?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge tag 'for-3.8-merge' of git://git.kernel.org/pub/scm/linux/kernel/git/jaegeuk...
[linux-2.6/libata-dev.git] / drivers / md / raid5.c
blob19d77a02663972c3daa9d74295c820d2bb85ba37
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 <trace/events/block.h>
57
58 #include "md.h"
59 #include "raid5.h"
60 #include "raid0.h"
61 #include "bitmap.h"
62
63 /*
64 * Stripe cache
65 */
66
67 #define NR_STRIPES 256
68 #define STRIPE_SIZE PAGE_SIZE
69 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
70 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
71 #define IO_THRESHOLD 1
72 #define BYPASS_THRESHOLD 1
73 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
74 #define HASH_MASK (NR_HASH - 1)
75
76 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
77 {
78 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
79 return &conf->stripe_hashtbl[hash];
80 }
81
82 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
83 * order without overlap. There may be several bio's per stripe+device, and
84 * a bio could span several devices.
85 * When walking this list for a particular stripe+device, we must never proceed
86 * beyond a bio that extends past this device, as the next bio might no longer
87 * be valid.
88 * This function is used to determine the 'next' bio in the list, given the sector
89 * of the current stripe+device
90 */
91 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
92 {
93 int sectors = bio->bi_size >> 9;
94 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
95 return bio->bi_next;
96 else
97 return NULL;
98 }
99
100 /*
101 * We maintain a biased count of active stripes in the bottom 16 bits of
102 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
103 */
104 static inline int raid5_bi_processed_stripes(struct bio *bio)
105 {
106 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
107 return (atomic_read(segments) >> 16) & 0xffff;
108 }
109
110 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
111 {
112 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
113 return atomic_sub_return(1, segments) & 0xffff;
114 }
115
116 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
117 {
118 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
119 atomic_inc(segments);
120 }
121
122 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
123 unsigned int cnt)
124 {
125 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
126 int old, new;
127
128 do {
129 old = atomic_read(segments);
130 new = (old & 0xffff) | (cnt << 16);
131 } while (atomic_cmpxchg(segments, old, new) != old);
132 }
133
134 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
135 {
136 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
137 atomic_set(segments, cnt);
138 }
139
140 /* Find first data disk in a raid6 stripe */
141 static inline int raid6_d0(struct stripe_head *sh)
142 {
143 if (sh->ddf_layout)
144 /* ddf always start from first device */
145 return 0;
146 /* md starts just after Q block */
147 if (sh->qd_idx == sh->disks - 1)
148 return 0;
149 else
150 return sh->qd_idx + 1;
151 }
152 static inline int raid6_next_disk(int disk, int raid_disks)
153 {
154 disk++;
155 return (disk < raid_disks) ? disk : 0;
156 }
157
158 /* When walking through the disks in a raid5, starting at raid6_d0,
159 * We need to map each disk to a 'slot', where the data disks are slot
160 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
161 * is raid_disks-1. This help does that mapping.
162 */
163 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
164 int *count, int syndrome_disks)
165 {
166 int slot = *count;
167
168 if (sh->ddf_layout)
169 (*count)++;
170 if (idx == sh->pd_idx)
171 return syndrome_disks;
172 if (idx == sh->qd_idx)
173 return syndrome_disks + 1;
174 if (!sh->ddf_layout)
175 (*count)++;
176 return slot;
177 }
178
179 static void return_io(struct bio *return_bi)
180 {
181 struct bio *bi = return_bi;
182 while (bi) {
183
184 return_bi = bi->bi_next;
185 bi->bi_next = NULL;
186 bi->bi_size = 0;
187 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
188 bi, 0);
189 bio_endio(bi, 0);
190 bi = return_bi;
191 }
192 }
193
194 static void print_raid5_conf (struct r5conf *conf);
195
196 static int stripe_operations_active(struct stripe_head *sh)
197 {
198 return sh->check_state || sh->reconstruct_state ||
199 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
200 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
201 }
202
203 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
204 {
205 BUG_ON(!list_empty(&sh->lru));
206 BUG_ON(atomic_read(&conf->active_stripes)==0);
207 if (test_bit(STRIPE_HANDLE, &sh->state)) {
208 if (test_bit(STRIPE_DELAYED, &sh->state) &&
209 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
210 list_add_tail(&sh->lru, &conf->delayed_list);
211 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
212 sh->bm_seq - conf->seq_write > 0)
213 list_add_tail(&sh->lru, &conf->bitmap_list);
214 else {
215 clear_bit(STRIPE_DELAYED, &sh->state);
216 clear_bit(STRIPE_BIT_DELAY, &sh->state);
217 list_add_tail(&sh->lru, &conf->handle_list);
218 }
219 md_wakeup_thread(conf->mddev->thread);
220 } else {
221 BUG_ON(stripe_operations_active(sh));
222 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
223 if (atomic_dec_return(&conf->preread_active_stripes)
224 < IO_THRESHOLD)
225 md_wakeup_thread(conf->mddev->thread);
226 atomic_dec(&conf->active_stripes);
227 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
228 list_add_tail(&sh->lru, &conf->inactive_list);
229 wake_up(&conf->wait_for_stripe);
230 if (conf->retry_read_aligned)
231 md_wakeup_thread(conf->mddev->thread);
232 }
233 }
234 }
235
236 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
237 {
238 if (atomic_dec_and_test(&sh->count))
239 do_release_stripe(conf, sh);
240 }
241
242 static void release_stripe(struct stripe_head *sh)
243 {
244 struct r5conf *conf = sh->raid_conf;
245 unsigned long flags;
246
247 local_irq_save(flags);
248 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
249 do_release_stripe(conf, sh);
250 spin_unlock(&conf->device_lock);
251 }
252 local_irq_restore(flags);
253 }
254
255 static inline void remove_hash(struct stripe_head *sh)
256 {
257 pr_debug("remove_hash(), stripe %llu\n",
258 (unsigned long long)sh->sector);
259
260 hlist_del_init(&sh->hash);
261 }
262
263 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
264 {
265 struct hlist_head *hp = stripe_hash(conf, sh->sector);
266
267 pr_debug("insert_hash(), stripe %llu\n",
268 (unsigned long long)sh->sector);
269
270 hlist_add_head(&sh->hash, hp);
271 }
272
273
274 /* find an idle stripe, make sure it is unhashed, and return it. */
275 static struct stripe_head *get_free_stripe(struct r5conf *conf)
276 {
277 struct stripe_head *sh = NULL;
278 struct list_head *first;
279
280 if (list_empty(&conf->inactive_list))
281 goto out;
282 first = conf->inactive_list.next;
283 sh = list_entry(first, struct stripe_head, lru);
284 list_del_init(first);
285 remove_hash(sh);
286 atomic_inc(&conf->active_stripes);
287 out:
288 return sh;
289 }
290
291 static void shrink_buffers(struct stripe_head *sh)
292 {
293 struct page *p;
294 int i;
295 int num = sh->raid_conf->pool_size;
296
297 for (i = 0; i < num ; i++) {
298 p = sh->dev[i].page;
299 if (!p)
300 continue;
301 sh->dev[i].page = NULL;
302 put_page(p);
303 }
304 }
305
306 static int grow_buffers(struct stripe_head *sh)
307 {
308 int i;
309 int num = sh->raid_conf->pool_size;
310
311 for (i = 0; i < num; i++) {
312 struct page *page;
313
314 if (!(page = alloc_page(GFP_KERNEL))) {
315 return 1;
316 }
317 sh->dev[i].page = page;
318 }
319 return 0;
320 }
321
322 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
323 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
324 struct stripe_head *sh);
325
326 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
327 {
328 struct r5conf *conf = sh->raid_conf;
329 int i;
330
331 BUG_ON(atomic_read(&sh->count) != 0);
332 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
333 BUG_ON(stripe_operations_active(sh));
334
335 pr_debug("init_stripe called, stripe %llu\n",
336 (unsigned long long)sh->sector);
337
338 remove_hash(sh);
339
340 sh->generation = conf->generation - previous;
341 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
342 sh->sector = sector;
343 stripe_set_idx(sector, conf, previous, sh);
344 sh->state = 0;
345
346
347 for (i = sh->disks; i--; ) {
348 struct r5dev *dev = &sh->dev[i];
349
350 if (dev->toread || dev->read || dev->towrite || dev->written ||
351 test_bit(R5_LOCKED, &dev->flags)) {
352 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
353 (unsigned long long)sh->sector, i, dev->toread,
354 dev->read, dev->towrite, dev->written,
355 test_bit(R5_LOCKED, &dev->flags));
356 WARN_ON(1);
357 }
358 dev->flags = 0;
359 raid5_build_block(sh, i, previous);
360 }
361 insert_hash(conf, sh);
362 }
363
364 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
365 short generation)
366 {
367 struct stripe_head *sh;
368 struct hlist_node *hn;
369
370 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
371 hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
372 if (sh->sector == sector && sh->generation == generation)
373 return sh;
374 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
375 return NULL;
376 }
377
378 /*
379 * Need to check if array has failed when deciding whether to:
380 * - start an array
381 * - remove non-faulty devices
382 * - add a spare
383 * - allow a reshape
384 * This determination is simple when no reshape is happening.
385 * However if there is a reshape, we need to carefully check
386 * both the before and after sections.
387 * This is because some failed devices may only affect one
388 * of the two sections, and some non-in_sync devices may
389 * be insync in the section most affected by failed devices.
390 */
391 static int calc_degraded(struct r5conf *conf)
392 {
393 int degraded, degraded2;
394 int i;
395
396 rcu_read_lock();
397 degraded = 0;
398 for (i = 0; i < conf->previous_raid_disks; i++) {
399 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
400 if (rdev && test_bit(Faulty, &rdev->flags))
401 rdev = rcu_dereference(conf->disks[i].replacement);
402 if (!rdev || test_bit(Faulty, &rdev->flags))
403 degraded++;
404 else if (test_bit(In_sync, &rdev->flags))
405 ;
406 else
407 /* not in-sync or faulty.
408 * If the reshape increases the number of devices,
409 * this is being recovered by the reshape, so
410 * this 'previous' section is not in_sync.
411 * If the number of devices is being reduced however,
412 * the device can only be part of the array if
413 * we are reverting a reshape, so this section will
414 * be in-sync.
415 */
416 if (conf->raid_disks >= conf->previous_raid_disks)
417 degraded++;
418 }
419 rcu_read_unlock();
420 if (conf->raid_disks == conf->previous_raid_disks)
421 return degraded;
422 rcu_read_lock();
423 degraded2 = 0;
424 for (i = 0; i < conf->raid_disks; i++) {
425 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
426 if (rdev && test_bit(Faulty, &rdev->flags))
427 rdev = rcu_dereference(conf->disks[i].replacement);
428 if (!rdev || test_bit(Faulty, &rdev->flags))
429 degraded2++;
430 else if (test_bit(In_sync, &rdev->flags))
431 ;
432 else
433 /* not in-sync or faulty.
434 * If reshape increases the number of devices, this
435 * section has already been recovered, else it
436 * almost certainly hasn't.
437 */
438 if (conf->raid_disks <= conf->previous_raid_disks)
439 degraded2++;
440 }
441 rcu_read_unlock();
442 if (degraded2 > degraded)
443 return degraded2;
444 return degraded;
445 }
446
447 static int has_failed(struct r5conf *conf)
448 {
449 int degraded;
450
451 if (conf->mddev->reshape_position == MaxSector)
452 return conf->mddev->degraded > conf->max_degraded;
453
454 degraded = calc_degraded(conf);
455 if (degraded > conf->max_degraded)
456 return 1;
457 return 0;
458 }
459
460 static struct stripe_head *
461 get_active_stripe(struct r5conf *conf, sector_t sector,
462 int previous, int noblock, int noquiesce)
463 {
464 struct stripe_head *sh;
465
466 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
467
468 spin_lock_irq(&conf->device_lock);
469
470 do {
471 wait_event_lock_irq(conf->wait_for_stripe,
472 conf->quiesce == 0 || noquiesce,
473 conf->device_lock);
474 sh = __find_stripe(conf, sector, conf->generation - previous);
475 if (!sh) {
476 if (!conf->inactive_blocked)
477 sh = get_free_stripe(conf);
478 if (noblock && sh == NULL)
479 break;
480 if (!sh) {
481 conf->inactive_blocked = 1;
482 wait_event_lock_irq(conf->wait_for_stripe,
483 !list_empty(&conf->inactive_list) &&
484 (atomic_read(&conf->active_stripes)
485 < (conf->max_nr_stripes *3/4)
486 || !conf->inactive_blocked),
487 conf->device_lock);
488 conf->inactive_blocked = 0;
489 } else
490 init_stripe(sh, sector, previous);
491 } else {
492 if (atomic_read(&sh->count)) {
493 BUG_ON(!list_empty(&sh->lru)
494 && !test_bit(STRIPE_EXPANDING, &sh->state)
495 && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state));
496 } else {
497 if (!test_bit(STRIPE_HANDLE, &sh->state))
498 atomic_inc(&conf->active_stripes);
499 if (list_empty(&sh->lru) &&
500 !test_bit(STRIPE_EXPANDING, &sh->state))
501 BUG();
502 list_del_init(&sh->lru);
503 }
504 }
505 } while (sh == NULL);
506
507 if (sh)
508 atomic_inc(&sh->count);
509
510 spin_unlock_irq(&conf->device_lock);
511 return sh;
512 }
513
514 /* Determine if 'data_offset' or 'new_data_offset' should be used
515 * in this stripe_head.
516 */
517 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
518 {
519 sector_t progress = conf->reshape_progress;
520 /* Need a memory barrier to make sure we see the value
521 * of conf->generation, or ->data_offset that was set before
522 * reshape_progress was updated.
523 */
524 smp_rmb();
525 if (progress == MaxSector)
526 return 0;
527 if (sh->generation == conf->generation - 1)
528 return 0;
529 /* We are in a reshape, and this is a new-generation stripe,
530 * so use new_data_offset.
531 */
532 return 1;
533 }
534
535 static void
536 raid5_end_read_request(struct bio *bi, int error);
537 static void
538 raid5_end_write_request(struct bio *bi, int error);
539
540 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
541 {
542 struct r5conf *conf = sh->raid_conf;
543 int i, disks = sh->disks;
544
545 might_sleep();
546
547 for (i = disks; i--; ) {
548 int rw;
549 int replace_only = 0;
550 struct bio *bi, *rbi;
551 struct md_rdev *rdev, *rrdev = NULL;
552 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
553 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
554 rw = WRITE_FUA;
555 else
556 rw = WRITE;
557 if (test_bit(R5_Discard, &sh->dev[i].flags))
558 rw |= REQ_DISCARD;
559 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
560 rw = READ;
561 else if (test_and_clear_bit(R5_WantReplace,
562 &sh->dev[i].flags)) {
563 rw = WRITE;
564 replace_only = 1;
565 } else
566 continue;
567 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
568 rw |= REQ_SYNC;
569
570 bi = &sh->dev[i].req;
571 rbi = &sh->dev[i].rreq; /* For writing to replacement */
572
573 bi->bi_rw = rw;
574 rbi->bi_rw = rw;
575 if (rw & WRITE) {
576 bi->bi_end_io = raid5_end_write_request;
577 rbi->bi_end_io = raid5_end_write_request;
578 } else
579 bi->bi_end_io = raid5_end_read_request;
580
581 rcu_read_lock();
582 rrdev = rcu_dereference(conf->disks[i].replacement);
583 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
584 rdev = rcu_dereference(conf->disks[i].rdev);
585 if (!rdev) {
586 rdev = rrdev;
587 rrdev = NULL;
588 }
589 if (rw & WRITE) {
590 if (replace_only)
591 rdev = NULL;
592 if (rdev == rrdev)
593 /* We raced and saw duplicates */
594 rrdev = NULL;
595 } else {
596 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
597 rdev = rrdev;
598 rrdev = NULL;
599 }
600
601 if (rdev && test_bit(Faulty, &rdev->flags))
602 rdev = NULL;
603 if (rdev)
604 atomic_inc(&rdev->nr_pending);
605 if (rrdev && test_bit(Faulty, &rrdev->flags))
606 rrdev = NULL;
607 if (rrdev)
608 atomic_inc(&rrdev->nr_pending);
609 rcu_read_unlock();
610
611 /* We have already checked bad blocks for reads. Now
612 * need to check for writes. We never accept write errors
613 * on the replacement, so we don't to check rrdev.
614 */
615 while ((rw & WRITE) && rdev &&
616 test_bit(WriteErrorSeen, &rdev->flags)) {
617 sector_t first_bad;
618 int bad_sectors;
619 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
620 &first_bad, &bad_sectors);
621 if (!bad)
622 break;
623
624 if (bad < 0) {
625 set_bit(BlockedBadBlocks, &rdev->flags);
626 if (!conf->mddev->external &&
627 conf->mddev->flags) {
628 /* It is very unlikely, but we might
629 * still need to write out the
630 * bad block log - better give it
631 * a chance*/
632 md_check_recovery(conf->mddev);
633 }
634 /*
635 * Because md_wait_for_blocked_rdev
636 * will dec nr_pending, we must
637 * increment it first.
638 */
639 atomic_inc(&rdev->nr_pending);
640 md_wait_for_blocked_rdev(rdev, conf->mddev);
641 } else {
642 /* Acknowledged bad block - skip the write */
643 rdev_dec_pending(rdev, conf->mddev);
644 rdev = NULL;
645 }
646 }
647
648 if (rdev) {
649 if (s->syncing || s->expanding || s->expanded
650 || s->replacing)
651 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
652
653 set_bit(STRIPE_IO_STARTED, &sh->state);
654
655 bi->bi_bdev = rdev->bdev;
656 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
657 __func__, (unsigned long long)sh->sector,
658 bi->bi_rw, i);
659 atomic_inc(&sh->count);
660 if (use_new_offset(conf, sh))
661 bi->bi_sector = (sh->sector
662 + rdev->new_data_offset);
663 else
664 bi->bi_sector = (sh->sector
665 + rdev->data_offset);
666 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
667 bi->bi_rw |= REQ_FLUSH;
668
669 bi->bi_flags = 1 << BIO_UPTODATE;
670 bi->bi_idx = 0;
671 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
672 bi->bi_io_vec[0].bv_offset = 0;
673 bi->bi_size = STRIPE_SIZE;
674 bi->bi_next = NULL;
675 if (rrdev)
676 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
677 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
678 bi, disk_devt(conf->mddev->gendisk),
679 sh->dev[i].sector);
680 generic_make_request(bi);
681 }
682 if (rrdev) {
683 if (s->syncing || s->expanding || s->expanded
684 || s->replacing)
685 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
686
687 set_bit(STRIPE_IO_STARTED, &sh->state);
688
689 rbi->bi_bdev = rrdev->bdev;
690 pr_debug("%s: for %llu schedule op %ld on "
691 "replacement disc %d\n",
692 __func__, (unsigned long long)sh->sector,
693 rbi->bi_rw, i);
694 atomic_inc(&sh->count);
695 if (use_new_offset(conf, sh))
696 rbi->bi_sector = (sh->sector
697 + rrdev->new_data_offset);
698 else
699 rbi->bi_sector = (sh->sector
700 + rrdev->data_offset);
701 rbi->bi_flags = 1 << BIO_UPTODATE;
702 rbi->bi_idx = 0;
703 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
704 rbi->bi_io_vec[0].bv_offset = 0;
705 rbi->bi_size = STRIPE_SIZE;
706 rbi->bi_next = NULL;
707 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
708 rbi, disk_devt(conf->mddev->gendisk),
709 sh->dev[i].sector);
710 generic_make_request(rbi);
711 }
712 if (!rdev && !rrdev) {
713 if (rw & WRITE)
714 set_bit(STRIPE_DEGRADED, &sh->state);
715 pr_debug("skip op %ld on disc %d for sector %llu\n",
716 bi->bi_rw, i, (unsigned long long)sh->sector);
717 clear_bit(R5_LOCKED, &sh->dev[i].flags);
718 set_bit(STRIPE_HANDLE, &sh->state);
719 }
720 }
721 }
722
723 static struct dma_async_tx_descriptor *
724 async_copy_data(int frombio, struct bio *bio, struct page *page,
725 sector_t sector, struct dma_async_tx_descriptor *tx)
726 {
727 struct bio_vec *bvl;
728 struct page *bio_page;
729 int i;
730 int page_offset;
731 struct async_submit_ctl submit;
732 enum async_tx_flags flags = 0;
733
734 if (bio->bi_sector >= sector)
735 page_offset = (signed)(bio->bi_sector - sector) * 512;
736 else
737 page_offset = (signed)(sector - bio->bi_sector) * -512;
738
739 if (frombio)
740 flags |= ASYNC_TX_FENCE;
741 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
742
743 bio_for_each_segment(bvl, bio, i) {
744 int len = bvl->bv_len;
745 int clen;
746 int b_offset = 0;
747
748 if (page_offset < 0) {
749 b_offset = -page_offset;
750 page_offset += b_offset;
751 len -= b_offset;
752 }
753
754 if (len > 0 && page_offset + len > STRIPE_SIZE)
755 clen = STRIPE_SIZE - page_offset;
756 else
757 clen = len;
758
759 if (clen > 0) {
760 b_offset += bvl->bv_offset;
761 bio_page = bvl->bv_page;
762 if (frombio)
763 tx = async_memcpy(page, bio_page, page_offset,
764 b_offset, clen, &submit);
765 else
766 tx = async_memcpy(bio_page, page, b_offset,
767 page_offset, clen, &submit);
768 }
769 /* chain the operations */
770 submit.depend_tx = tx;
771
772 if (clen < len) /* hit end of page */
773 break;
774 page_offset += len;
775 }
776
777 return tx;
778 }
779
780 static void ops_complete_biofill(void *stripe_head_ref)
781 {
782 struct stripe_head *sh = stripe_head_ref;
783 struct bio *return_bi = NULL;
784 int i;
785
786 pr_debug("%s: stripe %llu\n", __func__,
787 (unsigned long long)sh->sector);
788
789 /* clear completed biofills */
790 for (i = sh->disks; i--; ) {
791 struct r5dev *dev = &sh->dev[i];
792
793 /* acknowledge completion of a biofill operation */
794 /* and check if we need to reply to a read request,
795 * new R5_Wantfill requests are held off until
796 * !STRIPE_BIOFILL_RUN
797 */
798 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
799 struct bio *rbi, *rbi2;
800
801 BUG_ON(!dev->read);
802 rbi = dev->read;
803 dev->read = NULL;
804 while (rbi && rbi->bi_sector <
805 dev->sector + STRIPE_SECTORS) {
806 rbi2 = r5_next_bio(rbi, dev->sector);
807 if (!raid5_dec_bi_active_stripes(rbi)) {
808 rbi->bi_next = return_bi;
809 return_bi = rbi;
810 }
811 rbi = rbi2;
812 }
813 }
814 }
815 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
816
817 return_io(return_bi);
818
819 set_bit(STRIPE_HANDLE, &sh->state);
820 release_stripe(sh);
821 }
822
823 static void ops_run_biofill(struct stripe_head *sh)
824 {
825 struct dma_async_tx_descriptor *tx = NULL;
826 struct async_submit_ctl submit;
827 int i;
828
829 pr_debug("%s: stripe %llu\n", __func__,
830 (unsigned long long)sh->sector);
831
832 for (i = sh->disks; i--; ) {
833 struct r5dev *dev = &sh->dev[i];
834 if (test_bit(R5_Wantfill, &dev->flags)) {
835 struct bio *rbi;
836 spin_lock_irq(&sh->stripe_lock);
837 dev->read = rbi = dev->toread;
838 dev->toread = NULL;
839 spin_unlock_irq(&sh->stripe_lock);
840 while (rbi && rbi->bi_sector <
841 dev->sector + STRIPE_SECTORS) {
842 tx = async_copy_data(0, rbi, dev->page,
843 dev->sector, tx);
844 rbi = r5_next_bio(rbi, dev->sector);
845 }
846 }
847 }
848
849 atomic_inc(&sh->count);
850 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
851 async_trigger_callback(&submit);
852 }
853
854 static void mark_target_uptodate(struct stripe_head *sh, int target)
855 {
856 struct r5dev *tgt;
857
858 if (target < 0)
859 return;
860
861 tgt = &sh->dev[target];
862 set_bit(R5_UPTODATE, &tgt->flags);
863 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
864 clear_bit(R5_Wantcompute, &tgt->flags);
865 }
866
867 static void ops_complete_compute(void *stripe_head_ref)
868 {
869 struct stripe_head *sh = stripe_head_ref;
870
871 pr_debug("%s: stripe %llu\n", __func__,
872 (unsigned long long)sh->sector);
873
874 /* mark the computed target(s) as uptodate */
875 mark_target_uptodate(sh, sh->ops.target);
876 mark_target_uptodate(sh, sh->ops.target2);
877
878 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
879 if (sh->check_state == check_state_compute_run)
880 sh->check_state = check_state_compute_result;
881 set_bit(STRIPE_HANDLE, &sh->state);
882 release_stripe(sh);
883 }
884
885 /* return a pointer to the address conversion region of the scribble buffer */
886 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
887 struct raid5_percpu *percpu)
888 {
889 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
890 }
891
892 static struct dma_async_tx_descriptor *
893 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
894 {
895 int disks = sh->disks;
896 struct page **xor_srcs = percpu->scribble;
897 int target = sh->ops.target;
898 struct r5dev *tgt = &sh->dev[target];
899 struct page *xor_dest = tgt->page;
900 int count = 0;
901 struct dma_async_tx_descriptor *tx;
902 struct async_submit_ctl submit;
903 int i;
904
905 pr_debug("%s: stripe %llu block: %d\n",
906 __func__, (unsigned long long)sh->sector, target);
907 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
908
909 for (i = disks; i--; )
910 if (i != target)
911 xor_srcs[count++] = sh->dev[i].page;
912
913 atomic_inc(&sh->count);
914
915 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
916 ops_complete_compute, sh, to_addr_conv(sh, percpu));
917 if (unlikely(count == 1))
918 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
919 else
920 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
921
922 return tx;
923 }
924
925 /* set_syndrome_sources - populate source buffers for gen_syndrome
926 * @srcs - (struct page *) array of size sh->disks
927 * @sh - stripe_head to parse
928 *
929 * Populates srcs in proper layout order for the stripe and returns the
930 * 'count' of sources to be used in a call to async_gen_syndrome. The P
931 * destination buffer is recorded in srcs[count] and the Q destination
932 * is recorded in srcs[count+1]].
933 */
934 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
935 {
936 int disks = sh->disks;
937 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
938 int d0_idx = raid6_d0(sh);
939 int count;
940 int i;
941
942 for (i = 0; i < disks; i++)
943 srcs[i] = NULL;
944
945 count = 0;
946 i = d0_idx;
947 do {
948 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
949
950 srcs[slot] = sh->dev[i].page;
951 i = raid6_next_disk(i, disks);
952 } while (i != d0_idx);
953
954 return syndrome_disks;
955 }
956
957 static struct dma_async_tx_descriptor *
958 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
959 {
960 int disks = sh->disks;
961 struct page **blocks = percpu->scribble;
962 int target;
963 int qd_idx = sh->qd_idx;
964 struct dma_async_tx_descriptor *tx;
965 struct async_submit_ctl submit;
966 struct r5dev *tgt;
967 struct page *dest;
968 int i;
969 int count;
970
971 if (sh->ops.target < 0)
972 target = sh->ops.target2;
973 else if (sh->ops.target2 < 0)
974 target = sh->ops.target;
975 else
976 /* we should only have one valid target */
977 BUG();
978 BUG_ON(target < 0);
979 pr_debug("%s: stripe %llu block: %d\n",
980 __func__, (unsigned long long)sh->sector, target);
981
982 tgt = &sh->dev[target];
983 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
984 dest = tgt->page;
985
986 atomic_inc(&sh->count);
987
988 if (target == qd_idx) {
989 count = set_syndrome_sources(blocks, sh);
990 blocks[count] = NULL; /* regenerating p is not necessary */
991 BUG_ON(blocks[count+1] != dest); /* q should already be set */
992 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
993 ops_complete_compute, sh,
994 to_addr_conv(sh, percpu));
995 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
996 } else {
997 /* Compute any data- or p-drive using XOR */
998 count = 0;
999 for (i = disks; i-- ; ) {
1000 if (i == target || i == qd_idx)
1001 continue;
1002 blocks[count++] = sh->dev[i].page;
1003 }
1004
1005 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1006 NULL, ops_complete_compute, sh,
1007 to_addr_conv(sh, percpu));
1008 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1009 }
1010
1011 return tx;
1012 }
1013
1014 static struct dma_async_tx_descriptor *
1015 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1016 {
1017 int i, count, disks = sh->disks;
1018 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1019 int d0_idx = raid6_d0(sh);
1020 int faila = -1, failb = -1;
1021 int target = sh->ops.target;
1022 int target2 = sh->ops.target2;
1023 struct r5dev *tgt = &sh->dev[target];
1024 struct r5dev *tgt2 = &sh->dev[target2];
1025 struct dma_async_tx_descriptor *tx;
1026 struct page **blocks = percpu->scribble;
1027 struct async_submit_ctl submit;
1028
1029 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1030 __func__, (unsigned long long)sh->sector, target, target2);
1031 BUG_ON(target < 0 || target2 < 0);
1032 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1033 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1034
1035 /* we need to open-code set_syndrome_sources to handle the
1036 * slot number conversion for 'faila' and 'failb'
1037 */
1038 for (i = 0; i < disks ; i++)
1039 blocks[i] = NULL;
1040 count = 0;
1041 i = d0_idx;
1042 do {
1043 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1044
1045 blocks[slot] = sh->dev[i].page;
1046
1047 if (i == target)
1048 faila = slot;
1049 if (i == target2)
1050 failb = slot;
1051 i = raid6_next_disk(i, disks);
1052 } while (i != d0_idx);
1053
1054 BUG_ON(faila == failb);
1055 if (failb < faila)
1056 swap(faila, failb);
1057 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1058 __func__, (unsigned long long)sh->sector, faila, failb);
1059
1060 atomic_inc(&sh->count);
1061
1062 if (failb == syndrome_disks+1) {
1063 /* Q disk is one of the missing disks */
1064 if (faila == syndrome_disks) {
1065 /* Missing P+Q, just recompute */
1066 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1067 ops_complete_compute, sh,
1068 to_addr_conv(sh, percpu));
1069 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1070 STRIPE_SIZE, &submit);
1071 } else {
1072 struct page *dest;
1073 int data_target;
1074 int qd_idx = sh->qd_idx;
1075
1076 /* Missing D+Q: recompute D from P, then recompute Q */
1077 if (target == qd_idx)
1078 data_target = target2;
1079 else
1080 data_target = target;
1081
1082 count = 0;
1083 for (i = disks; i-- ; ) {
1084 if (i == data_target || i == qd_idx)
1085 continue;
1086 blocks[count++] = sh->dev[i].page;
1087 }
1088 dest = sh->dev[data_target].page;
1089 init_async_submit(&submit,
1090 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1091 NULL, NULL, NULL,
1092 to_addr_conv(sh, percpu));
1093 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1094 &submit);
1095
1096 count = set_syndrome_sources(blocks, sh);
1097 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1098 ops_complete_compute, sh,
1099 to_addr_conv(sh, percpu));
1100 return async_gen_syndrome(blocks, 0, count+2,
1101 STRIPE_SIZE, &submit);
1102 }
1103 } else {
1104 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1105 ops_complete_compute, sh,
1106 to_addr_conv(sh, percpu));
1107 if (failb == syndrome_disks) {
1108 /* We're missing D+P. */
1109 return async_raid6_datap_recov(syndrome_disks+2,
1110 STRIPE_SIZE, faila,
1111 blocks, &submit);
1112 } else {
1113 /* We're missing D+D. */
1114 return async_raid6_2data_recov(syndrome_disks+2,
1115 STRIPE_SIZE, faila, failb,
1116 blocks, &submit);
1117 }
1118 }
1119 }
1120
1121
1122 static void ops_complete_prexor(void *stripe_head_ref)
1123 {
1124 struct stripe_head *sh = stripe_head_ref;
1125
1126 pr_debug("%s: stripe %llu\n", __func__,
1127 (unsigned long long)sh->sector);
1128 }
1129
1130 static struct dma_async_tx_descriptor *
1131 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1132 struct dma_async_tx_descriptor *tx)
1133 {
1134 int disks = sh->disks;
1135 struct page **xor_srcs = percpu->scribble;
1136 int count = 0, pd_idx = sh->pd_idx, i;
1137 struct async_submit_ctl submit;
1138
1139 /* existing parity data subtracted */
1140 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1141
1142 pr_debug("%s: stripe %llu\n", __func__,
1143 (unsigned long long)sh->sector);
1144
1145 for (i = disks; i--; ) {
1146 struct r5dev *dev = &sh->dev[i];
1147 /* Only process blocks that are known to be uptodate */
1148 if (test_bit(R5_Wantdrain, &dev->flags))
1149 xor_srcs[count++] = dev->page;
1150 }
1151
1152 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1153 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1154 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1155
1156 return tx;
1157 }
1158
1159 static struct dma_async_tx_descriptor *
1160 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1161 {
1162 int disks = sh->disks;
1163 int i;
1164
1165 pr_debug("%s: stripe %llu\n", __func__,
1166 (unsigned long long)sh->sector);
1167
1168 for (i = disks; i--; ) {
1169 struct r5dev *dev = &sh->dev[i];
1170 struct bio *chosen;
1171
1172 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1173 struct bio *wbi;
1174
1175 spin_lock_irq(&sh->stripe_lock);
1176 chosen = dev->towrite;
1177 dev->towrite = NULL;
1178 BUG_ON(dev->written);
1179 wbi = dev->written = chosen;
1180 spin_unlock_irq(&sh->stripe_lock);
1181
1182 while (wbi && wbi->bi_sector <
1183 dev->sector + STRIPE_SECTORS) {
1184 if (wbi->bi_rw & REQ_FUA)
1185 set_bit(R5_WantFUA, &dev->flags);
1186 if (wbi->bi_rw & REQ_SYNC)
1187 set_bit(R5_SyncIO, &dev->flags);
1188 if (wbi->bi_rw & REQ_DISCARD)
1189 set_bit(R5_Discard, &dev->flags);
1190 else
1191 tx = async_copy_data(1, wbi, dev->page,
1192 dev->sector, tx);
1193 wbi = r5_next_bio(wbi, dev->sector);
1194 }
1195 }
1196 }
1197
1198 return tx;
1199 }
1200
1201 static void ops_complete_reconstruct(void *stripe_head_ref)
1202 {
1203 struct stripe_head *sh = stripe_head_ref;
1204 int disks = sh->disks;
1205 int pd_idx = sh->pd_idx;
1206 int qd_idx = sh->qd_idx;
1207 int i;
1208 bool fua = false, sync = false, discard = false;
1209
1210 pr_debug("%s: stripe %llu\n", __func__,
1211 (unsigned long long)sh->sector);
1212
1213 for (i = disks; i--; ) {
1214 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1215 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1216 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1217 }
1218
1219 for (i = disks; i--; ) {
1220 struct r5dev *dev = &sh->dev[i];
1221
1222 if (dev->written || i == pd_idx || i == qd_idx) {
1223 if (!discard)
1224 set_bit(R5_UPTODATE, &dev->flags);
1225 if (fua)
1226 set_bit(R5_WantFUA, &dev->flags);
1227 if (sync)
1228 set_bit(R5_SyncIO, &dev->flags);
1229 }
1230 }
1231
1232 if (sh->reconstruct_state == reconstruct_state_drain_run)
1233 sh->reconstruct_state = reconstruct_state_drain_result;
1234 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1235 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1236 else {
1237 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1238 sh->reconstruct_state = reconstruct_state_result;
1239 }
1240
1241 set_bit(STRIPE_HANDLE, &sh->state);
1242 release_stripe(sh);
1243 }
1244
1245 static void
1246 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1247 struct dma_async_tx_descriptor *tx)
1248 {
1249 int disks = sh->disks;
1250 struct page **xor_srcs = percpu->scribble;
1251 struct async_submit_ctl submit;
1252 int count = 0, pd_idx = sh->pd_idx, i;
1253 struct page *xor_dest;
1254 int prexor = 0;
1255 unsigned long flags;
1256
1257 pr_debug("%s: stripe %llu\n", __func__,
1258 (unsigned long long)sh->sector);
1259
1260 for (i = 0; i < sh->disks; i++) {
1261 if (pd_idx == i)
1262 continue;
1263 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1264 break;
1265 }
1266 if (i >= sh->disks) {
1267 atomic_inc(&sh->count);
1268 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1269 ops_complete_reconstruct(sh);
1270 return;
1271 }
1272 /* check if prexor is active which means only process blocks
1273 * that are part of a read-modify-write (written)
1274 */
1275 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1276 prexor = 1;
1277 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1278 for (i = disks; i--; ) {
1279 struct r5dev *dev = &sh->dev[i];
1280 if (dev->written)
1281 xor_srcs[count++] = dev->page;
1282 }
1283 } else {
1284 xor_dest = sh->dev[pd_idx].page;
1285 for (i = disks; i--; ) {
1286 struct r5dev *dev = &sh->dev[i];
1287 if (i != pd_idx)
1288 xor_srcs[count++] = dev->page;
1289 }
1290 }
1291
1292 /* 1/ if we prexor'd then the dest is reused as a source
1293 * 2/ if we did not prexor then we are redoing the parity
1294 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1295 * for the synchronous xor case
1296 */
1297 flags = ASYNC_TX_ACK |
1298 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1299
1300 atomic_inc(&sh->count);
1301
1302 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1303 to_addr_conv(sh, percpu));
1304 if (unlikely(count == 1))
1305 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1306 else
1307 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1308 }
1309
1310 static void
1311 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1312 struct dma_async_tx_descriptor *tx)
1313 {
1314 struct async_submit_ctl submit;
1315 struct page **blocks = percpu->scribble;
1316 int count, i;
1317
1318 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1319
1320 for (i = 0; i < sh->disks; i++) {
1321 if (sh->pd_idx == i || sh->qd_idx == i)
1322 continue;
1323 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1324 break;
1325 }
1326 if (i >= sh->disks) {
1327 atomic_inc(&sh->count);
1328 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1329 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1330 ops_complete_reconstruct(sh);
1331 return;
1332 }
1333
1334 count = set_syndrome_sources(blocks, sh);
1335
1336 atomic_inc(&sh->count);
1337
1338 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1339 sh, to_addr_conv(sh, percpu));
1340 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1341 }
1342
1343 static void ops_complete_check(void *stripe_head_ref)
1344 {
1345 struct stripe_head *sh = stripe_head_ref;
1346
1347 pr_debug("%s: stripe %llu\n", __func__,
1348 (unsigned long long)sh->sector);
1349
1350 sh->check_state = check_state_check_result;
1351 set_bit(STRIPE_HANDLE, &sh->state);
1352 release_stripe(sh);
1353 }
1354
1355 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1356 {
1357 int disks = sh->disks;
1358 int pd_idx = sh->pd_idx;
1359 int qd_idx = sh->qd_idx;
1360 struct page *xor_dest;
1361 struct page **xor_srcs = percpu->scribble;
1362 struct dma_async_tx_descriptor *tx;
1363 struct async_submit_ctl submit;
1364 int count;
1365 int i;
1366
1367 pr_debug("%s: stripe %llu\n", __func__,
1368 (unsigned long long)sh->sector);
1369
1370 count = 0;
1371 xor_dest = sh->dev[pd_idx].page;
1372 xor_srcs[count++] = xor_dest;
1373 for (i = disks; i--; ) {
1374 if (i == pd_idx || i == qd_idx)
1375 continue;
1376 xor_srcs[count++] = sh->dev[i].page;
1377 }
1378
1379 init_async_submit(&submit, 0, NULL, NULL, NULL,
1380 to_addr_conv(sh, percpu));
1381 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1382 &sh->ops.zero_sum_result, &submit);
1383
1384 atomic_inc(&sh->count);
1385 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1386 tx = async_trigger_callback(&submit);
1387 }
1388
1389 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1390 {
1391 struct page **srcs = percpu->scribble;
1392 struct async_submit_ctl submit;
1393 int count;
1394
1395 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1396 (unsigned long long)sh->sector, checkp);
1397
1398 count = set_syndrome_sources(srcs, sh);
1399 if (!checkp)
1400 srcs[count] = NULL;
1401
1402 atomic_inc(&sh->count);
1403 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1404 sh, to_addr_conv(sh, percpu));
1405 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1406 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1407 }
1408
1409 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1410 {
1411 int overlap_clear = 0, i, disks = sh->disks;
1412 struct dma_async_tx_descriptor *tx = NULL;
1413 struct r5conf *conf = sh->raid_conf;
1414 int level = conf->level;
1415 struct raid5_percpu *percpu;
1416 unsigned long cpu;
1417
1418 cpu = get_cpu();
1419 percpu = per_cpu_ptr(conf->percpu, cpu);
1420 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1421 ops_run_biofill(sh);
1422 overlap_clear++;
1423 }
1424
1425 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1426 if (level < 6)
1427 tx = ops_run_compute5(sh, percpu);
1428 else {
1429 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1430 tx = ops_run_compute6_1(sh, percpu);
1431 else
1432 tx = ops_run_compute6_2(sh, percpu);
1433 }
1434 /* terminate the chain if reconstruct is not set to be run */
1435 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1436 async_tx_ack(tx);
1437 }
1438
1439 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1440 tx = ops_run_prexor(sh, percpu, tx);
1441
1442 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1443 tx = ops_run_biodrain(sh, tx);
1444 overlap_clear++;
1445 }
1446
1447 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1448 if (level < 6)
1449 ops_run_reconstruct5(sh, percpu, tx);
1450 else
1451 ops_run_reconstruct6(sh, percpu, tx);
1452 }
1453
1454 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1455 if (sh->check_state == check_state_run)
1456 ops_run_check_p(sh, percpu);
1457 else if (sh->check_state == check_state_run_q)
1458 ops_run_check_pq(sh, percpu, 0);
1459 else if (sh->check_state == check_state_run_pq)
1460 ops_run_check_pq(sh, percpu, 1);
1461 else
1462 BUG();
1463 }
1464
1465 if (overlap_clear)
1466 for (i = disks; i--; ) {
1467 struct r5dev *dev = &sh->dev[i];
1468 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1469 wake_up(&sh->raid_conf->wait_for_overlap);
1470 }
1471 put_cpu();
1472 }
1473
1474 #ifdef CONFIG_MULTICORE_RAID456
1475 static void async_run_ops(void *param, async_cookie_t cookie)
1476 {
1477 struct stripe_head *sh = param;
1478 unsigned long ops_request = sh->ops.request;
1479
1480 clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1481 wake_up(&sh->ops.wait_for_ops);
1482
1483 __raid_run_ops(sh, ops_request);
1484 release_stripe(sh);
1485 }
1486
1487 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1488 {
1489 /* since handle_stripe can be called outside of raid5d context
1490 * we need to ensure sh->ops.request is de-staged before another
1491 * request arrives
1492 */
1493 wait_event(sh->ops.wait_for_ops,
1494 !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1495 sh->ops.request = ops_request;
1496
1497 atomic_inc(&sh->count);
1498 async_schedule(async_run_ops, sh);
1499 }
1500 #else
1501 #define raid_run_ops __raid_run_ops
1502 #endif
1503
1504 static int grow_one_stripe(struct r5conf *conf)
1505 {
1506 struct stripe_head *sh;
1507 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1508 if (!sh)
1509 return 0;
1510
1511 sh->raid_conf = conf;
1512 #ifdef CONFIG_MULTICORE_RAID456
1513 init_waitqueue_head(&sh->ops.wait_for_ops);
1514 #endif
1515
1516 spin_lock_init(&sh->stripe_lock);
1517
1518 if (grow_buffers(sh)) {
1519 shrink_buffers(sh);
1520 kmem_cache_free(conf->slab_cache, sh);
1521 return 0;
1522 }
1523 /* we just created an active stripe so... */
1524 atomic_set(&sh->count, 1);
1525 atomic_inc(&conf->active_stripes);
1526 INIT_LIST_HEAD(&sh->lru);
1527 release_stripe(sh);
1528 return 1;
1529 }
1530
1531 static int grow_stripes(struct r5conf *conf, int num)
1532 {
1533 struct kmem_cache *sc;
1534 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1535
1536 if (conf->mddev->gendisk)
1537 sprintf(conf->cache_name[0],
1538 "raid%d-%s", conf->level, mdname(conf->mddev));
1539 else
1540 sprintf(conf->cache_name[0],
1541 "raid%d-%p", conf->level, conf->mddev);
1542 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1543
1544 conf->active_name = 0;
1545 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1546 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1547 0, 0, NULL);
1548 if (!sc)
1549 return 1;
1550 conf->slab_cache = sc;
1551 conf->pool_size = devs;
1552 while (num--)
1553 if (!grow_one_stripe(conf))
1554 return 1;
1555 return 0;
1556 }
1557
1558 /**
1559 * scribble_len - return the required size of the scribble region
1560 * @num - total number of disks in the array
1561 *
1562 * The size must be enough to contain:
1563 * 1/ a struct page pointer for each device in the array +2
1564 * 2/ room to convert each entry in (1) to its corresponding dma
1565 * (dma_map_page()) or page (page_address()) address.
1566 *
1567 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1568 * calculate over all devices (not just the data blocks), using zeros in place
1569 * of the P and Q blocks.
1570 */
1571 static size_t scribble_len(int num)
1572 {
1573 size_t len;
1574
1575 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1576
1577 return len;
1578 }
1579
1580 static int resize_stripes(struct r5conf *conf, int newsize)
1581 {
1582 /* Make all the stripes able to hold 'newsize' devices.
1583 * New slots in each stripe get 'page' set to a new page.
1584 *
1585 * This happens in stages:
1586 * 1/ create a new kmem_cache and allocate the required number of
1587 * stripe_heads.
1588 * 2/ gather all the old stripe_heads and transfer the pages across
1589 * to the new stripe_heads. This will have the side effect of
1590 * freezing the array as once all stripe_heads have been collected,
1591 * no IO will be possible. Old stripe heads are freed once their
1592 * pages have been transferred over, and the old kmem_cache is
1593 * freed when all stripes are done.
1594 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1595 * we simple return a failre status - no need to clean anything up.
1596 * 4/ allocate new pages for the new slots in the new stripe_heads.
1597 * If this fails, we don't bother trying the shrink the
1598 * stripe_heads down again, we just leave them as they are.
1599 * As each stripe_head is processed the new one is released into
1600 * active service.
1601 *
1602 * Once step2 is started, we cannot afford to wait for a write,
1603 * so we use GFP_NOIO allocations.
1604 */
1605 struct stripe_head *osh, *nsh;
1606 LIST_HEAD(newstripes);
1607 struct disk_info *ndisks;
1608 unsigned long cpu;
1609 int err;
1610 struct kmem_cache *sc;
1611 int i;
1612
1613 if (newsize <= conf->pool_size)
1614 return 0; /* never bother to shrink */
1615
1616 err = md_allow_write(conf->mddev);
1617 if (err)
1618 return err;
1619
1620 /* Step 1 */
1621 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1622 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1623 0, 0, NULL);
1624 if (!sc)
1625 return -ENOMEM;
1626
1627 for (i = conf->max_nr_stripes; i; i--) {
1628 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1629 if (!nsh)
1630 break;
1631
1632 nsh->raid_conf = conf;
1633 #ifdef CONFIG_MULTICORE_RAID456
1634 init_waitqueue_head(&nsh->ops.wait_for_ops);
1635 #endif
1636 spin_lock_init(&nsh->stripe_lock);
1637
1638 list_add(&nsh->lru, &newstripes);
1639 }
1640 if (i) {
1641 /* didn't get enough, give up */
1642 while (!list_empty(&newstripes)) {
1643 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1644 list_del(&nsh->lru);
1645 kmem_cache_free(sc, nsh);
1646 }
1647 kmem_cache_destroy(sc);
1648 return -ENOMEM;
1649 }
1650 /* Step 2 - Must use GFP_NOIO now.
1651 * OK, we have enough stripes, start collecting inactive
1652 * stripes and copying them over
1653 */
1654 list_for_each_entry(nsh, &newstripes, lru) {
1655 spin_lock_irq(&conf->device_lock);
1656 wait_event_lock_irq(conf->wait_for_stripe,
1657 !list_empty(&conf->inactive_list),
1658 conf->device_lock);
1659 osh = get_free_stripe(conf);
1660 spin_unlock_irq(&conf->device_lock);
1661 atomic_set(&nsh->count, 1);
1662 for(i=0; i<conf->pool_size; i++)
1663 nsh->dev[i].page = osh->dev[i].page;
1664 for( ; i<newsize; i++)
1665 nsh->dev[i].page = NULL;
1666 kmem_cache_free(conf->slab_cache, osh);
1667 }
1668 kmem_cache_destroy(conf->slab_cache);
1669
1670 /* Step 3.
1671 * At this point, we are holding all the stripes so the array
1672 * is completely stalled, so now is a good time to resize
1673 * conf->disks and the scribble region
1674 */
1675 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1676 if (ndisks) {
1677 for (i=0; i<conf->raid_disks; i++)
1678 ndisks[i] = conf->disks[i];
1679 kfree(conf->disks);
1680 conf->disks = ndisks;
1681 } else
1682 err = -ENOMEM;
1683
1684 get_online_cpus();
1685 conf->scribble_len = scribble_len(newsize);
1686 for_each_present_cpu(cpu) {
1687 struct raid5_percpu *percpu;
1688 void *scribble;
1689
1690 percpu = per_cpu_ptr(conf->percpu, cpu);
1691 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1692
1693 if (scribble) {
1694 kfree(percpu->scribble);
1695 percpu->scribble = scribble;
1696 } else {
1697 err = -ENOMEM;
1698 break;
1699 }
1700 }
1701 put_online_cpus();
1702
1703 /* Step 4, return new stripes to service */
1704 while(!list_empty(&newstripes)) {
1705 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1706 list_del_init(&nsh->lru);
1707
1708 for (i=conf->raid_disks; i < newsize; i++)
1709 if (nsh->dev[i].page == NULL) {
1710 struct page *p = alloc_page(GFP_NOIO);
1711 nsh->dev[i].page = p;
1712 if (!p)
1713 err = -ENOMEM;
1714 }
1715 release_stripe(nsh);
1716 }
1717 /* critical section pass, GFP_NOIO no longer needed */
1718
1719 conf->slab_cache = sc;
1720 conf->active_name = 1-conf->active_name;
1721 conf->pool_size = newsize;
1722 return err;
1723 }
1724
1725 static int drop_one_stripe(struct r5conf *conf)
1726 {
1727 struct stripe_head *sh;
1728
1729 spin_lock_irq(&conf->device_lock);
1730 sh = get_free_stripe(conf);
1731 spin_unlock_irq(&conf->device_lock);
1732 if (!sh)
1733 return 0;
1734 BUG_ON(atomic_read(&sh->count));
1735 shrink_buffers(sh);
1736 kmem_cache_free(conf->slab_cache, sh);
1737 atomic_dec(&conf->active_stripes);
1738 return 1;
1739 }
1740
1741 static void shrink_stripes(struct r5conf *conf)
1742 {
1743 while (drop_one_stripe(conf))
1744 ;
1745
1746 if (conf->slab_cache)
1747 kmem_cache_destroy(conf->slab_cache);
1748 conf->slab_cache = NULL;
1749 }
1750
1751 static void raid5_end_read_request(struct bio * bi, int error)
1752 {
1753 struct stripe_head *sh = bi->bi_private;
1754 struct r5conf *conf = sh->raid_conf;
1755 int disks = sh->disks, i;
1756 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1757 char b[BDEVNAME_SIZE];
1758 struct md_rdev *rdev = NULL;
1759 sector_t s;
1760
1761 for (i=0 ; i<disks; i++)
1762 if (bi == &sh->dev[i].req)
1763 break;
1764
1765 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1766 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1767 uptodate);
1768 if (i == disks) {
1769 BUG();
1770 return;
1771 }
1772 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1773 /* If replacement finished while this request was outstanding,
1774 * 'replacement' might be NULL already.
1775 * In that case it moved down to 'rdev'.
1776 * rdev is not removed until all requests are finished.
1777 */
1778 rdev = conf->disks[i].replacement;
1779 if (!rdev)
1780 rdev = conf->disks[i].rdev;
1781
1782 if (use_new_offset(conf, sh))
1783 s = sh->sector + rdev->new_data_offset;
1784 else
1785 s = sh->sector + rdev->data_offset;
1786 if (uptodate) {
1787 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1788 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1789 /* Note that this cannot happen on a
1790 * replacement device. We just fail those on
1791 * any error
1792 */
1793 printk_ratelimited(
1794 KERN_INFO
1795 "md/raid:%s: read error corrected"
1796 " (%lu sectors at %llu on %s)\n",
1797 mdname(conf->mddev), STRIPE_SECTORS,
1798 (unsigned long long)s,
1799 bdevname(rdev->bdev, b));
1800 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1801 clear_bit(R5_ReadError, &sh->dev[i].flags);
1802 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1803 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1804 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1805
1806 if (atomic_read(&rdev->read_errors))
1807 atomic_set(&rdev->read_errors, 0);
1808 } else {
1809 const char *bdn = bdevname(rdev->bdev, b);
1810 int retry = 0;
1811 int set_bad = 0;
1812
1813 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1814 atomic_inc(&rdev->read_errors);
1815 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1816 printk_ratelimited(
1817 KERN_WARNING
1818 "md/raid:%s: read error on replacement device "
1819 "(sector %llu on %s).\n",
1820 mdname(conf->mddev),
1821 (unsigned long long)s,
1822 bdn);
1823 else if (conf->mddev->degraded >= conf->max_degraded) {
1824 set_bad = 1;
1825 printk_ratelimited(
1826 KERN_WARNING
1827 "md/raid:%s: read error not correctable "
1828 "(sector %llu on %s).\n",
1829 mdname(conf->mddev),
1830 (unsigned long long)s,
1831 bdn);
1832 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1833 /* Oh, no!!! */
1834 set_bad = 1;
1835 printk_ratelimited(
1836 KERN_WARNING
1837 "md/raid:%s: read error NOT corrected!! "
1838 "(sector %llu on %s).\n",
1839 mdname(conf->mddev),
1840 (unsigned long long)s,
1841 bdn);
1842 } else if (atomic_read(&rdev->read_errors)
1843 > conf->max_nr_stripes)
1844 printk(KERN_WARNING
1845 "md/raid:%s: Too many read errors, failing device %s.\n",
1846 mdname(conf->mddev), bdn);
1847 else
1848 retry = 1;
1849 if (retry)
1850 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1851 set_bit(R5_ReadError, &sh->dev[i].flags);
1852 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1853 } else
1854 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1855 else {
1856 clear_bit(R5_ReadError, &sh->dev[i].flags);
1857 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1858 if (!(set_bad
1859 && test_bit(In_sync, &rdev->flags)
1860 && rdev_set_badblocks(
1861 rdev, sh->sector, STRIPE_SECTORS, 0)))
1862 md_error(conf->mddev, rdev);
1863 }
1864 }
1865 rdev_dec_pending(rdev, conf->mddev);
1866 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1867 set_bit(STRIPE_HANDLE, &sh->state);
1868 release_stripe(sh);
1869 }
1870
1871 static void raid5_end_write_request(struct bio *bi, int error)
1872 {
1873 struct stripe_head *sh = bi->bi_private;
1874 struct r5conf *conf = sh->raid_conf;
1875 int disks = sh->disks, i;
1876 struct md_rdev *uninitialized_var(rdev);
1877 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1878 sector_t first_bad;
1879 int bad_sectors;
1880 int replacement = 0;
1881
1882 for (i = 0 ; i < disks; i++) {
1883 if (bi == &sh->dev[i].req) {
1884 rdev = conf->disks[i].rdev;
1885 break;
1886 }
1887 if (bi == &sh->dev[i].rreq) {
1888 rdev = conf->disks[i].replacement;
1889 if (rdev)
1890 replacement = 1;
1891 else
1892 /* rdev was removed and 'replacement'
1893 * replaced it. rdev is not removed
1894 * until all requests are finished.
1895 */
1896 rdev = conf->disks[i].rdev;
1897 break;
1898 }
1899 }
1900 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1901 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1902 uptodate);
1903 if (i == disks) {
1904 BUG();
1905 return;
1906 }
1907
1908 if (replacement) {
1909 if (!uptodate)
1910 md_error(conf->mddev, rdev);
1911 else if (is_badblock(rdev, sh->sector,
1912 STRIPE_SECTORS,
1913 &first_bad, &bad_sectors))
1914 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1915 } else {
1916 if (!uptodate) {
1917 set_bit(WriteErrorSeen, &rdev->flags);
1918 set_bit(R5_WriteError, &sh->dev[i].flags);
1919 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1920 set_bit(MD_RECOVERY_NEEDED,
1921 &rdev->mddev->recovery);
1922 } else if (is_badblock(rdev, sh->sector,
1923 STRIPE_SECTORS,
1924 &first_bad, &bad_sectors))
1925 set_bit(R5_MadeGood, &sh->dev[i].flags);
1926 }
1927 rdev_dec_pending(rdev, conf->mddev);
1928
1929 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1930 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1931 set_bit(STRIPE_HANDLE, &sh->state);
1932 release_stripe(sh);
1933 }
1934
1935 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1936
1937 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1938 {
1939 struct r5dev *dev = &sh->dev[i];
1940
1941 bio_init(&dev->req);
1942 dev->req.bi_io_vec = &dev->vec;
1943 dev->req.bi_vcnt++;
1944 dev->req.bi_max_vecs++;
1945 dev->req.bi_private = sh;
1946 dev->vec.bv_page = dev->page;
1947
1948 bio_init(&dev->rreq);
1949 dev->rreq.bi_io_vec = &dev->rvec;
1950 dev->rreq.bi_vcnt++;
1951 dev->rreq.bi_max_vecs++;
1952 dev->rreq.bi_private = sh;
1953 dev->rvec.bv_page = dev->page;
1954
1955 dev->flags = 0;
1956 dev->sector = compute_blocknr(sh, i, previous);
1957 }
1958
1959 static void error(struct mddev *mddev, struct md_rdev *rdev)
1960 {
1961 char b[BDEVNAME_SIZE];
1962 struct r5conf *conf = mddev->private;
1963 unsigned long flags;
1964 pr_debug("raid456: error called\n");
1965
1966 spin_lock_irqsave(&conf->device_lock, flags);
1967 clear_bit(In_sync, &rdev->flags);
1968 mddev->degraded = calc_degraded(conf);
1969 spin_unlock_irqrestore(&conf->device_lock, flags);
1970 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1971
1972 set_bit(Blocked, &rdev->flags);
1973 set_bit(Faulty, &rdev->flags);
1974 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1975 printk(KERN_ALERT
1976 "md/raid:%s: Disk failure on %s, disabling device.\n"
1977 "md/raid:%s: Operation continuing on %d devices.\n",
1978 mdname(mddev),
1979 bdevname(rdev->bdev, b),
1980 mdname(mddev),
1981 conf->raid_disks - mddev->degraded);
1982 }
1983
1984 /*
1985 * Input: a 'big' sector number,
1986 * Output: index of the data and parity disk, and the sector # in them.
1987 */
1988 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1989 int previous, int *dd_idx,
1990 struct stripe_head *sh)
1991 {
1992 sector_t stripe, stripe2;
1993 sector_t chunk_number;
1994 unsigned int chunk_offset;
1995 int pd_idx, qd_idx;
1996 int ddf_layout = 0;
1997 sector_t new_sector;
1998 int algorithm = previous ? conf->prev_algo
1999 : conf->algorithm;
2000 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2001 : conf->chunk_sectors;
2002 int raid_disks = previous ? conf->previous_raid_disks
2003 : conf->raid_disks;
2004 int data_disks = raid_disks - conf->max_degraded;
2005
2006 /* First compute the information on this sector */
2007
2008 /*
2009 * Compute the chunk number and the sector offset inside the chunk
2010 */
2011 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2012 chunk_number = r_sector;
2013
2014 /*
2015 * Compute the stripe number
2016 */
2017 stripe = chunk_number;
2018 *dd_idx = sector_div(stripe, data_disks);
2019 stripe2 = stripe;
2020 /*
2021 * Select the parity disk based on the user selected algorithm.
2022 */
2023 pd_idx = qd_idx = -1;
2024 switch(conf->level) {
2025 case 4:
2026 pd_idx = data_disks;
2027 break;
2028 case 5:
2029 switch (algorithm) {
2030 case ALGORITHM_LEFT_ASYMMETRIC:
2031 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2032 if (*dd_idx >= pd_idx)
2033 (*dd_idx)++;
2034 break;
2035 case ALGORITHM_RIGHT_ASYMMETRIC:
2036 pd_idx = sector_div(stripe2, raid_disks);
2037 if (*dd_idx >= pd_idx)
2038 (*dd_idx)++;
2039 break;
2040 case ALGORITHM_LEFT_SYMMETRIC:
2041 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2042 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2043 break;
2044 case ALGORITHM_RIGHT_SYMMETRIC:
2045 pd_idx = sector_div(stripe2, raid_disks);
2046 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2047 break;
2048 case ALGORITHM_PARITY_0:
2049 pd_idx = 0;
2050 (*dd_idx)++;
2051 break;
2052 case ALGORITHM_PARITY_N:
2053 pd_idx = data_disks;
2054 break;
2055 default:
2056 BUG();
2057 }
2058 break;
2059 case 6:
2060
2061 switch (algorithm) {
2062 case ALGORITHM_LEFT_ASYMMETRIC:
2063 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2064 qd_idx = pd_idx + 1;
2065 if (pd_idx == raid_disks-1) {
2066 (*dd_idx)++; /* Q D D D P */
2067 qd_idx = 0;
2068 } else if (*dd_idx >= pd_idx)
2069 (*dd_idx) += 2; /* D D P Q D */
2070 break;
2071 case ALGORITHM_RIGHT_ASYMMETRIC:
2072 pd_idx = sector_div(stripe2, raid_disks);
2073 qd_idx = pd_idx + 1;
2074 if (pd_idx == raid_disks-1) {
2075 (*dd_idx)++; /* Q D D D P */
2076 qd_idx = 0;
2077 } else if (*dd_idx >= pd_idx)
2078 (*dd_idx) += 2; /* D D P Q D */
2079 break;
2080 case ALGORITHM_LEFT_SYMMETRIC:
2081 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2082 qd_idx = (pd_idx + 1) % raid_disks;
2083 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2084 break;
2085 case ALGORITHM_RIGHT_SYMMETRIC:
2086 pd_idx = sector_div(stripe2, raid_disks);
2087 qd_idx = (pd_idx + 1) % raid_disks;
2088 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2089 break;
2090
2091 case ALGORITHM_PARITY_0:
2092 pd_idx = 0;
2093 qd_idx = 1;
2094 (*dd_idx) += 2;
2095 break;
2096 case ALGORITHM_PARITY_N:
2097 pd_idx = data_disks;
2098 qd_idx = data_disks + 1;
2099 break;
2100
2101 case ALGORITHM_ROTATING_ZERO_RESTART:
2102 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2103 * of blocks for computing Q is different.
2104 */
2105 pd_idx = sector_div(stripe2, raid_disks);
2106 qd_idx = pd_idx + 1;
2107 if (pd_idx == raid_disks-1) {
2108 (*dd_idx)++; /* Q D D D P */
2109 qd_idx = 0;
2110 } else if (*dd_idx >= pd_idx)
2111 (*dd_idx) += 2; /* D D P Q D */
2112 ddf_layout = 1;
2113 break;
2114
2115 case ALGORITHM_ROTATING_N_RESTART:
2116 /* Same a left_asymmetric, by first stripe is
2117 * D D D P Q rather than
2118 * Q D D D P
2119 */
2120 stripe2 += 1;
2121 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2122 qd_idx = pd_idx + 1;
2123 if (pd_idx == raid_disks-1) {
2124 (*dd_idx)++; /* Q D D D P */
2125 qd_idx = 0;
2126 } else if (*dd_idx >= pd_idx)
2127 (*dd_idx) += 2; /* D D P Q D */
2128 ddf_layout = 1;
2129 break;
2130
2131 case ALGORITHM_ROTATING_N_CONTINUE:
2132 /* Same as left_symmetric but Q is before P */
2133 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2134 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2135 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2136 ddf_layout = 1;
2137 break;
2138
2139 case ALGORITHM_LEFT_ASYMMETRIC_6:
2140 /* RAID5 left_asymmetric, with Q on last device */
2141 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2142 if (*dd_idx >= pd_idx)
2143 (*dd_idx)++;
2144 qd_idx = raid_disks - 1;
2145 break;
2146
2147 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2148 pd_idx = sector_div(stripe2, raid_disks-1);
2149 if (*dd_idx >= pd_idx)
2150 (*dd_idx)++;
2151 qd_idx = raid_disks - 1;
2152 break;
2153
2154 case ALGORITHM_LEFT_SYMMETRIC_6:
2155 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2156 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2157 qd_idx = raid_disks - 1;
2158 break;
2159
2160 case ALGORITHM_RIGHT_SYMMETRIC_6:
2161 pd_idx = sector_div(stripe2, raid_disks-1);
2162 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2163 qd_idx = raid_disks - 1;
2164 break;
2165
2166 case ALGORITHM_PARITY_0_6:
2167 pd_idx = 0;
2168 (*dd_idx)++;
2169 qd_idx = raid_disks - 1;
2170 break;
2171
2172 default:
2173 BUG();
2174 }
2175 break;
2176 }
2177
2178 if (sh) {
2179 sh->pd_idx = pd_idx;
2180 sh->qd_idx = qd_idx;
2181 sh->ddf_layout = ddf_layout;
2182 }
2183 /*
2184 * Finally, compute the new sector number
2185 */
2186 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2187 return new_sector;
2188 }
2189
2190
2191 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2192 {
2193 struct r5conf *conf = sh->raid_conf;
2194 int raid_disks = sh->disks;
2195 int data_disks = raid_disks - conf->max_degraded;
2196 sector_t new_sector = sh->sector, check;
2197 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2198 : conf->chunk_sectors;
2199 int algorithm = previous ? conf->prev_algo
2200 : conf->algorithm;
2201 sector_t stripe;
2202 int chunk_offset;
2203 sector_t chunk_number;
2204 int dummy1, dd_idx = i;
2205 sector_t r_sector;
2206 struct stripe_head sh2;
2207
2208
2209 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2210 stripe = new_sector;
2211
2212 if (i == sh->pd_idx)
2213 return 0;
2214 switch(conf->level) {
2215 case 4: break;
2216 case 5:
2217 switch (algorithm) {
2218 case ALGORITHM_LEFT_ASYMMETRIC:
2219 case ALGORITHM_RIGHT_ASYMMETRIC:
2220 if (i > sh->pd_idx)
2221 i--;
2222 break;
2223 case ALGORITHM_LEFT_SYMMETRIC:
2224 case ALGORITHM_RIGHT_SYMMETRIC:
2225 if (i < sh->pd_idx)
2226 i += raid_disks;
2227 i -= (sh->pd_idx + 1);
2228 break;
2229 case ALGORITHM_PARITY_0:
2230 i -= 1;
2231 break;
2232 case ALGORITHM_PARITY_N:
2233 break;
2234 default:
2235 BUG();
2236 }
2237 break;
2238 case 6:
2239 if (i == sh->qd_idx)
2240 return 0; /* It is the Q disk */
2241 switch (algorithm) {
2242 case ALGORITHM_LEFT_ASYMMETRIC:
2243 case ALGORITHM_RIGHT_ASYMMETRIC:
2244 case ALGORITHM_ROTATING_ZERO_RESTART:
2245 case ALGORITHM_ROTATING_N_RESTART:
2246 if (sh->pd_idx == raid_disks-1)
2247 i--; /* Q D D D P */
2248 else if (i > sh->pd_idx)
2249 i -= 2; /* D D P Q D */
2250 break;
2251 case ALGORITHM_LEFT_SYMMETRIC:
2252 case ALGORITHM_RIGHT_SYMMETRIC:
2253 if (sh->pd_idx == raid_disks-1)
2254 i--; /* Q D D D P */
2255 else {
2256 /* D D P Q D */
2257 if (i < sh->pd_idx)
2258 i += raid_disks;
2259 i -= (sh->pd_idx + 2);
2260 }
2261 break;
2262 case ALGORITHM_PARITY_0:
2263 i -= 2;
2264 break;
2265 case ALGORITHM_PARITY_N:
2266 break;
2267 case ALGORITHM_ROTATING_N_CONTINUE:
2268 /* Like left_symmetric, but P is before Q */
2269 if (sh->pd_idx == 0)
2270 i--; /* P D D D Q */
2271 else {
2272 /* D D Q P D */
2273 if (i < sh->pd_idx)
2274 i += raid_disks;
2275 i -= (sh->pd_idx + 1);
2276 }
2277 break;
2278 case ALGORITHM_LEFT_ASYMMETRIC_6:
2279 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2280 if (i > sh->pd_idx)
2281 i--;
2282 break;
2283 case ALGORITHM_LEFT_SYMMETRIC_6:
2284 case ALGORITHM_RIGHT_SYMMETRIC_6:
2285 if (i < sh->pd_idx)
2286 i += data_disks + 1;
2287 i -= (sh->pd_idx + 1);
2288 break;
2289 case ALGORITHM_PARITY_0_6:
2290 i -= 1;
2291 break;
2292 default:
2293 BUG();
2294 }
2295 break;
2296 }
2297
2298 chunk_number = stripe * data_disks + i;
2299 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2300
2301 check = raid5_compute_sector(conf, r_sector,
2302 previous, &dummy1, &sh2);
2303 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2304 || sh2.qd_idx != sh->qd_idx) {
2305 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2306 mdname(conf->mddev));
2307 return 0;
2308 }
2309 return r_sector;
2310 }
2311
2312
2313 static void
2314 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2315 int rcw, int expand)
2316 {
2317 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2318 struct r5conf *conf = sh->raid_conf;
2319 int level = conf->level;
2320
2321 if (rcw) {
2322 /* if we are not expanding this is a proper write request, and
2323 * there will be bios with new data to be drained into the
2324 * stripe cache
2325 */
2326 if (!expand) {
2327 sh->reconstruct_state = reconstruct_state_drain_run;
2328 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2329 } else
2330 sh->reconstruct_state = reconstruct_state_run;
2331
2332 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2333
2334 for (i = disks; i--; ) {
2335 struct r5dev *dev = &sh->dev[i];
2336
2337 if (dev->towrite) {
2338 set_bit(R5_LOCKED, &dev->flags);
2339 set_bit(R5_Wantdrain, &dev->flags);
2340 if (!expand)
2341 clear_bit(R5_UPTODATE, &dev->flags);
2342 s->locked++;
2343 }
2344 }
2345 if (s->locked + conf->max_degraded == disks)
2346 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2347 atomic_inc(&conf->pending_full_writes);
2348 } else {
2349 BUG_ON(level == 6);
2350 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2351 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2352
2353 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2354 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2355 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2356 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2357
2358 for (i = disks; i--; ) {
2359 struct r5dev *dev = &sh->dev[i];
2360 if (i == pd_idx)
2361 continue;
2362
2363 if (dev->towrite &&
2364 (test_bit(R5_UPTODATE, &dev->flags) ||
2365 test_bit(R5_Wantcompute, &dev->flags))) {
2366 set_bit(R5_Wantdrain, &dev->flags);
2367 set_bit(R5_LOCKED, &dev->flags);
2368 clear_bit(R5_UPTODATE, &dev->flags);
2369 s->locked++;
2370 }
2371 }
2372 }
2373
2374 /* keep the parity disk(s) locked while asynchronous operations
2375 * are in flight
2376 */
2377 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2378 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2379 s->locked++;
2380
2381 if (level == 6) {
2382 int qd_idx = sh->qd_idx;
2383 struct r5dev *dev = &sh->dev[qd_idx];
2384
2385 set_bit(R5_LOCKED, &dev->flags);
2386 clear_bit(R5_UPTODATE, &dev->flags);
2387 s->locked++;
2388 }
2389
2390 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2391 __func__, (unsigned long long)sh->sector,
2392 s->locked, s->ops_request);
2393 }
2394
2395 /*
2396 * Each stripe/dev can have one or more bion attached.
2397 * toread/towrite point to the first in a chain.
2398 * The bi_next chain must be in order.
2399 */
2400 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2401 {
2402 struct bio **bip;
2403 struct r5conf *conf = sh->raid_conf;
2404 int firstwrite=0;
2405
2406 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2407 (unsigned long long)bi->bi_sector,
2408 (unsigned long long)sh->sector);
2409
2410 /*
2411 * If several bio share a stripe. The bio bi_phys_segments acts as a
2412 * reference count to avoid race. The reference count should already be
2413 * increased before this function is called (for example, in
2414 * make_request()), so other bio sharing this stripe will not free the
2415 * stripe. If a stripe is owned by one stripe, the stripe lock will
2416 * protect it.
2417 */
2418 spin_lock_irq(&sh->stripe_lock);
2419 if (forwrite) {
2420 bip = &sh->dev[dd_idx].towrite;
2421 if (*bip == NULL)
2422 firstwrite = 1;
2423 } else
2424 bip = &sh->dev[dd_idx].toread;
2425 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2426 if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2427 goto overlap;
2428 bip = & (*bip)->bi_next;
2429 }
2430 if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2431 goto overlap;
2432
2433 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2434 if (*bip)
2435 bi->bi_next = *bip;
2436 *bip = bi;
2437 raid5_inc_bi_active_stripes(bi);
2438
2439 if (forwrite) {
2440 /* check if page is covered */
2441 sector_t sector = sh->dev[dd_idx].sector;
2442 for (bi=sh->dev[dd_idx].towrite;
2443 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2444 bi && bi->bi_sector <= sector;
2445 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2446 if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2447 sector = bi->bi_sector + (bi->bi_size>>9);
2448 }
2449 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2450 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2451 }
2452
2453 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2454 (unsigned long long)(*bip)->bi_sector,
2455 (unsigned long long)sh->sector, dd_idx);
2456 spin_unlock_irq(&sh->stripe_lock);
2457
2458 if (conf->mddev->bitmap && firstwrite) {
2459 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2460 STRIPE_SECTORS, 0);
2461 sh->bm_seq = conf->seq_flush+1;
2462 set_bit(STRIPE_BIT_DELAY, &sh->state);
2463 }
2464 return 1;
2465
2466 overlap:
2467 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2468 spin_unlock_irq(&sh->stripe_lock);
2469 return 0;
2470 }
2471
2472 static void end_reshape(struct r5conf *conf);
2473
2474 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2475 struct stripe_head *sh)
2476 {
2477 int sectors_per_chunk =
2478 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2479 int dd_idx;
2480 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2481 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2482
2483 raid5_compute_sector(conf,
2484 stripe * (disks - conf->max_degraded)
2485 *sectors_per_chunk + chunk_offset,
2486 previous,
2487 &dd_idx, sh);
2488 }
2489
2490 static void
2491 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2492 struct stripe_head_state *s, int disks,
2493 struct bio **return_bi)
2494 {
2495 int i;
2496 for (i = disks; i--; ) {
2497 struct bio *bi;
2498 int bitmap_end = 0;
2499
2500 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2501 struct md_rdev *rdev;
2502 rcu_read_lock();
2503 rdev = rcu_dereference(conf->disks[i].rdev);
2504 if (rdev && test_bit(In_sync, &rdev->flags))
2505 atomic_inc(&rdev->nr_pending);
2506 else
2507 rdev = NULL;
2508 rcu_read_unlock();
2509 if (rdev) {
2510 if (!rdev_set_badblocks(
2511 rdev,
2512 sh->sector,
2513 STRIPE_SECTORS, 0))
2514 md_error(conf->mddev, rdev);
2515 rdev_dec_pending(rdev, conf->mddev);
2516 }
2517 }
2518 spin_lock_irq(&sh->stripe_lock);
2519 /* fail all writes first */
2520 bi = sh->dev[i].towrite;
2521 sh->dev[i].towrite = NULL;
2522 spin_unlock_irq(&sh->stripe_lock);
2523 if (bi)
2524 bitmap_end = 1;
2525
2526 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2527 wake_up(&conf->wait_for_overlap);
2528
2529 while (bi && bi->bi_sector <
2530 sh->dev[i].sector + STRIPE_SECTORS) {
2531 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2532 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2533 if (!raid5_dec_bi_active_stripes(bi)) {
2534 md_write_end(conf->mddev);
2535 bi->bi_next = *return_bi;
2536 *return_bi = bi;
2537 }
2538 bi = nextbi;
2539 }
2540 if (bitmap_end)
2541 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2542 STRIPE_SECTORS, 0, 0);
2543 bitmap_end = 0;
2544 /* and fail all 'written' */
2545 bi = sh->dev[i].written;
2546 sh->dev[i].written = NULL;
2547 if (bi) bitmap_end = 1;
2548 while (bi && bi->bi_sector <
2549 sh->dev[i].sector + STRIPE_SECTORS) {
2550 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2551 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2552 if (!raid5_dec_bi_active_stripes(bi)) {
2553 md_write_end(conf->mddev);
2554 bi->bi_next = *return_bi;
2555 *return_bi = bi;
2556 }
2557 bi = bi2;
2558 }
2559
2560 /* fail any reads if this device is non-operational and
2561 * the data has not reached the cache yet.
2562 */
2563 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2564 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2565 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2566 spin_lock_irq(&sh->stripe_lock);
2567 bi = sh->dev[i].toread;
2568 sh->dev[i].toread = NULL;
2569 spin_unlock_irq(&sh->stripe_lock);
2570 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2571 wake_up(&conf->wait_for_overlap);
2572 while (bi && bi->bi_sector <
2573 sh->dev[i].sector + STRIPE_SECTORS) {
2574 struct bio *nextbi =
2575 r5_next_bio(bi, sh->dev[i].sector);
2576 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2577 if (!raid5_dec_bi_active_stripes(bi)) {
2578 bi->bi_next = *return_bi;
2579 *return_bi = bi;
2580 }
2581 bi = nextbi;
2582 }
2583 }
2584 if (bitmap_end)
2585 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2586 STRIPE_SECTORS, 0, 0);
2587 /* If we were in the middle of a write the parity block might
2588 * still be locked - so just clear all R5_LOCKED flags
2589 */
2590 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2591 }
2592
2593 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2594 if (atomic_dec_and_test(&conf->pending_full_writes))
2595 md_wakeup_thread(conf->mddev->thread);
2596 }
2597
2598 static void
2599 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2600 struct stripe_head_state *s)
2601 {
2602 int abort = 0;
2603 int i;
2604
2605 clear_bit(STRIPE_SYNCING, &sh->state);
2606 s->syncing = 0;
2607 s->replacing = 0;
2608 /* There is nothing more to do for sync/check/repair.
2609 * Don't even need to abort as that is handled elsewhere
2610 * if needed, and not always wanted e.g. if there is a known
2611 * bad block here.
2612 * For recover/replace we need to record a bad block on all
2613 * non-sync devices, or abort the recovery
2614 */
2615 if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2616 /* During recovery devices cannot be removed, so
2617 * locking and refcounting of rdevs is not needed
2618 */
2619 for (i = 0; i < conf->raid_disks; i++) {
2620 struct md_rdev *rdev = conf->disks[i].rdev;
2621 if (rdev
2622 && !test_bit(Faulty, &rdev->flags)
2623 && !test_bit(In_sync, &rdev->flags)
2624 && !rdev_set_badblocks(rdev, sh->sector,
2625 STRIPE_SECTORS, 0))
2626 abort = 1;
2627 rdev = conf->disks[i].replacement;
2628 if (rdev
2629 && !test_bit(Faulty, &rdev->flags)
2630 && !test_bit(In_sync, &rdev->flags)
2631 && !rdev_set_badblocks(rdev, sh->sector,
2632 STRIPE_SECTORS, 0))
2633 abort = 1;
2634 }
2635 if (abort)
2636 conf->recovery_disabled =
2637 conf->mddev->recovery_disabled;
2638 }
2639 md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2640 }
2641
2642 static int want_replace(struct stripe_head *sh, int disk_idx)
2643 {
2644 struct md_rdev *rdev;
2645 int rv = 0;
2646 /* Doing recovery so rcu locking not required */
2647 rdev = sh->raid_conf->disks[disk_idx].replacement;
2648 if (rdev
2649 && !test_bit(Faulty, &rdev->flags)
2650 && !test_bit(In_sync, &rdev->flags)
2651 && (rdev->recovery_offset <= sh->sector
2652 || rdev->mddev->recovery_cp <= sh->sector))
2653 rv = 1;
2654
2655 return rv;
2656 }
2657
2658 /* fetch_block - checks the given member device to see if its data needs
2659 * to be read or computed to satisfy a request.
2660 *
2661 * Returns 1 when no more member devices need to be checked, otherwise returns
2662 * 0 to tell the loop in handle_stripe_fill to continue
2663 */
2664 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2665 int disk_idx, int disks)
2666 {
2667 struct r5dev *dev = &sh->dev[disk_idx];
2668 struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2669 &sh->dev[s->failed_num[1]] };
2670
2671 /* is the data in this block needed, and can we get it? */
2672 if (!test_bit(R5_LOCKED, &dev->flags) &&
2673 !test_bit(R5_UPTODATE, &dev->flags) &&
2674 (dev->toread ||
2675 (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2676 s->syncing || s->expanding ||
2677 (s->replacing && want_replace(sh, disk_idx)) ||
2678 (s->failed >= 1 && fdev[0]->toread) ||
2679 (s->failed >= 2 && fdev[1]->toread) ||
2680 (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2681 !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2682 (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2683 /* we would like to get this block, possibly by computing it,
2684 * otherwise read it if the backing disk is insync
2685 */
2686 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2687 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2688 if ((s->uptodate == disks - 1) &&
2689 (s->failed && (disk_idx == s->failed_num[0] ||
2690 disk_idx == s->failed_num[1]))) {
2691 /* have disk failed, and we're requested to fetch it;
2692 * do compute it
2693 */
2694 pr_debug("Computing stripe %llu block %d\n",
2695 (unsigned long long)sh->sector, disk_idx);
2696 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2697 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2698 set_bit(R5_Wantcompute, &dev->flags);
2699 sh->ops.target = disk_idx;
2700 sh->ops.target2 = -1; /* no 2nd target */
2701 s->req_compute = 1;
2702 /* Careful: from this point on 'uptodate' is in the eye
2703 * of raid_run_ops which services 'compute' operations
2704 * before writes. R5_Wantcompute flags a block that will
2705 * be R5_UPTODATE by the time it is needed for a
2706 * subsequent operation.
2707 */
2708 s->uptodate++;
2709 return 1;
2710 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2711 /* Computing 2-failure is *very* expensive; only
2712 * do it if failed >= 2
2713 */
2714 int other;
2715 for (other = disks; other--; ) {
2716 if (other == disk_idx)
2717 continue;
2718 if (!test_bit(R5_UPTODATE,
2719 &sh->dev[other].flags))
2720 break;
2721 }
2722 BUG_ON(other < 0);
2723 pr_debug("Computing stripe %llu blocks %d,%d\n",
2724 (unsigned long long)sh->sector,
2725 disk_idx, other);
2726 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2727 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2728 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2729 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2730 sh->ops.target = disk_idx;
2731 sh->ops.target2 = other;
2732 s->uptodate += 2;
2733 s->req_compute = 1;
2734 return 1;
2735 } else if (test_bit(R5_Insync, &dev->flags)) {
2736 set_bit(R5_LOCKED, &dev->flags);
2737 set_bit(R5_Wantread, &dev->flags);
2738 s->locked++;
2739 pr_debug("Reading block %d (sync=%d)\n",
2740 disk_idx, s->syncing);
2741 }
2742 }
2743
2744 return 0;
2745 }
2746
2747 /**
2748 * handle_stripe_fill - read or compute data to satisfy pending requests.
2749 */
2750 static void handle_stripe_fill(struct stripe_head *sh,
2751 struct stripe_head_state *s,
2752 int disks)
2753 {
2754 int i;
2755
2756 /* look for blocks to read/compute, skip this if a compute
2757 * is already in flight, or if the stripe contents are in the
2758 * midst of changing due to a write
2759 */
2760 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2761 !sh->reconstruct_state)
2762 for (i = disks; i--; )
2763 if (fetch_block(sh, s, i, disks))
2764 break;
2765 set_bit(STRIPE_HANDLE, &sh->state);
2766 }
2767
2768
2769 /* handle_stripe_clean_event
2770 * any written block on an uptodate or failed drive can be returned.
2771 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2772 * never LOCKED, so we don't need to test 'failed' directly.
2773 */
2774 static void handle_stripe_clean_event(struct r5conf *conf,
2775 struct stripe_head *sh, int disks, struct bio **return_bi)
2776 {
2777 int i;
2778 struct r5dev *dev;
2779
2780 for (i = disks; i--; )
2781 if (sh->dev[i].written) {
2782 dev = &sh->dev[i];
2783 if (!test_bit(R5_LOCKED, &dev->flags) &&
2784 (test_bit(R5_UPTODATE, &dev->flags) ||
2785 test_bit(R5_Discard, &dev->flags))) {
2786 /* We can return any write requests */
2787 struct bio *wbi, *wbi2;
2788 pr_debug("Return write for disc %d\n", i);
2789 if (test_and_clear_bit(R5_Discard, &dev->flags))
2790 clear_bit(R5_UPTODATE, &dev->flags);
2791 wbi = dev->written;
2792 dev->written = NULL;
2793 while (wbi && wbi->bi_sector <
2794 dev->sector + STRIPE_SECTORS) {
2795 wbi2 = r5_next_bio(wbi, dev->sector);
2796 if (!raid5_dec_bi_active_stripes(wbi)) {
2797 md_write_end(conf->mddev);
2798 wbi->bi_next = *return_bi;
2799 *return_bi = wbi;
2800 }
2801 wbi = wbi2;
2802 }
2803 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2804 STRIPE_SECTORS,
2805 !test_bit(STRIPE_DEGRADED, &sh->state),
2806 0);
2807 }
2808 } else if (test_bit(R5_Discard, &sh->dev[i].flags))
2809 clear_bit(R5_Discard, &sh->dev[i].flags);
2810
2811 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2812 if (atomic_dec_and_test(&conf->pending_full_writes))
2813 md_wakeup_thread(conf->mddev->thread);
2814 }
2815
2816 static void handle_stripe_dirtying(struct r5conf *conf,
2817 struct stripe_head *sh,
2818 struct stripe_head_state *s,
2819 int disks)
2820 {
2821 int rmw = 0, rcw = 0, i;
2822 sector_t recovery_cp = conf->mddev->recovery_cp;
2823
2824 /* RAID6 requires 'rcw' in current implementation.
2825 * Otherwise, check whether resync is now happening or should start.
2826 * If yes, then the array is dirty (after unclean shutdown or
2827 * initial creation), so parity in some stripes might be inconsistent.
2828 * In this case, we need to always do reconstruct-write, to ensure
2829 * that in case of drive failure or read-error correction, we
2830 * generate correct data from the parity.
2831 */
2832 if (conf->max_degraded == 2 ||
2833 (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2834 /* Calculate the real rcw later - for now make it
2835 * look like rcw is cheaper
2836 */
2837 rcw = 1; rmw = 2;
2838 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2839 conf->max_degraded, (unsigned long long)recovery_cp,
2840 (unsigned long long)sh->sector);
2841 } else for (i = disks; i--; ) {
2842 /* would I have to read this buffer for read_modify_write */
2843 struct r5dev *dev = &sh->dev[i];
2844 if ((dev->towrite || i == sh->pd_idx) &&
2845 !test_bit(R5_LOCKED, &dev->flags) &&
2846 !(test_bit(R5_UPTODATE, &dev->flags) ||
2847 test_bit(R5_Wantcompute, &dev->flags))) {
2848 if (test_bit(R5_Insync, &dev->flags))
2849 rmw++;
2850 else
2851 rmw += 2*disks; /* cannot read it */
2852 }
2853 /* Would I have to read this buffer for reconstruct_write */
2854 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2855 !test_bit(R5_LOCKED, &dev->flags) &&
2856 !(test_bit(R5_UPTODATE, &dev->flags) ||
2857 test_bit(R5_Wantcompute, &dev->flags))) {
2858 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2859 else
2860 rcw += 2*disks;
2861 }
2862 }
2863 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2864 (unsigned long long)sh->sector, rmw, rcw);
2865 set_bit(STRIPE_HANDLE, &sh->state);
2866 if (rmw < rcw && rmw > 0) {
2867 /* prefer read-modify-write, but need to get some data */
2868 blk_add_trace_msg(conf->mddev->queue, "raid5 rmw %llu %d",
2869 (unsigned long long)sh->sector, rmw);
2870 for (i = disks; i--; ) {
2871 struct r5dev *dev = &sh->dev[i];
2872 if ((dev->towrite || i == sh->pd_idx) &&
2873 !test_bit(R5_LOCKED, &dev->flags) &&
2874 !(test_bit(R5_UPTODATE, &dev->flags) ||
2875 test_bit(R5_Wantcompute, &dev->flags)) &&
2876 test_bit(R5_Insync, &dev->flags)) {
2877 if (
2878 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2879 pr_debug("Read_old block "
2880 "%d for r-m-w\n", i);
2881 set_bit(R5_LOCKED, &dev->flags);
2882 set_bit(R5_Wantread, &dev->flags);
2883 s->locked++;
2884 } else {
2885 set_bit(STRIPE_DELAYED, &sh->state);
2886 set_bit(STRIPE_HANDLE, &sh->state);
2887 }
2888 }
2889 }
2890 }
2891 if (rcw <= rmw && rcw > 0) {
2892 /* want reconstruct write, but need to get some data */
2893 int qread =0;
2894 rcw = 0;
2895 for (i = disks; i--; ) {
2896 struct r5dev *dev = &sh->dev[i];
2897 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2898 i != sh->pd_idx && i != sh->qd_idx &&
2899 !test_bit(R5_LOCKED, &dev->flags) &&
2900 !(test_bit(R5_UPTODATE, &dev->flags) ||
2901 test_bit(R5_Wantcompute, &dev->flags))) {
2902 rcw++;
2903 if (!test_bit(R5_Insync, &dev->flags))
2904 continue; /* it's a failed drive */
2905 if (
2906 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2907 pr_debug("Read_old block "
2908 "%d for Reconstruct\n", i);
2909 set_bit(R5_LOCKED, &dev->flags);
2910 set_bit(R5_Wantread, &dev->flags);
2911 s->locked++;
2912 qread++;
2913 } else {
2914 set_bit(STRIPE_DELAYED, &sh->state);
2915 set_bit(STRIPE_HANDLE, &sh->state);
2916 }
2917 }
2918 }
2919 if (rcw)
2920 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
2921 (unsigned long long)sh->sector,
2922 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
2923 }
2924 /* now if nothing is locked, and if we have enough data,
2925 * we can start a write request
2926 */
2927 /* since handle_stripe can be called at any time we need to handle the
2928 * case where a compute block operation has been submitted and then a
2929 * subsequent call wants to start a write request. raid_run_ops only
2930 * handles the case where compute block and reconstruct are requested
2931 * simultaneously. If this is not the case then new writes need to be
2932 * held off until the compute completes.
2933 */
2934 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2935 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2936 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2937 schedule_reconstruction(sh, s, rcw == 0, 0);
2938 }
2939
2940 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2941 struct stripe_head_state *s, int disks)
2942 {
2943 struct r5dev *dev = NULL;
2944
2945 set_bit(STRIPE_HANDLE, &sh->state);
2946
2947 switch (sh->check_state) {
2948 case check_state_idle:
2949 /* start a new check operation if there are no failures */
2950 if (s->failed == 0) {
2951 BUG_ON(s->uptodate != disks);
2952 sh->check_state = check_state_run;
2953 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2954 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2955 s->uptodate--;
2956 break;
2957 }
2958 dev = &sh->dev[s->failed_num[0]];
2959 /* fall through */
2960 case check_state_compute_result:
2961 sh->check_state = check_state_idle;
2962 if (!dev)
2963 dev = &sh->dev[sh->pd_idx];
2964
2965 /* check that a write has not made the stripe insync */
2966 if (test_bit(STRIPE_INSYNC, &sh->state))
2967 break;
2968
2969 /* either failed parity check, or recovery is happening */
2970 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2971 BUG_ON(s->uptodate != disks);
2972
2973 set_bit(R5_LOCKED, &dev->flags);
2974 s->locked++;
2975 set_bit(R5_Wantwrite, &dev->flags);
2976
2977 clear_bit(STRIPE_DEGRADED, &sh->state);
2978 set_bit(STRIPE_INSYNC, &sh->state);
2979 break;
2980 case check_state_run:
2981 break; /* we will be called again upon completion */
2982 case check_state_check_result:
2983 sh->check_state = check_state_idle;
2984
2985 /* if a failure occurred during the check operation, leave
2986 * STRIPE_INSYNC not set and let the stripe be handled again
2987 */
2988 if (s->failed)
2989 break;
2990
2991 /* handle a successful check operation, if parity is correct
2992 * we are done. Otherwise update the mismatch count and repair
2993 * parity if !MD_RECOVERY_CHECK
2994 */
2995 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2996 /* parity is correct (on disc,
2997 * not in buffer any more)
2998 */
2999 set_bit(STRIPE_INSYNC, &sh->state);
3000 else {
3001 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3002 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3003 /* don't try to repair!! */
3004 set_bit(STRIPE_INSYNC, &sh->state);
3005 else {
3006 sh->check_state = check_state_compute_run;
3007 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3008 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3009 set_bit(R5_Wantcompute,
3010 &sh->dev[sh->pd_idx].flags);
3011 sh->ops.target = sh->pd_idx;
3012 sh->ops.target2 = -1;
3013 s->uptodate++;
3014 }
3015 }
3016 break;
3017 case check_state_compute_run:
3018 break;
3019 default:
3020 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3021 __func__, sh->check_state,
3022 (unsigned long long) sh->sector);
3023 BUG();
3024 }
3025 }
3026
3027
3028 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3029 struct stripe_head_state *s,
3030 int disks)
3031 {
3032 int pd_idx = sh->pd_idx;
3033 int qd_idx = sh->qd_idx;
3034 struct r5dev *dev;
3035
3036 set_bit(STRIPE_HANDLE, &sh->state);
3037
3038 BUG_ON(s->failed > 2);
3039
3040 /* Want to check and possibly repair P and Q.
3041 * However there could be one 'failed' device, in which
3042 * case we can only check one of them, possibly using the
3043 * other to generate missing data
3044 */
3045
3046 switch (sh->check_state) {
3047 case check_state_idle:
3048 /* start a new check operation if there are < 2 failures */
3049 if (s->failed == s->q_failed) {
3050 /* The only possible failed device holds Q, so it
3051 * makes sense to check P (If anything else were failed,
3052 * we would have used P to recreate it).
3053 */
3054 sh->check_state = check_state_run;
3055 }
3056 if (!s->q_failed && s->failed < 2) {
3057 /* Q is not failed, and we didn't use it to generate
3058 * anything, so it makes sense to check it
3059 */
3060 if (sh->check_state == check_state_run)
3061 sh->check_state = check_state_run_pq;
3062 else
3063 sh->check_state = check_state_run_q;
3064 }
3065
3066 /* discard potentially stale zero_sum_result */
3067 sh->ops.zero_sum_result = 0;
3068
3069 if (sh->check_state == check_state_run) {
3070 /* async_xor_zero_sum destroys the contents of P */
3071 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3072 s->uptodate--;
3073 }
3074 if (sh->check_state >= check_state_run &&
3075 sh->check_state <= check_state_run_pq) {
3076 /* async_syndrome_zero_sum preserves P and Q, so
3077 * no need to mark them !uptodate here
3078 */
3079 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3080 break;
3081 }
3082
3083 /* we have 2-disk failure */
3084 BUG_ON(s->failed != 2);
3085 /* fall through */
3086 case check_state_compute_result:
3087 sh->check_state = check_state_idle;
3088
3089 /* check that a write has not made the stripe insync */
3090 if (test_bit(STRIPE_INSYNC, &sh->state))
3091 break;
3092
3093 /* now write out any block on a failed drive,
3094 * or P or Q if they were recomputed
3095 */
3096 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3097 if (s->failed == 2) {
3098 dev = &sh->dev[s->failed_num[1]];
3099 s->locked++;
3100 set_bit(R5_LOCKED, &dev->flags);
3101 set_bit(R5_Wantwrite, &dev->flags);
3102 }
3103 if (s->failed >= 1) {
3104 dev = &sh->dev[s->failed_num[0]];
3105 s->locked++;
3106 set_bit(R5_LOCKED, &dev->flags);
3107 set_bit(R5_Wantwrite, &dev->flags);
3108 }
3109 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3110 dev = &sh->dev[pd_idx];
3111 s->locked++;
3112 set_bit(R5_LOCKED, &dev->flags);
3113 set_bit(R5_Wantwrite, &dev->flags);
3114 }
3115 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3116 dev = &sh->dev[qd_idx];
3117 s->locked++;
3118 set_bit(R5_LOCKED, &dev->flags);
3119 set_bit(R5_Wantwrite, &dev->flags);
3120 }
3121 clear_bit(STRIPE_DEGRADED, &sh->state);
3122
3123 set_bit(STRIPE_INSYNC, &sh->state);
3124 break;
3125 case check_state_run:
3126 case check_state_run_q:
3127 case check_state_run_pq:
3128 break; /* we will be called again upon completion */
3129 case check_state_check_result:
3130 sh->check_state = check_state_idle;
3131
3132 /* handle a successful check operation, if parity is correct
3133 * we are done. Otherwise update the mismatch count and repair
3134 * parity if !MD_RECOVERY_CHECK
3135 */
3136 if (sh->ops.zero_sum_result == 0) {
3137 /* both parities are correct */
3138 if (!s->failed)
3139 set_bit(STRIPE_INSYNC, &sh->state);
3140 else {
3141 /* in contrast to the raid5 case we can validate
3142 * parity, but still have a failure to write
3143 * back
3144 */
3145 sh->check_state = check_state_compute_result;
3146 /* Returning at this point means that we may go
3147 * off and bring p and/or q uptodate again so
3148 * we make sure to check zero_sum_result again
3149 * to verify if p or q need writeback
3150 */
3151 }
3152 } else {
3153 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3154 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3155 /* don't try to repair!! */
3156 set_bit(STRIPE_INSYNC, &sh->state);
3157 else {
3158 int *target = &sh->ops.target;
3159
3160 sh->ops.target = -1;
3161 sh->ops.target2 = -1;
3162 sh->check_state = check_state_compute_run;
3163 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3164 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3165 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3166 set_bit(R5_Wantcompute,
3167 &sh->dev[pd_idx].flags);
3168 *target = pd_idx;
3169 target = &sh->ops.target2;
3170 s->uptodate++;
3171 }
3172 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3173 set_bit(R5_Wantcompute,
3174 &sh->dev[qd_idx].flags);
3175 *target = qd_idx;
3176 s->uptodate++;
3177 }
3178 }
3179 }
3180 break;
3181 case check_state_compute_run:
3182 break;
3183 default:
3184 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3185 __func__, sh->check_state,
3186 (unsigned long long) sh->sector);
3187 BUG();
3188 }
3189 }
3190
3191 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3192 {
3193 int i;
3194
3195 /* We have read all the blocks in this stripe and now we need to
3196 * copy some of them into a target stripe for expand.
3197 */
3198 struct dma_async_tx_descriptor *tx = NULL;
3199 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3200 for (i = 0; i < sh->disks; i++)
3201 if (i != sh->pd_idx && i != sh->qd_idx) {
3202 int dd_idx, j;
3203 struct stripe_head *sh2;
3204 struct async_submit_ctl submit;
3205
3206 sector_t bn = compute_blocknr(sh, i, 1);
3207 sector_t s = raid5_compute_sector(conf, bn, 0,
3208 &dd_idx, NULL);
3209 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3210 if (sh2 == NULL)
3211 /* so far only the early blocks of this stripe
3212 * have been requested. When later blocks
3213 * get requested, we will try again
3214 */
3215 continue;
3216 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3217 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3218 /* must have already done this block */
3219 release_stripe(sh2);
3220 continue;
3221 }
3222
3223 /* place all the copies on one channel */
3224 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3225 tx = async_memcpy(sh2->dev[dd_idx].page,
3226 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3227 &submit);
3228
3229 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3230 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3231 for (j = 0; j < conf->raid_disks; j++)
3232 if (j != sh2->pd_idx &&
3233 j != sh2->qd_idx &&
3234 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3235 break;
3236 if (j == conf->raid_disks) {
3237 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3238 set_bit(STRIPE_HANDLE, &sh2->state);
3239 }
3240 release_stripe(sh2);
3241
3242 }
3243 /* done submitting copies, wait for them to complete */
3244 async_tx_quiesce(&tx);
3245 }
3246
3247 /*
3248 * handle_stripe - do things to a stripe.
3249 *
3250 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3251 * state of various bits to see what needs to be done.
3252 * Possible results:
3253 * return some read requests which now have data
3254 * return some write requests which are safely on storage
3255 * schedule a read on some buffers
3256 * schedule a write of some buffers
3257 * return confirmation of parity correctness
3258 *
3259 */
3260
3261 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3262 {
3263 struct r5conf *conf = sh->raid_conf;
3264 int disks = sh->disks;
3265 struct r5dev *dev;
3266 int i;
3267 int do_recovery = 0;
3268
3269 memset(s, 0, sizeof(*s));
3270
3271 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3272 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3273 s->failed_num[0] = -1;
3274 s->failed_num[1] = -1;
3275
3276 /* Now to look around and see what can be done */
3277 rcu_read_lock();
3278 for (i=disks; i--; ) {
3279 struct md_rdev *rdev;
3280 sector_t first_bad;
3281 int bad_sectors;
3282 int is_bad = 0;
3283
3284 dev = &sh->dev[i];
3285
3286 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3287 i, dev->flags,
3288 dev->toread, dev->towrite, dev->written);
3289 /* maybe we can reply to a read
3290 *
3291 * new wantfill requests are only permitted while
3292 * ops_complete_biofill is guaranteed to be inactive
3293 */
3294 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3295 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3296 set_bit(R5_Wantfill, &dev->flags);
3297
3298 /* now count some things */
3299 if (test_bit(R5_LOCKED, &dev->flags))
3300 s->locked++;
3301 if (test_bit(R5_UPTODATE, &dev->flags))
3302 s->uptodate++;
3303 if (test_bit(R5_Wantcompute, &dev->flags)) {
3304 s->compute++;
3305 BUG_ON(s->compute > 2);
3306 }
3307
3308 if (test_bit(R5_Wantfill, &dev->flags))
3309 s->to_fill++;
3310 else if (dev->toread)
3311 s->to_read++;
3312 if (dev->towrite) {
3313 s->to_write++;
3314 if (!test_bit(R5_OVERWRITE, &dev->flags))
3315 s->non_overwrite++;
3316 }
3317 if (dev->written)
3318 s->written++;
3319 /* Prefer to use the replacement for reads, but only
3320 * if it is recovered enough and has no bad blocks.
3321 */
3322 rdev = rcu_dereference(conf->disks[i].replacement);
3323 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3324 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3325 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3326 &first_bad, &bad_sectors))
3327 set_bit(R5_ReadRepl, &dev->flags);
3328 else {
3329 if (rdev)
3330 set_bit(R5_NeedReplace, &dev->flags);
3331 rdev = rcu_dereference(conf->disks[i].rdev);
3332 clear_bit(R5_ReadRepl, &dev->flags);
3333 }
3334 if (rdev && test_bit(Faulty, &rdev->flags))
3335 rdev = NULL;
3336 if (rdev) {
3337 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3338 &first_bad, &bad_sectors);
3339 if (s->blocked_rdev == NULL
3340 && (test_bit(Blocked, &rdev->flags)
3341 || is_bad < 0)) {
3342 if (is_bad < 0)
3343 set_bit(BlockedBadBlocks,
3344 &rdev->flags);
3345 s->blocked_rdev = rdev;
3346 atomic_inc(&rdev->nr_pending);
3347 }
3348 }
3349 clear_bit(R5_Insync, &dev->flags);
3350 if (!rdev)
3351 /* Not in-sync */;
3352 else if (is_bad) {
3353 /* also not in-sync */
3354 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3355 test_bit(R5_UPTODATE, &dev->flags)) {
3356 /* treat as in-sync, but with a read error
3357 * which we can now try to correct
3358 */
3359 set_bit(R5_Insync, &dev->flags);
3360 set_bit(R5_ReadError, &dev->flags);
3361 }
3362 } else if (test_bit(In_sync, &rdev->flags))
3363 set_bit(R5_Insync, &dev->flags);
3364 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3365 /* in sync if before recovery_offset */
3366 set_bit(R5_Insync, &dev->flags);
3367 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3368 test_bit(R5_Expanded, &dev->flags))
3369 /* If we've reshaped into here, we assume it is Insync.
3370 * We will shortly update recovery_offset to make
3371 * it official.
3372 */
3373 set_bit(R5_Insync, &dev->flags);
3374
3375 if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3376 /* This flag does not apply to '.replacement'
3377 * only to .rdev, so make sure to check that*/
3378 struct md_rdev *rdev2 = rcu_dereference(
3379 conf->disks[i].rdev);
3380 if (rdev2 == rdev)
3381 clear_bit(R5_Insync, &dev->flags);
3382 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3383 s->handle_bad_blocks = 1;
3384 atomic_inc(&rdev2->nr_pending);
3385 } else
3386 clear_bit(R5_WriteError, &dev->flags);
3387 }
3388 if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3389 /* This flag does not apply to '.replacement'
3390 * only to .rdev, so make sure to check that*/
3391 struct md_rdev *rdev2 = rcu_dereference(
3392 conf->disks[i].rdev);
3393 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3394 s->handle_bad_blocks = 1;
3395 atomic_inc(&rdev2->nr_pending);
3396 } else
3397 clear_bit(R5_MadeGood, &dev->flags);
3398 }
3399 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3400 struct md_rdev *rdev2 = rcu_dereference(
3401 conf->disks[i].replacement);
3402 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3403 s->handle_bad_blocks = 1;
3404 atomic_inc(&rdev2->nr_pending);
3405 } else
3406 clear_bit(R5_MadeGoodRepl, &dev->flags);
3407 }
3408 if (!test_bit(R5_Insync, &dev->flags)) {
3409 /* The ReadError flag will just be confusing now */
3410 clear_bit(R5_ReadError, &dev->flags);
3411 clear_bit(R5_ReWrite, &dev->flags);
3412 }
3413 if (test_bit(R5_ReadError, &dev->flags))
3414 clear_bit(R5_Insync, &dev->flags);
3415 if (!test_bit(R5_Insync, &dev->flags)) {
3416 if (s->failed < 2)
3417 s->failed_num[s->failed] = i;
3418 s->failed++;
3419 if (rdev && !test_bit(Faulty, &rdev->flags))
3420 do_recovery = 1;
3421 }
3422 }
3423 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3424 /* If there is a failed device being replaced,
3425 * we must be recovering.
3426 * else if we are after recovery_cp, we must be syncing
3427 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3428 * else we can only be replacing
3429 * sync and recovery both need to read all devices, and so
3430 * use the same flag.
3431 */
3432 if (do_recovery ||
3433 sh->sector >= conf->mddev->recovery_cp ||
3434 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3435 s->syncing = 1;
3436 else
3437 s->replacing = 1;
3438 }
3439 rcu_read_unlock();
3440 }
3441
3442 static void handle_stripe(struct stripe_head *sh)
3443 {
3444 struct stripe_head_state s;
3445 struct r5conf *conf = sh->raid_conf;
3446 int i;
3447 int prexor;
3448 int disks = sh->disks;
3449 struct r5dev *pdev, *qdev;
3450
3451 clear_bit(STRIPE_HANDLE, &sh->state);
3452 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3453 /* already being handled, ensure it gets handled
3454 * again when current action finishes */
3455 set_bit(STRIPE_HANDLE, &sh->state);
3456 return;
3457 }
3458
3459 if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3460 set_bit(STRIPE_SYNCING, &sh->state);
3461 clear_bit(STRIPE_INSYNC, &sh->state);
3462 }
3463 clear_bit(STRIPE_DELAYED, &sh->state);
3464
3465 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3466 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3467 (unsigned long long)sh->sector, sh->state,
3468 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3469 sh->check_state, sh->reconstruct_state);
3470
3471 analyse_stripe(sh, &s);
3472
3473 if (s.handle_bad_blocks) {
3474 set_bit(STRIPE_HANDLE, &sh->state);
3475 goto finish;
3476 }
3477
3478 if (unlikely(s.blocked_rdev)) {
3479 if (s.syncing || s.expanding || s.expanded ||
3480 s.replacing || s.to_write || s.written) {
3481 set_bit(STRIPE_HANDLE, &sh->state);
3482 goto finish;
3483 }
3484 /* There is nothing for the blocked_rdev to block */
3485 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3486 s.blocked_rdev = NULL;
3487 }
3488
3489 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3490 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3491 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3492 }
3493
3494 pr_debug("locked=%d uptodate=%d to_read=%d"
3495 " to_write=%d failed=%d failed_num=%d,%d\n",
3496 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3497 s.failed_num[0], s.failed_num[1]);
3498 /* check if the array has lost more than max_degraded devices and,
3499 * if so, some requests might need to be failed.
3500 */
3501 if (s.failed > conf->max_degraded) {
3502 sh->check_state = 0;
3503 sh->reconstruct_state = 0;
3504 if (s.to_read+s.to_write+s.written)
3505 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3506 if (s.syncing + s.replacing)
3507 handle_failed_sync(conf, sh, &s);
3508 }
3509
3510 /* Now we check to see if any write operations have recently
3511 * completed
3512 */
3513 prexor = 0;
3514 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3515 prexor = 1;
3516 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3517 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3518 sh->reconstruct_state = reconstruct_state_idle;
3519
3520 /* All the 'written' buffers and the parity block are ready to
3521 * be written back to disk
3522 */
3523 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3524 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3525 BUG_ON(sh->qd_idx >= 0 &&
3526 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3527 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3528 for (i = disks; i--; ) {
3529 struct r5dev *dev = &sh->dev[i];
3530 if (test_bit(R5_LOCKED, &dev->flags) &&
3531 (i == sh->pd_idx || i == sh->qd_idx ||
3532 dev->written)) {
3533 pr_debug("Writing block %d\n", i);
3534 set_bit(R5_Wantwrite, &dev->flags);
3535 if (prexor)
3536 continue;
3537 if (!test_bit(R5_Insync, &dev->flags) ||
3538 ((i == sh->pd_idx || i == sh->qd_idx) &&
3539 s.failed == 0))
3540 set_bit(STRIPE_INSYNC, &sh->state);
3541 }
3542 }
3543 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3544 s.dec_preread_active = 1;
3545 }
3546
3547 /*
3548 * might be able to return some write requests if the parity blocks
3549 * are safe, or on a failed drive
3550 */
3551 pdev = &sh->dev[sh->pd_idx];
3552 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3553 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3554 qdev = &sh->dev[sh->qd_idx];
3555 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3556 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3557 || conf->level < 6;
3558
3559 if (s.written &&
3560 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3561 && !test_bit(R5_LOCKED, &pdev->flags)
3562 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3563 test_bit(R5_Discard, &pdev->flags))))) &&
3564 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3565 && !test_bit(R5_LOCKED, &qdev->flags)
3566 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3567 test_bit(R5_Discard, &qdev->flags))))))
3568 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3569
3570 /* Now we might consider reading some blocks, either to check/generate
3571 * parity, or to satisfy requests
3572 * or to load a block that is being partially written.
3573 */
3574 if (s.to_read || s.non_overwrite
3575 || (conf->level == 6 && s.to_write && s.failed)
3576 || (s.syncing && (s.uptodate + s.compute < disks))
3577 || s.replacing
3578 || s.expanding)
3579 handle_stripe_fill(sh, &s, disks);
3580
3581 /* Now to consider new write requests and what else, if anything
3582 * should be read. We do not handle new writes when:
3583 * 1/ A 'write' operation (copy+xor) is already in flight.
3584 * 2/ A 'check' operation is in flight, as it may clobber the parity
3585 * block.
3586 */
3587 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3588 handle_stripe_dirtying(conf, sh, &s, disks);
3589
3590 /* maybe we need to check and possibly fix the parity for this stripe
3591 * Any reads will already have been scheduled, so we just see if enough
3592 * data is available. The parity check is held off while parity
3593 * dependent operations are in flight.
3594 */
3595 if (sh->check_state ||
3596 (s.syncing && s.locked == 0 &&
3597 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3598 !test_bit(STRIPE_INSYNC, &sh->state))) {
3599 if (conf->level == 6)
3600 handle_parity_checks6(conf, sh, &s, disks);
3601 else
3602 handle_parity_checks5(conf, sh, &s, disks);
3603 }
3604
3605 if (s.replacing && s.locked == 0
3606 && !test_bit(STRIPE_INSYNC, &sh->state)) {
3607 /* Write out to replacement devices where possible */
3608 for (i = 0; i < conf->raid_disks; i++)
3609 if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3610 test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3611 set_bit(R5_WantReplace, &sh->dev[i].flags);
3612 set_bit(R5_LOCKED, &sh->dev[i].flags);
3613 s.locked++;
3614 }
3615 set_bit(STRIPE_INSYNC, &sh->state);
3616 }
3617 if ((s.syncing || s.replacing) && s.locked == 0 &&
3618 test_bit(STRIPE_INSYNC, &sh->state)) {
3619 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3620 clear_bit(STRIPE_SYNCING, &sh->state);
3621 }
3622
3623 /* If the failed drives are just a ReadError, then we might need
3624 * to progress the repair/check process
3625 */
3626 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3627 for (i = 0; i < s.failed; i++) {
3628 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3629 if (test_bit(R5_ReadError, &dev->flags)
3630 && !test_bit(R5_LOCKED, &dev->flags)
3631 && test_bit(R5_UPTODATE, &dev->flags)
3632 ) {
3633 if (!test_bit(R5_ReWrite, &dev->flags)) {
3634 set_bit(R5_Wantwrite, &dev->flags);
3635 set_bit(R5_ReWrite, &dev->flags);
3636 set_bit(R5_LOCKED, &dev->flags);
3637 s.locked++;
3638 } else {
3639 /* let's read it back */
3640 set_bit(R5_Wantread, &dev->flags);
3641 set_bit(R5_LOCKED, &dev->flags);
3642 s.locked++;
3643 }
3644 }
3645 }
3646
3647
3648 /* Finish reconstruct operations initiated by the expansion process */
3649 if (sh->reconstruct_state == reconstruct_state_result) {
3650 struct stripe_head *sh_src
3651 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3652 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3653 /* sh cannot be written until sh_src has been read.
3654 * so arrange for sh to be delayed a little
3655 */
3656 set_bit(STRIPE_DELAYED, &sh->state);
3657 set_bit(STRIPE_HANDLE, &sh->state);
3658 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3659 &sh_src->state))
3660 atomic_inc(&conf->preread_active_stripes);
3661 release_stripe(sh_src);
3662 goto finish;
3663 }
3664 if (sh_src)
3665 release_stripe(sh_src);
3666
3667 sh->reconstruct_state = reconstruct_state_idle;
3668 clear_bit(STRIPE_EXPANDING, &sh->state);
3669 for (i = conf->raid_disks; i--; ) {
3670 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3671 set_bit(R5_LOCKED, &sh->dev[i].flags);
3672 s.locked++;
3673 }
3674 }
3675
3676 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3677 !sh->reconstruct_state) {
3678 /* Need to write out all blocks after computing parity */
3679 sh->disks = conf->raid_disks;
3680 stripe_set_idx(sh->sector, conf, 0, sh);
3681 schedule_reconstruction(sh, &s, 1, 1);
3682 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3683 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3684 atomic_dec(&conf->reshape_stripes);
3685 wake_up(&conf->wait_for_overlap);
3686 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3687 }
3688
3689 if (s.expanding && s.locked == 0 &&
3690 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3691 handle_stripe_expansion(conf, sh);
3692
3693 finish:
3694 /* wait for this device to become unblocked */
3695 if (unlikely(s.blocked_rdev)) {
3696 if (conf->mddev->external)
3697 md_wait_for_blocked_rdev(s.blocked_rdev,
3698 conf->mddev);
3699 else
3700 /* Internal metadata will immediately
3701 * be written by raid5d, so we don't
3702 * need to wait here.
3703 */
3704 rdev_dec_pending(s.blocked_rdev,
3705 conf->mddev);
3706 }
3707
3708 if (s.handle_bad_blocks)
3709 for (i = disks; i--; ) {
3710 struct md_rdev *rdev;
3711 struct r5dev *dev = &sh->dev[i];
3712 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3713 /* We own a safe reference to the rdev */
3714 rdev = conf->disks[i].rdev;
3715 if (!rdev_set_badblocks(rdev, sh->sector,
3716 STRIPE_SECTORS, 0))
3717 md_error(conf->mddev, rdev);
3718 rdev_dec_pending(rdev, conf->mddev);
3719 }
3720 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3721 rdev = conf->disks[i].rdev;
3722 rdev_clear_badblocks(rdev, sh->sector,
3723 STRIPE_SECTORS, 0);
3724 rdev_dec_pending(rdev, conf->mddev);
3725 }
3726 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3727 rdev = conf->disks[i].replacement;
3728 if (!rdev)
3729 /* rdev have been moved down */
3730 rdev = conf->disks[i].rdev;
3731 rdev_clear_badblocks(rdev, sh->sector,
3732 STRIPE_SECTORS, 0);
3733 rdev_dec_pending(rdev, conf->mddev);
3734 }
3735 }
3736
3737 if (s.ops_request)
3738 raid_run_ops(sh, s.ops_request);
3739
3740 ops_run_io(sh, &s);
3741
3742 if (s.dec_preread_active) {
3743 /* We delay this until after ops_run_io so that if make_request
3744 * is waiting on a flush, it won't continue until the writes
3745 * have actually been submitted.
3746 */
3747 atomic_dec(&conf->preread_active_stripes);
3748 if (atomic_read(&conf->preread_active_stripes) <
3749 IO_THRESHOLD)
3750 md_wakeup_thread(conf->mddev->thread);
3751 }
3752
3753 return_io(s.return_bi);
3754
3755 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3756 }
3757
3758 static void raid5_activate_delayed(struct r5conf *conf)
3759 {
3760 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3761 while (!list_empty(&conf->delayed_list)) {
3762 struct list_head *l = conf->delayed_list.next;
3763 struct stripe_head *sh;
3764 sh = list_entry(l, struct stripe_head, lru);
3765 list_del_init(l);
3766 clear_bit(STRIPE_DELAYED, &sh->state);
3767 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3768 atomic_inc(&conf->preread_active_stripes);
3769 list_add_tail(&sh->lru, &conf->hold_list);
3770 }
3771 }
3772 }
3773
3774 static void activate_bit_delay(struct r5conf *conf)
3775 {
3776 /* device_lock is held */
3777 struct list_head head;
3778 list_add(&head, &conf->bitmap_list);
3779 list_del_init(&conf->bitmap_list);
3780 while (!list_empty(&head)) {
3781 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3782 list_del_init(&sh->lru);
3783 atomic_inc(&sh->count);
3784 __release_stripe(conf, sh);
3785 }
3786 }
3787
3788 int md_raid5_congested(struct mddev *mddev, int bits)
3789 {
3790 struct r5conf *conf = mddev->private;
3791
3792 /* No difference between reads and writes. Just check
3793 * how busy the stripe_cache is
3794 */
3795
3796 if (conf->inactive_blocked)
3797 return 1;
3798 if (conf->quiesce)
3799 return 1;
3800 if (list_empty_careful(&conf->inactive_list))
3801 return 1;
3802
3803 return 0;
3804 }
3805 EXPORT_SYMBOL_GPL(md_raid5_congested);
3806
3807 static int raid5_congested(void *data, int bits)
3808 {
3809 struct mddev *mddev = data;
3810
3811 return mddev_congested(mddev, bits) ||
3812 md_raid5_congested(mddev, bits);
3813 }
3814
3815 /* We want read requests to align with chunks where possible,
3816 * but write requests don't need to.
3817 */
3818 static int raid5_mergeable_bvec(struct request_queue *q,
3819 struct bvec_merge_data *bvm,
3820 struct bio_vec *biovec)
3821 {
3822 struct mddev *mddev = q->queuedata;
3823 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3824 int max;
3825 unsigned int chunk_sectors = mddev->chunk_sectors;
3826 unsigned int bio_sectors = bvm->bi_size >> 9;
3827
3828 if ((bvm->bi_rw & 1) == WRITE)
3829 return biovec->bv_len; /* always allow writes to be mergeable */
3830
3831 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3832 chunk_sectors = mddev->new_chunk_sectors;
3833 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3834 if (max < 0) max = 0;
3835 if (max <= biovec->bv_len && bio_sectors == 0)
3836 return biovec->bv_len;
3837 else
3838 return max;
3839 }
3840
3841
3842 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3843 {
3844 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3845 unsigned int chunk_sectors = mddev->chunk_sectors;
3846 unsigned int bio_sectors = bio->bi_size >> 9;
3847
3848 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3849 chunk_sectors = mddev->new_chunk_sectors;
3850 return chunk_sectors >=
3851 ((sector & (chunk_sectors - 1)) + bio_sectors);
3852 }
3853
3854 /*
3855 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3856 * later sampled by raid5d.
3857 */
3858 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3859 {
3860 unsigned long flags;
3861
3862 spin_lock_irqsave(&conf->device_lock, flags);
3863
3864 bi->bi_next = conf->retry_read_aligned_list;
3865 conf->retry_read_aligned_list = bi;
3866
3867 spin_unlock_irqrestore(&conf->device_lock, flags);
3868 md_wakeup_thread(conf->mddev->thread);
3869 }
3870
3871
3872 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3873 {
3874 struct bio *bi;
3875
3876 bi = conf->retry_read_aligned;
3877 if (bi) {
3878 conf->retry_read_aligned = NULL;
3879 return bi;
3880 }
3881 bi = conf->retry_read_aligned_list;
3882 if(bi) {
3883 conf->retry_read_aligned_list = bi->bi_next;
3884 bi->bi_next = NULL;
3885 /*
3886 * this sets the active strip count to 1 and the processed
3887 * strip count to zero (upper 8 bits)
3888 */
3889 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
3890 }
3891
3892 return bi;
3893 }
3894
3895
3896 /*
3897 * The "raid5_align_endio" should check if the read succeeded and if it
3898 * did, call bio_endio on the original bio (having bio_put the new bio
3899 * first).
3900 * If the read failed..
3901 */
3902 static void raid5_align_endio(struct bio *bi, int error)
3903 {
3904 struct bio* raid_bi = bi->bi_private;
3905 struct mddev *mddev;
3906 struct r5conf *conf;
3907 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3908 struct md_rdev *rdev;
3909
3910 bio_put(bi);
3911
3912 rdev = (void*)raid_bi->bi_next;
3913 raid_bi->bi_next = NULL;
3914 mddev = rdev->mddev;
3915 conf = mddev->private;
3916
3917 rdev_dec_pending(rdev, conf->mddev);
3918
3919 if (!error && uptodate) {
3920 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
3921 raid_bi, 0);
3922 bio_endio(raid_bi, 0);
3923 if (atomic_dec_and_test(&conf->active_aligned_reads))
3924 wake_up(&conf->wait_for_stripe);
3925 return;
3926 }
3927
3928
3929 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3930
3931 add_bio_to_retry(raid_bi, conf);
3932 }
3933
3934 static int bio_fits_rdev(struct bio *bi)
3935 {
3936 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3937
3938 if ((bi->bi_size>>9) > queue_max_sectors(q))
3939 return 0;
3940 blk_recount_segments(q, bi);
3941 if (bi->bi_phys_segments > queue_max_segments(q))
3942 return 0;
3943
3944 if (q->merge_bvec_fn)
3945 /* it's too hard to apply the merge_bvec_fn at this stage,
3946 * just just give up
3947 */
3948 return 0;
3949
3950 return 1;
3951 }
3952
3953
3954 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3955 {
3956 struct r5conf *conf = mddev->private;
3957 int dd_idx;
3958 struct bio* align_bi;
3959 struct md_rdev *rdev;
3960 sector_t end_sector;
3961
3962 if (!in_chunk_boundary(mddev, raid_bio)) {
3963 pr_debug("chunk_aligned_read : non aligned\n");
3964 return 0;
3965 }
3966 /*
3967 * use bio_clone_mddev to make a copy of the bio
3968 */
3969 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3970 if (!align_bi)
3971 return 0;
3972 /*
3973 * set bi_end_io to a new function, and set bi_private to the
3974 * original bio.
3975 */
3976 align_bi->bi_end_io = raid5_align_endio;
3977 align_bi->bi_private = raid_bio;
3978 /*
3979 * compute position
3980 */
3981 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
3982 0,
3983 &dd_idx, NULL);
3984
3985 end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3986 rcu_read_lock();
3987 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3988 if (!rdev || test_bit(Faulty, &rdev->flags) ||
3989 rdev->recovery_offset < end_sector) {
3990 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3991 if (rdev &&
3992 (test_bit(Faulty, &rdev->flags) ||
3993 !(test_bit(In_sync, &rdev->flags) ||
3994 rdev->recovery_offset >= end_sector)))
3995 rdev = NULL;
3996 }
3997 if (rdev) {
3998 sector_t first_bad;
3999 int bad_sectors;
4000
4001 atomic_inc(&rdev->nr_pending);
4002 rcu_read_unlock();
4003 raid_bio->bi_next = (void*)rdev;
4004 align_bi->bi_bdev = rdev->bdev;
4005 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4006
4007 if (!bio_fits_rdev(align_bi) ||
4008 is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
4009 &first_bad, &bad_sectors)) {
4010 /* too big in some way, or has a known bad block */
4011 bio_put(align_bi);
4012 rdev_dec_pending(rdev, mddev);
4013 return 0;
4014 }
4015
4016 /* No reshape active, so we can trust rdev->data_offset */
4017 align_bi->bi_sector += rdev->data_offset;
4018
4019 spin_lock_irq(&conf->device_lock);
4020 wait_event_lock_irq(conf->wait_for_stripe,
4021 conf->quiesce == 0,
4022 conf->device_lock);
4023 atomic_inc(&conf->active_aligned_reads);
4024 spin_unlock_irq(&conf->device_lock);
4025
4026 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4027 align_bi, disk_devt(mddev->gendisk),
4028 raid_bio->bi_sector);
4029 generic_make_request(align_bi);
4030 return 1;
4031 } else {
4032 rcu_read_unlock();
4033 bio_put(align_bi);
4034 return 0;
4035 }
4036 }
4037
4038 /* __get_priority_stripe - get the next stripe to process
4039 *
4040 * Full stripe writes are allowed to pass preread active stripes up until
4041 * the bypass_threshold is exceeded. In general the bypass_count
4042 * increments when the handle_list is handled before the hold_list; however, it
4043 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4044 * stripe with in flight i/o. The bypass_count will be reset when the
4045 * head of the hold_list has changed, i.e. the head was promoted to the
4046 * handle_list.
4047 */
4048 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
4049 {
4050 struct stripe_head *sh;
4051
4052 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4053 __func__,
4054 list_empty(&conf->handle_list) ? "empty" : "busy",
4055 list_empty(&conf->hold_list) ? "empty" : "busy",
4056 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4057
4058 if (!list_empty(&conf->handle_list)) {
4059 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
4060
4061 if (list_empty(&conf->hold_list))
4062 conf->bypass_count = 0;
4063 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4064 if (conf->hold_list.next == conf->last_hold)
4065 conf->bypass_count++;
4066 else {
4067 conf->last_hold = conf->hold_list.next;
4068 conf->bypass_count -= conf->bypass_threshold;
4069 if (conf->bypass_count < 0)
4070 conf->bypass_count = 0;
4071 }
4072 }
4073 } else if (!list_empty(&conf->hold_list) &&
4074 ((conf->bypass_threshold &&
4075 conf->bypass_count > conf->bypass_threshold) ||
4076 atomic_read(&conf->pending_full_writes) == 0)) {
4077 sh = list_entry(conf->hold_list.next,
4078 typeof(*sh), lru);
4079 conf->bypass_count -= conf->bypass_threshold;
4080 if (conf->bypass_count < 0)
4081 conf->bypass_count = 0;
4082 } else
4083 return NULL;
4084
4085 list_del_init(&sh->lru);
4086 atomic_inc(&sh->count);
4087 BUG_ON(atomic_read(&sh->count) != 1);
4088 return sh;
4089 }
4090
4091 struct raid5_plug_cb {
4092 struct blk_plug_cb cb;
4093 struct list_head list;
4094 };
4095
4096 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4097 {
4098 struct raid5_plug_cb *cb = container_of(
4099 blk_cb, struct raid5_plug_cb, cb);
4100 struct stripe_head *sh;
4101 struct mddev *mddev = cb->cb.data;
4102 struct r5conf *conf = mddev->private;
4103 int cnt = 0;
4104
4105 if (cb->list.next && !list_empty(&cb->list)) {
4106 spin_lock_irq(&conf->device_lock);
4107 while (!list_empty(&cb->list)) {
4108 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4109 list_del_init(&sh->lru);
4110 /*
4111 * avoid race release_stripe_plug() sees
4112 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4113 * is still in our list
4114 */
4115 smp_mb__before_clear_bit();
4116 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4117 __release_stripe(conf, sh);
4118 cnt++;
4119 }
4120 spin_unlock_irq(&conf->device_lock);
4121 }
4122 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4123 kfree(cb);
4124 }
4125
4126 static void release_stripe_plug(struct mddev *mddev,
4127 struct stripe_head *sh)
4128 {
4129 struct blk_plug_cb *blk_cb = blk_check_plugged(
4130 raid5_unplug, mddev,
4131 sizeof(struct raid5_plug_cb));
4132 struct raid5_plug_cb *cb;
4133
4134 if (!blk_cb) {
4135 release_stripe(sh);
4136 return;
4137 }
4138
4139 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4140
4141 if (cb->list.next == NULL)
4142 INIT_LIST_HEAD(&cb->list);
4143
4144 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4145 list_add_tail(&sh->lru, &cb->list);
4146 else
4147 release_stripe(sh);
4148 }
4149
4150 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4151 {
4152 struct r5conf *conf = mddev->private;
4153 sector_t logical_sector, last_sector;
4154 struct stripe_head *sh;
4155 int remaining;
4156 int stripe_sectors;
4157
4158 if (mddev->reshape_position != MaxSector)
4159 /* Skip discard while reshape is happening */
4160 return;
4161
4162 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4163 last_sector = bi->bi_sector + (bi->bi_size>>9);
4164
4165 bi->bi_next = NULL;
4166 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4167
4168 stripe_sectors = conf->chunk_sectors *
4169 (conf->raid_disks - conf->max_degraded);
4170 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4171 stripe_sectors);
4172 sector_div(last_sector, stripe_sectors);
4173
4174 logical_sector *= conf->chunk_sectors;
4175 last_sector *= conf->chunk_sectors;
4176
4177 for (; logical_sector < last_sector;
4178 logical_sector += STRIPE_SECTORS) {
4179 DEFINE_WAIT(w);
4180 int d;
4181 again:
4182 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4183 prepare_to_wait(&conf->wait_for_overlap, &w,
4184 TASK_UNINTERRUPTIBLE);
4185 spin_lock_irq(&sh->stripe_lock);
4186 for (d = 0; d < conf->raid_disks; d++) {
4187 if (d == sh->pd_idx || d == sh->qd_idx)
4188 continue;
4189 if (sh->dev[d].towrite || sh->dev[d].toread) {
4190 set_bit(R5_Overlap, &sh->dev[d].flags);
4191 spin_unlock_irq(&sh->stripe_lock);
4192 release_stripe(sh);
4193 schedule();
4194 goto again;
4195 }
4196 }
4197 finish_wait(&conf->wait_for_overlap, &w);
4198 for (d = 0; d < conf->raid_disks; d++) {
4199 if (d == sh->pd_idx || d == sh->qd_idx)
4200 continue;
4201 sh->dev[d].towrite = bi;
4202 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4203 raid5_inc_bi_active_stripes(bi);
4204 }
4205 spin_unlock_irq(&sh->stripe_lock);
4206 if (conf->mddev->bitmap) {
4207 for (d = 0;
4208 d < conf->raid_disks - conf->max_degraded;
4209 d++)
4210 bitmap_startwrite(mddev->bitmap,
4211 sh->sector,
4212 STRIPE_SECTORS,
4213 0);
4214 sh->bm_seq = conf->seq_flush + 1;
4215 set_bit(STRIPE_BIT_DELAY, &sh->state);
4216 }
4217
4218 set_bit(STRIPE_HANDLE, &sh->state);
4219 clear_bit(STRIPE_DELAYED, &sh->state);
4220 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4221 atomic_inc(&conf->preread_active_stripes);
4222 release_stripe_plug(mddev, sh);
4223 }
4224
4225 remaining = raid5_dec_bi_active_stripes(bi);
4226 if (remaining == 0) {
4227 md_write_end(mddev);
4228 bio_endio(bi, 0);
4229 }
4230 }
4231
4232 static void make_request(struct mddev *mddev, struct bio * bi)
4233 {
4234 struct r5conf *conf = mddev->private;
4235 int dd_idx;
4236 sector_t new_sector;
4237 sector_t logical_sector, last_sector;
4238 struct stripe_head *sh;
4239 const int rw = bio_data_dir(bi);
4240 int remaining;
4241
4242 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4243 md_flush_request(mddev, bi);
4244 return;
4245 }
4246
4247 md_write_start(mddev, bi);
4248
4249 if (rw == READ &&
4250 mddev->reshape_position == MaxSector &&
4251 chunk_aligned_read(mddev,bi))
4252 return;
4253
4254 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4255 make_discard_request(mddev, bi);
4256 return;
4257 }
4258
4259 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4260 last_sector = bi->bi_sector + (bi->bi_size>>9);
4261 bi->bi_next = NULL;
4262 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4263
4264 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4265 DEFINE_WAIT(w);
4266 int previous;
4267
4268 retry:
4269 previous = 0;
4270 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4271 if (unlikely(conf->reshape_progress != MaxSector)) {
4272 /* spinlock is needed as reshape_progress may be
4273 * 64bit on a 32bit platform, and so it might be
4274 * possible to see a half-updated value
4275 * Of course reshape_progress could change after
4276 * the lock is dropped, so once we get a reference
4277 * to the stripe that we think it is, we will have
4278 * to check again.
4279 */
4280 spin_lock_irq(&conf->device_lock);
4281 if (mddev->reshape_backwards
4282 ? logical_sector < conf->reshape_progress
4283 : logical_sector >= conf->reshape_progress) {
4284 previous = 1;
4285 } else {
4286 if (mddev->reshape_backwards
4287 ? logical_sector < conf->reshape_safe
4288 : logical_sector >= conf->reshape_safe) {
4289 spin_unlock_irq(&conf->device_lock);
4290 schedule();
4291 goto retry;
4292 }
4293 }
4294 spin_unlock_irq(&conf->device_lock);
4295 }
4296
4297 new_sector = raid5_compute_sector(conf, logical_sector,
4298 previous,
4299 &dd_idx, NULL);
4300 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4301 (unsigned long long)new_sector,
4302 (unsigned long long)logical_sector);
4303
4304 sh = get_active_stripe(conf, new_sector, previous,
4305 (bi->bi_rw&RWA_MASK), 0);
4306 if (sh) {
4307 if (unlikely(previous)) {
4308 /* expansion might have moved on while waiting for a
4309 * stripe, so we must do the range check again.
4310 * Expansion could still move past after this
4311 * test, but as we are holding a reference to
4312 * 'sh', we know that if that happens,
4313 * STRIPE_EXPANDING will get set and the expansion
4314 * won't proceed until we finish with the stripe.
4315 */
4316 int must_retry = 0;
4317 spin_lock_irq(&conf->device_lock);
4318 if (mddev->reshape_backwards
4319 ? logical_sector >= conf->reshape_progress
4320 : logical_sector < conf->reshape_progress)
4321 /* mismatch, need to try again */
4322 must_retry = 1;
4323 spin_unlock_irq(&conf->device_lock);
4324 if (must_retry) {
4325 release_stripe(sh);
4326 schedule();
4327 goto retry;
4328 }
4329 }
4330
4331 if (rw == WRITE &&
4332 logical_sector >= mddev->suspend_lo &&
4333 logical_sector < mddev->suspend_hi) {
4334 release_stripe(sh);
4335 /* As the suspend_* range is controlled by
4336 * userspace, we want an interruptible
4337 * wait.
4338 */
4339 flush_signals(current);
4340 prepare_to_wait(&conf->wait_for_overlap,
4341 &w, TASK_INTERRUPTIBLE);
4342 if (logical_sector >= mddev->suspend_lo &&
4343 logical_sector < mddev->suspend_hi)
4344 schedule();
4345 goto retry;
4346 }
4347
4348 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4349 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4350 /* Stripe is busy expanding or
4351 * add failed due to overlap. Flush everything
4352 * and wait a while
4353 */
4354 md_wakeup_thread(mddev->thread);
4355 release_stripe(sh);
4356 schedule();
4357 goto retry;
4358 }
4359 finish_wait(&conf->wait_for_overlap, &w);
4360 set_bit(STRIPE_HANDLE, &sh->state);
4361 clear_bit(STRIPE_DELAYED, &sh->state);
4362 if ((bi->bi_rw & REQ_SYNC) &&
4363 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4364 atomic_inc(&conf->preread_active_stripes);
4365 release_stripe_plug(mddev, sh);
4366 } else {
4367 /* cannot get stripe for read-ahead, just give-up */
4368 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4369 finish_wait(&conf->wait_for_overlap, &w);
4370 break;
4371 }
4372 }
4373
4374 remaining = raid5_dec_bi_active_stripes(bi);
4375 if (remaining == 0) {
4376
4377 if ( rw == WRITE )
4378 md_write_end(mddev);
4379
4380 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4381 bi, 0);
4382 bio_endio(bi, 0);
4383 }
4384 }
4385
4386 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4387
4388 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4389 {
4390 /* reshaping is quite different to recovery/resync so it is
4391 * handled quite separately ... here.
4392 *
4393 * On each call to sync_request, we gather one chunk worth of
4394 * destination stripes and flag them as expanding.
4395 * Then we find all the source stripes and request reads.
4396 * As the reads complete, handle_stripe will copy the data
4397 * into the destination stripe and release that stripe.
4398 */
4399 struct r5conf *conf = mddev->private;
4400 struct stripe_head *sh;
4401 sector_t first_sector, last_sector;
4402 int raid_disks = conf->previous_raid_disks;
4403 int data_disks = raid_disks - conf->max_degraded;
4404 int new_data_disks = conf->raid_disks - conf->max_degraded;
4405 int i;
4406 int dd_idx;
4407 sector_t writepos, readpos, safepos;
4408 sector_t stripe_addr;
4409 int reshape_sectors;
4410 struct list_head stripes;
4411
4412 if (sector_nr == 0) {
4413 /* If restarting in the middle, skip the initial sectors */
4414 if (mddev->reshape_backwards &&
4415 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4416 sector_nr = raid5_size(mddev, 0, 0)
4417 - conf->reshape_progress;
4418 } else if (!mddev->reshape_backwards &&
4419 conf->reshape_progress > 0)
4420 sector_nr = conf->reshape_progress;
4421 sector_div(sector_nr, new_data_disks);
4422 if (sector_nr) {
4423 mddev->curr_resync_completed = sector_nr;
4424 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4425 *skipped = 1;
4426 return sector_nr;
4427 }
4428 }
4429
4430 /* We need to process a full chunk at a time.
4431 * If old and new chunk sizes differ, we need to process the
4432 * largest of these
4433 */
4434 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4435 reshape_sectors = mddev->new_chunk_sectors;
4436 else
4437 reshape_sectors = mddev->chunk_sectors;
4438
4439 /* We update the metadata at least every 10 seconds, or when
4440 * the data about to be copied would over-write the source of
4441 * the data at the front of the range. i.e. one new_stripe
4442 * along from reshape_progress new_maps to after where
4443 * reshape_safe old_maps to
4444 */
4445 writepos = conf->reshape_progress;
4446 sector_div(writepos, new_data_disks);
4447 readpos = conf->reshape_progress;
4448 sector_div(readpos, data_disks);
4449 safepos = conf->reshape_safe;
4450 sector_div(safepos, data_disks);
4451 if (mddev->reshape_backwards) {
4452 writepos -= min_t(sector_t, reshape_sectors, writepos);
4453 readpos += reshape_sectors;
4454 safepos += reshape_sectors;
4455 } else {
4456 writepos += reshape_sectors;
4457 readpos -= min_t(sector_t, reshape_sectors, readpos);
4458 safepos -= min_t(sector_t, reshape_sectors, safepos);
4459 }
4460
4461 /* Having calculated the 'writepos' possibly use it
4462 * to set 'stripe_addr' which is where we will write to.
4463 */
4464 if (mddev->reshape_backwards) {
4465 BUG_ON(conf->reshape_progress == 0);
4466 stripe_addr = writepos;
4467 BUG_ON((mddev->dev_sectors &
4468 ~((sector_t)reshape_sectors - 1))
4469 - reshape_sectors - stripe_addr
4470 != sector_nr);
4471 } else {
4472 BUG_ON(writepos != sector_nr + reshape_sectors);
4473 stripe_addr = sector_nr;
4474 }
4475
4476 /* 'writepos' is the most advanced device address we might write.
4477 * 'readpos' is the least advanced device address we might read.
4478 * 'safepos' is the least address recorded in the metadata as having
4479 * been reshaped.
4480 * If there is a min_offset_diff, these are adjusted either by
4481 * increasing the safepos/readpos if diff is negative, or
4482 * increasing writepos if diff is positive.
4483 * If 'readpos' is then behind 'writepos', there is no way that we can
4484 * ensure safety in the face of a crash - that must be done by userspace
4485 * making a backup of the data. So in that case there is no particular
4486 * rush to update metadata.
4487 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4488 * update the metadata to advance 'safepos' to match 'readpos' so that
4489 * we can be safe in the event of a crash.
4490 * So we insist on updating metadata if safepos is behind writepos and
4491 * readpos is beyond writepos.
4492 * In any case, update the metadata every 10 seconds.
4493 * Maybe that number should be configurable, but I'm not sure it is
4494 * worth it.... maybe it could be a multiple of safemode_delay???
4495 */
4496 if (conf->min_offset_diff < 0) {
4497 safepos += -conf->min_offset_diff;
4498 readpos += -conf->min_offset_diff;
4499 } else
4500 writepos += conf->min_offset_diff;
4501
4502 if ((mddev->reshape_backwards
4503 ? (safepos > writepos && readpos < writepos)
4504 : (safepos < writepos && readpos > writepos)) ||
4505 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4506 /* Cannot proceed until we've updated the superblock... */
4507 wait_event(conf->wait_for_overlap,
4508 atomic_read(&conf->reshape_stripes)==0);
4509 mddev->reshape_position = conf->reshape_progress;
4510 mddev->curr_resync_completed = sector_nr;
4511 conf->reshape_checkpoint = jiffies;
4512 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4513 md_wakeup_thread(mddev->thread);
4514 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4515 kthread_should_stop());
4516 spin_lock_irq(&conf->device_lock);
4517 conf->reshape_safe = mddev->reshape_position;
4518 spin_unlock_irq(&conf->device_lock);
4519 wake_up(&conf->wait_for_overlap);
4520 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4521 }
4522
4523 INIT_LIST_HEAD(&stripes);
4524 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4525 int j;
4526 int skipped_disk = 0;
4527 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4528 set_bit(STRIPE_EXPANDING, &sh->state);
4529 atomic_inc(&conf->reshape_stripes);
4530 /* If any of this stripe is beyond the end of the old
4531 * array, then we need to zero those blocks
4532 */
4533 for (j=sh->disks; j--;) {
4534 sector_t s;
4535 if (j == sh->pd_idx)
4536 continue;
4537 if (conf->level == 6 &&
4538 j == sh->qd_idx)
4539 continue;
4540 s = compute_blocknr(sh, j, 0);
4541 if (s < raid5_size(mddev, 0, 0)) {
4542 skipped_disk = 1;
4543 continue;
4544 }
4545 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4546 set_bit(R5_Expanded, &sh->dev[j].flags);
4547 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4548 }
4549 if (!skipped_disk) {
4550 set_bit(STRIPE_EXPAND_READY, &sh->state);
4551 set_bit(STRIPE_HANDLE, &sh->state);
4552 }
4553 list_add(&sh->lru, &stripes);
4554 }
4555 spin_lock_irq(&conf->device_lock);
4556 if (mddev->reshape_backwards)
4557 conf->reshape_progress -= reshape_sectors * new_data_disks;
4558 else
4559 conf->reshape_progress += reshape_sectors * new_data_disks;
4560 spin_unlock_irq(&conf->device_lock);
4561 /* Ok, those stripe are ready. We can start scheduling
4562 * reads on the source stripes.
4563 * The source stripes are determined by mapping the first and last
4564 * block on the destination stripes.
4565 */
4566 first_sector =
4567 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4568 1, &dd_idx, NULL);
4569 last_sector =
4570 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4571 * new_data_disks - 1),
4572 1, &dd_idx, NULL);
4573 if (last_sector >= mddev->dev_sectors)
4574 last_sector = mddev->dev_sectors - 1;
4575 while (first_sector <= last_sector) {
4576 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4577 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4578 set_bit(STRIPE_HANDLE, &sh->state);
4579 release_stripe(sh);
4580 first_sector += STRIPE_SECTORS;
4581 }
4582 /* Now that the sources are clearly marked, we can release
4583 * the destination stripes
4584 */
4585 while (!list_empty(&stripes)) {
4586 sh = list_entry(stripes.next, struct stripe_head, lru);
4587 list_del_init(&sh->lru);
4588 release_stripe(sh);
4589 }
4590 /* If this takes us to the resync_max point where we have to pause,
4591 * then we need to write out the superblock.
4592 */
4593 sector_nr += reshape_sectors;
4594 if ((sector_nr - mddev->curr_resync_completed) * 2
4595 >= mddev->resync_max - mddev->curr_resync_completed) {
4596 /* Cannot proceed until we've updated the superblock... */
4597 wait_event(conf->wait_for_overlap,
4598 atomic_read(&conf->reshape_stripes) == 0);
4599 mddev->reshape_position = conf->reshape_progress;
4600 mddev->curr_resync_completed = sector_nr;
4601 conf->reshape_checkpoint = jiffies;
4602 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4603 md_wakeup_thread(mddev->thread);
4604 wait_event(mddev->sb_wait,
4605 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4606 || kthread_should_stop());
4607 spin_lock_irq(&conf->device_lock);
4608 conf->reshape_safe = mddev->reshape_position;
4609 spin_unlock_irq(&conf->device_lock);
4610 wake_up(&conf->wait_for_overlap);
4611 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4612 }
4613 return reshape_sectors;
4614 }
4615
4616 /* FIXME go_faster isn't used */
4617 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4618 {
4619 struct r5conf *conf = mddev->private;
4620 struct stripe_head *sh;
4621 sector_t max_sector = mddev->dev_sectors;
4622 sector_t sync_blocks;
4623 int still_degraded = 0;
4624 int i;
4625
4626 if (sector_nr >= max_sector) {
4627 /* just being told to finish up .. nothing much to do */
4628
4629 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4630 end_reshape(conf);
4631 return 0;
4632 }
4633
4634 if (mddev->curr_resync < max_sector) /* aborted */
4635 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4636 &sync_blocks, 1);
4637 else /* completed sync */
4638 conf->fullsync = 0;
4639 bitmap_close_sync(mddev->bitmap);
4640
4641 return 0;
4642 }
4643
4644 /* Allow raid5_quiesce to complete */
4645 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4646
4647 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4648 return reshape_request(mddev, sector_nr, skipped);
4649
4650 /* No need to check resync_max as we never do more than one
4651 * stripe, and as resync_max will always be on a chunk boundary,
4652 * if the check in md_do_sync didn't fire, there is no chance
4653 * of overstepping resync_max here
4654 */
4655
4656 /* if there is too many failed drives and we are trying
4657 * to resync, then assert that we are finished, because there is
4658 * nothing we can do.
4659 */
4660 if (mddev->degraded >= conf->max_degraded &&
4661 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4662 sector_t rv = mddev->dev_sectors - sector_nr;
4663 *skipped = 1;
4664 return rv;
4665 }
4666 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4667 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4668 !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4669 /* we can skip this block, and probably more */
4670 sync_blocks /= STRIPE_SECTORS;
4671 *skipped = 1;
4672 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4673 }
4674
4675 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4676
4677 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4678 if (sh == NULL) {
4679 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4680 /* make sure we don't swamp the stripe cache if someone else
4681 * is trying to get access
4682 */
4683 schedule_timeout_uninterruptible(1);
4684 }
4685 /* Need to check if array will still be degraded after recovery/resync
4686 * We don't need to check the 'failed' flag as when that gets set,
4687 * recovery aborts.
4688 */
4689 for (i = 0; i < conf->raid_disks; i++)
4690 if (conf->disks[i].rdev == NULL)
4691 still_degraded = 1;
4692
4693 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4694
4695 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4696
4697 handle_stripe(sh);
4698 release_stripe(sh);
4699
4700 return STRIPE_SECTORS;
4701 }
4702
4703 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4704 {
4705 /* We may not be able to submit a whole bio at once as there
4706 * may not be enough stripe_heads available.
4707 * We cannot pre-allocate enough stripe_heads as we may need
4708 * more than exist in the cache (if we allow ever large chunks).
4709 * So we do one stripe head at a time and record in
4710 * ->bi_hw_segments how many have been done.
4711 *
4712 * We *know* that this entire raid_bio is in one chunk, so
4713 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4714 */
4715 struct stripe_head *sh;
4716 int dd_idx;
4717 sector_t sector, logical_sector, last_sector;
4718 int scnt = 0;
4719 int remaining;
4720 int handled = 0;
4721
4722 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4723 sector = raid5_compute_sector(conf, logical_sector,
4724 0, &dd_idx, NULL);
4725 last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4726
4727 for (; logical_sector < last_sector;
4728 logical_sector += STRIPE_SECTORS,
4729 sector += STRIPE_SECTORS,
4730 scnt++) {
4731
4732 if (scnt < raid5_bi_processed_stripes(raid_bio))
4733 /* already done this stripe */
4734 continue;
4735
4736 sh = get_active_stripe(conf, sector, 0, 1, 0);
4737
4738 if (!sh) {
4739 /* failed to get a stripe - must wait */
4740 raid5_set_bi_processed_stripes(raid_bio, scnt);
4741 conf->retry_read_aligned = raid_bio;
4742 return handled;
4743 }
4744
4745 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4746 release_stripe(sh);
4747 raid5_set_bi_processed_stripes(raid_bio, scnt);
4748 conf->retry_read_aligned = raid_bio;
4749 return handled;
4750 }
4751
4752 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4753 handle_stripe(sh);
4754 release_stripe(sh);
4755 handled++;
4756 }
4757 remaining = raid5_dec_bi_active_stripes(raid_bio);
4758 if (remaining == 0) {
4759 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
4760 raid_bio, 0);
4761 bio_endio(raid_bio, 0);
4762 }
4763 if (atomic_dec_and_test(&conf->active_aligned_reads))
4764 wake_up(&conf->wait_for_stripe);
4765 return handled;
4766 }
4767
4768 #define MAX_STRIPE_BATCH 8
4769 static int handle_active_stripes(struct r5conf *conf)
4770 {
4771 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4772 int i, batch_size = 0;
4773
4774 while (batch_size < MAX_STRIPE_BATCH &&
4775 (sh = __get_priority_stripe(conf)) != NULL)
4776 batch[batch_size++] = sh;
4777
4778 if (batch_size == 0)
4779 return batch_size;
4780 spin_unlock_irq(&conf->device_lock);
4781
4782 for (i = 0; i < batch_size; i++)
4783 handle_stripe(batch[i]);
4784
4785 cond_resched();
4786
4787 spin_lock_irq(&conf->device_lock);
4788 for (i = 0; i < batch_size; i++)
4789 __release_stripe(conf, batch[i]);
4790 return batch_size;
4791 }
4792
4793 /*
4794 * This is our raid5 kernel thread.
4795 *
4796 * We scan the hash table for stripes which can be handled now.
4797 * During the scan, completed stripes are saved for us by the interrupt
4798 * handler, so that they will not have to wait for our next wakeup.
4799 */
4800 static void raid5d(struct md_thread *thread)
4801 {
4802 struct mddev *mddev = thread->mddev;
4803 struct r5conf *conf = mddev->private;
4804 int handled;
4805 struct blk_plug plug;
4806
4807 pr_debug("+++ raid5d active\n");
4808
4809 md_check_recovery(mddev);
4810
4811 blk_start_plug(&plug);
4812 handled = 0;
4813 spin_lock_irq(&conf->device_lock);
4814 while (1) {
4815 struct bio *bio;
4816 int batch_size;
4817
4818 if (
4819 !list_empty(&conf->bitmap_list)) {
4820 /* Now is a good time to flush some bitmap updates */
4821 conf->seq_flush++;
4822 spin_unlock_irq(&conf->device_lock);
4823 bitmap_unplug(mddev->bitmap);
4824 spin_lock_irq(&conf->device_lock);
4825 conf->seq_write = conf->seq_flush;
4826 activate_bit_delay(conf);
4827 }
4828 raid5_activate_delayed(conf);
4829
4830 while ((bio = remove_bio_from_retry(conf))) {
4831 int ok;
4832 spin_unlock_irq(&conf->device_lock);
4833 ok = retry_aligned_read(conf, bio);
4834 spin_lock_irq(&conf->device_lock);
4835 if (!ok)
4836 break;
4837 handled++;
4838 }
4839
4840 batch_size = handle_active_stripes(conf);
4841 if (!batch_size)
4842 break;
4843 handled += batch_size;
4844
4845 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
4846 spin_unlock_irq(&conf->device_lock);
4847 md_check_recovery(mddev);
4848 spin_lock_irq(&conf->device_lock);
4849 }
4850 }
4851 pr_debug("%d stripes handled\n", handled);
4852
4853 spin_unlock_irq(&conf->device_lock);
4854
4855 async_tx_issue_pending_all();
4856 blk_finish_plug(&plug);
4857
4858 pr_debug("--- raid5d inactive\n");
4859 }
4860
4861 static ssize_t
4862 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4863 {
4864 struct r5conf *conf = mddev->private;
4865 if (conf)
4866 return sprintf(page, "%d\n", conf->max_nr_stripes);
4867 else
4868 return 0;
4869 }
4870
4871 int
4872 raid5_set_cache_size(struct mddev *mddev, int size)
4873 {
4874 struct r5conf *conf = mddev->private;
4875 int err;
4876
4877 if (size <= 16 || size > 32768)
4878 return -EINVAL;
4879 while (size < conf->max_nr_stripes) {
4880 if (drop_one_stripe(conf))
4881 conf->max_nr_stripes--;
4882 else
4883 break;
4884 }
4885 err = md_allow_write(mddev);
4886 if (err)
4887 return err;
4888 while (size > conf->max_nr_stripes) {
4889 if (grow_one_stripe(conf))
4890 conf->max_nr_stripes++;
4891 else break;
4892 }
4893 return 0;
4894 }
4895 EXPORT_SYMBOL(raid5_set_cache_size);
4896
4897 static ssize_t
4898 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4899 {
4900 struct r5conf *conf = mddev->private;
4901 unsigned long new;
4902 int err;
4903
4904 if (len >= PAGE_SIZE)
4905 return -EINVAL;
4906 if (!conf)
4907 return -ENODEV;
4908
4909 if (strict_strtoul(page, 10, &new))
4910 return -EINVAL;
4911 err = raid5_set_cache_size(mddev, new);
4912 if (err)
4913 return err;
4914 return len;
4915 }
4916
4917 static struct md_sysfs_entry
4918 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4919 raid5_show_stripe_cache_size,
4920 raid5_store_stripe_cache_size);
4921
4922 static ssize_t
4923 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4924 {
4925 struct r5conf *conf = mddev->private;
4926 if (conf)
4927 return sprintf(page, "%d\n", conf->bypass_threshold);
4928 else
4929 return 0;
4930 }
4931
4932 static ssize_t
4933 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4934 {
4935 struct r5conf *conf = mddev->private;
4936 unsigned long new;
4937 if (len >= PAGE_SIZE)
4938 return -EINVAL;
4939 if (!conf)
4940 return -ENODEV;
4941
4942 if (strict_strtoul(page, 10, &new))
4943 return -EINVAL;
4944 if (new > conf->max_nr_stripes)
4945 return -EINVAL;
4946 conf->bypass_threshold = new;
4947 return len;
4948 }
4949
4950 static struct md_sysfs_entry
4951 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4952 S_IRUGO | S_IWUSR,
4953 raid5_show_preread_threshold,
4954 raid5_store_preread_threshold);
4955
4956 static ssize_t
4957 stripe_cache_active_show(struct mddev *mddev, char *page)
4958 {
4959 struct r5conf *conf = mddev->private;
4960 if (conf)
4961 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4962 else
4963 return 0;
4964 }
4965
4966 static struct md_sysfs_entry
4967 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4968
4969 static struct attribute *raid5_attrs[] = {
4970 &raid5_stripecache_size.attr,
4971 &raid5_stripecache_active.attr,
4972 &raid5_preread_bypass_threshold.attr,
4973 NULL,
4974 };
4975 static struct attribute_group raid5_attrs_group = {
4976 .name = NULL,
4977 .attrs = raid5_attrs,
4978 };
4979
4980 static sector_t
4981 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4982 {
4983 struct r5conf *conf = mddev->private;
4984
4985 if (!sectors)
4986 sectors = mddev->dev_sectors;
4987 if (!raid_disks)
4988 /* size is defined by the smallest of previous and new size */
4989 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4990
4991 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4992 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4993 return sectors * (raid_disks - conf->max_degraded);
4994 }
4995
4996 static void raid5_free_percpu(struct r5conf *conf)
4997 {
4998 struct raid5_percpu *percpu;
4999 unsigned long cpu;
5000
5001 if (!conf->percpu)
5002 return;
5003
5004 get_online_cpus();
5005 for_each_possible_cpu(cpu) {
5006 percpu = per_cpu_ptr(conf->percpu, cpu);
5007 safe_put_page(percpu->spare_page);
5008 kfree(percpu->scribble);
5009 }
5010 #ifdef CONFIG_HOTPLUG_CPU
5011 unregister_cpu_notifier(&conf->cpu_notify);
5012 #endif
5013 put_online_cpus();
5014
5015 free_percpu(conf->percpu);
5016 }
5017
5018 static void free_conf(struct r5conf *conf)
5019 {
5020 shrink_stripes(conf);
5021 raid5_free_percpu(conf);
5022 kfree(conf->disks);
5023 kfree(conf->stripe_hashtbl);
5024 kfree(conf);
5025 }
5026
5027 #ifdef CONFIG_HOTPLUG_CPU
5028 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5029 void *hcpu)
5030 {
5031 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5032 long cpu = (long)hcpu;
5033 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5034
5035 switch (action) {
5036 case CPU_UP_PREPARE:
5037 case CPU_UP_PREPARE_FROZEN:
5038 if (conf->level == 6 && !percpu->spare_page)
5039 percpu->spare_page = alloc_page(GFP_KERNEL);
5040 if (!percpu->scribble)
5041 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5042
5043 if (!percpu->scribble ||
5044 (conf->level == 6 && !percpu->spare_page)) {
5045 safe_put_page(percpu->spare_page);
5046 kfree(percpu->scribble);
5047 pr_err("%s: failed memory allocation for cpu%ld\n",
5048 __func__, cpu);
5049 return notifier_from_errno(-ENOMEM);
5050 }
5051 break;
5052 case CPU_DEAD:
5053 case CPU_DEAD_FROZEN:
5054 safe_put_page(percpu->spare_page);
5055 kfree(percpu->scribble);
5056 percpu->spare_page = NULL;
5057 percpu->scribble = NULL;
5058 break;
5059 default:
5060 break;
5061 }
5062 return NOTIFY_OK;
5063 }
5064 #endif
5065
5066 static int raid5_alloc_percpu(struct r5conf *conf)
5067 {
5068 unsigned long cpu;
5069 struct page *spare_page;
5070 struct raid5_percpu __percpu *allcpus;
5071 void *scribble;
5072 int err;
5073
5074 allcpus = alloc_percpu(struct raid5_percpu);
5075 if (!allcpus)
5076 return -ENOMEM;
5077 conf->percpu = allcpus;
5078
5079 get_online_cpus();
5080 err = 0;
5081 for_each_present_cpu(cpu) {
5082 if (conf->level == 6) {
5083 spare_page = alloc_page(GFP_KERNEL);
5084 if (!spare_page) {
5085 err = -ENOMEM;
5086 break;
5087 }
5088 per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5089 }
5090 scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5091 if (!scribble) {
5092 err = -ENOMEM;
5093 break;
5094 }
5095 per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
5096 }
5097 #ifdef CONFIG_HOTPLUG_CPU
5098 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5099 conf->cpu_notify.priority = 0;
5100 if (err == 0)
5101 err = register_cpu_notifier(&conf->cpu_notify);
5102 #endif
5103 put_online_cpus();
5104
5105 return err;
5106 }
5107
5108 static struct r5conf *setup_conf(struct mddev *mddev)
5109 {
5110 struct r5conf *conf;
5111 int raid_disk, memory, max_disks;
5112 struct md_rdev *rdev;
5113 struct disk_info *disk;
5114 char pers_name[6];
5115
5116 if (mddev->new_level != 5
5117 && mddev->new_level != 4
5118 && mddev->new_level != 6) {
5119 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5120 mdname(mddev), mddev->new_level);
5121 return ERR_PTR(-EIO);
5122 }
5123 if ((mddev->new_level == 5
5124 && !algorithm_valid_raid5(mddev->new_layout)) ||
5125 (mddev->new_level == 6
5126 && !algorithm_valid_raid6(mddev->new_layout))) {
5127 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5128 mdname(mddev), mddev->new_layout);
5129 return ERR_PTR(-EIO);
5130 }
5131 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5132 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5133 mdname(mddev), mddev->raid_disks);
5134 return ERR_PTR(-EINVAL);
5135 }
5136
5137 if (!mddev->new_chunk_sectors ||
5138 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5139 !is_power_of_2(mddev->new_chunk_sectors)) {
5140 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5141 mdname(mddev), mddev->new_chunk_sectors << 9);
5142 return ERR_PTR(-EINVAL);
5143 }
5144
5145 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5146 if (conf == NULL)
5147 goto abort;
5148 spin_lock_init(&conf->device_lock);
5149 init_waitqueue_head(&conf->wait_for_stripe);
5150 init_waitqueue_head(&conf->wait_for_overlap);
5151 INIT_LIST_HEAD(&conf->handle_list);
5152 INIT_LIST_HEAD(&conf->hold_list);
5153 INIT_LIST_HEAD(&conf->delayed_list);
5154 INIT_LIST_HEAD(&conf->bitmap_list);
5155 INIT_LIST_HEAD(&conf->inactive_list);
5156 atomic_set(&conf->active_stripes, 0);
5157 atomic_set(&conf->preread_active_stripes, 0);
5158 atomic_set(&conf->active_aligned_reads, 0);
5159 conf->bypass_threshold = BYPASS_THRESHOLD;
5160 conf->recovery_disabled = mddev->recovery_disabled - 1;
5161
5162 conf->raid_disks = mddev->raid_disks;
5163 if (mddev->reshape_position == MaxSector)
5164 conf->previous_raid_disks = mddev->raid_disks;
5165 else
5166 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5167 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5168 conf->scribble_len = scribble_len(max_disks);
5169
5170 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5171 GFP_KERNEL);
5172 if (!conf->disks)
5173 goto abort;
5174
5175 conf->mddev = mddev;
5176
5177 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5178 goto abort;
5179
5180 conf->level = mddev->new_level;
5181 if (raid5_alloc_percpu(conf) != 0)
5182 goto abort;
5183
5184 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5185
5186 rdev_for_each(rdev, mddev) {
5187 raid_disk = rdev->raid_disk;
5188 if (raid_disk >= max_disks
5189 || raid_disk < 0)
5190 continue;
5191 disk = conf->disks + raid_disk;
5192
5193 if (test_bit(Replacement, &rdev->flags)) {
5194 if (disk->replacement)
5195 goto abort;
5196 disk->replacement = rdev;
5197 } else {
5198 if (disk->rdev)
5199 goto abort;
5200 disk->rdev = rdev;
5201 }
5202
5203 if (test_bit(In_sync, &rdev->flags)) {
5204 char b[BDEVNAME_SIZE];
5205 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5206 " disk %d\n",
5207 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5208 } else if (rdev->saved_raid_disk != raid_disk)
5209 /* Cannot rely on bitmap to complete recovery */
5210 conf->fullsync = 1;
5211 }
5212
5213 conf->chunk_sectors = mddev->new_chunk_sectors;
5214 conf->level = mddev->new_level;
5215 if (conf->level == 6)
5216 conf->max_degraded = 2;
5217 else
5218 conf->max_degraded = 1;
5219 conf->algorithm = mddev->new_layout;
5220 conf->max_nr_stripes = NR_STRIPES;
5221 conf->reshape_progress = mddev->reshape_position;
5222 if (conf->reshape_progress != MaxSector) {
5223 conf->prev_chunk_sectors = mddev->chunk_sectors;
5224 conf->prev_algo = mddev->layout;
5225 }
5226
5227 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5228 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5229 if (grow_stripes(conf, conf->max_nr_stripes)) {
5230 printk(KERN_ERR
5231 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5232 mdname(mddev), memory);
5233 goto abort;
5234 } else
5235 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5236 mdname(mddev), memory);
5237
5238 sprintf(pers_name, "raid%d", mddev->new_level);
5239 conf->thread = md_register_thread(raid5d, mddev, pers_name);
5240 if (!conf->thread) {
5241 printk(KERN_ERR
5242 "md/raid:%s: couldn't allocate thread.\n",
5243 mdname(mddev));
5244 goto abort;
5245 }
5246
5247 return conf;
5248
5249 abort:
5250 if (conf) {
5251 free_conf(conf);
5252 return ERR_PTR(-EIO);
5253 } else
5254 return ERR_PTR(-ENOMEM);
5255 }
5256
5257
5258 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5259 {
5260 switch (algo) {
5261 case ALGORITHM_PARITY_0:
5262 if (raid_disk < max_degraded)
5263 return 1;
5264 break;
5265 case ALGORITHM_PARITY_N:
5266 if (raid_disk >= raid_disks - max_degraded)
5267 return 1;
5268 break;
5269 case ALGORITHM_PARITY_0_6:
5270 if (raid_disk == 0 ||
5271 raid_disk == raid_disks - 1)
5272 return 1;
5273 break;
5274 case ALGORITHM_LEFT_ASYMMETRIC_6:
5275 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5276 case ALGORITHM_LEFT_SYMMETRIC_6:
5277 case ALGORITHM_RIGHT_SYMMETRIC_6:
5278 if (raid_disk == raid_disks - 1)
5279 return 1;
5280 }
5281 return 0;
5282 }
5283
5284 static int run(struct mddev *mddev)
5285 {
5286 struct r5conf *conf;
5287 int working_disks = 0;
5288 int dirty_parity_disks = 0;
5289 struct md_rdev *rdev;
5290 sector_t reshape_offset = 0;
5291 int i;
5292 long long min_offset_diff = 0;
5293 int first = 1;
5294
5295 if (mddev->recovery_cp != MaxSector)
5296 printk(KERN_NOTICE "md/raid:%s: not clean"
5297 " -- starting background reconstruction\n",
5298 mdname(mddev));
5299
5300 rdev_for_each(rdev, mddev) {
5301 long long diff;
5302 if (rdev->raid_disk < 0)
5303 continue;
5304 diff = (rdev->new_data_offset - rdev->data_offset);
5305 if (first) {
5306 min_offset_diff = diff;
5307 first = 0;
5308 } else if (mddev->reshape_backwards &&
5309 diff < min_offset_diff)
5310 min_offset_diff = diff;
5311 else if (!mddev->reshape_backwards &&
5312 diff > min_offset_diff)
5313 min_offset_diff = diff;
5314 }
5315
5316 if (mddev->reshape_position != MaxSector) {
5317 /* Check that we can continue the reshape.
5318 * Difficulties arise if the stripe we would write to
5319 * next is at or after the stripe we would read from next.
5320 * For a reshape that changes the number of devices, this
5321 * is only possible for a very short time, and mdadm makes
5322 * sure that time appears to have past before assembling
5323 * the array. So we fail if that time hasn't passed.
5324 * For a reshape that keeps the number of devices the same
5325 * mdadm must be monitoring the reshape can keeping the
5326 * critical areas read-only and backed up. It will start
5327 * the array in read-only mode, so we check for that.
5328 */
5329 sector_t here_new, here_old;
5330 int old_disks;
5331 int max_degraded = (mddev->level == 6 ? 2 : 1);
5332
5333 if (mddev->new_level != mddev->level) {
5334 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5335 "required - aborting.\n",
5336 mdname(mddev));
5337 return -EINVAL;
5338 }
5339 old_disks = mddev->raid_disks - mddev->delta_disks;
5340 /* reshape_position must be on a new-stripe boundary, and one
5341 * further up in new geometry must map after here in old
5342 * geometry.
5343 */
5344 here_new = mddev->reshape_position;
5345 if (sector_div(here_new, mddev->new_chunk_sectors *
5346 (mddev->raid_disks - max_degraded))) {
5347 printk(KERN_ERR "md/raid:%s: reshape_position not "
5348 "on a stripe boundary\n", mdname(mddev));
5349 return -EINVAL;
5350 }
5351 reshape_offset = here_new * mddev->new_chunk_sectors;
5352 /* here_new is the stripe we will write to */
5353 here_old = mddev->reshape_position;
5354 sector_div(here_old, mddev->chunk_sectors *
5355 (old_disks-max_degraded));
5356 /* here_old is the first stripe that we might need to read
5357 * from */
5358 if (mddev->delta_disks == 0) {
5359 if ((here_new * mddev->new_chunk_sectors !=
5360 here_old * mddev->chunk_sectors)) {
5361 printk(KERN_ERR "md/raid:%s: reshape position is"
5362 " confused - aborting\n", mdname(mddev));
5363 return -EINVAL;
5364 }
5365 /* We cannot be sure it is safe to start an in-place
5366 * reshape. It is only safe if user-space is monitoring
5367 * and taking constant backups.
5368 * mdadm always starts a situation like this in
5369 * readonly mode so it can take control before
5370 * allowing any writes. So just check for that.
5371 */
5372 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5373 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5374 /* not really in-place - so OK */;
5375 else if (mddev->ro == 0) {
5376 printk(KERN_ERR "md/raid:%s: in-place reshape "
5377 "must be started in read-only mode "
5378 "- aborting\n",
5379 mdname(mddev));
5380 return -EINVAL;
5381 }
5382 } else if (mddev->reshape_backwards
5383 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5384 here_old * mddev->chunk_sectors)
5385 : (here_new * mddev->new_chunk_sectors >=
5386 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5387 /* Reading from the same stripe as writing to - bad */
5388 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5389 "auto-recovery - aborting.\n",
5390 mdname(mddev));
5391 return -EINVAL;
5392 }
5393 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5394 mdname(mddev));
5395 /* OK, we should be able to continue; */
5396 } else {
5397 BUG_ON(mddev->level != mddev->new_level);
5398 BUG_ON(mddev->layout != mddev->new_layout);
5399 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5400 BUG_ON(mddev->delta_disks != 0);
5401 }
5402
5403 if (mddev->private == NULL)
5404 conf = setup_conf(mddev);
5405 else
5406 conf = mddev->private;
5407
5408 if (IS_ERR(conf))
5409 return PTR_ERR(conf);
5410
5411 conf->min_offset_diff = min_offset_diff;
5412 mddev->thread = conf->thread;
5413 conf->thread = NULL;
5414 mddev->private = conf;
5415
5416 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5417 i++) {
5418 rdev = conf->disks[i].rdev;
5419 if (!rdev && conf->disks[i].replacement) {
5420 /* The replacement is all we have yet */
5421 rdev = conf->disks[i].replacement;
5422 conf->disks[i].replacement = NULL;
5423 clear_bit(Replacement, &rdev->flags);
5424 conf->disks[i].rdev = rdev;
5425 }
5426 if (!rdev)
5427 continue;
5428 if (conf->disks[i].replacement &&
5429 conf->reshape_progress != MaxSector) {
5430 /* replacements and reshape simply do not mix. */
5431 printk(KERN_ERR "md: cannot handle concurrent "
5432 "replacement and reshape.\n");
5433 goto abort;
5434 }
5435 if (test_bit(In_sync, &rdev->flags)) {
5436 working_disks++;
5437 continue;
5438 }
5439 /* This disc is not fully in-sync. However if it
5440 * just stored parity (beyond the recovery_offset),
5441 * when we don't need to be concerned about the
5442 * array being dirty.
5443 * When reshape goes 'backwards', we never have
5444 * partially completed devices, so we only need
5445 * to worry about reshape going forwards.
5446 */
5447 /* Hack because v0.91 doesn't store recovery_offset properly. */
5448 if (mddev->major_version == 0 &&
5449 mddev->minor_version > 90)
5450 rdev->recovery_offset = reshape_offset;
5451
5452 if (rdev->recovery_offset < reshape_offset) {
5453 /* We need to check old and new layout */
5454 if (!only_parity(rdev->raid_disk,
5455 conf->algorithm,
5456 conf->raid_disks,
5457 conf->max_degraded))
5458 continue;
5459 }
5460 if (!only_parity(rdev->raid_disk,
5461 conf->prev_algo,
5462 conf->previous_raid_disks,
5463 conf->max_degraded))
5464 continue;
5465 dirty_parity_disks++;
5466 }
5467
5468 /*
5469 * 0 for a fully functional array, 1 or 2 for a degraded array.
5470 */
5471 mddev->degraded = calc_degraded(conf);
5472
5473 if (has_failed(conf)) {
5474 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5475 " (%d/%d failed)\n",
5476 mdname(mddev), mddev->degraded, conf->raid_disks);
5477 goto abort;
5478 }
5479
5480 /* device size must be a multiple of chunk size */
5481 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5482 mddev->resync_max_sectors = mddev->dev_sectors;
5483
5484 if (mddev->degraded > dirty_parity_disks &&
5485 mddev->recovery_cp != MaxSector) {
5486 if (mddev->ok_start_degraded)
5487 printk(KERN_WARNING
5488 "md/raid:%s: starting dirty degraded array"
5489 " - data corruption possible.\n",
5490 mdname(mddev));
5491 else {
5492 printk(KERN_ERR
5493 "md/raid:%s: cannot start dirty degraded array.\n",
5494 mdname(mddev));
5495 goto abort;
5496 }
5497 }
5498
5499 if (mddev->degraded == 0)
5500 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5501 " devices, algorithm %d\n", mdname(mddev), conf->level,
5502 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5503 mddev->new_layout);
5504 else
5505 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5506 " out of %d devices, algorithm %d\n",
5507 mdname(mddev), conf->level,
5508 mddev->raid_disks - mddev->degraded,
5509 mddev->raid_disks, mddev->new_layout);
5510
5511 print_raid5_conf(conf);
5512
5513 if (conf->reshape_progress != MaxSector) {
5514 conf->reshape_safe = conf->reshape_progress;
5515 atomic_set(&conf->reshape_stripes, 0);
5516 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5517 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5518 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5519 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5520 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5521 "reshape");
5522 }
5523
5524
5525 /* Ok, everything is just fine now */
5526 if (mddev->to_remove == &raid5_attrs_group)
5527 mddev->to_remove = NULL;
5528 else if (mddev->kobj.sd &&
5529 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5530 printk(KERN_WARNING
5531 "raid5: failed to create sysfs attributes for %s\n",
5532 mdname(mddev));
5533 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5534
5535 if (mddev->queue) {
5536 int chunk_size;
5537 bool discard_supported = true;
5538 /* read-ahead size must cover two whole stripes, which
5539 * is 2 * (datadisks) * chunksize where 'n' is the
5540 * number of raid devices
5541 */
5542 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5543 int stripe = data_disks *
5544 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5545 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5546 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5547
5548 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5549
5550 mddev->queue->backing_dev_info.congested_data = mddev;
5551 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5552
5553 chunk_size = mddev->chunk_sectors << 9;
5554 blk_queue_io_min(mddev->queue, chunk_size);
5555 blk_queue_io_opt(mddev->queue, chunk_size *
5556 (conf->raid_disks - conf->max_degraded));
5557 /*
5558 * We can only discard a whole stripe. It doesn't make sense to
5559 * discard data disk but write parity disk
5560 */
5561 stripe = stripe * PAGE_SIZE;
5562 /* Round up to power of 2, as discard handling
5563 * currently assumes that */
5564 while ((stripe-1) & stripe)
5565 stripe = (stripe | (stripe-1)) + 1;
5566 mddev->queue->limits.discard_alignment = stripe;
5567 mddev->queue->limits.discard_granularity = stripe;
5568 /*
5569 * unaligned part of discard request will be ignored, so can't
5570 * guarantee discard_zerors_data
5571 */
5572 mddev->queue->limits.discard_zeroes_data = 0;
5573
5574 rdev_for_each(rdev, mddev) {
5575 disk_stack_limits(mddev->gendisk, rdev->bdev,
5576 rdev->data_offset << 9);
5577 disk_stack_limits(mddev->gendisk, rdev->bdev,
5578 rdev->new_data_offset << 9);
5579 /*
5580 * discard_zeroes_data is required, otherwise data
5581 * could be lost. Consider a scenario: discard a stripe
5582 * (the stripe could be inconsistent if
5583 * discard_zeroes_data is 0); write one disk of the
5584 * stripe (the stripe could be inconsistent again
5585 * depending on which disks are used to calculate
5586 * parity); the disk is broken; The stripe data of this
5587 * disk is lost.
5588 */
5589 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5590 !bdev_get_queue(rdev->bdev)->
5591 limits.discard_zeroes_data)
5592 discard_supported = false;
5593 }
5594
5595 if (discard_supported &&
5596 mddev->queue->limits.max_discard_sectors >= stripe &&
5597 mddev->queue->limits.discard_granularity >= stripe)
5598 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5599 mddev->queue);
5600 else
5601 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5602 mddev->queue);
5603 }
5604
5605 return 0;
5606 abort:
5607 md_unregister_thread(&mddev->thread);
5608 print_raid5_conf(conf);
5609 free_conf(conf);
5610 mddev->private = NULL;
5611 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5612 return -EIO;
5613 }
5614
5615 static int stop(struct mddev *mddev)
5616 {
5617 struct r5conf *conf = mddev->private;
5618
5619 md_unregister_thread(&mddev->thread);
5620 if (mddev->queue)
5621 mddev->queue->backing_dev_info.congested_fn = NULL;
5622 free_conf(conf);
5623 mddev->private = NULL;
5624 mddev->to_remove = &raid5_attrs_group;
5625 return 0;
5626 }
5627
5628 static void status(struct seq_file *seq, struct mddev *mddev)
5629 {
5630 struct r5conf *conf = mddev->private;
5631 int i;
5632
5633 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5634 mddev->chunk_sectors / 2, mddev->layout);
5635 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5636 for (i = 0; i < conf->raid_disks; i++)
5637 seq_printf (seq, "%s",
5638 conf->disks[i].rdev &&
5639 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5640 seq_printf (seq, "]");
5641 }
5642
5643 static void print_raid5_conf (struct r5conf *conf)
5644 {
5645 int i;
5646 struct disk_info *tmp;
5647
5648 printk(KERN_DEBUG "RAID conf printout:\n");
5649 if (!conf) {
5650 printk("(conf==NULL)\n");
5651 return;
5652 }
5653 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5654 conf->raid_disks,
5655 conf->raid_disks - conf->mddev->degraded);
5656
5657 for (i = 0; i < conf->raid_disks; i++) {
5658 char b[BDEVNAME_SIZE];
5659 tmp = conf->disks + i;
5660 if (tmp->rdev)
5661 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5662 i, !test_bit(Faulty, &tmp->rdev->flags),
5663 bdevname(tmp->rdev->bdev, b));
5664 }
5665 }
5666
5667 static int raid5_spare_active(struct mddev *mddev)
5668 {
5669 int i;
5670 struct r5conf *conf = mddev->private;
5671 struct disk_info *tmp;
5672 int count = 0;
5673 unsigned long flags;
5674
5675 for (i = 0; i < conf->raid_disks; i++) {
5676 tmp = conf->disks + i;
5677 if (tmp->replacement
5678 && tmp->replacement->recovery_offset == MaxSector
5679 && !test_bit(Faulty, &tmp->replacement->flags)
5680 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5681 /* Replacement has just become active. */
5682 if (!tmp->rdev
5683 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5684 count++;
5685 if (tmp->rdev) {
5686 /* Replaced device not technically faulty,
5687 * but we need to be sure it gets removed
5688 * and never re-added.
5689 */
5690 set_bit(Faulty, &tmp->rdev->flags);
5691 sysfs_notify_dirent_safe(
5692 tmp->rdev->sysfs_state);
5693 }
5694 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5695 } else if (tmp->rdev
5696 && tmp->rdev->recovery_offset == MaxSector
5697 && !test_bit(Faulty, &tmp->rdev->flags)
5698 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5699 count++;
5700 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5701 }
5702 }
5703 spin_lock_irqsave(&conf->device_lock, flags);
5704 mddev->degraded = calc_degraded(conf);
5705 spin_unlock_irqrestore(&conf->device_lock, flags);
5706 print_raid5_conf(conf);
5707 return count;
5708 }
5709
5710 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5711 {
5712 struct r5conf *conf = mddev->private;
5713 int err = 0;
5714 int number = rdev->raid_disk;
5715 struct md_rdev **rdevp;
5716 struct disk_info *p = conf->disks + number;
5717
5718 print_raid5_conf(conf);
5719 if (rdev == p->rdev)
5720 rdevp = &p->rdev;
5721 else if (rdev == p->replacement)
5722 rdevp = &p->replacement;
5723 else
5724 return 0;
5725
5726 if (number >= conf->raid_disks &&
5727 conf->reshape_progress == MaxSector)
5728 clear_bit(In_sync, &rdev->flags);
5729
5730 if (test_bit(In_sync, &rdev->flags) ||
5731 atomic_read(&rdev->nr_pending)) {
5732 err = -EBUSY;
5733 goto abort;
5734 }
5735 /* Only remove non-faulty devices if recovery
5736 * isn't possible.
5737 */
5738 if (!test_bit(Faulty, &rdev->flags) &&
5739 mddev->recovery_disabled != conf->recovery_disabled &&
5740 !has_failed(conf) &&
5741 (!p->replacement || p->replacement == rdev) &&
5742 number < conf->raid_disks) {
5743 err = -EBUSY;
5744 goto abort;
5745 }
5746 *rdevp = NULL;
5747 synchronize_rcu();
5748 if (atomic_read(&rdev->nr_pending)) {
5749 /* lost the race, try later */
5750 err = -EBUSY;
5751 *rdevp = rdev;
5752 } else if (p->replacement) {
5753 /* We must have just cleared 'rdev' */
5754 p->rdev = p->replacement;
5755 clear_bit(Replacement, &p->replacement->flags);
5756 smp_mb(); /* Make sure other CPUs may see both as identical
5757 * but will never see neither - if they are careful
5758 */
5759 p->replacement = NULL;
5760 clear_bit(WantReplacement, &rdev->flags);
5761 } else
5762 /* We might have just removed the Replacement as faulty-
5763 * clear the bit just in case
5764 */
5765 clear_bit(WantReplacement, &rdev->flags);
5766 abort:
5767
5768 print_raid5_conf(conf);
5769 return err;
5770 }
5771
5772 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5773 {
5774 struct r5conf *conf = mddev->private;
5775 int err = -EEXIST;
5776 int disk;
5777 struct disk_info *p;
5778 int first = 0;
5779 int last = conf->raid_disks - 1;
5780
5781 if (mddev->recovery_disabled == conf->recovery_disabled)
5782 return -EBUSY;
5783
5784 if (rdev->saved_raid_disk < 0 && has_failed(conf))
5785 /* no point adding a device */
5786 return -EINVAL;
5787
5788 if (rdev->raid_disk >= 0)
5789 first = last = rdev->raid_disk;
5790
5791 /*
5792 * find the disk ... but prefer rdev->saved_raid_disk
5793 * if possible.
5794 */
5795 if (rdev->saved_raid_disk >= 0 &&
5796 rdev->saved_raid_disk >= first &&
5797 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5798 first = rdev->saved_raid_disk;
5799
5800 for (disk = first; disk <= last; disk++) {
5801 p = conf->disks + disk;
5802 if (p->rdev == NULL) {
5803 clear_bit(In_sync, &rdev->flags);
5804 rdev->raid_disk = disk;
5805 err = 0;
5806 if (rdev->saved_raid_disk != disk)
5807 conf->fullsync = 1;
5808 rcu_assign_pointer(p->rdev, rdev);
5809 goto out;
5810 }
5811 }
5812 for (disk = first; disk <= last; disk++) {
5813 p = conf->disks + disk;
5814 if (test_bit(WantReplacement, &p->rdev->flags) &&
5815 p->replacement == NULL) {
5816 clear_bit(In_sync, &rdev->flags);
5817 set_bit(Replacement, &rdev->flags);
5818 rdev->raid_disk = disk;
5819 err = 0;
5820 conf->fullsync = 1;
5821 rcu_assign_pointer(p->replacement, rdev);
5822 break;
5823 }
5824 }
5825 out:
5826 print_raid5_conf(conf);
5827 return err;
5828 }
5829
5830 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5831 {
5832 /* no resync is happening, and there is enough space
5833 * on all devices, so we can resize.
5834 * We need to make sure resync covers any new space.
5835 * If the array is shrinking we should possibly wait until
5836 * any io in the removed space completes, but it hardly seems
5837 * worth it.
5838 */
5839 sector_t newsize;
5840 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5841 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5842 if (mddev->external_size &&
5843 mddev->array_sectors > newsize)
5844 return -EINVAL;
5845 if (mddev->bitmap) {
5846 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5847 if (ret)
5848 return ret;
5849 }
5850 md_set_array_sectors(mddev, newsize);
5851 set_capacity(mddev->gendisk, mddev->array_sectors);
5852 revalidate_disk(mddev->gendisk);
5853 if (sectors > mddev->dev_sectors &&
5854 mddev->recovery_cp > mddev->dev_sectors) {
5855 mddev->recovery_cp = mddev->dev_sectors;
5856 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5857 }
5858 mddev->dev_sectors = sectors;
5859 mddev->resync_max_sectors = sectors;
5860 return 0;
5861 }
5862
5863 static int check_stripe_cache(struct mddev *mddev)
5864 {
5865 /* Can only proceed if there are plenty of stripe_heads.
5866 * We need a minimum of one full stripe,, and for sensible progress
5867 * it is best to have about 4 times that.
5868 * If we require 4 times, then the default 256 4K stripe_heads will
5869 * allow for chunk sizes up to 256K, which is probably OK.
5870 * If the chunk size is greater, user-space should request more
5871 * stripe_heads first.
5872 */
5873 struct r5conf *conf = mddev->private;
5874 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5875 > conf->max_nr_stripes ||
5876 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5877 > conf->max_nr_stripes) {
5878 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5879 mdname(mddev),
5880 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5881 / STRIPE_SIZE)*4);
5882 return 0;
5883 }
5884 return 1;
5885 }
5886
5887 static int check_reshape(struct mddev *mddev)
5888 {
5889 struct r5conf *conf = mddev->private;
5890
5891 if (mddev->delta_disks == 0 &&
5892 mddev->new_layout == mddev->layout &&
5893 mddev->new_chunk_sectors == mddev->chunk_sectors)
5894 return 0; /* nothing to do */
5895 if (has_failed(conf))
5896 return -EINVAL;
5897 if (mddev->delta_disks < 0) {
5898 /* We might be able to shrink, but the devices must
5899 * be made bigger first.
5900 * For raid6, 4 is the minimum size.
5901 * Otherwise 2 is the minimum
5902 */
5903 int min = 2;
5904 if (mddev->level == 6)
5905 min = 4;
5906 if (mddev->raid_disks + mddev->delta_disks < min)
5907 return -EINVAL;
5908 }
5909
5910 if (!check_stripe_cache(mddev))
5911 return -ENOSPC;
5912
5913 return resize_stripes(conf, (conf->previous_raid_disks
5914 + mddev->delta_disks));
5915 }
5916
5917 static int raid5_start_reshape(struct mddev *mddev)
5918 {
5919 struct r5conf *conf = mddev->private;
5920 struct md_rdev *rdev;
5921 int spares = 0;
5922 unsigned long flags;
5923
5924 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5925 return -EBUSY;
5926
5927 if (!check_stripe_cache(mddev))
5928 return -ENOSPC;
5929
5930 if (has_failed(conf))
5931 return -EINVAL;
5932
5933 rdev_for_each(rdev, mddev) {
5934 if (!test_bit(In_sync, &rdev->flags)
5935 && !test_bit(Faulty, &rdev->flags))
5936 spares++;
5937 }
5938
5939 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5940 /* Not enough devices even to make a degraded array
5941 * of that size
5942 */
5943 return -EINVAL;
5944
5945 /* Refuse to reduce size of the array. Any reductions in
5946 * array size must be through explicit setting of array_size
5947 * attribute.
5948 */
5949 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5950 < mddev->array_sectors) {
5951 printk(KERN_ERR "md/raid:%s: array size must be reduced "
5952 "before number of disks\n", mdname(mddev));
5953 return -EINVAL;
5954 }
5955
5956 atomic_set(&conf->reshape_stripes, 0);
5957 spin_lock_irq(&conf->device_lock);
5958 conf->previous_raid_disks = conf->raid_disks;
5959 conf->raid_disks += mddev->delta_disks;
5960 conf->prev_chunk_sectors = conf->chunk_sectors;
5961 conf->chunk_sectors = mddev->new_chunk_sectors;
5962 conf->prev_algo = conf->algorithm;
5963 conf->algorithm = mddev->new_layout;
5964 conf->generation++;
5965 /* Code that selects data_offset needs to see the generation update
5966 * if reshape_progress has been set - so a memory barrier needed.
5967 */
5968 smp_mb();
5969 if (mddev->reshape_backwards)
5970 conf->reshape_progress = raid5_size(mddev, 0, 0);
5971 else
5972 conf->reshape_progress = 0;
5973 conf->reshape_safe = conf->reshape_progress;
5974 spin_unlock_irq(&conf->device_lock);
5975
5976 /* Add some new drives, as many as will fit.
5977 * We know there are enough to make the newly sized array work.
5978 * Don't add devices if we are reducing the number of
5979 * devices in the array. This is because it is not possible
5980 * to correctly record the "partially reconstructed" state of
5981 * such devices during the reshape and confusion could result.
5982 */
5983 if (mddev->delta_disks >= 0) {
5984 rdev_for_each(rdev, mddev)
5985 if (rdev->raid_disk < 0 &&
5986 !test_bit(Faulty, &rdev->flags)) {
5987 if (raid5_add_disk(mddev, rdev) == 0) {
5988 if (rdev->raid_disk
5989 >= conf->previous_raid_disks)
5990 set_bit(In_sync, &rdev->flags);
5991 else
5992 rdev->recovery_offset = 0;
5993
5994 if (sysfs_link_rdev(mddev, rdev))
5995 /* Failure here is OK */;
5996 }
5997 } else if (rdev->raid_disk >= conf->previous_raid_disks
5998 && !test_bit(Faulty, &rdev->flags)) {
5999 /* This is a spare that was manually added */
6000 set_bit(In_sync, &rdev->flags);
6001 }
6002
6003 /* When a reshape changes the number of devices,
6004 * ->degraded is measured against the larger of the
6005 * pre and post number of devices.
6006 */
6007 spin_lock_irqsave(&conf->device_lock, flags);
6008 mddev->degraded = calc_degraded(conf);
6009 spin_unlock_irqrestore(&conf->device_lock, flags);
6010 }
6011 mddev->raid_disks = conf->raid_disks;
6012 mddev->reshape_position = conf->reshape_progress;
6013 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6014
6015 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6016 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6017 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6018 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6019 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6020 "reshape");
6021 if (!mddev->sync_thread) {
6022 mddev->recovery = 0;
6023 spin_lock_irq(&conf->device_lock);
6024 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6025 rdev_for_each(rdev, mddev)
6026 rdev->new_data_offset = rdev->data_offset;
6027 smp_wmb();
6028 conf->reshape_progress = MaxSector;
6029 mddev->reshape_position = MaxSector;
6030 spin_unlock_irq(&conf->device_lock);
6031 return -EAGAIN;
6032 }
6033 conf->reshape_checkpoint = jiffies;
6034 md_wakeup_thread(mddev->sync_thread);
6035 md_new_event(mddev);
6036 return 0;
6037 }
6038
6039 /* This is called from the reshape thread and should make any
6040 * changes needed in 'conf'
6041 */
6042 static void end_reshape(struct r5conf *conf)
6043 {
6044
6045 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6046 struct md_rdev *rdev;
6047
6048 spin_lock_irq(&conf->device_lock);
6049 conf->previous_raid_disks = conf->raid_disks;
6050 rdev_for_each(rdev, conf->mddev)
6051 rdev->data_offset = rdev->new_data_offset;
6052 smp_wmb();
6053 conf->reshape_progress = MaxSector;
6054 spin_unlock_irq(&conf->device_lock);
6055 wake_up(&conf->wait_for_overlap);
6056
6057 /* read-ahead size must cover two whole stripes, which is
6058 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6059 */
6060 if (conf->mddev->queue) {
6061 int data_disks = conf->raid_disks - conf->max_degraded;
6062 int stripe = data_disks * ((conf->chunk_sectors << 9)
6063 / PAGE_SIZE);
6064 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6065 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6066 }
6067 }
6068 }
6069
6070 /* This is called from the raid5d thread with mddev_lock held.
6071 * It makes config changes to the device.
6072 */
6073 static void raid5_finish_reshape(struct mddev *mddev)
6074 {
6075 struct r5conf *conf = mddev->private;
6076
6077 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6078
6079 if (mddev->delta_disks > 0) {
6080 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6081 set_capacity(mddev->gendisk, mddev->array_sectors);
6082 revalidate_disk(mddev->gendisk);
6083 } else {
6084 int d;
6085 spin_lock_irq(&conf->device_lock);
6086 mddev->degraded = calc_degraded(conf);
6087 spin_unlock_irq(&conf->device_lock);
6088 for (d = conf->raid_disks ;
6089 d < conf->raid_disks - mddev->delta_disks;
6090 d++) {
6091 struct md_rdev *rdev = conf->disks[d].rdev;
6092 if (rdev)
6093 clear_bit(In_sync, &rdev->flags);
6094 rdev = conf->disks[d].replacement;
6095 if (rdev)
6096 clear_bit(In_sync, &rdev->flags);
6097 }
6098 }
6099 mddev->layout = conf->algorithm;
6100 mddev->chunk_sectors = conf->chunk_sectors;
6101 mddev->reshape_position = MaxSector;
6102 mddev->delta_disks = 0;
6103 mddev->reshape_backwards = 0;
6104 }
6105 }
6106
6107 static void raid5_quiesce(struct mddev *mddev, int state)
6108 {
6109 struct r5conf *conf = mddev->private;
6110
6111 switch(state) {
6112 case 2: /* resume for a suspend */
6113 wake_up(&conf->wait_for_overlap);
6114 break;
6115
6116 case 1: /* stop all writes */
6117 spin_lock_irq(&conf->device_lock);
6118 /* '2' tells resync/reshape to pause so that all
6119 * active stripes can drain
6120 */
6121 conf->quiesce = 2;
6122 wait_event_lock_irq(conf->wait_for_stripe,
6123 atomic_read(&conf->active_stripes) == 0 &&
6124 atomic_read(&conf->active_aligned_reads) == 0,
6125 conf->device_lock);
6126 conf->quiesce = 1;
6127 spin_unlock_irq(&conf->device_lock);
6128 /* allow reshape to continue */
6129 wake_up(&conf->wait_for_overlap);
6130 break;
6131
6132 case 0: /* re-enable writes */
6133 spin_lock_irq(&conf->device_lock);
6134 conf->quiesce = 0;
6135 wake_up(&conf->wait_for_stripe);
6136 wake_up(&conf->wait_for_overlap);
6137 spin_unlock_irq(&conf->device_lock);
6138 break;
6139 }
6140 }
6141
6142
6143 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6144 {
6145 struct r0conf *raid0_conf = mddev->private;
6146 sector_t sectors;
6147
6148 /* for raid0 takeover only one zone is supported */
6149 if (raid0_conf->nr_strip_zones > 1) {
6150 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6151 mdname(mddev));
6152 return ERR_PTR(-EINVAL);
6153 }
6154
6155 sectors = raid0_conf->strip_zone[0].zone_end;
6156 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6157 mddev->dev_sectors = sectors;
6158 mddev->new_level = level;
6159 mddev->new_layout = ALGORITHM_PARITY_N;
6160 mddev->new_chunk_sectors = mddev->chunk_sectors;
6161 mddev->raid_disks += 1;
6162 mddev->delta_disks = 1;
6163 /* make sure it will be not marked as dirty */
6164 mddev->recovery_cp = MaxSector;
6165
6166 return setup_conf(mddev);
6167 }
6168
6169
6170 static void *raid5_takeover_raid1(struct mddev *mddev)
6171 {
6172 int chunksect;
6173
6174 if (mddev->raid_disks != 2 ||
6175 mddev->degraded > 1)
6176 return ERR_PTR(-EINVAL);
6177
6178 /* Should check if there are write-behind devices? */
6179
6180 chunksect = 64*2; /* 64K by default */
6181
6182 /* The array must be an exact multiple of chunksize */
6183 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6184 chunksect >>= 1;
6185
6186 if ((chunksect<<9) < STRIPE_SIZE)
6187 /* array size does not allow a suitable chunk size */
6188 return ERR_PTR(-EINVAL);
6189
6190 mddev->new_level = 5;
6191 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6192 mddev->new_chunk_sectors = chunksect;
6193
6194 return setup_conf(mddev);
6195 }
6196
6197 static void *raid5_takeover_raid6(struct mddev *mddev)
6198 {
6199 int new_layout;
6200
6201 switch (mddev->layout) {
6202 case ALGORITHM_LEFT_ASYMMETRIC_6:
6203 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6204 break;
6205 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6206 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6207 break;
6208 case ALGORITHM_LEFT_SYMMETRIC_6:
6209 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6210 break;
6211 case ALGORITHM_RIGHT_SYMMETRIC_6:
6212 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6213 break;
6214 case ALGORITHM_PARITY_0_6:
6215 new_layout = ALGORITHM_PARITY_0;
6216 break;
6217 case ALGORITHM_PARITY_N:
6218 new_layout = ALGORITHM_PARITY_N;
6219 break;
6220 default:
6221 return ERR_PTR(-EINVAL);
6222 }
6223 mddev->new_level = 5;
6224 mddev->new_layout = new_layout;
6225 mddev->delta_disks = -1;
6226 mddev->raid_disks -= 1;
6227 return setup_conf(mddev);
6228 }
6229
6230
6231 static int raid5_check_reshape(struct mddev *mddev)
6232 {
6233 /* For a 2-drive array, the layout and chunk size can be changed
6234 * immediately as not restriping is needed.
6235 * For larger arrays we record the new value - after validation
6236 * to be used by a reshape pass.
6237 */
6238 struct r5conf *conf = mddev->private;
6239 int new_chunk = mddev->new_chunk_sectors;
6240
6241 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6242 return -EINVAL;
6243 if (new_chunk > 0) {
6244 if (!is_power_of_2(new_chunk))
6245 return -EINVAL;
6246 if (new_chunk < (PAGE_SIZE>>9))
6247 return -EINVAL;
6248 if (mddev->array_sectors & (new_chunk-1))
6249 /* not factor of array size */
6250 return -EINVAL;
6251 }
6252
6253 /* They look valid */
6254
6255 if (mddev->raid_disks == 2) {
6256 /* can make the change immediately */
6257 if (mddev->new_layout >= 0) {
6258 conf->algorithm = mddev->new_layout;
6259 mddev->layout = mddev->new_layout;
6260 }
6261 if (new_chunk > 0) {
6262 conf->chunk_sectors = new_chunk ;
6263 mddev->chunk_sectors = new_chunk;
6264 }
6265 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6266 md_wakeup_thread(mddev->thread);
6267 }
6268 return check_reshape(mddev);
6269 }
6270
6271 static int raid6_check_reshape(struct mddev *mddev)
6272 {
6273 int new_chunk = mddev->new_chunk_sectors;
6274
6275 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6276 return -EINVAL;
6277 if (new_chunk > 0) {
6278 if (!is_power_of_2(new_chunk))
6279 return -EINVAL;
6280 if (new_chunk < (PAGE_SIZE >> 9))
6281 return -EINVAL;
6282 if (mddev->array_sectors & (new_chunk-1))
6283 /* not factor of array size */
6284 return -EINVAL;
6285 }
6286
6287 /* They look valid */
6288 return check_reshape(mddev);
6289 }
6290
6291 static void *raid5_takeover(struct mddev *mddev)
6292 {
6293 /* raid5 can take over:
6294 * raid0 - if there is only one strip zone - make it a raid4 layout
6295 * raid1 - if there are two drives. We need to know the chunk size
6296 * raid4 - trivial - just use a raid4 layout.
6297 * raid6 - Providing it is a *_6 layout
6298 */
6299 if (mddev->level == 0)
6300 return raid45_takeover_raid0(mddev, 5);
6301 if (mddev->level == 1)
6302 return raid5_takeover_raid1(mddev);
6303 if (mddev->level == 4) {
6304 mddev->new_layout = ALGORITHM_PARITY_N;
6305 mddev->new_level = 5;
6306 return setup_conf(mddev);
6307 }
6308 if (mddev->level == 6)
6309 return raid5_takeover_raid6(mddev);
6310
6311 return ERR_PTR(-EINVAL);
6312 }
6313
6314 static void *raid4_takeover(struct mddev *mddev)
6315 {
6316 /* raid4 can take over:
6317 * raid0 - if there is only one strip zone
6318 * raid5 - if layout is right
6319 */
6320 if (mddev->level == 0)
6321 return raid45_takeover_raid0(mddev, 4);
6322 if (mddev->level == 5 &&
6323 mddev->layout == ALGORITHM_PARITY_N) {
6324 mddev->new_layout = 0;
6325 mddev->new_level = 4;
6326 return setup_conf(mddev);
6327 }
6328 return ERR_PTR(-EINVAL);
6329 }
6330
6331 static struct md_personality raid5_personality;
6332
6333 static void *raid6_takeover(struct mddev *mddev)
6334 {
6335 /* Currently can only take over a raid5. We map the
6336 * personality to an equivalent raid6 personality
6337 * with the Q block at the end.
6338 */
6339 int new_layout;
6340
6341 if (mddev->pers != &raid5_personality)
6342 return ERR_PTR(-EINVAL);
6343 if (mddev->degraded > 1)
6344 return ERR_PTR(-EINVAL);
6345 if (mddev->raid_disks > 253)
6346 return ERR_PTR(-EINVAL);
6347 if (mddev->raid_disks < 3)
6348 return ERR_PTR(-EINVAL);
6349
6350 switch (mddev->layout) {
6351 case ALGORITHM_LEFT_ASYMMETRIC:
6352 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6353 break;
6354 case ALGORITHM_RIGHT_ASYMMETRIC:
6355 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6356 break;
6357 case ALGORITHM_LEFT_SYMMETRIC:
6358 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6359 break;
6360 case ALGORITHM_RIGHT_SYMMETRIC:
6361 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6362 break;
6363 case ALGORITHM_PARITY_0:
6364 new_layout = ALGORITHM_PARITY_0_6;
6365 break;
6366 case ALGORITHM_PARITY_N:
6367 new_layout = ALGORITHM_PARITY_N;
6368 break;
6369 default:
6370 return ERR_PTR(-EINVAL);
6371 }
6372 mddev->new_level = 6;
6373 mddev->new_layout = new_layout;
6374 mddev->delta_disks = 1;
6375 mddev->raid_disks += 1;
6376 return setup_conf(mddev);
6377 }
6378
6379
6380 static struct md_personality raid6_personality =
6381 {
6382 .name = "raid6",
6383 .level = 6,
6384 .owner = THIS_MODULE,
6385 .make_request = make_request,
6386 .run = run,
6387 .stop = stop,
6388 .status = status,
6389 .error_handler = error,
6390 .hot_add_disk = raid5_add_disk,
6391 .hot_remove_disk= raid5_remove_disk,
6392 .spare_active = raid5_spare_active,
6393 .sync_request = sync_request,
6394 .resize = raid5_resize,
6395 .size = raid5_size,
6396 .check_reshape = raid6_check_reshape,
6397 .start_reshape = raid5_start_reshape,
6398 .finish_reshape = raid5_finish_reshape,
6399 .quiesce = raid5_quiesce,
6400 .takeover = raid6_takeover,
6401 };
6402 static struct md_personality raid5_personality =
6403 {
6404 .name = "raid5",
6405 .level = 5,
6406 .owner = THIS_MODULE,
6407 .make_request = make_request,
6408 .run = run,
6409 .stop = stop,
6410 .status = status,
6411 .error_handler = error,
6412 .hot_add_disk = raid5_add_disk,
6413 .hot_remove_disk= raid5_remove_disk,
6414 .spare_active = raid5_spare_active,
6415 .sync_request = sync_request,
6416 .resize = raid5_resize,
6417 .size = raid5_size,
6418 .check_reshape = raid5_check_reshape,
6419 .start_reshape = raid5_start_reshape,
6420 .finish_reshape = raid5_finish_reshape,
6421 .quiesce = raid5_quiesce,
6422 .takeover = raid5_takeover,
6423 };
6424
6425 static struct md_personality raid4_personality =
6426 {
6427 .name = "raid4",
6428 .level = 4,
6429 .owner = THIS_MODULE,
6430 .make_request = make_request,
6431 .run = run,
6432 .stop = stop,
6433 .status = status,
6434 .error_handler = error,
6435 .hot_add_disk = raid5_add_disk,
6436 .hot_remove_disk= raid5_remove_disk,
6437 .spare_active = raid5_spare_active,
6438 .sync_request = sync_request,
6439 .resize = raid5_resize,
6440 .size = raid5_size,
6441 .check_reshape = raid5_check_reshape,
6442 .start_reshape = raid5_start_reshape,
6443 .finish_reshape = raid5_finish_reshape,
6444 .quiesce = raid5_quiesce,
6445 .takeover = raid4_takeover,
6446 };
6447
6448 static int __init raid5_init(void)
6449 {
6450 register_md_personality(&raid6_personality);
6451 register_md_personality(&raid5_personality);
6452 register_md_personality(&raid4_personality);
6453 return 0;
6454 }
6455
6456 static void raid5_exit(void)
6457 {
6458 unregister_md_personality(&raid6_personality);
6459 unregister_md_personality(&raid5_personality);
6460 unregister_md_personality(&raid4_personality);
6461 }
6462
6463 module_init(raid5_init);
6464 module_exit(raid5_exit);
6465 MODULE_LICENSE("GPL");
6466 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6467 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6468 MODULE_ALIAS("md-raid5");
6469 MODULE_ALIAS("md-raid4");
6470 MODULE_ALIAS("md-level-5");
6471 MODULE_ALIAS("md-level-4");
6472 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6473 MODULE_ALIAS("md-raid6");
6474 MODULE_ALIAS("md-level-6");
6475
6476 /* This used to be two separate modules, they were: */
6477 MODULE_ALIAS("raid5");
6478 MODULE_ALIAS("raid6");
Linux 2.6 libata-dev (mirror)