projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
Linux 3.17-rc7
[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
67static struct workqueue_struct *raid5_wq;
1da177e4
LT		 68/*
69 * Stripe cache
70 */
71
72#define NR_STRIPES 256
73#define STRIPE_SIZE PAGE_SIZE
74#define STRIPE_SHIFT (PAGE_SHIFT - 9)
75#define STRIPE_SECTORS (STRIPE_SIZE>>9)
76#define IO_THRESHOLD 1
8b3e6cdc 77#define BYPASS_THRESHOLD 1
fccddba0 78#define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
1da177e4 79#define HASH_MASK (NR_HASH - 1)
bfc90cb0 80#define MAX_STRIPE_BATCH 8
1da177e4 81
d1688a6d 82static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
db298e19
N		 83{
84 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
85 return &conf->stripe_hashtbl[hash];
86}
1da177e4 87
566c09c5
SL		 88static inline int stripe_hash_locks_hash(sector_t sect)
89{
90 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
91}
92
93static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
94{
95 spin_lock_irq(conf->hash_locks + hash);
96 spin_lock(&conf->device_lock);
97}
98
99static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
100{
101 spin_unlock(&conf->device_lock);
102 spin_unlock_irq(conf->hash_locks + hash);
103}
104
105static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
106{
107 int i;
108 local_irq_disable();
109 spin_lock(conf->hash_locks);
110 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
111 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
112 spin_lock(&conf->device_lock);
113}
114
115static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
116{
117 int i;
118 spin_unlock(&conf->device_lock);
119 for (i = NR_STRIPE_HASH_LOCKS; i; i--)
120 spin_unlock(conf->hash_locks + i - 1);
121 local_irq_enable();
122}
123
1da177e4
LT		 124/* bio's attached to a stripe+device for I/O are linked together in bi_sector
125 * order without overlap. There may be several bio's per stripe+device, and
126 * a bio could span several devices.
127 * When walking this list for a particular stripe+device, we must never proceed
128 * beyond a bio that extends past this device, as the next bio might no longer
129 * be valid.
db298e19 130 * This function is used to determine the 'next' bio in the list, given the sector
1da177e4
LT		 131 * of the current stripe+device
132 */
db298e19
N		 133static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
134{
aa8b57aa 135 int sectors = bio_sectors(bio);
4f024f37 136 if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
db298e19
N		 137 return bio->bi_next;
138 else
139 return NULL;
140}
1da177e4 141
960e739d 142/*
5b99c2ff
JA		 143 * We maintain a biased count of active stripes in the bottom 16 bits of
144 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
960e739d 145 */
e7836bd6 146static inline int raid5_bi_processed_stripes(struct bio *bio)
960e739d 147{
e7836bd6
SL		 148 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
149 return (atomic_read(segments) >> 16) & 0xffff;
960e739d
JA		 150}
151
e7836bd6 152static inline int raid5_dec_bi_active_stripes(struct bio *bio)
960e739d 153{
e7836bd6
SL		 154 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
155 return atomic_sub_return(1, segments) & 0xffff;
960e739d
JA		 156}
157
e7836bd6 158static inline void raid5_inc_bi_active_stripes(struct bio *bio)
960e739d 159{
e7836bd6
SL		 160 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
161 atomic_inc(segments);
960e739d
JA		 162}
163
e7836bd6
SL		 164static inline void raid5_set_bi_processed_stripes(struct bio *bio,
165 unsigned int cnt)
960e739d 166{
e7836bd6
SL		 167 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
168 int old, new;
960e739d 169
e7836bd6
SL		 170 do {
171 old = atomic_read(segments);
172 new = (old & 0xffff) | (cnt << 16);
173 } while (atomic_cmpxchg(segments, old, new) != old);
960e739d
JA		 174}
175
e7836bd6 176static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
960e739d 177{
e7836bd6
SL		 178 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
179 atomic_set(segments, cnt);
960e739d
JA		 180}
181
d0dabf7e
N		 182/* Find first data disk in a raid6 stripe */
183static inline int raid6_d0(struct stripe_head *sh)
184{
67cc2b81
N		 185 if (sh->ddf_layout)
186 /* ddf always start from first device */
187 return 0;
188 /* md starts just after Q block */
d0dabf7e
N		 189 if (sh->qd_idx == sh->disks - 1)
190 return 0;
191 else
192 return sh->qd_idx + 1;
193}
16a53ecc
N		 194static inline int raid6_next_disk(int disk, int raid_disks)
195{
196 disk++;
197 return (disk < raid_disks) ? disk : 0;
198}
a4456856 199
d0dabf7e
N		 200/* When walking through the disks in a raid5, starting at raid6_d0,
201 * We need to map each disk to a 'slot', where the data disks are slot
202 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
203 * is raid_disks-1. This help does that mapping.
204 */
67cc2b81
N		 205static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
206 int *count, int syndrome_disks)
d0dabf7e 207{
6629542e 208 int slot = *count;
67cc2b81 209
e4424fee 210 if (sh->ddf_layout)
6629542e 211 (*count)++;
d0dabf7e 212 if (idx == sh->pd_idx)
67cc2b81 213 return syndrome_disks;
d0dabf7e 214 if (idx == sh->qd_idx)
67cc2b81 215 return syndrome_disks + 1;
e4424fee 216 if (!sh->ddf_layout)
6629542e 217 (*count)++;
d0dabf7e
N		 218 return slot;
219}
220
a4456856
DW		 221static void return_io(struct bio *return_bi)
222{
223 struct bio *bi = return_bi;
224 while (bi) {
a4456856
DW		 225
226 return_bi = bi->bi_next;
227 bi->bi_next = NULL;
4f024f37 228 bi->bi_iter.bi_size = 0;
0a82a8d1
LT		 229 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
230 bi, 0);
0e13fe23 231 bio_endio(bi, 0);
a4456856
DW		 232 bi = return_bi;
233 }
234}
235
d1688a6d 236static void print_raid5_conf (struct r5conf *conf);
1da177e4 237
600aa109
DW		 238static int stripe_operations_active(struct stripe_head *sh)
239{
240 return sh->check_state || sh->reconstruct_state ||
241 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
242 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
243}
244
851c30c9
SL		 245static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
246{
247 struct r5conf *conf = sh->raid_conf;
248 struct r5worker_group *group;
bfc90cb0 249 int thread_cnt;
851c30c9
SL		 250 int i, cpu = sh->cpu;
251
252 if (!cpu_online(cpu)) {
253 cpu = cpumask_any(cpu_online_mask);
254 sh->cpu = cpu;
255 }
256
257 if (list_empty(&sh->lru)) {
258 struct r5worker_group *group;
259 group = conf->worker_groups + cpu_to_group(cpu);
260 list_add_tail(&sh->lru, &group->handle_list);
bfc90cb0
SL		 261 group->stripes_cnt++;
262 sh->group = group;
851c30c9
SL		 263 }
264
265 if (conf->worker_cnt_per_group == 0) {
266 md_wakeup_thread(conf->mddev->thread);
267 return;
268 }
269
270 group = conf->worker_groups + cpu_to_group(sh->cpu);
271
bfc90cb0
SL		 272 group->workers[0].working = true;
273 /* at least one worker should run to avoid race */
274 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
275
276 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
277 /* wakeup more workers */
278 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
279 if (group->workers[i].working == false) {
280 group->workers[i].working = true;
281 queue_work_on(sh->cpu, raid5_wq,
282 &group->workers[i].work);
283 thread_cnt--;
284 }
285 }
851c30c9
SL		 286}
287
566c09c5
SL		 288static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
289 struct list_head *temp_inactive_list)
1da177e4 290{
4eb788df
SL		 291 BUG_ON(!list_empty(&sh->lru));
292 BUG_ON(atomic_read(&conf->active_stripes)==0);
293 if (test_bit(STRIPE_HANDLE, &sh->state)) {
294 if (test_bit(STRIPE_DELAYED, &sh->state) &&
67f45548 295 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
4eb788df 296 list_add_tail(&sh->lru, &conf->delayed_list);
67f45548
N		 297 if (atomic_read(&conf->preread_active_stripes)
298 < IO_THRESHOLD)
299 md_wakeup_thread(conf->mddev->thread);
300 } else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
4eb788df
SL		 301 sh->bm_seq - conf->seq_write > 0)
302 list_add_tail(&sh->lru, &conf->bitmap_list);
303 else {
304 clear_bit(STRIPE_DELAYED, &sh->state);
305 clear_bit(STRIPE_BIT_DELAY, &sh->state);
851c30c9
SL		 306 if (conf->worker_cnt_per_group == 0) {
307 list_add_tail(&sh->lru, &conf->handle_list);
308 } else {
309 raid5_wakeup_stripe_thread(sh);
310 return;
311 }
4eb788df
SL		 312 }
313 md_wakeup_thread(conf->mddev->thread);
314 } else {
315 BUG_ON(stripe_operations_active(sh));
316 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
317 if (atomic_dec_return(&conf->preread_active_stripes)
318 < IO_THRESHOLD)
319 md_wakeup_thread(conf->mddev->thread);
320 atomic_dec(&conf->active_stripes);
566c09c5
SL		 321 if (!test_bit(STRIPE_EXPANDING, &sh->state))
322 list_add_tail(&sh->lru, temp_inactive_list);
1da177e4
LT		 323 }
324}
d0dabf7e 325
566c09c5
SL		 326static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
327 struct list_head *temp_inactive_list)
4eb788df
SL		 328{
329 if (atomic_dec_and_test(&sh->count))
566c09c5
SL		 330 do_release_stripe(conf, sh, temp_inactive_list);
331}
332
333/*
334 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
335 *
336 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
337 * given time. Adding stripes only takes device lock, while deleting stripes
338 * only takes hash lock.
339 */
340static void release_inactive_stripe_list(struct r5conf *conf,
341 struct list_head *temp_inactive_list,
342 int hash)
343{
344 int size;
345 bool do_wakeup = false;
346 unsigned long flags;
347
348 if (hash == NR_STRIPE_HASH_LOCKS) {
349 size = NR_STRIPE_HASH_LOCKS;
350 hash = NR_STRIPE_HASH_LOCKS - 1;
351 } else
352 size = 1;
353 while (size) {
354 struct list_head *list = &temp_inactive_list[size - 1];
355
356 /*
357 * We don't hold any lock here yet, get_active_stripe() might
358 * remove stripes from the list
359 */
360 if (!list_empty_careful(list)) {
361 spin_lock_irqsave(conf->hash_locks + hash, flags);
4bda556a
SL		 362 if (list_empty(conf->inactive_list + hash) &&
363 !list_empty(list))
364 atomic_dec(&conf->empty_inactive_list_nr);
566c09c5
SL		 365 list_splice_tail_init(list, conf->inactive_list + hash);
366 do_wakeup = true;
367 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
368 }
369 size--;
370 hash--;
371 }
372
373 if (do_wakeup) {
374 wake_up(&conf->wait_for_stripe);
375 if (conf->retry_read_aligned)
376 md_wakeup_thread(conf->mddev->thread);
377 }
4eb788df
SL		 378}
379
773ca82f 380/* should hold conf->device_lock already */
566c09c5
SL		 381static int release_stripe_list(struct r5conf *conf,
382 struct list_head *temp_inactive_list)
773ca82f
SL		 383{
384 struct stripe_head *sh;
385 int count = 0;
386 struct llist_node *head;
387
388 head = llist_del_all(&conf->released_stripes);
d265d9dc 389 head = llist_reverse_order(head);
773ca82f 390 while (head) {
566c09c5
SL		 391 int hash;
392
773ca82f
SL		 393 sh = llist_entry(head, struct stripe_head, release_list);
394 head = llist_next(head);
395 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
396 smp_mb();
397 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
398 /*
399 * Don't worry the bit is set here, because if the bit is set
400 * again, the count is always > 1. This is true for
401 * STRIPE_ON_UNPLUG_LIST bit too.
402 */
566c09c5
SL		 403 hash = sh->hash_lock_index;
404 __release_stripe(conf, sh, &temp_inactive_list[hash]);
773ca82f
SL		 405 count++;
406 }
407
408 return count;
409}
410
1da177e4
LT		 411static void release_stripe(struct stripe_head *sh)
412{
d1688a6d 413 struct r5conf *conf = sh->raid_conf;
1da177e4 414 unsigned long flags;
566c09c5
SL		 415 struct list_head list;
416 int hash;
773ca82f 417 bool wakeup;
16a53ecc 418
cf170f3f
ES		 419 /* Avoid release_list until the last reference.
420 */
421 if (atomic_add_unless(&sh->count, -1, 1))
422 return;
423
ad4068de 424 if (unlikely(!conf->mddev->thread) ||
425 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
773ca82f
SL		 426 goto slow_path;
427 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
428 if (wakeup)
429 md_wakeup_thread(conf->mddev->thread);
430 return;
431slow_path:
4eb788df 432 local_irq_save(flags);
773ca82f 433 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
4eb788df 434 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
566c09c5
SL		 435 INIT_LIST_HEAD(&list);
436 hash = sh->hash_lock_index;
437 do_release_stripe(conf, sh, &list);
4eb788df 438 spin_unlock(&conf->device_lock);
566c09c5 439 release_inactive_stripe_list(conf, &list, hash);
4eb788df
SL		 440 }
441 local_irq_restore(flags);
1da177e4
LT		 442}
443
fccddba0 444static inline void remove_hash(struct stripe_head *sh)
1da177e4 445{
45b4233c
DW		 446 pr_debug("remove_hash(), stripe %llu\n",
447 (unsigned long long)sh->sector);
1da177e4 448
fccddba0 449 hlist_del_init(&sh->hash);
1da177e4
LT		 450}
451
d1688a6d 452static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
1da177e4 453{
fccddba0 454 struct hlist_head *hp = stripe_hash(conf, sh->sector);
1da177e4 455
45b4233c
DW		 456 pr_debug("insert_hash(), stripe %llu\n",
457 (unsigned long long)sh->sector);
1da177e4 458
fccddba0 459 hlist_add_head(&sh->hash, hp);
1da177e4
LT		 460}
461
462
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",
1da177e4
LT		 530 (unsigned long long)sh->sector);
531
532 remove_hash(sh);
566c09c5
SL		 533retry:
534 seq = read_seqcount_begin(&conf->gen_lock);
86b42c71 535 sh->generation = conf->generation - previous;
b5663ba4 536 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
1da177e4 537 sh->sector = sector;
911d4ee8 538 stripe_set_idx(sector, conf, previous, sh);
1da177e4
LT		 539 sh->state = 0;
540
7ecaa1e6
N		 541
542 for (i = sh->disks; i--; ) {
1da177e4
LT		 543 struct r5dev *dev = &sh->dev[i];
544
d84e0f10 545 if (dev->toread || dev->read || dev->towrite || dev->written ||
1da177e4 546 test_bit(R5_LOCKED, &dev->flags)) {
d84e0f10 547 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
1da177e4 548 (unsigned long long)sh->sector, i, dev->toread,
d84e0f10 549 dev->read, dev->towrite, dev->written,
1da177e4 550 test_bit(R5_LOCKED, &dev->flags));
8cfa7b0f 551 WARN_ON(1);
1da177e4
LT		 552 }
553 dev->flags = 0;
784052ec 554 raid5_build_block(sh, i, previous);
1da177e4 555 }
566c09c5
SL		 556 if (read_seqcount_retry(&conf->gen_lock, seq))
557 goto retry;
1da177e4 558 insert_hash(conf, sh);
851c30c9 559 sh->cpu = smp_processor_id();
1da177e4
LT		 560}
561
d1688a6d 562static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
86b42c71 563 short generation)
1da177e4
LT		 564{
565 struct stripe_head *sh;
566
45b4233c 567 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
b67bfe0d 568 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
86b42c71 569 if (sh->sector == sector && sh->generation == generation)
1da177e4 570 return sh;
45b4233c 571 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
1da177e4
LT		 572 return NULL;
573}
574
674806d6
N		 575/*
576 * Need to check if array has failed when deciding whether to:
577 * - start an array
578 * - remove non-faulty devices
579 * - add a spare
580 * - allow a reshape
581 * This determination is simple when no reshape is happening.
582 * However if there is a reshape, we need to carefully check
583 * both the before and after sections.
584 * This is because some failed devices may only affect one
585 * of the two sections, and some non-in_sync devices may
586 * be insync in the section most affected by failed devices.
587 */
908f4fbd 588static int calc_degraded(struct r5conf *conf)
674806d6 589{
908f4fbd 590 int degraded, degraded2;
674806d6 591 int i;
674806d6
N		 592
593 rcu_read_lock();
594 degraded = 0;
595 for (i = 0; i < conf->previous_raid_disks; i++) {
3cb03002 596 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
e5c86471
N		 597 if (rdev && test_bit(Faulty, &rdev->flags))
598 rdev = rcu_dereference(conf->disks[i].replacement);
674806d6
N		 599 if (!rdev || test_bit(Faulty, &rdev->flags))
600 degraded++;
601 else if (test_bit(In_sync, &rdev->flags))
602 ;
603 else
604 /* not in-sync or faulty.
605 * If the reshape increases the number of devices,
606 * this is being recovered by the reshape, so
607 * this 'previous' section is not in_sync.
608 * If the number of devices is being reduced however,
609 * the device can only be part of the array if
610 * we are reverting a reshape, so this section will
611 * be in-sync.
612 */
613 if (conf->raid_disks >= conf->previous_raid_disks)
614 degraded++;
615 }
616 rcu_read_unlock();
908f4fbd
N		 617 if (conf->raid_disks == conf->previous_raid_disks)
618 return degraded;
674806d6 619 rcu_read_lock();
908f4fbd 620 degraded2 = 0;
674806d6 621 for (i = 0; i < conf->raid_disks; i++) {
3cb03002 622 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
e5c86471
N		 623 if (rdev && test_bit(Faulty, &rdev->flags))
624 rdev = rcu_dereference(conf->disks[i].replacement);
674806d6 625 if (!rdev || test_bit(Faulty, &rdev->flags))
908f4fbd 626 degraded2++;
674806d6
N		 627 else if (test_bit(In_sync, &rdev->flags))
628 ;
629 else
630 /* not in-sync or faulty.
631 * If reshape increases the number of devices, this
632 * section has already been recovered, else it
633 * almost certainly hasn't.
634 */
635 if (conf->raid_disks <= conf->previous_raid_disks)
908f4fbd 636 degraded2++;
674806d6
N		 637 }
638 rcu_read_unlock();
908f4fbd
N		 639 if (degraded2 > degraded)
640 return degraded2;
641 return degraded;
642}
643
644static int has_failed(struct r5conf *conf)
645{
646 int degraded;
647
648 if (conf->mddev->reshape_position == MaxSector)
649 return conf->mddev->degraded > conf->max_degraded;
650
651 degraded = calc_degraded(conf);
674806d6
N		 652 if (degraded > conf->max_degraded)
653 return 1;
654 return 0;
655}
656
b5663ba4 657static struct stripe_head *
d1688a6d 658get_active_stripe(struct r5conf *conf, sector_t sector,
a8c906ca 659 int previous, int noblock, int noquiesce)
1da177e4
LT		 660{
661 struct stripe_head *sh;
566c09c5 662 int hash = stripe_hash_locks_hash(sector);
1da177e4 663
45b4233c 664 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
1da177e4 665
566c09c5 666 spin_lock_irq(conf->hash_locks + hash);
1da177e4
LT		 667
668 do {
72626685 669 wait_event_lock_irq(conf->wait_for_stripe,
a8c906ca 670 conf->quiesce == 0 || noquiesce,
566c09c5 671 *(conf->hash_locks + hash));
86b42c71 672 sh = __find_stripe(conf, sector, conf->generation - previous);
1da177e4
LT		 673 if (!sh) {
674 if (!conf->inactive_blocked)
566c09c5 675 sh = get_free_stripe(conf, hash);
1da177e4
LT		 676 if (noblock && sh == NULL)
677 break;
678 if (!sh) {
679 conf->inactive_blocked = 1;
566c09c5
SL		 680 wait_event_lock_irq(
681 conf->wait_for_stripe,
682 !list_empty(conf->inactive_list + hash) &&
683 (atomic_read(&conf->active_stripes)
684 < (conf->max_nr_stripes * 3 / 4)
685 || !conf->inactive_blocked),
686 *(conf->hash_locks + hash));
1da177e4 687 conf->inactive_blocked = 0;
7da9d450 688 } else {
b5663ba4 689 init_stripe(sh, sector, previous);
7da9d450
N		 690 atomic_inc(&sh->count);
691 }
e240c183 692 } else if (!atomic_inc_not_zero(&sh->count)) {
6d183de4 693 spin_lock(&conf->device_lock);
e240c183 694 if (!atomic_read(&sh->count)) {
1da177e4
LT		 695 if (!test_bit(STRIPE_HANDLE, &sh->state))
696 atomic_inc(&conf->active_stripes);
5af9bef7
N		 697 BUG_ON(list_empty(&sh->lru) &&
698 !test_bit(STRIPE_EXPANDING, &sh->state));
16a53ecc 699 list_del_init(&sh->lru);
bfc90cb0
SL		 700 if (sh->group) {
701 sh->group->stripes_cnt--;
702 sh->group = NULL;
703 }
1da177e4 704 }
7da9d450 705 atomic_inc(&sh->count);
6d183de4 706 spin_unlock(&conf->device_lock);
1da177e4
LT		 707 }
708 } while (sh == NULL);
709
566c09c5 710 spin_unlock_irq(conf->hash_locks + hash);
1da177e4
LT		 711 return sh;
712}
713
05616be5
N		 714/* Determine if 'data_offset' or 'new_data_offset' should be used
715 * in this stripe_head.
716 */
717static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
718{
719 sector_t progress = conf->reshape_progress;
720 /* Need a memory barrier to make sure we see the value
721 * of conf->generation, or ->data_offset that was set before
722 * reshape_progress was updated.
723 */
724 smp_rmb();
725 if (progress == MaxSector)
726 return 0;
727 if (sh->generation == conf->generation - 1)
728 return 0;
729 /* We are in a reshape, and this is a new-generation stripe,
730 * so use new_data_offset.
731 */
732 return 1;
733}
734
6712ecf8
N		 735static void
736raid5_end_read_request(struct bio *bi, int error);
737static void
738raid5_end_write_request(struct bio *bi, int error);
91c00924 739
c4e5ac0a 740static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
91c00924 741{
d1688a6d 742 struct r5conf *conf = sh->raid_conf;
91c00924
DW		 743 int i, disks = sh->disks;
744
745 might_sleep();
746
747 for (i = disks; i--; ) {
748 int rw;
9a3e1101 749 int replace_only = 0;
977df362
N		 750 struct bio *bi, *rbi;
751 struct md_rdev *rdev, *rrdev = NULL;
e9c7469b
TH		 752 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
753 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
754 rw = WRITE_FUA;
755 else
756 rw = WRITE;
9e444768 757 if (test_bit(R5_Discard, &sh->dev[i].flags))
620125f2 758 rw |= REQ_DISCARD;
e9c7469b 759 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
91c00924 760 rw = READ;
9a3e1101
N		 761 else if (test_and_clear_bit(R5_WantReplace,
762 &sh->dev[i].flags)) {
763 rw = WRITE;
764 replace_only = 1;
765 } else
91c00924 766 continue;
bc0934f0
SL		 767 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
768 rw |= REQ_SYNC;
91c00924
DW		 769
770 bi = &sh->dev[i].req;
977df362 771 rbi = &sh->dev[i].rreq; /* For writing to replacement */
91c00924 772
91c00924 773 rcu_read_lock();
9a3e1101 774 rrdev = rcu_dereference(conf->disks[i].replacement);
dd054fce
N		 775 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
776 rdev = rcu_dereference(conf->disks[i].rdev);
777 if (!rdev) {
778 rdev = rrdev;
779 rrdev = NULL;
780 }
9a3e1101
N		 781 if (rw & WRITE) {
782 if (replace_only)
783 rdev = NULL;
dd054fce
N		 784 if (rdev == rrdev)
785 /* We raced and saw duplicates */
786 rrdev = NULL;
9a3e1101 787 } else {
dd054fce 788 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
9a3e1101
N		 789 rdev = rrdev;
790 rrdev = NULL;
791 }
977df362 792
91c00924
DW		 793 if (rdev && test_bit(Faulty, &rdev->flags))
794 rdev = NULL;
795 if (rdev)
796 atomic_inc(&rdev->nr_pending);
977df362
N		 797 if (rrdev && test_bit(Faulty, &rrdev->flags))
798 rrdev = NULL;
799 if (rrdev)
800 atomic_inc(&rrdev->nr_pending);
91c00924
DW		 801 rcu_read_unlock();
802
73e92e51 803 /* We have already checked bad blocks for reads. Now
977df362
N		 804 * need to check for writes. We never accept write errors
805 * on the replacement, so we don't to check rrdev.
73e92e51
N		 806 */
807 while ((rw & WRITE) && rdev &&
808 test_bit(WriteErrorSeen, &rdev->flags)) {
809 sector_t first_bad;
810 int bad_sectors;
811 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
812 &first_bad, &bad_sectors);
813 if (!bad)
814 break;
815
816 if (bad < 0) {
817 set_bit(BlockedBadBlocks, &rdev->flags);
818 if (!conf->mddev->external &&
819 conf->mddev->flags) {
820 /* It is very unlikely, but we might
821 * still need to write out the
822 * bad block log - better give it
823 * a chance*/
824 md_check_recovery(conf->mddev);
825 }
1850753d 826 /*
827 * Because md_wait_for_blocked_rdev
828 * will dec nr_pending, we must
829 * increment it first.
830 */
831 atomic_inc(&rdev->nr_pending);
73e92e51
N		 832 md_wait_for_blocked_rdev(rdev, conf->mddev);
833 } else {
834 /* Acknowledged bad block - skip the write */
835 rdev_dec_pending(rdev, conf->mddev);
836 rdev = NULL;
837 }
838 }
839
91c00924 840 if (rdev) {
9a3e1101
N		 841 if (s->syncing || s->expanding || s->expanded
842 || s->replacing)
91c00924
DW		 843 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
844
2b7497f0
DW		 845 set_bit(STRIPE_IO_STARTED, &sh->state);
846
2f6db2a7 847 bio_reset(bi);
91c00924 848 bi->bi_bdev = rdev->bdev;
2f6db2a7
KO		 849 bi->bi_rw = rw;
850 bi->bi_end_io = (rw & WRITE)
851 ? raid5_end_write_request
852 : raid5_end_read_request;
853 bi->bi_private = sh;
854
91c00924 855 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
e46b272b 856 __func__, (unsigned long long)sh->sector,
91c00924
DW		 857 bi->bi_rw, i);
858 atomic_inc(&sh->count);
05616be5 859 if (use_new_offset(conf, sh))
4f024f37 860 bi->bi_iter.bi_sector = (sh->sector
05616be5
N		 861 + rdev->new_data_offset);
862 else
4f024f37 863 bi->bi_iter.bi_sector = (sh->sector
05616be5 864 + rdev->data_offset);
3f9e7c14 865 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
e59aa23f 866 bi->bi_rw |= REQ_NOMERGE;
3f9e7c14 867
d592a996
SL		 868 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
869 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
870 sh->dev[i].vec.bv_page = sh->dev[i].page;
4997b72e 871 bi->bi_vcnt = 1;
91c00924
DW		 872 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
873 bi->bi_io_vec[0].bv_offset = 0;
4f024f37 874 bi->bi_iter.bi_size = STRIPE_SIZE;
37c61ff3
SL		 875 /*
876 * If this is discard request, set bi_vcnt 0. We don't
877 * want to confuse SCSI because SCSI will replace payload
878 */
879 if (rw & REQ_DISCARD)
880 bi->bi_vcnt = 0;
977df362
N		 881 if (rrdev)
882 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
e3620a3a
JB		 883
884 if (conf->mddev->gendisk)
885 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
886 bi, disk_devt(conf->mddev->gendisk),
887 sh->dev[i].sector);
91c00924 888 generic_make_request(bi);
977df362
N		 889 }
890 if (rrdev) {
9a3e1101
N		 891 if (s->syncing || s->expanding || s->expanded
892 || s->replacing)
977df362
N		 893 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
894
895 set_bit(STRIPE_IO_STARTED, &sh->state);
896
2f6db2a7 897 bio_reset(rbi);
977df362 898 rbi->bi_bdev = rrdev->bdev;
2f6db2a7
KO		 899 rbi->bi_rw = rw;
900 BUG_ON(!(rw & WRITE));
901 rbi->bi_end_io = raid5_end_write_request;
902 rbi->bi_private = sh;
903
977df362
N		 904 pr_debug("%s: for %llu schedule op %ld on "
905 "replacement disc %d\n",
906 __func__, (unsigned long long)sh->sector,
907 rbi->bi_rw, i);
908 atomic_inc(&sh->count);
05616be5 909 if (use_new_offset(conf, sh))
4f024f37 910 rbi->bi_iter.bi_sector = (sh->sector
05616be5
N		 911 + rrdev->new_data_offset);
912 else
4f024f37 913 rbi->bi_iter.bi_sector = (sh->sector
05616be5 914 + rrdev->data_offset);
d592a996
SL		 915 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
916 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
917 sh->dev[i].rvec.bv_page = sh->dev[i].page;
4997b72e 918 rbi->bi_vcnt = 1;
977df362
N		 919 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
920 rbi->bi_io_vec[0].bv_offset = 0;
4f024f37 921 rbi->bi_iter.bi_size = STRIPE_SIZE;
37c61ff3
SL		 922 /*
923 * If this is discard request, set bi_vcnt 0. We don't
924 * want to confuse SCSI because SCSI will replace payload
925 */
926 if (rw & REQ_DISCARD)
927 rbi->bi_vcnt = 0;
e3620a3a
JB		 928 if (conf->mddev->gendisk)
929 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
930 rbi, disk_devt(conf->mddev->gendisk),
931 sh->dev[i].sector);
977df362
N		 932 generic_make_request(rbi);
933 }
934 if (!rdev && !rrdev) {
b062962e 935 if (rw & WRITE)
91c00924
DW		 936 set_bit(STRIPE_DEGRADED, &sh->state);
937 pr_debug("skip op %ld on disc %d for sector %llu\n",
938 bi->bi_rw, i, (unsigned long long)sh->sector);
939 clear_bit(R5_LOCKED, &sh->dev[i].flags);
940 set_bit(STRIPE_HANDLE, &sh->state);
941 }
942 }
943}
944
945static struct dma_async_tx_descriptor *
d592a996
SL		 946async_copy_data(int frombio, struct bio *bio, struct page **page,
947 sector_t sector, struct dma_async_tx_descriptor *tx,
948 struct stripe_head *sh)
91c00924 949{
7988613b
KO		 950 struct bio_vec bvl;
951 struct bvec_iter iter;
91c00924 952 struct page *bio_page;
91c00924 953 int page_offset;
a08abd8c 954 struct async_submit_ctl submit;
0403e382 955 enum async_tx_flags flags = 0;
91c00924 956
4f024f37
KO		 957 if (bio->bi_iter.bi_sector >= sector)
958 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
91c00924 959 else
4f024f37 960 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
a08abd8c 961
0403e382
DW		 962 if (frombio)
963 flags |= ASYNC_TX_FENCE;
964 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
965
7988613b
KO		 966 bio_for_each_segment(bvl, bio, iter) {
967 int len = bvl.bv_len;
91c00924
DW		 968 int clen;
969 int b_offset = 0;
970
971 if (page_offset < 0) {
972 b_offset = -page_offset;
973 page_offset += b_offset;
974 len -= b_offset;
975 }
976
977 if (len > 0 && page_offset + len > STRIPE_SIZE)
978 clen = STRIPE_SIZE - page_offset;
979 else
980 clen = len;
981
982 if (clen > 0) {
7988613b
KO		 983 b_offset += bvl.bv_offset;
984 bio_page = bvl.bv_page;
d592a996
SL		 985 if (frombio) {
986 if (sh->raid_conf->skip_copy &&
987 b_offset == 0 && page_offset == 0 &&
988 clen == STRIPE_SIZE)
989 *page = bio_page;
990 else
991 tx = async_memcpy(*page, bio_page, page_offset,
a08abd8c 992 b_offset, clen, &submit);
d592a996
SL		 993 } else
994 tx = async_memcpy(bio_page, *page, b_offset,
a08abd8c 995 page_offset, clen, &submit);
91c00924 996 }
a08abd8c
DW		 997 /* chain the operations */
998 submit.depend_tx = tx;
999
91c00924
DW		 1000 if (clen < len) /* hit end of page */
1001 break;
1002 page_offset += len;
1003 }
1004
1005 return tx;
1006}
1007
1008static void ops_complete_biofill(void *stripe_head_ref)
1009{
1010 struct stripe_head *sh = stripe_head_ref;
1011 struct bio *return_bi = NULL;
e4d84909 1012 int i;
91c00924 1013
e46b272b 1014 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1015 (unsigned long long)sh->sector);
1016
1017 /* clear completed biofills */
1018 for (i = sh->disks; i--; ) {
1019 struct r5dev *dev = &sh->dev[i];
91c00924
DW		 1020
1021 /* acknowledge completion of a biofill operation */
e4d84909
DW		 1022 /* and check if we need to reply to a read request,
1023 * new R5_Wantfill requests are held off until
83de75cc 1024 * !STRIPE_BIOFILL_RUN
e4d84909
DW		 1025 */
1026 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
91c00924 1027 struct bio *rbi, *rbi2;
91c00924 1028
91c00924
DW		 1029 BUG_ON(!dev->read);
1030 rbi = dev->read;
1031 dev->read = NULL;
4f024f37 1032 while (rbi && rbi->bi_iter.bi_sector <
91c00924
DW		 1033 dev->sector + STRIPE_SECTORS) {
1034 rbi2 = r5_next_bio(rbi, dev->sector);
e7836bd6 1035 if (!raid5_dec_bi_active_stripes(rbi)) {
91c00924
DW		 1036 rbi->bi_next = return_bi;
1037 return_bi = rbi;
1038 }
91c00924
DW		 1039 rbi = rbi2;
1040 }
1041 }
1042 }
83de75cc 1043 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
91c00924
DW		 1044
1045 return_io(return_bi);
1046
e4d84909 1047 set_bit(STRIPE_HANDLE, &sh->state);
91c00924
DW		 1048 release_stripe(sh);
1049}
1050
1051static void ops_run_biofill(struct stripe_head *sh)
1052{
1053 struct dma_async_tx_descriptor *tx = NULL;
a08abd8c 1054 struct async_submit_ctl submit;
91c00924
DW		 1055 int i;
1056
e46b272b 1057 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1058 (unsigned long long)sh->sector);
1059
1060 for (i = sh->disks; i--; ) {
1061 struct r5dev *dev = &sh->dev[i];
1062 if (test_bit(R5_Wantfill, &dev->flags)) {
1063 struct bio *rbi;
b17459c0 1064 spin_lock_irq(&sh->stripe_lock);
91c00924
DW		 1065 dev->read = rbi = dev->toread;
1066 dev->toread = NULL;
b17459c0 1067 spin_unlock_irq(&sh->stripe_lock);
4f024f37 1068 while (rbi && rbi->bi_iter.bi_sector <
91c00924 1069 dev->sector + STRIPE_SECTORS) {
d592a996
SL		 1070 tx = async_copy_data(0, rbi, &dev->page,
1071 dev->sector, tx, sh);
91c00924
DW		 1072 rbi = r5_next_bio(rbi, dev->sector);
1073 }
1074 }
1075 }
1076
1077 atomic_inc(&sh->count);
a08abd8c
DW		 1078 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1079 async_trigger_callback(&submit);
91c00924
DW		 1080}
1081
4e7d2c0a 1082static void mark_target_uptodate(struct stripe_head *sh, int target)
91c00924 1083{
4e7d2c0a 1084 struct r5dev *tgt;
91c00924 1085
4e7d2c0a
DW		 1086 if (target < 0)
1087 return;
91c00924 1088
4e7d2c0a 1089 tgt = &sh->dev[target];
91c00924
DW		 1090 set_bit(R5_UPTODATE, &tgt->flags);
1091 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1092 clear_bit(R5_Wantcompute, &tgt->flags);
4e7d2c0a
DW		 1093}
1094
ac6b53b6 1095static void ops_complete_compute(void *stripe_head_ref)
91c00924
DW		 1096{
1097 struct stripe_head *sh = stripe_head_ref;
91c00924 1098
e46b272b 1099 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1100 (unsigned long long)sh->sector);
1101
ac6b53b6 1102 /* mark the computed target(s) as uptodate */
4e7d2c0a 1103 mark_target_uptodate(sh, sh->ops.target);
ac6b53b6 1104 mark_target_uptodate(sh, sh->ops.target2);
4e7d2c0a 1105
ecc65c9b
DW		 1106 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1107 if (sh->check_state == check_state_compute_run)
1108 sh->check_state = check_state_compute_result;
91c00924
DW		 1109 set_bit(STRIPE_HANDLE, &sh->state);
1110 release_stripe(sh);
1111}
1112
d6f38f31
DW		 1113/* return a pointer to the address conversion region of the scribble buffer */
1114static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1115 struct raid5_percpu *percpu)
1116{
1117 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
1118}
1119
1120static struct dma_async_tx_descriptor *
1121ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
91c00924 1122{
91c00924 1123 int disks = sh->disks;
d6f38f31 1124 struct page **xor_srcs = percpu->scribble;
91c00924
DW		 1125 int target = sh->ops.target;
1126 struct r5dev *tgt = &sh->dev[target];
1127 struct page *xor_dest = tgt->page;
1128 int count = 0;
1129 struct dma_async_tx_descriptor *tx;
a08abd8c 1130 struct async_submit_ctl submit;
91c00924
DW		 1131 int i;
1132
1133 pr_debug("%s: stripe %llu block: %d\n",
e46b272b 1134 __func__, (unsigned long long)sh->sector, target);
91c00924
DW		 1135 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1136
1137 for (i = disks; i--; )
1138 if (i != target)
1139 xor_srcs[count++] = sh->dev[i].page;
1140
1141 atomic_inc(&sh->count);
1142
0403e382 1143 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
ac6b53b6 1144 ops_complete_compute, sh, to_addr_conv(sh, percpu));
91c00924 1145 if (unlikely(count == 1))
a08abd8c 1146 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
91c00924 1147 else
a08abd8c 1148 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
91c00924 1149
91c00924
DW		 1150 return tx;
1151}
1152
ac6b53b6
DW		 1153/* set_syndrome_sources - populate source buffers for gen_syndrome
1154 * @srcs - (struct page *) array of size sh->disks
1155 * @sh - stripe_head to parse
1156 *
1157 * Populates srcs in proper layout order for the stripe and returns the
1158 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1159 * destination buffer is recorded in srcs[count] and the Q destination
1160 * is recorded in srcs[count+1]].
1161 */
1162static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
1163{
1164 int disks = sh->disks;
1165 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1166 int d0_idx = raid6_d0(sh);
1167 int count;
1168 int i;
1169
1170 for (i = 0; i < disks; i++)
5dd33c9a 1171 srcs[i] = NULL;
ac6b53b6
DW		 1172
1173 count = 0;
1174 i = d0_idx;
1175 do {
1176 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1177
1178 srcs[slot] = sh->dev[i].page;
1179 i = raid6_next_disk(i, disks);
1180 } while (i != d0_idx);
ac6b53b6 1181
e4424fee 1182 return syndrome_disks;
ac6b53b6
DW		 1183}
1184
1185static struct dma_async_tx_descriptor *
1186ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1187{
1188 int disks = sh->disks;
1189 struct page **blocks = percpu->scribble;
1190 int target;
1191 int qd_idx = sh->qd_idx;
1192 struct dma_async_tx_descriptor *tx;
1193 struct async_submit_ctl submit;
1194 struct r5dev *tgt;
1195 struct page *dest;
1196 int i;
1197 int count;
1198
1199 if (sh->ops.target < 0)
1200 target = sh->ops.target2;
1201 else if (sh->ops.target2 < 0)
1202 target = sh->ops.target;
91c00924 1203 else
ac6b53b6
DW		 1204 /* we should only have one valid target */
1205 BUG();
1206 BUG_ON(target < 0);
1207 pr_debug("%s: stripe %llu block: %d\n",
1208 __func__, (unsigned long long)sh->sector, target);
1209
1210 tgt = &sh->dev[target];
1211 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1212 dest = tgt->page;
1213
1214 atomic_inc(&sh->count);
1215
1216 if (target == qd_idx) {
1217 count = set_syndrome_sources(blocks, sh);
1218 blocks[count] = NULL; /* regenerating p is not necessary */
1219 BUG_ON(blocks[count+1] != dest); /* q should already be set */
0403e382
DW		 1220 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1221 ops_complete_compute, sh,
ac6b53b6
DW		 1222 to_addr_conv(sh, percpu));
1223 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1224 } else {
1225 /* Compute any data- or p-drive using XOR */
1226 count = 0;
1227 for (i = disks; i-- ; ) {
1228 if (i == target || i == qd_idx)
1229 continue;
1230 blocks[count++] = sh->dev[i].page;
1231 }
1232
0403e382
DW		 1233 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1234 NULL, ops_complete_compute, sh,
ac6b53b6
DW		 1235 to_addr_conv(sh, percpu));
1236 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1237 }
91c00924 1238
91c00924
DW		 1239 return tx;
1240}
1241
ac6b53b6
DW		 1242static struct dma_async_tx_descriptor *
1243ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1244{
1245 int i, count, disks = sh->disks;
1246 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1247 int d0_idx = raid6_d0(sh);
1248 int faila = -1, failb = -1;
1249 int target = sh->ops.target;
1250 int target2 = sh->ops.target2;
1251 struct r5dev *tgt = &sh->dev[target];
1252 struct r5dev *tgt2 = &sh->dev[target2];
1253 struct dma_async_tx_descriptor *tx;
1254 struct page **blocks = percpu->scribble;
1255 struct async_submit_ctl submit;
1256
1257 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1258 __func__, (unsigned long long)sh->sector, target, target2);
1259 BUG_ON(target < 0 || target2 < 0);
1260 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1261 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1262
6c910a78 1263 /* we need to open-code set_syndrome_sources to handle the
ac6b53b6
DW		 1264 * slot number conversion for 'faila' and 'failb'
1265 */
1266 for (i = 0; i < disks ; i++)
5dd33c9a 1267 blocks[i] = NULL;
ac6b53b6
DW		 1268 count = 0;
1269 i = d0_idx;
1270 do {
1271 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1272
1273 blocks[slot] = sh->dev[i].page;
1274
1275 if (i == target)
1276 faila = slot;
1277 if (i == target2)
1278 failb = slot;
1279 i = raid6_next_disk(i, disks);
1280 } while (i != d0_idx);
ac6b53b6
DW		 1281
1282 BUG_ON(faila == failb);
1283 if (failb < faila)
1284 swap(faila, failb);
1285 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1286 __func__, (unsigned long long)sh->sector, faila, failb);
1287
1288 atomic_inc(&sh->count);
1289
1290 if (failb == syndrome_disks+1) {
1291 /* Q disk is one of the missing disks */
1292 if (faila == syndrome_disks) {
1293 /* Missing P+Q, just recompute */
0403e382
DW		 1294 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1295 ops_complete_compute, sh,
1296 to_addr_conv(sh, percpu));
e4424fee 1297 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
ac6b53b6
DW		 1298 STRIPE_SIZE, &submit);
1299 } else {
1300 struct page *dest;
1301 int data_target;
1302 int qd_idx = sh->qd_idx;
1303
1304 /* Missing D+Q: recompute D from P, then recompute Q */
1305 if (target == qd_idx)
1306 data_target = target2;
1307 else
1308 data_target = target;
1309
1310 count = 0;
1311 for (i = disks; i-- ; ) {
1312 if (i == data_target || i == qd_idx)
1313 continue;
1314 blocks[count++] = sh->dev[i].page;
1315 }
1316 dest = sh->dev[data_target].page;
0403e382
DW		 1317 init_async_submit(&submit,
1318 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1319 NULL, NULL, NULL,
1320 to_addr_conv(sh, percpu));
ac6b53b6
DW		 1321 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1322 &submit);
1323
1324 count = set_syndrome_sources(blocks, sh);
0403e382
DW		 1325 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1326 ops_complete_compute, sh,
1327 to_addr_conv(sh, percpu));
ac6b53b6
DW		 1328 return async_gen_syndrome(blocks, 0, count+2,
1329 STRIPE_SIZE, &submit);
1330 }
ac6b53b6 1331 } else {
6c910a78
DW		 1332 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1333 ops_complete_compute, sh,
1334 to_addr_conv(sh, percpu));
1335 if (failb == syndrome_disks) {
1336 /* We're missing D+P. */
1337 return async_raid6_datap_recov(syndrome_disks+2,
1338 STRIPE_SIZE, faila,
1339 blocks, &submit);
1340 } else {
1341 /* We're missing D+D. */
1342 return async_raid6_2data_recov(syndrome_disks+2,
1343 STRIPE_SIZE, faila, failb,
1344 blocks, &submit);
1345 }
ac6b53b6
DW		 1346 }
1347}
1348
1349
91c00924
DW		 1350static void ops_complete_prexor(void *stripe_head_ref)
1351{
1352 struct stripe_head *sh = stripe_head_ref;
1353
e46b272b 1354 pr_debug("%s: stripe %llu\n", __func__,
91c00924 1355 (unsigned long long)sh->sector);
91c00924
DW		 1356}
1357
1358static struct dma_async_tx_descriptor *
d6f38f31
DW		 1359ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1360 struct dma_async_tx_descriptor *tx)
91c00924 1361{
91c00924 1362 int disks = sh->disks;
d6f38f31 1363 struct page **xor_srcs = percpu->scribble;
91c00924 1364 int count = 0, pd_idx = sh->pd_idx, i;
a08abd8c 1365 struct async_submit_ctl submit;
91c00924
DW		 1366
1367 /* existing parity data subtracted */
1368 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1369
e46b272b 1370 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1371 (unsigned long long)sh->sector);
1372
1373 for (i = disks; i--; ) {
1374 struct r5dev *dev = &sh->dev[i];
1375 /* Only process blocks that are known to be uptodate */
d8ee0728 1376 if (test_bit(R5_Wantdrain, &dev->flags))
91c00924
DW		 1377 xor_srcs[count++] = dev->page;
1378 }
1379
0403e382 1380 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
d6f38f31 1381 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
a08abd8c 1382 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
91c00924
DW		 1383
1384 return tx;
1385}
1386
1387static struct dma_async_tx_descriptor *
d8ee0728 1388ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
91c00924
DW		 1389{
1390 int disks = sh->disks;
d8ee0728 1391 int i;
91c00924 1392
e46b272b 1393 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1394 (unsigned long long)sh->sector);
1395
1396 for (i = disks; i--; ) {
1397 struct r5dev *dev = &sh->dev[i];
1398 struct bio *chosen;
91c00924 1399
d8ee0728 1400 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
91c00924
DW		 1401 struct bio *wbi;
1402
b17459c0 1403 spin_lock_irq(&sh->stripe_lock);
91c00924
DW		 1404 chosen = dev->towrite;
1405 dev->towrite = NULL;
1406 BUG_ON(dev->written);
1407 wbi = dev->written = chosen;
b17459c0 1408 spin_unlock_irq(&sh->stripe_lock);
d592a996 1409 WARN_ON(dev->page != dev->orig_page);
91c00924 1410
4f024f37 1411 while (wbi && wbi->bi_iter.bi_sector <
91c00924 1412 dev->sector + STRIPE_SECTORS) {
e9c7469b
TH		 1413 if (wbi->bi_rw & REQ_FUA)
1414 set_bit(R5_WantFUA, &dev->flags);
bc0934f0
SL		 1415 if (wbi->bi_rw & REQ_SYNC)
1416 set_bit(R5_SyncIO, &dev->flags);
9e444768 1417 if (wbi->bi_rw & REQ_DISCARD)
620125f2 1418 set_bit(R5_Discard, &dev->flags);
d592a996
SL		 1419 else {
1420 tx = async_copy_data(1, wbi, &dev->page,
1421 dev->sector, tx, sh);
1422 if (dev->page != dev->orig_page) {
1423 set_bit(R5_SkipCopy, &dev->flags);
1424 clear_bit(R5_UPTODATE, &dev->flags);
1425 clear_bit(R5_OVERWRITE, &dev->flags);
1426 }
1427 }
91c00924
DW		 1428 wbi = r5_next_bio(wbi, dev->sector);
1429 }
1430 }
1431 }
1432
1433 return tx;
1434}
1435
ac6b53b6 1436static void ops_complete_reconstruct(void *stripe_head_ref)
91c00924
DW		 1437{
1438 struct stripe_head *sh = stripe_head_ref;
ac6b53b6
DW		 1439 int disks = sh->disks;
1440 int pd_idx = sh->pd_idx;
1441 int qd_idx = sh->qd_idx;
1442 int i;
9e444768 1443 bool fua = false, sync = false, discard = false;
91c00924 1444
e46b272b 1445 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1446 (unsigned long long)sh->sector);
1447
bc0934f0 1448 for (i = disks; i--; ) {
e9c7469b 1449 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
bc0934f0 1450 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
9e444768 1451 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
bc0934f0 1452 }
e9c7469b 1453
91c00924
DW		 1454 for (i = disks; i--; ) {
1455 struct r5dev *dev = &sh->dev[i];
ac6b53b6 1456
e9c7469b 1457 if (dev->written || i == pd_idx || i == qd_idx) {
d592a996 1458 if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
9e444768 1459 set_bit(R5_UPTODATE, &dev->flags);
e9c7469b
TH		 1460 if (fua)
1461 set_bit(R5_WantFUA, &dev->flags);
bc0934f0
SL		 1462 if (sync)
1463 set_bit(R5_SyncIO, &dev->flags);
e9c7469b 1464 }
91c00924
DW		 1465 }
1466
d8ee0728
DW		 1467 if (sh->reconstruct_state == reconstruct_state_drain_run)
1468 sh->reconstruct_state = reconstruct_state_drain_result;
1469 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1470 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1471 else {
1472 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1473 sh->reconstruct_state = reconstruct_state_result;
1474 }
91c00924
DW		 1475
1476 set_bit(STRIPE_HANDLE, &sh->state);
1477 release_stripe(sh);
1478}
1479
1480static void
ac6b53b6
DW		 1481ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1482 struct dma_async_tx_descriptor *tx)
91c00924 1483{
91c00924 1484 int disks = sh->disks;
d6f38f31 1485 struct page **xor_srcs = percpu->scribble;
a08abd8c 1486 struct async_submit_ctl submit;
91c00924
DW		 1487 int count = 0, pd_idx = sh->pd_idx, i;
1488 struct page *xor_dest;
d8ee0728 1489 int prexor = 0;
91c00924 1490 unsigned long flags;
91c00924 1491
e46b272b 1492 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1493 (unsigned long long)sh->sector);
1494
620125f2
SL		 1495 for (i = 0; i < sh->disks; i++) {
1496 if (pd_idx == i)
1497 continue;
1498 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1499 break;
1500 }
1501 if (i >= sh->disks) {
1502 atomic_inc(&sh->count);
620125f2
SL		 1503 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1504 ops_complete_reconstruct(sh);
1505 return;
1506 }
91c00924
DW		 1507 /* check if prexor is active which means only process blocks
1508 * that are part of a read-modify-write (written)
1509 */
d8ee0728
DW		 1510 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1511 prexor = 1;
91c00924
DW		 1512 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1513 for (i = disks; i--; ) {
1514 struct r5dev *dev = &sh->dev[i];
1515 if (dev->written)
1516 xor_srcs[count++] = dev->page;
1517 }
1518 } else {
1519 xor_dest = sh->dev[pd_idx].page;
1520 for (i = disks; i--; ) {
1521 struct r5dev *dev = &sh->dev[i];
1522 if (i != pd_idx)
1523 xor_srcs[count++] = dev->page;
1524 }
1525 }
1526
91c00924
DW		 1527 /* 1/ if we prexor'd then the dest is reused as a source
1528 * 2/ if we did not prexor then we are redoing the parity
1529 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1530 * for the synchronous xor case
1531 */
88ba2aa5 1532 flags = ASYNC_TX_ACK |
91c00924
DW		 1533 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1534
1535 atomic_inc(&sh->count);
1536
ac6b53b6 1537 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
d6f38f31 1538 to_addr_conv(sh, percpu));
a08abd8c
DW		 1539 if (unlikely(count == 1))
1540 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1541 else
1542 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
91c00924
DW		 1543}
1544
ac6b53b6
DW		 1545static void
1546ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1547 struct dma_async_tx_descriptor *tx)
1548{
1549 struct async_submit_ctl submit;
1550 struct page **blocks = percpu->scribble;
620125f2 1551 int count, i;
ac6b53b6
DW		 1552
1553 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1554
620125f2
SL		 1555 for (i = 0; i < sh->disks; i++) {
1556 if (sh->pd_idx == i || sh->qd_idx == i)
1557 continue;
1558 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1559 break;
1560 }
1561 if (i >= sh->disks) {
1562 atomic_inc(&sh->count);
620125f2
SL		 1563 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1564 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1565 ops_complete_reconstruct(sh);
1566 return;
1567 }
1568
ac6b53b6
DW		 1569 count = set_syndrome_sources(blocks, sh);
1570
1571 atomic_inc(&sh->count);
1572
1573 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1574 sh, to_addr_conv(sh, percpu));
1575 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
91c00924
DW		 1576}
1577
1578static void ops_complete_check(void *stripe_head_ref)
1579{
1580 struct stripe_head *sh = stripe_head_ref;
91c00924 1581
e46b272b 1582 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1583 (unsigned long long)sh->sector);
1584
ecc65c9b 1585 sh->check_state = check_state_check_result;
91c00924
DW		 1586 set_bit(STRIPE_HANDLE, &sh->state);
1587 release_stripe(sh);
1588}
1589
ac6b53b6 1590static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
91c00924 1591{
91c00924 1592 int disks = sh->disks;
ac6b53b6
DW		 1593 int pd_idx = sh->pd_idx;
1594 int qd_idx = sh->qd_idx;
1595 struct page *xor_dest;
d6f38f31 1596 struct page **xor_srcs = percpu->scribble;
91c00924 1597 struct dma_async_tx_descriptor *tx;
a08abd8c 1598 struct async_submit_ctl submit;
ac6b53b6
DW		 1599 int count;
1600 int i;
91c00924 1601
e46b272b 1602 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1603 (unsigned long long)sh->sector);
1604
ac6b53b6
DW		 1605 count = 0;
1606 xor_dest = sh->dev[pd_idx].page;
1607 xor_srcs[count++] = xor_dest;
91c00924 1608 for (i = disks; i--; ) {
ac6b53b6
DW		 1609 if (i == pd_idx || i == qd_idx)
1610 continue;
1611 xor_srcs[count++] = sh->dev[i].page;
91c00924
DW		 1612 }
1613
d6f38f31
DW		 1614 init_async_submit(&submit, 0, NULL, NULL, NULL,
1615 to_addr_conv(sh, percpu));
099f53cb 1616 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
a08abd8c 1617 &sh->ops.zero_sum_result, &submit);
91c00924 1618
91c00924 1619 atomic_inc(&sh->count);
a08abd8c
DW		 1620 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1621 tx = async_trigger_callback(&submit);
91c00924
DW		 1622}
1623
ac6b53b6
DW		 1624static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1625{
1626 struct page **srcs = percpu->scribble;
1627 struct async_submit_ctl submit;
1628 int count;
1629
1630 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1631 (unsigned long long)sh->sector, checkp);
1632
1633 count = set_syndrome_sources(srcs, sh);
1634 if (!checkp)
1635 srcs[count] = NULL;
91c00924 1636
91c00924 1637 atomic_inc(&sh->count);
ac6b53b6
DW		 1638 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1639 sh, to_addr_conv(sh, percpu));
1640 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1641 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
91c00924
DW		 1642}
1643
51acbcec 1644static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
91c00924
DW		 1645{
1646 int overlap_clear = 0, i, disks = sh->disks;
1647 struct dma_async_tx_descriptor *tx = NULL;
d1688a6d 1648 struct r5conf *conf = sh->raid_conf;
ac6b53b6 1649 int level = conf->level;
d6f38f31
DW		 1650 struct raid5_percpu *percpu;
1651 unsigned long cpu;
91c00924 1652
d6f38f31
DW		 1653 cpu = get_cpu();
1654 percpu = per_cpu_ptr(conf->percpu, cpu);
83de75cc 1655 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
91c00924
DW		 1656 ops_run_biofill(sh);
1657 overlap_clear++;
1658 }
1659
7b3a871e 1660 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
ac6b53b6
DW		 1661 if (level < 6)
1662 tx = ops_run_compute5(sh, percpu);
1663 else {
1664 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1665 tx = ops_run_compute6_1(sh, percpu);
1666 else
1667 tx = ops_run_compute6_2(sh, percpu);
1668 }
1669 /* terminate the chain if reconstruct is not set to be run */
1670 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
7b3a871e
DW		 1671 async_tx_ack(tx);
1672 }
91c00924 1673
600aa109 1674 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
d6f38f31 1675 tx = ops_run_prexor(sh, percpu, tx);
91c00924 1676
600aa109 1677 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
d8ee0728 1678 tx = ops_run_biodrain(sh, tx);
91c00924
DW		 1679 overlap_clear++;
1680 }
1681
ac6b53b6
DW		 1682 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1683 if (level < 6)
1684 ops_run_reconstruct5(sh, percpu, tx);
1685 else
1686 ops_run_reconstruct6(sh, percpu, tx);
1687 }
91c00924 1688
ac6b53b6
DW		 1689 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1690 if (sh->check_state == check_state_run)
1691 ops_run_check_p(sh, percpu);
1692 else if (sh->check_state == check_state_run_q)
1693 ops_run_check_pq(sh, percpu, 0);
1694 else if (sh->check_state == check_state_run_pq)
1695 ops_run_check_pq(sh, percpu, 1);
1696 else
1697 BUG();
1698 }
91c00924 1699
91c00924
DW		 1700 if (overlap_clear)
1701 for (i = disks; i--; ) {
1702 struct r5dev *dev = &sh->dev[i];
1703 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1704 wake_up(&sh->raid_conf->wait_for_overlap);
1705 }
d6f38f31 1706 put_cpu();
91c00924
DW		 1707}
1708
566c09c5 1709static int grow_one_stripe(struct r5conf *conf, int hash)
1da177e4
LT		 1710{
1711 struct stripe_head *sh;
6ce32846 1712 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
3f294f4f
N		 1713 if (!sh)
1714 return 0;
6ce32846 1715
3f294f4f 1716 sh->raid_conf = conf;
3f294f4f 1717
b17459c0
SL		 1718 spin_lock_init(&sh->stripe_lock);
1719
e4e11e38
N		 1720 if (grow_buffers(sh)) {
1721 shrink_buffers(sh);
3f294f4f
N		 1722 kmem_cache_free(conf->slab_cache, sh);
1723 return 0;
1724 }
566c09c5 1725 sh->hash_lock_index = hash;
3f294f4f
N		 1726 /* we just created an active stripe so... */
1727 atomic_set(&sh->count, 1);
1728 atomic_inc(&conf->active_stripes);
1729 INIT_LIST_HEAD(&sh->lru);
1730 release_stripe(sh);
1731 return 1;
1732}
1733
d1688a6d 1734static int grow_stripes(struct r5conf *conf, int num)
3f294f4f 1735{
e18b890b 1736 struct kmem_cache *sc;
5e5e3e78 1737 int devs = max(conf->raid_disks, conf->previous_raid_disks);
566c09c5 1738 int hash;
1da177e4 1739
f4be6b43
N		 1740 if (conf->mddev->gendisk)
1741 sprintf(conf->cache_name[0],
1742 "raid%d-%s", conf->level, mdname(conf->mddev));
1743 else
1744 sprintf(conf->cache_name[0],
1745 "raid%d-%p", conf->level, conf->mddev);
1746 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1747
ad01c9e3
N		 1748 conf->active_name = 0;
1749 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1da177e4 1750 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
20c2df83 1751 0, 0, NULL);
1da177e4
LT		 1752 if (!sc)
1753 return 1;
1754 conf->slab_cache = sc;
ad01c9e3 1755 conf->pool_size = devs;
566c09c5
SL		 1756 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
1757 while (num--) {
1758 if (!grow_one_stripe(conf, hash))
1da177e4 1759 return 1;
566c09c5
SL		 1760 conf->max_nr_stripes++;
1761 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
1762 }
1da177e4
LT		 1763 return 0;
1764}
29269553 1765
d6f38f31
DW		 1766/**
1767 * scribble_len - return the required size of the scribble region
1768 * @num - total number of disks in the array
1769 *
1770 * The size must be enough to contain:
1771 * 1/ a struct page pointer for each device in the array +2
1772 * 2/ room to convert each entry in (1) to its corresponding dma
1773 * (dma_map_page()) or page (page_address()) address.
1774 *
1775 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1776 * calculate over all devices (not just the data blocks), using zeros in place
1777 * of the P and Q blocks.
1778 */
1779static size_t scribble_len(int num)
1780{
1781 size_t len;
1782
1783 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1784
1785 return len;
1786}
1787
d1688a6d 1788static int resize_stripes(struct r5conf *conf, int newsize)
ad01c9e3
N		 1789{
1790 /* Make all the stripes able to hold 'newsize' devices.
1791 * New slots in each stripe get 'page' set to a new page.
1792 *
1793 * This happens in stages:
1794 * 1/ create a new kmem_cache and allocate the required number of
1795 * stripe_heads.
83f0d77a 1796 * 2/ gather all the old stripe_heads and transfer the pages across
ad01c9e3
N		 1797 * to the new stripe_heads. This will have the side effect of
1798 * freezing the array as once all stripe_heads have been collected,
1799 * no IO will be possible. Old stripe heads are freed once their
1800 * pages have been transferred over, and the old kmem_cache is
1801 * freed when all stripes are done.
1802 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1803 * we simple return a failre status - no need to clean anything up.
1804 * 4/ allocate new pages for the new slots in the new stripe_heads.
1805 * If this fails, we don't bother trying the shrink the
1806 * stripe_heads down again, we just leave them as they are.
1807 * As each stripe_head is processed the new one is released into
1808 * active service.
1809 *
1810 * Once step2 is started, we cannot afford to wait for a write,
1811 * so we use GFP_NOIO allocations.
1812 */
1813 struct stripe_head *osh, *nsh;
1814 LIST_HEAD(newstripes);
1815 struct disk_info *ndisks;
d6f38f31 1816 unsigned long cpu;
b5470dc5 1817 int err;
e18b890b 1818 struct kmem_cache *sc;
ad01c9e3 1819 int i;
566c09c5 1820 int hash, cnt;
ad01c9e3
N		 1821
1822 if (newsize <= conf->pool_size)
1823 return 0; /* never bother to shrink */
1824
b5470dc5
DW		 1825 err = md_allow_write(conf->mddev);
1826 if (err)
1827 return err;
2a2275d6 1828
ad01c9e3
N		 1829 /* Step 1 */
1830 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1831 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
20c2df83 1832 0, 0, NULL);
ad01c9e3
N		 1833 if (!sc)
1834 return -ENOMEM;
1835
1836 for (i = conf->max_nr_stripes; i; i--) {
6ce32846 1837 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
ad01c9e3
N		 1838 if (!nsh)
1839 break;
1840
ad01c9e3 1841 nsh->raid_conf = conf;
cb13ff69 1842 spin_lock_init(&nsh->stripe_lock);
ad01c9e3
N		 1843
1844 list_add(&nsh->lru, &newstripes);
1845 }
1846 if (i) {
1847 /* didn't get enough, give up */
1848 while (!list_empty(&newstripes)) {
1849 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1850 list_del(&nsh->lru);
1851 kmem_cache_free(sc, nsh);
1852 }
1853 kmem_cache_destroy(sc);
1854 return -ENOMEM;
1855 }
1856 /* Step 2 - Must use GFP_NOIO now.
1857 * OK, we have enough stripes, start collecting inactive
1858 * stripes and copying them over
1859 */
566c09c5
SL		 1860 hash = 0;
1861 cnt = 0;
ad01c9e3 1862 list_for_each_entry(nsh, &newstripes, lru) {
566c09c5
SL		 1863 lock_device_hash_lock(conf, hash);
1864 wait_event_cmd(conf->wait_for_stripe,
1865 !list_empty(conf->inactive_list + hash),
1866 unlock_device_hash_lock(conf, hash),
1867 lock_device_hash_lock(conf, hash));
1868 osh = get_free_stripe(conf, hash);
1869 unlock_device_hash_lock(conf, hash);
ad01c9e3 1870 atomic_set(&nsh->count, 1);
d592a996 1871 for(i=0; i<conf->pool_size; i++) {
ad01c9e3 1872 nsh->dev[i].page = osh->dev[i].page;
d592a996
SL		 1873 nsh->dev[i].orig_page = osh->dev[i].page;
1874 }
ad01c9e3
N		 1875 for( ; i<newsize; i++)
1876 nsh->dev[i].page = NULL;
566c09c5 1877 nsh->hash_lock_index = hash;
ad01c9e3 1878 kmem_cache_free(conf->slab_cache, osh);
566c09c5
SL		 1879 cnt++;
1880 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
1881 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
1882 hash++;
1883 cnt = 0;
1884 }
ad01c9e3
N		 1885 }
1886 kmem_cache_destroy(conf->slab_cache);
1887
1888 /* Step 3.
1889 * At this point, we are holding all the stripes so the array
1890 * is completely stalled, so now is a good time to resize
d6f38f31 1891 * conf->disks and the scribble region
ad01c9e3
N		 1892 */
1893 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1894 if (ndisks) {
1895 for (i=0; i<conf->raid_disks; i++)
1896 ndisks[i] = conf->disks[i];
1897 kfree(conf->disks);
1898 conf->disks = ndisks;
1899 } else
1900 err = -ENOMEM;
1901
d6f38f31
DW		 1902 get_online_cpus();
1903 conf->scribble_len = scribble_len(newsize);
1904 for_each_present_cpu(cpu) {
1905 struct raid5_percpu *percpu;
1906 void *scribble;
1907
1908 percpu = per_cpu_ptr(conf->percpu, cpu);
1909 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1910
1911 if (scribble) {
1912 kfree(percpu->scribble);
1913 percpu->scribble = scribble;
1914 } else {
1915 err = -ENOMEM;
1916 break;
1917 }
1918 }
1919 put_online_cpus();
1920
ad01c9e3
N		 1921 /* Step 4, return new stripes to service */
1922 while(!list_empty(&newstripes)) {
1923 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1924 list_del_init(&nsh->lru);
d6f38f31 1925
ad01c9e3
N		 1926 for (i=conf->raid_disks; i < newsize; i++)
1927 if (nsh->dev[i].page == NULL) {
1928 struct page *p = alloc_page(GFP_NOIO);
1929 nsh->dev[i].page = p;
d592a996 1930 nsh->dev[i].orig_page = p;
ad01c9e3
N		 1931 if (!p)
1932 err = -ENOMEM;
1933 }
1934 release_stripe(nsh);
1935 }
1936 /* critical section pass, GFP_NOIO no longer needed */
1937
1938 conf->slab_cache = sc;
1939 conf->active_name = 1-conf->active_name;
1940 conf->pool_size = newsize;
1941 return err;
1942}
1da177e4 1943
566c09c5 1944static int drop_one_stripe(struct r5conf *conf, int hash)
1da177e4
LT		 1945{
1946 struct stripe_head *sh;
1947
566c09c5
SL		 1948 spin_lock_irq(conf->hash_locks + hash);
1949 sh = get_free_stripe(conf, hash);
1950 spin_unlock_irq(conf->hash_locks + hash);
3f294f4f
N		 1951 if (!sh)
1952 return 0;
78bafebd 1953 BUG_ON(atomic_read(&sh->count));
e4e11e38 1954 shrink_buffers(sh);
3f294f4f
N		 1955 kmem_cache_free(conf->slab_cache, sh);
1956 atomic_dec(&conf->active_stripes);
1957 return 1;
1958}
1959
d1688a6d 1960static void shrink_stripes(struct r5conf *conf)
3f294f4f 1961{
566c09c5
SL		 1962 int hash;
1963 for (hash = 0; hash < NR_STRIPE_HASH_LOCKS; hash++)
1964 while (drop_one_stripe(conf, hash))
1965 ;
3f294f4f 1966
29fc7e3e
N		 1967 if (conf->slab_cache)
1968 kmem_cache_destroy(conf->slab_cache);
1da177e4
LT		 1969 conf->slab_cache = NULL;
1970}
1971
6712ecf8 1972static void raid5_end_read_request(struct bio * bi, int error)
1da177e4 1973{
99c0fb5f 1974 struct stripe_head *sh = bi->bi_private;
d1688a6d 1975 struct r5conf *conf = sh->raid_conf;
7ecaa1e6 1976 int disks = sh->disks, i;
1da177e4 1977 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
d6950432 1978 char b[BDEVNAME_SIZE];
dd054fce 1979 struct md_rdev *rdev = NULL;
05616be5 1980 sector_t s;
1da177e4
LT		 1981
1982 for (i=0 ; i<disks; i++)
1983 if (bi == &sh->dev[i].req)
1984 break;
1985
45b4233c
DW		 1986 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1987 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1da177e4
LT		 1988 uptodate);
1989 if (i == disks) {
1990 BUG();
6712ecf8 1991 return;
1da177e4 1992 }
14a75d3e 1993 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
dd054fce
N		 1994 /* If replacement finished while this request was outstanding,
1995 * 'replacement' might be NULL already.
1996 * In that case it moved down to 'rdev'.
1997 * rdev is not removed until all requests are finished.
1998 */
14a75d3e 1999 rdev = conf->disks[i].replacement;
dd054fce 2000 if (!rdev)
14a75d3e 2001 rdev = conf->disks[i].rdev;
1da177e4 2002
05616be5
N		 2003 if (use_new_offset(conf, sh))
2004 s = sh->sector + rdev->new_data_offset;
2005 else
2006 s = sh->sector + rdev->data_offset;
1da177e4 2007 if (uptodate) {
1da177e4 2008 set_bit(R5_UPTODATE, &sh->dev[i].flags);
4e5314b5 2009 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
14a75d3e
N		 2010 /* Note that this cannot happen on a
2011 * replacement device. We just fail those on
2012 * any error
2013 */
8bda470e
CD		 2014 printk_ratelimited(
2015 KERN_INFO
2016 "md/raid:%s: read error corrected"
2017 " (%lu sectors at %llu on %s)\n",
2018 mdname(conf->mddev), STRIPE_SECTORS,
05616be5 2019 (unsigned long long)s,
8bda470e 2020 bdevname(rdev->bdev, b));
ddd5115f 2021 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
4e5314b5
N		 2022 clear_bit(R5_ReadError, &sh->dev[i].flags);
2023 clear_bit(R5_ReWrite, &sh->dev[i].flags);
3f9e7c14 2024 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2025 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2026
14a75d3e
N		 2027 if (atomic_read(&rdev->read_errors))
2028 atomic_set(&rdev->read_errors, 0);
1da177e4 2029 } else {
14a75d3e 2030 const char *bdn = bdevname(rdev->bdev, b);
ba22dcbf 2031 int retry = 0;
2e8ac303 2032 int set_bad = 0;
d6950432 2033
1da177e4 2034 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
d6950432 2035 atomic_inc(&rdev->read_errors);
14a75d3e
N		 2036 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2037 printk_ratelimited(
2038 KERN_WARNING
2039 "md/raid:%s: read error on replacement device "
2040 "(sector %llu on %s).\n",
2041 mdname(conf->mddev),
05616be5 2042 (unsigned long long)s,
14a75d3e 2043 bdn);
2e8ac303 2044 else if (conf->mddev->degraded >= conf->max_degraded) {
2045 set_bad = 1;
8bda470e
CD		 2046 printk_ratelimited(
2047 KERN_WARNING
2048 "md/raid:%s: read error not correctable "
2049 "(sector %llu on %s).\n",
2050 mdname(conf->mddev),
05616be5 2051 (unsigned long long)s,
8bda470e 2052 bdn);
2e8ac303 2053 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
4e5314b5 2054 /* Oh, no!!! */
2e8ac303 2055 set_bad = 1;
8bda470e
CD		 2056 printk_ratelimited(
2057 KERN_WARNING
2058 "md/raid:%s: read error NOT corrected!! "
2059 "(sector %llu on %s).\n",
2060 mdname(conf->mddev),
05616be5 2061 (unsigned long long)s,
8bda470e 2062 bdn);
2e8ac303 2063 } else if (atomic_read(&rdev->read_errors)
ba22dcbf 2064 > conf->max_nr_stripes)
14f8d26b 2065 printk(KERN_WARNING
0c55e022 2066 "md/raid:%s: Too many read errors, failing device %s.\n",
d6950432 2067 mdname(conf->mddev), bdn);
ba22dcbf
N		 2068 else
2069 retry = 1;
edfa1f65
BY		 2070 if (set_bad && test_bit(In_sync, &rdev->flags)
2071 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2072 retry = 1;
ba22dcbf 2073 if (retry)
3f9e7c14 2074 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2075 set_bit(R5_ReadError, &sh->dev[i].flags);
2076 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2077 } else
2078 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
ba22dcbf 2079 else {
4e5314b5
N		 2080 clear_bit(R5_ReadError, &sh->dev[i].flags);
2081 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2e8ac303 2082 if (!(set_bad
2083 && test_bit(In_sync, &rdev->flags)
2084 && rdev_set_badblocks(
2085 rdev, sh->sector, STRIPE_SECTORS, 0)))
2086 md_error(conf->mddev, rdev);
ba22dcbf 2087 }
1da177e4 2088 }
14a75d3e 2089 rdev_dec_pending(rdev, conf->mddev);
1da177e4
LT		 2090 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2091 set_bit(STRIPE_HANDLE, &sh->state);
2092 release_stripe(sh);
1da177e4
LT		 2093}
2094
d710e138 2095static void raid5_end_write_request(struct bio *bi, int error)
1da177e4 2096{
99c0fb5f 2097 struct stripe_head *sh = bi->bi_private;
d1688a6d 2098 struct r5conf *conf = sh->raid_conf;
7ecaa1e6 2099 int disks = sh->disks, i;
977df362 2100 struct md_rdev *uninitialized_var(rdev);
1da177e4 2101 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
b84db560
N		 2102 sector_t first_bad;
2103 int bad_sectors;
977df362 2104 int replacement = 0;
1da177e4 2105
977df362
N		 2106 for (i = 0 ; i < disks; i++) {
2107 if (bi == &sh->dev[i].req) {
2108 rdev = conf->disks[i].rdev;
1da177e4 2109 break;
977df362
N		 2110 }
2111 if (bi == &sh->dev[i].rreq) {
2112 rdev = conf->disks[i].replacement;
dd054fce
N		 2113 if (rdev)
2114 replacement = 1;
2115 else
2116 /* rdev was removed and 'replacement'
2117 * replaced it. rdev is not removed
2118 * until all requests are finished.
2119 */
2120 rdev = conf->disks[i].rdev;
977df362
N		 2121 break;
2122 }
2123 }
45b4233c 2124 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1da177e4
LT		 2125 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2126 uptodate);
2127 if (i == disks) {
2128 BUG();
6712ecf8 2129 return;
1da177e4
LT		 2130 }
2131
977df362
N		 2132 if (replacement) {
2133 if (!uptodate)
2134 md_error(conf->mddev, rdev);
2135 else if (is_badblock(rdev, sh->sector,
2136 STRIPE_SECTORS,
2137 &first_bad, &bad_sectors))
2138 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2139 } else {
2140 if (!uptodate) {
9f97e4b1 2141 set_bit(STRIPE_DEGRADED, &sh->state);
977df362
N		 2142 set_bit(WriteErrorSeen, &rdev->flags);
2143 set_bit(R5_WriteError, &sh->dev[i].flags);
3a6de292
N		 2144 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2145 set_bit(MD_RECOVERY_NEEDED,
2146 &rdev->mddev->recovery);
977df362
N		 2147 } else if (is_badblock(rdev, sh->sector,
2148 STRIPE_SECTORS,
c0b32972 2149 &first_bad, &bad_sectors)) {
977df362 2150 set_bit(R5_MadeGood, &sh->dev[i].flags);
c0b32972
N		 2151 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2152 /* That was a successful write so make
2153 * sure it looks like we already did
2154 * a re-write.
2155 */
2156 set_bit(R5_ReWrite, &sh->dev[i].flags);
2157 }
977df362
N		 2158 }
2159 rdev_dec_pending(rdev, conf->mddev);
1da177e4 2160
977df362
N		 2161 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2162 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1da177e4 2163 set_bit(STRIPE_HANDLE, &sh->state);
c04be0aa 2164 release_stripe(sh);
1da177e4
LT		 2165}
2166
784052ec 2167static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
d592a996 2168
784052ec 2169static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1da177e4
LT		 2170{
2171 struct r5dev *dev = &sh->dev[i];
2172
2173 bio_init(&dev->req);
2174 dev->req.bi_io_vec = &dev->vec;
d592a996 2175 dev->req.bi_max_vecs = 1;
1da177e4
LT		 2176 dev->req.bi_private = sh;
2177
977df362
N		 2178 bio_init(&dev->rreq);
2179 dev->rreq.bi_io_vec = &dev->rvec;
d592a996 2180 dev->rreq.bi_max_vecs = 1;
977df362 2181 dev->rreq.bi_private = sh;
977df362 2182
1da177e4 2183 dev->flags = 0;
784052ec 2184 dev->sector = compute_blocknr(sh, i, previous);
1da177e4
LT		 2185}
2186
fd01b88c 2187static void error(struct mddev *mddev, struct md_rdev *rdev)
1da177e4
LT		 2188{
2189 char b[BDEVNAME_SIZE];
d1688a6d 2190 struct r5conf *conf = mddev->private;
908f4fbd 2191 unsigned long flags;
0c55e022 2192 pr_debug("raid456: error called\n");
1da177e4 2193
908f4fbd
N		 2194 spin_lock_irqsave(&conf->device_lock, flags);
2195 clear_bit(In_sync, &rdev->flags);
2196 mddev->degraded = calc_degraded(conf);
2197 spin_unlock_irqrestore(&conf->device_lock, flags);
2198 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2199
de393cde 2200 set_bit(Blocked, &rdev->flags);
6f8d0c77
N		 2201 set_bit(Faulty, &rdev->flags);
2202 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2203 printk(KERN_ALERT
2204 "md/raid:%s: Disk failure on %s, disabling device.\n"
2205 "md/raid:%s: Operation continuing on %d devices.\n",
2206 mdname(mddev),
2207 bdevname(rdev->bdev, b),
2208 mdname(mddev),
2209 conf->raid_disks - mddev->degraded);
16a53ecc 2210}
1da177e4
LT		 2211
2212/*
2213 * Input: a 'big' sector number,
2214 * Output: index of the data and parity disk, and the sector # in them.
2215 */
d1688a6d 2216static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
911d4ee8
N		 2217 int previous, int *dd_idx,
2218 struct stripe_head *sh)
1da177e4 2219{
6e3b96ed 2220 sector_t stripe, stripe2;
35f2a591 2221 sector_t chunk_number;
1da177e4 2222 unsigned int chunk_offset;
911d4ee8 2223 int pd_idx, qd_idx;
67cc2b81 2224 int ddf_layout = 0;
1da177e4 2225 sector_t new_sector;
e183eaed
N		 2226 int algorithm = previous ? conf->prev_algo
2227 : conf->algorithm;
09c9e5fa
AN		 2228 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2229 : conf->chunk_sectors;
112bf897
N		 2230 int raid_disks = previous ? conf->previous_raid_disks
2231 : conf->raid_disks;
2232 int data_disks = raid_disks - conf->max_degraded;
1da177e4
LT		 2233
2234 /* First compute the information on this sector */
2235
2236 /*
2237 * Compute the chunk number and the sector offset inside the chunk
2238 */
2239 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2240 chunk_number = r_sector;
1da177e4
LT		 2241
2242 /*
2243 * Compute the stripe number
2244 */
35f2a591
N		 2245 stripe = chunk_number;
2246 *dd_idx = sector_div(stripe, data_disks);
6e3b96ed 2247 stripe2 = stripe;
1da177e4
LT		 2248 /*
2249 * Select the parity disk based on the user selected algorithm.
2250 */
84789554 2251 pd_idx = qd_idx = -1;
16a53ecc
N		 2252 switch(conf->level) {
2253 case 4:
911d4ee8 2254 pd_idx = data_disks;
16a53ecc
N		 2255 break;
2256 case 5:
e183eaed 2257 switch (algorithm) {
1da177e4 2258 case ALGORITHM_LEFT_ASYMMETRIC:
6e3b96ed 2259 pd_idx = data_disks - sector_div(stripe2, raid_disks);
911d4ee8 2260 if (*dd_idx >= pd_idx)
1da177e4
LT		 2261 (*dd_idx)++;
2262 break;
2263 case ALGORITHM_RIGHT_ASYMMETRIC:
6e3b96ed 2264 pd_idx = sector_div(stripe2, raid_disks);
911d4ee8 2265 if (*dd_idx >= pd_idx)
1da177e4
LT		 2266 (*dd_idx)++;
2267 break;
2268 case ALGORITHM_LEFT_SYMMETRIC:
6e3b96ed 2269 pd_idx = data_disks - sector_div(stripe2, raid_disks);
911d4ee8 2270 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1da177e4
LT		 2271 break;
2272 case ALGORITHM_RIGHT_SYMMETRIC:
6e3b96ed 2273 pd_idx = sector_div(stripe2, raid_disks);
911d4ee8 2274 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1da177e4 2275 break;
99c0fb5f
N		 2276 case ALGORITHM_PARITY_0:
2277 pd_idx = 0;
2278 (*dd_idx)++;
2279 break;
2280 case ALGORITHM_PARITY_N:
2281 pd_idx = data_disks;
2282 break;
1da177e4 2283 default:
99c0fb5f 2284 BUG();
16a53ecc
N		 2285 }
2286 break;
2287 case 6:
2288
e183eaed 2289 switch (algorithm) {
16a53ecc 2290 case ALGORITHM_LEFT_ASYMMETRIC:
6e3b96ed 2291 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
911d4ee8
N		 2292 qd_idx = pd_idx + 1;
2293 if (pd_idx == raid_disks-1) {
99c0fb5f 2294 (*dd_idx)++; /* Q D D D P */
911d4ee8
N		 2295 qd_idx = 0;
2296 } else if (*dd_idx >= pd_idx)
16a53ecc
N		 2297 (*dd_idx) += 2; /* D D P Q D */
2298 break;
2299 case ALGORITHM_RIGHT_ASYMMETRIC:
6e3b96ed 2300 pd_idx = sector_div(stripe2, raid_disks);
911d4ee8
N		 2301 qd_idx = pd_idx + 1;
2302 if (pd_idx == raid_disks-1) {
99c0fb5f 2303 (*dd_idx)++; /* Q D D D P */
911d4ee8
N		 2304 qd_idx = 0;
2305 } else if (*dd_idx >= pd_idx)
16a53ecc
N		 2306 (*dd_idx) += 2; /* D D P Q D */
2307 break;
2308 case ALGORITHM_LEFT_SYMMETRIC:
6e3b96ed 2309 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
911d4ee8
N		 2310 qd_idx = (pd_idx + 1) % raid_disks;
2311 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
16a53ecc
N		 2312 break;
2313 case ALGORITHM_RIGHT_SYMMETRIC:
6e3b96ed 2314 pd_idx = sector_div(stripe2, raid_disks);
911d4ee8
N		 2315 qd_idx = (pd_idx + 1) % raid_disks;
2316 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
16a53ecc 2317 break;
99c0fb5f
N		 2318
2319 case ALGORITHM_PARITY_0:
2320 pd_idx = 0;
2321 qd_idx = 1;
2322 (*dd_idx) += 2;
2323 break;
2324 case ALGORITHM_PARITY_N:
2325 pd_idx = data_disks;
2326 qd_idx = data_disks + 1;
2327 break;
2328
2329 case ALGORITHM_ROTATING_ZERO_RESTART:
2330 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2331 * of blocks for computing Q is different.
2332 */
6e3b96ed 2333 pd_idx = sector_div(stripe2, raid_disks);
99c0fb5f
N		 2334 qd_idx = pd_idx + 1;
2335 if (pd_idx == raid_disks-1) {
2336 (*dd_idx)++; /* Q D D D P */
2337 qd_idx = 0;
2338 } else if (*dd_idx >= pd_idx)
2339 (*dd_idx) += 2; /* D D P Q D */
67cc2b81 2340 ddf_layout = 1;
99c0fb5f
N		 2341 break;
2342
2343 case ALGORITHM_ROTATING_N_RESTART:
2344 /* Same a left_asymmetric, by first stripe is
2345 * D D D P Q rather than
2346 * Q D D D P
2347 */
6e3b96ed
N		 2348 stripe2 += 1;
2349 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
99c0fb5f
N		 2350 qd_idx = pd_idx + 1;
2351 if (pd_idx == raid_disks-1) {
2352 (*dd_idx)++; /* Q D D D P */
2353 qd_idx = 0;
2354 } else if (*dd_idx >= pd_idx)
2355 (*dd_idx) += 2; /* D D P Q D */
67cc2b81 2356 ddf_layout = 1;
99c0fb5f
N		 2357 break;
2358
2359 case ALGORITHM_ROTATING_N_CONTINUE:
2360 /* Same as left_symmetric but Q is before P */
6e3b96ed 2361 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
99c0fb5f
N		 2362 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2363 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
67cc2b81 2364 ddf_layout = 1;
99c0fb5f
N		 2365 break;
2366
2367 case ALGORITHM_LEFT_ASYMMETRIC_6:
2368 /* RAID5 left_asymmetric, with Q on last device */
6e3b96ed 2369 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
99c0fb5f
N		 2370 if (*dd_idx >= pd_idx)
2371 (*dd_idx)++;
2372 qd_idx = raid_disks - 1;
2373 break;
2374
2375 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6e3b96ed 2376 pd_idx = sector_div(stripe2, raid_disks-1);
99c0fb5f
N		 2377 if (*dd_idx >= pd_idx)
2378 (*dd_idx)++;
2379 qd_idx = raid_disks - 1;
2380 break;
2381
2382 case ALGORITHM_LEFT_SYMMETRIC_6:
6e3b96ed 2383 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
99c0fb5f
N		 2384 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2385 qd_idx = raid_disks - 1;
2386 break;
2387
2388 case ALGORITHM_RIGHT_SYMMETRIC_6:
6e3b96ed 2389 pd_idx = sector_div(stripe2, raid_disks-1);
99c0fb5f
N		 2390 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2391 qd_idx = raid_disks - 1;
2392 break;
2393
2394 case ALGORITHM_PARITY_0_6:
2395 pd_idx = 0;
2396 (*dd_idx)++;
2397 qd_idx = raid_disks - 1;
2398 break;
2399
16a53ecc 2400 default:
99c0fb5f 2401 BUG();
16a53ecc
N		 2402 }
2403 break;
1da177e4
LT		 2404 }
2405
911d4ee8
N		 2406 if (sh) {
2407 sh->pd_idx = pd_idx;
2408 sh->qd_idx = qd_idx;
67cc2b81 2409 sh->ddf_layout = ddf_layout;
911d4ee8 2410 }
1da177e4
LT		 2411 /*
2412 * Finally, compute the new sector number
2413 */
2414 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2415 return new_sector;
2416}
2417
2418
784052ec 2419static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1da177e4 2420{
d1688a6d 2421 struct r5conf *conf = sh->raid_conf;
b875e531
N		 2422 int raid_disks = sh->disks;
2423 int data_disks = raid_disks - conf->max_degraded;
1da177e4 2424 sector_t new_sector = sh->sector, check;
09c9e5fa
AN		 2425 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2426 : conf->chunk_sectors;
e183eaed
N		 2427 int algorithm = previous ? conf->prev_algo
2428 : conf->algorithm;
1da177e4
LT		 2429 sector_t stripe;
2430 int chunk_offset;
35f2a591
N		 2431 sector_t chunk_number;
2432 int dummy1, dd_idx = i;
1da177e4 2433 sector_t r_sector;
911d4ee8 2434 struct stripe_head sh2;
1da177e4 2435
16a53ecc 2436
1da177e4
LT		 2437 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2438 stripe = new_sector;
1da177e4 2439
16a53ecc
N		 2440 if (i == sh->pd_idx)
2441 return 0;
2442 switch(conf->level) {
2443 case 4: break;
2444 case 5:
e183eaed 2445 switch (algorithm) {
1da177e4
LT		 2446 case ALGORITHM_LEFT_ASYMMETRIC:
2447 case ALGORITHM_RIGHT_ASYMMETRIC:
2448 if (i > sh->pd_idx)
2449 i--;
2450 break;
2451 case ALGORITHM_LEFT_SYMMETRIC:
2452 case ALGORITHM_RIGHT_SYMMETRIC:
2453 if (i < sh->pd_idx)
2454 i += raid_disks;
2455 i -= (sh->pd_idx + 1);
2456 break;
99c0fb5f
N		 2457 case ALGORITHM_PARITY_0:
2458 i -= 1;
2459 break;
2460 case ALGORITHM_PARITY_N:
2461 break;
1da177e4 2462 default:
99c0fb5f 2463 BUG();
16a53ecc
N		 2464 }
2465 break;
2466 case 6:
d0dabf7e 2467 if (i == sh->qd_idx)
16a53ecc 2468 return 0; /* It is the Q disk */
e183eaed 2469 switch (algorithm) {
16a53ecc
N		 2470 case ALGORITHM_LEFT_ASYMMETRIC:
2471 case ALGORITHM_RIGHT_ASYMMETRIC:
99c0fb5f
N		 2472 case ALGORITHM_ROTATING_ZERO_RESTART:
2473 case ALGORITHM_ROTATING_N_RESTART:
2474 if (sh->pd_idx == raid_disks-1)
2475 i--; /* Q D D D P */
16a53ecc
N		 2476 else if (i > sh->pd_idx)
2477 i -= 2; /* D D P Q D */
2478 break;
2479 case ALGORITHM_LEFT_SYMMETRIC:
2480 case ALGORITHM_RIGHT_SYMMETRIC:
2481 if (sh->pd_idx == raid_disks-1)
2482 i--; /* Q D D D P */
2483 else {
2484 /* D D P Q D */
2485 if (i < sh->pd_idx)
2486 i += raid_disks;
2487 i -= (sh->pd_idx + 2);
2488 }
2489 break;
99c0fb5f
N		 2490 case ALGORITHM_PARITY_0:
2491 i -= 2;
2492 break;
2493 case ALGORITHM_PARITY_N:
2494 break;
2495 case ALGORITHM_ROTATING_N_CONTINUE:
e4424fee 2496 /* Like left_symmetric, but P is before Q */
99c0fb5f
N		 2497 if (sh->pd_idx == 0)
2498 i--; /* P D D D Q */
e4424fee
N		 2499 else {
2500 /* D D Q P D */
2501 if (i < sh->pd_idx)
2502 i += raid_disks;
2503 i -= (sh->pd_idx + 1);
2504 }
99c0fb5f
N		 2505 break;
2506 case ALGORITHM_LEFT_ASYMMETRIC_6:
2507 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2508 if (i > sh->pd_idx)
2509 i--;
2510 break;
2511 case ALGORITHM_LEFT_SYMMETRIC_6:
2512 case ALGORITHM_RIGHT_SYMMETRIC_6:
2513 if (i < sh->pd_idx)
2514 i += data_disks + 1;
2515 i -= (sh->pd_idx + 1);
2516 break;
2517 case ALGORITHM_PARITY_0_6:
2518 i -= 1;
2519 break;
16a53ecc 2520 default:
99c0fb5f 2521 BUG();
16a53ecc
N		 2522 }
2523 break;
1da177e4
LT		 2524 }
2525
2526 chunk_number = stripe * data_disks + i;
35f2a591 2527 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
1da177e4 2528
112bf897 2529 check = raid5_compute_sector(conf, r_sector,
784052ec 2530 previous, &dummy1, &sh2);
911d4ee8
N		 2531 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2532 || sh2.qd_idx != sh->qd_idx) {
0c55e022
N		 2533 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2534 mdname(conf->mddev));
1da177e4
LT		 2535 return 0;
2536 }
2537 return r_sector;
2538}
2539
2540
600aa109 2541static void
c0f7bddb 2542schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
600aa109 2543 int rcw, int expand)
e33129d8
DW		 2544{
2545 int i, pd_idx = sh->pd_idx, disks = sh->disks;
d1688a6d 2546 struct r5conf *conf = sh->raid_conf;
c0f7bddb 2547 int level = conf->level;
e33129d8
DW		 2548
2549 if (rcw) {
e33129d8
DW		 2550
2551 for (i = disks; i--; ) {
2552 struct r5dev *dev = &sh->dev[i];
2553
2554 if (dev->towrite) {
2555 set_bit(R5_LOCKED, &dev->flags);
d8ee0728 2556 set_bit(R5_Wantdrain, &dev->flags);
e33129d8
DW		 2557 if (!expand)
2558 clear_bit(R5_UPTODATE, &dev->flags);
600aa109 2559 s->locked++;
e33129d8
DW		 2560 }
2561 }
ce7d363a
N		 2562 /* if we are not expanding this is a proper write request, and
2563 * there will be bios with new data to be drained into the
2564 * stripe cache
2565 */
2566 if (!expand) {
2567 if (!s->locked)
2568 /* False alarm, nothing to do */
2569 return;
2570 sh->reconstruct_state = reconstruct_state_drain_run;
2571 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2572 } else
2573 sh->reconstruct_state = reconstruct_state_run;
2574
2575 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2576
c0f7bddb 2577 if (s->locked + conf->max_degraded == disks)
8b3e6cdc 2578 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
c0f7bddb 2579 atomic_inc(&conf->pending_full_writes);
e33129d8 2580 } else {
c0f7bddb 2581 BUG_ON(level == 6);
e33129d8
DW		 2582 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2583 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2584
e33129d8
DW		 2585 for (i = disks; i--; ) {
2586 struct r5dev *dev = &sh->dev[i];
2587 if (i == pd_idx)
2588 continue;
2589
e33129d8
DW		 2590 if (dev->towrite &&
2591 (test_bit(R5_UPTODATE, &dev->flags) ||
d8ee0728
DW		 2592 test_bit(R5_Wantcompute, &dev->flags))) {
2593 set_bit(R5_Wantdrain, &dev->flags);
e33129d8
DW		 2594 set_bit(R5_LOCKED, &dev->flags);
2595 clear_bit(R5_UPTODATE, &dev->flags);
600aa109 2596 s->locked++;
e33129d8
DW		 2597 }
2598 }
ce7d363a
N		 2599 if (!s->locked)
2600 /* False alarm - nothing to do */
2601 return;
2602 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2603 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2604 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2605 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
e33129d8
DW		 2606 }
2607
c0f7bddb 2608 /* keep the parity disk(s) locked while asynchronous operations
e33129d8
DW		 2609 * are in flight
2610 */
2611 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2612 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
600aa109 2613 s->locked++;
e33129d8 2614
c0f7bddb
YT		 2615 if (level == 6) {
2616 int qd_idx = sh->qd_idx;
2617 struct r5dev *dev = &sh->dev[qd_idx];
2618
2619 set_bit(R5_LOCKED, &dev->flags);
2620 clear_bit(R5_UPTODATE, &dev->flags);
2621 s->locked++;
2622 }
2623
600aa109 2624 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
e46b272b 2625 __func__, (unsigned long long)sh->sector,
600aa109 2626 s->locked, s->ops_request);
e33129d8 2627}
16a53ecc 2628
1da177e4
LT		 2629/*
2630 * Each stripe/dev can have one or more bion attached.
16a53ecc 2631 * toread/towrite point to the first in a chain.
1da177e4
LT		 2632 * The bi_next chain must be in order.
2633 */
2634static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2635{
2636 struct bio **bip;
d1688a6d 2637 struct r5conf *conf = sh->raid_conf;
72626685 2638 int firstwrite=0;
1da177e4 2639
cbe47ec5 2640 pr_debug("adding bi b#%llu to stripe s#%llu\n",
4f024f37 2641 (unsigned long long)bi->bi_iter.bi_sector,
1da177e4
LT		 2642 (unsigned long long)sh->sector);
2643
b17459c0
SL		 2644 /*
2645 * If several bio share a stripe. The bio bi_phys_segments acts as a
2646 * reference count to avoid race. The reference count should already be
2647 * increased before this function is called (for example, in
2648 * make_request()), so other bio sharing this stripe will not free the
2649 * stripe. If a stripe is owned by one stripe, the stripe lock will
2650 * protect it.
2651 */
2652 spin_lock_irq(&sh->stripe_lock);
72626685 2653 if (forwrite) {
1da177e4 2654 bip = &sh->dev[dd_idx].towrite;
7eaf7e8e 2655 if (*bip == NULL)
72626685
N		 2656 firstwrite = 1;
2657 } else
1da177e4 2658 bip = &sh->dev[dd_idx].toread;
4f024f37
KO		 2659 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2660 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
1da177e4
LT		 2661 goto overlap;
2662 bip = & (*bip)->bi_next;
2663 }
4f024f37 2664 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
1da177e4
LT		 2665 goto overlap;
2666
78bafebd 2667 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1da177e4
LT		 2668 if (*bip)
2669 bi->bi_next = *bip;
2670 *bip = bi;
e7836bd6 2671 raid5_inc_bi_active_stripes(bi);
72626685 2672
1da177e4
LT		 2673 if (forwrite) {
2674 /* check if page is covered */
2675 sector_t sector = sh->dev[dd_idx].sector;
2676 for (bi=sh->dev[dd_idx].towrite;
2677 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
4f024f37 2678 bi && bi->bi_iter.bi_sector <= sector;
1da177e4 2679 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
f73a1c7d
KO		 2680 if (bio_end_sector(bi) >= sector)
2681 sector = bio_end_sector(bi);
1da177e4
LT		 2682 }
2683 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2684 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2685 }
cbe47ec5
N		 2686
2687 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
4f024f37 2688 (unsigned long long)(*bip)->bi_iter.bi_sector,
cbe47ec5 2689 (unsigned long long)sh->sector, dd_idx);
b97390ae 2690 spin_unlock_irq(&sh->stripe_lock);
cbe47ec5
N		 2691
2692 if (conf->mddev->bitmap && firstwrite) {
2693 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2694 STRIPE_SECTORS, 0);
2695 sh->bm_seq = conf->seq_flush+1;
2696 set_bit(STRIPE_BIT_DELAY, &sh->state);
2697 }
1da177e4
LT		 2698 return 1;
2699
2700 overlap:
2701 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
b17459c0 2702 spin_unlock_irq(&sh->stripe_lock);
1da177e4
LT		 2703 return 0;
2704}
2705
d1688a6d 2706static void end_reshape(struct r5conf *conf);
29269553 2707
d1688a6d 2708static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
911d4ee8 2709 struct stripe_head *sh)
ccfcc3c1 2710{
784052ec 2711 int sectors_per_chunk =
09c9e5fa 2712 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
911d4ee8 2713 int dd_idx;
2d2063ce 2714 int chunk_offset = sector_div(stripe, sectors_per_chunk);
112bf897 2715 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2d2063ce 2716
112bf897
N		 2717 raid5_compute_sector(conf,
2718 stripe * (disks - conf->max_degraded)
b875e531 2719 *sectors_per_chunk + chunk_offset,
112bf897 2720 previous,
911d4ee8 2721 &dd_idx, sh);
ccfcc3c1
N		 2722}
2723
a4456856 2724static void
d1688a6d 2725handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
a4456856
DW		 2726 struct stripe_head_state *s, int disks,
2727 struct bio **return_bi)
2728{
2729 int i;
2730 for (i = disks; i--; ) {
2731 struct bio *bi;
2732 int bitmap_end = 0;
2733
2734 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3cb03002 2735 struct md_rdev *rdev;
a4456856
DW		 2736 rcu_read_lock();
2737 rdev = rcu_dereference(conf->disks[i].rdev);
2738 if (rdev && test_bit(In_sync, &rdev->flags))
7f0da59b
N		 2739 atomic_inc(&rdev->nr_pending);
2740 else
2741 rdev = NULL;
a4456856 2742 rcu_read_unlock();
7f0da59b
N		 2743 if (rdev) {
2744 if (!rdev_set_badblocks(
2745 rdev,
2746 sh->sector,
2747 STRIPE_SECTORS, 0))
2748 md_error(conf->mddev, rdev);
2749 rdev_dec_pending(rdev, conf->mddev);
2750 }
a4456856 2751 }
b17459c0 2752 spin_lock_irq(&sh->stripe_lock);
a4456856
DW		 2753 /* fail all writes first */
2754 bi = sh->dev[i].towrite;
2755 sh->dev[i].towrite = NULL;
b17459c0 2756 spin_unlock_irq(&sh->stripe_lock);
1ed850f3 2757 if (bi)
a4456856 2758 bitmap_end = 1;
a4456856
DW		 2759
2760 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2761 wake_up(&conf->wait_for_overlap);
2762
4f024f37 2763 while (bi && bi->bi_iter.bi_sector <
a4456856
DW		 2764 sh->dev[i].sector + STRIPE_SECTORS) {
2765 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2766 clear_bit(BIO_UPTODATE, &bi->bi_flags);
e7836bd6 2767 if (!raid5_dec_bi_active_stripes(bi)) {
a4456856
DW		 2768 md_write_end(conf->mddev);
2769 bi->bi_next = *return_bi;
2770 *return_bi = bi;
2771 }
2772 bi = nextbi;
2773 }
7eaf7e8e
SL		 2774 if (bitmap_end)
2775 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2776 STRIPE_SECTORS, 0, 0);
2777 bitmap_end = 0;
a4456856
DW		 2778 /* and fail all 'written' */
2779 bi = sh->dev[i].written;
2780 sh->dev[i].written = NULL;
d592a996
SL		 2781 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
2782 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
2783 sh->dev[i].page = sh->dev[i].orig_page;
2784 }
2785
a4456856 2786 if (bi) bitmap_end = 1;
4f024f37 2787 while (bi && bi->bi_iter.bi_sector <
a4456856
DW		 2788 sh->dev[i].sector + STRIPE_SECTORS) {
2789 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2790 clear_bit(BIO_UPTODATE, &bi->bi_flags);
e7836bd6 2791 if (!raid5_dec_bi_active_stripes(bi)) {
a4456856
DW		 2792 md_write_end(conf->mddev);
2793 bi->bi_next = *return_bi;
2794 *return_bi = bi;
2795 }
2796 bi = bi2;
2797 }
2798
b5e98d65
DW		 2799 /* fail any reads if this device is non-operational and
2800 * the data has not reached the cache yet.
2801 */
2802 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2803 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2804 test_bit(R5_ReadError, &sh->dev[i].flags))) {
143c4d05 2805 spin_lock_irq(&sh->stripe_lock);
a4456856
DW		 2806 bi = sh->dev[i].toread;
2807 sh->dev[i].toread = NULL;
143c4d05 2808 spin_unlock_irq(&sh->stripe_lock);
a4456856
DW		 2809 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2810 wake_up(&conf->wait_for_overlap);
4f024f37 2811 while (bi && bi->bi_iter.bi_sector <
a4456856
DW		 2812 sh->dev[i].sector + STRIPE_SECTORS) {
2813 struct bio *nextbi =
2814 r5_next_bio(bi, sh->dev[i].sector);
2815 clear_bit(BIO_UPTODATE, &bi->bi_flags);
e7836bd6 2816 if (!raid5_dec_bi_active_stripes(bi)) {
a4456856
DW		 2817 bi->bi_next = *return_bi;
2818 *return_bi = bi;
2819 }
2820 bi = nextbi;
2821 }
2822 }
a4456856
DW		 2823 if (bitmap_end)
2824 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2825 STRIPE_SECTORS, 0, 0);
8cfa7b0f
N		 2826 /* If we were in the middle of a write the parity block might
2827 * still be locked - so just clear all R5_LOCKED flags
2828 */
2829 clear_bit(R5_LOCKED, &sh->dev[i].flags);
a4456856
DW		 2830 }
2831
8b3e6cdc
DW		 2832 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2833 if (atomic_dec_and_test(&conf->pending_full_writes))
2834 md_wakeup_thread(conf->mddev->thread);
a4456856
DW		 2835}
2836
7f0da59b 2837static void
d1688a6d 2838handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
7f0da59b
N		 2839 struct stripe_head_state *s)
2840{
2841 int abort = 0;
2842 int i;
2843
7f0da59b 2844 clear_bit(STRIPE_SYNCING, &sh->state);
f8dfcffd
N		 2845 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2846 wake_up(&conf->wait_for_overlap);
7f0da59b 2847 s->syncing = 0;
9a3e1101 2848 s->replacing = 0;
7f0da59b 2849 /* There is nothing more to do for sync/check/repair.
18b9837e
N		 2850 * Don't even need to abort as that is handled elsewhere
2851 * if needed, and not always wanted e.g. if there is a known
2852 * bad block here.
9a3e1101 2853 * For recover/replace we need to record a bad block on all
7f0da59b
N		 2854 * non-sync devices, or abort the recovery
2855 */
18b9837e
N		 2856 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2857 /* During recovery devices cannot be removed, so
2858 * locking and refcounting of rdevs is not needed
2859 */
2860 for (i = 0; i < conf->raid_disks; i++) {
2861 struct md_rdev *rdev = conf->disks[i].rdev;
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 rdev = conf->disks[i].replacement;
2869 if (rdev
2870 && !test_bit(Faulty, &rdev->flags)
2871 && !test_bit(In_sync, &rdev->flags)
2872 && !rdev_set_badblocks(rdev, sh->sector,
2873 STRIPE_SECTORS, 0))
2874 abort = 1;
2875 }
2876 if (abort)
2877 conf->recovery_disabled =
2878 conf->mddev->recovery_disabled;
7f0da59b 2879 }
18b9837e 2880 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
7f0da59b
N		 2881}
2882
9a3e1101
N		 2883static int want_replace(struct stripe_head *sh, int disk_idx)
2884{
2885 struct md_rdev *rdev;
2886 int rv = 0;
2887 /* Doing recovery so rcu locking not required */
2888 rdev = sh->raid_conf->disks[disk_idx].replacement;
2889 if (rdev
2890 && !test_bit(Faulty, &rdev->flags)
2891 && !test_bit(In_sync, &rdev->flags)
2892 && (rdev->recovery_offset <= sh->sector
2893 || rdev->mddev->recovery_cp <= sh->sector))
2894 rv = 1;
2895
2896 return rv;
2897}
2898
93b3dbce 2899/* fetch_block - checks the given member device to see if its data needs
1fe797e6
DW		 2900 * to be read or computed to satisfy a request.
2901 *
2902 * Returns 1 when no more member devices need to be checked, otherwise returns
93b3dbce 2903 * 0 to tell the loop in handle_stripe_fill to continue
f38e1219 2904 */
93b3dbce
N		 2905static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2906 int disk_idx, int disks)
a4456856 2907{
5599becc 2908 struct r5dev *dev = &sh->dev[disk_idx];
f2b3b44d
N		 2909 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2910 &sh->dev[s->failed_num[1]] };
5599becc 2911
93b3dbce 2912 /* is the data in this block needed, and can we get it? */
5599becc
YT		 2913 if (!test_bit(R5_LOCKED, &dev->flags) &&
2914 !test_bit(R5_UPTODATE, &dev->flags) &&
2915 (dev->toread ||
2916 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2917 s->syncing || s->expanding ||
9a3e1101 2918 (s->replacing && want_replace(sh, disk_idx)) ||
5d35e09c
N		 2919 (s->failed >= 1 && fdev[0]->toread) ||
2920 (s->failed >= 2 && fdev[1]->toread) ||
93b3dbce 2921 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
67f45548 2922 (!test_bit(R5_Insync, &dev->flags) || test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) &&
93b3dbce 2923 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
67f45548 2924 (sh->raid_conf->level == 6 && s->failed && s->to_write &&
a40687ff 2925 s->to_write - s->non_overwrite < sh->raid_conf->raid_disks - 2 &&
67f45548 2926 (!test_bit(R5_Insync, &dev->flags) || test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))))) {
5599becc
YT		 2927 /* we would like to get this block, possibly by computing it,
2928 * otherwise read it if the backing disk is insync
2929 */
2930 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2931 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2932 if ((s->uptodate == disks - 1) &&
f2b3b44d
N		 2933 (s->failed && (disk_idx == s->failed_num[0] ||
2934 disk_idx == s->failed_num[1]))) {
5599becc
YT		 2935 /* have disk failed, and we're requested to fetch it;
2936 * do compute it
a4456856 2937 */
5599becc
YT		 2938 pr_debug("Computing stripe %llu block %d\n",
2939 (unsigned long long)sh->sector, disk_idx);
2940 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2941 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2942 set_bit(R5_Wantcompute, &dev->flags);
2943 sh->ops.target = disk_idx;
2944 sh->ops.target2 = -1; /* no 2nd target */
2945 s->req_compute = 1;
93b3dbce
N		 2946 /* Careful: from this point on 'uptodate' is in the eye
2947 * of raid_run_ops which services 'compute' operations
2948 * before writes. R5_Wantcompute flags a block that will
2949 * be R5_UPTODATE by the time it is needed for a
2950 * subsequent operation.
2951 */
5599becc
YT		 2952 s->uptodate++;
2953 return 1;
2954 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2955 /* Computing 2-failure is *very* expensive; only
2956 * do it if failed >= 2
2957 */
2958 int other;
2959 for (other = disks; other--; ) {
2960 if (other == disk_idx)
2961 continue;
2962 if (!test_bit(R5_UPTODATE,
2963 &sh->dev[other].flags))
2964 break;
a4456856 2965 }
5599becc
YT		 2966 BUG_ON(other < 0);
2967 pr_debug("Computing stripe %llu blocks %d,%d\n",
2968 (unsigned long long)sh->sector,
2969 disk_idx, other);
2970 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2971 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2972 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2973 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2974 sh->ops.target = disk_idx;
2975 sh->ops.target2 = other;
2976 s->uptodate += 2;
2977 s->req_compute = 1;
2978 return 1;
2979 } else if (test_bit(R5_Insync, &dev->flags)) {
2980 set_bit(R5_LOCKED, &dev->flags);
2981 set_bit(R5_Wantread, &dev->flags);
2982 s->locked++;
2983 pr_debug("Reading block %d (sync=%d)\n",
2984 disk_idx, s->syncing);
a4456856
DW		 2985 }
2986 }
5599becc
YT		 2987
2988 return 0;
2989}
2990
2991/**
93b3dbce 2992 * handle_stripe_fill - read or compute data to satisfy pending requests.
5599becc 2993 */
93b3dbce
N		 2994static void handle_stripe_fill(struct stripe_head *sh,
2995 struct stripe_head_state *s,
2996 int disks)
5599becc
YT		 2997{
2998 int i;
2999
3000 /* look for blocks to read/compute, skip this if a compute
3001 * is already in flight, or if the stripe contents are in the
3002 * midst of changing due to a write
3003 */
3004 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3005 !sh->reconstruct_state)
3006 for (i = disks; i--; )
93b3dbce 3007 if (fetch_block(sh, s, i, disks))
5599becc 3008 break;
a4456856
DW		 3009 set_bit(STRIPE_HANDLE, &sh->state);
3010}
3011
3012
1fe797e6 3013/* handle_stripe_clean_event
a4456856
DW		 3014 * any written block on an uptodate or failed drive can be returned.
3015 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3016 * never LOCKED, so we don't need to test 'failed' directly.
3017 */
d1688a6d 3018static void handle_stripe_clean_event(struct r5conf *conf,
a4456856
DW		 3019 struct stripe_head *sh, int disks, struct bio **return_bi)
3020{
3021 int i;
3022 struct r5dev *dev;
f8dfcffd 3023 int discard_pending = 0;
a4456856
DW		 3024
3025 for (i = disks; i--; )
3026 if (sh->dev[i].written) {
3027 dev = &sh->dev[i];
3028 if (!test_bit(R5_LOCKED, &dev->flags) &&
9e444768 3029 (test_bit(R5_UPTODATE, &dev->flags) ||
d592a996
SL		 3030 test_bit(R5_Discard, &dev->flags) ||
3031 test_bit(R5_SkipCopy, &dev->flags))) {
a4456856
DW		 3032 /* We can return any write requests */
3033 struct bio *wbi, *wbi2;
45b4233c 3034 pr_debug("Return write for disc %d\n", i);
ca64cae9
N		 3035 if (test_and_clear_bit(R5_Discard, &dev->flags))
3036 clear_bit(R5_UPTODATE, &dev->flags);
d592a996
SL		 3037 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3038 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3039 dev->page = dev->orig_page;
3040 }
a4456856
DW		 3041 wbi = dev->written;
3042 dev->written = NULL;
4f024f37 3043 while (wbi && wbi->bi_iter.bi_sector <
a4456856
DW		 3044 dev->sector + STRIPE_SECTORS) {
3045 wbi2 = r5_next_bio(wbi, dev->sector);
e7836bd6 3046 if (!raid5_dec_bi_active_stripes(wbi)) {
a4456856
DW		 3047 md_write_end(conf->mddev);
3048 wbi->bi_next = *return_bi;
3049 *return_bi = wbi;
3050 }
3051 wbi = wbi2;
3052 }
7eaf7e8e
SL		 3053 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3054 STRIPE_SECTORS,
a4456856 3055 !test_bit(STRIPE_DEGRADED, &sh->state),
7eaf7e8e 3056 0);
f8dfcffd
N		 3057 } else if (test_bit(R5_Discard, &dev->flags))
3058 discard_pending = 1;
d592a996
SL		 3059 WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3060 WARN_ON(dev->page != dev->orig_page);
f8dfcffd
N		 3061 }
3062 if (!discard_pending &&
3063 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3064 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3065 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3066 if (sh->qd_idx >= 0) {
3067 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3068 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3069 }
3070 /* now that discard is done we can proceed with any sync */
3071 clear_bit(STRIPE_DISCARD, &sh->state);
d47648fc
SL		 3072 /*
3073 * SCSI discard will change some bio fields and the stripe has
3074 * no updated data, so remove it from hash list and the stripe
3075 * will be reinitialized
3076 */
3077 spin_lock_irq(&conf->device_lock);
3078 remove_hash(sh);
3079 spin_unlock_irq(&conf->device_lock);
f8dfcffd
N		 3080 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3081 set_bit(STRIPE_HANDLE, &sh->state);
3082
3083 }
8b3e6cdc
DW		 3084
3085 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3086 if (atomic_dec_and_test(&conf->pending_full_writes))
3087 md_wakeup_thread(conf->mddev->thread);
a4456856
DW		 3088}
3089
d1688a6d 3090static void handle_stripe_dirtying(struct r5conf *conf,
c8ac1803
N		 3091 struct stripe_head *sh,
3092 struct stripe_head_state *s,
3093 int disks)
a4456856
DW		 3094{
3095 int rmw = 0, rcw = 0, i;
a7854487
AL		 3096 sector_t recovery_cp = conf->mddev->recovery_cp;
3097
3098 /* RAID6 requires 'rcw' in current implementation.
3099 * Otherwise, check whether resync is now happening or should start.
3100 * If yes, then the array is dirty (after unclean shutdown or
3101 * initial creation), so parity in some stripes might be inconsistent.
3102 * In this case, we need to always do reconstruct-write, to ensure
3103 * that in case of drive failure or read-error correction, we
3104 * generate correct data from the parity.
3105 */
3106 if (conf->max_degraded == 2 ||
3107 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
3108 /* Calculate the real rcw later - for now make it
c8ac1803
N		 3109 * look like rcw is cheaper
3110 */
3111 rcw = 1; rmw = 2;
a7854487
AL		 3112 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
3113 conf->max_degraded, (unsigned long long)recovery_cp,
3114 (unsigned long long)sh->sector);
c8ac1803 3115 } else for (i = disks; i--; ) {
a4456856
DW		 3116 /* would I have to read this buffer for read_modify_write */
3117 struct r5dev *dev = &sh->dev[i];
3118 if ((dev->towrite || i == sh->pd_idx) &&
3119 !test_bit(R5_LOCKED, &dev->flags) &&
f38e1219
DW		 3120 !(test_bit(R5_UPTODATE, &dev->flags) ||
3121 test_bit(R5_Wantcompute, &dev->flags))) {
a4456856
DW		 3122 if (test_bit(R5_Insync, &dev->flags))
3123 rmw++;
3124 else
3125 rmw += 2*disks; /* cannot read it */
3126 }
3127 /* Would I have to read this buffer for reconstruct_write */
3128 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
3129 !test_bit(R5_LOCKED, &dev->flags) &&
f38e1219
DW		 3130 !(test_bit(R5_UPTODATE, &dev->flags) ||
3131 test_bit(R5_Wantcompute, &dev->flags))) {
67f45548
N		 3132 if (test_bit(R5_Insync, &dev->flags))
3133 rcw++;
a4456856
DW		 3134 else
3135 rcw += 2*disks;
3136 }
3137 }
45b4233c 3138 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
a4456856
DW		 3139 (unsigned long long)sh->sector, rmw, rcw);
3140 set_bit(STRIPE_HANDLE, &sh->state);
a9add5d9 3141 if (rmw < rcw && rmw > 0) {
a4456856 3142 /* prefer read-modify-write, but need to get some data */
e3620a3a
JB		 3143 if (conf->mddev->queue)
3144 blk_add_trace_msg(conf->mddev->queue,
3145 "raid5 rmw %llu %d",
3146 (unsigned long long)sh->sector, rmw);
a4456856
DW		 3147 for (i = disks; i--; ) {
3148 struct r5dev *dev = &sh->dev[i];
3149 if ((dev->towrite || i == sh->pd_idx) &&
3150 !test_bit(R5_LOCKED, &dev->flags) &&
f38e1219
DW		 3151 !(test_bit(R5_UPTODATE, &dev->flags) ||
3152 test_bit(R5_Wantcompute, &dev->flags)) &&
a4456856 3153 test_bit(R5_Insync, &dev->flags)) {
67f45548
N		 3154 if (test_bit(STRIPE_PREREAD_ACTIVE,
3155 &sh->state)) {
3156 pr_debug("Read_old block %d for r-m-w\n",
3157 i);
a4456856
DW		 3158 set_bit(R5_LOCKED, &dev->flags);
3159 set_bit(R5_Wantread, &dev->flags);
3160 s->locked++;
3161 } else {
3162 set_bit(STRIPE_DELAYED, &sh->state);
3163 set_bit(STRIPE_HANDLE, &sh->state);
3164 }
3165 }
3166 }
a9add5d9 3167 }
c8ac1803 3168 if (rcw <= rmw && rcw > 0) {
a4456856 3169 /* want reconstruct write, but need to get some data */
a9add5d9 3170 int qread =0;
c8ac1803 3171 rcw = 0;
a4456856
DW		 3172 for (i = disks; i--; ) {
3173 struct r5dev *dev = &sh->dev[i];
3174 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
c8ac1803 3175 i != sh->pd_idx && i != sh->qd_idx &&
a4456856 3176 !test_bit(R5_LOCKED, &dev->flags) &&
f38e1219 3177 !(test_bit(R5_UPTODATE, &dev->flags) ||
c8ac1803
N		 3178 test_bit(R5_Wantcompute, &dev->flags))) {
3179 rcw++;
67f45548
N		 3180 if (test_bit(R5_Insync, &dev->flags) &&
3181 test_bit(STRIPE_PREREAD_ACTIVE,
3182 &sh->state)) {
45b4233c 3183 pr_debug("Read_old block "
a4456856
DW		 3184 "%d for Reconstruct\n", i);
3185 set_bit(R5_LOCKED, &dev->flags);
3186 set_bit(R5_Wantread, &dev->flags);
3187 s->locked++;
a9add5d9 3188 qread++;
a4456856
DW		 3189 } else {
3190 set_bit(STRIPE_DELAYED, &sh->state);
3191 set_bit(STRIPE_HANDLE, &sh->state);
3192 }
3193 }
3194 }
e3620a3a 3195 if (rcw && conf->mddev->queue)
a9add5d9
N		 3196 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3197 (unsigned long long)sh->sector,
3198 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
c8ac1803 3199 }
a4456856
DW		 3200 /* now if nothing is locked, and if we have enough data,
3201 * we can start a write request
3202 */
f38e1219
DW		 3203 /* since handle_stripe can be called at any time we need to handle the
3204 * case where a compute block operation has been submitted and then a
ac6b53b6
DW		 3205 * subsequent call wants to start a write request. raid_run_ops only
3206 * handles the case where compute block and reconstruct are requested
f38e1219
DW		 3207 * simultaneously. If this is not the case then new writes need to be
3208 * held off until the compute completes.
3209 */
976ea8d4
DW		 3210 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3211 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3212 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
c0f7bddb 3213 schedule_reconstruction(sh, s, rcw == 0, 0);
a4456856
DW		 3214}
3215
d1688a6d 3216static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
a4456856
DW		 3217 struct stripe_head_state *s, int disks)
3218{
ecc65c9b 3219 struct r5dev *dev = NULL;
bd2ab670 3220
a4456856 3221 set_bit(STRIPE_HANDLE, &sh->state);
e89f8962 3222
ecc65c9b
DW		 3223 switch (sh->check_state) {
3224 case check_state_idle:
3225 /* start a new check operation if there are no failures */
bd2ab670 3226 if (s->failed == 0) {
bd2ab670 3227 BUG_ON(s->uptodate != disks);
ecc65c9b
DW		 3228 sh->check_state = check_state_run;
3229 set_bit(STRIPE_OP_CHECK, &s->ops_request);
bd2ab670 3230 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
bd2ab670 3231 s->uptodate--;
ecc65c9b 3232 break;
bd2ab670 3233 }
f2b3b44d 3234 dev = &sh->dev[s->failed_num[0]];
ecc65c9b
DW		 3235 /* fall through */
3236 case check_state_compute_result:
3237 sh->check_state = check_state_idle;
3238 if (!dev)
3239 dev = &sh->dev[sh->pd_idx];
3240
3241 /* check that a write has not made the stripe insync */
3242 if (test_bit(STRIPE_INSYNC, &sh->state))
3243 break;
c8894419 3244
a4456856 3245 /* either failed parity check, or recovery is happening */
a4456856
DW		 3246 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3247 BUG_ON(s->uptodate != disks);
3248
3249 set_bit(R5_LOCKED, &dev->flags);
ecc65c9b 3250 s->locked++;
a4456856 3251 set_bit(R5_Wantwrite, &dev->flags);
830ea016 3252
a4456856 3253 clear_bit(STRIPE_DEGRADED, &sh->state);
a4456856 3254 set_bit(STRIPE_INSYNC, &sh->state);
ecc65c9b
DW		 3255 break;
3256 case check_state_run:
3257 break; /* we will be called again upon completion */
3258 case check_state_check_result:
3259 sh->check_state = check_state_idle;
3260
3261 /* if a failure occurred during the check operation, leave
3262 * STRIPE_INSYNC not set and let the stripe be handled again
3263 */
3264 if (s->failed)
3265 break;
3266
3267 /* handle a successful check operation, if parity is correct
3268 * we are done. Otherwise update the mismatch count and repair
3269 * parity if !MD_RECOVERY_CHECK
3270 */
ad283ea4 3271 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
ecc65c9b
DW		 3272 /* parity is correct (on disc,
3273 * not in buffer any more)
3274 */
3275 set_bit(STRIPE_INSYNC, &sh->state);
3276 else {
7f7583d4 3277 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
ecc65c9b
DW		 3278 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3279 /* don't try to repair!! */
3280 set_bit(STRIPE_INSYNC, &sh->state);
3281 else {
3282 sh->check_state = check_state_compute_run;
976ea8d4 3283 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
ecc65c9b
DW		 3284 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3285 set_bit(R5_Wantcompute,
3286 &sh->dev[sh->pd_idx].flags);
3287 sh->ops.target = sh->pd_idx;
ac6b53b6 3288 sh->ops.target2 = -1;
ecc65c9b
DW		 3289 s->uptodate++;
3290 }
3291 }
3292 break;
3293 case check_state_compute_run:
3294 break;
3295 default:
3296 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3297 __func__, sh->check_state,
3298 (unsigned long long) sh->sector);
3299 BUG();
a4456856
DW		 3300 }
3301}
3302
3303
d1688a6d 3304static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
36d1c647 3305 struct stripe_head_state *s,
f2b3b44d 3306 int disks)
a4456856 3307{
a4456856 3308 int pd_idx = sh->pd_idx;
34e04e87 3309 int qd_idx = sh->qd_idx;
d82dfee0 3310 struct r5dev *dev;
a4456856
DW		 3311
3312 set_bit(STRIPE_HANDLE, &sh->state);
3313
3314 BUG_ON(s->failed > 2);
d82dfee0 3315
a4456856
DW		 3316 /* Want to check and possibly repair P and Q.
3317 * However there could be one 'failed' device, in which
3318 * case we can only check one of them, possibly using the
3319 * other to generate missing data
3320 */
3321
d82dfee0
DW		 3322 switch (sh->check_state) {
3323 case check_state_idle:
3324 /* start a new check operation if there are < 2 failures */
f2b3b44d 3325 if (s->failed == s->q_failed) {
d82dfee0 3326 /* The only possible failed device holds Q, so it
a4456856
DW		 3327 * makes sense to check P (If anything else were failed,
3328 * we would have used P to recreate it).
3329 */
d82dfee0 3330 sh->check_state = check_state_run;
a4456856 3331 }
f2b3b44d 3332 if (!s->q_failed && s->failed < 2) {
d82dfee0 3333 /* Q is not failed, and we didn't use it to generate
a4456856
DW		 3334 * anything, so it makes sense to check it
3335 */
d82dfee0
DW		 3336 if (sh->check_state == check_state_run)
3337 sh->check_state = check_state_run_pq;
3338 else
3339 sh->check_state = check_state_run_q;
a4456856 3340 }
a4456856 3341
d82dfee0
DW		 3342 /* discard potentially stale zero_sum_result */
3343 sh->ops.zero_sum_result = 0;
a4456856 3344
d82dfee0
DW		 3345 if (sh->check_state == check_state_run) {
3346 /* async_xor_zero_sum destroys the contents of P */
3347 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3348 s->uptodate--;
a4456856 3349 }
d82dfee0
DW		 3350 if (sh->check_state >= check_state_run &&
3351 sh->check_state <= check_state_run_pq) {
3352 /* async_syndrome_zero_sum preserves P and Q, so
3353 * no need to mark them !uptodate here
3354 */
3355 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3356 break;
a4456856
DW		 3357 }
3358
d82dfee0
DW		 3359 /* we have 2-disk failure */
3360 BUG_ON(s->failed != 2);
3361 /* fall through */
3362 case check_state_compute_result:
3363 sh->check_state = check_state_idle;
a4456856 3364
d82dfee0
DW		 3365 /* check that a write has not made the stripe insync */
3366 if (test_bit(STRIPE_INSYNC, &sh->state))
3367 break;
a4456856
DW		 3368
3369 /* now write out any block on a failed drive,
d82dfee0 3370 * or P or Q if they were recomputed
a4456856 3371 */
d82dfee0 3372 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
a4456856 3373 if (s->failed == 2) {
f2b3b44d 3374 dev = &sh->dev[s->failed_num[1]];
a4456856
DW		 3375 s->locked++;
3376 set_bit(R5_LOCKED, &dev->flags);
3377 set_bit(R5_Wantwrite, &dev->flags);
3378 }
3379 if (s->failed >= 1) {
f2b3b44d 3380 dev = &sh->dev[s->failed_num[0]];
a4456856
DW		 3381 s->locked++;
3382 set_bit(R5_LOCKED, &dev->flags);
3383 set_bit(R5_Wantwrite, &dev->flags);
3384 }
d82dfee0 3385 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
a4456856
DW		 3386 dev = &sh->dev[pd_idx];
3387 s->locked++;
3388 set_bit(R5_LOCKED, &dev->flags);
3389 set_bit(R5_Wantwrite, &dev->flags);
3390 }
d82dfee0 3391 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
a4456856
DW		 3392 dev = &sh->dev[qd_idx];
3393 s->locked++;
3394 set_bit(R5_LOCKED, &dev->flags);
3395 set_bit(R5_Wantwrite, &dev->flags);
3396 }
3397 clear_bit(STRIPE_DEGRADED, &sh->state);
3398
3399 set_bit(STRIPE_INSYNC, &sh->state);
d82dfee0
DW		 3400 break;
3401 case check_state_run:
3402 case check_state_run_q:
3403 case check_state_run_pq:
3404 break; /* we will be called again upon completion */
3405 case check_state_check_result:
3406 sh->check_state = check_state_idle;
3407
3408 /* handle a successful check operation, if parity is correct
3409 * we are done. Otherwise update the mismatch count and repair
3410 * parity if !MD_RECOVERY_CHECK
3411 */
3412 if (sh->ops.zero_sum_result == 0) {
3413 /* both parities are correct */
3414 if (!s->failed)
3415 set_bit(STRIPE_INSYNC, &sh->state);
3416 else {
3417 /* in contrast to the raid5 case we can validate
3418 * parity, but still have a failure to write
3419 * back
3420 */
3421 sh->check_state = check_state_compute_result;
3422 /* Returning at this point means that we may go
3423 * off and bring p and/or q uptodate again so
3424 * we make sure to check zero_sum_result again
3425 * to verify if p or q need writeback
3426 */
3427 }
3428 } else {
7f7583d4 3429 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
d82dfee0
DW		 3430 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3431 /* don't try to repair!! */
3432 set_bit(STRIPE_INSYNC, &sh->state);
3433 else {
3434 int *target = &sh->ops.target;
3435
3436 sh->ops.target = -1;
3437 sh->ops.target2 = -1;
3438 sh->check_state = check_state_compute_run;
3439 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3440 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3441 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3442 set_bit(R5_Wantcompute,
3443 &sh->dev[pd_idx].flags);
3444 *target = pd_idx;
3445 target = &sh->ops.target2;
3446 s->uptodate++;
3447 }
3448 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3449 set_bit(R5_Wantcompute,
3450 &sh->dev[qd_idx].flags);
3451 *target = qd_idx;
3452 s->uptodate++;
3453 }
3454 }
3455 }
3456 break;
3457 case check_state_compute_run:
3458 break;
3459 default:
3460 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3461 __func__, sh->check_state,
3462 (unsigned long long) sh->sector);
3463 BUG();
a4456856
DW		 3464 }
3465}
3466
d1688a6d 3467static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
a4456856
DW		 3468{
3469 int i;
3470
3471 /* We have read all the blocks in this stripe and now we need to
3472 * copy some of them into a target stripe for expand.
3473 */
f0a50d37 3474 struct dma_async_tx_descriptor *tx = NULL;
a4456856
DW		 3475 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3476 for (i = 0; i < sh->disks; i++)
34e04e87 3477 if (i != sh->pd_idx && i != sh->qd_idx) {
911d4ee8 3478 int dd_idx, j;
a4456856 3479 struct stripe_head *sh2;
a08abd8c 3480 struct async_submit_ctl submit;
a4456856 3481
784052ec 3482 sector_t bn = compute_blocknr(sh, i, 1);
911d4ee8
N		 3483 sector_t s = raid5_compute_sector(conf, bn, 0,
3484 &dd_idx, NULL);
a8c906ca 3485 sh2 = get_active_stripe(conf, s, 0, 1, 1);
a4456856
DW		 3486 if (sh2 == NULL)
3487 /* so far only the early blocks of this stripe
3488 * have been requested. When later blocks
3489 * get requested, we will try again
3490 */
3491 continue;
3492 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3493 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3494 /* must have already done this block */
3495 release_stripe(sh2);
3496 continue;
3497 }
f0a50d37
DW		 3498
3499 /* place all the copies on one channel */
a08abd8c 3500 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
f0a50d37 3501 tx = async_memcpy(sh2->dev[dd_idx].page,
88ba2aa5 3502 sh->dev[i].page, 0, 0, STRIPE_SIZE,
a08abd8c 3503 &submit);
f0a50d37 3504
a4456856
DW		 3505 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3506 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3507 for (j = 0; j < conf->raid_disks; j++)
3508 if (j != sh2->pd_idx &&
86c374ba 3509 j != sh2->qd_idx &&
a4456856
DW		 3510 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3511 break;
3512 if (j == conf->raid_disks) {
3513 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3514 set_bit(STRIPE_HANDLE, &sh2->state);
3515 }
3516 release_stripe(sh2);
f0a50d37 3517
a4456856 3518 }
a2e08551 3519 /* done submitting copies, wait for them to complete */
749586b7 3520 async_tx_quiesce(&tx);
a4456856 3521}
1da177e4
LT		 3522
3523/*
3524 * handle_stripe - do things to a stripe.
3525 *
9a3e1101
N		 3526 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3527 * state of various bits to see what needs to be done.
1da177e4 3528 * Possible results:
9a3e1101
N		 3529 * return some read requests which now have data
3530 * return some write requests which are safely on storage
1da177e4
LT		 3531 * schedule a read on some buffers
3532 * schedule a write of some buffers
3533 * return confirmation of parity correctness
3534 *
1da177e4 3535 */
a4456856 3536
acfe726b 3537static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
1da177e4 3538{
d1688a6d 3539 struct r5conf *conf = sh->raid_conf;
f416885e 3540 int disks = sh->disks;
474af965
N		 3541 struct r5dev *dev;
3542 int i;
9a3e1101 3543 int do_recovery = 0;
1da177e4 3544
acfe726b
N		 3545 memset(s, 0, sizeof(*s));
3546
acfe726b
N		 3547 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3548 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3549 s->failed_num[0] = -1;
3550 s->failed_num[1] = -1;
1da177e4 3551
acfe726b 3552 /* Now to look around and see what can be done */
1da177e4 3553 rcu_read_lock();
16a53ecc 3554 for (i=disks; i--; ) {
3cb03002 3555 struct md_rdev *rdev;
31c176ec
N		 3556 sector_t first_bad;
3557 int bad_sectors;
3558 int is_bad = 0;
acfe726b 3559
16a53ecc 3560 dev = &sh->dev[i];
1da177e4 3561
45b4233c 3562 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
9a3e1101
N		 3563 i, dev->flags,
3564 dev->toread, dev->towrite, dev->written);
6c0069c0
YT		 3565 /* maybe we can reply to a read
3566 *
3567 * new wantfill requests are only permitted while
3568 * ops_complete_biofill is guaranteed to be inactive
3569 */
3570 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3571 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3572 set_bit(R5_Wantfill, &dev->flags);
1da177e4 3573
16a53ecc 3574 /* now count some things */
cc94015a
N		 3575 if (test_bit(R5_LOCKED, &dev->flags))
3576 s->locked++;
3577 if (test_bit(R5_UPTODATE, &dev->flags))
3578 s->uptodate++;
2d6e4ecc 3579 if (test_bit(R5_Wantcompute, &dev->flags)) {
cc94015a
N		 3580 s->compute++;
3581 BUG_ON(s->compute > 2);
2d6e4ecc 3582 }
1da177e4 3583
acfe726b 3584 if (test_bit(R5_Wantfill, &dev->flags))
cc94015a 3585 s->to_fill++;
acfe726b 3586 else if (dev->toread)
cc94015a 3587 s->to_read++;
16a53ecc 3588 if (dev->towrite) {
cc94015a 3589 s->to_write++;
16a53ecc 3590 if (!test_bit(R5_OVERWRITE, &dev->flags))
cc94015a 3591 s->non_overwrite++;
16a53ecc 3592 }
a4456856 3593 if (dev->written)
cc94015a 3594 s->written++;
14a75d3e
N		 3595 /* Prefer to use the replacement for reads, but only
3596 * if it is recovered enough and has no bad blocks.
3597 */
3598 rdev = rcu_dereference(conf->disks[i].replacement);
3599 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3600 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3601 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3602 &first_bad, &bad_sectors))
3603 set_bit(R5_ReadRepl, &dev->flags);
3604 else {
9a3e1101
N		 3605 if (rdev)
3606 set_bit(R5_NeedReplace, &dev->flags);
14a75d3e
N		 3607 rdev = rcu_dereference(conf->disks[i].rdev);
3608 clear_bit(R5_ReadRepl, &dev->flags);
3609 }
9283d8c5
N		 3610 if (rdev && test_bit(Faulty, &rdev->flags))
3611 rdev = NULL;
31c176ec
N		 3612 if (rdev) {
3613 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3614 &first_bad, &bad_sectors);
3615 if (s->blocked_rdev == NULL
3616 && (test_bit(Blocked, &rdev->flags)
3617 || is_bad < 0)) {
3618 if (is_bad < 0)
3619 set_bit(BlockedBadBlocks,
3620 &rdev->flags);
3621 s->blocked_rdev = rdev;
3622 atomic_inc(&rdev->nr_pending);
3623 }
6bfe0b49 3624 }
415e72d0
N		 3625 clear_bit(R5_Insync, &dev->flags);
3626 if (!rdev)
3627 /* Not in-sync */;
31c176ec
N		 3628 else if (is_bad) {
3629 /* also not in-sync */
18b9837e
N		 3630 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3631 test_bit(R5_UPTODATE, &dev->flags)) {
31c176ec
N		 3632 /* treat as in-sync, but with a read error
3633 * which we can now try to correct
3634 */
3635 set_bit(R5_Insync, &dev->flags);
3636 set_bit(R5_ReadError, &dev->flags);
3637 }
3638 } else if (test_bit(In_sync, &rdev->flags))
415e72d0 3639 set_bit(R5_Insync, &dev->flags);
30d7a483 3640 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
415e72d0 3641 /* in sync if before recovery_offset */
30d7a483
N		 3642 set_bit(R5_Insync, &dev->flags);
3643 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3644 test_bit(R5_Expanded, &dev->flags))
3645 /* If we've reshaped into here, we assume it is Insync.
3646 * We will shortly update recovery_offset to make
3647 * it official.
3648 */
3649 set_bit(R5_Insync, &dev->flags);
3650
1cc03eb9 3651 if (test_bit(R5_WriteError, &dev->flags)) {
14a75d3e
N		 3652 /* This flag does not apply to '.replacement'
3653 * only to .rdev, so make sure to check that*/
3654 struct md_rdev *rdev2 = rcu_dereference(
3655 conf->disks[i].rdev);
3656 if (rdev2 == rdev)
3657 clear_bit(R5_Insync, &dev->flags);
3658 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
bc2607f3 3659 s->handle_bad_blocks = 1;
14a75d3e 3660 atomic_inc(&rdev2->nr_pending);
bc2607f3
N		 3661 } else
3662 clear_bit(R5_WriteError, &dev->flags);
3663 }
1cc03eb9 3664 if (test_bit(R5_MadeGood, &dev->flags)) {
14a75d3e
N		 3665 /* This flag does not apply to '.replacement'
3666 * only to .rdev, so make sure to check that*/
3667 struct md_rdev *rdev2 = rcu_dereference(
3668 conf->disks[i].rdev);
3669 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
b84db560 3670 s->handle_bad_blocks = 1;
14a75d3e 3671 atomic_inc(&rdev2->nr_pending);
b84db560
N		 3672 } else
3673 clear_bit(R5_MadeGood, &dev->flags);
3674 }
977df362
N		 3675 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3676 struct md_rdev *rdev2 = rcu_dereference(
3677 conf->disks[i].replacement);
3678 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3679 s->handle_bad_blocks = 1;
3680 atomic_inc(&rdev2->nr_pending);
3681 } else
3682 clear_bit(R5_MadeGoodRepl, &dev->flags);
3683 }
415e72d0 3684 if (!test_bit(R5_Insync, &dev->flags)) {
16a53ecc
N		 3685 /* The ReadError flag will just be confusing now */
3686 clear_bit(R5_ReadError, &dev->flags);
3687 clear_bit(R5_ReWrite, &dev->flags);
1da177e4 3688 }
415e72d0
N		 3689 if (test_bit(R5_ReadError, &dev->flags))
3690 clear_bit(R5_Insync, &dev->flags);
3691 if (!test_bit(R5_Insync, &dev->flags)) {
cc94015a
N		 3692 if (s->failed < 2)
3693 s->failed_num[s->failed] = i;
3694 s->failed++;
9a3e1101
N		 3695 if (rdev && !test_bit(Faulty, &rdev->flags))
3696 do_recovery = 1;
415e72d0 3697 }
1da177e4 3698 }
9a3e1101
N		 3699 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3700 /* If there is a failed device being replaced,
3701 * we must be recovering.
3702 * else if we are after recovery_cp, we must be syncing
c6d2e084 3703 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
9a3e1101
N		 3704 * else we can only be replacing
3705 * sync and recovery both need to read all devices, and so
3706 * use the same flag.
3707 */
3708 if (do_recovery ||
c6d2e084 3709 sh->sector >= conf->mddev->recovery_cp ||
3710 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
9a3e1101
N		 3711 s->syncing = 1;
3712 else
3713 s->replacing = 1;
3714 }
1da177e4 3715 rcu_read_unlock();
cc94015a
N		 3716}
3717
3718static void handle_stripe(struct stripe_head *sh)
3719{
3720 struct stripe_head_state s;
d1688a6d 3721 struct r5conf *conf = sh->raid_conf;
3687c061 3722 int i;
84789554
N		 3723 int prexor;
3724 int disks = sh->disks;
474af965 3725 struct r5dev *pdev, *qdev;
cc94015a
N		 3726
3727 clear_bit(STRIPE_HANDLE, &sh->state);
257a4b42 3728 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
cc94015a
N		 3729 /* already being handled, ensure it gets handled
3730 * again when current action finishes */
3731 set_bit(STRIPE_HANDLE, &sh->state);
3732 return;
3733 }
3734
f8dfcffd
N		 3735 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3736 spin_lock(&sh->stripe_lock);
3737 /* Cannot process 'sync' concurrently with 'discard' */
3738 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3739 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3740 set_bit(STRIPE_SYNCING, &sh->state);
3741 clear_bit(STRIPE_INSYNC, &sh->state);
f94c0b66 3742 clear_bit(STRIPE_REPLACED, &sh->state);
f8dfcffd
N		 3743 }
3744 spin_unlock(&sh->stripe_lock);
cc94015a
N		 3745 }
3746 clear_bit(STRIPE_DELAYED, &sh->state);
3747
3748 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3749 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3750 (unsigned long long)sh->sector, sh->state,
3751 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3752 sh->check_state, sh->reconstruct_state);
3687c061 3753
acfe726b 3754 analyse_stripe(sh, &s);
c5a31000 3755
bc2607f3
N		 3756 if (s.handle_bad_blocks) {
3757 set_bit(STRIPE_HANDLE, &sh->state);
3758 goto finish;
3759 }
3760
474af965
N		 3761 if (unlikely(s.blocked_rdev)) {
3762 if (s.syncing || s.expanding || s.expanded ||
9a3e1101 3763 s.replacing || s.to_write || s.written) {
474af965
N		 3764 set_bit(STRIPE_HANDLE, &sh->state);
3765 goto finish;
3766 }
3767 /* There is nothing for the blocked_rdev to block */
3768 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3769 s.blocked_rdev = NULL;
3770 }
3771
3772 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3773 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3774 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3775 }
3776
3777 pr_debug("locked=%d uptodate=%d to_read=%d"
3778 " to_write=%d failed=%d failed_num=%d,%d\n",
3779 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3780 s.failed_num[0], s.failed_num[1]);
3781 /* check if the array has lost more than max_degraded devices and,
3782 * if so, some requests might need to be failed.
3783 */
9a3f530f
N		 3784 if (s.failed > conf->max_degraded) {
3785 sh->check_state = 0;
3786 sh->reconstruct_state = 0;
3787 if (s.to_read+s.to_write+s.written)
3788 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
9a3e1101 3789 if (s.syncing + s.replacing)
9a3f530f
N		 3790 handle_failed_sync(conf, sh, &s);
3791 }
474af965 3792
84789554
N		 3793 /* Now we check to see if any write operations have recently
3794 * completed
3795 */
3796 prexor = 0;
3797 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3798 prexor = 1;
3799 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3800 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3801 sh->reconstruct_state = reconstruct_state_idle;
3802
3803 /* All the 'written' buffers and the parity block are ready to
3804 * be written back to disk
3805 */
9e444768
SL		 3806 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3807 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
84789554 3808 BUG_ON(sh->qd_idx >= 0 &&
9e444768
SL		 3809 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3810 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
84789554
N		 3811 for (i = disks; i--; ) {
3812 struct r5dev *dev = &sh->dev[i];
3813 if (test_bit(R5_LOCKED, &dev->flags) &&
3814 (i == sh->pd_idx || i == sh->qd_idx ||
3815 dev->written)) {
3816 pr_debug("Writing block %d\n", i);
3817 set_bit(R5_Wantwrite, &dev->flags);
3818 if (prexor)
3819 continue;
9c4bdf69
N		 3820 if (s.failed > 1)
3821 continue;
84789554
N		 3822 if (!test_bit(R5_Insync, &dev->flags) ||
3823 ((i == sh->pd_idx || i == sh->qd_idx) &&
3824 s.failed == 0))
3825 set_bit(STRIPE_INSYNC, &sh->state);
3826 }
3827 }
3828 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3829 s.dec_preread_active = 1;
3830 }
3831
ef5b7c69
N		 3832 /*
3833 * might be able to return some write requests if the parity blocks
3834 * are safe, or on a failed drive
3835 */
3836 pdev = &sh->dev[sh->pd_idx];
3837 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3838 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3839 qdev = &sh->dev[sh->qd_idx];
3840 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3841 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3842 || conf->level < 6;
3843
3844 if (s.written &&
3845 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3846 && !test_bit(R5_LOCKED, &pdev->flags)
3847 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3848 test_bit(R5_Discard, &pdev->flags))))) &&
3849 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3850 && !test_bit(R5_LOCKED, &qdev->flags)
3851 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3852 test_bit(R5_Discard, &qdev->flags))))))
3853 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3854
3855 /* Now we might consider reading some blocks, either to check/generate
3856 * parity, or to satisfy requests
3857 * or to load a block that is being partially written.
3858 */
3859 if (s.to_read || s.non_overwrite
3860 || (conf->level == 6 && s.to_write && s.failed)
3861 || (s.syncing && (s.uptodate + s.compute < disks))
3862 || s.replacing
3863 || s.expanding)
3864 handle_stripe_fill(sh, &s, disks);
3865
84789554
N		 3866 /* Now to consider new write requests and what else, if anything
3867 * should be read. We do not handle new writes when:
3868 * 1/ A 'write' operation (copy+xor) is already in flight.
3869 * 2/ A 'check' operation is in flight, as it may clobber the parity
3870 * block.
3871 */
3872 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3873 handle_stripe_dirtying(conf, sh, &s, disks);
3874
3875 /* maybe we need to check and possibly fix the parity for this stripe
3876 * Any reads will already have been scheduled, so we just see if enough
3877 * data is available. The parity check is held off while parity
3878 * dependent operations are in flight.
3879 */
3880 if (sh->check_state ||
3881 (s.syncing && s.locked == 0 &&
3882 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3883 !test_bit(STRIPE_INSYNC, &sh->state))) {
3884 if (conf->level == 6)
3885 handle_parity_checks6(conf, sh, &s, disks);
3886 else
3887 handle_parity_checks5(conf, sh, &s, disks);
3888 }
c5a31000 3889
f94c0b66
N		 3890 if ((s.replacing || s.syncing) && s.locked == 0
3891 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3892 && !test_bit(STRIPE_REPLACED, &sh->state)) {
9a3e1101
N		 3893 /* Write out to replacement devices where possible */
3894 for (i = 0; i < conf->raid_disks; i++)
f94c0b66
N		 3895 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3896 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
9a3e1101
N		 3897 set_bit(R5_WantReplace, &sh->dev[i].flags);
3898 set_bit(R5_LOCKED, &sh->dev[i].flags);
3899 s.locked++;
3900 }
f94c0b66
N		 3901 if (s.replacing)
3902 set_bit(STRIPE_INSYNC, &sh->state);
3903 set_bit(STRIPE_REPLACED, &sh->state);
9a3e1101
N		 3904 }
3905 if ((s.syncing || s.replacing) && s.locked == 0 &&
f94c0b66 3906 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
9a3e1101 3907 test_bit(STRIPE_INSYNC, &sh->state)) {
c5a31000
N		 3908 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3909 clear_bit(STRIPE_SYNCING, &sh->state);
f8dfcffd
N		 3910 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3911 wake_up(&conf->wait_for_overlap);
c5a31000
N		 3912 }
3913
3914 /* If the failed drives are just a ReadError, then we might need
3915 * to progress the repair/check process
3916 */
3917 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3918 for (i = 0; i < s.failed; i++) {
3919 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3920 if (test_bit(R5_ReadError, &dev->flags)
3921 && !test_bit(R5_LOCKED, &dev->flags)
3922 && test_bit(R5_UPTODATE, &dev->flags)
3923 ) {
3924 if (!test_bit(R5_ReWrite, &dev->flags)) {
3925 set_bit(R5_Wantwrite, &dev->flags);
3926 set_bit(R5_ReWrite, &dev->flags);
3927 set_bit(R5_LOCKED, &dev->flags);
3928 s.locked++;
3929 } else {
3930 /* let's read it back */
3931 set_bit(R5_Wantread, &dev->flags);
3932 set_bit(R5_LOCKED, &dev->flags);
3933 s.locked++;
3934 }
3935 }
3936 }
3937
3938
3687c061
N		 3939 /* Finish reconstruct operations initiated by the expansion process */
3940 if (sh->reconstruct_state == reconstruct_state_result) {
3941 struct stripe_head *sh_src
3942 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3943 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3944 /* sh cannot be written until sh_src has been read.
3945 * so arrange for sh to be delayed a little
3946 */
3947 set_bit(STRIPE_DELAYED, &sh->state);
3948 set_bit(STRIPE_HANDLE, &sh->state);
3949 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3950 &sh_src->state))
3951 atomic_inc(&conf->preread_active_stripes);
3952 release_stripe(sh_src);
3953 goto finish;
3954 }
3955 if (sh_src)
3956 release_stripe(sh_src);
3957
3958 sh->reconstruct_state = reconstruct_state_idle;
3959 clear_bit(STRIPE_EXPANDING, &sh->state);
3960 for (i = conf->raid_disks; i--; ) {
3961 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3962 set_bit(R5_LOCKED, &sh->dev[i].flags);
3963 s.locked++;
3964 }
3965 }
f416885e 3966
3687c061
N		 3967 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3968 !sh->reconstruct_state) {
3969 /* Need to write out all blocks after computing parity */
3970 sh->disks = conf->raid_disks;
3971 stripe_set_idx(sh->sector, conf, 0, sh);
3972 schedule_reconstruction(sh, &s, 1, 1);
3973 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3974 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3975 atomic_dec(&conf->reshape_stripes);
3976 wake_up(&conf->wait_for_overlap);
3977 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3978 }
3979
3980 if (s.expanding && s.locked == 0 &&
3981 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3982 handle_stripe_expansion(conf, sh);
16a53ecc 3983
3687c061 3984finish:
6bfe0b49 3985 /* wait for this device to become unblocked */
5f066c63
N		 3986 if (unlikely(s.blocked_rdev)) {
3987 if (conf->mddev->external)
3988 md_wait_for_blocked_rdev(s.blocked_rdev,
3989 conf->mddev);
3990 else
3991 /* Internal metadata will immediately
3992 * be written by raid5d, so we don't
3993 * need to wait here.
3994 */
3995 rdev_dec_pending(s.blocked_rdev,
3996 conf->mddev);
3997 }
6bfe0b49 3998
bc2607f3
N		 3999 if (s.handle_bad_blocks)
4000 for (i = disks; i--; ) {
3cb03002 4001 struct md_rdev *rdev;
bc2607f3
N		 4002 struct r5dev *dev = &sh->dev[i];
4003 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4004 /* We own a safe reference to the rdev */
4005 rdev = conf->disks[i].rdev;
4006 if (!rdev_set_badblocks(rdev, sh->sector,
4007 STRIPE_SECTORS, 0))
4008 md_error(conf->mddev, rdev);
4009 rdev_dec_pending(rdev, conf->mddev);
4010 }
b84db560
N		 4011 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4012 rdev = conf->disks[i].rdev;
4013 rdev_clear_badblocks(rdev, sh->sector,
c6563a8c 4014 STRIPE_SECTORS, 0);
b84db560
N		 4015 rdev_dec_pending(rdev, conf->mddev);
4016 }
977df362
N		 4017 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4018 rdev = conf->disks[i].replacement;
dd054fce
N		 4019 if (!rdev)
4020 /* rdev have been moved down */
4021 rdev = conf->disks[i].rdev;
977df362 4022 rdev_clear_badblocks(rdev, sh->sector,
c6563a8c 4023 STRIPE_SECTORS, 0);
977df362
N		 4024 rdev_dec_pending(rdev, conf->mddev);
4025 }
bc2607f3
N		 4026 }
4027
6c0069c0
YT		 4028 if (s.ops_request)
4029 raid_run_ops(sh, s.ops_request);
4030
f0e43bcd 4031 ops_run_io(sh, &s);
16a53ecc 4032
c5709ef6 4033 if (s.dec_preread_active) {
729a1866 4034 /* We delay this until after ops_run_io so that if make_request
e9c7469b 4035 * is waiting on a flush, it won't continue until the writes
729a1866
N		 4036 * have actually been submitted.
4037 */
4038 atomic_dec(&conf->preread_active_stripes);
4039 if (atomic_read(&conf->preread_active_stripes) <
4040 IO_THRESHOLD)
4041 md_wakeup_thread(conf->mddev->thread);
4042 }
4043
c5709ef6 4044 return_io(s.return_bi);
16a53ecc 4045
257a4b42 4046 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
16a53ecc
N		 4047}
4048
d1688a6d 4049static void raid5_activate_delayed(struct r5conf *conf)
16a53ecc
N		 4050{
4051 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4052 while (!list_empty(&conf->delayed_list)) {
4053 struct list_head *l = conf->delayed_list.next;
4054 struct stripe_head *sh;
4055 sh = list_entry(l, struct stripe_head, lru);
4056 list_del_init(l);
4057 clear_bit(STRIPE_DELAYED, &sh->state);
4058 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4059 atomic_inc(&conf->preread_active_stripes);
8b3e6cdc 4060 list_add_tail(&sh->lru, &conf->hold_list);
851c30c9 4061 raid5_wakeup_stripe_thread(sh);
16a53ecc 4062 }
482c0834 4063 }
16a53ecc
N		 4064}
4065
566c09c5
SL		 4066static void activate_bit_delay(struct r5conf *conf,
4067 struct list_head *temp_inactive_list)
16a53ecc
N		 4068{
4069 /* device_lock is held */
4070 struct list_head head;
4071 list_add(&head, &conf->bitmap_list);
4072 list_del_init(&conf->bitmap_list);
4073 while (!list_empty(&head)) {
4074 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
566c09c5 4075 int hash;
16a53ecc
N		 4076 list_del_init(&sh->lru);
4077 atomic_inc(&sh->count);
566c09c5
SL		 4078 hash = sh->hash_lock_index;
4079 __release_stripe(conf, sh, &temp_inactive_list[hash]);
16a53ecc
N		 4080 }
4081}
4082
fd01b88c 4083int md_raid5_congested(struct mddev *mddev, int bits)
f022b2fd 4084{
d1688a6d 4085 struct r5conf *conf = mddev->private;
f022b2fd
N		 4086
4087 /* No difference between reads and writes. Just check
4088 * how busy the stripe_cache is
4089 */
3fa841d7 4090
f022b2fd
N		 4091 if (conf->inactive_blocked)
4092 return 1;
4093 if (conf->quiesce)
4094 return 1;
4bda556a 4095 if (atomic_read(&conf->empty_inactive_list_nr))
f022b2fd
N		 4096 return 1;
4097
4098 return 0;
4099}
11d8a6e3
N		 4100EXPORT_SYMBOL_GPL(md_raid5_congested);
4101
4102static int raid5_congested(void *data, int bits)
4103{
fd01b88c 4104 struct mddev *mddev = data;
11d8a6e3
N		 4105
4106 return mddev_congested(mddev, bits) ||
4107 md_raid5_congested(mddev, bits);
4108}
f022b2fd 4109
23032a0e
RBJ		 4110/* We want read requests to align with chunks where possible,
4111 * but write requests don't need to.
4112 */
cc371e66
AK		 4113static int raid5_mergeable_bvec(struct request_queue *q,
4114 struct bvec_merge_data *bvm,
4115 struct bio_vec *biovec)
23032a0e 4116{
fd01b88c 4117 struct mddev *mddev = q->queuedata;
cc371e66 4118 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
23032a0e 4119 int max;
9d8f0363 4120 unsigned int chunk_sectors = mddev->chunk_sectors;
cc371e66 4121 unsigned int bio_sectors = bvm->bi_size >> 9;
23032a0e 4122
cc371e66 4123 if ((bvm->bi_rw & 1) == WRITE)
23032a0e
RBJ		 4124 return biovec->bv_len; /* always allow writes to be mergeable */
4125
664e7c41
AN		 4126 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4127 chunk_sectors = mddev->new_chunk_sectors;
23032a0e
RBJ		 4128 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4129 if (max < 0) max = 0;
4130 if (max <= biovec->bv_len && bio_sectors == 0)
4131 return biovec->bv_len;
4132 else
4133 return max;
4134}
4135
f679623f 4136
fd01b88c 4137static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
f679623f 4138{
4f024f37 4139 sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
9d8f0363 4140 unsigned int chunk_sectors = mddev->chunk_sectors;
aa8b57aa 4141 unsigned int bio_sectors = bio_sectors(bio);
f679623f 4142
664e7c41
AN		 4143 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4144 chunk_sectors = mddev->new_chunk_sectors;
f679623f
RBJ		 4145 return chunk_sectors >=
4146 ((sector & (chunk_sectors - 1)) + bio_sectors);
4147}
4148
46031f9a
RBJ		 4149/*
4150 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
4151 * later sampled by raid5d.
4152 */
d1688a6d 4153static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
46031f9a
RBJ		 4154{
4155 unsigned long flags;
4156
4157 spin_lock_irqsave(&conf->device_lock, flags);
4158
4159 bi->bi_next = conf->retry_read_aligned_list;
4160 conf->retry_read_aligned_list = bi;
4161
4162 spin_unlock_irqrestore(&conf->device_lock, flags);
4163 md_wakeup_thread(conf->mddev->thread);
4164}
4165
4166
d1688a6d 4167static struct bio *remove_bio_from_retry(struct r5conf *conf)
46031f9a
RBJ		 4168{
4169 struct bio *bi;
4170
4171 bi = conf->retry_read_aligned;
4172 if (bi) {
4173 conf->retry_read_aligned = NULL;
4174 return bi;
4175 }
4176 bi = conf->retry_read_aligned_list;
4177 if(bi) {
387bb173 4178 conf->retry_read_aligned_list = bi->bi_next;
46031f9a 4179 bi->bi_next = NULL;
960e739d
JA		 4180 /*
4181 * this sets the active strip count to 1 and the processed
4182 * strip count to zero (upper 8 bits)
4183 */
e7836bd6 4184 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
46031f9a
RBJ		 4185 }
4186
4187 return bi;
4188}
4189
4190
f679623f
RBJ		 4191/*
4192 * The "raid5_align_endio" should check if the read succeeded and if it
4193 * did, call bio_endio on the original bio (having bio_put the new bio
4194 * first).
4195 * If the read failed..
4196 */
6712ecf8 4197static void raid5_align_endio(struct bio *bi, int error)
f679623f
RBJ		 4198{
4199 struct bio* raid_bi = bi->bi_private;
fd01b88c 4200 struct mddev *mddev;
d1688a6d 4201 struct r5conf *conf;
46031f9a 4202 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3cb03002 4203 struct md_rdev *rdev;
46031f9a 4204
f679623f 4205 bio_put(bi);
46031f9a 4206
46031f9a
RBJ		 4207 rdev = (void*)raid_bi->bi_next;
4208 raid_bi->bi_next = NULL;
2b7f2228
N		 4209 mddev = rdev->mddev;
4210 conf = mddev->private;
46031f9a
RBJ		 4211
4212 rdev_dec_pending(rdev, conf->mddev);
4213
4214 if (!error && uptodate) {
0a82a8d1
LT		 4215 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4216 raid_bi, 0);
6712ecf8 4217 bio_endio(raid_bi, 0);
46031f9a
RBJ		 4218 if (atomic_dec_and_test(&conf->active_aligned_reads))
4219 wake_up(&conf->wait_for_stripe);
6712ecf8 4220 return;
46031f9a
RBJ		 4221 }
4222
4223
45b4233c 4224 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
46031f9a
RBJ		 4225
4226 add_bio_to_retry(raid_bi, conf);
f679623f
RBJ		 4227}
4228
387bb173
NB		 4229static int bio_fits_rdev(struct bio *bi)
4230{
165125e1 4231 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
387bb173 4232
aa8b57aa 4233 if (bio_sectors(bi) > queue_max_sectors(q))
387bb173
NB		 4234 return 0;
4235 blk_recount_segments(q, bi);
8a78362c 4236 if (bi->bi_phys_segments > queue_max_segments(q))
387bb173
NB		 4237 return 0;
4238
4239 if (q->merge_bvec_fn)
4240 /* it's too hard to apply the merge_bvec_fn at this stage,
4241 * just just give up
4242 */
4243 return 0;
4244
4245 return 1;
4246}
4247
4248
fd01b88c 4249static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
f679623f 4250{
d1688a6d 4251 struct r5conf *conf = mddev->private;
8553fe7e 4252 int dd_idx;
f679623f 4253 struct bio* align_bi;
3cb03002 4254 struct md_rdev *rdev;
671488cc 4255 sector_t end_sector;
f679623f
RBJ		 4256
4257 if (!in_chunk_boundary(mddev, raid_bio)) {
45b4233c 4258 pr_debug("chunk_aligned_read : non aligned\n");
f679623f
RBJ		 4259 return 0;
4260 }
4261 /*
a167f663 4262 * use bio_clone_mddev to make a copy of the bio
f679623f 4263 */
a167f663 4264 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
f679623f
RBJ		 4265 if (!align_bi)
4266 return 0;
4267 /*
4268 * set bi_end_io to a new function, and set bi_private to the
4269 * original bio.
4270 */
4271 align_bi->bi_end_io = raid5_align_endio;
4272 align_bi->bi_private = raid_bio;
4273 /*
4274 * compute position
4275 */
4f024f37
KO		 4276 align_bi->bi_iter.bi_sector =
4277 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4278 0, &dd_idx, NULL);
f679623f 4279
f73a1c7d 4280 end_sector = bio_end_sector(align_bi);
f679623f 4281 rcu_read_lock();
671488cc
N		 4282 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4283 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4284 rdev->recovery_offset < end_sector) {
4285 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4286 if (rdev &&
4287 (test_bit(Faulty, &rdev->flags) ||
4288 !(test_bit(In_sync, &rdev->flags) ||
4289 rdev->recovery_offset >= end_sector)))
4290 rdev = NULL;
4291 }
4292 if (rdev) {
31c176ec
N		 4293 sector_t first_bad;
4294 int bad_sectors;
4295
f679623f
RBJ		 4296 atomic_inc(&rdev->nr_pending);
4297 rcu_read_unlock();
46031f9a
RBJ		 4298 raid_bio->bi_next = (void*)rdev;
4299 align_bi->bi_bdev = rdev->bdev;
4300 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
46031f9a 4301
31c176ec 4302 if (!bio_fits_rdev(align_bi) ||
4f024f37
KO		 4303 is_badblock(rdev, align_bi->bi_iter.bi_sector,
4304 bio_sectors(align_bi),
31c176ec
N		 4305 &first_bad, &bad_sectors)) {
4306 /* too big in some way, or has a known bad block */
387bb173
NB		 4307 bio_put(align_bi);
4308 rdev_dec_pending(rdev, mddev);
4309 return 0;
4310 }
4311
6c0544e2 4312 /* No reshape active, so we can trust rdev->data_offset */
4f024f37 4313 align_bi->bi_iter.bi_sector += rdev->data_offset;
6c0544e2 4314
46031f9a
RBJ		 4315 spin_lock_irq(&conf->device_lock);
4316 wait_event_lock_irq(conf->wait_for_stripe,
4317 conf->quiesce == 0,
eed8c02e 4318 conf->device_lock);
46031f9a
RBJ		 4319 atomic_inc(&conf->active_aligned_reads);
4320 spin_unlock_irq(&conf->device_lock);
4321
e3620a3a
JB		 4322 if (mddev->gendisk)
4323 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4324 align_bi, disk_devt(mddev->gendisk),
4f024f37 4325 raid_bio->bi_iter.bi_sector);
f679623f
RBJ		 4326 generic_make_request(align_bi);
4327 return 1;
4328 } else {
4329 rcu_read_unlock();
46031f9a 4330 bio_put(align_bi);
f679623f
RBJ		 4331 return 0;
4332 }
4333}
4334
8b3e6cdc
DW		 4335/* __get_priority_stripe - get the next stripe to process
4336 *
4337 * Full stripe writes are allowed to pass preread active stripes up until
4338 * the bypass_threshold is exceeded. In general the bypass_count
4339 * increments when the handle_list is handled before the hold_list; however, it
4340 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4341 * stripe with in flight i/o. The bypass_count will be reset when the
4342 * head of the hold_list has changed, i.e. the head was promoted to the
4343 * handle_list.
4344 */
851c30c9 4345static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
8b3e6cdc 4346{
851c30c9
SL		 4347 struct stripe_head *sh = NULL, *tmp;
4348 struct list_head *handle_list = NULL;
bfc90cb0 4349 struct r5worker_group *wg = NULL;
851c30c9
SL		 4350
4351 if (conf->worker_cnt_per_group == 0) {
4352 handle_list = &conf->handle_list;
4353 } else if (group != ANY_GROUP) {
4354 handle_list = &conf->worker_groups[group].handle_list;
bfc90cb0 4355 wg = &conf->worker_groups[group];
851c30c9
SL		 4356 } else {
4357 int i;
4358 for (i = 0; i < conf->group_cnt; i++) {
4359 handle_list = &conf->worker_groups[i].handle_list;
bfc90cb0 4360 wg = &conf->worker_groups[i];
851c30c9
SL		 4361 if (!list_empty(handle_list))
4362 break;
4363 }
4364 }
8b3e6cdc
DW		 4365
4366 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4367 __func__,
851c30c9 4368 list_empty(handle_list) ? "empty" : "busy",
8b3e6cdc
DW		 4369 list_empty(&conf->hold_list) ? "empty" : "busy",
4370 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4371
851c30c9
SL		 4372 if (!list_empty(handle_list)) {
4373 sh = list_entry(handle_list->next, typeof(*sh), lru);
8b3e6cdc
DW		 4374
4375 if (list_empty(&conf->hold_list))
4376 conf->bypass_count = 0;
4377 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4378 if (conf->hold_list.next == conf->last_hold)
4379 conf->bypass_count++;
4380 else {
4381 conf->last_hold = conf->hold_list.next;
4382 conf->bypass_count -= conf->bypass_threshold;
4383 if (conf->bypass_count < 0)
4384 conf->bypass_count = 0;
4385 }
4386 }
4387 } else if (!list_empty(&conf->hold_list) &&
4388 ((conf->bypass_threshold &&
4389 conf->bypass_count > conf->bypass_threshold) ||
4390 atomic_read(&conf->pending_full_writes) == 0)) {
851c30c9
SL		 4391
4392 list_for_each_entry(tmp, &conf->hold_list, lru) {
4393 if (conf->worker_cnt_per_group == 0 ||
4394 group == ANY_GROUP ||
4395 !cpu_online(tmp->cpu) ||
4396 cpu_to_group(tmp->cpu) == group) {
4397 sh = tmp;
4398 break;
4399 }
4400 }
4401
4402 if (sh) {
4403 conf->bypass_count -= conf->bypass_threshold;
4404 if (conf->bypass_count < 0)
4405 conf->bypass_count = 0;
4406 }
bfc90cb0 4407 wg = NULL;
851c30c9
SL		 4408 }
4409
4410 if (!sh)
8b3e6cdc
DW		 4411 return NULL;
4412
bfc90cb0
SL		 4413 if (wg) {
4414 wg->stripes_cnt--;
4415 sh->group = NULL;
4416 }
8b3e6cdc 4417 list_del_init(&sh->lru);
c7a6d35e 4418 BUG_ON(atomic_inc_return(&sh->count) != 1);
8b3e6cdc
DW		 4419 return sh;
4420}
f679623f 4421
8811b596
SL		 4422struct raid5_plug_cb {
4423 struct blk_plug_cb cb;
4424 struct list_head list;
566c09c5 4425 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
8811b596
SL		 4426};
4427
4428static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4429{
4430 struct raid5_plug_cb *cb = container_of(
4431 blk_cb, struct raid5_plug_cb, cb);
4432 struct stripe_head *sh;
4433 struct mddev *mddev = cb->cb.data;
4434 struct r5conf *conf = mddev->private;
a9add5d9 4435 int cnt = 0;
566c09c5 4436 int hash;
8811b596
SL		 4437
4438 if (cb->list.next && !list_empty(&cb->list)) {
4439 spin_lock_irq(&conf->device_lock);
4440 while (!list_empty(&cb->list)) {
4441 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4442 list_del_init(&sh->lru);
4443 /*
4444 * avoid race release_stripe_plug() sees
4445 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4446 * is still in our list
4447 */
4e857c58 4448 smp_mb__before_atomic();
8811b596 4449 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
773ca82f
SL		 4450 /*
4451 * STRIPE_ON_RELEASE_LIST could be set here. In that
4452 * case, the count is always > 1 here
4453 */
566c09c5
SL		 4454 hash = sh->hash_lock_index;
4455 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
a9add5d9 4456 cnt++;
8811b596
SL		 4457 }
4458 spin_unlock_irq(&conf->device_lock);
4459 }
566c09c5
SL		 4460 release_inactive_stripe_list(conf, cb->temp_inactive_list,
4461 NR_STRIPE_HASH_LOCKS);
e3620a3a
JB		 4462 if (mddev->queue)
4463 trace_block_unplug(mddev->queue, cnt, !from_schedule);
8811b596
SL		 4464 kfree(cb);
4465}
4466
4467static void release_stripe_plug(struct mddev *mddev,
4468 struct stripe_head *sh)
4469{
4470 struct blk_plug_cb *blk_cb = blk_check_plugged(
4471 raid5_unplug, mddev,
4472 sizeof(struct raid5_plug_cb));
4473 struct raid5_plug_cb *cb;
4474
4475 if (!blk_cb) {
4476 release_stripe(sh);
4477 return;
4478 }
4479
4480 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4481
566c09c5
SL		 4482 if (cb->list.next == NULL) {
4483 int i;
8811b596 4484 INIT_LIST_HEAD(&cb->list);
566c09c5
SL		 4485 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
4486 INIT_LIST_HEAD(cb->temp_inactive_list + i);
4487 }
8811b596
SL		 4488
4489 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4490 list_add_tail(&sh->lru, &cb->list);
4491 else
4492 release_stripe(sh);
4493}
4494
620125f2
SL		 4495static void make_discard_request(struct mddev *mddev, struct bio *bi)
4496{
4497 struct r5conf *conf = mddev->private;
4498 sector_t logical_sector, last_sector;
4499 struct stripe_head *sh;
4500 int remaining;
4501 int stripe_sectors;
4502
4503 if (mddev->reshape_position != MaxSector)
4504 /* Skip discard while reshape is happening */
4505 return;
4506
4f024f37
KO		 4507 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4508 last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
620125f2
SL		 4509
4510 bi->bi_next = NULL;
4511 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4512
4513 stripe_sectors = conf->chunk_sectors *
4514 (conf->raid_disks - conf->max_degraded);
4515 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4516 stripe_sectors);
4517 sector_div(last_sector, stripe_sectors);
4518
4519 logical_sector *= conf->chunk_sectors;
4520 last_sector *= conf->chunk_sectors;
4521
4522 for (; logical_sector < last_sector;
4523 logical_sector += STRIPE_SECTORS) {
4524 DEFINE_WAIT(w);
4525 int d;
4526 again:
4527 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4528 prepare_to_wait(&conf->wait_for_overlap, &w,
4529 TASK_UNINTERRUPTIBLE);
f8dfcffd
N		 4530 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4531 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4532 release_stripe(sh);
4533 schedule();
4534 goto again;
4535 }
4536 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
620125f2
SL		 4537 spin_lock_irq(&sh->stripe_lock);
4538 for (d = 0; d < conf->raid_disks; d++) {
4539 if (d == sh->pd_idx || d == sh->qd_idx)
4540 continue;
4541 if (sh->dev[d].towrite || sh->dev[d].toread) {
4542 set_bit(R5_Overlap, &sh->dev[d].flags);
4543 spin_unlock_irq(&sh->stripe_lock);
4544 release_stripe(sh);
4545 schedule();
4546 goto again;
4547 }
4548 }
f8dfcffd 4549 set_bit(STRIPE_DISCARD, &sh->state);
620125f2
SL		 4550 finish_wait(&conf->wait_for_overlap, &w);
4551 for (d = 0; d < conf->raid_disks; d++) {
4552 if (d == sh->pd_idx || d == sh->qd_idx)
4553 continue;
4554 sh->dev[d].towrite = bi;
4555 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4556 raid5_inc_bi_active_stripes(bi);
4557 }
4558 spin_unlock_irq(&sh->stripe_lock);
4559 if (conf->mddev->bitmap) {
4560 for (d = 0;
4561 d < conf->raid_disks - conf->max_degraded;
4562 d++)
4563 bitmap_startwrite(mddev->bitmap,
4564 sh->sector,
4565 STRIPE_SECTORS,
4566 0);
4567 sh->bm_seq = conf->seq_flush + 1;
4568 set_bit(STRIPE_BIT_DELAY, &sh->state);
4569 }
4570
4571 set_bit(STRIPE_HANDLE, &sh->state);
4572 clear_bit(STRIPE_DELAYED, &sh->state);
4573 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4574 atomic_inc(&conf->preread_active_stripes);
4575 release_stripe_plug(mddev, sh);
4576 }
4577
4578 remaining = raid5_dec_bi_active_stripes(bi);
4579 if (remaining == 0) {
4580 md_write_end(mddev);
4581 bio_endio(bi, 0);
4582 }
4583}
4584
b4fdcb02 4585static void make_request(struct mddev *mddev, struct bio * bi)
1da177e4 4586{
d1688a6d 4587 struct r5conf *conf = mddev->private;
911d4ee8 4588 int dd_idx;
1da177e4
LT		 4589 sector_t new_sector;
4590 sector_t logical_sector, last_sector;
4591 struct stripe_head *sh;
a362357b 4592 const int rw = bio_data_dir(bi);
49077326 4593 int remaining;
27c0f68f
SL		 4594 DEFINE_WAIT(w);
4595 bool do_prepare;
1da177e4 4596
e9c7469b
TH		 4597 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4598 md_flush_request(mddev, bi);
5a7bbad2 4599 return;
e5dcdd80
N		 4600 }
4601
3d310eb7 4602 md_write_start(mddev, bi);
06d91a5f 4603
802ba064 4604 if (rw == READ &&
52488615 4605 mddev->reshape_position == MaxSector &&
21a52c6d 4606 chunk_aligned_read(mddev,bi))
5a7bbad2 4607 return;
52488615 4608
620125f2
SL		 4609 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4610 make_discard_request(mddev, bi);
4611 return;
4612 }
4613
4f024f37 4614 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
f73a1c7d 4615 last_sector = bio_end_sector(bi);
1da177e4
LT		 4616 bi->bi_next = NULL;
4617 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
06d91a5f 4618
27c0f68f 4619 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
1da177e4 4620 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
b5663ba4 4621 int previous;
c46501b2 4622 int seq;
b578d55f 4623
27c0f68f 4624 do_prepare = false;
7ecaa1e6 4625 retry:
c46501b2 4626 seq = read_seqcount_begin(&conf->gen_lock);
b5663ba4 4627 previous = 0;
27c0f68f
SL		 4628 if (do_prepare)
4629 prepare_to_wait(&conf->wait_for_overlap, &w,
4630 TASK_UNINTERRUPTIBLE);
b0f9ec04 4631 if (unlikely(conf->reshape_progress != MaxSector)) {
fef9c61f 4632 /* spinlock is needed as reshape_progress may be
df8e7f76
N		 4633 * 64bit on a 32bit platform, and so it might be
4634 * possible to see a half-updated value
aeb878b0 4635 * Of course reshape_progress could change after
df8e7f76
N		 4636 * the lock is dropped, so once we get a reference
4637 * to the stripe that we think it is, we will have
4638 * to check again.
4639 */
7ecaa1e6 4640 spin_lock_irq(&conf->device_lock);
2c810cdd 4641 if (mddev->reshape_backwards
fef9c61f
N		 4642 ? logical_sector < conf->reshape_progress
4643 : logical_sector >= conf->reshape_progress) {
b5663ba4
N		 4644 previous = 1;
4645 } else {
2c810cdd 4646 if (mddev->reshape_backwards
fef9c61f
N		 4647 ? logical_sector < conf->reshape_safe
4648 : logical_sector >= conf->reshape_safe) {
b578d55f
N		 4649 spin_unlock_irq(&conf->device_lock);
4650 schedule();
27c0f68f 4651 do_prepare = true;
b578d55f
N		 4652 goto retry;
4653 }
4654 }
7ecaa1e6
N		 4655 spin_unlock_irq(&conf->device_lock);
4656 }
16a53ecc 4657
112bf897
N		 4658 new_sector = raid5_compute_sector(conf, logical_sector,
4659 previous,
911d4ee8 4660 &dd_idx, NULL);
0c55e022 4661 pr_debug("raid456: make_request, sector %llu logical %llu\n",
c46501b2 4662 (unsigned long long)new_sector,
1da177e4
LT		 4663 (unsigned long long)logical_sector);
4664
b5663ba4 4665 sh = get_active_stripe(conf, new_sector, previous,
a8c906ca 4666 (bi->bi_rw&RWA_MASK), 0);
1da177e4 4667 if (sh) {
b0f9ec04 4668 if (unlikely(previous)) {
7ecaa1e6 4669 /* expansion might have moved on while waiting for a
df8e7f76
N		 4670 * stripe, so we must do the range check again.
4671 * Expansion could still move past after this
4672 * test, but as we are holding a reference to
4673 * 'sh', we know that if that happens,
4674 * STRIPE_EXPANDING will get set and the expansion
4675 * won't proceed until we finish with the stripe.
7ecaa1e6
N		 4676 */
4677 int must_retry = 0;
4678 spin_lock_irq(&conf->device_lock);
2c810cdd 4679 if (mddev->reshape_backwards
b0f9ec04
N		 4680 ? logical_sector >= conf->reshape_progress
4681 : logical_sector < conf->reshape_progress)
7ecaa1e6
N		 4682 /* mismatch, need to try again */
4683 must_retry = 1;
4684 spin_unlock_irq(&conf->device_lock);
4685 if (must_retry) {
4686 release_stripe(sh);
7a3ab908 4687 schedule();
27c0f68f 4688 do_prepare = true;
7ecaa1e6
N		 4689 goto retry;
4690 }
4691 }
c46501b2
N		 4692 if (read_seqcount_retry(&conf->gen_lock, seq)) {
4693 /* Might have got the wrong stripe_head
4694 * by accident
4695 */
4696 release_stripe(sh);
4697 goto retry;
4698 }
e62e58a5 4699
ffd96e35 4700 if (rw == WRITE &&
a5c308d4 4701 logical_sector >= mddev->suspend_lo &&
e464eafd
N		 4702 logical_sector < mddev->suspend_hi) {
4703 release_stripe(sh);
e62e58a5
N		 4704 /* As the suspend_* range is controlled by
4705 * userspace, we want an interruptible
4706 * wait.
4707 */
4708 flush_signals(current);
4709 prepare_to_wait(&conf->wait_for_overlap,
4710 &w, TASK_INTERRUPTIBLE);
4711 if (logical_sector >= mddev->suspend_lo &&
27c0f68f 4712 logical_sector < mddev->suspend_hi) {
e62e58a5 4713 schedule();
27c0f68f
SL		 4714 do_prepare = true;
4715 }
e464eafd
N		 4716 goto retry;
4717 }
7ecaa1e6
N		 4718
4719 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
ffd96e35 4720 !add_stripe_bio(sh, bi, dd_idx, rw)) {
7ecaa1e6
N		 4721 /* Stripe is busy expanding or
4722 * add failed due to overlap. Flush everything
1da177e4
LT		 4723 * and wait a while
4724 */
482c0834 4725 md_wakeup_thread(mddev->thread);
1da177e4
LT		 4726 release_stripe(sh);
4727 schedule();
27c0f68f 4728 do_prepare = true;
1da177e4
LT		 4729 goto retry;
4730 }
6ed3003c
N		 4731 set_bit(STRIPE_HANDLE, &sh->state);
4732 clear_bit(STRIPE_DELAYED, &sh->state);
a852d7b8 4733 if ((bi->bi_rw & REQ_SYNC) &&
729a1866
N		 4734 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4735 atomic_inc(&conf->preread_active_stripes);
8811b596 4736 release_stripe_plug(mddev, sh);
1da177e4
LT		 4737 } else {
4738 /* cannot get stripe for read-ahead, just give-up */
4739 clear_bit(BIO_UPTODATE, &bi->bi_flags);
1da177e4
LT		 4740 break;
4741 }
1da177e4 4742 }
27c0f68f 4743 finish_wait(&conf->wait_for_overlap, &w);
7c13edc8 4744
e7836bd6 4745 remaining = raid5_dec_bi_active_stripes(bi);
f6344757 4746 if (remaining == 0) {
1da177e4 4747
16a53ecc 4748 if ( rw == WRITE )
1da177e4 4749 md_write_end(mddev);
6712ecf8 4750
0a82a8d1
LT		 4751 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4752 bi, 0);
0e13fe23 4753 bio_endio(bi, 0);
1da177e4 4754 }
1da177e4
LT		 4755}
4756
fd01b88c 4757static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
b522adcd 4758
fd01b88c 4759static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
1da177e4 4760{
52c03291
N		 4761 /* reshaping is quite different to recovery/resync so it is
4762 * handled quite separately ... here.
4763 *
4764 * On each call to sync_request, we gather one chunk worth of
4765 * destination stripes and flag them as expanding.
4766 * Then we find all the source stripes and request reads.
4767 * As the reads complete, handle_stripe will copy the data
4768 * into the destination stripe and release that stripe.
4769 */
d1688a6d 4770 struct r5conf *conf = mddev->private;
1da177e4 4771 struct stripe_head *sh;
ccfcc3c1 4772 sector_t first_sector, last_sector;
f416885e
N		 4773 int raid_disks = conf->previous_raid_disks;
4774 int data_disks = raid_disks - conf->max_degraded;
4775 int new_data_disks = conf->raid_disks - conf->max_degraded;
52c03291
N		 4776 int i;
4777 int dd_idx;
c8f517c4 4778 sector_t writepos, readpos, safepos;
ec32a2bd 4779 sector_t stripe_addr;
7a661381 4780 int reshape_sectors;
ab69ae12 4781 struct list_head stripes;
52c03291 4782
fef9c61f
N		 4783 if (sector_nr == 0) {
4784 /* If restarting in the middle, skip the initial sectors */
2c810cdd 4785 if (mddev->reshape_backwards &&
fef9c61f
N		 4786 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4787 sector_nr = raid5_size(mddev, 0, 0)
4788 - conf->reshape_progress;
2c810cdd 4789 } else if (!mddev->reshape_backwards &&
fef9c61f
N		 4790 conf->reshape_progress > 0)
4791 sector_nr = conf->reshape_progress;
f416885e 4792 sector_div(sector_nr, new_data_disks);
fef9c61f 4793 if (sector_nr) {
8dee7211
N		 4794 mddev->curr_resync_completed = sector_nr;
4795 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
fef9c61f
N		 4796 *skipped = 1;
4797 return sector_nr;
4798 }
52c03291
N		 4799 }
4800
7a661381
N		 4801 /* We need to process a full chunk at a time.
4802 * If old and new chunk sizes differ, we need to process the
4803 * largest of these
4804 */
664e7c41
AN		 4805 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4806 reshape_sectors = mddev->new_chunk_sectors;
7a661381 4807 else
9d8f0363 4808 reshape_sectors = mddev->chunk_sectors;
7a661381 4809
b5254dd5
N		 4810 /* We update the metadata at least every 10 seconds, or when
4811 * the data about to be copied would over-write the source of
4812 * the data at the front of the range. i.e. one new_stripe
4813 * along from reshape_progress new_maps to after where
4814 * reshape_safe old_maps to
52c03291 4815 */
fef9c61f 4816 writepos = conf->reshape_progress;
f416885e 4817 sector_div(writepos, new_data_disks);
c8f517c4
N		 4818 readpos = conf->reshape_progress;
4819 sector_div(readpos, data_disks);
fef9c61f 4820 safepos = conf->reshape_safe;
f416885e 4821 sector_div(safepos, data_disks);
2c810cdd 4822 if (mddev->reshape_backwards) {
ed37d83e 4823 writepos -= min_t(sector_t, reshape_sectors, writepos);
c8f517c4 4824 readpos += reshape_sectors;
7a661381 4825 safepos += reshape_sectors;
fef9c61f 4826 } else {
7a661381 4827 writepos += reshape_sectors;
ed37d83e
N		 4828 readpos -= min_t(sector_t, reshape_sectors, readpos);
4829 safepos -= min_t(sector_t, reshape_sectors, safepos);
fef9c61f 4830 }
52c03291 4831
b5254dd5
N		 4832 /* Having calculated the 'writepos' possibly use it
4833 * to set 'stripe_addr' which is where we will write to.
4834 */
4835 if (mddev->reshape_backwards) {
4836 BUG_ON(conf->reshape_progress == 0);
4837 stripe_addr = writepos;
4838 BUG_ON((mddev->dev_sectors &
4839 ~((sector_t)reshape_sectors - 1))
4840 - reshape_sectors - stripe_addr
4841 != sector_nr);
4842 } else {
4843 BUG_ON(writepos != sector_nr + reshape_sectors);
4844 stripe_addr = sector_nr;
4845 }
4846
c8f517c4
N		 4847 /* 'writepos' is the most advanced device address we might write.
4848 * 'readpos' is the least advanced device address we might read.
4849 * 'safepos' is the least address recorded in the metadata as having
4850 * been reshaped.
b5254dd5
N		 4851 * If there is a min_offset_diff, these are adjusted either by
4852 * increasing the safepos/readpos if diff is negative, or
4853 * increasing writepos if diff is positive.
4854 * If 'readpos' is then behind 'writepos', there is no way that we can
c8f517c4
N		 4855 * ensure safety in the face of a crash - that must be done by userspace
4856 * making a backup of the data. So in that case there is no particular
4857 * rush to update metadata.
4858 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4859 * update the metadata to advance 'safepos' to match 'readpos' so that
4860 * we can be safe in the event of a crash.
4861 * So we insist on updating metadata if safepos is behind writepos and
4862 * readpos is beyond writepos.
4863 * In any case, update the metadata every 10 seconds.
4864 * Maybe that number should be configurable, but I'm not sure it is
4865 * worth it.... maybe it could be a multiple of safemode_delay???
4866 */
b5254dd5
N		 4867 if (conf->min_offset_diff < 0) {
4868 safepos += -conf->min_offset_diff;
4869 readpos += -conf->min_offset_diff;
4870 } else
4871 writepos += conf->min_offset_diff;
4872
2c810cdd 4873 if ((mddev->reshape_backwards
c8f517c4
N		 4874 ? (safepos > writepos && readpos < writepos)
4875 : (safepos < writepos && readpos > writepos)) ||
4876 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
52c03291
N		 4877 /* Cannot proceed until we've updated the superblock... */
4878 wait_event(conf->wait_for_overlap,
c91abf5a
N		 4879 atomic_read(&conf->reshape_stripes)==0
4880 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4881 if (atomic_read(&conf->reshape_stripes) != 0)
4882 return 0;
fef9c61f 4883 mddev->reshape_position = conf->reshape_progress;
75d3da43 4884 mddev->curr_resync_completed = sector_nr;
c8f517c4 4885 conf->reshape_checkpoint = jiffies;
850b2b42 4886 set_bit(MD_CHANGE_DEVS, &mddev->flags);
52c03291 4887 md_wakeup_thread(mddev->thread);
850b2b42 4888 wait_event(mddev->sb_wait, mddev->flags == 0 ||
c91abf5a
N		 4889 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4890 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4891 return 0;
52c03291 4892 spin_lock_irq(&conf->device_lock);
fef9c61f 4893 conf->reshape_safe = mddev->reshape_position;
52c03291
N		 4894 spin_unlock_irq(&conf->device_lock);
4895 wake_up(&conf->wait_for_overlap);
acb180b0 4896 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
52c03291
N		 4897 }
4898
ab69ae12 4899 INIT_LIST_HEAD(&stripes);
7a661381 4900 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
52c03291 4901 int j;
a9f326eb 4902 int skipped_disk = 0;
a8c906ca 4903 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
52c03291
N		 4904 set_bit(STRIPE_EXPANDING, &sh->state);
4905 atomic_inc(&conf->reshape_stripes);
4906 /* If any of this stripe is beyond the end of the old
4907 * array, then we need to zero those blocks
4908 */
4909 for (j=sh->disks; j--;) {
4910 sector_t s;
4911 if (j == sh->pd_idx)
4912 continue;
f416885e 4913 if (conf->level == 6 &&
d0dabf7e 4914 j == sh->qd_idx)
f416885e 4915 continue;
784052ec 4916 s = compute_blocknr(sh, j, 0);
b522adcd 4917 if (s < raid5_size(mddev, 0, 0)) {
a9f326eb 4918 skipped_disk = 1;
52c03291
N		 4919 continue;
4920 }
4921 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4922 set_bit(R5_Expanded, &sh->dev[j].flags);
4923 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4924 }
a9f326eb 4925 if (!skipped_disk) {
52c03291
N		 4926 set_bit(STRIPE_EXPAND_READY, &sh->state);
4927 set_bit(STRIPE_HANDLE, &sh->state);
4928 }
ab69ae12 4929 list_add(&sh->lru, &stripes);
52c03291
N		 4930 }
4931 spin_lock_irq(&conf->device_lock);
2c810cdd 4932 if (mddev->reshape_backwards)
7a661381 4933 conf->reshape_progress -= reshape_sectors * new_data_disks;
fef9c61f 4934 else
7a661381 4935 conf->reshape_progress += reshape_sectors * new_data_disks;
52c03291
N		 4936 spin_unlock_irq(&conf->device_lock);
4937 /* Ok, those stripe are ready. We can start scheduling
4938 * reads on the source stripes.
4939 * The source stripes are determined by mapping the first and last
4940 * block on the destination stripes.
4941 */
52c03291 4942 first_sector =
ec32a2bd 4943 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
911d4ee8 4944 1, &dd_idx, NULL);
52c03291 4945 last_sector =
0e6e0271 4946 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
09c9e5fa 4947 * new_data_disks - 1),
911d4ee8 4948 1, &dd_idx, NULL);
58c0fed4
AN		 4949 if (last_sector >= mddev->dev_sectors)
4950 last_sector = mddev->dev_sectors - 1;
52c03291 4951 while (first_sector <= last_sector) {
a8c906ca 4952 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
52c03291
N		 4953 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4954 set_bit(STRIPE_HANDLE, &sh->state);
4955 release_stripe(sh);
4956 first_sector += STRIPE_SECTORS;
4957 }
ab69ae12
N		 4958 /* Now that the sources are clearly marked, we can release
4959 * the destination stripes
4960 */
4961 while (!list_empty(&stripes)) {
4962 sh = list_entry(stripes.next, struct stripe_head, lru);
4963 list_del_init(&sh->lru);
4964 release_stripe(sh);
4965 }
c6207277
N		 4966 /* If this takes us to the resync_max point where we have to pause,
4967 * then we need to write out the superblock.
4968 */
7a661381 4969 sector_nr += reshape_sectors;
c03f6a19
N		 4970 if ((sector_nr - mddev->curr_resync_completed) * 2
4971 >= mddev->resync_max - mddev->curr_resync_completed) {
c6207277
N		 4972 /* Cannot proceed until we've updated the superblock... */
4973 wait_event(conf->wait_for_overlap,
c91abf5a
N		 4974 atomic_read(&conf->reshape_stripes) == 0
4975 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4976 if (atomic_read(&conf->reshape_stripes) != 0)
4977 goto ret;
fef9c61f 4978 mddev->reshape_position = conf->reshape_progress;
75d3da43 4979 mddev->curr_resync_completed = sector_nr;
c8f517c4 4980 conf->reshape_checkpoint = jiffies;
c6207277
N		 4981 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4982 md_wakeup_thread(mddev->thread);
4983 wait_event(mddev->sb_wait,
4984 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
c91abf5a
N		 4985 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4986 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4987 goto ret;
c6207277 4988 spin_lock_irq(&conf->device_lock);
fef9c61f 4989 conf->reshape_safe = mddev->reshape_position;
c6207277
N		 4990 spin_unlock_irq(&conf->device_lock);
4991 wake_up(&conf->wait_for_overlap);
acb180b0 4992 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
c6207277 4993 }
c91abf5a 4994ret:
7a661381 4995 return reshape_sectors;
52c03291
N		 4996}
4997
4998/* FIXME go_faster isn't used */
fd01b88c 4999static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
52c03291 5000{
d1688a6d 5001 struct r5conf *conf = mddev->private;
52c03291 5002 struct stripe_head *sh;
58c0fed4 5003 sector_t max_sector = mddev->dev_sectors;
57dab0bd 5004 sector_t sync_blocks;
16a53ecc
N		 5005 int still_degraded = 0;
5006 int i;
1da177e4 5007
72626685 5008 if (sector_nr >= max_sector) {
1da177e4 5009 /* just being told to finish up .. nothing much to do */
cea9c228 5010
29269553
N		 5011 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5012 end_reshape(conf);
5013 return 0;
5014 }
72626685
N		 5015
5016 if (mddev->curr_resync < max_sector) /* aborted */
5017 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5018 &sync_blocks, 1);
16a53ecc 5019 else /* completed sync */
72626685
N		 5020 conf->fullsync = 0;
5021 bitmap_close_sync(mddev->bitmap);
5022
1da177e4
LT		 5023 return 0;
5024 }
ccfcc3c1 5025
64bd660b
N		 5026 /* Allow raid5_quiesce to complete */
5027 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5028
52c03291
N		 5029 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5030 return reshape_request(mddev, sector_nr, skipped);
f6705578 5031
c6207277
N		 5032 /* No need to check resync_max as we never do more than one
5033 * stripe, and as resync_max will always be on a chunk boundary,
5034 * if the check in md_do_sync didn't fire, there is no chance
5035 * of overstepping resync_max here
5036 */
5037
16a53ecc 5038 /* if there is too many failed drives and we are trying
1da177e4
LT		 5039 * to resync, then assert that we are finished, because there is
5040 * nothing we can do.
5041 */
3285edf1 5042 if (mddev->degraded >= conf->max_degraded &&
16a53ecc 5043 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
58c0fed4 5044 sector_t rv = mddev->dev_sectors - sector_nr;
57afd89f 5045 *skipped = 1;
1da177e4
LT		 5046 return rv;
5047 }
6f608040 5048 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5049 !conf->fullsync &&
5050 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5051 sync_blocks >= STRIPE_SECTORS) {
72626685
N		 5052 /* we can skip this block, and probably more */
5053 sync_blocks /= STRIPE_SECTORS;
5054 *skipped = 1;
5055 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5056 }
1da177e4 5057
b47490c9
N		 5058 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5059
a8c906ca 5060 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
1da177e4 5061 if (sh == NULL) {
a8c906ca 5062 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
1da177e4 5063 /* make sure we don't swamp the stripe cache if someone else
16a53ecc 5064 * is trying to get access
1da177e4 5065 */
66c006a5 5066 schedule_timeout_uninterruptible(1);
1da177e4 5067 }
16a53ecc
N		 5068 /* Need to check if array will still be degraded after recovery/resync
5069 * We don't need to check the 'failed' flag as when that gets set,
5070 * recovery aborts.
5071 */
f001a70c 5072 for (i = 0; i < conf->raid_disks; i++)
16a53ecc
N		 5073 if (conf->disks[i].rdev == NULL)
5074 still_degraded = 1;
5075
5076 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5077
83206d66 5078 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
053f5b65 5079 set_bit(STRIPE_HANDLE, &sh->state);
1da177e4 5080
1da177e4
LT		 5081 release_stripe(sh);
5082
5083 return STRIPE_SECTORS;
5084}
5085
d1688a6d 5086static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
46031f9a
RBJ		 5087{
5088 /* We may not be able to submit a whole bio at once as there
5089 * may not be enough stripe_heads available.
5090 * We cannot pre-allocate enough stripe_heads as we may need
5091 * more than exist in the cache (if we allow ever large chunks).
5092 * So we do one stripe head at a time and record in
5093 * ->bi_hw_segments how many have been done.
5094 *
5095 * We *know* that this entire raid_bio is in one chunk, so
5096 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5097 */
5098 struct stripe_head *sh;
911d4ee8 5099 int dd_idx;
46031f9a
RBJ		 5100 sector_t sector, logical_sector, last_sector;
5101 int scnt = 0;
5102 int remaining;
5103 int handled = 0;
5104
4f024f37
KO		 5105 logical_sector = raid_bio->bi_iter.bi_sector &
5106 ~((sector_t)STRIPE_SECTORS-1);
112bf897 5107 sector = raid5_compute_sector(conf, logical_sector,
911d4ee8 5108 0, &dd_idx, NULL);
f73a1c7d 5109 last_sector = bio_end_sector(raid_bio);
46031f9a
RBJ		 5110
5111 for (; logical_sector < last_sector;
387bb173
NB		 5112 logical_sector += STRIPE_SECTORS,
5113 sector += STRIPE_SECTORS,
5114 scnt++) {
46031f9a 5115
e7836bd6 5116 if (scnt < raid5_bi_processed_stripes(raid_bio))
46031f9a
RBJ		 5117 /* already done this stripe */
5118 continue;
5119
2844dc32 5120 sh = get_active_stripe(conf, sector, 0, 1, 1);
46031f9a
RBJ		 5121
5122 if (!sh) {
5123 /* failed to get a stripe - must wait */
e7836bd6 5124 raid5_set_bi_processed_stripes(raid_bio, scnt);
46031f9a
RBJ		 5125 conf->retry_read_aligned = raid_bio;
5126 return handled;
5127 }
5128
387bb173
NB		 5129 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
5130 release_stripe(sh);
e7836bd6 5131 raid5_set_bi_processed_stripes(raid_bio, scnt);
387bb173
NB		 5132 conf->retry_read_aligned = raid_bio;
5133 return handled;
5134 }
5135
3f9e7c14 5136 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
36d1c647 5137 handle_stripe(sh);
46031f9a
RBJ		 5138 release_stripe(sh);
5139 handled++;
5140 }
e7836bd6 5141 remaining = raid5_dec_bi_active_stripes(raid_bio);
0a82a8d1
LT		 5142 if (remaining == 0) {
5143 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5144 raid_bio, 0);
0e13fe23 5145 bio_endio(raid_bio, 0);
0a82a8d1 5146 }
46031f9a
RBJ		 5147 if (atomic_dec_and_test(&conf->active_aligned_reads))
5148 wake_up(&conf->wait_for_stripe);
5149 return handled;
5150}
5151
bfc90cb0 5152static int handle_active_stripes(struct r5conf *conf, int group,
566c09c5
SL		 5153 struct r5worker *worker,
5154 struct list_head *temp_inactive_list)
46a06401
SL		 5155{
5156 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
566c09c5
SL		 5157 int i, batch_size = 0, hash;
5158 bool release_inactive = false;
46a06401
SL		 5159
5160 while (batch_size < MAX_STRIPE_BATCH &&
851c30c9 5161 (sh = __get_priority_stripe(conf, group)) != NULL)
46a06401
SL		 5162 batch[batch_size++] = sh;
5163
566c09c5
SL		 5164 if (batch_size == 0) {
5165 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5166 if (!list_empty(temp_inactive_list + i))
5167 break;
5168 if (i == NR_STRIPE_HASH_LOCKS)
5169 return batch_size;
5170 release_inactive = true;
5171 }
46a06401
SL		 5172 spin_unlock_irq(&conf->device_lock);
5173
566c09c5
SL		 5174 release_inactive_stripe_list(conf, temp_inactive_list,
5175 NR_STRIPE_HASH_LOCKS);
5176
5177 if (release_inactive) {
5178 spin_lock_irq(&conf->device_lock);
5179 return 0;
5180 }
5181
46a06401
SL		 5182 for (i = 0; i < batch_size; i++)
5183 handle_stripe(batch[i]);
5184
5185 cond_resched();
5186
5187 spin_lock_irq(&conf->device_lock);
566c09c5
SL		 5188 for (i = 0; i < batch_size; i++) {
5189 hash = batch[i]->hash_lock_index;
5190 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5191 }
46a06401
SL		 5192 return batch_size;
5193}
46031f9a 5194
851c30c9
SL		 5195static void raid5_do_work(struct work_struct *work)
5196{
5197 struct r5worker *worker = container_of(work, struct r5worker, work);
5198 struct r5worker_group *group = worker->group;
5199 struct r5conf *conf = group->conf;
5200 int group_id = group - conf->worker_groups;
5201 int handled;
5202 struct blk_plug plug;
5203
5204 pr_debug("+++ raid5worker active\n");
5205
5206 blk_start_plug(&plug);
5207 handled = 0;
5208 spin_lock_irq(&conf->device_lock);
5209 while (1) {
5210 int batch_size, released;
5211
566c09c5 5212 released = release_stripe_list(conf, worker->temp_inactive_list);
851c30c9 5213
566c09c5
SL		 5214 batch_size = handle_active_stripes(conf, group_id, worker,
5215 worker->temp_inactive_list);
bfc90cb0 5216 worker->working = false;
851c30c9
SL		 5217 if (!batch_size && !released)
5218 break;
5219 handled += batch_size;
5220 }
5221 pr_debug("%d stripes handled\n", handled);
5222
5223 spin_unlock_irq(&conf->device_lock);
5224 blk_finish_plug(&plug);
5225
5226 pr_debug("--- raid5worker inactive\n");
5227}
5228
1da177e4
LT		 5229/*
5230 * This is our raid5 kernel thread.
5231 *
5232 * We scan the hash table for stripes which can be handled now.
5233 * During the scan, completed stripes are saved for us by the interrupt
5234 * handler, so that they will not have to wait for our next wakeup.
5235 */
4ed8731d 5236static void raid5d(struct md_thread *thread)
1da177e4 5237{
4ed8731d 5238 struct mddev *mddev = thread->mddev;
d1688a6d 5239 struct r5conf *conf = mddev->private;
1da177e4 5240 int handled;
e1dfa0a2 5241 struct blk_plug plug;
1da177e4 5242
45b4233c 5243 pr_debug("+++ raid5d active\n");
1da177e4
LT		 5244
5245 md_check_recovery(mddev);
1da177e4 5246
e1dfa0a2 5247 blk_start_plug(&plug);
1da177e4
LT		 5248 handled = 0;
5249 spin_lock_irq(&conf->device_lock);
5250 while (1) {
46031f9a 5251 struct bio *bio;
773ca82f
SL		 5252 int batch_size, released;
5253
566c09c5 5254 released = release_stripe_list(conf, conf->temp_inactive_list);
1da177e4 5255
0021b7bc 5256 if (
7c13edc8
N		 5257 !list_empty(&conf->bitmap_list)) {
5258 /* Now is a good time to flush some bitmap updates */
5259 conf->seq_flush++;
700e432d 5260 spin_unlock_irq(&conf->device_lock);
72626685 5261 bitmap_unplug(mddev->bitmap);
700e432d 5262 spin_lock_irq(&conf->device_lock);
7c13edc8 5263 conf->seq_write = conf->seq_flush;
566c09c5 5264 activate_bit_delay(conf, conf->temp_inactive_list);
72626685 5265 }
0021b7bc 5266 raid5_activate_delayed(conf);
72626685 5267
46031f9a
RBJ		 5268 while ((bio = remove_bio_from_retry(conf))) {
5269 int ok;
5270 spin_unlock_irq(&conf->device_lock);
5271 ok = retry_aligned_read(conf, bio);
5272 spin_lock_irq(&conf->device_lock);
5273 if (!ok)
5274 break;
5275 handled++;
5276 }
5277
566c09c5
SL		 5278 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5279 conf->temp_inactive_list);
773ca82f 5280 if (!batch_size && !released)
1da177e4 5281 break;
46a06401 5282 handled += batch_size;
1da177e4 5283
46a06401
SL		 5284 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5285 spin_unlock_irq(&conf->device_lock);
de393cde 5286 md_check_recovery(mddev);
46a06401
SL		 5287 spin_lock_irq(&conf->device_lock);
5288 }
1da177e4 5289 }
45b4233c 5290 pr_debug("%d stripes handled\n", handled);
1da177e4
LT		 5291
5292 spin_unlock_irq(&conf->device_lock);
5293
c9f21aaf 5294 async_tx_issue_pending_all();
e1dfa0a2 5295 blk_finish_plug(&plug);
1da177e4 5296
45b4233c 5297 pr_debug("--- raid5d inactive\n");
1da177e4
LT		 5298}
5299
3f294f4f 5300static ssize_t
fd01b88c 5301raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
3f294f4f 5302{
d1688a6d 5303 struct r5conf *conf = mddev->private;
96de1e66
N		 5304 if (conf)
5305 return sprintf(page, "%d\n", conf->max_nr_stripes);
5306 else
5307 return 0;
3f294f4f
N		 5308}
5309
c41d4ac4 5310int
fd01b88c 5311raid5_set_cache_size(struct mddev *mddev, int size)
3f294f4f 5312{
d1688a6d 5313 struct r5conf *conf = mddev->private;
b5470dc5 5314 int err;
566c09c5 5315 int hash;
b5470dc5 5316
c41d4ac4 5317 if (size <= 16 || size > 32768)
3f294f4f 5318 return -EINVAL;
566c09c5 5319 hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
c41d4ac4 5320 while (size < conf->max_nr_stripes) {
566c09c5 5321 if (drop_one_stripe(conf, hash))
3f294f4f
N		 5322 conf->max_nr_stripes--;
5323 else
5324 break;
566c09c5
SL		 5325 hash--;
5326 if (hash < 0)
5327 hash = NR_STRIPE_HASH_LOCKS - 1;
3f294f4f 5328 }
b5470dc5
DW		 5329 err = md_allow_write(mddev);
5330 if (err)
5331 return err;
566c09c5 5332 hash = conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
c41d4ac4 5333 while (size > conf->max_nr_stripes) {
566c09c5 5334 if (grow_one_stripe(conf, hash))
3f294f4f
N		 5335 conf->max_nr_stripes++;
5336 else break;
566c09c5 5337 hash = (hash + 1) % NR_STRIPE_HASH_LOCKS;
3f294f4f 5338 }
c41d4ac4
N		 5339 return 0;
5340}
5341EXPORT_SYMBOL(raid5_set_cache_size);
5342
5343static ssize_t
fd01b88c 5344raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
c41d4ac4 5345{
d1688a6d 5346 struct r5conf *conf = mddev->private;
c41d4ac4
N		 5347 unsigned long new;
5348 int err;
5349
5350 if (len >= PAGE_SIZE)
5351 return -EINVAL;
5352 if (!conf)
5353 return -ENODEV;
5354
b29bebd6 5355 if (kstrtoul(page, 10, &new))
c41d4ac4
N		 5356 return -EINVAL;
5357 err = raid5_set_cache_size(mddev, new);
5358 if (err)
5359 return err;
3f294f4f
N		 5360 return len;
5361}
007583c9 5362
96de1e66
N		 5363static struct md_sysfs_entry
5364raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5365 raid5_show_stripe_cache_size,
5366 raid5_store_stripe_cache_size);
3f294f4f 5367
8b3e6cdc 5368static ssize_t
fd01b88c 5369raid5_show_preread_threshold(struct mddev *mddev, char *page)
8b3e6cdc 5370{
d1688a6d 5371 struct r5conf *conf = mddev->private;
8b3e6cdc
DW		 5372 if (conf)
5373 return sprintf(page, "%d\n", conf->bypass_threshold);
5374 else
5375 return 0;
5376}
5377
5378static ssize_t
fd01b88c 5379raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
8b3e6cdc 5380{
d1688a6d 5381 struct r5conf *conf = mddev->private;
4ef197d8 5382 unsigned long new;
8b3e6cdc
DW		 5383 if (len >= PAGE_SIZE)
5384 return -EINVAL;
5385 if (!conf)
5386 return -ENODEV;
5387
b29bebd6 5388 if (kstrtoul(page, 10, &new))
8b3e6cdc 5389 return -EINVAL;
4ef197d8 5390 if (new > conf->max_nr_stripes)
8b3e6cdc
DW		 5391 return -EINVAL;
5392 conf->bypass_threshold = new;
5393 return len;
5394}
5395
5396static struct md_sysfs_entry
5397raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5398 S_IRUGO | S_IWUSR,
5399 raid5_show_preread_threshold,
5400 raid5_store_preread_threshold);
5401
d592a996
SL		 5402static ssize_t
5403raid5_show_skip_copy(struct mddev *mddev, char *page)
5404{
5405 struct r5conf *conf = mddev->private;
5406 if (conf)
5407 return sprintf(page, "%d\n", conf->skip_copy);
5408 else
5409 return 0;
5410}
5411
5412static ssize_t
5413raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
5414{
5415 struct r5conf *conf = mddev->private;
5416 unsigned long new;
5417 if (len >= PAGE_SIZE)
5418 return -EINVAL;
5419 if (!conf)
5420 return -ENODEV;
5421
5422 if (kstrtoul(page, 10, &new))
5423 return -EINVAL;
5424 new = !!new;
5425 if (new == conf->skip_copy)
5426 return len;
5427
5428 mddev_suspend(mddev);
5429 conf->skip_copy = new;
5430 if (new)
5431 mddev->queue->backing_dev_info.capabilities |=
5432 BDI_CAP_STABLE_WRITES;
5433 else
5434 mddev->queue->backing_dev_info.capabilities &=
5435 ~BDI_CAP_STABLE_WRITES;
5436 mddev_resume(mddev);
5437 return len;
5438}
5439
5440static struct md_sysfs_entry
5441raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
5442 raid5_show_skip_copy,
5443 raid5_store_skip_copy);
5444
5445
3f294f4f 5446static ssize_t
fd01b88c 5447stripe_cache_active_show(struct mddev *mddev, char *page)
3f294f4f 5448{
d1688a6d 5449 struct r5conf *conf = mddev->private;
96de1e66
N		 5450 if (conf)
5451 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5452 else
5453 return 0;
3f294f4f
N		 5454}
5455
96de1e66
N		 5456static struct md_sysfs_entry
5457raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
3f294f4f 5458
b721420e
SL		 5459static ssize_t
5460raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
5461{
5462 struct r5conf *conf = mddev->private;
5463 if (conf)
5464 return sprintf(page, "%d\n", conf->worker_cnt_per_group);
5465 else
5466 return 0;
5467}
5468
60aaf933 5469static int alloc_thread_groups(struct r5conf *conf, int cnt,
5470 int *group_cnt,
5471 int *worker_cnt_per_group,
5472 struct r5worker_group **worker_groups);
b721420e
SL		 5473static ssize_t
5474raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
5475{
5476 struct r5conf *conf = mddev->private;
5477 unsigned long new;
5478 int err;
60aaf933 5479 struct r5worker_group *new_groups, *old_groups;
5480 int group_cnt, worker_cnt_per_group;
b721420e
SL		 5481
5482 if (len >= PAGE_SIZE)
5483 return -EINVAL;
5484 if (!conf)
5485 return -ENODEV;
5486
5487 if (kstrtoul(page, 10, &new))
5488 return -EINVAL;
5489
5490 if (new == conf->worker_cnt_per_group)
5491 return len;
5492
5493 mddev_suspend(mddev);
5494
5495 old_groups = conf->worker_groups;
d206dcfa 5496 if (old_groups)
5497 flush_workqueue(raid5_wq);
5498
60aaf933 5499 err = alloc_thread_groups(conf, new,
5500 &group_cnt, &worker_cnt_per_group,
5501 &new_groups);
5502 if (!err) {
5503 spin_lock_irq(&conf->device_lock);
5504 conf->group_cnt = group_cnt;
5505 conf->worker_cnt_per_group = worker_cnt_per_group;
5506 conf->worker_groups = new_groups;
5507 spin_unlock_irq(&conf->device_lock);
b721420e 5508
b721420e
SL		 5509 if (old_groups)
5510 kfree(old_groups[0].workers);
5511 kfree(old_groups);
5512 }
5513
5514 mddev_resume(mddev);
5515
5516 if (err)
5517 return err;
5518 return len;
5519}
5520
5521static struct md_sysfs_entry
5522raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
5523 raid5_show_group_thread_cnt,
5524 raid5_store_group_thread_cnt);
5525
007583c9 5526static struct attribute *raid5_attrs[] = {
3f294f4f
N		 5527 &raid5_stripecache_size.attr,
5528 &raid5_stripecache_active.attr,
8b3e6cdc 5529 &raid5_preread_bypass_threshold.attr,
b721420e 5530 &raid5_group_thread_cnt.attr,
d592a996 5531 &raid5_skip_copy.attr,
3f294f4f
N		 5532 NULL,
5533};
007583c9
N		 5534static struct attribute_group raid5_attrs_group = {
5535 .name = NULL,
5536 .attrs = raid5_attrs,
3f294f4f
N		 5537};
5538
60aaf933 5539static int alloc_thread_groups(struct r5conf *conf, int cnt,
5540 int *group_cnt,
5541 int *worker_cnt_per_group,
5542 struct r5worker_group **worker_groups)
851c30c9 5543{
566c09c5 5544 int i, j, k;
851c30c9
SL		 5545 ssize_t size;
5546 struct r5worker *workers;
5547
60aaf933 5548 *worker_cnt_per_group = cnt;
851c30c9 5549 if (cnt == 0) {
60aaf933 5550 *group_cnt = 0;
5551 *worker_groups = NULL;
851c30c9
SL		 5552 return 0;
5553 }
60aaf933 5554 *group_cnt = num_possible_nodes();
851c30c9 5555 size = sizeof(struct r5worker) * cnt;
60aaf933 5556 workers = kzalloc(size * *group_cnt, GFP_NOIO);
5557 *worker_groups = kzalloc(sizeof(struct r5worker_group) *
5558 *group_cnt, GFP_NOIO);
5559 if (!*worker_groups || !workers) {
851c30c9 5560 kfree(workers);
60aaf933 5561 kfree(*worker_groups);
851c30c9
SL		 5562 return -ENOMEM;
5563 }
5564
60aaf933 5565 for (i = 0; i < *group_cnt; i++) {
851c30c9
SL		 5566 struct r5worker_group *group;
5567
0c775d52 5568 group = &(*worker_groups)[i];
851c30c9
SL		 5569 INIT_LIST_HEAD(&group->handle_list);
5570 group->conf = conf;
5571 group->workers = workers + i * cnt;
5572
5573 for (j = 0; j < cnt; j++) {
566c09c5
SL		 5574 struct r5worker *worker = group->workers + j;
5575 worker->group = group;
5576 INIT_WORK(&worker->work, raid5_do_work);
5577
5578 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
5579 INIT_LIST_HEAD(worker->temp_inactive_list + k);
851c30c9
SL		 5580 }
5581 }
5582
5583 return 0;
5584}
5585
5586static void free_thread_groups(struct r5conf *conf)
5587{
5588 if (conf->worker_groups)
5589 kfree(conf->worker_groups[0].workers);
5590 kfree(conf->worker_groups);
5591 conf->worker_groups = NULL;
5592}
5593
80c3a6ce 5594static sector_t
fd01b88c 5595raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
80c3a6ce 5596{
d1688a6d 5597 struct r5conf *conf = mddev->private;
80c3a6ce
DW		 5598
5599 if (!sectors)
5600 sectors = mddev->dev_sectors;
5e5e3e78 5601 if (!raid_disks)
7ec05478 5602 /* size is defined by the smallest of previous and new size */
5e5e3e78 5603 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
80c3a6ce 5604
9d8f0363 5605 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
664e7c41 5606 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
80c3a6ce
DW		 5607 return sectors * (raid_disks - conf->max_degraded);
5608}
5609
789b5e03
ON		 5610static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5611{
5612 safe_put_page(percpu->spare_page);
5613 kfree(percpu->scribble);
5614 percpu->spare_page = NULL;
5615 percpu->scribble = NULL;
5616}
5617
5618static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5619{
5620 if (conf->level == 6 && !percpu->spare_page)
5621 percpu->spare_page = alloc_page(GFP_KERNEL);
5622 if (!percpu->scribble)
5623 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5624
5625 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
5626 free_scratch_buffer(conf, percpu);
5627 return -ENOMEM;
5628 }
5629
5630 return 0;
5631}
5632
d1688a6d 5633static void raid5_free_percpu(struct r5conf *conf)
36d1c647 5634{
36d1c647
DW		 5635 unsigned long cpu;
5636
5637 if (!conf->percpu)
5638 return;
5639
36d1c647
DW		 5640#ifdef CONFIG_HOTPLUG_CPU
5641 unregister_cpu_notifier(&conf->cpu_notify);
5642#endif
789b5e03
ON		 5643
5644 get_online_cpus();
5645 for_each_possible_cpu(cpu)
5646 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
36d1c647
DW		 5647 put_online_cpus();
5648
5649 free_percpu(conf->percpu);
5650}
5651
d1688a6d 5652static void free_conf(struct r5conf *conf)
95fc17aa 5653{
851c30c9 5654 free_thread_groups(conf);
95fc17aa 5655 shrink_stripes(conf);
36d1c647 5656 raid5_free_percpu(conf);
95fc17aa
DW		 5657 kfree(conf->disks);
5658 kfree(conf->stripe_hashtbl);
5659 kfree(conf);
5660}
5661
36d1c647
DW		 5662#ifdef CONFIG_HOTPLUG_CPU
5663static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5664 void *hcpu)
5665{
d1688a6d 5666 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
36d1c647
DW		 5667 long cpu = (long)hcpu;
5668 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5669
5670 switch (action) {
5671 case CPU_UP_PREPARE:
5672 case CPU_UP_PREPARE_FROZEN:
789b5e03 5673 if (alloc_scratch_buffer(conf, percpu)) {
36d1c647
DW		 5674 pr_err("%s: failed memory allocation for cpu%ld\n",
5675 __func__, cpu);
55af6bb5 5676 return notifier_from_errno(-ENOMEM);
36d1c647
DW		 5677 }
5678 break;
5679 case CPU_DEAD:
5680 case CPU_DEAD_FROZEN:
789b5e03 5681 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
36d1c647
DW		 5682 break;
5683 default:
5684 break;
5685 }
5686 return NOTIFY_OK;
5687}
5688#endif
5689
d1688a6d 5690static int raid5_alloc_percpu(struct r5conf *conf)
36d1c647
DW		 5691{
5692 unsigned long cpu;
789b5e03 5693 int err = 0;
36d1c647 5694
789b5e03
ON		 5695 conf->percpu = alloc_percpu(struct raid5_percpu);
5696 if (!conf->percpu)
36d1c647 5697 return -ENOMEM;
789b5e03
ON		 5698
5699#ifdef CONFIG_HOTPLUG_CPU
5700 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5701 conf->cpu_notify.priority = 0;
5702 err = register_cpu_notifier(&conf->cpu_notify);
5703 if (err)
5704 return err;
5705#endif
36d1c647
DW		 5706
5707 get_online_cpus();
36d1c647 5708 for_each_present_cpu(cpu) {
789b5e03
ON		 5709 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5710 if (err) {
5711 pr_err("%s: failed memory allocation for cpu%ld\n",
5712 __func__, cpu);
36d1c647
DW		 5713 break;
5714 }
36d1c647 5715 }
36d1c647
DW		 5716 put_online_cpus();
5717
5718 return err;
5719}
5720
d1688a6d 5721static struct r5conf *setup_conf(struct mddev *mddev)
1da177e4 5722{
d1688a6d 5723 struct r5conf *conf;
5e5e3e78 5724 int raid_disk, memory, max_disks;
3cb03002 5725 struct md_rdev *rdev;
1da177e4 5726 struct disk_info *disk;
0232605d 5727 char pers_name[6];
566c09c5 5728 int i;
60aaf933 5729 int group_cnt, worker_cnt_per_group;
5730 struct r5worker_group *new_group;
1da177e4 5731
91adb564
N		 5732 if (mddev->new_level != 5
5733 && mddev->new_level != 4
5734 && mddev->new_level != 6) {
0c55e022 5735 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
91adb564
N		 5736 mdname(mddev), mddev->new_level);
5737 return ERR_PTR(-EIO);
1da177e4 5738 }
91adb564
N		 5739 if ((mddev->new_level == 5
5740 && !algorithm_valid_raid5(mddev->new_layout)) ||
5741 (mddev->new_level == 6
5742 && !algorithm_valid_raid6(mddev->new_layout))) {
0c55e022 5743 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
91adb564
N		 5744 mdname(mddev), mddev->new_layout);
5745 return ERR_PTR(-EIO);
99c0fb5f 5746 }
91adb564 5747 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
0c55e022 5748 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
91adb564
N		 5749 mdname(mddev), mddev->raid_disks);
5750 return ERR_PTR(-EINVAL);
4bbf3771
N		 5751 }
5752
664e7c41
AN		 5753 if (!mddev->new_chunk_sectors ||
5754 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5755 !is_power_of_2(mddev->new_chunk_sectors)) {
0c55e022
N		 5756 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5757 mdname(mddev), mddev->new_chunk_sectors << 9);
91adb564 5758 return ERR_PTR(-EINVAL);
f6705578
N		 5759 }
5760
d1688a6d 5761 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
91adb564 5762 if (conf == NULL)
1da177e4 5763 goto abort;
851c30c9 5764 /* Don't enable multi-threading by default*/
60aaf933 5765 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
5766 &new_group)) {
5767 conf->group_cnt = group_cnt;
5768 conf->worker_cnt_per_group = worker_cnt_per_group;
5769 conf->worker_groups = new_group;
5770 } else
851c30c9 5771 goto abort;
f5efd45a 5772 spin_lock_init(&conf->device_lock);
c46501b2 5773 seqcount_init(&conf->gen_lock);
f5efd45a
DW		 5774 init_waitqueue_head(&conf->wait_for_stripe);
5775 init_waitqueue_head(&conf->wait_for_overlap);
5776 INIT_LIST_HEAD(&conf->handle_list);
5777 INIT_LIST_HEAD(&conf->hold_list);
5778 INIT_LIST_HEAD(&conf->delayed_list);
5779 INIT_LIST_HEAD(&conf->bitmap_list);
773ca82f 5780 init_llist_head(&conf->released_stripes);
f5efd45a
DW		 5781 atomic_set(&conf->active_stripes, 0);
5782 atomic_set(&conf->preread_active_stripes, 0);
5783 atomic_set(&conf->active_aligned_reads, 0);
5784 conf->bypass_threshold = BYPASS_THRESHOLD;
d890fa2b 5785 conf->recovery_disabled = mddev->recovery_disabled - 1;
91adb564
N		 5786
5787 conf->raid_disks = mddev->raid_disks;
5788 if (mddev->reshape_position == MaxSector)
5789 conf->previous_raid_disks = mddev->raid_disks;
5790 else
f6705578 5791 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5e5e3e78
N		 5792 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5793 conf->scribble_len = scribble_len(max_disks);
f6705578 5794
5e5e3e78 5795 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
b55e6bfc
N		 5796 GFP_KERNEL);
5797 if (!conf->disks)
5798 goto abort;
9ffae0cf 5799
1da177e4
LT		 5800 conf->mddev = mddev;
5801
fccddba0 5802 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
1da177e4 5803 goto abort;
1da177e4 5804
566c09c5
SL		 5805 /* We init hash_locks[0] separately to that it can be used
5806 * as the reference lock in the spin_lock_nest_lock() call
5807 * in lock_all_device_hash_locks_irq in order to convince
5808 * lockdep that we know what we are doing.
5809 */
5810 spin_lock_init(conf->hash_locks);
5811 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
5812 spin_lock_init(conf->hash_locks + i);
5813
5814 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5815 INIT_LIST_HEAD(conf->inactive_list + i);
5816
5817 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5818 INIT_LIST_HEAD(conf->temp_inactive_list + i);
5819
36d1c647
DW		 5820 conf->level = mddev->new_level;
5821 if (raid5_alloc_percpu(conf) != 0)
5822 goto abort;
5823
0c55e022 5824 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
1da177e4 5825
dafb20fa 5826 rdev_for_each(rdev, mddev) {
1da177e4 5827 raid_disk = rdev->raid_disk;
5e5e3e78 5828 if (raid_disk >= max_disks
1da177e4
LT		 5829 || raid_disk < 0)
5830 continue;
5831 disk = conf->disks + raid_disk;
5832
17045f52
N		 5833 if (test_bit(Replacement, &rdev->flags)) {
5834 if (disk->replacement)
5835 goto abort;
5836 disk->replacement = rdev;
5837 } else {
5838 if (disk->rdev)
5839 goto abort;
5840 disk->rdev = rdev;
5841 }
1da177e4 5842
b2d444d7 5843 if (test_bit(In_sync, &rdev->flags)) {
1da177e4 5844 char b[BDEVNAME_SIZE];
0c55e022
N		 5845 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5846 " disk %d\n",
5847 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
d6b212f4 5848 } else if (rdev->saved_raid_disk != raid_disk)
8c2e870a
NB		 5849 /* Cannot rely on bitmap to complete recovery */
5850 conf->fullsync = 1;
1da177e4
LT		 5851 }
5852
09c9e5fa 5853 conf->chunk_sectors = mddev->new_chunk_sectors;
91adb564 5854 conf->level = mddev->new_level;
16a53ecc
N		 5855 if (conf->level == 6)
5856 conf->max_degraded = 2;
5857 else
5858 conf->max_degraded = 1;
91adb564 5859 conf->algorithm = mddev->new_layout;
fef9c61f 5860 conf->reshape_progress = mddev->reshape_position;
e183eaed 5861 if (conf->reshape_progress != MaxSector) {
09c9e5fa 5862 conf->prev_chunk_sectors = mddev->chunk_sectors;
e183eaed
N		 5863 conf->prev_algo = mddev->layout;
5864 }
1da177e4 5865
91adb564 5866 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5e5e3e78 5867 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4bda556a 5868 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
566c09c5 5869 if (grow_stripes(conf, NR_STRIPES)) {
91adb564 5870 printk(KERN_ERR
0c55e022
N		 5871 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5872 mdname(mddev), memory);
91adb564
N		 5873 goto abort;
5874 } else
0c55e022
N		 5875 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5876 mdname(mddev), memory);
1da177e4 5877
0232605d
N		 5878 sprintf(pers_name, "raid%d", mddev->new_level);
5879 conf->thread = md_register_thread(raid5d, mddev, pers_name);
91adb564
N		 5880 if (!conf->thread) {
5881 printk(KERN_ERR
0c55e022 5882 "md/raid:%s: couldn't allocate thread.\n",
91adb564 5883 mdname(mddev));
16a53ecc
N		 5884 goto abort;
5885 }
91adb564
N		 5886
5887 return conf;
5888
5889 abort:
5890 if (conf) {
95fc17aa 5891 free_conf(conf);
91adb564
N		 5892 return ERR_PTR(-EIO);
5893 } else
5894 return ERR_PTR(-ENOMEM);
5895}
5896
c148ffdc
N		 5897
5898static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5899{
5900 switch (algo) {
5901 case ALGORITHM_PARITY_0:
5902 if (raid_disk < max_degraded)
5903 return 1;
5904 break;
5905 case ALGORITHM_PARITY_N:
5906 if (raid_disk >= raid_disks - max_degraded)
5907 return 1;
5908 break;
5909 case ALGORITHM_PARITY_0_6:
5910 if (raid_disk == 0 ||
5911 raid_disk == raid_disks - 1)
5912 return 1;
5913 break;
5914 case ALGORITHM_LEFT_ASYMMETRIC_6:
5915 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5916 case ALGORITHM_LEFT_SYMMETRIC_6:
5917 case ALGORITHM_RIGHT_SYMMETRIC_6:
5918 if (raid_disk == raid_disks - 1)
5919 return 1;
5920 }
5921 return 0;
5922}
5923
fd01b88c 5924static int run(struct mddev *mddev)
91adb564 5925{
d1688a6d 5926 struct r5conf *conf;
9f7c2220 5927 int working_disks = 0;
c148ffdc 5928 int dirty_parity_disks = 0;
3cb03002 5929 struct md_rdev *rdev;
c148ffdc 5930 sector_t reshape_offset = 0;
17045f52 5931 int i;
b5254dd5
N		 5932 long long min_offset_diff = 0;
5933 int first = 1;
91adb564 5934
8c6ac868 5935 if (mddev->recovery_cp != MaxSector)
0c55e022 5936 printk(KERN_NOTICE "md/raid:%s: not clean"
8c6ac868
AN		 5937 " -- starting background reconstruction\n",
5938 mdname(mddev));
b5254dd5
N		 5939
5940 rdev_for_each(rdev, mddev) {
5941 long long diff;
5942 if (rdev->raid_disk < 0)
5943 continue;
5944 diff = (rdev->new_data_offset - rdev->data_offset);
5945 if (first) {
5946 min_offset_diff = diff;
5947 first = 0;
5948 } else if (mddev->reshape_backwards &&
5949 diff < min_offset_diff)
5950 min_offset_diff = diff;
5951 else if (!mddev->reshape_backwards &&
5952 diff > min_offset_diff)
5953 min_offset_diff = diff;
5954 }
5955
91adb564
N		 5956 if (mddev->reshape_position != MaxSector) {
5957 /* Check that we can continue the reshape.
b5254dd5
N		 5958 * Difficulties arise if the stripe we would write to
5959 * next is at or after the stripe we would read from next.
5960 * For a reshape that changes the number of devices, this
5961 * is only possible for a very short time, and mdadm makes
5962 * sure that time appears to have past before assembling
5963 * the array. So we fail if that time hasn't passed.
5964 * For a reshape that keeps the number of devices the same
5965 * mdadm must be monitoring the reshape can keeping the
5966 * critical areas read-only and backed up. It will start
5967 * the array in read-only mode, so we check for that.
91adb564
N		 5968 */
5969 sector_t here_new, here_old;
5970 int old_disks;
18b00334 5971 int max_degraded = (mddev->level == 6 ? 2 : 1);
91adb564 5972
88ce4930 5973 if (mddev->new_level != mddev->level) {
0c55e022 5974 printk(KERN_ERR "md/raid:%s: unsupported reshape "
91adb564
N		 5975 "required - aborting.\n",
5976 mdname(mddev));
5977 return -EINVAL;
5978 }
91adb564
N		 5979 old_disks = mddev->raid_disks - mddev->delta_disks;
5980 /* reshape_position must be on a new-stripe boundary, and one
5981 * further up in new geometry must map after here in old
5982 * geometry.
5983 */
5984 here_new = mddev->reshape_position;
664e7c41 5985 if (sector_div(here_new, mddev->new_chunk_sectors *
91adb564 5986 (mddev->raid_disks - max_degraded))) {
0c55e022
N		 5987 printk(KERN_ERR "md/raid:%s: reshape_position not "
5988 "on a stripe boundary\n", mdname(mddev));
91adb564
N		 5989 return -EINVAL;
5990 }
c148ffdc 5991 reshape_offset = here_new * mddev->new_chunk_sectors;
91adb564
N		 5992 /* here_new is the stripe we will write to */
5993 here_old = mddev->reshape_position;
9d8f0363 5994 sector_div(here_old, mddev->chunk_sectors *
91adb564
N		 5995 (old_disks-max_degraded));
5996 /* here_old is the first stripe that we might need to read
5997 * from */
67ac6011 5998 if (mddev->delta_disks == 0) {
b5254dd5
N		 5999 if ((here_new * mddev->new_chunk_sectors !=
6000 here_old * mddev->chunk_sectors)) {
6001 printk(KERN_ERR "md/raid:%s: reshape position is"
6002 " confused - aborting\n", mdname(mddev));
6003 return -EINVAL;
6004 }
67ac6011 6005 /* We cannot be sure it is safe to start an in-place
b5254dd5 6006 * reshape. It is only safe if user-space is monitoring
67ac6011
N		 6007 * and taking constant backups.
6008 * mdadm always starts a situation like this in
6009 * readonly mode so it can take control before
6010 * allowing any writes. So just check for that.
6011 */
b5254dd5
N		 6012 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6013 abs(min_offset_diff) >= mddev->new_chunk_sectors)
6014 /* not really in-place - so OK */;
6015 else if (mddev->ro == 0) {
6016 printk(KERN_ERR "md/raid:%s: in-place reshape "
6017 "must be started in read-only mode "
6018 "- aborting\n",
0c55e022 6019 mdname(mddev));
67ac6011
N		 6020 return -EINVAL;
6021 }
2c810cdd 6022 } else if (mddev->reshape_backwards
b5254dd5 6023 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
67ac6011
N		 6024 here_old * mddev->chunk_sectors)
6025 : (here_new * mddev->new_chunk_sectors >=
b5254dd5 6026 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
91adb564 6027 /* Reading from the same stripe as writing to - bad */
0c55e022
N		 6028 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6029 "auto-recovery - aborting.\n",
6030 mdname(mddev));
91adb564
N		 6031 return -EINVAL;
6032 }
0c55e022
N		 6033 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6034 mdname(mddev));
91adb564
N		 6035 /* OK, we should be able to continue; */
6036 } else {
6037 BUG_ON(mddev->level != mddev->new_level);
6038 BUG_ON(mddev->layout != mddev->new_layout);
664e7c41 6039 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
91adb564 6040 BUG_ON(mddev->delta_disks != 0);
1da177e4 6041 }
91adb564 6042
245f46c2
N		 6043 if (mddev->private == NULL)
6044 conf = setup_conf(mddev);
6045 else
6046 conf = mddev->private;
6047
91adb564
N		 6048 if (IS_ERR(conf))
6049 return PTR_ERR(conf);
6050
b5254dd5 6051 conf->min_offset_diff = min_offset_diff;
91adb564
N		 6052 mddev->thread = conf->thread;
6053 conf->thread = NULL;
6054 mddev->private = conf;
6055
17045f52
N		 6056 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6057 i++) {
6058 rdev = conf->disks[i].rdev;
6059 if (!rdev && conf->disks[i].replacement) {
6060 /* The replacement is all we have yet */
6061 rdev = conf->disks[i].replacement;
6062 conf->disks[i].replacement = NULL;
6063 clear_bit(Replacement, &rdev->flags);
6064 conf->disks[i].rdev = rdev;
6065 }
6066 if (!rdev)
c148ffdc 6067 continue;
17045f52
N		 6068 if (conf->disks[i].replacement &&
6069 conf->reshape_progress != MaxSector) {
6070 /* replacements and reshape simply do not mix. */
6071 printk(KERN_ERR "md: cannot handle concurrent "
6072 "replacement and reshape.\n");
6073 goto abort;
6074 }
2f115882 6075 if (test_bit(In_sync, &rdev->flags)) {
91adb564 6076 working_disks++;
2f115882
N		 6077 continue;
6078 }
c148ffdc
N		 6079 /* This disc is not fully in-sync. However if it
6080 * just stored parity (beyond the recovery_offset),
6081 * when we don't need to be concerned about the
6082 * array being dirty.
6083 * When reshape goes 'backwards', we never have
6084 * partially completed devices, so we only need
6085 * to worry about reshape going forwards.
6086 */
6087 /* Hack because v0.91 doesn't store recovery_offset properly. */
6088 if (mddev->major_version == 0 &&
6089 mddev->minor_version > 90)
6090 rdev->recovery_offset = reshape_offset;
5026d7a9 6091
c148ffdc
N		 6092 if (rdev->recovery_offset < reshape_offset) {
6093 /* We need to check old and new layout */
6094 if (!only_parity(rdev->raid_disk,
6095 conf->algorithm,
6096 conf->raid_disks,
6097 conf->max_degraded))
6098 continue;
6099 }
6100 if (!only_parity(rdev->raid_disk,
6101 conf->prev_algo,
6102 conf->previous_raid_disks,
6103 conf->max_degraded))
6104 continue;
6105 dirty_parity_disks++;
6106 }
91adb564 6107
17045f52
N		 6108 /*
6109 * 0 for a fully functional array, 1 or 2 for a degraded array.
6110 */
908f4fbd 6111 mddev->degraded = calc_degraded(conf);
91adb564 6112
674806d6 6113 if (has_failed(conf)) {
0c55e022 6114 printk(KERN_ERR "md/raid:%s: not enough operational devices"
1da177e4 6115 " (%d/%d failed)\n",
02c2de8c 6116 mdname(mddev), mddev->degraded, conf->raid_disks);
1da177e4
LT		 6117 goto abort;
6118 }
6119
91adb564 6120 /* device size must be a multiple of chunk size */
9d8f0363 6121 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
91adb564
N		 6122 mddev->resync_max_sectors = mddev->dev_sectors;
6123
c148ffdc 6124 if (mddev->degraded > dirty_parity_disks &&
1da177e4 6125 mddev->recovery_cp != MaxSector) {
6ff8d8ec
N		 6126 if (mddev->ok_start_degraded)
6127 printk(KERN_WARNING
0c55e022
N		 6128 "md/raid:%s: starting dirty degraded array"
6129 " - data corruption possible.\n",
6ff8d8ec
N		 6130 mdname(mddev));
6131 else {
6132 printk(KERN_ERR
0c55e022 6133 "md/raid:%s: cannot start dirty degraded array.\n",
6ff8d8ec
N		 6134 mdname(mddev));
6135 goto abort;
6136 }
1da177e4
LT		 6137 }
6138
1da177e4 6139 if (mddev->degraded == 0)
0c55e022
N		 6140 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6141 " devices, algorithm %d\n", mdname(mddev), conf->level,
e183eaed
N		 6142 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6143 mddev->new_layout);
1da177e4 6144 else
0c55e022
N		 6145 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6146 " out of %d devices, algorithm %d\n",
6147 mdname(mddev), conf->level,
6148 mddev->raid_disks - mddev->degraded,
6149 mddev->raid_disks, mddev->new_layout);
1da177e4
LT		 6150
6151 print_raid5_conf(conf);
6152
fef9c61f 6153 if (conf->reshape_progress != MaxSector) {
fef9c61f 6154 conf->reshape_safe = conf->reshape_progress;
f6705578
N		 6155 atomic_set(&conf->reshape_stripes, 0);
6156 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6157 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6158 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6159 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6160 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
0da3c619 6161 "reshape");
f6705578
N		 6162 }
6163
1da177e4
LT		 6164
6165 /* Ok, everything is just fine now */
a64c876f
N		 6166 if (mddev->to_remove == &raid5_attrs_group)
6167 mddev->to_remove = NULL;
00bcb4ac
N		 6168 else if (mddev->kobj.sd &&
6169 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5e55e2f5 6170 printk(KERN_WARNING
4a5add49 6171 "raid5: failed to create sysfs attributes for %s\n",
5e55e2f5 6172 mdname(mddev));
4a5add49 6173 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7a5febe9 6174
4a5add49 6175 if (mddev->queue) {
9f7c2220 6176 int chunk_size;
620125f2 6177 bool discard_supported = true;
4a5add49
N		 6178 /* read-ahead size must cover two whole stripes, which
6179 * is 2 * (datadisks) * chunksize where 'n' is the
6180 * number of raid devices
6181 */
6182 int data_disks = conf->previous_raid_disks - conf->max_degraded;
6183 int stripe = data_disks *
6184 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
6185 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6186 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
91adb564 6187
4a5add49 6188 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
f022b2fd 6189
11d8a6e3
N		 6190 mddev->queue->backing_dev_info.congested_data = mddev;
6191 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
7a5febe9 6192
9f7c2220
N		 6193 chunk_size = mddev->chunk_sectors << 9;
6194 blk_queue_io_min(mddev->queue, chunk_size);
6195 blk_queue_io_opt(mddev->queue, chunk_size *
6196 (conf->raid_disks - conf->max_degraded));
c78afc62 6197 mddev->queue->limits.raid_partial_stripes_expensive = 1;
620125f2
SL		 6198 /*
6199 * We can only discard a whole stripe. It doesn't make sense to
6200 * discard data disk but write parity disk
6201 */
6202 stripe = stripe * PAGE_SIZE;
4ac6875e
N		 6203 /* Round up to power of 2, as discard handling
6204 * currently assumes that */
6205 while ((stripe-1) & stripe)
6206 stripe = (stripe | (stripe-1)) + 1;
620125f2
SL		 6207 mddev->queue->limits.discard_alignment = stripe;
6208 mddev->queue->limits.discard_granularity = stripe;
6209 /*
6210 * unaligned part of discard request will be ignored, so can't
6211 * guarantee discard_zerors_data
6212 */
6213 mddev->queue->limits.discard_zeroes_data = 0;
8f6c2e4b 6214
5026d7a9
PA		 6215 blk_queue_max_write_same_sectors(mddev->queue, 0);
6216
05616be5 6217 rdev_for_each(rdev, mddev) {
9f7c2220
N		 6218 disk_stack_limits(mddev->gendisk, rdev->bdev,
6219 rdev->data_offset << 9);
05616be5
N		 6220 disk_stack_limits(mddev->gendisk, rdev->bdev,
6221 rdev->new_data_offset << 9);
620125f2
SL		 6222 /*
6223 * discard_zeroes_data is required, otherwise data
6224 * could be lost. Consider a scenario: discard a stripe
6225 * (the stripe could be inconsistent if
6226 * discard_zeroes_data is 0); write one disk of the
6227 * stripe (the stripe could be inconsistent again
6228 * depending on which disks are used to calculate
6229 * parity); the disk is broken; The stripe data of this
6230 * disk is lost.
6231 */
6232 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6233 !bdev_get_queue(rdev->bdev)->
6234 limits.discard_zeroes_data)
6235 discard_supported = false;
05616be5 6236 }
620125f2
SL		 6237
6238 if (discard_supported &&
6239 mddev->queue->limits.max_discard_sectors >= stripe &&
6240 mddev->queue->limits.discard_granularity >= stripe)
6241 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6242 mddev->queue);
6243 else
6244 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6245 mddev->queue);
9f7c2220 6246 }
23032a0e 6247
1da177e4
LT		 6248 return 0;
6249abort:
01f96c0a 6250 md_unregister_thread(&mddev->thread);
e4f869d9
N		 6251 print_raid5_conf(conf);
6252 free_conf(conf);
1da177e4 6253 mddev->private = NULL;
0c55e022 6254 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
1da177e4
LT		 6255 return -EIO;
6256}
6257
fd01b88c 6258static int stop(struct mddev *mddev)
1da177e4 6259{
d1688a6d 6260 struct r5conf *conf = mddev->private;
1da177e4 6261
01f96c0a 6262 md_unregister_thread(&mddev->thread);
11d8a6e3
N		 6263 if (mddev->queue)
6264 mddev->queue->backing_dev_info.congested_fn = NULL;
95fc17aa 6265 free_conf(conf);
a64c876f
N		 6266 mddev->private = NULL;
6267 mddev->to_remove = &raid5_attrs_group;
1da177e4
LT		 6268 return 0;
6269}
6270
fd01b88c 6271static void status(struct seq_file *seq, struct mddev *mddev)
1da177e4 6272{
d1688a6d 6273 struct r5conf *conf = mddev->private;
1da177e4
LT		 6274 int i;
6275
9d8f0363
AN		 6276 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6277 mddev->chunk_sectors / 2, mddev->layout);
02c2de8c 6278 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
1da177e4
LT		 6279 for (i = 0; i < conf->raid_disks; i++)
6280 seq_printf (seq, "%s",
6281 conf->disks[i].rdev &&
b2d444d7 6282 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
1da177e4 6283 seq_printf (seq, "]");
1da177e4
LT		 6284}
6285
d1688a6d 6286static void print_raid5_conf (struct r5conf *conf)
1da177e4
LT		 6287{
6288 int i;
6289 struct disk_info *tmp;
6290
0c55e022 6291 printk(KERN_DEBUG "RAID conf printout:\n");
1da177e4
LT		 6292 if (!conf) {
6293 printk("(conf==NULL)\n");
6294 return;
6295 }
0c55e022
N		 6296 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6297 conf->raid_disks,
6298 conf->raid_disks - conf->mddev->degraded);
1da177e4
LT		 6299
6300 for (i = 0; i < conf->raid_disks; i++) {
6301 char b[BDEVNAME_SIZE];
6302 tmp = conf->disks + i;
6303 if (tmp->rdev)
0c55e022
N		 6304 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6305 i, !test_bit(Faulty, &tmp->rdev->flags),
6306 bdevname(tmp->rdev->bdev, b));
1da177e4
LT		 6307 }
6308}
6309
fd01b88c 6310static int raid5_spare_active(struct mddev *mddev)
1da177e4
LT		 6311{
6312 int i;
d1688a6d 6313 struct r5conf *conf = mddev->private;
1da177e4 6314 struct disk_info *tmp;
6b965620
N		 6315 int count = 0;
6316 unsigned long flags;
1da177e4
LT		 6317
6318 for (i = 0; i < conf->raid_disks; i++) {
6319 tmp = conf->disks + i;
dd054fce
N		 6320 if (tmp->replacement
6321 && tmp->replacement->recovery_offset == MaxSector
6322 && !test_bit(Faulty, &tmp->replacement->flags)
6323 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
6324 /* Replacement has just become active. */
6325 if (!tmp->rdev
6326 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
6327 count++;
6328 if (tmp->rdev) {
6329 /* Replaced device not technically faulty,
6330 * but we need to be sure it gets removed
6331 * and never re-added.
6332 */
6333 set_bit(Faulty, &tmp->rdev->flags);
6334 sysfs_notify_dirent_safe(
6335 tmp->rdev->sysfs_state);
6336 }
6337 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
6338 } else if (tmp->rdev
70fffd0b 6339 && tmp->rdev->recovery_offset == MaxSector
b2d444d7 6340 && !test_bit(Faulty, &tmp->rdev->flags)
c04be0aa 6341 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
6b965620 6342 count++;
43c73ca4 6343 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1da177e4
LT		 6344 }
6345 }
6b965620 6346 spin_lock_irqsave(&conf->device_lock, flags);
908f4fbd 6347 mddev->degraded = calc_degraded(conf);
6b965620 6348 spin_unlock_irqrestore(&conf->device_lock, flags);
1da177e4 6349 print_raid5_conf(conf);
6b965620 6350 return count;
1da177e4
LT		 6351}
6352
b8321b68 6353static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1da177e4 6354{
d1688a6d 6355 struct r5conf *conf = mddev->private;
1da177e4 6356 int err = 0;
b8321b68 6357 int number = rdev->raid_disk;
657e3e4d 6358 struct md_rdev **rdevp;
1da177e4
LT		 6359 struct disk_info *p = conf->disks + number;
6360
6361 print_raid5_conf(conf);
657e3e4d
N		 6362 if (rdev == p->rdev)
6363 rdevp = &p->rdev;
6364 else if (rdev == p->replacement)
6365 rdevp = &p->replacement;
6366 else
6367 return 0;
6368
6369 if (number >= conf->raid_disks &&
6370 conf->reshape_progress == MaxSector)
6371 clear_bit(In_sync, &rdev->flags);
6372
6373 if (test_bit(In_sync, &rdev->flags) ||
6374 atomic_read(&rdev->nr_pending)) {
6375 err = -EBUSY;
6376 goto abort;
6377 }
6378 /* Only remove non-faulty devices if recovery
6379 * isn't possible.
6380 */
6381 if (!test_bit(Faulty, &rdev->flags) &&
6382 mddev->recovery_disabled != conf->recovery_disabled &&
6383 !has_failed(conf) &&
dd054fce 6384 (!p->replacement || p->replacement == rdev) &&
657e3e4d
N		 6385 number < conf->raid_disks) {
6386 err = -EBUSY;
6387 goto abort;
6388 }
6389 *rdevp = NULL;
6390 synchronize_rcu();
6391 if (atomic_read(&rdev->nr_pending)) {
6392 /* lost the race, try later */
6393 err = -EBUSY;
6394 *rdevp = rdev;
dd054fce
N		 6395 } else if (p->replacement) {
6396 /* We must have just cleared 'rdev' */
6397 p->rdev = p->replacement;
6398 clear_bit(Replacement, &p->replacement->flags);
6399 smp_mb(); /* Make sure other CPUs may see both as identical
6400 * but will never see neither - if they are careful
6401 */
6402 p->replacement = NULL;
6403 clear_bit(WantReplacement, &rdev->flags);
6404 } else
6405 /* We might have just removed the Replacement as faulty-
6406 * clear the bit just in case
6407 */
6408 clear_bit(WantReplacement, &rdev->flags);
1da177e4
LT		 6409abort:
6410
6411 print_raid5_conf(conf);
6412 return err;
6413}
6414
fd01b88c 6415static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1da177e4 6416{
d1688a6d 6417 struct r5conf *conf = mddev->private;
199050ea 6418 int err = -EEXIST;
1da177e4
LT		 6419 int disk;
6420 struct disk_info *p;
6c2fce2e
NB		 6421 int first = 0;
6422 int last = conf->raid_disks - 1;
1da177e4 6423
7f0da59b
N		 6424 if (mddev->recovery_disabled == conf->recovery_disabled)
6425 return -EBUSY;
6426
dc10c643 6427 if (rdev->saved_raid_disk < 0 && has_failed(conf))
1da177e4 6428 /* no point adding a device */
199050ea 6429 return -EINVAL;
1da177e4 6430
6c2fce2e
NB		 6431 if (rdev->raid_disk >= 0)
6432 first = last = rdev->raid_disk;
1da177e4
LT		 6433
6434 /*
16a53ecc
N		 6435 * find the disk ... but prefer rdev->saved_raid_disk
6436 * if possible.
1da177e4 6437 */
16a53ecc 6438 if (rdev->saved_raid_disk >= 0 &&
6c2fce2e 6439 rdev->saved_raid_disk >= first &&
16a53ecc 6440 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5cfb22a1
N		 6441 first = rdev->saved_raid_disk;
6442
6443 for (disk = first; disk <= last; disk++) {
7bfec5f3
N		 6444 p = conf->disks + disk;
6445 if (p->rdev == NULL) {
b2d444d7 6446 clear_bit(In_sync, &rdev->flags);
1da177e4 6447 rdev->raid_disk = disk;
199050ea 6448 err = 0;
72626685
N		 6449 if (rdev->saved_raid_disk != disk)
6450 conf->fullsync = 1;
d6065f7b 6451 rcu_assign_pointer(p->rdev, rdev);
5cfb22a1 6452 goto out;
1da177e4 6453 }
5cfb22a1
N		 6454 }
6455 for (disk = first; disk <= last; disk++) {
6456 p = conf->disks + disk;
7bfec5f3
N		 6457 if (test_bit(WantReplacement, &p->rdev->flags) &&
6458 p->replacement == NULL) {
6459 clear_bit(In_sync, &rdev->flags);
6460 set_bit(Replacement, &rdev->flags);
6461 rdev->raid_disk = disk;
6462 err = 0;
6463 conf->fullsync = 1;
6464 rcu_assign_pointer(p->replacement, rdev);
6465 break;
6466 }
6467 }
5cfb22a1 6468out:
1da177e4 6469 print_raid5_conf(conf);
199050ea 6470 return err;
1da177e4
LT		 6471}
6472
fd01b88c 6473static int raid5_resize(struct mddev *mddev, sector_t sectors)
1da177e4
LT		 6474{
6475 /* no resync is happening, and there is enough space
6476 * on all devices, so we can resize.
6477 * We need to make sure resync covers any new space.
6478 * If the array is shrinking we should possibly wait until
6479 * any io in the removed space completes, but it hardly seems
6480 * worth it.
6481 */
a4a6125a 6482 sector_t newsize;
9d8f0363 6483 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
a4a6125a
N		 6484 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
6485 if (mddev->external_size &&
6486 mddev->array_sectors > newsize)
b522adcd 6487 return -EINVAL;
a4a6125a
N		 6488 if (mddev->bitmap) {
6489 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
6490 if (ret)
6491 return ret;
6492 }
6493 md_set_array_sectors(mddev, newsize);
f233ea5c 6494 set_capacity(mddev->gendisk, mddev->array_sectors);
449aad3e 6495 revalidate_disk(mddev->gendisk);
b098636c
N		 6496 if (sectors > mddev->dev_sectors &&
6497 mddev->recovery_cp > mddev->dev_sectors) {
58c0fed4 6498 mddev->recovery_cp = mddev->dev_sectors;
1da177e4
LT		 6499 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
6500 }
58c0fed4 6501 mddev->dev_sectors = sectors;
4b5c7ae8 6502 mddev->resync_max_sectors = sectors;
1da177e4
LT		 6503 return 0;
6504}
6505
fd01b88c 6506static int check_stripe_cache(struct mddev *mddev)
01ee22b4
N		 6507{
6508 /* Can only proceed if there are plenty of stripe_heads.
6509 * We need a minimum of one full stripe,, and for sensible progress
6510 * it is best to have about 4 times that.
6511 * If we require 4 times, then the default 256 4K stripe_heads will
6512 * allow for chunk sizes up to 256K, which is probably OK.
6513 * If the chunk size is greater, user-space should request more
6514 * stripe_heads first.
6515 */
d1688a6d 6516 struct r5conf *conf = mddev->private;
01ee22b4
N		 6517 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
6518 > conf->max_nr_stripes ||
6519 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
6520 > conf->max_nr_stripes) {
0c55e022
N		 6521 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
6522 mdname(mddev),
01ee22b4
N		 6523 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
6524 / STRIPE_SIZE)*4);
6525 return 0;
6526 }
6527 return 1;
6528}
6529
fd01b88c 6530static int check_reshape(struct mddev *mddev)
29269553 6531{
d1688a6d 6532 struct r5conf *conf = mddev->private;
29269553 6533
88ce4930
N		 6534 if (mddev->delta_disks == 0 &&
6535 mddev->new_layout == mddev->layout &&
664e7c41 6536 mddev->new_chunk_sectors == mddev->chunk_sectors)
50ac168a 6537 return 0; /* nothing to do */
674806d6 6538 if (has_failed(conf))
ec32a2bd 6539 return -EINVAL;
fdcfbbb6 6540 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
ec32a2bd
N		 6541 /* We might be able to shrink, but the devices must
6542 * be made bigger first.
6543 * For raid6, 4 is the minimum size.
6544 * Otherwise 2 is the minimum
6545 */
6546 int min = 2;
6547 if (mddev->level == 6)
6548 min = 4;
6549 if (mddev->raid_disks + mddev->delta_disks < min)
6550 return -EINVAL;
6551 }
29269553 6552
01ee22b4 6553 if (!check_stripe_cache(mddev))
29269553 6554 return -ENOSPC;
29269553 6555
e56108d6
N		 6556 return resize_stripes(conf, (conf->previous_raid_disks
6557 + mddev->delta_disks));
63c70c4f
N		 6558}
6559
fd01b88c 6560static int raid5_start_reshape(struct mddev *mddev)
63c70c4f 6561{
d1688a6d 6562 struct r5conf *conf = mddev->private;
3cb03002 6563 struct md_rdev *rdev;
63c70c4f 6564 int spares = 0;
c04be0aa 6565 unsigned long flags;
63c70c4f 6566
f416885e 6567 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
63c70c4f
N		 6568 return -EBUSY;
6569
01ee22b4
N		 6570 if (!check_stripe_cache(mddev))
6571 return -ENOSPC;
6572
30b67645
N		 6573 if (has_failed(conf))
6574 return -EINVAL;
6575
c6563a8c 6576 rdev_for_each(rdev, mddev) {
469518a3
N		 6577 if (!test_bit(In_sync, &rdev->flags)
6578 && !test_bit(Faulty, &rdev->flags))
29269553 6579 spares++;
c6563a8c 6580 }
63c70c4f 6581
f416885e 6582 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
29269553
N		 6583 /* Not enough devices even to make a degraded array
6584 * of that size
6585 */
6586 return -EINVAL;
6587
ec32a2bd
N		 6588 /* Refuse to reduce size of the array. Any reductions in
6589 * array size must be through explicit setting of array_size
6590 * attribute.
6591 */
6592 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
6593 < mddev->array_sectors) {
0c55e022 6594 printk(KERN_ERR "md/raid:%s: array size must be reduced "
ec32a2bd
N		 6595 "before number of disks\n", mdname(mddev));
6596 return -EINVAL;
6597 }
6598
f6705578 6599 atomic_set(&conf->reshape_stripes, 0);
29269553 6600 spin_lock_irq(&conf->device_lock);
c46501b2 6601 write_seqcount_begin(&conf->gen_lock);
29269553 6602 conf->previous_raid_disks = conf->raid_disks;
63c70c4f 6603 conf->raid_disks += mddev->delta_disks;
09c9e5fa
AN		 6604 conf->prev_chunk_sectors = conf->chunk_sectors;
6605 conf->chunk_sectors = mddev->new_chunk_sectors;
88ce4930
N		 6606 conf->prev_algo = conf->algorithm;
6607 conf->algorithm = mddev->new_layout;
05616be5
N		 6608 conf->generation++;
6609 /* Code that selects data_offset needs to see the generation update
6610 * if reshape_progress has been set - so a memory barrier needed.
6611 */
6612 smp_mb();
2c810cdd 6613 if (mddev->reshape_backwards)
fef9c61f
N		 6614 conf->reshape_progress = raid5_size(mddev, 0, 0);
6615 else
6616 conf->reshape_progress = 0;
6617 conf->reshape_safe = conf->reshape_progress;
c46501b2 6618 write_seqcount_end(&conf->gen_lock);
29269553
N		 6619 spin_unlock_irq(&conf->device_lock);
6620
4d77e3ba
N		 6621 /* Now make sure any requests that proceeded on the assumption
6622 * the reshape wasn't running - like Discard or Read - have
6623 * completed.
6624 */
6625 mddev_suspend(mddev);
6626 mddev_resume(mddev);
6627
29269553
N		 6628 /* Add some new drives, as many as will fit.
6629 * We know there are enough to make the newly sized array work.
3424bf6a
N		 6630 * Don't add devices if we are reducing the number of
6631 * devices in the array. This is because it is not possible
6632 * to correctly record the "partially reconstructed" state of
6633 * such devices during the reshape and confusion could result.
29269553 6634 */
87a8dec9 6635 if (mddev->delta_disks >= 0) {
dafb20fa 6636 rdev_for_each(rdev, mddev)
87a8dec9
N		 6637 if (rdev->raid_disk < 0 &&
6638 !test_bit(Faulty, &rdev->flags)) {
6639 if (raid5_add_disk(mddev, rdev) == 0) {
87a8dec9 6640 if (rdev->raid_disk
9d4c7d87 6641 >= conf->previous_raid_disks)
87a8dec9 6642 set_bit(In_sync, &rdev->flags);
9d4c7d87 6643 else
87a8dec9 6644 rdev->recovery_offset = 0;
36fad858
NK		 6645
6646 if (sysfs_link_rdev(mddev, rdev))
87a8dec9 6647 /* Failure here is OK */;
50da0840 6648 }
87a8dec9
N		 6649 } else if (rdev->raid_disk >= conf->previous_raid_disks
6650 && !test_bit(Faulty, &rdev->flags)) {
6651 /* This is a spare that was manually added */
6652 set_bit(In_sync, &rdev->flags);
87a8dec9 6653 }
29269553 6654
87a8dec9
N		 6655 /* When a reshape changes the number of devices,
6656 * ->degraded is measured against the larger of the
6657 * pre and post number of devices.
6658 */
ec32a2bd 6659 spin_lock_irqsave(&conf->device_lock, flags);
908f4fbd 6660 mddev->degraded = calc_degraded(conf);
ec32a2bd
N		 6661 spin_unlock_irqrestore(&conf->device_lock, flags);
6662 }
63c70c4f 6663 mddev->raid_disks = conf->raid_disks;
e516402c 6664 mddev->reshape_position = conf->reshape_progress;
850b2b42 6665 set_bit(MD_CHANGE_DEVS, &mddev->flags);
f6705578 6666
29269553
N		 6667 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6668 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6669 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6670 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6671 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
0da3c619 6672 "reshape");
29269553
N		 6673 if (!mddev->sync_thread) {
6674 mddev->recovery = 0;
6675 spin_lock_irq(&conf->device_lock);
ba8805b9 6676 write_seqcount_begin(&conf->gen_lock);
29269553 6677 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
ba8805b9
N		 6678 mddev->new_chunk_sectors =
6679 conf->chunk_sectors = conf->prev_chunk_sectors;
6680 mddev->new_layout = conf->algorithm = conf->prev_algo;
05616be5
N		 6681 rdev_for_each(rdev, mddev)
6682 rdev->new_data_offset = rdev->data_offset;
6683 smp_wmb();
ba8805b9 6684 conf->generation --;
fef9c61f 6685 conf->reshape_progress = MaxSector;
1e3fa9bd 6686 mddev->reshape_position = MaxSector;
ba8805b9 6687 write_seqcount_end(&conf->gen_lock);
29269553
N		 6688 spin_unlock_irq(&conf->device_lock);
6689 return -EAGAIN;
6690 }
c8f517c4 6691 conf->reshape_checkpoint = jiffies;
29269553
N		 6692 md_wakeup_thread(mddev->sync_thread);
6693 md_new_event(mddev);
6694 return 0;
6695}
29269553 6696
ec32a2bd
N		 6697/* This is called from the reshape thread and should make any
6698 * changes needed in 'conf'
6699 */
d1688a6d 6700static void end_reshape(struct r5conf *conf)
29269553 6701{
29269553 6702
f6705578 6703 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
05616be5 6704 struct md_rdev *rdev;
f6705578 6705
f6705578 6706 spin_lock_irq(&conf->device_lock);
cea9c228 6707 conf->previous_raid_disks = conf->raid_disks;
05616be5
N		 6708 rdev_for_each(rdev, conf->mddev)
6709 rdev->data_offset = rdev->new_data_offset;
6710 smp_wmb();
fef9c61f 6711 conf->reshape_progress = MaxSector;
f6705578 6712 spin_unlock_irq(&conf->device_lock);
b0f9ec04 6713 wake_up(&conf->wait_for_overlap);
16a53ecc
N		 6714
6715 /* read-ahead size must cover two whole stripes, which is
6716 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6717 */
4a5add49 6718 if (conf->mddev->queue) {
cea9c228 6719 int data_disks = conf->raid_disks - conf->max_degraded;
09c9e5fa 6720 int stripe = data_disks * ((conf->chunk_sectors << 9)
cea9c228 6721 / PAGE_SIZE);
16a53ecc
N		 6722 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6723 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6724 }
29269553 6725 }
29269553
N		 6726}
6727
ec32a2bd
N		 6728/* This is called from the raid5d thread with mddev_lock held.
6729 * It makes config changes to the device.
6730 */
fd01b88c 6731static void raid5_finish_reshape(struct mddev *mddev)
cea9c228 6732{
d1688a6d 6733 struct r5conf *conf = mddev->private;
cea9c228
N		 6734
6735 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6736
ec32a2bd
N		 6737 if (mddev->delta_disks > 0) {
6738 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6739 set_capacity(mddev->gendisk, mddev->array_sectors);
449aad3e 6740 revalidate_disk(mddev->gendisk);
ec32a2bd
N		 6741 } else {
6742 int d;
908f4fbd
N		 6743 spin_lock_irq(&conf->device_lock);
6744 mddev->degraded = calc_degraded(conf);
6745 spin_unlock_irq(&conf->device_lock);
ec32a2bd
N		 6746 for (d = conf->raid_disks ;
6747 d < conf->raid_disks - mddev->delta_disks;
1a67dde0 6748 d++) {
3cb03002 6749 struct md_rdev *rdev = conf->disks[d].rdev;
da7613b8
N		 6750 if (rdev)
6751 clear_bit(In_sync, &rdev->flags);
6752 rdev = conf->disks[d].replacement;
6753 if (rdev)
6754 clear_bit(In_sync, &rdev->flags);
1a67dde0 6755 }
cea9c228 6756 }
88ce4930 6757 mddev->layout = conf->algorithm;
09c9e5fa 6758 mddev->chunk_sectors = conf->chunk_sectors;
ec32a2bd
N		 6759 mddev->reshape_position = MaxSector;
6760 mddev->delta_disks = 0;
2c810cdd 6761 mddev->reshape_backwards = 0;
cea9c228
N		 6762 }
6763}
6764
fd01b88c 6765static void raid5_quiesce(struct mddev *mddev, int state)
72626685 6766{
d1688a6d 6767 struct r5conf *conf = mddev->private;
72626685
N		 6768
6769 switch(state) {
e464eafd
N		 6770 case 2: /* resume for a suspend */
6771 wake_up(&conf->wait_for_overlap);
6772 break;
6773
72626685 6774 case 1: /* stop all writes */
566c09c5 6775 lock_all_device_hash_locks_irq(conf);
64bd660b
N		 6776 /* '2' tells resync/reshape to pause so that all
6777 * active stripes can drain
6778 */
6779 conf->quiesce = 2;
566c09c5 6780 wait_event_cmd(conf->wait_for_stripe,
46031f9a
RBJ		 6781 atomic_read(&conf->active_stripes) == 0 &&
6782 atomic_read(&conf->active_aligned_reads) == 0,
566c09c5
SL		 6783 unlock_all_device_hash_locks_irq(conf),
6784 lock_all_device_hash_locks_irq(conf));
64bd660b 6785 conf->quiesce = 1;
566c09c5 6786 unlock_all_device_hash_locks_irq(conf);
64bd660b
N		 6787 /* allow reshape to continue */
6788 wake_up(&conf->wait_for_overlap);
72626685
N		 6789 break;
6790
6791 case 0: /* re-enable writes */
566c09c5 6792 lock_all_device_hash_locks_irq(conf);
72626685
N		 6793 conf->quiesce = 0;
6794 wake_up(&conf->wait_for_stripe);
e464eafd 6795 wake_up(&conf->wait_for_overlap);
566c09c5 6796 unlock_all_device_hash_locks_irq(conf);
72626685
N		 6797 break;
6798 }
72626685 6799}
b15c2e57 6800
d562b0c4 6801
fd01b88c 6802static void *raid45_takeover_raid0(struct mddev *mddev, int level)
54071b38 6803{
e373ab10 6804 struct r0conf *raid0_conf = mddev->private;
d76c8420 6805 sector_t sectors;
54071b38 6806
f1b29bca 6807 /* for raid0 takeover only one zone is supported */
e373ab10 6808 if (raid0_conf->nr_strip_zones > 1) {
0c55e022
N		 6809 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6810 mdname(mddev));
f1b29bca
DW		 6811 return ERR_PTR(-EINVAL);
6812 }
6813
e373ab10
N		 6814 sectors = raid0_conf->strip_zone[0].zone_end;
6815 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
3b71bd93 6816 mddev->dev_sectors = sectors;
f1b29bca 6817 mddev->new_level = level;
54071b38
TM		 6818 mddev->new_layout = ALGORITHM_PARITY_N;
6819 mddev->new_chunk_sectors = mddev->chunk_sectors;
6820 mddev->raid_disks += 1;
6821 mddev->delta_disks = 1;
6822 /* make sure it will be not marked as dirty */
6823 mddev->recovery_cp = MaxSector;
6824
6825 return setup_conf(mddev);
6826}
6827
6828
fd01b88c 6829static void *raid5_takeover_raid1(struct mddev *mddev)
d562b0c4
N		 6830{
6831 int chunksect;
6832
6833 if (mddev->raid_disks != 2 ||
6834 mddev->degraded > 1)
6835 return ERR_PTR(-EINVAL);
6836
6837 /* Should check if there are write-behind devices? */
6838
6839 chunksect = 64*2; /* 64K by default */
6840
6841 /* The array must be an exact multiple of chunksize */
6842 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6843 chunksect >>= 1;
6844
6845 if ((chunksect<<9) < STRIPE_SIZE)
6846 /* array size does not allow a suitable chunk size */
6847 return ERR_PTR(-EINVAL);
6848
6849 mddev->new_level = 5;
6850 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
664e7c41 6851 mddev->new_chunk_sectors = chunksect;
d562b0c4
N		 6852
6853 return setup_conf(mddev);
6854}
6855
fd01b88c 6856static void *raid5_takeover_raid6(struct mddev *mddev)
fc9739c6
N		 6857{
6858 int new_layout;
6859
6860 switch (mddev->layout) {
6861 case ALGORITHM_LEFT_ASYMMETRIC_6:
6862 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6863 break;
6864 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6865 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6866 break;
6867 case ALGORITHM_LEFT_SYMMETRIC_6:
6868 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6869 break;
6870 case ALGORITHM_RIGHT_SYMMETRIC_6:
6871 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6872 break;
6873 case ALGORITHM_PARITY_0_6:
6874 new_layout = ALGORITHM_PARITY_0;
6875 break;
6876 case ALGORITHM_PARITY_N:
6877 new_layout = ALGORITHM_PARITY_N;
6878 break;
6879 default:
6880 return ERR_PTR(-EINVAL);
6881 }
6882 mddev->new_level = 5;
6883 mddev->new_layout = new_layout;
6884 mddev->delta_disks = -1;
6885 mddev->raid_disks -= 1;
6886 return setup_conf(mddev);
6887}
6888
d562b0c4 6889
fd01b88c 6890static int raid5_check_reshape(struct mddev *mddev)
b3546035 6891{
88ce4930
N		 6892 /* For a 2-drive array, the layout and chunk size can be changed
6893 * immediately as not restriping is needed.
6894 * For larger arrays we record the new value - after validation
6895 * to be used by a reshape pass.
b3546035 6896 */
d1688a6d 6897 struct r5conf *conf = mddev->private;
597a711b 6898 int new_chunk = mddev->new_chunk_sectors;
b3546035 6899
597a711b 6900 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
b3546035
N		 6901 return -EINVAL;
6902 if (new_chunk > 0) {
0ba459d2 6903 if (!is_power_of_2(new_chunk))
b3546035 6904 return -EINVAL;
597a711b 6905 if (new_chunk < (PAGE_SIZE>>9))
b3546035 6906 return -EINVAL;
597a711b 6907 if (mddev->array_sectors & (new_chunk-1))
b3546035
N		 6908 /* not factor of array size */
6909 return -EINVAL;
6910 }
6911
6912 /* They look valid */
6913
88ce4930 6914 if (mddev->raid_disks == 2) {
597a711b
N		 6915 /* can make the change immediately */
6916 if (mddev->new_layout >= 0) {
6917 conf->algorithm = mddev->new_layout;
6918 mddev->layout = mddev->new_layout;
88ce4930
N		 6919 }
6920 if (new_chunk > 0) {
597a711b
N		 6921 conf->chunk_sectors = new_chunk ;
6922 mddev->chunk_sectors = new_chunk;
88ce4930
N		 6923 }
6924 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6925 md_wakeup_thread(mddev->thread);
b3546035 6926 }
50ac168a 6927 return check_reshape(mddev);
88ce4930
N		 6928}
6929
fd01b88c 6930static int raid6_check_reshape(struct mddev *mddev)
88ce4930 6931{
597a711b 6932 int new_chunk = mddev->new_chunk_sectors;
50ac168a 6933
597a711b 6934 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
88ce4930 6935 return -EINVAL;
b3546035 6936 if (new_chunk > 0) {
0ba459d2 6937 if (!is_power_of_2(new_chunk))
88ce4930 6938 return -EINVAL;
597a711b 6939 if (new_chunk < (PAGE_SIZE >> 9))
88ce4930 6940 return -EINVAL;
597a711b 6941 if (mddev->array_sectors & (new_chunk-1))
88ce4930
N		 6942 /* not factor of array size */
6943 return -EINVAL;
b3546035 6944 }
88ce4930
N		 6945
6946 /* They look valid */
50ac168a 6947 return check_reshape(mddev);
b3546035
N		 6948}
6949
fd01b88c 6950static void *raid5_takeover(struct mddev *mddev)
d562b0c4
N		 6951{
6952 /* raid5 can take over:
f1b29bca 6953 * raid0 - if there is only one strip zone - make it a raid4 layout
d562b0c4
N		 6954 * raid1 - if there are two drives. We need to know the chunk size
6955 * raid4 - trivial - just use a raid4 layout.
6956 * raid6 - Providing it is a *_6 layout
d562b0c4 6957 */
f1b29bca
DW		 6958 if (mddev->level == 0)
6959 return raid45_takeover_raid0(mddev, 5);
d562b0c4
N		 6960 if (mddev->level == 1)
6961 return raid5_takeover_raid1(mddev);
e9d4758f
N		 6962 if (mddev->level == 4) {
6963 mddev->new_layout = ALGORITHM_PARITY_N;
6964 mddev->new_level = 5;
6965 return setup_conf(mddev);
6966 }
fc9739c6
N		 6967 if (mddev->level == 6)
6968 return raid5_takeover_raid6(mddev);
d562b0c4
N		 6969
6970 return ERR_PTR(-EINVAL);
6971}
6972
fd01b88c 6973static void *raid4_takeover(struct mddev *mddev)
a78d38a1 6974{
f1b29bca
DW		 6975 /* raid4 can take over:
6976 * raid0 - if there is only one strip zone
6977 * raid5 - if layout is right
a78d38a1 6978 */
f1b29bca
DW		 6979 if (mddev->level == 0)
6980 return raid45_takeover_raid0(mddev, 4);
a78d38a1
N		 6981 if (mddev->level == 5 &&
6982 mddev->layout == ALGORITHM_PARITY_N) {
6983 mddev->new_layout = 0;
6984 mddev->new_level = 4;
6985 return setup_conf(mddev);
6986 }
6987 return ERR_PTR(-EINVAL);
6988}
d562b0c4 6989
84fc4b56 6990static struct md_personality raid5_personality;
245f46c2 6991
fd01b88c 6992static void *raid6_takeover(struct mddev *mddev)
245f46c2
N		 6993{
6994 /* Currently can only take over a raid5. We map the
6995 * personality to an equivalent raid6 personality
6996 * with the Q block at the end.
6997 */
6998 int new_layout;
6999
7000 if (mddev->pers != &raid5_personality)
7001 return ERR_PTR(-EINVAL);
7002 if (mddev->degraded > 1)
7003 return ERR_PTR(-EINVAL);
7004 if (mddev->raid_disks > 253)
7005 return ERR_PTR(-EINVAL);
7006 if (mddev->raid_disks < 3)
7007 return ERR_PTR(-EINVAL);
7008
7009 switch (mddev->layout) {
7010 case ALGORITHM_LEFT_ASYMMETRIC:
7011 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7012 break;
7013 case ALGORITHM_RIGHT_ASYMMETRIC:
7014 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7015 break;
7016 case ALGORITHM_LEFT_SYMMETRIC:
7017 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7018 break;
7019 case ALGORITHM_RIGHT_SYMMETRIC:
7020 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7021 break;
7022 case ALGORITHM_PARITY_0:
7023 new_layout = ALGORITHM_PARITY_0_6;
7024 break;
7025 case ALGORITHM_PARITY_N:
7026 new_layout = ALGORITHM_PARITY_N;
7027 break;
7028 default:
7029 return ERR_PTR(-EINVAL);
7030 }
7031 mddev->new_level = 6;
7032 mddev->new_layout = new_layout;
7033 mddev->delta_disks = 1;
7034 mddev->raid_disks += 1;
7035 return setup_conf(mddev);
7036}
7037
7038
84fc4b56 7039static struct md_personality raid6_personality =
16a53ecc
N		 7040{
7041 .name = "raid6",
7042 .level = 6,
7043 .owner = THIS_MODULE,
7044 .make_request = make_request,
7045 .run = run,
7046 .stop = stop,
7047 .status = status,
7048 .error_handler = error,
7049 .hot_add_disk = raid5_add_disk,
7050 .hot_remove_disk= raid5_remove_disk,
7051 .spare_active = raid5_spare_active,
7052 .sync_request = sync_request,
7053 .resize = raid5_resize,
80c3a6ce 7054 .size = raid5_size,
50ac168a 7055 .check_reshape = raid6_check_reshape,
f416885e 7056 .start_reshape = raid5_start_reshape,
cea9c228 7057 .finish_reshape = raid5_finish_reshape,
16a53ecc 7058 .quiesce = raid5_quiesce,
245f46c2 7059 .takeover = raid6_takeover,
16a53ecc 7060};
84fc4b56 7061static struct md_personality raid5_personality =
1da177e4
LT		 7062{
7063 .name = "raid5",
2604b703 7064 .level = 5,
1da177e4
LT		 7065 .owner = THIS_MODULE,
7066 .make_request = make_request,
7067 .run = run,
7068 .stop = stop,
7069 .status = status,
7070 .error_handler = error,
7071 .hot_add_disk = raid5_add_disk,
7072 .hot_remove_disk= raid5_remove_disk,
7073 .spare_active = raid5_spare_active,
7074 .sync_request = sync_request,
7075 .resize = raid5_resize,
80c3a6ce 7076 .size = raid5_size,
63c70c4f
N		 7077 .check_reshape = raid5_check_reshape,
7078 .start_reshape = raid5_start_reshape,
cea9c228 7079 .finish_reshape = raid5_finish_reshape,
72626685 7080 .quiesce = raid5_quiesce,
d562b0c4 7081 .takeover = raid5_takeover,
1da177e4
LT		 7082};
7083
84fc4b56 7084static struct md_personality raid4_personality =
1da177e4 7085{
2604b703
N		 7086 .name = "raid4",
7087 .level = 4,
7088 .owner = THIS_MODULE,
7089 .make_request = make_request,
7090 .run = run,
7091 .stop = stop,
7092 .status = status,
7093 .error_handler = error,
7094 .hot_add_disk = raid5_add_disk,
7095 .hot_remove_disk= raid5_remove_disk,
7096 .spare_active = raid5_spare_active,
7097 .sync_request = sync_request,
7098 .resize = raid5_resize,
80c3a6ce 7099 .size = raid5_size,
3d37890b
N		 7100 .check_reshape = raid5_check_reshape,
7101 .start_reshape = raid5_start_reshape,
cea9c228 7102 .finish_reshape = raid5_finish_reshape,
2604b703 7103 .quiesce = raid5_quiesce,
a78d38a1 7104 .takeover = raid4_takeover,
2604b703
N		 7105};
7106
7107static int __init raid5_init(void)
7108{
851c30c9
SL		 7109 raid5_wq = alloc_workqueue("raid5wq",
7110 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7111 if (!raid5_wq)
7112 return -ENOMEM;
16a53ecc 7113 register_md_personality(&raid6_personality);
2604b703
N		 7114 register_md_personality(&raid5_personality);
7115 register_md_personality(&raid4_personality);
7116 return 0;
1da177e4
LT		 7117}
7118
2604b703 7119static void raid5_exit(void)
1da177e4 7120{
16a53ecc 7121 unregister_md_personality(&raid6_personality);
2604b703
N		 7122 unregister_md_personality(&raid5_personality);
7123 unregister_md_personality(&raid4_personality);
851c30c9 7124 destroy_workqueue(raid5_wq);
1da177e4
LT		 7125}
7126
7127module_init(raid5_init);
7128module_exit(raid5_exit);
7129MODULE_LICENSE("GPL");
0efb9e61 7130MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
1da177e4 7131MODULE_ALIAS("md-personality-4"); /* RAID5 */
d9d166c2
N		 7132MODULE_ALIAS("md-raid5");
7133MODULE_ALIAS("md-raid4");
2604b703
N		 7134MODULE_ALIAS("md-level-5");
7135MODULE_ALIAS("md-level-4");
16a53ecc
N		 7136MODULE_ALIAS("md-personality-8"); /* RAID6 */
7137MODULE_ALIAS("md-raid6");
7138MODULE_ALIAS("md-level-6");
7139
7140/* This used to be two separate modules, they were: */
7141MODULE_ALIAS("raid5");
7142MODULE_ALIAS("raid6");
Linux 6.x block layer and io_uring tree(s)