xref: /linux/drivers/md/raid5.c (revision d97b46a64674a267bc41c9e16132ee2a98c3347d)
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *	   Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *	   Copyright (C) 1999, 2000 Ingo Molnar
5  *	   Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20 
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->seq_write is the number of the last batch successfully written.
31  * conf->seq_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is seq_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45 
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include "md.h"
57 #include "raid5.h"
58 #include "raid0.h"
59 #include "bitmap.h"
60 
61 /*
62  * Stripe cache
63  */
64 
65 #define NR_STRIPES		256
66 #define STRIPE_SIZE		PAGE_SIZE
67 #define STRIPE_SHIFT		(PAGE_SHIFT - 9)
68 #define STRIPE_SECTORS		(STRIPE_SIZE>>9)
69 #define	IO_THRESHOLD		1
70 #define BYPASS_THRESHOLD	1
71 #define NR_HASH			(PAGE_SIZE / sizeof(struct hlist_head))
72 #define HASH_MASK		(NR_HASH - 1)
73 
74 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
75 {
76 	int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
77 	return &conf->stripe_hashtbl[hash];
78 }
79 
80 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
81  * order without overlap.  There may be several bio's per stripe+device, and
82  * a bio could span several devices.
83  * When walking this list for a particular stripe+device, we must never proceed
84  * beyond a bio that extends past this device, as the next bio might no longer
85  * be valid.
86  * This function is used to determine the 'next' bio in the list, given the sector
87  * of the current stripe+device
88  */
89 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
90 {
91 	int sectors = bio->bi_size >> 9;
92 	if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
93 		return bio->bi_next;
94 	else
95 		return NULL;
96 }
97 
98 /*
99  * We maintain a biased count of active stripes in the bottom 16 bits of
100  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
101  */
102 static inline int raid5_bi_phys_segments(struct bio *bio)
103 {
104 	return bio->bi_phys_segments & 0xffff;
105 }
106 
107 static inline int raid5_bi_hw_segments(struct bio *bio)
108 {
109 	return (bio->bi_phys_segments >> 16) & 0xffff;
110 }
111 
112 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
113 {
114 	--bio->bi_phys_segments;
115 	return raid5_bi_phys_segments(bio);
116 }
117 
118 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
119 {
120 	unsigned short val = raid5_bi_hw_segments(bio);
121 
122 	--val;
123 	bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
124 	return val;
125 }
126 
127 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
128 {
129 	bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
130 }
131 
132 /* Find first data disk in a raid6 stripe */
133 static inline int raid6_d0(struct stripe_head *sh)
134 {
135 	if (sh->ddf_layout)
136 		/* ddf always start from first device */
137 		return 0;
138 	/* md starts just after Q block */
139 	if (sh->qd_idx == sh->disks - 1)
140 		return 0;
141 	else
142 		return sh->qd_idx + 1;
143 }
144 static inline int raid6_next_disk(int disk, int raid_disks)
145 {
146 	disk++;
147 	return (disk < raid_disks) ? disk : 0;
148 }
149 
150 /* When walking through the disks in a raid5, starting at raid6_d0,
151  * We need to map each disk to a 'slot', where the data disks are slot
152  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
153  * is raid_disks-1.  This help does that mapping.
154  */
155 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
156 			     int *count, int syndrome_disks)
157 {
158 	int slot = *count;
159 
160 	if (sh->ddf_layout)
161 		(*count)++;
162 	if (idx == sh->pd_idx)
163 		return syndrome_disks;
164 	if (idx == sh->qd_idx)
165 		return syndrome_disks + 1;
166 	if (!sh->ddf_layout)
167 		(*count)++;
168 	return slot;
169 }
170 
171 static void return_io(struct bio *return_bi)
172 {
173 	struct bio *bi = return_bi;
174 	while (bi) {
175 
176 		return_bi = bi->bi_next;
177 		bi->bi_next = NULL;
178 		bi->bi_size = 0;
179 		bio_endio(bi, 0);
180 		bi = return_bi;
181 	}
182 }
183 
184 static void print_raid5_conf (struct r5conf *conf);
185 
186 static int stripe_operations_active(struct stripe_head *sh)
187 {
188 	return sh->check_state || sh->reconstruct_state ||
189 	       test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
190 	       test_bit(STRIPE_COMPUTE_RUN, &sh->state);
191 }
192 
193 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
194 {
195 	if (atomic_dec_and_test(&sh->count)) {
196 		BUG_ON(!list_empty(&sh->lru));
197 		BUG_ON(atomic_read(&conf->active_stripes)==0);
198 		if (test_bit(STRIPE_HANDLE, &sh->state)) {
199 			if (test_bit(STRIPE_DELAYED, &sh->state))
200 				list_add_tail(&sh->lru, &conf->delayed_list);
201 			else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
202 				   sh->bm_seq - conf->seq_write > 0)
203 				list_add_tail(&sh->lru, &conf->bitmap_list);
204 			else {
205 				clear_bit(STRIPE_BIT_DELAY, &sh->state);
206 				list_add_tail(&sh->lru, &conf->handle_list);
207 			}
208 			md_wakeup_thread(conf->mddev->thread);
209 		} else {
210 			BUG_ON(stripe_operations_active(sh));
211 			if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
212 				if (atomic_dec_return(&conf->preread_active_stripes)
213 				    < IO_THRESHOLD)
214 					md_wakeup_thread(conf->mddev->thread);
215 			atomic_dec(&conf->active_stripes);
216 			if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
217 				list_add_tail(&sh->lru, &conf->inactive_list);
218 				wake_up(&conf->wait_for_stripe);
219 				if (conf->retry_read_aligned)
220 					md_wakeup_thread(conf->mddev->thread);
221 			}
222 		}
223 	}
224 }
225 
226 static void release_stripe(struct stripe_head *sh)
227 {
228 	struct r5conf *conf = sh->raid_conf;
229 	unsigned long flags;
230 
231 	spin_lock_irqsave(&conf->device_lock, flags);
232 	__release_stripe(conf, sh);
233 	spin_unlock_irqrestore(&conf->device_lock, flags);
234 }
235 
236 static inline void remove_hash(struct stripe_head *sh)
237 {
238 	pr_debug("remove_hash(), stripe %llu\n",
239 		(unsigned long long)sh->sector);
240 
241 	hlist_del_init(&sh->hash);
242 }
243 
244 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
245 {
246 	struct hlist_head *hp = stripe_hash(conf, sh->sector);
247 
248 	pr_debug("insert_hash(), stripe %llu\n",
249 		(unsigned long long)sh->sector);
250 
251 	hlist_add_head(&sh->hash, hp);
252 }
253 
254 
255 /* find an idle stripe, make sure it is unhashed, and return it. */
256 static struct stripe_head *get_free_stripe(struct r5conf *conf)
257 {
258 	struct stripe_head *sh = NULL;
259 	struct list_head *first;
260 
261 	if (list_empty(&conf->inactive_list))
262 		goto out;
263 	first = conf->inactive_list.next;
264 	sh = list_entry(first, struct stripe_head, lru);
265 	list_del_init(first);
266 	remove_hash(sh);
267 	atomic_inc(&conf->active_stripes);
268 out:
269 	return sh;
270 }
271 
272 static void shrink_buffers(struct stripe_head *sh)
273 {
274 	struct page *p;
275 	int i;
276 	int num = sh->raid_conf->pool_size;
277 
278 	for (i = 0; i < num ; i++) {
279 		p = sh->dev[i].page;
280 		if (!p)
281 			continue;
282 		sh->dev[i].page = NULL;
283 		put_page(p);
284 	}
285 }
286 
287 static int grow_buffers(struct stripe_head *sh)
288 {
289 	int i;
290 	int num = sh->raid_conf->pool_size;
291 
292 	for (i = 0; i < num; i++) {
293 		struct page *page;
294 
295 		if (!(page = alloc_page(GFP_KERNEL))) {
296 			return 1;
297 		}
298 		sh->dev[i].page = page;
299 	}
300 	return 0;
301 }
302 
303 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
304 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
305 			    struct stripe_head *sh);
306 
307 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
308 {
309 	struct r5conf *conf = sh->raid_conf;
310 	int i;
311 
312 	BUG_ON(atomic_read(&sh->count) != 0);
313 	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
314 	BUG_ON(stripe_operations_active(sh));
315 
316 	pr_debug("init_stripe called, stripe %llu\n",
317 		(unsigned long long)sh->sector);
318 
319 	remove_hash(sh);
320 
321 	sh->generation = conf->generation - previous;
322 	sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
323 	sh->sector = sector;
324 	stripe_set_idx(sector, conf, previous, sh);
325 	sh->state = 0;
326 
327 
328 	for (i = sh->disks; i--; ) {
329 		struct r5dev *dev = &sh->dev[i];
330 
331 		if (dev->toread || dev->read || dev->towrite || dev->written ||
332 		    test_bit(R5_LOCKED, &dev->flags)) {
333 			printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
334 			       (unsigned long long)sh->sector, i, dev->toread,
335 			       dev->read, dev->towrite, dev->written,
336 			       test_bit(R5_LOCKED, &dev->flags));
337 			WARN_ON(1);
338 		}
339 		dev->flags = 0;
340 		raid5_build_block(sh, i, previous);
341 	}
342 	insert_hash(conf, sh);
343 }
344 
345 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
346 					 short generation)
347 {
348 	struct stripe_head *sh;
349 	struct hlist_node *hn;
350 
351 	pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
352 	hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
353 		if (sh->sector == sector && sh->generation == generation)
354 			return sh;
355 	pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
356 	return NULL;
357 }
358 
359 /*
360  * Need to check if array has failed when deciding whether to:
361  *  - start an array
362  *  - remove non-faulty devices
363  *  - add a spare
364  *  - allow a reshape
365  * This determination is simple when no reshape is happening.
366  * However if there is a reshape, we need to carefully check
367  * both the before and after sections.
368  * This is because some failed devices may only affect one
369  * of the two sections, and some non-in_sync devices may
370  * be insync in the section most affected by failed devices.
371  */
372 static int calc_degraded(struct r5conf *conf)
373 {
374 	int degraded, degraded2;
375 	int i;
376 
377 	rcu_read_lock();
378 	degraded = 0;
379 	for (i = 0; i < conf->previous_raid_disks; i++) {
380 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
381 		if (!rdev || test_bit(Faulty, &rdev->flags))
382 			degraded++;
383 		else if (test_bit(In_sync, &rdev->flags))
384 			;
385 		else
386 			/* not in-sync or faulty.
387 			 * If the reshape increases the number of devices,
388 			 * this is being recovered by the reshape, so
389 			 * this 'previous' section is not in_sync.
390 			 * If the number of devices is being reduced however,
391 			 * the device can only be part of the array if
392 			 * we are reverting a reshape, so this section will
393 			 * be in-sync.
394 			 */
395 			if (conf->raid_disks >= conf->previous_raid_disks)
396 				degraded++;
397 	}
398 	rcu_read_unlock();
399 	if (conf->raid_disks == conf->previous_raid_disks)
400 		return degraded;
401 	rcu_read_lock();
402 	degraded2 = 0;
403 	for (i = 0; i < conf->raid_disks; i++) {
404 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
405 		if (!rdev || test_bit(Faulty, &rdev->flags))
406 			degraded2++;
407 		else if (test_bit(In_sync, &rdev->flags))
408 			;
409 		else
410 			/* not in-sync or faulty.
411 			 * If reshape increases the number of devices, this
412 			 * section has already been recovered, else it
413 			 * almost certainly hasn't.
414 			 */
415 			if (conf->raid_disks <= conf->previous_raid_disks)
416 				degraded2++;
417 	}
418 	rcu_read_unlock();
419 	if (degraded2 > degraded)
420 		return degraded2;
421 	return degraded;
422 }
423 
424 static int has_failed(struct r5conf *conf)
425 {
426 	int degraded;
427 
428 	if (conf->mddev->reshape_position == MaxSector)
429 		return conf->mddev->degraded > conf->max_degraded;
430 
431 	degraded = calc_degraded(conf);
432 	if (degraded > conf->max_degraded)
433 		return 1;
434 	return 0;
435 }
436 
437 static struct stripe_head *
438 get_active_stripe(struct r5conf *conf, sector_t sector,
439 		  int previous, int noblock, int noquiesce)
440 {
441 	struct stripe_head *sh;
442 
443 	pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
444 
445 	spin_lock_irq(&conf->device_lock);
446 
447 	do {
448 		wait_event_lock_irq(conf->wait_for_stripe,
449 				    conf->quiesce == 0 || noquiesce,
450 				    conf->device_lock, /* nothing */);
451 		sh = __find_stripe(conf, sector, conf->generation - previous);
452 		if (!sh) {
453 			if (!conf->inactive_blocked)
454 				sh = get_free_stripe(conf);
455 			if (noblock && sh == NULL)
456 				break;
457 			if (!sh) {
458 				conf->inactive_blocked = 1;
459 				wait_event_lock_irq(conf->wait_for_stripe,
460 						    !list_empty(&conf->inactive_list) &&
461 						    (atomic_read(&conf->active_stripes)
462 						     < (conf->max_nr_stripes *3/4)
463 						     || !conf->inactive_blocked),
464 						    conf->device_lock,
465 						    );
466 				conf->inactive_blocked = 0;
467 			} else
468 				init_stripe(sh, sector, previous);
469 		} else {
470 			if (atomic_read(&sh->count)) {
471 				BUG_ON(!list_empty(&sh->lru)
472 				    && !test_bit(STRIPE_EXPANDING, &sh->state));
473 			} else {
474 				if (!test_bit(STRIPE_HANDLE, &sh->state))
475 					atomic_inc(&conf->active_stripes);
476 				if (list_empty(&sh->lru) &&
477 				    !test_bit(STRIPE_EXPANDING, &sh->state))
478 					BUG();
479 				list_del_init(&sh->lru);
480 			}
481 		}
482 	} while (sh == NULL);
483 
484 	if (sh)
485 		atomic_inc(&sh->count);
486 
487 	spin_unlock_irq(&conf->device_lock);
488 	return sh;
489 }
490 
491 /* Determine if 'data_offset' or 'new_data_offset' should be used
492  * in this stripe_head.
493  */
494 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
495 {
496 	sector_t progress = conf->reshape_progress;
497 	/* Need a memory barrier to make sure we see the value
498 	 * of conf->generation, or ->data_offset that was set before
499 	 * reshape_progress was updated.
500 	 */
501 	smp_rmb();
502 	if (progress == MaxSector)
503 		return 0;
504 	if (sh->generation == conf->generation - 1)
505 		return 0;
506 	/* We are in a reshape, and this is a new-generation stripe,
507 	 * so use new_data_offset.
508 	 */
509 	return 1;
510 }
511 
512 static void
513 raid5_end_read_request(struct bio *bi, int error);
514 static void
515 raid5_end_write_request(struct bio *bi, int error);
516 
517 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
518 {
519 	struct r5conf *conf = sh->raid_conf;
520 	int i, disks = sh->disks;
521 
522 	might_sleep();
523 
524 	for (i = disks; i--; ) {
525 		int rw;
526 		int replace_only = 0;
527 		struct bio *bi, *rbi;
528 		struct md_rdev *rdev, *rrdev = NULL;
529 		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
530 			if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
531 				rw = WRITE_FUA;
532 			else
533 				rw = WRITE;
534 		} else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
535 			rw = READ;
536 		else if (test_and_clear_bit(R5_WantReplace,
537 					    &sh->dev[i].flags)) {
538 			rw = WRITE;
539 			replace_only = 1;
540 		} else
541 			continue;
542 		if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
543 			rw |= REQ_SYNC;
544 
545 		bi = &sh->dev[i].req;
546 		rbi = &sh->dev[i].rreq; /* For writing to replacement */
547 
548 		bi->bi_rw = rw;
549 		rbi->bi_rw = rw;
550 		if (rw & WRITE) {
551 			bi->bi_end_io = raid5_end_write_request;
552 			rbi->bi_end_io = raid5_end_write_request;
553 		} else
554 			bi->bi_end_io = raid5_end_read_request;
555 
556 		rcu_read_lock();
557 		rrdev = rcu_dereference(conf->disks[i].replacement);
558 		smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
559 		rdev = rcu_dereference(conf->disks[i].rdev);
560 		if (!rdev) {
561 			rdev = rrdev;
562 			rrdev = NULL;
563 		}
564 		if (rw & WRITE) {
565 			if (replace_only)
566 				rdev = NULL;
567 			if (rdev == rrdev)
568 				/* We raced and saw duplicates */
569 				rrdev = NULL;
570 		} else {
571 			if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
572 				rdev = rrdev;
573 			rrdev = NULL;
574 		}
575 
576 		if (rdev && test_bit(Faulty, &rdev->flags))
577 			rdev = NULL;
578 		if (rdev)
579 			atomic_inc(&rdev->nr_pending);
580 		if (rrdev && test_bit(Faulty, &rrdev->flags))
581 			rrdev = NULL;
582 		if (rrdev)
583 			atomic_inc(&rrdev->nr_pending);
584 		rcu_read_unlock();
585 
586 		/* We have already checked bad blocks for reads.  Now
587 		 * need to check for writes.  We never accept write errors
588 		 * on the replacement, so we don't to check rrdev.
589 		 */
590 		while ((rw & WRITE) && rdev &&
591 		       test_bit(WriteErrorSeen, &rdev->flags)) {
592 			sector_t first_bad;
593 			int bad_sectors;
594 			int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
595 					      &first_bad, &bad_sectors);
596 			if (!bad)
597 				break;
598 
599 			if (bad < 0) {
600 				set_bit(BlockedBadBlocks, &rdev->flags);
601 				if (!conf->mddev->external &&
602 				    conf->mddev->flags) {
603 					/* It is very unlikely, but we might
604 					 * still need to write out the
605 					 * bad block log - better give it
606 					 * a chance*/
607 					md_check_recovery(conf->mddev);
608 				}
609 				md_wait_for_blocked_rdev(rdev, conf->mddev);
610 			} else {
611 				/* Acknowledged bad block - skip the write */
612 				rdev_dec_pending(rdev, conf->mddev);
613 				rdev = NULL;
614 			}
615 		}
616 
617 		if (rdev) {
618 			if (s->syncing || s->expanding || s->expanded
619 			    || s->replacing)
620 				md_sync_acct(rdev->bdev, STRIPE_SECTORS);
621 
622 			set_bit(STRIPE_IO_STARTED, &sh->state);
623 
624 			bi->bi_bdev = rdev->bdev;
625 			pr_debug("%s: for %llu schedule op %ld on disc %d\n",
626 				__func__, (unsigned long long)sh->sector,
627 				bi->bi_rw, i);
628 			atomic_inc(&sh->count);
629 			if (use_new_offset(conf, sh))
630 				bi->bi_sector = (sh->sector
631 						 + rdev->new_data_offset);
632 			else
633 				bi->bi_sector = (sh->sector
634 						 + rdev->data_offset);
635 			bi->bi_flags = 1 << BIO_UPTODATE;
636 			bi->bi_idx = 0;
637 			bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
638 			bi->bi_io_vec[0].bv_offset = 0;
639 			bi->bi_size = STRIPE_SIZE;
640 			bi->bi_next = NULL;
641 			if (rrdev)
642 				set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
643 			generic_make_request(bi);
644 		}
645 		if (rrdev) {
646 			if (s->syncing || s->expanding || s->expanded
647 			    || s->replacing)
648 				md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
649 
650 			set_bit(STRIPE_IO_STARTED, &sh->state);
651 
652 			rbi->bi_bdev = rrdev->bdev;
653 			pr_debug("%s: for %llu schedule op %ld on "
654 				 "replacement disc %d\n",
655 				__func__, (unsigned long long)sh->sector,
656 				rbi->bi_rw, i);
657 			atomic_inc(&sh->count);
658 			if (use_new_offset(conf, sh))
659 				rbi->bi_sector = (sh->sector
660 						  + rrdev->new_data_offset);
661 			else
662 				rbi->bi_sector = (sh->sector
663 						  + rrdev->data_offset);
664 			rbi->bi_flags = 1 << BIO_UPTODATE;
665 			rbi->bi_idx = 0;
666 			rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
667 			rbi->bi_io_vec[0].bv_offset = 0;
668 			rbi->bi_size = STRIPE_SIZE;
669 			rbi->bi_next = NULL;
670 			generic_make_request(rbi);
671 		}
672 		if (!rdev && !rrdev) {
673 			if (rw & WRITE)
674 				set_bit(STRIPE_DEGRADED, &sh->state);
675 			pr_debug("skip op %ld on disc %d for sector %llu\n",
676 				bi->bi_rw, i, (unsigned long long)sh->sector);
677 			clear_bit(R5_LOCKED, &sh->dev[i].flags);
678 			set_bit(STRIPE_HANDLE, &sh->state);
679 		}
680 	}
681 }
682 
683 static struct dma_async_tx_descriptor *
684 async_copy_data(int frombio, struct bio *bio, struct page *page,
685 	sector_t sector, struct dma_async_tx_descriptor *tx)
686 {
687 	struct bio_vec *bvl;
688 	struct page *bio_page;
689 	int i;
690 	int page_offset;
691 	struct async_submit_ctl submit;
692 	enum async_tx_flags flags = 0;
693 
694 	if (bio->bi_sector >= sector)
695 		page_offset = (signed)(bio->bi_sector - sector) * 512;
696 	else
697 		page_offset = (signed)(sector - bio->bi_sector) * -512;
698 
699 	if (frombio)
700 		flags |= ASYNC_TX_FENCE;
701 	init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
702 
703 	bio_for_each_segment(bvl, bio, i) {
704 		int len = bvl->bv_len;
705 		int clen;
706 		int b_offset = 0;
707 
708 		if (page_offset < 0) {
709 			b_offset = -page_offset;
710 			page_offset += b_offset;
711 			len -= b_offset;
712 		}
713 
714 		if (len > 0 && page_offset + len > STRIPE_SIZE)
715 			clen = STRIPE_SIZE - page_offset;
716 		else
717 			clen = len;
718 
719 		if (clen > 0) {
720 			b_offset += bvl->bv_offset;
721 			bio_page = bvl->bv_page;
722 			if (frombio)
723 				tx = async_memcpy(page, bio_page, page_offset,
724 						  b_offset, clen, &submit);
725 			else
726 				tx = async_memcpy(bio_page, page, b_offset,
727 						  page_offset, clen, &submit);
728 		}
729 		/* chain the operations */
730 		submit.depend_tx = tx;
731 
732 		if (clen < len) /* hit end of page */
733 			break;
734 		page_offset +=  len;
735 	}
736 
737 	return tx;
738 }
739 
740 static void ops_complete_biofill(void *stripe_head_ref)
741 {
742 	struct stripe_head *sh = stripe_head_ref;
743 	struct bio *return_bi = NULL;
744 	struct r5conf *conf = sh->raid_conf;
745 	int i;
746 
747 	pr_debug("%s: stripe %llu\n", __func__,
748 		(unsigned long long)sh->sector);
749 
750 	/* clear completed biofills */
751 	spin_lock_irq(&conf->device_lock);
752 	for (i = sh->disks; i--; ) {
753 		struct r5dev *dev = &sh->dev[i];
754 
755 		/* acknowledge completion of a biofill operation */
756 		/* and check if we need to reply to a read request,
757 		 * new R5_Wantfill requests are held off until
758 		 * !STRIPE_BIOFILL_RUN
759 		 */
760 		if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
761 			struct bio *rbi, *rbi2;
762 
763 			BUG_ON(!dev->read);
764 			rbi = dev->read;
765 			dev->read = NULL;
766 			while (rbi && rbi->bi_sector <
767 				dev->sector + STRIPE_SECTORS) {
768 				rbi2 = r5_next_bio(rbi, dev->sector);
769 				if (!raid5_dec_bi_phys_segments(rbi)) {
770 					rbi->bi_next = return_bi;
771 					return_bi = rbi;
772 				}
773 				rbi = rbi2;
774 			}
775 		}
776 	}
777 	spin_unlock_irq(&conf->device_lock);
778 	clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
779 
780 	return_io(return_bi);
781 
782 	set_bit(STRIPE_HANDLE, &sh->state);
783 	release_stripe(sh);
784 }
785 
786 static void ops_run_biofill(struct stripe_head *sh)
787 {
788 	struct dma_async_tx_descriptor *tx = NULL;
789 	struct r5conf *conf = sh->raid_conf;
790 	struct async_submit_ctl submit;
791 	int i;
792 
793 	pr_debug("%s: stripe %llu\n", __func__,
794 		(unsigned long long)sh->sector);
795 
796 	for (i = sh->disks; i--; ) {
797 		struct r5dev *dev = &sh->dev[i];
798 		if (test_bit(R5_Wantfill, &dev->flags)) {
799 			struct bio *rbi;
800 			spin_lock_irq(&conf->device_lock);
801 			dev->read = rbi = dev->toread;
802 			dev->toread = NULL;
803 			spin_unlock_irq(&conf->device_lock);
804 			while (rbi && rbi->bi_sector <
805 				dev->sector + STRIPE_SECTORS) {
806 				tx = async_copy_data(0, rbi, dev->page,
807 					dev->sector, tx);
808 				rbi = r5_next_bio(rbi, dev->sector);
809 			}
810 		}
811 	}
812 
813 	atomic_inc(&sh->count);
814 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
815 	async_trigger_callback(&submit);
816 }
817 
818 static void mark_target_uptodate(struct stripe_head *sh, int target)
819 {
820 	struct r5dev *tgt;
821 
822 	if (target < 0)
823 		return;
824 
825 	tgt = &sh->dev[target];
826 	set_bit(R5_UPTODATE, &tgt->flags);
827 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
828 	clear_bit(R5_Wantcompute, &tgt->flags);
829 }
830 
831 static void ops_complete_compute(void *stripe_head_ref)
832 {
833 	struct stripe_head *sh = stripe_head_ref;
834 
835 	pr_debug("%s: stripe %llu\n", __func__,
836 		(unsigned long long)sh->sector);
837 
838 	/* mark the computed target(s) as uptodate */
839 	mark_target_uptodate(sh, sh->ops.target);
840 	mark_target_uptodate(sh, sh->ops.target2);
841 
842 	clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
843 	if (sh->check_state == check_state_compute_run)
844 		sh->check_state = check_state_compute_result;
845 	set_bit(STRIPE_HANDLE, &sh->state);
846 	release_stripe(sh);
847 }
848 
849 /* return a pointer to the address conversion region of the scribble buffer */
850 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
851 				 struct raid5_percpu *percpu)
852 {
853 	return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
854 }
855 
856 static struct dma_async_tx_descriptor *
857 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
858 {
859 	int disks = sh->disks;
860 	struct page **xor_srcs = percpu->scribble;
861 	int target = sh->ops.target;
862 	struct r5dev *tgt = &sh->dev[target];
863 	struct page *xor_dest = tgt->page;
864 	int count = 0;
865 	struct dma_async_tx_descriptor *tx;
866 	struct async_submit_ctl submit;
867 	int i;
868 
869 	pr_debug("%s: stripe %llu block: %d\n",
870 		__func__, (unsigned long long)sh->sector, target);
871 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
872 
873 	for (i = disks; i--; )
874 		if (i != target)
875 			xor_srcs[count++] = sh->dev[i].page;
876 
877 	atomic_inc(&sh->count);
878 
879 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
880 			  ops_complete_compute, sh, to_addr_conv(sh, percpu));
881 	if (unlikely(count == 1))
882 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
883 	else
884 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
885 
886 	return tx;
887 }
888 
889 /* set_syndrome_sources - populate source buffers for gen_syndrome
890  * @srcs - (struct page *) array of size sh->disks
891  * @sh - stripe_head to parse
892  *
893  * Populates srcs in proper layout order for the stripe and returns the
894  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
895  * destination buffer is recorded in srcs[count] and the Q destination
896  * is recorded in srcs[count+1]].
897  */
898 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
899 {
900 	int disks = sh->disks;
901 	int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
902 	int d0_idx = raid6_d0(sh);
903 	int count;
904 	int i;
905 
906 	for (i = 0; i < disks; i++)
907 		srcs[i] = NULL;
908 
909 	count = 0;
910 	i = d0_idx;
911 	do {
912 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
913 
914 		srcs[slot] = sh->dev[i].page;
915 		i = raid6_next_disk(i, disks);
916 	} while (i != d0_idx);
917 
918 	return syndrome_disks;
919 }
920 
921 static struct dma_async_tx_descriptor *
922 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
923 {
924 	int disks = sh->disks;
925 	struct page **blocks = percpu->scribble;
926 	int target;
927 	int qd_idx = sh->qd_idx;
928 	struct dma_async_tx_descriptor *tx;
929 	struct async_submit_ctl submit;
930 	struct r5dev *tgt;
931 	struct page *dest;
932 	int i;
933 	int count;
934 
935 	if (sh->ops.target < 0)
936 		target = sh->ops.target2;
937 	else if (sh->ops.target2 < 0)
938 		target = sh->ops.target;
939 	else
940 		/* we should only have one valid target */
941 		BUG();
942 	BUG_ON(target < 0);
943 	pr_debug("%s: stripe %llu block: %d\n",
944 		__func__, (unsigned long long)sh->sector, target);
945 
946 	tgt = &sh->dev[target];
947 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
948 	dest = tgt->page;
949 
950 	atomic_inc(&sh->count);
951 
952 	if (target == qd_idx) {
953 		count = set_syndrome_sources(blocks, sh);
954 		blocks[count] = NULL; /* regenerating p is not necessary */
955 		BUG_ON(blocks[count+1] != dest); /* q should already be set */
956 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
957 				  ops_complete_compute, sh,
958 				  to_addr_conv(sh, percpu));
959 		tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
960 	} else {
961 		/* Compute any data- or p-drive using XOR */
962 		count = 0;
963 		for (i = disks; i-- ; ) {
964 			if (i == target || i == qd_idx)
965 				continue;
966 			blocks[count++] = sh->dev[i].page;
967 		}
968 
969 		init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
970 				  NULL, ops_complete_compute, sh,
971 				  to_addr_conv(sh, percpu));
972 		tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
973 	}
974 
975 	return tx;
976 }
977 
978 static struct dma_async_tx_descriptor *
979 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
980 {
981 	int i, count, disks = sh->disks;
982 	int syndrome_disks = sh->ddf_layout ? disks : disks-2;
983 	int d0_idx = raid6_d0(sh);
984 	int faila = -1, failb = -1;
985 	int target = sh->ops.target;
986 	int target2 = sh->ops.target2;
987 	struct r5dev *tgt = &sh->dev[target];
988 	struct r5dev *tgt2 = &sh->dev[target2];
989 	struct dma_async_tx_descriptor *tx;
990 	struct page **blocks = percpu->scribble;
991 	struct async_submit_ctl submit;
992 
993 	pr_debug("%s: stripe %llu block1: %d block2: %d\n",
994 		 __func__, (unsigned long long)sh->sector, target, target2);
995 	BUG_ON(target < 0 || target2 < 0);
996 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
997 	BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
998 
999 	/* we need to open-code set_syndrome_sources to handle the
1000 	 * slot number conversion for 'faila' and 'failb'
1001 	 */
1002 	for (i = 0; i < disks ; i++)
1003 		blocks[i] = NULL;
1004 	count = 0;
1005 	i = d0_idx;
1006 	do {
1007 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1008 
1009 		blocks[slot] = sh->dev[i].page;
1010 
1011 		if (i == target)
1012 			faila = slot;
1013 		if (i == target2)
1014 			failb = slot;
1015 		i = raid6_next_disk(i, disks);
1016 	} while (i != d0_idx);
1017 
1018 	BUG_ON(faila == failb);
1019 	if (failb < faila)
1020 		swap(faila, failb);
1021 	pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1022 		 __func__, (unsigned long long)sh->sector, faila, failb);
1023 
1024 	atomic_inc(&sh->count);
1025 
1026 	if (failb == syndrome_disks+1) {
1027 		/* Q disk is one of the missing disks */
1028 		if (faila == syndrome_disks) {
1029 			/* Missing P+Q, just recompute */
1030 			init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1031 					  ops_complete_compute, sh,
1032 					  to_addr_conv(sh, percpu));
1033 			return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1034 						  STRIPE_SIZE, &submit);
1035 		} else {
1036 			struct page *dest;
1037 			int data_target;
1038 			int qd_idx = sh->qd_idx;
1039 
1040 			/* Missing D+Q: recompute D from P, then recompute Q */
1041 			if (target == qd_idx)
1042 				data_target = target2;
1043 			else
1044 				data_target = target;
1045 
1046 			count = 0;
1047 			for (i = disks; i-- ; ) {
1048 				if (i == data_target || i == qd_idx)
1049 					continue;
1050 				blocks[count++] = sh->dev[i].page;
1051 			}
1052 			dest = sh->dev[data_target].page;
1053 			init_async_submit(&submit,
1054 					  ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1055 					  NULL, NULL, NULL,
1056 					  to_addr_conv(sh, percpu));
1057 			tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1058 				       &submit);
1059 
1060 			count = set_syndrome_sources(blocks, sh);
1061 			init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1062 					  ops_complete_compute, sh,
1063 					  to_addr_conv(sh, percpu));
1064 			return async_gen_syndrome(blocks, 0, count+2,
1065 						  STRIPE_SIZE, &submit);
1066 		}
1067 	} else {
1068 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1069 				  ops_complete_compute, sh,
1070 				  to_addr_conv(sh, percpu));
1071 		if (failb == syndrome_disks) {
1072 			/* We're missing D+P. */
1073 			return async_raid6_datap_recov(syndrome_disks+2,
1074 						       STRIPE_SIZE, faila,
1075 						       blocks, &submit);
1076 		} else {
1077 			/* We're missing D+D. */
1078 			return async_raid6_2data_recov(syndrome_disks+2,
1079 						       STRIPE_SIZE, faila, failb,
1080 						       blocks, &submit);
1081 		}
1082 	}
1083 }
1084 
1085 
1086 static void ops_complete_prexor(void *stripe_head_ref)
1087 {
1088 	struct stripe_head *sh = stripe_head_ref;
1089 
1090 	pr_debug("%s: stripe %llu\n", __func__,
1091 		(unsigned long long)sh->sector);
1092 }
1093 
1094 static struct dma_async_tx_descriptor *
1095 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1096 	       struct dma_async_tx_descriptor *tx)
1097 {
1098 	int disks = sh->disks;
1099 	struct page **xor_srcs = percpu->scribble;
1100 	int count = 0, pd_idx = sh->pd_idx, i;
1101 	struct async_submit_ctl submit;
1102 
1103 	/* existing parity data subtracted */
1104 	struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1105 
1106 	pr_debug("%s: stripe %llu\n", __func__,
1107 		(unsigned long long)sh->sector);
1108 
1109 	for (i = disks; i--; ) {
1110 		struct r5dev *dev = &sh->dev[i];
1111 		/* Only process blocks that are known to be uptodate */
1112 		if (test_bit(R5_Wantdrain, &dev->flags))
1113 			xor_srcs[count++] = dev->page;
1114 	}
1115 
1116 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1117 			  ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1118 	tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1119 
1120 	return tx;
1121 }
1122 
1123 static struct dma_async_tx_descriptor *
1124 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1125 {
1126 	int disks = sh->disks;
1127 	int i;
1128 
1129 	pr_debug("%s: stripe %llu\n", __func__,
1130 		(unsigned long long)sh->sector);
1131 
1132 	for (i = disks; i--; ) {
1133 		struct r5dev *dev = &sh->dev[i];
1134 		struct bio *chosen;
1135 
1136 		if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1137 			struct bio *wbi;
1138 
1139 			spin_lock_irq(&sh->raid_conf->device_lock);
1140 			chosen = dev->towrite;
1141 			dev->towrite = NULL;
1142 			BUG_ON(dev->written);
1143 			wbi = dev->written = chosen;
1144 			spin_unlock_irq(&sh->raid_conf->device_lock);
1145 
1146 			while (wbi && wbi->bi_sector <
1147 				dev->sector + STRIPE_SECTORS) {
1148 				if (wbi->bi_rw & REQ_FUA)
1149 					set_bit(R5_WantFUA, &dev->flags);
1150 				if (wbi->bi_rw & REQ_SYNC)
1151 					set_bit(R5_SyncIO, &dev->flags);
1152 				tx = async_copy_data(1, wbi, dev->page,
1153 					dev->sector, tx);
1154 				wbi = r5_next_bio(wbi, dev->sector);
1155 			}
1156 		}
1157 	}
1158 
1159 	return tx;
1160 }
1161 
1162 static void ops_complete_reconstruct(void *stripe_head_ref)
1163 {
1164 	struct stripe_head *sh = stripe_head_ref;
1165 	int disks = sh->disks;
1166 	int pd_idx = sh->pd_idx;
1167 	int qd_idx = sh->qd_idx;
1168 	int i;
1169 	bool fua = false, sync = false;
1170 
1171 	pr_debug("%s: stripe %llu\n", __func__,
1172 		(unsigned long long)sh->sector);
1173 
1174 	for (i = disks; i--; ) {
1175 		fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1176 		sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1177 	}
1178 
1179 	for (i = disks; i--; ) {
1180 		struct r5dev *dev = &sh->dev[i];
1181 
1182 		if (dev->written || i == pd_idx || i == qd_idx) {
1183 			set_bit(R5_UPTODATE, &dev->flags);
1184 			if (fua)
1185 				set_bit(R5_WantFUA, &dev->flags);
1186 			if (sync)
1187 				set_bit(R5_SyncIO, &dev->flags);
1188 		}
1189 	}
1190 
1191 	if (sh->reconstruct_state == reconstruct_state_drain_run)
1192 		sh->reconstruct_state = reconstruct_state_drain_result;
1193 	else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1194 		sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1195 	else {
1196 		BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1197 		sh->reconstruct_state = reconstruct_state_result;
1198 	}
1199 
1200 	set_bit(STRIPE_HANDLE, &sh->state);
1201 	release_stripe(sh);
1202 }
1203 
1204 static void
1205 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1206 		     struct dma_async_tx_descriptor *tx)
1207 {
1208 	int disks = sh->disks;
1209 	struct page **xor_srcs = percpu->scribble;
1210 	struct async_submit_ctl submit;
1211 	int count = 0, pd_idx = sh->pd_idx, i;
1212 	struct page *xor_dest;
1213 	int prexor = 0;
1214 	unsigned long flags;
1215 
1216 	pr_debug("%s: stripe %llu\n", __func__,
1217 		(unsigned long long)sh->sector);
1218 
1219 	/* check if prexor is active which means only process blocks
1220 	 * that are part of a read-modify-write (written)
1221 	 */
1222 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1223 		prexor = 1;
1224 		xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1225 		for (i = disks; i--; ) {
1226 			struct r5dev *dev = &sh->dev[i];
1227 			if (dev->written)
1228 				xor_srcs[count++] = dev->page;
1229 		}
1230 	} else {
1231 		xor_dest = sh->dev[pd_idx].page;
1232 		for (i = disks; i--; ) {
1233 			struct r5dev *dev = &sh->dev[i];
1234 			if (i != pd_idx)
1235 				xor_srcs[count++] = dev->page;
1236 		}
1237 	}
1238 
1239 	/* 1/ if we prexor'd then the dest is reused as a source
1240 	 * 2/ if we did not prexor then we are redoing the parity
1241 	 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1242 	 * for the synchronous xor case
1243 	 */
1244 	flags = ASYNC_TX_ACK |
1245 		(prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1246 
1247 	atomic_inc(&sh->count);
1248 
1249 	init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1250 			  to_addr_conv(sh, percpu));
1251 	if (unlikely(count == 1))
1252 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1253 	else
1254 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1255 }
1256 
1257 static void
1258 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1259 		     struct dma_async_tx_descriptor *tx)
1260 {
1261 	struct async_submit_ctl submit;
1262 	struct page **blocks = percpu->scribble;
1263 	int count;
1264 
1265 	pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1266 
1267 	count = set_syndrome_sources(blocks, sh);
1268 
1269 	atomic_inc(&sh->count);
1270 
1271 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1272 			  sh, to_addr_conv(sh, percpu));
1273 	async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1274 }
1275 
1276 static void ops_complete_check(void *stripe_head_ref)
1277 {
1278 	struct stripe_head *sh = stripe_head_ref;
1279 
1280 	pr_debug("%s: stripe %llu\n", __func__,
1281 		(unsigned long long)sh->sector);
1282 
1283 	sh->check_state = check_state_check_result;
1284 	set_bit(STRIPE_HANDLE, &sh->state);
1285 	release_stripe(sh);
1286 }
1287 
1288 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1289 {
1290 	int disks = sh->disks;
1291 	int pd_idx = sh->pd_idx;
1292 	int qd_idx = sh->qd_idx;
1293 	struct page *xor_dest;
1294 	struct page **xor_srcs = percpu->scribble;
1295 	struct dma_async_tx_descriptor *tx;
1296 	struct async_submit_ctl submit;
1297 	int count;
1298 	int i;
1299 
1300 	pr_debug("%s: stripe %llu\n", __func__,
1301 		(unsigned long long)sh->sector);
1302 
1303 	count = 0;
1304 	xor_dest = sh->dev[pd_idx].page;
1305 	xor_srcs[count++] = xor_dest;
1306 	for (i = disks; i--; ) {
1307 		if (i == pd_idx || i == qd_idx)
1308 			continue;
1309 		xor_srcs[count++] = sh->dev[i].page;
1310 	}
1311 
1312 	init_async_submit(&submit, 0, NULL, NULL, NULL,
1313 			  to_addr_conv(sh, percpu));
1314 	tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1315 			   &sh->ops.zero_sum_result, &submit);
1316 
1317 	atomic_inc(&sh->count);
1318 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1319 	tx = async_trigger_callback(&submit);
1320 }
1321 
1322 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1323 {
1324 	struct page **srcs = percpu->scribble;
1325 	struct async_submit_ctl submit;
1326 	int count;
1327 
1328 	pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1329 		(unsigned long long)sh->sector, checkp);
1330 
1331 	count = set_syndrome_sources(srcs, sh);
1332 	if (!checkp)
1333 		srcs[count] = NULL;
1334 
1335 	atomic_inc(&sh->count);
1336 	init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1337 			  sh, to_addr_conv(sh, percpu));
1338 	async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1339 			   &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1340 }
1341 
1342 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1343 {
1344 	int overlap_clear = 0, i, disks = sh->disks;
1345 	struct dma_async_tx_descriptor *tx = NULL;
1346 	struct r5conf *conf = sh->raid_conf;
1347 	int level = conf->level;
1348 	struct raid5_percpu *percpu;
1349 	unsigned long cpu;
1350 
1351 	cpu = get_cpu();
1352 	percpu = per_cpu_ptr(conf->percpu, cpu);
1353 	if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1354 		ops_run_biofill(sh);
1355 		overlap_clear++;
1356 	}
1357 
1358 	if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1359 		if (level < 6)
1360 			tx = ops_run_compute5(sh, percpu);
1361 		else {
1362 			if (sh->ops.target2 < 0 || sh->ops.target < 0)
1363 				tx = ops_run_compute6_1(sh, percpu);
1364 			else
1365 				tx = ops_run_compute6_2(sh, percpu);
1366 		}
1367 		/* terminate the chain if reconstruct is not set to be run */
1368 		if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1369 			async_tx_ack(tx);
1370 	}
1371 
1372 	if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1373 		tx = ops_run_prexor(sh, percpu, tx);
1374 
1375 	if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1376 		tx = ops_run_biodrain(sh, tx);
1377 		overlap_clear++;
1378 	}
1379 
1380 	if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1381 		if (level < 6)
1382 			ops_run_reconstruct5(sh, percpu, tx);
1383 		else
1384 			ops_run_reconstruct6(sh, percpu, tx);
1385 	}
1386 
1387 	if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1388 		if (sh->check_state == check_state_run)
1389 			ops_run_check_p(sh, percpu);
1390 		else if (sh->check_state == check_state_run_q)
1391 			ops_run_check_pq(sh, percpu, 0);
1392 		else if (sh->check_state == check_state_run_pq)
1393 			ops_run_check_pq(sh, percpu, 1);
1394 		else
1395 			BUG();
1396 	}
1397 
1398 	if (overlap_clear)
1399 		for (i = disks; i--; ) {
1400 			struct r5dev *dev = &sh->dev[i];
1401 			if (test_and_clear_bit(R5_Overlap, &dev->flags))
1402 				wake_up(&sh->raid_conf->wait_for_overlap);
1403 		}
1404 	put_cpu();
1405 }
1406 
1407 #ifdef CONFIG_MULTICORE_RAID456
1408 static void async_run_ops(void *param, async_cookie_t cookie)
1409 {
1410 	struct stripe_head *sh = param;
1411 	unsigned long ops_request = sh->ops.request;
1412 
1413 	clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1414 	wake_up(&sh->ops.wait_for_ops);
1415 
1416 	__raid_run_ops(sh, ops_request);
1417 	release_stripe(sh);
1418 }
1419 
1420 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1421 {
1422 	/* since handle_stripe can be called outside of raid5d context
1423 	 * we need to ensure sh->ops.request is de-staged before another
1424 	 * request arrives
1425 	 */
1426 	wait_event(sh->ops.wait_for_ops,
1427 		   !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1428 	sh->ops.request = ops_request;
1429 
1430 	atomic_inc(&sh->count);
1431 	async_schedule(async_run_ops, sh);
1432 }
1433 #else
1434 #define raid_run_ops __raid_run_ops
1435 #endif
1436 
1437 static int grow_one_stripe(struct r5conf *conf)
1438 {
1439 	struct stripe_head *sh;
1440 	sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1441 	if (!sh)
1442 		return 0;
1443 
1444 	sh->raid_conf = conf;
1445 	#ifdef CONFIG_MULTICORE_RAID456
1446 	init_waitqueue_head(&sh->ops.wait_for_ops);
1447 	#endif
1448 
1449 	if (grow_buffers(sh)) {
1450 		shrink_buffers(sh);
1451 		kmem_cache_free(conf->slab_cache, sh);
1452 		return 0;
1453 	}
1454 	/* we just created an active stripe so... */
1455 	atomic_set(&sh->count, 1);
1456 	atomic_inc(&conf->active_stripes);
1457 	INIT_LIST_HEAD(&sh->lru);
1458 	release_stripe(sh);
1459 	return 1;
1460 }
1461 
1462 static int grow_stripes(struct r5conf *conf, int num)
1463 {
1464 	struct kmem_cache *sc;
1465 	int devs = max(conf->raid_disks, conf->previous_raid_disks);
1466 
1467 	if (conf->mddev->gendisk)
1468 		sprintf(conf->cache_name[0],
1469 			"raid%d-%s", conf->level, mdname(conf->mddev));
1470 	else
1471 		sprintf(conf->cache_name[0],
1472 			"raid%d-%p", conf->level, conf->mddev);
1473 	sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1474 
1475 	conf->active_name = 0;
1476 	sc = kmem_cache_create(conf->cache_name[conf->active_name],
1477 			       sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1478 			       0, 0, NULL);
1479 	if (!sc)
1480 		return 1;
1481 	conf->slab_cache = sc;
1482 	conf->pool_size = devs;
1483 	while (num--)
1484 		if (!grow_one_stripe(conf))
1485 			return 1;
1486 	return 0;
1487 }
1488 
1489 /**
1490  * scribble_len - return the required size of the scribble region
1491  * @num - total number of disks in the array
1492  *
1493  * The size must be enough to contain:
1494  * 1/ a struct page pointer for each device in the array +2
1495  * 2/ room to convert each entry in (1) to its corresponding dma
1496  *    (dma_map_page()) or page (page_address()) address.
1497  *
1498  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1499  * calculate over all devices (not just the data blocks), using zeros in place
1500  * of the P and Q blocks.
1501  */
1502 static size_t scribble_len(int num)
1503 {
1504 	size_t len;
1505 
1506 	len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1507 
1508 	return len;
1509 }
1510 
1511 static int resize_stripes(struct r5conf *conf, int newsize)
1512 {
1513 	/* Make all the stripes able to hold 'newsize' devices.
1514 	 * New slots in each stripe get 'page' set to a new page.
1515 	 *
1516 	 * This happens in stages:
1517 	 * 1/ create a new kmem_cache and allocate the required number of
1518 	 *    stripe_heads.
1519 	 * 2/ gather all the old stripe_heads and tranfer the pages across
1520 	 *    to the new stripe_heads.  This will have the side effect of
1521 	 *    freezing the array as once all stripe_heads have been collected,
1522 	 *    no IO will be possible.  Old stripe heads are freed once their
1523 	 *    pages have been transferred over, and the old kmem_cache is
1524 	 *    freed when all stripes are done.
1525 	 * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1526 	 *    we simple return a failre status - no need to clean anything up.
1527 	 * 4/ allocate new pages for the new slots in the new stripe_heads.
1528 	 *    If this fails, we don't bother trying the shrink the
1529 	 *    stripe_heads down again, we just leave them as they are.
1530 	 *    As each stripe_head is processed the new one is released into
1531 	 *    active service.
1532 	 *
1533 	 * Once step2 is started, we cannot afford to wait for a write,
1534 	 * so we use GFP_NOIO allocations.
1535 	 */
1536 	struct stripe_head *osh, *nsh;
1537 	LIST_HEAD(newstripes);
1538 	struct disk_info *ndisks;
1539 	unsigned long cpu;
1540 	int err;
1541 	struct kmem_cache *sc;
1542 	int i;
1543 
1544 	if (newsize <= conf->pool_size)
1545 		return 0; /* never bother to shrink */
1546 
1547 	err = md_allow_write(conf->mddev);
1548 	if (err)
1549 		return err;
1550 
1551 	/* Step 1 */
1552 	sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1553 			       sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1554 			       0, 0, NULL);
1555 	if (!sc)
1556 		return -ENOMEM;
1557 
1558 	for (i = conf->max_nr_stripes; i; i--) {
1559 		nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1560 		if (!nsh)
1561 			break;
1562 
1563 		nsh->raid_conf = conf;
1564 		#ifdef CONFIG_MULTICORE_RAID456
1565 		init_waitqueue_head(&nsh->ops.wait_for_ops);
1566 		#endif
1567 
1568 		list_add(&nsh->lru, &newstripes);
1569 	}
1570 	if (i) {
1571 		/* didn't get enough, give up */
1572 		while (!list_empty(&newstripes)) {
1573 			nsh = list_entry(newstripes.next, struct stripe_head, lru);
1574 			list_del(&nsh->lru);
1575 			kmem_cache_free(sc, nsh);
1576 		}
1577 		kmem_cache_destroy(sc);
1578 		return -ENOMEM;
1579 	}
1580 	/* Step 2 - Must use GFP_NOIO now.
1581 	 * OK, we have enough stripes, start collecting inactive
1582 	 * stripes and copying them over
1583 	 */
1584 	list_for_each_entry(nsh, &newstripes, lru) {
1585 		spin_lock_irq(&conf->device_lock);
1586 		wait_event_lock_irq(conf->wait_for_stripe,
1587 				    !list_empty(&conf->inactive_list),
1588 				    conf->device_lock,
1589 				    );
1590 		osh = get_free_stripe(conf);
1591 		spin_unlock_irq(&conf->device_lock);
1592 		atomic_set(&nsh->count, 1);
1593 		for(i=0; i<conf->pool_size; i++)
1594 			nsh->dev[i].page = osh->dev[i].page;
1595 		for( ; i<newsize; i++)
1596 			nsh->dev[i].page = NULL;
1597 		kmem_cache_free(conf->slab_cache, osh);
1598 	}
1599 	kmem_cache_destroy(conf->slab_cache);
1600 
1601 	/* Step 3.
1602 	 * At this point, we are holding all the stripes so the array
1603 	 * is completely stalled, so now is a good time to resize
1604 	 * conf->disks and the scribble region
1605 	 */
1606 	ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1607 	if (ndisks) {
1608 		for (i=0; i<conf->raid_disks; i++)
1609 			ndisks[i] = conf->disks[i];
1610 		kfree(conf->disks);
1611 		conf->disks = ndisks;
1612 	} else
1613 		err = -ENOMEM;
1614 
1615 	get_online_cpus();
1616 	conf->scribble_len = scribble_len(newsize);
1617 	for_each_present_cpu(cpu) {
1618 		struct raid5_percpu *percpu;
1619 		void *scribble;
1620 
1621 		percpu = per_cpu_ptr(conf->percpu, cpu);
1622 		scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1623 
1624 		if (scribble) {
1625 			kfree(percpu->scribble);
1626 			percpu->scribble = scribble;
1627 		} else {
1628 			err = -ENOMEM;
1629 			break;
1630 		}
1631 	}
1632 	put_online_cpus();
1633 
1634 	/* Step 4, return new stripes to service */
1635 	while(!list_empty(&newstripes)) {
1636 		nsh = list_entry(newstripes.next, struct stripe_head, lru);
1637 		list_del_init(&nsh->lru);
1638 
1639 		for (i=conf->raid_disks; i < newsize; i++)
1640 			if (nsh->dev[i].page == NULL) {
1641 				struct page *p = alloc_page(GFP_NOIO);
1642 				nsh->dev[i].page = p;
1643 				if (!p)
1644 					err = -ENOMEM;
1645 			}
1646 		release_stripe(nsh);
1647 	}
1648 	/* critical section pass, GFP_NOIO no longer needed */
1649 
1650 	conf->slab_cache = sc;
1651 	conf->active_name = 1-conf->active_name;
1652 	conf->pool_size = newsize;
1653 	return err;
1654 }
1655 
1656 static int drop_one_stripe(struct r5conf *conf)
1657 {
1658 	struct stripe_head *sh;
1659 
1660 	spin_lock_irq(&conf->device_lock);
1661 	sh = get_free_stripe(conf);
1662 	spin_unlock_irq(&conf->device_lock);
1663 	if (!sh)
1664 		return 0;
1665 	BUG_ON(atomic_read(&sh->count));
1666 	shrink_buffers(sh);
1667 	kmem_cache_free(conf->slab_cache, sh);
1668 	atomic_dec(&conf->active_stripes);
1669 	return 1;
1670 }
1671 
1672 static void shrink_stripes(struct r5conf *conf)
1673 {
1674 	while (drop_one_stripe(conf))
1675 		;
1676 
1677 	if (conf->slab_cache)
1678 		kmem_cache_destroy(conf->slab_cache);
1679 	conf->slab_cache = NULL;
1680 }
1681 
1682 static void raid5_end_read_request(struct bio * bi, int error)
1683 {
1684 	struct stripe_head *sh = bi->bi_private;
1685 	struct r5conf *conf = sh->raid_conf;
1686 	int disks = sh->disks, i;
1687 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1688 	char b[BDEVNAME_SIZE];
1689 	struct md_rdev *rdev = NULL;
1690 	sector_t s;
1691 
1692 	for (i=0 ; i<disks; i++)
1693 		if (bi == &sh->dev[i].req)
1694 			break;
1695 
1696 	pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1697 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
1698 		uptodate);
1699 	if (i == disks) {
1700 		BUG();
1701 		return;
1702 	}
1703 	if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1704 		/* If replacement finished while this request was outstanding,
1705 		 * 'replacement' might be NULL already.
1706 		 * In that case it moved down to 'rdev'.
1707 		 * rdev is not removed until all requests are finished.
1708 		 */
1709 		rdev = conf->disks[i].replacement;
1710 	if (!rdev)
1711 		rdev = conf->disks[i].rdev;
1712 
1713 	if (use_new_offset(conf, sh))
1714 		s = sh->sector + rdev->new_data_offset;
1715 	else
1716 		s = sh->sector + rdev->data_offset;
1717 	if (uptodate) {
1718 		set_bit(R5_UPTODATE, &sh->dev[i].flags);
1719 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1720 			/* Note that this cannot happen on a
1721 			 * replacement device.  We just fail those on
1722 			 * any error
1723 			 */
1724 			printk_ratelimited(
1725 				KERN_INFO
1726 				"md/raid:%s: read error corrected"
1727 				" (%lu sectors at %llu on %s)\n",
1728 				mdname(conf->mddev), STRIPE_SECTORS,
1729 				(unsigned long long)s,
1730 				bdevname(rdev->bdev, b));
1731 			atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1732 			clear_bit(R5_ReadError, &sh->dev[i].flags);
1733 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
1734 		}
1735 		if (atomic_read(&rdev->read_errors))
1736 			atomic_set(&rdev->read_errors, 0);
1737 	} else {
1738 		const char *bdn = bdevname(rdev->bdev, b);
1739 		int retry = 0;
1740 
1741 		clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1742 		atomic_inc(&rdev->read_errors);
1743 		if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1744 			printk_ratelimited(
1745 				KERN_WARNING
1746 				"md/raid:%s: read error on replacement device "
1747 				"(sector %llu on %s).\n",
1748 				mdname(conf->mddev),
1749 				(unsigned long long)s,
1750 				bdn);
1751 		else if (conf->mddev->degraded >= conf->max_degraded)
1752 			printk_ratelimited(
1753 				KERN_WARNING
1754 				"md/raid:%s: read error not correctable "
1755 				"(sector %llu on %s).\n",
1756 				mdname(conf->mddev),
1757 				(unsigned long long)s,
1758 				bdn);
1759 		else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1760 			/* Oh, no!!! */
1761 			printk_ratelimited(
1762 				KERN_WARNING
1763 				"md/raid:%s: read error NOT corrected!! "
1764 				"(sector %llu on %s).\n",
1765 				mdname(conf->mddev),
1766 				(unsigned long long)s,
1767 				bdn);
1768 		else if (atomic_read(&rdev->read_errors)
1769 			 > conf->max_nr_stripes)
1770 			printk(KERN_WARNING
1771 			       "md/raid:%s: Too many read errors, failing device %s.\n",
1772 			       mdname(conf->mddev), bdn);
1773 		else
1774 			retry = 1;
1775 		if (retry)
1776 			set_bit(R5_ReadError, &sh->dev[i].flags);
1777 		else {
1778 			clear_bit(R5_ReadError, &sh->dev[i].flags);
1779 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
1780 			md_error(conf->mddev, rdev);
1781 		}
1782 	}
1783 	rdev_dec_pending(rdev, conf->mddev);
1784 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
1785 	set_bit(STRIPE_HANDLE, &sh->state);
1786 	release_stripe(sh);
1787 }
1788 
1789 static void raid5_end_write_request(struct bio *bi, int error)
1790 {
1791 	struct stripe_head *sh = bi->bi_private;
1792 	struct r5conf *conf = sh->raid_conf;
1793 	int disks = sh->disks, i;
1794 	struct md_rdev *uninitialized_var(rdev);
1795 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1796 	sector_t first_bad;
1797 	int bad_sectors;
1798 	int replacement = 0;
1799 
1800 	for (i = 0 ; i < disks; i++) {
1801 		if (bi == &sh->dev[i].req) {
1802 			rdev = conf->disks[i].rdev;
1803 			break;
1804 		}
1805 		if (bi == &sh->dev[i].rreq) {
1806 			rdev = conf->disks[i].replacement;
1807 			if (rdev)
1808 				replacement = 1;
1809 			else
1810 				/* rdev was removed and 'replacement'
1811 				 * replaced it.  rdev is not removed
1812 				 * until all requests are finished.
1813 				 */
1814 				rdev = conf->disks[i].rdev;
1815 			break;
1816 		}
1817 	}
1818 	pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1819 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
1820 		uptodate);
1821 	if (i == disks) {
1822 		BUG();
1823 		return;
1824 	}
1825 
1826 	if (replacement) {
1827 		if (!uptodate)
1828 			md_error(conf->mddev, rdev);
1829 		else if (is_badblock(rdev, sh->sector,
1830 				     STRIPE_SECTORS,
1831 				     &first_bad, &bad_sectors))
1832 			set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1833 	} else {
1834 		if (!uptodate) {
1835 			set_bit(WriteErrorSeen, &rdev->flags);
1836 			set_bit(R5_WriteError, &sh->dev[i].flags);
1837 			if (!test_and_set_bit(WantReplacement, &rdev->flags))
1838 				set_bit(MD_RECOVERY_NEEDED,
1839 					&rdev->mddev->recovery);
1840 		} else if (is_badblock(rdev, sh->sector,
1841 				       STRIPE_SECTORS,
1842 				       &first_bad, &bad_sectors))
1843 			set_bit(R5_MadeGood, &sh->dev[i].flags);
1844 	}
1845 	rdev_dec_pending(rdev, conf->mddev);
1846 
1847 	if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
1848 		clear_bit(R5_LOCKED, &sh->dev[i].flags);
1849 	set_bit(STRIPE_HANDLE, &sh->state);
1850 	release_stripe(sh);
1851 }
1852 
1853 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1854 
1855 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1856 {
1857 	struct r5dev *dev = &sh->dev[i];
1858 
1859 	bio_init(&dev->req);
1860 	dev->req.bi_io_vec = &dev->vec;
1861 	dev->req.bi_vcnt++;
1862 	dev->req.bi_max_vecs++;
1863 	dev->req.bi_private = sh;
1864 	dev->vec.bv_page = dev->page;
1865 
1866 	bio_init(&dev->rreq);
1867 	dev->rreq.bi_io_vec = &dev->rvec;
1868 	dev->rreq.bi_vcnt++;
1869 	dev->rreq.bi_max_vecs++;
1870 	dev->rreq.bi_private = sh;
1871 	dev->rvec.bv_page = dev->page;
1872 
1873 	dev->flags = 0;
1874 	dev->sector = compute_blocknr(sh, i, previous);
1875 }
1876 
1877 static void error(struct mddev *mddev, struct md_rdev *rdev)
1878 {
1879 	char b[BDEVNAME_SIZE];
1880 	struct r5conf *conf = mddev->private;
1881 	unsigned long flags;
1882 	pr_debug("raid456: error called\n");
1883 
1884 	spin_lock_irqsave(&conf->device_lock, flags);
1885 	clear_bit(In_sync, &rdev->flags);
1886 	mddev->degraded = calc_degraded(conf);
1887 	spin_unlock_irqrestore(&conf->device_lock, flags);
1888 	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1889 
1890 	set_bit(Blocked, &rdev->flags);
1891 	set_bit(Faulty, &rdev->flags);
1892 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
1893 	printk(KERN_ALERT
1894 	       "md/raid:%s: Disk failure on %s, disabling device.\n"
1895 	       "md/raid:%s: Operation continuing on %d devices.\n",
1896 	       mdname(mddev),
1897 	       bdevname(rdev->bdev, b),
1898 	       mdname(mddev),
1899 	       conf->raid_disks - mddev->degraded);
1900 }
1901 
1902 /*
1903  * Input: a 'big' sector number,
1904  * Output: index of the data and parity disk, and the sector # in them.
1905  */
1906 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1907 				     int previous, int *dd_idx,
1908 				     struct stripe_head *sh)
1909 {
1910 	sector_t stripe, stripe2;
1911 	sector_t chunk_number;
1912 	unsigned int chunk_offset;
1913 	int pd_idx, qd_idx;
1914 	int ddf_layout = 0;
1915 	sector_t new_sector;
1916 	int algorithm = previous ? conf->prev_algo
1917 				 : conf->algorithm;
1918 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1919 					 : conf->chunk_sectors;
1920 	int raid_disks = previous ? conf->previous_raid_disks
1921 				  : conf->raid_disks;
1922 	int data_disks = raid_disks - conf->max_degraded;
1923 
1924 	/* First compute the information on this sector */
1925 
1926 	/*
1927 	 * Compute the chunk number and the sector offset inside the chunk
1928 	 */
1929 	chunk_offset = sector_div(r_sector, sectors_per_chunk);
1930 	chunk_number = r_sector;
1931 
1932 	/*
1933 	 * Compute the stripe number
1934 	 */
1935 	stripe = chunk_number;
1936 	*dd_idx = sector_div(stripe, data_disks);
1937 	stripe2 = stripe;
1938 	/*
1939 	 * Select the parity disk based on the user selected algorithm.
1940 	 */
1941 	pd_idx = qd_idx = -1;
1942 	switch(conf->level) {
1943 	case 4:
1944 		pd_idx = data_disks;
1945 		break;
1946 	case 5:
1947 		switch (algorithm) {
1948 		case ALGORITHM_LEFT_ASYMMETRIC:
1949 			pd_idx = data_disks - sector_div(stripe2, raid_disks);
1950 			if (*dd_idx >= pd_idx)
1951 				(*dd_idx)++;
1952 			break;
1953 		case ALGORITHM_RIGHT_ASYMMETRIC:
1954 			pd_idx = sector_div(stripe2, raid_disks);
1955 			if (*dd_idx >= pd_idx)
1956 				(*dd_idx)++;
1957 			break;
1958 		case ALGORITHM_LEFT_SYMMETRIC:
1959 			pd_idx = data_disks - sector_div(stripe2, raid_disks);
1960 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1961 			break;
1962 		case ALGORITHM_RIGHT_SYMMETRIC:
1963 			pd_idx = sector_div(stripe2, raid_disks);
1964 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1965 			break;
1966 		case ALGORITHM_PARITY_0:
1967 			pd_idx = 0;
1968 			(*dd_idx)++;
1969 			break;
1970 		case ALGORITHM_PARITY_N:
1971 			pd_idx = data_disks;
1972 			break;
1973 		default:
1974 			BUG();
1975 		}
1976 		break;
1977 	case 6:
1978 
1979 		switch (algorithm) {
1980 		case ALGORITHM_LEFT_ASYMMETRIC:
1981 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1982 			qd_idx = pd_idx + 1;
1983 			if (pd_idx == raid_disks-1) {
1984 				(*dd_idx)++;	/* Q D D D P */
1985 				qd_idx = 0;
1986 			} else if (*dd_idx >= pd_idx)
1987 				(*dd_idx) += 2; /* D D P Q D */
1988 			break;
1989 		case ALGORITHM_RIGHT_ASYMMETRIC:
1990 			pd_idx = sector_div(stripe2, raid_disks);
1991 			qd_idx = pd_idx + 1;
1992 			if (pd_idx == raid_disks-1) {
1993 				(*dd_idx)++;	/* Q D D D P */
1994 				qd_idx = 0;
1995 			} else if (*dd_idx >= pd_idx)
1996 				(*dd_idx) += 2; /* D D P Q D */
1997 			break;
1998 		case ALGORITHM_LEFT_SYMMETRIC:
1999 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2000 			qd_idx = (pd_idx + 1) % raid_disks;
2001 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2002 			break;
2003 		case ALGORITHM_RIGHT_SYMMETRIC:
2004 			pd_idx = sector_div(stripe2, raid_disks);
2005 			qd_idx = (pd_idx + 1) % raid_disks;
2006 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2007 			break;
2008 
2009 		case ALGORITHM_PARITY_0:
2010 			pd_idx = 0;
2011 			qd_idx = 1;
2012 			(*dd_idx) += 2;
2013 			break;
2014 		case ALGORITHM_PARITY_N:
2015 			pd_idx = data_disks;
2016 			qd_idx = data_disks + 1;
2017 			break;
2018 
2019 		case ALGORITHM_ROTATING_ZERO_RESTART:
2020 			/* Exactly the same as RIGHT_ASYMMETRIC, but or
2021 			 * of blocks for computing Q is different.
2022 			 */
2023 			pd_idx = sector_div(stripe2, raid_disks);
2024 			qd_idx = pd_idx + 1;
2025 			if (pd_idx == raid_disks-1) {
2026 				(*dd_idx)++;	/* Q D D D P */
2027 				qd_idx = 0;
2028 			} else if (*dd_idx >= pd_idx)
2029 				(*dd_idx) += 2; /* D D P Q D */
2030 			ddf_layout = 1;
2031 			break;
2032 
2033 		case ALGORITHM_ROTATING_N_RESTART:
2034 			/* Same a left_asymmetric, by first stripe is
2035 			 * D D D P Q  rather than
2036 			 * Q D D D P
2037 			 */
2038 			stripe2 += 1;
2039 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2040 			qd_idx = pd_idx + 1;
2041 			if (pd_idx == raid_disks-1) {
2042 				(*dd_idx)++;	/* Q D D D P */
2043 				qd_idx = 0;
2044 			} else if (*dd_idx >= pd_idx)
2045 				(*dd_idx) += 2; /* D D P Q D */
2046 			ddf_layout = 1;
2047 			break;
2048 
2049 		case ALGORITHM_ROTATING_N_CONTINUE:
2050 			/* Same as left_symmetric but Q is before P */
2051 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2052 			qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2053 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2054 			ddf_layout = 1;
2055 			break;
2056 
2057 		case ALGORITHM_LEFT_ASYMMETRIC_6:
2058 			/* RAID5 left_asymmetric, with Q on last device */
2059 			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2060 			if (*dd_idx >= pd_idx)
2061 				(*dd_idx)++;
2062 			qd_idx = raid_disks - 1;
2063 			break;
2064 
2065 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
2066 			pd_idx = sector_div(stripe2, raid_disks-1);
2067 			if (*dd_idx >= pd_idx)
2068 				(*dd_idx)++;
2069 			qd_idx = raid_disks - 1;
2070 			break;
2071 
2072 		case ALGORITHM_LEFT_SYMMETRIC_6:
2073 			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2074 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2075 			qd_idx = raid_disks - 1;
2076 			break;
2077 
2078 		case ALGORITHM_RIGHT_SYMMETRIC_6:
2079 			pd_idx = sector_div(stripe2, raid_disks-1);
2080 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2081 			qd_idx = raid_disks - 1;
2082 			break;
2083 
2084 		case ALGORITHM_PARITY_0_6:
2085 			pd_idx = 0;
2086 			(*dd_idx)++;
2087 			qd_idx = raid_disks - 1;
2088 			break;
2089 
2090 		default:
2091 			BUG();
2092 		}
2093 		break;
2094 	}
2095 
2096 	if (sh) {
2097 		sh->pd_idx = pd_idx;
2098 		sh->qd_idx = qd_idx;
2099 		sh->ddf_layout = ddf_layout;
2100 	}
2101 	/*
2102 	 * Finally, compute the new sector number
2103 	 */
2104 	new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2105 	return new_sector;
2106 }
2107 
2108 
2109 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2110 {
2111 	struct r5conf *conf = sh->raid_conf;
2112 	int raid_disks = sh->disks;
2113 	int data_disks = raid_disks - conf->max_degraded;
2114 	sector_t new_sector = sh->sector, check;
2115 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2116 					 : conf->chunk_sectors;
2117 	int algorithm = previous ? conf->prev_algo
2118 				 : conf->algorithm;
2119 	sector_t stripe;
2120 	int chunk_offset;
2121 	sector_t chunk_number;
2122 	int dummy1, dd_idx = i;
2123 	sector_t r_sector;
2124 	struct stripe_head sh2;
2125 
2126 
2127 	chunk_offset = sector_div(new_sector, sectors_per_chunk);
2128 	stripe = new_sector;
2129 
2130 	if (i == sh->pd_idx)
2131 		return 0;
2132 	switch(conf->level) {
2133 	case 4: break;
2134 	case 5:
2135 		switch (algorithm) {
2136 		case ALGORITHM_LEFT_ASYMMETRIC:
2137 		case ALGORITHM_RIGHT_ASYMMETRIC:
2138 			if (i > sh->pd_idx)
2139 				i--;
2140 			break;
2141 		case ALGORITHM_LEFT_SYMMETRIC:
2142 		case ALGORITHM_RIGHT_SYMMETRIC:
2143 			if (i < sh->pd_idx)
2144 				i += raid_disks;
2145 			i -= (sh->pd_idx + 1);
2146 			break;
2147 		case ALGORITHM_PARITY_0:
2148 			i -= 1;
2149 			break;
2150 		case ALGORITHM_PARITY_N:
2151 			break;
2152 		default:
2153 			BUG();
2154 		}
2155 		break;
2156 	case 6:
2157 		if (i == sh->qd_idx)
2158 			return 0; /* It is the Q disk */
2159 		switch (algorithm) {
2160 		case ALGORITHM_LEFT_ASYMMETRIC:
2161 		case ALGORITHM_RIGHT_ASYMMETRIC:
2162 		case ALGORITHM_ROTATING_ZERO_RESTART:
2163 		case ALGORITHM_ROTATING_N_RESTART:
2164 			if (sh->pd_idx == raid_disks-1)
2165 				i--;	/* Q D D D P */
2166 			else if (i > sh->pd_idx)
2167 				i -= 2; /* D D P Q D */
2168 			break;
2169 		case ALGORITHM_LEFT_SYMMETRIC:
2170 		case ALGORITHM_RIGHT_SYMMETRIC:
2171 			if (sh->pd_idx == raid_disks-1)
2172 				i--; /* Q D D D P */
2173 			else {
2174 				/* D D P Q D */
2175 				if (i < sh->pd_idx)
2176 					i += raid_disks;
2177 				i -= (sh->pd_idx + 2);
2178 			}
2179 			break;
2180 		case ALGORITHM_PARITY_0:
2181 			i -= 2;
2182 			break;
2183 		case ALGORITHM_PARITY_N:
2184 			break;
2185 		case ALGORITHM_ROTATING_N_CONTINUE:
2186 			/* Like left_symmetric, but P is before Q */
2187 			if (sh->pd_idx == 0)
2188 				i--;	/* P D D D Q */
2189 			else {
2190 				/* D D Q P D */
2191 				if (i < sh->pd_idx)
2192 					i += raid_disks;
2193 				i -= (sh->pd_idx + 1);
2194 			}
2195 			break;
2196 		case ALGORITHM_LEFT_ASYMMETRIC_6:
2197 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
2198 			if (i > sh->pd_idx)
2199 				i--;
2200 			break;
2201 		case ALGORITHM_LEFT_SYMMETRIC_6:
2202 		case ALGORITHM_RIGHT_SYMMETRIC_6:
2203 			if (i < sh->pd_idx)
2204 				i += data_disks + 1;
2205 			i -= (sh->pd_idx + 1);
2206 			break;
2207 		case ALGORITHM_PARITY_0_6:
2208 			i -= 1;
2209 			break;
2210 		default:
2211 			BUG();
2212 		}
2213 		break;
2214 	}
2215 
2216 	chunk_number = stripe * data_disks + i;
2217 	r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2218 
2219 	check = raid5_compute_sector(conf, r_sector,
2220 				     previous, &dummy1, &sh2);
2221 	if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2222 		|| sh2.qd_idx != sh->qd_idx) {
2223 		printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2224 		       mdname(conf->mddev));
2225 		return 0;
2226 	}
2227 	return r_sector;
2228 }
2229 
2230 
2231 static void
2232 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2233 			 int rcw, int expand)
2234 {
2235 	int i, pd_idx = sh->pd_idx, disks = sh->disks;
2236 	struct r5conf *conf = sh->raid_conf;
2237 	int level = conf->level;
2238 
2239 	if (rcw) {
2240 		/* if we are not expanding this is a proper write request, and
2241 		 * there will be bios with new data to be drained into the
2242 		 * stripe cache
2243 		 */
2244 		if (!expand) {
2245 			sh->reconstruct_state = reconstruct_state_drain_run;
2246 			set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2247 		} else
2248 			sh->reconstruct_state = reconstruct_state_run;
2249 
2250 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2251 
2252 		for (i = disks; i--; ) {
2253 			struct r5dev *dev = &sh->dev[i];
2254 
2255 			if (dev->towrite) {
2256 				set_bit(R5_LOCKED, &dev->flags);
2257 				set_bit(R5_Wantdrain, &dev->flags);
2258 				if (!expand)
2259 					clear_bit(R5_UPTODATE, &dev->flags);
2260 				s->locked++;
2261 			}
2262 		}
2263 		if (s->locked + conf->max_degraded == disks)
2264 			if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2265 				atomic_inc(&conf->pending_full_writes);
2266 	} else {
2267 		BUG_ON(level == 6);
2268 		BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2269 			test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2270 
2271 		sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2272 		set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2273 		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2274 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2275 
2276 		for (i = disks; i--; ) {
2277 			struct r5dev *dev = &sh->dev[i];
2278 			if (i == pd_idx)
2279 				continue;
2280 
2281 			if (dev->towrite &&
2282 			    (test_bit(R5_UPTODATE, &dev->flags) ||
2283 			     test_bit(R5_Wantcompute, &dev->flags))) {
2284 				set_bit(R5_Wantdrain, &dev->flags);
2285 				set_bit(R5_LOCKED, &dev->flags);
2286 				clear_bit(R5_UPTODATE, &dev->flags);
2287 				s->locked++;
2288 			}
2289 		}
2290 	}
2291 
2292 	/* keep the parity disk(s) locked while asynchronous operations
2293 	 * are in flight
2294 	 */
2295 	set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2296 	clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2297 	s->locked++;
2298 
2299 	if (level == 6) {
2300 		int qd_idx = sh->qd_idx;
2301 		struct r5dev *dev = &sh->dev[qd_idx];
2302 
2303 		set_bit(R5_LOCKED, &dev->flags);
2304 		clear_bit(R5_UPTODATE, &dev->flags);
2305 		s->locked++;
2306 	}
2307 
2308 	pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2309 		__func__, (unsigned long long)sh->sector,
2310 		s->locked, s->ops_request);
2311 }
2312 
2313 /*
2314  * Each stripe/dev can have one or more bion attached.
2315  * toread/towrite point to the first in a chain.
2316  * The bi_next chain must be in order.
2317  */
2318 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2319 {
2320 	struct bio **bip;
2321 	struct r5conf *conf = sh->raid_conf;
2322 	int firstwrite=0;
2323 
2324 	pr_debug("adding bi b#%llu to stripe s#%llu\n",
2325 		(unsigned long long)bi->bi_sector,
2326 		(unsigned long long)sh->sector);
2327 
2328 
2329 	spin_lock_irq(&conf->device_lock);
2330 	if (forwrite) {
2331 		bip = &sh->dev[dd_idx].towrite;
2332 		if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2333 			firstwrite = 1;
2334 	} else
2335 		bip = &sh->dev[dd_idx].toread;
2336 	while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2337 		if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2338 			goto overlap;
2339 		bip = & (*bip)->bi_next;
2340 	}
2341 	if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2342 		goto overlap;
2343 
2344 	BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2345 	if (*bip)
2346 		bi->bi_next = *bip;
2347 	*bip = bi;
2348 	bi->bi_phys_segments++;
2349 
2350 	if (forwrite) {
2351 		/* check if page is covered */
2352 		sector_t sector = sh->dev[dd_idx].sector;
2353 		for (bi=sh->dev[dd_idx].towrite;
2354 		     sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2355 			     bi && bi->bi_sector <= sector;
2356 		     bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2357 			if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2358 				sector = bi->bi_sector + (bi->bi_size>>9);
2359 		}
2360 		if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2361 			set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2362 	}
2363 	spin_unlock_irq(&conf->device_lock);
2364 
2365 	pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2366 		(unsigned long long)(*bip)->bi_sector,
2367 		(unsigned long long)sh->sector, dd_idx);
2368 
2369 	if (conf->mddev->bitmap && firstwrite) {
2370 		bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2371 				  STRIPE_SECTORS, 0);
2372 		sh->bm_seq = conf->seq_flush+1;
2373 		set_bit(STRIPE_BIT_DELAY, &sh->state);
2374 	}
2375 	return 1;
2376 
2377  overlap:
2378 	set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2379 	spin_unlock_irq(&conf->device_lock);
2380 	return 0;
2381 }
2382 
2383 static void end_reshape(struct r5conf *conf);
2384 
2385 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2386 			    struct stripe_head *sh)
2387 {
2388 	int sectors_per_chunk =
2389 		previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2390 	int dd_idx;
2391 	int chunk_offset = sector_div(stripe, sectors_per_chunk);
2392 	int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2393 
2394 	raid5_compute_sector(conf,
2395 			     stripe * (disks - conf->max_degraded)
2396 			     *sectors_per_chunk + chunk_offset,
2397 			     previous,
2398 			     &dd_idx, sh);
2399 }
2400 
2401 static void
2402 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2403 				struct stripe_head_state *s, int disks,
2404 				struct bio **return_bi)
2405 {
2406 	int i;
2407 	for (i = disks; i--; ) {
2408 		struct bio *bi;
2409 		int bitmap_end = 0;
2410 
2411 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2412 			struct md_rdev *rdev;
2413 			rcu_read_lock();
2414 			rdev = rcu_dereference(conf->disks[i].rdev);
2415 			if (rdev && test_bit(In_sync, &rdev->flags))
2416 				atomic_inc(&rdev->nr_pending);
2417 			else
2418 				rdev = NULL;
2419 			rcu_read_unlock();
2420 			if (rdev) {
2421 				if (!rdev_set_badblocks(
2422 					    rdev,
2423 					    sh->sector,
2424 					    STRIPE_SECTORS, 0))
2425 					md_error(conf->mddev, rdev);
2426 				rdev_dec_pending(rdev, conf->mddev);
2427 			}
2428 		}
2429 		spin_lock_irq(&conf->device_lock);
2430 		/* fail all writes first */
2431 		bi = sh->dev[i].towrite;
2432 		sh->dev[i].towrite = NULL;
2433 		if (bi) {
2434 			s->to_write--;
2435 			bitmap_end = 1;
2436 		}
2437 
2438 		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2439 			wake_up(&conf->wait_for_overlap);
2440 
2441 		while (bi && bi->bi_sector <
2442 			sh->dev[i].sector + STRIPE_SECTORS) {
2443 			struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2444 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
2445 			if (!raid5_dec_bi_phys_segments(bi)) {
2446 				md_write_end(conf->mddev);
2447 				bi->bi_next = *return_bi;
2448 				*return_bi = bi;
2449 			}
2450 			bi = nextbi;
2451 		}
2452 		/* and fail all 'written' */
2453 		bi = sh->dev[i].written;
2454 		sh->dev[i].written = NULL;
2455 		if (bi) bitmap_end = 1;
2456 		while (bi && bi->bi_sector <
2457 		       sh->dev[i].sector + STRIPE_SECTORS) {
2458 			struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2459 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
2460 			if (!raid5_dec_bi_phys_segments(bi)) {
2461 				md_write_end(conf->mddev);
2462 				bi->bi_next = *return_bi;
2463 				*return_bi = bi;
2464 			}
2465 			bi = bi2;
2466 		}
2467 
2468 		/* fail any reads if this device is non-operational and
2469 		 * the data has not reached the cache yet.
2470 		 */
2471 		if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2472 		    (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2473 		      test_bit(R5_ReadError, &sh->dev[i].flags))) {
2474 			bi = sh->dev[i].toread;
2475 			sh->dev[i].toread = NULL;
2476 			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2477 				wake_up(&conf->wait_for_overlap);
2478 			if (bi) s->to_read--;
2479 			while (bi && bi->bi_sector <
2480 			       sh->dev[i].sector + STRIPE_SECTORS) {
2481 				struct bio *nextbi =
2482 					r5_next_bio(bi, sh->dev[i].sector);
2483 				clear_bit(BIO_UPTODATE, &bi->bi_flags);
2484 				if (!raid5_dec_bi_phys_segments(bi)) {
2485 					bi->bi_next = *return_bi;
2486 					*return_bi = bi;
2487 				}
2488 				bi = nextbi;
2489 			}
2490 		}
2491 		spin_unlock_irq(&conf->device_lock);
2492 		if (bitmap_end)
2493 			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2494 					STRIPE_SECTORS, 0, 0);
2495 		/* If we were in the middle of a write the parity block might
2496 		 * still be locked - so just clear all R5_LOCKED flags
2497 		 */
2498 		clear_bit(R5_LOCKED, &sh->dev[i].flags);
2499 	}
2500 
2501 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2502 		if (atomic_dec_and_test(&conf->pending_full_writes))
2503 			md_wakeup_thread(conf->mddev->thread);
2504 }
2505 
2506 static void
2507 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2508 		   struct stripe_head_state *s)
2509 {
2510 	int abort = 0;
2511 	int i;
2512 
2513 	clear_bit(STRIPE_SYNCING, &sh->state);
2514 	s->syncing = 0;
2515 	s->replacing = 0;
2516 	/* There is nothing more to do for sync/check/repair.
2517 	 * Don't even need to abort as that is handled elsewhere
2518 	 * if needed, and not always wanted e.g. if there is a known
2519 	 * bad block here.
2520 	 * For recover/replace we need to record a bad block on all
2521 	 * non-sync devices, or abort the recovery
2522 	 */
2523 	if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2524 		/* During recovery devices cannot be removed, so
2525 		 * locking and refcounting of rdevs is not needed
2526 		 */
2527 		for (i = 0; i < conf->raid_disks; i++) {
2528 			struct md_rdev *rdev = conf->disks[i].rdev;
2529 			if (rdev
2530 			    && !test_bit(Faulty, &rdev->flags)
2531 			    && !test_bit(In_sync, &rdev->flags)
2532 			    && !rdev_set_badblocks(rdev, sh->sector,
2533 						   STRIPE_SECTORS, 0))
2534 				abort = 1;
2535 			rdev = conf->disks[i].replacement;
2536 			if (rdev
2537 			    && !test_bit(Faulty, &rdev->flags)
2538 			    && !test_bit(In_sync, &rdev->flags)
2539 			    && !rdev_set_badblocks(rdev, sh->sector,
2540 						   STRIPE_SECTORS, 0))
2541 				abort = 1;
2542 		}
2543 		if (abort)
2544 			conf->recovery_disabled =
2545 				conf->mddev->recovery_disabled;
2546 	}
2547 	md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2548 }
2549 
2550 static int want_replace(struct stripe_head *sh, int disk_idx)
2551 {
2552 	struct md_rdev *rdev;
2553 	int rv = 0;
2554 	/* Doing recovery so rcu locking not required */
2555 	rdev = sh->raid_conf->disks[disk_idx].replacement;
2556 	if (rdev
2557 	    && !test_bit(Faulty, &rdev->flags)
2558 	    && !test_bit(In_sync, &rdev->flags)
2559 	    && (rdev->recovery_offset <= sh->sector
2560 		|| rdev->mddev->recovery_cp <= sh->sector))
2561 		rv = 1;
2562 
2563 	return rv;
2564 }
2565 
2566 /* fetch_block - checks the given member device to see if its data needs
2567  * to be read or computed to satisfy a request.
2568  *
2569  * Returns 1 when no more member devices need to be checked, otherwise returns
2570  * 0 to tell the loop in handle_stripe_fill to continue
2571  */
2572 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2573 		       int disk_idx, int disks)
2574 {
2575 	struct r5dev *dev = &sh->dev[disk_idx];
2576 	struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2577 				  &sh->dev[s->failed_num[1]] };
2578 
2579 	/* is the data in this block needed, and can we get it? */
2580 	if (!test_bit(R5_LOCKED, &dev->flags) &&
2581 	    !test_bit(R5_UPTODATE, &dev->flags) &&
2582 	    (dev->toread ||
2583 	     (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2584 	     s->syncing || s->expanding ||
2585 	     (s->replacing && want_replace(sh, disk_idx)) ||
2586 	     (s->failed >= 1 && fdev[0]->toread) ||
2587 	     (s->failed >= 2 && fdev[1]->toread) ||
2588 	     (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2589 	      !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2590 	     (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2591 		/* we would like to get this block, possibly by computing it,
2592 		 * otherwise read it if the backing disk is insync
2593 		 */
2594 		BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2595 		BUG_ON(test_bit(R5_Wantread, &dev->flags));
2596 		if ((s->uptodate == disks - 1) &&
2597 		    (s->failed && (disk_idx == s->failed_num[0] ||
2598 				   disk_idx == s->failed_num[1]))) {
2599 			/* have disk failed, and we're requested to fetch it;
2600 			 * do compute it
2601 			 */
2602 			pr_debug("Computing stripe %llu block %d\n",
2603 			       (unsigned long long)sh->sector, disk_idx);
2604 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2605 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2606 			set_bit(R5_Wantcompute, &dev->flags);
2607 			sh->ops.target = disk_idx;
2608 			sh->ops.target2 = -1; /* no 2nd target */
2609 			s->req_compute = 1;
2610 			/* Careful: from this point on 'uptodate' is in the eye
2611 			 * of raid_run_ops which services 'compute' operations
2612 			 * before writes. R5_Wantcompute flags a block that will
2613 			 * be R5_UPTODATE by the time it is needed for a
2614 			 * subsequent operation.
2615 			 */
2616 			s->uptodate++;
2617 			return 1;
2618 		} else if (s->uptodate == disks-2 && s->failed >= 2) {
2619 			/* Computing 2-failure is *very* expensive; only
2620 			 * do it if failed >= 2
2621 			 */
2622 			int other;
2623 			for (other = disks; other--; ) {
2624 				if (other == disk_idx)
2625 					continue;
2626 				if (!test_bit(R5_UPTODATE,
2627 				      &sh->dev[other].flags))
2628 					break;
2629 			}
2630 			BUG_ON(other < 0);
2631 			pr_debug("Computing stripe %llu blocks %d,%d\n",
2632 			       (unsigned long long)sh->sector,
2633 			       disk_idx, other);
2634 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2635 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2636 			set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2637 			set_bit(R5_Wantcompute, &sh->dev[other].flags);
2638 			sh->ops.target = disk_idx;
2639 			sh->ops.target2 = other;
2640 			s->uptodate += 2;
2641 			s->req_compute = 1;
2642 			return 1;
2643 		} else if (test_bit(R5_Insync, &dev->flags)) {
2644 			set_bit(R5_LOCKED, &dev->flags);
2645 			set_bit(R5_Wantread, &dev->flags);
2646 			s->locked++;
2647 			pr_debug("Reading block %d (sync=%d)\n",
2648 				disk_idx, s->syncing);
2649 		}
2650 	}
2651 
2652 	return 0;
2653 }
2654 
2655 /**
2656  * handle_stripe_fill - read or compute data to satisfy pending requests.
2657  */
2658 static void handle_stripe_fill(struct stripe_head *sh,
2659 			       struct stripe_head_state *s,
2660 			       int disks)
2661 {
2662 	int i;
2663 
2664 	/* look for blocks to read/compute, skip this if a compute
2665 	 * is already in flight, or if the stripe contents are in the
2666 	 * midst of changing due to a write
2667 	 */
2668 	if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2669 	    !sh->reconstruct_state)
2670 		for (i = disks; i--; )
2671 			if (fetch_block(sh, s, i, disks))
2672 				break;
2673 	set_bit(STRIPE_HANDLE, &sh->state);
2674 }
2675 
2676 
2677 /* handle_stripe_clean_event
2678  * any written block on an uptodate or failed drive can be returned.
2679  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2680  * never LOCKED, so we don't need to test 'failed' directly.
2681  */
2682 static void handle_stripe_clean_event(struct r5conf *conf,
2683 	struct stripe_head *sh, int disks, struct bio **return_bi)
2684 {
2685 	int i;
2686 	struct r5dev *dev;
2687 
2688 	for (i = disks; i--; )
2689 		if (sh->dev[i].written) {
2690 			dev = &sh->dev[i];
2691 			if (!test_bit(R5_LOCKED, &dev->flags) &&
2692 				test_bit(R5_UPTODATE, &dev->flags)) {
2693 				/* We can return any write requests */
2694 				struct bio *wbi, *wbi2;
2695 				int bitmap_end = 0;
2696 				pr_debug("Return write for disc %d\n", i);
2697 				spin_lock_irq(&conf->device_lock);
2698 				wbi = dev->written;
2699 				dev->written = NULL;
2700 				while (wbi && wbi->bi_sector <
2701 					dev->sector + STRIPE_SECTORS) {
2702 					wbi2 = r5_next_bio(wbi, dev->sector);
2703 					if (!raid5_dec_bi_phys_segments(wbi)) {
2704 						md_write_end(conf->mddev);
2705 						wbi->bi_next = *return_bi;
2706 						*return_bi = wbi;
2707 					}
2708 					wbi = wbi2;
2709 				}
2710 				if (dev->towrite == NULL)
2711 					bitmap_end = 1;
2712 				spin_unlock_irq(&conf->device_lock);
2713 				if (bitmap_end)
2714 					bitmap_endwrite(conf->mddev->bitmap,
2715 							sh->sector,
2716 							STRIPE_SECTORS,
2717 					 !test_bit(STRIPE_DEGRADED, &sh->state),
2718 							0);
2719 			}
2720 		}
2721 
2722 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2723 		if (atomic_dec_and_test(&conf->pending_full_writes))
2724 			md_wakeup_thread(conf->mddev->thread);
2725 }
2726 
2727 static void handle_stripe_dirtying(struct r5conf *conf,
2728 				   struct stripe_head *sh,
2729 				   struct stripe_head_state *s,
2730 				   int disks)
2731 {
2732 	int rmw = 0, rcw = 0, i;
2733 	if (conf->max_degraded == 2) {
2734 		/* RAID6 requires 'rcw' in current implementation
2735 		 * Calculate the real rcw later - for now fake it
2736 		 * look like rcw is cheaper
2737 		 */
2738 		rcw = 1; rmw = 2;
2739 	} else for (i = disks; i--; ) {
2740 		/* would I have to read this buffer for read_modify_write */
2741 		struct r5dev *dev = &sh->dev[i];
2742 		if ((dev->towrite || i == sh->pd_idx) &&
2743 		    !test_bit(R5_LOCKED, &dev->flags) &&
2744 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2745 		      test_bit(R5_Wantcompute, &dev->flags))) {
2746 			if (test_bit(R5_Insync, &dev->flags))
2747 				rmw++;
2748 			else
2749 				rmw += 2*disks;  /* cannot read it */
2750 		}
2751 		/* Would I have to read this buffer for reconstruct_write */
2752 		if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2753 		    !test_bit(R5_LOCKED, &dev->flags) &&
2754 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2755 		    test_bit(R5_Wantcompute, &dev->flags))) {
2756 			if (test_bit(R5_Insync, &dev->flags)) rcw++;
2757 			else
2758 				rcw += 2*disks;
2759 		}
2760 	}
2761 	pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2762 		(unsigned long long)sh->sector, rmw, rcw);
2763 	set_bit(STRIPE_HANDLE, &sh->state);
2764 	if (rmw < rcw && rmw > 0)
2765 		/* prefer read-modify-write, but need to get some data */
2766 		for (i = disks; i--; ) {
2767 			struct r5dev *dev = &sh->dev[i];
2768 			if ((dev->towrite || i == sh->pd_idx) &&
2769 			    !test_bit(R5_LOCKED, &dev->flags) &&
2770 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
2771 			    test_bit(R5_Wantcompute, &dev->flags)) &&
2772 			    test_bit(R5_Insync, &dev->flags)) {
2773 				if (
2774 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2775 					pr_debug("Read_old block "
2776 						"%d for r-m-w\n", i);
2777 					set_bit(R5_LOCKED, &dev->flags);
2778 					set_bit(R5_Wantread, &dev->flags);
2779 					s->locked++;
2780 				} else {
2781 					set_bit(STRIPE_DELAYED, &sh->state);
2782 					set_bit(STRIPE_HANDLE, &sh->state);
2783 				}
2784 			}
2785 		}
2786 	if (rcw <= rmw && rcw > 0) {
2787 		/* want reconstruct write, but need to get some data */
2788 		rcw = 0;
2789 		for (i = disks; i--; ) {
2790 			struct r5dev *dev = &sh->dev[i];
2791 			if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2792 			    i != sh->pd_idx && i != sh->qd_idx &&
2793 			    !test_bit(R5_LOCKED, &dev->flags) &&
2794 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
2795 			      test_bit(R5_Wantcompute, &dev->flags))) {
2796 				rcw++;
2797 				if (!test_bit(R5_Insync, &dev->flags))
2798 					continue; /* it's a failed drive */
2799 				if (
2800 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2801 					pr_debug("Read_old block "
2802 						"%d for Reconstruct\n", i);
2803 					set_bit(R5_LOCKED, &dev->flags);
2804 					set_bit(R5_Wantread, &dev->flags);
2805 					s->locked++;
2806 				} else {
2807 					set_bit(STRIPE_DELAYED, &sh->state);
2808 					set_bit(STRIPE_HANDLE, &sh->state);
2809 				}
2810 			}
2811 		}
2812 	}
2813 	/* now if nothing is locked, and if we have enough data,
2814 	 * we can start a write request
2815 	 */
2816 	/* since handle_stripe can be called at any time we need to handle the
2817 	 * case where a compute block operation has been submitted and then a
2818 	 * subsequent call wants to start a write request.  raid_run_ops only
2819 	 * handles the case where compute block and reconstruct are requested
2820 	 * simultaneously.  If this is not the case then new writes need to be
2821 	 * held off until the compute completes.
2822 	 */
2823 	if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2824 	    (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2825 	    !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2826 		schedule_reconstruction(sh, s, rcw == 0, 0);
2827 }
2828 
2829 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2830 				struct stripe_head_state *s, int disks)
2831 {
2832 	struct r5dev *dev = NULL;
2833 
2834 	set_bit(STRIPE_HANDLE, &sh->state);
2835 
2836 	switch (sh->check_state) {
2837 	case check_state_idle:
2838 		/* start a new check operation if there are no failures */
2839 		if (s->failed == 0) {
2840 			BUG_ON(s->uptodate != disks);
2841 			sh->check_state = check_state_run;
2842 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
2843 			clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2844 			s->uptodate--;
2845 			break;
2846 		}
2847 		dev = &sh->dev[s->failed_num[0]];
2848 		/* fall through */
2849 	case check_state_compute_result:
2850 		sh->check_state = check_state_idle;
2851 		if (!dev)
2852 			dev = &sh->dev[sh->pd_idx];
2853 
2854 		/* check that a write has not made the stripe insync */
2855 		if (test_bit(STRIPE_INSYNC, &sh->state))
2856 			break;
2857 
2858 		/* either failed parity check, or recovery is happening */
2859 		BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2860 		BUG_ON(s->uptodate != disks);
2861 
2862 		set_bit(R5_LOCKED, &dev->flags);
2863 		s->locked++;
2864 		set_bit(R5_Wantwrite, &dev->flags);
2865 
2866 		clear_bit(STRIPE_DEGRADED, &sh->state);
2867 		set_bit(STRIPE_INSYNC, &sh->state);
2868 		break;
2869 	case check_state_run:
2870 		break; /* we will be called again upon completion */
2871 	case check_state_check_result:
2872 		sh->check_state = check_state_idle;
2873 
2874 		/* if a failure occurred during the check operation, leave
2875 		 * STRIPE_INSYNC not set and let the stripe be handled again
2876 		 */
2877 		if (s->failed)
2878 			break;
2879 
2880 		/* handle a successful check operation, if parity is correct
2881 		 * we are done.  Otherwise update the mismatch count and repair
2882 		 * parity if !MD_RECOVERY_CHECK
2883 		 */
2884 		if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2885 			/* parity is correct (on disc,
2886 			 * not in buffer any more)
2887 			 */
2888 			set_bit(STRIPE_INSYNC, &sh->state);
2889 		else {
2890 			conf->mddev->resync_mismatches += STRIPE_SECTORS;
2891 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2892 				/* don't try to repair!! */
2893 				set_bit(STRIPE_INSYNC, &sh->state);
2894 			else {
2895 				sh->check_state = check_state_compute_run;
2896 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2897 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2898 				set_bit(R5_Wantcompute,
2899 					&sh->dev[sh->pd_idx].flags);
2900 				sh->ops.target = sh->pd_idx;
2901 				sh->ops.target2 = -1;
2902 				s->uptodate++;
2903 			}
2904 		}
2905 		break;
2906 	case check_state_compute_run:
2907 		break;
2908 	default:
2909 		printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2910 		       __func__, sh->check_state,
2911 		       (unsigned long long) sh->sector);
2912 		BUG();
2913 	}
2914 }
2915 
2916 
2917 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2918 				  struct stripe_head_state *s,
2919 				  int disks)
2920 {
2921 	int pd_idx = sh->pd_idx;
2922 	int qd_idx = sh->qd_idx;
2923 	struct r5dev *dev;
2924 
2925 	set_bit(STRIPE_HANDLE, &sh->state);
2926 
2927 	BUG_ON(s->failed > 2);
2928 
2929 	/* Want to check and possibly repair P and Q.
2930 	 * However there could be one 'failed' device, in which
2931 	 * case we can only check one of them, possibly using the
2932 	 * other to generate missing data
2933 	 */
2934 
2935 	switch (sh->check_state) {
2936 	case check_state_idle:
2937 		/* start a new check operation if there are < 2 failures */
2938 		if (s->failed == s->q_failed) {
2939 			/* The only possible failed device holds Q, so it
2940 			 * makes sense to check P (If anything else were failed,
2941 			 * we would have used P to recreate it).
2942 			 */
2943 			sh->check_state = check_state_run;
2944 		}
2945 		if (!s->q_failed && s->failed < 2) {
2946 			/* Q is not failed, and we didn't use it to generate
2947 			 * anything, so it makes sense to check it
2948 			 */
2949 			if (sh->check_state == check_state_run)
2950 				sh->check_state = check_state_run_pq;
2951 			else
2952 				sh->check_state = check_state_run_q;
2953 		}
2954 
2955 		/* discard potentially stale zero_sum_result */
2956 		sh->ops.zero_sum_result = 0;
2957 
2958 		if (sh->check_state == check_state_run) {
2959 			/* async_xor_zero_sum destroys the contents of P */
2960 			clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2961 			s->uptodate--;
2962 		}
2963 		if (sh->check_state >= check_state_run &&
2964 		    sh->check_state <= check_state_run_pq) {
2965 			/* async_syndrome_zero_sum preserves P and Q, so
2966 			 * no need to mark them !uptodate here
2967 			 */
2968 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
2969 			break;
2970 		}
2971 
2972 		/* we have 2-disk failure */
2973 		BUG_ON(s->failed != 2);
2974 		/* fall through */
2975 	case check_state_compute_result:
2976 		sh->check_state = check_state_idle;
2977 
2978 		/* check that a write has not made the stripe insync */
2979 		if (test_bit(STRIPE_INSYNC, &sh->state))
2980 			break;
2981 
2982 		/* now write out any block on a failed drive,
2983 		 * or P or Q if they were recomputed
2984 		 */
2985 		BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2986 		if (s->failed == 2) {
2987 			dev = &sh->dev[s->failed_num[1]];
2988 			s->locked++;
2989 			set_bit(R5_LOCKED, &dev->flags);
2990 			set_bit(R5_Wantwrite, &dev->flags);
2991 		}
2992 		if (s->failed >= 1) {
2993 			dev = &sh->dev[s->failed_num[0]];
2994 			s->locked++;
2995 			set_bit(R5_LOCKED, &dev->flags);
2996 			set_bit(R5_Wantwrite, &dev->flags);
2997 		}
2998 		if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2999 			dev = &sh->dev[pd_idx];
3000 			s->locked++;
3001 			set_bit(R5_LOCKED, &dev->flags);
3002 			set_bit(R5_Wantwrite, &dev->flags);
3003 		}
3004 		if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3005 			dev = &sh->dev[qd_idx];
3006 			s->locked++;
3007 			set_bit(R5_LOCKED, &dev->flags);
3008 			set_bit(R5_Wantwrite, &dev->flags);
3009 		}
3010 		clear_bit(STRIPE_DEGRADED, &sh->state);
3011 
3012 		set_bit(STRIPE_INSYNC, &sh->state);
3013 		break;
3014 	case check_state_run:
3015 	case check_state_run_q:
3016 	case check_state_run_pq:
3017 		break; /* we will be called again upon completion */
3018 	case check_state_check_result:
3019 		sh->check_state = check_state_idle;
3020 
3021 		/* handle a successful check operation, if parity is correct
3022 		 * we are done.  Otherwise update the mismatch count and repair
3023 		 * parity if !MD_RECOVERY_CHECK
3024 		 */
3025 		if (sh->ops.zero_sum_result == 0) {
3026 			/* both parities are correct */
3027 			if (!s->failed)
3028 				set_bit(STRIPE_INSYNC, &sh->state);
3029 			else {
3030 				/* in contrast to the raid5 case we can validate
3031 				 * parity, but still have a failure to write
3032 				 * back
3033 				 */
3034 				sh->check_state = check_state_compute_result;
3035 				/* Returning at this point means that we may go
3036 				 * off and bring p and/or q uptodate again so
3037 				 * we make sure to check zero_sum_result again
3038 				 * to verify if p or q need writeback
3039 				 */
3040 			}
3041 		} else {
3042 			conf->mddev->resync_mismatches += STRIPE_SECTORS;
3043 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3044 				/* don't try to repair!! */
3045 				set_bit(STRIPE_INSYNC, &sh->state);
3046 			else {
3047 				int *target = &sh->ops.target;
3048 
3049 				sh->ops.target = -1;
3050 				sh->ops.target2 = -1;
3051 				sh->check_state = check_state_compute_run;
3052 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3053 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3054 				if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3055 					set_bit(R5_Wantcompute,
3056 						&sh->dev[pd_idx].flags);
3057 					*target = pd_idx;
3058 					target = &sh->ops.target2;
3059 					s->uptodate++;
3060 				}
3061 				if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3062 					set_bit(R5_Wantcompute,
3063 						&sh->dev[qd_idx].flags);
3064 					*target = qd_idx;
3065 					s->uptodate++;
3066 				}
3067 			}
3068 		}
3069 		break;
3070 	case check_state_compute_run:
3071 		break;
3072 	default:
3073 		printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3074 		       __func__, sh->check_state,
3075 		       (unsigned long long) sh->sector);
3076 		BUG();
3077 	}
3078 }
3079 
3080 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3081 {
3082 	int i;
3083 
3084 	/* We have read all the blocks in this stripe and now we need to
3085 	 * copy some of them into a target stripe for expand.
3086 	 */
3087 	struct dma_async_tx_descriptor *tx = NULL;
3088 	clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3089 	for (i = 0; i < sh->disks; i++)
3090 		if (i != sh->pd_idx && i != sh->qd_idx) {
3091 			int dd_idx, j;
3092 			struct stripe_head *sh2;
3093 			struct async_submit_ctl submit;
3094 
3095 			sector_t bn = compute_blocknr(sh, i, 1);
3096 			sector_t s = raid5_compute_sector(conf, bn, 0,
3097 							  &dd_idx, NULL);
3098 			sh2 = get_active_stripe(conf, s, 0, 1, 1);
3099 			if (sh2 == NULL)
3100 				/* so far only the early blocks of this stripe
3101 				 * have been requested.  When later blocks
3102 				 * get requested, we will try again
3103 				 */
3104 				continue;
3105 			if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3106 			   test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3107 				/* must have already done this block */
3108 				release_stripe(sh2);
3109 				continue;
3110 			}
3111 
3112 			/* place all the copies on one channel */
3113 			init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3114 			tx = async_memcpy(sh2->dev[dd_idx].page,
3115 					  sh->dev[i].page, 0, 0, STRIPE_SIZE,
3116 					  &submit);
3117 
3118 			set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3119 			set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3120 			for (j = 0; j < conf->raid_disks; j++)
3121 				if (j != sh2->pd_idx &&
3122 				    j != sh2->qd_idx &&
3123 				    !test_bit(R5_Expanded, &sh2->dev[j].flags))
3124 					break;
3125 			if (j == conf->raid_disks) {
3126 				set_bit(STRIPE_EXPAND_READY, &sh2->state);
3127 				set_bit(STRIPE_HANDLE, &sh2->state);
3128 			}
3129 			release_stripe(sh2);
3130 
3131 		}
3132 	/* done submitting copies, wait for them to complete */
3133 	if (tx) {
3134 		async_tx_ack(tx);
3135 		dma_wait_for_async_tx(tx);
3136 	}
3137 }
3138 
3139 /*
3140  * handle_stripe - do things to a stripe.
3141  *
3142  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3143  * state of various bits to see what needs to be done.
3144  * Possible results:
3145  *    return some read requests which now have data
3146  *    return some write requests which are safely on storage
3147  *    schedule a read on some buffers
3148  *    schedule a write of some buffers
3149  *    return confirmation of parity correctness
3150  *
3151  */
3152 
3153 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3154 {
3155 	struct r5conf *conf = sh->raid_conf;
3156 	int disks = sh->disks;
3157 	struct r5dev *dev;
3158 	int i;
3159 	int do_recovery = 0;
3160 
3161 	memset(s, 0, sizeof(*s));
3162 
3163 	s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3164 	s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3165 	s->failed_num[0] = -1;
3166 	s->failed_num[1] = -1;
3167 
3168 	/* Now to look around and see what can be done */
3169 	rcu_read_lock();
3170 	spin_lock_irq(&conf->device_lock);
3171 	for (i=disks; i--; ) {
3172 		struct md_rdev *rdev;
3173 		sector_t first_bad;
3174 		int bad_sectors;
3175 		int is_bad = 0;
3176 
3177 		dev = &sh->dev[i];
3178 
3179 		pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3180 			 i, dev->flags,
3181 			 dev->toread, dev->towrite, dev->written);
3182 		/* maybe we can reply to a read
3183 		 *
3184 		 * new wantfill requests are only permitted while
3185 		 * ops_complete_biofill is guaranteed to be inactive
3186 		 */
3187 		if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3188 		    !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3189 			set_bit(R5_Wantfill, &dev->flags);
3190 
3191 		/* now count some things */
3192 		if (test_bit(R5_LOCKED, &dev->flags))
3193 			s->locked++;
3194 		if (test_bit(R5_UPTODATE, &dev->flags))
3195 			s->uptodate++;
3196 		if (test_bit(R5_Wantcompute, &dev->flags)) {
3197 			s->compute++;
3198 			BUG_ON(s->compute > 2);
3199 		}
3200 
3201 		if (test_bit(R5_Wantfill, &dev->flags))
3202 			s->to_fill++;
3203 		else if (dev->toread)
3204 			s->to_read++;
3205 		if (dev->towrite) {
3206 			s->to_write++;
3207 			if (!test_bit(R5_OVERWRITE, &dev->flags))
3208 				s->non_overwrite++;
3209 		}
3210 		if (dev->written)
3211 			s->written++;
3212 		/* Prefer to use the replacement for reads, but only
3213 		 * if it is recovered enough and has no bad blocks.
3214 		 */
3215 		rdev = rcu_dereference(conf->disks[i].replacement);
3216 		if (rdev && !test_bit(Faulty, &rdev->flags) &&
3217 		    rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3218 		    !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3219 				 &first_bad, &bad_sectors))
3220 			set_bit(R5_ReadRepl, &dev->flags);
3221 		else {
3222 			if (rdev)
3223 				set_bit(R5_NeedReplace, &dev->flags);
3224 			rdev = rcu_dereference(conf->disks[i].rdev);
3225 			clear_bit(R5_ReadRepl, &dev->flags);
3226 		}
3227 		if (rdev && test_bit(Faulty, &rdev->flags))
3228 			rdev = NULL;
3229 		if (rdev) {
3230 			is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3231 					     &first_bad, &bad_sectors);
3232 			if (s->blocked_rdev == NULL
3233 			    && (test_bit(Blocked, &rdev->flags)
3234 				|| is_bad < 0)) {
3235 				if (is_bad < 0)
3236 					set_bit(BlockedBadBlocks,
3237 						&rdev->flags);
3238 				s->blocked_rdev = rdev;
3239 				atomic_inc(&rdev->nr_pending);
3240 			}
3241 		}
3242 		clear_bit(R5_Insync, &dev->flags);
3243 		if (!rdev)
3244 			/* Not in-sync */;
3245 		else if (is_bad) {
3246 			/* also not in-sync */
3247 			if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3248 			    test_bit(R5_UPTODATE, &dev->flags)) {
3249 				/* treat as in-sync, but with a read error
3250 				 * which we can now try to correct
3251 				 */
3252 				set_bit(R5_Insync, &dev->flags);
3253 				set_bit(R5_ReadError, &dev->flags);
3254 			}
3255 		} else if (test_bit(In_sync, &rdev->flags))
3256 			set_bit(R5_Insync, &dev->flags);
3257 		else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3258 			/* in sync if before recovery_offset */
3259 			set_bit(R5_Insync, &dev->flags);
3260 		else if (test_bit(R5_UPTODATE, &dev->flags) &&
3261 			 test_bit(R5_Expanded, &dev->flags))
3262 			/* If we've reshaped into here, we assume it is Insync.
3263 			 * We will shortly update recovery_offset to make
3264 			 * it official.
3265 			 */
3266 			set_bit(R5_Insync, &dev->flags);
3267 
3268 		if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3269 			/* This flag does not apply to '.replacement'
3270 			 * only to .rdev, so make sure to check that*/
3271 			struct md_rdev *rdev2 = rcu_dereference(
3272 				conf->disks[i].rdev);
3273 			if (rdev2 == rdev)
3274 				clear_bit(R5_Insync, &dev->flags);
3275 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3276 				s->handle_bad_blocks = 1;
3277 				atomic_inc(&rdev2->nr_pending);
3278 			} else
3279 				clear_bit(R5_WriteError, &dev->flags);
3280 		}
3281 		if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3282 			/* This flag does not apply to '.replacement'
3283 			 * only to .rdev, so make sure to check that*/
3284 			struct md_rdev *rdev2 = rcu_dereference(
3285 				conf->disks[i].rdev);
3286 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3287 				s->handle_bad_blocks = 1;
3288 				atomic_inc(&rdev2->nr_pending);
3289 			} else
3290 				clear_bit(R5_MadeGood, &dev->flags);
3291 		}
3292 		if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3293 			struct md_rdev *rdev2 = rcu_dereference(
3294 				conf->disks[i].replacement);
3295 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3296 				s->handle_bad_blocks = 1;
3297 				atomic_inc(&rdev2->nr_pending);
3298 			} else
3299 				clear_bit(R5_MadeGoodRepl, &dev->flags);
3300 		}
3301 		if (!test_bit(R5_Insync, &dev->flags)) {
3302 			/* The ReadError flag will just be confusing now */
3303 			clear_bit(R5_ReadError, &dev->flags);
3304 			clear_bit(R5_ReWrite, &dev->flags);
3305 		}
3306 		if (test_bit(R5_ReadError, &dev->flags))
3307 			clear_bit(R5_Insync, &dev->flags);
3308 		if (!test_bit(R5_Insync, &dev->flags)) {
3309 			if (s->failed < 2)
3310 				s->failed_num[s->failed] = i;
3311 			s->failed++;
3312 			if (rdev && !test_bit(Faulty, &rdev->flags))
3313 				do_recovery = 1;
3314 		}
3315 	}
3316 	spin_unlock_irq(&conf->device_lock);
3317 	if (test_bit(STRIPE_SYNCING, &sh->state)) {
3318 		/* If there is a failed device being replaced,
3319 		 *     we must be recovering.
3320 		 * else if we are after recovery_cp, we must be syncing
3321 		 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3322 		 * else we can only be replacing
3323 		 * sync and recovery both need to read all devices, and so
3324 		 * use the same flag.
3325 		 */
3326 		if (do_recovery ||
3327 		    sh->sector >= conf->mddev->recovery_cp ||
3328 		    test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3329 			s->syncing = 1;
3330 		else
3331 			s->replacing = 1;
3332 	}
3333 	rcu_read_unlock();
3334 }
3335 
3336 static void handle_stripe(struct stripe_head *sh)
3337 {
3338 	struct stripe_head_state s;
3339 	struct r5conf *conf = sh->raid_conf;
3340 	int i;
3341 	int prexor;
3342 	int disks = sh->disks;
3343 	struct r5dev *pdev, *qdev;
3344 
3345 	clear_bit(STRIPE_HANDLE, &sh->state);
3346 	if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3347 		/* already being handled, ensure it gets handled
3348 		 * again when current action finishes */
3349 		set_bit(STRIPE_HANDLE, &sh->state);
3350 		return;
3351 	}
3352 
3353 	if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3354 		set_bit(STRIPE_SYNCING, &sh->state);
3355 		clear_bit(STRIPE_INSYNC, &sh->state);
3356 	}
3357 	clear_bit(STRIPE_DELAYED, &sh->state);
3358 
3359 	pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3360 		"pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3361 	       (unsigned long long)sh->sector, sh->state,
3362 	       atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3363 	       sh->check_state, sh->reconstruct_state);
3364 
3365 	analyse_stripe(sh, &s);
3366 
3367 	if (s.handle_bad_blocks) {
3368 		set_bit(STRIPE_HANDLE, &sh->state);
3369 		goto finish;
3370 	}
3371 
3372 	if (unlikely(s.blocked_rdev)) {
3373 		if (s.syncing || s.expanding || s.expanded ||
3374 		    s.replacing || s.to_write || s.written) {
3375 			set_bit(STRIPE_HANDLE, &sh->state);
3376 			goto finish;
3377 		}
3378 		/* There is nothing for the blocked_rdev to block */
3379 		rdev_dec_pending(s.blocked_rdev, conf->mddev);
3380 		s.blocked_rdev = NULL;
3381 	}
3382 
3383 	if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3384 		set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3385 		set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3386 	}
3387 
3388 	pr_debug("locked=%d uptodate=%d to_read=%d"
3389 	       " to_write=%d failed=%d failed_num=%d,%d\n",
3390 	       s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3391 	       s.failed_num[0], s.failed_num[1]);
3392 	/* check if the array has lost more than max_degraded devices and,
3393 	 * if so, some requests might need to be failed.
3394 	 */
3395 	if (s.failed > conf->max_degraded) {
3396 		sh->check_state = 0;
3397 		sh->reconstruct_state = 0;
3398 		if (s.to_read+s.to_write+s.written)
3399 			handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3400 		if (s.syncing + s.replacing)
3401 			handle_failed_sync(conf, sh, &s);
3402 	}
3403 
3404 	/*
3405 	 * might be able to return some write requests if the parity blocks
3406 	 * are safe, or on a failed drive
3407 	 */
3408 	pdev = &sh->dev[sh->pd_idx];
3409 	s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3410 		|| (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3411 	qdev = &sh->dev[sh->qd_idx];
3412 	s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3413 		|| (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3414 		|| conf->level < 6;
3415 
3416 	if (s.written &&
3417 	    (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3418 			     && !test_bit(R5_LOCKED, &pdev->flags)
3419 			     && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3420 	    (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3421 			     && !test_bit(R5_LOCKED, &qdev->flags)
3422 			     && test_bit(R5_UPTODATE, &qdev->flags)))))
3423 		handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3424 
3425 	/* Now we might consider reading some blocks, either to check/generate
3426 	 * parity, or to satisfy requests
3427 	 * or to load a block that is being partially written.
3428 	 */
3429 	if (s.to_read || s.non_overwrite
3430 	    || (conf->level == 6 && s.to_write && s.failed)
3431 	    || (s.syncing && (s.uptodate + s.compute < disks))
3432 	    || s.replacing
3433 	    || s.expanding)
3434 		handle_stripe_fill(sh, &s, disks);
3435 
3436 	/* Now we check to see if any write operations have recently
3437 	 * completed
3438 	 */
3439 	prexor = 0;
3440 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3441 		prexor = 1;
3442 	if (sh->reconstruct_state == reconstruct_state_drain_result ||
3443 	    sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3444 		sh->reconstruct_state = reconstruct_state_idle;
3445 
3446 		/* All the 'written' buffers and the parity block are ready to
3447 		 * be written back to disk
3448 		 */
3449 		BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3450 		BUG_ON(sh->qd_idx >= 0 &&
3451 		       !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3452 		for (i = disks; i--; ) {
3453 			struct r5dev *dev = &sh->dev[i];
3454 			if (test_bit(R5_LOCKED, &dev->flags) &&
3455 				(i == sh->pd_idx || i == sh->qd_idx ||
3456 				 dev->written)) {
3457 				pr_debug("Writing block %d\n", i);
3458 				set_bit(R5_Wantwrite, &dev->flags);
3459 				if (prexor)
3460 					continue;
3461 				if (!test_bit(R5_Insync, &dev->flags) ||
3462 				    ((i == sh->pd_idx || i == sh->qd_idx)  &&
3463 				     s.failed == 0))
3464 					set_bit(STRIPE_INSYNC, &sh->state);
3465 			}
3466 		}
3467 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3468 			s.dec_preread_active = 1;
3469 	}
3470 
3471 	/* Now to consider new write requests and what else, if anything
3472 	 * should be read.  We do not handle new writes when:
3473 	 * 1/ A 'write' operation (copy+xor) is already in flight.
3474 	 * 2/ A 'check' operation is in flight, as it may clobber the parity
3475 	 *    block.
3476 	 */
3477 	if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3478 		handle_stripe_dirtying(conf, sh, &s, disks);
3479 
3480 	/* maybe we need to check and possibly fix the parity for this stripe
3481 	 * Any reads will already have been scheduled, so we just see if enough
3482 	 * data is available.  The parity check is held off while parity
3483 	 * dependent operations are in flight.
3484 	 */
3485 	if (sh->check_state ||
3486 	    (s.syncing && s.locked == 0 &&
3487 	     !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3488 	     !test_bit(STRIPE_INSYNC, &sh->state))) {
3489 		if (conf->level == 6)
3490 			handle_parity_checks6(conf, sh, &s, disks);
3491 		else
3492 			handle_parity_checks5(conf, sh, &s, disks);
3493 	}
3494 
3495 	if (s.replacing && s.locked == 0
3496 	    && !test_bit(STRIPE_INSYNC, &sh->state)) {
3497 		/* Write out to replacement devices where possible */
3498 		for (i = 0; i < conf->raid_disks; i++)
3499 			if (test_bit(R5_UPTODATE, &sh->dev[i].flags) &&
3500 			    test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3501 				set_bit(R5_WantReplace, &sh->dev[i].flags);
3502 				set_bit(R5_LOCKED, &sh->dev[i].flags);
3503 				s.locked++;
3504 			}
3505 		set_bit(STRIPE_INSYNC, &sh->state);
3506 	}
3507 	if ((s.syncing || s.replacing) && s.locked == 0 &&
3508 	    test_bit(STRIPE_INSYNC, &sh->state)) {
3509 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3510 		clear_bit(STRIPE_SYNCING, &sh->state);
3511 	}
3512 
3513 	/* If the failed drives are just a ReadError, then we might need
3514 	 * to progress the repair/check process
3515 	 */
3516 	if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3517 		for (i = 0; i < s.failed; i++) {
3518 			struct r5dev *dev = &sh->dev[s.failed_num[i]];
3519 			if (test_bit(R5_ReadError, &dev->flags)
3520 			    && !test_bit(R5_LOCKED, &dev->flags)
3521 			    && test_bit(R5_UPTODATE, &dev->flags)
3522 				) {
3523 				if (!test_bit(R5_ReWrite, &dev->flags)) {
3524 					set_bit(R5_Wantwrite, &dev->flags);
3525 					set_bit(R5_ReWrite, &dev->flags);
3526 					set_bit(R5_LOCKED, &dev->flags);
3527 					s.locked++;
3528 				} else {
3529 					/* let's read it back */
3530 					set_bit(R5_Wantread, &dev->flags);
3531 					set_bit(R5_LOCKED, &dev->flags);
3532 					s.locked++;
3533 				}
3534 			}
3535 		}
3536 
3537 
3538 	/* Finish reconstruct operations initiated by the expansion process */
3539 	if (sh->reconstruct_state == reconstruct_state_result) {
3540 		struct stripe_head *sh_src
3541 			= get_active_stripe(conf, sh->sector, 1, 1, 1);
3542 		if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3543 			/* sh cannot be written until sh_src has been read.
3544 			 * so arrange for sh to be delayed a little
3545 			 */
3546 			set_bit(STRIPE_DELAYED, &sh->state);
3547 			set_bit(STRIPE_HANDLE, &sh->state);
3548 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3549 					      &sh_src->state))
3550 				atomic_inc(&conf->preread_active_stripes);
3551 			release_stripe(sh_src);
3552 			goto finish;
3553 		}
3554 		if (sh_src)
3555 			release_stripe(sh_src);
3556 
3557 		sh->reconstruct_state = reconstruct_state_idle;
3558 		clear_bit(STRIPE_EXPANDING, &sh->state);
3559 		for (i = conf->raid_disks; i--; ) {
3560 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
3561 			set_bit(R5_LOCKED, &sh->dev[i].flags);
3562 			s.locked++;
3563 		}
3564 	}
3565 
3566 	if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3567 	    !sh->reconstruct_state) {
3568 		/* Need to write out all blocks after computing parity */
3569 		sh->disks = conf->raid_disks;
3570 		stripe_set_idx(sh->sector, conf, 0, sh);
3571 		schedule_reconstruction(sh, &s, 1, 1);
3572 	} else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3573 		clear_bit(STRIPE_EXPAND_READY, &sh->state);
3574 		atomic_dec(&conf->reshape_stripes);
3575 		wake_up(&conf->wait_for_overlap);
3576 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3577 	}
3578 
3579 	if (s.expanding && s.locked == 0 &&
3580 	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3581 		handle_stripe_expansion(conf, sh);
3582 
3583 finish:
3584 	/* wait for this device to become unblocked */
3585 	if (conf->mddev->external && unlikely(s.blocked_rdev))
3586 		md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3587 
3588 	if (s.handle_bad_blocks)
3589 		for (i = disks; i--; ) {
3590 			struct md_rdev *rdev;
3591 			struct r5dev *dev = &sh->dev[i];
3592 			if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3593 				/* We own a safe reference to the rdev */
3594 				rdev = conf->disks[i].rdev;
3595 				if (!rdev_set_badblocks(rdev, sh->sector,
3596 							STRIPE_SECTORS, 0))
3597 					md_error(conf->mddev, rdev);
3598 				rdev_dec_pending(rdev, conf->mddev);
3599 			}
3600 			if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3601 				rdev = conf->disks[i].rdev;
3602 				rdev_clear_badblocks(rdev, sh->sector,
3603 						     STRIPE_SECTORS, 0);
3604 				rdev_dec_pending(rdev, conf->mddev);
3605 			}
3606 			if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3607 				rdev = conf->disks[i].replacement;
3608 				if (!rdev)
3609 					/* rdev have been moved down */
3610 					rdev = conf->disks[i].rdev;
3611 				rdev_clear_badblocks(rdev, sh->sector,
3612 						     STRIPE_SECTORS, 0);
3613 				rdev_dec_pending(rdev, conf->mddev);
3614 			}
3615 		}
3616 
3617 	if (s.ops_request)
3618 		raid_run_ops(sh, s.ops_request);
3619 
3620 	ops_run_io(sh, &s);
3621 
3622 	if (s.dec_preread_active) {
3623 		/* We delay this until after ops_run_io so that if make_request
3624 		 * is waiting on a flush, it won't continue until the writes
3625 		 * have actually been submitted.
3626 		 */
3627 		atomic_dec(&conf->preread_active_stripes);
3628 		if (atomic_read(&conf->preread_active_stripes) <
3629 		    IO_THRESHOLD)
3630 			md_wakeup_thread(conf->mddev->thread);
3631 	}
3632 
3633 	return_io(s.return_bi);
3634 
3635 	clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3636 }
3637 
3638 static void raid5_activate_delayed(struct r5conf *conf)
3639 {
3640 	if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3641 		while (!list_empty(&conf->delayed_list)) {
3642 			struct list_head *l = conf->delayed_list.next;
3643 			struct stripe_head *sh;
3644 			sh = list_entry(l, struct stripe_head, lru);
3645 			list_del_init(l);
3646 			clear_bit(STRIPE_DELAYED, &sh->state);
3647 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3648 				atomic_inc(&conf->preread_active_stripes);
3649 			list_add_tail(&sh->lru, &conf->hold_list);
3650 		}
3651 	}
3652 }
3653 
3654 static void activate_bit_delay(struct r5conf *conf)
3655 {
3656 	/* device_lock is held */
3657 	struct list_head head;
3658 	list_add(&head, &conf->bitmap_list);
3659 	list_del_init(&conf->bitmap_list);
3660 	while (!list_empty(&head)) {
3661 		struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3662 		list_del_init(&sh->lru);
3663 		atomic_inc(&sh->count);
3664 		__release_stripe(conf, sh);
3665 	}
3666 }
3667 
3668 int md_raid5_congested(struct mddev *mddev, int bits)
3669 {
3670 	struct r5conf *conf = mddev->private;
3671 
3672 	/* No difference between reads and writes.  Just check
3673 	 * how busy the stripe_cache is
3674 	 */
3675 
3676 	if (conf->inactive_blocked)
3677 		return 1;
3678 	if (conf->quiesce)
3679 		return 1;
3680 	if (list_empty_careful(&conf->inactive_list))
3681 		return 1;
3682 
3683 	return 0;
3684 }
3685 EXPORT_SYMBOL_GPL(md_raid5_congested);
3686 
3687 static int raid5_congested(void *data, int bits)
3688 {
3689 	struct mddev *mddev = data;
3690 
3691 	return mddev_congested(mddev, bits) ||
3692 		md_raid5_congested(mddev, bits);
3693 }
3694 
3695 /* We want read requests to align with chunks where possible,
3696  * but write requests don't need to.
3697  */
3698 static int raid5_mergeable_bvec(struct request_queue *q,
3699 				struct bvec_merge_data *bvm,
3700 				struct bio_vec *biovec)
3701 {
3702 	struct mddev *mddev = q->queuedata;
3703 	sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3704 	int max;
3705 	unsigned int chunk_sectors = mddev->chunk_sectors;
3706 	unsigned int bio_sectors = bvm->bi_size >> 9;
3707 
3708 	if ((bvm->bi_rw & 1) == WRITE)
3709 		return biovec->bv_len; /* always allow writes to be mergeable */
3710 
3711 	if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3712 		chunk_sectors = mddev->new_chunk_sectors;
3713 	max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3714 	if (max < 0) max = 0;
3715 	if (max <= biovec->bv_len && bio_sectors == 0)
3716 		return biovec->bv_len;
3717 	else
3718 		return max;
3719 }
3720 
3721 
3722 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3723 {
3724 	sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3725 	unsigned int chunk_sectors = mddev->chunk_sectors;
3726 	unsigned int bio_sectors = bio->bi_size >> 9;
3727 
3728 	if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3729 		chunk_sectors = mddev->new_chunk_sectors;
3730 	return  chunk_sectors >=
3731 		((sector & (chunk_sectors - 1)) + bio_sectors);
3732 }
3733 
3734 /*
3735  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3736  *  later sampled by raid5d.
3737  */
3738 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3739 {
3740 	unsigned long flags;
3741 
3742 	spin_lock_irqsave(&conf->device_lock, flags);
3743 
3744 	bi->bi_next = conf->retry_read_aligned_list;
3745 	conf->retry_read_aligned_list = bi;
3746 
3747 	spin_unlock_irqrestore(&conf->device_lock, flags);
3748 	md_wakeup_thread(conf->mddev->thread);
3749 }
3750 
3751 
3752 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3753 {
3754 	struct bio *bi;
3755 
3756 	bi = conf->retry_read_aligned;
3757 	if (bi) {
3758 		conf->retry_read_aligned = NULL;
3759 		return bi;
3760 	}
3761 	bi = conf->retry_read_aligned_list;
3762 	if(bi) {
3763 		conf->retry_read_aligned_list = bi->bi_next;
3764 		bi->bi_next = NULL;
3765 		/*
3766 		 * this sets the active strip count to 1 and the processed
3767 		 * strip count to zero (upper 8 bits)
3768 		 */
3769 		bi->bi_phys_segments = 1; /* biased count of active stripes */
3770 	}
3771 
3772 	return bi;
3773 }
3774 
3775 
3776 /*
3777  *  The "raid5_align_endio" should check if the read succeeded and if it
3778  *  did, call bio_endio on the original bio (having bio_put the new bio
3779  *  first).
3780  *  If the read failed..
3781  */
3782 static void raid5_align_endio(struct bio *bi, int error)
3783 {
3784 	struct bio* raid_bi  = bi->bi_private;
3785 	struct mddev *mddev;
3786 	struct r5conf *conf;
3787 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3788 	struct md_rdev *rdev;
3789 
3790 	bio_put(bi);
3791 
3792 	rdev = (void*)raid_bi->bi_next;
3793 	raid_bi->bi_next = NULL;
3794 	mddev = rdev->mddev;
3795 	conf = mddev->private;
3796 
3797 	rdev_dec_pending(rdev, conf->mddev);
3798 
3799 	if (!error && uptodate) {
3800 		bio_endio(raid_bi, 0);
3801 		if (atomic_dec_and_test(&conf->active_aligned_reads))
3802 			wake_up(&conf->wait_for_stripe);
3803 		return;
3804 	}
3805 
3806 
3807 	pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3808 
3809 	add_bio_to_retry(raid_bi, conf);
3810 }
3811 
3812 static int bio_fits_rdev(struct bio *bi)
3813 {
3814 	struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3815 
3816 	if ((bi->bi_size>>9) > queue_max_sectors(q))
3817 		return 0;
3818 	blk_recount_segments(q, bi);
3819 	if (bi->bi_phys_segments > queue_max_segments(q))
3820 		return 0;
3821 
3822 	if (q->merge_bvec_fn)
3823 		/* it's too hard to apply the merge_bvec_fn at this stage,
3824 		 * just just give up
3825 		 */
3826 		return 0;
3827 
3828 	return 1;
3829 }
3830 
3831 
3832 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3833 {
3834 	struct r5conf *conf = mddev->private;
3835 	int dd_idx;
3836 	struct bio* align_bi;
3837 	struct md_rdev *rdev;
3838 	sector_t end_sector;
3839 
3840 	if (!in_chunk_boundary(mddev, raid_bio)) {
3841 		pr_debug("chunk_aligned_read : non aligned\n");
3842 		return 0;
3843 	}
3844 	/*
3845 	 * use bio_clone_mddev to make a copy of the bio
3846 	 */
3847 	align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3848 	if (!align_bi)
3849 		return 0;
3850 	/*
3851 	 *   set bi_end_io to a new function, and set bi_private to the
3852 	 *     original bio.
3853 	 */
3854 	align_bi->bi_end_io  = raid5_align_endio;
3855 	align_bi->bi_private = raid_bio;
3856 	/*
3857 	 *	compute position
3858 	 */
3859 	align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3860 						    0,
3861 						    &dd_idx, NULL);
3862 
3863 	end_sector = align_bi->bi_sector + (align_bi->bi_size >> 9);
3864 	rcu_read_lock();
3865 	rdev = rcu_dereference(conf->disks[dd_idx].replacement);
3866 	if (!rdev || test_bit(Faulty, &rdev->flags) ||
3867 	    rdev->recovery_offset < end_sector) {
3868 		rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3869 		if (rdev &&
3870 		    (test_bit(Faulty, &rdev->flags) ||
3871 		    !(test_bit(In_sync, &rdev->flags) ||
3872 		      rdev->recovery_offset >= end_sector)))
3873 			rdev = NULL;
3874 	}
3875 	if (rdev) {
3876 		sector_t first_bad;
3877 		int bad_sectors;
3878 
3879 		atomic_inc(&rdev->nr_pending);
3880 		rcu_read_unlock();
3881 		raid_bio->bi_next = (void*)rdev;
3882 		align_bi->bi_bdev =  rdev->bdev;
3883 		align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3884 		/* No reshape active, so we can trust rdev->data_offset */
3885 		align_bi->bi_sector += rdev->data_offset;
3886 
3887 		if (!bio_fits_rdev(align_bi) ||
3888 		    is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3889 				&first_bad, &bad_sectors)) {
3890 			/* too big in some way, or has a known bad block */
3891 			bio_put(align_bi);
3892 			rdev_dec_pending(rdev, mddev);
3893 			return 0;
3894 		}
3895 
3896 		spin_lock_irq(&conf->device_lock);
3897 		wait_event_lock_irq(conf->wait_for_stripe,
3898 				    conf->quiesce == 0,
3899 				    conf->device_lock, /* nothing */);
3900 		atomic_inc(&conf->active_aligned_reads);
3901 		spin_unlock_irq(&conf->device_lock);
3902 
3903 		generic_make_request(align_bi);
3904 		return 1;
3905 	} else {
3906 		rcu_read_unlock();
3907 		bio_put(align_bi);
3908 		return 0;
3909 	}
3910 }
3911 
3912 /* __get_priority_stripe - get the next stripe to process
3913  *
3914  * Full stripe writes are allowed to pass preread active stripes up until
3915  * the bypass_threshold is exceeded.  In general the bypass_count
3916  * increments when the handle_list is handled before the hold_list; however, it
3917  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3918  * stripe with in flight i/o.  The bypass_count will be reset when the
3919  * head of the hold_list has changed, i.e. the head was promoted to the
3920  * handle_list.
3921  */
3922 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3923 {
3924 	struct stripe_head *sh;
3925 
3926 	pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3927 		  __func__,
3928 		  list_empty(&conf->handle_list) ? "empty" : "busy",
3929 		  list_empty(&conf->hold_list) ? "empty" : "busy",
3930 		  atomic_read(&conf->pending_full_writes), conf->bypass_count);
3931 
3932 	if (!list_empty(&conf->handle_list)) {
3933 		sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3934 
3935 		if (list_empty(&conf->hold_list))
3936 			conf->bypass_count = 0;
3937 		else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3938 			if (conf->hold_list.next == conf->last_hold)
3939 				conf->bypass_count++;
3940 			else {
3941 				conf->last_hold = conf->hold_list.next;
3942 				conf->bypass_count -= conf->bypass_threshold;
3943 				if (conf->bypass_count < 0)
3944 					conf->bypass_count = 0;
3945 			}
3946 		}
3947 	} else if (!list_empty(&conf->hold_list) &&
3948 		   ((conf->bypass_threshold &&
3949 		     conf->bypass_count > conf->bypass_threshold) ||
3950 		    atomic_read(&conf->pending_full_writes) == 0)) {
3951 		sh = list_entry(conf->hold_list.next,
3952 				typeof(*sh), lru);
3953 		conf->bypass_count -= conf->bypass_threshold;
3954 		if (conf->bypass_count < 0)
3955 			conf->bypass_count = 0;
3956 	} else
3957 		return NULL;
3958 
3959 	list_del_init(&sh->lru);
3960 	atomic_inc(&sh->count);
3961 	BUG_ON(atomic_read(&sh->count) != 1);
3962 	return sh;
3963 }
3964 
3965 static void make_request(struct mddev *mddev, struct bio * bi)
3966 {
3967 	struct r5conf *conf = mddev->private;
3968 	int dd_idx;
3969 	sector_t new_sector;
3970 	sector_t logical_sector, last_sector;
3971 	struct stripe_head *sh;
3972 	const int rw = bio_data_dir(bi);
3973 	int remaining;
3974 	int plugged;
3975 
3976 	if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3977 		md_flush_request(mddev, bi);
3978 		return;
3979 	}
3980 
3981 	md_write_start(mddev, bi);
3982 
3983 	if (rw == READ &&
3984 	     mddev->reshape_position == MaxSector &&
3985 	     chunk_aligned_read(mddev,bi))
3986 		return;
3987 
3988 	logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3989 	last_sector = bi->bi_sector + (bi->bi_size>>9);
3990 	bi->bi_next = NULL;
3991 	bi->bi_phys_segments = 1;	/* over-loaded to count active stripes */
3992 
3993 	plugged = mddev_check_plugged(mddev);
3994 	for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3995 		DEFINE_WAIT(w);
3996 		int previous;
3997 
3998 	retry:
3999 		previous = 0;
4000 		prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4001 		if (unlikely(conf->reshape_progress != MaxSector)) {
4002 			/* spinlock is needed as reshape_progress may be
4003 			 * 64bit on a 32bit platform, and so it might be
4004 			 * possible to see a half-updated value
4005 			 * Of course reshape_progress could change after
4006 			 * the lock is dropped, so once we get a reference
4007 			 * to the stripe that we think it is, we will have
4008 			 * to check again.
4009 			 */
4010 			spin_lock_irq(&conf->device_lock);
4011 			if (mddev->reshape_backwards
4012 			    ? logical_sector < conf->reshape_progress
4013 			    : logical_sector >= conf->reshape_progress) {
4014 				previous = 1;
4015 			} else {
4016 				if (mddev->reshape_backwards
4017 				    ? logical_sector < conf->reshape_safe
4018 				    : logical_sector >= conf->reshape_safe) {
4019 					spin_unlock_irq(&conf->device_lock);
4020 					schedule();
4021 					goto retry;
4022 				}
4023 			}
4024 			spin_unlock_irq(&conf->device_lock);
4025 		}
4026 
4027 		new_sector = raid5_compute_sector(conf, logical_sector,
4028 						  previous,
4029 						  &dd_idx, NULL);
4030 		pr_debug("raid456: make_request, sector %llu logical %llu\n",
4031 			(unsigned long long)new_sector,
4032 			(unsigned long long)logical_sector);
4033 
4034 		sh = get_active_stripe(conf, new_sector, previous,
4035 				       (bi->bi_rw&RWA_MASK), 0);
4036 		if (sh) {
4037 			if (unlikely(previous)) {
4038 				/* expansion might have moved on while waiting for a
4039 				 * stripe, so we must do the range check again.
4040 				 * Expansion could still move past after this
4041 				 * test, but as we are holding a reference to
4042 				 * 'sh', we know that if that happens,
4043 				 *  STRIPE_EXPANDING will get set and the expansion
4044 				 * won't proceed until we finish with the stripe.
4045 				 */
4046 				int must_retry = 0;
4047 				spin_lock_irq(&conf->device_lock);
4048 				if (mddev->reshape_backwards
4049 				    ? logical_sector >= conf->reshape_progress
4050 				    : logical_sector < conf->reshape_progress)
4051 					/* mismatch, need to try again */
4052 					must_retry = 1;
4053 				spin_unlock_irq(&conf->device_lock);
4054 				if (must_retry) {
4055 					release_stripe(sh);
4056 					schedule();
4057 					goto retry;
4058 				}
4059 			}
4060 
4061 			if (rw == WRITE &&
4062 			    logical_sector >= mddev->suspend_lo &&
4063 			    logical_sector < mddev->suspend_hi) {
4064 				release_stripe(sh);
4065 				/* As the suspend_* range is controlled by
4066 				 * userspace, we want an interruptible
4067 				 * wait.
4068 				 */
4069 				flush_signals(current);
4070 				prepare_to_wait(&conf->wait_for_overlap,
4071 						&w, TASK_INTERRUPTIBLE);
4072 				if (logical_sector >= mddev->suspend_lo &&
4073 				    logical_sector < mddev->suspend_hi)
4074 					schedule();
4075 				goto retry;
4076 			}
4077 
4078 			if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4079 			    !add_stripe_bio(sh, bi, dd_idx, rw)) {
4080 				/* Stripe is busy expanding or
4081 				 * add failed due to overlap.  Flush everything
4082 				 * and wait a while
4083 				 */
4084 				md_wakeup_thread(mddev->thread);
4085 				release_stripe(sh);
4086 				schedule();
4087 				goto retry;
4088 			}
4089 			finish_wait(&conf->wait_for_overlap, &w);
4090 			set_bit(STRIPE_HANDLE, &sh->state);
4091 			clear_bit(STRIPE_DELAYED, &sh->state);
4092 			if ((bi->bi_rw & REQ_SYNC) &&
4093 			    !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4094 				atomic_inc(&conf->preread_active_stripes);
4095 			release_stripe(sh);
4096 		} else {
4097 			/* cannot get stripe for read-ahead, just give-up */
4098 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
4099 			finish_wait(&conf->wait_for_overlap, &w);
4100 			break;
4101 		}
4102 
4103 	}
4104 	if (!plugged)
4105 		md_wakeup_thread(mddev->thread);
4106 
4107 	spin_lock_irq(&conf->device_lock);
4108 	remaining = raid5_dec_bi_phys_segments(bi);
4109 	spin_unlock_irq(&conf->device_lock);
4110 	if (remaining == 0) {
4111 
4112 		if ( rw == WRITE )
4113 			md_write_end(mddev);
4114 
4115 		bio_endio(bi, 0);
4116 	}
4117 }
4118 
4119 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4120 
4121 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4122 {
4123 	/* reshaping is quite different to recovery/resync so it is
4124 	 * handled quite separately ... here.
4125 	 *
4126 	 * On each call to sync_request, we gather one chunk worth of
4127 	 * destination stripes and flag them as expanding.
4128 	 * Then we find all the source stripes and request reads.
4129 	 * As the reads complete, handle_stripe will copy the data
4130 	 * into the destination stripe and release that stripe.
4131 	 */
4132 	struct r5conf *conf = mddev->private;
4133 	struct stripe_head *sh;
4134 	sector_t first_sector, last_sector;
4135 	int raid_disks = conf->previous_raid_disks;
4136 	int data_disks = raid_disks - conf->max_degraded;
4137 	int new_data_disks = conf->raid_disks - conf->max_degraded;
4138 	int i;
4139 	int dd_idx;
4140 	sector_t writepos, readpos, safepos;
4141 	sector_t stripe_addr;
4142 	int reshape_sectors;
4143 	struct list_head stripes;
4144 
4145 	if (sector_nr == 0) {
4146 		/* If restarting in the middle, skip the initial sectors */
4147 		if (mddev->reshape_backwards &&
4148 		    conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4149 			sector_nr = raid5_size(mddev, 0, 0)
4150 				- conf->reshape_progress;
4151 		} else if (!mddev->reshape_backwards &&
4152 			   conf->reshape_progress > 0)
4153 			sector_nr = conf->reshape_progress;
4154 		sector_div(sector_nr, new_data_disks);
4155 		if (sector_nr) {
4156 			mddev->curr_resync_completed = sector_nr;
4157 			sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4158 			*skipped = 1;
4159 			return sector_nr;
4160 		}
4161 	}
4162 
4163 	/* We need to process a full chunk at a time.
4164 	 * If old and new chunk sizes differ, we need to process the
4165 	 * largest of these
4166 	 */
4167 	if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4168 		reshape_sectors = mddev->new_chunk_sectors;
4169 	else
4170 		reshape_sectors = mddev->chunk_sectors;
4171 
4172 	/* We update the metadata at least every 10 seconds, or when
4173 	 * the data about to be copied would over-write the source of
4174 	 * the data at the front of the range.  i.e. one new_stripe
4175 	 * along from reshape_progress new_maps to after where
4176 	 * reshape_safe old_maps to
4177 	 */
4178 	writepos = conf->reshape_progress;
4179 	sector_div(writepos, new_data_disks);
4180 	readpos = conf->reshape_progress;
4181 	sector_div(readpos, data_disks);
4182 	safepos = conf->reshape_safe;
4183 	sector_div(safepos, data_disks);
4184 	if (mddev->reshape_backwards) {
4185 		writepos -= min_t(sector_t, reshape_sectors, writepos);
4186 		readpos += reshape_sectors;
4187 		safepos += reshape_sectors;
4188 	} else {
4189 		writepos += reshape_sectors;
4190 		readpos -= min_t(sector_t, reshape_sectors, readpos);
4191 		safepos -= min_t(sector_t, reshape_sectors, safepos);
4192 	}
4193 
4194 	/* Having calculated the 'writepos' possibly use it
4195 	 * to set 'stripe_addr' which is where we will write to.
4196 	 */
4197 	if (mddev->reshape_backwards) {
4198 		BUG_ON(conf->reshape_progress == 0);
4199 		stripe_addr = writepos;
4200 		BUG_ON((mddev->dev_sectors &
4201 			~((sector_t)reshape_sectors - 1))
4202 		       - reshape_sectors - stripe_addr
4203 		       != sector_nr);
4204 	} else {
4205 		BUG_ON(writepos != sector_nr + reshape_sectors);
4206 		stripe_addr = sector_nr;
4207 	}
4208 
4209 	/* 'writepos' is the most advanced device address we might write.
4210 	 * 'readpos' is the least advanced device address we might read.
4211 	 * 'safepos' is the least address recorded in the metadata as having
4212 	 *     been reshaped.
4213 	 * If there is a min_offset_diff, these are adjusted either by
4214 	 * increasing the safepos/readpos if diff is negative, or
4215 	 * increasing writepos if diff is positive.
4216 	 * If 'readpos' is then behind 'writepos', there is no way that we can
4217 	 * ensure safety in the face of a crash - that must be done by userspace
4218 	 * making a backup of the data.  So in that case there is no particular
4219 	 * rush to update metadata.
4220 	 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4221 	 * update the metadata to advance 'safepos' to match 'readpos' so that
4222 	 * we can be safe in the event of a crash.
4223 	 * So we insist on updating metadata if safepos is behind writepos and
4224 	 * readpos is beyond writepos.
4225 	 * In any case, update the metadata every 10 seconds.
4226 	 * Maybe that number should be configurable, but I'm not sure it is
4227 	 * worth it.... maybe it could be a multiple of safemode_delay???
4228 	 */
4229 	if (conf->min_offset_diff < 0) {
4230 		safepos += -conf->min_offset_diff;
4231 		readpos += -conf->min_offset_diff;
4232 	} else
4233 		writepos += conf->min_offset_diff;
4234 
4235 	if ((mddev->reshape_backwards
4236 	     ? (safepos > writepos && readpos < writepos)
4237 	     : (safepos < writepos && readpos > writepos)) ||
4238 	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4239 		/* Cannot proceed until we've updated the superblock... */
4240 		wait_event(conf->wait_for_overlap,
4241 			   atomic_read(&conf->reshape_stripes)==0);
4242 		mddev->reshape_position = conf->reshape_progress;
4243 		mddev->curr_resync_completed = sector_nr;
4244 		conf->reshape_checkpoint = jiffies;
4245 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
4246 		md_wakeup_thread(mddev->thread);
4247 		wait_event(mddev->sb_wait, mddev->flags == 0 ||
4248 			   kthread_should_stop());
4249 		spin_lock_irq(&conf->device_lock);
4250 		conf->reshape_safe = mddev->reshape_position;
4251 		spin_unlock_irq(&conf->device_lock);
4252 		wake_up(&conf->wait_for_overlap);
4253 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4254 	}
4255 
4256 	INIT_LIST_HEAD(&stripes);
4257 	for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4258 		int j;
4259 		int skipped_disk = 0;
4260 		sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4261 		set_bit(STRIPE_EXPANDING, &sh->state);
4262 		atomic_inc(&conf->reshape_stripes);
4263 		/* If any of this stripe is beyond the end of the old
4264 		 * array, then we need to zero those blocks
4265 		 */
4266 		for (j=sh->disks; j--;) {
4267 			sector_t s;
4268 			if (j == sh->pd_idx)
4269 				continue;
4270 			if (conf->level == 6 &&
4271 			    j == sh->qd_idx)
4272 				continue;
4273 			s = compute_blocknr(sh, j, 0);
4274 			if (s < raid5_size(mddev, 0, 0)) {
4275 				skipped_disk = 1;
4276 				continue;
4277 			}
4278 			memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4279 			set_bit(R5_Expanded, &sh->dev[j].flags);
4280 			set_bit(R5_UPTODATE, &sh->dev[j].flags);
4281 		}
4282 		if (!skipped_disk) {
4283 			set_bit(STRIPE_EXPAND_READY, &sh->state);
4284 			set_bit(STRIPE_HANDLE, &sh->state);
4285 		}
4286 		list_add(&sh->lru, &stripes);
4287 	}
4288 	spin_lock_irq(&conf->device_lock);
4289 	if (mddev->reshape_backwards)
4290 		conf->reshape_progress -= reshape_sectors * new_data_disks;
4291 	else
4292 		conf->reshape_progress += reshape_sectors * new_data_disks;
4293 	spin_unlock_irq(&conf->device_lock);
4294 	/* Ok, those stripe are ready. We can start scheduling
4295 	 * reads on the source stripes.
4296 	 * The source stripes are determined by mapping the first and last
4297 	 * block on the destination stripes.
4298 	 */
4299 	first_sector =
4300 		raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4301 				     1, &dd_idx, NULL);
4302 	last_sector =
4303 		raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4304 					    * new_data_disks - 1),
4305 				     1, &dd_idx, NULL);
4306 	if (last_sector >= mddev->dev_sectors)
4307 		last_sector = mddev->dev_sectors - 1;
4308 	while (first_sector <= last_sector) {
4309 		sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4310 		set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4311 		set_bit(STRIPE_HANDLE, &sh->state);
4312 		release_stripe(sh);
4313 		first_sector += STRIPE_SECTORS;
4314 	}
4315 	/* Now that the sources are clearly marked, we can release
4316 	 * the destination stripes
4317 	 */
4318 	while (!list_empty(&stripes)) {
4319 		sh = list_entry(stripes.next, struct stripe_head, lru);
4320 		list_del_init(&sh->lru);
4321 		release_stripe(sh);
4322 	}
4323 	/* If this takes us to the resync_max point where we have to pause,
4324 	 * then we need to write out the superblock.
4325 	 */
4326 	sector_nr += reshape_sectors;
4327 	if ((sector_nr - mddev->curr_resync_completed) * 2
4328 	    >= mddev->resync_max - mddev->curr_resync_completed) {
4329 		/* Cannot proceed until we've updated the superblock... */
4330 		wait_event(conf->wait_for_overlap,
4331 			   atomic_read(&conf->reshape_stripes) == 0);
4332 		mddev->reshape_position = conf->reshape_progress;
4333 		mddev->curr_resync_completed = sector_nr;
4334 		conf->reshape_checkpoint = jiffies;
4335 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
4336 		md_wakeup_thread(mddev->thread);
4337 		wait_event(mddev->sb_wait,
4338 			   !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4339 			   || kthread_should_stop());
4340 		spin_lock_irq(&conf->device_lock);
4341 		conf->reshape_safe = mddev->reshape_position;
4342 		spin_unlock_irq(&conf->device_lock);
4343 		wake_up(&conf->wait_for_overlap);
4344 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4345 	}
4346 	return reshape_sectors;
4347 }
4348 
4349 /* FIXME go_faster isn't used */
4350 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4351 {
4352 	struct r5conf *conf = mddev->private;
4353 	struct stripe_head *sh;
4354 	sector_t max_sector = mddev->dev_sectors;
4355 	sector_t sync_blocks;
4356 	int still_degraded = 0;
4357 	int i;
4358 
4359 	if (sector_nr >= max_sector) {
4360 		/* just being told to finish up .. nothing much to do */
4361 
4362 		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4363 			end_reshape(conf);
4364 			return 0;
4365 		}
4366 
4367 		if (mddev->curr_resync < max_sector) /* aborted */
4368 			bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4369 					&sync_blocks, 1);
4370 		else /* completed sync */
4371 			conf->fullsync = 0;
4372 		bitmap_close_sync(mddev->bitmap);
4373 
4374 		return 0;
4375 	}
4376 
4377 	/* Allow raid5_quiesce to complete */
4378 	wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4379 
4380 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4381 		return reshape_request(mddev, sector_nr, skipped);
4382 
4383 	/* No need to check resync_max as we never do more than one
4384 	 * stripe, and as resync_max will always be on a chunk boundary,
4385 	 * if the check in md_do_sync didn't fire, there is no chance
4386 	 * of overstepping resync_max here
4387 	 */
4388 
4389 	/* if there is too many failed drives and we are trying
4390 	 * to resync, then assert that we are finished, because there is
4391 	 * nothing we can do.
4392 	 */
4393 	if (mddev->degraded >= conf->max_degraded &&
4394 	    test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4395 		sector_t rv = mddev->dev_sectors - sector_nr;
4396 		*skipped = 1;
4397 		return rv;
4398 	}
4399 	if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4400 	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4401 	    !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4402 		/* we can skip this block, and probably more */
4403 		sync_blocks /= STRIPE_SECTORS;
4404 		*skipped = 1;
4405 		return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4406 	}
4407 
4408 	bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4409 
4410 	sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4411 	if (sh == NULL) {
4412 		sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4413 		/* make sure we don't swamp the stripe cache if someone else
4414 		 * is trying to get access
4415 		 */
4416 		schedule_timeout_uninterruptible(1);
4417 	}
4418 	/* Need to check if array will still be degraded after recovery/resync
4419 	 * We don't need to check the 'failed' flag as when that gets set,
4420 	 * recovery aborts.
4421 	 */
4422 	for (i = 0; i < conf->raid_disks; i++)
4423 		if (conf->disks[i].rdev == NULL)
4424 			still_degraded = 1;
4425 
4426 	bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4427 
4428 	set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4429 
4430 	handle_stripe(sh);
4431 	release_stripe(sh);
4432 
4433 	return STRIPE_SECTORS;
4434 }
4435 
4436 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4437 {
4438 	/* We may not be able to submit a whole bio at once as there
4439 	 * may not be enough stripe_heads available.
4440 	 * We cannot pre-allocate enough stripe_heads as we may need
4441 	 * more than exist in the cache (if we allow ever large chunks).
4442 	 * So we do one stripe head at a time and record in
4443 	 * ->bi_hw_segments how many have been done.
4444 	 *
4445 	 * We *know* that this entire raid_bio is in one chunk, so
4446 	 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4447 	 */
4448 	struct stripe_head *sh;
4449 	int dd_idx;
4450 	sector_t sector, logical_sector, last_sector;
4451 	int scnt = 0;
4452 	int remaining;
4453 	int handled = 0;
4454 
4455 	logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4456 	sector = raid5_compute_sector(conf, logical_sector,
4457 				      0, &dd_idx, NULL);
4458 	last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4459 
4460 	for (; logical_sector < last_sector;
4461 	     logical_sector += STRIPE_SECTORS,
4462 		     sector += STRIPE_SECTORS,
4463 		     scnt++) {
4464 
4465 		if (scnt < raid5_bi_hw_segments(raid_bio))
4466 			/* already done this stripe */
4467 			continue;
4468 
4469 		sh = get_active_stripe(conf, sector, 0, 1, 0);
4470 
4471 		if (!sh) {
4472 			/* failed to get a stripe - must wait */
4473 			raid5_set_bi_hw_segments(raid_bio, scnt);
4474 			conf->retry_read_aligned = raid_bio;
4475 			return handled;
4476 		}
4477 
4478 		if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4479 			release_stripe(sh);
4480 			raid5_set_bi_hw_segments(raid_bio, scnt);
4481 			conf->retry_read_aligned = raid_bio;
4482 			return handled;
4483 		}
4484 
4485 		handle_stripe(sh);
4486 		release_stripe(sh);
4487 		handled++;
4488 	}
4489 	spin_lock_irq(&conf->device_lock);
4490 	remaining = raid5_dec_bi_phys_segments(raid_bio);
4491 	spin_unlock_irq(&conf->device_lock);
4492 	if (remaining == 0)
4493 		bio_endio(raid_bio, 0);
4494 	if (atomic_dec_and_test(&conf->active_aligned_reads))
4495 		wake_up(&conf->wait_for_stripe);
4496 	return handled;
4497 }
4498 
4499 
4500 /*
4501  * This is our raid5 kernel thread.
4502  *
4503  * We scan the hash table for stripes which can be handled now.
4504  * During the scan, completed stripes are saved for us by the interrupt
4505  * handler, so that they will not have to wait for our next wakeup.
4506  */
4507 static void raid5d(struct mddev *mddev)
4508 {
4509 	struct stripe_head *sh;
4510 	struct r5conf *conf = mddev->private;
4511 	int handled;
4512 	struct blk_plug plug;
4513 
4514 	pr_debug("+++ raid5d active\n");
4515 
4516 	md_check_recovery(mddev);
4517 
4518 	blk_start_plug(&plug);
4519 	handled = 0;
4520 	spin_lock_irq(&conf->device_lock);
4521 	while (1) {
4522 		struct bio *bio;
4523 
4524 		if (atomic_read(&mddev->plug_cnt) == 0 &&
4525 		    !list_empty(&conf->bitmap_list)) {
4526 			/* Now is a good time to flush some bitmap updates */
4527 			conf->seq_flush++;
4528 			spin_unlock_irq(&conf->device_lock);
4529 			bitmap_unplug(mddev->bitmap);
4530 			spin_lock_irq(&conf->device_lock);
4531 			conf->seq_write = conf->seq_flush;
4532 			activate_bit_delay(conf);
4533 		}
4534 		if (atomic_read(&mddev->plug_cnt) == 0)
4535 			raid5_activate_delayed(conf);
4536 
4537 		while ((bio = remove_bio_from_retry(conf))) {
4538 			int ok;
4539 			spin_unlock_irq(&conf->device_lock);
4540 			ok = retry_aligned_read(conf, bio);
4541 			spin_lock_irq(&conf->device_lock);
4542 			if (!ok)
4543 				break;
4544 			handled++;
4545 		}
4546 
4547 		sh = __get_priority_stripe(conf);
4548 
4549 		if (!sh)
4550 			break;
4551 		spin_unlock_irq(&conf->device_lock);
4552 
4553 		handled++;
4554 		handle_stripe(sh);
4555 		release_stripe(sh);
4556 		cond_resched();
4557 
4558 		if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4559 			md_check_recovery(mddev);
4560 
4561 		spin_lock_irq(&conf->device_lock);
4562 	}
4563 	pr_debug("%d stripes handled\n", handled);
4564 
4565 	spin_unlock_irq(&conf->device_lock);
4566 
4567 	async_tx_issue_pending_all();
4568 	blk_finish_plug(&plug);
4569 
4570 	pr_debug("--- raid5d inactive\n");
4571 }
4572 
4573 static ssize_t
4574 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4575 {
4576 	struct r5conf *conf = mddev->private;
4577 	if (conf)
4578 		return sprintf(page, "%d\n", conf->max_nr_stripes);
4579 	else
4580 		return 0;
4581 }
4582 
4583 int
4584 raid5_set_cache_size(struct mddev *mddev, int size)
4585 {
4586 	struct r5conf *conf = mddev->private;
4587 	int err;
4588 
4589 	if (size <= 16 || size > 32768)
4590 		return -EINVAL;
4591 	while (size < conf->max_nr_stripes) {
4592 		if (drop_one_stripe(conf))
4593 			conf->max_nr_stripes--;
4594 		else
4595 			break;
4596 	}
4597 	err = md_allow_write(mddev);
4598 	if (err)
4599 		return err;
4600 	while (size > conf->max_nr_stripes) {
4601 		if (grow_one_stripe(conf))
4602 			conf->max_nr_stripes++;
4603 		else break;
4604 	}
4605 	return 0;
4606 }
4607 EXPORT_SYMBOL(raid5_set_cache_size);
4608 
4609 static ssize_t
4610 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4611 {
4612 	struct r5conf *conf = mddev->private;
4613 	unsigned long new;
4614 	int err;
4615 
4616 	if (len >= PAGE_SIZE)
4617 		return -EINVAL;
4618 	if (!conf)
4619 		return -ENODEV;
4620 
4621 	if (strict_strtoul(page, 10, &new))
4622 		return -EINVAL;
4623 	err = raid5_set_cache_size(mddev, new);
4624 	if (err)
4625 		return err;
4626 	return len;
4627 }
4628 
4629 static struct md_sysfs_entry
4630 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4631 				raid5_show_stripe_cache_size,
4632 				raid5_store_stripe_cache_size);
4633 
4634 static ssize_t
4635 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4636 {
4637 	struct r5conf *conf = mddev->private;
4638 	if (conf)
4639 		return sprintf(page, "%d\n", conf->bypass_threshold);
4640 	else
4641 		return 0;
4642 }
4643 
4644 static ssize_t
4645 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4646 {
4647 	struct r5conf *conf = mddev->private;
4648 	unsigned long new;
4649 	if (len >= PAGE_SIZE)
4650 		return -EINVAL;
4651 	if (!conf)
4652 		return -ENODEV;
4653 
4654 	if (strict_strtoul(page, 10, &new))
4655 		return -EINVAL;
4656 	if (new > conf->max_nr_stripes)
4657 		return -EINVAL;
4658 	conf->bypass_threshold = new;
4659 	return len;
4660 }
4661 
4662 static struct md_sysfs_entry
4663 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4664 					S_IRUGO | S_IWUSR,
4665 					raid5_show_preread_threshold,
4666 					raid5_store_preread_threshold);
4667 
4668 static ssize_t
4669 stripe_cache_active_show(struct mddev *mddev, char *page)
4670 {
4671 	struct r5conf *conf = mddev->private;
4672 	if (conf)
4673 		return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4674 	else
4675 		return 0;
4676 }
4677 
4678 static struct md_sysfs_entry
4679 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4680 
4681 static struct attribute *raid5_attrs[] =  {
4682 	&raid5_stripecache_size.attr,
4683 	&raid5_stripecache_active.attr,
4684 	&raid5_preread_bypass_threshold.attr,
4685 	NULL,
4686 };
4687 static struct attribute_group raid5_attrs_group = {
4688 	.name = NULL,
4689 	.attrs = raid5_attrs,
4690 };
4691 
4692 static sector_t
4693 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4694 {
4695 	struct r5conf *conf = mddev->private;
4696 
4697 	if (!sectors)
4698 		sectors = mddev->dev_sectors;
4699 	if (!raid_disks)
4700 		/* size is defined by the smallest of previous and new size */
4701 		raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4702 
4703 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4704 	sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4705 	return sectors * (raid_disks - conf->max_degraded);
4706 }
4707 
4708 static void raid5_free_percpu(struct r5conf *conf)
4709 {
4710 	struct raid5_percpu *percpu;
4711 	unsigned long cpu;
4712 
4713 	if (!conf->percpu)
4714 		return;
4715 
4716 	get_online_cpus();
4717 	for_each_possible_cpu(cpu) {
4718 		percpu = per_cpu_ptr(conf->percpu, cpu);
4719 		safe_put_page(percpu->spare_page);
4720 		kfree(percpu->scribble);
4721 	}
4722 #ifdef CONFIG_HOTPLUG_CPU
4723 	unregister_cpu_notifier(&conf->cpu_notify);
4724 #endif
4725 	put_online_cpus();
4726 
4727 	free_percpu(conf->percpu);
4728 }
4729 
4730 static void free_conf(struct r5conf *conf)
4731 {
4732 	shrink_stripes(conf);
4733 	raid5_free_percpu(conf);
4734 	kfree(conf->disks);
4735 	kfree(conf->stripe_hashtbl);
4736 	kfree(conf);
4737 }
4738 
4739 #ifdef CONFIG_HOTPLUG_CPU
4740 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4741 			      void *hcpu)
4742 {
4743 	struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4744 	long cpu = (long)hcpu;
4745 	struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4746 
4747 	switch (action) {
4748 	case CPU_UP_PREPARE:
4749 	case CPU_UP_PREPARE_FROZEN:
4750 		if (conf->level == 6 && !percpu->spare_page)
4751 			percpu->spare_page = alloc_page(GFP_KERNEL);
4752 		if (!percpu->scribble)
4753 			percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4754 
4755 		if (!percpu->scribble ||
4756 		    (conf->level == 6 && !percpu->spare_page)) {
4757 			safe_put_page(percpu->spare_page);
4758 			kfree(percpu->scribble);
4759 			pr_err("%s: failed memory allocation for cpu%ld\n",
4760 			       __func__, cpu);
4761 			return notifier_from_errno(-ENOMEM);
4762 		}
4763 		break;
4764 	case CPU_DEAD:
4765 	case CPU_DEAD_FROZEN:
4766 		safe_put_page(percpu->spare_page);
4767 		kfree(percpu->scribble);
4768 		percpu->spare_page = NULL;
4769 		percpu->scribble = NULL;
4770 		break;
4771 	default:
4772 		break;
4773 	}
4774 	return NOTIFY_OK;
4775 }
4776 #endif
4777 
4778 static int raid5_alloc_percpu(struct r5conf *conf)
4779 {
4780 	unsigned long cpu;
4781 	struct page *spare_page;
4782 	struct raid5_percpu __percpu *allcpus;
4783 	void *scribble;
4784 	int err;
4785 
4786 	allcpus = alloc_percpu(struct raid5_percpu);
4787 	if (!allcpus)
4788 		return -ENOMEM;
4789 	conf->percpu = allcpus;
4790 
4791 	get_online_cpus();
4792 	err = 0;
4793 	for_each_present_cpu(cpu) {
4794 		if (conf->level == 6) {
4795 			spare_page = alloc_page(GFP_KERNEL);
4796 			if (!spare_page) {
4797 				err = -ENOMEM;
4798 				break;
4799 			}
4800 			per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4801 		}
4802 		scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4803 		if (!scribble) {
4804 			err = -ENOMEM;
4805 			break;
4806 		}
4807 		per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4808 	}
4809 #ifdef CONFIG_HOTPLUG_CPU
4810 	conf->cpu_notify.notifier_call = raid456_cpu_notify;
4811 	conf->cpu_notify.priority = 0;
4812 	if (err == 0)
4813 		err = register_cpu_notifier(&conf->cpu_notify);
4814 #endif
4815 	put_online_cpus();
4816 
4817 	return err;
4818 }
4819 
4820 static struct r5conf *setup_conf(struct mddev *mddev)
4821 {
4822 	struct r5conf *conf;
4823 	int raid_disk, memory, max_disks;
4824 	struct md_rdev *rdev;
4825 	struct disk_info *disk;
4826 
4827 	if (mddev->new_level != 5
4828 	    && mddev->new_level != 4
4829 	    && mddev->new_level != 6) {
4830 		printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4831 		       mdname(mddev), mddev->new_level);
4832 		return ERR_PTR(-EIO);
4833 	}
4834 	if ((mddev->new_level == 5
4835 	     && !algorithm_valid_raid5(mddev->new_layout)) ||
4836 	    (mddev->new_level == 6
4837 	     && !algorithm_valid_raid6(mddev->new_layout))) {
4838 		printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4839 		       mdname(mddev), mddev->new_layout);
4840 		return ERR_PTR(-EIO);
4841 	}
4842 	if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4843 		printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4844 		       mdname(mddev), mddev->raid_disks);
4845 		return ERR_PTR(-EINVAL);
4846 	}
4847 
4848 	if (!mddev->new_chunk_sectors ||
4849 	    (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4850 	    !is_power_of_2(mddev->new_chunk_sectors)) {
4851 		printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4852 		       mdname(mddev), mddev->new_chunk_sectors << 9);
4853 		return ERR_PTR(-EINVAL);
4854 	}
4855 
4856 	conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4857 	if (conf == NULL)
4858 		goto abort;
4859 	spin_lock_init(&conf->device_lock);
4860 	init_waitqueue_head(&conf->wait_for_stripe);
4861 	init_waitqueue_head(&conf->wait_for_overlap);
4862 	INIT_LIST_HEAD(&conf->handle_list);
4863 	INIT_LIST_HEAD(&conf->hold_list);
4864 	INIT_LIST_HEAD(&conf->delayed_list);
4865 	INIT_LIST_HEAD(&conf->bitmap_list);
4866 	INIT_LIST_HEAD(&conf->inactive_list);
4867 	atomic_set(&conf->active_stripes, 0);
4868 	atomic_set(&conf->preread_active_stripes, 0);
4869 	atomic_set(&conf->active_aligned_reads, 0);
4870 	conf->bypass_threshold = BYPASS_THRESHOLD;
4871 	conf->recovery_disabled = mddev->recovery_disabled - 1;
4872 
4873 	conf->raid_disks = mddev->raid_disks;
4874 	if (mddev->reshape_position == MaxSector)
4875 		conf->previous_raid_disks = mddev->raid_disks;
4876 	else
4877 		conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4878 	max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4879 	conf->scribble_len = scribble_len(max_disks);
4880 
4881 	conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4882 			      GFP_KERNEL);
4883 	if (!conf->disks)
4884 		goto abort;
4885 
4886 	conf->mddev = mddev;
4887 
4888 	if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4889 		goto abort;
4890 
4891 	conf->level = mddev->new_level;
4892 	if (raid5_alloc_percpu(conf) != 0)
4893 		goto abort;
4894 
4895 	pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4896 
4897 	rdev_for_each(rdev, mddev) {
4898 		raid_disk = rdev->raid_disk;
4899 		if (raid_disk >= max_disks
4900 		    || raid_disk < 0)
4901 			continue;
4902 		disk = conf->disks + raid_disk;
4903 
4904 		if (test_bit(Replacement, &rdev->flags)) {
4905 			if (disk->replacement)
4906 				goto abort;
4907 			disk->replacement = rdev;
4908 		} else {
4909 			if (disk->rdev)
4910 				goto abort;
4911 			disk->rdev = rdev;
4912 		}
4913 
4914 		if (test_bit(In_sync, &rdev->flags)) {
4915 			char b[BDEVNAME_SIZE];
4916 			printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4917 			       " disk %d\n",
4918 			       mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4919 		} else if (rdev->saved_raid_disk != raid_disk)
4920 			/* Cannot rely on bitmap to complete recovery */
4921 			conf->fullsync = 1;
4922 	}
4923 
4924 	conf->chunk_sectors = mddev->new_chunk_sectors;
4925 	conf->level = mddev->new_level;
4926 	if (conf->level == 6)
4927 		conf->max_degraded = 2;
4928 	else
4929 		conf->max_degraded = 1;
4930 	conf->algorithm = mddev->new_layout;
4931 	conf->max_nr_stripes = NR_STRIPES;
4932 	conf->reshape_progress = mddev->reshape_position;
4933 	if (conf->reshape_progress != MaxSector) {
4934 		conf->prev_chunk_sectors = mddev->chunk_sectors;
4935 		conf->prev_algo = mddev->layout;
4936 	}
4937 
4938 	memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4939 		 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4940 	if (grow_stripes(conf, conf->max_nr_stripes)) {
4941 		printk(KERN_ERR
4942 		       "md/raid:%s: couldn't allocate %dkB for buffers\n",
4943 		       mdname(mddev), memory);
4944 		goto abort;
4945 	} else
4946 		printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4947 		       mdname(mddev), memory);
4948 
4949 	conf->thread = md_register_thread(raid5d, mddev, NULL);
4950 	if (!conf->thread) {
4951 		printk(KERN_ERR
4952 		       "md/raid:%s: couldn't allocate thread.\n",
4953 		       mdname(mddev));
4954 		goto abort;
4955 	}
4956 
4957 	return conf;
4958 
4959  abort:
4960 	if (conf) {
4961 		free_conf(conf);
4962 		return ERR_PTR(-EIO);
4963 	} else
4964 		return ERR_PTR(-ENOMEM);
4965 }
4966 
4967 
4968 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4969 {
4970 	switch (algo) {
4971 	case ALGORITHM_PARITY_0:
4972 		if (raid_disk < max_degraded)
4973 			return 1;
4974 		break;
4975 	case ALGORITHM_PARITY_N:
4976 		if (raid_disk >= raid_disks - max_degraded)
4977 			return 1;
4978 		break;
4979 	case ALGORITHM_PARITY_0_6:
4980 		if (raid_disk == 0 ||
4981 		    raid_disk == raid_disks - 1)
4982 			return 1;
4983 		break;
4984 	case ALGORITHM_LEFT_ASYMMETRIC_6:
4985 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
4986 	case ALGORITHM_LEFT_SYMMETRIC_6:
4987 	case ALGORITHM_RIGHT_SYMMETRIC_6:
4988 		if (raid_disk == raid_disks - 1)
4989 			return 1;
4990 	}
4991 	return 0;
4992 }
4993 
4994 static int run(struct mddev *mddev)
4995 {
4996 	struct r5conf *conf;
4997 	int working_disks = 0;
4998 	int dirty_parity_disks = 0;
4999 	struct md_rdev *rdev;
5000 	sector_t reshape_offset = 0;
5001 	int i;
5002 	long long min_offset_diff = 0;
5003 	int first = 1;
5004 
5005 	if (mddev->recovery_cp != MaxSector)
5006 		printk(KERN_NOTICE "md/raid:%s: not clean"
5007 		       " -- starting background reconstruction\n",
5008 		       mdname(mddev));
5009 
5010 	rdev_for_each(rdev, mddev) {
5011 		long long diff;
5012 		if (rdev->raid_disk < 0)
5013 			continue;
5014 		diff = (rdev->new_data_offset - rdev->data_offset);
5015 		if (first) {
5016 			min_offset_diff = diff;
5017 			first = 0;
5018 		} else if (mddev->reshape_backwards &&
5019 			 diff < min_offset_diff)
5020 			min_offset_diff = diff;
5021 		else if (!mddev->reshape_backwards &&
5022 			 diff > min_offset_diff)
5023 			min_offset_diff = diff;
5024 	}
5025 
5026 	if (mddev->reshape_position != MaxSector) {
5027 		/* Check that we can continue the reshape.
5028 		 * Difficulties arise if the stripe we would write to
5029 		 * next is at or after the stripe we would read from next.
5030 		 * For a reshape that changes the number of devices, this
5031 		 * is only possible for a very short time, and mdadm makes
5032 		 * sure that time appears to have past before assembling
5033 		 * the array.  So we fail if that time hasn't passed.
5034 		 * For a reshape that keeps the number of devices the same
5035 		 * mdadm must be monitoring the reshape can keeping the
5036 		 * critical areas read-only and backed up.  It will start
5037 		 * the array in read-only mode, so we check for that.
5038 		 */
5039 		sector_t here_new, here_old;
5040 		int old_disks;
5041 		int max_degraded = (mddev->level == 6 ? 2 : 1);
5042 
5043 		if (mddev->new_level != mddev->level) {
5044 			printk(KERN_ERR "md/raid:%s: unsupported reshape "
5045 			       "required - aborting.\n",
5046 			       mdname(mddev));
5047 			return -EINVAL;
5048 		}
5049 		old_disks = mddev->raid_disks - mddev->delta_disks;
5050 		/* reshape_position must be on a new-stripe boundary, and one
5051 		 * further up in new geometry must map after here in old
5052 		 * geometry.
5053 		 */
5054 		here_new = mddev->reshape_position;
5055 		if (sector_div(here_new, mddev->new_chunk_sectors *
5056 			       (mddev->raid_disks - max_degraded))) {
5057 			printk(KERN_ERR "md/raid:%s: reshape_position not "
5058 			       "on a stripe boundary\n", mdname(mddev));
5059 			return -EINVAL;
5060 		}
5061 		reshape_offset = here_new * mddev->new_chunk_sectors;
5062 		/* here_new is the stripe we will write to */
5063 		here_old = mddev->reshape_position;
5064 		sector_div(here_old, mddev->chunk_sectors *
5065 			   (old_disks-max_degraded));
5066 		/* here_old is the first stripe that we might need to read
5067 		 * from */
5068 		if (mddev->delta_disks == 0) {
5069 			if ((here_new * mddev->new_chunk_sectors !=
5070 			     here_old * mddev->chunk_sectors)) {
5071 				printk(KERN_ERR "md/raid:%s: reshape position is"
5072 				       " confused - aborting\n", mdname(mddev));
5073 				return -EINVAL;
5074 			}
5075 			/* We cannot be sure it is safe to start an in-place
5076 			 * reshape.  It is only safe if user-space is monitoring
5077 			 * and taking constant backups.
5078 			 * mdadm always starts a situation like this in
5079 			 * readonly mode so it can take control before
5080 			 * allowing any writes.  So just check for that.
5081 			 */
5082 			if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5083 			    abs(min_offset_diff) >= mddev->new_chunk_sectors)
5084 				/* not really in-place - so OK */;
5085 			else if (mddev->ro == 0) {
5086 				printk(KERN_ERR "md/raid:%s: in-place reshape "
5087 				       "must be started in read-only mode "
5088 				       "- aborting\n",
5089 				       mdname(mddev));
5090 				return -EINVAL;
5091 			}
5092 		} else if (mddev->reshape_backwards
5093 		    ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5094 		       here_old * mddev->chunk_sectors)
5095 		    : (here_new * mddev->new_chunk_sectors >=
5096 		       here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5097 			/* Reading from the same stripe as writing to - bad */
5098 			printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5099 			       "auto-recovery - aborting.\n",
5100 			       mdname(mddev));
5101 			return -EINVAL;
5102 		}
5103 		printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5104 		       mdname(mddev));
5105 		/* OK, we should be able to continue; */
5106 	} else {
5107 		BUG_ON(mddev->level != mddev->new_level);
5108 		BUG_ON(mddev->layout != mddev->new_layout);
5109 		BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5110 		BUG_ON(mddev->delta_disks != 0);
5111 	}
5112 
5113 	if (mddev->private == NULL)
5114 		conf = setup_conf(mddev);
5115 	else
5116 		conf = mddev->private;
5117 
5118 	if (IS_ERR(conf))
5119 		return PTR_ERR(conf);
5120 
5121 	conf->min_offset_diff = min_offset_diff;
5122 	mddev->thread = conf->thread;
5123 	conf->thread = NULL;
5124 	mddev->private = conf;
5125 
5126 	for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5127 	     i++) {
5128 		rdev = conf->disks[i].rdev;
5129 		if (!rdev && conf->disks[i].replacement) {
5130 			/* The replacement is all we have yet */
5131 			rdev = conf->disks[i].replacement;
5132 			conf->disks[i].replacement = NULL;
5133 			clear_bit(Replacement, &rdev->flags);
5134 			conf->disks[i].rdev = rdev;
5135 		}
5136 		if (!rdev)
5137 			continue;
5138 		if (conf->disks[i].replacement &&
5139 		    conf->reshape_progress != MaxSector) {
5140 			/* replacements and reshape simply do not mix. */
5141 			printk(KERN_ERR "md: cannot handle concurrent "
5142 			       "replacement and reshape.\n");
5143 			goto abort;
5144 		}
5145 		if (test_bit(In_sync, &rdev->flags)) {
5146 			working_disks++;
5147 			continue;
5148 		}
5149 		/* This disc is not fully in-sync.  However if it
5150 		 * just stored parity (beyond the recovery_offset),
5151 		 * when we don't need to be concerned about the
5152 		 * array being dirty.
5153 		 * When reshape goes 'backwards', we never have
5154 		 * partially completed devices, so we only need
5155 		 * to worry about reshape going forwards.
5156 		 */
5157 		/* Hack because v0.91 doesn't store recovery_offset properly. */
5158 		if (mddev->major_version == 0 &&
5159 		    mddev->minor_version > 90)
5160 			rdev->recovery_offset = reshape_offset;
5161 
5162 		if (rdev->recovery_offset < reshape_offset) {
5163 			/* We need to check old and new layout */
5164 			if (!only_parity(rdev->raid_disk,
5165 					 conf->algorithm,
5166 					 conf->raid_disks,
5167 					 conf->max_degraded))
5168 				continue;
5169 		}
5170 		if (!only_parity(rdev->raid_disk,
5171 				 conf->prev_algo,
5172 				 conf->previous_raid_disks,
5173 				 conf->max_degraded))
5174 			continue;
5175 		dirty_parity_disks++;
5176 	}
5177 
5178 	/*
5179 	 * 0 for a fully functional array, 1 or 2 for a degraded array.
5180 	 */
5181 	mddev->degraded = calc_degraded(conf);
5182 
5183 	if (has_failed(conf)) {
5184 		printk(KERN_ERR "md/raid:%s: not enough operational devices"
5185 			" (%d/%d failed)\n",
5186 			mdname(mddev), mddev->degraded, conf->raid_disks);
5187 		goto abort;
5188 	}
5189 
5190 	/* device size must be a multiple of chunk size */
5191 	mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5192 	mddev->resync_max_sectors = mddev->dev_sectors;
5193 
5194 	if (mddev->degraded > dirty_parity_disks &&
5195 	    mddev->recovery_cp != MaxSector) {
5196 		if (mddev->ok_start_degraded)
5197 			printk(KERN_WARNING
5198 			       "md/raid:%s: starting dirty degraded array"
5199 			       " - data corruption possible.\n",
5200 			       mdname(mddev));
5201 		else {
5202 			printk(KERN_ERR
5203 			       "md/raid:%s: cannot start dirty degraded array.\n",
5204 			       mdname(mddev));
5205 			goto abort;
5206 		}
5207 	}
5208 
5209 	if (mddev->degraded == 0)
5210 		printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5211 		       " devices, algorithm %d\n", mdname(mddev), conf->level,
5212 		       mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5213 		       mddev->new_layout);
5214 	else
5215 		printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5216 		       " out of %d devices, algorithm %d\n",
5217 		       mdname(mddev), conf->level,
5218 		       mddev->raid_disks - mddev->degraded,
5219 		       mddev->raid_disks, mddev->new_layout);
5220 
5221 	print_raid5_conf(conf);
5222 
5223 	if (conf->reshape_progress != MaxSector) {
5224 		conf->reshape_safe = conf->reshape_progress;
5225 		atomic_set(&conf->reshape_stripes, 0);
5226 		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5227 		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5228 		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5229 		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5230 		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5231 							"reshape");
5232 	}
5233 
5234 
5235 	/* Ok, everything is just fine now */
5236 	if (mddev->to_remove == &raid5_attrs_group)
5237 		mddev->to_remove = NULL;
5238 	else if (mddev->kobj.sd &&
5239 	    sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5240 		printk(KERN_WARNING
5241 		       "raid5: failed to create sysfs attributes for %s\n",
5242 		       mdname(mddev));
5243 	md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5244 
5245 	if (mddev->queue) {
5246 		int chunk_size;
5247 		/* read-ahead size must cover two whole stripes, which
5248 		 * is 2 * (datadisks) * chunksize where 'n' is the
5249 		 * number of raid devices
5250 		 */
5251 		int data_disks = conf->previous_raid_disks - conf->max_degraded;
5252 		int stripe = data_disks *
5253 			((mddev->chunk_sectors << 9) / PAGE_SIZE);
5254 		if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5255 			mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5256 
5257 		blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5258 
5259 		mddev->queue->backing_dev_info.congested_data = mddev;
5260 		mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5261 
5262 		chunk_size = mddev->chunk_sectors << 9;
5263 		blk_queue_io_min(mddev->queue, chunk_size);
5264 		blk_queue_io_opt(mddev->queue, chunk_size *
5265 				 (conf->raid_disks - conf->max_degraded));
5266 
5267 		rdev_for_each(rdev, mddev) {
5268 			disk_stack_limits(mddev->gendisk, rdev->bdev,
5269 					  rdev->data_offset << 9);
5270 			disk_stack_limits(mddev->gendisk, rdev->bdev,
5271 					  rdev->new_data_offset << 9);
5272 		}
5273 	}
5274 
5275 	return 0;
5276 abort:
5277 	md_unregister_thread(&mddev->thread);
5278 	print_raid5_conf(conf);
5279 	free_conf(conf);
5280 	mddev->private = NULL;
5281 	printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5282 	return -EIO;
5283 }
5284 
5285 static int stop(struct mddev *mddev)
5286 {
5287 	struct r5conf *conf = mddev->private;
5288 
5289 	md_unregister_thread(&mddev->thread);
5290 	if (mddev->queue)
5291 		mddev->queue->backing_dev_info.congested_fn = NULL;
5292 	free_conf(conf);
5293 	mddev->private = NULL;
5294 	mddev->to_remove = &raid5_attrs_group;
5295 	return 0;
5296 }
5297 
5298 static void status(struct seq_file *seq, struct mddev *mddev)
5299 {
5300 	struct r5conf *conf = mddev->private;
5301 	int i;
5302 
5303 	seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5304 		mddev->chunk_sectors / 2, mddev->layout);
5305 	seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5306 	for (i = 0; i < conf->raid_disks; i++)
5307 		seq_printf (seq, "%s",
5308 			       conf->disks[i].rdev &&
5309 			       test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5310 	seq_printf (seq, "]");
5311 }
5312 
5313 static void print_raid5_conf (struct r5conf *conf)
5314 {
5315 	int i;
5316 	struct disk_info *tmp;
5317 
5318 	printk(KERN_DEBUG "RAID conf printout:\n");
5319 	if (!conf) {
5320 		printk("(conf==NULL)\n");
5321 		return;
5322 	}
5323 	printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5324 	       conf->raid_disks,
5325 	       conf->raid_disks - conf->mddev->degraded);
5326 
5327 	for (i = 0; i < conf->raid_disks; i++) {
5328 		char b[BDEVNAME_SIZE];
5329 		tmp = conf->disks + i;
5330 		if (tmp->rdev)
5331 			printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
5332 			       i, !test_bit(Faulty, &tmp->rdev->flags),
5333 			       bdevname(tmp->rdev->bdev, b));
5334 	}
5335 }
5336 
5337 static int raid5_spare_active(struct mddev *mddev)
5338 {
5339 	int i;
5340 	struct r5conf *conf = mddev->private;
5341 	struct disk_info *tmp;
5342 	int count = 0;
5343 	unsigned long flags;
5344 
5345 	for (i = 0; i < conf->raid_disks; i++) {
5346 		tmp = conf->disks + i;
5347 		if (tmp->replacement
5348 		    && tmp->replacement->recovery_offset == MaxSector
5349 		    && !test_bit(Faulty, &tmp->replacement->flags)
5350 		    && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
5351 			/* Replacement has just become active. */
5352 			if (!tmp->rdev
5353 			    || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
5354 				count++;
5355 			if (tmp->rdev) {
5356 				/* Replaced device not technically faulty,
5357 				 * but we need to be sure it gets removed
5358 				 * and never re-added.
5359 				 */
5360 				set_bit(Faulty, &tmp->rdev->flags);
5361 				sysfs_notify_dirent_safe(
5362 					tmp->rdev->sysfs_state);
5363 			}
5364 			sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
5365 		} else if (tmp->rdev
5366 		    && tmp->rdev->recovery_offset == MaxSector
5367 		    && !test_bit(Faulty, &tmp->rdev->flags)
5368 		    && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5369 			count++;
5370 			sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5371 		}
5372 	}
5373 	spin_lock_irqsave(&conf->device_lock, flags);
5374 	mddev->degraded = calc_degraded(conf);
5375 	spin_unlock_irqrestore(&conf->device_lock, flags);
5376 	print_raid5_conf(conf);
5377 	return count;
5378 }
5379 
5380 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
5381 {
5382 	struct r5conf *conf = mddev->private;
5383 	int err = 0;
5384 	int number = rdev->raid_disk;
5385 	struct md_rdev **rdevp;
5386 	struct disk_info *p = conf->disks + number;
5387 
5388 	print_raid5_conf(conf);
5389 	if (rdev == p->rdev)
5390 		rdevp = &p->rdev;
5391 	else if (rdev == p->replacement)
5392 		rdevp = &p->replacement;
5393 	else
5394 		return 0;
5395 
5396 	if (number >= conf->raid_disks &&
5397 	    conf->reshape_progress == MaxSector)
5398 		clear_bit(In_sync, &rdev->flags);
5399 
5400 	if (test_bit(In_sync, &rdev->flags) ||
5401 	    atomic_read(&rdev->nr_pending)) {
5402 		err = -EBUSY;
5403 		goto abort;
5404 	}
5405 	/* Only remove non-faulty devices if recovery
5406 	 * isn't possible.
5407 	 */
5408 	if (!test_bit(Faulty, &rdev->flags) &&
5409 	    mddev->recovery_disabled != conf->recovery_disabled &&
5410 	    !has_failed(conf) &&
5411 	    (!p->replacement || p->replacement == rdev) &&
5412 	    number < conf->raid_disks) {
5413 		err = -EBUSY;
5414 		goto abort;
5415 	}
5416 	*rdevp = NULL;
5417 	synchronize_rcu();
5418 	if (atomic_read(&rdev->nr_pending)) {
5419 		/* lost the race, try later */
5420 		err = -EBUSY;
5421 		*rdevp = rdev;
5422 	} else if (p->replacement) {
5423 		/* We must have just cleared 'rdev' */
5424 		p->rdev = p->replacement;
5425 		clear_bit(Replacement, &p->replacement->flags);
5426 		smp_mb(); /* Make sure other CPUs may see both as identical
5427 			   * but will never see neither - if they are careful
5428 			   */
5429 		p->replacement = NULL;
5430 		clear_bit(WantReplacement, &rdev->flags);
5431 	} else
5432 		/* We might have just removed the Replacement as faulty-
5433 		 * clear the bit just in case
5434 		 */
5435 		clear_bit(WantReplacement, &rdev->flags);
5436 abort:
5437 
5438 	print_raid5_conf(conf);
5439 	return err;
5440 }
5441 
5442 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5443 {
5444 	struct r5conf *conf = mddev->private;
5445 	int err = -EEXIST;
5446 	int disk;
5447 	struct disk_info *p;
5448 	int first = 0;
5449 	int last = conf->raid_disks - 1;
5450 
5451 	if (mddev->recovery_disabled == conf->recovery_disabled)
5452 		return -EBUSY;
5453 
5454 	if (rdev->saved_raid_disk < 0 && has_failed(conf))
5455 		/* no point adding a device */
5456 		return -EINVAL;
5457 
5458 	if (rdev->raid_disk >= 0)
5459 		first = last = rdev->raid_disk;
5460 
5461 	/*
5462 	 * find the disk ... but prefer rdev->saved_raid_disk
5463 	 * if possible.
5464 	 */
5465 	if (rdev->saved_raid_disk >= 0 &&
5466 	    rdev->saved_raid_disk >= first &&
5467 	    conf->disks[rdev->saved_raid_disk].rdev == NULL)
5468 		disk = rdev->saved_raid_disk;
5469 	else
5470 		disk = first;
5471 	for ( ; disk <= last ; disk++) {
5472 		p = conf->disks + disk;
5473 		if (p->rdev == NULL) {
5474 			clear_bit(In_sync, &rdev->flags);
5475 			rdev->raid_disk = disk;
5476 			err = 0;
5477 			if (rdev->saved_raid_disk != disk)
5478 				conf->fullsync = 1;
5479 			rcu_assign_pointer(p->rdev, rdev);
5480 			break;
5481 		}
5482 		if (test_bit(WantReplacement, &p->rdev->flags) &&
5483 		    p->replacement == NULL) {
5484 			clear_bit(In_sync, &rdev->flags);
5485 			set_bit(Replacement, &rdev->flags);
5486 			rdev->raid_disk = disk;
5487 			err = 0;
5488 			conf->fullsync = 1;
5489 			rcu_assign_pointer(p->replacement, rdev);
5490 			break;
5491 		}
5492 	}
5493 	print_raid5_conf(conf);
5494 	return err;
5495 }
5496 
5497 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5498 {
5499 	/* no resync is happening, and there is enough space
5500 	 * on all devices, so we can resize.
5501 	 * We need to make sure resync covers any new space.
5502 	 * If the array is shrinking we should possibly wait until
5503 	 * any io in the removed space completes, but it hardly seems
5504 	 * worth it.
5505 	 */
5506 	sector_t newsize;
5507 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5508 	newsize = raid5_size(mddev, sectors, mddev->raid_disks);
5509 	if (mddev->external_size &&
5510 	    mddev->array_sectors > newsize)
5511 		return -EINVAL;
5512 	if (mddev->bitmap) {
5513 		int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
5514 		if (ret)
5515 			return ret;
5516 	}
5517 	md_set_array_sectors(mddev, newsize);
5518 	set_capacity(mddev->gendisk, mddev->array_sectors);
5519 	revalidate_disk(mddev->gendisk);
5520 	if (sectors > mddev->dev_sectors &&
5521 	    mddev->recovery_cp > mddev->dev_sectors) {
5522 		mddev->recovery_cp = mddev->dev_sectors;
5523 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5524 	}
5525 	mddev->dev_sectors = sectors;
5526 	mddev->resync_max_sectors = sectors;
5527 	return 0;
5528 }
5529 
5530 static int check_stripe_cache(struct mddev *mddev)
5531 {
5532 	/* Can only proceed if there are plenty of stripe_heads.
5533 	 * We need a minimum of one full stripe,, and for sensible progress
5534 	 * it is best to have about 4 times that.
5535 	 * If we require 4 times, then the default 256 4K stripe_heads will
5536 	 * allow for chunk sizes up to 256K, which is probably OK.
5537 	 * If the chunk size is greater, user-space should request more
5538 	 * stripe_heads first.
5539 	 */
5540 	struct r5conf *conf = mddev->private;
5541 	if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5542 	    > conf->max_nr_stripes ||
5543 	    ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5544 	    > conf->max_nr_stripes) {
5545 		printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
5546 		       mdname(mddev),
5547 		       ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5548 			/ STRIPE_SIZE)*4);
5549 		return 0;
5550 	}
5551 	return 1;
5552 }
5553 
5554 static int check_reshape(struct mddev *mddev)
5555 {
5556 	struct r5conf *conf = mddev->private;
5557 
5558 	if (mddev->delta_disks == 0 &&
5559 	    mddev->new_layout == mddev->layout &&
5560 	    mddev->new_chunk_sectors == mddev->chunk_sectors)
5561 		return 0; /* nothing to do */
5562 	if (has_failed(conf))
5563 		return -EINVAL;
5564 	if (mddev->delta_disks < 0) {
5565 		/* We might be able to shrink, but the devices must
5566 		 * be made bigger first.
5567 		 * For raid6, 4 is the minimum size.
5568 		 * Otherwise 2 is the minimum
5569 		 */
5570 		int min = 2;
5571 		if (mddev->level == 6)
5572 			min = 4;
5573 		if (mddev->raid_disks + mddev->delta_disks < min)
5574 			return -EINVAL;
5575 	}
5576 
5577 	if (!check_stripe_cache(mddev))
5578 		return -ENOSPC;
5579 
5580 	return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5581 }
5582 
5583 static int raid5_start_reshape(struct mddev *mddev)
5584 {
5585 	struct r5conf *conf = mddev->private;
5586 	struct md_rdev *rdev;
5587 	int spares = 0;
5588 	unsigned long flags;
5589 
5590 	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5591 		return -EBUSY;
5592 
5593 	if (!check_stripe_cache(mddev))
5594 		return -ENOSPC;
5595 
5596 	if (has_failed(conf))
5597 		return -EINVAL;
5598 
5599 	rdev_for_each(rdev, mddev) {
5600 		if (!test_bit(In_sync, &rdev->flags)
5601 		    && !test_bit(Faulty, &rdev->flags))
5602 			spares++;
5603 	}
5604 
5605 	if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5606 		/* Not enough devices even to make a degraded array
5607 		 * of that size
5608 		 */
5609 		return -EINVAL;
5610 
5611 	/* Refuse to reduce size of the array.  Any reductions in
5612 	 * array size must be through explicit setting of array_size
5613 	 * attribute.
5614 	 */
5615 	if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5616 	    < mddev->array_sectors) {
5617 		printk(KERN_ERR "md/raid:%s: array size must be reduced "
5618 		       "before number of disks\n", mdname(mddev));
5619 		return -EINVAL;
5620 	}
5621 
5622 	atomic_set(&conf->reshape_stripes, 0);
5623 	spin_lock_irq(&conf->device_lock);
5624 	conf->previous_raid_disks = conf->raid_disks;
5625 	conf->raid_disks += mddev->delta_disks;
5626 	conf->prev_chunk_sectors = conf->chunk_sectors;
5627 	conf->chunk_sectors = mddev->new_chunk_sectors;
5628 	conf->prev_algo = conf->algorithm;
5629 	conf->algorithm = mddev->new_layout;
5630 	conf->generation++;
5631 	/* Code that selects data_offset needs to see the generation update
5632 	 * if reshape_progress has been set - so a memory barrier needed.
5633 	 */
5634 	smp_mb();
5635 	if (mddev->reshape_backwards)
5636 		conf->reshape_progress = raid5_size(mddev, 0, 0);
5637 	else
5638 		conf->reshape_progress = 0;
5639 	conf->reshape_safe = conf->reshape_progress;
5640 	spin_unlock_irq(&conf->device_lock);
5641 
5642 	/* Add some new drives, as many as will fit.
5643 	 * We know there are enough to make the newly sized array work.
5644 	 * Don't add devices if we are reducing the number of
5645 	 * devices in the array.  This is because it is not possible
5646 	 * to correctly record the "partially reconstructed" state of
5647 	 * such devices during the reshape and confusion could result.
5648 	 */
5649 	if (mddev->delta_disks >= 0) {
5650 		rdev_for_each(rdev, mddev)
5651 			if (rdev->raid_disk < 0 &&
5652 			    !test_bit(Faulty, &rdev->flags)) {
5653 				if (raid5_add_disk(mddev, rdev) == 0) {
5654 					if (rdev->raid_disk
5655 					    >= conf->previous_raid_disks)
5656 						set_bit(In_sync, &rdev->flags);
5657 					else
5658 						rdev->recovery_offset = 0;
5659 
5660 					if (sysfs_link_rdev(mddev, rdev))
5661 						/* Failure here is OK */;
5662 				}
5663 			} else if (rdev->raid_disk >= conf->previous_raid_disks
5664 				   && !test_bit(Faulty, &rdev->flags)) {
5665 				/* This is a spare that was manually added */
5666 				set_bit(In_sync, &rdev->flags);
5667 			}
5668 
5669 		/* When a reshape changes the number of devices,
5670 		 * ->degraded is measured against the larger of the
5671 		 * pre and post number of devices.
5672 		 */
5673 		spin_lock_irqsave(&conf->device_lock, flags);
5674 		mddev->degraded = calc_degraded(conf);
5675 		spin_unlock_irqrestore(&conf->device_lock, flags);
5676 	}
5677 	mddev->raid_disks = conf->raid_disks;
5678 	mddev->reshape_position = conf->reshape_progress;
5679 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
5680 
5681 	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5682 	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5683 	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5684 	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5685 	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5686 						"reshape");
5687 	if (!mddev->sync_thread) {
5688 		mddev->recovery = 0;
5689 		spin_lock_irq(&conf->device_lock);
5690 		mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5691 		rdev_for_each(rdev, mddev)
5692 			rdev->new_data_offset = rdev->data_offset;
5693 		smp_wmb();
5694 		conf->reshape_progress = MaxSector;
5695 		mddev->reshape_position = MaxSector;
5696 		spin_unlock_irq(&conf->device_lock);
5697 		return -EAGAIN;
5698 	}
5699 	conf->reshape_checkpoint = jiffies;
5700 	md_wakeup_thread(mddev->sync_thread);
5701 	md_new_event(mddev);
5702 	return 0;
5703 }
5704 
5705 /* This is called from the reshape thread and should make any
5706  * changes needed in 'conf'
5707  */
5708 static void end_reshape(struct r5conf *conf)
5709 {
5710 
5711 	if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5712 		struct md_rdev *rdev;
5713 
5714 		spin_lock_irq(&conf->device_lock);
5715 		conf->previous_raid_disks = conf->raid_disks;
5716 		rdev_for_each(rdev, conf->mddev)
5717 			rdev->data_offset = rdev->new_data_offset;
5718 		smp_wmb();
5719 		conf->reshape_progress = MaxSector;
5720 		spin_unlock_irq(&conf->device_lock);
5721 		wake_up(&conf->wait_for_overlap);
5722 
5723 		/* read-ahead size must cover two whole stripes, which is
5724 		 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5725 		 */
5726 		if (conf->mddev->queue) {
5727 			int data_disks = conf->raid_disks - conf->max_degraded;
5728 			int stripe = data_disks * ((conf->chunk_sectors << 9)
5729 						   / PAGE_SIZE);
5730 			if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5731 				conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5732 		}
5733 	}
5734 }
5735 
5736 /* This is called from the raid5d thread with mddev_lock held.
5737  * It makes config changes to the device.
5738  */
5739 static void raid5_finish_reshape(struct mddev *mddev)
5740 {
5741 	struct r5conf *conf = mddev->private;
5742 
5743 	if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5744 
5745 		if (mddev->delta_disks > 0) {
5746 			md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5747 			set_capacity(mddev->gendisk, mddev->array_sectors);
5748 			revalidate_disk(mddev->gendisk);
5749 		} else {
5750 			int d;
5751 			spin_lock_irq(&conf->device_lock);
5752 			mddev->degraded = calc_degraded(conf);
5753 			spin_unlock_irq(&conf->device_lock);
5754 			for (d = conf->raid_disks ;
5755 			     d < conf->raid_disks - mddev->delta_disks;
5756 			     d++) {
5757 				struct md_rdev *rdev = conf->disks[d].rdev;
5758 				if (rdev)
5759 					clear_bit(In_sync, &rdev->flags);
5760 				rdev = conf->disks[d].replacement;
5761 				if (rdev)
5762 					clear_bit(In_sync, &rdev->flags);
5763 			}
5764 		}
5765 		mddev->layout = conf->algorithm;
5766 		mddev->chunk_sectors = conf->chunk_sectors;
5767 		mddev->reshape_position = MaxSector;
5768 		mddev->delta_disks = 0;
5769 		mddev->reshape_backwards = 0;
5770 	}
5771 }
5772 
5773 static void raid5_quiesce(struct mddev *mddev, int state)
5774 {
5775 	struct r5conf *conf = mddev->private;
5776 
5777 	switch(state) {
5778 	case 2: /* resume for a suspend */
5779 		wake_up(&conf->wait_for_overlap);
5780 		break;
5781 
5782 	case 1: /* stop all writes */
5783 		spin_lock_irq(&conf->device_lock);
5784 		/* '2' tells resync/reshape to pause so that all
5785 		 * active stripes can drain
5786 		 */
5787 		conf->quiesce = 2;
5788 		wait_event_lock_irq(conf->wait_for_stripe,
5789 				    atomic_read(&conf->active_stripes) == 0 &&
5790 				    atomic_read(&conf->active_aligned_reads) == 0,
5791 				    conf->device_lock, /* nothing */);
5792 		conf->quiesce = 1;
5793 		spin_unlock_irq(&conf->device_lock);
5794 		/* allow reshape to continue */
5795 		wake_up(&conf->wait_for_overlap);
5796 		break;
5797 
5798 	case 0: /* re-enable writes */
5799 		spin_lock_irq(&conf->device_lock);
5800 		conf->quiesce = 0;
5801 		wake_up(&conf->wait_for_stripe);
5802 		wake_up(&conf->wait_for_overlap);
5803 		spin_unlock_irq(&conf->device_lock);
5804 		break;
5805 	}
5806 }
5807 
5808 
5809 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5810 {
5811 	struct r0conf *raid0_conf = mddev->private;
5812 	sector_t sectors;
5813 
5814 	/* for raid0 takeover only one zone is supported */
5815 	if (raid0_conf->nr_strip_zones > 1) {
5816 		printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5817 		       mdname(mddev));
5818 		return ERR_PTR(-EINVAL);
5819 	}
5820 
5821 	sectors = raid0_conf->strip_zone[0].zone_end;
5822 	sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5823 	mddev->dev_sectors = sectors;
5824 	mddev->new_level = level;
5825 	mddev->new_layout = ALGORITHM_PARITY_N;
5826 	mddev->new_chunk_sectors = mddev->chunk_sectors;
5827 	mddev->raid_disks += 1;
5828 	mddev->delta_disks = 1;
5829 	/* make sure it will be not marked as dirty */
5830 	mddev->recovery_cp = MaxSector;
5831 
5832 	return setup_conf(mddev);
5833 }
5834 
5835 
5836 static void *raid5_takeover_raid1(struct mddev *mddev)
5837 {
5838 	int chunksect;
5839 
5840 	if (mddev->raid_disks != 2 ||
5841 	    mddev->degraded > 1)
5842 		return ERR_PTR(-EINVAL);
5843 
5844 	/* Should check if there are write-behind devices? */
5845 
5846 	chunksect = 64*2; /* 64K by default */
5847 
5848 	/* The array must be an exact multiple of chunksize */
5849 	while (chunksect && (mddev->array_sectors & (chunksect-1)))
5850 		chunksect >>= 1;
5851 
5852 	if ((chunksect<<9) < STRIPE_SIZE)
5853 		/* array size does not allow a suitable chunk size */
5854 		return ERR_PTR(-EINVAL);
5855 
5856 	mddev->new_level = 5;
5857 	mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5858 	mddev->new_chunk_sectors = chunksect;
5859 
5860 	return setup_conf(mddev);
5861 }
5862 
5863 static void *raid5_takeover_raid6(struct mddev *mddev)
5864 {
5865 	int new_layout;
5866 
5867 	switch (mddev->layout) {
5868 	case ALGORITHM_LEFT_ASYMMETRIC_6:
5869 		new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5870 		break;
5871 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
5872 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5873 		break;
5874 	case ALGORITHM_LEFT_SYMMETRIC_6:
5875 		new_layout = ALGORITHM_LEFT_SYMMETRIC;
5876 		break;
5877 	case ALGORITHM_RIGHT_SYMMETRIC_6:
5878 		new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5879 		break;
5880 	case ALGORITHM_PARITY_0_6:
5881 		new_layout = ALGORITHM_PARITY_0;
5882 		break;
5883 	case ALGORITHM_PARITY_N:
5884 		new_layout = ALGORITHM_PARITY_N;
5885 		break;
5886 	default:
5887 		return ERR_PTR(-EINVAL);
5888 	}
5889 	mddev->new_level = 5;
5890 	mddev->new_layout = new_layout;
5891 	mddev->delta_disks = -1;
5892 	mddev->raid_disks -= 1;
5893 	return setup_conf(mddev);
5894 }
5895 
5896 
5897 static int raid5_check_reshape(struct mddev *mddev)
5898 {
5899 	/* For a 2-drive array, the layout and chunk size can be changed
5900 	 * immediately as not restriping is needed.
5901 	 * For larger arrays we record the new value - after validation
5902 	 * to be used by a reshape pass.
5903 	 */
5904 	struct r5conf *conf = mddev->private;
5905 	int new_chunk = mddev->new_chunk_sectors;
5906 
5907 	if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5908 		return -EINVAL;
5909 	if (new_chunk > 0) {
5910 		if (!is_power_of_2(new_chunk))
5911 			return -EINVAL;
5912 		if (new_chunk < (PAGE_SIZE>>9))
5913 			return -EINVAL;
5914 		if (mddev->array_sectors & (new_chunk-1))
5915 			/* not factor of array size */
5916 			return -EINVAL;
5917 	}
5918 
5919 	/* They look valid */
5920 
5921 	if (mddev->raid_disks == 2) {
5922 		/* can make the change immediately */
5923 		if (mddev->new_layout >= 0) {
5924 			conf->algorithm = mddev->new_layout;
5925 			mddev->layout = mddev->new_layout;
5926 		}
5927 		if (new_chunk > 0) {
5928 			conf->chunk_sectors = new_chunk ;
5929 			mddev->chunk_sectors = new_chunk;
5930 		}
5931 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
5932 		md_wakeup_thread(mddev->thread);
5933 	}
5934 	return check_reshape(mddev);
5935 }
5936 
5937 static int raid6_check_reshape(struct mddev *mddev)
5938 {
5939 	int new_chunk = mddev->new_chunk_sectors;
5940 
5941 	if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5942 		return -EINVAL;
5943 	if (new_chunk > 0) {
5944 		if (!is_power_of_2(new_chunk))
5945 			return -EINVAL;
5946 		if (new_chunk < (PAGE_SIZE >> 9))
5947 			return -EINVAL;
5948 		if (mddev->array_sectors & (new_chunk-1))
5949 			/* not factor of array size */
5950 			return -EINVAL;
5951 	}
5952 
5953 	/* They look valid */
5954 	return check_reshape(mddev);
5955 }
5956 
5957 static void *raid5_takeover(struct mddev *mddev)
5958 {
5959 	/* raid5 can take over:
5960 	 *  raid0 - if there is only one strip zone - make it a raid4 layout
5961 	 *  raid1 - if there are two drives.  We need to know the chunk size
5962 	 *  raid4 - trivial - just use a raid4 layout.
5963 	 *  raid6 - Providing it is a *_6 layout
5964 	 */
5965 	if (mddev->level == 0)
5966 		return raid45_takeover_raid0(mddev, 5);
5967 	if (mddev->level == 1)
5968 		return raid5_takeover_raid1(mddev);
5969 	if (mddev->level == 4) {
5970 		mddev->new_layout = ALGORITHM_PARITY_N;
5971 		mddev->new_level = 5;
5972 		return setup_conf(mddev);
5973 	}
5974 	if (mddev->level == 6)
5975 		return raid5_takeover_raid6(mddev);
5976 
5977 	return ERR_PTR(-EINVAL);
5978 }
5979 
5980 static void *raid4_takeover(struct mddev *mddev)
5981 {
5982 	/* raid4 can take over:
5983 	 *  raid0 - if there is only one strip zone
5984 	 *  raid5 - if layout is right
5985 	 */
5986 	if (mddev->level == 0)
5987 		return raid45_takeover_raid0(mddev, 4);
5988 	if (mddev->level == 5 &&
5989 	    mddev->layout == ALGORITHM_PARITY_N) {
5990 		mddev->new_layout = 0;
5991 		mddev->new_level = 4;
5992 		return setup_conf(mddev);
5993 	}
5994 	return ERR_PTR(-EINVAL);
5995 }
5996 
5997 static struct md_personality raid5_personality;
5998 
5999 static void *raid6_takeover(struct mddev *mddev)
6000 {
6001 	/* Currently can only take over a raid5.  We map the
6002 	 * personality to an equivalent raid6 personality
6003 	 * with the Q block at the end.
6004 	 */
6005 	int new_layout;
6006 
6007 	if (mddev->pers != &raid5_personality)
6008 		return ERR_PTR(-EINVAL);
6009 	if (mddev->degraded > 1)
6010 		return ERR_PTR(-EINVAL);
6011 	if (mddev->raid_disks > 253)
6012 		return ERR_PTR(-EINVAL);
6013 	if (mddev->raid_disks < 3)
6014 		return ERR_PTR(-EINVAL);
6015 
6016 	switch (mddev->layout) {
6017 	case ALGORITHM_LEFT_ASYMMETRIC:
6018 		new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6019 		break;
6020 	case ALGORITHM_RIGHT_ASYMMETRIC:
6021 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6022 		break;
6023 	case ALGORITHM_LEFT_SYMMETRIC:
6024 		new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6025 		break;
6026 	case ALGORITHM_RIGHT_SYMMETRIC:
6027 		new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6028 		break;
6029 	case ALGORITHM_PARITY_0:
6030 		new_layout = ALGORITHM_PARITY_0_6;
6031 		break;
6032 	case ALGORITHM_PARITY_N:
6033 		new_layout = ALGORITHM_PARITY_N;
6034 		break;
6035 	default:
6036 		return ERR_PTR(-EINVAL);
6037 	}
6038 	mddev->new_level = 6;
6039 	mddev->new_layout = new_layout;
6040 	mddev->delta_disks = 1;
6041 	mddev->raid_disks += 1;
6042 	return setup_conf(mddev);
6043 }
6044 
6045 
6046 static struct md_personality raid6_personality =
6047 {
6048 	.name		= "raid6",
6049 	.level		= 6,
6050 	.owner		= THIS_MODULE,
6051 	.make_request	= make_request,
6052 	.run		= run,
6053 	.stop		= stop,
6054 	.status		= status,
6055 	.error_handler	= error,
6056 	.hot_add_disk	= raid5_add_disk,
6057 	.hot_remove_disk= raid5_remove_disk,
6058 	.spare_active	= raid5_spare_active,
6059 	.sync_request	= sync_request,
6060 	.resize		= raid5_resize,
6061 	.size		= raid5_size,
6062 	.check_reshape	= raid6_check_reshape,
6063 	.start_reshape  = raid5_start_reshape,
6064 	.finish_reshape = raid5_finish_reshape,
6065 	.quiesce	= raid5_quiesce,
6066 	.takeover	= raid6_takeover,
6067 };
6068 static struct md_personality raid5_personality =
6069 {
6070 	.name		= "raid5",
6071 	.level		= 5,
6072 	.owner		= THIS_MODULE,
6073 	.make_request	= make_request,
6074 	.run		= run,
6075 	.stop		= stop,
6076 	.status		= status,
6077 	.error_handler	= error,
6078 	.hot_add_disk	= raid5_add_disk,
6079 	.hot_remove_disk= raid5_remove_disk,
6080 	.spare_active	= raid5_spare_active,
6081 	.sync_request	= sync_request,
6082 	.resize		= raid5_resize,
6083 	.size		= raid5_size,
6084 	.check_reshape	= raid5_check_reshape,
6085 	.start_reshape  = raid5_start_reshape,
6086 	.finish_reshape = raid5_finish_reshape,
6087 	.quiesce	= raid5_quiesce,
6088 	.takeover	= raid5_takeover,
6089 };
6090 
6091 static struct md_personality raid4_personality =
6092 {
6093 	.name		= "raid4",
6094 	.level		= 4,
6095 	.owner		= THIS_MODULE,
6096 	.make_request	= make_request,
6097 	.run		= run,
6098 	.stop		= stop,
6099 	.status		= status,
6100 	.error_handler	= error,
6101 	.hot_add_disk	= raid5_add_disk,
6102 	.hot_remove_disk= raid5_remove_disk,
6103 	.spare_active	= raid5_spare_active,
6104 	.sync_request	= sync_request,
6105 	.resize		= raid5_resize,
6106 	.size		= raid5_size,
6107 	.check_reshape	= raid5_check_reshape,
6108 	.start_reshape  = raid5_start_reshape,
6109 	.finish_reshape = raid5_finish_reshape,
6110 	.quiesce	= raid5_quiesce,
6111 	.takeover	= raid4_takeover,
6112 };
6113 
6114 static int __init raid5_init(void)
6115 {
6116 	register_md_personality(&raid6_personality);
6117 	register_md_personality(&raid5_personality);
6118 	register_md_personality(&raid4_personality);
6119 	return 0;
6120 }
6121 
6122 static void raid5_exit(void)
6123 {
6124 	unregister_md_personality(&raid6_personality);
6125 	unregister_md_personality(&raid5_personality);
6126 	unregister_md_personality(&raid4_personality);
6127 }
6128 
6129 module_init(raid5_init);
6130 module_exit(raid5_exit);
6131 MODULE_LICENSE("GPL");
6132 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6133 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6134 MODULE_ALIAS("md-raid5");
6135 MODULE_ALIAS("md-raid4");
6136 MODULE_ALIAS("md-level-5");
6137 MODULE_ALIAS("md-level-4");
6138 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6139 MODULE_ALIAS("md-raid6");
6140 MODULE_ALIAS("md-level-6");
6141 
6142 /* This used to be two separate modules, they were: */
6143 MODULE_ALIAS("raid5");
6144 MODULE_ALIAS("raid6");
6145