blob: da5129a24b18142d39595c86415bdef40612e353 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * raid10.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 2000-2004 Neil Brown
5 *
6 * RAID-10 support for md.
7 *
8 * Base on code in raid1.c. See raid1.c for futher copyright information.
9 *
10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20
NeilBrown6cce3b22006-01-06 00:20:16 -080021#include "dm-bio-list.h"
Stephen Rothwell25570722008-10-15 09:09:21 +110022#include <linux/delay.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/raid/raid10.h>
NeilBrown6cce3b22006-01-06 00:20:16 -080024#include <linux/raid/bitmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025
26/*
27 * RAID10 provides a combination of RAID0 and RAID1 functionality.
28 * The layout of data is defined by
29 * chunk_size
30 * raid_disks
31 * near_copies (stored in low byte of layout)
32 * far_copies (stored in second byte of layout)
NeilBrownc93983b2006-06-26 00:27:41 -070033 * far_offset (stored in bit 16 of layout )
Linus Torvalds1da177e2005-04-16 15:20:36 -070034 *
35 * The data to be stored is divided into chunks using chunksize.
36 * Each device is divided into far_copies sections.
37 * In each section, chunks are laid out in a style similar to raid0, but
38 * near_copies copies of each chunk is stored (each on a different drive).
39 * The starting device for each section is offset near_copies from the starting
40 * device of the previous section.
NeilBrownc93983b2006-06-26 00:27:41 -070041 * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
Linus Torvalds1da177e2005-04-16 15:20:36 -070042 * drive.
43 * near_copies and far_copies must be at least one, and their product is at most
44 * raid_disks.
NeilBrownc93983b2006-06-26 00:27:41 -070045 *
46 * If far_offset is true, then the far_copies are handled a bit differently.
47 * The copies are still in different stripes, but instead of be very far apart
48 * on disk, there are adjacent stripes.
Linus Torvalds1da177e2005-04-16 15:20:36 -070049 */
50
51/*
52 * Number of guaranteed r10bios in case of extreme VM load:
53 */
54#define NR_RAID10_BIOS 256
55
56static void unplug_slaves(mddev_t *mddev);
57
NeilBrown0a27ec92006-01-06 00:20:13 -080058static void allow_barrier(conf_t *conf);
59static void lower_barrier(conf_t *conf);
60
Al Virodd0fc662005-10-07 07:46:04 +010061static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070062{
63 conf_t *conf = data;
64 r10bio_t *r10_bio;
65 int size = offsetof(struct r10bio_s, devs[conf->copies]);
66
67 /* allocate a r10bio with room for raid_disks entries in the bios array */
NeilBrown9ffae0c2006-01-06 00:20:32 -080068 r10_bio = kzalloc(size, gfp_flags);
69 if (!r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -070070 unplug_slaves(conf->mddev);
71
72 return r10_bio;
73}
74
75static void r10bio_pool_free(void *r10_bio, void *data)
76{
77 kfree(r10_bio);
78}
79
NeilBrown0310fa22008-08-05 15:54:14 +100080/* Maximum size of each resync request */
Linus Torvalds1da177e2005-04-16 15:20:36 -070081#define RESYNC_BLOCK_SIZE (64*1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -070082#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
NeilBrown0310fa22008-08-05 15:54:14 +100083/* amount of memory to reserve for resync requests */
84#define RESYNC_WINDOW (1024*1024)
85/* maximum number of concurrent requests, memory permitting */
86#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -070087
88/*
89 * When performing a resync, we need to read and compare, so
90 * we need as many pages are there are copies.
91 * When performing a recovery, we need 2 bios, one for read,
92 * one for write (we recover only one drive per r10buf)
93 *
94 */
Al Virodd0fc662005-10-07 07:46:04 +010095static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -070096{
97 conf_t *conf = data;
98 struct page *page;
99 r10bio_t *r10_bio;
100 struct bio *bio;
101 int i, j;
102 int nalloc;
103
104 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
105 if (!r10_bio) {
106 unplug_slaves(conf->mddev);
107 return NULL;
108 }
109
110 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
111 nalloc = conf->copies; /* resync */
112 else
113 nalloc = 2; /* recovery */
114
115 /*
116 * Allocate bios.
117 */
118 for (j = nalloc ; j-- ; ) {
119 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
120 if (!bio)
121 goto out_free_bio;
122 r10_bio->devs[j].bio = bio;
123 }
124 /*
125 * Allocate RESYNC_PAGES data pages and attach them
126 * where needed.
127 */
128 for (j = 0 ; j < nalloc; j++) {
129 bio = r10_bio->devs[j].bio;
130 for (i = 0; i < RESYNC_PAGES; i++) {
131 page = alloc_page(gfp_flags);
132 if (unlikely(!page))
133 goto out_free_pages;
134
135 bio->bi_io_vec[i].bv_page = page;
136 }
137 }
138
139 return r10_bio;
140
141out_free_pages:
142 for ( ; i > 0 ; i--)
NeilBrown1345b1d2006-01-06 00:20:40 -0800143 safe_put_page(bio->bi_io_vec[i-1].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 while (j--)
145 for (i = 0; i < RESYNC_PAGES ; i++)
NeilBrown1345b1d2006-01-06 00:20:40 -0800146 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147 j = -1;
148out_free_bio:
149 while ( ++j < nalloc )
150 bio_put(r10_bio->devs[j].bio);
151 r10bio_pool_free(r10_bio, conf);
152 return NULL;
153}
154
155static void r10buf_pool_free(void *__r10_bio, void *data)
156{
157 int i;
158 conf_t *conf = data;
159 r10bio_t *r10bio = __r10_bio;
160 int j;
161
162 for (j=0; j < conf->copies; j++) {
163 struct bio *bio = r10bio->devs[j].bio;
164 if (bio) {
165 for (i = 0; i < RESYNC_PAGES; i++) {
NeilBrown1345b1d2006-01-06 00:20:40 -0800166 safe_put_page(bio->bi_io_vec[i].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167 bio->bi_io_vec[i].bv_page = NULL;
168 }
169 bio_put(bio);
170 }
171 }
172 r10bio_pool_free(r10bio, conf);
173}
174
175static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
176{
177 int i;
178
179 for (i = 0; i < conf->copies; i++) {
180 struct bio **bio = & r10_bio->devs[i].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -0800181 if (*bio && *bio != IO_BLOCKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182 bio_put(*bio);
183 *bio = NULL;
184 }
185}
186
Arjan van de Ven858119e2006-01-14 13:20:43 -0800187static void free_r10bio(r10bio_t *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188{
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 conf_t *conf = mddev_to_conf(r10_bio->mddev);
190
191 /*
192 * Wake up any possible resync thread that waits for the device
193 * to go idle.
194 */
NeilBrown0a27ec92006-01-06 00:20:13 -0800195 allow_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700196
197 put_all_bios(conf, r10_bio);
198 mempool_free(r10_bio, conf->r10bio_pool);
199}
200
Arjan van de Ven858119e2006-01-14 13:20:43 -0800201static void put_buf(r10bio_t *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202{
203 conf_t *conf = mddev_to_conf(r10_bio->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204
205 mempool_free(r10_bio, conf->r10buf_pool);
206
NeilBrown0a27ec92006-01-06 00:20:13 -0800207 lower_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208}
209
210static void reschedule_retry(r10bio_t *r10_bio)
211{
212 unsigned long flags;
213 mddev_t *mddev = r10_bio->mddev;
214 conf_t *conf = mddev_to_conf(mddev);
215
216 spin_lock_irqsave(&conf->device_lock, flags);
217 list_add(&r10_bio->retry_list, &conf->retry_list);
NeilBrown4443ae12006-01-06 00:20:28 -0800218 conf->nr_queued ++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 spin_unlock_irqrestore(&conf->device_lock, flags);
220
Arthur Jones388667b2008-07-25 12:03:38 -0700221 /* wake up frozen array... */
222 wake_up(&conf->wait_barrier);
223
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224 md_wakeup_thread(mddev->thread);
225}
226
227/*
228 * raid_end_bio_io() is called when we have finished servicing a mirrored
229 * operation and are ready to return a success/failure code to the buffer
230 * cache layer.
231 */
232static void raid_end_bio_io(r10bio_t *r10_bio)
233{
234 struct bio *bio = r10_bio->master_bio;
235
NeilBrown6712ecf2007-09-27 12:47:43 +0200236 bio_endio(bio,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
238 free_r10bio(r10_bio);
239}
240
241/*
242 * Update disk head position estimator based on IRQ completion info.
243 */
244static inline void update_head_pos(int slot, r10bio_t *r10_bio)
245{
246 conf_t *conf = mddev_to_conf(r10_bio->mddev);
247
248 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
249 r10_bio->devs[slot].addr + (r10_bio->sectors);
250}
251
NeilBrown6712ecf2007-09-27 12:47:43 +0200252static void raid10_end_read_request(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253{
254 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
255 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
256 int slot, dev;
257 conf_t *conf = mddev_to_conf(r10_bio->mddev);
258
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259
260 slot = r10_bio->read_slot;
261 dev = r10_bio->devs[slot].devnum;
262 /*
263 * this branch is our 'one mirror IO has finished' event handler:
264 */
NeilBrown4443ae12006-01-06 00:20:28 -0800265 update_head_pos(slot, r10_bio);
266
267 if (uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268 /*
269 * Set R10BIO_Uptodate in our master bio, so that
270 * we will return a good error code to the higher
271 * levels even if IO on some other mirrored buffer fails.
272 *
273 * The 'master' represents the composite IO operation to
274 * user-side. So if something waits for IO, then it will
275 * wait for the 'master' bio.
276 */
277 set_bit(R10BIO_Uptodate, &r10_bio->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 raid_end_bio_io(r10_bio);
NeilBrown4443ae12006-01-06 00:20:28 -0800279 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 /*
281 * oops, read error:
282 */
283 char b[BDEVNAME_SIZE];
284 if (printk_ratelimit())
285 printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
286 bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
287 reschedule_retry(r10_bio);
288 }
289
290 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291}
292
NeilBrown6712ecf2007-09-27 12:47:43 +0200293static void raid10_end_write_request(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294{
295 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
296 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
297 int slot, dev;
298 conf_t *conf = mddev_to_conf(r10_bio->mddev);
299
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 for (slot = 0; slot < conf->copies; slot++)
301 if (r10_bio->devs[slot].bio == bio)
302 break;
303 dev = r10_bio->devs[slot].devnum;
304
305 /*
306 * this branch is our 'one mirror IO has finished' event handler:
307 */
NeilBrown6cce3b22006-01-06 00:20:16 -0800308 if (!uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
NeilBrown6cce3b22006-01-06 00:20:16 -0800310 /* an I/O failed, we can't clear the bitmap */
311 set_bit(R10BIO_Degraded, &r10_bio->state);
312 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313 /*
314 * Set R10BIO_Uptodate in our master bio, so that
315 * we will return a good error code for to the higher
316 * levels even if IO on some other mirrored buffer fails.
317 *
318 * The 'master' represents the composite IO operation to
319 * user-side. So if something waits for IO, then it will
320 * wait for the 'master' bio.
321 */
322 set_bit(R10BIO_Uptodate, &r10_bio->state);
323
324 update_head_pos(slot, r10_bio);
325
326 /*
327 *
328 * Let's see if all mirrored write operations have finished
329 * already.
330 */
331 if (atomic_dec_and_test(&r10_bio->remaining)) {
NeilBrown6cce3b22006-01-06 00:20:16 -0800332 /* clear the bitmap if all writes complete successfully */
333 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
334 r10_bio->sectors,
335 !test_bit(R10BIO_Degraded, &r10_bio->state),
336 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700337 md_write_end(r10_bio->mddev);
338 raid_end_bio_io(r10_bio);
339 }
340
341 rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342}
343
344
345/*
346 * RAID10 layout manager
347 * Aswell as the chunksize and raid_disks count, there are two
348 * parameters: near_copies and far_copies.
349 * near_copies * far_copies must be <= raid_disks.
350 * Normally one of these will be 1.
351 * If both are 1, we get raid0.
352 * If near_copies == raid_disks, we get raid1.
353 *
354 * Chunks are layed out in raid0 style with near_copies copies of the
355 * first chunk, followed by near_copies copies of the next chunk and
356 * so on.
357 * If far_copies > 1, then after 1/far_copies of the array has been assigned
358 * as described above, we start again with a device offset of near_copies.
359 * So we effectively have another copy of the whole array further down all
360 * the drives, but with blocks on different drives.
361 * With this layout, and block is never stored twice on the one device.
362 *
363 * raid10_find_phys finds the sector offset of a given virtual sector
NeilBrownc93983b2006-06-26 00:27:41 -0700364 * on each device that it is on.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365 *
366 * raid10_find_virt does the reverse mapping, from a device and a
367 * sector offset to a virtual address
368 */
369
370static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
371{
372 int n,f;
373 sector_t sector;
374 sector_t chunk;
375 sector_t stripe;
376 int dev;
377
378 int slot = 0;
379
380 /* now calculate first sector/dev */
381 chunk = r10bio->sector >> conf->chunk_shift;
382 sector = r10bio->sector & conf->chunk_mask;
383
384 chunk *= conf->near_copies;
385 stripe = chunk;
386 dev = sector_div(stripe, conf->raid_disks);
NeilBrownc93983b2006-06-26 00:27:41 -0700387 if (conf->far_offset)
388 stripe *= conf->far_copies;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389
390 sector += stripe << conf->chunk_shift;
391
392 /* and calculate all the others */
393 for (n=0; n < conf->near_copies; n++) {
394 int d = dev;
395 sector_t s = sector;
396 r10bio->devs[slot].addr = sector;
397 r10bio->devs[slot].devnum = d;
398 slot++;
399
400 for (f = 1; f < conf->far_copies; f++) {
401 d += conf->near_copies;
402 if (d >= conf->raid_disks)
403 d -= conf->raid_disks;
404 s += conf->stride;
405 r10bio->devs[slot].devnum = d;
406 r10bio->devs[slot].addr = s;
407 slot++;
408 }
409 dev++;
410 if (dev >= conf->raid_disks) {
411 dev = 0;
412 sector += (conf->chunk_mask + 1);
413 }
414 }
415 BUG_ON(slot != conf->copies);
416}
417
418static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
419{
420 sector_t offset, chunk, vchunk;
421
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 offset = sector & conf->chunk_mask;
NeilBrownc93983b2006-06-26 00:27:41 -0700423 if (conf->far_offset) {
424 int fc;
425 chunk = sector >> conf->chunk_shift;
426 fc = sector_div(chunk, conf->far_copies);
427 dev -= fc * conf->near_copies;
428 if (dev < 0)
429 dev += conf->raid_disks;
430 } else {
NeilBrown64a742b2007-02-28 20:11:18 -0800431 while (sector >= conf->stride) {
NeilBrownc93983b2006-06-26 00:27:41 -0700432 sector -= conf->stride;
433 if (dev < conf->near_copies)
434 dev += conf->raid_disks - conf->near_copies;
435 else
436 dev -= conf->near_copies;
437 }
438 chunk = sector >> conf->chunk_shift;
439 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 vchunk = chunk * conf->raid_disks + dev;
441 sector_div(vchunk, conf->near_copies);
442 return (vchunk << conf->chunk_shift) + offset;
443}
444
445/**
446 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
447 * @q: request queue
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200448 * @bvm: properties of new bio
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 * @biovec: the request that could be merged to it.
450 *
451 * Return amount of bytes we can accept at this offset
452 * If near_copies == raid_disk, there are no striping issues,
453 * but in that case, the function isn't called at all.
454 */
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200455static int raid10_mergeable_bvec(struct request_queue *q,
456 struct bvec_merge_data *bvm,
457 struct bio_vec *biovec)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458{
459 mddev_t *mddev = q->queuedata;
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200460 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 int max;
462 unsigned int chunk_sectors = mddev->chunk_size >> 9;
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200463 unsigned int bio_sectors = bvm->bi_size >> 9;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464
465 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
466 if (max < 0) max = 0; /* bio_add cannot handle a negative return */
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200467 if (max <= biovec->bv_len && bio_sectors == 0)
468 return biovec->bv_len;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 else
470 return max;
471}
472
473/*
474 * This routine returns the disk from which the requested read should
475 * be done. There is a per-array 'next expected sequential IO' sector
476 * number - if this matches on the next IO then we use the last disk.
477 * There is also a per-disk 'last know head position' sector that is
478 * maintained from IRQ contexts, both the normal and the resync IO
479 * completion handlers update this position correctly. If there is no
480 * perfect sequential match then we pick the disk whose head is closest.
481 *
482 * If there are 2 mirrors in the same 2 devices, performance degrades
483 * because position is mirror, not device based.
484 *
485 * The rdev for the device selected will have nr_pending incremented.
486 */
487
488/*
489 * FIXME: possibly should rethink readbalancing and do it differently
490 * depending on near_copies / far_copies geometry.
491 */
492static int read_balance(conf_t *conf, r10bio_t *r10_bio)
493{
494 const unsigned long this_sector = r10_bio->sector;
495 int disk, slot, nslot;
496 const int sectors = r10_bio->sectors;
497 sector_t new_distance, current_distance;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800498 mdk_rdev_t *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499
500 raid10_find_phys(conf, r10_bio);
501 rcu_read_lock();
502 /*
503 * Check if we can balance. We can balance on the whole
NeilBrown6cce3b22006-01-06 00:20:16 -0800504 * device if no resync is going on (recovery is ok), or below
505 * the resync window. We take the first readable disk when
506 * above the resync window.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 */
508 if (conf->mddev->recovery_cp < MaxSector
509 && (this_sector + sectors >= conf->next_resync)) {
510 /* make sure that disk is operational */
511 slot = 0;
512 disk = r10_bio->devs[slot].devnum;
513
Suzanne Woodd6065f72005-11-08 21:39:27 -0800514 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800515 r10_bio->devs[slot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800516 !test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 slot++;
518 if (slot == conf->copies) {
519 slot = 0;
520 disk = -1;
521 break;
522 }
523 disk = r10_bio->devs[slot].devnum;
524 }
525 goto rb_out;
526 }
527
528
529 /* make sure the disk is operational */
530 slot = 0;
531 disk = r10_bio->devs[slot].devnum;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800532 while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800533 r10_bio->devs[slot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800534 !test_bit(In_sync, &rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 slot ++;
536 if (slot == conf->copies) {
537 disk = -1;
538 goto rb_out;
539 }
540 disk = r10_bio->devs[slot].devnum;
541 }
542
543
NeilBrown3ec67ac2005-09-09 16:23:40 -0700544 current_distance = abs(r10_bio->devs[slot].addr -
545 conf->mirrors[disk].head_position);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546
Keld Simonsen8ed3a192008-03-04 14:29:34 -0800547 /* Find the disk whose head is closest,
548 * or - for far > 1 - find the closest to partition beginning */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549
550 for (nslot = slot; nslot < conf->copies; nslot++) {
551 int ndisk = r10_bio->devs[nslot].devnum;
552
553
Suzanne Woodd6065f72005-11-08 21:39:27 -0800554 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
NeilBrown0eb3ff12006-01-06 00:20:29 -0800555 r10_bio->devs[nslot].bio == IO_BLOCKED ||
NeilBrownb2d444d2005-11-08 21:39:31 -0800556 !test_bit(In_sync, &rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557 continue;
558
NeilBrown22dfdf52005-11-28 13:44:09 -0800559 /* This optimisation is debatable, and completely destroys
560 * sequential read speed for 'far copies' arrays. So only
561 * keep it for 'near' arrays, and review those later.
562 */
563 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 disk = ndisk;
565 slot = nslot;
566 break;
567 }
Keld Simonsen8ed3a192008-03-04 14:29:34 -0800568
569 /* for far > 1 always use the lowest address */
570 if (conf->far_copies > 1)
571 new_distance = r10_bio->devs[nslot].addr;
572 else
573 new_distance = abs(r10_bio->devs[nslot].addr -
574 conf->mirrors[ndisk].head_position);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 if (new_distance < current_distance) {
576 current_distance = new_distance;
577 disk = ndisk;
578 slot = nslot;
579 }
580 }
581
582rb_out:
583 r10_bio->read_slot = slot;
584/* conf->next_seq_sect = this_sector + sectors;*/
585
Suzanne Woodd6065f72005-11-08 21:39:27 -0800586 if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
NeilBrown29fc7e32006-02-03 03:03:41 -0800588 else
589 disk = -1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 rcu_read_unlock();
591
592 return disk;
593}
594
595static void unplug_slaves(mddev_t *mddev)
596{
597 conf_t *conf = mddev_to_conf(mddev);
598 int i;
599
600 rcu_read_lock();
601 for (i=0; i<mddev->raid_disks; i++) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800602 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
NeilBrownb2d444d2005-11-08 21:39:31 -0800603 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200604 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
606 atomic_inc(&rdev->nr_pending);
607 rcu_read_unlock();
608
Alan D. Brunelle2ad8b1e2007-11-07 14:26:56 -0500609 blk_unplug(r_queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610
611 rdev_dec_pending(rdev, mddev);
612 rcu_read_lock();
613 }
614 }
615 rcu_read_unlock();
616}
617
Jens Axboe165125e2007-07-24 09:28:11 +0200618static void raid10_unplug(struct request_queue *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619{
NeilBrown6cce3b22006-01-06 00:20:16 -0800620 mddev_t *mddev = q->queuedata;
621
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 unplug_slaves(q->queuedata);
NeilBrown6cce3b22006-01-06 00:20:16 -0800623 md_wakeup_thread(mddev->thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624}
625
NeilBrown0d129222006-10-03 01:15:54 -0700626static int raid10_congested(void *data, int bits)
627{
628 mddev_t *mddev = data;
629 conf_t *conf = mddev_to_conf(mddev);
630 int i, ret = 0;
631
632 rcu_read_lock();
633 for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
634 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
635 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200636 struct request_queue *q = bdev_get_queue(rdev->bdev);
NeilBrown0d129222006-10-03 01:15:54 -0700637
638 ret |= bdi_congested(&q->backing_dev_info, bits);
639 }
640 }
641 rcu_read_unlock();
642 return ret;
643}
644
NeilBrowna35e63e2008-03-04 14:29:29 -0800645static int flush_pending_writes(conf_t *conf)
646{
647 /* Any writes that have been queued but are awaiting
648 * bitmap updates get flushed here.
649 * We return 1 if any requests were actually submitted.
650 */
651 int rv = 0;
NeilBrown0d129222006-10-03 01:15:54 -0700652
NeilBrowna35e63e2008-03-04 14:29:29 -0800653 spin_lock_irq(&conf->device_lock);
654
655 if (conf->pending_bio_list.head) {
656 struct bio *bio;
657 bio = bio_list_get(&conf->pending_bio_list);
658 blk_remove_plug(conf->mddev->queue);
659 spin_unlock_irq(&conf->device_lock);
660 /* flush any pending bitmap writes to disk
661 * before proceeding w/ I/O */
662 bitmap_unplug(conf->mddev->bitmap);
663
664 while (bio) { /* submit pending writes */
665 struct bio *next = bio->bi_next;
666 bio->bi_next = NULL;
667 generic_make_request(bio);
668 bio = next;
669 }
670 rv = 1;
671 } else
672 spin_unlock_irq(&conf->device_lock);
673 return rv;
674}
NeilBrown0a27ec92006-01-06 00:20:13 -0800675/* Barriers....
676 * Sometimes we need to suspend IO while we do something else,
677 * either some resync/recovery, or reconfigure the array.
678 * To do this we raise a 'barrier'.
679 * The 'barrier' is a counter that can be raised multiple times
680 * to count how many activities are happening which preclude
681 * normal IO.
682 * We can only raise the barrier if there is no pending IO.
683 * i.e. if nr_pending == 0.
684 * We choose only to raise the barrier if no-one is waiting for the
685 * barrier to go down. This means that as soon as an IO request
686 * is ready, no other operations which require a barrier will start
687 * until the IO request has had a chance.
688 *
689 * So: regular IO calls 'wait_barrier'. When that returns there
690 * is no backgroup IO happening, It must arrange to call
691 * allow_barrier when it has finished its IO.
692 * backgroup IO calls must call raise_barrier. Once that returns
693 * there is no normal IO happeing. It must arrange to call
694 * lower_barrier when the particular background IO completes.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700695 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696
NeilBrown6cce3b22006-01-06 00:20:16 -0800697static void raise_barrier(conf_t *conf, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698{
NeilBrown6cce3b22006-01-06 00:20:16 -0800699 BUG_ON(force && !conf->barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 spin_lock_irq(&conf->resync_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701
NeilBrown6cce3b22006-01-06 00:20:16 -0800702 /* Wait until no block IO is waiting (unless 'force') */
703 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
NeilBrown0a27ec92006-01-06 00:20:13 -0800704 conf->resync_lock,
705 raid10_unplug(conf->mddev->queue));
706
707 /* block any new IO from starting */
708 conf->barrier++;
709
710 /* No wait for all pending IO to complete */
711 wait_event_lock_irq(conf->wait_barrier,
712 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
713 conf->resync_lock,
714 raid10_unplug(conf->mddev->queue));
715
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 spin_unlock_irq(&conf->resync_lock);
717}
718
NeilBrown0a27ec92006-01-06 00:20:13 -0800719static void lower_barrier(conf_t *conf)
720{
721 unsigned long flags;
722 spin_lock_irqsave(&conf->resync_lock, flags);
723 conf->barrier--;
724 spin_unlock_irqrestore(&conf->resync_lock, flags);
725 wake_up(&conf->wait_barrier);
726}
727
728static void wait_barrier(conf_t *conf)
729{
730 spin_lock_irq(&conf->resync_lock);
731 if (conf->barrier) {
732 conf->nr_waiting++;
733 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
734 conf->resync_lock,
735 raid10_unplug(conf->mddev->queue));
736 conf->nr_waiting--;
737 }
738 conf->nr_pending++;
739 spin_unlock_irq(&conf->resync_lock);
740}
741
742static void allow_barrier(conf_t *conf)
743{
744 unsigned long flags;
745 spin_lock_irqsave(&conf->resync_lock, flags);
746 conf->nr_pending--;
747 spin_unlock_irqrestore(&conf->resync_lock, flags);
748 wake_up(&conf->wait_barrier);
749}
750
NeilBrown4443ae12006-01-06 00:20:28 -0800751static void freeze_array(conf_t *conf)
752{
753 /* stop syncio and normal IO and wait for everything to
NeilBrownf1885932006-01-06 00:20:42 -0800754 * go quiet.
NeilBrown4443ae12006-01-06 00:20:28 -0800755 * We increment barrier and nr_waiting, and then
NeilBrown1c830532008-03-04 14:29:35 -0800756 * wait until nr_pending match nr_queued+1
757 * This is called in the context of one normal IO request
758 * that has failed. Thus any sync request that might be pending
759 * will be blocked by nr_pending, and we need to wait for
760 * pending IO requests to complete or be queued for re-try.
761 * Thus the number queued (nr_queued) plus this request (1)
762 * must match the number of pending IOs (nr_pending) before
763 * we continue.
NeilBrown4443ae12006-01-06 00:20:28 -0800764 */
765 spin_lock_irq(&conf->resync_lock);
766 conf->barrier++;
767 conf->nr_waiting++;
768 wait_event_lock_irq(conf->wait_barrier,
NeilBrown1c830532008-03-04 14:29:35 -0800769 conf->nr_pending == conf->nr_queued+1,
NeilBrown4443ae12006-01-06 00:20:28 -0800770 conf->resync_lock,
NeilBrowna35e63e2008-03-04 14:29:29 -0800771 ({ flush_pending_writes(conf);
772 raid10_unplug(conf->mddev->queue); }));
NeilBrown4443ae12006-01-06 00:20:28 -0800773 spin_unlock_irq(&conf->resync_lock);
774}
775
776static void unfreeze_array(conf_t *conf)
777{
778 /* reverse the effect of the freeze */
779 spin_lock_irq(&conf->resync_lock);
780 conf->barrier--;
781 conf->nr_waiting--;
782 wake_up(&conf->wait_barrier);
783 spin_unlock_irq(&conf->resync_lock);
784}
785
Jens Axboe165125e2007-07-24 09:28:11 +0200786static int make_request(struct request_queue *q, struct bio * bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787{
788 mddev_t *mddev = q->queuedata;
789 conf_t *conf = mddev_to_conf(mddev);
790 mirror_info_t *mirror;
791 r10bio_t *r10_bio;
792 struct bio *read_bio;
Tejun Heoc9959052008-08-25 19:47:21 +0900793 int cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 int i;
795 int chunk_sects = conf->chunk_mask + 1;
Jens Axboea3623572005-11-01 09:26:16 +0100796 const int rw = bio_data_dir(bio);
Lars Ellenberge3881a62007-01-10 23:15:37 -0800797 const int do_sync = bio_sync(bio);
NeilBrown6cce3b22006-01-06 00:20:16 -0800798 struct bio_list bl;
799 unsigned long flags;
Dan Williams6bfe0b42008-04-30 00:52:32 -0700800 mdk_rdev_t *blocked_rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801
NeilBrowne5dcdd82005-09-09 16:23:41 -0700802 if (unlikely(bio_barrier(bio))) {
NeilBrown6712ecf2007-09-27 12:47:43 +0200803 bio_endio(bio, -EOPNOTSUPP);
NeilBrowne5dcdd82005-09-09 16:23:41 -0700804 return 0;
805 }
806
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 /* If this request crosses a chunk boundary, we need to
808 * split it. This will only happen for 1 PAGE (or less) requests.
809 */
810 if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
811 > chunk_sects &&
812 conf->near_copies < conf->raid_disks)) {
813 struct bio_pair *bp;
814 /* Sanity check -- queue functions should prevent this happening */
815 if (bio->bi_vcnt != 1 ||
816 bio->bi_idx != 0)
817 goto bad_map;
818 /* This is a one page bio that upper layers
819 * refuse to split for us, so we need to split it.
820 */
Denis ChengRq6feef532008-10-09 08:57:05 +0200821 bp = bio_split(bio,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
823 if (make_request(q, &bp->bio1))
824 generic_make_request(&bp->bio1);
825 if (make_request(q, &bp->bio2))
826 generic_make_request(&bp->bio2);
827
828 bio_pair_release(bp);
829 return 0;
830 bad_map:
831 printk("raid10_make_request bug: can't convert block across chunks"
832 " or bigger than %dk %llu %d\n", chunk_sects/2,
833 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
834
NeilBrown6712ecf2007-09-27 12:47:43 +0200835 bio_io_error(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836 return 0;
837 }
838
NeilBrown3d310eb2005-06-21 17:17:26 -0700839 md_write_start(mddev, bio);
NeilBrown06d91a52005-06-21 17:17:12 -0700840
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 /*
842 * Register the new request and wait if the reconstruction
843 * thread has put up a bar for new requests.
844 * Continue immediately if no resync is active currently.
845 */
NeilBrown0a27ec92006-01-06 00:20:13 -0800846 wait_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700847
Tejun Heo074a7ac2008-08-25 19:56:14 +0900848 cpu = part_stat_lock();
849 part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
850 part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
851 bio_sectors(bio));
852 part_stat_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853
854 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
855
856 r10_bio->master_bio = bio;
857 r10_bio->sectors = bio->bi_size >> 9;
858
859 r10_bio->mddev = mddev;
860 r10_bio->sector = bio->bi_sector;
NeilBrown6cce3b22006-01-06 00:20:16 -0800861 r10_bio->state = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862
Jens Axboea3623572005-11-01 09:26:16 +0100863 if (rw == READ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864 /*
865 * read balancing logic:
866 */
867 int disk = read_balance(conf, r10_bio);
868 int slot = r10_bio->read_slot;
869 if (disk < 0) {
870 raid_end_bio_io(r10_bio);
871 return 0;
872 }
873 mirror = conf->mirrors + disk;
874
875 read_bio = bio_clone(bio, GFP_NOIO);
876
877 r10_bio->devs[slot].bio = read_bio;
878
879 read_bio->bi_sector = r10_bio->devs[slot].addr +
880 mirror->rdev->data_offset;
881 read_bio->bi_bdev = mirror->rdev->bdev;
882 read_bio->bi_end_io = raid10_end_read_request;
Lars Ellenberge3881a62007-01-10 23:15:37 -0800883 read_bio->bi_rw = READ | do_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 read_bio->bi_private = r10_bio;
885
886 generic_make_request(read_bio);
887 return 0;
888 }
889
890 /*
891 * WRITE:
892 */
Dan Williams6bfe0b42008-04-30 00:52:32 -0700893 /* first select target devices under rcu_lock and
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894 * inc refcount on their rdev. Record them by setting
895 * bios[x] to bio
896 */
897 raid10_find_phys(conf, r10_bio);
Dan Williams6bfe0b42008-04-30 00:52:32 -0700898 retry_write:
Harvey Harrisoncb6969e2008-05-06 20:42:32 -0700899 blocked_rdev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 rcu_read_lock();
901 for (i = 0; i < conf->copies; i++) {
902 int d = r10_bio->devs[i].devnum;
Suzanne Woodd6065f72005-11-08 21:39:27 -0800903 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
Dan Williams6bfe0b42008-04-30 00:52:32 -0700904 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
905 atomic_inc(&rdev->nr_pending);
906 blocked_rdev = rdev;
907 break;
908 }
909 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Suzanne Woodd6065f72005-11-08 21:39:27 -0800910 atomic_inc(&rdev->nr_pending);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 r10_bio->devs[i].bio = bio;
NeilBrown6cce3b22006-01-06 00:20:16 -0800912 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 r10_bio->devs[i].bio = NULL;
NeilBrown6cce3b22006-01-06 00:20:16 -0800914 set_bit(R10BIO_Degraded, &r10_bio->state);
915 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 }
917 rcu_read_unlock();
918
Dan Williams6bfe0b42008-04-30 00:52:32 -0700919 if (unlikely(blocked_rdev)) {
920 /* Have to wait for this device to get unblocked, then retry */
921 int j;
922 int d;
923
924 for (j = 0; j < i; j++)
925 if (r10_bio->devs[j].bio) {
926 d = r10_bio->devs[j].devnum;
927 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
928 }
929 allow_barrier(conf);
930 md_wait_for_blocked_rdev(blocked_rdev, mddev);
931 wait_barrier(conf);
932 goto retry_write;
933 }
934
NeilBrown6cce3b22006-01-06 00:20:16 -0800935 atomic_set(&r10_bio->remaining, 0);
NeilBrown06d91a52005-06-21 17:17:12 -0700936
NeilBrown6cce3b22006-01-06 00:20:16 -0800937 bio_list_init(&bl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 for (i = 0; i < conf->copies; i++) {
939 struct bio *mbio;
940 int d = r10_bio->devs[i].devnum;
941 if (!r10_bio->devs[i].bio)
942 continue;
943
944 mbio = bio_clone(bio, GFP_NOIO);
945 r10_bio->devs[i].bio = mbio;
946
947 mbio->bi_sector = r10_bio->devs[i].addr+
948 conf->mirrors[d].rdev->data_offset;
949 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
950 mbio->bi_end_io = raid10_end_write_request;
Lars Ellenberge3881a62007-01-10 23:15:37 -0800951 mbio->bi_rw = WRITE | do_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952 mbio->bi_private = r10_bio;
953
954 atomic_inc(&r10_bio->remaining);
NeilBrown6cce3b22006-01-06 00:20:16 -0800955 bio_list_add(&bl, mbio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 }
957
Arne Redlichf6f953a2007-07-31 00:37:57 -0700958 if (unlikely(!atomic_read(&r10_bio->remaining))) {
959 /* the array is dead */
960 md_write_end(mddev);
961 raid_end_bio_io(r10_bio);
962 return 0;
963 }
964
NeilBrown6cce3b22006-01-06 00:20:16 -0800965 bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
966 spin_lock_irqsave(&conf->device_lock, flags);
967 bio_list_merge(&conf->pending_bio_list, &bl);
968 blk_plug_device(mddev->queue);
969 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970
NeilBrowna35e63e2008-03-04 14:29:29 -0800971 /* In case raid10d snuck in to freeze_array */
972 wake_up(&conf->wait_barrier);
973
Lars Ellenberge3881a62007-01-10 23:15:37 -0800974 if (do_sync)
975 md_wakeup_thread(mddev->thread);
976
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 return 0;
978}
979
980static void status(struct seq_file *seq, mddev_t *mddev)
981{
982 conf_t *conf = mddev_to_conf(mddev);
983 int i;
984
985 if (conf->near_copies < conf->raid_disks)
986 seq_printf(seq, " %dK chunks", mddev->chunk_size/1024);
987 if (conf->near_copies > 1)
988 seq_printf(seq, " %d near-copies", conf->near_copies);
NeilBrownc93983b2006-06-26 00:27:41 -0700989 if (conf->far_copies > 1) {
990 if (conf->far_offset)
991 seq_printf(seq, " %d offset-copies", conf->far_copies);
992 else
993 seq_printf(seq, " %d far-copies", conf->far_copies);
994 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
NeilBrown76186dd2006-10-03 01:15:48 -0700996 conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 for (i = 0; i < conf->raid_disks; i++)
998 seq_printf(seq, "%s",
999 conf->mirrors[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001000 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001 seq_printf(seq, "]");
1002}
1003
1004static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1005{
1006 char b[BDEVNAME_SIZE];
1007 conf_t *conf = mddev_to_conf(mddev);
1008
1009 /*
1010 * If it is not operational, then we have already marked it as dead
1011 * else if it is the last working disks, ignore the error, let the
1012 * next level up know.
1013 * else mark the drive as failed
1014 */
NeilBrownb2d444d2005-11-08 21:39:31 -08001015 if (test_bit(In_sync, &rdev->flags)
NeilBrown76186dd2006-10-03 01:15:48 -07001016 && conf->raid_disks-mddev->degraded == 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017 /*
1018 * Don't fail the drive, just return an IO error.
1019 * The test should really be more sophisticated than
1020 * "working_disks == 1", but it isn't critical, and
1021 * can wait until we do more sophisticated "is the drive
1022 * really dead" tests...
1023 */
1024 return;
NeilBrownc04be0a2006-10-03 01:15:53 -07001025 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1026 unsigned long flags;
1027 spin_lock_irqsave(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001028 mddev->degraded++;
NeilBrownc04be0a2006-10-03 01:15:53 -07001029 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030 /*
1031 * if recovery is running, make sure it aborts.
1032 */
NeilBrowndfc70642008-05-23 13:04:39 -07001033 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034 }
NeilBrownb2d444d2005-11-08 21:39:31 -08001035 set_bit(Faulty, &rdev->flags);
NeilBrown850b2b42006-10-03 01:15:46 -07001036 set_bit(MD_CHANGE_DEVS, &mddev->flags);
Nick Andrewd7a420c2008-04-28 02:15:55 -07001037 printk(KERN_ALERT "raid10: Disk failure on %s, disabling device.\n"
1038 "raid10: Operation continuing on %d devices.\n",
NeilBrown76186dd2006-10-03 01:15:48 -07001039 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040}
1041
1042static void print_conf(conf_t *conf)
1043{
1044 int i;
1045 mirror_info_t *tmp;
1046
1047 printk("RAID10 conf printout:\n");
1048 if (!conf) {
1049 printk("(!conf)\n");
1050 return;
1051 }
NeilBrown76186dd2006-10-03 01:15:48 -07001052 printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053 conf->raid_disks);
1054
1055 for (i = 0; i < conf->raid_disks; i++) {
1056 char b[BDEVNAME_SIZE];
1057 tmp = conf->mirrors + i;
1058 if (tmp->rdev)
1059 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
NeilBrownb2d444d2005-11-08 21:39:31 -08001060 i, !test_bit(In_sync, &tmp->rdev->flags),
1061 !test_bit(Faulty, &tmp->rdev->flags),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 bdevname(tmp->rdev->bdev,b));
1063 }
1064}
1065
1066static void close_sync(conf_t *conf)
1067{
NeilBrown0a27ec92006-01-06 00:20:13 -08001068 wait_barrier(conf);
1069 allow_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070
1071 mempool_destroy(conf->r10buf_pool);
1072 conf->r10buf_pool = NULL;
1073}
1074
NeilBrown6d508242005-09-09 16:24:03 -07001075/* check if there are enough drives for
1076 * every block to appear on atleast one
1077 */
1078static int enough(conf_t *conf)
1079{
1080 int first = 0;
1081
1082 do {
1083 int n = conf->copies;
1084 int cnt = 0;
1085 while (n--) {
1086 if (conf->mirrors[first].rdev)
1087 cnt++;
1088 first = (first+1) % conf->raid_disks;
1089 }
1090 if (cnt == 0)
1091 return 0;
1092 } while (first != 0);
1093 return 1;
1094}
1095
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096static int raid10_spare_active(mddev_t *mddev)
1097{
1098 int i;
1099 conf_t *conf = mddev->private;
1100 mirror_info_t *tmp;
1101
1102 /*
1103 * Find all non-in_sync disks within the RAID10 configuration
1104 * and mark them in_sync
1105 */
1106 for (i = 0; i < conf->raid_disks; i++) {
1107 tmp = conf->mirrors + i;
1108 if (tmp->rdev
NeilBrownb2d444d2005-11-08 21:39:31 -08001109 && !test_bit(Faulty, &tmp->rdev->flags)
NeilBrownc04be0a2006-10-03 01:15:53 -07001110 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1111 unsigned long flags;
1112 spin_lock_irqsave(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 mddev->degraded--;
NeilBrownc04be0a2006-10-03 01:15:53 -07001114 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 }
1116 }
1117
1118 print_conf(conf);
1119 return 0;
1120}
1121
1122
1123static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1124{
1125 conf_t *conf = mddev->private;
Neil Brown199050e2008-06-28 08:31:33 +10001126 int err = -EEXIST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 int mirror;
1128 mirror_info_t *p;
Neil Brown6c2fce22008-06-28 08:31:31 +10001129 int first = 0;
1130 int last = mddev->raid_disks - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131
1132 if (mddev->recovery_cp < MaxSector)
1133 /* only hot-add to in-sync arrays, as recovery is
1134 * very different from resync
1135 */
Neil Brown199050e2008-06-28 08:31:33 +10001136 return -EBUSY;
NeilBrown6d508242005-09-09 16:24:03 -07001137 if (!enough(conf))
Neil Brown199050e2008-06-28 08:31:33 +10001138 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139
Neil Brown6c2fce22008-06-28 08:31:31 +10001140 if (rdev->raid_disk)
1141 first = last = rdev->raid_disk;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142
NeilBrown6cce3b22006-01-06 00:20:16 -08001143 if (rdev->saved_raid_disk >= 0 &&
Neil Brown6c2fce22008-06-28 08:31:31 +10001144 rdev->saved_raid_disk >= first &&
NeilBrown6cce3b22006-01-06 00:20:16 -08001145 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1146 mirror = rdev->saved_raid_disk;
1147 else
Neil Brown6c2fce22008-06-28 08:31:31 +10001148 mirror = first;
1149 for ( ; mirror <= last ; mirror++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150 if ( !(p=conf->mirrors+mirror)->rdev) {
1151
1152 blk_queue_stack_limits(mddev->queue,
1153 rdev->bdev->bd_disk->queue);
1154 /* as we don't honour merge_bvec_fn, we must never risk
1155 * violating it, so limit ->max_sector to one PAGE, as
1156 * a one page request is never in violation.
1157 */
1158 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1159 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1160 mddev->queue->max_sectors = (PAGE_SIZE>>9);
1161
1162 p->head_position = 0;
1163 rdev->raid_disk = mirror;
Neil Brown199050e2008-06-28 08:31:33 +10001164 err = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08001165 if (rdev->saved_raid_disk != mirror)
1166 conf->fullsync = 1;
Suzanne Woodd6065f72005-11-08 21:39:27 -08001167 rcu_assign_pointer(p->rdev, rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 break;
1169 }
1170
1171 print_conf(conf);
Neil Brown199050e2008-06-28 08:31:33 +10001172 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001173}
1174
1175static int raid10_remove_disk(mddev_t *mddev, int number)
1176{
1177 conf_t *conf = mddev->private;
1178 int err = 0;
1179 mdk_rdev_t *rdev;
1180 mirror_info_t *p = conf->mirrors+ number;
1181
1182 print_conf(conf);
1183 rdev = p->rdev;
1184 if (rdev) {
NeilBrownb2d444d2005-11-08 21:39:31 -08001185 if (test_bit(In_sync, &rdev->flags) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 atomic_read(&rdev->nr_pending)) {
1187 err = -EBUSY;
1188 goto abort;
1189 }
NeilBrowndfc70642008-05-23 13:04:39 -07001190 /* Only remove faulty devices in recovery
1191 * is not possible.
1192 */
1193 if (!test_bit(Faulty, &rdev->flags) &&
1194 enough(conf)) {
1195 err = -EBUSY;
1196 goto abort;
1197 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 p->rdev = NULL;
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07001199 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200 if (atomic_read(&rdev->nr_pending)) {
1201 /* lost the race, try later */
1202 err = -EBUSY;
1203 p->rdev = rdev;
1204 }
1205 }
1206abort:
1207
1208 print_conf(conf);
1209 return err;
1210}
1211
1212
NeilBrown6712ecf2007-09-27 12:47:43 +02001213static void end_sync_read(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1216 conf_t *conf = mddev_to_conf(r10_bio->mddev);
1217 int i,d;
1218
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 for (i=0; i<conf->copies; i++)
1220 if (r10_bio->devs[i].bio == bio)
1221 break;
Eric Sesterhennb6385482006-04-02 13:34:29 +02001222 BUG_ON(i == conf->copies);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001223 update_head_pos(i, r10_bio);
1224 d = r10_bio->devs[i].devnum;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001225
1226 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1227 set_bit(R10BIO_Uptodate, &r10_bio->state);
NeilBrown4dbcdc72006-01-06 00:20:52 -08001228 else {
1229 atomic_add(r10_bio->sectors,
1230 &conf->mirrors[d].rdev->corrected_errors);
1231 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1232 md_error(r10_bio->mddev,
1233 conf->mirrors[d].rdev);
1234 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235
1236 /* for reconstruct, we always reschedule after a read.
1237 * for resync, only after all reads
1238 */
1239 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1240 atomic_dec_and_test(&r10_bio->remaining)) {
1241 /* we have read all the blocks,
1242 * do the comparison in process context in raid10d
1243 */
1244 reschedule_retry(r10_bio);
1245 }
1246 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247}
1248
NeilBrown6712ecf2007-09-27 12:47:43 +02001249static void end_sync_write(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250{
1251 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1252 r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1253 mddev_t *mddev = r10_bio->mddev;
1254 conf_t *conf = mddev_to_conf(mddev);
1255 int i,d;
1256
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 for (i = 0; i < conf->copies; i++)
1258 if (r10_bio->devs[i].bio == bio)
1259 break;
1260 d = r10_bio->devs[i].devnum;
1261
1262 if (!uptodate)
1263 md_error(mddev, conf->mirrors[d].rdev);
NeilBrowndfc70642008-05-23 13:04:39 -07001264
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 update_head_pos(i, r10_bio);
1266
1267 while (atomic_dec_and_test(&r10_bio->remaining)) {
1268 if (r10_bio->master_bio == NULL) {
1269 /* the primary of several recovery bios */
1270 md_done_sync(mddev, r10_bio->sectors, 1);
1271 put_buf(r10_bio);
1272 break;
1273 } else {
1274 r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1275 put_buf(r10_bio);
1276 r10_bio = r10_bio2;
1277 }
1278 }
1279 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280}
1281
1282/*
1283 * Note: sync and recover and handled very differently for raid10
1284 * This code is for resync.
1285 * For resync, we read through virtual addresses and read all blocks.
1286 * If there is any error, we schedule a write. The lowest numbered
1287 * drive is authoritative.
1288 * However requests come for physical address, so we need to map.
1289 * For every physical address there are raid_disks/copies virtual addresses,
1290 * which is always are least one, but is not necessarly an integer.
1291 * This means that a physical address can span multiple chunks, so we may
1292 * have to submit multiple io requests for a single sync request.
1293 */
1294/*
1295 * We check if all blocks are in-sync and only write to blocks that
1296 * aren't in sync
1297 */
1298static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1299{
1300 conf_t *conf = mddev_to_conf(mddev);
1301 int i, first;
1302 struct bio *tbio, *fbio;
1303
1304 atomic_set(&r10_bio->remaining, 1);
1305
1306 /* find the first device with a block */
1307 for (i=0; i<conf->copies; i++)
1308 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1309 break;
1310
1311 if (i == conf->copies)
1312 goto done;
1313
1314 first = i;
1315 fbio = r10_bio->devs[i].bio;
1316
1317 /* now find blocks with errors */
NeilBrown0eb3ff12006-01-06 00:20:29 -08001318 for (i=0 ; i < conf->copies ; i++) {
1319 int j, d;
1320 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 tbio = r10_bio->devs[i].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001323
1324 if (tbio->bi_end_io != end_sync_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001325 continue;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001326 if (i == first)
1327 continue;
1328 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1329 /* We know that the bi_io_vec layout is the same for
1330 * both 'first' and 'i', so we just compare them.
1331 * All vec entries are PAGE_SIZE;
1332 */
1333 for (j = 0; j < vcnt; j++)
1334 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1335 page_address(tbio->bi_io_vec[j].bv_page),
1336 PAGE_SIZE))
1337 break;
1338 if (j == vcnt)
1339 continue;
1340 mddev->resync_mismatches += r10_bio->sectors;
1341 }
NeilBrown18f08812006-01-06 00:20:25 -08001342 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1343 /* Don't fix anything. */
1344 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 /* Ok, we need to write this bio
1346 * First we need to fixup bv_offset, bv_len and
1347 * bi_vecs, as the read request might have corrupted these
1348 */
1349 tbio->bi_vcnt = vcnt;
1350 tbio->bi_size = r10_bio->sectors << 9;
1351 tbio->bi_idx = 0;
1352 tbio->bi_phys_segments = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1354 tbio->bi_flags |= 1 << BIO_UPTODATE;
1355 tbio->bi_next = NULL;
1356 tbio->bi_rw = WRITE;
1357 tbio->bi_private = r10_bio;
1358 tbio->bi_sector = r10_bio->devs[i].addr;
1359
1360 for (j=0; j < vcnt ; j++) {
1361 tbio->bi_io_vec[j].bv_offset = 0;
1362 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1363
1364 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1365 page_address(fbio->bi_io_vec[j].bv_page),
1366 PAGE_SIZE);
1367 }
1368 tbio->bi_end_io = end_sync_write;
1369
1370 d = r10_bio->devs[i].devnum;
1371 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1372 atomic_inc(&r10_bio->remaining);
1373 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1374
1375 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1376 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1377 generic_make_request(tbio);
1378 }
1379
1380done:
1381 if (atomic_dec_and_test(&r10_bio->remaining)) {
1382 md_done_sync(mddev, r10_bio->sectors, 1);
1383 put_buf(r10_bio);
1384 }
1385}
1386
1387/*
1388 * Now for the recovery code.
1389 * Recovery happens across physical sectors.
1390 * We recover all non-is_sync drives by finding the virtual address of
1391 * each, and then choose a working drive that also has that virt address.
1392 * There is a separate r10_bio for each non-in_sync drive.
1393 * Only the first two slots are in use. The first for reading,
1394 * The second for writing.
1395 *
1396 */
1397
1398static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1399{
1400 conf_t *conf = mddev_to_conf(mddev);
1401 int i, d;
1402 struct bio *bio, *wbio;
1403
1404
1405 /* move the pages across to the second bio
1406 * and submit the write request
1407 */
1408 bio = r10_bio->devs[0].bio;
1409 wbio = r10_bio->devs[1].bio;
1410 for (i=0; i < wbio->bi_vcnt; i++) {
1411 struct page *p = bio->bi_io_vec[i].bv_page;
1412 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1413 wbio->bi_io_vec[i].bv_page = p;
1414 }
1415 d = r10_bio->devs[1].devnum;
1416
1417 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1418 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
NeilBrown0eb3ff12006-01-06 00:20:29 -08001419 if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1420 generic_make_request(wbio);
1421 else
NeilBrown6712ecf2007-09-27 12:47:43 +02001422 bio_endio(wbio, -EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423}
1424
1425
1426/*
1427 * This is a kernel thread which:
1428 *
1429 * 1. Retries failed read operations on working mirrors.
1430 * 2. Updates the raid superblock when problems encounter.
NeilBrown6814d532006-10-03 01:15:45 -07001431 * 3. Performs writes following reads for array synchronising.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 */
1433
NeilBrown6814d532006-10-03 01:15:45 -07001434static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1435{
1436 int sect = 0; /* Offset from r10_bio->sector */
1437 int sectors = r10_bio->sectors;
1438 mdk_rdev_t*rdev;
1439 while(sectors) {
1440 int s = sectors;
1441 int sl = r10_bio->read_slot;
1442 int success = 0;
1443 int start;
1444
1445 if (s > (PAGE_SIZE>>9))
1446 s = PAGE_SIZE >> 9;
1447
1448 rcu_read_lock();
1449 do {
1450 int d = r10_bio->devs[sl].devnum;
1451 rdev = rcu_dereference(conf->mirrors[d].rdev);
1452 if (rdev &&
1453 test_bit(In_sync, &rdev->flags)) {
1454 atomic_inc(&rdev->nr_pending);
1455 rcu_read_unlock();
1456 success = sync_page_io(rdev->bdev,
1457 r10_bio->devs[sl].addr +
1458 sect + rdev->data_offset,
1459 s<<9,
1460 conf->tmppage, READ);
1461 rdev_dec_pending(rdev, mddev);
1462 rcu_read_lock();
1463 if (success)
1464 break;
1465 }
1466 sl++;
1467 if (sl == conf->copies)
1468 sl = 0;
1469 } while (!success && sl != r10_bio->read_slot);
1470 rcu_read_unlock();
1471
1472 if (!success) {
1473 /* Cannot read from anywhere -- bye bye array */
1474 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1475 md_error(mddev, conf->mirrors[dn].rdev);
1476 break;
1477 }
1478
1479 start = sl;
1480 /* write it back and re-read */
1481 rcu_read_lock();
1482 while (sl != r10_bio->read_slot) {
1483 int d;
1484 if (sl==0)
1485 sl = conf->copies;
1486 sl--;
1487 d = r10_bio->devs[sl].devnum;
1488 rdev = rcu_dereference(conf->mirrors[d].rdev);
1489 if (rdev &&
1490 test_bit(In_sync, &rdev->flags)) {
1491 atomic_inc(&rdev->nr_pending);
1492 rcu_read_unlock();
1493 atomic_add(s, &rdev->corrected_errors);
1494 if (sync_page_io(rdev->bdev,
1495 r10_bio->devs[sl].addr +
1496 sect + rdev->data_offset,
1497 s<<9, conf->tmppage, WRITE)
1498 == 0)
1499 /* Well, this device is dead */
1500 md_error(mddev, rdev);
1501 rdev_dec_pending(rdev, mddev);
1502 rcu_read_lock();
1503 }
1504 }
1505 sl = start;
1506 while (sl != r10_bio->read_slot) {
1507 int d;
1508 if (sl==0)
1509 sl = conf->copies;
1510 sl--;
1511 d = r10_bio->devs[sl].devnum;
1512 rdev = rcu_dereference(conf->mirrors[d].rdev);
1513 if (rdev &&
1514 test_bit(In_sync, &rdev->flags)) {
1515 char b[BDEVNAME_SIZE];
1516 atomic_inc(&rdev->nr_pending);
1517 rcu_read_unlock();
1518 if (sync_page_io(rdev->bdev,
1519 r10_bio->devs[sl].addr +
1520 sect + rdev->data_offset,
1521 s<<9, conf->tmppage, READ) == 0)
1522 /* Well, this device is dead */
1523 md_error(mddev, rdev);
1524 else
1525 printk(KERN_INFO
1526 "raid10:%s: read error corrected"
1527 " (%d sectors at %llu on %s)\n",
1528 mdname(mddev), s,
Randy Dunlap969b7552006-10-28 10:38:32 -07001529 (unsigned long long)(sect+
1530 rdev->data_offset),
NeilBrown6814d532006-10-03 01:15:45 -07001531 bdevname(rdev->bdev, b));
1532
1533 rdev_dec_pending(rdev, mddev);
1534 rcu_read_lock();
1535 }
1536 }
1537 rcu_read_unlock();
1538
1539 sectors -= s;
1540 sect += s;
1541 }
1542}
1543
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544static void raid10d(mddev_t *mddev)
1545{
1546 r10bio_t *r10_bio;
1547 struct bio *bio;
1548 unsigned long flags;
1549 conf_t *conf = mddev_to_conf(mddev);
1550 struct list_head *head = &conf->retry_list;
1551 int unplug=0;
1552 mdk_rdev_t *rdev;
1553
1554 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555
1556 for (;;) {
1557 char b[BDEVNAME_SIZE];
NeilBrowna35e63e2008-03-04 14:29:29 -08001558
1559 unplug += flush_pending_writes(conf);
1560
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrowna35e63e2008-03-04 14:29:29 -08001562 if (list_empty(head)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001563 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564 break;
NeilBrowna35e63e2008-03-04 14:29:29 -08001565 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1567 list_del(head->prev);
NeilBrown4443ae12006-01-06 00:20:28 -08001568 conf->nr_queued--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 spin_unlock_irqrestore(&conf->device_lock, flags);
1570
1571 mddev = r10_bio->mddev;
1572 conf = mddev_to_conf(mddev);
1573 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1574 sync_request_write(mddev, r10_bio);
1575 unplug = 1;
1576 } else if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1577 recovery_request_write(mddev, r10_bio);
1578 unplug = 1;
1579 } else {
1580 int mirror;
NeilBrown4443ae12006-01-06 00:20:28 -08001581 /* we got a read error. Maybe the drive is bad. Maybe just
1582 * the block and we can fix it.
1583 * We freeze all other IO, and try reading the block from
1584 * other devices. When we find one, we re-write
1585 * and check it that fixes the read error.
1586 * This is all done synchronously while the array is
1587 * frozen.
1588 */
NeilBrown6814d532006-10-03 01:15:45 -07001589 if (mddev->ro == 0) {
1590 freeze_array(conf);
1591 fix_read_error(conf, mddev, r10_bio);
1592 unfreeze_array(conf);
NeilBrown4443ae12006-01-06 00:20:28 -08001593 }
1594
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595 bio = r10_bio->devs[r10_bio->read_slot].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -08001596 r10_bio->devs[r10_bio->read_slot].bio =
1597 mddev->ro ? IO_BLOCKED : NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598 mirror = read_balance(conf, r10_bio);
1599 if (mirror == -1) {
1600 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1601 " read error for block %llu\n",
1602 bdevname(bio->bi_bdev,b),
1603 (unsigned long long)r10_bio->sector);
1604 raid_end_bio_io(r10_bio);
Maik Hampel14e71342007-07-31 00:37:57 -07001605 bio_put(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 } else {
Lars Ellenberge3881a62007-01-10 23:15:37 -08001607 const int do_sync = bio_sync(r10_bio->master_bio);
Maik Hampel14e71342007-07-31 00:37:57 -07001608 bio_put(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609 rdev = conf->mirrors[mirror].rdev;
1610 if (printk_ratelimit())
1611 printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1612 " another mirror\n",
1613 bdevname(rdev->bdev,b),
1614 (unsigned long long)r10_bio->sector);
1615 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1616 r10_bio->devs[r10_bio->read_slot].bio = bio;
1617 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1618 + rdev->data_offset;
1619 bio->bi_bdev = rdev->bdev;
Lars Ellenberge3881a62007-01-10 23:15:37 -08001620 bio->bi_rw = READ | do_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 bio->bi_private = r10_bio;
1622 bio->bi_end_io = raid10_end_read_request;
1623 unplug = 1;
1624 generic_make_request(bio);
1625 }
1626 }
1627 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628 if (unplug)
1629 unplug_slaves(mddev);
1630}
1631
1632
1633static int init_resync(conf_t *conf)
1634{
1635 int buffs;
1636
1637 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
Eric Sesterhennb6385482006-04-02 13:34:29 +02001638 BUG_ON(conf->r10buf_pool);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1640 if (!conf->r10buf_pool)
1641 return -ENOMEM;
1642 conf->next_resync = 0;
1643 return 0;
1644}
1645
1646/*
1647 * perform a "sync" on one "block"
1648 *
1649 * We need to make sure that no normal I/O request - particularly write
1650 * requests - conflict with active sync requests.
1651 *
1652 * This is achieved by tracking pending requests and a 'barrier' concept
1653 * that can be installed to exclude normal IO requests.
1654 *
1655 * Resync and recovery are handled very differently.
1656 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1657 *
1658 * For resync, we iterate over virtual addresses, read all copies,
1659 * and update if there are differences. If only one copy is live,
1660 * skip it.
1661 * For recovery, we iterate over physical addresses, read a good
1662 * value for each non-in_sync drive, and over-write.
1663 *
1664 * So, for recovery we may have several outstanding complex requests for a
1665 * given address, one for each out-of-sync device. We model this by allocating
1666 * a number of r10_bio structures, one for each out-of-sync device.
1667 * As we setup these structures, we collect all bio's together into a list
1668 * which we then process collectively to add pages, and then process again
1669 * to pass to generic_make_request.
1670 *
1671 * The r10_bio structures are linked using a borrowed master_bio pointer.
1672 * This link is counted in ->remaining. When the r10_bio that points to NULL
1673 * has its remaining count decremented to 0, the whole complex operation
1674 * is complete.
1675 *
1676 */
1677
NeilBrown57afd892005-06-21 17:17:13 -07001678static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679{
1680 conf_t *conf = mddev_to_conf(mddev);
1681 r10bio_t *r10_bio;
1682 struct bio *biolist = NULL, *bio;
1683 sector_t max_sector, nr_sectors;
1684 int disk;
1685 int i;
NeilBrown6cce3b22006-01-06 00:20:16 -08001686 int max_sync;
1687 int sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688
1689 sector_t sectors_skipped = 0;
1690 int chunks_skipped = 0;
1691
1692 if (!conf->r10buf_pool)
1693 if (init_resync(conf))
NeilBrown57afd892005-06-21 17:17:13 -07001694 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695
1696 skipped:
1697 max_sector = mddev->size << 1;
1698 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1699 max_sector = mddev->resync_max_sectors;
1700 if (sector_nr >= max_sector) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001701 /* If we aborted, we need to abort the
1702 * sync on the 'current' bitmap chucks (there can
1703 * be several when recovering multiple devices).
1704 * as we may have started syncing it but not finished.
1705 * We can find the current address in
1706 * mddev->curr_resync, but for recovery,
1707 * we need to convert that to several
1708 * virtual addresses.
1709 */
1710 if (mddev->curr_resync < max_sector) { /* aborted */
1711 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1712 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1713 &sync_blocks, 1);
1714 else for (i=0; i<conf->raid_disks; i++) {
1715 sector_t sect =
1716 raid10_find_virt(conf, mddev->curr_resync, i);
1717 bitmap_end_sync(mddev->bitmap, sect,
1718 &sync_blocks, 1);
1719 }
1720 } else /* completed sync */
1721 conf->fullsync = 0;
1722
1723 bitmap_close_sync(mddev->bitmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 close_sync(conf);
NeilBrown57afd892005-06-21 17:17:13 -07001725 *skipped = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 return sectors_skipped;
1727 }
1728 if (chunks_skipped >= conf->raid_disks) {
1729 /* if there has been nothing to do on any drive,
1730 * then there is nothing to do at all..
1731 */
NeilBrown57afd892005-06-21 17:17:13 -07001732 *skipped = 1;
1733 return (max_sector - sector_nr) + sectors_skipped;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 }
1735
NeilBrownc6207272008-02-06 01:39:52 -08001736 if (max_sector > mddev->resync_max)
1737 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1738
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739 /* make sure whole request will fit in a chunk - if chunks
1740 * are meaningful
1741 */
1742 if (conf->near_copies < conf->raid_disks &&
1743 max_sector > (sector_nr | conf->chunk_mask))
1744 max_sector = (sector_nr | conf->chunk_mask) + 1;
1745 /*
1746 * If there is non-resync activity waiting for us then
1747 * put in a delay to throttle resync.
1748 */
NeilBrown0a27ec92006-01-06 00:20:13 -08001749 if (!go_faster && conf->nr_waiting)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 msleep_interruptible(1000);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751
NeilBrownb47490c2008-02-06 01:39:50 -08001752 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1753
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 /* Again, very different code for resync and recovery.
1755 * Both must result in an r10bio with a list of bios that
1756 * have bi_end_io, bi_sector, bi_bdev set,
1757 * and bi_private set to the r10bio.
1758 * For recovery, we may actually create several r10bios
1759 * with 2 bios in each, that correspond to the bios in the main one.
1760 * In this case, the subordinate r10bios link back through a
1761 * borrowed master_bio pointer, and the counter in the master
1762 * includes a ref from each subordinate.
1763 */
1764 /* First, we decide what to do and set ->bi_end_io
1765 * To end_sync_read if we want to read, and
1766 * end_sync_write if we will want to write.
1767 */
1768
NeilBrown6cce3b22006-01-06 00:20:16 -08001769 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1771 /* recovery... the complicated one */
1772 int i, j, k;
1773 r10_bio = NULL;
1774
1775 for (i=0 ; i<conf->raid_disks; i++)
1776 if (conf->mirrors[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001777 !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001778 int still_degraded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779 /* want to reconstruct this device */
1780 r10bio_t *rb2 = r10_bio;
NeilBrown6cce3b22006-01-06 00:20:16 -08001781 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1782 int must_sync;
1783 /* Unless we are doing a full sync, we only need
1784 * to recover the block if it is set in the bitmap
1785 */
1786 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1787 &sync_blocks, 1);
1788 if (sync_blocks < max_sync)
1789 max_sync = sync_blocks;
1790 if (!must_sync &&
1791 !conf->fullsync) {
1792 /* yep, skip the sync_blocks here, but don't assume
1793 * that there will never be anything to do here
1794 */
1795 chunks_skipped = -1;
1796 continue;
1797 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798
1799 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
NeilBrown6cce3b22006-01-06 00:20:16 -08001800 raise_barrier(conf, rb2 != NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801 atomic_set(&r10_bio->remaining, 0);
1802
1803 r10_bio->master_bio = (struct bio*)rb2;
1804 if (rb2)
1805 atomic_inc(&rb2->remaining);
1806 r10_bio->mddev = mddev;
1807 set_bit(R10BIO_IsRecover, &r10_bio->state);
NeilBrown6cce3b22006-01-06 00:20:16 -08001808 r10_bio->sector = sect;
1809
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 raid10_find_phys(conf, r10_bio);
NeilBrown6cce3b22006-01-06 00:20:16 -08001811 /* Need to check if this section will still be
1812 * degraded
1813 */
1814 for (j=0; j<conf->copies;j++) {
1815 int d = r10_bio->devs[j].devnum;
1816 if (conf->mirrors[d].rdev == NULL ||
NeilBrowna24a8dd2006-01-06 00:20:35 -08001817 test_bit(Faulty, &conf->mirrors[d].rdev->flags)) {
NeilBrown6cce3b22006-01-06 00:20:16 -08001818 still_degraded = 1;
NeilBrowna24a8dd2006-01-06 00:20:35 -08001819 break;
1820 }
NeilBrown6cce3b22006-01-06 00:20:16 -08001821 }
1822 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1823 &sync_blocks, still_degraded);
1824
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 for (j=0; j<conf->copies;j++) {
1826 int d = r10_bio->devs[j].devnum;
1827 if (conf->mirrors[d].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001828 test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 /* This is where we read from */
1830 bio = r10_bio->devs[0].bio;
1831 bio->bi_next = biolist;
1832 biolist = bio;
1833 bio->bi_private = r10_bio;
1834 bio->bi_end_io = end_sync_read;
NeilBrown802ba062006-12-13 00:34:13 -08001835 bio->bi_rw = READ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836 bio->bi_sector = r10_bio->devs[j].addr +
1837 conf->mirrors[d].rdev->data_offset;
1838 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1839 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1840 atomic_inc(&r10_bio->remaining);
1841 /* and we write to 'i' */
1842
1843 for (k=0; k<conf->copies; k++)
1844 if (r10_bio->devs[k].devnum == i)
1845 break;
NeilBrown64a742b2007-02-28 20:11:18 -08001846 BUG_ON(k == conf->copies);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 bio = r10_bio->devs[1].bio;
1848 bio->bi_next = biolist;
1849 biolist = bio;
1850 bio->bi_private = r10_bio;
1851 bio->bi_end_io = end_sync_write;
NeilBrown802ba062006-12-13 00:34:13 -08001852 bio->bi_rw = WRITE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 bio->bi_sector = r10_bio->devs[k].addr +
1854 conf->mirrors[i].rdev->data_offset;
1855 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1856
1857 r10_bio->devs[0].devnum = d;
1858 r10_bio->devs[1].devnum = i;
1859
1860 break;
1861 }
1862 }
1863 if (j == conf->copies) {
NeilBrown87fc7672005-09-09 16:24:04 -07001864 /* Cannot recover, so abort the recovery */
1865 put_buf(r10_bio);
K.Tanakaa07e6ab2008-03-04 14:29:37 -08001866 if (rb2)
1867 atomic_dec(&rb2->remaining);
NeilBrown87fc7672005-09-09 16:24:04 -07001868 r10_bio = rb2;
NeilBrowndfc70642008-05-23 13:04:39 -07001869 if (!test_and_set_bit(MD_RECOVERY_INTR,
1870 &mddev->recovery))
NeilBrown87fc7672005-09-09 16:24:04 -07001871 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1872 mdname(mddev));
1873 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001874 }
1875 }
1876 if (biolist == NULL) {
1877 while (r10_bio) {
1878 r10bio_t *rb2 = r10_bio;
1879 r10_bio = (r10bio_t*) rb2->master_bio;
1880 rb2->master_bio = NULL;
1881 put_buf(rb2);
1882 }
1883 goto giveup;
1884 }
1885 } else {
1886 /* resync. Schedule a read for every block at this virt offset */
1887 int count = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08001888
1889 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1890 &sync_blocks, mddev->degraded) &&
1891 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1892 /* We can skip this block */
1893 *skipped = 1;
1894 return sync_blocks + sectors_skipped;
1895 }
1896 if (sync_blocks < max_sync)
1897 max_sync = sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1899
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900 r10_bio->mddev = mddev;
1901 atomic_set(&r10_bio->remaining, 0);
NeilBrown6cce3b22006-01-06 00:20:16 -08001902 raise_barrier(conf, 0);
1903 conf->next_resync = sector_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904
1905 r10_bio->master_bio = NULL;
1906 r10_bio->sector = sector_nr;
1907 set_bit(R10BIO_IsSync, &r10_bio->state);
1908 raid10_find_phys(conf, r10_bio);
1909 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1910
1911 for (i=0; i<conf->copies; i++) {
1912 int d = r10_bio->devs[i].devnum;
1913 bio = r10_bio->devs[i].bio;
1914 bio->bi_end_io = NULL;
NeilBrownaf03b8e2007-06-16 10:16:06 -07001915 clear_bit(BIO_UPTODATE, &bio->bi_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916 if (conf->mirrors[d].rdev == NULL ||
NeilBrownb2d444d2005-11-08 21:39:31 -08001917 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918 continue;
1919 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1920 atomic_inc(&r10_bio->remaining);
1921 bio->bi_next = biolist;
1922 biolist = bio;
1923 bio->bi_private = r10_bio;
1924 bio->bi_end_io = end_sync_read;
NeilBrown802ba062006-12-13 00:34:13 -08001925 bio->bi_rw = READ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 bio->bi_sector = r10_bio->devs[i].addr +
1927 conf->mirrors[d].rdev->data_offset;
1928 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1929 count++;
1930 }
1931
1932 if (count < 2) {
1933 for (i=0; i<conf->copies; i++) {
1934 int d = r10_bio->devs[i].devnum;
1935 if (r10_bio->devs[i].bio->bi_end_io)
1936 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1937 }
1938 put_buf(r10_bio);
1939 biolist = NULL;
1940 goto giveup;
1941 }
1942 }
1943
1944 for (bio = biolist; bio ; bio=bio->bi_next) {
1945
1946 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1947 if (bio->bi_end_io)
1948 bio->bi_flags |= 1 << BIO_UPTODATE;
1949 bio->bi_vcnt = 0;
1950 bio->bi_idx = 0;
1951 bio->bi_phys_segments = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001952 bio->bi_size = 0;
1953 }
1954
1955 nr_sectors = 0;
NeilBrown6cce3b22006-01-06 00:20:16 -08001956 if (sector_nr + max_sync < max_sector)
1957 max_sector = sector_nr + max_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 do {
1959 struct page *page;
1960 int len = PAGE_SIZE;
1961 disk = 0;
1962 if (sector_nr + (len>>9) > max_sector)
1963 len = (max_sector - sector_nr) << 9;
1964 if (len == 0)
1965 break;
1966 for (bio= biolist ; bio ; bio=bio->bi_next) {
1967 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1968 if (bio_add_page(bio, page, len, 0) == 0) {
1969 /* stop here */
1970 struct bio *bio2;
1971 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1972 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1973 /* remove last page from this bio */
1974 bio2->bi_vcnt--;
1975 bio2->bi_size -= len;
1976 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1977 }
1978 goto bio_full;
1979 }
1980 disk = i;
1981 }
1982 nr_sectors += len>>9;
1983 sector_nr += len>>9;
1984 } while (biolist->bi_vcnt < RESYNC_PAGES);
1985 bio_full:
1986 r10_bio->sectors = nr_sectors;
1987
1988 while (biolist) {
1989 bio = biolist;
1990 biolist = biolist->bi_next;
1991
1992 bio->bi_next = NULL;
1993 r10_bio = bio->bi_private;
1994 r10_bio->sectors = nr_sectors;
1995
1996 if (bio->bi_end_io == end_sync_read) {
1997 md_sync_acct(bio->bi_bdev, nr_sectors);
1998 generic_make_request(bio);
1999 }
2000 }
2001
NeilBrown57afd892005-06-21 17:17:13 -07002002 if (sectors_skipped)
2003 /* pretend they weren't skipped, it makes
2004 * no important difference in this case
2005 */
2006 md_done_sync(mddev, sectors_skipped, 1);
2007
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008 return sectors_skipped + nr_sectors;
2009 giveup:
2010 /* There is nowhere to write, so all non-sync
2011 * drives must be failed, so try the next chunk...
2012 */
2013 {
NeilBrown57afd892005-06-21 17:17:13 -07002014 sector_t sec = max_sector - sector_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015 sectors_skipped += sec;
2016 chunks_skipped ++;
2017 sector_nr = max_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018 goto skipped;
2019 }
2020}
2021
2022static int run(mddev_t *mddev)
2023{
2024 conf_t *conf;
2025 int i, disk_idx;
2026 mirror_info_t *disk;
2027 mdk_rdev_t *rdev;
2028 struct list_head *tmp;
NeilBrownc93983b2006-06-26 00:27:41 -07002029 int nc, fc, fo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 sector_t stride, size;
2031
NeilBrown4bbf3772008-10-13 11:55:12 +11002032 if (mddev->chunk_size < PAGE_SIZE) {
2033 printk(KERN_ERR "md/raid10: chunk size must be "
2034 "at least PAGE_SIZE(%ld).\n", PAGE_SIZE);
NeilBrown2604b702006-01-06 00:20:36 -08002035 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 }
NeilBrown2604b702006-01-06 00:20:36 -08002037
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038 nc = mddev->layout & 255;
2039 fc = (mddev->layout >> 8) & 255;
NeilBrownc93983b2006-06-26 00:27:41 -07002040 fo = mddev->layout & (1<<16);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041 if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
NeilBrownc93983b2006-06-26 00:27:41 -07002042 (mddev->layout >> 17)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
2044 mdname(mddev), mddev->layout);
2045 goto out;
2046 }
2047 /*
2048 * copy the already verified devices into our private RAID10
2049 * bookkeeping area. [whatever we allocate in run(),
2050 * should be freed in stop()]
2051 */
NeilBrown4443ae12006-01-06 00:20:28 -08002052 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053 mddev->private = conf;
2054 if (!conf) {
2055 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2056 mdname(mddev));
2057 goto out;
2058 }
NeilBrown4443ae12006-01-06 00:20:28 -08002059 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060 GFP_KERNEL);
2061 if (!conf->mirrors) {
2062 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2063 mdname(mddev));
2064 goto out_free_conf;
2065 }
NeilBrown4443ae12006-01-06 00:20:28 -08002066
2067 conf->tmppage = alloc_page(GFP_KERNEL);
2068 if (!conf->tmppage)
2069 goto out_free_conf;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070
NeilBrown64a742b2007-02-28 20:11:18 -08002071 conf->mddev = mddev;
2072 conf->raid_disks = mddev->raid_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073 conf->near_copies = nc;
2074 conf->far_copies = fc;
2075 conf->copies = nc*fc;
NeilBrownc93983b2006-06-26 00:27:41 -07002076 conf->far_offset = fo;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 conf->chunk_mask = (sector_t)(mddev->chunk_size>>9)-1;
2078 conf->chunk_shift = ffz(~mddev->chunk_size) - 9;
NeilBrown64a742b2007-02-28 20:11:18 -08002079 size = mddev->size >> (conf->chunk_shift-1);
2080 sector_div(size, fc);
2081 size = size * conf->raid_disks;
2082 sector_div(size, nc);
2083 /* 'size' is now the number of chunks in the array */
2084 /* calculate "used chunks per device" in 'stride' */
2085 stride = size * conf->copies;
NeilBrownaf03b8e2007-06-16 10:16:06 -07002086
2087 /* We need to round up when dividing by raid_disks to
2088 * get the stride size.
2089 */
2090 stride += conf->raid_disks - 1;
NeilBrown64a742b2007-02-28 20:11:18 -08002091 sector_div(stride, conf->raid_disks);
2092 mddev->size = stride << (conf->chunk_shift-1);
2093
NeilBrownc93983b2006-06-26 00:27:41 -07002094 if (fo)
NeilBrown64a742b2007-02-28 20:11:18 -08002095 stride = 1;
2096 else
NeilBrownc93983b2006-06-26 00:27:41 -07002097 sector_div(stride, fc);
NeilBrown64a742b2007-02-28 20:11:18 -08002098 conf->stride = stride << conf->chunk_shift;
2099
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2101 r10bio_pool_free, conf);
2102 if (!conf->r10bio_pool) {
2103 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2104 mdname(mddev));
2105 goto out_free_conf;
2106 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107
Neil Browne7e72bf2008-05-14 16:05:54 -07002108 spin_lock_init(&conf->device_lock);
2109 mddev->queue->queue_lock = &conf->device_lock;
2110
NeilBrownd089c6a2008-02-06 01:39:59 -08002111 rdev_for_each(rdev, tmp, mddev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112 disk_idx = rdev->raid_disk;
2113 if (disk_idx >= mddev->raid_disks
2114 || disk_idx < 0)
2115 continue;
2116 disk = conf->mirrors + disk_idx;
2117
2118 disk->rdev = rdev;
2119
2120 blk_queue_stack_limits(mddev->queue,
2121 rdev->bdev->bd_disk->queue);
2122 /* as we don't honour merge_bvec_fn, we must never risk
2123 * violating it, so limit ->max_sector to one PAGE, as
2124 * a one page request is never in violation.
2125 */
2126 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2127 mddev->queue->max_sectors > (PAGE_SIZE>>9))
2128 mddev->queue->max_sectors = (PAGE_SIZE>>9);
2129
2130 disk->head_position = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132 INIT_LIST_HEAD(&conf->retry_list);
2133
2134 spin_lock_init(&conf->resync_lock);
NeilBrown0a27ec92006-01-06 00:20:13 -08002135 init_waitqueue_head(&conf->wait_barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136
NeilBrown6d508242005-09-09 16:24:03 -07002137 /* need to check that every block has at least one working mirror */
2138 if (!enough(conf)) {
2139 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2140 mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 goto out_free_conf;
2142 }
2143
2144 mddev->degraded = 0;
2145 for (i = 0; i < conf->raid_disks; i++) {
2146
2147 disk = conf->mirrors + i;
2148
NeilBrown5fd6c1d2006-06-26 00:27:40 -07002149 if (!disk->rdev ||
NeilBrown2e333e82006-10-21 10:24:07 -07002150 !test_bit(In_sync, &disk->rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151 disk->head_position = 0;
2152 mddev->degraded++;
Neil Brown8c2e8702008-06-28 08:30:52 +10002153 if (disk->rdev)
2154 conf->fullsync = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155 }
2156 }
2157
2158
2159 mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2160 if (!mddev->thread) {
2161 printk(KERN_ERR
2162 "raid10: couldn't allocate thread for %s\n",
2163 mdname(mddev));
2164 goto out_free_conf;
2165 }
2166
2167 printk(KERN_INFO
2168 "raid10: raid set %s active with %d out of %d devices\n",
2169 mdname(mddev), mddev->raid_disks - mddev->degraded,
2170 mddev->raid_disks);
2171 /*
2172 * Ok, everything is just fine now
2173 */
Andre Nollf233ea52008-07-21 17:05:22 +10002174 mddev->array_sectors = size << conf->chunk_shift;
NeilBrown64a742b2007-02-28 20:11:18 -08002175 mddev->resync_max_sectors = size << conf->chunk_shift;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176
NeilBrown7a5febe2005-05-16 21:53:16 -07002177 mddev->queue->unplug_fn = raid10_unplug;
NeilBrown0d129222006-10-03 01:15:54 -07002178 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2179 mddev->queue->backing_dev_info.congested_data = mddev;
NeilBrown7a5febe2005-05-16 21:53:16 -07002180
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181 /* Calculate max read-ahead size.
2182 * We need to readahead at least twice a whole stripe....
2183 * maybe...
2184 */
2185 {
NeilBrown8932c2e2006-06-26 00:27:36 -07002186 int stripe = conf->raid_disks * (mddev->chunk_size / PAGE_SIZE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187 stripe /= conf->near_copies;
2188 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2189 mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2190 }
2191
2192 if (conf->near_copies < mddev->raid_disks)
2193 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2194 return 0;
2195
2196out_free_conf:
2197 if (conf->r10bio_pool)
2198 mempool_destroy(conf->r10bio_pool);
NeilBrown1345b1d2006-01-06 00:20:40 -08002199 safe_put_page(conf->tmppage);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07002200 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201 kfree(conf);
2202 mddev->private = NULL;
2203out:
2204 return -EIO;
2205}
2206
2207static int stop(mddev_t *mddev)
2208{
2209 conf_t *conf = mddev_to_conf(mddev);
2210
2211 md_unregister_thread(mddev->thread);
2212 mddev->thread = NULL;
2213 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2214 if (conf->r10bio_pool)
2215 mempool_destroy(conf->r10bio_pool);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07002216 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 kfree(conf);
2218 mddev->private = NULL;
2219 return 0;
2220}
2221
NeilBrown6cce3b22006-01-06 00:20:16 -08002222static void raid10_quiesce(mddev_t *mddev, int state)
2223{
2224 conf_t *conf = mddev_to_conf(mddev);
2225
2226 switch(state) {
2227 case 1:
2228 raise_barrier(conf, 0);
2229 break;
2230 case 0:
2231 lower_barrier(conf);
2232 break;
2233 }
2234 if (mddev->thread) {
2235 if (mddev->bitmap)
2236 mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2237 else
2238 mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2239 md_wakeup_thread(mddev->thread);
2240 }
2241}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242
NeilBrown2604b702006-01-06 00:20:36 -08002243static struct mdk_personality raid10_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244{
2245 .name = "raid10",
NeilBrown2604b702006-01-06 00:20:36 -08002246 .level = 10,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247 .owner = THIS_MODULE,
2248 .make_request = make_request,
2249 .run = run,
2250 .stop = stop,
2251 .status = status,
2252 .error_handler = error,
2253 .hot_add_disk = raid10_add_disk,
2254 .hot_remove_disk= raid10_remove_disk,
2255 .spare_active = raid10_spare_active,
2256 .sync_request = sync_request,
NeilBrown6cce3b22006-01-06 00:20:16 -08002257 .quiesce = raid10_quiesce,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258};
2259
2260static int __init raid_init(void)
2261{
NeilBrown2604b702006-01-06 00:20:36 -08002262 return register_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263}
2264
2265static void raid_exit(void)
2266{
NeilBrown2604b702006-01-06 00:20:36 -08002267 unregister_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268}
2269
2270module_init(raid_init);
2271module_exit(raid_exit);
2272MODULE_LICENSE("GPL");
2273MODULE_ALIAS("md-personality-9"); /* RAID10 */
NeilBrownd9d166c2006-01-06 00:20:51 -08002274MODULE_ALIAS("md-raid10");
NeilBrown2604b702006-01-06 00:20:36 -08002275MODULE_ALIAS("md-level-10");