projects / linux-2.6-block.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
raid5: sysfs entry to control worker thread number
[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
LT		 79#define HASH_MASK (NR_HASH - 1)
80
d1688a6d 81static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
db298e19
N		 82{
83 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
84 return &conf->stripe_hashtbl[hash];
85}
1da177e4
LT		 86
87/* bio's attached to a stripe+device for I/O are linked together in bi_sector
88 * order without overlap. There may be several bio's per stripe+device, and
89 * a bio could span several devices.
90 * When walking this list for a particular stripe+device, we must never proceed
91 * beyond a bio that extends past this device, as the next bio might no longer
92 * be valid.
db298e19 93 * This function is used to determine the 'next' bio in the list, given the sector
1da177e4
LT		 94 * of the current stripe+device
95 */
db298e19
N		 96static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
97{
aa8b57aa 98 int sectors = bio_sectors(bio);
db298e19
N		 99 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
100 return bio->bi_next;
101 else
102 return NULL;
103}
1da177e4 104
960e739d 105/*
5b99c2ff
JA		 106 * We maintain a biased count of active stripes in the bottom 16 bits of
107 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
960e739d 108 */
e7836bd6 109static inline int raid5_bi_processed_stripes(struct bio *bio)
960e739d 110{
e7836bd6
SL		 111 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
112 return (atomic_read(segments) >> 16) & 0xffff;
960e739d
JA		 113}
114
e7836bd6 115static inline int raid5_dec_bi_active_stripes(struct bio *bio)
960e739d 116{
e7836bd6
SL		 117 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
118 return atomic_sub_return(1, segments) & 0xffff;
960e739d
JA		 119}
120
e7836bd6 121static inline void raid5_inc_bi_active_stripes(struct bio *bio)
960e739d 122{
e7836bd6
SL		 123 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
124 atomic_inc(segments);
960e739d
JA		 125}
126
e7836bd6
SL		 127static inline void raid5_set_bi_processed_stripes(struct bio *bio,
128 unsigned int cnt)
960e739d 129{
e7836bd6
SL		 130 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
131 int old, new;
960e739d 132
e7836bd6
SL		 133 do {
134 old = atomic_read(segments);
135 new = (old & 0xffff) | (cnt << 16);
136 } while (atomic_cmpxchg(segments, old, new) != old);
960e739d
JA		 137}
138
e7836bd6 139static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
960e739d 140{
e7836bd6
SL		 141 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
142 atomic_set(segments, cnt);
960e739d
JA		 143}
144
d0dabf7e
N		 145/* Find first data disk in a raid6 stripe */
146static inline int raid6_d0(struct stripe_head *sh)
147{
67cc2b81
N		 148 if (sh->ddf_layout)
149 /* ddf always start from first device */
150 return 0;
151 /* md starts just after Q block */
d0dabf7e
N		 152 if (sh->qd_idx == sh->disks - 1)
153 return 0;
154 else
155 return sh->qd_idx + 1;
156}
16a53ecc
N		 157static inline int raid6_next_disk(int disk, int raid_disks)
158{
159 disk++;
160 return (disk < raid_disks) ? disk : 0;
161}
a4456856 162
d0dabf7e
N		 163/* When walking through the disks in a raid5, starting at raid6_d0,
164 * We need to map each disk to a 'slot', where the data disks are slot
165 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
166 * is raid_disks-1. This help does that mapping.
167 */
67cc2b81
N		 168static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
169 int *count, int syndrome_disks)
d0dabf7e 170{
6629542e 171 int slot = *count;
67cc2b81 172
e4424fee 173 if (sh->ddf_layout)
6629542e 174 (*count)++;
d0dabf7e 175 if (idx == sh->pd_idx)
67cc2b81 176 return syndrome_disks;
d0dabf7e 177 if (idx == sh->qd_idx)
67cc2b81 178 return syndrome_disks + 1;
e4424fee 179 if (!sh->ddf_layout)
6629542e 180 (*count)++;
d0dabf7e
N		 181 return slot;
182}
183
a4456856
DW		 184static void return_io(struct bio *return_bi)
185{
186 struct bio *bi = return_bi;
187 while (bi) {
a4456856
DW		 188
189 return_bi = bi->bi_next;
190 bi->bi_next = NULL;
191 bi->bi_size = 0;
0a82a8d1
LT		 192 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
193 bi, 0);
0e13fe23 194 bio_endio(bi, 0);
a4456856
DW		 195 bi = return_bi;
196 }
197}
198
d1688a6d 199static void print_raid5_conf (struct r5conf *conf);
1da177e4 200
600aa109
DW		 201static int stripe_operations_active(struct stripe_head *sh)
202{
203 return sh->check_state || sh->reconstruct_state ||
204 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
205 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
206}
207
851c30c9
SL		 208static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
209{
210 struct r5conf *conf = sh->raid_conf;
211 struct r5worker_group *group;
212 int i, cpu = sh->cpu;
213
214 if (!cpu_online(cpu)) {
215 cpu = cpumask_any(cpu_online_mask);
216 sh->cpu = cpu;
217 }
218
219 if (list_empty(&sh->lru)) {
220 struct r5worker_group *group;
221 group = conf->worker_groups + cpu_to_group(cpu);
222 list_add_tail(&sh->lru, &group->handle_list);
223 }
224
225 if (conf->worker_cnt_per_group == 0) {
226 md_wakeup_thread(conf->mddev->thread);
227 return;
228 }
229
230 group = conf->worker_groups + cpu_to_group(sh->cpu);
231
232 for (i = 0; i < conf->worker_cnt_per_group; i++)
233 queue_work_on(sh->cpu, raid5_wq, &group->workers[i].work);
234}
235
4eb788df 236static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
1da177e4 237{
4eb788df
SL		 238 BUG_ON(!list_empty(&sh->lru));
239 BUG_ON(atomic_read(&conf->active_stripes)==0);
240 if (test_bit(STRIPE_HANDLE, &sh->state)) {
241 if (test_bit(STRIPE_DELAYED, &sh->state) &&
242 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
243 list_add_tail(&sh->lru, &conf->delayed_list);
244 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
245 sh->bm_seq - conf->seq_write > 0)
246 list_add_tail(&sh->lru, &conf->bitmap_list);
247 else {
248 clear_bit(STRIPE_DELAYED, &sh->state);
249 clear_bit(STRIPE_BIT_DELAY, &sh->state);
851c30c9
SL		 250 if (conf->worker_cnt_per_group == 0) {
251 list_add_tail(&sh->lru, &conf->handle_list);
252 } else {
253 raid5_wakeup_stripe_thread(sh);
254 return;
255 }
4eb788df
SL		 256 }
257 md_wakeup_thread(conf->mddev->thread);
258 } else {
259 BUG_ON(stripe_operations_active(sh));
260 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
261 if (atomic_dec_return(&conf->preread_active_stripes)
262 < IO_THRESHOLD)
263 md_wakeup_thread(conf->mddev->thread);
264 atomic_dec(&conf->active_stripes);
265 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
266 list_add_tail(&sh->lru, &conf->inactive_list);
267 wake_up(&conf->wait_for_stripe);
268 if (conf->retry_read_aligned)
269 md_wakeup_thread(conf->mddev->thread);
1da177e4
LT		 270 }
271 }
272}
d0dabf7e 273
4eb788df
SL		 274static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
275{
276 if (atomic_dec_and_test(&sh->count))
277 do_release_stripe(conf, sh);
278}
279
d265d9dc
SL		 280static struct llist_node *llist_reverse_order(struct llist_node *head)
281{
282 struct llist_node *new_head = NULL;
283
284 while (head) {
285 struct llist_node *tmp = head;
286 head = head->next;
287 tmp->next = new_head;
288 new_head = tmp;
289 }
290
291 return new_head;
292}
293
773ca82f
SL		 294/* should hold conf->device_lock already */
295static int release_stripe_list(struct r5conf *conf)
296{
297 struct stripe_head *sh;
298 int count = 0;
299 struct llist_node *head;
300
301 head = llist_del_all(&conf->released_stripes);
d265d9dc 302 head = llist_reverse_order(head);
773ca82f
SL		 303 while (head) {
304 sh = llist_entry(head, struct stripe_head, release_list);
305 head = llist_next(head);
306 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
307 smp_mb();
308 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
309 /*
310 * Don't worry the bit is set here, because if the bit is set
311 * again, the count is always > 1. This is true for
312 * STRIPE_ON_UNPLUG_LIST bit too.
313 */
314 __release_stripe(conf, sh);
315 count++;
316 }
317
318 return count;
319}
320
1da177e4
LT		 321static void release_stripe(struct stripe_head *sh)
322{
d1688a6d 323 struct r5conf *conf = sh->raid_conf;
1da177e4 324 unsigned long flags;
773ca82f 325 bool wakeup;
16a53ecc 326
773ca82f
SL		 327 if (test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
328 goto slow_path;
329 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
330 if (wakeup)
331 md_wakeup_thread(conf->mddev->thread);
332 return;
333slow_path:
4eb788df 334 local_irq_save(flags);
773ca82f 335 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
4eb788df
SL		 336 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
337 do_release_stripe(conf, sh);
338 spin_unlock(&conf->device_lock);
339 }
340 local_irq_restore(flags);
1da177e4
LT		 341}
342
fccddba0 343static inline void remove_hash(struct stripe_head *sh)
1da177e4 344{
45b4233c
DW		 345 pr_debug("remove_hash(), stripe %llu\n",
346 (unsigned long long)sh->sector);
1da177e4 347
fccddba0 348 hlist_del_init(&sh->hash);
1da177e4
LT		 349}
350
d1688a6d 351static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
1da177e4 352{
fccddba0 353 struct hlist_head *hp = stripe_hash(conf, sh->sector);
1da177e4 354
45b4233c
DW		 355 pr_debug("insert_hash(), stripe %llu\n",
356 (unsigned long long)sh->sector);
1da177e4 357
fccddba0 358 hlist_add_head(&sh->hash, hp);
1da177e4
LT		 359}
360
361
362/* find an idle stripe, make sure it is unhashed, and return it. */
d1688a6d 363static struct stripe_head *get_free_stripe(struct r5conf *conf)
1da177e4
LT		 364{
365 struct stripe_head *sh = NULL;
366 struct list_head *first;
367
1da177e4
LT		 368 if (list_empty(&conf->inactive_list))
369 goto out;
370 first = conf->inactive_list.next;
371 sh = list_entry(first, struct stripe_head, lru);
372 list_del_init(first);
373 remove_hash(sh);
374 atomic_inc(&conf->active_stripes);
375out:
376 return sh;
377}
378
e4e11e38 379static void shrink_buffers(struct stripe_head *sh)
1da177e4
LT		 380{
381 struct page *p;
382 int i;
e4e11e38 383 int num = sh->raid_conf->pool_size;
1da177e4 384
e4e11e38 385 for (i = 0; i < num ; i++) {
1da177e4
LT		 386 p = sh->dev[i].page;
387 if (!p)
388 continue;
389 sh->dev[i].page = NULL;
2d1f3b5d 390 put_page(p);
1da177e4
LT		 391 }
392}
393
e4e11e38 394static int grow_buffers(struct stripe_head *sh)
1da177e4
LT		 395{
396 int i;
e4e11e38 397 int num = sh->raid_conf->pool_size;
1da177e4 398
e4e11e38 399 for (i = 0; i < num; i++) {
1da177e4
LT		 400 struct page *page;
401
402 if (!(page = alloc_page(GFP_KERNEL))) {
403 return 1;
404 }
405 sh->dev[i].page = page;
406 }
407 return 0;
408}
409
784052ec 410static void raid5_build_block(struct stripe_head *sh, int i, int previous);
d1688a6d 411static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
911d4ee8 412 struct stripe_head *sh);
1da177e4 413
b5663ba4 414static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
1da177e4 415{
d1688a6d 416 struct r5conf *conf = sh->raid_conf;
7ecaa1e6 417 int i;
1da177e4 418
78bafebd
ES		 419 BUG_ON(atomic_read(&sh->count) != 0);
420 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
600aa109 421 BUG_ON(stripe_operations_active(sh));
d84e0f10 422
45b4233c 423 pr_debug("init_stripe called, stripe %llu\n",
1da177e4
LT		 424 (unsigned long long)sh->sector);
425
426 remove_hash(sh);
16a53ecc 427
86b42c71 428 sh->generation = conf->generation - previous;
b5663ba4 429 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
1da177e4 430 sh->sector = sector;
911d4ee8 431 stripe_set_idx(sector, conf, previous, sh);
1da177e4
LT		 432 sh->state = 0;
433
7ecaa1e6
N		 434
435 for (i = sh->disks; i--; ) {
1da177e4
LT		 436 struct r5dev *dev = &sh->dev[i];
437
d84e0f10 438 if (dev->toread || dev->read || dev->towrite || dev->written ||
1da177e4 439 test_bit(R5_LOCKED, &dev->flags)) {
d84e0f10 440 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
1da177e4 441 (unsigned long long)sh->sector, i, dev->toread,
d84e0f10 442 dev->read, dev->towrite, dev->written,
1da177e4 443 test_bit(R5_LOCKED, &dev->flags));
8cfa7b0f 444 WARN_ON(1);
1da177e4
LT		 445 }
446 dev->flags = 0;
784052ec 447 raid5_build_block(sh, i, previous);
1da177e4
LT		 448 }
449 insert_hash(conf, sh);
851c30c9 450 sh->cpu = smp_processor_id();
1da177e4
LT		 451}
452
d1688a6d 453static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
86b42c71 454 short generation)
1da177e4
LT		 455{
456 struct stripe_head *sh;
457
45b4233c 458 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
b67bfe0d 459 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
86b42c71 460 if (sh->sector == sector && sh->generation == generation)
1da177e4 461 return sh;
45b4233c 462 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
1da177e4
LT		 463 return NULL;
464}
465
674806d6
N		 466/*
467 * Need to check if array has failed when deciding whether to:
468 * - start an array
469 * - remove non-faulty devices
470 * - add a spare
471 * - allow a reshape
472 * This determination is simple when no reshape is happening.
473 * However if there is a reshape, we need to carefully check
474 * both the before and after sections.
475 * This is because some failed devices may only affect one
476 * of the two sections, and some non-in_sync devices may
477 * be insync in the section most affected by failed devices.
478 */
908f4fbd 479static int calc_degraded(struct r5conf *conf)
674806d6 480{
908f4fbd 481 int degraded, degraded2;
674806d6 482 int i;
674806d6
N		 483
484 rcu_read_lock();
485 degraded = 0;
486 for (i = 0; i < conf->previous_raid_disks; i++) {
3cb03002 487 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
e5c86471
N		 488 if (rdev && test_bit(Faulty, &rdev->flags))
489 rdev = rcu_dereference(conf->disks[i].replacement);
674806d6
N		 490 if (!rdev || test_bit(Faulty, &rdev->flags))
491 degraded++;
492 else if (test_bit(In_sync, &rdev->flags))
493 ;
494 else
495 /* not in-sync or faulty.
496 * If the reshape increases the number of devices,
497 * this is being recovered by the reshape, so
498 * this 'previous' section is not in_sync.
499 * If the number of devices is being reduced however,
500 * the device can only be part of the array if
501 * we are reverting a reshape, so this section will
502 * be in-sync.
503 */
504 if (conf->raid_disks >= conf->previous_raid_disks)
505 degraded++;
506 }
507 rcu_read_unlock();
908f4fbd
N		 508 if (conf->raid_disks == conf->previous_raid_disks)
509 return degraded;
674806d6 510 rcu_read_lock();
908f4fbd 511 degraded2 = 0;
674806d6 512 for (i = 0; i < conf->raid_disks; i++) {
3cb03002 513 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
e5c86471
N		 514 if (rdev && test_bit(Faulty, &rdev->flags))
515 rdev = rcu_dereference(conf->disks[i].replacement);
674806d6 516 if (!rdev || test_bit(Faulty, &rdev->flags))
908f4fbd 517 degraded2++;
674806d6
N		 518 else if (test_bit(In_sync, &rdev->flags))
519 ;
520 else
521 /* not in-sync or faulty.
522 * If reshape increases the number of devices, this
523 * section has already been recovered, else it
524 * almost certainly hasn't.
525 */
526 if (conf->raid_disks <= conf->previous_raid_disks)
908f4fbd 527 degraded2++;
674806d6
N		 528 }
529 rcu_read_unlock();
908f4fbd
N		 530 if (degraded2 > degraded)
531 return degraded2;
532 return degraded;
533}
534
535static int has_failed(struct r5conf *conf)
536{
537 int degraded;
538
539 if (conf->mddev->reshape_position == MaxSector)
540 return conf->mddev->degraded > conf->max_degraded;
541
542 degraded = calc_degraded(conf);
674806d6
N		 543 if (degraded > conf->max_degraded)
544 return 1;
545 return 0;
546}
547
b5663ba4 548static struct stripe_head *
d1688a6d 549get_active_stripe(struct r5conf *conf, sector_t sector,
a8c906ca 550 int previous, int noblock, int noquiesce)
1da177e4
LT		 551{
552 struct stripe_head *sh;
553
45b4233c 554 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
1da177e4
LT		 555
556 spin_lock_irq(&conf->device_lock);
557
558 do {
72626685 559 wait_event_lock_irq(conf->wait_for_stripe,
a8c906ca 560 conf->quiesce == 0 || noquiesce,
eed8c02e 561 conf->device_lock);
86b42c71 562 sh = __find_stripe(conf, sector, conf->generation - previous);
1da177e4
LT		 563 if (!sh) {
564 if (!conf->inactive_blocked)
565 sh = get_free_stripe(conf);
566 if (noblock && sh == NULL)
567 break;
568 if (!sh) {
569 conf->inactive_blocked = 1;
570 wait_event_lock_irq(conf->wait_for_stripe,
571 !list_empty(&conf->inactive_list) &&
5036805b
N		 572 (atomic_read(&conf->active_stripes)
573 < (conf->max_nr_stripes *3/4)
1da177e4 574 || !conf->inactive_blocked),
eed8c02e 575 conf->device_lock);
1da177e4
LT		 576 conf->inactive_blocked = 0;
577 } else
b5663ba4 578 init_stripe(sh, sector, previous);
1da177e4
LT		 579 } else {
580 if (atomic_read(&sh->count)) {
ab69ae12 581 BUG_ON(!list_empty(&sh->lru)
8811b596 582 && !test_bit(STRIPE_EXPANDING, &sh->state)
773ca82f
SL		 583 && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state)
584 && !test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1da177e4
LT		 585 } else {
586 if (!test_bit(STRIPE_HANDLE, &sh->state))
587 atomic_inc(&conf->active_stripes);
ff4e8d9a
N		 588 if (list_empty(&sh->lru) &&
589 !test_bit(STRIPE_EXPANDING, &sh->state))
16a53ecc
N		 590 BUG();
591 list_del_init(&sh->lru);
1da177e4
LT		 592 }
593 }
594 } while (sh == NULL);
595
596 if (sh)
597 atomic_inc(&sh->count);
598
599 spin_unlock_irq(&conf->device_lock);
600 return sh;
601}
602
05616be5
N		 603/* Determine if 'data_offset' or 'new_data_offset' should be used
604 * in this stripe_head.
605 */
606static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
607{
608 sector_t progress = conf->reshape_progress;
609 /* Need a memory barrier to make sure we see the value
610 * of conf->generation, or ->data_offset that was set before
611 * reshape_progress was updated.
612 */
613 smp_rmb();
614 if (progress == MaxSector)
615 return 0;
616 if (sh->generation == conf->generation - 1)
617 return 0;
618 /* We are in a reshape, and this is a new-generation stripe,
619 * so use new_data_offset.
620 */
621 return 1;
622}
623
6712ecf8
N		 624static void
625raid5_end_read_request(struct bio *bi, int error);
626static void
627raid5_end_write_request(struct bio *bi, int error);
91c00924 628
c4e5ac0a 629static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
91c00924 630{
d1688a6d 631 struct r5conf *conf = sh->raid_conf;
91c00924
DW		 632 int i, disks = sh->disks;
633
634 might_sleep();
635
636 for (i = disks; i--; ) {
637 int rw;
9a3e1101 638 int replace_only = 0;
977df362
N		 639 struct bio *bi, *rbi;
640 struct md_rdev *rdev, *rrdev = NULL;
e9c7469b
TH		 641 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
642 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
643 rw = WRITE_FUA;
644 else
645 rw = WRITE;
9e444768 646 if (test_bit(R5_Discard, &sh->dev[i].flags))
620125f2 647 rw |= REQ_DISCARD;
e9c7469b 648 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
91c00924 649 rw = READ;
9a3e1101
N		 650 else if (test_and_clear_bit(R5_WantReplace,
651 &sh->dev[i].flags)) {
652 rw = WRITE;
653 replace_only = 1;
654 } else
91c00924 655 continue;
bc0934f0
SL		 656 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
657 rw |= REQ_SYNC;
91c00924
DW		 658
659 bi = &sh->dev[i].req;
977df362 660 rbi = &sh->dev[i].rreq; /* For writing to replacement */
91c00924 661
91c00924 662 rcu_read_lock();
9a3e1101 663 rrdev = rcu_dereference(conf->disks[i].replacement);
dd054fce
N		 664 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
665 rdev = rcu_dereference(conf->disks[i].rdev);
666 if (!rdev) {
667 rdev = rrdev;
668 rrdev = NULL;
669 }
9a3e1101
N		 670 if (rw & WRITE) {
671 if (replace_only)
672 rdev = NULL;
dd054fce
N		 673 if (rdev == rrdev)
674 /* We raced and saw duplicates */
675 rrdev = NULL;
9a3e1101 676 } else {
dd054fce 677 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
9a3e1101
N		 678 rdev = rrdev;
679 rrdev = NULL;
680 }
977df362 681
91c00924
DW		 682 if (rdev && test_bit(Faulty, &rdev->flags))
683 rdev = NULL;
684 if (rdev)
685 atomic_inc(&rdev->nr_pending);
977df362
N		 686 if (rrdev && test_bit(Faulty, &rrdev->flags))
687 rrdev = NULL;
688 if (rrdev)
689 atomic_inc(&rrdev->nr_pending);
91c00924
DW		 690 rcu_read_unlock();
691
73e92e51 692 /* We have already checked bad blocks for reads. Now
977df362
N		 693 * need to check for writes. We never accept write errors
694 * on the replacement, so we don't to check rrdev.
73e92e51
N		 695 */
696 while ((rw & WRITE) && rdev &&
697 test_bit(WriteErrorSeen, &rdev->flags)) {
698 sector_t first_bad;
699 int bad_sectors;
700 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
701 &first_bad, &bad_sectors);
702 if (!bad)
703 break;
704
705 if (bad < 0) {
706 set_bit(BlockedBadBlocks, &rdev->flags);
707 if (!conf->mddev->external &&
708 conf->mddev->flags) {
709 /* It is very unlikely, but we might
710 * still need to write out the
711 * bad block log - better give it
712 * a chance*/
713 md_check_recovery(conf->mddev);
714 }
1850753d 715 /*
716 * Because md_wait_for_blocked_rdev
717 * will dec nr_pending, we must
718 * increment it first.
719 */
720 atomic_inc(&rdev->nr_pending);
73e92e51
N		 721 md_wait_for_blocked_rdev(rdev, conf->mddev);
722 } else {
723 /* Acknowledged bad block - skip the write */
724 rdev_dec_pending(rdev, conf->mddev);
725 rdev = NULL;
726 }
727 }
728
91c00924 729 if (rdev) {
9a3e1101
N		 730 if (s->syncing || s->expanding || s->expanded
731 || s->replacing)
91c00924
DW		 732 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
733
2b7497f0
DW		 734 set_bit(STRIPE_IO_STARTED, &sh->state);
735
2f6db2a7 736 bio_reset(bi);
91c00924 737 bi->bi_bdev = rdev->bdev;
2f6db2a7
KO		 738 bi->bi_rw = rw;
739 bi->bi_end_io = (rw & WRITE)
740 ? raid5_end_write_request
741 : raid5_end_read_request;
742 bi->bi_private = sh;
743
91c00924 744 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
e46b272b 745 __func__, (unsigned long long)sh->sector,
91c00924
DW		 746 bi->bi_rw, i);
747 atomic_inc(&sh->count);
05616be5
N		 748 if (use_new_offset(conf, sh))
749 bi->bi_sector = (sh->sector
750 + rdev->new_data_offset);
751 else
752 bi->bi_sector = (sh->sector
753 + rdev->data_offset);
3f9e7c14 754 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
755 bi->bi_rw |= REQ_FLUSH;
756
4997b72e 757 bi->bi_vcnt = 1;
91c00924
DW		 758 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
759 bi->bi_io_vec[0].bv_offset = 0;
760 bi->bi_size = STRIPE_SIZE;
977df362
N		 761 if (rrdev)
762 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
e3620a3a
JB		 763
764 if (conf->mddev->gendisk)
765 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
766 bi, disk_devt(conf->mddev->gendisk),
767 sh->dev[i].sector);
91c00924 768 generic_make_request(bi);
977df362
N		 769 }
770 if (rrdev) {
9a3e1101
N		 771 if (s->syncing || s->expanding || s->expanded
772 || s->replacing)
977df362
N		 773 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
774
775 set_bit(STRIPE_IO_STARTED, &sh->state);
776
2f6db2a7 777 bio_reset(rbi);
977df362 778 rbi->bi_bdev = rrdev->bdev;
2f6db2a7
KO		 779 rbi->bi_rw = rw;
780 BUG_ON(!(rw & WRITE));
781 rbi->bi_end_io = raid5_end_write_request;
782 rbi->bi_private = sh;
783
977df362
N		 784 pr_debug("%s: for %llu schedule op %ld on "
785 "replacement disc %d\n",
786 __func__, (unsigned long long)sh->sector,
787 rbi->bi_rw, i);
788 atomic_inc(&sh->count);
05616be5
N		 789 if (use_new_offset(conf, sh))
790 rbi->bi_sector = (sh->sector
791 + rrdev->new_data_offset);
792 else
793 rbi->bi_sector = (sh->sector
794 + rrdev->data_offset);
4997b72e 795 rbi->bi_vcnt = 1;
977df362
N		 796 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
797 rbi->bi_io_vec[0].bv_offset = 0;
798 rbi->bi_size = STRIPE_SIZE;
e3620a3a
JB		 799 if (conf->mddev->gendisk)
800 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
801 rbi, disk_devt(conf->mddev->gendisk),
802 sh->dev[i].sector);
977df362
N		 803 generic_make_request(rbi);
804 }
805 if (!rdev && !rrdev) {
b062962e 806 if (rw & WRITE)
91c00924
DW		 807 set_bit(STRIPE_DEGRADED, &sh->state);
808 pr_debug("skip op %ld on disc %d for sector %llu\n",
809 bi->bi_rw, i, (unsigned long long)sh->sector);
810 clear_bit(R5_LOCKED, &sh->dev[i].flags);
811 set_bit(STRIPE_HANDLE, &sh->state);
812 }
813 }
814}
815
816static struct dma_async_tx_descriptor *
817async_copy_data(int frombio, struct bio *bio, struct page *page,
818 sector_t sector, struct dma_async_tx_descriptor *tx)
819{
820 struct bio_vec *bvl;
821 struct page *bio_page;
822 int i;
823 int page_offset;
a08abd8c 824 struct async_submit_ctl submit;
0403e382 825 enum async_tx_flags flags = 0;
91c00924
DW		 826
827 if (bio->bi_sector >= sector)
828 page_offset = (signed)(bio->bi_sector - sector) * 512;
829 else
830 page_offset = (signed)(sector - bio->bi_sector) * -512;
a08abd8c 831
0403e382
DW		 832 if (frombio)
833 flags |= ASYNC_TX_FENCE;
834 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
835
91c00924 836 bio_for_each_segment(bvl, bio, i) {
fcde9075 837 int len = bvl->bv_len;
91c00924
DW		 838 int clen;
839 int b_offset = 0;
840
841 if (page_offset < 0) {
842 b_offset = -page_offset;
843 page_offset += b_offset;
844 len -= b_offset;
845 }
846
847 if (len > 0 && page_offset + len > STRIPE_SIZE)
848 clen = STRIPE_SIZE - page_offset;
849 else
850 clen = len;
851
852 if (clen > 0) {
fcde9075
NK		 853 b_offset += bvl->bv_offset;
854 bio_page = bvl->bv_page;
91c00924
DW		 855 if (frombio)
856 tx = async_memcpy(page, bio_page, page_offset,
a08abd8c 857 b_offset, clen, &submit);
91c00924
DW		 858 else
859 tx = async_memcpy(bio_page, page, b_offset,
a08abd8c 860 page_offset, clen, &submit);
91c00924 861 }
a08abd8c
DW		 862 /* chain the operations */
863 submit.depend_tx = tx;
864
91c00924
DW		 865 if (clen < len) /* hit end of page */
866 break;
867 page_offset += len;
868 }
869
870 return tx;
871}
872
873static void ops_complete_biofill(void *stripe_head_ref)
874{
875 struct stripe_head *sh = stripe_head_ref;
876 struct bio *return_bi = NULL;
e4d84909 877 int i;
91c00924 878
e46b272b 879 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 880 (unsigned long long)sh->sector);
881
882 /* clear completed biofills */
883 for (i = sh->disks; i--; ) {
884 struct r5dev *dev = &sh->dev[i];
91c00924
DW		 885
886 /* acknowledge completion of a biofill operation */
e4d84909
DW		 887 /* and check if we need to reply to a read request,
888 * new R5_Wantfill requests are held off until
83de75cc 889 * !STRIPE_BIOFILL_RUN
e4d84909
DW		 890 */
891 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
91c00924 892 struct bio *rbi, *rbi2;
91c00924 893
91c00924
DW		 894 BUG_ON(!dev->read);
895 rbi = dev->read;
896 dev->read = NULL;
897 while (rbi && rbi->bi_sector <
898 dev->sector + STRIPE_SECTORS) {
899 rbi2 = r5_next_bio(rbi, dev->sector);
e7836bd6 900 if (!raid5_dec_bi_active_stripes(rbi)) {
91c00924
DW		 901 rbi->bi_next = return_bi;
902 return_bi = rbi;
903 }
91c00924
DW		 904 rbi = rbi2;
905 }
906 }
907 }
83de75cc 908 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
91c00924
DW		 909
910 return_io(return_bi);
911
e4d84909 912 set_bit(STRIPE_HANDLE, &sh->state);
91c00924
DW		 913 release_stripe(sh);
914}
915
916static void ops_run_biofill(struct stripe_head *sh)
917{
918 struct dma_async_tx_descriptor *tx = NULL;
a08abd8c 919 struct async_submit_ctl submit;
91c00924
DW		 920 int i;
921
e46b272b 922 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 923 (unsigned long long)sh->sector);
924
925 for (i = sh->disks; i--; ) {
926 struct r5dev *dev = &sh->dev[i];
927 if (test_bit(R5_Wantfill, &dev->flags)) {
928 struct bio *rbi;
b17459c0 929 spin_lock_irq(&sh->stripe_lock);
91c00924
DW		 930 dev->read = rbi = dev->toread;
931 dev->toread = NULL;
b17459c0 932 spin_unlock_irq(&sh->stripe_lock);
91c00924
DW		 933 while (rbi && rbi->bi_sector <
934 dev->sector + STRIPE_SECTORS) {
935 tx = async_copy_data(0, rbi, dev->page,
936 dev->sector, tx);
937 rbi = r5_next_bio(rbi, dev->sector);
938 }
939 }
940 }
941
942 atomic_inc(&sh->count);
a08abd8c
DW		 943 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
944 async_trigger_callback(&submit);
91c00924
DW		 945}
946
4e7d2c0a 947static void mark_target_uptodate(struct stripe_head *sh, int target)
91c00924 948{
4e7d2c0a 949 struct r5dev *tgt;
91c00924 950
4e7d2c0a
DW		 951 if (target < 0)
952 return;
91c00924 953
4e7d2c0a 954 tgt = &sh->dev[target];
91c00924
DW		 955 set_bit(R5_UPTODATE, &tgt->flags);
956 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
957 clear_bit(R5_Wantcompute, &tgt->flags);
4e7d2c0a
DW		 958}
959
ac6b53b6 960static void ops_complete_compute(void *stripe_head_ref)
91c00924
DW		 961{
962 struct stripe_head *sh = stripe_head_ref;
91c00924 963
e46b272b 964 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 965 (unsigned long long)sh->sector);
966
ac6b53b6 967 /* mark the computed target(s) as uptodate */
4e7d2c0a 968 mark_target_uptodate(sh, sh->ops.target);
ac6b53b6 969 mark_target_uptodate(sh, sh->ops.target2);
4e7d2c0a 970
ecc65c9b
DW		 971 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
972 if (sh->check_state == check_state_compute_run)
973 sh->check_state = check_state_compute_result;
91c00924
DW		 974 set_bit(STRIPE_HANDLE, &sh->state);
975 release_stripe(sh);
976}
977
d6f38f31
DW		 978/* return a pointer to the address conversion region of the scribble buffer */
979static addr_conv_t *to_addr_conv(struct stripe_head *sh,
980 struct raid5_percpu *percpu)
981{
982 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
983}
984
985static struct dma_async_tx_descriptor *
986ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
91c00924 987{
91c00924 988 int disks = sh->disks;
d6f38f31 989 struct page **xor_srcs = percpu->scribble;
91c00924
DW		 990 int target = sh->ops.target;
991 struct r5dev *tgt = &sh->dev[target];
992 struct page *xor_dest = tgt->page;
993 int count = 0;
994 struct dma_async_tx_descriptor *tx;
a08abd8c 995 struct async_submit_ctl submit;
91c00924
DW		 996 int i;
997
998 pr_debug("%s: stripe %llu block: %d\n",
e46b272b 999 __func__, (unsigned long long)sh->sector, target);
91c00924
DW		 1000 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1001
1002 for (i = disks; i--; )
1003 if (i != target)
1004 xor_srcs[count++] = sh->dev[i].page;
1005
1006 atomic_inc(&sh->count);
1007
0403e382 1008 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
ac6b53b6 1009 ops_complete_compute, sh, to_addr_conv(sh, percpu));
91c00924 1010 if (unlikely(count == 1))
a08abd8c 1011 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
91c00924 1012 else
a08abd8c 1013 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
91c00924 1014
91c00924
DW		 1015 return tx;
1016}
1017
ac6b53b6
DW		 1018/* set_syndrome_sources - populate source buffers for gen_syndrome
1019 * @srcs - (struct page *) array of size sh->disks
1020 * @sh - stripe_head to parse
1021 *
1022 * Populates srcs in proper layout order for the stripe and returns the
1023 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1024 * destination buffer is recorded in srcs[count] and the Q destination
1025 * is recorded in srcs[count+1]].
1026 */
1027static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
1028{
1029 int disks = sh->disks;
1030 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1031 int d0_idx = raid6_d0(sh);
1032 int count;
1033 int i;
1034
1035 for (i = 0; i < disks; i++)
5dd33c9a 1036 srcs[i] = NULL;
ac6b53b6
DW		 1037
1038 count = 0;
1039 i = d0_idx;
1040 do {
1041 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1042
1043 srcs[slot] = sh->dev[i].page;
1044 i = raid6_next_disk(i, disks);
1045 } while (i != d0_idx);
ac6b53b6 1046
e4424fee 1047 return syndrome_disks;
ac6b53b6
DW		 1048}
1049
1050static struct dma_async_tx_descriptor *
1051ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1052{
1053 int disks = sh->disks;
1054 struct page **blocks = percpu->scribble;
1055 int target;
1056 int qd_idx = sh->qd_idx;
1057 struct dma_async_tx_descriptor *tx;
1058 struct async_submit_ctl submit;
1059 struct r5dev *tgt;
1060 struct page *dest;
1061 int i;
1062 int count;
1063
1064 if (sh->ops.target < 0)
1065 target = sh->ops.target2;
1066 else if (sh->ops.target2 < 0)
1067 target = sh->ops.target;
91c00924 1068 else
ac6b53b6
DW		 1069 /* we should only have one valid target */
1070 BUG();
1071 BUG_ON(target < 0);
1072 pr_debug("%s: stripe %llu block: %d\n",
1073 __func__, (unsigned long long)sh->sector, target);
1074
1075 tgt = &sh->dev[target];
1076 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1077 dest = tgt->page;
1078
1079 atomic_inc(&sh->count);
1080
1081 if (target == qd_idx) {
1082 count = set_syndrome_sources(blocks, sh);
1083 blocks[count] = NULL; /* regenerating p is not necessary */
1084 BUG_ON(blocks[count+1] != dest); /* q should already be set */
0403e382
DW		 1085 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1086 ops_complete_compute, sh,
ac6b53b6
DW		 1087 to_addr_conv(sh, percpu));
1088 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1089 } else {
1090 /* Compute any data- or p-drive using XOR */
1091 count = 0;
1092 for (i = disks; i-- ; ) {
1093 if (i == target || i == qd_idx)
1094 continue;
1095 blocks[count++] = sh->dev[i].page;
1096 }
1097
0403e382
DW		 1098 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1099 NULL, ops_complete_compute, sh,
ac6b53b6
DW		 1100 to_addr_conv(sh, percpu));
1101 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1102 }
91c00924 1103
91c00924
DW		 1104 return tx;
1105}
1106
ac6b53b6
DW		 1107static struct dma_async_tx_descriptor *
1108ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1109{
1110 int i, count, disks = sh->disks;
1111 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1112 int d0_idx = raid6_d0(sh);
1113 int faila = -1, failb = -1;
1114 int target = sh->ops.target;
1115 int target2 = sh->ops.target2;
1116 struct r5dev *tgt = &sh->dev[target];
1117 struct r5dev *tgt2 = &sh->dev[target2];
1118 struct dma_async_tx_descriptor *tx;
1119 struct page **blocks = percpu->scribble;
1120 struct async_submit_ctl submit;
1121
1122 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1123 __func__, (unsigned long long)sh->sector, target, target2);
1124 BUG_ON(target < 0 || target2 < 0);
1125 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1126 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1127
6c910a78 1128 /* we need to open-code set_syndrome_sources to handle the
ac6b53b6
DW		 1129 * slot number conversion for 'faila' and 'failb'
1130 */
1131 for (i = 0; i < disks ; i++)
5dd33c9a 1132 blocks[i] = NULL;
ac6b53b6
DW		 1133 count = 0;
1134 i = d0_idx;
1135 do {
1136 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1137
1138 blocks[slot] = sh->dev[i].page;
1139
1140 if (i == target)
1141 faila = slot;
1142 if (i == target2)
1143 failb = slot;
1144 i = raid6_next_disk(i, disks);
1145 } while (i != d0_idx);
ac6b53b6
DW		 1146
1147 BUG_ON(faila == failb);
1148 if (failb < faila)
1149 swap(faila, failb);
1150 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1151 __func__, (unsigned long long)sh->sector, faila, failb);
1152
1153 atomic_inc(&sh->count);
1154
1155 if (failb == syndrome_disks+1) {
1156 /* Q disk is one of the missing disks */
1157 if (faila == syndrome_disks) {
1158 /* Missing P+Q, just recompute */
0403e382
DW		 1159 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1160 ops_complete_compute, sh,
1161 to_addr_conv(sh, percpu));
e4424fee 1162 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
ac6b53b6
DW		 1163 STRIPE_SIZE, &submit);
1164 } else {
1165 struct page *dest;
1166 int data_target;
1167 int qd_idx = sh->qd_idx;
1168
1169 /* Missing D+Q: recompute D from P, then recompute Q */
1170 if (target == qd_idx)
1171 data_target = target2;
1172 else
1173 data_target = target;
1174
1175 count = 0;
1176 for (i = disks; i-- ; ) {
1177 if (i == data_target || i == qd_idx)
1178 continue;
1179 blocks[count++] = sh->dev[i].page;
1180 }
1181 dest = sh->dev[data_target].page;
0403e382
DW		 1182 init_async_submit(&submit,
1183 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1184 NULL, NULL, NULL,
1185 to_addr_conv(sh, percpu));
ac6b53b6
DW		 1186 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1187 &submit);
1188
1189 count = set_syndrome_sources(blocks, sh);
0403e382
DW		 1190 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1191 ops_complete_compute, sh,
1192 to_addr_conv(sh, percpu));
ac6b53b6
DW		 1193 return async_gen_syndrome(blocks, 0, count+2,
1194 STRIPE_SIZE, &submit);
1195 }
ac6b53b6 1196 } else {
6c910a78
DW		 1197 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1198 ops_complete_compute, sh,
1199 to_addr_conv(sh, percpu));
1200 if (failb == syndrome_disks) {
1201 /* We're missing D+P. */
1202 return async_raid6_datap_recov(syndrome_disks+2,
1203 STRIPE_SIZE, faila,
1204 blocks, &submit);
1205 } else {
1206 /* We're missing D+D. */
1207 return async_raid6_2data_recov(syndrome_disks+2,
1208 STRIPE_SIZE, faila, failb,
1209 blocks, &submit);
1210 }
ac6b53b6
DW		 1211 }
1212}
1213
1214
91c00924
DW		 1215static void ops_complete_prexor(void *stripe_head_ref)
1216{
1217 struct stripe_head *sh = stripe_head_ref;
1218
e46b272b 1219 pr_debug("%s: stripe %llu\n", __func__,
91c00924 1220 (unsigned long long)sh->sector);
91c00924
DW		 1221}
1222
1223static struct dma_async_tx_descriptor *
d6f38f31
DW		 1224ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1225 struct dma_async_tx_descriptor *tx)
91c00924 1226{
91c00924 1227 int disks = sh->disks;
d6f38f31 1228 struct page **xor_srcs = percpu->scribble;
91c00924 1229 int count = 0, pd_idx = sh->pd_idx, i;
a08abd8c 1230 struct async_submit_ctl submit;
91c00924
DW		 1231
1232 /* existing parity data subtracted */
1233 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1234
e46b272b 1235 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1236 (unsigned long long)sh->sector);
1237
1238 for (i = disks; i--; ) {
1239 struct r5dev *dev = &sh->dev[i];
1240 /* Only process blocks that are known to be uptodate */
d8ee0728 1241 if (test_bit(R5_Wantdrain, &dev->flags))
91c00924
DW		 1242 xor_srcs[count++] = dev->page;
1243 }
1244
0403e382 1245 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
d6f38f31 1246 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
a08abd8c 1247 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
91c00924
DW		 1248
1249 return tx;
1250}
1251
1252static struct dma_async_tx_descriptor *
d8ee0728 1253ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
91c00924
DW		 1254{
1255 int disks = sh->disks;
d8ee0728 1256 int i;
91c00924 1257
e46b272b 1258 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1259 (unsigned long long)sh->sector);
1260
1261 for (i = disks; i--; ) {
1262 struct r5dev *dev = &sh->dev[i];
1263 struct bio *chosen;
91c00924 1264
d8ee0728 1265 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
91c00924
DW		 1266 struct bio *wbi;
1267
b17459c0 1268 spin_lock_irq(&sh->stripe_lock);
91c00924
DW		 1269 chosen = dev->towrite;
1270 dev->towrite = NULL;
1271 BUG_ON(dev->written);
1272 wbi = dev->written = chosen;
b17459c0 1273 spin_unlock_irq(&sh->stripe_lock);
91c00924
DW		 1274
1275 while (wbi && wbi->bi_sector <
1276 dev->sector + STRIPE_SECTORS) {
e9c7469b
TH		 1277 if (wbi->bi_rw & REQ_FUA)
1278 set_bit(R5_WantFUA, &dev->flags);
bc0934f0
SL		 1279 if (wbi->bi_rw & REQ_SYNC)
1280 set_bit(R5_SyncIO, &dev->flags);
9e444768 1281 if (wbi->bi_rw & REQ_DISCARD)
620125f2 1282 set_bit(R5_Discard, &dev->flags);
9e444768 1283 else
620125f2
SL		 1284 tx = async_copy_data(1, wbi, dev->page,
1285 dev->sector, tx);
91c00924
DW		 1286 wbi = r5_next_bio(wbi, dev->sector);
1287 }
1288 }
1289 }
1290
1291 return tx;
1292}
1293
ac6b53b6 1294static void ops_complete_reconstruct(void *stripe_head_ref)
91c00924
DW		 1295{
1296 struct stripe_head *sh = stripe_head_ref;
ac6b53b6
DW		 1297 int disks = sh->disks;
1298 int pd_idx = sh->pd_idx;
1299 int qd_idx = sh->qd_idx;
1300 int i;
9e444768 1301 bool fua = false, sync = false, discard = false;
91c00924 1302
e46b272b 1303 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1304 (unsigned long long)sh->sector);
1305
bc0934f0 1306 for (i = disks; i--; ) {
e9c7469b 1307 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
bc0934f0 1308 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
9e444768 1309 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
bc0934f0 1310 }
e9c7469b 1311
91c00924
DW		 1312 for (i = disks; i--; ) {
1313 struct r5dev *dev = &sh->dev[i];
ac6b53b6 1314
e9c7469b 1315 if (dev->written || i == pd_idx || i == qd_idx) {
9e444768
SL		 1316 if (!discard)
1317 set_bit(R5_UPTODATE, &dev->flags);
e9c7469b
TH		 1318 if (fua)
1319 set_bit(R5_WantFUA, &dev->flags);
bc0934f0
SL		 1320 if (sync)
1321 set_bit(R5_SyncIO, &dev->flags);
e9c7469b 1322 }
91c00924
DW		 1323 }
1324
d8ee0728
DW		 1325 if (sh->reconstruct_state == reconstruct_state_drain_run)
1326 sh->reconstruct_state = reconstruct_state_drain_result;
1327 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1328 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1329 else {
1330 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1331 sh->reconstruct_state = reconstruct_state_result;
1332 }
91c00924
DW		 1333
1334 set_bit(STRIPE_HANDLE, &sh->state);
1335 release_stripe(sh);
1336}
1337
1338static void
ac6b53b6
DW		 1339ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1340 struct dma_async_tx_descriptor *tx)
91c00924 1341{
91c00924 1342 int disks = sh->disks;
d6f38f31 1343 struct page **xor_srcs = percpu->scribble;
a08abd8c 1344 struct async_submit_ctl submit;
91c00924
DW		 1345 int count = 0, pd_idx = sh->pd_idx, i;
1346 struct page *xor_dest;
d8ee0728 1347 int prexor = 0;
91c00924 1348 unsigned long flags;
91c00924 1349
e46b272b 1350 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1351 (unsigned long long)sh->sector);
1352
620125f2
SL		 1353 for (i = 0; i < sh->disks; i++) {
1354 if (pd_idx == i)
1355 continue;
1356 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1357 break;
1358 }
1359 if (i >= sh->disks) {
1360 atomic_inc(&sh->count);
620125f2
SL		 1361 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1362 ops_complete_reconstruct(sh);
1363 return;
1364 }
91c00924
DW		 1365 /* check if prexor is active which means only process blocks
1366 * that are part of a read-modify-write (written)
1367 */
d8ee0728
DW		 1368 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1369 prexor = 1;
91c00924
DW		 1370 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1371 for (i = disks; i--; ) {
1372 struct r5dev *dev = &sh->dev[i];
1373 if (dev->written)
1374 xor_srcs[count++] = dev->page;
1375 }
1376 } else {
1377 xor_dest = sh->dev[pd_idx].page;
1378 for (i = disks; i--; ) {
1379 struct r5dev *dev = &sh->dev[i];
1380 if (i != pd_idx)
1381 xor_srcs[count++] = dev->page;
1382 }
1383 }
1384
91c00924
DW		 1385 /* 1/ if we prexor'd then the dest is reused as a source
1386 * 2/ if we did not prexor then we are redoing the parity
1387 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1388 * for the synchronous xor case
1389 */
88ba2aa5 1390 flags = ASYNC_TX_ACK |
91c00924
DW		 1391 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1392
1393 atomic_inc(&sh->count);
1394
ac6b53b6 1395 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
d6f38f31 1396 to_addr_conv(sh, percpu));
a08abd8c
DW		 1397 if (unlikely(count == 1))
1398 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1399 else
1400 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
91c00924
DW		 1401}
1402
ac6b53b6
DW		 1403static void
1404ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1405 struct dma_async_tx_descriptor *tx)
1406{
1407 struct async_submit_ctl submit;
1408 struct page **blocks = percpu->scribble;
620125f2 1409 int count, i;
ac6b53b6
DW		 1410
1411 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1412
620125f2
SL		 1413 for (i = 0; i < sh->disks; i++) {
1414 if (sh->pd_idx == i || sh->qd_idx == i)
1415 continue;
1416 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1417 break;
1418 }
1419 if (i >= sh->disks) {
1420 atomic_inc(&sh->count);
620125f2
SL		 1421 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1422 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1423 ops_complete_reconstruct(sh);
1424 return;
1425 }
1426
ac6b53b6
DW		 1427 count = set_syndrome_sources(blocks, sh);
1428
1429 atomic_inc(&sh->count);
1430
1431 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1432 sh, to_addr_conv(sh, percpu));
1433 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
91c00924
DW		 1434}
1435
1436static void ops_complete_check(void *stripe_head_ref)
1437{
1438 struct stripe_head *sh = stripe_head_ref;
91c00924 1439
e46b272b 1440 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1441 (unsigned long long)sh->sector);
1442
ecc65c9b 1443 sh->check_state = check_state_check_result;
91c00924
DW		 1444 set_bit(STRIPE_HANDLE, &sh->state);
1445 release_stripe(sh);
1446}
1447
ac6b53b6 1448static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
91c00924 1449{
91c00924 1450 int disks = sh->disks;
ac6b53b6
DW		 1451 int pd_idx = sh->pd_idx;
1452 int qd_idx = sh->qd_idx;
1453 struct page *xor_dest;
d6f38f31 1454 struct page **xor_srcs = percpu->scribble;
91c00924 1455 struct dma_async_tx_descriptor *tx;
a08abd8c 1456 struct async_submit_ctl submit;
ac6b53b6
DW		 1457 int count;
1458 int i;
91c00924 1459
e46b272b 1460 pr_debug("%s: stripe %llu\n", __func__,
91c00924
DW		 1461 (unsigned long long)sh->sector);
1462
ac6b53b6
DW		 1463 count = 0;
1464 xor_dest = sh->dev[pd_idx].page;
1465 xor_srcs[count++] = xor_dest;
91c00924 1466 for (i = disks; i--; ) {
ac6b53b6
DW		 1467 if (i == pd_idx || i == qd_idx)
1468 continue;
1469 xor_srcs[count++] = sh->dev[i].page;
91c00924
DW		 1470 }
1471
d6f38f31
DW		 1472 init_async_submit(&submit, 0, NULL, NULL, NULL,
1473 to_addr_conv(sh, percpu));
099f53cb 1474 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
a08abd8c 1475 &sh->ops.zero_sum_result, &submit);
91c00924 1476
91c00924 1477 atomic_inc(&sh->count);
a08abd8c
DW		 1478 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1479 tx = async_trigger_callback(&submit);
91c00924
DW		 1480}
1481
ac6b53b6
DW		 1482static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1483{
1484 struct page **srcs = percpu->scribble;
1485 struct async_submit_ctl submit;
1486 int count;
1487
1488 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1489 (unsigned long long)sh->sector, checkp);
1490
1491 count = set_syndrome_sources(srcs, sh);
1492 if (!checkp)
1493 srcs[count] = NULL;
91c00924 1494
91c00924 1495 atomic_inc(&sh->count);
ac6b53b6
DW		 1496 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1497 sh, to_addr_conv(sh, percpu));
1498 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1499 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
91c00924
DW		 1500}
1501
51acbcec 1502static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
91c00924
DW		 1503{
1504 int overlap_clear = 0, i, disks = sh->disks;
1505 struct dma_async_tx_descriptor *tx = NULL;
d1688a6d 1506 struct r5conf *conf = sh->raid_conf;
ac6b53b6 1507 int level = conf->level;
d6f38f31
DW		 1508 struct raid5_percpu *percpu;
1509 unsigned long cpu;
91c00924 1510
d6f38f31
DW		 1511 cpu = get_cpu();
1512 percpu = per_cpu_ptr(conf->percpu, cpu);
83de75cc 1513 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
91c00924
DW		 1514 ops_run_biofill(sh);
1515 overlap_clear++;
1516 }
1517
7b3a871e 1518 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
ac6b53b6
DW		 1519 if (level < 6)
1520 tx = ops_run_compute5(sh, percpu);
1521 else {
1522 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1523 tx = ops_run_compute6_1(sh, percpu);
1524 else
1525 tx = ops_run_compute6_2(sh, percpu);
1526 }
1527 /* terminate the chain if reconstruct is not set to be run */
1528 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
7b3a871e
DW		 1529 async_tx_ack(tx);
1530 }
91c00924 1531
600aa109 1532 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
d6f38f31 1533 tx = ops_run_prexor(sh, percpu, tx);
91c00924 1534
600aa109 1535 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
d8ee0728 1536 tx = ops_run_biodrain(sh, tx);
91c00924
DW		 1537 overlap_clear++;
1538 }
1539
ac6b53b6
DW		 1540 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1541 if (level < 6)
1542 ops_run_reconstruct5(sh, percpu, tx);
1543 else
1544 ops_run_reconstruct6(sh, percpu, tx);
1545 }
91c00924 1546
ac6b53b6
DW		 1547 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1548 if (sh->check_state == check_state_run)
1549 ops_run_check_p(sh, percpu);
1550 else if (sh->check_state == check_state_run_q)
1551 ops_run_check_pq(sh, percpu, 0);
1552 else if (sh->check_state == check_state_run_pq)
1553 ops_run_check_pq(sh, percpu, 1);
1554 else
1555 BUG();
1556 }
91c00924 1557
91c00924
DW		 1558 if (overlap_clear)
1559 for (i = disks; i--; ) {
1560 struct r5dev *dev = &sh->dev[i];
1561 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1562 wake_up(&sh->raid_conf->wait_for_overlap);
1563 }
d6f38f31 1564 put_cpu();
91c00924
DW		 1565}
1566
d1688a6d 1567static int grow_one_stripe(struct r5conf *conf)
1da177e4
LT		 1568{
1569 struct stripe_head *sh;
6ce32846 1570 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
3f294f4f
N		 1571 if (!sh)
1572 return 0;
6ce32846 1573
3f294f4f 1574 sh->raid_conf = conf;
3f294f4f 1575
b17459c0
SL		 1576 spin_lock_init(&sh->stripe_lock);
1577
e4e11e38
N		 1578 if (grow_buffers(sh)) {
1579 shrink_buffers(sh);
3f294f4f
N		 1580 kmem_cache_free(conf->slab_cache, sh);
1581 return 0;
1582 }
1583 /* we just created an active stripe so... */
1584 atomic_set(&sh->count, 1);
1585 atomic_inc(&conf->active_stripes);
1586 INIT_LIST_HEAD(&sh->lru);
1587 release_stripe(sh);
1588 return 1;
1589}
1590
d1688a6d 1591static int grow_stripes(struct r5conf *conf, int num)
3f294f4f 1592{
e18b890b 1593 struct kmem_cache *sc;
5e5e3e78 1594 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1da177e4 1595
f4be6b43
N		 1596 if (conf->mddev->gendisk)
1597 sprintf(conf->cache_name[0],
1598 "raid%d-%s", conf->level, mdname(conf->mddev));
1599 else
1600 sprintf(conf->cache_name[0],
1601 "raid%d-%p", conf->level, conf->mddev);
1602 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1603
ad01c9e3
N		 1604 conf->active_name = 0;
1605 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1da177e4 1606 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
20c2df83 1607 0, 0, NULL);
1da177e4
LT		 1608 if (!sc)
1609 return 1;
1610 conf->slab_cache = sc;
ad01c9e3 1611 conf->pool_size = devs;
16a53ecc 1612 while (num--)
3f294f4f 1613 if (!grow_one_stripe(conf))
1da177e4 1614 return 1;
1da177e4
LT		 1615 return 0;
1616}
29269553 1617
d6f38f31
DW		 1618/**
1619 * scribble_len - return the required size of the scribble region
1620 * @num - total number of disks in the array
1621 *
1622 * The size must be enough to contain:
1623 * 1/ a struct page pointer for each device in the array +2
1624 * 2/ room to convert each entry in (1) to its corresponding dma
1625 * (dma_map_page()) or page (page_address()) address.
1626 *
1627 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1628 * calculate over all devices (not just the data blocks), using zeros in place
1629 * of the P and Q blocks.
1630 */
1631static size_t scribble_len(int num)
1632{
1633 size_t len;
1634
1635 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1636
1637 return len;
1638}
1639
d1688a6d 1640static int resize_stripes(struct r5conf *conf, int newsize)
ad01c9e3
N		 1641{
1642 /* Make all the stripes able to hold 'newsize' devices.
1643 * New slots in each stripe get 'page' set to a new page.
1644 *
1645 * This happens in stages:
1646 * 1/ create a new kmem_cache and allocate the required number of
1647 * stripe_heads.
83f0d77a 1648 * 2/ gather all the old stripe_heads and transfer the pages across
ad01c9e3
N		 1649 * to the new stripe_heads. This will have the side effect of
1650 * freezing the array as once all stripe_heads have been collected,
1651 * no IO will be possible. Old stripe heads are freed once their
1652 * pages have been transferred over, and the old kmem_cache is
1653 * freed when all stripes are done.
1654 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1655 * we simple return a failre status - no need to clean anything up.
1656 * 4/ allocate new pages for the new slots in the new stripe_heads.
1657 * If this fails, we don't bother trying the shrink the
1658 * stripe_heads down again, we just leave them as they are.
1659 * As each stripe_head is processed the new one is released into
1660 * active service.
1661 *
1662 * Once step2 is started, we cannot afford to wait for a write,
1663 * so we use GFP_NOIO allocations.
1664 */
1665 struct stripe_head *osh, *nsh;
1666 LIST_HEAD(newstripes);
1667 struct disk_info *ndisks;
d6f38f31 1668 unsigned long cpu;
b5470dc5 1669 int err;
e18b890b 1670 struct kmem_cache *sc;
ad01c9e3
N		 1671 int i;
1672
1673 if (newsize <= conf->pool_size)
1674 return 0; /* never bother to shrink */
1675
b5470dc5
DW		 1676 err = md_allow_write(conf->mddev);
1677 if (err)
1678 return err;
2a2275d6 1679
ad01c9e3
N		 1680 /* Step 1 */
1681 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1682 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
20c2df83 1683 0, 0, NULL);
ad01c9e3
N		 1684 if (!sc)
1685 return -ENOMEM;
1686
1687 for (i = conf->max_nr_stripes; i; i--) {
6ce32846 1688 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
ad01c9e3
N		 1689 if (!nsh)
1690 break;
1691
ad01c9e3 1692 nsh->raid_conf = conf;
cb13ff69 1693 spin_lock_init(&nsh->stripe_lock);
ad01c9e3
N		 1694
1695 list_add(&nsh->lru, &newstripes);
1696 }
1697 if (i) {
1698 /* didn't get enough, give up */
1699 while (!list_empty(&newstripes)) {
1700 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1701 list_del(&nsh->lru);
1702 kmem_cache_free(sc, nsh);
1703 }
1704 kmem_cache_destroy(sc);
1705 return -ENOMEM;
1706 }
1707 /* Step 2 - Must use GFP_NOIO now.
1708 * OK, we have enough stripes, start collecting inactive
1709 * stripes and copying them over
1710 */
1711 list_for_each_entry(nsh, &newstripes, lru) {
1712 spin_lock_irq(&conf->device_lock);
1713 wait_event_lock_irq(conf->wait_for_stripe,
1714 !list_empty(&conf->inactive_list),
eed8c02e 1715 conf->device_lock);
ad01c9e3
N		 1716 osh = get_free_stripe(conf);
1717 spin_unlock_irq(&conf->device_lock);
1718 atomic_set(&nsh->count, 1);
1719 for(i=0; i<conf->pool_size; i++)
1720 nsh->dev[i].page = osh->dev[i].page;
1721 for( ; i<newsize; i++)
1722 nsh->dev[i].page = NULL;
1723 kmem_cache_free(conf->slab_cache, osh);
1724 }
1725 kmem_cache_destroy(conf->slab_cache);
1726
1727 /* Step 3.
1728 * At this point, we are holding all the stripes so the array
1729 * is completely stalled, so now is a good time to resize
d6f38f31 1730 * conf->disks and the scribble region
ad01c9e3
N		 1731 */
1732 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1733 if (ndisks) {
1734 for (i=0; i<conf->raid_disks; i++)
1735 ndisks[i] = conf->disks[i];
1736 kfree(conf->disks);
1737 conf->disks = ndisks;
1738 } else
1739 err = -ENOMEM;
1740
d6f38f31
DW		 1741 get_online_cpus();
1742 conf->scribble_len = scribble_len(newsize);
1743 for_each_present_cpu(cpu) {
1744 struct raid5_percpu *percpu;
1745 void *scribble;
1746
1747 percpu = per_cpu_ptr(conf->percpu, cpu);
1748 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1749
1750 if (scribble) {
1751 kfree(percpu->scribble);
1752 percpu->scribble = scribble;
1753 } else {
1754 err = -ENOMEM;
1755 break;
1756 }
1757 }
1758 put_online_cpus();
1759
ad01c9e3
N		 1760 /* Step 4, return new stripes to service */
1761 while(!list_empty(&newstripes)) {
1762 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1763 list_del_init(&nsh->lru);
d6f38f31 1764
ad01c9e3
N		 1765 for (i=conf->raid_disks; i < newsize; i++)
1766 if (nsh->dev[i].page == NULL) {
1767 struct page *p = alloc_page(GFP_NOIO);
1768 nsh->dev[i].page = p;
1769 if (!p)
1770 err = -ENOMEM;
1771 }
1772 release_stripe(nsh);
1773 }
1774 /* critical section pass, GFP_NOIO no longer needed */
1775
1776 conf->slab_cache = sc;
1777 conf->active_name = 1-conf->active_name;
1778 conf->pool_size = newsize;
1779 return err;
1780}
1da177e4 1781
d1688a6d 1782static int drop_one_stripe(struct r5conf *conf)
1da177e4
LT		 1783{
1784 struct stripe_head *sh;
1785
3f294f4f
N		 1786 spin_lock_irq(&conf->device_lock);
1787 sh = get_free_stripe(conf);
1788 spin_unlock_irq(&conf->device_lock);
1789 if (!sh)
1790 return 0;
78bafebd 1791 BUG_ON(atomic_read(&sh->count));
e4e11e38 1792 shrink_buffers(sh);
3f294f4f
N		 1793 kmem_cache_free(conf->slab_cache, sh);
1794 atomic_dec(&conf->active_stripes);
1795 return 1;
1796}
1797
d1688a6d 1798static void shrink_stripes(struct r5conf *conf)
3f294f4f
N		 1799{
1800 while (drop_one_stripe(conf))
1801 ;
1802
29fc7e3e
N		 1803 if (conf->slab_cache)
1804 kmem_cache_destroy(conf->slab_cache);
1da177e4
LT		 1805 conf->slab_cache = NULL;
1806}
1807
6712ecf8 1808static void raid5_end_read_request(struct bio * bi, int error)
1da177e4 1809{
99c0fb5f 1810 struct stripe_head *sh = bi->bi_private;
d1688a6d 1811 struct r5conf *conf = sh->raid_conf;
7ecaa1e6 1812 int disks = sh->disks, i;
1da177e4 1813 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
d6950432 1814 char b[BDEVNAME_SIZE];
dd054fce 1815 struct md_rdev *rdev = NULL;
05616be5 1816 sector_t s;
1da177e4
LT		 1817
1818 for (i=0 ; i<disks; i++)
1819 if (bi == &sh->dev[i].req)
1820 break;
1821
45b4233c
DW		 1822 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1823 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1da177e4
LT		 1824 uptodate);
1825 if (i == disks) {
1826 BUG();
6712ecf8 1827 return;
1da177e4 1828 }
14a75d3e 1829 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
dd054fce
N		 1830 /* If replacement finished while this request was outstanding,
1831 * 'replacement' might be NULL already.
1832 * In that case it moved down to 'rdev'.
1833 * rdev is not removed until all requests are finished.
1834 */
14a75d3e 1835 rdev = conf->disks[i].replacement;
dd054fce 1836 if (!rdev)
14a75d3e 1837 rdev = conf->disks[i].rdev;
1da177e4 1838
05616be5
N		 1839 if (use_new_offset(conf, sh))
1840 s = sh->sector + rdev->new_data_offset;
1841 else
1842 s = sh->sector + rdev->data_offset;
1da177e4 1843 if (uptodate) {
1da177e4 1844 set_bit(R5_UPTODATE, &sh->dev[i].flags);
4e5314b5 1845 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
14a75d3e
N		 1846 /* Note that this cannot happen on a
1847 * replacement device. We just fail those on
1848 * any error
1849 */
8bda470e
CD		 1850 printk_ratelimited(
1851 KERN_INFO
1852 "md/raid:%s: read error corrected"
1853 " (%lu sectors at %llu on %s)\n",
1854 mdname(conf->mddev), STRIPE_SECTORS,
05616be5 1855 (unsigned long long)s,
8bda470e 1856 bdevname(rdev->bdev, b));
ddd5115f 1857 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
4e5314b5
N		 1858 clear_bit(R5_ReadError, &sh->dev[i].flags);
1859 clear_bit(R5_ReWrite, &sh->dev[i].flags);
3f9e7c14 1860 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1861 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1862
14a75d3e
N		 1863 if (atomic_read(&rdev->read_errors))
1864 atomic_set(&rdev->read_errors, 0);
1da177e4 1865 } else {
14a75d3e 1866 const char *bdn = bdevname(rdev->bdev, b);
ba22dcbf 1867 int retry = 0;
2e8ac303 1868 int set_bad = 0;
d6950432 1869
1da177e4 1870 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
d6950432 1871 atomic_inc(&rdev->read_errors);
14a75d3e
N		 1872 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1873 printk_ratelimited(
1874 KERN_WARNING
1875 "md/raid:%s: read error on replacement device "
1876 "(sector %llu on %s).\n",
1877 mdname(conf->mddev),
05616be5 1878 (unsigned long long)s,
14a75d3e 1879 bdn);
2e8ac303 1880 else if (conf->mddev->degraded >= conf->max_degraded) {
1881 set_bad = 1;
8bda470e
CD		 1882 printk_ratelimited(
1883 KERN_WARNING
1884 "md/raid:%s: read error not correctable "
1885 "(sector %llu on %s).\n",
1886 mdname(conf->mddev),
05616be5 1887 (unsigned long long)s,
8bda470e 1888 bdn);
2e8ac303 1889 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
4e5314b5 1890 /* Oh, no!!! */
2e8ac303 1891 set_bad = 1;
8bda470e
CD		 1892 printk_ratelimited(
1893 KERN_WARNING
1894 "md/raid:%s: read error NOT corrected!! "
1895 "(sector %llu on %s).\n",
1896 mdname(conf->mddev),
05616be5 1897 (unsigned long long)s,
8bda470e 1898 bdn);
2e8ac303 1899 } else if (atomic_read(&rdev->read_errors)
ba22dcbf 1900 > conf->max_nr_stripes)
14f8d26b 1901 printk(KERN_WARNING
0c55e022 1902 "md/raid:%s: Too many read errors, failing device %s.\n",
d6950432 1903 mdname(conf->mddev), bdn);
ba22dcbf
N		 1904 else
1905 retry = 1;
1906 if (retry)
3f9e7c14 1907 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1908 set_bit(R5_ReadError, &sh->dev[i].flags);
1909 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1910 } else
1911 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
ba22dcbf 1912 else {
4e5314b5
N		 1913 clear_bit(R5_ReadError, &sh->dev[i].flags);
1914 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2e8ac303 1915 if (!(set_bad
1916 && test_bit(In_sync, &rdev->flags)
1917 && rdev_set_badblocks(
1918 rdev, sh->sector, STRIPE_SECTORS, 0)))
1919 md_error(conf->mddev, rdev);
ba22dcbf 1920 }
1da177e4 1921 }
14a75d3e 1922 rdev_dec_pending(rdev, conf->mddev);
1da177e4
LT		 1923 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1924 set_bit(STRIPE_HANDLE, &sh->state);
1925 release_stripe(sh);
1da177e4
LT		 1926}
1927
d710e138 1928static void raid5_end_write_request(struct bio *bi, int error)
1da177e4 1929{
99c0fb5f 1930 struct stripe_head *sh = bi->bi_private;
d1688a6d 1931 struct r5conf *conf = sh->raid_conf;
7ecaa1e6 1932 int disks = sh->disks, i;
977df362 1933 struct md_rdev *uninitialized_var(rdev);
1da177e4 1934 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
b84db560
N		 1935 sector_t first_bad;
1936 int bad_sectors;
977df362 1937 int replacement = 0;
1da177e4 1938
977df362
N		 1939 for (i = 0 ; i < disks; i++) {
1940 if (bi == &sh->dev[i].req) {
1941 rdev = conf->disks[i].rdev;
1da177e4 1942 break;
977df362
N		 1943 }
1944 if (bi == &sh->dev[i].rreq) {
1945 rdev = conf->disks[i].replacement;
dd054fce
N		 1946 if (rdev)
1947 replacement = 1;
1948 else
1949 /* rdev was removed and 'replacement'
1950 * replaced it. rdev is not removed
1951 * until all requests are finished.
1952 */
1953 rdev = conf->disks[i].rdev;
977df362
N		 1954 break;
1955 }
1956 }
45b4233c 1957 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1da177e4
LT		 1958 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1959 uptodate);
1960 if (i == disks) {
1961 BUG();
6712ecf8 1962 return;
1da177e4
LT		 1963 }
1964
977df362
N		 1965 if (replacement) {
1966 if (!uptodate)
1967 md_error(conf->mddev, rdev);
1968 else if (is_badblock(rdev, sh->sector,
1969 STRIPE_SECTORS,
1970 &first_bad, &bad_sectors))
1971 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1972 } else {
1973 if (!uptodate) {
1974 set_bit(WriteErrorSeen, &rdev->flags);
1975 set_bit(R5_WriteError, &sh->dev[i].flags);
3a6de292
N		 1976 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1977 set_bit(MD_RECOVERY_NEEDED,
1978 &rdev->mddev->recovery);
977df362
N		 1979 } else if (is_badblock(rdev, sh->sector,
1980 STRIPE_SECTORS,
c0b32972 1981 &first_bad, &bad_sectors)) {
977df362 1982 set_bit(R5_MadeGood, &sh->dev[i].flags);
c0b32972
N		 1983 if (test_bit(R5_ReadError, &sh->dev[i].flags))
1984 /* That was a successful write so make
1985 * sure it looks like we already did
1986 * a re-write.
1987 */
1988 set_bit(R5_ReWrite, &sh->dev[i].flags);
1989 }
977df362
N		 1990 }
1991 rdev_dec_pending(rdev, conf->mddev);
1da177e4 1992
977df362
N		 1993 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1994 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1da177e4 1995 set_bit(STRIPE_HANDLE, &sh->state);
c04be0aa 1996 release_stripe(sh);
1da177e4
LT		 1997}
1998
784052ec 1999static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1da177e4 2000
784052ec 2001static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1da177e4
LT		 2002{
2003 struct r5dev *dev = &sh->dev[i];
2004
2005 bio_init(&dev->req);
2006 dev->req.bi_io_vec = &dev->vec;
2007 dev->req.bi_vcnt++;
2008 dev->req.bi_max_vecs++;
1da177e4 2009 dev->req.bi_private = sh;
995c4275 2010 dev->vec.bv_page = dev->page;
1da177e4 2011
977df362
N		 2012 bio_init(&dev->rreq);
2013 dev->rreq.bi_io_vec = &dev->rvec;
2014 dev->rreq.bi_vcnt++;
2015 dev->rreq.bi_max_vecs++;
2016 dev->rreq.bi_private = sh;
2017 dev->rvec.bv_page = dev->page;
2018
1da177e4 2019 dev->flags = 0;
784052ec 2020 dev->sector = compute_blocknr(sh, i, previous);
1da177e4
LT		 2021}
2022
fd01b88c 2023static void error(struct mddev *mddev, struct md_rdev *rdev)
1da177e4
LT		 2024{
2025 char b[BDEVNAME_SIZE];
d1688a6d 2026 struct r5conf *conf = mddev->private;
908f4fbd 2027 unsigned long flags;
0c55e022 2028 pr_debug("raid456: error called\n");
1da177e4 2029
908f4fbd
N		 2030 spin_lock_irqsave(&conf->device_lock, flags);
2031 clear_bit(In_sync, &rdev->flags);
2032 mddev->degraded = calc_degraded(conf);
2033 spin_unlock_irqrestore(&conf->device_lock, flags);
2034 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2035
de393cde 2036 set_bit(Blocked, &rdev->flags);
6f8d0c77
N		 2037 set_bit(Faulty, &rdev->flags);
2038 set_bit(MD_CHANGE_DEVS, &mddev->flags);
2039 printk(KERN_ALERT
2040 "md/raid:%s: Disk failure on %s, disabling device.\n"
2041 "md/raid:%s: Operation continuing on %d devices.\n",
2042 mdname(mddev),
2043 bdevname(rdev->bdev, b),
2044 mdname(mddev),
2045 conf->raid_disks - mddev->degraded);
16a53ecc 2046}
1da177e4
LT		 2047
2048/*
2049 * Input: a 'big' sector number,
2050 * Output: index of the data and parity disk, and the sector # in them.
2051 */
d1688a6d 2052static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
911d4ee8
N		 2053 int previous, int *dd_idx,
2054 struct stripe_head *sh)
1da177e4 2055{
6e3b96ed 2056 sector_t stripe, stripe2;
35f2a591 2057 sector_t chunk_number;
1da177e4 2058 unsigned int chunk_offset;
911d4ee8 2059 int pd_idx, qd_idx;
67cc2b81 2060 int ddf_layout = 0;
1da177e4 2061 sector_t new_sector;
e183eaed
N		 2062 int algorithm = previous ? conf->prev_algo
2063 : conf->algorithm;
09c9e5fa
AN		 2064 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2065 : conf->chunk_sectors;
112bf897
N		 2066 int raid_disks = previous ? conf->previous_raid_disks
2067 : conf->raid_disks;
2068 int data_disks = raid_disks - conf->max_degraded;
1da177e4
LT		 2069
2070 /* First compute the information on this sector */
2071
2072 /*
2073 * Compute the chunk number and the sector offset inside the chunk
2074 */
2075 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2076 chunk_number = r_sector;
1da177e4
LT		 2077
2078 /*
2079 * Compute the stripe number
2080 */
35f2a591
N		 2081 stripe = chunk_number;
2082 *dd_idx = sector_div(stripe, data_disks);
6e3b96ed 2083 stripe2 = stripe;
1da177e4
LT		 2084 /*
2085 * Select the parity disk based on the user selected algorithm.
2086 */
84789554 2087 pd_idx = qd_idx = -1;
16a53ecc
N		 2088 switch(conf->level) {
2089 case 4:
911d4ee8 2090 pd_idx = data_disks;
16a53ecc
N		 2091 break;
2092 case 5:
e183eaed 2093 switch (algorithm) {
1da177e4 2094 case ALGORITHM_LEFT_ASYMMETRIC:
6e3b96ed 2095 pd_idx = data_disks - sector_div(stripe2, raid_disks);
911d4ee8 2096 if (*dd_idx >= pd_idx)
1da177e4
LT		 2097 (*dd_idx)++;
2098 break;
2099 case ALGORITHM_RIGHT_ASYMMETRIC:
6e3b96ed 2100 pd_idx = sector_div(stripe2, raid_disks);
911d4ee8 2101 if (*dd_idx >= pd_idx)
1da177e4
LT		 2102 (*dd_idx)++;
2103 break;
2104 case ALGORITHM_LEFT_SYMMETRIC:
6e3b96ed 2105 pd_idx = data_disks - sector_div(stripe2, raid_disks);
911d4ee8 2106 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1da177e4
LT		 2107 break;
2108 case ALGORITHM_RIGHT_SYMMETRIC:
6e3b96ed 2109 pd_idx = sector_div(stripe2, raid_disks);
911d4ee8 2110 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1da177e4 2111 break;
99c0fb5f
N		 2112 case ALGORITHM_PARITY_0:
2113 pd_idx = 0;
2114 (*dd_idx)++;
2115 break;
2116 case ALGORITHM_PARITY_N:
2117 pd_idx = data_disks;
2118 break;
1da177e4 2119 default:
99c0fb5f 2120 BUG();
16a53ecc
N		 2121 }
2122 break;
2123 case 6:
2124
e183eaed 2125 switch (algorithm) {
16a53ecc 2126 case ALGORITHM_LEFT_ASYMMETRIC:
6e3b96ed 2127 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
911d4ee8
N		 2128 qd_idx = pd_idx + 1;
2129 if (pd_idx == raid_disks-1) {
99c0fb5f 2130 (*dd_idx)++; /* Q D D D P */
911d4ee8
N		 2131 qd_idx = 0;
2132 } else if (*dd_idx >= pd_idx)
16a53ecc
N		 2133 (*dd_idx) += 2; /* D D P Q D */
2134 break;
2135 case ALGORITHM_RIGHT_ASYMMETRIC:
6e3b96ed 2136 pd_idx = sector_div(stripe2, raid_disks);
911d4ee8
N		 2137 qd_idx = pd_idx + 1;
2138 if (pd_idx == raid_disks-1) {
99c0fb5f 2139 (*dd_idx)++; /* Q D D D P */
911d4ee8
N		 2140 qd_idx = 0;
2141 } else if (*dd_idx >= pd_idx)
16a53ecc
N		 2142 (*dd_idx) += 2; /* D D P Q D */
2143 break;
2144 case ALGORITHM_LEFT_SYMMETRIC:
6e3b96ed 2145 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
911d4ee8
N		 2146 qd_idx = (pd_idx + 1) % raid_disks;
2147 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
16a53ecc
N		 2148 break;
2149 case ALGORITHM_RIGHT_SYMMETRIC:
6e3b96ed 2150 pd_idx = sector_div(stripe2, raid_disks);
911d4ee8
N		 2151 qd_idx = (pd_idx + 1) % raid_disks;
2152 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
16a53ecc 2153 break;
99c0fb5f
N		 2154
2155 case ALGORITHM_PARITY_0:
2156 pd_idx = 0;
2157 qd_idx = 1;
2158 (*dd_idx) += 2;
2159 break;
2160 case ALGORITHM_PARITY_N:
2161 pd_idx = data_disks;
2162 qd_idx = data_disks + 1;
2163 break;
2164
2165 case ALGORITHM_ROTATING_ZERO_RESTART:
2166 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2167 * of blocks for computing Q is different.
2168 */
6e3b96ed 2169 pd_idx = sector_div(stripe2, raid_disks);
99c0fb5f
N		 2170 qd_idx = pd_idx + 1;
2171 if (pd_idx == raid_disks-1) {
2172 (*dd_idx)++; /* Q D D D P */
2173 qd_idx = 0;
2174 } else if (*dd_idx >= pd_idx)
2175 (*dd_idx) += 2; /* D D P Q D */
67cc2b81 2176 ddf_layout = 1;
99c0fb5f
N		 2177 break;
2178
2179 case ALGORITHM_ROTATING_N_RESTART:
2180 /* Same a left_asymmetric, by first stripe is
2181 * D D D P Q rather than
2182 * Q D D D P
2183 */
6e3b96ed
N		 2184 stripe2 += 1;
2185 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
99c0fb5f
N		 2186 qd_idx = pd_idx + 1;
2187 if (pd_idx == raid_disks-1) {
2188 (*dd_idx)++; /* Q D D D P */
2189 qd_idx = 0;
2190 } else if (*dd_idx >= pd_idx)
2191 (*dd_idx) += 2; /* D D P Q D */
67cc2b81 2192 ddf_layout = 1;
99c0fb5f
N		 2193 break;
2194
2195 case ALGORITHM_ROTATING_N_CONTINUE:
2196 /* Same as left_symmetric but Q is before P */
6e3b96ed 2197 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
99c0fb5f
N		 2198 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2199 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
67cc2b81 2200 ddf_layout = 1;
99c0fb5f
N		 2201 break;
2202
2203 case ALGORITHM_LEFT_ASYMMETRIC_6:
2204 /* RAID5 left_asymmetric, with Q on last device */
6e3b96ed 2205 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
99c0fb5f
N		 2206 if (*dd_idx >= pd_idx)
2207 (*dd_idx)++;
2208 qd_idx = raid_disks - 1;
2209 break;
2210
2211 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6e3b96ed 2212 pd_idx = sector_div(stripe2, raid_disks-1);
99c0fb5f
N		 2213 if (*dd_idx >= pd_idx)
2214 (*dd_idx)++;
2215 qd_idx = raid_disks - 1;
2216 break;
2217
2218 case ALGORITHM_LEFT_SYMMETRIC_6:
6e3b96ed 2219 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
99c0fb5f
N		 2220 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2221 qd_idx = raid_disks - 1;
2222 break;
2223
2224 case ALGORITHM_RIGHT_SYMMETRIC_6:
6e3b96ed 2225 pd_idx = sector_div(stripe2, raid_disks-1);
99c0fb5f
N		 2226 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2227 qd_idx = raid_disks - 1;
2228 break;
2229
2230 case ALGORITHM_PARITY_0_6:
2231 pd_idx = 0;
2232 (*dd_idx)++;
2233 qd_idx = raid_disks - 1;
2234 break;
2235
16a53ecc 2236 default:
99c0fb5f 2237 BUG();
16a53ecc
N		 2238 }
2239 break;
1da177e4
LT		 2240 }
2241
911d4ee8
N		 2242 if (sh) {
2243 sh->pd_idx = pd_idx;
2244 sh->qd_idx = qd_idx;
67cc2b81 2245 sh->ddf_layout = ddf_layout;
911d4ee8 2246 }
1da177e4
LT		 2247 /*
2248 * Finally, compute the new sector number
2249 */
2250 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2251 return new_sector;
2252}
2253
2254
784052ec 2255static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1da177e4 2256{
d1688a6d 2257 struct r5conf *conf = sh->raid_conf;
b875e531
N		 2258 int raid_disks = sh->disks;
2259 int data_disks = raid_disks - conf->max_degraded;
1da177e4 2260 sector_t new_sector = sh->sector, check;
09c9e5fa
AN		 2261 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2262 : conf->chunk_sectors;
e183eaed
N		 2263 int algorithm = previous ? conf->prev_algo
2264 : conf->algorithm;
1da177e4
LT		 2265 sector_t stripe;
2266 int chunk_offset;
35f2a591
N		 2267 sector_t chunk_number;
2268 int dummy1, dd_idx = i;
1da177e4 2269 sector_t r_sector;
911d4ee8 2270 struct stripe_head sh2;
1da177e4 2271
16a53ecc 2272
1da177e4
LT		 2273 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2274 stripe = new_sector;
1da177e4 2275
16a53ecc
N		 2276 if (i == sh->pd_idx)
2277 return 0;
2278 switch(conf->level) {
2279 case 4: break;
2280 case 5:
e183eaed 2281 switch (algorithm) {
1da177e4
LT		 2282 case ALGORITHM_LEFT_ASYMMETRIC:
2283 case ALGORITHM_RIGHT_ASYMMETRIC:
2284 if (i > sh->pd_idx)
2285 i--;
2286 break;
2287 case ALGORITHM_LEFT_SYMMETRIC:
2288 case ALGORITHM_RIGHT_SYMMETRIC:
2289 if (i < sh->pd_idx)
2290 i += raid_disks;
2291 i -= (sh->pd_idx + 1);
2292 break;
99c0fb5f
N		 2293 case ALGORITHM_PARITY_0:
2294 i -= 1;
2295 break;
2296 case ALGORITHM_PARITY_N:
2297 break;
1da177e4 2298 default:
99c0fb5f 2299 BUG();
16a53ecc
N		 2300 }
2301 break;
2302 case 6:
d0dabf7e 2303 if (i == sh->qd_idx)
16a53ecc 2304 return 0; /* It is the Q disk */
e183eaed 2305 switch (algorithm) {
16a53ecc
N		 2306 case ALGORITHM_LEFT_ASYMMETRIC:
2307 case ALGORITHM_RIGHT_ASYMMETRIC:
99c0fb5f
N		 2308 case ALGORITHM_ROTATING_ZERO_RESTART:
2309 case ALGORITHM_ROTATING_N_RESTART:
2310 if (sh->pd_idx == raid_disks-1)
2311 i--; /* Q D D D P */
16a53ecc
N		 2312 else if (i > sh->pd_idx)
2313 i -= 2; /* D D P Q D */
2314 break;
2315 case ALGORITHM_LEFT_SYMMETRIC:
2316 case ALGORITHM_RIGHT_SYMMETRIC:
2317 if (sh->pd_idx == raid_disks-1)
2318 i--; /* Q D D D P */
2319 else {
2320 /* D D P Q D */
2321 if (i < sh->pd_idx)
2322 i += raid_disks;
2323 i -= (sh->pd_idx + 2);
2324 }
2325 break;
99c0fb5f
N		 2326 case ALGORITHM_PARITY_0:
2327 i -= 2;
2328 break;
2329 case ALGORITHM_PARITY_N:
2330 break;
2331 case ALGORITHM_ROTATING_N_CONTINUE:
e4424fee 2332 /* Like left_symmetric, but P is before Q */
99c0fb5f
N		 2333 if (sh->pd_idx == 0)
2334 i--; /* P D D D Q */
e4424fee
N		 2335 else {
2336 /* D D Q P D */
2337 if (i < sh->pd_idx)
2338 i += raid_disks;
2339 i -= (sh->pd_idx + 1);
2340 }
99c0fb5f
N		 2341 break;
2342 case ALGORITHM_LEFT_ASYMMETRIC_6:
2343 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2344 if (i > sh->pd_idx)
2345 i--;
2346 break;
2347 case ALGORITHM_LEFT_SYMMETRIC_6:
2348 case ALGORITHM_RIGHT_SYMMETRIC_6:
2349 if (i < sh->pd_idx)
2350 i += data_disks + 1;
2351 i -= (sh->pd_idx + 1);
2352 break;
2353 case ALGORITHM_PARITY_0_6:
2354 i -= 1;
2355 break;
16a53ecc 2356 default:
99c0fb5f 2357 BUG();
16a53ecc
N		 2358 }
2359 break;
1da177e4
LT		 2360 }
2361
2362 chunk_number = stripe * data_disks + i;
35f2a591 2363 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
1da177e4 2364
112bf897 2365 check = raid5_compute_sector(conf, r_sector,
784052ec 2366 previous, &dummy1, &sh2);
911d4ee8
N		 2367 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2368 || sh2.qd_idx != sh->qd_idx) {
0c55e022
N		 2369 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2370 mdname(conf->mddev));
1da177e4
LT		 2371 return 0;
2372 }
2373 return r_sector;
2374}
2375
2376
600aa109 2377static void
c0f7bddb 2378schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
600aa109 2379 int rcw, int expand)
e33129d8
DW		 2380{
2381 int i, pd_idx = sh->pd_idx, disks = sh->disks;
d1688a6d 2382 struct r5conf *conf = sh->raid_conf;
c0f7bddb 2383 int level = conf->level;
e33129d8
DW		 2384
2385 if (rcw) {
e33129d8
DW		 2386
2387 for (i = disks; i--; ) {
2388 struct r5dev *dev = &sh->dev[i];
2389
2390 if (dev->towrite) {
2391 set_bit(R5_LOCKED, &dev->flags);
d8ee0728 2392 set_bit(R5_Wantdrain, &dev->flags);
e33129d8
DW		 2393 if (!expand)
2394 clear_bit(R5_UPTODATE, &dev->flags);
600aa109 2395 s->locked++;
e33129d8
DW		 2396 }
2397 }
ce7d363a
N		 2398 /* if we are not expanding this is a proper write request, and
2399 * there will be bios with new data to be drained into the
2400 * stripe cache
2401 */
2402 if (!expand) {
2403 if (!s->locked)
2404 /* False alarm, nothing to do */
2405 return;
2406 sh->reconstruct_state = reconstruct_state_drain_run;
2407 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2408 } else
2409 sh->reconstruct_state = reconstruct_state_run;
2410
2411 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2412
c0f7bddb 2413 if (s->locked + conf->max_degraded == disks)
8b3e6cdc 2414 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
c0f7bddb 2415 atomic_inc(&conf->pending_full_writes);
e33129d8 2416 } else {
c0f7bddb 2417 BUG_ON(level == 6);
e33129d8
DW		 2418 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2419 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2420
e33129d8
DW		 2421 for (i = disks; i--; ) {
2422 struct r5dev *dev = &sh->dev[i];
2423 if (i == pd_idx)
2424 continue;
2425
e33129d8
DW		 2426 if (dev->towrite &&
2427 (test_bit(R5_UPTODATE, &dev->flags) ||
d8ee0728
DW		 2428 test_bit(R5_Wantcompute, &dev->flags))) {
2429 set_bit(R5_Wantdrain, &dev->flags);
e33129d8
DW		 2430 set_bit(R5_LOCKED, &dev->flags);
2431 clear_bit(R5_UPTODATE, &dev->flags);
600aa109 2432 s->locked++;
e33129d8
DW		 2433 }
2434 }
ce7d363a
N		 2435 if (!s->locked)
2436 /* False alarm - nothing to do */
2437 return;
2438 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2439 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2440 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2441 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
e33129d8
DW		 2442 }
2443
c0f7bddb 2444 /* keep the parity disk(s) locked while asynchronous operations
e33129d8
DW		 2445 * are in flight
2446 */
2447 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2448 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
600aa109 2449 s->locked++;
e33129d8 2450
c0f7bddb
YT		 2451 if (level == 6) {
2452 int qd_idx = sh->qd_idx;
2453 struct r5dev *dev = &sh->dev[qd_idx];
2454
2455 set_bit(R5_LOCKED, &dev->flags);
2456 clear_bit(R5_UPTODATE, &dev->flags);
2457 s->locked++;
2458 }
2459
600aa109 2460 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
e46b272b 2461 __func__, (unsigned long long)sh->sector,
600aa109 2462 s->locked, s->ops_request);
e33129d8 2463}
16a53ecc 2464
1da177e4
LT		 2465/*
2466 * Each stripe/dev can have one or more bion attached.
16a53ecc 2467 * toread/towrite point to the first in a chain.
1da177e4
LT		 2468 * The bi_next chain must be in order.
2469 */
2470static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2471{
2472 struct bio **bip;
d1688a6d 2473 struct r5conf *conf = sh->raid_conf;
72626685 2474 int firstwrite=0;
1da177e4 2475
cbe47ec5 2476 pr_debug("adding bi b#%llu to stripe s#%llu\n",
1da177e4
LT		 2477 (unsigned long long)bi->bi_sector,
2478 (unsigned long long)sh->sector);
2479
b17459c0
SL		 2480 /*
2481 * If several bio share a stripe. The bio bi_phys_segments acts as a
2482 * reference count to avoid race. The reference count should already be
2483 * increased before this function is called (for example, in
2484 * make_request()), so other bio sharing this stripe will not free the
2485 * stripe. If a stripe is owned by one stripe, the stripe lock will
2486 * protect it.
2487 */
2488 spin_lock_irq(&sh->stripe_lock);
72626685 2489 if (forwrite) {
1da177e4 2490 bip = &sh->dev[dd_idx].towrite;
7eaf7e8e 2491 if (*bip == NULL)
72626685
N		 2492 firstwrite = 1;
2493 } else
1da177e4
LT		 2494 bip = &sh->dev[dd_idx].toread;
2495 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
f73a1c7d 2496 if (bio_end_sector(*bip) > bi->bi_sector)
1da177e4
LT		 2497 goto overlap;
2498 bip = & (*bip)->bi_next;
2499 }
f73a1c7d 2500 if (*bip && (*bip)->bi_sector < bio_end_sector(bi))
1da177e4
LT		 2501 goto overlap;
2502
78bafebd 2503 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1da177e4
LT		 2504 if (*bip)
2505 bi->bi_next = *bip;
2506 *bip = bi;
e7836bd6 2507 raid5_inc_bi_active_stripes(bi);
72626685 2508
1da177e4
LT		 2509 if (forwrite) {
2510 /* check if page is covered */
2511 sector_t sector = sh->dev[dd_idx].sector;
2512 for (bi=sh->dev[dd_idx].towrite;
2513 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2514 bi && bi->bi_sector <= sector;
2515 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
f73a1c7d
KO		 2516 if (bio_end_sector(bi) >= sector)
2517 sector = bio_end_sector(bi);
1da177e4
LT		 2518 }
2519 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2520 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2521 }
cbe47ec5
N		 2522
2523 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2524 (unsigned long long)(*bip)->bi_sector,
2525 (unsigned long long)sh->sector, dd_idx);
b97390ae 2526 spin_unlock_irq(&sh->stripe_lock);
cbe47ec5
N		 2527
2528 if (conf->mddev->bitmap && firstwrite) {
2529 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2530 STRIPE_SECTORS, 0);
2531 sh->bm_seq = conf->seq_flush+1;
2532 set_bit(STRIPE_BIT_DELAY, &sh->state);
2533 }
1da177e4
LT		 2534 return 1;
2535
2536 overlap:
2537 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
b17459c0 2538 spin_unlock_irq(&sh->stripe_lock);
1da177e4
LT		 2539 return 0;
2540}
2541
d1688a6d 2542static void end_reshape(struct r5conf *conf);
29269553 2543
d1688a6d 2544static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
911d4ee8 2545 struct stripe_head *sh)
ccfcc3c1 2546{
784052ec 2547 int sectors_per_chunk =
09c9e5fa 2548 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
911d4ee8 2549 int dd_idx;
2d2063ce 2550 int chunk_offset = sector_div(stripe, sectors_per_chunk);
112bf897 2551 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2d2063ce 2552
112bf897
N		 2553 raid5_compute_sector(conf,
2554 stripe * (disks - conf->max_degraded)
b875e531 2555 *sectors_per_chunk + chunk_offset,
112bf897 2556 previous,
911d4ee8 2557 &dd_idx, sh);
ccfcc3c1
N		 2558}
2559
a4456856 2560static void
d1688a6d 2561handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
a4456856
DW		 2562 struct stripe_head_state *s, int disks,
2563 struct bio **return_bi)
2564{
2565 int i;
2566 for (i = disks; i--; ) {
2567 struct bio *bi;
2568 int bitmap_end = 0;
2569
2570 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3cb03002 2571 struct md_rdev *rdev;
a4456856
DW		 2572 rcu_read_lock();
2573 rdev = rcu_dereference(conf->disks[i].rdev);
2574 if (rdev && test_bit(In_sync, &rdev->flags))
7f0da59b
N		 2575 atomic_inc(&rdev->nr_pending);
2576 else
2577 rdev = NULL;
a4456856 2578 rcu_read_unlock();
7f0da59b
N		 2579 if (rdev) {
2580 if (!rdev_set_badblocks(
2581 rdev,
2582 sh->sector,
2583 STRIPE_SECTORS, 0))
2584 md_error(conf->mddev, rdev);
2585 rdev_dec_pending(rdev, conf->mddev);
2586 }
a4456856 2587 }
b17459c0 2588 spin_lock_irq(&sh->stripe_lock);
a4456856
DW		 2589 /* fail all writes first */
2590 bi = sh->dev[i].towrite;
2591 sh->dev[i].towrite = NULL;
b17459c0 2592 spin_unlock_irq(&sh->stripe_lock);
1ed850f3 2593 if (bi)
a4456856 2594 bitmap_end = 1;
a4456856
DW		 2595
2596 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2597 wake_up(&conf->wait_for_overlap);
2598
2599 while (bi && bi->bi_sector <
2600 sh->dev[i].sector + STRIPE_SECTORS) {
2601 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2602 clear_bit(BIO_UPTODATE, &bi->bi_flags);
e7836bd6 2603 if (!raid5_dec_bi_active_stripes(bi)) {
a4456856
DW		 2604 md_write_end(conf->mddev);
2605 bi->bi_next = *return_bi;
2606 *return_bi = bi;
2607 }
2608 bi = nextbi;
2609 }
7eaf7e8e
SL		 2610 if (bitmap_end)
2611 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2612 STRIPE_SECTORS, 0, 0);
2613 bitmap_end = 0;
a4456856
DW		 2614 /* and fail all 'written' */
2615 bi = sh->dev[i].written;
2616 sh->dev[i].written = NULL;
2617 if (bi) bitmap_end = 1;
2618 while (bi && bi->bi_sector <
2619 sh->dev[i].sector + STRIPE_SECTORS) {
2620 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2621 clear_bit(BIO_UPTODATE, &bi->bi_flags);
e7836bd6 2622 if (!raid5_dec_bi_active_stripes(bi)) {
a4456856
DW		 2623 md_write_end(conf->mddev);
2624 bi->bi_next = *return_bi;
2625 *return_bi = bi;
2626 }
2627 bi = bi2;
2628 }
2629
b5e98d65
DW		 2630 /* fail any reads if this device is non-operational and
2631 * the data has not reached the cache yet.
2632 */
2633 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2634 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2635 test_bit(R5_ReadError, &sh->dev[i].flags))) {
143c4d05 2636 spin_lock_irq(&sh->stripe_lock);
a4456856
DW		 2637 bi = sh->dev[i].toread;
2638 sh->dev[i].toread = NULL;
143c4d05 2639 spin_unlock_irq(&sh->stripe_lock);
a4456856
DW		 2640 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2641 wake_up(&conf->wait_for_overlap);
a4456856
DW		 2642 while (bi && bi->bi_sector <
2643 sh->dev[i].sector + STRIPE_SECTORS) {
2644 struct bio *nextbi =
2645 r5_next_bio(bi, sh->dev[i].sector);
2646 clear_bit(BIO_UPTODATE, &bi->bi_flags);
e7836bd6 2647 if (!raid5_dec_bi_active_stripes(bi)) {
a4456856
DW		 2648 bi->bi_next = *return_bi;
2649 *return_bi = bi;
2650 }
2651 bi = nextbi;
2652 }
2653 }
a4456856
DW		 2654 if (bitmap_end)
2655 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2656 STRIPE_SECTORS, 0, 0);
8cfa7b0f
N		 2657 /* If we were in the middle of a write the parity block might
2658 * still be locked - so just clear all R5_LOCKED flags
2659 */
2660 clear_bit(R5_LOCKED, &sh->dev[i].flags);
a4456856
DW		 2661 }
2662
8b3e6cdc
DW		 2663 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2664 if (atomic_dec_and_test(&conf->pending_full_writes))
2665 md_wakeup_thread(conf->mddev->thread);
a4456856
DW		 2666}
2667
7f0da59b 2668static void
d1688a6d 2669handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
7f0da59b
N		 2670 struct stripe_head_state *s)
2671{
2672 int abort = 0;
2673 int i;
2674
7f0da59b 2675 clear_bit(STRIPE_SYNCING, &sh->state);
f8dfcffd
N		 2676 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2677 wake_up(&conf->wait_for_overlap);
7f0da59b 2678 s->syncing = 0;
9a3e1101 2679 s->replacing = 0;
7f0da59b 2680 /* There is nothing more to do for sync/check/repair.
18b9837e
N		 2681 * Don't even need to abort as that is handled elsewhere
2682 * if needed, and not always wanted e.g. if there is a known
2683 * bad block here.
9a3e1101 2684 * For recover/replace we need to record a bad block on all
7f0da59b
N		 2685 * non-sync devices, or abort the recovery
2686 */
18b9837e
N		 2687 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2688 /* During recovery devices cannot be removed, so
2689 * locking and refcounting of rdevs is not needed
2690 */
2691 for (i = 0; i < conf->raid_disks; i++) {
2692 struct md_rdev *rdev = conf->disks[i].rdev;
2693 if (rdev
2694 && !test_bit(Faulty, &rdev->flags)
2695 && !test_bit(In_sync, &rdev->flags)
2696 && !rdev_set_badblocks(rdev, sh->sector,
2697 STRIPE_SECTORS, 0))
2698 abort = 1;
2699 rdev = conf->disks[i].replacement;
2700 if (rdev
2701 && !test_bit(Faulty, &rdev->flags)
2702 && !test_bit(In_sync, &rdev->flags)
2703 && !rdev_set_badblocks(rdev, sh->sector,
2704 STRIPE_SECTORS, 0))
2705 abort = 1;
2706 }
2707 if (abort)
2708 conf->recovery_disabled =
2709 conf->mddev->recovery_disabled;
7f0da59b 2710 }
18b9837e 2711 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
7f0da59b
N		 2712}
2713
9a3e1101
N		 2714static int want_replace(struct stripe_head *sh, int disk_idx)
2715{
2716 struct md_rdev *rdev;
2717 int rv = 0;
2718 /* Doing recovery so rcu locking not required */
2719 rdev = sh->raid_conf->disks[disk_idx].replacement;
2720 if (rdev
2721 && !test_bit(Faulty, &rdev->flags)
2722 && !test_bit(In_sync, &rdev->flags)
2723 && (rdev->recovery_offset <= sh->sector
2724 || rdev->mddev->recovery_cp <= sh->sector))
2725 rv = 1;
2726
2727 return rv;
2728}
2729
93b3dbce 2730/* fetch_block - checks the given member device to see if its data needs
1fe797e6
DW		 2731 * to be read or computed to satisfy a request.
2732 *
2733 * Returns 1 when no more member devices need to be checked, otherwise returns
93b3dbce 2734 * 0 to tell the loop in handle_stripe_fill to continue
f38e1219 2735 */
93b3dbce
N		 2736static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2737 int disk_idx, int disks)
a4456856 2738{
5599becc 2739 struct r5dev *dev = &sh->dev[disk_idx];
f2b3b44d
N		 2740 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2741 &sh->dev[s->failed_num[1]] };
5599becc 2742
93b3dbce 2743 /* is the data in this block needed, and can we get it? */
5599becc
YT		 2744 if (!test_bit(R5_LOCKED, &dev->flags) &&
2745 !test_bit(R5_UPTODATE, &dev->flags) &&
2746 (dev->toread ||
2747 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2748 s->syncing || s->expanding ||
9a3e1101 2749 (s->replacing && want_replace(sh, disk_idx)) ||
5d35e09c
N		 2750 (s->failed >= 1 && fdev[0]->toread) ||
2751 (s->failed >= 2 && fdev[1]->toread) ||
93b3dbce
N		 2752 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2753 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2754 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
5599becc
YT		 2755 /* we would like to get this block, possibly by computing it,
2756 * otherwise read it if the backing disk is insync
2757 */
2758 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2759 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2760 if ((s->uptodate == disks - 1) &&
f2b3b44d
N		 2761 (s->failed && (disk_idx == s->failed_num[0] ||
2762 disk_idx == s->failed_num[1]))) {
5599becc
YT		 2763 /* have disk failed, and we're requested to fetch it;
2764 * do compute it
a4456856 2765 */
5599becc
YT		 2766 pr_debug("Computing stripe %llu block %d\n",
2767 (unsigned long long)sh->sector, disk_idx);
2768 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2769 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2770 set_bit(R5_Wantcompute, &dev->flags);
2771 sh->ops.target = disk_idx;
2772 sh->ops.target2 = -1; /* no 2nd target */
2773 s->req_compute = 1;
93b3dbce
N		 2774 /* Careful: from this point on 'uptodate' is in the eye
2775 * of raid_run_ops which services 'compute' operations
2776 * before writes. R5_Wantcompute flags a block that will
2777 * be R5_UPTODATE by the time it is needed for a
2778 * subsequent operation.
2779 */
5599becc
YT		 2780 s->uptodate++;
2781 return 1;
2782 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2783 /* Computing 2-failure is *very* expensive; only
2784 * do it if failed >= 2
2785 */
2786 int other;
2787 for (other = disks; other--; ) {
2788 if (other == disk_idx)
2789 continue;
2790 if (!test_bit(R5_UPTODATE,
2791 &sh->dev[other].flags))
2792 break;
a4456856 2793 }
5599becc
YT		 2794 BUG_ON(other < 0);
2795 pr_debug("Computing stripe %llu blocks %d,%d\n",
2796 (unsigned long long)sh->sector,
2797 disk_idx, other);
2798 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2799 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2800 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2801 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2802 sh->ops.target = disk_idx;
2803 sh->ops.target2 = other;
2804 s->uptodate += 2;
2805 s->req_compute = 1;
2806 return 1;
2807 } else if (test_bit(R5_Insync, &dev->flags)) {
2808 set_bit(R5_LOCKED, &dev->flags);
2809 set_bit(R5_Wantread, &dev->flags);
2810 s->locked++;
2811 pr_debug("Reading block %d (sync=%d)\n",
2812 disk_idx, s->syncing);
a4456856
DW		 2813 }
2814 }
5599becc
YT		 2815
2816 return 0;
2817}
2818
2819/**
93b3dbce 2820 * handle_stripe_fill - read or compute data to satisfy pending requests.
5599becc 2821 */
93b3dbce
N		 2822static void handle_stripe_fill(struct stripe_head *sh,
2823 struct stripe_head_state *s,
2824 int disks)
5599becc
YT		 2825{
2826 int i;
2827
2828 /* look for blocks to read/compute, skip this if a compute
2829 * is already in flight, or if the stripe contents are in the
2830 * midst of changing due to a write
2831 */
2832 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2833 !sh->reconstruct_state)
2834 for (i = disks; i--; )
93b3dbce 2835 if (fetch_block(sh, s, i, disks))
5599becc 2836 break;
a4456856
DW		 2837 set_bit(STRIPE_HANDLE, &sh->state);
2838}
2839
2840
1fe797e6 2841/* handle_stripe_clean_event
a4456856
DW		 2842 * any written block on an uptodate or failed drive can be returned.
2843 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2844 * never LOCKED, so we don't need to test 'failed' directly.
2845 */
d1688a6d 2846static void handle_stripe_clean_event(struct r5conf *conf,
a4456856
DW		 2847 struct stripe_head *sh, int disks, struct bio **return_bi)
2848{
2849 int i;
2850 struct r5dev *dev;
f8dfcffd 2851 int discard_pending = 0;
a4456856
DW		 2852
2853 for (i = disks; i--; )
2854 if (sh->dev[i].written) {
2855 dev = &sh->dev[i];
2856 if (!test_bit(R5_LOCKED, &dev->flags) &&
9e444768 2857 (test_bit(R5_UPTODATE, &dev->flags) ||
ca64cae9 2858 test_bit(R5_Discard, &dev->flags))) {
a4456856
DW		 2859 /* We can return any write requests */
2860 struct bio *wbi, *wbi2;
45b4233c 2861 pr_debug("Return write for disc %d\n", i);
ca64cae9
N		 2862 if (test_and_clear_bit(R5_Discard, &dev->flags))
2863 clear_bit(R5_UPTODATE, &dev->flags);
a4456856
DW		 2864 wbi = dev->written;
2865 dev->written = NULL;
2866 while (wbi && wbi->bi_sector <
2867 dev->sector + STRIPE_SECTORS) {
2868 wbi2 = r5_next_bio(wbi, dev->sector);
e7836bd6 2869 if (!raid5_dec_bi_active_stripes(wbi)) {
a4456856
DW		 2870 md_write_end(conf->mddev);
2871 wbi->bi_next = *return_bi;
2872 *return_bi = wbi;
2873 }
2874 wbi = wbi2;
2875 }
7eaf7e8e
SL		 2876 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2877 STRIPE_SECTORS,
a4456856 2878 !test_bit(STRIPE_DEGRADED, &sh->state),
7eaf7e8e 2879 0);
f8dfcffd
N		 2880 } else if (test_bit(R5_Discard, &dev->flags))
2881 discard_pending = 1;
2882 }
2883 if (!discard_pending &&
2884 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
2885 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2886 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2887 if (sh->qd_idx >= 0) {
2888 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2889 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
2890 }
2891 /* now that discard is done we can proceed with any sync */
2892 clear_bit(STRIPE_DISCARD, &sh->state);
2893 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2894 set_bit(STRIPE_HANDLE, &sh->state);
2895
2896 }
8b3e6cdc
DW		 2897
2898 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2899 if (atomic_dec_and_test(&conf->pending_full_writes))
2900 md_wakeup_thread(conf->mddev->thread);
a4456856
DW		 2901}
2902
d1688a6d 2903static void handle_stripe_dirtying(struct r5conf *conf,
c8ac1803
N		 2904 struct stripe_head *sh,
2905 struct stripe_head_state *s,
2906 int disks)
a4456856
DW		 2907{
2908 int rmw = 0, rcw = 0, i;
a7854487
AL		 2909 sector_t recovery_cp = conf->mddev->recovery_cp;
2910
2911 /* RAID6 requires 'rcw' in current implementation.
2912 * Otherwise, check whether resync is now happening or should start.
2913 * If yes, then the array is dirty (after unclean shutdown or
2914 * initial creation), so parity in some stripes might be inconsistent.
2915 * In this case, we need to always do reconstruct-write, to ensure
2916 * that in case of drive failure or read-error correction, we
2917 * generate correct data from the parity.
2918 */
2919 if (conf->max_degraded == 2 ||
2920 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2921 /* Calculate the real rcw later - for now make it
c8ac1803
N		 2922 * look like rcw is cheaper
2923 */
2924 rcw = 1; rmw = 2;
a7854487
AL		 2925 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2926 conf->max_degraded, (unsigned long long)recovery_cp,
2927 (unsigned long long)sh->sector);
c8ac1803 2928 } else for (i = disks; i--; ) {
a4456856
DW		 2929 /* would I have to read this buffer for read_modify_write */
2930 struct r5dev *dev = &sh->dev[i];
2931 if ((dev->towrite || i == sh->pd_idx) &&
2932 !test_bit(R5_LOCKED, &dev->flags) &&
f38e1219
DW		 2933 !(test_bit(R5_UPTODATE, &dev->flags) ||
2934 test_bit(R5_Wantcompute, &dev->flags))) {
a4456856
DW		 2935 if (test_bit(R5_Insync, &dev->flags))
2936 rmw++;
2937 else
2938 rmw += 2*disks; /* cannot read it */
2939 }
2940 /* Would I have to read this buffer for reconstruct_write */
2941 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2942 !test_bit(R5_LOCKED, &dev->flags) &&
f38e1219
DW		 2943 !(test_bit(R5_UPTODATE, &dev->flags) ||
2944 test_bit(R5_Wantcompute, &dev->flags))) {
2945 if (test_bit(R5_Insync, &dev->flags)) rcw++;
a4456856
DW		 2946 else
2947 rcw += 2*disks;
2948 }
2949 }
45b4233c 2950 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
a4456856
DW		 2951 (unsigned long long)sh->sector, rmw, rcw);
2952 set_bit(STRIPE_HANDLE, &sh->state);
a9add5d9 2953 if (rmw < rcw && rmw > 0) {
a4456856 2954 /* prefer read-modify-write, but need to get some data */
e3620a3a
JB		 2955 if (conf->mddev->queue)
2956 blk_add_trace_msg(conf->mddev->queue,
2957 "raid5 rmw %llu %d",
2958 (unsigned long long)sh->sector, rmw);
a4456856
DW		 2959 for (i = disks; i--; ) {
2960 struct r5dev *dev = &sh->dev[i];
2961 if ((dev->towrite || i == sh->pd_idx) &&
2962 !test_bit(R5_LOCKED, &dev->flags) &&
f38e1219
DW		 2963 !(test_bit(R5_UPTODATE, &dev->flags) ||
2964 test_bit(R5_Wantcompute, &dev->flags)) &&
a4456856
DW		 2965 test_bit(R5_Insync, &dev->flags)) {
2966 if (
2967 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
45b4233c 2968 pr_debug("Read_old block "
a9add5d9 2969 "%d for r-m-w\n", i);
a4456856
DW		 2970 set_bit(R5_LOCKED, &dev->flags);
2971 set_bit(R5_Wantread, &dev->flags);
2972 s->locked++;
2973 } else {
2974 set_bit(STRIPE_DELAYED, &sh->state);
2975 set_bit(STRIPE_HANDLE, &sh->state);
2976 }
2977 }
2978 }
a9add5d9 2979 }
c8ac1803 2980 if (rcw <= rmw && rcw > 0) {
a4456856 2981 /* want reconstruct write, but need to get some data */
a9add5d9 2982 int qread =0;
c8ac1803 2983 rcw = 0;
a4456856
DW		 2984 for (i = disks; i--; ) {
2985 struct r5dev *dev = &sh->dev[i];
2986 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
c8ac1803 2987 i != sh->pd_idx && i != sh->qd_idx &&
a4456856 2988 !test_bit(R5_LOCKED, &dev->flags) &&
f38e1219 2989 !(test_bit(R5_UPTODATE, &dev->flags) ||
c8ac1803
N		 2990 test_bit(R5_Wantcompute, &dev->flags))) {
2991 rcw++;
2992 if (!test_bit(R5_Insync, &dev->flags))
2993 continue; /* it's a failed drive */
a4456856
DW		 2994 if (
2995 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
45b4233c 2996 pr_debug("Read_old block "
a4456856
DW		 2997 "%d for Reconstruct\n", i);
2998 set_bit(R5_LOCKED, &dev->flags);
2999 set_bit(R5_Wantread, &dev->flags);
3000 s->locked++;
a9add5d9 3001 qread++;
a4456856
DW		 3002 } else {
3003 set_bit(STRIPE_DELAYED, &sh->state);
3004 set_bit(STRIPE_HANDLE, &sh->state);
3005 }
3006 }
3007 }
e3620a3a 3008 if (rcw && conf->mddev->queue)
a9add5d9
N		 3009 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3010 (unsigned long long)sh->sector,
3011 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
c8ac1803 3012 }
a4456856
DW		 3013 /* now if nothing is locked, and if we have enough data,
3014 * we can start a write request
3015 */
f38e1219
DW		 3016 /* since handle_stripe can be called at any time we need to handle the
3017 * case where a compute block operation has been submitted and then a
ac6b53b6
DW		 3018 * subsequent call wants to start a write request. raid_run_ops only
3019 * handles the case where compute block and reconstruct are requested
f38e1219
DW		 3020 * simultaneously. If this is not the case then new writes need to be
3021 * held off until the compute completes.
3022 */
976ea8d4
DW		 3023 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3024 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3025 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
c0f7bddb 3026 schedule_reconstruction(sh, s, rcw == 0, 0);
a4456856
DW		 3027}
3028
d1688a6d 3029static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
a4456856
DW		 3030 struct stripe_head_state *s, int disks)
3031{
ecc65c9b 3032 struct r5dev *dev = NULL;
bd2ab670 3033
a4456856 3034 set_bit(STRIPE_HANDLE, &sh->state);
e89f8962 3035
ecc65c9b
DW		 3036 switch (sh->check_state) {
3037 case check_state_idle:
3038 /* start a new check operation if there are no failures */
bd2ab670 3039 if (s->failed == 0) {
bd2ab670 3040 BUG_ON(s->uptodate != disks);
ecc65c9b
DW		 3041 sh->check_state = check_state_run;
3042 set_bit(STRIPE_OP_CHECK, &s->ops_request);
bd2ab670 3043 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
bd2ab670 3044 s->uptodate--;
ecc65c9b 3045 break;
bd2ab670 3046 }
f2b3b44d 3047 dev = &sh->dev[s->failed_num[0]];
ecc65c9b
DW		 3048 /* fall through */
3049 case check_state_compute_result:
3050 sh->check_state = check_state_idle;
3051 if (!dev)
3052 dev = &sh->dev[sh->pd_idx];
3053
3054 /* check that a write has not made the stripe insync */
3055 if (test_bit(STRIPE_INSYNC, &sh->state))
3056 break;
c8894419 3057
a4456856 3058 /* either failed parity check, or recovery is happening */
a4456856
DW		 3059 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3060 BUG_ON(s->uptodate != disks);
3061
3062 set_bit(R5_LOCKED, &dev->flags);
ecc65c9b 3063 s->locked++;
a4456856 3064 set_bit(R5_Wantwrite, &dev->flags);
830ea016 3065
a4456856 3066 clear_bit(STRIPE_DEGRADED, &sh->state);
a4456856 3067 set_bit(STRIPE_INSYNC, &sh->state);
ecc65c9b
DW		 3068 break;
3069 case check_state_run:
3070 break; /* we will be called again upon completion */
3071 case check_state_check_result:
3072 sh->check_state = check_state_idle;
3073
3074 /* if a failure occurred during the check operation, leave
3075 * STRIPE_INSYNC not set and let the stripe be handled again
3076 */
3077 if (s->failed)
3078 break;
3079
3080 /* handle a successful check operation, if parity is correct
3081 * we are done. Otherwise update the mismatch count and repair
3082 * parity if !MD_RECOVERY_CHECK
3083 */
ad283ea4 3084 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
ecc65c9b
DW		 3085 /* parity is correct (on disc,
3086 * not in buffer any more)
3087 */
3088 set_bit(STRIPE_INSYNC, &sh->state);
3089 else {
7f7583d4 3090 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
ecc65c9b
DW		 3091 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3092 /* don't try to repair!! */
3093 set_bit(STRIPE_INSYNC, &sh->state);
3094 else {
3095 sh->check_state = check_state_compute_run;
976ea8d4 3096 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
ecc65c9b
DW		 3097 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3098 set_bit(R5_Wantcompute,
3099 &sh->dev[sh->pd_idx].flags);
3100 sh->ops.target = sh->pd_idx;
ac6b53b6 3101 sh->ops.target2 = -1;
ecc65c9b
DW		 3102 s->uptodate++;
3103 }
3104 }
3105 break;
3106 case check_state_compute_run:
3107 break;
3108 default:
3109 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3110 __func__, sh->check_state,
3111 (unsigned long long) sh->sector);
3112 BUG();
a4456856
DW		 3113 }
3114}
3115
3116
d1688a6d 3117static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
36d1c647 3118 struct stripe_head_state *s,
f2b3b44d 3119 int disks)
a4456856 3120{
a4456856 3121 int pd_idx = sh->pd_idx;
34e04e87 3122 int qd_idx = sh->qd_idx;
d82dfee0 3123 struct r5dev *dev;
a4456856
DW		 3124
3125 set_bit(STRIPE_HANDLE, &sh->state);
3126
3127 BUG_ON(s->failed > 2);
d82dfee0 3128
a4456856
DW		 3129 /* Want to check and possibly repair P and Q.
3130 * However there could be one 'failed' device, in which
3131 * case we can only check one of them, possibly using the
3132 * other to generate missing data
3133 */
3134
d82dfee0
DW		 3135 switch (sh->check_state) {
3136 case check_state_idle:
3137 /* start a new check operation if there are < 2 failures */
f2b3b44d 3138 if (s->failed == s->q_failed) {
d82dfee0 3139 /* The only possible failed device holds Q, so it
a4456856
DW		 3140 * makes sense to check P (If anything else were failed,
3141 * we would have used P to recreate it).
3142 */
d82dfee0 3143 sh->check_state = check_state_run;
a4456856 3144 }
f2b3b44d 3145 if (!s->q_failed && s->failed < 2) {
d82dfee0 3146 /* Q is not failed, and we didn't use it to generate
a4456856
DW		 3147 * anything, so it makes sense to check it
3148 */
d82dfee0
DW		 3149 if (sh->check_state == check_state_run)
3150 sh->check_state = check_state_run_pq;
3151 else
3152 sh->check_state = check_state_run_q;
a4456856 3153 }
a4456856 3154
d82dfee0
DW		 3155 /* discard potentially stale zero_sum_result */
3156 sh->ops.zero_sum_result = 0;
a4456856 3157
d82dfee0
DW		 3158 if (sh->check_state == check_state_run) {
3159 /* async_xor_zero_sum destroys the contents of P */
3160 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3161 s->uptodate--;
a4456856 3162 }
d82dfee0
DW		 3163 if (sh->check_state >= check_state_run &&
3164 sh->check_state <= check_state_run_pq) {
3165 /* async_syndrome_zero_sum preserves P and Q, so
3166 * no need to mark them !uptodate here
3167 */
3168 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3169 break;
a4456856
DW		 3170 }
3171
d82dfee0
DW		 3172 /* we have 2-disk failure */
3173 BUG_ON(s->failed != 2);
3174 /* fall through */
3175 case check_state_compute_result:
3176 sh->check_state = check_state_idle;
a4456856 3177
d82dfee0
DW		 3178 /* check that a write has not made the stripe insync */
3179 if (test_bit(STRIPE_INSYNC, &sh->state))
3180 break;
a4456856
DW		 3181
3182 /* now write out any block on a failed drive,
d82dfee0 3183 * or P or Q if they were recomputed
a4456856 3184 */
d82dfee0 3185 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
a4456856 3186 if (s->failed == 2) {
f2b3b44d 3187 dev = &sh->dev[s->failed_num[1]];
a4456856
DW		 3188 s->locked++;
3189 set_bit(R5_LOCKED, &dev->flags);
3190 set_bit(R5_Wantwrite, &dev->flags);
3191 }
3192 if (s->failed >= 1) {
f2b3b44d 3193 dev = &sh->dev[s->failed_num[0]];
a4456856
DW		 3194 s->locked++;
3195 set_bit(R5_LOCKED, &dev->flags);
3196 set_bit(R5_Wantwrite, &dev->flags);
3197 }
d82dfee0 3198 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
a4456856
DW		 3199 dev = &sh->dev[pd_idx];
3200 s->locked++;
3201 set_bit(R5_LOCKED, &dev->flags);
3202 set_bit(R5_Wantwrite, &dev->flags);
3203 }
d82dfee0 3204 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
a4456856
DW		 3205 dev = &sh->dev[qd_idx];
3206 s->locked++;
3207 set_bit(R5_LOCKED, &dev->flags);
3208 set_bit(R5_Wantwrite, &dev->flags);
3209 }
3210 clear_bit(STRIPE_DEGRADED, &sh->state);
3211
3212 set_bit(STRIPE_INSYNC, &sh->state);
d82dfee0
DW		 3213 break;
3214 case check_state_run:
3215 case check_state_run_q:
3216 case check_state_run_pq:
3217 break; /* we will be called again upon completion */
3218 case check_state_check_result:
3219 sh->check_state = check_state_idle;
3220
3221 /* handle a successful check operation, if parity is correct
3222 * we are done. Otherwise update the mismatch count and repair
3223 * parity if !MD_RECOVERY_CHECK
3224 */
3225 if (sh->ops.zero_sum_result == 0) {
3226 /* both parities are correct */
3227 if (!s->failed)
3228 set_bit(STRIPE_INSYNC, &sh->state);
3229 else {
3230 /* in contrast to the raid5 case we can validate
3231 * parity, but still have a failure to write
3232 * back
3233 */
3234 sh->check_state = check_state_compute_result;
3235 /* Returning at this point means that we may go
3236 * off and bring p and/or q uptodate again so
3237 * we make sure to check zero_sum_result again
3238 * to verify if p or q need writeback
3239 */
3240 }
3241 } else {
7f7583d4 3242 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
d82dfee0
DW		 3243 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3244 /* don't try to repair!! */
3245 set_bit(STRIPE_INSYNC, &sh->state);
3246 else {
3247 int *target = &sh->ops.target;
3248
3249 sh->ops.target = -1;
3250 sh->ops.target2 = -1;
3251 sh->check_state = check_state_compute_run;
3252 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3253 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3254 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3255 set_bit(R5_Wantcompute,
3256 &sh->dev[pd_idx].flags);
3257 *target = pd_idx;
3258 target = &sh->ops.target2;
3259 s->uptodate++;
3260 }
3261 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3262 set_bit(R5_Wantcompute,
3263 &sh->dev[qd_idx].flags);
3264 *target = qd_idx;
3265 s->uptodate++;
3266 }
3267 }
3268 }
3269 break;
3270 case check_state_compute_run:
3271 break;
3272 default:
3273 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3274 __func__, sh->check_state,
3275 (unsigned long long) sh->sector);
3276 BUG();
a4456856
DW		 3277 }
3278}
3279
d1688a6d 3280static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
a4456856
DW		 3281{
3282 int i;
3283
3284 /* We have read all the blocks in this stripe and now we need to
3285 * copy some of them into a target stripe for expand.
3286 */
f0a50d37 3287 struct dma_async_tx_descriptor *tx = NULL;
a4456856
DW		 3288 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3289 for (i = 0; i < sh->disks; i++)
34e04e87 3290 if (i != sh->pd_idx && i != sh->qd_idx) {
911d4ee8 3291 int dd_idx, j;
a4456856 3292 struct stripe_head *sh2;
a08abd8c 3293 struct async_submit_ctl submit;
a4456856 3294
784052ec 3295 sector_t bn = compute_blocknr(sh, i, 1);
911d4ee8
N		 3296 sector_t s = raid5_compute_sector(conf, bn, 0,
3297 &dd_idx, NULL);
a8c906ca 3298 sh2 = get_active_stripe(conf, s, 0, 1, 1);
a4456856
DW		 3299 if (sh2 == NULL)
3300 /* so far only the early blocks of this stripe
3301 * have been requested. When later blocks
3302 * get requested, we will try again
3303 */
3304 continue;
3305 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3306 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3307 /* must have already done this block */
3308 release_stripe(sh2);
3309 continue;
3310 }
f0a50d37
DW		 3311
3312 /* place all the copies on one channel */
a08abd8c 3313 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
f0a50d37 3314 tx = async_memcpy(sh2->dev[dd_idx].page,
88ba2aa5 3315 sh->dev[i].page, 0, 0, STRIPE_SIZE,
a08abd8c 3316 &submit);
f0a50d37 3317
a4456856
DW		 3318 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3319 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3320 for (j = 0; j < conf->raid_disks; j++)
3321 if (j != sh2->pd_idx &&
86c374ba 3322 j != sh2->qd_idx &&
a4456856
DW		 3323 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3324 break;
3325 if (j == conf->raid_disks) {
3326 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3327 set_bit(STRIPE_HANDLE, &sh2->state);
3328 }
3329 release_stripe(sh2);
f0a50d37 3330
a4456856 3331 }
a2e08551 3332 /* done submitting copies, wait for them to complete */
749586b7 3333 async_tx_quiesce(&tx);
a4456856 3334}
1da177e4
LT		 3335
3336/*
3337 * handle_stripe - do things to a stripe.
3338 *
9a3e1101
N		 3339 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3340 * state of various bits to see what needs to be done.
1da177e4 3341 * Possible results:
9a3e1101
N		 3342 * return some read requests which now have data
3343 * return some write requests which are safely on storage
1da177e4
LT		 3344 * schedule a read on some buffers
3345 * schedule a write of some buffers
3346 * return confirmation of parity correctness
3347 *
1da177e4 3348 */
a4456856 3349
acfe726b 3350static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
1da177e4 3351{
d1688a6d 3352 struct r5conf *conf = sh->raid_conf;
f416885e 3353 int disks = sh->disks;
474af965
N		 3354 struct r5dev *dev;
3355 int i;
9a3e1101 3356 int do_recovery = 0;
1da177e4 3357
acfe726b
N		 3358 memset(s, 0, sizeof(*s));
3359
acfe726b
N		 3360 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3361 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3362 s->failed_num[0] = -1;
3363 s->failed_num[1] = -1;
1da177e4 3364
acfe726b 3365 /* Now to look around and see what can be done */
1da177e4 3366 rcu_read_lock();
16a53ecc 3367 for (i=disks; i--; ) {
3cb03002 3368 struct md_rdev *rdev;
31c176ec
N		 3369 sector_t first_bad;
3370 int bad_sectors;
3371 int is_bad = 0;
acfe726b 3372
16a53ecc 3373 dev = &sh->dev[i];
1da177e4 3374
45b4233c 3375 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
9a3e1101
N		 3376 i, dev->flags,
3377 dev->toread, dev->towrite, dev->written);
6c0069c0
YT		 3378 /* maybe we can reply to a read
3379 *
3380 * new wantfill requests are only permitted while
3381 * ops_complete_biofill is guaranteed to be inactive
3382 */
3383 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3384 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3385 set_bit(R5_Wantfill, &dev->flags);
1da177e4 3386
16a53ecc 3387 /* now count some things */
cc94015a
N		 3388 if (test_bit(R5_LOCKED, &dev->flags))
3389 s->locked++;
3390 if (test_bit(R5_UPTODATE, &dev->flags))
3391 s->uptodate++;
2d6e4ecc 3392 if (test_bit(R5_Wantcompute, &dev->flags)) {
cc94015a
N		 3393 s->compute++;
3394 BUG_ON(s->compute > 2);
2d6e4ecc 3395 }
1da177e4 3396
acfe726b 3397 if (test_bit(R5_Wantfill, &dev->flags))
cc94015a 3398 s->to_fill++;
acfe726b 3399 else if (dev->toread)
cc94015a 3400 s->to_read++;
16a53ecc 3401 if (dev->towrite) {
cc94015a 3402 s->to_write++;
16a53ecc 3403 if (!test_bit(R5_OVERWRITE, &dev->flags))
cc94015a 3404 s->non_overwrite++;
16a53ecc 3405 }
a4456856 3406 if (dev->written)
cc94015a 3407 s->written++;
14a75d3e
N		 3408 /* Prefer to use the replacement for reads, but only
3409 * if it is recovered enough and has no bad blocks.
3410 */
3411 rdev = rcu_dereference(conf->disks[i].replacement);
3412 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3413 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3414 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3415 &first_bad, &bad_sectors))
3416 set_bit(R5_ReadRepl, &dev->flags);
3417 else {
9a3e1101
N		 3418 if (rdev)
3419 set_bit(R5_NeedReplace, &dev->flags);
14a75d3e
N		 3420 rdev = rcu_dereference(conf->disks[i].rdev);
3421 clear_bit(R5_ReadRepl, &dev->flags);
3422 }
9283d8c5
N		 3423 if (rdev && test_bit(Faulty, &rdev->flags))
3424 rdev = NULL;
31c176ec
N		 3425 if (rdev) {
3426 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3427 &first_bad, &bad_sectors);
3428 if (s->blocked_rdev == NULL
3429 && (test_bit(Blocked, &rdev->flags)
3430 || is_bad < 0)) {
3431 if (is_bad < 0)
3432 set_bit(BlockedBadBlocks,
3433 &rdev->flags);
3434 s->blocked_rdev = rdev;
3435 atomic_inc(&rdev->nr_pending);
3436 }
6bfe0b49 3437 }
415e72d0
N		 3438 clear_bit(R5_Insync, &dev->flags);
3439 if (!rdev)
3440 /* Not in-sync */;
31c176ec
N		 3441 else if (is_bad) {
3442 /* also not in-sync */
18b9837e
N		 3443 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3444 test_bit(R5_UPTODATE, &dev->flags)) {
31c176ec
N		 3445 /* treat as in-sync, but with a read error
3446 * which we can now try to correct
3447 */
3448 set_bit(R5_Insync, &dev->flags);
3449 set_bit(R5_ReadError, &dev->flags);
3450 }
3451 } else if (test_bit(In_sync, &rdev->flags))
415e72d0 3452 set_bit(R5_Insync, &dev->flags);
30d7a483 3453 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
415e72d0 3454 /* in sync if before recovery_offset */
30d7a483
N		 3455 set_bit(R5_Insync, &dev->flags);
3456 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3457 test_bit(R5_Expanded, &dev->flags))
3458 /* If we've reshaped into here, we assume it is Insync.
3459 * We will shortly update recovery_offset to make
3460 * it official.
3461 */
3462 set_bit(R5_Insync, &dev->flags);
3463
5d8c71f9 3464 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
14a75d3e
N		 3465 /* This flag does not apply to '.replacement'
3466 * only to .rdev, so make sure to check that*/
3467 struct md_rdev *rdev2 = rcu_dereference(
3468 conf->disks[i].rdev);
3469 if (rdev2 == rdev)
3470 clear_bit(R5_Insync, &dev->flags);
3471 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
bc2607f3 3472 s->handle_bad_blocks = 1;
14a75d3e 3473 atomic_inc(&rdev2->nr_pending);
bc2607f3
N		 3474 } else
3475 clear_bit(R5_WriteError, &dev->flags);
3476 }
5d8c71f9 3477 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
14a75d3e
N		 3478 /* This flag does not apply to '.replacement'
3479 * only to .rdev, so make sure to check that*/
3480 struct md_rdev *rdev2 = rcu_dereference(
3481 conf->disks[i].rdev);
3482 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
b84db560 3483 s->handle_bad_blocks = 1;
14a75d3e 3484 atomic_inc(&rdev2->nr_pending);
b84db560
N		 3485 } else
3486 clear_bit(R5_MadeGood, &dev->flags);
3487 }
977df362
N		 3488 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3489 struct md_rdev *rdev2 = rcu_dereference(
3490 conf->disks[i].replacement);
3491 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3492 s->handle_bad_blocks = 1;
3493 atomic_inc(&rdev2->nr_pending);
3494 } else
3495 clear_bit(R5_MadeGoodRepl, &dev->flags);
3496 }
415e72d0 3497 if (!test_bit(R5_Insync, &dev->flags)) {
16a53ecc
N		 3498 /* The ReadError flag will just be confusing now */
3499 clear_bit(R5_ReadError, &dev->flags);
3500 clear_bit(R5_ReWrite, &dev->flags);
1da177e4 3501 }
415e72d0
N		 3502 if (test_bit(R5_ReadError, &dev->flags))
3503 clear_bit(R5_Insync, &dev->flags);
3504 if (!test_bit(R5_Insync, &dev->flags)) {
cc94015a
N		 3505 if (s->failed < 2)
3506 s->failed_num[s->failed] = i;
3507 s->failed++;
9a3e1101
N		 3508 if (rdev && !test_bit(Faulty, &rdev->flags))
3509 do_recovery = 1;
415e72d0 3510 }
1da177e4 3511 }
9a3e1101
N		 3512 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3513 /* If there is a failed device being replaced,
3514 * we must be recovering.
3515 * else if we are after recovery_cp, we must be syncing
c6d2e084 3516 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
9a3e1101
N		 3517 * else we can only be replacing
3518 * sync and recovery both need to read all devices, and so
3519 * use the same flag.
3520 */
3521 if (do_recovery ||
c6d2e084 3522 sh->sector >= conf->mddev->recovery_cp ||
3523 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
9a3e1101
N		 3524 s->syncing = 1;
3525 else
3526 s->replacing = 1;
3527 }
1da177e4 3528 rcu_read_unlock();
cc94015a
N		 3529}
3530
3531static void handle_stripe(struct stripe_head *sh)
3532{
3533 struct stripe_head_state s;
d1688a6d 3534 struct r5conf *conf = sh->raid_conf;
3687c061 3535 int i;
84789554
N		 3536 int prexor;
3537 int disks = sh->disks;
474af965 3538 struct r5dev *pdev, *qdev;
cc94015a
N		 3539
3540 clear_bit(STRIPE_HANDLE, &sh->state);
257a4b42 3541 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
cc94015a
N		 3542 /* already being handled, ensure it gets handled
3543 * again when current action finishes */
3544 set_bit(STRIPE_HANDLE, &sh->state);
3545 return;
3546 }
3547
f8dfcffd
N		 3548 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3549 spin_lock(&sh->stripe_lock);
3550 /* Cannot process 'sync' concurrently with 'discard' */
3551 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3552 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3553 set_bit(STRIPE_SYNCING, &sh->state);
3554 clear_bit(STRIPE_INSYNC, &sh->state);
f94c0b66 3555 clear_bit(STRIPE_REPLACED, &sh->state);
f8dfcffd
N		 3556 }
3557 spin_unlock(&sh->stripe_lock);
cc94015a
N		 3558 }
3559 clear_bit(STRIPE_DELAYED, &sh->state);
3560
3561 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3562 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3563 (unsigned long long)sh->sector, sh->state,
3564 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3565 sh->check_state, sh->reconstruct_state);
3687c061 3566
acfe726b 3567 analyse_stripe(sh, &s);
c5a31000 3568
bc2607f3
N		 3569 if (s.handle_bad_blocks) {
3570 set_bit(STRIPE_HANDLE, &sh->state);
3571 goto finish;
3572 }
3573
474af965
N		 3574 if (unlikely(s.blocked_rdev)) {
3575 if (s.syncing || s.expanding || s.expanded ||
9a3e1101 3576 s.replacing || s.to_write || s.written) {
474af965
N		 3577 set_bit(STRIPE_HANDLE, &sh->state);
3578 goto finish;
3579 }
3580 /* There is nothing for the blocked_rdev to block */
3581 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3582 s.blocked_rdev = NULL;
3583 }
3584
3585 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3586 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3587 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3588 }
3589
3590 pr_debug("locked=%d uptodate=%d to_read=%d"
3591 " to_write=%d failed=%d failed_num=%d,%d\n",
3592 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3593 s.failed_num[0], s.failed_num[1]);
3594 /* check if the array has lost more than max_degraded devices and,
3595 * if so, some requests might need to be failed.
3596 */
9a3f530f
N		 3597 if (s.failed > conf->max_degraded) {
3598 sh->check_state = 0;
3599 sh->reconstruct_state = 0;
3600 if (s.to_read+s.to_write+s.written)
3601 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
9a3e1101 3602 if (s.syncing + s.replacing)
9a3f530f
N		 3603 handle_failed_sync(conf, sh, &s);
3604 }
474af965 3605
84789554
N		 3606 /* Now we check to see if any write operations have recently
3607 * completed
3608 */
3609 prexor = 0;
3610 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3611 prexor = 1;
3612 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3613 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3614 sh->reconstruct_state = reconstruct_state_idle;
3615
3616 /* All the 'written' buffers and the parity block are ready to
3617 * be written back to disk
3618 */
9e444768
SL		 3619 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3620 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
84789554 3621 BUG_ON(sh->qd_idx >= 0 &&
9e444768
SL		 3622 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3623 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
84789554
N		 3624 for (i = disks; i--; ) {
3625 struct r5dev *dev = &sh->dev[i];
3626 if (test_bit(R5_LOCKED, &dev->flags) &&
3627 (i == sh->pd_idx || i == sh->qd_idx ||
3628 dev->written)) {
3629 pr_debug("Writing block %d\n", i);
3630 set_bit(R5_Wantwrite, &dev->flags);
3631 if (prexor)
3632 continue;
3633 if (!test_bit(R5_Insync, &dev->flags) ||
3634 ((i == sh->pd_idx || i == sh->qd_idx) &&
3635 s.failed == 0))
3636 set_bit(STRIPE_INSYNC, &sh->state);
3637 }
3638 }
3639 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3640 s.dec_preread_active = 1;
3641 }
3642
ef5b7c69
N		 3643 /*
3644 * might be able to return some write requests if the parity blocks
3645 * are safe, or on a failed drive
3646 */
3647 pdev = &sh->dev[sh->pd_idx];
3648 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3649 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3650 qdev = &sh->dev[sh->qd_idx];
3651 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3652 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3653 || conf->level < 6;
3654
3655 if (s.written &&
3656 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3657 && !test_bit(R5_LOCKED, &pdev->flags)
3658 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3659 test_bit(R5_Discard, &pdev->flags))))) &&
3660 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3661 && !test_bit(R5_LOCKED, &qdev->flags)
3662 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3663 test_bit(R5_Discard, &qdev->flags))))))
3664 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3665
3666 /* Now we might consider reading some blocks, either to check/generate
3667 * parity, or to satisfy requests
3668 * or to load a block that is being partially written.
3669 */
3670 if (s.to_read || s.non_overwrite
3671 || (conf->level == 6 && s.to_write && s.failed)
3672 || (s.syncing && (s.uptodate + s.compute < disks))
3673 || s.replacing
3674 || s.expanding)
3675 handle_stripe_fill(sh, &s, disks);
3676
84789554
N		 3677 /* Now to consider new write requests and what else, if anything
3678 * should be read. We do not handle new writes when:
3679 * 1/ A 'write' operation (copy+xor) is already in flight.
3680 * 2/ A 'check' operation is in flight, as it may clobber the parity
3681 * block.
3682 */
3683 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3684 handle_stripe_dirtying(conf, sh, &s, disks);
3685
3686 /* maybe we need to check and possibly fix the parity for this stripe
3687 * Any reads will already have been scheduled, so we just see if enough
3688 * data is available. The parity check is held off while parity
3689 * dependent operations are in flight.
3690 */
3691 if (sh->check_state ||
3692 (s.syncing && s.locked == 0 &&
3693 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3694 !test_bit(STRIPE_INSYNC, &sh->state))) {
3695 if (conf->level == 6)
3696 handle_parity_checks6(conf, sh, &s, disks);
3697 else
3698 handle_parity_checks5(conf, sh, &s, disks);
3699 }
c5a31000 3700
f94c0b66
N		 3701 if ((s.replacing || s.syncing) && s.locked == 0
3702 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3703 && !test_bit(STRIPE_REPLACED, &sh->state)) {
9a3e1101
N		 3704 /* Write out to replacement devices where possible */
3705 for (i = 0; i < conf->raid_disks; i++)
f94c0b66
N		 3706 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3707 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
9a3e1101
N		 3708 set_bit(R5_WantReplace, &sh->dev[i].flags);
3709 set_bit(R5_LOCKED, &sh->dev[i].flags);
3710 s.locked++;
3711 }
f94c0b66
N		 3712 if (s.replacing)
3713 set_bit(STRIPE_INSYNC, &sh->state);
3714 set_bit(STRIPE_REPLACED, &sh->state);
9a3e1101
N		 3715 }
3716 if ((s.syncing || s.replacing) && s.locked == 0 &&
f94c0b66 3717 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
9a3e1101 3718 test_bit(STRIPE_INSYNC, &sh->state)) {
c5a31000
N		 3719 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3720 clear_bit(STRIPE_SYNCING, &sh->state);
f8dfcffd
N		 3721 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3722 wake_up(&conf->wait_for_overlap);
c5a31000
N		 3723 }
3724
3725 /* If the failed drives are just a ReadError, then we might need
3726 * to progress the repair/check process
3727 */
3728 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3729 for (i = 0; i < s.failed; i++) {
3730 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3731 if (test_bit(R5_ReadError, &dev->flags)
3732 && !test_bit(R5_LOCKED, &dev->flags)
3733 && test_bit(R5_UPTODATE, &dev->flags)
3734 ) {
3735 if (!test_bit(R5_ReWrite, &dev->flags)) {
3736 set_bit(R5_Wantwrite, &dev->flags);
3737 set_bit(R5_ReWrite, &dev->flags);
3738 set_bit(R5_LOCKED, &dev->flags);
3739 s.locked++;
3740 } else {
3741 /* let's read it back */
3742 set_bit(R5_Wantread, &dev->flags);
3743 set_bit(R5_LOCKED, &dev->flags);
3744 s.locked++;
3745 }
3746 }
3747 }
3748
3749
3687c061
N		 3750 /* Finish reconstruct operations initiated by the expansion process */
3751 if (sh->reconstruct_state == reconstruct_state_result) {
3752 struct stripe_head *sh_src
3753 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3754 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3755 /* sh cannot be written until sh_src has been read.
3756 * so arrange for sh to be delayed a little
3757 */
3758 set_bit(STRIPE_DELAYED, &sh->state);
3759 set_bit(STRIPE_HANDLE, &sh->state);
3760 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3761 &sh_src->state))
3762 atomic_inc(&conf->preread_active_stripes);
3763 release_stripe(sh_src);
3764 goto finish;
3765 }
3766 if (sh_src)
3767 release_stripe(sh_src);
3768
3769 sh->reconstruct_state = reconstruct_state_idle;
3770 clear_bit(STRIPE_EXPANDING, &sh->state);
3771 for (i = conf->raid_disks; i--; ) {
3772 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3773 set_bit(R5_LOCKED, &sh->dev[i].flags);
3774 s.locked++;
3775 }
3776 }
f416885e 3777
3687c061
N		 3778 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3779 !sh->reconstruct_state) {
3780 /* Need to write out all blocks after computing parity */
3781 sh->disks = conf->raid_disks;
3782 stripe_set_idx(sh->sector, conf, 0, sh);
3783 schedule_reconstruction(sh, &s, 1, 1);
3784 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3785 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3786 atomic_dec(&conf->reshape_stripes);
3787 wake_up(&conf->wait_for_overlap);
3788 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3789 }
3790
3791 if (s.expanding && s.locked == 0 &&
3792 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3793 handle_stripe_expansion(conf, sh);
16a53ecc 3794
3687c061 3795finish:
6bfe0b49 3796 /* wait for this device to become unblocked */
5f066c63
N		 3797 if (unlikely(s.blocked_rdev)) {
3798 if (conf->mddev->external)
3799 md_wait_for_blocked_rdev(s.blocked_rdev,
3800 conf->mddev);
3801 else
3802 /* Internal metadata will immediately
3803 * be written by raid5d, so we don't
3804 * need to wait here.
3805 */
3806 rdev_dec_pending(s.blocked_rdev,
3807 conf->mddev);
3808 }
6bfe0b49 3809
bc2607f3
N		 3810 if (s.handle_bad_blocks)
3811 for (i = disks; i--; ) {
3cb03002 3812 struct md_rdev *rdev;
bc2607f3
N		 3813 struct r5dev *dev = &sh->dev[i];
3814 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3815 /* We own a safe reference to the rdev */
3816 rdev = conf->disks[i].rdev;
3817 if (!rdev_set_badblocks(rdev, sh->sector,
3818 STRIPE_SECTORS, 0))
3819 md_error(conf->mddev, rdev);
3820 rdev_dec_pending(rdev, conf->mddev);
3821 }
b84db560
N		 3822 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3823 rdev = conf->disks[i].rdev;
3824 rdev_clear_badblocks(rdev, sh->sector,
c6563a8c 3825 STRIPE_SECTORS, 0);
b84db560
N		 3826 rdev_dec_pending(rdev, conf->mddev);
3827 }
977df362
N		 3828 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3829 rdev = conf->disks[i].replacement;
dd054fce
N		 3830 if (!rdev)
3831 /* rdev have been moved down */
3832 rdev = conf->disks[i].rdev;
977df362 3833 rdev_clear_badblocks(rdev, sh->sector,
c6563a8c 3834 STRIPE_SECTORS, 0);
977df362
N		 3835 rdev_dec_pending(rdev, conf->mddev);
3836 }
bc2607f3
N		 3837 }
3838
6c0069c0
YT		 3839 if (s.ops_request)
3840 raid_run_ops(sh, s.ops_request);
3841
f0e43bcd 3842 ops_run_io(sh, &s);
16a53ecc 3843
c5709ef6 3844 if (s.dec_preread_active) {
729a1866 3845 /* We delay this until after ops_run_io so that if make_request
e9c7469b 3846 * is waiting on a flush, it won't continue until the writes
729a1866
N		 3847 * have actually been submitted.
3848 */
3849 atomic_dec(&conf->preread_active_stripes);
3850 if (atomic_read(&conf->preread_active_stripes) <
3851 IO_THRESHOLD)
3852 md_wakeup_thread(conf->mddev->thread);
3853 }
3854
c5709ef6 3855 return_io(s.return_bi);
16a53ecc 3856
257a4b42 3857 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
16a53ecc
N		 3858}
3859
d1688a6d 3860static void raid5_activate_delayed(struct r5conf *conf)
16a53ecc
N		 3861{
3862 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3863 while (!list_empty(&conf->delayed_list)) {
3864 struct list_head *l = conf->delayed_list.next;
3865 struct stripe_head *sh;
3866 sh = list_entry(l, struct stripe_head, lru);
3867 list_del_init(l);
3868 clear_bit(STRIPE_DELAYED, &sh->state);
3869 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3870 atomic_inc(&conf->preread_active_stripes);
8b3e6cdc 3871 list_add_tail(&sh->lru, &conf->hold_list);
851c30c9 3872 raid5_wakeup_stripe_thread(sh);
16a53ecc 3873 }
482c0834 3874 }
16a53ecc
N		 3875}
3876
d1688a6d 3877static void activate_bit_delay(struct r5conf *conf)
16a53ecc
N		 3878{
3879 /* device_lock is held */
3880 struct list_head head;
3881 list_add(&head, &conf->bitmap_list);
3882 list_del_init(&conf->bitmap_list);
3883 while (!list_empty(&head)) {
3884 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3885 list_del_init(&sh->lru);
3886 atomic_inc(&sh->count);
3887 __release_stripe(conf, sh);
3888 }
3889}
3890
fd01b88c 3891int md_raid5_congested(struct mddev *mddev, int bits)
f022b2fd 3892{
d1688a6d 3893 struct r5conf *conf = mddev->private;
f022b2fd
N		 3894
3895 /* No difference between reads and writes. Just check
3896 * how busy the stripe_cache is
3897 */
3fa841d7 3898
f022b2fd
N		 3899 if (conf->inactive_blocked)
3900 return 1;
3901 if (conf->quiesce)
3902 return 1;
3903 if (list_empty_careful(&conf->inactive_list))
3904 return 1;
3905
3906 return 0;
3907}
11d8a6e3
N		 3908EXPORT_SYMBOL_GPL(md_raid5_congested);
3909
3910static int raid5_congested(void *data, int bits)
3911{
fd01b88c 3912 struct mddev *mddev = data;
11d8a6e3
N		 3913
3914 return mddev_congested(mddev, bits) ||
3915 md_raid5_congested(mddev, bits);
3916}
f022b2fd 3917
23032a0e
RBJ		 3918/* We want read requests to align with chunks where possible,
3919 * but write requests don't need to.
3920 */
cc371e66
AK		 3921static int raid5_mergeable_bvec(struct request_queue *q,
3922 struct bvec_merge_data *bvm,
3923 struct bio_vec *biovec)
23032a0e 3924{
fd01b88c 3925 struct mddev *mddev = q->queuedata;
cc371e66 3926 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
23032a0e 3927 int max;
9d8f0363 3928 unsigned int chunk_sectors = mddev->chunk_sectors;
cc371e66 3929 unsigned int bio_sectors = bvm->bi_size >> 9;
23032a0e 3930
cc371e66 3931 if ((bvm->bi_rw & 1) == WRITE)
23032a0e
RBJ		 3932 return biovec->bv_len; /* always allow writes to be mergeable */
3933
664e7c41
AN		 3934 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3935 chunk_sectors = mddev->new_chunk_sectors;
23032a0e
RBJ		 3936 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3937 if (max < 0) max = 0;
3938 if (max <= biovec->bv_len && bio_sectors == 0)
3939 return biovec->bv_len;
3940 else
3941 return max;
3942}
3943
f679623f 3944
fd01b88c 3945static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
f679623f
RBJ		 3946{
3947 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
9d8f0363 3948 unsigned int chunk_sectors = mddev->chunk_sectors;
aa8b57aa 3949 unsigned int bio_sectors = bio_sectors(bio);
f679623f 3950
664e7c41
AN		 3951 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3952 chunk_sectors = mddev->new_chunk_sectors;
f679623f
RBJ		 3953 return chunk_sectors >=
3954 ((sector & (chunk_sectors - 1)) + bio_sectors);
3955}
3956
46031f9a
RBJ		 3957/*
3958 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3959 * later sampled by raid5d.
3960 */
d1688a6d 3961static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
46031f9a
RBJ		 3962{
3963 unsigned long flags;
3964
3965 spin_lock_irqsave(&conf->device_lock, flags);
3966
3967 bi->bi_next = conf->retry_read_aligned_list;
3968 conf->retry_read_aligned_list = bi;
3969
3970 spin_unlock_irqrestore(&conf->device_lock, flags);
3971 md_wakeup_thread(conf->mddev->thread);
3972}
3973
3974
d1688a6d 3975static struct bio *remove_bio_from_retry(struct r5conf *conf)
46031f9a
RBJ		 3976{
3977 struct bio *bi;
3978
3979 bi = conf->retry_read_aligned;
3980 if (bi) {
3981 conf->retry_read_aligned = NULL;
3982 return bi;
3983 }
3984 bi = conf->retry_read_aligned_list;
3985 if(bi) {
387bb173 3986 conf->retry_read_aligned_list = bi->bi_next;
46031f9a 3987 bi->bi_next = NULL;
960e739d
JA		 3988 /*
3989 * this sets the active strip count to 1 and the processed
3990 * strip count to zero (upper 8 bits)
3991 */
e7836bd6 3992 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
46031f9a
RBJ		 3993 }
3994
3995 return bi;
3996}
3997
3998
f679623f
RBJ		 3999/*
4000 * The "raid5_align_endio" should check if the read succeeded and if it
4001 * did, call bio_endio on the original bio (having bio_put the new bio
4002 * first).
4003 * If the read failed..
4004 */
6712ecf8 4005static void raid5_align_endio(struct bio *bi, int error)
f679623f
RBJ		 4006{
4007 struct bio* raid_bi = bi->bi_private;
fd01b88c 4008 struct mddev *mddev;
d1688a6d 4009 struct r5conf *conf;
46031f9a 4010 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3cb03002 4011 struct md_rdev *rdev;
46031f9a 4012
f679623f 4013 bio_put(bi);
46031f9a 4014
46031f9a
RBJ		 4015 rdev = (void*)raid_bi->bi_next;
4016 raid_bi->bi_next = NULL;
2b7f2228
N		 4017 mddev = rdev->mddev;
4018 conf = mddev->private;
46031f9a
RBJ		 4019
4020 rdev_dec_pending(rdev, conf->mddev);
4021
4022 if (!error && uptodate) {
0a82a8d1
LT		 4023 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4024 raid_bi, 0);
6712ecf8 4025 bio_endio(raid_bi, 0);
46031f9a
RBJ		 4026 if (atomic_dec_and_test(&conf->active_aligned_reads))
4027 wake_up(&conf->wait_for_stripe);
6712ecf8 4028 return;
46031f9a
RBJ		 4029 }
4030
4031
45b4233c 4032 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
46031f9a
RBJ		 4033
4034 add_bio_to_retry(raid_bi, conf);
f679623f
RBJ		 4035}
4036
387bb173
NB		 4037static int bio_fits_rdev(struct bio *bi)
4038{
165125e1 4039 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
387bb173 4040
aa8b57aa 4041 if (bio_sectors(bi) > queue_max_sectors(q))
387bb173
NB		 4042 return 0;
4043 blk_recount_segments(q, bi);
8a78362c 4044 if (bi->bi_phys_segments > queue_max_segments(q))
387bb173
NB		 4045 return 0;
4046
4047 if (q->merge_bvec_fn)
4048 /* it's too hard to apply the merge_bvec_fn at this stage,
4049 * just just give up
4050 */
4051 return 0;
4052
4053 return 1;
4054}
4055
4056
fd01b88c 4057static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
f679623f 4058{
d1688a6d 4059 struct r5conf *conf = mddev->private;
8553fe7e 4060 int dd_idx;
f679623f 4061 struct bio* align_bi;
3cb03002 4062 struct md_rdev *rdev;
671488cc 4063 sector_t end_sector;
f679623f
RBJ		 4064
4065 if (!in_chunk_boundary(mddev, raid_bio)) {
45b4233c 4066 pr_debug("chunk_aligned_read : non aligned\n");
f679623f
RBJ		 4067 return 0;
4068 }
4069 /*
a167f663 4070 * use bio_clone_mddev to make a copy of the bio
f679623f 4071 */
a167f663 4072 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
f679623f
RBJ		 4073 if (!align_bi)
4074 return 0;
4075 /*
4076 * set bi_end_io to a new function, and set bi_private to the
4077 * original bio.
4078 */
4079 align_bi->bi_end_io = raid5_align_endio;
4080 align_bi->bi_private = raid_bio;
4081 /*
4082 * compute position
4083 */
112bf897
N		 4084 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
4085 0,
911d4ee8 4086 &dd_idx, NULL);
f679623f 4087
f73a1c7d 4088 end_sector = bio_end_sector(align_bi);
f679623f 4089 rcu_read_lock();
671488cc
N		 4090 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4091 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4092 rdev->recovery_offset < end_sector) {
4093 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4094 if (rdev &&
4095 (test_bit(Faulty, &rdev->flags) ||
4096 !(test_bit(In_sync, &rdev->flags) ||
4097 rdev->recovery_offset >= end_sector)))
4098 rdev = NULL;
4099 }
4100 if (rdev) {
31c176ec
N		 4101 sector_t first_bad;
4102 int bad_sectors;
4103
f679623f
RBJ		 4104 atomic_inc(&rdev->nr_pending);
4105 rcu_read_unlock();
46031f9a
RBJ		 4106 raid_bio->bi_next = (void*)rdev;
4107 align_bi->bi_bdev = rdev->bdev;
4108 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
46031f9a 4109
31c176ec 4110 if (!bio_fits_rdev(align_bi) ||
aa8b57aa 4111 is_badblock(rdev, align_bi->bi_sector, bio_sectors(align_bi),
31c176ec
N		 4112 &first_bad, &bad_sectors)) {
4113 /* too big in some way, or has a known bad block */
387bb173
NB		 4114 bio_put(align_bi);
4115 rdev_dec_pending(rdev, mddev);
4116 return 0;
4117 }
4118
6c0544e2 4119 /* No reshape active, so we can trust rdev->data_offset */
4120 align_bi->bi_sector += rdev->data_offset;
4121
46031f9a
RBJ		 4122 spin_lock_irq(&conf->device_lock);
4123 wait_event_lock_irq(conf->wait_for_stripe,
4124 conf->quiesce == 0,
eed8c02e 4125 conf->device_lock);
46031f9a
RBJ		 4126 atomic_inc(&conf->active_aligned_reads);
4127 spin_unlock_irq(&conf->device_lock);
4128
e3620a3a
JB		 4129 if (mddev->gendisk)
4130 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4131 align_bi, disk_devt(mddev->gendisk),
4132 raid_bio->bi_sector);
f679623f
RBJ		 4133 generic_make_request(align_bi);
4134 return 1;
4135 } else {
4136 rcu_read_unlock();
46031f9a 4137 bio_put(align_bi);
f679623f
RBJ		 4138 return 0;
4139 }
4140}
4141
8b3e6cdc
DW		 4142/* __get_priority_stripe - get the next stripe to process
4143 *
4144 * Full stripe writes are allowed to pass preread active stripes up until
4145 * the bypass_threshold is exceeded. In general the bypass_count
4146 * increments when the handle_list is handled before the hold_list; however, it
4147 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4148 * stripe with in flight i/o. The bypass_count will be reset when the
4149 * head of the hold_list has changed, i.e. the head was promoted to the
4150 * handle_list.
4151 */
851c30c9 4152static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
8b3e6cdc 4153{
851c30c9
SL		 4154 struct stripe_head *sh = NULL, *tmp;
4155 struct list_head *handle_list = NULL;
4156
4157 if (conf->worker_cnt_per_group == 0) {
4158 handle_list = &conf->handle_list;
4159 } else if (group != ANY_GROUP) {
4160 handle_list = &conf->worker_groups[group].handle_list;
4161 } else {
4162 int i;
4163 for (i = 0; i < conf->group_cnt; i++) {
4164 handle_list = &conf->worker_groups[i].handle_list;
4165 if (!list_empty(handle_list))
4166 break;
4167 }
4168 }
8b3e6cdc
DW		 4169
4170 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4171 __func__,
851c30c9 4172 list_empty(handle_list) ? "empty" : "busy",
8b3e6cdc
DW		 4173 list_empty(&conf->hold_list) ? "empty" : "busy",
4174 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4175
851c30c9
SL		 4176 if (!list_empty(handle_list)) {
4177 sh = list_entry(handle_list->next, typeof(*sh), lru);
8b3e6cdc
DW		 4178
4179 if (list_empty(&conf->hold_list))
4180 conf->bypass_count = 0;
4181 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4182 if (conf->hold_list.next == conf->last_hold)
4183 conf->bypass_count++;
4184 else {
4185 conf->last_hold = conf->hold_list.next;
4186 conf->bypass_count -= conf->bypass_threshold;
4187 if (conf->bypass_count < 0)
4188 conf->bypass_count = 0;
4189 }
4190 }
4191 } else if (!list_empty(&conf->hold_list) &&
4192 ((conf->bypass_threshold &&
4193 conf->bypass_count > conf->bypass_threshold) ||
4194 atomic_read(&conf->pending_full_writes) == 0)) {
851c30c9
SL		 4195
4196 list_for_each_entry(tmp, &conf->hold_list, lru) {
4197 if (conf->worker_cnt_per_group == 0 ||
4198 group == ANY_GROUP ||
4199 !cpu_online(tmp->cpu) ||
4200 cpu_to_group(tmp->cpu) == group) {
4201 sh = tmp;
4202 break;
4203 }
4204 }
4205
4206 if (sh) {
4207 conf->bypass_count -= conf->bypass_threshold;
4208 if (conf->bypass_count < 0)
4209 conf->bypass_count = 0;
4210 }
4211 }
4212
4213 if (!sh)
8b3e6cdc
DW		 4214 return NULL;
4215
4216 list_del_init(&sh->lru);
4217 atomic_inc(&sh->count);
4218 BUG_ON(atomic_read(&sh->count) != 1);
4219 return sh;
4220}
f679623f 4221
8811b596
SL		 4222struct raid5_plug_cb {
4223 struct blk_plug_cb cb;
4224 struct list_head list;
4225};
4226
4227static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4228{
4229 struct raid5_plug_cb *cb = container_of(
4230 blk_cb, struct raid5_plug_cb, cb);
4231 struct stripe_head *sh;
4232 struct mddev *mddev = cb->cb.data;
4233 struct r5conf *conf = mddev->private;
a9add5d9 4234 int cnt = 0;
8811b596
SL		 4235
4236 if (cb->list.next && !list_empty(&cb->list)) {
4237 spin_lock_irq(&conf->device_lock);
4238 while (!list_empty(&cb->list)) {
4239 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4240 list_del_init(&sh->lru);
4241 /*
4242 * avoid race release_stripe_plug() sees
4243 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4244 * is still in our list
4245 */
4246 smp_mb__before_clear_bit();
4247 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
773ca82f
SL		 4248 /*
4249 * STRIPE_ON_RELEASE_LIST could be set here. In that
4250 * case, the count is always > 1 here
4251 */
8811b596 4252 __release_stripe(conf, sh);
a9add5d9 4253 cnt++;
8811b596
SL		 4254 }
4255 spin_unlock_irq(&conf->device_lock);
4256 }
e3620a3a
JB		 4257 if (mddev->queue)
4258 trace_block_unplug(mddev->queue, cnt, !from_schedule);
8811b596
SL		 4259 kfree(cb);
4260}
4261
4262static void release_stripe_plug(struct mddev *mddev,
4263 struct stripe_head *sh)
4264{
4265 struct blk_plug_cb *blk_cb = blk_check_plugged(
4266 raid5_unplug, mddev,
4267 sizeof(struct raid5_plug_cb));
4268 struct raid5_plug_cb *cb;
4269
4270 if (!blk_cb) {
4271 release_stripe(sh);
4272 return;
4273 }
4274
4275 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4276
4277 if (cb->list.next == NULL)
4278 INIT_LIST_HEAD(&cb->list);
4279
4280 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4281 list_add_tail(&sh->lru, &cb->list);
4282 else
4283 release_stripe(sh);
4284}
4285
620125f2
SL		 4286static void make_discard_request(struct mddev *mddev, struct bio *bi)
4287{
4288 struct r5conf *conf = mddev->private;
4289 sector_t logical_sector, last_sector;
4290 struct stripe_head *sh;
4291 int remaining;
4292 int stripe_sectors;
4293
4294 if (mddev->reshape_position != MaxSector)
4295 /* Skip discard while reshape is happening */
4296 return;
4297
4298 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4299 last_sector = bi->bi_sector + (bi->bi_size>>9);
4300
4301 bi->bi_next = NULL;
4302 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4303
4304 stripe_sectors = conf->chunk_sectors *
4305 (conf->raid_disks - conf->max_degraded);
4306 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4307 stripe_sectors);
4308 sector_div(last_sector, stripe_sectors);
4309
4310 logical_sector *= conf->chunk_sectors;
4311 last_sector *= conf->chunk_sectors;
4312
4313 for (; logical_sector < last_sector;
4314 logical_sector += STRIPE_SECTORS) {
4315 DEFINE_WAIT(w);
4316 int d;
4317 again:
4318 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4319 prepare_to_wait(&conf->wait_for_overlap, &w,
4320 TASK_UNINTERRUPTIBLE);
f8dfcffd
N		 4321 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4322 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4323 release_stripe(sh);
4324 schedule();
4325 goto again;
4326 }
4327 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
620125f2
SL		 4328 spin_lock_irq(&sh->stripe_lock);
4329 for (d = 0; d < conf->raid_disks; d++) {
4330 if (d == sh->pd_idx || d == sh->qd_idx)
4331 continue;
4332 if (sh->dev[d].towrite || sh->dev[d].toread) {
4333 set_bit(R5_Overlap, &sh->dev[d].flags);
4334 spin_unlock_irq(&sh->stripe_lock);
4335 release_stripe(sh);
4336 schedule();
4337 goto again;
4338 }
4339 }
f8dfcffd 4340 set_bit(STRIPE_DISCARD, &sh->state);
620125f2
SL		 4341 finish_wait(&conf->wait_for_overlap, &w);
4342 for (d = 0; d < conf->raid_disks; d++) {
4343 if (d == sh->pd_idx || d == sh->qd_idx)
4344 continue;
4345 sh->dev[d].towrite = bi;
4346 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4347 raid5_inc_bi_active_stripes(bi);
4348 }
4349 spin_unlock_irq(&sh->stripe_lock);
4350 if (conf->mddev->bitmap) {
4351 for (d = 0;
4352 d < conf->raid_disks - conf->max_degraded;
4353 d++)
4354 bitmap_startwrite(mddev->bitmap,
4355 sh->sector,
4356 STRIPE_SECTORS,
4357 0);
4358 sh->bm_seq = conf->seq_flush + 1;
4359 set_bit(STRIPE_BIT_DELAY, &sh->state);
4360 }
4361
4362 set_bit(STRIPE_HANDLE, &sh->state);
4363 clear_bit(STRIPE_DELAYED, &sh->state);
4364 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4365 atomic_inc(&conf->preread_active_stripes);
4366 release_stripe_plug(mddev, sh);
4367 }
4368
4369 remaining = raid5_dec_bi_active_stripes(bi);
4370 if (remaining == 0) {
4371 md_write_end(mddev);
4372 bio_endio(bi, 0);
4373 }
4374}
4375
b4fdcb02 4376static void make_request(struct mddev *mddev, struct bio * bi)
1da177e4 4377{
d1688a6d 4378 struct r5conf *conf = mddev->private;
911d4ee8 4379 int dd_idx;
1da177e4
LT		 4380 sector_t new_sector;
4381 sector_t logical_sector, last_sector;
4382 struct stripe_head *sh;
a362357b 4383 const int rw = bio_data_dir(bi);
49077326 4384 int remaining;
1da177e4 4385
e9c7469b
TH		 4386 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4387 md_flush_request(mddev, bi);
5a7bbad2 4388 return;
e5dcdd80
N		 4389 }
4390
3d310eb7 4391 md_write_start(mddev, bi);
06d91a5f 4392
802ba064 4393 if (rw == READ &&
52488615 4394 mddev->reshape_position == MaxSector &&
21a52c6d 4395 chunk_aligned_read(mddev,bi))
5a7bbad2 4396 return;
52488615 4397
620125f2
SL		 4398 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4399 make_discard_request(mddev, bi);
4400 return;
4401 }
4402
1da177e4 4403 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
f73a1c7d 4404 last_sector = bio_end_sector(bi);
1da177e4
LT		 4405 bi->bi_next = NULL;
4406 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
06d91a5f 4407
1da177e4
LT		 4408 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4409 DEFINE_WAIT(w);
b5663ba4 4410 int previous;
b578d55f 4411
7ecaa1e6 4412 retry:
b5663ba4 4413 previous = 0;
b578d55f 4414 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
b0f9ec04 4415 if (unlikely(conf->reshape_progress != MaxSector)) {
fef9c61f 4416 /* spinlock is needed as reshape_progress may be
df8e7f76
N		 4417 * 64bit on a 32bit platform, and so it might be
4418 * possible to see a half-updated value
aeb878b0 4419 * Of course reshape_progress could change after
df8e7f76
N		 4420 * the lock is dropped, so once we get a reference
4421 * to the stripe that we think it is, we will have
4422 * to check again.
4423 */
7ecaa1e6 4424 spin_lock_irq(&conf->device_lock);
2c810cdd 4425 if (mddev->reshape_backwards
fef9c61f
N		 4426 ? logical_sector < conf->reshape_progress
4427 : logical_sector >= conf->reshape_progress) {
b5663ba4
N		 4428 previous = 1;
4429 } else {
2c810cdd 4430 if (mddev->reshape_backwards
fef9c61f
N		 4431 ? logical_sector < conf->reshape_safe
4432 : logical_sector >= conf->reshape_safe) {
b578d55f
N		 4433 spin_unlock_irq(&conf->device_lock);
4434 schedule();
4435 goto retry;
4436 }
4437 }
7ecaa1e6
N		 4438 spin_unlock_irq(&conf->device_lock);
4439 }
16a53ecc 4440
112bf897
N		 4441 new_sector = raid5_compute_sector(conf, logical_sector,
4442 previous,
911d4ee8 4443 &dd_idx, NULL);
0c55e022 4444 pr_debug("raid456: make_request, sector %llu logical %llu\n",
1da177e4
LT		 4445 (unsigned long long)new_sector,
4446 (unsigned long long)logical_sector);
4447
b5663ba4 4448 sh = get_active_stripe(conf, new_sector, previous,
a8c906ca 4449 (bi->bi_rw&RWA_MASK), 0);
1da177e4 4450 if (sh) {
b0f9ec04 4451 if (unlikely(previous)) {
7ecaa1e6 4452 /* expansion might have moved on while waiting for a
df8e7f76
N		 4453 * stripe, so we must do the range check again.
4454 * Expansion could still move past after this
4455 * test, but as we are holding a reference to
4456 * 'sh', we know that if that happens,
4457 * STRIPE_EXPANDING will get set and the expansion
4458 * won't proceed until we finish with the stripe.
7ecaa1e6
N		 4459 */
4460 int must_retry = 0;
4461 spin_lock_irq(&conf->device_lock);
2c810cdd 4462 if (mddev->reshape_backwards
b0f9ec04
N		 4463 ? logical_sector >= conf->reshape_progress
4464 : logical_sector < conf->reshape_progress)
7ecaa1e6
N		 4465 /* mismatch, need to try again */
4466 must_retry = 1;
4467 spin_unlock_irq(&conf->device_lock);
4468 if (must_retry) {
4469 release_stripe(sh);
7a3ab908 4470 schedule();
7ecaa1e6
N		 4471 goto retry;
4472 }
4473 }
e62e58a5 4474
ffd96e35 4475 if (rw == WRITE &&
a5c308d4 4476 logical_sector >= mddev->suspend_lo &&
e464eafd
N		 4477 logical_sector < mddev->suspend_hi) {
4478 release_stripe(sh);
e62e58a5
N		 4479 /* As the suspend_* range is controlled by
4480 * userspace, we want an interruptible
4481 * wait.
4482 */
4483 flush_signals(current);
4484 prepare_to_wait(&conf->wait_for_overlap,
4485 &w, TASK_INTERRUPTIBLE);
4486 if (logical_sector >= mddev->suspend_lo &&
4487 logical_sector < mddev->suspend_hi)
4488 schedule();
e464eafd
N		 4489 goto retry;
4490 }
7ecaa1e6
N		 4491
4492 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
ffd96e35 4493 !add_stripe_bio(sh, bi, dd_idx, rw)) {
7ecaa1e6
N		 4494 /* Stripe is busy expanding or
4495 * add failed due to overlap. Flush everything
1da177e4
LT		 4496 * and wait a while
4497 */
482c0834 4498 md_wakeup_thread(mddev->thread);
1da177e4
LT		 4499 release_stripe(sh);
4500 schedule();
4501 goto retry;
4502 }
4503 finish_wait(&conf->wait_for_overlap, &w);
6ed3003c
N		 4504 set_bit(STRIPE_HANDLE, &sh->state);
4505 clear_bit(STRIPE_DELAYED, &sh->state);
a852d7b8 4506 if ((bi->bi_rw & REQ_SYNC) &&
729a1866
N		 4507 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4508 atomic_inc(&conf->preread_active_stripes);
8811b596 4509 release_stripe_plug(mddev, sh);
1da177e4
LT		 4510 } else {
4511 /* cannot get stripe for read-ahead, just give-up */
4512 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4513 finish_wait(&conf->wait_for_overlap, &w);
4514 break;
4515 }
1da177e4 4516 }
7c13edc8 4517
e7836bd6 4518 remaining = raid5_dec_bi_active_stripes(bi);
f6344757 4519 if (remaining == 0) {
1da177e4 4520
16a53ecc 4521 if ( rw == WRITE )
1da177e4 4522 md_write_end(mddev);
6712ecf8 4523
0a82a8d1
LT		 4524 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4525 bi, 0);
0e13fe23 4526 bio_endio(bi, 0);
1da177e4 4527 }
1da177e4
LT		 4528}
4529
fd01b88c 4530static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
b522adcd 4531
fd01b88c 4532static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
1da177e4 4533{
52c03291
N		 4534 /* reshaping is quite different to recovery/resync so it is
4535 * handled quite separately ... here.
4536 *
4537 * On each call to sync_request, we gather one chunk worth of
4538 * destination stripes and flag them as expanding.
4539 * Then we find all the source stripes and request reads.
4540 * As the reads complete, handle_stripe will copy the data
4541 * into the destination stripe and release that stripe.
4542 */
d1688a6d 4543 struct r5conf *conf = mddev->private;
1da177e4 4544 struct stripe_head *sh;
ccfcc3c1 4545 sector_t first_sector, last_sector;
f416885e
N		 4546 int raid_disks = conf->previous_raid_disks;
4547 int data_disks = raid_disks - conf->max_degraded;
4548 int new_data_disks = conf->raid_disks - conf->max_degraded;
52c03291
N		 4549 int i;
4550 int dd_idx;
c8f517c4 4551 sector_t writepos, readpos, safepos;
ec32a2bd 4552 sector_t stripe_addr;
7a661381 4553 int reshape_sectors;
ab69ae12 4554 struct list_head stripes;
52c03291 4555
fef9c61f
N		 4556 if (sector_nr == 0) {
4557 /* If restarting in the middle, skip the initial sectors */
2c810cdd 4558 if (mddev->reshape_backwards &&
fef9c61f
N		 4559 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4560 sector_nr = raid5_size(mddev, 0, 0)
4561 - conf->reshape_progress;
2c810cdd 4562 } else if (!mddev->reshape_backwards &&
fef9c61f
N		 4563 conf->reshape_progress > 0)
4564 sector_nr = conf->reshape_progress;
f416885e 4565 sector_div(sector_nr, new_data_disks);
fef9c61f 4566 if (sector_nr) {
8dee7211
N		 4567 mddev->curr_resync_completed = sector_nr;
4568 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
fef9c61f
N		 4569 *skipped = 1;
4570 return sector_nr;
4571 }
52c03291
N		 4572 }
4573
7a661381
N		 4574 /* We need to process a full chunk at a time.
4575 * If old and new chunk sizes differ, we need to process the
4576 * largest of these
4577 */
664e7c41
AN		 4578 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4579 reshape_sectors = mddev->new_chunk_sectors;
7a661381 4580 else
9d8f0363 4581 reshape_sectors = mddev->chunk_sectors;
7a661381 4582
b5254dd5
N		 4583 /* We update the metadata at least every 10 seconds, or when
4584 * the data about to be copied would over-write the source of
4585 * the data at the front of the range. i.e. one new_stripe
4586 * along from reshape_progress new_maps to after where
4587 * reshape_safe old_maps to
52c03291 4588 */
fef9c61f 4589 writepos = conf->reshape_progress;
f416885e 4590 sector_div(writepos, new_data_disks);
c8f517c4
N		 4591 readpos = conf->reshape_progress;
4592 sector_div(readpos, data_disks);
fef9c61f 4593 safepos = conf->reshape_safe;
f416885e 4594 sector_div(safepos, data_disks);
2c810cdd 4595 if (mddev->reshape_backwards) {
ed37d83e 4596 writepos -= min_t(sector_t, reshape_sectors, writepos);
c8f517c4 4597 readpos += reshape_sectors;
7a661381 4598 safepos += reshape_sectors;
fef9c61f 4599 } else {
7a661381 4600 writepos += reshape_sectors;
ed37d83e
N		 4601 readpos -= min_t(sector_t, reshape_sectors, readpos);
4602 safepos -= min_t(sector_t, reshape_sectors, safepos);
fef9c61f 4603 }
52c03291 4604
b5254dd5
N		 4605 /* Having calculated the 'writepos' possibly use it
4606 * to set 'stripe_addr' which is where we will write to.
4607 */
4608 if (mddev->reshape_backwards) {
4609 BUG_ON(conf->reshape_progress == 0);
4610 stripe_addr = writepos;
4611 BUG_ON((mddev->dev_sectors &
4612 ~((sector_t)reshape_sectors - 1))
4613 - reshape_sectors - stripe_addr
4614 != sector_nr);
4615 } else {
4616 BUG_ON(writepos != sector_nr + reshape_sectors);
4617 stripe_addr = sector_nr;
4618 }
4619
c8f517c4
N		 4620 /* 'writepos' is the most advanced device address we might write.
4621 * 'readpos' is the least advanced device address we might read.
4622 * 'safepos' is the least address recorded in the metadata as having
4623 * been reshaped.
b5254dd5
N		 4624 * If there is a min_offset_diff, these are adjusted either by
4625 * increasing the safepos/readpos if diff is negative, or
4626 * increasing writepos if diff is positive.
4627 * If 'readpos' is then behind 'writepos', there is no way that we can
c8f517c4
N		 4628 * ensure safety in the face of a crash - that must be done by userspace
4629 * making a backup of the data. So in that case there is no particular
4630 * rush to update metadata.
4631 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4632 * update the metadata to advance 'safepos' to match 'readpos' so that
4633 * we can be safe in the event of a crash.
4634 * So we insist on updating metadata if safepos is behind writepos and
4635 * readpos is beyond writepos.
4636 * In any case, update the metadata every 10 seconds.
4637 * Maybe that number should be configurable, but I'm not sure it is
4638 * worth it.... maybe it could be a multiple of safemode_delay???
4639 */
b5254dd5
N		 4640 if (conf->min_offset_diff < 0) {
4641 safepos += -conf->min_offset_diff;
4642 readpos += -conf->min_offset_diff;
4643 } else
4644 writepos += conf->min_offset_diff;
4645
2c810cdd 4646 if ((mddev->reshape_backwards
c8f517c4
N		 4647 ? (safepos > writepos && readpos < writepos)
4648 : (safepos < writepos && readpos > writepos)) ||
4649 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
52c03291
N		 4650 /* Cannot proceed until we've updated the superblock... */
4651 wait_event(conf->wait_for_overlap,
4652 atomic_read(&conf->reshape_stripes)==0);
fef9c61f 4653 mddev->reshape_position = conf->reshape_progress;
75d3da43 4654 mddev->curr_resync_completed = sector_nr;
c8f517c4 4655 conf->reshape_checkpoint = jiffies;
850b2b42 4656 set_bit(MD_CHANGE_DEVS, &mddev->flags);
52c03291 4657 md_wakeup_thread(mddev->thread);
850b2b42 4658 wait_event(mddev->sb_wait, mddev->flags == 0 ||
52c03291
N		 4659 kthread_should_stop());
4660 spin_lock_irq(&conf->device_lock);
fef9c61f 4661 conf->reshape_safe = mddev->reshape_position;
52c03291
N		 4662 spin_unlock_irq(&conf->device_lock);
4663 wake_up(&conf->wait_for_overlap);
acb180b0 4664 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
52c03291
N		 4665 }
4666
ab69ae12 4667 INIT_LIST_HEAD(&stripes);
7a661381 4668 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
52c03291 4669 int j;
a9f326eb 4670 int skipped_disk = 0;
a8c906ca 4671 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
52c03291
N		 4672 set_bit(STRIPE_EXPANDING, &sh->state);
4673 atomic_inc(&conf->reshape_stripes);
4674 /* If any of this stripe is beyond the end of the old
4675 * array, then we need to zero those blocks
4676 */
4677 for (j=sh->disks; j--;) {
4678 sector_t s;
4679 if (j == sh->pd_idx)
4680 continue;
f416885e 4681 if (conf->level == 6 &&
d0dabf7e 4682 j == sh->qd_idx)
f416885e 4683 continue;
784052ec 4684 s = compute_blocknr(sh, j, 0);
b522adcd 4685 if (s < raid5_size(mddev, 0, 0)) {
a9f326eb 4686 skipped_disk = 1;
52c03291
N		 4687 continue;
4688 }
4689 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4690 set_bit(R5_Expanded, &sh->dev[j].flags);
4691 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4692 }
a9f326eb 4693 if (!skipped_disk) {
52c03291
N		 4694 set_bit(STRIPE_EXPAND_READY, &sh->state);
4695 set_bit(STRIPE_HANDLE, &sh->state);
4696 }
ab69ae12 4697 list_add(&sh->lru, &stripes);
52c03291
N		 4698 }
4699 spin_lock_irq(&conf->device_lock);
2c810cdd 4700 if (mddev->reshape_backwards)
7a661381 4701 conf->reshape_progress -= reshape_sectors * new_data_disks;
fef9c61f 4702 else
7a661381 4703 conf->reshape_progress += reshape_sectors * new_data_disks;
52c03291
N		 4704 spin_unlock_irq(&conf->device_lock);
4705 /* Ok, those stripe are ready. We can start scheduling
4706 * reads on the source stripes.
4707 * The source stripes are determined by mapping the first and last
4708 * block on the destination stripes.
4709 */
52c03291 4710 first_sector =
ec32a2bd 4711 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
911d4ee8 4712 1, &dd_idx, NULL);
52c03291 4713 last_sector =
0e6e0271 4714 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
09c9e5fa 4715 * new_data_disks - 1),
911d4ee8 4716 1, &dd_idx, NULL);
58c0fed4
AN		 4717 if (last_sector >= mddev->dev_sectors)
4718 last_sector = mddev->dev_sectors - 1;
52c03291 4719 while (first_sector <= last_sector) {
a8c906ca 4720 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
52c03291
N		 4721 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4722 set_bit(STRIPE_HANDLE, &sh->state);
4723 release_stripe(sh);
4724 first_sector += STRIPE_SECTORS;
4725 }
ab69ae12
N		 4726 /* Now that the sources are clearly marked, we can release
4727 * the destination stripes
4728 */
4729 while (!list_empty(&stripes)) {
4730 sh = list_entry(stripes.next, struct stripe_head, lru);
4731 list_del_init(&sh->lru);
4732 release_stripe(sh);
4733 }
c6207277
N		 4734 /* If this takes us to the resync_max point where we have to pause,
4735 * then we need to write out the superblock.
4736 */
7a661381 4737 sector_nr += reshape_sectors;
c03f6a19
N		 4738 if ((sector_nr - mddev->curr_resync_completed) * 2
4739 >= mddev->resync_max - mddev->curr_resync_completed) {
c6207277
N		 4740 /* Cannot proceed until we've updated the superblock... */
4741 wait_event(conf->wait_for_overlap,
4742 atomic_read(&conf->reshape_stripes) == 0);
fef9c61f 4743 mddev->reshape_position = conf->reshape_progress;
75d3da43 4744 mddev->curr_resync_completed = sector_nr;
c8f517c4 4745 conf->reshape_checkpoint = jiffies;
c6207277
N		 4746 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4747 md_wakeup_thread(mddev->thread);
4748 wait_event(mddev->sb_wait,
4749 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4750 || kthread_should_stop());
4751 spin_lock_irq(&conf->device_lock);
fef9c61f 4752 conf->reshape_safe = mddev->reshape_position;
c6207277
N		 4753 spin_unlock_irq(&conf->device_lock);
4754 wake_up(&conf->wait_for_overlap);
acb180b0 4755 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
c6207277 4756 }
7a661381 4757 return reshape_sectors;
52c03291
N		 4758}
4759
4760/* FIXME go_faster isn't used */
fd01b88c 4761static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
52c03291 4762{
d1688a6d 4763 struct r5conf *conf = mddev->private;
52c03291 4764 struct stripe_head *sh;
58c0fed4 4765 sector_t max_sector = mddev->dev_sectors;
57dab0bd 4766 sector_t sync_blocks;
16a53ecc
N		 4767 int still_degraded = 0;
4768 int i;
1da177e4 4769
72626685 4770 if (sector_nr >= max_sector) {
1da177e4 4771 /* just being told to finish up .. nothing much to do */
cea9c228 4772
29269553
N		 4773 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4774 end_reshape(conf);
4775 return 0;
4776 }
72626685
N		 4777
4778 if (mddev->curr_resync < max_sector) /* aborted */
4779 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4780 &sync_blocks, 1);
16a53ecc 4781 else /* completed sync */
72626685
N		 4782 conf->fullsync = 0;
4783 bitmap_close_sync(mddev->bitmap);
4784
1da177e4
LT		 4785 return 0;
4786 }
ccfcc3c1 4787
64bd660b
N		 4788 /* Allow raid5_quiesce to complete */
4789 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4790
52c03291
N		 4791 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4792 return reshape_request(mddev, sector_nr, skipped);
f6705578 4793
c6207277
N		 4794 /* No need to check resync_max as we never do more than one
4795 * stripe, and as resync_max will always be on a chunk boundary,
4796 * if the check in md_do_sync didn't fire, there is no chance
4797 * of overstepping resync_max here
4798 */
4799
16a53ecc 4800 /* if there is too many failed drives and we are trying
1da177e4
LT		 4801 * to resync, then assert that we are finished, because there is
4802 * nothing we can do.
4803 */
3285edf1 4804 if (mddev->degraded >= conf->max_degraded &&
16a53ecc 4805 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
58c0fed4 4806 sector_t rv = mddev->dev_sectors - sector_nr;
57afd89f 4807 *skipped = 1;
1da177e4
LT		 4808 return rv;
4809 }
6f608040 4810 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4811 !conf->fullsync &&
4812 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4813 sync_blocks >= STRIPE_SECTORS) {
72626685
N		 4814 /* we can skip this block, and probably more */
4815 sync_blocks /= STRIPE_SECTORS;
4816 *skipped = 1;
4817 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4818 }
1da177e4 4819
b47490c9
N		 4820 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4821
a8c906ca 4822 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
1da177e4 4823 if (sh == NULL) {
a8c906ca 4824 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
1da177e4 4825 /* make sure we don't swamp the stripe cache if someone else
16a53ecc 4826 * is trying to get access
1da177e4 4827 */
66c006a5 4828 schedule_timeout_uninterruptible(1);
1da177e4 4829 }
16a53ecc
N		 4830 /* Need to check if array will still be degraded after recovery/resync
4831 * We don't need to check the 'failed' flag as when that gets set,
4832 * recovery aborts.
4833 */
f001a70c 4834 for (i = 0; i < conf->raid_disks; i++)
16a53ecc
N		 4835 if (conf->disks[i].rdev == NULL)
4836 still_degraded = 1;
4837
4838 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4839
83206d66 4840 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
1da177e4 4841
1442577b 4842 handle_stripe(sh);
1da177e4
LT		 4843 release_stripe(sh);
4844
4845 return STRIPE_SECTORS;
4846}
4847
d1688a6d 4848static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
46031f9a
RBJ		 4849{
4850 /* We may not be able to submit a whole bio at once as there
4851 * may not be enough stripe_heads available.
4852 * We cannot pre-allocate enough stripe_heads as we may need
4853 * more than exist in the cache (if we allow ever large chunks).
4854 * So we do one stripe head at a time and record in
4855 * ->bi_hw_segments how many have been done.
4856 *
4857 * We *know* that this entire raid_bio is in one chunk, so
4858 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4859 */
4860 struct stripe_head *sh;
911d4ee8 4861 int dd_idx;
46031f9a
RBJ		 4862 sector_t sector, logical_sector, last_sector;
4863 int scnt = 0;
4864 int remaining;
4865 int handled = 0;
4866
4867 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
112bf897 4868 sector = raid5_compute_sector(conf, logical_sector,
911d4ee8 4869 0, &dd_idx, NULL);
f73a1c7d 4870 last_sector = bio_end_sector(raid_bio);
46031f9a
RBJ		 4871
4872 for (; logical_sector < last_sector;
387bb173
NB		 4873 logical_sector += STRIPE_SECTORS,
4874 sector += STRIPE_SECTORS,
4875 scnt++) {
46031f9a 4876
e7836bd6 4877 if (scnt < raid5_bi_processed_stripes(raid_bio))
46031f9a
RBJ		 4878 /* already done this stripe */
4879 continue;
4880
a8c906ca 4881 sh = get_active_stripe(conf, sector, 0, 1, 0);
46031f9a
RBJ		 4882
4883 if (!sh) {
4884 /* failed to get a stripe - must wait */
e7836bd6 4885 raid5_set_bi_processed_stripes(raid_bio, scnt);
46031f9a
RBJ		 4886 conf->retry_read_aligned = raid_bio;
4887 return handled;
4888 }
4889
387bb173
NB		 4890 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4891 release_stripe(sh);
e7836bd6 4892 raid5_set_bi_processed_stripes(raid_bio, scnt);
387bb173
NB		 4893 conf->retry_read_aligned = raid_bio;
4894 return handled;
4895 }
4896
3f9e7c14 4897 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
36d1c647 4898 handle_stripe(sh);
46031f9a
RBJ		 4899 release_stripe(sh);
4900 handled++;
4901 }
e7836bd6 4902 remaining = raid5_dec_bi_active_stripes(raid_bio);
0a82a8d1
LT		 4903 if (remaining == 0) {
4904 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
4905 raid_bio, 0);
0e13fe23 4906 bio_endio(raid_bio, 0);
0a82a8d1 4907 }
46031f9a
RBJ		 4908 if (atomic_dec_and_test(&conf->active_aligned_reads))
4909 wake_up(&conf->wait_for_stripe);
4910 return handled;
4911}
4912
46a06401 4913#define MAX_STRIPE_BATCH 8
851c30c9 4914static int handle_active_stripes(struct r5conf *conf, int group)
46a06401
SL		 4915{
4916 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4917 int i, batch_size = 0;
4918
4919 while (batch_size < MAX_STRIPE_BATCH &&
851c30c9 4920 (sh = __get_priority_stripe(conf, group)) != NULL)
46a06401
SL		 4921 batch[batch_size++] = sh;
4922
4923 if (batch_size == 0)
4924 return batch_size;
4925 spin_unlock_irq(&conf->device_lock);
4926
4927 for (i = 0; i < batch_size; i++)
4928 handle_stripe(batch[i]);
4929
4930 cond_resched();
4931
4932 spin_lock_irq(&conf->device_lock);
4933 for (i = 0; i < batch_size; i++)
4934 __release_stripe(conf, batch[i]);
4935 return batch_size;
4936}
46031f9a 4937
851c30c9
SL		 4938static void raid5_do_work(struct work_struct *work)
4939{
4940 struct r5worker *worker = container_of(work, struct r5worker, work);
4941 struct r5worker_group *group = worker->group;
4942 struct r5conf *conf = group->conf;
4943 int group_id = group - conf->worker_groups;
4944 int handled;
4945 struct blk_plug plug;
4946
4947 pr_debug("+++ raid5worker active\n");
4948
4949 blk_start_plug(&plug);
4950 handled = 0;
4951 spin_lock_irq(&conf->device_lock);
4952 while (1) {
4953 int batch_size, released;
4954
4955 released = release_stripe_list(conf);
4956
4957 batch_size = handle_active_stripes(conf, group_id);
4958 if (!batch_size && !released)
4959 break;
4960 handled += batch_size;
4961 }
4962 pr_debug("%d stripes handled\n", handled);
4963
4964 spin_unlock_irq(&conf->device_lock);
4965 blk_finish_plug(&plug);
4966
4967 pr_debug("--- raid5worker inactive\n");
4968}
4969
1da177e4
LT		 4970/*
4971 * This is our raid5 kernel thread.
4972 *
4973 * We scan the hash table for stripes which can be handled now.
4974 * During the scan, completed stripes are saved for us by the interrupt
4975 * handler, so that they will not have to wait for our next wakeup.
4976 */
4ed8731d 4977static void raid5d(struct md_thread *thread)
1da177e4 4978{
4ed8731d 4979 struct mddev *mddev = thread->mddev;
d1688a6d 4980 struct r5conf *conf = mddev->private;
1da177e4 4981 int handled;
e1dfa0a2 4982 struct blk_plug plug;
1da177e4 4983
45b4233c 4984 pr_debug("+++ raid5d active\n");
1da177e4
LT		 4985
4986 md_check_recovery(mddev);
1da177e4 4987
e1dfa0a2 4988 blk_start_plug(&plug);
1da177e4
LT		 4989 handled = 0;
4990 spin_lock_irq(&conf->device_lock);
4991 while (1) {
46031f9a 4992 struct bio *bio;
773ca82f
SL		 4993 int batch_size, released;
4994
4995 released = release_stripe_list(conf);
1da177e4 4996
0021b7bc 4997 if (
7c13edc8
N		 4998 !list_empty(&conf->bitmap_list)) {
4999 /* Now is a good time to flush some bitmap updates */
5000 conf->seq_flush++;
700e432d 5001 spin_unlock_irq(&conf->device_lock);
72626685 5002 bitmap_unplug(mddev->bitmap);
700e432d 5003 spin_lock_irq(&conf->device_lock);
7c13edc8 5004 conf->seq_write = conf->seq_flush;
72626685
N		 5005 activate_bit_delay(conf);
5006 }
0021b7bc 5007 raid5_activate_delayed(conf);
72626685 5008
46031f9a
RBJ		 5009 while ((bio = remove_bio_from_retry(conf))) {
5010 int ok;
5011 spin_unlock_irq(&conf->device_lock);
5012 ok = retry_aligned_read(conf, bio);
5013 spin_lock_irq(&conf->device_lock);
5014 if (!ok)
5015 break;
5016 handled++;
5017 }
5018
851c30c9 5019 batch_size = handle_active_stripes(conf, ANY_GROUP);
773ca82f 5020 if (!batch_size && !released)
1da177e4 5021 break;
46a06401 5022 handled += batch_size;
1da177e4 5023
46a06401
SL		 5024 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5025 spin_unlock_irq(&conf->device_lock);
de393cde 5026 md_check_recovery(mddev);
46a06401
SL		 5027 spin_lock_irq(&conf->device_lock);
5028 }
1da177e4 5029 }
45b4233c 5030 pr_debug("%d stripes handled\n", handled);
1da177e4
LT		 5031
5032 spin_unlock_irq(&conf->device_lock);
5033
c9f21aaf 5034 async_tx_issue_pending_all();
e1dfa0a2 5035 blk_finish_plug(&plug);
1da177e4 5036
45b4233c 5037 pr_debug("--- raid5d inactive\n");
1da177e4
LT		 5038}
5039
3f294f4f 5040static ssize_t
fd01b88c 5041raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
3f294f4f 5042{
d1688a6d 5043 struct r5conf *conf = mddev->private;
96de1e66
N		 5044 if (conf)
5045 return sprintf(page, "%d\n", conf->max_nr_stripes);
5046 else
5047 return 0;
3f294f4f
N		 5048}
5049
c41d4ac4 5050int
fd01b88c 5051raid5_set_cache_size(struct mddev *mddev, int size)
3f294f4f 5052{
d1688a6d 5053 struct r5conf *conf = mddev->private;
b5470dc5
DW		 5054 int err;
5055
c41d4ac4 5056 if (size <= 16 || size > 32768)
3f294f4f 5057 return -EINVAL;
c41d4ac4 5058 while (size < conf->max_nr_stripes) {
3f294f4f
N		 5059 if (drop_one_stripe(conf))
5060 conf->max_nr_stripes--;
5061 else
5062 break;
5063 }
b5470dc5
DW		 5064 err = md_allow_write(mddev);
5065 if (err)
5066 return err;
c41d4ac4 5067 while (size > conf->max_nr_stripes) {
3f294f4f
N		 5068 if (grow_one_stripe(conf))
5069 conf->max_nr_stripes++;
5070 else break;
5071 }
c41d4ac4
N		 5072 return 0;
5073}
5074EXPORT_SYMBOL(raid5_set_cache_size);
5075
5076static ssize_t
fd01b88c 5077raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
c41d4ac4 5078{
d1688a6d 5079 struct r5conf *conf = mddev->private;
c41d4ac4
N		 5080 unsigned long new;
5081 int err;
5082
5083 if (len >= PAGE_SIZE)
5084 return -EINVAL;
5085 if (!conf)
5086 return -ENODEV;
5087
b29bebd6 5088 if (kstrtoul(page, 10, &new))
c41d4ac4
N		 5089 return -EINVAL;
5090 err = raid5_set_cache_size(mddev, new);
5091 if (err)
5092 return err;
3f294f4f
N		 5093 return len;
5094}
007583c9 5095
96de1e66
N		 5096static struct md_sysfs_entry
5097raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5098 raid5_show_stripe_cache_size,
5099 raid5_store_stripe_cache_size);
3f294f4f 5100
8b3e6cdc 5101static ssize_t
fd01b88c 5102raid5_show_preread_threshold(struct mddev *mddev, char *page)
8b3e6cdc 5103{
d1688a6d 5104 struct r5conf *conf = mddev->private;
8b3e6cdc
DW		 5105 if (conf)
5106 return sprintf(page, "%d\n", conf->bypass_threshold);
5107 else
5108 return 0;
5109}
5110
5111static ssize_t
fd01b88c 5112raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
8b3e6cdc 5113{
d1688a6d 5114 struct r5conf *conf = mddev->private;
4ef197d8 5115 unsigned long new;
8b3e6cdc
DW		 5116 if (len >= PAGE_SIZE)
5117 return -EINVAL;
5118 if (!conf)
5119 return -ENODEV;
5120
b29bebd6 5121 if (kstrtoul(page, 10, &new))
8b3e6cdc 5122 return -EINVAL;
4ef197d8 5123 if (new > conf->max_nr_stripes)
8b3e6cdc
DW		 5124 return -EINVAL;
5125 conf->bypass_threshold = new;
5126 return len;
5127}
5128
5129static struct md_sysfs_entry
5130raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5131 S_IRUGO | S_IWUSR,
5132 raid5_show_preread_threshold,
5133 raid5_store_preread_threshold);
5134
3f294f4f 5135static ssize_t
fd01b88c 5136stripe_cache_active_show(struct mddev *mddev, char *page)
3f294f4f 5137{
d1688a6d 5138 struct r5conf *conf = mddev->private;
96de1e66
N		 5139 if (conf)
5140 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5141 else
5142 return 0;
3f294f4f
N		 5143}
5144
96de1e66
N		 5145static struct md_sysfs_entry
5146raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
3f294f4f 5147
b721420e
SL		 5148static ssize_t
5149raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
5150{
5151 struct r5conf *conf = mddev->private;
5152 if (conf)
5153 return sprintf(page, "%d\n", conf->worker_cnt_per_group);
5154 else
5155 return 0;
5156}
5157
5158static int alloc_thread_groups(struct r5conf *conf, int cnt);
5159static ssize_t
5160raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
5161{
5162 struct r5conf *conf = mddev->private;
5163 unsigned long new;
5164 int err;
5165 struct r5worker_group *old_groups;
5166 int old_group_cnt;
5167
5168 if (len >= PAGE_SIZE)
5169 return -EINVAL;
5170 if (!conf)
5171 return -ENODEV;
5172
5173 if (kstrtoul(page, 10, &new))
5174 return -EINVAL;
5175
5176 if (new == conf->worker_cnt_per_group)
5177 return len;
5178
5179 mddev_suspend(mddev);
5180
5181 old_groups = conf->worker_groups;
5182 old_group_cnt = conf->worker_cnt_per_group;
5183
5184 conf->worker_groups = NULL;
5185 err = alloc_thread_groups(conf, new);
5186 if (err) {
5187 conf->worker_groups = old_groups;
5188 conf->worker_cnt_per_group = old_group_cnt;
5189 } else {
5190 if (old_groups)
5191 kfree(old_groups[0].workers);
5192 kfree(old_groups);
5193 }
5194
5195 mddev_resume(mddev);
5196
5197 if (err)
5198 return err;
5199 return len;
5200}
5201
5202static struct md_sysfs_entry
5203raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
5204 raid5_show_group_thread_cnt,
5205 raid5_store_group_thread_cnt);
5206
007583c9 5207static struct attribute *raid5_attrs[] = {
3f294f4f
N		 5208 &raid5_stripecache_size.attr,
5209 &raid5_stripecache_active.attr,
8b3e6cdc 5210 &raid5_preread_bypass_threshold.attr,
b721420e 5211 &raid5_group_thread_cnt.attr,
3f294f4f
N		 5212 NULL,
5213};
007583c9
N		 5214static struct attribute_group raid5_attrs_group = {
5215 .name = NULL,
5216 .attrs = raid5_attrs,
3f294f4f
N		 5217};
5218
851c30c9
SL		 5219static int alloc_thread_groups(struct r5conf *conf, int cnt)
5220{
5221 int i, j;
5222 ssize_t size;
5223 struct r5worker *workers;
5224
5225 conf->worker_cnt_per_group = cnt;
5226 if (cnt == 0) {
5227 conf->worker_groups = NULL;
5228 return 0;
5229 }
5230 conf->group_cnt = num_possible_nodes();
5231 size = sizeof(struct r5worker) * cnt;
5232 workers = kzalloc(size * conf->group_cnt, GFP_NOIO);
5233 conf->worker_groups = kzalloc(sizeof(struct r5worker_group) *
5234 conf->group_cnt, GFP_NOIO);
5235 if (!conf->worker_groups || !workers) {
5236 kfree(workers);
5237 kfree(conf->worker_groups);
5238 conf->worker_groups = NULL;
5239 return -ENOMEM;
5240 }
5241
5242 for (i = 0; i < conf->group_cnt; i++) {
5243 struct r5worker_group *group;
5244
5245 group = &conf->worker_groups[i];
5246 INIT_LIST_HEAD(&group->handle_list);
5247 group->conf = conf;
5248 group->workers = workers + i * cnt;
5249
5250 for (j = 0; j < cnt; j++) {
5251 group->workers[j].group = group;
5252 INIT_WORK(&group->workers[j].work, raid5_do_work);
5253 }
5254 }
5255
5256 return 0;
5257}
5258
5259static void free_thread_groups(struct r5conf *conf)
5260{
5261 if (conf->worker_groups)
5262 kfree(conf->worker_groups[0].workers);
5263 kfree(conf->worker_groups);
5264 conf->worker_groups = NULL;
5265}
5266
80c3a6ce 5267static sector_t
fd01b88c 5268raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
80c3a6ce 5269{
d1688a6d 5270 struct r5conf *conf = mddev->private;
80c3a6ce
DW		 5271
5272 if (!sectors)
5273 sectors = mddev->dev_sectors;
5e5e3e78 5274 if (!raid_disks)
7ec05478 5275 /* size is defined by the smallest of previous and new size */
5e5e3e78 5276 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
80c3a6ce 5277
9d8f0363 5278 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
664e7c41 5279 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
80c3a6ce
DW		 5280 return sectors * (raid_disks - conf->max_degraded);
5281}
5282
d1688a6d 5283static void raid5_free_percpu(struct r5conf *conf)
36d1c647
DW		 5284{
5285 struct raid5_percpu *percpu;
5286 unsigned long cpu;
5287
5288 if (!conf->percpu)
5289 return;
5290
5291 get_online_cpus();
5292 for_each_possible_cpu(cpu) {
5293 percpu = per_cpu_ptr(conf->percpu, cpu);
5294 safe_put_page(percpu->spare_page);
d6f38f31 5295 kfree(percpu->scribble);
36d1c647
DW		 5296 }
5297#ifdef CONFIG_HOTPLUG_CPU
5298 unregister_cpu_notifier(&conf->cpu_notify);
5299#endif
5300 put_online_cpus();
5301
5302 free_percpu(conf->percpu);
5303}
5304
d1688a6d 5305static void free_conf(struct r5conf *conf)
95fc17aa 5306{
851c30c9 5307 free_thread_groups(conf);
95fc17aa 5308 shrink_stripes(conf);
36d1c647 5309 raid5_free_percpu(conf);
95fc17aa
DW		 5310 kfree(conf->disks);
5311 kfree(conf->stripe_hashtbl);
5312 kfree(conf);
5313}
5314
36d1c647
DW		 5315#ifdef CONFIG_HOTPLUG_CPU
5316static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5317 void *hcpu)
5318{
d1688a6d 5319 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
36d1c647
DW		 5320 long cpu = (long)hcpu;
5321 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5322
5323 switch (action) {
5324 case CPU_UP_PREPARE:
5325 case CPU_UP_PREPARE_FROZEN:
d6f38f31 5326 if (conf->level == 6 && !percpu->spare_page)
36d1c647 5327 percpu->spare_page = alloc_page(GFP_KERNEL);
d6f38f31
DW		 5328 if (!percpu->scribble)
5329 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5330
5331 if (!percpu->scribble ||
5332 (conf->level == 6 && !percpu->spare_page)) {
5333 safe_put_page(percpu->spare_page);
5334 kfree(percpu->scribble);
36d1c647
DW		 5335 pr_err("%s: failed memory allocation for cpu%ld\n",
5336 __func__, cpu);
55af6bb5 5337 return notifier_from_errno(-ENOMEM);
36d1c647
DW		 5338 }
5339 break;
5340 case CPU_DEAD:
5341 case CPU_DEAD_FROZEN:
5342 safe_put_page(percpu->spare_page);
d6f38f31 5343 kfree(percpu->scribble);
36d1c647 5344 percpu->spare_page = NULL;
d6f38f31 5345 percpu->scribble = NULL;
36d1c647
DW		 5346 break;
5347 default:
5348 break;
5349 }
5350 return NOTIFY_OK;
5351}
5352#endif
5353
d1688a6d 5354static int raid5_alloc_percpu(struct r5conf *conf)
36d1c647
DW		 5355{
5356 unsigned long cpu;
5357 struct page *spare_page;
a29d8b8e 5358 struct raid5_percpu __percpu *allcpus;
d6f38f31 5359 void *scribble;
36d1c647
DW		 5360 int err;
5361
36d1c647
DW		 5362 allcpus = alloc_percpu(struct raid5_percpu);
5363 if (!allcpus)
5364 return -ENOMEM;
5365 conf->percpu = allcpus;
5366
5367 get_online_cpus();
5368 err = 0;
5369 for_each_present_cpu(cpu) {
d6f38f31
DW		 5370 if (conf->level == 6) {
5371 spare_page = alloc_page(GFP_KERNEL);
5372 if (!spare_page) {
5373 err = -ENOMEM;
5374 break;
5375 }
5376 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5377 }
5e5e3e78 5378 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
d6f38f31 5379 if (!scribble) {
36d1c647
DW		 5380 err = -ENOMEM;
5381 break;
5382 }
d6f38f31 5383 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
36d1c647
DW		 5384 }
5385#ifdef CONFIG_HOTPLUG_CPU
5386 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5387 conf->cpu_notify.priority = 0;
5388 if (err == 0)
5389 err = register_cpu_notifier(&conf->cpu_notify);
5390#endif
5391 put_online_cpus();
5392
5393 return err;
5394}
5395
d1688a6d 5396static struct r5conf *setup_conf(struct mddev *mddev)
1da177e4 5397{
d1688a6d 5398 struct r5conf *conf;
5e5e3e78 5399 int raid_disk, memory, max_disks;
3cb03002 5400 struct md_rdev *rdev;
1da177e4 5401 struct disk_info *disk;
0232605d 5402 char pers_name[6];
1da177e4 5403
91adb564
N		 5404 if (mddev->new_level != 5
5405 && mddev->new_level != 4
5406 && mddev->new_level != 6) {
0c55e022 5407 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
91adb564
N		 5408 mdname(mddev), mddev->new_level);
5409 return ERR_PTR(-EIO);
1da177e4 5410 }
91adb564
N		 5411 if ((mddev->new_level == 5
5412 && !algorithm_valid_raid5(mddev->new_layout)) ||
5413 (mddev->new_level == 6
5414 && !algorithm_valid_raid6(mddev->new_layout))) {
0c55e022 5415 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
91adb564
N		 5416 mdname(mddev), mddev->new_layout);
5417 return ERR_PTR(-EIO);
99c0fb5f 5418 }
91adb564 5419 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
0c55e022 5420 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
91adb564
N		 5421 mdname(mddev), mddev->raid_disks);
5422 return ERR_PTR(-EINVAL);
4bbf3771
N		 5423 }
5424
664e7c41
AN		 5425 if (!mddev->new_chunk_sectors ||
5426 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5427 !is_power_of_2(mddev->new_chunk_sectors)) {
0c55e022
N		 5428 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5429 mdname(mddev), mddev->new_chunk_sectors << 9);
91adb564 5430 return ERR_PTR(-EINVAL);
f6705578
N		 5431 }
5432
d1688a6d 5433 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
91adb564 5434 if (conf == NULL)
1da177e4 5435 goto abort;
851c30c9
SL		 5436 /* Don't enable multi-threading by default*/
5437 if (alloc_thread_groups(conf, 0))
5438 goto abort;
f5efd45a
DW		 5439 spin_lock_init(&conf->device_lock);
5440 init_waitqueue_head(&conf->wait_for_stripe);
5441 init_waitqueue_head(&conf->wait_for_overlap);
5442 INIT_LIST_HEAD(&conf->handle_list);
5443 INIT_LIST_HEAD(&conf->hold_list);
5444 INIT_LIST_HEAD(&conf->delayed_list);
5445 INIT_LIST_HEAD(&conf->bitmap_list);
5446 INIT_LIST_HEAD(&conf->inactive_list);
773ca82f 5447 init_llist_head(&conf->released_stripes);
f5efd45a
DW		 5448 atomic_set(&conf->active_stripes, 0);
5449 atomic_set(&conf->preread_active_stripes, 0);
5450 atomic_set(&conf->active_aligned_reads, 0);
5451 conf->bypass_threshold = BYPASS_THRESHOLD;
d890fa2b 5452 conf->recovery_disabled = mddev->recovery_disabled - 1;
91adb564
N		 5453
5454 conf->raid_disks = mddev->raid_disks;
5455 if (mddev->reshape_position == MaxSector)
5456 conf->previous_raid_disks = mddev->raid_disks;
5457 else
f6705578 5458 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5e5e3e78
N		 5459 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5460 conf->scribble_len = scribble_len(max_disks);
f6705578 5461
5e5e3e78 5462 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
b55e6bfc
N		 5463 GFP_KERNEL);
5464 if (!conf->disks)
5465 goto abort;
9ffae0cf 5466
1da177e4
LT		 5467 conf->mddev = mddev;
5468
fccddba0 5469 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
1da177e4 5470 goto abort;
1da177e4 5471
36d1c647
DW		 5472 conf->level = mddev->new_level;
5473 if (raid5_alloc_percpu(conf) != 0)
5474 goto abort;
5475
0c55e022 5476 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
1da177e4 5477
dafb20fa 5478 rdev_for_each(rdev, mddev) {
1da177e4 5479 raid_disk = rdev->raid_disk;
5e5e3e78 5480 if (raid_disk >= max_disks
1da177e4
LT		 5481 || raid_disk < 0)
5482 continue;
5483 disk = conf->disks + raid_disk;
5484
17045f52
N		 5485 if (test_bit(Replacement, &rdev->flags)) {
5486 if (disk->replacement)
5487 goto abort;
5488 disk->replacement = rdev;
5489 } else {
5490 if (disk->rdev)
5491 goto abort;
5492 disk->rdev = rdev;
5493 }
1da177e4 5494
b2d444d7 5495 if (test_bit(In_sync, &rdev->flags)) {
1da177e4 5496 char b[BDEVNAME_SIZE];
0c55e022
N		 5497 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5498 " disk %d\n",
5499 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
d6b212f4 5500 } else if (rdev->saved_raid_disk != raid_disk)
8c2e870a
NB		 5501 /* Cannot rely on bitmap to complete recovery */
5502 conf->fullsync = 1;
1da177e4
LT		 5503 }
5504
09c9e5fa 5505 conf->chunk_sectors = mddev->new_chunk_sectors;
91adb564 5506 conf->level = mddev->new_level;
16a53ecc
N		 5507 if (conf->level == 6)
5508 conf->max_degraded = 2;
5509 else
5510 conf->max_degraded = 1;
91adb564 5511 conf->algorithm = mddev->new_layout;
1da177e4 5512 conf->max_nr_stripes = NR_STRIPES;
fef9c61f 5513 conf->reshape_progress = mddev->reshape_position;
e183eaed 5514 if (conf->reshape_progress != MaxSector) {
09c9e5fa 5515 conf->prev_chunk_sectors = mddev->chunk_sectors;
e183eaed
N		 5516 conf->prev_algo = mddev->layout;
5517 }
1da177e4 5518
91adb564 5519 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5e5e3e78 5520 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
91adb564
N		 5521 if (grow_stripes(conf, conf->max_nr_stripes)) {
5522 printk(KERN_ERR
0c55e022
N		 5523 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5524 mdname(mddev), memory);
91adb564
N		 5525 goto abort;
5526 } else
0c55e022
N		 5527 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5528 mdname(mddev), memory);
1da177e4 5529
0232605d
N		 5530 sprintf(pers_name, "raid%d", mddev->new_level);
5531 conf->thread = md_register_thread(raid5d, mddev, pers_name);
91adb564
N		 5532 if (!conf->thread) {
5533 printk(KERN_ERR
0c55e022 5534 "md/raid:%s: couldn't allocate thread.\n",
91adb564 5535 mdname(mddev));
16a53ecc
N		 5536 goto abort;
5537 }
91adb564
N		 5538
5539 return conf;
5540
5541 abort:
5542 if (conf) {
95fc17aa 5543 free_conf(conf);
91adb564
N		 5544 return ERR_PTR(-EIO);
5545 } else
5546 return ERR_PTR(-ENOMEM);
5547}
5548
c148ffdc
N		 5549
5550static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5551{
5552 switch (algo) {
5553 case ALGORITHM_PARITY_0:
5554 if (raid_disk < max_degraded)
5555 return 1;
5556 break;
5557 case ALGORITHM_PARITY_N:
5558 if (raid_disk >= raid_disks - max_degraded)
5559 return 1;
5560 break;
5561 case ALGORITHM_PARITY_0_6:
5562 if (raid_disk == 0 ||
5563 raid_disk == raid_disks - 1)
5564 return 1;
5565 break;
5566 case ALGORITHM_LEFT_ASYMMETRIC_6:
5567 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5568 case ALGORITHM_LEFT_SYMMETRIC_6:
5569 case ALGORITHM_RIGHT_SYMMETRIC_6:
5570 if (raid_disk == raid_disks - 1)
5571 return 1;
5572 }
5573 return 0;
5574}
5575
fd01b88c 5576static int run(struct mddev *mddev)
91adb564 5577{
d1688a6d 5578 struct r5conf *conf;
9f7c2220 5579 int working_disks = 0;
c148ffdc 5580 int dirty_parity_disks = 0;
3cb03002 5581 struct md_rdev *rdev;
c148ffdc 5582 sector_t reshape_offset = 0;
17045f52 5583 int i;
b5254dd5
N		 5584 long long min_offset_diff = 0;
5585 int first = 1;
91adb564 5586
8c6ac868 5587 if (mddev->recovery_cp != MaxSector)
0c55e022 5588 printk(KERN_NOTICE "md/raid:%s: not clean"
8c6ac868
AN		 5589 " -- starting background reconstruction\n",
5590 mdname(mddev));
b5254dd5
N		 5591
5592 rdev_for_each(rdev, mddev) {
5593 long long diff;
5594 if (rdev->raid_disk < 0)
5595 continue;
5596 diff = (rdev->new_data_offset - rdev->data_offset);
5597 if (first) {
5598 min_offset_diff = diff;
5599 first = 0;
5600 } else if (mddev->reshape_backwards &&
5601 diff < min_offset_diff)
5602 min_offset_diff = diff;
5603 else if (!mddev->reshape_backwards &&
5604 diff > min_offset_diff)
5605 min_offset_diff = diff;
5606 }
5607
91adb564
N		 5608 if (mddev->reshape_position != MaxSector) {
5609 /* Check that we can continue the reshape.
b5254dd5
N		 5610 * Difficulties arise if the stripe we would write to
5611 * next is at or after the stripe we would read from next.
5612 * For a reshape that changes the number of devices, this
5613 * is only possible for a very short time, and mdadm makes
5614 * sure that time appears to have past before assembling
5615 * the array. So we fail if that time hasn't passed.
5616 * For a reshape that keeps the number of devices the same
5617 * mdadm must be monitoring the reshape can keeping the
5618 * critical areas read-only and backed up. It will start
5619 * the array in read-only mode, so we check for that.
91adb564
N		 5620 */
5621 sector_t here_new, here_old;
5622 int old_disks;
18b00334 5623 int max_degraded = (mddev->level == 6 ? 2 : 1);
91adb564 5624
88ce4930 5625 if (mddev->new_level != mddev->level) {
0c55e022 5626 printk(KERN_ERR "md/raid:%s: unsupported reshape "
91adb564
N		 5627 "required - aborting.\n",
5628 mdname(mddev));
5629 return -EINVAL;
5630 }
91adb564
N		 5631 old_disks = mddev->raid_disks - mddev->delta_disks;
5632 /* reshape_position must be on a new-stripe boundary, and one
5633 * further up in new geometry must map after here in old
5634 * geometry.
5635 */
5636 here_new = mddev->reshape_position;
664e7c41 5637 if (sector_div(here_new, mddev->new_chunk_sectors *
91adb564 5638 (mddev->raid_disks - max_degraded))) {
0c55e022
N		 5639 printk(KERN_ERR "md/raid:%s: reshape_position not "
5640 "on a stripe boundary\n", mdname(mddev));
91adb564
N		 5641 return -EINVAL;
5642 }
c148ffdc 5643 reshape_offset = here_new * mddev->new_chunk_sectors;
91adb564
N		 5644 /* here_new is the stripe we will write to */
5645 here_old = mddev->reshape_position;
9d8f0363 5646 sector_div(here_old, mddev->chunk_sectors *
91adb564
N		 5647 (old_disks-max_degraded));
5648 /* here_old is the first stripe that we might need to read
5649 * from */
67ac6011 5650 if (mddev->delta_disks == 0) {
b5254dd5
N		 5651 if ((here_new * mddev->new_chunk_sectors !=
5652 here_old * mddev->chunk_sectors)) {
5653 printk(KERN_ERR "md/raid:%s: reshape position is"
5654 " confused - aborting\n", mdname(mddev));
5655 return -EINVAL;
5656 }
67ac6011 5657 /* We cannot be sure it is safe to start an in-place
b5254dd5 5658 * reshape. It is only safe if user-space is monitoring
67ac6011
N		 5659 * and taking constant backups.
5660 * mdadm always starts a situation like this in
5661 * readonly mode so it can take control before
5662 * allowing any writes. So just check for that.
5663 */
b5254dd5
N		 5664 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5665 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5666 /* not really in-place - so OK */;
5667 else if (mddev->ro == 0) {
5668 printk(KERN_ERR "md/raid:%s: in-place reshape "
5669 "must be started in read-only mode "
5670 "- aborting\n",
0c55e022 5671 mdname(mddev));
67ac6011
N		 5672 return -EINVAL;
5673 }
2c810cdd 5674 } else if (mddev->reshape_backwards
b5254dd5 5675 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
67ac6011
N		 5676 here_old * mddev->chunk_sectors)
5677 : (here_new * mddev->new_chunk_sectors >=
b5254dd5 5678 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
91adb564 5679 /* Reading from the same stripe as writing to - bad */
0c55e022
N		 5680 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5681 "auto-recovery - aborting.\n",
5682 mdname(mddev));
91adb564
N		 5683 return -EINVAL;
5684 }
0c55e022
N		 5685 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5686 mdname(mddev));
91adb564
N		 5687 /* OK, we should be able to continue; */
5688 } else {
5689 BUG_ON(mddev->level != mddev->new_level);
5690 BUG_ON(mddev->layout != mddev->new_layout);
664e7c41 5691 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
91adb564 5692 BUG_ON(mddev->delta_disks != 0);
1da177e4 5693 }
91adb564 5694
245f46c2
N		 5695 if (mddev->private == NULL)
5696 conf = setup_conf(mddev);
5697 else
5698 conf = mddev->private;
5699
91adb564
N		 5700 if (IS_ERR(conf))
5701 return PTR_ERR(conf);
5702
b5254dd5 5703 conf->min_offset_diff = min_offset_diff;
91adb564
N		 5704 mddev->thread = conf->thread;
5705 conf->thread = NULL;
5706 mddev->private = conf;
5707
17045f52
N		 5708 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5709 i++) {
5710 rdev = conf->disks[i].rdev;
5711 if (!rdev && conf->disks[i].replacement) {
5712 /* The replacement is all we have yet */
5713 rdev = conf->disks[i].replacement;
5714 conf->disks[i].replacement = NULL;
5715 clear_bit(Replacement, &rdev->flags);
5716 conf->disks[i].rdev = rdev;
5717 }
5718 if (!rdev)
c148ffdc 5719 continue;
17045f52
N		 5720 if (conf->disks[i].replacement &&
5721 conf->reshape_progress != MaxSector) {
5722 /* replacements and reshape simply do not mix. */
5723 printk(KERN_ERR "md: cannot handle concurrent "
5724 "replacement and reshape.\n");
5725 goto abort;
5726 }
2f115882 5727 if (test_bit(In_sync, &rdev->flags)) {
91adb564 5728 working_disks++;
2f115882
N		 5729 continue;
5730 }
c148ffdc
N		 5731 /* This disc is not fully in-sync. However if it
5732 * just stored parity (beyond the recovery_offset),
5733 * when we don't need to be concerned about the
5734 * array being dirty.
5735 * When reshape goes 'backwards', we never have
5736 * partially completed devices, so we only need
5737 * to worry about reshape going forwards.
5738 */
5739 /* Hack because v0.91 doesn't store recovery_offset properly. */
5740 if (mddev->major_version == 0 &&
5741 mddev->minor_version > 90)
5742 rdev->recovery_offset = reshape_offset;
5026d7a9 5743
c148ffdc
N		 5744 if (rdev->recovery_offset < reshape_offset) {
5745 /* We need to check old and new layout */
5746 if (!only_parity(rdev->raid_disk,
5747 conf->algorithm,
5748 conf->raid_disks,
5749 conf->max_degraded))
5750 continue;
5751 }
5752 if (!only_parity(rdev->raid_disk,
5753 conf->prev_algo,
5754 conf->previous_raid_disks,
5755 conf->max_degraded))
5756 continue;
5757 dirty_parity_disks++;
5758 }
91adb564 5759
17045f52
N		 5760 /*
5761 * 0 for a fully functional array, 1 or 2 for a degraded array.
5762 */
908f4fbd 5763 mddev->degraded = calc_degraded(conf);
91adb564 5764
674806d6 5765 if (has_failed(conf)) {
0c55e022 5766 printk(KERN_ERR "md/raid:%s: not enough operational devices"
1da177e4 5767 " (%d/%d failed)\n",
02c2de8c 5768 mdname(mddev), mddev->degraded, conf->raid_disks);
1da177e4
LT		 5769 goto abort;
5770 }
5771
91adb564 5772 /* device size must be a multiple of chunk size */
9d8f0363 5773 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
91adb564
N		 5774 mddev->resync_max_sectors = mddev->dev_sectors;
5775
c148ffdc 5776 if (mddev->degraded > dirty_parity_disks &&
1da177e4 5777 mddev->recovery_cp != MaxSector) {
6ff8d8ec
N		 5778 if (mddev->ok_start_degraded)
5779 printk(KERN_WARNING
0c55e022
N		 5780 "md/raid:%s: starting dirty degraded array"
5781 " - data corruption possible.\n",
6ff8d8ec
N		 5782 mdname(mddev));
5783 else {
5784 printk(KERN_ERR
0c55e022 5785 "md/raid:%s: cannot start dirty degraded array.\n",
6ff8d8ec
N		 5786 mdname(mddev));
5787 goto abort;
5788 }
1da177e4
LT		 5789 }
5790
1da177e4 5791 if (mddev->degraded == 0)
0c55e022
N		 5792 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5793 " devices, algorithm %d\n", mdname(mddev), conf->level,
e183eaed
N		 5794 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5795 mddev->new_layout);
1da177e4 5796 else
0c55e022
N		 5797 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5798 " out of %d devices, algorithm %d\n",
5799 mdname(mddev), conf->level,
5800 mddev->raid_disks - mddev->degraded,
5801 mddev->raid_disks, mddev->new_layout);
1da177e4
LT		 5802
5803 print_raid5_conf(conf);
5804
fef9c61f 5805 if (conf->reshape_progress != MaxSector) {
fef9c61f 5806 conf->reshape_safe = conf->reshape_progress;
f6705578
N		 5807 atomic_set(&conf->reshape_stripes, 0);
5808 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5809 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5810 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5811 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5812 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
0da3c619 5813 "reshape");
f6705578
N		 5814 }
5815
1da177e4
LT		 5816
5817 /* Ok, everything is just fine now */
a64c876f
N		 5818 if (mddev->to_remove == &raid5_attrs_group)
5819 mddev->to_remove = NULL;
00bcb4ac
N		 5820 else if (mddev->kobj.sd &&
5821 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5e55e2f5 5822 printk(KERN_WARNING
4a5add49 5823 "raid5: failed to create sysfs attributes for %s\n",
5e55e2f5 5824 mdname(mddev));
4a5add49 5825 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7a5febe9 5826
4a5add49 5827 if (mddev->queue) {
9f7c2220 5828 int chunk_size;
620125f2 5829 bool discard_supported = true;
4a5add49
N		 5830 /* read-ahead size must cover two whole stripes, which
5831 * is 2 * (datadisks) * chunksize where 'n' is the
5832 * number of raid devices
5833 */
5834 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5835 int stripe = data_disks *
5836 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5837 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5838 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
91adb564 5839
4a5add49 5840 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
f022b2fd 5841
11d8a6e3
N		 5842 mddev->queue->backing_dev_info.congested_data = mddev;
5843 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
7a5febe9 5844
9f7c2220
N		 5845 chunk_size = mddev->chunk_sectors << 9;
5846 blk_queue_io_min(mddev->queue, chunk_size);
5847 blk_queue_io_opt(mddev->queue, chunk_size *
5848 (conf->raid_disks - conf->max_degraded));
620125f2
SL		 5849 /*
5850 * We can only discard a whole stripe. It doesn't make sense to
5851 * discard data disk but write parity disk
5852 */
5853 stripe = stripe * PAGE_SIZE;
4ac6875e
N		 5854 /* Round up to power of 2, as discard handling
5855 * currently assumes that */
5856 while ((stripe-1) & stripe)
5857 stripe = (stripe | (stripe-1)) + 1;
620125f2
SL		 5858 mddev->queue->limits.discard_alignment = stripe;
5859 mddev->queue->limits.discard_granularity = stripe;
5860 /*
5861 * unaligned part of discard request will be ignored, so can't
5862 * guarantee discard_zerors_data
5863 */
5864 mddev->queue->limits.discard_zeroes_data = 0;
8f6c2e4b 5865
5026d7a9
PA		 5866 blk_queue_max_write_same_sectors(mddev->queue, 0);
5867
05616be5 5868 rdev_for_each(rdev, mddev) {
9f7c2220
N		 5869 disk_stack_limits(mddev->gendisk, rdev->bdev,
5870 rdev->data_offset << 9);
05616be5
N		 5871 disk_stack_limits(mddev->gendisk, rdev->bdev,
5872 rdev->new_data_offset << 9);
620125f2
SL		 5873 /*
5874 * discard_zeroes_data is required, otherwise data
5875 * could be lost. Consider a scenario: discard a stripe
5876 * (the stripe could be inconsistent if
5877 * discard_zeroes_data is 0); write one disk of the
5878 * stripe (the stripe could be inconsistent again
5879 * depending on which disks are used to calculate
5880 * parity); the disk is broken; The stripe data of this
5881 * disk is lost.
5882 */
5883 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5884 !bdev_get_queue(rdev->bdev)->
5885 limits.discard_zeroes_data)
5886 discard_supported = false;
05616be5 5887 }
620125f2
SL		 5888
5889 if (discard_supported &&
5890 mddev->queue->limits.max_discard_sectors >= stripe &&
5891 mddev->queue->limits.discard_granularity >= stripe)
5892 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5893 mddev->queue);
5894 else
5895 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5896 mddev->queue);
9f7c2220 5897 }
23032a0e 5898
1da177e4
LT		 5899 return 0;
5900abort:
01f96c0a 5901 md_unregister_thread(&mddev->thread);
e4f869d9
N		 5902 print_raid5_conf(conf);
5903 free_conf(conf);
1da177e4 5904 mddev->private = NULL;
0c55e022 5905 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
1da177e4
LT		 5906 return -EIO;
5907}
5908
fd01b88c 5909static int stop(struct mddev *mddev)
1da177e4 5910{
d1688a6d 5911 struct r5conf *conf = mddev->private;
1da177e4 5912
01f96c0a 5913 md_unregister_thread(&mddev->thread);
11d8a6e3
N		 5914 if (mddev->queue)
5915 mddev->queue->backing_dev_info.congested_fn = NULL;
95fc17aa 5916 free_conf(conf);
a64c876f
N		 5917 mddev->private = NULL;
5918 mddev->to_remove = &raid5_attrs_group;
1da177e4
LT		 5919 return 0;
5920}
5921
fd01b88c 5922static void status(struct seq_file *seq, struct mddev *mddev)
1da177e4 5923{
d1688a6d 5924 struct r5conf *conf = mddev->private;
1da177e4
LT		 5925 int i;
5926
9d8f0363
AN		 5927 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5928 mddev->chunk_sectors / 2, mddev->layout);
02c2de8c 5929 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
1da177e4
LT		 5930 for (i = 0; i < conf->raid_disks; i++)
5931 seq_printf (seq, "%s",
5932 conf->disks[i].rdev &&
b2d444d7 5933 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
1da177e4 5934 seq_printf (seq, "]");
1da177e4
LT		 5935}
5936
d1688a6d 5937static void print_raid5_conf (struct r5conf *conf)
1da177e4
LT		 5938{
5939 int i;
5940 struct disk_info *tmp;
5941
0c55e022 5942 printk(KERN_DEBUG "RAID conf printout:\n");
1da177e4
LT		 5943 if (!conf) {
5944 printk("(conf==NULL)\n");
5945 return;
5946 }
0c55e022
N		 5947 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5948 conf->raid_disks,
5949 conf->raid_disks - conf->mddev->degraded);
1da177e4
LT		 5950
5951 for (i = 0; i < conf->raid_disks; i++) {
5952 char b[BDEVNAME_SIZE];
5953 tmp = conf->disks + i;
5954 if (tmp->rdev)
0c55e022
N		 5955 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5956 i, !test_bit(Faulty, &tmp->rdev->flags),
5957 bdevname(tmp->rdev->bdev, b));
1da177e4
LT		 5958 }
5959}
5960
fd01b88c 5961static int raid5_spare_active(struct mddev *mddev)
1da177e4
LT		 5962{
5963 int i;
d1688a6d 5964 struct r5conf *conf = mddev->private;
1da177e4 5965 struct disk_info *tmp;
6b965620
N		 5966 int count = 0;
5967 unsigned long flags;
1da177e4
LT		 5968
5969 for (i = 0; i < conf->raid_disks; i++) {
5970 tmp = conf->disks + i;
dd054fce
N		 5971 if (tmp->replacement
5972 && tmp->replacement->recovery_offset == MaxSector
5973 && !test_bit(Faulty, &tmp->replacement->flags)
5974 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5975 /* Replacement has just become active. */
5976 if (!tmp->rdev
5977 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5978 count++;
5979 if (tmp->rdev) {
5980 /* Replaced device not technically faulty,
5981 * but we need to be sure it gets removed
5982 * and never re-added.
5983 */
5984 set_bit(Faulty, &tmp->rdev->flags);
5985 sysfs_notify_dirent_safe(
5986 tmp->rdev->sysfs_state);
5987 }
5988 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5989 } else if (tmp->rdev
70fffd0b 5990 && tmp->rdev->recovery_offset == MaxSector
b2d444d7 5991 && !test_bit(Faulty, &tmp->rdev->flags)
c04be0aa 5992 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
6b965620 5993 count++;
43c73ca4 5994 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1da177e4
LT		 5995 }
5996 }
6b965620 5997 spin_lock_irqsave(&conf->device_lock, flags);
908f4fbd 5998 mddev->degraded = calc_degraded(conf);
6b965620 5999 spin_unlock_irqrestore(&conf->device_lock, flags);
1da177e4 6000 print_raid5_conf(conf);
6b965620 6001 return count;
1da177e4
LT		 6002}
6003
b8321b68 6004static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1da177e4 6005{
d1688a6d 6006 struct r5conf *conf = mddev->private;
1da177e4 6007 int err = 0;
b8321b68 6008 int number = rdev->raid_disk;
657e3e4d 6009 struct md_rdev **rdevp;
1da177e4
LT		 6010 struct disk_info *p = conf->disks + number;
6011
6012 print_raid5_conf(conf);
657e3e4d
N		 6013 if (rdev == p->rdev)
6014 rdevp = &p->rdev;
6015 else if (rdev == p->replacement)
6016 rdevp = &p->replacement;
6017 else
6018 return 0;
6019
6020 if (number >= conf->raid_disks &&
6021 conf->reshape_progress == MaxSector)
6022 clear_bit(In_sync, &rdev->flags);
6023
6024 if (test_bit(In_sync, &rdev->flags) ||
6025 atomic_read(&rdev->nr_pending)) {
6026 err = -EBUSY;
6027 goto abort;
6028 }
6029 /* Only remove non-faulty devices if recovery
6030 * isn't possible.
6031 */
6032 if (!test_bit(Faulty, &rdev->flags) &&
6033 mddev->recovery_disabled != conf->recovery_disabled &&
6034 !has_failed(conf) &&
dd054fce 6035 (!p->replacement || p->replacement == rdev) &&
657e3e4d
N		 6036 number < conf->raid_disks) {
6037 err = -EBUSY;
6038 goto abort;
6039 }
6040 *rdevp = NULL;
6041 synchronize_rcu();
6042 if (atomic_read(&rdev->nr_pending)) {
6043 /* lost the race, try later */
6044 err = -EBUSY;
6045 *rdevp = rdev;
dd054fce
N		 6046 } else if (p->replacement) {
6047 /* We must have just cleared 'rdev' */
6048 p->rdev = p->replacement;
6049 clear_bit(Replacement, &p->replacement->flags);
6050 smp_mb(); /* Make sure other CPUs may see both as identical
6051 * but will never see neither - if they are careful
6052 */
6053 p->replacement = NULL;
6054 clear_bit(WantReplacement, &rdev->flags);
6055 } else
6056 /* We might have just removed the Replacement as faulty-
6057 * clear the bit just in case
6058 */
6059 clear_bit(WantReplacement, &rdev->flags);
1da177e4
LT		 6060abort:
6061
6062 print_raid5_conf(conf);
6063 return err;
6064}
6065
fd01b88c 6066static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1da177e4 6067{
d1688a6d 6068 struct r5conf *conf = mddev->private;
199050ea 6069 int err = -EEXIST;
1da177e4
LT		 6070 int disk;
6071 struct disk_info *p;
6c2fce2e
NB		 6072 int first = 0;
6073 int last = conf->raid_disks - 1;
1da177e4 6074
7f0da59b
N		 6075 if (mddev->recovery_disabled == conf->recovery_disabled)
6076 return -EBUSY;
6077
dc10c643 6078 if (rdev->saved_raid_disk < 0 && has_failed(conf))
1da177e4 6079 /* no point adding a device */
199050ea 6080 return -EINVAL;
1da177e4 6081
6c2fce2e
NB		 6082 if (rdev->raid_disk >= 0)
6083 first = last = rdev->raid_disk;
1da177e4
LT		 6084
6085 /*
16a53ecc
N		 6086 * find the disk ... but prefer rdev->saved_raid_disk
6087 * if possible.
1da177e4 6088 */
16a53ecc 6089 if (rdev->saved_raid_disk >= 0 &&
6c2fce2e 6090 rdev->saved_raid_disk >= first &&
16a53ecc 6091 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5cfb22a1
N		 6092 first = rdev->saved_raid_disk;
6093
6094 for (disk = first; disk <= last; disk++) {
7bfec5f3
N		 6095 p = conf->disks + disk;
6096 if (p->rdev == NULL) {
b2d444d7 6097 clear_bit(In_sync, &rdev->flags);
1da177e4 6098 rdev->raid_disk = disk;
199050ea 6099 err = 0;
72626685
N		 6100 if (rdev->saved_raid_disk != disk)
6101 conf->fullsync = 1;
d6065f7b 6102 rcu_assign_pointer(p->rdev, rdev);
5cfb22a1 6103 goto out;
1da177e4 6104 }
5cfb22a1
N		 6105 }
6106 for (disk = first; disk <= last; disk++) {
6107 p = conf->disks + disk;
7bfec5f3
N		 6108 if (test_bit(WantReplacement, &p->rdev->flags) &&
6109 p->replacement == NULL) {
6110 clear_bit(In_sync, &rdev->flags);
6111 set_bit(Replacement, &rdev->flags);
6112 rdev->raid_disk = disk;
6113 err = 0;
6114 conf->fullsync = 1;
6115 rcu_assign_pointer(p->replacement, rdev);
6116 break;
6117 }
6118 }
5cfb22a1 6119out:
1da177e4 6120 print_raid5_conf(conf);
199050ea 6121 return err;
1da177e4
LT		 6122}
6123
fd01b88c 6124static int raid5_resize(struct mddev *mddev, sector_t sectors)
1da177e4
LT		 6125{
6126 /* no resync is happening, and there is enough space
6127 * on all devices, so we can resize.
6128 * We need to make sure resync covers any new space.
6129 * If the array is shrinking we should possibly wait until
6130 * any io in the removed space completes, but it hardly seems
6131 * worth it.
6132 */
a4a6125a 6133 sector_t newsize;
9d8f0363 6134 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
a4a6125a
N		 6135 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
6136 if (mddev->external_size &&
6137 mddev->array_sectors > newsize)
b522adcd 6138 return -EINVAL;
a4a6125a
N		 6139 if (mddev->bitmap) {
6140 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
6141 if (ret)
6142 return ret;
6143 }
6144 md_set_array_sectors(mddev, newsize);
f233ea5c 6145 set_capacity(mddev->gendisk, mddev->array_sectors);
449aad3e 6146 revalidate_disk(mddev->gendisk);
b098636c
N		 6147 if (sectors > mddev->dev_sectors &&
6148 mddev->recovery_cp > mddev->dev_sectors) {
58c0fed4 6149 mddev->recovery_cp = mddev->dev_sectors;
1da177e4
LT		 6150 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
6151 }
58c0fed4 6152 mddev->dev_sectors = sectors;
4b5c7ae8 6153 mddev->resync_max_sectors = sectors;
1da177e4
LT		 6154 return 0;
6155}
6156
fd01b88c 6157static int check_stripe_cache(struct mddev *mddev)
01ee22b4
N		 6158{
6159 /* Can only proceed if there are plenty of stripe_heads.
6160 * We need a minimum of one full stripe,, and for sensible progress
6161 * it is best to have about 4 times that.
6162 * If we require 4 times, then the default 256 4K stripe_heads will
6163 * allow for chunk sizes up to 256K, which is probably OK.
6164 * If the chunk size is greater, user-space should request more
6165 * stripe_heads first.
6166 */
d1688a6d 6167 struct r5conf *conf = mddev->private;
01ee22b4
N		 6168 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
6169 > conf->max_nr_stripes ||
6170 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
6171 > conf->max_nr_stripes) {
0c55e022
N		 6172 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
6173 mdname(mddev),
01ee22b4
N		 6174 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
6175 / STRIPE_SIZE)*4);
6176 return 0;
6177 }
6178 return 1;
6179}
6180
fd01b88c 6181static int check_reshape(struct mddev *mddev)
29269553 6182{
d1688a6d 6183 struct r5conf *conf = mddev->private;
29269553 6184
88ce4930
N		 6185 if (mddev->delta_disks == 0 &&
6186 mddev->new_layout == mddev->layout &&
664e7c41 6187 mddev->new_chunk_sectors == mddev->chunk_sectors)
50ac168a 6188 return 0; /* nothing to do */
674806d6 6189 if (has_failed(conf))
ec32a2bd 6190 return -EINVAL;
fdcfbbb6 6191 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
ec32a2bd
N		 6192 /* We might be able to shrink, but the devices must
6193 * be made bigger first.
6194 * For raid6, 4 is the minimum size.
6195 * Otherwise 2 is the minimum
6196 */
6197 int min = 2;
6198 if (mddev->level == 6)
6199 min = 4;
6200 if (mddev->raid_disks + mddev->delta_disks < min)
6201 return -EINVAL;
6202 }
29269553 6203
01ee22b4 6204 if (!check_stripe_cache(mddev))
29269553 6205 return -ENOSPC;
29269553 6206
e56108d6
N		 6207 return resize_stripes(conf, (conf->previous_raid_disks
6208 + mddev->delta_disks));
63c70c4f
N		 6209}
6210
fd01b88c 6211static int raid5_start_reshape(struct mddev *mddev)
63c70c4f 6212{
d1688a6d 6213 struct r5conf *conf = mddev->private;
3cb03002 6214 struct md_rdev *rdev;
63c70c4f 6215 int spares = 0;
c04be0aa 6216 unsigned long flags;
63c70c4f 6217
f416885e 6218 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
63c70c4f
N		 6219 return -EBUSY;
6220
01ee22b4
N		 6221 if (!check_stripe_cache(mddev))
6222 return -ENOSPC;
6223
30b67645
N		 6224 if (has_failed(conf))
6225 return -EINVAL;
6226
c6563a8c 6227 rdev_for_each(rdev, mddev) {
469518a3
N		 6228 if (!test_bit(In_sync, &rdev->flags)
6229 && !test_bit(Faulty, &rdev->flags))
29269553 6230 spares++;
c6563a8c 6231 }
63c70c4f 6232
f416885e 6233 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
29269553
N		 6234 /* Not enough devices even to make a degraded array
6235 * of that size
6236 */
6237 return -EINVAL;
6238
ec32a2bd
N		 6239 /* Refuse to reduce size of the array. Any reductions in
6240 * array size must be through explicit setting of array_size
6241 * attribute.
6242 */
6243 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
6244 < mddev->array_sectors) {
0c55e022 6245 printk(KERN_ERR "md/raid:%s: array size must be reduced "
ec32a2bd
N		 6246 "before number of disks\n", mdname(mddev));
6247 return -EINVAL;
6248 }
6249
f6705578 6250 atomic_set(&conf->reshape_stripes, 0);
29269553
N		 6251 spin_lock_irq(&conf->device_lock);
6252 conf->previous_raid_disks = conf->raid_disks;
63c70c4f 6253 conf->raid_disks += mddev->delta_disks;
09c9e5fa
AN		 6254 conf->prev_chunk_sectors = conf->chunk_sectors;
6255 conf->chunk_sectors = mddev->new_chunk_sectors;
88ce4930
N		 6256 conf->prev_algo = conf->algorithm;
6257 conf->algorithm = mddev->new_layout;
05616be5
N		 6258 conf->generation++;
6259 /* Code that selects data_offset needs to see the generation update
6260 * if reshape_progress has been set - so a memory barrier needed.
6261 */
6262 smp_mb();
2c810cdd 6263 if (mddev->reshape_backwards)
fef9c61f
N		 6264 conf->reshape_progress = raid5_size(mddev, 0, 0);
6265 else
6266 conf->reshape_progress = 0;
6267 conf->reshape_safe = conf->reshape_progress;
29269553
N		 6268 spin_unlock_irq(&conf->device_lock);
6269
6270 /* Add some new drives, as many as will fit.
6271 * We know there are enough to make the newly sized array work.
3424bf6a
N		 6272 * Don't add devices if we are reducing the number of
6273 * devices in the array. This is because it is not possible
6274 * to correctly record the "partially reconstructed" state of
6275 * such devices during the reshape and confusion could result.
29269553 6276 */
87a8dec9 6277 if (mddev->delta_disks >= 0) {
dafb20fa 6278 rdev_for_each(rdev, mddev)
87a8dec9
N		 6279 if (rdev->raid_disk < 0 &&
6280 !test_bit(Faulty, &rdev->flags)) {
6281 if (raid5_add_disk(mddev, rdev) == 0) {
87a8dec9 6282 if (rdev->raid_disk
9d4c7d87 6283 >= conf->previous_raid_disks)
87a8dec9 6284 set_bit(In_sync, &rdev->flags);
9d4c7d87 6285 else
87a8dec9 6286 rdev->recovery_offset = 0;
36fad858
NK		 6287
6288 if (sysfs_link_rdev(mddev, rdev))
87a8dec9 6289 /* Failure here is OK */;
50da0840 6290 }
87a8dec9
N		 6291 } else if (rdev->raid_disk >= conf->previous_raid_disks
6292 && !test_bit(Faulty, &rdev->flags)) {
6293 /* This is a spare that was manually added */
6294 set_bit(In_sync, &rdev->flags);
87a8dec9 6295 }
29269553 6296
87a8dec9
N		 6297 /* When a reshape changes the number of devices,
6298 * ->degraded is measured against the larger of the
6299 * pre and post number of devices.
6300 */
ec32a2bd 6301 spin_lock_irqsave(&conf->device_lock, flags);
908f4fbd 6302 mddev->degraded = calc_degraded(conf);
ec32a2bd
N		 6303 spin_unlock_irqrestore(&conf->device_lock, flags);
6304 }
63c70c4f 6305 mddev->raid_disks = conf->raid_disks;
e516402c 6306 mddev->reshape_position = conf->reshape_progress;
850b2b42 6307 set_bit(MD_CHANGE_DEVS, &mddev->flags);
f6705578 6308
29269553
N		 6309 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6310 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6311 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6312 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6313 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
0da3c619 6314 "reshape");
29269553
N		 6315 if (!mddev->sync_thread) {
6316 mddev->recovery = 0;
6317 spin_lock_irq(&conf->device_lock);
6318 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
05616be5
N		 6319 rdev_for_each(rdev, mddev)
6320 rdev->new_data_offset = rdev->data_offset;
6321 smp_wmb();
fef9c61f 6322 conf->reshape_progress = MaxSector;
1e3fa9bd 6323 mddev->reshape_position = MaxSector;
29269553
N		 6324 spin_unlock_irq(&conf->device_lock);
6325 return -EAGAIN;
6326 }
c8f517c4 6327 conf->reshape_checkpoint = jiffies;
29269553
N		 6328 md_wakeup_thread(mddev->sync_thread);
6329 md_new_event(mddev);
6330 return 0;
6331}
29269553 6332
ec32a2bd
N		 6333/* This is called from the reshape thread and should make any
6334 * changes needed in 'conf'
6335 */
d1688a6d 6336static void end_reshape(struct r5conf *conf)
29269553 6337{
29269553 6338
f6705578 6339 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
05616be5 6340 struct md_rdev *rdev;
f6705578 6341
f6705578 6342 spin_lock_irq(&conf->device_lock);
cea9c228 6343 conf->previous_raid_disks = conf->raid_disks;
05616be5
N		 6344 rdev_for_each(rdev, conf->mddev)
6345 rdev->data_offset = rdev->new_data_offset;
6346 smp_wmb();
fef9c61f 6347 conf->reshape_progress = MaxSector;
f6705578 6348 spin_unlock_irq(&conf->device_lock);
b0f9ec04 6349 wake_up(&conf->wait_for_overlap);
16a53ecc
N		 6350
6351 /* read-ahead size must cover two whole stripes, which is
6352 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6353 */
4a5add49 6354 if (conf->mddev->queue) {
cea9c228 6355 int data_disks = conf->raid_disks - conf->max_degraded;
09c9e5fa 6356 int stripe = data_disks * ((conf->chunk_sectors << 9)
cea9c228 6357 / PAGE_SIZE);
16a53ecc
N		 6358 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6359 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6360 }
29269553 6361 }
29269553
N		 6362}
6363
ec32a2bd
N		 6364/* This is called from the raid5d thread with mddev_lock held.
6365 * It makes config changes to the device.
6366 */
fd01b88c 6367static void raid5_finish_reshape(struct mddev *mddev)
cea9c228 6368{
d1688a6d 6369 struct r5conf *conf = mddev->private;
cea9c228
N		 6370
6371 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6372
ec32a2bd
N		 6373 if (mddev->delta_disks > 0) {
6374 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6375 set_capacity(mddev->gendisk, mddev->array_sectors);
449aad3e 6376 revalidate_disk(mddev->gendisk);
ec32a2bd
N		 6377 } else {
6378 int d;
908f4fbd
N		 6379 spin_lock_irq(&conf->device_lock);
6380 mddev->degraded = calc_degraded(conf);
6381 spin_unlock_irq(&conf->device_lock);
ec32a2bd
N		 6382 for (d = conf->raid_disks ;
6383 d < conf->raid_disks - mddev->delta_disks;
1a67dde0 6384 d++) {
3cb03002 6385 struct md_rdev *rdev = conf->disks[d].rdev;
da7613b8
N		 6386 if (rdev)
6387 clear_bit(In_sync, &rdev->flags);
6388 rdev = conf->disks[d].replacement;
6389 if (rdev)
6390 clear_bit(In_sync, &rdev->flags);
1a67dde0 6391 }
cea9c228 6392 }
88ce4930 6393 mddev->layout = conf->algorithm;
09c9e5fa 6394 mddev->chunk_sectors = conf->chunk_sectors;
ec32a2bd
N		 6395 mddev->reshape_position = MaxSector;
6396 mddev->delta_disks = 0;
2c810cdd 6397 mddev->reshape_backwards = 0;
cea9c228
N		 6398 }
6399}
6400
fd01b88c 6401static void raid5_quiesce(struct mddev *mddev, int state)
72626685 6402{
d1688a6d 6403 struct r5conf *conf = mddev->private;
72626685
N		 6404
6405 switch(state) {
e464eafd
N		 6406 case 2: /* resume for a suspend */
6407 wake_up(&conf->wait_for_overlap);
6408 break;
6409
72626685
N		 6410 case 1: /* stop all writes */
6411 spin_lock_irq(&conf->device_lock);
64bd660b
N		 6412 /* '2' tells resync/reshape to pause so that all
6413 * active stripes can drain
6414 */
6415 conf->quiesce = 2;
72626685 6416 wait_event_lock_irq(conf->wait_for_stripe,
46031f9a
RBJ		 6417 atomic_read(&conf->active_stripes) == 0 &&
6418 atomic_read(&conf->active_aligned_reads) == 0,
eed8c02e 6419 conf->device_lock);
64bd660b 6420 conf->quiesce = 1;
72626685 6421 spin_unlock_irq(&conf->device_lock);
64bd660b
N		 6422 /* allow reshape to continue */
6423 wake_up(&conf->wait_for_overlap);
72626685
N		 6424 break;
6425
6426 case 0: /* re-enable writes */
6427 spin_lock_irq(&conf->device_lock);
6428 conf->quiesce = 0;
6429 wake_up(&conf->wait_for_stripe);
e464eafd 6430 wake_up(&conf->wait_for_overlap);
72626685
N		 6431 spin_unlock_irq(&conf->device_lock);
6432 break;
6433 }
72626685 6434}
b15c2e57 6435
d562b0c4 6436
fd01b88c 6437static void *raid45_takeover_raid0(struct mddev *mddev, int level)
54071b38 6438{
e373ab10 6439 struct r0conf *raid0_conf = mddev->private;
d76c8420 6440 sector_t sectors;
54071b38 6441
f1b29bca 6442 /* for raid0 takeover only one zone is supported */
e373ab10 6443 if (raid0_conf->nr_strip_zones > 1) {
0c55e022
N		 6444 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6445 mdname(mddev));
f1b29bca
DW		 6446 return ERR_PTR(-EINVAL);
6447 }
6448
e373ab10
N		 6449 sectors = raid0_conf->strip_zone[0].zone_end;
6450 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
3b71bd93 6451 mddev->dev_sectors = sectors;
f1b29bca 6452 mddev->new_level = level;
54071b38
TM		 6453 mddev->new_layout = ALGORITHM_PARITY_N;
6454 mddev->new_chunk_sectors = mddev->chunk_sectors;
6455 mddev->raid_disks += 1;
6456 mddev->delta_disks = 1;
6457 /* make sure it will be not marked as dirty */
6458 mddev->recovery_cp = MaxSector;
6459
6460 return setup_conf(mddev);
6461}
6462
6463
fd01b88c 6464static void *raid5_takeover_raid1(struct mddev *mddev)
d562b0c4
N		 6465{
6466 int chunksect;
6467
6468 if (mddev->raid_disks != 2 ||
6469 mddev->degraded > 1)
6470 return ERR_PTR(-EINVAL);
6471
6472 /* Should check if there are write-behind devices? */
6473
6474 chunksect = 64*2; /* 64K by default */
6475
6476 /* The array must be an exact multiple of chunksize */
6477 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6478 chunksect >>= 1;
6479
6480 if ((chunksect<<9) < STRIPE_SIZE)
6481 /* array size does not allow a suitable chunk size */
6482 return ERR_PTR(-EINVAL);
6483
6484 mddev->new_level = 5;
6485 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
664e7c41 6486 mddev->new_chunk_sectors = chunksect;
d562b0c4
N		 6487
6488 return setup_conf(mddev);
6489}
6490
fd01b88c 6491static void *raid5_takeover_raid6(struct mddev *mddev)
fc9739c6
N		 6492{
6493 int new_layout;
6494
6495 switch (mddev->layout) {
6496 case ALGORITHM_LEFT_ASYMMETRIC_6:
6497 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6498 break;
6499 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6500 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6501 break;
6502 case ALGORITHM_LEFT_SYMMETRIC_6:
6503 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6504 break;
6505 case ALGORITHM_RIGHT_SYMMETRIC_6:
6506 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6507 break;
6508 case ALGORITHM_PARITY_0_6:
6509 new_layout = ALGORITHM_PARITY_0;
6510 break;
6511 case ALGORITHM_PARITY_N:
6512 new_layout = ALGORITHM_PARITY_N;
6513 break;
6514 default:
6515 return ERR_PTR(-EINVAL);
6516 }
6517 mddev->new_level = 5;
6518 mddev->new_layout = new_layout;
6519 mddev->delta_disks = -1;
6520 mddev->raid_disks -= 1;
6521 return setup_conf(mddev);
6522}
6523
d562b0c4 6524
fd01b88c 6525static int raid5_check_reshape(struct mddev *mddev)
b3546035 6526{
88ce4930
N		 6527 /* For a 2-drive array, the layout and chunk size can be changed
6528 * immediately as not restriping is needed.
6529 * For larger arrays we record the new value - after validation
6530 * to be used by a reshape pass.
b3546035 6531 */
d1688a6d 6532 struct r5conf *conf = mddev->private;
597a711b 6533 int new_chunk = mddev->new_chunk_sectors;
b3546035 6534
597a711b 6535 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
b3546035
N		 6536 return -EINVAL;
6537 if (new_chunk > 0) {
0ba459d2 6538 if (!is_power_of_2(new_chunk))
b3546035 6539 return -EINVAL;
597a711b 6540 if (new_chunk < (PAGE_SIZE>>9))
b3546035 6541 return -EINVAL;
597a711b 6542 if (mddev->array_sectors & (new_chunk-1))
b3546035
N		 6543 /* not factor of array size */
6544 return -EINVAL;
6545 }
6546
6547 /* They look valid */
6548
88ce4930 6549 if (mddev->raid_disks == 2) {
597a711b
N		 6550 /* can make the change immediately */
6551 if (mddev->new_layout >= 0) {
6552 conf->algorithm = mddev->new_layout;
6553 mddev->layout = mddev->new_layout;
88ce4930
N		 6554 }
6555 if (new_chunk > 0) {
597a711b
N		 6556 conf->chunk_sectors = new_chunk ;
6557 mddev->chunk_sectors = new_chunk;
88ce4930
N		 6558 }
6559 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6560 md_wakeup_thread(mddev->thread);
b3546035 6561 }
50ac168a 6562 return check_reshape(mddev);
88ce4930
N		 6563}
6564
fd01b88c 6565static int raid6_check_reshape(struct mddev *mddev)
88ce4930 6566{
597a711b 6567 int new_chunk = mddev->new_chunk_sectors;
50ac168a 6568
597a711b 6569 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
88ce4930 6570 return -EINVAL;
b3546035 6571 if (new_chunk > 0) {
0ba459d2 6572 if (!is_power_of_2(new_chunk))
88ce4930 6573 return -EINVAL;
597a711b 6574 if (new_chunk < (PAGE_SIZE >> 9))
88ce4930 6575 return -EINVAL;
597a711b 6576 if (mddev->array_sectors & (new_chunk-1))
88ce4930
N		 6577 /* not factor of array size */
6578 return -EINVAL;
b3546035 6579 }
88ce4930
N		 6580
6581 /* They look valid */
50ac168a 6582 return check_reshape(mddev);
b3546035
N		 6583}
6584
fd01b88c 6585static void *raid5_takeover(struct mddev *mddev)
d562b0c4
N		 6586{
6587 /* raid5 can take over:
f1b29bca 6588 * raid0 - if there is only one strip zone - make it a raid4 layout
d562b0c4
N		 6589 * raid1 - if there are two drives. We need to know the chunk size
6590 * raid4 - trivial - just use a raid4 layout.
6591 * raid6 - Providing it is a *_6 layout
d562b0c4 6592 */
f1b29bca
DW		 6593 if (mddev->level == 0)
6594 return raid45_takeover_raid0(mddev, 5);
d562b0c4
N		 6595 if (mddev->level == 1)
6596 return raid5_takeover_raid1(mddev);
e9d4758f
N		 6597 if (mddev->level == 4) {
6598 mddev->new_layout = ALGORITHM_PARITY_N;
6599 mddev->new_level = 5;
6600 return setup_conf(mddev);
6601 }
fc9739c6
N		 6602 if (mddev->level == 6)
6603 return raid5_takeover_raid6(mddev);
d562b0c4
N		 6604
6605 return ERR_PTR(-EINVAL);
6606}
6607
fd01b88c 6608static void *raid4_takeover(struct mddev *mddev)
a78d38a1 6609{
f1b29bca
DW		 6610 /* raid4 can take over:
6611 * raid0 - if there is only one strip zone
6612 * raid5 - if layout is right
a78d38a1 6613 */
f1b29bca
DW		 6614 if (mddev->level == 0)
6615 return raid45_takeover_raid0(mddev, 4);
a78d38a1
N		 6616 if (mddev->level == 5 &&
6617 mddev->layout == ALGORITHM_PARITY_N) {
6618 mddev->new_layout = 0;
6619 mddev->new_level = 4;
6620 return setup_conf(mddev);
6621 }
6622 return ERR_PTR(-EINVAL);
6623}
d562b0c4 6624
84fc4b56 6625static struct md_personality raid5_personality;
245f46c2 6626
fd01b88c 6627static void *raid6_takeover(struct mddev *mddev)
245f46c2
N		 6628{
6629 /* Currently can only take over a raid5. We map the
6630 * personality to an equivalent raid6 personality
6631 * with the Q block at the end.
6632 */
6633 int new_layout;
6634
6635 if (mddev->pers != &raid5_personality)
6636 return ERR_PTR(-EINVAL);
6637 if (mddev->degraded > 1)
6638 return ERR_PTR(-EINVAL);
6639 if (mddev->raid_disks > 253)
6640 return ERR_PTR(-EINVAL);
6641 if (mddev->raid_disks < 3)
6642 return ERR_PTR(-EINVAL);
6643
6644 switch (mddev->layout) {
6645 case ALGORITHM_LEFT_ASYMMETRIC:
6646 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6647 break;
6648 case ALGORITHM_RIGHT_ASYMMETRIC:
6649 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6650 break;
6651 case ALGORITHM_LEFT_SYMMETRIC:
6652 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6653 break;
6654 case ALGORITHM_RIGHT_SYMMETRIC:
6655 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6656 break;
6657 case ALGORITHM_PARITY_0:
6658 new_layout = ALGORITHM_PARITY_0_6;
6659 break;
6660 case ALGORITHM_PARITY_N:
6661 new_layout = ALGORITHM_PARITY_N;
6662 break;
6663 default:
6664 return ERR_PTR(-EINVAL);
6665 }
6666 mddev->new_level = 6;
6667 mddev->new_layout = new_layout;
6668 mddev->delta_disks = 1;
6669 mddev->raid_disks += 1;
6670 return setup_conf(mddev);
6671}
6672
6673
84fc4b56 6674static struct md_personality raid6_personality =
16a53ecc
N		 6675{
6676 .name = "raid6",
6677 .level = 6,
6678 .owner = THIS_MODULE,
6679 .make_request = make_request,
6680 .run = run,
6681 .stop = stop,
6682 .status = status,
6683 .error_handler = error,
6684 .hot_add_disk = raid5_add_disk,
6685 .hot_remove_disk= raid5_remove_disk,
6686 .spare_active = raid5_spare_active,
6687 .sync_request = sync_request,
6688 .resize = raid5_resize,
80c3a6ce 6689 .size = raid5_size,
50ac168a 6690 .check_reshape = raid6_check_reshape,
f416885e 6691 .start_reshape = raid5_start_reshape,
cea9c228 6692 .finish_reshape = raid5_finish_reshape,
16a53ecc 6693 .quiesce = raid5_quiesce,
245f46c2 6694 .takeover = raid6_takeover,
16a53ecc 6695};
84fc4b56 6696static struct md_personality raid5_personality =
1da177e4
LT		 6697{
6698 .name = "raid5",
2604b703 6699 .level = 5,
1da177e4
LT		 6700 .owner = THIS_MODULE,
6701 .make_request = make_request,
6702 .run = run,
6703 .stop = stop,
6704 .status = status,
6705 .error_handler = error,
6706 .hot_add_disk = raid5_add_disk,
6707 .hot_remove_disk= raid5_remove_disk,
6708 .spare_active = raid5_spare_active,
6709 .sync_request = sync_request,
6710 .resize = raid5_resize,
80c3a6ce 6711 .size = raid5_size,
63c70c4f
N		 6712 .check_reshape = raid5_check_reshape,
6713 .start_reshape = raid5_start_reshape,
cea9c228 6714 .finish_reshape = raid5_finish_reshape,
72626685 6715 .quiesce = raid5_quiesce,
d562b0c4 6716 .takeover = raid5_takeover,
1da177e4
LT		 6717};
6718
84fc4b56 6719static struct md_personality raid4_personality =
1da177e4 6720{
2604b703
N		 6721 .name = "raid4",
6722 .level = 4,
6723 .owner = THIS_MODULE,
6724 .make_request = make_request,
6725 .run = run,
6726 .stop = stop,
6727 .status = status,
6728 .error_handler = error,
6729 .hot_add_disk = raid5_add_disk,
6730 .hot_remove_disk= raid5_remove_disk,
6731 .spare_active = raid5_spare_active,
6732 .sync_request = sync_request,
6733 .resize = raid5_resize,
80c3a6ce 6734 .size = raid5_size,
3d37890b
N		 6735 .check_reshape = raid5_check_reshape,
6736 .start_reshape = raid5_start_reshape,
cea9c228 6737 .finish_reshape = raid5_finish_reshape,
2604b703 6738 .quiesce = raid5_quiesce,
a78d38a1 6739 .takeover = raid4_takeover,
2604b703
N		 6740};
6741
6742static int __init raid5_init(void)
6743{
851c30c9
SL		 6744 raid5_wq = alloc_workqueue("raid5wq",
6745 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
6746 if (!raid5_wq)
6747 return -ENOMEM;
16a53ecc 6748 register_md_personality(&raid6_personality);
2604b703
N		 6749 register_md_personality(&raid5_personality);
6750 register_md_personality(&raid4_personality);
6751 return 0;
1da177e4
LT		 6752}
6753
2604b703 6754static void raid5_exit(void)
1da177e4 6755{
16a53ecc 6756 unregister_md_personality(&raid6_personality);
2604b703
N		 6757 unregister_md_personality(&raid5_personality);
6758 unregister_md_personality(&raid4_personality);
851c30c9 6759 destroy_workqueue(raid5_wq);
1da177e4
LT		 6760}
6761
6762module_init(raid5_init);
6763module_exit(raid5_exit);
6764MODULE_LICENSE("GPL");
0efb9e61 6765MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
1da177e4 6766MODULE_ALIAS("md-personality-4"); /* RAID5 */
d9d166c2
N		 6767MODULE_ALIAS("md-raid5");
6768MODULE_ALIAS("md-raid4");
2604b703
N		 6769MODULE_ALIAS("md-level-5");
6770MODULE_ALIAS("md-level-4");
16a53ecc
N		 6771MODULE_ALIAS("md-personality-8"); /* RAID6 */
6772MODULE_ALIAS("md-raid6");
6773MODULE_ALIAS("md-level-6");
6774
6775/* This used to be two separate modules, they were: */
6776MODULE_ALIAS("raid5");
6777MODULE_ALIAS("raid6");
Linux 6.x block layer and io_uring tree(s)