projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
md: split detach operation out from ->stop.
[linux-2.6-block.git] / drivers / md / raid5.c
CommitLineData
1da177e4
LT		 1/*
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
16a53ecc 5 * Copyright (C) 2002, 2003 H. Peter Anvin
1da177e4 6 *
16a53ecc
N		 7 * RAID-4/5/6 management functions.
8 * Thanks to Penguin Computing for making the RAID-6 development possible
9 * by donating a test server!
1da177e4
LT		 10 *
11 * This program is free software; you can redistribute it and/or modify
12 * it under the terms of the GNU General Public License as published by
13 * the Free Software Foundation; either version 2, or (at your option)
14 * any later version.
15 *
16 * You should have received a copy of the GNU General Public License
17 * (for example /usr/src/linux/COPYING); if not, write to the Free
18 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19 */
20
ae3c20cc
N		 21/*
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.
7c13edc8
N		 30 * 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
ae3c20cc
N		 32 * 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
7c13edc8 35 * the number of the batch it will be in. This is seq_flush+1.
ae3c20cc
N		 36 * 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 */
1da177e4 45
bff61975 46#include <linux/blkdev.h>
f6705578 47#include <linux/kthread.h>
f701d589 48#include <linux/raid/pq.h>
91c00924 49#include <linux/async_tx.h>
056075c7 50#include <linux/module.h>
07a3b417 51#include <linux/async.h>
bff61975 52#include <linux/seq_file.h>
36d1c647 53#include <linux/cpu.h>
5a0e3ad6 54#include <linux/slab.h>
8bda470e 55#include <linux/ratelimit.h>
851c30c9 56#include <linux/nodemask.h>
a9add5d9
N		 57#include <trace/events/block.h>
58
43b2e5d8 59#include "md.h"
bff61975 60#include "raid5.h"
54071b38 61#include "raid0.h"
ef740c37 62#include "bitmap.h"
72626685 63
851c30c9
SL		 64#define cpu_to_group(cpu) cpu_to_node(cpu)
65#define ANY_GROUP NUMA_NO_NODE
66
8e0e99ba
N		 67static bool devices_handle_discard_safely = false;
68module_param(devices_handle_discard_safely, bool, 0644);
69MODULE_PARM_DESC(devices_handle_discard_safely,
70 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
851c30c9 71static struct workqueue_struct *raid5_wq;
1da177e4
LT		 72/*
73 * Stripe cache
74 */
75
76#define NR_STRIPES 256
77#define STRIPE_SIZE PAGE_SIZE
78#define STRIPE_SHIFT (PAGE_SHIFT - 9)
79#define STRIPE_SECTORS (STRIPE_SIZE>>9)
80#define IO_THRESHOLD 1
8b3e6cdc 81#define BYPASS_THRESHOLD 1
fccddba0 82#define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
1da177e4 83#define HASH_MASK (NR_HASH - 1)
bfc90cb0 84#define MAX_STRIPE_BATCH 8
1da177e4 85
d1688a6d 86static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
db298e19
N		 87{
88 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
89 return &conf->stripe_hashtbl[hash];
90}
1da177e4 91
566c09c5
SL		 92static inline int stripe_hash_locks_hash(sector_t sect)
93{
94 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
95}
96
97static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
98{
99 spin_lock_irq(conf->hash_locks + hash);
100 spin_lock(&conf->device_lock);
101}
102
103static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
104{
105 spin_unlock(&conf->device_lock);
106 spin_unlock_irq(conf->hash_locks + hash);
107}
108
109static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
110{
111 int i;
112 local_irq_disable();
113 spin_lock(conf->hash_locks);
114 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
115 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
116 spin_lock(&conf->device_lock);
117}
118
119static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
120{
121 int i;
122 spin_unlock(&conf->device_lock);
123 for (i = NR_STRIPE_HASH_LOCKS; i; i--)
124 spin_unlock(conf->hash_locks + i - 1);
125 local_irq_enable();
126}
127
1da177e4
LT		 128/* bio's attached to a stripe+device for I/O are linked together in bi_sector
129 * order without overlap. There may be several bio's per stripe+device, and
130 * a bio could span several devices.
131 * When walking this list for a particular stripe+device, we must never proceed
132 * beyond a bio that extends past this device, as the next bio might no longer
133 * be valid.
db298e19 134 * This function is used to determine the 'next' bio in the list, given the sector
1da177e4
LT		 135 * of the current stripe+device
136 */
db298e19
N		 137static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
138{
aa8b57aa 139 int sectors = bio_sectors(bio);
4f024f37 140 if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
db298e19
N		 141 return bio->bi_next;
142 else
143 return NULL;
144}
1da177e4 145
960e739d 146/*
5b99c2ff
JA		 147 * We maintain a biased count of active stripes in the bottom 16 bits of
148 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
960e739d 149 */
e7836bd6 150static inline int raid5_bi_processed_stripes(struct bio *bio)
960e739d 151{
e7836bd6
SL		 152 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
153 return (atomic_read(segments) >> 16) & 0xffff;
960e739d
JA		 154}
155
e7836bd6 156static inline int raid5_dec_bi_active_stripes(struct bio *bio)
960e739d 157{
e7836bd6
SL		 158 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
159 return atomic_sub_return(1, segments) & 0xffff;
960e739d
JA		 160}
161
e7836bd6 162static inline void raid5_inc_bi_active_stripes(struct bio *bio)
960e739d 163{
e7836bd6
SL		 164 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
165 atomic_inc(segments);
960e739d
JA		 166}
167
e7836bd6
SL		 168static inline void raid5_set_bi_processed_stripes(struct bio *bio,
169 unsigned int cnt)
960e739d 170{
e7836bd6
SL		 171 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
172 int old, new;
960e739d 173
e7836bd6
SL		 174 do {
175 old = atomic_read(segments);
176 new = (old & 0xffff) | (cnt << 16);
177 } while (atomic_cmpxchg(segments, old, new) != old);
960e739d
JA		 178}
179
e7836bd6 180static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
960e739d 181{
e7836bd6
SL		 182 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
183 atomic_set(segments, cnt);
960e739d
JA		 184}
185
d0dabf7e
N		 186/* Find first data disk in a raid6 stripe */
187static inline int raid6_d0(struct stripe_head *sh)
188{
67cc2b81
N		 189 if (sh->ddf_layout)
190 /* ddf always start from first device */
191 return 0;
192 /* md starts just after Q block */
d0dabf7e
N		 193 if (sh->qd_idx == sh->disks - 1)
194 return 0;
195 else
196 return sh->qd_idx + 1;
197}
16a53ecc
N		 198static inline int raid6_next_disk(int disk, int raid_disks)
199{
200 disk++;
201 return (disk < raid_disks) ? disk : 0;
202}
a4456856 203
d0dabf7e
N		 204/* When walking through the disks in a raid5, starting at raid6_d0,
205 * We need to map each disk to a 'slot', where the data disks are slot
206 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
207 * is raid_disks-1. This help does that mapping.
208 */
67cc2b81
N		 209static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
210 int *count, int syndrome_disks)
d0dabf7e 211{
6629542e 212 int slot = *count;
67cc2b81 213
e4424fee 214 if (sh->ddf_layout)
6629542e 215 (*count)++;
d0dabf7e 216 if (idx == sh->pd_idx)
67cc2b81 217 return syndrome_disks;
d0dabf7e 218 if (idx == sh->qd_idx)
67cc2b81 219 return syndrome_disks + 1;
e4424fee 220 if (!sh->ddf_layout)
6629542e 221 (*count)++;
d0dabf7e
N		 222 return slot;
223}
224
a4456856
DW		 225static void return_io(struct bio *return_bi)
226{
227 struct bio *bi = return_bi;
228 while (bi) {
a4456856
DW		 229
230 return_bi = bi->bi_next;
231 bi->bi_next = NULL;
4f024f37 232 bi->bi_iter.bi_size = 0;
0a82a8d1
LT		 233 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
234 bi, 0);
0e13fe23 235 bio_endio(bi, 0);
a4456856
DW		 236 bi = return_bi;
237 }
238}
239
d1688a6d 240static void print_raid5_conf (struct r5conf *conf);
1da177e4 241
600aa109
DW		 242static int stripe_operations_active(struct stripe_head *sh)
243{
244 return sh->check_state || sh->reconstruct_state ||
245 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
246 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
247}
248
851c30c9
SL		 249static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
250{
251 struct r5conf *conf = sh->raid_conf;
252 struct r5worker_group *group;
bfc90cb0 253 int thread_cnt;
851c30c9
SL		 254 int i, cpu = sh->cpu;
255
256 if (!cpu_online(cpu)) {
257 cpu = cpumask_any(cpu_online_mask);
258 sh->cpu = cpu;
259 }
260
261 if (list_empty(&sh->lru)) {
262 struct r5worker_group *group;
263 group = conf->worker_groups + cpu_to_group(cpu);
264 list_add_tail(&sh->lru, &group->handle_list);
bfc90cb0
SL		 265 group->stripes_cnt++;
266 sh->group = group;
851c30c9
SL		 267 }
268
269 if (conf->worker_cnt_per_group == 0) {
270 md_wakeup_thread(conf->mddev->thread);
271 return;
272 }
273
274 group = conf->worker_groups + cpu_to_group(sh->cpu);
275
bfc90cb0
SL		 276 group->workers[0].working = true;
277 /* at least one worker should run to avoid race */
278 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
279
280 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
281 /* wakeup more workers */
282 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
283 if (group->workers[i].working == false) {
284 group->workers[i].working = true;
285 queue_work_on(sh->cpu, raid5_wq,
286 &group->workers[i].work);
287 thread_cnt--;
288 }
289 }
851c30c9
SL		 290}
291
566c09c5
SL		 292static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
293 struct list_head *temp_inactive_list)
1da177e4 294{
4eb788df
SL		 295 BUG_ON(!list_empty(&sh->lru));
296 BUG_ON(atomic_read(&conf->active_stripes)==0);
297 if (test_bit(STRIPE_HANDLE, &sh->state)) {
298 if (test_bit(STRIPE_DELAYED, &sh->state) &&
ad3ab8b6 299 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4eb788df 300 list_add_tail(&sh->lru, &conf->delayed_list);
ad3ab8b6 301 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
4eb788df
SL		 302 sh->bm_seq - conf->seq_write > 0)
303 list_add_tail(&sh->lru, &conf->bitmap_list);
304 else {
305 clear_bit(STRIPE_DELAYED, &sh->state);
306 clear_bit(STRIPE_BIT_DELAY, &sh->state);
851c30c9
SL		 307 if (conf->worker_cnt_per_group == 0) {
308 list_add_tail(&sh->lru, &conf->handle_list);
309 } else {
310 raid5_wakeup_stripe_thread(sh);
311 return;
312 }
4eb788df
SL		 313 }
314 md_wakeup_thread(conf->mddev->thread);
315 } else {
316 BUG_ON(stripe_operations_active(sh));
317 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
318 if (atomic_dec_return(&conf->preread_active_stripes)
319 < IO_THRESHOLD)
320 md_wakeup_thread(conf->mddev->thread);
321 atomic_dec(&conf->active_stripes);
566c09c5
SL		 322 if (!test_bit(STRIPE_EXPANDING, &sh->state))
323 list_add_tail(&sh->lru, temp_inactive_list);
1da177e4
LT		 324 }
325}
d0dabf7e 326
566c09c5
SL		 327static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
328 struct list_head *temp_inactive_list)
4eb788df
SL		 329{
330 if (atomic_dec_and_test(&sh->count))
566c09c5
SL		 331 do_release_stripe(conf, sh, temp_inactive_list);
332}
333
334/*
335 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
336 *
337 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
338 * given time. Adding stripes only takes device lock, while deleting stripes
339 * only takes hash lock.
340 */
341static void release_inactive_stripe_list(struct r5conf *conf,
342 struct list_head *temp_inactive_list,
343 int hash)
344{
345 int size;
346 bool do_wakeup = false;
347 unsigned long flags;
348
349 if (hash == NR_STRIPE_HASH_LOCKS) {
350 size = NR_STRIPE_HASH_LOCKS;
351 hash = NR_STRIPE_HASH_LOCKS - 1;
352 } else
353 size = 1;
354 while (size) {
355 struct list_head *list = &temp_inactive_list[size - 1];
356
357 /*
358 * We don't hold any lock here yet, get_active_stripe() might
359 * remove stripes from the list
360 */
361 if (!list_empty_careful(list)) {
362 spin_lock_irqsave(conf->hash_locks + hash, flags);
4bda556a
SL		 363 if (list_empty(conf->inactive_list + hash) &&
364 !list_empty(list))
365 atomic_dec(&conf->empty_inactive_list_nr);
566c09c5
SL		 366 list_splice_tail_init(list, conf->inactive_list + hash);
367 do_wakeup = true;
368 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
369 }
370 size--;
371 hash--;
372 }
373
374 if (do_wakeup) {
375 wake_up(&conf->wait_for_stripe);
376 if (conf->retry_read_aligned)
377 md_wakeup_thread(conf->mddev->thread);
378 }
4eb788df
SL		 379}
380
773ca82f 381/* should hold conf->device_lock already */
566c09c5
SL		 382static int release_stripe_list(struct r5conf *conf,
383 struct list_head *temp_inactive_list)
773ca82f
SL		 384{
385 struct stripe_head *sh;
386 int count = 0;
387 struct llist_node *head;
388
389 head = llist_del_all(&conf->released_stripes);
d265d9dc 390 head = llist_reverse_order(head);
773ca82f 391 while (head) {
566c09c5
SL		 392 int hash;
393
773ca82f
SL		 394 sh = llist_entry(head, struct stripe_head, release_list);
395 head = llist_next(head);
396 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
397 smp_mb();
398 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
399 /*
400 * Don't worry the bit is set here, because if the bit is set
401 * again, the count is always > 1. This is true for
402 * STRIPE_ON_UNPLUG_LIST bit too.
403 */
566c09c5
SL		 404 hash = sh->hash_lock_index;
405 __release_stripe(conf, sh, &temp_inactive_list[hash]);
773ca82f
SL		 406 count++;
407 }
408
409 return count;
410}
411
1da177e4
LT		 412static void release_stripe(struct stripe_head *sh)
413{
d1688a6d 414 struct r5conf *conf = sh->raid_conf;
1da177e4 415 unsigned long flags;
566c09c5
SL		 416 struct list_head list;
417 int hash;
773ca82f 418 bool wakeup;
16a53ecc 419
cf170f3f
ES		 420 /* Avoid release_list until the last reference.
421 */
422 if (atomic_add_unless(&sh->count, -1, 1))
423 return;
424
ad4068de 425 if (unlikely(!conf->mddev->thread) ||
426 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
773ca82f
SL		 427 goto slow_path;
428 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
429 if (wakeup)
430 md_wakeup_thread(conf->mddev->thread);
431 return;
432slow_path:
4eb788df 433 local_irq_save(flags);
773ca82f 434 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
4eb788df 435 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
566c09c5
SL		 436 INIT_LIST_HEAD(&list);
437 hash = sh->hash_lock_index;
438 do_release_stripe(conf, sh, &list);
4eb788df 439 spin_unlock(&conf->device_lock);
566c09c5 440 release_inactive_stripe_list(conf, &list, hash);
4eb788df
SL		 441 }
442 local_irq_restore(flags);
1da177e4
LT		 443}
444
fccddba0 445static inline void remove_hash(struct stripe_head *sh)
1da177e4 446{
45b4233c
DW		 447 pr_debug("remove_hash(), stripe %llu\n",
448 (unsigned long long)sh->sector);
1da177e4 449
fccddba0 450 hlist_del_init(&sh->hash);
1da177e4
LT		 451}
452
d1688a6d 453static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
1da177e4 454{
fccddba0 455 struct hlist_head *hp = stripe_hash(conf, sh->sector);
1da177e4 456
45b4233c
DW		 457 pr_debug("insert_hash(), stripe %llu\n",
458 (unsigned long long)sh->sector);
1da177e4 459
fccddba0 460 hlist_add_head(&sh->hash, hp);
1da177e4
LT		 461}
462
1da177e4 463/* find an idle stripe, make sure it is unhashed, and return it. */
566c09c5 464static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
1da177e4
LT		 465{
466 struct stripe_head *sh = NULL;
467 struct list_head *first;
468
566c09c5 469 if (list_empty(conf->inactive_list + hash))
1da177e4 470 goto out;
566c09c5 471 first = (conf->inactive_list + hash)->next;
1da177e4
LT		 472 sh = list_entry(first, struct stripe_head, lru);
473 list_del_init(first);
474 remove_hash(sh);
475 atomic_inc(&conf->active_stripes);
566c09c5 476 BUG_ON(hash != sh->hash_lock_index);
4bda556a
SL		 477 if (list_empty(conf->inactive_list + hash))
478 atomic_inc(&conf->empty_inactive_list_nr);
1da177e4
LT		 479out:
480 return sh;
481}
482
e4e11e38 483static void shrink_buffers(struct stripe_head *sh)
1da177e4
LT		 484{
485 struct page *p;
486 int i;
e4e11e38 487 int num = sh->raid_conf->pool_size;
1da177e4 488
e4e11e38 489 for (i = 0; i < num ; i++) {
d592a996 490 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
1da177e4
LT		 491 p = sh->dev[i].page;
492 if (!p)
493 continue;
494 sh->dev[i].page = NULL;
2d1f3b5d 495 put_page(p);
1da177e4
LT		 496 }
497}
498
e4e11e38 499static int grow_buffers(struct stripe_head *sh)
1da177e4
LT		 500{
501 int i;
e4e11e38 502 int num = sh->raid_conf->pool_size;
1da177e4 503
e4e11e38 504 for (i = 0; i < num; i++) {
1da177e4
LT		 505 struct page *page;
506
507 if (!(page = alloc_page(GFP_KERNEL))) {
508 return 1;
509 }
510 sh->dev[i].page = page;
d592a996 511 sh->dev[i].orig_page = page;
1da177e4
LT		 512 }
513 return 0;
514}
515
784052ec 516static void raid5_build_block(struct stripe_head *sh, int i, int previous);
d1688a6d 517static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
911d4ee8 518 struct stripe_head *sh);
1da177e4 519
b5663ba4 520static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
1da177e4 521{
d1688a6d 522 struct r5conf *conf = sh->raid_conf;
566c09c5 523 int i, seq;
1da177e4 524
78bafebd
ES		 525 BUG_ON(atomic_read(&sh->count) != 0);
526 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
600aa109 527 BUG_ON(stripe_operations_active(sh));
d84e0f10 528
45b4233c 529 pr_debug("init_stripe called, stripe %llu\n",
b8e6a15a 530 (unsigned long long)sector);
566c09c5
SL		 531retry:
532 seq = read_seqcount_begin(&conf->gen_lock);
86b42c71 533 sh->generation = conf->generation - previous;
b5663ba4 534 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
1da177e4 535 sh->sector = sector;
911d4ee8 536 stripe_set_idx(sector, conf, previous, sh);
1da177e4
LT		 537 sh->state = 0;
538
7ecaa1e6 539 for (i = sh->disks; i--; ) {
1da177e4
LT		 540 struct r5dev *dev = &sh->dev[i];
541
d84e0f10 542 if (dev->toread || dev->read || dev->towrite || dev->written ||
1da177e4 543 test_bit(R5_LOCKED, &dev->flags)) {
d84e0f10 544 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
1da177e4 545 (unsigned long long)sh->sector, i, dev->toread,
d84e0f10 546 dev->read, dev->towrite, dev->written,
1da177e4 547 test_bit(R5_LOCKED, &dev->flags));
8cfa7b0f 548 WARN_ON(1);
1da177e4
LT		 549 }
550 dev->flags = 0;
784052ec 551 raid5_build_block(sh, i, previous);
1da177e4 552 }
566c09c5
SL		 553 if (read_seqcount_retry(&conf->gen_lock, seq))
554 goto retry;
1da177e4 555 insert_hash(conf, sh);
851c30c9 556 sh->cpu = smp_processor_id();
1da177e4
LT		 557}
558
d1688a6d 559static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
86b42c71 560 short generation)
1da177e4
LT		 561{
562 struct stripe_head *sh;
563
45b4233c 564 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
b67bfe0d 565 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
86b42c71 566 if (sh->sector == sector && sh->generation == generation)
1da177e4 567 return sh;
45b4233c 568 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
1da177e4
LT		 569 return NULL;
570}
571
674806d6
N		 572/*
573 * Need to check if array has failed when deciding whether to:
574 * - start an array
575 * - remove non-faulty devices
576 * - add a spare
577 * - allow a reshape
578 * This determination is simple when no reshape is happening.
579 * However if there is a reshape, we need to carefully check
580 * both the before and after sections.
581 * This is because some failed devices may only affect one
582 * of the two sections, and some non-in_sync devices may
583 * be insync in the section most affected by failed devices.
584 */
908f4fbd 585static int calc_degraded(struct r5conf *conf)
674806d6 586{
908f4fbd 587 int degraded, degraded2;
674806d6 588 int i;
674806d6
N		 589
590 rcu_read_lock();
591 degraded = 0;
592 for (i = 0; i < conf->previous_raid_disks; i++) {
3cb03002 593 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
e5c86471
N		 594 if (rdev && test_bit(Faulty, &rdev->flags))
595 rdev = rcu_dereference(conf->disks[i].replacement);
674806d6
N		 596 if (!rdev || test_bit(Faulty, &rdev->flags))
597 degraded++;
598 else if (test_bit(In_sync, &rdev->flags))
599 ;
600 else
601 /* not in-sync or faulty.
602 * If the reshape increases the number of devices,
603 * this is being recovered by the reshape, so
604 * this 'previous' section is not in_sync.
605 * If the number of devices is being reduced however,
606 * the device can only be part of the array if
607 * we are reverting a reshape, so this section will
608 * be in-sync.
609 */
610 if (conf->raid_disks >= conf->previous_raid_disks)
611 degraded++;
612 }
613 rcu_read_unlock();
908f4fbd
N		 614 if (conf->raid_disks == conf->previous_raid_disks)
615 return degraded;
674806d6 616 rcu_read_lock();
908f4fbd 617 degraded2 = 0;
674806d6 618 for (i = 0; i < conf->raid_disks; i++) {
3cb03002 619 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
e5c86471
N		 620 if (rdev && test_bit(Faulty, &rdev->flags))
621 rdev = rcu_dereference(conf->disks[i].replacement);
674806d6 622 if (!rdev || test_bit(Faulty, &rdev->flags))
908f4fbd 623 degraded2++;
674806d6
N		 624 else if (test_bit(In_sync, &rdev->flags))
625 ;
626 else
627 /* not in-sync or faulty.
628 * If reshape increases the number of devices, this
629 * section has already been recovered, else it
630 * almost certainly hasn't.
631 */
632 if (conf->raid_disks <= conf->previous_raid_disks)
908f4fbd 633 degraded2++;
674806d6
N		 634 }
635 rcu_read_unlock();
908f4fbd
N		 636 if (degraded2 > degraded)
637 return degraded2;
638 return degraded;
639}
640
641static int has_failed(struct r5conf *conf)
642{
643 int degraded;
644
645 if (conf->mddev->reshape_position == MaxSector)
646 return conf->mddev->degraded > conf->max_degraded;
647
648 degraded = calc_degraded(conf);
674806d6
N		 649 if (degraded > conf->max_degraded)
650 return 1;
651 return 0;
652}
653
b5663ba4 654static struct stripe_head *
d1688a6d 655get_active_stripe(struct r5conf *conf, sector_t sector,
a8c906ca 656 int previous, int noblock, int noquiesce)
1da177e4
LT		 657{
658 struct stripe_head *sh;
566c09c5 659 int hash = stripe_hash_locks_hash(sector);
1da177e4 660
45b4233c 661 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
1da177e4 662
566c09c5 663 spin_lock_irq(conf->hash_locks + hash);
1da177e4
LT		 664
665 do {
72626685 666 wait_event_lock_irq(conf->wait_for_stripe,
a8c906ca 667 conf->quiesce == 0 || noquiesce,
566c09c5 668 *(conf->hash_locks + hash));
86b42c71 669 sh = __find_stripe(conf, sector, conf->generation - previous);
1da177e4
LT		 670 if (!sh) {
671 if (!conf->inactive_blocked)
566c09c5 672 sh = get_free_stripe(conf, hash);
1da177e4
LT		 673 if (noblock && sh == NULL)
674 break;
675 if (!sh) {
676 conf->inactive_blocked = 1;
566c09c5
SL		 677 wait_event_lock_irq(
678 conf->wait_for_stripe,
679 !list_empty(conf->inactive_list + hash) &&
680 (atomic_read(&conf->active_stripes)
681 < (conf->max_nr_stripes * 3 / 4)
682 || !conf->inactive_blocked),
683 *(conf->hash_locks + hash));
1da177e4 684 conf->inactive_blocked = 0;
7da9d450 685 } else {
b5663ba4 686 init_stripe(sh, sector, previous);
7da9d450
N		 687 atomic_inc(&sh->count);
688 }
e240c183 689 } else if (!atomic_inc_not_zero(&sh->count)) {
6d183de4 690 spin_lock(&conf->device_lock);
e240c183 691 if (!atomic_read(&sh->count)) {
1da177e4
LT		 692 if (!test_bit(STRIPE_HANDLE, &sh->state))
693 atomic_inc(&conf->active_stripes);
5af9bef7
N		 694 BUG_ON(list_empty(&sh->lru) &&
695 !test_bit(STRIPE_EXPANDING, &sh->state));
16a53ecc 696 list_del_init(&sh->lru);
bfc90cb0
SL		 697 if (sh->group) {
698 sh->group->stripes_cnt--;
699 sh->group = NULL;
700 }
1da177e4 701 }
7da9d450 702 atomic_inc(&sh->count);
6d183de4 703 spin_unlock(&conf->device_lock);
1da177e4
LT		 704 }
705 } while (sh == NULL);
706
566c09c5 707 spin_unlock_irq(conf->hash_locks + hash);
1da177e4
LT		 708 return sh;
709}
710
05616be5
N		 711/* Determine if 'data_offset' or 'new_data_offset' should be used
712 * in this stripe_head.
713 */
714static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
715{
716 sector_t progress = conf->reshape_progress;
717 /* Need a memory barrier to make sure we see the value
718 * of conf->generation, or ->data_offset that was set before
719 * reshape_progress was updated.
720 */
721 smp_rmb();
722 if (progress == MaxSector)
723 return 0;
724 if (sh->generation == conf->generation - 1)
725 return 0;
726 /* We are in a reshape, and this is a new-generation stripe,
727 * so use new_data_offset.
728 */
729 return 1;
730}
731
6712ecf8
N		 732static void
733raid5_end_read_request(struct bio *bi, int error);
734static void
735raid5_end_write_request(struct bio *bi, int error);
91c00924 736
c4e5ac0a 737static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
91c00924 738{
d1688a6d 739 struct r5conf *conf = sh->raid_conf;
91c00924
DW		 740 int i, disks = sh->disks;
741
742 might_sleep();
743
744 for (i = disks; i--; ) {
745 int rw;
9a3e1101 746 int replace_only = 0;
977df362
N		 747 struct bio *bi, *rbi;
748 struct md_rdev *rdev, *rrdev = NULL;
e9c7469b
TH		 749 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
750 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
751 rw = WRITE_FUA;
752 else
753 rw = WRITE;
9e444768 754 if (test_bit(R5_Discard, &sh->dev[i].flags))
620125f2 755 rw |= REQ_DISCARD;
e9c7469b 756 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
91c00924 757 rw = READ;
9a3e1101
N		 758 else if (test_and_clear_bit(R5_WantReplace,
759 &sh->dev[i].flags)) {
760 rw = WRITE;
761 replace_only = 1;
762 } else
91c00924 763 continue;
bc0934f0
SL		 764 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
765 rw |= REQ_SYNC;
91c00924
DW		 766
767 bi = &sh->dev[i].req;
977df362 768 rbi = &sh->dev[i].rreq; /* For writing to replacement */
91c00924 769
91c00924 770 rcu_read_lock();
9a3e1101 771 rrdev = rcu_dereference(conf->disks[i].replacement);
dd054fce
N		 772 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
773 rdev = rcu_dereference(conf->disks[i].rdev);
774 if (!rdev) {
775 rdev = rrdev;
776 rrdev = NULL;
777 }
9a3e1101
N		 778 if (rw & WRITE) {
779 if (replace_only)
780 rdev = NULL;
dd054fce
N		 781 if (rdev == rrdev)
782 /* We raced and saw duplicates */
783 rrdev = NULL;
9a3e1101 784 } else {
dd054fce 785 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
9a3e1101
N		 786 rdev = rrdev;
787 rrdev = NULL;
788 }
977df362 789
91c00924
DW		 790 if (rdev && test_bit(Faulty, &rdev->flags))
791 rdev = NULL;
792 if (rdev)
793 atomic_inc(&rdev->nr_pending);
977df362
N		 794 if (rrdev && test_bit(Faulty, &rrdev->flags))
795 rrdev = NULL;
796 if (rrdev)
797 atomic_inc(&rrdev->nr_pending);
91c00924
DW		 798 rcu_read_unlock();
799
73e92e51 800 /* We have already checked bad blocks for reads. Now
977df362
N		 801 * need to check for writes. We never accept write errors
802 * on the replacement, so we don't to check rrdev.
73e92e51
N		 803 */
804 while ((rw & WRITE) && rdev &&
805 test_bit(WriteErrorSeen, &rdev->flags)) {
806 sector_t first_bad;
807 int bad_sectors;
808 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
809 &first_bad, &bad_sectors);
810 if (!bad)
811 break;
812
813 if (bad < 0) {
814 set_bit(BlockedBadBlocks, &rdev->flags);
815 if (!conf->mddev->external &&
816 conf->mddev->flags) {
817 /* It is very unlikely, but we might
818 * still need to write out the
819 * bad block log - better give it
820 * a chance*/
821 md_check_recovery(conf->mddev);
822 }
1850753d 823 /*
824 * Because md_wait_for_blocked_rdev
825 * will dec nr_pending, we must
826 * increment it first.
827 */
828 atomic_inc(&rdev->nr_pending);
73e92e51
N		 829 md_wait_for_blocked_rdev(rdev, conf->mddev);
830 } else {
831 /* Acknowledged bad block - skip the write */
832 rdev_dec_pending(rdev, conf->mddev);
833 rdev = NULL;
834 }
835 }
836
91c00924 837 if (rdev) {
9a3e1101
N		 838 if (s->syncing || s->expanding || s->expanded
839 || s->replacing)
91c00924
DW		 840 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
841
2b7497f0
DW		 842 set_bit(STRIPE_IO_STARTED, &sh->state);
843
2f6db2a7 844 bio_reset(bi);
91c00924 845 bi->bi_bdev = rdev->bdev;
2f6db2a7
KO		 846 bi->bi_rw = rw;
847 bi->bi_end_io = (rw & WRITE)
848 ? raid5_end_write_request
849 : raid5_end_read_request;
850 bi->bi_private = sh;
851
91c00924 852 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
e46b272b 853 __func__, (unsigned long long)sh->sector,
91c00924
DW		 854 bi->bi_rw, i);
855 atomic_inc(&sh->count);
05616be5 856 if (use_new_offset(conf, sh))
4f024f37 857 bi->bi_iter.bi_sector = (sh->sector
05616be5
N		 858 + rdev->new_data_offset);
859 else
4f024f37 860 bi->bi_iter.bi_sector = (sh->sector
05616be5 861 + rdev->data_offset);
3f9e7c14 862 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
e59aa23f 863 bi->bi_rw |= REQ_NOMERGE;
3f9e7c14 864
d592a996
SL		 865 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
866 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
867 sh->dev[i].vec.bv_page = sh->dev[i].page;
4997b72e 868 bi->bi_vcnt = 1;
91c00924
DW		 869 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
870 bi->bi_io_vec[0].bv_offset = 0;
4f024f37 871 bi->bi_iter.bi_size = STRIPE_SIZE;
37c61ff3
SL		 872 /*
873 * If this is discard request, set bi_vcnt 0. We don't
874 * want to confuse SCSI because SCSI will replace payload
875 */
876 if (rw & REQ_DISCARD)
877 bi->bi_vcnt = 0;
977df362
N		 878 if (rrdev)
879 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
e3620a3a
JB		 880
881 if (conf->mddev->gendisk)
882 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
883 bi, disk_devt(conf->mddev->gendisk),
884 sh->dev[i].sector);
91c00924 885 generic_make_request(bi);
977df362
N		 886 }
887 if (rrdev) {
9a3e1101
N		 888 if (s->syncing || s->expanding || s->expanded
889 || s->replacing)
977df362
N		 890 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
891
892 set_bit(STRIPE_IO_STARTED, &sh->state);
893
2f6db2a7 894 bio_reset(rbi);
977df362 895 rbi->bi_bdev = rrdev->bdev;
2f6db2a7
KO		 896 rbi->bi_rw = rw;
897 BUG_ON(!(rw & WRITE));
898 rbi->bi_end_io = raid5_end_write_request;
899 rbi->bi_private = sh;
900
977df362
N		 901 pr_debug("%s: for %llu schedule op %ld on "
902 "replacement disc %d\n",
903 __func__, (unsigned long long)sh->sector,
904 rbi->bi_rw, i);
905 atomic_inc(&sh->count);
05616be5 906 if (use_new_offset(conf, sh))
4f024f37 907 rbi->bi_iter.bi_sector = (sh->sector
05616be5
N		 908 + rrdev->new_data_offset);
909 else
4f024f37 910 rbi->bi_iter.bi_sector = (sh->sector
05616be5 911 + rrdev->data_offset);
d592a996
SL		 912 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
913 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
914 sh->dev[i].rvec.bv_page = sh->dev[i].page;
4997b72e 915 rbi->bi_vcnt = 1;
977df362
N		 916 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
917 rbi->bi_io_vec[0].bv_offset = 0;
4f024f37 918 rbi->bi_iter.bi_size = STRIPE_SIZE;
37c61ff3
SL		 919 /*
920 * If this is discard request, set bi_vcnt 0. We don't
921 * want to confuse SCSI because SCSI will replace payload
922 */
923 if (rw & REQ_DISCARD)
924 rbi->bi_vcnt = 0;
e3620a3a
JB		 925 if (conf->mddev->gendisk)
926 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
927 rbi, disk_devt(conf->mddev->gendisk),
928 sh->dev[i].sector);
977df362
N		 929 generic_make_request(rbi);
930 }
931 if (!rdev && !rrdev) {
b062962e 932 if (rw & WRITE)
91c00924
DW		 933 set_bit(STRIPE_DEGRADED, &sh->state);
934 pr_debug("skip op %ld on disc %d for sector %llu\n",
935 bi->bi_rw, i, (unsigned long long)sh->sector);
936 clear_bit(R5_LOCKED, &sh->dev[i].flags);
937 set_bit(STRIPE_HANDLE, &sh->state);
938 }
939 }
940}
941
942static struct dma_async_tx_descriptor *
d592a996
SL		 943async_copy_data(int frombio, struct bio *bio, struct page **page,
944 sector_t sector, struct dma_async_tx_descriptor *tx,
945 struct stripe_head *sh)
91c00924 946{
7988613b
KO		 947 struct bio_vec bvl;
948 struct bvec_iter iter;
91c00924 949 struct page *bio_page;
91c00924 950 int page_offset;
a08abd8c 951 struct async_submit_ctl submit;
0403e382 952 enum async_tx_flags flags = 0;
91c00924 953
4f024f37
KO		 954 if (bio->bi_iter.bi_sector >= sector)
955 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
91c00924 956 else
4f024f37 957 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
a08abd8c 958
0403e382
DW		 959 if (frombio)
960 flags |= ASYNC_TX_FENCE;
961 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
962
7988613b
KO		 963 bio_for_each_segment(bvl, bio, iter) {
964 int len = bvl.bv_len;
91c00924
DW		 965 int clen;
966 int b_offset = 0;
967
968 if (page_offset < 0) {
969 b_offset = -page_offset;
970 page_offset += b_offset;
971 len -= b_offset;
972 }
973
974 if (len > 0 && page_offset + len > STRIPE_SIZE)
975 clen = STRIPE_SIZE - page_offset;
976 else
977 clen = len;
978
979 if (clen > 0) {
7988613b
KO		 980 b_offset += bvl.bv_offset;
981 bio_page = bvl.bv_page;
d592a996
SL		 982 if (frombio) {
983 if (sh->raid_conf->skip_copy &&
984 b_offset == 0 && page_offset == 0 &&
985 clen == STRIPE_SIZE)
986 *page = bio_page;
987 else
988 tx = async_memcpy(*page, bio_page, page_offset,
a08abd8c 989 b_offset, clen, &submit);
d592a996
SL		 990 } else
991 tx = async_memcpy(bio_page, *page, b_offset,
a08abd8c 992 page_offset, clen, &submit);
91c00924 993 }
a08abd8c
DW		 994 /* chain the operations */
995 submit.depend_tx = tx;
996
91c00924
DW		 997 if (clen < len) /* hit end of page */
998 break;
999 page_offset += len;
1000 }
1001
1002 return tx;
1003}
1004
1005static void ops_complete_biofill(void *stripe_head_ref)
1006{
1007 struct stripe_head *sh = stripe_head_ref;
1008 struct bio *return_bi = NULL;
e4d84909 1009 int i;
91c00924 1010
e46b272b 1011 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1012 (unsigned long long)sh->sector);
1013
1014 /* clear completed biofills */
1015 for (i = sh->disks; i--; ) {
1016 struct r5dev *dev = &sh->dev[i];
91c00924
DW		 1017
1018 /* acknowledge completion of a biofill operation */
e4d84909
DW		 1019 /* and check if we need to reply to a read request,
1020 * new R5_Wantfill requests are held off until
83de75cc 1021 * !STRIPE_BIOFILL_RUN
e4d84909
DW		 1022 */
1023 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
91c00924 1024 struct bio *rbi, *rbi2;
91c00924 1025
91c00924
DW		 1026 BUG_ON(!dev->read);
1027 rbi = dev->read;
1028 dev->read = NULL;
4f024f37 1029 while (rbi && rbi->bi_iter.bi_sector <
91c00924
DW		 1030 dev->sector + STRIPE_SECTORS) {
1031 rbi2 = r5_next_bio(rbi, dev->sector);
e7836bd6 1032 if (!raid5_dec_bi_active_stripes(rbi)) {
91c00924
DW		 1033 rbi->bi_next = return_bi;
1034 return_bi = rbi;
1035 }
91c00924
DW		 1036 rbi = rbi2;
1037 }
1038 }
1039 }
83de75cc 1040 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
91c00924
DW		 1041
1042 return_io(return_bi);
1043
e4d84909 1044 set_bit(STRIPE_HANDLE, &sh->state);
91c00924
DW		 1045 release_stripe(sh);
1046}
1047
1048static void ops_run_biofill(struct stripe_head *sh)
1049{
1050 struct dma_async_tx_descriptor *tx = NULL;
a08abd8c 1051 struct async_submit_ctl submit;
91c00924
DW		 1052 int i;
1053
e46b272b 1054 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1055 (unsigned long long)sh->sector);
1056
1057 for (i = sh->disks; i--; ) {
1058 struct r5dev *dev = &sh->dev[i];
1059 if (test_bit(R5_Wantfill, &dev->flags)) {
1060 struct bio *rbi;
b17459c0 1061 spin_lock_irq(&sh->stripe_lock);
91c00924
DW		 1062 dev->read = rbi = dev->toread;
1063 dev->toread = NULL;
b17459c0 1064 spin_unlock_irq(&sh->stripe_lock);
4f024f37 1065 while (rbi && rbi->bi_iter.bi_sector <
91c00924 1066 dev->sector + STRIPE_SECTORS) {
d592a996
SL		 1067 tx = async_copy_data(0, rbi, &dev->page,
1068 dev->sector, tx, sh);
91c00924
DW		 1069 rbi = r5_next_bio(rbi, dev->sector);
1070 }
1071 }
1072 }
1073
1074 atomic_inc(&sh->count);
a08abd8c
DW		 1075 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1076 async_trigger_callback(&submit);
91c00924
DW		 1077}
1078
4e7d2c0a 1079static void mark_target_uptodate(struct stripe_head *sh, int target)
91c00924 1080{
4e7d2c0a 1081 struct r5dev *tgt;
91c00924 1082
4e7d2c0a
DW		 1083 if (target < 0)
1084 return;
91c00924 1085
4e7d2c0a 1086 tgt = &sh->dev[target];
91c00924
DW		 1087 set_bit(R5_UPTODATE, &tgt->flags);
1088 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1089 clear_bit(R5_Wantcompute, &tgt->flags);
4e7d2c0a
DW		 1090}
1091
ac6b53b6 1092static void ops_complete_compute(void *stripe_head_ref)
91c00924
DW		 1093{
1094 struct stripe_head *sh = stripe_head_ref;
91c00924 1095
e46b272b 1096 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1097 (unsigned long long)sh->sector);
1098
ac6b53b6 1099 /* mark the computed target(s) as uptodate */
4e7d2c0a 1100 mark_target_uptodate(sh, sh->ops.target);
ac6b53b6 1101 mark_target_uptodate(sh, sh->ops.target2);
4e7d2c0a 1102
ecc65c9b
DW		 1103 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1104 if (sh->check_state == check_state_compute_run)
1105 sh->check_state = check_state_compute_result;
91c00924
DW		 1106 set_bit(STRIPE_HANDLE, &sh->state);
1107 release_stripe(sh);
1108}
1109
d6f38f31
DW		 1110/* return a pointer to the address conversion region of the scribble buffer */
1111static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1112 struct raid5_percpu *percpu)
1113{
1114 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
1115}
1116
1117static struct dma_async_tx_descriptor *
1118ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
91c00924 1119{
91c00924 1120 int disks = sh->disks;
d6f38f31 1121 struct page **xor_srcs = percpu->scribble;
91c00924
DW		 1122 int target = sh->ops.target;
1123 struct r5dev *tgt = &sh->dev[target];
1124 struct page *xor_dest = tgt->page;
1125 int count = 0;
1126 struct dma_async_tx_descriptor *tx;
a08abd8c 1127 struct async_submit_ctl submit;
91c00924
DW		 1128 int i;
1129
1130 pr_debug("%s: stripe %llu block: %d\n",
e46b272b 1131 __func__, (unsigned long long)sh->sector, target);
91c00924
DW		 1132 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1133
1134 for (i = disks; i--; )
1135 if (i != target)
1136 xor_srcs[count++] = sh->dev[i].page;
1137
1138 atomic_inc(&sh->count);
1139
0403e382 1140 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
ac6b53b6 1141 ops_complete_compute, sh, to_addr_conv(sh, percpu));
91c00924 1142 if (unlikely(count == 1))
a08abd8c 1143 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
91c00924 1144 else
a08abd8c 1145 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
91c00924 1146
91c00924
DW		 1147 return tx;
1148}
1149
ac6b53b6
DW		 1150/* set_syndrome_sources - populate source buffers for gen_syndrome
1151 * @srcs - (struct page *) array of size sh->disks
1152 * @sh - stripe_head to parse
1153 *
1154 * Populates srcs in proper layout order for the stripe and returns the
1155 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1156 * destination buffer is recorded in srcs[count] and the Q destination
1157 * is recorded in srcs[count+1]].
1158 */
1159static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
1160{
1161 int disks = sh->disks;
1162 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1163 int d0_idx = raid6_d0(sh);
1164 int count;
1165 int i;
1166
1167 for (i = 0; i < disks; i++)
5dd33c9a 1168 srcs[i] = NULL;
ac6b53b6
DW		 1169
1170 count = 0;
1171 i = d0_idx;
1172 do {
1173 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1174
1175 srcs[slot] = sh->dev[i].page;
1176 i = raid6_next_disk(i, disks);
1177 } while (i != d0_idx);
ac6b53b6 1178
e4424fee 1179 return syndrome_disks;
ac6b53b6
DW		 1180}
1181
1182static struct dma_async_tx_descriptor *
1183ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1184{
1185 int disks = sh->disks;
1186 struct page **blocks = percpu->scribble;
1187 int target;
1188 int qd_idx = sh->qd_idx;
1189 struct dma_async_tx_descriptor *tx;
1190 struct async_submit_ctl submit;
1191 struct r5dev *tgt;
1192 struct page *dest;
1193 int i;
1194 int count;
1195
1196 if (sh->ops.target < 0)
1197 target = sh->ops.target2;
1198 else if (sh->ops.target2 < 0)
1199 target = sh->ops.target;
91c00924 1200 else
ac6b53b6
DW		 1201 /* we should only have one valid target */
1202 BUG();
1203 BUG_ON(target < 0);
1204 pr_debug("%s: stripe %llu block: %d\n",
1205 __func__, (unsigned long long)sh->sector, target);
1206
1207 tgt = &sh->dev[target];
1208 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1209 dest = tgt->page;
1210
1211 atomic_inc(&sh->count);
1212
1213 if (target == qd_idx) {
1214 count = set_syndrome_sources(blocks, sh);
1215 blocks[count] = NULL; /* regenerating p is not necessary */
1216 BUG_ON(blocks[count+1] != dest); /* q should already be set */
0403e382
DW		 1217 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1218 ops_complete_compute, sh,
ac6b53b6
DW		 1219 to_addr_conv(sh, percpu));
1220 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1221 } else {
1222 /* Compute any data- or p-drive using XOR */
1223 count = 0;
1224 for (i = disks; i-- ; ) {
1225 if (i == target || i == qd_idx)
1226 continue;
1227 blocks[count++] = sh->dev[i].page;
1228 }
1229
0403e382
DW		 1230 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1231 NULL, ops_complete_compute, sh,
ac6b53b6
DW		 1232 to_addr_conv(sh, percpu));
1233 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1234 }
91c00924 1235
91c00924
DW		 1236 return tx;
1237}
1238
ac6b53b6
DW		 1239static struct dma_async_tx_descriptor *
1240ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1241{
1242 int i, count, disks = sh->disks;
1243 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1244 int d0_idx = raid6_d0(sh);
1245 int faila = -1, failb = -1;
1246 int target = sh->ops.target;
1247 int target2 = sh->ops.target2;
1248 struct r5dev *tgt = &sh->dev[target];
1249 struct r5dev *tgt2 = &sh->dev[target2];
1250 struct dma_async_tx_descriptor *tx;
1251 struct page **blocks = percpu->scribble;
1252 struct async_submit_ctl submit;
1253
1254 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1255 __func__, (unsigned long long)sh->sector, target, target2);
1256 BUG_ON(target < 0 || target2 < 0);
1257 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1258 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1259
6c910a78 1260 /* we need to open-code set_syndrome_sources to handle the
ac6b53b6
DW		 1261 * slot number conversion for 'faila' and 'failb'
1262 */
1263 for (i = 0; i < disks ; i++)
5dd33c9a 1264 blocks[i] = NULL;
ac6b53b6
DW		 1265 count = 0;
1266 i = d0_idx;
1267 do {
1268 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1269
1270 blocks[slot] = sh->dev[i].page;
1271
1272 if (i == target)
1273 faila = slot;
1274 if (i == target2)
1275 failb = slot;
1276 i = raid6_next_disk(i, disks);
1277 } while (i != d0_idx);
ac6b53b6
DW		 1278
1279 BUG_ON(faila == failb);
1280 if (failb < faila)
1281 swap(faila, failb);
1282 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1283 __func__, (unsigned long long)sh->sector, faila, failb);
1284
1285 atomic_inc(&sh->count);
1286
1287 if (failb == syndrome_disks+1) {
1288 /* Q disk is one of the missing disks */
1289 if (faila == syndrome_disks) {
1290 /* Missing P+Q, just recompute */
0403e382
DW		 1291 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1292 ops_complete_compute, sh,
1293 to_addr_conv(sh, percpu));
e4424fee 1294 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
ac6b53b6
DW		 1295 STRIPE_SIZE, &submit);
1296 } else {
1297 struct page *dest;
1298 int data_target;
1299 int qd_idx = sh->qd_idx;
1300
1301 /* Missing D+Q: recompute D from P, then recompute Q */
1302 if (target == qd_idx)
1303 data_target = target2;
1304 else
1305 data_target = target;
1306
1307 count = 0;
1308 for (i = disks; i-- ; ) {
1309 if (i == data_target || i == qd_idx)
1310 continue;
1311 blocks[count++] = sh->dev[i].page;
1312 }
1313 dest = sh->dev[data_target].page;
0403e382
DW		 1314 init_async_submit(&submit,
1315 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1316 NULL, NULL, NULL,
1317 to_addr_conv(sh, percpu));
ac6b53b6
DW		 1318 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1319 &submit);
1320
1321 count = set_syndrome_sources(blocks, sh);
0403e382
DW		 1322 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1323 ops_complete_compute, sh,
1324 to_addr_conv(sh, percpu));
ac6b53b6
DW		 1325 return async_gen_syndrome(blocks, 0, count+2,
1326 STRIPE_SIZE, &submit);
1327 }
ac6b53b6 1328 } else {
6c910a78
DW		 1329 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1330 ops_complete_compute, sh,
1331 to_addr_conv(sh, percpu));
1332 if (failb == syndrome_disks) {
1333 /* We're missing D+P. */
1334 return async_raid6_datap_recov(syndrome_disks+2,
1335 STRIPE_SIZE, faila,
1336 blocks, &submit);
1337 } else {
1338 /* We're missing D+D. */
1339 return async_raid6_2data_recov(syndrome_disks+2,
1340 STRIPE_SIZE, faila, failb,
1341 blocks, &submit);
1342 }
ac6b53b6
DW		 1343 }
1344}
1345
91c00924
DW		 1346static void ops_complete_prexor(void *stripe_head_ref)
1347{
1348 struct stripe_head *sh = stripe_head_ref;
1349
e46b272b 1350 pr_debug("%s: stripe %llu\n", __func__,
91c00924 1351 (unsigned long long)sh->sector);
91c00924
DW		 1352}
1353
1354static struct dma_async_tx_descriptor *
d6f38f31
DW		 1355ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1356 struct dma_async_tx_descriptor *tx)
91c00924 1357{
91c00924 1358 int disks = sh->disks;
d6f38f31 1359 struct page **xor_srcs = percpu->scribble;
91c00924 1360 int count = 0, pd_idx = sh->pd_idx, i;
a08abd8c 1361 struct async_submit_ctl submit;
91c00924
DW		 1362
1363 /* existing parity data subtracted */
1364 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1365
e46b272b 1366 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1367 (unsigned long long)sh->sector);
1368
1369 for (i = disks; i--; ) {
1370 struct r5dev *dev = &sh->dev[i];
1371 /* Only process blocks that are known to be uptodate */
d8ee0728 1372 if (test_bit(R5_Wantdrain, &dev->flags))
91c00924
DW		 1373 xor_srcs[count++] = dev->page;
1374 }
1375
0403e382 1376 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
d6f38f31 1377 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
a08abd8c 1378 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
91c00924
DW		 1379
1380 return tx;
1381}
1382
1383static struct dma_async_tx_descriptor *
d8ee0728 1384ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
91c00924
DW		 1385{
1386 int disks = sh->disks;
d8ee0728 1387 int i;
91c00924 1388
e46b272b 1389 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1390 (unsigned long long)sh->sector);
1391
1392 for (i = disks; i--; ) {
1393 struct r5dev *dev = &sh->dev[i];
1394 struct bio *chosen;
91c00924 1395
d8ee0728 1396 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
91c00924
DW		 1397 struct bio *wbi;
1398
b17459c0 1399 spin_lock_irq(&sh->stripe_lock);
91c00924
DW		 1400 chosen = dev->towrite;
1401 dev->towrite = NULL;
1402 BUG_ON(dev->written);
1403 wbi = dev->written = chosen;
b17459c0 1404 spin_unlock_irq(&sh->stripe_lock);
d592a996 1405 WARN_ON(dev->page != dev->orig_page);
91c00924 1406
4f024f37 1407 while (wbi && wbi->bi_iter.bi_sector <
91c00924 1408 dev->sector + STRIPE_SECTORS) {
e9c7469b
TH		 1409 if (wbi->bi_rw & REQ_FUA)
1410 set_bit(R5_WantFUA, &dev->flags);
bc0934f0
SL		 1411 if (wbi->bi_rw & REQ_SYNC)
1412 set_bit(R5_SyncIO, &dev->flags);
9e444768 1413 if (wbi->bi_rw & REQ_DISCARD)
620125f2 1414 set_bit(R5_Discard, &dev->flags);
d592a996
SL		 1415 else {
1416 tx = async_copy_data(1, wbi, &dev->page,
1417 dev->sector, tx, sh);
1418 if (dev->page != dev->orig_page) {
1419 set_bit(R5_SkipCopy, &dev->flags);
1420 clear_bit(R5_UPTODATE, &dev->flags);
1421 clear_bit(R5_OVERWRITE, &dev->flags);
1422 }
1423 }
91c00924
DW		 1424 wbi = r5_next_bio(wbi, dev->sector);
1425 }
1426 }
1427 }
1428
1429 return tx;
1430}
1431
ac6b53b6 1432static void ops_complete_reconstruct(void *stripe_head_ref)
91c00924
DW		 1433{
1434 struct stripe_head *sh = stripe_head_ref;
ac6b53b6
DW		 1435 int disks = sh->disks;
1436 int pd_idx = sh->pd_idx;
1437 int qd_idx = sh->qd_idx;
1438 int i;
9e444768 1439 bool fua = false, sync = false, discard = false;
91c00924 1440
e46b272b 1441 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1442 (unsigned long long)sh->sector);
1443
bc0934f0 1444 for (i = disks; i--; ) {
e9c7469b 1445 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
bc0934f0 1446 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
9e444768 1447 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
bc0934f0 1448 }
e9c7469b 1449
91c00924
DW		 1450 for (i = disks; i--; ) {
1451 struct r5dev *dev = &sh->dev[i];
ac6b53b6 1452
e9c7469b 1453 if (dev->written || i == pd_idx || i == qd_idx) {
d592a996 1454 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
9e444768 1455 set_bit(R5_UPTODATE, &dev->flags);
e9c7469b
TH		 1456 if (fua)
1457 set_bit(R5_WantFUA, &dev->flags);
bc0934f0
SL		 1458 if (sync)
1459 set_bit(R5_SyncIO, &dev->flags);
e9c7469b 1460 }
91c00924
DW		 1461 }
1462
d8ee0728
DW		 1463 if (sh->reconstruct_state == reconstruct_state_drain_run)
1464 sh->reconstruct_state = reconstruct_state_drain_result;
1465 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1466 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1467 else {
1468 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1469 sh->reconstruct_state = reconstruct_state_result;
1470 }
91c00924
DW		 1471
1472 set_bit(STRIPE_HANDLE, &sh->state);
1473 release_stripe(sh);
1474}
1475
1476static void
ac6b53b6
DW		 1477ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1478 struct dma_async_tx_descriptor *tx)
91c00924 1479{
91c00924 1480 int disks = sh->disks;
d6f38f31 1481 struct page **xor_srcs = percpu->scribble;
a08abd8c 1482 struct async_submit_ctl submit;
91c00924
DW		 1483 int count = 0, pd_idx = sh->pd_idx, i;
1484 struct page *xor_dest;
d8ee0728 1485 int prexor = 0;
91c00924 1486 unsigned long flags;
91c00924 1487
e46b272b 1488 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1489 (unsigned long long)sh->sector);
1490
620125f2
SL		 1491 for (i = 0; i < sh->disks; i++) {
1492 if (pd_idx == i)
1493 continue;
1494 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1495 break;
1496 }
1497 if (i >= sh->disks) {
1498 atomic_inc(&sh->count);
620125f2
SL		 1499 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1500 ops_complete_reconstruct(sh);
1501 return;
1502 }
91c00924
DW		 1503 /* check if prexor is active which means only process blocks
1504 * that are part of a read-modify-write (written)
1505 */
d8ee0728
DW		 1506 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1507 prexor = 1;
91c00924
DW		 1508 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1509 for (i = disks; i--; ) {
1510 struct r5dev *dev = &sh->dev[i];
1511 if (dev->written)
1512 xor_srcs[count++] = dev->page;
1513 }
1514 } else {
1515 xor_dest = sh->dev[pd_idx].page;
1516 for (i = disks; i--; ) {
1517 struct r5dev *dev = &sh->dev[i];
1518 if (i != pd_idx)
1519 xor_srcs[count++] = dev->page;
1520 }
1521 }
1522
91c00924
DW		 1523 /* 1/ if we prexor'd then the dest is reused as a source
1524 * 2/ if we did not prexor then we are redoing the parity
1525 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1526 * for the synchronous xor case
1527 */
88ba2aa5 1528 flags = ASYNC_TX_ACK |
91c00924
DW		 1529 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1530
1531 atomic_inc(&sh->count);
1532
ac6b53b6 1533 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
d6f38f31 1534 to_addr_conv(sh, percpu));
a08abd8c
DW		 1535 if (unlikely(count == 1))
1536 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1537 else
1538 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
91c00924
DW		 1539}
1540
ac6b53b6
DW		 1541static void
1542ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1543 struct dma_async_tx_descriptor *tx)
1544{
1545 struct async_submit_ctl submit;
1546 struct page **blocks = percpu->scribble;
620125f2 1547 int count, i;
ac6b53b6
DW		 1548
1549 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1550
620125f2
SL		 1551 for (i = 0; i < sh->disks; i++) {
1552 if (sh->pd_idx == i || sh->qd_idx == i)
1553 continue;
1554 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1555 break;
1556 }
1557 if (i >= sh->disks) {
1558 atomic_inc(&sh->count);
620125f2
SL		 1559 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1560 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1561 ops_complete_reconstruct(sh);
1562 return;
1563 }
1564
ac6b53b6
DW		 1565 count = set_syndrome_sources(blocks, sh);
1566
1567 atomic_inc(&sh->count);
1568
1569 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1570 sh, to_addr_conv(sh, percpu));
1571 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
91c00924
DW		 1572}
1573
1574static void ops_complete_check(void *stripe_head_ref)
1575{
1576 struct stripe_head *sh = stripe_head_ref;
91c00924 1577
e46b272b 1578 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1579 (unsigned long long)sh->sector);
1580
ecc65c9b 1581 sh->check_state = check_state_check_result;
91c00924
DW		 1582 set_bit(STRIPE_HANDLE, &sh->state);
1583 release_stripe(sh);
1584}
1585
ac6b53b6 1586static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
91c00924 1587{
91c00924 1588 int disks = sh->disks;
ac6b53b6
DW		 1589 int pd_idx = sh->pd_idx;
1590 int qd_idx = sh->qd_idx;
1591 struct page *xor_dest;
d6f38f31 1592 struct page **xor_srcs = percpu->scribble;
91c00924 1593 struct dma_async_tx_descriptor *tx;
a08abd8c 1594 struct async_submit_ctl submit;
ac6b53b6
DW		 1595 int count;
1596 int i;
91c00924 1597
e46b272b 1598 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1599 (unsigned long long)sh->sector);
1600
ac6b53b6
DW		 1601 count = 0;
1602 xor_dest = sh->dev[pd_idx].page;
1603 xor_srcs[count++] = xor_dest;
91c00924 1604 for (i = disks; i--; ) {
ac6b53b6
DW		 1605 if (i == pd_idx || i == qd_idx)
1606 continue;
1607 xor_srcs[count++] = sh->dev[i].page;
91c00924
DW		 1608 }
1609
d6f38f31
DW		 1610 init_async_submit(&submit, 0, NULL, NULL, NULL,
1611 to_addr_conv(sh, percpu));
099f53cb 1612 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
a08abd8c 1613 &sh->ops.zero_sum_result, &submit);
91c00924 1614
91c00924 1615 atomic_inc(&sh->count);
a08abd8c
DW		 1616 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1617 tx = async_trigger_callback(&submit);
91c00924
DW		 1618}
1619
ac6b53b6
DW		 1620static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1621{
1622 struct page **srcs = percpu->scribble;
1623 struct async_submit_ctl submit;
1624 int count;
1625
1626 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1627 (unsigned long long)sh->sector, checkp);
1628
1629 count = set_syndrome_sources(srcs, sh);
1630 if (!checkp)
1631 srcs[count] = NULL;
91c00924 1632
91c00924 1633 atomic_inc(&sh->count);
ac6b53b6
DW		 1634 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1635 sh, to_addr_conv(sh, percpu));
1636 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1637 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
91c00924
DW		 1638}
1639
51acbcec 1640static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
91c00924
DW		 1641{
1642 int overlap_clear = 0, i, disks = sh->disks;
1643 struct dma_async_tx_descriptor *tx = NULL;
d1688a6d 1644 struct r5conf *conf = sh->raid_conf;
ac6b53b6 1645 int level = conf->level;
d6f38f31
DW		 1646 struct raid5_percpu *percpu;
1647 unsigned long cpu;
91c00924 1648
d6f38f31
DW		 1649 cpu = get_cpu();
1650 percpu = per_cpu_ptr(conf->percpu, cpu);
83de75cc 1651 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
91c00924
DW		 1652 ops_run_biofill(sh);
1653 overlap_clear++;
1654 }
1655
7b3a871e 1656 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
ac6b53b6
DW		 1657 if (level < 6)
1658 tx = ops_run_compute5(sh, percpu);
1659 else {
1660 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1661 tx = ops_run_compute6_1(sh, percpu);
1662 else
1663 tx = ops_run_compute6_2(sh, percpu);
1664 }
1665 /* terminate the chain if reconstruct is not set to be run */
1666 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
7b3a871e
DW		 1667 async_tx_ack(tx);
1668 }
91c00924 1669
600aa109 1670 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
d6f38f31 1671 tx = ops_run_prexor(sh, percpu, tx);
91c00924 1672
600aa109 1673 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
d8ee0728 1674 tx = ops_run_biodrain(sh, tx);
91c00924
DW		 1675 overlap_clear++;
1676 }
1677
ac6b53b6
DW		 1678 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1679 if (level < 6)
1680 ops_run_reconstruct5(sh, percpu, tx);
1681 else
1682 ops_run_reconstruct6(sh, percpu, tx);
1683 }
91c00924 1684
ac6b53b6
DW		 1685 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1686 if (sh->check_state == check_state_run)
1687 ops_run_check_p(sh, percpu);
1688 else if (sh->check_state == check_state_run_q)
1689 ops_run_check_pq(sh, percpu, 0);
1690 else if (sh->check_state == check_state_run_pq)
1691 ops_run_check_pq(sh, percpu, 1);
1692 else
1693 BUG();
1694 }
91c00924 1695
91c00924
DW		 1696 if (overlap_clear)
1697 for (i = disks; i--; ) {
1698 struct r5dev *dev = &sh->dev[i];
1699 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1700 wake_up(&sh->raid_conf->wait_for_overlap);
1701 }
d6f38f31 1702 put_cpu();
91c00924
DW		 1703}
1704
566c09c5 1705static int grow_one_stripe(struct r5conf *conf, int hash)
1da177e4
LT		 1706{
1707 struct stripe_head *sh;
6ce32846 1708 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
3f294f4f
N		 1709 if (!sh)
1710 return 0;
6ce32846 1711
3f294f4f 1712 sh->raid_conf = conf;
3f294f4f 1713
b17459c0
SL		 1714 spin_lock_init(&sh->stripe_lock);
1715
e4e11e38
N		 1716 if (grow_buffers(sh)) {
1717 shrink_buffers(sh);
3f294f4f
N		 1718 kmem_cache_free(conf->slab_cache, sh);
1719 return 0;
1720 }
566c09c5 1721 sh->hash_lock_index = hash;
3f294f4f
N		 1722 /* we just created an active stripe so... */
1723 atomic_set(&sh->count, 1);
1724 atomic_inc(&conf->active_stripes);
1725 INIT_LIST_HEAD(&sh->lru);
1726 release_stripe(sh);
1727 return 1;
1728}
1729
d1688a6d 1730static int grow_stripes(struct r5conf *conf, int num)
3f294f4f 1731{
e18b890b 1732 struct kmem_cache *sc;
5e5e3e78 1733 int devs = max(conf->raid_disks, conf->previous_raid_disks);
566c09c5 1734 int hash;
1da177e4 1735
f4be6b43
N		 1736 if (conf->mddev->gendisk)
1737 sprintf(conf->cache_name[0],
1738 "raid%d-%s", conf->level, mdname(conf->mddev));
1739 else
1740 sprintf(conf->cache_name[0],
1741 "raid%d-%p", conf->level, conf->mddev);
1742 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1743
ad01c9e3
N		 1744 conf->active_name = 0;
1745 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1da177e4 1746 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
20c2df83 1747 0, 0, NULL);
1da177e4
LT		 1748 if (!sc)
1749 return 1;
1750 conf->slab_cache = sc;
ad01c9e3 1751 conf->pool_size = devs;
566c09c5
SL		 1752 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
1753 while (num--) {
1754 if (!grow_one_stripe(conf, hash))
1da177e4 1755 return 1;
566c09c5
SL		 1756 conf->max_nr_stripes++;
1757 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
1758 }
1da177e4
LT		 1759 return 0;
1760}
29269553 1761
d6f38f31
DW		 1762/**
1763 * scribble_len - return the required size of the scribble region
1764 * @num - total number of disks in the array
1765 *
1766 * The size must be enough to contain:
1767 * 1/ a struct page pointer for each device in the array +2
1768 * 2/ room to convert each entry in (1) to its corresponding dma
1769 * (dma_map_page()) or page (page_address()) address.
1770 *
1771 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1772 * calculate over all devices (not just the data blocks), using zeros in place
1773 * of the P and Q blocks.
1774 */
1775static size_t scribble_len(int num)
1776{
1777 size_t len;
1778
1779 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1780
1781 return len;
1782}
1783
d1688a6d 1784static int resize_stripes(struct r5conf *conf, int newsize)
ad01c9e3
N		 1785{
1786 /* Make all the stripes able to hold 'newsize' devices.
1787 * New slots in each stripe get 'page' set to a new page.
1788 *
1789 * This happens in stages:
1790 * 1/ create a new kmem_cache and allocate the required number of
1791 * stripe_heads.
83f0d77a 1792 * 2/ gather all the old stripe_heads and transfer the pages across
ad01c9e3
N		 1793 * to the new stripe_heads. This will have the side effect of
1794 * freezing the array as once all stripe_heads have been collected,
1795 * no IO will be possible. Old stripe heads are freed once their
1796 * pages have been transferred over, and the old kmem_cache is
1797 * freed when all stripes are done.
1798 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1799 * we simple return a failre status - no need to clean anything up.
1800 * 4/ allocate new pages for the new slots in the new stripe_heads.
1801 * If this fails, we don't bother trying the shrink the
1802 * stripe_heads down again, we just leave them as they are.
1803 * As each stripe_head is processed the new one is released into
1804 * active service.
1805 *
1806 * Once step2 is started, we cannot afford to wait for a write,
1807 * so we use GFP_NOIO allocations.
1808 */
1809 struct stripe_head *osh, *nsh;
1810 LIST_HEAD(newstripes);
1811 struct disk_info *ndisks;
d6f38f31 1812 unsigned long cpu;
b5470dc5 1813 int err;
e18b890b 1814 struct kmem_cache *sc;
ad01c9e3 1815 int i;
566c09c5 1816 int hash, cnt;
ad01c9e3
N		 1817
1818 if (newsize <= conf->pool_size)
1819 return 0; /* never bother to shrink */
1820
b5470dc5
DW		 1821 err = md_allow_write(conf->mddev);
1822 if (err)
1823 return err;
2a2275d6 1824
ad01c9e3
N		 1825 /* Step 1 */
1826 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1827 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
20c2df83 1828 0, 0, NULL);
ad01c9e3
N		 1829 if (!sc)
1830 return -ENOMEM;
1831
1832 for (i = conf->max_nr_stripes; i; i--) {
6ce32846 1833 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
ad01c9e3
N		 1834 if (!nsh)
1835 break;
1836
ad01c9e3 1837 nsh->raid_conf = conf;
cb13ff69 1838 spin_lock_init(&nsh->stripe_lock);
ad01c9e3
N		 1839
1840 list_add(&nsh->lru, &newstripes);
1841 }
1842 if (i) {
1843 /* didn't get enough, give up */
1844 while (!list_empty(&newstripes)) {
1845 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1846 list_del(&nsh->lru);
1847 kmem_cache_free(sc, nsh);
1848 }
1849 kmem_cache_destroy(sc);
1850 return -ENOMEM;
1851 }
1852 /* Step 2 - Must use GFP_NOIO now.
1853 * OK, we have enough stripes, start collecting inactive
1854 * stripes and copying them over
1855 */
566c09c5
SL		 1856 hash = 0;
1857 cnt = 0;
ad01c9e3 1858 list_for_each_entry(nsh, &newstripes, lru) {
566c09c5
SL		 1859 lock_device_hash_lock(conf, hash);
1860 wait_event_cmd(conf->wait_for_stripe,
1861 !list_empty(conf->inactive_list + hash),
1862 unlock_device_hash_lock(conf, hash),
1863 lock_device_hash_lock(conf, hash));
1864 osh = get_free_stripe(conf, hash);
1865 unlock_device_hash_lock(conf, hash);
ad01c9e3 1866 atomic_set(&nsh->count, 1);
d592a996 1867 for(i=0; i<conf->pool_size; i++) {
ad01c9e3 1868 nsh->dev[i].page = osh->dev[i].page;
d592a996
SL		 1869 nsh->dev[i].orig_page = osh->dev[i].page;
1870 }
ad01c9e3
N		 1871 for( ; i<newsize; i++)
1872 nsh->dev[i].page = NULL;
566c09c5 1873 nsh->hash_lock_index = hash;
ad01c9e3 1874 kmem_cache_free(conf->slab_cache, osh);
566c09c5
SL		 1875 cnt++;
1876 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
1877 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
1878 hash++;
1879 cnt = 0;
1880 }
ad01c9e3
N		 1881 }
1882 kmem_cache_destroy(conf->slab_cache);
1883
1884 /* Step 3.
1885 * At this point, we are holding all the stripes so the array
1886 * is completely stalled, so now is a good time to resize
d6f38f31 1887 * conf->disks and the scribble region
ad01c9e3
N		 1888 */
1889 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1890 if (ndisks) {
1891 for (i=0; i<conf->raid_disks; i++)
1892 ndisks[i] = conf->disks[i];
1893 kfree(conf->disks);
1894 conf->disks = ndisks;
1895 } else
1896 err = -ENOMEM;
1897
d6f38f31
DW		 1898 get_online_cpus();
1899 conf->scribble_len = scribble_len(newsize);
1900 for_each_present_cpu(cpu) {
1901 struct raid5_percpu *percpu;
1902 void *scribble;
1903
1904 percpu = per_cpu_ptr(conf->percpu, cpu);
1905 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1906
1907 if (scribble) {
1908 kfree(percpu->scribble);
1909 percpu->scribble = scribble;
1910 } else {
1911 err = -ENOMEM;
1912 break;
1913 }
1914 }
1915 put_online_cpus();
1916
ad01c9e3
N		 1917 /* Step 4, return new stripes to service */
1918 while(!list_empty(&newstripes)) {
1919 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1920 list_del_init(&nsh->lru);
d6f38f31 1921
ad01c9e3
N		 1922 for (i=conf->raid_disks; i < newsize; i++)
1923 if (nsh->dev[i].page == NULL) {
1924 struct page *p = alloc_page(GFP_NOIO);
1925 nsh->dev[i].page = p;
d592a996 1926 nsh->dev[i].orig_page = p;
ad01c9e3
N		 1927 if (!p)
1928 err = -ENOMEM;
1929 }
1930 release_stripe(nsh);
1931 }
1932 /* critical section pass, GFP_NOIO no longer needed */
1933
1934 conf->slab_cache = sc;
1935 conf->active_name = 1-conf->active_name;
1936 conf->pool_size = newsize;
1937 return err;
1938}
1da177e4 1939
566c09c5 1940static int drop_one_stripe(struct r5conf *conf, int hash)
1da177e4
LT		 1941{
1942 struct stripe_head *sh;
1943
566c09c5
SL		 1944 spin_lock_irq(conf->hash_locks + hash);
1945 sh = get_free_stripe(conf, hash);
1946 spin_unlock_irq(conf->hash_locks + hash);
3f294f4f
N		 1947 if (!sh)
1948 return 0;
78bafebd 1949 BUG_ON(atomic_read(&sh->count));
e4e11e38 1950 shrink_buffers(sh);
3f294f4f
N		 1951 kmem_cache_free(conf->slab_cache, sh);
1952 atomic_dec(&conf->active_stripes);
1953 return 1;
1954}
1955
d1688a6d 1956static void shrink_stripes(struct r5conf *conf)
3f294f4f 1957{
566c09c5
SL		 1958 int hash;
1959 for (hash = 0; hash < NR_STRIPE_HASH_LOCKS; hash++)
1960 while (drop_one_stripe(conf, hash))
1961 ;
3f294f4f 1962
29fc7e3e
N		 1963 if (conf->slab_cache)
1964 kmem_cache_destroy(conf->slab_cache);
1da177e4
LT		 1965 conf->slab_cache = NULL;
1966}
1967
6712ecf8 1968static void raid5_end_read_request(struct bio * bi, int error)
1da177e4 1969{
99c0fb5f 1970 struct stripe_head *sh = bi->bi_private;
d1688a6d 1971 struct r5conf *conf = sh->raid_conf;
7ecaa1e6 1972 int disks = sh->disks, i;
1da177e4 1973 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
d6950432 1974 char b[BDEVNAME_SIZE];
dd054fce 1975 struct md_rdev *rdev = NULL;
05616be5 1976 sector_t s;
1da177e4
LT		 1977
1978 for (i=0 ; i<disks; i++)
1979 if (bi == &sh->dev[i].req)
1980 break;
1981
45b4233c
DW		 1982 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1983 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1da177e4
LT		 1984 uptodate);
1985 if (i == disks) {
1986 BUG();
6712ecf8 1987 return;
1da177e4 1988 }
14a75d3e 1989 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
dd054fce
N		 1990 /* If replacement finished while this request was outstanding,
1991 * 'replacement' might be NULL already.
1992 * In that case it moved down to 'rdev'.
1993 * rdev is not removed until all requests are finished.
1994 */
14a75d3e 1995 rdev = conf->disks[i].replacement;
dd054fce 1996 if (!rdev)
14a75d3e 1997 rdev = conf->disks[i].rdev;
1da177e4 1998
05616be5
N		 1999 if (use_new_offset(conf, sh))
2000 s = sh->sector + rdev->new_data_offset;
2001 else
2002 s = sh->sector + rdev->data_offset;
1da177e4 2003 if (uptodate) {
1da177e4 2004 set_bit(R5_UPTODATE, &sh->dev[i].flags);
4e5314b5 2005 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
14a75d3e
N		 2006 /* Note that this cannot happen on a
2007 * replacement device. We just fail those on
2008 * any error
2009 */
8bda470e
CD		 2010 printk_ratelimited(
2011 KERN_INFO
2012 "md/raid:%s: read error corrected"
2013 " (%lu sectors at %llu on %s)\n",
2014 mdname(conf->mddev), STRIPE_SECTORS,
05616be5 2015 (unsigned long long)s,
8bda470e 2016 bdevname(rdev->bdev, b));
ddd5115f 2017 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
4e5314b5
N		 2018 clear_bit(R5_ReadError, &sh->dev[i].flags);
2019 clear_bit(R5_ReWrite, &sh->dev[i].flags);
3f9e7c14 2020 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2021 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2022
14a75d3e
N		 2023 if (atomic_read(&rdev->read_errors))
2024 atomic_set(&rdev->read_errors, 0);
1da177e4 2025 } else {
14a75d3e 2026 const char *bdn = bdevname(rdev->bdev, b);
ba22dcbf 2027 int retry = 0;
2e8ac303 2028 int set_bad = 0;
d6950432 2029
1da177e4 2030 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
d6950432 2031 atomic_inc(&rdev->read_errors);
14a75d3e
N		 2032 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2033 printk_ratelimited(
2034 KERN_WARNING
2035 "md/raid:%s: read error on replacement device "
2036 "(sector %llu on %s).\n",
2037 mdname(conf->mddev),
05616be5 2038 (unsigned long long)s,
14a75d3e 2039 bdn);
2e8ac303 2040 else if (conf->mddev->degraded >= conf->max_degraded) {
2041 set_bad = 1;
8bda470e
CD		 2042 printk_ratelimited(
2043 KERN_WARNING
2044 "md/raid:%s: read error not correctable "
2045 "(sector %llu on %s).\n",
2046 mdname(conf->mddev),
05616be5 2047 (unsigned long long)s,
8bda470e 2048 bdn);
2e8ac303 2049 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
4e5314b5 2050 /* Oh, no!!! */
2e8ac303 2051 set_bad = 1;
8bda470e
CD		 2052 printk_ratelimited(
2053 KERN_WARNING
2054 "md/raid:%s: read error NOT corrected!! "
2055 "(sector %llu on %s).\n",
2056 mdname(conf->mddev),
05616be5 2057 (unsigned long long)s,
8bda470e 2058 bdn);
2e8ac303 2059 } else if (atomic_read(&rdev->read_errors)
ba22dcbf 2060 > conf->max_nr_stripes)
14f8d26b 2061 printk(KERN_WARNING
0c55e022 2062 "md/raid:%s: Too many read errors, failing device %s.\n",
d6950432 2063 mdname(conf->mddev), bdn);
ba22dcbf
N		 2064 else
2065 retry = 1;
edfa1f65
BY		 2066 if (set_bad && test_bit(In_sync, &rdev->flags)
2067 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2068 retry = 1;
ba22dcbf 2069 if (retry)
3f9e7c14 2070 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2071 set_bit(R5_ReadError, &sh->dev[i].flags);
2072 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2073 } else
2074 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
ba22dcbf 2075 else {
4e5314b5
N		 2076 clear_bit(R5_ReadError, &sh->dev[i].flags);
2077 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2e8ac303 2078 if (!(set_bad
2079 && test_bit(In_sync, &rdev->flags)
2080 && rdev_set_badblocks(
2081 rdev, sh->sector, STRIPE_SECTORS, 0)))
2082 md_error(conf->mddev, rdev);
ba22dcbf 2083 }
1da177e4 2084 }
14a75d3e 2085 rdev_dec_pending(rdev, conf->mddev);
1da177e4
LT		 2086 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2087 set_bit(STRIPE_HANDLE, &sh->state);
2088 release_stripe(sh);
1da177e4
LT		 2089}
2090
d710e138 2091static void raid5_end_write_request(struct bio *bi, int error)
1da177e4 2092{
99c0fb5f 2093 struct stripe_head *sh = bi->bi_private;
d1688a6d 2094 struct r5conf *conf = sh->raid_conf;
7ecaa1e6 2095 int disks = sh->disks, i;
977df362 2096 struct md_rdev *uninitialized_var(rdev);
1da177e4 2097 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
b84db560
N		 2098 sector_t first_bad;
2099 int bad_sectors;
977df362 2100 int replacement = 0;
1da177e4 2101
977df362
N		 2102 for (i = 0 ; i < disks; i++) {
2103 if (bi == &sh->dev[i].req) {
2104 rdev = conf->disks[i].rdev;
1da177e4 2105 break;
977df362
N		 2106 }
2107 if (bi == &sh->dev[i].rreq) {
2108 rdev = conf->disks[i].replacement;
dd054fce
N		 2109 if (rdev)
2110 replacement = 1;
2111 else
2112 /* rdev was removed and 'replacement'
2113 * replaced it. rdev is not removed
2114 * until all requests are finished.
2115 */
2116 rdev = conf->disks[i].rdev;
977df362
N		 2117 break;
2118 }
2119 }
45b4233c 2120 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1da177e4
LT		 2121 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2122 uptodate);
2123 if (i == disks) {
2124 BUG();
6712ecf8 2125 return;
1da177e4
LT		 2126 }
2127
977df362
N		 2128 if (replacement) {
2129 if (!uptodate)
2130 md_error(conf->mddev, rdev);
2131 else if (is_badblock(rdev, sh->sector,
2132 STRIPE_SECTORS,
2133 &first_bad, &bad_sectors))
2134 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2135 } else {
2136 if (!uptodate) {
9f97e4b1 2137 set_bit(STRIPE_DEGRADED, &sh->state);
977df362
N		 2138 set_bit(WriteErrorSeen, &rdev->flags);
2139 set_bit(R5_WriteError, &sh->dev[i].flags);
3a6de292
N		 2140 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2141 set_bit(MD_RECOVERY_NEEDED,
2142 &rdev->mddev->recovery);
977df362
N		 2143 } else if (is_badblock(rdev, sh->sector,
2144 STRIPE_SECTORS,
c0b32972 2145 &first_bad, &bad_sectors)) {
977df362 2146 set_bit(R5_MadeGood, &sh->dev[i].flags);
c0b32972
N		 2147 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2148 /* That was a successful write so make
2149 * sure it looks like we already did
2150 * a re-write.
2151 */
2152 set_bit(R5_ReWrite, &sh->dev[i].flags);
2153 }
977df362
N		 2154 }
2155 rdev_dec_pending(rdev, conf->mddev);
1da177e4 2156
977df362
N		 2157 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2158 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1da177e4 2159 set_bit(STRIPE_HANDLE, &sh->state);
c04be0aa 2160 release_stripe(sh);
1da177e4
LT		 2161}
2162
784052ec 2163static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
d592a996 2164
784052ec 2165static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1da177e4
LT		 2166{
2167 struct r5dev *dev = &sh->dev[i];
2168
2169 bio_init(&dev->req);
2170 dev->req.bi_io_vec = &dev->vec;
d592a996 2171 dev->req.bi_max_vecs = 1;
1da177e4
LT		 2172 dev->req.bi_private = sh;
2173
977df362
N		 2174 bio_init(&dev->rreq);
2175 dev->rreq.bi_io_vec = &dev->rvec;
d592a996 2176 dev->rreq.bi_max_vecs = 1;
977df362 2177 dev->rreq.bi_private = sh;
977df362 2178
1da177e4 2179 dev->flags = 0;
784052ec 2180 dev->sector = compute_blocknr(sh, i, previous);
1da177e4
LT		 2181}
2182
fd01b88c 2183static void error(struct mddev *mddev, struct md_rdev *rdev)
1da177e4
LT		 2184{
2185 char b[BDEVNAME_SIZE];
d1688a6d 2186 struct r5conf *conf = mddev->private;
908f4fbd 2187 unsigned long flags;
0c55e022 2188 pr_debug("raid456: error called\n");
1da177e4 2189
908f4fbd
N		 2190 spin_lock_irqsave(&conf->device_lock, flags);
2191 clear_bit(In_sync, &rdev->flags);
2192 mddev->degraded = calc_degraded(conf);
2193 spin_unlock_irqrestore(&conf->device_lock, flags);
2194 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2195
de393cde 2196 set_bit(Blocked, &rdev->flags);
6f8d0c77
N		 2197 set_bit(Faulty, &rdev->flags);
2198 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2199 printk(KERN_ALERT
2200 "md/raid:%s: Disk failure on %s, disabling device.\n"
2201 "md/raid:%s: Operation continuing on %d devices.\n",
2202 mdname(mddev),
2203 bdevname(rdev->bdev, b),
2204 mdname(mddev),
2205 conf->raid_disks - mddev->degraded);
16a53ecc 2206}
1da177e4
LT		 2207
2208/*
2209 * Input: a 'big' sector number,
2210 * Output: index of the data and parity disk, and the sector # in them.
2211 */
d1688a6d 2212static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
911d4ee8
N		 2213 int previous, int *dd_idx,
2214 struct stripe_head *sh)
1da177e4 2215{
6e3b96ed 2216 sector_t stripe, stripe2;
35f2a591 2217 sector_t chunk_number;
1da177e4 2218 unsigned int chunk_offset;
911d4ee8 2219 int pd_idx, qd_idx;
67cc2b81 2220 int ddf_layout = 0;
1da177e4 2221 sector_t new_sector;
e183eaed
N		 2222 int algorithm = previous ? conf->prev_algo
2223 : conf->algorithm;
09c9e5fa
AN		 2224 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2225 : conf->chunk_sectors;
112bf897
N		 2226 int raid_disks = previous ? conf->previous_raid_disks
2227 : conf->raid_disks;
2228 int data_disks = raid_disks - conf->max_degraded;
1da177e4
LT		 2229
2230 /* First compute the information on this sector */
2231
2232 /*
2233 * Compute the chunk number and the sector offset inside the chunk
2234 */
2235 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2236 chunk_number = r_sector;
1da177e4
LT		 2237
2238 /*
2239 * Compute the stripe number
2240 */
35f2a591
N		 2241 stripe = chunk_number;
2242 *dd_idx = sector_div(stripe, data_disks);
6e3b96ed 2243 stripe2 = stripe;
1da177e4
LT		 2244 /*
2245 * Select the parity disk based on the user selected algorithm.
2246 */
84789554 2247 pd_idx = qd_idx = -1;
16a53ecc
N		 2248 switch(conf->level) {
2249 case 4:
911d4ee8 2250 pd_idx = data_disks;
16a53ecc
N		 2251 break;
2252 case 5:
e183eaed 2253 switch (algorithm) {
1da177e4 2254 case ALGORITHM_LEFT_ASYMMETRIC:
6e3b96ed 2255 pd_idx = data_disks - sector_div(stripe2, raid_disks);
911d4ee8 2256 if (*dd_idx >= pd_idx)
1da177e4
LT		 2257 (*dd_idx)++;
2258 break;
2259 case ALGORITHM_RIGHT_ASYMMETRIC:
6e3b96ed 2260 pd_idx = sector_div(stripe2, raid_disks);
911d4ee8 2261 if (*dd_idx >= pd_idx)
1da177e4
LT		 2262 (*dd_idx)++;
2263 break;
2264 case ALGORITHM_LEFT_SYMMETRIC:
6e3b96ed 2265 pd_idx = data_disks - sector_div(stripe2, raid_disks);
911d4ee8 2266 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1da177e4
LT		 2267 break;
2268 case ALGORITHM_RIGHT_SYMMETRIC:
6e3b96ed 2269 pd_idx = sector_div(stripe2, raid_disks);
911d4ee8 2270 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1da177e4 2271 break;
99c0fb5f
N		 2272 case ALGORITHM_PARITY_0:
2273 pd_idx = 0;
2274 (*dd_idx)++;
2275 break;
2276 case ALGORITHM_PARITY_N:
2277 pd_idx = data_disks;
2278 break;
1da177e4 2279 default:
99c0fb5f 2280 BUG();
16a53ecc
N		 2281 }
2282 break;
2283 case 6:
2284
e183eaed 2285 switch (algorithm) {
16a53ecc 2286 case ALGORITHM_LEFT_ASYMMETRIC:
6e3b96ed 2287 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
911d4ee8
N		 2288 qd_idx = pd_idx + 1;
2289 if (pd_idx == raid_disks-1) {
99c0fb5f 2290 (*dd_idx)++; /* Q D D D P */
911d4ee8
N		 2291 qd_idx = 0;
2292 } else if (*dd_idx >= pd_idx)
16a53ecc
N		 2293 (*dd_idx) += 2; /* D D P Q D */
2294 break;
2295 case ALGORITHM_RIGHT_ASYMMETRIC:
6e3b96ed 2296 pd_idx = sector_div(stripe2, raid_disks);
911d4ee8
N		 2297 qd_idx = pd_idx + 1;
2298 if (pd_idx == raid_disks-1) {
99c0fb5f 2299 (*dd_idx)++; /* Q D D D P */
911d4ee8
N		 2300 qd_idx = 0;
2301 } else if (*dd_idx >= pd_idx)
16a53ecc
N		 2302 (*dd_idx) += 2; /* D D P Q D */
2303 break;
2304 case ALGORITHM_LEFT_SYMMETRIC:
6e3b96ed 2305 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
911d4ee8
N		 2306 qd_idx = (pd_idx + 1) % raid_disks;
2307 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
16a53ecc
N		 2308 break;
2309 case ALGORITHM_RIGHT_SYMMETRIC:
6e3b96ed 2310 pd_idx = sector_div(stripe2, raid_disks);
911d4ee8
N		 2311 qd_idx = (pd_idx + 1) % raid_disks;
2312 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
16a53ecc 2313 break;
99c0fb5f
N		 2314
2315 case ALGORITHM_PARITY_0:
2316 pd_idx = 0;
2317 qd_idx = 1;
2318 (*dd_idx) += 2;
2319 break;
2320 case ALGORITHM_PARITY_N:
2321 pd_idx = data_disks;
2322 qd_idx = data_disks + 1;
2323 break;
2324
2325 case ALGORITHM_ROTATING_ZERO_RESTART:
2326 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2327 * of blocks for computing Q is different.
2328 */
6e3b96ed 2329 pd_idx = sector_div(stripe2, raid_disks);
99c0fb5f
N		 2330 qd_idx = pd_idx + 1;
2331 if (pd_idx == raid_disks-1) {
2332 (*dd_idx)++; /* Q D D D P */
2333 qd_idx = 0;
2334 } else if (*dd_idx >= pd_idx)
2335 (*dd_idx) += 2; /* D D P Q D */
67cc2b81 2336 ddf_layout = 1;
99c0fb5f
N		 2337 break;
2338
2339 case ALGORITHM_ROTATING_N_RESTART:
2340 /* Same a left_asymmetric, by first stripe is
2341 * D D D P Q rather than
2342 * Q D D D P
2343 */
6e3b96ed
N		 2344 stripe2 += 1;
2345 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
99c0fb5f
N		 2346 qd_idx = pd_idx + 1;
2347 if (pd_idx == raid_disks-1) {
2348 (*dd_idx)++; /* Q D D D P */
2349 qd_idx = 0;
2350 } else if (*dd_idx >= pd_idx)
2351 (*dd_idx) += 2; /* D D P Q D */
67cc2b81 2352 ddf_layout = 1;
99c0fb5f
N		 2353 break;
2354
2355 case ALGORITHM_ROTATING_N_CONTINUE:
2356 /* Same as left_symmetric but Q is before P */
6e3b96ed 2357 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
99c0fb5f
N		 2358 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2359 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
67cc2b81 2360 ddf_layout = 1;
99c0fb5f
N		 2361 break;
2362
2363 case ALGORITHM_LEFT_ASYMMETRIC_6:
2364 /* RAID5 left_asymmetric, with Q on last device */
6e3b96ed 2365 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
99c0fb5f
N		 2366 if (*dd_idx >= pd_idx)
2367 (*dd_idx)++;
2368 qd_idx = raid_disks - 1;
2369 break;
2370
2371 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6e3b96ed 2372 pd_idx = sector_div(stripe2, raid_disks-1);
99c0fb5f
N		 2373 if (*dd_idx >= pd_idx)
2374 (*dd_idx)++;
2375 qd_idx = raid_disks - 1;
2376 break;
2377
2378 case ALGORITHM_LEFT_SYMMETRIC_6:
6e3b96ed 2379 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
99c0fb5f
N		 2380 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2381 qd_idx = raid_disks - 1;
2382 break;
2383
2384 case ALGORITHM_RIGHT_SYMMETRIC_6:
6e3b96ed 2385 pd_idx = sector_div(stripe2, raid_disks-1);
99c0fb5f
N		 2386 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2387 qd_idx = raid_disks - 1;
2388 break;
2389
2390 case ALGORITHM_PARITY_0_6:
2391 pd_idx = 0;
2392 (*dd_idx)++;
2393 qd_idx = raid_disks - 1;
2394 break;
2395
16a53ecc 2396 default:
99c0fb5f 2397 BUG();
16a53ecc
N		 2398 }
2399 break;
1da177e4
LT		 2400 }
2401
911d4ee8
N		 2402 if (sh) {
2403 sh->pd_idx = pd_idx;
2404 sh->qd_idx = qd_idx;
67cc2b81 2405 sh->ddf_layout = ddf_layout;
911d4ee8 2406 }
1da177e4
LT		 2407 /*
2408 * Finally, compute the new sector number
2409 */
2410 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2411 return new_sector;
2412}
2413
784052ec 2414static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1da177e4 2415{
d1688a6d 2416 struct r5conf *conf = sh->raid_conf;
b875e531
N		 2417 int raid_disks = sh->disks;
2418 int data_disks = raid_disks - conf->max_degraded;
1da177e4 2419 sector_t new_sector = sh->sector, check;
09c9e5fa
AN		 2420 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2421 : conf->chunk_sectors;
e183eaed
N		 2422 int algorithm = previous ? conf->prev_algo
2423 : conf->algorithm;
1da177e4
LT		 2424 sector_t stripe;
2425 int chunk_offset;
35f2a591
N		 2426 sector_t chunk_number;
2427 int dummy1, dd_idx = i;
1da177e4 2428 sector_t r_sector;
911d4ee8 2429 struct stripe_head sh2;
1da177e4
LT		 2430
2431 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2432 stripe = new_sector;
1da177e4 2433
16a53ecc
N		 2434 if (i == sh->pd_idx)
2435 return 0;
2436 switch(conf->level) {
2437 case 4: break;
2438 case 5:
e183eaed 2439 switch (algorithm) {
1da177e4
LT		 2440 case ALGORITHM_LEFT_ASYMMETRIC:
2441 case ALGORITHM_RIGHT_ASYMMETRIC:
2442 if (i > sh->pd_idx)
2443 i--;
2444 break;
2445 case ALGORITHM_LEFT_SYMMETRIC:
2446 case ALGORITHM_RIGHT_SYMMETRIC:
2447 if (i < sh->pd_idx)
2448 i += raid_disks;
2449 i -= (sh->pd_idx + 1);
2450 break;
99c0fb5f
N		 2451 case ALGORITHM_PARITY_0:
2452 i -= 1;
2453 break;
2454 case ALGORITHM_PARITY_N:
2455 break;
1da177e4 2456 default:
99c0fb5f 2457 BUG();
16a53ecc
N		 2458 }
2459 break;
2460 case 6:
d0dabf7e 2461 if (i == sh->qd_idx)
16a53ecc 2462 return 0; /* It is the Q disk */
e183eaed 2463 switch (algorithm) {
16a53ecc
N		 2464 case ALGORITHM_LEFT_ASYMMETRIC:
2465 case ALGORITHM_RIGHT_ASYMMETRIC:
99c0fb5f
N		 2466 case ALGORITHM_ROTATING_ZERO_RESTART:
2467 case ALGORITHM_ROTATING_N_RESTART:
2468 if (sh->pd_idx == raid_disks-1)
2469 i--; /* Q D D D P */
16a53ecc
N		 2470 else if (i > sh->pd_idx)
2471 i -= 2; /* D D P Q D */
2472 break;
2473 case ALGORITHM_LEFT_SYMMETRIC:
2474 case ALGORITHM_RIGHT_SYMMETRIC:
2475 if (sh->pd_idx == raid_disks-1)
2476 i--; /* Q D D D P */
2477 else {
2478 /* D D P Q D */
2479 if (i < sh->pd_idx)
2480 i += raid_disks;
2481 i -= (sh->pd_idx + 2);
2482 }
2483 break;
99c0fb5f
N		 2484 case ALGORITHM_PARITY_0:
2485 i -= 2;
2486 break;
2487 case ALGORITHM_PARITY_N:
2488 break;
2489 case ALGORITHM_ROTATING_N_CONTINUE:
e4424fee 2490 /* Like left_symmetric, but P is before Q */
99c0fb5f
N		 2491 if (sh->pd_idx == 0)
2492 i--; /* P D D D Q */
e4424fee
N		 2493 else {
2494 /* D D Q P D */
2495 if (i < sh->pd_idx)
2496 i += raid_disks;
2497 i -= (sh->pd_idx + 1);
2498 }
99c0fb5f
N		 2499 break;
2500 case ALGORITHM_LEFT_ASYMMETRIC_6:
2501 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2502 if (i > sh->pd_idx)
2503 i--;
2504 break;
2505 case ALGORITHM_LEFT_SYMMETRIC_6:
2506 case ALGORITHM_RIGHT_SYMMETRIC_6:
2507 if (i < sh->pd_idx)
2508 i += data_disks + 1;
2509 i -= (sh->pd_idx + 1);
2510 break;
2511 case ALGORITHM_PARITY_0_6:
2512 i -= 1;
2513 break;
16a53ecc 2514 default:
99c0fb5f 2515 BUG();
16a53ecc
N		 2516 }
2517 break;
1da177e4
LT		 2518 }
2519
2520 chunk_number = stripe * data_disks + i;
35f2a591 2521 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
1da177e4 2522
112bf897 2523 check = raid5_compute_sector(conf, r_sector,
784052ec 2524 previous, &dummy1, &sh2);
911d4ee8
N		 2525 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2526 || sh2.qd_idx != sh->qd_idx) {
0c55e022
N		 2527 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2528 mdname(conf->mddev));
1da177e4
LT		 2529 return 0;
2530 }
2531 return r_sector;
2532}
2533
600aa109 2534static void
c0f7bddb 2535schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
600aa109 2536 int rcw, int expand)
e33129d8
DW		 2537{
2538 int i, pd_idx = sh->pd_idx, disks = sh->disks;
d1688a6d 2539 struct r5conf *conf = sh->raid_conf;
c0f7bddb 2540 int level = conf->level;
e33129d8
DW		 2541
2542 if (rcw) {
e33129d8
DW		 2543
2544 for (i = disks; i--; ) {
2545 struct r5dev *dev = &sh->dev[i];
2546
2547 if (dev->towrite) {
2548 set_bit(R5_LOCKED, &dev->flags);
d8ee0728 2549 set_bit(R5_Wantdrain, &dev->flags);
e33129d8
DW		 2550 if (!expand)
2551 clear_bit(R5_UPTODATE, &dev->flags);
600aa109 2552 s->locked++;
e33129d8
DW		 2553 }
2554 }
ce7d363a
N		 2555 /* if we are not expanding this is a proper write request, and
2556 * there will be bios with new data to be drained into the
2557 * stripe cache
2558 */
2559 if (!expand) {
2560 if (!s->locked)
2561 /* False alarm, nothing to do */
2562 return;
2563 sh->reconstruct_state = reconstruct_state_drain_run;
2564 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2565 } else
2566 sh->reconstruct_state = reconstruct_state_run;
2567
2568 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2569
c0f7bddb 2570 if (s->locked + conf->max_degraded == disks)
8b3e6cdc 2571 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
c0f7bddb 2572 atomic_inc(&conf->pending_full_writes);
e33129d8 2573 } else {
c0f7bddb 2574 BUG_ON(level == 6);
e33129d8
DW		 2575 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2576 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2577
e33129d8
DW		 2578 for (i = disks; i--; ) {
2579 struct r5dev *dev = &sh->dev[i];
2580 if (i == pd_idx)
2581 continue;
2582
e33129d8
DW		 2583 if (dev->towrite &&
2584 (test_bit(R5_UPTODATE, &dev->flags) ||
d8ee0728
DW		 2585 test_bit(R5_Wantcompute, &dev->flags))) {
2586 set_bit(R5_Wantdrain, &dev->flags);
e33129d8
DW		 2587 set_bit(R5_LOCKED, &dev->flags);
2588 clear_bit(R5_UPTODATE, &dev->flags);
600aa109 2589 s->locked++;
e33129d8
DW		 2590 }
2591 }
ce7d363a
N		 2592 if (!s->locked)
2593 /* False alarm - nothing to do */
2594 return;
2595 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2596 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2597 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2598 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
e33129d8
DW		 2599 }
2600
c0f7bddb 2601 /* keep the parity disk(s) locked while asynchronous operations
e33129d8
DW		 2602 * are in flight
2603 */
2604 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2605 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
600aa109 2606 s->locked++;
e33129d8 2607
c0f7bddb
YT		 2608 if (level == 6) {
2609 int qd_idx = sh->qd_idx;
2610 struct r5dev *dev = &sh->dev[qd_idx];
2611
2612 set_bit(R5_LOCKED, &dev->flags);
2613 clear_bit(R5_UPTODATE, &dev->flags);
2614 s->locked++;
2615 }
2616
600aa109 2617 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
e46b272b 2618 __func__, (unsigned long long)sh->sector,
600aa109 2619 s->locked, s->ops_request);
e33129d8 2620}
16a53ecc 2621
1da177e4
LT		 2622/*
2623 * Each stripe/dev can have one or more bion attached.
16a53ecc 2624 * toread/towrite point to the first in a chain.
1da177e4
LT		 2625 * The bi_next chain must be in order.
2626 */
2627static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2628{
2629 struct bio **bip;
d1688a6d 2630 struct r5conf *conf = sh->raid_conf;
72626685 2631 int firstwrite=0;
1da177e4 2632
cbe47ec5 2633 pr_debug("adding bi b#%llu to stripe s#%llu\n",
4f024f37 2634 (unsigned long long)bi->bi_iter.bi_sector,
1da177e4
LT		 2635 (unsigned long long)sh->sector);
2636
b17459c0
SL		 2637 /*
2638 * If several bio share a stripe. The bio bi_phys_segments acts as a
2639 * reference count to avoid race. The reference count should already be
2640 * increased before this function is called (for example, in
2641 * make_request()), so other bio sharing this stripe will not free the
2642 * stripe. If a stripe is owned by one stripe, the stripe lock will
2643 * protect it.
2644 */
2645 spin_lock_irq(&sh->stripe_lock);
72626685 2646 if (forwrite) {
1da177e4 2647 bip = &sh->dev[dd_idx].towrite;
7eaf7e8e 2648 if (*bip == NULL)
72626685
N		 2649 firstwrite = 1;
2650 } else
1da177e4 2651 bip = &sh->dev[dd_idx].toread;
4f024f37
KO		 2652 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2653 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
1da177e4
LT		 2654 goto overlap;
2655 bip = & (*bip)->bi_next;
2656 }
4f024f37 2657 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
1da177e4
LT		 2658 goto overlap;
2659
78bafebd 2660 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1da177e4
LT		 2661 if (*bip)
2662 bi->bi_next = *bip;
2663 *bip = bi;
e7836bd6 2664 raid5_inc_bi_active_stripes(bi);
72626685 2665
1da177e4
LT		 2666 if (forwrite) {
2667 /* check if page is covered */
2668 sector_t sector = sh->dev[dd_idx].sector;
2669 for (bi=sh->dev[dd_idx].towrite;
2670 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
4f024f37 2671 bi && bi->bi_iter.bi_sector <= sector;
1da177e4 2672 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
f73a1c7d
KO		 2673 if (bio_end_sector(bi) >= sector)
2674 sector = bio_end_sector(bi);
1da177e4
LT		 2675 }
2676 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2677 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2678 }
cbe47ec5
N		 2679
2680 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
4f024f37 2681 (unsigned long long)(*bip)->bi_iter.bi_sector,
cbe47ec5 2682 (unsigned long long)sh->sector, dd_idx);
b97390ae 2683 spin_unlock_irq(&sh->stripe_lock);
cbe47ec5
N		 2684
2685 if (conf->mddev->bitmap && firstwrite) {
2686 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2687 STRIPE_SECTORS, 0);
2688 sh->bm_seq = conf->seq_flush+1;
2689 set_bit(STRIPE_BIT_DELAY, &sh->state);
2690 }
1da177e4
LT		 2691 return 1;
2692
2693 overlap:
2694 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
b17459c0 2695 spin_unlock_irq(&sh->stripe_lock);
1da177e4
LT		 2696 return 0;
2697}
2698
d1688a6d 2699static void end_reshape(struct r5conf *conf);
29269553 2700
d1688a6d 2701static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
911d4ee8 2702 struct stripe_head *sh)
ccfcc3c1 2703{
784052ec 2704 int sectors_per_chunk =
09c9e5fa 2705 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
911d4ee8 2706 int dd_idx;
2d2063ce 2707 int chunk_offset = sector_div(stripe, sectors_per_chunk);
112bf897 2708 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2d2063ce 2709
112bf897
N		 2710 raid5_compute_sector(conf,
2711 stripe * (disks - conf->max_degraded)
b875e531 2712 *sectors_per_chunk + chunk_offset,
112bf897 2713 previous,
911d4ee8 2714 &dd_idx, sh);
ccfcc3c1
N		 2715}
2716
a4456856 2717static void
d1688a6d 2718handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
a4456856
DW		 2719 struct stripe_head_state *s, int disks,
2720 struct bio **return_bi)
2721{
2722 int i;
2723 for (i = disks; i--; ) {
2724 struct bio *bi;
2725 int bitmap_end = 0;
2726
2727 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3cb03002 2728 struct md_rdev *rdev;
a4456856
DW		 2729 rcu_read_lock();
2730 rdev = rcu_dereference(conf->disks[i].rdev);
2731 if (rdev && test_bit(In_sync, &rdev->flags))
7f0da59b
N		 2732 atomic_inc(&rdev->nr_pending);
2733 else
2734 rdev = NULL;
a4456856 2735 rcu_read_unlock();
7f0da59b
N		 2736 if (rdev) {
2737 if (!rdev_set_badblocks(
2738 rdev,
2739 sh->sector,
2740 STRIPE_SECTORS, 0))
2741 md_error(conf->mddev, rdev);
2742 rdev_dec_pending(rdev, conf->mddev);
2743 }
a4456856 2744 }
b17459c0 2745 spin_lock_irq(&sh->stripe_lock);
a4456856
DW		 2746 /* fail all writes first */
2747 bi = sh->dev[i].towrite;
2748 sh->dev[i].towrite = NULL;
b17459c0 2749 spin_unlock_irq(&sh->stripe_lock);
1ed850f3 2750 if (bi)
a4456856 2751 bitmap_end = 1;
a4456856
DW		 2752
2753 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2754 wake_up(&conf->wait_for_overlap);
2755
4f024f37 2756 while (bi && bi->bi_iter.bi_sector <
a4456856
DW		 2757 sh->dev[i].sector + STRIPE_SECTORS) {
2758 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2759 clear_bit(BIO_UPTODATE, &bi->bi_flags);
e7836bd6 2760 if (!raid5_dec_bi_active_stripes(bi)) {
a4456856
DW		 2761 md_write_end(conf->mddev);
2762 bi->bi_next = *return_bi;
2763 *return_bi = bi;
2764 }
2765 bi = nextbi;
2766 }
7eaf7e8e
SL		 2767 if (bitmap_end)
2768 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2769 STRIPE_SECTORS, 0, 0);
2770 bitmap_end = 0;
a4456856
DW		 2771 /* and fail all 'written' */
2772 bi = sh->dev[i].written;
2773 sh->dev[i].written = NULL;
d592a996
SL		 2774 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
2775 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
2776 sh->dev[i].page = sh->dev[i].orig_page;
2777 }
2778
a4456856 2779 if (bi) bitmap_end = 1;
4f024f37 2780 while (bi && bi->bi_iter.bi_sector <
a4456856
DW		 2781 sh->dev[i].sector + STRIPE_SECTORS) {
2782 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2783 clear_bit(BIO_UPTODATE, &bi->bi_flags);
e7836bd6 2784 if (!raid5_dec_bi_active_stripes(bi)) {
a4456856
DW		 2785 md_write_end(conf->mddev);
2786 bi->bi_next = *return_bi;
2787 *return_bi = bi;
2788 }
2789 bi = bi2;
2790 }
2791
b5e98d65
DW		 2792 /* fail any reads if this device is non-operational and
2793 * the data has not reached the cache yet.
2794 */
2795 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2796 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2797 test_bit(R5_ReadError, &sh->dev[i].flags))) {
143c4d05 2798 spin_lock_irq(&sh->stripe_lock);
a4456856
DW		 2799 bi = sh->dev[i].toread;
2800 sh->dev[i].toread = NULL;
143c4d05 2801 spin_unlock_irq(&sh->stripe_lock);
a4456856
DW		 2802 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2803 wake_up(&conf->wait_for_overlap);
4f024f37 2804 while (bi && bi->bi_iter.bi_sector <
a4456856
DW		 2805 sh->dev[i].sector + STRIPE_SECTORS) {
2806 struct bio *nextbi =
2807 r5_next_bio(bi, sh->dev[i].sector);
2808 clear_bit(BIO_UPTODATE, &bi->bi_flags);
e7836bd6 2809 if (!raid5_dec_bi_active_stripes(bi)) {
a4456856
DW		 2810 bi->bi_next = *return_bi;
2811 *return_bi = bi;
2812 }
2813 bi = nextbi;
2814 }
2815 }
a4456856
DW		 2816 if (bitmap_end)
2817 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2818 STRIPE_SECTORS, 0, 0);
8cfa7b0f
N		 2819 /* If we were in the middle of a write the parity block might
2820 * still be locked - so just clear all R5_LOCKED flags
2821 */
2822 clear_bit(R5_LOCKED, &sh->dev[i].flags);
a4456856
DW		 2823 }
2824
8b3e6cdc
DW		 2825 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2826 if (atomic_dec_and_test(&conf->pending_full_writes))
2827 md_wakeup_thread(conf->mddev->thread);
a4456856
DW		 2828}
2829
7f0da59b 2830static void
d1688a6d 2831handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
7f0da59b
N		 2832 struct stripe_head_state *s)
2833{
2834 int abort = 0;
2835 int i;
2836
7f0da59b 2837 clear_bit(STRIPE_SYNCING, &sh->state);
f8dfcffd
N		 2838 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2839 wake_up(&conf->wait_for_overlap);
7f0da59b 2840 s->syncing = 0;
9a3e1101 2841 s->replacing = 0;
7f0da59b 2842 /* There is nothing more to do for sync/check/repair.
18b9837e
N		 2843 * Don't even need to abort as that is handled elsewhere
2844 * if needed, and not always wanted e.g. if there is a known
2845 * bad block here.
9a3e1101 2846 * For recover/replace we need to record a bad block on all
7f0da59b
N		 2847 * non-sync devices, or abort the recovery
2848 */
18b9837e
N		 2849 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2850 /* During recovery devices cannot be removed, so
2851 * locking and refcounting of rdevs is not needed
2852 */
2853 for (i = 0; i < conf->raid_disks; i++) {
2854 struct md_rdev *rdev = conf->disks[i].rdev;
2855 if (rdev
2856 && !test_bit(Faulty, &rdev->flags)
2857 && !test_bit(In_sync, &rdev->flags)
2858 && !rdev_set_badblocks(rdev, sh->sector,
2859 STRIPE_SECTORS, 0))
2860 abort = 1;
2861 rdev = conf->disks[i].replacement;
2862 if (rdev
2863 && !test_bit(Faulty, &rdev->flags)
2864 && !test_bit(In_sync, &rdev->flags)
2865 && !rdev_set_badblocks(rdev, sh->sector,
2866 STRIPE_SECTORS, 0))
2867 abort = 1;
2868 }
2869 if (abort)
2870 conf->recovery_disabled =
2871 conf->mddev->recovery_disabled;
7f0da59b 2872 }
18b9837e 2873 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
7f0da59b
N		 2874}
2875
9a3e1101
N		 2876static int want_replace(struct stripe_head *sh, int disk_idx)
2877{
2878 struct md_rdev *rdev;
2879 int rv = 0;
2880 /* Doing recovery so rcu locking not required */
2881 rdev = sh->raid_conf->disks[disk_idx].replacement;
2882 if (rdev
2883 && !test_bit(Faulty, &rdev->flags)
2884 && !test_bit(In_sync, &rdev->flags)
2885 && (rdev->recovery_offset <= sh->sector
2886 || rdev->mddev->recovery_cp <= sh->sector))
2887 rv = 1;
2888
2889 return rv;
2890}
2891
93b3dbce 2892/* fetch_block - checks the given member device to see if its data needs
1fe797e6
DW		 2893 * to be read or computed to satisfy a request.
2894 *
2895 * Returns 1 when no more member devices need to be checked, otherwise returns
93b3dbce 2896 * 0 to tell the loop in handle_stripe_fill to continue
f38e1219 2897 */
2c58f06e
N		 2898
2899static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
2900 int disk_idx, int disks)
a4456856 2901{
5599becc 2902 struct r5dev *dev = &sh->dev[disk_idx];
f2b3b44d
N		 2903 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2904 &sh->dev[s->failed_num[1]] };
ea664c82 2905 int i;
5599becc 2906
a79cfe12
N		 2907
2908 if (test_bit(R5_LOCKED, &dev->flags) ||
2909 test_bit(R5_UPTODATE, &dev->flags))
2910 /* No point reading this as we already have it or have
2911 * decided to get it.
2912 */
2913 return 0;
2914
2915 if (dev->toread ||
2916 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
2917 /* We need this block to directly satisfy a request */
2918 return 1;
2919
2920 if (s->syncing || s->expanding ||
2921 (s->replacing && want_replace(sh, disk_idx)))
2922 /* When syncing, or expanding we read everything.
2923 * When replacing, we need the replaced block.
2924 */
2925 return 1;
2926
2927 if ((s->failed >= 1 && fdev[0]->toread) ||
2928 (s->failed >= 2 && fdev[1]->toread))
2929 /* If we want to read from a failed device, then
2930 * we need to actually read every other device.
2931 */
2932 return 1;
2933
a9d56950
N		 2934 /* Sometimes neither read-modify-write nor reconstruct-write
2935 * cycles can work. In those cases we read every block we
2936 * can. Then the parity-update is certain to have enough to
2937 * work with.
2938 * This can only be a problem when we need to write something,
2939 * and some device has failed. If either of those tests
2940 * fail we need look no further.
2941 */
2942 if (!s->failed || !s->to_write)
2943 return 0;
2944
2945 if (test_bit(R5_Insync, &dev->flags) &&
2946 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
2947 /* Pre-reads at not permitted until after short delay
2948 * to gather multiple requests. However if this
2949 * device is no Insync, the block could only be be computed
2950 * and there is no need to delay that.
2951 */
2952 return 0;
ea664c82
N		 2953
2954 for (i = 0; i < s->failed; i++) {
2955 if (fdev[i]->towrite &&
2956 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
2957 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
2958 /* If we have a partial write to a failed
2959 * device, then we will need to reconstruct
2960 * the content of that device, so all other
2961 * devices must be read.
2962 */
2963 return 1;
2964 }
2965
2966 /* If we are forced to do a reconstruct-write, either because
2967 * the current RAID6 implementation only supports that, or
2968 * or because parity cannot be trusted and we are currently
2969 * recovering it, there is extra need to be careful.
2970 * If one of the devices that we would need to read, because
2971 * it is not being overwritten (and maybe not written at all)
2972 * is missing/faulty, then we need to read everything we can.
2973 */
2974 if (sh->raid_conf->level != 6 &&
2975 sh->sector < sh->raid_conf->mddev->recovery_cp)
2976 /* reconstruct-write isn't being forced */
2977 return 0;
2978 for (i = 0; i < s->failed; i++) {
2979 if (!test_bit(R5_UPTODATE, &fdev[i]->flags) &&
2980 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
2981 return 1;
2982 }
2983
2c58f06e
N		 2984 return 0;
2985}
2986
2987static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2988 int disk_idx, int disks)
2989{
2990 struct r5dev *dev = &sh->dev[disk_idx];
2991
2992 /* is the data in this block needed, and can we get it? */
2993 if (need_this_block(sh, s, disk_idx, disks)) {
5599becc
YT		 2994 /* we would like to get this block, possibly by computing it,
2995 * otherwise read it if the backing disk is insync
2996 */
2997 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2998 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2999 if ((s->uptodate == disks - 1) &&
f2b3b44d
N		 3000 (s->failed && (disk_idx == s->failed_num[0] ||
3001 disk_idx == s->failed_num[1]))) {
5599becc
YT		 3002 /* have disk failed, and we're requested to fetch it;
3003 * do compute it
a4456856 3004 */
5599becc
YT		 3005 pr_debug("Computing stripe %llu block %d\n",
3006 (unsigned long long)sh->sector, disk_idx);
3007 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3008 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3009 set_bit(R5_Wantcompute, &dev->flags);
3010 sh->ops.target = disk_idx;
3011 sh->ops.target2 = -1; /* no 2nd target */
3012 s->req_compute = 1;
93b3dbce
N		 3013 /* Careful: from this point on 'uptodate' is in the eye
3014 * of raid_run_ops which services 'compute' operations
3015 * before writes. R5_Wantcompute flags a block that will
3016 * be R5_UPTODATE by the time it is needed for a
3017 * subsequent operation.
3018 */
5599becc
YT		 3019 s->uptodate++;
3020 return 1;
3021 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3022 /* Computing 2-failure is *very* expensive; only
3023 * do it if failed >= 2
3024 */
3025 int other;
3026 for (other = disks; other--; ) {
3027 if (other == disk_idx)
3028 continue;
3029 if (!test_bit(R5_UPTODATE,
3030 &sh->dev[other].flags))
3031 break;
a4456856 3032 }
5599becc
YT		 3033 BUG_ON(other < 0);
3034 pr_debug("Computing stripe %llu blocks %d,%d\n",
3035 (unsigned long long)sh->sector,
3036 disk_idx, other);
3037 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3038 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3039 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3040 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3041 sh->ops.target = disk_idx;
3042 sh->ops.target2 = other;
3043 s->uptodate += 2;
3044 s->req_compute = 1;
3045 return 1;
3046 } else if (test_bit(R5_Insync, &dev->flags)) {
3047 set_bit(R5_LOCKED, &dev->flags);
3048 set_bit(R5_Wantread, &dev->flags);
3049 s->locked++;
3050 pr_debug("Reading block %d (sync=%d)\n",
3051 disk_idx, s->syncing);
a4456856
DW		 3052 }
3053 }
5599becc
YT		 3054
3055 return 0;
3056}
3057
3058/**
93b3dbce 3059 * handle_stripe_fill - read or compute data to satisfy pending requests.
5599becc 3060 */
93b3dbce
N		 3061static void handle_stripe_fill(struct stripe_head *sh,
3062 struct stripe_head_state *s,
3063 int disks)
5599becc
YT		 3064{
3065 int i;
3066
3067 /* look for blocks to read/compute, skip this if a compute
3068 * is already in flight, or if the stripe contents are in the
3069 * midst of changing due to a write
3070 */
3071 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3072 !sh->reconstruct_state)
3073 for (i = disks; i--; )
93b3dbce 3074 if (fetch_block(sh, s, i, disks))
5599becc 3075 break;
a4456856
DW		 3076 set_bit(STRIPE_HANDLE, &sh->state);
3077}
3078
1fe797e6 3079/* handle_stripe_clean_event
a4456856
DW		 3080 * any written block on an uptodate or failed drive can be returned.
3081 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3082 * never LOCKED, so we don't need to test 'failed' directly.
3083 */
d1688a6d 3084static void handle_stripe_clean_event(struct r5conf *conf,
a4456856
DW		 3085 struct stripe_head *sh, int disks, struct bio **return_bi)
3086{
3087 int i;
3088 struct r5dev *dev;
f8dfcffd 3089 int discard_pending = 0;
a4456856
DW		 3090
3091 for (i = disks; i--; )
3092 if (sh->dev[i].written) {
3093 dev = &sh->dev[i];
3094 if (!test_bit(R5_LOCKED, &dev->flags) &&
9e444768 3095 (test_bit(R5_UPTODATE, &dev->flags) ||
d592a996
SL		 3096 test_bit(R5_Discard, &dev->flags) ||
3097 test_bit(R5_SkipCopy, &dev->flags))) {
a4456856
DW		 3098 /* We can return any write requests */
3099 struct bio *wbi, *wbi2;
45b4233c 3100 pr_debug("Return write for disc %d\n", i);
ca64cae9
N		 3101 if (test_and_clear_bit(R5_Discard, &dev->flags))
3102 clear_bit(R5_UPTODATE, &dev->flags);
d592a996
SL		 3103 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3104 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3105 dev->page = dev->orig_page;
3106 }
a4456856
DW		 3107 wbi = dev->written;
3108 dev->written = NULL;
4f024f37 3109 while (wbi && wbi->bi_iter.bi_sector <
a4456856
DW		 3110 dev->sector + STRIPE_SECTORS) {
3111 wbi2 = r5_next_bio(wbi, dev->sector);
e7836bd6 3112 if (!raid5_dec_bi_active_stripes(wbi)) {
a4456856
DW		 3113 md_write_end(conf->mddev);
3114 wbi->bi_next = *return_bi;
3115 *return_bi = wbi;
3116 }
3117 wbi = wbi2;
3118 }
7eaf7e8e
SL		 3119 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3120 STRIPE_SECTORS,
a4456856 3121 !test_bit(STRIPE_DEGRADED, &sh->state),
7eaf7e8e 3122 0);
f8dfcffd
N		 3123 } else if (test_bit(R5_Discard, &dev->flags))
3124 discard_pending = 1;
d592a996
SL		 3125 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3126 WARN_ON(dev->page != dev->orig_page);
f8dfcffd
N		 3127 }
3128 if (!discard_pending &&
3129 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3130 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3131 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3132 if (sh->qd_idx >= 0) {
3133 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3134 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3135 }
3136 /* now that discard is done we can proceed with any sync */
3137 clear_bit(STRIPE_DISCARD, &sh->state);
d47648fc
SL		 3138 /*
3139 * SCSI discard will change some bio fields and the stripe has
3140 * no updated data, so remove it from hash list and the stripe
3141 * will be reinitialized
3142 */
3143 spin_lock_irq(&conf->device_lock);
3144 remove_hash(sh);
3145 spin_unlock_irq(&conf->device_lock);
f8dfcffd
N		 3146 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3147 set_bit(STRIPE_HANDLE, &sh->state);
3148
3149 }
8b3e6cdc
DW		 3150
3151 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3152 if (atomic_dec_and_test(&conf->pending_full_writes))
3153 md_wakeup_thread(conf->mddev->thread);
a4456856
DW		 3154}
3155
d1688a6d 3156static void handle_stripe_dirtying(struct r5conf *conf,
c8ac1803
N		 3157 struct stripe_head *sh,
3158 struct stripe_head_state *s,
3159 int disks)
a4456856
DW		 3160{
3161 int rmw = 0, rcw = 0, i;
a7854487
AL		 3162 sector_t recovery_cp = conf->mddev->recovery_cp;
3163
3164 /* RAID6 requires 'rcw' in current implementation.
3165 * Otherwise, check whether resync is now happening or should start.
3166 * If yes, then the array is dirty (after unclean shutdown or
3167 * initial creation), so parity in some stripes might be inconsistent.
3168 * In this case, we need to always do reconstruct-write, to ensure
3169 * that in case of drive failure or read-error correction, we
3170 * generate correct data from the parity.
3171 */
3172 if (conf->max_degraded == 2 ||
3173 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
3174 /* Calculate the real rcw later - for now make it
c8ac1803
N		 3175 * look like rcw is cheaper
3176 */
3177 rcw = 1; rmw = 2;
a7854487
AL		 3178 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
3179 conf->max_degraded, (unsigned long long)recovery_cp,
3180 (unsigned long long)sh->sector);
c8ac1803 3181 } else for (i = disks; i--; ) {
a4456856
DW		 3182 /* would I have to read this buffer for read_modify_write */
3183 struct r5dev *dev = &sh->dev[i];
3184 if ((dev->towrite || i == sh->pd_idx) &&
3185 !test_bit(R5_LOCKED, &dev->flags) &&
f38e1219
DW		 3186 !(test_bit(R5_UPTODATE, &dev->flags) ||
3187 test_bit(R5_Wantcompute, &dev->flags))) {
a4456856
DW		 3188 if (test_bit(R5_Insync, &dev->flags))
3189 rmw++;
3190 else
3191 rmw += 2*disks; /* cannot read it */
3192 }
3193 /* Would I have to read this buffer for reconstruct_write */
3194 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
3195 !test_bit(R5_LOCKED, &dev->flags) &&
f38e1219
DW		 3196 !(test_bit(R5_UPTODATE, &dev->flags) ||
3197 test_bit(R5_Wantcompute, &dev->flags))) {
67f45548
N		 3198 if (test_bit(R5_Insync, &dev->flags))
3199 rcw++;
a4456856
DW		 3200 else
3201 rcw += 2*disks;
3202 }
3203 }
45b4233c 3204 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
a4456856
DW		 3205 (unsigned long long)sh->sector, rmw, rcw);
3206 set_bit(STRIPE_HANDLE, &sh->state);
a9add5d9 3207 if (rmw < rcw && rmw > 0) {
a4456856 3208 /* prefer read-modify-write, but need to get some data */
e3620a3a
JB		 3209 if (conf->mddev->queue)
3210 blk_add_trace_msg(conf->mddev->queue,
3211 "raid5 rmw %llu %d",
3212 (unsigned long long)sh->sector, rmw);
a4456856
DW		 3213 for (i = disks; i--; ) {
3214 struct r5dev *dev = &sh->dev[i];
3215 if ((dev->towrite || i == sh->pd_idx) &&
3216 !test_bit(R5_LOCKED, &dev->flags) &&
f38e1219
DW		 3217 !(test_bit(R5_UPTODATE, &dev->flags) ||
3218 test_bit(R5_Wantcompute, &dev->flags)) &&
a4456856 3219 test_bit(R5_Insync, &dev->flags)) {
67f45548
N		 3220 if (test_bit(STRIPE_PREREAD_ACTIVE,
3221 &sh->state)) {
3222 pr_debug("Read_old block %d for r-m-w\n",
3223 i);
a4456856
DW		 3224 set_bit(R5_LOCKED, &dev->flags);
3225 set_bit(R5_Wantread, &dev->flags);
3226 s->locked++;
3227 } else {
3228 set_bit(STRIPE_DELAYED, &sh->state);
3229 set_bit(STRIPE_HANDLE, &sh->state);
3230 }
3231 }
3232 }
a9add5d9 3233 }
c8ac1803 3234 if (rcw <= rmw && rcw > 0) {
a4456856 3235 /* want reconstruct write, but need to get some data */
a9add5d9 3236 int qread =0;
c8ac1803 3237 rcw = 0;
a4456856
DW		 3238 for (i = disks; i--; ) {
3239 struct r5dev *dev = &sh->dev[i];
3240 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
c8ac1803 3241 i != sh->pd_idx && i != sh->qd_idx &&
a4456856 3242 !test_bit(R5_LOCKED, &dev->flags) &&
f38e1219 3243 !(test_bit(R5_UPTODATE, &dev->flags) ||
c8ac1803
N		 3244 test_bit(R5_Wantcompute, &dev->flags))) {
3245 rcw++;
67f45548
N		 3246 if (test_bit(R5_Insync, &dev->flags) &&
3247 test_bit(STRIPE_PREREAD_ACTIVE,
3248 &sh->state)) {
45b4233c 3249 pr_debug("Read_old block "
a4456856
DW		 3250 "%d for Reconstruct\n", i);
3251 set_bit(R5_LOCKED, &dev->flags);
3252 set_bit(R5_Wantread, &dev->flags);
3253 s->locked++;
a9add5d9 3254 qread++;
a4456856
DW		 3255 } else {
3256 set_bit(STRIPE_DELAYED, &sh->state);
3257 set_bit(STRIPE_HANDLE, &sh->state);
3258 }
3259 }
3260 }
e3620a3a 3261 if (rcw && conf->mddev->queue)
a9add5d9
N		 3262 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3263 (unsigned long long)sh->sector,
3264 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
c8ac1803 3265 }
b1b02fe9
N		 3266
3267 if (rcw > disks && rmw > disks &&
3268 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3269 set_bit(STRIPE_DELAYED, &sh->state);
3270
a4456856
DW		 3271 /* now if nothing is locked, and if we have enough data,
3272 * we can start a write request
3273 */
f38e1219
DW		 3274 /* since handle_stripe can be called at any time we need to handle the
3275 * case where a compute block operation has been submitted and then a
ac6b53b6
DW		 3276 * subsequent call wants to start a write request. raid_run_ops only
3277 * handles the case where compute block and reconstruct are requested
f38e1219
DW		 3278 * simultaneously. If this is not the case then new writes need to be
3279 * held off until the compute completes.
3280 */
976ea8d4
DW		 3281 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3282 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3283 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
c0f7bddb 3284 schedule_reconstruction(sh, s, rcw == 0, 0);
a4456856
DW		 3285}
3286
d1688a6d 3287static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
a4456856
DW		 3288 struct stripe_head_state *s, int disks)
3289{
ecc65c9b 3290 struct r5dev *dev = NULL;
bd2ab670 3291
a4456856 3292 set_bit(STRIPE_HANDLE, &sh->state);
e89f8962 3293
ecc65c9b
DW		 3294 switch (sh->check_state) {
3295 case check_state_idle:
3296 /* start a new check operation if there are no failures */
bd2ab670 3297 if (s->failed == 0) {
bd2ab670 3298 BUG_ON(s->uptodate != disks);
ecc65c9b
DW		 3299 sh->check_state = check_state_run;
3300 set_bit(STRIPE_OP_CHECK, &s->ops_request);
bd2ab670 3301 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
bd2ab670 3302 s->uptodate--;
ecc65c9b 3303 break;
bd2ab670 3304 }
f2b3b44d 3305 dev = &sh->dev[s->failed_num[0]];
ecc65c9b
DW		 3306 /* fall through */
3307 case check_state_compute_result:
3308 sh->check_state = check_state_idle;
3309 if (!dev)
3310 dev = &sh->dev[sh->pd_idx];
3311
3312 /* check that a write has not made the stripe insync */
3313 if (test_bit(STRIPE_INSYNC, &sh->state))
3314 break;
c8894419 3315
a4456856 3316 /* either failed parity check, or recovery is happening */
a4456856
DW		 3317 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3318 BUG_ON(s->uptodate != disks);
3319
3320 set_bit(R5_LOCKED, &dev->flags);
ecc65c9b 3321 s->locked++;
a4456856 3322 set_bit(R5_Wantwrite, &dev->flags);
830ea016 3323
a4456856 3324 clear_bit(STRIPE_DEGRADED, &sh->state);
a4456856 3325 set_bit(STRIPE_INSYNC, &sh->state);
ecc65c9b
DW		 3326 break;
3327 case check_state_run:
3328 break; /* we will be called again upon completion */
3329 case check_state_check_result:
3330 sh->check_state = check_state_idle;
3331
3332 /* if a failure occurred during the check operation, leave
3333 * STRIPE_INSYNC not set and let the stripe be handled again
3334 */
3335 if (s->failed)
3336 break;
3337
3338 /* handle a successful check operation, if parity is correct
3339 * we are done. Otherwise update the mismatch count and repair
3340 * parity if !MD_RECOVERY_CHECK
3341 */
ad283ea4 3342 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
ecc65c9b
DW		 3343 /* parity is correct (on disc,
3344 * not in buffer any more)
3345 */
3346 set_bit(STRIPE_INSYNC, &sh->state);
3347 else {
7f7583d4 3348 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
ecc65c9b
DW		 3349 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3350 /* don't try to repair!! */
3351 set_bit(STRIPE_INSYNC, &sh->state);
3352 else {
3353 sh->check_state = check_state_compute_run;
976ea8d4 3354 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
ecc65c9b
DW		 3355 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3356 set_bit(R5_Wantcompute,
3357 &sh->dev[sh->pd_idx].flags);
3358 sh->ops.target = sh->pd_idx;
ac6b53b6 3359 sh->ops.target2 = -1;
ecc65c9b
DW		 3360 s->uptodate++;
3361 }
3362 }
3363 break;
3364 case check_state_compute_run:
3365 break;
3366 default:
3367 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3368 __func__, sh->check_state,
3369 (unsigned long long) sh->sector);
3370 BUG();
a4456856
DW		 3371 }
3372}
3373
d1688a6d 3374static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
36d1c647 3375 struct stripe_head_state *s,
f2b3b44d 3376 int disks)
a4456856 3377{
a4456856 3378 int pd_idx = sh->pd_idx;
34e04e87 3379 int qd_idx = sh->qd_idx;
d82dfee0 3380 struct r5dev *dev;
a4456856
DW		 3381
3382 set_bit(STRIPE_HANDLE, &sh->state);
3383
3384 BUG_ON(s->failed > 2);
d82dfee0 3385
a4456856
DW		 3386 /* Want to check and possibly repair P and Q.
3387 * However there could be one 'failed' device, in which
3388 * case we can only check one of them, possibly using the
3389 * other to generate missing data
3390 */
3391
d82dfee0
DW		 3392 switch (sh->check_state) {
3393 case check_state_idle:
3394 /* start a new check operation if there are < 2 failures */
f2b3b44d 3395 if (s->failed == s->q_failed) {
d82dfee0 3396 /* The only possible failed device holds Q, so it
a4456856
DW		 3397 * makes sense to check P (If anything else were failed,
3398 * we would have used P to recreate it).
3399 */
d82dfee0 3400 sh->check_state = check_state_run;
a4456856 3401 }
f2b3b44d 3402 if (!s->q_failed && s->failed < 2) {
d82dfee0 3403 /* Q is not failed, and we didn't use it to generate
a4456856
DW		 3404 * anything, so it makes sense to check it
3405 */
d82dfee0
DW		 3406 if (sh->check_state == check_state_run)
3407 sh->check_state = check_state_run_pq;
3408 else
3409 sh->check_state = check_state_run_q;
a4456856 3410 }
a4456856 3411
d82dfee0
DW		 3412 /* discard potentially stale zero_sum_result */
3413 sh->ops.zero_sum_result = 0;
a4456856 3414
d82dfee0
DW		 3415 if (sh->check_state == check_state_run) {
3416 /* async_xor_zero_sum destroys the contents of P */
3417 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3418 s->uptodate--;
a4456856 3419 }
d82dfee0
DW		 3420 if (sh->check_state >= check_state_run &&
3421 sh->check_state <= check_state_run_pq) {
3422 /* async_syndrome_zero_sum preserves P and Q, so
3423 * no need to mark them !uptodate here
3424 */
3425 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3426 break;
a4456856
DW		 3427 }
3428
d82dfee0
DW		 3429 /* we have 2-disk failure */
3430 BUG_ON(s->failed != 2);
3431 /* fall through */
3432 case check_state_compute_result:
3433 sh->check_state = check_state_idle;
a4456856 3434
d82dfee0
DW		 3435 /* check that a write has not made the stripe insync */
3436 if (test_bit(STRIPE_INSYNC, &sh->state))
3437 break;
a4456856
DW		 3438
3439 /* now write out any block on a failed drive,
d82dfee0 3440 * or P or Q if they were recomputed
a4456856 3441 */
d82dfee0 3442 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
a4456856 3443 if (s->failed == 2) {
f2b3b44d 3444 dev = &sh->dev[s->failed_num[1]];
a4456856
DW		 3445 s->locked++;
3446 set_bit(R5_LOCKED, &dev->flags);
3447 set_bit(R5_Wantwrite, &dev->flags);
3448 }
3449 if (s->failed >= 1) {
f2b3b44d 3450 dev = &sh->dev[s->failed_num[0]];
a4456856
DW		 3451 s->locked++;
3452 set_bit(R5_LOCKED, &dev->flags);
3453 set_bit(R5_Wantwrite, &dev->flags);
3454 }
d82dfee0 3455 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
a4456856
DW		 3456 dev = &sh->dev[pd_idx];
3457 s->locked++;
3458 set_bit(R5_LOCKED, &dev->flags);
3459 set_bit(R5_Wantwrite, &dev->flags);
3460 }
d82dfee0 3461 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
a4456856
DW		 3462 dev = &sh->dev[qd_idx];
3463 s->locked++;
3464 set_bit(R5_LOCKED, &dev->flags);
3465 set_bit(R5_Wantwrite, &dev->flags);
3466 }
3467 clear_bit(STRIPE_DEGRADED, &sh->state);
3468
3469 set_bit(STRIPE_INSYNC, &sh->state);
d82dfee0
DW		 3470 break;
3471 case check_state_run:
3472 case check_state_run_q:
3473 case check_state_run_pq:
3474 break; /* we will be called again upon completion */
3475 case check_state_check_result:
3476 sh->check_state = check_state_idle;
3477
3478 /* handle a successful check operation, if parity is correct
3479 * we are done. Otherwise update the mismatch count and repair
3480 * parity if !MD_RECOVERY_CHECK
3481 */
3482 if (sh->ops.zero_sum_result == 0) {
3483 /* both parities are correct */
3484 if (!s->failed)
3485 set_bit(STRIPE_INSYNC, &sh->state);
3486 else {
3487 /* in contrast to the raid5 case we can validate
3488 * parity, but still have a failure to write
3489 * back
3490 */
3491 sh->check_state = check_state_compute_result;
3492 /* Returning at this point means that we may go
3493 * off and bring p and/or q uptodate again so
3494 * we make sure to check zero_sum_result again
3495 * to verify if p or q need writeback
3496 */
3497 }
3498 } else {
7f7583d4 3499 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
d82dfee0
DW		 3500 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3501 /* don't try to repair!! */
3502 set_bit(STRIPE_INSYNC, &sh->state);
3503 else {
3504 int *target = &sh->ops.target;
3505
3506 sh->ops.target = -1;
3507 sh->ops.target2 = -1;
3508 sh->check_state = check_state_compute_run;
3509 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3510 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3511 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3512 set_bit(R5_Wantcompute,
3513 &sh->dev[pd_idx].flags);
3514 *target = pd_idx;
3515 target = &sh->ops.target2;
3516 s->uptodate++;
3517 }
3518 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3519 set_bit(R5_Wantcompute,
3520 &sh->dev[qd_idx].flags);
3521 *target = qd_idx;
3522 s->uptodate++;
3523 }
3524 }
3525 }
3526 break;
3527 case check_state_compute_run:
3528 break;
3529 default:
3530 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3531 __func__, sh->check_state,
3532 (unsigned long long) sh->sector);
3533 BUG();
a4456856
DW		 3534 }
3535}
3536
d1688a6d 3537static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
a4456856
DW		 3538{
3539 int i;
3540
3541 /* We have read all the blocks in this stripe and now we need to
3542 * copy some of them into a target stripe for expand.
3543 */
f0a50d37 3544 struct dma_async_tx_descriptor *tx = NULL;
a4456856
DW		 3545 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3546 for (i = 0; i < sh->disks; i++)
34e04e87 3547 if (i != sh->pd_idx && i != sh->qd_idx) {
911d4ee8 3548 int dd_idx, j;
a4456856 3549 struct stripe_head *sh2;
a08abd8c 3550 struct async_submit_ctl submit;
a4456856 3551
784052ec 3552 sector_t bn = compute_blocknr(sh, i, 1);
911d4ee8
N		 3553 sector_t s = raid5_compute_sector(conf, bn, 0,
3554 &dd_idx, NULL);
a8c906ca 3555 sh2 = get_active_stripe(conf, s, 0, 1, 1);
a4456856
DW		 3556 if (sh2 == NULL)
3557 /* so far only the early blocks of this stripe
3558 * have been requested. When later blocks
3559 * get requested, we will try again
3560 */
3561 continue;
3562 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3563 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3564 /* must have already done this block */
3565 release_stripe(sh2);
3566 continue;
3567 }
f0a50d37
DW		 3568
3569 /* place all the copies on one channel */
a08abd8c 3570 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
f0a50d37 3571 tx = async_memcpy(sh2->dev[dd_idx].page,
88ba2aa5 3572 sh->dev[i].page, 0, 0, STRIPE_SIZE,
a08abd8c 3573 &submit);
f0a50d37 3574
a4456856
DW		 3575 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3576 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3577 for (j = 0; j < conf->raid_disks; j++)
3578 if (j != sh2->pd_idx &&
86c374ba 3579 j != sh2->qd_idx &&
a4456856
DW		 3580 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3581 break;
3582 if (j == conf->raid_disks) {
3583 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3584 set_bit(STRIPE_HANDLE, &sh2->state);
3585 }
3586 release_stripe(sh2);
f0a50d37 3587
a4456856 3588 }
a2e08551 3589 /* done submitting copies, wait for them to complete */
749586b7 3590 async_tx_quiesce(&tx);
a4456856 3591}
1da177e4
LT		 3592
3593/*
3594 * handle_stripe - do things to a stripe.
3595 *
9a3e1101
N		 3596 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3597 * state of various bits to see what needs to be done.
1da177e4 3598 * Possible results:
9a3e1101
N		 3599 * return some read requests which now have data
3600 * return some write requests which are safely on storage
1da177e4
LT		 3601 * schedule a read on some buffers
3602 * schedule a write of some buffers
3603 * return confirmation of parity correctness
3604 *
1da177e4 3605 */
a4456856 3606
acfe726b 3607static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
1da177e4 3608{
d1688a6d 3609 struct r5conf *conf = sh->raid_conf;
f416885e 3610 int disks = sh->disks;
474af965
N		 3611 struct r5dev *dev;
3612 int i;
9a3e1101 3613 int do_recovery = 0;
1da177e4 3614
acfe726b
N		 3615 memset(s, 0, sizeof(*s));
3616
acfe726b
N		 3617 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3618 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3619 s->failed_num[0] = -1;
3620 s->failed_num[1] = -1;
1da177e4 3621
acfe726b 3622 /* Now to look around and see what can be done */
1da177e4 3623 rcu_read_lock();
16a53ecc 3624 for (i=disks; i--; ) {
3cb03002 3625 struct md_rdev *rdev;
31c176ec
N		 3626 sector_t first_bad;
3627 int bad_sectors;
3628 int is_bad = 0;
acfe726b 3629
16a53ecc 3630 dev = &sh->dev[i];
1da177e4 3631
45b4233c 3632 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
9a3e1101
N		 3633 i, dev->flags,
3634 dev->toread, dev->towrite, dev->written);
6c0069c0
YT		 3635 /* maybe we can reply to a read
3636 *
3637 * new wantfill requests are only permitted while
3638 * ops_complete_biofill is guaranteed to be inactive
3639 */
3640 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3641 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3642 set_bit(R5_Wantfill, &dev->flags);
1da177e4 3643
16a53ecc 3644 /* now count some things */
cc94015a
N		 3645 if (test_bit(R5_LOCKED, &dev->flags))
3646 s->locked++;
3647 if (test_bit(R5_UPTODATE, &dev->flags))
3648 s->uptodate++;
2d6e4ecc 3649 if (test_bit(R5_Wantcompute, &dev->flags)) {
cc94015a
N		 3650 s->compute++;
3651 BUG_ON(s->compute > 2);
2d6e4ecc 3652 }
1da177e4 3653
acfe726b 3654 if (test_bit(R5_Wantfill, &dev->flags))
cc94015a 3655 s->to_fill++;
acfe726b 3656 else if (dev->toread)
cc94015a 3657 s->to_read++;
16a53ecc 3658 if (dev->towrite) {
cc94015a 3659 s->to_write++;
16a53ecc 3660 if (!test_bit(R5_OVERWRITE, &dev->flags))
cc94015a 3661 s->non_overwrite++;
16a53ecc 3662 }
a4456856 3663 if (dev->written)
cc94015a 3664 s->written++;
14a75d3e
N		 3665 /* Prefer to use the replacement for reads, but only
3666 * if it is recovered enough and has no bad blocks.
3667 */
3668 rdev = rcu_dereference(conf->disks[i].replacement);
3669 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3670 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3671 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3672 &first_bad, &bad_sectors))
3673 set_bit(R5_ReadRepl, &dev->flags);
3674 else {
9a3e1101
N		 3675 if (rdev)
3676 set_bit(R5_NeedReplace, &dev->flags);
14a75d3e
N		 3677 rdev = rcu_dereference(conf->disks[i].rdev);
3678 clear_bit(R5_ReadRepl, &dev->flags);
3679 }
9283d8c5
N		 3680 if (rdev && test_bit(Faulty, &rdev->flags))
3681 rdev = NULL;
31c176ec
N		 3682 if (rdev) {
3683 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3684 &first_bad, &bad_sectors);
3685 if (s->blocked_rdev == NULL
3686 && (test_bit(Blocked, &rdev->flags)
3687 || is_bad < 0)) {
3688 if (is_bad < 0)
3689 set_bit(BlockedBadBlocks,
3690 &rdev->flags);
3691 s->blocked_rdev = rdev;
3692 atomic_inc(&rdev->nr_pending);
3693 }
6bfe0b49 3694 }
415e72d0
N		 3695 clear_bit(R5_Insync, &dev->flags);
3696 if (!rdev)
3697 /* Not in-sync */;
31c176ec
N		 3698 else if (is_bad) {
3699 /* also not in-sync */
18b9837e
N		 3700 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3701 test_bit(R5_UPTODATE, &dev->flags)) {
31c176ec
N		 3702 /* treat as in-sync, but with a read error
3703 * which we can now try to correct
3704 */
3705 set_bit(R5_Insync, &dev->flags);
3706 set_bit(R5_ReadError, &dev->flags);
3707 }
3708 } else if (test_bit(In_sync, &rdev->flags))
415e72d0 3709 set_bit(R5_Insync, &dev->flags);
30d7a483 3710 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
415e72d0 3711 /* in sync if before recovery_offset */
30d7a483
N		 3712 set_bit(R5_Insync, &dev->flags);
3713 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3714 test_bit(R5_Expanded, &dev->flags))
3715 /* If we've reshaped into here, we assume it is Insync.
3716 * We will shortly update recovery_offset to make
3717 * it official.
3718 */
3719 set_bit(R5_Insync, &dev->flags);
3720
1cc03eb9 3721 if (test_bit(R5_WriteError, &dev->flags)) {
14a75d3e
N		 3722 /* This flag does not apply to '.replacement'
3723 * only to .rdev, so make sure to check that*/
3724 struct md_rdev *rdev2 = rcu_dereference(
3725 conf->disks[i].rdev);
3726 if (rdev2 == rdev)
3727 clear_bit(R5_Insync, &dev->flags);
3728 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
bc2607f3 3729 s->handle_bad_blocks = 1;
14a75d3e 3730 atomic_inc(&rdev2->nr_pending);
bc2607f3
N		 3731 } else
3732 clear_bit(R5_WriteError, &dev->flags);
3733 }
1cc03eb9 3734 if (test_bit(R5_MadeGood, &dev->flags)) {
14a75d3e
N		 3735 /* This flag does not apply to '.replacement'
3736 * only to .rdev, so make sure to check that*/
3737 struct md_rdev *rdev2 = rcu_dereference(
3738 conf->disks[i].rdev);
3739 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
b84db560 3740 s->handle_bad_blocks = 1;
14a75d3e 3741 atomic_inc(&rdev2->nr_pending);
b84db560
N		 3742 } else
3743 clear_bit(R5_MadeGood, &dev->flags);
3744 }
977df362
N		 3745 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3746 struct md_rdev *rdev2 = rcu_dereference(
3747 conf->disks[i].replacement);
3748 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3749 s->handle_bad_blocks = 1;
3750 atomic_inc(&rdev2->nr_pending);
3751 } else
3752 clear_bit(R5_MadeGoodRepl, &dev->flags);
3753 }
415e72d0 3754 if (!test_bit(R5_Insync, &dev->flags)) {
16a53ecc
N		 3755 /* The ReadError flag will just be confusing now */
3756 clear_bit(R5_ReadError, &dev->flags);
3757 clear_bit(R5_ReWrite, &dev->flags);
1da177e4 3758 }
415e72d0
N		 3759 if (test_bit(R5_ReadError, &dev->flags))
3760 clear_bit(R5_Insync, &dev->flags);
3761 if (!test_bit(R5_Insync, &dev->flags)) {
cc94015a
N		 3762 if (s->failed < 2)
3763 s->failed_num[s->failed] = i;
3764 s->failed++;
9a3e1101
N		 3765 if (rdev && !test_bit(Faulty, &rdev->flags))
3766 do_recovery = 1;
415e72d0 3767 }
1da177e4 3768 }
9a3e1101
N		 3769 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3770 /* If there is a failed device being replaced,
3771 * we must be recovering.
3772 * else if we are after recovery_cp, we must be syncing
c6d2e084 3773 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
9a3e1101
N		 3774 * else we can only be replacing
3775 * sync and recovery both need to read all devices, and so
3776 * use the same flag.
3777 */
3778 if (do_recovery ||
c6d2e084 3779 sh->sector >= conf->mddev->recovery_cp ||
3780 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
9a3e1101
N		 3781 s->syncing = 1;
3782 else
3783 s->replacing = 1;
3784 }
1da177e4 3785 rcu_read_unlock();
cc94015a
N		 3786}
3787
3788static void handle_stripe(struct stripe_head *sh)
3789{
3790 struct stripe_head_state s;
d1688a6d 3791 struct r5conf *conf = sh->raid_conf;
3687c061 3792 int i;
84789554
N		 3793 int prexor;
3794 int disks = sh->disks;
474af965 3795 struct r5dev *pdev, *qdev;
cc94015a
N		 3796
3797 clear_bit(STRIPE_HANDLE, &sh->state);
257a4b42 3798 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
cc94015a
N		 3799 /* already being handled, ensure it gets handled
3800 * again when current action finishes */
3801 set_bit(STRIPE_HANDLE, &sh->state);
3802 return;
3803 }
3804
f8dfcffd
N		 3805 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3806 spin_lock(&sh->stripe_lock);
3807 /* Cannot process 'sync' concurrently with 'discard' */
3808 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3809 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3810 set_bit(STRIPE_SYNCING, &sh->state);
3811 clear_bit(STRIPE_INSYNC, &sh->state);
f94c0b66 3812 clear_bit(STRIPE_REPLACED, &sh->state);
f8dfcffd
N		 3813 }
3814 spin_unlock(&sh->stripe_lock);
cc94015a
N		 3815 }
3816 clear_bit(STRIPE_DELAYED, &sh->state);
3817
3818 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3819 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3820 (unsigned long long)sh->sector, sh->state,
3821 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3822 sh->check_state, sh->reconstruct_state);
3687c061 3823
acfe726b 3824 analyse_stripe(sh, &s);
c5a31000 3825
bc2607f3
N		 3826 if (s.handle_bad_blocks) {
3827 set_bit(STRIPE_HANDLE, &sh->state);
3828 goto finish;
3829 }
3830
474af965
N		 3831 if (unlikely(s.blocked_rdev)) {
3832 if (s.syncing || s.expanding || s.expanded ||
9a3e1101 3833 s.replacing || s.to_write || s.written) {
474af965
N		 3834 set_bit(STRIPE_HANDLE, &sh->state);
3835 goto finish;
3836 }
3837 /* There is nothing for the blocked_rdev to block */
3838 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3839 s.blocked_rdev = NULL;
3840 }
3841
3842 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3843 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3844 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3845 }
3846
3847 pr_debug("locked=%d uptodate=%d to_read=%d"
3848 " to_write=%d failed=%d failed_num=%d,%d\n",
3849 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3850 s.failed_num[0], s.failed_num[1]);
3851 /* check if the array has lost more than max_degraded devices and,
3852 * if so, some requests might need to be failed.
3853 */
9a3f530f
N		 3854 if (s.failed > conf->max_degraded) {
3855 sh->check_state = 0;
3856 sh->reconstruct_state = 0;
3857 if (s.to_read+s.to_write+s.written)
3858 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
9a3e1101 3859 if (s.syncing + s.replacing)
9a3f530f
N		 3860 handle_failed_sync(conf, sh, &s);
3861 }
474af965 3862
84789554
N		 3863 /* Now we check to see if any write operations have recently
3864 * completed
3865 */
3866 prexor = 0;
3867 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3868 prexor = 1;
3869 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3870 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3871 sh->reconstruct_state = reconstruct_state_idle;
3872
3873 /* All the 'written' buffers and the parity block are ready to
3874 * be written back to disk
3875 */
9e444768
SL		 3876 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3877 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
84789554 3878 BUG_ON(sh->qd_idx >= 0 &&
9e444768
SL		 3879 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3880 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
84789554
N		 3881 for (i = disks; i--; ) {
3882 struct r5dev *dev = &sh->dev[i];
3883 if (test_bit(R5_LOCKED, &dev->flags) &&
3884 (i == sh->pd_idx || i == sh->qd_idx ||
3885 dev->written)) {
3886 pr_debug("Writing block %d\n", i);
3887 set_bit(R5_Wantwrite, &dev->flags);
3888 if (prexor)
3889 continue;
9c4bdf69
N		 3890 if (s.failed > 1)
3891 continue;
84789554
N		 3892 if (!test_bit(R5_Insync, &dev->flags) ||
3893 ((i == sh->pd_idx || i == sh->qd_idx) &&
3894 s.failed == 0))
3895 set_bit(STRIPE_INSYNC, &sh->state);
3896 }
3897 }
3898 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3899 s.dec_preread_active = 1;
3900 }
3901
ef5b7c69
N		 3902 /*
3903 * might be able to return some write requests if the parity blocks
3904 * are safe, or on a failed drive
3905 */
3906 pdev = &sh->dev[sh->pd_idx];
3907 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3908 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3909 qdev = &sh->dev[sh->qd_idx];
3910 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3911 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3912 || conf->level < 6;
3913
3914 if (s.written &&
3915 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3916 && !test_bit(R5_LOCKED, &pdev->flags)
3917 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3918 test_bit(R5_Discard, &pdev->flags))))) &&
3919 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3920 && !test_bit(R5_LOCKED, &qdev->flags)
3921 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3922 test_bit(R5_Discard, &qdev->flags))))))
3923 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3924
3925 /* Now we might consider reading some blocks, either to check/generate
3926 * parity, or to satisfy requests
3927 * or to load a block that is being partially written.
3928 */
3929 if (s.to_read || s.non_overwrite
3930 || (conf->level == 6 && s.to_write && s.failed)
3931 || (s.syncing && (s.uptodate + s.compute < disks))
3932 || s.replacing
3933 || s.expanding)
3934 handle_stripe_fill(sh, &s, disks);
3935
84789554
N		 3936 /* Now to consider new write requests and what else, if anything
3937 * should be read. We do not handle new writes when:
3938 * 1/ A 'write' operation (copy+xor) is already in flight.
3939 * 2/ A 'check' operation is in flight, as it may clobber the parity
3940 * block.
3941 */
3942 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3943 handle_stripe_dirtying(conf, sh, &s, disks);
3944
3945 /* maybe we need to check and possibly fix the parity for this stripe
3946 * Any reads will already have been scheduled, so we just see if enough
3947 * data is available. The parity check is held off while parity
3948 * dependent operations are in flight.
3949 */
3950 if (sh->check_state ||
3951 (s.syncing && s.locked == 0 &&
3952 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3953 !test_bit(STRIPE_INSYNC, &sh->state))) {
3954 if (conf->level == 6)
3955 handle_parity_checks6(conf, sh, &s, disks);
3956 else
3957 handle_parity_checks5(conf, sh, &s, disks);
3958 }
c5a31000 3959
f94c0b66
N		 3960 if ((s.replacing || s.syncing) && s.locked == 0
3961 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3962 && !test_bit(STRIPE_REPLACED, &sh->state)) {
9a3e1101
N		 3963 /* Write out to replacement devices where possible */
3964 for (i = 0; i < conf->raid_disks; i++)
f94c0b66
N		 3965 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3966 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
9a3e1101
N		 3967 set_bit(R5_WantReplace, &sh->dev[i].flags);
3968 set_bit(R5_LOCKED, &sh->dev[i].flags);
3969 s.locked++;
3970 }
f94c0b66
N		 3971 if (s.replacing)
3972 set_bit(STRIPE_INSYNC, &sh->state);
3973 set_bit(STRIPE_REPLACED, &sh->state);
9a3e1101
N		 3974 }
3975 if ((s.syncing || s.replacing) && s.locked == 0 &&
f94c0b66 3976 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
9a3e1101 3977 test_bit(STRIPE_INSYNC, &sh->state)) {
c5a31000
N		 3978 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3979 clear_bit(STRIPE_SYNCING, &sh->state);
f8dfcffd
N		 3980 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3981 wake_up(&conf->wait_for_overlap);
c5a31000
N		 3982 }
3983
3984 /* If the failed drives are just a ReadError, then we might need
3985 * to progress the repair/check process
3986 */
3987 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3988 for (i = 0; i < s.failed; i++) {
3989 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3990 if (test_bit(R5_ReadError, &dev->flags)
3991 && !test_bit(R5_LOCKED, &dev->flags)
3992 && test_bit(R5_UPTODATE, &dev->flags)
3993 ) {
3994 if (!test_bit(R5_ReWrite, &dev->flags)) {
3995 set_bit(R5_Wantwrite, &dev->flags);
3996 set_bit(R5_ReWrite, &dev->flags);
3997 set_bit(R5_LOCKED, &dev->flags);
3998 s.locked++;
3999 } else {
4000 /* let's read it back */
4001 set_bit(R5_Wantread, &dev->flags);
4002 set_bit(R5_LOCKED, &dev->flags);
4003 s.locked++;
4004 }
4005 }
4006 }
4007
3687c061
N		 4008 /* Finish reconstruct operations initiated by the expansion process */
4009 if (sh->reconstruct_state == reconstruct_state_result) {
4010 struct stripe_head *sh_src
4011 = get_active_stripe(conf, sh->sector, 1, 1, 1);
4012 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4013 /* sh cannot be written until sh_src has been read.
4014 * so arrange for sh to be delayed a little
4015 */
4016 set_bit(STRIPE_DELAYED, &sh->state);
4017 set_bit(STRIPE_HANDLE, &sh->state);
4018 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4019 &sh_src->state))
4020 atomic_inc(&conf->preread_active_stripes);
4021 release_stripe(sh_src);
4022 goto finish;
4023 }
4024 if (sh_src)
4025 release_stripe(sh_src);
4026
4027 sh->reconstruct_state = reconstruct_state_idle;
4028 clear_bit(STRIPE_EXPANDING, &sh->state);
4029 for (i = conf->raid_disks; i--; ) {
4030 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4031 set_bit(R5_LOCKED, &sh->dev[i].flags);
4032 s.locked++;
4033 }
4034 }
f416885e 4035
3687c061
N		 4036 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4037 !sh->reconstruct_state) {
4038 /* Need to write out all blocks after computing parity */
4039 sh->disks = conf->raid_disks;
4040 stripe_set_idx(sh->sector, conf, 0, sh);
4041 schedule_reconstruction(sh, &s, 1, 1);
4042 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4043 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4044 atomic_dec(&conf->reshape_stripes);
4045 wake_up(&conf->wait_for_overlap);
4046 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4047 }
4048
4049 if (s.expanding && s.locked == 0 &&
4050 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4051 handle_stripe_expansion(conf, sh);
16a53ecc 4052
3687c061 4053finish:
6bfe0b49 4054 /* wait for this device to become unblocked */
5f066c63
N		 4055 if (unlikely(s.blocked_rdev)) {
4056 if (conf->mddev->external)
4057 md_wait_for_blocked_rdev(s.blocked_rdev,
4058 conf->mddev);
4059 else
4060 /* Internal metadata will immediately
4061 * be written by raid5d, so we don't
4062 * need to wait here.
4063 */
4064 rdev_dec_pending(s.blocked_rdev,
4065 conf->mddev);
4066 }
6bfe0b49 4067
bc2607f3
N		 4068 if (s.handle_bad_blocks)
4069 for (i = disks; i--; ) {
3cb03002 4070 struct md_rdev *rdev;
bc2607f3
N		 4071 struct r5dev *dev = &sh->dev[i];
4072 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4073 /* We own a safe reference to the rdev */
4074 rdev = conf->disks[i].rdev;
4075 if (!rdev_set_badblocks(rdev, sh->sector,
4076 STRIPE_SECTORS, 0))
4077 md_error(conf->mddev, rdev);
4078 rdev_dec_pending(rdev, conf->mddev);
4079 }
b84db560
N		 4080 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4081 rdev = conf->disks[i].rdev;
4082 rdev_clear_badblocks(rdev, sh->sector,
c6563a8c 4083 STRIPE_SECTORS, 0);
b84db560
N		 4084 rdev_dec_pending(rdev, conf->mddev);
4085 }
977df362
N		 4086 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4087 rdev = conf->disks[i].replacement;
dd054fce
N		 4088 if (!rdev)
4089 /* rdev have been moved down */
4090 rdev = conf->disks[i].rdev;
977df362 4091 rdev_clear_badblocks(rdev, sh->sector,
c6563a8c 4092 STRIPE_SECTORS, 0);
977df362
N		 4093 rdev_dec_pending(rdev, conf->mddev);
4094 }
bc2607f3
N		 4095 }
4096
6c0069c0
YT		 4097 if (s.ops_request)
4098 raid_run_ops(sh, s.ops_request);
4099
f0e43bcd 4100 ops_run_io(sh, &s);
16a53ecc 4101
c5709ef6 4102 if (s.dec_preread_active) {
729a1866 4103 /* We delay this until after ops_run_io so that if make_request
e9c7469b 4104 * is waiting on a flush, it won't continue until the writes
729a1866
N		 4105 * have actually been submitted.
4106 */
4107 atomic_dec(&conf->preread_active_stripes);
4108 if (atomic_read(&conf->preread_active_stripes) <
4109 IO_THRESHOLD)
4110 md_wakeup_thread(conf->mddev->thread);
4111 }
4112
c5709ef6 4113 return_io(s.return_bi);
16a53ecc 4114
257a4b42 4115 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
16a53ecc
N		 4116}
4117
d1688a6d 4118static void raid5_activate_delayed(struct r5conf *conf)
16a53ecc
N		 4119{
4120 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4121 while (!list_empty(&conf->delayed_list)) {
4122 struct list_head *l = conf->delayed_list.next;
4123 struct stripe_head *sh;
4124 sh = list_entry(l, struct stripe_head, lru);
4125 list_del_init(l);
4126 clear_bit(STRIPE_DELAYED, &sh->state);
4127 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4128 atomic_inc(&conf->preread_active_stripes);
8b3e6cdc 4129 list_add_tail(&sh->lru, &conf->hold_list);
851c30c9 4130 raid5_wakeup_stripe_thread(sh);
16a53ecc 4131 }
482c0834 4132 }
16a53ecc
N		 4133}
4134
566c09c5
SL		 4135static void activate_bit_delay(struct r5conf *conf,
4136 struct list_head *temp_inactive_list)
16a53ecc
N		 4137{
4138 /* device_lock is held */
4139 struct list_head head;
4140 list_add(&head, &conf->bitmap_list);
4141 list_del_init(&conf->bitmap_list);
4142 while (!list_empty(&head)) {
4143 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
566c09c5 4144 int hash;
16a53ecc
N		 4145 list_del_init(&sh->lru);
4146 atomic_inc(&sh->count);
566c09c5
SL		 4147 hash = sh->hash_lock_index;
4148 __release_stripe(conf, sh, &temp_inactive_list[hash]);
16a53ecc
N		 4149 }
4150}
4151
5c675f83 4152static int raid5_congested(struct mddev *mddev, int bits)
f022b2fd 4153{
d1688a6d 4154 struct r5conf *conf = mddev->private;
f022b2fd
N		 4155
4156 /* No difference between reads and writes. Just check
4157 * how busy the stripe_cache is
4158 */
3fa841d7 4159
f022b2fd
N		 4160 if (conf->inactive_blocked)
4161 return 1;
4162 if (conf->quiesce)
4163 return 1;
4bda556a 4164 if (atomic_read(&conf->empty_inactive_list_nr))
f022b2fd
N		 4165 return 1;
4166
4167 return 0;
4168}
4169
23032a0e
RBJ		 4170/* We want read requests to align with chunks where possible,
4171 * but write requests don't need to.
4172 */
64590f45 4173static int raid5_mergeable_bvec(struct mddev *mddev,
cc371e66
AK		 4174 struct bvec_merge_data *bvm,
4175 struct bio_vec *biovec)
23032a0e 4176{
cc371e66 4177 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
23032a0e 4178 int max;
9d8f0363 4179 unsigned int chunk_sectors = mddev->chunk_sectors;
cc371e66 4180 unsigned int bio_sectors = bvm->bi_size >> 9;
23032a0e 4181
cc371e66 4182 if ((bvm->bi_rw & 1) == WRITE)
23032a0e
RBJ		 4183 return biovec->bv_len; /* always allow writes to be mergeable */
4184
664e7c41
AN		 4185 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4186 chunk_sectors = mddev->new_chunk_sectors;
23032a0e
RBJ		 4187 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4188 if (max < 0) max = 0;
4189 if (max <= biovec->bv_len && bio_sectors == 0)
4190 return biovec->bv_len;
4191 else
4192 return max;
4193}
4194
fd01b88c 4195static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
f679623f 4196{
4f024f37 4197 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
9d8f0363 4198 unsigned int chunk_sectors = mddev->chunk_sectors;
aa8b57aa 4199 unsigned int bio_sectors = bio_sectors(bio);
f679623f 4200
664e7c41
AN		 4201 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4202 chunk_sectors = mddev->new_chunk_sectors;
f679623f
RBJ		 4203 return chunk_sectors >=
4204 ((sector & (chunk_sectors - 1)) + bio_sectors);
4205}
4206
46031f9a
RBJ		 4207/*
4208 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4209 * later sampled by raid5d.
4210 */
d1688a6d 4211static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
46031f9a
RBJ		 4212{
4213 unsigned long flags;
4214
4215 spin_lock_irqsave(&conf->device_lock, flags);
4216
4217 bi->bi_next = conf->retry_read_aligned_list;
4218 conf->retry_read_aligned_list = bi;
4219
4220 spin_unlock_irqrestore(&conf->device_lock, flags);
4221 md_wakeup_thread(conf->mddev->thread);
4222}
4223
d1688a6d 4224static struct bio *remove_bio_from_retry(struct r5conf *conf)
46031f9a
RBJ		 4225{
4226 struct bio *bi;
4227
4228 bi = conf->retry_read_aligned;
4229 if (bi) {
4230 conf->retry_read_aligned = NULL;
4231 return bi;
4232 }
4233 bi = conf->retry_read_aligned_list;
4234 if(bi) {
387bb173 4235 conf->retry_read_aligned_list = bi->bi_next;
46031f9a 4236 bi->bi_next = NULL;
960e739d
JA		 4237 /*
4238 * this sets the active strip count to 1 and the processed
4239 * strip count to zero (upper 8 bits)
4240 */
e7836bd6 4241 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
46031f9a
RBJ		 4242 }
4243
4244 return bi;
4245}
4246
f679623f
RBJ		 4247/*
4248 * The "raid5_align_endio" should check if the read succeeded and if it
4249 * did, call bio_endio on the original bio (having bio_put the new bio
4250 * first).
4251 * If the read failed..
4252 */
6712ecf8 4253static void raid5_align_endio(struct bio *bi, int error)
f679623f
RBJ		 4254{
4255 struct bio* raid_bi = bi->bi_private;
fd01b88c 4256 struct mddev *mddev;
d1688a6d 4257 struct r5conf *conf;
46031f9a 4258 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3cb03002 4259 struct md_rdev *rdev;
46031f9a 4260
f679623f 4261 bio_put(bi);
46031f9a 4262
46031f9a
RBJ		 4263 rdev = (void*)raid_bi->bi_next;
4264 raid_bi->bi_next = NULL;
2b7f2228
N		 4265 mddev = rdev->mddev;
4266 conf = mddev->private;
46031f9a
RBJ		 4267
4268 rdev_dec_pending(rdev, conf->mddev);
4269
4270 if (!error && uptodate) {
0a82a8d1
LT		 4271 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4272 raid_bi, 0);
6712ecf8 4273 bio_endio(raid_bi, 0);
46031f9a
RBJ		 4274 if (atomic_dec_and_test(&conf->active_aligned_reads))
4275 wake_up(&conf->wait_for_stripe);
6712ecf8 4276 return;
46031f9a
RBJ		 4277 }
4278
45b4233c 4279 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
46031f9a
RBJ		 4280
4281 add_bio_to_retry(raid_bi, conf);
f679623f
RBJ		 4282}
4283
387bb173
NB		 4284static int bio_fits_rdev(struct bio *bi)
4285{
165125e1 4286 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
387bb173 4287
aa8b57aa 4288 if (bio_sectors(bi) > queue_max_sectors(q))
387bb173
NB		 4289 return 0;
4290 blk_recount_segments(q, bi);
8a78362c 4291 if (bi->bi_phys_segments > queue_max_segments(q))
387bb173
NB		 4292 return 0;
4293
4294 if (q->merge_bvec_fn)
4295 /* it's too hard to apply the merge_bvec_fn at this stage,
4296 * just just give up
4297 */
4298 return 0;
4299
4300 return 1;
4301}
4302
fd01b88c 4303static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
f679623f 4304{
d1688a6d 4305 struct r5conf *conf = mddev->private;
8553fe7e 4306 int dd_idx;
f679623f 4307 struct bio* align_bi;
3cb03002 4308 struct md_rdev *rdev;
671488cc 4309 sector_t end_sector;
f679623f
RBJ		 4310
4311 if (!in_chunk_boundary(mddev, raid_bio)) {
45b4233c 4312 pr_debug("chunk_aligned_read : non aligned\n");
f679623f
RBJ		 4313 return 0;
4314 }
4315 /*
a167f663 4316 * use bio_clone_mddev to make a copy of the bio
f679623f 4317 */
a167f663 4318 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
f679623f
RBJ		 4319 if (!align_bi)
4320 return 0;
4321 /*
4322 * set bi_end_io to a new function, and set bi_private to the
4323 * original bio.
4324 */
4325 align_bi->bi_end_io = raid5_align_endio;
4326 align_bi->bi_private = raid_bio;
4327 /*
4328 * compute position
4329 */
4f024f37
KO		 4330 align_bi->bi_iter.bi_sector =
4331 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4332 0, &dd_idx, NULL);
f679623f 4333
f73a1c7d 4334 end_sector = bio_end_sector(align_bi);
f679623f 4335 rcu_read_lock();
671488cc
N		 4336 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4337 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4338 rdev->recovery_offset < end_sector) {
4339 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4340 if (rdev &&
4341 (test_bit(Faulty, &rdev->flags) ||
4342 !(test_bit(In_sync, &rdev->flags) ||
4343 rdev->recovery_offset >= end_sector)))
4344 rdev = NULL;
4345 }
4346 if (rdev) {
31c176ec
N		 4347 sector_t first_bad;
4348 int bad_sectors;
4349
f679623f
RBJ		 4350 atomic_inc(&rdev->nr_pending);
4351 rcu_read_unlock();
46031f9a
RBJ		 4352 raid_bio->bi_next = (void*)rdev;
4353 align_bi->bi_bdev = rdev->bdev;
3fd83717 4354 __clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
46031f9a 4355
31c176ec 4356 if (!bio_fits_rdev(align_bi) ||
4f024f37
KO		 4357 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4358 bio_sectors(align_bi),
31c176ec
N		 4359 &first_bad, &bad_sectors)) {
4360 /* too big in some way, or has a known bad block */
387bb173
NB		 4361 bio_put(align_bi);
4362 rdev_dec_pending(rdev, mddev);
4363 return 0;
4364 }
4365
6c0544e2 4366 /* No reshape active, so we can trust rdev->data_offset */
4f024f37 4367 align_bi->bi_iter.bi_sector += rdev->data_offset;
6c0544e2 4368
46031f9a
RBJ		 4369 spin_lock_irq(&conf->device_lock);
4370 wait_event_lock_irq(conf->wait_for_stripe,
4371 conf->quiesce == 0,
eed8c02e 4372 conf->device_lock);
46031f9a
RBJ		 4373 atomic_inc(&conf->active_aligned_reads);
4374 spin_unlock_irq(&conf->device_lock);
4375
e3620a3a
JB		 4376 if (mddev->gendisk)
4377 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4378 align_bi, disk_devt(mddev->gendisk),
4f024f37 4379 raid_bio->bi_iter.bi_sector);
f679623f
RBJ		 4380 generic_make_request(align_bi);
4381 return 1;
4382 } else {
4383 rcu_read_unlock();
46031f9a 4384 bio_put(align_bi);
f679623f
RBJ		 4385 return 0;
4386 }
4387}
4388
8b3e6cdc
DW		 4389/* __get_priority_stripe - get the next stripe to process
4390 *
4391 * Full stripe writes are allowed to pass preread active stripes up until
4392 * the bypass_threshold is exceeded. In general the bypass_count
4393 * increments when the handle_list is handled before the hold_list; however, it
4394 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4395 * stripe with in flight i/o. The bypass_count will be reset when the
4396 * head of the hold_list has changed, i.e. the head was promoted to the
4397 * handle_list.
4398 */
851c30c9 4399static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
8b3e6cdc 4400{
851c30c9
SL		 4401 struct stripe_head *sh = NULL, *tmp;
4402 struct list_head *handle_list = NULL;
bfc90cb0 4403 struct r5worker_group *wg = NULL;
851c30c9
SL		 4404
4405 if (conf->worker_cnt_per_group == 0) {
4406 handle_list = &conf->handle_list;
4407 } else if (group != ANY_GROUP) {
4408 handle_list = &conf->worker_groups[group].handle_list;
bfc90cb0 4409 wg = &conf->worker_groups[group];
851c30c9
SL		 4410 } else {
4411 int i;
4412 for (i = 0; i < conf->group_cnt; i++) {
4413 handle_list = &conf->worker_groups[i].handle_list;
bfc90cb0 4414 wg = &conf->worker_groups[i];
851c30c9
SL		 4415 if (!list_empty(handle_list))
4416 break;
4417 }
4418 }
8b3e6cdc
DW		 4419
4420 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4421 __func__,
851c30c9 4422 list_empty(handle_list) ? "empty" : "busy",
8b3e6cdc
DW		 4423 list_empty(&conf->hold_list) ? "empty" : "busy",
4424 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4425
851c30c9
SL		 4426 if (!list_empty(handle_list)) {
4427 sh = list_entry(handle_list->next, typeof(*sh), lru);
8b3e6cdc
DW		 4428
4429 if (list_empty(&conf->hold_list))
4430 conf->bypass_count = 0;
4431 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4432 if (conf->hold_list.next == conf->last_hold)
4433 conf->bypass_count++;
4434 else {
4435 conf->last_hold = conf->hold_list.next;
4436 conf->bypass_count -= conf->bypass_threshold;
4437 if (conf->bypass_count < 0)
4438 conf->bypass_count = 0;
4439 }
4440 }
4441 } else if (!list_empty(&conf->hold_list) &&
4442 ((conf->bypass_threshold &&
4443 conf->bypass_count > conf->bypass_threshold) ||
4444 atomic_read(&conf->pending_full_writes) == 0)) {
851c30c9
SL		 4445
4446 list_for_each_entry(tmp, &conf->hold_list, lru) {
4447 if (conf->worker_cnt_per_group == 0 ||
4448 group == ANY_GROUP ||
4449 !cpu_online(tmp->cpu) ||
4450 cpu_to_group(tmp->cpu) == group) {
4451 sh = tmp;
4452 break;
4453 }
4454 }
4455
4456 if (sh) {
4457 conf->bypass_count -= conf->bypass_threshold;
4458 if (conf->bypass_count < 0)
4459 conf->bypass_count = 0;
4460 }
bfc90cb0 4461 wg = NULL;
851c30c9
SL		 4462 }
4463
4464 if (!sh)
8b3e6cdc
DW		 4465 return NULL;
4466
bfc90cb0
SL		 4467 if (wg) {
4468 wg->stripes_cnt--;
4469 sh->group = NULL;
4470 }
8b3e6cdc 4471 list_del_init(&sh->lru);
c7a6d35e 4472 BUG_ON(atomic_inc_return(&sh->count) != 1);
8b3e6cdc
DW		 4473 return sh;
4474}
f679623f 4475
8811b596
SL		 4476struct raid5_plug_cb {
4477 struct blk_plug_cb cb;
4478 struct list_head list;
566c09c5 4479 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
8811b596
SL		 4480};
4481
4482static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4483{
4484 struct raid5_plug_cb *cb = container_of(
4485 blk_cb, struct raid5_plug_cb, cb);
4486 struct stripe_head *sh;
4487 struct mddev *mddev = cb->cb.data;
4488 struct r5conf *conf = mddev->private;
a9add5d9 4489 int cnt = 0;
566c09c5 4490 int hash;
8811b596
SL		 4491
4492 if (cb->list.next && !list_empty(&cb->list)) {
4493 spin_lock_irq(&conf->device_lock);
4494 while (!list_empty(&cb->list)) {
4495 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4496 list_del_init(&sh->lru);
4497 /*
4498 * avoid race release_stripe_plug() sees
4499 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4500 * is still in our list
4501 */
4e857c58 4502 smp_mb__before_atomic();
8811b596 4503 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
773ca82f
SL		 4504 /*
4505 * STRIPE_ON_RELEASE_LIST could be set here. In that
4506 * case, the count is always > 1 here
4507 */
566c09c5
SL		 4508 hash = sh->hash_lock_index;
4509 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
a9add5d9 4510 cnt++;
8811b596
SL		 4511 }
4512 spin_unlock_irq(&conf->device_lock);
4513 }
566c09c5
SL		 4514 release_inactive_stripe_list(conf, cb->temp_inactive_list,
4515 NR_STRIPE_HASH_LOCKS);
e3620a3a
JB		 4516 if (mddev->queue)
4517 trace_block_unplug(mddev->queue, cnt, !from_schedule);
8811b596
SL		 4518 kfree(cb);
4519}
4520
4521static void release_stripe_plug(struct mddev *mddev,
4522 struct stripe_head *sh)
4523{
4524 struct blk_plug_cb *blk_cb = blk_check_plugged(
4525 raid5_unplug, mddev,
4526 sizeof(struct raid5_plug_cb));
4527 struct raid5_plug_cb *cb;
4528
4529 if (!blk_cb) {
4530 release_stripe(sh);
4531 return;
4532 }
4533
4534 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4535
566c09c5
SL		 4536 if (cb->list.next == NULL) {
4537 int i;
8811b596 4538 INIT_LIST_HEAD(&cb->list);
566c09c5
SL		 4539 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
4540 INIT_LIST_HEAD(cb->temp_inactive_list + i);
4541 }
8811b596
SL		 4542
4543 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4544 list_add_tail(&sh->lru, &cb->list);
4545 else
4546 release_stripe(sh);
4547}
4548
620125f2
SL		 4549static void make_discard_request(struct mddev *mddev, struct bio *bi)
4550{
4551 struct r5conf *conf = mddev->private;
4552 sector_t logical_sector, last_sector;
4553 struct stripe_head *sh;
4554 int remaining;
4555 int stripe_sectors;
4556
4557 if (mddev->reshape_position != MaxSector)
4558 /* Skip discard while reshape is happening */
4559 return;
4560
4f024f37
KO		 4561 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4562 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
620125f2
SL		 4563
4564 bi->bi_next = NULL;
4565 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4566
4567 stripe_sectors = conf->chunk_sectors *
4568 (conf->raid_disks - conf->max_degraded);
4569 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4570 stripe_sectors);
4571 sector_div(last_sector, stripe_sectors);
4572
4573 logical_sector *= conf->chunk_sectors;
4574 last_sector *= conf->chunk_sectors;
4575
4576 for (; logical_sector < last_sector;
4577 logical_sector += STRIPE_SECTORS) {
4578 DEFINE_WAIT(w);
4579 int d;
4580 again:
4581 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4582 prepare_to_wait(&conf->wait_for_overlap, &w,
4583 TASK_UNINTERRUPTIBLE);
f8dfcffd
N		 4584 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4585 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4586 release_stripe(sh);
4587 schedule();
4588 goto again;
4589 }
4590 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
620125f2
SL		 4591 spin_lock_irq(&sh->stripe_lock);
4592 for (d = 0; d < conf->raid_disks; d++) {
4593 if (d == sh->pd_idx || d == sh->qd_idx)
4594 continue;
4595 if (sh->dev[d].towrite || sh->dev[d].toread) {
4596 set_bit(R5_Overlap, &sh->dev[d].flags);
4597 spin_unlock_irq(&sh->stripe_lock);
4598 release_stripe(sh);
4599 schedule();
4600 goto again;
4601 }
4602 }
f8dfcffd 4603 set_bit(STRIPE_DISCARD, &sh->state);
620125f2
SL		 4604 finish_wait(&conf->wait_for_overlap, &w);
4605 for (d = 0; d < conf->raid_disks; d++) {
4606 if (d == sh->pd_idx || d == sh->qd_idx)
4607 continue;
4608 sh->dev[d].towrite = bi;
4609 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4610 raid5_inc_bi_active_stripes(bi);
4611 }
4612 spin_unlock_irq(&sh->stripe_lock);
4613 if (conf->mddev->bitmap) {
4614 for (d = 0;
4615 d < conf->raid_disks - conf->max_degraded;
4616 d++)
4617 bitmap_startwrite(mddev->bitmap,
4618 sh->sector,
4619 STRIPE_SECTORS,
4620 0);
4621 sh->bm_seq = conf->seq_flush + 1;
4622 set_bit(STRIPE_BIT_DELAY, &sh->state);
4623 }
4624
4625 set_bit(STRIPE_HANDLE, &sh->state);
4626 clear_bit(STRIPE_DELAYED, &sh->state);
4627 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4628 atomic_inc(&conf->preread_active_stripes);
4629 release_stripe_plug(mddev, sh);
4630 }
4631
4632 remaining = raid5_dec_bi_active_stripes(bi);
4633 if (remaining == 0) {
4634 md_write_end(mddev);
4635 bio_endio(bi, 0);
4636 }
4637}
4638
b4fdcb02 4639static void make_request(struct mddev *mddev, struct bio * bi)
1da177e4 4640{
d1688a6d 4641 struct r5conf *conf = mddev->private;
911d4ee8 4642 int dd_idx;
1da177e4
LT		 4643 sector_t new_sector;
4644 sector_t logical_sector, last_sector;
4645 struct stripe_head *sh;
a362357b 4646 const int rw = bio_data_dir(bi);
49077326 4647 int remaining;
27c0f68f
SL		 4648 DEFINE_WAIT(w);
4649 bool do_prepare;
1da177e4 4650
e9c7469b
TH		 4651 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4652 md_flush_request(mddev, bi);
5a7bbad2 4653 return;
e5dcdd80
N		 4654 }
4655
3d310eb7 4656 md_write_start(mddev, bi);
06d91a5f 4657
802ba064 4658 if (rw == READ &&
52488615 4659 mddev->reshape_position == MaxSector &&
21a52c6d 4660 chunk_aligned_read(mddev,bi))
5a7bbad2 4661 return;
52488615 4662
620125f2
SL		 4663 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4664 make_discard_request(mddev, bi);
4665 return;
4666 }
4667
4f024f37 4668 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
f73a1c7d 4669 last_sector = bio_end_sector(bi);
1da177e4
LT		 4670 bi->bi_next = NULL;
4671 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
06d91a5f 4672
27c0f68f 4673 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1da177e4 4674 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
b5663ba4 4675 int previous;
c46501b2 4676 int seq;
b578d55f 4677
27c0f68f 4678 do_prepare = false;
7ecaa1e6 4679 retry:
c46501b2 4680 seq = read_seqcount_begin(&conf->gen_lock);
b5663ba4 4681 previous = 0;
27c0f68f
SL		 4682 if (do_prepare)
4683 prepare_to_wait(&conf->wait_for_overlap, &w,
4684 TASK_UNINTERRUPTIBLE);
b0f9ec04 4685 if (unlikely(conf->reshape_progress != MaxSector)) {
fef9c61f 4686 /* spinlock is needed as reshape_progress may be
df8e7f76
N		 4687 * 64bit on a 32bit platform, and so it might be
4688 * possible to see a half-updated value
aeb878b0 4689 * Of course reshape_progress could change after
df8e7f76
N		 4690 * the lock is dropped, so once we get a reference
4691 * to the stripe that we think it is, we will have
4692 * to check again.
4693 */
7ecaa1e6 4694 spin_lock_irq(&conf->device_lock);
2c810cdd 4695 if (mddev->reshape_backwards
fef9c61f
N		 4696 ? logical_sector < conf->reshape_progress
4697 : logical_sector >= conf->reshape_progress) {
b5663ba4
N		 4698 previous = 1;
4699 } else {
2c810cdd 4700 if (mddev->reshape_backwards
fef9c61f
N		 4701 ? logical_sector < conf->reshape_safe
4702 : logical_sector >= conf->reshape_safe) {
b578d55f
N		 4703 spin_unlock_irq(&conf->device_lock);
4704 schedule();
27c0f68f 4705 do_prepare = true;
b578d55f
N		 4706 goto retry;
4707 }
4708 }
7ecaa1e6
N		 4709 spin_unlock_irq(&conf->device_lock);
4710 }
16a53ecc 4711
112bf897
N		 4712 new_sector = raid5_compute_sector(conf, logical_sector,
4713 previous,
911d4ee8 4714 &dd_idx, NULL);
0c55e022 4715 pr_debug("raid456: make_request, sector %llu logical %llu\n",
c46501b2 4716 (unsigned long long)new_sector,
1da177e4
LT		 4717 (unsigned long long)logical_sector);
4718
b5663ba4 4719 sh = get_active_stripe(conf, new_sector, previous,
a8c906ca 4720 (bi->bi_rw&RWA_MASK), 0);
1da177e4 4721 if (sh) {
b0f9ec04 4722 if (unlikely(previous)) {
7ecaa1e6 4723 /* expansion might have moved on while waiting for a
df8e7f76
N		 4724 * stripe, so we must do the range check again.
4725 * Expansion could still move past after this
4726 * test, but as we are holding a reference to
4727 * 'sh', we know that if that happens,
4728 * STRIPE_EXPANDING will get set and the expansion
4729 * won't proceed until we finish with the stripe.
7ecaa1e6
N		 4730 */
4731 int must_retry = 0;
4732 spin_lock_irq(&conf->device_lock);
2c810cdd 4733 if (mddev->reshape_backwards
b0f9ec04
N		 4734 ? logical_sector >= conf->reshape_progress
4735 : logical_sector < conf->reshape_progress)
7ecaa1e6
N		 4736 /* mismatch, need to try again */
4737 must_retry = 1;
4738 spin_unlock_irq(&conf->device_lock);
4739 if (must_retry) {
4740 release_stripe(sh);
7a3ab908 4741 schedule();
27c0f68f 4742 do_prepare = true;
7ecaa1e6
N		 4743 goto retry;
4744 }
4745 }
c46501b2
N		 4746 if (read_seqcount_retry(&conf->gen_lock, seq)) {
4747 /* Might have got the wrong stripe_head
4748 * by accident
4749 */
4750 release_stripe(sh);
4751 goto retry;
4752 }
e62e58a5 4753
ffd96e35 4754 if (rw == WRITE &&
a5c308d4 4755 logical_sector >= mddev->suspend_lo &&
e464eafd
N		 4756 logical_sector < mddev->suspend_hi) {
4757 release_stripe(sh);
e62e58a5
N		 4758 /* As the suspend_* range is controlled by
4759 * userspace, we want an interruptible
4760 * wait.
4761 */
4762 flush_signals(current);
4763 prepare_to_wait(&conf->wait_for_overlap,
4764 &w, TASK_INTERRUPTIBLE);
4765 if (logical_sector >= mddev->suspend_lo &&
27c0f68f 4766 logical_sector < mddev->suspend_hi) {
e62e58a5 4767 schedule();
27c0f68f
SL		 4768 do_prepare = true;
4769 }
e464eafd
N		 4770 goto retry;
4771 }
7ecaa1e6
N		 4772
4773 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
ffd96e35 4774 !add_stripe_bio(sh, bi, dd_idx, rw)) {
7ecaa1e6
N		 4775 /* Stripe is busy expanding or
4776 * add failed due to overlap. Flush everything
1da177e4
LT		 4777 * and wait a while
4778 */
482c0834 4779 md_wakeup_thread(mddev->thread);
1da177e4
LT		 4780 release_stripe(sh);
4781 schedule();
27c0f68f 4782 do_prepare = true;
1da177e4
LT		 4783 goto retry;
4784 }
6ed3003c
N		 4785 set_bit(STRIPE_HANDLE, &sh->state);
4786 clear_bit(STRIPE_DELAYED, &sh->state);
a852d7b8 4787 if ((bi->bi_rw & REQ_SYNC) &&
729a1866
N		 4788 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4789 atomic_inc(&conf->preread_active_stripes);
8811b596 4790 release_stripe_plug(mddev, sh);
1da177e4
LT		 4791 } else {
4792 /* cannot get stripe for read-ahead, just give-up */
4793 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1da177e4
LT		 4794 break;
4795 }
1da177e4 4796 }
27c0f68f 4797 finish_wait(&conf->wait_for_overlap, &w);
7c13edc8 4798
e7836bd6 4799 remaining = raid5_dec_bi_active_stripes(bi);
f6344757 4800 if (remaining == 0) {
1da177e4 4801
16a53ecc 4802 if ( rw == WRITE )
1da177e4 4803 md_write_end(mddev);
6712ecf8 4804
0a82a8d1
LT		 4805 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4806 bi, 0);
0e13fe23 4807 bio_endio(bi, 0);
1da177e4 4808 }
1da177e4
LT		 4809}
4810
fd01b88c 4811static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
b522adcd 4812
fd01b88c 4813static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
1da177e4 4814{
52c03291
N		 4815 /* reshaping is quite different to recovery/resync so it is
4816 * handled quite separately ... here.
4817 *
4818 * On each call to sync_request, we gather one chunk worth of
4819 * destination stripes and flag them as expanding.
4820 * Then we find all the source stripes and request reads.
4821 * As the reads complete, handle_stripe will copy the data
4822 * into the destination stripe and release that stripe.
4823 */
d1688a6d 4824 struct r5conf *conf = mddev->private;
1da177e4 4825 struct stripe_head *sh;
ccfcc3c1 4826 sector_t first_sector, last_sector;
f416885e
N		 4827 int raid_disks = conf->previous_raid_disks;
4828 int data_disks = raid_disks - conf->max_degraded;
4829 int new_data_disks = conf->raid_disks - conf->max_degraded;
52c03291
N		 4830 int i;
4831 int dd_idx;
c8f517c4 4832 sector_t writepos, readpos, safepos;
ec32a2bd 4833 sector_t stripe_addr;
7a661381 4834 int reshape_sectors;
ab69ae12 4835 struct list_head stripes;
52c03291 4836
fef9c61f
N		 4837 if (sector_nr == 0) {
4838 /* If restarting in the middle, skip the initial sectors */
2c810cdd 4839 if (mddev->reshape_backwards &&
fef9c61f
N		 4840 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4841 sector_nr = raid5_size(mddev, 0, 0)
4842 - conf->reshape_progress;
2c810cdd 4843 } else if (!mddev->reshape_backwards &&
fef9c61f
N		 4844 conf->reshape_progress > 0)
4845 sector_nr = conf->reshape_progress;
f416885e 4846 sector_div(sector_nr, new_data_disks);
fef9c61f 4847 if (sector_nr) {
8dee7211
N		 4848 mddev->curr_resync_completed = sector_nr;
4849 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
fef9c61f
N		 4850 *skipped = 1;
4851 return sector_nr;
4852 }
52c03291
N		 4853 }
4854
7a661381
N		 4855 /* We need to process a full chunk at a time.
4856 * If old and new chunk sizes differ, we need to process the
4857 * largest of these
4858 */
664e7c41
AN		 4859 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4860 reshape_sectors = mddev->new_chunk_sectors;
7a661381 4861 else
9d8f0363 4862 reshape_sectors = mddev->chunk_sectors;
7a661381 4863
b5254dd5
N		 4864 /* We update the metadata at least every 10 seconds, or when
4865 * the data about to be copied would over-write the source of
4866 * the data at the front of the range. i.e. one new_stripe
4867 * along from reshape_progress new_maps to after where
4868 * reshape_safe old_maps to
52c03291 4869 */
fef9c61f 4870 writepos = conf->reshape_progress;
f416885e 4871 sector_div(writepos, new_data_disks);
c8f517c4
N		 4872 readpos = conf->reshape_progress;
4873 sector_div(readpos, data_disks);
fef9c61f 4874 safepos = conf->reshape_safe;
f416885e 4875 sector_div(safepos, data_disks);
2c810cdd 4876 if (mddev->reshape_backwards) {
ed37d83e 4877 writepos -= min_t(sector_t, reshape_sectors, writepos);
c8f517c4 4878 readpos += reshape_sectors;
7a661381 4879 safepos += reshape_sectors;
fef9c61f 4880 } else {
7a661381 4881 writepos += reshape_sectors;
ed37d83e
N		 4882 readpos -= min_t(sector_t, reshape_sectors, readpos);
4883 safepos -= min_t(sector_t, reshape_sectors, safepos);
fef9c61f 4884 }
52c03291 4885
b5254dd5
N		 4886 /* Having calculated the 'writepos' possibly use it
4887 * to set 'stripe_addr' which is where we will write to.
4888 */
4889 if (mddev->reshape_backwards) {
4890 BUG_ON(conf->reshape_progress == 0);
4891 stripe_addr = writepos;
4892 BUG_ON((mddev->dev_sectors &
4893 ~((sector_t)reshape_sectors - 1))
4894 - reshape_sectors - stripe_addr
4895 != sector_nr);
4896 } else {
4897 BUG_ON(writepos != sector_nr + reshape_sectors);
4898 stripe_addr = sector_nr;
4899 }
4900
c8f517c4
N		 4901 /* 'writepos' is the most advanced device address we might write.
4902 * 'readpos' is the least advanced device address we might read.
4903 * 'safepos' is the least address recorded in the metadata as having
4904 * been reshaped.
b5254dd5
N		 4905 * If there is a min_offset_diff, these are adjusted either by
4906 * increasing the safepos/readpos if diff is negative, or
4907 * increasing writepos if diff is positive.
4908 * If 'readpos' is then behind 'writepos', there is no way that we can
c8f517c4
N		 4909 * ensure safety in the face of a crash - that must be done by userspace
4910 * making a backup of the data. So in that case there is no particular
4911 * rush to update metadata.
4912 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4913 * update the metadata to advance 'safepos' to match 'readpos' so that
4914 * we can be safe in the event of a crash.
4915 * So we insist on updating metadata if safepos is behind writepos and
4916 * readpos is beyond writepos.
4917 * In any case, update the metadata every 10 seconds.
4918 * Maybe that number should be configurable, but I'm not sure it is
4919 * worth it.... maybe it could be a multiple of safemode_delay???
4920 */
b5254dd5
N		 4921 if (conf->min_offset_diff < 0) {
4922 safepos += -conf->min_offset_diff;
4923 readpos += -conf->min_offset_diff;
4924 } else
4925 writepos += conf->min_offset_diff;
4926
2c810cdd 4927 if ((mddev->reshape_backwards
c8f517c4
N		 4928 ? (safepos > writepos && readpos < writepos)
4929 : (safepos < writepos && readpos > writepos)) ||
4930 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
52c03291
N		 4931 /* Cannot proceed until we've updated the superblock... */
4932 wait_event(conf->wait_for_overlap,
c91abf5a
N		 4933 atomic_read(&conf->reshape_stripes)==0
4934 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4935 if (atomic_read(&conf->reshape_stripes) != 0)
4936 return 0;
fef9c61f 4937 mddev->reshape_position = conf->reshape_progress;
75d3da43 4938 mddev->curr_resync_completed = sector_nr;
c8f517c4 4939 conf->reshape_checkpoint = jiffies;
850b2b42 4940 set_bit(MD_CHANGE_DEVS, &mddev->flags);
52c03291 4941 md_wakeup_thread(mddev->thread);
850b2b42 4942 wait_event(mddev->sb_wait, mddev->flags == 0 ||
c91abf5a
N		 4943 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4944 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4945 return 0;
52c03291 4946 spin_lock_irq(&conf->device_lock);
fef9c61f 4947 conf->reshape_safe = mddev->reshape_position;
52c03291
N		 4948 spin_unlock_irq(&conf->device_lock);
4949 wake_up(&conf->wait_for_overlap);
acb180b0 4950 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
52c03291
N		 4951 }
4952
ab69ae12 4953 INIT_LIST_HEAD(&stripes);
7a661381 4954 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
52c03291 4955 int j;
a9f326eb 4956 int skipped_disk = 0;
a8c906ca 4957 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
52c03291
N		 4958 set_bit(STRIPE_EXPANDING, &sh->state);
4959 atomic_inc(&conf->reshape_stripes);
4960 /* If any of this stripe is beyond the end of the old
4961 * array, then we need to zero those blocks
4962 */
4963 for (j=sh->disks; j--;) {
4964 sector_t s;
4965 if (j == sh->pd_idx)
4966 continue;
f416885e 4967 if (conf->level == 6 &&
d0dabf7e 4968 j == sh->qd_idx)
f416885e 4969 continue;
784052ec 4970 s = compute_blocknr(sh, j, 0);
b522adcd 4971 if (s < raid5_size(mddev, 0, 0)) {
a9f326eb 4972 skipped_disk = 1;
52c03291
N		 4973 continue;
4974 }
4975 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4976 set_bit(R5_Expanded, &sh->dev[j].flags);
4977 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4978 }
a9f326eb 4979 if (!skipped_disk) {
52c03291
N		 4980 set_bit(STRIPE_EXPAND_READY, &sh->state);
4981 set_bit(STRIPE_HANDLE, &sh->state);
4982 }
ab69ae12 4983 list_add(&sh->lru, &stripes);
52c03291
N		 4984 }
4985 spin_lock_irq(&conf->device_lock);
2c810cdd 4986 if (mddev->reshape_backwards)
7a661381 4987 conf->reshape_progress -= reshape_sectors * new_data_disks;
fef9c61f 4988 else
7a661381 4989 conf->reshape_progress += reshape_sectors * new_data_disks;
52c03291
N		 4990 spin_unlock_irq(&conf->device_lock);
4991 /* Ok, those stripe are ready. We can start scheduling
4992 * reads on the source stripes.
4993 * The source stripes are determined by mapping the first and last
4994 * block on the destination stripes.
4995 */
52c03291 4996 first_sector =
ec32a2bd 4997 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
911d4ee8 4998 1, &dd_idx, NULL);
52c03291 4999 last_sector =
0e6e0271 5000 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
09c9e5fa 5001 * new_data_disks - 1),
911d4ee8 5002 1, &dd_idx, NULL);
58c0fed4
AN		 5003 if (last_sector >= mddev->dev_sectors)
5004 last_sector = mddev->dev_sectors - 1;
52c03291 5005 while (first_sector <= last_sector) {
a8c906ca 5006 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
52c03291
N		 5007 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5008 set_bit(STRIPE_HANDLE, &sh->state);
5009 release_stripe(sh);
5010 first_sector += STRIPE_SECTORS;
5011 }
ab69ae12
N		 5012 /* Now that the sources are clearly marked, we can release
5013 * the destination stripes
5014 */
5015 while (!list_empty(&stripes)) {
5016 sh = list_entry(stripes.next, struct stripe_head, lru);
5017 list_del_init(&sh->lru);
5018 release_stripe(sh);
5019 }
c6207277
N		 5020 /* If this takes us to the resync_max point where we have to pause,
5021 * then we need to write out the superblock.
5022 */
7a661381 5023 sector_nr += reshape_sectors;
c03f6a19
N		 5024 if ((sector_nr - mddev->curr_resync_completed) * 2
5025 >= mddev->resync_max - mddev->curr_resync_completed) {
c6207277
N		 5026 /* Cannot proceed until we've updated the superblock... */
5027 wait_event(conf->wait_for_overlap,
c91abf5a
N		 5028 atomic_read(&conf->reshape_stripes) == 0
5029 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5030 if (atomic_read(&conf->reshape_stripes) != 0)
5031 goto ret;
fef9c61f 5032 mddev->reshape_position = conf->reshape_progress;
75d3da43 5033 mddev->curr_resync_completed = sector_nr;
c8f517c4 5034 conf->reshape_checkpoint = jiffies;
c6207277
N		 5035 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5036 md_wakeup_thread(mddev->thread);
5037 wait_event(mddev->sb_wait,
5038 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
c91abf5a
N		 5039 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5040 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5041 goto ret;
c6207277 5042 spin_lock_irq(&conf->device_lock);
fef9c61f 5043 conf->reshape_safe = mddev->reshape_position;
c6207277
N		 5044 spin_unlock_irq(&conf->device_lock);
5045 wake_up(&conf->wait_for_overlap);
acb180b0 5046 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
c6207277 5047 }
c91abf5a 5048ret:
7a661381 5049 return reshape_sectors;
52c03291
N		 5050}
5051
5052/* FIXME go_faster isn't used */
fd01b88c 5053static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
52c03291 5054{
d1688a6d 5055 struct r5conf *conf = mddev->private;
52c03291 5056 struct stripe_head *sh;
58c0fed4 5057 sector_t max_sector = mddev->dev_sectors;
57dab0bd 5058 sector_t sync_blocks;
16a53ecc
N		 5059 int still_degraded = 0;
5060 int i;
1da177e4 5061
72626685 5062 if (sector_nr >= max_sector) {
1da177e4 5063 /* just being told to finish up .. nothing much to do */
cea9c228 5064
29269553
N		 5065 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5066 end_reshape(conf);
5067 return 0;
5068 }
72626685
N		 5069
5070 if (mddev->curr_resync < max_sector) /* aborted */
5071 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5072 &sync_blocks, 1);
16a53ecc 5073 else /* completed sync */
72626685
N		 5074 conf->fullsync = 0;
5075 bitmap_close_sync(mddev->bitmap);
5076
1da177e4
LT		 5077 return 0;
5078 }
ccfcc3c1 5079
64bd660b
N		 5080 /* Allow raid5_quiesce to complete */
5081 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5082
52c03291
N		 5083 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5084 return reshape_request(mddev, sector_nr, skipped);
f6705578 5085
c6207277
N		 5086 /* No need to check resync_max as we never do more than one
5087 * stripe, and as resync_max will always be on a chunk boundary,
5088 * if the check in md_do_sync didn't fire, there is no chance
5089 * of overstepping resync_max here
5090 */
5091
16a53ecc 5092 /* if there is too many failed drives and we are trying
1da177e4
LT		 5093 * to resync, then assert that we are finished, because there is
5094 * nothing we can do.
5095 */
3285edf1 5096 if (mddev->degraded >= conf->max_degraded &&
16a53ecc 5097 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
58c0fed4 5098 sector_t rv = mddev->dev_sectors - sector_nr;
57afd89f 5099 *skipped = 1;
1da177e4
LT		 5100 return rv;
5101 }
6f608040 5102 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5103 !conf->fullsync &&
5104 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5105 sync_blocks >= STRIPE_SECTORS) {
72626685
N		 5106 /* we can skip this block, and probably more */
5107 sync_blocks /= STRIPE_SECTORS;
5108 *skipped = 1;
5109 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5110 }
1da177e4 5111
b47490c9
N		 5112 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5113
a8c906ca 5114 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
1da177e4 5115 if (sh == NULL) {
a8c906ca 5116 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
1da177e4 5117 /* make sure we don't swamp the stripe cache if someone else
16a53ecc 5118 * is trying to get access
1da177e4 5119 */
66c006a5 5120 schedule_timeout_uninterruptible(1);
1da177e4 5121 }
16a53ecc
N		 5122 /* Need to check if array will still be degraded after recovery/resync
5123 * We don't need to check the 'failed' flag as when that gets set,
5124 * recovery aborts.
5125 */
f001a70c 5126 for (i = 0; i < conf->raid_disks; i++)
16a53ecc
N		 5127 if (conf->disks[i].rdev == NULL)
5128 still_degraded = 1;
5129
5130 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5131
83206d66 5132 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
053f5b65 5133 set_bit(STRIPE_HANDLE, &sh->state);
1da177e4 5134
1da177e4
LT		 5135 release_stripe(sh);
5136
5137 return STRIPE_SECTORS;
5138}
5139
d1688a6d 5140static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
46031f9a
RBJ		 5141{
5142 /* We may not be able to submit a whole bio at once as there
5143 * may not be enough stripe_heads available.
5144 * We cannot pre-allocate enough stripe_heads as we may need
5145 * more than exist in the cache (if we allow ever large chunks).
5146 * So we do one stripe head at a time and record in
5147 * ->bi_hw_segments how many have been done.
5148 *
5149 * We *know* that this entire raid_bio is in one chunk, so
5150 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5151 */
5152 struct stripe_head *sh;
911d4ee8 5153 int dd_idx;
46031f9a
RBJ		 5154 sector_t sector, logical_sector, last_sector;
5155 int scnt = 0;
5156 int remaining;
5157 int handled = 0;
5158
4f024f37
KO		 5159 logical_sector = raid_bio->bi_iter.bi_sector &
5160 ~((sector_t)STRIPE_SECTORS-1);
112bf897 5161 sector = raid5_compute_sector(conf, logical_sector,
911d4ee8 5162 0, &dd_idx, NULL);
f73a1c7d 5163 last_sector = bio_end_sector(raid_bio);
46031f9a
RBJ		 5164
5165 for (; logical_sector < last_sector;
387bb173
NB		 5166 logical_sector += STRIPE_SECTORS,
5167 sector += STRIPE_SECTORS,
5168 scnt++) {
46031f9a 5169
e7836bd6 5170 if (scnt < raid5_bi_processed_stripes(raid_bio))
46031f9a
RBJ		 5171 /* already done this stripe */
5172 continue;
5173
2844dc32 5174 sh = get_active_stripe(conf, sector, 0, 1, 1);
46031f9a
RBJ		 5175
5176 if (!sh) {
5177 /* failed to get a stripe - must wait */
e7836bd6 5178 raid5_set_bi_processed_stripes(raid_bio, scnt);
46031f9a
RBJ		 5179 conf->retry_read_aligned = raid_bio;
5180 return handled;
5181 }
5182
387bb173
NB		 5183 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
5184 release_stripe(sh);
e7836bd6 5185 raid5_set_bi_processed_stripes(raid_bio, scnt);
387bb173
NB		 5186 conf->retry_read_aligned = raid_bio;
5187 return handled;
5188 }
5189
3f9e7c14 5190 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
36d1c647 5191 handle_stripe(sh);
46031f9a
RBJ		 5192 release_stripe(sh);
5193 handled++;
5194 }
e7836bd6 5195 remaining = raid5_dec_bi_active_stripes(raid_bio);
0a82a8d1
LT		 5196 if (remaining == 0) {
5197 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5198 raid_bio, 0);
0e13fe23 5199 bio_endio(raid_bio, 0);
0a82a8d1 5200 }
46031f9a
RBJ		 5201 if (atomic_dec_and_test(&conf->active_aligned_reads))
5202 wake_up(&conf->wait_for_stripe);
5203 return handled;
5204}
5205
bfc90cb0 5206static int handle_active_stripes(struct r5conf *conf, int group,
566c09c5
SL		 5207 struct r5worker *worker,
5208 struct list_head *temp_inactive_list)
46a06401
SL		 5209{
5210 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
566c09c5
SL		 5211 int i, batch_size = 0, hash;
5212 bool release_inactive = false;
46a06401
SL		 5213
5214 while (batch_size < MAX_STRIPE_BATCH &&
851c30c9 5215 (sh = __get_priority_stripe(conf, group)) != NULL)
46a06401
SL		 5216 batch[batch_size++] = sh;
5217
566c09c5
SL		 5218 if (batch_size == 0) {
5219 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5220 if (!list_empty(temp_inactive_list + i))
5221 break;
5222 if (i == NR_STRIPE_HASH_LOCKS)
5223 return batch_size;
5224 release_inactive = true;
5225 }
46a06401
SL		 5226 spin_unlock_irq(&conf->device_lock);
5227
566c09c5
SL		 5228 release_inactive_stripe_list(conf, temp_inactive_list,
5229 NR_STRIPE_HASH_LOCKS);
5230
5231 if (release_inactive) {
5232 spin_lock_irq(&conf->device_lock);
5233 return 0;
5234 }
5235
46a06401
SL		 5236 for (i = 0; i < batch_size; i++)
5237 handle_stripe(batch[i]);
5238
5239 cond_resched();
5240
5241 spin_lock_irq(&conf->device_lock);
566c09c5
SL		 5242 for (i = 0; i < batch_size; i++) {
5243 hash = batch[i]->hash_lock_index;
5244 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5245 }
46a06401
SL		 5246 return batch_size;
5247}
46031f9a 5248
851c30c9
SL		 5249static void raid5_do_work(struct work_struct *work)
5250{
5251 struct r5worker *worker = container_of(work, struct r5worker, work);
5252 struct r5worker_group *group = worker->group;
5253 struct r5conf *conf = group->conf;
5254 int group_id = group - conf->worker_groups;
5255 int handled;
5256 struct blk_plug plug;
5257
5258 pr_debug("+++ raid5worker active\n");
5259
5260 blk_start_plug(&plug);
5261 handled = 0;
5262 spin_lock_irq(&conf->device_lock);
5263 while (1) {
5264 int batch_size, released;
5265
566c09c5 5266 released = release_stripe_list(conf, worker->temp_inactive_list);
851c30c9 5267
566c09c5
SL		 5268 batch_size = handle_active_stripes(conf, group_id, worker,
5269 worker->temp_inactive_list);
bfc90cb0 5270 worker->working = false;
851c30c9
SL		 5271 if (!batch_size && !released)
5272 break;
5273 handled += batch_size;
5274 }
5275 pr_debug("%d stripes handled\n", handled);
5276
5277 spin_unlock_irq(&conf->device_lock);
5278 blk_finish_plug(&plug);
5279
5280 pr_debug("--- raid5worker inactive\n");
5281}
5282
1da177e4
LT		 5283/*
5284 * This is our raid5 kernel thread.
5285 *
5286 * We scan the hash table for stripes which can be handled now.
5287 * During the scan, completed stripes are saved for us by the interrupt
5288 * handler, so that they will not have to wait for our next wakeup.
5289 */
4ed8731d 5290static void raid5d(struct md_thread *thread)
1da177e4 5291{
4ed8731d 5292 struct mddev *mddev = thread->mddev;
d1688a6d 5293 struct r5conf *conf = mddev->private;
1da177e4 5294 int handled;
e1dfa0a2 5295 struct blk_plug plug;
1da177e4 5296
45b4233c 5297 pr_debug("+++ raid5d active\n");
1da177e4
LT		 5298
5299 md_check_recovery(mddev);
1da177e4 5300
e1dfa0a2 5301 blk_start_plug(&plug);
1da177e4
LT		 5302 handled = 0;
5303 spin_lock_irq(&conf->device_lock);
5304 while (1) {
46031f9a 5305 struct bio *bio;
773ca82f
SL		 5306 int batch_size, released;
5307
566c09c5 5308 released = release_stripe_list(conf, conf->temp_inactive_list);
1da177e4 5309
0021b7bc 5310 if (
7c13edc8
N		 5311 !list_empty(&conf->bitmap_list)) {
5312 /* Now is a good time to flush some bitmap updates */
5313 conf->seq_flush++;
700e432d 5314 spin_unlock_irq(&conf->device_lock);
72626685 5315 bitmap_unplug(mddev->bitmap);
700e432d 5316 spin_lock_irq(&conf->device_lock);
7c13edc8 5317 conf->seq_write = conf->seq_flush;
566c09c5 5318 activate_bit_delay(conf, conf->temp_inactive_list);
72626685 5319 }
0021b7bc 5320 raid5_activate_delayed(conf);
72626685 5321
46031f9a
RBJ		 5322 while ((bio = remove_bio_from_retry(conf))) {
5323 int ok;
5324 spin_unlock_irq(&conf->device_lock);
5325 ok = retry_aligned_read(conf, bio);
5326 spin_lock_irq(&conf->device_lock);
5327 if (!ok)
5328 break;
5329 handled++;
5330 }
5331
566c09c5
SL		 5332 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5333 conf->temp_inactive_list);
773ca82f 5334 if (!batch_size && !released)
1da177e4 5335 break;
46a06401 5336 handled += batch_size;
1da177e4 5337
46a06401
SL		 5338 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5339 spin_unlock_irq(&conf->device_lock);
de393cde 5340 md_check_recovery(mddev);
46a06401
SL		 5341 spin_lock_irq(&conf->device_lock);
5342 }
1da177e4 5343 }
45b4233c 5344 pr_debug("%d stripes handled\n", handled);
1da177e4
LT		 5345
5346 spin_unlock_irq(&conf->device_lock);
5347
c9f21aaf 5348 async_tx_issue_pending_all();
e1dfa0a2 5349 blk_finish_plug(&plug);
1da177e4 5350
45b4233c 5351 pr_debug("--- raid5d inactive\n");
1da177e4
LT		 5352}
5353
3f294f4f 5354static ssize_t
fd01b88c 5355raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
3f294f4f 5356{
d1688a6d 5357 struct r5conf *conf = mddev->private;
96de1e66
N		 5358 if (conf)
5359 return sprintf(page, "%d\n", conf->max_nr_stripes);
5360 else
5361 return 0;
3f294f4f
N		 5362}
5363
c41d4ac4 5364int
fd01b88c 5365raid5_set_cache_size(struct mddev *mddev, int size)
3f294f4f 5366{
d1688a6d 5367 struct r5conf *conf = mddev->private;
b5470dc5 5368 int err;
566c09c5 5369 int hash;
b5470dc5 5370
c41d4ac4 5371 if (size <= 16 || size > 32768)
3f294f4f 5372 return -EINVAL;
566c09c5 5373 hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
c41d4ac4 5374 while (size < conf->max_nr_stripes) {
566c09c5 5375 if (drop_one_stripe(conf, hash))
3f294f4f
N		 5376 conf->max_nr_stripes--;
5377 else
5378 break;
566c09c5
SL		 5379 hash--;
5380 if (hash < 0)
5381 hash = NR_STRIPE_HASH_LOCKS - 1;
3f294f4f 5382 }
b5470dc5
DW		 5383 err = md_allow_write(mddev);
5384 if (err)
5385 return err;
566c09c5 5386 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
c41d4ac4 5387 while (size > conf->max_nr_stripes) {
566c09c5 5388 if (grow_one_stripe(conf, hash))
3f294f4f
N		 5389 conf->max_nr_stripes++;
5390 else break;
566c09c5 5391 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
3f294f4f 5392 }
c41d4ac4
N		 5393 return 0;
5394}
5395EXPORT_SYMBOL(raid5_set_cache_size);
5396
5397static ssize_t
fd01b88c 5398raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
c41d4ac4 5399{
d1688a6d 5400 struct r5conf *conf = mddev->private;
c41d4ac4
N		 5401 unsigned long new;
5402 int err;
5403
5404 if (len >= PAGE_SIZE)
5405 return -EINVAL;
5406 if (!conf)
5407 return -ENODEV;
5408
b29bebd6 5409 if (kstrtoul(page, 10, &new))
c41d4ac4
N		 5410 return -EINVAL;
5411 err = raid5_set_cache_size(mddev, new);
5412 if (err)
5413 return err;
3f294f4f
N		 5414 return len;
5415}
007583c9 5416
96de1e66
N		 5417static struct md_sysfs_entry
5418raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5419 raid5_show_stripe_cache_size,
5420 raid5_store_stripe_cache_size);
3f294f4f 5421
8b3e6cdc 5422static ssize_t
fd01b88c 5423raid5_show_preread_threshold(struct mddev *mddev, char *page)
8b3e6cdc 5424{
d1688a6d 5425 struct r5conf *conf = mddev->private;
8b3e6cdc
DW		 5426 if (conf)
5427 return sprintf(page, "%d\n", conf->bypass_threshold);
5428 else
5429 return 0;
5430}
5431
5432static ssize_t
fd01b88c 5433raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
8b3e6cdc 5434{
d1688a6d 5435 struct r5conf *conf = mddev->private;
4ef197d8 5436 unsigned long new;
8b3e6cdc
DW		 5437 if (len >= PAGE_SIZE)
5438 return -EINVAL;
5439 if (!conf)
5440 return -ENODEV;
5441
b29bebd6 5442 if (kstrtoul(page, 10, &new))
8b3e6cdc 5443 return -EINVAL;
4ef197d8 5444 if (new > conf->max_nr_stripes)
8b3e6cdc
DW		 5445 return -EINVAL;
5446 conf->bypass_threshold = new;
5447 return len;
5448}
5449
5450static struct md_sysfs_entry
5451raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5452 S_IRUGO | S_IWUSR,
5453 raid5_show_preread_threshold,
5454 raid5_store_preread_threshold);
5455
d592a996
SL		 5456static ssize_t
5457raid5_show_skip_copy(struct mddev *mddev, char *page)
5458{
5459 struct r5conf *conf = mddev->private;
5460 if (conf)
5461 return sprintf(page, "%d\n", conf->skip_copy);
5462 else
5463 return 0;
5464}
5465
5466static ssize_t
5467raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
5468{
5469 struct r5conf *conf = mddev->private;
5470 unsigned long new;
5471 if (len >= PAGE_SIZE)
5472 return -EINVAL;
5473 if (!conf)
5474 return -ENODEV;
5475
5476 if (kstrtoul(page, 10, &new))
5477 return -EINVAL;
5478 new = !!new;
5479 if (new == conf->skip_copy)
5480 return len;
5481
5482 mddev_suspend(mddev);
5483 conf->skip_copy = new;
5484 if (new)
5485 mddev->queue->backing_dev_info.capabilities |=
5486 BDI_CAP_STABLE_WRITES;
5487 else
5488 mddev->queue->backing_dev_info.capabilities &=
5489 ~BDI_CAP_STABLE_WRITES;
5490 mddev_resume(mddev);
5491 return len;
5492}
5493
5494static struct md_sysfs_entry
5495raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
5496 raid5_show_skip_copy,
5497 raid5_store_skip_copy);
5498
3f294f4f 5499static ssize_t
fd01b88c 5500stripe_cache_active_show(struct mddev *mddev, char *page)
3f294f4f 5501{
d1688a6d 5502 struct r5conf *conf = mddev->private;
96de1e66
N		 5503 if (conf)
5504 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5505 else
5506 return 0;
3f294f4f
N		 5507}
5508
96de1e66
N		 5509static struct md_sysfs_entry
5510raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
3f294f4f 5511
b721420e
SL		 5512static ssize_t
5513raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
5514{
5515 struct r5conf *conf = mddev->private;
5516 if (conf)
5517 return sprintf(page, "%d\n", conf->worker_cnt_per_group);
5518 else
5519 return 0;
5520}
5521
60aaf933 5522static int alloc_thread_groups(struct r5conf *conf, int cnt,
5523 int *group_cnt,
5524 int *worker_cnt_per_group,
5525 struct r5worker_group **worker_groups);
b721420e
SL		 5526static ssize_t
5527raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
5528{
5529 struct r5conf *conf = mddev->private;
5530 unsigned long new;
5531 int err;
60aaf933 5532 struct r5worker_group *new_groups, *old_groups;
5533 int group_cnt, worker_cnt_per_group;
b721420e
SL		 5534
5535 if (len >= PAGE_SIZE)
5536 return -EINVAL;
5537 if (!conf)
5538 return -ENODEV;
5539
5540 if (kstrtoul(page, 10, &new))
5541 return -EINVAL;
5542
5543 if (new == conf->worker_cnt_per_group)
5544 return len;
5545
5546 mddev_suspend(mddev);
5547
5548 old_groups = conf->worker_groups;
d206dcfa 5549 if (old_groups)
5550 flush_workqueue(raid5_wq);
5551
60aaf933 5552 err = alloc_thread_groups(conf, new,
5553 &group_cnt, &worker_cnt_per_group,
5554 &new_groups);
5555 if (!err) {
5556 spin_lock_irq(&conf->device_lock);
5557 conf->group_cnt = group_cnt;
5558 conf->worker_cnt_per_group = worker_cnt_per_group;
5559 conf->worker_groups = new_groups;
5560 spin_unlock_irq(&conf->device_lock);
b721420e 5561
b721420e
SL		 5562 if (old_groups)
5563 kfree(old_groups[0].workers);
5564 kfree(old_groups);
5565 }
5566
5567 mddev_resume(mddev);
5568
5569 if (err)
5570 return err;
5571 return len;
5572}
5573
5574static struct md_sysfs_entry
5575raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
5576 raid5_show_group_thread_cnt,
5577 raid5_store_group_thread_cnt);
5578
007583c9 5579static struct attribute *raid5_attrs[] = {
3f294f4f
N		 5580 &raid5_stripecache_size.attr,
5581 &raid5_stripecache_active.attr,
8b3e6cdc 5582 &raid5_preread_bypass_threshold.attr,
b721420e 5583 &raid5_group_thread_cnt.attr,
d592a996 5584 &raid5_skip_copy.attr,
3f294f4f
N		 5585 NULL,
5586};
007583c9
N		 5587static struct attribute_group raid5_attrs_group = {
5588 .name = NULL,
5589 .attrs = raid5_attrs,
3f294f4f
N		 5590};
5591
60aaf933 5592static int alloc_thread_groups(struct r5conf *conf, int cnt,
5593 int *group_cnt,
5594 int *worker_cnt_per_group,
5595 struct r5worker_group **worker_groups)
851c30c9 5596{
566c09c5 5597 int i, j, k;
851c30c9
SL		 5598 ssize_t size;
5599 struct r5worker *workers;
5600
60aaf933 5601 *worker_cnt_per_group = cnt;
851c30c9 5602 if (cnt == 0) {
60aaf933 5603 *group_cnt = 0;
5604 *worker_groups = NULL;
851c30c9
SL		 5605 return 0;
5606 }
60aaf933 5607 *group_cnt = num_possible_nodes();
851c30c9 5608 size = sizeof(struct r5worker) * cnt;
60aaf933 5609 workers = kzalloc(size * *group_cnt, GFP_NOIO);
5610 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
5611 *group_cnt, GFP_NOIO);
5612 if (!*worker_groups || !workers) {
851c30c9 5613 kfree(workers);
60aaf933 5614 kfree(*worker_groups);
851c30c9
SL		 5615 return -ENOMEM;
5616 }
5617
60aaf933 5618 for (i = 0; i < *group_cnt; i++) {
851c30c9
SL		 5619 struct r5worker_group *group;
5620
0c775d52 5621 group = &(*worker_groups)[i];
851c30c9
SL		 5622 INIT_LIST_HEAD(&group->handle_list);
5623 group->conf = conf;
5624 group->workers = workers + i * cnt;
5625
5626 for (j = 0; j < cnt; j++) {
566c09c5
SL		 5627 struct r5worker *worker = group->workers + j;
5628 worker->group = group;
5629 INIT_WORK(&worker->work, raid5_do_work);
5630
5631 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
5632 INIT_LIST_HEAD(worker->temp_inactive_list + k);
851c30c9
SL		 5633 }
5634 }
5635
5636 return 0;
5637}
5638
5639static void free_thread_groups(struct r5conf *conf)
5640{
5641 if (conf->worker_groups)
5642 kfree(conf->worker_groups[0].workers);
5643 kfree(conf->worker_groups);
5644 conf->worker_groups = NULL;
5645}
5646
80c3a6ce 5647static sector_t
fd01b88c 5648raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
80c3a6ce 5649{
d1688a6d 5650 struct r5conf *conf = mddev->private;
80c3a6ce
DW		 5651
5652 if (!sectors)
5653 sectors = mddev->dev_sectors;
5e5e3e78 5654 if (!raid_disks)
7ec05478 5655 /* size is defined by the smallest of previous and new size */
5e5e3e78 5656 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
80c3a6ce 5657
9d8f0363 5658 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
664e7c41 5659 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
80c3a6ce
DW		 5660 return sectors * (raid_disks - conf->max_degraded);
5661}
5662
789b5e03
ON		 5663static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5664{
5665 safe_put_page(percpu->spare_page);
5666 kfree(percpu->scribble);
5667 percpu->spare_page = NULL;
5668 percpu->scribble = NULL;
5669}
5670
5671static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5672{
5673 if (conf->level == 6 && !percpu->spare_page)
5674 percpu->spare_page = alloc_page(GFP_KERNEL);
5675 if (!percpu->scribble)
5676 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5677
5678 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
5679 free_scratch_buffer(conf, percpu);
5680 return -ENOMEM;
5681 }
5682
5683 return 0;
5684}
5685
d1688a6d 5686static void raid5_free_percpu(struct r5conf *conf)
36d1c647 5687{
36d1c647
DW		 5688 unsigned long cpu;
5689
5690 if (!conf->percpu)
5691 return;
5692
36d1c647
DW		 5693#ifdef CONFIG_HOTPLUG_CPU
5694 unregister_cpu_notifier(&conf->cpu_notify);
5695#endif
789b5e03
ON		 5696
5697 get_online_cpus();
5698 for_each_possible_cpu(cpu)
5699 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
36d1c647
DW		 5700 put_online_cpus();
5701
5702 free_percpu(conf->percpu);
5703}
5704
d1688a6d 5705static void free_conf(struct r5conf *conf)
95fc17aa 5706{
851c30c9 5707 free_thread_groups(conf);
95fc17aa 5708 shrink_stripes(conf);
36d1c647 5709 raid5_free_percpu(conf);
95fc17aa
DW		 5710 kfree(conf->disks);
5711 kfree(conf->stripe_hashtbl);
5712 kfree(conf);
5713}
5714
36d1c647
DW		 5715#ifdef CONFIG_HOTPLUG_CPU
5716static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5717 void *hcpu)
5718{
d1688a6d 5719 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
36d1c647
DW		 5720 long cpu = (long)hcpu;
5721 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5722
5723 switch (action) {
5724 case CPU_UP_PREPARE:
5725 case CPU_UP_PREPARE_FROZEN:
789b5e03 5726 if (alloc_scratch_buffer(conf, percpu)) {
36d1c647
DW		 5727 pr_err("%s: failed memory allocation for cpu%ld\n",
5728 __func__, cpu);
55af6bb5 5729 return notifier_from_errno(-ENOMEM);
36d1c647
DW		 5730 }
5731 break;
5732 case CPU_DEAD:
5733 case CPU_DEAD_FROZEN:
789b5e03 5734 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
36d1c647
DW		 5735 break;
5736 default:
5737 break;
5738 }
5739 return NOTIFY_OK;
5740}
5741#endif
5742
d1688a6d 5743static int raid5_alloc_percpu(struct r5conf *conf)
36d1c647
DW		 5744{
5745 unsigned long cpu;
789b5e03 5746 int err = 0;
36d1c647 5747
789b5e03
ON		 5748 conf->percpu = alloc_percpu(struct raid5_percpu);
5749 if (!conf->percpu)
36d1c647 5750 return -ENOMEM;
789b5e03
ON		 5751
5752#ifdef CONFIG_HOTPLUG_CPU
5753 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5754 conf->cpu_notify.priority = 0;
5755 err = register_cpu_notifier(&conf->cpu_notify);
5756 if (err)
5757 return err;
5758#endif
36d1c647
DW		 5759
5760 get_online_cpus();
36d1c647 5761 for_each_present_cpu(cpu) {
789b5e03
ON		 5762 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5763 if (err) {
5764 pr_err("%s: failed memory allocation for cpu%ld\n",
5765 __func__, cpu);
36d1c647
DW		 5766 break;
5767 }
36d1c647 5768 }
36d1c647
DW		 5769 put_online_cpus();
5770
5771 return err;
5772}
5773
d1688a6d 5774static struct r5conf *setup_conf(struct mddev *mddev)
1da177e4 5775{
d1688a6d 5776 struct r5conf *conf;
5e5e3e78 5777 int raid_disk, memory, max_disks;
3cb03002 5778 struct md_rdev *rdev;
1da177e4 5779 struct disk_info *disk;
0232605d 5780 char pers_name[6];
566c09c5 5781 int i;
60aaf933 5782 int group_cnt, worker_cnt_per_group;
5783 struct r5worker_group *new_group;
1da177e4 5784
91adb564
N		 5785 if (mddev->new_level != 5
5786 && mddev->new_level != 4
5787 && mddev->new_level != 6) {
0c55e022 5788 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
91adb564
N		 5789 mdname(mddev), mddev->new_level);
5790 return ERR_PTR(-EIO);
1da177e4 5791 }
91adb564
N		 5792 if ((mddev->new_level == 5
5793 && !algorithm_valid_raid5(mddev->new_layout)) ||
5794 (mddev->new_level == 6
5795 && !algorithm_valid_raid6(mddev->new_layout))) {
0c55e022 5796 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
91adb564
N		 5797 mdname(mddev), mddev->new_layout);
5798 return ERR_PTR(-EIO);
99c0fb5f 5799 }
91adb564 5800 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
0c55e022 5801 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
91adb564
N		 5802 mdname(mddev), mddev->raid_disks);
5803 return ERR_PTR(-EINVAL);
4bbf3771
N		 5804 }
5805
664e7c41
AN		 5806 if (!mddev->new_chunk_sectors ||
5807 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5808 !is_power_of_2(mddev->new_chunk_sectors)) {
0c55e022
N		 5809 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5810 mdname(mddev), mddev->new_chunk_sectors << 9);
91adb564 5811 return ERR_PTR(-EINVAL);
f6705578
N		 5812 }
5813
d1688a6d 5814 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
91adb564 5815 if (conf == NULL)
1da177e4 5816 goto abort;
851c30c9 5817 /* Don't enable multi-threading by default*/
60aaf933 5818 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
5819 &new_group)) {
5820 conf->group_cnt = group_cnt;
5821 conf->worker_cnt_per_group = worker_cnt_per_group;
5822 conf->worker_groups = new_group;
5823 } else
851c30c9 5824 goto abort;
f5efd45a 5825 spin_lock_init(&conf->device_lock);
c46501b2 5826 seqcount_init(&conf->gen_lock);
f5efd45a
DW		 5827 init_waitqueue_head(&conf->wait_for_stripe);
5828 init_waitqueue_head(&conf->wait_for_overlap);
5829 INIT_LIST_HEAD(&conf->handle_list);
5830 INIT_LIST_HEAD(&conf->hold_list);
5831 INIT_LIST_HEAD(&conf->delayed_list);
5832 INIT_LIST_HEAD(&conf->bitmap_list);
773ca82f 5833 init_llist_head(&conf->released_stripes);
f5efd45a
DW		 5834 atomic_set(&conf->active_stripes, 0);
5835 atomic_set(&conf->preread_active_stripes, 0);
5836 atomic_set(&conf->active_aligned_reads, 0);
5837 conf->bypass_threshold = BYPASS_THRESHOLD;
d890fa2b 5838 conf->recovery_disabled = mddev->recovery_disabled - 1;
91adb564
N		 5839
5840 conf->raid_disks = mddev->raid_disks;
5841 if (mddev->reshape_position == MaxSector)
5842 conf->previous_raid_disks = mddev->raid_disks;
5843 else
f6705578 5844 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5e5e3e78
N		 5845 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5846 conf->scribble_len = scribble_len(max_disks);
f6705578 5847
5e5e3e78 5848 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
b55e6bfc
N		 5849 GFP_KERNEL);
5850 if (!conf->disks)
5851 goto abort;
9ffae0cf 5852
1da177e4
LT		 5853 conf->mddev = mddev;
5854
fccddba0 5855 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
1da177e4 5856 goto abort;
1da177e4 5857
566c09c5
SL		 5858 /* We init hash_locks[0] separately to that it can be used
5859 * as the reference lock in the spin_lock_nest_lock() call
5860 * in lock_all_device_hash_locks_irq in order to convince
5861 * lockdep that we know what we are doing.
5862 */
5863 spin_lock_init(conf->hash_locks);
5864 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
5865 spin_lock_init(conf->hash_locks + i);
5866
5867 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5868 INIT_LIST_HEAD(conf->inactive_list + i);
5869
5870 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5871 INIT_LIST_HEAD(conf->temp_inactive_list + i);
5872
36d1c647
DW		 5873 conf->level = mddev->new_level;
5874 if (raid5_alloc_percpu(conf) != 0)
5875 goto abort;
5876
0c55e022 5877 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
1da177e4 5878
dafb20fa 5879 rdev_for_each(rdev, mddev) {
1da177e4 5880 raid_disk = rdev->raid_disk;
5e5e3e78 5881 if (raid_disk >= max_disks
1da177e4
LT		 5882 || raid_disk < 0)
5883 continue;
5884 disk = conf->disks + raid_disk;
5885
17045f52
N		 5886 if (test_bit(Replacement, &rdev->flags)) {
5887 if (disk->replacement)
5888 goto abort;
5889 disk->replacement = rdev;
5890 } else {
5891 if (disk->rdev)
5892 goto abort;
5893 disk->rdev = rdev;
5894 }
1da177e4 5895
b2d444d7 5896 if (test_bit(In_sync, &rdev->flags)) {
1da177e4 5897 char b[BDEVNAME_SIZE];
0c55e022
N		 5898 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5899 " disk %d\n",
5900 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
d6b212f4 5901 } else if (rdev->saved_raid_disk != raid_disk)
8c2e870a
NB		 5902 /* Cannot rely on bitmap to complete recovery */
5903 conf->fullsync = 1;
1da177e4
LT		 5904 }
5905
09c9e5fa 5906 conf->chunk_sectors = mddev->new_chunk_sectors;
91adb564 5907 conf->level = mddev->new_level;
16a53ecc
N		 5908 if (conf->level == 6)
5909 conf->max_degraded = 2;
5910 else
5911 conf->max_degraded = 1;
91adb564 5912 conf->algorithm = mddev->new_layout;
fef9c61f 5913 conf->reshape_progress = mddev->reshape_position;
e183eaed 5914 if (conf->reshape_progress != MaxSector) {
09c9e5fa 5915 conf->prev_chunk_sectors = mddev->chunk_sectors;
e183eaed
N		 5916 conf->prev_algo = mddev->layout;
5917 }
1da177e4 5918
91adb564 5919 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5e5e3e78 5920 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4bda556a 5921 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
566c09c5 5922 if (grow_stripes(conf, NR_STRIPES)) {
91adb564 5923 printk(KERN_ERR
0c55e022
N		 5924 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5925 mdname(mddev), memory);
91adb564
N		 5926 goto abort;
5927 } else
0c55e022
N		 5928 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5929 mdname(mddev), memory);
1da177e4 5930
0232605d
N		 5931 sprintf(pers_name, "raid%d", mddev->new_level);
5932 conf->thread = md_register_thread(raid5d, mddev, pers_name);
91adb564
N		 5933 if (!conf->thread) {
5934 printk(KERN_ERR
0c55e022 5935 "md/raid:%s: couldn't allocate thread.\n",
91adb564 5936 mdname(mddev));
16a53ecc
N		 5937 goto abort;
5938 }
91adb564
N		 5939
5940 return conf;
5941
5942 abort:
5943 if (conf) {
95fc17aa 5944 free_conf(conf);
91adb564
N		 5945 return ERR_PTR(-EIO);
5946 } else
5947 return ERR_PTR(-ENOMEM);
5948}
5949
c148ffdc
N		 5950static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5951{
5952 switch (algo) {
5953 case ALGORITHM_PARITY_0:
5954 if (raid_disk < max_degraded)
5955 return 1;
5956 break;
5957 case ALGORITHM_PARITY_N:
5958 if (raid_disk >= raid_disks - max_degraded)
5959 return 1;
5960 break;
5961 case ALGORITHM_PARITY_0_6:
f72ffdd6 5962 if (raid_disk == 0 ||
c148ffdc
N		 5963 raid_disk == raid_disks - 1)
5964 return 1;
5965 break;
5966 case ALGORITHM_LEFT_ASYMMETRIC_6:
5967 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5968 case ALGORITHM_LEFT_SYMMETRIC_6:
5969 case ALGORITHM_RIGHT_SYMMETRIC_6:
5970 if (raid_disk == raid_disks - 1)
5971 return 1;
5972 }
5973 return 0;
5974}
5975
fd01b88c 5976static int run(struct mddev *mddev)
91adb564 5977{
d1688a6d 5978 struct r5conf *conf;
9f7c2220 5979 int working_disks = 0;
c148ffdc 5980 int dirty_parity_disks = 0;
3cb03002 5981 struct md_rdev *rdev;
c148ffdc 5982 sector_t reshape_offset = 0;
17045f52 5983 int i;
b5254dd5
N		 5984 long long min_offset_diff = 0;
5985 int first = 1;
91adb564 5986
8c6ac868 5987 if (mddev->recovery_cp != MaxSector)
0c55e022 5988 printk(KERN_NOTICE "md/raid:%s: not clean"
8c6ac868
AN		 5989 " -- starting background reconstruction\n",
5990 mdname(mddev));
b5254dd5
N		 5991
5992 rdev_for_each(rdev, mddev) {
5993 long long diff;
5994 if (rdev->raid_disk < 0)
5995 continue;
5996 diff = (rdev->new_data_offset - rdev->data_offset);
5997 if (first) {
5998 min_offset_diff = diff;
5999 first = 0;
6000 } else if (mddev->reshape_backwards &&
6001 diff < min_offset_diff)
6002 min_offset_diff = diff;
6003 else if (!mddev->reshape_backwards &&
6004 diff > min_offset_diff)
6005 min_offset_diff = diff;
6006 }
6007
91adb564
N		 6008 if (mddev->reshape_position != MaxSector) {
6009 /* Check that we can continue the reshape.
b5254dd5
N		 6010 * Difficulties arise if the stripe we would write to
6011 * next is at or after the stripe we would read from next.
6012 * For a reshape that changes the number of devices, this
6013 * is only possible for a very short time, and mdadm makes
6014 * sure that time appears to have past before assembling
6015 * the array. So we fail if that time hasn't passed.
6016 * For a reshape that keeps the number of devices the same
6017 * mdadm must be monitoring the reshape can keeping the
6018 * critical areas read-only and backed up. It will start
6019 * the array in read-only mode, so we check for that.
91adb564
N		 6020 */
6021 sector_t here_new, here_old;
6022 int old_disks;
18b00334 6023 int max_degraded = (mddev->level == 6 ? 2 : 1);
91adb564 6024
88ce4930 6025 if (mddev->new_level != mddev->level) {
0c55e022 6026 printk(KERN_ERR "md/raid:%s: unsupported reshape "
91adb564
N		 6027 "required - aborting.\n",
6028 mdname(mddev));
6029 return -EINVAL;
6030 }
91adb564
N		 6031 old_disks = mddev->raid_disks - mddev->delta_disks;
6032 /* reshape_position must be on a new-stripe boundary, and one
6033 * further up in new geometry must map after here in old
6034 * geometry.
6035 */
6036 here_new = mddev->reshape_position;
664e7c41 6037 if (sector_div(here_new, mddev->new_chunk_sectors *
91adb564 6038 (mddev->raid_disks - max_degraded))) {
0c55e022
N		 6039 printk(KERN_ERR "md/raid:%s: reshape_position not "
6040 "on a stripe boundary\n", mdname(mddev));
91adb564
N		 6041 return -EINVAL;
6042 }
c148ffdc 6043 reshape_offset = here_new * mddev->new_chunk_sectors;
91adb564
N		 6044 /* here_new is the stripe we will write to */
6045 here_old = mddev->reshape_position;
9d8f0363 6046 sector_div(here_old, mddev->chunk_sectors *
91adb564
N		 6047 (old_disks-max_degraded));
6048 /* here_old is the first stripe that we might need to read
6049 * from */
67ac6011 6050 if (mddev->delta_disks == 0) {
b5254dd5
N		 6051 if ((here_new * mddev->new_chunk_sectors !=
6052 here_old * mddev->chunk_sectors)) {
6053 printk(KERN_ERR "md/raid:%s: reshape position is"
6054 " confused - aborting\n", mdname(mddev));
6055 return -EINVAL;
6056 }
67ac6011 6057 /* We cannot be sure it is safe to start an in-place
b5254dd5 6058 * reshape. It is only safe if user-space is monitoring
67ac6011
N		 6059 * and taking constant backups.
6060 * mdadm always starts a situation like this in
6061 * readonly mode so it can take control before
6062 * allowing any writes. So just check for that.
6063 */
b5254dd5
N		 6064 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6065 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6066 /* not really in-place - so OK */;
6067 else if (mddev->ro == 0) {
6068 printk(KERN_ERR "md/raid:%s: in-place reshape "
6069 "must be started in read-only mode "
6070 "- aborting\n",
0c55e022 6071 mdname(mddev));
67ac6011
N		 6072 return -EINVAL;
6073 }
2c810cdd 6074 } else if (mddev->reshape_backwards
b5254dd5 6075 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
67ac6011
N		 6076 here_old * mddev->chunk_sectors)
6077 : (here_new * mddev->new_chunk_sectors >=
b5254dd5 6078 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
91adb564 6079 /* Reading from the same stripe as writing to - bad */
0c55e022
N		 6080 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6081 "auto-recovery - aborting.\n",
6082 mdname(mddev));
91adb564
N		 6083 return -EINVAL;
6084 }
0c55e022
N		 6085 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6086 mdname(mddev));
91adb564
N		 6087 /* OK, we should be able to continue; */
6088 } else {
6089 BUG_ON(mddev->level != mddev->new_level);
6090 BUG_ON(mddev->layout != mddev->new_layout);
664e7c41 6091 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
91adb564 6092 BUG_ON(mddev->delta_disks != 0);
1da177e4 6093 }
91adb564 6094
245f46c2
N		 6095 if (mddev->private == NULL)
6096 conf = setup_conf(mddev);
6097 else
6098 conf = mddev->private;
6099
91adb564
N		 6100 if (IS_ERR(conf))
6101 return PTR_ERR(conf);
6102
b5254dd5 6103 conf->min_offset_diff = min_offset_diff;
91adb564
N		 6104 mddev->thread = conf->thread;
6105 conf->thread = NULL;
6106 mddev->private = conf;
6107
17045f52
N		 6108 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6109 i++) {
6110 rdev = conf->disks[i].rdev;
6111 if (!rdev && conf->disks[i].replacement) {
6112 /* The replacement is all we have yet */
6113 rdev = conf->disks[i].replacement;
6114 conf->disks[i].replacement = NULL;
6115 clear_bit(Replacement, &rdev->flags);
6116 conf->disks[i].rdev = rdev;
6117 }
6118 if (!rdev)
c148ffdc 6119 continue;
17045f52
N		 6120 if (conf->disks[i].replacement &&
6121 conf->reshape_progress != MaxSector) {
6122 /* replacements and reshape simply do not mix. */
6123 printk(KERN_ERR "md: cannot handle concurrent "
6124 "replacement and reshape.\n");
6125 goto abort;
6126 }
2f115882 6127 if (test_bit(In_sync, &rdev->flags)) {
91adb564 6128 working_disks++;
2f115882
N		 6129 continue;
6130 }
c148ffdc
N		 6131 /* This disc is not fully in-sync. However if it
6132 * just stored parity (beyond the recovery_offset),
6133 * when we don't need to be concerned about the
6134 * array being dirty.
6135 * When reshape goes 'backwards', we never have
6136 * partially completed devices, so we only need
6137 * to worry about reshape going forwards.
6138 */
6139 /* Hack because v0.91 doesn't store recovery_offset properly. */
6140 if (mddev->major_version == 0 &&
6141 mddev->minor_version > 90)
6142 rdev->recovery_offset = reshape_offset;
5026d7a9 6143
c148ffdc
N		 6144 if (rdev->recovery_offset < reshape_offset) {
6145 /* We need to check old and new layout */
6146 if (!only_parity(rdev->raid_disk,
6147 conf->algorithm,
6148 conf->raid_disks,
6149 conf->max_degraded))
6150 continue;
6151 }
6152 if (!only_parity(rdev->raid_disk,
6153 conf->prev_algo,
6154 conf->previous_raid_disks,
6155 conf->max_degraded))
6156 continue;
6157 dirty_parity_disks++;
6158 }
91adb564 6159
17045f52
N		 6160 /*
6161 * 0 for a fully functional array, 1 or 2 for a degraded array.
6162 */
908f4fbd 6163 mddev->degraded = calc_degraded(conf);
91adb564 6164
674806d6 6165 if (has_failed(conf)) {
0c55e022 6166 printk(KERN_ERR "md/raid:%s: not enough operational devices"
1da177e4 6167 " (%d/%d failed)\n",
02c2de8c 6168 mdname(mddev), mddev->degraded, conf->raid_disks);
1da177e4
LT		 6169 goto abort;
6170 }
6171
91adb564 6172 /* device size must be a multiple of chunk size */
9d8f0363 6173 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
91adb564
N		 6174 mddev->resync_max_sectors = mddev->dev_sectors;
6175
c148ffdc 6176 if (mddev->degraded > dirty_parity_disks &&
1da177e4 6177 mddev->recovery_cp != MaxSector) {
6ff8d8ec
N		 6178 if (mddev->ok_start_degraded)
6179 printk(KERN_WARNING
0c55e022
N		 6180 "md/raid:%s: starting dirty degraded array"
6181 " - data corruption possible.\n",
6ff8d8ec
N		 6182 mdname(mddev));
6183 else {
6184 printk(KERN_ERR
0c55e022 6185 "md/raid:%s: cannot start dirty degraded array.\n",
6ff8d8ec
N		 6186 mdname(mddev));
6187 goto abort;
6188 }
1da177e4
LT		 6189 }
6190
1da177e4 6191 if (mddev->degraded == 0)
0c55e022
N		 6192 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6193 " devices, algorithm %d\n", mdname(mddev), conf->level,
e183eaed
N		 6194 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6195 mddev->new_layout);
1da177e4 6196 else
0c55e022
N		 6197 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6198 " out of %d devices, algorithm %d\n",
6199 mdname(mddev), conf->level,
6200 mddev->raid_disks - mddev->degraded,
6201 mddev->raid_disks, mddev->new_layout);
1da177e4
LT		 6202
6203 print_raid5_conf(conf);
6204
fef9c61f 6205 if (conf->reshape_progress != MaxSector) {
fef9c61f 6206 conf->reshape_safe = conf->reshape_progress;
f6705578
N		 6207 atomic_set(&conf->reshape_stripes, 0);
6208 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6209 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6210 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6211 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6212 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
0da3c619 6213 "reshape");
f6705578
N		 6214 }
6215
1da177e4 6216 /* Ok, everything is just fine now */
a64c876f
N		 6217 if (mddev->to_remove == &raid5_attrs_group)
6218 mddev->to_remove = NULL;
00bcb4ac
N		 6219 else if (mddev->kobj.sd &&
6220 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5e55e2f5 6221 printk(KERN_WARNING
4a5add49 6222 "raid5: failed to create sysfs attributes for %s\n",
5e55e2f5 6223 mdname(mddev));
4a5add49 6224 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7a5febe9 6225
4a5add49 6226 if (mddev->queue) {
9f7c2220 6227 int chunk_size;
620125f2 6228 bool discard_supported = true;
4a5add49
N		 6229 /* read-ahead size must cover two whole stripes, which
6230 * is 2 * (datadisks) * chunksize where 'n' is the
6231 * number of raid devices
6232 */
6233 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6234 int stripe = data_disks *
6235 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6236 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6237 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
91adb564 6238
9f7c2220
N		 6239 chunk_size = mddev->chunk_sectors << 9;
6240 blk_queue_io_min(mddev->queue, chunk_size);
6241 blk_queue_io_opt(mddev->queue, chunk_size *
6242 (conf->raid_disks - conf->max_degraded));
c78afc62 6243 mddev->queue->limits.raid_partial_stripes_expensive = 1;
620125f2
SL		 6244 /*
6245 * We can only discard a whole stripe. It doesn't make sense to
6246 * discard data disk but write parity disk
6247 */
6248 stripe = stripe * PAGE_SIZE;
4ac6875e
N		 6249 /* Round up to power of 2, as discard handling
6250 * currently assumes that */
6251 while ((stripe-1) & stripe)
6252 stripe = (stripe | (stripe-1)) + 1;
620125f2
SL		 6253 mddev->queue->limits.discard_alignment = stripe;
6254 mddev->queue->limits.discard_granularity = stripe;
6255 /*
6256 * unaligned part of discard request will be ignored, so can't
8e0e99ba 6257 * guarantee discard_zeroes_data
620125f2
SL		 6258 */
6259 mddev->queue->limits.discard_zeroes_data = 0;
8f6c2e4b 6260
5026d7a9
PA		 6261 blk_queue_max_write_same_sectors(mddev->queue, 0);
6262
05616be5 6263 rdev_for_each(rdev, mddev) {
9f7c2220
N		 6264 disk_stack_limits(mddev->gendisk, rdev->bdev,
6265 rdev->data_offset << 9);
05616be5
N		 6266 disk_stack_limits(mddev->gendisk, rdev->bdev,
6267 rdev->new_data_offset << 9);
620125f2
SL		 6268 /*
6269 * discard_zeroes_data is required, otherwise data
6270 * could be lost. Consider a scenario: discard a stripe
6271 * (the stripe could be inconsistent if
6272 * discard_zeroes_data is 0); write one disk of the
6273 * stripe (the stripe could be inconsistent again
6274 * depending on which disks are used to calculate
6275 * parity); the disk is broken; The stripe data of this
6276 * disk is lost.
6277 */
6278 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6279 !bdev_get_queue(rdev->bdev)->
6280 limits.discard_zeroes_data)
6281 discard_supported = false;
8e0e99ba
N		 6282 /* Unfortunately, discard_zeroes_data is not currently
6283 * a guarantee - just a hint. So we only allow DISCARD
6284 * if the sysadmin has confirmed that only safe devices
6285 * are in use by setting a module parameter.
6286 */
6287 if (!devices_handle_discard_safely) {
6288 if (discard_supported) {
6289 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6290 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6291 }
6292 discard_supported = false;
6293 }
05616be5 6294 }
620125f2
SL		 6295
6296 if (discard_supported &&
6297 mddev->queue->limits.max_discard_sectors >= stripe &&
6298 mddev->queue->limits.discard_granularity >= stripe)
6299 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6300 mddev->queue);
6301 else
6302 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6303 mddev->queue);
9f7c2220 6304 }
23032a0e 6305
1da177e4
LT		 6306 return 0;
6307abort:
01f96c0a 6308 md_unregister_thread(&mddev->thread);
e4f869d9
N		 6309 print_raid5_conf(conf);
6310 free_conf(conf);
1da177e4 6311 mddev->private = NULL;
0c55e022 6312 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
1da177e4
LT		 6313 return -EIO;
6314}
6315
fd01b88c 6316static int stop(struct mddev *mddev)
1da177e4 6317{
d1688a6d 6318 struct r5conf *conf = mddev->private;
1da177e4 6319
95fc17aa 6320 free_conf(conf);
a64c876f
N		 6321 mddev->private = NULL;
6322 mddev->to_remove = &raid5_attrs_group;
1da177e4
LT		 6323 return 0;
6324}
6325
fd01b88c 6326static void status(struct seq_file *seq, struct mddev *mddev)
1da177e4 6327{
d1688a6d 6328 struct r5conf *conf = mddev->private;
1da177e4
LT		 6329 int i;
6330
9d8f0363
AN		 6331 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6332 mddev->chunk_sectors / 2, mddev->layout);
02c2de8c 6333 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
1da177e4
LT		 6334 for (i = 0; i < conf->raid_disks; i++)
6335 seq_printf (seq, "%s",
6336 conf->disks[i].rdev &&
b2d444d7 6337 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
1da177e4 6338 seq_printf (seq, "]");
1da177e4
LT		 6339}
6340
d1688a6d 6341static void print_raid5_conf (struct r5conf *conf)
1da177e4
LT		 6342{
6343 int i;
6344 struct disk_info *tmp;
6345
0c55e022 6346 printk(KERN_DEBUG "RAID conf printout:\n");
1da177e4
LT		 6347 if (!conf) {
6348 printk("(conf==NULL)\n");
6349 return;
6350 }
0c55e022
N		 6351 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6352 conf->raid_disks,
6353 conf->raid_disks - conf->mddev->degraded);
1da177e4
LT		 6354
6355 for (i = 0; i < conf->raid_disks; i++) {
6356 char b[BDEVNAME_SIZE];
6357 tmp = conf->disks + i;
6358 if (tmp->rdev)
0c55e022
N		 6359 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6360 i, !test_bit(Faulty, &tmp->rdev->flags),
6361 bdevname(tmp->rdev->bdev, b));
1da177e4
LT		 6362 }
6363}
6364
fd01b88c 6365static int raid5_spare_active(struct mddev *mddev)
1da177e4
LT		 6366{
6367 int i;
d1688a6d 6368 struct r5conf *conf = mddev->private;
1da177e4 6369 struct disk_info *tmp;
6b965620
N		 6370 int count = 0;
6371 unsigned long flags;
1da177e4
LT		 6372
6373 for (i = 0; i < conf->raid_disks; i++) {
6374 tmp = conf->disks + i;
dd054fce
N		 6375 if (tmp->replacement
6376 && tmp->replacement->recovery_offset == MaxSector
6377 && !test_bit(Faulty, &tmp->replacement->flags)
6378 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
6379 /* Replacement has just become active. */
6380 if (!tmp->rdev
6381 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
6382 count++;
6383 if (tmp->rdev) {
6384 /* Replaced device not technically faulty,
6385 * but we need to be sure it gets removed
6386 * and never re-added.
6387 */
6388 set_bit(Faulty, &tmp->rdev->flags);
6389 sysfs_notify_dirent_safe(
6390 tmp->rdev->sysfs_state);
6391 }
6392 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
6393 } else if (tmp->rdev
70fffd0b 6394 && tmp->rdev->recovery_offset == MaxSector
b2d444d7 6395 && !test_bit(Faulty, &tmp->rdev->flags)
c04be0aa 6396 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
6b965620 6397 count++;
43c73ca4 6398 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1da177e4
LT		 6399 }
6400 }
6b965620 6401 spin_lock_irqsave(&conf->device_lock, flags);
908f4fbd 6402 mddev->degraded = calc_degraded(conf);
6b965620 6403 spin_unlock_irqrestore(&conf->device_lock, flags);
1da177e4 6404 print_raid5_conf(conf);
6b965620 6405 return count;
1da177e4
LT		 6406}
6407
b8321b68 6408static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1da177e4 6409{
d1688a6d 6410 struct r5conf *conf = mddev->private;
1da177e4 6411 int err = 0;
b8321b68 6412 int number = rdev->raid_disk;
657e3e4d 6413 struct md_rdev **rdevp;
1da177e4
LT		 6414 struct disk_info *p = conf->disks + number;
6415
6416 print_raid5_conf(conf);
657e3e4d
N		 6417 if (rdev == p->rdev)
6418 rdevp = &p->rdev;
6419 else if (rdev == p->replacement)
6420 rdevp = &p->replacement;
6421 else
6422 return 0;
6423
6424 if (number >= conf->raid_disks &&
6425 conf->reshape_progress == MaxSector)
6426 clear_bit(In_sync, &rdev->flags);
6427
6428 if (test_bit(In_sync, &rdev->flags) ||
6429 atomic_read(&rdev->nr_pending)) {
6430 err = -EBUSY;
6431 goto abort;
6432 }
6433 /* Only remove non-faulty devices if recovery
6434 * isn't possible.
6435 */
6436 if (!test_bit(Faulty, &rdev->flags) &&
6437 mddev->recovery_disabled != conf->recovery_disabled &&
6438 !has_failed(conf) &&
dd054fce 6439 (!p->replacement || p->replacement == rdev) &&
657e3e4d
N		 6440 number < conf->raid_disks) {
6441 err = -EBUSY;
6442 goto abort;
6443 }
6444 *rdevp = NULL;
6445 synchronize_rcu();
6446 if (atomic_read(&rdev->nr_pending)) {
6447 /* lost the race, try later */
6448 err = -EBUSY;
6449 *rdevp = rdev;
dd054fce
N		 6450 } else if (p->replacement) {
6451 /* We must have just cleared 'rdev' */
6452 p->rdev = p->replacement;
6453 clear_bit(Replacement, &p->replacement->flags);
6454 smp_mb(); /* Make sure other CPUs may see both as identical
6455 * but will never see neither - if they are careful
6456 */
6457 p->replacement = NULL;
6458 clear_bit(WantReplacement, &rdev->flags);
6459 } else
6460 /* We might have just removed the Replacement as faulty-
6461 * clear the bit just in case
6462 */
6463 clear_bit(WantReplacement, &rdev->flags);
1da177e4
LT		 6464abort:
6465
6466 print_raid5_conf(conf);
6467 return err;
6468}
6469
fd01b88c 6470static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1da177e4 6471{
d1688a6d 6472 struct r5conf *conf = mddev->private;
199050ea 6473 int err = -EEXIST;
1da177e4
LT		 6474 int disk;
6475 struct disk_info *p;
6c2fce2e
NB		 6476 int first = 0;
6477 int last = conf->raid_disks - 1;
1da177e4 6478
7f0da59b
N		 6479 if (mddev->recovery_disabled == conf->recovery_disabled)
6480 return -EBUSY;
6481
dc10c643 6482 if (rdev->saved_raid_disk < 0 && has_failed(conf))
1da177e4 6483 /* no point adding a device */
199050ea 6484 return -EINVAL;
1da177e4 6485
6c2fce2e
NB		 6486 if (rdev->raid_disk >= 0)
6487 first = last = rdev->raid_disk;
1da177e4
LT		 6488
6489 /*
16a53ecc
N		 6490 * find the disk ... but prefer rdev->saved_raid_disk
6491 * if possible.
1da177e4 6492 */
16a53ecc 6493 if (rdev->saved_raid_disk >= 0 &&
6c2fce2e 6494 rdev->saved_raid_disk >= first &&
16a53ecc 6495 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5cfb22a1
N		 6496 first = rdev->saved_raid_disk;
6497
6498 for (disk = first; disk <= last; disk++) {
7bfec5f3
N		 6499 p = conf->disks + disk;
6500 if (p->rdev == NULL) {
b2d444d7 6501 clear_bit(In_sync, &rdev->flags);
1da177e4 6502 rdev->raid_disk = disk;
199050ea 6503 err = 0;
72626685
N		 6504 if (rdev->saved_raid_disk != disk)
6505 conf->fullsync = 1;
d6065f7b 6506 rcu_assign_pointer(p->rdev, rdev);
5cfb22a1 6507 goto out;
1da177e4 6508 }
5cfb22a1
N		 6509 }
6510 for (disk = first; disk <= last; disk++) {
6511 p = conf->disks + disk;
7bfec5f3
N		 6512 if (test_bit(WantReplacement, &p->rdev->flags) &&
6513 p->replacement == NULL) {
6514 clear_bit(In_sync, &rdev->flags);
6515 set_bit(Replacement, &rdev->flags);
6516 rdev->raid_disk = disk;
6517 err = 0;
6518 conf->fullsync = 1;
6519 rcu_assign_pointer(p->replacement, rdev);
6520 break;
6521 }
6522 }
5cfb22a1 6523out:
1da177e4 6524 print_raid5_conf(conf);
199050ea 6525 return err;
1da177e4
LT		 6526}
6527
fd01b88c 6528static int raid5_resize(struct mddev *mddev, sector_t sectors)
1da177e4
LT		 6529{
6530 /* no resync is happening, and there is enough space
6531 * on all devices, so we can resize.
6532 * We need to make sure resync covers any new space.
6533 * If the array is shrinking we should possibly wait until
6534 * any io in the removed space completes, but it hardly seems
6535 * worth it.
6536 */
a4a6125a 6537 sector_t newsize;
9d8f0363 6538 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
a4a6125a
N		 6539 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
6540 if (mddev->external_size &&
6541 mddev->array_sectors > newsize)
b522adcd 6542 return -EINVAL;
a4a6125a
N		 6543 if (mddev->bitmap) {
6544 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
6545 if (ret)
6546 return ret;
6547 }
6548 md_set_array_sectors(mddev, newsize);
f233ea5c 6549 set_capacity(mddev->gendisk, mddev->array_sectors);
449aad3e 6550 revalidate_disk(mddev->gendisk);
b098636c
N		 6551 if (sectors > mddev->dev_sectors &&
6552 mddev->recovery_cp > mddev->dev_sectors) {
58c0fed4 6553 mddev->recovery_cp = mddev->dev_sectors;
1da177e4
LT		 6554 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
6555 }
58c0fed4 6556 mddev->dev_sectors = sectors;
4b5c7ae8 6557 mddev->resync_max_sectors = sectors;
1da177e4
LT		 6558 return 0;
6559}
6560
fd01b88c 6561static int check_stripe_cache(struct mddev *mddev)
01ee22b4
N		 6562{
6563 /* Can only proceed if there are plenty of stripe_heads.
6564 * We need a minimum of one full stripe,, and for sensible progress
6565 * it is best to have about 4 times that.
6566 * If we require 4 times, then the default 256 4K stripe_heads will
6567 * allow for chunk sizes up to 256K, which is probably OK.
6568 * If the chunk size is greater, user-space should request more
6569 * stripe_heads first.
6570 */
d1688a6d 6571 struct r5conf *conf = mddev->private;
01ee22b4
N		 6572 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
6573 > conf->max_nr_stripes ||
6574 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
6575 > conf->max_nr_stripes) {
0c55e022
N		 6576 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
6577 mdname(mddev),
01ee22b4
N		 6578 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
6579 / STRIPE_SIZE)*4);
6580 return 0;
6581 }
6582 return 1;
6583}
6584
fd01b88c 6585static int check_reshape(struct mddev *mddev)
29269553 6586{
d1688a6d 6587 struct r5conf *conf = mddev->private;
29269553 6588
88ce4930
N		 6589 if (mddev->delta_disks == 0 &&
6590 mddev->new_layout == mddev->layout &&
664e7c41 6591 mddev->new_chunk_sectors == mddev->chunk_sectors)
50ac168a 6592 return 0; /* nothing to do */
674806d6 6593 if (has_failed(conf))
ec32a2bd 6594 return -EINVAL;
fdcfbbb6 6595 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
ec32a2bd
N		 6596 /* We might be able to shrink, but the devices must
6597 * be made bigger first.
6598 * For raid6, 4 is the minimum size.
6599 * Otherwise 2 is the minimum
6600 */
6601 int min = 2;
6602 if (mddev->level == 6)
6603 min = 4;
6604 if (mddev->raid_disks + mddev->delta_disks < min)
6605 return -EINVAL;
6606 }
29269553 6607
01ee22b4 6608 if (!check_stripe_cache(mddev))
29269553 6609 return -ENOSPC;
29269553 6610
e56108d6
N		 6611 return resize_stripes(conf, (conf->previous_raid_disks
6612 + mddev->delta_disks));
63c70c4f
N		 6613}
6614
fd01b88c 6615static int raid5_start_reshape(struct mddev *mddev)
63c70c4f 6616{
d1688a6d 6617 struct r5conf *conf = mddev->private;
3cb03002 6618 struct md_rdev *rdev;
63c70c4f 6619 int spares = 0;
c04be0aa 6620 unsigned long flags;
63c70c4f 6621
f416885e 6622 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
63c70c4f
N		 6623 return -EBUSY;
6624
01ee22b4
N		 6625 if (!check_stripe_cache(mddev))
6626 return -ENOSPC;
6627
30b67645
N		 6628 if (has_failed(conf))
6629 return -EINVAL;
6630
c6563a8c 6631 rdev_for_each(rdev, mddev) {
469518a3
N		 6632 if (!test_bit(In_sync, &rdev->flags)
6633 && !test_bit(Faulty, &rdev->flags))
29269553 6634 spares++;
c6563a8c 6635 }
63c70c4f 6636
f416885e 6637 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
29269553
N		 6638 /* Not enough devices even to make a degraded array
6639 * of that size
6640 */
6641 return -EINVAL;
6642
ec32a2bd
N		 6643 /* Refuse to reduce size of the array. Any reductions in
6644 * array size must be through explicit setting of array_size
6645 * attribute.
6646 */
6647 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
6648 < mddev->array_sectors) {
0c55e022 6649 printk(KERN_ERR "md/raid:%s: array size must be reduced "
ec32a2bd
N		 6650 "before number of disks\n", mdname(mddev));
6651 return -EINVAL;
6652 }
6653
f6705578 6654 atomic_set(&conf->reshape_stripes, 0);
29269553 6655 spin_lock_irq(&conf->device_lock);
c46501b2 6656 write_seqcount_begin(&conf->gen_lock);
29269553 6657 conf->previous_raid_disks = conf->raid_disks;
63c70c4f 6658 conf->raid_disks += mddev->delta_disks;
09c9e5fa
AN		 6659 conf->prev_chunk_sectors = conf->chunk_sectors;
6660 conf->chunk_sectors = mddev->new_chunk_sectors;
88ce4930
N		 6661 conf->prev_algo = conf->algorithm;
6662 conf->algorithm = mddev->new_layout;
05616be5
N		 6663 conf->generation++;
6664 /* Code that selects data_offset needs to see the generation update
6665 * if reshape_progress has been set - so a memory barrier needed.
6666 */
6667 smp_mb();
2c810cdd 6668 if (mddev->reshape_backwards)
fef9c61f
N		 6669 conf->reshape_progress = raid5_size(mddev, 0, 0);
6670 else
6671 conf->reshape_progress = 0;
6672 conf->reshape_safe = conf->reshape_progress;
c46501b2 6673 write_seqcount_end(&conf->gen_lock);
29269553
N		 6674 spin_unlock_irq(&conf->device_lock);
6675
4d77e3ba
N		 6676 /* Now make sure any requests that proceeded on the assumption
6677 * the reshape wasn't running - like Discard or Read - have
6678 * completed.
6679 */
6680 mddev_suspend(mddev);
6681 mddev_resume(mddev);
6682
29269553
N		 6683 /* Add some new drives, as many as will fit.
6684 * We know there are enough to make the newly sized array work.
3424bf6a
N		 6685 * Don't add devices if we are reducing the number of
6686 * devices in the array. This is because it is not possible
6687 * to correctly record the "partially reconstructed" state of
6688 * such devices during the reshape and confusion could result.
29269553 6689 */
87a8dec9 6690 if (mddev->delta_disks >= 0) {
dafb20fa 6691 rdev_for_each(rdev, mddev)
87a8dec9
N		 6692 if (rdev->raid_disk < 0 &&
6693 !test_bit(Faulty, &rdev->flags)) {
6694 if (raid5_add_disk(mddev, rdev) == 0) {
87a8dec9 6695 if (rdev->raid_disk
9d4c7d87 6696 >= conf->previous_raid_disks)
87a8dec9 6697 set_bit(In_sync, &rdev->flags);
9d4c7d87 6698 else
87a8dec9 6699 rdev->recovery_offset = 0;
36fad858
NK		 6700
6701 if (sysfs_link_rdev(mddev, rdev))
87a8dec9 6702 /* Failure here is OK */;
50da0840 6703 }
87a8dec9
N		 6704 } else if (rdev->raid_disk >= conf->previous_raid_disks
6705 && !test_bit(Faulty, &rdev->flags)) {
6706 /* This is a spare that was manually added */
6707 set_bit(In_sync, &rdev->flags);
87a8dec9 6708 }
29269553 6709
87a8dec9
N		 6710 /* When a reshape changes the number of devices,
6711 * ->degraded is measured against the larger of the
6712 * pre and post number of devices.
6713 */
ec32a2bd 6714 spin_lock_irqsave(&conf->device_lock, flags);
908f4fbd 6715 mddev->degraded = calc_degraded(conf);
ec32a2bd
N		 6716 spin_unlock_irqrestore(&conf->device_lock, flags);
6717 }
63c70c4f 6718 mddev->raid_disks = conf->raid_disks;
e516402c 6719 mddev->reshape_position = conf->reshape_progress;
850b2b42 6720 set_bit(MD_CHANGE_DEVS, &mddev->flags);
f6705578 6721
29269553
N		 6722 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6723 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6724 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6725 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6726 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
0da3c619 6727 "reshape");
29269553
N		 6728 if (!mddev->sync_thread) {
6729 mddev->recovery = 0;
6730 spin_lock_irq(&conf->device_lock);
ba8805b9 6731 write_seqcount_begin(&conf->gen_lock);
29269553 6732 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
ba8805b9
N		 6733 mddev->new_chunk_sectors =
6734 conf->chunk_sectors = conf->prev_chunk_sectors;
6735 mddev->new_layout = conf->algorithm = conf->prev_algo;
05616be5
N		 6736 rdev_for_each(rdev, mddev)
6737 rdev->new_data_offset = rdev->data_offset;
6738 smp_wmb();
ba8805b9 6739 conf->generation --;
fef9c61f 6740 conf->reshape_progress = MaxSector;
1e3fa9bd 6741 mddev->reshape_position = MaxSector;
ba8805b9 6742 write_seqcount_end(&conf->gen_lock);
29269553
N		 6743 spin_unlock_irq(&conf->device_lock);
6744 return -EAGAIN;
6745 }
c8f517c4 6746 conf->reshape_checkpoint = jiffies;
29269553
N		 6747 md_wakeup_thread(mddev->sync_thread);
6748 md_new_event(mddev);
6749 return 0;
6750}
29269553 6751
ec32a2bd
N		 6752/* This is called from the reshape thread and should make any
6753 * changes needed in 'conf'
6754 */
d1688a6d 6755static void end_reshape(struct r5conf *conf)
29269553 6756{
29269553 6757
f6705578 6758 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
05616be5 6759 struct md_rdev *rdev;
f6705578 6760
f6705578 6761 spin_lock_irq(&conf->device_lock);
cea9c228 6762 conf->previous_raid_disks = conf->raid_disks;
05616be5
N		 6763 rdev_for_each(rdev, conf->mddev)
6764 rdev->data_offset = rdev->new_data_offset;
6765 smp_wmb();
fef9c61f 6766 conf->reshape_progress = MaxSector;
f6705578 6767 spin_unlock_irq(&conf->device_lock);
b0f9ec04 6768 wake_up(&conf->wait_for_overlap);
16a53ecc
N		 6769
6770 /* read-ahead size must cover two whole stripes, which is
6771 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6772 */
4a5add49 6773 if (conf->mddev->queue) {
cea9c228 6774 int data_disks = conf->raid_disks - conf->max_degraded;
09c9e5fa 6775 int stripe = data_disks * ((conf->chunk_sectors << 9)
cea9c228 6776 / PAGE_SIZE);
16a53ecc
N		 6777 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6778 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6779 }
29269553 6780 }
29269553
N		 6781}
6782
ec32a2bd
N		 6783/* This is called from the raid5d thread with mddev_lock held.
6784 * It makes config changes to the device.
6785 */
fd01b88c 6786static void raid5_finish_reshape(struct mddev *mddev)
cea9c228 6787{
d1688a6d 6788 struct r5conf *conf = mddev->private;
cea9c228
N		 6789
6790 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6791
ec32a2bd
N		 6792 if (mddev->delta_disks > 0) {
6793 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6794 set_capacity(mddev->gendisk, mddev->array_sectors);
449aad3e 6795 revalidate_disk(mddev->gendisk);
ec32a2bd
N		 6796 } else {
6797 int d;
908f4fbd
N		 6798 spin_lock_irq(&conf->device_lock);
6799 mddev->degraded = calc_degraded(conf);
6800 spin_unlock_irq(&conf->device_lock);
ec32a2bd
N		 6801 for (d = conf->raid_disks ;
6802 d < conf->raid_disks - mddev->delta_disks;
1a67dde0 6803 d++) {
3cb03002 6804 struct md_rdev *rdev = conf->disks[d].rdev;
da7613b8
N		 6805 if (rdev)
6806 clear_bit(In_sync, &rdev->flags);
6807 rdev = conf->disks[d].replacement;
6808 if (rdev)
6809 clear_bit(In_sync, &rdev->flags);
1a67dde0 6810 }
cea9c228 6811 }
88ce4930 6812 mddev->layout = conf->algorithm;
09c9e5fa 6813 mddev->chunk_sectors = conf->chunk_sectors;
ec32a2bd
N		 6814 mddev->reshape_position = MaxSector;
6815 mddev->delta_disks = 0;
2c810cdd 6816 mddev->reshape_backwards = 0;
cea9c228
N		 6817 }
6818}
6819
fd01b88c 6820static void raid5_quiesce(struct mddev *mddev, int state)
72626685 6821{
d1688a6d 6822 struct r5conf *conf = mddev->private;
72626685
N		 6823
6824 switch(state) {
e464eafd
N		 6825 case 2: /* resume for a suspend */
6826 wake_up(&conf->wait_for_overlap);
6827 break;
6828
72626685 6829 case 1: /* stop all writes */
566c09c5 6830 lock_all_device_hash_locks_irq(conf);
64bd660b
N		 6831 /* '2' tells resync/reshape to pause so that all
6832 * active stripes can drain
6833 */
6834 conf->quiesce = 2;
566c09c5 6835 wait_event_cmd(conf->wait_for_stripe,
46031f9a
RBJ		 6836 atomic_read(&conf->active_stripes) == 0 &&
6837 atomic_read(&conf->active_aligned_reads) == 0,
566c09c5
SL		 6838 unlock_all_device_hash_locks_irq(conf),
6839 lock_all_device_hash_locks_irq(conf));
64bd660b 6840 conf->quiesce = 1;
566c09c5 6841 unlock_all_device_hash_locks_irq(conf);
64bd660b
N		 6842 /* allow reshape to continue */
6843 wake_up(&conf->wait_for_overlap);
72626685
N		 6844 break;
6845
6846 case 0: /* re-enable writes */
566c09c5 6847 lock_all_device_hash_locks_irq(conf);
72626685
N		 6848 conf->quiesce = 0;
6849 wake_up(&conf->wait_for_stripe);
e464eafd 6850 wake_up(&conf->wait_for_overlap);
566c09c5 6851 unlock_all_device_hash_locks_irq(conf);
72626685
N		 6852 break;
6853 }
72626685 6854}
b15c2e57 6855
fd01b88c 6856static void *raid45_takeover_raid0(struct mddev *mddev, int level)
54071b38 6857{
e373ab10 6858 struct r0conf *raid0_conf = mddev->private;
d76c8420 6859 sector_t sectors;
54071b38 6860
f1b29bca 6861 /* for raid0 takeover only one zone is supported */
e373ab10 6862 if (raid0_conf->nr_strip_zones > 1) {
0c55e022
N		 6863 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6864 mdname(mddev));
f1b29bca
DW		 6865 return ERR_PTR(-EINVAL);
6866 }
6867
e373ab10
N		 6868 sectors = raid0_conf->strip_zone[0].zone_end;
6869 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
3b71bd93 6870 mddev->dev_sectors = sectors;
f1b29bca 6871 mddev->new_level = level;
54071b38
TM		 6872 mddev->new_layout = ALGORITHM_PARITY_N;
6873 mddev->new_chunk_sectors = mddev->chunk_sectors;
6874 mddev->raid_disks += 1;
6875 mddev->delta_disks = 1;
6876 /* make sure it will be not marked as dirty */
6877 mddev->recovery_cp = MaxSector;
6878
6879 return setup_conf(mddev);
6880}
6881
fd01b88c 6882static void *raid5_takeover_raid1(struct mddev *mddev)
d562b0c4
N		 6883{
6884 int chunksect;
6885
6886 if (mddev->raid_disks != 2 ||
6887 mddev->degraded > 1)
6888 return ERR_PTR(-EINVAL);
6889
6890 /* Should check if there are write-behind devices? */
6891
6892 chunksect = 64*2; /* 64K by default */
6893
6894 /* The array must be an exact multiple of chunksize */
6895 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6896 chunksect >>= 1;
6897
6898 if ((chunksect<<9) < STRIPE_SIZE)
6899 /* array size does not allow a suitable chunk size */
6900 return ERR_PTR(-EINVAL);
6901
6902 mddev->new_level = 5;
6903 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
664e7c41 6904 mddev->new_chunk_sectors = chunksect;
d562b0c4
N		 6905
6906 return setup_conf(mddev);
6907}
6908
fd01b88c 6909static void *raid5_takeover_raid6(struct mddev *mddev)
fc9739c6
N		 6910{
6911 int new_layout;
6912
6913 switch (mddev->layout) {
6914 case ALGORITHM_LEFT_ASYMMETRIC_6:
6915 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6916 break;
6917 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6918 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6919 break;
6920 case ALGORITHM_LEFT_SYMMETRIC_6:
6921 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6922 break;
6923 case ALGORITHM_RIGHT_SYMMETRIC_6:
6924 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6925 break;
6926 case ALGORITHM_PARITY_0_6:
6927 new_layout = ALGORITHM_PARITY_0;
6928 break;
6929 case ALGORITHM_PARITY_N:
6930 new_layout = ALGORITHM_PARITY_N;
6931 break;
6932 default:
6933 return ERR_PTR(-EINVAL);
6934 }
6935 mddev->new_level = 5;
6936 mddev->new_layout = new_layout;
6937 mddev->delta_disks = -1;
6938 mddev->raid_disks -= 1;
6939 return setup_conf(mddev);
6940}
6941
fd01b88c 6942static int raid5_check_reshape(struct mddev *mddev)
b3546035 6943{
88ce4930
N		 6944 /* For a 2-drive array, the layout and chunk size can be changed
6945 * immediately as not restriping is needed.
6946 * For larger arrays we record the new value - after validation
6947 * to be used by a reshape pass.
b3546035 6948 */
d1688a6d 6949 struct r5conf *conf = mddev->private;
597a711b 6950 int new_chunk = mddev->new_chunk_sectors;
b3546035 6951
597a711b 6952 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
b3546035
N		 6953 return -EINVAL;
6954 if (new_chunk > 0) {
0ba459d2 6955 if (!is_power_of_2(new_chunk))
b3546035 6956 return -EINVAL;
597a711b 6957 if (new_chunk < (PAGE_SIZE>>9))
b3546035 6958 return -EINVAL;
597a711b 6959 if (mddev->array_sectors & (new_chunk-1))
b3546035
N		 6960 /* not factor of array size */
6961 return -EINVAL;
6962 }
6963
6964 /* They look valid */
6965
88ce4930 6966 if (mddev->raid_disks == 2) {
597a711b
N		 6967 /* can make the change immediately */
6968 if (mddev->new_layout >= 0) {
6969 conf->algorithm = mddev->new_layout;
6970 mddev->layout = mddev->new_layout;
88ce4930
N		 6971 }
6972 if (new_chunk > 0) {
597a711b
N		 6973 conf->chunk_sectors = new_chunk ;
6974 mddev->chunk_sectors = new_chunk;
88ce4930
N		 6975 }
6976 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6977 md_wakeup_thread(mddev->thread);
b3546035 6978 }
50ac168a 6979 return check_reshape(mddev);
88ce4930
N		 6980}
6981
fd01b88c 6982static int raid6_check_reshape(struct mddev *mddev)
88ce4930 6983{
597a711b 6984 int new_chunk = mddev->new_chunk_sectors;
50ac168a 6985
597a711b 6986 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
88ce4930 6987 return -EINVAL;
b3546035 6988 if (new_chunk > 0) {
0ba459d2 6989 if (!is_power_of_2(new_chunk))
88ce4930 6990 return -EINVAL;
597a711b 6991 if (new_chunk < (PAGE_SIZE >> 9))
88ce4930 6992 return -EINVAL;
597a711b 6993 if (mddev->array_sectors & (new_chunk-1))
88ce4930
N		 6994 /* not factor of array size */
6995 return -EINVAL;
b3546035 6996 }
88ce4930
N		 6997
6998 /* They look valid */
50ac168a 6999 return check_reshape(mddev);
b3546035
N		 7000}
7001
fd01b88c 7002static void *raid5_takeover(struct mddev *mddev)
d562b0c4
N		 7003{
7004 /* raid5 can take over:
f1b29bca 7005 * raid0 - if there is only one strip zone - make it a raid4 layout
d562b0c4
N		 7006 * raid1 - if there are two drives. We need to know the chunk size
7007 * raid4 - trivial - just use a raid4 layout.
7008 * raid6 - Providing it is a *_6 layout
d562b0c4 7009 */
f1b29bca
DW		 7010 if (mddev->level == 0)
7011 return raid45_takeover_raid0(mddev, 5);
d562b0c4
N		 7012 if (mddev->level == 1)
7013 return raid5_takeover_raid1(mddev);
e9d4758f
N		 7014 if (mddev->level == 4) {
7015 mddev->new_layout = ALGORITHM_PARITY_N;
7016 mddev->new_level = 5;
7017 return setup_conf(mddev);
7018 }
fc9739c6
N		 7019 if (mddev->level == 6)
7020 return raid5_takeover_raid6(mddev);
d562b0c4
N		 7021
7022 return ERR_PTR(-EINVAL);
7023}
7024
fd01b88c 7025static void *raid4_takeover(struct mddev *mddev)
a78d38a1 7026{
f1b29bca
DW		 7027 /* raid4 can take over:
7028 * raid0 - if there is only one strip zone
7029 * raid5 - if layout is right
a78d38a1 7030 */
f1b29bca
DW		 7031 if (mddev->level == 0)
7032 return raid45_takeover_raid0(mddev, 4);
a78d38a1
N		 7033 if (mddev->level == 5 &&
7034 mddev->layout == ALGORITHM_PARITY_N) {
7035 mddev->new_layout = 0;
7036 mddev->new_level = 4;
7037 return setup_conf(mddev);
7038 }
7039 return ERR_PTR(-EINVAL);
7040}
d562b0c4 7041
84fc4b56 7042static struct md_personality raid5_personality;
245f46c2 7043
fd01b88c 7044static void *raid6_takeover(struct mddev *mddev)
245f46c2
N		 7045{
7046 /* Currently can only take over a raid5. We map the
7047 * personality to an equivalent raid6 personality
7048 * with the Q block at the end.
7049 */
7050 int new_layout;
7051
7052 if (mddev->pers != &raid5_personality)
7053 return ERR_PTR(-EINVAL);
7054 if (mddev->degraded > 1)
7055 return ERR_PTR(-EINVAL);
7056 if (mddev->raid_disks > 253)
7057 return ERR_PTR(-EINVAL);
7058 if (mddev->raid_disks < 3)
7059 return ERR_PTR(-EINVAL);
7060
7061 switch (mddev->layout) {
7062 case ALGORITHM_LEFT_ASYMMETRIC:
7063 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7064 break;
7065 case ALGORITHM_RIGHT_ASYMMETRIC:
7066 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7067 break;
7068 case ALGORITHM_LEFT_SYMMETRIC:
7069 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7070 break;
7071 case ALGORITHM_RIGHT_SYMMETRIC:
7072 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7073 break;
7074 case ALGORITHM_PARITY_0:
7075 new_layout = ALGORITHM_PARITY_0_6;
7076 break;
7077 case ALGORITHM_PARITY_N:
7078 new_layout = ALGORITHM_PARITY_N;
7079 break;
7080 default:
7081 return ERR_PTR(-EINVAL);
7082 }
7083 mddev->new_level = 6;
7084 mddev->new_layout = new_layout;
7085 mddev->delta_disks = 1;
7086 mddev->raid_disks += 1;
7087 return setup_conf(mddev);
7088}
7089
84fc4b56 7090static struct md_personality raid6_personality =
16a53ecc
N		 7091{
7092 .name = "raid6",
7093 .level = 6,
7094 .owner = THIS_MODULE,
7095 .make_request = make_request,
7096 .run = run,
7097 .stop = stop,
7098 .status = status,
7099 .error_handler = error,
7100 .hot_add_disk = raid5_add_disk,
7101 .hot_remove_disk= raid5_remove_disk,
7102 .spare_active = raid5_spare_active,
7103 .sync_request = sync_request,
7104 .resize = raid5_resize,
80c3a6ce 7105 .size = raid5_size,
50ac168a 7106 .check_reshape = raid6_check_reshape,
f416885e 7107 .start_reshape = raid5_start_reshape,
cea9c228 7108 .finish_reshape = raid5_finish_reshape,
16a53ecc 7109 .quiesce = raid5_quiesce,
245f46c2 7110 .takeover = raid6_takeover,
5c675f83 7111 .congested = raid5_congested,
64590f45 7112 .mergeable_bvec = raid5_mergeable_bvec,
16a53ecc 7113};
84fc4b56 7114static struct md_personality raid5_personality =
1da177e4
LT		 7115{
7116 .name = "raid5",
2604b703 7117 .level = 5,
1da177e4
LT		 7118 .owner = THIS_MODULE,
7119 .make_request = make_request,
7120 .run = run,
7121 .stop = stop,
7122 .status = status,
7123 .error_handler = error,
7124 .hot_add_disk = raid5_add_disk,
7125 .hot_remove_disk= raid5_remove_disk,
7126 .spare_active = raid5_spare_active,
7127 .sync_request = sync_request,
7128 .resize = raid5_resize,
80c3a6ce 7129 .size = raid5_size,
63c70c4f
N		 7130 .check_reshape = raid5_check_reshape,
7131 .start_reshape = raid5_start_reshape,
cea9c228 7132 .finish_reshape = raid5_finish_reshape,
72626685 7133 .quiesce = raid5_quiesce,
d562b0c4 7134 .takeover = raid5_takeover,
5c675f83 7135 .congested = raid5_congested,
64590f45 7136 .mergeable_bvec = raid5_mergeable_bvec,
1da177e4
LT		 7137};
7138
84fc4b56 7139static struct md_personality raid4_personality =
1da177e4 7140{
2604b703
N		 7141 .name = "raid4",
7142 .level = 4,
7143 .owner = THIS_MODULE,
7144 .make_request = make_request,
7145 .run = run,
7146 .stop = stop,
7147 .status = status,
7148 .error_handler = error,
7149 .hot_add_disk = raid5_add_disk,
7150 .hot_remove_disk= raid5_remove_disk,
7151 .spare_active = raid5_spare_active,
7152 .sync_request = sync_request,
7153 .resize = raid5_resize,
80c3a6ce 7154 .size = raid5_size,
3d37890b
N		 7155 .check_reshape = raid5_check_reshape,
7156 .start_reshape = raid5_start_reshape,
cea9c228 7157 .finish_reshape = raid5_finish_reshape,
2604b703 7158 .quiesce = raid5_quiesce,
a78d38a1 7159 .takeover = raid4_takeover,
5c675f83 7160 .congested = raid5_congested,
64590f45 7161 .mergeable_bvec = raid5_mergeable_bvec,
2604b703
N		 7162};
7163
7164static int __init raid5_init(void)
7165{
851c30c9
SL		 7166 raid5_wq = alloc_workqueue("raid5wq",
7167 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7168 if (!raid5_wq)
7169 return -ENOMEM;
16a53ecc 7170 register_md_personality(&raid6_personality);
2604b703
N		 7171 register_md_personality(&raid5_personality);
7172 register_md_personality(&raid4_personality);
7173 return 0;
1da177e4
LT		 7174}
7175
2604b703 7176static void raid5_exit(void)
1da177e4 7177{
16a53ecc 7178 unregister_md_personality(&raid6_personality);
2604b703
N		 7179 unregister_md_personality(&raid5_personality);
7180 unregister_md_personality(&raid4_personality);
851c30c9 7181 destroy_workqueue(raid5_wq);
1da177e4
LT		 7182}
7183
7184module_init(raid5_init);
7185module_exit(raid5_exit);
7186MODULE_LICENSE("GPL");
0efb9e61 7187MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
1da177e4 7188MODULE_ALIAS("md-personality-4"); /* RAID5 */
d9d166c2
N		 7189MODULE_ALIAS("md-raid5");
7190MODULE_ALIAS("md-raid4");
2604b703
N		 7191MODULE_ALIAS("md-level-5");
7192MODULE_ALIAS("md-level-4");
16a53ecc
N		 7193MODULE_ALIAS("md-personality-8"); /* RAID6 */
7194MODULE_ALIAS("md-raid6");
7195MODULE_ALIAS("md-level-6");
7196
7197/* This used to be two separate modules, they were: */
7198MODULE_ALIAS("raid5");
7199MODULE_ALIAS("raid6");
Linux 6.x block layer and io_uring tree(s)