, mddev->curr_resync, i);
1849 bitmap_end_sync(mddev->bitmap, sect,
1850 &sync_blocks, 1);
1851 }
1852 } else /* completed sync */
1853 conf->fullsync = 0;
1854
1855 bitmap_close_sync(mddev->bitmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 close_sync(conf);
NeilBrown57afd892005-06-21 17:17:13 -07001857 *skipped = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858 return sectors_skipped;
1859 }
1860 if (chunks_skipped >= conf->raid_disks) {
1861 /* if there has been nothing to do on any drive,
1862 * then there is nothing to do at all..
1863 */
NeilBrown57afd892005-06-21 17:17:13 -07001864 *skipped = 1;
1865 return (max_sector - sector_nr) + sectors_skipped;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 }
1867
NeilBrownc6207272008-02-06 01:39:52 -08001868 if (max_sector > mddev->resync_max)
1869 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1870
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871 /* make sure whole request will fit in a chunk - if chunks
1872 * are meaningful
1873 */
1874 if (conf->near_copies < conf->raid_disks &&
1875 max_sector > (sector_nr | conf->chunk_mask))
1876 max_sector = (sector_nr | conf->chunk_mask) + 1;
1877 /*
1878 * If there is non-resync activity waiting for us then
1879 * put in a delay to throttle resync.
1880 */
NeilBrown0a27ec92006-01-06 00:20:13 -08001881 if (!go_faster && conf->nr_waiting)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 msleep_interruptible(1000);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883
1884 /* Again, very different code for resync and recovery.
1885 * Both must result in an r10bio with a list of bios that
1886 * have bi_end_io, bi_sector, bi_bdev set,
1887 * and bi_private set to the r10bio.
1888 * For recovery, we may actually create several r10bios
1889 * with 2 bios in each, that correspond to the bios in the main one.
1890 * In this case, the subordinate r10bios link back through a
1891 * borrowed master_bio pointer, and the counter in the master
1892 * includes a ref from each subordinate.
1893 */
1894 /* First, we decide what to do and set ->bi_end_io
1895 * To end_sync_read if we want to read, and
1896 * end_sync_write if we will want to write.
1897 */
1898
NeilBrown6cce3b22006-01-06 00:20:16 -08001899 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1901 /* recovery... the complicated one */
NeilBrowna9f326e2009-09-23 18:06:41 +10001902 int j, k;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903 r10_bio = NULL;
1904
1905 for (i=0 ; i<conf->raid_disks; i++)
1906 if (conf->mirrors[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001907 !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001908 int still_degraded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909 /* want to reconstruct this device */
1910 r10bio_t *rb2 = r10_bio;
NeilBrown6cce3b22006-01-06 00:20:16 -08001911 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1912 int must_sync;
1913 /* Unless we are doing a full sync, we only need
1914 * to recover the block if it is set in the bitmap
1915 */
1916 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1917 &sync_blocks, 1);
1918 if (sync_blocks < max_sync)
1919 max_sync = sync_blocks;
1920 if (!must_sync &&
1921 !conf->fullsync) {
1922 /* yep, skip the sync_blocks here, but don't assume
1923 * that there will never be anything to do here
1924 */
1925 chunks_skipped = -1;
1926 continue;
1927 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928
1929 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
NeilBrown6cce3b22006-01-06 00:20:16 -08001930 raise_barrier(conf, rb2 != NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931 atomic_set(&r10_bio->remaining, 0);
1932
1933 r10_bio->master_bio = (struct bio*)rb2;
1934 if (rb2)
1935 atomic_inc(&rb2->remaining);
1936 r10_bio->mddev = mddev;
1937 set_bit(R10BIO_IsRecover, &r10_bio->state);
NeilBrown6cce3b22006-01-06 00:20:16 -08001938 r10_bio->sector = sect;
1939
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 raid10_find_phys(conf, r10_bio);
NeilBrown18055562009-05-07 12:48:10 +10001941
1942 /* Need to check if the array will still be
NeilBrown6cce3b22006-01-06 00:20:16 -08001943 * degraded
1944 */
NeilBrown18055562009-05-07 12:48:10 +10001945 for (j=0; j<conf->raid_disks; j++)
1946 if (conf->mirrors[j].rdev == NULL ||
1947 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001948 still_degraded = 1;
NeilBrowna24a8dd2006-01-06 00:20:35 -08001949 break;
1950 }
NeilBrown18055562009-05-07 12:48:10 +10001951
NeilBrown6cce3b22006-01-06 00:20:16 -08001952 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1953 &sync_blocks, still_degraded);
1954
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955 for (j=0; j<conf->copies;j++) {
1956 int d = r10_bio->devs[j].devnum;
1957 if (conf->mirrors[d].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001958 test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 /* This is where we read from */
1960 bio = r10_bio->devs[0].bio;
1961 bio->bi_next = biolist;
1962 biolist = bio;
1963 bio->bi_private = r10_bio;
1964 bio->bi_end_io = end_sync_read;
NeilBrown802ba062006-12-13 00:34:13 -08001965 bio->bi_rw = READ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966 bio->bi_sector = r10_bio->devs[j].addr +
1967 conf->mirrors[d].rdev->data_offset;
1968 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1969 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1970 atomic_inc(&r10_bio->remaining);
1971 /* and we write to 'i' */
1972
1973 for (k=0; k<conf->copies; k++)
1974 if (r10_bio->devs[k].devnum == i)
1975 break;
NeilBrown64a742b2007-02-28 20:11:18 -08001976 BUG_ON(k == conf->copies);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 bio = r10_bio->devs[1].bio;
1978 bio->bi_next = biolist;
1979 biolist = bio;
1980 bio->bi_private = r10_bio;
1981 bio->bi_end_io = end_sync_write;
NeilBrown802ba062006-12-13 00:34:13 -08001982 bio->bi_rw = WRITE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983 bio->bi_sector = r10_bio->devs[k].addr +
1984 conf->mirrors[i].rdev->data_offset;
1985 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1986
1987 r10_bio->devs[0].devnum = d;
1988 r10_bio->devs[1].devnum = i;
1989
1990 break;
1991 }
1992 }
1993 if (j == conf->copies) {
NeilBrown87fc7672005-09-09 16:24:04 -07001994 /* Cannot recover, so abort the recovery */
1995 put_buf(r10_bio);
K.Tanakaa07e6ab2008-03-04 14:29:37 -08001996 if (rb2)
1997 atomic_dec(&rb2->remaining);
NeilBrown87fc7672005-09-09 16:24:04 -07001998 r10_bio = rb2;
NeilBrowndfc70642008-05-23 13:04:39 -07001999 if (!test_and_set_bit(MD_RECOVERY_INTR,
2000 &mddev->recovery))
NeilBrown128595e2010-05-03 14:47:14 +10002001 printk(KERN_INFO "md/raid10:%s: insufficient "
2002 "working devices for recovery.\n",
NeilBrown87fc7672005-09-09 16:24:04 -07002003 mdname(mddev));
2004 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005 }
2006 }
2007 if (biolist == NULL) {
2008 while (r10_bio) {
2009 r10bio_t *rb2 = r10_bio;
2010 r10_bio = (r10bio_t*) rb2->master_bio;
2011 rb2->master_bio = NULL;
2012 put_buf(rb2);
2013 }
2014 goto giveup;
2015 }
2016 } else {
2017 /* resync. Schedule a read for every block at this virt offset */
2018 int count = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08002019
NeilBrown78200d42009-02-25 13:18:47 +11002020 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2021
NeilBrown6cce3b22006-01-06 00:20:16 -08002022 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2023 &sync_blocks, mddev->degraded) &&
2024 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2025 /* We can skip this block */
2026 *skipped = 1;
2027 return sync_blocks + sectors_skipped;
2028 }
2029 if (sync_blocks < max_sync)
2030 max_sync = sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2032
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 r10_bio->mddev = mddev;
2034 atomic_set(&r10_bio->remaining, 0);
NeilBrown6cce3b22006-01-06 00:20:16 -08002035 raise_barrier(conf, 0);
2036 conf->next_resync = sector_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037
2038 r10_bio->master_bio = NULL;
2039 r10_bio->sector = sector_nr;
2040 set_bit(R10BIO_IsSync, &r10_bio->state);
2041 raid10_find_phys(conf, r10_bio);
2042 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
2043
2044 for (i=0; i<conf->copies; i++) {
2045 int d = r10_bio->devs[i].devnum;
2046 bio = r10_bio->devs[i].bio;
2047 bio->bi_end_io = NULL;
NeilBrownaf03b8e2007-06-16 10:16:06 -07002048 clear_bit(BIO_UPTODATE, &bio->bi_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049 if (conf->mirrors[d].rdev == NULL ||
NeilBrownb2d444d2005-11-08 21:39:31 -08002050 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051 continue;
2052 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2053 atomic_inc(&r10_bio->remaining);
2054 bio->bi_next = biolist;
2055 biolist = bio;
2056 bio->bi_private = r10_bio;
2057 bio->bi_end_io = end_sync_read;
NeilBrown802ba062006-12-13 00:34:13 -08002058 bio->bi_rw = READ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059 bio->bi_sector = r10_bio->devs[i].addr +
2060 conf->mirrors[d].rdev->data_offset;
2061 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
2062 count++;
2063 }
2064
2065 if (count < 2) {
2066 for (i=0; i<conf->copies; i++) {
2067 int d = r10_bio->devs[i].devnum;
2068 if (r10_bio->devs[i].bio->bi_end_io)
2069 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
2070 }
2071 put_buf(r10_bio);
2072 biolist = NULL;
2073 goto giveup;
2074 }
2075 }
2076
2077 for (bio = biolist; bio ; bio=bio->bi_next) {
2078
2079 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
2080 if (bio->bi_end_io)
2081 bio->bi_flags |= 1 << BIO_UPTODATE;
2082 bio->bi_vcnt = 0;
2083 bio->bi_idx = 0;
2084 bio->bi_phys_segments = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 bio->bi_size = 0;
2086 }
2087
2088 nr_sectors = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08002089 if (sector_nr + max_sync < max_sector)
2090 max_sector = sector_nr + max_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 do {
2092 struct page *page;
2093 int len = PAGE_SIZE;
2094 disk = 0;
2095 if (sector_nr + (len>>9) > max_sector)
2096 len = (max_sector - sector_nr) << 9;
2097 if (len == 0)
2098 break;
2099 for (bio= biolist ; bio ; bio=bio->bi_next) {
2100 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2101 if (bio_add_page(bio, page, len, 0) == 0) {
2102 /* stop here */
2103 struct bio *bio2;
2104 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2105 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
2106 /* remove last page from this bio */
2107 bio2->bi_vcnt--;
2108 bio2->bi_size -= len;
2109 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
2110 }
2111 goto bio_full;
2112 }
2113 disk = i;
2114 }
2115 nr_sectors += len>>9;
2116 sector_nr += len>>9;
2117 } while (biolist->bi_vcnt < RESYNC_PAGES);
2118 bio_full:
2119 r10_bio->sectors = nr_sectors;
2120
2121 while (biolist) {
2122 bio = biolist;
2123 biolist = biolist->bi_next;
2124
2125 bio->bi_next = NULL;
2126 r10_bio = bio->bi_private;
2127 r10_bio->sectors = nr_sectors;
2128
2129 if (bio->bi_end_io == end_sync_read) {
2130 md_sync_acct(bio->bi_bdev, nr_sectors);
2131 generic_make_request(bio);
2132 }
2133 }
2134
NeilBrown57afd892005-06-21 17:17:13 -07002135 if (sectors_skipped)
2136 /* pretend they weren't skipped, it makes
2137 * no important difference in this case
2138 */
2139 md_done_sync(mddev, sectors_skipped, 1);
2140
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 return sectors_skipped + nr_sectors;
2142 giveup:
2143 /* There is nowhere to write, so all non-sync
2144 * drives must be failed, so try the next chunk...
2145 */
NeilBrown09b40682009-02-25 13:18:47 +11002146 if (sector_nr + max_sync < max_sector)
2147 max_sector = sector_nr + max_sync;
2148
2149 sectors_skipped += (max_sector - sector_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150 chunks_skipped ++;
2151 sector_nr = max_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 goto skipped;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153}
2154
Dan Williams80c3a6c2009-03-17 18:10:40 -07002155static sector_t
2156raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2157{
2158 sector_t size;
NeilBrown070ec552009-06-16 16:54:21 +10002159 conf_t *conf = mddev->private;
Dan Williams80c3a6c2009-03-17 18:10:40 -07002160
2161 if (!raid_disks)
NeilBrown84707f32010-03-16 17:23:35 +11002162 raid_disks = conf->raid_disks;
Dan Williams80c3a6c2009-03-17 18:10:40 -07002163 if (!sectors)
Trela, Maciejdab8b292010-03-08 16:02:45 +11002164 sectors = conf->dev_sectors;
Dan Williams80c3a6c2009-03-17 18:10:40 -07002165
2166 size = sectors >> conf->chunk_shift;
2167 sector_div(size, conf->far_copies);
2168 size = size * raid_disks;
2169 sector_div(size, conf->near_copies);
2170
2171 return size << conf->chunk_shift;
2172}
2173
Trela, Maciejdab8b292010-03-08 16:02:45 +11002174
2175static conf_t *setup_conf(mddev_t *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176{
Trela, Maciejdab8b292010-03-08 16:02:45 +11002177 conf_t *conf = NULL;
NeilBrownc93983b2006-06-26 00:27:41 -07002178 int nc, fc, fo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179 sector_t stride, size;
Trela, Maciejdab8b292010-03-08 16:02:45 +11002180 int err = -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181
Maciej Trelaf73ea872010-06-16 11:46:29 +01002182 if (mddev->new_chunk_sectors < (PAGE_SIZE >> 9) ||
2183 !is_power_of_2(mddev->new_chunk_sectors)) {
NeilBrown128595e2010-05-03 14:47:14 +10002184 printk(KERN_ERR "md/raid10:%s: chunk size must be "
2185 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
2186 mdname(mddev), PAGE_SIZE);
Trela, Maciejdab8b292010-03-08 16:02:45 +11002187 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188 }
NeilBrown2604b702006-01-06 00:20:36 -08002189
Maciej Trelaf73ea872010-06-16 11:46:29 +01002190 nc = mddev->new_layout & 255;
2191 fc = (mddev->new_layout >> 8) & 255;
2192 fo = mddev->new_layout & (1<<16);
Trela, Maciejdab8b292010-03-08 16:02:45 +11002193
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
Maciej Trelaf73ea872010-06-16 11:46:29 +01002195 (mddev->new_layout >> 17)) {
NeilBrown128595e2010-05-03 14:47:14 +10002196 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
Maciej Trelaf73ea872010-06-16 11:46:29 +01002197 mdname(mddev), mddev->new_layout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 goto out;
2199 }
Trela, Maciejdab8b292010-03-08 16:02:45 +11002200
2201 err = -ENOMEM;
NeilBrown4443ae12006-01-06 00:20:28 -08002202 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
Trela, Maciejdab8b292010-03-08 16:02:45 +11002203 if (!conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 goto out;
Trela, Maciejdab8b292010-03-08 16:02:45 +11002205
NeilBrown4443ae12006-01-06 00:20:28 -08002206 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
Trela, Maciejdab8b292010-03-08 16:02:45 +11002207 GFP_KERNEL);
2208 if (!conf->mirrors)
2209 goto out;
NeilBrown4443ae12006-01-06 00:20:28 -08002210
2211 conf->tmppage = alloc_page(GFP_KERNEL);
2212 if (!conf->tmppage)
Trela, Maciejdab8b292010-03-08 16:02:45 +11002213 goto out;
2214
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215
NeilBrown64a742b2007-02-28 20:11:18 -08002216 conf->raid_disks = mddev->raid_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 conf->near_copies = nc;
2218 conf->far_copies = fc;
2219 conf->copies = nc*fc;
NeilBrownc93983b2006-06-26 00:27:41 -07002220 conf->far_offset = fo;
Trela, Maciejdab8b292010-03-08 16:02:45 +11002221 conf->chunk_mask = mddev->new_chunk_sectors - 1;
2222 conf->chunk_shift = ffz(~mddev->new_chunk_sectors);
2223
2224 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2225 r10bio_pool_free, conf);
2226 if (!conf->r10bio_pool)
2227 goto out;
2228
Andre Noll58c0fed2009-03-31 14:33:13 +11002229 size = mddev->dev_sectors >> conf->chunk_shift;
NeilBrown64a742b2007-02-28 20:11:18 -08002230 sector_div(size, fc);
2231 size = size * conf->raid_disks;
2232 sector_div(size, nc);
2233 /* 'size' is now the number of chunks in the array */
2234 /* calculate "used chunks per device" in 'stride' */
2235 stride = size * conf->copies;
NeilBrownaf03b8e2007-06-16 10:16:06 -07002236
2237 /* We need to round up when dividing by raid_disks to
2238 * get the stride size.
2239 */
2240 stride += conf->raid_disks - 1;
NeilBrown64a742b2007-02-28 20:11:18 -08002241 sector_div(stride, conf->raid_disks);
Trela, Maciejdab8b292010-03-08 16:02:45 +11002242
2243 conf->dev_sectors = stride << conf->chunk_shift;
NeilBrown64a742b2007-02-28 20:11:18 -08002244
NeilBrownc93983b2006-06-26 00:27:41 -07002245 if (fo)
NeilBrown64a742b2007-02-28 20:11:18 -08002246 stride = 1;
2247 else
NeilBrownc93983b2006-06-26 00:27:41 -07002248 sector_div(stride, fc);
NeilBrown64a742b2007-02-28 20:11:18 -08002249 conf->stride = stride << conf->chunk_shift;
2250
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251
Neil Browne7e72bf2008-05-14 16:05:54 -07002252 spin_lock_init(&conf->device_lock);
Trela, Maciejdab8b292010-03-08 16:02:45 +11002253 INIT_LIST_HEAD(&conf->retry_list);
2254
2255 spin_lock_init(&conf->resync_lock);
2256 init_waitqueue_head(&conf->wait_barrier);
2257
2258 conf->thread = md_register_thread(raid10d, mddev, NULL);
2259 if (!conf->thread)
2260 goto out;
2261
Trela, Maciejdab8b292010-03-08 16:02:45 +11002262 conf->mddev = mddev;
2263 return conf;
2264
2265 out:
NeilBrown128595e2010-05-03 14:47:14 +10002266 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
Trela, Maciejdab8b292010-03-08 16:02:45 +11002267 mdname(mddev));
2268 if (conf) {
2269 if (conf->r10bio_pool)
2270 mempool_destroy(conf->r10bio_pool);
2271 kfree(conf->mirrors);
2272 safe_put_page(conf->tmppage);
2273 kfree(conf);
2274 }
2275 return ERR_PTR(err);
2276}
2277
2278static int run(mddev_t *mddev)
2279{
2280 conf_t *conf;
2281 int i, disk_idx, chunk_size;
2282 mirror_info_t *disk;
2283 mdk_rdev_t *rdev;
2284 sector_t size;
2285
2286 /*
2287 * copy the already verified devices into our private RAID10
2288 * bookkeeping area. [whatever we allocate in run(),
2289 * should be freed in stop()]
2290 */
2291
2292 if (mddev->private == NULL) {
2293 conf = setup_conf(mddev);
2294 if (IS_ERR(conf))
2295 return PTR_ERR(conf);
2296 mddev->private = conf;
2297 }
2298 conf = mddev->private;
2299 if (!conf)
2300 goto out;
2301
Neil Browne7e72bf2008-05-14 16:05:54 -07002302 mddev->queue->queue_lock = &conf->device_lock;
2303
Trela, Maciejdab8b292010-03-08 16:02:45 +11002304 mddev->thread = conf->thread;
2305 conf->thread = NULL;
2306
Martin K. Petersen8f6c2e42009-07-01 11:13:45 +10002307 chunk_size = mddev->chunk_sectors << 9;
2308 blk_queue_io_min(mddev->queue, chunk_size);
2309 if (conf->raid_disks % conf->near_copies)
2310 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2311 else
2312 blk_queue_io_opt(mddev->queue, chunk_size *
2313 (conf->raid_disks / conf->near_copies));
2314
Cheng Renquan159ec1f2009-01-09 08:31:08 +11002315 list_for_each_entry(rdev, &mddev->disks, same_set) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316 disk_idx = rdev->raid_disk;
NeilBrown84707f32010-03-16 17:23:35 +11002317 if (disk_idx >= conf->raid_disks
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318 || disk_idx < 0)
2319 continue;
2320 disk = conf->mirrors + disk_idx;
2321
2322 disk->rdev = rdev;
Martin K. Petersen8f6c2e42009-07-01 11:13:45 +10002323 disk_stack_limits(mddev->gendisk, rdev->bdev,
2324 rdev->data_offset << 9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325 /* as we don't honour merge_bvec_fn, we must never risk
NeilBrown627a2d32010-03-08 16:44:38 +11002326 * violating it, so limit max_segments to 1 lying
2327 * within a single page.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328 */
NeilBrown627a2d32010-03-08 16:44:38 +11002329 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2330 blk_queue_max_segments(mddev->queue, 1);
2331 blk_queue_segment_boundary(mddev->queue,
2332 PAGE_CACHE_SIZE - 1);
2333 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334
2335 disk->head_position = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 }
NeilBrown6d508242005-09-09 16:24:03 -07002337 /* need to check that every block has at least one working mirror */
2338 if (!enough(conf)) {
NeilBrown128595e2010-05-03 14:47:14 +10002339 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
NeilBrown6d508242005-09-09 16:24:03 -07002340 mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 goto out_free_conf;
2342 }
2343
2344 mddev->degraded = 0;
2345 for (i = 0; i < conf->raid_disks; i++) {
2346
2347 disk = conf->mirrors + i;
2348
NeilBrown5fd6c1d2006-06-26 00:27:40 -07002349 if (!disk->rdev ||
NeilBrown2e333e82006-10-21 10:24:07 -07002350 !test_bit(In_sync, &disk->rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351 disk->head_position = 0;
2352 mddev->degraded++;
Neil Brown8c2e8702008-06-28 08:30:52 +10002353 if (disk->rdev)
2354 conf->fullsync = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355 }
2356 }
2357
Andre Noll8c6ac862009-06-18 08:48:06 +10002358 if (mddev->recovery_cp != MaxSector)
NeilBrown128595e2010-05-03 14:47:14 +10002359 printk(KERN_NOTICE "md/raid10:%s: not clean"
Andre Noll8c6ac862009-06-18 08:48:06 +10002360 " -- starting background reconstruction\n",
2361 mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 printk(KERN_INFO
NeilBrown128595e2010-05-03 14:47:14 +10002363 "md/raid10:%s: active with %d out of %d devices\n",
NeilBrown84707f32010-03-16 17:23:35 +11002364 mdname(mddev), conf->raid_disks - mddev->degraded,
2365 conf->raid_disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 /*
2367 * Ok, everything is just fine now
2368 */
Trela, Maciejdab8b292010-03-08 16:02:45 +11002369 mddev->dev_sectors = conf->dev_sectors;
2370 size = raid10_size(mddev, 0, 0);
2371 md_set_array_sectors(mddev, size);
2372 mddev->resync_max_sectors = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373
NeilBrown7a5febe2005-05-16 21:53:16 -07002374 mddev->queue->unplug_fn = raid10_unplug;
NeilBrown0d129222006-10-03 01:15:54 -07002375 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2376 mddev->queue->backing_dev_info.congested_data = mddev;
NeilBrown7a5febe2005-05-16 21:53:16 -07002377
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378 /* Calculate max read-ahead size.
2379 * We need to readahead at least twice a whole stripe....
2380 * maybe...
2381 */
2382 {
Andre Noll9d8f0362009-06-18 08:45:01 +10002383 int stripe = conf->raid_disks *
2384 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385 stripe /= conf->near_copies;
2386 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2387 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2388 }
2389
NeilBrown84707f32010-03-16 17:23:35 +11002390 if (conf->near_copies < conf->raid_disks)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
Andre Nollac5e7112009-08-03 10:59:47 +10002392 md_integrity_register(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393 return 0;
2394
2395out_free_conf:
2396 if (conf->r10bio_pool)
2397 mempool_destroy(conf->r10bio_pool);
NeilBrown1345b1d2006-01-06 00:20:40 -08002398 safe_put_page(conf->tmppage);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07002399 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400 kfree(conf);
2401 mddev->private = NULL;
Trela, Maciejdab8b292010-03-08 16:02:45 +11002402 md_unregister_thread(mddev->thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403out:
2404 return -EIO;
2405}
2406
2407static int stop(mddev_t *mddev)
2408{
NeilBrown070ec552009-06-16 16:54:21 +10002409 conf_t *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410
NeilBrown409c57f2009-03-31 14:39:39 +11002411 raise_barrier(conf, 0);
2412 lower_barrier(conf);
2413
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414 md_unregister_thread(mddev->thread);
2415 mddev->thread = NULL;
2416 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2417 if (conf->r10bio_pool)
2418 mempool_destroy(conf->r10bio_pool);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07002419 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420 kfree(conf);
2421 mddev->private = NULL;
2422 return 0;
2423}
2424
NeilBrown6cce3b22006-01-06 00:20:16 -08002425static void raid10_quiesce(mddev_t *mddev, int state)
2426{
NeilBrown070ec552009-06-16 16:54:21 +10002427 conf_t *conf = mddev->private;
NeilBrown6cce3b22006-01-06 00:20:16 -08002428
2429 switch(state) {
2430 case 1:
2431 raise_barrier(conf, 0);
2432 break;
2433 case 0:
2434 lower_barrier(conf);
2435 break;
2436 }
NeilBrown6cce3b22006-01-06 00:20:16 -08002437}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438
Trela, Maciejdab8b292010-03-08 16:02:45 +11002439static void *raid10_takeover_raid0(mddev_t *mddev)
2440{
2441 mdk_rdev_t *rdev;
2442 conf_t *conf;
2443
2444 if (mddev->degraded > 0) {
NeilBrown128595e2010-05-03 14:47:14 +10002445 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
2446 mdname(mddev));
Trela, Maciejdab8b292010-03-08 16:02:45 +11002447 return ERR_PTR(-EINVAL);
2448 }
2449
Trela, Maciejdab8b292010-03-08 16:02:45 +11002450 /* Set new parameters */
2451 mddev->new_level = 10;
2452 /* new layout: far_copies = 1, near_copies = 2 */
2453 mddev->new_layout = (1<<8) + 2;
2454 mddev->new_chunk_sectors = mddev->chunk_sectors;
2455 mddev->delta_disks = mddev->raid_disks;
Trela, Maciejdab8b292010-03-08 16:02:45 +11002456 mddev->raid_disks *= 2;
2457 /* make sure it will be not marked as dirty */
2458 mddev->recovery_cp = MaxSector;
2459
2460 conf = setup_conf(mddev);
NeilBrowne93f68a2010-06-15 09:36:03 +01002461 if (!IS_ERR(conf))
2462 list_for_each_entry(rdev, &mddev->disks, same_set)
2463 if (rdev->raid_disk >= 0)
2464 rdev->new_raid_disk = rdev->raid_disk * 2;
2465
Trela, Maciejdab8b292010-03-08 16:02:45 +11002466 return conf;
2467}
2468
2469static void *raid10_takeover(mddev_t *mddev)
2470{
2471 struct raid0_private_data *raid0_priv;
2472
2473 /* raid10 can take over:
2474 * raid0 - providing it has only two drives
2475 */
2476 if (mddev->level == 0) {
2477 /* for raid0 takeover only one zone is supported */
2478 raid0_priv = mddev->private;
2479 if (raid0_priv->nr_strip_zones > 1) {
NeilBrown128595e2010-05-03 14:47:14 +10002480 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
2481 " with more than one zone.\n",
2482 mdname(mddev));
Trela, Maciejdab8b292010-03-08 16:02:45 +11002483 return ERR_PTR(-EINVAL);
2484 }
2485 return raid10_takeover_raid0(mddev);
2486 }
2487 return ERR_PTR(-EINVAL);
2488}
2489
NeilBrown2604b702006-01-06 00:20:36 -08002490static struct mdk_personality raid10_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491{
2492 .name = "raid10",
NeilBrown2604b702006-01-06 00:20:36 -08002493 .level = 10,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494 .owner = THIS_MODULE,
2495 .make_request = make_request,
2496 .run = run,
2497 .stop = stop,
2498 .status = status,
2499 .error_handler = error,
2500 .hot_add_disk = raid10_add_disk,
2501 .hot_remove_disk= raid10_remove_disk,
2502 .spare_active = raid10_spare_active,
2503 .sync_request = sync_request,
NeilBrown6cce3b22006-01-06 00:20:16 -08002504 .quiesce = raid10_quiesce,
Dan Williams80c3a6c2009-03-17 18:10:40 -07002505 .size = raid10_size,
Trela, Maciejdab8b292010-03-08 16:02:45 +11002506 .takeover = raid10_takeover,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507};
2508
2509static int __init raid_init(void)
2510{
NeilBrown2604b702006-01-06 00:20:36 -08002511 return register_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512}
2513
2514static void raid_exit(void)
2515{
NeilBrown2604b702006-01-06 00:20:36 -08002516 unregister_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517}
2518
2519module_init(raid_init);
2520module_exit(raid_exit);
2521MODULE_LICENSE("GPL");
NeilBrown0efb9e62009-12-14 12:49:58 +11002522MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523MODULE_ALIAS("md-personality-9"); /* RAID10 */
NeilBrownd9d166c2006-01-06 00:20:51 -08002524MODULE_ALIAS("md-raid10");
NeilBrown2604b702006-01-06 00:20:36 -08002525MODULE_ALIAS("md-level-10");