projects / GitHub / mt8127 / android_kernel_alcatel_ttab.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
drivers: power: report battery voltage in AOSP compatible format
[GitHub/mt8127/android_kernel_alcatel_ttab.git] / drivers / md / raid5.c
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 static bool devices_handle_discard_safely = false;
64 module_param(devices_handle_discard_safely, bool, 0644);
65 MODULE_PARM_DESC(devices_handle_discard_safely,
66 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
67 /*
68 * Stripe cache
69 */
70
71 #define NR_STRIPES 256
72 #define STRIPE_SIZE PAGE_SIZE
73 #define STRIPE_SHIFT (PAGE_SHIFT - 9)
74 #define STRIPE_SECTORS (STRIPE_SIZE>>9)
75 #define IO_THRESHOLD 1
76 #define BYPASS_THRESHOLD 1
77 #define NR_HASH (PAGE_SIZE / sizeof(struct hlist_head))
78 #define HASH_MASK (NR_HASH - 1)
79
80 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
81 {
82 int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
83 return &conf->stripe_hashtbl[hash];
84 }
85
86 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
87 * order without overlap. There may be several bio's per stripe+device, and
88 * a bio could span several devices.
89 * When walking this list for a particular stripe+device, we must never proceed
90 * beyond a bio that extends past this device, as the next bio might no longer
91 * be valid.
92 * This function is used to determine the 'next' bio in the list, given the sector
93 * of the current stripe+device
94 */
95 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
96 {
97 int sectors = bio_sectors(bio);
98 if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
99 return bio->bi_next;
100 else
101 return NULL;
102 }
103
104 /*
105 * We maintain a biased count of active stripes in the bottom 16 bits of
106 * bi_phys_segments, and a count of processed stripes in the upper 16 bits
107 */
108 static inline int raid5_bi_processed_stripes(struct bio *bio)
109 {
110 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
111 return (atomic_read(segments) >> 16) & 0xffff;
112 }
113
114 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
115 {
116 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
117 return atomic_sub_return(1, segments) & 0xffff;
118 }
119
120 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
121 {
122 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
123 atomic_inc(segments);
124 }
125
126 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
127 unsigned int cnt)
128 {
129 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
130 int old, new;
131
132 do {
133 old = atomic_read(segments);
134 new = (old & 0xffff) | (cnt << 16);
135 } while (atomic_cmpxchg(segments, old, new) != old);
136 }
137
138 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
139 {
140 atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
141 atomic_set(segments, cnt);
142 }
143
144 /* Find first data disk in a raid6 stripe */
145 static inline int raid6_d0(struct stripe_head *sh)
146 {
147 if (sh->ddf_layout)
148 /* ddf always start from first device */
149 return 0;
150 /* md starts just after Q block */
151 if (sh->qd_idx == sh->disks - 1)
152 return 0;
153 else
154 return sh->qd_idx + 1;
155 }
156 static inline int raid6_next_disk(int disk, int raid_disks)
157 {
158 disk++;
159 return (disk < raid_disks) ? disk : 0;
160 }
161
162 /* When walking through the disks in a raid5, starting at raid6_d0,
163 * We need to map each disk to a 'slot', where the data disks are slot
164 * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
165 * is raid_disks-1. This help does that mapping.
166 */
167 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
168 int *count, int syndrome_disks)
169 {
170 int slot = *count;
171
172 if (sh->ddf_layout)
173 (*count)++;
174 if (idx == sh->pd_idx)
175 return syndrome_disks;
176 if (idx == sh->qd_idx)
177 return syndrome_disks + 1;
178 if (!sh->ddf_layout)
179 (*count)++;
180 return slot;
181 }
182
183 static void return_io(struct bio *return_bi)
184 {
185 struct bio *bi = return_bi;
186 while (bi) {
187
188 return_bi = bi->bi_next;
189 bi->bi_next = NULL;
190 bi->bi_size = 0;
191 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
192 bi, 0);
193 bio_endio(bi, 0);
194 bi = return_bi;
195 }
196 }
197
198 static void print_raid5_conf (struct r5conf *conf);
199
200 static int stripe_operations_active(struct stripe_head *sh)
201 {
202 return sh->check_state || sh->reconstruct_state ||
203 test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
204 test_bit(STRIPE_COMPUTE_RUN, &sh->state);
205 }
206
207 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
208 {
209 BUG_ON(!list_empty(&sh->lru));
210 BUG_ON(atomic_read(&conf->active_stripes)==0);
211 if (test_bit(STRIPE_HANDLE, &sh->state)) {
212 if (test_bit(STRIPE_DELAYED, &sh->state) &&
213 !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
214 list_add_tail(&sh->lru, &conf->delayed_list);
215 else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
216 sh->bm_seq - conf->seq_write > 0)
217 list_add_tail(&sh->lru, &conf->bitmap_list);
218 else {
219 clear_bit(STRIPE_DELAYED, &sh->state);
220 clear_bit(STRIPE_BIT_DELAY, &sh->state);
221 list_add_tail(&sh->lru, &conf->handle_list);
222 }
223 md_wakeup_thread(conf->mddev->thread);
224 } else {
225 BUG_ON(stripe_operations_active(sh));
226 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
227 if (atomic_dec_return(&conf->preread_active_stripes)
228 < IO_THRESHOLD)
229 md_wakeup_thread(conf->mddev->thread);
230 atomic_dec(&conf->active_stripes);
231 if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
232 list_add_tail(&sh->lru, &conf->inactive_list);
233 wake_up(&conf->wait_for_stripe);
234 if (conf->retry_read_aligned)
235 md_wakeup_thread(conf->mddev->thread);
236 }
237 }
238 }
239
240 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
241 {
242 if (atomic_dec_and_test(&sh->count))
243 do_release_stripe(conf, sh);
244 }
245
246 static void release_stripe(struct stripe_head *sh)
247 {
248 struct r5conf *conf = sh->raid_conf;
249 unsigned long flags;
250
251 local_irq_save(flags);
252 if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
253 do_release_stripe(conf, sh);
254 spin_unlock(&conf->device_lock);
255 }
256 local_irq_restore(flags);
257 }
258
259 static inline void remove_hash(struct stripe_head *sh)
260 {
261 pr_debug("remove_hash(), stripe %llu\n",
262 (unsigned long long)sh->sector);
263
264 hlist_del_init(&sh->hash);
265 }
266
267 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
268 {
269 struct hlist_head *hp = stripe_hash(conf, sh->sector);
270
271 pr_debug("insert_hash(), stripe %llu\n",
272 (unsigned long long)sh->sector);
273
274 hlist_add_head(&sh->hash, hp);
275 }
276
277
278 /* find an idle stripe, make sure it is unhashed, and return it. */
279 static struct stripe_head *get_free_stripe(struct r5conf *conf)
280 {
281 struct stripe_head *sh = NULL;
282 struct list_head *first;
283
284 if (list_empty(&conf->inactive_list))
285 goto out;
286 first = conf->inactive_list.next;
287 sh = list_entry(first, struct stripe_head, lru);
288 list_del_init(first);
289 remove_hash(sh);
290 atomic_inc(&conf->active_stripes);
291 out:
292 return sh;
293 }
294
295 static void shrink_buffers(struct stripe_head *sh)
296 {
297 struct page *p;
298 int i;
299 int num = sh->raid_conf->pool_size;
300
301 for (i = 0; i < num ; i++) {
302 p = sh->dev[i].page;
303 if (!p)
304 continue;
305 sh->dev[i].page = NULL;
306 put_page(p);
307 }
308 }
309
310 static int grow_buffers(struct stripe_head *sh)
311 {
312 int i;
313 int num = sh->raid_conf->pool_size;
314
315 for (i = 0; i < num; i++) {
316 struct page *page;
317
318 if (!(page = alloc_page(GFP_KERNEL))) {
319 return 1;
320 }
321 sh->dev[i].page = page;
322 }
323 return 0;
324 }
325
326 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
327 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
328 struct stripe_head *sh);
329
330 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
331 {
332 struct r5conf *conf = sh->raid_conf;
333 int i;
334
335 BUG_ON(atomic_read(&sh->count) != 0);
336 BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
337 BUG_ON(stripe_operations_active(sh));
338
339 pr_debug("init_stripe called, stripe %llu\n",
340 (unsigned long long)sh->sector);
341
342 remove_hash(sh);
343
344 sh->generation = conf->generation - previous;
345 sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
346 sh->sector = sector;
347 stripe_set_idx(sector, conf, previous, sh);
348 sh->state = 0;
349
350
351 for (i = sh->disks; i--; ) {
352 struct r5dev *dev = &sh->dev[i];
353
354 if (dev->toread || dev->read || dev->towrite || dev->written ||
355 test_bit(R5_LOCKED, &dev->flags)) {
356 printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
357 (unsigned long long)sh->sector, i, dev->toread,
358 dev->read, dev->towrite, dev->written,
359 test_bit(R5_LOCKED, &dev->flags));
360 WARN_ON(1);
361 }
362 dev->flags = 0;
363 raid5_build_block(sh, i, previous);
364 }
365 insert_hash(conf, sh);
366 }
367
368 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
369 short generation)
370 {
371 struct stripe_head *sh;
372
373 pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
374 hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
375 if (sh->sector == sector && sh->generation == generation)
376 return sh;
377 pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
378 return NULL;
379 }
380
381 /*
382 * Need to check if array has failed when deciding whether to:
383 * - start an array
384 * - remove non-faulty devices
385 * - add a spare
386 * - allow a reshape
387 * This determination is simple when no reshape is happening.
388 * However if there is a reshape, we need to carefully check
389 * both the before and after sections.
390 * This is because some failed devices may only affect one
391 * of the two sections, and some non-in_sync devices may
392 * be insync in the section most affected by failed devices.
393 */
394 static int calc_degraded(struct r5conf *conf)
395 {
396 int degraded, degraded2;
397 int i;
398
399 rcu_read_lock();
400 degraded = 0;
401 for (i = 0; i < conf->previous_raid_disks; i++) {
402 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
403 if (rdev && test_bit(Faulty, &rdev->flags))
404 rdev = rcu_dereference(conf->disks[i].replacement);
405 if (!rdev || test_bit(Faulty, &rdev->flags))
406 degraded++;
407 else if (test_bit(In_sync, &rdev->flags))
408 ;
409 else
410 /* not in-sync or faulty.
411 * If the reshape increases the number of devices,
412 * this is being recovered by the reshape, so
413 * this 'previous' section is not in_sync.
414 * If the number of devices is being reduced however,
415 * the device can only be part of the array if
416 * we are reverting a reshape, so this section will
417 * be in-sync.
418 */
419 if (conf->raid_disks >= conf->previous_raid_disks)
420 degraded++;
421 }
422 rcu_read_unlock();
423 if (conf->raid_disks == conf->previous_raid_disks)
424 return degraded;
425 rcu_read_lock();
426 degraded2 = 0;
427 for (i = 0; i < conf->raid_disks; i++) {
428 struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
429 if (rdev && test_bit(Faulty, &rdev->flags))
430 rdev = rcu_dereference(conf->disks[i].replacement);
431 if (!rdev || test_bit(Faulty, &rdev->flags))
432 degraded2++;
433 else if (test_bit(In_sync, &rdev->flags))
434 ;
435 else
436 /* not in-sync or faulty.
437 * If reshape increases the number of devices, this
438 * section has already been recovered, else it
439 * almost certainly hasn't.
440 */
441 if (conf->raid_disks <= conf->previous_raid_disks)
442 degraded2++;
443 }
444 rcu_read_unlock();
445 if (degraded2 > degraded)
446 return degraded2;
447 return degraded;
448 }
449
450 static int has_failed(struct r5conf *conf)
451 {
452 int degraded;
453
454 if (conf->mddev->reshape_position == MaxSector)
455 return conf->mddev->degraded > conf->max_degraded;
456
457 degraded = calc_degraded(conf);
458 if (degraded > conf->max_degraded)
459 return 1;
460 return 0;
461 }
462
463 static struct stripe_head *
464 get_active_stripe(struct r5conf *conf, sector_t sector,
465 int previous, int noblock, int noquiesce)
466 {
467 struct stripe_head *sh;
468
469 pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
470
471 spin_lock_irq(&conf->device_lock);
472
473 do {
474 wait_event_lock_irq(conf->wait_for_stripe,
475 conf->quiesce == 0 || noquiesce,
476 conf->device_lock);
477 sh = __find_stripe(conf, sector, conf->generation - previous);
478 if (!sh) {
479 if (!conf->inactive_blocked)
480 sh = get_free_stripe(conf);
481 if (noblock && sh == NULL)
482 break;
483 if (!sh) {
484 conf->inactive_blocked = 1;
485 wait_event_lock_irq(conf->wait_for_stripe,
486 !list_empty(&conf->inactive_list) &&
487 (atomic_read(&conf->active_stripes)
488 < (conf->max_nr_stripes *3/4)
489 || !conf->inactive_blocked),
490 conf->device_lock);
491 conf->inactive_blocked = 0;
492 } else
493 init_stripe(sh, sector, previous);
494 } else {
495 if (atomic_read(&sh->count)) {
496 BUG_ON(!list_empty(&sh->lru)
497 && !test_bit(STRIPE_EXPANDING, &sh->state)
498 && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state));
499 } else {
500 if (!test_bit(STRIPE_HANDLE, &sh->state))
501 atomic_inc(&conf->active_stripes);
502 if (list_empty(&sh->lru) &&
503 !test_bit(STRIPE_EXPANDING, &sh->state))
504 BUG();
505 list_del_init(&sh->lru);
506 }
507 }
508 } while (sh == NULL);
509
510 if (sh)
511 atomic_inc(&sh->count);
512
513 spin_unlock_irq(&conf->device_lock);
514 return sh;
515 }
516
517 /* Determine if 'data_offset' or 'new_data_offset' should be used
518 * in this stripe_head.
519 */
520 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
521 {
522 sector_t progress = conf->reshape_progress;
523 /* Need a memory barrier to make sure we see the value
524 * of conf->generation, or ->data_offset that was set before
525 * reshape_progress was updated.
526 */
527 smp_rmb();
528 if (progress == MaxSector)
529 return 0;
530 if (sh->generation == conf->generation - 1)
531 return 0;
532 /* We are in a reshape, and this is a new-generation stripe,
533 * so use new_data_offset.
534 */
535 return 1;
536 }
537
538 static void
539 raid5_end_read_request(struct bio *bi, int error);
540 static void
541 raid5_end_write_request(struct bio *bi, int error);
542
543 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
544 {
545 struct r5conf *conf = sh->raid_conf;
546 int i, disks = sh->disks;
547
548 might_sleep();
549
550 for (i = disks; i--; ) {
551 int rw;
552 int replace_only = 0;
553 struct bio *bi, *rbi;
554 struct md_rdev *rdev, *rrdev = NULL;
555 if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
556 if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
557 rw = WRITE_FUA;
558 else
559 rw = WRITE;
560 if (test_bit(R5_Discard, &sh->dev[i].flags))
561 rw |= REQ_DISCARD;
562 } else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
563 rw = READ;
564 else if (test_and_clear_bit(R5_WantReplace,
565 &sh->dev[i].flags)) {
566 rw = WRITE;
567 replace_only = 1;
568 } else
569 continue;
570 if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
571 rw |= REQ_SYNC;
572
573 bi = &sh->dev[i].req;
574 rbi = &sh->dev[i].rreq; /* For writing to replacement */
575
576 rcu_read_lock();
577 rrdev = rcu_dereference(conf->disks[i].replacement);
578 smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
579 rdev = rcu_dereference(conf->disks[i].rdev);
580 if (!rdev) {
581 rdev = rrdev;
582 rrdev = NULL;
583 }
584 if (rw & WRITE) {
585 if (replace_only)
586 rdev = NULL;
587 if (rdev == rrdev)
588 /* We raced and saw duplicates */
589 rrdev = NULL;
590 } else {
591 if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
592 rdev = rrdev;
593 rrdev = NULL;
594 }
595
596 if (rdev && test_bit(Faulty, &rdev->flags))
597 rdev = NULL;
598 if (rdev)
599 atomic_inc(&rdev->nr_pending);
600 if (rrdev && test_bit(Faulty, &rrdev->flags))
601 rrdev = NULL;
602 if (rrdev)
603 atomic_inc(&rrdev->nr_pending);
604 rcu_read_unlock();
605
606 /* We have already checked bad blocks for reads. Now
607 * need to check for writes. We never accept write errors
608 * on the replacement, so we don't to check rrdev.
609 */
610 while ((rw & WRITE) && rdev &&
611 test_bit(WriteErrorSeen, &rdev->flags)) {
612 sector_t first_bad;
613 int bad_sectors;
614 int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
615 &first_bad, &bad_sectors);
616 if (!bad)
617 break;
618
619 if (bad < 0) {
620 set_bit(BlockedBadBlocks, &rdev->flags);
621 if (!conf->mddev->external &&
622 conf->mddev->flags) {
623 /* It is very unlikely, but we might
624 * still need to write out the
625 * bad block log - better give it
626 * a chance*/
627 md_check_recovery(conf->mddev);
628 }
629 /*
630 * Because md_wait_for_blocked_rdev
631 * will dec nr_pending, we must
632 * increment it first.
633 */
634 atomic_inc(&rdev->nr_pending);
635 md_wait_for_blocked_rdev(rdev, conf->mddev);
636 } else {
637 /* Acknowledged bad block - skip the write */
638 rdev_dec_pending(rdev, conf->mddev);
639 rdev = NULL;
640 }
641 }
642
643 if (rdev) {
644 if (s->syncing || s->expanding || s->expanded
645 || s->replacing)
646 md_sync_acct(rdev->bdev, STRIPE_SECTORS);
647
648 set_bit(STRIPE_IO_STARTED, &sh->state);
649
650 bio_reset(bi);
651 bi->bi_bdev = rdev->bdev;
652 bi->bi_rw = rw;
653 bi->bi_end_io = (rw & WRITE)
654 ? raid5_end_write_request
655 : raid5_end_read_request;
656 bi->bi_private = sh;
657
658 pr_debug("%s: for %llu schedule op %ld on disc %d\n",
659 __func__, (unsigned long long)sh->sector,
660 bi->bi_rw, i);
661 atomic_inc(&sh->count);
662 if (use_new_offset(conf, sh))
663 bi->bi_sector = (sh->sector
664 + rdev->new_data_offset);
665 else
666 bi->bi_sector = (sh->sector
667 + rdev->data_offset);
668 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
669 bi->bi_rw |= REQ_FLUSH;
670
671 bi->bi_vcnt = 1;
672 bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
673 bi->bi_io_vec[0].bv_offset = 0;
674 bi->bi_size = STRIPE_SIZE;
675 /*
676 * If this is discard request, set bi_vcnt 0. We don't
677 * want to confuse SCSI because SCSI will replace payload
678 */
679 if (rw & REQ_DISCARD)
680 bi->bi_vcnt = 0;
681 if (rrdev)
682 set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
683
684 if (conf->mddev->gendisk)
685 trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
686 bi, disk_devt(conf->mddev->gendisk),
687 sh->dev[i].sector);
688 generic_make_request(bi);
689 }
690 if (rrdev) {
691 if (s->syncing || s->expanding || s->expanded
692 || s->replacing)
693 md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
694
695 set_bit(STRIPE_IO_STARTED, &sh->state);
696
697 bio_reset(rbi);
698 rbi->bi_bdev = rrdev->bdev;
699 rbi->bi_rw = rw;
700 BUG_ON(!(rw & WRITE));
701 rbi->bi_end_io = raid5_end_write_request;
702 rbi->bi_private = sh;
703
704 pr_debug("%s: for %llu schedule op %ld on "
705 "replacement disc %d\n",
706 __func__, (unsigned long long)sh->sector,
707 rbi->bi_rw, i);
708 atomic_inc(&sh->count);
709 if (use_new_offset(conf, sh))
710 rbi->bi_sector = (sh->sector
711 + rrdev->new_data_offset);
712 else
713 rbi->bi_sector = (sh->sector
714 + rrdev->data_offset);
715 rbi->bi_vcnt = 1;
716 rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
717 rbi->bi_io_vec[0].bv_offset = 0;
718 rbi->bi_size = STRIPE_SIZE;
719 /*
720 * If this is discard request, set bi_vcnt 0. We don't
721 * want to confuse SCSI because SCSI will replace payload
722 */
723 if (rw & REQ_DISCARD)
724 rbi->bi_vcnt = 0;
725 if (conf->mddev->gendisk)
726 trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
727 rbi, disk_devt(conf->mddev->gendisk),
728 sh->dev[i].sector);
729 generic_make_request(rbi);
730 }
731 if (!rdev && !rrdev) {
732 if (rw & WRITE)
733 set_bit(STRIPE_DEGRADED, &sh->state);
734 pr_debug("skip op %ld on disc %d for sector %llu\n",
735 bi->bi_rw, i, (unsigned long long)sh->sector);
736 clear_bit(R5_LOCKED, &sh->dev[i].flags);
737 set_bit(STRIPE_HANDLE, &sh->state);
738 }
739 }
740 }
741
742 static struct dma_async_tx_descriptor *
743 async_copy_data(int frombio, struct bio *bio, struct page *page,
744 sector_t sector, struct dma_async_tx_descriptor *tx)
745 {
746 struct bio_vec *bvl;
747 struct page *bio_page;
748 int i;
749 int page_offset;
750 struct async_submit_ctl submit;
751 enum async_tx_flags flags = 0;
752
753 if (bio->bi_sector >= sector)
754 page_offset = (signed)(bio->bi_sector - sector) * 512;
755 else
756 page_offset = (signed)(sector - bio->bi_sector) * -512;
757
758 if (frombio)
759 flags |= ASYNC_TX_FENCE;
760 init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
761
762 bio_for_each_segment(bvl, bio, i) {
763 int len = bvl->bv_len;
764 int clen;
765 int b_offset = 0;
766
767 if (page_offset < 0) {
768 b_offset = -page_offset;
769 page_offset += b_offset;
770 len -= b_offset;
771 }
772
773 if (len > 0 && page_offset + len > STRIPE_SIZE)
774 clen = STRIPE_SIZE - page_offset;
775 else
776 clen = len;
777
778 if (clen > 0) {
779 b_offset += bvl->bv_offset;
780 bio_page = bvl->bv_page;
781 if (frombio)
782 tx = async_memcpy(page, bio_page, page_offset,
783 b_offset, clen, &submit);
784 else
785 tx = async_memcpy(bio_page, page, b_offset,
786 page_offset, clen, &submit);
787 }
788 /* chain the operations */
789 submit.depend_tx = tx;
790
791 if (clen < len) /* hit end of page */
792 break;
793 page_offset += len;
794 }
795
796 return tx;
797 }
798
799 static void ops_complete_biofill(void *stripe_head_ref)
800 {
801 struct stripe_head *sh = stripe_head_ref;
802 struct bio *return_bi = NULL;
803 int i;
804
805 pr_debug("%s: stripe %llu\n", __func__,
806 (unsigned long long)sh->sector);
807
808 /* clear completed biofills */
809 for (i = sh->disks; i--; ) {
810 struct r5dev *dev = &sh->dev[i];
811
812 /* acknowledge completion of a biofill operation */
813 /* and check if we need to reply to a read request,
814 * new R5_Wantfill requests are held off until
815 * !STRIPE_BIOFILL_RUN
816 */
817 if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
818 struct bio *rbi, *rbi2;
819
820 BUG_ON(!dev->read);
821 rbi = dev->read;
822 dev->read = NULL;
823 while (rbi && rbi->bi_sector <
824 dev->sector + STRIPE_SECTORS) {
825 rbi2 = r5_next_bio(rbi, dev->sector);
826 if (!raid5_dec_bi_active_stripes(rbi)) {
827 rbi->bi_next = return_bi;
828 return_bi = rbi;
829 }
830 rbi = rbi2;
831 }
832 }
833 }
834 clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
835
836 return_io(return_bi);
837
838 set_bit(STRIPE_HANDLE, &sh->state);
839 release_stripe(sh);
840 }
841
842 static void ops_run_biofill(struct stripe_head *sh)
843 {
844 struct dma_async_tx_descriptor *tx = NULL;
845 struct async_submit_ctl submit;
846 int i;
847
848 pr_debug("%s: stripe %llu\n", __func__,
849 (unsigned long long)sh->sector);
850
851 for (i = sh->disks; i--; ) {
852 struct r5dev *dev = &sh->dev[i];
853 if (test_bit(R5_Wantfill, &dev->flags)) {
854 struct bio *rbi;
855 spin_lock_irq(&sh->stripe_lock);
856 dev->read = rbi = dev->toread;
857 dev->toread = NULL;
858 spin_unlock_irq(&sh->stripe_lock);
859 while (rbi && rbi->bi_sector <
860 dev->sector + STRIPE_SECTORS) {
861 tx = async_copy_data(0, rbi, dev->page,
862 dev->sector, tx);
863 rbi = r5_next_bio(rbi, dev->sector);
864 }
865 }
866 }
867
868 atomic_inc(&sh->count);
869 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
870 async_trigger_callback(&submit);
871 }
872
873 static void mark_target_uptodate(struct stripe_head *sh, int target)
874 {
875 struct r5dev *tgt;
876
877 if (target < 0)
878 return;
879
880 tgt = &sh->dev[target];
881 set_bit(R5_UPTODATE, &tgt->flags);
882 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
883 clear_bit(R5_Wantcompute, &tgt->flags);
884 }
885
886 static void ops_complete_compute(void *stripe_head_ref)
887 {
888 struct stripe_head *sh = stripe_head_ref;
889
890 pr_debug("%s: stripe %llu\n", __func__,
891 (unsigned long long)sh->sector);
892
893 /* mark the computed target(s) as uptodate */
894 mark_target_uptodate(sh, sh->ops.target);
895 mark_target_uptodate(sh, sh->ops.target2);
896
897 clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
898 if (sh->check_state == check_state_compute_run)
899 sh->check_state = check_state_compute_result;
900 set_bit(STRIPE_HANDLE, &sh->state);
901 release_stripe(sh);
902 }
903
904 /* return a pointer to the address conversion region of the scribble buffer */
905 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
906 struct raid5_percpu *percpu)
907 {
908 return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
909 }
910
911 static struct dma_async_tx_descriptor *
912 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
913 {
914 int disks = sh->disks;
915 struct page **xor_srcs = percpu->scribble;
916 int target = sh->ops.target;
917 struct r5dev *tgt = &sh->dev[target];
918 struct page *xor_dest = tgt->page;
919 int count = 0;
920 struct dma_async_tx_descriptor *tx;
921 struct async_submit_ctl submit;
922 int i;
923
924 pr_debug("%s: stripe %llu block: %d\n",
925 __func__, (unsigned long long)sh->sector, target);
926 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
927
928 for (i = disks; i--; )
929 if (i != target)
930 xor_srcs[count++] = sh->dev[i].page;
931
932 atomic_inc(&sh->count);
933
934 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
935 ops_complete_compute, sh, to_addr_conv(sh, percpu));
936 if (unlikely(count == 1))
937 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
938 else
939 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
940
941 return tx;
942 }
943
944 /* set_syndrome_sources - populate source buffers for gen_syndrome
945 * @srcs - (struct page *) array of size sh->disks
946 * @sh - stripe_head to parse
947 *
948 * Populates srcs in proper layout order for the stripe and returns the
949 * 'count' of sources to be used in a call to async_gen_syndrome. The P
950 * destination buffer is recorded in srcs[count] and the Q destination
951 * is recorded in srcs[count+1]].
952 */
953 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
954 {
955 int disks = sh->disks;
956 int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
957 int d0_idx = raid6_d0(sh);
958 int count;
959 int i;
960
961 for (i = 0; i < disks; i++)
962 srcs[i] = NULL;
963
964 count = 0;
965 i = d0_idx;
966 do {
967 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
968
969 srcs[slot] = sh->dev[i].page;
970 i = raid6_next_disk(i, disks);
971 } while (i != d0_idx);
972
973 return syndrome_disks;
974 }
975
976 static struct dma_async_tx_descriptor *
977 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
978 {
979 int disks = sh->disks;
980 struct page **blocks = percpu->scribble;
981 int target;
982 int qd_idx = sh->qd_idx;
983 struct dma_async_tx_descriptor *tx;
984 struct async_submit_ctl submit;
985 struct r5dev *tgt;
986 struct page *dest;
987 int i;
988 int count;
989
990 if (sh->ops.target < 0)
991 target = sh->ops.target2;
992 else if (sh->ops.target2 < 0)
993 target = sh->ops.target;
994 else
995 /* we should only have one valid target */
996 BUG();
997 BUG_ON(target < 0);
998 pr_debug("%s: stripe %llu block: %d\n",
999 __func__, (unsigned long long)sh->sector, target);
1000
1001 tgt = &sh->dev[target];
1002 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1003 dest = tgt->page;
1004
1005 atomic_inc(&sh->count);
1006
1007 if (target == qd_idx) {
1008 count = set_syndrome_sources(blocks, sh);
1009 blocks[count] = NULL; /* regenerating p is not necessary */
1010 BUG_ON(blocks[count+1] != dest); /* q should already be set */
1011 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1012 ops_complete_compute, sh,
1013 to_addr_conv(sh, percpu));
1014 tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1015 } else {
1016 /* Compute any data- or p-drive using XOR */
1017 count = 0;
1018 for (i = disks; i-- ; ) {
1019 if (i == target || i == qd_idx)
1020 continue;
1021 blocks[count++] = sh->dev[i].page;
1022 }
1023
1024 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1025 NULL, ops_complete_compute, sh,
1026 to_addr_conv(sh, percpu));
1027 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1028 }
1029
1030 return tx;
1031 }
1032
1033 static struct dma_async_tx_descriptor *
1034 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1035 {
1036 int i, count, disks = sh->disks;
1037 int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1038 int d0_idx = raid6_d0(sh);
1039 int faila = -1, failb = -1;
1040 int target = sh->ops.target;
1041 int target2 = sh->ops.target2;
1042 struct r5dev *tgt = &sh->dev[target];
1043 struct r5dev *tgt2 = &sh->dev[target2];
1044 struct dma_async_tx_descriptor *tx;
1045 struct page **blocks = percpu->scribble;
1046 struct async_submit_ctl submit;
1047
1048 pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1049 __func__, (unsigned long long)sh->sector, target, target2);
1050 BUG_ON(target < 0 || target2 < 0);
1051 BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1052 BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1053
1054 /* we need to open-code set_syndrome_sources to handle the
1055 * slot number conversion for 'faila' and 'failb'
1056 */
1057 for (i = 0; i < disks ; i++)
1058 blocks[i] = NULL;
1059 count = 0;
1060 i = d0_idx;
1061 do {
1062 int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1063
1064 blocks[slot] = sh->dev[i].page;
1065
1066 if (i == target)
1067 faila = slot;
1068 if (i == target2)
1069 failb = slot;
1070 i = raid6_next_disk(i, disks);
1071 } while (i != d0_idx);
1072
1073 BUG_ON(faila == failb);
1074 if (failb < faila)
1075 swap(faila, failb);
1076 pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1077 __func__, (unsigned long long)sh->sector, faila, failb);
1078
1079 atomic_inc(&sh->count);
1080
1081 if (failb == syndrome_disks+1) {
1082 /* Q disk is one of the missing disks */
1083 if (faila == syndrome_disks) {
1084 /* Missing P+Q, just recompute */
1085 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1086 ops_complete_compute, sh,
1087 to_addr_conv(sh, percpu));
1088 return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1089 STRIPE_SIZE, &submit);
1090 } else {
1091 struct page *dest;
1092 int data_target;
1093 int qd_idx = sh->qd_idx;
1094
1095 /* Missing D+Q: recompute D from P, then recompute Q */
1096 if (target == qd_idx)
1097 data_target = target2;
1098 else
1099 data_target = target;
1100
1101 count = 0;
1102 for (i = disks; i-- ; ) {
1103 if (i == data_target || i == qd_idx)
1104 continue;
1105 blocks[count++] = sh->dev[i].page;
1106 }
1107 dest = sh->dev[data_target].page;
1108 init_async_submit(&submit,
1109 ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1110 NULL, NULL, NULL,
1111 to_addr_conv(sh, percpu));
1112 tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1113 &submit);
1114
1115 count = set_syndrome_sources(blocks, sh);
1116 init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1117 ops_complete_compute, sh,
1118 to_addr_conv(sh, percpu));
1119 return async_gen_syndrome(blocks, 0, count+2,
1120 STRIPE_SIZE, &submit);
1121 }
1122 } else {
1123 init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1124 ops_complete_compute, sh,
1125 to_addr_conv(sh, percpu));
1126 if (failb == syndrome_disks) {
1127 /* We're missing D+P. */
1128 return async_raid6_datap_recov(syndrome_disks+2,
1129 STRIPE_SIZE, faila,
1130 blocks, &submit);
1131 } else {
1132 /* We're missing D+D. */
1133 return async_raid6_2data_recov(syndrome_disks+2,
1134 STRIPE_SIZE, faila, failb,
1135 blocks, &submit);
1136 }
1137 }
1138 }
1139
1140
1141 static void ops_complete_prexor(void *stripe_head_ref)
1142 {
1143 struct stripe_head *sh = stripe_head_ref;
1144
1145 pr_debug("%s: stripe %llu\n", __func__,
1146 (unsigned long long)sh->sector);
1147 }
1148
1149 static struct dma_async_tx_descriptor *
1150 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1151 struct dma_async_tx_descriptor *tx)
1152 {
1153 int disks = sh->disks;
1154 struct page **xor_srcs = percpu->scribble;
1155 int count = 0, pd_idx = sh->pd_idx, i;
1156 struct async_submit_ctl submit;
1157
1158 /* existing parity data subtracted */
1159 struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1160
1161 pr_debug("%s: stripe %llu\n", __func__,
1162 (unsigned long long)sh->sector);
1163
1164 for (i = disks; i--; ) {
1165 struct r5dev *dev = &sh->dev[i];
1166 /* Only process blocks that are known to be uptodate */
1167 if (test_bit(R5_Wantdrain, &dev->flags))
1168 xor_srcs[count++] = dev->page;
1169 }
1170
1171 init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1172 ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1173 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1174
1175 return tx;
1176 }
1177
1178 static struct dma_async_tx_descriptor *
1179 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1180 {
1181 int disks = sh->disks;
1182 int i;
1183
1184 pr_debug("%s: stripe %llu\n", __func__,
1185 (unsigned long long)sh->sector);
1186
1187 for (i = disks; i--; ) {
1188 struct r5dev *dev = &sh->dev[i];
1189 struct bio *chosen;
1190
1191 if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1192 struct bio *wbi;
1193
1194 spin_lock_irq(&sh->stripe_lock);
1195 chosen = dev->towrite;
1196 dev->towrite = NULL;
1197 BUG_ON(dev->written);
1198 wbi = dev->written = chosen;
1199 spin_unlock_irq(&sh->stripe_lock);
1200
1201 while (wbi && wbi->bi_sector <
1202 dev->sector + STRIPE_SECTORS) {
1203 if (wbi->bi_rw & REQ_FUA)
1204 set_bit(R5_WantFUA, &dev->flags);
1205 if (wbi->bi_rw & REQ_SYNC)
1206 set_bit(R5_SyncIO, &dev->flags);
1207 if (wbi->bi_rw & REQ_DISCARD)
1208 set_bit(R5_Discard, &dev->flags);
1209 else
1210 tx = async_copy_data(1, wbi, dev->page,
1211 dev->sector, tx);
1212 wbi = r5_next_bio(wbi, dev->sector);
1213 }
1214 }
1215 }
1216
1217 return tx;
1218 }
1219
1220 static void ops_complete_reconstruct(void *stripe_head_ref)
1221 {
1222 struct stripe_head *sh = stripe_head_ref;
1223 int disks = sh->disks;
1224 int pd_idx = sh->pd_idx;
1225 int qd_idx = sh->qd_idx;
1226 int i;
1227 bool fua = false, sync = false, discard = false;
1228
1229 pr_debug("%s: stripe %llu\n", __func__,
1230 (unsigned long long)sh->sector);
1231
1232 for (i = disks; i--; ) {
1233 fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1234 sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1235 discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1236 }
1237
1238 for (i = disks; i--; ) {
1239 struct r5dev *dev = &sh->dev[i];
1240
1241 if (dev->written || i == pd_idx || i == qd_idx) {
1242 if (!discard)
1243 set_bit(R5_UPTODATE, &dev->flags);
1244 if (fua)
1245 set_bit(R5_WantFUA, &dev->flags);
1246 if (sync)
1247 set_bit(R5_SyncIO, &dev->flags);
1248 }
1249 }
1250
1251 if (sh->reconstruct_state == reconstruct_state_drain_run)
1252 sh->reconstruct_state = reconstruct_state_drain_result;
1253 else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1254 sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1255 else {
1256 BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1257 sh->reconstruct_state = reconstruct_state_result;
1258 }
1259
1260 set_bit(STRIPE_HANDLE, &sh->state);
1261 release_stripe(sh);
1262 }
1263
1264 static void
1265 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1266 struct dma_async_tx_descriptor *tx)
1267 {
1268 int disks = sh->disks;
1269 struct page **xor_srcs = percpu->scribble;
1270 struct async_submit_ctl submit;
1271 int count = 0, pd_idx = sh->pd_idx, i;
1272 struct page *xor_dest;
1273 int prexor = 0;
1274 unsigned long flags;
1275
1276 pr_debug("%s: stripe %llu\n", __func__,
1277 (unsigned long long)sh->sector);
1278
1279 for (i = 0; i < sh->disks; i++) {
1280 if (pd_idx == i)
1281 continue;
1282 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1283 break;
1284 }
1285 if (i >= sh->disks) {
1286 atomic_inc(&sh->count);
1287 set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1288 ops_complete_reconstruct(sh);
1289 return;
1290 }
1291 /* check if prexor is active which means only process blocks
1292 * that are part of a read-modify-write (written)
1293 */
1294 if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1295 prexor = 1;
1296 xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1297 for (i = disks; i--; ) {
1298 struct r5dev *dev = &sh->dev[i];
1299 if (dev->written)
1300 xor_srcs[count++] = dev->page;
1301 }
1302 } else {
1303 xor_dest = sh->dev[pd_idx].page;
1304 for (i = disks; i--; ) {
1305 struct r5dev *dev = &sh->dev[i];
1306 if (i != pd_idx)
1307 xor_srcs[count++] = dev->page;
1308 }
1309 }
1310
1311 /* 1/ if we prexor'd then the dest is reused as a source
1312 * 2/ if we did not prexor then we are redoing the parity
1313 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1314 * for the synchronous xor case
1315 */
1316 flags = ASYNC_TX_ACK |
1317 (prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1318
1319 atomic_inc(&sh->count);
1320
1321 init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1322 to_addr_conv(sh, percpu));
1323 if (unlikely(count == 1))
1324 tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1325 else
1326 tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1327 }
1328
1329 static void
1330 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1331 struct dma_async_tx_descriptor *tx)
1332 {
1333 struct async_submit_ctl submit;
1334 struct page **blocks = percpu->scribble;
1335 int count, i;
1336
1337 pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1338
1339 for (i = 0; i < sh->disks; i++) {
1340 if (sh->pd_idx == i || sh->qd_idx == i)
1341 continue;
1342 if (!test_bit(R5_Discard, &sh->dev[i].flags))
1343 break;
1344 }
1345 if (i >= sh->disks) {
1346 atomic_inc(&sh->count);
1347 set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1348 set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1349 ops_complete_reconstruct(sh);
1350 return;
1351 }
1352
1353 count = set_syndrome_sources(blocks, sh);
1354
1355 atomic_inc(&sh->count);
1356
1357 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1358 sh, to_addr_conv(sh, percpu));
1359 async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1360 }
1361
1362 static void ops_complete_check(void *stripe_head_ref)
1363 {
1364 struct stripe_head *sh = stripe_head_ref;
1365
1366 pr_debug("%s: stripe %llu\n", __func__,
1367 (unsigned long long)sh->sector);
1368
1369 sh->check_state = check_state_check_result;
1370 set_bit(STRIPE_HANDLE, &sh->state);
1371 release_stripe(sh);
1372 }
1373
1374 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1375 {
1376 int disks = sh->disks;
1377 int pd_idx = sh->pd_idx;
1378 int qd_idx = sh->qd_idx;
1379 struct page *xor_dest;
1380 struct page **xor_srcs = percpu->scribble;
1381 struct dma_async_tx_descriptor *tx;
1382 struct async_submit_ctl submit;
1383 int count;
1384 int i;
1385
1386 pr_debug("%s: stripe %llu\n", __func__,
1387 (unsigned long long)sh->sector);
1388
1389 count = 0;
1390 xor_dest = sh->dev[pd_idx].page;
1391 xor_srcs[count++] = xor_dest;
1392 for (i = disks; i--; ) {
1393 if (i == pd_idx || i == qd_idx)
1394 continue;
1395 xor_srcs[count++] = sh->dev[i].page;
1396 }
1397
1398 init_async_submit(&submit, 0, NULL, NULL, NULL,
1399 to_addr_conv(sh, percpu));
1400 tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1401 &sh->ops.zero_sum_result, &submit);
1402
1403 atomic_inc(&sh->count);
1404 init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1405 tx = async_trigger_callback(&submit);
1406 }
1407
1408 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1409 {
1410 struct page **srcs = percpu->scribble;
1411 struct async_submit_ctl submit;
1412 int count;
1413
1414 pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1415 (unsigned long long)sh->sector, checkp);
1416
1417 count = set_syndrome_sources(srcs, sh);
1418 if (!checkp)
1419 srcs[count] = NULL;
1420
1421 atomic_inc(&sh->count);
1422 init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1423 sh, to_addr_conv(sh, percpu));
1424 async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1425 &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1426 }
1427
1428 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1429 {
1430 int overlap_clear = 0, i, disks = sh->disks;
1431 struct dma_async_tx_descriptor *tx = NULL;
1432 struct r5conf *conf = sh->raid_conf;
1433 int level = conf->level;
1434 struct raid5_percpu *percpu;
1435 unsigned long cpu;
1436
1437 cpu = get_cpu();
1438 percpu = per_cpu_ptr(conf->percpu, cpu);
1439 if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1440 ops_run_biofill(sh);
1441 overlap_clear++;
1442 }
1443
1444 if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1445 if (level < 6)
1446 tx = ops_run_compute5(sh, percpu);
1447 else {
1448 if (sh->ops.target2 < 0 || sh->ops.target < 0)
1449 tx = ops_run_compute6_1(sh, percpu);
1450 else
1451 tx = ops_run_compute6_2(sh, percpu);
1452 }
1453 /* terminate the chain if reconstruct is not set to be run */
1454 if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1455 async_tx_ack(tx);
1456 }
1457
1458 if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1459 tx = ops_run_prexor(sh, percpu, tx);
1460
1461 if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1462 tx = ops_run_biodrain(sh, tx);
1463 overlap_clear++;
1464 }
1465
1466 if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1467 if (level < 6)
1468 ops_run_reconstruct5(sh, percpu, tx);
1469 else
1470 ops_run_reconstruct6(sh, percpu, tx);
1471 }
1472
1473 if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1474 if (sh->check_state == check_state_run)
1475 ops_run_check_p(sh, percpu);
1476 else if (sh->check_state == check_state_run_q)
1477 ops_run_check_pq(sh, percpu, 0);
1478 else if (sh->check_state == check_state_run_pq)
1479 ops_run_check_pq(sh, percpu, 1);
1480 else
1481 BUG();
1482 }
1483
1484 if (overlap_clear)
1485 for (i = disks; i--; ) {
1486 struct r5dev *dev = &sh->dev[i];
1487 if (test_and_clear_bit(R5_Overlap, &dev->flags))
1488 wake_up(&sh->raid_conf->wait_for_overlap);
1489 }
1490 put_cpu();
1491 }
1492
1493 static int grow_one_stripe(struct r5conf *conf)
1494 {
1495 struct stripe_head *sh;
1496 sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1497 if (!sh)
1498 return 0;
1499
1500 sh->raid_conf = conf;
1501
1502 spin_lock_init(&sh->stripe_lock);
1503
1504 if (grow_buffers(sh)) {
1505 shrink_buffers(sh);
1506 kmem_cache_free(conf->slab_cache, sh);
1507 return 0;
1508 }
1509 /* we just created an active stripe so... */
1510 atomic_set(&sh->count, 1);
1511 atomic_inc(&conf->active_stripes);
1512 INIT_LIST_HEAD(&sh->lru);
1513 release_stripe(sh);
1514 return 1;
1515 }
1516
1517 static int grow_stripes(struct r5conf *conf, int num)
1518 {
1519 struct kmem_cache *sc;
1520 int devs = max(conf->raid_disks, conf->previous_raid_disks);
1521
1522 if (conf->mddev->gendisk)
1523 sprintf(conf->cache_name[0],
1524 "raid%d-%s", conf->level, mdname(conf->mddev));
1525 else
1526 sprintf(conf->cache_name[0],
1527 "raid%d-%p", conf->level, conf->mddev);
1528 sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1529
1530 conf->active_name = 0;
1531 sc = kmem_cache_create(conf->cache_name[conf->active_name],
1532 sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1533 0, 0, NULL);
1534 if (!sc)
1535 return 1;
1536 conf->slab_cache = sc;
1537 conf->pool_size = devs;
1538 while (num--)
1539 if (!grow_one_stripe(conf))
1540 return 1;
1541 return 0;
1542 }
1543
1544 /**
1545 * scribble_len - return the required size of the scribble region
1546 * @num - total number of disks in the array
1547 *
1548 * The size must be enough to contain:
1549 * 1/ a struct page pointer for each device in the array +2
1550 * 2/ room to convert each entry in (1) to its corresponding dma
1551 * (dma_map_page()) or page (page_address()) address.
1552 *
1553 * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1554 * calculate over all devices (not just the data blocks), using zeros in place
1555 * of the P and Q blocks.
1556 */
1557 static size_t scribble_len(int num)
1558 {
1559 size_t len;
1560
1561 len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1562
1563 return len;
1564 }
1565
1566 static int resize_stripes(struct r5conf *conf, int newsize)
1567 {
1568 /* Make all the stripes able to hold 'newsize' devices.
1569 * New slots in each stripe get 'page' set to a new page.
1570 *
1571 * This happens in stages:
1572 * 1/ create a new kmem_cache and allocate the required number of
1573 * stripe_heads.
1574 * 2/ gather all the old stripe_heads and transfer the pages across
1575 * to the new stripe_heads. This will have the side effect of
1576 * freezing the array as once all stripe_heads have been collected,
1577 * no IO will be possible. Old stripe heads are freed once their
1578 * pages have been transferred over, and the old kmem_cache is
1579 * freed when all stripes are done.
1580 * 3/ reallocate conf->disks to be suitable bigger. If this fails,
1581 * we simple return a failre status - no need to clean anything up.
1582 * 4/ allocate new pages for the new slots in the new stripe_heads.
1583 * If this fails, we don't bother trying the shrink the
1584 * stripe_heads down again, we just leave them as they are.
1585 * As each stripe_head is processed the new one is released into
1586 * active service.
1587 *
1588 * Once step2 is started, we cannot afford to wait for a write,
1589 * so we use GFP_NOIO allocations.
1590 */
1591 struct stripe_head *osh, *nsh;
1592 LIST_HEAD(newstripes);
1593 struct disk_info *ndisks;
1594 unsigned long cpu;
1595 int err;
1596 struct kmem_cache *sc;
1597 int i;
1598
1599 if (newsize <= conf->pool_size)
1600 return 0; /* never bother to shrink */
1601
1602 err = md_allow_write(conf->mddev);
1603 if (err)
1604 return err;
1605
1606 /* Step 1 */
1607 sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1608 sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1609 0, 0, NULL);
1610 if (!sc)
1611 return -ENOMEM;
1612
1613 for (i = conf->max_nr_stripes; i; i--) {
1614 nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1615 if (!nsh)
1616 break;
1617
1618 nsh->raid_conf = conf;
1619 spin_lock_init(&nsh->stripe_lock);
1620
1621 list_add(&nsh->lru, &newstripes);
1622 }
1623 if (i) {
1624 /* didn't get enough, give up */
1625 while (!list_empty(&newstripes)) {
1626 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1627 list_del(&nsh->lru);
1628 kmem_cache_free(sc, nsh);
1629 }
1630 kmem_cache_destroy(sc);
1631 return -ENOMEM;
1632 }
1633 /* Step 2 - Must use GFP_NOIO now.
1634 * OK, we have enough stripes, start collecting inactive
1635 * stripes and copying them over
1636 */
1637 list_for_each_entry(nsh, &newstripes, lru) {
1638 spin_lock_irq(&conf->device_lock);
1639 wait_event_lock_irq(conf->wait_for_stripe,
1640 !list_empty(&conf->inactive_list),
1641 conf->device_lock);
1642 osh = get_free_stripe(conf);
1643 spin_unlock_irq(&conf->device_lock);
1644 atomic_set(&nsh->count, 1);
1645 for(i=0; i<conf->pool_size; i++)
1646 nsh->dev[i].page = osh->dev[i].page;
1647 for( ; i<newsize; i++)
1648 nsh->dev[i].page = NULL;
1649 kmem_cache_free(conf->slab_cache, osh);
1650 }
1651 kmem_cache_destroy(conf->slab_cache);
1652
1653 /* Step 3.
1654 * At this point, we are holding all the stripes so the array
1655 * is completely stalled, so now is a good time to resize
1656 * conf->disks and the scribble region
1657 */
1658 ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1659 if (ndisks) {
1660 for (i=0; i<conf->raid_disks; i++)
1661 ndisks[i] = conf->disks[i];
1662 kfree(conf->disks);
1663 conf->disks = ndisks;
1664 } else
1665 err = -ENOMEM;
1666
1667 get_online_cpus();
1668 conf->scribble_len = scribble_len(newsize);
1669 for_each_present_cpu(cpu) {
1670 struct raid5_percpu *percpu;
1671 void *scribble;
1672
1673 percpu = per_cpu_ptr(conf->percpu, cpu);
1674 scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1675
1676 if (scribble) {
1677 kfree(percpu->scribble);
1678 percpu->scribble = scribble;
1679 } else {
1680 err = -ENOMEM;
1681 break;
1682 }
1683 }
1684 put_online_cpus();
1685
1686 /* Step 4, return new stripes to service */
1687 while(!list_empty(&newstripes)) {
1688 nsh = list_entry(newstripes.next, struct stripe_head, lru);
1689 list_del_init(&nsh->lru);
1690
1691 for (i=conf->raid_disks; i < newsize; i++)
1692 if (nsh->dev[i].page == NULL) {
1693 struct page *p = alloc_page(GFP_NOIO);
1694 nsh->dev[i].page = p;
1695 if (!p)
1696 err = -ENOMEM;
1697 }
1698 release_stripe(nsh);
1699 }
1700 /* critical section pass, GFP_NOIO no longer needed */
1701
1702 conf->slab_cache = sc;
1703 conf->active_name = 1-conf->active_name;
1704 if (!err)
1705 conf->pool_size = newsize;
1706 return err;
1707 }
1708
1709 static int drop_one_stripe(struct r5conf *conf)
1710 {
1711 struct stripe_head *sh;
1712
1713 spin_lock_irq(&conf->device_lock);
1714 sh = get_free_stripe(conf);
1715 spin_unlock_irq(&conf->device_lock);
1716 if (!sh)
1717 return 0;
1718 BUG_ON(atomic_read(&sh->count));
1719 shrink_buffers(sh);
1720 kmem_cache_free(conf->slab_cache, sh);
1721 atomic_dec(&conf->active_stripes);
1722 return 1;
1723 }
1724
1725 static void shrink_stripes(struct r5conf *conf)
1726 {
1727 while (drop_one_stripe(conf))
1728 ;
1729
1730 if (conf->slab_cache)
1731 kmem_cache_destroy(conf->slab_cache);
1732 conf->slab_cache = NULL;
1733 }
1734
1735 static void raid5_end_read_request(struct bio * bi, int error)
1736 {
1737 struct stripe_head *sh = bi->bi_private;
1738 struct r5conf *conf = sh->raid_conf;
1739 int disks = sh->disks, i;
1740 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1741 char b[BDEVNAME_SIZE];
1742 struct md_rdev *rdev = NULL;
1743 sector_t s;
1744
1745 for (i=0 ; i<disks; i++)
1746 if (bi == &sh->dev[i].req)
1747 break;
1748
1749 pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1750 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1751 uptodate);
1752 if (i == disks) {
1753 BUG();
1754 return;
1755 }
1756 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1757 /* If replacement finished while this request was outstanding,
1758 * 'replacement' might be NULL already.
1759 * In that case it moved down to 'rdev'.
1760 * rdev is not removed until all requests are finished.
1761 */
1762 rdev = conf->disks[i].replacement;
1763 if (!rdev)
1764 rdev = conf->disks[i].rdev;
1765
1766 if (use_new_offset(conf, sh))
1767 s = sh->sector + rdev->new_data_offset;
1768 else
1769 s = sh->sector + rdev->data_offset;
1770 if (uptodate) {
1771 set_bit(R5_UPTODATE, &sh->dev[i].flags);
1772 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1773 /* Note that this cannot happen on a
1774 * replacement device. We just fail those on
1775 * any error
1776 */
1777 printk_ratelimited(
1778 KERN_INFO
1779 "md/raid:%s: read error corrected"
1780 " (%lu sectors at %llu on %s)\n",
1781 mdname(conf->mddev), STRIPE_SECTORS,
1782 (unsigned long long)s,
1783 bdevname(rdev->bdev, b));
1784 atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1785 clear_bit(R5_ReadError, &sh->dev[i].flags);
1786 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1787 } else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1788 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1789
1790 if (atomic_read(&rdev->read_errors))
1791 atomic_set(&rdev->read_errors, 0);
1792 } else {
1793 const char *bdn = bdevname(rdev->bdev, b);
1794 int retry = 0;
1795 int set_bad = 0;
1796
1797 clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1798 atomic_inc(&rdev->read_errors);
1799 if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1800 printk_ratelimited(
1801 KERN_WARNING
1802 "md/raid:%s: read error on replacement device "
1803 "(sector %llu on %s).\n",
1804 mdname(conf->mddev),
1805 (unsigned long long)s,
1806 bdn);
1807 else if (conf->mddev->degraded >= conf->max_degraded) {
1808 set_bad = 1;
1809 printk_ratelimited(
1810 KERN_WARNING
1811 "md/raid:%s: read error not correctable "
1812 "(sector %llu on %s).\n",
1813 mdname(conf->mddev),
1814 (unsigned long long)s,
1815 bdn);
1816 } else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1817 /* Oh, no!!! */
1818 set_bad = 1;
1819 printk_ratelimited(
1820 KERN_WARNING
1821 "md/raid:%s: read error NOT corrected!! "
1822 "(sector %llu on %s).\n",
1823 mdname(conf->mddev),
1824 (unsigned long long)s,
1825 bdn);
1826 } else if (atomic_read(&rdev->read_errors)
1827 > conf->max_nr_stripes)
1828 printk(KERN_WARNING
1829 "md/raid:%s: Too many read errors, failing device %s.\n",
1830 mdname(conf->mddev), bdn);
1831 else
1832 retry = 1;
1833 if (retry)
1834 if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1835 set_bit(R5_ReadError, &sh->dev[i].flags);
1836 clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1837 } else
1838 set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1839 else {
1840 clear_bit(R5_ReadError, &sh->dev[i].flags);
1841 clear_bit(R5_ReWrite, &sh->dev[i].flags);
1842 if (!(set_bad
1843 && test_bit(In_sync, &rdev->flags)
1844 && rdev_set_badblocks(
1845 rdev, sh->sector, STRIPE_SECTORS, 0)))
1846 md_error(conf->mddev, rdev);
1847 }
1848 }
1849 rdev_dec_pending(rdev, conf->mddev);
1850 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1851 set_bit(STRIPE_HANDLE, &sh->state);
1852 release_stripe(sh);
1853 }
1854
1855 static void raid5_end_write_request(struct bio *bi, int error)
1856 {
1857 struct stripe_head *sh = bi->bi_private;
1858 struct r5conf *conf = sh->raid_conf;
1859 int disks = sh->disks, i;
1860 struct md_rdev *uninitialized_var(rdev);
1861 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1862 sector_t first_bad;
1863 int bad_sectors;
1864 int replacement = 0;
1865
1866 for (i = 0 ; i < disks; i++) {
1867 if (bi == &sh->dev[i].req) {
1868 rdev = conf->disks[i].rdev;
1869 break;
1870 }
1871 if (bi == &sh->dev[i].rreq) {
1872 rdev = conf->disks[i].replacement;
1873 if (rdev)
1874 replacement = 1;
1875 else
1876 /* rdev was removed and 'replacement'
1877 * replaced it. rdev is not removed
1878 * until all requests are finished.
1879 */
1880 rdev = conf->disks[i].rdev;
1881 break;
1882 }
1883 }
1884 pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1885 (unsigned long long)sh->sector, i, atomic_read(&sh->count),
1886 uptodate);
1887 if (i == disks) {
1888 BUG();
1889 return;
1890 }
1891
1892 if (replacement) {
1893 if (!uptodate)
1894 md_error(conf->mddev, rdev);
1895 else if (is_badblock(rdev, sh->sector,
1896 STRIPE_SECTORS,
1897 &first_bad, &bad_sectors))
1898 set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1899 } else {
1900 if (!uptodate) {
1901 set_bit(STRIPE_DEGRADED, &sh->state);
1902 set_bit(WriteErrorSeen, &rdev->flags);
1903 set_bit(R5_WriteError, &sh->dev[i].flags);
1904 if (!test_and_set_bit(WantReplacement, &rdev->flags))
1905 set_bit(MD_RECOVERY_NEEDED,
1906 &rdev->mddev->recovery);
1907 } else if (is_badblock(rdev, sh->sector,
1908 STRIPE_SECTORS,
1909 &first_bad, &bad_sectors)) {
1910 set_bit(R5_MadeGood, &sh->dev[i].flags);
1911 if (test_bit(R5_ReadError, &sh->dev[i].flags))
1912 /* That was a successful write so make
1913 * sure it looks like we already did
1914 * a re-write.
1915 */
1916 set_bit(R5_ReWrite, &sh->dev[i].flags);
1917 }
1918 }
1919 rdev_dec_pending(rdev, conf->mddev);
1920
1921 if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1922 clear_bit(R5_LOCKED, &sh->dev[i].flags);
1923 set_bit(STRIPE_HANDLE, &sh->state);
1924 release_stripe(sh);
1925 }
1926
1927 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1928
1929 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1930 {
1931 struct r5dev *dev = &sh->dev[i];
1932
1933 bio_init(&dev->req);
1934 dev->req.bi_io_vec = &dev->vec;
1935 dev->req.bi_vcnt++;
1936 dev->req.bi_max_vecs++;
1937 dev->req.bi_private = sh;
1938 dev->vec.bv_page = dev->page;
1939
1940 bio_init(&dev->rreq);
1941 dev->rreq.bi_io_vec = &dev->rvec;
1942 dev->rreq.bi_vcnt++;
1943 dev->rreq.bi_max_vecs++;
1944 dev->rreq.bi_private = sh;
1945 dev->rvec.bv_page = dev->page;
1946
1947 dev->flags = 0;
1948 dev->sector = compute_blocknr(sh, i, previous);
1949 }
1950
1951 static void error(struct mddev *mddev, struct md_rdev *rdev)
1952 {
1953 char b[BDEVNAME_SIZE];
1954 struct r5conf *conf = mddev->private;
1955 unsigned long flags;
1956 pr_debug("raid456: error called\n");
1957
1958 spin_lock_irqsave(&conf->device_lock, flags);
1959 clear_bit(In_sync, &rdev->flags);
1960 mddev->degraded = calc_degraded(conf);
1961 spin_unlock_irqrestore(&conf->device_lock, flags);
1962 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1963
1964 set_bit(Blocked, &rdev->flags);
1965 set_bit(Faulty, &rdev->flags);
1966 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1967 printk(KERN_ALERT
1968 "md/raid:%s: Disk failure on %s, disabling device.\n"
1969 "md/raid:%s: Operation continuing on %d devices.\n",
1970 mdname(mddev),
1971 bdevname(rdev->bdev, b),
1972 mdname(mddev),
1973 conf->raid_disks - mddev->degraded);
1974 }
1975
1976 /*
1977 * Input: a 'big' sector number,
1978 * Output: index of the data and parity disk, and the sector # in them.
1979 */
1980 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1981 int previous, int *dd_idx,
1982 struct stripe_head *sh)
1983 {
1984 sector_t stripe, stripe2;
1985 sector_t chunk_number;
1986 unsigned int chunk_offset;
1987 int pd_idx, qd_idx;
1988 int ddf_layout = 0;
1989 sector_t new_sector;
1990 int algorithm = previous ? conf->prev_algo
1991 : conf->algorithm;
1992 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1993 : conf->chunk_sectors;
1994 int raid_disks = previous ? conf->previous_raid_disks
1995 : conf->raid_disks;
1996 int data_disks = raid_disks - conf->max_degraded;
1997
1998 /* First compute the information on this sector */
1999
2000 /*
2001 * Compute the chunk number and the sector offset inside the chunk
2002 */
2003 chunk_offset = sector_div(r_sector, sectors_per_chunk);
2004 chunk_number = r_sector;
2005
2006 /*
2007 * Compute the stripe number
2008 */
2009 stripe = chunk_number;
2010 *dd_idx = sector_div(stripe, data_disks);
2011 stripe2 = stripe;
2012 /*
2013 * Select the parity disk based on the user selected algorithm.
2014 */
2015 pd_idx = qd_idx = -1;
2016 switch(conf->level) {
2017 case 4:
2018 pd_idx = data_disks;
2019 break;
2020 case 5:
2021 switch (algorithm) {
2022 case ALGORITHM_LEFT_ASYMMETRIC:
2023 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2024 if (*dd_idx >= pd_idx)
2025 (*dd_idx)++;
2026 break;
2027 case ALGORITHM_RIGHT_ASYMMETRIC:
2028 pd_idx = sector_div(stripe2, raid_disks);
2029 if (*dd_idx >= pd_idx)
2030 (*dd_idx)++;
2031 break;
2032 case ALGORITHM_LEFT_SYMMETRIC:
2033 pd_idx = data_disks - sector_div(stripe2, raid_disks);
2034 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2035 break;
2036 case ALGORITHM_RIGHT_SYMMETRIC:
2037 pd_idx = sector_div(stripe2, raid_disks);
2038 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2039 break;
2040 case ALGORITHM_PARITY_0:
2041 pd_idx = 0;
2042 (*dd_idx)++;
2043 break;
2044 case ALGORITHM_PARITY_N:
2045 pd_idx = data_disks;
2046 break;
2047 default:
2048 BUG();
2049 }
2050 break;
2051 case 6:
2052
2053 switch (algorithm) {
2054 case ALGORITHM_LEFT_ASYMMETRIC:
2055 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2056 qd_idx = pd_idx + 1;
2057 if (pd_idx == raid_disks-1) {
2058 (*dd_idx)++; /* Q D D D P */
2059 qd_idx = 0;
2060 } else if (*dd_idx >= pd_idx)
2061 (*dd_idx) += 2; /* D D P Q D */
2062 break;
2063 case ALGORITHM_RIGHT_ASYMMETRIC:
2064 pd_idx = sector_div(stripe2, raid_disks);
2065 qd_idx = pd_idx + 1;
2066 if (pd_idx == raid_disks-1) {
2067 (*dd_idx)++; /* Q D D D P */
2068 qd_idx = 0;
2069 } else if (*dd_idx >= pd_idx)
2070 (*dd_idx) += 2; /* D D P Q D */
2071 break;
2072 case ALGORITHM_LEFT_SYMMETRIC:
2073 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2074 qd_idx = (pd_idx + 1) % raid_disks;
2075 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2076 break;
2077 case ALGORITHM_RIGHT_SYMMETRIC:
2078 pd_idx = sector_div(stripe2, raid_disks);
2079 qd_idx = (pd_idx + 1) % raid_disks;
2080 *dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2081 break;
2082
2083 case ALGORITHM_PARITY_0:
2084 pd_idx = 0;
2085 qd_idx = 1;
2086 (*dd_idx) += 2;
2087 break;
2088 case ALGORITHM_PARITY_N:
2089 pd_idx = data_disks;
2090 qd_idx = data_disks + 1;
2091 break;
2092
2093 case ALGORITHM_ROTATING_ZERO_RESTART:
2094 /* Exactly the same as RIGHT_ASYMMETRIC, but or
2095 * of blocks for computing Q is different.
2096 */
2097 pd_idx = sector_div(stripe2, raid_disks);
2098 qd_idx = pd_idx + 1;
2099 if (pd_idx == raid_disks-1) {
2100 (*dd_idx)++; /* Q D D D P */
2101 qd_idx = 0;
2102 } else if (*dd_idx >= pd_idx)
2103 (*dd_idx) += 2; /* D D P Q D */
2104 ddf_layout = 1;
2105 break;
2106
2107 case ALGORITHM_ROTATING_N_RESTART:
2108 /* Same a left_asymmetric, by first stripe is
2109 * D D D P Q rather than
2110 * Q D D D P
2111 */
2112 stripe2 += 1;
2113 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2114 qd_idx = pd_idx + 1;
2115 if (pd_idx == raid_disks-1) {
2116 (*dd_idx)++; /* Q D D D P */
2117 qd_idx = 0;
2118 } else if (*dd_idx >= pd_idx)
2119 (*dd_idx) += 2; /* D D P Q D */
2120 ddf_layout = 1;
2121 break;
2122
2123 case ALGORITHM_ROTATING_N_CONTINUE:
2124 /* Same as left_symmetric but Q is before P */
2125 pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2126 qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2127 *dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2128 ddf_layout = 1;
2129 break;
2130
2131 case ALGORITHM_LEFT_ASYMMETRIC_6:
2132 /* RAID5 left_asymmetric, with Q on last device */
2133 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2134 if (*dd_idx >= pd_idx)
2135 (*dd_idx)++;
2136 qd_idx = raid_disks - 1;
2137 break;
2138
2139 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2140 pd_idx = sector_div(stripe2, raid_disks-1);
2141 if (*dd_idx >= pd_idx)
2142 (*dd_idx)++;
2143 qd_idx = raid_disks - 1;
2144 break;
2145
2146 case ALGORITHM_LEFT_SYMMETRIC_6:
2147 pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2148 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2149 qd_idx = raid_disks - 1;
2150 break;
2151
2152 case ALGORITHM_RIGHT_SYMMETRIC_6:
2153 pd_idx = sector_div(stripe2, raid_disks-1);
2154 *dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2155 qd_idx = raid_disks - 1;
2156 break;
2157
2158 case ALGORITHM_PARITY_0_6:
2159 pd_idx = 0;
2160 (*dd_idx)++;
2161 qd_idx = raid_disks - 1;
2162 break;
2163
2164 default:
2165 BUG();
2166 }
2167 break;
2168 }
2169
2170 if (sh) {
2171 sh->pd_idx = pd_idx;
2172 sh->qd_idx = qd_idx;
2173 sh->ddf_layout = ddf_layout;
2174 }
2175 /*
2176 * Finally, compute the new sector number
2177 */
2178 new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2179 return new_sector;
2180 }
2181
2182
2183 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2184 {
2185 struct r5conf *conf = sh->raid_conf;
2186 int raid_disks = sh->disks;
2187 int data_disks = raid_disks - conf->max_degraded;
2188 sector_t new_sector = sh->sector, check;
2189 int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2190 : conf->chunk_sectors;
2191 int algorithm = previous ? conf->prev_algo
2192 : conf->algorithm;
2193 sector_t stripe;
2194 int chunk_offset;
2195 sector_t chunk_number;
2196 int dummy1, dd_idx = i;
2197 sector_t r_sector;
2198 struct stripe_head sh2;
2199
2200
2201 chunk_offset = sector_div(new_sector, sectors_per_chunk);
2202 stripe = new_sector;
2203
2204 if (i == sh->pd_idx)
2205 return 0;
2206 switch(conf->level) {
2207 case 4: break;
2208 case 5:
2209 switch (algorithm) {
2210 case ALGORITHM_LEFT_ASYMMETRIC:
2211 case ALGORITHM_RIGHT_ASYMMETRIC:
2212 if (i > sh->pd_idx)
2213 i--;
2214 break;
2215 case ALGORITHM_LEFT_SYMMETRIC:
2216 case ALGORITHM_RIGHT_SYMMETRIC:
2217 if (i < sh->pd_idx)
2218 i += raid_disks;
2219 i -= (sh->pd_idx + 1);
2220 break;
2221 case ALGORITHM_PARITY_0:
2222 i -= 1;
2223 break;
2224 case ALGORITHM_PARITY_N:
2225 break;
2226 default:
2227 BUG();
2228 }
2229 break;
2230 case 6:
2231 if (i == sh->qd_idx)
2232 return 0; /* It is the Q disk */
2233 switch (algorithm) {
2234 case ALGORITHM_LEFT_ASYMMETRIC:
2235 case ALGORITHM_RIGHT_ASYMMETRIC:
2236 case ALGORITHM_ROTATING_ZERO_RESTART:
2237 case ALGORITHM_ROTATING_N_RESTART:
2238 if (sh->pd_idx == raid_disks-1)
2239 i--; /* Q D D D P */
2240 else if (i > sh->pd_idx)
2241 i -= 2; /* D D P Q D */
2242 break;
2243 case ALGORITHM_LEFT_SYMMETRIC:
2244 case ALGORITHM_RIGHT_SYMMETRIC:
2245 if (sh->pd_idx == raid_disks-1)
2246 i--; /* Q D D D P */
2247 else {
2248 /* D D P Q D */
2249 if (i < sh->pd_idx)
2250 i += raid_disks;
2251 i -= (sh->pd_idx + 2);
2252 }
2253 break;
2254 case ALGORITHM_PARITY_0:
2255 i -= 2;
2256 break;
2257 case ALGORITHM_PARITY_N:
2258 break;
2259 case ALGORITHM_ROTATING_N_CONTINUE:
2260 /* Like left_symmetric, but P is before Q */
2261 if (sh->pd_idx == 0)
2262 i--; /* P D D D Q */
2263 else {
2264 /* D D Q P D */
2265 if (i < sh->pd_idx)
2266 i += raid_disks;
2267 i -= (sh->pd_idx + 1);
2268 }
2269 break;
2270 case ALGORITHM_LEFT_ASYMMETRIC_6:
2271 case ALGORITHM_RIGHT_ASYMMETRIC_6:
2272 if (i > sh->pd_idx)
2273 i--;
2274 break;
2275 case ALGORITHM_LEFT_SYMMETRIC_6:
2276 case ALGORITHM_RIGHT_SYMMETRIC_6:
2277 if (i < sh->pd_idx)
2278 i += data_disks + 1;
2279 i -= (sh->pd_idx + 1);
2280 break;
2281 case ALGORITHM_PARITY_0_6:
2282 i -= 1;
2283 break;
2284 default:
2285 BUG();
2286 }
2287 break;
2288 }
2289
2290 chunk_number = stripe * data_disks + i;
2291 r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2292
2293 check = raid5_compute_sector(conf, r_sector,
2294 previous, &dummy1, &sh2);
2295 if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2296 || sh2.qd_idx != sh->qd_idx) {
2297 printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2298 mdname(conf->mddev));
2299 return 0;
2300 }
2301 return r_sector;
2302 }
2303
2304
2305 static void
2306 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2307 int rcw, int expand)
2308 {
2309 int i, pd_idx = sh->pd_idx, disks = sh->disks;
2310 struct r5conf *conf = sh->raid_conf;
2311 int level = conf->level;
2312
2313 if (rcw) {
2314
2315 for (i = disks; i--; ) {
2316 struct r5dev *dev = &sh->dev[i];
2317
2318 if (dev->towrite) {
2319 set_bit(R5_LOCKED, &dev->flags);
2320 set_bit(R5_Wantdrain, &dev->flags);
2321 if (!expand)
2322 clear_bit(R5_UPTODATE, &dev->flags);
2323 s->locked++;
2324 }
2325 }
2326 /* if we are not expanding this is a proper write request, and
2327 * there will be bios with new data to be drained into the
2328 * stripe cache
2329 */
2330 if (!expand) {
2331 if (!s->locked)
2332 /* False alarm, nothing to do */
2333 return;
2334 sh->reconstruct_state = reconstruct_state_drain_run;
2335 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2336 } else
2337 sh->reconstruct_state = reconstruct_state_run;
2338
2339 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2340
2341 if (s->locked + conf->max_degraded == disks)
2342 if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2343 atomic_inc(&conf->pending_full_writes);
2344 } else {
2345 BUG_ON(level == 6);
2346 BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2347 test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2348
2349 for (i = disks; i--; ) {
2350 struct r5dev *dev = &sh->dev[i];
2351 if (i == pd_idx)
2352 continue;
2353
2354 if (dev->towrite &&
2355 (test_bit(R5_UPTODATE, &dev->flags) ||
2356 test_bit(R5_Wantcompute, &dev->flags))) {
2357 set_bit(R5_Wantdrain, &dev->flags);
2358 set_bit(R5_LOCKED, &dev->flags);
2359 clear_bit(R5_UPTODATE, &dev->flags);
2360 s->locked++;
2361 }
2362 }
2363 if (!s->locked)
2364 /* False alarm - nothing to do */
2365 return;
2366 sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2367 set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2368 set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2369 set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2370 }
2371
2372 /* keep the parity disk(s) locked while asynchronous operations
2373 * are in flight
2374 */
2375 set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2376 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2377 s->locked++;
2378
2379 if (level == 6) {
2380 int qd_idx = sh->qd_idx;
2381 struct r5dev *dev = &sh->dev[qd_idx];
2382
2383 set_bit(R5_LOCKED, &dev->flags);
2384 clear_bit(R5_UPTODATE, &dev->flags);
2385 s->locked++;
2386 }
2387
2388 pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2389 __func__, (unsigned long long)sh->sector,
2390 s->locked, s->ops_request);
2391 }
2392
2393 /*
2394 * Each stripe/dev can have one or more bion attached.
2395 * toread/towrite point to the first in a chain.
2396 * The bi_next chain must be in order.
2397 */
2398 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2399 {
2400 struct bio **bip;
2401 struct r5conf *conf = sh->raid_conf;
2402 int firstwrite=0;
2403
2404 pr_debug("adding bi b#%llu to stripe s#%llu\n",
2405 (unsigned long long)bi->bi_sector,
2406 (unsigned long long)sh->sector);
2407
2408 /*
2409 * If several bio share a stripe. The bio bi_phys_segments acts as a
2410 * reference count to avoid race. The reference count should already be
2411 * increased before this function is called (for example, in
2412 * make_request()), so other bio sharing this stripe will not free the
2413 * stripe. If a stripe is owned by one stripe, the stripe lock will
2414 * protect it.
2415 */
2416 spin_lock_irq(&sh->stripe_lock);
2417 if (forwrite) {
2418 bip = &sh->dev[dd_idx].towrite;
2419 if (*bip == NULL)
2420 firstwrite = 1;
2421 } else
2422 bip = &sh->dev[dd_idx].toread;
2423 while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2424 if (bio_end_sector(*bip) > bi->bi_sector)
2425 goto overlap;
2426 bip = & (*bip)->bi_next;
2427 }
2428 if (*bip && (*bip)->bi_sector < bio_end_sector(bi))
2429 goto overlap;
2430
2431 BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2432 if (*bip)
2433 bi->bi_next = *bip;
2434 *bip = bi;
2435 raid5_inc_bi_active_stripes(bi);
2436
2437 if (forwrite) {
2438 /* check if page is covered */
2439 sector_t sector = sh->dev[dd_idx].sector;
2440 for (bi=sh->dev[dd_idx].towrite;
2441 sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2442 bi && bi->bi_sector <= sector;
2443 bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2444 if (bio_end_sector(bi) >= sector)
2445 sector = bio_end_sector(bi);
2446 }
2447 if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2448 set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2449 }
2450
2451 pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2452 (unsigned long long)(*bip)->bi_sector,
2453 (unsigned long long)sh->sector, dd_idx);
2454 spin_unlock_irq(&sh->stripe_lock);
2455
2456 if (conf->mddev->bitmap && firstwrite) {
2457 bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2458 STRIPE_SECTORS, 0);
2459 sh->bm_seq = conf->seq_flush+1;
2460 set_bit(STRIPE_BIT_DELAY, &sh->state);
2461 }
2462 return 1;
2463
2464 overlap:
2465 set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2466 spin_unlock_irq(&sh->stripe_lock);
2467 return 0;
2468 }
2469
2470 static void end_reshape(struct r5conf *conf);
2471
2472 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2473 struct stripe_head *sh)
2474 {
2475 int sectors_per_chunk =
2476 previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2477 int dd_idx;
2478 int chunk_offset = sector_div(stripe, sectors_per_chunk);
2479 int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2480
2481 raid5_compute_sector(conf,
2482 stripe * (disks - conf->max_degraded)
2483 *sectors_per_chunk + chunk_offset,
2484 previous,
2485 &dd_idx, sh);
2486 }
2487
2488 static void
2489 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2490 struct stripe_head_state *s, int disks,
2491 struct bio **return_bi)
2492 {
2493 int i;
2494 for (i = disks; i--; ) {
2495 struct bio *bi;
2496 int bitmap_end = 0;
2497
2498 if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2499 struct md_rdev *rdev;
2500 rcu_read_lock();
2501 rdev = rcu_dereference(conf->disks[i].rdev);
2502 if (rdev && test_bit(In_sync, &rdev->flags))
2503 atomic_inc(&rdev->nr_pending);
2504 else
2505 rdev = NULL;
2506 rcu_read_unlock();
2507 if (rdev) {
2508 if (!rdev_set_badblocks(
2509 rdev,
2510 sh->sector,
2511 STRIPE_SECTORS, 0))
2512 md_error(conf->mddev, rdev);
2513 rdev_dec_pending(rdev, conf->mddev);
2514 }
2515 }
2516 spin_lock_irq(&sh->stripe_lock);
2517 /* fail all writes first */
2518 bi = sh->dev[i].towrite;
2519 sh->dev[i].towrite = NULL;
2520 spin_unlock_irq(&sh->stripe_lock);
2521 if (bi)
2522 bitmap_end = 1;
2523
2524 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2525 wake_up(&conf->wait_for_overlap);
2526
2527 while (bi && bi->bi_sector <
2528 sh->dev[i].sector + STRIPE_SECTORS) {
2529 struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2530 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2531 if (!raid5_dec_bi_active_stripes(bi)) {
2532 md_write_end(conf->mddev);
2533 bi->bi_next = *return_bi;
2534 *return_bi = bi;
2535 }
2536 bi = nextbi;
2537 }
2538 if (bitmap_end)
2539 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2540 STRIPE_SECTORS, 0, 0);
2541 bitmap_end = 0;
2542 /* and fail all 'written' */
2543 bi = sh->dev[i].written;
2544 sh->dev[i].written = NULL;
2545 if (bi) bitmap_end = 1;
2546 while (bi && bi->bi_sector <
2547 sh->dev[i].sector + STRIPE_SECTORS) {
2548 struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2549 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2550 if (!raid5_dec_bi_active_stripes(bi)) {
2551 md_write_end(conf->mddev);
2552 bi->bi_next = *return_bi;
2553 *return_bi = bi;
2554 }
2555 bi = bi2;
2556 }
2557
2558 /* fail any reads if this device is non-operational and
2559 * the data has not reached the cache yet.
2560 */
2561 if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2562 (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2563 test_bit(R5_ReadError, &sh->dev[i].flags))) {
2564 spin_lock_irq(&sh->stripe_lock);
2565 bi = sh->dev[i].toread;
2566 sh->dev[i].toread = NULL;
2567 spin_unlock_irq(&sh->stripe_lock);
2568 if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2569 wake_up(&conf->wait_for_overlap);
2570 while (bi && bi->bi_sector <
2571 sh->dev[i].sector + STRIPE_SECTORS) {
2572 struct bio *nextbi =
2573 r5_next_bio(bi, sh->dev[i].sector);
2574 clear_bit(BIO_UPTODATE, &bi->bi_flags);
2575 if (!raid5_dec_bi_active_stripes(bi)) {
2576 bi->bi_next = *return_bi;
2577 *return_bi = bi;
2578 }
2579 bi = nextbi;
2580 }
2581 }
2582 if (bitmap_end)
2583 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2584 STRIPE_SECTORS, 0, 0);
2585 /* If we were in the middle of a write the parity block might
2586 * still be locked - so just clear all R5_LOCKED flags
2587 */
2588 clear_bit(R5_LOCKED, &sh->dev[i].flags);
2589 }
2590
2591 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2592 if (atomic_dec_and_test(&conf->pending_full_writes))
2593 md_wakeup_thread(conf->mddev->thread);
2594 }
2595
2596 static void
2597 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2598 struct stripe_head_state *s)
2599 {
2600 int abort = 0;
2601 int i;
2602
2603 clear_bit(STRIPE_SYNCING, &sh->state);
2604 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2605 wake_up(&conf->wait_for_overlap);
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 || sh->sector >= sh->raid_conf->mddev->recovery_cp)
2683 && s->failed && s->to_write))) {
2684 /* we would like to get this block, possibly by computing it,
2685 * otherwise read it if the backing disk is insync
2686 */
2687 BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2688 BUG_ON(test_bit(R5_Wantread, &dev->flags));
2689 if ((s->uptodate == disks - 1) &&
2690 (s->failed && (disk_idx == s->failed_num[0] ||
2691 disk_idx == s->failed_num[1]))) {
2692 /* have disk failed, and we're requested to fetch it;
2693 * do compute it
2694 */
2695 pr_debug("Computing stripe %llu block %d\n",
2696 (unsigned long long)sh->sector, disk_idx);
2697 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2698 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2699 set_bit(R5_Wantcompute, &dev->flags);
2700 sh->ops.target = disk_idx;
2701 sh->ops.target2 = -1; /* no 2nd target */
2702 s->req_compute = 1;
2703 /* Careful: from this point on 'uptodate' is in the eye
2704 * of raid_run_ops which services 'compute' operations
2705 * before writes. R5_Wantcompute flags a block that will
2706 * be R5_UPTODATE by the time it is needed for a
2707 * subsequent operation.
2708 */
2709 s->uptodate++;
2710 return 1;
2711 } else if (s->uptodate == disks-2 && s->failed >= 2) {
2712 /* Computing 2-failure is *very* expensive; only
2713 * do it if failed >= 2
2714 */
2715 int other;
2716 for (other = disks; other--; ) {
2717 if (other == disk_idx)
2718 continue;
2719 if (!test_bit(R5_UPTODATE,
2720 &sh->dev[other].flags))
2721 break;
2722 }
2723 BUG_ON(other < 0);
2724 pr_debug("Computing stripe %llu blocks %d,%d\n",
2725 (unsigned long long)sh->sector,
2726 disk_idx, other);
2727 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2728 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2729 set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2730 set_bit(R5_Wantcompute, &sh->dev[other].flags);
2731 sh->ops.target = disk_idx;
2732 sh->ops.target2 = other;
2733 s->uptodate += 2;
2734 s->req_compute = 1;
2735 return 1;
2736 } else if (test_bit(R5_Insync, &dev->flags)) {
2737 set_bit(R5_LOCKED, &dev->flags);
2738 set_bit(R5_Wantread, &dev->flags);
2739 s->locked++;
2740 pr_debug("Reading block %d (sync=%d)\n",
2741 disk_idx, s->syncing);
2742 }
2743 }
2744
2745 return 0;
2746 }
2747
2748 /**
2749 * handle_stripe_fill - read or compute data to satisfy pending requests.
2750 */
2751 static void handle_stripe_fill(struct stripe_head *sh,
2752 struct stripe_head_state *s,
2753 int disks)
2754 {
2755 int i;
2756
2757 /* look for blocks to read/compute, skip this if a compute
2758 * is already in flight, or if the stripe contents are in the
2759 * midst of changing due to a write
2760 */
2761 if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2762 !sh->reconstruct_state)
2763 for (i = disks; i--; )
2764 if (fetch_block(sh, s, i, disks))
2765 break;
2766 set_bit(STRIPE_HANDLE, &sh->state);
2767 }
2768
2769
2770 /* handle_stripe_clean_event
2771 * any written block on an uptodate or failed drive can be returned.
2772 * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2773 * never LOCKED, so we don't need to test 'failed' directly.
2774 */
2775 static void handle_stripe_clean_event(struct r5conf *conf,
2776 struct stripe_head *sh, int disks, struct bio **return_bi)
2777 {
2778 int i;
2779 struct r5dev *dev;
2780 int discard_pending = 0;
2781
2782 for (i = disks; i--; )
2783 if (sh->dev[i].written) {
2784 dev = &sh->dev[i];
2785 if (!test_bit(R5_LOCKED, &dev->flags) &&
2786 (test_bit(R5_UPTODATE, &dev->flags) ||
2787 test_bit(R5_Discard, &dev->flags))) {
2788 /* We can return any write requests */
2789 struct bio *wbi, *wbi2;
2790 pr_debug("Return write for disc %d\n", i);
2791 if (test_and_clear_bit(R5_Discard, &dev->flags))
2792 clear_bit(R5_UPTODATE, &dev->flags);
2793 wbi = dev->written;
2794 dev->written = NULL;
2795 while (wbi && wbi->bi_sector <
2796 dev->sector + STRIPE_SECTORS) {
2797 wbi2 = r5_next_bio(wbi, dev->sector);
2798 if (!raid5_dec_bi_active_stripes(wbi)) {
2799 md_write_end(conf->mddev);
2800 wbi->bi_next = *return_bi;
2801 *return_bi = wbi;
2802 }
2803 wbi = wbi2;
2804 }
2805 bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2806 STRIPE_SECTORS,
2807 !test_bit(STRIPE_DEGRADED, &sh->state),
2808 0);
2809 } else if (test_bit(R5_Discard, &dev->flags))
2810 discard_pending = 1;
2811 }
2812 if (!discard_pending &&
2813 test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
2814 clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2815 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2816 if (sh->qd_idx >= 0) {
2817 clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2818 clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
2819 }
2820 /* now that discard is done we can proceed with any sync */
2821 clear_bit(STRIPE_DISCARD, &sh->state);
2822 /*
2823 * SCSI discard will change some bio fields and the stripe has
2824 * no updated data, so remove it from hash list and the stripe
2825 * will be reinitialized
2826 */
2827 spin_lock_irq(&conf->device_lock);
2828 remove_hash(sh);
2829 spin_unlock_irq(&conf->device_lock);
2830 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2831 set_bit(STRIPE_HANDLE, &sh->state);
2832
2833 }
2834
2835 if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2836 if (atomic_dec_and_test(&conf->pending_full_writes))
2837 md_wakeup_thread(conf->mddev->thread);
2838 }
2839
2840 static void handle_stripe_dirtying(struct r5conf *conf,
2841 struct stripe_head *sh,
2842 struct stripe_head_state *s,
2843 int disks)
2844 {
2845 int rmw = 0, rcw = 0, i;
2846 sector_t recovery_cp = conf->mddev->recovery_cp;
2847
2848 /* RAID6 requires 'rcw' in current implementation.
2849 * Otherwise, check whether resync is now happening or should start.
2850 * If yes, then the array is dirty (after unclean shutdown or
2851 * initial creation), so parity in some stripes might be inconsistent.
2852 * In this case, we need to always do reconstruct-write, to ensure
2853 * that in case of drive failure or read-error correction, we
2854 * generate correct data from the parity.
2855 */
2856 if (conf->max_degraded == 2 ||
2857 (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
2858 s->failed == 0)) {
2859 /* Calculate the real rcw later - for now make it
2860 * look like rcw is cheaper
2861 */
2862 rcw = 1; rmw = 2;
2863 pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2864 conf->max_degraded, (unsigned long long)recovery_cp,
2865 (unsigned long long)sh->sector);
2866 } else for (i = disks; i--; ) {
2867 /* would I have to read this buffer for read_modify_write */
2868 struct r5dev *dev = &sh->dev[i];
2869 if ((dev->towrite || i == sh->pd_idx) &&
2870 !test_bit(R5_LOCKED, &dev->flags) &&
2871 !(test_bit(R5_UPTODATE, &dev->flags) ||
2872 test_bit(R5_Wantcompute, &dev->flags))) {
2873 if (test_bit(R5_Insync, &dev->flags))
2874 rmw++;
2875 else
2876 rmw += 2*disks; /* cannot read it */
2877 }
2878 /* Would I have to read this buffer for reconstruct_write */
2879 if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2880 !test_bit(R5_LOCKED, &dev->flags) &&
2881 !(test_bit(R5_UPTODATE, &dev->flags) ||
2882 test_bit(R5_Wantcompute, &dev->flags))) {
2883 if (test_bit(R5_Insync, &dev->flags)) rcw++;
2884 else
2885 rcw += 2*disks;
2886 }
2887 }
2888 pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2889 (unsigned long long)sh->sector, rmw, rcw);
2890 set_bit(STRIPE_HANDLE, &sh->state);
2891 if (rmw < rcw && rmw > 0) {
2892 /* prefer read-modify-write, but need to get some data */
2893 if (conf->mddev->queue)
2894 blk_add_trace_msg(conf->mddev->queue,
2895 "raid5 rmw %llu %d",
2896 (unsigned long long)sh->sector, rmw);
2897 for (i = disks; i--; ) {
2898 struct r5dev *dev = &sh->dev[i];
2899 if ((dev->towrite || i == sh->pd_idx) &&
2900 !test_bit(R5_LOCKED, &dev->flags) &&
2901 !(test_bit(R5_UPTODATE, &dev->flags) ||
2902 test_bit(R5_Wantcompute, &dev->flags)) &&
2903 test_bit(R5_Insync, &dev->flags)) {
2904 if (
2905 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2906 pr_debug("Read_old block "
2907 "%d for r-m-w\n", i);
2908 set_bit(R5_LOCKED, &dev->flags);
2909 set_bit(R5_Wantread, &dev->flags);
2910 s->locked++;
2911 } else {
2912 set_bit(STRIPE_DELAYED, &sh->state);
2913 set_bit(STRIPE_HANDLE, &sh->state);
2914 }
2915 }
2916 }
2917 }
2918 if (rcw <= rmw && rcw > 0) {
2919 /* want reconstruct write, but need to get some data */
2920 int qread =0;
2921 rcw = 0;
2922 for (i = disks; i--; ) {
2923 struct r5dev *dev = &sh->dev[i];
2924 if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2925 i != sh->pd_idx && i != sh->qd_idx &&
2926 !test_bit(R5_LOCKED, &dev->flags) &&
2927 !(test_bit(R5_UPTODATE, &dev->flags) ||
2928 test_bit(R5_Wantcompute, &dev->flags))) {
2929 rcw++;
2930 if (!test_bit(R5_Insync, &dev->flags))
2931 continue; /* it's a failed drive */
2932 if (
2933 test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2934 pr_debug("Read_old block "
2935 "%d for Reconstruct\n", i);
2936 set_bit(R5_LOCKED, &dev->flags);
2937 set_bit(R5_Wantread, &dev->flags);
2938 s->locked++;
2939 qread++;
2940 } else {
2941 set_bit(STRIPE_DELAYED, &sh->state);
2942 set_bit(STRIPE_HANDLE, &sh->state);
2943 }
2944 }
2945 }
2946 if (rcw && conf->mddev->queue)
2947 blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
2948 (unsigned long long)sh->sector,
2949 rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
2950 }
2951 /* now if nothing is locked, and if we have enough data,
2952 * we can start a write request
2953 */
2954 /* since handle_stripe can be called at any time we need to handle the
2955 * case where a compute block operation has been submitted and then a
2956 * subsequent call wants to start a write request. raid_run_ops only
2957 * handles the case where compute block and reconstruct are requested
2958 * simultaneously. If this is not the case then new writes need to be
2959 * held off until the compute completes.
2960 */
2961 if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2962 (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2963 !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2964 schedule_reconstruction(sh, s, rcw == 0, 0);
2965 }
2966
2967 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2968 struct stripe_head_state *s, int disks)
2969 {
2970 struct r5dev *dev = NULL;
2971
2972 set_bit(STRIPE_HANDLE, &sh->state);
2973
2974 switch (sh->check_state) {
2975 case check_state_idle:
2976 /* start a new check operation if there are no failures */
2977 if (s->failed == 0) {
2978 BUG_ON(s->uptodate != disks);
2979 sh->check_state = check_state_run;
2980 set_bit(STRIPE_OP_CHECK, &s->ops_request);
2981 clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2982 s->uptodate--;
2983 break;
2984 }
2985 dev = &sh->dev[s->failed_num[0]];
2986 /* fall through */
2987 case check_state_compute_result:
2988 sh->check_state = check_state_idle;
2989 if (!dev)
2990 dev = &sh->dev[sh->pd_idx];
2991
2992 /* check that a write has not made the stripe insync */
2993 if (test_bit(STRIPE_INSYNC, &sh->state))
2994 break;
2995
2996 /* either failed parity check, or recovery is happening */
2997 BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2998 BUG_ON(s->uptodate != disks);
2999
3000 set_bit(R5_LOCKED, &dev->flags);
3001 s->locked++;
3002 set_bit(R5_Wantwrite, &dev->flags);
3003
3004 clear_bit(STRIPE_DEGRADED, &sh->state);
3005 set_bit(STRIPE_INSYNC, &sh->state);
3006 break;
3007 case check_state_run:
3008 break; /* we will be called again upon completion */
3009 case check_state_check_result:
3010 sh->check_state = check_state_idle;
3011
3012 /* if a failure occurred during the check operation, leave
3013 * STRIPE_INSYNC not set and let the stripe be handled again
3014 */
3015 if (s->failed)
3016 break;
3017
3018 /* handle a successful check operation, if parity is correct
3019 * we are done. Otherwise update the mismatch count and repair
3020 * parity if !MD_RECOVERY_CHECK
3021 */
3022 if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3023 /* parity is correct (on disc,
3024 * not in buffer any more)
3025 */
3026 set_bit(STRIPE_INSYNC, &sh->state);
3027 else {
3028 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3029 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3030 /* don't try to repair!! */
3031 set_bit(STRIPE_INSYNC, &sh->state);
3032 else {
3033 sh->check_state = check_state_compute_run;
3034 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3035 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3036 set_bit(R5_Wantcompute,
3037 &sh->dev[sh->pd_idx].flags);
3038 sh->ops.target = sh->pd_idx;
3039 sh->ops.target2 = -1;
3040 s->uptodate++;
3041 }
3042 }
3043 break;
3044 case check_state_compute_run:
3045 break;
3046 default:
3047 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3048 __func__, sh->check_state,
3049 (unsigned long long) sh->sector);
3050 BUG();
3051 }
3052 }
3053
3054
3055 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3056 struct stripe_head_state *s,
3057 int disks)
3058 {
3059 int pd_idx = sh->pd_idx;
3060 int qd_idx = sh->qd_idx;
3061 struct r5dev *dev;
3062
3063 set_bit(STRIPE_HANDLE, &sh->state);
3064
3065 BUG_ON(s->failed > 2);
3066
3067 /* Want to check and possibly repair P and Q.
3068 * However there could be one 'failed' device, in which
3069 * case we can only check one of them, possibly using the
3070 * other to generate missing data
3071 */
3072
3073 switch (sh->check_state) {
3074 case check_state_idle:
3075 /* start a new check operation if there are < 2 failures */
3076 if (s->failed == s->q_failed) {
3077 /* The only possible failed device holds Q, so it
3078 * makes sense to check P (If anything else were failed,
3079 * we would have used P to recreate it).
3080 */
3081 sh->check_state = check_state_run;
3082 }
3083 if (!s->q_failed && s->failed < 2) {
3084 /* Q is not failed, and we didn't use it to generate
3085 * anything, so it makes sense to check it
3086 */
3087 if (sh->check_state == check_state_run)
3088 sh->check_state = check_state_run_pq;
3089 else
3090 sh->check_state = check_state_run_q;
3091 }
3092
3093 /* discard potentially stale zero_sum_result */
3094 sh->ops.zero_sum_result = 0;
3095
3096 if (sh->check_state == check_state_run) {
3097 /* async_xor_zero_sum destroys the contents of P */
3098 clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3099 s->uptodate--;
3100 }
3101 if (sh->check_state >= check_state_run &&
3102 sh->check_state <= check_state_run_pq) {
3103 /* async_syndrome_zero_sum preserves P and Q, so
3104 * no need to mark them !uptodate here
3105 */
3106 set_bit(STRIPE_OP_CHECK, &s->ops_request);
3107 break;
3108 }
3109
3110 /* we have 2-disk failure */
3111 BUG_ON(s->failed != 2);
3112 /* fall through */
3113 case check_state_compute_result:
3114 sh->check_state = check_state_idle;
3115
3116 /* check that a write has not made the stripe insync */
3117 if (test_bit(STRIPE_INSYNC, &sh->state))
3118 break;
3119
3120 /* now write out any block on a failed drive,
3121 * or P or Q if they were recomputed
3122 */
3123 BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3124 if (s->failed == 2) {
3125 dev = &sh->dev[s->failed_num[1]];
3126 s->locked++;
3127 set_bit(R5_LOCKED, &dev->flags);
3128 set_bit(R5_Wantwrite, &dev->flags);
3129 }
3130 if (s->failed >= 1) {
3131 dev = &sh->dev[s->failed_num[0]];
3132 s->locked++;
3133 set_bit(R5_LOCKED, &dev->flags);
3134 set_bit(R5_Wantwrite, &dev->flags);
3135 }
3136 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3137 dev = &sh->dev[pd_idx];
3138 s->locked++;
3139 set_bit(R5_LOCKED, &dev->flags);
3140 set_bit(R5_Wantwrite, &dev->flags);
3141 }
3142 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3143 dev = &sh->dev[qd_idx];
3144 s->locked++;
3145 set_bit(R5_LOCKED, &dev->flags);
3146 set_bit(R5_Wantwrite, &dev->flags);
3147 }
3148 clear_bit(STRIPE_DEGRADED, &sh->state);
3149
3150 set_bit(STRIPE_INSYNC, &sh->state);
3151 break;
3152 case check_state_run:
3153 case check_state_run_q:
3154 case check_state_run_pq:
3155 break; /* we will be called again upon completion */
3156 case check_state_check_result:
3157 sh->check_state = check_state_idle;
3158
3159 /* handle a successful check operation, if parity is correct
3160 * we are done. Otherwise update the mismatch count and repair
3161 * parity if !MD_RECOVERY_CHECK
3162 */
3163 if (sh->ops.zero_sum_result == 0) {
3164 /* both parities are correct */
3165 if (!s->failed)
3166 set_bit(STRIPE_INSYNC, &sh->state);
3167 else {
3168 /* in contrast to the raid5 case we can validate
3169 * parity, but still have a failure to write
3170 * back
3171 */
3172 sh->check_state = check_state_compute_result;
3173 /* Returning at this point means that we may go
3174 * off and bring p and/or q uptodate again so
3175 * we make sure to check zero_sum_result again
3176 * to verify if p or q need writeback
3177 */
3178 }
3179 } else {
3180 atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3181 if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3182 /* don't try to repair!! */
3183 set_bit(STRIPE_INSYNC, &sh->state);
3184 else {
3185 int *target = &sh->ops.target;
3186
3187 sh->ops.target = -1;
3188 sh->ops.target2 = -1;
3189 sh->check_state = check_state_compute_run;
3190 set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3191 set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3192 if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3193 set_bit(R5_Wantcompute,
3194 &sh->dev[pd_idx].flags);
3195 *target = pd_idx;
3196 target = &sh->ops.target2;
3197 s->uptodate++;
3198 }
3199 if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3200 set_bit(R5_Wantcompute,
3201 &sh->dev[qd_idx].flags);
3202 *target = qd_idx;
3203 s->uptodate++;
3204 }
3205 }
3206 }
3207 break;
3208 case check_state_compute_run:
3209 break;
3210 default:
3211 printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3212 __func__, sh->check_state,
3213 (unsigned long long) sh->sector);
3214 BUG();
3215 }
3216 }
3217
3218 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3219 {
3220 int i;
3221
3222 /* We have read all the blocks in this stripe and now we need to
3223 * copy some of them into a target stripe for expand.
3224 */
3225 struct dma_async_tx_descriptor *tx = NULL;
3226 clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3227 for (i = 0; i < sh->disks; i++)
3228 if (i != sh->pd_idx && i != sh->qd_idx) {
3229 int dd_idx, j;
3230 struct stripe_head *sh2;
3231 struct async_submit_ctl submit;
3232
3233 sector_t bn = compute_blocknr(sh, i, 1);
3234 sector_t s = raid5_compute_sector(conf, bn, 0,
3235 &dd_idx, NULL);
3236 sh2 = get_active_stripe(conf, s, 0, 1, 1);
3237 if (sh2 == NULL)
3238 /* so far only the early blocks of this stripe
3239 * have been requested. When later blocks
3240 * get requested, we will try again
3241 */
3242 continue;
3243 if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3244 test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3245 /* must have already done this block */
3246 release_stripe(sh2);
3247 continue;
3248 }
3249
3250 /* place all the copies on one channel */
3251 init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3252 tx = async_memcpy(sh2->dev[dd_idx].page,
3253 sh->dev[i].page, 0, 0, STRIPE_SIZE,
3254 &submit);
3255
3256 set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3257 set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3258 for (j = 0; j < conf->raid_disks; j++)
3259 if (j != sh2->pd_idx &&
3260 j != sh2->qd_idx &&
3261 !test_bit(R5_Expanded, &sh2->dev[j].flags))
3262 break;
3263 if (j == conf->raid_disks) {
3264 set_bit(STRIPE_EXPAND_READY, &sh2->state);
3265 set_bit(STRIPE_HANDLE, &sh2->state);
3266 }
3267 release_stripe(sh2);
3268
3269 }
3270 /* done submitting copies, wait for them to complete */
3271 async_tx_quiesce(&tx);
3272 }
3273
3274 /*
3275 * handle_stripe - do things to a stripe.
3276 *
3277 * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3278 * state of various bits to see what needs to be done.
3279 * Possible results:
3280 * return some read requests which now have data
3281 * return some write requests which are safely on storage
3282 * schedule a read on some buffers
3283 * schedule a write of some buffers
3284 * return confirmation of parity correctness
3285 *
3286 */
3287
3288 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3289 {
3290 struct r5conf *conf = sh->raid_conf;
3291 int disks = sh->disks;
3292 struct r5dev *dev;
3293 int i;
3294 int do_recovery = 0;
3295
3296 memset(s, 0, sizeof(*s));
3297
3298 s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3299 s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3300 s->failed_num[0] = -1;
3301 s->failed_num[1] = -1;
3302
3303 /* Now to look around and see what can be done */
3304 rcu_read_lock();
3305 for (i=disks; i--; ) {
3306 struct md_rdev *rdev;
3307 sector_t first_bad;
3308 int bad_sectors;
3309 int is_bad = 0;
3310
3311 dev = &sh->dev[i];
3312
3313 pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3314 i, dev->flags,
3315 dev->toread, dev->towrite, dev->written);
3316 /* maybe we can reply to a read
3317 *
3318 * new wantfill requests are only permitted while
3319 * ops_complete_biofill is guaranteed to be inactive
3320 */
3321 if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3322 !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3323 set_bit(R5_Wantfill, &dev->flags);
3324
3325 /* now count some things */
3326 if (test_bit(R5_LOCKED, &dev->flags))
3327 s->locked++;
3328 if (test_bit(R5_UPTODATE, &dev->flags))
3329 s->uptodate++;
3330 if (test_bit(R5_Wantcompute, &dev->flags)) {
3331 s->compute++;
3332 BUG_ON(s->compute > 2);
3333 }
3334
3335 if (test_bit(R5_Wantfill, &dev->flags))
3336 s->to_fill++;
3337 else if (dev->toread)
3338 s->to_read++;
3339 if (dev->towrite) {
3340 s->to_write++;
3341 if (!test_bit(R5_OVERWRITE, &dev->flags))
3342 s->non_overwrite++;
3343 }
3344 if (dev->written)
3345 s->written++;
3346 /* Prefer to use the replacement for reads, but only
3347 * if it is recovered enough and has no bad blocks.
3348 */
3349 rdev = rcu_dereference(conf->disks[i].replacement);
3350 if (rdev && !test_bit(Faulty, &rdev->flags) &&
3351 rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3352 !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3353 &first_bad, &bad_sectors))
3354 set_bit(R5_ReadRepl, &dev->flags);
3355 else {
3356 if (rdev)
3357 set_bit(R5_NeedReplace, &dev->flags);
3358 rdev = rcu_dereference(conf->disks[i].rdev);
3359 clear_bit(R5_ReadRepl, &dev->flags);
3360 }
3361 if (rdev && test_bit(Faulty, &rdev->flags))
3362 rdev = NULL;
3363 if (rdev) {
3364 is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3365 &first_bad, &bad_sectors);
3366 if (s->blocked_rdev == NULL
3367 && (test_bit(Blocked, &rdev->flags)
3368 || is_bad < 0)) {
3369 if (is_bad < 0)
3370 set_bit(BlockedBadBlocks,
3371 &rdev->flags);
3372 s->blocked_rdev = rdev;
3373 atomic_inc(&rdev->nr_pending);
3374 }
3375 }
3376 clear_bit(R5_Insync, &dev->flags);
3377 if (!rdev)
3378 /* Not in-sync */;
3379 else if (is_bad) {
3380 /* also not in-sync */
3381 if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3382 test_bit(R5_UPTODATE, &dev->flags)) {
3383 /* treat as in-sync, but with a read error
3384 * which we can now try to correct
3385 */
3386 set_bit(R5_Insync, &dev->flags);
3387 set_bit(R5_ReadError, &dev->flags);
3388 }
3389 } else if (test_bit(In_sync, &rdev->flags))
3390 set_bit(R5_Insync, &dev->flags);
3391 else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3392 /* in sync if before recovery_offset */
3393 set_bit(R5_Insync, &dev->flags);
3394 else if (test_bit(R5_UPTODATE, &dev->flags) &&
3395 test_bit(R5_Expanded, &dev->flags))
3396 /* If we've reshaped into here, we assume it is Insync.
3397 * We will shortly update recovery_offset to make
3398 * it official.
3399 */
3400 set_bit(R5_Insync, &dev->flags);
3401
3402 if (test_bit(R5_WriteError, &dev->flags)) {
3403 /* This flag does not apply to '.replacement'
3404 * only to .rdev, so make sure to check that*/
3405 struct md_rdev *rdev2 = rcu_dereference(
3406 conf->disks[i].rdev);
3407 if (rdev2 == rdev)
3408 clear_bit(R5_Insync, &dev->flags);
3409 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3410 s->handle_bad_blocks = 1;
3411 atomic_inc(&rdev2->nr_pending);
3412 } else
3413 clear_bit(R5_WriteError, &dev->flags);
3414 }
3415 if (test_bit(R5_MadeGood, &dev->flags)) {
3416 /* This flag does not apply to '.replacement'
3417 * only to .rdev, so make sure to check that*/
3418 struct md_rdev *rdev2 = rcu_dereference(
3419 conf->disks[i].rdev);
3420 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3421 s->handle_bad_blocks = 1;
3422 atomic_inc(&rdev2->nr_pending);
3423 } else
3424 clear_bit(R5_MadeGood, &dev->flags);
3425 }
3426 if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3427 struct md_rdev *rdev2 = rcu_dereference(
3428 conf->disks[i].replacement);
3429 if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3430 s->handle_bad_blocks = 1;
3431 atomic_inc(&rdev2->nr_pending);
3432 } else
3433 clear_bit(R5_MadeGoodRepl, &dev->flags);
3434 }
3435 if (!test_bit(R5_Insync, &dev->flags)) {
3436 /* The ReadError flag will just be confusing now */
3437 clear_bit(R5_ReadError, &dev->flags);
3438 clear_bit(R5_ReWrite, &dev->flags);
3439 }
3440 if (test_bit(R5_ReadError, &dev->flags))
3441 clear_bit(R5_Insync, &dev->flags);
3442 if (!test_bit(R5_Insync, &dev->flags)) {
3443 if (s->failed < 2)
3444 s->failed_num[s->failed] = i;
3445 s->failed++;
3446 if (rdev && !test_bit(Faulty, &rdev->flags))
3447 do_recovery = 1;
3448 }
3449 }
3450 if (test_bit(STRIPE_SYNCING, &sh->state)) {
3451 /* If there is a failed device being replaced,
3452 * we must be recovering.
3453 * else if we are after recovery_cp, we must be syncing
3454 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3455 * else we can only be replacing
3456 * sync and recovery both need to read all devices, and so
3457 * use the same flag.
3458 */
3459 if (do_recovery ||
3460 sh->sector >= conf->mddev->recovery_cp ||
3461 test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3462 s->syncing = 1;
3463 else
3464 s->replacing = 1;
3465 }
3466 rcu_read_unlock();
3467 }
3468
3469 static void handle_stripe(struct stripe_head *sh)
3470 {
3471 struct stripe_head_state s;
3472 struct r5conf *conf = sh->raid_conf;
3473 int i;
3474 int prexor;
3475 int disks = sh->disks;
3476 struct r5dev *pdev, *qdev;
3477
3478 clear_bit(STRIPE_HANDLE, &sh->state);
3479 if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3480 /* already being handled, ensure it gets handled
3481 * again when current action finishes */
3482 set_bit(STRIPE_HANDLE, &sh->state);
3483 return;
3484 }
3485
3486 if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3487 spin_lock(&sh->stripe_lock);
3488 /* Cannot process 'sync' concurrently with 'discard' */
3489 if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3490 test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3491 set_bit(STRIPE_SYNCING, &sh->state);
3492 clear_bit(STRIPE_INSYNC, &sh->state);
3493 clear_bit(STRIPE_REPLACED, &sh->state);
3494 }
3495 spin_unlock(&sh->stripe_lock);
3496 }
3497 clear_bit(STRIPE_DELAYED, &sh->state);
3498
3499 pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3500 "pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3501 (unsigned long long)sh->sector, sh->state,
3502 atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3503 sh->check_state, sh->reconstruct_state);
3504
3505 analyse_stripe(sh, &s);
3506
3507 if (s.handle_bad_blocks) {
3508 set_bit(STRIPE_HANDLE, &sh->state);
3509 goto finish;
3510 }
3511
3512 if (unlikely(s.blocked_rdev)) {
3513 if (s.syncing || s.expanding || s.expanded ||
3514 s.replacing || s.to_write || s.written) {
3515 set_bit(STRIPE_HANDLE, &sh->state);
3516 goto finish;
3517 }
3518 /* There is nothing for the blocked_rdev to block */
3519 rdev_dec_pending(s.blocked_rdev, conf->mddev);
3520 s.blocked_rdev = NULL;
3521 }
3522
3523 if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3524 set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3525 set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3526 }
3527
3528 pr_debug("locked=%d uptodate=%d to_read=%d"
3529 " to_write=%d failed=%d failed_num=%d,%d\n",
3530 s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3531 s.failed_num[0], s.failed_num[1]);
3532 /* check if the array has lost more than max_degraded devices and,
3533 * if so, some requests might need to be failed.
3534 */
3535 if (s.failed > conf->max_degraded) {
3536 sh->check_state = 0;
3537 sh->reconstruct_state = 0;
3538 if (s.to_read+s.to_write+s.written)
3539 handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3540 if (s.syncing + s.replacing)
3541 handle_failed_sync(conf, sh, &s);
3542 }
3543
3544 /* Now we check to see if any write operations have recently
3545 * completed
3546 */
3547 prexor = 0;
3548 if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3549 prexor = 1;
3550 if (sh->reconstruct_state == reconstruct_state_drain_result ||
3551 sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3552 sh->reconstruct_state = reconstruct_state_idle;
3553
3554 /* All the 'written' buffers and the parity block are ready to
3555 * be written back to disk
3556 */
3557 BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3558 !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3559 BUG_ON(sh->qd_idx >= 0 &&
3560 !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3561 !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3562 for (i = disks; i--; ) {
3563 struct r5dev *dev = &sh->dev[i];
3564 if (test_bit(R5_LOCKED, &dev->flags) &&
3565 (i == sh->pd_idx || i == sh->qd_idx ||
3566 dev->written)) {
3567 pr_debug("Writing block %d\n", i);
3568 set_bit(R5_Wantwrite, &dev->flags);
3569 if (prexor)
3570 continue;
3571 if (s.failed > 1)
3572 continue;
3573 if (!test_bit(R5_Insync, &dev->flags) ||
3574 ((i == sh->pd_idx || i == sh->qd_idx) &&
3575 s.failed == 0))
3576 set_bit(STRIPE_INSYNC, &sh->state);
3577 }
3578 }
3579 if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3580 s.dec_preread_active = 1;
3581 }
3582
3583 /*
3584 * might be able to return some write requests if the parity blocks
3585 * are safe, or on a failed drive
3586 */
3587 pdev = &sh->dev[sh->pd_idx];
3588 s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3589 || (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3590 qdev = &sh->dev[sh->qd_idx];
3591 s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3592 || (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3593 || conf->level < 6;
3594
3595 if (s.written &&
3596 (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3597 && !test_bit(R5_LOCKED, &pdev->flags)
3598 && (test_bit(R5_UPTODATE, &pdev->flags) ||
3599 test_bit(R5_Discard, &pdev->flags))))) &&
3600 (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3601 && !test_bit(R5_LOCKED, &qdev->flags)
3602 && (test_bit(R5_UPTODATE, &qdev->flags) ||
3603 test_bit(R5_Discard, &qdev->flags))))))
3604 handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3605
3606 /* Now we might consider reading some blocks, either to check/generate
3607 * parity, or to satisfy requests
3608 * or to load a block that is being partially written.
3609 */
3610 if (s.to_read || s.non_overwrite
3611 || (conf->level == 6 && s.to_write && s.failed)
3612 || (s.syncing && (s.uptodate + s.compute < disks))
3613 || s.replacing
3614 || s.expanding)
3615 handle_stripe_fill(sh, &s, disks);
3616
3617 /* Now to consider new write requests and what else, if anything
3618 * should be read. We do not handle new writes when:
3619 * 1/ A 'write' operation (copy+xor) is already in flight.
3620 * 2/ A 'check' operation is in flight, as it may clobber the parity
3621 * block.
3622 */
3623 if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3624 handle_stripe_dirtying(conf, sh, &s, disks);
3625
3626 /* maybe we need to check and possibly fix the parity for this stripe
3627 * Any reads will already have been scheduled, so we just see if enough
3628 * data is available. The parity check is held off while parity
3629 * dependent operations are in flight.
3630 */
3631 if (sh->check_state ||
3632 (s.syncing && s.locked == 0 &&
3633 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3634 !test_bit(STRIPE_INSYNC, &sh->state))) {
3635 if (conf->level == 6)
3636 handle_parity_checks6(conf, sh, &s, disks);
3637 else
3638 handle_parity_checks5(conf, sh, &s, disks);
3639 }
3640
3641 if ((s.replacing || s.syncing) && s.locked == 0
3642 && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3643 && !test_bit(STRIPE_REPLACED, &sh->state)) {
3644 /* Write out to replacement devices where possible */
3645 for (i = 0; i < conf->raid_disks; i++)
3646 if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3647 WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
3648 set_bit(R5_WantReplace, &sh->dev[i].flags);
3649 set_bit(R5_LOCKED, &sh->dev[i].flags);
3650 s.locked++;
3651 }
3652 if (s.replacing)
3653 set_bit(STRIPE_INSYNC, &sh->state);
3654 set_bit(STRIPE_REPLACED, &sh->state);
3655 }
3656 if ((s.syncing || s.replacing) && s.locked == 0 &&
3657 !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3658 test_bit(STRIPE_INSYNC, &sh->state)) {
3659 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3660 clear_bit(STRIPE_SYNCING, &sh->state);
3661 if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3662 wake_up(&conf->wait_for_overlap);
3663 }
3664
3665 /* If the failed drives are just a ReadError, then we might need
3666 * to progress the repair/check process
3667 */
3668 if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3669 for (i = 0; i < s.failed; i++) {
3670 struct r5dev *dev = &sh->dev[s.failed_num[i]];
3671 if (test_bit(R5_ReadError, &dev->flags)
3672 && !test_bit(R5_LOCKED, &dev->flags)
3673 && test_bit(R5_UPTODATE, &dev->flags)
3674 ) {
3675 if (!test_bit(R5_ReWrite, &dev->flags)) {
3676 set_bit(R5_Wantwrite, &dev->flags);
3677 set_bit(R5_ReWrite, &dev->flags);
3678 set_bit(R5_LOCKED, &dev->flags);
3679 s.locked++;
3680 } else {
3681 /* let's read it back */
3682 set_bit(R5_Wantread, &dev->flags);
3683 set_bit(R5_LOCKED, &dev->flags);
3684 s.locked++;
3685 }
3686 }
3687 }
3688
3689
3690 /* Finish reconstruct operations initiated by the expansion process */
3691 if (sh->reconstruct_state == reconstruct_state_result) {
3692 struct stripe_head *sh_src
3693 = get_active_stripe(conf, sh->sector, 1, 1, 1);
3694 if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3695 /* sh cannot be written until sh_src has been read.
3696 * so arrange for sh to be delayed a little
3697 */
3698 set_bit(STRIPE_DELAYED, &sh->state);
3699 set_bit(STRIPE_HANDLE, &sh->state);
3700 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3701 &sh_src->state))
3702 atomic_inc(&conf->preread_active_stripes);
3703 release_stripe(sh_src);
3704 goto finish;
3705 }
3706 if (sh_src)
3707 release_stripe(sh_src);
3708
3709 sh->reconstruct_state = reconstruct_state_idle;
3710 clear_bit(STRIPE_EXPANDING, &sh->state);
3711 for (i = conf->raid_disks; i--; ) {
3712 set_bit(R5_Wantwrite, &sh->dev[i].flags);
3713 set_bit(R5_LOCKED, &sh->dev[i].flags);
3714 s.locked++;
3715 }
3716 }
3717
3718 if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3719 !sh->reconstruct_state) {
3720 /* Need to write out all blocks after computing parity */
3721 sh->disks = conf->raid_disks;
3722 stripe_set_idx(sh->sector, conf, 0, sh);
3723 schedule_reconstruction(sh, &s, 1, 1);
3724 } else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3725 clear_bit(STRIPE_EXPAND_READY, &sh->state);
3726 atomic_dec(&conf->reshape_stripes);
3727 wake_up(&conf->wait_for_overlap);
3728 md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3729 }
3730
3731 if (s.expanding && s.locked == 0 &&
3732 !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3733 handle_stripe_expansion(conf, sh);
3734
3735 finish:
3736 /* wait for this device to become unblocked */
3737 if (unlikely(s.blocked_rdev)) {
3738 if (conf->mddev->external)
3739 md_wait_for_blocked_rdev(s.blocked_rdev,
3740 conf->mddev);
3741 else
3742 /* Internal metadata will immediately
3743 * be written by raid5d, so we don't
3744 * need to wait here.
3745 */
3746 rdev_dec_pending(s.blocked_rdev,
3747 conf->mddev);
3748 }
3749
3750 if (s.handle_bad_blocks)
3751 for (i = disks; i--; ) {
3752 struct md_rdev *rdev;
3753 struct r5dev *dev = &sh->dev[i];
3754 if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3755 /* We own a safe reference to the rdev */
3756 rdev = conf->disks[i].rdev;
3757 if (!rdev_set_badblocks(rdev, sh->sector,
3758 STRIPE_SECTORS, 0))
3759 md_error(conf->mddev, rdev);
3760 rdev_dec_pending(rdev, conf->mddev);
3761 }
3762 if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3763 rdev = conf->disks[i].rdev;
3764 rdev_clear_badblocks(rdev, sh->sector,
3765 STRIPE_SECTORS, 0);
3766 rdev_dec_pending(rdev, conf->mddev);
3767 }
3768 if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3769 rdev = conf->disks[i].replacement;
3770 if (!rdev)
3771 /* rdev have been moved down */
3772 rdev = conf->disks[i].rdev;
3773 rdev_clear_badblocks(rdev, sh->sector,
3774 STRIPE_SECTORS, 0);
3775 rdev_dec_pending(rdev, conf->mddev);
3776 }
3777 }
3778
3779 if (s.ops_request)
3780 raid_run_ops(sh, s.ops_request);
3781
3782 ops_run_io(sh, &s);
3783
3784 if (s.dec_preread_active) {
3785 /* We delay this until after ops_run_io so that if make_request
3786 * is waiting on a flush, it won't continue until the writes
3787 * have actually been submitted.
3788 */
3789 atomic_dec(&conf->preread_active_stripes);
3790 if (atomic_read(&conf->preread_active_stripes) <
3791 IO_THRESHOLD)
3792 md_wakeup_thread(conf->mddev->thread);
3793 }
3794
3795 return_io(s.return_bi);
3796
3797 clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3798 }
3799
3800 static void raid5_activate_delayed(struct r5conf *conf)
3801 {
3802 if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3803 while (!list_empty(&conf->delayed_list)) {
3804 struct list_head *l = conf->delayed_list.next;
3805 struct stripe_head *sh;
3806 sh = list_entry(l, struct stripe_head, lru);
3807 list_del_init(l);
3808 clear_bit(STRIPE_DELAYED, &sh->state);
3809 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3810 atomic_inc(&conf->preread_active_stripes);
3811 list_add_tail(&sh->lru, &conf->hold_list);
3812 }
3813 }
3814 }
3815
3816 static void activate_bit_delay(struct r5conf *conf)
3817 {
3818 /* device_lock is held */
3819 struct list_head head;
3820 list_add(&head, &conf->bitmap_list);
3821 list_del_init(&conf->bitmap_list);
3822 while (!list_empty(&head)) {
3823 struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3824 list_del_init(&sh->lru);
3825 atomic_inc(&sh->count);
3826 __release_stripe(conf, sh);
3827 }
3828 }
3829
3830 int md_raid5_congested(struct mddev *mddev, int bits)
3831 {
3832 struct r5conf *conf = mddev->private;
3833
3834 /* No difference between reads and writes. Just check
3835 * how busy the stripe_cache is
3836 */
3837
3838 if (conf->inactive_blocked)
3839 return 1;
3840 if (conf->quiesce)
3841 return 1;
3842 if (list_empty_careful(&conf->inactive_list))
3843 return 1;
3844
3845 return 0;
3846 }
3847 EXPORT_SYMBOL_GPL(md_raid5_congested);
3848
3849 static int raid5_congested(void *data, int bits)
3850 {
3851 struct mddev *mddev = data;
3852
3853 return mddev_congested(mddev, bits) ||
3854 md_raid5_congested(mddev, bits);
3855 }
3856
3857 /* We want read requests to align with chunks where possible,
3858 * but write requests don't need to.
3859 */
3860 static int raid5_mergeable_bvec(struct request_queue *q,
3861 struct bvec_merge_data *bvm,
3862 struct bio_vec *biovec)
3863 {
3864 struct mddev *mddev = q->queuedata;
3865 sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3866 int max;
3867 unsigned int chunk_sectors = mddev->chunk_sectors;
3868 unsigned int bio_sectors = bvm->bi_size >> 9;
3869
3870 if ((bvm->bi_rw & 1) == WRITE)
3871 return biovec->bv_len; /* always allow writes to be mergeable */
3872
3873 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3874 chunk_sectors = mddev->new_chunk_sectors;
3875 max = (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3876 if (max < 0) max = 0;
3877 if (max <= biovec->bv_len && bio_sectors == 0)
3878 return biovec->bv_len;
3879 else
3880 return max;
3881 }
3882
3883
3884 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3885 {
3886 sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3887 unsigned int chunk_sectors = mddev->chunk_sectors;
3888 unsigned int bio_sectors = bio_sectors(bio);
3889
3890 if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3891 chunk_sectors = mddev->new_chunk_sectors;
3892 return chunk_sectors >=
3893 ((sector & (chunk_sectors - 1)) + bio_sectors);
3894 }
3895
3896 /*
3897 * add bio to the retry LIFO ( in O(1) ... we are in interrupt )
3898 * later sampled by raid5d.
3899 */
3900 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3901 {
3902 unsigned long flags;
3903
3904 spin_lock_irqsave(&conf->device_lock, flags);
3905
3906 bi->bi_next = conf->retry_read_aligned_list;
3907 conf->retry_read_aligned_list = bi;
3908
3909 spin_unlock_irqrestore(&conf->device_lock, flags);
3910 md_wakeup_thread(conf->mddev->thread);
3911 }
3912
3913
3914 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3915 {
3916 struct bio *bi;
3917
3918 bi = conf->retry_read_aligned;
3919 if (bi) {
3920 conf->retry_read_aligned = NULL;
3921 return bi;
3922 }
3923 bi = conf->retry_read_aligned_list;
3924 if(bi) {
3925 conf->retry_read_aligned_list = bi->bi_next;
3926 bi->bi_next = NULL;
3927 /*
3928 * this sets the active strip count to 1 and the processed
3929 * strip count to zero (upper 8 bits)
3930 */
3931 raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
3932 }
3933
3934 return bi;
3935 }
3936
3937
3938 /*
3939 * The "raid5_align_endio" should check if the read succeeded and if it
3940 * did, call bio_endio on the original bio (having bio_put the new bio
3941 * first).
3942 * If the read failed..
3943 */
3944 static void raid5_align_endio(struct bio *bi, int error)
3945 {
3946 struct bio* raid_bi = bi->bi_private;
3947 struct mddev *mddev;
3948 struct r5conf *conf;
3949 int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3950 struct md_rdev *rdev;
3951
3952 bio_put(bi);
3953
3954 rdev = (void*)raid_bi->bi_next;
3955 raid_bi->bi_next = NULL;
3956 mddev = rdev->mddev;
3957 conf = mddev->private;
3958
3959 rdev_dec_pending(rdev, conf->mddev);
3960
3961 if (!error && uptodate) {
3962 trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
3963 raid_bi, 0);
3964 bio_endio(raid_bi, 0);
3965 if (atomic_dec_and_test(&conf->active_aligned_reads))
3966 wake_up(&conf->wait_for_stripe);
3967 return;
3968 }
3969
3970
3971 pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3972
3973 add_bio_to_retry(raid_bi, conf);
3974 }
3975
3976 static int bio_fits_rdev(struct bio *bi)
3977 {
3978 struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3979
3980 if (bio_sectors(bi) > queue_max_sectors(q))
3981 return 0;
3982 blk_recount_segments(q, bi);
3983 if (bi->bi_phys_segments > queue_max_segments(q))
3984 return 0;
3985
3986 if (q->merge_bvec_fn)
3987 /* it's too hard to apply the merge_bvec_fn at this stage,
3988 * just just give up
3989 */
3990 return 0;
3991
3992 return 1;
3993 }
3994
3995
3996 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3997 {
3998 struct r5conf *conf = mddev->private;
3999 int dd_idx;
4000 struct bio* align_bi;
4001 struct md_rdev *rdev;
4002 sector_t end_sector;
4003
4004 if (!in_chunk_boundary(mddev, raid_bio)) {
4005 pr_debug("chunk_aligned_read : non aligned\n");
4006 return 0;
4007 }
4008 /*
4009 * use bio_clone_mddev to make a copy of the bio
4010 */
4011 align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4012 if (!align_bi)
4013 return 0;
4014 /*
4015 * set bi_end_io to a new function, and set bi_private to the
4016 * original bio.
4017 */
4018 align_bi->bi_end_io = raid5_align_endio;
4019 align_bi->bi_private = raid_bio;
4020 /*
4021 * compute position
4022 */
4023 align_bi->bi_sector = raid5_compute_sector(conf, raid_bio->bi_sector,
4024 0,
4025 &dd_idx, NULL);
4026
4027 end_sector = bio_end_sector(align_bi);
4028 rcu_read_lock();
4029 rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4030 if (!rdev || test_bit(Faulty, &rdev->flags) ||
4031 rdev->recovery_offset < end_sector) {
4032 rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4033 if (rdev &&
4034 (test_bit(Faulty, &rdev->flags) ||
4035 !(test_bit(In_sync, &rdev->flags) ||
4036 rdev->recovery_offset >= end_sector)))
4037 rdev = NULL;
4038 }
4039 if (rdev) {
4040 sector_t first_bad;
4041 int bad_sectors;
4042
4043 atomic_inc(&rdev->nr_pending);
4044 rcu_read_unlock();
4045 raid_bio->bi_next = (void*)rdev;
4046 align_bi->bi_bdev = rdev->bdev;
4047 align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4048
4049 if (!bio_fits_rdev(align_bi) ||
4050 is_badblock(rdev, align_bi->bi_sector, bio_sectors(align_bi),
4051 &first_bad, &bad_sectors)) {
4052 /* too big in some way, or has a known bad block */
4053 bio_put(align_bi);
4054 rdev_dec_pending(rdev, mddev);
4055 return 0;
4056 }
4057
4058 /* No reshape active, so we can trust rdev->data_offset */
4059 align_bi->bi_sector += rdev->data_offset;
4060
4061 spin_lock_irq(&conf->device_lock);
4062 wait_event_lock_irq(conf->wait_for_stripe,
4063 conf->quiesce == 0,
4064 conf->device_lock);
4065 atomic_inc(&conf->active_aligned_reads);
4066 spin_unlock_irq(&conf->device_lock);
4067
4068 if (mddev->gendisk)
4069 trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4070 align_bi, disk_devt(mddev->gendisk),
4071 raid_bio->bi_sector);
4072 generic_make_request(align_bi);
4073 return 1;
4074 } else {
4075 rcu_read_unlock();
4076 bio_put(align_bi);
4077 return 0;
4078 }
4079 }
4080
4081 /* __get_priority_stripe - get the next stripe to process
4082 *
4083 * Full stripe writes are allowed to pass preread active stripes up until
4084 * the bypass_threshold is exceeded. In general the bypass_count
4085 * increments when the handle_list is handled before the hold_list; however, it
4086 * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4087 * stripe with in flight i/o. The bypass_count will be reset when the
4088 * head of the hold_list has changed, i.e. the head was promoted to the
4089 * handle_list.
4090 */
4091 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
4092 {
4093 struct stripe_head *sh;
4094
4095 pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4096 __func__,
4097 list_empty(&conf->handle_list) ? "empty" : "busy",
4098 list_empty(&conf->hold_list) ? "empty" : "busy",
4099 atomic_read(&conf->pending_full_writes), conf->bypass_count);
4100
4101 if (!list_empty(&conf->handle_list)) {
4102 sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
4103
4104 if (list_empty(&conf->hold_list))
4105 conf->bypass_count = 0;
4106 else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4107 if (conf->hold_list.next == conf->last_hold)
4108 conf->bypass_count++;
4109 else {
4110 conf->last_hold = conf->hold_list.next;
4111 conf->bypass_count -= conf->bypass_threshold;
4112 if (conf->bypass_count < 0)
4113 conf->bypass_count = 0;
4114 }
4115 }
4116 } else if (!list_empty(&conf->hold_list) &&
4117 ((conf->bypass_threshold &&
4118 conf->bypass_count > conf->bypass_threshold) ||
4119 atomic_read(&conf->pending_full_writes) == 0)) {
4120 sh = list_entry(conf->hold_list.next,
4121 typeof(*sh), lru);
4122 conf->bypass_count -= conf->bypass_threshold;
4123 if (conf->bypass_count < 0)
4124 conf->bypass_count = 0;
4125 } else
4126 return NULL;
4127
4128 list_del_init(&sh->lru);
4129 atomic_inc(&sh->count);
4130 BUG_ON(atomic_read(&sh->count) != 1);
4131 return sh;
4132 }
4133
4134 struct raid5_plug_cb {
4135 struct blk_plug_cb cb;
4136 struct list_head list;
4137 };
4138
4139 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4140 {
4141 struct raid5_plug_cb *cb = container_of(
4142 blk_cb, struct raid5_plug_cb, cb);
4143 struct stripe_head *sh;
4144 struct mddev *mddev = cb->cb.data;
4145 struct r5conf *conf = mddev->private;
4146 int cnt = 0;
4147
4148 if (cb->list.next && !list_empty(&cb->list)) {
4149 spin_lock_irq(&conf->device_lock);
4150 while (!list_empty(&cb->list)) {
4151 sh = list_first_entry(&cb->list, struct stripe_head, lru);
4152 list_del_init(&sh->lru);
4153 /*
4154 * avoid race release_stripe_plug() sees
4155 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4156 * is still in our list
4157 */
4158 smp_mb__before_clear_bit();
4159 clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4160 __release_stripe(conf, sh);
4161 cnt++;
4162 }
4163 spin_unlock_irq(&conf->device_lock);
4164 }
4165 if (mddev->queue)
4166 trace_block_unplug(mddev->queue, cnt, !from_schedule);
4167 kfree(cb);
4168 }
4169
4170 static void release_stripe_plug(struct mddev *mddev,
4171 struct stripe_head *sh)
4172 {
4173 struct blk_plug_cb *blk_cb = blk_check_plugged(
4174 raid5_unplug, mddev,
4175 sizeof(struct raid5_plug_cb));
4176 struct raid5_plug_cb *cb;
4177
4178 if (!blk_cb) {
4179 release_stripe(sh);
4180 return;
4181 }
4182
4183 cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4184
4185 if (cb->list.next == NULL)
4186 INIT_LIST_HEAD(&cb->list);
4187
4188 if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4189 list_add_tail(&sh->lru, &cb->list);
4190 else
4191 release_stripe(sh);
4192 }
4193
4194 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4195 {
4196 struct r5conf *conf = mddev->private;
4197 sector_t logical_sector, last_sector;
4198 struct stripe_head *sh;
4199 int remaining;
4200 int stripe_sectors;
4201
4202 if (mddev->reshape_position != MaxSector)
4203 /* Skip discard while reshape is happening */
4204 return;
4205
4206 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4207 last_sector = bi->bi_sector + (bi->bi_size>>9);
4208
4209 bi->bi_next = NULL;
4210 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4211
4212 stripe_sectors = conf->chunk_sectors *
4213 (conf->raid_disks - conf->max_degraded);
4214 logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4215 stripe_sectors);
4216 sector_div(last_sector, stripe_sectors);
4217
4218 logical_sector *= conf->chunk_sectors;
4219 last_sector *= conf->chunk_sectors;
4220
4221 for (; logical_sector < last_sector;
4222 logical_sector += STRIPE_SECTORS) {
4223 DEFINE_WAIT(w);
4224 int d;
4225 again:
4226 sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4227 prepare_to_wait(&conf->wait_for_overlap, &w,
4228 TASK_UNINTERRUPTIBLE);
4229 set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4230 if (test_bit(STRIPE_SYNCING, &sh->state)) {
4231 release_stripe(sh);
4232 schedule();
4233 goto again;
4234 }
4235 clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4236 spin_lock_irq(&sh->stripe_lock);
4237 for (d = 0; d < conf->raid_disks; d++) {
4238 if (d == sh->pd_idx || d == sh->qd_idx)
4239 continue;
4240 if (sh->dev[d].towrite || sh->dev[d].toread) {
4241 set_bit(R5_Overlap, &sh->dev[d].flags);
4242 spin_unlock_irq(&sh->stripe_lock);
4243 release_stripe(sh);
4244 schedule();
4245 goto again;
4246 }
4247 }
4248 set_bit(STRIPE_DISCARD, &sh->state);
4249 finish_wait(&conf->wait_for_overlap, &w);
4250 for (d = 0; d < conf->raid_disks; d++) {
4251 if (d == sh->pd_idx || d == sh->qd_idx)
4252 continue;
4253 sh->dev[d].towrite = bi;
4254 set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4255 raid5_inc_bi_active_stripes(bi);
4256 }
4257 spin_unlock_irq(&sh->stripe_lock);
4258 if (conf->mddev->bitmap) {
4259 for (d = 0;
4260 d < conf->raid_disks - conf->max_degraded;
4261 d++)
4262 bitmap_startwrite(mddev->bitmap,
4263 sh->sector,
4264 STRIPE_SECTORS,
4265 0);
4266 sh->bm_seq = conf->seq_flush + 1;
4267 set_bit(STRIPE_BIT_DELAY, &sh->state);
4268 }
4269
4270 set_bit(STRIPE_HANDLE, &sh->state);
4271 clear_bit(STRIPE_DELAYED, &sh->state);
4272 if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4273 atomic_inc(&conf->preread_active_stripes);
4274 release_stripe_plug(mddev, sh);
4275 }
4276
4277 remaining = raid5_dec_bi_active_stripes(bi);
4278 if (remaining == 0) {
4279 md_write_end(mddev);
4280 bio_endio(bi, 0);
4281 }
4282 }
4283
4284 static void make_request(struct mddev *mddev, struct bio * bi)
4285 {
4286 struct r5conf *conf = mddev->private;
4287 int dd_idx;
4288 sector_t new_sector;
4289 sector_t logical_sector, last_sector;
4290 struct stripe_head *sh;
4291 const int rw = bio_data_dir(bi);
4292 int remaining;
4293
4294 if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4295 md_flush_request(mddev, bi);
4296 return;
4297 }
4298
4299 md_write_start(mddev, bi);
4300
4301 if (rw == READ &&
4302 mddev->reshape_position == MaxSector &&
4303 chunk_aligned_read(mddev,bi))
4304 return;
4305
4306 if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4307 make_discard_request(mddev, bi);
4308 return;
4309 }
4310
4311 logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4312 last_sector = bio_end_sector(bi);
4313 bi->bi_next = NULL;
4314 bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4315
4316 for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4317 DEFINE_WAIT(w);
4318 int previous;
4319
4320 retry:
4321 previous = 0;
4322 prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4323 if (unlikely(conf->reshape_progress != MaxSector)) {
4324 /* spinlock is needed as reshape_progress may be
4325 * 64bit on a 32bit platform, and so it might be
4326 * possible to see a half-updated value
4327 * Of course reshape_progress could change after
4328 * the lock is dropped, so once we get a reference
4329 * to the stripe that we think it is, we will have
4330 * to check again.
4331 */
4332 spin_lock_irq(&conf->device_lock);
4333 if (mddev->reshape_backwards
4334 ? logical_sector < conf->reshape_progress
4335 : logical_sector >= conf->reshape_progress) {
4336 previous = 1;
4337 } else {
4338 if (mddev->reshape_backwards
4339 ? logical_sector < conf->reshape_safe
4340 : logical_sector >= conf->reshape_safe) {
4341 spin_unlock_irq(&conf->device_lock);
4342 schedule();
4343 goto retry;
4344 }
4345 }
4346 spin_unlock_irq(&conf->device_lock);
4347 }
4348
4349 new_sector = raid5_compute_sector(conf, logical_sector,
4350 previous,
4351 &dd_idx, NULL);
4352 pr_debug("raid456: make_request, sector %llu logical %llu\n",
4353 (unsigned long long)new_sector,
4354 (unsigned long long)logical_sector);
4355
4356 sh = get_active_stripe(conf, new_sector, previous,
4357 (bi->bi_rw&RWA_MASK), 0);
4358 if (sh) {
4359 if (unlikely(previous)) {
4360 /* expansion might have moved on while waiting for a
4361 * stripe, so we must do the range check again.
4362 * Expansion could still move past after this
4363 * test, but as we are holding a reference to
4364 * 'sh', we know that if that happens,
4365 * STRIPE_EXPANDING will get set and the expansion
4366 * won't proceed until we finish with the stripe.
4367 */
4368 int must_retry = 0;
4369 spin_lock_irq(&conf->device_lock);
4370 if (mddev->reshape_backwards
4371 ? logical_sector >= conf->reshape_progress
4372 : logical_sector < conf->reshape_progress)
4373 /* mismatch, need to try again */
4374 must_retry = 1;
4375 spin_unlock_irq(&conf->device_lock);
4376 if (must_retry) {
4377 release_stripe(sh);
4378 schedule();
4379 goto retry;
4380 }
4381 }
4382
4383 if (rw == WRITE &&
4384 logical_sector >= mddev->suspend_lo &&
4385 logical_sector < mddev->suspend_hi) {
4386 release_stripe(sh);
4387 /* As the suspend_* range is controlled by
4388 * userspace, we want an interruptible
4389 * wait.
4390 */
4391 flush_signals(current);
4392 prepare_to_wait(&conf->wait_for_overlap,
4393 &w, TASK_INTERRUPTIBLE);
4394 if (logical_sector >= mddev->suspend_lo &&
4395 logical_sector < mddev->suspend_hi)
4396 schedule();
4397 goto retry;
4398 }
4399
4400 if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4401 !add_stripe_bio(sh, bi, dd_idx, rw)) {
4402 /* Stripe is busy expanding or
4403 * add failed due to overlap. Flush everything
4404 * and wait a while
4405 */
4406 md_wakeup_thread(mddev->thread);
4407 release_stripe(sh);
4408 schedule();
4409 goto retry;
4410 }
4411 finish_wait(&conf->wait_for_overlap, &w);
4412 set_bit(STRIPE_HANDLE, &sh->state);
4413 clear_bit(STRIPE_DELAYED, &sh->state);
4414 if ((bi->bi_rw & REQ_SYNC) &&
4415 !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4416 atomic_inc(&conf->preread_active_stripes);
4417 release_stripe_plug(mddev, sh);
4418 } else {
4419 /* cannot get stripe for read-ahead, just give-up */
4420 clear_bit(BIO_UPTODATE, &bi->bi_flags);
4421 finish_wait(&conf->wait_for_overlap, &w);
4422 break;
4423 }
4424 }
4425
4426 remaining = raid5_dec_bi_active_stripes(bi);
4427 if (remaining == 0) {
4428
4429 if ( rw == WRITE )
4430 md_write_end(mddev);
4431
4432 trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4433 bi, 0);
4434 bio_endio(bi, 0);
4435 }
4436 }
4437
4438 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4439
4440 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4441 {
4442 /* reshaping is quite different to recovery/resync so it is
4443 * handled quite separately ... here.
4444 *
4445 * On each call to sync_request, we gather one chunk worth of
4446 * destination stripes and flag them as expanding.
4447 * Then we find all the source stripes and request reads.
4448 * As the reads complete, handle_stripe will copy the data
4449 * into the destination stripe and release that stripe.
4450 */
4451 struct r5conf *conf = mddev->private;
4452 struct stripe_head *sh;
4453 sector_t first_sector, last_sector;
4454 int raid_disks = conf->previous_raid_disks;
4455 int data_disks = raid_disks - conf->max_degraded;
4456 int new_data_disks = conf->raid_disks - conf->max_degraded;
4457 int i;
4458 int dd_idx;
4459 sector_t writepos, readpos, safepos;
4460 sector_t stripe_addr;
4461 int reshape_sectors;
4462 struct list_head stripes;
4463
4464 if (sector_nr == 0) {
4465 /* If restarting in the middle, skip the initial sectors */
4466 if (mddev->reshape_backwards &&
4467 conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4468 sector_nr = raid5_size(mddev, 0, 0)
4469 - conf->reshape_progress;
4470 } else if (!mddev->reshape_backwards &&
4471 conf->reshape_progress > 0)
4472 sector_nr = conf->reshape_progress;
4473 sector_div(sector_nr, new_data_disks);
4474 if (sector_nr) {
4475 mddev->curr_resync_completed = sector_nr;
4476 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4477 *skipped = 1;
4478 return sector_nr;
4479 }
4480 }
4481
4482 /* We need to process a full chunk at a time.
4483 * If old and new chunk sizes differ, we need to process the
4484 * largest of these
4485 */
4486 if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4487 reshape_sectors = mddev->new_chunk_sectors;
4488 else
4489 reshape_sectors = mddev->chunk_sectors;
4490
4491 /* We update the metadata at least every 10 seconds, or when
4492 * the data about to be copied would over-write the source of
4493 * the data at the front of the range. i.e. one new_stripe
4494 * along from reshape_progress new_maps to after where
4495 * reshape_safe old_maps to
4496 */
4497 writepos = conf->reshape_progress;
4498 sector_div(writepos, new_data_disks);
4499 readpos = conf->reshape_progress;
4500 sector_div(readpos, data_disks);
4501 safepos = conf->reshape_safe;
4502 sector_div(safepos, data_disks);
4503 if (mddev->reshape_backwards) {
4504 writepos -= min_t(sector_t, reshape_sectors, writepos);
4505 readpos += reshape_sectors;
4506 safepos += reshape_sectors;
4507 } else {
4508 writepos += reshape_sectors;
4509 readpos -= min_t(sector_t, reshape_sectors, readpos);
4510 safepos -= min_t(sector_t, reshape_sectors, safepos);
4511 }
4512
4513 /* Having calculated the 'writepos' possibly use it
4514 * to set 'stripe_addr' which is where we will write to.
4515 */
4516 if (mddev->reshape_backwards) {
4517 BUG_ON(conf->reshape_progress == 0);
4518 stripe_addr = writepos;
4519 BUG_ON((mddev->dev_sectors &
4520 ~((sector_t)reshape_sectors - 1))
4521 - reshape_sectors - stripe_addr
4522 != sector_nr);
4523 } else {
4524 BUG_ON(writepos != sector_nr + reshape_sectors);
4525 stripe_addr = sector_nr;
4526 }
4527
4528 /* 'writepos' is the most advanced device address we might write.
4529 * 'readpos' is the least advanced device address we might read.
4530 * 'safepos' is the least address recorded in the metadata as having
4531 * been reshaped.
4532 * If there is a min_offset_diff, these are adjusted either by
4533 * increasing the safepos/readpos if diff is negative, or
4534 * increasing writepos if diff is positive.
4535 * If 'readpos' is then behind 'writepos', there is no way that we can
4536 * ensure safety in the face of a crash - that must be done by userspace
4537 * making a backup of the data. So in that case there is no particular
4538 * rush to update metadata.
4539 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4540 * update the metadata to advance 'safepos' to match 'readpos' so that
4541 * we can be safe in the event of a crash.
4542 * So we insist on updating metadata if safepos is behind writepos and
4543 * readpos is beyond writepos.
4544 * In any case, update the metadata every 10 seconds.
4545 * Maybe that number should be configurable, but I'm not sure it is
4546 * worth it.... maybe it could be a multiple of safemode_delay???
4547 */
4548 if (conf->min_offset_diff < 0) {
4549 safepos += -conf->min_offset_diff;
4550 readpos += -conf->min_offset_diff;
4551 } else
4552 writepos += conf->min_offset_diff;
4553
4554 if ((mddev->reshape_backwards
4555 ? (safepos > writepos && readpos < writepos)
4556 : (safepos < writepos && readpos > writepos)) ||
4557 time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4558 /* Cannot proceed until we've updated the superblock... */
4559 wait_event(conf->wait_for_overlap,
4560 atomic_read(&conf->reshape_stripes)==0);
4561 mddev->reshape_position = conf->reshape_progress;
4562 mddev->curr_resync_completed = sector_nr;
4563 conf->reshape_checkpoint = jiffies;
4564 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4565 md_wakeup_thread(mddev->thread);
4566 wait_event(mddev->sb_wait, mddev->flags == 0 ||
4567 kthread_should_stop());
4568 spin_lock_irq(&conf->device_lock);
4569 conf->reshape_safe = mddev->reshape_position;
4570 spin_unlock_irq(&conf->device_lock);
4571 wake_up(&conf->wait_for_overlap);
4572 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4573 }
4574
4575 INIT_LIST_HEAD(&stripes);
4576 for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4577 int j;
4578 int skipped_disk = 0;
4579 sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4580 set_bit(STRIPE_EXPANDING, &sh->state);
4581 atomic_inc(&conf->reshape_stripes);
4582 /* If any of this stripe is beyond the end of the old
4583 * array, then we need to zero those blocks
4584 */
4585 for (j=sh->disks; j--;) {
4586 sector_t s;
4587 if (j == sh->pd_idx)
4588 continue;
4589 if (conf->level == 6 &&
4590 j == sh->qd_idx)
4591 continue;
4592 s = compute_blocknr(sh, j, 0);
4593 if (s < raid5_size(mddev, 0, 0)) {
4594 skipped_disk = 1;
4595 continue;
4596 }
4597 memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4598 set_bit(R5_Expanded, &sh->dev[j].flags);
4599 set_bit(R5_UPTODATE, &sh->dev[j].flags);
4600 }
4601 if (!skipped_disk) {
4602 set_bit(STRIPE_EXPAND_READY, &sh->state);
4603 set_bit(STRIPE_HANDLE, &sh->state);
4604 }
4605 list_add(&sh->lru, &stripes);
4606 }
4607 spin_lock_irq(&conf->device_lock);
4608 if (mddev->reshape_backwards)
4609 conf->reshape_progress -= reshape_sectors * new_data_disks;
4610 else
4611 conf->reshape_progress += reshape_sectors * new_data_disks;
4612 spin_unlock_irq(&conf->device_lock);
4613 /* Ok, those stripe are ready. We can start scheduling
4614 * reads on the source stripes.
4615 * The source stripes are determined by mapping the first and last
4616 * block on the destination stripes.
4617 */
4618 first_sector =
4619 raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4620 1, &dd_idx, NULL);
4621 last_sector =
4622 raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4623 * new_data_disks - 1),
4624 1, &dd_idx, NULL);
4625 if (last_sector >= mddev->dev_sectors)
4626 last_sector = mddev->dev_sectors - 1;
4627 while (first_sector <= last_sector) {
4628 sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4629 set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4630 set_bit(STRIPE_HANDLE, &sh->state);
4631 release_stripe(sh);
4632 first_sector += STRIPE_SECTORS;
4633 }
4634 /* Now that the sources are clearly marked, we can release
4635 * the destination stripes
4636 */
4637 while (!list_empty(&stripes)) {
4638 sh = list_entry(stripes.next, struct stripe_head, lru);
4639 list_del_init(&sh->lru);
4640 release_stripe(sh);
4641 }
4642 /* If this takes us to the resync_max point where we have to pause,
4643 * then we need to write out the superblock.
4644 */
4645 sector_nr += reshape_sectors;
4646 if ((sector_nr - mddev->curr_resync_completed) * 2
4647 >= mddev->resync_max - mddev->curr_resync_completed) {
4648 /* Cannot proceed until we've updated the superblock... */
4649 wait_event(conf->wait_for_overlap,
4650 atomic_read(&conf->reshape_stripes) == 0);
4651 mddev->reshape_position = conf->reshape_progress;
4652 mddev->curr_resync_completed = sector_nr;
4653 conf->reshape_checkpoint = jiffies;
4654 set_bit(MD_CHANGE_DEVS, &mddev->flags);
4655 md_wakeup_thread(mddev->thread);
4656 wait_event(mddev->sb_wait,
4657 !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4658 || kthread_should_stop());
4659 spin_lock_irq(&conf->device_lock);
4660 conf->reshape_safe = mddev->reshape_position;
4661 spin_unlock_irq(&conf->device_lock);
4662 wake_up(&conf->wait_for_overlap);
4663 sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4664 }
4665 return reshape_sectors;
4666 }
4667
4668 /* FIXME go_faster isn't used */
4669 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4670 {
4671 struct r5conf *conf = mddev->private;
4672 struct stripe_head *sh;
4673 sector_t max_sector = mddev->dev_sectors;
4674 sector_t sync_blocks;
4675 int still_degraded = 0;
4676 int i;
4677
4678 if (sector_nr >= max_sector) {
4679 /* just being told to finish up .. nothing much to do */
4680
4681 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4682 end_reshape(conf);
4683 return 0;
4684 }
4685
4686 if (mddev->curr_resync < max_sector) /* aborted */
4687 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4688 &sync_blocks, 1);
4689 else /* completed sync */
4690 conf->fullsync = 0;
4691 bitmap_close_sync(mddev->bitmap);
4692
4693 return 0;
4694 }
4695
4696 /* Allow raid5_quiesce to complete */
4697 wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4698
4699 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4700 return reshape_request(mddev, sector_nr, skipped);
4701
4702 /* No need to check resync_max as we never do more than one
4703 * stripe, and as resync_max will always be on a chunk boundary,
4704 * if the check in md_do_sync didn't fire, there is no chance
4705 * of overstepping resync_max here
4706 */
4707
4708 /* if there is too many failed drives and we are trying
4709 * to resync, then assert that we are finished, because there is
4710 * nothing we can do.
4711 */
4712 if (mddev->degraded >= conf->max_degraded &&
4713 test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4714 sector_t rv = mddev->dev_sectors - sector_nr;
4715 *skipped = 1;
4716 return rv;
4717 }
4718 if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4719 !conf->fullsync &&
4720 !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4721 sync_blocks >= STRIPE_SECTORS) {
4722 /* we can skip this block, and probably more */
4723 sync_blocks /= STRIPE_SECTORS;
4724 *skipped = 1;
4725 return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4726 }
4727
4728 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4729
4730 sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4731 if (sh == NULL) {
4732 sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4733 /* make sure we don't swamp the stripe cache if someone else
4734 * is trying to get access
4735 */
4736 schedule_timeout_uninterruptible(1);
4737 }
4738 /* Need to check if array will still be degraded after recovery/resync
4739 * We don't need to check the 'failed' flag as when that gets set,
4740 * recovery aborts.
4741 */
4742 for (i = 0; i < conf->raid_disks; i++)
4743 if (conf->disks[i].rdev == NULL)
4744 still_degraded = 1;
4745
4746 bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4747
4748 set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4749
4750 handle_stripe(sh);
4751 release_stripe(sh);
4752
4753 return STRIPE_SECTORS;
4754 }
4755
4756 static int retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4757 {
4758 /* We may not be able to submit a whole bio at once as there
4759 * may not be enough stripe_heads available.
4760 * We cannot pre-allocate enough stripe_heads as we may need
4761 * more than exist in the cache (if we allow ever large chunks).
4762 * So we do one stripe head at a time and record in
4763 * ->bi_hw_segments how many have been done.
4764 *
4765 * We *know* that this entire raid_bio is in one chunk, so
4766 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4767 */
4768 struct stripe_head *sh;
4769 int dd_idx;
4770 sector_t sector, logical_sector, last_sector;
4771 int scnt = 0;
4772 int remaining;
4773 int handled = 0;
4774
4775 logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4776 sector = raid5_compute_sector(conf, logical_sector,
4777 0, &dd_idx, NULL);
4778 last_sector = bio_end_sector(raid_bio);
4779
4780 for (; logical_sector < last_sector;
4781 logical_sector += STRIPE_SECTORS,
4782 sector += STRIPE_SECTORS,
4783 scnt++) {
4784
4785 if (scnt < raid5_bi_processed_stripes(raid_bio))
4786 /* already done this stripe */
4787 continue;
4788
4789 sh = get_active_stripe(conf, sector, 0, 1, 0);
4790
4791 if (!sh) {
4792 /* failed to get a stripe - must wait */
4793 raid5_set_bi_processed_stripes(raid_bio, scnt);
4794 conf->retry_read_aligned = raid_bio;
4795 return handled;
4796 }
4797
4798 if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4799 release_stripe(sh);
4800 raid5_set_bi_processed_stripes(raid_bio, scnt);
4801 conf->retry_read_aligned = raid_bio;
4802 return handled;
4803 }
4804
4805 set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4806 handle_stripe(sh);
4807 release_stripe(sh);
4808 handled++;
4809 }
4810 remaining = raid5_dec_bi_active_stripes(raid_bio);
4811 if (remaining == 0) {
4812 trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
4813 raid_bio, 0);
4814 bio_endio(raid_bio, 0);
4815 }
4816 if (atomic_dec_and_test(&conf->active_aligned_reads))
4817 wake_up(&conf->wait_for_stripe);
4818 return handled;
4819 }
4820
4821 #define MAX_STRIPE_BATCH 8
4822 static int handle_active_stripes(struct r5conf *conf)
4823 {
4824 struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4825 int i, batch_size = 0;
4826
4827 while (batch_size < MAX_STRIPE_BATCH &&
4828 (sh = __get_priority_stripe(conf)) != NULL)
4829 batch[batch_size++] = sh;
4830
4831 if (batch_size == 0)
4832 return batch_size;
4833 spin_unlock_irq(&conf->device_lock);
4834
4835 for (i = 0; i < batch_size; i++)
4836 handle_stripe(batch[i]);
4837
4838 cond_resched();
4839
4840 spin_lock_irq(&conf->device_lock);
4841 for (i = 0; i < batch_size; i++)
4842 __release_stripe(conf, batch[i]);
4843 return batch_size;
4844 }
4845
4846 /*
4847 * This is our raid5 kernel thread.
4848 *
4849 * We scan the hash table for stripes which can be handled now.
4850 * During the scan, completed stripes are saved for us by the interrupt
4851 * handler, so that they will not have to wait for our next wakeup.
4852 */
4853 static void raid5d(struct md_thread *thread)
4854 {
4855 struct mddev *mddev = thread->mddev;
4856 struct r5conf *conf = mddev->private;
4857 int handled;
4858 struct blk_plug plug;
4859
4860 pr_debug("+++ raid5d active\n");
4861
4862 md_check_recovery(mddev);
4863
4864 blk_start_plug(&plug);
4865 handled = 0;
4866 spin_lock_irq(&conf->device_lock);
4867 while (1) {
4868 struct bio *bio;
4869 int batch_size;
4870
4871 if (
4872 !list_empty(&conf->bitmap_list)) {
4873 /* Now is a good time to flush some bitmap updates */
4874 conf->seq_flush++;
4875 spin_unlock_irq(&conf->device_lock);
4876 bitmap_unplug(mddev->bitmap);
4877 spin_lock_irq(&conf->device_lock);
4878 conf->seq_write = conf->seq_flush;
4879 activate_bit_delay(conf);
4880 }
4881 raid5_activate_delayed(conf);
4882
4883 while ((bio = remove_bio_from_retry(conf))) {
4884 int ok;
4885 spin_unlock_irq(&conf->device_lock);
4886 ok = retry_aligned_read(conf, bio);
4887 spin_lock_irq(&conf->device_lock);
4888 if (!ok)
4889 break;
4890 handled++;
4891 }
4892
4893 batch_size = handle_active_stripes(conf);
4894 if (!batch_size)
4895 break;
4896 handled += batch_size;
4897
4898 if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
4899 spin_unlock_irq(&conf->device_lock);
4900 md_check_recovery(mddev);
4901 spin_lock_irq(&conf->device_lock);
4902 }
4903 }
4904 pr_debug("%d stripes handled\n", handled);
4905
4906 spin_unlock_irq(&conf->device_lock);
4907
4908 async_tx_issue_pending_all();
4909 blk_finish_plug(&plug);
4910
4911 pr_debug("--- raid5d inactive\n");
4912 }
4913
4914 static ssize_t
4915 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4916 {
4917 struct r5conf *conf = mddev->private;
4918 if (conf)
4919 return sprintf(page, "%d\n", conf->max_nr_stripes);
4920 else
4921 return 0;
4922 }
4923
4924 int
4925 raid5_set_cache_size(struct mddev *mddev, int size)
4926 {
4927 struct r5conf *conf = mddev->private;
4928 int err;
4929
4930 if (size <= 16 || size > 32768)
4931 return -EINVAL;
4932 while (size < conf->max_nr_stripes) {
4933 if (drop_one_stripe(conf))
4934 conf->max_nr_stripes--;
4935 else
4936 break;
4937 }
4938 err = md_allow_write(mddev);
4939 if (err)
4940 return err;
4941 while (size > conf->max_nr_stripes) {
4942 if (grow_one_stripe(conf))
4943 conf->max_nr_stripes++;
4944 else break;
4945 }
4946 return 0;
4947 }
4948 EXPORT_SYMBOL(raid5_set_cache_size);
4949
4950 static ssize_t
4951 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4952 {
4953 struct r5conf *conf = mddev->private;
4954 unsigned long new;
4955 int err;
4956
4957 if (len >= PAGE_SIZE)
4958 return -EINVAL;
4959 if (!conf)
4960 return -ENODEV;
4961
4962 if (strict_strtoul(page, 10, &new))
4963 return -EINVAL;
4964 err = raid5_set_cache_size(mddev, new);
4965 if (err)
4966 return err;
4967 return len;
4968 }
4969
4970 static struct md_sysfs_entry
4971 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4972 raid5_show_stripe_cache_size,
4973 raid5_store_stripe_cache_size);
4974
4975 static ssize_t
4976 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4977 {
4978 struct r5conf *conf = mddev->private;
4979 if (conf)
4980 return sprintf(page, "%d\n", conf->bypass_threshold);
4981 else
4982 return 0;
4983 }
4984
4985 static ssize_t
4986 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4987 {
4988 struct r5conf *conf = mddev->private;
4989 unsigned long new;
4990 if (len >= PAGE_SIZE)
4991 return -EINVAL;
4992 if (!conf)
4993 return -ENODEV;
4994
4995 if (strict_strtoul(page, 10, &new))
4996 return -EINVAL;
4997 if (new > conf->max_nr_stripes)
4998 return -EINVAL;
4999 conf->bypass_threshold = new;
5000 return len;
5001 }
5002
5003 static struct md_sysfs_entry
5004 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5005 S_IRUGO | S_IWUSR,
5006 raid5_show_preread_threshold,
5007 raid5_store_preread_threshold);
5008
5009 static ssize_t
5010 stripe_cache_active_show(struct mddev *mddev, char *page)
5011 {
5012 struct r5conf *conf = mddev->private;
5013 if (conf)
5014 return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5015 else
5016 return 0;
5017 }
5018
5019 static struct md_sysfs_entry
5020 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
5021
5022 static struct attribute *raid5_attrs[] = {
5023 &raid5_stripecache_size.attr,
5024 &raid5_stripecache_active.attr,
5025 &raid5_preread_bypass_threshold.attr,
5026 NULL,
5027 };
5028 static struct attribute_group raid5_attrs_group = {
5029 .name = NULL,
5030 .attrs = raid5_attrs,
5031 };
5032
5033 static sector_t
5034 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
5035 {
5036 struct r5conf *conf = mddev->private;
5037
5038 if (!sectors)
5039 sectors = mddev->dev_sectors;
5040 if (!raid_disks)
5041 /* size is defined by the smallest of previous and new size */
5042 raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
5043
5044 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5045 sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
5046 return sectors * (raid_disks - conf->max_degraded);
5047 }
5048
5049 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5050 {
5051 safe_put_page(percpu->spare_page);
5052 kfree(percpu->scribble);
5053 percpu->spare_page = NULL;
5054 percpu->scribble = NULL;
5055 }
5056
5057 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
5058 {
5059 if (conf->level == 6 && !percpu->spare_page)
5060 percpu->spare_page = alloc_page(GFP_KERNEL);
5061 if (!percpu->scribble)
5062 percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5063
5064 if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
5065 free_scratch_buffer(conf, percpu);
5066 return -ENOMEM;
5067 }
5068
5069 return 0;
5070 }
5071
5072 static void raid5_free_percpu(struct r5conf *conf)
5073 {
5074 unsigned long cpu;
5075
5076 if (!conf->percpu)
5077 return;
5078
5079 #ifdef CONFIG_HOTPLUG_CPU
5080 unregister_cpu_notifier(&conf->cpu_notify);
5081 #endif
5082
5083 get_online_cpus();
5084 for_each_possible_cpu(cpu)
5085 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5086 put_online_cpus();
5087
5088 free_percpu(conf->percpu);
5089 }
5090
5091 static void free_conf(struct r5conf *conf)
5092 {
5093 shrink_stripes(conf);
5094 raid5_free_percpu(conf);
5095 kfree(conf->disks);
5096 kfree(conf->stripe_hashtbl);
5097 kfree(conf);
5098 }
5099
5100 #ifdef CONFIG_HOTPLUG_CPU
5101 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5102 void *hcpu)
5103 {
5104 struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5105 long cpu = (long)hcpu;
5106 struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5107
5108 switch (action) {
5109 case CPU_UP_PREPARE:
5110 case CPU_UP_PREPARE_FROZEN:
5111 if (alloc_scratch_buffer(conf, percpu)) {
5112 pr_err("%s: failed memory allocation for cpu%ld\n",
5113 __func__, cpu);
5114 return notifier_from_errno(-ENOMEM);
5115 }
5116 break;
5117 case CPU_DEAD:
5118 case CPU_DEAD_FROZEN:
5119 free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5120 break;
5121 default:
5122 break;
5123 }
5124 return NOTIFY_OK;
5125 }
5126 #endif
5127
5128 static int raid5_alloc_percpu(struct r5conf *conf)
5129 {
5130 unsigned long cpu;
5131 int err = 0;
5132
5133 conf->percpu = alloc_percpu(struct raid5_percpu);
5134 if (!conf->percpu)
5135 return -ENOMEM;
5136
5137 #ifdef CONFIG_HOTPLUG_CPU
5138 conf->cpu_notify.notifier_call = raid456_cpu_notify;
5139 conf->cpu_notify.priority = 0;
5140 err = register_cpu_notifier(&conf->cpu_notify);
5141 if (err)
5142 return err;
5143 #endif
5144
5145 get_online_cpus();
5146 for_each_present_cpu(cpu) {
5147 err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
5148 if (err) {
5149 pr_err("%s: failed memory allocation for cpu%ld\n",
5150 __func__, cpu);
5151 break;
5152 }
5153 }
5154 put_online_cpus();
5155
5156 return err;
5157 }
5158
5159 static struct r5conf *setup_conf(struct mddev *mddev)
5160 {
5161 struct r5conf *conf;
5162 int raid_disk, memory, max_disks;
5163 struct md_rdev *rdev;
5164 struct disk_info *disk;
5165 char pers_name[6];
5166
5167 if (mddev->new_level != 5
5168 && mddev->new_level != 4
5169 && mddev->new_level != 6) {
5170 printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5171 mdname(mddev), mddev->new_level);
5172 return ERR_PTR(-EIO);
5173 }
5174 if ((mddev->new_level == 5
5175 && !algorithm_valid_raid5(mddev->new_layout)) ||
5176 (mddev->new_level == 6
5177 && !algorithm_valid_raid6(mddev->new_layout))) {
5178 printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5179 mdname(mddev), mddev->new_layout);
5180 return ERR_PTR(-EIO);
5181 }
5182 if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5183 printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5184 mdname(mddev), mddev->raid_disks);
5185 return ERR_PTR(-EINVAL);
5186 }
5187
5188 if (!mddev->new_chunk_sectors ||
5189 (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5190 !is_power_of_2(mddev->new_chunk_sectors)) {
5191 printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5192 mdname(mddev), mddev->new_chunk_sectors << 9);
5193 return ERR_PTR(-EINVAL);
5194 }
5195
5196 conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5197 if (conf == NULL)
5198 goto abort;
5199 spin_lock_init(&conf->device_lock);
5200 init_waitqueue_head(&conf->wait_for_stripe);
5201 init_waitqueue_head(&conf->wait_for_overlap);
5202 INIT_LIST_HEAD(&conf->handle_list);
5203 INIT_LIST_HEAD(&conf->hold_list);
5204 INIT_LIST_HEAD(&conf->delayed_list);
5205 INIT_LIST_HEAD(&conf->bitmap_list);
5206 INIT_LIST_HEAD(&conf->inactive_list);
5207 atomic_set(&conf->active_stripes, 0);
5208 atomic_set(&conf->preread_active_stripes, 0);
5209 atomic_set(&conf->active_aligned_reads, 0);
5210 conf->bypass_threshold = BYPASS_THRESHOLD;
5211 conf->recovery_disabled = mddev->recovery_disabled - 1;
5212
5213 conf->raid_disks = mddev->raid_disks;
5214 if (mddev->reshape_position == MaxSector)
5215 conf->previous_raid_disks = mddev->raid_disks;
5216 else
5217 conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5218 max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5219 conf->scribble_len = scribble_len(max_disks);
5220
5221 conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5222 GFP_KERNEL);
5223 if (!conf->disks)
5224 goto abort;
5225
5226 conf->mddev = mddev;
5227
5228 if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5229 goto abort;
5230
5231 conf->level = mddev->new_level;
5232 if (raid5_alloc_percpu(conf) != 0)
5233 goto abort;
5234
5235 pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5236
5237 rdev_for_each(rdev, mddev) {
5238 raid_disk = rdev->raid_disk;
5239 if (raid_disk >= max_disks
5240 || raid_disk < 0)
5241 continue;
5242 disk = conf->disks + raid_disk;
5243
5244 if (test_bit(Replacement, &rdev->flags)) {
5245 if (disk->replacement)
5246 goto abort;
5247 disk->replacement = rdev;
5248 } else {
5249 if (disk->rdev)
5250 goto abort;
5251 disk->rdev = rdev;
5252 }
5253
5254 if (test_bit(In_sync, &rdev->flags)) {
5255 char b[BDEVNAME_SIZE];
5256 printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5257 " disk %d\n",
5258 mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5259 } else if (rdev->saved_raid_disk != raid_disk)
5260 /* Cannot rely on bitmap to complete recovery */
5261 conf->fullsync = 1;
5262 }
5263
5264 conf->chunk_sectors = mddev->new_chunk_sectors;
5265 conf->level = mddev->new_level;
5266 if (conf->level == 6)
5267 conf->max_degraded = 2;
5268 else
5269 conf->max_degraded = 1;
5270 conf->algorithm = mddev->new_layout;
5271 conf->max_nr_stripes = NR_STRIPES;
5272 conf->reshape_progress = mddev->reshape_position;
5273 if (conf->reshape_progress != MaxSector) {
5274 conf->prev_chunk_sectors = mddev->chunk_sectors;
5275 conf->prev_algo = mddev->layout;
5276 }
5277
5278 memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5279 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5280 if (grow_stripes(conf, conf->max_nr_stripes)) {
5281 printk(KERN_ERR
5282 "md/raid:%s: couldn't allocate %dkB for buffers\n",
5283 mdname(mddev), memory);
5284 goto abort;
5285 } else
5286 printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5287 mdname(mddev), memory);
5288
5289 sprintf(pers_name, "raid%d", mddev->new_level);
5290 conf->thread = md_register_thread(raid5d, mddev, pers_name);
5291 if (!conf->thread) {
5292 printk(KERN_ERR
5293 "md/raid:%s: couldn't allocate thread.\n",
5294 mdname(mddev));
5295 goto abort;
5296 }
5297
5298 return conf;
5299
5300 abort:
5301 if (conf) {
5302 free_conf(conf);
5303 return ERR_PTR(-EIO);
5304 } else
5305 return ERR_PTR(-ENOMEM);
5306 }
5307
5308
5309 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5310 {
5311 switch (algo) {
5312 case ALGORITHM_PARITY_0:
5313 if (raid_disk < max_degraded)
5314 return 1;
5315 break;
5316 case ALGORITHM_PARITY_N:
5317 if (raid_disk >= raid_disks - max_degraded)
5318 return 1;
5319 break;
5320 case ALGORITHM_PARITY_0_6:
5321 if (raid_disk == 0 ||
5322 raid_disk == raid_disks - 1)
5323 return 1;
5324 break;
5325 case ALGORITHM_LEFT_ASYMMETRIC_6:
5326 case ALGORITHM_RIGHT_ASYMMETRIC_6:
5327 case ALGORITHM_LEFT_SYMMETRIC_6:
5328 case ALGORITHM_RIGHT_SYMMETRIC_6:
5329 if (raid_disk == raid_disks - 1)
5330 return 1;
5331 }
5332 return 0;
5333 }
5334
5335 static int run(struct mddev *mddev)
5336 {
5337 struct r5conf *conf;
5338 int working_disks = 0;
5339 int dirty_parity_disks = 0;
5340 struct md_rdev *rdev;
5341 sector_t reshape_offset = 0;
5342 int i;
5343 long long min_offset_diff = 0;
5344 int first = 1;
5345
5346 if (mddev->recovery_cp != MaxSector)
5347 printk(KERN_NOTICE "md/raid:%s: not clean"
5348 " -- starting background reconstruction\n",
5349 mdname(mddev));
5350
5351 rdev_for_each(rdev, mddev) {
5352 long long diff;
5353 if (rdev->raid_disk < 0)
5354 continue;
5355 diff = (rdev->new_data_offset - rdev->data_offset);
5356 if (first) {
5357 min_offset_diff = diff;
5358 first = 0;
5359 } else if (mddev->reshape_backwards &&
5360 diff < min_offset_diff)
5361 min_offset_diff = diff;
5362 else if (!mddev->reshape_backwards &&
5363 diff > min_offset_diff)
5364 min_offset_diff = diff;
5365 }
5366
5367 if (mddev->reshape_position != MaxSector) {
5368 /* Check that we can continue the reshape.
5369 * Difficulties arise if the stripe we would write to
5370 * next is at or after the stripe we would read from next.
5371 * For a reshape that changes the number of devices, this
5372 * is only possible for a very short time, and mdadm makes
5373 * sure that time appears to have past before assembling
5374 * the array. So we fail if that time hasn't passed.
5375 * For a reshape that keeps the number of devices the same
5376 * mdadm must be monitoring the reshape can keeping the
5377 * critical areas read-only and backed up. It will start
5378 * the array in read-only mode, so we check for that.
5379 */
5380 sector_t here_new, here_old;
5381 int old_disks;
5382 int max_degraded = (mddev->level == 6 ? 2 : 1);
5383
5384 if (mddev->new_level != mddev->level) {
5385 printk(KERN_ERR "md/raid:%s: unsupported reshape "
5386 "required - aborting.\n",
5387 mdname(mddev));
5388 return -EINVAL;
5389 }
5390 old_disks = mddev->raid_disks - mddev->delta_disks;
5391 /* reshape_position must be on a new-stripe boundary, and one
5392 * further up in new geometry must map after here in old
5393 * geometry.
5394 */
5395 here_new = mddev->reshape_position;
5396 if (sector_div(here_new, mddev->new_chunk_sectors *
5397 (mddev->raid_disks - max_degraded))) {
5398 printk(KERN_ERR "md/raid:%s: reshape_position not "
5399 "on a stripe boundary\n", mdname(mddev));
5400 return -EINVAL;
5401 }
5402 reshape_offset = here_new * mddev->new_chunk_sectors;
5403 /* here_new is the stripe we will write to */
5404 here_old = mddev->reshape_position;
5405 sector_div(here_old, mddev->chunk_sectors *
5406 (old_disks-max_degraded));
5407 /* here_old is the first stripe that we might need to read
5408 * from */
5409 if (mddev->delta_disks == 0) {
5410 if ((here_new * mddev->new_chunk_sectors !=
5411 here_old * mddev->chunk_sectors)) {
5412 printk(KERN_ERR "md/raid:%s: reshape position is"
5413 " confused - aborting\n", mdname(mddev));
5414 return -EINVAL;
5415 }
5416 /* We cannot be sure it is safe to start an in-place
5417 * reshape. It is only safe if user-space is monitoring
5418 * and taking constant backups.
5419 * mdadm always starts a situation like this in
5420 * readonly mode so it can take control before
5421 * allowing any writes. So just check for that.
5422 */
5423 if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5424 abs(min_offset_diff) >= mddev->new_chunk_sectors)
5425 /* not really in-place - so OK */;
5426 else if (mddev->ro == 0) {
5427 printk(KERN_ERR "md/raid:%s: in-place reshape "
5428 "must be started in read-only mode "
5429 "- aborting\n",
5430 mdname(mddev));
5431 return -EINVAL;
5432 }
5433 } else if (mddev->reshape_backwards
5434 ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5435 here_old * mddev->chunk_sectors)
5436 : (here_new * mddev->new_chunk_sectors >=
5437 here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5438 /* Reading from the same stripe as writing to - bad */
5439 printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5440 "auto-recovery - aborting.\n",
5441 mdname(mddev));
5442 return -EINVAL;
5443 }
5444 printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5445 mdname(mddev));
5446 /* OK, we should be able to continue; */
5447 } else {
5448 BUG_ON(mddev->level != mddev->new_level);
5449 BUG_ON(mddev->layout != mddev->new_layout);
5450 BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5451 BUG_ON(mddev->delta_disks != 0);
5452 }
5453
5454 if (mddev->private == NULL)
5455 conf = setup_conf(mddev);
5456 else
5457 conf = mddev->private;
5458
5459 if (IS_ERR(conf))
5460 return PTR_ERR(conf);
5461
5462 conf->min_offset_diff = min_offset_diff;
5463 mddev->thread = conf->thread;
5464 conf->thread = NULL;
5465 mddev->private = conf;
5466
5467 for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5468 i++) {
5469 rdev = conf->disks[i].rdev;
5470 if (!rdev && conf->disks[i].replacement) {
5471 /* The replacement is all we have yet */
5472 rdev = conf->disks[i].replacement;
5473 conf->disks[i].replacement = NULL;
5474 clear_bit(Replacement, &rdev->flags);
5475 conf->disks[i].rdev = rdev;
5476 }
5477 if (!rdev)
5478 continue;
5479 if (conf->disks[i].replacement &&
5480 conf->reshape_progress != MaxSector) {
5481 /* replacements and reshape simply do not mix. */
5482 printk(KERN_ERR "md: cannot handle concurrent "
5483 "replacement and reshape.\n");
5484 goto abort;
5485 }
5486 if (test_bit(In_sync, &rdev->flags)) {
5487 working_disks++;
5488 continue;
5489 }
5490 /* This disc is not fully in-sync. However if it
5491 * just stored parity (beyond the recovery_offset),
5492 * when we don't need to be concerned about the
5493 * array being dirty.
5494 * When reshape goes 'backwards', we never have
5495 * partially completed devices, so we only need
5496 * to worry about reshape going forwards.
5497 */
5498 /* Hack because v0.91 doesn't store recovery_offset properly. */
5499 if (mddev->major_version == 0 &&
5500 mddev->minor_version > 90)
5501 rdev->recovery_offset = reshape_offset;
5502
5503 if (rdev->recovery_offset < reshape_offset) {
5504 /* We need to check old and new layout */
5505 if (!only_parity(rdev->raid_disk,
5506 conf->algorithm,
5507 conf->raid_disks,
5508 conf->max_degraded))
5509 continue;
5510 }
5511 if (!only_parity(rdev->raid_disk,
5512 conf->prev_algo,
5513 conf->previous_raid_disks,
5514 conf->max_degraded))
5515 continue;
5516 dirty_parity_disks++;
5517 }
5518
5519 /*
5520 * 0 for a fully functional array, 1 or 2 for a degraded array.
5521 */
5522 mddev->degraded = calc_degraded(conf);
5523
5524 if (has_failed(conf)) {
5525 printk(KERN_ERR "md/raid:%s: not enough operational devices"
5526 " (%d/%d failed)\n",
5527 mdname(mddev), mddev->degraded, conf->raid_disks);
5528 goto abort;
5529 }
5530
5531 /* device size must be a multiple of chunk size */
5532 mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5533 mddev->resync_max_sectors = mddev->dev_sectors;
5534
5535 if (mddev->degraded > dirty_parity_disks &&
5536 mddev->recovery_cp != MaxSector) {
5537 if (mddev->ok_start_degraded)
5538 printk(KERN_WARNING
5539 "md/raid:%s: starting dirty degraded array"
5540 " - data corruption possible.\n",
5541 mdname(mddev));
5542 else {
5543 printk(KERN_ERR
5544 "md/raid:%s: cannot start dirty degraded array.\n",
5545 mdname(mddev));
5546 goto abort;
5547 }
5548 }
5549
5550 if (mddev->degraded == 0)
5551 printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5552 " devices, algorithm %d\n", mdname(mddev), conf->level,
5553 mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5554 mddev->new_layout);
5555 else
5556 printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5557 " out of %d devices, algorithm %d\n",
5558 mdname(mddev), conf->level,
5559 mddev->raid_disks - mddev->degraded,
5560 mddev->raid_disks, mddev->new_layout);
5561
5562 print_raid5_conf(conf);
5563
5564 if (conf->reshape_progress != MaxSector) {
5565 conf->reshape_safe = conf->reshape_progress;
5566 atomic_set(&conf->reshape_stripes, 0);
5567 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5568 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5569 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5570 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5571 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5572 "reshape");
5573 }
5574
5575
5576 /* Ok, everything is just fine now */
5577 if (mddev->to_remove == &raid5_attrs_group)
5578 mddev->to_remove = NULL;
5579 else if (mddev->kobj.sd &&
5580 sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5581 printk(KERN_WARNING
5582 "raid5: failed to create sysfs attributes for %s\n",
5583 mdname(mddev));
5584 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5585
5586 if (mddev->queue) {
5587 int chunk_size;
5588 bool discard_supported = true;
5589 /* read-ahead size must cover two whole stripes, which
5590 * is 2 * (datadisks) * chunksize where 'n' is the
5591 * number of raid devices
5592 */
5593 int data_disks = conf->previous_raid_disks - conf->max_degraded;
5594 int stripe = data_disks *
5595 ((mddev->chunk_sectors << 9) / PAGE_SIZE);
5596 if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5597 mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5598
5599 blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5600
5601 mddev->queue->backing_dev_info.congested_data = mddev;
5602 mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5603
5604 chunk_size = mddev->chunk_sectors << 9;
5605 blk_queue_io_min(mddev->queue, chunk_size);
5606 blk_queue_io_opt(mddev->queue, chunk_size *
5607 (conf->raid_disks - conf->max_degraded));
5608 /*
5609 * We can only discard a whole stripe. It doesn't make sense to
5610 * discard data disk but write parity disk
5611 */
5612 stripe = stripe * PAGE_SIZE;
5613 /* Round up to power of 2, as discard handling
5614 * currently assumes that */
5615 while ((stripe-1) & stripe)
5616 stripe = (stripe | (stripe-1)) + 1;
5617 mddev->queue->limits.discard_alignment = stripe;
5618 mddev->queue->limits.discard_granularity = stripe;
5619
5620 /*
5621 * We use 16-bit counter of active stripes in bi_phys_segments
5622 * (minus one for over-loaded initialization)
5623 */
5624 blk_queue_max_hw_sectors(mddev->queue, 0xfffe * STRIPE_SECTORS);
5625 blk_queue_max_discard_sectors(mddev->queue,
5626 0xfffe * STRIPE_SECTORS);
5627
5628 /*
5629 * unaligned part of discard request will be ignored, so can't
5630 * guarantee discard_zeroes_data
5631 */
5632 mddev->queue->limits.discard_zeroes_data = 0;
5633
5634 blk_queue_max_write_same_sectors(mddev->queue, 0);
5635
5636 rdev_for_each(rdev, mddev) {
5637 disk_stack_limits(mddev->gendisk, rdev->bdev,
5638 rdev->data_offset << 9);
5639 disk_stack_limits(mddev->gendisk, rdev->bdev,
5640 rdev->new_data_offset << 9);
5641 /*
5642 * discard_zeroes_data is required, otherwise data
5643 * could be lost. Consider a scenario: discard a stripe
5644 * (the stripe could be inconsistent if
5645 * discard_zeroes_data is 0); write one disk of the
5646 * stripe (the stripe could be inconsistent again
5647 * depending on which disks are used to calculate
5648 * parity); the disk is broken; The stripe data of this
5649 * disk is lost.
5650 */
5651 if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5652 !bdev_get_queue(rdev->bdev)->
5653 limits.discard_zeroes_data)
5654 discard_supported = false;
5655 /* Unfortunately, discard_zeroes_data is not currently
5656 * a guarantee - just a hint. So we only allow DISCARD
5657 * if the sysadmin has confirmed that only safe devices
5658 * are in use by setting a module parameter.
5659 */
5660 if (!devices_handle_discard_safely) {
5661 if (discard_supported) {
5662 pr_info("md/raid456: discard support disabled due to uncertainty.\n");
5663 pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
5664 }
5665 discard_supported = false;
5666 }
5667 }
5668
5669 if (discard_supported &&
5670 mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
5671 mddev->queue->limits.discard_granularity >= stripe)
5672 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5673 mddev->queue);
5674 else
5675 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5676 mddev->queue);
5677 }
5678
5679 return 0;
5680 abort:
5681 md_unregister_thread(&mddev->thread);
5682 print_raid5_conf(conf);
5683 free_conf(conf);
5684 mddev->private = NULL;
5685 printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5686 return -EIO;
5687 }
5688
5689 static int stop(struct mddev *mddev)
5690 {
5691 struct r5conf *conf = mddev->private;
5692
5693 md_unregister_thread(&mddev->thread);
5694 if (mddev->queue)
5695 mddev->queue->backing_dev_info.congested_fn = NULL;
5696 free_conf(conf);
5697 mddev->private = NULL;
5698 mddev->to_remove = &raid5_attrs_group;
5699 return 0;
5700 }
5701
5702 static void status(struct seq_file *seq, struct mddev *mddev)
5703 {
5704 struct r5conf *conf = mddev->private;
5705 int i;
5706
5707 seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5708 mddev->chunk_sectors / 2, mddev->layout);
5709 seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5710 for (i = 0; i < conf->raid_disks; i++)
5711 seq_printf (seq, "%s",
5712 conf->disks[i].rdev &&
5713 test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5714 seq_printf (seq, "]");
5715 }
5716
5717 static void print_raid5_conf (struct r5conf *conf)
5718 {
5719 int i;
5720 struct disk_info *tmp;
5721
5722 printk(KERN_DEBUG "RAID conf printout:\n");
5723 if (!conf) {
5724 printk("(conf==NULL)\n");
5725 return;
5726 }
5727 printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5728 conf->raid_disks,
5729 conf->raid_disks - conf->mddev->degraded);
5730
5731 for (i = 0; i < conf->raid_disks; i++) {
5732 char b[BDEVNAME_SIZE];
5733 tmp = conf->disks + i;
5734 if (tmp->rdev)
5735 printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5736 i, !test_bit(Faulty, &tmp->rdev->flags),
5737 bdevname(tmp->rdev->bdev, b));
5738 }
5739 }
5740
5741 static int raid5_spare_active(struct mddev *mddev)
5742 {
5743 int i;
5744 struct r5conf *conf = mddev->private;
5745 struct disk_info *tmp;
5746 int count = 0;
5747 unsigned long flags;
5748
5749 for (i = 0; i < conf->raid_disks; i++) {
5750 tmp = conf->disks + i;
5751 if (tmp->replacement
5752 && tmp->replacement->recovery_offset == MaxSector
5753 && !test_bit(Faulty, &tmp->replacement->flags)
5754 && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5755 /* Replacement has just become active. */
5756 if (!tmp->rdev
5757 || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5758 count++;
5759 if (tmp->rdev) {
5760 /* Replaced device not technically faulty,
5761 * but we need to be sure it gets removed
5762 * and never re-added.
5763 */
5764 set_bit(Faulty, &tmp->rdev->flags);
5765 sysfs_notify_dirent_safe(
5766 tmp->rdev->sysfs_state);
5767 }
5768 sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5769 } else if (tmp->rdev
5770 && tmp->rdev->recovery_offset == MaxSector
5771 && !test_bit(Faulty, &tmp->rdev->flags)
5772 && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5773 count++;
5774 sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5775 }
5776 }
5777 spin_lock_irqsave(&conf->device_lock, flags);
5778 mddev->degraded = calc_degraded(conf);
5779 spin_unlock_irqrestore(&conf->device_lock, flags);
5780 print_raid5_conf(conf);
5781 return count;
5782 }
5783
5784 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5785 {
5786 struct r5conf *conf = mddev->private;
5787 int err = 0;
5788 int number = rdev->raid_disk;
5789 struct md_rdev **rdevp;
5790 struct disk_info *p = conf->disks + number;
5791
5792 print_raid5_conf(conf);
5793 if (rdev == p->rdev)
5794 rdevp = &p->rdev;
5795 else if (rdev == p->replacement)
5796 rdevp = &p->replacement;
5797 else
5798 return 0;
5799
5800 if (number >= conf->raid_disks &&
5801 conf->reshape_progress == MaxSector)
5802 clear_bit(In_sync, &rdev->flags);
5803
5804 if (test_bit(In_sync, &rdev->flags) ||
5805 atomic_read(&rdev->nr_pending)) {
5806 err = -EBUSY;
5807 goto abort;
5808 }
5809 /* Only remove non-faulty devices if recovery
5810 * isn't possible.
5811 */
5812 if (!test_bit(Faulty, &rdev->flags) &&
5813 mddev->recovery_disabled != conf->recovery_disabled &&
5814 !has_failed(conf) &&
5815 (!p->replacement || p->replacement == rdev) &&
5816 number < conf->raid_disks) {
5817 err = -EBUSY;
5818 goto abort;
5819 }
5820 *rdevp = NULL;
5821 synchronize_rcu();
5822 if (atomic_read(&rdev->nr_pending)) {
5823 /* lost the race, try later */
5824 err = -EBUSY;
5825 *rdevp = rdev;
5826 } else if (p->replacement) {
5827 /* We must have just cleared 'rdev' */
5828 p->rdev = p->replacement;
5829 clear_bit(Replacement, &p->replacement->flags);
5830 smp_mb(); /* Make sure other CPUs may see both as identical
5831 * but will never see neither - if they are careful
5832 */
5833 p->replacement = NULL;
5834 clear_bit(WantReplacement, &rdev->flags);
5835 } else
5836 /* We might have just removed the Replacement as faulty-
5837 * clear the bit just in case
5838 */
5839 clear_bit(WantReplacement, &rdev->flags);
5840 abort:
5841
5842 print_raid5_conf(conf);
5843 return err;
5844 }
5845
5846 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5847 {
5848 struct r5conf *conf = mddev->private;
5849 int err = -EEXIST;
5850 int disk;
5851 struct disk_info *p;
5852 int first = 0;
5853 int last = conf->raid_disks - 1;
5854
5855 if (mddev->recovery_disabled == conf->recovery_disabled)
5856 return -EBUSY;
5857
5858 if (rdev->saved_raid_disk < 0 && has_failed(conf))
5859 /* no point adding a device */
5860 return -EINVAL;
5861
5862 if (rdev->raid_disk >= 0)
5863 first = last = rdev->raid_disk;
5864
5865 /*
5866 * find the disk ... but prefer rdev->saved_raid_disk
5867 * if possible.
5868 */
5869 if (rdev->saved_raid_disk >= 0 &&
5870 rdev->saved_raid_disk >= first &&
5871 conf->disks[rdev->saved_raid_disk].rdev == NULL)
5872 first = rdev->saved_raid_disk;
5873
5874 for (disk = first; disk <= last; disk++) {
5875 p = conf->disks + disk;
5876 if (p->rdev == NULL) {
5877 clear_bit(In_sync, &rdev->flags);
5878 rdev->raid_disk = disk;
5879 err = 0;
5880 if (rdev->saved_raid_disk != disk)
5881 conf->fullsync = 1;
5882 rcu_assign_pointer(p->rdev, rdev);
5883 goto out;
5884 }
5885 }
5886 for (disk = first; disk <= last; disk++) {
5887 p = conf->disks + disk;
5888 if (test_bit(WantReplacement, &p->rdev->flags) &&
5889 p->replacement == NULL) {
5890 clear_bit(In_sync, &rdev->flags);
5891 set_bit(Replacement, &rdev->flags);
5892 rdev->raid_disk = disk;
5893 err = 0;
5894 conf->fullsync = 1;
5895 rcu_assign_pointer(p->replacement, rdev);
5896 break;
5897 }
5898 }
5899 out:
5900 print_raid5_conf(conf);
5901 return err;
5902 }
5903
5904 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5905 {
5906 /* no resync is happening, and there is enough space
5907 * on all devices, so we can resize.
5908 * We need to make sure resync covers any new space.
5909 * If the array is shrinking we should possibly wait until
5910 * any io in the removed space completes, but it hardly seems
5911 * worth it.
5912 */
5913 sector_t newsize;
5914 sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5915 newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5916 if (mddev->external_size &&
5917 mddev->array_sectors > newsize)
5918 return -EINVAL;
5919 if (mddev->bitmap) {
5920 int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5921 if (ret)
5922 return ret;
5923 }
5924 md_set_array_sectors(mddev, newsize);
5925 set_capacity(mddev->gendisk, mddev->array_sectors);
5926 revalidate_disk(mddev->gendisk);
5927 if (sectors > mddev->dev_sectors &&
5928 mddev->recovery_cp > mddev->dev_sectors) {
5929 mddev->recovery_cp = mddev->dev_sectors;
5930 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5931 }
5932 mddev->dev_sectors = sectors;
5933 mddev->resync_max_sectors = sectors;
5934 return 0;
5935 }
5936
5937 static int check_stripe_cache(struct mddev *mddev)
5938 {
5939 /* Can only proceed if there are plenty of stripe_heads.
5940 * We need a minimum of one full stripe,, and for sensible progress
5941 * it is best to have about 4 times that.
5942 * If we require 4 times, then the default 256 4K stripe_heads will
5943 * allow for chunk sizes up to 256K, which is probably OK.
5944 * If the chunk size is greater, user-space should request more
5945 * stripe_heads first.
5946 */
5947 struct r5conf *conf = mddev->private;
5948 if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5949 > conf->max_nr_stripes ||
5950 ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5951 > conf->max_nr_stripes) {
5952 printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes. Needed %lu\n",
5953 mdname(mddev),
5954 ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5955 / STRIPE_SIZE)*4);
5956 return 0;
5957 }
5958 return 1;
5959 }
5960
5961 static int check_reshape(struct mddev *mddev)
5962 {
5963 struct r5conf *conf = mddev->private;
5964
5965 if (mddev->delta_disks == 0 &&
5966 mddev->new_layout == mddev->layout &&
5967 mddev->new_chunk_sectors == mddev->chunk_sectors)
5968 return 0; /* nothing to do */
5969 if (has_failed(conf))
5970 return -EINVAL;
5971 if (mddev->delta_disks < 0) {
5972 /* We might be able to shrink, but the devices must
5973 * be made bigger first.
5974 * For raid6, 4 is the minimum size.
5975 * Otherwise 2 is the minimum
5976 */
5977 int min = 2;
5978 if (mddev->level == 6)
5979 min = 4;
5980 if (mddev->raid_disks + mddev->delta_disks < min)
5981 return -EINVAL;
5982 }
5983
5984 if (!check_stripe_cache(mddev))
5985 return -ENOSPC;
5986
5987 return resize_stripes(conf, (conf->previous_raid_disks
5988 + mddev->delta_disks));
5989 }
5990
5991 static int raid5_start_reshape(struct mddev *mddev)
5992 {
5993 struct r5conf *conf = mddev->private;
5994 struct md_rdev *rdev;
5995 int spares = 0;
5996 unsigned long flags;
5997
5998 if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5999 return -EBUSY;
6000
6001 if (!check_stripe_cache(mddev))
6002 return -ENOSPC;
6003
6004 if (has_failed(conf))
6005 return -EINVAL;
6006
6007 rdev_for_each(rdev, mddev) {
6008 if (!test_bit(In_sync, &rdev->flags)
6009 && !test_bit(Faulty, &rdev->flags))
6010 spares++;
6011 }
6012
6013 if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
6014 /* Not enough devices even to make a degraded array
6015 * of that size
6016 */
6017 return -EINVAL;
6018
6019 /* Refuse to reduce size of the array. Any reductions in
6020 * array size must be through explicit setting of array_size
6021 * attribute.
6022 */
6023 if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
6024 < mddev->array_sectors) {
6025 printk(KERN_ERR "md/raid:%s: array size must be reduced "
6026 "before number of disks\n", mdname(mddev));
6027 return -EINVAL;
6028 }
6029
6030 atomic_set(&conf->reshape_stripes, 0);
6031 spin_lock_irq(&conf->device_lock);
6032 conf->previous_raid_disks = conf->raid_disks;
6033 conf->raid_disks += mddev->delta_disks;
6034 conf->prev_chunk_sectors = conf->chunk_sectors;
6035 conf->chunk_sectors = mddev->new_chunk_sectors;
6036 conf->prev_algo = conf->algorithm;
6037 conf->algorithm = mddev->new_layout;
6038 conf->generation++;
6039 /* Code that selects data_offset needs to see the generation update
6040 * if reshape_progress has been set - so a memory barrier needed.
6041 */
6042 smp_mb();
6043 if (mddev->reshape_backwards)
6044 conf->reshape_progress = raid5_size(mddev, 0, 0);
6045 else
6046 conf->reshape_progress = 0;
6047 conf->reshape_safe = conf->reshape_progress;
6048 spin_unlock_irq(&conf->device_lock);
6049
6050 /* Add some new drives, as many as will fit.
6051 * We know there are enough to make the newly sized array work.
6052 * Don't add devices if we are reducing the number of
6053 * devices in the array. This is because it is not possible
6054 * to correctly record the "partially reconstructed" state of
6055 * such devices during the reshape and confusion could result.
6056 */
6057 if (mddev->delta_disks >= 0) {
6058 rdev_for_each(rdev, mddev)
6059 if (rdev->raid_disk < 0 &&
6060 !test_bit(Faulty, &rdev->flags)) {
6061 if (raid5_add_disk(mddev, rdev) == 0) {
6062 if (rdev->raid_disk
6063 >= conf->previous_raid_disks)
6064 set_bit(In_sync, &rdev->flags);
6065 else
6066 rdev->recovery_offset = 0;
6067
6068 if (sysfs_link_rdev(mddev, rdev))
6069 /* Failure here is OK */;
6070 }
6071 } else if (rdev->raid_disk >= conf->previous_raid_disks
6072 && !test_bit(Faulty, &rdev->flags)) {
6073 /* This is a spare that was manually added */
6074 set_bit(In_sync, &rdev->flags);
6075 }
6076
6077 /* When a reshape changes the number of devices,
6078 * ->degraded is measured against the larger of the
6079 * pre and post number of devices.
6080 */
6081 spin_lock_irqsave(&conf->device_lock, flags);
6082 mddev->degraded = calc_degraded(conf);
6083 spin_unlock_irqrestore(&conf->device_lock, flags);
6084 }
6085 mddev->raid_disks = conf->raid_disks;
6086 mddev->reshape_position = conf->reshape_progress;
6087 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6088
6089 clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6090 clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6091 set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6092 set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6093 mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6094 "reshape");
6095 if (!mddev->sync_thread) {
6096 mddev->recovery = 0;
6097 spin_lock_irq(&conf->device_lock);
6098 mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6099 rdev_for_each(rdev, mddev)
6100 rdev->new_data_offset = rdev->data_offset;
6101 smp_wmb();
6102 conf->reshape_progress = MaxSector;
6103 mddev->reshape_position = MaxSector;
6104 spin_unlock_irq(&conf->device_lock);
6105 return -EAGAIN;
6106 }
6107 conf->reshape_checkpoint = jiffies;
6108 md_wakeup_thread(mddev->sync_thread);
6109 md_new_event(mddev);
6110 return 0;
6111 }
6112
6113 /* This is called from the reshape thread and should make any
6114 * changes needed in 'conf'
6115 */
6116 static void end_reshape(struct r5conf *conf)
6117 {
6118
6119 if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6120 struct md_rdev *rdev;
6121
6122 spin_lock_irq(&conf->device_lock);
6123 conf->previous_raid_disks = conf->raid_disks;
6124 rdev_for_each(rdev, conf->mddev)
6125 rdev->data_offset = rdev->new_data_offset;
6126 smp_wmb();
6127 conf->reshape_progress = MaxSector;
6128 spin_unlock_irq(&conf->device_lock);
6129 wake_up(&conf->wait_for_overlap);
6130
6131 /* read-ahead size must cover two whole stripes, which is
6132 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6133 */
6134 if (conf->mddev->queue) {
6135 int data_disks = conf->raid_disks - conf->max_degraded;
6136 int stripe = data_disks * ((conf->chunk_sectors << 9)
6137 / PAGE_SIZE);
6138 if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6139 conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6140 }
6141 }
6142 }
6143
6144 /* This is called from the raid5d thread with mddev_lock held.
6145 * It makes config changes to the device.
6146 */
6147 static void raid5_finish_reshape(struct mddev *mddev)
6148 {
6149 struct r5conf *conf = mddev->private;
6150
6151 if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6152
6153 if (mddev->delta_disks > 0) {
6154 md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6155 set_capacity(mddev->gendisk, mddev->array_sectors);
6156 revalidate_disk(mddev->gendisk);
6157 } else {
6158 int d;
6159 spin_lock_irq(&conf->device_lock);
6160 mddev->degraded = calc_degraded(conf);
6161 spin_unlock_irq(&conf->device_lock);
6162 for (d = conf->raid_disks ;
6163 d < conf->raid_disks - mddev->delta_disks;
6164 d++) {
6165 struct md_rdev *rdev = conf->disks[d].rdev;
6166 if (rdev)
6167 clear_bit(In_sync, &rdev->flags);
6168 rdev = conf->disks[d].replacement;
6169 if (rdev)
6170 clear_bit(In_sync, &rdev->flags);
6171 }
6172 }
6173 mddev->layout = conf->algorithm;
6174 mddev->chunk_sectors = conf->chunk_sectors;
6175 mddev->reshape_position = MaxSector;
6176 mddev->delta_disks = 0;
6177 mddev->reshape_backwards = 0;
6178 }
6179 }
6180
6181 static void raid5_quiesce(struct mddev *mddev, int state)
6182 {
6183 struct r5conf *conf = mddev->private;
6184
6185 switch(state) {
6186 case 2: /* resume for a suspend */
6187 wake_up(&conf->wait_for_overlap);
6188 break;
6189
6190 case 1: /* stop all writes */
6191 spin_lock_irq(&conf->device_lock);
6192 /* '2' tells resync/reshape to pause so that all
6193 * active stripes can drain
6194 */
6195 conf->quiesce = 2;
6196 wait_event_lock_irq(conf->wait_for_stripe,
6197 atomic_read(&conf->active_stripes) == 0 &&
6198 atomic_read(&conf->active_aligned_reads) == 0,
6199 conf->device_lock);
6200 conf->quiesce = 1;
6201 spin_unlock_irq(&conf->device_lock);
6202 /* allow reshape to continue */
6203 wake_up(&conf->wait_for_overlap);
6204 break;
6205
6206 case 0: /* re-enable writes */
6207 spin_lock_irq(&conf->device_lock);
6208 conf->quiesce = 0;
6209 wake_up(&conf->wait_for_stripe);
6210 wake_up(&conf->wait_for_overlap);
6211 spin_unlock_irq(&conf->device_lock);
6212 break;
6213 }
6214 }
6215
6216
6217 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6218 {
6219 struct r0conf *raid0_conf = mddev->private;
6220 sector_t sectors;
6221
6222 /* for raid0 takeover only one zone is supported */
6223 if (raid0_conf->nr_strip_zones > 1) {
6224 printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6225 mdname(mddev));
6226 return ERR_PTR(-EINVAL);
6227 }
6228
6229 sectors = raid0_conf->strip_zone[0].zone_end;
6230 sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6231 mddev->dev_sectors = sectors;
6232 mddev->new_level = level;
6233 mddev->new_layout = ALGORITHM_PARITY_N;
6234 mddev->new_chunk_sectors = mddev->chunk_sectors;
6235 mddev->raid_disks += 1;
6236 mddev->delta_disks = 1;
6237 /* make sure it will be not marked as dirty */
6238 mddev->recovery_cp = MaxSector;
6239
6240 return setup_conf(mddev);
6241 }
6242
6243
6244 static void *raid5_takeover_raid1(struct mddev *mddev)
6245 {
6246 int chunksect;
6247
6248 if (mddev->raid_disks != 2 ||
6249 mddev->degraded > 1)
6250 return ERR_PTR(-EINVAL);
6251
6252 /* Should check if there are write-behind devices? */
6253
6254 chunksect = 64*2; /* 64K by default */
6255
6256 /* The array must be an exact multiple of chunksize */
6257 while (chunksect && (mddev->array_sectors & (chunksect-1)))
6258 chunksect >>= 1;
6259
6260 if ((chunksect<<9) < STRIPE_SIZE)
6261 /* array size does not allow a suitable chunk size */
6262 return ERR_PTR(-EINVAL);
6263
6264 mddev->new_level = 5;
6265 mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6266 mddev->new_chunk_sectors = chunksect;
6267
6268 return setup_conf(mddev);
6269 }
6270
6271 static void *raid5_takeover_raid6(struct mddev *mddev)
6272 {
6273 int new_layout;
6274
6275 switch (mddev->layout) {
6276 case ALGORITHM_LEFT_ASYMMETRIC_6:
6277 new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6278 break;
6279 case ALGORITHM_RIGHT_ASYMMETRIC_6:
6280 new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6281 break;
6282 case ALGORITHM_LEFT_SYMMETRIC_6:
6283 new_layout = ALGORITHM_LEFT_SYMMETRIC;
6284 break;
6285 case ALGORITHM_RIGHT_SYMMETRIC_6:
6286 new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6287 break;
6288 case ALGORITHM_PARITY_0_6:
6289 new_layout = ALGORITHM_PARITY_0;
6290 break;
6291 case ALGORITHM_PARITY_N:
6292 new_layout = ALGORITHM_PARITY_N;
6293 break;
6294 default:
6295 return ERR_PTR(-EINVAL);
6296 }
6297 mddev->new_level = 5;
6298 mddev->new_layout = new_layout;
6299 mddev->delta_disks = -1;
6300 mddev->raid_disks -= 1;
6301 return setup_conf(mddev);
6302 }
6303
6304
6305 static int raid5_check_reshape(struct mddev *mddev)
6306 {
6307 /* For a 2-drive array, the layout and chunk size can be changed
6308 * immediately as not restriping is needed.
6309 * For larger arrays we record the new value - after validation
6310 * to be used by a reshape pass.
6311 */
6312 struct r5conf *conf = mddev->private;
6313 int new_chunk = mddev->new_chunk_sectors;
6314
6315 if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6316 return -EINVAL;
6317 if (new_chunk > 0) {
6318 if (!is_power_of_2(new_chunk))
6319 return -EINVAL;
6320 if (new_chunk < (PAGE_SIZE>>9))
6321 return -EINVAL;
6322 if (mddev->array_sectors & (new_chunk-1))
6323 /* not factor of array size */
6324 return -EINVAL;
6325 }
6326
6327 /* They look valid */
6328
6329 if (mddev->raid_disks == 2) {
6330 /* can make the change immediately */
6331 if (mddev->new_layout >= 0) {
6332 conf->algorithm = mddev->new_layout;
6333 mddev->layout = mddev->new_layout;
6334 }
6335 if (new_chunk > 0) {
6336 conf->chunk_sectors = new_chunk ;
6337 mddev->chunk_sectors = new_chunk;
6338 }
6339 set_bit(MD_CHANGE_DEVS, &mddev->flags);
6340 md_wakeup_thread(mddev->thread);
6341 }
6342 return check_reshape(mddev);
6343 }
6344
6345 static int raid6_check_reshape(struct mddev *mddev)
6346 {
6347 int new_chunk = mddev->new_chunk_sectors;
6348
6349 if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6350 return -EINVAL;
6351 if (new_chunk > 0) {
6352 if (!is_power_of_2(new_chunk))
6353 return -EINVAL;
6354 if (new_chunk < (PAGE_SIZE >> 9))
6355 return -EINVAL;
6356 if (mddev->array_sectors & (new_chunk-1))
6357 /* not factor of array size */
6358 return -EINVAL;
6359 }
6360
6361 /* They look valid */
6362 return check_reshape(mddev);
6363 }
6364
6365 static void *raid5_takeover(struct mddev *mddev)
6366 {
6367 /* raid5 can take over:
6368 * raid0 - if there is only one strip zone - make it a raid4 layout
6369 * raid1 - if there are two drives. We need to know the chunk size
6370 * raid4 - trivial - just use a raid4 layout.
6371 * raid6 - Providing it is a *_6 layout
6372 */
6373 if (mddev->level == 0)
6374 return raid45_takeover_raid0(mddev, 5);
6375 if (mddev->level == 1)
6376 return raid5_takeover_raid1(mddev);
6377 if (mddev->level == 4) {
6378 mddev->new_layout = ALGORITHM_PARITY_N;
6379 mddev->new_level = 5;
6380 return setup_conf(mddev);
6381 }
6382 if (mddev->level == 6)
6383 return raid5_takeover_raid6(mddev);
6384
6385 return ERR_PTR(-EINVAL);
6386 }
6387
6388 static void *raid4_takeover(struct mddev *mddev)
6389 {
6390 /* raid4 can take over:
6391 * raid0 - if there is only one strip zone
6392 * raid5 - if layout is right
6393 */
6394 if (mddev->level == 0)
6395 return raid45_takeover_raid0(mddev, 4);
6396 if (mddev->level == 5 &&
6397 mddev->layout == ALGORITHM_PARITY_N) {
6398 mddev->new_layout = 0;
6399 mddev->new_level = 4;
6400 return setup_conf(mddev);
6401 }
6402 return ERR_PTR(-EINVAL);
6403 }
6404
6405 static struct md_personality raid5_personality;
6406
6407 static void *raid6_takeover(struct mddev *mddev)
6408 {
6409 /* Currently can only take over a raid5. We map the
6410 * personality to an equivalent raid6 personality
6411 * with the Q block at the end.
6412 */
6413 int new_layout;
6414
6415 if (mddev->pers != &raid5_personality)
6416 return ERR_PTR(-EINVAL);
6417 if (mddev->degraded > 1)
6418 return ERR_PTR(-EINVAL);
6419 if (mddev->raid_disks > 253)
6420 return ERR_PTR(-EINVAL);
6421 if (mddev->raid_disks < 3)
6422 return ERR_PTR(-EINVAL);
6423
6424 switch (mddev->layout) {
6425 case ALGORITHM_LEFT_ASYMMETRIC:
6426 new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6427 break;
6428 case ALGORITHM_RIGHT_ASYMMETRIC:
6429 new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6430 break;
6431 case ALGORITHM_LEFT_SYMMETRIC:
6432 new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6433 break;
6434 case ALGORITHM_RIGHT_SYMMETRIC:
6435 new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6436 break;
6437 case ALGORITHM_PARITY_0:
6438 new_layout = ALGORITHM_PARITY_0_6;
6439 break;
6440 case ALGORITHM_PARITY_N:
6441 new_layout = ALGORITHM_PARITY_N;
6442 break;
6443 default:
6444 return ERR_PTR(-EINVAL);
6445 }
6446 mddev->new_level = 6;
6447 mddev->new_layout = new_layout;
6448 mddev->delta_disks = 1;
6449 mddev->raid_disks += 1;
6450 return setup_conf(mddev);
6451 }
6452
6453
6454 static struct md_personality raid6_personality =
6455 {
6456 .name = "raid6",
6457 .level = 6,
6458 .owner = THIS_MODULE,
6459 .make_request = make_request,
6460 .run = run,
6461 .stop = stop,
6462 .status = status,
6463 .error_handler = error,
6464 .hot_add_disk = raid5_add_disk,
6465 .hot_remove_disk= raid5_remove_disk,
6466 .spare_active = raid5_spare_active,
6467 .sync_request = sync_request,
6468 .resize = raid5_resize,
6469 .size = raid5_size,
6470 .check_reshape = raid6_check_reshape,
6471 .start_reshape = raid5_start_reshape,
6472 .finish_reshape = raid5_finish_reshape,
6473 .quiesce = raid5_quiesce,
6474 .takeover = raid6_takeover,
6475 };
6476 static struct md_personality raid5_personality =
6477 {
6478 .name = "raid5",
6479 .level = 5,
6480 .owner = THIS_MODULE,
6481 .make_request = make_request,
6482 .run = run,
6483 .stop = stop,
6484 .status = status,
6485 .error_handler = error,
6486 .hot_add_disk = raid5_add_disk,
6487 .hot_remove_disk= raid5_remove_disk,
6488 .spare_active = raid5_spare_active,
6489 .sync_request = sync_request,
6490 .resize = raid5_resize,
6491 .size = raid5_size,
6492 .check_reshape = raid5_check_reshape,
6493 .start_reshape = raid5_start_reshape,
6494 .finish_reshape = raid5_finish_reshape,
6495 .quiesce = raid5_quiesce,
6496 .takeover = raid5_takeover,
6497 };
6498
6499 static struct md_personality raid4_personality =
6500 {
6501 .name = "raid4",
6502 .level = 4,
6503 .owner = THIS_MODULE,
6504 .make_request = make_request,
6505 .run = run,
6506 .stop = stop,
6507 .status = status,
6508 .error_handler = error,
6509 .hot_add_disk = raid5_add_disk,
6510 .hot_remove_disk= raid5_remove_disk,
6511 .spare_active = raid5_spare_active,
6512 .sync_request = sync_request,
6513 .resize = raid5_resize,
6514 .size = raid5_size,
6515 .check_reshape = raid5_check_reshape,
6516 .start_reshape = raid5_start_reshape,
6517 .finish_reshape = raid5_finish_reshape,
6518 .quiesce = raid5_quiesce,
6519 .takeover = raid4_takeover,
6520 };
6521
6522 static int __init raid5_init(void)
6523 {
6524 register_md_personality(&raid6_personality);
6525 register_md_personality(&raid5_personality);
6526 register_md_personality(&raid4_personality);
6527 return 0;
6528 }
6529
6530 static void raid5_exit(void)
6531 {
6532 unregister_md_personality(&raid6_personality);
6533 unregister_md_personality(&raid5_personality);
6534 unregister_md_personality(&raid4_personality);
6535 }
6536
6537 module_init(raid5_init);
6538 module_exit(raid5_exit);
6539 MODULE_LICENSE("GPL");
6540 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6541 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6542 MODULE_ALIAS("md-raid5");
6543 MODULE_ALIAS("md-raid4");
6544 MODULE_ALIAS("md-level-5");
6545 MODULE_ALIAS("md-level-4");
6546 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6547 MODULE_ALIAS("md-raid6");
6548 MODULE_ALIAS("md-level-6");
6549
6550 /* This used to be two separate modules, they were: */
6551 MODULE_ALIAS("raid5");
6552 MODULE_ALIAS("raid6");
kernel source for telekom puls (alcatel/mediatek)