xref: /openbmc/linux/drivers/md/raid5.c (revision cd5d5810)
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 <linux/nodemask.h>
57 #include <trace/events/block.h>
58 
59 #include "md.h"
60 #include "raid5.h"
61 #include "raid0.h"
62 #include "bitmap.h"
63 
64 #define cpu_to_group(cpu) cpu_to_node(cpu)
65 #define ANY_GROUP NUMA_NO_NODE
66 
67 static struct workqueue_struct *raid5_wq;
68 /*
69  * Stripe cache
70  */
71 
72 #define NR_STRIPES		256
73 #define STRIPE_SIZE		PAGE_SIZE
74 #define STRIPE_SHIFT		(PAGE_SHIFT - 9)
75 #define STRIPE_SECTORS		(STRIPE_SIZE>>9)
76 #define	IO_THRESHOLD		1
77 #define BYPASS_THRESHOLD	1
78 #define NR_HASH			(PAGE_SIZE / sizeof(struct hlist_head))
79 #define HASH_MASK		(NR_HASH - 1)
80 #define MAX_STRIPE_BATCH	8
81 
82 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
83 {
84 	int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
85 	return &conf->stripe_hashtbl[hash];
86 }
87 
88 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
89  * order without overlap.  There may be several bio's per stripe+device, and
90  * a bio could span several devices.
91  * When walking this list for a particular stripe+device, we must never proceed
92  * beyond a bio that extends past this device, as the next bio might no longer
93  * be valid.
94  * This function is used to determine the 'next' bio in the list, given the sector
95  * of the current stripe+device
96  */
97 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
98 {
99 	int sectors = bio_sectors(bio);
100 	if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
101 		return bio->bi_next;
102 	else
103 		return NULL;
104 }
105 
106 /*
107  * We maintain a biased count of active stripes in the bottom 16 bits of
108  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
109  */
110 static inline int raid5_bi_processed_stripes(struct bio *bio)
111 {
112 	atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
113 	return (atomic_read(segments) >> 16) & 0xffff;
114 }
115 
116 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
117 {
118 	atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
119 	return atomic_sub_return(1, segments) & 0xffff;
120 }
121 
122 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
123 {
124 	atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
125 	atomic_inc(segments);
126 }
127 
128 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
129 	unsigned int cnt)
130 {
131 	atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
132 	int old, new;
133 
134 	do {
135 		old = atomic_read(segments);
136 		new = (old & 0xffff) | (cnt << 16);
137 	} while (atomic_cmpxchg(segments, old, new) != old);
138 }
139 
140 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
141 {
142 	atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
143 	atomic_set(segments, cnt);
144 }
145 
146 /* Find first data disk in a raid6 stripe */
147 static inline int raid6_d0(struct stripe_head *sh)
148 {
149 	if (sh->ddf_layout)
150 		/* ddf always start from first device */
151 		return 0;
152 	/* md starts just after Q block */
153 	if (sh->qd_idx == sh->disks - 1)
154 		return 0;
155 	else
156 		return sh->qd_idx + 1;
157 }
158 static inline int raid6_next_disk(int disk, int raid_disks)
159 {
160 	disk++;
161 	return (disk < raid_disks) ? disk : 0;
162 }
163 
164 /* When walking through the disks in a raid5, starting at raid6_d0,
165  * We need to map each disk to a 'slot', where the data disks are slot
166  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
167  * is raid_disks-1.  This help does that mapping.
168  */
169 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
170 			     int *count, int syndrome_disks)
171 {
172 	int slot = *count;
173 
174 	if (sh->ddf_layout)
175 		(*count)++;
176 	if (idx == sh->pd_idx)
177 		return syndrome_disks;
178 	if (idx == sh->qd_idx)
179 		return syndrome_disks + 1;
180 	if (!sh->ddf_layout)
181 		(*count)++;
182 	return slot;
183 }
184 
185 static void return_io(struct bio *return_bi)
186 {
187 	struct bio *bi = return_bi;
188 	while (bi) {
189 
190 		return_bi = bi->bi_next;
191 		bi->bi_next = NULL;
192 		bi->bi_size = 0;
193 		trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
194 					 bi, 0);
195 		bio_endio(bi, 0);
196 		bi = return_bi;
197 	}
198 }
199 
200 static void print_raid5_conf (struct r5conf *conf);
201 
202 static int stripe_operations_active(struct stripe_head *sh)
203 {
204 	return sh->check_state || sh->reconstruct_state ||
205 	       test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
206 	       test_bit(STRIPE_COMPUTE_RUN, &sh->state);
207 }
208 
209 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
210 {
211 	struct r5conf *conf = sh->raid_conf;
212 	struct r5worker_group *group;
213 	int thread_cnt;
214 	int i, cpu = sh->cpu;
215 
216 	if (!cpu_online(cpu)) {
217 		cpu = cpumask_any(cpu_online_mask);
218 		sh->cpu = cpu;
219 	}
220 
221 	if (list_empty(&sh->lru)) {
222 		struct r5worker_group *group;
223 		group = conf->worker_groups + cpu_to_group(cpu);
224 		list_add_tail(&sh->lru, &group->handle_list);
225 		group->stripes_cnt++;
226 		sh->group = group;
227 	}
228 
229 	if (conf->worker_cnt_per_group == 0) {
230 		md_wakeup_thread(conf->mddev->thread);
231 		return;
232 	}
233 
234 	group = conf->worker_groups + cpu_to_group(sh->cpu);
235 
236 	group->workers[0].working = true;
237 	/* at least one worker should run to avoid race */
238 	queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
239 
240 	thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
241 	/* wakeup more workers */
242 	for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
243 		if (group->workers[i].working == false) {
244 			group->workers[i].working = true;
245 			queue_work_on(sh->cpu, raid5_wq,
246 				      &group->workers[i].work);
247 			thread_cnt--;
248 		}
249 	}
250 }
251 
252 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh)
253 {
254 	BUG_ON(!list_empty(&sh->lru));
255 	BUG_ON(atomic_read(&conf->active_stripes)==0);
256 	if (test_bit(STRIPE_HANDLE, &sh->state)) {
257 		if (test_bit(STRIPE_DELAYED, &sh->state) &&
258 		    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
259 			list_add_tail(&sh->lru, &conf->delayed_list);
260 		else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
261 			   sh->bm_seq - conf->seq_write > 0)
262 			list_add_tail(&sh->lru, &conf->bitmap_list);
263 		else {
264 			clear_bit(STRIPE_DELAYED, &sh->state);
265 			clear_bit(STRIPE_BIT_DELAY, &sh->state);
266 			if (conf->worker_cnt_per_group == 0) {
267 				list_add_tail(&sh->lru, &conf->handle_list);
268 			} else {
269 				raid5_wakeup_stripe_thread(sh);
270 				return;
271 			}
272 		}
273 		md_wakeup_thread(conf->mddev->thread);
274 	} else {
275 		BUG_ON(stripe_operations_active(sh));
276 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
277 			if (atomic_dec_return(&conf->preread_active_stripes)
278 			    < IO_THRESHOLD)
279 				md_wakeup_thread(conf->mddev->thread);
280 		atomic_dec(&conf->active_stripes);
281 		if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
282 			list_add_tail(&sh->lru, &conf->inactive_list);
283 			wake_up(&conf->wait_for_stripe);
284 			if (conf->retry_read_aligned)
285 				md_wakeup_thread(conf->mddev->thread);
286 		}
287 	}
288 }
289 
290 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
291 {
292 	if (atomic_dec_and_test(&sh->count))
293 		do_release_stripe(conf, sh);
294 }
295 
296 static struct llist_node *llist_reverse_order(struct llist_node *head)
297 {
298 	struct llist_node *new_head = NULL;
299 
300 	while (head) {
301 		struct llist_node *tmp = head;
302 		head = head->next;
303 		tmp->next = new_head;
304 		new_head = tmp;
305 	}
306 
307 	return new_head;
308 }
309 
310 /* should hold conf->device_lock already */
311 static int release_stripe_list(struct r5conf *conf)
312 {
313 	struct stripe_head *sh;
314 	int count = 0;
315 	struct llist_node *head;
316 
317 	head = llist_del_all(&conf->released_stripes);
318 	head = llist_reverse_order(head);
319 	while (head) {
320 		sh = llist_entry(head, struct stripe_head, release_list);
321 		head = llist_next(head);
322 		/* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
323 		smp_mb();
324 		clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
325 		/*
326 		 * Don't worry the bit is set here, because if the bit is set
327 		 * again, the count is always > 1. This is true for
328 		 * STRIPE_ON_UNPLUG_LIST bit too.
329 		 */
330 		__release_stripe(conf, sh);
331 		count++;
332 	}
333 
334 	return count;
335 }
336 
337 static void release_stripe(struct stripe_head *sh)
338 {
339 	struct r5conf *conf = sh->raid_conf;
340 	unsigned long flags;
341 	bool wakeup;
342 
343 	if (test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
344 		goto slow_path;
345 	wakeup = llist_add(&sh->release_list, &conf->released_stripes);
346 	if (wakeup)
347 		md_wakeup_thread(conf->mddev->thread);
348 	return;
349 slow_path:
350 	local_irq_save(flags);
351 	/* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
352 	if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
353 		do_release_stripe(conf, sh);
354 		spin_unlock(&conf->device_lock);
355 	}
356 	local_irq_restore(flags);
357 }
358 
359 static inline void remove_hash(struct stripe_head *sh)
360 {
361 	pr_debug("remove_hash(), stripe %llu\n",
362 		(unsigned long long)sh->sector);
363 
364 	hlist_del_init(&sh->hash);
365 }
366 
367 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
368 {
369 	struct hlist_head *hp = stripe_hash(conf, sh->sector);
370 
371 	pr_debug("insert_hash(), stripe %llu\n",
372 		(unsigned long long)sh->sector);
373 
374 	hlist_add_head(&sh->hash, hp);
375 }
376 
377 
378 /* find an idle stripe, make sure it is unhashed, and return it. */
379 static struct stripe_head *get_free_stripe(struct r5conf *conf)
380 {
381 	struct stripe_head *sh = NULL;
382 	struct list_head *first;
383 
384 	if (list_empty(&conf->inactive_list))
385 		goto out;
386 	first = conf->inactive_list.next;
387 	sh = list_entry(first, struct stripe_head, lru);
388 	list_del_init(first);
389 	remove_hash(sh);
390 	atomic_inc(&conf->active_stripes);
391 out:
392 	return sh;
393 }
394 
395 static void shrink_buffers(struct stripe_head *sh)
396 {
397 	struct page *p;
398 	int i;
399 	int num = sh->raid_conf->pool_size;
400 
401 	for (i = 0; i < num ; i++) {
402 		p = sh->dev[i].page;
403 		if (!p)
404 			continue;
405 		sh->dev[i].page = NULL;
406 		put_page(p);
407 	}
408 }
409 
410 static int grow_buffers(struct stripe_head *sh)
411 {
412 	int i;
413 	int num = sh->raid_conf->pool_size;
414 
415 	for (i = 0; i < num; i++) {
416 		struct page *page;
417 
418 		if (!(page = alloc_page(GFP_KERNEL))) {
419 			return 1;
420 		}
421 		sh->dev[i].page = page;
422 	}
423 	return 0;
424 }
425 
426 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
427 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
428 			    struct stripe_head *sh);
429 
430 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
431 {
432 	struct r5conf *conf = sh->raid_conf;
433 	int i;
434 
435 	BUG_ON(atomic_read(&sh->count) != 0);
436 	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
437 	BUG_ON(stripe_operations_active(sh));
438 
439 	pr_debug("init_stripe called, stripe %llu\n",
440 		(unsigned long long)sh->sector);
441 
442 	remove_hash(sh);
443 
444 	sh->generation = conf->generation - previous;
445 	sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
446 	sh->sector = sector;
447 	stripe_set_idx(sector, conf, previous, sh);
448 	sh->state = 0;
449 
450 
451 	for (i = sh->disks; i--; ) {
452 		struct r5dev *dev = &sh->dev[i];
453 
454 		if (dev->toread || dev->read || dev->towrite || dev->written ||
455 		    test_bit(R5_LOCKED, &dev->flags)) {
456 			printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
457 			       (unsigned long long)sh->sector, i, dev->toread,
458 			       dev->read, dev->towrite, dev->written,
459 			       test_bit(R5_LOCKED, &dev->flags));
460 			WARN_ON(1);
461 		}
462 		dev->flags = 0;
463 		raid5_build_block(sh, i, previous);
464 	}
465 	insert_hash(conf, sh);
466 	sh->cpu = smp_processor_id();
467 }
468 
469 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
470 					 short generation)
471 {
472 	struct stripe_head *sh;
473 
474 	pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
475 	hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
476 		if (sh->sector == sector && sh->generation == generation)
477 			return sh;
478 	pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
479 	return NULL;
480 }
481 
482 /*
483  * Need to check if array has failed when deciding whether to:
484  *  - start an array
485  *  - remove non-faulty devices
486  *  - add a spare
487  *  - allow a reshape
488  * This determination is simple when no reshape is happening.
489  * However if there is a reshape, we need to carefully check
490  * both the before and after sections.
491  * This is because some failed devices may only affect one
492  * of the two sections, and some non-in_sync devices may
493  * be insync in the section most affected by failed devices.
494  */
495 static int calc_degraded(struct r5conf *conf)
496 {
497 	int degraded, degraded2;
498 	int i;
499 
500 	rcu_read_lock();
501 	degraded = 0;
502 	for (i = 0; i < conf->previous_raid_disks; i++) {
503 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
504 		if (rdev && test_bit(Faulty, &rdev->flags))
505 			rdev = rcu_dereference(conf->disks[i].replacement);
506 		if (!rdev || test_bit(Faulty, &rdev->flags))
507 			degraded++;
508 		else if (test_bit(In_sync, &rdev->flags))
509 			;
510 		else
511 			/* not in-sync or faulty.
512 			 * If the reshape increases the number of devices,
513 			 * this is being recovered by the reshape, so
514 			 * this 'previous' section is not in_sync.
515 			 * If the number of devices is being reduced however,
516 			 * the device can only be part of the array if
517 			 * we are reverting a reshape, so this section will
518 			 * be in-sync.
519 			 */
520 			if (conf->raid_disks >= conf->previous_raid_disks)
521 				degraded++;
522 	}
523 	rcu_read_unlock();
524 	if (conf->raid_disks == conf->previous_raid_disks)
525 		return degraded;
526 	rcu_read_lock();
527 	degraded2 = 0;
528 	for (i = 0; i < conf->raid_disks; i++) {
529 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
530 		if (rdev && test_bit(Faulty, &rdev->flags))
531 			rdev = rcu_dereference(conf->disks[i].replacement);
532 		if (!rdev || test_bit(Faulty, &rdev->flags))
533 			degraded2++;
534 		else if (test_bit(In_sync, &rdev->flags))
535 			;
536 		else
537 			/* not in-sync or faulty.
538 			 * If reshape increases the number of devices, this
539 			 * section has already been recovered, else it
540 			 * almost certainly hasn't.
541 			 */
542 			if (conf->raid_disks <= conf->previous_raid_disks)
543 				degraded2++;
544 	}
545 	rcu_read_unlock();
546 	if (degraded2 > degraded)
547 		return degraded2;
548 	return degraded;
549 }
550 
551 static int has_failed(struct r5conf *conf)
552 {
553 	int degraded;
554 
555 	if (conf->mddev->reshape_position == MaxSector)
556 		return conf->mddev->degraded > conf->max_degraded;
557 
558 	degraded = calc_degraded(conf);
559 	if (degraded > conf->max_degraded)
560 		return 1;
561 	return 0;
562 }
563 
564 static struct stripe_head *
565 get_active_stripe(struct r5conf *conf, sector_t sector,
566 		  int previous, int noblock, int noquiesce)
567 {
568 	struct stripe_head *sh;
569 
570 	pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
571 
572 	spin_lock_irq(&conf->device_lock);
573 
574 	do {
575 		wait_event_lock_irq(conf->wait_for_stripe,
576 				    conf->quiesce == 0 || noquiesce,
577 				    conf->device_lock);
578 		sh = __find_stripe(conf, sector, conf->generation - previous);
579 		if (!sh) {
580 			if (!conf->inactive_blocked)
581 				sh = get_free_stripe(conf);
582 			if (noblock && sh == NULL)
583 				break;
584 			if (!sh) {
585 				conf->inactive_blocked = 1;
586 				wait_event_lock_irq(conf->wait_for_stripe,
587 						    !list_empty(&conf->inactive_list) &&
588 						    (atomic_read(&conf->active_stripes)
589 						     < (conf->max_nr_stripes *3/4)
590 						     || !conf->inactive_blocked),
591 						    conf->device_lock);
592 				conf->inactive_blocked = 0;
593 			} else
594 				init_stripe(sh, sector, previous);
595 		} else {
596 			if (atomic_read(&sh->count)) {
597 				BUG_ON(!list_empty(&sh->lru)
598 				    && !test_bit(STRIPE_EXPANDING, &sh->state)
599 				    && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state)
600 				    && !test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
601 			} else {
602 				if (!test_bit(STRIPE_HANDLE, &sh->state))
603 					atomic_inc(&conf->active_stripes);
604 				if (list_empty(&sh->lru) &&
605 				    !test_bit(STRIPE_EXPANDING, &sh->state))
606 					BUG();
607 				list_del_init(&sh->lru);
608 				if (sh->group) {
609 					sh->group->stripes_cnt--;
610 					sh->group = NULL;
611 				}
612 			}
613 		}
614 	} while (sh == NULL);
615 
616 	if (sh)
617 		atomic_inc(&sh->count);
618 
619 	spin_unlock_irq(&conf->device_lock);
620 	return sh;
621 }
622 
623 /* Determine if 'data_offset' or 'new_data_offset' should be used
624  * in this stripe_head.
625  */
626 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
627 {
628 	sector_t progress = conf->reshape_progress;
629 	/* Need a memory barrier to make sure we see the value
630 	 * of conf->generation, or ->data_offset that was set before
631 	 * reshape_progress was updated.
632 	 */
633 	smp_rmb();
634 	if (progress == MaxSector)
635 		return 0;
636 	if (sh->generation == conf->generation - 1)
637 		return 0;
638 	/* We are in a reshape, and this is a new-generation stripe,
639 	 * so use new_data_offset.
640 	 */
641 	return 1;
642 }
643 
644 static void
645 raid5_end_read_request(struct bio *bi, int error);
646 static void
647 raid5_end_write_request(struct bio *bi, int error);
648 
649 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
650 {
651 	struct r5conf *conf = sh->raid_conf;
652 	int i, disks = sh->disks;
653 
654 	might_sleep();
655 
656 	for (i = disks; i--; ) {
657 		int rw;
658 		int replace_only = 0;
659 		struct bio *bi, *rbi;
660 		struct md_rdev *rdev, *rrdev = NULL;
661 		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
662 			if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
663 				rw = WRITE_FUA;
664 			else
665 				rw = WRITE;
666 			if (test_bit(R5_Discard, &sh->dev[i].flags))
667 				rw |= REQ_DISCARD;
668 		} else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
669 			rw = READ;
670 		else if (test_and_clear_bit(R5_WantReplace,
671 					    &sh->dev[i].flags)) {
672 			rw = WRITE;
673 			replace_only = 1;
674 		} else
675 			continue;
676 		if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
677 			rw |= REQ_SYNC;
678 
679 		bi = &sh->dev[i].req;
680 		rbi = &sh->dev[i].rreq; /* For writing to replacement */
681 
682 		rcu_read_lock();
683 		rrdev = rcu_dereference(conf->disks[i].replacement);
684 		smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
685 		rdev = rcu_dereference(conf->disks[i].rdev);
686 		if (!rdev) {
687 			rdev = rrdev;
688 			rrdev = NULL;
689 		}
690 		if (rw & WRITE) {
691 			if (replace_only)
692 				rdev = NULL;
693 			if (rdev == rrdev)
694 				/* We raced and saw duplicates */
695 				rrdev = NULL;
696 		} else {
697 			if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
698 				rdev = rrdev;
699 			rrdev = NULL;
700 		}
701 
702 		if (rdev && test_bit(Faulty, &rdev->flags))
703 			rdev = NULL;
704 		if (rdev)
705 			atomic_inc(&rdev->nr_pending);
706 		if (rrdev && test_bit(Faulty, &rrdev->flags))
707 			rrdev = NULL;
708 		if (rrdev)
709 			atomic_inc(&rrdev->nr_pending);
710 		rcu_read_unlock();
711 
712 		/* We have already checked bad blocks for reads.  Now
713 		 * need to check for writes.  We never accept write errors
714 		 * on the replacement, so we don't to check rrdev.
715 		 */
716 		while ((rw & WRITE) && rdev &&
717 		       test_bit(WriteErrorSeen, &rdev->flags)) {
718 			sector_t first_bad;
719 			int bad_sectors;
720 			int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
721 					      &first_bad, &bad_sectors);
722 			if (!bad)
723 				break;
724 
725 			if (bad < 0) {
726 				set_bit(BlockedBadBlocks, &rdev->flags);
727 				if (!conf->mddev->external &&
728 				    conf->mddev->flags) {
729 					/* It is very unlikely, but we might
730 					 * still need to write out the
731 					 * bad block log - better give it
732 					 * a chance*/
733 					md_check_recovery(conf->mddev);
734 				}
735 				/*
736 				 * Because md_wait_for_blocked_rdev
737 				 * will dec nr_pending, we must
738 				 * increment it first.
739 				 */
740 				atomic_inc(&rdev->nr_pending);
741 				md_wait_for_blocked_rdev(rdev, conf->mddev);
742 			} else {
743 				/* Acknowledged bad block - skip the write */
744 				rdev_dec_pending(rdev, conf->mddev);
745 				rdev = NULL;
746 			}
747 		}
748 
749 		if (rdev) {
750 			if (s->syncing || s->expanding || s->expanded
751 			    || s->replacing)
752 				md_sync_acct(rdev->bdev, STRIPE_SECTORS);
753 
754 			set_bit(STRIPE_IO_STARTED, &sh->state);
755 
756 			bio_reset(bi);
757 			bi->bi_bdev = rdev->bdev;
758 			bi->bi_rw = rw;
759 			bi->bi_end_io = (rw & WRITE)
760 				? raid5_end_write_request
761 				: raid5_end_read_request;
762 			bi->bi_private = sh;
763 
764 			pr_debug("%s: for %llu schedule op %ld on disc %d\n",
765 				__func__, (unsigned long long)sh->sector,
766 				bi->bi_rw, i);
767 			atomic_inc(&sh->count);
768 			if (use_new_offset(conf, sh))
769 				bi->bi_sector = (sh->sector
770 						 + rdev->new_data_offset);
771 			else
772 				bi->bi_sector = (sh->sector
773 						 + rdev->data_offset);
774 			if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
775 				bi->bi_rw |= REQ_FLUSH;
776 
777 			bi->bi_vcnt = 1;
778 			bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
779 			bi->bi_io_vec[0].bv_offset = 0;
780 			bi->bi_size = STRIPE_SIZE;
781 			/*
782 			 * If this is discard request, set bi_vcnt 0. We don't
783 			 * want to confuse SCSI because SCSI will replace payload
784 			 */
785 			if (rw & REQ_DISCARD)
786 				bi->bi_vcnt = 0;
787 			if (rrdev)
788 				set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
789 
790 			if (conf->mddev->gendisk)
791 				trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
792 						      bi, disk_devt(conf->mddev->gendisk),
793 						      sh->dev[i].sector);
794 			generic_make_request(bi);
795 		}
796 		if (rrdev) {
797 			if (s->syncing || s->expanding || s->expanded
798 			    || s->replacing)
799 				md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
800 
801 			set_bit(STRIPE_IO_STARTED, &sh->state);
802 
803 			bio_reset(rbi);
804 			rbi->bi_bdev = rrdev->bdev;
805 			rbi->bi_rw = rw;
806 			BUG_ON(!(rw & WRITE));
807 			rbi->bi_end_io = raid5_end_write_request;
808 			rbi->bi_private = sh;
809 
810 			pr_debug("%s: for %llu schedule op %ld on "
811 				 "replacement disc %d\n",
812 				__func__, (unsigned long long)sh->sector,
813 				rbi->bi_rw, i);
814 			atomic_inc(&sh->count);
815 			if (use_new_offset(conf, sh))
816 				rbi->bi_sector = (sh->sector
817 						  + rrdev->new_data_offset);
818 			else
819 				rbi->bi_sector = (sh->sector
820 						  + rrdev->data_offset);
821 			rbi->bi_vcnt = 1;
822 			rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
823 			rbi->bi_io_vec[0].bv_offset = 0;
824 			rbi->bi_size = STRIPE_SIZE;
825 			/*
826 			 * If this is discard request, set bi_vcnt 0. We don't
827 			 * want to confuse SCSI because SCSI will replace payload
828 			 */
829 			if (rw & REQ_DISCARD)
830 				rbi->bi_vcnt = 0;
831 			if (conf->mddev->gendisk)
832 				trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
833 						      rbi, disk_devt(conf->mddev->gendisk),
834 						      sh->dev[i].sector);
835 			generic_make_request(rbi);
836 		}
837 		if (!rdev && !rrdev) {
838 			if (rw & WRITE)
839 				set_bit(STRIPE_DEGRADED, &sh->state);
840 			pr_debug("skip op %ld on disc %d for sector %llu\n",
841 				bi->bi_rw, i, (unsigned long long)sh->sector);
842 			clear_bit(R5_LOCKED, &sh->dev[i].flags);
843 			set_bit(STRIPE_HANDLE, &sh->state);
844 		}
845 	}
846 }
847 
848 static struct dma_async_tx_descriptor *
849 async_copy_data(int frombio, struct bio *bio, struct page *page,
850 	sector_t sector, struct dma_async_tx_descriptor *tx)
851 {
852 	struct bio_vec *bvl;
853 	struct page *bio_page;
854 	int i;
855 	int page_offset;
856 	struct async_submit_ctl submit;
857 	enum async_tx_flags flags = 0;
858 
859 	if (bio->bi_sector >= sector)
860 		page_offset = (signed)(bio->bi_sector - sector) * 512;
861 	else
862 		page_offset = (signed)(sector - bio->bi_sector) * -512;
863 
864 	if (frombio)
865 		flags |= ASYNC_TX_FENCE;
866 	init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
867 
868 	bio_for_each_segment(bvl, bio, i) {
869 		int len = bvl->bv_len;
870 		int clen;
871 		int b_offset = 0;
872 
873 		if (page_offset < 0) {
874 			b_offset = -page_offset;
875 			page_offset += b_offset;
876 			len -= b_offset;
877 		}
878 
879 		if (len > 0 && page_offset + len > STRIPE_SIZE)
880 			clen = STRIPE_SIZE - page_offset;
881 		else
882 			clen = len;
883 
884 		if (clen > 0) {
885 			b_offset += bvl->bv_offset;
886 			bio_page = bvl->bv_page;
887 			if (frombio)
888 				tx = async_memcpy(page, bio_page, page_offset,
889 						  b_offset, clen, &submit);
890 			else
891 				tx = async_memcpy(bio_page, page, b_offset,
892 						  page_offset, clen, &submit);
893 		}
894 		/* chain the operations */
895 		submit.depend_tx = tx;
896 
897 		if (clen < len) /* hit end of page */
898 			break;
899 		page_offset +=  len;
900 	}
901 
902 	return tx;
903 }
904 
905 static void ops_complete_biofill(void *stripe_head_ref)
906 {
907 	struct stripe_head *sh = stripe_head_ref;
908 	struct bio *return_bi = NULL;
909 	int i;
910 
911 	pr_debug("%s: stripe %llu\n", __func__,
912 		(unsigned long long)sh->sector);
913 
914 	/* clear completed biofills */
915 	for (i = sh->disks; i--; ) {
916 		struct r5dev *dev = &sh->dev[i];
917 
918 		/* acknowledge completion of a biofill operation */
919 		/* and check if we need to reply to a read request,
920 		 * new R5_Wantfill requests are held off until
921 		 * !STRIPE_BIOFILL_RUN
922 		 */
923 		if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
924 			struct bio *rbi, *rbi2;
925 
926 			BUG_ON(!dev->read);
927 			rbi = dev->read;
928 			dev->read = NULL;
929 			while (rbi && rbi->bi_sector <
930 				dev->sector + STRIPE_SECTORS) {
931 				rbi2 = r5_next_bio(rbi, dev->sector);
932 				if (!raid5_dec_bi_active_stripes(rbi)) {
933 					rbi->bi_next = return_bi;
934 					return_bi = rbi;
935 				}
936 				rbi = rbi2;
937 			}
938 		}
939 	}
940 	clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
941 
942 	return_io(return_bi);
943 
944 	set_bit(STRIPE_HANDLE, &sh->state);
945 	release_stripe(sh);
946 }
947 
948 static void ops_run_biofill(struct stripe_head *sh)
949 {
950 	struct dma_async_tx_descriptor *tx = NULL;
951 	struct async_submit_ctl submit;
952 	int i;
953 
954 	pr_debug("%s: stripe %llu\n", __func__,
955 		(unsigned long long)sh->sector);
956 
957 	for (i = sh->disks; i--; ) {
958 		struct r5dev *dev = &sh->dev[i];
959 		if (test_bit(R5_Wantfill, &dev->flags)) {
960 			struct bio *rbi;
961 			spin_lock_irq(&sh->stripe_lock);
962 			dev->read = rbi = dev->toread;
963 			dev->toread = NULL;
964 			spin_unlock_irq(&sh->stripe_lock);
965 			while (rbi && rbi->bi_sector <
966 				dev->sector + STRIPE_SECTORS) {
967 				tx = async_copy_data(0, rbi, dev->page,
968 					dev->sector, tx);
969 				rbi = r5_next_bio(rbi, dev->sector);
970 			}
971 		}
972 	}
973 
974 	atomic_inc(&sh->count);
975 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
976 	async_trigger_callback(&submit);
977 }
978 
979 static void mark_target_uptodate(struct stripe_head *sh, int target)
980 {
981 	struct r5dev *tgt;
982 
983 	if (target < 0)
984 		return;
985 
986 	tgt = &sh->dev[target];
987 	set_bit(R5_UPTODATE, &tgt->flags);
988 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
989 	clear_bit(R5_Wantcompute, &tgt->flags);
990 }
991 
992 static void ops_complete_compute(void *stripe_head_ref)
993 {
994 	struct stripe_head *sh = stripe_head_ref;
995 
996 	pr_debug("%s: stripe %llu\n", __func__,
997 		(unsigned long long)sh->sector);
998 
999 	/* mark the computed target(s) as uptodate */
1000 	mark_target_uptodate(sh, sh->ops.target);
1001 	mark_target_uptodate(sh, sh->ops.target2);
1002 
1003 	clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1004 	if (sh->check_state == check_state_compute_run)
1005 		sh->check_state = check_state_compute_result;
1006 	set_bit(STRIPE_HANDLE, &sh->state);
1007 	release_stripe(sh);
1008 }
1009 
1010 /* return a pointer to the address conversion region of the scribble buffer */
1011 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1012 				 struct raid5_percpu *percpu)
1013 {
1014 	return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
1015 }
1016 
1017 static struct dma_async_tx_descriptor *
1018 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1019 {
1020 	int disks = sh->disks;
1021 	struct page **xor_srcs = percpu->scribble;
1022 	int target = sh->ops.target;
1023 	struct r5dev *tgt = &sh->dev[target];
1024 	struct page *xor_dest = tgt->page;
1025 	int count = 0;
1026 	struct dma_async_tx_descriptor *tx;
1027 	struct async_submit_ctl submit;
1028 	int i;
1029 
1030 	pr_debug("%s: stripe %llu block: %d\n",
1031 		__func__, (unsigned long long)sh->sector, target);
1032 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1033 
1034 	for (i = disks; i--; )
1035 		if (i != target)
1036 			xor_srcs[count++] = sh->dev[i].page;
1037 
1038 	atomic_inc(&sh->count);
1039 
1040 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1041 			  ops_complete_compute, sh, to_addr_conv(sh, percpu));
1042 	if (unlikely(count == 1))
1043 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1044 	else
1045 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1046 
1047 	return tx;
1048 }
1049 
1050 /* set_syndrome_sources - populate source buffers for gen_syndrome
1051  * @srcs - (struct page *) array of size sh->disks
1052  * @sh - stripe_head to parse
1053  *
1054  * Populates srcs in proper layout order for the stripe and returns the
1055  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1056  * destination buffer is recorded in srcs[count] and the Q destination
1057  * is recorded in srcs[count+1]].
1058  */
1059 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
1060 {
1061 	int disks = sh->disks;
1062 	int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1063 	int d0_idx = raid6_d0(sh);
1064 	int count;
1065 	int i;
1066 
1067 	for (i = 0; i < disks; i++)
1068 		srcs[i] = NULL;
1069 
1070 	count = 0;
1071 	i = d0_idx;
1072 	do {
1073 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1074 
1075 		srcs[slot] = sh->dev[i].page;
1076 		i = raid6_next_disk(i, disks);
1077 	} while (i != d0_idx);
1078 
1079 	return syndrome_disks;
1080 }
1081 
1082 static struct dma_async_tx_descriptor *
1083 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1084 {
1085 	int disks = sh->disks;
1086 	struct page **blocks = percpu->scribble;
1087 	int target;
1088 	int qd_idx = sh->qd_idx;
1089 	struct dma_async_tx_descriptor *tx;
1090 	struct async_submit_ctl submit;
1091 	struct r5dev *tgt;
1092 	struct page *dest;
1093 	int i;
1094 	int count;
1095 
1096 	if (sh->ops.target < 0)
1097 		target = sh->ops.target2;
1098 	else if (sh->ops.target2 < 0)
1099 		target = sh->ops.target;
1100 	else
1101 		/* we should only have one valid target */
1102 		BUG();
1103 	BUG_ON(target < 0);
1104 	pr_debug("%s: stripe %llu block: %d\n",
1105 		__func__, (unsigned long long)sh->sector, target);
1106 
1107 	tgt = &sh->dev[target];
1108 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1109 	dest = tgt->page;
1110 
1111 	atomic_inc(&sh->count);
1112 
1113 	if (target == qd_idx) {
1114 		count = set_syndrome_sources(blocks, sh);
1115 		blocks[count] = NULL; /* regenerating p is not necessary */
1116 		BUG_ON(blocks[count+1] != dest); /* q should already be set */
1117 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1118 				  ops_complete_compute, sh,
1119 				  to_addr_conv(sh, percpu));
1120 		tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1121 	} else {
1122 		/* Compute any data- or p-drive using XOR */
1123 		count = 0;
1124 		for (i = disks; i-- ; ) {
1125 			if (i == target || i == qd_idx)
1126 				continue;
1127 			blocks[count++] = sh->dev[i].page;
1128 		}
1129 
1130 		init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1131 				  NULL, ops_complete_compute, sh,
1132 				  to_addr_conv(sh, percpu));
1133 		tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1134 	}
1135 
1136 	return tx;
1137 }
1138 
1139 static struct dma_async_tx_descriptor *
1140 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1141 {
1142 	int i, count, disks = sh->disks;
1143 	int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1144 	int d0_idx = raid6_d0(sh);
1145 	int faila = -1, failb = -1;
1146 	int target = sh->ops.target;
1147 	int target2 = sh->ops.target2;
1148 	struct r5dev *tgt = &sh->dev[target];
1149 	struct r5dev *tgt2 = &sh->dev[target2];
1150 	struct dma_async_tx_descriptor *tx;
1151 	struct page **blocks = percpu->scribble;
1152 	struct async_submit_ctl submit;
1153 
1154 	pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1155 		 __func__, (unsigned long long)sh->sector, target, target2);
1156 	BUG_ON(target < 0 || target2 < 0);
1157 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1158 	BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1159 
1160 	/* we need to open-code set_syndrome_sources to handle the
1161 	 * slot number conversion for 'faila' and 'failb'
1162 	 */
1163 	for (i = 0; i < disks ; i++)
1164 		blocks[i] = NULL;
1165 	count = 0;
1166 	i = d0_idx;
1167 	do {
1168 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1169 
1170 		blocks[slot] = sh->dev[i].page;
1171 
1172 		if (i == target)
1173 			faila = slot;
1174 		if (i == target2)
1175 			failb = slot;
1176 		i = raid6_next_disk(i, disks);
1177 	} while (i != d0_idx);
1178 
1179 	BUG_ON(faila == failb);
1180 	if (failb < faila)
1181 		swap(faila, failb);
1182 	pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1183 		 __func__, (unsigned long long)sh->sector, faila, failb);
1184 
1185 	atomic_inc(&sh->count);
1186 
1187 	if (failb == syndrome_disks+1) {
1188 		/* Q disk is one of the missing disks */
1189 		if (faila == syndrome_disks) {
1190 			/* Missing P+Q, just recompute */
1191 			init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1192 					  ops_complete_compute, sh,
1193 					  to_addr_conv(sh, percpu));
1194 			return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1195 						  STRIPE_SIZE, &submit);
1196 		} else {
1197 			struct page *dest;
1198 			int data_target;
1199 			int qd_idx = sh->qd_idx;
1200 
1201 			/* Missing D+Q: recompute D from P, then recompute Q */
1202 			if (target == qd_idx)
1203 				data_target = target2;
1204 			else
1205 				data_target = target;
1206 
1207 			count = 0;
1208 			for (i = disks; i-- ; ) {
1209 				if (i == data_target || i == qd_idx)
1210 					continue;
1211 				blocks[count++] = sh->dev[i].page;
1212 			}
1213 			dest = sh->dev[data_target].page;
1214 			init_async_submit(&submit,
1215 					  ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1216 					  NULL, NULL, NULL,
1217 					  to_addr_conv(sh, percpu));
1218 			tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1219 				       &submit);
1220 
1221 			count = set_syndrome_sources(blocks, sh);
1222 			init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1223 					  ops_complete_compute, sh,
1224 					  to_addr_conv(sh, percpu));
1225 			return async_gen_syndrome(blocks, 0, count+2,
1226 						  STRIPE_SIZE, &submit);
1227 		}
1228 	} else {
1229 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1230 				  ops_complete_compute, sh,
1231 				  to_addr_conv(sh, percpu));
1232 		if (failb == syndrome_disks) {
1233 			/* We're missing D+P. */
1234 			return async_raid6_datap_recov(syndrome_disks+2,
1235 						       STRIPE_SIZE, faila,
1236 						       blocks, &submit);
1237 		} else {
1238 			/* We're missing D+D. */
1239 			return async_raid6_2data_recov(syndrome_disks+2,
1240 						       STRIPE_SIZE, faila, failb,
1241 						       blocks, &submit);
1242 		}
1243 	}
1244 }
1245 
1246 
1247 static void ops_complete_prexor(void *stripe_head_ref)
1248 {
1249 	struct stripe_head *sh = stripe_head_ref;
1250 
1251 	pr_debug("%s: stripe %llu\n", __func__,
1252 		(unsigned long long)sh->sector);
1253 }
1254 
1255 static struct dma_async_tx_descriptor *
1256 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1257 	       struct dma_async_tx_descriptor *tx)
1258 {
1259 	int disks = sh->disks;
1260 	struct page **xor_srcs = percpu->scribble;
1261 	int count = 0, pd_idx = sh->pd_idx, i;
1262 	struct async_submit_ctl submit;
1263 
1264 	/* existing parity data subtracted */
1265 	struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1266 
1267 	pr_debug("%s: stripe %llu\n", __func__,
1268 		(unsigned long long)sh->sector);
1269 
1270 	for (i = disks; i--; ) {
1271 		struct r5dev *dev = &sh->dev[i];
1272 		/* Only process blocks that are known to be uptodate */
1273 		if (test_bit(R5_Wantdrain, &dev->flags))
1274 			xor_srcs[count++] = dev->page;
1275 	}
1276 
1277 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1278 			  ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1279 	tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1280 
1281 	return tx;
1282 }
1283 
1284 static struct dma_async_tx_descriptor *
1285 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1286 {
1287 	int disks = sh->disks;
1288 	int i;
1289 
1290 	pr_debug("%s: stripe %llu\n", __func__,
1291 		(unsigned long long)sh->sector);
1292 
1293 	for (i = disks; i--; ) {
1294 		struct r5dev *dev = &sh->dev[i];
1295 		struct bio *chosen;
1296 
1297 		if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1298 			struct bio *wbi;
1299 
1300 			spin_lock_irq(&sh->stripe_lock);
1301 			chosen = dev->towrite;
1302 			dev->towrite = NULL;
1303 			BUG_ON(dev->written);
1304 			wbi = dev->written = chosen;
1305 			spin_unlock_irq(&sh->stripe_lock);
1306 
1307 			while (wbi && wbi->bi_sector <
1308 				dev->sector + STRIPE_SECTORS) {
1309 				if (wbi->bi_rw & REQ_FUA)
1310 					set_bit(R5_WantFUA, &dev->flags);
1311 				if (wbi->bi_rw & REQ_SYNC)
1312 					set_bit(R5_SyncIO, &dev->flags);
1313 				if (wbi->bi_rw & REQ_DISCARD)
1314 					set_bit(R5_Discard, &dev->flags);
1315 				else
1316 					tx = async_copy_data(1, wbi, dev->page,
1317 						dev->sector, tx);
1318 				wbi = r5_next_bio(wbi, dev->sector);
1319 			}
1320 		}
1321 	}
1322 
1323 	return tx;
1324 }
1325 
1326 static void ops_complete_reconstruct(void *stripe_head_ref)
1327 {
1328 	struct stripe_head *sh = stripe_head_ref;
1329 	int disks = sh->disks;
1330 	int pd_idx = sh->pd_idx;
1331 	int qd_idx = sh->qd_idx;
1332 	int i;
1333 	bool fua = false, sync = false, discard = false;
1334 
1335 	pr_debug("%s: stripe %llu\n", __func__,
1336 		(unsigned long long)sh->sector);
1337 
1338 	for (i = disks; i--; ) {
1339 		fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1340 		sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1341 		discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1342 	}
1343 
1344 	for (i = disks; i--; ) {
1345 		struct r5dev *dev = &sh->dev[i];
1346 
1347 		if (dev->written || i == pd_idx || i == qd_idx) {
1348 			if (!discard)
1349 				set_bit(R5_UPTODATE, &dev->flags);
1350 			if (fua)
1351 				set_bit(R5_WantFUA, &dev->flags);
1352 			if (sync)
1353 				set_bit(R5_SyncIO, &dev->flags);
1354 		}
1355 	}
1356 
1357 	if (sh->reconstruct_state == reconstruct_state_drain_run)
1358 		sh->reconstruct_state = reconstruct_state_drain_result;
1359 	else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1360 		sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1361 	else {
1362 		BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1363 		sh->reconstruct_state = reconstruct_state_result;
1364 	}
1365 
1366 	set_bit(STRIPE_HANDLE, &sh->state);
1367 	release_stripe(sh);
1368 }
1369 
1370 static void
1371 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1372 		     struct dma_async_tx_descriptor *tx)
1373 {
1374 	int disks = sh->disks;
1375 	struct page **xor_srcs = percpu->scribble;
1376 	struct async_submit_ctl submit;
1377 	int count = 0, pd_idx = sh->pd_idx, i;
1378 	struct page *xor_dest;
1379 	int prexor = 0;
1380 	unsigned long flags;
1381 
1382 	pr_debug("%s: stripe %llu\n", __func__,
1383 		(unsigned long long)sh->sector);
1384 
1385 	for (i = 0; i < sh->disks; i++) {
1386 		if (pd_idx == i)
1387 			continue;
1388 		if (!test_bit(R5_Discard, &sh->dev[i].flags))
1389 			break;
1390 	}
1391 	if (i >= sh->disks) {
1392 		atomic_inc(&sh->count);
1393 		set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1394 		ops_complete_reconstruct(sh);
1395 		return;
1396 	}
1397 	/* check if prexor is active which means only process blocks
1398 	 * that are part of a read-modify-write (written)
1399 	 */
1400 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1401 		prexor = 1;
1402 		xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1403 		for (i = disks; i--; ) {
1404 			struct r5dev *dev = &sh->dev[i];
1405 			if (dev->written)
1406 				xor_srcs[count++] = dev->page;
1407 		}
1408 	} else {
1409 		xor_dest = sh->dev[pd_idx].page;
1410 		for (i = disks; i--; ) {
1411 			struct r5dev *dev = &sh->dev[i];
1412 			if (i != pd_idx)
1413 				xor_srcs[count++] = dev->page;
1414 		}
1415 	}
1416 
1417 	/* 1/ if we prexor'd then the dest is reused as a source
1418 	 * 2/ if we did not prexor then we are redoing the parity
1419 	 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1420 	 * for the synchronous xor case
1421 	 */
1422 	flags = ASYNC_TX_ACK |
1423 		(prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1424 
1425 	atomic_inc(&sh->count);
1426 
1427 	init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1428 			  to_addr_conv(sh, percpu));
1429 	if (unlikely(count == 1))
1430 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1431 	else
1432 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1433 }
1434 
1435 static void
1436 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1437 		     struct dma_async_tx_descriptor *tx)
1438 {
1439 	struct async_submit_ctl submit;
1440 	struct page **blocks = percpu->scribble;
1441 	int count, i;
1442 
1443 	pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1444 
1445 	for (i = 0; i < sh->disks; i++) {
1446 		if (sh->pd_idx == i || sh->qd_idx == i)
1447 			continue;
1448 		if (!test_bit(R5_Discard, &sh->dev[i].flags))
1449 			break;
1450 	}
1451 	if (i >= sh->disks) {
1452 		atomic_inc(&sh->count);
1453 		set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1454 		set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1455 		ops_complete_reconstruct(sh);
1456 		return;
1457 	}
1458 
1459 	count = set_syndrome_sources(blocks, sh);
1460 
1461 	atomic_inc(&sh->count);
1462 
1463 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1464 			  sh, to_addr_conv(sh, percpu));
1465 	async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1466 }
1467 
1468 static void ops_complete_check(void *stripe_head_ref)
1469 {
1470 	struct stripe_head *sh = stripe_head_ref;
1471 
1472 	pr_debug("%s: stripe %llu\n", __func__,
1473 		(unsigned long long)sh->sector);
1474 
1475 	sh->check_state = check_state_check_result;
1476 	set_bit(STRIPE_HANDLE, &sh->state);
1477 	release_stripe(sh);
1478 }
1479 
1480 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1481 {
1482 	int disks = sh->disks;
1483 	int pd_idx = sh->pd_idx;
1484 	int qd_idx = sh->qd_idx;
1485 	struct page *xor_dest;
1486 	struct page **xor_srcs = percpu->scribble;
1487 	struct dma_async_tx_descriptor *tx;
1488 	struct async_submit_ctl submit;
1489 	int count;
1490 	int i;
1491 
1492 	pr_debug("%s: stripe %llu\n", __func__,
1493 		(unsigned long long)sh->sector);
1494 
1495 	count = 0;
1496 	xor_dest = sh->dev[pd_idx].page;
1497 	xor_srcs[count++] = xor_dest;
1498 	for (i = disks; i--; ) {
1499 		if (i == pd_idx || i == qd_idx)
1500 			continue;
1501 		xor_srcs[count++] = sh->dev[i].page;
1502 	}
1503 
1504 	init_async_submit(&submit, 0, NULL, NULL, NULL,
1505 			  to_addr_conv(sh, percpu));
1506 	tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1507 			   &sh->ops.zero_sum_result, &submit);
1508 
1509 	atomic_inc(&sh->count);
1510 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1511 	tx = async_trigger_callback(&submit);
1512 }
1513 
1514 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1515 {
1516 	struct page **srcs = percpu->scribble;
1517 	struct async_submit_ctl submit;
1518 	int count;
1519 
1520 	pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1521 		(unsigned long long)sh->sector, checkp);
1522 
1523 	count = set_syndrome_sources(srcs, sh);
1524 	if (!checkp)
1525 		srcs[count] = NULL;
1526 
1527 	atomic_inc(&sh->count);
1528 	init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1529 			  sh, to_addr_conv(sh, percpu));
1530 	async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1531 			   &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1532 }
1533 
1534 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1535 {
1536 	int overlap_clear = 0, i, disks = sh->disks;
1537 	struct dma_async_tx_descriptor *tx = NULL;
1538 	struct r5conf *conf = sh->raid_conf;
1539 	int level = conf->level;
1540 	struct raid5_percpu *percpu;
1541 	unsigned long cpu;
1542 
1543 	cpu = get_cpu();
1544 	percpu = per_cpu_ptr(conf->percpu, cpu);
1545 	if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1546 		ops_run_biofill(sh);
1547 		overlap_clear++;
1548 	}
1549 
1550 	if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1551 		if (level < 6)
1552 			tx = ops_run_compute5(sh, percpu);
1553 		else {
1554 			if (sh->ops.target2 < 0 || sh->ops.target < 0)
1555 				tx = ops_run_compute6_1(sh, percpu);
1556 			else
1557 				tx = ops_run_compute6_2(sh, percpu);
1558 		}
1559 		/* terminate the chain if reconstruct is not set to be run */
1560 		if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1561 			async_tx_ack(tx);
1562 	}
1563 
1564 	if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1565 		tx = ops_run_prexor(sh, percpu, tx);
1566 
1567 	if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1568 		tx = ops_run_biodrain(sh, tx);
1569 		overlap_clear++;
1570 	}
1571 
1572 	if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1573 		if (level < 6)
1574 			ops_run_reconstruct5(sh, percpu, tx);
1575 		else
1576 			ops_run_reconstruct6(sh, percpu, tx);
1577 	}
1578 
1579 	if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1580 		if (sh->check_state == check_state_run)
1581 			ops_run_check_p(sh, percpu);
1582 		else if (sh->check_state == check_state_run_q)
1583 			ops_run_check_pq(sh, percpu, 0);
1584 		else if (sh->check_state == check_state_run_pq)
1585 			ops_run_check_pq(sh, percpu, 1);
1586 		else
1587 			BUG();
1588 	}
1589 
1590 	if (overlap_clear)
1591 		for (i = disks; i--; ) {
1592 			struct r5dev *dev = &sh->dev[i];
1593 			if (test_and_clear_bit(R5_Overlap, &dev->flags))
1594 				wake_up(&sh->raid_conf->wait_for_overlap);
1595 		}
1596 	put_cpu();
1597 }
1598 
1599 static int grow_one_stripe(struct r5conf *conf)
1600 {
1601 	struct stripe_head *sh;
1602 	sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1603 	if (!sh)
1604 		return 0;
1605 
1606 	sh->raid_conf = conf;
1607 
1608 	spin_lock_init(&sh->stripe_lock);
1609 
1610 	if (grow_buffers(sh)) {
1611 		shrink_buffers(sh);
1612 		kmem_cache_free(conf->slab_cache, sh);
1613 		return 0;
1614 	}
1615 	/* we just created an active stripe so... */
1616 	atomic_set(&sh->count, 1);
1617 	atomic_inc(&conf->active_stripes);
1618 	INIT_LIST_HEAD(&sh->lru);
1619 	release_stripe(sh);
1620 	return 1;
1621 }
1622 
1623 static int grow_stripes(struct r5conf *conf, int num)
1624 {
1625 	struct kmem_cache *sc;
1626 	int devs = max(conf->raid_disks, conf->previous_raid_disks);
1627 
1628 	if (conf->mddev->gendisk)
1629 		sprintf(conf->cache_name[0],
1630 			"raid%d-%s", conf->level, mdname(conf->mddev));
1631 	else
1632 		sprintf(conf->cache_name[0],
1633 			"raid%d-%p", conf->level, conf->mddev);
1634 	sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1635 
1636 	conf->active_name = 0;
1637 	sc = kmem_cache_create(conf->cache_name[conf->active_name],
1638 			       sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1639 			       0, 0, NULL);
1640 	if (!sc)
1641 		return 1;
1642 	conf->slab_cache = sc;
1643 	conf->pool_size = devs;
1644 	while (num--)
1645 		if (!grow_one_stripe(conf))
1646 			return 1;
1647 	return 0;
1648 }
1649 
1650 /**
1651  * scribble_len - return the required size of the scribble region
1652  * @num - total number of disks in the array
1653  *
1654  * The size must be enough to contain:
1655  * 1/ a struct page pointer for each device in the array +2
1656  * 2/ room to convert each entry in (1) to its corresponding dma
1657  *    (dma_map_page()) or page (page_address()) address.
1658  *
1659  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1660  * calculate over all devices (not just the data blocks), using zeros in place
1661  * of the P and Q blocks.
1662  */
1663 static size_t scribble_len(int num)
1664 {
1665 	size_t len;
1666 
1667 	len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1668 
1669 	return len;
1670 }
1671 
1672 static int resize_stripes(struct r5conf *conf, int newsize)
1673 {
1674 	/* Make all the stripes able to hold 'newsize' devices.
1675 	 * New slots in each stripe get 'page' set to a new page.
1676 	 *
1677 	 * This happens in stages:
1678 	 * 1/ create a new kmem_cache and allocate the required number of
1679 	 *    stripe_heads.
1680 	 * 2/ gather all the old stripe_heads and transfer the pages across
1681 	 *    to the new stripe_heads.  This will have the side effect of
1682 	 *    freezing the array as once all stripe_heads have been collected,
1683 	 *    no IO will be possible.  Old stripe heads are freed once their
1684 	 *    pages have been transferred over, and the old kmem_cache is
1685 	 *    freed when all stripes are done.
1686 	 * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1687 	 *    we simple return a failre status - no need to clean anything up.
1688 	 * 4/ allocate new pages for the new slots in the new stripe_heads.
1689 	 *    If this fails, we don't bother trying the shrink the
1690 	 *    stripe_heads down again, we just leave them as they are.
1691 	 *    As each stripe_head is processed the new one is released into
1692 	 *    active service.
1693 	 *
1694 	 * Once step2 is started, we cannot afford to wait for a write,
1695 	 * so we use GFP_NOIO allocations.
1696 	 */
1697 	struct stripe_head *osh, *nsh;
1698 	LIST_HEAD(newstripes);
1699 	struct disk_info *ndisks;
1700 	unsigned long cpu;
1701 	int err;
1702 	struct kmem_cache *sc;
1703 	int i;
1704 
1705 	if (newsize <= conf->pool_size)
1706 		return 0; /* never bother to shrink */
1707 
1708 	err = md_allow_write(conf->mddev);
1709 	if (err)
1710 		return err;
1711 
1712 	/* Step 1 */
1713 	sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1714 			       sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1715 			       0, 0, NULL);
1716 	if (!sc)
1717 		return -ENOMEM;
1718 
1719 	for (i = conf->max_nr_stripes; i; i--) {
1720 		nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1721 		if (!nsh)
1722 			break;
1723 
1724 		nsh->raid_conf = conf;
1725 		spin_lock_init(&nsh->stripe_lock);
1726 
1727 		list_add(&nsh->lru, &newstripes);
1728 	}
1729 	if (i) {
1730 		/* didn't get enough, give up */
1731 		while (!list_empty(&newstripes)) {
1732 			nsh = list_entry(newstripes.next, struct stripe_head, lru);
1733 			list_del(&nsh->lru);
1734 			kmem_cache_free(sc, nsh);
1735 		}
1736 		kmem_cache_destroy(sc);
1737 		return -ENOMEM;
1738 	}
1739 	/* Step 2 - Must use GFP_NOIO now.
1740 	 * OK, we have enough stripes, start collecting inactive
1741 	 * stripes and copying them over
1742 	 */
1743 	list_for_each_entry(nsh, &newstripes, lru) {
1744 		spin_lock_irq(&conf->device_lock);
1745 		wait_event_lock_irq(conf->wait_for_stripe,
1746 				    !list_empty(&conf->inactive_list),
1747 				    conf->device_lock);
1748 		osh = get_free_stripe(conf);
1749 		spin_unlock_irq(&conf->device_lock);
1750 		atomic_set(&nsh->count, 1);
1751 		for(i=0; i<conf->pool_size; i++)
1752 			nsh->dev[i].page = osh->dev[i].page;
1753 		for( ; i<newsize; i++)
1754 			nsh->dev[i].page = NULL;
1755 		kmem_cache_free(conf->slab_cache, osh);
1756 	}
1757 	kmem_cache_destroy(conf->slab_cache);
1758 
1759 	/* Step 3.
1760 	 * At this point, we are holding all the stripes so the array
1761 	 * is completely stalled, so now is a good time to resize
1762 	 * conf->disks and the scribble region
1763 	 */
1764 	ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1765 	if (ndisks) {
1766 		for (i=0; i<conf->raid_disks; i++)
1767 			ndisks[i] = conf->disks[i];
1768 		kfree(conf->disks);
1769 		conf->disks = ndisks;
1770 	} else
1771 		err = -ENOMEM;
1772 
1773 	get_online_cpus();
1774 	conf->scribble_len = scribble_len(newsize);
1775 	for_each_present_cpu(cpu) {
1776 		struct raid5_percpu *percpu;
1777 		void *scribble;
1778 
1779 		percpu = per_cpu_ptr(conf->percpu, cpu);
1780 		scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1781 
1782 		if (scribble) {
1783 			kfree(percpu->scribble);
1784 			percpu->scribble = scribble;
1785 		} else {
1786 			err = -ENOMEM;
1787 			break;
1788 		}
1789 	}
1790 	put_online_cpus();
1791 
1792 	/* Step 4, return new stripes to service */
1793 	while(!list_empty(&newstripes)) {
1794 		nsh = list_entry(newstripes.next, struct stripe_head, lru);
1795 		list_del_init(&nsh->lru);
1796 
1797 		for (i=conf->raid_disks; i < newsize; i++)
1798 			if (nsh->dev[i].page == NULL) {
1799 				struct page *p = alloc_page(GFP_NOIO);
1800 				nsh->dev[i].page = p;
1801 				if (!p)
1802 					err = -ENOMEM;
1803 			}
1804 		release_stripe(nsh);
1805 	}
1806 	/* critical section pass, GFP_NOIO no longer needed */
1807 
1808 	conf->slab_cache = sc;
1809 	conf->active_name = 1-conf->active_name;
1810 	conf->pool_size = newsize;
1811 	return err;
1812 }
1813 
1814 static int drop_one_stripe(struct r5conf *conf)
1815 {
1816 	struct stripe_head *sh;
1817 
1818 	spin_lock_irq(&conf->device_lock);
1819 	sh = get_free_stripe(conf);
1820 	spin_unlock_irq(&conf->device_lock);
1821 	if (!sh)
1822 		return 0;
1823 	BUG_ON(atomic_read(&sh->count));
1824 	shrink_buffers(sh);
1825 	kmem_cache_free(conf->slab_cache, sh);
1826 	atomic_dec(&conf->active_stripes);
1827 	return 1;
1828 }
1829 
1830 static void shrink_stripes(struct r5conf *conf)
1831 {
1832 	while (drop_one_stripe(conf))
1833 		;
1834 
1835 	if (conf->slab_cache)
1836 		kmem_cache_destroy(conf->slab_cache);
1837 	conf->slab_cache = NULL;
1838 }
1839 
1840 static void raid5_end_read_request(struct bio * bi, int error)
1841 {
1842 	struct stripe_head *sh = bi->bi_private;
1843 	struct r5conf *conf = sh->raid_conf;
1844 	int disks = sh->disks, i;
1845 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1846 	char b[BDEVNAME_SIZE];
1847 	struct md_rdev *rdev = NULL;
1848 	sector_t s;
1849 
1850 	for (i=0 ; i<disks; i++)
1851 		if (bi == &sh->dev[i].req)
1852 			break;
1853 
1854 	pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1855 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
1856 		uptodate);
1857 	if (i == disks) {
1858 		BUG();
1859 		return;
1860 	}
1861 	if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1862 		/* If replacement finished while this request was outstanding,
1863 		 * 'replacement' might be NULL already.
1864 		 * In that case it moved down to 'rdev'.
1865 		 * rdev is not removed until all requests are finished.
1866 		 */
1867 		rdev = conf->disks[i].replacement;
1868 	if (!rdev)
1869 		rdev = conf->disks[i].rdev;
1870 
1871 	if (use_new_offset(conf, sh))
1872 		s = sh->sector + rdev->new_data_offset;
1873 	else
1874 		s = sh->sector + rdev->data_offset;
1875 	if (uptodate) {
1876 		set_bit(R5_UPTODATE, &sh->dev[i].flags);
1877 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1878 			/* Note that this cannot happen on a
1879 			 * replacement device.  We just fail those on
1880 			 * any error
1881 			 */
1882 			printk_ratelimited(
1883 				KERN_INFO
1884 				"md/raid:%s: read error corrected"
1885 				" (%lu sectors at %llu on %s)\n",
1886 				mdname(conf->mddev), STRIPE_SECTORS,
1887 				(unsigned long long)s,
1888 				bdevname(rdev->bdev, b));
1889 			atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1890 			clear_bit(R5_ReadError, &sh->dev[i].flags);
1891 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
1892 		} else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1893 			clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1894 
1895 		if (atomic_read(&rdev->read_errors))
1896 			atomic_set(&rdev->read_errors, 0);
1897 	} else {
1898 		const char *bdn = bdevname(rdev->bdev, b);
1899 		int retry = 0;
1900 		int set_bad = 0;
1901 
1902 		clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1903 		atomic_inc(&rdev->read_errors);
1904 		if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1905 			printk_ratelimited(
1906 				KERN_WARNING
1907 				"md/raid:%s: read error on replacement device "
1908 				"(sector %llu on %s).\n",
1909 				mdname(conf->mddev),
1910 				(unsigned long long)s,
1911 				bdn);
1912 		else if (conf->mddev->degraded >= conf->max_degraded) {
1913 			set_bad = 1;
1914 			printk_ratelimited(
1915 				KERN_WARNING
1916 				"md/raid:%s: read error not correctable "
1917 				"(sector %llu on %s).\n",
1918 				mdname(conf->mddev),
1919 				(unsigned long long)s,
1920 				bdn);
1921 		} else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1922 			/* Oh, no!!! */
1923 			set_bad = 1;
1924 			printk_ratelimited(
1925 				KERN_WARNING
1926 				"md/raid:%s: read error NOT corrected!! "
1927 				"(sector %llu on %s).\n",
1928 				mdname(conf->mddev),
1929 				(unsigned long long)s,
1930 				bdn);
1931 		} else if (atomic_read(&rdev->read_errors)
1932 			 > conf->max_nr_stripes)
1933 			printk(KERN_WARNING
1934 			       "md/raid:%s: Too many read errors, failing device %s.\n",
1935 			       mdname(conf->mddev), bdn);
1936 		else
1937 			retry = 1;
1938 		if (retry)
1939 			if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1940 				set_bit(R5_ReadError, &sh->dev[i].flags);
1941 				clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1942 			} else
1943 				set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1944 		else {
1945 			clear_bit(R5_ReadError, &sh->dev[i].flags);
1946 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
1947 			if (!(set_bad
1948 			      && test_bit(In_sync, &rdev->flags)
1949 			      && rdev_set_badblocks(
1950 				      rdev, sh->sector, STRIPE_SECTORS, 0)))
1951 				md_error(conf->mddev, rdev);
1952 		}
1953 	}
1954 	rdev_dec_pending(rdev, conf->mddev);
1955 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
1956 	set_bit(STRIPE_HANDLE, &sh->state);
1957 	release_stripe(sh);
1958 }
1959 
1960 static void raid5_end_write_request(struct bio *bi, int error)
1961 {
1962 	struct stripe_head *sh = bi->bi_private;
1963 	struct r5conf *conf = sh->raid_conf;
1964 	int disks = sh->disks, i;
1965 	struct md_rdev *uninitialized_var(rdev);
1966 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1967 	sector_t first_bad;
1968 	int bad_sectors;
1969 	int replacement = 0;
1970 
1971 	for (i = 0 ; i < disks; i++) {
1972 		if (bi == &sh->dev[i].req) {
1973 			rdev = conf->disks[i].rdev;
1974 			break;
1975 		}
1976 		if (bi == &sh->dev[i].rreq) {
1977 			rdev = conf->disks[i].replacement;
1978 			if (rdev)
1979 				replacement = 1;
1980 			else
1981 				/* rdev was removed and 'replacement'
1982 				 * replaced it.  rdev is not removed
1983 				 * until all requests are finished.
1984 				 */
1985 				rdev = conf->disks[i].rdev;
1986 			break;
1987 		}
1988 	}
1989 	pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1990 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
1991 		uptodate);
1992 	if (i == disks) {
1993 		BUG();
1994 		return;
1995 	}
1996 
1997 	if (replacement) {
1998 		if (!uptodate)
1999 			md_error(conf->mddev, rdev);
2000 		else if (is_badblock(rdev, sh->sector,
2001 				     STRIPE_SECTORS,
2002 				     &first_bad, &bad_sectors))
2003 			set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2004 	} else {
2005 		if (!uptodate) {
2006 			set_bit(WriteErrorSeen, &rdev->flags);
2007 			set_bit(R5_WriteError, &sh->dev[i].flags);
2008 			if (!test_and_set_bit(WantReplacement, &rdev->flags))
2009 				set_bit(MD_RECOVERY_NEEDED,
2010 					&rdev->mddev->recovery);
2011 		} else if (is_badblock(rdev, sh->sector,
2012 				       STRIPE_SECTORS,
2013 				       &first_bad, &bad_sectors)) {
2014 			set_bit(R5_MadeGood, &sh->dev[i].flags);
2015 			if (test_bit(R5_ReadError, &sh->dev[i].flags))
2016 				/* That was a successful write so make
2017 				 * sure it looks like we already did
2018 				 * a re-write.
2019 				 */
2020 				set_bit(R5_ReWrite, &sh->dev[i].flags);
2021 		}
2022 	}
2023 	rdev_dec_pending(rdev, conf->mddev);
2024 
2025 	if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2026 		clear_bit(R5_LOCKED, &sh->dev[i].flags);
2027 	set_bit(STRIPE_HANDLE, &sh->state);
2028 	release_stripe(sh);
2029 }
2030 
2031 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2032 
2033 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2034 {
2035 	struct r5dev *dev = &sh->dev[i];
2036 
2037 	bio_init(&dev->req);
2038 	dev->req.bi_io_vec = &dev->vec;
2039 	dev->req.bi_vcnt++;
2040 	dev->req.bi_max_vecs++;
2041 	dev->req.bi_private = sh;
2042 	dev->vec.bv_page = dev->page;
2043 
2044 	bio_init(&dev->rreq);
2045 	dev->rreq.bi_io_vec = &dev->rvec;
2046 	dev->rreq.bi_vcnt++;
2047 	dev->rreq.bi_max_vecs++;
2048 	dev->rreq.bi_private = sh;
2049 	dev->rvec.bv_page = dev->page;
2050 
2051 	dev->flags = 0;
2052 	dev->sector = compute_blocknr(sh, i, previous);
2053 }
2054 
2055 static void error(struct mddev *mddev, struct md_rdev *rdev)
2056 {
2057 	char b[BDEVNAME_SIZE];
2058 	struct r5conf *conf = mddev->private;
2059 	unsigned long flags;
2060 	pr_debug("raid456: error called\n");
2061 
2062 	spin_lock_irqsave(&conf->device_lock, flags);
2063 	clear_bit(In_sync, &rdev->flags);
2064 	mddev->degraded = calc_degraded(conf);
2065 	spin_unlock_irqrestore(&conf->device_lock, flags);
2066 	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2067 
2068 	set_bit(Blocked, &rdev->flags);
2069 	set_bit(Faulty, &rdev->flags);
2070 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
2071 	printk(KERN_ALERT
2072 	       "md/raid:%s: Disk failure on %s, disabling device.\n"
2073 	       "md/raid:%s: Operation continuing on %d devices.\n",
2074 	       mdname(mddev),
2075 	       bdevname(rdev->bdev, b),
2076 	       mdname(mddev),
2077 	       conf->raid_disks - mddev->degraded);
2078 }
2079 
2080 /*
2081  * Input: a 'big' sector number,
2082  * Output: index of the data and parity disk, and the sector # in them.
2083  */
2084 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2085 				     int previous, int *dd_idx,
2086 				     struct stripe_head *sh)
2087 {
2088 	sector_t stripe, stripe2;
2089 	sector_t chunk_number;
2090 	unsigned int chunk_offset;
2091 	int pd_idx, qd_idx;
2092 	int ddf_layout = 0;
2093 	sector_t new_sector;
2094 	int algorithm = previous ? conf->prev_algo
2095 				 : conf->algorithm;
2096 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2097 					 : conf->chunk_sectors;
2098 	int raid_disks = previous ? conf->previous_raid_disks
2099 				  : conf->raid_disks;
2100 	int data_disks = raid_disks - conf->max_degraded;
2101 
2102 	/* First compute the information on this sector */
2103 
2104 	/*
2105 	 * Compute the chunk number and the sector offset inside the chunk
2106 	 */
2107 	chunk_offset = sector_div(r_sector, sectors_per_chunk);
2108 	chunk_number = r_sector;
2109 
2110 	/*
2111 	 * Compute the stripe number
2112 	 */
2113 	stripe = chunk_number;
2114 	*dd_idx = sector_div(stripe, data_disks);
2115 	stripe2 = stripe;
2116 	/*
2117 	 * Select the parity disk based on the user selected algorithm.
2118 	 */
2119 	pd_idx = qd_idx = -1;
2120 	switch(conf->level) {
2121 	case 4:
2122 		pd_idx = data_disks;
2123 		break;
2124 	case 5:
2125 		switch (algorithm) {
2126 		case ALGORITHM_LEFT_ASYMMETRIC:
2127 			pd_idx = data_disks - sector_div(stripe2, raid_disks);
2128 			if (*dd_idx >= pd_idx)
2129 				(*dd_idx)++;
2130 			break;
2131 		case ALGORITHM_RIGHT_ASYMMETRIC:
2132 			pd_idx = sector_div(stripe2, raid_disks);
2133 			if (*dd_idx >= pd_idx)
2134 				(*dd_idx)++;
2135 			break;
2136 		case ALGORITHM_LEFT_SYMMETRIC:
2137 			pd_idx = data_disks - sector_div(stripe2, raid_disks);
2138 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2139 			break;
2140 		case ALGORITHM_RIGHT_SYMMETRIC:
2141 			pd_idx = sector_div(stripe2, raid_disks);
2142 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2143 			break;
2144 		case ALGORITHM_PARITY_0:
2145 			pd_idx = 0;
2146 			(*dd_idx)++;
2147 			break;
2148 		case ALGORITHM_PARITY_N:
2149 			pd_idx = data_disks;
2150 			break;
2151 		default:
2152 			BUG();
2153 		}
2154 		break;
2155 	case 6:
2156 
2157 		switch (algorithm) {
2158 		case ALGORITHM_LEFT_ASYMMETRIC:
2159 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2160 			qd_idx = pd_idx + 1;
2161 			if (pd_idx == raid_disks-1) {
2162 				(*dd_idx)++;	/* Q D D D P */
2163 				qd_idx = 0;
2164 			} else if (*dd_idx >= pd_idx)
2165 				(*dd_idx) += 2; /* D D P Q D */
2166 			break;
2167 		case ALGORITHM_RIGHT_ASYMMETRIC:
2168 			pd_idx = sector_div(stripe2, raid_disks);
2169 			qd_idx = pd_idx + 1;
2170 			if (pd_idx == raid_disks-1) {
2171 				(*dd_idx)++;	/* Q D D D P */
2172 				qd_idx = 0;
2173 			} else if (*dd_idx >= pd_idx)
2174 				(*dd_idx) += 2; /* D D P Q D */
2175 			break;
2176 		case ALGORITHM_LEFT_SYMMETRIC:
2177 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2178 			qd_idx = (pd_idx + 1) % raid_disks;
2179 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2180 			break;
2181 		case ALGORITHM_RIGHT_SYMMETRIC:
2182 			pd_idx = sector_div(stripe2, raid_disks);
2183 			qd_idx = (pd_idx + 1) % raid_disks;
2184 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2185 			break;
2186 
2187 		case ALGORITHM_PARITY_0:
2188 			pd_idx = 0;
2189 			qd_idx = 1;
2190 			(*dd_idx) += 2;
2191 			break;
2192 		case ALGORITHM_PARITY_N:
2193 			pd_idx = data_disks;
2194 			qd_idx = data_disks + 1;
2195 			break;
2196 
2197 		case ALGORITHM_ROTATING_ZERO_RESTART:
2198 			/* Exactly the same as RIGHT_ASYMMETRIC, but or
2199 			 * of blocks for computing Q is different.
2200 			 */
2201 			pd_idx = sector_div(stripe2, raid_disks);
2202 			qd_idx = pd_idx + 1;
2203 			if (pd_idx == raid_disks-1) {
2204 				(*dd_idx)++;	/* Q D D D P */
2205 				qd_idx = 0;
2206 			} else if (*dd_idx >= pd_idx)
2207 				(*dd_idx) += 2; /* D D P Q D */
2208 			ddf_layout = 1;
2209 			break;
2210 
2211 		case ALGORITHM_ROTATING_N_RESTART:
2212 			/* Same a left_asymmetric, by first stripe is
2213 			 * D D D P Q  rather than
2214 			 * Q D D D P
2215 			 */
2216 			stripe2 += 1;
2217 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2218 			qd_idx = pd_idx + 1;
2219 			if (pd_idx == raid_disks-1) {
2220 				(*dd_idx)++;	/* Q D D D P */
2221 				qd_idx = 0;
2222 			} else if (*dd_idx >= pd_idx)
2223 				(*dd_idx) += 2; /* D D P Q D */
2224 			ddf_layout = 1;
2225 			break;
2226 
2227 		case ALGORITHM_ROTATING_N_CONTINUE:
2228 			/* Same as left_symmetric but Q is before P */
2229 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2230 			qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2231 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2232 			ddf_layout = 1;
2233 			break;
2234 
2235 		case ALGORITHM_LEFT_ASYMMETRIC_6:
2236 			/* RAID5 left_asymmetric, with Q on last device */
2237 			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2238 			if (*dd_idx >= pd_idx)
2239 				(*dd_idx)++;
2240 			qd_idx = raid_disks - 1;
2241 			break;
2242 
2243 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
2244 			pd_idx = sector_div(stripe2, raid_disks-1);
2245 			if (*dd_idx >= pd_idx)
2246 				(*dd_idx)++;
2247 			qd_idx = raid_disks - 1;
2248 			break;
2249 
2250 		case ALGORITHM_LEFT_SYMMETRIC_6:
2251 			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2252 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2253 			qd_idx = raid_disks - 1;
2254 			break;
2255 
2256 		case ALGORITHM_RIGHT_SYMMETRIC_6:
2257 			pd_idx = sector_div(stripe2, raid_disks-1);
2258 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2259 			qd_idx = raid_disks - 1;
2260 			break;
2261 
2262 		case ALGORITHM_PARITY_0_6:
2263 			pd_idx = 0;
2264 			(*dd_idx)++;
2265 			qd_idx = raid_disks - 1;
2266 			break;
2267 
2268 		default:
2269 			BUG();
2270 		}
2271 		break;
2272 	}
2273 
2274 	if (sh) {
2275 		sh->pd_idx = pd_idx;
2276 		sh->qd_idx = qd_idx;
2277 		sh->ddf_layout = ddf_layout;
2278 	}
2279 	/*
2280 	 * Finally, compute the new sector number
2281 	 */
2282 	new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2283 	return new_sector;
2284 }
2285 
2286 
2287 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2288 {
2289 	struct r5conf *conf = sh->raid_conf;
2290 	int raid_disks = sh->disks;
2291 	int data_disks = raid_disks - conf->max_degraded;
2292 	sector_t new_sector = sh->sector, check;
2293 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2294 					 : conf->chunk_sectors;
2295 	int algorithm = previous ? conf->prev_algo
2296 				 : conf->algorithm;
2297 	sector_t stripe;
2298 	int chunk_offset;
2299 	sector_t chunk_number;
2300 	int dummy1, dd_idx = i;
2301 	sector_t r_sector;
2302 	struct stripe_head sh2;
2303 
2304 
2305 	chunk_offset = sector_div(new_sector, sectors_per_chunk);
2306 	stripe = new_sector;
2307 
2308 	if (i == sh->pd_idx)
2309 		return 0;
2310 	switch(conf->level) {
2311 	case 4: break;
2312 	case 5:
2313 		switch (algorithm) {
2314 		case ALGORITHM_LEFT_ASYMMETRIC:
2315 		case ALGORITHM_RIGHT_ASYMMETRIC:
2316 			if (i > sh->pd_idx)
2317 				i--;
2318 			break;
2319 		case ALGORITHM_LEFT_SYMMETRIC:
2320 		case ALGORITHM_RIGHT_SYMMETRIC:
2321 			if (i < sh->pd_idx)
2322 				i += raid_disks;
2323 			i -= (sh->pd_idx + 1);
2324 			break;
2325 		case ALGORITHM_PARITY_0:
2326 			i -= 1;
2327 			break;
2328 		case ALGORITHM_PARITY_N:
2329 			break;
2330 		default:
2331 			BUG();
2332 		}
2333 		break;
2334 	case 6:
2335 		if (i == sh->qd_idx)
2336 			return 0; /* It is the Q disk */
2337 		switch (algorithm) {
2338 		case ALGORITHM_LEFT_ASYMMETRIC:
2339 		case ALGORITHM_RIGHT_ASYMMETRIC:
2340 		case ALGORITHM_ROTATING_ZERO_RESTART:
2341 		case ALGORITHM_ROTATING_N_RESTART:
2342 			if (sh->pd_idx == raid_disks-1)
2343 				i--;	/* Q D D D P */
2344 			else if (i > sh->pd_idx)
2345 				i -= 2; /* D D P Q D */
2346 			break;
2347 		case ALGORITHM_LEFT_SYMMETRIC:
2348 		case ALGORITHM_RIGHT_SYMMETRIC:
2349 			if (sh->pd_idx == raid_disks-1)
2350 				i--; /* Q D D D P */
2351 			else {
2352 				/* D D P Q D */
2353 				if (i < sh->pd_idx)
2354 					i += raid_disks;
2355 				i -= (sh->pd_idx + 2);
2356 			}
2357 			break;
2358 		case ALGORITHM_PARITY_0:
2359 			i -= 2;
2360 			break;
2361 		case ALGORITHM_PARITY_N:
2362 			break;
2363 		case ALGORITHM_ROTATING_N_CONTINUE:
2364 			/* Like left_symmetric, but P is before Q */
2365 			if (sh->pd_idx == 0)
2366 				i--;	/* P D D D Q */
2367 			else {
2368 				/* D D Q P D */
2369 				if (i < sh->pd_idx)
2370 					i += raid_disks;
2371 				i -= (sh->pd_idx + 1);
2372 			}
2373 			break;
2374 		case ALGORITHM_LEFT_ASYMMETRIC_6:
2375 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
2376 			if (i > sh->pd_idx)
2377 				i--;
2378 			break;
2379 		case ALGORITHM_LEFT_SYMMETRIC_6:
2380 		case ALGORITHM_RIGHT_SYMMETRIC_6:
2381 			if (i < sh->pd_idx)
2382 				i += data_disks + 1;
2383 			i -= (sh->pd_idx + 1);
2384 			break;
2385 		case ALGORITHM_PARITY_0_6:
2386 			i -= 1;
2387 			break;
2388 		default:
2389 			BUG();
2390 		}
2391 		break;
2392 	}
2393 
2394 	chunk_number = stripe * data_disks + i;
2395 	r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2396 
2397 	check = raid5_compute_sector(conf, r_sector,
2398 				     previous, &dummy1, &sh2);
2399 	if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2400 		|| sh2.qd_idx != sh->qd_idx) {
2401 		printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2402 		       mdname(conf->mddev));
2403 		return 0;
2404 	}
2405 	return r_sector;
2406 }
2407 
2408 
2409 static void
2410 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2411 			 int rcw, int expand)
2412 {
2413 	int i, pd_idx = sh->pd_idx, disks = sh->disks;
2414 	struct r5conf *conf = sh->raid_conf;
2415 	int level = conf->level;
2416 
2417 	if (rcw) {
2418 
2419 		for (i = disks; i--; ) {
2420 			struct r5dev *dev = &sh->dev[i];
2421 
2422 			if (dev->towrite) {
2423 				set_bit(R5_LOCKED, &dev->flags);
2424 				set_bit(R5_Wantdrain, &dev->flags);
2425 				if (!expand)
2426 					clear_bit(R5_UPTODATE, &dev->flags);
2427 				s->locked++;
2428 			}
2429 		}
2430 		/* if we are not expanding this is a proper write request, and
2431 		 * there will be bios with new data to be drained into the
2432 		 * stripe cache
2433 		 */
2434 		if (!expand) {
2435 			if (!s->locked)
2436 				/* False alarm, nothing to do */
2437 				return;
2438 			sh->reconstruct_state = reconstruct_state_drain_run;
2439 			set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2440 		} else
2441 			sh->reconstruct_state = reconstruct_state_run;
2442 
2443 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2444 
2445 		if (s->locked + conf->max_degraded == disks)
2446 			if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2447 				atomic_inc(&conf->pending_full_writes);
2448 	} else {
2449 		BUG_ON(level == 6);
2450 		BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2451 			test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2452 
2453 		for (i = disks; i--; ) {
2454 			struct r5dev *dev = &sh->dev[i];
2455 			if (i == pd_idx)
2456 				continue;
2457 
2458 			if (dev->towrite &&
2459 			    (test_bit(R5_UPTODATE, &dev->flags) ||
2460 			     test_bit(R5_Wantcompute, &dev->flags))) {
2461 				set_bit(R5_Wantdrain, &dev->flags);
2462 				set_bit(R5_LOCKED, &dev->flags);
2463 				clear_bit(R5_UPTODATE, &dev->flags);
2464 				s->locked++;
2465 			}
2466 		}
2467 		if (!s->locked)
2468 			/* False alarm - nothing to do */
2469 			return;
2470 		sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2471 		set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2472 		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2473 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2474 	}
2475 
2476 	/* keep the parity disk(s) locked while asynchronous operations
2477 	 * are in flight
2478 	 */
2479 	set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2480 	clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2481 	s->locked++;
2482 
2483 	if (level == 6) {
2484 		int qd_idx = sh->qd_idx;
2485 		struct r5dev *dev = &sh->dev[qd_idx];
2486 
2487 		set_bit(R5_LOCKED, &dev->flags);
2488 		clear_bit(R5_UPTODATE, &dev->flags);
2489 		s->locked++;
2490 	}
2491 
2492 	pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2493 		__func__, (unsigned long long)sh->sector,
2494 		s->locked, s->ops_request);
2495 }
2496 
2497 /*
2498  * Each stripe/dev can have one or more bion attached.
2499  * toread/towrite point to the first in a chain.
2500  * The bi_next chain must be in order.
2501  */
2502 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2503 {
2504 	struct bio **bip;
2505 	struct r5conf *conf = sh->raid_conf;
2506 	int firstwrite=0;
2507 
2508 	pr_debug("adding bi b#%llu to stripe s#%llu\n",
2509 		(unsigned long long)bi->bi_sector,
2510 		(unsigned long long)sh->sector);
2511 
2512 	/*
2513 	 * If several bio share a stripe. The bio bi_phys_segments acts as a
2514 	 * reference count to avoid race. The reference count should already be
2515 	 * increased before this function is called (for example, in
2516 	 * make_request()), so other bio sharing this stripe will not free the
2517 	 * stripe. If a stripe is owned by one stripe, the stripe lock will
2518 	 * protect it.
2519 	 */
2520 	spin_lock_irq(&sh->stripe_lock);
2521 	if (forwrite) {
2522 		bip = &sh->dev[dd_idx].towrite;
2523 		if (*bip == NULL)
2524 			firstwrite = 1;
2525 	} else
2526 		bip = &sh->dev[dd_idx].toread;
2527 	while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2528 		if (bio_end_sector(*bip) > bi->bi_sector)
2529 			goto overlap;
2530 		bip = & (*bip)->bi_next;
2531 	}
2532 	if (*bip && (*bip)->bi_sector < bio_end_sector(bi))
2533 		goto overlap;
2534 
2535 	BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2536 	if (*bip)
2537 		bi->bi_next = *bip;
2538 	*bip = bi;
2539 	raid5_inc_bi_active_stripes(bi);
2540 
2541 	if (forwrite) {
2542 		/* check if page is covered */
2543 		sector_t sector = sh->dev[dd_idx].sector;
2544 		for (bi=sh->dev[dd_idx].towrite;
2545 		     sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2546 			     bi && bi->bi_sector <= sector;
2547 		     bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2548 			if (bio_end_sector(bi) >= sector)
2549 				sector = bio_end_sector(bi);
2550 		}
2551 		if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2552 			set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2553 	}
2554 
2555 	pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2556 		(unsigned long long)(*bip)->bi_sector,
2557 		(unsigned long long)sh->sector, dd_idx);
2558 	spin_unlock_irq(&sh->stripe_lock);
2559 
2560 	if (conf->mddev->bitmap && firstwrite) {
2561 		bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2562 				  STRIPE_SECTORS, 0);
2563 		sh->bm_seq = conf->seq_flush+1;
2564 		set_bit(STRIPE_BIT_DELAY, &sh->state);
2565 	}
2566 	return 1;
2567 
2568  overlap:
2569 	set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2570 	spin_unlock_irq(&sh->stripe_lock);
2571 	return 0;
2572 }
2573 
2574 static void end_reshape(struct r5conf *conf);
2575 
2576 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2577 			    struct stripe_head *sh)
2578 {
2579 	int sectors_per_chunk =
2580 		previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2581 	int dd_idx;
2582 	int chunk_offset = sector_div(stripe, sectors_per_chunk);
2583 	int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2584 
2585 	raid5_compute_sector(conf,
2586 			     stripe * (disks - conf->max_degraded)
2587 			     *sectors_per_chunk + chunk_offset,
2588 			     previous,
2589 			     &dd_idx, sh);
2590 }
2591 
2592 static void
2593 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2594 				struct stripe_head_state *s, int disks,
2595 				struct bio **return_bi)
2596 {
2597 	int i;
2598 	for (i = disks; i--; ) {
2599 		struct bio *bi;
2600 		int bitmap_end = 0;
2601 
2602 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2603 			struct md_rdev *rdev;
2604 			rcu_read_lock();
2605 			rdev = rcu_dereference(conf->disks[i].rdev);
2606 			if (rdev && test_bit(In_sync, &rdev->flags))
2607 				atomic_inc(&rdev->nr_pending);
2608 			else
2609 				rdev = NULL;
2610 			rcu_read_unlock();
2611 			if (rdev) {
2612 				if (!rdev_set_badblocks(
2613 					    rdev,
2614 					    sh->sector,
2615 					    STRIPE_SECTORS, 0))
2616 					md_error(conf->mddev, rdev);
2617 				rdev_dec_pending(rdev, conf->mddev);
2618 			}
2619 		}
2620 		spin_lock_irq(&sh->stripe_lock);
2621 		/* fail all writes first */
2622 		bi = sh->dev[i].towrite;
2623 		sh->dev[i].towrite = NULL;
2624 		spin_unlock_irq(&sh->stripe_lock);
2625 		if (bi)
2626 			bitmap_end = 1;
2627 
2628 		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2629 			wake_up(&conf->wait_for_overlap);
2630 
2631 		while (bi && bi->bi_sector <
2632 			sh->dev[i].sector + STRIPE_SECTORS) {
2633 			struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2634 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
2635 			if (!raid5_dec_bi_active_stripes(bi)) {
2636 				md_write_end(conf->mddev);
2637 				bi->bi_next = *return_bi;
2638 				*return_bi = bi;
2639 			}
2640 			bi = nextbi;
2641 		}
2642 		if (bitmap_end)
2643 			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2644 				STRIPE_SECTORS, 0, 0);
2645 		bitmap_end = 0;
2646 		/* and fail all 'written' */
2647 		bi = sh->dev[i].written;
2648 		sh->dev[i].written = NULL;
2649 		if (bi) bitmap_end = 1;
2650 		while (bi && bi->bi_sector <
2651 		       sh->dev[i].sector + STRIPE_SECTORS) {
2652 			struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2653 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
2654 			if (!raid5_dec_bi_active_stripes(bi)) {
2655 				md_write_end(conf->mddev);
2656 				bi->bi_next = *return_bi;
2657 				*return_bi = bi;
2658 			}
2659 			bi = bi2;
2660 		}
2661 
2662 		/* fail any reads if this device is non-operational and
2663 		 * the data has not reached the cache yet.
2664 		 */
2665 		if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2666 		    (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2667 		      test_bit(R5_ReadError, &sh->dev[i].flags))) {
2668 			spin_lock_irq(&sh->stripe_lock);
2669 			bi = sh->dev[i].toread;
2670 			sh->dev[i].toread = NULL;
2671 			spin_unlock_irq(&sh->stripe_lock);
2672 			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2673 				wake_up(&conf->wait_for_overlap);
2674 			while (bi && bi->bi_sector <
2675 			       sh->dev[i].sector + STRIPE_SECTORS) {
2676 				struct bio *nextbi =
2677 					r5_next_bio(bi, sh->dev[i].sector);
2678 				clear_bit(BIO_UPTODATE, &bi->bi_flags);
2679 				if (!raid5_dec_bi_active_stripes(bi)) {
2680 					bi->bi_next = *return_bi;
2681 					*return_bi = bi;
2682 				}
2683 				bi = nextbi;
2684 			}
2685 		}
2686 		if (bitmap_end)
2687 			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2688 					STRIPE_SECTORS, 0, 0);
2689 		/* If we were in the middle of a write the parity block might
2690 		 * still be locked - so just clear all R5_LOCKED flags
2691 		 */
2692 		clear_bit(R5_LOCKED, &sh->dev[i].flags);
2693 	}
2694 
2695 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2696 		if (atomic_dec_and_test(&conf->pending_full_writes))
2697 			md_wakeup_thread(conf->mddev->thread);
2698 }
2699 
2700 static void
2701 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2702 		   struct stripe_head_state *s)
2703 {
2704 	int abort = 0;
2705 	int i;
2706 
2707 	clear_bit(STRIPE_SYNCING, &sh->state);
2708 	if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2709 		wake_up(&conf->wait_for_overlap);
2710 	s->syncing = 0;
2711 	s->replacing = 0;
2712 	/* There is nothing more to do for sync/check/repair.
2713 	 * Don't even need to abort as that is handled elsewhere
2714 	 * if needed, and not always wanted e.g. if there is a known
2715 	 * bad block here.
2716 	 * For recover/replace we need to record a bad block on all
2717 	 * non-sync devices, or abort the recovery
2718 	 */
2719 	if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2720 		/* During recovery devices cannot be removed, so
2721 		 * locking and refcounting of rdevs is not needed
2722 		 */
2723 		for (i = 0; i < conf->raid_disks; i++) {
2724 			struct md_rdev *rdev = conf->disks[i].rdev;
2725 			if (rdev
2726 			    && !test_bit(Faulty, &rdev->flags)
2727 			    && !test_bit(In_sync, &rdev->flags)
2728 			    && !rdev_set_badblocks(rdev, sh->sector,
2729 						   STRIPE_SECTORS, 0))
2730 				abort = 1;
2731 			rdev = conf->disks[i].replacement;
2732 			if (rdev
2733 			    && !test_bit(Faulty, &rdev->flags)
2734 			    && !test_bit(In_sync, &rdev->flags)
2735 			    && !rdev_set_badblocks(rdev, sh->sector,
2736 						   STRIPE_SECTORS, 0))
2737 				abort = 1;
2738 		}
2739 		if (abort)
2740 			conf->recovery_disabled =
2741 				conf->mddev->recovery_disabled;
2742 	}
2743 	md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2744 }
2745 
2746 static int want_replace(struct stripe_head *sh, int disk_idx)
2747 {
2748 	struct md_rdev *rdev;
2749 	int rv = 0;
2750 	/* Doing recovery so rcu locking not required */
2751 	rdev = sh->raid_conf->disks[disk_idx].replacement;
2752 	if (rdev
2753 	    && !test_bit(Faulty, &rdev->flags)
2754 	    && !test_bit(In_sync, &rdev->flags)
2755 	    && (rdev->recovery_offset <= sh->sector
2756 		|| rdev->mddev->recovery_cp <= sh->sector))
2757 		rv = 1;
2758 
2759 	return rv;
2760 }
2761 
2762 /* fetch_block - checks the given member device to see if its data needs
2763  * to be read or computed to satisfy a request.
2764  *
2765  * Returns 1 when no more member devices need to be checked, otherwise returns
2766  * 0 to tell the loop in handle_stripe_fill to continue
2767  */
2768 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2769 		       int disk_idx, int disks)
2770 {
2771 	struct r5dev *dev = &sh->dev[disk_idx];
2772 	struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2773 				  &sh->dev[s->failed_num[1]] };
2774 
2775 	/* is the data in this block needed, and can we get it? */
2776 	if (!test_bit(R5_LOCKED, &dev->flags) &&
2777 	    !test_bit(R5_UPTODATE, &dev->flags) &&
2778 	    (dev->toread ||
2779 	     (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2780 	     s->syncing || s->expanding ||
2781 	     (s->replacing && want_replace(sh, disk_idx)) ||
2782 	     (s->failed >= 1 && fdev[0]->toread) ||
2783 	     (s->failed >= 2 && fdev[1]->toread) ||
2784 	     (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2785 	      !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2786 	     (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2787 		/* we would like to get this block, possibly by computing it,
2788 		 * otherwise read it if the backing disk is insync
2789 		 */
2790 		BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2791 		BUG_ON(test_bit(R5_Wantread, &dev->flags));
2792 		if ((s->uptodate == disks - 1) &&
2793 		    (s->failed && (disk_idx == s->failed_num[0] ||
2794 				   disk_idx == s->failed_num[1]))) {
2795 			/* have disk failed, and we're requested to fetch it;
2796 			 * do compute it
2797 			 */
2798 			pr_debug("Computing stripe %llu block %d\n",
2799 			       (unsigned long long)sh->sector, disk_idx);
2800 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2801 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2802 			set_bit(R5_Wantcompute, &dev->flags);
2803 			sh->ops.target = disk_idx;
2804 			sh->ops.target2 = -1; /* no 2nd target */
2805 			s->req_compute = 1;
2806 			/* Careful: from this point on 'uptodate' is in the eye
2807 			 * of raid_run_ops which services 'compute' operations
2808 			 * before writes. R5_Wantcompute flags a block that will
2809 			 * be R5_UPTODATE by the time it is needed for a
2810 			 * subsequent operation.
2811 			 */
2812 			s->uptodate++;
2813 			return 1;
2814 		} else if (s->uptodate == disks-2 && s->failed >= 2) {
2815 			/* Computing 2-failure is *very* expensive; only
2816 			 * do it if failed >= 2
2817 			 */
2818 			int other;
2819 			for (other = disks; other--; ) {
2820 				if (other == disk_idx)
2821 					continue;
2822 				if (!test_bit(R5_UPTODATE,
2823 				      &sh->dev[other].flags))
2824 					break;
2825 			}
2826 			BUG_ON(other < 0);
2827 			pr_debug("Computing stripe %llu blocks %d,%d\n",
2828 			       (unsigned long long)sh->sector,
2829 			       disk_idx, other);
2830 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2831 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2832 			set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2833 			set_bit(R5_Wantcompute, &sh->dev[other].flags);
2834 			sh->ops.target = disk_idx;
2835 			sh->ops.target2 = other;
2836 			s->uptodate += 2;
2837 			s->req_compute = 1;
2838 			return 1;
2839 		} else if (test_bit(R5_Insync, &dev->flags)) {
2840 			set_bit(R5_LOCKED, &dev->flags);
2841 			set_bit(R5_Wantread, &dev->flags);
2842 			s->locked++;
2843 			pr_debug("Reading block %d (sync=%d)\n",
2844 				disk_idx, s->syncing);
2845 		}
2846 	}
2847 
2848 	return 0;
2849 }
2850 
2851 /**
2852  * handle_stripe_fill - read or compute data to satisfy pending requests.
2853  */
2854 static void handle_stripe_fill(struct stripe_head *sh,
2855 			       struct stripe_head_state *s,
2856 			       int disks)
2857 {
2858 	int i;
2859 
2860 	/* look for blocks to read/compute, skip this if a compute
2861 	 * is already in flight, or if the stripe contents are in the
2862 	 * midst of changing due to a write
2863 	 */
2864 	if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2865 	    !sh->reconstruct_state)
2866 		for (i = disks; i--; )
2867 			if (fetch_block(sh, s, i, disks))
2868 				break;
2869 	set_bit(STRIPE_HANDLE, &sh->state);
2870 }
2871 
2872 
2873 /* handle_stripe_clean_event
2874  * any written block on an uptodate or failed drive can be returned.
2875  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2876  * never LOCKED, so we don't need to test 'failed' directly.
2877  */
2878 static void handle_stripe_clean_event(struct r5conf *conf,
2879 	struct stripe_head *sh, int disks, struct bio **return_bi)
2880 {
2881 	int i;
2882 	struct r5dev *dev;
2883 	int discard_pending = 0;
2884 
2885 	for (i = disks; i--; )
2886 		if (sh->dev[i].written) {
2887 			dev = &sh->dev[i];
2888 			if (!test_bit(R5_LOCKED, &dev->flags) &&
2889 			    (test_bit(R5_UPTODATE, &dev->flags) ||
2890 			     test_bit(R5_Discard, &dev->flags))) {
2891 				/* We can return any write requests */
2892 				struct bio *wbi, *wbi2;
2893 				pr_debug("Return write for disc %d\n", i);
2894 				if (test_and_clear_bit(R5_Discard, &dev->flags))
2895 					clear_bit(R5_UPTODATE, &dev->flags);
2896 				wbi = dev->written;
2897 				dev->written = NULL;
2898 				while (wbi && wbi->bi_sector <
2899 					dev->sector + STRIPE_SECTORS) {
2900 					wbi2 = r5_next_bio(wbi, dev->sector);
2901 					if (!raid5_dec_bi_active_stripes(wbi)) {
2902 						md_write_end(conf->mddev);
2903 						wbi->bi_next = *return_bi;
2904 						*return_bi = wbi;
2905 					}
2906 					wbi = wbi2;
2907 				}
2908 				bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2909 						STRIPE_SECTORS,
2910 					 !test_bit(STRIPE_DEGRADED, &sh->state),
2911 						0);
2912 			} else if (test_bit(R5_Discard, &dev->flags))
2913 				discard_pending = 1;
2914 		}
2915 	if (!discard_pending &&
2916 	    test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
2917 		clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2918 		clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2919 		if (sh->qd_idx >= 0) {
2920 			clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2921 			clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
2922 		}
2923 		/* now that discard is done we can proceed with any sync */
2924 		clear_bit(STRIPE_DISCARD, &sh->state);
2925 		/*
2926 		 * SCSI discard will change some bio fields and the stripe has
2927 		 * no updated data, so remove it from hash list and the stripe
2928 		 * will be reinitialized
2929 		 */
2930 		spin_lock_irq(&conf->device_lock);
2931 		remove_hash(sh);
2932 		spin_unlock_irq(&conf->device_lock);
2933 		if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2934 			set_bit(STRIPE_HANDLE, &sh->state);
2935 
2936 	}
2937 
2938 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2939 		if (atomic_dec_and_test(&conf->pending_full_writes))
2940 			md_wakeup_thread(conf->mddev->thread);
2941 }
2942 
2943 static void handle_stripe_dirtying(struct r5conf *conf,
2944 				   struct stripe_head *sh,
2945 				   struct stripe_head_state *s,
2946 				   int disks)
2947 {
2948 	int rmw = 0, rcw = 0, i;
2949 	sector_t recovery_cp = conf->mddev->recovery_cp;
2950 
2951 	/* RAID6 requires 'rcw' in current implementation.
2952 	 * Otherwise, check whether resync is now happening or should start.
2953 	 * If yes, then the array is dirty (after unclean shutdown or
2954 	 * initial creation), so parity in some stripes might be inconsistent.
2955 	 * In this case, we need to always do reconstruct-write, to ensure
2956 	 * that in case of drive failure or read-error correction, we
2957 	 * generate correct data from the parity.
2958 	 */
2959 	if (conf->max_degraded == 2 ||
2960 	    (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2961 		/* Calculate the real rcw later - for now make it
2962 		 * look like rcw is cheaper
2963 		 */
2964 		rcw = 1; rmw = 2;
2965 		pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2966 			 conf->max_degraded, (unsigned long long)recovery_cp,
2967 			 (unsigned long long)sh->sector);
2968 	} else for (i = disks; i--; ) {
2969 		/* would I have to read this buffer for read_modify_write */
2970 		struct r5dev *dev = &sh->dev[i];
2971 		if ((dev->towrite || i == sh->pd_idx) &&
2972 		    !test_bit(R5_LOCKED, &dev->flags) &&
2973 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2974 		      test_bit(R5_Wantcompute, &dev->flags))) {
2975 			if (test_bit(R5_Insync, &dev->flags))
2976 				rmw++;
2977 			else
2978 				rmw += 2*disks;  /* cannot read it */
2979 		}
2980 		/* Would I have to read this buffer for reconstruct_write */
2981 		if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2982 		    !test_bit(R5_LOCKED, &dev->flags) &&
2983 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2984 		    test_bit(R5_Wantcompute, &dev->flags))) {
2985 			if (test_bit(R5_Insync, &dev->flags)) rcw++;
2986 			else
2987 				rcw += 2*disks;
2988 		}
2989 	}
2990 	pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2991 		(unsigned long long)sh->sector, rmw, rcw);
2992 	set_bit(STRIPE_HANDLE, &sh->state);
2993 	if (rmw < rcw && rmw > 0) {
2994 		/* prefer read-modify-write, but need to get some data */
2995 		if (conf->mddev->queue)
2996 			blk_add_trace_msg(conf->mddev->queue,
2997 					  "raid5 rmw %llu %d",
2998 					  (unsigned long long)sh->sector, rmw);
2999 		for (i = disks; i--; ) {
3000 			struct r5dev *dev = &sh->dev[i];
3001 			if ((dev->towrite || i == sh->pd_idx) &&
3002 			    !test_bit(R5_LOCKED, &dev->flags) &&
3003 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
3004 			    test_bit(R5_Wantcompute, &dev->flags)) &&
3005 			    test_bit(R5_Insync, &dev->flags)) {
3006 				if (
3007 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3008 					pr_debug("Read_old block "
3009 						 "%d for r-m-w\n", i);
3010 					set_bit(R5_LOCKED, &dev->flags);
3011 					set_bit(R5_Wantread, &dev->flags);
3012 					s->locked++;
3013 				} else {
3014 					set_bit(STRIPE_DELAYED, &sh->state);
3015 					set_bit(STRIPE_HANDLE, &sh->state);
3016 				}
3017 			}
3018 		}
3019 	}
3020 	if (rcw <= rmw && rcw > 0) {
3021 		/* want reconstruct write, but need to get some data */
3022 		int qread =0;
3023 		rcw = 0;
3024 		for (i = disks; i--; ) {
3025 			struct r5dev *dev = &sh->dev[i];
3026 			if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3027 			    i != sh->pd_idx && i != sh->qd_idx &&
3028 			    !test_bit(R5_LOCKED, &dev->flags) &&
3029 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
3030 			      test_bit(R5_Wantcompute, &dev->flags))) {
3031 				rcw++;
3032 				if (!test_bit(R5_Insync, &dev->flags))
3033 					continue; /* it's a failed drive */
3034 				if (
3035 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3036 					pr_debug("Read_old block "
3037 						"%d for Reconstruct\n", i);
3038 					set_bit(R5_LOCKED, &dev->flags);
3039 					set_bit(R5_Wantread, &dev->flags);
3040 					s->locked++;
3041 					qread++;
3042 				} else {
3043 					set_bit(STRIPE_DELAYED, &sh->state);
3044 					set_bit(STRIPE_HANDLE, &sh->state);
3045 				}
3046 			}
3047 		}
3048 		if (rcw && conf->mddev->queue)
3049 			blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3050 					  (unsigned long long)sh->sector,
3051 					  rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3052 	}
3053 	/* now if nothing is locked, and if we have enough data,
3054 	 * we can start a write request
3055 	 */
3056 	/* since handle_stripe can be called at any time we need to handle the
3057 	 * case where a compute block operation has been submitted and then a
3058 	 * subsequent call wants to start a write request.  raid_run_ops only
3059 	 * handles the case where compute block and reconstruct are requested
3060 	 * simultaneously.  If this is not the case then new writes need to be
3061 	 * held off until the compute completes.
3062 	 */
3063 	if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3064 	    (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3065 	    !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3066 		schedule_reconstruction(sh, s, rcw == 0, 0);
3067 }
3068 
3069 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3070 				struct stripe_head_state *s, int disks)
3071 {
3072 	struct r5dev *dev = NULL;
3073 
3074 	set_bit(STRIPE_HANDLE, &sh->state);
3075 
3076 	switch (sh->check_state) {
3077 	case check_state_idle:
3078 		/* start a new check operation if there are no failures */
3079 		if (s->failed == 0) {
3080 			BUG_ON(s->uptodate != disks);
3081 			sh->check_state = check_state_run;
3082 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
3083 			clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3084 			s->uptodate--;
3085 			break;
3086 		}
3087 		dev = &sh->dev[s->failed_num[0]];
3088 		/* fall through */
3089 	case check_state_compute_result:
3090 		sh->check_state = check_state_idle;
3091 		if (!dev)
3092 			dev = &sh->dev[sh->pd_idx];
3093 
3094 		/* check that a write has not made the stripe insync */
3095 		if (test_bit(STRIPE_INSYNC, &sh->state))
3096 			break;
3097 
3098 		/* either failed parity check, or recovery is happening */
3099 		BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3100 		BUG_ON(s->uptodate != disks);
3101 
3102 		set_bit(R5_LOCKED, &dev->flags);
3103 		s->locked++;
3104 		set_bit(R5_Wantwrite, &dev->flags);
3105 
3106 		clear_bit(STRIPE_DEGRADED, &sh->state);
3107 		set_bit(STRIPE_INSYNC, &sh->state);
3108 		break;
3109 	case check_state_run:
3110 		break; /* we will be called again upon completion */
3111 	case check_state_check_result:
3112 		sh->check_state = check_state_idle;
3113 
3114 		/* if a failure occurred during the check operation, leave
3115 		 * STRIPE_INSYNC not set and let the stripe be handled again
3116 		 */
3117 		if (s->failed)
3118 			break;
3119 
3120 		/* handle a successful check operation, if parity is correct
3121 		 * we are done.  Otherwise update the mismatch count and repair
3122 		 * parity if !MD_RECOVERY_CHECK
3123 		 */
3124 		if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3125 			/* parity is correct (on disc,
3126 			 * not in buffer any more)
3127 			 */
3128 			set_bit(STRIPE_INSYNC, &sh->state);
3129 		else {
3130 			atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3131 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3132 				/* don't try to repair!! */
3133 				set_bit(STRIPE_INSYNC, &sh->state);
3134 			else {
3135 				sh->check_state = check_state_compute_run;
3136 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3137 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3138 				set_bit(R5_Wantcompute,
3139 					&sh->dev[sh->pd_idx].flags);
3140 				sh->ops.target = sh->pd_idx;
3141 				sh->ops.target2 = -1;
3142 				s->uptodate++;
3143 			}
3144 		}
3145 		break;
3146 	case check_state_compute_run:
3147 		break;
3148 	default:
3149 		printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3150 		       __func__, sh->check_state,
3151 		       (unsigned long long) sh->sector);
3152 		BUG();
3153 	}
3154 }
3155 
3156 
3157 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3158 				  struct stripe_head_state *s,
3159 				  int disks)
3160 {
3161 	int pd_idx = sh->pd_idx;
3162 	int qd_idx = sh->qd_idx;
3163 	struct r5dev *dev;
3164 
3165 	set_bit(STRIPE_HANDLE, &sh->state);
3166 
3167 	BUG_ON(s->failed > 2);
3168 
3169 	/* Want to check and possibly repair P and Q.
3170 	 * However there could be one 'failed' device, in which
3171 	 * case we can only check one of them, possibly using the
3172 	 * other to generate missing data
3173 	 */
3174 
3175 	switch (sh->check_state) {
3176 	case check_state_idle:
3177 		/* start a new check operation if there are < 2 failures */
3178 		if (s->failed == s->q_failed) {
3179 			/* The only possible failed device holds Q, so it
3180 			 * makes sense to check P (If anything else were failed,
3181 			 * we would have used P to recreate it).
3182 			 */
3183 			sh->check_state = check_state_run;
3184 		}
3185 		if (!s->q_failed && s->failed < 2) {
3186 			/* Q is not failed, and we didn't use it to generate
3187 			 * anything, so it makes sense to check it
3188 			 */
3189 			if (sh->check_state == check_state_run)
3190 				sh->check_state = check_state_run_pq;
3191 			else
3192 				sh->check_state = check_state_run_q;
3193 		}
3194 
3195 		/* discard potentially stale zero_sum_result */
3196 		sh->ops.zero_sum_result = 0;
3197 
3198 		if (sh->check_state == check_state_run) {
3199 			/* async_xor_zero_sum destroys the contents of P */
3200 			clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3201 			s->uptodate--;
3202 		}
3203 		if (sh->check_state >= check_state_run &&
3204 		    sh->check_state <= check_state_run_pq) {
3205 			/* async_syndrome_zero_sum preserves P and Q, so
3206 			 * no need to mark them !uptodate here
3207 			 */
3208 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
3209 			break;
3210 		}
3211 
3212 		/* we have 2-disk failure */
3213 		BUG_ON(s->failed != 2);
3214 		/* fall through */
3215 	case check_state_compute_result:
3216 		sh->check_state = check_state_idle;
3217 
3218 		/* check that a write has not made the stripe insync */
3219 		if (test_bit(STRIPE_INSYNC, &sh->state))
3220 			break;
3221 
3222 		/* now write out any block on a failed drive,
3223 		 * or P or Q if they were recomputed
3224 		 */
3225 		BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3226 		if (s->failed == 2) {
3227 			dev = &sh->dev[s->failed_num[1]];
3228 			s->locked++;
3229 			set_bit(R5_LOCKED, &dev->flags);
3230 			set_bit(R5_Wantwrite, &dev->flags);
3231 		}
3232 		if (s->failed >= 1) {
3233 			dev = &sh->dev[s->failed_num[0]];
3234 			s->locked++;
3235 			set_bit(R5_LOCKED, &dev->flags);
3236 			set_bit(R5_Wantwrite, &dev->flags);
3237 		}
3238 		if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3239 			dev = &sh->dev[pd_idx];
3240 			s->locked++;
3241 			set_bit(R5_LOCKED, &dev->flags);
3242 			set_bit(R5_Wantwrite, &dev->flags);
3243 		}
3244 		if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3245 			dev = &sh->dev[qd_idx];
3246 			s->locked++;
3247 			set_bit(R5_LOCKED, &dev->flags);
3248 			set_bit(R5_Wantwrite, &dev->flags);
3249 		}
3250 		clear_bit(STRIPE_DEGRADED, &sh->state);
3251 
3252 		set_bit(STRIPE_INSYNC, &sh->state);
3253 		break;
3254 	case check_state_run:
3255 	case check_state_run_q:
3256 	case check_state_run_pq:
3257 		break; /* we will be called again upon completion */
3258 	case check_state_check_result:
3259 		sh->check_state = check_state_idle;
3260 
3261 		/* handle a successful check operation, if parity is correct
3262 		 * we are done.  Otherwise update the mismatch count and repair
3263 		 * parity if !MD_RECOVERY_CHECK
3264 		 */
3265 		if (sh->ops.zero_sum_result == 0) {
3266 			/* both parities are correct */
3267 			if (!s->failed)
3268 				set_bit(STRIPE_INSYNC, &sh->state);
3269 			else {
3270 				/* in contrast to the raid5 case we can validate
3271 				 * parity, but still have a failure to write
3272 				 * back
3273 				 */
3274 				sh->check_state = check_state_compute_result;
3275 				/* Returning at this point means that we may go
3276 				 * off and bring p and/or q uptodate again so
3277 				 * we make sure to check zero_sum_result again
3278 				 * to verify if p or q need writeback
3279 				 */
3280 			}
3281 		} else {
3282 			atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3283 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3284 				/* don't try to repair!! */
3285 				set_bit(STRIPE_INSYNC, &sh->state);
3286 			else {
3287 				int *target = &sh->ops.target;
3288 
3289 				sh->ops.target = -1;
3290 				sh->ops.target2 = -1;
3291 				sh->check_state = check_state_compute_run;
3292 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3293 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3294 				if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3295 					set_bit(R5_Wantcompute,
3296 						&sh->dev[pd_idx].flags);
3297 					*target = pd_idx;
3298 					target = &sh->ops.target2;
3299 					s->uptodate++;
3300 				}
3301 				if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3302 					set_bit(R5_Wantcompute,
3303 						&sh->dev[qd_idx].flags);
3304 					*target = qd_idx;
3305 					s->uptodate++;
3306 				}
3307 			}
3308 		}
3309 		break;
3310 	case check_state_compute_run:
3311 		break;
3312 	default:
3313 		printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3314 		       __func__, sh->check_state,
3315 		       (unsigned long long) sh->sector);
3316 		BUG();
3317 	}
3318 }
3319 
3320 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3321 {
3322 	int i;
3323 
3324 	/* We have read all the blocks in this stripe and now we need to
3325 	 * copy some of them into a target stripe for expand.
3326 	 */
3327 	struct dma_async_tx_descriptor *tx = NULL;
3328 	clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3329 	for (i = 0; i < sh->disks; i++)
3330 		if (i != sh->pd_idx && i != sh->qd_idx) {
3331 			int dd_idx, j;
3332 			struct stripe_head *sh2;
3333 			struct async_submit_ctl submit;
3334 
3335 			sector_t bn = compute_blocknr(sh, i, 1);
3336 			sector_t s = raid5_compute_sector(conf, bn, 0,
3337 							  &dd_idx, NULL);
3338 			sh2 = get_active_stripe(conf, s, 0, 1, 1);
3339 			if (sh2 == NULL)
3340 				/* so far only the early blocks of this stripe
3341 				 * have been requested.  When later blocks
3342 				 * get requested, we will try again
3343 				 */
3344 				continue;
3345 			if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3346 			   test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3347 				/* must have already done this block */
3348 				release_stripe(sh2);
3349 				continue;
3350 			}
3351 
3352 			/* place all the copies on one channel */
3353 			init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3354 			tx = async_memcpy(sh2->dev[dd_idx].page,
3355 					  sh->dev[i].page, 0, 0, STRIPE_SIZE,
3356 					  &submit);
3357 
3358 			set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3359 			set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3360 			for (j = 0; j < conf->raid_disks; j++)
3361 				if (j != sh2->pd_idx &&
3362 				    j != sh2->qd_idx &&
3363 				    !test_bit(R5_Expanded, &sh2->dev[j].flags))
3364 					break;
3365 			if (j == conf->raid_disks) {
3366 				set_bit(STRIPE_EXPAND_READY, &sh2->state);
3367 				set_bit(STRIPE_HANDLE, &sh2->state);
3368 			}
3369 			release_stripe(sh2);
3370 
3371 		}
3372 	/* done submitting copies, wait for them to complete */
3373 	async_tx_quiesce(&tx);
3374 }
3375 
3376 /*
3377  * handle_stripe - do things to a stripe.
3378  *
3379  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3380  * state of various bits to see what needs to be done.
3381  * Possible results:
3382  *    return some read requests which now have data
3383  *    return some write requests which are safely on storage
3384  *    schedule a read on some buffers
3385  *    schedule a write of some buffers
3386  *    return confirmation of parity correctness
3387  *
3388  */
3389 
3390 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3391 {
3392 	struct r5conf *conf = sh->raid_conf;
3393 	int disks = sh->disks;
3394 	struct r5dev *dev;
3395 	int i;
3396 	int do_recovery = 0;
3397 
3398 	memset(s, 0, sizeof(*s));
3399 
3400 	s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3401 	s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3402 	s->failed_num[0] = -1;
3403 	s->failed_num[1] = -1;
3404 
3405 	/* Now to look around and see what can be done */
3406 	rcu_read_lock();
3407 	for (i=disks; i--; ) {
3408 		struct md_rdev *rdev;
3409 		sector_t first_bad;
3410 		int bad_sectors;
3411 		int is_bad = 0;
3412 
3413 		dev = &sh->dev[i];
3414 
3415 		pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3416 			 i, dev->flags,
3417 			 dev->toread, dev->towrite, dev->written);
3418 		/* maybe we can reply to a read
3419 		 *
3420 		 * new wantfill requests are only permitted while
3421 		 * ops_complete_biofill is guaranteed to be inactive
3422 		 */
3423 		if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3424 		    !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3425 			set_bit(R5_Wantfill, &dev->flags);
3426 
3427 		/* now count some things */
3428 		if (test_bit(R5_LOCKED, &dev->flags))
3429 			s->locked++;
3430 		if (test_bit(R5_UPTODATE, &dev->flags))
3431 			s->uptodate++;
3432 		if (test_bit(R5_Wantcompute, &dev->flags)) {
3433 			s->compute++;
3434 			BUG_ON(s->compute > 2);
3435 		}
3436 
3437 		if (test_bit(R5_Wantfill, &dev->flags))
3438 			s->to_fill++;
3439 		else if (dev->toread)
3440 			s->to_read++;
3441 		if (dev->towrite) {
3442 			s->to_write++;
3443 			if (!test_bit(R5_OVERWRITE, &dev->flags))
3444 				s->non_overwrite++;
3445 		}
3446 		if (dev->written)
3447 			s->written++;
3448 		/* Prefer to use the replacement for reads, but only
3449 		 * if it is recovered enough and has no bad blocks.
3450 		 */
3451 		rdev = rcu_dereference(conf->disks[i].replacement);
3452 		if (rdev && !test_bit(Faulty, &rdev->flags) &&
3453 		    rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3454 		    !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3455 				 &first_bad, &bad_sectors))
3456 			set_bit(R5_ReadRepl, &dev->flags);
3457 		else {
3458 			if (rdev)
3459 				set_bit(R5_NeedReplace, &dev->flags);
3460 			rdev = rcu_dereference(conf->disks[i].rdev);
3461 			clear_bit(R5_ReadRepl, &dev->flags);
3462 		}
3463 		if (rdev && test_bit(Faulty, &rdev->flags))
3464 			rdev = NULL;
3465 		if (rdev) {
3466 			is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3467 					     &first_bad, &bad_sectors);
3468 			if (s->blocked_rdev == NULL
3469 			    && (test_bit(Blocked, &rdev->flags)
3470 				|| is_bad < 0)) {
3471 				if (is_bad < 0)
3472 					set_bit(BlockedBadBlocks,
3473 						&rdev->flags);
3474 				s->blocked_rdev = rdev;
3475 				atomic_inc(&rdev->nr_pending);
3476 			}
3477 		}
3478 		clear_bit(R5_Insync, &dev->flags);
3479 		if (!rdev)
3480 			/* Not in-sync */;
3481 		else if (is_bad) {
3482 			/* also not in-sync */
3483 			if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3484 			    test_bit(R5_UPTODATE, &dev->flags)) {
3485 				/* treat as in-sync, but with a read error
3486 				 * which we can now try to correct
3487 				 */
3488 				set_bit(R5_Insync, &dev->flags);
3489 				set_bit(R5_ReadError, &dev->flags);
3490 			}
3491 		} else if (test_bit(In_sync, &rdev->flags))
3492 			set_bit(R5_Insync, &dev->flags);
3493 		else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3494 			/* in sync if before recovery_offset */
3495 			set_bit(R5_Insync, &dev->flags);
3496 		else if (test_bit(R5_UPTODATE, &dev->flags) &&
3497 			 test_bit(R5_Expanded, &dev->flags))
3498 			/* If we've reshaped into here, we assume it is Insync.
3499 			 * We will shortly update recovery_offset to make
3500 			 * it official.
3501 			 */
3502 			set_bit(R5_Insync, &dev->flags);
3503 
3504 		if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3505 			/* This flag does not apply to '.replacement'
3506 			 * only to .rdev, so make sure to check that*/
3507 			struct md_rdev *rdev2 = rcu_dereference(
3508 				conf->disks[i].rdev);
3509 			if (rdev2 == rdev)
3510 				clear_bit(R5_Insync, &dev->flags);
3511 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3512 				s->handle_bad_blocks = 1;
3513 				atomic_inc(&rdev2->nr_pending);
3514 			} else
3515 				clear_bit(R5_WriteError, &dev->flags);
3516 		}
3517 		if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3518 			/* This flag does not apply to '.replacement'
3519 			 * only to .rdev, so make sure to check that*/
3520 			struct md_rdev *rdev2 = rcu_dereference(
3521 				conf->disks[i].rdev);
3522 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3523 				s->handle_bad_blocks = 1;
3524 				atomic_inc(&rdev2->nr_pending);
3525 			} else
3526 				clear_bit(R5_MadeGood, &dev->flags);
3527 		}
3528 		if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3529 			struct md_rdev *rdev2 = rcu_dereference(
3530 				conf->disks[i].replacement);
3531 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3532 				s->handle_bad_blocks = 1;
3533 				atomic_inc(&rdev2->nr_pending);
3534 			} else
3535 				clear_bit(R5_MadeGoodRepl, &dev->flags);
3536 		}
3537 		if (!test_bit(R5_Insync, &dev->flags)) {
3538 			/* The ReadError flag will just be confusing now */
3539 			clear_bit(R5_ReadError, &dev->flags);
3540 			clear_bit(R5_ReWrite, &dev->flags);
3541 		}
3542 		if (test_bit(R5_ReadError, &dev->flags))
3543 			clear_bit(R5_Insync, &dev->flags);
3544 		if (!test_bit(R5_Insync, &dev->flags)) {
3545 			if (s->failed < 2)
3546 				s->failed_num[s->failed] = i;
3547 			s->failed++;
3548 			if (rdev && !test_bit(Faulty, &rdev->flags))
3549 				do_recovery = 1;
3550 		}
3551 	}
3552 	if (test_bit(STRIPE_SYNCING, &sh->state)) {
3553 		/* If there is a failed device being replaced,
3554 		 *     we must be recovering.
3555 		 * else if we are after recovery_cp, we must be syncing
3556 		 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3557 		 * else we can only be replacing
3558 		 * sync and recovery both need to read all devices, and so
3559 		 * use the same flag.
3560 		 */
3561 		if (do_recovery ||
3562 		    sh->sector >= conf->mddev->recovery_cp ||
3563 		    test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3564 			s->syncing = 1;
3565 		else
3566 			s->replacing = 1;
3567 	}
3568 	rcu_read_unlock();
3569 }
3570 
3571 static void handle_stripe(struct stripe_head *sh)
3572 {
3573 	struct stripe_head_state s;
3574 	struct r5conf *conf = sh->raid_conf;
3575 	int i;
3576 	int prexor;
3577 	int disks = sh->disks;
3578 	struct r5dev *pdev, *qdev;
3579 
3580 	clear_bit(STRIPE_HANDLE, &sh->state);
3581 	if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3582 		/* already being handled, ensure it gets handled
3583 		 * again when current action finishes */
3584 		set_bit(STRIPE_HANDLE, &sh->state);
3585 		return;
3586 	}
3587 
3588 	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3589 		spin_lock(&sh->stripe_lock);
3590 		/* Cannot process 'sync' concurrently with 'discard' */
3591 		if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3592 		    test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3593 			set_bit(STRIPE_SYNCING, &sh->state);
3594 			clear_bit(STRIPE_INSYNC, &sh->state);
3595 			clear_bit(STRIPE_REPLACED, &sh->state);
3596 		}
3597 		spin_unlock(&sh->stripe_lock);
3598 	}
3599 	clear_bit(STRIPE_DELAYED, &sh->state);
3600 
3601 	pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3602 		"pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3603 	       (unsigned long long)sh->sector, sh->state,
3604 	       atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3605 	       sh->check_state, sh->reconstruct_state);
3606 
3607 	analyse_stripe(sh, &s);
3608 
3609 	if (s.handle_bad_blocks) {
3610 		set_bit(STRIPE_HANDLE, &sh->state);
3611 		goto finish;
3612 	}
3613 
3614 	if (unlikely(s.blocked_rdev)) {
3615 		if (s.syncing || s.expanding || s.expanded ||
3616 		    s.replacing || s.to_write || s.written) {
3617 			set_bit(STRIPE_HANDLE, &sh->state);
3618 			goto finish;
3619 		}
3620 		/* There is nothing for the blocked_rdev to block */
3621 		rdev_dec_pending(s.blocked_rdev, conf->mddev);
3622 		s.blocked_rdev = NULL;
3623 	}
3624 
3625 	if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3626 		set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3627 		set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3628 	}
3629 
3630 	pr_debug("locked=%d uptodate=%d to_read=%d"
3631 	       " to_write=%d failed=%d failed_num=%d,%d\n",
3632 	       s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3633 	       s.failed_num[0], s.failed_num[1]);
3634 	/* check if the array has lost more than max_degraded devices and,
3635 	 * if so, some requests might need to be failed.
3636 	 */
3637 	if (s.failed > conf->max_degraded) {
3638 		sh->check_state = 0;
3639 		sh->reconstruct_state = 0;
3640 		if (s.to_read+s.to_write+s.written)
3641 			handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3642 		if (s.syncing + s.replacing)
3643 			handle_failed_sync(conf, sh, &s);
3644 	}
3645 
3646 	/* Now we check to see if any write operations have recently
3647 	 * completed
3648 	 */
3649 	prexor = 0;
3650 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3651 		prexor = 1;
3652 	if (sh->reconstruct_state == reconstruct_state_drain_result ||
3653 	    sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3654 		sh->reconstruct_state = reconstruct_state_idle;
3655 
3656 		/* All the 'written' buffers and the parity block are ready to
3657 		 * be written back to disk
3658 		 */
3659 		BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3660 		       !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3661 		BUG_ON(sh->qd_idx >= 0 &&
3662 		       !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3663 		       !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3664 		for (i = disks; i--; ) {
3665 			struct r5dev *dev = &sh->dev[i];
3666 			if (test_bit(R5_LOCKED, &dev->flags) &&
3667 				(i == sh->pd_idx || i == sh->qd_idx ||
3668 				 dev->written)) {
3669 				pr_debug("Writing block %d\n", i);
3670 				set_bit(R5_Wantwrite, &dev->flags);
3671 				if (prexor)
3672 					continue;
3673 				if (!test_bit(R5_Insync, &dev->flags) ||
3674 				    ((i == sh->pd_idx || i == sh->qd_idx)  &&
3675 				     s.failed == 0))
3676 					set_bit(STRIPE_INSYNC, &sh->state);
3677 			}
3678 		}
3679 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3680 			s.dec_preread_active = 1;
3681 	}
3682 
3683 	/*
3684 	 * might be able to return some write requests if the parity blocks
3685 	 * are safe, or on a failed drive
3686 	 */
3687 	pdev = &sh->dev[sh->pd_idx];
3688 	s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3689 		|| (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3690 	qdev = &sh->dev[sh->qd_idx];
3691 	s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3692 		|| (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3693 		|| conf->level < 6;
3694 
3695 	if (s.written &&
3696 	    (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3697 			     && !test_bit(R5_LOCKED, &pdev->flags)
3698 			     && (test_bit(R5_UPTODATE, &pdev->flags) ||
3699 				 test_bit(R5_Discard, &pdev->flags))))) &&
3700 	    (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3701 			     && !test_bit(R5_LOCKED, &qdev->flags)
3702 			     && (test_bit(R5_UPTODATE, &qdev->flags) ||
3703 				 test_bit(R5_Discard, &qdev->flags))))))
3704 		handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3705 
3706 	/* Now we might consider reading some blocks, either to check/generate
3707 	 * parity, or to satisfy requests
3708 	 * or to load a block that is being partially written.
3709 	 */
3710 	if (s.to_read || s.non_overwrite
3711 	    || (conf->level == 6 && s.to_write && s.failed)
3712 	    || (s.syncing && (s.uptodate + s.compute < disks))
3713 	    || s.replacing
3714 	    || s.expanding)
3715 		handle_stripe_fill(sh, &s, disks);
3716 
3717 	/* Now to consider new write requests and what else, if anything
3718 	 * should be read.  We do not handle new writes when:
3719 	 * 1/ A 'write' operation (copy+xor) is already in flight.
3720 	 * 2/ A 'check' operation is in flight, as it may clobber the parity
3721 	 *    block.
3722 	 */
3723 	if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3724 		handle_stripe_dirtying(conf, sh, &s, disks);
3725 
3726 	/* maybe we need to check and possibly fix the parity for this stripe
3727 	 * Any reads will already have been scheduled, so we just see if enough
3728 	 * data is available.  The parity check is held off while parity
3729 	 * dependent operations are in flight.
3730 	 */
3731 	if (sh->check_state ||
3732 	    (s.syncing && s.locked == 0 &&
3733 	     !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3734 	     !test_bit(STRIPE_INSYNC, &sh->state))) {
3735 		if (conf->level == 6)
3736 			handle_parity_checks6(conf, sh, &s, disks);
3737 		else
3738 			handle_parity_checks5(conf, sh, &s, disks);
3739 	}
3740 
3741 	if ((s.replacing || s.syncing) && s.locked == 0
3742 	    && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3743 	    && !test_bit(STRIPE_REPLACED, &sh->state)) {
3744 		/* Write out to replacement devices where possible */
3745 		for (i = 0; i < conf->raid_disks; i++)
3746 			if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3747 				WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
3748 				set_bit(R5_WantReplace, &sh->dev[i].flags);
3749 				set_bit(R5_LOCKED, &sh->dev[i].flags);
3750 				s.locked++;
3751 			}
3752 		if (s.replacing)
3753 			set_bit(STRIPE_INSYNC, &sh->state);
3754 		set_bit(STRIPE_REPLACED, &sh->state);
3755 	}
3756 	if ((s.syncing || s.replacing) && s.locked == 0 &&
3757 	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3758 	    test_bit(STRIPE_INSYNC, &sh->state)) {
3759 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3760 		clear_bit(STRIPE_SYNCING, &sh->state);
3761 		if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3762 			wake_up(&conf->wait_for_overlap);
3763 	}
3764 
3765 	/* If the failed drives are just a ReadError, then we might need
3766 	 * to progress the repair/check process
3767 	 */
3768 	if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3769 		for (i = 0; i < s.failed; i++) {
3770 			struct r5dev *dev = &sh->dev[s.failed_num[i]];
3771 			if (test_bit(R5_ReadError, &dev->flags)
3772 			    && !test_bit(R5_LOCKED, &dev->flags)
3773 			    && test_bit(R5_UPTODATE, &dev->flags)
3774 				) {
3775 				if (!test_bit(R5_ReWrite, &dev->flags)) {
3776 					set_bit(R5_Wantwrite, &dev->flags);
3777 					set_bit(R5_ReWrite, &dev->flags);
3778 					set_bit(R5_LOCKED, &dev->flags);
3779 					s.locked++;
3780 				} else {
3781 					/* let's read it back */
3782 					set_bit(R5_Wantread, &dev->flags);
3783 					set_bit(R5_LOCKED, &dev->flags);
3784 					s.locked++;
3785 				}
3786 			}
3787 		}
3788 
3789 
3790 	/* Finish reconstruct operations initiated by the expansion process */
3791 	if (sh->reconstruct_state == reconstruct_state_result) {
3792 		struct stripe_head *sh_src
3793 			= get_active_stripe(conf, sh->sector, 1, 1, 1);
3794 		if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3795 			/* sh cannot be written until sh_src has been read.
3796 			 * so arrange for sh to be delayed a little
3797 			 */
3798 			set_bit(STRIPE_DELAYED, &sh->state);
3799 			set_bit(STRIPE_HANDLE, &sh->state);
3800 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3801 					      &sh_src->state))
3802 				atomic_inc(&conf->preread_active_stripes);
3803 			release_stripe(sh_src);
3804 			goto finish;
3805 		}
3806 		if (sh_src)
3807 			release_stripe(sh_src);
3808 
3809 		sh->reconstruct_state = reconstruct_state_idle;
3810 		clear_bit(STRIPE_EXPANDING, &sh->state);
3811 		for (i = conf->raid_disks; i--; ) {
3812 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
3813 			set_bit(R5_LOCKED, &sh->dev[i].flags);
3814 			s.locked++;
3815 		}
3816 	}
3817 
3818 	if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3819 	    !sh->reconstruct_state) {
3820 		/* Need to write out all blocks after computing parity */
3821 		sh->disks = conf->raid_disks;
3822 		stripe_set_idx(sh->sector, conf, 0, sh);
3823 		schedule_reconstruction(sh, &s, 1, 1);
3824 	} else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3825 		clear_bit(STRIPE_EXPAND_READY, &sh->state);
3826 		atomic_dec(&conf->reshape_stripes);
3827 		wake_up(&conf->wait_for_overlap);
3828 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3829 	}
3830 
3831 	if (s.expanding && s.locked == 0 &&
3832 	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3833 		handle_stripe_expansion(conf, sh);
3834 
3835 finish:
3836 	/* wait for this device to become unblocked */
3837 	if (unlikely(s.blocked_rdev)) {
3838 		if (conf->mddev->external)
3839 			md_wait_for_blocked_rdev(s.blocked_rdev,
3840 						 conf->mddev);
3841 		else
3842 			/* Internal metadata will immediately
3843 			 * be written by raid5d, so we don't
3844 			 * need to wait here.
3845 			 */
3846 			rdev_dec_pending(s.blocked_rdev,
3847 					 conf->mddev);
3848 	}
3849 
3850 	if (s.handle_bad_blocks)
3851 		for (i = disks; i--; ) {
3852 			struct md_rdev *rdev;
3853 			struct r5dev *dev = &sh->dev[i];
3854 			if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3855 				/* We own a safe reference to the rdev */
3856 				rdev = conf->disks[i].rdev;
3857 				if (!rdev_set_badblocks(rdev, sh->sector,
3858 							STRIPE_SECTORS, 0))
3859 					md_error(conf->mddev, rdev);
3860 				rdev_dec_pending(rdev, conf->mddev);
3861 			}
3862 			if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3863 				rdev = conf->disks[i].rdev;
3864 				rdev_clear_badblocks(rdev, sh->sector,
3865 						     STRIPE_SECTORS, 0);
3866 				rdev_dec_pending(rdev, conf->mddev);
3867 			}
3868 			if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3869 				rdev = conf->disks[i].replacement;
3870 				if (!rdev)
3871 					/* rdev have been moved down */
3872 					rdev = conf->disks[i].rdev;
3873 				rdev_clear_badblocks(rdev, sh->sector,
3874 						     STRIPE_SECTORS, 0);
3875 				rdev_dec_pending(rdev, conf->mddev);
3876 			}
3877 		}
3878 
3879 	if (s.ops_request)
3880 		raid_run_ops(sh, s.ops_request);
3881 
3882 	ops_run_io(sh, &s);
3883 
3884 	if (s.dec_preread_active) {
3885 		/* We delay this until after ops_run_io so that if make_request
3886 		 * is waiting on a flush, it won't continue until the writes
3887 		 * have actually been submitted.
3888 		 */
3889 		atomic_dec(&conf->preread_active_stripes);
3890 		if (atomic_read(&conf->preread_active_stripes) <
3891 		    IO_THRESHOLD)
3892 			md_wakeup_thread(conf->mddev->thread);
3893 	}
3894 
3895 	return_io(s.return_bi);
3896 
3897 	clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3898 }
3899 
3900 static void raid5_activate_delayed(struct r5conf *conf)
3901 {
3902 	if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3903 		while (!list_empty(&conf->delayed_list)) {
3904 			struct list_head *l = conf->delayed_list.next;
3905 			struct stripe_head *sh;
3906 			sh = list_entry(l, struct stripe_head, lru);
3907 			list_del_init(l);
3908 			clear_bit(STRIPE_DELAYED, &sh->state);
3909 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3910 				atomic_inc(&conf->preread_active_stripes);
3911 			list_add_tail(&sh->lru, &conf->hold_list);
3912 			raid5_wakeup_stripe_thread(sh);
3913 		}
3914 	}
3915 }
3916 
3917 static void activate_bit_delay(struct r5conf *conf)
3918 {
3919 	/* device_lock is held */
3920 	struct list_head head;
3921 	list_add(&head, &conf->bitmap_list);
3922 	list_del_init(&conf->bitmap_list);
3923 	while (!list_empty(&head)) {
3924 		struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3925 		list_del_init(&sh->lru);
3926 		atomic_inc(&sh->count);
3927 		__release_stripe(conf, sh);
3928 	}
3929 }
3930 
3931 int md_raid5_congested(struct mddev *mddev, int bits)
3932 {
3933 	struct r5conf *conf = mddev->private;
3934 
3935 	/* No difference between reads and writes.  Just check
3936 	 * how busy the stripe_cache is
3937 	 */
3938 
3939 	if (conf->inactive_blocked)
3940 		return 1;
3941 	if (conf->quiesce)
3942 		return 1;
3943 	if (list_empty_careful(&conf->inactive_list))
3944 		return 1;
3945 
3946 	return 0;
3947 }
3948 EXPORT_SYMBOL_GPL(md_raid5_congested);
3949 
3950 static int raid5_congested(void *data, int bits)
3951 {
3952 	struct mddev *mddev = data;
3953 
3954 	return mddev_congested(mddev, bits) ||
3955 		md_raid5_congested(mddev, bits);
3956 }
3957 
3958 /* We want read requests to align with chunks where possible,
3959  * but write requests don't need to.
3960  */
3961 static int raid5_mergeable_bvec(struct request_queue *q,
3962 				struct bvec_merge_data *bvm,
3963 				struct bio_vec *biovec)
3964 {
3965 	struct mddev *mddev = q->queuedata;
3966 	sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3967 	int max;
3968 	unsigned int chunk_sectors = mddev->chunk_sectors;
3969 	unsigned int bio_sectors = bvm->bi_size >> 9;
3970 
3971 	if ((bvm->bi_rw & 1) == WRITE)
3972 		return biovec->bv_len; /* always allow writes to be mergeable */
3973 
3974 	if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3975 		chunk_sectors = mddev->new_chunk_sectors;
3976 	max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3977 	if (max < 0) max = 0;
3978 	if (max <= biovec->bv_len && bio_sectors == 0)
3979 		return biovec->bv_len;
3980 	else
3981 		return max;
3982 }
3983 
3984 
3985 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3986 {
3987 	sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3988 	unsigned int chunk_sectors = mddev->chunk_sectors;
3989 	unsigned int bio_sectors = bio_sectors(bio);
3990 
3991 	if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3992 		chunk_sectors = mddev->new_chunk_sectors;
3993 	return  chunk_sectors >=
3994 		((sector & (chunk_sectors - 1)) + bio_sectors);
3995 }
3996 
3997 /*
3998  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3999  *  later sampled by raid5d.
4000  */
4001 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4002 {
4003 	unsigned long flags;
4004 
4005 	spin_lock_irqsave(&conf->device_lock, flags);
4006 
4007 	bi->bi_next = conf->retry_read_aligned_list;
4008 	conf->retry_read_aligned_list = bi;
4009 
4010 	spin_unlock_irqrestore(&conf->device_lock, flags);
4011 	md_wakeup_thread(conf->mddev->thread);
4012 }
4013 
4014 
4015 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4016 {
4017 	struct bio *bi;
4018 
4019 	bi = conf->retry_read_aligned;
4020 	if (bi) {
4021 		conf->retry_read_aligned = NULL;
4022 		return bi;
4023 	}
4024 	bi = conf->retry_read_aligned_list;
4025 	if(bi) {
4026 		conf->retry_read_aligned_list = bi->bi_next;
4027 		bi->bi_next = NULL;
4028 		/*
4029 		 * this sets the active strip count to 1 and the processed
4030 		 * strip count to zero (upper 8 bits)
4031 		 */
4032 		raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4033 	}
4034 
4035 	return bi;
4036 }
4037 
4038 
4039 /*
4040  *  The "raid5_align_endio" should check if the read succeeded and if it
4041  *  did, call bio_endio on the original bio (having bio_put the new bio
4042  *  first).
4043  *  If the read failed..
4044  */
4045 static void raid5_align_endio(struct bio *bi, int error)
4046 {
4047 	struct bio* raid_bi  = bi->bi_private;
4048 	struct mddev *mddev;
4049 	struct r5conf *conf;
4050 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4051 	struct md_rdev *rdev;
4052 
4053 	bio_put(bi);
4054 
4055 	rdev = (void*)raid_bi->bi_next;
4056 	raid_bi->bi_next = NULL;
4057 	mddev = rdev->mddev;
4058 	conf = mddev->private;
4059 
4060 	rdev_dec_pending(rdev, conf->mddev);
4061 
4062 	if (!error && uptodate) {
4063 		trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4064 					 raid_bi, 0);
4065 		bio_endio(raid_bi, 0);
4066 		if (atomic_dec_and_test(&conf->active_aligned_reads))
4067 			wake_up(&conf->wait_for_stripe);
4068 		return;
4069 	}
4070 
4071 
4072 	pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4073 
4074 	add_bio_to_retry(raid_bi, conf);
4075 }
4076 
4077 static int bio_fits_rdev(struct bio *bi)
4078 {
4079 	struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4080 
4081 	if (bio_sectors(bi) > queue_max_sectors(q))
4082 		return 0;
4083 	blk_recount_segments(q, bi);
4084 	if (bi->bi_phys_segments > queue_max_segments(q))
4085 		return 0;
4086 
4087 	if (q->merge_bvec_fn)
4088 		/* it's too hard to apply the merge_bvec_fn at this stage,
4089 		 * just just give up
4090 		 */
4091 		return 0;
4092 
4093 	return 1;
4094 }
4095 
4096 
4097 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4098 {
4099 	struct r5conf *conf = mddev->private;
4100 	int dd_idx;
4101 	struct bio* align_bi;
4102 	struct md_rdev *rdev;
4103 	sector_t end_sector;
4104 
4105 	if (!in_chunk_boundary(mddev, raid_bio)) {
4106 		pr_debug("chunk_aligned_read : non aligned\n");
4107 		return 0;
4108 	}
4109 	/*
4110 	 * use bio_clone_mddev to make a copy of the bio
4111 	 */
4112 	align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4113 	if (!align_bi)
4114 		return 0;
4115 	/*
4116 	 *   set bi_end_io to a new function, and set bi_private to the
4117 	 *     original bio.
4118 	 */
4119 	align_bi->bi_end_io  = raid5_align_endio;
4120 	align_bi->bi_private = raid_bio;
4121 	/*
4122 	 *	compute position
4123 	 */
4124 	align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
4125 						    0,
4126 						    &dd_idx, NULL);
4127 
4128 	end_sector = bio_end_sector(align_bi);
4129 	rcu_read_lock();
4130 	rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4131 	if (!rdev || test_bit(Faulty, &rdev->flags) ||
4132 	    rdev->recovery_offset < end_sector) {
4133 		rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4134 		if (rdev &&
4135 		    (test_bit(Faulty, &rdev->flags) ||
4136 		    !(test_bit(In_sync, &rdev->flags) ||
4137 		      rdev->recovery_offset >= end_sector)))
4138 			rdev = NULL;
4139 	}
4140 	if (rdev) {
4141 		sector_t first_bad;
4142 		int bad_sectors;
4143 
4144 		atomic_inc(&rdev->nr_pending);
4145 		rcu_read_unlock();
4146 		raid_bio->bi_next = (void*)rdev;
4147 		align_bi->bi_bdev =  rdev->bdev;
4148 		align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4149 
4150 		if (!bio_fits_rdev(align_bi) ||
4151 		    is_badblock(rdev, align_bi->bi_sector, bio_sectors(align_bi),
4152 				&first_bad, &bad_sectors)) {
4153 			/* too big in some way, or has a known bad block */
4154 			bio_put(align_bi);
4155 			rdev_dec_pending(rdev, mddev);
4156 			return 0;
4157 		}
4158 
4159 		/* No reshape active, so we can trust rdev->data_offset */
4160 		align_bi->bi_sector += rdev->data_offset;
4161 
4162 		spin_lock_irq(&conf->device_lock);
4163 		wait_event_lock_irq(conf->wait_for_stripe,
4164 				    conf->quiesce == 0,
4165 				    conf->device_lock);
4166 		atomic_inc(&conf->active_aligned_reads);
4167 		spin_unlock_irq(&conf->device_lock);
4168 
4169 		if (mddev->gendisk)
4170 			trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4171 					      align_bi, disk_devt(mddev->gendisk),
4172 					      raid_bio->bi_sector);
4173 		generic_make_request(align_bi);
4174 		return 1;
4175 	} else {
4176 		rcu_read_unlock();
4177 		bio_put(align_bi);
4178 		return 0;
4179 	}
4180 }
4181 
4182 /* __get_priority_stripe - get the next stripe to process
4183  *
4184  * Full stripe writes are allowed to pass preread active stripes up until
4185  * the bypass_threshold is exceeded.  In general the bypass_count
4186  * increments when the handle_list is handled before the hold_list; however, it
4187  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4188  * stripe with in flight i/o.  The bypass_count will be reset when the
4189  * head of the hold_list has changed, i.e. the head was promoted to the
4190  * handle_list.
4191  */
4192 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4193 {
4194 	struct stripe_head *sh = NULL, *tmp;
4195 	struct list_head *handle_list = NULL;
4196 	struct r5worker_group *wg = NULL;
4197 
4198 	if (conf->worker_cnt_per_group == 0) {
4199 		handle_list = &conf->handle_list;
4200 	} else if (group != ANY_GROUP) {
4201 		handle_list = &conf->worker_groups[group].handle_list;
4202 		wg = &conf->worker_groups[group];
4203 	} else {
4204 		int i;
4205 		for (i = 0; i < conf->group_cnt; i++) {
4206 			handle_list = &conf->worker_groups[i].handle_list;
4207 			wg = &conf->worker_groups[i];
4208 			if (!list_empty(handle_list))
4209 				break;
4210 		}
4211 	}
4212 
4213 	pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4214 		  __func__,
4215 		  list_empty(handle_list) ? "empty" : "busy",
4216 		  list_empty(&conf->hold_list) ? "empty" : "busy",
4217 		  atomic_read(&conf->pending_full_writes), conf->bypass_count);
4218 
4219 	if (!list_empty(handle_list)) {
4220 		sh = list_entry(handle_list->next, typeof(*sh), lru);
4221 
4222 		if (list_empty(&conf->hold_list))
4223 			conf->bypass_count = 0;
4224 		else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4225 			if (conf->hold_list.next == conf->last_hold)
4226 				conf->bypass_count++;
4227 			else {
4228 				conf->last_hold = conf->hold_list.next;
4229 				conf->bypass_count -= conf->bypass_threshold;
4230 				if (conf->bypass_count < 0)
4231 					conf->bypass_count = 0;
4232 			}
4233 		}
4234 	} else if (!list_empty(&conf->hold_list) &&
4235 		   ((conf->bypass_threshold &&
4236 		     conf->bypass_count > conf->bypass_threshold) ||
4237 		    atomic_read(&conf->pending_full_writes) == 0)) {
4238 
4239 		list_for_each_entry(tmp, &conf->hold_list,  lru) {
4240 			if (conf->worker_cnt_per_group == 0 ||
4241 			    group == ANY_GROUP ||
4242 			    !cpu_online(tmp->cpu) ||
4243 			    cpu_to_group(tmp->cpu) == group) {
4244 				sh = tmp;
4245 				break;
4246 			}
4247 		}
4248 
4249 		if (sh) {
4250 			conf->bypass_count -= conf->bypass_threshold;
4251 			if (conf->bypass_count < 0)
4252 				conf->bypass_count = 0;
4253 		}
4254 		wg = NULL;
4255 	}
4256 
4257 	if (!sh)
4258 		return NULL;
4259 
4260 	if (wg) {
4261 		wg->stripes_cnt--;
4262 		sh->group = NULL;
4263 	}
4264 	list_del_init(&sh->lru);
4265 	atomic_inc(&sh->count);
4266 	BUG_ON(atomic_read(&sh->count) != 1);
4267 	return sh;
4268 }
4269 
4270 struct raid5_plug_cb {
4271 	struct blk_plug_cb	cb;
4272 	struct list_head	list;
4273 };
4274 
4275 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4276 {
4277 	struct raid5_plug_cb *cb = container_of(
4278 		blk_cb, struct raid5_plug_cb, cb);
4279 	struct stripe_head *sh;
4280 	struct mddev *mddev = cb->cb.data;
4281 	struct r5conf *conf = mddev->private;
4282 	int cnt = 0;
4283 
4284 	if (cb->list.next && !list_empty(&cb->list)) {
4285 		spin_lock_irq(&conf->device_lock);
4286 		while (!list_empty(&cb->list)) {
4287 			sh = list_first_entry(&cb->list, struct stripe_head, lru);
4288 			list_del_init(&sh->lru);
4289 			/*
4290 			 * avoid race release_stripe_plug() sees
4291 			 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4292 			 * is still in our list
4293 			 */
4294 			smp_mb__before_clear_bit();
4295 			clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4296 			/*
4297 			 * STRIPE_ON_RELEASE_LIST could be set here. In that
4298 			 * case, the count is always > 1 here
4299 			 */
4300 			__release_stripe(conf, sh);
4301 			cnt++;
4302 		}
4303 		spin_unlock_irq(&conf->device_lock);
4304 	}
4305 	if (mddev->queue)
4306 		trace_block_unplug(mddev->queue, cnt, !from_schedule);
4307 	kfree(cb);
4308 }
4309 
4310 static void release_stripe_plug(struct mddev *mddev,
4311 				struct stripe_head *sh)
4312 {
4313 	struct blk_plug_cb *blk_cb = blk_check_plugged(
4314 		raid5_unplug, mddev,
4315 		sizeof(struct raid5_plug_cb));
4316 	struct raid5_plug_cb *cb;
4317 
4318 	if (!blk_cb) {
4319 		release_stripe(sh);
4320 		return;
4321 	}
4322 
4323 	cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4324 
4325 	if (cb->list.next == NULL)
4326 		INIT_LIST_HEAD(&cb->list);
4327 
4328 	if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4329 		list_add_tail(&sh->lru, &cb->list);
4330 	else
4331 		release_stripe(sh);
4332 }
4333 
4334 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4335 {
4336 	struct r5conf *conf = mddev->private;
4337 	sector_t logical_sector, last_sector;
4338 	struct stripe_head *sh;
4339 	int remaining;
4340 	int stripe_sectors;
4341 
4342 	if (mddev->reshape_position != MaxSector)
4343 		/* Skip discard while reshape is happening */
4344 		return;
4345 
4346 	logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4347 	last_sector = bi->bi_sector + (bi->bi_size>>9);
4348 
4349 	bi->bi_next = NULL;
4350 	bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4351 
4352 	stripe_sectors = conf->chunk_sectors *
4353 		(conf->raid_disks - conf->max_degraded);
4354 	logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4355 					       stripe_sectors);
4356 	sector_div(last_sector, stripe_sectors);
4357 
4358 	logical_sector *= conf->chunk_sectors;
4359 	last_sector *= conf->chunk_sectors;
4360 
4361 	for (; logical_sector < last_sector;
4362 	     logical_sector += STRIPE_SECTORS) {
4363 		DEFINE_WAIT(w);
4364 		int d;
4365 	again:
4366 		sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4367 		prepare_to_wait(&conf->wait_for_overlap, &w,
4368 				TASK_UNINTERRUPTIBLE);
4369 		set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4370 		if (test_bit(STRIPE_SYNCING, &sh->state)) {
4371 			release_stripe(sh);
4372 			schedule();
4373 			goto again;
4374 		}
4375 		clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4376 		spin_lock_irq(&sh->stripe_lock);
4377 		for (d = 0; d < conf->raid_disks; d++) {
4378 			if (d == sh->pd_idx || d == sh->qd_idx)
4379 				continue;
4380 			if (sh->dev[d].towrite || sh->dev[d].toread) {
4381 				set_bit(R5_Overlap, &sh->dev[d].flags);
4382 				spin_unlock_irq(&sh->stripe_lock);
4383 				release_stripe(sh);
4384 				schedule();
4385 				goto again;
4386 			}
4387 		}
4388 		set_bit(STRIPE_DISCARD, &sh->state);
4389 		finish_wait(&conf->wait_for_overlap, &w);
4390 		for (d = 0; d < conf->raid_disks; d++) {
4391 			if (d == sh->pd_idx || d == sh->qd_idx)
4392 				continue;
4393 			sh->dev[d].towrite = bi;
4394 			set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4395 			raid5_inc_bi_active_stripes(bi);
4396 		}
4397 		spin_unlock_irq(&sh->stripe_lock);
4398 		if (conf->mddev->bitmap) {
4399 			for (d = 0;
4400 			     d < conf->raid_disks - conf->max_degraded;
4401 			     d++)
4402 				bitmap_startwrite(mddev->bitmap,
4403 						  sh->sector,
4404 						  STRIPE_SECTORS,
4405 						  0);
4406 			sh->bm_seq = conf->seq_flush + 1;
4407 			set_bit(STRIPE_BIT_DELAY, &sh->state);
4408 		}
4409 
4410 		set_bit(STRIPE_HANDLE, &sh->state);
4411 		clear_bit(STRIPE_DELAYED, &sh->state);
4412 		if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4413 			atomic_inc(&conf->preread_active_stripes);
4414 		release_stripe_plug(mddev, sh);
4415 	}
4416 
4417 	remaining = raid5_dec_bi_active_stripes(bi);
4418 	if (remaining == 0) {
4419 		md_write_end(mddev);
4420 		bio_endio(bi, 0);
4421 	}
4422 }
4423 
4424 static void make_request(struct mddev *mddev, struct bio * bi)
4425 {
4426 	struct r5conf *conf = mddev->private;
4427 	int dd_idx;
4428 	sector_t new_sector;
4429 	sector_t logical_sector, last_sector;
4430 	struct stripe_head *sh;
4431 	const int rw = bio_data_dir(bi);
4432 	int remaining;
4433 
4434 	if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4435 		md_flush_request(mddev, bi);
4436 		return;
4437 	}
4438 
4439 	md_write_start(mddev, bi);
4440 
4441 	if (rw == READ &&
4442 	     mddev->reshape_position == MaxSector &&
4443 	     chunk_aligned_read(mddev,bi))
4444 		return;
4445 
4446 	if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4447 		make_discard_request(mddev, bi);
4448 		return;
4449 	}
4450 
4451 	logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4452 	last_sector = bio_end_sector(bi);
4453 	bi->bi_next = NULL;
4454 	bi->bi_phys_segments = 1;	/* over-loaded to count active stripes */
4455 
4456 	for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4457 		DEFINE_WAIT(w);
4458 		int previous;
4459 		int seq;
4460 
4461 	retry:
4462 		seq = read_seqcount_begin(&conf->gen_lock);
4463 		previous = 0;
4464 		prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4465 		if (unlikely(conf->reshape_progress != MaxSector)) {
4466 			/* spinlock is needed as reshape_progress may be
4467 			 * 64bit on a 32bit platform, and so it might be
4468 			 * possible to see a half-updated value
4469 			 * Of course reshape_progress could change after
4470 			 * the lock is dropped, so once we get a reference
4471 			 * to the stripe that we think it is, we will have
4472 			 * to check again.
4473 			 */
4474 			spin_lock_irq(&conf->device_lock);
4475 			if (mddev->reshape_backwards
4476 			    ? logical_sector < conf->reshape_progress
4477 			    : logical_sector >= conf->reshape_progress) {
4478 				previous = 1;
4479 			} else {
4480 				if (mddev->reshape_backwards
4481 				    ? logical_sector < conf->reshape_safe
4482 				    : logical_sector >= conf->reshape_safe) {
4483 					spin_unlock_irq(&conf->device_lock);
4484 					schedule();
4485 					goto retry;
4486 				}
4487 			}
4488 			spin_unlock_irq(&conf->device_lock);
4489 		}
4490 
4491 		new_sector = raid5_compute_sector(conf, logical_sector,
4492 						  previous,
4493 						  &dd_idx, NULL);
4494 		pr_debug("raid456: make_request, sector %llu logical %llu\n",
4495 			(unsigned long long)new_sector,
4496 			(unsigned long long)logical_sector);
4497 
4498 		sh = get_active_stripe(conf, new_sector, previous,
4499 				       (bi->bi_rw&RWA_MASK), 0);
4500 		if (sh) {
4501 			if (unlikely(previous)) {
4502 				/* expansion might have moved on while waiting for a
4503 				 * stripe, so we must do the range check again.
4504 				 * Expansion could still move past after this
4505 				 * test, but as we are holding a reference to
4506 				 * 'sh', we know that if that happens,
4507 				 *  STRIPE_EXPANDING will get set and the expansion
4508 				 * won't proceed until we finish with the stripe.
4509 				 */
4510 				int must_retry = 0;
4511 				spin_lock_irq(&conf->device_lock);
4512 				if (mddev->reshape_backwards
4513 				    ? logical_sector >= conf->reshape_progress
4514 				    : logical_sector < conf->reshape_progress)
4515 					/* mismatch, need to try again */
4516 					must_retry = 1;
4517 				spin_unlock_irq(&conf->device_lock);
4518 				if (must_retry) {
4519 					release_stripe(sh);
4520 					schedule();
4521 					goto retry;
4522 				}
4523 			}
4524 			if (read_seqcount_retry(&conf->gen_lock, seq)) {
4525 				/* Might have got the wrong stripe_head
4526 				 * by accident
4527 				 */
4528 				release_stripe(sh);
4529 				goto retry;
4530 			}
4531 
4532 			if (rw == WRITE &&
4533 			    logical_sector >= mddev->suspend_lo &&
4534 			    logical_sector < mddev->suspend_hi) {
4535 				release_stripe(sh);
4536 				/* As the suspend_* range is controlled by
4537 				 * userspace, we want an interruptible
4538 				 * wait.
4539 				 */
4540 				flush_signals(current);
4541 				prepare_to_wait(&conf->wait_for_overlap,
4542 						&w, TASK_INTERRUPTIBLE);
4543 				if (logical_sector >= mddev->suspend_lo &&
4544 				    logical_sector < mddev->suspend_hi)
4545 					schedule();
4546 				goto retry;
4547 			}
4548 
4549 			if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4550 			    !add_stripe_bio(sh, bi, dd_idx, rw)) {
4551 				/* Stripe is busy expanding or
4552 				 * add failed due to overlap.  Flush everything
4553 				 * and wait a while
4554 				 */
4555 				md_wakeup_thread(mddev->thread);
4556 				release_stripe(sh);
4557 				schedule();
4558 				goto retry;
4559 			}
4560 			finish_wait(&conf->wait_for_overlap, &w);
4561 			set_bit(STRIPE_HANDLE, &sh->state);
4562 			clear_bit(STRIPE_DELAYED, &sh->state);
4563 			if ((bi->bi_rw & REQ_SYNC) &&
4564 			    !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4565 				atomic_inc(&conf->preread_active_stripes);
4566 			release_stripe_plug(mddev, sh);
4567 		} else {
4568 			/* cannot get stripe for read-ahead, just give-up */
4569 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
4570 			finish_wait(&conf->wait_for_overlap, &w);
4571 			break;
4572 		}
4573 	}
4574 
4575 	remaining = raid5_dec_bi_active_stripes(bi);
4576 	if (remaining == 0) {
4577 
4578 		if ( rw == WRITE )
4579 			md_write_end(mddev);
4580 
4581 		trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4582 					 bi, 0);
4583 		bio_endio(bi, 0);
4584 	}
4585 }
4586 
4587 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4588 
4589 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4590 {
4591 	/* reshaping is quite different to recovery/resync so it is
4592 	 * handled quite separately ... here.
4593 	 *
4594 	 * On each call to sync_request, we gather one chunk worth of
4595 	 * destination stripes and flag them as expanding.
4596 	 * Then we find all the source stripes and request reads.
4597 	 * As the reads complete, handle_stripe will copy the data
4598 	 * into the destination stripe and release that stripe.
4599 	 */
4600 	struct r5conf *conf = mddev->private;
4601 	struct stripe_head *sh;
4602 	sector_t first_sector, last_sector;
4603 	int raid_disks = conf->previous_raid_disks;
4604 	int data_disks = raid_disks - conf->max_degraded;
4605 	int new_data_disks = conf->raid_disks - conf->max_degraded;
4606 	int i;
4607 	int dd_idx;
4608 	sector_t writepos, readpos, safepos;
4609 	sector_t stripe_addr;
4610 	int reshape_sectors;
4611 	struct list_head stripes;
4612 
4613 	if (sector_nr == 0) {
4614 		/* If restarting in the middle, skip the initial sectors */
4615 		if (mddev->reshape_backwards &&
4616 		    conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4617 			sector_nr = raid5_size(mddev, 0, 0)
4618 				- conf->reshape_progress;
4619 		} else if (!mddev->reshape_backwards &&
4620 			   conf->reshape_progress > 0)
4621 			sector_nr = conf->reshape_progress;
4622 		sector_div(sector_nr, new_data_disks);
4623 		if (sector_nr) {
4624 			mddev->curr_resync_completed = sector_nr;
4625 			sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4626 			*skipped = 1;
4627 			return sector_nr;
4628 		}
4629 	}
4630 
4631 	/* We need to process a full chunk at a time.
4632 	 * If old and new chunk sizes differ, we need to process the
4633 	 * largest of these
4634 	 */
4635 	if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4636 		reshape_sectors = mddev->new_chunk_sectors;
4637 	else
4638 		reshape_sectors = mddev->chunk_sectors;
4639 
4640 	/* We update the metadata at least every 10 seconds, or when
4641 	 * the data about to be copied would over-write the source of
4642 	 * the data at the front of the range.  i.e. one new_stripe
4643 	 * along from reshape_progress new_maps to after where
4644 	 * reshape_safe old_maps to
4645 	 */
4646 	writepos = conf->reshape_progress;
4647 	sector_div(writepos, new_data_disks);
4648 	readpos = conf->reshape_progress;
4649 	sector_div(readpos, data_disks);
4650 	safepos = conf->reshape_safe;
4651 	sector_div(safepos, data_disks);
4652 	if (mddev->reshape_backwards) {
4653 		writepos -= min_t(sector_t, reshape_sectors, writepos);
4654 		readpos += reshape_sectors;
4655 		safepos += reshape_sectors;
4656 	} else {
4657 		writepos += reshape_sectors;
4658 		readpos -= min_t(sector_t, reshape_sectors, readpos);
4659 		safepos -= min_t(sector_t, reshape_sectors, safepos);
4660 	}
4661 
4662 	/* Having calculated the 'writepos' possibly use it
4663 	 * to set 'stripe_addr' which is where we will write to.
4664 	 */
4665 	if (mddev->reshape_backwards) {
4666 		BUG_ON(conf->reshape_progress == 0);
4667 		stripe_addr = writepos;
4668 		BUG_ON((mddev->dev_sectors &
4669 			~((sector_t)reshape_sectors - 1))
4670 		       - reshape_sectors - stripe_addr
4671 		       != sector_nr);
4672 	} else {
4673 		BUG_ON(writepos != sector_nr + reshape_sectors);
4674 		stripe_addr = sector_nr;
4675 	}
4676 
4677 	/* 'writepos' is the most advanced device address we might write.
4678 	 * 'readpos' is the least advanced device address we might read.
4679 	 * 'safepos' is the least address recorded in the metadata as having
4680 	 *     been reshaped.
4681 	 * If there is a min_offset_diff, these are adjusted either by
4682 	 * increasing the safepos/readpos if diff is negative, or
4683 	 * increasing writepos if diff is positive.
4684 	 * If 'readpos' is then behind 'writepos', there is no way that we can
4685 	 * ensure safety in the face of a crash - that must be done by userspace
4686 	 * making a backup of the data.  So in that case there is no particular
4687 	 * rush to update metadata.
4688 	 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4689 	 * update the metadata to advance 'safepos' to match 'readpos' so that
4690 	 * we can be safe in the event of a crash.
4691 	 * So we insist on updating metadata if safepos is behind writepos and
4692 	 * readpos is beyond writepos.
4693 	 * In any case, update the metadata every 10 seconds.
4694 	 * Maybe that number should be configurable, but I'm not sure it is
4695 	 * worth it.... maybe it could be a multiple of safemode_delay???
4696 	 */
4697 	if (conf->min_offset_diff < 0) {
4698 		safepos += -conf->min_offset_diff;
4699 		readpos += -conf->min_offset_diff;
4700 	} else
4701 		writepos += conf->min_offset_diff;
4702 
4703 	if ((mddev->reshape_backwards
4704 	     ? (safepos > writepos && readpos < writepos)
4705 	     : (safepos < writepos && readpos > writepos)) ||
4706 	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4707 		/* Cannot proceed until we've updated the superblock... */
4708 		wait_event(conf->wait_for_overlap,
4709 			   atomic_read(&conf->reshape_stripes)==0);
4710 		mddev->reshape_position = conf->reshape_progress;
4711 		mddev->curr_resync_completed = sector_nr;
4712 		conf->reshape_checkpoint = jiffies;
4713 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
4714 		md_wakeup_thread(mddev->thread);
4715 		wait_event(mddev->sb_wait, mddev->flags == 0 ||
4716 			   kthread_should_stop());
4717 		spin_lock_irq(&conf->device_lock);
4718 		conf->reshape_safe = mddev->reshape_position;
4719 		spin_unlock_irq(&conf->device_lock);
4720 		wake_up(&conf->wait_for_overlap);
4721 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4722 	}
4723 
4724 	INIT_LIST_HEAD(&stripes);
4725 	for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4726 		int j;
4727 		int skipped_disk = 0;
4728 		sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4729 		set_bit(STRIPE_EXPANDING, &sh->state);
4730 		atomic_inc(&conf->reshape_stripes);
4731 		/* If any of this stripe is beyond the end of the old
4732 		 * array, then we need to zero those blocks
4733 		 */
4734 		for (j=sh->disks; j--;) {
4735 			sector_t s;
4736 			if (j == sh->pd_idx)
4737 				continue;
4738 			if (conf->level == 6 &&
4739 			    j == sh->qd_idx)
4740 				continue;
4741 			s = compute_blocknr(sh, j, 0);
4742 			if (s < raid5_size(mddev, 0, 0)) {
4743 				skipped_disk = 1;
4744 				continue;
4745 			}
4746 			memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4747 			set_bit(R5_Expanded, &sh->dev[j].flags);
4748 			set_bit(R5_UPTODATE, &sh->dev[j].flags);
4749 		}
4750 		if (!skipped_disk) {
4751 			set_bit(STRIPE_EXPAND_READY, &sh->state);
4752 			set_bit(STRIPE_HANDLE, &sh->state);
4753 		}
4754 		list_add(&sh->lru, &stripes);
4755 	}
4756 	spin_lock_irq(&conf->device_lock);
4757 	if (mddev->reshape_backwards)
4758 		conf->reshape_progress -= reshape_sectors * new_data_disks;
4759 	else
4760 		conf->reshape_progress += reshape_sectors * new_data_disks;
4761 	spin_unlock_irq(&conf->device_lock);
4762 	/* Ok, those stripe are ready. We can start scheduling
4763 	 * reads on the source stripes.
4764 	 * The source stripes are determined by mapping the first and last
4765 	 * block on the destination stripes.
4766 	 */
4767 	first_sector =
4768 		raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4769 				     1, &dd_idx, NULL);
4770 	last_sector =
4771 		raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4772 					    * new_data_disks - 1),
4773 				     1, &dd_idx, NULL);
4774 	if (last_sector >= mddev->dev_sectors)
4775 		last_sector = mddev->dev_sectors - 1;
4776 	while (first_sector <= last_sector) {
4777 		sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4778 		set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4779 		set_bit(STRIPE_HANDLE, &sh->state);
4780 		release_stripe(sh);
4781 		first_sector += STRIPE_SECTORS;
4782 	}
4783 	/* Now that the sources are clearly marked, we can release
4784 	 * the destination stripes
4785 	 */
4786 	while (!list_empty(&stripes)) {
4787 		sh = list_entry(stripes.next, struct stripe_head, lru);
4788 		list_del_init(&sh->lru);
4789 		release_stripe(sh);
4790 	}
4791 	/* If this takes us to the resync_max point where we have to pause,
4792 	 * then we need to write out the superblock.
4793 	 */
4794 	sector_nr += reshape_sectors;
4795 	if ((sector_nr - mddev->curr_resync_completed) * 2
4796 	    >= mddev->resync_max - mddev->curr_resync_completed) {
4797 		/* Cannot proceed until we've updated the superblock... */
4798 		wait_event(conf->wait_for_overlap,
4799 			   atomic_read(&conf->reshape_stripes) == 0);
4800 		mddev->reshape_position = conf->reshape_progress;
4801 		mddev->curr_resync_completed = sector_nr;
4802 		conf->reshape_checkpoint = jiffies;
4803 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
4804 		md_wakeup_thread(mddev->thread);
4805 		wait_event(mddev->sb_wait,
4806 			   !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4807 			   || kthread_should_stop());
4808 		spin_lock_irq(&conf->device_lock);
4809 		conf->reshape_safe = mddev->reshape_position;
4810 		spin_unlock_irq(&conf->device_lock);
4811 		wake_up(&conf->wait_for_overlap);
4812 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4813 	}
4814 	return reshape_sectors;
4815 }
4816 
4817 /* FIXME go_faster isn't used */
4818 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4819 {
4820 	struct r5conf *conf = mddev->private;
4821 	struct stripe_head *sh;
4822 	sector_t max_sector = mddev->dev_sectors;
4823 	sector_t sync_blocks;
4824 	int still_degraded = 0;
4825 	int i;
4826 
4827 	if (sector_nr >= max_sector) {
4828 		/* just being told to finish up .. nothing much to do */
4829 
4830 		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4831 			end_reshape(conf);
4832 			return 0;
4833 		}
4834 
4835 		if (mddev->curr_resync < max_sector) /* aborted */
4836 			bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4837 					&sync_blocks, 1);
4838 		else /* completed sync */
4839 			conf->fullsync = 0;
4840 		bitmap_close_sync(mddev->bitmap);
4841 
4842 		return 0;
4843 	}
4844 
4845 	/* Allow raid5_quiesce to complete */
4846 	wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4847 
4848 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4849 		return reshape_request(mddev, sector_nr, skipped);
4850 
4851 	/* No need to check resync_max as we never do more than one
4852 	 * stripe, and as resync_max will always be on a chunk boundary,
4853 	 * if the check in md_do_sync didn't fire, there is no chance
4854 	 * of overstepping resync_max here
4855 	 */
4856 
4857 	/* if there is too many failed drives and we are trying
4858 	 * to resync, then assert that we are finished, because there is
4859 	 * nothing we can do.
4860 	 */
4861 	if (mddev->degraded >= conf->max_degraded &&
4862 	    test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4863 		sector_t rv = mddev->dev_sectors - sector_nr;
4864 		*skipped = 1;
4865 		return rv;
4866 	}
4867 	if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4868 	    !conf->fullsync &&
4869 	    !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4870 	    sync_blocks >= STRIPE_SECTORS) {
4871 		/* we can skip this block, and probably more */
4872 		sync_blocks /= STRIPE_SECTORS;
4873 		*skipped = 1;
4874 		return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4875 	}
4876 
4877 	bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4878 
4879 	sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4880 	if (sh == NULL) {
4881 		sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4882 		/* make sure we don't swamp the stripe cache if someone else
4883 		 * is trying to get access
4884 		 */
4885 		schedule_timeout_uninterruptible(1);
4886 	}
4887 	/* Need to check if array will still be degraded after recovery/resync
4888 	 * We don't need to check the 'failed' flag as when that gets set,
4889 	 * recovery aborts.
4890 	 */
4891 	for (i = 0; i < conf->raid_disks; i++)
4892 		if (conf->disks[i].rdev == NULL)
4893 			still_degraded = 1;
4894 
4895 	bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4896 
4897 	set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4898 
4899 	handle_stripe(sh);
4900 	release_stripe(sh);
4901 
4902 	return STRIPE_SECTORS;
4903 }
4904 
4905 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4906 {
4907 	/* We may not be able to submit a whole bio at once as there
4908 	 * may not be enough stripe_heads available.
4909 	 * We cannot pre-allocate enough stripe_heads as we may need
4910 	 * more than exist in the cache (if we allow ever large chunks).
4911 	 * So we do one stripe head at a time and record in
4912 	 * ->bi_hw_segments how many have been done.
4913 	 *
4914 	 * We *know* that this entire raid_bio is in one chunk, so
4915 	 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4916 	 */
4917 	struct stripe_head *sh;
4918 	int dd_idx;
4919 	sector_t sector, logical_sector, last_sector;
4920 	int scnt = 0;
4921 	int remaining;
4922 	int handled = 0;
4923 
4924 	logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4925 	sector = raid5_compute_sector(conf, logical_sector,
4926 				      0, &dd_idx, NULL);
4927 	last_sector = bio_end_sector(raid_bio);
4928 
4929 	for (; logical_sector < last_sector;
4930 	     logical_sector += STRIPE_SECTORS,
4931 		     sector += STRIPE_SECTORS,
4932 		     scnt++) {
4933 
4934 		if (scnt < raid5_bi_processed_stripes(raid_bio))
4935 			/* already done this stripe */
4936 			continue;
4937 
4938 		sh = get_active_stripe(conf, sector, 0, 1, 0);
4939 
4940 		if (!sh) {
4941 			/* failed to get a stripe - must wait */
4942 			raid5_set_bi_processed_stripes(raid_bio, scnt);
4943 			conf->retry_read_aligned = raid_bio;
4944 			return handled;
4945 		}
4946 
4947 		if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4948 			release_stripe(sh);
4949 			raid5_set_bi_processed_stripes(raid_bio, scnt);
4950 			conf->retry_read_aligned = raid_bio;
4951 			return handled;
4952 		}
4953 
4954 		set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4955 		handle_stripe(sh);
4956 		release_stripe(sh);
4957 		handled++;
4958 	}
4959 	remaining = raid5_dec_bi_active_stripes(raid_bio);
4960 	if (remaining == 0) {
4961 		trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
4962 					 raid_bio, 0);
4963 		bio_endio(raid_bio, 0);
4964 	}
4965 	if (atomic_dec_and_test(&conf->active_aligned_reads))
4966 		wake_up(&conf->wait_for_stripe);
4967 	return handled;
4968 }
4969 
4970 static int handle_active_stripes(struct r5conf *conf, int group,
4971 				 struct r5worker *worker)
4972 {
4973 	struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4974 	int i, batch_size = 0;
4975 
4976 	while (batch_size < MAX_STRIPE_BATCH &&
4977 			(sh = __get_priority_stripe(conf, group)) != NULL)
4978 		batch[batch_size++] = sh;
4979 
4980 	if (batch_size == 0)
4981 		return batch_size;
4982 	spin_unlock_irq(&conf->device_lock);
4983 
4984 	for (i = 0; i < batch_size; i++)
4985 		handle_stripe(batch[i]);
4986 
4987 	cond_resched();
4988 
4989 	spin_lock_irq(&conf->device_lock);
4990 	for (i = 0; i < batch_size; i++)
4991 		__release_stripe(conf, batch[i]);
4992 	return batch_size;
4993 }
4994 
4995 static void raid5_do_work(struct work_struct *work)
4996 {
4997 	struct r5worker *worker = container_of(work, struct r5worker, work);
4998 	struct r5worker_group *group = worker->group;
4999 	struct r5conf *conf = group->conf;
5000 	int group_id = group - conf->worker_groups;
5001 	int handled;
5002 	struct blk_plug plug;
5003 
5004 	pr_debug("+++ raid5worker active\n");
5005 
5006 	blk_start_plug(&plug);
5007 	handled = 0;
5008 	spin_lock_irq(&conf->device_lock);
5009 	while (1) {
5010 		int batch_size, released;
5011 
5012 		released = release_stripe_list(conf);
5013 
5014 		batch_size = handle_active_stripes(conf, group_id, worker);
5015 		worker->working = false;
5016 		if (!batch_size && !released)
5017 			break;
5018 		handled += batch_size;
5019 	}
5020 	pr_debug("%d stripes handled\n", handled);
5021 
5022 	spin_unlock_irq(&conf->device_lock);
5023 	blk_finish_plug(&plug);
5024 
5025 	pr_debug("--- raid5worker inactive\n");
5026 }
5027 
5028 /*
5029  * This is our raid5 kernel thread.
5030  *
5031  * We scan the hash table for stripes which can be handled now.
5032  * During the scan, completed stripes are saved for us by the interrupt
5033  * handler, so that they will not have to wait for our next wakeup.
5034  */
5035 static void raid5d(struct md_thread *thread)
5036 {
5037 	struct mddev *mddev = thread->mddev;
5038 	struct r5conf *conf = mddev->private;
5039 	int handled;
5040 	struct blk_plug plug;
5041 
5042 	pr_debug("+++ raid5d active\n");
5043 
5044 	md_check_recovery(mddev);
5045 
5046 	blk_start_plug(&plug);
5047 	handled = 0;
5048 	spin_lock_irq(&conf->device_lock);
5049 	while (1) {
5050 		struct bio *bio;
5051 		int batch_size, released;
5052 
5053 		released = release_stripe_list(conf);
5054 
5055 		if (
5056 		    !list_empty(&conf->bitmap_list)) {
5057 			/* Now is a good time to flush some bitmap updates */
5058 			conf->seq_flush++;
5059 			spin_unlock_irq(&conf->device_lock);
5060 			bitmap_unplug(mddev->bitmap);
5061 			spin_lock_irq(&conf->device_lock);
5062 			conf->seq_write = conf->seq_flush;
5063 			activate_bit_delay(conf);
5064 		}
5065 		raid5_activate_delayed(conf);
5066 
5067 		while ((bio = remove_bio_from_retry(conf))) {
5068 			int ok;
5069 			spin_unlock_irq(&conf->device_lock);
5070 			ok = retry_aligned_read(conf, bio);
5071 			spin_lock_irq(&conf->device_lock);
5072 			if (!ok)
5073 				break;
5074 			handled++;
5075 		}
5076 
5077 		batch_size = handle_active_stripes(conf, ANY_GROUP, NULL);
5078 		if (!batch_size && !released)
5079 			break;
5080 		handled += batch_size;
5081 
5082 		if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5083 			spin_unlock_irq(&conf->device_lock);
5084 			md_check_recovery(mddev);
5085 			spin_lock_irq(&conf->device_lock);
5086 		}
5087 	}
5088 	pr_debug("%d stripes handled\n", handled);
5089 
5090 	spin_unlock_irq(&conf->device_lock);
5091 
5092 	async_tx_issue_pending_all();
5093 	blk_finish_plug(&plug);
5094 
5095 	pr_debug("--- raid5d inactive\n");
5096 }
5097 
5098 static ssize_t
5099 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5100 {
5101 	struct r5conf *conf = mddev->private;
5102 	if (conf)
5103 		return sprintf(page, "%d\n", conf->max_nr_stripes);
5104 	else
5105 		return 0;
5106 }
5107 
5108 int
5109 raid5_set_cache_size(struct mddev *mddev, int size)
5110 {
5111 	struct r5conf *conf = mddev->private;
5112 	int err;
5113 
5114 	if (size <= 16 || size > 32768)
5115 		return -EINVAL;
5116 	while (size < conf->max_nr_stripes) {
5117 		if (drop_one_stripe(conf))
5118 			conf->max_nr_stripes--;
5119 		else
5120 			break;
5121 	}
5122 	err = md_allow_write(mddev);
5123 	if (err)
5124 		return err;
5125 	while (size > conf->max_nr_stripes) {
5126 		if (grow_one_stripe(conf))
5127 			conf->max_nr_stripes++;
5128 		else break;
5129 	}
5130 	return 0;
5131 }
5132 EXPORT_SYMBOL(raid5_set_cache_size);
5133 
5134 static ssize_t
5135 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5136 {
5137 	struct r5conf *conf = mddev->private;
5138 	unsigned long new;
5139 	int err;
5140 
5141 	if (len >= PAGE_SIZE)
5142 		return -EINVAL;
5143 	if (!conf)
5144 		return -ENODEV;
5145 
5146 	if (kstrtoul(page, 10, &new))
5147 		return -EINVAL;
5148 	err = raid5_set_cache_size(mddev, new);
5149 	if (err)
5150 		return err;
5151 	return len;
5152 }
5153 
5154 static struct md_sysfs_entry
5155 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5156 				raid5_show_stripe_cache_size,
5157 				raid5_store_stripe_cache_size);
5158 
5159 static ssize_t
5160 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5161 {
5162 	struct r5conf *conf = mddev->private;
5163 	if (conf)
5164 		return sprintf(page, "%d\n", conf->bypass_threshold);
5165 	else
5166 		return 0;
5167 }
5168 
5169 static ssize_t
5170 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
5171 {
5172 	struct r5conf *conf = mddev->private;
5173 	unsigned long new;
5174 	if (len >= PAGE_SIZE)
5175 		return -EINVAL;
5176 	if (!conf)
5177 		return -ENODEV;
5178 
5179 	if (kstrtoul(page, 10, &new))
5180 		return -EINVAL;
5181 	if (new > conf->max_nr_stripes)
5182 		return -EINVAL;
5183 	conf->bypass_threshold = new;
5184 	return len;
5185 }
5186 
5187 static struct md_sysfs_entry
5188 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5189 					S_IRUGO | S_IWUSR,
5190 					raid5_show_preread_threshold,
5191 					raid5_store_preread_threshold);
5192 
5193 static ssize_t
5194 stripe_cache_active_show(struct mddev *mddev, char *page)
5195 {
5196 	struct r5conf *conf = mddev->private;
5197 	if (conf)
5198 		return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5199 	else
5200 		return 0;
5201 }
5202 
5203 static struct md_sysfs_entry
5204 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
5205 
5206 static ssize_t
5207 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
5208 {
5209 	struct r5conf *conf = mddev->private;
5210 	if (conf)
5211 		return sprintf(page, "%d\n", conf->worker_cnt_per_group);
5212 	else
5213 		return 0;
5214 }
5215 
5216 static int alloc_thread_groups(struct r5conf *conf, int cnt);
5217 static ssize_t
5218 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
5219 {
5220 	struct r5conf *conf = mddev->private;
5221 	unsigned long new;
5222 	int err;
5223 	struct r5worker_group *old_groups;
5224 	int old_group_cnt;
5225 
5226 	if (len >= PAGE_SIZE)
5227 		return -EINVAL;
5228 	if (!conf)
5229 		return -ENODEV;
5230 
5231 	if (kstrtoul(page, 10, &new))
5232 		return -EINVAL;
5233 
5234 	if (new == conf->worker_cnt_per_group)
5235 		return len;
5236 
5237 	mddev_suspend(mddev);
5238 
5239 	old_groups = conf->worker_groups;
5240 	old_group_cnt = conf->worker_cnt_per_group;
5241 
5242 	conf->worker_groups = NULL;
5243 	err = alloc_thread_groups(conf, new);
5244 	if (err) {
5245 		conf->worker_groups = old_groups;
5246 		conf->worker_cnt_per_group = old_group_cnt;
5247 	} else {
5248 		if (old_groups)
5249 			kfree(old_groups[0].workers);
5250 		kfree(old_groups);
5251 	}
5252 
5253 	mddev_resume(mddev);
5254 
5255 	if (err)
5256 		return err;
5257 	return len;
5258 }
5259 
5260 static struct md_sysfs_entry
5261 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
5262 				raid5_show_group_thread_cnt,
5263 				raid5_store_group_thread_cnt);
5264 
5265 static struct attribute *raid5_attrs[] =  {
5266 	&raid5_stripecache_size.attr,
5267 	&raid5_stripecache_active.attr,
5268 	&raid5_preread_bypass_threshold.attr,
5269 	&raid5_group_thread_cnt.attr,
5270 	NULL,
5271 };
5272 static struct attribute_group raid5_attrs_group = {
5273 	.name = NULL,
5274 	.attrs = raid5_attrs,
5275 };
5276 
5277 static int alloc_thread_groups(struct r5conf *conf, int cnt)
5278 {
5279 	int i, j;
5280 	ssize_t size;
5281 	struct r5worker *workers;
5282 
5283 	conf->worker_cnt_per_group = cnt;
5284 	if (cnt == 0) {
5285 		conf->worker_groups = NULL;
5286 		return 0;
5287 	}
5288 	conf->group_cnt = num_possible_nodes();
5289 	size = sizeof(struct r5worker) * cnt;
5290 	workers = kzalloc(size * conf->group_cnt, GFP_NOIO);
5291 	conf->worker_groups = kzalloc(sizeof(struct r5worker_group) *
5292 				conf->group_cnt, GFP_NOIO);
5293 	if (!conf->worker_groups || !workers) {
5294 		kfree(workers);
5295 		kfree(conf->worker_groups);
5296 		conf->worker_groups = NULL;
5297 		return -ENOMEM;
5298 	}
5299 
5300 	for (i = 0; i < conf->group_cnt; i++) {
5301 		struct r5worker_group *group;
5302 
5303 		group = &conf->worker_groups[i];
5304 		INIT_LIST_HEAD(&group->handle_list);
5305 		group->conf = conf;
5306 		group->workers = workers + i * cnt;
5307 
5308 		for (j = 0; j < cnt; j++) {
5309 			group->workers[j].group = group;
5310 			INIT_WORK(&group->workers[j].work, raid5_do_work);
5311 		}
5312 	}
5313 
5314 	return 0;
5315 }
5316 
5317 static void free_thread_groups(struct r5conf *conf)
5318 {
5319 	if (conf->worker_groups)
5320 		kfree(conf->worker_groups[0].workers);
5321 	kfree(conf->worker_groups);
5322 	conf->worker_groups = NULL;
5323 }
5324 
5325 static sector_t
5326 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
5327 {
5328 	struct r5conf *conf = mddev->private;
5329 
5330 	if (!sectors)
5331 		sectors = mddev->dev_sectors;
5332 	if (!raid_disks)
5333 		/* size is defined by the smallest of previous and new size */
5334 		raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
5335 
5336 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5337 	sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
5338 	return sectors * (raid_disks - conf->max_degraded);
5339 }
5340 
5341 static void raid5_free_percpu(struct r5conf *conf)
5342 {
5343 	struct raid5_percpu *percpu;
5344 	unsigned long cpu;
5345 
5346 	if (!conf->percpu)
5347 		return;
5348 
5349 	get_online_cpus();
5350 	for_each_possible_cpu(cpu) {
5351 		percpu = per_cpu_ptr(conf->percpu, cpu);
5352 		safe_put_page(percpu->spare_page);
5353 		kfree(percpu->scribble);
5354 	}
5355 #ifdef CONFIG_HOTPLUG_CPU
5356 	unregister_cpu_notifier(&conf->cpu_notify);
5357 #endif
5358 	put_online_cpus();
5359 
5360 	free_percpu(conf->percpu);
5361 }
5362 
5363 static void free_conf(struct r5conf *conf)
5364 {
5365 	free_thread_groups(conf);
5366 	shrink_stripes(conf);
5367 	raid5_free_percpu(conf);
5368 	kfree(conf->disks);
5369 	kfree(conf->stripe_hashtbl);
5370 	kfree(conf);
5371 }
5372 
5373 #ifdef CONFIG_HOTPLUG_CPU
5374 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5375 			      void *hcpu)
5376 {
5377 	struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5378 	long cpu = (long)hcpu;
5379 	struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5380 
5381 	switch (action) {
5382 	case CPU_UP_PREPARE:
5383 	case CPU_UP_PREPARE_FROZEN:
5384 		if (conf->level == 6 && !percpu->spare_page)
5385 			percpu->spare_page = alloc_page(GFP_KERNEL);
5386 		if (!percpu->scribble)
5387 			percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5388 
5389 		if (!percpu->scribble ||
5390 		    (conf->level == 6 && !percpu->spare_page)) {
5391 			safe_put_page(percpu->spare_page);
5392 			kfree(percpu->scribble);
5393 			pr_err("%s: failed memory allocation for cpu%ld\n",
5394 			       __func__, cpu);
5395 			return notifier_from_errno(-ENOMEM);
5396 		}
5397 		break;
5398 	case CPU_DEAD:
5399 	case CPU_DEAD_FROZEN:
5400 		safe_put_page(percpu->spare_page);
5401 		kfree(percpu->scribble);
5402 		percpu->spare_page = NULL;
5403 		percpu->scribble = NULL;
5404 		break;
5405 	default:
5406 		break;
5407 	}
5408 	return NOTIFY_OK;
5409 }
5410 #endif
5411 
5412 static int raid5_alloc_percpu(struct r5conf *conf)
5413 {
5414 	unsigned long cpu;
5415 	struct page *spare_page;
5416 	struct raid5_percpu __percpu *allcpus;
5417 	void *scribble;
5418 	int err;
5419 
5420 	allcpus = alloc_percpu(struct raid5_percpu);
5421 	if (!allcpus)
5422 		return -ENOMEM;
5423 	conf->percpu = allcpus;
5424 
5425 	get_online_cpus();
5426 	err = 0;
5427 	for_each_present_cpu(cpu) {
5428 		if (conf->level == 6) {
5429 			spare_page = alloc_page(GFP_KERNEL);
5430 			if (!spare_page) {
5431 				err = -ENOMEM;
5432 				break;
5433 			}
5434 			per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5435 		}
5436 		scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5437 		if (!scribble) {
5438 			err = -ENOMEM;
5439 			break;
5440 		}
5441 		per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
5442 	}
5443 #ifdef CONFIG_HOTPLUG_CPU
5444 	conf->cpu_notify.notifier_call = raid456_cpu_notify;
5445 	conf->cpu_notify.priority = 0;
5446 	if (err == 0)
5447 		err = register_cpu_notifier(&conf->cpu_notify);
5448 #endif
5449 	put_online_cpus();
5450 
5451 	return err;
5452 }
5453 
5454 static struct r5conf *setup_conf(struct mddev *mddev)
5455 {
5456 	struct r5conf *conf;
5457 	int raid_disk, memory, max_disks;
5458 	struct md_rdev *rdev;
5459 	struct disk_info *disk;
5460 	char pers_name[6];
5461 
5462 	if (mddev->new_level != 5
5463 	    && mddev->new_level != 4
5464 	    && mddev->new_level != 6) {
5465 		printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5466 		       mdname(mddev), mddev->new_level);
5467 		return ERR_PTR(-EIO);
5468 	}
5469 	if ((mddev->new_level == 5
5470 	     && !algorithm_valid_raid5(mddev->new_layout)) ||
5471 	    (mddev->new_level == 6
5472 	     && !algorithm_valid_raid6(mddev->new_layout))) {
5473 		printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5474 		       mdname(mddev), mddev->new_layout);
5475 		return ERR_PTR(-EIO);
5476 	}
5477 	if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5478 		printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5479 		       mdname(mddev), mddev->raid_disks);
5480 		return ERR_PTR(-EINVAL);
5481 	}
5482 
5483 	if (!mddev->new_chunk_sectors ||
5484 	    (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5485 	    !is_power_of_2(mddev->new_chunk_sectors)) {
5486 		printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5487 		       mdname(mddev), mddev->new_chunk_sectors << 9);
5488 		return ERR_PTR(-EINVAL);
5489 	}
5490 
5491 	conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5492 	if (conf == NULL)
5493 		goto abort;
5494 	/* Don't enable multi-threading by default*/
5495 	if (alloc_thread_groups(conf, 0))
5496 		goto abort;
5497 	spin_lock_init(&conf->device_lock);
5498 	seqcount_init(&conf->gen_lock);
5499 	init_waitqueue_head(&conf->wait_for_stripe);
5500 	init_waitqueue_head(&conf->wait_for_overlap);
5501 	INIT_LIST_HEAD(&conf->handle_list);
5502 	INIT_LIST_HEAD(&conf->hold_list);
5503 	INIT_LIST_HEAD(&conf->delayed_list);
5504 	INIT_LIST_HEAD(&conf->bitmap_list);
5505 	INIT_LIST_HEAD(&conf->inactive_list);
5506 	init_llist_head(&conf->released_stripes);
5507 	atomic_set(&conf->active_stripes, 0);
5508 	atomic_set(&conf->preread_active_stripes, 0);
5509 	atomic_set(&conf->active_aligned_reads, 0);
5510 	conf->bypass_threshold = BYPASS_THRESHOLD;
5511 	conf->recovery_disabled = mddev->recovery_disabled - 1;
5512 
5513 	conf->raid_disks = mddev->raid_disks;
5514 	if (mddev->reshape_position == MaxSector)
5515 		conf->previous_raid_disks = mddev->raid_disks;
5516 	else
5517 		conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5518 	max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5519 	conf->scribble_len = scribble_len(max_disks);
5520 
5521 	conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5522 			      GFP_KERNEL);
5523 	if (!conf->disks)
5524 		goto abort;
5525 
5526 	conf->mddev = mddev;
5527 
5528 	if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5529 		goto abort;
5530 
5531 	conf->level = mddev->new_level;
5532 	if (raid5_alloc_percpu(conf) != 0)
5533 		goto abort;
5534 
5535 	pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5536 
5537 	rdev_for_each(rdev, mddev) {
5538 		raid_disk = rdev->raid_disk;
5539 		if (raid_disk >= max_disks
5540 		    || raid_disk < 0)
5541 			continue;
5542 		disk = conf->disks + raid_disk;
5543 
5544 		if (test_bit(Replacement, &rdev->flags)) {
5545 			if (disk->replacement)
5546 				goto abort;
5547 			disk->replacement = rdev;
5548 		} else {
5549 			if (disk->rdev)
5550 				goto abort;
5551 			disk->rdev = rdev;
5552 		}
5553 
5554 		if (test_bit(In_sync, &rdev->flags)) {
5555 			char b[BDEVNAME_SIZE];
5556 			printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5557 			       " disk %d\n",
5558 			       mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5559 		} else if (rdev->saved_raid_disk != raid_disk)
5560 			/* Cannot rely on bitmap to complete recovery */
5561 			conf->fullsync = 1;
5562 	}
5563 
5564 	conf->chunk_sectors = mddev->new_chunk_sectors;
5565 	conf->level = mddev->new_level;
5566 	if (conf->level == 6)
5567 		conf->max_degraded = 2;
5568 	else
5569 		conf->max_degraded = 1;
5570 	conf->algorithm = mddev->new_layout;
5571 	conf->max_nr_stripes = NR_STRIPES;
5572 	conf->reshape_progress = mddev->reshape_position;
5573 	if (conf->reshape_progress != MaxSector) {
5574 		conf->prev_chunk_sectors = mddev->chunk_sectors;
5575 		conf->prev_algo = mddev->layout;
5576 	}
5577 
5578 	memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5579 		 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5580 	if (grow_stripes(conf, conf->max_nr_stripes)) {
5581 		printk(KERN_ERR
5582 		       "md/raid:%s: couldn't allocate %dkB for buffers\n",
5583 		       mdname(mddev), memory);
5584 		goto abort;
5585 	} else
5586 		printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5587 		       mdname(mddev), memory);
5588 
5589 	sprintf(pers_name, "raid%d", mddev->new_level);
5590 	conf->thread = md_register_thread(raid5d, mddev, pers_name);
5591 	if (!conf->thread) {
5592 		printk(KERN_ERR
5593 		       "md/raid:%s: couldn't allocate thread.\n",
5594 		       mdname(mddev));
5595 		goto abort;
5596 	}
5597 
5598 	return conf;
5599 
5600  abort:
5601 	if (conf) {
5602 		free_conf(conf);
5603 		return ERR_PTR(-EIO);
5604 	} else
5605 		return ERR_PTR(-ENOMEM);
5606 }
5607 
5608 
5609 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5610 {
5611 	switch (algo) {
5612 	case ALGORITHM_PARITY_0:
5613 		if (raid_disk < max_degraded)
5614 			return 1;
5615 		break;
5616 	case ALGORITHM_PARITY_N:
5617 		if (raid_disk >= raid_disks - max_degraded)
5618 			return 1;
5619 		break;
5620 	case ALGORITHM_PARITY_0_6:
5621 		if (raid_disk == 0 ||
5622 		    raid_disk == raid_disks - 1)
5623 			return 1;
5624 		break;
5625 	case ALGORITHM_LEFT_ASYMMETRIC_6:
5626 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
5627 	case ALGORITHM_LEFT_SYMMETRIC_6:
5628 	case ALGORITHM_RIGHT_SYMMETRIC_6:
5629 		if (raid_disk == raid_disks - 1)
5630 			return 1;
5631 	}
5632 	return 0;
5633 }
5634 
5635 static int run(struct mddev *mddev)
5636 {
5637 	struct r5conf *conf;
5638 	int working_disks = 0;
5639 	int dirty_parity_disks = 0;
5640 	struct md_rdev *rdev;
5641 	sector_t reshape_offset = 0;
5642 	int i;
5643 	long long min_offset_diff = 0;
5644 	int first = 1;
5645 
5646 	if (mddev->recovery_cp != MaxSector)
5647 		printk(KERN_NOTICE "md/raid:%s: not clean"
5648 		       " -- starting background reconstruction\n",
5649 		       mdname(mddev));
5650 
5651 	rdev_for_each(rdev, mddev) {
5652 		long long diff;
5653 		if (rdev->raid_disk < 0)
5654 			continue;
5655 		diff = (rdev->new_data_offset - rdev->data_offset);
5656 		if (first) {
5657 			min_offset_diff = diff;
5658 			first = 0;
5659 		} else if (mddev->reshape_backwards &&
5660 			 diff < min_offset_diff)
5661 			min_offset_diff = diff;
5662 		else if (!mddev->reshape_backwards &&
5663 			 diff > min_offset_diff)
5664 			min_offset_diff = diff;
5665 	}
5666 
5667 	if (mddev->reshape_position != MaxSector) {
5668 		/* Check that we can continue the reshape.
5669 		 * Difficulties arise if the stripe we would write to
5670 		 * next is at or after the stripe we would read from next.
5671 		 * For a reshape that changes the number of devices, this
5672 		 * is only possible for a very short time, and mdadm makes
5673 		 * sure that time appears to have past before assembling
5674 		 * the array.  So we fail if that time hasn't passed.
5675 		 * For a reshape that keeps the number of devices the same
5676 		 * mdadm must be monitoring the reshape can keeping the
5677 		 * critical areas read-only and backed up.  It will start
5678 		 * the array in read-only mode, so we check for that.
5679 		 */
5680 		sector_t here_new, here_old;
5681 		int old_disks;
5682 		int max_degraded = (mddev->level == 6 ? 2 : 1);
5683 
5684 		if (mddev->new_level != mddev->level) {
5685 			printk(KERN_ERR "md/raid:%s: unsupported reshape "
5686 			       "required - aborting.\n",
5687 			       mdname(mddev));
5688 			return -EINVAL;
5689 		}
5690 		old_disks = mddev->raid_disks - mddev->delta_disks;
5691 		/* reshape_position must be on a new-stripe boundary, and one
5692 		 * further up in new geometry must map after here in old
5693 		 * geometry.
5694 		 */
5695 		here_new = mddev->reshape_position;
5696 		if (sector_div(here_new, mddev->new_chunk_sectors *
5697 			       (mddev->raid_disks - max_degraded))) {
5698 			printk(KERN_ERR "md/raid:%s: reshape_position not "
5699 			       "on a stripe boundary\n", mdname(mddev));
5700 			return -EINVAL;
5701 		}
5702 		reshape_offset = here_new * mddev->new_chunk_sectors;
5703 		/* here_new is the stripe we will write to */
5704 		here_old = mddev->reshape_position;
5705 		sector_div(here_old, mddev->chunk_sectors *
5706 			   (old_disks-max_degraded));
5707 		/* here_old is the first stripe that we might need to read
5708 		 * from */
5709 		if (mddev->delta_disks == 0) {
5710 			if ((here_new * mddev->new_chunk_sectors !=
5711 			     here_old * mddev->chunk_sectors)) {
5712 				printk(KERN_ERR "md/raid:%s: reshape position is"
5713 				       " confused - aborting\n", mdname(mddev));
5714 				return -EINVAL;
5715 			}
5716 			/* We cannot be sure it is safe to start an in-place
5717 			 * reshape.  It is only safe if user-space is monitoring
5718 			 * and taking constant backups.
5719 			 * mdadm always starts a situation like this in
5720 			 * readonly mode so it can take control before
5721 			 * allowing any writes.  So just check for that.
5722 			 */
5723 			if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5724 			    abs(min_offset_diff) >= mddev->new_chunk_sectors)
5725 				/* not really in-place - so OK */;
5726 			else if (mddev->ro == 0) {
5727 				printk(KERN_ERR "md/raid:%s: in-place reshape "
5728 				       "must be started in read-only mode "
5729 				       "- aborting\n",
5730 				       mdname(mddev));
5731 				return -EINVAL;
5732 			}
5733 		} else if (mddev->reshape_backwards
5734 		    ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5735 		       here_old * mddev->chunk_sectors)
5736 		    : (here_new * mddev->new_chunk_sectors >=
5737 		       here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5738 			/* Reading from the same stripe as writing to - bad */
5739 			printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5740 			       "auto-recovery - aborting.\n",
5741 			       mdname(mddev));
5742 			return -EINVAL;
5743 		}
5744 		printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5745 		       mdname(mddev));
5746 		/* OK, we should be able to continue; */
5747 	} else {
5748 		BUG_ON(mddev->level != mddev->new_level);
5749 		BUG_ON(mddev->layout != mddev->new_layout);
5750 		BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5751 		BUG_ON(mddev->delta_disks != 0);
5752 	}
5753 
5754 	if (mddev->private == NULL)
5755 		conf = setup_conf(mddev);
5756 	else
5757 		conf = mddev->private;
5758 
5759 	if (IS_ERR(conf))
5760 		return PTR_ERR(conf);
5761 
5762 	conf->min_offset_diff = min_offset_diff;
5763 	mddev->thread = conf->thread;
5764 	conf->thread = NULL;
5765 	mddev->private = conf;
5766 
5767 	for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5768 	     i++) {
5769 		rdev = conf->disks[i].rdev;
5770 		if (!rdev && conf->disks[i].replacement) {
5771 			/* The replacement is all we have yet */
5772 			rdev = conf->disks[i].replacement;
5773 			conf->disks[i].replacement = NULL;
5774 			clear_bit(Replacement, &rdev->flags);
5775 			conf->disks[i].rdev = rdev;
5776 		}
5777 		if (!rdev)
5778 			continue;
5779 		if (conf->disks[i].replacement &&
5780 		    conf->reshape_progress != MaxSector) {
5781 			/* replacements and reshape simply do not mix. */
5782 			printk(KERN_ERR "md: cannot handle concurrent "
5783 			       "replacement and reshape.\n");
5784 			goto abort;
5785 		}
5786 		if (test_bit(In_sync, &rdev->flags)) {
5787 			working_disks++;
5788 			continue;
5789 		}
5790 		/* This disc is not fully in-sync.  However if it
5791 		 * just stored parity (beyond the recovery_offset),
5792 		 * when we don't need to be concerned about the
5793 		 * array being dirty.
5794 		 * When reshape goes 'backwards', we never have
5795 		 * partially completed devices, so we only need
5796 		 * to worry about reshape going forwards.
5797 		 */
5798 		/* Hack because v0.91 doesn't store recovery_offset properly. */
5799 		if (mddev->major_version == 0 &&
5800 		    mddev->minor_version > 90)
5801 			rdev->recovery_offset = reshape_offset;
5802 
5803 		if (rdev->recovery_offset < reshape_offset) {
5804 			/* We need to check old and new layout */
5805 			if (!only_parity(rdev->raid_disk,
5806 					 conf->algorithm,
5807 					 conf->raid_disks,
5808 					 conf->max_degraded))
5809 				continue;
5810 		}
5811 		if (!only_parity(rdev->raid_disk,
5812 				 conf->prev_algo,
5813 				 conf->previous_raid_disks,
5814 				 conf->max_degraded))
5815 			continue;
5816 		dirty_parity_disks++;
5817 	}
5818 
5819 	/*
5820 	 * 0 for a fully functional array, 1 or 2 for a degraded array.
5821 	 */
5822 	mddev->degraded = calc_degraded(conf);
5823 
5824 	if (has_failed(conf)) {
5825 		printk(KERN_ERR "md/raid:%s: not enough operational devices"
5826 			" (%d/%d failed)\n",
5827 			mdname(mddev), mddev->degraded, conf->raid_disks);
5828 		goto abort;
5829 	}
5830 
5831 	/* device size must be a multiple of chunk size */
5832 	mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5833 	mddev->resync_max_sectors = mddev->dev_sectors;
5834 
5835 	if (mddev->degraded > dirty_parity_disks &&
5836 	    mddev->recovery_cp != MaxSector) {
5837 		if (mddev->ok_start_degraded)
5838 			printk(KERN_WARNING
5839 			       "md/raid:%s: starting dirty degraded array"
5840 			       " - data corruption possible.\n",
5841 			       mdname(mddev));
5842 		else {
5843 			printk(KERN_ERR
5844 			       "md/raid:%s: cannot start dirty degraded array.\n",
5845 			       mdname(mddev));
5846 			goto abort;
5847 		}
5848 	}
5849 
5850 	if (mddev->degraded == 0)
5851 		printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5852 		       " devices, algorithm %d\n", mdname(mddev), conf->level,
5853 		       mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5854 		       mddev->new_layout);
5855 	else
5856 		printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5857 		       " out of %d devices, algorithm %d\n",
5858 		       mdname(mddev), conf->level,
5859 		       mddev->raid_disks - mddev->degraded,
5860 		       mddev->raid_disks, mddev->new_layout);
5861 
5862 	print_raid5_conf(conf);
5863 
5864 	if (conf->reshape_progress != MaxSector) {
5865 		conf->reshape_safe = conf->reshape_progress;
5866 		atomic_set(&conf->reshape_stripes, 0);
5867 		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5868 		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5869 		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5870 		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5871 		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5872 							"reshape");
5873 	}
5874 
5875 
5876 	/* Ok, everything is just fine now */
5877 	if (mddev->to_remove == &raid5_attrs_group)
5878 		mddev->to_remove = NULL;
5879 	else if (mddev->kobj.sd &&
5880 	    sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5881 		printk(KERN_WARNING
5882 		       "raid5: failed to create sysfs attributes for %s\n",
5883 		       mdname(mddev));
5884 	md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5885 
5886 	if (mddev->queue) {
5887 		int chunk_size;
5888 		bool discard_supported = true;
5889 		/* read-ahead size must cover two whole stripes, which
5890 		 * is 2 * (datadisks) * chunksize where 'n' is the
5891 		 * number of raid devices
5892 		 */
5893 		int data_disks = conf->previous_raid_disks - conf->max_degraded;
5894 		int stripe = data_disks *
5895 			((mddev->chunk_sectors << 9) / PAGE_SIZE);
5896 		if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5897 			mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5898 
5899 		blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5900 
5901 		mddev->queue->backing_dev_info.congested_data = mddev;
5902 		mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5903 
5904 		chunk_size = mddev->chunk_sectors << 9;
5905 		blk_queue_io_min(mddev->queue, chunk_size);
5906 		blk_queue_io_opt(mddev->queue, chunk_size *
5907 				 (conf->raid_disks - conf->max_degraded));
5908 		/*
5909 		 * We can only discard a whole stripe. It doesn't make sense to
5910 		 * discard data disk but write parity disk
5911 		 */
5912 		stripe = stripe * PAGE_SIZE;
5913 		/* Round up to power of 2, as discard handling
5914 		 * currently assumes that */
5915 		while ((stripe-1) & stripe)
5916 			stripe = (stripe | (stripe-1)) + 1;
5917 		mddev->queue->limits.discard_alignment = stripe;
5918 		mddev->queue->limits.discard_granularity = stripe;
5919 		/*
5920 		 * unaligned part of discard request will be ignored, so can't
5921 		 * guarantee discard_zerors_data
5922 		 */
5923 		mddev->queue->limits.discard_zeroes_data = 0;
5924 
5925 		blk_queue_max_write_same_sectors(mddev->queue, 0);
5926 
5927 		rdev_for_each(rdev, mddev) {
5928 			disk_stack_limits(mddev->gendisk, rdev->bdev,
5929 					  rdev->data_offset << 9);
5930 			disk_stack_limits(mddev->gendisk, rdev->bdev,
5931 					  rdev->new_data_offset << 9);
5932 			/*
5933 			 * discard_zeroes_data is required, otherwise data
5934 			 * could be lost. Consider a scenario: discard a stripe
5935 			 * (the stripe could be inconsistent if
5936 			 * discard_zeroes_data is 0); write one disk of the
5937 			 * stripe (the stripe could be inconsistent again
5938 			 * depending on which disks are used to calculate
5939 			 * parity); the disk is broken; The stripe data of this
5940 			 * disk is lost.
5941 			 */
5942 			if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5943 			    !bdev_get_queue(rdev->bdev)->
5944 						limits.discard_zeroes_data)
5945 				discard_supported = false;
5946 		}
5947 
5948 		if (discard_supported &&
5949 		   mddev->queue->limits.max_discard_sectors >= stripe &&
5950 		   mddev->queue->limits.discard_granularity >= stripe)
5951 			queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5952 						mddev->queue);
5953 		else
5954 			queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5955 						mddev->queue);
5956 	}
5957 
5958 	return 0;
5959 abort:
5960 	md_unregister_thread(&mddev->thread);
5961 	print_raid5_conf(conf);
5962 	free_conf(conf);
5963 	mddev->private = NULL;
5964 	printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5965 	return -EIO;
5966 }
5967 
5968 static int stop(struct mddev *mddev)
5969 {
5970 	struct r5conf *conf = mddev->private;
5971 
5972 	md_unregister_thread(&mddev->thread);
5973 	if (mddev->queue)
5974 		mddev->queue->backing_dev_info.congested_fn = NULL;
5975 	free_conf(conf);
5976 	mddev->private = NULL;
5977 	mddev->to_remove = &raid5_attrs_group;
5978 	return 0;
5979 }
5980 
5981 static void status(struct seq_file *seq, struct mddev *mddev)
5982 {
5983 	struct r5conf *conf = mddev->private;
5984 	int i;
5985 
5986 	seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5987 		mddev->chunk_sectors / 2, mddev->layout);
5988 	seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5989 	for (i = 0; i < conf->raid_disks; i++)
5990 		seq_printf (seq, "%s",
5991 			       conf->disks[i].rdev &&
5992 			       test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5993 	seq_printf (seq, "]");
5994 }
5995 
5996 static void print_raid5_conf (struct r5conf *conf)
5997 {
5998 	int i;
5999 	struct disk_info *tmp;
6000 
6001 	printk(KERN_DEBUG "RAID conf printout:\n");
6002 	if (!conf) {
6003 		printk("(conf==NULL)\n");
6004 		return;
6005 	}
6006 	printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6007 	       conf->raid_disks,
6008 	       conf->raid_disks - conf->mddev->degraded);
6009 
6010 	for (i = 0; i < conf->raid_disks; i++) {
6011 		char b[BDEVNAME_SIZE];
6012 		tmp = conf->disks + i;
6013 		if (tmp->rdev)
6014 			printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6015 			       i, !test_bit(Faulty, &tmp->rdev->flags),
6016 			       bdevname(tmp->rdev->bdev, b));
6017 	}
6018 }
6019 
6020 static int raid5_spare_active(struct mddev *mddev)
6021 {
6022 	int i;
6023 	struct r5conf *conf = mddev->private;
6024 	struct disk_info *tmp;
6025 	int count = 0;
6026 	unsigned long flags;
6027 
6028 	for (i = 0; i < conf->raid_disks; i++) {
6029 		tmp = conf->disks + i;
6030 		if (tmp->replacement
6031 		    && tmp->replacement->recovery_offset == MaxSector
6032 		    && !test_bit(Faulty, &tmp->replacement->flags)
6033 		    && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
6034 			/* Replacement has just become active. */
6035 			if (!tmp->rdev
6036 			    || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
6037 				count++;
6038 			if (tmp->rdev) {
6039 				/* Replaced device not technically faulty,
6040 				 * but we need to be sure it gets removed
6041 				 * and never re-added.
6042 				 */
6043 				set_bit(Faulty, &tmp->rdev->flags);
6044 				sysfs_notify_dirent_safe(
6045 					tmp->rdev->sysfs_state);
6046 			}
6047 			sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
6048 		} else if (tmp->rdev
6049 		    && tmp->rdev->recovery_offset == MaxSector
6050 		    && !test_bit(Faulty, &tmp->rdev->flags)
6051 		    && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
6052 			count++;
6053 			sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
6054 		}
6055 	}
6056 	spin_lock_irqsave(&conf->device_lock, flags);
6057 	mddev->degraded = calc_degraded(conf);
6058 	spin_unlock_irqrestore(&conf->device_lock, flags);
6059 	print_raid5_conf(conf);
6060 	return count;
6061 }
6062 
6063 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
6064 {
6065 	struct r5conf *conf = mddev->private;
6066 	int err = 0;
6067 	int number = rdev->raid_disk;
6068 	struct md_rdev **rdevp;
6069 	struct disk_info *p = conf->disks + number;
6070 
6071 	print_raid5_conf(conf);
6072 	if (rdev == p->rdev)
6073 		rdevp = &p->rdev;
6074 	else if (rdev == p->replacement)
6075 		rdevp = &p->replacement;
6076 	else
6077 		return 0;
6078 
6079 	if (number >= conf->raid_disks &&
6080 	    conf->reshape_progress == MaxSector)
6081 		clear_bit(In_sync, &rdev->flags);
6082 
6083 	if (test_bit(In_sync, &rdev->flags) ||
6084 	    atomic_read(&rdev->nr_pending)) {
6085 		err = -EBUSY;
6086 		goto abort;
6087 	}
6088 	/* Only remove non-faulty devices if recovery
6089 	 * isn't possible.
6090 	 */
6091 	if (!test_bit(Faulty, &rdev->flags) &&
6092 	    mddev->recovery_disabled != conf->recovery_disabled &&
6093 	    !has_failed(conf) &&
6094 	    (!p->replacement || p->replacement == rdev) &&
6095 	    number < conf->raid_disks) {
6096 		err = -EBUSY;
6097 		goto abort;
6098 	}
6099 	*rdevp = NULL;
6100 	synchronize_rcu();
6101 	if (atomic_read(&rdev->nr_pending)) {
6102 		/* lost the race, try later */
6103 		err = -EBUSY;
6104 		*rdevp = rdev;
6105 	} else if (p->replacement) {
6106 		/* We must have just cleared 'rdev' */
6107 		p->rdev = p->replacement;
6108 		clear_bit(Replacement, &p->replacement->flags);
6109 		smp_mb(); /* Make sure other CPUs may see both as identical
6110 			   * but will never see neither - if they are careful
6111 			   */
6112 		p->replacement = NULL;
6113 		clear_bit(WantReplacement, &rdev->flags);
6114 	} else
6115 		/* We might have just removed the Replacement as faulty-
6116 		 * clear the bit just in case
6117 		 */
6118 		clear_bit(WantReplacement, &rdev->flags);
6119 abort:
6120 
6121 	print_raid5_conf(conf);
6122 	return err;
6123 }
6124 
6125 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
6126 {
6127 	struct r5conf *conf = mddev->private;
6128 	int err = -EEXIST;
6129 	int disk;
6130 	struct disk_info *p;
6131 	int first = 0;
6132 	int last = conf->raid_disks - 1;
6133 
6134 	if (mddev->recovery_disabled == conf->recovery_disabled)
6135 		return -EBUSY;
6136 
6137 	if (rdev->saved_raid_disk < 0 && has_failed(conf))
6138 		/* no point adding a device */
6139 		return -EINVAL;
6140 
6141 	if (rdev->raid_disk >= 0)
6142 		first = last = rdev->raid_disk;
6143 
6144 	/*
6145 	 * find the disk ... but prefer rdev->saved_raid_disk
6146 	 * if possible.
6147 	 */
6148 	if (rdev->saved_raid_disk >= 0 &&
6149 	    rdev->saved_raid_disk >= first &&
6150 	    conf->disks[rdev->saved_raid_disk].rdev == NULL)
6151 		first = rdev->saved_raid_disk;
6152 
6153 	for (disk = first; disk <= last; disk++) {
6154 		p = conf->disks + disk;
6155 		if (p->rdev == NULL) {
6156 			clear_bit(In_sync, &rdev->flags);
6157 			rdev->raid_disk = disk;
6158 			err = 0;
6159 			if (rdev->saved_raid_disk != disk)
6160 				conf->fullsync = 1;
6161 			rcu_assign_pointer(p->rdev, rdev);
6162 			goto out;
6163 		}
6164 	}
6165 	for (disk = first; disk <= last; disk++) {
6166 		p = conf->disks + disk;
6167 		if (test_bit(WantReplacement, &p->rdev->flags) &&
6168 		    p->replacement == NULL) {
6169 			clear_bit(In_sync, &rdev->flags);
6170 			set_bit(Replacement, &rdev->flags);
6171 			rdev->raid_disk = disk;
6172 			err = 0;
6173 			conf->fullsync = 1;
6174 			rcu_assign_pointer(p->replacement, rdev);
6175 			break;
6176 		}
6177 	}
6178 out:
6179 	print_raid5_conf(conf);
6180 	return err;
6181 }
6182 
6183 static int raid5_resize(struct mddev *mddev, sector_t sectors)
6184 {
6185 	/* no resync is happening, and there is enough space
6186 	 * on all devices, so we can resize.
6187 	 * We need to make sure resync covers any new space.
6188 	 * If the array is shrinking we should possibly wait until
6189 	 * any io in the removed space completes, but it hardly seems
6190 	 * worth it.
6191 	 */
6192 	sector_t newsize;
6193 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6194 	newsize = raid5_size(mddev, sectors, mddev->raid_disks);
6195 	if (mddev->external_size &&
6196 	    mddev->array_sectors > newsize)
6197 		return -EINVAL;
6198 	if (mddev->bitmap) {
6199 		int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
6200 		if (ret)
6201 			return ret;
6202 	}
6203 	md_set_array_sectors(mddev, newsize);
6204 	set_capacity(mddev->gendisk, mddev->array_sectors);
6205 	revalidate_disk(mddev->gendisk);
6206 	if (sectors > mddev->dev_sectors &&
6207 	    mddev->recovery_cp > mddev->dev_sectors) {
6208 		mddev->recovery_cp = mddev->dev_sectors;
6209 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
6210 	}
6211 	mddev->dev_sectors = sectors;
6212 	mddev->resync_max_sectors = sectors;
6213 	return 0;
6214 }
6215 
6216 static int check_stripe_cache(struct mddev *mddev)
6217 {
6218 	/* Can only proceed if there are plenty of stripe_heads.
6219 	 * We need a minimum of one full stripe,, and for sensible progress
6220 	 * it is best to have about 4 times that.
6221 	 * If we require 4 times, then the default 256 4K stripe_heads will
6222 	 * allow for chunk sizes up to 256K, which is probably OK.
6223 	 * If the chunk size is greater, user-space should request more
6224 	 * stripe_heads first.
6225 	 */
6226 	struct r5conf *conf = mddev->private;
6227 	if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
6228 	    > conf->max_nr_stripes ||
6229 	    ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
6230 	    > conf->max_nr_stripes) {
6231 		printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
6232 		       mdname(mddev),
6233 		       ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
6234 			/ STRIPE_SIZE)*4);
6235 		return 0;
6236 	}
6237 	return 1;
6238 }
6239 
6240 static int check_reshape(struct mddev *mddev)
6241 {
6242 	struct r5conf *conf = mddev->private;
6243 
6244 	if (mddev->delta_disks == 0 &&
6245 	    mddev->new_layout == mddev->layout &&
6246 	    mddev->new_chunk_sectors == mddev->chunk_sectors)
6247 		return 0; /* nothing to do */
6248 	if (has_failed(conf))
6249 		return -EINVAL;
6250 	if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
6251 		/* We might be able to shrink, but the devices must
6252 		 * be made bigger first.
6253 		 * For raid6, 4 is the minimum size.
6254 		 * Otherwise 2 is the minimum
6255 		 */
6256 		int min = 2;
6257 		if (mddev->level == 6)
6258 			min = 4;
6259 		if (mddev->raid_disks + mddev->delta_disks < min)
6260 			return -EINVAL;
6261 	}
6262 
6263 	if (!check_stripe_cache(mddev))
6264 		return -ENOSPC;
6265 
6266 	return resize_stripes(conf, (conf->previous_raid_disks
6267 				     + mddev->delta_disks));
6268 }
6269 
6270 static int raid5_start_reshape(struct mddev *mddev)
6271 {
6272 	struct r5conf *conf = mddev->private;
6273 	struct md_rdev *rdev;
6274 	int spares = 0;
6275 	unsigned long flags;
6276 
6277 	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
6278 		return -EBUSY;
6279 
6280 	if (!check_stripe_cache(mddev))
6281 		return -ENOSPC;
6282 
6283 	if (has_failed(conf))
6284 		return -EINVAL;
6285 
6286 	rdev_for_each(rdev, mddev) {
6287 		if (!test_bit(In_sync, &rdev->flags)
6288 		    && !test_bit(Faulty, &rdev->flags))
6289 			spares++;
6290 	}
6291 
6292 	if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
6293 		/* Not enough devices even to make a degraded array
6294 		 * of that size
6295 		 */
6296 		return -EINVAL;
6297 
6298 	/* Refuse to reduce size of the array.  Any reductions in
6299 	 * array size must be through explicit setting of array_size
6300 	 * attribute.
6301 	 */
6302 	if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
6303 	    < mddev->array_sectors) {
6304 		printk(KERN_ERR "md/raid:%s: array size must be reduced "
6305 		       "before number of disks\n", mdname(mddev));
6306 		return -EINVAL;
6307 	}
6308 
6309 	atomic_set(&conf->reshape_stripes, 0);
6310 	spin_lock_irq(&conf->device_lock);
6311 	write_seqcount_begin(&conf->gen_lock);
6312 	conf->previous_raid_disks = conf->raid_disks;
6313 	conf->raid_disks += mddev->delta_disks;
6314 	conf->prev_chunk_sectors = conf->chunk_sectors;
6315 	conf->chunk_sectors = mddev->new_chunk_sectors;
6316 	conf->prev_algo = conf->algorithm;
6317 	conf->algorithm = mddev->new_layout;
6318 	conf->generation++;
6319 	/* Code that selects data_offset needs to see the generation update
6320 	 * if reshape_progress has been set - so a memory barrier needed.
6321 	 */
6322 	smp_mb();
6323 	if (mddev->reshape_backwards)
6324 		conf->reshape_progress = raid5_size(mddev, 0, 0);
6325 	else
6326 		conf->reshape_progress = 0;
6327 	conf->reshape_safe = conf->reshape_progress;
6328 	write_seqcount_end(&conf->gen_lock);
6329 	spin_unlock_irq(&conf->device_lock);
6330 
6331 	/* Now make sure any requests that proceeded on the assumption
6332 	 * the reshape wasn't running - like Discard or Read - have
6333 	 * completed.
6334 	 */
6335 	mddev_suspend(mddev);
6336 	mddev_resume(mddev);
6337 
6338 	/* Add some new drives, as many as will fit.
6339 	 * We know there are enough to make the newly sized array work.
6340 	 * Don't add devices if we are reducing the number of
6341 	 * devices in the array.  This is because it is not possible
6342 	 * to correctly record the "partially reconstructed" state of
6343 	 * such devices during the reshape and confusion could result.
6344 	 */
6345 	if (mddev->delta_disks >= 0) {
6346 		rdev_for_each(rdev, mddev)
6347 			if (rdev->raid_disk < 0 &&
6348 			    !test_bit(Faulty, &rdev->flags)) {
6349 				if (raid5_add_disk(mddev, rdev) == 0) {
6350 					if (rdev->raid_disk
6351 					    >= conf->previous_raid_disks)
6352 						set_bit(In_sync, &rdev->flags);
6353 					else
6354 						rdev->recovery_offset = 0;
6355 
6356 					if (sysfs_link_rdev(mddev, rdev))
6357 						/* Failure here is OK */;
6358 				}
6359 			} else if (rdev->raid_disk >= conf->previous_raid_disks
6360 				   && !test_bit(Faulty, &rdev->flags)) {
6361 				/* This is a spare that was manually added */
6362 				set_bit(In_sync, &rdev->flags);
6363 			}
6364 
6365 		/* When a reshape changes the number of devices,
6366 		 * ->degraded is measured against the larger of the
6367 		 * pre and post number of devices.
6368 		 */
6369 		spin_lock_irqsave(&conf->device_lock, flags);
6370 		mddev->degraded = calc_degraded(conf);
6371 		spin_unlock_irqrestore(&conf->device_lock, flags);
6372 	}
6373 	mddev->raid_disks = conf->raid_disks;
6374 	mddev->reshape_position = conf->reshape_progress;
6375 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
6376 
6377 	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6378 	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6379 	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6380 	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6381 	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6382 						"reshape");
6383 	if (!mddev->sync_thread) {
6384 		mddev->recovery = 0;
6385 		spin_lock_irq(&conf->device_lock);
6386 		mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6387 		rdev_for_each(rdev, mddev)
6388 			rdev->new_data_offset = rdev->data_offset;
6389 		smp_wmb();
6390 		conf->reshape_progress = MaxSector;
6391 		mddev->reshape_position = MaxSector;
6392 		spin_unlock_irq(&conf->device_lock);
6393 		return -EAGAIN;
6394 	}
6395 	conf->reshape_checkpoint = jiffies;
6396 	md_wakeup_thread(mddev->sync_thread);
6397 	md_new_event(mddev);
6398 	return 0;
6399 }
6400 
6401 /* This is called from the reshape thread and should make any
6402  * changes needed in 'conf'
6403  */
6404 static void end_reshape(struct r5conf *conf)
6405 {
6406 
6407 	if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6408 		struct md_rdev *rdev;
6409 
6410 		spin_lock_irq(&conf->device_lock);
6411 		conf->previous_raid_disks = conf->raid_disks;
6412 		rdev_for_each(rdev, conf->mddev)
6413 			rdev->data_offset = rdev->new_data_offset;
6414 		smp_wmb();
6415 		conf->reshape_progress = MaxSector;
6416 		spin_unlock_irq(&conf->device_lock);
6417 		wake_up(&conf->wait_for_overlap);
6418 
6419 		/* read-ahead size must cover two whole stripes, which is
6420 		 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6421 		 */
6422 		if (conf->mddev->queue) {
6423 			int data_disks = conf->raid_disks - conf->max_degraded;
6424 			int stripe = data_disks * ((conf->chunk_sectors << 9)
6425 						   / PAGE_SIZE);
6426 			if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6427 				conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6428 		}
6429 	}
6430 }
6431 
6432 /* This is called from the raid5d thread with mddev_lock held.
6433  * It makes config changes to the device.
6434  */
6435 static void raid5_finish_reshape(struct mddev *mddev)
6436 {
6437 	struct r5conf *conf = mddev->private;
6438 
6439 	if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6440 
6441 		if (mddev->delta_disks > 0) {
6442 			md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6443 			set_capacity(mddev->gendisk, mddev->array_sectors);
6444 			revalidate_disk(mddev->gendisk);
6445 		} else {
6446 			int d;
6447 			spin_lock_irq(&conf->device_lock);
6448 			mddev->degraded = calc_degraded(conf);
6449 			spin_unlock_irq(&conf->device_lock);
6450 			for (d = conf->raid_disks ;
6451 			     d < conf->raid_disks - mddev->delta_disks;
6452 			     d++) {
6453 				struct md_rdev *rdev = conf->disks[d].rdev;
6454 				if (rdev)
6455 					clear_bit(In_sync, &rdev->flags);
6456 				rdev = conf->disks[d].replacement;
6457 				if (rdev)
6458 					clear_bit(In_sync, &rdev->flags);
6459 			}
6460 		}
6461 		mddev->layout = conf->algorithm;
6462 		mddev->chunk_sectors = conf->chunk_sectors;
6463 		mddev->reshape_position = MaxSector;
6464 		mddev->delta_disks = 0;
6465 		mddev->reshape_backwards = 0;
6466 	}
6467 }
6468 
6469 static void raid5_quiesce(struct mddev *mddev, int state)
6470 {
6471 	struct r5conf *conf = mddev->private;
6472 
6473 	switch(state) {
6474 	case 2: /* resume for a suspend */
6475 		wake_up(&conf->wait_for_overlap);
6476 		break;
6477 
6478 	case 1: /* stop all writes */
6479 		spin_lock_irq(&conf->device_lock);
6480 		/* '2' tells resync/reshape to pause so that all
6481 		 * active stripes can drain
6482 		 */
6483 		conf->quiesce = 2;
6484 		wait_event_lock_irq(conf->wait_for_stripe,
6485 				    atomic_read(&conf->active_stripes) == 0 &&
6486 				    atomic_read(&conf->active_aligned_reads) == 0,
6487 				    conf->device_lock);
6488 		conf->quiesce = 1;
6489 		spin_unlock_irq(&conf->device_lock);
6490 		/* allow reshape to continue */
6491 		wake_up(&conf->wait_for_overlap);
6492 		break;
6493 
6494 	case 0: /* re-enable writes */
6495 		spin_lock_irq(&conf->device_lock);
6496 		conf->quiesce = 0;
6497 		wake_up(&conf->wait_for_stripe);
6498 		wake_up(&conf->wait_for_overlap);
6499 		spin_unlock_irq(&conf->device_lock);
6500 		break;
6501 	}
6502 }
6503 
6504 
6505 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6506 {
6507 	struct r0conf *raid0_conf = mddev->private;
6508 	sector_t sectors;
6509 
6510 	/* for raid0 takeover only one zone is supported */
6511 	if (raid0_conf->nr_strip_zones > 1) {
6512 		printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6513 		       mdname(mddev));
6514 		return ERR_PTR(-EINVAL);
6515 	}
6516 
6517 	sectors = raid0_conf->strip_zone[0].zone_end;
6518 	sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6519 	mddev->dev_sectors = sectors;
6520 	mddev->new_level = level;
6521 	mddev->new_layout = ALGORITHM_PARITY_N;
6522 	mddev->new_chunk_sectors = mddev->chunk_sectors;
6523 	mddev->raid_disks += 1;
6524 	mddev->delta_disks = 1;
6525 	/* make sure it will be not marked as dirty */
6526 	mddev->recovery_cp = MaxSector;
6527 
6528 	return setup_conf(mddev);
6529 }
6530 
6531 
6532 static void *raid5_takeover_raid1(struct mddev *mddev)
6533 {
6534 	int chunksect;
6535 
6536 	if (mddev->raid_disks != 2 ||
6537 	    mddev->degraded > 1)
6538 		return ERR_PTR(-EINVAL);
6539 
6540 	/* Should check if there are write-behind devices? */
6541 
6542 	chunksect = 64*2; /* 64K by default */
6543 
6544 	/* The array must be an exact multiple of chunksize */
6545 	while (chunksect && (mddev->array_sectors & (chunksect-1)))
6546 		chunksect >>= 1;
6547 
6548 	if ((chunksect<<9) < STRIPE_SIZE)
6549 		/* array size does not allow a suitable chunk size */
6550 		return ERR_PTR(-EINVAL);
6551 
6552 	mddev->new_level = 5;
6553 	mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6554 	mddev->new_chunk_sectors = chunksect;
6555 
6556 	return setup_conf(mddev);
6557 }
6558 
6559 static void *raid5_takeover_raid6(struct mddev *mddev)
6560 {
6561 	int new_layout;
6562 
6563 	switch (mddev->layout) {
6564 	case ALGORITHM_LEFT_ASYMMETRIC_6:
6565 		new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6566 		break;
6567 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
6568 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6569 		break;
6570 	case ALGORITHM_LEFT_SYMMETRIC_6:
6571 		new_layout = ALGORITHM_LEFT_SYMMETRIC;
6572 		break;
6573 	case ALGORITHM_RIGHT_SYMMETRIC_6:
6574 		new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6575 		break;
6576 	case ALGORITHM_PARITY_0_6:
6577 		new_layout = ALGORITHM_PARITY_0;
6578 		break;
6579 	case ALGORITHM_PARITY_N:
6580 		new_layout = ALGORITHM_PARITY_N;
6581 		break;
6582 	default:
6583 		return ERR_PTR(-EINVAL);
6584 	}
6585 	mddev->new_level = 5;
6586 	mddev->new_layout = new_layout;
6587 	mddev->delta_disks = -1;
6588 	mddev->raid_disks -= 1;
6589 	return setup_conf(mddev);
6590 }
6591 
6592 
6593 static int raid5_check_reshape(struct mddev *mddev)
6594 {
6595 	/* For a 2-drive array, the layout and chunk size can be changed
6596 	 * immediately as not restriping is needed.
6597 	 * For larger arrays we record the new value - after validation
6598 	 * to be used by a reshape pass.
6599 	 */
6600 	struct r5conf *conf = mddev->private;
6601 	int new_chunk = mddev->new_chunk_sectors;
6602 
6603 	if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6604 		return -EINVAL;
6605 	if (new_chunk > 0) {
6606 		if (!is_power_of_2(new_chunk))
6607 			return -EINVAL;
6608 		if (new_chunk < (PAGE_SIZE>>9))
6609 			return -EINVAL;
6610 		if (mddev->array_sectors & (new_chunk-1))
6611 			/* not factor of array size */
6612 			return -EINVAL;
6613 	}
6614 
6615 	/* They look valid */
6616 
6617 	if (mddev->raid_disks == 2) {
6618 		/* can make the change immediately */
6619 		if (mddev->new_layout >= 0) {
6620 			conf->algorithm = mddev->new_layout;
6621 			mddev->layout = mddev->new_layout;
6622 		}
6623 		if (new_chunk > 0) {
6624 			conf->chunk_sectors = new_chunk ;
6625 			mddev->chunk_sectors = new_chunk;
6626 		}
6627 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
6628 		md_wakeup_thread(mddev->thread);
6629 	}
6630 	return check_reshape(mddev);
6631 }
6632 
6633 static int raid6_check_reshape(struct mddev *mddev)
6634 {
6635 	int new_chunk = mddev->new_chunk_sectors;
6636 
6637 	if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6638 		return -EINVAL;
6639 	if (new_chunk > 0) {
6640 		if (!is_power_of_2(new_chunk))
6641 			return -EINVAL;
6642 		if (new_chunk < (PAGE_SIZE >> 9))
6643 			return -EINVAL;
6644 		if (mddev->array_sectors & (new_chunk-1))
6645 			/* not factor of array size */
6646 			return -EINVAL;
6647 	}
6648 
6649 	/* They look valid */
6650 	return check_reshape(mddev);
6651 }
6652 
6653 static void *raid5_takeover(struct mddev *mddev)
6654 {
6655 	/* raid5 can take over:
6656 	 *  raid0 - if there is only one strip zone - make it a raid4 layout
6657 	 *  raid1 - if there are two drives.  We need to know the chunk size
6658 	 *  raid4 - trivial - just use a raid4 layout.
6659 	 *  raid6 - Providing it is a *_6 layout
6660 	 */
6661 	if (mddev->level == 0)
6662 		return raid45_takeover_raid0(mddev, 5);
6663 	if (mddev->level == 1)
6664 		return raid5_takeover_raid1(mddev);
6665 	if (mddev->level == 4) {
6666 		mddev->new_layout = ALGORITHM_PARITY_N;
6667 		mddev->new_level = 5;
6668 		return setup_conf(mddev);
6669 	}
6670 	if (mddev->level == 6)
6671 		return raid5_takeover_raid6(mddev);
6672 
6673 	return ERR_PTR(-EINVAL);
6674 }
6675 
6676 static void *raid4_takeover(struct mddev *mddev)
6677 {
6678 	/* raid4 can take over:
6679 	 *  raid0 - if there is only one strip zone
6680 	 *  raid5 - if layout is right
6681 	 */
6682 	if (mddev->level == 0)
6683 		return raid45_takeover_raid0(mddev, 4);
6684 	if (mddev->level == 5 &&
6685 	    mddev->layout == ALGORITHM_PARITY_N) {
6686 		mddev->new_layout = 0;
6687 		mddev->new_level = 4;
6688 		return setup_conf(mddev);
6689 	}
6690 	return ERR_PTR(-EINVAL);
6691 }
6692 
6693 static struct md_personality raid5_personality;
6694 
6695 static void *raid6_takeover(struct mddev *mddev)
6696 {
6697 	/* Currently can only take over a raid5.  We map the
6698 	 * personality to an equivalent raid6 personality
6699 	 * with the Q block at the end.
6700 	 */
6701 	int new_layout;
6702 
6703 	if (mddev->pers != &raid5_personality)
6704 		return ERR_PTR(-EINVAL);
6705 	if (mddev->degraded > 1)
6706 		return ERR_PTR(-EINVAL);
6707 	if (mddev->raid_disks > 253)
6708 		return ERR_PTR(-EINVAL);
6709 	if (mddev->raid_disks < 3)
6710 		return ERR_PTR(-EINVAL);
6711 
6712 	switch (mddev->layout) {
6713 	case ALGORITHM_LEFT_ASYMMETRIC:
6714 		new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6715 		break;
6716 	case ALGORITHM_RIGHT_ASYMMETRIC:
6717 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6718 		break;
6719 	case ALGORITHM_LEFT_SYMMETRIC:
6720 		new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6721 		break;
6722 	case ALGORITHM_RIGHT_SYMMETRIC:
6723 		new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6724 		break;
6725 	case ALGORITHM_PARITY_0:
6726 		new_layout = ALGORITHM_PARITY_0_6;
6727 		break;
6728 	case ALGORITHM_PARITY_N:
6729 		new_layout = ALGORITHM_PARITY_N;
6730 		break;
6731 	default:
6732 		return ERR_PTR(-EINVAL);
6733 	}
6734 	mddev->new_level = 6;
6735 	mddev->new_layout = new_layout;
6736 	mddev->delta_disks = 1;
6737 	mddev->raid_disks += 1;
6738 	return setup_conf(mddev);
6739 }
6740 
6741 
6742 static struct md_personality raid6_personality =
6743 {
6744 	.name		= "raid6",
6745 	.level		= 6,
6746 	.owner		= THIS_MODULE,
6747 	.make_request	= make_request,
6748 	.run		= run,
6749 	.stop		= stop,
6750 	.status		= status,
6751 	.error_handler	= error,
6752 	.hot_add_disk	= raid5_add_disk,
6753 	.hot_remove_disk= raid5_remove_disk,
6754 	.spare_active	= raid5_spare_active,
6755 	.sync_request	= sync_request,
6756 	.resize		= raid5_resize,
6757 	.size		= raid5_size,
6758 	.check_reshape	= raid6_check_reshape,
6759 	.start_reshape  = raid5_start_reshape,
6760 	.finish_reshape = raid5_finish_reshape,
6761 	.quiesce	= raid5_quiesce,
6762 	.takeover	= raid6_takeover,
6763 };
6764 static struct md_personality raid5_personality =
6765 {
6766 	.name		= "raid5",
6767 	.level		= 5,
6768 	.owner		= THIS_MODULE,
6769 	.make_request	= make_request,
6770 	.run		= run,
6771 	.stop		= stop,
6772 	.status		= status,
6773 	.error_handler	= error,
6774 	.hot_add_disk	= raid5_add_disk,
6775 	.hot_remove_disk= raid5_remove_disk,
6776 	.spare_active	= raid5_spare_active,
6777 	.sync_request	= sync_request,
6778 	.resize		= raid5_resize,
6779 	.size		= raid5_size,
6780 	.check_reshape	= raid5_check_reshape,
6781 	.start_reshape  = raid5_start_reshape,
6782 	.finish_reshape = raid5_finish_reshape,
6783 	.quiesce	= raid5_quiesce,
6784 	.takeover	= raid5_takeover,
6785 };
6786 
6787 static struct md_personality raid4_personality =
6788 {
6789 	.name		= "raid4",
6790 	.level		= 4,
6791 	.owner		= THIS_MODULE,
6792 	.make_request	= make_request,
6793 	.run		= run,
6794 	.stop		= stop,
6795 	.status		= status,
6796 	.error_handler	= error,
6797 	.hot_add_disk	= raid5_add_disk,
6798 	.hot_remove_disk= raid5_remove_disk,
6799 	.spare_active	= raid5_spare_active,
6800 	.sync_request	= sync_request,
6801 	.resize		= raid5_resize,
6802 	.size		= raid5_size,
6803 	.check_reshape	= raid5_check_reshape,
6804 	.start_reshape  = raid5_start_reshape,
6805 	.finish_reshape = raid5_finish_reshape,
6806 	.quiesce	= raid5_quiesce,
6807 	.takeover	= raid4_takeover,
6808 };
6809 
6810 static int __init raid5_init(void)
6811 {
6812 	raid5_wq = alloc_workqueue("raid5wq",
6813 		WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
6814 	if (!raid5_wq)
6815 		return -ENOMEM;
6816 	register_md_personality(&raid6_personality);
6817 	register_md_personality(&raid5_personality);
6818 	register_md_personality(&raid4_personality);
6819 	return 0;
6820 }
6821 
6822 static void raid5_exit(void)
6823 {
6824 	unregister_md_personality(&raid6_personality);
6825 	unregister_md_personality(&raid5_personality);
6826 	unregister_md_personality(&raid4_personality);
6827 	destroy_workqueue(raid5_wq);
6828 }
6829 
6830 module_init(raid5_init);
6831 module_exit(raid5_exit);
6832 MODULE_LICENSE("GPL");
6833 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6834 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6835 MODULE_ALIAS("md-raid5");
6836 MODULE_ALIAS("md-raid4");
6837 MODULE_ALIAS("md-level-5");
6838 MODULE_ALIAS("md-level-4");
6839 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6840 MODULE_ALIAS("md-raid6");
6841 MODULE_ALIAS("md-level-6");
6842 
6843 /* This used to be two separate modules, they were: */
6844 MODULE_ALIAS("raid5");
6845 MODULE_ALIAS("raid6");
6846