blob: 2332b5ced0dd0e9069237779e5edee246352c9aa [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * raid5.c : Multiple Devices driver for Linux
3 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4 * Copyright (C) 1999, 2000 Ingo Molnar
NeilBrown16a53ec2006-06-26 00:27:38 -07005 * Copyright (C) 2002, 2003 H. Peter Anvin
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
NeilBrown16a53ec2006-06-26 00:27:38 -07007 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
Linus Torvalds1da177e2005-04-16 15:20:36 -070010 *
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
NeilBrownae3c20c2006-07-10 04:44:17 -070021/*
22 * BITMAP UNPLUGGING:
23 *
24 * The sequencing for updating the bitmap reliably is a little
25 * subtle (and I got it wrong the first time) so it deserves some
26 * explanation.
27 *
28 * We group bitmap updates into batches. Each batch has a number.
29 * We may write out several batches at once, but that isn't very important.
NeilBrown7c13edc2011-04-18 18:25:43 +100030 * conf->seq_write is the number of the last batch successfully written.
31 * conf->seq_flush is the number of the last batch that was closed to
NeilBrownae3c20c2006-07-10 04:44:17 -070032 * new additions.
33 * When we discover that we will need to write to any block in a stripe
34 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
NeilBrown7c13edc2011-04-18 18:25:43 +100035 * the number of the batch it will be in. This is seq_flush+1.
NeilBrownae3c20c2006-07-10 04:44:17 -070036 * When we are ready to do a write, if that batch hasn't been written yet,
37 * we plug the array and queue the stripe for later.
38 * When an unplug happens, we increment bm_flush, thus closing the current
39 * batch.
40 * When we notice that bm_flush > bm_write, we write out all pending updates
41 * to the bitmap, and advance bm_write to where bm_flush was.
42 * This may occasionally write a bit out twice, but is sure never to
43 * miss any bits.
44 */
Linus Torvalds1da177e2005-04-16 15:20:36 -070045
NeilBrownbff61972009-03-31 14:33:13 +110046#include <linux/blkdev.h>
NeilBrownf6705572006-03-27 01:18:11 -080047#include <linux/kthread.h>
Dan Williamsf701d582009-03-31 15:09:39 +110048#include <linux/raid/pq.h>
Dan Williams91c00922007-01-02 13:52:30 -070049#include <linux/async_tx.h>
Paul Gortmaker056075c2011-07-03 13:58:33 -040050#include <linux/module.h>
Dan Williams07a3b412009-08-29 19:13:13 -070051#include <linux/async.h>
NeilBrownbff61972009-03-31 14:33:13 +110052#include <linux/seq_file.h>
Dan Williams36d1c642009-07-14 11:48:22 -070053#include <linux/cpu.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090054#include <linux/slab.h>
Christian Dietrich8bda4702011-07-27 11:00:36 +100055#include <linux/ratelimit.h>
NeilBrowna9add5d2012-10-31 11:59:09 +110056#include <trace/events/block.h>
57
NeilBrown43b2e5d2009-03-31 14:33:13 +110058#include "md.h"
NeilBrownbff61972009-03-31 14:33:13 +110059#include "raid5.h"
Trela Maciej54071b32010-03-08 16:02:42 +110060#include "raid0.h"
Christoph Hellwigef740c32009-03-31 14:27:03 +110061#include "bitmap.h"
NeilBrown72626682005-09-09 16:23:54 -070062
NeilBrown06905ff2014-10-02 13:45:00 +100063static bool devices_handle_discard_safely = false;
64module_param(devices_handle_discard_safely, bool, 0644);
65MODULE_PARM_DESC(devices_handle_discard_safely,
66 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
Linus Torvalds1da177e2005-04-16 15:20:36 -070067/*
68 * Stripe cache
69 */
70
71#define NR_STRIPES 256
72#define STRIPE_SIZE PAGE_SIZE
73#define STRIPE_SHIFT (PAGE_SHIFT - 9)
74#define STRIPE_SECTORS (STRIPE_SIZE>>9)
75#define IO_THRESHOLD 1
Dan Williams8b3e6cd2008-04-28 02:15:53 -070076#define BYPASS_THRESHOLD 1
NeilBrownfccddba2006-01-06 00:20:33 -080077#define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
Linus Torvalds1da177e2005-04-16 15:20:36 -070078#define HASH_MASK (NR_HASH - 1)
79
NeilBrownd1688a62011-10-11 16:49:52 +110080static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
NeilBrowndb298e12011-10-07 14:23:00 +110081{
82 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
83 return &conf->stripe_hashtbl[hash];
84}
Linus Torvalds1da177e2005-04-16 15:20:36 -070085
86/* bio's attached to a stripe+device for I/O are linked together in bi_sector
87 * order without overlap. There may be several bio's per stripe+device, and
88 * a bio could span several devices.
89 * When walking this list for a particular stripe+device, we must never proceed
90 * beyond a bio that extends past this device, as the next bio might no longer
91 * be valid.
NeilBrowndb298e12011-10-07 14:23:00 +110092 * This function is used to determine the 'next' bio in the list, given the sector
Linus Torvalds1da177e2005-04-16 15:20:36 -070093 * of the current stripe+device
94 */
NeilBrowndb298e12011-10-07 14:23:00 +110095static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
96{
Kent Overstreetaa8b57a2013-02-05 15:19:29 -080097 int sectors = bio_sectors(bio);
NeilBrowndb298e12011-10-07 14:23:00 +110098 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
99 return bio->bi_next;
100 else
101 return NULL;
102}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103
Jens Axboe960e7392008-08-15 10:41:18 +0200104/*
Jens Axboe5b99c2f2008-08-15 10:56:11 +0200105 * We maintain a biased count of active stripes in the bottom 16 bits of
106 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
Jens Axboe960e7392008-08-15 10:41:18 +0200107 */
Shaohua Lie7836bd62012-07-19 16:01:31 +1000108static inline int raid5_bi_processed_stripes(struct bio *bio)
Jens Axboe960e7392008-08-15 10:41:18 +0200109{
Shaohua Lie7836bd62012-07-19 16:01:31 +1000110 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
111 return (atomic_read(segments) >> 16) & 0xffff;
Jens Axboe960e7392008-08-15 10:41:18 +0200112}
113
Shaohua Lie7836bd62012-07-19 16:01:31 +1000114static inline int raid5_dec_bi_active_stripes(struct bio *bio)
Jens Axboe960e7392008-08-15 10:41:18 +0200115{
Shaohua Lie7836bd62012-07-19 16:01:31 +1000116 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
117 return atomic_sub_return(1, segments) & 0xffff;
Jens Axboe960e7392008-08-15 10:41:18 +0200118}
119
Shaohua Lie7836bd62012-07-19 16:01:31 +1000120static inline void raid5_inc_bi_active_stripes(struct bio *bio)
Jens Axboe960e7392008-08-15 10:41:18 +0200121{
Shaohua Lie7836bd62012-07-19 16:01:31 +1000122 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
123 atomic_inc(segments);
Jens Axboe960e7392008-08-15 10:41:18 +0200124}
125
Shaohua Lie7836bd62012-07-19 16:01:31 +1000126static inline void raid5_set_bi_processed_stripes(struct bio *bio,
127 unsigned int cnt)
Jens Axboe960e7392008-08-15 10:41:18 +0200128{
Shaohua Lie7836bd62012-07-19 16:01:31 +1000129 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
130 int old, new;
Jens Axboe960e7392008-08-15 10:41:18 +0200131
Shaohua Lie7836bd62012-07-19 16:01:31 +1000132 do {
133 old = atomic_read(segments);
134 new = (old & 0xffff) | (cnt << 16);
135 } while (atomic_cmpxchg(segments, old, new) != old);
Jens Axboe960e7392008-08-15 10:41:18 +0200136}
137
Shaohua Lie7836bd62012-07-19 16:01:31 +1000138static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
Jens Axboe960e7392008-08-15 10:41:18 +0200139{
Shaohua Lie7836bd62012-07-19 16:01:31 +1000140 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
141 atomic_set(segments, cnt);
Jens Axboe960e7392008-08-15 10:41:18 +0200142}
143
NeilBrownd0dabf72009-03-31 14:39:38 +1100144/* Find first data disk in a raid6 stripe */
145static inline int raid6_d0(struct stripe_head *sh)
146{
NeilBrown67cc2b82009-03-31 14:39:38 +1100147 if (sh->ddf_layout)
148 /* ddf always start from first device */
149 return 0;
150 /* md starts just after Q block */
NeilBrownd0dabf72009-03-31 14:39:38 +1100151 if (sh->qd_idx == sh->disks - 1)
152 return 0;
153 else
154 return sh->qd_idx + 1;
155}
NeilBrown16a53ec2006-06-26 00:27:38 -0700156static inline int raid6_next_disk(int disk, int raid_disks)
157{
158 disk++;
159 return (disk < raid_disks) ? disk : 0;
160}
Dan Williamsa4456852007-07-09 11:56:43 -0700161
NeilBrownd0dabf72009-03-31 14:39:38 +1100162/* When walking through the disks in a raid5, starting at raid6_d0,
163 * We need to map each disk to a 'slot', where the data disks are slot
164 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
165 * is raid_disks-1. This help does that mapping.
166 */
NeilBrown67cc2b82009-03-31 14:39:38 +1100167static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
168 int *count, int syndrome_disks)
NeilBrownd0dabf72009-03-31 14:39:38 +1100169{
Dan Williams66295422009-10-19 18:09:32 -0700170 int slot = *count;
NeilBrown67cc2b82009-03-31 14:39:38 +1100171
NeilBrowne4424fe2009-10-16 16:27:34 +1100172 if (sh->ddf_layout)
Dan Williams66295422009-10-19 18:09:32 -0700173 (*count)++;
NeilBrownd0dabf72009-03-31 14:39:38 +1100174 if (idx == sh->pd_idx)
NeilBrown67cc2b82009-03-31 14:39:38 +1100175 return syndrome_disks;
NeilBrownd0dabf72009-03-31 14:39:38 +1100176 if (idx == sh->qd_idx)
NeilBrown67cc2b82009-03-31 14:39:38 +1100177 return syndrome_disks + 1;
NeilBrowne4424fe2009-10-16 16:27:34 +1100178 if (!sh->ddf_layout)
Dan Williams66295422009-10-19 18:09:32 -0700179 (*count)++;
NeilBrownd0dabf72009-03-31 14:39:38 +1100180 return slot;
181}
182
Dan Williamsa4456852007-07-09 11:56:43 -0700183static void return_io(struct bio *return_bi)
184{
185 struct bio *bi = return_bi;
186 while (bi) {
Dan Williamsa4456852007-07-09 11:56:43 -0700187
188 return_bi = bi->bi_next;
189 bi->bi_next = NULL;
190 bi->bi_size = 0;
Linus Torvalds0a82a8d2013-04-18 09:00:26 -0700191 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
192 bi, 0);
Neil Brown0e13fe232008-06-28 08:31:20 +1000193 bio_endio(bi, 0);
Dan Williamsa4456852007-07-09 11:56:43 -0700194 bi = return_bi;
195 }
196}
197
NeilBrownd1688a62011-10-11 16:49:52 +1100198static void print_raid5_conf (struct r5conf *conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700199
Dan Williams600aa102008-06-28 08:32:05 +1000200static int stripe_operations_active(struct stripe_head *sh)
201{
202 return sh->check_state || sh->reconstruct_state ||
203 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
204 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
205}
206
Shaohua Li4eb788d2012-07-19 16:01:31 +1000207static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208{
Shaohua Li4eb788d2012-07-19 16:01:31 +1000209 BUG_ON(!list_empty(&sh->lru));
210 BUG_ON(atomic_read(&conf->active_stripes)==0);
211 if (test_bit(STRIPE_HANDLE, &sh->state)) {
212 if (test_bit(STRIPE_DELAYED, &sh->state) &&
213 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
214 list_add_tail(&sh->lru, &conf->delayed_list);
215 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
216 sh->bm_seq - conf->seq_write > 0)
217 list_add_tail(&sh->lru, &conf->bitmap_list);
218 else {
219 clear_bit(STRIPE_DELAYED, &sh->state);
220 clear_bit(STRIPE_BIT_DELAY, &sh->state);
221 list_add_tail(&sh->lru, &conf->handle_list);
222 }
223 md_wakeup_thread(conf->mddev->thread);
224 } else {
225 BUG_ON(stripe_operations_active(sh));
226 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
227 if (atomic_dec_return(&conf->preread_active_stripes)
228 < IO_THRESHOLD)
229 md_wakeup_thread(conf->mddev->thread);
230 atomic_dec(&conf->active_stripes);
231 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
232 list_add_tail(&sh->lru, &conf->inactive_list);
233 wake_up(&conf->wait_for_stripe);
234 if (conf->retry_read_aligned)
235 md_wakeup_thread(conf->mddev->thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236 }
237 }
238}
NeilBrownd0dabf72009-03-31 14:39:38 +1100239
Shaohua Li4eb788d2012-07-19 16:01:31 +1000240static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
241{
242 if (atomic_dec_and_test(&sh->count))
243 do_release_stripe(conf, sh);
244}
245
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246static void release_stripe(struct stripe_head *sh)
247{
NeilBrownd1688a62011-10-11 16:49:52 +1100248 struct r5conf *conf = sh->raid_conf;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 unsigned long flags;
NeilBrown16a53ec2006-06-26 00:27:38 -0700250
Shaohua Li4eb788d2012-07-19 16:01:31 +1000251 local_irq_save(flags);
252 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
253 do_release_stripe(conf, sh);
254 spin_unlock(&conf->device_lock);
255 }
256 local_irq_restore(flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257}
258
NeilBrownfccddba2006-01-06 00:20:33 -0800259static inline void remove_hash(struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700260{
Dan Williams45b42332007-07-09 11:56:43 -0700261 pr_debug("remove_hash(), stripe %llu\n",
262 (unsigned long long)sh->sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263
NeilBrownfccddba2006-01-06 00:20:33 -0800264 hlist_del_init(&sh->hash);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265}
266
NeilBrownd1688a62011-10-11 16:49:52 +1100267static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700268{
NeilBrownfccddba2006-01-06 00:20:33 -0800269 struct hlist_head *hp = stripe_hash(conf, sh->sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270
Dan Williams45b42332007-07-09 11:56:43 -0700271 pr_debug("insert_hash(), stripe %llu\n",
272 (unsigned long long)sh->sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273
NeilBrownfccddba2006-01-06 00:20:33 -0800274 hlist_add_head(&sh->hash, hp);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275}
276
277
278/* find an idle stripe, make sure it is unhashed, and return it. */
NeilBrownd1688a62011-10-11 16:49:52 +1100279static struct stripe_head *get_free_stripe(struct r5conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280{
281 struct stripe_head *sh = NULL;
282 struct list_head *first;
283
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284 if (list_empty(&conf->inactive_list))
285 goto out;
286 first = conf->inactive_list.next;
287 sh = list_entry(first, struct stripe_head, lru);
288 list_del_init(first);
289 remove_hash(sh);
290 atomic_inc(&conf->active_stripes);
291out:
292 return sh;
293}
294
NeilBrowne4e11e32010-06-16 16:45:16 +1000295static void shrink_buffers(struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700296{
297 struct page *p;
298 int i;
NeilBrowne4e11e32010-06-16 16:45:16 +1000299 int num = sh->raid_conf->pool_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300
NeilBrowne4e11e32010-06-16 16:45:16 +1000301 for (i = 0; i < num ; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302 p = sh->dev[i].page;
303 if (!p)
304 continue;
305 sh->dev[i].page = NULL;
NeilBrown2d1f3b52006-01-06 00:20:31 -0800306 put_page(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700307 }
308}
309
NeilBrowne4e11e32010-06-16 16:45:16 +1000310static int grow_buffers(struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311{
312 int i;
NeilBrowne4e11e32010-06-16 16:45:16 +1000313 int num = sh->raid_conf->pool_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314
NeilBrowne4e11e32010-06-16 16:45:16 +1000315 for (i = 0; i < num; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316 struct page *page;
317
318 if (!(page = alloc_page(GFP_KERNEL))) {
319 return 1;
320 }
321 sh->dev[i].page = page;
322 }
323 return 0;
324}
325
NeilBrown784052e2009-03-31 15:19:07 +1100326static void raid5_build_block(struct stripe_head *sh, int i, int previous);
NeilBrownd1688a62011-10-11 16:49:52 +1100327static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
NeilBrown911d4ee2009-03-31 14:39:38 +1100328 struct stripe_head *sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329
NeilBrownb5663ba2009-03-31 14:39:38 +1100330static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331{
NeilBrownd1688a62011-10-11 16:49:52 +1100332 struct r5conf *conf = sh->raid_conf;
NeilBrown7ecaa1e2006-03-27 01:18:08 -0800333 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700334
Eric Sesterhenn78bafeb2006-04-02 13:31:42 +0200335 BUG_ON(atomic_read(&sh->count) != 0);
336 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
Dan Williams600aa102008-06-28 08:32:05 +1000337 BUG_ON(stripe_operations_active(sh));
Dan Williamsd84e0f12007-01-02 13:52:30 -0700338
Dan Williams45b42332007-07-09 11:56:43 -0700339 pr_debug("init_stripe called, stripe %llu\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 (unsigned long long)sh->sector);
341
342 remove_hash(sh);
NeilBrown16a53ec2006-06-26 00:27:38 -0700343
NeilBrown86b42c72009-03-31 15:19:03 +1100344 sh->generation = conf->generation - previous;
NeilBrownb5663ba2009-03-31 14:39:38 +1100345 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346 sh->sector = sector;
NeilBrown911d4ee2009-03-31 14:39:38 +1100347 stripe_set_idx(sector, conf, previous, sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 sh->state = 0;
349
NeilBrown7ecaa1e2006-03-27 01:18:08 -0800350
351 for (i = sh->disks; i--; ) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352 struct r5dev *dev = &sh->dev[i];
353
Dan Williamsd84e0f12007-01-02 13:52:30 -0700354 if (dev->toread || dev->read || dev->towrite || dev->written ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700355 test_bit(R5_LOCKED, &dev->flags)) {
Dan Williamsd84e0f12007-01-02 13:52:30 -0700356 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357 (unsigned long long)sh->sector, i, dev->toread,
Dan Williamsd84e0f12007-01-02 13:52:30 -0700358 dev->read, dev->towrite, dev->written,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 test_bit(R5_LOCKED, &dev->flags));
NeilBrown8cfa7b02011-07-27 11:00:36 +1000360 WARN_ON(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 }
362 dev->flags = 0;
NeilBrown784052e2009-03-31 15:19:07 +1100363 raid5_build_block(sh, i, previous);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364 }
365 insert_hash(conf, sh);
366}
367
NeilBrownd1688a62011-10-11 16:49:52 +1100368static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
NeilBrown86b42c72009-03-31 15:19:03 +1100369 short generation)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370{
371 struct stripe_head *sh;
372
Dan Williams45b42332007-07-09 11:56:43 -0700373 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
Sasha Levinb67bfe02013-02-27 17:06:00 -0800374 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
NeilBrown86b42c72009-03-31 15:19:03 +1100375 if (sh->sector == sector && sh->generation == generation)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 return sh;
Dan Williams45b42332007-07-09 11:56:43 -0700377 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 return NULL;
379}
380
NeilBrown674806d2010-06-16 17:17:53 +1000381/*
382 * Need to check if array has failed when deciding whether to:
383 * - start an array
384 * - remove non-faulty devices
385 * - add a spare
386 * - allow a reshape
387 * This determination is simple when no reshape is happening.
388 * However if there is a reshape, we need to carefully check
389 * both the before and after sections.
390 * This is because some failed devices may only affect one
391 * of the two sections, and some non-in_sync devices may
392 * be insync in the section most affected by failed devices.
393 */
NeilBrown908f4fb2011-12-23 10:17:50 +1100394static int calc_degraded(struct r5conf *conf)
NeilBrown674806d2010-06-16 17:17:53 +1000395{
NeilBrown908f4fb2011-12-23 10:17:50 +1100396 int degraded, degraded2;
NeilBrown674806d2010-06-16 17:17:53 +1000397 int i;
NeilBrown674806d2010-06-16 17:17:53 +1000398
399 rcu_read_lock();
400 degraded = 0;
401 for (i = 0; i < conf->previous_raid_disks; i++) {
NeilBrown3cb03002011-10-11 16:45:26 +1100402 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
NeilBrowne5c86472012-09-19 12:52:30 +1000403 if (rdev && test_bit(Faulty, &rdev->flags))
404 rdev = rcu_dereference(conf->disks[i].replacement);
NeilBrown674806d2010-06-16 17:17:53 +1000405 if (!rdev || test_bit(Faulty, &rdev->flags))
406 degraded++;
407 else if (test_bit(In_sync, &rdev->flags))
408 ;
409 else
410 /* not in-sync or faulty.
411 * If the reshape increases the number of devices,
412 * this is being recovered by the reshape, so
413 * this 'previous' section is not in_sync.
414 * If the number of devices is being reduced however,
415 * the device can only be part of the array if
416 * we are reverting a reshape, so this section will
417 * be in-sync.
418 */
419 if (conf->raid_disks >= conf->previous_raid_disks)
420 degraded++;
421 }
422 rcu_read_unlock();
NeilBrown908f4fb2011-12-23 10:17:50 +1100423 if (conf->raid_disks == conf->previous_raid_disks)
424 return degraded;
NeilBrown674806d2010-06-16 17:17:53 +1000425 rcu_read_lock();
NeilBrown908f4fb2011-12-23 10:17:50 +1100426 degraded2 = 0;
NeilBrown674806d2010-06-16 17:17:53 +1000427 for (i = 0; i < conf->raid_disks; i++) {
NeilBrown3cb03002011-10-11 16:45:26 +1100428 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
NeilBrowne5c86472012-09-19 12:52:30 +1000429 if (rdev && test_bit(Faulty, &rdev->flags))
430 rdev = rcu_dereference(conf->disks[i].replacement);
NeilBrown674806d2010-06-16 17:17:53 +1000431 if (!rdev || test_bit(Faulty, &rdev->flags))
NeilBrown908f4fb2011-12-23 10:17:50 +1100432 degraded2++;
NeilBrown674806d2010-06-16 17:17:53 +1000433 else if (test_bit(In_sync, &rdev->flags))
434 ;
435 else
436 /* not in-sync or faulty.
437 * If reshape increases the number of devices, this
438 * section has already been recovered, else it
439 * almost certainly hasn't.
440 */
441 if (conf->raid_disks <= conf->previous_raid_disks)
NeilBrown908f4fb2011-12-23 10:17:50 +1100442 degraded2++;
NeilBrown674806d2010-06-16 17:17:53 +1000443 }
444 rcu_read_unlock();
NeilBrown908f4fb2011-12-23 10:17:50 +1100445 if (degraded2 > degraded)
446 return degraded2;
447 return degraded;
448}
449
450static int has_failed(struct r5conf *conf)
451{
452 int degraded;
453
454 if (conf->mddev->reshape_position == MaxSector)
455 return conf->mddev->degraded > conf->max_degraded;
456
457 degraded = calc_degraded(conf);
NeilBrown674806d2010-06-16 17:17:53 +1000458 if (degraded > conf->max_degraded)
459 return 1;
460 return 0;
461}
462
NeilBrownb5663ba2009-03-31 14:39:38 +1100463static struct stripe_head *
NeilBrownd1688a62011-10-11 16:49:52 +1100464get_active_stripe(struct r5conf *conf, sector_t sector,
NeilBrowna8c906c2009-06-09 14:39:59 +1000465 int previous, int noblock, int noquiesce)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466{
467 struct stripe_head *sh;
468
Dan Williams45b42332007-07-09 11:56:43 -0700469 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470
471 spin_lock_irq(&conf->device_lock);
472
473 do {
NeilBrown72626682005-09-09 16:23:54 -0700474 wait_event_lock_irq(conf->wait_for_stripe,
NeilBrowna8c906c2009-06-09 14:39:59 +1000475 conf->quiesce == 0 || noquiesce,
Lukas Czernereed8c022012-11-30 11:42:40 +0100476 conf->device_lock);
NeilBrown86b42c72009-03-31 15:19:03 +1100477 sh = __find_stripe(conf, sector, conf->generation - previous);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 if (!sh) {
479 if (!conf->inactive_blocked)
480 sh = get_free_stripe(conf);
481 if (noblock && sh == NULL)
482 break;
483 if (!sh) {
484 conf->inactive_blocked = 1;
485 wait_event_lock_irq(conf->wait_for_stripe,
486 !list_empty(&conf->inactive_list) &&
NeilBrown50368052005-12-12 02:39:17 -0800487 (atomic_read(&conf->active_stripes)
488 < (conf->max_nr_stripes *3/4)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 || !conf->inactive_blocked),
Lukas Czernereed8c022012-11-30 11:42:40 +0100490 conf->device_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491 conf->inactive_blocked = 0;
492 } else
NeilBrownb5663ba2009-03-31 14:39:38 +1100493 init_stripe(sh, sector, previous);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700494 } else {
495 if (atomic_read(&sh->count)) {
NeilBrownab69ae12009-03-31 15:26:47 +1100496 BUG_ON(!list_empty(&sh->lru)
Shaohua Li8811b592012-08-02 08:33:00 +1000497 && !test_bit(STRIPE_EXPANDING, &sh->state)
498 && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499 } else {
500 if (!test_bit(STRIPE_HANDLE, &sh->state))
501 atomic_inc(&conf->active_stripes);
NeilBrownff4e8d92006-07-10 04:44:16 -0700502 if (list_empty(&sh->lru) &&
503 !test_bit(STRIPE_EXPANDING, &sh->state))
NeilBrown16a53ec2006-06-26 00:27:38 -0700504 BUG();
505 list_del_init(&sh->lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 }
507 }
508 } while (sh == NULL);
509
510 if (sh)
511 atomic_inc(&sh->count);
512
513 spin_unlock_irq(&conf->device_lock);
514 return sh;
515}
516
NeilBrown05616be2012-05-21 09:27:00 +1000517/* Determine if 'data_offset' or 'new_data_offset' should be used
518 * in this stripe_head.
519 */
520static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
521{
522 sector_t progress = conf->reshape_progress;
523 /* Need a memory barrier to make sure we see the value
524 * of conf->generation, or ->data_offset that was set before
525 * reshape_progress was updated.
526 */
527 smp_rmb();
528 if (progress == MaxSector)
529 return 0;
530 if (sh->generation == conf->generation - 1)
531 return 0;
532 /* We are in a reshape, and this is a new-generation stripe,
533 * so use new_data_offset.
534 */
535 return 1;
536}
537
NeilBrown6712ecf2007-09-27 12:47:43 +0200538static void
539raid5_end_read_request(struct bio *bi, int error);
540static void
541raid5_end_write_request(struct bio *bi, int error);
Dan Williams91c00922007-01-02 13:52:30 -0700542
Dan Williamsc4e5ac02008-06-28 08:31:53 +1000543static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
Dan Williams91c00922007-01-02 13:52:30 -0700544{
NeilBrownd1688a62011-10-11 16:49:52 +1100545 struct r5conf *conf = sh->raid_conf;
Dan Williams91c00922007-01-02 13:52:30 -0700546 int i, disks = sh->disks;
547
548 might_sleep();
549
550 for (i = disks; i--; ) {
551 int rw;
NeilBrown9a3e1102011-12-23 10:17:53 +1100552 int replace_only = 0;
NeilBrown977df362011-12-23 10:17:53 +1100553 struct bio *bi, *rbi;
554 struct md_rdev *rdev, *rrdev = NULL;
Tejun Heoe9c74692010-09-03 11:56:18 +0200555 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
556 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
557 rw = WRITE_FUA;
558 else
559 rw = WRITE;
Shaohua Li9e4447682012-10-11 13:49:49 +1100560 if (test_bit(R5_Discard, &sh->dev[i].flags))
Shaohua Li620125f2012-10-11 13:49:05 +1100561 rw |= REQ_DISCARD;
Tejun Heoe9c74692010-09-03 11:56:18 +0200562 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
Dan Williams91c00922007-01-02 13:52:30 -0700563 rw = READ;
NeilBrown9a3e1102011-12-23 10:17:53 +1100564 else if (test_and_clear_bit(R5_WantReplace,
565 &sh->dev[i].flags)) {
566 rw = WRITE;
567 replace_only = 1;
568 } else
Dan Williams91c00922007-01-02 13:52:30 -0700569 continue;
Shaohua Libc0934f2012-05-22 13:55:05 +1000570 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
571 rw |= REQ_SYNC;
Dan Williams91c00922007-01-02 13:52:30 -0700572
573 bi = &sh->dev[i].req;
NeilBrown977df362011-12-23 10:17:53 +1100574 rbi = &sh->dev[i].rreq; /* For writing to replacement */
Dan Williams91c00922007-01-02 13:52:30 -0700575
Dan Williams91c00922007-01-02 13:52:30 -0700576 rcu_read_lock();
NeilBrown9a3e1102011-12-23 10:17:53 +1100577 rrdev = rcu_dereference(conf->disks[i].replacement);
NeilBrowndd054fc2011-12-23 10:17:53 +1100578 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
579 rdev = rcu_dereference(conf->disks[i].rdev);
580 if (!rdev) {
581 rdev = rrdev;
582 rrdev = NULL;
583 }
NeilBrown9a3e1102011-12-23 10:17:53 +1100584 if (rw & WRITE) {
585 if (replace_only)
586 rdev = NULL;
NeilBrowndd054fc2011-12-23 10:17:53 +1100587 if (rdev == rrdev)
588 /* We raced and saw duplicates */
589 rrdev = NULL;
NeilBrown9a3e1102011-12-23 10:17:53 +1100590 } else {
NeilBrowndd054fc2011-12-23 10:17:53 +1100591 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
NeilBrown9a3e1102011-12-23 10:17:53 +1100592 rdev = rrdev;
593 rrdev = NULL;
594 }
NeilBrown977df362011-12-23 10:17:53 +1100595
Dan Williams91c00922007-01-02 13:52:30 -0700596 if (rdev && test_bit(Faulty, &rdev->flags))
597 rdev = NULL;
598 if (rdev)
599 atomic_inc(&rdev->nr_pending);
NeilBrown977df362011-12-23 10:17:53 +1100600 if (rrdev && test_bit(Faulty, &rrdev->flags))
601 rrdev = NULL;
602 if (rrdev)
603 atomic_inc(&rrdev->nr_pending);
Dan Williams91c00922007-01-02 13:52:30 -0700604 rcu_read_unlock();
605
NeilBrown73e92e52011-07-28 11:39:22 +1000606 /* We have already checked bad blocks for reads. Now
NeilBrown977df362011-12-23 10:17:53 +1100607 * need to check for writes. We never accept write errors
608 * on the replacement, so we don't to check rrdev.
NeilBrown73e92e52011-07-28 11:39:22 +1000609 */
610 while ((rw & WRITE) && rdev &&
611 test_bit(WriteErrorSeen, &rdev->flags)) {
612 sector_t first_bad;
613 int bad_sectors;
614 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
615 &first_bad, &bad_sectors);
616 if (!bad)
617 break;
618
619 if (bad < 0) {
620 set_bit(BlockedBadBlocks, &rdev->flags);
621 if (!conf->mddev->external &&
622 conf->mddev->flags) {
623 /* It is very unlikely, but we might
624 * still need to write out the
625 * bad block log - better give it
626 * a chance*/
627 md_check_recovery(conf->mddev);
628 }
majianpeng18507532012-07-03 12:11:54 +1000629 /*
630 * Because md_wait_for_blocked_rdev
631 * will dec nr_pending, we must
632 * increment it first.
633 */
634 atomic_inc(&rdev->nr_pending);
NeilBrown73e92e52011-07-28 11:39:22 +1000635 md_wait_for_blocked_rdev(rdev, conf->mddev);
636 } else {
637 /* Acknowledged bad block - skip the write */
638 rdev_dec_pending(rdev, conf->mddev);
639 rdev = NULL;
640 }
641 }
642
Dan Williams91c00922007-01-02 13:52:30 -0700643 if (rdev) {
NeilBrown9a3e1102011-12-23 10:17:53 +1100644 if (s->syncing || s->expanding || s->expanded
645 || s->replacing)
Dan Williams91c00922007-01-02 13:52:30 -0700646 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
647
Dan Williams2b7497f2008-06-28 08:31:52 +1000648 set_bit(STRIPE_IO_STARTED, &sh->state);
649
Kent Overstreet2f6db2a2012-09-11 12:26:38 -0700650 bio_reset(bi);
Dan Williams91c00922007-01-02 13:52:30 -0700651 bi->bi_bdev = rdev->bdev;
Kent Overstreet2f6db2a2012-09-11 12:26:38 -0700652 bi->bi_rw = rw;
653 bi->bi_end_io = (rw & WRITE)
654 ? raid5_end_write_request
655 : raid5_end_read_request;
656 bi->bi_private = sh;
657
Dan Williams91c00922007-01-02 13:52:30 -0700658 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
Harvey Harrisone46b2722008-04-28 02:15:50 -0700659 __func__, (unsigned long long)sh->sector,
Dan Williams91c00922007-01-02 13:52:30 -0700660 bi->bi_rw, i);
661 atomic_inc(&sh->count);
NeilBrown05616be2012-05-21 09:27:00 +1000662 if (use_new_offset(conf, sh))
663 bi->bi_sector = (sh->sector
664 + rdev->new_data_offset);
665 else
666 bi->bi_sector = (sh->sector
667 + rdev->data_offset);
majianpeng3f9e7c12012-07-31 10:04:21 +1000668 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
669 bi->bi_rw |= REQ_FLUSH;
670
Kent Overstreet4997b722013-05-30 08:44:39 +0200671 bi->bi_vcnt = 1;
Dan Williams91c00922007-01-02 13:52:30 -0700672 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
673 bi->bi_io_vec[0].bv_offset = 0;
674 bi->bi_size = STRIPE_SIZE;
Shaohua Li7e44a922013-10-19 14:50:28 +0800675 /*
676 * If this is discard request, set bi_vcnt 0. We don't
677 * want to confuse SCSI because SCSI will replace payload
678 */
679 if (rw & REQ_DISCARD)
680 bi->bi_vcnt = 0;
NeilBrown977df362011-12-23 10:17:53 +1100681 if (rrdev)
682 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
Jonathan Brassowe3620a32013-03-07 16:22:01 -0600683
684 if (conf->mddev->gendisk)
685 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
686 bi, disk_devt(conf->mddev->gendisk),
687 sh->dev[i].sector);
Dan Williams91c00922007-01-02 13:52:30 -0700688 generic_make_request(bi);
NeilBrown977df362011-12-23 10:17:53 +1100689 }
690 if (rrdev) {
NeilBrown9a3e1102011-12-23 10:17:53 +1100691 if (s->syncing || s->expanding || s->expanded
692 || s->replacing)
NeilBrown977df362011-12-23 10:17:53 +1100693 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
694
695 set_bit(STRIPE_IO_STARTED, &sh->state);
696
Kent Overstreet2f6db2a2012-09-11 12:26:38 -0700697 bio_reset(rbi);
NeilBrown977df362011-12-23 10:17:53 +1100698 rbi->bi_bdev = rrdev->bdev;
Kent Overstreet2f6db2a2012-09-11 12:26:38 -0700699 rbi->bi_rw = rw;
700 BUG_ON(!(rw & WRITE));
701 rbi->bi_end_io = raid5_end_write_request;
702 rbi->bi_private = sh;
703
NeilBrown977df362011-12-23 10:17:53 +1100704 pr_debug("%s: for %llu schedule op %ld on "
705 "replacement disc %d\n",
706 __func__, (unsigned long long)sh->sector,
707 rbi->bi_rw, i);
708 atomic_inc(&sh->count);
NeilBrown05616be2012-05-21 09:27:00 +1000709 if (use_new_offset(conf, sh))
710 rbi->bi_sector = (sh->sector
711 + rrdev->new_data_offset);
712 else
713 rbi->bi_sector = (sh->sector
714 + rrdev->data_offset);
Kent Overstreet4997b722013-05-30 08:44:39 +0200715 rbi->bi_vcnt = 1;
NeilBrown977df362011-12-23 10:17:53 +1100716 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
717 rbi->bi_io_vec[0].bv_offset = 0;
718 rbi->bi_size = STRIPE_SIZE;
Shaohua Li7e44a922013-10-19 14:50:28 +0800719 /*
720 * If this is discard request, set bi_vcnt 0. We don't
721 * want to confuse SCSI because SCSI will replace payload
722 */
723 if (rw & REQ_DISCARD)
724 rbi->bi_vcnt = 0;
Jonathan Brassowe3620a32013-03-07 16:22:01 -0600725 if (conf->mddev->gendisk)
726 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
727 rbi, disk_devt(conf->mddev->gendisk),
728 sh->dev[i].sector);
NeilBrown977df362011-12-23 10:17:53 +1100729 generic_make_request(rbi);
730 }
731 if (!rdev && !rrdev) {
Namhyung Kimb0629622011-06-14 14:20:19 +1000732 if (rw & WRITE)
Dan Williams91c00922007-01-02 13:52:30 -0700733 set_bit(STRIPE_DEGRADED, &sh->state);
734 pr_debug("skip op %ld on disc %d for sector %llu\n",
735 bi->bi_rw, i, (unsigned long long)sh->sector);
736 clear_bit(R5_LOCKED, &sh->dev[i].flags);
737 set_bit(STRIPE_HANDLE, &sh->state);
738 }
739 }
740}
741
742static struct dma_async_tx_descriptor *
743async_copy_data(int frombio, struct bio *bio, struct page *page,
744 sector_t sector, struct dma_async_tx_descriptor *tx)
745{
746 struct bio_vec *bvl;
747 struct page *bio_page;
748 int i;
749 int page_offset;
Dan Williamsa08abd82009-06-03 11:43:59 -0700750 struct async_submit_ctl submit;
Dan Williams0403e382009-09-08 17:42:50 -0700751 enum async_tx_flags flags = 0;
Dan Williams91c00922007-01-02 13:52:30 -0700752
753 if (bio->bi_sector >= sector)
754 page_offset = (signed)(bio->bi_sector - sector) * 512;
755 else
756 page_offset = (signed)(sector - bio->bi_sector) * -512;
Dan Williamsa08abd82009-06-03 11:43:59 -0700757
Dan Williams0403e382009-09-08 17:42:50 -0700758 if (frombio)
759 flags |= ASYNC_TX_FENCE;
760 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
761
Dan Williams91c00922007-01-02 13:52:30 -0700762 bio_for_each_segment(bvl, bio, i) {
Namhyung Kimfcde9072011-06-14 14:23:57 +1000763 int len = bvl->bv_len;
Dan Williams91c00922007-01-02 13:52:30 -0700764 int clen;
765 int b_offset = 0;
766
767 if (page_offset < 0) {
768 b_offset = -page_offset;
769 page_offset += b_offset;
770 len -= b_offset;
771 }
772
773 if (len > 0 && page_offset + len > STRIPE_SIZE)
774 clen = STRIPE_SIZE - page_offset;
775 else
776 clen = len;
777
778 if (clen > 0) {
Namhyung Kimfcde9072011-06-14 14:23:57 +1000779 b_offset += bvl->bv_offset;
780 bio_page = bvl->bv_page;
Dan Williams91c00922007-01-02 13:52:30 -0700781 if (frombio)
782 tx = async_memcpy(page, bio_page, page_offset,
Dan Williamsa08abd82009-06-03 11:43:59 -0700783 b_offset, clen, &submit);
Dan Williams91c00922007-01-02 13:52:30 -0700784 else
785 tx = async_memcpy(bio_page, page, b_offset,
Dan Williamsa08abd82009-06-03 11:43:59 -0700786 page_offset, clen, &submit);
Dan Williams91c00922007-01-02 13:52:30 -0700787 }
Dan Williamsa08abd82009-06-03 11:43:59 -0700788 /* chain the operations */
789 submit.depend_tx = tx;
790
Dan Williams91c00922007-01-02 13:52:30 -0700791 if (clen < len) /* hit end of page */
792 break;
793 page_offset += len;
794 }
795
796 return tx;
797}
798
799static void ops_complete_biofill(void *stripe_head_ref)
800{
801 struct stripe_head *sh = stripe_head_ref;
802 struct bio *return_bi = NULL;
Dan Williamse4d84902007-09-24 10:06:13 -0700803 int i;
Dan Williams91c00922007-01-02 13:52:30 -0700804
Harvey Harrisone46b2722008-04-28 02:15:50 -0700805 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -0700806 (unsigned long long)sh->sector);
807
808 /* clear completed biofills */
809 for (i = sh->disks; i--; ) {
810 struct r5dev *dev = &sh->dev[i];
Dan Williams91c00922007-01-02 13:52:30 -0700811
812 /* acknowledge completion of a biofill operation */
Dan Williamse4d84902007-09-24 10:06:13 -0700813 /* and check if we need to reply to a read request,
814 * new R5_Wantfill requests are held off until
Dan Williams83de75c2008-06-28 08:31:58 +1000815 * !STRIPE_BIOFILL_RUN
Dan Williamse4d84902007-09-24 10:06:13 -0700816 */
817 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
Dan Williams91c00922007-01-02 13:52:30 -0700818 struct bio *rbi, *rbi2;
Dan Williams91c00922007-01-02 13:52:30 -0700819
Dan Williams91c00922007-01-02 13:52:30 -0700820 BUG_ON(!dev->read);
821 rbi = dev->read;
822 dev->read = NULL;
823 while (rbi && rbi->bi_sector <
824 dev->sector + STRIPE_SECTORS) {
825 rbi2 = r5_next_bio(rbi, dev->sector);
Shaohua Lie7836bd62012-07-19 16:01:31 +1000826 if (!raid5_dec_bi_active_stripes(rbi)) {
Dan Williams91c00922007-01-02 13:52:30 -0700827 rbi->bi_next = return_bi;
828 return_bi = rbi;
829 }
Dan Williams91c00922007-01-02 13:52:30 -0700830 rbi = rbi2;
831 }
832 }
833 }
Dan Williams83de75c2008-06-28 08:31:58 +1000834 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
Dan Williams91c00922007-01-02 13:52:30 -0700835
836 return_io(return_bi);
837
Dan Williamse4d84902007-09-24 10:06:13 -0700838 set_bit(STRIPE_HANDLE, &sh->state);
Dan Williams91c00922007-01-02 13:52:30 -0700839 release_stripe(sh);
840}
841
842static void ops_run_biofill(struct stripe_head *sh)
843{
844 struct dma_async_tx_descriptor *tx = NULL;
Dan Williamsa08abd82009-06-03 11:43:59 -0700845 struct async_submit_ctl submit;
Dan Williams91c00922007-01-02 13:52:30 -0700846 int i;
847
Harvey Harrisone46b2722008-04-28 02:15:50 -0700848 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -0700849 (unsigned long long)sh->sector);
850
851 for (i = sh->disks; i--; ) {
852 struct r5dev *dev = &sh->dev[i];
853 if (test_bit(R5_Wantfill, &dev->flags)) {
854 struct bio *rbi;
Shaohua Lib17459c2012-07-19 16:01:31 +1000855 spin_lock_irq(&sh->stripe_lock);
Dan Williams91c00922007-01-02 13:52:30 -0700856 dev->read = rbi = dev->toread;
857 dev->toread = NULL;
Shaohua Lib17459c2012-07-19 16:01:31 +1000858 spin_unlock_irq(&sh->stripe_lock);
Dan Williams91c00922007-01-02 13:52:30 -0700859 while (rbi && rbi->bi_sector <
860 dev->sector + STRIPE_SECTORS) {
861 tx = async_copy_data(0, rbi, dev->page,
862 dev->sector, tx);
863 rbi = r5_next_bio(rbi, dev->sector);
864 }
865 }
866 }
867
868 atomic_inc(&sh->count);
Dan Williamsa08abd82009-06-03 11:43:59 -0700869 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
870 async_trigger_callback(&submit);
Dan Williams91c00922007-01-02 13:52:30 -0700871}
872
Dan Williams4e7d2c02009-08-29 19:13:11 -0700873static void mark_target_uptodate(struct stripe_head *sh, int target)
874{
875 struct r5dev *tgt;
876
877 if (target < 0)
878 return;
879
880 tgt = &sh->dev[target];
881 set_bit(R5_UPTODATE, &tgt->flags);
882 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
883 clear_bit(R5_Wantcompute, &tgt->flags);
884}
885
Dan Williamsac6b53b2009-07-14 13:40:19 -0700886static void ops_complete_compute(void *stripe_head_ref)
Dan Williams91c00922007-01-02 13:52:30 -0700887{
888 struct stripe_head *sh = stripe_head_ref;
Dan Williams91c00922007-01-02 13:52:30 -0700889
Harvey Harrisone46b2722008-04-28 02:15:50 -0700890 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -0700891 (unsigned long long)sh->sector);
892
Dan Williamsac6b53b2009-07-14 13:40:19 -0700893 /* mark the computed target(s) as uptodate */
Dan Williams4e7d2c02009-08-29 19:13:11 -0700894 mark_target_uptodate(sh, sh->ops.target);
Dan Williamsac6b53b2009-07-14 13:40:19 -0700895 mark_target_uptodate(sh, sh->ops.target2);
Dan Williams4e7d2c02009-08-29 19:13:11 -0700896
Dan Williamsecc65c92008-06-28 08:31:57 +1000897 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
898 if (sh->check_state == check_state_compute_run)
899 sh->check_state = check_state_compute_result;
Dan Williams91c00922007-01-02 13:52:30 -0700900 set_bit(STRIPE_HANDLE, &sh->state);
901 release_stripe(sh);
902}
903
Dan Williamsd6f38f32009-07-14 11:50:52 -0700904/* return a pointer to the address conversion region of the scribble buffer */
905static addr_conv_t *to_addr_conv(struct stripe_head *sh,
906 struct raid5_percpu *percpu)
Dan Williams91c00922007-01-02 13:52:30 -0700907{
Dan Williamsd6f38f32009-07-14 11:50:52 -0700908 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
909}
910
911static struct dma_async_tx_descriptor *
912ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
913{
Dan Williams91c00922007-01-02 13:52:30 -0700914 int disks = sh->disks;
Dan Williamsd6f38f32009-07-14 11:50:52 -0700915 struct page **xor_srcs = percpu->scribble;
Dan Williams91c00922007-01-02 13:52:30 -0700916 int target = sh->ops.target;
917 struct r5dev *tgt = &sh->dev[target];
918 struct page *xor_dest = tgt->page;
919 int count = 0;
920 struct dma_async_tx_descriptor *tx;
Dan Williamsa08abd82009-06-03 11:43:59 -0700921 struct async_submit_ctl submit;
Dan Williams91c00922007-01-02 13:52:30 -0700922 int i;
923
924 pr_debug("%s: stripe %llu block: %d\n",
Harvey Harrisone46b2722008-04-28 02:15:50 -0700925 __func__, (unsigned long long)sh->sector, target);
Dan Williams91c00922007-01-02 13:52:30 -0700926 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
927
928 for (i = disks; i--; )
929 if (i != target)
930 xor_srcs[count++] = sh->dev[i].page;
931
932 atomic_inc(&sh->count);
933
Dan Williams0403e382009-09-08 17:42:50 -0700934 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
Dan Williamsac6b53b2009-07-14 13:40:19 -0700935 ops_complete_compute, sh, to_addr_conv(sh, percpu));
Dan Williams91c00922007-01-02 13:52:30 -0700936 if (unlikely(count == 1))
Dan Williamsa08abd82009-06-03 11:43:59 -0700937 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
Dan Williams91c00922007-01-02 13:52:30 -0700938 else
Dan Williamsa08abd82009-06-03 11:43:59 -0700939 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
Dan Williams91c00922007-01-02 13:52:30 -0700940
Dan Williams91c00922007-01-02 13:52:30 -0700941 return tx;
942}
943
Dan Williamsac6b53b2009-07-14 13:40:19 -0700944/* set_syndrome_sources - populate source buffers for gen_syndrome
945 * @srcs - (struct page *) array of size sh->disks
946 * @sh - stripe_head to parse
947 *
948 * Populates srcs in proper layout order for the stripe and returns the
949 * 'count' of sources to be used in a call to async_gen_syndrome. The P
950 * destination buffer is recorded in srcs[count] and the Q destination
951 * is recorded in srcs[count+1]].
952 */
953static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
954{
955 int disks = sh->disks;
956 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
957 int d0_idx = raid6_d0(sh);
958 int count;
959 int i;
960
961 for (i = 0; i < disks; i++)
NeilBrown5dd33c92009-10-16 16:40:25 +1100962 srcs[i] = NULL;
Dan Williamsac6b53b2009-07-14 13:40:19 -0700963
964 count = 0;
965 i = d0_idx;
966 do {
967 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
968
969 srcs[slot] = sh->dev[i].page;
970 i = raid6_next_disk(i, disks);
971 } while (i != d0_idx);
Dan Williamsac6b53b2009-07-14 13:40:19 -0700972
NeilBrowne4424fe2009-10-16 16:27:34 +1100973 return syndrome_disks;
Dan Williamsac6b53b2009-07-14 13:40:19 -0700974}
975
976static struct dma_async_tx_descriptor *
977ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
978{
979 int disks = sh->disks;
980 struct page **blocks = percpu->scribble;
981 int target;
982 int qd_idx = sh->qd_idx;
983 struct dma_async_tx_descriptor *tx;
984 struct async_submit_ctl submit;
985 struct r5dev *tgt;
986 struct page *dest;
987 int i;
988 int count;
989
990 if (sh->ops.target < 0)
991 target = sh->ops.target2;
992 else if (sh->ops.target2 < 0)
993 target = sh->ops.target;
994 else
995 /* we should only have one valid target */
996 BUG();
997 BUG_ON(target < 0);
998 pr_debug("%s: stripe %llu block: %d\n",
999 __func__, (unsigned long long)sh->sector, target);
1000
1001 tgt = &sh->dev[target];
1002 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1003 dest = tgt->page;
1004
1005 atomic_inc(&sh->count);
1006
1007 if (target == qd_idx) {
1008 count = set_syndrome_sources(blocks, sh);
1009 blocks[count] = NULL; /* regenerating p is not necessary */
1010 BUG_ON(blocks[count+1] != dest); /* q should already be set */
Dan Williams0403e382009-09-08 17:42:50 -07001011 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1012 ops_complete_compute, sh,
Dan Williamsac6b53b2009-07-14 13:40:19 -07001013 to_addr_conv(sh, percpu));
1014 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1015 } else {
1016 /* Compute any data- or p-drive using XOR */
1017 count = 0;
1018 for (i = disks; i-- ; ) {
1019 if (i == target || i == qd_idx)
1020 continue;
1021 blocks[count++] = sh->dev[i].page;
1022 }
1023
Dan Williams0403e382009-09-08 17:42:50 -07001024 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1025 NULL, ops_complete_compute, sh,
Dan Williamsac6b53b2009-07-14 13:40:19 -07001026 to_addr_conv(sh, percpu));
1027 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1028 }
1029
1030 return tx;
1031}
1032
1033static struct dma_async_tx_descriptor *
1034ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1035{
1036 int i, count, disks = sh->disks;
1037 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1038 int d0_idx = raid6_d0(sh);
1039 int faila = -1, failb = -1;
1040 int target = sh->ops.target;
1041 int target2 = sh->ops.target2;
1042 struct r5dev *tgt = &sh->dev[target];
1043 struct r5dev *tgt2 = &sh->dev[target2];
1044 struct dma_async_tx_descriptor *tx;
1045 struct page **blocks = percpu->scribble;
1046 struct async_submit_ctl submit;
1047
1048 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1049 __func__, (unsigned long long)sh->sector, target, target2);
1050 BUG_ON(target < 0 || target2 < 0);
1051 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1052 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1053
Dan Williams6c910a72009-09-16 12:24:54 -07001054 /* we need to open-code set_syndrome_sources to handle the
Dan Williamsac6b53b2009-07-14 13:40:19 -07001055 * slot number conversion for 'faila' and 'failb'
1056 */
1057 for (i = 0; i < disks ; i++)
NeilBrown5dd33c92009-10-16 16:40:25 +11001058 blocks[i] = NULL;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001059 count = 0;
1060 i = d0_idx;
1061 do {
1062 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1063
1064 blocks[slot] = sh->dev[i].page;
1065
1066 if (i == target)
1067 faila = slot;
1068 if (i == target2)
1069 failb = slot;
1070 i = raid6_next_disk(i, disks);
1071 } while (i != d0_idx);
Dan Williamsac6b53b2009-07-14 13:40:19 -07001072
1073 BUG_ON(faila == failb);
1074 if (failb < faila)
1075 swap(faila, failb);
1076 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1077 __func__, (unsigned long long)sh->sector, faila, failb);
1078
1079 atomic_inc(&sh->count);
1080
1081 if (failb == syndrome_disks+1) {
1082 /* Q disk is one of the missing disks */
1083 if (faila == syndrome_disks) {
1084 /* Missing P+Q, just recompute */
Dan Williams0403e382009-09-08 17:42:50 -07001085 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1086 ops_complete_compute, sh,
1087 to_addr_conv(sh, percpu));
NeilBrowne4424fe2009-10-16 16:27:34 +11001088 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
Dan Williamsac6b53b2009-07-14 13:40:19 -07001089 STRIPE_SIZE, &submit);
1090 } else {
1091 struct page *dest;
1092 int data_target;
1093 int qd_idx = sh->qd_idx;
1094
1095 /* Missing D+Q: recompute D from P, then recompute Q */
1096 if (target == qd_idx)
1097 data_target = target2;
1098 else
1099 data_target = target;
1100
1101 count = 0;
1102 for (i = disks; i-- ; ) {
1103 if (i == data_target || i == qd_idx)
1104 continue;
1105 blocks[count++] = sh->dev[i].page;
1106 }
1107 dest = sh->dev[data_target].page;
Dan Williams0403e382009-09-08 17:42:50 -07001108 init_async_submit(&submit,
1109 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1110 NULL, NULL, NULL,
1111 to_addr_conv(sh, percpu));
Dan Williamsac6b53b2009-07-14 13:40:19 -07001112 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1113 &submit);
1114
1115 count = set_syndrome_sources(blocks, sh);
Dan Williams0403e382009-09-08 17:42:50 -07001116 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1117 ops_complete_compute, sh,
1118 to_addr_conv(sh, percpu));
Dan Williamsac6b53b2009-07-14 13:40:19 -07001119 return async_gen_syndrome(blocks, 0, count+2,
1120 STRIPE_SIZE, &submit);
1121 }
Dan Williamsac6b53b2009-07-14 13:40:19 -07001122 } else {
Dan Williams6c910a72009-09-16 12:24:54 -07001123 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1124 ops_complete_compute, sh,
1125 to_addr_conv(sh, percpu));
1126 if (failb == syndrome_disks) {
1127 /* We're missing D+P. */
1128 return async_raid6_datap_recov(syndrome_disks+2,
1129 STRIPE_SIZE, faila,
1130 blocks, &submit);
1131 } else {
1132 /* We're missing D+D. */
1133 return async_raid6_2data_recov(syndrome_disks+2,
1134 STRIPE_SIZE, faila, failb,
1135 blocks, &submit);
1136 }
Dan Williamsac6b53b2009-07-14 13:40:19 -07001137 }
1138}
1139
1140
Dan Williams91c00922007-01-02 13:52:30 -07001141static void ops_complete_prexor(void *stripe_head_ref)
1142{
1143 struct stripe_head *sh = stripe_head_ref;
1144
Harvey Harrisone46b2722008-04-28 02:15:50 -07001145 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001146 (unsigned long long)sh->sector);
Dan Williams91c00922007-01-02 13:52:30 -07001147}
1148
1149static struct dma_async_tx_descriptor *
Dan Williamsd6f38f32009-07-14 11:50:52 -07001150ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1151 struct dma_async_tx_descriptor *tx)
Dan Williams91c00922007-01-02 13:52:30 -07001152{
Dan Williams91c00922007-01-02 13:52:30 -07001153 int disks = sh->disks;
Dan Williamsd6f38f32009-07-14 11:50:52 -07001154 struct page **xor_srcs = percpu->scribble;
Dan Williams91c00922007-01-02 13:52:30 -07001155 int count = 0, pd_idx = sh->pd_idx, i;
Dan Williamsa08abd82009-06-03 11:43:59 -07001156 struct async_submit_ctl submit;
Dan Williams91c00922007-01-02 13:52:30 -07001157
1158 /* existing parity data subtracted */
1159 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1160
Harvey Harrisone46b2722008-04-28 02:15:50 -07001161 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001162 (unsigned long long)sh->sector);
1163
1164 for (i = disks; i--; ) {
1165 struct r5dev *dev = &sh->dev[i];
1166 /* Only process blocks that are known to be uptodate */
Dan Williamsd8ee0722008-06-28 08:32:06 +10001167 if (test_bit(R5_Wantdrain, &dev->flags))
Dan Williams91c00922007-01-02 13:52:30 -07001168 xor_srcs[count++] = dev->page;
1169 }
1170
Dan Williams0403e382009-09-08 17:42:50 -07001171 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
Dan Williamsd6f38f32009-07-14 11:50:52 -07001172 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
Dan Williamsa08abd82009-06-03 11:43:59 -07001173 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
Dan Williams91c00922007-01-02 13:52:30 -07001174
1175 return tx;
1176}
1177
1178static struct dma_async_tx_descriptor *
Dan Williamsd8ee0722008-06-28 08:32:06 +10001179ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
Dan Williams91c00922007-01-02 13:52:30 -07001180{
1181 int disks = sh->disks;
Dan Williamsd8ee0722008-06-28 08:32:06 +10001182 int i;
Dan Williams91c00922007-01-02 13:52:30 -07001183
Harvey Harrisone46b2722008-04-28 02:15:50 -07001184 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001185 (unsigned long long)sh->sector);
1186
1187 for (i = disks; i--; ) {
1188 struct r5dev *dev = &sh->dev[i];
1189 struct bio *chosen;
Dan Williams91c00922007-01-02 13:52:30 -07001190
Dan Williamsd8ee0722008-06-28 08:32:06 +10001191 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
Dan Williams91c00922007-01-02 13:52:30 -07001192 struct bio *wbi;
1193
Shaohua Lib17459c2012-07-19 16:01:31 +10001194 spin_lock_irq(&sh->stripe_lock);
Dan Williams91c00922007-01-02 13:52:30 -07001195 chosen = dev->towrite;
1196 dev->towrite = NULL;
1197 BUG_ON(dev->written);
1198 wbi = dev->written = chosen;
Shaohua Lib17459c2012-07-19 16:01:31 +10001199 spin_unlock_irq(&sh->stripe_lock);
Dan Williams91c00922007-01-02 13:52:30 -07001200
1201 while (wbi && wbi->bi_sector <
1202 dev->sector + STRIPE_SECTORS) {
Tejun Heoe9c74692010-09-03 11:56:18 +02001203 if (wbi->bi_rw & REQ_FUA)
1204 set_bit(R5_WantFUA, &dev->flags);
Shaohua Libc0934f2012-05-22 13:55:05 +10001205 if (wbi->bi_rw & REQ_SYNC)
1206 set_bit(R5_SyncIO, &dev->flags);
Shaohua Li9e4447682012-10-11 13:49:49 +11001207 if (wbi->bi_rw & REQ_DISCARD)
Shaohua Li620125f2012-10-11 13:49:05 +11001208 set_bit(R5_Discard, &dev->flags);
Shaohua Li9e4447682012-10-11 13:49:49 +11001209 else
Shaohua Li620125f2012-10-11 13:49:05 +11001210 tx = async_copy_data(1, wbi, dev->page,
1211 dev->sector, tx);
Dan Williams91c00922007-01-02 13:52:30 -07001212 wbi = r5_next_bio(wbi, dev->sector);
1213 }
1214 }
1215 }
1216
1217 return tx;
1218}
1219
Dan Williamsac6b53b2009-07-14 13:40:19 -07001220static void ops_complete_reconstruct(void *stripe_head_ref)
Dan Williams91c00922007-01-02 13:52:30 -07001221{
1222 struct stripe_head *sh = stripe_head_ref;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001223 int disks = sh->disks;
1224 int pd_idx = sh->pd_idx;
1225 int qd_idx = sh->qd_idx;
1226 int i;
Shaohua Li9e4447682012-10-11 13:49:49 +11001227 bool fua = false, sync = false, discard = false;
Dan Williams91c00922007-01-02 13:52:30 -07001228
Harvey Harrisone46b2722008-04-28 02:15:50 -07001229 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001230 (unsigned long long)sh->sector);
1231
Shaohua Libc0934f2012-05-22 13:55:05 +10001232 for (i = disks; i--; ) {
Tejun Heoe9c74692010-09-03 11:56:18 +02001233 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
Shaohua Libc0934f2012-05-22 13:55:05 +10001234 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
Shaohua Li9e4447682012-10-11 13:49:49 +11001235 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
Shaohua Libc0934f2012-05-22 13:55:05 +10001236 }
Tejun Heoe9c74692010-09-03 11:56:18 +02001237
Dan Williams91c00922007-01-02 13:52:30 -07001238 for (i = disks; i--; ) {
1239 struct r5dev *dev = &sh->dev[i];
Dan Williamsac6b53b2009-07-14 13:40:19 -07001240
Tejun Heoe9c74692010-09-03 11:56:18 +02001241 if (dev->written || i == pd_idx || i == qd_idx) {
Shaohua Li9e4447682012-10-11 13:49:49 +11001242 if (!discard)
1243 set_bit(R5_UPTODATE, &dev->flags);
Tejun Heoe9c74692010-09-03 11:56:18 +02001244 if (fua)
1245 set_bit(R5_WantFUA, &dev->flags);
Shaohua Libc0934f2012-05-22 13:55:05 +10001246 if (sync)
1247 set_bit(R5_SyncIO, &dev->flags);
Tejun Heoe9c74692010-09-03 11:56:18 +02001248 }
Dan Williams91c00922007-01-02 13:52:30 -07001249 }
1250
Dan Williamsd8ee0722008-06-28 08:32:06 +10001251 if (sh->reconstruct_state == reconstruct_state_drain_run)
1252 sh->reconstruct_state = reconstruct_state_drain_result;
1253 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1254 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1255 else {
1256 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1257 sh->reconstruct_state = reconstruct_state_result;
1258 }
Dan Williams91c00922007-01-02 13:52:30 -07001259
1260 set_bit(STRIPE_HANDLE, &sh->state);
1261 release_stripe(sh);
1262}
1263
1264static void
Dan Williamsac6b53b2009-07-14 13:40:19 -07001265ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1266 struct dma_async_tx_descriptor *tx)
Dan Williams91c00922007-01-02 13:52:30 -07001267{
Dan Williams91c00922007-01-02 13:52:30 -07001268 int disks = sh->disks;
Dan Williamsd6f38f32009-07-14 11:50:52 -07001269 struct page **xor_srcs = percpu->scribble;
Dan Williamsa08abd82009-06-03 11:43:59 -07001270 struct async_submit_ctl submit;
Dan Williams91c00922007-01-02 13:52:30 -07001271 int count = 0, pd_idx = sh->pd_idx, i;
1272 struct page *xor_dest;
Dan Williamsd8ee0722008-06-28 08:32:06 +10001273 int prexor = 0;
Dan Williams91c00922007-01-02 13:52:30 -07001274 unsigned long flags;
Dan Williams91c00922007-01-02 13:52:30 -07001275
Harvey Harrisone46b2722008-04-28 02:15:50 -07001276 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001277 (unsigned long long)sh->sector);
1278
Shaohua Li620125f2012-10-11 13:49:05 +11001279 for (i = 0; i < sh->disks; i++) {
1280 if (pd_idx == i)
1281 continue;
1282 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1283 break;
1284 }
1285 if (i >= sh->disks) {
1286 atomic_inc(&sh->count);
Shaohua Li620125f2012-10-11 13:49:05 +11001287 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1288 ops_complete_reconstruct(sh);
1289 return;
1290 }
Dan Williams91c00922007-01-02 13:52:30 -07001291 /* check if prexor is active which means only process blocks
1292 * that are part of a read-modify-write (written)
1293 */
Dan Williamsd8ee0722008-06-28 08:32:06 +10001294 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1295 prexor = 1;
Dan Williams91c00922007-01-02 13:52:30 -07001296 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1297 for (i = disks; i--; ) {
1298 struct r5dev *dev = &sh->dev[i];
1299 if (dev->written)
1300 xor_srcs[count++] = dev->page;
1301 }
1302 } else {
1303 xor_dest = sh->dev[pd_idx].page;
1304 for (i = disks; i--; ) {
1305 struct r5dev *dev = &sh->dev[i];
1306 if (i != pd_idx)
1307 xor_srcs[count++] = dev->page;
1308 }
1309 }
1310
Dan Williams91c00922007-01-02 13:52:30 -07001311 /* 1/ if we prexor'd then the dest is reused as a source
1312 * 2/ if we did not prexor then we are redoing the parity
1313 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1314 * for the synchronous xor case
1315 */
Dan Williams88ba2aa2009-04-09 16:16:18 -07001316 flags = ASYNC_TX_ACK |
Dan Williams91c00922007-01-02 13:52:30 -07001317 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1318
1319 atomic_inc(&sh->count);
1320
Dan Williamsac6b53b2009-07-14 13:40:19 -07001321 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
Dan Williamsd6f38f32009-07-14 11:50:52 -07001322 to_addr_conv(sh, percpu));
Dan Williamsa08abd82009-06-03 11:43:59 -07001323 if (unlikely(count == 1))
1324 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1325 else
1326 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
Dan Williams91c00922007-01-02 13:52:30 -07001327}
1328
Dan Williamsac6b53b2009-07-14 13:40:19 -07001329static void
1330ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1331 struct dma_async_tx_descriptor *tx)
1332{
1333 struct async_submit_ctl submit;
1334 struct page **blocks = percpu->scribble;
Shaohua Li620125f2012-10-11 13:49:05 +11001335 int count, i;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001336
1337 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1338
Shaohua Li620125f2012-10-11 13:49:05 +11001339 for (i = 0; i < sh->disks; i++) {
1340 if (sh->pd_idx == i || sh->qd_idx == i)
1341 continue;
1342 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1343 break;
1344 }
1345 if (i >= sh->disks) {
1346 atomic_inc(&sh->count);
Shaohua Li620125f2012-10-11 13:49:05 +11001347 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1348 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1349 ops_complete_reconstruct(sh);
1350 return;
1351 }
1352
Dan Williamsac6b53b2009-07-14 13:40:19 -07001353 count = set_syndrome_sources(blocks, sh);
1354
1355 atomic_inc(&sh->count);
1356
1357 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1358 sh, to_addr_conv(sh, percpu));
1359 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
Dan Williams91c00922007-01-02 13:52:30 -07001360}
1361
1362static void ops_complete_check(void *stripe_head_ref)
1363{
1364 struct stripe_head *sh = stripe_head_ref;
Dan Williams91c00922007-01-02 13:52:30 -07001365
Harvey Harrisone46b2722008-04-28 02:15:50 -07001366 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001367 (unsigned long long)sh->sector);
1368
Dan Williamsecc65c92008-06-28 08:31:57 +10001369 sh->check_state = check_state_check_result;
Dan Williams91c00922007-01-02 13:52:30 -07001370 set_bit(STRIPE_HANDLE, &sh->state);
1371 release_stripe(sh);
1372}
1373
Dan Williamsac6b53b2009-07-14 13:40:19 -07001374static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
Dan Williams91c00922007-01-02 13:52:30 -07001375{
Dan Williams91c00922007-01-02 13:52:30 -07001376 int disks = sh->disks;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001377 int pd_idx = sh->pd_idx;
1378 int qd_idx = sh->qd_idx;
1379 struct page *xor_dest;
Dan Williamsd6f38f32009-07-14 11:50:52 -07001380 struct page **xor_srcs = percpu->scribble;
Dan Williams91c00922007-01-02 13:52:30 -07001381 struct dma_async_tx_descriptor *tx;
Dan Williamsa08abd82009-06-03 11:43:59 -07001382 struct async_submit_ctl submit;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001383 int count;
1384 int i;
Dan Williams91c00922007-01-02 13:52:30 -07001385
Harvey Harrisone46b2722008-04-28 02:15:50 -07001386 pr_debug("%s: stripe %llu\n", __func__,
Dan Williams91c00922007-01-02 13:52:30 -07001387 (unsigned long long)sh->sector);
1388
Dan Williamsac6b53b2009-07-14 13:40:19 -07001389 count = 0;
1390 xor_dest = sh->dev[pd_idx].page;
1391 xor_srcs[count++] = xor_dest;
Dan Williams91c00922007-01-02 13:52:30 -07001392 for (i = disks; i--; ) {
Dan Williamsac6b53b2009-07-14 13:40:19 -07001393 if (i == pd_idx || i == qd_idx)
1394 continue;
1395 xor_srcs[count++] = sh->dev[i].page;
Dan Williams91c00922007-01-02 13:52:30 -07001396 }
1397
Dan Williamsd6f38f32009-07-14 11:50:52 -07001398 init_async_submit(&submit, 0, NULL, NULL, NULL,
1399 to_addr_conv(sh, percpu));
Dan Williams099f53c2009-04-08 14:28:37 -07001400 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
Dan Williamsa08abd82009-06-03 11:43:59 -07001401 &sh->ops.zero_sum_result, &submit);
Dan Williams91c00922007-01-02 13:52:30 -07001402
Dan Williams91c00922007-01-02 13:52:30 -07001403 atomic_inc(&sh->count);
Dan Williamsa08abd82009-06-03 11:43:59 -07001404 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1405 tx = async_trigger_callback(&submit);
Dan Williams91c00922007-01-02 13:52:30 -07001406}
1407
Dan Williamsac6b53b2009-07-14 13:40:19 -07001408static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1409{
1410 struct page **srcs = percpu->scribble;
1411 struct async_submit_ctl submit;
1412 int count;
1413
1414 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1415 (unsigned long long)sh->sector, checkp);
1416
1417 count = set_syndrome_sources(srcs, sh);
1418 if (!checkp)
1419 srcs[count] = NULL;
1420
1421 atomic_inc(&sh->count);
1422 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1423 sh, to_addr_conv(sh, percpu));
1424 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1425 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1426}
1427
NeilBrown51acbce2013-02-28 09:08:34 +11001428static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
Dan Williams91c00922007-01-02 13:52:30 -07001429{
1430 int overlap_clear = 0, i, disks = sh->disks;
1431 struct dma_async_tx_descriptor *tx = NULL;
NeilBrownd1688a62011-10-11 16:49:52 +11001432 struct r5conf *conf = sh->raid_conf;
Dan Williamsac6b53b2009-07-14 13:40:19 -07001433 int level = conf->level;
Dan Williamsd6f38f32009-07-14 11:50:52 -07001434 struct raid5_percpu *percpu;
1435 unsigned long cpu;
Dan Williams91c00922007-01-02 13:52:30 -07001436
Dan Williamsd6f38f32009-07-14 11:50:52 -07001437 cpu = get_cpu();
1438 percpu = per_cpu_ptr(conf->percpu, cpu);
Dan Williams83de75c2008-06-28 08:31:58 +10001439 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
Dan Williams91c00922007-01-02 13:52:30 -07001440 ops_run_biofill(sh);
1441 overlap_clear++;
1442 }
1443
Dan Williams7b3a8712008-06-28 08:32:09 +10001444 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
Dan Williamsac6b53b2009-07-14 13:40:19 -07001445 if (level < 6)
1446 tx = ops_run_compute5(sh, percpu);
1447 else {
1448 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1449 tx = ops_run_compute6_1(sh, percpu);
1450 else
1451 tx = ops_run_compute6_2(sh, percpu);
1452 }
1453 /* terminate the chain if reconstruct is not set to be run */
1454 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
Dan Williams7b3a8712008-06-28 08:32:09 +10001455 async_tx_ack(tx);
1456 }
Dan Williams91c00922007-01-02 13:52:30 -07001457
Dan Williams600aa102008-06-28 08:32:05 +10001458 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
Dan Williamsd6f38f32009-07-14 11:50:52 -07001459 tx = ops_run_prexor(sh, percpu, tx);
Dan Williams91c00922007-01-02 13:52:30 -07001460
Dan Williams600aa102008-06-28 08:32:05 +10001461 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
Dan Williamsd8ee0722008-06-28 08:32:06 +10001462 tx = ops_run_biodrain(sh, tx);
Dan Williams91c00922007-01-02 13:52:30 -07001463 overlap_clear++;
1464 }
1465
Dan Williamsac6b53b2009-07-14 13:40:19 -07001466 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1467 if (level < 6)
1468 ops_run_reconstruct5(sh, percpu, tx);
1469 else
1470 ops_run_reconstruct6(sh, percpu, tx);
1471 }
Dan Williams91c00922007-01-02 13:52:30 -07001472
Dan Williamsac6b53b2009-07-14 13:40:19 -07001473 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1474 if (sh->check_state == check_state_run)
1475 ops_run_check_p(sh, percpu);
1476 else if (sh->check_state == check_state_run_q)
1477 ops_run_check_pq(sh, percpu, 0);
1478 else if (sh->check_state == check_state_run_pq)
1479 ops_run_check_pq(sh, percpu, 1);
1480 else
1481 BUG();
1482 }
Dan Williams91c00922007-01-02 13:52:30 -07001483
Dan Williams91c00922007-01-02 13:52:30 -07001484 if (overlap_clear)
1485 for (i = disks; i--; ) {
1486 struct r5dev *dev = &sh->dev[i];
1487 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1488 wake_up(&sh->raid_conf->wait_for_overlap);
1489 }
Dan Williamsd6f38f32009-07-14 11:50:52 -07001490 put_cpu();
Dan Williams91c00922007-01-02 13:52:30 -07001491}
1492
NeilBrownd1688a62011-10-11 16:49:52 +11001493static int grow_one_stripe(struct r5conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494{
1495 struct stripe_head *sh;
Namhyung Kim6ce32842011-07-18 17:38:50 +10001496 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
NeilBrown3f294f42005-11-08 21:39:25 -08001497 if (!sh)
1498 return 0;
Namhyung Kim6ce32842011-07-18 17:38:50 +10001499
NeilBrown3f294f42005-11-08 21:39:25 -08001500 sh->raid_conf = conf;
NeilBrown3f294f42005-11-08 21:39:25 -08001501
Shaohua Lib17459c2012-07-19 16:01:31 +10001502 spin_lock_init(&sh->stripe_lock);
1503
NeilBrowne4e11e32010-06-16 16:45:16 +10001504 if (grow_buffers(sh)) {
1505 shrink_buffers(sh);
NeilBrown3f294f42005-11-08 21:39:25 -08001506 kmem_cache_free(conf->slab_cache, sh);
1507 return 0;
1508 }
1509 /* we just created an active stripe so... */
1510 atomic_set(&sh->count, 1);
1511 atomic_inc(&conf->active_stripes);
1512 INIT_LIST_HEAD(&sh->lru);
1513 release_stripe(sh);
1514 return 1;
1515}
1516
NeilBrownd1688a62011-10-11 16:49:52 +11001517static int grow_stripes(struct r5conf *conf, int num)
NeilBrown3f294f42005-11-08 21:39:25 -08001518{
Christoph Lametere18b8902006-12-06 20:33:20 -08001519 struct kmem_cache *sc;
NeilBrown5e5e3e72009-10-16 16:35:30 +11001520 int devs = max(conf->raid_disks, conf->previous_raid_disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521
NeilBrownf4be6b42010-06-01 19:37:25 +10001522 if (conf->mddev->gendisk)
1523 sprintf(conf->cache_name[0],
1524 "raid%d-%s", conf->level, mdname(conf->mddev));
1525 else
1526 sprintf(conf->cache_name[0],
1527 "raid%d-%p", conf->level, conf->mddev);
1528 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1529
NeilBrownad01c9e2006-03-27 01:18:07 -08001530 conf->active_name = 0;
1531 sc = kmem_cache_create(conf->cache_name[conf->active_name],
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
Paul Mundt20c2df82007-07-20 10:11:58 +09001533 0, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 if (!sc)
1535 return 1;
1536 conf->slab_cache = sc;
NeilBrownad01c9e2006-03-27 01:18:07 -08001537 conf->pool_size = devs;
NeilBrown16a53ec2006-06-26 00:27:38 -07001538 while (num--)
NeilBrown3f294f42005-11-08 21:39:25 -08001539 if (!grow_one_stripe(conf))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541 return 0;
1542}
NeilBrown29269552006-03-27 01:18:10 -08001543
Dan Williamsd6f38f32009-07-14 11:50:52 -07001544/**
1545 * scribble_len - return the required size of the scribble region
1546 * @num - total number of disks in the array
1547 *
1548 * The size must be enough to contain:
1549 * 1/ a struct page pointer for each device in the array +2
1550 * 2/ room to convert each entry in (1) to its corresponding dma
1551 * (dma_map_page()) or page (page_address()) address.
1552 *
1553 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1554 * calculate over all devices (not just the data blocks), using zeros in place
1555 * of the P and Q blocks.
1556 */
1557static size_t scribble_len(int num)
1558{
1559 size_t len;
1560
1561 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1562
1563 return len;
1564}
1565
NeilBrownd1688a62011-10-11 16:49:52 +11001566static int resize_stripes(struct r5conf *conf, int newsize)
NeilBrownad01c9e2006-03-27 01:18:07 -08001567{
1568 /* Make all the stripes able to hold 'newsize' devices.
1569 * New slots in each stripe get 'page' set to a new page.
1570 *
1571 * This happens in stages:
1572 * 1/ create a new kmem_cache and allocate the required number of
1573 * stripe_heads.
Masanari Iida83f0d772012-10-30 00:18:08 +09001574 * 2/ gather all the old stripe_heads and transfer the pages across
NeilBrownad01c9e2006-03-27 01:18:07 -08001575 * to the new stripe_heads. This will have the side effect of
1576 * freezing the array as once all stripe_heads have been collected,
1577 * no IO will be possible. Old stripe heads are freed once their
1578 * pages have been transferred over, and the old kmem_cache is
1579 * freed when all stripes are done.
1580 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1581 * we simple return a failre status - no need to clean anything up.
1582 * 4/ allocate new pages for the new slots in the new stripe_heads.
1583 * If this fails, we don't bother trying the shrink the
1584 * stripe_heads down again, we just leave them as they are.
1585 * As each stripe_head is processed the new one is released into
1586 * active service.
1587 *
1588 * Once step2 is started, we cannot afford to wait for a write,
1589 * so we use GFP_NOIO allocations.
1590 */
1591 struct stripe_head *osh, *nsh;
1592 LIST_HEAD(newstripes);
1593 struct disk_info *ndisks;
Dan Williamsd6f38f32009-07-14 11:50:52 -07001594 unsigned long cpu;
Dan Williamsb5470dc2008-06-27 21:44:04 -07001595 int err;
Christoph Lametere18b8902006-12-06 20:33:20 -08001596 struct kmem_cache *sc;
NeilBrownad01c9e2006-03-27 01:18:07 -08001597 int i;
1598
1599 if (newsize <= conf->pool_size)
1600 return 0; /* never bother to shrink */
1601
Dan Williamsb5470dc2008-06-27 21:44:04 -07001602 err = md_allow_write(conf->mddev);
1603 if (err)
1604 return err;
NeilBrown2a2275d2007-01-26 00:57:11 -08001605
NeilBrownad01c9e2006-03-27 01:18:07 -08001606 /* Step 1 */
1607 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1608 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
Paul Mundt20c2df82007-07-20 10:11:58 +09001609 0, 0, NULL);
NeilBrownad01c9e2006-03-27 01:18:07 -08001610 if (!sc)
1611 return -ENOMEM;
1612
1613 for (i = conf->max_nr_stripes; i; i--) {
Namhyung Kim6ce32842011-07-18 17:38:50 +10001614 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
NeilBrownad01c9e2006-03-27 01:18:07 -08001615 if (!nsh)
1616 break;
1617
NeilBrownad01c9e2006-03-27 01:18:07 -08001618 nsh->raid_conf = conf;
NeilBrowncb13ff62012-09-24 16:27:20 +10001619 spin_lock_init(&nsh->stripe_lock);
NeilBrownad01c9e2006-03-27 01:18:07 -08001620
1621 list_add(&nsh->lru, &newstripes);
1622 }
1623 if (i) {
1624 /* didn't get enough, give up */
1625 while (!list_empty(&newstripes)) {
1626 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1627 list_del(&nsh->lru);
1628 kmem_cache_free(sc, nsh);
1629 }
1630 kmem_cache_destroy(sc);
1631 return -ENOMEM;
1632 }
1633 /* Step 2 - Must use GFP_NOIO now.
1634 * OK, we have enough stripes, start collecting inactive
1635 * stripes and copying them over
1636 */
1637 list_for_each_entry(nsh, &newstripes, lru) {
1638 spin_lock_irq(&conf->device_lock);
1639 wait_event_lock_irq(conf->wait_for_stripe,
1640 !list_empty(&conf->inactive_list),
Lukas Czernereed8c022012-11-30 11:42:40 +01001641 conf->device_lock);
NeilBrownad01c9e2006-03-27 01:18:07 -08001642 osh = get_free_stripe(conf);
1643 spin_unlock_irq(&conf->device_lock);
1644 atomic_set(&nsh->count, 1);
1645 for(i=0; i<conf->pool_size; i++)
1646 nsh->dev[i].page = osh->dev[i].page;
1647 for( ; i<newsize; i++)
1648 nsh->dev[i].page = NULL;
1649 kmem_cache_free(conf->slab_cache, osh);
1650 }
1651 kmem_cache_destroy(conf->slab_cache);
1652
1653 /* Step 3.
1654 * At this point, we are holding all the stripes so the array
1655 * is completely stalled, so now is a good time to resize
Dan Williamsd6f38f32009-07-14 11:50:52 -07001656 * conf->disks and the scribble region
NeilBrownad01c9e2006-03-27 01:18:07 -08001657 */
1658 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1659 if (ndisks) {
1660 for (i=0; i<conf->raid_disks; i++)
1661 ndisks[i] = conf->disks[i];
1662 kfree(conf->disks);
1663 conf->disks = ndisks;
1664 } else
1665 err = -ENOMEM;
1666
Dan Williamsd6f38f32009-07-14 11:50:52 -07001667 get_online_cpus();
1668 conf->scribble_len = scribble_len(newsize);
1669 for_each_present_cpu(cpu) {
1670 struct raid5_percpu *percpu;
1671 void *scribble;
1672
1673 percpu = per_cpu_ptr(conf->percpu, cpu);
1674 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1675
1676 if (scribble) {
1677 kfree(percpu->scribble);
1678 percpu->scribble = scribble;
1679 } else {
1680 err = -ENOMEM;
1681 break;
1682 }
1683 }
1684 put_online_cpus();
1685
NeilBrownad01c9e2006-03-27 01:18:07 -08001686 /* Step 4, return new stripes to service */
1687 while(!list_empty(&newstripes)) {
1688 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1689 list_del_init(&nsh->lru);
Dan Williamsd6f38f32009-07-14 11:50:52 -07001690
NeilBrownad01c9e2006-03-27 01:18:07 -08001691 for (i=conf->raid_disks; i < newsize; i++)
1692 if (nsh->dev[i].page == NULL) {
1693 struct page *p = alloc_page(GFP_NOIO);
1694 nsh->dev[i].page = p;
1695 if (!p)
1696 err = -ENOMEM;
1697 }
1698 release_stripe(nsh);
1699 }
1700 /* critical section pass, GFP_NOIO no longer needed */
1701
1702 conf->slab_cache = sc;
1703 conf->active_name = 1-conf->active_name;
1704 conf->pool_size = newsize;
1705 return err;
1706}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707
NeilBrownd1688a62011-10-11 16:49:52 +11001708static int drop_one_stripe(struct r5conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709{
1710 struct stripe_head *sh;
1711
NeilBrown3f294f42005-11-08 21:39:25 -08001712 spin_lock_irq(&conf->device_lock);
1713 sh = get_free_stripe(conf);
1714 spin_unlock_irq(&conf->device_lock);
1715 if (!sh)
1716 return 0;
Eric Sesterhenn78bafeb2006-04-02 13:31:42 +02001717 BUG_ON(atomic_read(&sh->count));
NeilBrowne4e11e32010-06-16 16:45:16 +10001718 shrink_buffers(sh);
NeilBrown3f294f42005-11-08 21:39:25 -08001719 kmem_cache_free(conf->slab_cache, sh);
1720 atomic_dec(&conf->active_stripes);
1721 return 1;
1722}
1723
NeilBrownd1688a62011-10-11 16:49:52 +11001724static void shrink_stripes(struct r5conf *conf)
NeilBrown3f294f42005-11-08 21:39:25 -08001725{
1726 while (drop_one_stripe(conf))
1727 ;
1728
NeilBrown29fc7e32006-02-03 03:03:41 -08001729 if (conf->slab_cache)
1730 kmem_cache_destroy(conf->slab_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731 conf->slab_cache = NULL;
1732}
1733
NeilBrown6712ecf2007-09-27 12:47:43 +02001734static void raid5_end_read_request(struct bio * bi, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735{
NeilBrown99c0fb52009-03-31 14:39:38 +11001736 struct stripe_head *sh = bi->bi_private;
NeilBrownd1688a62011-10-11 16:49:52 +11001737 struct r5conf *conf = sh->raid_conf;
NeilBrown7ecaa1e2006-03-27 01:18:08 -08001738 int disks = sh->disks, i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
NeilBrownd6950432006-07-10 04:44:20 -07001740 char b[BDEVNAME_SIZE];
NeilBrowndd054fc2011-12-23 10:17:53 +11001741 struct md_rdev *rdev = NULL;
NeilBrown05616be2012-05-21 09:27:00 +10001742 sector_t s;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743
1744 for (i=0 ; i<disks; i++)
1745 if (bi == &sh->dev[i].req)
1746 break;
1747
Dan Williams45b42332007-07-09 11:56:43 -07001748 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1749 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 uptodate);
1751 if (i == disks) {
1752 BUG();
NeilBrown6712ecf2007-09-27 12:47:43 +02001753 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 }
NeilBrown14a75d32011-12-23 10:17:52 +11001755 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
NeilBrowndd054fc2011-12-23 10:17:53 +11001756 /* If replacement finished while this request was outstanding,
1757 * 'replacement' might be NULL already.
1758 * In that case it moved down to 'rdev'.
1759 * rdev is not removed until all requests are finished.
1760 */
NeilBrown14a75d32011-12-23 10:17:52 +11001761 rdev = conf->disks[i].replacement;
NeilBrowndd054fc2011-12-23 10:17:53 +11001762 if (!rdev)
NeilBrown14a75d32011-12-23 10:17:52 +11001763 rdev = conf->disks[i].rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764
NeilBrown05616be2012-05-21 09:27:00 +10001765 if (use_new_offset(conf, sh))
1766 s = sh->sector + rdev->new_data_offset;
1767 else
1768 s = sh->sector + rdev->data_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 if (uptodate) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770 set_bit(R5_UPTODATE, &sh->dev[i].flags);
NeilBrown4e5314b2005-11-08 21:39:22 -08001771 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
NeilBrown14a75d32011-12-23 10:17:52 +11001772 /* Note that this cannot happen on a
1773 * replacement device. We just fail those on
1774 * any error
1775 */
Christian Dietrich8bda4702011-07-27 11:00:36 +10001776 printk_ratelimited(
1777 KERN_INFO
1778 "md/raid:%s: read error corrected"
1779 " (%lu sectors at %llu on %s)\n",
1780 mdname(conf->mddev), STRIPE_SECTORS,
NeilBrown05616be2012-05-21 09:27:00 +10001781 (unsigned long long)s,
Christian Dietrich8bda4702011-07-27 11:00:36 +10001782 bdevname(rdev->bdev, b));
Namhyung Kimddd51152011-07-27 11:00:36 +10001783 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
NeilBrown4e5314b2005-11-08 21:39:22 -08001784 clear_bit(R5_ReadError, &sh->dev[i].flags);
1785 clear_bit(R5_ReWrite, &sh->dev[i].flags);
majianpeng3f9e7c12012-07-31 10:04:21 +10001786 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1787 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1788
NeilBrown14a75d32011-12-23 10:17:52 +11001789 if (atomic_read(&rdev->read_errors))
1790 atomic_set(&rdev->read_errors, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791 } else {
NeilBrown14a75d32011-12-23 10:17:52 +11001792 const char *bdn = bdevname(rdev->bdev, b);
NeilBrownba22dcb2005-11-08 21:39:31 -08001793 int retry = 0;
majianpeng2e8ac3032012-07-03 15:57:02 +10001794 int set_bad = 0;
NeilBrownd6950432006-07-10 04:44:20 -07001795
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
NeilBrownd6950432006-07-10 04:44:20 -07001797 atomic_inc(&rdev->read_errors);
NeilBrown14a75d32011-12-23 10:17:52 +11001798 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1799 printk_ratelimited(
1800 KERN_WARNING
1801 "md/raid:%s: read error on replacement device "
1802 "(sector %llu on %s).\n",
1803 mdname(conf->mddev),
NeilBrown05616be2012-05-21 09:27:00 +10001804 (unsigned long long)s,
NeilBrown14a75d32011-12-23 10:17:52 +11001805 bdn);
majianpeng2e8ac3032012-07-03 15:57:02 +10001806 else if (conf->mddev->degraded >= conf->max_degraded) {
1807 set_bad = 1;
Christian Dietrich8bda4702011-07-27 11:00:36 +10001808 printk_ratelimited(
1809 KERN_WARNING
1810 "md/raid:%s: read error not correctable "
1811 "(sector %llu on %s).\n",
1812 mdname(conf->mddev),
NeilBrown05616be2012-05-21 09:27:00 +10001813 (unsigned long long)s,
Christian Dietrich8bda4702011-07-27 11:00:36 +10001814 bdn);
majianpeng2e8ac3032012-07-03 15:57:02 +10001815 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
NeilBrown4e5314b2005-11-08 21:39:22 -08001816 /* Oh, no!!! */
majianpeng2e8ac3032012-07-03 15:57:02 +10001817 set_bad = 1;
Christian Dietrich8bda4702011-07-27 11:00:36 +10001818 printk_ratelimited(
1819 KERN_WARNING
1820 "md/raid:%s: read error NOT corrected!! "
1821 "(sector %llu on %s).\n",
1822 mdname(conf->mddev),
NeilBrown05616be2012-05-21 09:27:00 +10001823 (unsigned long long)s,
Christian Dietrich8bda4702011-07-27 11:00:36 +10001824 bdn);
majianpeng2e8ac3032012-07-03 15:57:02 +10001825 } else if (atomic_read(&rdev->read_errors)
NeilBrownba22dcb2005-11-08 21:39:31 -08001826 > conf->max_nr_stripes)
NeilBrown14f8d262006-01-06 00:20:14 -08001827 printk(KERN_WARNING
NeilBrown0c55e022010-05-03 14:09:02 +10001828 "md/raid:%s: Too many read errors, failing device %s.\n",
NeilBrownd6950432006-07-10 04:44:20 -07001829 mdname(conf->mddev), bdn);
NeilBrownba22dcb2005-11-08 21:39:31 -08001830 else
1831 retry = 1;
1832 if (retry)
majianpeng3f9e7c12012-07-31 10:04:21 +10001833 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1834 set_bit(R5_ReadError, &sh->dev[i].flags);
1835 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1836 } else
1837 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
NeilBrownba22dcb2005-11-08 21:39:31 -08001838 else {
NeilBrown4e5314b2005-11-08 21:39:22 -08001839 clear_bit(R5_ReadError, &sh->dev[i].flags);
1840 clear_bit(R5_ReWrite, &sh->dev[i].flags);
majianpeng2e8ac3032012-07-03 15:57:02 +10001841 if (!(set_bad
1842 && test_bit(In_sync, &rdev->flags)
1843 && rdev_set_badblocks(
1844 rdev, sh->sector, STRIPE_SECTORS, 0)))
1845 md_error(conf->mddev, rdev);
NeilBrownba22dcb2005-11-08 21:39:31 -08001846 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 }
NeilBrown14a75d32011-12-23 10:17:52 +11001848 rdev_dec_pending(rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1850 set_bit(STRIPE_HANDLE, &sh->state);
1851 release_stripe(sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852}
1853
NeilBrownd710e132008-10-13 11:55:12 +11001854static void raid5_end_write_request(struct bio *bi, int error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855{
NeilBrown99c0fb52009-03-31 14:39:38 +11001856 struct stripe_head *sh = bi->bi_private;
NeilBrownd1688a62011-10-11 16:49:52 +11001857 struct r5conf *conf = sh->raid_conf;
NeilBrown7ecaa1e2006-03-27 01:18:08 -08001858 int disks = sh->disks, i;
NeilBrown977df362011-12-23 10:17:53 +11001859 struct md_rdev *uninitialized_var(rdev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
NeilBrownb84db562011-07-28 11:39:23 +10001861 sector_t first_bad;
1862 int bad_sectors;
NeilBrown977df362011-12-23 10:17:53 +11001863 int replacement = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864
NeilBrown977df362011-12-23 10:17:53 +11001865 for (i = 0 ; i < disks; i++) {
1866 if (bi == &sh->dev[i].req) {
1867 rdev = conf->disks[i].rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868 break;
NeilBrown977df362011-12-23 10:17:53 +11001869 }
1870 if (bi == &sh->dev[i].rreq) {
1871 rdev = conf->disks[i].replacement;
NeilBrowndd054fc2011-12-23 10:17:53 +11001872 if (rdev)
1873 replacement = 1;
1874 else
1875 /* rdev was removed and 'replacement'
1876 * replaced it. rdev is not removed
1877 * until all requests are finished.
1878 */
1879 rdev = conf->disks[i].rdev;
NeilBrown977df362011-12-23 10:17:53 +11001880 break;
1881 }
1882 }
Dan Williams45b42332007-07-09 11:56:43 -07001883 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1885 uptodate);
1886 if (i == disks) {
1887 BUG();
NeilBrown6712ecf2007-09-27 12:47:43 +02001888 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 }
1890
NeilBrown977df362011-12-23 10:17:53 +11001891 if (replacement) {
1892 if (!uptodate)
1893 md_error(conf->mddev, rdev);
1894 else if (is_badblock(rdev, sh->sector,
1895 STRIPE_SECTORS,
1896 &first_bad, &bad_sectors))
1897 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1898 } else {
1899 if (!uptodate) {
NeilBrown6ba854e2014-01-16 09:35:38 +11001900 set_bit(STRIPE_DEGRADED, &sh->state);
NeilBrown977df362011-12-23 10:17:53 +11001901 set_bit(WriteErrorSeen, &rdev->flags);
1902 set_bit(R5_WriteError, &sh->dev[i].flags);
NeilBrown3a6de292011-12-23 10:17:54 +11001903 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1904 set_bit(MD_RECOVERY_NEEDED,
1905 &rdev->mddev->recovery);
NeilBrown977df362011-12-23 10:17:53 +11001906 } else if (is_badblock(rdev, sh->sector,
1907 STRIPE_SECTORS,
NeilBrownc0b32972013-04-24 11:42:42 +10001908 &first_bad, &bad_sectors)) {
NeilBrown977df362011-12-23 10:17:53 +11001909 set_bit(R5_MadeGood, &sh->dev[i].flags);
NeilBrownc0b32972013-04-24 11:42:42 +10001910 if (test_bit(R5_ReadError, &sh->dev[i].flags))
1911 /* That was a successful write so make
1912 * sure it looks like we already did
1913 * a re-write.
1914 */
1915 set_bit(R5_ReWrite, &sh->dev[i].flags);
1916 }
NeilBrown977df362011-12-23 10:17:53 +11001917 }
1918 rdev_dec_pending(rdev, conf->mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919
NeilBrown977df362011-12-23 10:17:53 +11001920 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1921 clear_bit(R5_LOCKED, &sh->dev[i].flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922 set_bit(STRIPE_HANDLE, &sh->state);
NeilBrownc04be0a2006-10-03 01:15:53 -07001923 release_stripe(sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924}
1925
NeilBrown784052e2009-03-31 15:19:07 +11001926static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927
NeilBrown784052e2009-03-31 15:19:07 +11001928static void raid5_build_block(struct stripe_head *sh, int i, int previous)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929{
1930 struct r5dev *dev = &sh->dev[i];
1931
1932 bio_init(&dev->req);
1933 dev->req.bi_io_vec = &dev->vec;
1934 dev->req.bi_vcnt++;
1935 dev->req.bi_max_vecs++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936 dev->req.bi_private = sh;
NeilBrown995c4272011-12-23 10:17:52 +11001937 dev->vec.bv_page = dev->page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938
NeilBrown977df362011-12-23 10:17:53 +11001939 bio_init(&dev->rreq);
1940 dev->rreq.bi_io_vec = &dev->rvec;
1941 dev->rreq.bi_vcnt++;
1942 dev->rreq.bi_max_vecs++;
1943 dev->rreq.bi_private = sh;
1944 dev->rvec.bv_page = dev->page;
1945
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946 dev->flags = 0;
NeilBrown784052e2009-03-31 15:19:07 +11001947 dev->sector = compute_blocknr(sh, i, previous);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948}
1949
NeilBrownfd01b882011-10-11 16:47:53 +11001950static void error(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951{
1952 char b[BDEVNAME_SIZE];
NeilBrownd1688a62011-10-11 16:49:52 +11001953 struct r5conf *conf = mddev->private;
NeilBrown908f4fb2011-12-23 10:17:50 +11001954 unsigned long flags;
NeilBrown0c55e022010-05-03 14:09:02 +10001955 pr_debug("raid456: error called\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956
NeilBrown908f4fb2011-12-23 10:17:50 +11001957 spin_lock_irqsave(&conf->device_lock, flags);
1958 clear_bit(In_sync, &rdev->flags);
1959 mddev->degraded = calc_degraded(conf);
1960 spin_unlock_irqrestore(&conf->device_lock, flags);
1961 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1962
NeilBrownde393cd2011-07-28 11:31:48 +10001963 set_bit(Blocked, &rdev->flags);
NeilBrown6f8d0c72011-05-11 14:38:44 +10001964 set_bit(Faulty, &rdev->flags);
1965 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1966 printk(KERN_ALERT
1967 "md/raid:%s: Disk failure on %s, disabling device.\n"
1968 "md/raid:%s: Operation continuing on %d devices.\n",
1969 mdname(mddev),
1970 bdevname(rdev->bdev, b),
1971 mdname(mddev),
1972 conf->raid_disks - mddev->degraded);
NeilBrown16a53ec2006-06-26 00:27:38 -07001973}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974
1975/*
1976 * Input: a 'big' sector number,
1977 * Output: index of the data and parity disk, and the sector # in them.
1978 */
NeilBrownd1688a62011-10-11 16:49:52 +11001979static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
NeilBrown911d4ee2009-03-31 14:39:38 +11001980 int previous, int *dd_idx,
1981 struct stripe_head *sh)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982{
NeilBrown6e3b96e2010-04-23 07:08:28 +10001983 sector_t stripe, stripe2;
NeilBrown35f2a592010-04-20 14:13:34 +10001984 sector_t chunk_number;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 unsigned int chunk_offset;
NeilBrown911d4ee2009-03-31 14:39:38 +11001986 int pd_idx, qd_idx;
NeilBrown67cc2b82009-03-31 14:39:38 +11001987 int ddf_layout = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988 sector_t new_sector;
NeilBrowne183eae2009-03-31 15:20:22 +11001989 int algorithm = previous ? conf->prev_algo
1990 : conf->algorithm;
Andre Noll09c9e5f2009-06-18 08:45:55 +10001991 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1992 : conf->chunk_sectors;
NeilBrown112bf892009-03-31 14:39:38 +11001993 int raid_disks = previous ? conf->previous_raid_disks
1994 : conf->raid_disks;
1995 int data_disks = raid_disks - conf->max_degraded;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996
1997 /* First compute the information on this sector */
1998
1999 /*
2000 * Compute the chunk number and the sector offset inside the chunk
2001 */
2002 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2003 chunk_number = r_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004
2005 /*
2006 * Compute the stripe number
2007 */
NeilBrown35f2a592010-04-20 14:13:34 +10002008 stripe = chunk_number;
2009 *dd_idx = sector_div(stripe, data_disks);
NeilBrown6e3b96e2010-04-23 07:08:28 +10002010 stripe2 = stripe;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011 /*
2012 * Select the parity disk based on the user selected algorithm.
2013 */
NeilBrown84789552011-07-27 11:00:36 +10002014 pd_idx = qd_idx = -1;
NeilBrown16a53ec2006-06-26 00:27:38 -07002015 switch(conf->level) {
2016 case 4:
NeilBrown911d4ee2009-03-31 14:39:38 +11002017 pd_idx = data_disks;
NeilBrown16a53ec2006-06-26 00:27:38 -07002018 break;
2019 case 5:
NeilBrowne183eae2009-03-31 15:20:22 +11002020 switch (algorithm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021 case ALGORITHM_LEFT_ASYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002022 pd_idx = data_disks - sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002023 if (*dd_idx >= pd_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002024 (*dd_idx)++;
2025 break;
2026 case ALGORITHM_RIGHT_ASYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002027 pd_idx = sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002028 if (*dd_idx >= pd_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029 (*dd_idx)++;
2030 break;
2031 case ALGORITHM_LEFT_SYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002032 pd_idx = data_disks - sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002033 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034 break;
2035 case ALGORITHM_RIGHT_SYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002036 pd_idx = sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002037 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038 break;
NeilBrown99c0fb52009-03-31 14:39:38 +11002039 case ALGORITHM_PARITY_0:
2040 pd_idx = 0;
2041 (*dd_idx)++;
2042 break;
2043 case ALGORITHM_PARITY_N:
2044 pd_idx = data_disks;
2045 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046 default:
NeilBrown99c0fb52009-03-31 14:39:38 +11002047 BUG();
NeilBrown16a53ec2006-06-26 00:27:38 -07002048 }
2049 break;
2050 case 6:
2051
NeilBrowne183eae2009-03-31 15:20:22 +11002052 switch (algorithm) {
NeilBrown16a53ec2006-06-26 00:27:38 -07002053 case ALGORITHM_LEFT_ASYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002054 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002055 qd_idx = pd_idx + 1;
2056 if (pd_idx == raid_disks-1) {
NeilBrown99c0fb52009-03-31 14:39:38 +11002057 (*dd_idx)++; /* Q D D D P */
NeilBrown911d4ee2009-03-31 14:39:38 +11002058 qd_idx = 0;
2059 } else if (*dd_idx >= pd_idx)
NeilBrown16a53ec2006-06-26 00:27:38 -07002060 (*dd_idx) += 2; /* D D P Q D */
2061 break;
2062 case ALGORITHM_RIGHT_ASYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002063 pd_idx = sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002064 qd_idx = pd_idx + 1;
2065 if (pd_idx == raid_disks-1) {
NeilBrown99c0fb52009-03-31 14:39:38 +11002066 (*dd_idx)++; /* Q D D D P */
NeilBrown911d4ee2009-03-31 14:39:38 +11002067 qd_idx = 0;
2068 } else if (*dd_idx >= pd_idx)
NeilBrown16a53ec2006-06-26 00:27:38 -07002069 (*dd_idx) += 2; /* D D P Q D */
2070 break;
2071 case ALGORITHM_LEFT_SYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002072 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002073 qd_idx = (pd_idx + 1) % raid_disks;
2074 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
NeilBrown16a53ec2006-06-26 00:27:38 -07002075 break;
2076 case ALGORITHM_RIGHT_SYMMETRIC:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002077 pd_idx = sector_div(stripe2, raid_disks);
NeilBrown911d4ee2009-03-31 14:39:38 +11002078 qd_idx = (pd_idx + 1) % raid_disks;
2079 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
NeilBrown16a53ec2006-06-26 00:27:38 -07002080 break;
NeilBrown99c0fb52009-03-31 14:39:38 +11002081
2082 case ALGORITHM_PARITY_0:
2083 pd_idx = 0;
2084 qd_idx = 1;
2085 (*dd_idx) += 2;
2086 break;
2087 case ALGORITHM_PARITY_N:
2088 pd_idx = data_disks;
2089 qd_idx = data_disks + 1;
2090 break;
2091
2092 case ALGORITHM_ROTATING_ZERO_RESTART:
2093 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2094 * of blocks for computing Q is different.
2095 */
NeilBrown6e3b96e2010-04-23 07:08:28 +10002096 pd_idx = sector_div(stripe2, raid_disks);
NeilBrown99c0fb52009-03-31 14:39:38 +11002097 qd_idx = pd_idx + 1;
2098 if (pd_idx == raid_disks-1) {
2099 (*dd_idx)++; /* Q D D D P */
2100 qd_idx = 0;
2101 } else if (*dd_idx >= pd_idx)
2102 (*dd_idx) += 2; /* D D P Q D */
NeilBrown67cc2b82009-03-31 14:39:38 +11002103 ddf_layout = 1;
NeilBrown99c0fb52009-03-31 14:39:38 +11002104 break;
2105
2106 case ALGORITHM_ROTATING_N_RESTART:
2107 /* Same a left_asymmetric, by first stripe is
2108 * D D D P Q rather than
2109 * Q D D D P
2110 */
NeilBrown6e3b96e2010-04-23 07:08:28 +10002111 stripe2 += 1;
2112 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
NeilBrown99c0fb52009-03-31 14:39:38 +11002113 qd_idx = pd_idx + 1;
2114 if (pd_idx == raid_disks-1) {
2115 (*dd_idx)++; /* Q D D D P */
2116 qd_idx = 0;
2117 } else if (*dd_idx >= pd_idx)
2118 (*dd_idx) += 2; /* D D P Q D */
NeilBrown67cc2b82009-03-31 14:39:38 +11002119 ddf_layout = 1;
NeilBrown99c0fb52009-03-31 14:39:38 +11002120 break;
2121
2122 case ALGORITHM_ROTATING_N_CONTINUE:
2123 /* Same as left_symmetric but Q is before P */
NeilBrown6e3b96e2010-04-23 07:08:28 +10002124 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
NeilBrown99c0fb52009-03-31 14:39:38 +11002125 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2126 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
NeilBrown67cc2b82009-03-31 14:39:38 +11002127 ddf_layout = 1;
NeilBrown99c0fb52009-03-31 14:39:38 +11002128 break;
2129
2130 case ALGORITHM_LEFT_ASYMMETRIC_6:
2131 /* RAID5 left_asymmetric, with Q on last device */
NeilBrown6e3b96e2010-04-23 07:08:28 +10002132 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
NeilBrown99c0fb52009-03-31 14:39:38 +11002133 if (*dd_idx >= pd_idx)
2134 (*dd_idx)++;
2135 qd_idx = raid_disks - 1;
2136 break;
2137
2138 case ALGORITHM_RIGHT_ASYMMETRIC_6:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002139 pd_idx = sector_div(stripe2, raid_disks-1);
NeilBrown99c0fb52009-03-31 14:39:38 +11002140 if (*dd_idx >= pd_idx)
2141 (*dd_idx)++;
2142 qd_idx = raid_disks - 1;
2143 break;
2144
2145 case ALGORITHM_LEFT_SYMMETRIC_6:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002146 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
NeilBrown99c0fb52009-03-31 14:39:38 +11002147 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2148 qd_idx = raid_disks - 1;
2149 break;
2150
2151 case ALGORITHM_RIGHT_SYMMETRIC_6:
NeilBrown6e3b96e2010-04-23 07:08:28 +10002152 pd_idx = sector_div(stripe2, raid_disks-1);
NeilBrown99c0fb52009-03-31 14:39:38 +11002153 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2154 qd_idx = raid_disks - 1;
2155 break;
2156
2157 case ALGORITHM_PARITY_0_6:
2158 pd_idx = 0;
2159 (*dd_idx)++;
2160 qd_idx = raid_disks - 1;
2161 break;
2162
NeilBrown16a53ec2006-06-26 00:27:38 -07002163 default:
NeilBrown99c0fb52009-03-31 14:39:38 +11002164 BUG();
NeilBrown16a53ec2006-06-26 00:27:38 -07002165 }
2166 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 }
2168
NeilBrown911d4ee2009-03-31 14:39:38 +11002169 if (sh) {
2170 sh->pd_idx = pd_idx;
2171 sh->qd_idx = qd_idx;
NeilBrown67cc2b82009-03-31 14:39:38 +11002172 sh->ddf_layout = ddf_layout;
NeilBrown911d4ee2009-03-31 14:39:38 +11002173 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 /*
2175 * Finally, compute the new sector number
2176 */
2177 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2178 return new_sector;
2179}
2180
2181
NeilBrown784052e2009-03-31 15:19:07 +11002182static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183{
NeilBrownd1688a62011-10-11 16:49:52 +11002184 struct r5conf *conf = sh->raid_conf;
NeilBrownb875e532006-12-10 02:20:49 -08002185 int raid_disks = sh->disks;
2186 int data_disks = raid_disks - conf->max_degraded;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187 sector_t new_sector = sh->sector, check;
Andre Noll09c9e5f2009-06-18 08:45:55 +10002188 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2189 : conf->chunk_sectors;
NeilBrowne183eae2009-03-31 15:20:22 +11002190 int algorithm = previous ? conf->prev_algo
2191 : conf->algorithm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192 sector_t stripe;
2193 int chunk_offset;
NeilBrown35f2a592010-04-20 14:13:34 +10002194 sector_t chunk_number;
2195 int dummy1, dd_idx = i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196 sector_t r_sector;
NeilBrown911d4ee2009-03-31 14:39:38 +11002197 struct stripe_head sh2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198
NeilBrown16a53ec2006-06-26 00:27:38 -07002199
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2201 stripe = new_sector;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202
NeilBrown16a53ec2006-06-26 00:27:38 -07002203 if (i == sh->pd_idx)
2204 return 0;
2205 switch(conf->level) {
2206 case 4: break;
2207 case 5:
NeilBrowne183eae2009-03-31 15:20:22 +11002208 switch (algorithm) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 case ALGORITHM_LEFT_ASYMMETRIC:
2210 case ALGORITHM_RIGHT_ASYMMETRIC:
2211 if (i > sh->pd_idx)
2212 i--;
2213 break;
2214 case ALGORITHM_LEFT_SYMMETRIC:
2215 case ALGORITHM_RIGHT_SYMMETRIC:
2216 if (i < sh->pd_idx)
2217 i += raid_disks;
2218 i -= (sh->pd_idx + 1);
2219 break;
NeilBrown99c0fb52009-03-31 14:39:38 +11002220 case ALGORITHM_PARITY_0:
2221 i -= 1;
2222 break;
2223 case ALGORITHM_PARITY_N:
2224 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 default:
NeilBrown99c0fb52009-03-31 14:39:38 +11002226 BUG();
NeilBrown16a53ec2006-06-26 00:27:38 -07002227 }
2228 break;
2229 case 6:
NeilBrownd0dabf72009-03-31 14:39:38 +11002230 if (i == sh->qd_idx)
NeilBrown16a53ec2006-06-26 00:27:38 -07002231 return 0; /* It is the Q disk */
NeilBrowne183eae2009-03-31 15:20:22 +11002232 switch (algorithm) {
NeilBrown16a53ec2006-06-26 00:27:38 -07002233 case ALGORITHM_LEFT_ASYMMETRIC:
2234 case ALGORITHM_RIGHT_ASYMMETRIC:
NeilBrown99c0fb52009-03-31 14:39:38 +11002235 case ALGORITHM_ROTATING_ZERO_RESTART:
2236 case ALGORITHM_ROTATING_N_RESTART:
2237 if (sh->pd_idx == raid_disks-1)
2238 i--; /* Q D D D P */
NeilBrown16a53ec2006-06-26 00:27:38 -07002239 else if (i > sh->pd_idx)
2240 i -= 2; /* D D P Q D */
2241 break;
2242 case ALGORITHM_LEFT_SYMMETRIC:
2243 case ALGORITHM_RIGHT_SYMMETRIC:
2244 if (sh->pd_idx == raid_disks-1)
2245 i--; /* Q D D D P */
2246 else {
2247 /* D D P Q D */
2248 if (i < sh->pd_idx)
2249 i += raid_disks;
2250 i -= (sh->pd_idx + 2);
2251 }
2252 break;
NeilBrown99c0fb52009-03-31 14:39:38 +11002253 case ALGORITHM_PARITY_0:
2254 i -= 2;
2255 break;
2256 case ALGORITHM_PARITY_N:
2257 break;
2258 case ALGORITHM_ROTATING_N_CONTINUE:
NeilBrowne4424fe2009-10-16 16:27:34 +11002259 /* Like left_symmetric, but P is before Q */
NeilBrown99c0fb52009-03-31 14:39:38 +11002260 if (sh->pd_idx == 0)
2261 i--; /* P D D D Q */
NeilBrowne4424fe2009-10-16 16:27:34 +11002262 else {
2263 /* D D Q P D */
2264 if (i < sh->pd_idx)
2265 i += raid_disks;
2266 i -= (sh->pd_idx + 1);
2267 }
NeilBrown99c0fb52009-03-31 14:39:38 +11002268 break;
2269 case ALGORITHM_LEFT_ASYMMETRIC_6:
2270 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2271 if (i > sh->pd_idx)
2272 i--;
2273 break;
2274 case ALGORITHM_LEFT_SYMMETRIC_6:
2275 case ALGORITHM_RIGHT_SYMMETRIC_6:
2276 if (i < sh->pd_idx)
2277 i += data_disks + 1;
2278 i -= (sh->pd_idx + 1);
2279 break;
2280 case ALGORITHM_PARITY_0_6:
2281 i -= 1;
2282 break;
NeilBrown16a53ec2006-06-26 00:27:38 -07002283 default:
NeilBrown99c0fb52009-03-31 14:39:38 +11002284 BUG();
NeilBrown16a53ec2006-06-26 00:27:38 -07002285 }
2286 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 }
2288
2289 chunk_number = stripe * data_disks + i;
NeilBrown35f2a592010-04-20 14:13:34 +10002290 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291
NeilBrown112bf892009-03-31 14:39:38 +11002292 check = raid5_compute_sector(conf, r_sector,
NeilBrown784052e2009-03-31 15:19:07 +11002293 previous, &dummy1, &sh2);
NeilBrown911d4ee2009-03-31 14:39:38 +11002294 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2295 || sh2.qd_idx != sh->qd_idx) {
NeilBrown0c55e022010-05-03 14:09:02 +10002296 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2297 mdname(conf->mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 return 0;
2299 }
2300 return r_sector;
2301}
2302
2303
Dan Williams600aa102008-06-28 08:32:05 +10002304static void
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002305schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
Dan Williams600aa102008-06-28 08:32:05 +10002306 int rcw, int expand)
Dan Williamse33129d2007-01-02 13:52:30 -07002307{
2308 int i, pd_idx = sh->pd_idx, disks = sh->disks;
NeilBrownd1688a62011-10-11 16:49:52 +11002309 struct r5conf *conf = sh->raid_conf;
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002310 int level = conf->level;
NeilBrown16a53ec2006-06-26 00:27:38 -07002311
Dan Williamse33129d2007-01-02 13:52:30 -07002312 if (rcw) {
Dan Williamse33129d2007-01-02 13:52:30 -07002313
2314 for (i = disks; i--; ) {
2315 struct r5dev *dev = &sh->dev[i];
2316
2317 if (dev->towrite) {
2318 set_bit(R5_LOCKED, &dev->flags);
Dan Williamsd8ee0722008-06-28 08:32:06 +10002319 set_bit(R5_Wantdrain, &dev->flags);
Dan Williamse33129d2007-01-02 13:52:30 -07002320 if (!expand)
2321 clear_bit(R5_UPTODATE, &dev->flags);
Dan Williams600aa102008-06-28 08:32:05 +10002322 s->locked++;
Dan Williamse33129d2007-01-02 13:52:30 -07002323 }
2324 }
NeilBrownce7d3632013-03-04 12:37:14 +11002325 /* if we are not expanding this is a proper write request, and
2326 * there will be bios with new data to be drained into the
2327 * stripe cache
2328 */
2329 if (!expand) {
2330 if (!s->locked)
2331 /* False alarm, nothing to do */
2332 return;
2333 sh->reconstruct_state = reconstruct_state_drain_run;
2334 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2335 } else
2336 sh->reconstruct_state = reconstruct_state_run;
2337
2338 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2339
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002340 if (s->locked + conf->max_degraded == disks)
Dan Williams8b3e6cd2008-04-28 02:15:53 -07002341 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002342 atomic_inc(&conf->pending_full_writes);
Dan Williamse33129d2007-01-02 13:52:30 -07002343 } else {
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002344 BUG_ON(level == 6);
Dan Williamse33129d2007-01-02 13:52:30 -07002345 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2346 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2347
Dan Williamse33129d2007-01-02 13:52:30 -07002348 for (i = disks; i--; ) {
2349 struct r5dev *dev = &sh->dev[i];
2350 if (i == pd_idx)
2351 continue;
2352
Dan Williamse33129d2007-01-02 13:52:30 -07002353 if (dev->towrite &&
2354 (test_bit(R5_UPTODATE, &dev->flags) ||
Dan Williamsd8ee0722008-06-28 08:32:06 +10002355 test_bit(R5_Wantcompute, &dev->flags))) {
2356 set_bit(R5_Wantdrain, &dev->flags);
Dan Williamse33129d2007-01-02 13:52:30 -07002357 set_bit(R5_LOCKED, &dev->flags);
2358 clear_bit(R5_UPTODATE, &dev->flags);
Dan Williams600aa102008-06-28 08:32:05 +10002359 s->locked++;
Dan Williamse33129d2007-01-02 13:52:30 -07002360 }
2361 }
NeilBrownce7d3632013-03-04 12:37:14 +11002362 if (!s->locked)
2363 /* False alarm - nothing to do */
2364 return;
2365 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2366 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2367 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2368 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
Dan Williamse33129d2007-01-02 13:52:30 -07002369 }
2370
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002371 /* keep the parity disk(s) locked while asynchronous operations
Dan Williamse33129d2007-01-02 13:52:30 -07002372 * are in flight
2373 */
2374 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2375 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
Dan Williams600aa102008-06-28 08:32:05 +10002376 s->locked++;
Dan Williamse33129d2007-01-02 13:52:30 -07002377
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002378 if (level == 6) {
2379 int qd_idx = sh->qd_idx;
2380 struct r5dev *dev = &sh->dev[qd_idx];
2381
2382 set_bit(R5_LOCKED, &dev->flags);
2383 clear_bit(R5_UPTODATE, &dev->flags);
2384 s->locked++;
2385 }
2386
Dan Williams600aa102008-06-28 08:32:05 +10002387 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
Harvey Harrisone46b2722008-04-28 02:15:50 -07002388 __func__, (unsigned long long)sh->sector,
Dan Williams600aa102008-06-28 08:32:05 +10002389 s->locked, s->ops_request);
Dan Williamse33129d2007-01-02 13:52:30 -07002390}
NeilBrown16a53ec2006-06-26 00:27:38 -07002391
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392/*
2393 * Each stripe/dev can have one or more bion attached.
NeilBrown16a53ec2006-06-26 00:27:38 -07002394 * toread/towrite point to the first in a chain.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395 * The bi_next chain must be in order.
2396 */
2397static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2398{
2399 struct bio **bip;
NeilBrownd1688a62011-10-11 16:49:52 +11002400 struct r5conf *conf = sh->raid_conf;
NeilBrown72626682005-09-09 16:23:54 -07002401 int firstwrite=0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402
NeilBrowncbe47ec2011-07-26 11:20:35 +10002403 pr_debug("adding bi b#%llu to stripe s#%llu\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404 (unsigned long long)bi->bi_sector,
2405 (unsigned long long)sh->sector);
2406
Shaohua Lib17459c2012-07-19 16:01:31 +10002407 /*
2408 * If several bio share a stripe. The bio bi_phys_segments acts as a
2409 * reference count to avoid race. The reference count should already be
2410 * increased before this function is called (for example, in
2411 * make_request()), so other bio sharing this stripe will not free the
2412 * stripe. If a stripe is owned by one stripe, the stripe lock will
2413 * protect it.
2414 */
2415 spin_lock_irq(&sh->stripe_lock);
NeilBrown72626682005-09-09 16:23:54 -07002416 if (forwrite) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 bip = &sh->dev[dd_idx].towrite;
Shaohua Li7eaf7e82012-07-19 16:01:31 +10002418 if (*bip == NULL)
NeilBrown72626682005-09-09 16:23:54 -07002419 firstwrite = 1;
2420 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421 bip = &sh->dev[dd_idx].toread;
2422 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
Kent Overstreetf73a1c72012-09-25 15:05:12 -07002423 if (bio_end_sector(*bip) > bi->bi_sector)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424 goto overlap;
2425 bip = & (*bip)->bi_next;
2426 }
Kent Overstreetf73a1c72012-09-25 15:05:12 -07002427 if (*bip && (*bip)->bi_sector < bio_end_sector(bi))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 goto overlap;
2429
Eric Sesterhenn78bafeb2006-04-02 13:31:42 +02002430 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431 if (*bip)
2432 bi->bi_next = *bip;
2433 *bip = bi;
Shaohua Lie7836bd62012-07-19 16:01:31 +10002434 raid5_inc_bi_active_stripes(bi);
NeilBrown72626682005-09-09 16:23:54 -07002435
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436 if (forwrite) {
2437 /* check if page is covered */
2438 sector_t sector = sh->dev[dd_idx].sector;
2439 for (bi=sh->dev[dd_idx].towrite;
2440 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2441 bi && bi->bi_sector <= sector;
2442 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
Kent Overstreetf73a1c72012-09-25 15:05:12 -07002443 if (bio_end_sector(bi) >= sector)
2444 sector = bio_end_sector(bi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445 }
2446 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2447 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2448 }
NeilBrowncbe47ec2011-07-26 11:20:35 +10002449
2450 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2451 (unsigned long long)(*bip)->bi_sector,
2452 (unsigned long long)sh->sector, dd_idx);
NeilBrownb97390a2012-10-11 13:50:12 +11002453 spin_unlock_irq(&sh->stripe_lock);
NeilBrowncbe47ec2011-07-26 11:20:35 +10002454
2455 if (conf->mddev->bitmap && firstwrite) {
2456 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2457 STRIPE_SECTORS, 0);
2458 sh->bm_seq = conf->seq_flush+1;
2459 set_bit(STRIPE_BIT_DELAY, &sh->state);
2460 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461 return 1;
2462
2463 overlap:
2464 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
Shaohua Lib17459c2012-07-19 16:01:31 +10002465 spin_unlock_irq(&sh->stripe_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466 return 0;
2467}
2468
NeilBrownd1688a62011-10-11 16:49:52 +11002469static void end_reshape(struct r5conf *conf);
NeilBrown29269552006-03-27 01:18:10 -08002470
NeilBrownd1688a62011-10-11 16:49:52 +11002471static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
NeilBrown911d4ee2009-03-31 14:39:38 +11002472 struct stripe_head *sh)
NeilBrownccfcc3c2006-03-27 01:18:09 -08002473{
NeilBrown784052e2009-03-31 15:19:07 +11002474 int sectors_per_chunk =
Andre Noll09c9e5f2009-06-18 08:45:55 +10002475 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
NeilBrown911d4ee2009-03-31 14:39:38 +11002476 int dd_idx;
Coywolf Qi Hunt2d2063c2006-10-03 01:15:50 -07002477 int chunk_offset = sector_div(stripe, sectors_per_chunk);
NeilBrown112bf892009-03-31 14:39:38 +11002478 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
Coywolf Qi Hunt2d2063c2006-10-03 01:15:50 -07002479
NeilBrown112bf892009-03-31 14:39:38 +11002480 raid5_compute_sector(conf,
2481 stripe * (disks - conf->max_degraded)
NeilBrownb875e532006-12-10 02:20:49 -08002482 *sectors_per_chunk + chunk_offset,
NeilBrown112bf892009-03-31 14:39:38 +11002483 previous,
NeilBrown911d4ee2009-03-31 14:39:38 +11002484 &dd_idx, sh);
NeilBrownccfcc3c2006-03-27 01:18:09 -08002485}
2486
Dan Williamsa4456852007-07-09 11:56:43 -07002487static void
NeilBrownd1688a62011-10-11 16:49:52 +11002488handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
Dan Williamsa4456852007-07-09 11:56:43 -07002489 struct stripe_head_state *s, int disks,
2490 struct bio **return_bi)
2491{
2492 int i;
2493 for (i = disks; i--; ) {
2494 struct bio *bi;
2495 int bitmap_end = 0;
2496
2497 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
NeilBrown3cb03002011-10-11 16:45:26 +11002498 struct md_rdev *rdev;
Dan Williamsa4456852007-07-09 11:56:43 -07002499 rcu_read_lock();
2500 rdev = rcu_dereference(conf->disks[i].rdev);
2501 if (rdev && test_bit(In_sync, &rdev->flags))
NeilBrown7f0da592011-07-28 11:39:22 +10002502 atomic_inc(&rdev->nr_pending);
2503 else
2504 rdev = NULL;
Dan Williamsa4456852007-07-09 11:56:43 -07002505 rcu_read_unlock();
NeilBrown7f0da592011-07-28 11:39:22 +10002506 if (rdev) {
2507 if (!rdev_set_badblocks(
2508 rdev,
2509 sh->sector,
2510 STRIPE_SECTORS, 0))
2511 md_error(conf->mddev, rdev);
2512 rdev_dec_pending(rdev, conf->mddev);
2513 }
Dan Williamsa4456852007-07-09 11:56:43 -07002514 }
Shaohua Lib17459c2012-07-19 16:01:31 +10002515 spin_lock_irq(&sh->stripe_lock);
Dan Williamsa4456852007-07-09 11:56:43 -07002516 /* fail all writes first */
2517 bi = sh->dev[i].towrite;
2518 sh->dev[i].towrite = NULL;
Shaohua Lib17459c2012-07-19 16:01:31 +10002519 spin_unlock_irq(&sh->stripe_lock);
NeilBrown1ed850f2012-10-11 13:50:13 +11002520 if (bi)
Dan Williamsa4456852007-07-09 11:56:43 -07002521 bitmap_end = 1;
Dan Williamsa4456852007-07-09 11:56:43 -07002522
2523 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2524 wake_up(&conf->wait_for_overlap);
2525
2526 while (bi && bi->bi_sector <
2527 sh->dev[i].sector + STRIPE_SECTORS) {
2528 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2529 clear_bit(BIO_UPTODATE, &bi->bi_flags);
Shaohua Lie7836bd62012-07-19 16:01:31 +10002530 if (!raid5_dec_bi_active_stripes(bi)) {
Dan Williamsa4456852007-07-09 11:56:43 -07002531 md_write_end(conf->mddev);
2532 bi->bi_next = *return_bi;
2533 *return_bi = bi;
2534 }
2535 bi = nextbi;
2536 }
Shaohua Li7eaf7e82012-07-19 16:01:31 +10002537 if (bitmap_end)
2538 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2539 STRIPE_SECTORS, 0, 0);
2540 bitmap_end = 0;
Dan Williamsa4456852007-07-09 11:56:43 -07002541 /* and fail all 'written' */
2542 bi = sh->dev[i].written;
2543 sh->dev[i].written = NULL;
2544 if (bi) bitmap_end = 1;
2545 while (bi && bi->bi_sector <
2546 sh->dev[i].sector + STRIPE_SECTORS) {
2547 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2548 clear_bit(BIO_UPTODATE, &bi->bi_flags);
Shaohua Lie7836bd62012-07-19 16:01:31 +10002549 if (!raid5_dec_bi_active_stripes(bi)) {
Dan Williamsa4456852007-07-09 11:56:43 -07002550 md_write_end(conf->mddev);
2551 bi->bi_next = *return_bi;
2552 *return_bi = bi;
2553 }
2554 bi = bi2;
2555 }
2556
Dan Williamsb5e98d62007-01-02 13:52:31 -07002557 /* fail any reads if this device is non-operational and
2558 * the data has not reached the cache yet.
2559 */
2560 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2561 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2562 test_bit(R5_ReadError, &sh->dev[i].flags))) {
NeilBrown143c4d02012-10-11 13:50:12 +11002563 spin_lock_irq(&sh->stripe_lock);
Dan Williamsa4456852007-07-09 11:56:43 -07002564 bi = sh->dev[i].toread;
2565 sh->dev[i].toread = NULL;
NeilBrown143c4d02012-10-11 13:50:12 +11002566 spin_unlock_irq(&sh->stripe_lock);
Dan Williamsa4456852007-07-09 11:56:43 -07002567 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2568 wake_up(&conf->wait_for_overlap);
Dan Williamsa4456852007-07-09 11:56:43 -07002569 while (bi && bi->bi_sector <
2570 sh->dev[i].sector + STRIPE_SECTORS) {
2571 struct bio *nextbi =
2572 r5_next_bio(bi, sh->dev[i].sector);
2573 clear_bit(BIO_UPTODATE, &bi->bi_flags);
Shaohua Lie7836bd62012-07-19 16:01:31 +10002574 if (!raid5_dec_bi_active_stripes(bi)) {
Dan Williamsa4456852007-07-09 11:56:43 -07002575 bi->bi_next = *return_bi;
2576 *return_bi = bi;
2577 }
2578 bi = nextbi;
2579 }
2580 }
Dan Williamsa4456852007-07-09 11:56:43 -07002581 if (bitmap_end)
2582 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2583 STRIPE_SECTORS, 0, 0);
NeilBrown8cfa7b02011-07-27 11:00:36 +10002584 /* If we were in the middle of a write the parity block might
2585 * still be locked - so just clear all R5_LOCKED flags
2586 */
2587 clear_bit(R5_LOCKED, &sh->dev[i].flags);
Dan Williamsa4456852007-07-09 11:56:43 -07002588 }
2589
Dan Williams8b3e6cd2008-04-28 02:15:53 -07002590 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2591 if (atomic_dec_and_test(&conf->pending_full_writes))
2592 md_wakeup_thread(conf->mddev->thread);
Dan Williamsa4456852007-07-09 11:56:43 -07002593}
2594
NeilBrown7f0da592011-07-28 11:39:22 +10002595static void
NeilBrownd1688a62011-10-11 16:49:52 +11002596handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
NeilBrown7f0da592011-07-28 11:39:22 +10002597 struct stripe_head_state *s)
2598{
2599 int abort = 0;
2600 int i;
2601
NeilBrown7f0da592011-07-28 11:39:22 +10002602 clear_bit(STRIPE_SYNCING, &sh->state);
NeilBrownf8dfcff2013-03-12 12:18:06 +11002603 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2604 wake_up(&conf->wait_for_overlap);
NeilBrown7f0da592011-07-28 11:39:22 +10002605 s->syncing = 0;
NeilBrown9a3e1102011-12-23 10:17:53 +11002606 s->replacing = 0;
NeilBrown7f0da592011-07-28 11:39:22 +10002607 /* There is nothing more to do for sync/check/repair.
NeilBrown18b98372012-04-01 23:48:38 +10002608 * Don't even need to abort as that is handled elsewhere
2609 * if needed, and not always wanted e.g. if there is a known
2610 * bad block here.
NeilBrown9a3e1102011-12-23 10:17:53 +11002611 * For recover/replace we need to record a bad block on all
NeilBrown7f0da592011-07-28 11:39:22 +10002612 * non-sync devices, or abort the recovery
2613 */
NeilBrown18b98372012-04-01 23:48:38 +10002614 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2615 /* During recovery devices cannot be removed, so
2616 * locking and refcounting of rdevs is not needed
2617 */
2618 for (i = 0; i < conf->raid_disks; i++) {
2619 struct md_rdev *rdev = conf->disks[i].rdev;
2620 if (rdev
2621 && !test_bit(Faulty, &rdev->flags)
2622 && !test_bit(In_sync, &rdev->flags)
2623 && !rdev_set_badblocks(rdev, sh->sector,
2624 STRIPE_SECTORS, 0))
2625 abort = 1;
2626 rdev = conf->disks[i].replacement;
2627 if (rdev
2628 && !test_bit(Faulty, &rdev->flags)
2629 && !test_bit(In_sync, &rdev->flags)
2630 && !rdev_set_badblocks(rdev, sh->sector,
2631 STRIPE_SECTORS, 0))
2632 abort = 1;
2633 }
2634 if (abort)
2635 conf->recovery_disabled =
2636 conf->mddev->recovery_disabled;
NeilBrown7f0da592011-07-28 11:39:22 +10002637 }
NeilBrown18b98372012-04-01 23:48:38 +10002638 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
NeilBrown7f0da592011-07-28 11:39:22 +10002639}
2640
NeilBrown9a3e1102011-12-23 10:17:53 +11002641static int want_replace(struct stripe_head *sh, int disk_idx)
2642{
2643 struct md_rdev *rdev;
2644 int rv = 0;
2645 /* Doing recovery so rcu locking not required */
2646 rdev = sh->raid_conf->disks[disk_idx].replacement;
2647 if (rdev
2648 && !test_bit(Faulty, &rdev->flags)
2649 && !test_bit(In_sync, &rdev->flags)
2650 && (rdev->recovery_offset <= sh->sector
2651 || rdev->mddev->recovery_cp <= sh->sector))
2652 rv = 1;
2653
2654 return rv;
2655}
2656
NeilBrown93b3dbc2011-07-27 11:00:36 +10002657/* fetch_block - checks the given member device to see if its data needs
Dan Williams1fe797e2008-06-28 09:16:30 +10002658 * to be read or computed to satisfy a request.
2659 *
2660 * Returns 1 when no more member devices need to be checked, otherwise returns
NeilBrown93b3dbc2011-07-27 11:00:36 +10002661 * 0 to tell the loop in handle_stripe_fill to continue
Dan Williamsf38e1212007-01-02 13:52:30 -07002662 */
NeilBrown93b3dbc2011-07-27 11:00:36 +10002663static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2664 int disk_idx, int disks)
Dan Williamsf38e1212007-01-02 13:52:30 -07002665{
2666 struct r5dev *dev = &sh->dev[disk_idx];
NeilBrown93b3dbc2011-07-27 11:00:36 +10002667 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2668 &sh->dev[s->failed_num[1]] };
Dan Williamsf38e1212007-01-02 13:52:30 -07002669
Dan Williamsf38e1212007-01-02 13:52:30 -07002670 /* is the data in this block needed, and can we get it? */
2671 if (!test_bit(R5_LOCKED, &dev->flags) &&
Dan Williams1fe797e2008-06-28 09:16:30 +10002672 !test_bit(R5_UPTODATE, &dev->flags) &&
2673 (dev->toread ||
2674 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2675 s->syncing || s->expanding ||
NeilBrown9a3e1102011-12-23 10:17:53 +11002676 (s->replacing && want_replace(sh, disk_idx)) ||
NeilBrown5d35e092011-07-27 11:00:36 +10002677 (s->failed >= 1 && fdev[0]->toread) ||
2678 (s->failed >= 2 && fdev[1]->toread) ||
NeilBrown93b3dbc2011-07-27 11:00:36 +10002679 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2680 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2681 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
Yuri Tikhonov5599bec2009-08-29 19:13:12 -07002682 /* we would like to get this block, possibly by computing it,
2683 * otherwise read it if the backing disk is insync
2684 */
2685 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2686 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2687 if ((s->uptodate == disks - 1) &&
NeilBrownf2b3b442011-07-26 11:35:19 +10002688 (s->failed && (disk_idx == s->failed_num[0] ||
2689 disk_idx == s->failed_num[1]))) {
Yuri Tikhonov5599bec2009-08-29 19:13:12 -07002690 /* have disk failed, and we're requested to fetch it;
2691 * do compute it
2692 */
2693 pr_debug("Computing stripe %llu block %d\n",
2694 (unsigned long long)sh->sector, disk_idx);
2695 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2696 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2697 set_bit(R5_Wantcompute, &dev->flags);
2698 sh->ops.target = disk_idx;
2699 sh->ops.target2 = -1; /* no 2nd target */
2700 s->req_compute = 1;
NeilBrown93b3dbc2011-07-27 11:00:36 +10002701 /* Careful: from this point on 'uptodate' is in the eye
2702 * of raid_run_ops which services 'compute' operations
2703 * before writes. R5_Wantcompute flags a block that will
2704 * be R5_UPTODATE by the time it is needed for a
2705 * subsequent operation.
2706 */
Yuri Tikhonov5599bec2009-08-29 19:13:12 -07002707 s->uptodate++;
2708 return 1;
2709 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2710 /* Computing 2-failure is *very* expensive; only
2711 * do it if failed >= 2
2712 */
2713 int other;
2714 for (other = disks; other--; ) {
2715 if (other == disk_idx)
2716 continue;
2717 if (!test_bit(R5_UPTODATE,
2718 &sh->dev[other].flags))
2719 break;
2720 }
2721 BUG_ON(other < 0);
2722 pr_debug("Computing stripe %llu blocks %d,%d\n",
2723 (unsigned long long)sh->sector,
2724 disk_idx, other);
2725 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2726 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2727 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2728 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2729 sh->ops.target = disk_idx;
2730 sh->ops.target2 = other;
2731 s->uptodate += 2;
2732 s->req_compute = 1;
2733 return 1;
2734 } else if (test_bit(R5_Insync, &dev->flags)) {
2735 set_bit(R5_LOCKED, &dev->flags);
2736 set_bit(R5_Wantread, &dev->flags);
2737 s->locked++;
2738 pr_debug("Reading block %d (sync=%d)\n",
2739 disk_idx, s->syncing);
2740 }
2741 }
2742
2743 return 0;
2744}
2745
2746/**
NeilBrown93b3dbc2011-07-27 11:00:36 +10002747 * handle_stripe_fill - read or compute data to satisfy pending requests.
Yuri Tikhonov5599bec2009-08-29 19:13:12 -07002748 */
NeilBrown93b3dbc2011-07-27 11:00:36 +10002749static void handle_stripe_fill(struct stripe_head *sh,
2750 struct stripe_head_state *s,
2751 int disks)
Dan Williamsa4456852007-07-09 11:56:43 -07002752{
2753 int i;
Yuri Tikhonov5599bec2009-08-29 19:13:12 -07002754
2755 /* look for blocks to read/compute, skip this if a compute
2756 * is already in flight, or if the stripe contents are in the
2757 * midst of changing due to a write
2758 */
2759 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2760 !sh->reconstruct_state)
2761 for (i = disks; i--; )
NeilBrown93b3dbc2011-07-27 11:00:36 +10002762 if (fetch_block(sh, s, i, disks))
Yuri Tikhonov5599bec2009-08-29 19:13:12 -07002763 break;
Dan Williamsa4456852007-07-09 11:56:43 -07002764 set_bit(STRIPE_HANDLE, &sh->state);
2765}
2766
2767
Dan Williams1fe797e2008-06-28 09:16:30 +10002768/* handle_stripe_clean_event
Dan Williamsa4456852007-07-09 11:56:43 -07002769 * any written block on an uptodate or failed drive can be returned.
2770 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2771 * never LOCKED, so we don't need to test 'failed' directly.
2772 */
NeilBrownd1688a62011-10-11 16:49:52 +11002773static void handle_stripe_clean_event(struct r5conf *conf,
Dan Williamsa4456852007-07-09 11:56:43 -07002774 struct stripe_head *sh, int disks, struct bio **return_bi)
2775{
2776 int i;
2777 struct r5dev *dev;
NeilBrownf8dfcff2013-03-12 12:18:06 +11002778 int discard_pending = 0;
Dan Williamsa4456852007-07-09 11:56:43 -07002779
2780 for (i = disks; i--; )
2781 if (sh->dev[i].written) {
2782 dev = &sh->dev[i];
2783 if (!test_bit(R5_LOCKED, &dev->flags) &&
Shaohua Li9e4447682012-10-11 13:49:49 +11002784 (test_bit(R5_UPTODATE, &dev->flags) ||
NeilBrownca64cae2012-11-21 16:33:40 +11002785 test_bit(R5_Discard, &dev->flags))) {
Dan Williamsa4456852007-07-09 11:56:43 -07002786 /* We can return any write requests */
2787 struct bio *wbi, *wbi2;
Dan Williams45b42332007-07-09 11:56:43 -07002788 pr_debug("Return write for disc %d\n", i);
NeilBrownca64cae2012-11-21 16:33:40 +11002789 if (test_and_clear_bit(R5_Discard, &dev->flags))
2790 clear_bit(R5_UPTODATE, &dev->flags);
Dan Williamsa4456852007-07-09 11:56:43 -07002791 wbi = dev->written;
2792 dev->written = NULL;
2793 while (wbi && wbi->bi_sector <
2794 dev->sector + STRIPE_SECTORS) {
2795 wbi2 = r5_next_bio(wbi, dev->sector);
Shaohua Lie7836bd62012-07-19 16:01:31 +10002796 if (!raid5_dec_bi_active_stripes(wbi)) {
Dan Williamsa4456852007-07-09 11:56:43 -07002797 md_write_end(conf->mddev);
2798 wbi->bi_next = *return_bi;
2799 *return_bi = wbi;
2800 }
2801 wbi = wbi2;
2802 }
Shaohua Li7eaf7e82012-07-19 16:01:31 +10002803 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2804 STRIPE_SECTORS,
Dan Williamsa4456852007-07-09 11:56:43 -07002805 !test_bit(STRIPE_DEGRADED, &sh->state),
Shaohua Li7eaf7e82012-07-19 16:01:31 +10002806 0);
NeilBrownf8dfcff2013-03-12 12:18:06 +11002807 } else if (test_bit(R5_Discard, &dev->flags))
2808 discard_pending = 1;
2809 }
2810 if (!discard_pending &&
2811 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
2812 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2813 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2814 if (sh->qd_idx >= 0) {
2815 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2816 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
2817 }
2818 /* now that discard is done we can proceed with any sync */
2819 clear_bit(STRIPE_DISCARD, &sh->state);
Shaohua Li01e608d2013-10-19 14:51:42 +08002820 /*
2821 * SCSI discard will change some bio fields and the stripe has
2822 * no updated data, so remove it from hash list and the stripe
2823 * will be reinitialized
2824 */
2825 spin_lock_irq(&conf->device_lock);
2826 remove_hash(sh);
2827 spin_unlock_irq(&conf->device_lock);
NeilBrownf8dfcff2013-03-12 12:18:06 +11002828 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2829 set_bit(STRIPE_HANDLE, &sh->state);
2830
2831 }
Dan Williams8b3e6cd2008-04-28 02:15:53 -07002832
2833 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2834 if (atomic_dec_and_test(&conf->pending_full_writes))
2835 md_wakeup_thread(conf->mddev->thread);
Dan Williamsa4456852007-07-09 11:56:43 -07002836}
2837
NeilBrownd1688a62011-10-11 16:49:52 +11002838static void handle_stripe_dirtying(struct r5conf *conf,
NeilBrownc8ac1802011-07-27 11:00:36 +10002839 struct stripe_head *sh,
2840 struct stripe_head_state *s,
2841 int disks)
Dan Williamsa4456852007-07-09 11:56:43 -07002842{
2843 int rmw = 0, rcw = 0, i;
Alexander Lyakasa7854482012-10-11 13:50:12 +11002844 sector_t recovery_cp = conf->mddev->recovery_cp;
2845
2846 /* RAID6 requires 'rcw' in current implementation.
2847 * Otherwise, check whether resync is now happening or should start.
2848 * If yes, then the array is dirty (after unclean shutdown or
2849 * initial creation), so parity in some stripes might be inconsistent.
2850 * In this case, we need to always do reconstruct-write, to ensure
2851 * that in case of drive failure or read-error correction, we
2852 * generate correct data from the parity.
2853 */
2854 if (conf->max_degraded == 2 ||
2855 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2856 /* Calculate the real rcw later - for now make it
NeilBrownc8ac1802011-07-27 11:00:36 +10002857 * look like rcw is cheaper
2858 */
2859 rcw = 1; rmw = 2;
Alexander Lyakasa7854482012-10-11 13:50:12 +11002860 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2861 conf->max_degraded, (unsigned long long)recovery_cp,
2862 (unsigned long long)sh->sector);
NeilBrownc8ac1802011-07-27 11:00:36 +10002863 } else for (i = disks; i--; ) {
Dan Williamsa4456852007-07-09 11:56:43 -07002864 /* would I have to read this buffer for read_modify_write */
2865 struct r5dev *dev = &sh->dev[i];
2866 if ((dev->towrite || i == sh->pd_idx) &&
2867 !test_bit(R5_LOCKED, &dev->flags) &&
Dan Williamsf38e1212007-01-02 13:52:30 -07002868 !(test_bit(R5_UPTODATE, &dev->flags) ||
2869 test_bit(R5_Wantcompute, &dev->flags))) {
Dan Williamsa4456852007-07-09 11:56:43 -07002870 if (test_bit(R5_Insync, &dev->flags))
2871 rmw++;
2872 else
2873 rmw += 2*disks; /* cannot read it */
2874 }
2875 /* Would I have to read this buffer for reconstruct_write */
2876 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2877 !test_bit(R5_LOCKED, &dev->flags) &&
Dan Williamsf38e1212007-01-02 13:52:30 -07002878 !(test_bit(R5_UPTODATE, &dev->flags) ||
2879 test_bit(R5_Wantcompute, &dev->flags))) {
2880 if (test_bit(R5_Insync, &dev->flags)) rcw++;
Dan Williamsa4456852007-07-09 11:56:43 -07002881 else
2882 rcw += 2*disks;
2883 }
2884 }
Dan Williams45b42332007-07-09 11:56:43 -07002885 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
Dan Williamsa4456852007-07-09 11:56:43 -07002886 (unsigned long long)sh->sector, rmw, rcw);
2887 set_bit(STRIPE_HANDLE, &sh->state);
NeilBrowna9add5d2012-10-31 11:59:09 +11002888 if (rmw < rcw && rmw > 0) {
Dan Williamsa4456852007-07-09 11:56:43 -07002889 /* prefer read-modify-write, but need to get some data */
Jonathan Brassowe3620a32013-03-07 16:22:01 -06002890 if (conf->mddev->queue)
2891 blk_add_trace_msg(conf->mddev->queue,
2892 "raid5 rmw %llu %d",
2893 (unsigned long long)sh->sector, rmw);
Dan Williamsa4456852007-07-09 11:56:43 -07002894 for (i = disks; i--; ) {
2895 struct r5dev *dev = &sh->dev[i];
2896 if ((dev->towrite || i == sh->pd_idx) &&
2897 !test_bit(R5_LOCKED, &dev->flags) &&
Dan Williamsf38e1212007-01-02 13:52:30 -07002898 !(test_bit(R5_UPTODATE, &dev->flags) ||
2899 test_bit(R5_Wantcompute, &dev->flags)) &&
Dan Williamsa4456852007-07-09 11:56:43 -07002900 test_bit(R5_Insync, &dev->flags)) {
2901 if (
2902 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
Dan Williams45b42332007-07-09 11:56:43 -07002903 pr_debug("Read_old block "
NeilBrowna9add5d2012-10-31 11:59:09 +11002904 "%d for r-m-w\n", i);
Dan Williamsa4456852007-07-09 11:56:43 -07002905 set_bit(R5_LOCKED, &dev->flags);
2906 set_bit(R5_Wantread, &dev->flags);
2907 s->locked++;
2908 } else {
2909 set_bit(STRIPE_DELAYED, &sh->state);
2910 set_bit(STRIPE_HANDLE, &sh->state);
2911 }
2912 }
2913 }
NeilBrowna9add5d2012-10-31 11:59:09 +11002914 }
NeilBrownc8ac1802011-07-27 11:00:36 +10002915 if (rcw <= rmw && rcw > 0) {
Dan Williamsa4456852007-07-09 11:56:43 -07002916 /* want reconstruct write, but need to get some data */
NeilBrowna9add5d2012-10-31 11:59:09 +11002917 int qread =0;
NeilBrownc8ac1802011-07-27 11:00:36 +10002918 rcw = 0;
Dan Williamsa4456852007-07-09 11:56:43 -07002919 for (i = disks; i--; ) {
2920 struct r5dev *dev = &sh->dev[i];
2921 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
NeilBrownc8ac1802011-07-27 11:00:36 +10002922 i != sh->pd_idx && i != sh->qd_idx &&
Dan Williamsa4456852007-07-09 11:56:43 -07002923 !test_bit(R5_LOCKED, &dev->flags) &&
Dan Williamsf38e1212007-01-02 13:52:30 -07002924 !(test_bit(R5_UPTODATE, &dev->flags) ||
NeilBrownc8ac1802011-07-27 11:00:36 +10002925 test_bit(R5_Wantcompute, &dev->flags))) {
2926 rcw++;
2927 if (!test_bit(R5_Insync, &dev->flags))
2928 continue; /* it's a failed drive */
Dan Williamsa4456852007-07-09 11:56:43 -07002929 if (
2930 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
Dan Williams45b42332007-07-09 11:56:43 -07002931 pr_debug("Read_old block "
Dan Williamsa4456852007-07-09 11:56:43 -07002932 "%d for Reconstruct\n", i);
2933 set_bit(R5_LOCKED, &dev->flags);
2934 set_bit(R5_Wantread, &dev->flags);
2935 s->locked++;
NeilBrowna9add5d2012-10-31 11:59:09 +11002936 qread++;
Dan Williamsa4456852007-07-09 11:56:43 -07002937 } else {
2938 set_bit(STRIPE_DELAYED, &sh->state);
2939 set_bit(STRIPE_HANDLE, &sh->state);
2940 }
2941 }
2942 }
Jonathan Brassowe3620a32013-03-07 16:22:01 -06002943 if (rcw && conf->mddev->queue)
NeilBrowna9add5d2012-10-31 11:59:09 +11002944 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
2945 (unsigned long long)sh->sector,
2946 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
NeilBrownc8ac1802011-07-27 11:00:36 +10002947 }
Dan Williamsa4456852007-07-09 11:56:43 -07002948 /* now if nothing is locked, and if we have enough data,
2949 * we can start a write request
2950 */
Dan Williamsf38e1212007-01-02 13:52:30 -07002951 /* since handle_stripe can be called at any time we need to handle the
2952 * case where a compute block operation has been submitted and then a
Dan Williamsac6b53b2009-07-14 13:40:19 -07002953 * subsequent call wants to start a write request. raid_run_ops only
2954 * handles the case where compute block and reconstruct are requested
Dan Williamsf38e1212007-01-02 13:52:30 -07002955 * simultaneously. If this is not the case then new writes need to be
2956 * held off until the compute completes.
2957 */
Dan Williams976ea8d2008-06-28 08:32:03 +10002958 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2959 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2960 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
Yuri Tikhonovc0f7bdd2009-08-29 19:13:12 -07002961 schedule_reconstruction(sh, s, rcw == 0, 0);
Dan Williamsa4456852007-07-09 11:56:43 -07002962}
2963
NeilBrownd1688a62011-10-11 16:49:52 +11002964static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
Dan Williamsa4456852007-07-09 11:56:43 -07002965 struct stripe_head_state *s, int disks)
2966{
Dan Williamsecc65c92008-06-28 08:31:57 +10002967 struct r5dev *dev = NULL;
Dan Williamse89f8962007-01-02 13:52:31 -07002968
Dan Williamsbd2ab672008-04-10 21:29:27 -07002969 set_bit(STRIPE_HANDLE, &sh->state);
2970
Dan Williamsecc65c92008-06-28 08:31:57 +10002971 switch (sh->check_state) {
2972 case check_state_idle:
2973 /* start a new check operation if there are no failures */
Dan Williamsbd2ab672008-04-10 21:29:27 -07002974 if (s->failed == 0) {
Dan Williamsbd2ab672008-04-10 21:29:27 -07002975 BUG_ON(s->uptodate != disks);
Dan Williamsecc65c92008-06-28 08:31:57 +10002976 sh->check_state = check_state_run;
2977 set_bit(STRIPE_OP_CHECK, &s->ops_request);
Dan Williamsbd2ab672008-04-10 21:29:27 -07002978 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
Dan Williamsbd2ab672008-04-10 21:29:27 -07002979 s->uptodate--;
Dan Williamsecc65c92008-06-28 08:31:57 +10002980 break;
Dan Williamsbd2ab672008-04-10 21:29:27 -07002981 }
NeilBrownf2b3b442011-07-26 11:35:19 +10002982 dev = &sh->dev[s->failed_num[0]];
Dan Williamsecc65c92008-06-28 08:31:57 +10002983 /* fall through */
2984 case check_state_compute_result:
2985 sh->check_state = check_state_idle;
2986 if (!dev)
2987 dev = &sh->dev[sh->pd_idx];
2988
2989 /* check that a write has not made the stripe insync */
2990 if (test_bit(STRIPE_INSYNC, &sh->state))
2991 break;
2992
2993 /* either failed parity check, or recovery is happening */
Dan Williamsa4456852007-07-09 11:56:43 -07002994 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2995 BUG_ON(s->uptodate != disks);
2996
2997 set_bit(R5_LOCKED, &dev->flags);
Dan Williamsecc65c92008-06-28 08:31:57 +10002998 s->locked++;
Dan Williamsa4456852007-07-09 11:56:43 -07002999 set_bit(R5_Wantwrite, &dev->flags);
Dan Williams830ea012007-01-02 13:52:31 -07003000
Dan Williamsa4456852007-07-09 11:56:43 -07003001 clear_bit(STRIPE_DEGRADED, &sh->state);
Dan Williamsa4456852007-07-09 11:56:43 -07003002 set_bit(STRIPE_INSYNC, &sh->state);
Dan Williamsecc65c92008-06-28 08:31:57 +10003003 break;
3004 case check_state_run:
3005 break; /* we will be called again upon completion */
3006 case check_state_check_result:
3007 sh->check_state = check_state_idle;
3008
3009 /* if a failure occurred during the check operation, leave
3010 * STRIPE_INSYNC not set and let the stripe be handled again
3011 */
3012 if (s->failed)
3013 break;
3014
3015 /* handle a successful check operation, if parity is correct
3016 * we are done. Otherwise update the mismatch count and repair
3017 * parity if !MD_RECOVERY_CHECK
3018 */
Dan Williamsad283ea2009-08-29 19:09:26 -07003019 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
Dan Williamsecc65c92008-06-28 08:31:57 +10003020 /* parity is correct (on disc,
3021 * not in buffer any more)
3022 */
3023 set_bit(STRIPE_INSYNC, &sh->state);
3024 else {
Jianpeng Ma7f7583d2012-10-11 14:17:59 +11003025 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
Dan Williamsecc65c92008-06-28 08:31:57 +10003026 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3027 /* don't try to repair!! */
3028 set_bit(STRIPE_INSYNC, &sh->state);
3029 else {
3030 sh->check_state = check_state_compute_run;
Dan Williams976ea8d2008-06-28 08:32:03 +10003031 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
Dan Williamsecc65c92008-06-28 08:31:57 +10003032 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3033 set_bit(R5_Wantcompute,
3034 &sh->dev[sh->pd_idx].flags);
3035 sh->ops.target = sh->pd_idx;
Dan Williamsac6b53b2009-07-14 13:40:19 -07003036 sh->ops.target2 = -1;
Dan Williamsecc65c92008-06-28 08:31:57 +10003037 s->uptodate++;
3038 }
3039 }
3040 break;
3041 case check_state_compute_run:
3042 break;
3043 default:
3044 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3045 __func__, sh->check_state,
3046 (unsigned long long) sh->sector);
3047 BUG();
Dan Williamsa4456852007-07-09 11:56:43 -07003048 }
3049}
3050
3051
NeilBrownd1688a62011-10-11 16:49:52 +11003052static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
Dan Williams36d1c642009-07-14 11:48:22 -07003053 struct stripe_head_state *s,
NeilBrownf2b3b442011-07-26 11:35:19 +10003054 int disks)
Dan Williamsa4456852007-07-09 11:56:43 -07003055{
Dan Williamsa4456852007-07-09 11:56:43 -07003056 int pd_idx = sh->pd_idx;
NeilBrown34e04e82009-03-31 15:10:16 +11003057 int qd_idx = sh->qd_idx;
Dan Williamsd82dfee2009-07-14 13:40:57 -07003058 struct r5dev *dev;
Dan Williamsa4456852007-07-09 11:56:43 -07003059
3060 set_bit(STRIPE_HANDLE, &sh->state);
3061
3062 BUG_ON(s->failed > 2);
Dan Williamsd82dfee2009-07-14 13:40:57 -07003063
Dan Williamsa4456852007-07-09 11:56:43 -07003064 /* Want to check and possibly repair P and Q.
3065 * However there could be one 'failed' device, in which
3066 * case we can only check one of them, possibly using the
3067 * other to generate missing data
3068 */
3069
Dan Williamsd82dfee2009-07-14 13:40:57 -07003070 switch (sh->check_state) {
3071 case check_state_idle:
3072 /* start a new check operation if there are < 2 failures */
NeilBrownf2b3b442011-07-26 11:35:19 +10003073 if (s->failed == s->q_failed) {
Dan Williamsd82dfee2009-07-14 13:40:57 -07003074 /* The only possible failed device holds Q, so it
Dan Williamsa4456852007-07-09 11:56:43 -07003075 * makes sense to check P (If anything else were failed,
3076 * we would have used P to recreate it).
3077 */
Dan Williamsd82dfee2009-07-14 13:40:57 -07003078 sh->check_state = check_state_run;
Dan Williamsa4456852007-07-09 11:56:43 -07003079 }
NeilBrownf2b3b442011-07-26 11:35:19 +10003080 if (!s->q_failed && s->failed < 2) {
Dan Williamsd82dfee2009-07-14 13:40:57 -07003081 /* Q is not failed, and we didn't use it to generate
Dan Williamsa4456852007-07-09 11:56:43 -07003082 * anything, so it makes sense to check it
3083 */
Dan Williamsd82dfee2009-07-14 13:40:57 -07003084 if (sh->check_state == check_state_run)
3085 sh->check_state = check_state_run_pq;
3086 else
3087 sh->check_state = check_state_run_q;
Dan Williamsa4456852007-07-09 11:56:43 -07003088 }
Dan Williams36d1c642009-07-14 11:48:22 -07003089
Dan Williamsd82dfee2009-07-14 13:40:57 -07003090 /* discard potentially stale zero_sum_result */
3091 sh->ops.zero_sum_result = 0;
Dan Williams36d1c642009-07-14 11:48:22 -07003092
Dan Williamsd82dfee2009-07-14 13:40:57 -07003093 if (sh->check_state == check_state_run) {
3094 /* async_xor_zero_sum destroys the contents of P */
3095 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3096 s->uptodate--;
Dan Williamsa4456852007-07-09 11:56:43 -07003097 }
Dan Williamsd82dfee2009-07-14 13:40:57 -07003098 if (sh->check_state >= check_state_run &&
3099 sh->check_state <= check_state_run_pq) {
3100 /* async_syndrome_zero_sum preserves P and Q, so
3101 * no need to mark them !uptodate here
3102 */
3103 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3104 break;
3105 }
Dan Williams36d1c642009-07-14 11:48:22 -07003106
Dan Williamsd82dfee2009-07-14 13:40:57 -07003107 /* we have 2-disk failure */
3108 BUG_ON(s->failed != 2);
3109 /* fall through */
3110 case check_state_compute_result:
3111 sh->check_state = check_state_idle;
Dan Williams36d1c642009-07-14 11:48:22 -07003112
Dan Williamsd82dfee2009-07-14 13:40:57 -07003113 /* check that a write has not made the stripe insync */
3114 if (test_bit(STRIPE_INSYNC, &sh->state))
3115 break;
Dan Williamsa4456852007-07-09 11:56:43 -07003116
3117 /* now write out any block on a failed drive,
Dan Williamsd82dfee2009-07-14 13:40:57 -07003118 * or P or Q if they were recomputed
Dan Williamsa4456852007-07-09 11:56:43 -07003119 */
Dan Williamsd82dfee2009-07-14 13:40:57 -07003120 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
Dan Williamsa4456852007-07-09 11:56:43 -07003121 if (s->failed == 2) {
NeilBrownf2b3b442011-07-26 11:35:19 +10003122 dev = &sh->dev[s->failed_num[1]];
Dan Williamsa4456852007-07-09 11:56:43 -07003123 s->locked++;
3124 set_bit(R5_LOCKED, &dev->flags);
3125 set_bit(R5_Wantwrite, &dev->flags);
3126 }
3127 if (s->failed >= 1) {
NeilBrownf2b3b442011-07-26 11:35:19 +10003128 dev = &sh->dev[s->failed_num[0]];
Dan Williamsa4456852007-07-09 11:56:43 -07003129 s->locked++;
3130 set_bit(R5_LOCKED, &dev->flags);
3131 set_bit(R5_Wantwrite, &dev->flags);
3132 }
Dan Williamsd82dfee2009-07-14 13:40:57 -07003133 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
Dan Williamsa4456852007-07-09 11:56:43 -07003134 dev = &sh->dev[pd_idx];
3135 s->locked++;
3136 set_bit(R5_LOCKED, &dev->flags);
3137 set_bit(R5_Wantwrite, &dev->flags);
3138 }
Dan Williamsd82dfee2009-07-14 13:40:57 -07003139 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
Dan Williamsa4456852007-07-09 11:56:43 -07003140 dev = &sh->dev[qd_idx];
3141 s->locked++;
3142 set_bit(R5_LOCKED, &dev->flags);
3143 set_bit(R5_Wantwrite, &dev->flags);
3144 }
3145 clear_bit(STRIPE_DEGRADED, &sh->state);
3146
3147 set_bit(STRIPE_INSYNC, &sh->state);
Dan Williamsd82dfee2009-07-14 13:40:57 -07003148 break;
3149 case check_state_run:
3150 case check_state_run_q:
3151 case check_state_run_pq:
3152 break; /* we will be called again upon completion */
3153 case check_state_check_result:
3154 sh->check_state = check_state_idle;
3155
3156 /* handle a successful check operation, if parity is correct
3157 * we are done. Otherwise update the mismatch count and repair
3158 * parity if !MD_RECOVERY_CHECK
3159 */
3160 if (sh->ops.zero_sum_result == 0) {
3161 /* both parities are correct */
3162 if (!s->failed)
3163 set_bit(STRIPE_INSYNC, &sh->state);
3164 else {
3165 /* in contrast to the raid5 case we can validate
3166 * parity, but still have a failure to write
3167 * back
3168 */
3169 sh->check_state = check_state_compute_result;
3170 /* Returning at this point means that we may go
3171 * off and bring p and/or q uptodate again so
3172 * we make sure to check zero_sum_result again
3173 * to verify if p or q need writeback
3174 */
3175 }
3176 } else {
Jianpeng Ma7f7583d2012-10-11 14:17:59 +11003177 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
Dan Williamsd82dfee2009-07-14 13:40:57 -07003178 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3179 /* don't try to repair!! */
3180 set_bit(STRIPE_INSYNC, &sh->state);
3181 else {
3182 int *target = &sh->ops.target;
3183
3184 sh->ops.target = -1;
3185 sh->ops.target2 = -1;
3186 sh->check_state = check_state_compute_run;
3187 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3188 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3189 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3190 set_bit(R5_Wantcompute,
3191 &sh->dev[pd_idx].flags);
3192 *target = pd_idx;
3193 target = &sh->ops.target2;
3194 s->uptodate++;
3195 }
3196 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3197 set_bit(R5_Wantcompute,
3198 &sh->dev[qd_idx].flags);
3199 *target = qd_idx;
3200 s->uptodate++;
3201 }
3202 }
3203 }
3204 break;
3205 case check_state_compute_run:
3206 break;
3207 default:
3208 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3209 __func__, sh->check_state,
3210 (unsigned long long) sh->sector);
3211 BUG();
Dan Williamsa4456852007-07-09 11:56:43 -07003212 }
3213}
3214
NeilBrownd1688a62011-10-11 16:49:52 +11003215static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
Dan Williamsa4456852007-07-09 11:56:43 -07003216{
3217 int i;
3218
3219 /* We have read all the blocks in this stripe and now we need to
3220 * copy some of them into a target stripe for expand.
3221 */
Dan Williamsf0a50d32007-01-02 13:52:31 -07003222 struct dma_async_tx_descriptor *tx = NULL;
Dan Williamsa4456852007-07-09 11:56:43 -07003223 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3224 for (i = 0; i < sh->disks; i++)
NeilBrown34e04e82009-03-31 15:10:16 +11003225 if (i != sh->pd_idx && i != sh->qd_idx) {
NeilBrown911d4ee2009-03-31 14:39:38 +11003226 int dd_idx, j;
Dan Williamsa4456852007-07-09 11:56:43 -07003227 struct stripe_head *sh2;
Dan Williamsa08abd82009-06-03 11:43:59 -07003228 struct async_submit_ctl submit;
Dan Williamsa4456852007-07-09 11:56:43 -07003229
NeilBrown784052e2009-03-31 15:19:07 +11003230 sector_t bn = compute_blocknr(sh, i, 1);
NeilBrown911d4ee2009-03-31 14:39:38 +11003231 sector_t s = raid5_compute_sector(conf, bn, 0,
3232 &dd_idx, NULL);
NeilBrowna8c906c2009-06-09 14:39:59 +10003233 sh2 = get_active_stripe(conf, s, 0, 1, 1);
Dan Williamsa4456852007-07-09 11:56:43 -07003234 if (sh2 == NULL)
3235 /* so far only the early blocks of this stripe
3236 * have been requested. When later blocks
3237 * get requested, we will try again
3238 */
3239 continue;
3240 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3241 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3242 /* must have already done this block */
3243 release_stripe(sh2);
3244 continue;
3245 }
Dan Williamsf0a50d32007-01-02 13:52:31 -07003246
3247 /* place all the copies on one channel */
Dan Williamsa08abd82009-06-03 11:43:59 -07003248 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
Dan Williamsf0a50d32007-01-02 13:52:31 -07003249 tx = async_memcpy(sh2->dev[dd_idx].page,
Dan Williams88ba2aa2009-04-09 16:16:18 -07003250 sh->dev[i].page, 0, 0, STRIPE_SIZE,
Dan Williamsa08abd82009-06-03 11:43:59 -07003251 &submit);
Dan Williamsf0a50d32007-01-02 13:52:31 -07003252
Dan Williamsa4456852007-07-09 11:56:43 -07003253 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3254 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3255 for (j = 0; j < conf->raid_disks; j++)
3256 if (j != sh2->pd_idx &&
NeilBrown86c374b2011-07-27 11:00:36 +10003257 j != sh2->qd_idx &&
Dan Williamsa4456852007-07-09 11:56:43 -07003258 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3259 break;
3260 if (j == conf->raid_disks) {
3261 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3262 set_bit(STRIPE_HANDLE, &sh2->state);
3263 }
3264 release_stripe(sh2);
Dan Williamsf0a50d32007-01-02 13:52:31 -07003265
Dan Williamsa4456852007-07-09 11:56:43 -07003266 }
NeilBrowna2e08552007-09-11 15:23:36 -07003267 /* done submitting copies, wait for them to complete */
NeilBrown749586b2012-11-20 14:11:15 +11003268 async_tx_quiesce(&tx);
Dan Williamsa4456852007-07-09 11:56:43 -07003269}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003270
3271/*
3272 * handle_stripe - do things to a stripe.
3273 *
NeilBrown9a3e1102011-12-23 10:17:53 +11003274 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3275 * state of various bits to see what needs to be done.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276 * Possible results:
NeilBrown9a3e1102011-12-23 10:17:53 +11003277 * return some read requests which now have data
3278 * return some write requests which are safely on storage
Linus Torvalds1da177e2005-04-16 15:20:36 -07003279 * schedule a read on some buffers
3280 * schedule a write of some buffers
3281 * return confirmation of parity correctness
3282 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003283 */
Dan Williamsa4456852007-07-09 11:56:43 -07003284
NeilBrownacfe7262011-07-27 11:00:36 +10003285static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
NeilBrown16a53ec2006-06-26 00:27:38 -07003286{
NeilBrownd1688a62011-10-11 16:49:52 +11003287 struct r5conf *conf = sh->raid_conf;
NeilBrownf4168852007-02-28 20:11:53 -08003288 int disks = sh->disks;
NeilBrown474af965fe2011-07-27 11:00:36 +10003289 struct r5dev *dev;
3290 int i;
NeilBrown9a3e1102011-12-23 10:17:53 +11003291 int do_recovery = 0;
NeilBrown16a53ec2006-06-26 00:27:38 -07003292
NeilBrownacfe7262011-07-27 11:00:36 +10003293 memset(s, 0, sizeof(*s));
NeilBrown16a53ec2006-06-26 00:27:38 -07003294
NeilBrownacfe7262011-07-27 11:00:36 +10003295 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3296 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3297 s->failed_num[0] = -1;
3298 s->failed_num[1] = -1;
3299
3300 /* Now to look around and see what can be done */
NeilBrown16a53ec2006-06-26 00:27:38 -07003301 rcu_read_lock();
3302 for (i=disks; i--; ) {
NeilBrown3cb03002011-10-11 16:45:26 +11003303 struct md_rdev *rdev;
NeilBrown31c176e2011-07-28 11:39:22 +10003304 sector_t first_bad;
3305 int bad_sectors;
3306 int is_bad = 0;
NeilBrownacfe7262011-07-27 11:00:36 +10003307
NeilBrown16a53ec2006-06-26 00:27:38 -07003308 dev = &sh->dev[i];
NeilBrown16a53ec2006-06-26 00:27:38 -07003309
Dan Williams45b42332007-07-09 11:56:43 -07003310 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
NeilBrown9a3e1102011-12-23 10:17:53 +11003311 i, dev->flags,
3312 dev->toread, dev->towrite, dev->written);
Yuri Tikhonov6c0069c2009-08-29 19:13:13 -07003313 /* maybe we can reply to a read
3314 *
3315 * new wantfill requests are only permitted while
3316 * ops_complete_biofill is guaranteed to be inactive
3317 */
3318 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3319 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3320 set_bit(R5_Wantfill, &dev->flags);
NeilBrown16a53ec2006-06-26 00:27:38 -07003321
3322 /* now count some things */
NeilBrowncc940152011-07-26 11:35:35 +10003323 if (test_bit(R5_LOCKED, &dev->flags))
3324 s->locked++;
3325 if (test_bit(R5_UPTODATE, &dev->flags))
3326 s->uptodate++;
Dan Williams2d6e4ec2009-09-16 12:11:54 -07003327 if (test_bit(R5_Wantcompute, &dev->flags)) {
NeilBrowncc940152011-07-26 11:35:35 +10003328 s->compute++;
3329 BUG_ON(s->compute > 2);
Dan Williams2d6e4ec2009-09-16 12:11:54 -07003330 }
NeilBrown16a53ec2006-06-26 00:27:38 -07003331
NeilBrownacfe7262011-07-27 11:00:36 +10003332 if (test_bit(R5_Wantfill, &dev->flags))
NeilBrowncc940152011-07-26 11:35:35 +10003333 s->to_fill++;
NeilBrownacfe7262011-07-27 11:00:36 +10003334 else if (dev->toread)
NeilBrowncc940152011-07-26 11:35:35 +10003335 s->to_read++;
NeilBrown16a53ec2006-06-26 00:27:38 -07003336 if (dev->towrite) {
NeilBrowncc940152011-07-26 11:35:35 +10003337 s->to_write++;
NeilBrown16a53ec2006-06-26 00:27:38 -07003338 if (!test_bit(R5_OVERWRITE, &dev->flags))
NeilBrowncc940152011-07-26 11:35:35 +10003339 s->non_overwrite++;
NeilBrown16a53ec2006-06-26 00:27:38 -07003340 }
Dan Williamsa4456852007-07-09 11:56:43 -07003341 if (dev->written)
NeilBrowncc940152011-07-26 11:35:35 +10003342 s->written++;
NeilBrown14a75d32011-12-23 10:17:52 +11003343 /* Prefer to use the replacement for reads, but only
3344 * if it is recovered enough and has no bad blocks.
3345 */
3346 rdev = rcu_dereference(conf->disks[i].replacement);
3347 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3348 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3349 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3350 &first_bad, &bad_sectors))
3351 set_bit(R5_ReadRepl, &dev->flags);
3352 else {
NeilBrown9a3e1102011-12-23 10:17:53 +11003353 if (rdev)
3354 set_bit(R5_NeedReplace, &dev->flags);
NeilBrown14a75d32011-12-23 10:17:52 +11003355 rdev = rcu_dereference(conf->disks[i].rdev);
3356 clear_bit(R5_ReadRepl, &dev->flags);
3357 }
NeilBrown9283d8c2011-12-08 16:27:57 +11003358 if (rdev && test_bit(Faulty, &rdev->flags))
3359 rdev = NULL;
NeilBrown31c176e2011-07-28 11:39:22 +10003360 if (rdev) {
3361 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3362 &first_bad, &bad_sectors);
3363 if (s->blocked_rdev == NULL
3364 && (test_bit(Blocked, &rdev->flags)
3365 || is_bad < 0)) {
3366 if (is_bad < 0)
3367 set_bit(BlockedBadBlocks,
3368 &rdev->flags);
3369 s->blocked_rdev = rdev;
3370 atomic_inc(&rdev->nr_pending);
3371 }
Dan Williams6bfe0b42008-04-30 00:52:32 -07003372 }
NeilBrown415e72d2010-06-17 17:25:21 +10003373 clear_bit(R5_Insync, &dev->flags);
3374 if (!rdev)
3375 /* Not in-sync */;
NeilBrown31c176e2011-07-28 11:39:22 +10003376 else if (is_bad) {
3377 /* also not in-sync */
NeilBrown18b98372012-04-01 23:48:38 +10003378 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3379 test_bit(R5_UPTODATE, &dev->flags)) {
NeilBrown31c176e2011-07-28 11:39:22 +10003380 /* treat as in-sync, but with a read error
3381 * which we can now try to correct
3382 */
3383 set_bit(R5_Insync, &dev->flags);
3384 set_bit(R5_ReadError, &dev->flags);
3385 }
3386 } else if (test_bit(In_sync, &rdev->flags))
NeilBrown415e72d2010-06-17 17:25:21 +10003387 set_bit(R5_Insync, &dev->flags);
NeilBrown30d7a482011-12-23 09:57:00 +11003388 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
NeilBrown415e72d2010-06-17 17:25:21 +10003389 /* in sync if before recovery_offset */
NeilBrown30d7a482011-12-23 09:57:00 +11003390 set_bit(R5_Insync, &dev->flags);
3391 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3392 test_bit(R5_Expanded, &dev->flags))
3393 /* If we've reshaped into here, we assume it is Insync.
3394 * We will shortly update recovery_offset to make
3395 * it official.
3396 */
3397 set_bit(R5_Insync, &dev->flags);
3398
NeilBrownc44bc442014-01-06 13:19:42 +11003399 if (test_bit(R5_WriteError, &dev->flags)) {
NeilBrown14a75d32011-12-23 10:17:52 +11003400 /* This flag does not apply to '.replacement'
3401 * only to .rdev, so make sure to check that*/
3402 struct md_rdev *rdev2 = rcu_dereference(
3403 conf->disks[i].rdev);
3404 if (rdev2 == rdev)
3405 clear_bit(R5_Insync, &dev->flags);
3406 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
NeilBrownbc2607f2011-07-28 11:39:22 +10003407 s->handle_bad_blocks = 1;
NeilBrown14a75d32011-12-23 10:17:52 +11003408 atomic_inc(&rdev2->nr_pending);
NeilBrownbc2607f2011-07-28 11:39:22 +10003409 } else
3410 clear_bit(R5_WriteError, &dev->flags);
3411 }
NeilBrownc44bc442014-01-06 13:19:42 +11003412 if (test_bit(R5_MadeGood, &dev->flags)) {
NeilBrown14a75d32011-12-23 10:17:52 +11003413 /* This flag does not apply to '.replacement'
3414 * only to .rdev, so make sure to check that*/
3415 struct md_rdev *rdev2 = rcu_dereference(
3416 conf->disks[i].rdev);
3417 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
NeilBrownb84db562011-07-28 11:39:23 +10003418 s->handle_bad_blocks = 1;
NeilBrown14a75d32011-12-23 10:17:52 +11003419 atomic_inc(&rdev2->nr_pending);
NeilBrownb84db562011-07-28 11:39:23 +10003420 } else
3421 clear_bit(R5_MadeGood, &dev->flags);
3422 }
NeilBrown977df362011-12-23 10:17:53 +11003423 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3424 struct md_rdev *rdev2 = rcu_dereference(
3425 conf->disks[i].replacement);
3426 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3427 s->handle_bad_blocks = 1;
3428 atomic_inc(&rdev2->nr_pending);
3429 } else
3430 clear_bit(R5_MadeGoodRepl, &dev->flags);
3431 }
NeilBrown415e72d2010-06-17 17:25:21 +10003432 if (!test_bit(R5_Insync, &dev->flags)) {
NeilBrown16a53ec2006-06-26 00:27:38 -07003433 /* The ReadError flag will just be confusing now */
3434 clear_bit(R5_ReadError, &dev->flags);
3435 clear_bit(R5_ReWrite, &dev->flags);
3436 }
NeilBrown415e72d2010-06-17 17:25:21 +10003437 if (test_bit(R5_ReadError, &dev->flags))
3438 clear_bit(R5_Insync, &dev->flags);
3439 if (!test_bit(R5_Insync, &dev->flags)) {
NeilBrowncc940152011-07-26 11:35:35 +10003440 if (s->failed < 2)
3441 s->failed_num[s->failed] = i;
3442 s->failed++;
NeilBrown9a3e1102011-12-23 10:17:53 +11003443 if (rdev && !test_bit(Faulty, &rdev->flags))
3444 do_recovery = 1;
NeilBrown415e72d2010-06-17 17:25:21 +10003445 }
NeilBrown16a53ec2006-06-26 00:27:38 -07003446 }
NeilBrown9a3e1102011-12-23 10:17:53 +11003447 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3448 /* If there is a failed device being replaced,
3449 * we must be recovering.
3450 * else if we are after recovery_cp, we must be syncing
majianpengc6d2e082012-04-02 01:16:59 +10003451 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
NeilBrown9a3e1102011-12-23 10:17:53 +11003452 * else we can only be replacing
3453 * sync and recovery both need to read all devices, and so
3454 * use the same flag.
3455 */
3456 if (do_recovery ||
majianpengc6d2e082012-04-02 01:16:59 +10003457 sh->sector >= conf->mddev->recovery_cp ||
3458 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
NeilBrown9a3e1102011-12-23 10:17:53 +11003459 s->syncing = 1;
3460 else
3461 s->replacing = 1;
3462 }
NeilBrown16a53ec2006-06-26 00:27:38 -07003463 rcu_read_unlock();
NeilBrowncc940152011-07-26 11:35:35 +10003464}
NeilBrownf4168852007-02-28 20:11:53 -08003465
NeilBrowncc940152011-07-26 11:35:35 +10003466static void handle_stripe(struct stripe_head *sh)
3467{
3468 struct stripe_head_state s;
NeilBrownd1688a62011-10-11 16:49:52 +11003469 struct r5conf *conf = sh->raid_conf;
NeilBrown3687c062011-07-27 11:00:36 +10003470 int i;
NeilBrown84789552011-07-27 11:00:36 +10003471 int prexor;
3472 int disks = sh->disks;
NeilBrown474af965fe2011-07-27 11:00:36 +10003473 struct r5dev *pdev, *qdev;
NeilBrowncc940152011-07-26 11:35:35 +10003474
3475 clear_bit(STRIPE_HANDLE, &sh->state);
Dan Williams257a4b42011-11-08 16:22:06 +11003476 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
NeilBrowncc940152011-07-26 11:35:35 +10003477 /* already being handled, ensure it gets handled
3478 * again when current action finishes */
3479 set_bit(STRIPE_HANDLE, &sh->state);
3480 return;
3481 }
3482
NeilBrownf8dfcff2013-03-12 12:18:06 +11003483 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3484 spin_lock(&sh->stripe_lock);
3485 /* Cannot process 'sync' concurrently with 'discard' */
3486 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3487 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3488 set_bit(STRIPE_SYNCING, &sh->state);
3489 clear_bit(STRIPE_INSYNC, &sh->state);
NeilBrownc1dadcc2013-07-22 12:57:21 +10003490 clear_bit(STRIPE_REPLACED, &sh->state);
NeilBrownf8dfcff2013-03-12 12:18:06 +11003491 }
3492 spin_unlock(&sh->stripe_lock);
NeilBrowncc940152011-07-26 11:35:35 +10003493 }
3494 clear_bit(STRIPE_DELAYED, &sh->state);
3495
3496 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3497 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3498 (unsigned long long)sh->sector, sh->state,
3499 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3500 sh->check_state, sh->reconstruct_state);
NeilBrowncc940152011-07-26 11:35:35 +10003501
NeilBrownacfe7262011-07-27 11:00:36 +10003502 analyse_stripe(sh, &s);
NeilBrownc5a31002011-07-27 11:00:36 +10003503
NeilBrownbc2607f2011-07-28 11:39:22 +10003504 if (s.handle_bad_blocks) {
3505 set_bit(STRIPE_HANDLE, &sh->state);
3506 goto finish;
3507 }
3508
NeilBrown474af965fe2011-07-27 11:00:36 +10003509 if (unlikely(s.blocked_rdev)) {
3510 if (s.syncing || s.expanding || s.expanded ||
NeilBrown9a3e1102011-12-23 10:17:53 +11003511 s.replacing || s.to_write || s.written) {
NeilBrown474af965fe2011-07-27 11:00:36 +10003512 set_bit(STRIPE_HANDLE, &sh->state);
3513 goto finish;
3514 }
3515 /* There is nothing for the blocked_rdev to block */
3516 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3517 s.blocked_rdev = NULL;
3518 }
3519
3520 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3521 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3522 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3523 }
3524
3525 pr_debug("locked=%d uptodate=%d to_read=%d"
3526 " to_write=%d failed=%d failed_num=%d,%d\n",
3527 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3528 s.failed_num[0], s.failed_num[1]);
3529 /* check if the array has lost more than max_degraded devices and,
3530 * if so, some requests might need to be failed.
3531 */
NeilBrown9a3f5302011-11-08 16:22:01 +11003532 if (s.failed > conf->max_degraded) {
3533 sh->check_state = 0;
3534 sh->reconstruct_state = 0;
3535 if (s.to_read+s.to_write+s.written)
3536 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
NeilBrown9a3e1102011-12-23 10:17:53 +11003537 if (s.syncing + s.replacing)
NeilBrown9a3f5302011-11-08 16:22:01 +11003538 handle_failed_sync(conf, sh, &s);
3539 }
NeilBrown474af965fe2011-07-27 11:00:36 +10003540
NeilBrown84789552011-07-27 11:00:36 +10003541 /* Now we check to see if any write operations have recently
3542 * completed
3543 */
3544 prexor = 0;
3545 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3546 prexor = 1;
3547 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3548 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3549 sh->reconstruct_state = reconstruct_state_idle;
3550
3551 /* All the 'written' buffers and the parity block are ready to
3552 * be written back to disk
3553 */
Shaohua Li9e4447682012-10-11 13:49:49 +11003554 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3555 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
NeilBrown84789552011-07-27 11:00:36 +10003556 BUG_ON(sh->qd_idx >= 0 &&
Shaohua Li9e4447682012-10-11 13:49:49 +11003557 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3558 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
NeilBrown84789552011-07-27 11:00:36 +10003559 for (i = disks; i--; ) {
3560 struct r5dev *dev = &sh->dev[i];
3561 if (test_bit(R5_LOCKED, &dev->flags) &&
3562 (i == sh->pd_idx || i == sh->qd_idx ||
3563 dev->written)) {
3564 pr_debug("Writing block %d\n", i);
3565 set_bit(R5_Wantwrite, &dev->flags);
3566 if (prexor)
3567 continue;
NeilBrown318a3d52014-08-13 09:57:07 +10003568 if (s.failed > 1)
3569 continue;
NeilBrown84789552011-07-27 11:00:36 +10003570 if (!test_bit(R5_Insync, &dev->flags) ||
3571 ((i == sh->pd_idx || i == sh->qd_idx) &&
3572 s.failed == 0))
3573 set_bit(STRIPE_INSYNC, &sh->state);
3574 }
3575 }
3576 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3577 s.dec_preread_active = 1;
3578 }
3579
NeilBrownef5b7c62012-11-22 09:13:36 +11003580 /*
3581 * might be able to return some write requests if the parity blocks
3582 * are safe, or on a failed drive
3583 */
3584 pdev = &sh->dev[sh->pd_idx];
3585 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3586 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3587 qdev = &sh->dev[sh->qd_idx];
3588 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3589 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3590 || conf->level < 6;
3591
3592 if (s.written &&
3593 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3594 && !test_bit(R5_LOCKED, &pdev->flags)
3595 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3596 test_bit(R5_Discard, &pdev->flags))))) &&
3597 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3598 && !test_bit(R5_LOCKED, &qdev->flags)
3599 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3600 test_bit(R5_Discard, &qdev->flags))))))
3601 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3602
3603 /* Now we might consider reading some blocks, either to check/generate
3604 * parity, or to satisfy requests
3605 * or to load a block that is being partially written.
3606 */
3607 if (s.to_read || s.non_overwrite
3608 || (conf->level == 6 && s.to_write && s.failed)
3609 || (s.syncing && (s.uptodate + s.compute < disks))
3610 || s.replacing
3611 || s.expanding)
3612 handle_stripe_fill(sh, &s, disks);
3613
NeilBrown84789552011-07-27 11:00:36 +10003614 /* Now to consider new write requests and what else, if anything
3615 * should be read. We do not handle new writes when:
3616 * 1/ A 'write' operation (copy+xor) is already in flight.
3617 * 2/ A 'check' operation is in flight, as it may clobber the parity
3618 * block.
3619 */
3620 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3621 handle_stripe_dirtying(conf, sh, &s, disks);
3622
3623 /* maybe we need to check and possibly fix the parity for this stripe
3624 * Any reads will already have been scheduled, so we just see if enough
3625 * data is available. The parity check is held off while parity
3626 * dependent operations are in flight.
3627 */
3628 if (sh->check_state ||
3629 (s.syncing && s.locked == 0 &&
3630 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3631 !test_bit(STRIPE_INSYNC, &sh->state))) {
3632 if (conf->level == 6)
3633 handle_parity_checks6(conf, sh, &s, disks);
3634 else
3635 handle_parity_checks5(conf, sh, &s, disks);
3636 }
NeilBrownc5a31002011-07-27 11:00:36 +10003637
NeilBrownc1dadcc2013-07-22 12:57:21 +10003638 if ((s.replacing || s.syncing) && s.locked == 0
3639 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3640 && !test_bit(STRIPE_REPLACED, &sh->state)) {
NeilBrown9a3e1102011-12-23 10:17:53 +11003641 /* Write out to replacement devices where possible */
3642 for (i = 0; i < conf->raid_disks; i++)
NeilBrownc1dadcc2013-07-22 12:57:21 +10003643 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3644 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
NeilBrown9a3e1102011-12-23 10:17:53 +11003645 set_bit(R5_WantReplace, &sh->dev[i].flags);
3646 set_bit(R5_LOCKED, &sh->dev[i].flags);
3647 s.locked++;
3648 }
NeilBrownc1dadcc2013-07-22 12:57:21 +10003649 if (s.replacing)
3650 set_bit(STRIPE_INSYNC, &sh->state);
3651 set_bit(STRIPE_REPLACED, &sh->state);
NeilBrown9a3e1102011-12-23 10:17:53 +11003652 }
3653 if ((s.syncing || s.replacing) && s.locked == 0 &&
NeilBrownc1dadcc2013-07-22 12:57:21 +10003654 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
NeilBrown9a3e1102011-12-23 10:17:53 +11003655 test_bit(STRIPE_INSYNC, &sh->state)) {
NeilBrownc5a31002011-07-27 11:00:36 +10003656 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3657 clear_bit(STRIPE_SYNCING, &sh->state);
NeilBrownf8dfcff2013-03-12 12:18:06 +11003658 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3659 wake_up(&conf->wait_for_overlap);
NeilBrownc5a31002011-07-27 11:00:36 +10003660 }
3661
3662 /* If the failed drives are just a ReadError, then we might need
3663 * to progress the repair/check process
3664 */
3665 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3666 for (i = 0; i < s.failed; i++) {
3667 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3668 if (test_bit(R5_ReadError, &dev->flags)
3669 && !test_bit(R5_LOCKED, &dev->flags)
3670 && test_bit(R5_UPTODATE, &dev->flags)
3671 ) {
3672 if (!test_bit(R5_ReWrite, &dev->flags)) {
3673 set_bit(R5_Wantwrite, &dev->flags);
3674 set_bit(R5_ReWrite, &dev->flags);
3675 set_bit(R5_LOCKED, &dev->flags);
3676 s.locked++;
3677 } else {
3678 /* let's read it back */
3679 set_bit(R5_Wantread, &dev->flags);
3680 set_bit(R5_LOCKED, &dev->flags);
3681 s.locked++;
3682 }
3683 }
3684 }
3685
3686
NeilBrown3687c062011-07-27 11:00:36 +10003687 /* Finish reconstruct operations initiated by the expansion process */
3688 if (sh->reconstruct_state == reconstruct_state_result) {
3689 struct stripe_head *sh_src
3690 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3691 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3692 /* sh cannot be written until sh_src has been read.
3693 * so arrange for sh to be delayed a little
3694 */
3695 set_bit(STRIPE_DELAYED, &sh->state);
3696 set_bit(STRIPE_HANDLE, &sh->state);
3697 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3698 &sh_src->state))
3699 atomic_inc(&conf->preread_active_stripes);
3700 release_stripe(sh_src);
3701 goto finish;
3702 }
3703 if (sh_src)
3704 release_stripe(sh_src);
NeilBrown16a53ec2006-06-26 00:27:38 -07003705
NeilBrown3687c062011-07-27 11:00:36 +10003706 sh->reconstruct_state = reconstruct_state_idle;
3707 clear_bit(STRIPE_EXPANDING, &sh->state);
3708 for (i = conf->raid_disks; i--; ) {
3709 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3710 set_bit(R5_LOCKED, &sh->dev[i].flags);
3711 s.locked++;
3712 }
3713 }
3714
3715 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3716 !sh->reconstruct_state) {
3717 /* Need to write out all blocks after computing parity */
3718 sh->disks = conf->raid_disks;
3719 stripe_set_idx(sh->sector, conf, 0, sh);
3720 schedule_reconstruction(sh, &s, 1, 1);
3721 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3722 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3723 atomic_dec(&conf->reshape_stripes);
3724 wake_up(&conf->wait_for_overlap);
3725 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3726 }
3727
3728 if (s.expanding && s.locked == 0 &&
3729 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3730 handle_stripe_expansion(conf, sh);
3731
3732finish:
Dan Williams6bfe0b42008-04-30 00:52:32 -07003733 /* wait for this device to become unblocked */
NeilBrown5f066c62012-07-03 12:13:29 +10003734 if (unlikely(s.blocked_rdev)) {
3735 if (conf->mddev->external)
3736 md_wait_for_blocked_rdev(s.blocked_rdev,
3737 conf->mddev);
3738 else
3739 /* Internal metadata will immediately
3740 * be written by raid5d, so we don't
3741 * need to wait here.
3742 */
3743 rdev_dec_pending(s.blocked_rdev,
3744 conf->mddev);
3745 }
Dan Williams6bfe0b42008-04-30 00:52:32 -07003746
NeilBrownbc2607f2011-07-28 11:39:22 +10003747 if (s.handle_bad_blocks)
3748 for (i = disks; i--; ) {
NeilBrown3cb03002011-10-11 16:45:26 +11003749 struct md_rdev *rdev;
NeilBrownbc2607f2011-07-28 11:39:22 +10003750 struct r5dev *dev = &sh->dev[i];
3751 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3752 /* We own a safe reference to the rdev */
3753 rdev = conf->disks[i].rdev;
3754 if (!rdev_set_badblocks(rdev, sh->sector,
3755 STRIPE_SECTORS, 0))
3756 md_error(conf->mddev, rdev);
3757 rdev_dec_pending(rdev, conf->mddev);
3758 }
NeilBrownb84db562011-07-28 11:39:23 +10003759 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3760 rdev = conf->disks[i].rdev;
3761 rdev_clear_badblocks(rdev, sh->sector,
NeilBrownc6563a82012-05-21 09:27:00 +10003762 STRIPE_SECTORS, 0);
NeilBrownb84db562011-07-28 11:39:23 +10003763 rdev_dec_pending(rdev, conf->mddev);
3764 }
NeilBrown977df362011-12-23 10:17:53 +11003765 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3766 rdev = conf->disks[i].replacement;
NeilBrowndd054fc2011-12-23 10:17:53 +11003767 if (!rdev)
3768 /* rdev have been moved down */
3769 rdev = conf->disks[i].rdev;
NeilBrown977df362011-12-23 10:17:53 +11003770 rdev_clear_badblocks(rdev, sh->sector,
NeilBrownc6563a82012-05-21 09:27:00 +10003771 STRIPE_SECTORS, 0);
NeilBrown977df362011-12-23 10:17:53 +11003772 rdev_dec_pending(rdev, conf->mddev);
3773 }
NeilBrownbc2607f2011-07-28 11:39:22 +10003774 }
3775
Yuri Tikhonov6c0069c2009-08-29 19:13:13 -07003776 if (s.ops_request)
3777 raid_run_ops(sh, s.ops_request);
3778
Dan Williamsf0e43bc2008-06-28 08:31:55 +10003779 ops_run_io(sh, &s);
3780
NeilBrownc5709ef2011-07-26 11:35:20 +10003781 if (s.dec_preread_active) {
NeilBrown729a1862009-12-14 12:49:50 +11003782 /* We delay this until after ops_run_io so that if make_request
Tejun Heoe9c74692010-09-03 11:56:18 +02003783 * is waiting on a flush, it won't continue until the writes
NeilBrown729a1862009-12-14 12:49:50 +11003784 * have actually been submitted.
3785 */
3786 atomic_dec(&conf->preread_active_stripes);
3787 if (atomic_read(&conf->preread_active_stripes) <
3788 IO_THRESHOLD)
3789 md_wakeup_thread(conf->mddev->thread);
3790 }
3791
NeilBrownc5709ef2011-07-26 11:35:20 +10003792 return_io(s.return_bi);
NeilBrown16a53ec2006-06-26 00:27:38 -07003793
Dan Williams257a4b42011-11-08 16:22:06 +11003794 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
NeilBrown16a53ec2006-06-26 00:27:38 -07003795}
3796
NeilBrownd1688a62011-10-11 16:49:52 +11003797static void raid5_activate_delayed(struct r5conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798{
3799 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3800 while (!list_empty(&conf->delayed_list)) {
3801 struct list_head *l = conf->delayed_list.next;
3802 struct stripe_head *sh;
3803 sh = list_entry(l, struct stripe_head, lru);
3804 list_del_init(l);
3805 clear_bit(STRIPE_DELAYED, &sh->state);
3806 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3807 atomic_inc(&conf->preread_active_stripes);
Dan Williams8b3e6cd2008-04-28 02:15:53 -07003808 list_add_tail(&sh->lru, &conf->hold_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003809 }
NeilBrown482c0832011-04-18 18:25:42 +10003810 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003811}
3812
NeilBrownd1688a62011-10-11 16:49:52 +11003813static void activate_bit_delay(struct r5conf *conf)
NeilBrown72626682005-09-09 16:23:54 -07003814{
3815 /* device_lock is held */
3816 struct list_head head;
3817 list_add(&head, &conf->bitmap_list);
3818 list_del_init(&conf->bitmap_list);
3819 while (!list_empty(&head)) {
3820 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3821 list_del_init(&sh->lru);
3822 atomic_inc(&sh->count);
3823 __release_stripe(conf, sh);
3824 }
3825}
3826
NeilBrownfd01b882011-10-11 16:47:53 +11003827int md_raid5_congested(struct mddev *mddev, int bits)
NeilBrownf022b2f2006-10-03 01:15:56 -07003828{
NeilBrownd1688a62011-10-11 16:49:52 +11003829 struct r5conf *conf = mddev->private;
NeilBrownf022b2f2006-10-03 01:15:56 -07003830
3831 /* No difference between reads and writes. Just check
3832 * how busy the stripe_cache is
3833 */
NeilBrown3fa841d2009-09-23 18:10:29 +10003834
NeilBrownf022b2f2006-10-03 01:15:56 -07003835 if (conf->inactive_blocked)
3836 return 1;
3837 if (conf->quiesce)
3838 return 1;
3839 if (list_empty_careful(&conf->inactive_list))
3840 return 1;
3841
3842 return 0;
3843}
NeilBrown11d8a6e2010-07-26 11:57:07 +10003844EXPORT_SYMBOL_GPL(md_raid5_congested);
3845
3846static int raid5_congested(void *data, int bits)
3847{
NeilBrownfd01b882011-10-11 16:47:53 +11003848 struct mddev *mddev = data;
NeilBrown11d8a6e2010-07-26 11:57:07 +10003849
3850 return mddev_congested(mddev, bits) ||
3851 md_raid5_congested(mddev, bits);
3852}
NeilBrownf022b2f2006-10-03 01:15:56 -07003853
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08003854/* We want read requests to align with chunks where possible,
3855 * but write requests don't need to.
3856 */
Alasdair G Kergoncc371e62008-07-03 09:53:43 +02003857static int raid5_mergeable_bvec(struct request_queue *q,
3858 struct bvec_merge_data *bvm,
3859 struct bio_vec *biovec)
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08003860{
NeilBrownfd01b882011-10-11 16:47:53 +11003861 struct mddev *mddev = q->queuedata;
Alasdair G Kergoncc371e62008-07-03 09:53:43 +02003862 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08003863 int max;
Andre Noll9d8f0362009-06-18 08:45:01 +10003864 unsigned int chunk_sectors = mddev->chunk_sectors;
Alasdair G Kergoncc371e62008-07-03 09:53:43 +02003865 unsigned int bio_sectors = bvm->bi_size >> 9;
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08003866
Alasdair G Kergoncc371e62008-07-03 09:53:43 +02003867 if ((bvm->bi_rw & 1) == WRITE)
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08003868 return biovec->bv_len; /* always allow writes to be mergeable */
3869
Andre Noll664e7c42009-06-18 08:45:27 +10003870 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3871 chunk_sectors = mddev->new_chunk_sectors;
Raz Ben-Jehuda(caro)23032a02006-12-10 02:20:45 -08003872 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3873 if (max < 0) max = 0;
3874 if (max <= biovec->bv_len && bio_sectors == 0)
3875 return biovec->bv_len;
3876 else
3877 return max;
3878}
3879
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003880
NeilBrownfd01b882011-10-11 16:47:53 +11003881static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003882{
3883 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
Andre Noll9d8f0362009-06-18 08:45:01 +10003884 unsigned int chunk_sectors = mddev->chunk_sectors;
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08003885 unsigned int bio_sectors = bio_sectors(bio);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003886
Andre Noll664e7c42009-06-18 08:45:27 +10003887 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3888 chunk_sectors = mddev->new_chunk_sectors;
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003889 return chunk_sectors >=
3890 ((sector & (chunk_sectors - 1)) + bio_sectors);
3891}
3892
3893/*
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003894 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3895 * later sampled by raid5d.
3896 */
NeilBrownd1688a62011-10-11 16:49:52 +11003897static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003898{
3899 unsigned long flags;
3900
3901 spin_lock_irqsave(&conf->device_lock, flags);
3902
3903 bi->bi_next = conf->retry_read_aligned_list;
3904 conf->retry_read_aligned_list = bi;
3905
3906 spin_unlock_irqrestore(&conf->device_lock, flags);
3907 md_wakeup_thread(conf->mddev->thread);
3908}
3909
3910
NeilBrownd1688a62011-10-11 16:49:52 +11003911static struct bio *remove_bio_from_retry(struct r5conf *conf)
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003912{
3913 struct bio *bi;
3914
3915 bi = conf->retry_read_aligned;
3916 if (bi) {
3917 conf->retry_read_aligned = NULL;
3918 return bi;
3919 }
3920 bi = conf->retry_read_aligned_list;
3921 if(bi) {
Neil Brown387bb172007-02-08 14:20:29 -08003922 conf->retry_read_aligned_list = bi->bi_next;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003923 bi->bi_next = NULL;
Jens Axboe960e7392008-08-15 10:41:18 +02003924 /*
3925 * this sets the active strip count to 1 and the processed
3926 * strip count to zero (upper 8 bits)
3927 */
Shaohua Lie7836bd62012-07-19 16:01:31 +10003928 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003929 }
3930
3931 return bi;
3932}
3933
3934
3935/*
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003936 * The "raid5_align_endio" should check if the read succeeded and if it
3937 * did, call bio_endio on the original bio (having bio_put the new bio
3938 * first).
3939 * If the read failed..
3940 */
NeilBrown6712ecf2007-09-27 12:47:43 +02003941static void raid5_align_endio(struct bio *bi, int error)
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003942{
3943 struct bio* raid_bi = bi->bi_private;
NeilBrownfd01b882011-10-11 16:47:53 +11003944 struct mddev *mddev;
NeilBrownd1688a62011-10-11 16:49:52 +11003945 struct r5conf *conf;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003946 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
NeilBrown3cb03002011-10-11 16:45:26 +11003947 struct md_rdev *rdev;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003948
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003949 bio_put(bi);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003950
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003951 rdev = (void*)raid_bi->bi_next;
3952 raid_bi->bi_next = NULL;
NeilBrown2b7f2222010-03-25 16:06:03 +11003953 mddev = rdev->mddev;
3954 conf = mddev->private;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003955
3956 rdev_dec_pending(rdev, conf->mddev);
3957
3958 if (!error && uptodate) {
Linus Torvalds0a82a8d2013-04-18 09:00:26 -07003959 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
3960 raid_bi, 0);
NeilBrown6712ecf2007-09-27 12:47:43 +02003961 bio_endio(raid_bi, 0);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003962 if (atomic_dec_and_test(&conf->active_aligned_reads))
3963 wake_up(&conf->wait_for_stripe);
NeilBrown6712ecf2007-09-27 12:47:43 +02003964 return;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003965 }
3966
3967
Dan Williams45b42332007-07-09 11:56:43 -07003968 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08003969
3970 add_bio_to_retry(raid_bi, conf);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003971}
3972
Neil Brown387bb172007-02-08 14:20:29 -08003973static int bio_fits_rdev(struct bio *bi)
3974{
Jens Axboe165125e2007-07-24 09:28:11 +02003975 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
Neil Brown387bb172007-02-08 14:20:29 -08003976
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08003977 if (bio_sectors(bi) > queue_max_sectors(q))
Neil Brown387bb172007-02-08 14:20:29 -08003978 return 0;
3979 blk_recount_segments(q, bi);
Martin K. Petersen8a783622010-02-26 00:20:39 -05003980 if (bi->bi_phys_segments > queue_max_segments(q))
Neil Brown387bb172007-02-08 14:20:29 -08003981 return 0;
3982
3983 if (q->merge_bvec_fn)
3984 /* it's too hard to apply the merge_bvec_fn at this stage,
3985 * just just give up
3986 */
3987 return 0;
3988
3989 return 1;
3990}
3991
3992
NeilBrownfd01b882011-10-11 16:47:53 +11003993static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003994{
NeilBrownd1688a62011-10-11 16:49:52 +11003995 struct r5conf *conf = mddev->private;
NeilBrown8553fe7ec2009-12-14 12:49:47 +11003996 int dd_idx;
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08003997 struct bio* align_bi;
NeilBrown3cb03002011-10-11 16:45:26 +11003998 struct md_rdev *rdev;
NeilBrown671488c2011-12-23 10:17:52 +11003999 sector_t end_sector;
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004000
4001 if (!in_chunk_boundary(mddev, raid_bio)) {
Dan Williams45b42332007-07-09 11:56:43 -07004002 pr_debug("chunk_aligned_read : non aligned\n");
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004003 return 0;
4004 }
4005 /*
NeilBrowna167f662010-10-26 18:31:13 +11004006 * use bio_clone_mddev to make a copy of the bio
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004007 */
NeilBrowna167f662010-10-26 18:31:13 +11004008 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004009 if (!align_bi)
4010 return 0;
4011 /*
4012 * set bi_end_io to a new function, and set bi_private to the
4013 * original bio.
4014 */
4015 align_bi->bi_end_io = raid5_align_endio;
4016 align_bi->bi_private = raid_bio;
4017 /*
4018 * compute position
4019 */
NeilBrown112bf892009-03-31 14:39:38 +11004020 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
4021 0,
NeilBrown911d4ee2009-03-31 14:39:38 +11004022 &dd_idx, NULL);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004023
Kent Overstreetf73a1c72012-09-25 15:05:12 -07004024 end_sector = bio_end_sector(align_bi);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004025 rcu_read_lock();
NeilBrown671488c2011-12-23 10:17:52 +11004026 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4027 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4028 rdev->recovery_offset < end_sector) {
4029 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4030 if (rdev &&
4031 (test_bit(Faulty, &rdev->flags) ||
4032 !(test_bit(In_sync, &rdev->flags) ||
4033 rdev->recovery_offset >= end_sector)))
4034 rdev = NULL;
4035 }
4036 if (rdev) {
NeilBrown31c176e2011-07-28 11:39:22 +10004037 sector_t first_bad;
4038 int bad_sectors;
4039
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004040 atomic_inc(&rdev->nr_pending);
4041 rcu_read_unlock();
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004042 raid_bio->bi_next = (void*)rdev;
4043 align_bi->bi_bdev = rdev->bdev;
4044 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004045
NeilBrown31c176e2011-07-28 11:39:22 +10004046 if (!bio_fits_rdev(align_bi) ||
Kent Overstreetaa8b57a2013-02-05 15:19:29 -08004047 is_badblock(rdev, align_bi->bi_sector, bio_sectors(align_bi),
NeilBrown31c176e2011-07-28 11:39:22 +10004048 &first_bad, &bad_sectors)) {
4049 /* too big in some way, or has a known bad block */
Neil Brown387bb172007-02-08 14:20:29 -08004050 bio_put(align_bi);
4051 rdev_dec_pending(rdev, mddev);
4052 return 0;
4053 }
4054
majianpeng6c0544e2012-06-12 08:31:10 +08004055 /* No reshape active, so we can trust rdev->data_offset */
4056 align_bi->bi_sector += rdev->data_offset;
4057
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004058 spin_lock_irq(&conf->device_lock);
4059 wait_event_lock_irq(conf->wait_for_stripe,
4060 conf->quiesce == 0,
Lukas Czernereed8c022012-11-30 11:42:40 +01004061 conf->device_lock);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004062 atomic_inc(&conf->active_aligned_reads);
4063 spin_unlock_irq(&conf->device_lock);
4064
Jonathan Brassowe3620a32013-03-07 16:22:01 -06004065 if (mddev->gendisk)
4066 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4067 align_bi, disk_devt(mddev->gendisk),
4068 raid_bio->bi_sector);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004069 generic_make_request(align_bi);
4070 return 1;
4071 } else {
4072 rcu_read_unlock();
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004073 bio_put(align_bi);
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004074 return 0;
4075 }
4076}
4077
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004078/* __get_priority_stripe - get the next stripe to process
4079 *
4080 * Full stripe writes are allowed to pass preread active stripes up until
4081 * the bypass_threshold is exceeded. In general the bypass_count
4082 * increments when the handle_list is handled before the hold_list; however, it
4083 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4084 * stripe with in flight i/o. The bypass_count will be reset when the
4085 * head of the hold_list has changed, i.e. the head was promoted to the
4086 * handle_list.
4087 */
NeilBrownd1688a62011-10-11 16:49:52 +11004088static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004089{
4090 struct stripe_head *sh;
4091
4092 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4093 __func__,
4094 list_empty(&conf->handle_list) ? "empty" : "busy",
4095 list_empty(&conf->hold_list) ? "empty" : "busy",
4096 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4097
4098 if (!list_empty(&conf->handle_list)) {
4099 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
4100
4101 if (list_empty(&conf->hold_list))
4102 conf->bypass_count = 0;
4103 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4104 if (conf->hold_list.next == conf->last_hold)
4105 conf->bypass_count++;
4106 else {
4107 conf->last_hold = conf->hold_list.next;
4108 conf->bypass_count -= conf->bypass_threshold;
4109 if (conf->bypass_count < 0)
4110 conf->bypass_count = 0;
4111 }
4112 }
4113 } else if (!list_empty(&conf->hold_list) &&
4114 ((conf->bypass_threshold &&
4115 conf->bypass_count > conf->bypass_threshold) ||
4116 atomic_read(&conf->pending_full_writes) == 0)) {
4117 sh = list_entry(conf->hold_list.next,
4118 typeof(*sh), lru);
4119 conf->bypass_count -= conf->bypass_threshold;
4120 if (conf->bypass_count < 0)
4121 conf->bypass_count = 0;
4122 } else
4123 return NULL;
4124
4125 list_del_init(&sh->lru);
4126 atomic_inc(&sh->count);
4127 BUG_ON(atomic_read(&sh->count) != 1);
4128 return sh;
4129}
Raz Ben-Jehuda(caro)f6796232006-12-10 02:20:46 -08004130
Shaohua Li8811b592012-08-02 08:33:00 +10004131struct raid5_plug_cb {
4132 struct blk_plug_cb cb;
4133 struct list_head list;
4134};
4135
4136static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4137{
4138 struct raid5_plug_cb *cb = container_of(
4139 blk_cb, struct raid5_plug_cb, cb);
4140 struct stripe_head *sh;
4141 struct mddev *mddev = cb->cb.data;
4142 struct r5conf *conf = mddev->private;
NeilBrowna9add5d2012-10-31 11:59:09 +11004143 int cnt = 0;
Shaohua Li8811b592012-08-02 08:33:00 +10004144
4145 if (cb->list.next && !list_empty(&cb->list)) {
4146 spin_lock_irq(&conf->device_lock);
4147 while (!list_empty(&cb->list)) {
4148 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4149 list_del_init(&sh->lru);
4150 /*
4151 * avoid race release_stripe_plug() sees
4152 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4153 * is still in our list
4154 */
4155 smp_mb__before_clear_bit();
4156 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4157 __release_stripe(conf, sh);
NeilBrowna9add5d2012-10-31 11:59:09 +11004158 cnt++;
Shaohua Li8811b592012-08-02 08:33:00 +10004159 }
4160 spin_unlock_irq(&conf->device_lock);
4161 }
Jonathan Brassowe3620a32013-03-07 16:22:01 -06004162 if (mddev->queue)
4163 trace_block_unplug(mddev->queue, cnt, !from_schedule);
Shaohua Li8811b592012-08-02 08:33:00 +10004164 kfree(cb);
4165}
4166
4167static void release_stripe_plug(struct mddev *mddev,
4168 struct stripe_head *sh)
4169{
4170 struct blk_plug_cb *blk_cb = blk_check_plugged(
4171 raid5_unplug, mddev,
4172 sizeof(struct raid5_plug_cb));
4173 struct raid5_plug_cb *cb;
4174
4175 if (!blk_cb) {
4176 release_stripe(sh);
4177 return;
4178 }
4179
4180 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4181
4182 if (cb->list.next == NULL)
4183 INIT_LIST_HEAD(&cb->list);
4184
4185 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4186 list_add_tail(&sh->lru, &cb->list);
4187 else
4188 release_stripe(sh);
4189}
4190
Shaohua Li620125f2012-10-11 13:49:05 +11004191static void make_discard_request(struct mddev *mddev, struct bio *bi)
4192{
4193 struct r5conf *conf = mddev->private;
4194 sector_t logical_sector, last_sector;
4195 struct stripe_head *sh;
4196 int remaining;
4197 int stripe_sectors;
4198
4199 if (mddev->reshape_position != MaxSector)
4200 /* Skip discard while reshape is happening */
4201 return;
4202
4203 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4204 last_sector = bi->bi_sector + (bi->bi_size>>9);
4205
4206 bi->bi_next = NULL;
4207 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4208
4209 stripe_sectors = conf->chunk_sectors *
4210 (conf->raid_disks - conf->max_degraded);
4211 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4212 stripe_sectors);
4213 sector_div(last_sector, stripe_sectors);
4214
4215 logical_sector *= conf->chunk_sectors;
4216 last_sector *= conf->chunk_sectors;
4217
4218 for (; logical_sector < last_sector;
4219 logical_sector += STRIPE_SECTORS) {
4220 DEFINE_WAIT(w);
4221 int d;
4222 again:
4223 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4224 prepare_to_wait(&conf->wait_for_overlap, &w,
4225 TASK_UNINTERRUPTIBLE);
NeilBrownf8dfcff2013-03-12 12:18:06 +11004226 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4227 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4228 release_stripe(sh);
4229 schedule();
4230 goto again;
4231 }
4232 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
Shaohua Li620125f2012-10-11 13:49:05 +11004233 spin_lock_irq(&sh->stripe_lock);
4234 for (d = 0; d < conf->raid_disks; d++) {
4235 if (d == sh->pd_idx || d == sh->qd_idx)
4236 continue;
4237 if (sh->dev[d].towrite || sh->dev[d].toread) {
4238 set_bit(R5_Overlap, &sh->dev[d].flags);
4239 spin_unlock_irq(&sh->stripe_lock);
4240 release_stripe(sh);
4241 schedule();
4242 goto again;
4243 }
4244 }
NeilBrownf8dfcff2013-03-12 12:18:06 +11004245 set_bit(STRIPE_DISCARD, &sh->state);
Shaohua Li620125f2012-10-11 13:49:05 +11004246 finish_wait(&conf->wait_for_overlap, &w);
4247 for (d = 0; d < conf->raid_disks; d++) {
4248 if (d == sh->pd_idx || d == sh->qd_idx)
4249 continue;
4250 sh->dev[d].towrite = bi;
4251 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4252 raid5_inc_bi_active_stripes(bi);
4253 }
4254 spin_unlock_irq(&sh->stripe_lock);
4255 if (conf->mddev->bitmap) {
4256 for (d = 0;
4257 d < conf->raid_disks - conf->max_degraded;
4258 d++)
4259 bitmap_startwrite(mddev->bitmap,
4260 sh->sector,
4261 STRIPE_SECTORS,
4262 0);
4263 sh->bm_seq = conf->seq_flush + 1;
4264 set_bit(STRIPE_BIT_DELAY, &sh->state);
4265 }
4266
4267 set_bit(STRIPE_HANDLE, &sh->state);
4268 clear_bit(STRIPE_DELAYED, &sh->state);
4269 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4270 atomic_inc(&conf->preread_active_stripes);
4271 release_stripe_plug(mddev, sh);
4272 }
4273
4274 remaining = raid5_dec_bi_active_stripes(bi);
4275 if (remaining == 0) {
4276 md_write_end(mddev);
4277 bio_endio(bi, 0);
4278 }
4279}
4280
Linus Torvaldsb4fdcb02011-11-04 17:06:58 -07004281static void make_request(struct mddev *mddev, struct bio * bi)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004282{
NeilBrownd1688a62011-10-11 16:49:52 +11004283 struct r5conf *conf = mddev->private;
NeilBrown911d4ee2009-03-31 14:39:38 +11004284 int dd_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004285 sector_t new_sector;
4286 sector_t logical_sector, last_sector;
4287 struct stripe_head *sh;
Jens Axboea3623572005-11-01 09:26:16 +01004288 const int rw = bio_data_dir(bi);
NeilBrown49077322010-03-25 16:20:56 +11004289 int remaining;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004290
Tejun Heoe9c74692010-09-03 11:56:18 +02004291 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4292 md_flush_request(mddev, bi);
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +02004293 return;
NeilBrowne5dcdd82005-09-09 16:23:41 -07004294 }
4295
NeilBrown3d310eb2005-06-21 17:17:26 -07004296 md_write_start(mddev, bi);
NeilBrown06d91a52005-06-21 17:17:12 -07004297
NeilBrown802ba062006-12-13 00:34:13 -08004298 if (rw == READ &&
Raz Ben-Jehuda(caro)52488612006-12-10 02:20:48 -08004299 mddev->reshape_position == MaxSector &&
NeilBrown21a52c62010-04-01 15:02:13 +11004300 chunk_aligned_read(mddev,bi))
Christoph Hellwig5a7bbad2011-09-12 12:12:01 +02004301 return;
Raz Ben-Jehuda(caro)52488612006-12-10 02:20:48 -08004302
Shaohua Li620125f2012-10-11 13:49:05 +11004303 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4304 make_discard_request(mddev, bi);
4305 return;
4306 }
4307
Linus Torvalds1da177e2005-04-16 15:20:36 -07004308 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
Kent Overstreetf73a1c72012-09-25 15:05:12 -07004309 last_sector = bio_end_sector(bi);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004310 bi->bi_next = NULL;
4311 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
NeilBrown06d91a52005-06-21 17:17:12 -07004312
Linus Torvalds1da177e2005-04-16 15:20:36 -07004313 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4314 DEFINE_WAIT(w);
NeilBrownb5663ba2009-03-31 14:39:38 +11004315 int previous;
NeilBrownb578d552006-03-27 01:18:12 -08004316
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004317 retry:
NeilBrownb5663ba2009-03-31 14:39:38 +11004318 previous = 0;
NeilBrownb578d552006-03-27 01:18:12 -08004319 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
NeilBrownb0f9ec02009-03-31 15:27:18 +11004320 if (unlikely(conf->reshape_progress != MaxSector)) {
NeilBrownfef9c612009-03-31 15:16:46 +11004321 /* spinlock is needed as reshape_progress may be
NeilBrowndf8e7f72006-03-27 01:18:15 -08004322 * 64bit on a 32bit platform, and so it might be
4323 * possible to see a half-updated value
Jesper Juhlaeb878b2011-04-10 18:06:17 +02004324 * Of course reshape_progress could change after
NeilBrowndf8e7f72006-03-27 01:18:15 -08004325 * the lock is dropped, so once we get a reference
4326 * to the stripe that we think it is, we will have
4327 * to check again.
4328 */
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004329 spin_lock_irq(&conf->device_lock);
NeilBrown2c810cd2012-05-21 09:27:00 +10004330 if (mddev->reshape_backwards
NeilBrownfef9c612009-03-31 15:16:46 +11004331 ? logical_sector < conf->reshape_progress
4332 : logical_sector >= conf->reshape_progress) {
NeilBrownb5663ba2009-03-31 14:39:38 +11004333 previous = 1;
4334 } else {
NeilBrown2c810cd2012-05-21 09:27:00 +10004335 if (mddev->reshape_backwards
NeilBrownfef9c612009-03-31 15:16:46 +11004336 ? logical_sector < conf->reshape_safe
4337 : logical_sector >= conf->reshape_safe) {
NeilBrownb578d552006-03-27 01:18:12 -08004338 spin_unlock_irq(&conf->device_lock);
4339 schedule();
4340 goto retry;
4341 }
4342 }
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004343 spin_unlock_irq(&conf->device_lock);
4344 }
NeilBrown16a53ec2006-06-26 00:27:38 -07004345
NeilBrown112bf892009-03-31 14:39:38 +11004346 new_sector = raid5_compute_sector(conf, logical_sector,
4347 previous,
NeilBrown911d4ee2009-03-31 14:39:38 +11004348 &dd_idx, NULL);
NeilBrown0c55e022010-05-03 14:09:02 +10004349 pr_debug("raid456: make_request, sector %llu logical %llu\n",
Linus Torvalds1da177e2005-04-16 15:20:36 -07004350 (unsigned long long)new_sector,
4351 (unsigned long long)logical_sector);
4352
NeilBrownb5663ba2009-03-31 14:39:38 +11004353 sh = get_active_stripe(conf, new_sector, previous,
NeilBrowna8c906c2009-06-09 14:39:59 +10004354 (bi->bi_rw&RWA_MASK), 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004355 if (sh) {
NeilBrownb0f9ec02009-03-31 15:27:18 +11004356 if (unlikely(previous)) {
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004357 /* expansion might have moved on while waiting for a
NeilBrowndf8e7f72006-03-27 01:18:15 -08004358 * stripe, so we must do the range check again.
4359 * Expansion could still move past after this
4360 * test, but as we are holding a reference to
4361 * 'sh', we know that if that happens,
4362 * STRIPE_EXPANDING will get set and the expansion
4363 * won't proceed until we finish with the stripe.
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004364 */
4365 int must_retry = 0;
4366 spin_lock_irq(&conf->device_lock);
NeilBrown2c810cd2012-05-21 09:27:00 +10004367 if (mddev->reshape_backwards
NeilBrownb0f9ec02009-03-31 15:27:18 +11004368 ? logical_sector >= conf->reshape_progress
4369 : logical_sector < conf->reshape_progress)
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004370 /* mismatch, need to try again */
4371 must_retry = 1;
4372 spin_unlock_irq(&conf->device_lock);
4373 if (must_retry) {
4374 release_stripe(sh);
Dan Williams7a3ab902009-06-16 16:00:33 -07004375 schedule();
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004376 goto retry;
4377 }
4378 }
NeilBrowne62e58a2009-07-01 13:15:35 +10004379
Namhyung Kimffd96e32011-07-18 17:38:51 +10004380 if (rw == WRITE &&
NeilBrowna5c308d2009-07-01 13:15:35 +10004381 logical_sector >= mddev->suspend_lo &&
NeilBrowne464eaf2006-03-27 01:18:14 -08004382 logical_sector < mddev->suspend_hi) {
4383 release_stripe(sh);
NeilBrowne62e58a2009-07-01 13:15:35 +10004384 /* As the suspend_* range is controlled by
4385 * userspace, we want an interruptible
4386 * wait.
4387 */
4388 flush_signals(current);
4389 prepare_to_wait(&conf->wait_for_overlap,
4390 &w, TASK_INTERRUPTIBLE);
4391 if (logical_sector >= mddev->suspend_lo &&
4392 logical_sector < mddev->suspend_hi)
4393 schedule();
NeilBrowne464eaf2006-03-27 01:18:14 -08004394 goto retry;
4395 }
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004396
4397 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
Namhyung Kimffd96e32011-07-18 17:38:51 +10004398 !add_stripe_bio(sh, bi, dd_idx, rw)) {
NeilBrown7ecaa1e2006-03-27 01:18:08 -08004399 /* Stripe is busy expanding or
4400 * add failed due to overlap. Flush everything
Linus Torvalds1da177e2005-04-16 15:20:36 -07004401 * and wait a while
4402 */
NeilBrown482c0832011-04-18 18:25:42 +10004403 md_wakeup_thread(mddev->thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004404 release_stripe(sh);
4405 schedule();
4406 goto retry;
4407 }
4408 finish_wait(&conf->wait_for_overlap, &w);
NeilBrown6ed30032008-02-06 01:40:00 -08004409 set_bit(STRIPE_HANDLE, &sh->state);
4410 clear_bit(STRIPE_DELAYED, &sh->state);
NeilBrowna852d7b2012-09-19 12:48:30 +10004411 if ((bi->bi_rw & REQ_SYNC) &&
NeilBrown729a1862009-12-14 12:49:50 +11004412 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4413 atomic_inc(&conf->preread_active_stripes);
Shaohua Li8811b592012-08-02 08:33:00 +10004414 release_stripe_plug(mddev, sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004415 } else {
4416 /* cannot get stripe for read-ahead, just give-up */
4417 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4418 finish_wait(&conf->wait_for_overlap, &w);
4419 break;
4420 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004421 }
NeilBrown7c13edc2011-04-18 18:25:43 +10004422
Shaohua Lie7836bd62012-07-19 16:01:31 +10004423 remaining = raid5_dec_bi_active_stripes(bi);
NeilBrownf6344752006-03-27 01:18:17 -08004424 if (remaining == 0) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004425
NeilBrown16a53ec2006-06-26 00:27:38 -07004426 if ( rw == WRITE )
Linus Torvalds1da177e2005-04-16 15:20:36 -07004427 md_write_end(mddev);
NeilBrown6712ecf2007-09-27 12:47:43 +02004428
Linus Torvalds0a82a8d2013-04-18 09:00:26 -07004429 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4430 bi, 0);
Neil Brown0e13fe232008-06-28 08:31:20 +10004431 bio_endio(bi, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004432 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004433}
4434
NeilBrownfd01b882011-10-11 16:47:53 +11004435static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
Dan Williamsb522adc2009-03-31 15:00:31 +11004436
NeilBrownfd01b882011-10-11 16:47:53 +11004437static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004438{
NeilBrown52c03292006-06-26 00:27:43 -07004439 /* reshaping is quite different to recovery/resync so it is
4440 * handled quite separately ... here.
4441 *
4442 * On each call to sync_request, we gather one chunk worth of
4443 * destination stripes and flag them as expanding.
4444 * Then we find all the source stripes and request reads.
4445 * As the reads complete, handle_stripe will copy the data
4446 * into the destination stripe and release that stripe.
4447 */
NeilBrownd1688a62011-10-11 16:49:52 +11004448 struct r5conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004449 struct stripe_head *sh;
NeilBrownccfcc3c2006-03-27 01:18:09 -08004450 sector_t first_sector, last_sector;
NeilBrownf4168852007-02-28 20:11:53 -08004451 int raid_disks = conf->previous_raid_disks;
4452 int data_disks = raid_disks - conf->max_degraded;
4453 int new_data_disks = conf->raid_disks - conf->max_degraded;
NeilBrown52c03292006-06-26 00:27:43 -07004454 int i;
4455 int dd_idx;
NeilBrownc8f517c2009-03-31 15:28:40 +11004456 sector_t writepos, readpos, safepos;
NeilBrownec32a2b2009-03-31 15:17:38 +11004457 sector_t stripe_addr;
NeilBrown7a661382009-03-31 15:21:40 +11004458 int reshape_sectors;
NeilBrownab69ae12009-03-31 15:26:47 +11004459 struct list_head stripes;
NeilBrown52c03292006-06-26 00:27:43 -07004460
NeilBrownfef9c612009-03-31 15:16:46 +11004461 if (sector_nr == 0) {
4462 /* If restarting in the middle, skip the initial sectors */
NeilBrown2c810cd2012-05-21 09:27:00 +10004463 if (mddev->reshape_backwards &&
NeilBrownfef9c612009-03-31 15:16:46 +11004464 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4465 sector_nr = raid5_size(mddev, 0, 0)
4466 - conf->reshape_progress;
NeilBrown2c810cd2012-05-21 09:27:00 +10004467 } else if (!mddev->reshape_backwards &&
NeilBrownfef9c612009-03-31 15:16:46 +11004468 conf->reshape_progress > 0)
4469 sector_nr = conf->reshape_progress;
NeilBrownf4168852007-02-28 20:11:53 -08004470 sector_div(sector_nr, new_data_disks);
NeilBrownfef9c612009-03-31 15:16:46 +11004471 if (sector_nr) {
NeilBrown8dee7212009-11-06 14:59:29 +11004472 mddev->curr_resync_completed = sector_nr;
4473 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
NeilBrownfef9c612009-03-31 15:16:46 +11004474 *skipped = 1;
4475 return sector_nr;
4476 }
NeilBrown52c03292006-06-26 00:27:43 -07004477 }
4478
NeilBrown7a661382009-03-31 15:21:40 +11004479 /* We need to process a full chunk at a time.
4480 * If old and new chunk sizes differ, we need to process the
4481 * largest of these
4482 */
Andre Noll664e7c42009-06-18 08:45:27 +10004483 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4484 reshape_sectors = mddev->new_chunk_sectors;
NeilBrown7a661382009-03-31 15:21:40 +11004485 else
Andre Noll9d8f0362009-06-18 08:45:01 +10004486 reshape_sectors = mddev->chunk_sectors;
NeilBrown7a661382009-03-31 15:21:40 +11004487
NeilBrownb5254dd2012-05-21 09:27:01 +10004488 /* We update the metadata at least every 10 seconds, or when
4489 * the data about to be copied would over-write the source of
4490 * the data at the front of the range. i.e. one new_stripe
4491 * along from reshape_progress new_maps to after where
4492 * reshape_safe old_maps to
NeilBrown52c03292006-06-26 00:27:43 -07004493 */
NeilBrownfef9c612009-03-31 15:16:46 +11004494 writepos = conf->reshape_progress;
NeilBrownf4168852007-02-28 20:11:53 -08004495 sector_div(writepos, new_data_disks);
NeilBrownc8f517c2009-03-31 15:28:40 +11004496 readpos = conf->reshape_progress;
4497 sector_div(readpos, data_disks);
NeilBrownfef9c612009-03-31 15:16:46 +11004498 safepos = conf->reshape_safe;
NeilBrownf4168852007-02-28 20:11:53 -08004499 sector_div(safepos, data_disks);
NeilBrown2c810cd2012-05-21 09:27:00 +10004500 if (mddev->reshape_backwards) {
NeilBrowned37d832009-05-27 21:39:05 +10004501 writepos -= min_t(sector_t, reshape_sectors, writepos);
NeilBrownc8f517c2009-03-31 15:28:40 +11004502 readpos += reshape_sectors;
NeilBrown7a661382009-03-31 15:21:40 +11004503 safepos += reshape_sectors;
NeilBrownfef9c612009-03-31 15:16:46 +11004504 } else {
NeilBrown7a661382009-03-31 15:21:40 +11004505 writepos += reshape_sectors;
NeilBrowned37d832009-05-27 21:39:05 +10004506 readpos -= min_t(sector_t, reshape_sectors, readpos);
4507 safepos -= min_t(sector_t, reshape_sectors, safepos);
NeilBrownfef9c612009-03-31 15:16:46 +11004508 }
NeilBrown52c03292006-06-26 00:27:43 -07004509
NeilBrownb5254dd2012-05-21 09:27:01 +10004510 /* Having calculated the 'writepos' possibly use it
4511 * to set 'stripe_addr' which is where we will write to.
4512 */
4513 if (mddev->reshape_backwards) {
4514 BUG_ON(conf->reshape_progress == 0);
4515 stripe_addr = writepos;
4516 BUG_ON((mddev->dev_sectors &
4517 ~((sector_t)reshape_sectors - 1))
4518 - reshape_sectors - stripe_addr
4519 != sector_nr);
4520 } else {
4521 BUG_ON(writepos != sector_nr + reshape_sectors);
4522 stripe_addr = sector_nr;
4523 }
4524
NeilBrownc8f517c2009-03-31 15:28:40 +11004525 /* 'writepos' is the most advanced device address we might write.
4526 * 'readpos' is the least advanced device address we might read.
4527 * 'safepos' is the least address recorded in the metadata as having
4528 * been reshaped.
NeilBrownb5254dd2012-05-21 09:27:01 +10004529 * If there is a min_offset_diff, these are adjusted either by
4530 * increasing the safepos/readpos if diff is negative, or
4531 * increasing writepos if diff is positive.
4532 * If 'readpos' is then behind 'writepos', there is no way that we can
NeilBrownc8f517c2009-03-31 15:28:40 +11004533 * ensure safety in the face of a crash - that must be done by userspace
4534 * making a backup of the data. So in that case there is no particular
4535 * rush to update metadata.
4536 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4537 * update the metadata to advance 'safepos' to match 'readpos' so that
4538 * we can be safe in the event of a crash.
4539 * So we insist on updating metadata if safepos is behind writepos and
4540 * readpos is beyond writepos.
4541 * In any case, update the metadata every 10 seconds.
4542 * Maybe that number should be configurable, but I'm not sure it is
4543 * worth it.... maybe it could be a multiple of safemode_delay???
4544 */
NeilBrownb5254dd2012-05-21 09:27:01 +10004545 if (conf->min_offset_diff < 0) {
4546 safepos += -conf->min_offset_diff;
4547 readpos += -conf->min_offset_diff;
4548 } else
4549 writepos += conf->min_offset_diff;
4550
NeilBrown2c810cd2012-05-21 09:27:00 +10004551 if ((mddev->reshape_backwards
NeilBrownc8f517c2009-03-31 15:28:40 +11004552 ? (safepos > writepos && readpos < writepos)
4553 : (safepos < writepos && readpos > writepos)) ||
4554 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
NeilBrown52c03292006-06-26 00:27:43 -07004555 /* Cannot proceed until we've updated the superblock... */
4556 wait_event(conf->wait_for_overlap,
4557 atomic_read(&conf->reshape_stripes)==0);
NeilBrownfef9c612009-03-31 15:16:46 +11004558 mddev->reshape_position = conf->reshape_progress;
NeilBrown75d3da42011-01-14 09:14:34 +11004559 mddev->curr_resync_completed = sector_nr;
NeilBrownc8f517c2009-03-31 15:28:40 +11004560 conf->reshape_checkpoint = jiffies;
NeilBrown850b2b42006-10-03 01:15:46 -07004561 set_bit(MD_CHANGE_DEVS, &mddev->flags);
NeilBrown52c03292006-06-26 00:27:43 -07004562 md_wakeup_thread(mddev->thread);
NeilBrown850b2b42006-10-03 01:15:46 -07004563 wait_event(mddev->sb_wait, mddev->flags == 0 ||
NeilBrown52c03292006-06-26 00:27:43 -07004564 kthread_should_stop());
4565 spin_lock_irq(&conf->device_lock);
NeilBrownfef9c612009-03-31 15:16:46 +11004566 conf->reshape_safe = mddev->reshape_position;
NeilBrown52c03292006-06-26 00:27:43 -07004567 spin_unlock_irq(&conf->device_lock);
4568 wake_up(&conf->wait_for_overlap);
NeilBrownacb180b2009-04-14 16:28:34 +10004569 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
NeilBrown52c03292006-06-26 00:27:43 -07004570 }
4571
NeilBrownab69ae12009-03-31 15:26:47 +11004572 INIT_LIST_HEAD(&stripes);
NeilBrown7a661382009-03-31 15:21:40 +11004573 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
NeilBrown52c03292006-06-26 00:27:43 -07004574 int j;
NeilBrowna9f326e2009-09-23 18:06:41 +10004575 int skipped_disk = 0;
NeilBrowna8c906c2009-06-09 14:39:59 +10004576 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
NeilBrown52c03292006-06-26 00:27:43 -07004577 set_bit(STRIPE_EXPANDING, &sh->state);
4578 atomic_inc(&conf->reshape_stripes);
4579 /* If any of this stripe is beyond the end of the old
4580 * array, then we need to zero those blocks
4581 */
4582 for (j=sh->disks; j--;) {
4583 sector_t s;
4584 if (j == sh->pd_idx)
4585 continue;
NeilBrownf4168852007-02-28 20:11:53 -08004586 if (conf->level == 6 &&
NeilBrownd0dabf72009-03-31 14:39:38 +11004587 j == sh->qd_idx)
NeilBrownf4168852007-02-28 20:11:53 -08004588 continue;
NeilBrown784052e2009-03-31 15:19:07 +11004589 s = compute_blocknr(sh, j, 0);
Dan Williamsb522adc2009-03-31 15:00:31 +11004590 if (s < raid5_size(mddev, 0, 0)) {
NeilBrowna9f326e2009-09-23 18:06:41 +10004591 skipped_disk = 1;
NeilBrown52c03292006-06-26 00:27:43 -07004592 continue;
4593 }
4594 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4595 set_bit(R5_Expanded, &sh->dev[j].flags);
4596 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4597 }
NeilBrowna9f326e2009-09-23 18:06:41 +10004598 if (!skipped_disk) {
NeilBrown52c03292006-06-26 00:27:43 -07004599 set_bit(STRIPE_EXPAND_READY, &sh->state);
4600 set_bit(STRIPE_HANDLE, &sh->state);
4601 }
NeilBrownab69ae12009-03-31 15:26:47 +11004602 list_add(&sh->lru, &stripes);
NeilBrown52c03292006-06-26 00:27:43 -07004603 }
4604 spin_lock_irq(&conf->device_lock);
NeilBrown2c810cd2012-05-21 09:27:00 +10004605 if (mddev->reshape_backwards)
NeilBrown7a661382009-03-31 15:21:40 +11004606 conf->reshape_progress -= reshape_sectors * new_data_disks;
NeilBrownfef9c612009-03-31 15:16:46 +11004607 else
NeilBrown7a661382009-03-31 15:21:40 +11004608 conf->reshape_progress += reshape_sectors * new_data_disks;
NeilBrown52c03292006-06-26 00:27:43 -07004609 spin_unlock_irq(&conf->device_lock);
4610 /* Ok, those stripe are ready. We can start scheduling
4611 * reads on the source stripes.
4612 * The source stripes are determined by mapping the first and last
4613 * block on the destination stripes.
4614 */
NeilBrown52c03292006-06-26 00:27:43 -07004615 first_sector =
NeilBrownec32a2b2009-03-31 15:17:38 +11004616 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
NeilBrown911d4ee2009-03-31 14:39:38 +11004617 1, &dd_idx, NULL);
NeilBrown52c03292006-06-26 00:27:43 -07004618 last_sector =
NeilBrown0e6e0272009-06-09 16:32:22 +10004619 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
Andre Noll09c9e5f2009-06-18 08:45:55 +10004620 * new_data_disks - 1),
NeilBrown911d4ee2009-03-31 14:39:38 +11004621 1, &dd_idx, NULL);
Andre Noll58c0fed2009-03-31 14:33:13 +11004622 if (last_sector >= mddev->dev_sectors)
4623 last_sector = mddev->dev_sectors - 1;
NeilBrown52c03292006-06-26 00:27:43 -07004624 while (first_sector <= last_sector) {
NeilBrowna8c906c2009-06-09 14:39:59 +10004625 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
NeilBrown52c03292006-06-26 00:27:43 -07004626 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4627 set_bit(STRIPE_HANDLE, &sh->state);
4628 release_stripe(sh);
4629 first_sector += STRIPE_SECTORS;
4630 }
NeilBrownab69ae12009-03-31 15:26:47 +11004631 /* Now that the sources are clearly marked, we can release
4632 * the destination stripes
4633 */
4634 while (!list_empty(&stripes)) {
4635 sh = list_entry(stripes.next, struct stripe_head, lru);
4636 list_del_init(&sh->lru);
4637 release_stripe(sh);
4638 }
NeilBrownc6207272008-02-06 01:39:52 -08004639 /* If this takes us to the resync_max point where we have to pause,
4640 * then we need to write out the superblock.
4641 */
NeilBrown7a661382009-03-31 15:21:40 +11004642 sector_nr += reshape_sectors;
NeilBrownc03f6a12009-04-17 11:06:30 +10004643 if ((sector_nr - mddev->curr_resync_completed) * 2
4644 >= mddev->resync_max - mddev->curr_resync_completed) {
NeilBrownc6207272008-02-06 01:39:52 -08004645 /* Cannot proceed until we've updated the superblock... */
4646 wait_event(conf->wait_for_overlap,
4647 atomic_read(&conf->reshape_stripes) == 0);
NeilBrownfef9c612009-03-31 15:16:46 +11004648 mddev->reshape_position = conf->reshape_progress;
NeilBrown75d3da42011-01-14 09:14:34 +11004649 mddev->curr_resync_completed = sector_nr;
NeilBrownc8f517c2009-03-31 15:28:40 +11004650 conf->reshape_checkpoint = jiffies;
NeilBrownc6207272008-02-06 01:39:52 -08004651 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4652 md_wakeup_thread(mddev->thread);
4653 wait_event(mddev->sb_wait,
4654 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4655 || kthread_should_stop());
4656 spin_lock_irq(&conf->device_lock);
NeilBrownfef9c612009-03-31 15:16:46 +11004657 conf->reshape_safe = mddev->reshape_position;
NeilBrownc6207272008-02-06 01:39:52 -08004658 spin_unlock_irq(&conf->device_lock);
4659 wake_up(&conf->wait_for_overlap);
NeilBrownacb180b2009-04-14 16:28:34 +10004660 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
NeilBrownc6207272008-02-06 01:39:52 -08004661 }
NeilBrown7a661382009-03-31 15:21:40 +11004662 return reshape_sectors;
NeilBrown52c03292006-06-26 00:27:43 -07004663}
4664
4665/* FIXME go_faster isn't used */
NeilBrownfd01b882011-10-11 16:47:53 +11004666static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
NeilBrown52c03292006-06-26 00:27:43 -07004667{
NeilBrownd1688a62011-10-11 16:49:52 +11004668 struct r5conf *conf = mddev->private;
NeilBrown52c03292006-06-26 00:27:43 -07004669 struct stripe_head *sh;
Andre Noll58c0fed2009-03-31 14:33:13 +11004670 sector_t max_sector = mddev->dev_sectors;
NeilBrown57dab0b2010-10-19 10:03:39 +11004671 sector_t sync_blocks;
NeilBrown16a53ec2006-06-26 00:27:38 -07004672 int still_degraded = 0;
4673 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004674
NeilBrown72626682005-09-09 16:23:54 -07004675 if (sector_nr >= max_sector) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004676 /* just being told to finish up .. nothing much to do */
NeilBrowncea9c222009-03-31 15:15:05 +11004677
NeilBrown29269552006-03-27 01:18:10 -08004678 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4679 end_reshape(conf);
4680 return 0;
4681 }
NeilBrown72626682005-09-09 16:23:54 -07004682
4683 if (mddev->curr_resync < max_sector) /* aborted */
4684 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4685 &sync_blocks, 1);
NeilBrown16a53ec2006-06-26 00:27:38 -07004686 else /* completed sync */
NeilBrown72626682005-09-09 16:23:54 -07004687 conf->fullsync = 0;
4688 bitmap_close_sync(mddev->bitmap);
4689
Linus Torvalds1da177e2005-04-16 15:20:36 -07004690 return 0;
4691 }
NeilBrownccfcc3c2006-03-27 01:18:09 -08004692
NeilBrown64bd6602009-08-03 10:59:58 +10004693 /* Allow raid5_quiesce to complete */
4694 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4695
NeilBrown52c03292006-06-26 00:27:43 -07004696 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4697 return reshape_request(mddev, sector_nr, skipped);
NeilBrownf6705572006-03-27 01:18:11 -08004698
NeilBrownc6207272008-02-06 01:39:52 -08004699 /* No need to check resync_max as we never do more than one
4700 * stripe, and as resync_max will always be on a chunk boundary,
4701 * if the check in md_do_sync didn't fire, there is no chance
4702 * of overstepping resync_max here
4703 */
4704
NeilBrown16a53ec2006-06-26 00:27:38 -07004705 /* if there is too many failed drives and we are trying
Linus Torvalds1da177e2005-04-16 15:20:36 -07004706 * to resync, then assert that we are finished, because there is
4707 * nothing we can do.
4708 */
NeilBrown3285edf2006-06-26 00:27:55 -07004709 if (mddev->degraded >= conf->max_degraded &&
NeilBrown16a53ec2006-06-26 00:27:38 -07004710 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
Andre Noll58c0fed2009-03-31 14:33:13 +11004711 sector_t rv = mddev->dev_sectors - sector_nr;
NeilBrown57afd892005-06-21 17:17:13 -07004712 *skipped = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004713 return rv;
4714 }
majianpeng6f608042013-04-24 11:42:41 +10004715 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4716 !conf->fullsync &&
4717 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4718 sync_blocks >= STRIPE_SECTORS) {
NeilBrown72626682005-09-09 16:23:54 -07004719 /* we can skip this block, and probably more */
4720 sync_blocks /= STRIPE_SECTORS;
4721 *skipped = 1;
4722 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4723 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004724
NeilBrownb47490c2008-02-06 01:39:50 -08004725 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4726
NeilBrowna8c906c2009-06-09 14:39:59 +10004727 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004728 if (sh == NULL) {
NeilBrowna8c906c2009-06-09 14:39:59 +10004729 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004730 /* make sure we don't swamp the stripe cache if someone else
NeilBrown16a53ec2006-06-26 00:27:38 -07004731 * is trying to get access
Linus Torvalds1da177e2005-04-16 15:20:36 -07004732 */
Nishanth Aravamudan66c006a2005-11-07 01:01:17 -08004733 schedule_timeout_uninterruptible(1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004734 }
NeilBrown16a53ec2006-06-26 00:27:38 -07004735 /* Need to check if array will still be degraded after recovery/resync
4736 * We don't need to check the 'failed' flag as when that gets set,
4737 * recovery aborts.
4738 */
NeilBrownf001a702009-06-09 14:30:31 +10004739 for (i = 0; i < conf->raid_disks; i++)
NeilBrown16a53ec2006-06-26 00:27:38 -07004740 if (conf->disks[i].rdev == NULL)
4741 still_degraded = 1;
4742
4743 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4744
NeilBrown83206d62011-07-26 11:19:49 +10004745 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004746
NeilBrown14425772009-10-16 15:55:25 +11004747 handle_stripe(sh);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004748 release_stripe(sh);
4749
4750 return STRIPE_SECTORS;
4751}
4752
NeilBrownd1688a62011-10-11 16:49:52 +11004753static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004754{
4755 /* We may not be able to submit a whole bio at once as there
4756 * may not be enough stripe_heads available.
4757 * We cannot pre-allocate enough stripe_heads as we may need
4758 * more than exist in the cache (if we allow ever large chunks).
4759 * So we do one stripe head at a time and record in
4760 * ->bi_hw_segments how many have been done.
4761 *
4762 * We *know* that this entire raid_bio is in one chunk, so
4763 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4764 */
4765 struct stripe_head *sh;
NeilBrown911d4ee2009-03-31 14:39:38 +11004766 int dd_idx;
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004767 sector_t sector, logical_sector, last_sector;
4768 int scnt = 0;
4769 int remaining;
4770 int handled = 0;
4771
4772 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
NeilBrown112bf892009-03-31 14:39:38 +11004773 sector = raid5_compute_sector(conf, logical_sector,
NeilBrown911d4ee2009-03-31 14:39:38 +11004774 0, &dd_idx, NULL);
Kent Overstreetf73a1c72012-09-25 15:05:12 -07004775 last_sector = bio_end_sector(raid_bio);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004776
4777 for (; logical_sector < last_sector;
Neil Brown387bb172007-02-08 14:20:29 -08004778 logical_sector += STRIPE_SECTORS,
4779 sector += STRIPE_SECTORS,
4780 scnt++) {
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004781
Shaohua Lie7836bd62012-07-19 16:01:31 +10004782 if (scnt < raid5_bi_processed_stripes(raid_bio))
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004783 /* already done this stripe */
4784 continue;
4785
NeilBrowna8c906c2009-06-09 14:39:59 +10004786 sh = get_active_stripe(conf, sector, 0, 1, 0);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004787
4788 if (!sh) {
4789 /* failed to get a stripe - must wait */
Shaohua Lie7836bd62012-07-19 16:01:31 +10004790 raid5_set_bi_processed_stripes(raid_bio, scnt);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004791 conf->retry_read_aligned = raid_bio;
4792 return handled;
4793 }
4794
Neil Brown387bb172007-02-08 14:20:29 -08004795 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4796 release_stripe(sh);
Shaohua Lie7836bd62012-07-19 16:01:31 +10004797 raid5_set_bi_processed_stripes(raid_bio, scnt);
Neil Brown387bb172007-02-08 14:20:29 -08004798 conf->retry_read_aligned = raid_bio;
4799 return handled;
4800 }
4801
majianpeng3f9e7c12012-07-31 10:04:21 +10004802 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
Dan Williams36d1c642009-07-14 11:48:22 -07004803 handle_stripe(sh);
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004804 release_stripe(sh);
4805 handled++;
4806 }
Shaohua Lie7836bd62012-07-19 16:01:31 +10004807 remaining = raid5_dec_bi_active_stripes(raid_bio);
Linus Torvalds0a82a8d2013-04-18 09:00:26 -07004808 if (remaining == 0) {
4809 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
4810 raid_bio, 0);
Neil Brown0e13fe232008-06-28 08:31:20 +10004811 bio_endio(raid_bio, 0);
Linus Torvalds0a82a8d2013-04-18 09:00:26 -07004812 }
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004813 if (atomic_dec_and_test(&conf->active_aligned_reads))
4814 wake_up(&conf->wait_for_stripe);
4815 return handled;
4816}
4817
Shaohua Li46a06402012-08-02 08:33:15 +10004818#define MAX_STRIPE_BATCH 8
4819static int handle_active_stripes(struct r5conf *conf)
4820{
4821 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4822 int i, batch_size = 0;
4823
4824 while (batch_size < MAX_STRIPE_BATCH &&
4825 (sh = __get_priority_stripe(conf)) != NULL)
4826 batch[batch_size++] = sh;
4827
4828 if (batch_size == 0)
4829 return batch_size;
4830 spin_unlock_irq(&conf->device_lock);
4831
4832 for (i = 0; i < batch_size; i++)
4833 handle_stripe(batch[i]);
4834
4835 cond_resched();
4836
4837 spin_lock_irq(&conf->device_lock);
4838 for (i = 0; i < batch_size; i++)
4839 __release_stripe(conf, batch[i]);
4840 return batch_size;
4841}
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004842
Linus Torvalds1da177e2005-04-16 15:20:36 -07004843/*
4844 * This is our raid5 kernel thread.
4845 *
4846 * We scan the hash table for stripes which can be handled now.
4847 * During the scan, completed stripes are saved for us by the interrupt
4848 * handler, so that they will not have to wait for our next wakeup.
4849 */
Shaohua Li4ed87312012-10-11 13:34:00 +11004850static void raid5d(struct md_thread *thread)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004851{
Shaohua Li4ed87312012-10-11 13:34:00 +11004852 struct mddev *mddev = thread->mddev;
NeilBrownd1688a62011-10-11 16:49:52 +11004853 struct r5conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004854 int handled;
NeilBrowne1dfa0a2011-04-18 18:25:41 +10004855 struct blk_plug plug;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004856
Dan Williams45b42332007-07-09 11:56:43 -07004857 pr_debug("+++ raid5d active\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004858
4859 md_check_recovery(mddev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004860
NeilBrowne1dfa0a2011-04-18 18:25:41 +10004861 blk_start_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004862 handled = 0;
4863 spin_lock_irq(&conf->device_lock);
4864 while (1) {
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004865 struct bio *bio;
Shaohua Li46a06402012-08-02 08:33:15 +10004866 int batch_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004867
NeilBrown0021b7b2012-07-31 09:08:14 +02004868 if (
NeilBrown7c13edc2011-04-18 18:25:43 +10004869 !list_empty(&conf->bitmap_list)) {
4870 /* Now is a good time to flush some bitmap updates */
4871 conf->seq_flush++;
NeilBrown700e4322005-11-28 13:44:10 -08004872 spin_unlock_irq(&conf->device_lock);
NeilBrown72626682005-09-09 16:23:54 -07004873 bitmap_unplug(mddev->bitmap);
NeilBrown700e4322005-11-28 13:44:10 -08004874 spin_lock_irq(&conf->device_lock);
NeilBrown7c13edc2011-04-18 18:25:43 +10004875 conf->seq_write = conf->seq_flush;
NeilBrown72626682005-09-09 16:23:54 -07004876 activate_bit_delay(conf);
4877 }
NeilBrown0021b7b2012-07-31 09:08:14 +02004878 raid5_activate_delayed(conf);
NeilBrown72626682005-09-09 16:23:54 -07004879
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08004880 while ((bio = remove_bio_from_retry(conf))) {
4881 int ok;
4882 spin_unlock_irq(&conf->device_lock);
4883 ok = retry_aligned_read(conf, bio);
4884 spin_lock_irq(&conf->device_lock);
4885 if (!ok)
4886 break;
4887 handled++;
4888 }
4889
Shaohua Li46a06402012-08-02 08:33:15 +10004890 batch_size = handle_active_stripes(conf);
4891 if (!batch_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004892 break;
Shaohua Li46a06402012-08-02 08:33:15 +10004893 handled += batch_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004894
Shaohua Li46a06402012-08-02 08:33:15 +10004895 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
4896 spin_unlock_irq(&conf->device_lock);
NeilBrownde393cd2011-07-28 11:31:48 +10004897 md_check_recovery(mddev);
Shaohua Li46a06402012-08-02 08:33:15 +10004898 spin_lock_irq(&conf->device_lock);
4899 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004900 }
Dan Williams45b42332007-07-09 11:56:43 -07004901 pr_debug("%d stripes handled\n", handled);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004902
4903 spin_unlock_irq(&conf->device_lock);
4904
Dan Williamsc9f21aa2008-07-23 12:05:51 -07004905 async_tx_issue_pending_all();
NeilBrowne1dfa0a2011-04-18 18:25:41 +10004906 blk_finish_plug(&plug);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004907
Dan Williams45b42332007-07-09 11:56:43 -07004908 pr_debug("--- raid5d inactive\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07004909}
4910
NeilBrown3f294f42005-11-08 21:39:25 -08004911static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004912raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
NeilBrown3f294f42005-11-08 21:39:25 -08004913{
NeilBrownd1688a62011-10-11 16:49:52 +11004914 struct r5conf *conf = mddev->private;
NeilBrown96de1e62005-11-08 21:39:39 -08004915 if (conf)
4916 return sprintf(page, "%d\n", conf->max_nr_stripes);
4917 else
4918 return 0;
NeilBrown3f294f42005-11-08 21:39:25 -08004919}
4920
NeilBrownc41d4ac2010-06-01 19:37:24 +10004921int
NeilBrownfd01b882011-10-11 16:47:53 +11004922raid5_set_cache_size(struct mddev *mddev, int size)
NeilBrownc41d4ac2010-06-01 19:37:24 +10004923{
NeilBrownd1688a62011-10-11 16:49:52 +11004924 struct r5conf *conf = mddev->private;
NeilBrownc41d4ac2010-06-01 19:37:24 +10004925 int err;
4926
4927 if (size <= 16 || size > 32768)
4928 return -EINVAL;
4929 while (size < conf->max_nr_stripes) {
4930 if (drop_one_stripe(conf))
4931 conf->max_nr_stripes--;
4932 else
4933 break;
4934 }
4935 err = md_allow_write(mddev);
4936 if (err)
4937 return err;
4938 while (size > conf->max_nr_stripes) {
4939 if (grow_one_stripe(conf))
4940 conf->max_nr_stripes++;
4941 else break;
4942 }
4943 return 0;
4944}
4945EXPORT_SYMBOL(raid5_set_cache_size);
4946
NeilBrown3f294f42005-11-08 21:39:25 -08004947static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004948raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
NeilBrown3f294f42005-11-08 21:39:25 -08004949{
NeilBrownd1688a62011-10-11 16:49:52 +11004950 struct r5conf *conf = mddev->private;
Dan Williams4ef197d82008-04-28 02:15:54 -07004951 unsigned long new;
Dan Williamsb5470dc2008-06-27 21:44:04 -07004952 int err;
4953
NeilBrown3f294f42005-11-08 21:39:25 -08004954 if (len >= PAGE_SIZE)
4955 return -EINVAL;
NeilBrown96de1e62005-11-08 21:39:39 -08004956 if (!conf)
4957 return -ENODEV;
NeilBrown3f294f42005-11-08 21:39:25 -08004958
Dan Williams4ef197d82008-04-28 02:15:54 -07004959 if (strict_strtoul(page, 10, &new))
NeilBrown3f294f42005-11-08 21:39:25 -08004960 return -EINVAL;
NeilBrownc41d4ac2010-06-01 19:37:24 +10004961 err = raid5_set_cache_size(mddev, new);
Dan Williamsb5470dc2008-06-27 21:44:04 -07004962 if (err)
4963 return err;
NeilBrown3f294f42005-11-08 21:39:25 -08004964 return len;
4965}
NeilBrown007583c2005-11-08 21:39:30 -08004966
NeilBrown96de1e62005-11-08 21:39:39 -08004967static struct md_sysfs_entry
4968raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4969 raid5_show_stripe_cache_size,
4970 raid5_store_stripe_cache_size);
NeilBrown3f294f42005-11-08 21:39:25 -08004971
4972static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004973raid5_show_preread_threshold(struct mddev *mddev, char *page)
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004974{
NeilBrownd1688a62011-10-11 16:49:52 +11004975 struct r5conf *conf = mddev->private;
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004976 if (conf)
4977 return sprintf(page, "%d\n", conf->bypass_threshold);
4978 else
4979 return 0;
4980}
4981
4982static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11004983raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004984{
NeilBrownd1688a62011-10-11 16:49:52 +11004985 struct r5conf *conf = mddev->private;
Dan Williams4ef197d82008-04-28 02:15:54 -07004986 unsigned long new;
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004987 if (len >= PAGE_SIZE)
4988 return -EINVAL;
4989 if (!conf)
4990 return -ENODEV;
4991
Dan Williams4ef197d82008-04-28 02:15:54 -07004992 if (strict_strtoul(page, 10, &new))
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004993 return -EINVAL;
Dan Williams4ef197d82008-04-28 02:15:54 -07004994 if (new > conf->max_nr_stripes)
Dan Williams8b3e6cd2008-04-28 02:15:53 -07004995 return -EINVAL;
4996 conf->bypass_threshold = new;
4997 return len;
4998}
4999
5000static struct md_sysfs_entry
5001raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5002 S_IRUGO | S_IWUSR,
5003 raid5_show_preread_threshold,
5004 raid5_store_preread_threshold);
5005
5006static ssize_t
NeilBrownfd01b882011-10-11 16:47:53 +11005007stripe_cache_active_show(struct mddev *mddev, char *page)
NeilBrown3f294f42005-11-08 21:39:25 -08005008{
NeilBrownd1688a62011-10-11 16:49:52 +11005009 struct r5conf *conf = mddev->private;
NeilBrown96de1e62005-11-08 21:39:39 -08005010 if (conf)
5011 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5012 else
5013 return 0;
NeilBrown3f294f42005-11-08 21:39:25 -08005014}
5015
NeilBrown96de1e62005-11-08 21:39:39 -08005016static struct md_sysfs_entry
5017raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
NeilBrown3f294f42005-11-08 21:39:25 -08005018
NeilBrown007583c2005-11-08 21:39:30 -08005019static struct attribute *raid5_attrs[] = {
NeilBrown3f294f42005-11-08 21:39:25 -08005020 &raid5_stripecache_size.attr,
5021 &raid5_stripecache_active.attr,
Dan Williams8b3e6cd2008-04-28 02:15:53 -07005022 &raid5_preread_bypass_threshold.attr,
NeilBrown3f294f42005-11-08 21:39:25 -08005023 NULL,
5024};
NeilBrown007583c2005-11-08 21:39:30 -08005025static struct attribute_group raid5_attrs_group = {
5026 .name = NULL,
5027 .attrs = raid5_attrs,
NeilBrown3f294f42005-11-08 21:39:25 -08005028};
5029
Dan Williams80c3a6c2009-03-17 18:10:40 -07005030static sector_t
NeilBrownfd01b882011-10-11 16:47:53 +11005031raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
Dan Williams80c3a6c2009-03-17 18:10:40 -07005032{
NeilBrownd1688a62011-10-11 16:49:52 +11005033 struct r5conf *conf = mddev->private;
Dan Williams80c3a6c2009-03-17 18:10:40 -07005034
5035 if (!sectors)
5036 sectors = mddev->dev_sectors;
NeilBrown5e5e3e72009-10-16 16:35:30 +11005037 if (!raid_disks)
NeilBrown7ec05472009-03-31 15:10:36 +11005038 /* size is defined by the smallest of previous and new size */
NeilBrown5e5e3e72009-10-16 16:35:30 +11005039 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
Dan Williams80c3a6c2009-03-17 18:10:40 -07005040
Andre Noll9d8f0362009-06-18 08:45:01 +10005041 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
Andre Noll664e7c42009-06-18 08:45:27 +10005042 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
Dan Williams80c3a6c2009-03-17 18:10:40 -07005043 return sectors * (raid_disks - conf->max_degraded);
5044}
5045
Oleg Nesterov4d4ef862014-02-06 03:42:45 +05305046static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5047{
5048 safe_put_page(percpu->spare_page);
5049 kfree(percpu->scribble);
5050 percpu->spare_page = NULL;
5051 percpu->scribble = NULL;
5052}
5053
5054static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5055{
5056 if (conf->level == 6 && !percpu->spare_page)
5057 percpu->spare_page = alloc_page(GFP_KERNEL);
5058 if (!percpu->scribble)
5059 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5060
5061 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
5062 free_scratch_buffer(conf, percpu);
5063 return -ENOMEM;
5064 }
5065
5066 return 0;
5067}
5068
NeilBrownd1688a62011-10-11 16:49:52 +11005069static void raid5_free_percpu(struct r5conf *conf)
Dan Williams36d1c642009-07-14 11:48:22 -07005070{
Dan Williams36d1c642009-07-14 11:48:22 -07005071 unsigned long cpu;
5072
5073 if (!conf->percpu)
5074 return;
5075
Dan Williams36d1c642009-07-14 11:48:22 -07005076#ifdef CONFIG_HOTPLUG_CPU
5077 unregister_cpu_notifier(&conf->cpu_notify);
5078#endif
Oleg Nesterov4d4ef862014-02-06 03:42:45 +05305079
5080 get_online_cpus();
5081 for_each_possible_cpu(cpu)
5082 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
Dan Williams36d1c642009-07-14 11:48:22 -07005083 put_online_cpus();
5084
5085 free_percpu(conf->percpu);
5086}
5087
NeilBrownd1688a62011-10-11 16:49:52 +11005088static void free_conf(struct r5conf *conf)
Dan Williams95fc17a2009-07-31 12:39:15 +10005089{
5090 shrink_stripes(conf);
Dan Williams36d1c642009-07-14 11:48:22 -07005091 raid5_free_percpu(conf);
Dan Williams95fc17a2009-07-31 12:39:15 +10005092 kfree(conf->disks);
5093 kfree(conf->stripe_hashtbl);
5094 kfree(conf);
5095}
5096
Dan Williams36d1c642009-07-14 11:48:22 -07005097#ifdef CONFIG_HOTPLUG_CPU
5098static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5099 void *hcpu)
5100{
NeilBrownd1688a62011-10-11 16:49:52 +11005101 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
Dan Williams36d1c642009-07-14 11:48:22 -07005102 long cpu = (long)hcpu;
5103 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5104
5105 switch (action) {
5106 case CPU_UP_PREPARE:
5107 case CPU_UP_PREPARE_FROZEN:
Oleg Nesterov4d4ef862014-02-06 03:42:45 +05305108 if (alloc_scratch_buffer(conf, percpu)) {
Dan Williams36d1c642009-07-14 11:48:22 -07005109 pr_err("%s: failed memory allocation for cpu%ld\n",
5110 __func__, cpu);
Akinobu Mita55af6bb2010-05-26 14:43:35 -07005111 return notifier_from_errno(-ENOMEM);
Dan Williams36d1c642009-07-14 11:48:22 -07005112 }
5113 break;
5114 case CPU_DEAD:
5115 case CPU_DEAD_FROZEN:
Oleg Nesterov4d4ef862014-02-06 03:42:45 +05305116 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
Dan Williams36d1c642009-07-14 11:48:22 -07005117 break;
5118 default:
5119 break;
5120 }
5121 return NOTIFY_OK;
5122}
5123#endif
5124
NeilBrownd1688a62011-10-11 16:49:52 +11005125static int raid5_alloc_percpu(struct r5conf *conf)
Dan Williams36d1c642009-07-14 11:48:22 -07005126{
5127 unsigned long cpu;
Oleg Nesterov4d4ef862014-02-06 03:42:45 +05305128 int err = 0;
Dan Williams36d1c642009-07-14 11:48:22 -07005129
Oleg Nesterov4d4ef862014-02-06 03:42:45 +05305130 conf->percpu = alloc_percpu(struct raid5_percpu);
5131 if (!conf->percpu)
Dan Williams36d1c642009-07-14 11:48:22 -07005132 return -ENOMEM;
Dan Williams36d1c642009-07-14 11:48:22 -07005133
Dan Williams36d1c642009-07-14 11:48:22 -07005134#ifdef CONFIG_HOTPLUG_CPU
5135 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5136 conf->cpu_notify.priority = 0;
Oleg Nesterov4d4ef862014-02-06 03:42:45 +05305137 err = register_cpu_notifier(&conf->cpu_notify);
5138 if (err)
5139 return err;
Dan Williams36d1c642009-07-14 11:48:22 -07005140#endif
Oleg Nesterov4d4ef862014-02-06 03:42:45 +05305141
5142 get_online_cpus();
5143 for_each_present_cpu(cpu) {
5144 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5145 if (err) {
5146 pr_err("%s: failed memory allocation for cpu%ld\n",
5147 __func__, cpu);
5148 break;
5149 }
5150 }
Dan Williams36d1c642009-07-14 11:48:22 -07005151 put_online_cpus();
5152
5153 return err;
5154}
5155
NeilBrownd1688a62011-10-11 16:49:52 +11005156static struct r5conf *setup_conf(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005157{
NeilBrownd1688a62011-10-11 16:49:52 +11005158 struct r5conf *conf;
NeilBrown5e5e3e72009-10-16 16:35:30 +11005159 int raid_disk, memory, max_disks;
NeilBrown3cb03002011-10-11 16:45:26 +11005160 struct md_rdev *rdev;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005161 struct disk_info *disk;
NeilBrown02326052012-07-03 15:56:52 +10005162 char pers_name[6];
Linus Torvalds1da177e2005-04-16 15:20:36 -07005163
NeilBrown91adb562009-03-31 14:39:39 +11005164 if (mddev->new_level != 5
5165 && mddev->new_level != 4
5166 && mddev->new_level != 6) {
NeilBrown0c55e022010-05-03 14:09:02 +10005167 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
NeilBrown91adb562009-03-31 14:39:39 +11005168 mdname(mddev), mddev->new_level);
5169 return ERR_PTR(-EIO);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005170 }
NeilBrown91adb562009-03-31 14:39:39 +11005171 if ((mddev->new_level == 5
5172 && !algorithm_valid_raid5(mddev->new_layout)) ||
5173 (mddev->new_level == 6
5174 && !algorithm_valid_raid6(mddev->new_layout))) {
NeilBrown0c55e022010-05-03 14:09:02 +10005175 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
NeilBrown91adb562009-03-31 14:39:39 +11005176 mdname(mddev), mddev->new_layout);
5177 return ERR_PTR(-EIO);
5178 }
5179 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
NeilBrown0c55e022010-05-03 14:09:02 +10005180 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
NeilBrown91adb562009-03-31 14:39:39 +11005181 mdname(mddev), mddev->raid_disks);
5182 return ERR_PTR(-EINVAL);
NeilBrown99c0fb52009-03-31 14:39:38 +11005183 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005184
Andre Noll664e7c42009-06-18 08:45:27 +10005185 if (!mddev->new_chunk_sectors ||
5186 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5187 !is_power_of_2(mddev->new_chunk_sectors)) {
NeilBrown0c55e022010-05-03 14:09:02 +10005188 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5189 mdname(mddev), mddev->new_chunk_sectors << 9);
NeilBrown91adb562009-03-31 14:39:39 +11005190 return ERR_PTR(-EINVAL);
NeilBrown4bbf3772008-10-13 11:55:12 +11005191 }
5192
NeilBrownd1688a62011-10-11 16:49:52 +11005193 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
NeilBrown91adb562009-03-31 14:39:39 +11005194 if (conf == NULL)
5195 goto abort;
Dan Williamsf5efd452009-10-16 15:55:38 +11005196 spin_lock_init(&conf->device_lock);
5197 init_waitqueue_head(&conf->wait_for_stripe);
5198 init_waitqueue_head(&conf->wait_for_overlap);
5199 INIT_LIST_HEAD(&conf->handle_list);
5200 INIT_LIST_HEAD(&conf->hold_list);
5201 INIT_LIST_HEAD(&conf->delayed_list);
5202 INIT_LIST_HEAD(&conf->bitmap_list);
5203 INIT_LIST_HEAD(&conf->inactive_list);
5204 atomic_set(&conf->active_stripes, 0);
5205 atomic_set(&conf->preread_active_stripes, 0);
5206 atomic_set(&conf->active_aligned_reads, 0);
5207 conf->bypass_threshold = BYPASS_THRESHOLD;
NeilBrownd890fa22011-10-26 11:54:39 +11005208 conf->recovery_disabled = mddev->recovery_disabled - 1;
NeilBrown91adb562009-03-31 14:39:39 +11005209
5210 conf->raid_disks = mddev->raid_disks;
5211 if (mddev->reshape_position == MaxSector)
5212 conf->previous_raid_disks = mddev->raid_disks;
5213 else
5214 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
NeilBrown5e5e3e72009-10-16 16:35:30 +11005215 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5216 conf->scribble_len = scribble_len(max_disks);
NeilBrown91adb562009-03-31 14:39:39 +11005217
NeilBrown5e5e3e72009-10-16 16:35:30 +11005218 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
NeilBrown91adb562009-03-31 14:39:39 +11005219 GFP_KERNEL);
5220 if (!conf->disks)
5221 goto abort;
5222
5223 conf->mddev = mddev;
5224
5225 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5226 goto abort;
5227
Dan Williams36d1c642009-07-14 11:48:22 -07005228 conf->level = mddev->new_level;
5229 if (raid5_alloc_percpu(conf) != 0)
5230 goto abort;
5231
NeilBrown0c55e022010-05-03 14:09:02 +10005232 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
NeilBrown91adb562009-03-31 14:39:39 +11005233
NeilBrowndafb20f2012-03-19 12:46:39 +11005234 rdev_for_each(rdev, mddev) {
NeilBrown91adb562009-03-31 14:39:39 +11005235 raid_disk = rdev->raid_disk;
NeilBrown5e5e3e72009-10-16 16:35:30 +11005236 if (raid_disk >= max_disks
NeilBrown91adb562009-03-31 14:39:39 +11005237 || raid_disk < 0)
5238 continue;
5239 disk = conf->disks + raid_disk;
5240
NeilBrown17045f52011-12-23 10:17:53 +11005241 if (test_bit(Replacement, &rdev->flags)) {
5242 if (disk->replacement)
5243 goto abort;
5244 disk->replacement = rdev;
5245 } else {
5246 if (disk->rdev)
5247 goto abort;
5248 disk->rdev = rdev;
5249 }
NeilBrown91adb562009-03-31 14:39:39 +11005250
5251 if (test_bit(In_sync, &rdev->flags)) {
5252 char b[BDEVNAME_SIZE];
NeilBrown0c55e022010-05-03 14:09:02 +10005253 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5254 " disk %d\n",
5255 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
Jonathan Brassowd6b212f2011-06-08 18:00:28 -05005256 } else if (rdev->saved_raid_disk != raid_disk)
NeilBrown91adb562009-03-31 14:39:39 +11005257 /* Cannot rely on bitmap to complete recovery */
5258 conf->fullsync = 1;
5259 }
5260
Andre Noll09c9e5f2009-06-18 08:45:55 +10005261 conf->chunk_sectors = mddev->new_chunk_sectors;
NeilBrown91adb562009-03-31 14:39:39 +11005262 conf->level = mddev->new_level;
5263 if (conf->level == 6)
5264 conf->max_degraded = 2;
5265 else
5266 conf->max_degraded = 1;
5267 conf->algorithm = mddev->new_layout;
5268 conf->max_nr_stripes = NR_STRIPES;
NeilBrownfef9c612009-03-31 15:16:46 +11005269 conf->reshape_progress = mddev->reshape_position;
NeilBrowne183eae2009-03-31 15:20:22 +11005270 if (conf->reshape_progress != MaxSector) {
Andre Noll09c9e5f2009-06-18 08:45:55 +10005271 conf->prev_chunk_sectors = mddev->chunk_sectors;
NeilBrowne183eae2009-03-31 15:20:22 +11005272 conf->prev_algo = mddev->layout;
5273 }
NeilBrown91adb562009-03-31 14:39:39 +11005274
5275 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
NeilBrown5e5e3e72009-10-16 16:35:30 +11005276 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
NeilBrown91adb562009-03-31 14:39:39 +11005277 if (grow_stripes(conf, conf->max_nr_stripes)) {
5278 printk(KERN_ERR
NeilBrown0c55e022010-05-03 14:09:02 +10005279 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5280 mdname(mddev), memory);
NeilBrown91adb562009-03-31 14:39:39 +11005281 goto abort;
5282 } else
NeilBrown0c55e022010-05-03 14:09:02 +10005283 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5284 mdname(mddev), memory);
NeilBrown91adb562009-03-31 14:39:39 +11005285
NeilBrown02326052012-07-03 15:56:52 +10005286 sprintf(pers_name, "raid%d", mddev->new_level);
5287 conf->thread = md_register_thread(raid5d, mddev, pers_name);
NeilBrown91adb562009-03-31 14:39:39 +11005288 if (!conf->thread) {
5289 printk(KERN_ERR
NeilBrown0c55e022010-05-03 14:09:02 +10005290 "md/raid:%s: couldn't allocate thread.\n",
NeilBrown91adb562009-03-31 14:39:39 +11005291 mdname(mddev));
5292 goto abort;
5293 }
5294
5295 return conf;
5296
5297 abort:
5298 if (conf) {
Dan Williams95fc17a2009-07-31 12:39:15 +10005299 free_conf(conf);
NeilBrown91adb562009-03-31 14:39:39 +11005300 return ERR_PTR(-EIO);
5301 } else
5302 return ERR_PTR(-ENOMEM);
5303}
5304
NeilBrownc148ffd2009-11-13 17:47:00 +11005305
5306static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5307{
5308 switch (algo) {
5309 case ALGORITHM_PARITY_0:
5310 if (raid_disk < max_degraded)
5311 return 1;
5312 break;
5313 case ALGORITHM_PARITY_N:
5314 if (raid_disk >= raid_disks - max_degraded)
5315 return 1;
5316 break;
5317 case ALGORITHM_PARITY_0_6:
5318 if (raid_disk == 0 ||
5319 raid_disk == raid_disks - 1)
5320 return 1;
5321 break;
5322 case ALGORITHM_LEFT_ASYMMETRIC_6:
5323 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5324 case ALGORITHM_LEFT_SYMMETRIC_6:
5325 case ALGORITHM_RIGHT_SYMMETRIC_6:
5326 if (raid_disk == raid_disks - 1)
5327 return 1;
5328 }
5329 return 0;
5330}
5331
NeilBrownfd01b882011-10-11 16:47:53 +11005332static int run(struct mddev *mddev)
NeilBrown91adb562009-03-31 14:39:39 +11005333{
NeilBrownd1688a62011-10-11 16:49:52 +11005334 struct r5conf *conf;
NeilBrown9f7c2222010-07-26 12:04:13 +10005335 int working_disks = 0;
NeilBrownc148ffd2009-11-13 17:47:00 +11005336 int dirty_parity_disks = 0;
NeilBrown3cb03002011-10-11 16:45:26 +11005337 struct md_rdev *rdev;
NeilBrownc148ffd2009-11-13 17:47:00 +11005338 sector_t reshape_offset = 0;
NeilBrown17045f52011-12-23 10:17:53 +11005339 int i;
NeilBrownb5254dd2012-05-21 09:27:01 +10005340 long long min_offset_diff = 0;
5341 int first = 1;
NeilBrown91adb562009-03-31 14:39:39 +11005342
Andre Noll8c6ac862009-06-18 08:48:06 +10005343 if (mddev->recovery_cp != MaxSector)
NeilBrown0c55e022010-05-03 14:09:02 +10005344 printk(KERN_NOTICE "md/raid:%s: not clean"
Andre Noll8c6ac862009-06-18 08:48:06 +10005345 " -- starting background reconstruction\n",
5346 mdname(mddev));
NeilBrownb5254dd2012-05-21 09:27:01 +10005347
5348 rdev_for_each(rdev, mddev) {
5349 long long diff;
5350 if (rdev->raid_disk < 0)
5351 continue;
5352 diff = (rdev->new_data_offset - rdev->data_offset);
5353 if (first) {
5354 min_offset_diff = diff;
5355 first = 0;
5356 } else if (mddev->reshape_backwards &&
5357 diff < min_offset_diff)
5358 min_offset_diff = diff;
5359 else if (!mddev->reshape_backwards &&
5360 diff > min_offset_diff)
5361 min_offset_diff = diff;
5362 }
5363
NeilBrownf6705572006-03-27 01:18:11 -08005364 if (mddev->reshape_position != MaxSector) {
5365 /* Check that we can continue the reshape.
NeilBrownb5254dd2012-05-21 09:27:01 +10005366 * Difficulties arise if the stripe we would write to
5367 * next is at or after the stripe we would read from next.
5368 * For a reshape that changes the number of devices, this
5369 * is only possible for a very short time, and mdadm makes
5370 * sure that time appears to have past before assembling
5371 * the array. So we fail if that time hasn't passed.
5372 * For a reshape that keeps the number of devices the same
5373 * mdadm must be monitoring the reshape can keeping the
5374 * critical areas read-only and backed up. It will start
5375 * the array in read-only mode, so we check for that.
NeilBrownf6705572006-03-27 01:18:11 -08005376 */
5377 sector_t here_new, here_old;
5378 int old_disks;
Andre Noll18b00332009-03-31 15:00:56 +11005379 int max_degraded = (mddev->level == 6 ? 2 : 1);
NeilBrownf6705572006-03-27 01:18:11 -08005380
NeilBrown88ce4932009-03-31 15:24:23 +11005381 if (mddev->new_level != mddev->level) {
NeilBrown0c55e022010-05-03 14:09:02 +10005382 printk(KERN_ERR "md/raid:%s: unsupported reshape "
NeilBrownf4168852007-02-28 20:11:53 -08005383 "required - aborting.\n",
NeilBrownf6705572006-03-27 01:18:11 -08005384 mdname(mddev));
5385 return -EINVAL;
5386 }
NeilBrownf6705572006-03-27 01:18:11 -08005387 old_disks = mddev->raid_disks - mddev->delta_disks;
5388 /* reshape_position must be on a new-stripe boundary, and one
NeilBrownf4168852007-02-28 20:11:53 -08005389 * further up in new geometry must map after here in old
5390 * geometry.
NeilBrownf6705572006-03-27 01:18:11 -08005391 */
5392 here_new = mddev->reshape_position;
Andre Noll664e7c42009-06-18 08:45:27 +10005393 if (sector_div(here_new, mddev->new_chunk_sectors *
NeilBrownf4168852007-02-28 20:11:53 -08005394 (mddev->raid_disks - max_degraded))) {
NeilBrown0c55e022010-05-03 14:09:02 +10005395 printk(KERN_ERR "md/raid:%s: reshape_position not "
5396 "on a stripe boundary\n", mdname(mddev));
NeilBrownf6705572006-03-27 01:18:11 -08005397 return -EINVAL;
5398 }
NeilBrownc148ffd2009-11-13 17:47:00 +11005399 reshape_offset = here_new * mddev->new_chunk_sectors;
NeilBrownf6705572006-03-27 01:18:11 -08005400 /* here_new is the stripe we will write to */
5401 here_old = mddev->reshape_position;
Andre Noll9d8f0362009-06-18 08:45:01 +10005402 sector_div(here_old, mddev->chunk_sectors *
NeilBrownf4168852007-02-28 20:11:53 -08005403 (old_disks-max_degraded));
5404 /* here_old is the first stripe that we might need to read
5405 * from */
NeilBrown67ac6012009-08-13 10:06:24 +10005406 if (mddev->delta_disks == 0) {
NeilBrownb5254dd2012-05-21 09:27:01 +10005407 if ((here_new * mddev->new_chunk_sectors !=
5408 here_old * mddev->chunk_sectors)) {
5409 printk(KERN_ERR "md/raid:%s: reshape position is"
5410 " confused - aborting\n", mdname(mddev));
5411 return -EINVAL;
5412 }
NeilBrown67ac6012009-08-13 10:06:24 +10005413 /* We cannot be sure it is safe to start an in-place
NeilBrownb5254dd2012-05-21 09:27:01 +10005414 * reshape. It is only safe if user-space is monitoring
NeilBrown67ac6012009-08-13 10:06:24 +10005415 * and taking constant backups.
5416 * mdadm always starts a situation like this in
5417 * readonly mode so it can take control before
5418 * allowing any writes. So just check for that.
5419 */
NeilBrownb5254dd2012-05-21 09:27:01 +10005420 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5421 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5422 /* not really in-place - so OK */;
5423 else if (mddev->ro == 0) {
5424 printk(KERN_ERR "md/raid:%s: in-place reshape "
5425 "must be started in read-only mode "
5426 "- aborting\n",
NeilBrown0c55e022010-05-03 14:09:02 +10005427 mdname(mddev));
NeilBrown67ac6012009-08-13 10:06:24 +10005428 return -EINVAL;
5429 }
NeilBrown2c810cd2012-05-21 09:27:00 +10005430 } else if (mddev->reshape_backwards
NeilBrownb5254dd2012-05-21 09:27:01 +10005431 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
NeilBrown67ac6012009-08-13 10:06:24 +10005432 here_old * mddev->chunk_sectors)
5433 : (here_new * mddev->new_chunk_sectors >=
NeilBrownb5254dd2012-05-21 09:27:01 +10005434 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
NeilBrownf6705572006-03-27 01:18:11 -08005435 /* Reading from the same stripe as writing to - bad */
NeilBrown0c55e022010-05-03 14:09:02 +10005436 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5437 "auto-recovery - aborting.\n",
5438 mdname(mddev));
NeilBrownf6705572006-03-27 01:18:11 -08005439 return -EINVAL;
5440 }
NeilBrown0c55e022010-05-03 14:09:02 +10005441 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5442 mdname(mddev));
NeilBrownf6705572006-03-27 01:18:11 -08005443 /* OK, we should be able to continue; */
NeilBrownf6705572006-03-27 01:18:11 -08005444 } else {
NeilBrown91adb562009-03-31 14:39:39 +11005445 BUG_ON(mddev->level != mddev->new_level);
5446 BUG_ON(mddev->layout != mddev->new_layout);
Andre Noll664e7c42009-06-18 08:45:27 +10005447 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
NeilBrown91adb562009-03-31 14:39:39 +11005448 BUG_ON(mddev->delta_disks != 0);
NeilBrownf6705572006-03-27 01:18:11 -08005449 }
5450
NeilBrown245f46c2009-03-31 14:39:39 +11005451 if (mddev->private == NULL)
5452 conf = setup_conf(mddev);
5453 else
5454 conf = mddev->private;
5455
NeilBrown91adb562009-03-31 14:39:39 +11005456 if (IS_ERR(conf))
5457 return PTR_ERR(conf);
NeilBrown9ffae0c2006-01-06 00:20:32 -08005458
NeilBrownb5254dd2012-05-21 09:27:01 +10005459 conf->min_offset_diff = min_offset_diff;
NeilBrown91adb562009-03-31 14:39:39 +11005460 mddev->thread = conf->thread;
5461 conf->thread = NULL;
5462 mddev->private = conf;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005463
NeilBrown17045f52011-12-23 10:17:53 +11005464 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5465 i++) {
5466 rdev = conf->disks[i].rdev;
5467 if (!rdev && conf->disks[i].replacement) {
5468 /* The replacement is all we have yet */
5469 rdev = conf->disks[i].replacement;
5470 conf->disks[i].replacement = NULL;
5471 clear_bit(Replacement, &rdev->flags);
5472 conf->disks[i].rdev = rdev;
5473 }
5474 if (!rdev)
NeilBrownc148ffd2009-11-13 17:47:00 +11005475 continue;
NeilBrown17045f52011-12-23 10:17:53 +11005476 if (conf->disks[i].replacement &&
5477 conf->reshape_progress != MaxSector) {
5478 /* replacements and reshape simply do not mix. */
5479 printk(KERN_ERR "md: cannot handle concurrent "
5480 "replacement and reshape.\n");
5481 goto abort;
5482 }
NeilBrown2f115882010-06-17 17:41:03 +10005483 if (test_bit(In_sync, &rdev->flags)) {
NeilBrown91adb562009-03-31 14:39:39 +11005484 working_disks++;
NeilBrown2f115882010-06-17 17:41:03 +10005485 continue;
5486 }
NeilBrownc148ffd2009-11-13 17:47:00 +11005487 /* This disc is not fully in-sync. However if it
5488 * just stored parity (beyond the recovery_offset),
5489 * when we don't need to be concerned about the
5490 * array being dirty.
5491 * When reshape goes 'backwards', we never have
5492 * partially completed devices, so we only need
5493 * to worry about reshape going forwards.
5494 */
5495 /* Hack because v0.91 doesn't store recovery_offset properly. */
5496 if (mddev->major_version == 0 &&
5497 mddev->minor_version > 90)
5498 rdev->recovery_offset = reshape_offset;
H. Peter Anvin5026d7a2013-06-12 07:37:43 -07005499
NeilBrownc148ffd2009-11-13 17:47:00 +11005500 if (rdev->recovery_offset < reshape_offset) {
5501 /* We need to check old and new layout */
5502 if (!only_parity(rdev->raid_disk,
5503 conf->algorithm,
5504 conf->raid_disks,
5505 conf->max_degraded))
5506 continue;
5507 }
5508 if (!only_parity(rdev->raid_disk,
5509 conf->prev_algo,
5510 conf->previous_raid_disks,
5511 conf->max_degraded))
5512 continue;
5513 dirty_parity_disks++;
5514 }
NeilBrown91adb562009-03-31 14:39:39 +11005515
NeilBrown17045f52011-12-23 10:17:53 +11005516 /*
5517 * 0 for a fully functional array, 1 or 2 for a degraded array.
5518 */
NeilBrown908f4fb2011-12-23 10:17:50 +11005519 mddev->degraded = calc_degraded(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005520
NeilBrown674806d2010-06-16 17:17:53 +10005521 if (has_failed(conf)) {
NeilBrown0c55e022010-05-03 14:09:02 +10005522 printk(KERN_ERR "md/raid:%s: not enough operational devices"
Linus Torvalds1da177e2005-04-16 15:20:36 -07005523 " (%d/%d failed)\n",
NeilBrown02c2de82006-10-03 01:15:47 -07005524 mdname(mddev), mddev->degraded, conf->raid_disks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005525 goto abort;
5526 }
5527
NeilBrown91adb562009-03-31 14:39:39 +11005528 /* device size must be a multiple of chunk size */
Andre Noll9d8f0362009-06-18 08:45:01 +10005529 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
NeilBrown91adb562009-03-31 14:39:39 +11005530 mddev->resync_max_sectors = mddev->dev_sectors;
5531
NeilBrownc148ffd2009-11-13 17:47:00 +11005532 if (mddev->degraded > dirty_parity_disks &&
Linus Torvalds1da177e2005-04-16 15:20:36 -07005533 mddev->recovery_cp != MaxSector) {
NeilBrown6ff8d8ec2006-01-06 00:20:15 -08005534 if (mddev->ok_start_degraded)
5535 printk(KERN_WARNING
NeilBrown0c55e022010-05-03 14:09:02 +10005536 "md/raid:%s: starting dirty degraded array"
5537 " - data corruption possible.\n",
NeilBrown6ff8d8ec2006-01-06 00:20:15 -08005538 mdname(mddev));
5539 else {
5540 printk(KERN_ERR
NeilBrown0c55e022010-05-03 14:09:02 +10005541 "md/raid:%s: cannot start dirty degraded array.\n",
NeilBrown6ff8d8ec2006-01-06 00:20:15 -08005542 mdname(mddev));
5543 goto abort;
5544 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005545 }
5546
Linus Torvalds1da177e2005-04-16 15:20:36 -07005547 if (mddev->degraded == 0)
NeilBrown0c55e022010-05-03 14:09:02 +10005548 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5549 " devices, algorithm %d\n", mdname(mddev), conf->level,
NeilBrowne183eae2009-03-31 15:20:22 +11005550 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5551 mddev->new_layout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005552 else
NeilBrown0c55e022010-05-03 14:09:02 +10005553 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5554 " out of %d devices, algorithm %d\n",
5555 mdname(mddev), conf->level,
5556 mddev->raid_disks - mddev->degraded,
5557 mddev->raid_disks, mddev->new_layout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005558
5559 print_raid5_conf(conf);
5560
NeilBrownfef9c612009-03-31 15:16:46 +11005561 if (conf->reshape_progress != MaxSector) {
NeilBrownfef9c612009-03-31 15:16:46 +11005562 conf->reshape_safe = conf->reshape_progress;
NeilBrownf6705572006-03-27 01:18:11 -08005563 atomic_set(&conf->reshape_stripes, 0);
5564 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5565 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5566 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5567 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5568 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
NeilBrown0da3c612009-09-23 18:09:45 +10005569 "reshape");
NeilBrownf6705572006-03-27 01:18:11 -08005570 }
5571
Linus Torvalds1da177e2005-04-16 15:20:36 -07005572
5573 /* Ok, everything is just fine now */
NeilBrowna64c8762010-04-14 17:15:37 +10005574 if (mddev->to_remove == &raid5_attrs_group)
5575 mddev->to_remove = NULL;
NeilBrown00bcb4a2010-06-01 19:37:23 +10005576 else if (mddev->kobj.sd &&
5577 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
NeilBrown5e55e2f2007-03-26 21:32:14 -08005578 printk(KERN_WARNING
NeilBrown4a5add42010-06-01 19:37:28 +10005579 "raid5: failed to create sysfs attributes for %s\n",
NeilBrown5e55e2f2007-03-26 21:32:14 -08005580 mdname(mddev));
NeilBrown4a5add42010-06-01 19:37:28 +10005581 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5582
5583 if (mddev->queue) {
NeilBrown9f7c2222010-07-26 12:04:13 +10005584 int chunk_size;
Shaohua Li620125f2012-10-11 13:49:05 +11005585 bool discard_supported = true;
NeilBrown4a5add42010-06-01 19:37:28 +10005586 /* read-ahead size must cover two whole stripes, which
5587 * is 2 * (datadisks) * chunksize where 'n' is the
5588 * number of raid devices
5589 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005590 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5591 int stripe = data_disks *
5592 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5593 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5594 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
NeilBrown4a5add42010-06-01 19:37:28 +10005595
5596 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
NeilBrown11d8a6e2010-07-26 11:57:07 +10005597
5598 mddev->queue->backing_dev_info.congested_data = mddev;
5599 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
NeilBrown9f7c2222010-07-26 12:04:13 +10005600
5601 chunk_size = mddev->chunk_sectors << 9;
5602 blk_queue_io_min(mddev->queue, chunk_size);
5603 blk_queue_io_opt(mddev->queue, chunk_size *
5604 (conf->raid_disks - conf->max_degraded));
Shaohua Li620125f2012-10-11 13:49:05 +11005605 /*
5606 * We can only discard a whole stripe. It doesn't make sense to
5607 * discard data disk but write parity disk
5608 */
5609 stripe = stripe * PAGE_SIZE;
NeilBrown4ac68752012-11-19 13:11:26 +11005610 /* Round up to power of 2, as discard handling
5611 * currently assumes that */
5612 while ((stripe-1) & stripe)
5613 stripe = (stripe | (stripe-1)) + 1;
Shaohua Li620125f2012-10-11 13:49:05 +11005614 mddev->queue->limits.discard_alignment = stripe;
5615 mddev->queue->limits.discard_granularity = stripe;
5616 /*
5617 * unaligned part of discard request will be ignored, so can't
NeilBrown06905ff2014-10-02 13:45:00 +10005618 * guarantee discard_zeroes_data
Shaohua Li620125f2012-10-11 13:49:05 +11005619 */
5620 mddev->queue->limits.discard_zeroes_data = 0;
NeilBrown9f7c2222010-07-26 12:04:13 +10005621
H. Peter Anvin5026d7a2013-06-12 07:37:43 -07005622 blk_queue_max_write_same_sectors(mddev->queue, 0);
5623
NeilBrown05616be2012-05-21 09:27:00 +10005624 rdev_for_each(rdev, mddev) {
NeilBrown9f7c2222010-07-26 12:04:13 +10005625 disk_stack_limits(mddev->gendisk, rdev->bdev,
5626 rdev->data_offset << 9);
NeilBrown05616be2012-05-21 09:27:00 +10005627 disk_stack_limits(mddev->gendisk, rdev->bdev,
5628 rdev->new_data_offset << 9);
Shaohua Li620125f2012-10-11 13:49:05 +11005629 /*
5630 * discard_zeroes_data is required, otherwise data
5631 * could be lost. Consider a scenario: discard a stripe
5632 * (the stripe could be inconsistent if
5633 * discard_zeroes_data is 0); write one disk of the
5634 * stripe (the stripe could be inconsistent again
5635 * depending on which disks are used to calculate
5636 * parity); the disk is broken; The stripe data of this
5637 * disk is lost.
5638 */
5639 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5640 !bdev_get_queue(rdev->bdev)->
5641 limits.discard_zeroes_data)
5642 discard_supported = false;
NeilBrown06905ff2014-10-02 13:45:00 +10005643 /* Unfortunately, discard_zeroes_data is not currently
5644 * a guarantee - just a hint. So we only allow DISCARD
5645 * if the sysadmin has confirmed that only safe devices
5646 * are in use by setting a module parameter.
5647 */
5648 if (!devices_handle_discard_safely) {
5649 if (discard_supported) {
5650 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
5651 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
5652 }
5653 discard_supported = false;
5654 }
NeilBrown05616be2012-05-21 09:27:00 +10005655 }
Shaohua Li620125f2012-10-11 13:49:05 +11005656
5657 if (discard_supported &&
5658 mddev->queue->limits.max_discard_sectors >= stripe &&
5659 mddev->queue->limits.discard_granularity >= stripe)
5660 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5661 mddev->queue);
5662 else
5663 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5664 mddev->queue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005665 }
5666
Linus Torvalds1da177e2005-04-16 15:20:36 -07005667 return 0;
5668abort:
NeilBrown01f96c02011-09-21 15:30:20 +10005669 md_unregister_thread(&mddev->thread);
NeilBrowne4f869d2011-10-07 14:22:49 +11005670 print_raid5_conf(conf);
5671 free_conf(conf);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005672 mddev->private = NULL;
NeilBrown0c55e022010-05-03 14:09:02 +10005673 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005674 return -EIO;
5675}
5676
NeilBrownfd01b882011-10-11 16:47:53 +11005677static int stop(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005678{
NeilBrownd1688a62011-10-11 16:49:52 +11005679 struct r5conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005680
NeilBrown01f96c02011-09-21 15:30:20 +10005681 md_unregister_thread(&mddev->thread);
NeilBrown11d8a6e2010-07-26 11:57:07 +10005682 if (mddev->queue)
5683 mddev->queue->backing_dev_info.congested_fn = NULL;
Dan Williams95fc17a2009-07-31 12:39:15 +10005684 free_conf(conf);
NeilBrowna64c8762010-04-14 17:15:37 +10005685 mddev->private = NULL;
5686 mddev->to_remove = &raid5_attrs_group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005687 return 0;
5688}
5689
NeilBrownfd01b882011-10-11 16:47:53 +11005690static void status(struct seq_file *seq, struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005691{
NeilBrownd1688a62011-10-11 16:49:52 +11005692 struct r5conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005693 int i;
5694
Andre Noll9d8f0362009-06-18 08:45:01 +10005695 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5696 mddev->chunk_sectors / 2, mddev->layout);
NeilBrown02c2de82006-10-03 01:15:47 -07005697 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005698 for (i = 0; i < conf->raid_disks; i++)
5699 seq_printf (seq, "%s",
5700 conf->disks[i].rdev &&
NeilBrownb2d444d2005-11-08 21:39:31 -08005701 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005702 seq_printf (seq, "]");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005703}
5704
NeilBrownd1688a62011-10-11 16:49:52 +11005705static void print_raid5_conf (struct r5conf *conf)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005706{
5707 int i;
5708 struct disk_info *tmp;
5709
NeilBrown0c55e022010-05-03 14:09:02 +10005710 printk(KERN_DEBUG "RAID conf printout:\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005711 if (!conf) {
5712 printk("(conf==NULL)\n");
5713 return;
5714 }
NeilBrown0c55e022010-05-03 14:09:02 +10005715 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5716 conf->raid_disks,
5717 conf->raid_disks - conf->mddev->degraded);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005718
5719 for (i = 0; i < conf->raid_disks; i++) {
5720 char b[BDEVNAME_SIZE];
5721 tmp = conf->disks + i;
5722 if (tmp->rdev)
NeilBrown0c55e022010-05-03 14:09:02 +10005723 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5724 i, !test_bit(Faulty, &tmp->rdev->flags),
5725 bdevname(tmp->rdev->bdev, b));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005726 }
5727}
5728
NeilBrownfd01b882011-10-11 16:47:53 +11005729static int raid5_spare_active(struct mddev *mddev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005730{
5731 int i;
NeilBrownd1688a62011-10-11 16:49:52 +11005732 struct r5conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005733 struct disk_info *tmp;
NeilBrown6b965622010-08-18 11:56:59 +10005734 int count = 0;
5735 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005736
5737 for (i = 0; i < conf->raid_disks; i++) {
5738 tmp = conf->disks + i;
NeilBrowndd054fc2011-12-23 10:17:53 +11005739 if (tmp->replacement
5740 && tmp->replacement->recovery_offset == MaxSector
5741 && !test_bit(Faulty, &tmp->replacement->flags)
5742 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5743 /* Replacement has just become active. */
5744 if (!tmp->rdev
5745 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5746 count++;
5747 if (tmp->rdev) {
5748 /* Replaced device not technically faulty,
5749 * but we need to be sure it gets removed
5750 * and never re-added.
5751 */
5752 set_bit(Faulty, &tmp->rdev->flags);
5753 sysfs_notify_dirent_safe(
5754 tmp->rdev->sysfs_state);
5755 }
5756 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5757 } else if (tmp->rdev
NeilBrown70fffd02010-06-16 17:01:25 +10005758 && tmp->rdev->recovery_offset == MaxSector
NeilBrownb2d444d2005-11-08 21:39:31 -08005759 && !test_bit(Faulty, &tmp->rdev->flags)
NeilBrownc04be0a2006-10-03 01:15:53 -07005760 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
NeilBrown6b965622010-08-18 11:56:59 +10005761 count++;
Jonathan Brassow43c73ca2011-01-14 09:14:33 +11005762 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005763 }
5764 }
NeilBrown6b965622010-08-18 11:56:59 +10005765 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrown908f4fb2011-12-23 10:17:50 +11005766 mddev->degraded = calc_degraded(conf);
NeilBrown6b965622010-08-18 11:56:59 +10005767 spin_unlock_irqrestore(&conf->device_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005768 print_raid5_conf(conf);
NeilBrown6b965622010-08-18 11:56:59 +10005769 return count;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005770}
5771
NeilBrownb8321b62011-12-23 10:17:51 +11005772static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005773{
NeilBrownd1688a62011-10-11 16:49:52 +11005774 struct r5conf *conf = mddev->private;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005775 int err = 0;
NeilBrownb8321b62011-12-23 10:17:51 +11005776 int number = rdev->raid_disk;
NeilBrown657e3e42011-12-23 10:17:52 +11005777 struct md_rdev **rdevp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005778 struct disk_info *p = conf->disks + number;
5779
5780 print_raid5_conf(conf);
NeilBrown657e3e42011-12-23 10:17:52 +11005781 if (rdev == p->rdev)
5782 rdevp = &p->rdev;
5783 else if (rdev == p->replacement)
5784 rdevp = &p->replacement;
5785 else
5786 return 0;
NeilBrownec32a2b2009-03-31 15:17:38 +11005787
NeilBrown657e3e42011-12-23 10:17:52 +11005788 if (number >= conf->raid_disks &&
5789 conf->reshape_progress == MaxSector)
5790 clear_bit(In_sync, &rdev->flags);
5791
5792 if (test_bit(In_sync, &rdev->flags) ||
5793 atomic_read(&rdev->nr_pending)) {
5794 err = -EBUSY;
5795 goto abort;
5796 }
5797 /* Only remove non-faulty devices if recovery
5798 * isn't possible.
5799 */
5800 if (!test_bit(Faulty, &rdev->flags) &&
5801 mddev->recovery_disabled != conf->recovery_disabled &&
5802 !has_failed(conf) &&
NeilBrowndd054fc2011-12-23 10:17:53 +11005803 (!p->replacement || p->replacement == rdev) &&
NeilBrown657e3e42011-12-23 10:17:52 +11005804 number < conf->raid_disks) {
5805 err = -EBUSY;
5806 goto abort;
5807 }
5808 *rdevp = NULL;
5809 synchronize_rcu();
5810 if (atomic_read(&rdev->nr_pending)) {
5811 /* lost the race, try later */
5812 err = -EBUSY;
5813 *rdevp = rdev;
NeilBrowndd054fc2011-12-23 10:17:53 +11005814 } else if (p->replacement) {
5815 /* We must have just cleared 'rdev' */
5816 p->rdev = p->replacement;
5817 clear_bit(Replacement, &p->replacement->flags);
5818 smp_mb(); /* Make sure other CPUs may see both as identical
5819 * but will never see neither - if they are careful
5820 */
5821 p->replacement = NULL;
5822 clear_bit(WantReplacement, &rdev->flags);
5823 } else
5824 /* We might have just removed the Replacement as faulty-
5825 * clear the bit just in case
5826 */
5827 clear_bit(WantReplacement, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005828abort:
5829
5830 print_raid5_conf(conf);
5831 return err;
5832}
5833
NeilBrownfd01b882011-10-11 16:47:53 +11005834static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005835{
NeilBrownd1688a62011-10-11 16:49:52 +11005836 struct r5conf *conf = mddev->private;
Neil Brown199050e2008-06-28 08:31:33 +10005837 int err = -EEXIST;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005838 int disk;
5839 struct disk_info *p;
Neil Brown6c2fce22008-06-28 08:31:31 +10005840 int first = 0;
5841 int last = conf->raid_disks - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005842
NeilBrown7f0da592011-07-28 11:39:22 +10005843 if (mddev->recovery_disabled == conf->recovery_disabled)
5844 return -EBUSY;
5845
NeilBrowndc10c642012-03-19 12:46:37 +11005846 if (rdev->saved_raid_disk < 0 && has_failed(conf))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005847 /* no point adding a device */
Neil Brown199050e2008-06-28 08:31:33 +10005848 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005849
Neil Brown6c2fce22008-06-28 08:31:31 +10005850 if (rdev->raid_disk >= 0)
5851 first = last = rdev->raid_disk;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005852
5853 /*
NeilBrown16a53ec2006-06-26 00:27:38 -07005854 * find the disk ... but prefer rdev->saved_raid_disk
5855 * if possible.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005856 */
NeilBrown16a53ec2006-06-26 00:27:38 -07005857 if (rdev->saved_raid_disk >= 0 &&
Neil Brown6c2fce22008-06-28 08:31:31 +10005858 rdev->saved_raid_disk >= first &&
NeilBrown16a53ec2006-06-26 00:27:38 -07005859 conf->disks[rdev->saved_raid_disk].rdev == NULL)
NeilBrown5cfb22a2012-07-03 11:46:53 +10005860 first = rdev->saved_raid_disk;
5861
5862 for (disk = first; disk <= last; disk++) {
NeilBrown7bfec5f2011-12-23 10:17:53 +11005863 p = conf->disks + disk;
5864 if (p->rdev == NULL) {
NeilBrownb2d444d2005-11-08 21:39:31 -08005865 clear_bit(In_sync, &rdev->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005866 rdev->raid_disk = disk;
Neil Brown199050e2008-06-28 08:31:33 +10005867 err = 0;
NeilBrown72626682005-09-09 16:23:54 -07005868 if (rdev->saved_raid_disk != disk)
5869 conf->fullsync = 1;
Suzanne Woodd6065f72005-11-08 21:39:27 -08005870 rcu_assign_pointer(p->rdev, rdev);
NeilBrown5cfb22a2012-07-03 11:46:53 +10005871 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005872 }
NeilBrown5cfb22a2012-07-03 11:46:53 +10005873 }
5874 for (disk = first; disk <= last; disk++) {
5875 p = conf->disks + disk;
NeilBrown7bfec5f2011-12-23 10:17:53 +11005876 if (test_bit(WantReplacement, &p->rdev->flags) &&
5877 p->replacement == NULL) {
5878 clear_bit(In_sync, &rdev->flags);
5879 set_bit(Replacement, &rdev->flags);
5880 rdev->raid_disk = disk;
5881 err = 0;
5882 conf->fullsync = 1;
5883 rcu_assign_pointer(p->replacement, rdev);
5884 break;
5885 }
5886 }
NeilBrown5cfb22a2012-07-03 11:46:53 +10005887out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005888 print_raid5_conf(conf);
Neil Brown199050e2008-06-28 08:31:33 +10005889 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005890}
5891
NeilBrownfd01b882011-10-11 16:47:53 +11005892static int raid5_resize(struct mddev *mddev, sector_t sectors)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005893{
5894 /* no resync is happening, and there is enough space
5895 * on all devices, so we can resize.
5896 * We need to make sure resync covers any new space.
5897 * If the array is shrinking we should possibly wait until
5898 * any io in the removed space completes, but it hardly seems
5899 * worth it.
5900 */
NeilBrowna4a61252012-05-22 13:55:27 +10005901 sector_t newsize;
Andre Noll9d8f0362009-06-18 08:45:01 +10005902 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
NeilBrowna4a61252012-05-22 13:55:27 +10005903 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5904 if (mddev->external_size &&
5905 mddev->array_sectors > newsize)
Dan Williamsb522adc2009-03-31 15:00:31 +11005906 return -EINVAL;
NeilBrowna4a61252012-05-22 13:55:27 +10005907 if (mddev->bitmap) {
5908 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5909 if (ret)
5910 return ret;
5911 }
5912 md_set_array_sectors(mddev, newsize);
Andre Nollf233ea52008-07-21 17:05:22 +10005913 set_capacity(mddev->gendisk, mddev->array_sectors);
NeilBrown449aad32009-08-03 10:59:58 +10005914 revalidate_disk(mddev->gendisk);
NeilBrownb0986362011-05-11 15:52:21 +10005915 if (sectors > mddev->dev_sectors &&
5916 mddev->recovery_cp > mddev->dev_sectors) {
Andre Noll58c0fed2009-03-31 14:33:13 +11005917 mddev->recovery_cp = mddev->dev_sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005918 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5919 }
Andre Noll58c0fed2009-03-31 14:33:13 +11005920 mddev->dev_sectors = sectors;
NeilBrown4b5c7ae2005-07-27 11:43:28 -07005921 mddev->resync_max_sectors = sectors;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005922 return 0;
5923}
5924
NeilBrownfd01b882011-10-11 16:47:53 +11005925static int check_stripe_cache(struct mddev *mddev)
NeilBrown01ee22b2009-06-18 08:47:20 +10005926{
5927 /* Can only proceed if there are plenty of stripe_heads.
5928 * We need a minimum of one full stripe,, and for sensible progress
5929 * it is best to have about 4 times that.
5930 * If we require 4 times, then the default 256 4K stripe_heads will
5931 * allow for chunk sizes up to 256K, which is probably OK.
5932 * If the chunk size is greater, user-space should request more
5933 * stripe_heads first.
5934 */
NeilBrownd1688a62011-10-11 16:49:52 +11005935 struct r5conf *conf = mddev->private;
NeilBrown01ee22b2009-06-18 08:47:20 +10005936 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5937 > conf->max_nr_stripes ||
5938 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5939 > conf->max_nr_stripes) {
NeilBrown0c55e022010-05-03 14:09:02 +10005940 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5941 mdname(mddev),
NeilBrown01ee22b2009-06-18 08:47:20 +10005942 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5943 / STRIPE_SIZE)*4);
5944 return 0;
5945 }
5946 return 1;
5947}
5948
NeilBrownfd01b882011-10-11 16:47:53 +11005949static int check_reshape(struct mddev *mddev)
NeilBrown29269552006-03-27 01:18:10 -08005950{
NeilBrownd1688a62011-10-11 16:49:52 +11005951 struct r5conf *conf = mddev->private;
NeilBrown29269552006-03-27 01:18:10 -08005952
NeilBrown88ce4932009-03-31 15:24:23 +11005953 if (mddev->delta_disks == 0 &&
5954 mddev->new_layout == mddev->layout &&
Andre Noll664e7c42009-06-18 08:45:27 +10005955 mddev->new_chunk_sectors == mddev->chunk_sectors)
NeilBrown50ac1682009-06-18 08:47:55 +10005956 return 0; /* nothing to do */
NeilBrown674806d2010-06-16 17:17:53 +10005957 if (has_failed(conf))
NeilBrownec32a2b2009-03-31 15:17:38 +11005958 return -EINVAL;
5959 if (mddev->delta_disks < 0) {
5960 /* We might be able to shrink, but the devices must
5961 * be made bigger first.
5962 * For raid6, 4 is the minimum size.
5963 * Otherwise 2 is the minimum
5964 */
5965 int min = 2;
5966 if (mddev->level == 6)
5967 min = 4;
5968 if (mddev->raid_disks + mddev->delta_disks < min)
5969 return -EINVAL;
5970 }
NeilBrown29269552006-03-27 01:18:10 -08005971
NeilBrown01ee22b2009-06-18 08:47:20 +10005972 if (!check_stripe_cache(mddev))
NeilBrown29269552006-03-27 01:18:10 -08005973 return -ENOSPC;
NeilBrown29269552006-03-27 01:18:10 -08005974
NeilBrowne56108d62012-10-11 14:24:13 +11005975 return resize_stripes(conf, (conf->previous_raid_disks
5976 + mddev->delta_disks));
NeilBrown63c70c42006-03-27 01:18:13 -08005977}
5978
NeilBrownfd01b882011-10-11 16:47:53 +11005979static int raid5_start_reshape(struct mddev *mddev)
NeilBrown63c70c42006-03-27 01:18:13 -08005980{
NeilBrownd1688a62011-10-11 16:49:52 +11005981 struct r5conf *conf = mddev->private;
NeilBrown3cb03002011-10-11 16:45:26 +11005982 struct md_rdev *rdev;
NeilBrown63c70c42006-03-27 01:18:13 -08005983 int spares = 0;
NeilBrownc04be0a2006-10-03 01:15:53 -07005984 unsigned long flags;
NeilBrown63c70c42006-03-27 01:18:13 -08005985
NeilBrownf4168852007-02-28 20:11:53 -08005986 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
NeilBrown63c70c42006-03-27 01:18:13 -08005987 return -EBUSY;
5988
NeilBrown01ee22b2009-06-18 08:47:20 +10005989 if (!check_stripe_cache(mddev))
5990 return -ENOSPC;
5991
NeilBrown30b67642012-05-22 13:55:28 +10005992 if (has_failed(conf))
5993 return -EINVAL;
5994
NeilBrownc6563a82012-05-21 09:27:00 +10005995 rdev_for_each(rdev, mddev) {
NeilBrown469518a2011-01-31 11:57:43 +11005996 if (!test_bit(In_sync, &rdev->flags)
5997 && !test_bit(Faulty, &rdev->flags))
NeilBrown29269552006-03-27 01:18:10 -08005998 spares++;
NeilBrownc6563a82012-05-21 09:27:00 +10005999 }
NeilBrown63c70c42006-03-27 01:18:13 -08006000
NeilBrownf4168852007-02-28 20:11:53 -08006001 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
NeilBrown29269552006-03-27 01:18:10 -08006002 /* Not enough devices even to make a degraded array
6003 * of that size
6004 */
6005 return -EINVAL;
6006
NeilBrownec32a2b2009-03-31 15:17:38 +11006007 /* Refuse to reduce size of the array. Any reductions in
6008 * array size must be through explicit setting of array_size
6009 * attribute.
6010 */
6011 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
6012 < mddev->array_sectors) {
NeilBrown0c55e022010-05-03 14:09:02 +10006013 printk(KERN_ERR "md/raid:%s: array size must be reduced "
NeilBrownec32a2b2009-03-31 15:17:38 +11006014 "before number of disks\n", mdname(mddev));
6015 return -EINVAL;
6016 }
6017
NeilBrownf6705572006-03-27 01:18:11 -08006018 atomic_set(&conf->reshape_stripes, 0);
NeilBrown29269552006-03-27 01:18:10 -08006019 spin_lock_irq(&conf->device_lock);
6020 conf->previous_raid_disks = conf->raid_disks;
NeilBrown63c70c42006-03-27 01:18:13 -08006021 conf->raid_disks += mddev->delta_disks;
Andre Noll09c9e5f2009-06-18 08:45:55 +10006022 conf->prev_chunk_sectors = conf->chunk_sectors;
6023 conf->chunk_sectors = mddev->new_chunk_sectors;
NeilBrown88ce4932009-03-31 15:24:23 +11006024 conf->prev_algo = conf->algorithm;
6025 conf->algorithm = mddev->new_layout;
NeilBrown05616be2012-05-21 09:27:00 +10006026 conf->generation++;
6027 /* Code that selects data_offset needs to see the generation update
6028 * if reshape_progress has been set - so a memory barrier needed.
6029 */
6030 smp_mb();
NeilBrown2c810cd2012-05-21 09:27:00 +10006031 if (mddev->reshape_backwards)
NeilBrownfef9c612009-03-31 15:16:46 +11006032 conf->reshape_progress = raid5_size(mddev, 0, 0);
6033 else
6034 conf->reshape_progress = 0;
6035 conf->reshape_safe = conf->reshape_progress;
NeilBrown29269552006-03-27 01:18:10 -08006036 spin_unlock_irq(&conf->device_lock);
6037
6038 /* Add some new drives, as many as will fit.
6039 * We know there are enough to make the newly sized array work.
NeilBrown3424bf62010-06-17 17:48:26 +10006040 * Don't add devices if we are reducing the number of
6041 * devices in the array. This is because it is not possible
6042 * to correctly record the "partially reconstructed" state of
6043 * such devices during the reshape and confusion could result.
NeilBrown29269552006-03-27 01:18:10 -08006044 */
NeilBrown87a8dec2011-01-31 11:57:43 +11006045 if (mddev->delta_disks >= 0) {
NeilBrowndafb20f2012-03-19 12:46:39 +11006046 rdev_for_each(rdev, mddev)
NeilBrown87a8dec2011-01-31 11:57:43 +11006047 if (rdev->raid_disk < 0 &&
6048 !test_bit(Faulty, &rdev->flags)) {
6049 if (raid5_add_disk(mddev, rdev) == 0) {
NeilBrown87a8dec2011-01-31 11:57:43 +11006050 if (rdev->raid_disk
NeilBrown9d4c7d82012-03-13 11:21:21 +11006051 >= conf->previous_raid_disks)
NeilBrown87a8dec2011-01-31 11:57:43 +11006052 set_bit(In_sync, &rdev->flags);
NeilBrown9d4c7d82012-03-13 11:21:21 +11006053 else
NeilBrown87a8dec2011-01-31 11:57:43 +11006054 rdev->recovery_offset = 0;
Namhyung Kim36fad852011-07-27 11:00:36 +10006055
6056 if (sysfs_link_rdev(mddev, rdev))
NeilBrown87a8dec2011-01-31 11:57:43 +11006057 /* Failure here is OK */;
NeilBrown50da0842011-01-31 11:57:43 +11006058 }
NeilBrown87a8dec2011-01-31 11:57:43 +11006059 } else if (rdev->raid_disk >= conf->previous_raid_disks
6060 && !test_bit(Faulty, &rdev->flags)) {
6061 /* This is a spare that was manually added */
6062 set_bit(In_sync, &rdev->flags);
NeilBrown87a8dec2011-01-31 11:57:43 +11006063 }
NeilBrown29269552006-03-27 01:18:10 -08006064
NeilBrown87a8dec2011-01-31 11:57:43 +11006065 /* When a reshape changes the number of devices,
6066 * ->degraded is measured against the larger of the
6067 * pre and post number of devices.
6068 */
NeilBrownec32a2b2009-03-31 15:17:38 +11006069 spin_lock_irqsave(&conf->device_lock, flags);
NeilBrown908f4fb2011-12-23 10:17:50 +11006070 mddev->degraded = calc_degraded(conf);
NeilBrownec32a2b2009-03-31 15:17:38 +11006071 spin_unlock_irqrestore(&conf->device_lock, flags);
6072 }
NeilBrown63c70c42006-03-27 01:18:13 -08006073 mddev->raid_disks = conf->raid_disks;
NeilBrowne5164022009-08-03 10:59:57 +10006074 mddev->reshape_position = conf->reshape_progress;
NeilBrown850b2b42006-10-03 01:15:46 -07006075 set_bit(MD_CHANGE_DEVS, &mddev->flags);
NeilBrownf6705572006-03-27 01:18:11 -08006076
NeilBrown29269552006-03-27 01:18:10 -08006077 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6078 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6079 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6080 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6081 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
NeilBrown0da3c612009-09-23 18:09:45 +10006082 "reshape");
NeilBrown29269552006-03-27 01:18:10 -08006083 if (!mddev->sync_thread) {
6084 mddev->recovery = 0;
6085 spin_lock_irq(&conf->device_lock);
6086 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
NeilBrown05616be2012-05-21 09:27:00 +10006087 rdev_for_each(rdev, mddev)
6088 rdev->new_data_offset = rdev->data_offset;
6089 smp_wmb();
NeilBrownfef9c612009-03-31 15:16:46 +11006090 conf->reshape_progress = MaxSector;
NeilBrown1e3fa9b2012-03-13 11:21:18 +11006091 mddev->reshape_position = MaxSector;
NeilBrown29269552006-03-27 01:18:10 -08006092 spin_unlock_irq(&conf->device_lock);
6093 return -EAGAIN;
6094 }
NeilBrownc8f517c2009-03-31 15:28:40 +11006095 conf->reshape_checkpoint = jiffies;
NeilBrown29269552006-03-27 01:18:10 -08006096 md_wakeup_thread(mddev->sync_thread);
6097 md_new_event(mddev);
6098 return 0;
6099}
NeilBrown29269552006-03-27 01:18:10 -08006100
NeilBrownec32a2b2009-03-31 15:17:38 +11006101/* This is called from the reshape thread and should make any
6102 * changes needed in 'conf'
6103 */
NeilBrownd1688a62011-10-11 16:49:52 +11006104static void end_reshape(struct r5conf *conf)
NeilBrown29269552006-03-27 01:18:10 -08006105{
NeilBrown29269552006-03-27 01:18:10 -08006106
NeilBrownf6705572006-03-27 01:18:11 -08006107 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
NeilBrown05616be2012-05-21 09:27:00 +10006108 struct md_rdev *rdev;
Dan Williams80c3a6c2009-03-17 18:10:40 -07006109
NeilBrownf6705572006-03-27 01:18:11 -08006110 spin_lock_irq(&conf->device_lock);
NeilBrowncea9c222009-03-31 15:15:05 +11006111 conf->previous_raid_disks = conf->raid_disks;
NeilBrown05616be2012-05-21 09:27:00 +10006112 rdev_for_each(rdev, conf->mddev)
6113 rdev->data_offset = rdev->new_data_offset;
6114 smp_wmb();
NeilBrownfef9c612009-03-31 15:16:46 +11006115 conf->reshape_progress = MaxSector;
NeilBrownf6705572006-03-27 01:18:11 -08006116 spin_unlock_irq(&conf->device_lock);
NeilBrownb0f9ec02009-03-31 15:27:18 +11006117 wake_up(&conf->wait_for_overlap);
NeilBrown16a53ec2006-06-26 00:27:38 -07006118
6119 /* read-ahead size must cover two whole stripes, which is
6120 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6121 */
NeilBrown4a5add42010-06-01 19:37:28 +10006122 if (conf->mddev->queue) {
NeilBrowncea9c222009-03-31 15:15:05 +11006123 int data_disks = conf->raid_disks - conf->max_degraded;
Andre Noll09c9e5f2009-06-18 08:45:55 +10006124 int stripe = data_disks * ((conf->chunk_sectors << 9)
NeilBrowncea9c222009-03-31 15:15:05 +11006125 / PAGE_SIZE);
NeilBrown16a53ec2006-06-26 00:27:38 -07006126 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6127 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6128 }
NeilBrown29269552006-03-27 01:18:10 -08006129 }
NeilBrown29269552006-03-27 01:18:10 -08006130}
6131
NeilBrownec32a2b2009-03-31 15:17:38 +11006132/* This is called from the raid5d thread with mddev_lock held.
6133 * It makes config changes to the device.
6134 */
NeilBrownfd01b882011-10-11 16:47:53 +11006135static void raid5_finish_reshape(struct mddev *mddev)
NeilBrowncea9c222009-03-31 15:15:05 +11006136{
NeilBrownd1688a62011-10-11 16:49:52 +11006137 struct r5conf *conf = mddev->private;
NeilBrowncea9c222009-03-31 15:15:05 +11006138
6139 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6140
NeilBrownec32a2b2009-03-31 15:17:38 +11006141 if (mddev->delta_disks > 0) {
6142 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6143 set_capacity(mddev->gendisk, mddev->array_sectors);
NeilBrown449aad32009-08-03 10:59:58 +10006144 revalidate_disk(mddev->gendisk);
NeilBrownec32a2b2009-03-31 15:17:38 +11006145 } else {
6146 int d;
NeilBrown908f4fb2011-12-23 10:17:50 +11006147 spin_lock_irq(&conf->device_lock);
6148 mddev->degraded = calc_degraded(conf);
6149 spin_unlock_irq(&conf->device_lock);
NeilBrownec32a2b2009-03-31 15:17:38 +11006150 for (d = conf->raid_disks ;
6151 d < conf->raid_disks - mddev->delta_disks;
NeilBrown1a67dde2009-08-13 10:41:49 +10006152 d++) {
NeilBrown3cb03002011-10-11 16:45:26 +11006153 struct md_rdev *rdev = conf->disks[d].rdev;
NeilBrownda7613b2012-05-22 13:55:33 +10006154 if (rdev)
6155 clear_bit(In_sync, &rdev->flags);
6156 rdev = conf->disks[d].replacement;
6157 if (rdev)
6158 clear_bit(In_sync, &rdev->flags);
NeilBrown1a67dde2009-08-13 10:41:49 +10006159 }
NeilBrowncea9c222009-03-31 15:15:05 +11006160 }
NeilBrown88ce4932009-03-31 15:24:23 +11006161 mddev->layout = conf->algorithm;
Andre Noll09c9e5f2009-06-18 08:45:55 +10006162 mddev->chunk_sectors = conf->chunk_sectors;
NeilBrownec32a2b2009-03-31 15:17:38 +11006163 mddev->reshape_position = MaxSector;
6164 mddev->delta_disks = 0;
NeilBrown2c810cd2012-05-21 09:27:00 +10006165 mddev->reshape_backwards = 0;
NeilBrowncea9c222009-03-31 15:15:05 +11006166 }
6167}
6168
NeilBrownfd01b882011-10-11 16:47:53 +11006169static void raid5_quiesce(struct mddev *mddev, int state)
NeilBrown72626682005-09-09 16:23:54 -07006170{
NeilBrownd1688a62011-10-11 16:49:52 +11006171 struct r5conf *conf = mddev->private;
NeilBrown72626682005-09-09 16:23:54 -07006172
6173 switch(state) {
NeilBrowne464eaf2006-03-27 01:18:14 -08006174 case 2: /* resume for a suspend */
6175 wake_up(&conf->wait_for_overlap);
6176 break;
6177
NeilBrown72626682005-09-09 16:23:54 -07006178 case 1: /* stop all writes */
6179 spin_lock_irq(&conf->device_lock);
NeilBrown64bd6602009-08-03 10:59:58 +10006180 /* '2' tells resync/reshape to pause so that all
6181 * active stripes can drain
6182 */
6183 conf->quiesce = 2;
NeilBrown72626682005-09-09 16:23:54 -07006184 wait_event_lock_irq(conf->wait_for_stripe,
Raz Ben-Jehuda(caro)46031f92006-12-10 02:20:47 -08006185 atomic_read(&conf->active_stripes) == 0 &&
6186 atomic_read(&conf->active_aligned_reads) == 0,
Lukas Czernereed8c022012-11-30 11:42:40 +01006187 conf->device_lock);
NeilBrown64bd6602009-08-03 10:59:58 +10006188 conf->quiesce = 1;
NeilBrown72626682005-09-09 16:23:54 -07006189 spin_unlock_irq(&conf->device_lock);
NeilBrown64bd6602009-08-03 10:59:58 +10006190 /* allow reshape to continue */
6191 wake_up(&conf->wait_for_overlap);
NeilBrown72626682005-09-09 16:23:54 -07006192 break;
6193
6194 case 0: /* re-enable writes */
6195 spin_lock_irq(&conf->device_lock);
6196 conf->quiesce = 0;
6197 wake_up(&conf->wait_for_stripe);
NeilBrowne464eaf2006-03-27 01:18:14 -08006198 wake_up(&conf->wait_for_overlap);
NeilBrown72626682005-09-09 16:23:54 -07006199 spin_unlock_irq(&conf->device_lock);
6200 break;
6201 }
NeilBrown72626682005-09-09 16:23:54 -07006202}
NeilBrownb15c2e52006-01-06 00:20:16 -08006203
NeilBrownd562b0c2009-03-31 14:39:39 +11006204
NeilBrownfd01b882011-10-11 16:47:53 +11006205static void *raid45_takeover_raid0(struct mddev *mddev, int level)
Trela Maciej54071b32010-03-08 16:02:42 +11006206{
NeilBrowne373ab12011-10-11 16:48:59 +11006207 struct r0conf *raid0_conf = mddev->private;
Randy Dunlapd76c8422011-04-21 09:07:26 -07006208 sector_t sectors;
Trela Maciej54071b32010-03-08 16:02:42 +11006209
Dan Williamsf1b29bc2010-05-01 18:09:05 -07006210 /* for raid0 takeover only one zone is supported */
NeilBrowne373ab12011-10-11 16:48:59 +11006211 if (raid0_conf->nr_strip_zones > 1) {
NeilBrown0c55e022010-05-03 14:09:02 +10006212 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6213 mdname(mddev));
Dan Williamsf1b29bc2010-05-01 18:09:05 -07006214 return ERR_PTR(-EINVAL);
6215 }
6216
NeilBrowne373ab12011-10-11 16:48:59 +11006217 sectors = raid0_conf->strip_zone[0].zone_end;
6218 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
NeilBrown3b71bd92011-04-20 15:38:18 +10006219 mddev->dev_sectors = sectors;
Dan Williamsf1b29bc2010-05-01 18:09:05 -07006220 mddev->new_level = level;
Trela Maciej54071b32010-03-08 16:02:42 +11006221 mddev->new_layout = ALGORITHM_PARITY_N;
6222 mddev->new_chunk_sectors = mddev->chunk_sectors;
6223 mddev->raid_disks += 1;
6224 mddev->delta_disks = 1;
6225 /* make sure it will be not marked as dirty */
6226 mddev->recovery_cp = MaxSector;
6227
6228 return setup_conf(mddev);
6229}
6230
6231
NeilBrownfd01b882011-10-11 16:47:53 +11006232static void *raid5_takeover_raid1(struct mddev *mddev)
NeilBrownd562b0c2009-03-31 14:39:39 +11006233{
6234 int chunksect;
6235
6236 if (mddev->raid_disks != 2 ||
6237 mddev->degraded > 1)
6238 return ERR_PTR(-EINVAL);
6239
6240 /* Should check if there are write-behind devices? */
6241
6242 chunksect = 64*2; /* 64K by default */
6243
6244 /* The array must be an exact multiple of chunksize */
6245 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6246 chunksect >>= 1;
6247
6248 if ((chunksect<<9) < STRIPE_SIZE)
6249 /* array size does not allow a suitable chunk size */
6250 return ERR_PTR(-EINVAL);
6251
6252 mddev->new_level = 5;
6253 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
Andre Noll664e7c42009-06-18 08:45:27 +10006254 mddev->new_chunk_sectors = chunksect;
NeilBrownd562b0c2009-03-31 14:39:39 +11006255
6256 return setup_conf(mddev);
6257}
6258
NeilBrownfd01b882011-10-11 16:47:53 +11006259static void *raid5_takeover_raid6(struct mddev *mddev)
NeilBrownfc9739c2009-03-31 14:57:20 +11006260{
6261 int new_layout;
6262
6263 switch (mddev->layout) {
6264 case ALGORITHM_LEFT_ASYMMETRIC_6:
6265 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6266 break;
6267 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6268 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6269 break;
6270 case ALGORITHM_LEFT_SYMMETRIC_6:
6271 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6272 break;
6273 case ALGORITHM_RIGHT_SYMMETRIC_6:
6274 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6275 break;
6276 case ALGORITHM_PARITY_0_6:
6277 new_layout = ALGORITHM_PARITY_0;
6278 break;
6279 case ALGORITHM_PARITY_N:
6280 new_layout = ALGORITHM_PARITY_N;
6281 break;
6282 default:
6283 return ERR_PTR(-EINVAL);
6284 }
6285 mddev->new_level = 5;
6286 mddev->new_layout = new_layout;
6287 mddev->delta_disks = -1;
6288 mddev->raid_disks -= 1;
6289 return setup_conf(mddev);
6290}
6291
NeilBrownd562b0c2009-03-31 14:39:39 +11006292
NeilBrownfd01b882011-10-11 16:47:53 +11006293static int raid5_check_reshape(struct mddev *mddev)
NeilBrownb3546032009-03-31 14:56:41 +11006294{
NeilBrown88ce4932009-03-31 15:24:23 +11006295 /* For a 2-drive array, the layout and chunk size can be changed
6296 * immediately as not restriping is needed.
6297 * For larger arrays we record the new value - after validation
6298 * to be used by a reshape pass.
NeilBrownb3546032009-03-31 14:56:41 +11006299 */
NeilBrownd1688a62011-10-11 16:49:52 +11006300 struct r5conf *conf = mddev->private;
NeilBrown597a7112009-06-18 08:47:42 +10006301 int new_chunk = mddev->new_chunk_sectors;
NeilBrownb3546032009-03-31 14:56:41 +11006302
NeilBrown597a7112009-06-18 08:47:42 +10006303 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
NeilBrownb3546032009-03-31 14:56:41 +11006304 return -EINVAL;
6305 if (new_chunk > 0) {
Andre Noll0ba459d2009-06-18 08:46:10 +10006306 if (!is_power_of_2(new_chunk))
NeilBrownb3546032009-03-31 14:56:41 +11006307 return -EINVAL;
NeilBrown597a7112009-06-18 08:47:42 +10006308 if (new_chunk < (PAGE_SIZE>>9))
NeilBrownb3546032009-03-31 14:56:41 +11006309 return -EINVAL;
NeilBrown597a7112009-06-18 08:47:42 +10006310 if (mddev->array_sectors & (new_chunk-1))
NeilBrownb3546032009-03-31 14:56:41 +11006311 /* not factor of array size */
6312 return -EINVAL;
6313 }
6314
6315 /* They look valid */
6316
NeilBrown88ce4932009-03-31 15:24:23 +11006317 if (mddev->raid_disks == 2) {
NeilBrown597a7112009-06-18 08:47:42 +10006318 /* can make the change immediately */
6319 if (mddev->new_layout >= 0) {
6320 conf->algorithm = mddev->new_layout;
6321 mddev->layout = mddev->new_layout;
NeilBrown88ce4932009-03-31 15:24:23 +11006322 }
6323 if (new_chunk > 0) {
NeilBrown597a7112009-06-18 08:47:42 +10006324 conf->chunk_sectors = new_chunk ;
6325 mddev->chunk_sectors = new_chunk;
NeilBrown88ce4932009-03-31 15:24:23 +11006326 }
6327 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6328 md_wakeup_thread(mddev->thread);
NeilBrownb3546032009-03-31 14:56:41 +11006329 }
NeilBrown50ac1682009-06-18 08:47:55 +10006330 return check_reshape(mddev);
NeilBrown88ce4932009-03-31 15:24:23 +11006331}
6332
NeilBrownfd01b882011-10-11 16:47:53 +11006333static int raid6_check_reshape(struct mddev *mddev)
NeilBrown88ce4932009-03-31 15:24:23 +11006334{
NeilBrown597a7112009-06-18 08:47:42 +10006335 int new_chunk = mddev->new_chunk_sectors;
NeilBrown50ac1682009-06-18 08:47:55 +10006336
NeilBrown597a7112009-06-18 08:47:42 +10006337 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
NeilBrown88ce4932009-03-31 15:24:23 +11006338 return -EINVAL;
NeilBrownb3546032009-03-31 14:56:41 +11006339 if (new_chunk > 0) {
Andre Noll0ba459d2009-06-18 08:46:10 +10006340 if (!is_power_of_2(new_chunk))
NeilBrown88ce4932009-03-31 15:24:23 +11006341 return -EINVAL;
NeilBrown597a7112009-06-18 08:47:42 +10006342 if (new_chunk < (PAGE_SIZE >> 9))
NeilBrown88ce4932009-03-31 15:24:23 +11006343 return -EINVAL;
NeilBrown597a7112009-06-18 08:47:42 +10006344 if (mddev->array_sectors & (new_chunk-1))
NeilBrown88ce4932009-03-31 15:24:23 +11006345 /* not factor of array size */
6346 return -EINVAL;
NeilBrownb3546032009-03-31 14:56:41 +11006347 }
NeilBrown88ce4932009-03-31 15:24:23 +11006348
6349 /* They look valid */
NeilBrown50ac1682009-06-18 08:47:55 +10006350 return check_reshape(mddev);
NeilBrownb3546032009-03-31 14:56:41 +11006351}
6352
NeilBrownfd01b882011-10-11 16:47:53 +11006353static void *raid5_takeover(struct mddev *mddev)
NeilBrownd562b0c2009-03-31 14:39:39 +11006354{
6355 /* raid5 can take over:
Dan Williamsf1b29bc2010-05-01 18:09:05 -07006356 * raid0 - if there is only one strip zone - make it a raid4 layout
NeilBrownd562b0c2009-03-31 14:39:39 +11006357 * raid1 - if there are two drives. We need to know the chunk size
6358 * raid4 - trivial - just use a raid4 layout.
6359 * raid6 - Providing it is a *_6 layout
NeilBrownd562b0c2009-03-31 14:39:39 +11006360 */
Dan Williamsf1b29bc2010-05-01 18:09:05 -07006361 if (mddev->level == 0)
6362 return raid45_takeover_raid0(mddev, 5);
NeilBrownd562b0c2009-03-31 14:39:39 +11006363 if (mddev->level == 1)
6364 return raid5_takeover_raid1(mddev);
NeilBrowne9d47582009-03-31 14:57:09 +11006365 if (mddev->level == 4) {
6366 mddev->new_layout = ALGORITHM_PARITY_N;
6367 mddev->new_level = 5;
6368 return setup_conf(mddev);
6369 }
NeilBrownfc9739c2009-03-31 14:57:20 +11006370 if (mddev->level == 6)
6371 return raid5_takeover_raid6(mddev);
NeilBrownd562b0c2009-03-31 14:39:39 +11006372
6373 return ERR_PTR(-EINVAL);
6374}
6375
NeilBrownfd01b882011-10-11 16:47:53 +11006376static void *raid4_takeover(struct mddev *mddev)
NeilBrowna78d38a2010-03-22 16:53:49 +11006377{
Dan Williamsf1b29bc2010-05-01 18:09:05 -07006378 /* raid4 can take over:
6379 * raid0 - if there is only one strip zone
6380 * raid5 - if layout is right
NeilBrowna78d38a2010-03-22 16:53:49 +11006381 */
Dan Williamsf1b29bc2010-05-01 18:09:05 -07006382 if (mddev->level == 0)
6383 return raid45_takeover_raid0(mddev, 4);
NeilBrowna78d38a2010-03-22 16:53:49 +11006384 if (mddev->level == 5 &&
6385 mddev->layout == ALGORITHM_PARITY_N) {
6386 mddev->new_layout = 0;
6387 mddev->new_level = 4;
6388 return setup_conf(mddev);
6389 }
6390 return ERR_PTR(-EINVAL);
6391}
NeilBrownd562b0c2009-03-31 14:39:39 +11006392
NeilBrown84fc4b52011-10-11 16:49:58 +11006393static struct md_personality raid5_personality;
NeilBrown245f46c2009-03-31 14:39:39 +11006394
NeilBrownfd01b882011-10-11 16:47:53 +11006395static void *raid6_takeover(struct mddev *mddev)
NeilBrown245f46c2009-03-31 14:39:39 +11006396{
6397 /* Currently can only take over a raid5. We map the
6398 * personality to an equivalent raid6 personality
6399 * with the Q block at the end.
6400 */
6401 int new_layout;
6402
6403 if (mddev->pers != &raid5_personality)
6404 return ERR_PTR(-EINVAL);
6405 if (mddev->degraded > 1)
6406 return ERR_PTR(-EINVAL);
6407 if (mddev->raid_disks > 253)
6408 return ERR_PTR(-EINVAL);
6409 if (mddev->raid_disks < 3)
6410 return ERR_PTR(-EINVAL);
6411
6412 switch (mddev->layout) {
6413 case ALGORITHM_LEFT_ASYMMETRIC:
6414 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6415 break;
6416 case ALGORITHM_RIGHT_ASYMMETRIC:
6417 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6418 break;
6419 case ALGORITHM_LEFT_SYMMETRIC:
6420 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6421 break;
6422 case ALGORITHM_RIGHT_SYMMETRIC:
6423 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6424 break;
6425 case ALGORITHM_PARITY_0:
6426 new_layout = ALGORITHM_PARITY_0_6;
6427 break;
6428 case ALGORITHM_PARITY_N:
6429 new_layout = ALGORITHM_PARITY_N;
6430 break;
6431 default:
6432 return ERR_PTR(-EINVAL);
6433 }
6434 mddev->new_level = 6;
6435 mddev->new_layout = new_layout;
6436 mddev->delta_disks = 1;
6437 mddev->raid_disks += 1;
6438 return setup_conf(mddev);
6439}
6440
6441
NeilBrown84fc4b52011-10-11 16:49:58 +11006442static struct md_personality raid6_personality =
NeilBrown16a53ec2006-06-26 00:27:38 -07006443{
6444 .name = "raid6",
6445 .level = 6,
6446 .owner = THIS_MODULE,
6447 .make_request = make_request,
6448 .run = run,
6449 .stop = stop,
6450 .status = status,
6451 .error_handler = error,
6452 .hot_add_disk = raid5_add_disk,
6453 .hot_remove_disk= raid5_remove_disk,
6454 .spare_active = raid5_spare_active,
6455 .sync_request = sync_request,
6456 .resize = raid5_resize,
Dan Williams80c3a6c2009-03-17 18:10:40 -07006457 .size = raid5_size,
NeilBrown50ac1682009-06-18 08:47:55 +10006458 .check_reshape = raid6_check_reshape,
NeilBrownf4168852007-02-28 20:11:53 -08006459 .start_reshape = raid5_start_reshape,
NeilBrowncea9c222009-03-31 15:15:05 +11006460 .finish_reshape = raid5_finish_reshape,
NeilBrown16a53ec2006-06-26 00:27:38 -07006461 .quiesce = raid5_quiesce,
NeilBrown245f46c2009-03-31 14:39:39 +11006462 .takeover = raid6_takeover,
NeilBrown16a53ec2006-06-26 00:27:38 -07006463};
NeilBrown84fc4b52011-10-11 16:49:58 +11006464static struct md_personality raid5_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07006465{
6466 .name = "raid5",
NeilBrown2604b702006-01-06 00:20:36 -08006467 .level = 5,
Linus Torvalds1da177e2005-04-16 15:20:36 -07006468 .owner = THIS_MODULE,
6469 .make_request = make_request,
6470 .run = run,
6471 .stop = stop,
6472 .status = status,
6473 .error_handler = error,
6474 .hot_add_disk = raid5_add_disk,
6475 .hot_remove_disk= raid5_remove_disk,
6476 .spare_active = raid5_spare_active,
6477 .sync_request = sync_request,
6478 .resize = raid5_resize,
Dan Williams80c3a6c2009-03-17 18:10:40 -07006479 .size = raid5_size,
NeilBrown63c70c42006-03-27 01:18:13 -08006480 .check_reshape = raid5_check_reshape,
6481 .start_reshape = raid5_start_reshape,
NeilBrowncea9c222009-03-31 15:15:05 +11006482 .finish_reshape = raid5_finish_reshape,
NeilBrown72626682005-09-09 16:23:54 -07006483 .quiesce = raid5_quiesce,
NeilBrownd562b0c2009-03-31 14:39:39 +11006484 .takeover = raid5_takeover,
Linus Torvalds1da177e2005-04-16 15:20:36 -07006485};
6486
NeilBrown84fc4b52011-10-11 16:49:58 +11006487static struct md_personality raid4_personality =
Linus Torvalds1da177e2005-04-16 15:20:36 -07006488{
NeilBrown2604b702006-01-06 00:20:36 -08006489 .name = "raid4",
6490 .level = 4,
6491 .owner = THIS_MODULE,
6492 .make_request = make_request,
6493 .run = run,
6494 .stop = stop,
6495 .status = status,
6496 .error_handler = error,
6497 .hot_add_disk = raid5_add_disk,
6498 .hot_remove_disk= raid5_remove_disk,
6499 .spare_active = raid5_spare_active,
6500 .sync_request = sync_request,
6501 .resize = raid5_resize,
Dan Williams80c3a6c2009-03-17 18:10:40 -07006502 .size = raid5_size,
NeilBrown3d378902007-03-26 21:32:13 -08006503 .check_reshape = raid5_check_reshape,
6504 .start_reshape = raid5_start_reshape,
NeilBrowncea9c222009-03-31 15:15:05 +11006505 .finish_reshape = raid5_finish_reshape,
NeilBrown2604b702006-01-06 00:20:36 -08006506 .quiesce = raid5_quiesce,
NeilBrowna78d38a2010-03-22 16:53:49 +11006507 .takeover = raid4_takeover,
NeilBrown2604b702006-01-06 00:20:36 -08006508};
6509
6510static int __init raid5_init(void)
6511{
NeilBrown16a53ec2006-06-26 00:27:38 -07006512 register_md_personality(&raid6_personality);
NeilBrown2604b702006-01-06 00:20:36 -08006513 register_md_personality(&raid5_personality);
6514 register_md_personality(&raid4_personality);
6515 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006516}
6517
NeilBrown2604b702006-01-06 00:20:36 -08006518static void raid5_exit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006519{
NeilBrown16a53ec2006-06-26 00:27:38 -07006520 unregister_md_personality(&raid6_personality);
NeilBrown2604b702006-01-06 00:20:36 -08006521 unregister_md_personality(&raid5_personality);
6522 unregister_md_personality(&raid4_personality);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006523}
6524
6525module_init(raid5_init);
6526module_exit(raid5_exit);
6527MODULE_LICENSE("GPL");
NeilBrown0efb9e62009-12-14 12:49:58 +11006528MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006529MODULE_ALIAS("md-personality-4"); /* RAID5 */
NeilBrownd9d166c2006-01-06 00:20:51 -08006530MODULE_ALIAS("md-raid5");
6531MODULE_ALIAS("md-raid4");
NeilBrown2604b702006-01-06 00:20:36 -08006532MODULE_ALIAS("md-level-5");
6533MODULE_ALIAS("md-level-4");
NeilBrown16a53ec2006-06-26 00:27:38 -07006534MODULE_ALIAS("md-personality-8"); /* RAID6 */
6535MODULE_ALIAS("md-raid6");
6536MODULE_ALIAS("md-level-6");
6537
6538/* This used to be two separate modules, they were: */
6539MODULE_ALIAS("raid5");
6540MODULE_ALIAS("raid6");