Loading...
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
5 * Copyright (C) 2002, 2003 H. Peter Anvin
6 *
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!
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
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.
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
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
35 * the number of the batch it will be in. This is seq_flush+1.
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 */
45
46#include <linux/blkdev.h>
47#include <linux/kthread.h>
48#include <linux/raid/pq.h>
49#include <linux/async_tx.h>
50#include <linux/module.h>
51#include <linux/async.h>
52#include <linux/seq_file.h>
53#include <linux/cpu.h>
54#include <linux/slab.h>
55#include <linux/ratelimit.h>
56#include "md.h"
57#include "raid5.h"
58#include "raid0.h"
59#include "bitmap.h"
60
61/*
62 * Stripe cache
63 */
64
65#define NR_STRIPES 256
66#define STRIPE_SIZE PAGE_SIZE
67#define STRIPE_SHIFT (PAGE_SHIFT - 9)
68#define STRIPE_SECTORS (STRIPE_SIZE>>9)
69#define IO_THRESHOLD 1
70#define BYPASS_THRESHOLD 1
71#define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
72#define HASH_MASK (NR_HASH - 1)
73
74static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
75{
76 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
77 return &conf->stripe_hashtbl[hash];
78}
79
80/* bio's attached to a stripe+device for I/O are linked together in bi_sector
81 * order without overlap. There may be several bio's per stripe+device, and
82 * a bio could span several devices.
83 * When walking this list for a particular stripe+device, we must never proceed
84 * beyond a bio that extends past this device, as the next bio might no longer
85 * be valid.
86 * This function is used to determine the 'next' bio in the list, given the sector
87 * of the current stripe+device
88 */
89static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
90{
91 int sectors = bio->bi_size >> 9;
92 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
93 return bio->bi_next;
94 else
95 return NULL;
96}
97
98/*
99 * We maintain a biased count of active stripes in the bottom 16 bits of
100 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
101 */
102static inline int raid5_bi_phys_segments(struct bio *bio)
103{
104 return bio->bi_phys_segments & 0xffff;
105}
106
107static inline int raid5_bi_hw_segments(struct bio *bio)
108{
109 return (bio->bi_phys_segments >> 16) & 0xffff;
110}
111
112static inline int raid5_dec_bi_phys_segments(struct bio *bio)
113{
114 --bio->bi_phys_segments;
115 return raid5_bi_phys_segments(bio);
116}
117
118static inline int raid5_dec_bi_hw_segments(struct bio *bio)
119{
120 unsigned short val = raid5_bi_hw_segments(bio);
121
122 --val;
123 bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
124 return val;
125}
126
127static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
128{
129 bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
130}
131
132/* Find first data disk in a raid6 stripe */
133static inline int raid6_d0(struct stripe_head *sh)
134{
135 if (sh->ddf_layout)
136 /* ddf always start from first device */
137 return 0;
138 /* md starts just after Q block */
139 if (sh->qd_idx == sh->disks - 1)
140 return 0;
141 else
142 return sh->qd_idx + 1;
143}
144static inline int raid6_next_disk(int disk, int raid_disks)
145{
146 disk++;
147 return (disk < raid_disks) ? disk : 0;
148}
149
150/* When walking through the disks in a raid5, starting at raid6_d0,
151 * We need to map each disk to a 'slot', where the data disks are slot
152 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
153 * is raid_disks-1. This help does that mapping.
154 */
155static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
156 int *count, int syndrome_disks)
157{
158 int slot = *count;
159
160 if (sh->ddf_layout)
161 (*count)++;
162 if (idx == sh->pd_idx)
163 return syndrome_disks;
164 if (idx == sh->qd_idx)
165 return syndrome_disks + 1;
166 if (!sh->ddf_layout)
167 (*count)++;
168 return slot;
169}
170
171static void return_io(struct bio *return_bi)
172{
173 struct bio *bi = return_bi;
174 while (bi) {
175
176 return_bi = bi->bi_next;
177 bi->bi_next = NULL;
178 bi->bi_size = 0;
179 bio_endio(bi, 0);
180 bi = return_bi;
181 }
182}
183
184static void print_raid5_conf (struct r5conf *conf);
185
186static int stripe_operations_active(struct stripe_head *sh)
187{
188 return sh->check_state || sh->reconstruct_state ||
189 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
190 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
191}
192
193static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
194{
195 if (atomic_dec_and_test(&sh->count)) {
196 BUG_ON(!list_empty(&sh->lru));
197 BUG_ON(atomic_read(&conf->active_stripes)==0);
198 if (test_bit(STRIPE_HANDLE, &sh->state)) {
199 if (test_bit(STRIPE_DELAYED, &sh->state) &&
200 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
201 list_add_tail(&sh->lru, &conf->delayed_list);
202 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
203 sh->bm_seq - conf->seq_write > 0)
204 list_add_tail(&sh->lru, &conf->bitmap_list);
205 else {
206 clear_bit(STRIPE_DELAYED, &sh->state);
207 clear_bit(STRIPE_BIT_DELAY, &sh->state);
208 list_add_tail(&sh->lru, &conf->handle_list);
209 }
210 md_wakeup_thread(conf->mddev->thread);
211 } else {
212 BUG_ON(stripe_operations_active(sh));
213 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
214 if (atomic_dec_return(&conf->preread_active_stripes)
215 < IO_THRESHOLD)
216 md_wakeup_thread(conf->mddev->thread);
217 atomic_dec(&conf->active_stripes);
218 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
219 list_add_tail(&sh->lru, &conf->inactive_list);
220 wake_up(&conf->wait_for_stripe);
221 if (conf->retry_read_aligned)
222 md_wakeup_thread(conf->mddev->thread);
223 }
224 }
225 }
226}
227
228static void release_stripe(struct stripe_head *sh)
229{
230 struct r5conf *conf = sh->raid_conf;
231 unsigned long flags;
232
233 spin_lock_irqsave(&conf->device_lock, flags);
234 __release_stripe(conf, sh);
235 spin_unlock_irqrestore(&conf->device_lock, flags);
236}
237
238static inline void remove_hash(struct stripe_head *sh)
239{
240 pr_debug("remove_hash(), stripe %llu\n",
241 (unsigned long long)sh->sector);
242
243 hlist_del_init(&sh->hash);
244}
245
246static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
247{
248 struct hlist_head *hp = stripe_hash(conf, sh->sector);
249
250 pr_debug("insert_hash(), stripe %llu\n",
251 (unsigned long long)sh->sector);
252
253 hlist_add_head(&sh->hash, hp);
254}
255
256
257/* find an idle stripe, make sure it is unhashed, and return it. */
258static struct stripe_head *get_free_stripe(struct r5conf *conf)
259{
260 struct stripe_head *sh = NULL;
261 struct list_head *first;
262
263 if (list_empty(&conf->inactive_list))
264 goto out;
265 first = conf->inactive_list.next;
266 sh = list_entry(first, struct stripe_head, lru);
267 list_del_init(first);
268 remove_hash(sh);
269 atomic_inc(&conf->active_stripes);
270out:
271 return sh;
272}
273
274static void shrink_buffers(struct stripe_head *sh)
275{
276 struct page *p;
277 int i;
278 int num = sh->raid_conf->pool_size;
279
280 for (i = 0; i < num ; i++) {
281 p = sh->dev[i].page;
282 if (!p)
283 continue;
284 sh->dev[i].page = NULL;
285 put_page(p);
286 }
287}
288
289static int grow_buffers(struct stripe_head *sh)
290{
291 int i;
292 int num = sh->raid_conf->pool_size;
293
294 for (i = 0; i < num; i++) {
295 struct page *page;
296
297 if (!(page = alloc_page(GFP_KERNEL))) {
298 return 1;
299 }
300 sh->dev[i].page = page;
301 }
302 return 0;
303}
304
305static void raid5_build_block(struct stripe_head *sh, int i, int previous);
306static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
307 struct stripe_head *sh);
308
309static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
310{
311 struct r5conf *conf = sh->raid_conf;
312 int i;
313
314 BUG_ON(atomic_read(&sh->count) != 0);
315 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
316 BUG_ON(stripe_operations_active(sh));
317
318 pr_debug("init_stripe called, stripe %llu\n",
319 (unsigned long long)sh->sector);
320
321 remove_hash(sh);
322
323 sh->generation = conf->generation - previous;
324 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
325 sh->sector = sector;
326 stripe_set_idx(sector, conf, previous, sh);
327 sh->state = 0;
328
329
330 for (i = sh->disks; i--; ) {
331 struct r5dev *dev = &sh->dev[i];
332
333 if (dev->toread || dev->read || dev->towrite || dev->written ||
334 test_bit(R5_LOCKED, &dev->flags)) {
335 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
336 (unsigned long long)sh->sector, i, dev->toread,
337 dev->read, dev->towrite, dev->written,
338 test_bit(R5_LOCKED, &dev->flags));
339 WARN_ON(1);
340 }
341 dev->flags = 0;
342 raid5_build_block(sh, i, previous);
343 }
344 insert_hash(conf, sh);
345}
346
347static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
348 short generation)
349{
350 struct stripe_head *sh;
351 struct hlist_node *hn;
352
353 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
354 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
355 if (sh->sector == sector && sh->generation == generation)
356 return sh;
357 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
358 return NULL;
359}
360
361/*
362 * Need to check if array has failed when deciding whether to:
363 * - start an array
364 * - remove non-faulty devices
365 * - add a spare
366 * - allow a reshape
367 * This determination is simple when no reshape is happening.
368 * However if there is a reshape, we need to carefully check
369 * both the before and after sections.
370 * This is because some failed devices may only affect one
371 * of the two sections, and some non-in_sync devices may
372 * be insync in the section most affected by failed devices.
373 */
374static int calc_degraded(struct r5conf *conf)
375{
376 int degraded, degraded2;
377 int i;
378
379 rcu_read_lock();
380 degraded = 0;
381 for (i = 0; i < conf->previous_raid_disks; i++) {
382 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
383 if (rdev && test_bit(Faulty, &rdev->flags))
384 rdev = rcu_dereference(conf->disks[i].replacement);
385 if (!rdev || test_bit(Faulty, &rdev->flags))
386 degraded++;
387 else if (test_bit(In_sync, &rdev->flags))
388 ;
389 else
390 /* not in-sync or faulty.
391 * If the reshape increases the number of devices,
392 * this is being recovered by the reshape, so
393 * this 'previous' section is not in_sync.
394 * If the number of devices is being reduced however,
395 * the device can only be part of the array if
396 * we are reverting a reshape, so this section will
397 * be in-sync.
398 */
399 if (conf->raid_disks >= conf->previous_raid_disks)
400 degraded++;
401 }
402 rcu_read_unlock();
403 if (conf->raid_disks == conf->previous_raid_disks)
404 return degraded;
405 rcu_read_lock();
406 degraded2 = 0;
407 for (i = 0; i < conf->raid_disks; i++) {
408 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
409 if (rdev && test_bit(Faulty, &rdev->flags))
410 rdev = rcu_dereference(conf->disks[i].replacement);
411 if (!rdev || test_bit(Faulty, &rdev->flags))
412 degraded2++;
413 else if (test_bit(In_sync, &rdev->flags))
414 ;
415 else
416 /* not in-sync or faulty.
417 * If reshape increases the number of devices, this
418 * section has already been recovered, else it
419 * almost certainly hasn't.
420 */
421 if (conf->raid_disks <= conf->previous_raid_disks)
422 degraded2++;
423 }
424 rcu_read_unlock();
425 if (degraded2 > degraded)
426 return degraded2;
427 return degraded;
428}
429
430static int has_failed(struct r5conf *conf)
431{
432 int degraded;
433
434 if (conf->mddev->reshape_position == MaxSector)
435 return conf->mddev->degraded > conf->max_degraded;
436
437 degraded = calc_degraded(conf);
438 if (degraded > conf->max_degraded)
439 return 1;
440 return 0;
441}
442
443static struct stripe_head *
444get_active_stripe(struct r5conf *conf, sector_t sector,
445 int previous, int noblock, int noquiesce)
446{
447 struct stripe_head *sh;
448
449 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
450
451 spin_lock_irq(&conf->device_lock);
452
453 do {
454 wait_event_lock_irq(conf->wait_for_stripe,
455 conf->quiesce == 0 || noquiesce,
456 conf->device_lock, /* nothing */);
457 sh = __find_stripe(conf, sector, conf->generation - previous);
458 if (!sh) {
459 if (!conf->inactive_blocked)
460 sh = get_free_stripe(conf);
461 if (noblock && sh == NULL)
462 break;
463 if (!sh) {
464 conf->inactive_blocked = 1;
465 wait_event_lock_irq(conf->wait_for_stripe,
466 !list_empty(&conf->inactive_list) &&
467 (atomic_read(&conf->active_stripes)
468 < (conf->max_nr_stripes *3/4)
469 || !conf->inactive_blocked),
470 conf->device_lock,
471 );
472 conf->inactive_blocked = 0;
473 } else
474 init_stripe(sh, sector, previous);
475 } else {
476 if (atomic_read(&sh->count)) {
477 BUG_ON(!list_empty(&sh->lru)
478 && !test_bit(STRIPE_EXPANDING, &sh->state));
479 } else {
480 if (!test_bit(STRIPE_HANDLE, &sh->state))
481 atomic_inc(&conf->active_stripes);
482 if (list_empty(&sh->lru) &&
483 !test_bit(STRIPE_EXPANDING, &sh->state))
484 BUG();
485 list_del_init(&sh->lru);
486 }
487 }
488 } while (sh == NULL);
489
490 if (sh)
491 atomic_inc(&sh->count);
492
493 spin_unlock_irq(&conf->device_lock);
494 return sh;
495}
496
497/* Determine if 'data_offset' or 'new_data_offset' should be used
498 * in this stripe_head.
499 */
500static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
501{
502 sector_t progress = conf->reshape_progress;
503 /* Need a memory barrier to make sure we see the value
504 * of conf->generation, or ->data_offset that was set before
505 * reshape_progress was updated.
506 */
507 smp_rmb();
508 if (progress == MaxSector)
509 return 0;
510 if (sh->generation == conf->generation - 1)
511 return 0;
512 /* We are in a reshape, and this is a new-generation stripe,
513 * so use new_data_offset.
514 */
515 return 1;
516}
517
518static void
519raid5_end_read_request(struct bio *bi, int error);
520static void
521raid5_end_write_request(struct bio *bi, int error);
522
523static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
524{
525 struct r5conf *conf = sh->raid_conf;
526 int i, disks = sh->disks;
527
528 might_sleep();
529
530 for (i = disks; i--; ) {
531 int rw;
532 int replace_only = 0;
533 struct bio *bi, *rbi;
534 struct md_rdev *rdev, *rrdev = NULL;
535 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
536 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
537 rw = WRITE_FUA;
538 else
539 rw = WRITE;
540 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
541 rw = READ;
542 else if (test_and_clear_bit(R5_WantReplace,
543 &sh->dev[i].flags)) {
544 rw = WRITE;
545 replace_only = 1;
546 } else
547 continue;
548 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
549 rw |= REQ_SYNC;
550
551 bi = &sh->dev[i].req;
552 rbi = &sh->dev[i].rreq; /* For writing to replacement */
553
554 bi->bi_rw = rw;
555 rbi->bi_rw = rw;
556 if (rw & WRITE) {
557 bi->bi_end_io = raid5_end_write_request;
558 rbi->bi_end_io = raid5_end_write_request;
559 } else
560 bi->bi_end_io = raid5_end_read_request;
561
562 rcu_read_lock();
563 rrdev = rcu_dereference(conf->disks[i].replacement);
564 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
565 rdev = rcu_dereference(conf->disks[i].rdev);
566 if (!rdev) {
567 rdev = rrdev;
568 rrdev = NULL;
569 }
570 if (rw & WRITE) {
571 if (replace_only)
572 rdev = NULL;
573 if (rdev == rrdev)
574 /* We raced and saw duplicates */
575 rrdev = NULL;
576 } else {
577 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
578 rdev = rrdev;
579 rrdev = NULL;
580 }
581
582 if (rdev && test_bit(Faulty, &rdev->flags))
583 rdev = NULL;
584 if (rdev)
585 atomic_inc(&rdev->nr_pending);
586 if (rrdev && test_bit(Faulty, &rrdev->flags))
587 rrdev = NULL;
588 if (rrdev)
589 atomic_inc(&rrdev->nr_pending);
590 rcu_read_unlock();
591
592 /* We have already checked bad blocks for reads. Now
593 * need to check for writes. We never accept write errors
594 * on the replacement, so we don't to check rrdev.
595 */
596 while ((rw & WRITE) && rdev &&
597 test_bit(WriteErrorSeen, &rdev->flags)) {
598 sector_t first_bad;
599 int bad_sectors;
600 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
601 &first_bad, &bad_sectors);
602 if (!bad)
603 break;
604
605 if (bad < 0) {
606 set_bit(BlockedBadBlocks, &rdev->flags);
607 if (!conf->mddev->external &&
608 conf->mddev->flags) {
609 /* It is very unlikely, but we might
610 * still need to write out the
611 * bad block log - better give it
612 * a chance*/
613 md_check_recovery(conf->mddev);
614 }
615 /*
616 * Because md_wait_for_blocked_rdev
617 * will dec nr_pending, we must
618 * increment it first.
619 */
620 atomic_inc(&rdev->nr_pending);
621 md_wait_for_blocked_rdev(rdev, conf->mddev);
622 } else {
623 /* Acknowledged bad block - skip the write */
624 rdev_dec_pending(rdev, conf->mddev);
625 rdev = NULL;
626 }
627 }
628
629 if (rdev) {
630 if (s->syncing || s->expanding || s->expanded
631 || s->replacing)
632 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
633
634 set_bit(STRIPE_IO_STARTED, &sh->state);
635
636 bi->bi_bdev = rdev->bdev;
637 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
638 __func__, (unsigned long long)sh->sector,
639 bi->bi_rw, i);
640 atomic_inc(&sh->count);
641 if (use_new_offset(conf, sh))
642 bi->bi_sector = (sh->sector
643 + rdev->new_data_offset);
644 else
645 bi->bi_sector = (sh->sector
646 + rdev->data_offset);
647 bi->bi_flags = 1 << BIO_UPTODATE;
648 bi->bi_idx = 0;
649 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
650 bi->bi_io_vec[0].bv_offset = 0;
651 bi->bi_size = STRIPE_SIZE;
652 bi->bi_next = NULL;
653 if (rrdev)
654 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
655 generic_make_request(bi);
656 }
657 if (rrdev) {
658 if (s->syncing || s->expanding || s->expanded
659 || s->replacing)
660 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
661
662 set_bit(STRIPE_IO_STARTED, &sh->state);
663
664 rbi->bi_bdev = rrdev->bdev;
665 pr_debug("%s: for %llu schedule op %ld on "
666 "replacement disc %d\n",
667 __func__, (unsigned long long)sh->sector,
668 rbi->bi_rw, i);
669 atomic_inc(&sh->count);
670 if (use_new_offset(conf, sh))
671 rbi->bi_sector = (sh->sector
672 + rrdev->new_data_offset);
673 else
674 rbi->bi_sector = (sh->sector
675 + rrdev->data_offset);
676 rbi->bi_flags = 1 << BIO_UPTODATE;
677 rbi->bi_idx = 0;
678 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
679 rbi->bi_io_vec[0].bv_offset = 0;
680 rbi->bi_size = STRIPE_SIZE;
681 rbi->bi_next = NULL;
682 generic_make_request(rbi);
683 }
684 if (!rdev && !rrdev) {
685 if (rw & WRITE)
686 set_bit(STRIPE_DEGRADED, &sh->state);
687 pr_debug("skip op %ld on disc %d for sector %llu\n",
688 bi->bi_rw, i, (unsigned long long)sh->sector);
689 clear_bit(R5_LOCKED, &sh->dev[i].flags);
690 set_bit(STRIPE_HANDLE, &sh->state);
691 }
692 }
693}
694
695static struct dma_async_tx_descriptor *
696async_copy_data(int frombio, struct bio *bio, struct page *page,
697 sector_t sector, struct dma_async_tx_descriptor *tx)
698{
699 struct bio_vec *bvl;
700 struct page *bio_page;
701 int i;
702 int page_offset;
703 struct async_submit_ctl submit;
704 enum async_tx_flags flags = 0;
705
706 if (bio->bi_sector >= sector)
707 page_offset = (signed)(bio->bi_sector - sector) * 512;
708 else
709 page_offset = (signed)(sector - bio->bi_sector) * -512;
710
711 if (frombio)
712 flags |= ASYNC_TX_FENCE;
713 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
714
715 bio_for_each_segment(bvl, bio, i) {
716 int len = bvl->bv_len;
717 int clen;
718 int b_offset = 0;
719
720 if (page_offset < 0) {
721 b_offset = -page_offset;
722 page_offset += b_offset;
723 len -= b_offset;
724 }
725
726 if (len > 0 && page_offset + len > STRIPE_SIZE)
727 clen = STRIPE_SIZE - page_offset;
728 else
729 clen = len;
730
731 if (clen > 0) {
732 b_offset += bvl->bv_offset;
733 bio_page = bvl->bv_page;
734 if (frombio)
735 tx = async_memcpy(page, bio_page, page_offset,
736 b_offset, clen, &submit);
737 else
738 tx = async_memcpy(bio_page, page, b_offset,
739 page_offset, clen, &submit);
740 }
741 /* chain the operations */
742 submit.depend_tx = tx;
743
744 if (clen < len) /* hit end of page */
745 break;
746 page_offset += len;
747 }
748
749 return tx;
750}
751
752static void ops_complete_biofill(void *stripe_head_ref)
753{
754 struct stripe_head *sh = stripe_head_ref;
755 struct bio *return_bi = NULL;
756 struct r5conf *conf = sh->raid_conf;
757 int i;
758
759 pr_debug("%s: stripe %llu\n", __func__,
760 (unsigned long long)sh->sector);
761
762 /* clear completed biofills */
763 spin_lock_irq(&conf->device_lock);
764 for (i = sh->disks; i--; ) {
765 struct r5dev *dev = &sh->dev[i];
766
767 /* acknowledge completion of a biofill operation */
768 /* and check if we need to reply to a read request,
769 * new R5_Wantfill requests are held off until
770 * !STRIPE_BIOFILL_RUN
771 */
772 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
773 struct bio *rbi, *rbi2;
774
775 BUG_ON(!dev->read);
776 rbi = dev->read;
777 dev->read = NULL;
778 while (rbi && rbi->bi_sector <
779 dev->sector + STRIPE_SECTORS) {
780 rbi2 = r5_next_bio(rbi, dev->sector);
781 if (!raid5_dec_bi_phys_segments(rbi)) {
782 rbi->bi_next = return_bi;
783 return_bi = rbi;
784 }
785 rbi = rbi2;
786 }
787 }
788 }
789 spin_unlock_irq(&conf->device_lock);
790 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
791
792 return_io(return_bi);
793
794 set_bit(STRIPE_HANDLE, &sh->state);
795 release_stripe(sh);
796}
797
798static void ops_run_biofill(struct stripe_head *sh)
799{
800 struct dma_async_tx_descriptor *tx = NULL;
801 struct r5conf *conf = sh->raid_conf;
802 struct async_submit_ctl submit;
803 int i;
804
805 pr_debug("%s: stripe %llu\n", __func__,
806 (unsigned long long)sh->sector);
807
808 for (i = sh->disks; i--; ) {
809 struct r5dev *dev = &sh->dev[i];
810 if (test_bit(R5_Wantfill, &dev->flags)) {
811 struct bio *rbi;
812 spin_lock_irq(&conf->device_lock);
813 dev->read = rbi = dev->toread;
814 dev->toread = NULL;
815 spin_unlock_irq(&conf->device_lock);
816 while (rbi && rbi->bi_sector <
817 dev->sector + STRIPE_SECTORS) {
818 tx = async_copy_data(0, rbi, dev->page,
819 dev->sector, tx);
820 rbi = r5_next_bio(rbi, dev->sector);
821 }
822 }
823 }
824
825 atomic_inc(&sh->count);
826 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
827 async_trigger_callback(&submit);
828}
829
830static void mark_target_uptodate(struct stripe_head *sh, int target)
831{
832 struct r5dev *tgt;
833
834 if (target < 0)
835 return;
836
837 tgt = &sh->dev[target];
838 set_bit(R5_UPTODATE, &tgt->flags);
839 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
840 clear_bit(R5_Wantcompute, &tgt->flags);
841}
842
843static void ops_complete_compute(void *stripe_head_ref)
844{
845 struct stripe_head *sh = stripe_head_ref;
846
847 pr_debug("%s: stripe %llu\n", __func__,
848 (unsigned long long)sh->sector);
849
850 /* mark the computed target(s) as uptodate */
851 mark_target_uptodate(sh, sh->ops.target);
852 mark_target_uptodate(sh, sh->ops.target2);
853
854 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
855 if (sh->check_state == check_state_compute_run)
856 sh->check_state = check_state_compute_result;
857 set_bit(STRIPE_HANDLE, &sh->state);
858 release_stripe(sh);
859}
860
861/* return a pointer to the address conversion region of the scribble buffer */
862static addr_conv_t *to_addr_conv(struct stripe_head *sh,
863 struct raid5_percpu *percpu)
864{
865 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
866}
867
868static struct dma_async_tx_descriptor *
869ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
870{
871 int disks = sh->disks;
872 struct page **xor_srcs = percpu->scribble;
873 int target = sh->ops.target;
874 struct r5dev *tgt = &sh->dev[target];
875 struct page *xor_dest = tgt->page;
876 int count = 0;
877 struct dma_async_tx_descriptor *tx;
878 struct async_submit_ctl submit;
879 int i;
880
881 pr_debug("%s: stripe %llu block: %d\n",
882 __func__, (unsigned long long)sh->sector, target);
883 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
884
885 for (i = disks; i--; )
886 if (i != target)
887 xor_srcs[count++] = sh->dev[i].page;
888
889 atomic_inc(&sh->count);
890
891 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
892 ops_complete_compute, sh, to_addr_conv(sh, percpu));
893 if (unlikely(count == 1))
894 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
895 else
896 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
897
898 return tx;
899}
900
901/* set_syndrome_sources - populate source buffers for gen_syndrome
902 * @srcs - (struct page *) array of size sh->disks
903 * @sh - stripe_head to parse
904 *
905 * Populates srcs in proper layout order for the stripe and returns the
906 * 'count' of sources to be used in a call to async_gen_syndrome. The P
907 * destination buffer is recorded in srcs[count] and the Q destination
908 * is recorded in srcs[count+1]].
909 */
910static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
911{
912 int disks = sh->disks;
913 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
914 int d0_idx = raid6_d0(sh);
915 int count;
916 int i;
917
918 for (i = 0; i < disks; i++)
919 srcs[i] = NULL;
920
921 count = 0;
922 i = d0_idx;
923 do {
924 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
925
926 srcs[slot] = sh->dev[i].page;
927 i = raid6_next_disk(i, disks);
928 } while (i != d0_idx);
929
930 return syndrome_disks;
931}
932
933static struct dma_async_tx_descriptor *
934ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
935{
936 int disks = sh->disks;
937 struct page **blocks = percpu->scribble;
938 int target;
939 int qd_idx = sh->qd_idx;
940 struct dma_async_tx_descriptor *tx;
941 struct async_submit_ctl submit;
942 struct r5dev *tgt;
943 struct page *dest;
944 int i;
945 int count;
946
947 if (sh->ops.target < 0)
948 target = sh->ops.target2;
949 else if (sh->ops.target2 < 0)
950 target = sh->ops.target;
951 else
952 /* we should only have one valid target */
953 BUG();
954 BUG_ON(target < 0);
955 pr_debug("%s: stripe %llu block: %d\n",
956 __func__, (unsigned long long)sh->sector, target);
957
958 tgt = &sh->dev[target];
959 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
960 dest = tgt->page;
961
962 atomic_inc(&sh->count);
963
964 if (target == qd_idx) {
965 count = set_syndrome_sources(blocks, sh);
966 blocks[count] = NULL; /* regenerating p is not necessary */
967 BUG_ON(blocks[count+1] != dest); /* q should already be set */
968 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
969 ops_complete_compute, sh,
970 to_addr_conv(sh, percpu));
971 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
972 } else {
973 /* Compute any data- or p-drive using XOR */
974 count = 0;
975 for (i = disks; i-- ; ) {
976 if (i == target || i == qd_idx)
977 continue;
978 blocks[count++] = sh->dev[i].page;
979 }
980
981 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
982 NULL, ops_complete_compute, sh,
983 to_addr_conv(sh, percpu));
984 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
985 }
986
987 return tx;
988}
989
990static struct dma_async_tx_descriptor *
991ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
992{
993 int i, count, disks = sh->disks;
994 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
995 int d0_idx = raid6_d0(sh);
996 int faila = -1, failb = -1;
997 int target = sh->ops.target;
998 int target2 = sh->ops.target2;
999 struct r5dev *tgt = &sh->dev[target];
1000 struct r5dev *tgt2 = &sh->dev[target2];
1001 struct dma_async_tx_descriptor *tx;
1002 struct page **blocks = percpu->scribble;
1003 struct async_submit_ctl submit;
1004
1005 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1006 __func__, (unsigned long long)sh->sector, target, target2);
1007 BUG_ON(target < 0 || target2 < 0);
1008 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1009 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1010
1011 /* we need to open-code set_syndrome_sources to handle the
1012 * slot number conversion for 'faila' and 'failb'
1013 */
1014 for (i = 0; i < disks ; i++)
1015 blocks[i] = NULL;
1016 count = 0;
1017 i = d0_idx;
1018 do {
1019 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1020
1021 blocks[slot] = sh->dev[i].page;
1022
1023 if (i == target)
1024 faila = slot;
1025 if (i == target2)
1026 failb = slot;
1027 i = raid6_next_disk(i, disks);
1028 } while (i != d0_idx);
1029
1030 BUG_ON(faila == failb);
1031 if (failb < faila)
1032 swap(faila, failb);
1033 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1034 __func__, (unsigned long long)sh->sector, faila, failb);
1035
1036 atomic_inc(&sh->count);
1037
1038 if (failb == syndrome_disks+1) {
1039 /* Q disk is one of the missing disks */
1040 if (faila == syndrome_disks) {
1041 /* Missing P+Q, just recompute */
1042 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1043 ops_complete_compute, sh,
1044 to_addr_conv(sh, percpu));
1045 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1046 STRIPE_SIZE, &submit);
1047 } else {
1048 struct page *dest;
1049 int data_target;
1050 int qd_idx = sh->qd_idx;
1051
1052 /* Missing D+Q: recompute D from P, then recompute Q */
1053 if (target == qd_idx)
1054 data_target = target2;
1055 else
1056 data_target = target;
1057
1058 count = 0;
1059 for (i = disks; i-- ; ) {
1060 if (i == data_target || i == qd_idx)
1061 continue;
1062 blocks[count++] = sh->dev[i].page;
1063 }
1064 dest = sh->dev[data_target].page;
1065 init_async_submit(&submit,
1066 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1067 NULL, NULL, NULL,
1068 to_addr_conv(sh, percpu));
1069 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1070 &submit);
1071
1072 count = set_syndrome_sources(blocks, sh);
1073 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1074 ops_complete_compute, sh,
1075 to_addr_conv(sh, percpu));
1076 return async_gen_syndrome(blocks, 0, count+2,
1077 STRIPE_SIZE, &submit);
1078 }
1079 } else {
1080 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1081 ops_complete_compute, sh,
1082 to_addr_conv(sh, percpu));
1083 if (failb == syndrome_disks) {
1084 /* We're missing D+P. */
1085 return async_raid6_datap_recov(syndrome_disks+2,
1086 STRIPE_SIZE, faila,
1087 blocks, &submit);
1088 } else {
1089 /* We're missing D+D. */
1090 return async_raid6_2data_recov(syndrome_disks+2,
1091 STRIPE_SIZE, faila, failb,
1092 blocks, &submit);
1093 }
1094 }
1095}
1096
1097
1098static void ops_complete_prexor(void *stripe_head_ref)
1099{
1100 struct stripe_head *sh = stripe_head_ref;
1101
1102 pr_debug("%s: stripe %llu\n", __func__,
1103 (unsigned long long)sh->sector);
1104}
1105
1106static struct dma_async_tx_descriptor *
1107ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1108 struct dma_async_tx_descriptor *tx)
1109{
1110 int disks = sh->disks;
1111 struct page **xor_srcs = percpu->scribble;
1112 int count = 0, pd_idx = sh->pd_idx, i;
1113 struct async_submit_ctl submit;
1114
1115 /* existing parity data subtracted */
1116 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1117
1118 pr_debug("%s: stripe %llu\n", __func__,
1119 (unsigned long long)sh->sector);
1120
1121 for (i = disks; i--; ) {
1122 struct r5dev *dev = &sh->dev[i];
1123 /* Only process blocks that are known to be uptodate */
1124 if (test_bit(R5_Wantdrain, &dev->flags))
1125 xor_srcs[count++] = dev->page;
1126 }
1127
1128 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1129 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1130 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1131
1132 return tx;
1133}
1134
1135static struct dma_async_tx_descriptor *
1136ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1137{
1138 int disks = sh->disks;
1139 int i;
1140
1141 pr_debug("%s: stripe %llu\n", __func__,
1142 (unsigned long long)sh->sector);
1143
1144 for (i = disks; i--; ) {
1145 struct r5dev *dev = &sh->dev[i];
1146 struct bio *chosen;
1147
1148 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1149 struct bio *wbi;
1150
1151 spin_lock_irq(&sh->raid_conf->device_lock);
1152 chosen = dev->towrite;
1153 dev->towrite = NULL;
1154 BUG_ON(dev->written);
1155 wbi = dev->written = chosen;
1156 spin_unlock_irq(&sh->raid_conf->device_lock);
1157
1158 while (wbi && wbi->bi_sector <
1159 dev->sector + STRIPE_SECTORS) {
1160 if (wbi->bi_rw & REQ_FUA)
1161 set_bit(R5_WantFUA, &dev->flags);
1162 if (wbi->bi_rw & REQ_SYNC)
1163 set_bit(R5_SyncIO, &dev->flags);
1164 tx = async_copy_data(1, wbi, dev->page,
1165 dev->sector, tx);
1166 wbi = r5_next_bio(wbi, dev->sector);
1167 }
1168 }
1169 }
1170
1171 return tx;
1172}
1173
1174static void ops_complete_reconstruct(void *stripe_head_ref)
1175{
1176 struct stripe_head *sh = stripe_head_ref;
1177 int disks = sh->disks;
1178 int pd_idx = sh->pd_idx;
1179 int qd_idx = sh->qd_idx;
1180 int i;
1181 bool fua = false, sync = false;
1182
1183 pr_debug("%s: stripe %llu\n", __func__,
1184 (unsigned long long)sh->sector);
1185
1186 for (i = disks; i--; ) {
1187 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1188 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1189 }
1190
1191 for (i = disks; i--; ) {
1192 struct r5dev *dev = &sh->dev[i];
1193
1194 if (dev->written || i == pd_idx || i == qd_idx) {
1195 set_bit(R5_UPTODATE, &dev->flags);
1196 if (fua)
1197 set_bit(R5_WantFUA, &dev->flags);
1198 if (sync)
1199 set_bit(R5_SyncIO, &dev->flags);
1200 }
1201 }
1202
1203 if (sh->reconstruct_state == reconstruct_state_drain_run)
1204 sh->reconstruct_state = reconstruct_state_drain_result;
1205 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1206 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1207 else {
1208 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1209 sh->reconstruct_state = reconstruct_state_result;
1210 }
1211
1212 set_bit(STRIPE_HANDLE, &sh->state);
1213 release_stripe(sh);
1214}
1215
1216static void
1217ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1218 struct dma_async_tx_descriptor *tx)
1219{
1220 int disks = sh->disks;
1221 struct page **xor_srcs = percpu->scribble;
1222 struct async_submit_ctl submit;
1223 int count = 0, pd_idx = sh->pd_idx, i;
1224 struct page *xor_dest;
1225 int prexor = 0;
1226 unsigned long flags;
1227
1228 pr_debug("%s: stripe %llu\n", __func__,
1229 (unsigned long long)sh->sector);
1230
1231 /* check if prexor is active which means only process blocks
1232 * that are part of a read-modify-write (written)
1233 */
1234 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1235 prexor = 1;
1236 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1237 for (i = disks; i--; ) {
1238 struct r5dev *dev = &sh->dev[i];
1239 if (dev->written)
1240 xor_srcs[count++] = dev->page;
1241 }
1242 } else {
1243 xor_dest = sh->dev[pd_idx].page;
1244 for (i = disks; i--; ) {
1245 struct r5dev *dev = &sh->dev[i];
1246 if (i != pd_idx)
1247 xor_srcs[count++] = dev->page;
1248 }
1249 }
1250
1251 /* 1/ if we prexor'd then the dest is reused as a source
1252 * 2/ if we did not prexor then we are redoing the parity
1253 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1254 * for the synchronous xor case
1255 */
1256 flags = ASYNC_TX_ACK |
1257 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1258
1259 atomic_inc(&sh->count);
1260
1261 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1262 to_addr_conv(sh, percpu));
1263 if (unlikely(count == 1))
1264 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1265 else
1266 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1267}
1268
1269static void
1270ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1271 struct dma_async_tx_descriptor *tx)
1272{
1273 struct async_submit_ctl submit;
1274 struct page **blocks = percpu->scribble;
1275 int count;
1276
1277 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1278
1279 count = set_syndrome_sources(blocks, sh);
1280
1281 atomic_inc(&sh->count);
1282
1283 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1284 sh, to_addr_conv(sh, percpu));
1285 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1286}
1287
1288static void ops_complete_check(void *stripe_head_ref)
1289{
1290 struct stripe_head *sh = stripe_head_ref;
1291
1292 pr_debug("%s: stripe %llu\n", __func__,
1293 (unsigned long long)sh->sector);
1294
1295 sh->check_state = check_state_check_result;
1296 set_bit(STRIPE_HANDLE, &sh->state);
1297 release_stripe(sh);
1298}
1299
1300static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1301{
1302 int disks = sh->disks;
1303 int pd_idx = sh->pd_idx;
1304 int qd_idx = sh->qd_idx;
1305 struct page *xor_dest;
1306 struct page **xor_srcs = percpu->scribble;
1307 struct dma_async_tx_descriptor *tx;
1308 struct async_submit_ctl submit;
1309 int count;
1310 int i;
1311
1312 pr_debug("%s: stripe %llu\n", __func__,
1313 (unsigned long long)sh->sector);
1314
1315 count = 0;
1316 xor_dest = sh->dev[pd_idx].page;
1317 xor_srcs[count++] = xor_dest;
1318 for (i = disks; i--; ) {
1319 if (i == pd_idx || i == qd_idx)
1320 continue;
1321 xor_srcs[count++] = sh->dev[i].page;
1322 }
1323
1324 init_async_submit(&submit, 0, NULL, NULL, NULL,
1325 to_addr_conv(sh, percpu));
1326 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1327 &sh->ops.zero_sum_result, &submit);
1328
1329 atomic_inc(&sh->count);
1330 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1331 tx = async_trigger_callback(&submit);
1332}
1333
1334static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1335{
1336 struct page **srcs = percpu->scribble;
1337 struct async_submit_ctl submit;
1338 int count;
1339
1340 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1341 (unsigned long long)sh->sector, checkp);
1342
1343 count = set_syndrome_sources(srcs, sh);
1344 if (!checkp)
1345 srcs[count] = NULL;
1346
1347 atomic_inc(&sh->count);
1348 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1349 sh, to_addr_conv(sh, percpu));
1350 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1351 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1352}
1353
1354static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1355{
1356 int overlap_clear = 0, i, disks = sh->disks;
1357 struct dma_async_tx_descriptor *tx = NULL;
1358 struct r5conf *conf = sh->raid_conf;
1359 int level = conf->level;
1360 struct raid5_percpu *percpu;
1361 unsigned long cpu;
1362
1363 cpu = get_cpu();
1364 percpu = per_cpu_ptr(conf->percpu, cpu);
1365 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1366 ops_run_biofill(sh);
1367 overlap_clear++;
1368 }
1369
1370 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1371 if (level < 6)
1372 tx = ops_run_compute5(sh, percpu);
1373 else {
1374 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1375 tx = ops_run_compute6_1(sh, percpu);
1376 else
1377 tx = ops_run_compute6_2(sh, percpu);
1378 }
1379 /* terminate the chain if reconstruct is not set to be run */
1380 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1381 async_tx_ack(tx);
1382 }
1383
1384 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1385 tx = ops_run_prexor(sh, percpu, tx);
1386
1387 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1388 tx = ops_run_biodrain(sh, tx);
1389 overlap_clear++;
1390 }
1391
1392 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1393 if (level < 6)
1394 ops_run_reconstruct5(sh, percpu, tx);
1395 else
1396 ops_run_reconstruct6(sh, percpu, tx);
1397 }
1398
1399 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1400 if (sh->check_state == check_state_run)
1401 ops_run_check_p(sh, percpu);
1402 else if (sh->check_state == check_state_run_q)
1403 ops_run_check_pq(sh, percpu, 0);
1404 else if (sh->check_state == check_state_run_pq)
1405 ops_run_check_pq(sh, percpu, 1);
1406 else
1407 BUG();
1408 }
1409
1410 if (overlap_clear)
1411 for (i = disks; i--; ) {
1412 struct r5dev *dev = &sh->dev[i];
1413 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1414 wake_up(&sh->raid_conf->wait_for_overlap);
1415 }
1416 put_cpu();
1417}
1418
1419#ifdef CONFIG_MULTICORE_RAID456
1420static void async_run_ops(void *param, async_cookie_t cookie)
1421{
1422 struct stripe_head *sh = param;
1423 unsigned long ops_request = sh->ops.request;
1424
1425 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1426 wake_up(&sh->ops.wait_for_ops);
1427
1428 __raid_run_ops(sh, ops_request);
1429 release_stripe(sh);
1430}
1431
1432static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1433{
1434 /* since handle_stripe can be called outside of raid5d context
1435 * we need to ensure sh->ops.request is de-staged before another
1436 * request arrives
1437 */
1438 wait_event(sh->ops.wait_for_ops,
1439 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1440 sh->ops.request = ops_request;
1441
1442 atomic_inc(&sh->count);
1443 async_schedule(async_run_ops, sh);
1444}
1445#else
1446#define raid_run_ops __raid_run_ops
1447#endif
1448
1449static int grow_one_stripe(struct r5conf *conf)
1450{
1451 struct stripe_head *sh;
1452 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1453 if (!sh)
1454 return 0;
1455
1456 sh->raid_conf = conf;
1457 #ifdef CONFIG_MULTICORE_RAID456
1458 init_waitqueue_head(&sh->ops.wait_for_ops);
1459 #endif
1460
1461 if (grow_buffers(sh)) {
1462 shrink_buffers(sh);
1463 kmem_cache_free(conf->slab_cache, sh);
1464 return 0;
1465 }
1466 /* we just created an active stripe so... */
1467 atomic_set(&sh->count, 1);
1468 atomic_inc(&conf->active_stripes);
1469 INIT_LIST_HEAD(&sh->lru);
1470 release_stripe(sh);
1471 return 1;
1472}
1473
1474static int grow_stripes(struct r5conf *conf, int num)
1475{
1476 struct kmem_cache *sc;
1477 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1478
1479 if (conf->mddev->gendisk)
1480 sprintf(conf->cache_name[0],
1481 "raid%d-%s", conf->level, mdname(conf->mddev));
1482 else
1483 sprintf(conf->cache_name[0],
1484 "raid%d-%p", conf->level, conf->mddev);
1485 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1486
1487 conf->active_name = 0;
1488 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1489 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1490 0, 0, NULL);
1491 if (!sc)
1492 return 1;
1493 conf->slab_cache = sc;
1494 conf->pool_size = devs;
1495 while (num--)
1496 if (!grow_one_stripe(conf))
1497 return 1;
1498 return 0;
1499}
1500
1501/**
1502 * scribble_len - return the required size of the scribble region
1503 * @num - total number of disks in the array
1504 *
1505 * The size must be enough to contain:
1506 * 1/ a struct page pointer for each device in the array +2
1507 * 2/ room to convert each entry in (1) to its corresponding dma
1508 * (dma_map_page()) or page (page_address()) address.
1509 *
1510 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1511 * calculate over all devices (not just the data blocks), using zeros in place
1512 * of the P and Q blocks.
1513 */
1514static size_t scribble_len(int num)
1515{
1516 size_t len;
1517
1518 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1519
1520 return len;
1521}
1522
1523static int resize_stripes(struct r5conf *conf, int newsize)
1524{
1525 /* Make all the stripes able to hold 'newsize' devices.
1526 * New slots in each stripe get 'page' set to a new page.
1527 *
1528 * This happens in stages:
1529 * 1/ create a new kmem_cache and allocate the required number of
1530 * stripe_heads.
1531 * 2/ gather all the old stripe_heads and tranfer the pages across
1532 * to the new stripe_heads. This will have the side effect of
1533 * freezing the array as once all stripe_heads have been collected,
1534 * no IO will be possible. Old stripe heads are freed once their
1535 * pages have been transferred over, and the old kmem_cache is
1536 * freed when all stripes are done.
1537 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1538 * we simple return a failre status - no need to clean anything up.
1539 * 4/ allocate new pages for the new slots in the new stripe_heads.
1540 * If this fails, we don't bother trying the shrink the
1541 * stripe_heads down again, we just leave them as they are.
1542 * As each stripe_head is processed the new one is released into
1543 * active service.
1544 *
1545 * Once step2 is started, we cannot afford to wait for a write,
1546 * so we use GFP_NOIO allocations.
1547 */
1548 struct stripe_head *osh, *nsh;
1549 LIST_HEAD(newstripes);
1550 struct disk_info *ndisks;
1551 unsigned long cpu;
1552 int err;
1553 struct kmem_cache *sc;
1554 int i;
1555
1556 if (newsize <= conf->pool_size)
1557 return 0; /* never bother to shrink */
1558
1559 err = md_allow_write(conf->mddev);
1560 if (err)
1561 return err;
1562
1563 /* Step 1 */
1564 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1565 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1566 0, 0, NULL);
1567 if (!sc)
1568 return -ENOMEM;
1569
1570 for (i = conf->max_nr_stripes; i; i--) {
1571 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1572 if (!nsh)
1573 break;
1574
1575 nsh->raid_conf = conf;
1576 #ifdef CONFIG_MULTICORE_RAID456
1577 init_waitqueue_head(&nsh->ops.wait_for_ops);
1578 #endif
1579
1580 list_add(&nsh->lru, &newstripes);
1581 }
1582 if (i) {
1583 /* didn't get enough, give up */
1584 while (!list_empty(&newstripes)) {
1585 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1586 list_del(&nsh->lru);
1587 kmem_cache_free(sc, nsh);
1588 }
1589 kmem_cache_destroy(sc);
1590 return -ENOMEM;
1591 }
1592 /* Step 2 - Must use GFP_NOIO now.
1593 * OK, we have enough stripes, start collecting inactive
1594 * stripes and copying them over
1595 */
1596 list_for_each_entry(nsh, &newstripes, lru) {
1597 spin_lock_irq(&conf->device_lock);
1598 wait_event_lock_irq(conf->wait_for_stripe,
1599 !list_empty(&conf->inactive_list),
1600 conf->device_lock,
1601 );
1602 osh = get_free_stripe(conf);
1603 spin_unlock_irq(&conf->device_lock);
1604 atomic_set(&nsh->count, 1);
1605 for(i=0; i<conf->pool_size; i++)
1606 nsh->dev[i].page = osh->dev[i].page;
1607 for( ; i<newsize; i++)
1608 nsh->dev[i].page = NULL;
1609 kmem_cache_free(conf->slab_cache, osh);
1610 }
1611 kmem_cache_destroy(conf->slab_cache);
1612
1613 /* Step 3.
1614 * At this point, we are holding all the stripes so the array
1615 * is completely stalled, so now is a good time to resize
1616 * conf->disks and the scribble region
1617 */
1618 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1619 if (ndisks) {
1620 for (i=0; i<conf->raid_disks; i++)
1621 ndisks[i] = conf->disks[i];
1622 kfree(conf->disks);
1623 conf->disks = ndisks;
1624 } else
1625 err = -ENOMEM;
1626
1627 get_online_cpus();
1628 conf->scribble_len = scribble_len(newsize);
1629 for_each_present_cpu(cpu) {
1630 struct raid5_percpu *percpu;
1631 void *scribble;
1632
1633 percpu = per_cpu_ptr(conf->percpu, cpu);
1634 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1635
1636 if (scribble) {
1637 kfree(percpu->scribble);
1638 percpu->scribble = scribble;
1639 } else {
1640 err = -ENOMEM;
1641 break;
1642 }
1643 }
1644 put_online_cpus();
1645
1646 /* Step 4, return new stripes to service */
1647 while(!list_empty(&newstripes)) {
1648 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1649 list_del_init(&nsh->lru);
1650
1651 for (i=conf->raid_disks; i < newsize; i++)
1652 if (nsh->dev[i].page == NULL) {
1653 struct page *p = alloc_page(GFP_NOIO);
1654 nsh->dev[i].page = p;
1655 if (!p)
1656 err = -ENOMEM;
1657 }
1658 release_stripe(nsh);
1659 }
1660 /* critical section pass, GFP_NOIO no longer needed */
1661
1662 conf->slab_cache = sc;
1663 conf->active_name = 1-conf->active_name;
1664 conf->pool_size = newsize;
1665 return err;
1666}
1667
1668static int drop_one_stripe(struct r5conf *conf)
1669{
1670 struct stripe_head *sh;
1671
1672 spin_lock_irq(&conf->device_lock);
1673 sh = get_free_stripe(conf);
1674 spin_unlock_irq(&conf->device_lock);
1675 if (!sh)
1676 return 0;
1677 BUG_ON(atomic_read(&sh->count));
1678 shrink_buffers(sh);
1679 kmem_cache_free(conf->slab_cache, sh);
1680 atomic_dec(&conf->active_stripes);
1681 return 1;
1682}
1683
1684static void shrink_stripes(struct r5conf *conf)
1685{
1686 while (drop_one_stripe(conf))
1687 ;
1688
1689 if (conf->slab_cache)
1690 kmem_cache_destroy(conf->slab_cache);
1691 conf->slab_cache = NULL;
1692}
1693
1694static void raid5_end_read_request(struct bio * bi, int error)
1695{
1696 struct stripe_head *sh = bi->bi_private;
1697 struct r5conf *conf = sh->raid_conf;
1698 int disks = sh->disks, i;
1699 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1700 char b[BDEVNAME_SIZE];
1701 struct md_rdev *rdev = NULL;
1702 sector_t s;
1703
1704 for (i=0 ; i<disks; i++)
1705 if (bi == &sh->dev[i].req)
1706 break;
1707
1708 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1709 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1710 uptodate);
1711 if (i == disks) {
1712 BUG();
1713 return;
1714 }
1715 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1716 /* If replacement finished while this request was outstanding,
1717 * 'replacement' might be NULL already.
1718 * In that case it moved down to 'rdev'.
1719 * rdev is not removed until all requests are finished.
1720 */
1721 rdev = conf->disks[i].replacement;
1722 if (!rdev)
1723 rdev = conf->disks[i].rdev;
1724
1725 if (use_new_offset(conf, sh))
1726 s = sh->sector + rdev->new_data_offset;
1727 else
1728 s = sh->sector + rdev->data_offset;
1729 if (uptodate) {
1730 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1731 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1732 /* Note that this cannot happen on a
1733 * replacement device. We just fail those on
1734 * any error
1735 */
1736 printk_ratelimited(
1737 KERN_INFO
1738 "md/raid:%s: read error corrected"
1739 " (%lu sectors at %llu on %s)\n",
1740 mdname(conf->mddev), STRIPE_SECTORS,
1741 (unsigned long long)s,
1742 bdevname(rdev->bdev, b));
1743 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1744 clear_bit(R5_ReadError, &sh->dev[i].flags);
1745 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1746 }
1747 if (atomic_read(&rdev->read_errors))
1748 atomic_set(&rdev->read_errors, 0);
1749 } else {
1750 const char *bdn = bdevname(rdev->bdev, b);
1751 int retry = 0;
1752 int set_bad = 0;
1753
1754 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1755 atomic_inc(&rdev->read_errors);
1756 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1757 printk_ratelimited(
1758 KERN_WARNING
1759 "md/raid:%s: read error on replacement device "
1760 "(sector %llu on %s).\n",
1761 mdname(conf->mddev),
1762 (unsigned long long)s,
1763 bdn);
1764 else if (conf->mddev->degraded >= conf->max_degraded) {
1765 set_bad = 1;
1766 printk_ratelimited(
1767 KERN_WARNING
1768 "md/raid:%s: read error not correctable "
1769 "(sector %llu on %s).\n",
1770 mdname(conf->mddev),
1771 (unsigned long long)s,
1772 bdn);
1773 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1774 /* Oh, no!!! */
1775 set_bad = 1;
1776 printk_ratelimited(
1777 KERN_WARNING
1778 "md/raid:%s: read error NOT corrected!! "
1779 "(sector %llu on %s).\n",
1780 mdname(conf->mddev),
1781 (unsigned long long)s,
1782 bdn);
1783 } else if (atomic_read(&rdev->read_errors)
1784 > conf->max_nr_stripes)
1785 printk(KERN_WARNING
1786 "md/raid:%s: Too many read errors, failing device %s.\n",
1787 mdname(conf->mddev), bdn);
1788 else
1789 retry = 1;
1790 if (retry)
1791 set_bit(R5_ReadError, &sh->dev[i].flags);
1792 else {
1793 clear_bit(R5_ReadError, &sh->dev[i].flags);
1794 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1795 if (!(set_bad
1796 && test_bit(In_sync, &rdev->flags)
1797 && rdev_set_badblocks(
1798 rdev, sh->sector, STRIPE_SECTORS, 0)))
1799 md_error(conf->mddev, rdev);
1800 }
1801 }
1802 rdev_dec_pending(rdev, conf->mddev);
1803 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1804 set_bit(STRIPE_HANDLE, &sh->state);
1805 release_stripe(sh);
1806}
1807
1808static void raid5_end_write_request(struct bio *bi, int error)
1809{
1810 struct stripe_head *sh = bi->bi_private;
1811 struct r5conf *conf = sh->raid_conf;
1812 int disks = sh->disks, i;
1813 struct md_rdev *uninitialized_var(rdev);
1814 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1815 sector_t first_bad;
1816 int bad_sectors;
1817 int replacement = 0;
1818
1819 for (i = 0 ; i < disks; i++) {
1820 if (bi == &sh->dev[i].req) {
1821 rdev = conf->disks[i].rdev;
1822 break;
1823 }
1824 if (bi == &sh->dev[i].rreq) {
1825 rdev = conf->disks[i].replacement;
1826 if (rdev)
1827 replacement = 1;
1828 else
1829 /* rdev was removed and 'replacement'
1830 * replaced it. rdev is not removed
1831 * until all requests are finished.
1832 */
1833 rdev = conf->disks[i].rdev;
1834 break;
1835 }
1836 }
1837 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1838 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1839 uptodate);
1840 if (i == disks) {
1841 BUG();
1842 return;
1843 }
1844
1845 if (replacement) {
1846 if (!uptodate)
1847 md_error(conf->mddev, rdev);
1848 else if (is_badblock(rdev, sh->sector,
1849 STRIPE_SECTORS,
1850 &first_bad, &bad_sectors))
1851 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1852 } else {
1853 if (!uptodate) {
1854 set_bit(WriteErrorSeen, &rdev->flags);
1855 set_bit(R5_WriteError, &sh->dev[i].flags);
1856 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1857 set_bit(MD_RECOVERY_NEEDED,
1858 &rdev->mddev->recovery);
1859 } else if (is_badblock(rdev, sh->sector,
1860 STRIPE_SECTORS,
1861 &first_bad, &bad_sectors))
1862 set_bit(R5_MadeGood, &sh->dev[i].flags);
1863 }
1864 rdev_dec_pending(rdev, conf->mddev);
1865
1866 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1867 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1868 set_bit(STRIPE_HANDLE, &sh->state);
1869 release_stripe(sh);
1870}
1871
1872static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1873
1874static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1875{
1876 struct r5dev *dev = &sh->dev[i];
1877
1878 bio_init(&dev->req);
1879 dev->req.bi_io_vec = &dev->vec;
1880 dev->req.bi_vcnt++;
1881 dev->req.bi_max_vecs++;
1882 dev->req.bi_private = sh;
1883 dev->vec.bv_page = dev->page;
1884
1885 bio_init(&dev->rreq);
1886 dev->rreq.bi_io_vec = &dev->rvec;
1887 dev->rreq.bi_vcnt++;
1888 dev->rreq.bi_max_vecs++;
1889 dev->rreq.bi_private = sh;
1890 dev->rvec.bv_page = dev->page;
1891
1892 dev->flags = 0;
1893 dev->sector = compute_blocknr(sh, i, previous);
1894}
1895
1896static void error(struct mddev *mddev, struct md_rdev *rdev)
1897{
1898 char b[BDEVNAME_SIZE];
1899 struct r5conf *conf = mddev->private;
1900 unsigned long flags;
1901 pr_debug("raid456: error called\n");
1902
1903 spin_lock_irqsave(&conf->device_lock, flags);
1904 clear_bit(In_sync, &rdev->flags);
1905 mddev->degraded = calc_degraded(conf);
1906 spin_unlock_irqrestore(&conf->device_lock, flags);
1907 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1908
1909 set_bit(Blocked, &rdev->flags);
1910 set_bit(Faulty, &rdev->flags);
1911 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1912 printk(KERN_ALERT
1913 "md/raid:%s: Disk failure on %s, disabling device.\n"
1914 "md/raid:%s: Operation continuing on %d devices.\n",
1915 mdname(mddev),
1916 bdevname(rdev->bdev, b),
1917 mdname(mddev),
1918 conf->raid_disks - mddev->degraded);
1919}
1920
1921/*
1922 * Input: a 'big' sector number,
1923 * Output: index of the data and parity disk, and the sector # in them.
1924 */
1925static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1926 int previous, int *dd_idx,
1927 struct stripe_head *sh)
1928{
1929 sector_t stripe, stripe2;
1930 sector_t chunk_number;
1931 unsigned int chunk_offset;
1932 int pd_idx, qd_idx;
1933 int ddf_layout = 0;
1934 sector_t new_sector;
1935 int algorithm = previous ? conf->prev_algo
1936 : conf->algorithm;
1937 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1938 : conf->chunk_sectors;
1939 int raid_disks = previous ? conf->previous_raid_disks
1940 : conf->raid_disks;
1941 int data_disks = raid_disks - conf->max_degraded;
1942
1943 /* First compute the information on this sector */
1944
1945 /*
1946 * Compute the chunk number and the sector offset inside the chunk
1947 */
1948 chunk_offset = sector_div(r_sector, sectors_per_chunk);
1949 chunk_number = r_sector;
1950
1951 /*
1952 * Compute the stripe number
1953 */
1954 stripe = chunk_number;
1955 *dd_idx = sector_div(stripe, data_disks);
1956 stripe2 = stripe;
1957 /*
1958 * Select the parity disk based on the user selected algorithm.
1959 */
1960 pd_idx = qd_idx = -1;
1961 switch(conf->level) {
1962 case 4:
1963 pd_idx = data_disks;
1964 break;
1965 case 5:
1966 switch (algorithm) {
1967 case ALGORITHM_LEFT_ASYMMETRIC:
1968 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1969 if (*dd_idx >= pd_idx)
1970 (*dd_idx)++;
1971 break;
1972 case ALGORITHM_RIGHT_ASYMMETRIC:
1973 pd_idx = sector_div(stripe2, raid_disks);
1974 if (*dd_idx >= pd_idx)
1975 (*dd_idx)++;
1976 break;
1977 case ALGORITHM_LEFT_SYMMETRIC:
1978 pd_idx = data_disks - sector_div(stripe2, raid_disks);
1979 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1980 break;
1981 case ALGORITHM_RIGHT_SYMMETRIC:
1982 pd_idx = sector_div(stripe2, raid_disks);
1983 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1984 break;
1985 case ALGORITHM_PARITY_0:
1986 pd_idx = 0;
1987 (*dd_idx)++;
1988 break;
1989 case ALGORITHM_PARITY_N:
1990 pd_idx = data_disks;
1991 break;
1992 default:
1993 BUG();
1994 }
1995 break;
1996 case 6:
1997
1998 switch (algorithm) {
1999 case ALGORITHM_LEFT_ASYMMETRIC:
2000 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2001 qd_idx = pd_idx + 1;
2002 if (pd_idx == raid_disks-1) {
2003 (*dd_idx)++; /* Q D D D P */
2004 qd_idx = 0;
2005 } else if (*dd_idx >= pd_idx)
2006 (*dd_idx) += 2; /* D D P Q D */
2007 break;
2008 case ALGORITHM_RIGHT_ASYMMETRIC:
2009 pd_idx = sector_div(stripe2, raid_disks);
2010 qd_idx = pd_idx + 1;
2011 if (pd_idx == raid_disks-1) {
2012 (*dd_idx)++; /* Q D D D P */
2013 qd_idx = 0;
2014 } else if (*dd_idx >= pd_idx)
2015 (*dd_idx) += 2; /* D D P Q D */
2016 break;
2017 case ALGORITHM_LEFT_SYMMETRIC:
2018 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2019 qd_idx = (pd_idx + 1) % raid_disks;
2020 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2021 break;
2022 case ALGORITHM_RIGHT_SYMMETRIC:
2023 pd_idx = sector_div(stripe2, raid_disks);
2024 qd_idx = (pd_idx + 1) % raid_disks;
2025 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2026 break;
2027
2028 case ALGORITHM_PARITY_0:
2029 pd_idx = 0;
2030 qd_idx = 1;
2031 (*dd_idx) += 2;
2032 break;
2033 case ALGORITHM_PARITY_N:
2034 pd_idx = data_disks;
2035 qd_idx = data_disks + 1;
2036 break;
2037
2038 case ALGORITHM_ROTATING_ZERO_RESTART:
2039 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2040 * of blocks for computing Q is different.
2041 */
2042 pd_idx = sector_div(stripe2, raid_disks);
2043 qd_idx = pd_idx + 1;
2044 if (pd_idx == raid_disks-1) {
2045 (*dd_idx)++; /* Q D D D P */
2046 qd_idx = 0;
2047 } else if (*dd_idx >= pd_idx)
2048 (*dd_idx) += 2; /* D D P Q D */
2049 ddf_layout = 1;
2050 break;
2051
2052 case ALGORITHM_ROTATING_N_RESTART:
2053 /* Same a left_asymmetric, by first stripe is
2054 * D D D P Q rather than
2055 * Q D D D P
2056 */
2057 stripe2 += 1;
2058 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2059 qd_idx = pd_idx + 1;
2060 if (pd_idx == raid_disks-1) {
2061 (*dd_idx)++; /* Q D D D P */
2062 qd_idx = 0;
2063 } else if (*dd_idx >= pd_idx)
2064 (*dd_idx) += 2; /* D D P Q D */
2065 ddf_layout = 1;
2066 break;
2067
2068 case ALGORITHM_ROTATING_N_CONTINUE:
2069 /* Same as left_symmetric but Q is before P */
2070 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2071 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2072 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2073 ddf_layout = 1;
2074 break;
2075
2076 case ALGORITHM_LEFT_ASYMMETRIC_6:
2077 /* RAID5 left_asymmetric, with Q on last device */
2078 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2079 if (*dd_idx >= pd_idx)
2080 (*dd_idx)++;
2081 qd_idx = raid_disks - 1;
2082 break;
2083
2084 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2085 pd_idx = sector_div(stripe2, raid_disks-1);
2086 if (*dd_idx >= pd_idx)
2087 (*dd_idx)++;
2088 qd_idx = raid_disks - 1;
2089 break;
2090
2091 case ALGORITHM_LEFT_SYMMETRIC_6:
2092 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2093 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2094 qd_idx = raid_disks - 1;
2095 break;
2096
2097 case ALGORITHM_RIGHT_SYMMETRIC_6:
2098 pd_idx = sector_div(stripe2, raid_disks-1);
2099 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2100 qd_idx = raid_disks - 1;
2101 break;
2102
2103 case ALGORITHM_PARITY_0_6:
2104 pd_idx = 0;
2105 (*dd_idx)++;
2106 qd_idx = raid_disks - 1;
2107 break;
2108
2109 default:
2110 BUG();
2111 }
2112 break;
2113 }
2114
2115 if (sh) {
2116 sh->pd_idx = pd_idx;
2117 sh->qd_idx = qd_idx;
2118 sh->ddf_layout = ddf_layout;
2119 }
2120 /*
2121 * Finally, compute the new sector number
2122 */
2123 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2124 return new_sector;
2125}
2126
2127
2128static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2129{
2130 struct r5conf *conf = sh->raid_conf;
2131 int raid_disks = sh->disks;
2132 int data_disks = raid_disks - conf->max_degraded;
2133 sector_t new_sector = sh->sector, check;
2134 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2135 : conf->chunk_sectors;
2136 int algorithm = previous ? conf->prev_algo
2137 : conf->algorithm;
2138 sector_t stripe;
2139 int chunk_offset;
2140 sector_t chunk_number;
2141 int dummy1, dd_idx = i;
2142 sector_t r_sector;
2143 struct stripe_head sh2;
2144
2145
2146 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2147 stripe = new_sector;
2148
2149 if (i == sh->pd_idx)
2150 return 0;
2151 switch(conf->level) {
2152 case 4: break;
2153 case 5:
2154 switch (algorithm) {
2155 case ALGORITHM_LEFT_ASYMMETRIC:
2156 case ALGORITHM_RIGHT_ASYMMETRIC:
2157 if (i > sh->pd_idx)
2158 i--;
2159 break;
2160 case ALGORITHM_LEFT_SYMMETRIC:
2161 case ALGORITHM_RIGHT_SYMMETRIC:
2162 if (i < sh->pd_idx)
2163 i += raid_disks;
2164 i -= (sh->pd_idx + 1);
2165 break;
2166 case ALGORITHM_PARITY_0:
2167 i -= 1;
2168 break;
2169 case ALGORITHM_PARITY_N:
2170 break;
2171 default:
2172 BUG();
2173 }
2174 break;
2175 case 6:
2176 if (i == sh->qd_idx)
2177 return 0; /* It is the Q disk */
2178 switch (algorithm) {
2179 case ALGORITHM_LEFT_ASYMMETRIC:
2180 case ALGORITHM_RIGHT_ASYMMETRIC:
2181 case ALGORITHM_ROTATING_ZERO_RESTART:
2182 case ALGORITHM_ROTATING_N_RESTART:
2183 if (sh->pd_idx == raid_disks-1)
2184 i--; /* Q D D D P */
2185 else if (i > sh->pd_idx)
2186 i -= 2; /* D D P Q D */
2187 break;
2188 case ALGORITHM_LEFT_SYMMETRIC:
2189 case ALGORITHM_RIGHT_SYMMETRIC:
2190 if (sh->pd_idx == raid_disks-1)
2191 i--; /* Q D D D P */
2192 else {
2193 /* D D P Q D */
2194 if (i < sh->pd_idx)
2195 i += raid_disks;
2196 i -= (sh->pd_idx + 2);
2197 }
2198 break;
2199 case ALGORITHM_PARITY_0:
2200 i -= 2;
2201 break;
2202 case ALGORITHM_PARITY_N:
2203 break;
2204 case ALGORITHM_ROTATING_N_CONTINUE:
2205 /* Like left_symmetric, but P is before Q */
2206 if (sh->pd_idx == 0)
2207 i--; /* P D D D Q */
2208 else {
2209 /* D D Q P D */
2210 if (i < sh->pd_idx)
2211 i += raid_disks;
2212 i -= (sh->pd_idx + 1);
2213 }
2214 break;
2215 case ALGORITHM_LEFT_ASYMMETRIC_6:
2216 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2217 if (i > sh->pd_idx)
2218 i--;
2219 break;
2220 case ALGORITHM_LEFT_SYMMETRIC_6:
2221 case ALGORITHM_RIGHT_SYMMETRIC_6:
2222 if (i < sh->pd_idx)
2223 i += data_disks + 1;
2224 i -= (sh->pd_idx + 1);
2225 break;
2226 case ALGORITHM_PARITY_0_6:
2227 i -= 1;
2228 break;
2229 default:
2230 BUG();
2231 }
2232 break;
2233 }
2234
2235 chunk_number = stripe * data_disks + i;
2236 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2237
2238 check = raid5_compute_sector(conf, r_sector,
2239 previous, &dummy1, &sh2);
2240 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2241 || sh2.qd_idx != sh->qd_idx) {
2242 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2243 mdname(conf->mddev));
2244 return 0;
2245 }
2246 return r_sector;
2247}
2248
2249
2250static void
2251schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2252 int rcw, int expand)
2253{
2254 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2255 struct r5conf *conf = sh->raid_conf;
2256 int level = conf->level;
2257
2258 if (rcw) {
2259 /* if we are not expanding this is a proper write request, and
2260 * there will be bios with new data to be drained into the
2261 * stripe cache
2262 */
2263 if (!expand) {
2264 sh->reconstruct_state = reconstruct_state_drain_run;
2265 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2266 } else
2267 sh->reconstruct_state = reconstruct_state_run;
2268
2269 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2270
2271 for (i = disks; i--; ) {
2272 struct r5dev *dev = &sh->dev[i];
2273
2274 if (dev->towrite) {
2275 set_bit(R5_LOCKED, &dev->flags);
2276 set_bit(R5_Wantdrain, &dev->flags);
2277 if (!expand)
2278 clear_bit(R5_UPTODATE, &dev->flags);
2279 s->locked++;
2280 }
2281 }
2282 if (s->locked + conf->max_degraded == disks)
2283 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2284 atomic_inc(&conf->pending_full_writes);
2285 } else {
2286 BUG_ON(level == 6);
2287 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2288 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2289
2290 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2291 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2292 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2293 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2294
2295 for (i = disks; i--; ) {
2296 struct r5dev *dev = &sh->dev[i];
2297 if (i == pd_idx)
2298 continue;
2299
2300 if (dev->towrite &&
2301 (test_bit(R5_UPTODATE, &dev->flags) ||
2302 test_bit(R5_Wantcompute, &dev->flags))) {
2303 set_bit(R5_Wantdrain, &dev->flags);
2304 set_bit(R5_LOCKED, &dev->flags);
2305 clear_bit(R5_UPTODATE, &dev->flags);
2306 s->locked++;
2307 }
2308 }
2309 }
2310
2311 /* keep the parity disk(s) locked while asynchronous operations
2312 * are in flight
2313 */
2314 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2315 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2316 s->locked++;
2317
2318 if (level == 6) {
2319 int qd_idx = sh->qd_idx;
2320 struct r5dev *dev = &sh->dev[qd_idx];
2321
2322 set_bit(R5_LOCKED, &dev->flags);
2323 clear_bit(R5_UPTODATE, &dev->flags);
2324 s->locked++;
2325 }
2326
2327 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2328 __func__, (unsigned long long)sh->sector,
2329 s->locked, s->ops_request);
2330}
2331
2332/*
2333 * Each stripe/dev can have one or more bion attached.
2334 * toread/towrite point to the first in a chain.
2335 * The bi_next chain must be in order.
2336 */
2337static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2338{
2339 struct bio **bip;
2340 struct r5conf *conf = sh->raid_conf;
2341 int firstwrite=0;
2342
2343 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2344 (unsigned long long)bi->bi_sector,
2345 (unsigned long long)sh->sector);
2346
2347
2348 spin_lock_irq(&conf->device_lock);
2349 if (forwrite) {
2350 bip = &sh->dev[dd_idx].towrite;
2351 if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2352 firstwrite = 1;
2353 } else
2354 bip = &sh->dev[dd_idx].toread;
2355 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2356 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2357 goto overlap;
2358 bip = & (*bip)->bi_next;
2359 }
2360 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2361 goto overlap;
2362
2363 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2364 if (*bip)
2365 bi->bi_next = *bip;
2366 *bip = bi;
2367 bi->bi_phys_segments++;
2368
2369 if (forwrite) {
2370 /* check if page is covered */
2371 sector_t sector = sh->dev[dd_idx].sector;
2372 for (bi=sh->dev[dd_idx].towrite;
2373 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2374 bi && bi->bi_sector <= sector;
2375 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2376 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2377 sector = bi->bi_sector + (bi->bi_size>>9);
2378 }
2379 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2380 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2381 }
2382 spin_unlock_irq(&conf->device_lock);
2383
2384 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2385 (unsigned long long)(*bip)->bi_sector,
2386 (unsigned long long)sh->sector, dd_idx);
2387
2388 if (conf->mddev->bitmap && firstwrite) {
2389 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2390 STRIPE_SECTORS, 0);
2391 sh->bm_seq = conf->seq_flush+1;
2392 set_bit(STRIPE_BIT_DELAY, &sh->state);
2393 }
2394 return 1;
2395
2396 overlap:
2397 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2398 spin_unlock_irq(&conf->device_lock);
2399 return 0;
2400}
2401
2402static void end_reshape(struct r5conf *conf);
2403
2404static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2405 struct stripe_head *sh)
2406{
2407 int sectors_per_chunk =
2408 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2409 int dd_idx;
2410 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2411 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2412
2413 raid5_compute_sector(conf,
2414 stripe * (disks - conf->max_degraded)
2415 *sectors_per_chunk + chunk_offset,
2416 previous,
2417 &dd_idx, sh);
2418}
2419
2420static void
2421handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2422 struct stripe_head_state *s, int disks,
2423 struct bio **return_bi)
2424{
2425 int i;
2426 for (i = disks; i--; ) {
2427 struct bio *bi;
2428 int bitmap_end = 0;
2429
2430 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2431 struct md_rdev *rdev;
2432 rcu_read_lock();
2433 rdev = rcu_dereference(conf->disks[i].rdev);
2434 if (rdev && test_bit(In_sync, &rdev->flags))
2435 atomic_inc(&rdev->nr_pending);
2436 else
2437 rdev = NULL;
2438 rcu_read_unlock();
2439 if (rdev) {
2440 if (!rdev_set_badblocks(
2441 rdev,
2442 sh->sector,
2443 STRIPE_SECTORS, 0))
2444 md_error(conf->mddev, rdev);
2445 rdev_dec_pending(rdev, conf->mddev);
2446 }
2447 }
2448 spin_lock_irq(&conf->device_lock);
2449 /* fail all writes first */
2450 bi = sh->dev[i].towrite;
2451 sh->dev[i].towrite = NULL;
2452 if (bi) {
2453 s->to_write--;
2454 bitmap_end = 1;
2455 }
2456
2457 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2458 wake_up(&conf->wait_for_overlap);
2459
2460 while (bi && bi->bi_sector <
2461 sh->dev[i].sector + STRIPE_SECTORS) {
2462 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2463 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2464 if (!raid5_dec_bi_phys_segments(bi)) {
2465 md_write_end(conf->mddev);
2466 bi->bi_next = *return_bi;
2467 *return_bi = bi;
2468 }
2469 bi = nextbi;
2470 }
2471 /* and fail all 'written' */
2472 bi = sh->dev[i].written;
2473 sh->dev[i].written = NULL;
2474 if (bi) bitmap_end = 1;
2475 while (bi && bi->bi_sector <
2476 sh->dev[i].sector + STRIPE_SECTORS) {
2477 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2478 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2479 if (!raid5_dec_bi_phys_segments(bi)) {
2480 md_write_end(conf->mddev);
2481 bi->bi_next = *return_bi;
2482 *return_bi = bi;
2483 }
2484 bi = bi2;
2485 }
2486
2487 /* fail any reads if this device is non-operational and
2488 * the data has not reached the cache yet.
2489 */
2490 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2491 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2492 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2493 bi = sh->dev[i].toread;
2494 sh->dev[i].toread = NULL;
2495 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2496 wake_up(&conf->wait_for_overlap);
2497 if (bi) s->to_read--;
2498 while (bi && bi->bi_sector <
2499 sh->dev[i].sector + STRIPE_SECTORS) {
2500 struct bio *nextbi =
2501 r5_next_bio(bi, sh->dev[i].sector);
2502 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2503 if (!raid5_dec_bi_phys_segments(bi)) {
2504 bi->bi_next = *return_bi;
2505 *return_bi = bi;
2506 }
2507 bi = nextbi;
2508 }
2509 }
2510 spin_unlock_irq(&conf->device_lock);
2511 if (bitmap_end)
2512 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2513 STRIPE_SECTORS, 0, 0);
2514 /* If we were in the middle of a write the parity block might
2515 * still be locked - so just clear all R5_LOCKED flags
2516 */
2517 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2518 }
2519
2520 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2521 if (atomic_dec_and_test(&conf->pending_full_writes))
2522 md_wakeup_thread(conf->mddev->thread);
2523}
2524
2525static void
2526handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2527 struct stripe_head_state *s)
2528{
2529 int abort = 0;
2530 int i;
2531
2532 clear_bit(STRIPE_SYNCING, &sh->state);
2533 s->syncing = 0;
2534 s->replacing = 0;
2535 /* There is nothing more to do for sync/check/repair.
2536 * Don't even need to abort as that is handled elsewhere
2537 * if needed, and not always wanted e.g. if there is a known
2538 * bad block here.
2539 * For recover/replace we need to record a bad block on all
2540 * non-sync devices, or abort the recovery
2541 */
2542 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2543 /* During recovery devices cannot be removed, so
2544 * locking and refcounting of rdevs is not needed
2545 */
2546 for (i = 0; i < conf->raid_disks; i++) {
2547 struct md_rdev *rdev = conf->disks[i].rdev;
2548 if (rdev
2549 && !test_bit(Faulty, &rdev->flags)
2550 && !test_bit(In_sync, &rdev->flags)
2551 && !rdev_set_badblocks(rdev, sh->sector,
2552 STRIPE_SECTORS, 0))
2553 abort = 1;
2554 rdev = conf->disks[i].replacement;
2555 if (rdev
2556 && !test_bit(Faulty, &rdev->flags)
2557 && !test_bit(In_sync, &rdev->flags)
2558 && !rdev_set_badblocks(rdev, sh->sector,
2559 STRIPE_SECTORS, 0))
2560 abort = 1;
2561 }
2562 if (abort)
2563 conf->recovery_disabled =
2564 conf->mddev->recovery_disabled;
2565 }
2566 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2567}
2568
2569static int want_replace(struct stripe_head *sh, int disk_idx)
2570{
2571 struct md_rdev *rdev;
2572 int rv = 0;
2573 /* Doing recovery so rcu locking not required */
2574 rdev = sh->raid_conf->disks[disk_idx].replacement;
2575 if (rdev
2576 && !test_bit(Faulty, &rdev->flags)
2577 && !test_bit(In_sync, &rdev->flags)
2578 && (rdev->recovery_offset <= sh->sector
2579 || rdev->mddev->recovery_cp <= sh->sector))
2580 rv = 1;
2581
2582 return rv;
2583}
2584
2585/* fetch_block - checks the given member device to see if its data needs
2586 * to be read or computed to satisfy a request.
2587 *
2588 * Returns 1 when no more member devices need to be checked, otherwise returns
2589 * 0 to tell the loop in handle_stripe_fill to continue
2590 */
2591static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2592 int disk_idx, int disks)
2593{
2594 struct r5dev *dev = &sh->dev[disk_idx];
2595 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2596 &sh->dev[s->failed_num[1]] };
2597
2598 /* is the data in this block needed, and can we get it? */
2599 if (!test_bit(R5_LOCKED, &dev->flags) &&
2600 !test_bit(R5_UPTODATE, &dev->flags) &&
2601 (dev->toread ||
2602 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2603 s->syncing || s->expanding ||
2604 (s->replacing && want_replace(sh, disk_idx)) ||
2605 (s->failed >= 1 && fdev[0]->toread) ||
2606 (s->failed >= 2 && fdev[1]->toread) ||
2607 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2608 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2609 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2610 /* we would like to get this block, possibly by computing it,
2611 * otherwise read it if the backing disk is insync
2612 */
2613 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2614 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2615 if ((s->uptodate == disks - 1) &&
2616 (s->failed && (disk_idx == s->failed_num[0] ||
2617 disk_idx == s->failed_num[1]))) {
2618 /* have disk failed, and we're requested to fetch it;
2619 * do compute it
2620 */
2621 pr_debug("Computing stripe %llu block %d\n",
2622 (unsigned long long)sh->sector, disk_idx);
2623 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2624 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2625 set_bit(R5_Wantcompute, &dev->flags);
2626 sh->ops.target = disk_idx;
2627 sh->ops.target2 = -1; /* no 2nd target */
2628 s->req_compute = 1;
2629 /* Careful: from this point on 'uptodate' is in the eye
2630 * of raid_run_ops which services 'compute' operations
2631 * before writes. R5_Wantcompute flags a block that will
2632 * be R5_UPTODATE by the time it is needed for a
2633 * subsequent operation.
2634 */
2635 s->uptodate++;
2636 return 1;
2637 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2638 /* Computing 2-failure is *very* expensive; only
2639 * do it if failed >= 2
2640 */
2641 int other;
2642 for (other = disks; other--; ) {
2643 if (other == disk_idx)
2644 continue;
2645 if (!test_bit(R5_UPTODATE,
2646 &sh->dev[other].flags))
2647 break;
2648 }
2649 BUG_ON(other < 0);
2650 pr_debug("Computing stripe %llu blocks %d,%d\n",
2651 (unsigned long long)sh->sector,
2652 disk_idx, other);
2653 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2654 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2655 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2656 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2657 sh->ops.target = disk_idx;
2658 sh->ops.target2 = other;
2659 s->uptodate += 2;
2660 s->req_compute = 1;
2661 return 1;
2662 } else if (test_bit(R5_Insync, &dev->flags)) {
2663 set_bit(R5_LOCKED, &dev->flags);
2664 set_bit(R5_Wantread, &dev->flags);
2665 s->locked++;
2666 pr_debug("Reading block %d (sync=%d)\n",
2667 disk_idx, s->syncing);
2668 }
2669 }
2670
2671 return 0;
2672}
2673
2674/**
2675 * handle_stripe_fill - read or compute data to satisfy pending requests.
2676 */
2677static void handle_stripe_fill(struct stripe_head *sh,
2678 struct stripe_head_state *s,
2679 int disks)
2680{
2681 int i;
2682
2683 /* look for blocks to read/compute, skip this if a compute
2684 * is already in flight, or if the stripe contents are in the
2685 * midst of changing due to a write
2686 */
2687 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2688 !sh->reconstruct_state)
2689 for (i = disks; i--; )
2690 if (fetch_block(sh, s, i, disks))
2691 break;
2692 set_bit(STRIPE_HANDLE, &sh->state);
2693}
2694
2695
2696/* handle_stripe_clean_event
2697 * any written block on an uptodate or failed drive can be returned.
2698 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2699 * never LOCKED, so we don't need to test 'failed' directly.
2700 */
2701static void handle_stripe_clean_event(struct r5conf *conf,
2702 struct stripe_head *sh, int disks, struct bio **return_bi)
2703{
2704 int i;
2705 struct r5dev *dev;
2706
2707 for (i = disks; i--; )
2708 if (sh->dev[i].written) {
2709 dev = &sh->dev[i];
2710 if (!test_bit(R5_LOCKED, &dev->flags) &&
2711 test_bit(R5_UPTODATE, &dev->flags)) {
2712 /* We can return any write requests */
2713 struct bio *wbi, *wbi2;
2714 int bitmap_end = 0;
2715 pr_debug("Return write for disc %d\n", i);
2716 spin_lock_irq(&conf->device_lock);
2717 wbi = dev->written;
2718 dev->written = NULL;
2719 while (wbi && wbi->bi_sector <
2720 dev->sector + STRIPE_SECTORS) {
2721 wbi2 = r5_next_bio(wbi, dev->sector);
2722 if (!raid5_dec_bi_phys_segments(wbi)) {
2723 md_write_end(conf->mddev);
2724 wbi->bi_next = *return_bi;
2725 *return_bi = wbi;
2726 }
2727 wbi = wbi2;
2728 }
2729 if (dev->towrite == NULL)
2730 bitmap_end = 1;
2731 spin_unlock_irq(&conf->device_lock);
2732 if (bitmap_end)
2733 bitmap_endwrite(conf->mddev->bitmap,
2734 sh->sector,
2735 STRIPE_SECTORS,
2736 !test_bit(STRIPE_DEGRADED, &sh->state),
2737 0);
2738 }
2739 }
2740
2741 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2742 if (atomic_dec_and_test(&conf->pending_full_writes))
2743 md_wakeup_thread(conf->mddev->thread);
2744}
2745
2746static void handle_stripe_dirtying(struct r5conf *conf,
2747 struct stripe_head *sh,
2748 struct stripe_head_state *s,
2749 int disks)
2750{
2751 int rmw = 0, rcw = 0, i;
2752 if (conf->max_degraded == 2) {
2753 /* RAID6 requires 'rcw' in current implementation
2754 * Calculate the real rcw later - for now fake it
2755 * look like rcw is cheaper
2756 */
2757 rcw = 1; rmw = 2;
2758 } else for (i = disks; i--; ) {
2759 /* would I have to read this buffer for read_modify_write */
2760 struct r5dev *dev = &sh->dev[i];
2761 if ((dev->towrite || i == sh->pd_idx) &&
2762 !test_bit(R5_LOCKED, &dev->flags) &&
2763 !(test_bit(R5_UPTODATE, &dev->flags) ||
2764 test_bit(R5_Wantcompute, &dev->flags))) {
2765 if (test_bit(R5_Insync, &dev->flags))
2766 rmw++;
2767 else
2768 rmw += 2*disks; /* cannot read it */
2769 }
2770 /* Would I have to read this buffer for reconstruct_write */
2771 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2772 !test_bit(R5_LOCKED, &dev->flags) &&
2773 !(test_bit(R5_UPTODATE, &dev->flags) ||
2774 test_bit(R5_Wantcompute, &dev->flags))) {
2775 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2776 else
2777 rcw += 2*disks;
2778 }
2779 }
2780 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2781 (unsigned long long)sh->sector, rmw, rcw);
2782 set_bit(STRIPE_HANDLE, &sh->state);
2783 if (rmw < rcw && rmw > 0)
2784 /* prefer read-modify-write, but need to get some data */
2785 for (i = disks; i--; ) {
2786 struct r5dev *dev = &sh->dev[i];
2787 if ((dev->towrite || i == sh->pd_idx) &&
2788 !test_bit(R5_LOCKED, &dev->flags) &&
2789 !(test_bit(R5_UPTODATE, &dev->flags) ||
2790 test_bit(R5_Wantcompute, &dev->flags)) &&
2791 test_bit(R5_Insync, &dev->flags)) {
2792 if (
2793 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2794 pr_debug("Read_old block "
2795 "%d for r-m-w\n", i);
2796 set_bit(R5_LOCKED, &dev->flags);
2797 set_bit(R5_Wantread, &dev->flags);
2798 s->locked++;
2799 } else {
2800 set_bit(STRIPE_DELAYED, &sh->state);
2801 set_bit(STRIPE_HANDLE, &sh->state);
2802 }
2803 }
2804 }
2805 if (rcw <= rmw && rcw > 0) {
2806 /* want reconstruct write, but need to get some data */
2807 rcw = 0;
2808 for (i = disks; i--; ) {
2809 struct r5dev *dev = &sh->dev[i];
2810 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2811 i != sh->pd_idx && i != sh->qd_idx &&
2812 !test_bit(R5_LOCKED, &dev->flags) &&
2813 !(test_bit(R5_UPTODATE, &dev->flags) ||
2814 test_bit(R5_Wantcompute, &dev->flags))) {
2815 rcw++;
2816 if (!test_bit(R5_Insync, &dev->flags))
2817 continue; /* it's a failed drive */
2818 if (
2819 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2820 pr_debug("Read_old block "
2821 "%d for Reconstruct\n", i);
2822 set_bit(R5_LOCKED, &dev->flags);
2823 set_bit(R5_Wantread, &dev->flags);
2824 s->locked++;
2825 } else {
2826 set_bit(STRIPE_DELAYED, &sh->state);
2827 set_bit(STRIPE_HANDLE, &sh->state);
2828 }
2829 }
2830 }
2831 }
2832 /* now if nothing is locked, and if we have enough data,
2833 * we can start a write request
2834 */
2835 /* since handle_stripe can be called at any time we need to handle the
2836 * case where a compute block operation has been submitted and then a
2837 * subsequent call wants to start a write request. raid_run_ops only
2838 * handles the case where compute block and reconstruct are requested
2839 * simultaneously. If this is not the case then new writes need to be
2840 * held off until the compute completes.
2841 */
2842 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2843 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2844 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2845 schedule_reconstruction(sh, s, rcw == 0, 0);
2846}
2847
2848static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2849 struct stripe_head_state *s, int disks)
2850{
2851 struct r5dev *dev = NULL;
2852
2853 set_bit(STRIPE_HANDLE, &sh->state);
2854
2855 switch (sh->check_state) {
2856 case check_state_idle:
2857 /* start a new check operation if there are no failures */
2858 if (s->failed == 0) {
2859 BUG_ON(s->uptodate != disks);
2860 sh->check_state = check_state_run;
2861 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2862 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2863 s->uptodate--;
2864 break;
2865 }
2866 dev = &sh->dev[s->failed_num[0]];
2867 /* fall through */
2868 case check_state_compute_result:
2869 sh->check_state = check_state_idle;
2870 if (!dev)
2871 dev = &sh->dev[sh->pd_idx];
2872
2873 /* check that a write has not made the stripe insync */
2874 if (test_bit(STRIPE_INSYNC, &sh->state))
2875 break;
2876
2877 /* either failed parity check, or recovery is happening */
2878 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2879 BUG_ON(s->uptodate != disks);
2880
2881 set_bit(R5_LOCKED, &dev->flags);
2882 s->locked++;
2883 set_bit(R5_Wantwrite, &dev->flags);
2884
2885 clear_bit(STRIPE_DEGRADED, &sh->state);
2886 set_bit(STRIPE_INSYNC, &sh->state);
2887 break;
2888 case check_state_run:
2889 break; /* we will be called again upon completion */
2890 case check_state_check_result:
2891 sh->check_state = check_state_idle;
2892
2893 /* if a failure occurred during the check operation, leave
2894 * STRIPE_INSYNC not set and let the stripe be handled again
2895 */
2896 if (s->failed)
2897 break;
2898
2899 /* handle a successful check operation, if parity is correct
2900 * we are done. Otherwise update the mismatch count and repair
2901 * parity if !MD_RECOVERY_CHECK
2902 */
2903 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2904 /* parity is correct (on disc,
2905 * not in buffer any more)
2906 */
2907 set_bit(STRIPE_INSYNC, &sh->state);
2908 else {
2909 conf->mddev->resync_mismatches += STRIPE_SECTORS;
2910 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2911 /* don't try to repair!! */
2912 set_bit(STRIPE_INSYNC, &sh->state);
2913 else {
2914 sh->check_state = check_state_compute_run;
2915 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2916 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2917 set_bit(R5_Wantcompute,
2918 &sh->dev[sh->pd_idx].flags);
2919 sh->ops.target = sh->pd_idx;
2920 sh->ops.target2 = -1;
2921 s->uptodate++;
2922 }
2923 }
2924 break;
2925 case check_state_compute_run:
2926 break;
2927 default:
2928 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2929 __func__, sh->check_state,
2930 (unsigned long long) sh->sector);
2931 BUG();
2932 }
2933}
2934
2935
2936static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2937 struct stripe_head_state *s,
2938 int disks)
2939{
2940 int pd_idx = sh->pd_idx;
2941 int qd_idx = sh->qd_idx;
2942 struct r5dev *dev;
2943
2944 set_bit(STRIPE_HANDLE, &sh->state);
2945
2946 BUG_ON(s->failed > 2);
2947
2948 /* Want to check and possibly repair P and Q.
2949 * However there could be one 'failed' device, in which
2950 * case we can only check one of them, possibly using the
2951 * other to generate missing data
2952 */
2953
2954 switch (sh->check_state) {
2955 case check_state_idle:
2956 /* start a new check operation if there are < 2 failures */
2957 if (s->failed == s->q_failed) {
2958 /* The only possible failed device holds Q, so it
2959 * makes sense to check P (If anything else were failed,
2960 * we would have used P to recreate it).
2961 */
2962 sh->check_state = check_state_run;
2963 }
2964 if (!s->q_failed && s->failed < 2) {
2965 /* Q is not failed, and we didn't use it to generate
2966 * anything, so it makes sense to check it
2967 */
2968 if (sh->check_state == check_state_run)
2969 sh->check_state = check_state_run_pq;
2970 else
2971 sh->check_state = check_state_run_q;
2972 }
2973
2974 /* discard potentially stale zero_sum_result */
2975 sh->ops.zero_sum_result = 0;
2976
2977 if (sh->check_state == check_state_run) {
2978 /* async_xor_zero_sum destroys the contents of P */
2979 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2980 s->uptodate--;
2981 }
2982 if (sh->check_state >= check_state_run &&
2983 sh->check_state <= check_state_run_pq) {
2984 /* async_syndrome_zero_sum preserves P and Q, so
2985 * no need to mark them !uptodate here
2986 */
2987 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2988 break;
2989 }
2990
2991 /* we have 2-disk failure */
2992 BUG_ON(s->failed != 2);
2993 /* fall through */
2994 case check_state_compute_result:
2995 sh->check_state = check_state_idle;
2996
2997 /* check that a write has not made the stripe insync */
2998 if (test_bit(STRIPE_INSYNC, &sh->state))
2999 break;
3000
3001 /* now write out any block on a failed drive,
3002 * or P or Q if they were recomputed
3003 */
3004 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3005 if (s->failed == 2) {
3006 dev = &sh->dev[s->failed_num[1]];
3007 s->locked++;
3008 set_bit(R5_LOCKED, &dev->flags);
3009 set_bit(R5_Wantwrite, &dev->flags);
3010 }
3011 if (s->failed >= 1) {
3012 dev = &sh->dev[s->failed_num[0]];
3013 s->locked++;
3014 set_bit(R5_LOCKED, &dev->flags);
3015 set_bit(R5_Wantwrite, &dev->flags);
3016 }
3017 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3018 dev = &sh->dev[pd_idx];
3019 s->locked++;
3020 set_bit(R5_LOCKED, &dev->flags);
3021 set_bit(R5_Wantwrite, &dev->flags);
3022 }
3023 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3024 dev = &sh->dev[qd_idx];
3025 s->locked++;
3026 set_bit(R5_LOCKED, &dev->flags);
3027 set_bit(R5_Wantwrite, &dev->flags);
3028 }
3029 clear_bit(STRIPE_DEGRADED, &sh->state);
3030
3031 set_bit(STRIPE_INSYNC, &sh->state);
3032 break;
3033 case check_state_run:
3034 case check_state_run_q:
3035 case check_state_run_pq:
3036 break; /* we will be called again upon completion */
3037 case check_state_check_result:
3038 sh->check_state = check_state_idle;
3039
3040 /* handle a successful check operation, if parity is correct
3041 * we are done. Otherwise update the mismatch count and repair
3042 * parity if !MD_RECOVERY_CHECK
3043 */
3044 if (sh->ops.zero_sum_result == 0) {
3045 /* both parities are correct */
3046 if (!s->failed)
3047 set_bit(STRIPE_INSYNC, &sh->state);
3048 else {
3049 /* in contrast to the raid5 case we can validate
3050 * parity, but still have a failure to write
3051 * back
3052 */
3053 sh->check_state = check_state_compute_result;
3054 /* Returning at this point means that we may go
3055 * off and bring p and/or q uptodate again so
3056 * we make sure to check zero_sum_result again
3057 * to verify if p or q need writeback
3058 */
3059 }
3060 } else {
3061 conf->mddev->resync_mismatches += STRIPE_SECTORS;
3062 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3063 /* don't try to repair!! */
3064 set_bit(STRIPE_INSYNC, &sh->state);
3065 else {
3066 int *target = &sh->ops.target;
3067
3068 sh->ops.target = -1;
3069 sh->ops.target2 = -1;
3070 sh->check_state = check_state_compute_run;
3071 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3072 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3073 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3074 set_bit(R5_Wantcompute,
3075 &sh->dev[pd_idx].flags);
3076 *target = pd_idx;
3077 target = &sh->ops.target2;
3078 s->uptodate++;
3079 }
3080 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3081 set_bit(R5_Wantcompute,
3082 &sh->dev[qd_idx].flags);
3083 *target = qd_idx;
3084 s->uptodate++;
3085 }
3086 }
3087 }
3088 break;
3089 case check_state_compute_run:
3090 break;
3091 default:
3092 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3093 __func__, sh->check_state,
3094 (unsigned long long) sh->sector);
3095 BUG();
3096 }
3097}
3098
3099static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3100{
3101 int i;
3102
3103 /* We have read all the blocks in this stripe and now we need to
3104 * copy some of them into a target stripe for expand.
3105 */
3106 struct dma_async_tx_descriptor *tx = NULL;
3107 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3108 for (i = 0; i < sh->disks; i++)
3109 if (i != sh->pd_idx && i != sh->qd_idx) {
3110 int dd_idx, j;
3111 struct stripe_head *sh2;
3112 struct async_submit_ctl submit;
3113
3114 sector_t bn = compute_blocknr(sh, i, 1);
3115 sector_t s = raid5_compute_sector(conf, bn, 0,
3116 &dd_idx, NULL);
3117 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3118 if (sh2 == NULL)
3119 /* so far only the early blocks of this stripe
3120 * have been requested. When later blocks
3121 * get requested, we will try again
3122 */
3123 continue;
3124 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3125 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3126 /* must have already done this block */
3127 release_stripe(sh2);
3128 continue;
3129 }
3130
3131 /* place all the copies on one channel */
3132 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3133 tx = async_memcpy(sh2->dev[dd_idx].page,
3134 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3135 &submit);
3136
3137 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3138 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3139 for (j = 0; j < conf->raid_disks; j++)
3140 if (j != sh2->pd_idx &&
3141 j != sh2->qd_idx &&
3142 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3143 break;
3144 if (j == conf->raid_disks) {
3145 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3146 set_bit(STRIPE_HANDLE, &sh2->state);
3147 }
3148 release_stripe(sh2);
3149
3150 }
3151 /* done submitting copies, wait for them to complete */
3152 if (tx) {
3153 async_tx_ack(tx);
3154 dma_wait_for_async_tx(tx);
3155 }
3156}
3157
3158/*
3159 * handle_stripe - do things to a stripe.
3160 *
3161 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3162 * state of various bits to see what needs to be done.
3163 * Possible results:
3164 * return some read requests which now have data
3165 * return some write requests which are safely on storage
3166 * schedule a read on some buffers
3167 * schedule a write of some buffers
3168 * return confirmation of parity correctness
3169 *
3170 */
3171
3172static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3173{
3174 struct r5conf *conf = sh->raid_conf;
3175 int disks = sh->disks;
3176 struct r5dev *dev;
3177 int i;
3178 int do_recovery = 0;
3179
3180 memset(s, 0, sizeof(*s));
3181
3182 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3183 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3184 s->failed_num[0] = -1;
3185 s->failed_num[1] = -1;
3186
3187 /* Now to look around and see what can be done */
3188 rcu_read_lock();
3189 spin_lock_irq(&conf->device_lock);
3190 for (i=disks; i--; ) {
3191 struct md_rdev *rdev;
3192 sector_t first_bad;
3193 int bad_sectors;
3194 int is_bad = 0;
3195
3196 dev = &sh->dev[i];
3197
3198 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3199 i, dev->flags,
3200 dev->toread, dev->towrite, dev->written);
3201 /* maybe we can reply to a read
3202 *
3203 * new wantfill requests are only permitted while
3204 * ops_complete_biofill is guaranteed to be inactive
3205 */
3206 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3207 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3208 set_bit(R5_Wantfill, &dev->flags);
3209
3210 /* now count some things */
3211 if (test_bit(R5_LOCKED, &dev->flags))
3212 s->locked++;
3213 if (test_bit(R5_UPTODATE, &dev->flags))
3214 s->uptodate++;
3215 if (test_bit(R5_Wantcompute, &dev->flags)) {
3216 s->compute++;
3217 BUG_ON(s->compute > 2);
3218 }
3219
3220 if (test_bit(R5_Wantfill, &dev->flags))
3221 s->to_fill++;
3222 else if (dev->toread)
3223 s->to_read++;
3224 if (dev->towrite) {
3225 s->to_write++;
3226 if (!test_bit(R5_OVERWRITE, &dev->flags))
3227 s->non_overwrite++;
3228 }
3229 if (dev->written)
3230 s->written++;
3231 /* Prefer to use the replacement for reads, but only
3232 * if it is recovered enough and has no bad blocks.
3233 */
3234 rdev = rcu_dereference(conf->disks[i].replacement);
3235 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3236 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3237 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3238 &first_bad, &bad_sectors))
3239 set_bit(R5_ReadRepl, &dev->flags);
3240 else {
3241 if (rdev)
3242 set_bit(R5_NeedReplace, &dev->flags);
3243 rdev = rcu_dereference(conf->disks[i].rdev);
3244 clear_bit(R5_ReadRepl, &dev->flags);
3245 }
3246 if (rdev && test_bit(Faulty, &rdev->flags))
3247 rdev = NULL;
3248 if (rdev) {
3249 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3250 &first_bad, &bad_sectors);
3251 if (s->blocked_rdev == NULL
3252 && (test_bit(Blocked, &rdev->flags)
3253 || is_bad < 0)) {
3254 if (is_bad < 0)
3255 set_bit(BlockedBadBlocks,
3256 &rdev->flags);
3257 s->blocked_rdev = rdev;
3258 atomic_inc(&rdev->nr_pending);
3259 }
3260 }
3261 clear_bit(R5_Insync, &dev->flags);
3262 if (!rdev)
3263 /* Not in-sync */;
3264 else if (is_bad) {
3265 /* also not in-sync */
3266 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3267 test_bit(R5_UPTODATE, &dev->flags)) {
3268 /* treat as in-sync, but with a read error
3269 * which we can now try to correct
3270 */
3271 set_bit(R5_Insync, &dev->flags);
3272 set_bit(R5_ReadError, &dev->flags);
3273 }
3274 } else if (test_bit(In_sync, &rdev->flags))
3275 set_bit(R5_Insync, &dev->flags);
3276 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3277 /* in sync if before recovery_offset */
3278 set_bit(R5_Insync, &dev->flags);
3279 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3280 test_bit(R5_Expanded, &dev->flags))
3281 /* If we've reshaped into here, we assume it is Insync.
3282 * We will shortly update recovery_offset to make
3283 * it official.
3284 */
3285 set_bit(R5_Insync, &dev->flags);
3286
3287 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3288 /* This flag does not apply to '.replacement'
3289 * only to .rdev, so make sure to check that*/
3290 struct md_rdev *rdev2 = rcu_dereference(
3291 conf->disks[i].rdev);
3292 if (rdev2 == rdev)
3293 clear_bit(R5_Insync, &dev->flags);
3294 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3295 s->handle_bad_blocks = 1;
3296 atomic_inc(&rdev2->nr_pending);
3297 } else
3298 clear_bit(R5_WriteError, &dev->flags);
3299 }
3300 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3301 /* This flag does not apply to '.replacement'
3302 * only to .rdev, so make sure to check that*/
3303 struct md_rdev *rdev2 = rcu_dereference(
3304 conf->disks[i].rdev);
3305 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3306 s->handle_bad_blocks = 1;
3307 atomic_inc(&rdev2->nr_pending);
3308 } else
3309 clear_bit(R5_MadeGood, &dev->flags);
3310 }
3311 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3312 struct md_rdev *rdev2 = rcu_dereference(
3313 conf->disks[i].replacement);
3314 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3315 s->handle_bad_blocks = 1;
3316 atomic_inc(&rdev2->nr_pending);
3317 } else
3318 clear_bit(R5_MadeGoodRepl, &dev->flags);
3319 }
3320 if (!test_bit(R5_Insync, &dev->flags)) {
3321 /* The ReadError flag will just be confusing now */
3322 clear_bit(R5_ReadError, &dev->flags);
3323 clear_bit(R5_ReWrite, &dev->flags);
3324 }
3325 if (test_bit(R5_ReadError, &dev->flags))
3326 clear_bit(R5_Insync, &dev->flags);
3327 if (!test_bit(R5_Insync, &dev->flags)) {
3328 if (s->failed < 2)
3329 s->failed_num[s->failed] = i;
3330 s->failed++;
3331 if (rdev && !test_bit(Faulty, &rdev->flags))
3332 do_recovery = 1;
3333 }
3334 }
3335 spin_unlock_irq(&conf->device_lock);
3336 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3337 /* If there is a failed device being replaced,
3338 * we must be recovering.
3339 * else if we are after recovery_cp, we must be syncing
3340 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3341 * else we can only be replacing
3342 * sync and recovery both need to read all devices, and so
3343 * use the same flag.
3344 */
3345 if (do_recovery ||
3346 sh->sector >= conf->mddev->recovery_cp ||
3347 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3348 s->syncing = 1;
3349 else
3350 s->replacing = 1;
3351 }
3352 rcu_read_unlock();
3353}
3354
3355static void handle_stripe(struct stripe_head *sh)
3356{
3357 struct stripe_head_state s;
3358 struct r5conf *conf = sh->raid_conf;
3359 int i;
3360 int prexor;
3361 int disks = sh->disks;
3362 struct r5dev *pdev, *qdev;
3363
3364 clear_bit(STRIPE_HANDLE, &sh->state);
3365 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3366 /* already being handled, ensure it gets handled
3367 * again when current action finishes */
3368 set_bit(STRIPE_HANDLE, &sh->state);
3369 return;
3370 }
3371
3372 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3373 set_bit(STRIPE_SYNCING, &sh->state);
3374 clear_bit(STRIPE_INSYNC, &sh->state);
3375 }
3376 clear_bit(STRIPE_DELAYED, &sh->state);
3377
3378 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3379 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3380 (unsigned long long)sh->sector, sh->state,
3381 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3382 sh->check_state, sh->reconstruct_state);
3383
3384 analyse_stripe(sh, &s);
3385
3386 if (s.handle_bad_blocks) {
3387 set_bit(STRIPE_HANDLE, &sh->state);
3388 goto finish;
3389 }
3390
3391 if (unlikely(s.blocked_rdev)) {
3392 if (s.syncing || s.expanding || s.expanded ||
3393 s.replacing || s.to_write || s.written) {
3394 set_bit(STRIPE_HANDLE, &sh->state);
3395 goto finish;
3396 }
3397 /* There is nothing for the blocked_rdev to block */
3398 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3399 s.blocked_rdev = NULL;
3400 }
3401
3402 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3403 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3404 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3405 }
3406
3407 pr_debug("locked=%d uptodate=%d to_read=%d"
3408 " to_write=%d failed=%d failed_num=%d,%d\n",
3409 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3410 s.failed_num[0], s.failed_num[1]);
3411 /* check if the array has lost more than max_degraded devices and,
3412 * if so, some requests might need to be failed.
3413 */
3414 if (s.failed > conf->max_degraded) {
3415 sh->check_state = 0;
3416 sh->reconstruct_state = 0;
3417 if (s.to_read+s.to_write+s.written)
3418 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3419 if (s.syncing + s.replacing)
3420 handle_failed_sync(conf, sh, &s);
3421 }
3422
3423 /*
3424 * might be able to return some write requests if the parity blocks
3425 * are safe, or on a failed drive
3426 */
3427 pdev = &sh->dev[sh->pd_idx];
3428 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3429 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3430 qdev = &sh->dev[sh->qd_idx];
3431 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3432 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3433 || conf->level < 6;
3434
3435 if (s.written &&
3436 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3437 && !test_bit(R5_LOCKED, &pdev->flags)
3438 && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3439 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3440 && !test_bit(R5_LOCKED, &qdev->flags)
3441 && test_bit(R5_UPTODATE, &qdev->flags)))))
3442 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3443
3444 /* Now we might consider reading some blocks, either to check/generate
3445 * parity, or to satisfy requests
3446 * or to load a block that is being partially written.
3447 */
3448 if (s.to_read || s.non_overwrite
3449 || (conf->level == 6 && s.to_write && s.failed)
3450 || (s.syncing && (s.uptodate + s.compute < disks))
3451 || s.replacing
3452 || s.expanding)
3453 handle_stripe_fill(sh, &s, disks);
3454
3455 /* Now we check to see if any write operations have recently
3456 * completed
3457 */
3458 prexor = 0;
3459 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3460 prexor = 1;
3461 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3462 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3463 sh->reconstruct_state = reconstruct_state_idle;
3464
3465 /* All the 'written' buffers and the parity block are ready to
3466 * be written back to disk
3467 */
3468 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3469 BUG_ON(sh->qd_idx >= 0 &&
3470 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3471 for (i = disks; i--; ) {
3472 struct r5dev *dev = &sh->dev[i];
3473 if (test_bit(R5_LOCKED, &dev->flags) &&
3474 (i == sh->pd_idx || i == sh->qd_idx ||
3475 dev->written)) {
3476 pr_debug("Writing block %d\n", i);
3477 set_bit(R5_Wantwrite, &dev->flags);
3478 if (prexor)
3479 continue;
3480 if (!test_bit(R5_Insync, &dev->flags) ||
3481 ((i == sh->pd_idx || i == sh->qd_idx) &&
3482 s.failed == 0))
3483 set_bit(STRIPE_INSYNC, &sh->state);
3484 }
3485 }
3486 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3487 s.dec_preread_active = 1;
3488 }
3489
3490 /* Now to consider new write requests and what else, if anything
3491 * should be read. We do not handle new writes when:
3492 * 1/ A 'write' operation (copy+xor) is already in flight.
3493 * 2/ A 'check' operation is in flight, as it may clobber the parity
3494 * block.
3495 */
3496 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3497 handle_stripe_dirtying(conf, sh, &s, disks);
3498
3499 /* maybe we need to check and possibly fix the parity for this stripe
3500 * Any reads will already have been scheduled, so we just see if enough
3501 * data is available. The parity check is held off while parity
3502 * dependent operations are in flight.
3503 */
3504 if (sh->check_state ||
3505 (s.syncing && s.locked == 0 &&
3506 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3507 !test_bit(STRIPE_INSYNC, &sh->state))) {
3508 if (conf->level == 6)
3509 handle_parity_checks6(conf, sh, &s, disks);
3510 else
3511 handle_parity_checks5(conf, sh, &s, disks);
3512 }
3513
3514 if (s.replacing && s.locked == 0
3515 && !test_bit(STRIPE_INSYNC, &sh->state)) {
3516 /* Write out to replacement devices where possible */
3517 for (i = 0; i < conf->raid_disks; i++)
3518 if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3519 test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3520 set_bit(R5_WantReplace, &sh->dev[i].flags);
3521 set_bit(R5_LOCKED, &sh->dev[i].flags);
3522 s.locked++;
3523 }
3524 set_bit(STRIPE_INSYNC, &sh->state);
3525 }
3526 if ((s.syncing || s.replacing) && s.locked == 0 &&
3527 test_bit(STRIPE_INSYNC, &sh->state)) {
3528 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3529 clear_bit(STRIPE_SYNCING, &sh->state);
3530 }
3531
3532 /* If the failed drives are just a ReadError, then we might need
3533 * to progress the repair/check process
3534 */
3535 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3536 for (i = 0; i < s.failed; i++) {
3537 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3538 if (test_bit(R5_ReadError, &dev->flags)
3539 && !test_bit(R5_LOCKED, &dev->flags)
3540 && test_bit(R5_UPTODATE, &dev->flags)
3541 ) {
3542 if (!test_bit(R5_ReWrite, &dev->flags)) {
3543 set_bit(R5_Wantwrite, &dev->flags);
3544 set_bit(R5_ReWrite, &dev->flags);
3545 set_bit(R5_LOCKED, &dev->flags);
3546 s.locked++;
3547 } else {
3548 /* let's read it back */
3549 set_bit(R5_Wantread, &dev->flags);
3550 set_bit(R5_LOCKED, &dev->flags);
3551 s.locked++;
3552 }
3553 }
3554 }
3555
3556
3557 /* Finish reconstruct operations initiated by the expansion process */
3558 if (sh->reconstruct_state == reconstruct_state_result) {
3559 struct stripe_head *sh_src
3560 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3561 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3562 /* sh cannot be written until sh_src has been read.
3563 * so arrange for sh to be delayed a little
3564 */
3565 set_bit(STRIPE_DELAYED, &sh->state);
3566 set_bit(STRIPE_HANDLE, &sh->state);
3567 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3568 &sh_src->state))
3569 atomic_inc(&conf->preread_active_stripes);
3570 release_stripe(sh_src);
3571 goto finish;
3572 }
3573 if (sh_src)
3574 release_stripe(sh_src);
3575
3576 sh->reconstruct_state = reconstruct_state_idle;
3577 clear_bit(STRIPE_EXPANDING, &sh->state);
3578 for (i = conf->raid_disks; i--; ) {
3579 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3580 set_bit(R5_LOCKED, &sh->dev[i].flags);
3581 s.locked++;
3582 }
3583 }
3584
3585 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3586 !sh->reconstruct_state) {
3587 /* Need to write out all blocks after computing parity */
3588 sh->disks = conf->raid_disks;
3589 stripe_set_idx(sh->sector, conf, 0, sh);
3590 schedule_reconstruction(sh, &s, 1, 1);
3591 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3592 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3593 atomic_dec(&conf->reshape_stripes);
3594 wake_up(&conf->wait_for_overlap);
3595 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3596 }
3597
3598 if (s.expanding && s.locked == 0 &&
3599 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3600 handle_stripe_expansion(conf, sh);
3601
3602finish:
3603 /* wait for this device to become unblocked */
3604 if (unlikely(s.blocked_rdev)) {
3605 if (conf->mddev->external)
3606 md_wait_for_blocked_rdev(s.blocked_rdev,
3607 conf->mddev);
3608 else
3609 /* Internal metadata will immediately
3610 * be written by raid5d, so we don't
3611 * need to wait here.
3612 */
3613 rdev_dec_pending(s.blocked_rdev,
3614 conf->mddev);
3615 }
3616
3617 if (s.handle_bad_blocks)
3618 for (i = disks; i--; ) {
3619 struct md_rdev *rdev;
3620 struct r5dev *dev = &sh->dev[i];
3621 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3622 /* We own a safe reference to the rdev */
3623 rdev = conf->disks[i].rdev;
3624 if (!rdev_set_badblocks(rdev, sh->sector,
3625 STRIPE_SECTORS, 0))
3626 md_error(conf->mddev, rdev);
3627 rdev_dec_pending(rdev, conf->mddev);
3628 }
3629 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3630 rdev = conf->disks[i].rdev;
3631 rdev_clear_badblocks(rdev, sh->sector,
3632 STRIPE_SECTORS, 0);
3633 rdev_dec_pending(rdev, conf->mddev);
3634 }
3635 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3636 rdev = conf->disks[i].replacement;
3637 if (!rdev)
3638 /* rdev have been moved down */
3639 rdev = conf->disks[i].rdev;
3640 rdev_clear_badblocks(rdev, sh->sector,
3641 STRIPE_SECTORS, 0);
3642 rdev_dec_pending(rdev, conf->mddev);
3643 }
3644 }
3645
3646 if (s.ops_request)
3647 raid_run_ops(sh, s.ops_request);
3648
3649 ops_run_io(sh, &s);
3650
3651 if (s.dec_preread_active) {
3652 /* We delay this until after ops_run_io so that if make_request
3653 * is waiting on a flush, it won't continue until the writes
3654 * have actually been submitted.
3655 */
3656 atomic_dec(&conf->preread_active_stripes);
3657 if (atomic_read(&conf->preread_active_stripes) <
3658 IO_THRESHOLD)
3659 md_wakeup_thread(conf->mddev->thread);
3660 }
3661
3662 return_io(s.return_bi);
3663
3664 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3665}
3666
3667static void raid5_activate_delayed(struct r5conf *conf)
3668{
3669 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3670 while (!list_empty(&conf->delayed_list)) {
3671 struct list_head *l = conf->delayed_list.next;
3672 struct stripe_head *sh;
3673 sh = list_entry(l, struct stripe_head, lru);
3674 list_del_init(l);
3675 clear_bit(STRIPE_DELAYED, &sh->state);
3676 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3677 atomic_inc(&conf->preread_active_stripes);
3678 list_add_tail(&sh->lru, &conf->hold_list);
3679 }
3680 }
3681}
3682
3683static void activate_bit_delay(struct r5conf *conf)
3684{
3685 /* device_lock is held */
3686 struct list_head head;
3687 list_add(&head, &conf->bitmap_list);
3688 list_del_init(&conf->bitmap_list);
3689 while (!list_empty(&head)) {
3690 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3691 list_del_init(&sh->lru);
3692 atomic_inc(&sh->count);
3693 __release_stripe(conf, sh);
3694 }
3695}
3696
3697int md_raid5_congested(struct mddev *mddev, int bits)
3698{
3699 struct r5conf *conf = mddev->private;
3700
3701 /* No difference between reads and writes. Just check
3702 * how busy the stripe_cache is
3703 */
3704
3705 if (conf->inactive_blocked)
3706 return 1;
3707 if (conf->quiesce)
3708 return 1;
3709 if (list_empty_careful(&conf->inactive_list))
3710 return 1;
3711
3712 return 0;
3713}
3714EXPORT_SYMBOL_GPL(md_raid5_congested);
3715
3716static int raid5_congested(void *data, int bits)
3717{
3718 struct mddev *mddev = data;
3719
3720 return mddev_congested(mddev, bits) ||
3721 md_raid5_congested(mddev, bits);
3722}
3723
3724/* We want read requests to align with chunks where possible,
3725 * but write requests don't need to.
3726 */
3727static int raid5_mergeable_bvec(struct request_queue *q,
3728 struct bvec_merge_data *bvm,
3729 struct bio_vec *biovec)
3730{
3731 struct mddev *mddev = q->queuedata;
3732 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3733 int max;
3734 unsigned int chunk_sectors = mddev->chunk_sectors;
3735 unsigned int bio_sectors = bvm->bi_size >> 9;
3736
3737 if ((bvm->bi_rw & 1) == WRITE)
3738 return biovec->bv_len; /* always allow writes to be mergeable */
3739
3740 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3741 chunk_sectors = mddev->new_chunk_sectors;
3742 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3743 if (max < 0) max = 0;
3744 if (max <= biovec->bv_len && bio_sectors == 0)
3745 return biovec->bv_len;
3746 else
3747 return max;
3748}
3749
3750
3751static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3752{
3753 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3754 unsigned int chunk_sectors = mddev->chunk_sectors;
3755 unsigned int bio_sectors = bio->bi_size >> 9;
3756
3757 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3758 chunk_sectors = mddev->new_chunk_sectors;
3759 return chunk_sectors >=
3760 ((sector & (chunk_sectors - 1)) + bio_sectors);
3761}
3762
3763/*
3764 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3765 * later sampled by raid5d.
3766 */
3767static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3768{
3769 unsigned long flags;
3770
3771 spin_lock_irqsave(&conf->device_lock, flags);
3772
3773 bi->bi_next = conf->retry_read_aligned_list;
3774 conf->retry_read_aligned_list = bi;
3775
3776 spin_unlock_irqrestore(&conf->device_lock, flags);
3777 md_wakeup_thread(conf->mddev->thread);
3778}
3779
3780
3781static struct bio *remove_bio_from_retry(struct r5conf *conf)
3782{
3783 struct bio *bi;
3784
3785 bi = conf->retry_read_aligned;
3786 if (bi) {
3787 conf->retry_read_aligned = NULL;
3788 return bi;
3789 }
3790 bi = conf->retry_read_aligned_list;
3791 if(bi) {
3792 conf->retry_read_aligned_list = bi->bi_next;
3793 bi->bi_next = NULL;
3794 /*
3795 * this sets the active strip count to 1 and the processed
3796 * strip count to zero (upper 8 bits)
3797 */
3798 bi->bi_phys_segments = 1; /* biased count of active stripes */
3799 }
3800
3801 return bi;
3802}
3803
3804
3805/*
3806 * The "raid5_align_endio" should check if the read succeeded and if it
3807 * did, call bio_endio on the original bio (having bio_put the new bio
3808 * first).
3809 * If the read failed..
3810 */
3811static void raid5_align_endio(struct bio *bi, int error)
3812{
3813 struct bio* raid_bi = bi->bi_private;
3814 struct mddev *mddev;
3815 struct r5conf *conf;
3816 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3817 struct md_rdev *rdev;
3818
3819 bio_put(bi);
3820
3821 rdev = (void*)raid_bi->bi_next;
3822 raid_bi->bi_next = NULL;
3823 mddev = rdev->mddev;
3824 conf = mddev->private;
3825
3826 rdev_dec_pending(rdev, conf->mddev);
3827
3828 if (!error && uptodate) {
3829 bio_endio(raid_bi, 0);
3830 if (atomic_dec_and_test(&conf->active_aligned_reads))
3831 wake_up(&conf->wait_for_stripe);
3832 return;
3833 }
3834
3835
3836 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3837
3838 add_bio_to_retry(raid_bi, conf);
3839}
3840
3841static int bio_fits_rdev(struct bio *bi)
3842{
3843 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3844
3845 if ((bi->bi_size>>9) > queue_max_sectors(q))
3846 return 0;
3847 blk_recount_segments(q, bi);
3848 if (bi->bi_phys_segments > queue_max_segments(q))
3849 return 0;
3850
3851 if (q->merge_bvec_fn)
3852 /* it's too hard to apply the merge_bvec_fn at this stage,
3853 * just just give up
3854 */
3855 return 0;
3856
3857 return 1;
3858}
3859
3860
3861static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3862{
3863 struct r5conf *conf = mddev->private;
3864 int dd_idx;
3865 struct bio* align_bi;
3866 struct md_rdev *rdev;
3867 sector_t end_sector;
3868
3869 if (!in_chunk_boundary(mddev, raid_bio)) {
3870 pr_debug("chunk_aligned_read : non aligned\n");
3871 return 0;
3872 }
3873 /*
3874 * use bio_clone_mddev to make a copy of the bio
3875 */
3876 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3877 if (!align_bi)
3878 return 0;
3879 /*
3880 * set bi_end_io to a new function, and set bi_private to the
3881 * original bio.
3882 */
3883 align_bi->bi_end_io = raid5_align_endio;
3884 align_bi->bi_private = raid_bio;
3885 /*
3886 * compute position
3887 */
3888 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3889 0,
3890 &dd_idx, NULL);
3891
3892 end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3893 rcu_read_lock();
3894 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3895 if (!rdev || test_bit(Faulty, &rdev->flags) ||
3896 rdev->recovery_offset < end_sector) {
3897 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3898 if (rdev &&
3899 (test_bit(Faulty, &rdev->flags) ||
3900 !(test_bit(In_sync, &rdev->flags) ||
3901 rdev->recovery_offset >= end_sector)))
3902 rdev = NULL;
3903 }
3904 if (rdev) {
3905 sector_t first_bad;
3906 int bad_sectors;
3907
3908 atomic_inc(&rdev->nr_pending);
3909 rcu_read_unlock();
3910 raid_bio->bi_next = (void*)rdev;
3911 align_bi->bi_bdev = rdev->bdev;
3912 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3913
3914 if (!bio_fits_rdev(align_bi) ||
3915 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3916 &first_bad, &bad_sectors)) {
3917 /* too big in some way, or has a known bad block */
3918 bio_put(align_bi);
3919 rdev_dec_pending(rdev, mddev);
3920 return 0;
3921 }
3922
3923 /* No reshape active, so we can trust rdev->data_offset */
3924 align_bi->bi_sector += rdev->data_offset;
3925
3926 spin_lock_irq(&conf->device_lock);
3927 wait_event_lock_irq(conf->wait_for_stripe,
3928 conf->quiesce == 0,
3929 conf->device_lock, /* nothing */);
3930 atomic_inc(&conf->active_aligned_reads);
3931 spin_unlock_irq(&conf->device_lock);
3932
3933 generic_make_request(align_bi);
3934 return 1;
3935 } else {
3936 rcu_read_unlock();
3937 bio_put(align_bi);
3938 return 0;
3939 }
3940}
3941
3942/* __get_priority_stripe - get the next stripe to process
3943 *
3944 * Full stripe writes are allowed to pass preread active stripes up until
3945 * the bypass_threshold is exceeded. In general the bypass_count
3946 * increments when the handle_list is handled before the hold_list; however, it
3947 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3948 * stripe with in flight i/o. The bypass_count will be reset when the
3949 * head of the hold_list has changed, i.e. the head was promoted to the
3950 * handle_list.
3951 */
3952static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3953{
3954 struct stripe_head *sh;
3955
3956 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3957 __func__,
3958 list_empty(&conf->handle_list) ? "empty" : "busy",
3959 list_empty(&conf->hold_list) ? "empty" : "busy",
3960 atomic_read(&conf->pending_full_writes), conf->bypass_count);
3961
3962 if (!list_empty(&conf->handle_list)) {
3963 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3964
3965 if (list_empty(&conf->hold_list))
3966 conf->bypass_count = 0;
3967 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3968 if (conf->hold_list.next == conf->last_hold)
3969 conf->bypass_count++;
3970 else {
3971 conf->last_hold = conf->hold_list.next;
3972 conf->bypass_count -= conf->bypass_threshold;
3973 if (conf->bypass_count < 0)
3974 conf->bypass_count = 0;
3975 }
3976 }
3977 } else if (!list_empty(&conf->hold_list) &&
3978 ((conf->bypass_threshold &&
3979 conf->bypass_count > conf->bypass_threshold) ||
3980 atomic_read(&conf->pending_full_writes) == 0)) {
3981 sh = list_entry(conf->hold_list.next,
3982 typeof(*sh), lru);
3983 conf->bypass_count -= conf->bypass_threshold;
3984 if (conf->bypass_count < 0)
3985 conf->bypass_count = 0;
3986 } else
3987 return NULL;
3988
3989 list_del_init(&sh->lru);
3990 atomic_inc(&sh->count);
3991 BUG_ON(atomic_read(&sh->count) != 1);
3992 return sh;
3993}
3994
3995static void make_request(struct mddev *mddev, struct bio * bi)
3996{
3997 struct r5conf *conf = mddev->private;
3998 int dd_idx;
3999 sector_t new_sector;
4000 sector_t logical_sector, last_sector;
4001 struct stripe_head *sh;
4002 const int rw = bio_data_dir(bi);
4003 int remaining;
4004
4005 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4006 md_flush_request(mddev, bi);
4007 return;
4008 }
4009
4010 md_write_start(mddev, bi);
4011
4012 if (rw == READ &&
4013 mddev->reshape_position == MaxSector &&
4014 chunk_aligned_read(mddev,bi))
4015 return;
4016
4017 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4018 last_sector = bi->bi_sector + (bi->bi_size>>9);
4019 bi->bi_next = NULL;
4020 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4021
4022 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4023 DEFINE_WAIT(w);
4024 int previous;
4025
4026 retry:
4027 previous = 0;
4028 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4029 if (unlikely(conf->reshape_progress != MaxSector)) {
4030 /* spinlock is needed as reshape_progress may be
4031 * 64bit on a 32bit platform, and so it might be
4032 * possible to see a half-updated value
4033 * Of course reshape_progress could change after
4034 * the lock is dropped, so once we get a reference
4035 * to the stripe that we think it is, we will have
4036 * to check again.
4037 */
4038 spin_lock_irq(&conf->device_lock);
4039 if (mddev->reshape_backwards
4040 ? logical_sector < conf->reshape_progress
4041 : logical_sector >= conf->reshape_progress) {
4042 previous = 1;
4043 } else {
4044 if (mddev->reshape_backwards
4045 ? logical_sector < conf->reshape_safe
4046 : logical_sector >= conf->reshape_safe) {
4047 spin_unlock_irq(&conf->device_lock);
4048 schedule();
4049 goto retry;
4050 }
4051 }
4052 spin_unlock_irq(&conf->device_lock);
4053 }
4054
4055 new_sector = raid5_compute_sector(conf, logical_sector,
4056 previous,
4057 &dd_idx, NULL);
4058 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4059 (unsigned long long)new_sector,
4060 (unsigned long long)logical_sector);
4061
4062 sh = get_active_stripe(conf, new_sector, previous,
4063 (bi->bi_rw&RWA_MASK), 0);
4064 if (sh) {
4065 if (unlikely(previous)) {
4066 /* expansion might have moved on while waiting for a
4067 * stripe, so we must do the range check again.
4068 * Expansion could still move past after this
4069 * test, but as we are holding a reference to
4070 * 'sh', we know that if that happens,
4071 * STRIPE_EXPANDING will get set and the expansion
4072 * won't proceed until we finish with the stripe.
4073 */
4074 int must_retry = 0;
4075 spin_lock_irq(&conf->device_lock);
4076 if (mddev->reshape_backwards
4077 ? logical_sector >= conf->reshape_progress
4078 : logical_sector < conf->reshape_progress)
4079 /* mismatch, need to try again */
4080 must_retry = 1;
4081 spin_unlock_irq(&conf->device_lock);
4082 if (must_retry) {
4083 release_stripe(sh);
4084 schedule();
4085 goto retry;
4086 }
4087 }
4088
4089 if (rw == WRITE &&
4090 logical_sector >= mddev->suspend_lo &&
4091 logical_sector < mddev->suspend_hi) {
4092 release_stripe(sh);
4093 /* As the suspend_* range is controlled by
4094 * userspace, we want an interruptible
4095 * wait.
4096 */
4097 flush_signals(current);
4098 prepare_to_wait(&conf->wait_for_overlap,
4099 &w, TASK_INTERRUPTIBLE);
4100 if (logical_sector >= mddev->suspend_lo &&
4101 logical_sector < mddev->suspend_hi)
4102 schedule();
4103 goto retry;
4104 }
4105
4106 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4107 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4108 /* Stripe is busy expanding or
4109 * add failed due to overlap. Flush everything
4110 * and wait a while
4111 */
4112 md_wakeup_thread(mddev->thread);
4113 release_stripe(sh);
4114 schedule();
4115 goto retry;
4116 }
4117 finish_wait(&conf->wait_for_overlap, &w);
4118 set_bit(STRIPE_HANDLE, &sh->state);
4119 clear_bit(STRIPE_DELAYED, &sh->state);
4120 if ((bi->bi_rw & REQ_SYNC) &&
4121 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4122 atomic_inc(&conf->preread_active_stripes);
4123 mddev_check_plugged(mddev);
4124 release_stripe(sh);
4125 } else {
4126 /* cannot get stripe for read-ahead, just give-up */
4127 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4128 finish_wait(&conf->wait_for_overlap, &w);
4129 break;
4130 }
4131 }
4132
4133 spin_lock_irq(&conf->device_lock);
4134 remaining = raid5_dec_bi_phys_segments(bi);
4135 spin_unlock_irq(&conf->device_lock);
4136 if (remaining == 0) {
4137
4138 if ( rw == WRITE )
4139 md_write_end(mddev);
4140
4141 bio_endio(bi, 0);
4142 }
4143}
4144
4145static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4146
4147static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4148{
4149 /* reshaping is quite different to recovery/resync so it is
4150 * handled quite separately ... here.
4151 *
4152 * On each call to sync_request, we gather one chunk worth of
4153 * destination stripes and flag them as expanding.
4154 * Then we find all the source stripes and request reads.
4155 * As the reads complete, handle_stripe will copy the data
4156 * into the destination stripe and release that stripe.
4157 */
4158 struct r5conf *conf = mddev->private;
4159 struct stripe_head *sh;
4160 sector_t first_sector, last_sector;
4161 int raid_disks = conf->previous_raid_disks;
4162 int data_disks = raid_disks - conf->max_degraded;
4163 int new_data_disks = conf->raid_disks - conf->max_degraded;
4164 int i;
4165 int dd_idx;
4166 sector_t writepos, readpos, safepos;
4167 sector_t stripe_addr;
4168 int reshape_sectors;
4169 struct list_head stripes;
4170
4171 if (sector_nr == 0) {
4172 /* If restarting in the middle, skip the initial sectors */
4173 if (mddev->reshape_backwards &&
4174 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4175 sector_nr = raid5_size(mddev, 0, 0)
4176 - conf->reshape_progress;
4177 } else if (!mddev->reshape_backwards &&
4178 conf->reshape_progress > 0)
4179 sector_nr = conf->reshape_progress;
4180 sector_div(sector_nr, new_data_disks);
4181 if (sector_nr) {
4182 mddev->curr_resync_completed = sector_nr;
4183 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4184 *skipped = 1;
4185 return sector_nr;
4186 }
4187 }
4188
4189 /* We need to process a full chunk at a time.
4190 * If old and new chunk sizes differ, we need to process the
4191 * largest of these
4192 */
4193 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4194 reshape_sectors = mddev->new_chunk_sectors;
4195 else
4196 reshape_sectors = mddev->chunk_sectors;
4197
4198 /* We update the metadata at least every 10 seconds, or when
4199 * the data about to be copied would over-write the source of
4200 * the data at the front of the range. i.e. one new_stripe
4201 * along from reshape_progress new_maps to after where
4202 * reshape_safe old_maps to
4203 */
4204 writepos = conf->reshape_progress;
4205 sector_div(writepos, new_data_disks);
4206 readpos = conf->reshape_progress;
4207 sector_div(readpos, data_disks);
4208 safepos = conf->reshape_safe;
4209 sector_div(safepos, data_disks);
4210 if (mddev->reshape_backwards) {
4211 writepos -= min_t(sector_t, reshape_sectors, writepos);
4212 readpos += reshape_sectors;
4213 safepos += reshape_sectors;
4214 } else {
4215 writepos += reshape_sectors;
4216 readpos -= min_t(sector_t, reshape_sectors, readpos);
4217 safepos -= min_t(sector_t, reshape_sectors, safepos);
4218 }
4219
4220 /* Having calculated the 'writepos' possibly use it
4221 * to set 'stripe_addr' which is where we will write to.
4222 */
4223 if (mddev->reshape_backwards) {
4224 BUG_ON(conf->reshape_progress == 0);
4225 stripe_addr = writepos;
4226 BUG_ON((mddev->dev_sectors &
4227 ~((sector_t)reshape_sectors - 1))
4228 - reshape_sectors - stripe_addr
4229 != sector_nr);
4230 } else {
4231 BUG_ON(writepos != sector_nr + reshape_sectors);
4232 stripe_addr = sector_nr;
4233 }
4234
4235 /* 'writepos' is the most advanced device address we might write.
4236 * 'readpos' is the least advanced device address we might read.
4237 * 'safepos' is the least address recorded in the metadata as having
4238 * been reshaped.
4239 * If there is a min_offset_diff, these are adjusted either by
4240 * increasing the safepos/readpos if diff is negative, or
4241 * increasing writepos if diff is positive.
4242 * If 'readpos' is then behind 'writepos', there is no way that we can
4243 * ensure safety in the face of a crash - that must be done by userspace
4244 * making a backup of the data. So in that case there is no particular
4245 * rush to update metadata.
4246 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4247 * update the metadata to advance 'safepos' to match 'readpos' so that
4248 * we can be safe in the event of a crash.
4249 * So we insist on updating metadata if safepos is behind writepos and
4250 * readpos is beyond writepos.
4251 * In any case, update the metadata every 10 seconds.
4252 * Maybe that number should be configurable, but I'm not sure it is
4253 * worth it.... maybe it could be a multiple of safemode_delay???
4254 */
4255 if (conf->min_offset_diff < 0) {
4256 safepos += -conf->min_offset_diff;
4257 readpos += -conf->min_offset_diff;
4258 } else
4259 writepos += conf->min_offset_diff;
4260
4261 if ((mddev->reshape_backwards
4262 ? (safepos > writepos && readpos < writepos)
4263 : (safepos < writepos && readpos > writepos)) ||
4264 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4265 /* Cannot proceed until we've updated the superblock... */
4266 wait_event(conf->wait_for_overlap,
4267 atomic_read(&conf->reshape_stripes)==0);
4268 mddev->reshape_position = conf->reshape_progress;
4269 mddev->curr_resync_completed = sector_nr;
4270 conf->reshape_checkpoint = jiffies;
4271 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4272 md_wakeup_thread(mddev->thread);
4273 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4274 kthread_should_stop());
4275 spin_lock_irq(&conf->device_lock);
4276 conf->reshape_safe = mddev->reshape_position;
4277 spin_unlock_irq(&conf->device_lock);
4278 wake_up(&conf->wait_for_overlap);
4279 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4280 }
4281
4282 INIT_LIST_HEAD(&stripes);
4283 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4284 int j;
4285 int skipped_disk = 0;
4286 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4287 set_bit(STRIPE_EXPANDING, &sh->state);
4288 atomic_inc(&conf->reshape_stripes);
4289 /* If any of this stripe is beyond the end of the old
4290 * array, then we need to zero those blocks
4291 */
4292 for (j=sh->disks; j--;) {
4293 sector_t s;
4294 if (j == sh->pd_idx)
4295 continue;
4296 if (conf->level == 6 &&
4297 j == sh->qd_idx)
4298 continue;
4299 s = compute_blocknr(sh, j, 0);
4300 if (s < raid5_size(mddev, 0, 0)) {
4301 skipped_disk = 1;
4302 continue;
4303 }
4304 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4305 set_bit(R5_Expanded, &sh->dev[j].flags);
4306 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4307 }
4308 if (!skipped_disk) {
4309 set_bit(STRIPE_EXPAND_READY, &sh->state);
4310 set_bit(STRIPE_HANDLE, &sh->state);
4311 }
4312 list_add(&sh->lru, &stripes);
4313 }
4314 spin_lock_irq(&conf->device_lock);
4315 if (mddev->reshape_backwards)
4316 conf->reshape_progress -= reshape_sectors * new_data_disks;
4317 else
4318 conf->reshape_progress += reshape_sectors * new_data_disks;
4319 spin_unlock_irq(&conf->device_lock);
4320 /* Ok, those stripe are ready. We can start scheduling
4321 * reads on the source stripes.
4322 * The source stripes are determined by mapping the first and last
4323 * block on the destination stripes.
4324 */
4325 first_sector =
4326 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4327 1, &dd_idx, NULL);
4328 last_sector =
4329 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4330 * new_data_disks - 1),
4331 1, &dd_idx, NULL);
4332 if (last_sector >= mddev->dev_sectors)
4333 last_sector = mddev->dev_sectors - 1;
4334 while (first_sector <= last_sector) {
4335 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4336 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4337 set_bit(STRIPE_HANDLE, &sh->state);
4338 release_stripe(sh);
4339 first_sector += STRIPE_SECTORS;
4340 }
4341 /* Now that the sources are clearly marked, we can release
4342 * the destination stripes
4343 */
4344 while (!list_empty(&stripes)) {
4345 sh = list_entry(stripes.next, struct stripe_head, lru);
4346 list_del_init(&sh->lru);
4347 release_stripe(sh);
4348 }
4349 /* If this takes us to the resync_max point where we have to pause,
4350 * then we need to write out the superblock.
4351 */
4352 sector_nr += reshape_sectors;
4353 if ((sector_nr - mddev->curr_resync_completed) * 2
4354 >= mddev->resync_max - mddev->curr_resync_completed) {
4355 /* Cannot proceed until we've updated the superblock... */
4356 wait_event(conf->wait_for_overlap,
4357 atomic_read(&conf->reshape_stripes) == 0);
4358 mddev->reshape_position = conf->reshape_progress;
4359 mddev->curr_resync_completed = sector_nr;
4360 conf->reshape_checkpoint = jiffies;
4361 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4362 md_wakeup_thread(mddev->thread);
4363 wait_event(mddev->sb_wait,
4364 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4365 || kthread_should_stop());
4366 spin_lock_irq(&conf->device_lock);
4367 conf->reshape_safe = mddev->reshape_position;
4368 spin_unlock_irq(&conf->device_lock);
4369 wake_up(&conf->wait_for_overlap);
4370 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4371 }
4372 return reshape_sectors;
4373}
4374
4375/* FIXME go_faster isn't used */
4376static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4377{
4378 struct r5conf *conf = mddev->private;
4379 struct stripe_head *sh;
4380 sector_t max_sector = mddev->dev_sectors;
4381 sector_t sync_blocks;
4382 int still_degraded = 0;
4383 int i;
4384
4385 if (sector_nr >= max_sector) {
4386 /* just being told to finish up .. nothing much to do */
4387
4388 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4389 end_reshape(conf);
4390 return 0;
4391 }
4392
4393 if (mddev->curr_resync < max_sector) /* aborted */
4394 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4395 &sync_blocks, 1);
4396 else /* completed sync */
4397 conf->fullsync = 0;
4398 bitmap_close_sync(mddev->bitmap);
4399
4400 return 0;
4401 }
4402
4403 /* Allow raid5_quiesce to complete */
4404 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4405
4406 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4407 return reshape_request(mddev, sector_nr, skipped);
4408
4409 /* No need to check resync_max as we never do more than one
4410 * stripe, and as resync_max will always be on a chunk boundary,
4411 * if the check in md_do_sync didn't fire, there is no chance
4412 * of overstepping resync_max here
4413 */
4414
4415 /* if there is too many failed drives and we are trying
4416 * to resync, then assert that we are finished, because there is
4417 * nothing we can do.
4418 */
4419 if (mddev->degraded >= conf->max_degraded &&
4420 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4421 sector_t rv = mddev->dev_sectors - sector_nr;
4422 *skipped = 1;
4423 return rv;
4424 }
4425 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4426 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4427 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4428 /* we can skip this block, and probably more */
4429 sync_blocks /= STRIPE_SECTORS;
4430 *skipped = 1;
4431 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4432 }
4433
4434 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4435
4436 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4437 if (sh == NULL) {
4438 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4439 /* make sure we don't swamp the stripe cache if someone else
4440 * is trying to get access
4441 */
4442 schedule_timeout_uninterruptible(1);
4443 }
4444 /* Need to check if array will still be degraded after recovery/resync
4445 * We don't need to check the 'failed' flag as when that gets set,
4446 * recovery aborts.
4447 */
4448 for (i = 0; i < conf->raid_disks; i++)
4449 if (conf->disks[i].rdev == NULL)
4450 still_degraded = 1;
4451
4452 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4453
4454 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4455
4456 handle_stripe(sh);
4457 release_stripe(sh);
4458
4459 return STRIPE_SECTORS;
4460}
4461
4462static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4463{
4464 /* We may not be able to submit a whole bio at once as there
4465 * may not be enough stripe_heads available.
4466 * We cannot pre-allocate enough stripe_heads as we may need
4467 * more than exist in the cache (if we allow ever large chunks).
4468 * So we do one stripe head at a time and record in
4469 * ->bi_hw_segments how many have been done.
4470 *
4471 * We *know* that this entire raid_bio is in one chunk, so
4472 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4473 */
4474 struct stripe_head *sh;
4475 int dd_idx;
4476 sector_t sector, logical_sector, last_sector;
4477 int scnt = 0;
4478 int remaining;
4479 int handled = 0;
4480
4481 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4482 sector = raid5_compute_sector(conf, logical_sector,
4483 0, &dd_idx, NULL);
4484 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4485
4486 for (; logical_sector < last_sector;
4487 logical_sector += STRIPE_SECTORS,
4488 sector += STRIPE_SECTORS,
4489 scnt++) {
4490
4491 if (scnt < raid5_bi_hw_segments(raid_bio))
4492 /* already done this stripe */
4493 continue;
4494
4495 sh = get_active_stripe(conf, sector, 0, 1, 0);
4496
4497 if (!sh) {
4498 /* failed to get a stripe - must wait */
4499 raid5_set_bi_hw_segments(raid_bio, scnt);
4500 conf->retry_read_aligned = raid_bio;
4501 return handled;
4502 }
4503
4504 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4505 release_stripe(sh);
4506 raid5_set_bi_hw_segments(raid_bio, scnt);
4507 conf->retry_read_aligned = raid_bio;
4508 return handled;
4509 }
4510
4511 handle_stripe(sh);
4512 release_stripe(sh);
4513 handled++;
4514 }
4515 spin_lock_irq(&conf->device_lock);
4516 remaining = raid5_dec_bi_phys_segments(raid_bio);
4517 spin_unlock_irq(&conf->device_lock);
4518 if (remaining == 0)
4519 bio_endio(raid_bio, 0);
4520 if (atomic_dec_and_test(&conf->active_aligned_reads))
4521 wake_up(&conf->wait_for_stripe);
4522 return handled;
4523}
4524
4525
4526/*
4527 * This is our raid5 kernel thread.
4528 *
4529 * We scan the hash table for stripes which can be handled now.
4530 * During the scan, completed stripes are saved for us by the interrupt
4531 * handler, so that they will not have to wait for our next wakeup.
4532 */
4533static void raid5d(struct mddev *mddev)
4534{
4535 struct stripe_head *sh;
4536 struct r5conf *conf = mddev->private;
4537 int handled;
4538 struct blk_plug plug;
4539
4540 pr_debug("+++ raid5d active\n");
4541
4542 md_check_recovery(mddev);
4543
4544 blk_start_plug(&plug);
4545 handled = 0;
4546 spin_lock_irq(&conf->device_lock);
4547 while (1) {
4548 struct bio *bio;
4549
4550 if (atomic_read(&mddev->plug_cnt) == 0 &&
4551 !list_empty(&conf->bitmap_list)) {
4552 /* Now is a good time to flush some bitmap updates */
4553 conf->seq_flush++;
4554 spin_unlock_irq(&conf->device_lock);
4555 bitmap_unplug(mddev->bitmap);
4556 spin_lock_irq(&conf->device_lock);
4557 conf->seq_write = conf->seq_flush;
4558 activate_bit_delay(conf);
4559 }
4560 if (atomic_read(&mddev->plug_cnt) == 0)
4561 raid5_activate_delayed(conf);
4562
4563 while ((bio = remove_bio_from_retry(conf))) {
4564 int ok;
4565 spin_unlock_irq(&conf->device_lock);
4566 ok = retry_aligned_read(conf, bio);
4567 spin_lock_irq(&conf->device_lock);
4568 if (!ok)
4569 break;
4570 handled++;
4571 }
4572
4573 sh = __get_priority_stripe(conf);
4574
4575 if (!sh)
4576 break;
4577 spin_unlock_irq(&conf->device_lock);
4578
4579 handled++;
4580 handle_stripe(sh);
4581 release_stripe(sh);
4582 cond_resched();
4583
4584 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4585 md_check_recovery(mddev);
4586
4587 spin_lock_irq(&conf->device_lock);
4588 }
4589 pr_debug("%d stripes handled\n", handled);
4590
4591 spin_unlock_irq(&conf->device_lock);
4592
4593 async_tx_issue_pending_all();
4594 blk_finish_plug(&plug);
4595
4596 pr_debug("--- raid5d inactive\n");
4597}
4598
4599static ssize_t
4600raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4601{
4602 struct r5conf *conf = mddev->private;
4603 if (conf)
4604 return sprintf(page, "%d\n", conf->max_nr_stripes);
4605 else
4606 return 0;
4607}
4608
4609int
4610raid5_set_cache_size(struct mddev *mddev, int size)
4611{
4612 struct r5conf *conf = mddev->private;
4613 int err;
4614
4615 if (size <= 16 || size > 32768)
4616 return -EINVAL;
4617 while (size < conf->max_nr_stripes) {
4618 if (drop_one_stripe(conf))
4619 conf->max_nr_stripes--;
4620 else
4621 break;
4622 }
4623 err = md_allow_write(mddev);
4624 if (err)
4625 return err;
4626 while (size > conf->max_nr_stripes) {
4627 if (grow_one_stripe(conf))
4628 conf->max_nr_stripes++;
4629 else break;
4630 }
4631 return 0;
4632}
4633EXPORT_SYMBOL(raid5_set_cache_size);
4634
4635static ssize_t
4636raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4637{
4638 struct r5conf *conf = mddev->private;
4639 unsigned long new;
4640 int err;
4641
4642 if (len >= PAGE_SIZE)
4643 return -EINVAL;
4644 if (!conf)
4645 return -ENODEV;
4646
4647 if (strict_strtoul(page, 10, &new))
4648 return -EINVAL;
4649 err = raid5_set_cache_size(mddev, new);
4650 if (err)
4651 return err;
4652 return len;
4653}
4654
4655static struct md_sysfs_entry
4656raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4657 raid5_show_stripe_cache_size,
4658 raid5_store_stripe_cache_size);
4659
4660static ssize_t
4661raid5_show_preread_threshold(struct mddev *mddev, char *page)
4662{
4663 struct r5conf *conf = mddev->private;
4664 if (conf)
4665 return sprintf(page, "%d\n", conf->bypass_threshold);
4666 else
4667 return 0;
4668}
4669
4670static ssize_t
4671raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4672{
4673 struct r5conf *conf = mddev->private;
4674 unsigned long new;
4675 if (len >= PAGE_SIZE)
4676 return -EINVAL;
4677 if (!conf)
4678 return -ENODEV;
4679
4680 if (strict_strtoul(page, 10, &new))
4681 return -EINVAL;
4682 if (new > conf->max_nr_stripes)
4683 return -EINVAL;
4684 conf->bypass_threshold = new;
4685 return len;
4686}
4687
4688static struct md_sysfs_entry
4689raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4690 S_IRUGO | S_IWUSR,
4691 raid5_show_preread_threshold,
4692 raid5_store_preread_threshold);
4693
4694static ssize_t
4695stripe_cache_active_show(struct mddev *mddev, char *page)
4696{
4697 struct r5conf *conf = mddev->private;
4698 if (conf)
4699 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4700 else
4701 return 0;
4702}
4703
4704static struct md_sysfs_entry
4705raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4706
4707static struct attribute *raid5_attrs[] = {
4708 &raid5_stripecache_size.attr,
4709 &raid5_stripecache_active.attr,
4710 &raid5_preread_bypass_threshold.attr,
4711 NULL,
4712};
4713static struct attribute_group raid5_attrs_group = {
4714 .name = NULL,
4715 .attrs = raid5_attrs,
4716};
4717
4718static sector_t
4719raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4720{
4721 struct r5conf *conf = mddev->private;
4722
4723 if (!sectors)
4724 sectors = mddev->dev_sectors;
4725 if (!raid_disks)
4726 /* size is defined by the smallest of previous and new size */
4727 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4728
4729 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4730 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4731 return sectors * (raid_disks - conf->max_degraded);
4732}
4733
4734static void raid5_free_percpu(struct r5conf *conf)
4735{
4736 struct raid5_percpu *percpu;
4737 unsigned long cpu;
4738
4739 if (!conf->percpu)
4740 return;
4741
4742 get_online_cpus();
4743 for_each_possible_cpu(cpu) {
4744 percpu = per_cpu_ptr(conf->percpu, cpu);
4745 safe_put_page(percpu->spare_page);
4746 kfree(percpu->scribble);
4747 }
4748#ifdef CONFIG_HOTPLUG_CPU
4749 unregister_cpu_notifier(&conf->cpu_notify);
4750#endif
4751 put_online_cpus();
4752
4753 free_percpu(conf->percpu);
4754}
4755
4756static void free_conf(struct r5conf *conf)
4757{
4758 shrink_stripes(conf);
4759 raid5_free_percpu(conf);
4760 kfree(conf->disks);
4761 kfree(conf->stripe_hashtbl);
4762 kfree(conf);
4763}
4764
4765#ifdef CONFIG_HOTPLUG_CPU
4766static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4767 void *hcpu)
4768{
4769 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4770 long cpu = (long)hcpu;
4771 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4772
4773 switch (action) {
4774 case CPU_UP_PREPARE:
4775 case CPU_UP_PREPARE_FROZEN:
4776 if (conf->level == 6 && !percpu->spare_page)
4777 percpu->spare_page = alloc_page(GFP_KERNEL);
4778 if (!percpu->scribble)
4779 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4780
4781 if (!percpu->scribble ||
4782 (conf->level == 6 && !percpu->spare_page)) {
4783 safe_put_page(percpu->spare_page);
4784 kfree(percpu->scribble);
4785 pr_err("%s: failed memory allocation for cpu%ld\n",
4786 __func__, cpu);
4787 return notifier_from_errno(-ENOMEM);
4788 }
4789 break;
4790 case CPU_DEAD:
4791 case CPU_DEAD_FROZEN:
4792 safe_put_page(percpu->spare_page);
4793 kfree(percpu->scribble);
4794 percpu->spare_page = NULL;
4795 percpu->scribble = NULL;
4796 break;
4797 default:
4798 break;
4799 }
4800 return NOTIFY_OK;
4801}
4802#endif
4803
4804static int raid5_alloc_percpu(struct r5conf *conf)
4805{
4806 unsigned long cpu;
4807 struct page *spare_page;
4808 struct raid5_percpu __percpu *allcpus;
4809 void *scribble;
4810 int err;
4811
4812 allcpus = alloc_percpu(struct raid5_percpu);
4813 if (!allcpus)
4814 return -ENOMEM;
4815 conf->percpu = allcpus;
4816
4817 get_online_cpus();
4818 err = 0;
4819 for_each_present_cpu(cpu) {
4820 if (conf->level == 6) {
4821 spare_page = alloc_page(GFP_KERNEL);
4822 if (!spare_page) {
4823 err = -ENOMEM;
4824 break;
4825 }
4826 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4827 }
4828 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4829 if (!scribble) {
4830 err = -ENOMEM;
4831 break;
4832 }
4833 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4834 }
4835#ifdef CONFIG_HOTPLUG_CPU
4836 conf->cpu_notify.notifier_call = raid456_cpu_notify;
4837 conf->cpu_notify.priority = 0;
4838 if (err == 0)
4839 err = register_cpu_notifier(&conf->cpu_notify);
4840#endif
4841 put_online_cpus();
4842
4843 return err;
4844}
4845
4846static struct r5conf *setup_conf(struct mddev *mddev)
4847{
4848 struct r5conf *conf;
4849 int raid_disk, memory, max_disks;
4850 struct md_rdev *rdev;
4851 struct disk_info *disk;
4852 char pers_name[6];
4853
4854 if (mddev->new_level != 5
4855 && mddev->new_level != 4
4856 && mddev->new_level != 6) {
4857 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4858 mdname(mddev), mddev->new_level);
4859 return ERR_PTR(-EIO);
4860 }
4861 if ((mddev->new_level == 5
4862 && !algorithm_valid_raid5(mddev->new_layout)) ||
4863 (mddev->new_level == 6
4864 && !algorithm_valid_raid6(mddev->new_layout))) {
4865 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4866 mdname(mddev), mddev->new_layout);
4867 return ERR_PTR(-EIO);
4868 }
4869 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4870 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4871 mdname(mddev), mddev->raid_disks);
4872 return ERR_PTR(-EINVAL);
4873 }
4874
4875 if (!mddev->new_chunk_sectors ||
4876 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4877 !is_power_of_2(mddev->new_chunk_sectors)) {
4878 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4879 mdname(mddev), mddev->new_chunk_sectors << 9);
4880 return ERR_PTR(-EINVAL);
4881 }
4882
4883 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4884 if (conf == NULL)
4885 goto abort;
4886 spin_lock_init(&conf->device_lock);
4887 init_waitqueue_head(&conf->wait_for_stripe);
4888 init_waitqueue_head(&conf->wait_for_overlap);
4889 INIT_LIST_HEAD(&conf->handle_list);
4890 INIT_LIST_HEAD(&conf->hold_list);
4891 INIT_LIST_HEAD(&conf->delayed_list);
4892 INIT_LIST_HEAD(&conf->bitmap_list);
4893 INIT_LIST_HEAD(&conf->inactive_list);
4894 atomic_set(&conf->active_stripes, 0);
4895 atomic_set(&conf->preread_active_stripes, 0);
4896 atomic_set(&conf->active_aligned_reads, 0);
4897 conf->bypass_threshold = BYPASS_THRESHOLD;
4898 conf->recovery_disabled = mddev->recovery_disabled - 1;
4899
4900 conf->raid_disks = mddev->raid_disks;
4901 if (mddev->reshape_position == MaxSector)
4902 conf->previous_raid_disks = mddev->raid_disks;
4903 else
4904 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4905 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4906 conf->scribble_len = scribble_len(max_disks);
4907
4908 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4909 GFP_KERNEL);
4910 if (!conf->disks)
4911 goto abort;
4912
4913 conf->mddev = mddev;
4914
4915 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4916 goto abort;
4917
4918 conf->level = mddev->new_level;
4919 if (raid5_alloc_percpu(conf) != 0)
4920 goto abort;
4921
4922 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4923
4924 rdev_for_each(rdev, mddev) {
4925 raid_disk = rdev->raid_disk;
4926 if (raid_disk >= max_disks
4927 || raid_disk < 0)
4928 continue;
4929 disk = conf->disks + raid_disk;
4930
4931 if (test_bit(Replacement, &rdev->flags)) {
4932 if (disk->replacement)
4933 goto abort;
4934 disk->replacement = rdev;
4935 } else {
4936 if (disk->rdev)
4937 goto abort;
4938 disk->rdev = rdev;
4939 }
4940
4941 if (test_bit(In_sync, &rdev->flags)) {
4942 char b[BDEVNAME_SIZE];
4943 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4944 " disk %d\n",
4945 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4946 } else if (rdev->saved_raid_disk != raid_disk)
4947 /* Cannot rely on bitmap to complete recovery */
4948 conf->fullsync = 1;
4949 }
4950
4951 conf->chunk_sectors = mddev->new_chunk_sectors;
4952 conf->level = mddev->new_level;
4953 if (conf->level == 6)
4954 conf->max_degraded = 2;
4955 else
4956 conf->max_degraded = 1;
4957 conf->algorithm = mddev->new_layout;
4958 conf->max_nr_stripes = NR_STRIPES;
4959 conf->reshape_progress = mddev->reshape_position;
4960 if (conf->reshape_progress != MaxSector) {
4961 conf->prev_chunk_sectors = mddev->chunk_sectors;
4962 conf->prev_algo = mddev->layout;
4963 }
4964
4965 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4966 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4967 if (grow_stripes(conf, conf->max_nr_stripes)) {
4968 printk(KERN_ERR
4969 "md/raid:%s: couldn't allocate %dkB for buffers\n",
4970 mdname(mddev), memory);
4971 goto abort;
4972 } else
4973 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4974 mdname(mddev), memory);
4975
4976 sprintf(pers_name, "raid%d", mddev->new_level);
4977 conf->thread = md_register_thread(raid5d, mddev, pers_name);
4978 if (!conf->thread) {
4979 printk(KERN_ERR
4980 "md/raid:%s: couldn't allocate thread.\n",
4981 mdname(mddev));
4982 goto abort;
4983 }
4984
4985 return conf;
4986
4987 abort:
4988 if (conf) {
4989 free_conf(conf);
4990 return ERR_PTR(-EIO);
4991 } else
4992 return ERR_PTR(-ENOMEM);
4993}
4994
4995
4996static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4997{
4998 switch (algo) {
4999 case ALGORITHM_PARITY_0:
5000 if (raid_disk < max_degraded)
5001 return 1;
5002 break;
5003 case ALGORITHM_PARITY_N:
5004 if (raid_disk >= raid_disks - max_degraded)
5005 return 1;
5006 break;
5007 case ALGORITHM_PARITY_0_6:
5008 if (raid_disk == 0 ||
5009 raid_disk == raid_disks - 1)
5010 return 1;
5011 break;
5012 case ALGORITHM_LEFT_ASYMMETRIC_6:
5013 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5014 case ALGORITHM_LEFT_SYMMETRIC_6:
5015 case ALGORITHM_RIGHT_SYMMETRIC_6:
5016 if (raid_disk == raid_disks - 1)
5017 return 1;
5018 }
5019 return 0;
5020}
5021
5022static int run(struct mddev *mddev)
5023{
5024 struct r5conf *conf;
5025 int working_disks = 0;
5026 int dirty_parity_disks = 0;
5027 struct md_rdev *rdev;
5028 sector_t reshape_offset = 0;
5029 int i;
5030 long long min_offset_diff = 0;
5031 int first = 1;
5032
5033 if (mddev->recovery_cp != MaxSector)
5034 printk(KERN_NOTICE "md/raid:%s: not clean"
5035 " -- starting background reconstruction\n",
5036 mdname(mddev));
5037
5038 rdev_for_each(rdev, mddev) {
5039 long long diff;
5040 if (rdev->raid_disk < 0)
5041 continue;
5042 diff = (rdev->new_data_offset - rdev->data_offset);
5043 if (first) {
5044 min_offset_diff = diff;
5045 first = 0;
5046 } else if (mddev->reshape_backwards &&
5047 diff < min_offset_diff)
5048 min_offset_diff = diff;
5049 else if (!mddev->reshape_backwards &&
5050 diff > min_offset_diff)
5051 min_offset_diff = diff;
5052 }
5053
5054 if (mddev->reshape_position != MaxSector) {
5055 /* Check that we can continue the reshape.
5056 * Difficulties arise if the stripe we would write to
5057 * next is at or after the stripe we would read from next.
5058 * For a reshape that changes the number of devices, this
5059 * is only possible for a very short time, and mdadm makes
5060 * sure that time appears to have past before assembling
5061 * the array. So we fail if that time hasn't passed.
5062 * For a reshape that keeps the number of devices the same
5063 * mdadm must be monitoring the reshape can keeping the
5064 * critical areas read-only and backed up. It will start
5065 * the array in read-only mode, so we check for that.
5066 */
5067 sector_t here_new, here_old;
5068 int old_disks;
5069 int max_degraded = (mddev->level == 6 ? 2 : 1);
5070
5071 if (mddev->new_level != mddev->level) {
5072 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5073 "required - aborting.\n",
5074 mdname(mddev));
5075 return -EINVAL;
5076 }
5077 old_disks = mddev->raid_disks - mddev->delta_disks;
5078 /* reshape_position must be on a new-stripe boundary, and one
5079 * further up in new geometry must map after here in old
5080 * geometry.
5081 */
5082 here_new = mddev->reshape_position;
5083 if (sector_div(here_new, mddev->new_chunk_sectors *
5084 (mddev->raid_disks - max_degraded))) {
5085 printk(KERN_ERR "md/raid:%s: reshape_position not "
5086 "on a stripe boundary\n", mdname(mddev));
5087 return -EINVAL;
5088 }
5089 reshape_offset = here_new * mddev->new_chunk_sectors;
5090 /* here_new is the stripe we will write to */
5091 here_old = mddev->reshape_position;
5092 sector_div(here_old, mddev->chunk_sectors *
5093 (old_disks-max_degraded));
5094 /* here_old is the first stripe that we might need to read
5095 * from */
5096 if (mddev->delta_disks == 0) {
5097 if ((here_new * mddev->new_chunk_sectors !=
5098 here_old * mddev->chunk_sectors)) {
5099 printk(KERN_ERR "md/raid:%s: reshape position is"
5100 " confused - aborting\n", mdname(mddev));
5101 return -EINVAL;
5102 }
5103 /* We cannot be sure it is safe to start an in-place
5104 * reshape. It is only safe if user-space is monitoring
5105 * and taking constant backups.
5106 * mdadm always starts a situation like this in
5107 * readonly mode so it can take control before
5108 * allowing any writes. So just check for that.
5109 */
5110 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5111 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5112 /* not really in-place - so OK */;
5113 else if (mddev->ro == 0) {
5114 printk(KERN_ERR "md/raid:%s: in-place reshape "
5115 "must be started in read-only mode "
5116 "- aborting\n",
5117 mdname(mddev));
5118 return -EINVAL;
5119 }
5120 } else if (mddev->reshape_backwards
5121 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5122 here_old * mddev->chunk_sectors)
5123 : (here_new * mddev->new_chunk_sectors >=
5124 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5125 /* Reading from the same stripe as writing to - bad */
5126 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5127 "auto-recovery - aborting.\n",
5128 mdname(mddev));
5129 return -EINVAL;
5130 }
5131 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5132 mdname(mddev));
5133 /* OK, we should be able to continue; */
5134 } else {
5135 BUG_ON(mddev->level != mddev->new_level);
5136 BUG_ON(mddev->layout != mddev->new_layout);
5137 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5138 BUG_ON(mddev->delta_disks != 0);
5139 }
5140
5141 if (mddev->private == NULL)
5142 conf = setup_conf(mddev);
5143 else
5144 conf = mddev->private;
5145
5146 if (IS_ERR(conf))
5147 return PTR_ERR(conf);
5148
5149 conf->min_offset_diff = min_offset_diff;
5150 mddev->thread = conf->thread;
5151 conf->thread = NULL;
5152 mddev->private = conf;
5153
5154 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5155 i++) {
5156 rdev = conf->disks[i].rdev;
5157 if (!rdev && conf->disks[i].replacement) {
5158 /* The replacement is all we have yet */
5159 rdev = conf->disks[i].replacement;
5160 conf->disks[i].replacement = NULL;
5161 clear_bit(Replacement, &rdev->flags);
5162 conf->disks[i].rdev = rdev;
5163 }
5164 if (!rdev)
5165 continue;
5166 if (conf->disks[i].replacement &&
5167 conf->reshape_progress != MaxSector) {
5168 /* replacements and reshape simply do not mix. */
5169 printk(KERN_ERR "md: cannot handle concurrent "
5170 "replacement and reshape.\n");
5171 goto abort;
5172 }
5173 if (test_bit(In_sync, &rdev->flags)) {
5174 working_disks++;
5175 continue;
5176 }
5177 /* This disc is not fully in-sync. However if it
5178 * just stored parity (beyond the recovery_offset),
5179 * when we don't need to be concerned about the
5180 * array being dirty.
5181 * When reshape goes 'backwards', we never have
5182 * partially completed devices, so we only need
5183 * to worry about reshape going forwards.
5184 */
5185 /* Hack because v0.91 doesn't store recovery_offset properly. */
5186 if (mddev->major_version == 0 &&
5187 mddev->minor_version > 90)
5188 rdev->recovery_offset = reshape_offset;
5189
5190 if (rdev->recovery_offset < reshape_offset) {
5191 /* We need to check old and new layout */
5192 if (!only_parity(rdev->raid_disk,
5193 conf->algorithm,
5194 conf->raid_disks,
5195 conf->max_degraded))
5196 continue;
5197 }
5198 if (!only_parity(rdev->raid_disk,
5199 conf->prev_algo,
5200 conf->previous_raid_disks,
5201 conf->max_degraded))
5202 continue;
5203 dirty_parity_disks++;
5204 }
5205
5206 /*
5207 * 0 for a fully functional array, 1 or 2 for a degraded array.
5208 */
5209 mddev->degraded = calc_degraded(conf);
5210
5211 if (has_failed(conf)) {
5212 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5213 " (%d/%d failed)\n",
5214 mdname(mddev), mddev->degraded, conf->raid_disks);
5215 goto abort;
5216 }
5217
5218 /* device size must be a multiple of chunk size */
5219 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5220 mddev->resync_max_sectors = mddev->dev_sectors;
5221
5222 if (mddev->degraded > dirty_parity_disks &&
5223 mddev->recovery_cp != MaxSector) {
5224 if (mddev->ok_start_degraded)
5225 printk(KERN_WARNING
5226 "md/raid:%s: starting dirty degraded array"
5227 " - data corruption possible.\n",
5228 mdname(mddev));
5229 else {
5230 printk(KERN_ERR
5231 "md/raid:%s: cannot start dirty degraded array.\n",
5232 mdname(mddev));
5233 goto abort;
5234 }
5235 }
5236
5237 if (mddev->degraded == 0)
5238 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5239 " devices, algorithm %d\n", mdname(mddev), conf->level,
5240 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5241 mddev->new_layout);
5242 else
5243 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5244 " out of %d devices, algorithm %d\n",
5245 mdname(mddev), conf->level,
5246 mddev->raid_disks - mddev->degraded,
5247 mddev->raid_disks, mddev->new_layout);
5248
5249 print_raid5_conf(conf);
5250
5251 if (conf->reshape_progress != MaxSector) {
5252 conf->reshape_safe = conf->reshape_progress;
5253 atomic_set(&conf->reshape_stripes, 0);
5254 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5255 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5256 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5257 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5258 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5259 "reshape");
5260 }
5261
5262
5263 /* Ok, everything is just fine now */
5264 if (mddev->to_remove == &raid5_attrs_group)
5265 mddev->to_remove = NULL;
5266 else if (mddev->kobj.sd &&
5267 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5268 printk(KERN_WARNING
5269 "raid5: failed to create sysfs attributes for %s\n",
5270 mdname(mddev));
5271 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5272
5273 if (mddev->queue) {
5274 int chunk_size;
5275 /* read-ahead size must cover two whole stripes, which
5276 * is 2 * (datadisks) * chunksize where 'n' is the
5277 * number of raid devices
5278 */
5279 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5280 int stripe = data_disks *
5281 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5282 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5283 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5284
5285 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5286
5287 mddev->queue->backing_dev_info.congested_data = mddev;
5288 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5289
5290 chunk_size = mddev->chunk_sectors << 9;
5291 blk_queue_io_min(mddev->queue, chunk_size);
5292 blk_queue_io_opt(mddev->queue, chunk_size *
5293 (conf->raid_disks - conf->max_degraded));
5294
5295 rdev_for_each(rdev, mddev) {
5296 disk_stack_limits(mddev->gendisk, rdev->bdev,
5297 rdev->data_offset << 9);
5298 disk_stack_limits(mddev->gendisk, rdev->bdev,
5299 rdev->new_data_offset << 9);
5300 }
5301 }
5302
5303 return 0;
5304abort:
5305 md_unregister_thread(&mddev->thread);
5306 print_raid5_conf(conf);
5307 free_conf(conf);
5308 mddev->private = NULL;
5309 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5310 return -EIO;
5311}
5312
5313static int stop(struct mddev *mddev)
5314{
5315 struct r5conf *conf = mddev->private;
5316
5317 md_unregister_thread(&mddev->thread);
5318 if (mddev->queue)
5319 mddev->queue->backing_dev_info.congested_fn = NULL;
5320 free_conf(conf);
5321 mddev->private = NULL;
5322 mddev->to_remove = &raid5_attrs_group;
5323 return 0;
5324}
5325
5326static void status(struct seq_file *seq, struct mddev *mddev)
5327{
5328 struct r5conf *conf = mddev->private;
5329 int i;
5330
5331 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5332 mddev->chunk_sectors / 2, mddev->layout);
5333 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5334 for (i = 0; i < conf->raid_disks; i++)
5335 seq_printf (seq, "%s",
5336 conf->disks[i].rdev &&
5337 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5338 seq_printf (seq, "]");
5339}
5340
5341static void print_raid5_conf (struct r5conf *conf)
5342{
5343 int i;
5344 struct disk_info *tmp;
5345
5346 printk(KERN_DEBUG "RAID conf printout:\n");
5347 if (!conf) {
5348 printk("(conf==NULL)\n");
5349 return;
5350 }
5351 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5352 conf->raid_disks,
5353 conf->raid_disks - conf->mddev->degraded);
5354
5355 for (i = 0; i < conf->raid_disks; i++) {
5356 char b[BDEVNAME_SIZE];
5357 tmp = conf->disks + i;
5358 if (tmp->rdev)
5359 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5360 i, !test_bit(Faulty, &tmp->rdev->flags),
5361 bdevname(tmp->rdev->bdev, b));
5362 }
5363}
5364
5365static int raid5_spare_active(struct mddev *mddev)
5366{
5367 int i;
5368 struct r5conf *conf = mddev->private;
5369 struct disk_info *tmp;
5370 int count = 0;
5371 unsigned long flags;
5372
5373 for (i = 0; i < conf->raid_disks; i++) {
5374 tmp = conf->disks + i;
5375 if (tmp->replacement
5376 && tmp->replacement->recovery_offset == MaxSector
5377 && !test_bit(Faulty, &tmp->replacement->flags)
5378 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5379 /* Replacement has just become active. */
5380 if (!tmp->rdev
5381 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5382 count++;
5383 if (tmp->rdev) {
5384 /* Replaced device not technically faulty,
5385 * but we need to be sure it gets removed
5386 * and never re-added.
5387 */
5388 set_bit(Faulty, &tmp->rdev->flags);
5389 sysfs_notify_dirent_safe(
5390 tmp->rdev->sysfs_state);
5391 }
5392 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5393 } else if (tmp->rdev
5394 && tmp->rdev->recovery_offset == MaxSector
5395 && !test_bit(Faulty, &tmp->rdev->flags)
5396 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5397 count++;
5398 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5399 }
5400 }
5401 spin_lock_irqsave(&conf->device_lock, flags);
5402 mddev->degraded = calc_degraded(conf);
5403 spin_unlock_irqrestore(&conf->device_lock, flags);
5404 print_raid5_conf(conf);
5405 return count;
5406}
5407
5408static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5409{
5410 struct r5conf *conf = mddev->private;
5411 int err = 0;
5412 int number = rdev->raid_disk;
5413 struct md_rdev **rdevp;
5414 struct disk_info *p = conf->disks + number;
5415
5416 print_raid5_conf(conf);
5417 if (rdev == p->rdev)
5418 rdevp = &p->rdev;
5419 else if (rdev == p->replacement)
5420 rdevp = &p->replacement;
5421 else
5422 return 0;
5423
5424 if (number >= conf->raid_disks &&
5425 conf->reshape_progress == MaxSector)
5426 clear_bit(In_sync, &rdev->flags);
5427
5428 if (test_bit(In_sync, &rdev->flags) ||
5429 atomic_read(&rdev->nr_pending)) {
5430 err = -EBUSY;
5431 goto abort;
5432 }
5433 /* Only remove non-faulty devices if recovery
5434 * isn't possible.
5435 */
5436 if (!test_bit(Faulty, &rdev->flags) &&
5437 mddev->recovery_disabled != conf->recovery_disabled &&
5438 !has_failed(conf) &&
5439 (!p->replacement || p->replacement == rdev) &&
5440 number < conf->raid_disks) {
5441 err = -EBUSY;
5442 goto abort;
5443 }
5444 *rdevp = NULL;
5445 synchronize_rcu();
5446 if (atomic_read(&rdev->nr_pending)) {
5447 /* lost the race, try later */
5448 err = -EBUSY;
5449 *rdevp = rdev;
5450 } else if (p->replacement) {
5451 /* We must have just cleared 'rdev' */
5452 p->rdev = p->replacement;
5453 clear_bit(Replacement, &p->replacement->flags);
5454 smp_mb(); /* Make sure other CPUs may see both as identical
5455 * but will never see neither - if they are careful
5456 */
5457 p->replacement = NULL;
5458 clear_bit(WantReplacement, &rdev->flags);
5459 } else
5460 /* We might have just removed the Replacement as faulty-
5461 * clear the bit just in case
5462 */
5463 clear_bit(WantReplacement, &rdev->flags);
5464abort:
5465
5466 print_raid5_conf(conf);
5467 return err;
5468}
5469
5470static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5471{
5472 struct r5conf *conf = mddev->private;
5473 int err = -EEXIST;
5474 int disk;
5475 struct disk_info *p;
5476 int first = 0;
5477 int last = conf->raid_disks - 1;
5478
5479 if (mddev->recovery_disabled == conf->recovery_disabled)
5480 return -EBUSY;
5481
5482 if (rdev->saved_raid_disk < 0 && has_failed(conf))
5483 /* no point adding a device */
5484 return -EINVAL;
5485
5486 if (rdev->raid_disk >= 0)
5487 first = last = rdev->raid_disk;
5488
5489 /*
5490 * find the disk ... but prefer rdev->saved_raid_disk
5491 * if possible.
5492 */
5493 if (rdev->saved_raid_disk >= 0 &&
5494 rdev->saved_raid_disk >= first &&
5495 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5496 first = rdev->saved_raid_disk;
5497
5498 for (disk = first; disk <= last; disk++) {
5499 p = conf->disks + disk;
5500 if (p->rdev == NULL) {
5501 clear_bit(In_sync, &rdev->flags);
5502 rdev->raid_disk = disk;
5503 err = 0;
5504 if (rdev->saved_raid_disk != disk)
5505 conf->fullsync = 1;
5506 rcu_assign_pointer(p->rdev, rdev);
5507 goto out;
5508 }
5509 }
5510 for (disk = first; disk <= last; disk++) {
5511 p = conf->disks + disk;
5512 if (test_bit(WantReplacement, &p->rdev->flags) &&
5513 p->replacement == NULL) {
5514 clear_bit(In_sync, &rdev->flags);
5515 set_bit(Replacement, &rdev->flags);
5516 rdev->raid_disk = disk;
5517 err = 0;
5518 conf->fullsync = 1;
5519 rcu_assign_pointer(p->replacement, rdev);
5520 break;
5521 }
5522 }
5523out:
5524 print_raid5_conf(conf);
5525 return err;
5526}
5527
5528static int raid5_resize(struct mddev *mddev, sector_t sectors)
5529{
5530 /* no resync is happening, and there is enough space
5531 * on all devices, so we can resize.
5532 * We need to make sure resync covers any new space.
5533 * If the array is shrinking we should possibly wait until
5534 * any io in the removed space completes, but it hardly seems
5535 * worth it.
5536 */
5537 sector_t newsize;
5538 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5539 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5540 if (mddev->external_size &&
5541 mddev->array_sectors > newsize)
5542 return -EINVAL;
5543 if (mddev->bitmap) {
5544 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5545 if (ret)
5546 return ret;
5547 }
5548 md_set_array_sectors(mddev, newsize);
5549 set_capacity(mddev->gendisk, mddev->array_sectors);
5550 revalidate_disk(mddev->gendisk);
5551 if (sectors > mddev->dev_sectors &&
5552 mddev->recovery_cp > mddev->dev_sectors) {
5553 mddev->recovery_cp = mddev->dev_sectors;
5554 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5555 }
5556 mddev->dev_sectors = sectors;
5557 mddev->resync_max_sectors = sectors;
5558 return 0;
5559}
5560
5561static int check_stripe_cache(struct mddev *mddev)
5562{
5563 /* Can only proceed if there are plenty of stripe_heads.
5564 * We need a minimum of one full stripe,, and for sensible progress
5565 * it is best to have about 4 times that.
5566 * If we require 4 times, then the default 256 4K stripe_heads will
5567 * allow for chunk sizes up to 256K, which is probably OK.
5568 * If the chunk size is greater, user-space should request more
5569 * stripe_heads first.
5570 */
5571 struct r5conf *conf = mddev->private;
5572 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5573 > conf->max_nr_stripes ||
5574 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5575 > conf->max_nr_stripes) {
5576 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5577 mdname(mddev),
5578 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5579 / STRIPE_SIZE)*4);
5580 return 0;
5581 }
5582 return 1;
5583}
5584
5585static int check_reshape(struct mddev *mddev)
5586{
5587 struct r5conf *conf = mddev->private;
5588
5589 if (mddev->delta_disks == 0 &&
5590 mddev->new_layout == mddev->layout &&
5591 mddev->new_chunk_sectors == mddev->chunk_sectors)
5592 return 0; /* nothing to do */
5593 if (has_failed(conf))
5594 return -EINVAL;
5595 if (mddev->delta_disks < 0) {
5596 /* We might be able to shrink, but the devices must
5597 * be made bigger first.
5598 * For raid6, 4 is the minimum size.
5599 * Otherwise 2 is the minimum
5600 */
5601 int min = 2;
5602 if (mddev->level == 6)
5603 min = 4;
5604 if (mddev->raid_disks + mddev->delta_disks < min)
5605 return -EINVAL;
5606 }
5607
5608 if (!check_stripe_cache(mddev))
5609 return -ENOSPC;
5610
5611 return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5612}
5613
5614static int raid5_start_reshape(struct mddev *mddev)
5615{
5616 struct r5conf *conf = mddev->private;
5617 struct md_rdev *rdev;
5618 int spares = 0;
5619 unsigned long flags;
5620
5621 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5622 return -EBUSY;
5623
5624 if (!check_stripe_cache(mddev))
5625 return -ENOSPC;
5626
5627 if (has_failed(conf))
5628 return -EINVAL;
5629
5630 rdev_for_each(rdev, mddev) {
5631 if (!test_bit(In_sync, &rdev->flags)
5632 && !test_bit(Faulty, &rdev->flags))
5633 spares++;
5634 }
5635
5636 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5637 /* Not enough devices even to make a degraded array
5638 * of that size
5639 */
5640 return -EINVAL;
5641
5642 /* Refuse to reduce size of the array. Any reductions in
5643 * array size must be through explicit setting of array_size
5644 * attribute.
5645 */
5646 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5647 < mddev->array_sectors) {
5648 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5649 "before number of disks\n", mdname(mddev));
5650 return -EINVAL;
5651 }
5652
5653 atomic_set(&conf->reshape_stripes, 0);
5654 spin_lock_irq(&conf->device_lock);
5655 conf->previous_raid_disks = conf->raid_disks;
5656 conf->raid_disks += mddev->delta_disks;
5657 conf->prev_chunk_sectors = conf->chunk_sectors;
5658 conf->chunk_sectors = mddev->new_chunk_sectors;
5659 conf->prev_algo = conf->algorithm;
5660 conf->algorithm = mddev->new_layout;
5661 conf->generation++;
5662 /* Code that selects data_offset needs to see the generation update
5663 * if reshape_progress has been set - so a memory barrier needed.
5664 */
5665 smp_mb();
5666 if (mddev->reshape_backwards)
5667 conf->reshape_progress = raid5_size(mddev, 0, 0);
5668 else
5669 conf->reshape_progress = 0;
5670 conf->reshape_safe = conf->reshape_progress;
5671 spin_unlock_irq(&conf->device_lock);
5672
5673 /* Add some new drives, as many as will fit.
5674 * We know there are enough to make the newly sized array work.
5675 * Don't add devices if we are reducing the number of
5676 * devices in the array. This is because it is not possible
5677 * to correctly record the "partially reconstructed" state of
5678 * such devices during the reshape and confusion could result.
5679 */
5680 if (mddev->delta_disks >= 0) {
5681 rdev_for_each(rdev, mddev)
5682 if (rdev->raid_disk < 0 &&
5683 !test_bit(Faulty, &rdev->flags)) {
5684 if (raid5_add_disk(mddev, rdev) == 0) {
5685 if (rdev->raid_disk
5686 >= conf->previous_raid_disks)
5687 set_bit(In_sync, &rdev->flags);
5688 else
5689 rdev->recovery_offset = 0;
5690
5691 if (sysfs_link_rdev(mddev, rdev))
5692 /* Failure here is OK */;
5693 }
5694 } else if (rdev->raid_disk >= conf->previous_raid_disks
5695 && !test_bit(Faulty, &rdev->flags)) {
5696 /* This is a spare that was manually added */
5697 set_bit(In_sync, &rdev->flags);
5698 }
5699
5700 /* When a reshape changes the number of devices,
5701 * ->degraded is measured against the larger of the
5702 * pre and post number of devices.
5703 */
5704 spin_lock_irqsave(&conf->device_lock, flags);
5705 mddev->degraded = calc_degraded(conf);
5706 spin_unlock_irqrestore(&conf->device_lock, flags);
5707 }
5708 mddev->raid_disks = conf->raid_disks;
5709 mddev->reshape_position = conf->reshape_progress;
5710 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5711
5712 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5713 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5714 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5715 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5716 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5717 "reshape");
5718 if (!mddev->sync_thread) {
5719 mddev->recovery = 0;
5720 spin_lock_irq(&conf->device_lock);
5721 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5722 rdev_for_each(rdev, mddev)
5723 rdev->new_data_offset = rdev->data_offset;
5724 smp_wmb();
5725 conf->reshape_progress = MaxSector;
5726 mddev->reshape_position = MaxSector;
5727 spin_unlock_irq(&conf->device_lock);
5728 return -EAGAIN;
5729 }
5730 conf->reshape_checkpoint = jiffies;
5731 md_wakeup_thread(mddev->sync_thread);
5732 md_new_event(mddev);
5733 return 0;
5734}
5735
5736/* This is called from the reshape thread and should make any
5737 * changes needed in 'conf'
5738 */
5739static void end_reshape(struct r5conf *conf)
5740{
5741
5742 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5743 struct md_rdev *rdev;
5744
5745 spin_lock_irq(&conf->device_lock);
5746 conf->previous_raid_disks = conf->raid_disks;
5747 rdev_for_each(rdev, conf->mddev)
5748 rdev->data_offset = rdev->new_data_offset;
5749 smp_wmb();
5750 conf->reshape_progress = MaxSector;
5751 spin_unlock_irq(&conf->device_lock);
5752 wake_up(&conf->wait_for_overlap);
5753
5754 /* read-ahead size must cover two whole stripes, which is
5755 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5756 */
5757 if (conf->mddev->queue) {
5758 int data_disks = conf->raid_disks - conf->max_degraded;
5759 int stripe = data_disks * ((conf->chunk_sectors << 9)
5760 / PAGE_SIZE);
5761 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5762 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5763 }
5764 }
5765}
5766
5767/* This is called from the raid5d thread with mddev_lock held.
5768 * It makes config changes to the device.
5769 */
5770static void raid5_finish_reshape(struct mddev *mddev)
5771{
5772 struct r5conf *conf = mddev->private;
5773
5774 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5775
5776 if (mddev->delta_disks > 0) {
5777 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5778 set_capacity(mddev->gendisk, mddev->array_sectors);
5779 revalidate_disk(mddev->gendisk);
5780 } else {
5781 int d;
5782 spin_lock_irq(&conf->device_lock);
5783 mddev->degraded = calc_degraded(conf);
5784 spin_unlock_irq(&conf->device_lock);
5785 for (d = conf->raid_disks ;
5786 d < conf->raid_disks - mddev->delta_disks;
5787 d++) {
5788 struct md_rdev *rdev = conf->disks[d].rdev;
5789 if (rdev)
5790 clear_bit(In_sync, &rdev->flags);
5791 rdev = conf->disks[d].replacement;
5792 if (rdev)
5793 clear_bit(In_sync, &rdev->flags);
5794 }
5795 }
5796 mddev->layout = conf->algorithm;
5797 mddev->chunk_sectors = conf->chunk_sectors;
5798 mddev->reshape_position = MaxSector;
5799 mddev->delta_disks = 0;
5800 mddev->reshape_backwards = 0;
5801 }
5802}
5803
5804static void raid5_quiesce(struct mddev *mddev, int state)
5805{
5806 struct r5conf *conf = mddev->private;
5807
5808 switch(state) {
5809 case 2: /* resume for a suspend */
5810 wake_up(&conf->wait_for_overlap);
5811 break;
5812
5813 case 1: /* stop all writes */
5814 spin_lock_irq(&conf->device_lock);
5815 /* '2' tells resync/reshape to pause so that all
5816 * active stripes can drain
5817 */
5818 conf->quiesce = 2;
5819 wait_event_lock_irq(conf->wait_for_stripe,
5820 atomic_read(&conf->active_stripes) == 0 &&
5821 atomic_read(&conf->active_aligned_reads) == 0,
5822 conf->device_lock, /* nothing */);
5823 conf->quiesce = 1;
5824 spin_unlock_irq(&conf->device_lock);
5825 /* allow reshape to continue */
5826 wake_up(&conf->wait_for_overlap);
5827 break;
5828
5829 case 0: /* re-enable writes */
5830 spin_lock_irq(&conf->device_lock);
5831 conf->quiesce = 0;
5832 wake_up(&conf->wait_for_stripe);
5833 wake_up(&conf->wait_for_overlap);
5834 spin_unlock_irq(&conf->device_lock);
5835 break;
5836 }
5837}
5838
5839
5840static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5841{
5842 struct r0conf *raid0_conf = mddev->private;
5843 sector_t sectors;
5844
5845 /* for raid0 takeover only one zone is supported */
5846 if (raid0_conf->nr_strip_zones > 1) {
5847 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5848 mdname(mddev));
5849 return ERR_PTR(-EINVAL);
5850 }
5851
5852 sectors = raid0_conf->strip_zone[0].zone_end;
5853 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5854 mddev->dev_sectors = sectors;
5855 mddev->new_level = level;
5856 mddev->new_layout = ALGORITHM_PARITY_N;
5857 mddev->new_chunk_sectors = mddev->chunk_sectors;
5858 mddev->raid_disks += 1;
5859 mddev->delta_disks = 1;
5860 /* make sure it will be not marked as dirty */
5861 mddev->recovery_cp = MaxSector;
5862
5863 return setup_conf(mddev);
5864}
5865
5866
5867static void *raid5_takeover_raid1(struct mddev *mddev)
5868{
5869 int chunksect;
5870
5871 if (mddev->raid_disks != 2 ||
5872 mddev->degraded > 1)
5873 return ERR_PTR(-EINVAL);
5874
5875 /* Should check if there are write-behind devices? */
5876
5877 chunksect = 64*2; /* 64K by default */
5878
5879 /* The array must be an exact multiple of chunksize */
5880 while (chunksect && (mddev->array_sectors & (chunksect-1)))
5881 chunksect >>= 1;
5882
5883 if ((chunksect<<9) < STRIPE_SIZE)
5884 /* array size does not allow a suitable chunk size */
5885 return ERR_PTR(-EINVAL);
5886
5887 mddev->new_level = 5;
5888 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5889 mddev->new_chunk_sectors = chunksect;
5890
5891 return setup_conf(mddev);
5892}
5893
5894static void *raid5_takeover_raid6(struct mddev *mddev)
5895{
5896 int new_layout;
5897
5898 switch (mddev->layout) {
5899 case ALGORITHM_LEFT_ASYMMETRIC_6:
5900 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5901 break;
5902 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5903 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5904 break;
5905 case ALGORITHM_LEFT_SYMMETRIC_6:
5906 new_layout = ALGORITHM_LEFT_SYMMETRIC;
5907 break;
5908 case ALGORITHM_RIGHT_SYMMETRIC_6:
5909 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5910 break;
5911 case ALGORITHM_PARITY_0_6:
5912 new_layout = ALGORITHM_PARITY_0;
5913 break;
5914 case ALGORITHM_PARITY_N:
5915 new_layout = ALGORITHM_PARITY_N;
5916 break;
5917 default:
5918 return ERR_PTR(-EINVAL);
5919 }
5920 mddev->new_level = 5;
5921 mddev->new_layout = new_layout;
5922 mddev->delta_disks = -1;
5923 mddev->raid_disks -= 1;
5924 return setup_conf(mddev);
5925}
5926
5927
5928static int raid5_check_reshape(struct mddev *mddev)
5929{
5930 /* For a 2-drive array, the layout and chunk size can be changed
5931 * immediately as not restriping is needed.
5932 * For larger arrays we record the new value - after validation
5933 * to be used by a reshape pass.
5934 */
5935 struct r5conf *conf = mddev->private;
5936 int new_chunk = mddev->new_chunk_sectors;
5937
5938 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5939 return -EINVAL;
5940 if (new_chunk > 0) {
5941 if (!is_power_of_2(new_chunk))
5942 return -EINVAL;
5943 if (new_chunk < (PAGE_SIZE>>9))
5944 return -EINVAL;
5945 if (mddev->array_sectors & (new_chunk-1))
5946 /* not factor of array size */
5947 return -EINVAL;
5948 }
5949
5950 /* They look valid */
5951
5952 if (mddev->raid_disks == 2) {
5953 /* can make the change immediately */
5954 if (mddev->new_layout >= 0) {
5955 conf->algorithm = mddev->new_layout;
5956 mddev->layout = mddev->new_layout;
5957 }
5958 if (new_chunk > 0) {
5959 conf->chunk_sectors = new_chunk ;
5960 mddev->chunk_sectors = new_chunk;
5961 }
5962 set_bit(MD_CHANGE_DEVS, &mddev->flags);
5963 md_wakeup_thread(mddev->thread);
5964 }
5965 return check_reshape(mddev);
5966}
5967
5968static int raid6_check_reshape(struct mddev *mddev)
5969{
5970 int new_chunk = mddev->new_chunk_sectors;
5971
5972 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5973 return -EINVAL;
5974 if (new_chunk > 0) {
5975 if (!is_power_of_2(new_chunk))
5976 return -EINVAL;
5977 if (new_chunk < (PAGE_SIZE >> 9))
5978 return -EINVAL;
5979 if (mddev->array_sectors & (new_chunk-1))
5980 /* not factor of array size */
5981 return -EINVAL;
5982 }
5983
5984 /* They look valid */
5985 return check_reshape(mddev);
5986}
5987
5988static void *raid5_takeover(struct mddev *mddev)
5989{
5990 /* raid5 can take over:
5991 * raid0 - if there is only one strip zone - make it a raid4 layout
5992 * raid1 - if there are two drives. We need to know the chunk size
5993 * raid4 - trivial - just use a raid4 layout.
5994 * raid6 - Providing it is a *_6 layout
5995 */
5996 if (mddev->level == 0)
5997 return raid45_takeover_raid0(mddev, 5);
5998 if (mddev->level == 1)
5999 return raid5_takeover_raid1(mddev);
6000 if (mddev->level == 4) {
6001 mddev->new_layout = ALGORITHM_PARITY_N;
6002 mddev->new_level = 5;
6003 return setup_conf(mddev);
6004 }
6005 if (mddev->level == 6)
6006 return raid5_takeover_raid6(mddev);
6007
6008 return ERR_PTR(-EINVAL);
6009}
6010
6011static void *raid4_takeover(struct mddev *mddev)
6012{
6013 /* raid4 can take over:
6014 * raid0 - if there is only one strip zone
6015 * raid5 - if layout is right
6016 */
6017 if (mddev->level == 0)
6018 return raid45_takeover_raid0(mddev, 4);
6019 if (mddev->level == 5 &&
6020 mddev->layout == ALGORITHM_PARITY_N) {
6021 mddev->new_layout = 0;
6022 mddev->new_level = 4;
6023 return setup_conf(mddev);
6024 }
6025 return ERR_PTR(-EINVAL);
6026}
6027
6028static struct md_personality raid5_personality;
6029
6030static void *raid6_takeover(struct mddev *mddev)
6031{
6032 /* Currently can only take over a raid5. We map the
6033 * personality to an equivalent raid6 personality
6034 * with the Q block at the end.
6035 */
6036 int new_layout;
6037
6038 if (mddev->pers != &raid5_personality)
6039 return ERR_PTR(-EINVAL);
6040 if (mddev->degraded > 1)
6041 return ERR_PTR(-EINVAL);
6042 if (mddev->raid_disks > 253)
6043 return ERR_PTR(-EINVAL);
6044 if (mddev->raid_disks < 3)
6045 return ERR_PTR(-EINVAL);
6046
6047 switch (mddev->layout) {
6048 case ALGORITHM_LEFT_ASYMMETRIC:
6049 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6050 break;
6051 case ALGORITHM_RIGHT_ASYMMETRIC:
6052 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6053 break;
6054 case ALGORITHM_LEFT_SYMMETRIC:
6055 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6056 break;
6057 case ALGORITHM_RIGHT_SYMMETRIC:
6058 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6059 break;
6060 case ALGORITHM_PARITY_0:
6061 new_layout = ALGORITHM_PARITY_0_6;
6062 break;
6063 case ALGORITHM_PARITY_N:
6064 new_layout = ALGORITHM_PARITY_N;
6065 break;
6066 default:
6067 return ERR_PTR(-EINVAL);
6068 }
6069 mddev->new_level = 6;
6070 mddev->new_layout = new_layout;
6071 mddev->delta_disks = 1;
6072 mddev->raid_disks += 1;
6073 return setup_conf(mddev);
6074}
6075
6076
6077static struct md_personality raid6_personality =
6078{
6079 .name = "raid6",
6080 .level = 6,
6081 .owner = THIS_MODULE,
6082 .make_request = make_request,
6083 .run = run,
6084 .stop = stop,
6085 .status = status,
6086 .error_handler = error,
6087 .hot_add_disk = raid5_add_disk,
6088 .hot_remove_disk= raid5_remove_disk,
6089 .spare_active = raid5_spare_active,
6090 .sync_request = sync_request,
6091 .resize = raid5_resize,
6092 .size = raid5_size,
6093 .check_reshape = raid6_check_reshape,
6094 .start_reshape = raid5_start_reshape,
6095 .finish_reshape = raid5_finish_reshape,
6096 .quiesce = raid5_quiesce,
6097 .takeover = raid6_takeover,
6098};
6099static struct md_personality raid5_personality =
6100{
6101 .name = "raid5",
6102 .level = 5,
6103 .owner = THIS_MODULE,
6104 .make_request = make_request,
6105 .run = run,
6106 .stop = stop,
6107 .status = status,
6108 .error_handler = error,
6109 .hot_add_disk = raid5_add_disk,
6110 .hot_remove_disk= raid5_remove_disk,
6111 .spare_active = raid5_spare_active,
6112 .sync_request = sync_request,
6113 .resize = raid5_resize,
6114 .size = raid5_size,
6115 .check_reshape = raid5_check_reshape,
6116 .start_reshape = raid5_start_reshape,
6117 .finish_reshape = raid5_finish_reshape,
6118 .quiesce = raid5_quiesce,
6119 .takeover = raid5_takeover,
6120};
6121
6122static struct md_personality raid4_personality =
6123{
6124 .name = "raid4",
6125 .level = 4,
6126 .owner = THIS_MODULE,
6127 .make_request = make_request,
6128 .run = run,
6129 .stop = stop,
6130 .status = status,
6131 .error_handler = error,
6132 .hot_add_disk = raid5_add_disk,
6133 .hot_remove_disk= raid5_remove_disk,
6134 .spare_active = raid5_spare_active,
6135 .sync_request = sync_request,
6136 .resize = raid5_resize,
6137 .size = raid5_size,
6138 .check_reshape = raid5_check_reshape,
6139 .start_reshape = raid5_start_reshape,
6140 .finish_reshape = raid5_finish_reshape,
6141 .quiesce = raid5_quiesce,
6142 .takeover = raid4_takeover,
6143};
6144
6145static int __init raid5_init(void)
6146{
6147 register_md_personality(&raid6_personality);
6148 register_md_personality(&raid5_personality);
6149 register_md_personality(&raid4_personality);
6150 return 0;
6151}
6152
6153static void raid5_exit(void)
6154{
6155 unregister_md_personality(&raid6_personality);
6156 unregister_md_personality(&raid5_personality);
6157 unregister_md_personality(&raid4_personality);
6158}
6159
6160module_init(raid5_init);
6161module_exit(raid5_exit);
6162MODULE_LICENSE("GPL");
6163MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6164MODULE_ALIAS("md-personality-4"); /* RAID5 */
6165MODULE_ALIAS("md-raid5");
6166MODULE_ALIAS("md-raid4");
6167MODULE_ALIAS("md-level-5");
6168MODULE_ALIAS("md-level-4");
6169MODULE_ALIAS("md-personality-8"); /* RAID6 */
6170MODULE_ALIAS("md-raid6");
6171MODULE_ALIAS("md-level-6");
6172
6173/* This used to be two separate modules, they were: */
6174MODULE_ALIAS("raid5");
6175MODULE_ALIAS("raid6");
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * raid5.c : Multiple Devices driver for Linux
4 * Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
5 * Copyright (C) 1999, 2000 Ingo Molnar
6 * Copyright (C) 2002, 2003 H. Peter Anvin
7 *
8 * RAID-4/5/6 management functions.
9 * Thanks to Penguin Computing for making the RAID-6 development possible
10 * by donating a test server!
11 */
12
13/*
14 * BITMAP UNPLUGGING:
15 *
16 * The sequencing for updating the bitmap reliably is a little
17 * subtle (and I got it wrong the first time) so it deserves some
18 * explanation.
19 *
20 * We group bitmap updates into batches. Each batch has a number.
21 * We may write out several batches at once, but that isn't very important.
22 * conf->seq_write is the number of the last batch successfully written.
23 * conf->seq_flush is the number of the last batch that was closed to
24 * new additions.
25 * When we discover that we will need to write to any block in a stripe
26 * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
27 * the number of the batch it will be in. This is seq_flush+1.
28 * When we are ready to do a write, if that batch hasn't been written yet,
29 * we plug the array and queue the stripe for later.
30 * When an unplug happens, we increment bm_flush, thus closing the current
31 * batch.
32 * When we notice that bm_flush > bm_write, we write out all pending updates
33 * to the bitmap, and advance bm_write to where bm_flush was.
34 * This may occasionally write a bit out twice, but is sure never to
35 * miss any bits.
36 */
37
38#include <linux/blkdev.h>
39#include <linux/kthread.h>
40#include <linux/raid/pq.h>
41#include <linux/async_tx.h>
42#include <linux/module.h>
43#include <linux/async.h>
44#include <linux/seq_file.h>
45#include <linux/cpu.h>
46#include <linux/slab.h>
47#include <linux/ratelimit.h>
48#include <linux/nodemask.h>
49
50#include <trace/events/block.h>
51#include <linux/list_sort.h>
52
53#include "md.h"
54#include "raid5.h"
55#include "raid0.h"
56#include "md-bitmap.h"
57#include "raid5-log.h"
58
59#define UNSUPPORTED_MDDEV_FLAGS (1L << MD_FAILFAST_SUPPORTED)
60
61#define cpu_to_group(cpu) cpu_to_node(cpu)
62#define ANY_GROUP NUMA_NO_NODE
63
64static bool devices_handle_discard_safely = false;
65module_param(devices_handle_discard_safely, bool, 0644);
66MODULE_PARM_DESC(devices_handle_discard_safely,
67 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
68static struct workqueue_struct *raid5_wq;
69
70static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
71{
72 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
73 return &conf->stripe_hashtbl[hash];
74}
75
76static inline int stripe_hash_locks_hash(sector_t sect)
77{
78 return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
79}
80
81static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
82{
83 spin_lock_irq(conf->hash_locks + hash);
84 spin_lock(&conf->device_lock);
85}
86
87static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
88{
89 spin_unlock(&conf->device_lock);
90 spin_unlock_irq(conf->hash_locks + hash);
91}
92
93static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
94{
95 int i;
96 spin_lock_irq(conf->hash_locks);
97 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
98 spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
99 spin_lock(&conf->device_lock);
100}
101
102static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
103{
104 int i;
105 spin_unlock(&conf->device_lock);
106 for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
107 spin_unlock(conf->hash_locks + i);
108 spin_unlock_irq(conf->hash_locks);
109}
110
111/* Find first data disk in a raid6 stripe */
112static inline int raid6_d0(struct stripe_head *sh)
113{
114 if (sh->ddf_layout)
115 /* ddf always start from first device */
116 return 0;
117 /* md starts just after Q block */
118 if (sh->qd_idx == sh->disks - 1)
119 return 0;
120 else
121 return sh->qd_idx + 1;
122}
123static inline int raid6_next_disk(int disk, int raid_disks)
124{
125 disk++;
126 return (disk < raid_disks) ? disk : 0;
127}
128
129/* When walking through the disks in a raid5, starting at raid6_d0,
130 * We need to map each disk to a 'slot', where the data disks are slot
131 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
132 * is raid_disks-1. This help does that mapping.
133 */
134static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
135 int *count, int syndrome_disks)
136{
137 int slot = *count;
138
139 if (sh->ddf_layout)
140 (*count)++;
141 if (idx == sh->pd_idx)
142 return syndrome_disks;
143 if (idx == sh->qd_idx)
144 return syndrome_disks + 1;
145 if (!sh->ddf_layout)
146 (*count)++;
147 return slot;
148}
149
150static void print_raid5_conf (struct r5conf *conf);
151
152static int stripe_operations_active(struct stripe_head *sh)
153{
154 return sh->check_state || sh->reconstruct_state ||
155 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
156 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
157}
158
159static bool stripe_is_lowprio(struct stripe_head *sh)
160{
161 return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
162 test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
163 !test_bit(STRIPE_R5C_CACHING, &sh->state);
164}
165
166static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
167{
168 struct r5conf *conf = sh->raid_conf;
169 struct r5worker_group *group;
170 int thread_cnt;
171 int i, cpu = sh->cpu;
172
173 if (!cpu_online(cpu)) {
174 cpu = cpumask_any(cpu_online_mask);
175 sh->cpu = cpu;
176 }
177
178 if (list_empty(&sh->lru)) {
179 struct r5worker_group *group;
180 group = conf->worker_groups + cpu_to_group(cpu);
181 if (stripe_is_lowprio(sh))
182 list_add_tail(&sh->lru, &group->loprio_list);
183 else
184 list_add_tail(&sh->lru, &group->handle_list);
185 group->stripes_cnt++;
186 sh->group = group;
187 }
188
189 if (conf->worker_cnt_per_group == 0) {
190 md_wakeup_thread(conf->mddev->thread);
191 return;
192 }
193
194 group = conf->worker_groups + cpu_to_group(sh->cpu);
195
196 group->workers[0].working = true;
197 /* at least one worker should run to avoid race */
198 queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
199
200 thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
201 /* wakeup more workers */
202 for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
203 if (group->workers[i].working == false) {
204 group->workers[i].working = true;
205 queue_work_on(sh->cpu, raid5_wq,
206 &group->workers[i].work);
207 thread_cnt--;
208 }
209 }
210}
211
212static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
213 struct list_head *temp_inactive_list)
214{
215 int i;
216 int injournal = 0; /* number of date pages with R5_InJournal */
217
218 BUG_ON(!list_empty(&sh->lru));
219 BUG_ON(atomic_read(&conf->active_stripes)==0);
220
221 if (r5c_is_writeback(conf->log))
222 for (i = sh->disks; i--; )
223 if (test_bit(R5_InJournal, &sh->dev[i].flags))
224 injournal++;
225 /*
226 * In the following cases, the stripe cannot be released to cached
227 * lists. Therefore, we make the stripe write out and set
228 * STRIPE_HANDLE:
229 * 1. when quiesce in r5c write back;
230 * 2. when resync is requested fot the stripe.
231 */
232 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
233 (conf->quiesce && r5c_is_writeback(conf->log) &&
234 !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
235 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
236 r5c_make_stripe_write_out(sh);
237 set_bit(STRIPE_HANDLE, &sh->state);
238 }
239
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);
250 if (conf->worker_cnt_per_group == 0) {
251 if (stripe_is_lowprio(sh))
252 list_add_tail(&sh->lru,
253 &conf->loprio_list);
254 else
255 list_add_tail(&sh->lru,
256 &conf->handle_list);
257 } else {
258 raid5_wakeup_stripe_thread(sh);
259 return;
260 }
261 }
262 md_wakeup_thread(conf->mddev->thread);
263 } else {
264 BUG_ON(stripe_operations_active(sh));
265 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
266 if (atomic_dec_return(&conf->preread_active_stripes)
267 < IO_THRESHOLD)
268 md_wakeup_thread(conf->mddev->thread);
269 atomic_dec(&conf->active_stripes);
270 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
271 if (!r5c_is_writeback(conf->log))
272 list_add_tail(&sh->lru, temp_inactive_list);
273 else {
274 WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
275 if (injournal == 0)
276 list_add_tail(&sh->lru, temp_inactive_list);
277 else if (injournal == conf->raid_disks - conf->max_degraded) {
278 /* full stripe */
279 if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
280 atomic_inc(&conf->r5c_cached_full_stripes);
281 if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
282 atomic_dec(&conf->r5c_cached_partial_stripes);
283 list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
284 r5c_check_cached_full_stripe(conf);
285 } else
286 /*
287 * STRIPE_R5C_PARTIAL_STRIPE is set in
288 * r5c_try_caching_write(). No need to
289 * set it again.
290 */
291 list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
292 }
293 }
294 }
295}
296
297static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
298 struct list_head *temp_inactive_list)
299{
300 if (atomic_dec_and_test(&sh->count))
301 do_release_stripe(conf, sh, temp_inactive_list);
302}
303
304/*
305 * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
306 *
307 * Be careful: Only one task can add/delete stripes from temp_inactive_list at
308 * given time. Adding stripes only takes device lock, while deleting stripes
309 * only takes hash lock.
310 */
311static void release_inactive_stripe_list(struct r5conf *conf,
312 struct list_head *temp_inactive_list,
313 int hash)
314{
315 int size;
316 bool do_wakeup = false;
317 unsigned long flags;
318
319 if (hash == NR_STRIPE_HASH_LOCKS) {
320 size = NR_STRIPE_HASH_LOCKS;
321 hash = NR_STRIPE_HASH_LOCKS - 1;
322 } else
323 size = 1;
324 while (size) {
325 struct list_head *list = &temp_inactive_list[size - 1];
326
327 /*
328 * We don't hold any lock here yet, raid5_get_active_stripe() might
329 * remove stripes from the list
330 */
331 if (!list_empty_careful(list)) {
332 spin_lock_irqsave(conf->hash_locks + hash, flags);
333 if (list_empty(conf->inactive_list + hash) &&
334 !list_empty(list))
335 atomic_dec(&conf->empty_inactive_list_nr);
336 list_splice_tail_init(list, conf->inactive_list + hash);
337 do_wakeup = true;
338 spin_unlock_irqrestore(conf->hash_locks + hash, flags);
339 }
340 size--;
341 hash--;
342 }
343
344 if (do_wakeup) {
345 wake_up(&conf->wait_for_stripe);
346 if (atomic_read(&conf->active_stripes) == 0)
347 wake_up(&conf->wait_for_quiescent);
348 if (conf->retry_read_aligned)
349 md_wakeup_thread(conf->mddev->thread);
350 }
351}
352
353/* should hold conf->device_lock already */
354static int release_stripe_list(struct r5conf *conf,
355 struct list_head *temp_inactive_list)
356{
357 struct stripe_head *sh, *t;
358 int count = 0;
359 struct llist_node *head;
360
361 head = llist_del_all(&conf->released_stripes);
362 head = llist_reverse_order(head);
363 llist_for_each_entry_safe(sh, t, head, release_list) {
364 int hash;
365
366 /* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
367 smp_mb();
368 clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
369 /*
370 * Don't worry the bit is set here, because if the bit is set
371 * again, the count is always > 1. This is true for
372 * STRIPE_ON_UNPLUG_LIST bit too.
373 */
374 hash = sh->hash_lock_index;
375 __release_stripe(conf, sh, &temp_inactive_list[hash]);
376 count++;
377 }
378
379 return count;
380}
381
382void raid5_release_stripe(struct stripe_head *sh)
383{
384 struct r5conf *conf = sh->raid_conf;
385 unsigned long flags;
386 struct list_head list;
387 int hash;
388 bool wakeup;
389
390 /* Avoid release_list until the last reference.
391 */
392 if (atomic_add_unless(&sh->count, -1, 1))
393 return;
394
395 if (unlikely(!conf->mddev->thread) ||
396 test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
397 goto slow_path;
398 wakeup = llist_add(&sh->release_list, &conf->released_stripes);
399 if (wakeup)
400 md_wakeup_thread(conf->mddev->thread);
401 return;
402slow_path:
403 /* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
404 if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
405 INIT_LIST_HEAD(&list);
406 hash = sh->hash_lock_index;
407 do_release_stripe(conf, sh, &list);
408 spin_unlock_irqrestore(&conf->device_lock, flags);
409 release_inactive_stripe_list(conf, &list, hash);
410 }
411}
412
413static inline void remove_hash(struct stripe_head *sh)
414{
415 pr_debug("remove_hash(), stripe %llu\n",
416 (unsigned long long)sh->sector);
417
418 hlist_del_init(&sh->hash);
419}
420
421static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
422{
423 struct hlist_head *hp = stripe_hash(conf, sh->sector);
424
425 pr_debug("insert_hash(), stripe %llu\n",
426 (unsigned long long)sh->sector);
427
428 hlist_add_head(&sh->hash, hp);
429}
430
431/* find an idle stripe, make sure it is unhashed, and return it. */
432static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
433{
434 struct stripe_head *sh = NULL;
435 struct list_head *first;
436
437 if (list_empty(conf->inactive_list + hash))
438 goto out;
439 first = (conf->inactive_list + hash)->next;
440 sh = list_entry(first, struct stripe_head, lru);
441 list_del_init(first);
442 remove_hash(sh);
443 atomic_inc(&conf->active_stripes);
444 BUG_ON(hash != sh->hash_lock_index);
445 if (list_empty(conf->inactive_list + hash))
446 atomic_inc(&conf->empty_inactive_list_nr);
447out:
448 return sh;
449}
450
451static void shrink_buffers(struct stripe_head *sh)
452{
453 struct page *p;
454 int i;
455 int num = sh->raid_conf->pool_size;
456
457 for (i = 0; i < num ; i++) {
458 WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
459 p = sh->dev[i].page;
460 if (!p)
461 continue;
462 sh->dev[i].page = NULL;
463 put_page(p);
464 }
465}
466
467static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
468{
469 int i;
470 int num = sh->raid_conf->pool_size;
471
472 for (i = 0; i < num; i++) {
473 struct page *page;
474
475 if (!(page = alloc_page(gfp))) {
476 return 1;
477 }
478 sh->dev[i].page = page;
479 sh->dev[i].orig_page = page;
480 }
481
482 return 0;
483}
484
485static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
486 struct stripe_head *sh);
487
488static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
489{
490 struct r5conf *conf = sh->raid_conf;
491 int i, seq;
492
493 BUG_ON(atomic_read(&sh->count) != 0);
494 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
495 BUG_ON(stripe_operations_active(sh));
496 BUG_ON(sh->batch_head);
497
498 pr_debug("init_stripe called, stripe %llu\n",
499 (unsigned long long)sector);
500retry:
501 seq = read_seqcount_begin(&conf->gen_lock);
502 sh->generation = conf->generation - previous;
503 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
504 sh->sector = sector;
505 stripe_set_idx(sector, conf, previous, sh);
506 sh->state = 0;
507
508 for (i = sh->disks; i--; ) {
509 struct r5dev *dev = &sh->dev[i];
510
511 if (dev->toread || dev->read || dev->towrite || dev->written ||
512 test_bit(R5_LOCKED, &dev->flags)) {
513 pr_err("sector=%llx i=%d %p %p %p %p %d\n",
514 (unsigned long long)sh->sector, i, dev->toread,
515 dev->read, dev->towrite, dev->written,
516 test_bit(R5_LOCKED, &dev->flags));
517 WARN_ON(1);
518 }
519 dev->flags = 0;
520 dev->sector = raid5_compute_blocknr(sh, i, previous);
521 }
522 if (read_seqcount_retry(&conf->gen_lock, seq))
523 goto retry;
524 sh->overwrite_disks = 0;
525 insert_hash(conf, sh);
526 sh->cpu = smp_processor_id();
527 set_bit(STRIPE_BATCH_READY, &sh->state);
528}
529
530static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
531 short generation)
532{
533 struct stripe_head *sh;
534
535 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
536 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
537 if (sh->sector == sector && sh->generation == generation)
538 return sh;
539 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
540 return NULL;
541}
542
543/*
544 * Need to check if array has failed when deciding whether to:
545 * - start an array
546 * - remove non-faulty devices
547 * - add a spare
548 * - allow a reshape
549 * This determination is simple when no reshape is happening.
550 * However if there is a reshape, we need to carefully check
551 * both the before and after sections.
552 * This is because some failed devices may only affect one
553 * of the two sections, and some non-in_sync devices may
554 * be insync in the section most affected by failed devices.
555 */
556int raid5_calc_degraded(struct r5conf *conf)
557{
558 int degraded, degraded2;
559 int i;
560
561 rcu_read_lock();
562 degraded = 0;
563 for (i = 0; i < conf->previous_raid_disks; i++) {
564 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
565 if (rdev && test_bit(Faulty, &rdev->flags))
566 rdev = rcu_dereference(conf->disks[i].replacement);
567 if (!rdev || test_bit(Faulty, &rdev->flags))
568 degraded++;
569 else if (test_bit(In_sync, &rdev->flags))
570 ;
571 else
572 /* not in-sync or faulty.
573 * If the reshape increases the number of devices,
574 * this is being recovered by the reshape, so
575 * this 'previous' section is not in_sync.
576 * If the number of devices is being reduced however,
577 * the device can only be part of the array if
578 * we are reverting a reshape, so this section will
579 * be in-sync.
580 */
581 if (conf->raid_disks >= conf->previous_raid_disks)
582 degraded++;
583 }
584 rcu_read_unlock();
585 if (conf->raid_disks == conf->previous_raid_disks)
586 return degraded;
587 rcu_read_lock();
588 degraded2 = 0;
589 for (i = 0; i < conf->raid_disks; i++) {
590 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
591 if (rdev && test_bit(Faulty, &rdev->flags))
592 rdev = rcu_dereference(conf->disks[i].replacement);
593 if (!rdev || test_bit(Faulty, &rdev->flags))
594 degraded2++;
595 else if (test_bit(In_sync, &rdev->flags))
596 ;
597 else
598 /* not in-sync or faulty.
599 * If reshape increases the number of devices, this
600 * section has already been recovered, else it
601 * almost certainly hasn't.
602 */
603 if (conf->raid_disks <= conf->previous_raid_disks)
604 degraded2++;
605 }
606 rcu_read_unlock();
607 if (degraded2 > degraded)
608 return degraded2;
609 return degraded;
610}
611
612static int has_failed(struct r5conf *conf)
613{
614 int degraded;
615
616 if (conf->mddev->reshape_position == MaxSector)
617 return conf->mddev->degraded > conf->max_degraded;
618
619 degraded = raid5_calc_degraded(conf);
620 if (degraded > conf->max_degraded)
621 return 1;
622 return 0;
623}
624
625struct stripe_head *
626raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
627 int previous, int noblock, int noquiesce)
628{
629 struct stripe_head *sh;
630 int hash = stripe_hash_locks_hash(sector);
631 int inc_empty_inactive_list_flag;
632
633 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
634
635 spin_lock_irq(conf->hash_locks + hash);
636
637 do {
638 wait_event_lock_irq(conf->wait_for_quiescent,
639 conf->quiesce == 0 || noquiesce,
640 *(conf->hash_locks + hash));
641 sh = __find_stripe(conf, sector, conf->generation - previous);
642 if (!sh) {
643 if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
644 sh = get_free_stripe(conf, hash);
645 if (!sh && !test_bit(R5_DID_ALLOC,
646 &conf->cache_state))
647 set_bit(R5_ALLOC_MORE,
648 &conf->cache_state);
649 }
650 if (noblock && sh == NULL)
651 break;
652
653 r5c_check_stripe_cache_usage(conf);
654 if (!sh) {
655 set_bit(R5_INACTIVE_BLOCKED,
656 &conf->cache_state);
657 r5l_wake_reclaim(conf->log, 0);
658 wait_event_lock_irq(
659 conf->wait_for_stripe,
660 !list_empty(conf->inactive_list + hash) &&
661 (atomic_read(&conf->active_stripes)
662 < (conf->max_nr_stripes * 3 / 4)
663 || !test_bit(R5_INACTIVE_BLOCKED,
664 &conf->cache_state)),
665 *(conf->hash_locks + hash));
666 clear_bit(R5_INACTIVE_BLOCKED,
667 &conf->cache_state);
668 } else {
669 init_stripe(sh, sector, previous);
670 atomic_inc(&sh->count);
671 }
672 } else if (!atomic_inc_not_zero(&sh->count)) {
673 spin_lock(&conf->device_lock);
674 if (!atomic_read(&sh->count)) {
675 if (!test_bit(STRIPE_HANDLE, &sh->state))
676 atomic_inc(&conf->active_stripes);
677 BUG_ON(list_empty(&sh->lru) &&
678 !test_bit(STRIPE_EXPANDING, &sh->state));
679 inc_empty_inactive_list_flag = 0;
680 if (!list_empty(conf->inactive_list + hash))
681 inc_empty_inactive_list_flag = 1;
682 list_del_init(&sh->lru);
683 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
684 atomic_inc(&conf->empty_inactive_list_nr);
685 if (sh->group) {
686 sh->group->stripes_cnt--;
687 sh->group = NULL;
688 }
689 }
690 atomic_inc(&sh->count);
691 spin_unlock(&conf->device_lock);
692 }
693 } while (sh == NULL);
694
695 spin_unlock_irq(conf->hash_locks + hash);
696 return sh;
697}
698
699static bool is_full_stripe_write(struct stripe_head *sh)
700{
701 BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
702 return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
703}
704
705static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
706 __acquires(&sh1->stripe_lock)
707 __acquires(&sh2->stripe_lock)
708{
709 if (sh1 > sh2) {
710 spin_lock_irq(&sh2->stripe_lock);
711 spin_lock_nested(&sh1->stripe_lock, 1);
712 } else {
713 spin_lock_irq(&sh1->stripe_lock);
714 spin_lock_nested(&sh2->stripe_lock, 1);
715 }
716}
717
718static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
719 __releases(&sh1->stripe_lock)
720 __releases(&sh2->stripe_lock)
721{
722 spin_unlock(&sh1->stripe_lock);
723 spin_unlock_irq(&sh2->stripe_lock);
724}
725
726/* Only freshly new full stripe normal write stripe can be added to a batch list */
727static bool stripe_can_batch(struct stripe_head *sh)
728{
729 struct r5conf *conf = sh->raid_conf;
730
731 if (raid5_has_log(conf) || raid5_has_ppl(conf))
732 return false;
733 return test_bit(STRIPE_BATCH_READY, &sh->state) &&
734 !test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
735 is_full_stripe_write(sh);
736}
737
738/* we only do back search */
739static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
740{
741 struct stripe_head *head;
742 sector_t head_sector, tmp_sec;
743 int hash;
744 int dd_idx;
745 int inc_empty_inactive_list_flag;
746
747 /* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
748 tmp_sec = sh->sector;
749 if (!sector_div(tmp_sec, conf->chunk_sectors))
750 return;
751 head_sector = sh->sector - STRIPE_SECTORS;
752
753 hash = stripe_hash_locks_hash(head_sector);
754 spin_lock_irq(conf->hash_locks + hash);
755 head = __find_stripe(conf, head_sector, conf->generation);
756 if (head && !atomic_inc_not_zero(&head->count)) {
757 spin_lock(&conf->device_lock);
758 if (!atomic_read(&head->count)) {
759 if (!test_bit(STRIPE_HANDLE, &head->state))
760 atomic_inc(&conf->active_stripes);
761 BUG_ON(list_empty(&head->lru) &&
762 !test_bit(STRIPE_EXPANDING, &head->state));
763 inc_empty_inactive_list_flag = 0;
764 if (!list_empty(conf->inactive_list + hash))
765 inc_empty_inactive_list_flag = 1;
766 list_del_init(&head->lru);
767 if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
768 atomic_inc(&conf->empty_inactive_list_nr);
769 if (head->group) {
770 head->group->stripes_cnt--;
771 head->group = NULL;
772 }
773 }
774 atomic_inc(&head->count);
775 spin_unlock(&conf->device_lock);
776 }
777 spin_unlock_irq(conf->hash_locks + hash);
778
779 if (!head)
780 return;
781 if (!stripe_can_batch(head))
782 goto out;
783
784 lock_two_stripes(head, sh);
785 /* clear_batch_ready clear the flag */
786 if (!stripe_can_batch(head) || !stripe_can_batch(sh))
787 goto unlock_out;
788
789 if (sh->batch_head)
790 goto unlock_out;
791
792 dd_idx = 0;
793 while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
794 dd_idx++;
795 if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
796 bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
797 goto unlock_out;
798
799 if (head->batch_head) {
800 spin_lock(&head->batch_head->batch_lock);
801 /* This batch list is already running */
802 if (!stripe_can_batch(head)) {
803 spin_unlock(&head->batch_head->batch_lock);
804 goto unlock_out;
805 }
806 /*
807 * We must assign batch_head of this stripe within the
808 * batch_lock, otherwise clear_batch_ready of batch head
809 * stripe could clear BATCH_READY bit of this stripe and
810 * this stripe->batch_head doesn't get assigned, which
811 * could confuse clear_batch_ready for this stripe
812 */
813 sh->batch_head = head->batch_head;
814
815 /*
816 * at this point, head's BATCH_READY could be cleared, but we
817 * can still add the stripe to batch list
818 */
819 list_add(&sh->batch_list, &head->batch_list);
820 spin_unlock(&head->batch_head->batch_lock);
821 } else {
822 head->batch_head = head;
823 sh->batch_head = head->batch_head;
824 spin_lock(&head->batch_lock);
825 list_add_tail(&sh->batch_list, &head->batch_list);
826 spin_unlock(&head->batch_lock);
827 }
828
829 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
830 if (atomic_dec_return(&conf->preread_active_stripes)
831 < IO_THRESHOLD)
832 md_wakeup_thread(conf->mddev->thread);
833
834 if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
835 int seq = sh->bm_seq;
836 if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
837 sh->batch_head->bm_seq > seq)
838 seq = sh->batch_head->bm_seq;
839 set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
840 sh->batch_head->bm_seq = seq;
841 }
842
843 atomic_inc(&sh->count);
844unlock_out:
845 unlock_two_stripes(head, sh);
846out:
847 raid5_release_stripe(head);
848}
849
850/* Determine if 'data_offset' or 'new_data_offset' should be used
851 * in this stripe_head.
852 */
853static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
854{
855 sector_t progress = conf->reshape_progress;
856 /* Need a memory barrier to make sure we see the value
857 * of conf->generation, or ->data_offset that was set before
858 * reshape_progress was updated.
859 */
860 smp_rmb();
861 if (progress == MaxSector)
862 return 0;
863 if (sh->generation == conf->generation - 1)
864 return 0;
865 /* We are in a reshape, and this is a new-generation stripe,
866 * so use new_data_offset.
867 */
868 return 1;
869}
870
871static void dispatch_bio_list(struct bio_list *tmp)
872{
873 struct bio *bio;
874
875 while ((bio = bio_list_pop(tmp)))
876 generic_make_request(bio);
877}
878
879static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
880{
881 const struct r5pending_data *da = list_entry(a,
882 struct r5pending_data, sibling);
883 const struct r5pending_data *db = list_entry(b,
884 struct r5pending_data, sibling);
885 if (da->sector > db->sector)
886 return 1;
887 if (da->sector < db->sector)
888 return -1;
889 return 0;
890}
891
892static void dispatch_defer_bios(struct r5conf *conf, int target,
893 struct bio_list *list)
894{
895 struct r5pending_data *data;
896 struct list_head *first, *next = NULL;
897 int cnt = 0;
898
899 if (conf->pending_data_cnt == 0)
900 return;
901
902 list_sort(NULL, &conf->pending_list, cmp_stripe);
903
904 first = conf->pending_list.next;
905
906 /* temporarily move the head */
907 if (conf->next_pending_data)
908 list_move_tail(&conf->pending_list,
909 &conf->next_pending_data->sibling);
910
911 while (!list_empty(&conf->pending_list)) {
912 data = list_first_entry(&conf->pending_list,
913 struct r5pending_data, sibling);
914 if (&data->sibling == first)
915 first = data->sibling.next;
916 next = data->sibling.next;
917
918 bio_list_merge(list, &data->bios);
919 list_move(&data->sibling, &conf->free_list);
920 cnt++;
921 if (cnt >= target)
922 break;
923 }
924 conf->pending_data_cnt -= cnt;
925 BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
926
927 if (next != &conf->pending_list)
928 conf->next_pending_data = list_entry(next,
929 struct r5pending_data, sibling);
930 else
931 conf->next_pending_data = NULL;
932 /* list isn't empty */
933 if (first != &conf->pending_list)
934 list_move_tail(&conf->pending_list, first);
935}
936
937static void flush_deferred_bios(struct r5conf *conf)
938{
939 struct bio_list tmp = BIO_EMPTY_LIST;
940
941 if (conf->pending_data_cnt == 0)
942 return;
943
944 spin_lock(&conf->pending_bios_lock);
945 dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
946 BUG_ON(conf->pending_data_cnt != 0);
947 spin_unlock(&conf->pending_bios_lock);
948
949 dispatch_bio_list(&tmp);
950}
951
952static void defer_issue_bios(struct r5conf *conf, sector_t sector,
953 struct bio_list *bios)
954{
955 struct bio_list tmp = BIO_EMPTY_LIST;
956 struct r5pending_data *ent;
957
958 spin_lock(&conf->pending_bios_lock);
959 ent = list_first_entry(&conf->free_list, struct r5pending_data,
960 sibling);
961 list_move_tail(&ent->sibling, &conf->pending_list);
962 ent->sector = sector;
963 bio_list_init(&ent->bios);
964 bio_list_merge(&ent->bios, bios);
965 conf->pending_data_cnt++;
966 if (conf->pending_data_cnt >= PENDING_IO_MAX)
967 dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
968
969 spin_unlock(&conf->pending_bios_lock);
970
971 dispatch_bio_list(&tmp);
972}
973
974static void
975raid5_end_read_request(struct bio *bi);
976static void
977raid5_end_write_request(struct bio *bi);
978
979static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
980{
981 struct r5conf *conf = sh->raid_conf;
982 int i, disks = sh->disks;
983 struct stripe_head *head_sh = sh;
984 struct bio_list pending_bios = BIO_EMPTY_LIST;
985 bool should_defer;
986
987 might_sleep();
988
989 if (log_stripe(sh, s) == 0)
990 return;
991
992 should_defer = conf->batch_bio_dispatch && conf->group_cnt;
993
994 for (i = disks; i--; ) {
995 int op, op_flags = 0;
996 int replace_only = 0;
997 struct bio *bi, *rbi;
998 struct md_rdev *rdev, *rrdev = NULL;
999
1000 sh = head_sh;
1001 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1002 op = REQ_OP_WRITE;
1003 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1004 op_flags = REQ_FUA;
1005 if (test_bit(R5_Discard, &sh->dev[i].flags))
1006 op = REQ_OP_DISCARD;
1007 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1008 op = REQ_OP_READ;
1009 else if (test_and_clear_bit(R5_WantReplace,
1010 &sh->dev[i].flags)) {
1011 op = REQ_OP_WRITE;
1012 replace_only = 1;
1013 } else
1014 continue;
1015 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1016 op_flags |= REQ_SYNC;
1017
1018again:
1019 bi = &sh->dev[i].req;
1020 rbi = &sh->dev[i].rreq; /* For writing to replacement */
1021
1022 rcu_read_lock();
1023 rrdev = rcu_dereference(conf->disks[i].replacement);
1024 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1025 rdev = rcu_dereference(conf->disks[i].rdev);
1026 if (!rdev) {
1027 rdev = rrdev;
1028 rrdev = NULL;
1029 }
1030 if (op_is_write(op)) {
1031 if (replace_only)
1032 rdev = NULL;
1033 if (rdev == rrdev)
1034 /* We raced and saw duplicates */
1035 rrdev = NULL;
1036 } else {
1037 if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1038 rdev = rrdev;
1039 rrdev = NULL;
1040 }
1041
1042 if (rdev && test_bit(Faulty, &rdev->flags))
1043 rdev = NULL;
1044 if (rdev)
1045 atomic_inc(&rdev->nr_pending);
1046 if (rrdev && test_bit(Faulty, &rrdev->flags))
1047 rrdev = NULL;
1048 if (rrdev)
1049 atomic_inc(&rrdev->nr_pending);
1050 rcu_read_unlock();
1051
1052 /* We have already checked bad blocks for reads. Now
1053 * need to check for writes. We never accept write errors
1054 * on the replacement, so we don't to check rrdev.
1055 */
1056 while (op_is_write(op) && rdev &&
1057 test_bit(WriteErrorSeen, &rdev->flags)) {
1058 sector_t first_bad;
1059 int bad_sectors;
1060 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1061 &first_bad, &bad_sectors);
1062 if (!bad)
1063 break;
1064
1065 if (bad < 0) {
1066 set_bit(BlockedBadBlocks, &rdev->flags);
1067 if (!conf->mddev->external &&
1068 conf->mddev->sb_flags) {
1069 /* It is very unlikely, but we might
1070 * still need to write out the
1071 * bad block log - better give it
1072 * a chance*/
1073 md_check_recovery(conf->mddev);
1074 }
1075 /*
1076 * Because md_wait_for_blocked_rdev
1077 * will dec nr_pending, we must
1078 * increment it first.
1079 */
1080 atomic_inc(&rdev->nr_pending);
1081 md_wait_for_blocked_rdev(rdev, conf->mddev);
1082 } else {
1083 /* Acknowledged bad block - skip the write */
1084 rdev_dec_pending(rdev, conf->mddev);
1085 rdev = NULL;
1086 }
1087 }
1088
1089 if (rdev) {
1090 if (s->syncing || s->expanding || s->expanded
1091 || s->replacing)
1092 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1093
1094 set_bit(STRIPE_IO_STARTED, &sh->state);
1095
1096 bio_set_dev(bi, rdev->bdev);
1097 bio_set_op_attrs(bi, op, op_flags);
1098 bi->bi_end_io = op_is_write(op)
1099 ? raid5_end_write_request
1100 : raid5_end_read_request;
1101 bi->bi_private = sh;
1102
1103 pr_debug("%s: for %llu schedule op %d on disc %d\n",
1104 __func__, (unsigned long long)sh->sector,
1105 bi->bi_opf, i);
1106 atomic_inc(&sh->count);
1107 if (sh != head_sh)
1108 atomic_inc(&head_sh->count);
1109 if (use_new_offset(conf, sh))
1110 bi->bi_iter.bi_sector = (sh->sector
1111 + rdev->new_data_offset);
1112 else
1113 bi->bi_iter.bi_sector = (sh->sector
1114 + rdev->data_offset);
1115 if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1116 bi->bi_opf |= REQ_NOMERGE;
1117
1118 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1119 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1120
1121 if (!op_is_write(op) &&
1122 test_bit(R5_InJournal, &sh->dev[i].flags))
1123 /*
1124 * issuing read for a page in journal, this
1125 * must be preparing for prexor in rmw; read
1126 * the data into orig_page
1127 */
1128 sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1129 else
1130 sh->dev[i].vec.bv_page = sh->dev[i].page;
1131 bi->bi_vcnt = 1;
1132 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1133 bi->bi_io_vec[0].bv_offset = 0;
1134 bi->bi_iter.bi_size = STRIPE_SIZE;
1135 bi->bi_write_hint = sh->dev[i].write_hint;
1136 if (!rrdev)
1137 sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1138 /*
1139 * If this is discard request, set bi_vcnt 0. We don't
1140 * want to confuse SCSI because SCSI will replace payload
1141 */
1142 if (op == REQ_OP_DISCARD)
1143 bi->bi_vcnt = 0;
1144 if (rrdev)
1145 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1146
1147 if (conf->mddev->gendisk)
1148 trace_block_bio_remap(bi->bi_disk->queue,
1149 bi, disk_devt(conf->mddev->gendisk),
1150 sh->dev[i].sector);
1151 if (should_defer && op_is_write(op))
1152 bio_list_add(&pending_bios, bi);
1153 else
1154 generic_make_request(bi);
1155 }
1156 if (rrdev) {
1157 if (s->syncing || s->expanding || s->expanded
1158 || s->replacing)
1159 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1160
1161 set_bit(STRIPE_IO_STARTED, &sh->state);
1162
1163 bio_set_dev(rbi, rrdev->bdev);
1164 bio_set_op_attrs(rbi, op, op_flags);
1165 BUG_ON(!op_is_write(op));
1166 rbi->bi_end_io = raid5_end_write_request;
1167 rbi->bi_private = sh;
1168
1169 pr_debug("%s: for %llu schedule op %d on "
1170 "replacement disc %d\n",
1171 __func__, (unsigned long long)sh->sector,
1172 rbi->bi_opf, i);
1173 atomic_inc(&sh->count);
1174 if (sh != head_sh)
1175 atomic_inc(&head_sh->count);
1176 if (use_new_offset(conf, sh))
1177 rbi->bi_iter.bi_sector = (sh->sector
1178 + rrdev->new_data_offset);
1179 else
1180 rbi->bi_iter.bi_sector = (sh->sector
1181 + rrdev->data_offset);
1182 if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1183 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1184 sh->dev[i].rvec.bv_page = sh->dev[i].page;
1185 rbi->bi_vcnt = 1;
1186 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1187 rbi->bi_io_vec[0].bv_offset = 0;
1188 rbi->bi_iter.bi_size = STRIPE_SIZE;
1189 rbi->bi_write_hint = sh->dev[i].write_hint;
1190 sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1191 /*
1192 * If this is discard request, set bi_vcnt 0. We don't
1193 * want to confuse SCSI because SCSI will replace payload
1194 */
1195 if (op == REQ_OP_DISCARD)
1196 rbi->bi_vcnt = 0;
1197 if (conf->mddev->gendisk)
1198 trace_block_bio_remap(rbi->bi_disk->queue,
1199 rbi, disk_devt(conf->mddev->gendisk),
1200 sh->dev[i].sector);
1201 if (should_defer && op_is_write(op))
1202 bio_list_add(&pending_bios, rbi);
1203 else
1204 generic_make_request(rbi);
1205 }
1206 if (!rdev && !rrdev) {
1207 if (op_is_write(op))
1208 set_bit(STRIPE_DEGRADED, &sh->state);
1209 pr_debug("skip op %d on disc %d for sector %llu\n",
1210 bi->bi_opf, i, (unsigned long long)sh->sector);
1211 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1212 set_bit(STRIPE_HANDLE, &sh->state);
1213 }
1214
1215 if (!head_sh->batch_head)
1216 continue;
1217 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1218 batch_list);
1219 if (sh != head_sh)
1220 goto again;
1221 }
1222
1223 if (should_defer && !bio_list_empty(&pending_bios))
1224 defer_issue_bios(conf, head_sh->sector, &pending_bios);
1225}
1226
1227static struct dma_async_tx_descriptor *
1228async_copy_data(int frombio, struct bio *bio, struct page **page,
1229 sector_t sector, struct dma_async_tx_descriptor *tx,
1230 struct stripe_head *sh, int no_skipcopy)
1231{
1232 struct bio_vec bvl;
1233 struct bvec_iter iter;
1234 struct page *bio_page;
1235 int page_offset;
1236 struct async_submit_ctl submit;
1237 enum async_tx_flags flags = 0;
1238
1239 if (bio->bi_iter.bi_sector >= sector)
1240 page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1241 else
1242 page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1243
1244 if (frombio)
1245 flags |= ASYNC_TX_FENCE;
1246 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1247
1248 bio_for_each_segment(bvl, bio, iter) {
1249 int len = bvl.bv_len;
1250 int clen;
1251 int b_offset = 0;
1252
1253 if (page_offset < 0) {
1254 b_offset = -page_offset;
1255 page_offset += b_offset;
1256 len -= b_offset;
1257 }
1258
1259 if (len > 0 && page_offset + len > STRIPE_SIZE)
1260 clen = STRIPE_SIZE - page_offset;
1261 else
1262 clen = len;
1263
1264 if (clen > 0) {
1265 b_offset += bvl.bv_offset;
1266 bio_page = bvl.bv_page;
1267 if (frombio) {
1268 if (sh->raid_conf->skip_copy &&
1269 b_offset == 0 && page_offset == 0 &&
1270 clen == STRIPE_SIZE &&
1271 !no_skipcopy)
1272 *page = bio_page;
1273 else
1274 tx = async_memcpy(*page, bio_page, page_offset,
1275 b_offset, clen, &submit);
1276 } else
1277 tx = async_memcpy(bio_page, *page, b_offset,
1278 page_offset, clen, &submit);
1279 }
1280 /* chain the operations */
1281 submit.depend_tx = tx;
1282
1283 if (clen < len) /* hit end of page */
1284 break;
1285 page_offset += len;
1286 }
1287
1288 return tx;
1289}
1290
1291static void ops_complete_biofill(void *stripe_head_ref)
1292{
1293 struct stripe_head *sh = stripe_head_ref;
1294 int i;
1295
1296 pr_debug("%s: stripe %llu\n", __func__,
1297 (unsigned long long)sh->sector);
1298
1299 /* clear completed biofills */
1300 for (i = sh->disks; i--; ) {
1301 struct r5dev *dev = &sh->dev[i];
1302
1303 /* acknowledge completion of a biofill operation */
1304 /* and check if we need to reply to a read request,
1305 * new R5_Wantfill requests are held off until
1306 * !STRIPE_BIOFILL_RUN
1307 */
1308 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1309 struct bio *rbi, *rbi2;
1310
1311 BUG_ON(!dev->read);
1312 rbi = dev->read;
1313 dev->read = NULL;
1314 while (rbi && rbi->bi_iter.bi_sector <
1315 dev->sector + STRIPE_SECTORS) {
1316 rbi2 = r5_next_bio(rbi, dev->sector);
1317 bio_endio(rbi);
1318 rbi = rbi2;
1319 }
1320 }
1321 }
1322 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1323
1324 set_bit(STRIPE_HANDLE, &sh->state);
1325 raid5_release_stripe(sh);
1326}
1327
1328static void ops_run_biofill(struct stripe_head *sh)
1329{
1330 struct dma_async_tx_descriptor *tx = NULL;
1331 struct async_submit_ctl submit;
1332 int i;
1333
1334 BUG_ON(sh->batch_head);
1335 pr_debug("%s: stripe %llu\n", __func__,
1336 (unsigned long long)sh->sector);
1337
1338 for (i = sh->disks; i--; ) {
1339 struct r5dev *dev = &sh->dev[i];
1340 if (test_bit(R5_Wantfill, &dev->flags)) {
1341 struct bio *rbi;
1342 spin_lock_irq(&sh->stripe_lock);
1343 dev->read = rbi = dev->toread;
1344 dev->toread = NULL;
1345 spin_unlock_irq(&sh->stripe_lock);
1346 while (rbi && rbi->bi_iter.bi_sector <
1347 dev->sector + STRIPE_SECTORS) {
1348 tx = async_copy_data(0, rbi, &dev->page,
1349 dev->sector, tx, sh, 0);
1350 rbi = r5_next_bio(rbi, dev->sector);
1351 }
1352 }
1353 }
1354
1355 atomic_inc(&sh->count);
1356 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1357 async_trigger_callback(&submit);
1358}
1359
1360static void mark_target_uptodate(struct stripe_head *sh, int target)
1361{
1362 struct r5dev *tgt;
1363
1364 if (target < 0)
1365 return;
1366
1367 tgt = &sh->dev[target];
1368 set_bit(R5_UPTODATE, &tgt->flags);
1369 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1370 clear_bit(R5_Wantcompute, &tgt->flags);
1371}
1372
1373static void ops_complete_compute(void *stripe_head_ref)
1374{
1375 struct stripe_head *sh = stripe_head_ref;
1376
1377 pr_debug("%s: stripe %llu\n", __func__,
1378 (unsigned long long)sh->sector);
1379
1380 /* mark the computed target(s) as uptodate */
1381 mark_target_uptodate(sh, sh->ops.target);
1382 mark_target_uptodate(sh, sh->ops.target2);
1383
1384 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1385 if (sh->check_state == check_state_compute_run)
1386 sh->check_state = check_state_compute_result;
1387 set_bit(STRIPE_HANDLE, &sh->state);
1388 raid5_release_stripe(sh);
1389}
1390
1391/* return a pointer to the address conversion region of the scribble buffer */
1392static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1393{
1394 return percpu->scribble + i * percpu->scribble_obj_size;
1395}
1396
1397/* return a pointer to the address conversion region of the scribble buffer */
1398static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1399 struct raid5_percpu *percpu, int i)
1400{
1401 return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1402}
1403
1404static struct dma_async_tx_descriptor *
1405ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1406{
1407 int disks = sh->disks;
1408 struct page **xor_srcs = to_addr_page(percpu, 0);
1409 int target = sh->ops.target;
1410 struct r5dev *tgt = &sh->dev[target];
1411 struct page *xor_dest = tgt->page;
1412 int count = 0;
1413 struct dma_async_tx_descriptor *tx;
1414 struct async_submit_ctl submit;
1415 int i;
1416
1417 BUG_ON(sh->batch_head);
1418
1419 pr_debug("%s: stripe %llu block: %d\n",
1420 __func__, (unsigned long long)sh->sector, target);
1421 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1422
1423 for (i = disks; i--; )
1424 if (i != target)
1425 xor_srcs[count++] = sh->dev[i].page;
1426
1427 atomic_inc(&sh->count);
1428
1429 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1430 ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1431 if (unlikely(count == 1))
1432 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1433 else
1434 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1435
1436 return tx;
1437}
1438
1439/* set_syndrome_sources - populate source buffers for gen_syndrome
1440 * @srcs - (struct page *) array of size sh->disks
1441 * @sh - stripe_head to parse
1442 *
1443 * Populates srcs in proper layout order for the stripe and returns the
1444 * 'count' of sources to be used in a call to async_gen_syndrome. The P
1445 * destination buffer is recorded in srcs[count] and the Q destination
1446 * is recorded in srcs[count+1]].
1447 */
1448static int set_syndrome_sources(struct page **srcs,
1449 struct stripe_head *sh,
1450 int srctype)
1451{
1452 int disks = sh->disks;
1453 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1454 int d0_idx = raid6_d0(sh);
1455 int count;
1456 int i;
1457
1458 for (i = 0; i < disks; i++)
1459 srcs[i] = NULL;
1460
1461 count = 0;
1462 i = d0_idx;
1463 do {
1464 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1465 struct r5dev *dev = &sh->dev[i];
1466
1467 if (i == sh->qd_idx || i == sh->pd_idx ||
1468 (srctype == SYNDROME_SRC_ALL) ||
1469 (srctype == SYNDROME_SRC_WANT_DRAIN &&
1470 (test_bit(R5_Wantdrain, &dev->flags) ||
1471 test_bit(R5_InJournal, &dev->flags))) ||
1472 (srctype == SYNDROME_SRC_WRITTEN &&
1473 (dev->written ||
1474 test_bit(R5_InJournal, &dev->flags)))) {
1475 if (test_bit(R5_InJournal, &dev->flags))
1476 srcs[slot] = sh->dev[i].orig_page;
1477 else
1478 srcs[slot] = sh->dev[i].page;
1479 }
1480 i = raid6_next_disk(i, disks);
1481 } while (i != d0_idx);
1482
1483 return syndrome_disks;
1484}
1485
1486static struct dma_async_tx_descriptor *
1487ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1488{
1489 int disks = sh->disks;
1490 struct page **blocks = to_addr_page(percpu, 0);
1491 int target;
1492 int qd_idx = sh->qd_idx;
1493 struct dma_async_tx_descriptor *tx;
1494 struct async_submit_ctl submit;
1495 struct r5dev *tgt;
1496 struct page *dest;
1497 int i;
1498 int count;
1499
1500 BUG_ON(sh->batch_head);
1501 if (sh->ops.target < 0)
1502 target = sh->ops.target2;
1503 else if (sh->ops.target2 < 0)
1504 target = sh->ops.target;
1505 else
1506 /* we should only have one valid target */
1507 BUG();
1508 BUG_ON(target < 0);
1509 pr_debug("%s: stripe %llu block: %d\n",
1510 __func__, (unsigned long long)sh->sector, target);
1511
1512 tgt = &sh->dev[target];
1513 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1514 dest = tgt->page;
1515
1516 atomic_inc(&sh->count);
1517
1518 if (target == qd_idx) {
1519 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1520 blocks[count] = NULL; /* regenerating p is not necessary */
1521 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1522 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1523 ops_complete_compute, sh,
1524 to_addr_conv(sh, percpu, 0));
1525 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1526 } else {
1527 /* Compute any data- or p-drive using XOR */
1528 count = 0;
1529 for (i = disks; i-- ; ) {
1530 if (i == target || i == qd_idx)
1531 continue;
1532 blocks[count++] = sh->dev[i].page;
1533 }
1534
1535 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1536 NULL, ops_complete_compute, sh,
1537 to_addr_conv(sh, percpu, 0));
1538 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1539 }
1540
1541 return tx;
1542}
1543
1544static struct dma_async_tx_descriptor *
1545ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1546{
1547 int i, count, disks = sh->disks;
1548 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1549 int d0_idx = raid6_d0(sh);
1550 int faila = -1, failb = -1;
1551 int target = sh->ops.target;
1552 int target2 = sh->ops.target2;
1553 struct r5dev *tgt = &sh->dev[target];
1554 struct r5dev *tgt2 = &sh->dev[target2];
1555 struct dma_async_tx_descriptor *tx;
1556 struct page **blocks = to_addr_page(percpu, 0);
1557 struct async_submit_ctl submit;
1558
1559 BUG_ON(sh->batch_head);
1560 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1561 __func__, (unsigned long long)sh->sector, target, target2);
1562 BUG_ON(target < 0 || target2 < 0);
1563 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1564 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1565
1566 /* we need to open-code set_syndrome_sources to handle the
1567 * slot number conversion for 'faila' and 'failb'
1568 */
1569 for (i = 0; i < disks ; i++)
1570 blocks[i] = NULL;
1571 count = 0;
1572 i = d0_idx;
1573 do {
1574 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1575
1576 blocks[slot] = sh->dev[i].page;
1577
1578 if (i == target)
1579 faila = slot;
1580 if (i == target2)
1581 failb = slot;
1582 i = raid6_next_disk(i, disks);
1583 } while (i != d0_idx);
1584
1585 BUG_ON(faila == failb);
1586 if (failb < faila)
1587 swap(faila, failb);
1588 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1589 __func__, (unsigned long long)sh->sector, faila, failb);
1590
1591 atomic_inc(&sh->count);
1592
1593 if (failb == syndrome_disks+1) {
1594 /* Q disk is one of the missing disks */
1595 if (faila == syndrome_disks) {
1596 /* Missing P+Q, just recompute */
1597 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1598 ops_complete_compute, sh,
1599 to_addr_conv(sh, percpu, 0));
1600 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1601 STRIPE_SIZE, &submit);
1602 } else {
1603 struct page *dest;
1604 int data_target;
1605 int qd_idx = sh->qd_idx;
1606
1607 /* Missing D+Q: recompute D from P, then recompute Q */
1608 if (target == qd_idx)
1609 data_target = target2;
1610 else
1611 data_target = target;
1612
1613 count = 0;
1614 for (i = disks; i-- ; ) {
1615 if (i == data_target || i == qd_idx)
1616 continue;
1617 blocks[count++] = sh->dev[i].page;
1618 }
1619 dest = sh->dev[data_target].page;
1620 init_async_submit(&submit,
1621 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1622 NULL, NULL, NULL,
1623 to_addr_conv(sh, percpu, 0));
1624 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1625 &submit);
1626
1627 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1628 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1629 ops_complete_compute, sh,
1630 to_addr_conv(sh, percpu, 0));
1631 return async_gen_syndrome(blocks, 0, count+2,
1632 STRIPE_SIZE, &submit);
1633 }
1634 } else {
1635 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1636 ops_complete_compute, sh,
1637 to_addr_conv(sh, percpu, 0));
1638 if (failb == syndrome_disks) {
1639 /* We're missing D+P. */
1640 return async_raid6_datap_recov(syndrome_disks+2,
1641 STRIPE_SIZE, faila,
1642 blocks, &submit);
1643 } else {
1644 /* We're missing D+D. */
1645 return async_raid6_2data_recov(syndrome_disks+2,
1646 STRIPE_SIZE, faila, failb,
1647 blocks, &submit);
1648 }
1649 }
1650}
1651
1652static void ops_complete_prexor(void *stripe_head_ref)
1653{
1654 struct stripe_head *sh = stripe_head_ref;
1655
1656 pr_debug("%s: stripe %llu\n", __func__,
1657 (unsigned long long)sh->sector);
1658
1659 if (r5c_is_writeback(sh->raid_conf->log))
1660 /*
1661 * raid5-cache write back uses orig_page during prexor.
1662 * After prexor, it is time to free orig_page
1663 */
1664 r5c_release_extra_page(sh);
1665}
1666
1667static struct dma_async_tx_descriptor *
1668ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1669 struct dma_async_tx_descriptor *tx)
1670{
1671 int disks = sh->disks;
1672 struct page **xor_srcs = to_addr_page(percpu, 0);
1673 int count = 0, pd_idx = sh->pd_idx, i;
1674 struct async_submit_ctl submit;
1675
1676 /* existing parity data subtracted */
1677 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1678
1679 BUG_ON(sh->batch_head);
1680 pr_debug("%s: stripe %llu\n", __func__,
1681 (unsigned long long)sh->sector);
1682
1683 for (i = disks; i--; ) {
1684 struct r5dev *dev = &sh->dev[i];
1685 /* Only process blocks that are known to be uptodate */
1686 if (test_bit(R5_InJournal, &dev->flags))
1687 xor_srcs[count++] = dev->orig_page;
1688 else if (test_bit(R5_Wantdrain, &dev->flags))
1689 xor_srcs[count++] = dev->page;
1690 }
1691
1692 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1693 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1694 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1695
1696 return tx;
1697}
1698
1699static struct dma_async_tx_descriptor *
1700ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1701 struct dma_async_tx_descriptor *tx)
1702{
1703 struct page **blocks = to_addr_page(percpu, 0);
1704 int count;
1705 struct async_submit_ctl submit;
1706
1707 pr_debug("%s: stripe %llu\n", __func__,
1708 (unsigned long long)sh->sector);
1709
1710 count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1711
1712 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1713 ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1714 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1715
1716 return tx;
1717}
1718
1719static struct dma_async_tx_descriptor *
1720ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1721{
1722 struct r5conf *conf = sh->raid_conf;
1723 int disks = sh->disks;
1724 int i;
1725 struct stripe_head *head_sh = sh;
1726
1727 pr_debug("%s: stripe %llu\n", __func__,
1728 (unsigned long long)sh->sector);
1729
1730 for (i = disks; i--; ) {
1731 struct r5dev *dev;
1732 struct bio *chosen;
1733
1734 sh = head_sh;
1735 if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1736 struct bio *wbi;
1737
1738again:
1739 dev = &sh->dev[i];
1740 /*
1741 * clear R5_InJournal, so when rewriting a page in
1742 * journal, it is not skipped by r5l_log_stripe()
1743 */
1744 clear_bit(R5_InJournal, &dev->flags);
1745 spin_lock_irq(&sh->stripe_lock);
1746 chosen = dev->towrite;
1747 dev->towrite = NULL;
1748 sh->overwrite_disks = 0;
1749 BUG_ON(dev->written);
1750 wbi = dev->written = chosen;
1751 spin_unlock_irq(&sh->stripe_lock);
1752 WARN_ON(dev->page != dev->orig_page);
1753
1754 while (wbi && wbi->bi_iter.bi_sector <
1755 dev->sector + STRIPE_SECTORS) {
1756 if (wbi->bi_opf & REQ_FUA)
1757 set_bit(R5_WantFUA, &dev->flags);
1758 if (wbi->bi_opf & REQ_SYNC)
1759 set_bit(R5_SyncIO, &dev->flags);
1760 if (bio_op(wbi) == REQ_OP_DISCARD)
1761 set_bit(R5_Discard, &dev->flags);
1762 else {
1763 tx = async_copy_data(1, wbi, &dev->page,
1764 dev->sector, tx, sh,
1765 r5c_is_writeback(conf->log));
1766 if (dev->page != dev->orig_page &&
1767 !r5c_is_writeback(conf->log)) {
1768 set_bit(R5_SkipCopy, &dev->flags);
1769 clear_bit(R5_UPTODATE, &dev->flags);
1770 clear_bit(R5_OVERWRITE, &dev->flags);
1771 }
1772 }
1773 wbi = r5_next_bio(wbi, dev->sector);
1774 }
1775
1776 if (head_sh->batch_head) {
1777 sh = list_first_entry(&sh->batch_list,
1778 struct stripe_head,
1779 batch_list);
1780 if (sh == head_sh)
1781 continue;
1782 goto again;
1783 }
1784 }
1785 }
1786
1787 return tx;
1788}
1789
1790static void ops_complete_reconstruct(void *stripe_head_ref)
1791{
1792 struct stripe_head *sh = stripe_head_ref;
1793 int disks = sh->disks;
1794 int pd_idx = sh->pd_idx;
1795 int qd_idx = sh->qd_idx;
1796 int i;
1797 bool fua = false, sync = false, discard = false;
1798
1799 pr_debug("%s: stripe %llu\n", __func__,
1800 (unsigned long long)sh->sector);
1801
1802 for (i = disks; i--; ) {
1803 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1804 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1805 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1806 }
1807
1808 for (i = disks; i--; ) {
1809 struct r5dev *dev = &sh->dev[i];
1810
1811 if (dev->written || i == pd_idx || i == qd_idx) {
1812 if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1813 set_bit(R5_UPTODATE, &dev->flags);
1814 if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1815 set_bit(R5_Expanded, &dev->flags);
1816 }
1817 if (fua)
1818 set_bit(R5_WantFUA, &dev->flags);
1819 if (sync)
1820 set_bit(R5_SyncIO, &dev->flags);
1821 }
1822 }
1823
1824 if (sh->reconstruct_state == reconstruct_state_drain_run)
1825 sh->reconstruct_state = reconstruct_state_drain_result;
1826 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1827 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1828 else {
1829 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1830 sh->reconstruct_state = reconstruct_state_result;
1831 }
1832
1833 set_bit(STRIPE_HANDLE, &sh->state);
1834 raid5_release_stripe(sh);
1835}
1836
1837static void
1838ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1839 struct dma_async_tx_descriptor *tx)
1840{
1841 int disks = sh->disks;
1842 struct page **xor_srcs;
1843 struct async_submit_ctl submit;
1844 int count, pd_idx = sh->pd_idx, i;
1845 struct page *xor_dest;
1846 int prexor = 0;
1847 unsigned long flags;
1848 int j = 0;
1849 struct stripe_head *head_sh = sh;
1850 int last_stripe;
1851
1852 pr_debug("%s: stripe %llu\n", __func__,
1853 (unsigned long long)sh->sector);
1854
1855 for (i = 0; i < sh->disks; i++) {
1856 if (pd_idx == i)
1857 continue;
1858 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1859 break;
1860 }
1861 if (i >= sh->disks) {
1862 atomic_inc(&sh->count);
1863 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1864 ops_complete_reconstruct(sh);
1865 return;
1866 }
1867again:
1868 count = 0;
1869 xor_srcs = to_addr_page(percpu, j);
1870 /* check if prexor is active which means only process blocks
1871 * that are part of a read-modify-write (written)
1872 */
1873 if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1874 prexor = 1;
1875 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1876 for (i = disks; i--; ) {
1877 struct r5dev *dev = &sh->dev[i];
1878 if (head_sh->dev[i].written ||
1879 test_bit(R5_InJournal, &head_sh->dev[i].flags))
1880 xor_srcs[count++] = dev->page;
1881 }
1882 } else {
1883 xor_dest = sh->dev[pd_idx].page;
1884 for (i = disks; i--; ) {
1885 struct r5dev *dev = &sh->dev[i];
1886 if (i != pd_idx)
1887 xor_srcs[count++] = dev->page;
1888 }
1889 }
1890
1891 /* 1/ if we prexor'd then the dest is reused as a source
1892 * 2/ if we did not prexor then we are redoing the parity
1893 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1894 * for the synchronous xor case
1895 */
1896 last_stripe = !head_sh->batch_head ||
1897 list_first_entry(&sh->batch_list,
1898 struct stripe_head, batch_list) == head_sh;
1899 if (last_stripe) {
1900 flags = ASYNC_TX_ACK |
1901 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1902
1903 atomic_inc(&head_sh->count);
1904 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1905 to_addr_conv(sh, percpu, j));
1906 } else {
1907 flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1908 init_async_submit(&submit, flags, tx, NULL, NULL,
1909 to_addr_conv(sh, percpu, j));
1910 }
1911
1912 if (unlikely(count == 1))
1913 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1914 else
1915 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1916 if (!last_stripe) {
1917 j++;
1918 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1919 batch_list);
1920 goto again;
1921 }
1922}
1923
1924static void
1925ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1926 struct dma_async_tx_descriptor *tx)
1927{
1928 struct async_submit_ctl submit;
1929 struct page **blocks;
1930 int count, i, j = 0;
1931 struct stripe_head *head_sh = sh;
1932 int last_stripe;
1933 int synflags;
1934 unsigned long txflags;
1935
1936 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1937
1938 for (i = 0; i < sh->disks; i++) {
1939 if (sh->pd_idx == i || sh->qd_idx == i)
1940 continue;
1941 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1942 break;
1943 }
1944 if (i >= sh->disks) {
1945 atomic_inc(&sh->count);
1946 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1947 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1948 ops_complete_reconstruct(sh);
1949 return;
1950 }
1951
1952again:
1953 blocks = to_addr_page(percpu, j);
1954
1955 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1956 synflags = SYNDROME_SRC_WRITTEN;
1957 txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1958 } else {
1959 synflags = SYNDROME_SRC_ALL;
1960 txflags = ASYNC_TX_ACK;
1961 }
1962
1963 count = set_syndrome_sources(blocks, sh, synflags);
1964 last_stripe = !head_sh->batch_head ||
1965 list_first_entry(&sh->batch_list,
1966 struct stripe_head, batch_list) == head_sh;
1967
1968 if (last_stripe) {
1969 atomic_inc(&head_sh->count);
1970 init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1971 head_sh, to_addr_conv(sh, percpu, j));
1972 } else
1973 init_async_submit(&submit, 0, tx, NULL, NULL,
1974 to_addr_conv(sh, percpu, j));
1975 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1976 if (!last_stripe) {
1977 j++;
1978 sh = list_first_entry(&sh->batch_list, struct stripe_head,
1979 batch_list);
1980 goto again;
1981 }
1982}
1983
1984static void ops_complete_check(void *stripe_head_ref)
1985{
1986 struct stripe_head *sh = stripe_head_ref;
1987
1988 pr_debug("%s: stripe %llu\n", __func__,
1989 (unsigned long long)sh->sector);
1990
1991 sh->check_state = check_state_check_result;
1992 set_bit(STRIPE_HANDLE, &sh->state);
1993 raid5_release_stripe(sh);
1994}
1995
1996static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1997{
1998 int disks = sh->disks;
1999 int pd_idx = sh->pd_idx;
2000 int qd_idx = sh->qd_idx;
2001 struct page *xor_dest;
2002 struct page **xor_srcs = to_addr_page(percpu, 0);
2003 struct dma_async_tx_descriptor *tx;
2004 struct async_submit_ctl submit;
2005 int count;
2006 int i;
2007
2008 pr_debug("%s: stripe %llu\n", __func__,
2009 (unsigned long long)sh->sector);
2010
2011 BUG_ON(sh->batch_head);
2012 count = 0;
2013 xor_dest = sh->dev[pd_idx].page;
2014 xor_srcs[count++] = xor_dest;
2015 for (i = disks; i--; ) {
2016 if (i == pd_idx || i == qd_idx)
2017 continue;
2018 xor_srcs[count++] = sh->dev[i].page;
2019 }
2020
2021 init_async_submit(&submit, 0, NULL, NULL, NULL,
2022 to_addr_conv(sh, percpu, 0));
2023 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2024 &sh->ops.zero_sum_result, &submit);
2025
2026 atomic_inc(&sh->count);
2027 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2028 tx = async_trigger_callback(&submit);
2029}
2030
2031static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2032{
2033 struct page **srcs = to_addr_page(percpu, 0);
2034 struct async_submit_ctl submit;
2035 int count;
2036
2037 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2038 (unsigned long long)sh->sector, checkp);
2039
2040 BUG_ON(sh->batch_head);
2041 count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2042 if (!checkp)
2043 srcs[count] = NULL;
2044
2045 atomic_inc(&sh->count);
2046 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2047 sh, to_addr_conv(sh, percpu, 0));
2048 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2049 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2050}
2051
2052static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2053{
2054 int overlap_clear = 0, i, disks = sh->disks;
2055 struct dma_async_tx_descriptor *tx = NULL;
2056 struct r5conf *conf = sh->raid_conf;
2057 int level = conf->level;
2058 struct raid5_percpu *percpu;
2059 unsigned long cpu;
2060
2061 cpu = get_cpu();
2062 percpu = per_cpu_ptr(conf->percpu, cpu);
2063 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2064 ops_run_biofill(sh);
2065 overlap_clear++;
2066 }
2067
2068 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2069 if (level < 6)
2070 tx = ops_run_compute5(sh, percpu);
2071 else {
2072 if (sh->ops.target2 < 0 || sh->ops.target < 0)
2073 tx = ops_run_compute6_1(sh, percpu);
2074 else
2075 tx = ops_run_compute6_2(sh, percpu);
2076 }
2077 /* terminate the chain if reconstruct is not set to be run */
2078 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2079 async_tx_ack(tx);
2080 }
2081
2082 if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2083 if (level < 6)
2084 tx = ops_run_prexor5(sh, percpu, tx);
2085 else
2086 tx = ops_run_prexor6(sh, percpu, tx);
2087 }
2088
2089 if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2090 tx = ops_run_partial_parity(sh, percpu, tx);
2091
2092 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2093 tx = ops_run_biodrain(sh, tx);
2094 overlap_clear++;
2095 }
2096
2097 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2098 if (level < 6)
2099 ops_run_reconstruct5(sh, percpu, tx);
2100 else
2101 ops_run_reconstruct6(sh, percpu, tx);
2102 }
2103
2104 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2105 if (sh->check_state == check_state_run)
2106 ops_run_check_p(sh, percpu);
2107 else if (sh->check_state == check_state_run_q)
2108 ops_run_check_pq(sh, percpu, 0);
2109 else if (sh->check_state == check_state_run_pq)
2110 ops_run_check_pq(sh, percpu, 1);
2111 else
2112 BUG();
2113 }
2114
2115 if (overlap_clear && !sh->batch_head)
2116 for (i = disks; i--; ) {
2117 struct r5dev *dev = &sh->dev[i];
2118 if (test_and_clear_bit(R5_Overlap, &dev->flags))
2119 wake_up(&sh->raid_conf->wait_for_overlap);
2120 }
2121 put_cpu();
2122}
2123
2124static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2125{
2126 if (sh->ppl_page)
2127 __free_page(sh->ppl_page);
2128 kmem_cache_free(sc, sh);
2129}
2130
2131static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2132 int disks, struct r5conf *conf)
2133{
2134 struct stripe_head *sh;
2135 int i;
2136
2137 sh = kmem_cache_zalloc(sc, gfp);
2138 if (sh) {
2139 spin_lock_init(&sh->stripe_lock);
2140 spin_lock_init(&sh->batch_lock);
2141 INIT_LIST_HEAD(&sh->batch_list);
2142 INIT_LIST_HEAD(&sh->lru);
2143 INIT_LIST_HEAD(&sh->r5c);
2144 INIT_LIST_HEAD(&sh->log_list);
2145 atomic_set(&sh->count, 1);
2146 sh->raid_conf = conf;
2147 sh->log_start = MaxSector;
2148 for (i = 0; i < disks; i++) {
2149 struct r5dev *dev = &sh->dev[i];
2150
2151 bio_init(&dev->req, &dev->vec, 1);
2152 bio_init(&dev->rreq, &dev->rvec, 1);
2153 }
2154
2155 if (raid5_has_ppl(conf)) {
2156 sh->ppl_page = alloc_page(gfp);
2157 if (!sh->ppl_page) {
2158 free_stripe(sc, sh);
2159 sh = NULL;
2160 }
2161 }
2162 }
2163 return sh;
2164}
2165static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2166{
2167 struct stripe_head *sh;
2168
2169 sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2170 if (!sh)
2171 return 0;
2172
2173 if (grow_buffers(sh, gfp)) {
2174 shrink_buffers(sh);
2175 free_stripe(conf->slab_cache, sh);
2176 return 0;
2177 }
2178 sh->hash_lock_index =
2179 conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2180 /* we just created an active stripe so... */
2181 atomic_inc(&conf->active_stripes);
2182
2183 raid5_release_stripe(sh);
2184 conf->max_nr_stripes++;
2185 return 1;
2186}
2187
2188static int grow_stripes(struct r5conf *conf, int num)
2189{
2190 struct kmem_cache *sc;
2191 size_t namelen = sizeof(conf->cache_name[0]);
2192 int devs = max(conf->raid_disks, conf->previous_raid_disks);
2193
2194 if (conf->mddev->gendisk)
2195 snprintf(conf->cache_name[0], namelen,
2196 "raid%d-%s", conf->level, mdname(conf->mddev));
2197 else
2198 snprintf(conf->cache_name[0], namelen,
2199 "raid%d-%p", conf->level, conf->mddev);
2200 snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2201
2202 conf->active_name = 0;
2203 sc = kmem_cache_create(conf->cache_name[conf->active_name],
2204 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2205 0, 0, NULL);
2206 if (!sc)
2207 return 1;
2208 conf->slab_cache = sc;
2209 conf->pool_size = devs;
2210 while (num--)
2211 if (!grow_one_stripe(conf, GFP_KERNEL))
2212 return 1;
2213
2214 return 0;
2215}
2216
2217/**
2218 * scribble_len - return the required size of the scribble region
2219 * @num - total number of disks in the array
2220 *
2221 * The size must be enough to contain:
2222 * 1/ a struct page pointer for each device in the array +2
2223 * 2/ room to convert each entry in (1) to its corresponding dma
2224 * (dma_map_page()) or page (page_address()) address.
2225 *
2226 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2227 * calculate over all devices (not just the data blocks), using zeros in place
2228 * of the P and Q blocks.
2229 */
2230static int scribble_alloc(struct raid5_percpu *percpu,
2231 int num, int cnt, gfp_t flags)
2232{
2233 size_t obj_size =
2234 sizeof(struct page *) * (num+2) +
2235 sizeof(addr_conv_t) * (num+2);
2236 void *scribble;
2237
2238 scribble = kvmalloc_array(cnt, obj_size, flags);
2239 if (!scribble)
2240 return -ENOMEM;
2241
2242 kvfree(percpu->scribble);
2243
2244 percpu->scribble = scribble;
2245 percpu->scribble_obj_size = obj_size;
2246 return 0;
2247}
2248
2249static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2250{
2251 unsigned long cpu;
2252 int err = 0;
2253
2254 /*
2255 * Never shrink. And mddev_suspend() could deadlock if this is called
2256 * from raid5d. In that case, scribble_disks and scribble_sectors
2257 * should equal to new_disks and new_sectors
2258 */
2259 if (conf->scribble_disks >= new_disks &&
2260 conf->scribble_sectors >= new_sectors)
2261 return 0;
2262 mddev_suspend(conf->mddev);
2263 get_online_cpus();
2264
2265 for_each_present_cpu(cpu) {
2266 struct raid5_percpu *percpu;
2267
2268 percpu = per_cpu_ptr(conf->percpu, cpu);
2269 err = scribble_alloc(percpu, new_disks,
2270 new_sectors / STRIPE_SECTORS,
2271 GFP_NOIO);
2272 if (err)
2273 break;
2274 }
2275
2276 put_online_cpus();
2277 mddev_resume(conf->mddev);
2278 if (!err) {
2279 conf->scribble_disks = new_disks;
2280 conf->scribble_sectors = new_sectors;
2281 }
2282 return err;
2283}
2284
2285static int resize_stripes(struct r5conf *conf, int newsize)
2286{
2287 /* Make all the stripes able to hold 'newsize' devices.
2288 * New slots in each stripe get 'page' set to a new page.
2289 *
2290 * This happens in stages:
2291 * 1/ create a new kmem_cache and allocate the required number of
2292 * stripe_heads.
2293 * 2/ gather all the old stripe_heads and transfer the pages across
2294 * to the new stripe_heads. This will have the side effect of
2295 * freezing the array as once all stripe_heads have been collected,
2296 * no IO will be possible. Old stripe heads are freed once their
2297 * pages have been transferred over, and the old kmem_cache is
2298 * freed when all stripes are done.
2299 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
2300 * we simple return a failure status - no need to clean anything up.
2301 * 4/ allocate new pages for the new slots in the new stripe_heads.
2302 * If this fails, we don't bother trying the shrink the
2303 * stripe_heads down again, we just leave them as they are.
2304 * As each stripe_head is processed the new one is released into
2305 * active service.
2306 *
2307 * Once step2 is started, we cannot afford to wait for a write,
2308 * so we use GFP_NOIO allocations.
2309 */
2310 struct stripe_head *osh, *nsh;
2311 LIST_HEAD(newstripes);
2312 struct disk_info *ndisks;
2313 int err = 0;
2314 struct kmem_cache *sc;
2315 int i;
2316 int hash, cnt;
2317
2318 md_allow_write(conf->mddev);
2319
2320 /* Step 1 */
2321 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2322 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2323 0, 0, NULL);
2324 if (!sc)
2325 return -ENOMEM;
2326
2327 /* Need to ensure auto-resizing doesn't interfere */
2328 mutex_lock(&conf->cache_size_mutex);
2329
2330 for (i = conf->max_nr_stripes; i; i--) {
2331 nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2332 if (!nsh)
2333 break;
2334
2335 list_add(&nsh->lru, &newstripes);
2336 }
2337 if (i) {
2338 /* didn't get enough, give up */
2339 while (!list_empty(&newstripes)) {
2340 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2341 list_del(&nsh->lru);
2342 free_stripe(sc, nsh);
2343 }
2344 kmem_cache_destroy(sc);
2345 mutex_unlock(&conf->cache_size_mutex);
2346 return -ENOMEM;
2347 }
2348 /* Step 2 - Must use GFP_NOIO now.
2349 * OK, we have enough stripes, start collecting inactive
2350 * stripes and copying them over
2351 */
2352 hash = 0;
2353 cnt = 0;
2354 list_for_each_entry(nsh, &newstripes, lru) {
2355 lock_device_hash_lock(conf, hash);
2356 wait_event_cmd(conf->wait_for_stripe,
2357 !list_empty(conf->inactive_list + hash),
2358 unlock_device_hash_lock(conf, hash),
2359 lock_device_hash_lock(conf, hash));
2360 osh = get_free_stripe(conf, hash);
2361 unlock_device_hash_lock(conf, hash);
2362
2363 for(i=0; i<conf->pool_size; i++) {
2364 nsh->dev[i].page = osh->dev[i].page;
2365 nsh->dev[i].orig_page = osh->dev[i].page;
2366 }
2367 nsh->hash_lock_index = hash;
2368 free_stripe(conf->slab_cache, osh);
2369 cnt++;
2370 if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2371 !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2372 hash++;
2373 cnt = 0;
2374 }
2375 }
2376 kmem_cache_destroy(conf->slab_cache);
2377
2378 /* Step 3.
2379 * At this point, we are holding all the stripes so the array
2380 * is completely stalled, so now is a good time to resize
2381 * conf->disks and the scribble region
2382 */
2383 ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2384 if (ndisks) {
2385 for (i = 0; i < conf->pool_size; i++)
2386 ndisks[i] = conf->disks[i];
2387
2388 for (i = conf->pool_size; i < newsize; i++) {
2389 ndisks[i].extra_page = alloc_page(GFP_NOIO);
2390 if (!ndisks[i].extra_page)
2391 err = -ENOMEM;
2392 }
2393
2394 if (err) {
2395 for (i = conf->pool_size; i < newsize; i++)
2396 if (ndisks[i].extra_page)
2397 put_page(ndisks[i].extra_page);
2398 kfree(ndisks);
2399 } else {
2400 kfree(conf->disks);
2401 conf->disks = ndisks;
2402 }
2403 } else
2404 err = -ENOMEM;
2405
2406 mutex_unlock(&conf->cache_size_mutex);
2407
2408 conf->slab_cache = sc;
2409 conf->active_name = 1-conf->active_name;
2410
2411 /* Step 4, return new stripes to service */
2412 while(!list_empty(&newstripes)) {
2413 nsh = list_entry(newstripes.next, struct stripe_head, lru);
2414 list_del_init(&nsh->lru);
2415
2416 for (i=conf->raid_disks; i < newsize; i++)
2417 if (nsh->dev[i].page == NULL) {
2418 struct page *p = alloc_page(GFP_NOIO);
2419 nsh->dev[i].page = p;
2420 nsh->dev[i].orig_page = p;
2421 if (!p)
2422 err = -ENOMEM;
2423 }
2424 raid5_release_stripe(nsh);
2425 }
2426 /* critical section pass, GFP_NOIO no longer needed */
2427
2428 if (!err)
2429 conf->pool_size = newsize;
2430 return err;
2431}
2432
2433static int drop_one_stripe(struct r5conf *conf)
2434{
2435 struct stripe_head *sh;
2436 int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2437
2438 spin_lock_irq(conf->hash_locks + hash);
2439 sh = get_free_stripe(conf, hash);
2440 spin_unlock_irq(conf->hash_locks + hash);
2441 if (!sh)
2442 return 0;
2443 BUG_ON(atomic_read(&sh->count));
2444 shrink_buffers(sh);
2445 free_stripe(conf->slab_cache, sh);
2446 atomic_dec(&conf->active_stripes);
2447 conf->max_nr_stripes--;
2448 return 1;
2449}
2450
2451static void shrink_stripes(struct r5conf *conf)
2452{
2453 while (conf->max_nr_stripes &&
2454 drop_one_stripe(conf))
2455 ;
2456
2457 kmem_cache_destroy(conf->slab_cache);
2458 conf->slab_cache = NULL;
2459}
2460
2461static void raid5_end_read_request(struct bio * bi)
2462{
2463 struct stripe_head *sh = bi->bi_private;
2464 struct r5conf *conf = sh->raid_conf;
2465 int disks = sh->disks, i;
2466 char b[BDEVNAME_SIZE];
2467 struct md_rdev *rdev = NULL;
2468 sector_t s;
2469
2470 for (i=0 ; i<disks; i++)
2471 if (bi == &sh->dev[i].req)
2472 break;
2473
2474 pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2475 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2476 bi->bi_status);
2477 if (i == disks) {
2478 bio_reset(bi);
2479 BUG();
2480 return;
2481 }
2482 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2483 /* If replacement finished while this request was outstanding,
2484 * 'replacement' might be NULL already.
2485 * In that case it moved down to 'rdev'.
2486 * rdev is not removed until all requests are finished.
2487 */
2488 rdev = conf->disks[i].replacement;
2489 if (!rdev)
2490 rdev = conf->disks[i].rdev;
2491
2492 if (use_new_offset(conf, sh))
2493 s = sh->sector + rdev->new_data_offset;
2494 else
2495 s = sh->sector + rdev->data_offset;
2496 if (!bi->bi_status) {
2497 set_bit(R5_UPTODATE, &sh->dev[i].flags);
2498 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2499 /* Note that this cannot happen on a
2500 * replacement device. We just fail those on
2501 * any error
2502 */
2503 pr_info_ratelimited(
2504 "md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2505 mdname(conf->mddev), STRIPE_SECTORS,
2506 (unsigned long long)s,
2507 bdevname(rdev->bdev, b));
2508 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2509 clear_bit(R5_ReadError, &sh->dev[i].flags);
2510 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2511 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2512 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2513
2514 if (test_bit(R5_InJournal, &sh->dev[i].flags))
2515 /*
2516 * end read for a page in journal, this
2517 * must be preparing for prexor in rmw
2518 */
2519 set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2520
2521 if (atomic_read(&rdev->read_errors))
2522 atomic_set(&rdev->read_errors, 0);
2523 } else {
2524 const char *bdn = bdevname(rdev->bdev, b);
2525 int retry = 0;
2526 int set_bad = 0;
2527
2528 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2529 if (!(bi->bi_status == BLK_STS_PROTECTION))
2530 atomic_inc(&rdev->read_errors);
2531 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2532 pr_warn_ratelimited(
2533 "md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2534 mdname(conf->mddev),
2535 (unsigned long long)s,
2536 bdn);
2537 else if (conf->mddev->degraded >= conf->max_degraded) {
2538 set_bad = 1;
2539 pr_warn_ratelimited(
2540 "md/raid:%s: read error not correctable (sector %llu on %s).\n",
2541 mdname(conf->mddev),
2542 (unsigned long long)s,
2543 bdn);
2544 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2545 /* Oh, no!!! */
2546 set_bad = 1;
2547 pr_warn_ratelimited(
2548 "md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2549 mdname(conf->mddev),
2550 (unsigned long long)s,
2551 bdn);
2552 } else if (atomic_read(&rdev->read_errors)
2553 > conf->max_nr_stripes) {
2554 if (!test_bit(Faulty, &rdev->flags)) {
2555 pr_warn("md/raid:%s: %d read_errors > %d stripes\n",
2556 mdname(conf->mddev),
2557 atomic_read(&rdev->read_errors),
2558 conf->max_nr_stripes);
2559 pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2560 mdname(conf->mddev), bdn);
2561 }
2562 } else
2563 retry = 1;
2564 if (set_bad && test_bit(In_sync, &rdev->flags)
2565 && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2566 retry = 1;
2567 if (retry)
2568 if (sh->qd_idx >= 0 && sh->pd_idx == i)
2569 set_bit(R5_ReadError, &sh->dev[i].flags);
2570 else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2571 set_bit(R5_ReadError, &sh->dev[i].flags);
2572 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2573 } else
2574 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2575 else {
2576 clear_bit(R5_ReadError, &sh->dev[i].flags);
2577 clear_bit(R5_ReWrite, &sh->dev[i].flags);
2578 if (!(set_bad
2579 && test_bit(In_sync, &rdev->flags)
2580 && rdev_set_badblocks(
2581 rdev, sh->sector, STRIPE_SECTORS, 0)))
2582 md_error(conf->mddev, rdev);
2583 }
2584 }
2585 rdev_dec_pending(rdev, conf->mddev);
2586 bio_reset(bi);
2587 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2588 set_bit(STRIPE_HANDLE, &sh->state);
2589 raid5_release_stripe(sh);
2590}
2591
2592static void raid5_end_write_request(struct bio *bi)
2593{
2594 struct stripe_head *sh = bi->bi_private;
2595 struct r5conf *conf = sh->raid_conf;
2596 int disks = sh->disks, i;
2597 struct md_rdev *uninitialized_var(rdev);
2598 sector_t first_bad;
2599 int bad_sectors;
2600 int replacement = 0;
2601
2602 for (i = 0 ; i < disks; i++) {
2603 if (bi == &sh->dev[i].req) {
2604 rdev = conf->disks[i].rdev;
2605 break;
2606 }
2607 if (bi == &sh->dev[i].rreq) {
2608 rdev = conf->disks[i].replacement;
2609 if (rdev)
2610 replacement = 1;
2611 else
2612 /* rdev was removed and 'replacement'
2613 * replaced it. rdev is not removed
2614 * until all requests are finished.
2615 */
2616 rdev = conf->disks[i].rdev;
2617 break;
2618 }
2619 }
2620 pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2621 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
2622 bi->bi_status);
2623 if (i == disks) {
2624 bio_reset(bi);
2625 BUG();
2626 return;
2627 }
2628
2629 if (replacement) {
2630 if (bi->bi_status)
2631 md_error(conf->mddev, rdev);
2632 else if (is_badblock(rdev, sh->sector,
2633 STRIPE_SECTORS,
2634 &first_bad, &bad_sectors))
2635 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2636 } else {
2637 if (bi->bi_status) {
2638 set_bit(STRIPE_DEGRADED, &sh->state);
2639 set_bit(WriteErrorSeen, &rdev->flags);
2640 set_bit(R5_WriteError, &sh->dev[i].flags);
2641 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2642 set_bit(MD_RECOVERY_NEEDED,
2643 &rdev->mddev->recovery);
2644 } else if (is_badblock(rdev, sh->sector,
2645 STRIPE_SECTORS,
2646 &first_bad, &bad_sectors)) {
2647 set_bit(R5_MadeGood, &sh->dev[i].flags);
2648 if (test_bit(R5_ReadError, &sh->dev[i].flags))
2649 /* That was a successful write so make
2650 * sure it looks like we already did
2651 * a re-write.
2652 */
2653 set_bit(R5_ReWrite, &sh->dev[i].flags);
2654 }
2655 }
2656 rdev_dec_pending(rdev, conf->mddev);
2657
2658 if (sh->batch_head && bi->bi_status && !replacement)
2659 set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2660
2661 bio_reset(bi);
2662 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2663 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2664 set_bit(STRIPE_HANDLE, &sh->state);
2665 raid5_release_stripe(sh);
2666
2667 if (sh->batch_head && sh != sh->batch_head)
2668 raid5_release_stripe(sh->batch_head);
2669}
2670
2671static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2672{
2673 char b[BDEVNAME_SIZE];
2674 struct r5conf *conf = mddev->private;
2675 unsigned long flags;
2676 pr_debug("raid456: error called\n");
2677
2678 spin_lock_irqsave(&conf->device_lock, flags);
2679
2680 if (test_bit(In_sync, &rdev->flags) &&
2681 mddev->degraded == conf->max_degraded) {
2682 /*
2683 * Don't allow to achieve failed state
2684 * Don't try to recover this device
2685 */
2686 conf->recovery_disabled = mddev->recovery_disabled;
2687 spin_unlock_irqrestore(&conf->device_lock, flags);
2688 return;
2689 }
2690
2691 set_bit(Faulty, &rdev->flags);
2692 clear_bit(In_sync, &rdev->flags);
2693 mddev->degraded = raid5_calc_degraded(conf);
2694 spin_unlock_irqrestore(&conf->device_lock, flags);
2695 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2696
2697 set_bit(Blocked, &rdev->flags);
2698 set_mask_bits(&mddev->sb_flags, 0,
2699 BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2700 pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2701 "md/raid:%s: Operation continuing on %d devices.\n",
2702 mdname(mddev),
2703 bdevname(rdev->bdev, b),
2704 mdname(mddev),
2705 conf->raid_disks - mddev->degraded);
2706 r5c_update_on_rdev_error(mddev, rdev);
2707}
2708
2709/*
2710 * Input: a 'big' sector number,
2711 * Output: index of the data and parity disk, and the sector # in them.
2712 */
2713sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2714 int previous, int *dd_idx,
2715 struct stripe_head *sh)
2716{
2717 sector_t stripe, stripe2;
2718 sector_t chunk_number;
2719 unsigned int chunk_offset;
2720 int pd_idx, qd_idx;
2721 int ddf_layout = 0;
2722 sector_t new_sector;
2723 int algorithm = previous ? conf->prev_algo
2724 : conf->algorithm;
2725 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2726 : conf->chunk_sectors;
2727 int raid_disks = previous ? conf->previous_raid_disks
2728 : conf->raid_disks;
2729 int data_disks = raid_disks - conf->max_degraded;
2730
2731 /* First compute the information on this sector */
2732
2733 /*
2734 * Compute the chunk number and the sector offset inside the chunk
2735 */
2736 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2737 chunk_number = r_sector;
2738
2739 /*
2740 * Compute the stripe number
2741 */
2742 stripe = chunk_number;
2743 *dd_idx = sector_div(stripe, data_disks);
2744 stripe2 = stripe;
2745 /*
2746 * Select the parity disk based on the user selected algorithm.
2747 */
2748 pd_idx = qd_idx = -1;
2749 switch(conf->level) {
2750 case 4:
2751 pd_idx = data_disks;
2752 break;
2753 case 5:
2754 switch (algorithm) {
2755 case ALGORITHM_LEFT_ASYMMETRIC:
2756 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2757 if (*dd_idx >= pd_idx)
2758 (*dd_idx)++;
2759 break;
2760 case ALGORITHM_RIGHT_ASYMMETRIC:
2761 pd_idx = sector_div(stripe2, raid_disks);
2762 if (*dd_idx >= pd_idx)
2763 (*dd_idx)++;
2764 break;
2765 case ALGORITHM_LEFT_SYMMETRIC:
2766 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2767 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2768 break;
2769 case ALGORITHM_RIGHT_SYMMETRIC:
2770 pd_idx = sector_div(stripe2, raid_disks);
2771 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2772 break;
2773 case ALGORITHM_PARITY_0:
2774 pd_idx = 0;
2775 (*dd_idx)++;
2776 break;
2777 case ALGORITHM_PARITY_N:
2778 pd_idx = data_disks;
2779 break;
2780 default:
2781 BUG();
2782 }
2783 break;
2784 case 6:
2785
2786 switch (algorithm) {
2787 case ALGORITHM_LEFT_ASYMMETRIC:
2788 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2789 qd_idx = pd_idx + 1;
2790 if (pd_idx == raid_disks-1) {
2791 (*dd_idx)++; /* Q D D D P */
2792 qd_idx = 0;
2793 } else if (*dd_idx >= pd_idx)
2794 (*dd_idx) += 2; /* D D P Q D */
2795 break;
2796 case ALGORITHM_RIGHT_ASYMMETRIC:
2797 pd_idx = sector_div(stripe2, raid_disks);
2798 qd_idx = pd_idx + 1;
2799 if (pd_idx == raid_disks-1) {
2800 (*dd_idx)++; /* Q D D D P */
2801 qd_idx = 0;
2802 } else if (*dd_idx >= pd_idx)
2803 (*dd_idx) += 2; /* D D P Q D */
2804 break;
2805 case ALGORITHM_LEFT_SYMMETRIC:
2806 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2807 qd_idx = (pd_idx + 1) % raid_disks;
2808 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2809 break;
2810 case ALGORITHM_RIGHT_SYMMETRIC:
2811 pd_idx = sector_div(stripe2, raid_disks);
2812 qd_idx = (pd_idx + 1) % raid_disks;
2813 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2814 break;
2815
2816 case ALGORITHM_PARITY_0:
2817 pd_idx = 0;
2818 qd_idx = 1;
2819 (*dd_idx) += 2;
2820 break;
2821 case ALGORITHM_PARITY_N:
2822 pd_idx = data_disks;
2823 qd_idx = data_disks + 1;
2824 break;
2825
2826 case ALGORITHM_ROTATING_ZERO_RESTART:
2827 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2828 * of blocks for computing Q is different.
2829 */
2830 pd_idx = sector_div(stripe2, raid_disks);
2831 qd_idx = pd_idx + 1;
2832 if (pd_idx == raid_disks-1) {
2833 (*dd_idx)++; /* Q D D D P */
2834 qd_idx = 0;
2835 } else if (*dd_idx >= pd_idx)
2836 (*dd_idx) += 2; /* D D P Q D */
2837 ddf_layout = 1;
2838 break;
2839
2840 case ALGORITHM_ROTATING_N_RESTART:
2841 /* Same a left_asymmetric, by first stripe is
2842 * D D D P Q rather than
2843 * Q D D D P
2844 */
2845 stripe2 += 1;
2846 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2847 qd_idx = pd_idx + 1;
2848 if (pd_idx == raid_disks-1) {
2849 (*dd_idx)++; /* Q D D D P */
2850 qd_idx = 0;
2851 } else if (*dd_idx >= pd_idx)
2852 (*dd_idx) += 2; /* D D P Q D */
2853 ddf_layout = 1;
2854 break;
2855
2856 case ALGORITHM_ROTATING_N_CONTINUE:
2857 /* Same as left_symmetric but Q is before P */
2858 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2859 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2860 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2861 ddf_layout = 1;
2862 break;
2863
2864 case ALGORITHM_LEFT_ASYMMETRIC_6:
2865 /* RAID5 left_asymmetric, with Q on last device */
2866 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2867 if (*dd_idx >= pd_idx)
2868 (*dd_idx)++;
2869 qd_idx = raid_disks - 1;
2870 break;
2871
2872 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2873 pd_idx = sector_div(stripe2, raid_disks-1);
2874 if (*dd_idx >= pd_idx)
2875 (*dd_idx)++;
2876 qd_idx = raid_disks - 1;
2877 break;
2878
2879 case ALGORITHM_LEFT_SYMMETRIC_6:
2880 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2881 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2882 qd_idx = raid_disks - 1;
2883 break;
2884
2885 case ALGORITHM_RIGHT_SYMMETRIC_6:
2886 pd_idx = sector_div(stripe2, raid_disks-1);
2887 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2888 qd_idx = raid_disks - 1;
2889 break;
2890
2891 case ALGORITHM_PARITY_0_6:
2892 pd_idx = 0;
2893 (*dd_idx)++;
2894 qd_idx = raid_disks - 1;
2895 break;
2896
2897 default:
2898 BUG();
2899 }
2900 break;
2901 }
2902
2903 if (sh) {
2904 sh->pd_idx = pd_idx;
2905 sh->qd_idx = qd_idx;
2906 sh->ddf_layout = ddf_layout;
2907 }
2908 /*
2909 * Finally, compute the new sector number
2910 */
2911 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2912 return new_sector;
2913}
2914
2915sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2916{
2917 struct r5conf *conf = sh->raid_conf;
2918 int raid_disks = sh->disks;
2919 int data_disks = raid_disks - conf->max_degraded;
2920 sector_t new_sector = sh->sector, check;
2921 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2922 : conf->chunk_sectors;
2923 int algorithm = previous ? conf->prev_algo
2924 : conf->algorithm;
2925 sector_t stripe;
2926 int chunk_offset;
2927 sector_t chunk_number;
2928 int dummy1, dd_idx = i;
2929 sector_t r_sector;
2930 struct stripe_head sh2;
2931
2932 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2933 stripe = new_sector;
2934
2935 if (i == sh->pd_idx)
2936 return 0;
2937 switch(conf->level) {
2938 case 4: break;
2939 case 5:
2940 switch (algorithm) {
2941 case ALGORITHM_LEFT_ASYMMETRIC:
2942 case ALGORITHM_RIGHT_ASYMMETRIC:
2943 if (i > sh->pd_idx)
2944 i--;
2945 break;
2946 case ALGORITHM_LEFT_SYMMETRIC:
2947 case ALGORITHM_RIGHT_SYMMETRIC:
2948 if (i < sh->pd_idx)
2949 i += raid_disks;
2950 i -= (sh->pd_idx + 1);
2951 break;
2952 case ALGORITHM_PARITY_0:
2953 i -= 1;
2954 break;
2955 case ALGORITHM_PARITY_N:
2956 break;
2957 default:
2958 BUG();
2959 }
2960 break;
2961 case 6:
2962 if (i == sh->qd_idx)
2963 return 0; /* It is the Q disk */
2964 switch (algorithm) {
2965 case ALGORITHM_LEFT_ASYMMETRIC:
2966 case ALGORITHM_RIGHT_ASYMMETRIC:
2967 case ALGORITHM_ROTATING_ZERO_RESTART:
2968 case ALGORITHM_ROTATING_N_RESTART:
2969 if (sh->pd_idx == raid_disks-1)
2970 i--; /* Q D D D P */
2971 else if (i > sh->pd_idx)
2972 i -= 2; /* D D P Q D */
2973 break;
2974 case ALGORITHM_LEFT_SYMMETRIC:
2975 case ALGORITHM_RIGHT_SYMMETRIC:
2976 if (sh->pd_idx == raid_disks-1)
2977 i--; /* Q D D D P */
2978 else {
2979 /* D D P Q D */
2980 if (i < sh->pd_idx)
2981 i += raid_disks;
2982 i -= (sh->pd_idx + 2);
2983 }
2984 break;
2985 case ALGORITHM_PARITY_0:
2986 i -= 2;
2987 break;
2988 case ALGORITHM_PARITY_N:
2989 break;
2990 case ALGORITHM_ROTATING_N_CONTINUE:
2991 /* Like left_symmetric, but P is before Q */
2992 if (sh->pd_idx == 0)
2993 i--; /* P D D D Q */
2994 else {
2995 /* D D Q P D */
2996 if (i < sh->pd_idx)
2997 i += raid_disks;
2998 i -= (sh->pd_idx + 1);
2999 }
3000 break;
3001 case ALGORITHM_LEFT_ASYMMETRIC_6:
3002 case ALGORITHM_RIGHT_ASYMMETRIC_6:
3003 if (i > sh->pd_idx)
3004 i--;
3005 break;
3006 case ALGORITHM_LEFT_SYMMETRIC_6:
3007 case ALGORITHM_RIGHT_SYMMETRIC_6:
3008 if (i < sh->pd_idx)
3009 i += data_disks + 1;
3010 i -= (sh->pd_idx + 1);
3011 break;
3012 case ALGORITHM_PARITY_0_6:
3013 i -= 1;
3014 break;
3015 default:
3016 BUG();
3017 }
3018 break;
3019 }
3020
3021 chunk_number = stripe * data_disks + i;
3022 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3023
3024 check = raid5_compute_sector(conf, r_sector,
3025 previous, &dummy1, &sh2);
3026 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3027 || sh2.qd_idx != sh->qd_idx) {
3028 pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3029 mdname(conf->mddev));
3030 return 0;
3031 }
3032 return r_sector;
3033}
3034
3035/*
3036 * There are cases where we want handle_stripe_dirtying() and
3037 * schedule_reconstruction() to delay towrite to some dev of a stripe.
3038 *
3039 * This function checks whether we want to delay the towrite. Specifically,
3040 * we delay the towrite when:
3041 *
3042 * 1. degraded stripe has a non-overwrite to the missing dev, AND this
3043 * stripe has data in journal (for other devices).
3044 *
3045 * In this case, when reading data for the non-overwrite dev, it is
3046 * necessary to handle complex rmw of write back cache (prexor with
3047 * orig_page, and xor with page). To keep read path simple, we would
3048 * like to flush data in journal to RAID disks first, so complex rmw
3049 * is handled in the write patch (handle_stripe_dirtying).
3050 *
3051 * 2. when journal space is critical (R5C_LOG_CRITICAL=1)
3052 *
3053 * It is important to be able to flush all stripes in raid5-cache.
3054 * Therefore, we need reserve some space on the journal device for
3055 * these flushes. If flush operation includes pending writes to the
3056 * stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3057 * for the flush out. If we exclude these pending writes from flush
3058 * operation, we only need (conf->max_degraded + 1) pages per stripe.
3059 * Therefore, excluding pending writes in these cases enables more
3060 * efficient use of the journal device.
3061 *
3062 * Note: To make sure the stripe makes progress, we only delay
3063 * towrite for stripes with data already in journal (injournal > 0).
3064 * When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3065 * no_space_stripes list.
3066 *
3067 * 3. during journal failure
3068 * In journal failure, we try to flush all cached data to raid disks
3069 * based on data in stripe cache. The array is read-only to upper
3070 * layers, so we would skip all pending writes.
3071 *
3072 */
3073static inline bool delay_towrite(struct r5conf *conf,
3074 struct r5dev *dev,
3075 struct stripe_head_state *s)
3076{
3077 /* case 1 above */
3078 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3079 !test_bit(R5_Insync, &dev->flags) && s->injournal)
3080 return true;
3081 /* case 2 above */
3082 if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3083 s->injournal > 0)
3084 return true;
3085 /* case 3 above */
3086 if (s->log_failed && s->injournal)
3087 return true;
3088 return false;
3089}
3090
3091static void
3092schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3093 int rcw, int expand)
3094{
3095 int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3096 struct r5conf *conf = sh->raid_conf;
3097 int level = conf->level;
3098
3099 if (rcw) {
3100 /*
3101 * In some cases, handle_stripe_dirtying initially decided to
3102 * run rmw and allocates extra page for prexor. However, rcw is
3103 * cheaper later on. We need to free the extra page now,
3104 * because we won't be able to do that in ops_complete_prexor().
3105 */
3106 r5c_release_extra_page(sh);
3107
3108 for (i = disks; i--; ) {
3109 struct r5dev *dev = &sh->dev[i];
3110
3111 if (dev->towrite && !delay_towrite(conf, dev, s)) {
3112 set_bit(R5_LOCKED, &dev->flags);
3113 set_bit(R5_Wantdrain, &dev->flags);
3114 if (!expand)
3115 clear_bit(R5_UPTODATE, &dev->flags);
3116 s->locked++;
3117 } else if (test_bit(R5_InJournal, &dev->flags)) {
3118 set_bit(R5_LOCKED, &dev->flags);
3119 s->locked++;
3120 }
3121 }
3122 /* if we are not expanding this is a proper write request, and
3123 * there will be bios with new data to be drained into the
3124 * stripe cache
3125 */
3126 if (!expand) {
3127 if (!s->locked)
3128 /* False alarm, nothing to do */
3129 return;
3130 sh->reconstruct_state = reconstruct_state_drain_run;
3131 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3132 } else
3133 sh->reconstruct_state = reconstruct_state_run;
3134
3135 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3136
3137 if (s->locked + conf->max_degraded == disks)
3138 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3139 atomic_inc(&conf->pending_full_writes);
3140 } else {
3141 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3142 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3143 BUG_ON(level == 6 &&
3144 (!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3145 test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3146
3147 for (i = disks; i--; ) {
3148 struct r5dev *dev = &sh->dev[i];
3149 if (i == pd_idx || i == qd_idx)
3150 continue;
3151
3152 if (dev->towrite &&
3153 (test_bit(R5_UPTODATE, &dev->flags) ||
3154 test_bit(R5_Wantcompute, &dev->flags))) {
3155 set_bit(R5_Wantdrain, &dev->flags);
3156 set_bit(R5_LOCKED, &dev->flags);
3157 clear_bit(R5_UPTODATE, &dev->flags);
3158 s->locked++;
3159 } else if (test_bit(R5_InJournal, &dev->flags)) {
3160 set_bit(R5_LOCKED, &dev->flags);
3161 s->locked++;
3162 }
3163 }
3164 if (!s->locked)
3165 /* False alarm - nothing to do */
3166 return;
3167 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3168 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3169 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3170 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3171 }
3172
3173 /* keep the parity disk(s) locked while asynchronous operations
3174 * are in flight
3175 */
3176 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3177 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3178 s->locked++;
3179
3180 if (level == 6) {
3181 int qd_idx = sh->qd_idx;
3182 struct r5dev *dev = &sh->dev[qd_idx];
3183
3184 set_bit(R5_LOCKED, &dev->flags);
3185 clear_bit(R5_UPTODATE, &dev->flags);
3186 s->locked++;
3187 }
3188
3189 if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3190 test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3191 !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3192 test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3193 set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3194
3195 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3196 __func__, (unsigned long long)sh->sector,
3197 s->locked, s->ops_request);
3198}
3199
3200/*
3201 * Each stripe/dev can have one or more bion attached.
3202 * toread/towrite point to the first in a chain.
3203 * The bi_next chain must be in order.
3204 */
3205static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3206 int forwrite, int previous)
3207{
3208 struct bio **bip;
3209 struct r5conf *conf = sh->raid_conf;
3210 int firstwrite=0;
3211
3212 pr_debug("adding bi b#%llu to stripe s#%llu\n",
3213 (unsigned long long)bi->bi_iter.bi_sector,
3214 (unsigned long long)sh->sector);
3215
3216 spin_lock_irq(&sh->stripe_lock);
3217 sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3218 /* Don't allow new IO added to stripes in batch list */
3219 if (sh->batch_head)
3220 goto overlap;
3221 if (forwrite) {
3222 bip = &sh->dev[dd_idx].towrite;
3223 if (*bip == NULL)
3224 firstwrite = 1;
3225 } else
3226 bip = &sh->dev[dd_idx].toread;
3227 while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3228 if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3229 goto overlap;
3230 bip = & (*bip)->bi_next;
3231 }
3232 if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3233 goto overlap;
3234
3235 if (forwrite && raid5_has_ppl(conf)) {
3236 /*
3237 * With PPL only writes to consecutive data chunks within a
3238 * stripe are allowed because for a single stripe_head we can
3239 * only have one PPL entry at a time, which describes one data
3240 * range. Not really an overlap, but wait_for_overlap can be
3241 * used to handle this.
3242 */
3243 sector_t sector;
3244 sector_t first = 0;
3245 sector_t last = 0;
3246 int count = 0;
3247 int i;
3248
3249 for (i = 0; i < sh->disks; i++) {
3250 if (i != sh->pd_idx &&
3251 (i == dd_idx || sh->dev[i].towrite)) {
3252 sector = sh->dev[i].sector;
3253 if (count == 0 || sector < first)
3254 first = sector;
3255 if (sector > last)
3256 last = sector;
3257 count++;
3258 }
3259 }
3260
3261 if (first + conf->chunk_sectors * (count - 1) != last)
3262 goto overlap;
3263 }
3264
3265 if (!forwrite || previous)
3266 clear_bit(STRIPE_BATCH_READY, &sh->state);
3267
3268 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3269 if (*bip)
3270 bi->bi_next = *bip;
3271 *bip = bi;
3272 bio_inc_remaining(bi);
3273 md_write_inc(conf->mddev, bi);
3274
3275 if (forwrite) {
3276 /* check if page is covered */
3277 sector_t sector = sh->dev[dd_idx].sector;
3278 for (bi=sh->dev[dd_idx].towrite;
3279 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3280 bi && bi->bi_iter.bi_sector <= sector;
3281 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3282 if (bio_end_sector(bi) >= sector)
3283 sector = bio_end_sector(bi);
3284 }
3285 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3286 if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3287 sh->overwrite_disks++;
3288 }
3289
3290 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3291 (unsigned long long)(*bip)->bi_iter.bi_sector,
3292 (unsigned long long)sh->sector, dd_idx);
3293
3294 if (conf->mddev->bitmap && firstwrite) {
3295 /* Cannot hold spinlock over bitmap_startwrite,
3296 * but must ensure this isn't added to a batch until
3297 * we have added to the bitmap and set bm_seq.
3298 * So set STRIPE_BITMAP_PENDING to prevent
3299 * batching.
3300 * If multiple add_stripe_bio() calls race here they
3301 * much all set STRIPE_BITMAP_PENDING. So only the first one
3302 * to complete "bitmap_startwrite" gets to set
3303 * STRIPE_BIT_DELAY. This is important as once a stripe
3304 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3305 * any more.
3306 */
3307 set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3308 spin_unlock_irq(&sh->stripe_lock);
3309 md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3310 STRIPE_SECTORS, 0);
3311 spin_lock_irq(&sh->stripe_lock);
3312 clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3313 if (!sh->batch_head) {
3314 sh->bm_seq = conf->seq_flush+1;
3315 set_bit(STRIPE_BIT_DELAY, &sh->state);
3316 }
3317 }
3318 spin_unlock_irq(&sh->stripe_lock);
3319
3320 if (stripe_can_batch(sh))
3321 stripe_add_to_batch_list(conf, sh);
3322 return 1;
3323
3324 overlap:
3325 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3326 spin_unlock_irq(&sh->stripe_lock);
3327 return 0;
3328}
3329
3330static void end_reshape(struct r5conf *conf);
3331
3332static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3333 struct stripe_head *sh)
3334{
3335 int sectors_per_chunk =
3336 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3337 int dd_idx;
3338 int chunk_offset = sector_div(stripe, sectors_per_chunk);
3339 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3340
3341 raid5_compute_sector(conf,
3342 stripe * (disks - conf->max_degraded)
3343 *sectors_per_chunk + chunk_offset,
3344 previous,
3345 &dd_idx, sh);
3346}
3347
3348static void
3349handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3350 struct stripe_head_state *s, int disks)
3351{
3352 int i;
3353 BUG_ON(sh->batch_head);
3354 for (i = disks; i--; ) {
3355 struct bio *bi;
3356 int bitmap_end = 0;
3357
3358 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3359 struct md_rdev *rdev;
3360 rcu_read_lock();
3361 rdev = rcu_dereference(conf->disks[i].rdev);
3362 if (rdev && test_bit(In_sync, &rdev->flags) &&
3363 !test_bit(Faulty, &rdev->flags))
3364 atomic_inc(&rdev->nr_pending);
3365 else
3366 rdev = NULL;
3367 rcu_read_unlock();
3368 if (rdev) {
3369 if (!rdev_set_badblocks(
3370 rdev,
3371 sh->sector,
3372 STRIPE_SECTORS, 0))
3373 md_error(conf->mddev, rdev);
3374 rdev_dec_pending(rdev, conf->mddev);
3375 }
3376 }
3377 spin_lock_irq(&sh->stripe_lock);
3378 /* fail all writes first */
3379 bi = sh->dev[i].towrite;
3380 sh->dev[i].towrite = NULL;
3381 sh->overwrite_disks = 0;
3382 spin_unlock_irq(&sh->stripe_lock);
3383 if (bi)
3384 bitmap_end = 1;
3385
3386 log_stripe_write_finished(sh);
3387
3388 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3389 wake_up(&conf->wait_for_overlap);
3390
3391 while (bi && bi->bi_iter.bi_sector <
3392 sh->dev[i].sector + STRIPE_SECTORS) {
3393 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3394
3395 md_write_end(conf->mddev);
3396 bio_io_error(bi);
3397 bi = nextbi;
3398 }
3399 if (bitmap_end)
3400 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3401 STRIPE_SECTORS, 0, 0);
3402 bitmap_end = 0;
3403 /* and fail all 'written' */
3404 bi = sh->dev[i].written;
3405 sh->dev[i].written = NULL;
3406 if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3407 WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3408 sh->dev[i].page = sh->dev[i].orig_page;
3409 }
3410
3411 if (bi) bitmap_end = 1;
3412 while (bi && bi->bi_iter.bi_sector <
3413 sh->dev[i].sector + STRIPE_SECTORS) {
3414 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3415
3416 md_write_end(conf->mddev);
3417 bio_io_error(bi);
3418 bi = bi2;
3419 }
3420
3421 /* fail any reads if this device is non-operational and
3422 * the data has not reached the cache yet.
3423 */
3424 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3425 s->failed > conf->max_degraded &&
3426 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3427 test_bit(R5_ReadError, &sh->dev[i].flags))) {
3428 spin_lock_irq(&sh->stripe_lock);
3429 bi = sh->dev[i].toread;
3430 sh->dev[i].toread = NULL;
3431 spin_unlock_irq(&sh->stripe_lock);
3432 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3433 wake_up(&conf->wait_for_overlap);
3434 if (bi)
3435 s->to_read--;
3436 while (bi && bi->bi_iter.bi_sector <
3437 sh->dev[i].sector + STRIPE_SECTORS) {
3438 struct bio *nextbi =
3439 r5_next_bio(bi, sh->dev[i].sector);
3440
3441 bio_io_error(bi);
3442 bi = nextbi;
3443 }
3444 }
3445 if (bitmap_end)
3446 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3447 STRIPE_SECTORS, 0, 0);
3448 /* If we were in the middle of a write the parity block might
3449 * still be locked - so just clear all R5_LOCKED flags
3450 */
3451 clear_bit(R5_LOCKED, &sh->dev[i].flags);
3452 }
3453 s->to_write = 0;
3454 s->written = 0;
3455
3456 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3457 if (atomic_dec_and_test(&conf->pending_full_writes))
3458 md_wakeup_thread(conf->mddev->thread);
3459}
3460
3461static void
3462handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3463 struct stripe_head_state *s)
3464{
3465 int abort = 0;
3466 int i;
3467
3468 BUG_ON(sh->batch_head);
3469 clear_bit(STRIPE_SYNCING, &sh->state);
3470 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3471 wake_up(&conf->wait_for_overlap);
3472 s->syncing = 0;
3473 s->replacing = 0;
3474 /* There is nothing more to do for sync/check/repair.
3475 * Don't even need to abort as that is handled elsewhere
3476 * if needed, and not always wanted e.g. if there is a known
3477 * bad block here.
3478 * For recover/replace we need to record a bad block on all
3479 * non-sync devices, or abort the recovery
3480 */
3481 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3482 /* During recovery devices cannot be removed, so
3483 * locking and refcounting of rdevs is not needed
3484 */
3485 rcu_read_lock();
3486 for (i = 0; i < conf->raid_disks; i++) {
3487 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3488 if (rdev
3489 && !test_bit(Faulty, &rdev->flags)
3490 && !test_bit(In_sync, &rdev->flags)
3491 && !rdev_set_badblocks(rdev, sh->sector,
3492 STRIPE_SECTORS, 0))
3493 abort = 1;
3494 rdev = rcu_dereference(conf->disks[i].replacement);
3495 if (rdev
3496 && !test_bit(Faulty, &rdev->flags)
3497 && !test_bit(In_sync, &rdev->flags)
3498 && !rdev_set_badblocks(rdev, sh->sector,
3499 STRIPE_SECTORS, 0))
3500 abort = 1;
3501 }
3502 rcu_read_unlock();
3503 if (abort)
3504 conf->recovery_disabled =
3505 conf->mddev->recovery_disabled;
3506 }
3507 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3508}
3509
3510static int want_replace(struct stripe_head *sh, int disk_idx)
3511{
3512 struct md_rdev *rdev;
3513 int rv = 0;
3514
3515 rcu_read_lock();
3516 rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3517 if (rdev
3518 && !test_bit(Faulty, &rdev->flags)
3519 && !test_bit(In_sync, &rdev->flags)
3520 && (rdev->recovery_offset <= sh->sector
3521 || rdev->mddev->recovery_cp <= sh->sector))
3522 rv = 1;
3523 rcu_read_unlock();
3524 return rv;
3525}
3526
3527static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3528 int disk_idx, int disks)
3529{
3530 struct r5dev *dev = &sh->dev[disk_idx];
3531 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3532 &sh->dev[s->failed_num[1]] };
3533 int i;
3534
3535
3536 if (test_bit(R5_LOCKED, &dev->flags) ||
3537 test_bit(R5_UPTODATE, &dev->flags))
3538 /* No point reading this as we already have it or have
3539 * decided to get it.
3540 */
3541 return 0;
3542
3543 if (dev->toread ||
3544 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3545 /* We need this block to directly satisfy a request */
3546 return 1;
3547
3548 if (s->syncing || s->expanding ||
3549 (s->replacing && want_replace(sh, disk_idx)))
3550 /* When syncing, or expanding we read everything.
3551 * When replacing, we need the replaced block.
3552 */
3553 return 1;
3554
3555 if ((s->failed >= 1 && fdev[0]->toread) ||
3556 (s->failed >= 2 && fdev[1]->toread))
3557 /* If we want to read from a failed device, then
3558 * we need to actually read every other device.
3559 */
3560 return 1;
3561
3562 /* Sometimes neither read-modify-write nor reconstruct-write
3563 * cycles can work. In those cases we read every block we
3564 * can. Then the parity-update is certain to have enough to
3565 * work with.
3566 * This can only be a problem when we need to write something,
3567 * and some device has failed. If either of those tests
3568 * fail we need look no further.
3569 */
3570 if (!s->failed || !s->to_write)
3571 return 0;
3572
3573 if (test_bit(R5_Insync, &dev->flags) &&
3574 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3575 /* Pre-reads at not permitted until after short delay
3576 * to gather multiple requests. However if this
3577 * device is no Insync, the block could only be computed
3578 * and there is no need to delay that.
3579 */
3580 return 0;
3581
3582 for (i = 0; i < s->failed && i < 2; i++) {
3583 if (fdev[i]->towrite &&
3584 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3585 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3586 /* If we have a partial write to a failed
3587 * device, then we will need to reconstruct
3588 * the content of that device, so all other
3589 * devices must be read.
3590 */
3591 return 1;
3592 }
3593
3594 /* If we are forced to do a reconstruct-write, either because
3595 * the current RAID6 implementation only supports that, or
3596 * because parity cannot be trusted and we are currently
3597 * recovering it, there is extra need to be careful.
3598 * If one of the devices that we would need to read, because
3599 * it is not being overwritten (and maybe not written at all)
3600 * is missing/faulty, then we need to read everything we can.
3601 */
3602 if (sh->raid_conf->level != 6 &&
3603 sh->sector < sh->raid_conf->mddev->recovery_cp)
3604 /* reconstruct-write isn't being forced */
3605 return 0;
3606 for (i = 0; i < s->failed && i < 2; i++) {
3607 if (s->failed_num[i] != sh->pd_idx &&
3608 s->failed_num[i] != sh->qd_idx &&
3609 !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3610 !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3611 return 1;
3612 }
3613
3614 return 0;
3615}
3616
3617/* fetch_block - checks the given member device to see if its data needs
3618 * to be read or computed to satisfy a request.
3619 *
3620 * Returns 1 when no more member devices need to be checked, otherwise returns
3621 * 0 to tell the loop in handle_stripe_fill to continue
3622 */
3623static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3624 int disk_idx, int disks)
3625{
3626 struct r5dev *dev = &sh->dev[disk_idx];
3627
3628 /* is the data in this block needed, and can we get it? */
3629 if (need_this_block(sh, s, disk_idx, disks)) {
3630 /* we would like to get this block, possibly by computing it,
3631 * otherwise read it if the backing disk is insync
3632 */
3633 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3634 BUG_ON(test_bit(R5_Wantread, &dev->flags));
3635 BUG_ON(sh->batch_head);
3636
3637 /*
3638 * In the raid6 case if the only non-uptodate disk is P
3639 * then we already trusted P to compute the other failed
3640 * drives. It is safe to compute rather than re-read P.
3641 * In other cases we only compute blocks from failed
3642 * devices, otherwise check/repair might fail to detect
3643 * a real inconsistency.
3644 */
3645
3646 if ((s->uptodate == disks - 1) &&
3647 ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3648 (s->failed && (disk_idx == s->failed_num[0] ||
3649 disk_idx == s->failed_num[1])))) {
3650 /* have disk failed, and we're requested to fetch it;
3651 * do compute it
3652 */
3653 pr_debug("Computing stripe %llu block %d\n",
3654 (unsigned long long)sh->sector, disk_idx);
3655 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3656 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3657 set_bit(R5_Wantcompute, &dev->flags);
3658 sh->ops.target = disk_idx;
3659 sh->ops.target2 = -1; /* no 2nd target */
3660 s->req_compute = 1;
3661 /* Careful: from this point on 'uptodate' is in the eye
3662 * of raid_run_ops which services 'compute' operations
3663 * before writes. R5_Wantcompute flags a block that will
3664 * be R5_UPTODATE by the time it is needed for a
3665 * subsequent operation.
3666 */
3667 s->uptodate++;
3668 return 1;
3669 } else if (s->uptodate == disks-2 && s->failed >= 2) {
3670 /* Computing 2-failure is *very* expensive; only
3671 * do it if failed >= 2
3672 */
3673 int other;
3674 for (other = disks; other--; ) {
3675 if (other == disk_idx)
3676 continue;
3677 if (!test_bit(R5_UPTODATE,
3678 &sh->dev[other].flags))
3679 break;
3680 }
3681 BUG_ON(other < 0);
3682 pr_debug("Computing stripe %llu blocks %d,%d\n",
3683 (unsigned long long)sh->sector,
3684 disk_idx, other);
3685 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3686 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3687 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3688 set_bit(R5_Wantcompute, &sh->dev[other].flags);
3689 sh->ops.target = disk_idx;
3690 sh->ops.target2 = other;
3691 s->uptodate += 2;
3692 s->req_compute = 1;
3693 return 1;
3694 } else if (test_bit(R5_Insync, &dev->flags)) {
3695 set_bit(R5_LOCKED, &dev->flags);
3696 set_bit(R5_Wantread, &dev->flags);
3697 s->locked++;
3698 pr_debug("Reading block %d (sync=%d)\n",
3699 disk_idx, s->syncing);
3700 }
3701 }
3702
3703 return 0;
3704}
3705
3706/**
3707 * handle_stripe_fill - read or compute data to satisfy pending requests.
3708 */
3709static void handle_stripe_fill(struct stripe_head *sh,
3710 struct stripe_head_state *s,
3711 int disks)
3712{
3713 int i;
3714
3715 /* look for blocks to read/compute, skip this if a compute
3716 * is already in flight, or if the stripe contents are in the
3717 * midst of changing due to a write
3718 */
3719 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3720 !sh->reconstruct_state) {
3721
3722 /*
3723 * For degraded stripe with data in journal, do not handle
3724 * read requests yet, instead, flush the stripe to raid
3725 * disks first, this avoids handling complex rmw of write
3726 * back cache (prexor with orig_page, and then xor with
3727 * page) in the read path
3728 */
3729 if (s->injournal && s->failed) {
3730 if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3731 r5c_make_stripe_write_out(sh);
3732 goto out;
3733 }
3734
3735 for (i = disks; i--; )
3736 if (fetch_block(sh, s, i, disks))
3737 break;
3738 }
3739out:
3740 set_bit(STRIPE_HANDLE, &sh->state);
3741}
3742
3743static void break_stripe_batch_list(struct stripe_head *head_sh,
3744 unsigned long handle_flags);
3745/* handle_stripe_clean_event
3746 * any written block on an uptodate or failed drive can be returned.
3747 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3748 * never LOCKED, so we don't need to test 'failed' directly.
3749 */
3750static void handle_stripe_clean_event(struct r5conf *conf,
3751 struct stripe_head *sh, int disks)
3752{
3753 int i;
3754 struct r5dev *dev;
3755 int discard_pending = 0;
3756 struct stripe_head *head_sh = sh;
3757 bool do_endio = false;
3758
3759 for (i = disks; i--; )
3760 if (sh->dev[i].written) {
3761 dev = &sh->dev[i];
3762 if (!test_bit(R5_LOCKED, &dev->flags) &&
3763 (test_bit(R5_UPTODATE, &dev->flags) ||
3764 test_bit(R5_Discard, &dev->flags) ||
3765 test_bit(R5_SkipCopy, &dev->flags))) {
3766 /* We can return any write requests */
3767 struct bio *wbi, *wbi2;
3768 pr_debug("Return write for disc %d\n", i);
3769 if (test_and_clear_bit(R5_Discard, &dev->flags))
3770 clear_bit(R5_UPTODATE, &dev->flags);
3771 if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3772 WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3773 }
3774 do_endio = true;
3775
3776returnbi:
3777 dev->page = dev->orig_page;
3778 wbi = dev->written;
3779 dev->written = NULL;
3780 while (wbi && wbi->bi_iter.bi_sector <
3781 dev->sector + STRIPE_SECTORS) {
3782 wbi2 = r5_next_bio(wbi, dev->sector);
3783 md_write_end(conf->mddev);
3784 bio_endio(wbi);
3785 wbi = wbi2;
3786 }
3787 md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3788 STRIPE_SECTORS,
3789 !test_bit(STRIPE_DEGRADED, &sh->state),
3790 0);
3791 if (head_sh->batch_head) {
3792 sh = list_first_entry(&sh->batch_list,
3793 struct stripe_head,
3794 batch_list);
3795 if (sh != head_sh) {
3796 dev = &sh->dev[i];
3797 goto returnbi;
3798 }
3799 }
3800 sh = head_sh;
3801 dev = &sh->dev[i];
3802 } else if (test_bit(R5_Discard, &dev->flags))
3803 discard_pending = 1;
3804 }
3805
3806 log_stripe_write_finished(sh);
3807
3808 if (!discard_pending &&
3809 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3810 int hash;
3811 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3812 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3813 if (sh->qd_idx >= 0) {
3814 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3815 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3816 }
3817 /* now that discard is done we can proceed with any sync */
3818 clear_bit(STRIPE_DISCARD, &sh->state);
3819 /*
3820 * SCSI discard will change some bio fields and the stripe has
3821 * no updated data, so remove it from hash list and the stripe
3822 * will be reinitialized
3823 */
3824unhash:
3825 hash = sh->hash_lock_index;
3826 spin_lock_irq(conf->hash_locks + hash);
3827 remove_hash(sh);
3828 spin_unlock_irq(conf->hash_locks + hash);
3829 if (head_sh->batch_head) {
3830 sh = list_first_entry(&sh->batch_list,
3831 struct stripe_head, batch_list);
3832 if (sh != head_sh)
3833 goto unhash;
3834 }
3835 sh = head_sh;
3836
3837 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3838 set_bit(STRIPE_HANDLE, &sh->state);
3839
3840 }
3841
3842 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3843 if (atomic_dec_and_test(&conf->pending_full_writes))
3844 md_wakeup_thread(conf->mddev->thread);
3845
3846 if (head_sh->batch_head && do_endio)
3847 break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3848}
3849
3850/*
3851 * For RMW in write back cache, we need extra page in prexor to store the
3852 * old data. This page is stored in dev->orig_page.
3853 *
3854 * This function checks whether we have data for prexor. The exact logic
3855 * is:
3856 * R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3857 */
3858static inline bool uptodate_for_rmw(struct r5dev *dev)
3859{
3860 return (test_bit(R5_UPTODATE, &dev->flags)) &&
3861 (!test_bit(R5_InJournal, &dev->flags) ||
3862 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3863}
3864
3865static int handle_stripe_dirtying(struct r5conf *conf,
3866 struct stripe_head *sh,
3867 struct stripe_head_state *s,
3868 int disks)
3869{
3870 int rmw = 0, rcw = 0, i;
3871 sector_t recovery_cp = conf->mddev->recovery_cp;
3872
3873 /* Check whether resync is now happening or should start.
3874 * If yes, then the array is dirty (after unclean shutdown or
3875 * initial creation), so parity in some stripes might be inconsistent.
3876 * In this case, we need to always do reconstruct-write, to ensure
3877 * that in case of drive failure or read-error correction, we
3878 * generate correct data from the parity.
3879 */
3880 if (conf->rmw_level == PARITY_DISABLE_RMW ||
3881 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3882 s->failed == 0)) {
3883 /* Calculate the real rcw later - for now make it
3884 * look like rcw is cheaper
3885 */
3886 rcw = 1; rmw = 2;
3887 pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3888 conf->rmw_level, (unsigned long long)recovery_cp,
3889 (unsigned long long)sh->sector);
3890 } else for (i = disks; i--; ) {
3891 /* would I have to read this buffer for read_modify_write */
3892 struct r5dev *dev = &sh->dev[i];
3893 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3894 i == sh->pd_idx || i == sh->qd_idx ||
3895 test_bit(R5_InJournal, &dev->flags)) &&
3896 !test_bit(R5_LOCKED, &dev->flags) &&
3897 !(uptodate_for_rmw(dev) ||
3898 test_bit(R5_Wantcompute, &dev->flags))) {
3899 if (test_bit(R5_Insync, &dev->flags))
3900 rmw++;
3901 else
3902 rmw += 2*disks; /* cannot read it */
3903 }
3904 /* Would I have to read this buffer for reconstruct_write */
3905 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3906 i != sh->pd_idx && i != sh->qd_idx &&
3907 !test_bit(R5_LOCKED, &dev->flags) &&
3908 !(test_bit(R5_UPTODATE, &dev->flags) ||
3909 test_bit(R5_Wantcompute, &dev->flags))) {
3910 if (test_bit(R5_Insync, &dev->flags))
3911 rcw++;
3912 else
3913 rcw += 2*disks;
3914 }
3915 }
3916
3917 pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3918 (unsigned long long)sh->sector, sh->state, rmw, rcw);
3919 set_bit(STRIPE_HANDLE, &sh->state);
3920 if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3921 /* prefer read-modify-write, but need to get some data */
3922 if (conf->mddev->queue)
3923 blk_add_trace_msg(conf->mddev->queue,
3924 "raid5 rmw %llu %d",
3925 (unsigned long long)sh->sector, rmw);
3926 for (i = disks; i--; ) {
3927 struct r5dev *dev = &sh->dev[i];
3928 if (test_bit(R5_InJournal, &dev->flags) &&
3929 dev->page == dev->orig_page &&
3930 !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3931 /* alloc page for prexor */
3932 struct page *p = alloc_page(GFP_NOIO);
3933
3934 if (p) {
3935 dev->orig_page = p;
3936 continue;
3937 }
3938
3939 /*
3940 * alloc_page() failed, try use
3941 * disk_info->extra_page
3942 */
3943 if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3944 &conf->cache_state)) {
3945 r5c_use_extra_page(sh);
3946 break;
3947 }
3948
3949 /* extra_page in use, add to delayed_list */
3950 set_bit(STRIPE_DELAYED, &sh->state);
3951 s->waiting_extra_page = 1;
3952 return -EAGAIN;
3953 }
3954 }
3955
3956 for (i = disks; i--; ) {
3957 struct r5dev *dev = &sh->dev[i];
3958 if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3959 i == sh->pd_idx || i == sh->qd_idx ||
3960 test_bit(R5_InJournal, &dev->flags)) &&
3961 !test_bit(R5_LOCKED, &dev->flags) &&
3962 !(uptodate_for_rmw(dev) ||
3963 test_bit(R5_Wantcompute, &dev->flags)) &&
3964 test_bit(R5_Insync, &dev->flags)) {
3965 if (test_bit(STRIPE_PREREAD_ACTIVE,
3966 &sh->state)) {
3967 pr_debug("Read_old block %d for r-m-w\n",
3968 i);
3969 set_bit(R5_LOCKED, &dev->flags);
3970 set_bit(R5_Wantread, &dev->flags);
3971 s->locked++;
3972 } else {
3973 set_bit(STRIPE_DELAYED, &sh->state);
3974 set_bit(STRIPE_HANDLE, &sh->state);
3975 }
3976 }
3977 }
3978 }
3979 if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3980 /* want reconstruct write, but need to get some data */
3981 int qread =0;
3982 rcw = 0;
3983 for (i = disks; i--; ) {
3984 struct r5dev *dev = &sh->dev[i];
3985 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3986 i != sh->pd_idx && i != sh->qd_idx &&
3987 !test_bit(R5_LOCKED, &dev->flags) &&
3988 !(test_bit(R5_UPTODATE, &dev->flags) ||
3989 test_bit(R5_Wantcompute, &dev->flags))) {
3990 rcw++;
3991 if (test_bit(R5_Insync, &dev->flags) &&
3992 test_bit(STRIPE_PREREAD_ACTIVE,
3993 &sh->state)) {
3994 pr_debug("Read_old block "
3995 "%d for Reconstruct\n", i);
3996 set_bit(R5_LOCKED, &dev->flags);
3997 set_bit(R5_Wantread, &dev->flags);
3998 s->locked++;
3999 qread++;
4000 } else {
4001 set_bit(STRIPE_DELAYED, &sh->state);
4002 set_bit(STRIPE_HANDLE, &sh->state);
4003 }
4004 }
4005 }
4006 if (rcw && conf->mddev->queue)
4007 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4008 (unsigned long long)sh->sector,
4009 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4010 }
4011
4012 if (rcw > disks && rmw > disks &&
4013 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4014 set_bit(STRIPE_DELAYED, &sh->state);
4015
4016 /* now if nothing is locked, and if we have enough data,
4017 * we can start a write request
4018 */
4019 /* since handle_stripe can be called at any time we need to handle the
4020 * case where a compute block operation has been submitted and then a
4021 * subsequent call wants to start a write request. raid_run_ops only
4022 * handles the case where compute block and reconstruct are requested
4023 * simultaneously. If this is not the case then new writes need to be
4024 * held off until the compute completes.
4025 */
4026 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4027 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4028 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4029 schedule_reconstruction(sh, s, rcw == 0, 0);
4030 return 0;
4031}
4032
4033static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4034 struct stripe_head_state *s, int disks)
4035{
4036 struct r5dev *dev = NULL;
4037
4038 BUG_ON(sh->batch_head);
4039 set_bit(STRIPE_HANDLE, &sh->state);
4040
4041 switch (sh->check_state) {
4042 case check_state_idle:
4043 /* start a new check operation if there are no failures */
4044 if (s->failed == 0) {
4045 BUG_ON(s->uptodate != disks);
4046 sh->check_state = check_state_run;
4047 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4048 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4049 s->uptodate--;
4050 break;
4051 }
4052 dev = &sh->dev[s->failed_num[0]];
4053 /* fall through */
4054 case check_state_compute_result:
4055 sh->check_state = check_state_idle;
4056 if (!dev)
4057 dev = &sh->dev[sh->pd_idx];
4058
4059 /* check that a write has not made the stripe insync */
4060 if (test_bit(STRIPE_INSYNC, &sh->state))
4061 break;
4062
4063 /* either failed parity check, or recovery is happening */
4064 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4065 BUG_ON(s->uptodate != disks);
4066
4067 set_bit(R5_LOCKED, &dev->flags);
4068 s->locked++;
4069 set_bit(R5_Wantwrite, &dev->flags);
4070
4071 clear_bit(STRIPE_DEGRADED, &sh->state);
4072 set_bit(STRIPE_INSYNC, &sh->state);
4073 break;
4074 case check_state_run:
4075 break; /* we will be called again upon completion */
4076 case check_state_check_result:
4077 sh->check_state = check_state_idle;
4078
4079 /* if a failure occurred during the check operation, leave
4080 * STRIPE_INSYNC not set and let the stripe be handled again
4081 */
4082 if (s->failed)
4083 break;
4084
4085 /* handle a successful check operation, if parity is correct
4086 * we are done. Otherwise update the mismatch count and repair
4087 * parity if !MD_RECOVERY_CHECK
4088 */
4089 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4090 /* parity is correct (on disc,
4091 * not in buffer any more)
4092 */
4093 set_bit(STRIPE_INSYNC, &sh->state);
4094 else {
4095 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4096 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4097 /* don't try to repair!! */
4098 set_bit(STRIPE_INSYNC, &sh->state);
4099 pr_warn_ratelimited("%s: mismatch sector in range "
4100 "%llu-%llu\n", mdname(conf->mddev),
4101 (unsigned long long) sh->sector,
4102 (unsigned long long) sh->sector +
4103 STRIPE_SECTORS);
4104 } else {
4105 sh->check_state = check_state_compute_run;
4106 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4107 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4108 set_bit(R5_Wantcompute,
4109 &sh->dev[sh->pd_idx].flags);
4110 sh->ops.target = sh->pd_idx;
4111 sh->ops.target2 = -1;
4112 s->uptodate++;
4113 }
4114 }
4115 break;
4116 case check_state_compute_run:
4117 break;
4118 default:
4119 pr_err("%s: unknown check_state: %d sector: %llu\n",
4120 __func__, sh->check_state,
4121 (unsigned long long) sh->sector);
4122 BUG();
4123 }
4124}
4125
4126static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4127 struct stripe_head_state *s,
4128 int disks)
4129{
4130 int pd_idx = sh->pd_idx;
4131 int qd_idx = sh->qd_idx;
4132 struct r5dev *dev;
4133
4134 BUG_ON(sh->batch_head);
4135 set_bit(STRIPE_HANDLE, &sh->state);
4136
4137 BUG_ON(s->failed > 2);
4138
4139 /* Want to check and possibly repair P and Q.
4140 * However there could be one 'failed' device, in which
4141 * case we can only check one of them, possibly using the
4142 * other to generate missing data
4143 */
4144
4145 switch (sh->check_state) {
4146 case check_state_idle:
4147 /* start a new check operation if there are < 2 failures */
4148 if (s->failed == s->q_failed) {
4149 /* The only possible failed device holds Q, so it
4150 * makes sense to check P (If anything else were failed,
4151 * we would have used P to recreate it).
4152 */
4153 sh->check_state = check_state_run;
4154 }
4155 if (!s->q_failed && s->failed < 2) {
4156 /* Q is not failed, and we didn't use it to generate
4157 * anything, so it makes sense to check it
4158 */
4159 if (sh->check_state == check_state_run)
4160 sh->check_state = check_state_run_pq;
4161 else
4162 sh->check_state = check_state_run_q;
4163 }
4164
4165 /* discard potentially stale zero_sum_result */
4166 sh->ops.zero_sum_result = 0;
4167
4168 if (sh->check_state == check_state_run) {
4169 /* async_xor_zero_sum destroys the contents of P */
4170 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4171 s->uptodate--;
4172 }
4173 if (sh->check_state >= check_state_run &&
4174 sh->check_state <= check_state_run_pq) {
4175 /* async_syndrome_zero_sum preserves P and Q, so
4176 * no need to mark them !uptodate here
4177 */
4178 set_bit(STRIPE_OP_CHECK, &s->ops_request);
4179 break;
4180 }
4181
4182 /* we have 2-disk failure */
4183 BUG_ON(s->failed != 2);
4184 /* fall through */
4185 case check_state_compute_result:
4186 sh->check_state = check_state_idle;
4187
4188 /* check that a write has not made the stripe insync */
4189 if (test_bit(STRIPE_INSYNC, &sh->state))
4190 break;
4191
4192 /* now write out any block on a failed drive,
4193 * or P or Q if they were recomputed
4194 */
4195 dev = NULL;
4196 if (s->failed == 2) {
4197 dev = &sh->dev[s->failed_num[1]];
4198 s->locked++;
4199 set_bit(R5_LOCKED, &dev->flags);
4200 set_bit(R5_Wantwrite, &dev->flags);
4201 }
4202 if (s->failed >= 1) {
4203 dev = &sh->dev[s->failed_num[0]];
4204 s->locked++;
4205 set_bit(R5_LOCKED, &dev->flags);
4206 set_bit(R5_Wantwrite, &dev->flags);
4207 }
4208 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4209 dev = &sh->dev[pd_idx];
4210 s->locked++;
4211 set_bit(R5_LOCKED, &dev->flags);
4212 set_bit(R5_Wantwrite, &dev->flags);
4213 }
4214 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4215 dev = &sh->dev[qd_idx];
4216 s->locked++;
4217 set_bit(R5_LOCKED, &dev->flags);
4218 set_bit(R5_Wantwrite, &dev->flags);
4219 }
4220 if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4221 "%s: disk%td not up to date\n",
4222 mdname(conf->mddev),
4223 dev - (struct r5dev *) &sh->dev)) {
4224 clear_bit(R5_LOCKED, &dev->flags);
4225 clear_bit(R5_Wantwrite, &dev->flags);
4226 s->locked--;
4227 }
4228 clear_bit(STRIPE_DEGRADED, &sh->state);
4229
4230 set_bit(STRIPE_INSYNC, &sh->state);
4231 break;
4232 case check_state_run:
4233 case check_state_run_q:
4234 case check_state_run_pq:
4235 break; /* we will be called again upon completion */
4236 case check_state_check_result:
4237 sh->check_state = check_state_idle;
4238
4239 /* handle a successful check operation, if parity is correct
4240 * we are done. Otherwise update the mismatch count and repair
4241 * parity if !MD_RECOVERY_CHECK
4242 */
4243 if (sh->ops.zero_sum_result == 0) {
4244 /* both parities are correct */
4245 if (!s->failed)
4246 set_bit(STRIPE_INSYNC, &sh->state);
4247 else {
4248 /* in contrast to the raid5 case we can validate
4249 * parity, but still have a failure to write
4250 * back
4251 */
4252 sh->check_state = check_state_compute_result;
4253 /* Returning at this point means that we may go
4254 * off and bring p and/or q uptodate again so
4255 * we make sure to check zero_sum_result again
4256 * to verify if p or q need writeback
4257 */
4258 }
4259 } else {
4260 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4261 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4262 /* don't try to repair!! */
4263 set_bit(STRIPE_INSYNC, &sh->state);
4264 pr_warn_ratelimited("%s: mismatch sector in range "
4265 "%llu-%llu\n", mdname(conf->mddev),
4266 (unsigned long long) sh->sector,
4267 (unsigned long long) sh->sector +
4268 STRIPE_SECTORS);
4269 } else {
4270 int *target = &sh->ops.target;
4271
4272 sh->ops.target = -1;
4273 sh->ops.target2 = -1;
4274 sh->check_state = check_state_compute_run;
4275 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4276 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4277 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4278 set_bit(R5_Wantcompute,
4279 &sh->dev[pd_idx].flags);
4280 *target = pd_idx;
4281 target = &sh->ops.target2;
4282 s->uptodate++;
4283 }
4284 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4285 set_bit(R5_Wantcompute,
4286 &sh->dev[qd_idx].flags);
4287 *target = qd_idx;
4288 s->uptodate++;
4289 }
4290 }
4291 }
4292 break;
4293 case check_state_compute_run:
4294 break;
4295 default:
4296 pr_warn("%s: unknown check_state: %d sector: %llu\n",
4297 __func__, sh->check_state,
4298 (unsigned long long) sh->sector);
4299 BUG();
4300 }
4301}
4302
4303static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4304{
4305 int i;
4306
4307 /* We have read all the blocks in this stripe and now we need to
4308 * copy some of them into a target stripe for expand.
4309 */
4310 struct dma_async_tx_descriptor *tx = NULL;
4311 BUG_ON(sh->batch_head);
4312 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4313 for (i = 0; i < sh->disks; i++)
4314 if (i != sh->pd_idx && i != sh->qd_idx) {
4315 int dd_idx, j;
4316 struct stripe_head *sh2;
4317 struct async_submit_ctl submit;
4318
4319 sector_t bn = raid5_compute_blocknr(sh, i, 1);
4320 sector_t s = raid5_compute_sector(conf, bn, 0,
4321 &dd_idx, NULL);
4322 sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4323 if (sh2 == NULL)
4324 /* so far only the early blocks of this stripe
4325 * have been requested. When later blocks
4326 * get requested, we will try again
4327 */
4328 continue;
4329 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4330 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4331 /* must have already done this block */
4332 raid5_release_stripe(sh2);
4333 continue;
4334 }
4335
4336 /* place all the copies on one channel */
4337 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4338 tx = async_memcpy(sh2->dev[dd_idx].page,
4339 sh->dev[i].page, 0, 0, STRIPE_SIZE,
4340 &submit);
4341
4342 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4343 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4344 for (j = 0; j < conf->raid_disks; j++)
4345 if (j != sh2->pd_idx &&
4346 j != sh2->qd_idx &&
4347 !test_bit(R5_Expanded, &sh2->dev[j].flags))
4348 break;
4349 if (j == conf->raid_disks) {
4350 set_bit(STRIPE_EXPAND_READY, &sh2->state);
4351 set_bit(STRIPE_HANDLE, &sh2->state);
4352 }
4353 raid5_release_stripe(sh2);
4354
4355 }
4356 /* done submitting copies, wait for them to complete */
4357 async_tx_quiesce(&tx);
4358}
4359
4360/*
4361 * handle_stripe - do things to a stripe.
4362 *
4363 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4364 * state of various bits to see what needs to be done.
4365 * Possible results:
4366 * return some read requests which now have data
4367 * return some write requests which are safely on storage
4368 * schedule a read on some buffers
4369 * schedule a write of some buffers
4370 * return confirmation of parity correctness
4371 *
4372 */
4373
4374static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4375{
4376 struct r5conf *conf = sh->raid_conf;
4377 int disks = sh->disks;
4378 struct r5dev *dev;
4379 int i;
4380 int do_recovery = 0;
4381
4382 memset(s, 0, sizeof(*s));
4383
4384 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4385 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4386 s->failed_num[0] = -1;
4387 s->failed_num[1] = -1;
4388 s->log_failed = r5l_log_disk_error(conf);
4389
4390 /* Now to look around and see what can be done */
4391 rcu_read_lock();
4392 for (i=disks; i--; ) {
4393 struct md_rdev *rdev;
4394 sector_t first_bad;
4395 int bad_sectors;
4396 int is_bad = 0;
4397
4398 dev = &sh->dev[i];
4399
4400 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4401 i, dev->flags,
4402 dev->toread, dev->towrite, dev->written);
4403 /* maybe we can reply to a read
4404 *
4405 * new wantfill requests are only permitted while
4406 * ops_complete_biofill is guaranteed to be inactive
4407 */
4408 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4409 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4410 set_bit(R5_Wantfill, &dev->flags);
4411
4412 /* now count some things */
4413 if (test_bit(R5_LOCKED, &dev->flags))
4414 s->locked++;
4415 if (test_bit(R5_UPTODATE, &dev->flags))
4416 s->uptodate++;
4417 if (test_bit(R5_Wantcompute, &dev->flags)) {
4418 s->compute++;
4419 BUG_ON(s->compute > 2);
4420 }
4421
4422 if (test_bit(R5_Wantfill, &dev->flags))
4423 s->to_fill++;
4424 else if (dev->toread)
4425 s->to_read++;
4426 if (dev->towrite) {
4427 s->to_write++;
4428 if (!test_bit(R5_OVERWRITE, &dev->flags))
4429 s->non_overwrite++;
4430 }
4431 if (dev->written)
4432 s->written++;
4433 /* Prefer to use the replacement for reads, but only
4434 * if it is recovered enough and has no bad blocks.
4435 */
4436 rdev = rcu_dereference(conf->disks[i].replacement);
4437 if (rdev && !test_bit(Faulty, &rdev->flags) &&
4438 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4439 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4440 &first_bad, &bad_sectors))
4441 set_bit(R5_ReadRepl, &dev->flags);
4442 else {
4443 if (rdev && !test_bit(Faulty, &rdev->flags))
4444 set_bit(R5_NeedReplace, &dev->flags);
4445 else
4446 clear_bit(R5_NeedReplace, &dev->flags);
4447 rdev = rcu_dereference(conf->disks[i].rdev);
4448 clear_bit(R5_ReadRepl, &dev->flags);
4449 }
4450 if (rdev && test_bit(Faulty, &rdev->flags))
4451 rdev = NULL;
4452 if (rdev) {
4453 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4454 &first_bad, &bad_sectors);
4455 if (s->blocked_rdev == NULL
4456 && (test_bit(Blocked, &rdev->flags)
4457 || is_bad < 0)) {
4458 if (is_bad < 0)
4459 set_bit(BlockedBadBlocks,
4460 &rdev->flags);
4461 s->blocked_rdev = rdev;
4462 atomic_inc(&rdev->nr_pending);
4463 }
4464 }
4465 clear_bit(R5_Insync, &dev->flags);
4466 if (!rdev)
4467 /* Not in-sync */;
4468 else if (is_bad) {
4469 /* also not in-sync */
4470 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4471 test_bit(R5_UPTODATE, &dev->flags)) {
4472 /* treat as in-sync, but with a read error
4473 * which we can now try to correct
4474 */
4475 set_bit(R5_Insync, &dev->flags);
4476 set_bit(R5_ReadError, &dev->flags);
4477 }
4478 } else if (test_bit(In_sync, &rdev->flags))
4479 set_bit(R5_Insync, &dev->flags);
4480 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4481 /* in sync if before recovery_offset */
4482 set_bit(R5_Insync, &dev->flags);
4483 else if (test_bit(R5_UPTODATE, &dev->flags) &&
4484 test_bit(R5_Expanded, &dev->flags))
4485 /* If we've reshaped into here, we assume it is Insync.
4486 * We will shortly update recovery_offset to make
4487 * it official.
4488 */
4489 set_bit(R5_Insync, &dev->flags);
4490
4491 if (test_bit(R5_WriteError, &dev->flags)) {
4492 /* This flag does not apply to '.replacement'
4493 * only to .rdev, so make sure to check that*/
4494 struct md_rdev *rdev2 = rcu_dereference(
4495 conf->disks[i].rdev);
4496 if (rdev2 == rdev)
4497 clear_bit(R5_Insync, &dev->flags);
4498 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4499 s->handle_bad_blocks = 1;
4500 atomic_inc(&rdev2->nr_pending);
4501 } else
4502 clear_bit(R5_WriteError, &dev->flags);
4503 }
4504 if (test_bit(R5_MadeGood, &dev->flags)) {
4505 /* This flag does not apply to '.replacement'
4506 * only to .rdev, so make sure to check that*/
4507 struct md_rdev *rdev2 = rcu_dereference(
4508 conf->disks[i].rdev);
4509 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4510 s->handle_bad_blocks = 1;
4511 atomic_inc(&rdev2->nr_pending);
4512 } else
4513 clear_bit(R5_MadeGood, &dev->flags);
4514 }
4515 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4516 struct md_rdev *rdev2 = rcu_dereference(
4517 conf->disks[i].replacement);
4518 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4519 s->handle_bad_blocks = 1;
4520 atomic_inc(&rdev2->nr_pending);
4521 } else
4522 clear_bit(R5_MadeGoodRepl, &dev->flags);
4523 }
4524 if (!test_bit(R5_Insync, &dev->flags)) {
4525 /* The ReadError flag will just be confusing now */
4526 clear_bit(R5_ReadError, &dev->flags);
4527 clear_bit(R5_ReWrite, &dev->flags);
4528 }
4529 if (test_bit(R5_ReadError, &dev->flags))
4530 clear_bit(R5_Insync, &dev->flags);
4531 if (!test_bit(R5_Insync, &dev->flags)) {
4532 if (s->failed < 2)
4533 s->failed_num[s->failed] = i;
4534 s->failed++;
4535 if (rdev && !test_bit(Faulty, &rdev->flags))
4536 do_recovery = 1;
4537 else if (!rdev) {
4538 rdev = rcu_dereference(
4539 conf->disks[i].replacement);
4540 if (rdev && !test_bit(Faulty, &rdev->flags))
4541 do_recovery = 1;
4542 }
4543 }
4544
4545 if (test_bit(R5_InJournal, &dev->flags))
4546 s->injournal++;
4547 if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4548 s->just_cached++;
4549 }
4550 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4551 /* If there is a failed device being replaced,
4552 * we must be recovering.
4553 * else if we are after recovery_cp, we must be syncing
4554 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4555 * else we can only be replacing
4556 * sync and recovery both need to read all devices, and so
4557 * use the same flag.
4558 */
4559 if (do_recovery ||
4560 sh->sector >= conf->mddev->recovery_cp ||
4561 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4562 s->syncing = 1;
4563 else
4564 s->replacing = 1;
4565 }
4566 rcu_read_unlock();
4567}
4568
4569static int clear_batch_ready(struct stripe_head *sh)
4570{
4571 /* Return '1' if this is a member of batch, or
4572 * '0' if it is a lone stripe or a head which can now be
4573 * handled.
4574 */
4575 struct stripe_head *tmp;
4576 if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4577 return (sh->batch_head && sh->batch_head != sh);
4578 spin_lock(&sh->stripe_lock);
4579 if (!sh->batch_head) {
4580 spin_unlock(&sh->stripe_lock);
4581 return 0;
4582 }
4583
4584 /*
4585 * this stripe could be added to a batch list before we check
4586 * BATCH_READY, skips it
4587 */
4588 if (sh->batch_head != sh) {
4589 spin_unlock(&sh->stripe_lock);
4590 return 1;
4591 }
4592 spin_lock(&sh->batch_lock);
4593 list_for_each_entry(tmp, &sh->batch_list, batch_list)
4594 clear_bit(STRIPE_BATCH_READY, &tmp->state);
4595 spin_unlock(&sh->batch_lock);
4596 spin_unlock(&sh->stripe_lock);
4597
4598 /*
4599 * BATCH_READY is cleared, no new stripes can be added.
4600 * batch_list can be accessed without lock
4601 */
4602 return 0;
4603}
4604
4605static void break_stripe_batch_list(struct stripe_head *head_sh,
4606 unsigned long handle_flags)
4607{
4608 struct stripe_head *sh, *next;
4609 int i;
4610 int do_wakeup = 0;
4611
4612 list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4613
4614 list_del_init(&sh->batch_list);
4615
4616 WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4617 (1 << STRIPE_SYNCING) |
4618 (1 << STRIPE_REPLACED) |
4619 (1 << STRIPE_DELAYED) |
4620 (1 << STRIPE_BIT_DELAY) |
4621 (1 << STRIPE_FULL_WRITE) |
4622 (1 << STRIPE_BIOFILL_RUN) |
4623 (1 << STRIPE_COMPUTE_RUN) |
4624 (1 << STRIPE_DISCARD) |
4625 (1 << STRIPE_BATCH_READY) |
4626 (1 << STRIPE_BATCH_ERR) |
4627 (1 << STRIPE_BITMAP_PENDING)),
4628 "stripe state: %lx\n", sh->state);
4629 WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4630 (1 << STRIPE_REPLACED)),
4631 "head stripe state: %lx\n", head_sh->state);
4632
4633 set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4634 (1 << STRIPE_PREREAD_ACTIVE) |
4635 (1 << STRIPE_DEGRADED) |
4636 (1 << STRIPE_ON_UNPLUG_LIST)),
4637 head_sh->state & (1 << STRIPE_INSYNC));
4638
4639 sh->check_state = head_sh->check_state;
4640 sh->reconstruct_state = head_sh->reconstruct_state;
4641 spin_lock_irq(&sh->stripe_lock);
4642 sh->batch_head = NULL;
4643 spin_unlock_irq(&sh->stripe_lock);
4644 for (i = 0; i < sh->disks; i++) {
4645 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4646 do_wakeup = 1;
4647 sh->dev[i].flags = head_sh->dev[i].flags &
4648 (~((1 << R5_WriteError) | (1 << R5_Overlap)));
4649 }
4650 if (handle_flags == 0 ||
4651 sh->state & handle_flags)
4652 set_bit(STRIPE_HANDLE, &sh->state);
4653 raid5_release_stripe(sh);
4654 }
4655 spin_lock_irq(&head_sh->stripe_lock);
4656 head_sh->batch_head = NULL;
4657 spin_unlock_irq(&head_sh->stripe_lock);
4658 for (i = 0; i < head_sh->disks; i++)
4659 if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4660 do_wakeup = 1;
4661 if (head_sh->state & handle_flags)
4662 set_bit(STRIPE_HANDLE, &head_sh->state);
4663
4664 if (do_wakeup)
4665 wake_up(&head_sh->raid_conf->wait_for_overlap);
4666}
4667
4668static void handle_stripe(struct stripe_head *sh)
4669{
4670 struct stripe_head_state s;
4671 struct r5conf *conf = sh->raid_conf;
4672 int i;
4673 int prexor;
4674 int disks = sh->disks;
4675 struct r5dev *pdev, *qdev;
4676
4677 clear_bit(STRIPE_HANDLE, &sh->state);
4678 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4679 /* already being handled, ensure it gets handled
4680 * again when current action finishes */
4681 set_bit(STRIPE_HANDLE, &sh->state);
4682 return;
4683 }
4684
4685 if (clear_batch_ready(sh) ) {
4686 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4687 return;
4688 }
4689
4690 if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4691 break_stripe_batch_list(sh, 0);
4692
4693 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4694 spin_lock(&sh->stripe_lock);
4695 /*
4696 * Cannot process 'sync' concurrently with 'discard'.
4697 * Flush data in r5cache before 'sync'.
4698 */
4699 if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4700 !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4701 !test_bit(STRIPE_DISCARD, &sh->state) &&
4702 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4703 set_bit(STRIPE_SYNCING, &sh->state);
4704 clear_bit(STRIPE_INSYNC, &sh->state);
4705 clear_bit(STRIPE_REPLACED, &sh->state);
4706 }
4707 spin_unlock(&sh->stripe_lock);
4708 }
4709 clear_bit(STRIPE_DELAYED, &sh->state);
4710
4711 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4712 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4713 (unsigned long long)sh->sector, sh->state,
4714 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4715 sh->check_state, sh->reconstruct_state);
4716
4717 analyse_stripe(sh, &s);
4718
4719 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4720 goto finish;
4721
4722 if (s.handle_bad_blocks ||
4723 test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4724 set_bit(STRIPE_HANDLE, &sh->state);
4725 goto finish;
4726 }
4727
4728 if (unlikely(s.blocked_rdev)) {
4729 if (s.syncing || s.expanding || s.expanded ||
4730 s.replacing || s.to_write || s.written) {
4731 set_bit(STRIPE_HANDLE, &sh->state);
4732 goto finish;
4733 }
4734 /* There is nothing for the blocked_rdev to block */
4735 rdev_dec_pending(s.blocked_rdev, conf->mddev);
4736 s.blocked_rdev = NULL;
4737 }
4738
4739 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4740 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4741 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4742 }
4743
4744 pr_debug("locked=%d uptodate=%d to_read=%d"
4745 " to_write=%d failed=%d failed_num=%d,%d\n",
4746 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4747 s.failed_num[0], s.failed_num[1]);
4748 /*
4749 * check if the array has lost more than max_degraded devices and,
4750 * if so, some requests might need to be failed.
4751 *
4752 * When journal device failed (log_failed), we will only process
4753 * the stripe if there is data need write to raid disks
4754 */
4755 if (s.failed > conf->max_degraded ||
4756 (s.log_failed && s.injournal == 0)) {
4757 sh->check_state = 0;
4758 sh->reconstruct_state = 0;
4759 break_stripe_batch_list(sh, 0);
4760 if (s.to_read+s.to_write+s.written)
4761 handle_failed_stripe(conf, sh, &s, disks);
4762 if (s.syncing + s.replacing)
4763 handle_failed_sync(conf, sh, &s);
4764 }
4765
4766 /* Now we check to see if any write operations have recently
4767 * completed
4768 */
4769 prexor = 0;
4770 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4771 prexor = 1;
4772 if (sh->reconstruct_state == reconstruct_state_drain_result ||
4773 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4774 sh->reconstruct_state = reconstruct_state_idle;
4775
4776 /* All the 'written' buffers and the parity block are ready to
4777 * be written back to disk
4778 */
4779 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4780 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4781 BUG_ON(sh->qd_idx >= 0 &&
4782 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4783 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4784 for (i = disks; i--; ) {
4785 struct r5dev *dev = &sh->dev[i];
4786 if (test_bit(R5_LOCKED, &dev->flags) &&
4787 (i == sh->pd_idx || i == sh->qd_idx ||
4788 dev->written || test_bit(R5_InJournal,
4789 &dev->flags))) {
4790 pr_debug("Writing block %d\n", i);
4791 set_bit(R5_Wantwrite, &dev->flags);
4792 if (prexor)
4793 continue;
4794 if (s.failed > 1)
4795 continue;
4796 if (!test_bit(R5_Insync, &dev->flags) ||
4797 ((i == sh->pd_idx || i == sh->qd_idx) &&
4798 s.failed == 0))
4799 set_bit(STRIPE_INSYNC, &sh->state);
4800 }
4801 }
4802 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4803 s.dec_preread_active = 1;
4804 }
4805
4806 /*
4807 * might be able to return some write requests if the parity blocks
4808 * are safe, or on a failed drive
4809 */
4810 pdev = &sh->dev[sh->pd_idx];
4811 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4812 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4813 qdev = &sh->dev[sh->qd_idx];
4814 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4815 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4816 || conf->level < 6;
4817
4818 if (s.written &&
4819 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4820 && !test_bit(R5_LOCKED, &pdev->flags)
4821 && (test_bit(R5_UPTODATE, &pdev->flags) ||
4822 test_bit(R5_Discard, &pdev->flags))))) &&
4823 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4824 && !test_bit(R5_LOCKED, &qdev->flags)
4825 && (test_bit(R5_UPTODATE, &qdev->flags) ||
4826 test_bit(R5_Discard, &qdev->flags))))))
4827 handle_stripe_clean_event(conf, sh, disks);
4828
4829 if (s.just_cached)
4830 r5c_handle_cached_data_endio(conf, sh, disks);
4831 log_stripe_write_finished(sh);
4832
4833 /* Now we might consider reading some blocks, either to check/generate
4834 * parity, or to satisfy requests
4835 * or to load a block that is being partially written.
4836 */
4837 if (s.to_read || s.non_overwrite
4838 || (conf->level == 6 && s.to_write && s.failed)
4839 || (s.syncing && (s.uptodate + s.compute < disks))
4840 || s.replacing
4841 || s.expanding)
4842 handle_stripe_fill(sh, &s, disks);
4843
4844 /*
4845 * When the stripe finishes full journal write cycle (write to journal
4846 * and raid disk), this is the clean up procedure so it is ready for
4847 * next operation.
4848 */
4849 r5c_finish_stripe_write_out(conf, sh, &s);
4850
4851 /*
4852 * Now to consider new write requests, cache write back and what else,
4853 * if anything should be read. We do not handle new writes when:
4854 * 1/ A 'write' operation (copy+xor) is already in flight.
4855 * 2/ A 'check' operation is in flight, as it may clobber the parity
4856 * block.
4857 * 3/ A r5c cache log write is in flight.
4858 */
4859
4860 if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4861 if (!r5c_is_writeback(conf->log)) {
4862 if (s.to_write)
4863 handle_stripe_dirtying(conf, sh, &s, disks);
4864 } else { /* write back cache */
4865 int ret = 0;
4866
4867 /* First, try handle writes in caching phase */
4868 if (s.to_write)
4869 ret = r5c_try_caching_write(conf, sh, &s,
4870 disks);
4871 /*
4872 * If caching phase failed: ret == -EAGAIN
4873 * OR
4874 * stripe under reclaim: !caching && injournal
4875 *
4876 * fall back to handle_stripe_dirtying()
4877 */
4878 if (ret == -EAGAIN ||
4879 /* stripe under reclaim: !caching && injournal */
4880 (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4881 s.injournal > 0)) {
4882 ret = handle_stripe_dirtying(conf, sh, &s,
4883 disks);
4884 if (ret == -EAGAIN)
4885 goto finish;
4886 }
4887 }
4888 }
4889
4890 /* maybe we need to check and possibly fix the parity for this stripe
4891 * Any reads will already have been scheduled, so we just see if enough
4892 * data is available. The parity check is held off while parity
4893 * dependent operations are in flight.
4894 */
4895 if (sh->check_state ||
4896 (s.syncing && s.locked == 0 &&
4897 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4898 !test_bit(STRIPE_INSYNC, &sh->state))) {
4899 if (conf->level == 6)
4900 handle_parity_checks6(conf, sh, &s, disks);
4901 else
4902 handle_parity_checks5(conf, sh, &s, disks);
4903 }
4904
4905 if ((s.replacing || s.syncing) && s.locked == 0
4906 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4907 && !test_bit(STRIPE_REPLACED, &sh->state)) {
4908 /* Write out to replacement devices where possible */
4909 for (i = 0; i < conf->raid_disks; i++)
4910 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4911 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4912 set_bit(R5_WantReplace, &sh->dev[i].flags);
4913 set_bit(R5_LOCKED, &sh->dev[i].flags);
4914 s.locked++;
4915 }
4916 if (s.replacing)
4917 set_bit(STRIPE_INSYNC, &sh->state);
4918 set_bit(STRIPE_REPLACED, &sh->state);
4919 }
4920 if ((s.syncing || s.replacing) && s.locked == 0 &&
4921 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4922 test_bit(STRIPE_INSYNC, &sh->state)) {
4923 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4924 clear_bit(STRIPE_SYNCING, &sh->state);
4925 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4926 wake_up(&conf->wait_for_overlap);
4927 }
4928
4929 /* If the failed drives are just a ReadError, then we might need
4930 * to progress the repair/check process
4931 */
4932 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4933 for (i = 0; i < s.failed; i++) {
4934 struct r5dev *dev = &sh->dev[s.failed_num[i]];
4935 if (test_bit(R5_ReadError, &dev->flags)
4936 && !test_bit(R5_LOCKED, &dev->flags)
4937 && test_bit(R5_UPTODATE, &dev->flags)
4938 ) {
4939 if (!test_bit(R5_ReWrite, &dev->flags)) {
4940 set_bit(R5_Wantwrite, &dev->flags);
4941 set_bit(R5_ReWrite, &dev->flags);
4942 set_bit(R5_LOCKED, &dev->flags);
4943 s.locked++;
4944 } else {
4945 /* let's read it back */
4946 set_bit(R5_Wantread, &dev->flags);
4947 set_bit(R5_LOCKED, &dev->flags);
4948 s.locked++;
4949 }
4950 }
4951 }
4952
4953 /* Finish reconstruct operations initiated by the expansion process */
4954 if (sh->reconstruct_state == reconstruct_state_result) {
4955 struct stripe_head *sh_src
4956 = raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4957 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4958 /* sh cannot be written until sh_src has been read.
4959 * so arrange for sh to be delayed a little
4960 */
4961 set_bit(STRIPE_DELAYED, &sh->state);
4962 set_bit(STRIPE_HANDLE, &sh->state);
4963 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4964 &sh_src->state))
4965 atomic_inc(&conf->preread_active_stripes);
4966 raid5_release_stripe(sh_src);
4967 goto finish;
4968 }
4969 if (sh_src)
4970 raid5_release_stripe(sh_src);
4971
4972 sh->reconstruct_state = reconstruct_state_idle;
4973 clear_bit(STRIPE_EXPANDING, &sh->state);
4974 for (i = conf->raid_disks; i--; ) {
4975 set_bit(R5_Wantwrite, &sh->dev[i].flags);
4976 set_bit(R5_LOCKED, &sh->dev[i].flags);
4977 s.locked++;
4978 }
4979 }
4980
4981 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4982 !sh->reconstruct_state) {
4983 /* Need to write out all blocks after computing parity */
4984 sh->disks = conf->raid_disks;
4985 stripe_set_idx(sh->sector, conf, 0, sh);
4986 schedule_reconstruction(sh, &s, 1, 1);
4987 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4988 clear_bit(STRIPE_EXPAND_READY, &sh->state);
4989 atomic_dec(&conf->reshape_stripes);
4990 wake_up(&conf->wait_for_overlap);
4991 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4992 }
4993
4994 if (s.expanding && s.locked == 0 &&
4995 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4996 handle_stripe_expansion(conf, sh);
4997
4998finish:
4999 /* wait for this device to become unblocked */
5000 if (unlikely(s.blocked_rdev)) {
5001 if (conf->mddev->external)
5002 md_wait_for_blocked_rdev(s.blocked_rdev,
5003 conf->mddev);
5004 else
5005 /* Internal metadata will immediately
5006 * be written by raid5d, so we don't
5007 * need to wait here.
5008 */
5009 rdev_dec_pending(s.blocked_rdev,
5010 conf->mddev);
5011 }
5012
5013 if (s.handle_bad_blocks)
5014 for (i = disks; i--; ) {
5015 struct md_rdev *rdev;
5016 struct r5dev *dev = &sh->dev[i];
5017 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5018 /* We own a safe reference to the rdev */
5019 rdev = conf->disks[i].rdev;
5020 if (!rdev_set_badblocks(rdev, sh->sector,
5021 STRIPE_SECTORS, 0))
5022 md_error(conf->mddev, rdev);
5023 rdev_dec_pending(rdev, conf->mddev);
5024 }
5025 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5026 rdev = conf->disks[i].rdev;
5027 rdev_clear_badblocks(rdev, sh->sector,
5028 STRIPE_SECTORS, 0);
5029 rdev_dec_pending(rdev, conf->mddev);
5030 }
5031 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5032 rdev = conf->disks[i].replacement;
5033 if (!rdev)
5034 /* rdev have been moved down */
5035 rdev = conf->disks[i].rdev;
5036 rdev_clear_badblocks(rdev, sh->sector,
5037 STRIPE_SECTORS, 0);
5038 rdev_dec_pending(rdev, conf->mddev);
5039 }
5040 }
5041
5042 if (s.ops_request)
5043 raid_run_ops(sh, s.ops_request);
5044
5045 ops_run_io(sh, &s);
5046
5047 if (s.dec_preread_active) {
5048 /* We delay this until after ops_run_io so that if make_request
5049 * is waiting on a flush, it won't continue until the writes
5050 * have actually been submitted.
5051 */
5052 atomic_dec(&conf->preread_active_stripes);
5053 if (atomic_read(&conf->preread_active_stripes) <
5054 IO_THRESHOLD)
5055 md_wakeup_thread(conf->mddev->thread);
5056 }
5057
5058 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5059}
5060
5061static void raid5_activate_delayed(struct r5conf *conf)
5062{
5063 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5064 while (!list_empty(&conf->delayed_list)) {
5065 struct list_head *l = conf->delayed_list.next;
5066 struct stripe_head *sh;
5067 sh = list_entry(l, struct stripe_head, lru);
5068 list_del_init(l);
5069 clear_bit(STRIPE_DELAYED, &sh->state);
5070 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5071 atomic_inc(&conf->preread_active_stripes);
5072 list_add_tail(&sh->lru, &conf->hold_list);
5073 raid5_wakeup_stripe_thread(sh);
5074 }
5075 }
5076}
5077
5078static void activate_bit_delay(struct r5conf *conf,
5079 struct list_head *temp_inactive_list)
5080{
5081 /* device_lock is held */
5082 struct list_head head;
5083 list_add(&head, &conf->bitmap_list);
5084 list_del_init(&conf->bitmap_list);
5085 while (!list_empty(&head)) {
5086 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5087 int hash;
5088 list_del_init(&sh->lru);
5089 atomic_inc(&sh->count);
5090 hash = sh->hash_lock_index;
5091 __release_stripe(conf, sh, &temp_inactive_list[hash]);
5092 }
5093}
5094
5095static int raid5_congested(struct mddev *mddev, int bits)
5096{
5097 struct r5conf *conf = mddev->private;
5098
5099 /* No difference between reads and writes. Just check
5100 * how busy the stripe_cache is
5101 */
5102
5103 if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5104 return 1;
5105
5106 /* Also checks whether there is pressure on r5cache log space */
5107 if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5108 return 1;
5109 if (conf->quiesce)
5110 return 1;
5111 if (atomic_read(&conf->empty_inactive_list_nr))
5112 return 1;
5113
5114 return 0;
5115}
5116
5117static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5118{
5119 struct r5conf *conf = mddev->private;
5120 sector_t sector = bio->bi_iter.bi_sector;
5121 unsigned int chunk_sectors;
5122 unsigned int bio_sectors = bio_sectors(bio);
5123
5124 WARN_ON_ONCE(bio->bi_partno);
5125
5126 chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5127 return chunk_sectors >=
5128 ((sector & (chunk_sectors - 1)) + bio_sectors);
5129}
5130
5131/*
5132 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
5133 * later sampled by raid5d.
5134 */
5135static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5136{
5137 unsigned long flags;
5138
5139 spin_lock_irqsave(&conf->device_lock, flags);
5140
5141 bi->bi_next = conf->retry_read_aligned_list;
5142 conf->retry_read_aligned_list = bi;
5143
5144 spin_unlock_irqrestore(&conf->device_lock, flags);
5145 md_wakeup_thread(conf->mddev->thread);
5146}
5147
5148static struct bio *remove_bio_from_retry(struct r5conf *conf,
5149 unsigned int *offset)
5150{
5151 struct bio *bi;
5152
5153 bi = conf->retry_read_aligned;
5154 if (bi) {
5155 *offset = conf->retry_read_offset;
5156 conf->retry_read_aligned = NULL;
5157 return bi;
5158 }
5159 bi = conf->retry_read_aligned_list;
5160 if(bi) {
5161 conf->retry_read_aligned_list = bi->bi_next;
5162 bi->bi_next = NULL;
5163 *offset = 0;
5164 }
5165
5166 return bi;
5167}
5168
5169/*
5170 * The "raid5_align_endio" should check if the read succeeded and if it
5171 * did, call bio_endio on the original bio (having bio_put the new bio
5172 * first).
5173 * If the read failed..
5174 */
5175static void raid5_align_endio(struct bio *bi)
5176{
5177 struct bio* raid_bi = bi->bi_private;
5178 struct mddev *mddev;
5179 struct r5conf *conf;
5180 struct md_rdev *rdev;
5181 blk_status_t error = bi->bi_status;
5182
5183 bio_put(bi);
5184
5185 rdev = (void*)raid_bi->bi_next;
5186 raid_bi->bi_next = NULL;
5187 mddev = rdev->mddev;
5188 conf = mddev->private;
5189
5190 rdev_dec_pending(rdev, conf->mddev);
5191
5192 if (!error) {
5193 bio_endio(raid_bi);
5194 if (atomic_dec_and_test(&conf->active_aligned_reads))
5195 wake_up(&conf->wait_for_quiescent);
5196 return;
5197 }
5198
5199 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5200
5201 add_bio_to_retry(raid_bi, conf);
5202}
5203
5204static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5205{
5206 struct r5conf *conf = mddev->private;
5207 int dd_idx;
5208 struct bio* align_bi;
5209 struct md_rdev *rdev;
5210 sector_t end_sector;
5211
5212 if (!in_chunk_boundary(mddev, raid_bio)) {
5213 pr_debug("%s: non aligned\n", __func__);
5214 return 0;
5215 }
5216 /*
5217 * use bio_clone_fast to make a copy of the bio
5218 */
5219 align_bi = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
5220 if (!align_bi)
5221 return 0;
5222 /*
5223 * set bi_end_io to a new function, and set bi_private to the
5224 * original bio.
5225 */
5226 align_bi->bi_end_io = raid5_align_endio;
5227 align_bi->bi_private = raid_bio;
5228 /*
5229 * compute position
5230 */
5231 align_bi->bi_iter.bi_sector =
5232 raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5233 0, &dd_idx, NULL);
5234
5235 end_sector = bio_end_sector(align_bi);
5236 rcu_read_lock();
5237 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5238 if (!rdev || test_bit(Faulty, &rdev->flags) ||
5239 rdev->recovery_offset < end_sector) {
5240 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5241 if (rdev &&
5242 (test_bit(Faulty, &rdev->flags) ||
5243 !(test_bit(In_sync, &rdev->flags) ||
5244 rdev->recovery_offset >= end_sector)))
5245 rdev = NULL;
5246 }
5247
5248 if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5249 rcu_read_unlock();
5250 bio_put(align_bi);
5251 return 0;
5252 }
5253
5254 if (rdev) {
5255 sector_t first_bad;
5256 int bad_sectors;
5257
5258 atomic_inc(&rdev->nr_pending);
5259 rcu_read_unlock();
5260 raid_bio->bi_next = (void*)rdev;
5261 bio_set_dev(align_bi, rdev->bdev);
5262
5263 if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5264 bio_sectors(align_bi),
5265 &first_bad, &bad_sectors)) {
5266 bio_put(align_bi);
5267 rdev_dec_pending(rdev, mddev);
5268 return 0;
5269 }
5270
5271 /* No reshape active, so we can trust rdev->data_offset */
5272 align_bi->bi_iter.bi_sector += rdev->data_offset;
5273
5274 spin_lock_irq(&conf->device_lock);
5275 wait_event_lock_irq(conf->wait_for_quiescent,
5276 conf->quiesce == 0,
5277 conf->device_lock);
5278 atomic_inc(&conf->active_aligned_reads);
5279 spin_unlock_irq(&conf->device_lock);
5280
5281 if (mddev->gendisk)
5282 trace_block_bio_remap(align_bi->bi_disk->queue,
5283 align_bi, disk_devt(mddev->gendisk),
5284 raid_bio->bi_iter.bi_sector);
5285 generic_make_request(align_bi);
5286 return 1;
5287 } else {
5288 rcu_read_unlock();
5289 bio_put(align_bi);
5290 return 0;
5291 }
5292}
5293
5294static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5295{
5296 struct bio *split;
5297 sector_t sector = raid_bio->bi_iter.bi_sector;
5298 unsigned chunk_sects = mddev->chunk_sectors;
5299 unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5300
5301 if (sectors < bio_sectors(raid_bio)) {
5302 struct r5conf *conf = mddev->private;
5303 split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5304 bio_chain(split, raid_bio);
5305 generic_make_request(raid_bio);
5306 raid_bio = split;
5307 }
5308
5309 if (!raid5_read_one_chunk(mddev, raid_bio))
5310 return raid_bio;
5311
5312 return NULL;
5313}
5314
5315/* __get_priority_stripe - get the next stripe to process
5316 *
5317 * Full stripe writes are allowed to pass preread active stripes up until
5318 * the bypass_threshold is exceeded. In general the bypass_count
5319 * increments when the handle_list is handled before the hold_list; however, it
5320 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5321 * stripe with in flight i/o. The bypass_count will be reset when the
5322 * head of the hold_list has changed, i.e. the head was promoted to the
5323 * handle_list.
5324 */
5325static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5326{
5327 struct stripe_head *sh, *tmp;
5328 struct list_head *handle_list = NULL;
5329 struct r5worker_group *wg;
5330 bool second_try = !r5c_is_writeback(conf->log) &&
5331 !r5l_log_disk_error(conf);
5332 bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5333 r5l_log_disk_error(conf);
5334
5335again:
5336 wg = NULL;
5337 sh = NULL;
5338 if (conf->worker_cnt_per_group == 0) {
5339 handle_list = try_loprio ? &conf->loprio_list :
5340 &conf->handle_list;
5341 } else if (group != ANY_GROUP) {
5342 handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5343 &conf->worker_groups[group].handle_list;
5344 wg = &conf->worker_groups[group];
5345 } else {
5346 int i;
5347 for (i = 0; i < conf->group_cnt; i++) {
5348 handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5349 &conf->worker_groups[i].handle_list;
5350 wg = &conf->worker_groups[i];
5351 if (!list_empty(handle_list))
5352 break;
5353 }
5354 }
5355
5356 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5357 __func__,
5358 list_empty(handle_list) ? "empty" : "busy",
5359 list_empty(&conf->hold_list) ? "empty" : "busy",
5360 atomic_read(&conf->pending_full_writes), conf->bypass_count);
5361
5362 if (!list_empty(handle_list)) {
5363 sh = list_entry(handle_list->next, typeof(*sh), lru);
5364
5365 if (list_empty(&conf->hold_list))
5366 conf->bypass_count = 0;
5367 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5368 if (conf->hold_list.next == conf->last_hold)
5369 conf->bypass_count++;
5370 else {
5371 conf->last_hold = conf->hold_list.next;
5372 conf->bypass_count -= conf->bypass_threshold;
5373 if (conf->bypass_count < 0)
5374 conf->bypass_count = 0;
5375 }
5376 }
5377 } else if (!list_empty(&conf->hold_list) &&
5378 ((conf->bypass_threshold &&
5379 conf->bypass_count > conf->bypass_threshold) ||
5380 atomic_read(&conf->pending_full_writes) == 0)) {
5381
5382 list_for_each_entry(tmp, &conf->hold_list, lru) {
5383 if (conf->worker_cnt_per_group == 0 ||
5384 group == ANY_GROUP ||
5385 !cpu_online(tmp->cpu) ||
5386 cpu_to_group(tmp->cpu) == group) {
5387 sh = tmp;
5388 break;
5389 }
5390 }
5391
5392 if (sh) {
5393 conf->bypass_count -= conf->bypass_threshold;
5394 if (conf->bypass_count < 0)
5395 conf->bypass_count = 0;
5396 }
5397 wg = NULL;
5398 }
5399
5400 if (!sh) {
5401 if (second_try)
5402 return NULL;
5403 second_try = true;
5404 try_loprio = !try_loprio;
5405 goto again;
5406 }
5407
5408 if (wg) {
5409 wg->stripes_cnt--;
5410 sh->group = NULL;
5411 }
5412 list_del_init(&sh->lru);
5413 BUG_ON(atomic_inc_return(&sh->count) != 1);
5414 return sh;
5415}
5416
5417struct raid5_plug_cb {
5418 struct blk_plug_cb cb;
5419 struct list_head list;
5420 struct list_head temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5421};
5422
5423static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5424{
5425 struct raid5_plug_cb *cb = container_of(
5426 blk_cb, struct raid5_plug_cb, cb);
5427 struct stripe_head *sh;
5428 struct mddev *mddev = cb->cb.data;
5429 struct r5conf *conf = mddev->private;
5430 int cnt = 0;
5431 int hash;
5432
5433 if (cb->list.next && !list_empty(&cb->list)) {
5434 spin_lock_irq(&conf->device_lock);
5435 while (!list_empty(&cb->list)) {
5436 sh = list_first_entry(&cb->list, struct stripe_head, lru);
5437 list_del_init(&sh->lru);
5438 /*
5439 * avoid race release_stripe_plug() sees
5440 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5441 * is still in our list
5442 */
5443 smp_mb__before_atomic();
5444 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5445 /*
5446 * STRIPE_ON_RELEASE_LIST could be set here. In that
5447 * case, the count is always > 1 here
5448 */
5449 hash = sh->hash_lock_index;
5450 __release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5451 cnt++;
5452 }
5453 spin_unlock_irq(&conf->device_lock);
5454 }
5455 release_inactive_stripe_list(conf, cb->temp_inactive_list,
5456 NR_STRIPE_HASH_LOCKS);
5457 if (mddev->queue)
5458 trace_block_unplug(mddev->queue, cnt, !from_schedule);
5459 kfree(cb);
5460}
5461
5462static void release_stripe_plug(struct mddev *mddev,
5463 struct stripe_head *sh)
5464{
5465 struct blk_plug_cb *blk_cb = blk_check_plugged(
5466 raid5_unplug, mddev,
5467 sizeof(struct raid5_plug_cb));
5468 struct raid5_plug_cb *cb;
5469
5470 if (!blk_cb) {
5471 raid5_release_stripe(sh);
5472 return;
5473 }
5474
5475 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5476
5477 if (cb->list.next == NULL) {
5478 int i;
5479 INIT_LIST_HEAD(&cb->list);
5480 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5481 INIT_LIST_HEAD(cb->temp_inactive_list + i);
5482 }
5483
5484 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5485 list_add_tail(&sh->lru, &cb->list);
5486 else
5487 raid5_release_stripe(sh);
5488}
5489
5490static void make_discard_request(struct mddev *mddev, struct bio *bi)
5491{
5492 struct r5conf *conf = mddev->private;
5493 sector_t logical_sector, last_sector;
5494 struct stripe_head *sh;
5495 int stripe_sectors;
5496
5497 if (mddev->reshape_position != MaxSector)
5498 /* Skip discard while reshape is happening */
5499 return;
5500
5501 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5502 last_sector = bio_end_sector(bi);
5503
5504 bi->bi_next = NULL;
5505
5506 stripe_sectors = conf->chunk_sectors *
5507 (conf->raid_disks - conf->max_degraded);
5508 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5509 stripe_sectors);
5510 sector_div(last_sector, stripe_sectors);
5511
5512 logical_sector *= conf->chunk_sectors;
5513 last_sector *= conf->chunk_sectors;
5514
5515 for (; logical_sector < last_sector;
5516 logical_sector += STRIPE_SECTORS) {
5517 DEFINE_WAIT(w);
5518 int d;
5519 again:
5520 sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5521 prepare_to_wait(&conf->wait_for_overlap, &w,
5522 TASK_UNINTERRUPTIBLE);
5523 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5524 if (test_bit(STRIPE_SYNCING, &sh->state)) {
5525 raid5_release_stripe(sh);
5526 schedule();
5527 goto again;
5528 }
5529 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5530 spin_lock_irq(&sh->stripe_lock);
5531 for (d = 0; d < conf->raid_disks; d++) {
5532 if (d == sh->pd_idx || d == sh->qd_idx)
5533 continue;
5534 if (sh->dev[d].towrite || sh->dev[d].toread) {
5535 set_bit(R5_Overlap, &sh->dev[d].flags);
5536 spin_unlock_irq(&sh->stripe_lock);
5537 raid5_release_stripe(sh);
5538 schedule();
5539 goto again;
5540 }
5541 }
5542 set_bit(STRIPE_DISCARD, &sh->state);
5543 finish_wait(&conf->wait_for_overlap, &w);
5544 sh->overwrite_disks = 0;
5545 for (d = 0; d < conf->raid_disks; d++) {
5546 if (d == sh->pd_idx || d == sh->qd_idx)
5547 continue;
5548 sh->dev[d].towrite = bi;
5549 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5550 bio_inc_remaining(bi);
5551 md_write_inc(mddev, bi);
5552 sh->overwrite_disks++;
5553 }
5554 spin_unlock_irq(&sh->stripe_lock);
5555 if (conf->mddev->bitmap) {
5556 for (d = 0;
5557 d < conf->raid_disks - conf->max_degraded;
5558 d++)
5559 md_bitmap_startwrite(mddev->bitmap,
5560 sh->sector,
5561 STRIPE_SECTORS,
5562 0);
5563 sh->bm_seq = conf->seq_flush + 1;
5564 set_bit(STRIPE_BIT_DELAY, &sh->state);
5565 }
5566
5567 set_bit(STRIPE_HANDLE, &sh->state);
5568 clear_bit(STRIPE_DELAYED, &sh->state);
5569 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5570 atomic_inc(&conf->preread_active_stripes);
5571 release_stripe_plug(mddev, sh);
5572 }
5573
5574 bio_endio(bi);
5575}
5576
5577static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5578{
5579 struct r5conf *conf = mddev->private;
5580 int dd_idx;
5581 sector_t new_sector;
5582 sector_t logical_sector, last_sector;
5583 struct stripe_head *sh;
5584 const int rw = bio_data_dir(bi);
5585 DEFINE_WAIT(w);
5586 bool do_prepare;
5587 bool do_flush = false;
5588
5589 if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5590 int ret = log_handle_flush_request(conf, bi);
5591
5592 if (ret == 0)
5593 return true;
5594 if (ret == -ENODEV) {
5595 md_flush_request(mddev, bi);
5596 return true;
5597 }
5598 /* ret == -EAGAIN, fallback */
5599 /*
5600 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5601 * we need to flush journal device
5602 */
5603 do_flush = bi->bi_opf & REQ_PREFLUSH;
5604 }
5605
5606 if (!md_write_start(mddev, bi))
5607 return false;
5608 /*
5609 * If array is degraded, better not do chunk aligned read because
5610 * later we might have to read it again in order to reconstruct
5611 * data on failed drives.
5612 */
5613 if (rw == READ && mddev->degraded == 0 &&
5614 mddev->reshape_position == MaxSector) {
5615 bi = chunk_aligned_read(mddev, bi);
5616 if (!bi)
5617 return true;
5618 }
5619
5620 if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5621 make_discard_request(mddev, bi);
5622 md_write_end(mddev);
5623 return true;
5624 }
5625
5626 logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5627 last_sector = bio_end_sector(bi);
5628 bi->bi_next = NULL;
5629
5630 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5631 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5632 int previous;
5633 int seq;
5634
5635 do_prepare = false;
5636 retry:
5637 seq = read_seqcount_begin(&conf->gen_lock);
5638 previous = 0;
5639 if (do_prepare)
5640 prepare_to_wait(&conf->wait_for_overlap, &w,
5641 TASK_UNINTERRUPTIBLE);
5642 if (unlikely(conf->reshape_progress != MaxSector)) {
5643 /* spinlock is needed as reshape_progress may be
5644 * 64bit on a 32bit platform, and so it might be
5645 * possible to see a half-updated value
5646 * Of course reshape_progress could change after
5647 * the lock is dropped, so once we get a reference
5648 * to the stripe that we think it is, we will have
5649 * to check again.
5650 */
5651 spin_lock_irq(&conf->device_lock);
5652 if (mddev->reshape_backwards
5653 ? logical_sector < conf->reshape_progress
5654 : logical_sector >= conf->reshape_progress) {
5655 previous = 1;
5656 } else {
5657 if (mddev->reshape_backwards
5658 ? logical_sector < conf->reshape_safe
5659 : logical_sector >= conf->reshape_safe) {
5660 spin_unlock_irq(&conf->device_lock);
5661 schedule();
5662 do_prepare = true;
5663 goto retry;
5664 }
5665 }
5666 spin_unlock_irq(&conf->device_lock);
5667 }
5668
5669 new_sector = raid5_compute_sector(conf, logical_sector,
5670 previous,
5671 &dd_idx, NULL);
5672 pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5673 (unsigned long long)new_sector,
5674 (unsigned long long)logical_sector);
5675
5676 sh = raid5_get_active_stripe(conf, new_sector, previous,
5677 (bi->bi_opf & REQ_RAHEAD), 0);
5678 if (sh) {
5679 if (unlikely(previous)) {
5680 /* expansion might have moved on while waiting for a
5681 * stripe, so we must do the range check again.
5682 * Expansion could still move past after this
5683 * test, but as we are holding a reference to
5684 * 'sh', we know that if that happens,
5685 * STRIPE_EXPANDING will get set and the expansion
5686 * won't proceed until we finish with the stripe.
5687 */
5688 int must_retry = 0;
5689 spin_lock_irq(&conf->device_lock);
5690 if (mddev->reshape_backwards
5691 ? logical_sector >= conf->reshape_progress
5692 : logical_sector < conf->reshape_progress)
5693 /* mismatch, need to try again */
5694 must_retry = 1;
5695 spin_unlock_irq(&conf->device_lock);
5696 if (must_retry) {
5697 raid5_release_stripe(sh);
5698 schedule();
5699 do_prepare = true;
5700 goto retry;
5701 }
5702 }
5703 if (read_seqcount_retry(&conf->gen_lock, seq)) {
5704 /* Might have got the wrong stripe_head
5705 * by accident
5706 */
5707 raid5_release_stripe(sh);
5708 goto retry;
5709 }
5710
5711 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5712 !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5713 /* Stripe is busy expanding or
5714 * add failed due to overlap. Flush everything
5715 * and wait a while
5716 */
5717 md_wakeup_thread(mddev->thread);
5718 raid5_release_stripe(sh);
5719 schedule();
5720 do_prepare = true;
5721 goto retry;
5722 }
5723 if (do_flush) {
5724 set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5725 /* we only need flush for one stripe */
5726 do_flush = false;
5727 }
5728
5729 if (!sh->batch_head)
5730 set_bit(STRIPE_HANDLE, &sh->state);
5731 clear_bit(STRIPE_DELAYED, &sh->state);
5732 if ((!sh->batch_head || sh == sh->batch_head) &&
5733 (bi->bi_opf & REQ_SYNC) &&
5734 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5735 atomic_inc(&conf->preread_active_stripes);
5736 release_stripe_plug(mddev, sh);
5737 } else {
5738 /* cannot get stripe for read-ahead, just give-up */
5739 bi->bi_status = BLK_STS_IOERR;
5740 break;
5741 }
5742 }
5743 finish_wait(&conf->wait_for_overlap, &w);
5744
5745 if (rw == WRITE)
5746 md_write_end(mddev);
5747 bio_endio(bi);
5748 return true;
5749}
5750
5751static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5752
5753static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5754{
5755 /* reshaping is quite different to recovery/resync so it is
5756 * handled quite separately ... here.
5757 *
5758 * On each call to sync_request, we gather one chunk worth of
5759 * destination stripes and flag them as expanding.
5760 * Then we find all the source stripes and request reads.
5761 * As the reads complete, handle_stripe will copy the data
5762 * into the destination stripe and release that stripe.
5763 */
5764 struct r5conf *conf = mddev->private;
5765 struct stripe_head *sh;
5766 struct md_rdev *rdev;
5767 sector_t first_sector, last_sector;
5768 int raid_disks = conf->previous_raid_disks;
5769 int data_disks = raid_disks - conf->max_degraded;
5770 int new_data_disks = conf->raid_disks - conf->max_degraded;
5771 int i;
5772 int dd_idx;
5773 sector_t writepos, readpos, safepos;
5774 sector_t stripe_addr;
5775 int reshape_sectors;
5776 struct list_head stripes;
5777 sector_t retn;
5778
5779 if (sector_nr == 0) {
5780 /* If restarting in the middle, skip the initial sectors */
5781 if (mddev->reshape_backwards &&
5782 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5783 sector_nr = raid5_size(mddev, 0, 0)
5784 - conf->reshape_progress;
5785 } else if (mddev->reshape_backwards &&
5786 conf->reshape_progress == MaxSector) {
5787 /* shouldn't happen, but just in case, finish up.*/
5788 sector_nr = MaxSector;
5789 } else if (!mddev->reshape_backwards &&
5790 conf->reshape_progress > 0)
5791 sector_nr = conf->reshape_progress;
5792 sector_div(sector_nr, new_data_disks);
5793 if (sector_nr) {
5794 mddev->curr_resync_completed = sector_nr;
5795 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5796 *skipped = 1;
5797 retn = sector_nr;
5798 goto finish;
5799 }
5800 }
5801
5802 /* We need to process a full chunk at a time.
5803 * If old and new chunk sizes differ, we need to process the
5804 * largest of these
5805 */
5806
5807 reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5808
5809 /* We update the metadata at least every 10 seconds, or when
5810 * the data about to be copied would over-write the source of
5811 * the data at the front of the range. i.e. one new_stripe
5812 * along from reshape_progress new_maps to after where
5813 * reshape_safe old_maps to
5814 */
5815 writepos = conf->reshape_progress;
5816 sector_div(writepos, new_data_disks);
5817 readpos = conf->reshape_progress;
5818 sector_div(readpos, data_disks);
5819 safepos = conf->reshape_safe;
5820 sector_div(safepos, data_disks);
5821 if (mddev->reshape_backwards) {
5822 BUG_ON(writepos < reshape_sectors);
5823 writepos -= reshape_sectors;
5824 readpos += reshape_sectors;
5825 safepos += reshape_sectors;
5826 } else {
5827 writepos += reshape_sectors;
5828 /* readpos and safepos are worst-case calculations.
5829 * A negative number is overly pessimistic, and causes
5830 * obvious problems for unsigned storage. So clip to 0.
5831 */
5832 readpos -= min_t(sector_t, reshape_sectors, readpos);
5833 safepos -= min_t(sector_t, reshape_sectors, safepos);
5834 }
5835
5836 /* Having calculated the 'writepos' possibly use it
5837 * to set 'stripe_addr' which is where we will write to.
5838 */
5839 if (mddev->reshape_backwards) {
5840 BUG_ON(conf->reshape_progress == 0);
5841 stripe_addr = writepos;
5842 BUG_ON((mddev->dev_sectors &
5843 ~((sector_t)reshape_sectors - 1))
5844 - reshape_sectors - stripe_addr
5845 != sector_nr);
5846 } else {
5847 BUG_ON(writepos != sector_nr + reshape_sectors);
5848 stripe_addr = sector_nr;
5849 }
5850
5851 /* 'writepos' is the most advanced device address we might write.
5852 * 'readpos' is the least advanced device address we might read.
5853 * 'safepos' is the least address recorded in the metadata as having
5854 * been reshaped.
5855 * If there is a min_offset_diff, these are adjusted either by
5856 * increasing the safepos/readpos if diff is negative, or
5857 * increasing writepos if diff is positive.
5858 * If 'readpos' is then behind 'writepos', there is no way that we can
5859 * ensure safety in the face of a crash - that must be done by userspace
5860 * making a backup of the data. So in that case there is no particular
5861 * rush to update metadata.
5862 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5863 * update the metadata to advance 'safepos' to match 'readpos' so that
5864 * we can be safe in the event of a crash.
5865 * So we insist on updating metadata if safepos is behind writepos and
5866 * readpos is beyond writepos.
5867 * In any case, update the metadata every 10 seconds.
5868 * Maybe that number should be configurable, but I'm not sure it is
5869 * worth it.... maybe it could be a multiple of safemode_delay???
5870 */
5871 if (conf->min_offset_diff < 0) {
5872 safepos += -conf->min_offset_diff;
5873 readpos += -conf->min_offset_diff;
5874 } else
5875 writepos += conf->min_offset_diff;
5876
5877 if ((mddev->reshape_backwards
5878 ? (safepos > writepos && readpos < writepos)
5879 : (safepos < writepos && readpos > writepos)) ||
5880 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5881 /* Cannot proceed until we've updated the superblock... */
5882 wait_event(conf->wait_for_overlap,
5883 atomic_read(&conf->reshape_stripes)==0
5884 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5885 if (atomic_read(&conf->reshape_stripes) != 0)
5886 return 0;
5887 mddev->reshape_position = conf->reshape_progress;
5888 mddev->curr_resync_completed = sector_nr;
5889 if (!mddev->reshape_backwards)
5890 /* Can update recovery_offset */
5891 rdev_for_each(rdev, mddev)
5892 if (rdev->raid_disk >= 0 &&
5893 !test_bit(Journal, &rdev->flags) &&
5894 !test_bit(In_sync, &rdev->flags) &&
5895 rdev->recovery_offset < sector_nr)
5896 rdev->recovery_offset = sector_nr;
5897
5898 conf->reshape_checkpoint = jiffies;
5899 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5900 md_wakeup_thread(mddev->thread);
5901 wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5902 test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5903 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5904 return 0;
5905 spin_lock_irq(&conf->device_lock);
5906 conf->reshape_safe = mddev->reshape_position;
5907 spin_unlock_irq(&conf->device_lock);
5908 wake_up(&conf->wait_for_overlap);
5909 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5910 }
5911
5912 INIT_LIST_HEAD(&stripes);
5913 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5914 int j;
5915 int skipped_disk = 0;
5916 sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5917 set_bit(STRIPE_EXPANDING, &sh->state);
5918 atomic_inc(&conf->reshape_stripes);
5919 /* If any of this stripe is beyond the end of the old
5920 * array, then we need to zero those blocks
5921 */
5922 for (j=sh->disks; j--;) {
5923 sector_t s;
5924 if (j == sh->pd_idx)
5925 continue;
5926 if (conf->level == 6 &&
5927 j == sh->qd_idx)
5928 continue;
5929 s = raid5_compute_blocknr(sh, j, 0);
5930 if (s < raid5_size(mddev, 0, 0)) {
5931 skipped_disk = 1;
5932 continue;
5933 }
5934 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5935 set_bit(R5_Expanded, &sh->dev[j].flags);
5936 set_bit(R5_UPTODATE, &sh->dev[j].flags);
5937 }
5938 if (!skipped_disk) {
5939 set_bit(STRIPE_EXPAND_READY, &sh->state);
5940 set_bit(STRIPE_HANDLE, &sh->state);
5941 }
5942 list_add(&sh->lru, &stripes);
5943 }
5944 spin_lock_irq(&conf->device_lock);
5945 if (mddev->reshape_backwards)
5946 conf->reshape_progress -= reshape_sectors * new_data_disks;
5947 else
5948 conf->reshape_progress += reshape_sectors * new_data_disks;
5949 spin_unlock_irq(&conf->device_lock);
5950 /* Ok, those stripe are ready. We can start scheduling
5951 * reads on the source stripes.
5952 * The source stripes are determined by mapping the first and last
5953 * block on the destination stripes.
5954 */
5955 first_sector =
5956 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5957 1, &dd_idx, NULL);
5958 last_sector =
5959 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5960 * new_data_disks - 1),
5961 1, &dd_idx, NULL);
5962 if (last_sector >= mddev->dev_sectors)
5963 last_sector = mddev->dev_sectors - 1;
5964 while (first_sector <= last_sector) {
5965 sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5966 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5967 set_bit(STRIPE_HANDLE, &sh->state);
5968 raid5_release_stripe(sh);
5969 first_sector += STRIPE_SECTORS;
5970 }
5971 /* Now that the sources are clearly marked, we can release
5972 * the destination stripes
5973 */
5974 while (!list_empty(&stripes)) {
5975 sh = list_entry(stripes.next, struct stripe_head, lru);
5976 list_del_init(&sh->lru);
5977 raid5_release_stripe(sh);
5978 }
5979 /* If this takes us to the resync_max point where we have to pause,
5980 * then we need to write out the superblock.
5981 */
5982 sector_nr += reshape_sectors;
5983 retn = reshape_sectors;
5984finish:
5985 if (mddev->curr_resync_completed > mddev->resync_max ||
5986 (sector_nr - mddev->curr_resync_completed) * 2
5987 >= mddev->resync_max - mddev->curr_resync_completed) {
5988 /* Cannot proceed until we've updated the superblock... */
5989 wait_event(conf->wait_for_overlap,
5990 atomic_read(&conf->reshape_stripes) == 0
5991 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5992 if (atomic_read(&conf->reshape_stripes) != 0)
5993 goto ret;
5994 mddev->reshape_position = conf->reshape_progress;
5995 mddev->curr_resync_completed = sector_nr;
5996 if (!mddev->reshape_backwards)
5997 /* Can update recovery_offset */
5998 rdev_for_each(rdev, mddev)
5999 if (rdev->raid_disk >= 0 &&
6000 !test_bit(Journal, &rdev->flags) &&
6001 !test_bit(In_sync, &rdev->flags) &&
6002 rdev->recovery_offset < sector_nr)
6003 rdev->recovery_offset = sector_nr;
6004 conf->reshape_checkpoint = jiffies;
6005 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6006 md_wakeup_thread(mddev->thread);
6007 wait_event(mddev->sb_wait,
6008 !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6009 || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6010 if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6011 goto ret;
6012 spin_lock_irq(&conf->device_lock);
6013 conf->reshape_safe = mddev->reshape_position;
6014 spin_unlock_irq(&conf->device_lock);
6015 wake_up(&conf->wait_for_overlap);
6016 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
6017 }
6018ret:
6019 return retn;
6020}
6021
6022static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6023 int *skipped)
6024{
6025 struct r5conf *conf = mddev->private;
6026 struct stripe_head *sh;
6027 sector_t max_sector = mddev->dev_sectors;
6028 sector_t sync_blocks;
6029 int still_degraded = 0;
6030 int i;
6031
6032 if (sector_nr >= max_sector) {
6033 /* just being told to finish up .. nothing much to do */
6034
6035 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6036 end_reshape(conf);
6037 return 0;
6038 }
6039
6040 if (mddev->curr_resync < max_sector) /* aborted */
6041 md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6042 &sync_blocks, 1);
6043 else /* completed sync */
6044 conf->fullsync = 0;
6045 md_bitmap_close_sync(mddev->bitmap);
6046
6047 return 0;
6048 }
6049
6050 /* Allow raid5_quiesce to complete */
6051 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6052
6053 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6054 return reshape_request(mddev, sector_nr, skipped);
6055
6056 /* No need to check resync_max as we never do more than one
6057 * stripe, and as resync_max will always be on a chunk boundary,
6058 * if the check in md_do_sync didn't fire, there is no chance
6059 * of overstepping resync_max here
6060 */
6061
6062 /* if there is too many failed drives and we are trying
6063 * to resync, then assert that we are finished, because there is
6064 * nothing we can do.
6065 */
6066 if (mddev->degraded >= conf->max_degraded &&
6067 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6068 sector_t rv = mddev->dev_sectors - sector_nr;
6069 *skipped = 1;
6070 return rv;
6071 }
6072 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6073 !conf->fullsync &&
6074 !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6075 sync_blocks >= STRIPE_SECTORS) {
6076 /* we can skip this block, and probably more */
6077 sync_blocks /= STRIPE_SECTORS;
6078 *skipped = 1;
6079 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6080 }
6081
6082 md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6083
6084 sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6085 if (sh == NULL) {
6086 sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6087 /* make sure we don't swamp the stripe cache if someone else
6088 * is trying to get access
6089 */
6090 schedule_timeout_uninterruptible(1);
6091 }
6092 /* Need to check if array will still be degraded after recovery/resync
6093 * Note in case of > 1 drive failures it's possible we're rebuilding
6094 * one drive while leaving another faulty drive in array.
6095 */
6096 rcu_read_lock();
6097 for (i = 0; i < conf->raid_disks; i++) {
6098 struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6099
6100 if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6101 still_degraded = 1;
6102 }
6103 rcu_read_unlock();
6104
6105 md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6106
6107 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6108 set_bit(STRIPE_HANDLE, &sh->state);
6109
6110 raid5_release_stripe(sh);
6111
6112 return STRIPE_SECTORS;
6113}
6114
6115static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6116 unsigned int offset)
6117{
6118 /* We may not be able to submit a whole bio at once as there
6119 * may not be enough stripe_heads available.
6120 * We cannot pre-allocate enough stripe_heads as we may need
6121 * more than exist in the cache (if we allow ever large chunks).
6122 * So we do one stripe head at a time and record in
6123 * ->bi_hw_segments how many have been done.
6124 *
6125 * We *know* that this entire raid_bio is in one chunk, so
6126 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6127 */
6128 struct stripe_head *sh;
6129 int dd_idx;
6130 sector_t sector, logical_sector, last_sector;
6131 int scnt = 0;
6132 int handled = 0;
6133
6134 logical_sector = raid_bio->bi_iter.bi_sector &
6135 ~((sector_t)STRIPE_SECTORS-1);
6136 sector = raid5_compute_sector(conf, logical_sector,
6137 0, &dd_idx, NULL);
6138 last_sector = bio_end_sector(raid_bio);
6139
6140 for (; logical_sector < last_sector;
6141 logical_sector += STRIPE_SECTORS,
6142 sector += STRIPE_SECTORS,
6143 scnt++) {
6144
6145 if (scnt < offset)
6146 /* already done this stripe */
6147 continue;
6148
6149 sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6150
6151 if (!sh) {
6152 /* failed to get a stripe - must wait */
6153 conf->retry_read_aligned = raid_bio;
6154 conf->retry_read_offset = scnt;
6155 return handled;
6156 }
6157
6158 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6159 raid5_release_stripe(sh);
6160 conf->retry_read_aligned = raid_bio;
6161 conf->retry_read_offset = scnt;
6162 return handled;
6163 }
6164
6165 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6166 handle_stripe(sh);
6167 raid5_release_stripe(sh);
6168 handled++;
6169 }
6170
6171 bio_endio(raid_bio);
6172
6173 if (atomic_dec_and_test(&conf->active_aligned_reads))
6174 wake_up(&conf->wait_for_quiescent);
6175 return handled;
6176}
6177
6178static int handle_active_stripes(struct r5conf *conf, int group,
6179 struct r5worker *worker,
6180 struct list_head *temp_inactive_list)
6181 __releases(&conf->device_lock)
6182 __acquires(&conf->device_lock)
6183{
6184 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6185 int i, batch_size = 0, hash;
6186 bool release_inactive = false;
6187
6188 while (batch_size < MAX_STRIPE_BATCH &&
6189 (sh = __get_priority_stripe(conf, group)) != NULL)
6190 batch[batch_size++] = sh;
6191
6192 if (batch_size == 0) {
6193 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6194 if (!list_empty(temp_inactive_list + i))
6195 break;
6196 if (i == NR_STRIPE_HASH_LOCKS) {
6197 spin_unlock_irq(&conf->device_lock);
6198 log_flush_stripe_to_raid(conf);
6199 spin_lock_irq(&conf->device_lock);
6200 return batch_size;
6201 }
6202 release_inactive = true;
6203 }
6204 spin_unlock_irq(&conf->device_lock);
6205
6206 release_inactive_stripe_list(conf, temp_inactive_list,
6207 NR_STRIPE_HASH_LOCKS);
6208
6209 r5l_flush_stripe_to_raid(conf->log);
6210 if (release_inactive) {
6211 spin_lock_irq(&conf->device_lock);
6212 return 0;
6213 }
6214
6215 for (i = 0; i < batch_size; i++)
6216 handle_stripe(batch[i]);
6217 log_write_stripe_run(conf);
6218
6219 cond_resched();
6220
6221 spin_lock_irq(&conf->device_lock);
6222 for (i = 0; i < batch_size; i++) {
6223 hash = batch[i]->hash_lock_index;
6224 __release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6225 }
6226 return batch_size;
6227}
6228
6229static void raid5_do_work(struct work_struct *work)
6230{
6231 struct r5worker *worker = container_of(work, struct r5worker, work);
6232 struct r5worker_group *group = worker->group;
6233 struct r5conf *conf = group->conf;
6234 struct mddev *mddev = conf->mddev;
6235 int group_id = group - conf->worker_groups;
6236 int handled;
6237 struct blk_plug plug;
6238
6239 pr_debug("+++ raid5worker active\n");
6240
6241 blk_start_plug(&plug);
6242 handled = 0;
6243 spin_lock_irq(&conf->device_lock);
6244 while (1) {
6245 int batch_size, released;
6246
6247 released = release_stripe_list(conf, worker->temp_inactive_list);
6248
6249 batch_size = handle_active_stripes(conf, group_id, worker,
6250 worker->temp_inactive_list);
6251 worker->working = false;
6252 if (!batch_size && !released)
6253 break;
6254 handled += batch_size;
6255 wait_event_lock_irq(mddev->sb_wait,
6256 !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6257 conf->device_lock);
6258 }
6259 pr_debug("%d stripes handled\n", handled);
6260
6261 spin_unlock_irq(&conf->device_lock);
6262
6263 flush_deferred_bios(conf);
6264
6265 r5l_flush_stripe_to_raid(conf->log);
6266
6267 async_tx_issue_pending_all();
6268 blk_finish_plug(&plug);
6269
6270 pr_debug("--- raid5worker inactive\n");
6271}
6272
6273/*
6274 * This is our raid5 kernel thread.
6275 *
6276 * We scan the hash table for stripes which can be handled now.
6277 * During the scan, completed stripes are saved for us by the interrupt
6278 * handler, so that they will not have to wait for our next wakeup.
6279 */
6280static void raid5d(struct md_thread *thread)
6281{
6282 struct mddev *mddev = thread->mddev;
6283 struct r5conf *conf = mddev->private;
6284 int handled;
6285 struct blk_plug plug;
6286
6287 pr_debug("+++ raid5d active\n");
6288
6289 md_check_recovery(mddev);
6290
6291 blk_start_plug(&plug);
6292 handled = 0;
6293 spin_lock_irq(&conf->device_lock);
6294 while (1) {
6295 struct bio *bio;
6296 int batch_size, released;
6297 unsigned int offset;
6298
6299 released = release_stripe_list(conf, conf->temp_inactive_list);
6300 if (released)
6301 clear_bit(R5_DID_ALLOC, &conf->cache_state);
6302
6303 if (
6304 !list_empty(&conf->bitmap_list)) {
6305 /* Now is a good time to flush some bitmap updates */
6306 conf->seq_flush++;
6307 spin_unlock_irq(&conf->device_lock);
6308 md_bitmap_unplug(mddev->bitmap);
6309 spin_lock_irq(&conf->device_lock);
6310 conf->seq_write = conf->seq_flush;
6311 activate_bit_delay(conf, conf->temp_inactive_list);
6312 }
6313 raid5_activate_delayed(conf);
6314
6315 while ((bio = remove_bio_from_retry(conf, &offset))) {
6316 int ok;
6317 spin_unlock_irq(&conf->device_lock);
6318 ok = retry_aligned_read(conf, bio, offset);
6319 spin_lock_irq(&conf->device_lock);
6320 if (!ok)
6321 break;
6322 handled++;
6323 }
6324
6325 batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6326 conf->temp_inactive_list);
6327 if (!batch_size && !released)
6328 break;
6329 handled += batch_size;
6330
6331 if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6332 spin_unlock_irq(&conf->device_lock);
6333 md_check_recovery(mddev);
6334 spin_lock_irq(&conf->device_lock);
6335 }
6336 }
6337 pr_debug("%d stripes handled\n", handled);
6338
6339 spin_unlock_irq(&conf->device_lock);
6340 if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6341 mutex_trylock(&conf->cache_size_mutex)) {
6342 grow_one_stripe(conf, __GFP_NOWARN);
6343 /* Set flag even if allocation failed. This helps
6344 * slow down allocation requests when mem is short
6345 */
6346 set_bit(R5_DID_ALLOC, &conf->cache_state);
6347 mutex_unlock(&conf->cache_size_mutex);
6348 }
6349
6350 flush_deferred_bios(conf);
6351
6352 r5l_flush_stripe_to_raid(conf->log);
6353
6354 async_tx_issue_pending_all();
6355 blk_finish_plug(&plug);
6356
6357 pr_debug("--- raid5d inactive\n");
6358}
6359
6360static ssize_t
6361raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6362{
6363 struct r5conf *conf;
6364 int ret = 0;
6365 spin_lock(&mddev->lock);
6366 conf = mddev->private;
6367 if (conf)
6368 ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6369 spin_unlock(&mddev->lock);
6370 return ret;
6371}
6372
6373int
6374raid5_set_cache_size(struct mddev *mddev, int size)
6375{
6376 int result = 0;
6377 struct r5conf *conf = mddev->private;
6378
6379 if (size <= 16 || size > 32768)
6380 return -EINVAL;
6381
6382 conf->min_nr_stripes = size;
6383 mutex_lock(&conf->cache_size_mutex);
6384 while (size < conf->max_nr_stripes &&
6385 drop_one_stripe(conf))
6386 ;
6387 mutex_unlock(&conf->cache_size_mutex);
6388
6389 md_allow_write(mddev);
6390
6391 mutex_lock(&conf->cache_size_mutex);
6392 while (size > conf->max_nr_stripes)
6393 if (!grow_one_stripe(conf, GFP_KERNEL)) {
6394 conf->min_nr_stripes = conf->max_nr_stripes;
6395 result = -ENOMEM;
6396 break;
6397 }
6398 mutex_unlock(&conf->cache_size_mutex);
6399
6400 return result;
6401}
6402EXPORT_SYMBOL(raid5_set_cache_size);
6403
6404static ssize_t
6405raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6406{
6407 struct r5conf *conf;
6408 unsigned long new;
6409 int err;
6410
6411 if (len >= PAGE_SIZE)
6412 return -EINVAL;
6413 if (kstrtoul(page, 10, &new))
6414 return -EINVAL;
6415 err = mddev_lock(mddev);
6416 if (err)
6417 return err;
6418 conf = mddev->private;
6419 if (!conf)
6420 err = -ENODEV;
6421 else
6422 err = raid5_set_cache_size(mddev, new);
6423 mddev_unlock(mddev);
6424
6425 return err ?: len;
6426}
6427
6428static struct md_sysfs_entry
6429raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6430 raid5_show_stripe_cache_size,
6431 raid5_store_stripe_cache_size);
6432
6433static ssize_t
6434raid5_show_rmw_level(struct mddev *mddev, char *page)
6435{
6436 struct r5conf *conf = mddev->private;
6437 if (conf)
6438 return sprintf(page, "%d\n", conf->rmw_level);
6439 else
6440 return 0;
6441}
6442
6443static ssize_t
6444raid5_store_rmw_level(struct mddev *mddev, const char *page, size_t len)
6445{
6446 struct r5conf *conf = mddev->private;
6447 unsigned long new;
6448
6449 if (!conf)
6450 return -ENODEV;
6451
6452 if (len >= PAGE_SIZE)
6453 return -EINVAL;
6454
6455 if (kstrtoul(page, 10, &new))
6456 return -EINVAL;
6457
6458 if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6459 return -EINVAL;
6460
6461 if (new != PARITY_DISABLE_RMW &&
6462 new != PARITY_ENABLE_RMW &&
6463 new != PARITY_PREFER_RMW)
6464 return -EINVAL;
6465
6466 conf->rmw_level = new;
6467 return len;
6468}
6469
6470static struct md_sysfs_entry
6471raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6472 raid5_show_rmw_level,
6473 raid5_store_rmw_level);
6474
6475
6476static ssize_t
6477raid5_show_preread_threshold(struct mddev *mddev, char *page)
6478{
6479 struct r5conf *conf;
6480 int ret = 0;
6481 spin_lock(&mddev->lock);
6482 conf = mddev->private;
6483 if (conf)
6484 ret = sprintf(page, "%d\n", conf->bypass_threshold);
6485 spin_unlock(&mddev->lock);
6486 return ret;
6487}
6488
6489static ssize_t
6490raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6491{
6492 struct r5conf *conf;
6493 unsigned long new;
6494 int err;
6495
6496 if (len >= PAGE_SIZE)
6497 return -EINVAL;
6498 if (kstrtoul(page, 10, &new))
6499 return -EINVAL;
6500
6501 err = mddev_lock(mddev);
6502 if (err)
6503 return err;
6504 conf = mddev->private;
6505 if (!conf)
6506 err = -ENODEV;
6507 else if (new > conf->min_nr_stripes)
6508 err = -EINVAL;
6509 else
6510 conf->bypass_threshold = new;
6511 mddev_unlock(mddev);
6512 return err ?: len;
6513}
6514
6515static struct md_sysfs_entry
6516raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6517 S_IRUGO | S_IWUSR,
6518 raid5_show_preread_threshold,
6519 raid5_store_preread_threshold);
6520
6521static ssize_t
6522raid5_show_skip_copy(struct mddev *mddev, char *page)
6523{
6524 struct r5conf *conf;
6525 int ret = 0;
6526 spin_lock(&mddev->lock);
6527 conf = mddev->private;
6528 if (conf)
6529 ret = sprintf(page, "%d\n", conf->skip_copy);
6530 spin_unlock(&mddev->lock);
6531 return ret;
6532}
6533
6534static ssize_t
6535raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6536{
6537 struct r5conf *conf;
6538 unsigned long new;
6539 int err;
6540
6541 if (len >= PAGE_SIZE)
6542 return -EINVAL;
6543 if (kstrtoul(page, 10, &new))
6544 return -EINVAL;
6545 new = !!new;
6546
6547 err = mddev_lock(mddev);
6548 if (err)
6549 return err;
6550 conf = mddev->private;
6551 if (!conf)
6552 err = -ENODEV;
6553 else if (new != conf->skip_copy) {
6554 mddev_suspend(mddev);
6555 conf->skip_copy = new;
6556 if (new)
6557 mddev->queue->backing_dev_info->capabilities |=
6558 BDI_CAP_STABLE_WRITES;
6559 else
6560 mddev->queue->backing_dev_info->capabilities &=
6561 ~BDI_CAP_STABLE_WRITES;
6562 mddev_resume(mddev);
6563 }
6564 mddev_unlock(mddev);
6565 return err ?: len;
6566}
6567
6568static struct md_sysfs_entry
6569raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6570 raid5_show_skip_copy,
6571 raid5_store_skip_copy);
6572
6573static ssize_t
6574stripe_cache_active_show(struct mddev *mddev, char *page)
6575{
6576 struct r5conf *conf = mddev->private;
6577 if (conf)
6578 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6579 else
6580 return 0;
6581}
6582
6583static struct md_sysfs_entry
6584raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6585
6586static ssize_t
6587raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6588{
6589 struct r5conf *conf;
6590 int ret = 0;
6591 spin_lock(&mddev->lock);
6592 conf = mddev->private;
6593 if (conf)
6594 ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6595 spin_unlock(&mddev->lock);
6596 return ret;
6597}
6598
6599static int alloc_thread_groups(struct r5conf *conf, int cnt,
6600 int *group_cnt,
6601 int *worker_cnt_per_group,
6602 struct r5worker_group **worker_groups);
6603static ssize_t
6604raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6605{
6606 struct r5conf *conf;
6607 unsigned int new;
6608 int err;
6609 struct r5worker_group *new_groups, *old_groups;
6610 int group_cnt, worker_cnt_per_group;
6611
6612 if (len >= PAGE_SIZE)
6613 return -EINVAL;
6614 if (kstrtouint(page, 10, &new))
6615 return -EINVAL;
6616 /* 8192 should be big enough */
6617 if (new > 8192)
6618 return -EINVAL;
6619
6620 err = mddev_lock(mddev);
6621 if (err)
6622 return err;
6623 conf = mddev->private;
6624 if (!conf)
6625 err = -ENODEV;
6626 else if (new != conf->worker_cnt_per_group) {
6627 mddev_suspend(mddev);
6628
6629 old_groups = conf->worker_groups;
6630 if (old_groups)
6631 flush_workqueue(raid5_wq);
6632
6633 err = alloc_thread_groups(conf, new,
6634 &group_cnt, &worker_cnt_per_group,
6635 &new_groups);
6636 if (!err) {
6637 spin_lock_irq(&conf->device_lock);
6638 conf->group_cnt = group_cnt;
6639 conf->worker_cnt_per_group = worker_cnt_per_group;
6640 conf->worker_groups = new_groups;
6641 spin_unlock_irq(&conf->device_lock);
6642
6643 if (old_groups)
6644 kfree(old_groups[0].workers);
6645 kfree(old_groups);
6646 }
6647 mddev_resume(mddev);
6648 }
6649 mddev_unlock(mddev);
6650
6651 return err ?: len;
6652}
6653
6654static struct md_sysfs_entry
6655raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6656 raid5_show_group_thread_cnt,
6657 raid5_store_group_thread_cnt);
6658
6659static struct attribute *raid5_attrs[] = {
6660 &raid5_stripecache_size.attr,
6661 &raid5_stripecache_active.attr,
6662 &raid5_preread_bypass_threshold.attr,
6663 &raid5_group_thread_cnt.attr,
6664 &raid5_skip_copy.attr,
6665 &raid5_rmw_level.attr,
6666 &r5c_journal_mode.attr,
6667 &ppl_write_hint.attr,
6668 NULL,
6669};
6670static struct attribute_group raid5_attrs_group = {
6671 .name = NULL,
6672 .attrs = raid5_attrs,
6673};
6674
6675static int alloc_thread_groups(struct r5conf *conf, int cnt,
6676 int *group_cnt,
6677 int *worker_cnt_per_group,
6678 struct r5worker_group **worker_groups)
6679{
6680 int i, j, k;
6681 ssize_t size;
6682 struct r5worker *workers;
6683
6684 *worker_cnt_per_group = cnt;
6685 if (cnt == 0) {
6686 *group_cnt = 0;
6687 *worker_groups = NULL;
6688 return 0;
6689 }
6690 *group_cnt = num_possible_nodes();
6691 size = sizeof(struct r5worker) * cnt;
6692 workers = kcalloc(size, *group_cnt, GFP_NOIO);
6693 *worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6694 GFP_NOIO);
6695 if (!*worker_groups || !workers) {
6696 kfree(workers);
6697 kfree(*worker_groups);
6698 return -ENOMEM;
6699 }
6700
6701 for (i = 0; i < *group_cnt; i++) {
6702 struct r5worker_group *group;
6703
6704 group = &(*worker_groups)[i];
6705 INIT_LIST_HEAD(&group->handle_list);
6706 INIT_LIST_HEAD(&group->loprio_list);
6707 group->conf = conf;
6708 group->workers = workers + i * cnt;
6709
6710 for (j = 0; j < cnt; j++) {
6711 struct r5worker *worker = group->workers + j;
6712 worker->group = group;
6713 INIT_WORK(&worker->work, raid5_do_work);
6714
6715 for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6716 INIT_LIST_HEAD(worker->temp_inactive_list + k);
6717 }
6718 }
6719
6720 return 0;
6721}
6722
6723static void free_thread_groups(struct r5conf *conf)
6724{
6725 if (conf->worker_groups)
6726 kfree(conf->worker_groups[0].workers);
6727 kfree(conf->worker_groups);
6728 conf->worker_groups = NULL;
6729}
6730
6731static sector_t
6732raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6733{
6734 struct r5conf *conf = mddev->private;
6735
6736 if (!sectors)
6737 sectors = mddev->dev_sectors;
6738 if (!raid_disks)
6739 /* size is defined by the smallest of previous and new size */
6740 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6741
6742 sectors &= ~((sector_t)conf->chunk_sectors - 1);
6743 sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6744 return sectors * (raid_disks - conf->max_degraded);
6745}
6746
6747static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6748{
6749 safe_put_page(percpu->spare_page);
6750 percpu->spare_page = NULL;
6751 kvfree(percpu->scribble);
6752 percpu->scribble = NULL;
6753}
6754
6755static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6756{
6757 if (conf->level == 6 && !percpu->spare_page) {
6758 percpu->spare_page = alloc_page(GFP_KERNEL);
6759 if (!percpu->spare_page)
6760 return -ENOMEM;
6761 }
6762
6763 if (scribble_alloc(percpu,
6764 max(conf->raid_disks,
6765 conf->previous_raid_disks),
6766 max(conf->chunk_sectors,
6767 conf->prev_chunk_sectors)
6768 / STRIPE_SECTORS,
6769 GFP_KERNEL)) {
6770 free_scratch_buffer(conf, percpu);
6771 return -ENOMEM;
6772 }
6773
6774 return 0;
6775}
6776
6777static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6778{
6779 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6780
6781 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6782 return 0;
6783}
6784
6785static void raid5_free_percpu(struct r5conf *conf)
6786{
6787 if (!conf->percpu)
6788 return;
6789
6790 cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6791 free_percpu(conf->percpu);
6792}
6793
6794static void free_conf(struct r5conf *conf)
6795{
6796 int i;
6797
6798 log_exit(conf);
6799
6800 unregister_shrinker(&conf->shrinker);
6801 free_thread_groups(conf);
6802 shrink_stripes(conf);
6803 raid5_free_percpu(conf);
6804 for (i = 0; i < conf->pool_size; i++)
6805 if (conf->disks[i].extra_page)
6806 put_page(conf->disks[i].extra_page);
6807 kfree(conf->disks);
6808 bioset_exit(&conf->bio_split);
6809 kfree(conf->stripe_hashtbl);
6810 kfree(conf->pending_data);
6811 kfree(conf);
6812}
6813
6814static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6815{
6816 struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6817 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6818
6819 if (alloc_scratch_buffer(conf, percpu)) {
6820 pr_warn("%s: failed memory allocation for cpu%u\n",
6821 __func__, cpu);
6822 return -ENOMEM;
6823 }
6824 return 0;
6825}
6826
6827static int raid5_alloc_percpu(struct r5conf *conf)
6828{
6829 int err = 0;
6830
6831 conf->percpu = alloc_percpu(struct raid5_percpu);
6832 if (!conf->percpu)
6833 return -ENOMEM;
6834
6835 err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6836 if (!err) {
6837 conf->scribble_disks = max(conf->raid_disks,
6838 conf->previous_raid_disks);
6839 conf->scribble_sectors = max(conf->chunk_sectors,
6840 conf->prev_chunk_sectors);
6841 }
6842 return err;
6843}
6844
6845static unsigned long raid5_cache_scan(struct shrinker *shrink,
6846 struct shrink_control *sc)
6847{
6848 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6849 unsigned long ret = SHRINK_STOP;
6850
6851 if (mutex_trylock(&conf->cache_size_mutex)) {
6852 ret= 0;
6853 while (ret < sc->nr_to_scan &&
6854 conf->max_nr_stripes > conf->min_nr_stripes) {
6855 if (drop_one_stripe(conf) == 0) {
6856 ret = SHRINK_STOP;
6857 break;
6858 }
6859 ret++;
6860 }
6861 mutex_unlock(&conf->cache_size_mutex);
6862 }
6863 return ret;
6864}
6865
6866static unsigned long raid5_cache_count(struct shrinker *shrink,
6867 struct shrink_control *sc)
6868{
6869 struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6870
6871 if (conf->max_nr_stripes < conf->min_nr_stripes)
6872 /* unlikely, but not impossible */
6873 return 0;
6874 return conf->max_nr_stripes - conf->min_nr_stripes;
6875}
6876
6877static struct r5conf *setup_conf(struct mddev *mddev)
6878{
6879 struct r5conf *conf;
6880 int raid_disk, memory, max_disks;
6881 struct md_rdev *rdev;
6882 struct disk_info *disk;
6883 char pers_name[6];
6884 int i;
6885 int group_cnt, worker_cnt_per_group;
6886 struct r5worker_group *new_group;
6887 int ret;
6888
6889 if (mddev->new_level != 5
6890 && mddev->new_level != 4
6891 && mddev->new_level != 6) {
6892 pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6893 mdname(mddev), mddev->new_level);
6894 return ERR_PTR(-EIO);
6895 }
6896 if ((mddev->new_level == 5
6897 && !algorithm_valid_raid5(mddev->new_layout)) ||
6898 (mddev->new_level == 6
6899 && !algorithm_valid_raid6(mddev->new_layout))) {
6900 pr_warn("md/raid:%s: layout %d not supported\n",
6901 mdname(mddev), mddev->new_layout);
6902 return ERR_PTR(-EIO);
6903 }
6904 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6905 pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6906 mdname(mddev), mddev->raid_disks);
6907 return ERR_PTR(-EINVAL);
6908 }
6909
6910 if (!mddev->new_chunk_sectors ||
6911 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6912 !is_power_of_2(mddev->new_chunk_sectors)) {
6913 pr_warn("md/raid:%s: invalid chunk size %d\n",
6914 mdname(mddev), mddev->new_chunk_sectors << 9);
6915 return ERR_PTR(-EINVAL);
6916 }
6917
6918 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6919 if (conf == NULL)
6920 goto abort;
6921 INIT_LIST_HEAD(&conf->free_list);
6922 INIT_LIST_HEAD(&conf->pending_list);
6923 conf->pending_data = kcalloc(PENDING_IO_MAX,
6924 sizeof(struct r5pending_data),
6925 GFP_KERNEL);
6926 if (!conf->pending_data)
6927 goto abort;
6928 for (i = 0; i < PENDING_IO_MAX; i++)
6929 list_add(&conf->pending_data[i].sibling, &conf->free_list);
6930 /* Don't enable multi-threading by default*/
6931 if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6932 &new_group)) {
6933 conf->group_cnt = group_cnt;
6934 conf->worker_cnt_per_group = worker_cnt_per_group;
6935 conf->worker_groups = new_group;
6936 } else
6937 goto abort;
6938 spin_lock_init(&conf->device_lock);
6939 seqcount_init(&conf->gen_lock);
6940 mutex_init(&conf->cache_size_mutex);
6941 init_waitqueue_head(&conf->wait_for_quiescent);
6942 init_waitqueue_head(&conf->wait_for_stripe);
6943 init_waitqueue_head(&conf->wait_for_overlap);
6944 INIT_LIST_HEAD(&conf->handle_list);
6945 INIT_LIST_HEAD(&conf->loprio_list);
6946 INIT_LIST_HEAD(&conf->hold_list);
6947 INIT_LIST_HEAD(&conf->delayed_list);
6948 INIT_LIST_HEAD(&conf->bitmap_list);
6949 init_llist_head(&conf->released_stripes);
6950 atomic_set(&conf->active_stripes, 0);
6951 atomic_set(&conf->preread_active_stripes, 0);
6952 atomic_set(&conf->active_aligned_reads, 0);
6953 spin_lock_init(&conf->pending_bios_lock);
6954 conf->batch_bio_dispatch = true;
6955 rdev_for_each(rdev, mddev) {
6956 if (test_bit(Journal, &rdev->flags))
6957 continue;
6958 if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6959 conf->batch_bio_dispatch = false;
6960 break;
6961 }
6962 }
6963
6964 conf->bypass_threshold = BYPASS_THRESHOLD;
6965 conf->recovery_disabled = mddev->recovery_disabled - 1;
6966
6967 conf->raid_disks = mddev->raid_disks;
6968 if (mddev->reshape_position == MaxSector)
6969 conf->previous_raid_disks = mddev->raid_disks;
6970 else
6971 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6972 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6973
6974 conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
6975 GFP_KERNEL);
6976
6977 if (!conf->disks)
6978 goto abort;
6979
6980 for (i = 0; i < max_disks; i++) {
6981 conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6982 if (!conf->disks[i].extra_page)
6983 goto abort;
6984 }
6985
6986 ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
6987 if (ret)
6988 goto abort;
6989 conf->mddev = mddev;
6990
6991 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6992 goto abort;
6993
6994 /* We init hash_locks[0] separately to that it can be used
6995 * as the reference lock in the spin_lock_nest_lock() call
6996 * in lock_all_device_hash_locks_irq in order to convince
6997 * lockdep that we know what we are doing.
6998 */
6999 spin_lock_init(conf->hash_locks);
7000 for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7001 spin_lock_init(conf->hash_locks + i);
7002
7003 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7004 INIT_LIST_HEAD(conf->inactive_list + i);
7005
7006 for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7007 INIT_LIST_HEAD(conf->temp_inactive_list + i);
7008
7009 atomic_set(&conf->r5c_cached_full_stripes, 0);
7010 INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7011 atomic_set(&conf->r5c_cached_partial_stripes, 0);
7012 INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7013 atomic_set(&conf->r5c_flushing_full_stripes, 0);
7014 atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7015
7016 conf->level = mddev->new_level;
7017 conf->chunk_sectors = mddev->new_chunk_sectors;
7018 if (raid5_alloc_percpu(conf) != 0)
7019 goto abort;
7020
7021 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7022
7023 rdev_for_each(rdev, mddev) {
7024 raid_disk = rdev->raid_disk;
7025 if (raid_disk >= max_disks
7026 || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7027 continue;
7028 disk = conf->disks + raid_disk;
7029
7030 if (test_bit(Replacement, &rdev->flags)) {
7031 if (disk->replacement)
7032 goto abort;
7033 disk->replacement = rdev;
7034 } else {
7035 if (disk->rdev)
7036 goto abort;
7037 disk->rdev = rdev;
7038 }
7039
7040 if (test_bit(In_sync, &rdev->flags)) {
7041 char b[BDEVNAME_SIZE];
7042 pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7043 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7044 } else if (rdev->saved_raid_disk != raid_disk)
7045 /* Cannot rely on bitmap to complete recovery */
7046 conf->fullsync = 1;
7047 }
7048
7049 conf->level = mddev->new_level;
7050 if (conf->level == 6) {
7051 conf->max_degraded = 2;
7052 if (raid6_call.xor_syndrome)
7053 conf->rmw_level = PARITY_ENABLE_RMW;
7054 else
7055 conf->rmw_level = PARITY_DISABLE_RMW;
7056 } else {
7057 conf->max_degraded = 1;
7058 conf->rmw_level = PARITY_ENABLE_RMW;
7059 }
7060 conf->algorithm = mddev->new_layout;
7061 conf->reshape_progress = mddev->reshape_position;
7062 if (conf->reshape_progress != MaxSector) {
7063 conf->prev_chunk_sectors = mddev->chunk_sectors;
7064 conf->prev_algo = mddev->layout;
7065 } else {
7066 conf->prev_chunk_sectors = conf->chunk_sectors;
7067 conf->prev_algo = conf->algorithm;
7068 }
7069
7070 conf->min_nr_stripes = NR_STRIPES;
7071 if (mddev->reshape_position != MaxSector) {
7072 int stripes = max_t(int,
7073 ((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7074 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7075 conf->min_nr_stripes = max(NR_STRIPES, stripes);
7076 if (conf->min_nr_stripes != NR_STRIPES)
7077 pr_info("md/raid:%s: force stripe size %d for reshape\n",
7078 mdname(mddev), conf->min_nr_stripes);
7079 }
7080 memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7081 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7082 atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7083 if (grow_stripes(conf, conf->min_nr_stripes)) {
7084 pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7085 mdname(mddev), memory);
7086 goto abort;
7087 } else
7088 pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7089 /*
7090 * Losing a stripe head costs more than the time to refill it,
7091 * it reduces the queue depth and so can hurt throughput.
7092 * So set it rather large, scaled by number of devices.
7093 */
7094 conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7095 conf->shrinker.scan_objects = raid5_cache_scan;
7096 conf->shrinker.count_objects = raid5_cache_count;
7097 conf->shrinker.batch = 128;
7098 conf->shrinker.flags = 0;
7099 if (register_shrinker(&conf->shrinker)) {
7100 pr_warn("md/raid:%s: couldn't register shrinker.\n",
7101 mdname(mddev));
7102 goto abort;
7103 }
7104
7105 sprintf(pers_name, "raid%d", mddev->new_level);
7106 conf->thread = md_register_thread(raid5d, mddev, pers_name);
7107 if (!conf->thread) {
7108 pr_warn("md/raid:%s: couldn't allocate thread.\n",
7109 mdname(mddev));
7110 goto abort;
7111 }
7112
7113 return conf;
7114
7115 abort:
7116 if (conf) {
7117 free_conf(conf);
7118 return ERR_PTR(-EIO);
7119 } else
7120 return ERR_PTR(-ENOMEM);
7121}
7122
7123static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7124{
7125 switch (algo) {
7126 case ALGORITHM_PARITY_0:
7127 if (raid_disk < max_degraded)
7128 return 1;
7129 break;
7130 case ALGORITHM_PARITY_N:
7131 if (raid_disk >= raid_disks - max_degraded)
7132 return 1;
7133 break;
7134 case ALGORITHM_PARITY_0_6:
7135 if (raid_disk == 0 ||
7136 raid_disk == raid_disks - 1)
7137 return 1;
7138 break;
7139 case ALGORITHM_LEFT_ASYMMETRIC_6:
7140 case ALGORITHM_RIGHT_ASYMMETRIC_6:
7141 case ALGORITHM_LEFT_SYMMETRIC_6:
7142 case ALGORITHM_RIGHT_SYMMETRIC_6:
7143 if (raid_disk == raid_disks - 1)
7144 return 1;
7145 }
7146 return 0;
7147}
7148
7149static int raid5_run(struct mddev *mddev)
7150{
7151 struct r5conf *conf;
7152 int working_disks = 0;
7153 int dirty_parity_disks = 0;
7154 struct md_rdev *rdev;
7155 struct md_rdev *journal_dev = NULL;
7156 sector_t reshape_offset = 0;
7157 int i;
7158 long long min_offset_diff = 0;
7159 int first = 1;
7160
7161 if (mddev_init_writes_pending(mddev) < 0)
7162 return -ENOMEM;
7163
7164 if (mddev->recovery_cp != MaxSector)
7165 pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7166 mdname(mddev));
7167
7168 rdev_for_each(rdev, mddev) {
7169 long long diff;
7170
7171 if (test_bit(Journal, &rdev->flags)) {
7172 journal_dev = rdev;
7173 continue;
7174 }
7175 if (rdev->raid_disk < 0)
7176 continue;
7177 diff = (rdev->new_data_offset - rdev->data_offset);
7178 if (first) {
7179 min_offset_diff = diff;
7180 first = 0;
7181 } else if (mddev->reshape_backwards &&
7182 diff < min_offset_diff)
7183 min_offset_diff = diff;
7184 else if (!mddev->reshape_backwards &&
7185 diff > min_offset_diff)
7186 min_offset_diff = diff;
7187 }
7188
7189 if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7190 (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7191 pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7192 mdname(mddev));
7193 return -EINVAL;
7194 }
7195
7196 if (mddev->reshape_position != MaxSector) {
7197 /* Check that we can continue the reshape.
7198 * Difficulties arise if the stripe we would write to
7199 * next is at or after the stripe we would read from next.
7200 * For a reshape that changes the number of devices, this
7201 * is only possible for a very short time, and mdadm makes
7202 * sure that time appears to have past before assembling
7203 * the array. So we fail if that time hasn't passed.
7204 * For a reshape that keeps the number of devices the same
7205 * mdadm must be monitoring the reshape can keeping the
7206 * critical areas read-only and backed up. It will start
7207 * the array in read-only mode, so we check for that.
7208 */
7209 sector_t here_new, here_old;
7210 int old_disks;
7211 int max_degraded = (mddev->level == 6 ? 2 : 1);
7212 int chunk_sectors;
7213 int new_data_disks;
7214
7215 if (journal_dev) {
7216 pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7217 mdname(mddev));
7218 return -EINVAL;
7219 }
7220
7221 if (mddev->new_level != mddev->level) {
7222 pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7223 mdname(mddev));
7224 return -EINVAL;
7225 }
7226 old_disks = mddev->raid_disks - mddev->delta_disks;
7227 /* reshape_position must be on a new-stripe boundary, and one
7228 * further up in new geometry must map after here in old
7229 * geometry.
7230 * If the chunk sizes are different, then as we perform reshape
7231 * in units of the largest of the two, reshape_position needs
7232 * be a multiple of the largest chunk size times new data disks.
7233 */
7234 here_new = mddev->reshape_position;
7235 chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7236 new_data_disks = mddev->raid_disks - max_degraded;
7237 if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7238 pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7239 mdname(mddev));
7240 return -EINVAL;
7241 }
7242 reshape_offset = here_new * chunk_sectors;
7243 /* here_new is the stripe we will write to */
7244 here_old = mddev->reshape_position;
7245 sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7246 /* here_old is the first stripe that we might need to read
7247 * from */
7248 if (mddev->delta_disks == 0) {
7249 /* We cannot be sure it is safe to start an in-place
7250 * reshape. It is only safe if user-space is monitoring
7251 * and taking constant backups.
7252 * mdadm always starts a situation like this in
7253 * readonly mode so it can take control before
7254 * allowing any writes. So just check for that.
7255 */
7256 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7257 abs(min_offset_diff) >= mddev->new_chunk_sectors)
7258 /* not really in-place - so OK */;
7259 else if (mddev->ro == 0) {
7260 pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7261 mdname(mddev));
7262 return -EINVAL;
7263 }
7264 } else if (mddev->reshape_backwards
7265 ? (here_new * chunk_sectors + min_offset_diff <=
7266 here_old * chunk_sectors)
7267 : (here_new * chunk_sectors >=
7268 here_old * chunk_sectors + (-min_offset_diff))) {
7269 /* Reading from the same stripe as writing to - bad */
7270 pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7271 mdname(mddev));
7272 return -EINVAL;
7273 }
7274 pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7275 /* OK, we should be able to continue; */
7276 } else {
7277 BUG_ON(mddev->level != mddev->new_level);
7278 BUG_ON(mddev->layout != mddev->new_layout);
7279 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7280 BUG_ON(mddev->delta_disks != 0);
7281 }
7282
7283 if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7284 test_bit(MD_HAS_PPL, &mddev->flags)) {
7285 pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7286 mdname(mddev));
7287 clear_bit(MD_HAS_PPL, &mddev->flags);
7288 clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7289 }
7290
7291 if (mddev->private == NULL)
7292 conf = setup_conf(mddev);
7293 else
7294 conf = mddev->private;
7295
7296 if (IS_ERR(conf))
7297 return PTR_ERR(conf);
7298
7299 if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7300 if (!journal_dev) {
7301 pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7302 mdname(mddev));
7303 mddev->ro = 1;
7304 set_disk_ro(mddev->gendisk, 1);
7305 } else if (mddev->recovery_cp == MaxSector)
7306 set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7307 }
7308
7309 conf->min_offset_diff = min_offset_diff;
7310 mddev->thread = conf->thread;
7311 conf->thread = NULL;
7312 mddev->private = conf;
7313
7314 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7315 i++) {
7316 rdev = conf->disks[i].rdev;
7317 if (!rdev && conf->disks[i].replacement) {
7318 /* The replacement is all we have yet */
7319 rdev = conf->disks[i].replacement;
7320 conf->disks[i].replacement = NULL;
7321 clear_bit(Replacement, &rdev->flags);
7322 conf->disks[i].rdev = rdev;
7323 }
7324 if (!rdev)
7325 continue;
7326 if (conf->disks[i].replacement &&
7327 conf->reshape_progress != MaxSector) {
7328 /* replacements and reshape simply do not mix. */
7329 pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7330 goto abort;
7331 }
7332 if (test_bit(In_sync, &rdev->flags)) {
7333 working_disks++;
7334 continue;
7335 }
7336 /* This disc is not fully in-sync. However if it
7337 * just stored parity (beyond the recovery_offset),
7338 * when we don't need to be concerned about the
7339 * array being dirty.
7340 * When reshape goes 'backwards', we never have
7341 * partially completed devices, so we only need
7342 * to worry about reshape going forwards.
7343 */
7344 /* Hack because v0.91 doesn't store recovery_offset properly. */
7345 if (mddev->major_version == 0 &&
7346 mddev->minor_version > 90)
7347 rdev->recovery_offset = reshape_offset;
7348
7349 if (rdev->recovery_offset < reshape_offset) {
7350 /* We need to check old and new layout */
7351 if (!only_parity(rdev->raid_disk,
7352 conf->algorithm,
7353 conf->raid_disks,
7354 conf->max_degraded))
7355 continue;
7356 }
7357 if (!only_parity(rdev->raid_disk,
7358 conf->prev_algo,
7359 conf->previous_raid_disks,
7360 conf->max_degraded))
7361 continue;
7362 dirty_parity_disks++;
7363 }
7364
7365 /*
7366 * 0 for a fully functional array, 1 or 2 for a degraded array.
7367 */
7368 mddev->degraded = raid5_calc_degraded(conf);
7369
7370 if (has_failed(conf)) {
7371 pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7372 mdname(mddev), mddev->degraded, conf->raid_disks);
7373 goto abort;
7374 }
7375
7376 /* device size must be a multiple of chunk size */
7377 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7378 mddev->resync_max_sectors = mddev->dev_sectors;
7379
7380 if (mddev->degraded > dirty_parity_disks &&
7381 mddev->recovery_cp != MaxSector) {
7382 if (test_bit(MD_HAS_PPL, &mddev->flags))
7383 pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7384 mdname(mddev));
7385 else if (mddev->ok_start_degraded)
7386 pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7387 mdname(mddev));
7388 else {
7389 pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7390 mdname(mddev));
7391 goto abort;
7392 }
7393 }
7394
7395 pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7396 mdname(mddev), conf->level,
7397 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7398 mddev->new_layout);
7399
7400 print_raid5_conf(conf);
7401
7402 if (conf->reshape_progress != MaxSector) {
7403 conf->reshape_safe = conf->reshape_progress;
7404 atomic_set(&conf->reshape_stripes, 0);
7405 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7406 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7407 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7408 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7409 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7410 "reshape");
7411 if (!mddev->sync_thread)
7412 goto abort;
7413 }
7414
7415 /* Ok, everything is just fine now */
7416 if (mddev->to_remove == &raid5_attrs_group)
7417 mddev->to_remove = NULL;
7418 else if (mddev->kobj.sd &&
7419 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7420 pr_warn("raid5: failed to create sysfs attributes for %s\n",
7421 mdname(mddev));
7422 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7423
7424 if (mddev->queue) {
7425 int chunk_size;
7426 /* read-ahead size must cover two whole stripes, which
7427 * is 2 * (datadisks) * chunksize where 'n' is the
7428 * number of raid devices
7429 */
7430 int data_disks = conf->previous_raid_disks - conf->max_degraded;
7431 int stripe = data_disks *
7432 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
7433 if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7434 mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7435
7436 chunk_size = mddev->chunk_sectors << 9;
7437 blk_queue_io_min(mddev->queue, chunk_size);
7438 blk_queue_io_opt(mddev->queue, chunk_size *
7439 (conf->raid_disks - conf->max_degraded));
7440 mddev->queue->limits.raid_partial_stripes_expensive = 1;
7441 /*
7442 * We can only discard a whole stripe. It doesn't make sense to
7443 * discard data disk but write parity disk
7444 */
7445 stripe = stripe * PAGE_SIZE;
7446 /* Round up to power of 2, as discard handling
7447 * currently assumes that */
7448 while ((stripe-1) & stripe)
7449 stripe = (stripe | (stripe-1)) + 1;
7450 mddev->queue->limits.discard_alignment = stripe;
7451 mddev->queue->limits.discard_granularity = stripe;
7452
7453 blk_queue_max_write_same_sectors(mddev->queue, 0);
7454 blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7455
7456 rdev_for_each(rdev, mddev) {
7457 disk_stack_limits(mddev->gendisk, rdev->bdev,
7458 rdev->data_offset << 9);
7459 disk_stack_limits(mddev->gendisk, rdev->bdev,
7460 rdev->new_data_offset << 9);
7461 }
7462
7463 /*
7464 * zeroing is required, otherwise data
7465 * could be lost. Consider a scenario: discard a stripe
7466 * (the stripe could be inconsistent if
7467 * discard_zeroes_data is 0); write one disk of the
7468 * stripe (the stripe could be inconsistent again
7469 * depending on which disks are used to calculate
7470 * parity); the disk is broken; The stripe data of this
7471 * disk is lost.
7472 *
7473 * We only allow DISCARD if the sysadmin has confirmed that
7474 * only safe devices are in use by setting a module parameter.
7475 * A better idea might be to turn DISCARD into WRITE_ZEROES
7476 * requests, as that is required to be safe.
7477 */
7478 if (devices_handle_discard_safely &&
7479 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7480 mddev->queue->limits.discard_granularity >= stripe)
7481 blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7482 mddev->queue);
7483 else
7484 blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7485 mddev->queue);
7486
7487 blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7488 }
7489
7490 if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7491 goto abort;
7492
7493 return 0;
7494abort:
7495 md_unregister_thread(&mddev->thread);
7496 print_raid5_conf(conf);
7497 free_conf(conf);
7498 mddev->private = NULL;
7499 pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7500 return -EIO;
7501}
7502
7503static void raid5_free(struct mddev *mddev, void *priv)
7504{
7505 struct r5conf *conf = priv;
7506
7507 free_conf(conf);
7508 mddev->to_remove = &raid5_attrs_group;
7509}
7510
7511static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7512{
7513 struct r5conf *conf = mddev->private;
7514 int i;
7515
7516 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7517 conf->chunk_sectors / 2, mddev->layout);
7518 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7519 rcu_read_lock();
7520 for (i = 0; i < conf->raid_disks; i++) {
7521 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7522 seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7523 }
7524 rcu_read_unlock();
7525 seq_printf (seq, "]");
7526}
7527
7528static void print_raid5_conf (struct r5conf *conf)
7529{
7530 int i;
7531 struct disk_info *tmp;
7532
7533 pr_debug("RAID conf printout:\n");
7534 if (!conf) {
7535 pr_debug("(conf==NULL)\n");
7536 return;
7537 }
7538 pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7539 conf->raid_disks,
7540 conf->raid_disks - conf->mddev->degraded);
7541
7542 for (i = 0; i < conf->raid_disks; i++) {
7543 char b[BDEVNAME_SIZE];
7544 tmp = conf->disks + i;
7545 if (tmp->rdev)
7546 pr_debug(" disk %d, o:%d, dev:%s\n",
7547 i, !test_bit(Faulty, &tmp->rdev->flags),
7548 bdevname(tmp->rdev->bdev, b));
7549 }
7550}
7551
7552static int raid5_spare_active(struct mddev *mddev)
7553{
7554 int i;
7555 struct r5conf *conf = mddev->private;
7556 struct disk_info *tmp;
7557 int count = 0;
7558 unsigned long flags;
7559
7560 for (i = 0; i < conf->raid_disks; i++) {
7561 tmp = conf->disks + i;
7562 if (tmp->replacement
7563 && tmp->replacement->recovery_offset == MaxSector
7564 && !test_bit(Faulty, &tmp->replacement->flags)
7565 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7566 /* Replacement has just become active. */
7567 if (!tmp->rdev
7568 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7569 count++;
7570 if (tmp->rdev) {
7571 /* Replaced device not technically faulty,
7572 * but we need to be sure it gets removed
7573 * and never re-added.
7574 */
7575 set_bit(Faulty, &tmp->rdev->flags);
7576 sysfs_notify_dirent_safe(
7577 tmp->rdev->sysfs_state);
7578 }
7579 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7580 } else if (tmp->rdev
7581 && tmp->rdev->recovery_offset == MaxSector
7582 && !test_bit(Faulty, &tmp->rdev->flags)
7583 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7584 count++;
7585 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7586 }
7587 }
7588 spin_lock_irqsave(&conf->device_lock, flags);
7589 mddev->degraded = raid5_calc_degraded(conf);
7590 spin_unlock_irqrestore(&conf->device_lock, flags);
7591 print_raid5_conf(conf);
7592 return count;
7593}
7594
7595static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7596{
7597 struct r5conf *conf = mddev->private;
7598 int err = 0;
7599 int number = rdev->raid_disk;
7600 struct md_rdev **rdevp;
7601 struct disk_info *p = conf->disks + number;
7602
7603 print_raid5_conf(conf);
7604 if (test_bit(Journal, &rdev->flags) && conf->log) {
7605 /*
7606 * we can't wait pending write here, as this is called in
7607 * raid5d, wait will deadlock.
7608 * neilb: there is no locking about new writes here,
7609 * so this cannot be safe.
7610 */
7611 if (atomic_read(&conf->active_stripes) ||
7612 atomic_read(&conf->r5c_cached_full_stripes) ||
7613 atomic_read(&conf->r5c_cached_partial_stripes)) {
7614 return -EBUSY;
7615 }
7616 log_exit(conf);
7617 return 0;
7618 }
7619 if (rdev == p->rdev)
7620 rdevp = &p->rdev;
7621 else if (rdev == p->replacement)
7622 rdevp = &p->replacement;
7623 else
7624 return 0;
7625
7626 if (number >= conf->raid_disks &&
7627 conf->reshape_progress == MaxSector)
7628 clear_bit(In_sync, &rdev->flags);
7629
7630 if (test_bit(In_sync, &rdev->flags) ||
7631 atomic_read(&rdev->nr_pending)) {
7632 err = -EBUSY;
7633 goto abort;
7634 }
7635 /* Only remove non-faulty devices if recovery
7636 * isn't possible.
7637 */
7638 if (!test_bit(Faulty, &rdev->flags) &&
7639 mddev->recovery_disabled != conf->recovery_disabled &&
7640 !has_failed(conf) &&
7641 (!p->replacement || p->replacement == rdev) &&
7642 number < conf->raid_disks) {
7643 err = -EBUSY;
7644 goto abort;
7645 }
7646 *rdevp = NULL;
7647 if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7648 synchronize_rcu();
7649 if (atomic_read(&rdev->nr_pending)) {
7650 /* lost the race, try later */
7651 err = -EBUSY;
7652 *rdevp = rdev;
7653 }
7654 }
7655 if (!err) {
7656 err = log_modify(conf, rdev, false);
7657 if (err)
7658 goto abort;
7659 }
7660 if (p->replacement) {
7661 /* We must have just cleared 'rdev' */
7662 p->rdev = p->replacement;
7663 clear_bit(Replacement, &p->replacement->flags);
7664 smp_mb(); /* Make sure other CPUs may see both as identical
7665 * but will never see neither - if they are careful
7666 */
7667 p->replacement = NULL;
7668
7669 if (!err)
7670 err = log_modify(conf, p->rdev, true);
7671 }
7672
7673 clear_bit(WantReplacement, &rdev->flags);
7674abort:
7675
7676 print_raid5_conf(conf);
7677 return err;
7678}
7679
7680static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7681{
7682 struct r5conf *conf = mddev->private;
7683 int ret, err = -EEXIST;
7684 int disk;
7685 struct disk_info *p;
7686 int first = 0;
7687 int last = conf->raid_disks - 1;
7688
7689 if (test_bit(Journal, &rdev->flags)) {
7690 if (conf->log)
7691 return -EBUSY;
7692
7693 rdev->raid_disk = 0;
7694 /*
7695 * The array is in readonly mode if journal is missing, so no
7696 * write requests running. We should be safe
7697 */
7698 ret = log_init(conf, rdev, false);
7699 if (ret)
7700 return ret;
7701
7702 ret = r5l_start(conf->log);
7703 if (ret)
7704 return ret;
7705
7706 return 0;
7707 }
7708 if (mddev->recovery_disabled == conf->recovery_disabled)
7709 return -EBUSY;
7710
7711 if (rdev->saved_raid_disk < 0 && has_failed(conf))
7712 /* no point adding a device */
7713 return -EINVAL;
7714
7715 if (rdev->raid_disk >= 0)
7716 first = last = rdev->raid_disk;
7717
7718 /*
7719 * find the disk ... but prefer rdev->saved_raid_disk
7720 * if possible.
7721 */
7722 if (rdev->saved_raid_disk >= 0 &&
7723 rdev->saved_raid_disk >= first &&
7724 conf->disks[rdev->saved_raid_disk].rdev == NULL)
7725 first = rdev->saved_raid_disk;
7726
7727 for (disk = first; disk <= last; disk++) {
7728 p = conf->disks + disk;
7729 if (p->rdev == NULL) {
7730 clear_bit(In_sync, &rdev->flags);
7731 rdev->raid_disk = disk;
7732 if (rdev->saved_raid_disk != disk)
7733 conf->fullsync = 1;
7734 rcu_assign_pointer(p->rdev, rdev);
7735
7736 err = log_modify(conf, rdev, true);
7737
7738 goto out;
7739 }
7740 }
7741 for (disk = first; disk <= last; disk++) {
7742 p = conf->disks + disk;
7743 if (test_bit(WantReplacement, &p->rdev->flags) &&
7744 p->replacement == NULL) {
7745 clear_bit(In_sync, &rdev->flags);
7746 set_bit(Replacement, &rdev->flags);
7747 rdev->raid_disk = disk;
7748 err = 0;
7749 conf->fullsync = 1;
7750 rcu_assign_pointer(p->replacement, rdev);
7751 break;
7752 }
7753 }
7754out:
7755 print_raid5_conf(conf);
7756 return err;
7757}
7758
7759static int raid5_resize(struct mddev *mddev, sector_t sectors)
7760{
7761 /* no resync is happening, and there is enough space
7762 * on all devices, so we can resize.
7763 * We need to make sure resync covers any new space.
7764 * If the array is shrinking we should possibly wait until
7765 * any io in the removed space completes, but it hardly seems
7766 * worth it.
7767 */
7768 sector_t newsize;
7769 struct r5conf *conf = mddev->private;
7770
7771 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7772 return -EINVAL;
7773 sectors &= ~((sector_t)conf->chunk_sectors - 1);
7774 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7775 if (mddev->external_size &&
7776 mddev->array_sectors > newsize)
7777 return -EINVAL;
7778 if (mddev->bitmap) {
7779 int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
7780 if (ret)
7781 return ret;
7782 }
7783 md_set_array_sectors(mddev, newsize);
7784 if (sectors > mddev->dev_sectors &&
7785 mddev->recovery_cp > mddev->dev_sectors) {
7786 mddev->recovery_cp = mddev->dev_sectors;
7787 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7788 }
7789 mddev->dev_sectors = sectors;
7790 mddev->resync_max_sectors = sectors;
7791 return 0;
7792}
7793
7794static int check_stripe_cache(struct mddev *mddev)
7795{
7796 /* Can only proceed if there are plenty of stripe_heads.
7797 * We need a minimum of one full stripe,, and for sensible progress
7798 * it is best to have about 4 times that.
7799 * If we require 4 times, then the default 256 4K stripe_heads will
7800 * allow for chunk sizes up to 256K, which is probably OK.
7801 * If the chunk size is greater, user-space should request more
7802 * stripe_heads first.
7803 */
7804 struct r5conf *conf = mddev->private;
7805 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7806 > conf->min_nr_stripes ||
7807 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7808 > conf->min_nr_stripes) {
7809 pr_warn("md/raid:%s: reshape: not enough stripes. Needed %lu\n",
7810 mdname(mddev),
7811 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7812 / STRIPE_SIZE)*4);
7813 return 0;
7814 }
7815 return 1;
7816}
7817
7818static int check_reshape(struct mddev *mddev)
7819{
7820 struct r5conf *conf = mddev->private;
7821
7822 if (raid5_has_log(conf) || raid5_has_ppl(conf))
7823 return -EINVAL;
7824 if (mddev->delta_disks == 0 &&
7825 mddev->new_layout == mddev->layout &&
7826 mddev->new_chunk_sectors == mddev->chunk_sectors)
7827 return 0; /* nothing to do */
7828 if (has_failed(conf))
7829 return -EINVAL;
7830 if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7831 /* We might be able to shrink, but the devices must
7832 * be made bigger first.
7833 * For raid6, 4 is the minimum size.
7834 * Otherwise 2 is the minimum
7835 */
7836 int min = 2;
7837 if (mddev->level == 6)
7838 min = 4;
7839 if (mddev->raid_disks + mddev->delta_disks < min)
7840 return -EINVAL;
7841 }
7842
7843 if (!check_stripe_cache(mddev))
7844 return -ENOSPC;
7845
7846 if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7847 mddev->delta_disks > 0)
7848 if (resize_chunks(conf,
7849 conf->previous_raid_disks
7850 + max(0, mddev->delta_disks),
7851 max(mddev->new_chunk_sectors,
7852 mddev->chunk_sectors)
7853 ) < 0)
7854 return -ENOMEM;
7855
7856 if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7857 return 0; /* never bother to shrink */
7858 return resize_stripes(conf, (conf->previous_raid_disks
7859 + mddev->delta_disks));
7860}
7861
7862static int raid5_start_reshape(struct mddev *mddev)
7863{
7864 struct r5conf *conf = mddev->private;
7865 struct md_rdev *rdev;
7866 int spares = 0;
7867 unsigned long flags;
7868
7869 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7870 return -EBUSY;
7871
7872 if (!check_stripe_cache(mddev))
7873 return -ENOSPC;
7874
7875 if (has_failed(conf))
7876 return -EINVAL;
7877
7878 rdev_for_each(rdev, mddev) {
7879 if (!test_bit(In_sync, &rdev->flags)
7880 && !test_bit(Faulty, &rdev->flags))
7881 spares++;
7882 }
7883
7884 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7885 /* Not enough devices even to make a degraded array
7886 * of that size
7887 */
7888 return -EINVAL;
7889
7890 /* Refuse to reduce size of the array. Any reductions in
7891 * array size must be through explicit setting of array_size
7892 * attribute.
7893 */
7894 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7895 < mddev->array_sectors) {
7896 pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7897 mdname(mddev));
7898 return -EINVAL;
7899 }
7900
7901 atomic_set(&conf->reshape_stripes, 0);
7902 spin_lock_irq(&conf->device_lock);
7903 write_seqcount_begin(&conf->gen_lock);
7904 conf->previous_raid_disks = conf->raid_disks;
7905 conf->raid_disks += mddev->delta_disks;
7906 conf->prev_chunk_sectors = conf->chunk_sectors;
7907 conf->chunk_sectors = mddev->new_chunk_sectors;
7908 conf->prev_algo = conf->algorithm;
7909 conf->algorithm = mddev->new_layout;
7910 conf->generation++;
7911 /* Code that selects data_offset needs to see the generation update
7912 * if reshape_progress has been set - so a memory barrier needed.
7913 */
7914 smp_mb();
7915 if (mddev->reshape_backwards)
7916 conf->reshape_progress = raid5_size(mddev, 0, 0);
7917 else
7918 conf->reshape_progress = 0;
7919 conf->reshape_safe = conf->reshape_progress;
7920 write_seqcount_end(&conf->gen_lock);
7921 spin_unlock_irq(&conf->device_lock);
7922
7923 /* Now make sure any requests that proceeded on the assumption
7924 * the reshape wasn't running - like Discard or Read - have
7925 * completed.
7926 */
7927 mddev_suspend(mddev);
7928 mddev_resume(mddev);
7929
7930 /* Add some new drives, as many as will fit.
7931 * We know there are enough to make the newly sized array work.
7932 * Don't add devices if we are reducing the number of
7933 * devices in the array. This is because it is not possible
7934 * to correctly record the "partially reconstructed" state of
7935 * such devices during the reshape and confusion could result.
7936 */
7937 if (mddev->delta_disks >= 0) {
7938 rdev_for_each(rdev, mddev)
7939 if (rdev->raid_disk < 0 &&
7940 !test_bit(Faulty, &rdev->flags)) {
7941 if (raid5_add_disk(mddev, rdev) == 0) {
7942 if (rdev->raid_disk
7943 >= conf->previous_raid_disks)
7944 set_bit(In_sync, &rdev->flags);
7945 else
7946 rdev->recovery_offset = 0;
7947
7948 if (sysfs_link_rdev(mddev, rdev))
7949 /* Failure here is OK */;
7950 }
7951 } else if (rdev->raid_disk >= conf->previous_raid_disks
7952 && !test_bit(Faulty, &rdev->flags)) {
7953 /* This is a spare that was manually added */
7954 set_bit(In_sync, &rdev->flags);
7955 }
7956
7957 /* When a reshape changes the number of devices,
7958 * ->degraded is measured against the larger of the
7959 * pre and post number of devices.
7960 */
7961 spin_lock_irqsave(&conf->device_lock, flags);
7962 mddev->degraded = raid5_calc_degraded(conf);
7963 spin_unlock_irqrestore(&conf->device_lock, flags);
7964 }
7965 mddev->raid_disks = conf->raid_disks;
7966 mddev->reshape_position = conf->reshape_progress;
7967 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7968
7969 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7970 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7971 clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7972 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7973 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7974 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7975 "reshape");
7976 if (!mddev->sync_thread) {
7977 mddev->recovery = 0;
7978 spin_lock_irq(&conf->device_lock);
7979 write_seqcount_begin(&conf->gen_lock);
7980 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7981 mddev->new_chunk_sectors =
7982 conf->chunk_sectors = conf->prev_chunk_sectors;
7983 mddev->new_layout = conf->algorithm = conf->prev_algo;
7984 rdev_for_each(rdev, mddev)
7985 rdev->new_data_offset = rdev->data_offset;
7986 smp_wmb();
7987 conf->generation --;
7988 conf->reshape_progress = MaxSector;
7989 mddev->reshape_position = MaxSector;
7990 write_seqcount_end(&conf->gen_lock);
7991 spin_unlock_irq(&conf->device_lock);
7992 return -EAGAIN;
7993 }
7994 conf->reshape_checkpoint = jiffies;
7995 md_wakeup_thread(mddev->sync_thread);
7996 md_new_event(mddev);
7997 return 0;
7998}
7999
8000/* This is called from the reshape thread and should make any
8001 * changes needed in 'conf'
8002 */
8003static void end_reshape(struct r5conf *conf)
8004{
8005
8006 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8007 struct md_rdev *rdev;
8008
8009 spin_lock_irq(&conf->device_lock);
8010 conf->previous_raid_disks = conf->raid_disks;
8011 md_finish_reshape(conf->mddev);
8012 smp_wmb();
8013 conf->reshape_progress = MaxSector;
8014 conf->mddev->reshape_position = MaxSector;
8015 rdev_for_each(rdev, conf->mddev)
8016 if (rdev->raid_disk >= 0 &&
8017 !test_bit(Journal, &rdev->flags) &&
8018 !test_bit(In_sync, &rdev->flags))
8019 rdev->recovery_offset = MaxSector;
8020 spin_unlock_irq(&conf->device_lock);
8021 wake_up(&conf->wait_for_overlap);
8022
8023 /* read-ahead size must cover two whole stripes, which is
8024 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
8025 */
8026 if (conf->mddev->queue) {
8027 int data_disks = conf->raid_disks - conf->max_degraded;
8028 int stripe = data_disks * ((conf->chunk_sectors << 9)
8029 / PAGE_SIZE);
8030 if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
8031 conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
8032 }
8033 }
8034}
8035
8036/* This is called from the raid5d thread with mddev_lock held.
8037 * It makes config changes to the device.
8038 */
8039static void raid5_finish_reshape(struct mddev *mddev)
8040{
8041 struct r5conf *conf = mddev->private;
8042
8043 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8044
8045 if (mddev->delta_disks <= 0) {
8046 int d;
8047 spin_lock_irq(&conf->device_lock);
8048 mddev->degraded = raid5_calc_degraded(conf);
8049 spin_unlock_irq(&conf->device_lock);
8050 for (d = conf->raid_disks ;
8051 d < conf->raid_disks - mddev->delta_disks;
8052 d++) {
8053 struct md_rdev *rdev = conf->disks[d].rdev;
8054 if (rdev)
8055 clear_bit(In_sync, &rdev->flags);
8056 rdev = conf->disks[d].replacement;
8057 if (rdev)
8058 clear_bit(In_sync, &rdev->flags);
8059 }
8060 }
8061 mddev->layout = conf->algorithm;
8062 mddev->chunk_sectors = conf->chunk_sectors;
8063 mddev->reshape_position = MaxSector;
8064 mddev->delta_disks = 0;
8065 mddev->reshape_backwards = 0;
8066 }
8067}
8068
8069static void raid5_quiesce(struct mddev *mddev, int quiesce)
8070{
8071 struct r5conf *conf = mddev->private;
8072
8073 if (quiesce) {
8074 /* stop all writes */
8075 lock_all_device_hash_locks_irq(conf);
8076 /* '2' tells resync/reshape to pause so that all
8077 * active stripes can drain
8078 */
8079 r5c_flush_cache(conf, INT_MAX);
8080 conf->quiesce = 2;
8081 wait_event_cmd(conf->wait_for_quiescent,
8082 atomic_read(&conf->active_stripes) == 0 &&
8083 atomic_read(&conf->active_aligned_reads) == 0,
8084 unlock_all_device_hash_locks_irq(conf),
8085 lock_all_device_hash_locks_irq(conf));
8086 conf->quiesce = 1;
8087 unlock_all_device_hash_locks_irq(conf);
8088 /* allow reshape to continue */
8089 wake_up(&conf->wait_for_overlap);
8090 } else {
8091 /* re-enable writes */
8092 lock_all_device_hash_locks_irq(conf);
8093 conf->quiesce = 0;
8094 wake_up(&conf->wait_for_quiescent);
8095 wake_up(&conf->wait_for_overlap);
8096 unlock_all_device_hash_locks_irq(conf);
8097 }
8098 log_quiesce(conf, quiesce);
8099}
8100
8101static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8102{
8103 struct r0conf *raid0_conf = mddev->private;
8104 sector_t sectors;
8105
8106 /* for raid0 takeover only one zone is supported */
8107 if (raid0_conf->nr_strip_zones > 1) {
8108 pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8109 mdname(mddev));
8110 return ERR_PTR(-EINVAL);
8111 }
8112
8113 sectors = raid0_conf->strip_zone[0].zone_end;
8114 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8115 mddev->dev_sectors = sectors;
8116 mddev->new_level = level;
8117 mddev->new_layout = ALGORITHM_PARITY_N;
8118 mddev->new_chunk_sectors = mddev->chunk_sectors;
8119 mddev->raid_disks += 1;
8120 mddev->delta_disks = 1;
8121 /* make sure it will be not marked as dirty */
8122 mddev->recovery_cp = MaxSector;
8123
8124 return setup_conf(mddev);
8125}
8126
8127static void *raid5_takeover_raid1(struct mddev *mddev)
8128{
8129 int chunksect;
8130 void *ret;
8131
8132 if (mddev->raid_disks != 2 ||
8133 mddev->degraded > 1)
8134 return ERR_PTR(-EINVAL);
8135
8136 /* Should check if there are write-behind devices? */
8137
8138 chunksect = 64*2; /* 64K by default */
8139
8140 /* The array must be an exact multiple of chunksize */
8141 while (chunksect && (mddev->array_sectors & (chunksect-1)))
8142 chunksect >>= 1;
8143
8144 if ((chunksect<<9) < STRIPE_SIZE)
8145 /* array size does not allow a suitable chunk size */
8146 return ERR_PTR(-EINVAL);
8147
8148 mddev->new_level = 5;
8149 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8150 mddev->new_chunk_sectors = chunksect;
8151
8152 ret = setup_conf(mddev);
8153 if (!IS_ERR(ret))
8154 mddev_clear_unsupported_flags(mddev,
8155 UNSUPPORTED_MDDEV_FLAGS);
8156 return ret;
8157}
8158
8159static void *raid5_takeover_raid6(struct mddev *mddev)
8160{
8161 int new_layout;
8162
8163 switch (mddev->layout) {
8164 case ALGORITHM_LEFT_ASYMMETRIC_6:
8165 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8166 break;
8167 case ALGORITHM_RIGHT_ASYMMETRIC_6:
8168 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8169 break;
8170 case ALGORITHM_LEFT_SYMMETRIC_6:
8171 new_layout = ALGORITHM_LEFT_SYMMETRIC;
8172 break;
8173 case ALGORITHM_RIGHT_SYMMETRIC_6:
8174 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8175 break;
8176 case ALGORITHM_PARITY_0_6:
8177 new_layout = ALGORITHM_PARITY_0;
8178 break;
8179 case ALGORITHM_PARITY_N:
8180 new_layout = ALGORITHM_PARITY_N;
8181 break;
8182 default:
8183 return ERR_PTR(-EINVAL);
8184 }
8185 mddev->new_level = 5;
8186 mddev->new_layout = new_layout;
8187 mddev->delta_disks = -1;
8188 mddev->raid_disks -= 1;
8189 return setup_conf(mddev);
8190}
8191
8192static int raid5_check_reshape(struct mddev *mddev)
8193{
8194 /* For a 2-drive array, the layout and chunk size can be changed
8195 * immediately as not restriping is needed.
8196 * For larger arrays we record the new value - after validation
8197 * to be used by a reshape pass.
8198 */
8199 struct r5conf *conf = mddev->private;
8200 int new_chunk = mddev->new_chunk_sectors;
8201
8202 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8203 return -EINVAL;
8204 if (new_chunk > 0) {
8205 if (!is_power_of_2(new_chunk))
8206 return -EINVAL;
8207 if (new_chunk < (PAGE_SIZE>>9))
8208 return -EINVAL;
8209 if (mddev->array_sectors & (new_chunk-1))
8210 /* not factor of array size */
8211 return -EINVAL;
8212 }
8213
8214 /* They look valid */
8215
8216 if (mddev->raid_disks == 2) {
8217 /* can make the change immediately */
8218 if (mddev->new_layout >= 0) {
8219 conf->algorithm = mddev->new_layout;
8220 mddev->layout = mddev->new_layout;
8221 }
8222 if (new_chunk > 0) {
8223 conf->chunk_sectors = new_chunk ;
8224 mddev->chunk_sectors = new_chunk;
8225 }
8226 set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8227 md_wakeup_thread(mddev->thread);
8228 }
8229 return check_reshape(mddev);
8230}
8231
8232static int raid6_check_reshape(struct mddev *mddev)
8233{
8234 int new_chunk = mddev->new_chunk_sectors;
8235
8236 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8237 return -EINVAL;
8238 if (new_chunk > 0) {
8239 if (!is_power_of_2(new_chunk))
8240 return -EINVAL;
8241 if (new_chunk < (PAGE_SIZE >> 9))
8242 return -EINVAL;
8243 if (mddev->array_sectors & (new_chunk-1))
8244 /* not factor of array size */
8245 return -EINVAL;
8246 }
8247
8248 /* They look valid */
8249 return check_reshape(mddev);
8250}
8251
8252static void *raid5_takeover(struct mddev *mddev)
8253{
8254 /* raid5 can take over:
8255 * raid0 - if there is only one strip zone - make it a raid4 layout
8256 * raid1 - if there are two drives. We need to know the chunk size
8257 * raid4 - trivial - just use a raid4 layout.
8258 * raid6 - Providing it is a *_6 layout
8259 */
8260 if (mddev->level == 0)
8261 return raid45_takeover_raid0(mddev, 5);
8262 if (mddev->level == 1)
8263 return raid5_takeover_raid1(mddev);
8264 if (mddev->level == 4) {
8265 mddev->new_layout = ALGORITHM_PARITY_N;
8266 mddev->new_level = 5;
8267 return setup_conf(mddev);
8268 }
8269 if (mddev->level == 6)
8270 return raid5_takeover_raid6(mddev);
8271
8272 return ERR_PTR(-EINVAL);
8273}
8274
8275static void *raid4_takeover(struct mddev *mddev)
8276{
8277 /* raid4 can take over:
8278 * raid0 - if there is only one strip zone
8279 * raid5 - if layout is right
8280 */
8281 if (mddev->level == 0)
8282 return raid45_takeover_raid0(mddev, 4);
8283 if (mddev->level == 5 &&
8284 mddev->layout == ALGORITHM_PARITY_N) {
8285 mddev->new_layout = 0;
8286 mddev->new_level = 4;
8287 return setup_conf(mddev);
8288 }
8289 return ERR_PTR(-EINVAL);
8290}
8291
8292static struct md_personality raid5_personality;
8293
8294static void *raid6_takeover(struct mddev *mddev)
8295{
8296 /* Currently can only take over a raid5. We map the
8297 * personality to an equivalent raid6 personality
8298 * with the Q block at the end.
8299 */
8300 int new_layout;
8301
8302 if (mddev->pers != &raid5_personality)
8303 return ERR_PTR(-EINVAL);
8304 if (mddev->degraded > 1)
8305 return ERR_PTR(-EINVAL);
8306 if (mddev->raid_disks > 253)
8307 return ERR_PTR(-EINVAL);
8308 if (mddev->raid_disks < 3)
8309 return ERR_PTR(-EINVAL);
8310
8311 switch (mddev->layout) {
8312 case ALGORITHM_LEFT_ASYMMETRIC:
8313 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8314 break;
8315 case ALGORITHM_RIGHT_ASYMMETRIC:
8316 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8317 break;
8318 case ALGORITHM_LEFT_SYMMETRIC:
8319 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8320 break;
8321 case ALGORITHM_RIGHT_SYMMETRIC:
8322 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8323 break;
8324 case ALGORITHM_PARITY_0:
8325 new_layout = ALGORITHM_PARITY_0_6;
8326 break;
8327 case ALGORITHM_PARITY_N:
8328 new_layout = ALGORITHM_PARITY_N;
8329 break;
8330 default:
8331 return ERR_PTR(-EINVAL);
8332 }
8333 mddev->new_level = 6;
8334 mddev->new_layout = new_layout;
8335 mddev->delta_disks = 1;
8336 mddev->raid_disks += 1;
8337 return setup_conf(mddev);
8338}
8339
8340static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8341{
8342 struct r5conf *conf;
8343 int err;
8344
8345 err = mddev_lock(mddev);
8346 if (err)
8347 return err;
8348 conf = mddev->private;
8349 if (!conf) {
8350 mddev_unlock(mddev);
8351 return -ENODEV;
8352 }
8353
8354 if (strncmp(buf, "ppl", 3) == 0) {
8355 /* ppl only works with RAID 5 */
8356 if (!raid5_has_ppl(conf) && conf->level == 5) {
8357 err = log_init(conf, NULL, true);
8358 if (!err) {
8359 err = resize_stripes(conf, conf->pool_size);
8360 if (err)
8361 log_exit(conf);
8362 }
8363 } else
8364 err = -EINVAL;
8365 } else if (strncmp(buf, "resync", 6) == 0) {
8366 if (raid5_has_ppl(conf)) {
8367 mddev_suspend(mddev);
8368 log_exit(conf);
8369 mddev_resume(mddev);
8370 err = resize_stripes(conf, conf->pool_size);
8371 } else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8372 r5l_log_disk_error(conf)) {
8373 bool journal_dev_exists = false;
8374 struct md_rdev *rdev;
8375
8376 rdev_for_each(rdev, mddev)
8377 if (test_bit(Journal, &rdev->flags)) {
8378 journal_dev_exists = true;
8379 break;
8380 }
8381
8382 if (!journal_dev_exists) {
8383 mddev_suspend(mddev);
8384 clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8385 mddev_resume(mddev);
8386 } else /* need remove journal device first */
8387 err = -EBUSY;
8388 } else
8389 err = -EINVAL;
8390 } else {
8391 err = -EINVAL;
8392 }
8393
8394 if (!err)
8395 md_update_sb(mddev, 1);
8396
8397 mddev_unlock(mddev);
8398
8399 return err;
8400}
8401
8402static int raid5_start(struct mddev *mddev)
8403{
8404 struct r5conf *conf = mddev->private;
8405
8406 return r5l_start(conf->log);
8407}
8408
8409static struct md_personality raid6_personality =
8410{
8411 .name = "raid6",
8412 .level = 6,
8413 .owner = THIS_MODULE,
8414 .make_request = raid5_make_request,
8415 .run = raid5_run,
8416 .start = raid5_start,
8417 .free = raid5_free,
8418 .status = raid5_status,
8419 .error_handler = raid5_error,
8420 .hot_add_disk = raid5_add_disk,
8421 .hot_remove_disk= raid5_remove_disk,
8422 .spare_active = raid5_spare_active,
8423 .sync_request = raid5_sync_request,
8424 .resize = raid5_resize,
8425 .size = raid5_size,
8426 .check_reshape = raid6_check_reshape,
8427 .start_reshape = raid5_start_reshape,
8428 .finish_reshape = raid5_finish_reshape,
8429 .quiesce = raid5_quiesce,
8430 .takeover = raid6_takeover,
8431 .congested = raid5_congested,
8432 .change_consistency_policy = raid5_change_consistency_policy,
8433};
8434static struct md_personality raid5_personality =
8435{
8436 .name = "raid5",
8437 .level = 5,
8438 .owner = THIS_MODULE,
8439 .make_request = raid5_make_request,
8440 .run = raid5_run,
8441 .start = raid5_start,
8442 .free = raid5_free,
8443 .status = raid5_status,
8444 .error_handler = raid5_error,
8445 .hot_add_disk = raid5_add_disk,
8446 .hot_remove_disk= raid5_remove_disk,
8447 .spare_active = raid5_spare_active,
8448 .sync_request = raid5_sync_request,
8449 .resize = raid5_resize,
8450 .size = raid5_size,
8451 .check_reshape = raid5_check_reshape,
8452 .start_reshape = raid5_start_reshape,
8453 .finish_reshape = raid5_finish_reshape,
8454 .quiesce = raid5_quiesce,
8455 .takeover = raid5_takeover,
8456 .congested = raid5_congested,
8457 .change_consistency_policy = raid5_change_consistency_policy,
8458};
8459
8460static struct md_personality raid4_personality =
8461{
8462 .name = "raid4",
8463 .level = 4,
8464 .owner = THIS_MODULE,
8465 .make_request = raid5_make_request,
8466 .run = raid5_run,
8467 .start = raid5_start,
8468 .free = raid5_free,
8469 .status = raid5_status,
8470 .error_handler = raid5_error,
8471 .hot_add_disk = raid5_add_disk,
8472 .hot_remove_disk= raid5_remove_disk,
8473 .spare_active = raid5_spare_active,
8474 .sync_request = raid5_sync_request,
8475 .resize = raid5_resize,
8476 .size = raid5_size,
8477 .check_reshape = raid5_check_reshape,
8478 .start_reshape = raid5_start_reshape,
8479 .finish_reshape = raid5_finish_reshape,
8480 .quiesce = raid5_quiesce,
8481 .takeover = raid4_takeover,
8482 .congested = raid5_congested,
8483 .change_consistency_policy = raid5_change_consistency_policy,
8484};
8485
8486static int __init raid5_init(void)
8487{
8488 int ret;
8489
8490 raid5_wq = alloc_workqueue("raid5wq",
8491 WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8492 if (!raid5_wq)
8493 return -ENOMEM;
8494
8495 ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8496 "md/raid5:prepare",
8497 raid456_cpu_up_prepare,
8498 raid456_cpu_dead);
8499 if (ret) {
8500 destroy_workqueue(raid5_wq);
8501 return ret;
8502 }
8503 register_md_personality(&raid6_personality);
8504 register_md_personality(&raid5_personality);
8505 register_md_personality(&raid4_personality);
8506 return 0;
8507}
8508
8509static void raid5_exit(void)
8510{
8511 unregister_md_personality(&raid6_personality);
8512 unregister_md_personality(&raid5_personality);
8513 unregister_md_personality(&raid4_personality);
8514 cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8515 destroy_workqueue(raid5_wq);
8516}
8517
8518module_init(raid5_init);
8519module_exit(raid5_exit);
8520MODULE_LICENSE("GPL");
8521MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8522MODULE_ALIAS("md-personality-4"); /* RAID5 */
8523MODULE_ALIAS("md-raid5");
8524MODULE_ALIAS("md-raid4");
8525MODULE_ALIAS("md-level-5");
8526MODULE_ALIAS("md-level-4");
8527MODULE_ALIAS("md-personality-8"); /* RAID6 */
8528MODULE_ALIAS("md-raid6");
8529MODULE_ALIAS("md-level-6");
8530
8531/* This used to be two separate modules, they were: */
8532MODULE_ALIAS("raid5");
8533MODULE_ALIAS("raid6");