xref: /linux/drivers/md/raid5.c (revision 498d319bb512992ef0784c278fa03679f2f5649d)
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 /* should hold conf->device_lock already */
297 static int release_stripe_list(struct r5conf *conf)
298 {
299 	struct stripe_head *sh;
300 	int count = 0;
301 	struct llist_node *head;
302 
303 	head = llist_del_all(&conf->released_stripes);
304 	head = llist_reverse_order(head);
305 	while (head) {
306 		sh = llist_entry(head, struct stripe_head, release_list);
307 		head = llist_next(head);
308 		/* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
309 		smp_mb();
310 		clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
311 		/*
312 		 * Don't worry the bit is set here, because if the bit is set
313 		 * again, the count is always > 1. This is true for
314 		 * STRIPE_ON_UNPLUG_LIST bit too.
315 		 */
316 		__release_stripe(conf, sh);
317 		count++;
318 	}
319 
320 	return count;
321 }
322 
323 static void release_stripe(struct stripe_head *sh)
324 {
325 	struct r5conf *conf = sh->raid_conf;
326 	unsigned long flags;
327 	bool wakeup;
328 
329 	if (test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
330 		goto slow_path;
331 	wakeup = llist_add(&sh->release_list, &conf->released_stripes);
332 	if (wakeup)
333 		md_wakeup_thread(conf->mddev->thread);
334 	return;
335 slow_path:
336 	local_irq_save(flags);
337 	/* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
338 	if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
339 		do_release_stripe(conf, sh);
340 		spin_unlock(&conf->device_lock);
341 	}
342 	local_irq_restore(flags);
343 }
344 
345 static inline void remove_hash(struct stripe_head *sh)
346 {
347 	pr_debug("remove_hash(), stripe %llu\n",
348 		(unsigned long long)sh->sector);
349 
350 	hlist_del_init(&sh->hash);
351 }
352 
353 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
354 {
355 	struct hlist_head *hp = stripe_hash(conf, sh->sector);
356 
357 	pr_debug("insert_hash(), stripe %llu\n",
358 		(unsigned long long)sh->sector);
359 
360 	hlist_add_head(&sh->hash, hp);
361 }
362 
363 
364 /* find an idle stripe, make sure it is unhashed, and return it. */
365 static struct stripe_head *get_free_stripe(struct r5conf *conf)
366 {
367 	struct stripe_head *sh = NULL;
368 	struct list_head *first;
369 
370 	if (list_empty(&conf->inactive_list))
371 		goto out;
372 	first = conf->inactive_list.next;
373 	sh = list_entry(first, struct stripe_head, lru);
374 	list_del_init(first);
375 	remove_hash(sh);
376 	atomic_inc(&conf->active_stripes);
377 out:
378 	return sh;
379 }
380 
381 static void shrink_buffers(struct stripe_head *sh)
382 {
383 	struct page *p;
384 	int i;
385 	int num = sh->raid_conf->pool_size;
386 
387 	for (i = 0; i < num ; i++) {
388 		p = sh->dev[i].page;
389 		if (!p)
390 			continue;
391 		sh->dev[i].page = NULL;
392 		put_page(p);
393 	}
394 }
395 
396 static int grow_buffers(struct stripe_head *sh)
397 {
398 	int i;
399 	int num = sh->raid_conf->pool_size;
400 
401 	for (i = 0; i < num; i++) {
402 		struct page *page;
403 
404 		if (!(page = alloc_page(GFP_KERNEL))) {
405 			return 1;
406 		}
407 		sh->dev[i].page = page;
408 	}
409 	return 0;
410 }
411 
412 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
413 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
414 			    struct stripe_head *sh);
415 
416 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
417 {
418 	struct r5conf *conf = sh->raid_conf;
419 	int i;
420 
421 	BUG_ON(atomic_read(&sh->count) != 0);
422 	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
423 	BUG_ON(stripe_operations_active(sh));
424 
425 	pr_debug("init_stripe called, stripe %llu\n",
426 		(unsigned long long)sh->sector);
427 
428 	remove_hash(sh);
429 
430 	sh->generation = conf->generation - previous;
431 	sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
432 	sh->sector = sector;
433 	stripe_set_idx(sector, conf, previous, sh);
434 	sh->state = 0;
435 
436 
437 	for (i = sh->disks; i--; ) {
438 		struct r5dev *dev = &sh->dev[i];
439 
440 		if (dev->toread || dev->read || dev->towrite || dev->written ||
441 		    test_bit(R5_LOCKED, &dev->flags)) {
442 			printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
443 			       (unsigned long long)sh->sector, i, dev->toread,
444 			       dev->read, dev->towrite, dev->written,
445 			       test_bit(R5_LOCKED, &dev->flags));
446 			WARN_ON(1);
447 		}
448 		dev->flags = 0;
449 		raid5_build_block(sh, i, previous);
450 	}
451 	insert_hash(conf, sh);
452 	sh->cpu = smp_processor_id();
453 }
454 
455 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
456 					 short generation)
457 {
458 	struct stripe_head *sh;
459 
460 	pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
461 	hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
462 		if (sh->sector == sector && sh->generation == generation)
463 			return sh;
464 	pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
465 	return NULL;
466 }
467 
468 /*
469  * Need to check if array has failed when deciding whether to:
470  *  - start an array
471  *  - remove non-faulty devices
472  *  - add a spare
473  *  - allow a reshape
474  * This determination is simple when no reshape is happening.
475  * However if there is a reshape, we need to carefully check
476  * both the before and after sections.
477  * This is because some failed devices may only affect one
478  * of the two sections, and some non-in_sync devices may
479  * be insync in the section most affected by failed devices.
480  */
481 static int calc_degraded(struct r5conf *conf)
482 {
483 	int degraded, degraded2;
484 	int i;
485 
486 	rcu_read_lock();
487 	degraded = 0;
488 	for (i = 0; i < conf->previous_raid_disks; i++) {
489 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
490 		if (rdev && test_bit(Faulty, &rdev->flags))
491 			rdev = rcu_dereference(conf->disks[i].replacement);
492 		if (!rdev || test_bit(Faulty, &rdev->flags))
493 			degraded++;
494 		else if (test_bit(In_sync, &rdev->flags))
495 			;
496 		else
497 			/* not in-sync or faulty.
498 			 * If the reshape increases the number of devices,
499 			 * this is being recovered by the reshape, so
500 			 * this 'previous' section is not in_sync.
501 			 * If the number of devices is being reduced however,
502 			 * the device can only be part of the array if
503 			 * we are reverting a reshape, so this section will
504 			 * be in-sync.
505 			 */
506 			if (conf->raid_disks >= conf->previous_raid_disks)
507 				degraded++;
508 	}
509 	rcu_read_unlock();
510 	if (conf->raid_disks == conf->previous_raid_disks)
511 		return degraded;
512 	rcu_read_lock();
513 	degraded2 = 0;
514 	for (i = 0; i < conf->raid_disks; i++) {
515 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
516 		if (rdev && test_bit(Faulty, &rdev->flags))
517 			rdev = rcu_dereference(conf->disks[i].replacement);
518 		if (!rdev || test_bit(Faulty, &rdev->flags))
519 			degraded2++;
520 		else if (test_bit(In_sync, &rdev->flags))
521 			;
522 		else
523 			/* not in-sync or faulty.
524 			 * If reshape increases the number of devices, this
525 			 * section has already been recovered, else it
526 			 * almost certainly hasn't.
527 			 */
528 			if (conf->raid_disks <= conf->previous_raid_disks)
529 				degraded2++;
530 	}
531 	rcu_read_unlock();
532 	if (degraded2 > degraded)
533 		return degraded2;
534 	return degraded;
535 }
536 
537 static int has_failed(struct r5conf *conf)
538 {
539 	int degraded;
540 
541 	if (conf->mddev->reshape_position == MaxSector)
542 		return conf->mddev->degraded > conf->max_degraded;
543 
544 	degraded = calc_degraded(conf);
545 	if (degraded > conf->max_degraded)
546 		return 1;
547 	return 0;
548 }
549 
550 static struct stripe_head *
551 get_active_stripe(struct r5conf *conf, sector_t sector,
552 		  int previous, int noblock, int noquiesce)
553 {
554 	struct stripe_head *sh;
555 
556 	pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
557 
558 	spin_lock_irq(&conf->device_lock);
559 
560 	do {
561 		wait_event_lock_irq(conf->wait_for_stripe,
562 				    conf->quiesce == 0 || noquiesce,
563 				    conf->device_lock);
564 		sh = __find_stripe(conf, sector, conf->generation - previous);
565 		if (!sh) {
566 			if (!conf->inactive_blocked)
567 				sh = get_free_stripe(conf);
568 			if (noblock && sh == NULL)
569 				break;
570 			if (!sh) {
571 				conf->inactive_blocked = 1;
572 				wait_event_lock_irq(conf->wait_for_stripe,
573 						    !list_empty(&conf->inactive_list) &&
574 						    (atomic_read(&conf->active_stripes)
575 						     < (conf->max_nr_stripes *3/4)
576 						     || !conf->inactive_blocked),
577 						    conf->device_lock);
578 				conf->inactive_blocked = 0;
579 			} else
580 				init_stripe(sh, sector, previous);
581 		} else {
582 			if (atomic_read(&sh->count)) {
583 				BUG_ON(!list_empty(&sh->lru)
584 				    && !test_bit(STRIPE_EXPANDING, &sh->state)
585 				    && !test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state)
586 				    && !test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
587 			} else {
588 				if (!test_bit(STRIPE_HANDLE, &sh->state))
589 					atomic_inc(&conf->active_stripes);
590 				if (list_empty(&sh->lru) &&
591 				    !test_bit(STRIPE_EXPANDING, &sh->state))
592 					BUG();
593 				list_del_init(&sh->lru);
594 				if (sh->group) {
595 					sh->group->stripes_cnt--;
596 					sh->group = NULL;
597 				}
598 			}
599 		}
600 	} while (sh == NULL);
601 
602 	if (sh)
603 		atomic_inc(&sh->count);
604 
605 	spin_unlock_irq(&conf->device_lock);
606 	return sh;
607 }
608 
609 /* Determine if 'data_offset' or 'new_data_offset' should be used
610  * in this stripe_head.
611  */
612 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
613 {
614 	sector_t progress = conf->reshape_progress;
615 	/* Need a memory barrier to make sure we see the value
616 	 * of conf->generation, or ->data_offset that was set before
617 	 * reshape_progress was updated.
618 	 */
619 	smp_rmb();
620 	if (progress == MaxSector)
621 		return 0;
622 	if (sh->generation == conf->generation - 1)
623 		return 0;
624 	/* We are in a reshape, and this is a new-generation stripe,
625 	 * so use new_data_offset.
626 	 */
627 	return 1;
628 }
629 
630 static void
631 raid5_end_read_request(struct bio *bi, int error);
632 static void
633 raid5_end_write_request(struct bio *bi, int error);
634 
635 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
636 {
637 	struct r5conf *conf = sh->raid_conf;
638 	int i, disks = sh->disks;
639 
640 	might_sleep();
641 
642 	for (i = disks; i--; ) {
643 		int rw;
644 		int replace_only = 0;
645 		struct bio *bi, *rbi;
646 		struct md_rdev *rdev, *rrdev = NULL;
647 		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
648 			if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
649 				rw = WRITE_FUA;
650 			else
651 				rw = WRITE;
652 			if (test_bit(R5_Discard, &sh->dev[i].flags))
653 				rw |= REQ_DISCARD;
654 		} else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
655 			rw = READ;
656 		else if (test_and_clear_bit(R5_WantReplace,
657 					    &sh->dev[i].flags)) {
658 			rw = WRITE;
659 			replace_only = 1;
660 		} else
661 			continue;
662 		if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
663 			rw |= REQ_SYNC;
664 
665 		bi = &sh->dev[i].req;
666 		rbi = &sh->dev[i].rreq; /* For writing to replacement */
667 
668 		rcu_read_lock();
669 		rrdev = rcu_dereference(conf->disks[i].replacement);
670 		smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
671 		rdev = rcu_dereference(conf->disks[i].rdev);
672 		if (!rdev) {
673 			rdev = rrdev;
674 			rrdev = NULL;
675 		}
676 		if (rw & WRITE) {
677 			if (replace_only)
678 				rdev = NULL;
679 			if (rdev == rrdev)
680 				/* We raced and saw duplicates */
681 				rrdev = NULL;
682 		} else {
683 			if (test_bit(R5_ReadRepl, &sh->dev[i].flags) && rrdev)
684 				rdev = rrdev;
685 			rrdev = NULL;
686 		}
687 
688 		if (rdev && test_bit(Faulty, &rdev->flags))
689 			rdev = NULL;
690 		if (rdev)
691 			atomic_inc(&rdev->nr_pending);
692 		if (rrdev && test_bit(Faulty, &rrdev->flags))
693 			rrdev = NULL;
694 		if (rrdev)
695 			atomic_inc(&rrdev->nr_pending);
696 		rcu_read_unlock();
697 
698 		/* We have already checked bad blocks for reads.  Now
699 		 * need to check for writes.  We never accept write errors
700 		 * on the replacement, so we don't to check rrdev.
701 		 */
702 		while ((rw & WRITE) && rdev &&
703 		       test_bit(WriteErrorSeen, &rdev->flags)) {
704 			sector_t first_bad;
705 			int bad_sectors;
706 			int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
707 					      &first_bad, &bad_sectors);
708 			if (!bad)
709 				break;
710 
711 			if (bad < 0) {
712 				set_bit(BlockedBadBlocks, &rdev->flags);
713 				if (!conf->mddev->external &&
714 				    conf->mddev->flags) {
715 					/* It is very unlikely, but we might
716 					 * still need to write out the
717 					 * bad block log - better give it
718 					 * a chance*/
719 					md_check_recovery(conf->mddev);
720 				}
721 				/*
722 				 * Because md_wait_for_blocked_rdev
723 				 * will dec nr_pending, we must
724 				 * increment it first.
725 				 */
726 				atomic_inc(&rdev->nr_pending);
727 				md_wait_for_blocked_rdev(rdev, conf->mddev);
728 			} else {
729 				/* Acknowledged bad block - skip the write */
730 				rdev_dec_pending(rdev, conf->mddev);
731 				rdev = NULL;
732 			}
733 		}
734 
735 		if (rdev) {
736 			if (s->syncing || s->expanding || s->expanded
737 			    || s->replacing)
738 				md_sync_acct(rdev->bdev, STRIPE_SECTORS);
739 
740 			set_bit(STRIPE_IO_STARTED, &sh->state);
741 
742 			bio_reset(bi);
743 			bi->bi_bdev = rdev->bdev;
744 			bi->bi_rw = rw;
745 			bi->bi_end_io = (rw & WRITE)
746 				? raid5_end_write_request
747 				: raid5_end_read_request;
748 			bi->bi_private = sh;
749 
750 			pr_debug("%s: for %llu schedule op %ld on disc %d\n",
751 				__func__, (unsigned long long)sh->sector,
752 				bi->bi_rw, i);
753 			atomic_inc(&sh->count);
754 			if (use_new_offset(conf, sh))
755 				bi->bi_sector = (sh->sector
756 						 + rdev->new_data_offset);
757 			else
758 				bi->bi_sector = (sh->sector
759 						 + rdev->data_offset);
760 			if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
761 				bi->bi_rw |= REQ_FLUSH;
762 
763 			bi->bi_vcnt = 1;
764 			bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
765 			bi->bi_io_vec[0].bv_offset = 0;
766 			bi->bi_size = STRIPE_SIZE;
767 			/*
768 			 * If this is discard request, set bi_vcnt 0. We don't
769 			 * want to confuse SCSI because SCSI will replace payload
770 			 */
771 			if (rw & REQ_DISCARD)
772 				bi->bi_vcnt = 0;
773 			if (rrdev)
774 				set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
775 
776 			if (conf->mddev->gendisk)
777 				trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
778 						      bi, disk_devt(conf->mddev->gendisk),
779 						      sh->dev[i].sector);
780 			generic_make_request(bi);
781 		}
782 		if (rrdev) {
783 			if (s->syncing || s->expanding || s->expanded
784 			    || s->replacing)
785 				md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
786 
787 			set_bit(STRIPE_IO_STARTED, &sh->state);
788 
789 			bio_reset(rbi);
790 			rbi->bi_bdev = rrdev->bdev;
791 			rbi->bi_rw = rw;
792 			BUG_ON(!(rw & WRITE));
793 			rbi->bi_end_io = raid5_end_write_request;
794 			rbi->bi_private = sh;
795 
796 			pr_debug("%s: for %llu schedule op %ld on "
797 				 "replacement disc %d\n",
798 				__func__, (unsigned long long)sh->sector,
799 				rbi->bi_rw, i);
800 			atomic_inc(&sh->count);
801 			if (use_new_offset(conf, sh))
802 				rbi->bi_sector = (sh->sector
803 						  + rrdev->new_data_offset);
804 			else
805 				rbi->bi_sector = (sh->sector
806 						  + rrdev->data_offset);
807 			rbi->bi_vcnt = 1;
808 			rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
809 			rbi->bi_io_vec[0].bv_offset = 0;
810 			rbi->bi_size = STRIPE_SIZE;
811 			/*
812 			 * If this is discard request, set bi_vcnt 0. We don't
813 			 * want to confuse SCSI because SCSI will replace payload
814 			 */
815 			if (rw & REQ_DISCARD)
816 				rbi->bi_vcnt = 0;
817 			if (conf->mddev->gendisk)
818 				trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
819 						      rbi, disk_devt(conf->mddev->gendisk),
820 						      sh->dev[i].sector);
821 			generic_make_request(rbi);
822 		}
823 		if (!rdev && !rrdev) {
824 			if (rw & WRITE)
825 				set_bit(STRIPE_DEGRADED, &sh->state);
826 			pr_debug("skip op %ld on disc %d for sector %llu\n",
827 				bi->bi_rw, i, (unsigned long long)sh->sector);
828 			clear_bit(R5_LOCKED, &sh->dev[i].flags);
829 			set_bit(STRIPE_HANDLE, &sh->state);
830 		}
831 	}
832 }
833 
834 static struct dma_async_tx_descriptor *
835 async_copy_data(int frombio, struct bio *bio, struct page *page,
836 	sector_t sector, struct dma_async_tx_descriptor *tx)
837 {
838 	struct bio_vec *bvl;
839 	struct page *bio_page;
840 	int i;
841 	int page_offset;
842 	struct async_submit_ctl submit;
843 	enum async_tx_flags flags = 0;
844 
845 	if (bio->bi_sector >= sector)
846 		page_offset = (signed)(bio->bi_sector - sector) * 512;
847 	else
848 		page_offset = (signed)(sector - bio->bi_sector) * -512;
849 
850 	if (frombio)
851 		flags |= ASYNC_TX_FENCE;
852 	init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
853 
854 	bio_for_each_segment(bvl, bio, i) {
855 		int len = bvl->bv_len;
856 		int clen;
857 		int b_offset = 0;
858 
859 		if (page_offset < 0) {
860 			b_offset = -page_offset;
861 			page_offset += b_offset;
862 			len -= b_offset;
863 		}
864 
865 		if (len > 0 && page_offset + len > STRIPE_SIZE)
866 			clen = STRIPE_SIZE - page_offset;
867 		else
868 			clen = len;
869 
870 		if (clen > 0) {
871 			b_offset += bvl->bv_offset;
872 			bio_page = bvl->bv_page;
873 			if (frombio)
874 				tx = async_memcpy(page, bio_page, page_offset,
875 						  b_offset, clen, &submit);
876 			else
877 				tx = async_memcpy(bio_page, page, b_offset,
878 						  page_offset, clen, &submit);
879 		}
880 		/* chain the operations */
881 		submit.depend_tx = tx;
882 
883 		if (clen < len) /* hit end of page */
884 			break;
885 		page_offset +=  len;
886 	}
887 
888 	return tx;
889 }
890 
891 static void ops_complete_biofill(void *stripe_head_ref)
892 {
893 	struct stripe_head *sh = stripe_head_ref;
894 	struct bio *return_bi = NULL;
895 	int i;
896 
897 	pr_debug("%s: stripe %llu\n", __func__,
898 		(unsigned long long)sh->sector);
899 
900 	/* clear completed biofills */
901 	for (i = sh->disks; i--; ) {
902 		struct r5dev *dev = &sh->dev[i];
903 
904 		/* acknowledge completion of a biofill operation */
905 		/* and check if we need to reply to a read request,
906 		 * new R5_Wantfill requests are held off until
907 		 * !STRIPE_BIOFILL_RUN
908 		 */
909 		if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
910 			struct bio *rbi, *rbi2;
911 
912 			BUG_ON(!dev->read);
913 			rbi = dev->read;
914 			dev->read = NULL;
915 			while (rbi && rbi->bi_sector <
916 				dev->sector + STRIPE_SECTORS) {
917 				rbi2 = r5_next_bio(rbi, dev->sector);
918 				if (!raid5_dec_bi_active_stripes(rbi)) {
919 					rbi->bi_next = return_bi;
920 					return_bi = rbi;
921 				}
922 				rbi = rbi2;
923 			}
924 		}
925 	}
926 	clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
927 
928 	return_io(return_bi);
929 
930 	set_bit(STRIPE_HANDLE, &sh->state);
931 	release_stripe(sh);
932 }
933 
934 static void ops_run_biofill(struct stripe_head *sh)
935 {
936 	struct dma_async_tx_descriptor *tx = NULL;
937 	struct async_submit_ctl submit;
938 	int i;
939 
940 	pr_debug("%s: stripe %llu\n", __func__,
941 		(unsigned long long)sh->sector);
942 
943 	for (i = sh->disks; i--; ) {
944 		struct r5dev *dev = &sh->dev[i];
945 		if (test_bit(R5_Wantfill, &dev->flags)) {
946 			struct bio *rbi;
947 			spin_lock_irq(&sh->stripe_lock);
948 			dev->read = rbi = dev->toread;
949 			dev->toread = NULL;
950 			spin_unlock_irq(&sh->stripe_lock);
951 			while (rbi && rbi->bi_sector <
952 				dev->sector + STRIPE_SECTORS) {
953 				tx = async_copy_data(0, rbi, dev->page,
954 					dev->sector, tx);
955 				rbi = r5_next_bio(rbi, dev->sector);
956 			}
957 		}
958 	}
959 
960 	atomic_inc(&sh->count);
961 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
962 	async_trigger_callback(&submit);
963 }
964 
965 static void mark_target_uptodate(struct stripe_head *sh, int target)
966 {
967 	struct r5dev *tgt;
968 
969 	if (target < 0)
970 		return;
971 
972 	tgt = &sh->dev[target];
973 	set_bit(R5_UPTODATE, &tgt->flags);
974 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
975 	clear_bit(R5_Wantcompute, &tgt->flags);
976 }
977 
978 static void ops_complete_compute(void *stripe_head_ref)
979 {
980 	struct stripe_head *sh = stripe_head_ref;
981 
982 	pr_debug("%s: stripe %llu\n", __func__,
983 		(unsigned long long)sh->sector);
984 
985 	/* mark the computed target(s) as uptodate */
986 	mark_target_uptodate(sh, sh->ops.target);
987 	mark_target_uptodate(sh, sh->ops.target2);
988 
989 	clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
990 	if (sh->check_state == check_state_compute_run)
991 		sh->check_state = check_state_compute_result;
992 	set_bit(STRIPE_HANDLE, &sh->state);
993 	release_stripe(sh);
994 }
995 
996 /* return a pointer to the address conversion region of the scribble buffer */
997 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
998 				 struct raid5_percpu *percpu)
999 {
1000 	return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
1001 }
1002 
1003 static struct dma_async_tx_descriptor *
1004 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1005 {
1006 	int disks = sh->disks;
1007 	struct page **xor_srcs = percpu->scribble;
1008 	int target = sh->ops.target;
1009 	struct r5dev *tgt = &sh->dev[target];
1010 	struct page *xor_dest = tgt->page;
1011 	int count = 0;
1012 	struct dma_async_tx_descriptor *tx;
1013 	struct async_submit_ctl submit;
1014 	int i;
1015 
1016 	pr_debug("%s: stripe %llu block: %d\n",
1017 		__func__, (unsigned long long)sh->sector, target);
1018 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1019 
1020 	for (i = disks; i--; )
1021 		if (i != target)
1022 			xor_srcs[count++] = sh->dev[i].page;
1023 
1024 	atomic_inc(&sh->count);
1025 
1026 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1027 			  ops_complete_compute, sh, to_addr_conv(sh, percpu));
1028 	if (unlikely(count == 1))
1029 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1030 	else
1031 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1032 
1033 	return tx;
1034 }
1035 
1036 /* set_syndrome_sources - populate source buffers for gen_syndrome
1037  * @srcs - (struct page *) array of size sh->disks
1038  * @sh - stripe_head to parse
1039  *
1040  * Populates srcs in proper layout order for the stripe and returns the
1041  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1042  * destination buffer is recorded in srcs[count] and the Q destination
1043  * is recorded in srcs[count+1]].
1044  */
1045 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
1046 {
1047 	int disks = sh->disks;
1048 	int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1049 	int d0_idx = raid6_d0(sh);
1050 	int count;
1051 	int i;
1052 
1053 	for (i = 0; i < disks; i++)
1054 		srcs[i] = NULL;
1055 
1056 	count = 0;
1057 	i = d0_idx;
1058 	do {
1059 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1060 
1061 		srcs[slot] = sh->dev[i].page;
1062 		i = raid6_next_disk(i, disks);
1063 	} while (i != d0_idx);
1064 
1065 	return syndrome_disks;
1066 }
1067 
1068 static struct dma_async_tx_descriptor *
1069 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1070 {
1071 	int disks = sh->disks;
1072 	struct page **blocks = percpu->scribble;
1073 	int target;
1074 	int qd_idx = sh->qd_idx;
1075 	struct dma_async_tx_descriptor *tx;
1076 	struct async_submit_ctl submit;
1077 	struct r5dev *tgt;
1078 	struct page *dest;
1079 	int i;
1080 	int count;
1081 
1082 	if (sh->ops.target < 0)
1083 		target = sh->ops.target2;
1084 	else if (sh->ops.target2 < 0)
1085 		target = sh->ops.target;
1086 	else
1087 		/* we should only have one valid target */
1088 		BUG();
1089 	BUG_ON(target < 0);
1090 	pr_debug("%s: stripe %llu block: %d\n",
1091 		__func__, (unsigned long long)sh->sector, target);
1092 
1093 	tgt = &sh->dev[target];
1094 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1095 	dest = tgt->page;
1096 
1097 	atomic_inc(&sh->count);
1098 
1099 	if (target == qd_idx) {
1100 		count = set_syndrome_sources(blocks, sh);
1101 		blocks[count] = NULL; /* regenerating p is not necessary */
1102 		BUG_ON(blocks[count+1] != dest); /* q should already be set */
1103 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1104 				  ops_complete_compute, sh,
1105 				  to_addr_conv(sh, percpu));
1106 		tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1107 	} else {
1108 		/* Compute any data- or p-drive using XOR */
1109 		count = 0;
1110 		for (i = disks; i-- ; ) {
1111 			if (i == target || i == qd_idx)
1112 				continue;
1113 			blocks[count++] = sh->dev[i].page;
1114 		}
1115 
1116 		init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1117 				  NULL, ops_complete_compute, sh,
1118 				  to_addr_conv(sh, percpu));
1119 		tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1120 	}
1121 
1122 	return tx;
1123 }
1124 
1125 static struct dma_async_tx_descriptor *
1126 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1127 {
1128 	int i, count, disks = sh->disks;
1129 	int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1130 	int d0_idx = raid6_d0(sh);
1131 	int faila = -1, failb = -1;
1132 	int target = sh->ops.target;
1133 	int target2 = sh->ops.target2;
1134 	struct r5dev *tgt = &sh->dev[target];
1135 	struct r5dev *tgt2 = &sh->dev[target2];
1136 	struct dma_async_tx_descriptor *tx;
1137 	struct page **blocks = percpu->scribble;
1138 	struct async_submit_ctl submit;
1139 
1140 	pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1141 		 __func__, (unsigned long long)sh->sector, target, target2);
1142 	BUG_ON(target < 0 || target2 < 0);
1143 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1144 	BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1145 
1146 	/* we need to open-code set_syndrome_sources to handle the
1147 	 * slot number conversion for 'faila' and 'failb'
1148 	 */
1149 	for (i = 0; i < disks ; i++)
1150 		blocks[i] = NULL;
1151 	count = 0;
1152 	i = d0_idx;
1153 	do {
1154 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1155 
1156 		blocks[slot] = sh->dev[i].page;
1157 
1158 		if (i == target)
1159 			faila = slot;
1160 		if (i == target2)
1161 			failb = slot;
1162 		i = raid6_next_disk(i, disks);
1163 	} while (i != d0_idx);
1164 
1165 	BUG_ON(faila == failb);
1166 	if (failb < faila)
1167 		swap(faila, failb);
1168 	pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1169 		 __func__, (unsigned long long)sh->sector, faila, failb);
1170 
1171 	atomic_inc(&sh->count);
1172 
1173 	if (failb == syndrome_disks+1) {
1174 		/* Q disk is one of the missing disks */
1175 		if (faila == syndrome_disks) {
1176 			/* Missing P+Q, just recompute */
1177 			init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1178 					  ops_complete_compute, sh,
1179 					  to_addr_conv(sh, percpu));
1180 			return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1181 						  STRIPE_SIZE, &submit);
1182 		} else {
1183 			struct page *dest;
1184 			int data_target;
1185 			int qd_idx = sh->qd_idx;
1186 
1187 			/* Missing D+Q: recompute D from P, then recompute Q */
1188 			if (target == qd_idx)
1189 				data_target = target2;
1190 			else
1191 				data_target = target;
1192 
1193 			count = 0;
1194 			for (i = disks; i-- ; ) {
1195 				if (i == data_target || i == qd_idx)
1196 					continue;
1197 				blocks[count++] = sh->dev[i].page;
1198 			}
1199 			dest = sh->dev[data_target].page;
1200 			init_async_submit(&submit,
1201 					  ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1202 					  NULL, NULL, NULL,
1203 					  to_addr_conv(sh, percpu));
1204 			tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1205 				       &submit);
1206 
1207 			count = set_syndrome_sources(blocks, sh);
1208 			init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1209 					  ops_complete_compute, sh,
1210 					  to_addr_conv(sh, percpu));
1211 			return async_gen_syndrome(blocks, 0, count+2,
1212 						  STRIPE_SIZE, &submit);
1213 		}
1214 	} else {
1215 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1216 				  ops_complete_compute, sh,
1217 				  to_addr_conv(sh, percpu));
1218 		if (failb == syndrome_disks) {
1219 			/* We're missing D+P. */
1220 			return async_raid6_datap_recov(syndrome_disks+2,
1221 						       STRIPE_SIZE, faila,
1222 						       blocks, &submit);
1223 		} else {
1224 			/* We're missing D+D. */
1225 			return async_raid6_2data_recov(syndrome_disks+2,
1226 						       STRIPE_SIZE, faila, failb,
1227 						       blocks, &submit);
1228 		}
1229 	}
1230 }
1231 
1232 
1233 static void ops_complete_prexor(void *stripe_head_ref)
1234 {
1235 	struct stripe_head *sh = stripe_head_ref;
1236 
1237 	pr_debug("%s: stripe %llu\n", __func__,
1238 		(unsigned long long)sh->sector);
1239 }
1240 
1241 static struct dma_async_tx_descriptor *
1242 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
1243 	       struct dma_async_tx_descriptor *tx)
1244 {
1245 	int disks = sh->disks;
1246 	struct page **xor_srcs = percpu->scribble;
1247 	int count = 0, pd_idx = sh->pd_idx, i;
1248 	struct async_submit_ctl submit;
1249 
1250 	/* existing parity data subtracted */
1251 	struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1252 
1253 	pr_debug("%s: stripe %llu\n", __func__,
1254 		(unsigned long long)sh->sector);
1255 
1256 	for (i = disks; i--; ) {
1257 		struct r5dev *dev = &sh->dev[i];
1258 		/* Only process blocks that are known to be uptodate */
1259 		if (test_bit(R5_Wantdrain, &dev->flags))
1260 			xor_srcs[count++] = dev->page;
1261 	}
1262 
1263 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1264 			  ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1265 	tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1266 
1267 	return tx;
1268 }
1269 
1270 static struct dma_async_tx_descriptor *
1271 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1272 {
1273 	int disks = sh->disks;
1274 	int i;
1275 
1276 	pr_debug("%s: stripe %llu\n", __func__,
1277 		(unsigned long long)sh->sector);
1278 
1279 	for (i = disks; i--; ) {
1280 		struct r5dev *dev = &sh->dev[i];
1281 		struct bio *chosen;
1282 
1283 		if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1284 			struct bio *wbi;
1285 
1286 			spin_lock_irq(&sh->stripe_lock);
1287 			chosen = dev->towrite;
1288 			dev->towrite = NULL;
1289 			BUG_ON(dev->written);
1290 			wbi = dev->written = chosen;
1291 			spin_unlock_irq(&sh->stripe_lock);
1292 
1293 			while (wbi && wbi->bi_sector <
1294 				dev->sector + STRIPE_SECTORS) {
1295 				if (wbi->bi_rw & REQ_FUA)
1296 					set_bit(R5_WantFUA, &dev->flags);
1297 				if (wbi->bi_rw & REQ_SYNC)
1298 					set_bit(R5_SyncIO, &dev->flags);
1299 				if (wbi->bi_rw & REQ_DISCARD)
1300 					set_bit(R5_Discard, &dev->flags);
1301 				else
1302 					tx = async_copy_data(1, wbi, dev->page,
1303 						dev->sector, tx);
1304 				wbi = r5_next_bio(wbi, dev->sector);
1305 			}
1306 		}
1307 	}
1308 
1309 	return tx;
1310 }
1311 
1312 static void ops_complete_reconstruct(void *stripe_head_ref)
1313 {
1314 	struct stripe_head *sh = stripe_head_ref;
1315 	int disks = sh->disks;
1316 	int pd_idx = sh->pd_idx;
1317 	int qd_idx = sh->qd_idx;
1318 	int i;
1319 	bool fua = false, sync = false, discard = false;
1320 
1321 	pr_debug("%s: stripe %llu\n", __func__,
1322 		(unsigned long long)sh->sector);
1323 
1324 	for (i = disks; i--; ) {
1325 		fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1326 		sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1327 		discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1328 	}
1329 
1330 	for (i = disks; i--; ) {
1331 		struct r5dev *dev = &sh->dev[i];
1332 
1333 		if (dev->written || i == pd_idx || i == qd_idx) {
1334 			if (!discard)
1335 				set_bit(R5_UPTODATE, &dev->flags);
1336 			if (fua)
1337 				set_bit(R5_WantFUA, &dev->flags);
1338 			if (sync)
1339 				set_bit(R5_SyncIO, &dev->flags);
1340 		}
1341 	}
1342 
1343 	if (sh->reconstruct_state == reconstruct_state_drain_run)
1344 		sh->reconstruct_state = reconstruct_state_drain_result;
1345 	else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1346 		sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1347 	else {
1348 		BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1349 		sh->reconstruct_state = reconstruct_state_result;
1350 	}
1351 
1352 	set_bit(STRIPE_HANDLE, &sh->state);
1353 	release_stripe(sh);
1354 }
1355 
1356 static void
1357 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1358 		     struct dma_async_tx_descriptor *tx)
1359 {
1360 	int disks = sh->disks;
1361 	struct page **xor_srcs = percpu->scribble;
1362 	struct async_submit_ctl submit;
1363 	int count = 0, pd_idx = sh->pd_idx, i;
1364 	struct page *xor_dest;
1365 	int prexor = 0;
1366 	unsigned long flags;
1367 
1368 	pr_debug("%s: stripe %llu\n", __func__,
1369 		(unsigned long long)sh->sector);
1370 
1371 	for (i = 0; i < sh->disks; i++) {
1372 		if (pd_idx == i)
1373 			continue;
1374 		if (!test_bit(R5_Discard, &sh->dev[i].flags))
1375 			break;
1376 	}
1377 	if (i >= sh->disks) {
1378 		atomic_inc(&sh->count);
1379 		set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1380 		ops_complete_reconstruct(sh);
1381 		return;
1382 	}
1383 	/* check if prexor is active which means only process blocks
1384 	 * that are part of a read-modify-write (written)
1385 	 */
1386 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1387 		prexor = 1;
1388 		xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1389 		for (i = disks; i--; ) {
1390 			struct r5dev *dev = &sh->dev[i];
1391 			if (dev->written)
1392 				xor_srcs[count++] = dev->page;
1393 		}
1394 	} else {
1395 		xor_dest = sh->dev[pd_idx].page;
1396 		for (i = disks; i--; ) {
1397 			struct r5dev *dev = &sh->dev[i];
1398 			if (i != pd_idx)
1399 				xor_srcs[count++] = dev->page;
1400 		}
1401 	}
1402 
1403 	/* 1/ if we prexor'd then the dest is reused as a source
1404 	 * 2/ if we did not prexor then we are redoing the parity
1405 	 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1406 	 * for the synchronous xor case
1407 	 */
1408 	flags = ASYNC_TX_ACK |
1409 		(prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1410 
1411 	atomic_inc(&sh->count);
1412 
1413 	init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1414 			  to_addr_conv(sh, percpu));
1415 	if (unlikely(count == 1))
1416 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1417 	else
1418 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1419 }
1420 
1421 static void
1422 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1423 		     struct dma_async_tx_descriptor *tx)
1424 {
1425 	struct async_submit_ctl submit;
1426 	struct page **blocks = percpu->scribble;
1427 	int count, i;
1428 
1429 	pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1430 
1431 	for (i = 0; i < sh->disks; i++) {
1432 		if (sh->pd_idx == i || sh->qd_idx == i)
1433 			continue;
1434 		if (!test_bit(R5_Discard, &sh->dev[i].flags))
1435 			break;
1436 	}
1437 	if (i >= sh->disks) {
1438 		atomic_inc(&sh->count);
1439 		set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1440 		set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1441 		ops_complete_reconstruct(sh);
1442 		return;
1443 	}
1444 
1445 	count = set_syndrome_sources(blocks, sh);
1446 
1447 	atomic_inc(&sh->count);
1448 
1449 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1450 			  sh, to_addr_conv(sh, percpu));
1451 	async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1452 }
1453 
1454 static void ops_complete_check(void *stripe_head_ref)
1455 {
1456 	struct stripe_head *sh = stripe_head_ref;
1457 
1458 	pr_debug("%s: stripe %llu\n", __func__,
1459 		(unsigned long long)sh->sector);
1460 
1461 	sh->check_state = check_state_check_result;
1462 	set_bit(STRIPE_HANDLE, &sh->state);
1463 	release_stripe(sh);
1464 }
1465 
1466 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1467 {
1468 	int disks = sh->disks;
1469 	int pd_idx = sh->pd_idx;
1470 	int qd_idx = sh->qd_idx;
1471 	struct page *xor_dest;
1472 	struct page **xor_srcs = percpu->scribble;
1473 	struct dma_async_tx_descriptor *tx;
1474 	struct async_submit_ctl submit;
1475 	int count;
1476 	int i;
1477 
1478 	pr_debug("%s: stripe %llu\n", __func__,
1479 		(unsigned long long)sh->sector);
1480 
1481 	count = 0;
1482 	xor_dest = sh->dev[pd_idx].page;
1483 	xor_srcs[count++] = xor_dest;
1484 	for (i = disks; i--; ) {
1485 		if (i == pd_idx || i == qd_idx)
1486 			continue;
1487 		xor_srcs[count++] = sh->dev[i].page;
1488 	}
1489 
1490 	init_async_submit(&submit, 0, NULL, NULL, NULL,
1491 			  to_addr_conv(sh, percpu));
1492 	tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1493 			   &sh->ops.zero_sum_result, &submit);
1494 
1495 	atomic_inc(&sh->count);
1496 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1497 	tx = async_trigger_callback(&submit);
1498 }
1499 
1500 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1501 {
1502 	struct page **srcs = percpu->scribble;
1503 	struct async_submit_ctl submit;
1504 	int count;
1505 
1506 	pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1507 		(unsigned long long)sh->sector, checkp);
1508 
1509 	count = set_syndrome_sources(srcs, sh);
1510 	if (!checkp)
1511 		srcs[count] = NULL;
1512 
1513 	atomic_inc(&sh->count);
1514 	init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1515 			  sh, to_addr_conv(sh, percpu));
1516 	async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1517 			   &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1518 }
1519 
1520 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1521 {
1522 	int overlap_clear = 0, i, disks = sh->disks;
1523 	struct dma_async_tx_descriptor *tx = NULL;
1524 	struct r5conf *conf = sh->raid_conf;
1525 	int level = conf->level;
1526 	struct raid5_percpu *percpu;
1527 	unsigned long cpu;
1528 
1529 	cpu = get_cpu();
1530 	percpu = per_cpu_ptr(conf->percpu, cpu);
1531 	if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1532 		ops_run_biofill(sh);
1533 		overlap_clear++;
1534 	}
1535 
1536 	if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1537 		if (level < 6)
1538 			tx = ops_run_compute5(sh, percpu);
1539 		else {
1540 			if (sh->ops.target2 < 0 || sh->ops.target < 0)
1541 				tx = ops_run_compute6_1(sh, percpu);
1542 			else
1543 				tx = ops_run_compute6_2(sh, percpu);
1544 		}
1545 		/* terminate the chain if reconstruct is not set to be run */
1546 		if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1547 			async_tx_ack(tx);
1548 	}
1549 
1550 	if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1551 		tx = ops_run_prexor(sh, percpu, tx);
1552 
1553 	if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1554 		tx = ops_run_biodrain(sh, tx);
1555 		overlap_clear++;
1556 	}
1557 
1558 	if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1559 		if (level < 6)
1560 			ops_run_reconstruct5(sh, percpu, tx);
1561 		else
1562 			ops_run_reconstruct6(sh, percpu, tx);
1563 	}
1564 
1565 	if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1566 		if (sh->check_state == check_state_run)
1567 			ops_run_check_p(sh, percpu);
1568 		else if (sh->check_state == check_state_run_q)
1569 			ops_run_check_pq(sh, percpu, 0);
1570 		else if (sh->check_state == check_state_run_pq)
1571 			ops_run_check_pq(sh, percpu, 1);
1572 		else
1573 			BUG();
1574 	}
1575 
1576 	if (overlap_clear)
1577 		for (i = disks; i--; ) {
1578 			struct r5dev *dev = &sh->dev[i];
1579 			if (test_and_clear_bit(R5_Overlap, &dev->flags))
1580 				wake_up(&sh->raid_conf->wait_for_overlap);
1581 		}
1582 	put_cpu();
1583 }
1584 
1585 static int grow_one_stripe(struct r5conf *conf)
1586 {
1587 	struct stripe_head *sh;
1588 	sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1589 	if (!sh)
1590 		return 0;
1591 
1592 	sh->raid_conf = conf;
1593 
1594 	spin_lock_init(&sh->stripe_lock);
1595 
1596 	if (grow_buffers(sh)) {
1597 		shrink_buffers(sh);
1598 		kmem_cache_free(conf->slab_cache, sh);
1599 		return 0;
1600 	}
1601 	/* we just created an active stripe so... */
1602 	atomic_set(&sh->count, 1);
1603 	atomic_inc(&conf->active_stripes);
1604 	INIT_LIST_HEAD(&sh->lru);
1605 	release_stripe(sh);
1606 	return 1;
1607 }
1608 
1609 static int grow_stripes(struct r5conf *conf, int num)
1610 {
1611 	struct kmem_cache *sc;
1612 	int devs = max(conf->raid_disks, conf->previous_raid_disks);
1613 
1614 	if (conf->mddev->gendisk)
1615 		sprintf(conf->cache_name[0],
1616 			"raid%d-%s", conf->level, mdname(conf->mddev));
1617 	else
1618 		sprintf(conf->cache_name[0],
1619 			"raid%d-%p", conf->level, conf->mddev);
1620 	sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1621 
1622 	conf->active_name = 0;
1623 	sc = kmem_cache_create(conf->cache_name[conf->active_name],
1624 			       sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1625 			       0, 0, NULL);
1626 	if (!sc)
1627 		return 1;
1628 	conf->slab_cache = sc;
1629 	conf->pool_size = devs;
1630 	while (num--)
1631 		if (!grow_one_stripe(conf))
1632 			return 1;
1633 	return 0;
1634 }
1635 
1636 /**
1637  * scribble_len - return the required size of the scribble region
1638  * @num - total number of disks in the array
1639  *
1640  * The size must be enough to contain:
1641  * 1/ a struct page pointer for each device in the array +2
1642  * 2/ room to convert each entry in (1) to its corresponding dma
1643  *    (dma_map_page()) or page (page_address()) address.
1644  *
1645  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1646  * calculate over all devices (not just the data blocks), using zeros in place
1647  * of the P and Q blocks.
1648  */
1649 static size_t scribble_len(int num)
1650 {
1651 	size_t len;
1652 
1653 	len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1654 
1655 	return len;
1656 }
1657 
1658 static int resize_stripes(struct r5conf *conf, int newsize)
1659 {
1660 	/* Make all the stripes able to hold 'newsize' devices.
1661 	 * New slots in each stripe get 'page' set to a new page.
1662 	 *
1663 	 * This happens in stages:
1664 	 * 1/ create a new kmem_cache and allocate the required number of
1665 	 *    stripe_heads.
1666 	 * 2/ gather all the old stripe_heads and transfer the pages across
1667 	 *    to the new stripe_heads.  This will have the side effect of
1668 	 *    freezing the array as once all stripe_heads have been collected,
1669 	 *    no IO will be possible.  Old stripe heads are freed once their
1670 	 *    pages have been transferred over, and the old kmem_cache is
1671 	 *    freed when all stripes are done.
1672 	 * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1673 	 *    we simple return a failre status - no need to clean anything up.
1674 	 * 4/ allocate new pages for the new slots in the new stripe_heads.
1675 	 *    If this fails, we don't bother trying the shrink the
1676 	 *    stripe_heads down again, we just leave them as they are.
1677 	 *    As each stripe_head is processed the new one is released into
1678 	 *    active service.
1679 	 *
1680 	 * Once step2 is started, we cannot afford to wait for a write,
1681 	 * so we use GFP_NOIO allocations.
1682 	 */
1683 	struct stripe_head *osh, *nsh;
1684 	LIST_HEAD(newstripes);
1685 	struct disk_info *ndisks;
1686 	unsigned long cpu;
1687 	int err;
1688 	struct kmem_cache *sc;
1689 	int i;
1690 
1691 	if (newsize <= conf->pool_size)
1692 		return 0; /* never bother to shrink */
1693 
1694 	err = md_allow_write(conf->mddev);
1695 	if (err)
1696 		return err;
1697 
1698 	/* Step 1 */
1699 	sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1700 			       sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1701 			       0, 0, NULL);
1702 	if (!sc)
1703 		return -ENOMEM;
1704 
1705 	for (i = conf->max_nr_stripes; i; i--) {
1706 		nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1707 		if (!nsh)
1708 			break;
1709 
1710 		nsh->raid_conf = conf;
1711 		spin_lock_init(&nsh->stripe_lock);
1712 
1713 		list_add(&nsh->lru, &newstripes);
1714 	}
1715 	if (i) {
1716 		/* didn't get enough, give up */
1717 		while (!list_empty(&newstripes)) {
1718 			nsh = list_entry(newstripes.next, struct stripe_head, lru);
1719 			list_del(&nsh->lru);
1720 			kmem_cache_free(sc, nsh);
1721 		}
1722 		kmem_cache_destroy(sc);
1723 		return -ENOMEM;
1724 	}
1725 	/* Step 2 - Must use GFP_NOIO now.
1726 	 * OK, we have enough stripes, start collecting inactive
1727 	 * stripes and copying them over
1728 	 */
1729 	list_for_each_entry(nsh, &newstripes, lru) {
1730 		spin_lock_irq(&conf->device_lock);
1731 		wait_event_lock_irq(conf->wait_for_stripe,
1732 				    !list_empty(&conf->inactive_list),
1733 				    conf->device_lock);
1734 		osh = get_free_stripe(conf);
1735 		spin_unlock_irq(&conf->device_lock);
1736 		atomic_set(&nsh->count, 1);
1737 		for(i=0; i<conf->pool_size; i++)
1738 			nsh->dev[i].page = osh->dev[i].page;
1739 		for( ; i<newsize; i++)
1740 			nsh->dev[i].page = NULL;
1741 		kmem_cache_free(conf->slab_cache, osh);
1742 	}
1743 	kmem_cache_destroy(conf->slab_cache);
1744 
1745 	/* Step 3.
1746 	 * At this point, we are holding all the stripes so the array
1747 	 * is completely stalled, so now is a good time to resize
1748 	 * conf->disks and the scribble region
1749 	 */
1750 	ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1751 	if (ndisks) {
1752 		for (i=0; i<conf->raid_disks; i++)
1753 			ndisks[i] = conf->disks[i];
1754 		kfree(conf->disks);
1755 		conf->disks = ndisks;
1756 	} else
1757 		err = -ENOMEM;
1758 
1759 	get_online_cpus();
1760 	conf->scribble_len = scribble_len(newsize);
1761 	for_each_present_cpu(cpu) {
1762 		struct raid5_percpu *percpu;
1763 		void *scribble;
1764 
1765 		percpu = per_cpu_ptr(conf->percpu, cpu);
1766 		scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1767 
1768 		if (scribble) {
1769 			kfree(percpu->scribble);
1770 			percpu->scribble = scribble;
1771 		} else {
1772 			err = -ENOMEM;
1773 			break;
1774 		}
1775 	}
1776 	put_online_cpus();
1777 
1778 	/* Step 4, return new stripes to service */
1779 	while(!list_empty(&newstripes)) {
1780 		nsh = list_entry(newstripes.next, struct stripe_head, lru);
1781 		list_del_init(&nsh->lru);
1782 
1783 		for (i=conf->raid_disks; i < newsize; i++)
1784 			if (nsh->dev[i].page == NULL) {
1785 				struct page *p = alloc_page(GFP_NOIO);
1786 				nsh->dev[i].page = p;
1787 				if (!p)
1788 					err = -ENOMEM;
1789 			}
1790 		release_stripe(nsh);
1791 	}
1792 	/* critical section pass, GFP_NOIO no longer needed */
1793 
1794 	conf->slab_cache = sc;
1795 	conf->active_name = 1-conf->active_name;
1796 	conf->pool_size = newsize;
1797 	return err;
1798 }
1799 
1800 static int drop_one_stripe(struct r5conf *conf)
1801 {
1802 	struct stripe_head *sh;
1803 
1804 	spin_lock_irq(&conf->device_lock);
1805 	sh = get_free_stripe(conf);
1806 	spin_unlock_irq(&conf->device_lock);
1807 	if (!sh)
1808 		return 0;
1809 	BUG_ON(atomic_read(&sh->count));
1810 	shrink_buffers(sh);
1811 	kmem_cache_free(conf->slab_cache, sh);
1812 	atomic_dec(&conf->active_stripes);
1813 	return 1;
1814 }
1815 
1816 static void shrink_stripes(struct r5conf *conf)
1817 {
1818 	while (drop_one_stripe(conf))
1819 		;
1820 
1821 	if (conf->slab_cache)
1822 		kmem_cache_destroy(conf->slab_cache);
1823 	conf->slab_cache = NULL;
1824 }
1825 
1826 static void raid5_end_read_request(struct bio * bi, int error)
1827 {
1828 	struct stripe_head *sh = bi->bi_private;
1829 	struct r5conf *conf = sh->raid_conf;
1830 	int disks = sh->disks, i;
1831 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1832 	char b[BDEVNAME_SIZE];
1833 	struct md_rdev *rdev = NULL;
1834 	sector_t s;
1835 
1836 	for (i=0 ; i<disks; i++)
1837 		if (bi == &sh->dev[i].req)
1838 			break;
1839 
1840 	pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1841 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
1842 		uptodate);
1843 	if (i == disks) {
1844 		BUG();
1845 		return;
1846 	}
1847 	if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1848 		/* If replacement finished while this request was outstanding,
1849 		 * 'replacement' might be NULL already.
1850 		 * In that case it moved down to 'rdev'.
1851 		 * rdev is not removed until all requests are finished.
1852 		 */
1853 		rdev = conf->disks[i].replacement;
1854 	if (!rdev)
1855 		rdev = conf->disks[i].rdev;
1856 
1857 	if (use_new_offset(conf, sh))
1858 		s = sh->sector + rdev->new_data_offset;
1859 	else
1860 		s = sh->sector + rdev->data_offset;
1861 	if (uptodate) {
1862 		set_bit(R5_UPTODATE, &sh->dev[i].flags);
1863 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1864 			/* Note that this cannot happen on a
1865 			 * replacement device.  We just fail those on
1866 			 * any error
1867 			 */
1868 			printk_ratelimited(
1869 				KERN_INFO
1870 				"md/raid:%s: read error corrected"
1871 				" (%lu sectors at %llu on %s)\n",
1872 				mdname(conf->mddev), STRIPE_SECTORS,
1873 				(unsigned long long)s,
1874 				bdevname(rdev->bdev, b));
1875 			atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1876 			clear_bit(R5_ReadError, &sh->dev[i].flags);
1877 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
1878 		} else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
1879 			clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1880 
1881 		if (atomic_read(&rdev->read_errors))
1882 			atomic_set(&rdev->read_errors, 0);
1883 	} else {
1884 		const char *bdn = bdevname(rdev->bdev, b);
1885 		int retry = 0;
1886 		int set_bad = 0;
1887 
1888 		clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1889 		atomic_inc(&rdev->read_errors);
1890 		if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
1891 			printk_ratelimited(
1892 				KERN_WARNING
1893 				"md/raid:%s: read error on replacement device "
1894 				"(sector %llu on %s).\n",
1895 				mdname(conf->mddev),
1896 				(unsigned long long)s,
1897 				bdn);
1898 		else if (conf->mddev->degraded >= conf->max_degraded) {
1899 			set_bad = 1;
1900 			printk_ratelimited(
1901 				KERN_WARNING
1902 				"md/raid:%s: read error not correctable "
1903 				"(sector %llu on %s).\n",
1904 				mdname(conf->mddev),
1905 				(unsigned long long)s,
1906 				bdn);
1907 		} else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
1908 			/* Oh, no!!! */
1909 			set_bad = 1;
1910 			printk_ratelimited(
1911 				KERN_WARNING
1912 				"md/raid:%s: read error NOT corrected!! "
1913 				"(sector %llu on %s).\n",
1914 				mdname(conf->mddev),
1915 				(unsigned long long)s,
1916 				bdn);
1917 		} else if (atomic_read(&rdev->read_errors)
1918 			 > conf->max_nr_stripes)
1919 			printk(KERN_WARNING
1920 			       "md/raid:%s: Too many read errors, failing device %s.\n",
1921 			       mdname(conf->mddev), bdn);
1922 		else
1923 			retry = 1;
1924 		if (retry)
1925 			if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
1926 				set_bit(R5_ReadError, &sh->dev[i].flags);
1927 				clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1928 			} else
1929 				set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
1930 		else {
1931 			clear_bit(R5_ReadError, &sh->dev[i].flags);
1932 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
1933 			if (!(set_bad
1934 			      && test_bit(In_sync, &rdev->flags)
1935 			      && rdev_set_badblocks(
1936 				      rdev, sh->sector, STRIPE_SECTORS, 0)))
1937 				md_error(conf->mddev, rdev);
1938 		}
1939 	}
1940 	rdev_dec_pending(rdev, conf->mddev);
1941 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
1942 	set_bit(STRIPE_HANDLE, &sh->state);
1943 	release_stripe(sh);
1944 }
1945 
1946 static void raid5_end_write_request(struct bio *bi, int error)
1947 {
1948 	struct stripe_head *sh = bi->bi_private;
1949 	struct r5conf *conf = sh->raid_conf;
1950 	int disks = sh->disks, i;
1951 	struct md_rdev *uninitialized_var(rdev);
1952 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1953 	sector_t first_bad;
1954 	int bad_sectors;
1955 	int replacement = 0;
1956 
1957 	for (i = 0 ; i < disks; i++) {
1958 		if (bi == &sh->dev[i].req) {
1959 			rdev = conf->disks[i].rdev;
1960 			break;
1961 		}
1962 		if (bi == &sh->dev[i].rreq) {
1963 			rdev = conf->disks[i].replacement;
1964 			if (rdev)
1965 				replacement = 1;
1966 			else
1967 				/* rdev was removed and 'replacement'
1968 				 * replaced it.  rdev is not removed
1969 				 * until all requests are finished.
1970 				 */
1971 				rdev = conf->disks[i].rdev;
1972 			break;
1973 		}
1974 	}
1975 	pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1976 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
1977 		uptodate);
1978 	if (i == disks) {
1979 		BUG();
1980 		return;
1981 	}
1982 
1983 	if (replacement) {
1984 		if (!uptodate)
1985 			md_error(conf->mddev, rdev);
1986 		else if (is_badblock(rdev, sh->sector,
1987 				     STRIPE_SECTORS,
1988 				     &first_bad, &bad_sectors))
1989 			set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
1990 	} else {
1991 		if (!uptodate) {
1992 			set_bit(WriteErrorSeen, &rdev->flags);
1993 			set_bit(R5_WriteError, &sh->dev[i].flags);
1994 			if (!test_and_set_bit(WantReplacement, &rdev->flags))
1995 				set_bit(MD_RECOVERY_NEEDED,
1996 					&rdev->mddev->recovery);
1997 		} else if (is_badblock(rdev, sh->sector,
1998 				       STRIPE_SECTORS,
1999 				       &first_bad, &bad_sectors)) {
2000 			set_bit(R5_MadeGood, &sh->dev[i].flags);
2001 			if (test_bit(R5_ReadError, &sh->dev[i].flags))
2002 				/* That was a successful write so make
2003 				 * sure it looks like we already did
2004 				 * a re-write.
2005 				 */
2006 				set_bit(R5_ReWrite, &sh->dev[i].flags);
2007 		}
2008 	}
2009 	rdev_dec_pending(rdev, conf->mddev);
2010 
2011 	if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2012 		clear_bit(R5_LOCKED, &sh->dev[i].flags);
2013 	set_bit(STRIPE_HANDLE, &sh->state);
2014 	release_stripe(sh);
2015 }
2016 
2017 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2018 
2019 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2020 {
2021 	struct r5dev *dev = &sh->dev[i];
2022 
2023 	bio_init(&dev->req);
2024 	dev->req.bi_io_vec = &dev->vec;
2025 	dev->req.bi_vcnt++;
2026 	dev->req.bi_max_vecs++;
2027 	dev->req.bi_private = sh;
2028 	dev->vec.bv_page = dev->page;
2029 
2030 	bio_init(&dev->rreq);
2031 	dev->rreq.bi_io_vec = &dev->rvec;
2032 	dev->rreq.bi_vcnt++;
2033 	dev->rreq.bi_max_vecs++;
2034 	dev->rreq.bi_private = sh;
2035 	dev->rvec.bv_page = dev->page;
2036 
2037 	dev->flags = 0;
2038 	dev->sector = compute_blocknr(sh, i, previous);
2039 }
2040 
2041 static void error(struct mddev *mddev, struct md_rdev *rdev)
2042 {
2043 	char b[BDEVNAME_SIZE];
2044 	struct r5conf *conf = mddev->private;
2045 	unsigned long flags;
2046 	pr_debug("raid456: error called\n");
2047 
2048 	spin_lock_irqsave(&conf->device_lock, flags);
2049 	clear_bit(In_sync, &rdev->flags);
2050 	mddev->degraded = calc_degraded(conf);
2051 	spin_unlock_irqrestore(&conf->device_lock, flags);
2052 	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2053 
2054 	set_bit(Blocked, &rdev->flags);
2055 	set_bit(Faulty, &rdev->flags);
2056 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
2057 	printk(KERN_ALERT
2058 	       "md/raid:%s: Disk failure on %s, disabling device.\n"
2059 	       "md/raid:%s: Operation continuing on %d devices.\n",
2060 	       mdname(mddev),
2061 	       bdevname(rdev->bdev, b),
2062 	       mdname(mddev),
2063 	       conf->raid_disks - mddev->degraded);
2064 }
2065 
2066 /*
2067  * Input: a 'big' sector number,
2068  * Output: index of the data and parity disk, and the sector # in them.
2069  */
2070 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2071 				     int previous, int *dd_idx,
2072 				     struct stripe_head *sh)
2073 {
2074 	sector_t stripe, stripe2;
2075 	sector_t chunk_number;
2076 	unsigned int chunk_offset;
2077 	int pd_idx, qd_idx;
2078 	int ddf_layout = 0;
2079 	sector_t new_sector;
2080 	int algorithm = previous ? conf->prev_algo
2081 				 : conf->algorithm;
2082 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2083 					 : conf->chunk_sectors;
2084 	int raid_disks = previous ? conf->previous_raid_disks
2085 				  : conf->raid_disks;
2086 	int data_disks = raid_disks - conf->max_degraded;
2087 
2088 	/* First compute the information on this sector */
2089 
2090 	/*
2091 	 * Compute the chunk number and the sector offset inside the chunk
2092 	 */
2093 	chunk_offset = sector_div(r_sector, sectors_per_chunk);
2094 	chunk_number = r_sector;
2095 
2096 	/*
2097 	 * Compute the stripe number
2098 	 */
2099 	stripe = chunk_number;
2100 	*dd_idx = sector_div(stripe, data_disks);
2101 	stripe2 = stripe;
2102 	/*
2103 	 * Select the parity disk based on the user selected algorithm.
2104 	 */
2105 	pd_idx = qd_idx = -1;
2106 	switch(conf->level) {
2107 	case 4:
2108 		pd_idx = data_disks;
2109 		break;
2110 	case 5:
2111 		switch (algorithm) {
2112 		case ALGORITHM_LEFT_ASYMMETRIC:
2113 			pd_idx = data_disks - sector_div(stripe2, raid_disks);
2114 			if (*dd_idx >= pd_idx)
2115 				(*dd_idx)++;
2116 			break;
2117 		case ALGORITHM_RIGHT_ASYMMETRIC:
2118 			pd_idx = sector_div(stripe2, raid_disks);
2119 			if (*dd_idx >= pd_idx)
2120 				(*dd_idx)++;
2121 			break;
2122 		case ALGORITHM_LEFT_SYMMETRIC:
2123 			pd_idx = data_disks - sector_div(stripe2, raid_disks);
2124 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2125 			break;
2126 		case ALGORITHM_RIGHT_SYMMETRIC:
2127 			pd_idx = sector_div(stripe2, raid_disks);
2128 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2129 			break;
2130 		case ALGORITHM_PARITY_0:
2131 			pd_idx = 0;
2132 			(*dd_idx)++;
2133 			break;
2134 		case ALGORITHM_PARITY_N:
2135 			pd_idx = data_disks;
2136 			break;
2137 		default:
2138 			BUG();
2139 		}
2140 		break;
2141 	case 6:
2142 
2143 		switch (algorithm) {
2144 		case ALGORITHM_LEFT_ASYMMETRIC:
2145 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2146 			qd_idx = pd_idx + 1;
2147 			if (pd_idx == raid_disks-1) {
2148 				(*dd_idx)++;	/* Q D D D P */
2149 				qd_idx = 0;
2150 			} else if (*dd_idx >= pd_idx)
2151 				(*dd_idx) += 2; /* D D P Q D */
2152 			break;
2153 		case ALGORITHM_RIGHT_ASYMMETRIC:
2154 			pd_idx = sector_div(stripe2, raid_disks);
2155 			qd_idx = pd_idx + 1;
2156 			if (pd_idx == raid_disks-1) {
2157 				(*dd_idx)++;	/* Q D D D P */
2158 				qd_idx = 0;
2159 			} else if (*dd_idx >= pd_idx)
2160 				(*dd_idx) += 2; /* D D P Q D */
2161 			break;
2162 		case ALGORITHM_LEFT_SYMMETRIC:
2163 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2164 			qd_idx = (pd_idx + 1) % raid_disks;
2165 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2166 			break;
2167 		case ALGORITHM_RIGHT_SYMMETRIC:
2168 			pd_idx = sector_div(stripe2, raid_disks);
2169 			qd_idx = (pd_idx + 1) % raid_disks;
2170 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2171 			break;
2172 
2173 		case ALGORITHM_PARITY_0:
2174 			pd_idx = 0;
2175 			qd_idx = 1;
2176 			(*dd_idx) += 2;
2177 			break;
2178 		case ALGORITHM_PARITY_N:
2179 			pd_idx = data_disks;
2180 			qd_idx = data_disks + 1;
2181 			break;
2182 
2183 		case ALGORITHM_ROTATING_ZERO_RESTART:
2184 			/* Exactly the same as RIGHT_ASYMMETRIC, but or
2185 			 * of blocks for computing Q is different.
2186 			 */
2187 			pd_idx = sector_div(stripe2, raid_disks);
2188 			qd_idx = pd_idx + 1;
2189 			if (pd_idx == raid_disks-1) {
2190 				(*dd_idx)++;	/* Q D D D P */
2191 				qd_idx = 0;
2192 			} else if (*dd_idx >= pd_idx)
2193 				(*dd_idx) += 2; /* D D P Q D */
2194 			ddf_layout = 1;
2195 			break;
2196 
2197 		case ALGORITHM_ROTATING_N_RESTART:
2198 			/* Same a left_asymmetric, by first stripe is
2199 			 * D D D P Q  rather than
2200 			 * Q D D D P
2201 			 */
2202 			stripe2 += 1;
2203 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2204 			qd_idx = pd_idx + 1;
2205 			if (pd_idx == raid_disks-1) {
2206 				(*dd_idx)++;	/* Q D D D P */
2207 				qd_idx = 0;
2208 			} else if (*dd_idx >= pd_idx)
2209 				(*dd_idx) += 2; /* D D P Q D */
2210 			ddf_layout = 1;
2211 			break;
2212 
2213 		case ALGORITHM_ROTATING_N_CONTINUE:
2214 			/* Same as left_symmetric but Q is before P */
2215 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2216 			qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2217 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2218 			ddf_layout = 1;
2219 			break;
2220 
2221 		case ALGORITHM_LEFT_ASYMMETRIC_6:
2222 			/* RAID5 left_asymmetric, with Q on last device */
2223 			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2224 			if (*dd_idx >= pd_idx)
2225 				(*dd_idx)++;
2226 			qd_idx = raid_disks - 1;
2227 			break;
2228 
2229 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
2230 			pd_idx = sector_div(stripe2, raid_disks-1);
2231 			if (*dd_idx >= pd_idx)
2232 				(*dd_idx)++;
2233 			qd_idx = raid_disks - 1;
2234 			break;
2235 
2236 		case ALGORITHM_LEFT_SYMMETRIC_6:
2237 			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2238 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2239 			qd_idx = raid_disks - 1;
2240 			break;
2241 
2242 		case ALGORITHM_RIGHT_SYMMETRIC_6:
2243 			pd_idx = sector_div(stripe2, raid_disks-1);
2244 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2245 			qd_idx = raid_disks - 1;
2246 			break;
2247 
2248 		case ALGORITHM_PARITY_0_6:
2249 			pd_idx = 0;
2250 			(*dd_idx)++;
2251 			qd_idx = raid_disks - 1;
2252 			break;
2253 
2254 		default:
2255 			BUG();
2256 		}
2257 		break;
2258 	}
2259 
2260 	if (sh) {
2261 		sh->pd_idx = pd_idx;
2262 		sh->qd_idx = qd_idx;
2263 		sh->ddf_layout = ddf_layout;
2264 	}
2265 	/*
2266 	 * Finally, compute the new sector number
2267 	 */
2268 	new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2269 	return new_sector;
2270 }
2271 
2272 
2273 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2274 {
2275 	struct r5conf *conf = sh->raid_conf;
2276 	int raid_disks = sh->disks;
2277 	int data_disks = raid_disks - conf->max_degraded;
2278 	sector_t new_sector = sh->sector, check;
2279 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2280 					 : conf->chunk_sectors;
2281 	int algorithm = previous ? conf->prev_algo
2282 				 : conf->algorithm;
2283 	sector_t stripe;
2284 	int chunk_offset;
2285 	sector_t chunk_number;
2286 	int dummy1, dd_idx = i;
2287 	sector_t r_sector;
2288 	struct stripe_head sh2;
2289 
2290 
2291 	chunk_offset = sector_div(new_sector, sectors_per_chunk);
2292 	stripe = new_sector;
2293 
2294 	if (i == sh->pd_idx)
2295 		return 0;
2296 	switch(conf->level) {
2297 	case 4: break;
2298 	case 5:
2299 		switch (algorithm) {
2300 		case ALGORITHM_LEFT_ASYMMETRIC:
2301 		case ALGORITHM_RIGHT_ASYMMETRIC:
2302 			if (i > sh->pd_idx)
2303 				i--;
2304 			break;
2305 		case ALGORITHM_LEFT_SYMMETRIC:
2306 		case ALGORITHM_RIGHT_SYMMETRIC:
2307 			if (i < sh->pd_idx)
2308 				i += raid_disks;
2309 			i -= (sh->pd_idx + 1);
2310 			break;
2311 		case ALGORITHM_PARITY_0:
2312 			i -= 1;
2313 			break;
2314 		case ALGORITHM_PARITY_N:
2315 			break;
2316 		default:
2317 			BUG();
2318 		}
2319 		break;
2320 	case 6:
2321 		if (i == sh->qd_idx)
2322 			return 0; /* It is the Q disk */
2323 		switch (algorithm) {
2324 		case ALGORITHM_LEFT_ASYMMETRIC:
2325 		case ALGORITHM_RIGHT_ASYMMETRIC:
2326 		case ALGORITHM_ROTATING_ZERO_RESTART:
2327 		case ALGORITHM_ROTATING_N_RESTART:
2328 			if (sh->pd_idx == raid_disks-1)
2329 				i--;	/* Q D D D P */
2330 			else if (i > sh->pd_idx)
2331 				i -= 2; /* D D P Q D */
2332 			break;
2333 		case ALGORITHM_LEFT_SYMMETRIC:
2334 		case ALGORITHM_RIGHT_SYMMETRIC:
2335 			if (sh->pd_idx == raid_disks-1)
2336 				i--; /* Q D D D P */
2337 			else {
2338 				/* D D P Q D */
2339 				if (i < sh->pd_idx)
2340 					i += raid_disks;
2341 				i -= (sh->pd_idx + 2);
2342 			}
2343 			break;
2344 		case ALGORITHM_PARITY_0:
2345 			i -= 2;
2346 			break;
2347 		case ALGORITHM_PARITY_N:
2348 			break;
2349 		case ALGORITHM_ROTATING_N_CONTINUE:
2350 			/* Like left_symmetric, but P is before Q */
2351 			if (sh->pd_idx == 0)
2352 				i--;	/* P D D D Q */
2353 			else {
2354 				/* D D Q P D */
2355 				if (i < sh->pd_idx)
2356 					i += raid_disks;
2357 				i -= (sh->pd_idx + 1);
2358 			}
2359 			break;
2360 		case ALGORITHM_LEFT_ASYMMETRIC_6:
2361 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
2362 			if (i > sh->pd_idx)
2363 				i--;
2364 			break;
2365 		case ALGORITHM_LEFT_SYMMETRIC_6:
2366 		case ALGORITHM_RIGHT_SYMMETRIC_6:
2367 			if (i < sh->pd_idx)
2368 				i += data_disks + 1;
2369 			i -= (sh->pd_idx + 1);
2370 			break;
2371 		case ALGORITHM_PARITY_0_6:
2372 			i -= 1;
2373 			break;
2374 		default:
2375 			BUG();
2376 		}
2377 		break;
2378 	}
2379 
2380 	chunk_number = stripe * data_disks + i;
2381 	r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2382 
2383 	check = raid5_compute_sector(conf, r_sector,
2384 				     previous, &dummy1, &sh2);
2385 	if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2386 		|| sh2.qd_idx != sh->qd_idx) {
2387 		printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2388 		       mdname(conf->mddev));
2389 		return 0;
2390 	}
2391 	return r_sector;
2392 }
2393 
2394 
2395 static void
2396 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2397 			 int rcw, int expand)
2398 {
2399 	int i, pd_idx = sh->pd_idx, disks = sh->disks;
2400 	struct r5conf *conf = sh->raid_conf;
2401 	int level = conf->level;
2402 
2403 	if (rcw) {
2404 
2405 		for (i = disks; i--; ) {
2406 			struct r5dev *dev = &sh->dev[i];
2407 
2408 			if (dev->towrite) {
2409 				set_bit(R5_LOCKED, &dev->flags);
2410 				set_bit(R5_Wantdrain, &dev->flags);
2411 				if (!expand)
2412 					clear_bit(R5_UPTODATE, &dev->flags);
2413 				s->locked++;
2414 			}
2415 		}
2416 		/* if we are not expanding this is a proper write request, and
2417 		 * there will be bios with new data to be drained into the
2418 		 * stripe cache
2419 		 */
2420 		if (!expand) {
2421 			if (!s->locked)
2422 				/* False alarm, nothing to do */
2423 				return;
2424 			sh->reconstruct_state = reconstruct_state_drain_run;
2425 			set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2426 		} else
2427 			sh->reconstruct_state = reconstruct_state_run;
2428 
2429 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2430 
2431 		if (s->locked + conf->max_degraded == disks)
2432 			if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2433 				atomic_inc(&conf->pending_full_writes);
2434 	} else {
2435 		BUG_ON(level == 6);
2436 		BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2437 			test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2438 
2439 		for (i = disks; i--; ) {
2440 			struct r5dev *dev = &sh->dev[i];
2441 			if (i == pd_idx)
2442 				continue;
2443 
2444 			if (dev->towrite &&
2445 			    (test_bit(R5_UPTODATE, &dev->flags) ||
2446 			     test_bit(R5_Wantcompute, &dev->flags))) {
2447 				set_bit(R5_Wantdrain, &dev->flags);
2448 				set_bit(R5_LOCKED, &dev->flags);
2449 				clear_bit(R5_UPTODATE, &dev->flags);
2450 				s->locked++;
2451 			}
2452 		}
2453 		if (!s->locked)
2454 			/* False alarm - nothing to do */
2455 			return;
2456 		sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2457 		set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2458 		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2459 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2460 	}
2461 
2462 	/* keep the parity disk(s) locked while asynchronous operations
2463 	 * are in flight
2464 	 */
2465 	set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2466 	clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2467 	s->locked++;
2468 
2469 	if (level == 6) {
2470 		int qd_idx = sh->qd_idx;
2471 		struct r5dev *dev = &sh->dev[qd_idx];
2472 
2473 		set_bit(R5_LOCKED, &dev->flags);
2474 		clear_bit(R5_UPTODATE, &dev->flags);
2475 		s->locked++;
2476 	}
2477 
2478 	pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2479 		__func__, (unsigned long long)sh->sector,
2480 		s->locked, s->ops_request);
2481 }
2482 
2483 /*
2484  * Each stripe/dev can have one or more bion attached.
2485  * toread/towrite point to the first in a chain.
2486  * The bi_next chain must be in order.
2487  */
2488 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2489 {
2490 	struct bio **bip;
2491 	struct r5conf *conf = sh->raid_conf;
2492 	int firstwrite=0;
2493 
2494 	pr_debug("adding bi b#%llu to stripe s#%llu\n",
2495 		(unsigned long long)bi->bi_sector,
2496 		(unsigned long long)sh->sector);
2497 
2498 	/*
2499 	 * If several bio share a stripe. The bio bi_phys_segments acts as a
2500 	 * reference count to avoid race. The reference count should already be
2501 	 * increased before this function is called (for example, in
2502 	 * make_request()), so other bio sharing this stripe will not free the
2503 	 * stripe. If a stripe is owned by one stripe, the stripe lock will
2504 	 * protect it.
2505 	 */
2506 	spin_lock_irq(&sh->stripe_lock);
2507 	if (forwrite) {
2508 		bip = &sh->dev[dd_idx].towrite;
2509 		if (*bip == NULL)
2510 			firstwrite = 1;
2511 	} else
2512 		bip = &sh->dev[dd_idx].toread;
2513 	while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2514 		if (bio_end_sector(*bip) > bi->bi_sector)
2515 			goto overlap;
2516 		bip = & (*bip)->bi_next;
2517 	}
2518 	if (*bip && (*bip)->bi_sector < bio_end_sector(bi))
2519 		goto overlap;
2520 
2521 	BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2522 	if (*bip)
2523 		bi->bi_next = *bip;
2524 	*bip = bi;
2525 	raid5_inc_bi_active_stripes(bi);
2526 
2527 	if (forwrite) {
2528 		/* check if page is covered */
2529 		sector_t sector = sh->dev[dd_idx].sector;
2530 		for (bi=sh->dev[dd_idx].towrite;
2531 		     sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2532 			     bi && bi->bi_sector <= sector;
2533 		     bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2534 			if (bio_end_sector(bi) >= sector)
2535 				sector = bio_end_sector(bi);
2536 		}
2537 		if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2538 			set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2539 	}
2540 
2541 	pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2542 		(unsigned long long)(*bip)->bi_sector,
2543 		(unsigned long long)sh->sector, dd_idx);
2544 	spin_unlock_irq(&sh->stripe_lock);
2545 
2546 	if (conf->mddev->bitmap && firstwrite) {
2547 		bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2548 				  STRIPE_SECTORS, 0);
2549 		sh->bm_seq = conf->seq_flush+1;
2550 		set_bit(STRIPE_BIT_DELAY, &sh->state);
2551 	}
2552 	return 1;
2553 
2554  overlap:
2555 	set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2556 	spin_unlock_irq(&sh->stripe_lock);
2557 	return 0;
2558 }
2559 
2560 static void end_reshape(struct r5conf *conf);
2561 
2562 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2563 			    struct stripe_head *sh)
2564 {
2565 	int sectors_per_chunk =
2566 		previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2567 	int dd_idx;
2568 	int chunk_offset = sector_div(stripe, sectors_per_chunk);
2569 	int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2570 
2571 	raid5_compute_sector(conf,
2572 			     stripe * (disks - conf->max_degraded)
2573 			     *sectors_per_chunk + chunk_offset,
2574 			     previous,
2575 			     &dd_idx, sh);
2576 }
2577 
2578 static void
2579 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2580 				struct stripe_head_state *s, int disks,
2581 				struct bio **return_bi)
2582 {
2583 	int i;
2584 	for (i = disks; i--; ) {
2585 		struct bio *bi;
2586 		int bitmap_end = 0;
2587 
2588 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2589 			struct md_rdev *rdev;
2590 			rcu_read_lock();
2591 			rdev = rcu_dereference(conf->disks[i].rdev);
2592 			if (rdev && test_bit(In_sync, &rdev->flags))
2593 				atomic_inc(&rdev->nr_pending);
2594 			else
2595 				rdev = NULL;
2596 			rcu_read_unlock();
2597 			if (rdev) {
2598 				if (!rdev_set_badblocks(
2599 					    rdev,
2600 					    sh->sector,
2601 					    STRIPE_SECTORS, 0))
2602 					md_error(conf->mddev, rdev);
2603 				rdev_dec_pending(rdev, conf->mddev);
2604 			}
2605 		}
2606 		spin_lock_irq(&sh->stripe_lock);
2607 		/* fail all writes first */
2608 		bi = sh->dev[i].towrite;
2609 		sh->dev[i].towrite = NULL;
2610 		spin_unlock_irq(&sh->stripe_lock);
2611 		if (bi)
2612 			bitmap_end = 1;
2613 
2614 		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2615 			wake_up(&conf->wait_for_overlap);
2616 
2617 		while (bi && bi->bi_sector <
2618 			sh->dev[i].sector + STRIPE_SECTORS) {
2619 			struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2620 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
2621 			if (!raid5_dec_bi_active_stripes(bi)) {
2622 				md_write_end(conf->mddev);
2623 				bi->bi_next = *return_bi;
2624 				*return_bi = bi;
2625 			}
2626 			bi = nextbi;
2627 		}
2628 		if (bitmap_end)
2629 			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2630 				STRIPE_SECTORS, 0, 0);
2631 		bitmap_end = 0;
2632 		/* and fail all 'written' */
2633 		bi = sh->dev[i].written;
2634 		sh->dev[i].written = NULL;
2635 		if (bi) bitmap_end = 1;
2636 		while (bi && bi->bi_sector <
2637 		       sh->dev[i].sector + STRIPE_SECTORS) {
2638 			struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2639 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
2640 			if (!raid5_dec_bi_active_stripes(bi)) {
2641 				md_write_end(conf->mddev);
2642 				bi->bi_next = *return_bi;
2643 				*return_bi = bi;
2644 			}
2645 			bi = bi2;
2646 		}
2647 
2648 		/* fail any reads if this device is non-operational and
2649 		 * the data has not reached the cache yet.
2650 		 */
2651 		if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2652 		    (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2653 		      test_bit(R5_ReadError, &sh->dev[i].flags))) {
2654 			spin_lock_irq(&sh->stripe_lock);
2655 			bi = sh->dev[i].toread;
2656 			sh->dev[i].toread = NULL;
2657 			spin_unlock_irq(&sh->stripe_lock);
2658 			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2659 				wake_up(&conf->wait_for_overlap);
2660 			while (bi && bi->bi_sector <
2661 			       sh->dev[i].sector + STRIPE_SECTORS) {
2662 				struct bio *nextbi =
2663 					r5_next_bio(bi, sh->dev[i].sector);
2664 				clear_bit(BIO_UPTODATE, &bi->bi_flags);
2665 				if (!raid5_dec_bi_active_stripes(bi)) {
2666 					bi->bi_next = *return_bi;
2667 					*return_bi = bi;
2668 				}
2669 				bi = nextbi;
2670 			}
2671 		}
2672 		if (bitmap_end)
2673 			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2674 					STRIPE_SECTORS, 0, 0);
2675 		/* If we were in the middle of a write the parity block might
2676 		 * still be locked - so just clear all R5_LOCKED flags
2677 		 */
2678 		clear_bit(R5_LOCKED, &sh->dev[i].flags);
2679 	}
2680 
2681 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2682 		if (atomic_dec_and_test(&conf->pending_full_writes))
2683 			md_wakeup_thread(conf->mddev->thread);
2684 }
2685 
2686 static void
2687 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2688 		   struct stripe_head_state *s)
2689 {
2690 	int abort = 0;
2691 	int i;
2692 
2693 	clear_bit(STRIPE_SYNCING, &sh->state);
2694 	if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
2695 		wake_up(&conf->wait_for_overlap);
2696 	s->syncing = 0;
2697 	s->replacing = 0;
2698 	/* There is nothing more to do for sync/check/repair.
2699 	 * Don't even need to abort as that is handled elsewhere
2700 	 * if needed, and not always wanted e.g. if there is a known
2701 	 * bad block here.
2702 	 * For recover/replace we need to record a bad block on all
2703 	 * non-sync devices, or abort the recovery
2704 	 */
2705 	if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
2706 		/* During recovery devices cannot be removed, so
2707 		 * locking and refcounting of rdevs is not needed
2708 		 */
2709 		for (i = 0; i < conf->raid_disks; i++) {
2710 			struct md_rdev *rdev = conf->disks[i].rdev;
2711 			if (rdev
2712 			    && !test_bit(Faulty, &rdev->flags)
2713 			    && !test_bit(In_sync, &rdev->flags)
2714 			    && !rdev_set_badblocks(rdev, sh->sector,
2715 						   STRIPE_SECTORS, 0))
2716 				abort = 1;
2717 			rdev = conf->disks[i].replacement;
2718 			if (rdev
2719 			    && !test_bit(Faulty, &rdev->flags)
2720 			    && !test_bit(In_sync, &rdev->flags)
2721 			    && !rdev_set_badblocks(rdev, sh->sector,
2722 						   STRIPE_SECTORS, 0))
2723 				abort = 1;
2724 		}
2725 		if (abort)
2726 			conf->recovery_disabled =
2727 				conf->mddev->recovery_disabled;
2728 	}
2729 	md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
2730 }
2731 
2732 static int want_replace(struct stripe_head *sh, int disk_idx)
2733 {
2734 	struct md_rdev *rdev;
2735 	int rv = 0;
2736 	/* Doing recovery so rcu locking not required */
2737 	rdev = sh->raid_conf->disks[disk_idx].replacement;
2738 	if (rdev
2739 	    && !test_bit(Faulty, &rdev->flags)
2740 	    && !test_bit(In_sync, &rdev->flags)
2741 	    && (rdev->recovery_offset <= sh->sector
2742 		|| rdev->mddev->recovery_cp <= sh->sector))
2743 		rv = 1;
2744 
2745 	return rv;
2746 }
2747 
2748 /* fetch_block - checks the given member device to see if its data needs
2749  * to be read or computed to satisfy a request.
2750  *
2751  * Returns 1 when no more member devices need to be checked, otherwise returns
2752  * 0 to tell the loop in handle_stripe_fill to continue
2753  */
2754 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2755 		       int disk_idx, int disks)
2756 {
2757 	struct r5dev *dev = &sh->dev[disk_idx];
2758 	struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2759 				  &sh->dev[s->failed_num[1]] };
2760 
2761 	/* is the data in this block needed, and can we get it? */
2762 	if (!test_bit(R5_LOCKED, &dev->flags) &&
2763 	    !test_bit(R5_UPTODATE, &dev->flags) &&
2764 	    (dev->toread ||
2765 	     (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2766 	     s->syncing || s->expanding ||
2767 	     (s->replacing && want_replace(sh, disk_idx)) ||
2768 	     (s->failed >= 1 && fdev[0]->toread) ||
2769 	     (s->failed >= 2 && fdev[1]->toread) ||
2770 	     (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2771 	      !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2772 	     (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2773 		/* we would like to get this block, possibly by computing it,
2774 		 * otherwise read it if the backing disk is insync
2775 		 */
2776 		BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2777 		BUG_ON(test_bit(R5_Wantread, &dev->flags));
2778 		if ((s->uptodate == disks - 1) &&
2779 		    (s->failed && (disk_idx == s->failed_num[0] ||
2780 				   disk_idx == s->failed_num[1]))) {
2781 			/* have disk failed, and we're requested to fetch it;
2782 			 * do compute it
2783 			 */
2784 			pr_debug("Computing stripe %llu block %d\n",
2785 			       (unsigned long long)sh->sector, disk_idx);
2786 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2787 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2788 			set_bit(R5_Wantcompute, &dev->flags);
2789 			sh->ops.target = disk_idx;
2790 			sh->ops.target2 = -1; /* no 2nd target */
2791 			s->req_compute = 1;
2792 			/* Careful: from this point on 'uptodate' is in the eye
2793 			 * of raid_run_ops which services 'compute' operations
2794 			 * before writes. R5_Wantcompute flags a block that will
2795 			 * be R5_UPTODATE by the time it is needed for a
2796 			 * subsequent operation.
2797 			 */
2798 			s->uptodate++;
2799 			return 1;
2800 		} else if (s->uptodate == disks-2 && s->failed >= 2) {
2801 			/* Computing 2-failure is *very* expensive; only
2802 			 * do it if failed >= 2
2803 			 */
2804 			int other;
2805 			for (other = disks; other--; ) {
2806 				if (other == disk_idx)
2807 					continue;
2808 				if (!test_bit(R5_UPTODATE,
2809 				      &sh->dev[other].flags))
2810 					break;
2811 			}
2812 			BUG_ON(other < 0);
2813 			pr_debug("Computing stripe %llu blocks %d,%d\n",
2814 			       (unsigned long long)sh->sector,
2815 			       disk_idx, other);
2816 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2817 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2818 			set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2819 			set_bit(R5_Wantcompute, &sh->dev[other].flags);
2820 			sh->ops.target = disk_idx;
2821 			sh->ops.target2 = other;
2822 			s->uptodate += 2;
2823 			s->req_compute = 1;
2824 			return 1;
2825 		} else if (test_bit(R5_Insync, &dev->flags)) {
2826 			set_bit(R5_LOCKED, &dev->flags);
2827 			set_bit(R5_Wantread, &dev->flags);
2828 			s->locked++;
2829 			pr_debug("Reading block %d (sync=%d)\n",
2830 				disk_idx, s->syncing);
2831 		}
2832 	}
2833 
2834 	return 0;
2835 }
2836 
2837 /**
2838  * handle_stripe_fill - read or compute data to satisfy pending requests.
2839  */
2840 static void handle_stripe_fill(struct stripe_head *sh,
2841 			       struct stripe_head_state *s,
2842 			       int disks)
2843 {
2844 	int i;
2845 
2846 	/* look for blocks to read/compute, skip this if a compute
2847 	 * is already in flight, or if the stripe contents are in the
2848 	 * midst of changing due to a write
2849 	 */
2850 	if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2851 	    !sh->reconstruct_state)
2852 		for (i = disks; i--; )
2853 			if (fetch_block(sh, s, i, disks))
2854 				break;
2855 	set_bit(STRIPE_HANDLE, &sh->state);
2856 }
2857 
2858 
2859 /* handle_stripe_clean_event
2860  * any written block on an uptodate or failed drive can be returned.
2861  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2862  * never LOCKED, so we don't need to test 'failed' directly.
2863  */
2864 static void handle_stripe_clean_event(struct r5conf *conf,
2865 	struct stripe_head *sh, int disks, struct bio **return_bi)
2866 {
2867 	int i;
2868 	struct r5dev *dev;
2869 	int discard_pending = 0;
2870 
2871 	for (i = disks; i--; )
2872 		if (sh->dev[i].written) {
2873 			dev = &sh->dev[i];
2874 			if (!test_bit(R5_LOCKED, &dev->flags) &&
2875 			    (test_bit(R5_UPTODATE, &dev->flags) ||
2876 			     test_bit(R5_Discard, &dev->flags))) {
2877 				/* We can return any write requests */
2878 				struct bio *wbi, *wbi2;
2879 				pr_debug("Return write for disc %d\n", i);
2880 				if (test_and_clear_bit(R5_Discard, &dev->flags))
2881 					clear_bit(R5_UPTODATE, &dev->flags);
2882 				wbi = dev->written;
2883 				dev->written = NULL;
2884 				while (wbi && wbi->bi_sector <
2885 					dev->sector + STRIPE_SECTORS) {
2886 					wbi2 = r5_next_bio(wbi, dev->sector);
2887 					if (!raid5_dec_bi_active_stripes(wbi)) {
2888 						md_write_end(conf->mddev);
2889 						wbi->bi_next = *return_bi;
2890 						*return_bi = wbi;
2891 					}
2892 					wbi = wbi2;
2893 				}
2894 				bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2895 						STRIPE_SECTORS,
2896 					 !test_bit(STRIPE_DEGRADED, &sh->state),
2897 						0);
2898 			} else if (test_bit(R5_Discard, &dev->flags))
2899 				discard_pending = 1;
2900 		}
2901 	if (!discard_pending &&
2902 	    test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
2903 		clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
2904 		clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2905 		if (sh->qd_idx >= 0) {
2906 			clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
2907 			clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
2908 		}
2909 		/* now that discard is done we can proceed with any sync */
2910 		clear_bit(STRIPE_DISCARD, &sh->state);
2911 		/*
2912 		 * SCSI discard will change some bio fields and the stripe has
2913 		 * no updated data, so remove it from hash list and the stripe
2914 		 * will be reinitialized
2915 		 */
2916 		spin_lock_irq(&conf->device_lock);
2917 		remove_hash(sh);
2918 		spin_unlock_irq(&conf->device_lock);
2919 		if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2920 			set_bit(STRIPE_HANDLE, &sh->state);
2921 
2922 	}
2923 
2924 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2925 		if (atomic_dec_and_test(&conf->pending_full_writes))
2926 			md_wakeup_thread(conf->mddev->thread);
2927 }
2928 
2929 static void handle_stripe_dirtying(struct r5conf *conf,
2930 				   struct stripe_head *sh,
2931 				   struct stripe_head_state *s,
2932 				   int disks)
2933 {
2934 	int rmw = 0, rcw = 0, i;
2935 	sector_t recovery_cp = conf->mddev->recovery_cp;
2936 
2937 	/* RAID6 requires 'rcw' in current implementation.
2938 	 * Otherwise, check whether resync is now happening or should start.
2939 	 * If yes, then the array is dirty (after unclean shutdown or
2940 	 * initial creation), so parity in some stripes might be inconsistent.
2941 	 * In this case, we need to always do reconstruct-write, to ensure
2942 	 * that in case of drive failure or read-error correction, we
2943 	 * generate correct data from the parity.
2944 	 */
2945 	if (conf->max_degraded == 2 ||
2946 	    (recovery_cp < MaxSector && sh->sector >= recovery_cp)) {
2947 		/* Calculate the real rcw later - for now make it
2948 		 * look like rcw is cheaper
2949 		 */
2950 		rcw = 1; rmw = 2;
2951 		pr_debug("force RCW max_degraded=%u, recovery_cp=%llu sh->sector=%llu\n",
2952 			 conf->max_degraded, (unsigned long long)recovery_cp,
2953 			 (unsigned long long)sh->sector);
2954 	} else for (i = disks; i--; ) {
2955 		/* would I have to read this buffer for read_modify_write */
2956 		struct r5dev *dev = &sh->dev[i];
2957 		if ((dev->towrite || i == sh->pd_idx) &&
2958 		    !test_bit(R5_LOCKED, &dev->flags) &&
2959 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2960 		      test_bit(R5_Wantcompute, &dev->flags))) {
2961 			if (test_bit(R5_Insync, &dev->flags))
2962 				rmw++;
2963 			else
2964 				rmw += 2*disks;  /* cannot read it */
2965 		}
2966 		/* Would I have to read this buffer for reconstruct_write */
2967 		if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2968 		    !test_bit(R5_LOCKED, &dev->flags) &&
2969 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2970 		    test_bit(R5_Wantcompute, &dev->flags))) {
2971 			if (test_bit(R5_Insync, &dev->flags)) rcw++;
2972 			else
2973 				rcw += 2*disks;
2974 		}
2975 	}
2976 	pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2977 		(unsigned long long)sh->sector, rmw, rcw);
2978 	set_bit(STRIPE_HANDLE, &sh->state);
2979 	if (rmw < rcw && rmw > 0) {
2980 		/* prefer read-modify-write, but need to get some data */
2981 		if (conf->mddev->queue)
2982 			blk_add_trace_msg(conf->mddev->queue,
2983 					  "raid5 rmw %llu %d",
2984 					  (unsigned long long)sh->sector, rmw);
2985 		for (i = disks; i--; ) {
2986 			struct r5dev *dev = &sh->dev[i];
2987 			if ((dev->towrite || i == sh->pd_idx) &&
2988 			    !test_bit(R5_LOCKED, &dev->flags) &&
2989 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
2990 			    test_bit(R5_Wantcompute, &dev->flags)) &&
2991 			    test_bit(R5_Insync, &dev->flags)) {
2992 				if (
2993 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2994 					pr_debug("Read_old block "
2995 						 "%d for r-m-w\n", i);
2996 					set_bit(R5_LOCKED, &dev->flags);
2997 					set_bit(R5_Wantread, &dev->flags);
2998 					s->locked++;
2999 				} else {
3000 					set_bit(STRIPE_DELAYED, &sh->state);
3001 					set_bit(STRIPE_HANDLE, &sh->state);
3002 				}
3003 			}
3004 		}
3005 	}
3006 	if (rcw <= rmw && rcw > 0) {
3007 		/* want reconstruct write, but need to get some data */
3008 		int qread =0;
3009 		rcw = 0;
3010 		for (i = disks; i--; ) {
3011 			struct r5dev *dev = &sh->dev[i];
3012 			if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3013 			    i != sh->pd_idx && i != sh->qd_idx &&
3014 			    !test_bit(R5_LOCKED, &dev->flags) &&
3015 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
3016 			      test_bit(R5_Wantcompute, &dev->flags))) {
3017 				rcw++;
3018 				if (!test_bit(R5_Insync, &dev->flags))
3019 					continue; /* it's a failed drive */
3020 				if (
3021 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3022 					pr_debug("Read_old block "
3023 						"%d for Reconstruct\n", i);
3024 					set_bit(R5_LOCKED, &dev->flags);
3025 					set_bit(R5_Wantread, &dev->flags);
3026 					s->locked++;
3027 					qread++;
3028 				} else {
3029 					set_bit(STRIPE_DELAYED, &sh->state);
3030 					set_bit(STRIPE_HANDLE, &sh->state);
3031 				}
3032 			}
3033 		}
3034 		if (rcw && conf->mddev->queue)
3035 			blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3036 					  (unsigned long long)sh->sector,
3037 					  rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3038 	}
3039 	/* now if nothing is locked, and if we have enough data,
3040 	 * we can start a write request
3041 	 */
3042 	/* since handle_stripe can be called at any time we need to handle the
3043 	 * case where a compute block operation has been submitted and then a
3044 	 * subsequent call wants to start a write request.  raid_run_ops only
3045 	 * handles the case where compute block and reconstruct are requested
3046 	 * simultaneously.  If this is not the case then new writes need to be
3047 	 * held off until the compute completes.
3048 	 */
3049 	if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3050 	    (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3051 	    !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3052 		schedule_reconstruction(sh, s, rcw == 0, 0);
3053 }
3054 
3055 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3056 				struct stripe_head_state *s, int disks)
3057 {
3058 	struct r5dev *dev = NULL;
3059 
3060 	set_bit(STRIPE_HANDLE, &sh->state);
3061 
3062 	switch (sh->check_state) {
3063 	case check_state_idle:
3064 		/* start a new check operation if there are no failures */
3065 		if (s->failed == 0) {
3066 			BUG_ON(s->uptodate != disks);
3067 			sh->check_state = check_state_run;
3068 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
3069 			clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3070 			s->uptodate--;
3071 			break;
3072 		}
3073 		dev = &sh->dev[s->failed_num[0]];
3074 		/* fall through */
3075 	case check_state_compute_result:
3076 		sh->check_state = check_state_idle;
3077 		if (!dev)
3078 			dev = &sh->dev[sh->pd_idx];
3079 
3080 		/* check that a write has not made the stripe insync */
3081 		if (test_bit(STRIPE_INSYNC, &sh->state))
3082 			break;
3083 
3084 		/* either failed parity check, or recovery is happening */
3085 		BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3086 		BUG_ON(s->uptodate != disks);
3087 
3088 		set_bit(R5_LOCKED, &dev->flags);
3089 		s->locked++;
3090 		set_bit(R5_Wantwrite, &dev->flags);
3091 
3092 		clear_bit(STRIPE_DEGRADED, &sh->state);
3093 		set_bit(STRIPE_INSYNC, &sh->state);
3094 		break;
3095 	case check_state_run:
3096 		break; /* we will be called again upon completion */
3097 	case check_state_check_result:
3098 		sh->check_state = check_state_idle;
3099 
3100 		/* if a failure occurred during the check operation, leave
3101 		 * STRIPE_INSYNC not set and let the stripe be handled again
3102 		 */
3103 		if (s->failed)
3104 			break;
3105 
3106 		/* handle a successful check operation, if parity is correct
3107 		 * we are done.  Otherwise update the mismatch count and repair
3108 		 * parity if !MD_RECOVERY_CHECK
3109 		 */
3110 		if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3111 			/* parity is correct (on disc,
3112 			 * not in buffer any more)
3113 			 */
3114 			set_bit(STRIPE_INSYNC, &sh->state);
3115 		else {
3116 			atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3117 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3118 				/* don't try to repair!! */
3119 				set_bit(STRIPE_INSYNC, &sh->state);
3120 			else {
3121 				sh->check_state = check_state_compute_run;
3122 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3123 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3124 				set_bit(R5_Wantcompute,
3125 					&sh->dev[sh->pd_idx].flags);
3126 				sh->ops.target = sh->pd_idx;
3127 				sh->ops.target2 = -1;
3128 				s->uptodate++;
3129 			}
3130 		}
3131 		break;
3132 	case check_state_compute_run:
3133 		break;
3134 	default:
3135 		printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3136 		       __func__, sh->check_state,
3137 		       (unsigned long long) sh->sector);
3138 		BUG();
3139 	}
3140 }
3141 
3142 
3143 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3144 				  struct stripe_head_state *s,
3145 				  int disks)
3146 {
3147 	int pd_idx = sh->pd_idx;
3148 	int qd_idx = sh->qd_idx;
3149 	struct r5dev *dev;
3150 
3151 	set_bit(STRIPE_HANDLE, &sh->state);
3152 
3153 	BUG_ON(s->failed > 2);
3154 
3155 	/* Want to check and possibly repair P and Q.
3156 	 * However there could be one 'failed' device, in which
3157 	 * case we can only check one of them, possibly using the
3158 	 * other to generate missing data
3159 	 */
3160 
3161 	switch (sh->check_state) {
3162 	case check_state_idle:
3163 		/* start a new check operation if there are < 2 failures */
3164 		if (s->failed == s->q_failed) {
3165 			/* The only possible failed device holds Q, so it
3166 			 * makes sense to check P (If anything else were failed,
3167 			 * we would have used P to recreate it).
3168 			 */
3169 			sh->check_state = check_state_run;
3170 		}
3171 		if (!s->q_failed && s->failed < 2) {
3172 			/* Q is not failed, and we didn't use it to generate
3173 			 * anything, so it makes sense to check it
3174 			 */
3175 			if (sh->check_state == check_state_run)
3176 				sh->check_state = check_state_run_pq;
3177 			else
3178 				sh->check_state = check_state_run_q;
3179 		}
3180 
3181 		/* discard potentially stale zero_sum_result */
3182 		sh->ops.zero_sum_result = 0;
3183 
3184 		if (sh->check_state == check_state_run) {
3185 			/* async_xor_zero_sum destroys the contents of P */
3186 			clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3187 			s->uptodate--;
3188 		}
3189 		if (sh->check_state >= check_state_run &&
3190 		    sh->check_state <= check_state_run_pq) {
3191 			/* async_syndrome_zero_sum preserves P and Q, so
3192 			 * no need to mark them !uptodate here
3193 			 */
3194 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
3195 			break;
3196 		}
3197 
3198 		/* we have 2-disk failure */
3199 		BUG_ON(s->failed != 2);
3200 		/* fall through */
3201 	case check_state_compute_result:
3202 		sh->check_state = check_state_idle;
3203 
3204 		/* check that a write has not made the stripe insync */
3205 		if (test_bit(STRIPE_INSYNC, &sh->state))
3206 			break;
3207 
3208 		/* now write out any block on a failed drive,
3209 		 * or P or Q if they were recomputed
3210 		 */
3211 		BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3212 		if (s->failed == 2) {
3213 			dev = &sh->dev[s->failed_num[1]];
3214 			s->locked++;
3215 			set_bit(R5_LOCKED, &dev->flags);
3216 			set_bit(R5_Wantwrite, &dev->flags);
3217 		}
3218 		if (s->failed >= 1) {
3219 			dev = &sh->dev[s->failed_num[0]];
3220 			s->locked++;
3221 			set_bit(R5_LOCKED, &dev->flags);
3222 			set_bit(R5_Wantwrite, &dev->flags);
3223 		}
3224 		if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3225 			dev = &sh->dev[pd_idx];
3226 			s->locked++;
3227 			set_bit(R5_LOCKED, &dev->flags);
3228 			set_bit(R5_Wantwrite, &dev->flags);
3229 		}
3230 		if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3231 			dev = &sh->dev[qd_idx];
3232 			s->locked++;
3233 			set_bit(R5_LOCKED, &dev->flags);
3234 			set_bit(R5_Wantwrite, &dev->flags);
3235 		}
3236 		clear_bit(STRIPE_DEGRADED, &sh->state);
3237 
3238 		set_bit(STRIPE_INSYNC, &sh->state);
3239 		break;
3240 	case check_state_run:
3241 	case check_state_run_q:
3242 	case check_state_run_pq:
3243 		break; /* we will be called again upon completion */
3244 	case check_state_check_result:
3245 		sh->check_state = check_state_idle;
3246 
3247 		/* handle a successful check operation, if parity is correct
3248 		 * we are done.  Otherwise update the mismatch count and repair
3249 		 * parity if !MD_RECOVERY_CHECK
3250 		 */
3251 		if (sh->ops.zero_sum_result == 0) {
3252 			/* both parities are correct */
3253 			if (!s->failed)
3254 				set_bit(STRIPE_INSYNC, &sh->state);
3255 			else {
3256 				/* in contrast to the raid5 case we can validate
3257 				 * parity, but still have a failure to write
3258 				 * back
3259 				 */
3260 				sh->check_state = check_state_compute_result;
3261 				/* Returning at this point means that we may go
3262 				 * off and bring p and/or q uptodate again so
3263 				 * we make sure to check zero_sum_result again
3264 				 * to verify if p or q need writeback
3265 				 */
3266 			}
3267 		} else {
3268 			atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3269 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3270 				/* don't try to repair!! */
3271 				set_bit(STRIPE_INSYNC, &sh->state);
3272 			else {
3273 				int *target = &sh->ops.target;
3274 
3275 				sh->ops.target = -1;
3276 				sh->ops.target2 = -1;
3277 				sh->check_state = check_state_compute_run;
3278 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3279 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3280 				if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3281 					set_bit(R5_Wantcompute,
3282 						&sh->dev[pd_idx].flags);
3283 					*target = pd_idx;
3284 					target = &sh->ops.target2;
3285 					s->uptodate++;
3286 				}
3287 				if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3288 					set_bit(R5_Wantcompute,
3289 						&sh->dev[qd_idx].flags);
3290 					*target = qd_idx;
3291 					s->uptodate++;
3292 				}
3293 			}
3294 		}
3295 		break;
3296 	case check_state_compute_run:
3297 		break;
3298 	default:
3299 		printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3300 		       __func__, sh->check_state,
3301 		       (unsigned long long) sh->sector);
3302 		BUG();
3303 	}
3304 }
3305 
3306 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3307 {
3308 	int i;
3309 
3310 	/* We have read all the blocks in this stripe and now we need to
3311 	 * copy some of them into a target stripe for expand.
3312 	 */
3313 	struct dma_async_tx_descriptor *tx = NULL;
3314 	clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3315 	for (i = 0; i < sh->disks; i++)
3316 		if (i != sh->pd_idx && i != sh->qd_idx) {
3317 			int dd_idx, j;
3318 			struct stripe_head *sh2;
3319 			struct async_submit_ctl submit;
3320 
3321 			sector_t bn = compute_blocknr(sh, i, 1);
3322 			sector_t s = raid5_compute_sector(conf, bn, 0,
3323 							  &dd_idx, NULL);
3324 			sh2 = get_active_stripe(conf, s, 0, 1, 1);
3325 			if (sh2 == NULL)
3326 				/* so far only the early blocks of this stripe
3327 				 * have been requested.  When later blocks
3328 				 * get requested, we will try again
3329 				 */
3330 				continue;
3331 			if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3332 			   test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3333 				/* must have already done this block */
3334 				release_stripe(sh2);
3335 				continue;
3336 			}
3337 
3338 			/* place all the copies on one channel */
3339 			init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3340 			tx = async_memcpy(sh2->dev[dd_idx].page,
3341 					  sh->dev[i].page, 0, 0, STRIPE_SIZE,
3342 					  &submit);
3343 
3344 			set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3345 			set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3346 			for (j = 0; j < conf->raid_disks; j++)
3347 				if (j != sh2->pd_idx &&
3348 				    j != sh2->qd_idx &&
3349 				    !test_bit(R5_Expanded, &sh2->dev[j].flags))
3350 					break;
3351 			if (j == conf->raid_disks) {
3352 				set_bit(STRIPE_EXPAND_READY, &sh2->state);
3353 				set_bit(STRIPE_HANDLE, &sh2->state);
3354 			}
3355 			release_stripe(sh2);
3356 
3357 		}
3358 	/* done submitting copies, wait for them to complete */
3359 	async_tx_quiesce(&tx);
3360 }
3361 
3362 /*
3363  * handle_stripe - do things to a stripe.
3364  *
3365  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3366  * state of various bits to see what needs to be done.
3367  * Possible results:
3368  *    return some read requests which now have data
3369  *    return some write requests which are safely on storage
3370  *    schedule a read on some buffers
3371  *    schedule a write of some buffers
3372  *    return confirmation of parity correctness
3373  *
3374  */
3375 
3376 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3377 {
3378 	struct r5conf *conf = sh->raid_conf;
3379 	int disks = sh->disks;
3380 	struct r5dev *dev;
3381 	int i;
3382 	int do_recovery = 0;
3383 
3384 	memset(s, 0, sizeof(*s));
3385 
3386 	s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3387 	s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3388 	s->failed_num[0] = -1;
3389 	s->failed_num[1] = -1;
3390 
3391 	/* Now to look around and see what can be done */
3392 	rcu_read_lock();
3393 	for (i=disks; i--; ) {
3394 		struct md_rdev *rdev;
3395 		sector_t first_bad;
3396 		int bad_sectors;
3397 		int is_bad = 0;
3398 
3399 		dev = &sh->dev[i];
3400 
3401 		pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3402 			 i, dev->flags,
3403 			 dev->toread, dev->towrite, dev->written);
3404 		/* maybe we can reply to a read
3405 		 *
3406 		 * new wantfill requests are only permitted while
3407 		 * ops_complete_biofill is guaranteed to be inactive
3408 		 */
3409 		if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3410 		    !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3411 			set_bit(R5_Wantfill, &dev->flags);
3412 
3413 		/* now count some things */
3414 		if (test_bit(R5_LOCKED, &dev->flags))
3415 			s->locked++;
3416 		if (test_bit(R5_UPTODATE, &dev->flags))
3417 			s->uptodate++;
3418 		if (test_bit(R5_Wantcompute, &dev->flags)) {
3419 			s->compute++;
3420 			BUG_ON(s->compute > 2);
3421 		}
3422 
3423 		if (test_bit(R5_Wantfill, &dev->flags))
3424 			s->to_fill++;
3425 		else if (dev->toread)
3426 			s->to_read++;
3427 		if (dev->towrite) {
3428 			s->to_write++;
3429 			if (!test_bit(R5_OVERWRITE, &dev->flags))
3430 				s->non_overwrite++;
3431 		}
3432 		if (dev->written)
3433 			s->written++;
3434 		/* Prefer to use the replacement for reads, but only
3435 		 * if it is recovered enough and has no bad blocks.
3436 		 */
3437 		rdev = rcu_dereference(conf->disks[i].replacement);
3438 		if (rdev && !test_bit(Faulty, &rdev->flags) &&
3439 		    rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
3440 		    !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3441 				 &first_bad, &bad_sectors))
3442 			set_bit(R5_ReadRepl, &dev->flags);
3443 		else {
3444 			if (rdev)
3445 				set_bit(R5_NeedReplace, &dev->flags);
3446 			rdev = rcu_dereference(conf->disks[i].rdev);
3447 			clear_bit(R5_ReadRepl, &dev->flags);
3448 		}
3449 		if (rdev && test_bit(Faulty, &rdev->flags))
3450 			rdev = NULL;
3451 		if (rdev) {
3452 			is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3453 					     &first_bad, &bad_sectors);
3454 			if (s->blocked_rdev == NULL
3455 			    && (test_bit(Blocked, &rdev->flags)
3456 				|| is_bad < 0)) {
3457 				if (is_bad < 0)
3458 					set_bit(BlockedBadBlocks,
3459 						&rdev->flags);
3460 				s->blocked_rdev = rdev;
3461 				atomic_inc(&rdev->nr_pending);
3462 			}
3463 		}
3464 		clear_bit(R5_Insync, &dev->flags);
3465 		if (!rdev)
3466 			/* Not in-sync */;
3467 		else if (is_bad) {
3468 			/* also not in-sync */
3469 			if (!test_bit(WriteErrorSeen, &rdev->flags) &&
3470 			    test_bit(R5_UPTODATE, &dev->flags)) {
3471 				/* treat as in-sync, but with a read error
3472 				 * which we can now try to correct
3473 				 */
3474 				set_bit(R5_Insync, &dev->flags);
3475 				set_bit(R5_ReadError, &dev->flags);
3476 			}
3477 		} else if (test_bit(In_sync, &rdev->flags))
3478 			set_bit(R5_Insync, &dev->flags);
3479 		else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3480 			/* in sync if before recovery_offset */
3481 			set_bit(R5_Insync, &dev->flags);
3482 		else if (test_bit(R5_UPTODATE, &dev->flags) &&
3483 			 test_bit(R5_Expanded, &dev->flags))
3484 			/* If we've reshaped into here, we assume it is Insync.
3485 			 * We will shortly update recovery_offset to make
3486 			 * it official.
3487 			 */
3488 			set_bit(R5_Insync, &dev->flags);
3489 
3490 		if (rdev && test_bit(R5_WriteError, &dev->flags)) {
3491 			/* This flag does not apply to '.replacement'
3492 			 * only to .rdev, so make sure to check that*/
3493 			struct md_rdev *rdev2 = rcu_dereference(
3494 				conf->disks[i].rdev);
3495 			if (rdev2 == rdev)
3496 				clear_bit(R5_Insync, &dev->flags);
3497 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3498 				s->handle_bad_blocks = 1;
3499 				atomic_inc(&rdev2->nr_pending);
3500 			} else
3501 				clear_bit(R5_WriteError, &dev->flags);
3502 		}
3503 		if (rdev && test_bit(R5_MadeGood, &dev->flags)) {
3504 			/* This flag does not apply to '.replacement'
3505 			 * only to .rdev, so make sure to check that*/
3506 			struct md_rdev *rdev2 = rcu_dereference(
3507 				conf->disks[i].rdev);
3508 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3509 				s->handle_bad_blocks = 1;
3510 				atomic_inc(&rdev2->nr_pending);
3511 			} else
3512 				clear_bit(R5_MadeGood, &dev->flags);
3513 		}
3514 		if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
3515 			struct md_rdev *rdev2 = rcu_dereference(
3516 				conf->disks[i].replacement);
3517 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
3518 				s->handle_bad_blocks = 1;
3519 				atomic_inc(&rdev2->nr_pending);
3520 			} else
3521 				clear_bit(R5_MadeGoodRepl, &dev->flags);
3522 		}
3523 		if (!test_bit(R5_Insync, &dev->flags)) {
3524 			/* The ReadError flag will just be confusing now */
3525 			clear_bit(R5_ReadError, &dev->flags);
3526 			clear_bit(R5_ReWrite, &dev->flags);
3527 		}
3528 		if (test_bit(R5_ReadError, &dev->flags))
3529 			clear_bit(R5_Insync, &dev->flags);
3530 		if (!test_bit(R5_Insync, &dev->flags)) {
3531 			if (s->failed < 2)
3532 				s->failed_num[s->failed] = i;
3533 			s->failed++;
3534 			if (rdev && !test_bit(Faulty, &rdev->flags))
3535 				do_recovery = 1;
3536 		}
3537 	}
3538 	if (test_bit(STRIPE_SYNCING, &sh->state)) {
3539 		/* If there is a failed device being replaced,
3540 		 *     we must be recovering.
3541 		 * else if we are after recovery_cp, we must be syncing
3542 		 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
3543 		 * else we can only be replacing
3544 		 * sync and recovery both need to read all devices, and so
3545 		 * use the same flag.
3546 		 */
3547 		if (do_recovery ||
3548 		    sh->sector >= conf->mddev->recovery_cp ||
3549 		    test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
3550 			s->syncing = 1;
3551 		else
3552 			s->replacing = 1;
3553 	}
3554 	rcu_read_unlock();
3555 }
3556 
3557 static void handle_stripe(struct stripe_head *sh)
3558 {
3559 	struct stripe_head_state s;
3560 	struct r5conf *conf = sh->raid_conf;
3561 	int i;
3562 	int prexor;
3563 	int disks = sh->disks;
3564 	struct r5dev *pdev, *qdev;
3565 
3566 	clear_bit(STRIPE_HANDLE, &sh->state);
3567 	if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3568 		/* already being handled, ensure it gets handled
3569 		 * again when current action finishes */
3570 		set_bit(STRIPE_HANDLE, &sh->state);
3571 		return;
3572 	}
3573 
3574 	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3575 		spin_lock(&sh->stripe_lock);
3576 		/* Cannot process 'sync' concurrently with 'discard' */
3577 		if (!test_bit(STRIPE_DISCARD, &sh->state) &&
3578 		    test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3579 			set_bit(STRIPE_SYNCING, &sh->state);
3580 			clear_bit(STRIPE_INSYNC, &sh->state);
3581 			clear_bit(STRIPE_REPLACED, &sh->state);
3582 		}
3583 		spin_unlock(&sh->stripe_lock);
3584 	}
3585 	clear_bit(STRIPE_DELAYED, &sh->state);
3586 
3587 	pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3588 		"pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3589 	       (unsigned long long)sh->sector, sh->state,
3590 	       atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3591 	       sh->check_state, sh->reconstruct_state);
3592 
3593 	analyse_stripe(sh, &s);
3594 
3595 	if (s.handle_bad_blocks) {
3596 		set_bit(STRIPE_HANDLE, &sh->state);
3597 		goto finish;
3598 	}
3599 
3600 	if (unlikely(s.blocked_rdev)) {
3601 		if (s.syncing || s.expanding || s.expanded ||
3602 		    s.replacing || s.to_write || s.written) {
3603 			set_bit(STRIPE_HANDLE, &sh->state);
3604 			goto finish;
3605 		}
3606 		/* There is nothing for the blocked_rdev to block */
3607 		rdev_dec_pending(s.blocked_rdev, conf->mddev);
3608 		s.blocked_rdev = NULL;
3609 	}
3610 
3611 	if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3612 		set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3613 		set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3614 	}
3615 
3616 	pr_debug("locked=%d uptodate=%d to_read=%d"
3617 	       " to_write=%d failed=%d failed_num=%d,%d\n",
3618 	       s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3619 	       s.failed_num[0], s.failed_num[1]);
3620 	/* check if the array has lost more than max_degraded devices and,
3621 	 * if so, some requests might need to be failed.
3622 	 */
3623 	if (s.failed > conf->max_degraded) {
3624 		sh->check_state = 0;
3625 		sh->reconstruct_state = 0;
3626 		if (s.to_read+s.to_write+s.written)
3627 			handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3628 		if (s.syncing + s.replacing)
3629 			handle_failed_sync(conf, sh, &s);
3630 	}
3631 
3632 	/* Now we check to see if any write operations have recently
3633 	 * completed
3634 	 */
3635 	prexor = 0;
3636 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3637 		prexor = 1;
3638 	if (sh->reconstruct_state == reconstruct_state_drain_result ||
3639 	    sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3640 		sh->reconstruct_state = reconstruct_state_idle;
3641 
3642 		/* All the 'written' buffers and the parity block are ready to
3643 		 * be written back to disk
3644 		 */
3645 		BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
3646 		       !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
3647 		BUG_ON(sh->qd_idx >= 0 &&
3648 		       !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
3649 		       !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
3650 		for (i = disks; i--; ) {
3651 			struct r5dev *dev = &sh->dev[i];
3652 			if (test_bit(R5_LOCKED, &dev->flags) &&
3653 				(i == sh->pd_idx || i == sh->qd_idx ||
3654 				 dev->written)) {
3655 				pr_debug("Writing block %d\n", i);
3656 				set_bit(R5_Wantwrite, &dev->flags);
3657 				if (prexor)
3658 					continue;
3659 				if (!test_bit(R5_Insync, &dev->flags) ||
3660 				    ((i == sh->pd_idx || i == sh->qd_idx)  &&
3661 				     s.failed == 0))
3662 					set_bit(STRIPE_INSYNC, &sh->state);
3663 			}
3664 		}
3665 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3666 			s.dec_preread_active = 1;
3667 	}
3668 
3669 	/*
3670 	 * might be able to return some write requests if the parity blocks
3671 	 * are safe, or on a failed drive
3672 	 */
3673 	pdev = &sh->dev[sh->pd_idx];
3674 	s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3675 		|| (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3676 	qdev = &sh->dev[sh->qd_idx];
3677 	s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3678 		|| (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3679 		|| conf->level < 6;
3680 
3681 	if (s.written &&
3682 	    (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3683 			     && !test_bit(R5_LOCKED, &pdev->flags)
3684 			     && (test_bit(R5_UPTODATE, &pdev->flags) ||
3685 				 test_bit(R5_Discard, &pdev->flags))))) &&
3686 	    (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3687 			     && !test_bit(R5_LOCKED, &qdev->flags)
3688 			     && (test_bit(R5_UPTODATE, &qdev->flags) ||
3689 				 test_bit(R5_Discard, &qdev->flags))))))
3690 		handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3691 
3692 	/* Now we might consider reading some blocks, either to check/generate
3693 	 * parity, or to satisfy requests
3694 	 * or to load a block that is being partially written.
3695 	 */
3696 	if (s.to_read || s.non_overwrite
3697 	    || (conf->level == 6 && s.to_write && s.failed)
3698 	    || (s.syncing && (s.uptodate + s.compute < disks))
3699 	    || s.replacing
3700 	    || s.expanding)
3701 		handle_stripe_fill(sh, &s, disks);
3702 
3703 	/* Now to consider new write requests and what else, if anything
3704 	 * should be read.  We do not handle new writes when:
3705 	 * 1/ A 'write' operation (copy+xor) is already in flight.
3706 	 * 2/ A 'check' operation is in flight, as it may clobber the parity
3707 	 *    block.
3708 	 */
3709 	if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3710 		handle_stripe_dirtying(conf, sh, &s, disks);
3711 
3712 	/* maybe we need to check and possibly fix the parity for this stripe
3713 	 * Any reads will already have been scheduled, so we just see if enough
3714 	 * data is available.  The parity check is held off while parity
3715 	 * dependent operations are in flight.
3716 	 */
3717 	if (sh->check_state ||
3718 	    (s.syncing && s.locked == 0 &&
3719 	     !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3720 	     !test_bit(STRIPE_INSYNC, &sh->state))) {
3721 		if (conf->level == 6)
3722 			handle_parity_checks6(conf, sh, &s, disks);
3723 		else
3724 			handle_parity_checks5(conf, sh, &s, disks);
3725 	}
3726 
3727 	if ((s.replacing || s.syncing) && s.locked == 0
3728 	    && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
3729 	    && !test_bit(STRIPE_REPLACED, &sh->state)) {
3730 		/* Write out to replacement devices where possible */
3731 		for (i = 0; i < conf->raid_disks; i++)
3732 			if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
3733 				WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
3734 				set_bit(R5_WantReplace, &sh->dev[i].flags);
3735 				set_bit(R5_LOCKED, &sh->dev[i].flags);
3736 				s.locked++;
3737 			}
3738 		if (s.replacing)
3739 			set_bit(STRIPE_INSYNC, &sh->state);
3740 		set_bit(STRIPE_REPLACED, &sh->state);
3741 	}
3742 	if ((s.syncing || s.replacing) && s.locked == 0 &&
3743 	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3744 	    test_bit(STRIPE_INSYNC, &sh->state)) {
3745 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3746 		clear_bit(STRIPE_SYNCING, &sh->state);
3747 		if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3748 			wake_up(&conf->wait_for_overlap);
3749 	}
3750 
3751 	/* If the failed drives are just a ReadError, then we might need
3752 	 * to progress the repair/check process
3753 	 */
3754 	if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3755 		for (i = 0; i < s.failed; i++) {
3756 			struct r5dev *dev = &sh->dev[s.failed_num[i]];
3757 			if (test_bit(R5_ReadError, &dev->flags)
3758 			    && !test_bit(R5_LOCKED, &dev->flags)
3759 			    && test_bit(R5_UPTODATE, &dev->flags)
3760 				) {
3761 				if (!test_bit(R5_ReWrite, &dev->flags)) {
3762 					set_bit(R5_Wantwrite, &dev->flags);
3763 					set_bit(R5_ReWrite, &dev->flags);
3764 					set_bit(R5_LOCKED, &dev->flags);
3765 					s.locked++;
3766 				} else {
3767 					/* let's read it back */
3768 					set_bit(R5_Wantread, &dev->flags);
3769 					set_bit(R5_LOCKED, &dev->flags);
3770 					s.locked++;
3771 				}
3772 			}
3773 		}
3774 
3775 
3776 	/* Finish reconstruct operations initiated by the expansion process */
3777 	if (sh->reconstruct_state == reconstruct_state_result) {
3778 		struct stripe_head *sh_src
3779 			= get_active_stripe(conf, sh->sector, 1, 1, 1);
3780 		if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3781 			/* sh cannot be written until sh_src has been read.
3782 			 * so arrange for sh to be delayed a little
3783 			 */
3784 			set_bit(STRIPE_DELAYED, &sh->state);
3785 			set_bit(STRIPE_HANDLE, &sh->state);
3786 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3787 					      &sh_src->state))
3788 				atomic_inc(&conf->preread_active_stripes);
3789 			release_stripe(sh_src);
3790 			goto finish;
3791 		}
3792 		if (sh_src)
3793 			release_stripe(sh_src);
3794 
3795 		sh->reconstruct_state = reconstruct_state_idle;
3796 		clear_bit(STRIPE_EXPANDING, &sh->state);
3797 		for (i = conf->raid_disks; i--; ) {
3798 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
3799 			set_bit(R5_LOCKED, &sh->dev[i].flags);
3800 			s.locked++;
3801 		}
3802 	}
3803 
3804 	if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3805 	    !sh->reconstruct_state) {
3806 		/* Need to write out all blocks after computing parity */
3807 		sh->disks = conf->raid_disks;
3808 		stripe_set_idx(sh->sector, conf, 0, sh);
3809 		schedule_reconstruction(sh, &s, 1, 1);
3810 	} else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3811 		clear_bit(STRIPE_EXPAND_READY, &sh->state);
3812 		atomic_dec(&conf->reshape_stripes);
3813 		wake_up(&conf->wait_for_overlap);
3814 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3815 	}
3816 
3817 	if (s.expanding && s.locked == 0 &&
3818 	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3819 		handle_stripe_expansion(conf, sh);
3820 
3821 finish:
3822 	/* wait for this device to become unblocked */
3823 	if (unlikely(s.blocked_rdev)) {
3824 		if (conf->mddev->external)
3825 			md_wait_for_blocked_rdev(s.blocked_rdev,
3826 						 conf->mddev);
3827 		else
3828 			/* Internal metadata will immediately
3829 			 * be written by raid5d, so we don't
3830 			 * need to wait here.
3831 			 */
3832 			rdev_dec_pending(s.blocked_rdev,
3833 					 conf->mddev);
3834 	}
3835 
3836 	if (s.handle_bad_blocks)
3837 		for (i = disks; i--; ) {
3838 			struct md_rdev *rdev;
3839 			struct r5dev *dev = &sh->dev[i];
3840 			if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3841 				/* We own a safe reference to the rdev */
3842 				rdev = conf->disks[i].rdev;
3843 				if (!rdev_set_badblocks(rdev, sh->sector,
3844 							STRIPE_SECTORS, 0))
3845 					md_error(conf->mddev, rdev);
3846 				rdev_dec_pending(rdev, conf->mddev);
3847 			}
3848 			if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3849 				rdev = conf->disks[i].rdev;
3850 				rdev_clear_badblocks(rdev, sh->sector,
3851 						     STRIPE_SECTORS, 0);
3852 				rdev_dec_pending(rdev, conf->mddev);
3853 			}
3854 			if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
3855 				rdev = conf->disks[i].replacement;
3856 				if (!rdev)
3857 					/* rdev have been moved down */
3858 					rdev = conf->disks[i].rdev;
3859 				rdev_clear_badblocks(rdev, sh->sector,
3860 						     STRIPE_SECTORS, 0);
3861 				rdev_dec_pending(rdev, conf->mddev);
3862 			}
3863 		}
3864 
3865 	if (s.ops_request)
3866 		raid_run_ops(sh, s.ops_request);
3867 
3868 	ops_run_io(sh, &s);
3869 
3870 	if (s.dec_preread_active) {
3871 		/* We delay this until after ops_run_io so that if make_request
3872 		 * is waiting on a flush, it won't continue until the writes
3873 		 * have actually been submitted.
3874 		 */
3875 		atomic_dec(&conf->preread_active_stripes);
3876 		if (atomic_read(&conf->preread_active_stripes) <
3877 		    IO_THRESHOLD)
3878 			md_wakeup_thread(conf->mddev->thread);
3879 	}
3880 
3881 	return_io(s.return_bi);
3882 
3883 	clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3884 }
3885 
3886 static void raid5_activate_delayed(struct r5conf *conf)
3887 {
3888 	if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3889 		while (!list_empty(&conf->delayed_list)) {
3890 			struct list_head *l = conf->delayed_list.next;
3891 			struct stripe_head *sh;
3892 			sh = list_entry(l, struct stripe_head, lru);
3893 			list_del_init(l);
3894 			clear_bit(STRIPE_DELAYED, &sh->state);
3895 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3896 				atomic_inc(&conf->preread_active_stripes);
3897 			list_add_tail(&sh->lru, &conf->hold_list);
3898 			raid5_wakeup_stripe_thread(sh);
3899 		}
3900 	}
3901 }
3902 
3903 static void activate_bit_delay(struct r5conf *conf)
3904 {
3905 	/* device_lock is held */
3906 	struct list_head head;
3907 	list_add(&head, &conf->bitmap_list);
3908 	list_del_init(&conf->bitmap_list);
3909 	while (!list_empty(&head)) {
3910 		struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3911 		list_del_init(&sh->lru);
3912 		atomic_inc(&sh->count);
3913 		__release_stripe(conf, sh);
3914 	}
3915 }
3916 
3917 int md_raid5_congested(struct mddev *mddev, int bits)
3918 {
3919 	struct r5conf *conf = mddev->private;
3920 
3921 	/* No difference between reads and writes.  Just check
3922 	 * how busy the stripe_cache is
3923 	 */
3924 
3925 	if (conf->inactive_blocked)
3926 		return 1;
3927 	if (conf->quiesce)
3928 		return 1;
3929 	if (list_empty_careful(&conf->inactive_list))
3930 		return 1;
3931 
3932 	return 0;
3933 }
3934 EXPORT_SYMBOL_GPL(md_raid5_congested);
3935 
3936 static int raid5_congested(void *data, int bits)
3937 {
3938 	struct mddev *mddev = data;
3939 
3940 	return mddev_congested(mddev, bits) ||
3941 		md_raid5_congested(mddev, bits);
3942 }
3943 
3944 /* We want read requests to align with chunks where possible,
3945  * but write requests don't need to.
3946  */
3947 static int raid5_mergeable_bvec(struct request_queue *q,
3948 				struct bvec_merge_data *bvm,
3949 				struct bio_vec *biovec)
3950 {
3951 	struct mddev *mddev = q->queuedata;
3952 	sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3953 	int max;
3954 	unsigned int chunk_sectors = mddev->chunk_sectors;
3955 	unsigned int bio_sectors = bvm->bi_size >> 9;
3956 
3957 	if ((bvm->bi_rw & 1) == WRITE)
3958 		return biovec->bv_len; /* always allow writes to be mergeable */
3959 
3960 	if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3961 		chunk_sectors = mddev->new_chunk_sectors;
3962 	max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3963 	if (max < 0) max = 0;
3964 	if (max <= biovec->bv_len && bio_sectors == 0)
3965 		return biovec->bv_len;
3966 	else
3967 		return max;
3968 }
3969 
3970 
3971 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3972 {
3973 	sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3974 	unsigned int chunk_sectors = mddev->chunk_sectors;
3975 	unsigned int bio_sectors = bio_sectors(bio);
3976 
3977 	if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3978 		chunk_sectors = mddev->new_chunk_sectors;
3979 	return  chunk_sectors >=
3980 		((sector & (chunk_sectors - 1)) + bio_sectors);
3981 }
3982 
3983 /*
3984  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3985  *  later sampled by raid5d.
3986  */
3987 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3988 {
3989 	unsigned long flags;
3990 
3991 	spin_lock_irqsave(&conf->device_lock, flags);
3992 
3993 	bi->bi_next = conf->retry_read_aligned_list;
3994 	conf->retry_read_aligned_list = bi;
3995 
3996 	spin_unlock_irqrestore(&conf->device_lock, flags);
3997 	md_wakeup_thread(conf->mddev->thread);
3998 }
3999 
4000 
4001 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4002 {
4003 	struct bio *bi;
4004 
4005 	bi = conf->retry_read_aligned;
4006 	if (bi) {
4007 		conf->retry_read_aligned = NULL;
4008 		return bi;
4009 	}
4010 	bi = conf->retry_read_aligned_list;
4011 	if(bi) {
4012 		conf->retry_read_aligned_list = bi->bi_next;
4013 		bi->bi_next = NULL;
4014 		/*
4015 		 * this sets the active strip count to 1 and the processed
4016 		 * strip count to zero (upper 8 bits)
4017 		 */
4018 		raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4019 	}
4020 
4021 	return bi;
4022 }
4023 
4024 
4025 /*
4026  *  The "raid5_align_endio" should check if the read succeeded and if it
4027  *  did, call bio_endio on the original bio (having bio_put the new bio
4028  *  first).
4029  *  If the read failed..
4030  */
4031 static void raid5_align_endio(struct bio *bi, int error)
4032 {
4033 	struct bio* raid_bi  = bi->bi_private;
4034 	struct mddev *mddev;
4035 	struct r5conf *conf;
4036 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4037 	struct md_rdev *rdev;
4038 
4039 	bio_put(bi);
4040 
4041 	rdev = (void*)raid_bi->bi_next;
4042 	raid_bi->bi_next = NULL;
4043 	mddev = rdev->mddev;
4044 	conf = mddev->private;
4045 
4046 	rdev_dec_pending(rdev, conf->mddev);
4047 
4048 	if (!error && uptodate) {
4049 		trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4050 					 raid_bi, 0);
4051 		bio_endio(raid_bi, 0);
4052 		if (atomic_dec_and_test(&conf->active_aligned_reads))
4053 			wake_up(&conf->wait_for_stripe);
4054 		return;
4055 	}
4056 
4057 
4058 	pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4059 
4060 	add_bio_to_retry(raid_bi, conf);
4061 }
4062 
4063 static int bio_fits_rdev(struct bio *bi)
4064 {
4065 	struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4066 
4067 	if (bio_sectors(bi) > queue_max_sectors(q))
4068 		return 0;
4069 	blk_recount_segments(q, bi);
4070 	if (bi->bi_phys_segments > queue_max_segments(q))
4071 		return 0;
4072 
4073 	if (q->merge_bvec_fn)
4074 		/* it's too hard to apply the merge_bvec_fn at this stage,
4075 		 * just just give up
4076 		 */
4077 		return 0;
4078 
4079 	return 1;
4080 }
4081 
4082 
4083 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4084 {
4085 	struct r5conf *conf = mddev->private;
4086 	int dd_idx;
4087 	struct bio* align_bi;
4088 	struct md_rdev *rdev;
4089 	sector_t end_sector;
4090 
4091 	if (!in_chunk_boundary(mddev, raid_bio)) {
4092 		pr_debug("chunk_aligned_read : non aligned\n");
4093 		return 0;
4094 	}
4095 	/*
4096 	 * use bio_clone_mddev to make a copy of the bio
4097 	 */
4098 	align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4099 	if (!align_bi)
4100 		return 0;
4101 	/*
4102 	 *   set bi_end_io to a new function, and set bi_private to the
4103 	 *     original bio.
4104 	 */
4105 	align_bi->bi_end_io  = raid5_align_endio;
4106 	align_bi->bi_private = raid_bio;
4107 	/*
4108 	 *	compute position
4109 	 */
4110 	align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
4111 						    0,
4112 						    &dd_idx, NULL);
4113 
4114 	end_sector = bio_end_sector(align_bi);
4115 	rcu_read_lock();
4116 	rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4117 	if (!rdev || test_bit(Faulty, &rdev->flags) ||
4118 	    rdev->recovery_offset < end_sector) {
4119 		rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4120 		if (rdev &&
4121 		    (test_bit(Faulty, &rdev->flags) ||
4122 		    !(test_bit(In_sync, &rdev->flags) ||
4123 		      rdev->recovery_offset >= end_sector)))
4124 			rdev = NULL;
4125 	}
4126 	if (rdev) {
4127 		sector_t first_bad;
4128 		int bad_sectors;
4129 
4130 		atomic_inc(&rdev->nr_pending);
4131 		rcu_read_unlock();
4132 		raid_bio->bi_next = (void*)rdev;
4133 		align_bi->bi_bdev =  rdev->bdev;
4134 		align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
4135 
4136 		if (!bio_fits_rdev(align_bi) ||
4137 		    is_badblock(rdev, align_bi->bi_sector, bio_sectors(align_bi),
4138 				&first_bad, &bad_sectors)) {
4139 			/* too big in some way, or has a known bad block */
4140 			bio_put(align_bi);
4141 			rdev_dec_pending(rdev, mddev);
4142 			return 0;
4143 		}
4144 
4145 		/* No reshape active, so we can trust rdev->data_offset */
4146 		align_bi->bi_sector += rdev->data_offset;
4147 
4148 		spin_lock_irq(&conf->device_lock);
4149 		wait_event_lock_irq(conf->wait_for_stripe,
4150 				    conf->quiesce == 0,
4151 				    conf->device_lock);
4152 		atomic_inc(&conf->active_aligned_reads);
4153 		spin_unlock_irq(&conf->device_lock);
4154 
4155 		if (mddev->gendisk)
4156 			trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4157 					      align_bi, disk_devt(mddev->gendisk),
4158 					      raid_bio->bi_sector);
4159 		generic_make_request(align_bi);
4160 		return 1;
4161 	} else {
4162 		rcu_read_unlock();
4163 		bio_put(align_bi);
4164 		return 0;
4165 	}
4166 }
4167 
4168 /* __get_priority_stripe - get the next stripe to process
4169  *
4170  * Full stripe writes are allowed to pass preread active stripes up until
4171  * the bypass_threshold is exceeded.  In general the bypass_count
4172  * increments when the handle_list is handled before the hold_list; however, it
4173  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4174  * stripe with in flight i/o.  The bypass_count will be reset when the
4175  * head of the hold_list has changed, i.e. the head was promoted to the
4176  * handle_list.
4177  */
4178 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4179 {
4180 	struct stripe_head *sh = NULL, *tmp;
4181 	struct list_head *handle_list = NULL;
4182 	struct r5worker_group *wg = NULL;
4183 
4184 	if (conf->worker_cnt_per_group == 0) {
4185 		handle_list = &conf->handle_list;
4186 	} else if (group != ANY_GROUP) {
4187 		handle_list = &conf->worker_groups[group].handle_list;
4188 		wg = &conf->worker_groups[group];
4189 	} else {
4190 		int i;
4191 		for (i = 0; i < conf->group_cnt; i++) {
4192 			handle_list = &conf->worker_groups[i].handle_list;
4193 			wg = &conf->worker_groups[i];
4194 			if (!list_empty(handle_list))
4195 				break;
4196 		}
4197 	}
4198 
4199 	pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4200 		  __func__,
4201 		  list_empty(handle_list) ? "empty" : "busy",
4202 		  list_empty(&conf->hold_list) ? "empty" : "busy",
4203 		  atomic_read(&conf->pending_full_writes), conf->bypass_count);
4204 
4205 	if (!list_empty(handle_list)) {
4206 		sh = list_entry(handle_list->next, typeof(*sh), lru);
4207 
4208 		if (list_empty(&conf->hold_list))
4209 			conf->bypass_count = 0;
4210 		else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4211 			if (conf->hold_list.next == conf->last_hold)
4212 				conf->bypass_count++;
4213 			else {
4214 				conf->last_hold = conf->hold_list.next;
4215 				conf->bypass_count -= conf->bypass_threshold;
4216 				if (conf->bypass_count < 0)
4217 					conf->bypass_count = 0;
4218 			}
4219 		}
4220 	} else if (!list_empty(&conf->hold_list) &&
4221 		   ((conf->bypass_threshold &&
4222 		     conf->bypass_count > conf->bypass_threshold) ||
4223 		    atomic_read(&conf->pending_full_writes) == 0)) {
4224 
4225 		list_for_each_entry(tmp, &conf->hold_list,  lru) {
4226 			if (conf->worker_cnt_per_group == 0 ||
4227 			    group == ANY_GROUP ||
4228 			    !cpu_online(tmp->cpu) ||
4229 			    cpu_to_group(tmp->cpu) == group) {
4230 				sh = tmp;
4231 				break;
4232 			}
4233 		}
4234 
4235 		if (sh) {
4236 			conf->bypass_count -= conf->bypass_threshold;
4237 			if (conf->bypass_count < 0)
4238 				conf->bypass_count = 0;
4239 		}
4240 		wg = NULL;
4241 	}
4242 
4243 	if (!sh)
4244 		return NULL;
4245 
4246 	if (wg) {
4247 		wg->stripes_cnt--;
4248 		sh->group = NULL;
4249 	}
4250 	list_del_init(&sh->lru);
4251 	atomic_inc(&sh->count);
4252 	BUG_ON(atomic_read(&sh->count) != 1);
4253 	return sh;
4254 }
4255 
4256 struct raid5_plug_cb {
4257 	struct blk_plug_cb	cb;
4258 	struct list_head	list;
4259 };
4260 
4261 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4262 {
4263 	struct raid5_plug_cb *cb = container_of(
4264 		blk_cb, struct raid5_plug_cb, cb);
4265 	struct stripe_head *sh;
4266 	struct mddev *mddev = cb->cb.data;
4267 	struct r5conf *conf = mddev->private;
4268 	int cnt = 0;
4269 
4270 	if (cb->list.next && !list_empty(&cb->list)) {
4271 		spin_lock_irq(&conf->device_lock);
4272 		while (!list_empty(&cb->list)) {
4273 			sh = list_first_entry(&cb->list, struct stripe_head, lru);
4274 			list_del_init(&sh->lru);
4275 			/*
4276 			 * avoid race release_stripe_plug() sees
4277 			 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4278 			 * is still in our list
4279 			 */
4280 			smp_mb__before_clear_bit();
4281 			clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4282 			/*
4283 			 * STRIPE_ON_RELEASE_LIST could be set here. In that
4284 			 * case, the count is always > 1 here
4285 			 */
4286 			__release_stripe(conf, sh);
4287 			cnt++;
4288 		}
4289 		spin_unlock_irq(&conf->device_lock);
4290 	}
4291 	if (mddev->queue)
4292 		trace_block_unplug(mddev->queue, cnt, !from_schedule);
4293 	kfree(cb);
4294 }
4295 
4296 static void release_stripe_plug(struct mddev *mddev,
4297 				struct stripe_head *sh)
4298 {
4299 	struct blk_plug_cb *blk_cb = blk_check_plugged(
4300 		raid5_unplug, mddev,
4301 		sizeof(struct raid5_plug_cb));
4302 	struct raid5_plug_cb *cb;
4303 
4304 	if (!blk_cb) {
4305 		release_stripe(sh);
4306 		return;
4307 	}
4308 
4309 	cb = container_of(blk_cb, struct raid5_plug_cb, cb);
4310 
4311 	if (cb->list.next == NULL)
4312 		INIT_LIST_HEAD(&cb->list);
4313 
4314 	if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
4315 		list_add_tail(&sh->lru, &cb->list);
4316 	else
4317 		release_stripe(sh);
4318 }
4319 
4320 static void make_discard_request(struct mddev *mddev, struct bio *bi)
4321 {
4322 	struct r5conf *conf = mddev->private;
4323 	sector_t logical_sector, last_sector;
4324 	struct stripe_head *sh;
4325 	int remaining;
4326 	int stripe_sectors;
4327 
4328 	if (mddev->reshape_position != MaxSector)
4329 		/* Skip discard while reshape is happening */
4330 		return;
4331 
4332 	logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4333 	last_sector = bi->bi_sector + (bi->bi_size>>9);
4334 
4335 	bi->bi_next = NULL;
4336 	bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
4337 
4338 	stripe_sectors = conf->chunk_sectors *
4339 		(conf->raid_disks - conf->max_degraded);
4340 	logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
4341 					       stripe_sectors);
4342 	sector_div(last_sector, stripe_sectors);
4343 
4344 	logical_sector *= conf->chunk_sectors;
4345 	last_sector *= conf->chunk_sectors;
4346 
4347 	for (; logical_sector < last_sector;
4348 	     logical_sector += STRIPE_SECTORS) {
4349 		DEFINE_WAIT(w);
4350 		int d;
4351 	again:
4352 		sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
4353 		prepare_to_wait(&conf->wait_for_overlap, &w,
4354 				TASK_UNINTERRUPTIBLE);
4355 		set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4356 		if (test_bit(STRIPE_SYNCING, &sh->state)) {
4357 			release_stripe(sh);
4358 			schedule();
4359 			goto again;
4360 		}
4361 		clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
4362 		spin_lock_irq(&sh->stripe_lock);
4363 		for (d = 0; d < conf->raid_disks; d++) {
4364 			if (d == sh->pd_idx || d == sh->qd_idx)
4365 				continue;
4366 			if (sh->dev[d].towrite || sh->dev[d].toread) {
4367 				set_bit(R5_Overlap, &sh->dev[d].flags);
4368 				spin_unlock_irq(&sh->stripe_lock);
4369 				release_stripe(sh);
4370 				schedule();
4371 				goto again;
4372 			}
4373 		}
4374 		set_bit(STRIPE_DISCARD, &sh->state);
4375 		finish_wait(&conf->wait_for_overlap, &w);
4376 		for (d = 0; d < conf->raid_disks; d++) {
4377 			if (d == sh->pd_idx || d == sh->qd_idx)
4378 				continue;
4379 			sh->dev[d].towrite = bi;
4380 			set_bit(R5_OVERWRITE, &sh->dev[d].flags);
4381 			raid5_inc_bi_active_stripes(bi);
4382 		}
4383 		spin_unlock_irq(&sh->stripe_lock);
4384 		if (conf->mddev->bitmap) {
4385 			for (d = 0;
4386 			     d < conf->raid_disks - conf->max_degraded;
4387 			     d++)
4388 				bitmap_startwrite(mddev->bitmap,
4389 						  sh->sector,
4390 						  STRIPE_SECTORS,
4391 						  0);
4392 			sh->bm_seq = conf->seq_flush + 1;
4393 			set_bit(STRIPE_BIT_DELAY, &sh->state);
4394 		}
4395 
4396 		set_bit(STRIPE_HANDLE, &sh->state);
4397 		clear_bit(STRIPE_DELAYED, &sh->state);
4398 		if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4399 			atomic_inc(&conf->preread_active_stripes);
4400 		release_stripe_plug(mddev, sh);
4401 	}
4402 
4403 	remaining = raid5_dec_bi_active_stripes(bi);
4404 	if (remaining == 0) {
4405 		md_write_end(mddev);
4406 		bio_endio(bi, 0);
4407 	}
4408 }
4409 
4410 static void make_request(struct mddev *mddev, struct bio * bi)
4411 {
4412 	struct r5conf *conf = mddev->private;
4413 	int dd_idx;
4414 	sector_t new_sector;
4415 	sector_t logical_sector, last_sector;
4416 	struct stripe_head *sh;
4417 	const int rw = bio_data_dir(bi);
4418 	int remaining;
4419 
4420 	if (unlikely(bi->bi_rw & REQ_FLUSH)) {
4421 		md_flush_request(mddev, bi);
4422 		return;
4423 	}
4424 
4425 	md_write_start(mddev, bi);
4426 
4427 	if (rw == READ &&
4428 	     mddev->reshape_position == MaxSector &&
4429 	     chunk_aligned_read(mddev,bi))
4430 		return;
4431 
4432 	if (unlikely(bi->bi_rw & REQ_DISCARD)) {
4433 		make_discard_request(mddev, bi);
4434 		return;
4435 	}
4436 
4437 	logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4438 	last_sector = bio_end_sector(bi);
4439 	bi->bi_next = NULL;
4440 	bi->bi_phys_segments = 1;	/* over-loaded to count active stripes */
4441 
4442 	for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
4443 		DEFINE_WAIT(w);
4444 		int previous;
4445 		int seq;
4446 
4447 	retry:
4448 		seq = read_seqcount_begin(&conf->gen_lock);
4449 		previous = 0;
4450 		prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
4451 		if (unlikely(conf->reshape_progress != MaxSector)) {
4452 			/* spinlock is needed as reshape_progress may be
4453 			 * 64bit on a 32bit platform, and so it might be
4454 			 * possible to see a half-updated value
4455 			 * Of course reshape_progress could change after
4456 			 * the lock is dropped, so once we get a reference
4457 			 * to the stripe that we think it is, we will have
4458 			 * to check again.
4459 			 */
4460 			spin_lock_irq(&conf->device_lock);
4461 			if (mddev->reshape_backwards
4462 			    ? logical_sector < conf->reshape_progress
4463 			    : logical_sector >= conf->reshape_progress) {
4464 				previous = 1;
4465 			} else {
4466 				if (mddev->reshape_backwards
4467 				    ? logical_sector < conf->reshape_safe
4468 				    : logical_sector >= conf->reshape_safe) {
4469 					spin_unlock_irq(&conf->device_lock);
4470 					schedule();
4471 					goto retry;
4472 				}
4473 			}
4474 			spin_unlock_irq(&conf->device_lock);
4475 		}
4476 
4477 		new_sector = raid5_compute_sector(conf, logical_sector,
4478 						  previous,
4479 						  &dd_idx, NULL);
4480 		pr_debug("raid456: make_request, sector %llu logical %llu\n",
4481 			(unsigned long long)new_sector,
4482 			(unsigned long long)logical_sector);
4483 
4484 		sh = get_active_stripe(conf, new_sector, previous,
4485 				       (bi->bi_rw&RWA_MASK), 0);
4486 		if (sh) {
4487 			if (unlikely(previous)) {
4488 				/* expansion might have moved on while waiting for a
4489 				 * stripe, so we must do the range check again.
4490 				 * Expansion could still move past after this
4491 				 * test, but as we are holding a reference to
4492 				 * 'sh', we know that if that happens,
4493 				 *  STRIPE_EXPANDING will get set and the expansion
4494 				 * won't proceed until we finish with the stripe.
4495 				 */
4496 				int must_retry = 0;
4497 				spin_lock_irq(&conf->device_lock);
4498 				if (mddev->reshape_backwards
4499 				    ? logical_sector >= conf->reshape_progress
4500 				    : logical_sector < conf->reshape_progress)
4501 					/* mismatch, need to try again */
4502 					must_retry = 1;
4503 				spin_unlock_irq(&conf->device_lock);
4504 				if (must_retry) {
4505 					release_stripe(sh);
4506 					schedule();
4507 					goto retry;
4508 				}
4509 			}
4510 			if (read_seqcount_retry(&conf->gen_lock, seq)) {
4511 				/* Might have got the wrong stripe_head
4512 				 * by accident
4513 				 */
4514 				release_stripe(sh);
4515 				goto retry;
4516 			}
4517 
4518 			if (rw == WRITE &&
4519 			    logical_sector >= mddev->suspend_lo &&
4520 			    logical_sector < mddev->suspend_hi) {
4521 				release_stripe(sh);
4522 				/* As the suspend_* range is controlled by
4523 				 * userspace, we want an interruptible
4524 				 * wait.
4525 				 */
4526 				flush_signals(current);
4527 				prepare_to_wait(&conf->wait_for_overlap,
4528 						&w, TASK_INTERRUPTIBLE);
4529 				if (logical_sector >= mddev->suspend_lo &&
4530 				    logical_sector < mddev->suspend_hi)
4531 					schedule();
4532 				goto retry;
4533 			}
4534 
4535 			if (test_bit(STRIPE_EXPANDING, &sh->state) ||
4536 			    !add_stripe_bio(sh, bi, dd_idx, rw)) {
4537 				/* Stripe is busy expanding or
4538 				 * add failed due to overlap.  Flush everything
4539 				 * and wait a while
4540 				 */
4541 				md_wakeup_thread(mddev->thread);
4542 				release_stripe(sh);
4543 				schedule();
4544 				goto retry;
4545 			}
4546 			finish_wait(&conf->wait_for_overlap, &w);
4547 			set_bit(STRIPE_HANDLE, &sh->state);
4548 			clear_bit(STRIPE_DELAYED, &sh->state);
4549 			if ((bi->bi_rw & REQ_SYNC) &&
4550 			    !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4551 				atomic_inc(&conf->preread_active_stripes);
4552 			release_stripe_plug(mddev, sh);
4553 		} else {
4554 			/* cannot get stripe for read-ahead, just give-up */
4555 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
4556 			finish_wait(&conf->wait_for_overlap, &w);
4557 			break;
4558 		}
4559 	}
4560 
4561 	remaining = raid5_dec_bi_active_stripes(bi);
4562 	if (remaining == 0) {
4563 
4564 		if ( rw == WRITE )
4565 			md_write_end(mddev);
4566 
4567 		trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
4568 					 bi, 0);
4569 		bio_endio(bi, 0);
4570 	}
4571 }
4572 
4573 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
4574 
4575 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
4576 {
4577 	/* reshaping is quite different to recovery/resync so it is
4578 	 * handled quite separately ... here.
4579 	 *
4580 	 * On each call to sync_request, we gather one chunk worth of
4581 	 * destination stripes and flag them as expanding.
4582 	 * Then we find all the source stripes and request reads.
4583 	 * As the reads complete, handle_stripe will copy the data
4584 	 * into the destination stripe and release that stripe.
4585 	 */
4586 	struct r5conf *conf = mddev->private;
4587 	struct stripe_head *sh;
4588 	sector_t first_sector, last_sector;
4589 	int raid_disks = conf->previous_raid_disks;
4590 	int data_disks = raid_disks - conf->max_degraded;
4591 	int new_data_disks = conf->raid_disks - conf->max_degraded;
4592 	int i;
4593 	int dd_idx;
4594 	sector_t writepos, readpos, safepos;
4595 	sector_t stripe_addr;
4596 	int reshape_sectors;
4597 	struct list_head stripes;
4598 
4599 	if (sector_nr == 0) {
4600 		/* If restarting in the middle, skip the initial sectors */
4601 		if (mddev->reshape_backwards &&
4602 		    conf->reshape_progress < raid5_size(mddev, 0, 0)) {
4603 			sector_nr = raid5_size(mddev, 0, 0)
4604 				- conf->reshape_progress;
4605 		} else if (!mddev->reshape_backwards &&
4606 			   conf->reshape_progress > 0)
4607 			sector_nr = conf->reshape_progress;
4608 		sector_div(sector_nr, new_data_disks);
4609 		if (sector_nr) {
4610 			mddev->curr_resync_completed = sector_nr;
4611 			sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4612 			*skipped = 1;
4613 			return sector_nr;
4614 		}
4615 	}
4616 
4617 	/* We need to process a full chunk at a time.
4618 	 * If old and new chunk sizes differ, we need to process the
4619 	 * largest of these
4620 	 */
4621 	if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4622 		reshape_sectors = mddev->new_chunk_sectors;
4623 	else
4624 		reshape_sectors = mddev->chunk_sectors;
4625 
4626 	/* We update the metadata at least every 10 seconds, or when
4627 	 * the data about to be copied would over-write the source of
4628 	 * the data at the front of the range.  i.e. one new_stripe
4629 	 * along from reshape_progress new_maps to after where
4630 	 * reshape_safe old_maps to
4631 	 */
4632 	writepos = conf->reshape_progress;
4633 	sector_div(writepos, new_data_disks);
4634 	readpos = conf->reshape_progress;
4635 	sector_div(readpos, data_disks);
4636 	safepos = conf->reshape_safe;
4637 	sector_div(safepos, data_disks);
4638 	if (mddev->reshape_backwards) {
4639 		writepos -= min_t(sector_t, reshape_sectors, writepos);
4640 		readpos += reshape_sectors;
4641 		safepos += reshape_sectors;
4642 	} else {
4643 		writepos += reshape_sectors;
4644 		readpos -= min_t(sector_t, reshape_sectors, readpos);
4645 		safepos -= min_t(sector_t, reshape_sectors, safepos);
4646 	}
4647 
4648 	/* Having calculated the 'writepos' possibly use it
4649 	 * to set 'stripe_addr' which is where we will write to.
4650 	 */
4651 	if (mddev->reshape_backwards) {
4652 		BUG_ON(conf->reshape_progress == 0);
4653 		stripe_addr = writepos;
4654 		BUG_ON((mddev->dev_sectors &
4655 			~((sector_t)reshape_sectors - 1))
4656 		       - reshape_sectors - stripe_addr
4657 		       != sector_nr);
4658 	} else {
4659 		BUG_ON(writepos != sector_nr + reshape_sectors);
4660 		stripe_addr = sector_nr;
4661 	}
4662 
4663 	/* 'writepos' is the most advanced device address we might write.
4664 	 * 'readpos' is the least advanced device address we might read.
4665 	 * 'safepos' is the least address recorded in the metadata as having
4666 	 *     been reshaped.
4667 	 * If there is a min_offset_diff, these are adjusted either by
4668 	 * increasing the safepos/readpos if diff is negative, or
4669 	 * increasing writepos if diff is positive.
4670 	 * If 'readpos' is then behind 'writepos', there is no way that we can
4671 	 * ensure safety in the face of a crash - that must be done by userspace
4672 	 * making a backup of the data.  So in that case there is no particular
4673 	 * rush to update metadata.
4674 	 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4675 	 * update the metadata to advance 'safepos' to match 'readpos' so that
4676 	 * we can be safe in the event of a crash.
4677 	 * So we insist on updating metadata if safepos is behind writepos and
4678 	 * readpos is beyond writepos.
4679 	 * In any case, update the metadata every 10 seconds.
4680 	 * Maybe that number should be configurable, but I'm not sure it is
4681 	 * worth it.... maybe it could be a multiple of safemode_delay???
4682 	 */
4683 	if (conf->min_offset_diff < 0) {
4684 		safepos += -conf->min_offset_diff;
4685 		readpos += -conf->min_offset_diff;
4686 	} else
4687 		writepos += conf->min_offset_diff;
4688 
4689 	if ((mddev->reshape_backwards
4690 	     ? (safepos > writepos && readpos < writepos)
4691 	     : (safepos < writepos && readpos > writepos)) ||
4692 	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4693 		/* Cannot proceed until we've updated the superblock... */
4694 		wait_event(conf->wait_for_overlap,
4695 			   atomic_read(&conf->reshape_stripes)==0);
4696 		mddev->reshape_position = conf->reshape_progress;
4697 		mddev->curr_resync_completed = sector_nr;
4698 		conf->reshape_checkpoint = jiffies;
4699 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
4700 		md_wakeup_thread(mddev->thread);
4701 		wait_event(mddev->sb_wait, mddev->flags == 0 ||
4702 			   kthread_should_stop());
4703 		spin_lock_irq(&conf->device_lock);
4704 		conf->reshape_safe = mddev->reshape_position;
4705 		spin_unlock_irq(&conf->device_lock);
4706 		wake_up(&conf->wait_for_overlap);
4707 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4708 	}
4709 
4710 	INIT_LIST_HEAD(&stripes);
4711 	for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4712 		int j;
4713 		int skipped_disk = 0;
4714 		sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4715 		set_bit(STRIPE_EXPANDING, &sh->state);
4716 		atomic_inc(&conf->reshape_stripes);
4717 		/* If any of this stripe is beyond the end of the old
4718 		 * array, then we need to zero those blocks
4719 		 */
4720 		for (j=sh->disks; j--;) {
4721 			sector_t s;
4722 			if (j == sh->pd_idx)
4723 				continue;
4724 			if (conf->level == 6 &&
4725 			    j == sh->qd_idx)
4726 				continue;
4727 			s = compute_blocknr(sh, j, 0);
4728 			if (s < raid5_size(mddev, 0, 0)) {
4729 				skipped_disk = 1;
4730 				continue;
4731 			}
4732 			memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4733 			set_bit(R5_Expanded, &sh->dev[j].flags);
4734 			set_bit(R5_UPTODATE, &sh->dev[j].flags);
4735 		}
4736 		if (!skipped_disk) {
4737 			set_bit(STRIPE_EXPAND_READY, &sh->state);
4738 			set_bit(STRIPE_HANDLE, &sh->state);
4739 		}
4740 		list_add(&sh->lru, &stripes);
4741 	}
4742 	spin_lock_irq(&conf->device_lock);
4743 	if (mddev->reshape_backwards)
4744 		conf->reshape_progress -= reshape_sectors * new_data_disks;
4745 	else
4746 		conf->reshape_progress += reshape_sectors * new_data_disks;
4747 	spin_unlock_irq(&conf->device_lock);
4748 	/* Ok, those stripe are ready. We can start scheduling
4749 	 * reads on the source stripes.
4750 	 * The source stripes are determined by mapping the first and last
4751 	 * block on the destination stripes.
4752 	 */
4753 	first_sector =
4754 		raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4755 				     1, &dd_idx, NULL);
4756 	last_sector =
4757 		raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4758 					    * new_data_disks - 1),
4759 				     1, &dd_idx, NULL);
4760 	if (last_sector >= mddev->dev_sectors)
4761 		last_sector = mddev->dev_sectors - 1;
4762 	while (first_sector <= last_sector) {
4763 		sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4764 		set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4765 		set_bit(STRIPE_HANDLE, &sh->state);
4766 		release_stripe(sh);
4767 		first_sector += STRIPE_SECTORS;
4768 	}
4769 	/* Now that the sources are clearly marked, we can release
4770 	 * the destination stripes
4771 	 */
4772 	while (!list_empty(&stripes)) {
4773 		sh = list_entry(stripes.next, struct stripe_head, lru);
4774 		list_del_init(&sh->lru);
4775 		release_stripe(sh);
4776 	}
4777 	/* If this takes us to the resync_max point where we have to pause,
4778 	 * then we need to write out the superblock.
4779 	 */
4780 	sector_nr += reshape_sectors;
4781 	if ((sector_nr - mddev->curr_resync_completed) * 2
4782 	    >= mddev->resync_max - mddev->curr_resync_completed) {
4783 		/* Cannot proceed until we've updated the superblock... */
4784 		wait_event(conf->wait_for_overlap,
4785 			   atomic_read(&conf->reshape_stripes) == 0);
4786 		mddev->reshape_position = conf->reshape_progress;
4787 		mddev->curr_resync_completed = sector_nr;
4788 		conf->reshape_checkpoint = jiffies;
4789 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
4790 		md_wakeup_thread(mddev->thread);
4791 		wait_event(mddev->sb_wait,
4792 			   !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4793 			   || kthread_should_stop());
4794 		spin_lock_irq(&conf->device_lock);
4795 		conf->reshape_safe = mddev->reshape_position;
4796 		spin_unlock_irq(&conf->device_lock);
4797 		wake_up(&conf->wait_for_overlap);
4798 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4799 	}
4800 	return reshape_sectors;
4801 }
4802 
4803 /* FIXME go_faster isn't used */
4804 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4805 {
4806 	struct r5conf *conf = mddev->private;
4807 	struct stripe_head *sh;
4808 	sector_t max_sector = mddev->dev_sectors;
4809 	sector_t sync_blocks;
4810 	int still_degraded = 0;
4811 	int i;
4812 
4813 	if (sector_nr >= max_sector) {
4814 		/* just being told to finish up .. nothing much to do */
4815 
4816 		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4817 			end_reshape(conf);
4818 			return 0;
4819 		}
4820 
4821 		if (mddev->curr_resync < max_sector) /* aborted */
4822 			bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4823 					&sync_blocks, 1);
4824 		else /* completed sync */
4825 			conf->fullsync = 0;
4826 		bitmap_close_sync(mddev->bitmap);
4827 
4828 		return 0;
4829 	}
4830 
4831 	/* Allow raid5_quiesce to complete */
4832 	wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4833 
4834 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4835 		return reshape_request(mddev, sector_nr, skipped);
4836 
4837 	/* No need to check resync_max as we never do more than one
4838 	 * stripe, and as resync_max will always be on a chunk boundary,
4839 	 * if the check in md_do_sync didn't fire, there is no chance
4840 	 * of overstepping resync_max here
4841 	 */
4842 
4843 	/* if there is too many failed drives and we are trying
4844 	 * to resync, then assert that we are finished, because there is
4845 	 * nothing we can do.
4846 	 */
4847 	if (mddev->degraded >= conf->max_degraded &&
4848 	    test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4849 		sector_t rv = mddev->dev_sectors - sector_nr;
4850 		*skipped = 1;
4851 		return rv;
4852 	}
4853 	if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4854 	    !conf->fullsync &&
4855 	    !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4856 	    sync_blocks >= STRIPE_SECTORS) {
4857 		/* we can skip this block, and probably more */
4858 		sync_blocks /= STRIPE_SECTORS;
4859 		*skipped = 1;
4860 		return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4861 	}
4862 
4863 	bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4864 
4865 	sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4866 	if (sh == NULL) {
4867 		sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4868 		/* make sure we don't swamp the stripe cache if someone else
4869 		 * is trying to get access
4870 		 */
4871 		schedule_timeout_uninterruptible(1);
4872 	}
4873 	/* Need to check if array will still be degraded after recovery/resync
4874 	 * We don't need to check the 'failed' flag as when that gets set,
4875 	 * recovery aborts.
4876 	 */
4877 	for (i = 0; i < conf->raid_disks; i++)
4878 		if (conf->disks[i].rdev == NULL)
4879 			still_degraded = 1;
4880 
4881 	bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4882 
4883 	set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4884 
4885 	handle_stripe(sh);
4886 	release_stripe(sh);
4887 
4888 	return STRIPE_SECTORS;
4889 }
4890 
4891 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4892 {
4893 	/* We may not be able to submit a whole bio at once as there
4894 	 * may not be enough stripe_heads available.
4895 	 * We cannot pre-allocate enough stripe_heads as we may need
4896 	 * more than exist in the cache (if we allow ever large chunks).
4897 	 * So we do one stripe head at a time and record in
4898 	 * ->bi_hw_segments how many have been done.
4899 	 *
4900 	 * We *know* that this entire raid_bio is in one chunk, so
4901 	 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4902 	 */
4903 	struct stripe_head *sh;
4904 	int dd_idx;
4905 	sector_t sector, logical_sector, last_sector;
4906 	int scnt = 0;
4907 	int remaining;
4908 	int handled = 0;
4909 
4910 	logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4911 	sector = raid5_compute_sector(conf, logical_sector,
4912 				      0, &dd_idx, NULL);
4913 	last_sector = bio_end_sector(raid_bio);
4914 
4915 	for (; logical_sector < last_sector;
4916 	     logical_sector += STRIPE_SECTORS,
4917 		     sector += STRIPE_SECTORS,
4918 		     scnt++) {
4919 
4920 		if (scnt < raid5_bi_processed_stripes(raid_bio))
4921 			/* already done this stripe */
4922 			continue;
4923 
4924 		sh = get_active_stripe(conf, sector, 0, 1, 0);
4925 
4926 		if (!sh) {
4927 			/* failed to get a stripe - must wait */
4928 			raid5_set_bi_processed_stripes(raid_bio, scnt);
4929 			conf->retry_read_aligned = raid_bio;
4930 			return handled;
4931 		}
4932 
4933 		if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4934 			release_stripe(sh);
4935 			raid5_set_bi_processed_stripes(raid_bio, scnt);
4936 			conf->retry_read_aligned = raid_bio;
4937 			return handled;
4938 		}
4939 
4940 		set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
4941 		handle_stripe(sh);
4942 		release_stripe(sh);
4943 		handled++;
4944 	}
4945 	remaining = raid5_dec_bi_active_stripes(raid_bio);
4946 	if (remaining == 0) {
4947 		trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
4948 					 raid_bio, 0);
4949 		bio_endio(raid_bio, 0);
4950 	}
4951 	if (atomic_dec_and_test(&conf->active_aligned_reads))
4952 		wake_up(&conf->wait_for_stripe);
4953 	return handled;
4954 }
4955 
4956 static int handle_active_stripes(struct r5conf *conf, int group,
4957 				 struct r5worker *worker)
4958 {
4959 	struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
4960 	int i, batch_size = 0;
4961 
4962 	while (batch_size < MAX_STRIPE_BATCH &&
4963 			(sh = __get_priority_stripe(conf, group)) != NULL)
4964 		batch[batch_size++] = sh;
4965 
4966 	if (batch_size == 0)
4967 		return batch_size;
4968 	spin_unlock_irq(&conf->device_lock);
4969 
4970 	for (i = 0; i < batch_size; i++)
4971 		handle_stripe(batch[i]);
4972 
4973 	cond_resched();
4974 
4975 	spin_lock_irq(&conf->device_lock);
4976 	for (i = 0; i < batch_size; i++)
4977 		__release_stripe(conf, batch[i]);
4978 	return batch_size;
4979 }
4980 
4981 static void raid5_do_work(struct work_struct *work)
4982 {
4983 	struct r5worker *worker = container_of(work, struct r5worker, work);
4984 	struct r5worker_group *group = worker->group;
4985 	struct r5conf *conf = group->conf;
4986 	int group_id = group - conf->worker_groups;
4987 	int handled;
4988 	struct blk_plug plug;
4989 
4990 	pr_debug("+++ raid5worker active\n");
4991 
4992 	blk_start_plug(&plug);
4993 	handled = 0;
4994 	spin_lock_irq(&conf->device_lock);
4995 	while (1) {
4996 		int batch_size, released;
4997 
4998 		released = release_stripe_list(conf);
4999 
5000 		batch_size = handle_active_stripes(conf, group_id, worker);
5001 		worker->working = false;
5002 		if (!batch_size && !released)
5003 			break;
5004 		handled += batch_size;
5005 	}
5006 	pr_debug("%d stripes handled\n", handled);
5007 
5008 	spin_unlock_irq(&conf->device_lock);
5009 	blk_finish_plug(&plug);
5010 
5011 	pr_debug("--- raid5worker inactive\n");
5012 }
5013 
5014 /*
5015  * This is our raid5 kernel thread.
5016  *
5017  * We scan the hash table for stripes which can be handled now.
5018  * During the scan, completed stripes are saved for us by the interrupt
5019  * handler, so that they will not have to wait for our next wakeup.
5020  */
5021 static void raid5d(struct md_thread *thread)
5022 {
5023 	struct mddev *mddev = thread->mddev;
5024 	struct r5conf *conf = mddev->private;
5025 	int handled;
5026 	struct blk_plug plug;
5027 
5028 	pr_debug("+++ raid5d active\n");
5029 
5030 	md_check_recovery(mddev);
5031 
5032 	blk_start_plug(&plug);
5033 	handled = 0;
5034 	spin_lock_irq(&conf->device_lock);
5035 	while (1) {
5036 		struct bio *bio;
5037 		int batch_size, released;
5038 
5039 		released = release_stripe_list(conf);
5040 
5041 		if (
5042 		    !list_empty(&conf->bitmap_list)) {
5043 			/* Now is a good time to flush some bitmap updates */
5044 			conf->seq_flush++;
5045 			spin_unlock_irq(&conf->device_lock);
5046 			bitmap_unplug(mddev->bitmap);
5047 			spin_lock_irq(&conf->device_lock);
5048 			conf->seq_write = conf->seq_flush;
5049 			activate_bit_delay(conf);
5050 		}
5051 		raid5_activate_delayed(conf);
5052 
5053 		while ((bio = remove_bio_from_retry(conf))) {
5054 			int ok;
5055 			spin_unlock_irq(&conf->device_lock);
5056 			ok = retry_aligned_read(conf, bio);
5057 			spin_lock_irq(&conf->device_lock);
5058 			if (!ok)
5059 				break;
5060 			handled++;
5061 		}
5062 
5063 		batch_size = handle_active_stripes(conf, ANY_GROUP, NULL);
5064 		if (!batch_size && !released)
5065 			break;
5066 		handled += batch_size;
5067 
5068 		if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5069 			spin_unlock_irq(&conf->device_lock);
5070 			md_check_recovery(mddev);
5071 			spin_lock_irq(&conf->device_lock);
5072 		}
5073 	}
5074 	pr_debug("%d stripes handled\n", handled);
5075 
5076 	spin_unlock_irq(&conf->device_lock);
5077 
5078 	async_tx_issue_pending_all();
5079 	blk_finish_plug(&plug);
5080 
5081 	pr_debug("--- raid5d inactive\n");
5082 }
5083 
5084 static ssize_t
5085 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5086 {
5087 	struct r5conf *conf = mddev->private;
5088 	if (conf)
5089 		return sprintf(page, "%d\n", conf->max_nr_stripes);
5090 	else
5091 		return 0;
5092 }
5093 
5094 int
5095 raid5_set_cache_size(struct mddev *mddev, int size)
5096 {
5097 	struct r5conf *conf = mddev->private;
5098 	int err;
5099 
5100 	if (size <= 16 || size > 32768)
5101 		return -EINVAL;
5102 	while (size < conf->max_nr_stripes) {
5103 		if (drop_one_stripe(conf))
5104 			conf->max_nr_stripes--;
5105 		else
5106 			break;
5107 	}
5108 	err = md_allow_write(mddev);
5109 	if (err)
5110 		return err;
5111 	while (size > conf->max_nr_stripes) {
5112 		if (grow_one_stripe(conf))
5113 			conf->max_nr_stripes++;
5114 		else break;
5115 	}
5116 	return 0;
5117 }
5118 EXPORT_SYMBOL(raid5_set_cache_size);
5119 
5120 static ssize_t
5121 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5122 {
5123 	struct r5conf *conf = mddev->private;
5124 	unsigned long new;
5125 	int err;
5126 
5127 	if (len >= PAGE_SIZE)
5128 		return -EINVAL;
5129 	if (!conf)
5130 		return -ENODEV;
5131 
5132 	if (kstrtoul(page, 10, &new))
5133 		return -EINVAL;
5134 	err = raid5_set_cache_size(mddev, new);
5135 	if (err)
5136 		return err;
5137 	return len;
5138 }
5139 
5140 static struct md_sysfs_entry
5141 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5142 				raid5_show_stripe_cache_size,
5143 				raid5_store_stripe_cache_size);
5144 
5145 static ssize_t
5146 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5147 {
5148 	struct r5conf *conf = mddev->private;
5149 	if (conf)
5150 		return sprintf(page, "%d\n", conf->bypass_threshold);
5151 	else
5152 		return 0;
5153 }
5154 
5155 static ssize_t
5156 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
5157 {
5158 	struct r5conf *conf = mddev->private;
5159 	unsigned long new;
5160 	if (len >= PAGE_SIZE)
5161 		return -EINVAL;
5162 	if (!conf)
5163 		return -ENODEV;
5164 
5165 	if (kstrtoul(page, 10, &new))
5166 		return -EINVAL;
5167 	if (new > conf->max_nr_stripes)
5168 		return -EINVAL;
5169 	conf->bypass_threshold = new;
5170 	return len;
5171 }
5172 
5173 static struct md_sysfs_entry
5174 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
5175 					S_IRUGO | S_IWUSR,
5176 					raid5_show_preread_threshold,
5177 					raid5_store_preread_threshold);
5178 
5179 static ssize_t
5180 stripe_cache_active_show(struct mddev *mddev, char *page)
5181 {
5182 	struct r5conf *conf = mddev->private;
5183 	if (conf)
5184 		return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
5185 	else
5186 		return 0;
5187 }
5188 
5189 static struct md_sysfs_entry
5190 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
5191 
5192 static ssize_t
5193 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
5194 {
5195 	struct r5conf *conf = mddev->private;
5196 	if (conf)
5197 		return sprintf(page, "%d\n", conf->worker_cnt_per_group);
5198 	else
5199 		return 0;
5200 }
5201 
5202 static int alloc_thread_groups(struct r5conf *conf, int cnt);
5203 static ssize_t
5204 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
5205 {
5206 	struct r5conf *conf = mddev->private;
5207 	unsigned long new;
5208 	int err;
5209 	struct r5worker_group *old_groups;
5210 	int old_group_cnt;
5211 
5212 	if (len >= PAGE_SIZE)
5213 		return -EINVAL;
5214 	if (!conf)
5215 		return -ENODEV;
5216 
5217 	if (kstrtoul(page, 10, &new))
5218 		return -EINVAL;
5219 
5220 	if (new == conf->worker_cnt_per_group)
5221 		return len;
5222 
5223 	mddev_suspend(mddev);
5224 
5225 	old_groups = conf->worker_groups;
5226 	old_group_cnt = conf->worker_cnt_per_group;
5227 
5228 	conf->worker_groups = NULL;
5229 	err = alloc_thread_groups(conf, new);
5230 	if (err) {
5231 		conf->worker_groups = old_groups;
5232 		conf->worker_cnt_per_group = old_group_cnt;
5233 	} else {
5234 		if (old_groups)
5235 			kfree(old_groups[0].workers);
5236 		kfree(old_groups);
5237 	}
5238 
5239 	mddev_resume(mddev);
5240 
5241 	if (err)
5242 		return err;
5243 	return len;
5244 }
5245 
5246 static struct md_sysfs_entry
5247 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
5248 				raid5_show_group_thread_cnt,
5249 				raid5_store_group_thread_cnt);
5250 
5251 static struct attribute *raid5_attrs[] =  {
5252 	&raid5_stripecache_size.attr,
5253 	&raid5_stripecache_active.attr,
5254 	&raid5_preread_bypass_threshold.attr,
5255 	&raid5_group_thread_cnt.attr,
5256 	NULL,
5257 };
5258 static struct attribute_group raid5_attrs_group = {
5259 	.name = NULL,
5260 	.attrs = raid5_attrs,
5261 };
5262 
5263 static int alloc_thread_groups(struct r5conf *conf, int cnt)
5264 {
5265 	int i, j;
5266 	ssize_t size;
5267 	struct r5worker *workers;
5268 
5269 	conf->worker_cnt_per_group = cnt;
5270 	if (cnt == 0) {
5271 		conf->worker_groups = NULL;
5272 		return 0;
5273 	}
5274 	conf->group_cnt = num_possible_nodes();
5275 	size = sizeof(struct r5worker) * cnt;
5276 	workers = kzalloc(size * conf->group_cnt, GFP_NOIO);
5277 	conf->worker_groups = kzalloc(sizeof(struct r5worker_group) *
5278 				conf->group_cnt, GFP_NOIO);
5279 	if (!conf->worker_groups || !workers) {
5280 		kfree(workers);
5281 		kfree(conf->worker_groups);
5282 		conf->worker_groups = NULL;
5283 		return -ENOMEM;
5284 	}
5285 
5286 	for (i = 0; i < conf->group_cnt; i++) {
5287 		struct r5worker_group *group;
5288 
5289 		group = &conf->worker_groups[i];
5290 		INIT_LIST_HEAD(&group->handle_list);
5291 		group->conf = conf;
5292 		group->workers = workers + i * cnt;
5293 
5294 		for (j = 0; j < cnt; j++) {
5295 			group->workers[j].group = group;
5296 			INIT_WORK(&group->workers[j].work, raid5_do_work);
5297 		}
5298 	}
5299 
5300 	return 0;
5301 }
5302 
5303 static void free_thread_groups(struct r5conf *conf)
5304 {
5305 	if (conf->worker_groups)
5306 		kfree(conf->worker_groups[0].workers);
5307 	kfree(conf->worker_groups);
5308 	conf->worker_groups = NULL;
5309 }
5310 
5311 static sector_t
5312 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
5313 {
5314 	struct r5conf *conf = mddev->private;
5315 
5316 	if (!sectors)
5317 		sectors = mddev->dev_sectors;
5318 	if (!raid_disks)
5319 		/* size is defined by the smallest of previous and new size */
5320 		raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
5321 
5322 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5323 	sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
5324 	return sectors * (raid_disks - conf->max_degraded);
5325 }
5326 
5327 static void raid5_free_percpu(struct r5conf *conf)
5328 {
5329 	struct raid5_percpu *percpu;
5330 	unsigned long cpu;
5331 
5332 	if (!conf->percpu)
5333 		return;
5334 
5335 	get_online_cpus();
5336 	for_each_possible_cpu(cpu) {
5337 		percpu = per_cpu_ptr(conf->percpu, cpu);
5338 		safe_put_page(percpu->spare_page);
5339 		kfree(percpu->scribble);
5340 	}
5341 #ifdef CONFIG_HOTPLUG_CPU
5342 	unregister_cpu_notifier(&conf->cpu_notify);
5343 #endif
5344 	put_online_cpus();
5345 
5346 	free_percpu(conf->percpu);
5347 }
5348 
5349 static void free_conf(struct r5conf *conf)
5350 {
5351 	free_thread_groups(conf);
5352 	shrink_stripes(conf);
5353 	raid5_free_percpu(conf);
5354 	kfree(conf->disks);
5355 	kfree(conf->stripe_hashtbl);
5356 	kfree(conf);
5357 }
5358 
5359 #ifdef CONFIG_HOTPLUG_CPU
5360 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
5361 			      void *hcpu)
5362 {
5363 	struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
5364 	long cpu = (long)hcpu;
5365 	struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
5366 
5367 	switch (action) {
5368 	case CPU_UP_PREPARE:
5369 	case CPU_UP_PREPARE_FROZEN:
5370 		if (conf->level == 6 && !percpu->spare_page)
5371 			percpu->spare_page = alloc_page(GFP_KERNEL);
5372 		if (!percpu->scribble)
5373 			percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5374 
5375 		if (!percpu->scribble ||
5376 		    (conf->level == 6 && !percpu->spare_page)) {
5377 			safe_put_page(percpu->spare_page);
5378 			kfree(percpu->scribble);
5379 			pr_err("%s: failed memory allocation for cpu%ld\n",
5380 			       __func__, cpu);
5381 			return notifier_from_errno(-ENOMEM);
5382 		}
5383 		break;
5384 	case CPU_DEAD:
5385 	case CPU_DEAD_FROZEN:
5386 		safe_put_page(percpu->spare_page);
5387 		kfree(percpu->scribble);
5388 		percpu->spare_page = NULL;
5389 		percpu->scribble = NULL;
5390 		break;
5391 	default:
5392 		break;
5393 	}
5394 	return NOTIFY_OK;
5395 }
5396 #endif
5397 
5398 static int raid5_alloc_percpu(struct r5conf *conf)
5399 {
5400 	unsigned long cpu;
5401 	struct page *spare_page;
5402 	struct raid5_percpu __percpu *allcpus;
5403 	void *scribble;
5404 	int err;
5405 
5406 	allcpus = alloc_percpu(struct raid5_percpu);
5407 	if (!allcpus)
5408 		return -ENOMEM;
5409 	conf->percpu = allcpus;
5410 
5411 	get_online_cpus();
5412 	err = 0;
5413 	for_each_present_cpu(cpu) {
5414 		if (conf->level == 6) {
5415 			spare_page = alloc_page(GFP_KERNEL);
5416 			if (!spare_page) {
5417 				err = -ENOMEM;
5418 				break;
5419 			}
5420 			per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
5421 		}
5422 		scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
5423 		if (!scribble) {
5424 			err = -ENOMEM;
5425 			break;
5426 		}
5427 		per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
5428 	}
5429 #ifdef CONFIG_HOTPLUG_CPU
5430 	conf->cpu_notify.notifier_call = raid456_cpu_notify;
5431 	conf->cpu_notify.priority = 0;
5432 	if (err == 0)
5433 		err = register_cpu_notifier(&conf->cpu_notify);
5434 #endif
5435 	put_online_cpus();
5436 
5437 	return err;
5438 }
5439 
5440 static struct r5conf *setup_conf(struct mddev *mddev)
5441 {
5442 	struct r5conf *conf;
5443 	int raid_disk, memory, max_disks;
5444 	struct md_rdev *rdev;
5445 	struct disk_info *disk;
5446 	char pers_name[6];
5447 
5448 	if (mddev->new_level != 5
5449 	    && mddev->new_level != 4
5450 	    && mddev->new_level != 6) {
5451 		printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
5452 		       mdname(mddev), mddev->new_level);
5453 		return ERR_PTR(-EIO);
5454 	}
5455 	if ((mddev->new_level == 5
5456 	     && !algorithm_valid_raid5(mddev->new_layout)) ||
5457 	    (mddev->new_level == 6
5458 	     && !algorithm_valid_raid6(mddev->new_layout))) {
5459 		printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
5460 		       mdname(mddev), mddev->new_layout);
5461 		return ERR_PTR(-EIO);
5462 	}
5463 	if (mddev->new_level == 6 && mddev->raid_disks < 4) {
5464 		printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
5465 		       mdname(mddev), mddev->raid_disks);
5466 		return ERR_PTR(-EINVAL);
5467 	}
5468 
5469 	if (!mddev->new_chunk_sectors ||
5470 	    (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
5471 	    !is_power_of_2(mddev->new_chunk_sectors)) {
5472 		printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
5473 		       mdname(mddev), mddev->new_chunk_sectors << 9);
5474 		return ERR_PTR(-EINVAL);
5475 	}
5476 
5477 	conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
5478 	if (conf == NULL)
5479 		goto abort;
5480 	/* Don't enable multi-threading by default*/
5481 	if (alloc_thread_groups(conf, 0))
5482 		goto abort;
5483 	spin_lock_init(&conf->device_lock);
5484 	seqcount_init(&conf->gen_lock);
5485 	init_waitqueue_head(&conf->wait_for_stripe);
5486 	init_waitqueue_head(&conf->wait_for_overlap);
5487 	INIT_LIST_HEAD(&conf->handle_list);
5488 	INIT_LIST_HEAD(&conf->hold_list);
5489 	INIT_LIST_HEAD(&conf->delayed_list);
5490 	INIT_LIST_HEAD(&conf->bitmap_list);
5491 	INIT_LIST_HEAD(&conf->inactive_list);
5492 	init_llist_head(&conf->released_stripes);
5493 	atomic_set(&conf->active_stripes, 0);
5494 	atomic_set(&conf->preread_active_stripes, 0);
5495 	atomic_set(&conf->active_aligned_reads, 0);
5496 	conf->bypass_threshold = BYPASS_THRESHOLD;
5497 	conf->recovery_disabled = mddev->recovery_disabled - 1;
5498 
5499 	conf->raid_disks = mddev->raid_disks;
5500 	if (mddev->reshape_position == MaxSector)
5501 		conf->previous_raid_disks = mddev->raid_disks;
5502 	else
5503 		conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
5504 	max_disks = max(conf->raid_disks, conf->previous_raid_disks);
5505 	conf->scribble_len = scribble_len(max_disks);
5506 
5507 	conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
5508 			      GFP_KERNEL);
5509 	if (!conf->disks)
5510 		goto abort;
5511 
5512 	conf->mddev = mddev;
5513 
5514 	if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
5515 		goto abort;
5516 
5517 	conf->level = mddev->new_level;
5518 	if (raid5_alloc_percpu(conf) != 0)
5519 		goto abort;
5520 
5521 	pr_debug("raid456: run(%s) called.\n", mdname(mddev));
5522 
5523 	rdev_for_each(rdev, mddev) {
5524 		raid_disk = rdev->raid_disk;
5525 		if (raid_disk >= max_disks
5526 		    || raid_disk < 0)
5527 			continue;
5528 		disk = conf->disks + raid_disk;
5529 
5530 		if (test_bit(Replacement, &rdev->flags)) {
5531 			if (disk->replacement)
5532 				goto abort;
5533 			disk->replacement = rdev;
5534 		} else {
5535 			if (disk->rdev)
5536 				goto abort;
5537 			disk->rdev = rdev;
5538 		}
5539 
5540 		if (test_bit(In_sync, &rdev->flags)) {
5541 			char b[BDEVNAME_SIZE];
5542 			printk(KERN_INFO "md/raid:%s: device %s operational as raid"
5543 			       " disk %d\n",
5544 			       mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
5545 		} else if (rdev->saved_raid_disk != raid_disk)
5546 			/* Cannot rely on bitmap to complete recovery */
5547 			conf->fullsync = 1;
5548 	}
5549 
5550 	conf->chunk_sectors = mddev->new_chunk_sectors;
5551 	conf->level = mddev->new_level;
5552 	if (conf->level == 6)
5553 		conf->max_degraded = 2;
5554 	else
5555 		conf->max_degraded = 1;
5556 	conf->algorithm = mddev->new_layout;
5557 	conf->max_nr_stripes = NR_STRIPES;
5558 	conf->reshape_progress = mddev->reshape_position;
5559 	if (conf->reshape_progress != MaxSector) {
5560 		conf->prev_chunk_sectors = mddev->chunk_sectors;
5561 		conf->prev_algo = mddev->layout;
5562 	}
5563 
5564 	memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
5565 		 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
5566 	if (grow_stripes(conf, conf->max_nr_stripes)) {
5567 		printk(KERN_ERR
5568 		       "md/raid:%s: couldn't allocate %dkB for buffers\n",
5569 		       mdname(mddev), memory);
5570 		goto abort;
5571 	} else
5572 		printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
5573 		       mdname(mddev), memory);
5574 
5575 	sprintf(pers_name, "raid%d", mddev->new_level);
5576 	conf->thread = md_register_thread(raid5d, mddev, pers_name);
5577 	if (!conf->thread) {
5578 		printk(KERN_ERR
5579 		       "md/raid:%s: couldn't allocate thread.\n",
5580 		       mdname(mddev));
5581 		goto abort;
5582 	}
5583 
5584 	return conf;
5585 
5586  abort:
5587 	if (conf) {
5588 		free_conf(conf);
5589 		return ERR_PTR(-EIO);
5590 	} else
5591 		return ERR_PTR(-ENOMEM);
5592 }
5593 
5594 
5595 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
5596 {
5597 	switch (algo) {
5598 	case ALGORITHM_PARITY_0:
5599 		if (raid_disk < max_degraded)
5600 			return 1;
5601 		break;
5602 	case ALGORITHM_PARITY_N:
5603 		if (raid_disk >= raid_disks - max_degraded)
5604 			return 1;
5605 		break;
5606 	case ALGORITHM_PARITY_0_6:
5607 		if (raid_disk == 0 ||
5608 		    raid_disk == raid_disks - 1)
5609 			return 1;
5610 		break;
5611 	case ALGORITHM_LEFT_ASYMMETRIC_6:
5612 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
5613 	case ALGORITHM_LEFT_SYMMETRIC_6:
5614 	case ALGORITHM_RIGHT_SYMMETRIC_6:
5615 		if (raid_disk == raid_disks - 1)
5616 			return 1;
5617 	}
5618 	return 0;
5619 }
5620 
5621 static int run(struct mddev *mddev)
5622 {
5623 	struct r5conf *conf;
5624 	int working_disks = 0;
5625 	int dirty_parity_disks = 0;
5626 	struct md_rdev *rdev;
5627 	sector_t reshape_offset = 0;
5628 	int i;
5629 	long long min_offset_diff = 0;
5630 	int first = 1;
5631 
5632 	if (mddev->recovery_cp != MaxSector)
5633 		printk(KERN_NOTICE "md/raid:%s: not clean"
5634 		       " -- starting background reconstruction\n",
5635 		       mdname(mddev));
5636 
5637 	rdev_for_each(rdev, mddev) {
5638 		long long diff;
5639 		if (rdev->raid_disk < 0)
5640 			continue;
5641 		diff = (rdev->new_data_offset - rdev->data_offset);
5642 		if (first) {
5643 			min_offset_diff = diff;
5644 			first = 0;
5645 		} else if (mddev->reshape_backwards &&
5646 			 diff < min_offset_diff)
5647 			min_offset_diff = diff;
5648 		else if (!mddev->reshape_backwards &&
5649 			 diff > min_offset_diff)
5650 			min_offset_diff = diff;
5651 	}
5652 
5653 	if (mddev->reshape_position != MaxSector) {
5654 		/* Check that we can continue the reshape.
5655 		 * Difficulties arise if the stripe we would write to
5656 		 * next is at or after the stripe we would read from next.
5657 		 * For a reshape that changes the number of devices, this
5658 		 * is only possible for a very short time, and mdadm makes
5659 		 * sure that time appears to have past before assembling
5660 		 * the array.  So we fail if that time hasn't passed.
5661 		 * For a reshape that keeps the number of devices the same
5662 		 * mdadm must be monitoring the reshape can keeping the
5663 		 * critical areas read-only and backed up.  It will start
5664 		 * the array in read-only mode, so we check for that.
5665 		 */
5666 		sector_t here_new, here_old;
5667 		int old_disks;
5668 		int max_degraded = (mddev->level == 6 ? 2 : 1);
5669 
5670 		if (mddev->new_level != mddev->level) {
5671 			printk(KERN_ERR "md/raid:%s: unsupported reshape "
5672 			       "required - aborting.\n",
5673 			       mdname(mddev));
5674 			return -EINVAL;
5675 		}
5676 		old_disks = mddev->raid_disks - mddev->delta_disks;
5677 		/* reshape_position must be on a new-stripe boundary, and one
5678 		 * further up in new geometry must map after here in old
5679 		 * geometry.
5680 		 */
5681 		here_new = mddev->reshape_position;
5682 		if (sector_div(here_new, mddev->new_chunk_sectors *
5683 			       (mddev->raid_disks - max_degraded))) {
5684 			printk(KERN_ERR "md/raid:%s: reshape_position not "
5685 			       "on a stripe boundary\n", mdname(mddev));
5686 			return -EINVAL;
5687 		}
5688 		reshape_offset = here_new * mddev->new_chunk_sectors;
5689 		/* here_new is the stripe we will write to */
5690 		here_old = mddev->reshape_position;
5691 		sector_div(here_old, mddev->chunk_sectors *
5692 			   (old_disks-max_degraded));
5693 		/* here_old is the first stripe that we might need to read
5694 		 * from */
5695 		if (mddev->delta_disks == 0) {
5696 			if ((here_new * mddev->new_chunk_sectors !=
5697 			     here_old * mddev->chunk_sectors)) {
5698 				printk(KERN_ERR "md/raid:%s: reshape position is"
5699 				       " confused - aborting\n", mdname(mddev));
5700 				return -EINVAL;
5701 			}
5702 			/* We cannot be sure it is safe to start an in-place
5703 			 * reshape.  It is only safe if user-space is monitoring
5704 			 * and taking constant backups.
5705 			 * mdadm always starts a situation like this in
5706 			 * readonly mode so it can take control before
5707 			 * allowing any writes.  So just check for that.
5708 			 */
5709 			if (abs(min_offset_diff) >= mddev->chunk_sectors &&
5710 			    abs(min_offset_diff) >= mddev->new_chunk_sectors)
5711 				/* not really in-place - so OK */;
5712 			else if (mddev->ro == 0) {
5713 				printk(KERN_ERR "md/raid:%s: in-place reshape "
5714 				       "must be started in read-only mode "
5715 				       "- aborting\n",
5716 				       mdname(mddev));
5717 				return -EINVAL;
5718 			}
5719 		} else if (mddev->reshape_backwards
5720 		    ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
5721 		       here_old * mddev->chunk_sectors)
5722 		    : (here_new * mddev->new_chunk_sectors >=
5723 		       here_old * mddev->chunk_sectors + (-min_offset_diff))) {
5724 			/* Reading from the same stripe as writing to - bad */
5725 			printk(KERN_ERR "md/raid:%s: reshape_position too early for "
5726 			       "auto-recovery - aborting.\n",
5727 			       mdname(mddev));
5728 			return -EINVAL;
5729 		}
5730 		printk(KERN_INFO "md/raid:%s: reshape will continue\n",
5731 		       mdname(mddev));
5732 		/* OK, we should be able to continue; */
5733 	} else {
5734 		BUG_ON(mddev->level != mddev->new_level);
5735 		BUG_ON(mddev->layout != mddev->new_layout);
5736 		BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
5737 		BUG_ON(mddev->delta_disks != 0);
5738 	}
5739 
5740 	if (mddev->private == NULL)
5741 		conf = setup_conf(mddev);
5742 	else
5743 		conf = mddev->private;
5744 
5745 	if (IS_ERR(conf))
5746 		return PTR_ERR(conf);
5747 
5748 	conf->min_offset_diff = min_offset_diff;
5749 	mddev->thread = conf->thread;
5750 	conf->thread = NULL;
5751 	mddev->private = conf;
5752 
5753 	for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
5754 	     i++) {
5755 		rdev = conf->disks[i].rdev;
5756 		if (!rdev && conf->disks[i].replacement) {
5757 			/* The replacement is all we have yet */
5758 			rdev = conf->disks[i].replacement;
5759 			conf->disks[i].replacement = NULL;
5760 			clear_bit(Replacement, &rdev->flags);
5761 			conf->disks[i].rdev = rdev;
5762 		}
5763 		if (!rdev)
5764 			continue;
5765 		if (conf->disks[i].replacement &&
5766 		    conf->reshape_progress != MaxSector) {
5767 			/* replacements and reshape simply do not mix. */
5768 			printk(KERN_ERR "md: cannot handle concurrent "
5769 			       "replacement and reshape.\n");
5770 			goto abort;
5771 		}
5772 		if (test_bit(In_sync, &rdev->flags)) {
5773 			working_disks++;
5774 			continue;
5775 		}
5776 		/* This disc is not fully in-sync.  However if it
5777 		 * just stored parity (beyond the recovery_offset),
5778 		 * when we don't need to be concerned about the
5779 		 * array being dirty.
5780 		 * When reshape goes 'backwards', we never have
5781 		 * partially completed devices, so we only need
5782 		 * to worry about reshape going forwards.
5783 		 */
5784 		/* Hack because v0.91 doesn't store recovery_offset properly. */
5785 		if (mddev->major_version == 0 &&
5786 		    mddev->minor_version > 90)
5787 			rdev->recovery_offset = reshape_offset;
5788 
5789 		if (rdev->recovery_offset < reshape_offset) {
5790 			/* We need to check old and new layout */
5791 			if (!only_parity(rdev->raid_disk,
5792 					 conf->algorithm,
5793 					 conf->raid_disks,
5794 					 conf->max_degraded))
5795 				continue;
5796 		}
5797 		if (!only_parity(rdev->raid_disk,
5798 				 conf->prev_algo,
5799 				 conf->previous_raid_disks,
5800 				 conf->max_degraded))
5801 			continue;
5802 		dirty_parity_disks++;
5803 	}
5804 
5805 	/*
5806 	 * 0 for a fully functional array, 1 or 2 for a degraded array.
5807 	 */
5808 	mddev->degraded = calc_degraded(conf);
5809 
5810 	if (has_failed(conf)) {
5811 		printk(KERN_ERR "md/raid:%s: not enough operational devices"
5812 			" (%d/%d failed)\n",
5813 			mdname(mddev), mddev->degraded, conf->raid_disks);
5814 		goto abort;
5815 	}
5816 
5817 	/* device size must be a multiple of chunk size */
5818 	mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
5819 	mddev->resync_max_sectors = mddev->dev_sectors;
5820 
5821 	if (mddev->degraded > dirty_parity_disks &&
5822 	    mddev->recovery_cp != MaxSector) {
5823 		if (mddev->ok_start_degraded)
5824 			printk(KERN_WARNING
5825 			       "md/raid:%s: starting dirty degraded array"
5826 			       " - data corruption possible.\n",
5827 			       mdname(mddev));
5828 		else {
5829 			printk(KERN_ERR
5830 			       "md/raid:%s: cannot start dirty degraded array.\n",
5831 			       mdname(mddev));
5832 			goto abort;
5833 		}
5834 	}
5835 
5836 	if (mddev->degraded == 0)
5837 		printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
5838 		       " devices, algorithm %d\n", mdname(mddev), conf->level,
5839 		       mddev->raid_disks-mddev->degraded, mddev->raid_disks,
5840 		       mddev->new_layout);
5841 	else
5842 		printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
5843 		       " out of %d devices, algorithm %d\n",
5844 		       mdname(mddev), conf->level,
5845 		       mddev->raid_disks - mddev->degraded,
5846 		       mddev->raid_disks, mddev->new_layout);
5847 
5848 	print_raid5_conf(conf);
5849 
5850 	if (conf->reshape_progress != MaxSector) {
5851 		conf->reshape_safe = conf->reshape_progress;
5852 		atomic_set(&conf->reshape_stripes, 0);
5853 		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5854 		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5855 		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5856 		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5857 		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5858 							"reshape");
5859 	}
5860 
5861 
5862 	/* Ok, everything is just fine now */
5863 	if (mddev->to_remove == &raid5_attrs_group)
5864 		mddev->to_remove = NULL;
5865 	else if (mddev->kobj.sd &&
5866 	    sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
5867 		printk(KERN_WARNING
5868 		       "raid5: failed to create sysfs attributes for %s\n",
5869 		       mdname(mddev));
5870 	md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5871 
5872 	if (mddev->queue) {
5873 		int chunk_size;
5874 		bool discard_supported = true;
5875 		/* read-ahead size must cover two whole stripes, which
5876 		 * is 2 * (datadisks) * chunksize where 'n' is the
5877 		 * number of raid devices
5878 		 */
5879 		int data_disks = conf->previous_raid_disks - conf->max_degraded;
5880 		int stripe = data_disks *
5881 			((mddev->chunk_sectors << 9) / PAGE_SIZE);
5882 		if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5883 			mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5884 
5885 		blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
5886 
5887 		mddev->queue->backing_dev_info.congested_data = mddev;
5888 		mddev->queue->backing_dev_info.congested_fn = raid5_congested;
5889 
5890 		chunk_size = mddev->chunk_sectors << 9;
5891 		blk_queue_io_min(mddev->queue, chunk_size);
5892 		blk_queue_io_opt(mddev->queue, chunk_size *
5893 				 (conf->raid_disks - conf->max_degraded));
5894 		/*
5895 		 * We can only discard a whole stripe. It doesn't make sense to
5896 		 * discard data disk but write parity disk
5897 		 */
5898 		stripe = stripe * PAGE_SIZE;
5899 		/* Round up to power of 2, as discard handling
5900 		 * currently assumes that */
5901 		while ((stripe-1) & stripe)
5902 			stripe = (stripe | (stripe-1)) + 1;
5903 		mddev->queue->limits.discard_alignment = stripe;
5904 		mddev->queue->limits.discard_granularity = stripe;
5905 		/*
5906 		 * unaligned part of discard request will be ignored, so can't
5907 		 * guarantee discard_zerors_data
5908 		 */
5909 		mddev->queue->limits.discard_zeroes_data = 0;
5910 
5911 		blk_queue_max_write_same_sectors(mddev->queue, 0);
5912 
5913 		rdev_for_each(rdev, mddev) {
5914 			disk_stack_limits(mddev->gendisk, rdev->bdev,
5915 					  rdev->data_offset << 9);
5916 			disk_stack_limits(mddev->gendisk, rdev->bdev,
5917 					  rdev->new_data_offset << 9);
5918 			/*
5919 			 * discard_zeroes_data is required, otherwise data
5920 			 * could be lost. Consider a scenario: discard a stripe
5921 			 * (the stripe could be inconsistent if
5922 			 * discard_zeroes_data is 0); write one disk of the
5923 			 * stripe (the stripe could be inconsistent again
5924 			 * depending on which disks are used to calculate
5925 			 * parity); the disk is broken; The stripe data of this
5926 			 * disk is lost.
5927 			 */
5928 			if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
5929 			    !bdev_get_queue(rdev->bdev)->
5930 						limits.discard_zeroes_data)
5931 				discard_supported = false;
5932 		}
5933 
5934 		if (discard_supported &&
5935 		   mddev->queue->limits.max_discard_sectors >= stripe &&
5936 		   mddev->queue->limits.discard_granularity >= stripe)
5937 			queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
5938 						mddev->queue);
5939 		else
5940 			queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
5941 						mddev->queue);
5942 	}
5943 
5944 	return 0;
5945 abort:
5946 	md_unregister_thread(&mddev->thread);
5947 	print_raid5_conf(conf);
5948 	free_conf(conf);
5949 	mddev->private = NULL;
5950 	printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
5951 	return -EIO;
5952 }
5953 
5954 static int stop(struct mddev *mddev)
5955 {
5956 	struct r5conf *conf = mddev->private;
5957 
5958 	md_unregister_thread(&mddev->thread);
5959 	if (mddev->queue)
5960 		mddev->queue->backing_dev_info.congested_fn = NULL;
5961 	free_conf(conf);
5962 	mddev->private = NULL;
5963 	mddev->to_remove = &raid5_attrs_group;
5964 	return 0;
5965 }
5966 
5967 static void status(struct seq_file *seq, struct mddev *mddev)
5968 {
5969 	struct r5conf *conf = mddev->private;
5970 	int i;
5971 
5972 	seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5973 		mddev->chunk_sectors / 2, mddev->layout);
5974 	seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5975 	for (i = 0; i < conf->raid_disks; i++)
5976 		seq_printf (seq, "%s",
5977 			       conf->disks[i].rdev &&
5978 			       test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5979 	seq_printf (seq, "]");
5980 }
5981 
5982 static void print_raid5_conf (struct r5conf *conf)
5983 {
5984 	int i;
5985 	struct disk_info *tmp;
5986 
5987 	printk(KERN_DEBUG "RAID conf printout:\n");
5988 	if (!conf) {
5989 		printk("(conf==NULL)\n");
5990 		return;
5991 	}
5992 	printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
5993 	       conf->raid_disks,
5994 	       conf->raid_disks - conf->mddev->degraded);
5995 
5996 	for (i = 0; i < conf->raid_disks; i++) {
5997 		char b[BDEVNAME_SIZE];
5998 		tmp = conf->disks + i;
5999 		if (tmp->rdev)
6000 			printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6001 			       i, !test_bit(Faulty, &tmp->rdev->flags),
6002 			       bdevname(tmp->rdev->bdev, b));
6003 	}
6004 }
6005 
6006 static int raid5_spare_active(struct mddev *mddev)
6007 {
6008 	int i;
6009 	struct r5conf *conf = mddev->private;
6010 	struct disk_info *tmp;
6011 	int count = 0;
6012 	unsigned long flags;
6013 
6014 	for (i = 0; i < conf->raid_disks; i++) {
6015 		tmp = conf->disks + i;
6016 		if (tmp->replacement
6017 		    && tmp->replacement->recovery_offset == MaxSector
6018 		    && !test_bit(Faulty, &tmp->replacement->flags)
6019 		    && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
6020 			/* Replacement has just become active. */
6021 			if (!tmp->rdev
6022 			    || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
6023 				count++;
6024 			if (tmp->rdev) {
6025 				/* Replaced device not technically faulty,
6026 				 * but we need to be sure it gets removed
6027 				 * and never re-added.
6028 				 */
6029 				set_bit(Faulty, &tmp->rdev->flags);
6030 				sysfs_notify_dirent_safe(
6031 					tmp->rdev->sysfs_state);
6032 			}
6033 			sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
6034 		} else if (tmp->rdev
6035 		    && tmp->rdev->recovery_offset == MaxSector
6036 		    && !test_bit(Faulty, &tmp->rdev->flags)
6037 		    && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
6038 			count++;
6039 			sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
6040 		}
6041 	}
6042 	spin_lock_irqsave(&conf->device_lock, flags);
6043 	mddev->degraded = calc_degraded(conf);
6044 	spin_unlock_irqrestore(&conf->device_lock, flags);
6045 	print_raid5_conf(conf);
6046 	return count;
6047 }
6048 
6049 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
6050 {
6051 	struct r5conf *conf = mddev->private;
6052 	int err = 0;
6053 	int number = rdev->raid_disk;
6054 	struct md_rdev **rdevp;
6055 	struct disk_info *p = conf->disks + number;
6056 
6057 	print_raid5_conf(conf);
6058 	if (rdev == p->rdev)
6059 		rdevp = &p->rdev;
6060 	else if (rdev == p->replacement)
6061 		rdevp = &p->replacement;
6062 	else
6063 		return 0;
6064 
6065 	if (number >= conf->raid_disks &&
6066 	    conf->reshape_progress == MaxSector)
6067 		clear_bit(In_sync, &rdev->flags);
6068 
6069 	if (test_bit(In_sync, &rdev->flags) ||
6070 	    atomic_read(&rdev->nr_pending)) {
6071 		err = -EBUSY;
6072 		goto abort;
6073 	}
6074 	/* Only remove non-faulty devices if recovery
6075 	 * isn't possible.
6076 	 */
6077 	if (!test_bit(Faulty, &rdev->flags) &&
6078 	    mddev->recovery_disabled != conf->recovery_disabled &&
6079 	    !has_failed(conf) &&
6080 	    (!p->replacement || p->replacement == rdev) &&
6081 	    number < conf->raid_disks) {
6082 		err = -EBUSY;
6083 		goto abort;
6084 	}
6085 	*rdevp = NULL;
6086 	synchronize_rcu();
6087 	if (atomic_read(&rdev->nr_pending)) {
6088 		/* lost the race, try later */
6089 		err = -EBUSY;
6090 		*rdevp = rdev;
6091 	} else if (p->replacement) {
6092 		/* We must have just cleared 'rdev' */
6093 		p->rdev = p->replacement;
6094 		clear_bit(Replacement, &p->replacement->flags);
6095 		smp_mb(); /* Make sure other CPUs may see both as identical
6096 			   * but will never see neither - if they are careful
6097 			   */
6098 		p->replacement = NULL;
6099 		clear_bit(WantReplacement, &rdev->flags);
6100 	} else
6101 		/* We might have just removed the Replacement as faulty-
6102 		 * clear the bit just in case
6103 		 */
6104 		clear_bit(WantReplacement, &rdev->flags);
6105 abort:
6106 
6107 	print_raid5_conf(conf);
6108 	return err;
6109 }
6110 
6111 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
6112 {
6113 	struct r5conf *conf = mddev->private;
6114 	int err = -EEXIST;
6115 	int disk;
6116 	struct disk_info *p;
6117 	int first = 0;
6118 	int last = conf->raid_disks - 1;
6119 
6120 	if (mddev->recovery_disabled == conf->recovery_disabled)
6121 		return -EBUSY;
6122 
6123 	if (rdev->saved_raid_disk < 0 && has_failed(conf))
6124 		/* no point adding a device */
6125 		return -EINVAL;
6126 
6127 	if (rdev->raid_disk >= 0)
6128 		first = last = rdev->raid_disk;
6129 
6130 	/*
6131 	 * find the disk ... but prefer rdev->saved_raid_disk
6132 	 * if possible.
6133 	 */
6134 	if (rdev->saved_raid_disk >= 0 &&
6135 	    rdev->saved_raid_disk >= first &&
6136 	    conf->disks[rdev->saved_raid_disk].rdev == NULL)
6137 		first = rdev->saved_raid_disk;
6138 
6139 	for (disk = first; disk <= last; disk++) {
6140 		p = conf->disks + disk;
6141 		if (p->rdev == NULL) {
6142 			clear_bit(In_sync, &rdev->flags);
6143 			rdev->raid_disk = disk;
6144 			err = 0;
6145 			if (rdev->saved_raid_disk != disk)
6146 				conf->fullsync = 1;
6147 			rcu_assign_pointer(p->rdev, rdev);
6148 			goto out;
6149 		}
6150 	}
6151 	for (disk = first; disk <= last; disk++) {
6152 		p = conf->disks + disk;
6153 		if (test_bit(WantReplacement, &p->rdev->flags) &&
6154 		    p->replacement == NULL) {
6155 			clear_bit(In_sync, &rdev->flags);
6156 			set_bit(Replacement, &rdev->flags);
6157 			rdev->raid_disk = disk;
6158 			err = 0;
6159 			conf->fullsync = 1;
6160 			rcu_assign_pointer(p->replacement, rdev);
6161 			break;
6162 		}
6163 	}
6164 out:
6165 	print_raid5_conf(conf);
6166 	return err;
6167 }
6168 
6169 static int raid5_resize(struct mddev *mddev, sector_t sectors)
6170 {
6171 	/* no resync is happening, and there is enough space
6172 	 * on all devices, so we can resize.
6173 	 * We need to make sure resync covers any new space.
6174 	 * If the array is shrinking we should possibly wait until
6175 	 * any io in the removed space completes, but it hardly seems
6176 	 * worth it.
6177 	 */
6178 	sector_t newsize;
6179 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6180 	newsize = raid5_size(mddev, sectors, mddev->raid_disks);
6181 	if (mddev->external_size &&
6182 	    mddev->array_sectors > newsize)
6183 		return -EINVAL;
6184 	if (mddev->bitmap) {
6185 		int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
6186 		if (ret)
6187 			return ret;
6188 	}
6189 	md_set_array_sectors(mddev, newsize);
6190 	set_capacity(mddev->gendisk, mddev->array_sectors);
6191 	revalidate_disk(mddev->gendisk);
6192 	if (sectors > mddev->dev_sectors &&
6193 	    mddev->recovery_cp > mddev->dev_sectors) {
6194 		mddev->recovery_cp = mddev->dev_sectors;
6195 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
6196 	}
6197 	mddev->dev_sectors = sectors;
6198 	mddev->resync_max_sectors = sectors;
6199 	return 0;
6200 }
6201 
6202 static int check_stripe_cache(struct mddev *mddev)
6203 {
6204 	/* Can only proceed if there are plenty of stripe_heads.
6205 	 * We need a minimum of one full stripe,, and for sensible progress
6206 	 * it is best to have about 4 times that.
6207 	 * If we require 4 times, then the default 256 4K stripe_heads will
6208 	 * allow for chunk sizes up to 256K, which is probably OK.
6209 	 * If the chunk size is greater, user-space should request more
6210 	 * stripe_heads first.
6211 	 */
6212 	struct r5conf *conf = mddev->private;
6213 	if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
6214 	    > conf->max_nr_stripes ||
6215 	    ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
6216 	    > conf->max_nr_stripes) {
6217 		printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
6218 		       mdname(mddev),
6219 		       ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
6220 			/ STRIPE_SIZE)*4);
6221 		return 0;
6222 	}
6223 	return 1;
6224 }
6225 
6226 static int check_reshape(struct mddev *mddev)
6227 {
6228 	struct r5conf *conf = mddev->private;
6229 
6230 	if (mddev->delta_disks == 0 &&
6231 	    mddev->new_layout == mddev->layout &&
6232 	    mddev->new_chunk_sectors == mddev->chunk_sectors)
6233 		return 0; /* nothing to do */
6234 	if (has_failed(conf))
6235 		return -EINVAL;
6236 	if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
6237 		/* We might be able to shrink, but the devices must
6238 		 * be made bigger first.
6239 		 * For raid6, 4 is the minimum size.
6240 		 * Otherwise 2 is the minimum
6241 		 */
6242 		int min = 2;
6243 		if (mddev->level == 6)
6244 			min = 4;
6245 		if (mddev->raid_disks + mddev->delta_disks < min)
6246 			return -EINVAL;
6247 	}
6248 
6249 	if (!check_stripe_cache(mddev))
6250 		return -ENOSPC;
6251 
6252 	return resize_stripes(conf, (conf->previous_raid_disks
6253 				     + mddev->delta_disks));
6254 }
6255 
6256 static int raid5_start_reshape(struct mddev *mddev)
6257 {
6258 	struct r5conf *conf = mddev->private;
6259 	struct md_rdev *rdev;
6260 	int spares = 0;
6261 	unsigned long flags;
6262 
6263 	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
6264 		return -EBUSY;
6265 
6266 	if (!check_stripe_cache(mddev))
6267 		return -ENOSPC;
6268 
6269 	if (has_failed(conf))
6270 		return -EINVAL;
6271 
6272 	rdev_for_each(rdev, mddev) {
6273 		if (!test_bit(In_sync, &rdev->flags)
6274 		    && !test_bit(Faulty, &rdev->flags))
6275 			spares++;
6276 	}
6277 
6278 	if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
6279 		/* Not enough devices even to make a degraded array
6280 		 * of that size
6281 		 */
6282 		return -EINVAL;
6283 
6284 	/* Refuse to reduce size of the array.  Any reductions in
6285 	 * array size must be through explicit setting of array_size
6286 	 * attribute.
6287 	 */
6288 	if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
6289 	    < mddev->array_sectors) {
6290 		printk(KERN_ERR "md/raid:%s: array size must be reduced "
6291 		       "before number of disks\n", mdname(mddev));
6292 		return -EINVAL;
6293 	}
6294 
6295 	atomic_set(&conf->reshape_stripes, 0);
6296 	spin_lock_irq(&conf->device_lock);
6297 	write_seqcount_begin(&conf->gen_lock);
6298 	conf->previous_raid_disks = conf->raid_disks;
6299 	conf->raid_disks += mddev->delta_disks;
6300 	conf->prev_chunk_sectors = conf->chunk_sectors;
6301 	conf->chunk_sectors = mddev->new_chunk_sectors;
6302 	conf->prev_algo = conf->algorithm;
6303 	conf->algorithm = mddev->new_layout;
6304 	conf->generation++;
6305 	/* Code that selects data_offset needs to see the generation update
6306 	 * if reshape_progress has been set - so a memory barrier needed.
6307 	 */
6308 	smp_mb();
6309 	if (mddev->reshape_backwards)
6310 		conf->reshape_progress = raid5_size(mddev, 0, 0);
6311 	else
6312 		conf->reshape_progress = 0;
6313 	conf->reshape_safe = conf->reshape_progress;
6314 	write_seqcount_end(&conf->gen_lock);
6315 	spin_unlock_irq(&conf->device_lock);
6316 
6317 	/* Now make sure any requests that proceeded on the assumption
6318 	 * the reshape wasn't running - like Discard or Read - have
6319 	 * completed.
6320 	 */
6321 	mddev_suspend(mddev);
6322 	mddev_resume(mddev);
6323 
6324 	/* Add some new drives, as many as will fit.
6325 	 * We know there are enough to make the newly sized array work.
6326 	 * Don't add devices if we are reducing the number of
6327 	 * devices in the array.  This is because it is not possible
6328 	 * to correctly record the "partially reconstructed" state of
6329 	 * such devices during the reshape and confusion could result.
6330 	 */
6331 	if (mddev->delta_disks >= 0) {
6332 		rdev_for_each(rdev, mddev)
6333 			if (rdev->raid_disk < 0 &&
6334 			    !test_bit(Faulty, &rdev->flags)) {
6335 				if (raid5_add_disk(mddev, rdev) == 0) {
6336 					if (rdev->raid_disk
6337 					    >= conf->previous_raid_disks)
6338 						set_bit(In_sync, &rdev->flags);
6339 					else
6340 						rdev->recovery_offset = 0;
6341 
6342 					if (sysfs_link_rdev(mddev, rdev))
6343 						/* Failure here is OK */;
6344 				}
6345 			} else if (rdev->raid_disk >= conf->previous_raid_disks
6346 				   && !test_bit(Faulty, &rdev->flags)) {
6347 				/* This is a spare that was manually added */
6348 				set_bit(In_sync, &rdev->flags);
6349 			}
6350 
6351 		/* When a reshape changes the number of devices,
6352 		 * ->degraded is measured against the larger of the
6353 		 * pre and post number of devices.
6354 		 */
6355 		spin_lock_irqsave(&conf->device_lock, flags);
6356 		mddev->degraded = calc_degraded(conf);
6357 		spin_unlock_irqrestore(&conf->device_lock, flags);
6358 	}
6359 	mddev->raid_disks = conf->raid_disks;
6360 	mddev->reshape_position = conf->reshape_progress;
6361 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
6362 
6363 	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6364 	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6365 	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6366 	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6367 	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6368 						"reshape");
6369 	if (!mddev->sync_thread) {
6370 		mddev->recovery = 0;
6371 		spin_lock_irq(&conf->device_lock);
6372 		mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
6373 		rdev_for_each(rdev, mddev)
6374 			rdev->new_data_offset = rdev->data_offset;
6375 		smp_wmb();
6376 		conf->reshape_progress = MaxSector;
6377 		mddev->reshape_position = MaxSector;
6378 		spin_unlock_irq(&conf->device_lock);
6379 		return -EAGAIN;
6380 	}
6381 	conf->reshape_checkpoint = jiffies;
6382 	md_wakeup_thread(mddev->sync_thread);
6383 	md_new_event(mddev);
6384 	return 0;
6385 }
6386 
6387 /* This is called from the reshape thread and should make any
6388  * changes needed in 'conf'
6389  */
6390 static void end_reshape(struct r5conf *conf)
6391 {
6392 
6393 	if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
6394 		struct md_rdev *rdev;
6395 
6396 		spin_lock_irq(&conf->device_lock);
6397 		conf->previous_raid_disks = conf->raid_disks;
6398 		rdev_for_each(rdev, conf->mddev)
6399 			rdev->data_offset = rdev->new_data_offset;
6400 		smp_wmb();
6401 		conf->reshape_progress = MaxSector;
6402 		spin_unlock_irq(&conf->device_lock);
6403 		wake_up(&conf->wait_for_overlap);
6404 
6405 		/* read-ahead size must cover two whole stripes, which is
6406 		 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
6407 		 */
6408 		if (conf->mddev->queue) {
6409 			int data_disks = conf->raid_disks - conf->max_degraded;
6410 			int stripe = data_disks * ((conf->chunk_sectors << 9)
6411 						   / PAGE_SIZE);
6412 			if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6413 				conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6414 		}
6415 	}
6416 }
6417 
6418 /* This is called from the raid5d thread with mddev_lock held.
6419  * It makes config changes to the device.
6420  */
6421 static void raid5_finish_reshape(struct mddev *mddev)
6422 {
6423 	struct r5conf *conf = mddev->private;
6424 
6425 	if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
6426 
6427 		if (mddev->delta_disks > 0) {
6428 			md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6429 			set_capacity(mddev->gendisk, mddev->array_sectors);
6430 			revalidate_disk(mddev->gendisk);
6431 		} else {
6432 			int d;
6433 			spin_lock_irq(&conf->device_lock);
6434 			mddev->degraded = calc_degraded(conf);
6435 			spin_unlock_irq(&conf->device_lock);
6436 			for (d = conf->raid_disks ;
6437 			     d < conf->raid_disks - mddev->delta_disks;
6438 			     d++) {
6439 				struct md_rdev *rdev = conf->disks[d].rdev;
6440 				if (rdev)
6441 					clear_bit(In_sync, &rdev->flags);
6442 				rdev = conf->disks[d].replacement;
6443 				if (rdev)
6444 					clear_bit(In_sync, &rdev->flags);
6445 			}
6446 		}
6447 		mddev->layout = conf->algorithm;
6448 		mddev->chunk_sectors = conf->chunk_sectors;
6449 		mddev->reshape_position = MaxSector;
6450 		mddev->delta_disks = 0;
6451 		mddev->reshape_backwards = 0;
6452 	}
6453 }
6454 
6455 static void raid5_quiesce(struct mddev *mddev, int state)
6456 {
6457 	struct r5conf *conf = mddev->private;
6458 
6459 	switch(state) {
6460 	case 2: /* resume for a suspend */
6461 		wake_up(&conf->wait_for_overlap);
6462 		break;
6463 
6464 	case 1: /* stop all writes */
6465 		spin_lock_irq(&conf->device_lock);
6466 		/* '2' tells resync/reshape to pause so that all
6467 		 * active stripes can drain
6468 		 */
6469 		conf->quiesce = 2;
6470 		wait_event_lock_irq(conf->wait_for_stripe,
6471 				    atomic_read(&conf->active_stripes) == 0 &&
6472 				    atomic_read(&conf->active_aligned_reads) == 0,
6473 				    conf->device_lock);
6474 		conf->quiesce = 1;
6475 		spin_unlock_irq(&conf->device_lock);
6476 		/* allow reshape to continue */
6477 		wake_up(&conf->wait_for_overlap);
6478 		break;
6479 
6480 	case 0: /* re-enable writes */
6481 		spin_lock_irq(&conf->device_lock);
6482 		conf->quiesce = 0;
6483 		wake_up(&conf->wait_for_stripe);
6484 		wake_up(&conf->wait_for_overlap);
6485 		spin_unlock_irq(&conf->device_lock);
6486 		break;
6487 	}
6488 }
6489 
6490 
6491 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
6492 {
6493 	struct r0conf *raid0_conf = mddev->private;
6494 	sector_t sectors;
6495 
6496 	/* for raid0 takeover only one zone is supported */
6497 	if (raid0_conf->nr_strip_zones > 1) {
6498 		printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
6499 		       mdname(mddev));
6500 		return ERR_PTR(-EINVAL);
6501 	}
6502 
6503 	sectors = raid0_conf->strip_zone[0].zone_end;
6504 	sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
6505 	mddev->dev_sectors = sectors;
6506 	mddev->new_level = level;
6507 	mddev->new_layout = ALGORITHM_PARITY_N;
6508 	mddev->new_chunk_sectors = mddev->chunk_sectors;
6509 	mddev->raid_disks += 1;
6510 	mddev->delta_disks = 1;
6511 	/* make sure it will be not marked as dirty */
6512 	mddev->recovery_cp = MaxSector;
6513 
6514 	return setup_conf(mddev);
6515 }
6516 
6517 
6518 static void *raid5_takeover_raid1(struct mddev *mddev)
6519 {
6520 	int chunksect;
6521 
6522 	if (mddev->raid_disks != 2 ||
6523 	    mddev->degraded > 1)
6524 		return ERR_PTR(-EINVAL);
6525 
6526 	/* Should check if there are write-behind devices? */
6527 
6528 	chunksect = 64*2; /* 64K by default */
6529 
6530 	/* The array must be an exact multiple of chunksize */
6531 	while (chunksect && (mddev->array_sectors & (chunksect-1)))
6532 		chunksect >>= 1;
6533 
6534 	if ((chunksect<<9) < STRIPE_SIZE)
6535 		/* array size does not allow a suitable chunk size */
6536 		return ERR_PTR(-EINVAL);
6537 
6538 	mddev->new_level = 5;
6539 	mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
6540 	mddev->new_chunk_sectors = chunksect;
6541 
6542 	return setup_conf(mddev);
6543 }
6544 
6545 static void *raid5_takeover_raid6(struct mddev *mddev)
6546 {
6547 	int new_layout;
6548 
6549 	switch (mddev->layout) {
6550 	case ALGORITHM_LEFT_ASYMMETRIC_6:
6551 		new_layout = ALGORITHM_LEFT_ASYMMETRIC;
6552 		break;
6553 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
6554 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
6555 		break;
6556 	case ALGORITHM_LEFT_SYMMETRIC_6:
6557 		new_layout = ALGORITHM_LEFT_SYMMETRIC;
6558 		break;
6559 	case ALGORITHM_RIGHT_SYMMETRIC_6:
6560 		new_layout = ALGORITHM_RIGHT_SYMMETRIC;
6561 		break;
6562 	case ALGORITHM_PARITY_0_6:
6563 		new_layout = ALGORITHM_PARITY_0;
6564 		break;
6565 	case ALGORITHM_PARITY_N:
6566 		new_layout = ALGORITHM_PARITY_N;
6567 		break;
6568 	default:
6569 		return ERR_PTR(-EINVAL);
6570 	}
6571 	mddev->new_level = 5;
6572 	mddev->new_layout = new_layout;
6573 	mddev->delta_disks = -1;
6574 	mddev->raid_disks -= 1;
6575 	return setup_conf(mddev);
6576 }
6577 
6578 
6579 static int raid5_check_reshape(struct mddev *mddev)
6580 {
6581 	/* For a 2-drive array, the layout and chunk size can be changed
6582 	 * immediately as not restriping is needed.
6583 	 * For larger arrays we record the new value - after validation
6584 	 * to be used by a reshape pass.
6585 	 */
6586 	struct r5conf *conf = mddev->private;
6587 	int new_chunk = mddev->new_chunk_sectors;
6588 
6589 	if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
6590 		return -EINVAL;
6591 	if (new_chunk > 0) {
6592 		if (!is_power_of_2(new_chunk))
6593 			return -EINVAL;
6594 		if (new_chunk < (PAGE_SIZE>>9))
6595 			return -EINVAL;
6596 		if (mddev->array_sectors & (new_chunk-1))
6597 			/* not factor of array size */
6598 			return -EINVAL;
6599 	}
6600 
6601 	/* They look valid */
6602 
6603 	if (mddev->raid_disks == 2) {
6604 		/* can make the change immediately */
6605 		if (mddev->new_layout >= 0) {
6606 			conf->algorithm = mddev->new_layout;
6607 			mddev->layout = mddev->new_layout;
6608 		}
6609 		if (new_chunk > 0) {
6610 			conf->chunk_sectors = new_chunk ;
6611 			mddev->chunk_sectors = new_chunk;
6612 		}
6613 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
6614 		md_wakeup_thread(mddev->thread);
6615 	}
6616 	return check_reshape(mddev);
6617 }
6618 
6619 static int raid6_check_reshape(struct mddev *mddev)
6620 {
6621 	int new_chunk = mddev->new_chunk_sectors;
6622 
6623 	if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
6624 		return -EINVAL;
6625 	if (new_chunk > 0) {
6626 		if (!is_power_of_2(new_chunk))
6627 			return -EINVAL;
6628 		if (new_chunk < (PAGE_SIZE >> 9))
6629 			return -EINVAL;
6630 		if (mddev->array_sectors & (new_chunk-1))
6631 			/* not factor of array size */
6632 			return -EINVAL;
6633 	}
6634 
6635 	/* They look valid */
6636 	return check_reshape(mddev);
6637 }
6638 
6639 static void *raid5_takeover(struct mddev *mddev)
6640 {
6641 	/* raid5 can take over:
6642 	 *  raid0 - if there is only one strip zone - make it a raid4 layout
6643 	 *  raid1 - if there are two drives.  We need to know the chunk size
6644 	 *  raid4 - trivial - just use a raid4 layout.
6645 	 *  raid6 - Providing it is a *_6 layout
6646 	 */
6647 	if (mddev->level == 0)
6648 		return raid45_takeover_raid0(mddev, 5);
6649 	if (mddev->level == 1)
6650 		return raid5_takeover_raid1(mddev);
6651 	if (mddev->level == 4) {
6652 		mddev->new_layout = ALGORITHM_PARITY_N;
6653 		mddev->new_level = 5;
6654 		return setup_conf(mddev);
6655 	}
6656 	if (mddev->level == 6)
6657 		return raid5_takeover_raid6(mddev);
6658 
6659 	return ERR_PTR(-EINVAL);
6660 }
6661 
6662 static void *raid4_takeover(struct mddev *mddev)
6663 {
6664 	/* raid4 can take over:
6665 	 *  raid0 - if there is only one strip zone
6666 	 *  raid5 - if layout is right
6667 	 */
6668 	if (mddev->level == 0)
6669 		return raid45_takeover_raid0(mddev, 4);
6670 	if (mddev->level == 5 &&
6671 	    mddev->layout == ALGORITHM_PARITY_N) {
6672 		mddev->new_layout = 0;
6673 		mddev->new_level = 4;
6674 		return setup_conf(mddev);
6675 	}
6676 	return ERR_PTR(-EINVAL);
6677 }
6678 
6679 static struct md_personality raid5_personality;
6680 
6681 static void *raid6_takeover(struct mddev *mddev)
6682 {
6683 	/* Currently can only take over a raid5.  We map the
6684 	 * personality to an equivalent raid6 personality
6685 	 * with the Q block at the end.
6686 	 */
6687 	int new_layout;
6688 
6689 	if (mddev->pers != &raid5_personality)
6690 		return ERR_PTR(-EINVAL);
6691 	if (mddev->degraded > 1)
6692 		return ERR_PTR(-EINVAL);
6693 	if (mddev->raid_disks > 253)
6694 		return ERR_PTR(-EINVAL);
6695 	if (mddev->raid_disks < 3)
6696 		return ERR_PTR(-EINVAL);
6697 
6698 	switch (mddev->layout) {
6699 	case ALGORITHM_LEFT_ASYMMETRIC:
6700 		new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
6701 		break;
6702 	case ALGORITHM_RIGHT_ASYMMETRIC:
6703 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
6704 		break;
6705 	case ALGORITHM_LEFT_SYMMETRIC:
6706 		new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
6707 		break;
6708 	case ALGORITHM_RIGHT_SYMMETRIC:
6709 		new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
6710 		break;
6711 	case ALGORITHM_PARITY_0:
6712 		new_layout = ALGORITHM_PARITY_0_6;
6713 		break;
6714 	case ALGORITHM_PARITY_N:
6715 		new_layout = ALGORITHM_PARITY_N;
6716 		break;
6717 	default:
6718 		return ERR_PTR(-EINVAL);
6719 	}
6720 	mddev->new_level = 6;
6721 	mddev->new_layout = new_layout;
6722 	mddev->delta_disks = 1;
6723 	mddev->raid_disks += 1;
6724 	return setup_conf(mddev);
6725 }
6726 
6727 
6728 static struct md_personality raid6_personality =
6729 {
6730 	.name		= "raid6",
6731 	.level		= 6,
6732 	.owner		= THIS_MODULE,
6733 	.make_request	= make_request,
6734 	.run		= run,
6735 	.stop		= stop,
6736 	.status		= status,
6737 	.error_handler	= error,
6738 	.hot_add_disk	= raid5_add_disk,
6739 	.hot_remove_disk= raid5_remove_disk,
6740 	.spare_active	= raid5_spare_active,
6741 	.sync_request	= sync_request,
6742 	.resize		= raid5_resize,
6743 	.size		= raid5_size,
6744 	.check_reshape	= raid6_check_reshape,
6745 	.start_reshape  = raid5_start_reshape,
6746 	.finish_reshape = raid5_finish_reshape,
6747 	.quiesce	= raid5_quiesce,
6748 	.takeover	= raid6_takeover,
6749 };
6750 static struct md_personality raid5_personality =
6751 {
6752 	.name		= "raid5",
6753 	.level		= 5,
6754 	.owner		= THIS_MODULE,
6755 	.make_request	= make_request,
6756 	.run		= run,
6757 	.stop		= stop,
6758 	.status		= status,
6759 	.error_handler	= error,
6760 	.hot_add_disk	= raid5_add_disk,
6761 	.hot_remove_disk= raid5_remove_disk,
6762 	.spare_active	= raid5_spare_active,
6763 	.sync_request	= sync_request,
6764 	.resize		= raid5_resize,
6765 	.size		= raid5_size,
6766 	.check_reshape	= raid5_check_reshape,
6767 	.start_reshape  = raid5_start_reshape,
6768 	.finish_reshape = raid5_finish_reshape,
6769 	.quiesce	= raid5_quiesce,
6770 	.takeover	= raid5_takeover,
6771 };
6772 
6773 static struct md_personality raid4_personality =
6774 {
6775 	.name		= "raid4",
6776 	.level		= 4,
6777 	.owner		= THIS_MODULE,
6778 	.make_request	= make_request,
6779 	.run		= run,
6780 	.stop		= stop,
6781 	.status		= status,
6782 	.error_handler	= error,
6783 	.hot_add_disk	= raid5_add_disk,
6784 	.hot_remove_disk= raid5_remove_disk,
6785 	.spare_active	= raid5_spare_active,
6786 	.sync_request	= sync_request,
6787 	.resize		= raid5_resize,
6788 	.size		= raid5_size,
6789 	.check_reshape	= raid5_check_reshape,
6790 	.start_reshape  = raid5_start_reshape,
6791 	.finish_reshape = raid5_finish_reshape,
6792 	.quiesce	= raid5_quiesce,
6793 	.takeover	= raid4_takeover,
6794 };
6795 
6796 static int __init raid5_init(void)
6797 {
6798 	raid5_wq = alloc_workqueue("raid5wq",
6799 		WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
6800 	if (!raid5_wq)
6801 		return -ENOMEM;
6802 	register_md_personality(&raid6_personality);
6803 	register_md_personality(&raid5_personality);
6804 	register_md_personality(&raid4_personality);
6805 	return 0;
6806 }
6807 
6808 static void raid5_exit(void)
6809 {
6810 	unregister_md_personality(&raid6_personality);
6811 	unregister_md_personality(&raid5_personality);
6812 	unregister_md_personality(&raid4_personality);
6813 	destroy_workqueue(raid5_wq);
6814 }
6815 
6816 module_init(raid5_init);
6817 module_exit(raid5_exit);
6818 MODULE_LICENSE("GPL");
6819 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
6820 MODULE_ALIAS("md-personality-4"); /* RAID5 */
6821 MODULE_ALIAS("md-raid5");
6822 MODULE_ALIAS("md-raid4");
6823 MODULE_ALIAS("md-level-5");
6824 MODULE_ALIAS("md-level-4");
6825 MODULE_ALIAS("md-personality-8"); /* RAID6 */
6826 MODULE_ALIAS("md-raid6");
6827 MODULE_ALIAS("md-level-6");
6828 
6829 /* This used to be two separate modules, they were: */
6830 MODULE_ALIAS("raid5");
6831 MODULE_ALIAS("raid6");
6832