blob: 70b58b4bcf896b229bff7e50c18e62b671c0283f [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 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -03008 * Base on code in raid1.c. See raid1.c for further copyright information.
Linus Torvalds1da177e2005-04-16 15:20:36 -07009 *
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>
Paul Gortmaker056075c2011-07-03 13:58:33 -040024#include <linux/module.h>
NeilBrownbff61972009-03-31 14:33:13 +110025#include <linux/seq_file.h>
Christian Dietrich8bda4702011-07-27 11:00:36 +100026#include <linux/ratelimit.h>
NeilBrown3ea7daa2012-05-22 13:53:47 +100027#include <linux/kthread.h>
NeilBrown43b2e5d2009-03-31 14:33:13 +110028#include "md.h"
Christoph Hellwigef740c32009-03-31 14:27:03 +110029#include "raid10.h"
Trela, Maciejdab8b292010-03-08 16:02:45 +110030#include "raid0.h"
Christoph Hellwigef740c32009-03-31 14:27:03 +110031#include "bitmap.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070032
33/*
34 * RAID10 provides a combination of RAID0 and RAID1 functionality.
35 * The layout of data is defined by
36 * chunk_size
37 * raid_disks
38 * near_copies (stored in low byte of layout)
39 * far_copies (stored in second byte of layout)
NeilBrownc93983b2006-06-26 00:27:41 -070040 * far_offset (stored in bit 16 of layout )
Jonathan Brassow475901a2013-02-21 13:28:10 +110041 * use_far_sets (stored in bit 17 of layout )
Linus Torvalds1da177e2005-04-16 15:20:36 -070042 *
Jonathan Brassow475901a2013-02-21 13:28:10 +110043 * The data to be stored is divided into chunks using chunksize. Each device
44 * is divided into far_copies sections. In each section, chunks are laid out
45 * in a style similar to raid0, but near_copies copies of each chunk is stored
46 * (each on a different drive). The starting device for each section is offset
47 * near_copies from the starting device of the previous section. Thus there
48 * are (near_copies * far_copies) of each chunk, and each is on a different
49 * drive. near_copies and far_copies must be at least one, and their product
50 * is at most raid_disks.
NeilBrownc93983b2006-06-26 00:27:41 -070051 *
52 * If far_offset is true, then the far_copies are handled a bit differently.
Jonathan Brassow475901a2013-02-21 13:28:10 +110053 * The copies are still in different stripes, but instead of being very far
54 * apart on disk, there are adjacent stripes.
55 *
56 * The far and offset algorithms are handled slightly differently if
57 * 'use_far_sets' is true. In this case, the array's devices are grouped into
58 * sets that are (near_copies * far_copies) in size. The far copied stripes
59 * are still shifted by 'near_copies' devices, but this shifting stays confined
60 * to the set rather than the entire array. This is done to improve the number
61 * of device combinations that can fail without causing the array to fail.
62 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
63 * on a device):
64 * A B C D A B C D E
65 * ... ...
66 * D A B C E A B C D
67 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
68 * [A B] [C D] [A B] [C D E]
69 * |...| |...| |...| | ... |
70 * [B A] [D C] [B A] [E C D]
Linus Torvalds1da177e2005-04-16 15:20:36 -070071 */
72
73/*
74 * Number of guaranteed r10bios in case of extreme VM load:
75 */
76#define NR_RAID10_BIOS 256
77
Jonathan Brassow473e87c2012-07-31 10:03:52 +100078/* when we get a read error on a read-only array, we redirect to another
79 * device without failing the first device, or trying to over-write to
80 * correct the read error. To keep track of bad blocks on a per-bio
81 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
82 */
83#define IO_BLOCKED ((struct bio *)1)
84/* When we successfully write to a known bad-block, we need to remove the
85 * bad-block marking which must be done from process context. So we record
86 * the success by setting devs[n].bio to IO_MADE_GOOD
87 */
88#define IO_MADE_GOOD ((struct bio *)2)
89
90#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
91
92/* When there are this many requests queued to be written by
NeilBrown34db0cd2011-10-11 16:50:01 +110093 * the raid10 thread, we become 'congested' to provide back-pressure
94 * for writeback.
95 */
96static int max_queued_requests = 1024;
97
NeilBrowne879a872011-10-11 16:49:02 +110098static void allow_barrier(struct r10conf *conf);
99static void lower_barrier(struct r10conf *conf);
NeilBrownfae8cc5e2012-02-14 11:10:10 +1100100static int enough(struct r10conf *conf, int ignore);
NeilBrown3ea7daa2012-05-22 13:53:47 +1000101static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
102 int *skipped);
103static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
104static void end_reshape_write(struct bio *bio, int error);
105static void end_reshape(struct r10conf *conf);
NeilBrown0a27ec92006-01-06 00:20:13 -0800106
Al Virodd0fc662005-10-07 07:46:04 +0100107static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108{
NeilBrowne879a872011-10-11 16:49:02 +1100109 struct r10conf *conf = data;
NeilBrown9f2c9d12011-10-11 16:48:43 +1100110 int size = offsetof(struct r10bio, devs[conf->copies]);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111
NeilBrown69335ef2011-12-23 10:17:54 +1100112 /* allocate a r10bio with room for raid_disks entries in the
113 * bios array */
Jens Axboe7eaceac2011-03-10 08:52:07 +0100114 return kzalloc(size, gfp_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115}
116
117static void r10bio_pool_free(void *r10_bio, void *data)
118{
119 kfree(r10_bio);
120}
121
NeilBrown0310fa22008-08-05 15:54:14 +1000122/* Maximum size of each resync request */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123#define RESYNC_BLOCK_SIZE (64*1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
NeilBrown0310fa22008-08-05 15:54:14 +1000125/* amount of memory to reserve for resync requests */
126#define RESYNC_WINDOW (1024*1024)
127/* maximum number of concurrent requests, memory permitting */
128#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129
130/*
131 * When performing a resync, we need to read and compare, so
132 * we need as many pages are there are copies.
133 * When performing a recovery, we need 2 bios, one for read,
134 * one for write (we recover only one drive per r10buf)
135 *
136 */
Al Virodd0fc662005-10-07 07:46:04 +0100137static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138{
NeilBrowne879a872011-10-11 16:49:02 +1100139 struct r10conf *conf = data;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140 struct page *page;
NeilBrown9f2c9d12011-10-11 16:48:43 +1100141 struct r10bio *r10_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142 struct bio *bio;
143 int i, j;
144 int nalloc;
145
146 r10_bio = r10bio_pool_alloc(gfp_flags, conf);
Jens Axboe7eaceac2011-03-10 08:52:07 +0100147 if (!r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149
NeilBrown3ea7daa2012-05-22 13:53:47 +1000150 if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
151 test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 nalloc = conf->copies; /* resync */
153 else
154 nalloc = 2; /* recovery */
155
156 /*
157 * Allocate bios.
158 */
159 for (j = nalloc ; j-- ; ) {
NeilBrown67465572010-10-26 17:33:54 +1100160 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161 if (!bio)
162 goto out_free_bio;
163 r10_bio->devs[j].bio = bio;
NeilBrown69335ef2011-12-23 10:17:54 +1100164 if (!conf->have_replacement)
165 continue;
166 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
167 if (!bio)
168 goto out_free_bio;
169 r10_bio->devs[j].repl_bio = bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170 }
171 /*
172 * Allocate RESYNC_PAGES data pages and attach them
173 * where needed.
174 */
175 for (j = 0 ; j < nalloc; j++) {
NeilBrown69335ef2011-12-23 10:17:54 +1100176 struct bio *rbio = r10_bio->devs[j].repl_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 bio = r10_bio->devs[j].bio;
178 for (i = 0; i < RESYNC_PAGES; i++) {
NeilBrown3ea7daa2012-05-22 13:53:47 +1000179 if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
180 &conf->mddev->recovery)) {
181 /* we can share bv_page's during recovery
182 * and reshape */
Namhyung Kimc65060a2011-07-18 17:38:49 +1000183 struct bio *rbio = r10_bio->devs[0].bio;
184 page = rbio->bi_io_vec[i].bv_page;
185 get_page(page);
186 } else
187 page = alloc_page(gfp_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188 if (unlikely(!page))
189 goto out_free_pages;
190
191 bio->bi_io_vec[i].bv_page = page;
NeilBrown69335ef2011-12-23 10:17:54 +1100192 if (rbio)
193 rbio->bi_io_vec[i].bv_page = page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194 }
195 }
196
197 return r10_bio;
198
199out_free_pages:
200 for ( ; i > 0 ; i--)
NeilBrown1345b1d2006-01-06 00:20:40 -0800201 safe_put_page(bio->bi_io_vec[i-1].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202 while (j--)
203 for (i = 0; i < RESYNC_PAGES ; i++)
NeilBrown1345b1d2006-01-06 00:20:40 -0800204 safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
majianpeng5fdd2cf2012-05-22 13:55:03 +1000205 j = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206out_free_bio:
majianpeng5fdd2cf2012-05-22 13:55:03 +1000207 for ( ; j < nalloc; j++) {
208 if (r10_bio->devs[j].bio)
209 bio_put(r10_bio->devs[j].bio);
NeilBrown69335ef2011-12-23 10:17:54 +1100210 if (r10_bio->devs[j].repl_bio)
211 bio_put(r10_bio->devs[j].repl_bio);
212 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213 r10bio_pool_free(r10_bio, conf);
214 return NULL;
215}
216
217static void r10buf_pool_free(void *__r10_bio, void *data)
218{
219 int i;
NeilBrowne879a872011-10-11 16:49:02 +1100220 struct r10conf *conf = data;
NeilBrown9f2c9d12011-10-11 16:48:43 +1100221 struct r10bio *r10bio = __r10_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222 int j;
223
224 for (j=0; j < conf->copies; j++) {
225 struct bio *bio = r10bio->devs[j].bio;
226 if (bio) {
227 for (i = 0; i < RESYNC_PAGES; i++) {
NeilBrown1345b1d2006-01-06 00:20:40 -0800228 safe_put_page(bio->bi_io_vec[i].bv_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229 bio->bi_io_vec[i].bv_page = NULL;
230 }
231 bio_put(bio);
232 }
NeilBrown69335ef2011-12-23 10:17:54 +1100233 bio = r10bio->devs[j].repl_bio;
234 if (bio)
235 bio_put(bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236 }
237 r10bio_pool_free(r10bio, conf);
238}
239
NeilBrowne879a872011-10-11 16:49:02 +1100240static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700241{
242 int i;
243
244 for (i = 0; i < conf->copies; i++) {
245 struct bio **bio = & r10_bio->devs[i].bio;
NeilBrown749c55e2011-07-28 11:39:24 +1000246 if (!BIO_SPECIAL(*bio))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 bio_put(*bio);
248 *bio = NULL;
NeilBrown69335ef2011-12-23 10:17:54 +1100249 bio = &r10_bio->devs[i].repl_bio;
250 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
251 bio_put(*bio);
252 *bio = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 }
254}
255
NeilBrown9f2c9d12011-10-11 16:48:43 +1100256static void free_r10bio(struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257{
NeilBrowne879a872011-10-11 16:49:02 +1100258 struct r10conf *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260 put_all_bios(conf, r10_bio);
261 mempool_free(r10_bio, conf->r10bio_pool);
262}
263
NeilBrown9f2c9d12011-10-11 16:48:43 +1100264static void put_buf(struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265{
NeilBrowne879a872011-10-11 16:49:02 +1100266 struct r10conf *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700267
268 mempool_free(r10_bio, conf->r10buf_pool);
269
NeilBrown0a27ec92006-01-06 00:20:13 -0800270 lower_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271}
272
NeilBrown9f2c9d12011-10-11 16:48:43 +1100273static void reschedule_retry(struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274{
275 unsigned long flags;
NeilBrownfd01b882011-10-11 16:47:53 +1100276 struct mddev *mddev = r10_bio->mddev;
NeilBrowne879a872011-10-11 16:49:02 +1100277 struct r10conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278
279 spin_lock_irqsave(&conf->device_lock, flags);
280 list_add(&r10_bio->retry_list, &conf->retry_list);
NeilBrown4443ae12006-01-06 00:20:28 -0800281 conf->nr_queued ++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 spin_unlock_irqrestore(&conf->device_lock, flags);
283
Arthur Jones388667b2008-07-25 12:03:38 -0700284 /* wake up frozen array... */
285 wake_up(&conf->wait_barrier);
286
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 md_wakeup_thread(mddev->thread);
288}
289
290/*
291 * raid_end_bio_io() is called when we have finished servicing a mirrored
292 * operation and are ready to return a success/failure code to the buffer
293 * cache layer.
294 */
NeilBrown9f2c9d12011-10-11 16:48:43 +1100295static void raid_end_bio_io(struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296{
297 struct bio *bio = r10_bio->master_bio;
NeilBrown856e08e2011-07-28 11:39:23 +1000298 int done;
NeilBrowne879a872011-10-11 16:49:02 +1100299 struct r10conf *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300
NeilBrown856e08e2011-07-28 11:39:23 +1000301 if (bio->bi_phys_segments) {
302 unsigned long flags;
303 spin_lock_irqsave(&conf->device_lock, flags);
304 bio->bi_phys_segments--;
305 done = (bio->bi_phys_segments == 0);
306 spin_unlock_irqrestore(&conf->device_lock, flags);
307 } else
308 done = 1;
309 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
310 clear_bit(BIO_UPTODATE, &bio->bi_flags);
311 if (done) {
312 bio_endio(bio, 0);
313 /*
314 * Wake up any possible resync thread that waits for the device
315 * to go idle.
316 */
317 allow_barrier(conf);
318 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319 free_r10bio(r10_bio);
320}
321
322/*
323 * Update disk head position estimator based on IRQ completion info.
324 */
NeilBrown9f2c9d12011-10-11 16:48:43 +1100325static inline void update_head_pos(int slot, struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326{
NeilBrowne879a872011-10-11 16:49:02 +1100327 struct r10conf *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328
329 conf->mirrors[r10_bio->devs[slot].devnum].head_position =
330 r10_bio->devs[slot].addr + (r10_bio->sectors);
331}
332
Namhyung Kim778ca012011-07-18 17:38:47 +1000333/*
334 * Find the disk number which triggered given bio
335 */
NeilBrowne879a872011-10-11 16:49:02 +1100336static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
NeilBrown69335ef2011-12-23 10:17:54 +1100337 struct bio *bio, int *slotp, int *replp)
Namhyung Kim778ca012011-07-18 17:38:47 +1000338{
339 int slot;
NeilBrown69335ef2011-12-23 10:17:54 +1100340 int repl = 0;
Namhyung Kim778ca012011-07-18 17:38:47 +1000341
NeilBrown69335ef2011-12-23 10:17:54 +1100342 for (slot = 0; slot < conf->copies; slot++) {
Namhyung Kim778ca012011-07-18 17:38:47 +1000343 if (r10_bio->devs[slot].bio == bio)
344 break;
NeilBrown69335ef2011-12-23 10:17:54 +1100345 if (r10_bio->devs[slot].repl_bio == bio) {
346 repl = 1;
347 break;
348 }
349 }
Namhyung Kim778ca012011-07-18 17:38:47 +1000350
351 BUG_ON(slot == conf->copies);
352 update_head_pos(slot, r10_bio);
353
NeilBrown749c55e2011-07-28 11:39:24 +1000354 if (slotp)
355 *slotp = slot;
NeilBrown69335ef2011-12-23 10:17:54 +1100356 if (replp)
357 *replp = repl;
Namhyung Kim778ca012011-07-18 17:38:47 +1000358 return r10_bio->devs[slot].devnum;
359}
360
NeilBrown6712ecf2007-09-27 12:47:43 +0200361static void raid10_end_read_request(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362{
363 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
NeilBrown9f2c9d12011-10-11 16:48:43 +1100364 struct r10bio *r10_bio = bio->bi_private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365 int slot, dev;
NeilBrownabbf0982011-12-23 10:17:54 +1100366 struct md_rdev *rdev;
NeilBrowne879a872011-10-11 16:49:02 +1100367 struct r10conf *conf = r10_bio->mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369
370 slot = r10_bio->read_slot;
371 dev = r10_bio->devs[slot].devnum;
NeilBrownabbf0982011-12-23 10:17:54 +1100372 rdev = r10_bio->devs[slot].rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373 /*
374 * this branch is our 'one mirror IO has finished' event handler:
375 */
NeilBrown4443ae12006-01-06 00:20:28 -0800376 update_head_pos(slot, r10_bio);
377
378 if (uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 /*
380 * Set R10BIO_Uptodate in our master bio, so that
381 * we will return a good error code to the higher
382 * levels even if IO on some other mirrored buffer fails.
383 *
384 * The 'master' represents the composite IO operation to
385 * user-side. So if something waits for IO, then it will
386 * wait for the 'master' bio.
387 */
388 set_bit(R10BIO_Uptodate, &r10_bio->state);
NeilBrownfae8cc5e2012-02-14 11:10:10 +1100389 } else {
390 /* If all other devices that store this block have
391 * failed, we want to return the error upwards rather
392 * than fail the last device. Here we redefine
393 * "uptodate" to mean "Don't want to retry"
394 */
395 unsigned long flags;
396 spin_lock_irqsave(&conf->device_lock, flags);
397 if (!enough(conf, rdev->raid_disk))
398 uptodate = 1;
399 spin_unlock_irqrestore(&conf->device_lock, flags);
400 }
401 if (uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402 raid_end_bio_io(r10_bio);
NeilBrownabbf0982011-12-23 10:17:54 +1100403 rdev_dec_pending(rdev, conf->mddev);
NeilBrown4443ae12006-01-06 00:20:28 -0800404 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 /*
NeilBrown7c4e06f2011-05-11 14:53:17 +1000406 * oops, read error - keep the refcount on the rdev
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407 */
408 char b[BDEVNAME_SIZE];
Christian Dietrich8bda4702011-07-27 11:00:36 +1000409 printk_ratelimited(KERN_ERR
410 "md/raid10:%s: %s: rescheduling sector %llu\n",
411 mdname(conf->mddev),
NeilBrownabbf0982011-12-23 10:17:54 +1100412 bdevname(rdev->bdev, b),
Christian Dietrich8bda4702011-07-27 11:00:36 +1000413 (unsigned long long)r10_bio->sector);
NeilBrown856e08e2011-07-28 11:39:23 +1000414 set_bit(R10BIO_ReadError, &r10_bio->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 reschedule_retry(r10_bio);
416 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417}
418
NeilBrown9f2c9d12011-10-11 16:48:43 +1100419static void close_write(struct r10bio *r10_bio)
NeilBrownbd870a12011-07-28 11:39:24 +1000420{
421 /* clear the bitmap if all writes complete successfully */
422 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
423 r10_bio->sectors,
424 !test_bit(R10BIO_Degraded, &r10_bio->state),
425 0);
426 md_write_end(r10_bio->mddev);
427}
428
NeilBrown9f2c9d12011-10-11 16:48:43 +1100429static void one_write_done(struct r10bio *r10_bio)
NeilBrown19d5f832011-09-10 17:21:17 +1000430{
431 if (atomic_dec_and_test(&r10_bio->remaining)) {
432 if (test_bit(R10BIO_WriteError, &r10_bio->state))
433 reschedule_retry(r10_bio);
434 else {
435 close_write(r10_bio);
436 if (test_bit(R10BIO_MadeGood, &r10_bio->state))
437 reschedule_retry(r10_bio);
438 else
439 raid_end_bio_io(r10_bio);
440 }
441 }
442}
443
NeilBrown6712ecf2007-09-27 12:47:43 +0200444static void raid10_end_write_request(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445{
446 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
NeilBrown9f2c9d12011-10-11 16:48:43 +1100447 struct r10bio *r10_bio = bio->bi_private;
Namhyung Kim778ca012011-07-18 17:38:47 +1000448 int dev;
NeilBrown749c55e2011-07-28 11:39:24 +1000449 int dec_rdev = 1;
NeilBrowne879a872011-10-11 16:49:02 +1100450 struct r10conf *conf = r10_bio->mddev->private;
NeilBrown475b0322011-12-23 10:17:55 +1100451 int slot, repl;
NeilBrown4ca40c22011-12-23 10:17:55 +1100452 struct md_rdev *rdev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453
NeilBrown475b0322011-12-23 10:17:55 +1100454 dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455
NeilBrown475b0322011-12-23 10:17:55 +1100456 if (repl)
457 rdev = conf->mirrors[dev].replacement;
NeilBrown4ca40c22011-12-23 10:17:55 +1100458 if (!rdev) {
459 smp_rmb();
460 repl = 0;
NeilBrown475b0322011-12-23 10:17:55 +1100461 rdev = conf->mirrors[dev].rdev;
NeilBrown4ca40c22011-12-23 10:17:55 +1100462 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 /*
464 * this branch is our 'one mirror IO has finished' event handler:
465 */
NeilBrown6cce3b232006-01-06 00:20:16 -0800466 if (!uptodate) {
NeilBrown475b0322011-12-23 10:17:55 +1100467 if (repl)
468 /* Never record new bad blocks to replacement,
469 * just fail it.
470 */
471 md_error(rdev->mddev, rdev);
472 else {
473 set_bit(WriteErrorSeen, &rdev->flags);
NeilBrownb7044d42011-12-23 10:17:56 +1100474 if (!test_and_set_bit(WantReplacement, &rdev->flags))
475 set_bit(MD_RECOVERY_NEEDED,
476 &rdev->mddev->recovery);
NeilBrown475b0322011-12-23 10:17:55 +1100477 set_bit(R10BIO_WriteError, &r10_bio->state);
478 dec_rdev = 0;
479 }
NeilBrown749c55e2011-07-28 11:39:24 +1000480 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 /*
482 * Set R10BIO_Uptodate in our master bio, so that
483 * we will return a good error code for to the higher
484 * levels even if IO on some other mirrored buffer fails.
485 *
486 * The 'master' represents the composite IO operation to
487 * user-side. So if something waits for IO, then it will
488 * wait for the 'master' bio.
489 */
NeilBrown749c55e2011-07-28 11:39:24 +1000490 sector_t first_bad;
491 int bad_sectors;
492
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493 set_bit(R10BIO_Uptodate, &r10_bio->state);
494
NeilBrown749c55e2011-07-28 11:39:24 +1000495 /* Maybe we can clear some bad blocks. */
NeilBrown475b0322011-12-23 10:17:55 +1100496 if (is_badblock(rdev,
NeilBrown749c55e2011-07-28 11:39:24 +1000497 r10_bio->devs[slot].addr,
498 r10_bio->sectors,
499 &first_bad, &bad_sectors)) {
500 bio_put(bio);
NeilBrown475b0322011-12-23 10:17:55 +1100501 if (repl)
502 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
503 else
504 r10_bio->devs[slot].bio = IO_MADE_GOOD;
NeilBrown749c55e2011-07-28 11:39:24 +1000505 dec_rdev = 0;
506 set_bit(R10BIO_MadeGood, &r10_bio->state);
507 }
508 }
509
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 /*
511 *
512 * Let's see if all mirrored write operations have finished
513 * already.
514 */
NeilBrown19d5f832011-09-10 17:21:17 +1000515 one_write_done(r10_bio);
NeilBrown749c55e2011-07-28 11:39:24 +1000516 if (dec_rdev)
NeilBrown884162d2012-11-22 15:12:09 +1100517 rdev_dec_pending(rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518}
519
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520/*
521 * RAID10 layout manager
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300522 * As well as the chunksize and raid_disks count, there are two
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 * parameters: near_copies and far_copies.
524 * near_copies * far_copies must be <= raid_disks.
525 * Normally one of these will be 1.
526 * If both are 1, we get raid0.
527 * If near_copies == raid_disks, we get raid1.
528 *
Lucas De Marchi25985ed2011-03-30 22:57:33 -0300529 * Chunks are laid out in raid0 style with near_copies copies of the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530 * first chunk, followed by near_copies copies of the next chunk and
531 * so on.
532 * If far_copies > 1, then after 1/far_copies of the array has been assigned
533 * as described above, we start again with a device offset of near_copies.
534 * So we effectively have another copy of the whole array further down all
535 * the drives, but with blocks on different drives.
536 * With this layout, and block is never stored twice on the one device.
537 *
538 * raid10_find_phys finds the sector offset of a given virtual sector
NeilBrownc93983b2006-06-26 00:27:41 -0700539 * on each device that it is on.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540 *
541 * raid10_find_virt does the reverse mapping, from a device and a
542 * sector offset to a virtual address
543 */
544
NeilBrownf8c9e742012-05-21 09:28:33 +1000545static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546{
547 int n,f;
548 sector_t sector;
549 sector_t chunk;
550 sector_t stripe;
551 int dev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552 int slot = 0;
553
554 /* now calculate first sector/dev */
NeilBrown5cf00fc2012-05-21 09:28:20 +1000555 chunk = r10bio->sector >> geo->chunk_shift;
556 sector = r10bio->sector & geo->chunk_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557
NeilBrown5cf00fc2012-05-21 09:28:20 +1000558 chunk *= geo->near_copies;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 stripe = chunk;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000560 dev = sector_div(stripe, geo->raid_disks);
561 if (geo->far_offset)
562 stripe *= geo->far_copies;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563
NeilBrown5cf00fc2012-05-21 09:28:20 +1000564 sector += stripe << geo->chunk_shift;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565
566 /* and calculate all the others */
NeilBrown5cf00fc2012-05-21 09:28:20 +1000567 for (n = 0; n < geo->near_copies; n++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 int d = dev;
Jonathan Brassow475901a2013-02-21 13:28:10 +1100569 int set;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 sector_t s = sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 r10bio->devs[slot].devnum = d;
Jonathan Brassow4c0ca262013-02-21 13:28:09 +1100572 r10bio->devs[slot].addr = s;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 slot++;
574
NeilBrown5cf00fc2012-05-21 09:28:20 +1000575 for (f = 1; f < geo->far_copies; f++) {
Jonathan Brassow475901a2013-02-21 13:28:10 +1100576 set = d / geo->far_set_size;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000577 d += geo->near_copies;
Jonathan Brassow475901a2013-02-21 13:28:10 +1100578 d %= geo->far_set_size;
579 d += geo->far_set_size * set;
580
NeilBrown5cf00fc2012-05-21 09:28:20 +1000581 s += geo->stride;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 r10bio->devs[slot].devnum = d;
583 r10bio->devs[slot].addr = s;
584 slot++;
585 }
586 dev++;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000587 if (dev >= geo->raid_disks) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 dev = 0;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000589 sector += (geo->chunk_mask + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 }
591 }
NeilBrownf8c9e742012-05-21 09:28:33 +1000592}
593
594static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
595{
596 struct geom *geo = &conf->geo;
597
598 if (conf->reshape_progress != MaxSector &&
599 ((r10bio->sector >= conf->reshape_progress) !=
600 conf->mddev->reshape_backwards)) {
601 set_bit(R10BIO_Previous, &r10bio->state);
602 geo = &conf->prev;
603 } else
604 clear_bit(R10BIO_Previous, &r10bio->state);
605
606 __raid10_find_phys(geo, r10bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607}
608
NeilBrowne879a872011-10-11 16:49:02 +1100609static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610{
611 sector_t offset, chunk, vchunk;
NeilBrownf8c9e742012-05-21 09:28:33 +1000612 /* Never use conf->prev as this is only called during resync
613 * or recovery, so reshape isn't happening
614 */
NeilBrown5cf00fc2012-05-21 09:28:20 +1000615 struct geom *geo = &conf->geo;
Jonathan Brassow475901a2013-02-21 13:28:10 +1100616 int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
617 int far_set_size = geo->far_set_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618
NeilBrown5cf00fc2012-05-21 09:28:20 +1000619 offset = sector & geo->chunk_mask;
620 if (geo->far_offset) {
NeilBrownc93983b2006-06-26 00:27:41 -0700621 int fc;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000622 chunk = sector >> geo->chunk_shift;
623 fc = sector_div(chunk, geo->far_copies);
624 dev -= fc * geo->near_copies;
Jonathan Brassow475901a2013-02-21 13:28:10 +1100625 if (dev < far_set_start)
626 dev += far_set_size;
NeilBrownc93983b2006-06-26 00:27:41 -0700627 } else {
NeilBrown5cf00fc2012-05-21 09:28:20 +1000628 while (sector >= geo->stride) {
629 sector -= geo->stride;
Jonathan Brassow475901a2013-02-21 13:28:10 +1100630 if (dev < (geo->near_copies + far_set_start))
631 dev += far_set_size - geo->near_copies;
NeilBrownc93983b2006-06-26 00:27:41 -0700632 else
NeilBrown5cf00fc2012-05-21 09:28:20 +1000633 dev -= geo->near_copies;
NeilBrownc93983b2006-06-26 00:27:41 -0700634 }
NeilBrown5cf00fc2012-05-21 09:28:20 +1000635 chunk = sector >> geo->chunk_shift;
NeilBrownc93983b2006-06-26 00:27:41 -0700636 }
NeilBrown5cf00fc2012-05-21 09:28:20 +1000637 vchunk = chunk * geo->raid_disks + dev;
638 sector_div(vchunk, geo->near_copies);
639 return (vchunk << geo->chunk_shift) + offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640}
641
642/**
643 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
644 * @q: request queue
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200645 * @bvm: properties of new bio
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646 * @biovec: the request that could be merged to it.
647 *
648 * Return amount of bytes we can accept at this offset
NeilBrown050b6612012-03-19 12:46:39 +1100649 * This requires checking for end-of-chunk if near_copies != raid_disks,
650 * and for subordinate merge_bvec_fns if merge_check_needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 */
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200652static int raid10_mergeable_bvec(struct request_queue *q,
653 struct bvec_merge_data *bvm,
654 struct bio_vec *biovec)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655{
NeilBrownfd01b882011-10-11 16:47:53 +1100656 struct mddev *mddev = q->queuedata;
NeilBrown050b6612012-03-19 12:46:39 +1100657 struct r10conf *conf = mddev->private;
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200658 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 int max;
NeilBrown3ea7daa2012-05-22 13:53:47 +1000660 unsigned int chunk_sectors;
Alasdair G Kergoncc371e62008-07-03 09:53:43 +0200661 unsigned int bio_sectors = bvm->bi_size >> 9;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000662 struct geom *geo = &conf->geo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663
NeilBrown3ea7daa2012-05-22 13:53:47 +1000664 chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
NeilBrownf8c9e742012-05-21 09:28:33 +1000665 if (conf->reshape_progress != MaxSector &&
666 ((sector >= conf->reshape_progress) !=
667 conf->mddev->reshape_backwards))
668 geo = &conf->prev;
669
NeilBrown5cf00fc2012-05-21 09:28:20 +1000670 if (geo->near_copies < geo->raid_disks) {
NeilBrown050b6612012-03-19 12:46:39 +1100671 max = (chunk_sectors - ((sector & (chunk_sectors - 1))
672 + bio_sectors)) << 9;
673 if (max < 0)
674 /* bio_add cannot handle a negative return */
675 max = 0;
676 if (max <= biovec->bv_len && bio_sectors == 0)
677 return biovec->bv_len;
678 } else
679 max = biovec->bv_len;
680
681 if (mddev->merge_check_needed) {
NeilBrowne0ee7782012-08-18 09:51:42 +1000682 struct {
683 struct r10bio r10_bio;
684 struct r10dev devs[conf->copies];
685 } on_stack;
686 struct r10bio *r10_bio = &on_stack.r10_bio;
NeilBrown050b6612012-03-19 12:46:39 +1100687 int s;
NeilBrownf8c9e742012-05-21 09:28:33 +1000688 if (conf->reshape_progress != MaxSector) {
689 /* Cannot give any guidance during reshape */
690 if (max <= biovec->bv_len && bio_sectors == 0)
691 return biovec->bv_len;
692 return 0;
693 }
NeilBrowne0ee7782012-08-18 09:51:42 +1000694 r10_bio->sector = sector;
695 raid10_find_phys(conf, r10_bio);
NeilBrown050b6612012-03-19 12:46:39 +1100696 rcu_read_lock();
697 for (s = 0; s < conf->copies; s++) {
NeilBrowne0ee7782012-08-18 09:51:42 +1000698 int disk = r10_bio->devs[s].devnum;
NeilBrown050b6612012-03-19 12:46:39 +1100699 struct md_rdev *rdev = rcu_dereference(
700 conf->mirrors[disk].rdev);
701 if (rdev && !test_bit(Faulty, &rdev->flags)) {
702 struct request_queue *q =
703 bdev_get_queue(rdev->bdev);
704 if (q->merge_bvec_fn) {
NeilBrowne0ee7782012-08-18 09:51:42 +1000705 bvm->bi_sector = r10_bio->devs[s].addr
NeilBrown050b6612012-03-19 12:46:39 +1100706 + rdev->data_offset;
707 bvm->bi_bdev = rdev->bdev;
708 max = min(max, q->merge_bvec_fn(
709 q, bvm, biovec));
710 }
711 }
712 rdev = rcu_dereference(conf->mirrors[disk].replacement);
713 if (rdev && !test_bit(Faulty, &rdev->flags)) {
714 struct request_queue *q =
715 bdev_get_queue(rdev->bdev);
716 if (q->merge_bvec_fn) {
NeilBrowne0ee7782012-08-18 09:51:42 +1000717 bvm->bi_sector = r10_bio->devs[s].addr
NeilBrown050b6612012-03-19 12:46:39 +1100718 + rdev->data_offset;
719 bvm->bi_bdev = rdev->bdev;
720 max = min(max, q->merge_bvec_fn(
721 q, bvm, biovec));
722 }
723 }
724 }
725 rcu_read_unlock();
726 }
727 return max;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728}
729
730/*
731 * This routine returns the disk from which the requested read should
732 * be done. There is a per-array 'next expected sequential IO' sector
733 * number - if this matches on the next IO then we use the last disk.
734 * There is also a per-disk 'last know head position' sector that is
735 * maintained from IRQ contexts, both the normal and the resync IO
736 * completion handlers update this position correctly. If there is no
737 * perfect sequential match then we pick the disk whose head is closest.
738 *
739 * If there are 2 mirrors in the same 2 devices, performance degrades
740 * because position is mirror, not device based.
741 *
742 * The rdev for the device selected will have nr_pending incremented.
743 */
744
745/*
746 * FIXME: possibly should rethink readbalancing and do it differently
747 * depending on near_copies / far_copies geometry.
748 */
NeilBrown96c3fd12011-12-23 10:17:54 +1100749static struct md_rdev *read_balance(struct r10conf *conf,
750 struct r10bio *r10_bio,
751 int *max_sectors)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752{
NeilBrownaf3a2cd2010-05-08 08:20:17 +1000753 const sector_t this_sector = r10_bio->sector;
NeilBrown56d99122011-05-11 14:27:03 +1000754 int disk, slot;
NeilBrown856e08e2011-07-28 11:39:23 +1000755 int sectors = r10_bio->sectors;
756 int best_good_sectors;
NeilBrown56d99122011-05-11 14:27:03 +1000757 sector_t new_distance, best_dist;
Jonathan Brassow3bbae042012-07-31 10:03:52 +1000758 struct md_rdev *best_rdev, *rdev = NULL;
NeilBrown56d99122011-05-11 14:27:03 +1000759 int do_balance;
760 int best_slot;
NeilBrown5cf00fc2012-05-21 09:28:20 +1000761 struct geom *geo = &conf->geo;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762
763 raid10_find_phys(conf, r10_bio);
764 rcu_read_lock();
NeilBrown56d99122011-05-11 14:27:03 +1000765retry:
NeilBrown856e08e2011-07-28 11:39:23 +1000766 sectors = r10_bio->sectors;
NeilBrown56d99122011-05-11 14:27:03 +1000767 best_slot = -1;
NeilBrownabbf0982011-12-23 10:17:54 +1100768 best_rdev = NULL;
NeilBrown56d99122011-05-11 14:27:03 +1000769 best_dist = MaxSector;
NeilBrown856e08e2011-07-28 11:39:23 +1000770 best_good_sectors = 0;
NeilBrown56d99122011-05-11 14:27:03 +1000771 do_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 /*
773 * Check if we can balance. We can balance on the whole
NeilBrown6cce3b232006-01-06 00:20:16 -0800774 * device if no resync is going on (recovery is ok), or below
775 * the resync window. We take the first readable disk when
776 * above the resync window.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 */
778 if (conf->mddev->recovery_cp < MaxSector
NeilBrown56d99122011-05-11 14:27:03 +1000779 && (this_sector + sectors >= conf->next_resync))
780 do_balance = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781
NeilBrown56d99122011-05-11 14:27:03 +1000782 for (slot = 0; slot < conf->copies ; slot++) {
NeilBrown856e08e2011-07-28 11:39:23 +1000783 sector_t first_bad;
784 int bad_sectors;
785 sector_t dev_sector;
786
NeilBrown56d99122011-05-11 14:27:03 +1000787 if (r10_bio->devs[slot].bio == IO_BLOCKED)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 continue;
NeilBrown56d99122011-05-11 14:27:03 +1000789 disk = r10_bio->devs[slot].devnum;
NeilBrownabbf0982011-12-23 10:17:54 +1100790 rdev = rcu_dereference(conf->mirrors[disk].replacement);
791 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
NeilBrown050b6612012-03-19 12:46:39 +1100792 test_bit(Unmerged, &rdev->flags) ||
NeilBrownabbf0982011-12-23 10:17:54 +1100793 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
794 rdev = rcu_dereference(conf->mirrors[disk].rdev);
NeilBrown050b6612012-03-19 12:46:39 +1100795 if (rdev == NULL ||
796 test_bit(Faulty, &rdev->flags) ||
797 test_bit(Unmerged, &rdev->flags))
NeilBrownabbf0982011-12-23 10:17:54 +1100798 continue;
799 if (!test_bit(In_sync, &rdev->flags) &&
800 r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
NeilBrown56d99122011-05-11 14:27:03 +1000801 continue;
802
NeilBrown856e08e2011-07-28 11:39:23 +1000803 dev_sector = r10_bio->devs[slot].addr;
804 if (is_badblock(rdev, dev_sector, sectors,
805 &first_bad, &bad_sectors)) {
806 if (best_dist < MaxSector)
807 /* Already have a better slot */
808 continue;
809 if (first_bad <= dev_sector) {
810 /* Cannot read here. If this is the
811 * 'primary' device, then we must not read
812 * beyond 'bad_sectors' from another device.
813 */
814 bad_sectors -= (dev_sector - first_bad);
815 if (!do_balance && sectors > bad_sectors)
816 sectors = bad_sectors;
817 if (best_good_sectors > sectors)
818 best_good_sectors = sectors;
819 } else {
820 sector_t good_sectors =
821 first_bad - dev_sector;
822 if (good_sectors > best_good_sectors) {
823 best_good_sectors = good_sectors;
824 best_slot = slot;
NeilBrownabbf0982011-12-23 10:17:54 +1100825 best_rdev = rdev;
NeilBrown856e08e2011-07-28 11:39:23 +1000826 }
827 if (!do_balance)
828 /* Must read from here */
829 break;
830 }
831 continue;
832 } else
833 best_good_sectors = sectors;
834
NeilBrown56d99122011-05-11 14:27:03 +1000835 if (!do_balance)
836 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837
NeilBrown22dfdf52005-11-28 13:44:09 -0800838 /* This optimisation is debatable, and completely destroys
839 * sequential read speed for 'far copies' arrays. So only
840 * keep it for 'near' arrays, and review those later.
841 */
NeilBrown5cf00fc2012-05-21 09:28:20 +1000842 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 break;
Keld Simonsen8ed3a192008-03-04 14:29:34 -0800844
845 /* for far > 1 always use the lowest address */
NeilBrown5cf00fc2012-05-21 09:28:20 +1000846 if (geo->far_copies > 1)
NeilBrown56d99122011-05-11 14:27:03 +1000847 new_distance = r10_bio->devs[slot].addr;
Keld Simonsen8ed3a192008-03-04 14:29:34 -0800848 else
NeilBrown56d99122011-05-11 14:27:03 +1000849 new_distance = abs(r10_bio->devs[slot].addr -
850 conf->mirrors[disk].head_position);
851 if (new_distance < best_dist) {
852 best_dist = new_distance;
853 best_slot = slot;
NeilBrownabbf0982011-12-23 10:17:54 +1100854 best_rdev = rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 }
856 }
NeilBrownabbf0982011-12-23 10:17:54 +1100857 if (slot >= conf->copies) {
NeilBrown56d99122011-05-11 14:27:03 +1000858 slot = best_slot;
NeilBrownabbf0982011-12-23 10:17:54 +1100859 rdev = best_rdev;
860 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861
NeilBrown56d99122011-05-11 14:27:03 +1000862 if (slot >= 0) {
NeilBrown56d99122011-05-11 14:27:03 +1000863 atomic_inc(&rdev->nr_pending);
864 if (test_bit(Faulty, &rdev->flags)) {
865 /* Cannot risk returning a device that failed
866 * before we inc'ed nr_pending
867 */
868 rdev_dec_pending(rdev, conf->mddev);
869 goto retry;
870 }
871 r10_bio->read_slot = slot;
872 } else
NeilBrown96c3fd12011-12-23 10:17:54 +1100873 rdev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874 rcu_read_unlock();
NeilBrown856e08e2011-07-28 11:39:23 +1000875 *max_sectors = best_good_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876
NeilBrown96c3fd12011-12-23 10:17:54 +1100877 return rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878}
879
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +1000880int md_raid10_congested(struct mddev *mddev, int bits)
NeilBrown0d129222006-10-03 01:15:54 -0700881{
NeilBrowne879a872011-10-11 16:49:02 +1100882 struct r10conf *conf = mddev->private;
NeilBrown0d129222006-10-03 01:15:54 -0700883 int i, ret = 0;
884
NeilBrown34db0cd2011-10-11 16:50:01 +1100885 if ((bits & (1 << BDI_async_congested)) &&
886 conf->pending_count >= max_queued_requests)
887 return 1;
888
NeilBrown0d129222006-10-03 01:15:54 -0700889 rcu_read_lock();
NeilBrownf8c9e742012-05-21 09:28:33 +1000890 for (i = 0;
891 (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
892 && ret == 0;
893 i++) {
NeilBrown3cb03002011-10-11 16:45:26 +1100894 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
NeilBrown0d129222006-10-03 01:15:54 -0700895 if (rdev && !test_bit(Faulty, &rdev->flags)) {
Jens Axboe165125e2007-07-24 09:28:11 +0200896 struct request_queue *q = bdev_get_queue(rdev->bdev);
NeilBrown0d129222006-10-03 01:15:54 -0700897
898 ret |= bdi_congested(&q->backing_dev_info, bits);
899 }
900 }
901 rcu_read_unlock();
902 return ret;
903}
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +1000904EXPORT_SYMBOL_GPL(md_raid10_congested);
905
906static int raid10_congested(void *data, int bits)
907{
908 struct mddev *mddev = data;
909
910 return mddev_congested(mddev, bits) ||
911 md_raid10_congested(mddev, bits);
912}
NeilBrown0d129222006-10-03 01:15:54 -0700913
NeilBrowne879a872011-10-11 16:49:02 +1100914static void flush_pending_writes(struct r10conf *conf)
NeilBrowna35e63e2008-03-04 14:29:29 -0800915{
916 /* Any writes that have been queued but are awaiting
917 * bitmap updates get flushed here.
NeilBrowna35e63e2008-03-04 14:29:29 -0800918 */
NeilBrowna35e63e2008-03-04 14:29:29 -0800919 spin_lock_irq(&conf->device_lock);
920
921 if (conf->pending_bio_list.head) {
922 struct bio *bio;
923 bio = bio_list_get(&conf->pending_bio_list);
NeilBrown34db0cd2011-10-11 16:50:01 +1100924 conf->pending_count = 0;
NeilBrowna35e63e2008-03-04 14:29:29 -0800925 spin_unlock_irq(&conf->device_lock);
926 /* flush any pending bitmap writes to disk
927 * before proceeding w/ I/O */
928 bitmap_unplug(conf->mddev->bitmap);
NeilBrown34db0cd2011-10-11 16:50:01 +1100929 wake_up(&conf->wait_barrier);
NeilBrowna35e63e2008-03-04 14:29:29 -0800930
931 while (bio) { /* submit pending writes */
932 struct bio *next = bio->bi_next;
933 bio->bi_next = NULL;
Shaohua Li532a2a32012-10-11 13:30:52 +1100934 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
935 !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
936 /* Just ignore it */
937 bio_endio(bio, 0);
938 else
939 generic_make_request(bio);
NeilBrowna35e63e2008-03-04 14:29:29 -0800940 bio = next;
941 }
NeilBrowna35e63e2008-03-04 14:29:29 -0800942 } else
943 spin_unlock_irq(&conf->device_lock);
NeilBrowna35e63e2008-03-04 14:29:29 -0800944}
Jens Axboe7eaceac2011-03-10 08:52:07 +0100945
NeilBrown0a27ec92006-01-06 00:20:13 -0800946/* Barriers....
947 * Sometimes we need to suspend IO while we do something else,
948 * either some resync/recovery, or reconfigure the array.
949 * To do this we raise a 'barrier'.
950 * The 'barrier' is a counter that can be raised multiple times
951 * to count how many activities are happening which preclude
952 * normal IO.
953 * We can only raise the barrier if there is no pending IO.
954 * i.e. if nr_pending == 0.
955 * We choose only to raise the barrier if no-one is waiting for the
956 * barrier to go down. This means that as soon as an IO request
957 * is ready, no other operations which require a barrier will start
958 * until the IO request has had a chance.
959 *
960 * So: regular IO calls 'wait_barrier'. When that returns there
961 * is no backgroup IO happening, It must arrange to call
962 * allow_barrier when it has finished its IO.
963 * backgroup IO calls must call raise_barrier. Once that returns
964 * there is no normal IO happeing. It must arrange to call
965 * lower_barrier when the particular background IO completes.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700966 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967
NeilBrowne879a872011-10-11 16:49:02 +1100968static void raise_barrier(struct r10conf *conf, int force)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969{
NeilBrown6cce3b232006-01-06 00:20:16 -0800970 BUG_ON(force && !conf->barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 spin_lock_irq(&conf->resync_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972
NeilBrown6cce3b232006-01-06 00:20:16 -0800973 /* Wait until no block IO is waiting (unless 'force') */
974 wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
Lukas Czernereed8c022012-11-30 11:42:40 +0100975 conf->resync_lock);
NeilBrown0a27ec92006-01-06 00:20:13 -0800976
977 /* block any new IO from starting */
978 conf->barrier++;
979
NeilBrownc3b328a2011-04-18 18:25:43 +1000980 /* Now wait for all pending IO to complete */
NeilBrown0a27ec92006-01-06 00:20:13 -0800981 wait_event_lock_irq(conf->wait_barrier,
982 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
Lukas Czernereed8c022012-11-30 11:42:40 +0100983 conf->resync_lock);
NeilBrown0a27ec92006-01-06 00:20:13 -0800984
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 spin_unlock_irq(&conf->resync_lock);
986}
987
NeilBrowne879a872011-10-11 16:49:02 +1100988static void lower_barrier(struct r10conf *conf)
NeilBrown0a27ec92006-01-06 00:20:13 -0800989{
990 unsigned long flags;
991 spin_lock_irqsave(&conf->resync_lock, flags);
992 conf->barrier--;
993 spin_unlock_irqrestore(&conf->resync_lock, flags);
994 wake_up(&conf->wait_barrier);
995}
996
NeilBrowne879a872011-10-11 16:49:02 +1100997static void wait_barrier(struct r10conf *conf)
NeilBrown0a27ec92006-01-06 00:20:13 -0800998{
999 spin_lock_irq(&conf->resync_lock);
1000 if (conf->barrier) {
1001 conf->nr_waiting++;
NeilBrownd6b42dc2012-03-19 12:46:38 +11001002 /* Wait for the barrier to drop.
1003 * However if there are already pending
1004 * requests (preventing the barrier from
1005 * rising completely), and the
1006 * pre-process bio queue isn't empty,
1007 * then don't wait, as we need to empty
1008 * that queue to get the nr_pending
1009 * count down.
1010 */
1011 wait_event_lock_irq(conf->wait_barrier,
1012 !conf->barrier ||
1013 (conf->nr_pending &&
1014 current->bio_list &&
1015 !bio_list_empty(current->bio_list)),
Lukas Czernereed8c022012-11-30 11:42:40 +01001016 conf->resync_lock);
NeilBrown0a27ec92006-01-06 00:20:13 -08001017 conf->nr_waiting--;
1018 }
1019 conf->nr_pending++;
1020 spin_unlock_irq(&conf->resync_lock);
1021}
1022
NeilBrowne879a872011-10-11 16:49:02 +11001023static void allow_barrier(struct r10conf *conf)
NeilBrown0a27ec92006-01-06 00:20:13 -08001024{
1025 unsigned long flags;
1026 spin_lock_irqsave(&conf->resync_lock, flags);
1027 conf->nr_pending--;
1028 spin_unlock_irqrestore(&conf->resync_lock, flags);
1029 wake_up(&conf->wait_barrier);
1030}
1031
NeilBrowne879a872011-10-11 16:49:02 +11001032static void freeze_array(struct r10conf *conf)
NeilBrown4443ae12006-01-06 00:20:28 -08001033{
1034 /* stop syncio and normal IO and wait for everything to
NeilBrownf1885932006-01-06 00:20:42 -08001035 * go quiet.
NeilBrown4443ae12006-01-06 00:20:28 -08001036 * We increment barrier and nr_waiting, and then
NeilBrown1c830532008-03-04 14:29:35 -08001037 * wait until nr_pending match nr_queued+1
1038 * This is called in the context of one normal IO request
1039 * that has failed. Thus any sync request that might be pending
1040 * will be blocked by nr_pending, and we need to wait for
1041 * pending IO requests to complete or be queued for re-try.
1042 * Thus the number queued (nr_queued) plus this request (1)
1043 * must match the number of pending IOs (nr_pending) before
1044 * we continue.
NeilBrown4443ae12006-01-06 00:20:28 -08001045 */
1046 spin_lock_irq(&conf->resync_lock);
1047 conf->barrier++;
1048 conf->nr_waiting++;
Lukas Czernereed8c022012-11-30 11:42:40 +01001049 wait_event_lock_irq_cmd(conf->wait_barrier,
1050 conf->nr_pending == conf->nr_queued+1,
1051 conf->resync_lock,
1052 flush_pending_writes(conf));
NeilBrownc3b328a2011-04-18 18:25:43 +10001053
NeilBrown4443ae12006-01-06 00:20:28 -08001054 spin_unlock_irq(&conf->resync_lock);
1055}
1056
NeilBrowne879a872011-10-11 16:49:02 +11001057static void unfreeze_array(struct r10conf *conf)
NeilBrown4443ae12006-01-06 00:20:28 -08001058{
1059 /* reverse the effect of the freeze */
1060 spin_lock_irq(&conf->resync_lock);
1061 conf->barrier--;
1062 conf->nr_waiting--;
1063 wake_up(&conf->wait_barrier);
1064 spin_unlock_irq(&conf->resync_lock);
1065}
1066
NeilBrownf8c9e742012-05-21 09:28:33 +10001067static sector_t choose_data_offset(struct r10bio *r10_bio,
1068 struct md_rdev *rdev)
1069{
1070 if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1071 test_bit(R10BIO_Previous, &r10_bio->state))
1072 return rdev->data_offset;
1073 else
1074 return rdev->new_data_offset;
1075}
1076
NeilBrown57c67df2012-10-11 13:32:13 +11001077struct raid10_plug_cb {
1078 struct blk_plug_cb cb;
1079 struct bio_list pending;
1080 int pending_cnt;
1081};
1082
1083static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1084{
1085 struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1086 cb);
1087 struct mddev *mddev = plug->cb.data;
1088 struct r10conf *conf = mddev->private;
1089 struct bio *bio;
1090
NeilBrown874807a2012-11-27 12:14:40 +11001091 if (from_schedule || current->bio_list) {
NeilBrown57c67df2012-10-11 13:32:13 +11001092 spin_lock_irq(&conf->device_lock);
1093 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1094 conf->pending_count += plug->pending_cnt;
1095 spin_unlock_irq(&conf->device_lock);
1096 md_wakeup_thread(mddev->thread);
1097 kfree(plug);
1098 return;
1099 }
1100
1101 /* we aren't scheduling, so we can do the write-out directly. */
1102 bio = bio_list_get(&plug->pending);
1103 bitmap_unplug(mddev->bitmap);
1104 wake_up(&conf->wait_barrier);
1105
1106 while (bio) { /* submit pending writes */
1107 struct bio *next = bio->bi_next;
1108 bio->bi_next = NULL;
1109 generic_make_request(bio);
1110 bio = next;
1111 }
1112 kfree(plug);
1113}
1114
Linus Torvaldsb4fdcb02011-11-04 17:06:58 -07001115static void make_request(struct mddev *mddev, struct bio * bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116{
NeilBrowne879a872011-10-11 16:49:02 +11001117 struct r10conf *conf = mddev->private;
NeilBrown9f2c9d12011-10-11 16:48:43 +11001118 struct r10bio *r10_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 struct bio *read_bio;
1120 int i;
NeilBrownf8c9e742012-05-21 09:28:33 +10001121 sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
NeilBrown5cf00fc2012-05-21 09:28:20 +10001122 int chunk_sects = chunk_mask + 1;
Jens Axboea3623572005-11-01 09:26:16 +01001123 const int rw = bio_data_dir(bio);
NeilBrown2c7d46e2010-08-18 16:16:05 +10001124 const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
Tejun Heoe9c74692010-09-03 11:56:18 +02001125 const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
Shaohua Li532a2a32012-10-11 13:30:52 +11001126 const unsigned long do_discard = (bio->bi_rw
1127 & (REQ_DISCARD | REQ_SECURE));
Joe Lawrencec8dc9c62013-02-21 13:28:09 +11001128 const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
NeilBrown6cce3b232006-01-06 00:20:16 -08001129 unsigned long flags;
NeilBrown3cb03002011-10-11 16:45:26 +11001130 struct md_rdev *blocked_rdev;
NeilBrown57c67df2012-10-11 13:32:13 +11001131 struct blk_plug_cb *cb;
1132 struct raid10_plug_cb *plug = NULL;
NeilBrownd4432c22011-07-28 11:39:24 +10001133 int sectors_handled;
1134 int max_sectors;
NeilBrown3ea7daa2012-05-22 13:53:47 +10001135 int sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136
Tejun Heoe9c74692010-09-03 11:56:18 +02001137 if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1138 md_flush_request(mddev, bio);
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +02001139 return;
NeilBrowne5dcdd82005-09-09 16:23:41 -07001140 }
1141
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 /* If this request crosses a chunk boundary, we need to
1143 * split it. This will only happen for 1 PAGE (or less) requests.
1144 */
NeilBrown5cf00fc2012-05-21 09:28:20 +10001145 if (unlikely((bio->bi_sector & chunk_mask) + (bio->bi_size >> 9)
1146 > chunk_sects
NeilBrownf8c9e742012-05-21 09:28:33 +10001147 && (conf->geo.near_copies < conf->geo.raid_disks
1148 || conf->prev.near_copies < conf->prev.raid_disks))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 struct bio_pair *bp;
1150 /* Sanity check -- queue functions should prevent this happening */
Shaohua Li532a2a32012-10-11 13:30:52 +11001151 if ((bio->bi_vcnt != 1 && bio->bi_vcnt != 0) ||
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 bio->bi_idx != 0)
1153 goto bad_map;
1154 /* This is a one page bio that upper layers
1155 * refuse to split for us, so we need to split it.
1156 */
Denis ChengRq6feef532008-10-09 08:57:05 +02001157 bp = bio_split(bio,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158 chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
NeilBrown51e9ac72010-08-07 21:17:00 +10001159
1160 /* Each of these 'make_request' calls will call 'wait_barrier'.
1161 * If the first succeeds but the second blocks due to the resync
1162 * thread raising the barrier, we will deadlock because the
1163 * IO to the underlying device will be queued in generic_make_request
1164 * and will never complete, so will never reduce nr_pending.
1165 * So increment nr_waiting here so no new raise_barriers will
1166 * succeed, and so the second wait_barrier cannot block.
1167 */
1168 spin_lock_irq(&conf->resync_lock);
1169 conf->nr_waiting++;
1170 spin_unlock_irq(&conf->resync_lock);
1171
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +02001172 make_request(mddev, &bp->bio1);
1173 make_request(mddev, &bp->bio2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174
NeilBrown51e9ac72010-08-07 21:17:00 +10001175 spin_lock_irq(&conf->resync_lock);
1176 conf->nr_waiting--;
1177 wake_up(&conf->wait_barrier);
1178 spin_unlock_irq(&conf->resync_lock);
1179
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 bio_pair_release(bp);
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +02001181 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182 bad_map:
NeilBrown128595e2010-05-03 14:47:14 +10001183 printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1184 " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185 (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
1186
NeilBrown6712ecf2007-09-27 12:47:43 +02001187 bio_io_error(bio);
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +02001188 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001189 }
1190
NeilBrown3d310eb2005-06-21 17:17:26 -07001191 md_write_start(mddev, bio);
NeilBrown06d91a52005-06-21 17:17:12 -07001192
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 /*
1194 * Register the new request and wait if the reconstruction
1195 * thread has put up a bar for new requests.
1196 * Continue immediately if no resync is active currently.
1197 */
NeilBrown0a27ec92006-01-06 00:20:13 -08001198 wait_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199
NeilBrown3ea7daa2012-05-22 13:53:47 +10001200 sectors = bio->bi_size >> 9;
1201 while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1202 bio->bi_sector < conf->reshape_progress &&
1203 bio->bi_sector + sectors > conf->reshape_progress) {
1204 /* IO spans the reshape position. Need to wait for
1205 * reshape to pass
1206 */
1207 allow_barrier(conf);
1208 wait_event(conf->wait_barrier,
1209 conf->reshape_progress <= bio->bi_sector ||
1210 conf->reshape_progress >= bio->bi_sector + sectors);
1211 wait_barrier(conf);
1212 }
1213 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1214 bio_data_dir(bio) == WRITE &&
1215 (mddev->reshape_backwards
1216 ? (bio->bi_sector < conf->reshape_safe &&
1217 bio->bi_sector + sectors > conf->reshape_progress)
1218 : (bio->bi_sector + sectors > conf->reshape_safe &&
1219 bio->bi_sector < conf->reshape_progress))) {
1220 /* Need to update reshape_position in metadata */
1221 mddev->reshape_position = conf->reshape_progress;
1222 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1223 set_bit(MD_CHANGE_PENDING, &mddev->flags);
1224 md_wakeup_thread(mddev->thread);
1225 wait_event(mddev->sb_wait,
1226 !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1227
1228 conf->reshape_safe = mddev->reshape_position;
1229 }
1230
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1232
1233 r10_bio->master_bio = bio;
NeilBrown3ea7daa2012-05-22 13:53:47 +10001234 r10_bio->sectors = sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235
1236 r10_bio->mddev = mddev;
1237 r10_bio->sector = bio->bi_sector;
NeilBrown6cce3b232006-01-06 00:20:16 -08001238 r10_bio->state = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239
NeilBrown856e08e2011-07-28 11:39:23 +10001240 /* We might need to issue multiple reads to different
1241 * devices if there are bad blocks around, so we keep
1242 * track of the number of reads in bio->bi_phys_segments.
1243 * If this is 0, there is only one r10_bio and no locking
1244 * will be needed when the request completes. If it is
1245 * non-zero, then it is the number of not-completed requests.
1246 */
1247 bio->bi_phys_segments = 0;
1248 clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1249
Jens Axboea3623572005-11-01 09:26:16 +01001250 if (rw == READ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 /*
1252 * read balancing logic:
1253 */
NeilBrown96c3fd12011-12-23 10:17:54 +11001254 struct md_rdev *rdev;
NeilBrown856e08e2011-07-28 11:39:23 +10001255 int slot;
1256
1257read_again:
NeilBrown96c3fd12011-12-23 10:17:54 +11001258 rdev = read_balance(conf, r10_bio, &max_sectors);
1259 if (!rdev) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001260 raid_end_bio_io(r10_bio);
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +02001261 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262 }
NeilBrown96c3fd12011-12-23 10:17:54 +11001263 slot = r10_bio->read_slot;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264
NeilBrowna167f662010-10-26 18:31:13 +11001265 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
NeilBrown856e08e2011-07-28 11:39:23 +10001266 md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1267 max_sectors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268
1269 r10_bio->devs[slot].bio = read_bio;
NeilBrownabbf0982011-12-23 10:17:54 +11001270 r10_bio->devs[slot].rdev = rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001271
1272 read_bio->bi_sector = r10_bio->devs[slot].addr +
NeilBrownf8c9e742012-05-21 09:28:33 +10001273 choose_data_offset(r10_bio, rdev);
NeilBrown96c3fd12011-12-23 10:17:54 +11001274 read_bio->bi_bdev = rdev->bdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 read_bio->bi_end_io = raid10_end_read_request;
Christoph Hellwig7b6d91d2010-08-07 18:20:39 +02001276 read_bio->bi_rw = READ | do_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 read_bio->bi_private = r10_bio;
1278
NeilBrown856e08e2011-07-28 11:39:23 +10001279 if (max_sectors < r10_bio->sectors) {
1280 /* Could not read all from this device, so we will
1281 * need another r10_bio.
1282 */
NeilBrown856e08e2011-07-28 11:39:23 +10001283 sectors_handled = (r10_bio->sectors + max_sectors
1284 - bio->bi_sector);
1285 r10_bio->sectors = max_sectors;
1286 spin_lock_irq(&conf->device_lock);
1287 if (bio->bi_phys_segments == 0)
1288 bio->bi_phys_segments = 2;
1289 else
1290 bio->bi_phys_segments++;
1291 spin_unlock(&conf->device_lock);
1292 /* Cannot call generic_make_request directly
1293 * as that will be queued in __generic_make_request
1294 * and subsequent mempool_alloc might block
1295 * waiting for it. so hand bio over to raid10d.
1296 */
1297 reschedule_retry(r10_bio);
1298
1299 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1300
1301 r10_bio->master_bio = bio;
1302 r10_bio->sectors = ((bio->bi_size >> 9)
1303 - sectors_handled);
1304 r10_bio->state = 0;
1305 r10_bio->mddev = mddev;
1306 r10_bio->sector = bio->bi_sector + sectors_handled;
1307 goto read_again;
1308 } else
1309 generic_make_request(read_bio);
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +02001310 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311 }
1312
1313 /*
1314 * WRITE:
1315 */
NeilBrown34db0cd2011-10-11 16:50:01 +11001316 if (conf->pending_count >= max_queued_requests) {
1317 md_wakeup_thread(mddev->thread);
1318 wait_event(conf->wait_barrier,
1319 conf->pending_count < max_queued_requests);
1320 }
Dan Williams6bfe0b42008-04-30 00:52:32 -07001321 /* first select target devices under rcu_lock and
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 * inc refcount on their rdev. Record them by setting
1323 * bios[x] to bio
NeilBrownd4432c22011-07-28 11:39:24 +10001324 * If there are known/acknowledged bad blocks on any device
1325 * on which we have seen a write error, we want to avoid
1326 * writing to those blocks. This potentially requires several
1327 * writes to write around the bad blocks. Each set of writes
1328 * gets its own r10_bio with a set of bios attached. The number
1329 * of r10_bios is recored in bio->bi_phys_segments just as with
1330 * the read case.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331 */
NeilBrownc3b328a2011-04-18 18:25:43 +10001332
NeilBrown69335ef2011-12-23 10:17:54 +11001333 r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 raid10_find_phys(conf, r10_bio);
NeilBrownd4432c22011-07-28 11:39:24 +10001335retry_write:
Harvey Harrisoncb6969e2008-05-06 20:42:32 -07001336 blocked_rdev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 rcu_read_lock();
NeilBrownd4432c22011-07-28 11:39:24 +10001338 max_sectors = r10_bio->sectors;
1339
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 for (i = 0; i < conf->copies; i++) {
1341 int d = r10_bio->devs[i].devnum;
NeilBrown3cb03002011-10-11 16:45:26 +11001342 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
NeilBrown475b0322011-12-23 10:17:55 +11001343 struct md_rdev *rrdev = rcu_dereference(
1344 conf->mirrors[d].replacement);
NeilBrown4ca40c22011-12-23 10:17:55 +11001345 if (rdev == rrdev)
1346 rrdev = NULL;
Dan Williams6bfe0b42008-04-30 00:52:32 -07001347 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1348 atomic_inc(&rdev->nr_pending);
1349 blocked_rdev = rdev;
1350 break;
1351 }
NeilBrown475b0322011-12-23 10:17:55 +11001352 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1353 atomic_inc(&rrdev->nr_pending);
1354 blocked_rdev = rrdev;
1355 break;
1356 }
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001357 if (rdev && (test_bit(Faulty, &rdev->flags)
1358 || test_bit(Unmerged, &rdev->flags)))
1359 rdev = NULL;
NeilBrown050b6612012-03-19 12:46:39 +11001360 if (rrdev && (test_bit(Faulty, &rrdev->flags)
1361 || test_bit(Unmerged, &rrdev->flags)))
NeilBrown475b0322011-12-23 10:17:55 +11001362 rrdev = NULL;
1363
NeilBrownd4432c22011-07-28 11:39:24 +10001364 r10_bio->devs[i].bio = NULL;
NeilBrown475b0322011-12-23 10:17:55 +11001365 r10_bio->devs[i].repl_bio = NULL;
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001366
1367 if (!rdev && !rrdev) {
NeilBrown6cce3b232006-01-06 00:20:16 -08001368 set_bit(R10BIO_Degraded, &r10_bio->state);
NeilBrownd4432c22011-07-28 11:39:24 +10001369 continue;
NeilBrown6cce3b232006-01-06 00:20:16 -08001370 }
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001371 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
NeilBrownd4432c22011-07-28 11:39:24 +10001372 sector_t first_bad;
1373 sector_t dev_sector = r10_bio->devs[i].addr;
1374 int bad_sectors;
1375 int is_bad;
1376
1377 is_bad = is_badblock(rdev, dev_sector,
1378 max_sectors,
1379 &first_bad, &bad_sectors);
1380 if (is_bad < 0) {
1381 /* Mustn't write here until the bad block
1382 * is acknowledged
1383 */
1384 atomic_inc(&rdev->nr_pending);
1385 set_bit(BlockedBadBlocks, &rdev->flags);
1386 blocked_rdev = rdev;
1387 break;
1388 }
1389 if (is_bad && first_bad <= dev_sector) {
1390 /* Cannot write here at all */
1391 bad_sectors -= (dev_sector - first_bad);
1392 if (bad_sectors < max_sectors)
1393 /* Mustn't write more than bad_sectors
1394 * to other devices yet
1395 */
1396 max_sectors = bad_sectors;
1397 /* We don't set R10BIO_Degraded as that
1398 * only applies if the disk is missing,
1399 * so it might be re-added, and we want to
1400 * know to recover this chunk.
1401 * In this case the device is here, and the
1402 * fact that this chunk is not in-sync is
1403 * recorded in the bad block log.
1404 */
1405 continue;
1406 }
1407 if (is_bad) {
1408 int good_sectors = first_bad - dev_sector;
1409 if (good_sectors < max_sectors)
1410 max_sectors = good_sectors;
1411 }
1412 }
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001413 if (rdev) {
1414 r10_bio->devs[i].bio = bio;
1415 atomic_inc(&rdev->nr_pending);
1416 }
NeilBrown475b0322011-12-23 10:17:55 +11001417 if (rrdev) {
1418 r10_bio->devs[i].repl_bio = bio;
1419 atomic_inc(&rrdev->nr_pending);
1420 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 }
1422 rcu_read_unlock();
1423
Dan Williams6bfe0b42008-04-30 00:52:32 -07001424 if (unlikely(blocked_rdev)) {
1425 /* Have to wait for this device to get unblocked, then retry */
1426 int j;
1427 int d;
1428
NeilBrown475b0322011-12-23 10:17:55 +11001429 for (j = 0; j < i; j++) {
Dan Williams6bfe0b42008-04-30 00:52:32 -07001430 if (r10_bio->devs[j].bio) {
1431 d = r10_bio->devs[j].devnum;
1432 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1433 }
NeilBrown475b0322011-12-23 10:17:55 +11001434 if (r10_bio->devs[j].repl_bio) {
NeilBrown4ca40c22011-12-23 10:17:55 +11001435 struct md_rdev *rdev;
NeilBrown475b0322011-12-23 10:17:55 +11001436 d = r10_bio->devs[j].devnum;
NeilBrown4ca40c22011-12-23 10:17:55 +11001437 rdev = conf->mirrors[d].replacement;
1438 if (!rdev) {
1439 /* Race with remove_disk */
1440 smp_mb();
1441 rdev = conf->mirrors[d].rdev;
1442 }
1443 rdev_dec_pending(rdev, mddev);
NeilBrown475b0322011-12-23 10:17:55 +11001444 }
1445 }
Dan Williams6bfe0b42008-04-30 00:52:32 -07001446 allow_barrier(conf);
1447 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1448 wait_barrier(conf);
1449 goto retry_write;
1450 }
1451
NeilBrownd4432c22011-07-28 11:39:24 +10001452 if (max_sectors < r10_bio->sectors) {
1453 /* We are splitting this into multiple parts, so
1454 * we need to prepare for allocating another r10_bio.
1455 */
1456 r10_bio->sectors = max_sectors;
1457 spin_lock_irq(&conf->device_lock);
1458 if (bio->bi_phys_segments == 0)
1459 bio->bi_phys_segments = 2;
1460 else
1461 bio->bi_phys_segments++;
1462 spin_unlock_irq(&conf->device_lock);
1463 }
1464 sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1465
NeilBrown4e780642010-10-19 12:54:01 +11001466 atomic_set(&r10_bio->remaining, 1);
NeilBrownd4432c22011-07-28 11:39:24 +10001467 bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
NeilBrown06d91a52005-06-21 17:17:12 -07001468
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 for (i = 0; i < conf->copies; i++) {
1470 struct bio *mbio;
1471 int d = r10_bio->devs[i].devnum;
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001472 if (r10_bio->devs[i].bio) {
1473 struct md_rdev *rdev = conf->mirrors[d].rdev;
1474 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1475 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1476 max_sectors);
1477 r10_bio->devs[i].bio = mbio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001479 mbio->bi_sector = (r10_bio->devs[i].addr+
1480 choose_data_offset(r10_bio,
1481 rdev));
1482 mbio->bi_bdev = rdev->bdev;
1483 mbio->bi_end_io = raid10_end_write_request;
Joe Lawrencec8dc9c62013-02-21 13:28:09 +11001484 mbio->bi_rw =
1485 WRITE | do_sync | do_fua | do_discard | do_same;
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001486 mbio->bi_private = r10_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001488 atomic_inc(&r10_bio->remaining);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001490 cb = blk_check_plugged(raid10_unplug, mddev,
1491 sizeof(*plug));
1492 if (cb)
1493 plug = container_of(cb, struct raid10_plug_cb,
1494 cb);
1495 else
1496 plug = NULL;
1497 spin_lock_irqsave(&conf->device_lock, flags);
1498 if (plug) {
1499 bio_list_add(&plug->pending, mbio);
1500 plug->pending_cnt++;
1501 } else {
1502 bio_list_add(&conf->pending_bio_list, mbio);
1503 conf->pending_count++;
1504 }
1505 spin_unlock_irqrestore(&conf->device_lock, flags);
1506 if (!plug)
1507 md_wakeup_thread(mddev->thread);
1508 }
NeilBrown57c67df2012-10-11 13:32:13 +11001509
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001510 if (r10_bio->devs[i].repl_bio) {
1511 struct md_rdev *rdev = conf->mirrors[d].replacement;
1512 if (rdev == NULL) {
1513 /* Replacement just got moved to main 'rdev' */
1514 smp_mb();
1515 rdev = conf->mirrors[d].rdev;
1516 }
1517 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1518 md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1519 max_sectors);
1520 r10_bio->devs[i].repl_bio = mbio;
1521
1522 mbio->bi_sector = (r10_bio->devs[i].addr +
1523 choose_data_offset(
1524 r10_bio, rdev));
1525 mbio->bi_bdev = rdev->bdev;
1526 mbio->bi_end_io = raid10_end_write_request;
Joe Lawrencec8dc9c62013-02-21 13:28:09 +11001527 mbio->bi_rw =
1528 WRITE | do_sync | do_fua | do_discard | do_same;
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001529 mbio->bi_private = r10_bio;
1530
1531 atomic_inc(&r10_bio->remaining);
1532 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrown57c67df2012-10-11 13:32:13 +11001533 bio_list_add(&conf->pending_bio_list, mbio);
1534 conf->pending_count++;
NeilBrowne7c0c3f2012-11-22 14:42:49 +11001535 spin_unlock_irqrestore(&conf->device_lock, flags);
1536 if (!mddev_check_plugged(mddev))
1537 md_wakeup_thread(mddev->thread);
NeilBrown57c67df2012-10-11 13:32:13 +11001538 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 }
1540
NeilBrown079fa162011-09-10 17:21:23 +10001541 /* Don't remove the bias on 'remaining' (one_write_done) until
1542 * after checking if we need to go around again.
1543 */
NeilBrowna35e63e2008-03-04 14:29:29 -08001544
NeilBrownd4432c22011-07-28 11:39:24 +10001545 if (sectors_handled < (bio->bi_size >> 9)) {
NeilBrown079fa162011-09-10 17:21:23 +10001546 one_write_done(r10_bio);
NeilBrown5e570282011-07-28 11:39:25 +10001547 /* We need another r10_bio. It has already been counted
NeilBrownd4432c22011-07-28 11:39:24 +10001548 * in bio->bi_phys_segments.
1549 */
1550 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1551
1552 r10_bio->master_bio = bio;
1553 r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1554
1555 r10_bio->mddev = mddev;
1556 r10_bio->sector = bio->bi_sector + sectors_handled;
1557 r10_bio->state = 0;
1558 goto retry_write;
1559 }
NeilBrown079fa162011-09-10 17:21:23 +10001560 one_write_done(r10_bio);
1561
1562 /* In case raid10d snuck in to freeze_array */
1563 wake_up(&conf->wait_barrier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564}
1565
NeilBrownfd01b882011-10-11 16:47:53 +11001566static void status(struct seq_file *seq, struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567{
NeilBrowne879a872011-10-11 16:49:02 +11001568 struct r10conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 int i;
1570
NeilBrown5cf00fc2012-05-21 09:28:20 +10001571 if (conf->geo.near_copies < conf->geo.raid_disks)
Andre Noll9d8f0362009-06-18 08:45:01 +10001572 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
NeilBrown5cf00fc2012-05-21 09:28:20 +10001573 if (conf->geo.near_copies > 1)
1574 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1575 if (conf->geo.far_copies > 1) {
1576 if (conf->geo.far_offset)
1577 seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
NeilBrownc93983b2006-06-26 00:27:41 -07001578 else
NeilBrown5cf00fc2012-05-21 09:28:20 +10001579 seq_printf(seq, " %d far-copies", conf->geo.far_copies);
NeilBrownc93983b2006-06-26 00:27:41 -07001580 }
NeilBrown5cf00fc2012-05-21 09:28:20 +10001581 seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1582 conf->geo.raid_disks - mddev->degraded);
1583 for (i = 0; i < conf->geo.raid_disks; i++)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 seq_printf(seq, "%s",
1585 conf->mirrors[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08001586 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587 seq_printf(seq, "]");
1588}
1589
NeilBrown700c7212011-07-27 11:00:36 +10001590/* check if there are enough drives for
1591 * every block to appear on atleast one.
1592 * Don't consider the device numbered 'ignore'
1593 * as we might be about to remove it.
1594 */
NeilBrownf8c9e742012-05-21 09:28:33 +10001595static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
NeilBrown700c7212011-07-27 11:00:36 +10001596{
1597 int first = 0;
1598
1599 do {
1600 int n = conf->copies;
1601 int cnt = 0;
NeilBrown80b48122012-09-27 12:35:21 +10001602 int this = first;
NeilBrown700c7212011-07-27 11:00:36 +10001603 while (n--) {
NeilBrown80b48122012-09-27 12:35:21 +10001604 if (conf->mirrors[this].rdev &&
1605 this != ignore)
NeilBrown700c7212011-07-27 11:00:36 +10001606 cnt++;
NeilBrown80b48122012-09-27 12:35:21 +10001607 this = (this+1) % geo->raid_disks;
NeilBrown700c7212011-07-27 11:00:36 +10001608 }
1609 if (cnt == 0)
1610 return 0;
NeilBrown80b48122012-09-27 12:35:21 +10001611 first = (first + geo->near_copies) % geo->raid_disks;
NeilBrown700c7212011-07-27 11:00:36 +10001612 } while (first != 0);
1613 return 1;
1614}
1615
NeilBrownf8c9e742012-05-21 09:28:33 +10001616static int enough(struct r10conf *conf, int ignore)
1617{
1618 return _enough(conf, &conf->geo, ignore) &&
1619 _enough(conf, &conf->prev, ignore);
1620}
1621
NeilBrownfd01b882011-10-11 16:47:53 +11001622static void error(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623{
1624 char b[BDEVNAME_SIZE];
NeilBrowne879a872011-10-11 16:49:02 +11001625 struct r10conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626
1627 /*
1628 * If it is not operational, then we have already marked it as dead
1629 * else if it is the last working disks, ignore the error, let the
1630 * next level up know.
1631 * else mark the drive as failed
1632 */
NeilBrownb2d444d2005-11-08 21:39:31 -08001633 if (test_bit(In_sync, &rdev->flags)
NeilBrown700c7212011-07-27 11:00:36 +10001634 && !enough(conf, rdev->raid_disk))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 /*
1636 * Don't fail the drive, just return an IO error.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637 */
1638 return;
NeilBrownc04be0a2006-10-03 01:15:53 -07001639 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1640 unsigned long flags;
1641 spin_lock_irqsave(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 mddev->degraded++;
NeilBrownc04be0a2006-10-03 01:15:53 -07001643 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 /*
1645 * if recovery is running, make sure it aborts.
1646 */
NeilBrowndfc70642008-05-23 13:04:39 -07001647 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 }
NeilBrownde393cd2011-07-28 11:31:48 +10001649 set_bit(Blocked, &rdev->flags);
NeilBrownb2d444d2005-11-08 21:39:31 -08001650 set_bit(Faulty, &rdev->flags);
NeilBrown850b2b42006-10-03 01:15:46 -07001651 set_bit(MD_CHANGE_DEVS, &mddev->flags);
Joe Perches067032b2011-01-14 09:14:33 +11001652 printk(KERN_ALERT
1653 "md/raid10:%s: Disk failure on %s, disabling device.\n"
1654 "md/raid10:%s: Operation continuing on %d devices.\n",
NeilBrown128595e2010-05-03 14:47:14 +10001655 mdname(mddev), bdevname(rdev->bdev, b),
NeilBrown5cf00fc2012-05-21 09:28:20 +10001656 mdname(mddev), conf->geo.raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657}
1658
NeilBrowne879a872011-10-11 16:49:02 +11001659static void print_conf(struct r10conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660{
1661 int i;
Jonathan Brassowdc280d982012-07-31 10:03:52 +10001662 struct raid10_info *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663
NeilBrown128595e2010-05-03 14:47:14 +10001664 printk(KERN_DEBUG "RAID10 conf printout:\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665 if (!conf) {
NeilBrown128595e2010-05-03 14:47:14 +10001666 printk(KERN_DEBUG "(!conf)\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 return;
1668 }
NeilBrown5cf00fc2012-05-21 09:28:20 +10001669 printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1670 conf->geo.raid_disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671
NeilBrown5cf00fc2012-05-21 09:28:20 +10001672 for (i = 0; i < conf->geo.raid_disks; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 char b[BDEVNAME_SIZE];
1674 tmp = conf->mirrors + i;
1675 if (tmp->rdev)
NeilBrown128595e2010-05-03 14:47:14 +10001676 printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
NeilBrownb2d444d2005-11-08 21:39:31 -08001677 i, !test_bit(In_sync, &tmp->rdev->flags),
1678 !test_bit(Faulty, &tmp->rdev->flags),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679 bdevname(tmp->rdev->bdev,b));
1680 }
1681}
1682
NeilBrowne879a872011-10-11 16:49:02 +11001683static void close_sync(struct r10conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684{
NeilBrown0a27ec92006-01-06 00:20:13 -08001685 wait_barrier(conf);
1686 allow_barrier(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687
1688 mempool_destroy(conf->r10buf_pool);
1689 conf->r10buf_pool = NULL;
1690}
1691
NeilBrownfd01b882011-10-11 16:47:53 +11001692static int raid10_spare_active(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693{
1694 int i;
NeilBrowne879a872011-10-11 16:49:02 +11001695 struct r10conf *conf = mddev->private;
Jonathan Brassowdc280d982012-07-31 10:03:52 +10001696 struct raid10_info *tmp;
NeilBrown6b965622010-08-18 11:56:59 +10001697 int count = 0;
1698 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699
1700 /*
1701 * Find all non-in_sync disks within the RAID10 configuration
1702 * and mark them in_sync
1703 */
NeilBrown5cf00fc2012-05-21 09:28:20 +10001704 for (i = 0; i < conf->geo.raid_disks; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705 tmp = conf->mirrors + i;
NeilBrown4ca40c22011-12-23 10:17:55 +11001706 if (tmp->replacement
1707 && tmp->replacement->recovery_offset == MaxSector
1708 && !test_bit(Faulty, &tmp->replacement->flags)
1709 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1710 /* Replacement has just become active */
1711 if (!tmp->rdev
1712 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1713 count++;
1714 if (tmp->rdev) {
1715 /* Replaced device not technically faulty,
1716 * but we need to be sure it gets removed
1717 * and never re-added.
1718 */
1719 set_bit(Faulty, &tmp->rdev->flags);
1720 sysfs_notify_dirent_safe(
1721 tmp->rdev->sysfs_state);
1722 }
1723 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1724 } else if (tmp->rdev
1725 && !test_bit(Faulty, &tmp->rdev->flags)
1726 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
NeilBrown6b965622010-08-18 11:56:59 +10001727 count++;
Jonathan Brassow2863b9e2012-10-11 13:38:58 +11001728 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 }
1730 }
NeilBrown6b965622010-08-18 11:56:59 +10001731 spin_lock_irqsave(&conf->device_lock, flags);
1732 mddev->degraded -= count;
1733 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734
1735 print_conf(conf);
NeilBrown6b965622010-08-18 11:56:59 +10001736 return count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737}
1738
1739
NeilBrownfd01b882011-10-11 16:47:53 +11001740static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741{
NeilBrowne879a872011-10-11 16:49:02 +11001742 struct r10conf *conf = mddev->private;
Neil Brown199050e2008-06-28 08:31:33 +10001743 int err = -EEXIST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 int mirror;
Neil Brown6c2fce22008-06-28 08:31:31 +10001745 int first = 0;
NeilBrown5cf00fc2012-05-21 09:28:20 +10001746 int last = conf->geo.raid_disks - 1;
NeilBrown050b6612012-03-19 12:46:39 +11001747 struct request_queue *q = bdev_get_queue(rdev->bdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748
1749 if (mddev->recovery_cp < MaxSector)
1750 /* only hot-add to in-sync arrays, as recovery is
1751 * very different from resync
1752 */
Neil Brown199050e2008-06-28 08:31:33 +10001753 return -EBUSY;
NeilBrownf8c9e742012-05-21 09:28:33 +10001754 if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
Neil Brown199050e2008-06-28 08:31:33 +10001755 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756
NeilBrowna53a6c82008-11-06 17:28:20 +11001757 if (rdev->raid_disk >= 0)
Neil Brown6c2fce22008-06-28 08:31:31 +10001758 first = last = rdev->raid_disk;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759
NeilBrown050b6612012-03-19 12:46:39 +11001760 if (q->merge_bvec_fn) {
1761 set_bit(Unmerged, &rdev->flags);
1762 mddev->merge_check_needed = 1;
1763 }
1764
Namhyung Kim2c4193d2011-07-18 17:38:43 +10001765 if (rdev->saved_raid_disk >= first &&
NeilBrown6cce3b232006-01-06 00:20:16 -08001766 conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1767 mirror = rdev->saved_raid_disk;
1768 else
Neil Brown6c2fce22008-06-28 08:31:31 +10001769 mirror = first;
NeilBrown2bb77732011-07-27 11:00:36 +10001770 for ( ; mirror <= last ; mirror++) {
Jonathan Brassowdc280d982012-07-31 10:03:52 +10001771 struct raid10_info *p = &conf->mirrors[mirror];
NeilBrown2bb77732011-07-27 11:00:36 +10001772 if (p->recovery_disabled == mddev->recovery_disabled)
1773 continue;
NeilBrownb7044d42011-12-23 10:17:56 +11001774 if (p->rdev) {
1775 if (!test_bit(WantReplacement, &p->rdev->flags) ||
1776 p->replacement != NULL)
1777 continue;
1778 clear_bit(In_sync, &rdev->flags);
1779 set_bit(Replacement, &rdev->flags);
1780 rdev->raid_disk = mirror;
1781 err = 0;
1782 disk_stack_limits(mddev->gendisk, rdev->bdev,
1783 rdev->data_offset << 9);
NeilBrownb7044d42011-12-23 10:17:56 +11001784 conf->fullsync = 1;
1785 rcu_assign_pointer(p->replacement, rdev);
1786 break;
1787 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788
NeilBrown2bb77732011-07-27 11:00:36 +10001789 disk_stack_limits(mddev->gendisk, rdev->bdev,
1790 rdev->data_offset << 9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791
NeilBrown2bb77732011-07-27 11:00:36 +10001792 p->head_position = 0;
NeilBrownd890fa22011-10-26 11:54:39 +11001793 p->recovery_disabled = mddev->recovery_disabled - 1;
NeilBrown2bb77732011-07-27 11:00:36 +10001794 rdev->raid_disk = mirror;
1795 err = 0;
1796 if (rdev->saved_raid_disk != mirror)
1797 conf->fullsync = 1;
1798 rcu_assign_pointer(p->rdev, rdev);
1799 break;
1800 }
NeilBrown050b6612012-03-19 12:46:39 +11001801 if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1802 /* Some requests might not have seen this new
1803 * merge_bvec_fn. We must wait for them to complete
1804 * before merging the device fully.
1805 * First we make sure any code which has tested
1806 * our function has submitted the request, then
1807 * we wait for all outstanding requests to complete.
1808 */
1809 synchronize_sched();
1810 raise_barrier(conf, 0);
1811 lower_barrier(conf);
1812 clear_bit(Unmerged, &rdev->flags);
1813 }
Andre Nollac5e7112009-08-03 10:59:47 +10001814 md_integrity_add_rdev(rdev, mddev);
Jonathan Brassowed30be02012-10-31 11:42:30 +11001815 if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
Shaohua Li532a2a32012-10-11 13:30:52 +11001816 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1817
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818 print_conf(conf);
Neil Brown199050e2008-06-28 08:31:33 +10001819 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820}
1821
NeilBrownb8321b62011-12-23 10:17:51 +11001822static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823{
NeilBrowne879a872011-10-11 16:49:02 +11001824 struct r10conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 int err = 0;
NeilBrownb8321b62011-12-23 10:17:51 +11001826 int number = rdev->raid_disk;
NeilBrownc8ab9032011-12-23 10:17:54 +11001827 struct md_rdev **rdevp;
Jonathan Brassowdc280d982012-07-31 10:03:52 +10001828 struct raid10_info *p = conf->mirrors + number;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829
1830 print_conf(conf);
NeilBrownc8ab9032011-12-23 10:17:54 +11001831 if (rdev == p->rdev)
1832 rdevp = &p->rdev;
1833 else if (rdev == p->replacement)
1834 rdevp = &p->replacement;
1835 else
1836 return 0;
1837
1838 if (test_bit(In_sync, &rdev->flags) ||
1839 atomic_read(&rdev->nr_pending)) {
1840 err = -EBUSY;
1841 goto abort;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842 }
NeilBrownc8ab9032011-12-23 10:17:54 +11001843 /* Only remove faulty devices if recovery
1844 * is not possible.
1845 */
1846 if (!test_bit(Faulty, &rdev->flags) &&
1847 mddev->recovery_disabled != p->recovery_disabled &&
NeilBrown4ca40c22011-12-23 10:17:55 +11001848 (!p->replacement || p->replacement == rdev) &&
NeilBrown63aced62012-05-22 13:55:33 +10001849 number < conf->geo.raid_disks &&
NeilBrownc8ab9032011-12-23 10:17:54 +11001850 enough(conf, -1)) {
1851 err = -EBUSY;
1852 goto abort;
1853 }
1854 *rdevp = NULL;
1855 synchronize_rcu();
1856 if (atomic_read(&rdev->nr_pending)) {
1857 /* lost the race, try later */
1858 err = -EBUSY;
1859 *rdevp = rdev;
1860 goto abort;
NeilBrown4ca40c22011-12-23 10:17:55 +11001861 } else if (p->replacement) {
1862 /* We must have just cleared 'rdev' */
1863 p->rdev = p->replacement;
1864 clear_bit(Replacement, &p->replacement->flags);
1865 smp_mb(); /* Make sure other CPUs may see both as identical
1866 * but will never see neither -- if they are careful.
1867 */
1868 p->replacement = NULL;
1869 clear_bit(WantReplacement, &rdev->flags);
1870 } else
1871 /* We might have just remove the Replacement as faulty
1872 * Clear the flag just in case
1873 */
1874 clear_bit(WantReplacement, &rdev->flags);
1875
NeilBrownc8ab9032011-12-23 10:17:54 +11001876 err = md_integrity_register(mddev);
1877
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878abort:
1879
1880 print_conf(conf);
1881 return err;
1882}
1883
1884
NeilBrown6712ecf2007-09-27 12:47:43 +02001885static void end_sync_read(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886{
NeilBrown9f2c9d12011-10-11 16:48:43 +11001887 struct r10bio *r10_bio = bio->bi_private;
NeilBrowne879a872011-10-11 16:49:02 +11001888 struct r10conf *conf = r10_bio->mddev->private;
Namhyung Kim778ca012011-07-18 17:38:47 +10001889 int d;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890
NeilBrown3ea7daa2012-05-22 13:53:47 +10001891 if (bio == r10_bio->master_bio) {
1892 /* this is a reshape read */
1893 d = r10_bio->read_slot; /* really the read dev */
1894 } else
1895 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
NeilBrown0eb3ff12006-01-06 00:20:29 -08001896
1897 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1898 set_bit(R10BIO_Uptodate, &r10_bio->state);
NeilBrowne684e412011-07-28 11:39:25 +10001899 else
1900 /* The write handler will notice the lack of
1901 * R10BIO_Uptodate and record any errors etc
1902 */
NeilBrown4dbcdc72006-01-06 00:20:52 -08001903 atomic_add(r10_bio->sectors,
1904 &conf->mirrors[d].rdev->corrected_errors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905
1906 /* for reconstruct, we always reschedule after a read.
1907 * for resync, only after all reads
1908 */
NeilBrown73d5c382009-02-25 13:18:47 +11001909 rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1911 atomic_dec_and_test(&r10_bio->remaining)) {
1912 /* we have read all the blocks,
1913 * do the comparison in process context in raid10d
1914 */
1915 reschedule_retry(r10_bio);
1916 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917}
1918
NeilBrown9f2c9d12011-10-11 16:48:43 +11001919static void end_sync_request(struct r10bio *r10_bio)
NeilBrown5e570282011-07-28 11:39:25 +10001920{
NeilBrownfd01b882011-10-11 16:47:53 +11001921 struct mddev *mddev = r10_bio->mddev;
NeilBrown5e570282011-07-28 11:39:25 +10001922
1923 while (atomic_dec_and_test(&r10_bio->remaining)) {
1924 if (r10_bio->master_bio == NULL) {
1925 /* the primary of several recovery bios */
1926 sector_t s = r10_bio->sectors;
1927 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1928 test_bit(R10BIO_WriteError, &r10_bio->state))
1929 reschedule_retry(r10_bio);
1930 else
1931 put_buf(r10_bio);
1932 md_done_sync(mddev, s, 1);
1933 break;
1934 } else {
NeilBrown9f2c9d12011-10-11 16:48:43 +11001935 struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
NeilBrown5e570282011-07-28 11:39:25 +10001936 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1937 test_bit(R10BIO_WriteError, &r10_bio->state))
1938 reschedule_retry(r10_bio);
1939 else
1940 put_buf(r10_bio);
1941 r10_bio = r10_bio2;
1942 }
1943 }
1944}
1945
NeilBrown6712ecf2007-09-27 12:47:43 +02001946static void end_sync_write(struct bio *bio, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947{
1948 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
NeilBrown9f2c9d12011-10-11 16:48:43 +11001949 struct r10bio *r10_bio = bio->bi_private;
NeilBrownfd01b882011-10-11 16:47:53 +11001950 struct mddev *mddev = r10_bio->mddev;
NeilBrowne879a872011-10-11 16:49:02 +11001951 struct r10conf *conf = mddev->private;
Namhyung Kim778ca012011-07-18 17:38:47 +10001952 int d;
NeilBrown749c55e2011-07-28 11:39:24 +10001953 sector_t first_bad;
1954 int bad_sectors;
1955 int slot;
NeilBrown9ad1aef2011-12-23 10:17:55 +11001956 int repl;
NeilBrown4ca40c22011-12-23 10:17:55 +11001957 struct md_rdev *rdev = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958
NeilBrown9ad1aef2011-12-23 10:17:55 +11001959 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1960 if (repl)
1961 rdev = conf->mirrors[d].replacement;
NeilBrown547414d2012-03-13 11:21:20 +11001962 else
NeilBrown9ad1aef2011-12-23 10:17:55 +11001963 rdev = conf->mirrors[d].rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964
NeilBrown1a0b7cd2011-07-28 11:39:25 +10001965 if (!uptodate) {
NeilBrown9ad1aef2011-12-23 10:17:55 +11001966 if (repl)
1967 md_error(mddev, rdev);
1968 else {
1969 set_bit(WriteErrorSeen, &rdev->flags);
NeilBrownb7044d42011-12-23 10:17:56 +11001970 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1971 set_bit(MD_RECOVERY_NEEDED,
1972 &rdev->mddev->recovery);
NeilBrown9ad1aef2011-12-23 10:17:55 +11001973 set_bit(R10BIO_WriteError, &r10_bio->state);
1974 }
1975 } else if (is_badblock(rdev,
NeilBrown749c55e2011-07-28 11:39:24 +10001976 r10_bio->devs[slot].addr,
1977 r10_bio->sectors,
1978 &first_bad, &bad_sectors))
1979 set_bit(R10BIO_MadeGood, &r10_bio->state);
NeilBrowndfc70642008-05-23 13:04:39 -07001980
NeilBrown9ad1aef2011-12-23 10:17:55 +11001981 rdev_dec_pending(rdev, mddev);
NeilBrown5e570282011-07-28 11:39:25 +10001982
1983 end_sync_request(r10_bio);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984}
1985
1986/*
1987 * Note: sync and recover and handled very differently for raid10
1988 * This code is for resync.
1989 * For resync, we read through virtual addresses and read all blocks.
1990 * If there is any error, we schedule a write. The lowest numbered
1991 * drive is authoritative.
1992 * However requests come for physical address, so we need to map.
1993 * For every physical address there are raid_disks/copies virtual addresses,
1994 * which is always are least one, but is not necessarly an integer.
1995 * This means that a physical address can span multiple chunks, so we may
1996 * have to submit multiple io requests for a single sync request.
1997 */
1998/*
1999 * We check if all blocks are in-sync and only write to blocks that
2000 * aren't in sync
2001 */
NeilBrown9f2c9d12011-10-11 16:48:43 +11002002static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003{
NeilBrowne879a872011-10-11 16:49:02 +11002004 struct r10conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005 int i, first;
2006 struct bio *tbio, *fbio;
majianpengf4380a92012-04-12 16:04:47 +10002007 int vcnt;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008
2009 atomic_set(&r10_bio->remaining, 1);
2010
2011 /* find the first device with a block */
2012 for (i=0; i<conf->copies; i++)
2013 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2014 break;
2015
2016 if (i == conf->copies)
2017 goto done;
2018
2019 first = i;
2020 fbio = r10_bio->devs[i].bio;
2021
majianpengf4380a92012-04-12 16:04:47 +10002022 vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 /* now find blocks with errors */
NeilBrown0eb3ff12006-01-06 00:20:29 -08002024 for (i=0 ; i < conf->copies ; i++) {
2025 int j, d;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 tbio = r10_bio->devs[i].bio;
NeilBrown0eb3ff12006-01-06 00:20:29 -08002028
2029 if (tbio->bi_end_io != end_sync_read)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030 continue;
NeilBrown0eb3ff12006-01-06 00:20:29 -08002031 if (i == first)
2032 continue;
2033 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2034 /* We know that the bi_io_vec layout is the same for
2035 * both 'first' and 'i', so we just compare them.
2036 * All vec entries are PAGE_SIZE;
2037 */
2038 for (j = 0; j < vcnt; j++)
2039 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2040 page_address(tbio->bi_io_vec[j].bv_page),
NeilBrown5020ad72012-04-02 01:39:05 +10002041 fbio->bi_io_vec[j].bv_len))
NeilBrown0eb3ff12006-01-06 00:20:29 -08002042 break;
2043 if (j == vcnt)
2044 continue;
Jianpeng Ma7f7583d2012-10-11 14:17:59 +11002045 atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
NeilBrownf84ee362011-07-28 11:39:25 +10002046 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2047 /* Don't fix anything. */
2048 continue;
NeilBrown0eb3ff12006-01-06 00:20:29 -08002049 }
NeilBrownf84ee362011-07-28 11:39:25 +10002050 /* Ok, we need to write this bio, either to correct an
2051 * inconsistency or to correct an unreadable block.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 * First we need to fixup bv_offset, bv_len and
2053 * bi_vecs, as the read request might have corrupted these
2054 */
2055 tbio->bi_vcnt = vcnt;
2056 tbio->bi_size = r10_bio->sectors << 9;
2057 tbio->bi_idx = 0;
2058 tbio->bi_phys_segments = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
2060 tbio->bi_flags |= 1 << BIO_UPTODATE;
2061 tbio->bi_next = NULL;
2062 tbio->bi_rw = WRITE;
2063 tbio->bi_private = r10_bio;
2064 tbio->bi_sector = r10_bio->devs[i].addr;
2065
2066 for (j=0; j < vcnt ; j++) {
2067 tbio->bi_io_vec[j].bv_offset = 0;
2068 tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2069
2070 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2071 page_address(fbio->bi_io_vec[j].bv_page),
2072 PAGE_SIZE);
2073 }
2074 tbio->bi_end_io = end_sync_write;
2075
2076 d = r10_bio->devs[i].devnum;
2077 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2078 atomic_inc(&r10_bio->remaining);
2079 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
2080
2081 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
2082 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2083 generic_make_request(tbio);
2084 }
2085
NeilBrown9ad1aef2011-12-23 10:17:55 +11002086 /* Now write out to any replacement devices
2087 * that are active
2088 */
2089 for (i = 0; i < conf->copies; i++) {
2090 int j, d;
NeilBrown9ad1aef2011-12-23 10:17:55 +11002091
2092 tbio = r10_bio->devs[i].repl_bio;
2093 if (!tbio || !tbio->bi_end_io)
2094 continue;
2095 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2096 && r10_bio->devs[i].bio != fbio)
2097 for (j = 0; j < vcnt; j++)
2098 memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2099 page_address(fbio->bi_io_vec[j].bv_page),
2100 PAGE_SIZE);
2101 d = r10_bio->devs[i].devnum;
2102 atomic_inc(&r10_bio->remaining);
2103 md_sync_acct(conf->mirrors[d].replacement->bdev,
2104 tbio->bi_size >> 9);
2105 generic_make_request(tbio);
2106 }
2107
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108done:
2109 if (atomic_dec_and_test(&r10_bio->remaining)) {
2110 md_done_sync(mddev, r10_bio->sectors, 1);
2111 put_buf(r10_bio);
2112 }
2113}
2114
2115/*
2116 * Now for the recovery code.
2117 * Recovery happens across physical sectors.
2118 * We recover all non-is_sync drives by finding the virtual address of
2119 * each, and then choose a working drive that also has that virt address.
2120 * There is a separate r10_bio for each non-in_sync drive.
2121 * Only the first two slots are in use. The first for reading,
2122 * The second for writing.
2123 *
2124 */
NeilBrown9f2c9d12011-10-11 16:48:43 +11002125static void fix_recovery_read_error(struct r10bio *r10_bio)
NeilBrown5e570282011-07-28 11:39:25 +10002126{
2127 /* We got a read error during recovery.
2128 * We repeat the read in smaller page-sized sections.
2129 * If a read succeeds, write it to the new device or record
2130 * a bad block if we cannot.
2131 * If a read fails, record a bad block on both old and
2132 * new devices.
2133 */
NeilBrownfd01b882011-10-11 16:47:53 +11002134 struct mddev *mddev = r10_bio->mddev;
NeilBrowne879a872011-10-11 16:49:02 +11002135 struct r10conf *conf = mddev->private;
NeilBrown5e570282011-07-28 11:39:25 +10002136 struct bio *bio = r10_bio->devs[0].bio;
2137 sector_t sect = 0;
2138 int sectors = r10_bio->sectors;
2139 int idx = 0;
2140 int dr = r10_bio->devs[0].devnum;
2141 int dw = r10_bio->devs[1].devnum;
2142
2143 while (sectors) {
2144 int s = sectors;
NeilBrown3cb03002011-10-11 16:45:26 +11002145 struct md_rdev *rdev;
NeilBrown5e570282011-07-28 11:39:25 +10002146 sector_t addr;
2147 int ok;
2148
2149 if (s > (PAGE_SIZE>>9))
2150 s = PAGE_SIZE >> 9;
2151
2152 rdev = conf->mirrors[dr].rdev;
2153 addr = r10_bio->devs[0].addr + sect,
2154 ok = sync_page_io(rdev,
2155 addr,
2156 s << 9,
2157 bio->bi_io_vec[idx].bv_page,
2158 READ, false);
2159 if (ok) {
2160 rdev = conf->mirrors[dw].rdev;
2161 addr = r10_bio->devs[1].addr + sect;
2162 ok = sync_page_io(rdev,
2163 addr,
2164 s << 9,
2165 bio->bi_io_vec[idx].bv_page,
2166 WRITE, false);
NeilBrownb7044d42011-12-23 10:17:56 +11002167 if (!ok) {
NeilBrown5e570282011-07-28 11:39:25 +10002168 set_bit(WriteErrorSeen, &rdev->flags);
NeilBrownb7044d42011-12-23 10:17:56 +11002169 if (!test_and_set_bit(WantReplacement,
2170 &rdev->flags))
2171 set_bit(MD_RECOVERY_NEEDED,
2172 &rdev->mddev->recovery);
2173 }
NeilBrown5e570282011-07-28 11:39:25 +10002174 }
2175 if (!ok) {
2176 /* We don't worry if we cannot set a bad block -
2177 * it really is bad so there is no loss in not
2178 * recording it yet
2179 */
2180 rdev_set_badblocks(rdev, addr, s, 0);
2181
2182 if (rdev != conf->mirrors[dw].rdev) {
2183 /* need bad block on destination too */
NeilBrown3cb03002011-10-11 16:45:26 +11002184 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
NeilBrown5e570282011-07-28 11:39:25 +10002185 addr = r10_bio->devs[1].addr + sect;
2186 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2187 if (!ok) {
2188 /* just abort the recovery */
2189 printk(KERN_NOTICE
2190 "md/raid10:%s: recovery aborted"
2191 " due to read error\n",
2192 mdname(mddev));
2193
2194 conf->mirrors[dw].recovery_disabled
2195 = mddev->recovery_disabled;
2196 set_bit(MD_RECOVERY_INTR,
2197 &mddev->recovery);
2198 break;
2199 }
2200 }
2201 }
2202
2203 sectors -= s;
2204 sect += s;
2205 idx++;
2206 }
2207}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208
NeilBrown9f2c9d12011-10-11 16:48:43 +11002209static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210{
NeilBrowne879a872011-10-11 16:49:02 +11002211 struct r10conf *conf = mddev->private;
Namhyung Kimc65060a2011-07-18 17:38:49 +10002212 int d;
NeilBrown24afd802011-12-23 10:17:55 +11002213 struct bio *wbio, *wbio2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214
NeilBrown5e570282011-07-28 11:39:25 +10002215 if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2216 fix_recovery_read_error(r10_bio);
2217 end_sync_request(r10_bio);
2218 return;
2219 }
2220
Namhyung Kimc65060a2011-07-18 17:38:49 +10002221 /*
2222 * share the pages with the first bio
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 * and submit the write request
2224 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 d = r10_bio->devs[1].devnum;
NeilBrown24afd802011-12-23 10:17:55 +11002226 wbio = r10_bio->devs[1].bio;
2227 wbio2 = r10_bio->devs[1].repl_bio;
2228 if (wbio->bi_end_io) {
2229 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2230 md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
2231 generic_make_request(wbio);
2232 }
2233 if (wbio2 && wbio2->bi_end_io) {
2234 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2235 md_sync_acct(conf->mirrors[d].replacement->bdev,
2236 wbio2->bi_size >> 9);
2237 generic_make_request(wbio2);
2238 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239}
2240
2241
2242/*
Robert Becker1e509152009-12-14 12:49:58 +11002243 * Used by fix_read_error() to decay the per rdev read_errors.
2244 * We halve the read error count for every hour that has elapsed
2245 * since the last recorded read error.
2246 *
2247 */
NeilBrownfd01b882011-10-11 16:47:53 +11002248static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
Robert Becker1e509152009-12-14 12:49:58 +11002249{
2250 struct timespec cur_time_mon;
2251 unsigned long hours_since_last;
2252 unsigned int read_errors = atomic_read(&rdev->read_errors);
2253
2254 ktime_get_ts(&cur_time_mon);
2255
2256 if (rdev->last_read_error.tv_sec == 0 &&
2257 rdev->last_read_error.tv_nsec == 0) {
2258 /* first time we've seen a read error */
2259 rdev->last_read_error = cur_time_mon;
2260 return;
2261 }
2262
2263 hours_since_last = (cur_time_mon.tv_sec -
2264 rdev->last_read_error.tv_sec) / 3600;
2265
2266 rdev->last_read_error = cur_time_mon;
2267
2268 /*
2269 * if hours_since_last is > the number of bits in read_errors
2270 * just set read errors to 0. We do this to avoid
2271 * overflowing the shift of read_errors by hours_since_last.
2272 */
2273 if (hours_since_last >= 8 * sizeof(read_errors))
2274 atomic_set(&rdev->read_errors, 0);
2275 else
2276 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2277}
2278
NeilBrown3cb03002011-10-11 16:45:26 +11002279static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
NeilBrown58c54fc2011-07-28 11:39:25 +10002280 int sectors, struct page *page, int rw)
2281{
2282 sector_t first_bad;
2283 int bad_sectors;
2284
2285 if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2286 && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2287 return -1;
2288 if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2289 /* success */
2290 return 1;
NeilBrownb7044d42011-12-23 10:17:56 +11002291 if (rw == WRITE) {
NeilBrown58c54fc2011-07-28 11:39:25 +10002292 set_bit(WriteErrorSeen, &rdev->flags);
NeilBrownb7044d42011-12-23 10:17:56 +11002293 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2294 set_bit(MD_RECOVERY_NEEDED,
2295 &rdev->mddev->recovery);
2296 }
NeilBrown58c54fc2011-07-28 11:39:25 +10002297 /* need to record an error - either for the block or the device */
2298 if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2299 md_error(rdev->mddev, rdev);
2300 return 0;
2301}
2302
Robert Becker1e509152009-12-14 12:49:58 +11002303/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 * This is a kernel thread which:
2305 *
2306 * 1. Retries failed read operations on working mirrors.
2307 * 2. Updates the raid superblock when problems encounter.
NeilBrown6814d532006-10-03 01:15:45 -07002308 * 3. Performs writes following reads for array synchronising.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309 */
2310
NeilBrowne879a872011-10-11 16:49:02 +11002311static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
NeilBrown6814d532006-10-03 01:15:45 -07002312{
2313 int sect = 0; /* Offset from r10_bio->sector */
2314 int sectors = r10_bio->sectors;
NeilBrown3cb03002011-10-11 16:45:26 +11002315 struct md_rdev*rdev;
Robert Becker1e509152009-12-14 12:49:58 +11002316 int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10002317 int d = r10_bio->devs[r10_bio->read_slot].devnum;
Robert Becker1e509152009-12-14 12:49:58 +11002318
NeilBrown7c4e06f2011-05-11 14:53:17 +10002319 /* still own a reference to this rdev, so it cannot
2320 * have been cleared recently.
2321 */
2322 rdev = conf->mirrors[d].rdev;
Robert Becker1e509152009-12-14 12:49:58 +11002323
NeilBrown7c4e06f2011-05-11 14:53:17 +10002324 if (test_bit(Faulty, &rdev->flags))
2325 /* drive has already been failed, just ignore any
2326 more fix_read_error() attempts */
2327 return;
2328
2329 check_decay_read_errors(mddev, rdev);
2330 atomic_inc(&rdev->read_errors);
2331 if (atomic_read(&rdev->read_errors) > max_read_errors) {
2332 char b[BDEVNAME_SIZE];
Robert Becker1e509152009-12-14 12:49:58 +11002333 bdevname(rdev->bdev, b);
2334
NeilBrown7c4e06f2011-05-11 14:53:17 +10002335 printk(KERN_NOTICE
2336 "md/raid10:%s: %s: Raid device exceeded "
2337 "read_error threshold [cur %d:max %d]\n",
2338 mdname(mddev), b,
2339 atomic_read(&rdev->read_errors), max_read_errors);
2340 printk(KERN_NOTICE
2341 "md/raid10:%s: %s: Failing raid device\n",
2342 mdname(mddev), b);
2343 md_error(mddev, conf->mirrors[d].rdev);
NeilBrownfae8cc5e2012-02-14 11:10:10 +11002344 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
NeilBrown7c4e06f2011-05-11 14:53:17 +10002345 return;
Robert Becker1e509152009-12-14 12:49:58 +11002346 }
Robert Becker1e509152009-12-14 12:49:58 +11002347
NeilBrown6814d532006-10-03 01:15:45 -07002348 while(sectors) {
2349 int s = sectors;
2350 int sl = r10_bio->read_slot;
2351 int success = 0;
2352 int start;
2353
2354 if (s > (PAGE_SIZE>>9))
2355 s = PAGE_SIZE >> 9;
2356
2357 rcu_read_lock();
2358 do {
NeilBrown8dbed5c2011-07-28 11:39:24 +10002359 sector_t first_bad;
2360 int bad_sectors;
2361
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10002362 d = r10_bio->devs[sl].devnum;
NeilBrown6814d532006-10-03 01:15:45 -07002363 rdev = rcu_dereference(conf->mirrors[d].rdev);
2364 if (rdev &&
NeilBrown050b6612012-03-19 12:46:39 +11002365 !test_bit(Unmerged, &rdev->flags) &&
NeilBrown8dbed5c2011-07-28 11:39:24 +10002366 test_bit(In_sync, &rdev->flags) &&
2367 is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2368 &first_bad, &bad_sectors) == 0) {
NeilBrown6814d532006-10-03 01:15:45 -07002369 atomic_inc(&rdev->nr_pending);
2370 rcu_read_unlock();
NeilBrown2b193362010-10-27 15:16:40 +11002371 success = sync_page_io(rdev,
NeilBrown6814d532006-10-03 01:15:45 -07002372 r10_bio->devs[sl].addr +
Jonathan Brassowccebd4c2011-01-14 09:14:33 +11002373 sect,
NeilBrown6814d532006-10-03 01:15:45 -07002374 s<<9,
Jonathan Brassowccebd4c2011-01-14 09:14:33 +11002375 conf->tmppage, READ, false);
NeilBrown6814d532006-10-03 01:15:45 -07002376 rdev_dec_pending(rdev, mddev);
2377 rcu_read_lock();
2378 if (success)
2379 break;
2380 }
2381 sl++;
2382 if (sl == conf->copies)
2383 sl = 0;
2384 } while (!success && sl != r10_bio->read_slot);
2385 rcu_read_unlock();
2386
2387 if (!success) {
NeilBrown58c54fc2011-07-28 11:39:25 +10002388 /* Cannot read from anywhere, just mark the block
2389 * as bad on the first device to discourage future
2390 * reads.
2391 */
NeilBrown6814d532006-10-03 01:15:45 -07002392 int dn = r10_bio->devs[r10_bio->read_slot].devnum;
NeilBrown58c54fc2011-07-28 11:39:25 +10002393 rdev = conf->mirrors[dn].rdev;
2394
2395 if (!rdev_set_badblocks(
2396 rdev,
2397 r10_bio->devs[r10_bio->read_slot].addr
2398 + sect,
NeilBrownfae8cc5e2012-02-14 11:10:10 +11002399 s, 0)) {
NeilBrown58c54fc2011-07-28 11:39:25 +10002400 md_error(mddev, rdev);
NeilBrownfae8cc5e2012-02-14 11:10:10 +11002401 r10_bio->devs[r10_bio->read_slot].bio
2402 = IO_BLOCKED;
2403 }
NeilBrown6814d532006-10-03 01:15:45 -07002404 break;
2405 }
2406
2407 start = sl;
2408 /* write it back and re-read */
2409 rcu_read_lock();
2410 while (sl != r10_bio->read_slot) {
Robert Becker67b8dc42009-12-14 12:49:57 +11002411 char b[BDEVNAME_SIZE];
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10002412
NeilBrown6814d532006-10-03 01:15:45 -07002413 if (sl==0)
2414 sl = conf->copies;
2415 sl--;
2416 d = r10_bio->devs[sl].devnum;
2417 rdev = rcu_dereference(conf->mirrors[d].rdev);
NeilBrown1294b9c2011-07-28 11:39:23 +10002418 if (!rdev ||
NeilBrown050b6612012-03-19 12:46:39 +11002419 test_bit(Unmerged, &rdev->flags) ||
NeilBrown1294b9c2011-07-28 11:39:23 +10002420 !test_bit(In_sync, &rdev->flags))
2421 continue;
2422
2423 atomic_inc(&rdev->nr_pending);
2424 rcu_read_unlock();
NeilBrown58c54fc2011-07-28 11:39:25 +10002425 if (r10_sync_page_io(rdev,
2426 r10_bio->devs[sl].addr +
2427 sect,
NeilBrown055d3742012-07-03 15:55:33 +10002428 s, conf->tmppage, WRITE)
NeilBrown1294b9c2011-07-28 11:39:23 +10002429 == 0) {
2430 /* Well, this device is dead */
2431 printk(KERN_NOTICE
2432 "md/raid10:%s: read correction "
2433 "write failed"
2434 " (%d sectors at %llu on %s)\n",
2435 mdname(mddev), s,
2436 (unsigned long long)(
NeilBrownf8c9e742012-05-21 09:28:33 +10002437 sect +
2438 choose_data_offset(r10_bio,
2439 rdev)),
NeilBrown1294b9c2011-07-28 11:39:23 +10002440 bdevname(rdev->bdev, b));
2441 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2442 "drive\n",
2443 mdname(mddev),
2444 bdevname(rdev->bdev, b));
NeilBrown6814d532006-10-03 01:15:45 -07002445 }
NeilBrown1294b9c2011-07-28 11:39:23 +10002446 rdev_dec_pending(rdev, mddev);
2447 rcu_read_lock();
NeilBrown6814d532006-10-03 01:15:45 -07002448 }
2449 sl = start;
2450 while (sl != r10_bio->read_slot) {
NeilBrown1294b9c2011-07-28 11:39:23 +10002451 char b[BDEVNAME_SIZE];
Prasanna S. Panchamukhi0544a212010-06-24 13:31:03 +10002452
NeilBrown6814d532006-10-03 01:15:45 -07002453 if (sl==0)
2454 sl = conf->copies;
2455 sl--;
2456 d = r10_bio->devs[sl].devnum;
2457 rdev = rcu_dereference(conf->mirrors[d].rdev);
NeilBrown1294b9c2011-07-28 11:39:23 +10002458 if (!rdev ||
2459 !test_bit(In_sync, &rdev->flags))
2460 continue;
Robert Becker67b8dc42009-12-14 12:49:57 +11002461
NeilBrown1294b9c2011-07-28 11:39:23 +10002462 atomic_inc(&rdev->nr_pending);
2463 rcu_read_unlock();
NeilBrown58c54fc2011-07-28 11:39:25 +10002464 switch (r10_sync_page_io(rdev,
2465 r10_bio->devs[sl].addr +
2466 sect,
NeilBrown055d3742012-07-03 15:55:33 +10002467 s, conf->tmppage,
NeilBrown58c54fc2011-07-28 11:39:25 +10002468 READ)) {
2469 case 0:
NeilBrown1294b9c2011-07-28 11:39:23 +10002470 /* Well, this device is dead */
2471 printk(KERN_NOTICE
2472 "md/raid10:%s: unable to read back "
2473 "corrected sectors"
2474 " (%d sectors at %llu on %s)\n",
2475 mdname(mddev), s,
2476 (unsigned long long)(
NeilBrownf8c9e742012-05-21 09:28:33 +10002477 sect +
2478 choose_data_offset(r10_bio, rdev)),
NeilBrown1294b9c2011-07-28 11:39:23 +10002479 bdevname(rdev->bdev, b));
2480 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2481 "drive\n",
2482 mdname(mddev),
2483 bdevname(rdev->bdev, b));
NeilBrown58c54fc2011-07-28 11:39:25 +10002484 break;
2485 case 1:
NeilBrown1294b9c2011-07-28 11:39:23 +10002486 printk(KERN_INFO
2487 "md/raid10:%s: read error corrected"
2488 " (%d sectors at %llu on %s)\n",
2489 mdname(mddev), s,
2490 (unsigned long long)(
NeilBrownf8c9e742012-05-21 09:28:33 +10002491 sect +
2492 choose_data_offset(r10_bio, rdev)),
NeilBrown1294b9c2011-07-28 11:39:23 +10002493 bdevname(rdev->bdev, b));
2494 atomic_add(s, &rdev->corrected_errors);
NeilBrown6814d532006-10-03 01:15:45 -07002495 }
NeilBrown1294b9c2011-07-28 11:39:23 +10002496
2497 rdev_dec_pending(rdev, mddev);
2498 rcu_read_lock();
NeilBrown6814d532006-10-03 01:15:45 -07002499 }
2500 rcu_read_unlock();
2501
2502 sectors -= s;
2503 sect += s;
2504 }
2505}
2506
NeilBrownbd870a12011-07-28 11:39:24 +10002507static void bi_complete(struct bio *bio, int error)
2508{
2509 complete((struct completion *)bio->bi_private);
2510}
2511
2512static int submit_bio_wait(int rw, struct bio *bio)
2513{
2514 struct completion event;
2515 rw |= REQ_SYNC;
2516
2517 init_completion(&event);
2518 bio->bi_private = &event;
2519 bio->bi_end_io = bi_complete;
2520 submit_bio(rw, bio);
2521 wait_for_completion(&event);
2522
2523 return test_bit(BIO_UPTODATE, &bio->bi_flags);
2524}
2525
NeilBrown9f2c9d12011-10-11 16:48:43 +11002526static int narrow_write_error(struct r10bio *r10_bio, int i)
NeilBrownbd870a12011-07-28 11:39:24 +10002527{
2528 struct bio *bio = r10_bio->master_bio;
NeilBrownfd01b882011-10-11 16:47:53 +11002529 struct mddev *mddev = r10_bio->mddev;
NeilBrowne879a872011-10-11 16:49:02 +11002530 struct r10conf *conf = mddev->private;
NeilBrown3cb03002011-10-11 16:45:26 +11002531 struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
NeilBrownbd870a12011-07-28 11:39:24 +10002532 /* bio has the data to be written to slot 'i' where
2533 * we just recently had a write error.
2534 * We repeatedly clone the bio and trim down to one block,
2535 * then try the write. Where the write fails we record
2536 * a bad block.
2537 * It is conceivable that the bio doesn't exactly align with
2538 * blocks. We must handle this.
2539 *
2540 * We currently own a reference to the rdev.
2541 */
2542
2543 int block_sectors;
2544 sector_t sector;
2545 int sectors;
2546 int sect_to_write = r10_bio->sectors;
2547 int ok = 1;
2548
2549 if (rdev->badblocks.shift < 0)
2550 return 0;
2551
2552 block_sectors = 1 << rdev->badblocks.shift;
2553 sector = r10_bio->sector;
2554 sectors = ((r10_bio->sector + block_sectors)
2555 & ~(sector_t)(block_sectors - 1))
2556 - sector;
2557
2558 while (sect_to_write) {
2559 struct bio *wbio;
2560 if (sectors > sect_to_write)
2561 sectors = sect_to_write;
2562 /* Write at 'sector' for 'sectors' */
2563 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2564 md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2565 wbio->bi_sector = (r10_bio->devs[i].addr+
NeilBrownf8c9e742012-05-21 09:28:33 +10002566 choose_data_offset(r10_bio, rdev) +
NeilBrownbd870a12011-07-28 11:39:24 +10002567 (sector - r10_bio->sector));
2568 wbio->bi_bdev = rdev->bdev;
2569 if (submit_bio_wait(WRITE, wbio) == 0)
2570 /* Failure! */
2571 ok = rdev_set_badblocks(rdev, sector,
2572 sectors, 0)
2573 && ok;
2574
2575 bio_put(wbio);
2576 sect_to_write -= sectors;
2577 sector += sectors;
2578 sectors = block_sectors;
2579 }
2580 return ok;
2581}
2582
NeilBrown9f2c9d12011-10-11 16:48:43 +11002583static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
NeilBrown560f8e52011-07-28 11:39:23 +10002584{
2585 int slot = r10_bio->read_slot;
NeilBrown560f8e52011-07-28 11:39:23 +10002586 struct bio *bio;
NeilBrowne879a872011-10-11 16:49:02 +11002587 struct r10conf *conf = mddev->private;
NeilBrownabbf0982011-12-23 10:17:54 +11002588 struct md_rdev *rdev = r10_bio->devs[slot].rdev;
NeilBrown560f8e52011-07-28 11:39:23 +10002589 char b[BDEVNAME_SIZE];
2590 unsigned long do_sync;
NeilBrown856e08e2011-07-28 11:39:23 +10002591 int max_sectors;
NeilBrown560f8e52011-07-28 11:39:23 +10002592
2593 /* we got a read error. Maybe the drive is bad. Maybe just
2594 * the block and we can fix it.
2595 * We freeze all other IO, and try reading the block from
2596 * other devices. When we find one, we re-write
2597 * and check it that fixes the read error.
2598 * This is all done synchronously while the array is
2599 * frozen.
2600 */
NeilBrownfae8cc5e2012-02-14 11:10:10 +11002601 bio = r10_bio->devs[slot].bio;
2602 bdevname(bio->bi_bdev, b);
2603 bio_put(bio);
2604 r10_bio->devs[slot].bio = NULL;
2605
NeilBrown560f8e52011-07-28 11:39:23 +10002606 if (mddev->ro == 0) {
2607 freeze_array(conf);
2608 fix_read_error(conf, mddev, r10_bio);
2609 unfreeze_array(conf);
NeilBrownfae8cc5e2012-02-14 11:10:10 +11002610 } else
2611 r10_bio->devs[slot].bio = IO_BLOCKED;
2612
NeilBrownabbf0982011-12-23 10:17:54 +11002613 rdev_dec_pending(rdev, mddev);
NeilBrown560f8e52011-07-28 11:39:23 +10002614
NeilBrown7399c312011-07-28 11:39:23 +10002615read_more:
NeilBrown96c3fd12011-12-23 10:17:54 +11002616 rdev = read_balance(conf, r10_bio, &max_sectors);
2617 if (rdev == NULL) {
NeilBrown560f8e52011-07-28 11:39:23 +10002618 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2619 " read error for block %llu\n",
NeilBrown7399c312011-07-28 11:39:23 +10002620 mdname(mddev), b,
NeilBrown560f8e52011-07-28 11:39:23 +10002621 (unsigned long long)r10_bio->sector);
2622 raid_end_bio_io(r10_bio);
NeilBrown560f8e52011-07-28 11:39:23 +10002623 return;
2624 }
2625
2626 do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
NeilBrown560f8e52011-07-28 11:39:23 +10002627 slot = r10_bio->read_slot;
NeilBrown560f8e52011-07-28 11:39:23 +10002628 printk_ratelimited(
2629 KERN_ERR
NeilBrown055d3742012-07-03 15:55:33 +10002630 "md/raid10:%s: %s: redirecting "
NeilBrown560f8e52011-07-28 11:39:23 +10002631 "sector %llu to another mirror\n",
2632 mdname(mddev),
2633 bdevname(rdev->bdev, b),
2634 (unsigned long long)r10_bio->sector);
2635 bio = bio_clone_mddev(r10_bio->master_bio,
2636 GFP_NOIO, mddev);
NeilBrown7399c312011-07-28 11:39:23 +10002637 md_trim_bio(bio,
2638 r10_bio->sector - bio->bi_sector,
2639 max_sectors);
NeilBrown560f8e52011-07-28 11:39:23 +10002640 r10_bio->devs[slot].bio = bio;
NeilBrownabbf0982011-12-23 10:17:54 +11002641 r10_bio->devs[slot].rdev = rdev;
NeilBrown560f8e52011-07-28 11:39:23 +10002642 bio->bi_sector = r10_bio->devs[slot].addr
NeilBrownf8c9e742012-05-21 09:28:33 +10002643 + choose_data_offset(r10_bio, rdev);
NeilBrown560f8e52011-07-28 11:39:23 +10002644 bio->bi_bdev = rdev->bdev;
2645 bio->bi_rw = READ | do_sync;
2646 bio->bi_private = r10_bio;
2647 bio->bi_end_io = raid10_end_read_request;
NeilBrown7399c312011-07-28 11:39:23 +10002648 if (max_sectors < r10_bio->sectors) {
2649 /* Drat - have to split this up more */
2650 struct bio *mbio = r10_bio->master_bio;
2651 int sectors_handled =
2652 r10_bio->sector + max_sectors
2653 - mbio->bi_sector;
2654 r10_bio->sectors = max_sectors;
2655 spin_lock_irq(&conf->device_lock);
2656 if (mbio->bi_phys_segments == 0)
2657 mbio->bi_phys_segments = 2;
2658 else
2659 mbio->bi_phys_segments++;
2660 spin_unlock_irq(&conf->device_lock);
2661 generic_make_request(bio);
NeilBrown7399c312011-07-28 11:39:23 +10002662
2663 r10_bio = mempool_alloc(conf->r10bio_pool,
2664 GFP_NOIO);
2665 r10_bio->master_bio = mbio;
2666 r10_bio->sectors = (mbio->bi_size >> 9)
2667 - sectors_handled;
2668 r10_bio->state = 0;
2669 set_bit(R10BIO_ReadError,
2670 &r10_bio->state);
2671 r10_bio->mddev = mddev;
2672 r10_bio->sector = mbio->bi_sector
2673 + sectors_handled;
2674
2675 goto read_more;
2676 } else
2677 generic_make_request(bio);
NeilBrown560f8e52011-07-28 11:39:23 +10002678}
2679
NeilBrowne879a872011-10-11 16:49:02 +11002680static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
NeilBrown749c55e2011-07-28 11:39:24 +10002681{
2682 /* Some sort of write request has finished and it
2683 * succeeded in writing where we thought there was a
2684 * bad block. So forget the bad block.
NeilBrown1a0b7cd2011-07-28 11:39:25 +10002685 * Or possibly if failed and we need to record
2686 * a bad block.
NeilBrown749c55e2011-07-28 11:39:24 +10002687 */
2688 int m;
NeilBrown3cb03002011-10-11 16:45:26 +11002689 struct md_rdev *rdev;
NeilBrown749c55e2011-07-28 11:39:24 +10002690
2691 if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2692 test_bit(R10BIO_IsRecover, &r10_bio->state)) {
NeilBrown1a0b7cd2011-07-28 11:39:25 +10002693 for (m = 0; m < conf->copies; m++) {
2694 int dev = r10_bio->devs[m].devnum;
2695 rdev = conf->mirrors[dev].rdev;
2696 if (r10_bio->devs[m].bio == NULL)
2697 continue;
2698 if (test_bit(BIO_UPTODATE,
NeilBrown749c55e2011-07-28 11:39:24 +10002699 &r10_bio->devs[m].bio->bi_flags)) {
NeilBrown749c55e2011-07-28 11:39:24 +10002700 rdev_clear_badblocks(
2701 rdev,
2702 r10_bio->devs[m].addr,
NeilBrownc6563a82012-05-21 09:27:00 +10002703 r10_bio->sectors, 0);
NeilBrown1a0b7cd2011-07-28 11:39:25 +10002704 } else {
2705 if (!rdev_set_badblocks(
2706 rdev,
2707 r10_bio->devs[m].addr,
2708 r10_bio->sectors, 0))
2709 md_error(conf->mddev, rdev);
NeilBrown749c55e2011-07-28 11:39:24 +10002710 }
NeilBrown9ad1aef2011-12-23 10:17:55 +11002711 rdev = conf->mirrors[dev].replacement;
2712 if (r10_bio->devs[m].repl_bio == NULL)
2713 continue;
2714 if (test_bit(BIO_UPTODATE,
2715 &r10_bio->devs[m].repl_bio->bi_flags)) {
2716 rdev_clear_badblocks(
2717 rdev,
2718 r10_bio->devs[m].addr,
NeilBrownc6563a82012-05-21 09:27:00 +10002719 r10_bio->sectors, 0);
NeilBrown9ad1aef2011-12-23 10:17:55 +11002720 } else {
2721 if (!rdev_set_badblocks(
2722 rdev,
2723 r10_bio->devs[m].addr,
2724 r10_bio->sectors, 0))
2725 md_error(conf->mddev, rdev);
2726 }
NeilBrown1a0b7cd2011-07-28 11:39:25 +10002727 }
NeilBrown749c55e2011-07-28 11:39:24 +10002728 put_buf(r10_bio);
2729 } else {
NeilBrownbd870a12011-07-28 11:39:24 +10002730 for (m = 0; m < conf->copies; m++) {
2731 int dev = r10_bio->devs[m].devnum;
2732 struct bio *bio = r10_bio->devs[m].bio;
2733 rdev = conf->mirrors[dev].rdev;
2734 if (bio == IO_MADE_GOOD) {
NeilBrown749c55e2011-07-28 11:39:24 +10002735 rdev_clear_badblocks(
2736 rdev,
2737 r10_bio->devs[m].addr,
NeilBrownc6563a82012-05-21 09:27:00 +10002738 r10_bio->sectors, 0);
NeilBrown749c55e2011-07-28 11:39:24 +10002739 rdev_dec_pending(rdev, conf->mddev);
NeilBrownbd870a12011-07-28 11:39:24 +10002740 } else if (bio != NULL &&
2741 !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2742 if (!narrow_write_error(r10_bio, m)) {
2743 md_error(conf->mddev, rdev);
2744 set_bit(R10BIO_Degraded,
2745 &r10_bio->state);
2746 }
2747 rdev_dec_pending(rdev, conf->mddev);
NeilBrown749c55e2011-07-28 11:39:24 +10002748 }
NeilBrown475b0322011-12-23 10:17:55 +11002749 bio = r10_bio->devs[m].repl_bio;
2750 rdev = conf->mirrors[dev].replacement;
NeilBrown4ca40c22011-12-23 10:17:55 +11002751 if (rdev && bio == IO_MADE_GOOD) {
NeilBrown475b0322011-12-23 10:17:55 +11002752 rdev_clear_badblocks(
2753 rdev,
2754 r10_bio->devs[m].addr,
NeilBrownc6563a82012-05-21 09:27:00 +10002755 r10_bio->sectors, 0);
NeilBrown475b0322011-12-23 10:17:55 +11002756 rdev_dec_pending(rdev, conf->mddev);
2757 }
NeilBrownbd870a12011-07-28 11:39:24 +10002758 }
2759 if (test_bit(R10BIO_WriteError,
2760 &r10_bio->state))
2761 close_write(r10_bio);
NeilBrown749c55e2011-07-28 11:39:24 +10002762 raid_end_bio_io(r10_bio);
2763 }
2764}
2765
Shaohua Li4ed87312012-10-11 13:34:00 +11002766static void raid10d(struct md_thread *thread)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767{
Shaohua Li4ed87312012-10-11 13:34:00 +11002768 struct mddev *mddev = thread->mddev;
NeilBrown9f2c9d12011-10-11 16:48:43 +11002769 struct r10bio *r10_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770 unsigned long flags;
NeilBrowne879a872011-10-11 16:49:02 +11002771 struct r10conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 struct list_head *head = &conf->retry_list;
NeilBrowne1dfa0a2011-04-18 18:25:41 +10002773 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774
2775 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776
NeilBrowne1dfa0a2011-04-18 18:25:41 +10002777 blk_start_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 for (;;) {
NeilBrowna35e63e2008-03-04 14:29:29 -08002779
NeilBrown0021b7b2012-07-31 09:08:14 +02002780 flush_pending_writes(conf);
NeilBrowna35e63e2008-03-04 14:29:29 -08002781
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrowna35e63e2008-03-04 14:29:29 -08002783 if (list_empty(head)) {
NeilBrown6cce3b232006-01-06 00:20:16 -08002784 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785 break;
NeilBrowna35e63e2008-03-04 14:29:29 -08002786 }
NeilBrown9f2c9d12011-10-11 16:48:43 +11002787 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788 list_del(head->prev);
NeilBrown4443ae12006-01-06 00:20:28 -08002789 conf->nr_queued--;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 spin_unlock_irqrestore(&conf->device_lock, flags);
2791
2792 mddev = r10_bio->mddev;
NeilBrown070ec552009-06-16 16:54:21 +10002793 conf = mddev->private;
NeilBrownbd870a12011-07-28 11:39:24 +10002794 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2795 test_bit(R10BIO_WriteError, &r10_bio->state))
NeilBrown749c55e2011-07-28 11:39:24 +10002796 handle_write_completed(conf, r10_bio);
NeilBrown3ea7daa2012-05-22 13:53:47 +10002797 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2798 reshape_request_write(mddev, r10_bio);
NeilBrown749c55e2011-07-28 11:39:24 +10002799 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 sync_request_write(mddev, r10_bio);
Jens Axboe7eaceac2011-03-10 08:52:07 +01002801 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802 recovery_request_write(mddev, r10_bio);
NeilBrown856e08e2011-07-28 11:39:23 +10002803 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
NeilBrown560f8e52011-07-28 11:39:23 +10002804 handle_read_error(mddev, r10_bio);
NeilBrown856e08e2011-07-28 11:39:23 +10002805 else {
2806 /* just a partial read to be scheduled from a
2807 * separate context
2808 */
2809 int slot = r10_bio->read_slot;
2810 generic_make_request(r10_bio->devs[slot].bio);
2811 }
NeilBrown4443ae12006-01-06 00:20:28 -08002812
NeilBrown1d9d5242009-10-16 15:55:32 +11002813 cond_resched();
NeilBrownde393cd2011-07-28 11:31:48 +10002814 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2815 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816 }
NeilBrowne1dfa0a2011-04-18 18:25:41 +10002817 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818}
2819
2820
NeilBrowne879a872011-10-11 16:49:02 +11002821static int init_resync(struct r10conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822{
2823 int buffs;
NeilBrown69335ef2011-12-23 10:17:54 +11002824 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825
2826 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
Eric Sesterhennb6385482006-04-02 13:34:29 +02002827 BUG_ON(conf->r10buf_pool);
NeilBrown69335ef2011-12-23 10:17:54 +11002828 conf->have_replacement = 0;
NeilBrown5cf00fc2012-05-21 09:28:20 +10002829 for (i = 0; i < conf->geo.raid_disks; i++)
NeilBrown69335ef2011-12-23 10:17:54 +11002830 if (conf->mirrors[i].replacement)
2831 conf->have_replacement = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832 conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2833 if (!conf->r10buf_pool)
2834 return -ENOMEM;
2835 conf->next_resync = 0;
2836 return 0;
2837}
2838
2839/*
2840 * perform a "sync" on one "block"
2841 *
2842 * We need to make sure that no normal I/O request - particularly write
2843 * requests - conflict with active sync requests.
2844 *
2845 * This is achieved by tracking pending requests and a 'barrier' concept
2846 * that can be installed to exclude normal IO requests.
2847 *
2848 * Resync and recovery are handled very differently.
2849 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2850 *
2851 * For resync, we iterate over virtual addresses, read all copies,
2852 * and update if there are differences. If only one copy is live,
2853 * skip it.
2854 * For recovery, we iterate over physical addresses, read a good
2855 * value for each non-in_sync drive, and over-write.
2856 *
2857 * So, for recovery we may have several outstanding complex requests for a
2858 * given address, one for each out-of-sync device. We model this by allocating
2859 * a number of r10_bio structures, one for each out-of-sync device.
2860 * As we setup these structures, we collect all bio's together into a list
2861 * which we then process collectively to add pages, and then process again
2862 * to pass to generic_make_request.
2863 *
2864 * The r10_bio structures are linked using a borrowed master_bio pointer.
2865 * This link is counted in ->remaining. When the r10_bio that points to NULL
2866 * has its remaining count decremented to 0, the whole complex operation
2867 * is complete.
2868 *
2869 */
2870
NeilBrownfd01b882011-10-11 16:47:53 +11002871static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
NeilBrownab9d47e2011-05-11 14:54:41 +10002872 int *skipped, int go_faster)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873{
NeilBrowne879a872011-10-11 16:49:02 +11002874 struct r10conf *conf = mddev->private;
NeilBrown9f2c9d12011-10-11 16:48:43 +11002875 struct r10bio *r10_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876 struct bio *biolist = NULL, *bio;
2877 sector_t max_sector, nr_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878 int i;
NeilBrown6cce3b232006-01-06 00:20:16 -08002879 int max_sync;
NeilBrown57dab0b2010-10-19 10:03:39 +11002880 sector_t sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881 sector_t sectors_skipped = 0;
2882 int chunks_skipped = 0;
NeilBrown5cf00fc2012-05-21 09:28:20 +10002883 sector_t chunk_mask = conf->geo.chunk_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884
2885 if (!conf->r10buf_pool)
2886 if (init_resync(conf))
NeilBrown57afd892005-06-21 17:17:13 -07002887 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888
2889 skipped:
Andre Noll58c0fed2009-03-31 14:33:13 +11002890 max_sector = mddev->dev_sectors;
NeilBrown3ea7daa2012-05-22 13:53:47 +10002891 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2892 test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893 max_sector = mddev->resync_max_sectors;
2894 if (sector_nr >= max_sector) {
NeilBrown6cce3b232006-01-06 00:20:16 -08002895 /* If we aborted, we need to abort the
2896 * sync on the 'current' bitmap chucks (there can
2897 * be several when recovering multiple devices).
2898 * as we may have started syncing it but not finished.
2899 * We can find the current address in
2900 * mddev->curr_resync, but for recovery,
2901 * we need to convert that to several
2902 * virtual addresses.
2903 */
NeilBrown3ea7daa2012-05-22 13:53:47 +10002904 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2905 end_reshape(conf);
2906 return 0;
2907 }
2908
NeilBrown6cce3b232006-01-06 00:20:16 -08002909 if (mddev->curr_resync < max_sector) { /* aborted */
2910 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2911 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2912 &sync_blocks, 1);
NeilBrown5cf00fc2012-05-21 09:28:20 +10002913 else for (i = 0; i < conf->geo.raid_disks; i++) {
NeilBrown6cce3b232006-01-06 00:20:16 -08002914 sector_t sect =
2915 raid10_find_virt(conf, mddev->curr_resync, i);
2916 bitmap_end_sync(mddev->bitmap, sect,
2917 &sync_blocks, 1);
2918 }
NeilBrown9ad1aef2011-12-23 10:17:55 +11002919 } else {
2920 /* completed sync */
2921 if ((!mddev->bitmap || conf->fullsync)
2922 && conf->have_replacement
2923 && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2924 /* Completed a full sync so the replacements
2925 * are now fully recovered.
2926 */
NeilBrown5cf00fc2012-05-21 09:28:20 +10002927 for (i = 0; i < conf->geo.raid_disks; i++)
NeilBrown9ad1aef2011-12-23 10:17:55 +11002928 if (conf->mirrors[i].replacement)
2929 conf->mirrors[i].replacement
2930 ->recovery_offset
2931 = MaxSector;
2932 }
NeilBrown6cce3b232006-01-06 00:20:16 -08002933 conf->fullsync = 0;
NeilBrown9ad1aef2011-12-23 10:17:55 +11002934 }
NeilBrown6cce3b232006-01-06 00:20:16 -08002935 bitmap_close_sync(mddev->bitmap);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 close_sync(conf);
NeilBrown57afd892005-06-21 17:17:13 -07002937 *skipped = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 return sectors_skipped;
2939 }
NeilBrown3ea7daa2012-05-22 13:53:47 +10002940
2941 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2942 return reshape_request(mddev, sector_nr, skipped);
2943
NeilBrown5cf00fc2012-05-21 09:28:20 +10002944 if (chunks_skipped >= conf->geo.raid_disks) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945 /* if there has been nothing to do on any drive,
2946 * then there is nothing to do at all..
2947 */
NeilBrown57afd892005-06-21 17:17:13 -07002948 *skipped = 1;
2949 return (max_sector - sector_nr) + sectors_skipped;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 }
2951
NeilBrownc6207272008-02-06 01:39:52 -08002952 if (max_sector > mddev->resync_max)
2953 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2954
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955 /* make sure whole request will fit in a chunk - if chunks
2956 * are meaningful
2957 */
NeilBrown5cf00fc2012-05-21 09:28:20 +10002958 if (conf->geo.near_copies < conf->geo.raid_disks &&
2959 max_sector > (sector_nr | chunk_mask))
2960 max_sector = (sector_nr | chunk_mask) + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961 /*
2962 * If there is non-resync activity waiting for us then
2963 * put in a delay to throttle resync.
2964 */
NeilBrown0a27ec92006-01-06 00:20:13 -08002965 if (!go_faster && conf->nr_waiting)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002966 msleep_interruptible(1000);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002967
2968 /* Again, very different code for resync and recovery.
2969 * Both must result in an r10bio with a list of bios that
2970 * have bi_end_io, bi_sector, bi_bdev set,
2971 * and bi_private set to the r10bio.
2972 * For recovery, we may actually create several r10bios
2973 * with 2 bios in each, that correspond to the bios in the main one.
2974 * In this case, the subordinate r10bios link back through a
2975 * borrowed master_bio pointer, and the counter in the master
2976 * includes a ref from each subordinate.
2977 */
2978 /* First, we decide what to do and set ->bi_end_io
2979 * To end_sync_read if we want to read, and
2980 * end_sync_write if we will want to write.
2981 */
2982
NeilBrown6cce3b232006-01-06 00:20:16 -08002983 max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002984 if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2985 /* recovery... the complicated one */
NeilBrowne875ece2011-07-28 11:39:24 +10002986 int j;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002987 r10_bio = NULL;
2988
NeilBrown5cf00fc2012-05-21 09:28:20 +10002989 for (i = 0 ; i < conf->geo.raid_disks; i++) {
NeilBrownab9d47e2011-05-11 14:54:41 +10002990 int still_degraded;
NeilBrown9f2c9d12011-10-11 16:48:43 +11002991 struct r10bio *rb2;
NeilBrownab9d47e2011-05-11 14:54:41 +10002992 sector_t sect;
2993 int must_sync;
NeilBrowne875ece2011-07-28 11:39:24 +10002994 int any_working;
Jonathan Brassowdc280d982012-07-31 10:03:52 +10002995 struct raid10_info *mirror = &conf->mirrors[i];
NeilBrownab9d47e2011-05-11 14:54:41 +10002996
NeilBrown24afd802011-12-23 10:17:55 +11002997 if ((mirror->rdev == NULL ||
2998 test_bit(In_sync, &mirror->rdev->flags))
2999 &&
3000 (mirror->replacement == NULL ||
3001 test_bit(Faulty,
3002 &mirror->replacement->flags)))
NeilBrownab9d47e2011-05-11 14:54:41 +10003003 continue;
3004
3005 still_degraded = 0;
3006 /* want to reconstruct this device */
3007 rb2 = r10_bio;
3008 sect = raid10_find_virt(conf, sector_nr, i);
NeilBrownfc448a12012-07-03 10:37:30 +10003009 if (sect >= mddev->resync_max_sectors) {
3010 /* last stripe is not complete - don't
3011 * try to recover this sector.
3012 */
3013 continue;
3014 }
NeilBrown24afd802011-12-23 10:17:55 +11003015 /* Unless we are doing a full sync, or a replacement
3016 * we only need to recover the block if it is set in
3017 * the bitmap
NeilBrownab9d47e2011-05-11 14:54:41 +10003018 */
3019 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3020 &sync_blocks, 1);
3021 if (sync_blocks < max_sync)
3022 max_sync = sync_blocks;
3023 if (!must_sync &&
NeilBrown24afd802011-12-23 10:17:55 +11003024 mirror->replacement == NULL &&
NeilBrownab9d47e2011-05-11 14:54:41 +10003025 !conf->fullsync) {
3026 /* yep, skip the sync_blocks here, but don't assume
3027 * that there will never be anything to do here
NeilBrown6cce3b232006-01-06 00:20:16 -08003028 */
NeilBrownab9d47e2011-05-11 14:54:41 +10003029 chunks_skipped = -1;
3030 continue;
3031 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032
NeilBrownab9d47e2011-05-11 14:54:41 +10003033 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3034 raise_barrier(conf, rb2 != NULL);
3035 atomic_set(&r10_bio->remaining, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036
NeilBrownab9d47e2011-05-11 14:54:41 +10003037 r10_bio->master_bio = (struct bio*)rb2;
3038 if (rb2)
3039 atomic_inc(&rb2->remaining);
3040 r10_bio->mddev = mddev;
3041 set_bit(R10BIO_IsRecover, &r10_bio->state);
3042 r10_bio->sector = sect;
NeilBrown6cce3b232006-01-06 00:20:16 -08003043
NeilBrownab9d47e2011-05-11 14:54:41 +10003044 raid10_find_phys(conf, r10_bio);
NeilBrown18055562009-05-07 12:48:10 +10003045
NeilBrownab9d47e2011-05-11 14:54:41 +10003046 /* Need to check if the array will still be
3047 * degraded
3048 */
NeilBrown5cf00fc2012-05-21 09:28:20 +10003049 for (j = 0; j < conf->geo.raid_disks; j++)
NeilBrownab9d47e2011-05-11 14:54:41 +10003050 if (conf->mirrors[j].rdev == NULL ||
3051 test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3052 still_degraded = 1;
NeilBrown87fc7672005-09-09 16:24:04 -07003053 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003054 }
NeilBrownab9d47e2011-05-11 14:54:41 +10003055
3056 must_sync = bitmap_start_sync(mddev->bitmap, sect,
3057 &sync_blocks, still_degraded);
3058
NeilBrowne875ece2011-07-28 11:39:24 +10003059 any_working = 0;
NeilBrownab9d47e2011-05-11 14:54:41 +10003060 for (j=0; j<conf->copies;j++) {
NeilBrowne875ece2011-07-28 11:39:24 +10003061 int k;
NeilBrownab9d47e2011-05-11 14:54:41 +10003062 int d = r10_bio->devs[j].devnum;
NeilBrown5e570282011-07-28 11:39:25 +10003063 sector_t from_addr, to_addr;
NeilBrown3cb03002011-10-11 16:45:26 +11003064 struct md_rdev *rdev;
NeilBrown40c356c2011-07-28 11:39:24 +10003065 sector_t sector, first_bad;
3066 int bad_sectors;
NeilBrownab9d47e2011-05-11 14:54:41 +10003067 if (!conf->mirrors[d].rdev ||
3068 !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3069 continue;
3070 /* This is where we read from */
NeilBrowne875ece2011-07-28 11:39:24 +10003071 any_working = 1;
NeilBrown40c356c2011-07-28 11:39:24 +10003072 rdev = conf->mirrors[d].rdev;
3073 sector = r10_bio->devs[j].addr;
3074
3075 if (is_badblock(rdev, sector, max_sync,
3076 &first_bad, &bad_sectors)) {
3077 if (first_bad > sector)
3078 max_sync = first_bad - sector;
3079 else {
3080 bad_sectors -= (sector
3081 - first_bad);
3082 if (max_sync > bad_sectors)
3083 max_sync = bad_sectors;
3084 continue;
3085 }
3086 }
NeilBrownab9d47e2011-05-11 14:54:41 +10003087 bio = r10_bio->devs[0].bio;
3088 bio->bi_next = biolist;
3089 biolist = bio;
3090 bio->bi_private = r10_bio;
3091 bio->bi_end_io = end_sync_read;
3092 bio->bi_rw = READ;
NeilBrown5e570282011-07-28 11:39:25 +10003093 from_addr = r10_bio->devs[j].addr;
NeilBrown24afd802011-12-23 10:17:55 +11003094 bio->bi_sector = from_addr + rdev->data_offset;
3095 bio->bi_bdev = rdev->bdev;
3096 atomic_inc(&rdev->nr_pending);
3097 /* and we write to 'i' (if not in_sync) */
NeilBrownab9d47e2011-05-11 14:54:41 +10003098
3099 for (k=0; k<conf->copies; k++)
3100 if (r10_bio->devs[k].devnum == i)
3101 break;
3102 BUG_ON(k == conf->copies);
NeilBrown5e570282011-07-28 11:39:25 +10003103 to_addr = r10_bio->devs[k].addr;
NeilBrownab9d47e2011-05-11 14:54:41 +10003104 r10_bio->devs[0].devnum = d;
NeilBrown5e570282011-07-28 11:39:25 +10003105 r10_bio->devs[0].addr = from_addr;
NeilBrownab9d47e2011-05-11 14:54:41 +10003106 r10_bio->devs[1].devnum = i;
NeilBrown5e570282011-07-28 11:39:25 +10003107 r10_bio->devs[1].addr = to_addr;
NeilBrownab9d47e2011-05-11 14:54:41 +10003108
NeilBrown24afd802011-12-23 10:17:55 +11003109 rdev = mirror->rdev;
3110 if (!test_bit(In_sync, &rdev->flags)) {
3111 bio = r10_bio->devs[1].bio;
3112 bio->bi_next = biolist;
3113 biolist = bio;
3114 bio->bi_private = r10_bio;
3115 bio->bi_end_io = end_sync_write;
3116 bio->bi_rw = WRITE;
3117 bio->bi_sector = to_addr
3118 + rdev->data_offset;
3119 bio->bi_bdev = rdev->bdev;
3120 atomic_inc(&r10_bio->remaining);
3121 } else
3122 r10_bio->devs[1].bio->bi_end_io = NULL;
3123
3124 /* and maybe write to replacement */
3125 bio = r10_bio->devs[1].repl_bio;
3126 if (bio)
3127 bio->bi_end_io = NULL;
3128 rdev = mirror->replacement;
3129 /* Note: if rdev != NULL, then bio
3130 * cannot be NULL as r10buf_pool_alloc will
3131 * have allocated it.
3132 * So the second test here is pointless.
3133 * But it keeps semantic-checkers happy, and
3134 * this comment keeps human reviewers
3135 * happy.
3136 */
3137 if (rdev == NULL || bio == NULL ||
3138 test_bit(Faulty, &rdev->flags))
3139 break;
3140 bio->bi_next = biolist;
3141 biolist = bio;
3142 bio->bi_private = r10_bio;
3143 bio->bi_end_io = end_sync_write;
3144 bio->bi_rw = WRITE;
3145 bio->bi_sector = to_addr + rdev->data_offset;
3146 bio->bi_bdev = rdev->bdev;
3147 atomic_inc(&r10_bio->remaining);
NeilBrownab9d47e2011-05-11 14:54:41 +10003148 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003149 }
NeilBrownab9d47e2011-05-11 14:54:41 +10003150 if (j == conf->copies) {
NeilBrowne875ece2011-07-28 11:39:24 +10003151 /* Cannot recover, so abort the recovery or
3152 * record a bad block */
NeilBrownab9d47e2011-05-11 14:54:41 +10003153 put_buf(r10_bio);
3154 if (rb2)
3155 atomic_dec(&rb2->remaining);
3156 r10_bio = rb2;
NeilBrowne875ece2011-07-28 11:39:24 +10003157 if (any_working) {
3158 /* problem is that there are bad blocks
3159 * on other device(s)
3160 */
3161 int k;
3162 for (k = 0; k < conf->copies; k++)
3163 if (r10_bio->devs[k].devnum == i)
3164 break;
NeilBrown24afd802011-12-23 10:17:55 +11003165 if (!test_bit(In_sync,
3166 &mirror->rdev->flags)
3167 && !rdev_set_badblocks(
3168 mirror->rdev,
3169 r10_bio->devs[k].addr,
3170 max_sync, 0))
3171 any_working = 0;
3172 if (mirror->replacement &&
3173 !rdev_set_badblocks(
3174 mirror->replacement,
NeilBrowne875ece2011-07-28 11:39:24 +10003175 r10_bio->devs[k].addr,
3176 max_sync, 0))
3177 any_working = 0;
3178 }
3179 if (!any_working) {
3180 if (!test_and_set_bit(MD_RECOVERY_INTR,
3181 &mddev->recovery))
3182 printk(KERN_INFO "md/raid10:%s: insufficient "
3183 "working devices for recovery.\n",
3184 mdname(mddev));
NeilBrown24afd802011-12-23 10:17:55 +11003185 mirror->recovery_disabled
NeilBrowne875ece2011-07-28 11:39:24 +10003186 = mddev->recovery_disabled;
3187 }
NeilBrownab9d47e2011-05-11 14:54:41 +10003188 break;
3189 }
3190 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003191 if (biolist == NULL) {
3192 while (r10_bio) {
NeilBrown9f2c9d12011-10-11 16:48:43 +11003193 struct r10bio *rb2 = r10_bio;
3194 r10_bio = (struct r10bio*) rb2->master_bio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003195 rb2->master_bio = NULL;
3196 put_buf(rb2);
3197 }
3198 goto giveup;
3199 }
3200 } else {
3201 /* resync. Schedule a read for every block at this virt offset */
3202 int count = 0;
NeilBrown6cce3b232006-01-06 00:20:16 -08003203
NeilBrown78200d42009-02-25 13:18:47 +11003204 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3205
NeilBrown6cce3b232006-01-06 00:20:16 -08003206 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3207 &sync_blocks, mddev->degraded) &&
NeilBrownab9d47e2011-05-11 14:54:41 +10003208 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3209 &mddev->recovery)) {
NeilBrown6cce3b232006-01-06 00:20:16 -08003210 /* We can skip this block */
3211 *skipped = 1;
3212 return sync_blocks + sectors_skipped;
3213 }
3214 if (sync_blocks < max_sync)
3215 max_sync = sync_blocks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3217
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218 r10_bio->mddev = mddev;
3219 atomic_set(&r10_bio->remaining, 0);
NeilBrown6cce3b232006-01-06 00:20:16 -08003220 raise_barrier(conf, 0);
3221 conf->next_resync = sector_nr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003222
3223 r10_bio->master_bio = NULL;
3224 r10_bio->sector = sector_nr;
3225 set_bit(R10BIO_IsSync, &r10_bio->state);
3226 raid10_find_phys(conf, r10_bio);
NeilBrown5cf00fc2012-05-21 09:28:20 +10003227 r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003228
NeilBrown5cf00fc2012-05-21 09:28:20 +10003229 for (i = 0; i < conf->copies; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230 int d = r10_bio->devs[i].devnum;
NeilBrown40c356c2011-07-28 11:39:24 +10003231 sector_t first_bad, sector;
3232 int bad_sectors;
3233
NeilBrown9ad1aef2011-12-23 10:17:55 +11003234 if (r10_bio->devs[i].repl_bio)
3235 r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3236
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237 bio = r10_bio->devs[i].bio;
3238 bio->bi_end_io = NULL;
NeilBrownaf03b8e2007-06-16 10:16:06 -07003239 clear_bit(BIO_UPTODATE, &bio->bi_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003240 if (conf->mirrors[d].rdev == NULL ||
NeilBrownb2d444d2005-11-08 21:39:31 -08003241 test_bit(Faulty, &conf->mirrors[d].rdev->flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242 continue;
NeilBrown40c356c2011-07-28 11:39:24 +10003243 sector = r10_bio->devs[i].addr;
3244 if (is_badblock(conf->mirrors[d].rdev,
3245 sector, max_sync,
3246 &first_bad, &bad_sectors)) {
3247 if (first_bad > sector)
3248 max_sync = first_bad - sector;
3249 else {
3250 bad_sectors -= (sector - first_bad);
3251 if (max_sync > bad_sectors)
Dan Carpenter91502f02012-10-11 14:20:58 +11003252 max_sync = bad_sectors;
NeilBrown40c356c2011-07-28 11:39:24 +10003253 continue;
3254 }
3255 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3257 atomic_inc(&r10_bio->remaining);
3258 bio->bi_next = biolist;
3259 biolist = bio;
3260 bio->bi_private = r10_bio;
3261 bio->bi_end_io = end_sync_read;
NeilBrown802ba062006-12-13 00:34:13 -08003262 bio->bi_rw = READ;
NeilBrown40c356c2011-07-28 11:39:24 +10003263 bio->bi_sector = sector +
Linus Torvalds1da177e2005-04-16 15:20:36 -07003264 conf->mirrors[d].rdev->data_offset;
3265 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3266 count++;
NeilBrown9ad1aef2011-12-23 10:17:55 +11003267
3268 if (conf->mirrors[d].replacement == NULL ||
3269 test_bit(Faulty,
3270 &conf->mirrors[d].replacement->flags))
3271 continue;
3272
3273 /* Need to set up for writing to the replacement */
3274 bio = r10_bio->devs[i].repl_bio;
3275 clear_bit(BIO_UPTODATE, &bio->bi_flags);
3276
3277 sector = r10_bio->devs[i].addr;
3278 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3279 bio->bi_next = biolist;
3280 biolist = bio;
3281 bio->bi_private = r10_bio;
3282 bio->bi_end_io = end_sync_write;
3283 bio->bi_rw = WRITE;
3284 bio->bi_sector = sector +
3285 conf->mirrors[d].replacement->data_offset;
3286 bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3287 count++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003288 }
3289
3290 if (count < 2) {
3291 for (i=0; i<conf->copies; i++) {
3292 int d = r10_bio->devs[i].devnum;
3293 if (r10_bio->devs[i].bio->bi_end_io)
NeilBrownab9d47e2011-05-11 14:54:41 +10003294 rdev_dec_pending(conf->mirrors[d].rdev,
3295 mddev);
NeilBrown9ad1aef2011-12-23 10:17:55 +11003296 if (r10_bio->devs[i].repl_bio &&
3297 r10_bio->devs[i].repl_bio->bi_end_io)
3298 rdev_dec_pending(
3299 conf->mirrors[d].replacement,
3300 mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003301 }
3302 put_buf(r10_bio);
3303 biolist = NULL;
3304 goto giveup;
3305 }
3306 }
3307
3308 for (bio = biolist; bio ; bio=bio->bi_next) {
3309
3310 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
3311 if (bio->bi_end_io)
3312 bio->bi_flags |= 1 << BIO_UPTODATE;
3313 bio->bi_vcnt = 0;
3314 bio->bi_idx = 0;
3315 bio->bi_phys_segments = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003316 bio->bi_size = 0;
3317 }
3318
3319 nr_sectors = 0;
NeilBrown6cce3b232006-01-06 00:20:16 -08003320 if (sector_nr + max_sync < max_sector)
3321 max_sector = sector_nr + max_sync;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322 do {
3323 struct page *page;
3324 int len = PAGE_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003325 if (sector_nr + (len>>9) > max_sector)
3326 len = (max_sector - sector_nr) << 9;
3327 if (len == 0)
3328 break;
3329 for (bio= biolist ; bio ; bio=bio->bi_next) {
NeilBrownab9d47e2011-05-11 14:54:41 +10003330 struct bio *bio2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003331 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
NeilBrownab9d47e2011-05-11 14:54:41 +10003332 if (bio_add_page(bio, page, len, 0))
3333 continue;
3334
3335 /* stop here */
3336 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3337 for (bio2 = biolist;
3338 bio2 && bio2 != bio;
3339 bio2 = bio2->bi_next) {
3340 /* remove last page from this bio */
3341 bio2->bi_vcnt--;
3342 bio2->bi_size -= len;
3343 bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003344 }
NeilBrownab9d47e2011-05-11 14:54:41 +10003345 goto bio_full;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346 }
3347 nr_sectors += len>>9;
3348 sector_nr += len>>9;
3349 } while (biolist->bi_vcnt < RESYNC_PAGES);
3350 bio_full:
3351 r10_bio->sectors = nr_sectors;
3352
3353 while (biolist) {
3354 bio = biolist;
3355 biolist = biolist->bi_next;
3356
3357 bio->bi_next = NULL;
3358 r10_bio = bio->bi_private;
3359 r10_bio->sectors = nr_sectors;
3360
3361 if (bio->bi_end_io == end_sync_read) {
3362 md_sync_acct(bio->bi_bdev, nr_sectors);
3363 generic_make_request(bio);
3364 }
3365 }
3366
NeilBrown57afd892005-06-21 17:17:13 -07003367 if (sectors_skipped)
3368 /* pretend they weren't skipped, it makes
3369 * no important difference in this case
3370 */
3371 md_done_sync(mddev, sectors_skipped, 1);
3372
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373 return sectors_skipped + nr_sectors;
3374 giveup:
3375 /* There is nowhere to write, so all non-sync
NeilBrowne875ece2011-07-28 11:39:24 +10003376 * drives must be failed or in resync, all drives
3377 * have a bad block, so try the next chunk...
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378 */
NeilBrown09b40682009-02-25 13:18:47 +11003379 if (sector_nr + max_sync < max_sector)
3380 max_sector = sector_nr + max_sync;
3381
3382 sectors_skipped += (max_sector - sector_nr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383 chunks_skipped ++;
3384 sector_nr = max_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385 goto skipped;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386}
3387
Dan Williams80c3a6c2009-03-17 18:10:40 -07003388static sector_t
NeilBrownfd01b882011-10-11 16:47:53 +11003389raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
Dan Williams80c3a6c2009-03-17 18:10:40 -07003390{
3391 sector_t size;
NeilBrowne879a872011-10-11 16:49:02 +11003392 struct r10conf *conf = mddev->private;
Dan Williams80c3a6c2009-03-17 18:10:40 -07003393
3394 if (!raid_disks)
NeilBrown3ea7daa2012-05-22 13:53:47 +10003395 raid_disks = min(conf->geo.raid_disks,
3396 conf->prev.raid_disks);
Dan Williams80c3a6c2009-03-17 18:10:40 -07003397 if (!sectors)
Trela, Maciejdab8b292010-03-08 16:02:45 +11003398 sectors = conf->dev_sectors;
Dan Williams80c3a6c2009-03-17 18:10:40 -07003399
NeilBrown5cf00fc2012-05-21 09:28:20 +10003400 size = sectors >> conf->geo.chunk_shift;
3401 sector_div(size, conf->geo.far_copies);
Dan Williams80c3a6c2009-03-17 18:10:40 -07003402 size = size * raid_disks;
NeilBrown5cf00fc2012-05-21 09:28:20 +10003403 sector_div(size, conf->geo.near_copies);
Dan Williams80c3a6c2009-03-17 18:10:40 -07003404
NeilBrown5cf00fc2012-05-21 09:28:20 +10003405 return size << conf->geo.chunk_shift;
Dan Williams80c3a6c2009-03-17 18:10:40 -07003406}
3407
NeilBrown6508fdb2012-05-17 10:08:45 +10003408static void calc_sectors(struct r10conf *conf, sector_t size)
3409{
3410 /* Calculate the number of sectors-per-device that will
3411 * actually be used, and set conf->dev_sectors and
3412 * conf->stride
3413 */
3414
NeilBrown5cf00fc2012-05-21 09:28:20 +10003415 size = size >> conf->geo.chunk_shift;
3416 sector_div(size, conf->geo.far_copies);
3417 size = size * conf->geo.raid_disks;
3418 sector_div(size, conf->geo.near_copies);
NeilBrown6508fdb2012-05-17 10:08:45 +10003419 /* 'size' is now the number of chunks in the array */
3420 /* calculate "used chunks per device" */
3421 size = size * conf->copies;
3422
3423 /* We need to round up when dividing by raid_disks to
3424 * get the stride size.
3425 */
NeilBrown5cf00fc2012-05-21 09:28:20 +10003426 size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
NeilBrown6508fdb2012-05-17 10:08:45 +10003427
NeilBrown5cf00fc2012-05-21 09:28:20 +10003428 conf->dev_sectors = size << conf->geo.chunk_shift;
NeilBrown6508fdb2012-05-17 10:08:45 +10003429
NeilBrown5cf00fc2012-05-21 09:28:20 +10003430 if (conf->geo.far_offset)
3431 conf->geo.stride = 1 << conf->geo.chunk_shift;
NeilBrown6508fdb2012-05-17 10:08:45 +10003432 else {
NeilBrown5cf00fc2012-05-21 09:28:20 +10003433 sector_div(size, conf->geo.far_copies);
3434 conf->geo.stride = size << conf->geo.chunk_shift;
NeilBrown6508fdb2012-05-17 10:08:45 +10003435 }
3436}
Trela, Maciejdab8b292010-03-08 16:02:45 +11003437
NeilBrowndeb200d2012-05-21 09:28:33 +10003438enum geo_type {geo_new, geo_old, geo_start};
3439static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3440{
3441 int nc, fc, fo;
3442 int layout, chunk, disks;
3443 switch (new) {
3444 case geo_old:
3445 layout = mddev->layout;
3446 chunk = mddev->chunk_sectors;
3447 disks = mddev->raid_disks - mddev->delta_disks;
3448 break;
3449 case geo_new:
3450 layout = mddev->new_layout;
3451 chunk = mddev->new_chunk_sectors;
3452 disks = mddev->raid_disks;
3453 break;
3454 default: /* avoid 'may be unused' warnings */
3455 case geo_start: /* new when starting reshape - raid_disks not
3456 * updated yet. */
3457 layout = mddev->new_layout;
3458 chunk = mddev->new_chunk_sectors;
3459 disks = mddev->raid_disks + mddev->delta_disks;
3460 break;
3461 }
Jonathan Brassow475901a2013-02-21 13:28:10 +11003462 if (layout >> 18)
NeilBrowndeb200d2012-05-21 09:28:33 +10003463 return -1;
3464 if (chunk < (PAGE_SIZE >> 9) ||
3465 !is_power_of_2(chunk))
3466 return -2;
3467 nc = layout & 255;
3468 fc = (layout >> 8) & 255;
3469 fo = layout & (1<<16);
3470 geo->raid_disks = disks;
3471 geo->near_copies = nc;
3472 geo->far_copies = fc;
3473 geo->far_offset = fo;
Jonathan Brassow475901a2013-02-21 13:28:10 +11003474 geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
NeilBrowndeb200d2012-05-21 09:28:33 +10003475 geo->chunk_mask = chunk - 1;
3476 geo->chunk_shift = ffz(~chunk);
3477 return nc*fc;
3478}
3479
NeilBrowne879a872011-10-11 16:49:02 +11003480static struct r10conf *setup_conf(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481{
NeilBrowne879a872011-10-11 16:49:02 +11003482 struct r10conf *conf = NULL;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003483 int err = -EINVAL;
NeilBrowndeb200d2012-05-21 09:28:33 +10003484 struct geom geo;
3485 int copies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003486
NeilBrowndeb200d2012-05-21 09:28:33 +10003487 copies = setup_geo(&geo, mddev, geo_new);
3488
3489 if (copies == -2) {
NeilBrown128595e2010-05-03 14:47:14 +10003490 printk(KERN_ERR "md/raid10:%s: chunk size must be "
3491 "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3492 mdname(mddev), PAGE_SIZE);
Trela, Maciejdab8b292010-03-08 16:02:45 +11003493 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 }
NeilBrown2604b702006-01-06 00:20:36 -08003495
NeilBrowndeb200d2012-05-21 09:28:33 +10003496 if (copies < 2 || copies > mddev->raid_disks) {
NeilBrown128595e2010-05-03 14:47:14 +10003497 printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
Maciej Trelaf73ea872010-06-16 11:46:29 +01003498 mdname(mddev), mddev->new_layout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499 goto out;
3500 }
Trela, Maciejdab8b292010-03-08 16:02:45 +11003501
3502 err = -ENOMEM;
NeilBrowne879a872011-10-11 16:49:02 +11003503 conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
Trela, Maciejdab8b292010-03-08 16:02:45 +11003504 if (!conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 goto out;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003506
NeilBrown3ea7daa2012-05-22 13:53:47 +10003507 /* FIXME calc properly */
Jonathan Brassowdc280d982012-07-31 10:03:52 +10003508 conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
NeilBrown3ea7daa2012-05-22 13:53:47 +10003509 max(0,mddev->delta_disks)),
Trela, Maciejdab8b292010-03-08 16:02:45 +11003510 GFP_KERNEL);
3511 if (!conf->mirrors)
3512 goto out;
NeilBrown4443ae12006-01-06 00:20:28 -08003513
3514 conf->tmppage = alloc_page(GFP_KERNEL);
3515 if (!conf->tmppage)
Trela, Maciejdab8b292010-03-08 16:02:45 +11003516 goto out;
3517
NeilBrowndeb200d2012-05-21 09:28:33 +10003518 conf->geo = geo;
3519 conf->copies = copies;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003520 conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3521 r10bio_pool_free, conf);
3522 if (!conf->r10bio_pool)
3523 goto out;
3524
NeilBrown6508fdb2012-05-17 10:08:45 +10003525 calc_sectors(conf, mddev->dev_sectors);
NeilBrown3ea7daa2012-05-22 13:53:47 +10003526 if (mddev->reshape_position == MaxSector) {
3527 conf->prev = conf->geo;
3528 conf->reshape_progress = MaxSector;
3529 } else {
3530 if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3531 err = -EINVAL;
3532 goto out;
3533 }
3534 conf->reshape_progress = mddev->reshape_position;
3535 if (conf->prev.far_offset)
3536 conf->prev.stride = 1 << conf->prev.chunk_shift;
3537 else
3538 /* far_copies must be 1 */
3539 conf->prev.stride = conf->dev_sectors;
3540 }
Neil Browne7e72bf2008-05-14 16:05:54 -07003541 spin_lock_init(&conf->device_lock);
Trela, Maciejdab8b292010-03-08 16:02:45 +11003542 INIT_LIST_HEAD(&conf->retry_list);
3543
3544 spin_lock_init(&conf->resync_lock);
3545 init_waitqueue_head(&conf->wait_barrier);
3546
NeilBrown02326052012-07-03 15:56:52 +10003547 conf->thread = md_register_thread(raid10d, mddev, "raid10");
Trela, Maciejdab8b292010-03-08 16:02:45 +11003548 if (!conf->thread)
3549 goto out;
3550
Trela, Maciejdab8b292010-03-08 16:02:45 +11003551 conf->mddev = mddev;
3552 return conf;
3553
3554 out:
NeilBrown3ea7daa2012-05-22 13:53:47 +10003555 if (err == -ENOMEM)
3556 printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3557 mdname(mddev));
Trela, Maciejdab8b292010-03-08 16:02:45 +11003558 if (conf) {
3559 if (conf->r10bio_pool)
3560 mempool_destroy(conf->r10bio_pool);
3561 kfree(conf->mirrors);
3562 safe_put_page(conf->tmppage);
3563 kfree(conf);
3564 }
3565 return ERR_PTR(err);
3566}
3567
NeilBrownfd01b882011-10-11 16:47:53 +11003568static int run(struct mddev *mddev)
Trela, Maciejdab8b292010-03-08 16:02:45 +11003569{
NeilBrowne879a872011-10-11 16:49:02 +11003570 struct r10conf *conf;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003571 int i, disk_idx, chunk_size;
Jonathan Brassowdc280d982012-07-31 10:03:52 +10003572 struct raid10_info *disk;
NeilBrown3cb03002011-10-11 16:45:26 +11003573 struct md_rdev *rdev;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003574 sector_t size;
NeilBrown3ea7daa2012-05-22 13:53:47 +10003575 sector_t min_offset_diff = 0;
3576 int first = 1;
Shaohua Li532a2a32012-10-11 13:30:52 +11003577 bool discard_supported = false;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003578
3579 if (mddev->private == NULL) {
3580 conf = setup_conf(mddev);
3581 if (IS_ERR(conf))
3582 return PTR_ERR(conf);
3583 mddev->private = conf;
3584 }
3585 conf = mddev->private;
3586 if (!conf)
3587 goto out;
3588
Trela, Maciejdab8b292010-03-08 16:02:45 +11003589 mddev->thread = conf->thread;
3590 conf->thread = NULL;
3591
Martin K. Petersen8f6c2e42009-07-01 11:13:45 +10003592 chunk_size = mddev->chunk_sectors << 9;
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +10003593 if (mddev->queue) {
Shaohua Li532a2a32012-10-11 13:30:52 +11003594 blk_queue_max_discard_sectors(mddev->queue,
3595 mddev->chunk_sectors);
Joe Lawrencec8dc9c62013-02-21 13:28:09 +11003596 blk_queue_max_write_same_sectors(mddev->queue,
3597 mddev->chunk_sectors);
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +10003598 blk_queue_io_min(mddev->queue, chunk_size);
3599 if (conf->geo.raid_disks % conf->geo.near_copies)
3600 blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3601 else
3602 blk_queue_io_opt(mddev->queue, chunk_size *
3603 (conf->geo.raid_disks / conf->geo.near_copies));
3604 }
Martin K. Petersen8f6c2e42009-07-01 11:13:45 +10003605
NeilBrowndafb20f2012-03-19 12:46:39 +11003606 rdev_for_each(rdev, mddev) {
NeilBrown3ea7daa2012-05-22 13:53:47 +10003607 long long diff;
NeilBrownaba336b2012-05-31 15:39:11 +10003608 struct request_queue *q;
NeilBrown34b343c2011-07-28 11:31:47 +10003609
Linus Torvalds1da177e2005-04-16 15:20:36 -07003610 disk_idx = rdev->raid_disk;
NeilBrownf8c9e742012-05-21 09:28:33 +10003611 if (disk_idx < 0)
3612 continue;
3613 if (disk_idx >= conf->geo.raid_disks &&
3614 disk_idx >= conf->prev.raid_disks)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003615 continue;
3616 disk = conf->mirrors + disk_idx;
3617
NeilBrown56a25592011-12-23 10:17:55 +11003618 if (test_bit(Replacement, &rdev->flags)) {
3619 if (disk->replacement)
3620 goto out_free_conf;
3621 disk->replacement = rdev;
3622 } else {
3623 if (disk->rdev)
3624 goto out_free_conf;
3625 disk->rdev = rdev;
3626 }
NeilBrownaba336b2012-05-31 15:39:11 +10003627 q = bdev_get_queue(rdev->bdev);
3628 if (q->merge_bvec_fn)
3629 mddev->merge_check_needed = 1;
NeilBrown3ea7daa2012-05-22 13:53:47 +10003630 diff = (rdev->new_data_offset - rdev->data_offset);
3631 if (!mddev->reshape_backwards)
3632 diff = -diff;
3633 if (diff < 0)
3634 diff = 0;
3635 if (first || diff < min_offset_diff)
3636 min_offset_diff = diff;
NeilBrown56a25592011-12-23 10:17:55 +11003637
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +10003638 if (mddev->gendisk)
3639 disk_stack_limits(mddev->gendisk, rdev->bdev,
3640 rdev->data_offset << 9);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003641
3642 disk->head_position = 0;
Shaohua Li532a2a32012-10-11 13:30:52 +11003643
3644 if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3645 discard_supported = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003646 }
NeilBrown3ea7daa2012-05-22 13:53:47 +10003647
Jonathan Brassowed30be02012-10-31 11:42:30 +11003648 if (mddev->queue) {
3649 if (discard_supported)
3650 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3651 mddev->queue);
3652 else
3653 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3654 mddev->queue);
3655 }
NeilBrown6d508242005-09-09 16:24:03 -07003656 /* need to check that every block has at least one working mirror */
NeilBrown700c7212011-07-27 11:00:36 +10003657 if (!enough(conf, -1)) {
NeilBrown128595e2010-05-03 14:47:14 +10003658 printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
NeilBrown6d508242005-09-09 16:24:03 -07003659 mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660 goto out_free_conf;
3661 }
3662
NeilBrown3ea7daa2012-05-22 13:53:47 +10003663 if (conf->reshape_progress != MaxSector) {
3664 /* must ensure that shape change is supported */
3665 if (conf->geo.far_copies != 1 &&
3666 conf->geo.far_offset == 0)
3667 goto out_free_conf;
3668 if (conf->prev.far_copies != 1 &&
3669 conf->geo.far_offset == 0)
3670 goto out_free_conf;
3671 }
3672
Linus Torvalds1da177e2005-04-16 15:20:36 -07003673 mddev->degraded = 0;
NeilBrownf8c9e742012-05-21 09:28:33 +10003674 for (i = 0;
3675 i < conf->geo.raid_disks
3676 || i < conf->prev.raid_disks;
3677 i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678
3679 disk = conf->mirrors + i;
3680
NeilBrown56a25592011-12-23 10:17:55 +11003681 if (!disk->rdev && disk->replacement) {
3682 /* The replacement is all we have - use it */
3683 disk->rdev = disk->replacement;
3684 disk->replacement = NULL;
3685 clear_bit(Replacement, &disk->rdev->flags);
3686 }
3687
NeilBrown5fd6c1d2006-06-26 00:27:40 -07003688 if (!disk->rdev ||
NeilBrown2e333e82006-10-21 10:24:07 -07003689 !test_bit(In_sync, &disk->rdev->flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003690 disk->head_position = 0;
3691 mddev->degraded++;
Neil Brown8c2e8702008-06-28 08:30:52 +10003692 if (disk->rdev)
3693 conf->fullsync = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694 }
NeilBrownd890fa22011-10-26 11:54:39 +11003695 disk->recovery_disabled = mddev->recovery_disabled - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003696 }
3697
Andre Noll8c6ac862009-06-18 08:48:06 +10003698 if (mddev->recovery_cp != MaxSector)
NeilBrown128595e2010-05-03 14:47:14 +10003699 printk(KERN_NOTICE "md/raid10:%s: not clean"
Andre Noll8c6ac862009-06-18 08:48:06 +10003700 " -- starting background reconstruction\n",
3701 mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702 printk(KERN_INFO
NeilBrown128595e2010-05-03 14:47:14 +10003703 "md/raid10:%s: active with %d out of %d devices\n",
NeilBrown5cf00fc2012-05-21 09:28:20 +10003704 mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3705 conf->geo.raid_disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706 /*
3707 * Ok, everything is just fine now
3708 */
Trela, Maciejdab8b292010-03-08 16:02:45 +11003709 mddev->dev_sectors = conf->dev_sectors;
3710 size = raid10_size(mddev, 0, 0);
3711 md_set_array_sectors(mddev, size);
3712 mddev->resync_max_sectors = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003713
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +10003714 if (mddev->queue) {
NeilBrown5cf00fc2012-05-21 09:28:20 +10003715 int stripe = conf->geo.raid_disks *
Andre Noll9d8f0362009-06-18 08:45:01 +10003716 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +10003717 mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3718 mddev->queue->backing_dev_info.congested_data = mddev;
3719
3720 /* Calculate max read-ahead size.
3721 * We need to readahead at least twice a whole stripe....
3722 * maybe...
3723 */
NeilBrown5cf00fc2012-05-21 09:28:20 +10003724 stripe /= conf->geo.near_copies;
NeilBrown3ea7daa2012-05-22 13:53:47 +10003725 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3726 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +10003727 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003728 }
3729
Martin K. Petersena91a2782011-03-17 11:11:05 +01003730
3731 if (md_integrity_register(mddev))
3732 goto out_free_conf;
3733
NeilBrown3ea7daa2012-05-22 13:53:47 +10003734 if (conf->reshape_progress != MaxSector) {
3735 unsigned long before_length, after_length;
3736
3737 before_length = ((1 << conf->prev.chunk_shift) *
3738 conf->prev.far_copies);
3739 after_length = ((1 << conf->geo.chunk_shift) *
3740 conf->geo.far_copies);
3741
3742 if (max(before_length, after_length) > min_offset_diff) {
3743 /* This cannot work */
3744 printk("md/raid10: offset difference not enough to continue reshape\n");
3745 goto out_free_conf;
3746 }
3747 conf->offset_diff = min_offset_diff;
3748
3749 conf->reshape_safe = conf->reshape_progress;
3750 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3751 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3752 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3753 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3754 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3755 "reshape");
3756 }
3757
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758 return 0;
3759
3760out_free_conf:
NeilBrown01f96c02011-09-21 15:30:20 +10003761 md_unregister_thread(&mddev->thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003762 if (conf->r10bio_pool)
3763 mempool_destroy(conf->r10bio_pool);
NeilBrown1345b1d2006-01-06 00:20:40 -08003764 safe_put_page(conf->tmppage);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07003765 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766 kfree(conf);
3767 mddev->private = NULL;
3768out:
3769 return -EIO;
3770}
3771
NeilBrownfd01b882011-10-11 16:47:53 +11003772static int stop(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003773{
NeilBrowne879a872011-10-11 16:49:02 +11003774 struct r10conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775
NeilBrown409c57f2009-03-31 14:39:39 +11003776 raise_barrier(conf, 0);
3777 lower_barrier(conf);
3778
NeilBrown01f96c02011-09-21 15:30:20 +10003779 md_unregister_thread(&mddev->thread);
Jonathan Brassowcc4d1ef2012-07-31 10:03:53 +10003780 if (mddev->queue)
3781 /* the unplug fn references 'conf'*/
3782 blk_sync_queue(mddev->queue);
3783
Linus Torvalds1da177e2005-04-16 15:20:36 -07003784 if (conf->r10bio_pool)
3785 mempool_destroy(conf->r10bio_pool);
Jesper Juhl990a8ba2005-06-21 17:17:30 -07003786 kfree(conf->mirrors);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003787 kfree(conf);
3788 mddev->private = NULL;
3789 return 0;
3790}
3791
NeilBrownfd01b882011-10-11 16:47:53 +11003792static void raid10_quiesce(struct mddev *mddev, int state)
NeilBrown6cce3b232006-01-06 00:20:16 -08003793{
NeilBrowne879a872011-10-11 16:49:02 +11003794 struct r10conf *conf = mddev->private;
NeilBrown6cce3b232006-01-06 00:20:16 -08003795
3796 switch(state) {
3797 case 1:
3798 raise_barrier(conf, 0);
3799 break;
3800 case 0:
3801 lower_barrier(conf);
3802 break;
3803 }
NeilBrown6cce3b232006-01-06 00:20:16 -08003804}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003805
NeilBrown006a09a2012-03-19 12:46:40 +11003806static int raid10_resize(struct mddev *mddev, sector_t sectors)
3807{
3808 /* Resize of 'far' arrays is not supported.
3809 * For 'near' and 'offset' arrays we can set the
3810 * number of sectors used to be an appropriate multiple
3811 * of the chunk size.
3812 * For 'offset', this is far_copies*chunksize.
3813 * For 'near' the multiplier is the LCM of
3814 * near_copies and raid_disks.
3815 * So if far_copies > 1 && !far_offset, fail.
3816 * Else find LCM(raid_disks, near_copy)*far_copies and
3817 * multiply by chunk_size. Then round to this number.
3818 * This is mostly done by raid10_size()
3819 */
3820 struct r10conf *conf = mddev->private;
3821 sector_t oldsize, size;
3822
NeilBrownf8c9e742012-05-21 09:28:33 +10003823 if (mddev->reshape_position != MaxSector)
3824 return -EBUSY;
3825
NeilBrown5cf00fc2012-05-21 09:28:20 +10003826 if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
NeilBrown006a09a2012-03-19 12:46:40 +11003827 return -EINVAL;
3828
3829 oldsize = raid10_size(mddev, 0, 0);
3830 size = raid10_size(mddev, sectors, 0);
NeilBrowna4a61252012-05-22 13:55:27 +10003831 if (mddev->external_size &&
3832 mddev->array_sectors > size)
NeilBrown006a09a2012-03-19 12:46:40 +11003833 return -EINVAL;
NeilBrowna4a61252012-05-22 13:55:27 +10003834 if (mddev->bitmap) {
3835 int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3836 if (ret)
3837 return ret;
3838 }
3839 md_set_array_sectors(mddev, size);
NeilBrown006a09a2012-03-19 12:46:40 +11003840 set_capacity(mddev->gendisk, mddev->array_sectors);
3841 revalidate_disk(mddev->gendisk);
3842 if (sectors > mddev->dev_sectors &&
3843 mddev->recovery_cp > oldsize) {
3844 mddev->recovery_cp = oldsize;
3845 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3846 }
NeilBrown6508fdb2012-05-17 10:08:45 +10003847 calc_sectors(conf, sectors);
3848 mddev->dev_sectors = conf->dev_sectors;
NeilBrown006a09a2012-03-19 12:46:40 +11003849 mddev->resync_max_sectors = size;
3850 return 0;
3851}
3852
NeilBrownfd01b882011-10-11 16:47:53 +11003853static void *raid10_takeover_raid0(struct mddev *mddev)
Trela, Maciejdab8b292010-03-08 16:02:45 +11003854{
NeilBrown3cb03002011-10-11 16:45:26 +11003855 struct md_rdev *rdev;
NeilBrowne879a872011-10-11 16:49:02 +11003856 struct r10conf *conf;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003857
3858 if (mddev->degraded > 0) {
NeilBrown128595e2010-05-03 14:47:14 +10003859 printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3860 mdname(mddev));
Trela, Maciejdab8b292010-03-08 16:02:45 +11003861 return ERR_PTR(-EINVAL);
3862 }
3863
Trela, Maciejdab8b292010-03-08 16:02:45 +11003864 /* Set new parameters */
3865 mddev->new_level = 10;
3866 /* new layout: far_copies = 1, near_copies = 2 */
3867 mddev->new_layout = (1<<8) + 2;
3868 mddev->new_chunk_sectors = mddev->chunk_sectors;
3869 mddev->delta_disks = mddev->raid_disks;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003870 mddev->raid_disks *= 2;
3871 /* make sure it will be not marked as dirty */
3872 mddev->recovery_cp = MaxSector;
3873
3874 conf = setup_conf(mddev);
Krzysztof Wojcik02214dc2011-02-04 14:18:26 +01003875 if (!IS_ERR(conf)) {
NeilBrowndafb20f2012-03-19 12:46:39 +11003876 rdev_for_each(rdev, mddev)
NeilBrowne93f68a2010-06-15 09:36:03 +01003877 if (rdev->raid_disk >= 0)
3878 rdev->new_raid_disk = rdev->raid_disk * 2;
Krzysztof Wojcik02214dc2011-02-04 14:18:26 +01003879 conf->barrier = 1;
3880 }
3881
Trela, Maciejdab8b292010-03-08 16:02:45 +11003882 return conf;
3883}
3884
NeilBrownfd01b882011-10-11 16:47:53 +11003885static void *raid10_takeover(struct mddev *mddev)
Trela, Maciejdab8b292010-03-08 16:02:45 +11003886{
NeilBrowne373ab12011-10-11 16:48:59 +11003887 struct r0conf *raid0_conf;
Trela, Maciejdab8b292010-03-08 16:02:45 +11003888
3889 /* raid10 can take over:
3890 * raid0 - providing it has only two drives
3891 */
3892 if (mddev->level == 0) {
3893 /* for raid0 takeover only one zone is supported */
NeilBrowne373ab12011-10-11 16:48:59 +11003894 raid0_conf = mddev->private;
3895 if (raid0_conf->nr_strip_zones > 1) {
NeilBrown128595e2010-05-03 14:47:14 +10003896 printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3897 " with more than one zone.\n",
3898 mdname(mddev));
Trela, Maciejdab8b292010-03-08 16:02:45 +11003899 return ERR_PTR(-EINVAL);
3900 }
3901 return raid10_takeover_raid0(mddev);
3902 }
3903 return ERR_PTR(-EINVAL);
3904}
3905
NeilBrown3ea7daa2012-05-22 13:53:47 +10003906static int raid10_check_reshape(struct mddev *mddev)
3907{
3908 /* Called when there is a request to change
3909 * - layout (to ->new_layout)
3910 * - chunk size (to ->new_chunk_sectors)
3911 * - raid_disks (by delta_disks)
3912 * or when trying to restart a reshape that was ongoing.
3913 *
3914 * We need to validate the request and possibly allocate
3915 * space if that might be an issue later.
3916 *
3917 * Currently we reject any reshape of a 'far' mode array,
3918 * allow chunk size to change if new is generally acceptable,
3919 * allow raid_disks to increase, and allow
3920 * a switch between 'near' mode and 'offset' mode.
3921 */
3922 struct r10conf *conf = mddev->private;
3923 struct geom geo;
3924
3925 if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3926 return -EINVAL;
3927
3928 if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3929 /* mustn't change number of copies */
3930 return -EINVAL;
3931 if (geo.far_copies > 1 && !geo.far_offset)
3932 /* Cannot switch to 'far' mode */
3933 return -EINVAL;
3934
3935 if (mddev->array_sectors & geo.chunk_mask)
3936 /* not factor of array size */
3937 return -EINVAL;
3938
NeilBrown3ea7daa2012-05-22 13:53:47 +10003939 if (!enough(conf, -1))
3940 return -EINVAL;
3941
3942 kfree(conf->mirrors_new);
3943 conf->mirrors_new = NULL;
3944 if (mddev->delta_disks > 0) {
3945 /* allocate new 'mirrors' list */
3946 conf->mirrors_new = kzalloc(
Jonathan Brassowdc280d982012-07-31 10:03:52 +10003947 sizeof(struct raid10_info)
NeilBrown3ea7daa2012-05-22 13:53:47 +10003948 *(mddev->raid_disks +
3949 mddev->delta_disks),
3950 GFP_KERNEL);
3951 if (!conf->mirrors_new)
3952 return -ENOMEM;
3953 }
3954 return 0;
3955}
3956
3957/*
3958 * Need to check if array has failed when deciding whether to:
3959 * - start an array
3960 * - remove non-faulty devices
3961 * - add a spare
3962 * - allow a reshape
3963 * This determination is simple when no reshape is happening.
3964 * However if there is a reshape, we need to carefully check
3965 * both the before and after sections.
3966 * This is because some failed devices may only affect one
3967 * of the two sections, and some non-in_sync devices may
3968 * be insync in the section most affected by failed devices.
3969 */
3970static int calc_degraded(struct r10conf *conf)
3971{
3972 int degraded, degraded2;
3973 int i;
3974
3975 rcu_read_lock();
3976 degraded = 0;
3977 /* 'prev' section first */
3978 for (i = 0; i < conf->prev.raid_disks; i++) {
3979 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3980 if (!rdev || test_bit(Faulty, &rdev->flags))
3981 degraded++;
3982 else if (!test_bit(In_sync, &rdev->flags))
3983 /* When we can reduce the number of devices in
3984 * an array, this might not contribute to
3985 * 'degraded'. It does now.
3986 */
3987 degraded++;
3988 }
3989 rcu_read_unlock();
3990 if (conf->geo.raid_disks == conf->prev.raid_disks)
3991 return degraded;
3992 rcu_read_lock();
3993 degraded2 = 0;
3994 for (i = 0; i < conf->geo.raid_disks; i++) {
3995 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3996 if (!rdev || test_bit(Faulty, &rdev->flags))
3997 degraded2++;
3998 else if (!test_bit(In_sync, &rdev->flags)) {
3999 /* If reshape is increasing the number of devices,
4000 * this section has already been recovered, so
4001 * it doesn't contribute to degraded.
4002 * else it does.
4003 */
4004 if (conf->geo.raid_disks <= conf->prev.raid_disks)
4005 degraded2++;
4006 }
4007 }
4008 rcu_read_unlock();
4009 if (degraded2 > degraded)
4010 return degraded2;
4011 return degraded;
4012}
4013
4014static int raid10_start_reshape(struct mddev *mddev)
4015{
4016 /* A 'reshape' has been requested. This commits
4017 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4018 * This also checks if there are enough spares and adds them
4019 * to the array.
4020 * We currently require enough spares to make the final
4021 * array non-degraded. We also require that the difference
4022 * between old and new data_offset - on each device - is
4023 * enough that we never risk over-writing.
4024 */
4025
4026 unsigned long before_length, after_length;
4027 sector_t min_offset_diff = 0;
4028 int first = 1;
4029 struct geom new;
4030 struct r10conf *conf = mddev->private;
4031 struct md_rdev *rdev;
4032 int spares = 0;
NeilBrownbb63a702012-05-22 13:55:28 +10004033 int ret;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004034
4035 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4036 return -EBUSY;
4037
4038 if (setup_geo(&new, mddev, geo_start) != conf->copies)
4039 return -EINVAL;
4040
4041 before_length = ((1 << conf->prev.chunk_shift) *
4042 conf->prev.far_copies);
4043 after_length = ((1 << conf->geo.chunk_shift) *
4044 conf->geo.far_copies);
4045
4046 rdev_for_each(rdev, mddev) {
4047 if (!test_bit(In_sync, &rdev->flags)
4048 && !test_bit(Faulty, &rdev->flags))
4049 spares++;
4050 if (rdev->raid_disk >= 0) {
4051 long long diff = (rdev->new_data_offset
4052 - rdev->data_offset);
4053 if (!mddev->reshape_backwards)
4054 diff = -diff;
4055 if (diff < 0)
4056 diff = 0;
4057 if (first || diff < min_offset_diff)
4058 min_offset_diff = diff;
4059 }
4060 }
4061
4062 if (max(before_length, after_length) > min_offset_diff)
4063 return -EINVAL;
4064
4065 if (spares < mddev->delta_disks)
4066 return -EINVAL;
4067
4068 conf->offset_diff = min_offset_diff;
4069 spin_lock_irq(&conf->device_lock);
4070 if (conf->mirrors_new) {
4071 memcpy(conf->mirrors_new, conf->mirrors,
Jonathan Brassowdc280d982012-07-31 10:03:52 +10004072 sizeof(struct raid10_info)*conf->prev.raid_disks);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004073 smp_mb();
4074 kfree(conf->mirrors_old); /* FIXME and elsewhere */
4075 conf->mirrors_old = conf->mirrors;
4076 conf->mirrors = conf->mirrors_new;
4077 conf->mirrors_new = NULL;
4078 }
4079 setup_geo(&conf->geo, mddev, geo_start);
4080 smp_mb();
4081 if (mddev->reshape_backwards) {
4082 sector_t size = raid10_size(mddev, 0, 0);
4083 if (size < mddev->array_sectors) {
4084 spin_unlock_irq(&conf->device_lock);
4085 printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4086 mdname(mddev));
4087 return -EINVAL;
4088 }
4089 mddev->resync_max_sectors = size;
4090 conf->reshape_progress = size;
4091 } else
4092 conf->reshape_progress = 0;
4093 spin_unlock_irq(&conf->device_lock);
4094
NeilBrownbb63a702012-05-22 13:55:28 +10004095 if (mddev->delta_disks && mddev->bitmap) {
4096 ret = bitmap_resize(mddev->bitmap,
4097 raid10_size(mddev, 0,
4098 conf->geo.raid_disks),
4099 0, 0);
4100 if (ret)
4101 goto abort;
4102 }
NeilBrown3ea7daa2012-05-22 13:53:47 +10004103 if (mddev->delta_disks > 0) {
4104 rdev_for_each(rdev, mddev)
4105 if (rdev->raid_disk < 0 &&
4106 !test_bit(Faulty, &rdev->flags)) {
4107 if (raid10_add_disk(mddev, rdev) == 0) {
4108 if (rdev->raid_disk >=
4109 conf->prev.raid_disks)
4110 set_bit(In_sync, &rdev->flags);
4111 else
4112 rdev->recovery_offset = 0;
4113
4114 if (sysfs_link_rdev(mddev, rdev))
4115 /* Failure here is OK */;
4116 }
4117 } else if (rdev->raid_disk >= conf->prev.raid_disks
4118 && !test_bit(Faulty, &rdev->flags)) {
4119 /* This is a spare that was manually added */
4120 set_bit(In_sync, &rdev->flags);
4121 }
4122 }
4123 /* When a reshape changes the number of devices,
4124 * ->degraded is measured against the larger of the
4125 * pre and post numbers.
4126 */
4127 spin_lock_irq(&conf->device_lock);
4128 mddev->degraded = calc_degraded(conf);
4129 spin_unlock_irq(&conf->device_lock);
4130 mddev->raid_disks = conf->geo.raid_disks;
4131 mddev->reshape_position = conf->reshape_progress;
4132 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4133
4134 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4135 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4136 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4137 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4138
4139 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4140 "reshape");
4141 if (!mddev->sync_thread) {
NeilBrownbb63a702012-05-22 13:55:28 +10004142 ret = -EAGAIN;
4143 goto abort;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004144 }
4145 conf->reshape_checkpoint = jiffies;
4146 md_wakeup_thread(mddev->sync_thread);
4147 md_new_event(mddev);
4148 return 0;
NeilBrownbb63a702012-05-22 13:55:28 +10004149
4150abort:
4151 mddev->recovery = 0;
4152 spin_lock_irq(&conf->device_lock);
4153 conf->geo = conf->prev;
4154 mddev->raid_disks = conf->geo.raid_disks;
4155 rdev_for_each(rdev, mddev)
4156 rdev->new_data_offset = rdev->data_offset;
4157 smp_wmb();
4158 conf->reshape_progress = MaxSector;
4159 mddev->reshape_position = MaxSector;
4160 spin_unlock_irq(&conf->device_lock);
4161 return ret;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004162}
4163
4164/* Calculate the last device-address that could contain
4165 * any block from the chunk that includes the array-address 's'
4166 * and report the next address.
4167 * i.e. the address returned will be chunk-aligned and after
4168 * any data that is in the chunk containing 's'.
4169 */
4170static sector_t last_dev_address(sector_t s, struct geom *geo)
4171{
4172 s = (s | geo->chunk_mask) + 1;
4173 s >>= geo->chunk_shift;
4174 s *= geo->near_copies;
4175 s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4176 s *= geo->far_copies;
4177 s <<= geo->chunk_shift;
4178 return s;
4179}
4180
4181/* Calculate the first device-address that could contain
4182 * any block from the chunk that includes the array-address 's'.
4183 * This too will be the start of a chunk
4184 */
4185static sector_t first_dev_address(sector_t s, struct geom *geo)
4186{
4187 s >>= geo->chunk_shift;
4188 s *= geo->near_copies;
4189 sector_div(s, geo->raid_disks);
4190 s *= geo->far_copies;
4191 s <<= geo->chunk_shift;
4192 return s;
4193}
4194
4195static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4196 int *skipped)
4197{
4198 /* We simply copy at most one chunk (smallest of old and new)
4199 * at a time, possibly less if that exceeds RESYNC_PAGES,
4200 * or we hit a bad block or something.
4201 * This might mean we pause for normal IO in the middle of
4202 * a chunk, but that is not a problem was mddev->reshape_position
4203 * can record any location.
4204 *
4205 * If we will want to write to a location that isn't
4206 * yet recorded as 'safe' (i.e. in metadata on disk) then
4207 * we need to flush all reshape requests and update the metadata.
4208 *
4209 * When reshaping forwards (e.g. to more devices), we interpret
4210 * 'safe' as the earliest block which might not have been copied
4211 * down yet. We divide this by previous stripe size and multiply
4212 * by previous stripe length to get lowest device offset that we
4213 * cannot write to yet.
4214 * We interpret 'sector_nr' as an address that we want to write to.
4215 * From this we use last_device_address() to find where we might
4216 * write to, and first_device_address on the 'safe' position.
4217 * If this 'next' write position is after the 'safe' position,
4218 * we must update the metadata to increase the 'safe' position.
4219 *
4220 * When reshaping backwards, we round in the opposite direction
4221 * and perform the reverse test: next write position must not be
4222 * less than current safe position.
4223 *
4224 * In all this the minimum difference in data offsets
4225 * (conf->offset_diff - always positive) allows a bit of slack,
4226 * so next can be after 'safe', but not by more than offset_disk
4227 *
4228 * We need to prepare all the bios here before we start any IO
4229 * to ensure the size we choose is acceptable to all devices.
4230 * The means one for each copy for write-out and an extra one for
4231 * read-in.
4232 * We store the read-in bio in ->master_bio and the others in
4233 * ->devs[x].bio and ->devs[x].repl_bio.
4234 */
4235 struct r10conf *conf = mddev->private;
4236 struct r10bio *r10_bio;
4237 sector_t next, safe, last;
4238 int max_sectors;
4239 int nr_sectors;
4240 int s;
4241 struct md_rdev *rdev;
4242 int need_flush = 0;
4243 struct bio *blist;
4244 struct bio *bio, *read_bio;
4245 int sectors_done = 0;
4246
4247 if (sector_nr == 0) {
4248 /* If restarting in the middle, skip the initial sectors */
4249 if (mddev->reshape_backwards &&
4250 conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4251 sector_nr = (raid10_size(mddev, 0, 0)
4252 - conf->reshape_progress);
4253 } else if (!mddev->reshape_backwards &&
4254 conf->reshape_progress > 0)
4255 sector_nr = conf->reshape_progress;
4256 if (sector_nr) {
4257 mddev->curr_resync_completed = sector_nr;
4258 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4259 *skipped = 1;
4260 return sector_nr;
4261 }
4262 }
4263
4264 /* We don't use sector_nr to track where we are up to
4265 * as that doesn't work well for ->reshape_backwards.
4266 * So just use ->reshape_progress.
4267 */
4268 if (mddev->reshape_backwards) {
4269 /* 'next' is the earliest device address that we might
4270 * write to for this chunk in the new layout
4271 */
4272 next = first_dev_address(conf->reshape_progress - 1,
4273 &conf->geo);
4274
4275 /* 'safe' is the last device address that we might read from
4276 * in the old layout after a restart
4277 */
4278 safe = last_dev_address(conf->reshape_safe - 1,
4279 &conf->prev);
4280
4281 if (next + conf->offset_diff < safe)
4282 need_flush = 1;
4283
4284 last = conf->reshape_progress - 1;
4285 sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4286 & conf->prev.chunk_mask);
4287 if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4288 sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4289 } else {
4290 /* 'next' is after the last device address that we
4291 * might write to for this chunk in the new layout
4292 */
4293 next = last_dev_address(conf->reshape_progress, &conf->geo);
4294
4295 /* 'safe' is the earliest device address that we might
4296 * read from in the old layout after a restart
4297 */
4298 safe = first_dev_address(conf->reshape_safe, &conf->prev);
4299
4300 /* Need to update metadata if 'next' might be beyond 'safe'
4301 * as that would possibly corrupt data
4302 */
4303 if (next > safe + conf->offset_diff)
4304 need_flush = 1;
4305
4306 sector_nr = conf->reshape_progress;
4307 last = sector_nr | (conf->geo.chunk_mask
4308 & conf->prev.chunk_mask);
4309
4310 if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4311 last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4312 }
4313
4314 if (need_flush ||
4315 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4316 /* Need to update reshape_position in metadata */
4317 wait_barrier(conf);
4318 mddev->reshape_position = conf->reshape_progress;
4319 if (mddev->reshape_backwards)
4320 mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4321 - conf->reshape_progress;
4322 else
4323 mddev->curr_resync_completed = conf->reshape_progress;
4324 conf->reshape_checkpoint = jiffies;
4325 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4326 md_wakeup_thread(mddev->thread);
4327 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4328 kthread_should_stop());
4329 conf->reshape_safe = mddev->reshape_position;
4330 allow_barrier(conf);
4331 }
4332
4333read_more:
4334 /* Now schedule reads for blocks from sector_nr to last */
4335 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4336 raise_barrier(conf, sectors_done != 0);
4337 atomic_set(&r10_bio->remaining, 0);
4338 r10_bio->mddev = mddev;
4339 r10_bio->sector = sector_nr;
4340 set_bit(R10BIO_IsReshape, &r10_bio->state);
4341 r10_bio->sectors = last - sector_nr + 1;
4342 rdev = read_balance(conf, r10_bio, &max_sectors);
4343 BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4344
4345 if (!rdev) {
4346 /* Cannot read from here, so need to record bad blocks
4347 * on all the target devices.
4348 */
4349 // FIXME
4350 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4351 return sectors_done;
4352 }
4353
4354 read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4355
4356 read_bio->bi_bdev = rdev->bdev;
4357 read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4358 + rdev->data_offset);
4359 read_bio->bi_private = r10_bio;
4360 read_bio->bi_end_io = end_sync_read;
4361 read_bio->bi_rw = READ;
4362 read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4363 read_bio->bi_flags |= 1 << BIO_UPTODATE;
4364 read_bio->bi_vcnt = 0;
4365 read_bio->bi_idx = 0;
4366 read_bio->bi_size = 0;
4367 r10_bio->master_bio = read_bio;
4368 r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4369
4370 /* Now find the locations in the new layout */
4371 __raid10_find_phys(&conf->geo, r10_bio);
4372
4373 blist = read_bio;
4374 read_bio->bi_next = NULL;
4375
4376 for (s = 0; s < conf->copies*2; s++) {
4377 struct bio *b;
4378 int d = r10_bio->devs[s/2].devnum;
4379 struct md_rdev *rdev2;
4380 if (s&1) {
4381 rdev2 = conf->mirrors[d].replacement;
4382 b = r10_bio->devs[s/2].repl_bio;
4383 } else {
4384 rdev2 = conf->mirrors[d].rdev;
4385 b = r10_bio->devs[s/2].bio;
4386 }
4387 if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4388 continue;
4389 b->bi_bdev = rdev2->bdev;
4390 b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4391 b->bi_private = r10_bio;
4392 b->bi_end_io = end_reshape_write;
4393 b->bi_rw = WRITE;
4394 b->bi_flags &= ~(BIO_POOL_MASK - 1);
4395 b->bi_flags |= 1 << BIO_UPTODATE;
4396 b->bi_next = blist;
4397 b->bi_vcnt = 0;
4398 b->bi_idx = 0;
4399 b->bi_size = 0;
4400 blist = b;
4401 }
4402
4403 /* Now add as many pages as possible to all of these bios. */
4404
4405 nr_sectors = 0;
4406 for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4407 struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4408 int len = (max_sectors - s) << 9;
4409 if (len > PAGE_SIZE)
4410 len = PAGE_SIZE;
4411 for (bio = blist; bio ; bio = bio->bi_next) {
4412 struct bio *bio2;
4413 if (bio_add_page(bio, page, len, 0))
4414 continue;
4415
4416 /* Didn't fit, must stop */
4417 for (bio2 = blist;
4418 bio2 && bio2 != bio;
4419 bio2 = bio2->bi_next) {
4420 /* Remove last page from this bio */
4421 bio2->bi_vcnt--;
4422 bio2->bi_size -= len;
4423 bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4424 }
4425 goto bio_full;
4426 }
4427 sector_nr += len >> 9;
4428 nr_sectors += len >> 9;
4429 }
4430bio_full:
4431 r10_bio->sectors = nr_sectors;
4432
4433 /* Now submit the read */
4434 md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4435 atomic_inc(&r10_bio->remaining);
4436 read_bio->bi_next = NULL;
4437 generic_make_request(read_bio);
4438 sector_nr += nr_sectors;
4439 sectors_done += nr_sectors;
4440 if (sector_nr <= last)
4441 goto read_more;
4442
4443 /* Now that we have done the whole section we can
4444 * update reshape_progress
4445 */
4446 if (mddev->reshape_backwards)
4447 conf->reshape_progress -= sectors_done;
4448 else
4449 conf->reshape_progress += sectors_done;
4450
4451 return sectors_done;
4452}
4453
4454static void end_reshape_request(struct r10bio *r10_bio);
4455static int handle_reshape_read_error(struct mddev *mddev,
4456 struct r10bio *r10_bio);
4457static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4458{
4459 /* Reshape read completed. Hopefully we have a block
4460 * to write out.
4461 * If we got a read error then we do sync 1-page reads from
4462 * elsewhere until we find the data - or give up.
4463 */
4464 struct r10conf *conf = mddev->private;
4465 int s;
4466
4467 if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4468 if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4469 /* Reshape has been aborted */
4470 md_done_sync(mddev, r10_bio->sectors, 0);
4471 return;
4472 }
4473
4474 /* We definitely have the data in the pages, schedule the
4475 * writes.
4476 */
4477 atomic_set(&r10_bio->remaining, 1);
4478 for (s = 0; s < conf->copies*2; s++) {
4479 struct bio *b;
4480 int d = r10_bio->devs[s/2].devnum;
4481 struct md_rdev *rdev;
4482 if (s&1) {
4483 rdev = conf->mirrors[d].replacement;
4484 b = r10_bio->devs[s/2].repl_bio;
4485 } else {
4486 rdev = conf->mirrors[d].rdev;
4487 b = r10_bio->devs[s/2].bio;
4488 }
4489 if (!rdev || test_bit(Faulty, &rdev->flags))
4490 continue;
4491 atomic_inc(&rdev->nr_pending);
4492 md_sync_acct(b->bi_bdev, r10_bio->sectors);
4493 atomic_inc(&r10_bio->remaining);
4494 b->bi_next = NULL;
4495 generic_make_request(b);
4496 }
4497 end_reshape_request(r10_bio);
4498}
4499
4500static void end_reshape(struct r10conf *conf)
4501{
4502 if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4503 return;
4504
4505 spin_lock_irq(&conf->device_lock);
4506 conf->prev = conf->geo;
4507 md_finish_reshape(conf->mddev);
4508 smp_wmb();
4509 conf->reshape_progress = MaxSector;
4510 spin_unlock_irq(&conf->device_lock);
4511
4512 /* read-ahead size must cover two whole stripes, which is
4513 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4514 */
4515 if (conf->mddev->queue) {
4516 int stripe = conf->geo.raid_disks *
4517 ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4518 stripe /= conf->geo.near_copies;
4519 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4520 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4521 }
4522 conf->fullsync = 0;
4523}
4524
4525
4526static int handle_reshape_read_error(struct mddev *mddev,
4527 struct r10bio *r10_bio)
4528{
4529 /* Use sync reads to get the blocks from somewhere else */
4530 int sectors = r10_bio->sectors;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004531 struct r10conf *conf = mddev->private;
NeilBrowne0ee7782012-08-18 09:51:42 +10004532 struct {
4533 struct r10bio r10_bio;
4534 struct r10dev devs[conf->copies];
4535 } on_stack;
4536 struct r10bio *r10b = &on_stack.r10_bio;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004537 int slot = 0;
4538 int idx = 0;
4539 struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4540
NeilBrowne0ee7782012-08-18 09:51:42 +10004541 r10b->sector = r10_bio->sector;
4542 __raid10_find_phys(&conf->prev, r10b);
NeilBrown3ea7daa2012-05-22 13:53:47 +10004543
4544 while (sectors) {
4545 int s = sectors;
4546 int success = 0;
4547 int first_slot = slot;
4548
4549 if (s > (PAGE_SIZE >> 9))
4550 s = PAGE_SIZE >> 9;
4551
4552 while (!success) {
NeilBrowne0ee7782012-08-18 09:51:42 +10004553 int d = r10b->devs[slot].devnum;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004554 struct md_rdev *rdev = conf->mirrors[d].rdev;
4555 sector_t addr;
4556 if (rdev == NULL ||
4557 test_bit(Faulty, &rdev->flags) ||
4558 !test_bit(In_sync, &rdev->flags))
4559 goto failed;
4560
NeilBrowne0ee7782012-08-18 09:51:42 +10004561 addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
NeilBrown3ea7daa2012-05-22 13:53:47 +10004562 success = sync_page_io(rdev,
4563 addr,
4564 s << 9,
4565 bvec[idx].bv_page,
4566 READ, false);
4567 if (success)
4568 break;
4569 failed:
4570 slot++;
4571 if (slot >= conf->copies)
4572 slot = 0;
4573 if (slot == first_slot)
4574 break;
4575 }
4576 if (!success) {
4577 /* couldn't read this block, must give up */
4578 set_bit(MD_RECOVERY_INTR,
4579 &mddev->recovery);
4580 return -EIO;
4581 }
4582 sectors -= s;
4583 idx++;
4584 }
4585 return 0;
4586}
4587
4588static void end_reshape_write(struct bio *bio, int error)
4589{
4590 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4591 struct r10bio *r10_bio = bio->bi_private;
4592 struct mddev *mddev = r10_bio->mddev;
4593 struct r10conf *conf = mddev->private;
4594 int d;
4595 int slot;
4596 int repl;
4597 struct md_rdev *rdev = NULL;
4598
4599 d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4600 if (repl)
4601 rdev = conf->mirrors[d].replacement;
4602 if (!rdev) {
4603 smp_mb();
4604 rdev = conf->mirrors[d].rdev;
4605 }
4606
4607 if (!uptodate) {
4608 /* FIXME should record badblock */
4609 md_error(mddev, rdev);
4610 }
4611
4612 rdev_dec_pending(rdev, mddev);
4613 end_reshape_request(r10_bio);
4614}
4615
4616static void end_reshape_request(struct r10bio *r10_bio)
4617{
4618 if (!atomic_dec_and_test(&r10_bio->remaining))
4619 return;
4620 md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4621 bio_put(r10_bio->master_bio);
4622 put_buf(r10_bio);
4623}
4624
4625static void raid10_finish_reshape(struct mddev *mddev)
4626{
4627 struct r10conf *conf = mddev->private;
4628
4629 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4630 return;
4631
4632 if (mddev->delta_disks > 0) {
4633 sector_t size = raid10_size(mddev, 0, 0);
4634 md_set_array_sectors(mddev, size);
4635 if (mddev->recovery_cp > mddev->resync_max_sectors) {
4636 mddev->recovery_cp = mddev->resync_max_sectors;
4637 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4638 }
4639 mddev->resync_max_sectors = size;
4640 set_capacity(mddev->gendisk, mddev->array_sectors);
4641 revalidate_disk(mddev->gendisk);
NeilBrown63aced62012-05-22 13:55:33 +10004642 } else {
4643 int d;
4644 for (d = conf->geo.raid_disks ;
4645 d < conf->geo.raid_disks - mddev->delta_disks;
4646 d++) {
4647 struct md_rdev *rdev = conf->mirrors[d].rdev;
4648 if (rdev)
4649 clear_bit(In_sync, &rdev->flags);
4650 rdev = conf->mirrors[d].replacement;
4651 if (rdev)
4652 clear_bit(In_sync, &rdev->flags);
4653 }
NeilBrown3ea7daa2012-05-22 13:53:47 +10004654 }
4655 mddev->layout = mddev->new_layout;
4656 mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4657 mddev->reshape_position = MaxSector;
4658 mddev->delta_disks = 0;
4659 mddev->reshape_backwards = 0;
4660}
4661
NeilBrown84fc4b52011-10-11 16:49:58 +11004662static struct md_personality raid10_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07004663{
4664 .name = "raid10",
NeilBrown2604b702006-01-06 00:20:36 -08004665 .level = 10,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004666 .owner = THIS_MODULE,
4667 .make_request = make_request,
4668 .run = run,
4669 .stop = stop,
4670 .status = status,
4671 .error_handler = error,
4672 .hot_add_disk = raid10_add_disk,
4673 .hot_remove_disk= raid10_remove_disk,
4674 .spare_active = raid10_spare_active,
4675 .sync_request = sync_request,
NeilBrown6cce3b232006-01-06 00:20:16 -08004676 .quiesce = raid10_quiesce,
Dan Williams80c3a6c2009-03-17 18:10:40 -07004677 .size = raid10_size,
NeilBrown006a09a2012-03-19 12:46:40 +11004678 .resize = raid10_resize,
Trela, Maciejdab8b292010-03-08 16:02:45 +11004679 .takeover = raid10_takeover,
NeilBrown3ea7daa2012-05-22 13:53:47 +10004680 .check_reshape = raid10_check_reshape,
4681 .start_reshape = raid10_start_reshape,
4682 .finish_reshape = raid10_finish_reshape,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004683};
4684
4685static int __init raid_init(void)
4686{
NeilBrown2604b702006-01-06 00:20:36 -08004687 return register_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004688}
4689
4690static void raid_exit(void)
4691{
NeilBrown2604b702006-01-06 00:20:36 -08004692 unregister_md_personality(&raid10_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004693}
4694
4695module_init(raid_init);
4696module_exit(raid_exit);
4697MODULE_LICENSE("GPL");
NeilBrown0efb9e62009-12-14 12:49:58 +11004698MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004699MODULE_ALIAS("md-personality-9"); /* RAID10 */
NeilBrownd9d166c2006-01-06 00:20:51 -08004700MODULE_ALIAS("md-raid10");
NeilBrown2604b702006-01-06 00:20:36 -08004701MODULE_ALIAS("md-level-10");
NeilBrown34db0cd2011-10-11 16:50:01 +11004702
4703module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);