xref: /openbmc/linux/drivers/md/raid5.c (revision 52fb57e7)
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 <linux/flex_array.h>
58 #include <trace/events/block.h>
59 
60 #include "md.h"
61 #include "raid5.h"
62 #include "raid0.h"
63 #include "bitmap.h"
64 
65 #define cpu_to_group(cpu) cpu_to_node(cpu)
66 #define ANY_GROUP NUMA_NO_NODE
67 
68 static bool devices_handle_discard_safely = false;
69 module_param(devices_handle_discard_safely, bool, 0644);
70 MODULE_PARM_DESC(devices_handle_discard_safely,
71 		 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
72 static struct workqueue_struct *raid5_wq;
73 /*
74  * Stripe cache
75  */
76 
77 #define NR_STRIPES		256
78 #define STRIPE_SIZE		PAGE_SIZE
79 #define STRIPE_SHIFT		(PAGE_SHIFT - 9)
80 #define STRIPE_SECTORS		(STRIPE_SIZE>>9)
81 #define	IO_THRESHOLD		1
82 #define BYPASS_THRESHOLD	1
83 #define NR_HASH			(PAGE_SIZE / sizeof(struct hlist_head))
84 #define HASH_MASK		(NR_HASH - 1)
85 #define MAX_STRIPE_BATCH	8
86 
87 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
88 {
89 	int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
90 	return &conf->stripe_hashtbl[hash];
91 }
92 
93 static inline int stripe_hash_locks_hash(sector_t sect)
94 {
95 	return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
96 }
97 
98 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
99 {
100 	spin_lock_irq(conf->hash_locks + hash);
101 	spin_lock(&conf->device_lock);
102 }
103 
104 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
105 {
106 	spin_unlock(&conf->device_lock);
107 	spin_unlock_irq(conf->hash_locks + hash);
108 }
109 
110 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
111 {
112 	int i;
113 	local_irq_disable();
114 	spin_lock(conf->hash_locks);
115 	for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
116 		spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
117 	spin_lock(&conf->device_lock);
118 }
119 
120 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
121 {
122 	int i;
123 	spin_unlock(&conf->device_lock);
124 	for (i = NR_STRIPE_HASH_LOCKS; i; i--)
125 		spin_unlock(conf->hash_locks + i - 1);
126 	local_irq_enable();
127 }
128 
129 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
130  * order without overlap.  There may be several bio's per stripe+device, and
131  * a bio could span several devices.
132  * When walking this list for a particular stripe+device, we must never proceed
133  * beyond a bio that extends past this device, as the next bio might no longer
134  * be valid.
135  * This function is used to determine the 'next' bio in the list, given the sector
136  * of the current stripe+device
137  */
138 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
139 {
140 	int sectors = bio_sectors(bio);
141 	if (bio->bi_iter.bi_sector + sectors < sector + STRIPE_SECTORS)
142 		return bio->bi_next;
143 	else
144 		return NULL;
145 }
146 
147 /*
148  * We maintain a biased count of active stripes in the bottom 16 bits of
149  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
150  */
151 static inline int raid5_bi_processed_stripes(struct bio *bio)
152 {
153 	atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
154 	return (atomic_read(segments) >> 16) & 0xffff;
155 }
156 
157 static inline int raid5_dec_bi_active_stripes(struct bio *bio)
158 {
159 	atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
160 	return atomic_sub_return(1, segments) & 0xffff;
161 }
162 
163 static inline void raid5_inc_bi_active_stripes(struct bio *bio)
164 {
165 	atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
166 	atomic_inc(segments);
167 }
168 
169 static inline void raid5_set_bi_processed_stripes(struct bio *bio,
170 	unsigned int cnt)
171 {
172 	atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
173 	int old, new;
174 
175 	do {
176 		old = atomic_read(segments);
177 		new = (old & 0xffff) | (cnt << 16);
178 	} while (atomic_cmpxchg(segments, old, new) != old);
179 }
180 
181 static inline void raid5_set_bi_stripes(struct bio *bio, unsigned int cnt)
182 {
183 	atomic_t *segments = (atomic_t *)&bio->bi_phys_segments;
184 	atomic_set(segments, cnt);
185 }
186 
187 /* Find first data disk in a raid6 stripe */
188 static inline int raid6_d0(struct stripe_head *sh)
189 {
190 	if (sh->ddf_layout)
191 		/* ddf always start from first device */
192 		return 0;
193 	/* md starts just after Q block */
194 	if (sh->qd_idx == sh->disks - 1)
195 		return 0;
196 	else
197 		return sh->qd_idx + 1;
198 }
199 static inline int raid6_next_disk(int disk, int raid_disks)
200 {
201 	disk++;
202 	return (disk < raid_disks) ? disk : 0;
203 }
204 
205 /* When walking through the disks in a raid5, starting at raid6_d0,
206  * We need to map each disk to a 'slot', where the data disks are slot
207  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
208  * is raid_disks-1.  This help does that mapping.
209  */
210 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
211 			     int *count, int syndrome_disks)
212 {
213 	int slot = *count;
214 
215 	if (sh->ddf_layout)
216 		(*count)++;
217 	if (idx == sh->pd_idx)
218 		return syndrome_disks;
219 	if (idx == sh->qd_idx)
220 		return syndrome_disks + 1;
221 	if (!sh->ddf_layout)
222 		(*count)++;
223 	return slot;
224 }
225 
226 static void return_io(struct bio *return_bi)
227 {
228 	struct bio *bi = return_bi;
229 	while (bi) {
230 
231 		return_bi = bi->bi_next;
232 		bi->bi_next = NULL;
233 		bi->bi_iter.bi_size = 0;
234 		trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
235 					 bi, 0);
236 		bio_endio(bi, 0);
237 		bi = return_bi;
238 	}
239 }
240 
241 static void print_raid5_conf (struct r5conf *conf);
242 
243 static int stripe_operations_active(struct stripe_head *sh)
244 {
245 	return sh->check_state || sh->reconstruct_state ||
246 	       test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
247 	       test_bit(STRIPE_COMPUTE_RUN, &sh->state);
248 }
249 
250 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
251 {
252 	struct r5conf *conf = sh->raid_conf;
253 	struct r5worker_group *group;
254 	int thread_cnt;
255 	int i, cpu = sh->cpu;
256 
257 	if (!cpu_online(cpu)) {
258 		cpu = cpumask_any(cpu_online_mask);
259 		sh->cpu = cpu;
260 	}
261 
262 	if (list_empty(&sh->lru)) {
263 		struct r5worker_group *group;
264 		group = conf->worker_groups + cpu_to_group(cpu);
265 		list_add_tail(&sh->lru, &group->handle_list);
266 		group->stripes_cnt++;
267 		sh->group = group;
268 	}
269 
270 	if (conf->worker_cnt_per_group == 0) {
271 		md_wakeup_thread(conf->mddev->thread);
272 		return;
273 	}
274 
275 	group = conf->worker_groups + cpu_to_group(sh->cpu);
276 
277 	group->workers[0].working = true;
278 	/* at least one worker should run to avoid race */
279 	queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
280 
281 	thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
282 	/* wakeup more workers */
283 	for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
284 		if (group->workers[i].working == false) {
285 			group->workers[i].working = true;
286 			queue_work_on(sh->cpu, raid5_wq,
287 				      &group->workers[i].work);
288 			thread_cnt--;
289 		}
290 	}
291 }
292 
293 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
294 			      struct list_head *temp_inactive_list)
295 {
296 	BUG_ON(!list_empty(&sh->lru));
297 	BUG_ON(atomic_read(&conf->active_stripes)==0);
298 	if (test_bit(STRIPE_HANDLE, &sh->state)) {
299 		if (test_bit(STRIPE_DELAYED, &sh->state) &&
300 		    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
301 			list_add_tail(&sh->lru, &conf->delayed_list);
302 		else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
303 			   sh->bm_seq - conf->seq_write > 0)
304 			list_add_tail(&sh->lru, &conf->bitmap_list);
305 		else {
306 			clear_bit(STRIPE_DELAYED, &sh->state);
307 			clear_bit(STRIPE_BIT_DELAY, &sh->state);
308 			if (conf->worker_cnt_per_group == 0) {
309 				list_add_tail(&sh->lru, &conf->handle_list);
310 			} else {
311 				raid5_wakeup_stripe_thread(sh);
312 				return;
313 			}
314 		}
315 		md_wakeup_thread(conf->mddev->thread);
316 	} else {
317 		BUG_ON(stripe_operations_active(sh));
318 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
319 			if (atomic_dec_return(&conf->preread_active_stripes)
320 			    < IO_THRESHOLD)
321 				md_wakeup_thread(conf->mddev->thread);
322 		atomic_dec(&conf->active_stripes);
323 		if (!test_bit(STRIPE_EXPANDING, &sh->state))
324 			list_add_tail(&sh->lru, temp_inactive_list);
325 	}
326 }
327 
328 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
329 			     struct list_head *temp_inactive_list)
330 {
331 	if (atomic_dec_and_test(&sh->count))
332 		do_release_stripe(conf, sh, temp_inactive_list);
333 }
334 
335 /*
336  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
337  *
338  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
339  * given time. Adding stripes only takes device lock, while deleting stripes
340  * only takes hash lock.
341  */
342 static void release_inactive_stripe_list(struct r5conf *conf,
343 					 struct list_head *temp_inactive_list,
344 					 int hash)
345 {
346 	int size;
347 	bool do_wakeup = false;
348 	unsigned long flags;
349 
350 	if (hash == NR_STRIPE_HASH_LOCKS) {
351 		size = NR_STRIPE_HASH_LOCKS;
352 		hash = NR_STRIPE_HASH_LOCKS - 1;
353 	} else
354 		size = 1;
355 	while (size) {
356 		struct list_head *list = &temp_inactive_list[size - 1];
357 
358 		/*
359 		 * We don't hold any lock here yet, get_active_stripe() might
360 		 * remove stripes from the list
361 		 */
362 		if (!list_empty_careful(list)) {
363 			spin_lock_irqsave(conf->hash_locks + hash, flags);
364 			if (list_empty(conf->inactive_list + hash) &&
365 			    !list_empty(list))
366 				atomic_dec(&conf->empty_inactive_list_nr);
367 			list_splice_tail_init(list, conf->inactive_list + hash);
368 			do_wakeup = true;
369 			spin_unlock_irqrestore(conf->hash_locks + hash, flags);
370 		}
371 		size--;
372 		hash--;
373 	}
374 
375 	if (do_wakeup) {
376 		wake_up(&conf->wait_for_stripe);
377 		if (conf->retry_read_aligned)
378 			md_wakeup_thread(conf->mddev->thread);
379 	}
380 }
381 
382 /* should hold conf->device_lock already */
383 static int release_stripe_list(struct r5conf *conf,
384 			       struct list_head *temp_inactive_list)
385 {
386 	struct stripe_head *sh;
387 	int count = 0;
388 	struct llist_node *head;
389 
390 	head = llist_del_all(&conf->released_stripes);
391 	head = llist_reverse_order(head);
392 	while (head) {
393 		int hash;
394 
395 		sh = llist_entry(head, struct stripe_head, release_list);
396 		head = llist_next(head);
397 		/* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
398 		smp_mb();
399 		clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
400 		/*
401 		 * Don't worry the bit is set here, because if the bit is set
402 		 * again, the count is always > 1. This is true for
403 		 * STRIPE_ON_UNPLUG_LIST bit too.
404 		 */
405 		hash = sh->hash_lock_index;
406 		__release_stripe(conf, sh, &temp_inactive_list[hash]);
407 		count++;
408 	}
409 
410 	return count;
411 }
412 
413 static void release_stripe(struct stripe_head *sh)
414 {
415 	struct r5conf *conf = sh->raid_conf;
416 	unsigned long flags;
417 	struct list_head list;
418 	int hash;
419 	bool wakeup;
420 
421 	/* Avoid release_list until the last reference.
422 	 */
423 	if (atomic_add_unless(&sh->count, -1, 1))
424 		return;
425 
426 	if (unlikely(!conf->mddev->thread) ||
427 		test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
428 		goto slow_path;
429 	wakeup = llist_add(&sh->release_list, &conf->released_stripes);
430 	if (wakeup)
431 		md_wakeup_thread(conf->mddev->thread);
432 	return;
433 slow_path:
434 	local_irq_save(flags);
435 	/* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
436 	if (atomic_dec_and_lock(&sh->count, &conf->device_lock)) {
437 		INIT_LIST_HEAD(&list);
438 		hash = sh->hash_lock_index;
439 		do_release_stripe(conf, sh, &list);
440 		spin_unlock(&conf->device_lock);
441 		release_inactive_stripe_list(conf, &list, hash);
442 	}
443 	local_irq_restore(flags);
444 }
445 
446 static inline void remove_hash(struct stripe_head *sh)
447 {
448 	pr_debug("remove_hash(), stripe %llu\n",
449 		(unsigned long long)sh->sector);
450 
451 	hlist_del_init(&sh->hash);
452 }
453 
454 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
455 {
456 	struct hlist_head *hp = stripe_hash(conf, sh->sector);
457 
458 	pr_debug("insert_hash(), stripe %llu\n",
459 		(unsigned long long)sh->sector);
460 
461 	hlist_add_head(&sh->hash, hp);
462 }
463 
464 /* find an idle stripe, make sure it is unhashed, and return it. */
465 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
466 {
467 	struct stripe_head *sh = NULL;
468 	struct list_head *first;
469 
470 	if (list_empty(conf->inactive_list + hash))
471 		goto out;
472 	first = (conf->inactive_list + hash)->next;
473 	sh = list_entry(first, struct stripe_head, lru);
474 	list_del_init(first);
475 	remove_hash(sh);
476 	atomic_inc(&conf->active_stripes);
477 	BUG_ON(hash != sh->hash_lock_index);
478 	if (list_empty(conf->inactive_list + hash))
479 		atomic_inc(&conf->empty_inactive_list_nr);
480 out:
481 	return sh;
482 }
483 
484 static void shrink_buffers(struct stripe_head *sh)
485 {
486 	struct page *p;
487 	int i;
488 	int num = sh->raid_conf->pool_size;
489 
490 	for (i = 0; i < num ; i++) {
491 		WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
492 		p = sh->dev[i].page;
493 		if (!p)
494 			continue;
495 		sh->dev[i].page = NULL;
496 		put_page(p);
497 	}
498 }
499 
500 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
501 {
502 	int i;
503 	int num = sh->raid_conf->pool_size;
504 
505 	for (i = 0; i < num; i++) {
506 		struct page *page;
507 
508 		if (!(page = alloc_page(gfp))) {
509 			return 1;
510 		}
511 		sh->dev[i].page = page;
512 		sh->dev[i].orig_page = page;
513 	}
514 	return 0;
515 }
516 
517 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
518 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
519 			    struct stripe_head *sh);
520 
521 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
522 {
523 	struct r5conf *conf = sh->raid_conf;
524 	int i, seq;
525 
526 	BUG_ON(atomic_read(&sh->count) != 0);
527 	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
528 	BUG_ON(stripe_operations_active(sh));
529 	BUG_ON(sh->batch_head);
530 
531 	pr_debug("init_stripe called, stripe %llu\n",
532 		(unsigned long long)sector);
533 retry:
534 	seq = read_seqcount_begin(&conf->gen_lock);
535 	sh->generation = conf->generation - previous;
536 	sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
537 	sh->sector = sector;
538 	stripe_set_idx(sector, conf, previous, sh);
539 	sh->state = 0;
540 
541 	for (i = sh->disks; i--; ) {
542 		struct r5dev *dev = &sh->dev[i];
543 
544 		if (dev->toread || dev->read || dev->towrite || dev->written ||
545 		    test_bit(R5_LOCKED, &dev->flags)) {
546 			printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
547 			       (unsigned long long)sh->sector, i, dev->toread,
548 			       dev->read, dev->towrite, dev->written,
549 			       test_bit(R5_LOCKED, &dev->flags));
550 			WARN_ON(1);
551 		}
552 		dev->flags = 0;
553 		raid5_build_block(sh, i, previous);
554 	}
555 	if (read_seqcount_retry(&conf->gen_lock, seq))
556 		goto retry;
557 	sh->overwrite_disks = 0;
558 	insert_hash(conf, sh);
559 	sh->cpu = smp_processor_id();
560 	set_bit(STRIPE_BATCH_READY, &sh->state);
561 }
562 
563 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
564 					 short generation)
565 {
566 	struct stripe_head *sh;
567 
568 	pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
569 	hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
570 		if (sh->sector == sector && sh->generation == generation)
571 			return sh;
572 	pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
573 	return NULL;
574 }
575 
576 /*
577  * Need to check if array has failed when deciding whether to:
578  *  - start an array
579  *  - remove non-faulty devices
580  *  - add a spare
581  *  - allow a reshape
582  * This determination is simple when no reshape is happening.
583  * However if there is a reshape, we need to carefully check
584  * both the before and after sections.
585  * This is because some failed devices may only affect one
586  * of the two sections, and some non-in_sync devices may
587  * be insync in the section most affected by failed devices.
588  */
589 static int calc_degraded(struct r5conf *conf)
590 {
591 	int degraded, degraded2;
592 	int i;
593 
594 	rcu_read_lock();
595 	degraded = 0;
596 	for (i = 0; i < conf->previous_raid_disks; i++) {
597 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
598 		if (rdev && test_bit(Faulty, &rdev->flags))
599 			rdev = rcu_dereference(conf->disks[i].replacement);
600 		if (!rdev || test_bit(Faulty, &rdev->flags))
601 			degraded++;
602 		else if (test_bit(In_sync, &rdev->flags))
603 			;
604 		else
605 			/* not in-sync or faulty.
606 			 * If the reshape increases the number of devices,
607 			 * this is being recovered by the reshape, so
608 			 * this 'previous' section is not in_sync.
609 			 * If the number of devices is being reduced however,
610 			 * the device can only be part of the array if
611 			 * we are reverting a reshape, so this section will
612 			 * be in-sync.
613 			 */
614 			if (conf->raid_disks >= conf->previous_raid_disks)
615 				degraded++;
616 	}
617 	rcu_read_unlock();
618 	if (conf->raid_disks == conf->previous_raid_disks)
619 		return degraded;
620 	rcu_read_lock();
621 	degraded2 = 0;
622 	for (i = 0; i < conf->raid_disks; i++) {
623 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
624 		if (rdev && test_bit(Faulty, &rdev->flags))
625 			rdev = rcu_dereference(conf->disks[i].replacement);
626 		if (!rdev || test_bit(Faulty, &rdev->flags))
627 			degraded2++;
628 		else if (test_bit(In_sync, &rdev->flags))
629 			;
630 		else
631 			/* not in-sync or faulty.
632 			 * If reshape increases the number of devices, this
633 			 * section has already been recovered, else it
634 			 * almost certainly hasn't.
635 			 */
636 			if (conf->raid_disks <= conf->previous_raid_disks)
637 				degraded2++;
638 	}
639 	rcu_read_unlock();
640 	if (degraded2 > degraded)
641 		return degraded2;
642 	return degraded;
643 }
644 
645 static int has_failed(struct r5conf *conf)
646 {
647 	int degraded;
648 
649 	if (conf->mddev->reshape_position == MaxSector)
650 		return conf->mddev->degraded > conf->max_degraded;
651 
652 	degraded = calc_degraded(conf);
653 	if (degraded > conf->max_degraded)
654 		return 1;
655 	return 0;
656 }
657 
658 static struct stripe_head *
659 get_active_stripe(struct r5conf *conf, sector_t sector,
660 		  int previous, int noblock, int noquiesce)
661 {
662 	struct stripe_head *sh;
663 	int hash = stripe_hash_locks_hash(sector);
664 
665 	pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
666 
667 	spin_lock_irq(conf->hash_locks + hash);
668 
669 	do {
670 		wait_event_lock_irq(conf->wait_for_stripe,
671 				    conf->quiesce == 0 || noquiesce,
672 				    *(conf->hash_locks + hash));
673 		sh = __find_stripe(conf, sector, conf->generation - previous);
674 		if (!sh) {
675 			if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
676 				sh = get_free_stripe(conf, hash);
677 				if (!sh && llist_empty(&conf->released_stripes) &&
678 				    !test_bit(R5_DID_ALLOC, &conf->cache_state))
679 					set_bit(R5_ALLOC_MORE,
680 						&conf->cache_state);
681 			}
682 			if (noblock && sh == NULL)
683 				break;
684 			if (!sh) {
685 				set_bit(R5_INACTIVE_BLOCKED,
686 					&conf->cache_state);
687 				wait_event_lock_irq(
688 					conf->wait_for_stripe,
689 					!list_empty(conf->inactive_list + hash) &&
690 					(atomic_read(&conf->active_stripes)
691 					 < (conf->max_nr_stripes * 3 / 4)
692 					 || !test_bit(R5_INACTIVE_BLOCKED,
693 						      &conf->cache_state)),
694 					*(conf->hash_locks + hash));
695 				clear_bit(R5_INACTIVE_BLOCKED,
696 					  &conf->cache_state);
697 			} else {
698 				init_stripe(sh, sector, previous);
699 				atomic_inc(&sh->count);
700 			}
701 		} else if (!atomic_inc_not_zero(&sh->count)) {
702 			spin_lock(&conf->device_lock);
703 			if (!atomic_read(&sh->count)) {
704 				if (!test_bit(STRIPE_HANDLE, &sh->state))
705 					atomic_inc(&conf->active_stripes);
706 				BUG_ON(list_empty(&sh->lru) &&
707 				       !test_bit(STRIPE_EXPANDING, &sh->state));
708 				list_del_init(&sh->lru);
709 				if (sh->group) {
710 					sh->group->stripes_cnt--;
711 					sh->group = NULL;
712 				}
713 			}
714 			atomic_inc(&sh->count);
715 			spin_unlock(&conf->device_lock);
716 		}
717 	} while (sh == NULL);
718 
719 	spin_unlock_irq(conf->hash_locks + hash);
720 	return sh;
721 }
722 
723 static bool is_full_stripe_write(struct stripe_head *sh)
724 {
725 	BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
726 	return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
727 }
728 
729 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
730 {
731 	local_irq_disable();
732 	if (sh1 > sh2) {
733 		spin_lock(&sh2->stripe_lock);
734 		spin_lock_nested(&sh1->stripe_lock, 1);
735 	} else {
736 		spin_lock(&sh1->stripe_lock);
737 		spin_lock_nested(&sh2->stripe_lock, 1);
738 	}
739 }
740 
741 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
742 {
743 	spin_unlock(&sh1->stripe_lock);
744 	spin_unlock(&sh2->stripe_lock);
745 	local_irq_enable();
746 }
747 
748 /* Only freshly new full stripe normal write stripe can be added to a batch list */
749 static bool stripe_can_batch(struct stripe_head *sh)
750 {
751 	return test_bit(STRIPE_BATCH_READY, &sh->state) &&
752 		!test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
753 		is_full_stripe_write(sh);
754 }
755 
756 /* we only do back search */
757 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
758 {
759 	struct stripe_head *head;
760 	sector_t head_sector, tmp_sec;
761 	int hash;
762 	int dd_idx;
763 
764 	if (!stripe_can_batch(sh))
765 		return;
766 	/* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
767 	tmp_sec = sh->sector;
768 	if (!sector_div(tmp_sec, conf->chunk_sectors))
769 		return;
770 	head_sector = sh->sector - STRIPE_SECTORS;
771 
772 	hash = stripe_hash_locks_hash(head_sector);
773 	spin_lock_irq(conf->hash_locks + hash);
774 	head = __find_stripe(conf, head_sector, conf->generation);
775 	if (head && !atomic_inc_not_zero(&head->count)) {
776 		spin_lock(&conf->device_lock);
777 		if (!atomic_read(&head->count)) {
778 			if (!test_bit(STRIPE_HANDLE, &head->state))
779 				atomic_inc(&conf->active_stripes);
780 			BUG_ON(list_empty(&head->lru) &&
781 			       !test_bit(STRIPE_EXPANDING, &head->state));
782 			list_del_init(&head->lru);
783 			if (head->group) {
784 				head->group->stripes_cnt--;
785 				head->group = NULL;
786 			}
787 		}
788 		atomic_inc(&head->count);
789 		spin_unlock(&conf->device_lock);
790 	}
791 	spin_unlock_irq(conf->hash_locks + hash);
792 
793 	if (!head)
794 		return;
795 	if (!stripe_can_batch(head))
796 		goto out;
797 
798 	lock_two_stripes(head, sh);
799 	/* clear_batch_ready clear the flag */
800 	if (!stripe_can_batch(head) || !stripe_can_batch(sh))
801 		goto unlock_out;
802 
803 	if (sh->batch_head)
804 		goto unlock_out;
805 
806 	dd_idx = 0;
807 	while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
808 		dd_idx++;
809 	if (head->dev[dd_idx].towrite->bi_rw != sh->dev[dd_idx].towrite->bi_rw)
810 		goto unlock_out;
811 
812 	if (head->batch_head) {
813 		spin_lock(&head->batch_head->batch_lock);
814 		/* This batch list is already running */
815 		if (!stripe_can_batch(head)) {
816 			spin_unlock(&head->batch_head->batch_lock);
817 			goto unlock_out;
818 		}
819 
820 		/*
821 		 * at this point, head's BATCH_READY could be cleared, but we
822 		 * can still add the stripe to batch list
823 		 */
824 		list_add(&sh->batch_list, &head->batch_list);
825 		spin_unlock(&head->batch_head->batch_lock);
826 
827 		sh->batch_head = head->batch_head;
828 	} else {
829 		head->batch_head = head;
830 		sh->batch_head = head->batch_head;
831 		spin_lock(&head->batch_lock);
832 		list_add_tail(&sh->batch_list, &head->batch_list);
833 		spin_unlock(&head->batch_lock);
834 	}
835 
836 	if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
837 		if (atomic_dec_return(&conf->preread_active_stripes)
838 		    < IO_THRESHOLD)
839 			md_wakeup_thread(conf->mddev->thread);
840 
841 	if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
842 		int seq = sh->bm_seq;
843 		if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
844 		    sh->batch_head->bm_seq > seq)
845 			seq = sh->batch_head->bm_seq;
846 		set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
847 		sh->batch_head->bm_seq = seq;
848 	}
849 
850 	atomic_inc(&sh->count);
851 unlock_out:
852 	unlock_two_stripes(head, sh);
853 out:
854 	release_stripe(head);
855 }
856 
857 /* Determine if 'data_offset' or 'new_data_offset' should be used
858  * in this stripe_head.
859  */
860 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
861 {
862 	sector_t progress = conf->reshape_progress;
863 	/* Need a memory barrier to make sure we see the value
864 	 * of conf->generation, or ->data_offset that was set before
865 	 * reshape_progress was updated.
866 	 */
867 	smp_rmb();
868 	if (progress == MaxSector)
869 		return 0;
870 	if (sh->generation == conf->generation - 1)
871 		return 0;
872 	/* We are in a reshape, and this is a new-generation stripe,
873 	 * so use new_data_offset.
874 	 */
875 	return 1;
876 }
877 
878 static void
879 raid5_end_read_request(struct bio *bi, int error);
880 static void
881 raid5_end_write_request(struct bio *bi, int error);
882 
883 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
884 {
885 	struct r5conf *conf = sh->raid_conf;
886 	int i, disks = sh->disks;
887 	struct stripe_head *head_sh = sh;
888 
889 	might_sleep();
890 
891 	for (i = disks; i--; ) {
892 		int rw;
893 		int replace_only = 0;
894 		struct bio *bi, *rbi;
895 		struct md_rdev *rdev, *rrdev = NULL;
896 
897 		sh = head_sh;
898 		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
899 			if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
900 				rw = WRITE_FUA;
901 			else
902 				rw = WRITE;
903 			if (test_bit(R5_Discard, &sh->dev[i].flags))
904 				rw |= REQ_DISCARD;
905 		} else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
906 			rw = READ;
907 		else if (test_and_clear_bit(R5_WantReplace,
908 					    &sh->dev[i].flags)) {
909 			rw = WRITE;
910 			replace_only = 1;
911 		} else
912 			continue;
913 		if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
914 			rw |= REQ_SYNC;
915 
916 again:
917 		bi = &sh->dev[i].req;
918 		rbi = &sh->dev[i].rreq; /* For writing to replacement */
919 
920 		rcu_read_lock();
921 		rrdev = rcu_dereference(conf->disks[i].replacement);
922 		smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
923 		rdev = rcu_dereference(conf->disks[i].rdev);
924 		if (!rdev) {
925 			rdev = rrdev;
926 			rrdev = NULL;
927 		}
928 		if (rw & WRITE) {
929 			if (replace_only)
930 				rdev = NULL;
931 			if (rdev == rrdev)
932 				/* We raced and saw duplicates */
933 				rrdev = NULL;
934 		} else {
935 			if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
936 				rdev = rrdev;
937 			rrdev = NULL;
938 		}
939 
940 		if (rdev && test_bit(Faulty, &rdev->flags))
941 			rdev = NULL;
942 		if (rdev)
943 			atomic_inc(&rdev->nr_pending);
944 		if (rrdev && test_bit(Faulty, &rrdev->flags))
945 			rrdev = NULL;
946 		if (rrdev)
947 			atomic_inc(&rrdev->nr_pending);
948 		rcu_read_unlock();
949 
950 		/* We have already checked bad blocks for reads.  Now
951 		 * need to check for writes.  We never accept write errors
952 		 * on the replacement, so we don't to check rrdev.
953 		 */
954 		while ((rw & WRITE) && rdev &&
955 		       test_bit(WriteErrorSeen, &rdev->flags)) {
956 			sector_t first_bad;
957 			int bad_sectors;
958 			int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
959 					      &first_bad, &bad_sectors);
960 			if (!bad)
961 				break;
962 
963 			if (bad < 0) {
964 				set_bit(BlockedBadBlocks, &rdev->flags);
965 				if (!conf->mddev->external &&
966 				    conf->mddev->flags) {
967 					/* It is very unlikely, but we might
968 					 * still need to write out the
969 					 * bad block log - better give it
970 					 * a chance*/
971 					md_check_recovery(conf->mddev);
972 				}
973 				/*
974 				 * Because md_wait_for_blocked_rdev
975 				 * will dec nr_pending, we must
976 				 * increment it first.
977 				 */
978 				atomic_inc(&rdev->nr_pending);
979 				md_wait_for_blocked_rdev(rdev, conf->mddev);
980 			} else {
981 				/* Acknowledged bad block - skip the write */
982 				rdev_dec_pending(rdev, conf->mddev);
983 				rdev = NULL;
984 			}
985 		}
986 
987 		if (rdev) {
988 			if (s->syncing || s->expanding || s->expanded
989 			    || s->replacing)
990 				md_sync_acct(rdev->bdev, STRIPE_SECTORS);
991 
992 			set_bit(STRIPE_IO_STARTED, &sh->state);
993 
994 			bio_reset(bi);
995 			bi->bi_bdev = rdev->bdev;
996 			bi->bi_rw = rw;
997 			bi->bi_end_io = (rw & WRITE)
998 				? raid5_end_write_request
999 				: raid5_end_read_request;
1000 			bi->bi_private = sh;
1001 
1002 			pr_debug("%s: for %llu schedule op %ld on disc %d\n",
1003 				__func__, (unsigned long long)sh->sector,
1004 				bi->bi_rw, i);
1005 			atomic_inc(&sh->count);
1006 			if (sh != head_sh)
1007 				atomic_inc(&head_sh->count);
1008 			if (use_new_offset(conf, sh))
1009 				bi->bi_iter.bi_sector = (sh->sector
1010 						 + rdev->new_data_offset);
1011 			else
1012 				bi->bi_iter.bi_sector = (sh->sector
1013 						 + rdev->data_offset);
1014 			if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1015 				bi->bi_rw |= REQ_NOMERGE;
1016 
1017 			if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1018 				WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1019 			sh->dev[i].vec.bv_page = sh->dev[i].page;
1020 			bi->bi_vcnt = 1;
1021 			bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1022 			bi->bi_io_vec[0].bv_offset = 0;
1023 			bi->bi_iter.bi_size = STRIPE_SIZE;
1024 			/*
1025 			 * If this is discard request, set bi_vcnt 0. We don't
1026 			 * want to confuse SCSI because SCSI will replace payload
1027 			 */
1028 			if (rw & REQ_DISCARD)
1029 				bi->bi_vcnt = 0;
1030 			if (rrdev)
1031 				set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1032 
1033 			if (conf->mddev->gendisk)
1034 				trace_block_bio_remap(bdev_get_queue(bi->bi_bdev),
1035 						      bi, disk_devt(conf->mddev->gendisk),
1036 						      sh->dev[i].sector);
1037 			generic_make_request(bi);
1038 		}
1039 		if (rrdev) {
1040 			if (s->syncing || s->expanding || s->expanded
1041 			    || s->replacing)
1042 				md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1043 
1044 			set_bit(STRIPE_IO_STARTED, &sh->state);
1045 
1046 			bio_reset(rbi);
1047 			rbi->bi_bdev = rrdev->bdev;
1048 			rbi->bi_rw = rw;
1049 			BUG_ON(!(rw & WRITE));
1050 			rbi->bi_end_io = raid5_end_write_request;
1051 			rbi->bi_private = sh;
1052 
1053 			pr_debug("%s: for %llu schedule op %ld on "
1054 				 "replacement disc %d\n",
1055 				__func__, (unsigned long long)sh->sector,
1056 				rbi->bi_rw, i);
1057 			atomic_inc(&sh->count);
1058 			if (sh != head_sh)
1059 				atomic_inc(&head_sh->count);
1060 			if (use_new_offset(conf, sh))
1061 				rbi->bi_iter.bi_sector = (sh->sector
1062 						  + rrdev->new_data_offset);
1063 			else
1064 				rbi->bi_iter.bi_sector = (sh->sector
1065 						  + rrdev->data_offset);
1066 			if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1067 				WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1068 			sh->dev[i].rvec.bv_page = sh->dev[i].page;
1069 			rbi->bi_vcnt = 1;
1070 			rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1071 			rbi->bi_io_vec[0].bv_offset = 0;
1072 			rbi->bi_iter.bi_size = STRIPE_SIZE;
1073 			/*
1074 			 * If this is discard request, set bi_vcnt 0. We don't
1075 			 * want to confuse SCSI because SCSI will replace payload
1076 			 */
1077 			if (rw & REQ_DISCARD)
1078 				rbi->bi_vcnt = 0;
1079 			if (conf->mddev->gendisk)
1080 				trace_block_bio_remap(bdev_get_queue(rbi->bi_bdev),
1081 						      rbi, disk_devt(conf->mddev->gendisk),
1082 						      sh->dev[i].sector);
1083 			generic_make_request(rbi);
1084 		}
1085 		if (!rdev && !rrdev) {
1086 			if (rw & WRITE)
1087 				set_bit(STRIPE_DEGRADED, &sh->state);
1088 			pr_debug("skip op %ld on disc %d for sector %llu\n",
1089 				bi->bi_rw, i, (unsigned long long)sh->sector);
1090 			clear_bit(R5_LOCKED, &sh->dev[i].flags);
1091 			set_bit(STRIPE_HANDLE, &sh->state);
1092 		}
1093 
1094 		if (!head_sh->batch_head)
1095 			continue;
1096 		sh = list_first_entry(&sh->batch_list, struct stripe_head,
1097 				      batch_list);
1098 		if (sh != head_sh)
1099 			goto again;
1100 	}
1101 }
1102 
1103 static struct dma_async_tx_descriptor *
1104 async_copy_data(int frombio, struct bio *bio, struct page **page,
1105 	sector_t sector, struct dma_async_tx_descriptor *tx,
1106 	struct stripe_head *sh)
1107 {
1108 	struct bio_vec bvl;
1109 	struct bvec_iter iter;
1110 	struct page *bio_page;
1111 	int page_offset;
1112 	struct async_submit_ctl submit;
1113 	enum async_tx_flags flags = 0;
1114 
1115 	if (bio->bi_iter.bi_sector >= sector)
1116 		page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1117 	else
1118 		page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1119 
1120 	if (frombio)
1121 		flags |= ASYNC_TX_FENCE;
1122 	init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1123 
1124 	bio_for_each_segment(bvl, bio, iter) {
1125 		int len = bvl.bv_len;
1126 		int clen;
1127 		int b_offset = 0;
1128 
1129 		if (page_offset < 0) {
1130 			b_offset = -page_offset;
1131 			page_offset += b_offset;
1132 			len -= b_offset;
1133 		}
1134 
1135 		if (len > 0 && page_offset + len > STRIPE_SIZE)
1136 			clen = STRIPE_SIZE - page_offset;
1137 		else
1138 			clen = len;
1139 
1140 		if (clen > 0) {
1141 			b_offset += bvl.bv_offset;
1142 			bio_page = bvl.bv_page;
1143 			if (frombio) {
1144 				if (sh->raid_conf->skip_copy &&
1145 				    b_offset == 0 && page_offset == 0 &&
1146 				    clen == STRIPE_SIZE)
1147 					*page = bio_page;
1148 				else
1149 					tx = async_memcpy(*page, bio_page, page_offset,
1150 						  b_offset, clen, &submit);
1151 			} else
1152 				tx = async_memcpy(bio_page, *page, b_offset,
1153 						  page_offset, clen, &submit);
1154 		}
1155 		/* chain the operations */
1156 		submit.depend_tx = tx;
1157 
1158 		if (clen < len) /* hit end of page */
1159 			break;
1160 		page_offset +=  len;
1161 	}
1162 
1163 	return tx;
1164 }
1165 
1166 static void ops_complete_biofill(void *stripe_head_ref)
1167 {
1168 	struct stripe_head *sh = stripe_head_ref;
1169 	struct bio *return_bi = NULL;
1170 	int i;
1171 
1172 	pr_debug("%s: stripe %llu\n", __func__,
1173 		(unsigned long long)sh->sector);
1174 
1175 	/* clear completed biofills */
1176 	for (i = sh->disks; i--; ) {
1177 		struct r5dev *dev = &sh->dev[i];
1178 
1179 		/* acknowledge completion of a biofill operation */
1180 		/* and check if we need to reply to a read request,
1181 		 * new R5_Wantfill requests are held off until
1182 		 * !STRIPE_BIOFILL_RUN
1183 		 */
1184 		if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1185 			struct bio *rbi, *rbi2;
1186 
1187 			BUG_ON(!dev->read);
1188 			rbi = dev->read;
1189 			dev->read = NULL;
1190 			while (rbi && rbi->bi_iter.bi_sector <
1191 				dev->sector + STRIPE_SECTORS) {
1192 				rbi2 = r5_next_bio(rbi, dev->sector);
1193 				if (!raid5_dec_bi_active_stripes(rbi)) {
1194 					rbi->bi_next = return_bi;
1195 					return_bi = rbi;
1196 				}
1197 				rbi = rbi2;
1198 			}
1199 		}
1200 	}
1201 	clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1202 
1203 	return_io(return_bi);
1204 
1205 	set_bit(STRIPE_HANDLE, &sh->state);
1206 	release_stripe(sh);
1207 }
1208 
1209 static void ops_run_biofill(struct stripe_head *sh)
1210 {
1211 	struct dma_async_tx_descriptor *tx = NULL;
1212 	struct async_submit_ctl submit;
1213 	int i;
1214 
1215 	BUG_ON(sh->batch_head);
1216 	pr_debug("%s: stripe %llu\n", __func__,
1217 		(unsigned long long)sh->sector);
1218 
1219 	for (i = sh->disks; i--; ) {
1220 		struct r5dev *dev = &sh->dev[i];
1221 		if (test_bit(R5_Wantfill, &dev->flags)) {
1222 			struct bio *rbi;
1223 			spin_lock_irq(&sh->stripe_lock);
1224 			dev->read = rbi = dev->toread;
1225 			dev->toread = NULL;
1226 			spin_unlock_irq(&sh->stripe_lock);
1227 			while (rbi && rbi->bi_iter.bi_sector <
1228 				dev->sector + STRIPE_SECTORS) {
1229 				tx = async_copy_data(0, rbi, &dev->page,
1230 					dev->sector, tx, sh);
1231 				rbi = r5_next_bio(rbi, dev->sector);
1232 			}
1233 		}
1234 	}
1235 
1236 	atomic_inc(&sh->count);
1237 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1238 	async_trigger_callback(&submit);
1239 }
1240 
1241 static void mark_target_uptodate(struct stripe_head *sh, int target)
1242 {
1243 	struct r5dev *tgt;
1244 
1245 	if (target < 0)
1246 		return;
1247 
1248 	tgt = &sh->dev[target];
1249 	set_bit(R5_UPTODATE, &tgt->flags);
1250 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1251 	clear_bit(R5_Wantcompute, &tgt->flags);
1252 }
1253 
1254 static void ops_complete_compute(void *stripe_head_ref)
1255 {
1256 	struct stripe_head *sh = stripe_head_ref;
1257 
1258 	pr_debug("%s: stripe %llu\n", __func__,
1259 		(unsigned long long)sh->sector);
1260 
1261 	/* mark the computed target(s) as uptodate */
1262 	mark_target_uptodate(sh, sh->ops.target);
1263 	mark_target_uptodate(sh, sh->ops.target2);
1264 
1265 	clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1266 	if (sh->check_state == check_state_compute_run)
1267 		sh->check_state = check_state_compute_result;
1268 	set_bit(STRIPE_HANDLE, &sh->state);
1269 	release_stripe(sh);
1270 }
1271 
1272 /* return a pointer to the address conversion region of the scribble buffer */
1273 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1274 				 struct raid5_percpu *percpu, int i)
1275 {
1276 	void *addr;
1277 
1278 	addr = flex_array_get(percpu->scribble, i);
1279 	return addr + sizeof(struct page *) * (sh->disks + 2);
1280 }
1281 
1282 /* return a pointer to the address conversion region of the scribble buffer */
1283 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1284 {
1285 	void *addr;
1286 
1287 	addr = flex_array_get(percpu->scribble, i);
1288 	return addr;
1289 }
1290 
1291 static struct dma_async_tx_descriptor *
1292 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1293 {
1294 	int disks = sh->disks;
1295 	struct page **xor_srcs = to_addr_page(percpu, 0);
1296 	int target = sh->ops.target;
1297 	struct r5dev *tgt = &sh->dev[target];
1298 	struct page *xor_dest = tgt->page;
1299 	int count = 0;
1300 	struct dma_async_tx_descriptor *tx;
1301 	struct async_submit_ctl submit;
1302 	int i;
1303 
1304 	BUG_ON(sh->batch_head);
1305 
1306 	pr_debug("%s: stripe %llu block: %d\n",
1307 		__func__, (unsigned long long)sh->sector, target);
1308 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1309 
1310 	for (i = disks; i--; )
1311 		if (i != target)
1312 			xor_srcs[count++] = sh->dev[i].page;
1313 
1314 	atomic_inc(&sh->count);
1315 
1316 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1317 			  ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1318 	if (unlikely(count == 1))
1319 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1320 	else
1321 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1322 
1323 	return tx;
1324 }
1325 
1326 /* set_syndrome_sources - populate source buffers for gen_syndrome
1327  * @srcs - (struct page *) array of size sh->disks
1328  * @sh - stripe_head to parse
1329  *
1330  * Populates srcs in proper layout order for the stripe and returns the
1331  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1332  * destination buffer is recorded in srcs[count] and the Q destination
1333  * is recorded in srcs[count+1]].
1334  */
1335 static int set_syndrome_sources(struct page **srcs,
1336 				struct stripe_head *sh,
1337 				int srctype)
1338 {
1339 	int disks = sh->disks;
1340 	int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1341 	int d0_idx = raid6_d0(sh);
1342 	int count;
1343 	int i;
1344 
1345 	for (i = 0; i < disks; i++)
1346 		srcs[i] = NULL;
1347 
1348 	count = 0;
1349 	i = d0_idx;
1350 	do {
1351 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1352 		struct r5dev *dev = &sh->dev[i];
1353 
1354 		if (i == sh->qd_idx || i == sh->pd_idx ||
1355 		    (srctype == SYNDROME_SRC_ALL) ||
1356 		    (srctype == SYNDROME_SRC_WANT_DRAIN &&
1357 		     test_bit(R5_Wantdrain, &dev->flags)) ||
1358 		    (srctype == SYNDROME_SRC_WRITTEN &&
1359 		     dev->written))
1360 			srcs[slot] = sh->dev[i].page;
1361 		i = raid6_next_disk(i, disks);
1362 	} while (i != d0_idx);
1363 
1364 	return syndrome_disks;
1365 }
1366 
1367 static struct dma_async_tx_descriptor *
1368 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1369 {
1370 	int disks = sh->disks;
1371 	struct page **blocks = to_addr_page(percpu, 0);
1372 	int target;
1373 	int qd_idx = sh->qd_idx;
1374 	struct dma_async_tx_descriptor *tx;
1375 	struct async_submit_ctl submit;
1376 	struct r5dev *tgt;
1377 	struct page *dest;
1378 	int i;
1379 	int count;
1380 
1381 	BUG_ON(sh->batch_head);
1382 	if (sh->ops.target < 0)
1383 		target = sh->ops.target2;
1384 	else if (sh->ops.target2 < 0)
1385 		target = sh->ops.target;
1386 	else
1387 		/* we should only have one valid target */
1388 		BUG();
1389 	BUG_ON(target < 0);
1390 	pr_debug("%s: stripe %llu block: %d\n",
1391 		__func__, (unsigned long long)sh->sector, target);
1392 
1393 	tgt = &sh->dev[target];
1394 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1395 	dest = tgt->page;
1396 
1397 	atomic_inc(&sh->count);
1398 
1399 	if (target == qd_idx) {
1400 		count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1401 		blocks[count] = NULL; /* regenerating p is not necessary */
1402 		BUG_ON(blocks[count+1] != dest); /* q should already be set */
1403 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1404 				  ops_complete_compute, sh,
1405 				  to_addr_conv(sh, percpu, 0));
1406 		tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1407 	} else {
1408 		/* Compute any data- or p-drive using XOR */
1409 		count = 0;
1410 		for (i = disks; i-- ; ) {
1411 			if (i == target || i == qd_idx)
1412 				continue;
1413 			blocks[count++] = sh->dev[i].page;
1414 		}
1415 
1416 		init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1417 				  NULL, ops_complete_compute, sh,
1418 				  to_addr_conv(sh, percpu, 0));
1419 		tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1420 	}
1421 
1422 	return tx;
1423 }
1424 
1425 static struct dma_async_tx_descriptor *
1426 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1427 {
1428 	int i, count, disks = sh->disks;
1429 	int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1430 	int d0_idx = raid6_d0(sh);
1431 	int faila = -1, failb = -1;
1432 	int target = sh->ops.target;
1433 	int target2 = sh->ops.target2;
1434 	struct r5dev *tgt = &sh->dev[target];
1435 	struct r5dev *tgt2 = &sh->dev[target2];
1436 	struct dma_async_tx_descriptor *tx;
1437 	struct page **blocks = to_addr_page(percpu, 0);
1438 	struct async_submit_ctl submit;
1439 
1440 	BUG_ON(sh->batch_head);
1441 	pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1442 		 __func__, (unsigned long long)sh->sector, target, target2);
1443 	BUG_ON(target < 0 || target2 < 0);
1444 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1445 	BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1446 
1447 	/* we need to open-code set_syndrome_sources to handle the
1448 	 * slot number conversion for 'faila' and 'failb'
1449 	 */
1450 	for (i = 0; i < disks ; i++)
1451 		blocks[i] = NULL;
1452 	count = 0;
1453 	i = d0_idx;
1454 	do {
1455 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1456 
1457 		blocks[slot] = sh->dev[i].page;
1458 
1459 		if (i == target)
1460 			faila = slot;
1461 		if (i == target2)
1462 			failb = slot;
1463 		i = raid6_next_disk(i, disks);
1464 	} while (i != d0_idx);
1465 
1466 	BUG_ON(faila == failb);
1467 	if (failb < faila)
1468 		swap(faila, failb);
1469 	pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1470 		 __func__, (unsigned long long)sh->sector, faila, failb);
1471 
1472 	atomic_inc(&sh->count);
1473 
1474 	if (failb == syndrome_disks+1) {
1475 		/* Q disk is one of the missing disks */
1476 		if (faila == syndrome_disks) {
1477 			/* Missing P+Q, just recompute */
1478 			init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1479 					  ops_complete_compute, sh,
1480 					  to_addr_conv(sh, percpu, 0));
1481 			return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1482 						  STRIPE_SIZE, &submit);
1483 		} else {
1484 			struct page *dest;
1485 			int data_target;
1486 			int qd_idx = sh->qd_idx;
1487 
1488 			/* Missing D+Q: recompute D from P, then recompute Q */
1489 			if (target == qd_idx)
1490 				data_target = target2;
1491 			else
1492 				data_target = target;
1493 
1494 			count = 0;
1495 			for (i = disks; i-- ; ) {
1496 				if (i == data_target || i == qd_idx)
1497 					continue;
1498 				blocks[count++] = sh->dev[i].page;
1499 			}
1500 			dest = sh->dev[data_target].page;
1501 			init_async_submit(&submit,
1502 					  ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1503 					  NULL, NULL, NULL,
1504 					  to_addr_conv(sh, percpu, 0));
1505 			tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1506 				       &submit);
1507 
1508 			count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1509 			init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1510 					  ops_complete_compute, sh,
1511 					  to_addr_conv(sh, percpu, 0));
1512 			return async_gen_syndrome(blocks, 0, count+2,
1513 						  STRIPE_SIZE, &submit);
1514 		}
1515 	} else {
1516 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1517 				  ops_complete_compute, sh,
1518 				  to_addr_conv(sh, percpu, 0));
1519 		if (failb == syndrome_disks) {
1520 			/* We're missing D+P. */
1521 			return async_raid6_datap_recov(syndrome_disks+2,
1522 						       STRIPE_SIZE, faila,
1523 						       blocks, &submit);
1524 		} else {
1525 			/* We're missing D+D. */
1526 			return async_raid6_2data_recov(syndrome_disks+2,
1527 						       STRIPE_SIZE, faila, failb,
1528 						       blocks, &submit);
1529 		}
1530 	}
1531 }
1532 
1533 static void ops_complete_prexor(void *stripe_head_ref)
1534 {
1535 	struct stripe_head *sh = stripe_head_ref;
1536 
1537 	pr_debug("%s: stripe %llu\n", __func__,
1538 		(unsigned long long)sh->sector);
1539 }
1540 
1541 static struct dma_async_tx_descriptor *
1542 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1543 		struct dma_async_tx_descriptor *tx)
1544 {
1545 	int disks = sh->disks;
1546 	struct page **xor_srcs = to_addr_page(percpu, 0);
1547 	int count = 0, pd_idx = sh->pd_idx, i;
1548 	struct async_submit_ctl submit;
1549 
1550 	/* existing parity data subtracted */
1551 	struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1552 
1553 	BUG_ON(sh->batch_head);
1554 	pr_debug("%s: stripe %llu\n", __func__,
1555 		(unsigned long long)sh->sector);
1556 
1557 	for (i = disks; i--; ) {
1558 		struct r5dev *dev = &sh->dev[i];
1559 		/* Only process blocks that are known to be uptodate */
1560 		if (test_bit(R5_Wantdrain, &dev->flags))
1561 			xor_srcs[count++] = dev->page;
1562 	}
1563 
1564 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1565 			  ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1566 	tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1567 
1568 	return tx;
1569 }
1570 
1571 static struct dma_async_tx_descriptor *
1572 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1573 		struct dma_async_tx_descriptor *tx)
1574 {
1575 	struct page **blocks = to_addr_page(percpu, 0);
1576 	int count;
1577 	struct async_submit_ctl submit;
1578 
1579 	pr_debug("%s: stripe %llu\n", __func__,
1580 		(unsigned long long)sh->sector);
1581 
1582 	count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1583 
1584 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1585 			  ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1586 	tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1587 
1588 	return tx;
1589 }
1590 
1591 static struct dma_async_tx_descriptor *
1592 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1593 {
1594 	int disks = sh->disks;
1595 	int i;
1596 	struct stripe_head *head_sh = sh;
1597 
1598 	pr_debug("%s: stripe %llu\n", __func__,
1599 		(unsigned long long)sh->sector);
1600 
1601 	for (i = disks; i--; ) {
1602 		struct r5dev *dev;
1603 		struct bio *chosen;
1604 
1605 		sh = head_sh;
1606 		if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1607 			struct bio *wbi;
1608 
1609 again:
1610 			dev = &sh->dev[i];
1611 			spin_lock_irq(&sh->stripe_lock);
1612 			chosen = dev->towrite;
1613 			dev->towrite = NULL;
1614 			sh->overwrite_disks = 0;
1615 			BUG_ON(dev->written);
1616 			wbi = dev->written = chosen;
1617 			spin_unlock_irq(&sh->stripe_lock);
1618 			WARN_ON(dev->page != dev->orig_page);
1619 
1620 			while (wbi && wbi->bi_iter.bi_sector <
1621 				dev->sector + STRIPE_SECTORS) {
1622 				if (wbi->bi_rw & REQ_FUA)
1623 					set_bit(R5_WantFUA, &dev->flags);
1624 				if (wbi->bi_rw & REQ_SYNC)
1625 					set_bit(R5_SyncIO, &dev->flags);
1626 				if (wbi->bi_rw & REQ_DISCARD)
1627 					set_bit(R5_Discard, &dev->flags);
1628 				else {
1629 					tx = async_copy_data(1, wbi, &dev->page,
1630 						dev->sector, tx, sh);
1631 					if (dev->page != dev->orig_page) {
1632 						set_bit(R5_SkipCopy, &dev->flags);
1633 						clear_bit(R5_UPTODATE, &dev->flags);
1634 						clear_bit(R5_OVERWRITE, &dev->flags);
1635 					}
1636 				}
1637 				wbi = r5_next_bio(wbi, dev->sector);
1638 			}
1639 
1640 			if (head_sh->batch_head) {
1641 				sh = list_first_entry(&sh->batch_list,
1642 						      struct stripe_head,
1643 						      batch_list);
1644 				if (sh == head_sh)
1645 					continue;
1646 				goto again;
1647 			}
1648 		}
1649 	}
1650 
1651 	return tx;
1652 }
1653 
1654 static void ops_complete_reconstruct(void *stripe_head_ref)
1655 {
1656 	struct stripe_head *sh = stripe_head_ref;
1657 	int disks = sh->disks;
1658 	int pd_idx = sh->pd_idx;
1659 	int qd_idx = sh->qd_idx;
1660 	int i;
1661 	bool fua = false, sync = false, discard = false;
1662 
1663 	pr_debug("%s: stripe %llu\n", __func__,
1664 		(unsigned long long)sh->sector);
1665 
1666 	for (i = disks; i--; ) {
1667 		fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1668 		sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1669 		discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1670 	}
1671 
1672 	for (i = disks; i--; ) {
1673 		struct r5dev *dev = &sh->dev[i];
1674 
1675 		if (dev->written || i == pd_idx || i == qd_idx) {
1676 			if (!discard && !test_bit(R5_SkipCopy, &dev->flags))
1677 				set_bit(R5_UPTODATE, &dev->flags);
1678 			if (fua)
1679 				set_bit(R5_WantFUA, &dev->flags);
1680 			if (sync)
1681 				set_bit(R5_SyncIO, &dev->flags);
1682 		}
1683 	}
1684 
1685 	if (sh->reconstruct_state == reconstruct_state_drain_run)
1686 		sh->reconstruct_state = reconstruct_state_drain_result;
1687 	else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1688 		sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1689 	else {
1690 		BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1691 		sh->reconstruct_state = reconstruct_state_result;
1692 	}
1693 
1694 	set_bit(STRIPE_HANDLE, &sh->state);
1695 	release_stripe(sh);
1696 }
1697 
1698 static void
1699 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1700 		     struct dma_async_tx_descriptor *tx)
1701 {
1702 	int disks = sh->disks;
1703 	struct page **xor_srcs;
1704 	struct async_submit_ctl submit;
1705 	int count, pd_idx = sh->pd_idx, i;
1706 	struct page *xor_dest;
1707 	int prexor = 0;
1708 	unsigned long flags;
1709 	int j = 0;
1710 	struct stripe_head *head_sh = sh;
1711 	int last_stripe;
1712 
1713 	pr_debug("%s: stripe %llu\n", __func__,
1714 		(unsigned long long)sh->sector);
1715 
1716 	for (i = 0; i < sh->disks; i++) {
1717 		if (pd_idx == i)
1718 			continue;
1719 		if (!test_bit(R5_Discard, &sh->dev[i].flags))
1720 			break;
1721 	}
1722 	if (i >= sh->disks) {
1723 		atomic_inc(&sh->count);
1724 		set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1725 		ops_complete_reconstruct(sh);
1726 		return;
1727 	}
1728 again:
1729 	count = 0;
1730 	xor_srcs = to_addr_page(percpu, j);
1731 	/* check if prexor is active which means only process blocks
1732 	 * that are part of a read-modify-write (written)
1733 	 */
1734 	if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1735 		prexor = 1;
1736 		xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1737 		for (i = disks; i--; ) {
1738 			struct r5dev *dev = &sh->dev[i];
1739 			if (head_sh->dev[i].written)
1740 				xor_srcs[count++] = dev->page;
1741 		}
1742 	} else {
1743 		xor_dest = sh->dev[pd_idx].page;
1744 		for (i = disks; i--; ) {
1745 			struct r5dev *dev = &sh->dev[i];
1746 			if (i != pd_idx)
1747 				xor_srcs[count++] = dev->page;
1748 		}
1749 	}
1750 
1751 	/* 1/ if we prexor'd then the dest is reused as a source
1752 	 * 2/ if we did not prexor then we are redoing the parity
1753 	 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1754 	 * for the synchronous xor case
1755 	 */
1756 	last_stripe = !head_sh->batch_head ||
1757 		list_first_entry(&sh->batch_list,
1758 				 struct stripe_head, batch_list) == head_sh;
1759 	if (last_stripe) {
1760 		flags = ASYNC_TX_ACK |
1761 			(prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1762 
1763 		atomic_inc(&head_sh->count);
1764 		init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1765 				  to_addr_conv(sh, percpu, j));
1766 	} else {
1767 		flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1768 		init_async_submit(&submit, flags, tx, NULL, NULL,
1769 				  to_addr_conv(sh, percpu, j));
1770 	}
1771 
1772 	if (unlikely(count == 1))
1773 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1774 	else
1775 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1776 	if (!last_stripe) {
1777 		j++;
1778 		sh = list_first_entry(&sh->batch_list, struct stripe_head,
1779 				      batch_list);
1780 		goto again;
1781 	}
1782 }
1783 
1784 static void
1785 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1786 		     struct dma_async_tx_descriptor *tx)
1787 {
1788 	struct async_submit_ctl submit;
1789 	struct page **blocks;
1790 	int count, i, j = 0;
1791 	struct stripe_head *head_sh = sh;
1792 	int last_stripe;
1793 	int synflags;
1794 	unsigned long txflags;
1795 
1796 	pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1797 
1798 	for (i = 0; i < sh->disks; i++) {
1799 		if (sh->pd_idx == i || sh->qd_idx == i)
1800 			continue;
1801 		if (!test_bit(R5_Discard, &sh->dev[i].flags))
1802 			break;
1803 	}
1804 	if (i >= sh->disks) {
1805 		atomic_inc(&sh->count);
1806 		set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1807 		set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1808 		ops_complete_reconstruct(sh);
1809 		return;
1810 	}
1811 
1812 again:
1813 	blocks = to_addr_page(percpu, j);
1814 
1815 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1816 		synflags = SYNDROME_SRC_WRITTEN;
1817 		txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1818 	} else {
1819 		synflags = SYNDROME_SRC_ALL;
1820 		txflags = ASYNC_TX_ACK;
1821 	}
1822 
1823 	count = set_syndrome_sources(blocks, sh, synflags);
1824 	last_stripe = !head_sh->batch_head ||
1825 		list_first_entry(&sh->batch_list,
1826 				 struct stripe_head, batch_list) == head_sh;
1827 
1828 	if (last_stripe) {
1829 		atomic_inc(&head_sh->count);
1830 		init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1831 				  head_sh, to_addr_conv(sh, percpu, j));
1832 	} else
1833 		init_async_submit(&submit, 0, tx, NULL, NULL,
1834 				  to_addr_conv(sh, percpu, j));
1835 	tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1836 	if (!last_stripe) {
1837 		j++;
1838 		sh = list_first_entry(&sh->batch_list, struct stripe_head,
1839 				      batch_list);
1840 		goto again;
1841 	}
1842 }
1843 
1844 static void ops_complete_check(void *stripe_head_ref)
1845 {
1846 	struct stripe_head *sh = stripe_head_ref;
1847 
1848 	pr_debug("%s: stripe %llu\n", __func__,
1849 		(unsigned long long)sh->sector);
1850 
1851 	sh->check_state = check_state_check_result;
1852 	set_bit(STRIPE_HANDLE, &sh->state);
1853 	release_stripe(sh);
1854 }
1855 
1856 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1857 {
1858 	int disks = sh->disks;
1859 	int pd_idx = sh->pd_idx;
1860 	int qd_idx = sh->qd_idx;
1861 	struct page *xor_dest;
1862 	struct page **xor_srcs = to_addr_page(percpu, 0);
1863 	struct dma_async_tx_descriptor *tx;
1864 	struct async_submit_ctl submit;
1865 	int count;
1866 	int i;
1867 
1868 	pr_debug("%s: stripe %llu\n", __func__,
1869 		(unsigned long long)sh->sector);
1870 
1871 	BUG_ON(sh->batch_head);
1872 	count = 0;
1873 	xor_dest = sh->dev[pd_idx].page;
1874 	xor_srcs[count++] = xor_dest;
1875 	for (i = disks; i--; ) {
1876 		if (i == pd_idx || i == qd_idx)
1877 			continue;
1878 		xor_srcs[count++] = sh->dev[i].page;
1879 	}
1880 
1881 	init_async_submit(&submit, 0, NULL, NULL, NULL,
1882 			  to_addr_conv(sh, percpu, 0));
1883 	tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1884 			   &sh->ops.zero_sum_result, &submit);
1885 
1886 	atomic_inc(&sh->count);
1887 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1888 	tx = async_trigger_callback(&submit);
1889 }
1890 
1891 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1892 {
1893 	struct page **srcs = to_addr_page(percpu, 0);
1894 	struct async_submit_ctl submit;
1895 	int count;
1896 
1897 	pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1898 		(unsigned long long)sh->sector, checkp);
1899 
1900 	BUG_ON(sh->batch_head);
1901 	count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
1902 	if (!checkp)
1903 		srcs[count] = NULL;
1904 
1905 	atomic_inc(&sh->count);
1906 	init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1907 			  sh, to_addr_conv(sh, percpu, 0));
1908 	async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1909 			   &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1910 }
1911 
1912 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1913 {
1914 	int overlap_clear = 0, i, disks = sh->disks;
1915 	struct dma_async_tx_descriptor *tx = NULL;
1916 	struct r5conf *conf = sh->raid_conf;
1917 	int level = conf->level;
1918 	struct raid5_percpu *percpu;
1919 	unsigned long cpu;
1920 
1921 	cpu = get_cpu();
1922 	percpu = per_cpu_ptr(conf->percpu, cpu);
1923 	if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1924 		ops_run_biofill(sh);
1925 		overlap_clear++;
1926 	}
1927 
1928 	if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1929 		if (level < 6)
1930 			tx = ops_run_compute5(sh, percpu);
1931 		else {
1932 			if (sh->ops.target2 < 0 || sh->ops.target < 0)
1933 				tx = ops_run_compute6_1(sh, percpu);
1934 			else
1935 				tx = ops_run_compute6_2(sh, percpu);
1936 		}
1937 		/* terminate the chain if reconstruct is not set to be run */
1938 		if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1939 			async_tx_ack(tx);
1940 	}
1941 
1942 	if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
1943 		if (level < 6)
1944 			tx = ops_run_prexor5(sh, percpu, tx);
1945 		else
1946 			tx = ops_run_prexor6(sh, percpu, tx);
1947 	}
1948 
1949 	if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1950 		tx = ops_run_biodrain(sh, tx);
1951 		overlap_clear++;
1952 	}
1953 
1954 	if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1955 		if (level < 6)
1956 			ops_run_reconstruct5(sh, percpu, tx);
1957 		else
1958 			ops_run_reconstruct6(sh, percpu, tx);
1959 	}
1960 
1961 	if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1962 		if (sh->check_state == check_state_run)
1963 			ops_run_check_p(sh, percpu);
1964 		else if (sh->check_state == check_state_run_q)
1965 			ops_run_check_pq(sh, percpu, 0);
1966 		else if (sh->check_state == check_state_run_pq)
1967 			ops_run_check_pq(sh, percpu, 1);
1968 		else
1969 			BUG();
1970 	}
1971 
1972 	if (overlap_clear && !sh->batch_head)
1973 		for (i = disks; i--; ) {
1974 			struct r5dev *dev = &sh->dev[i];
1975 			if (test_and_clear_bit(R5_Overlap, &dev->flags))
1976 				wake_up(&sh->raid_conf->wait_for_overlap);
1977 		}
1978 	put_cpu();
1979 }
1980 
1981 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp)
1982 {
1983 	struct stripe_head *sh;
1984 
1985 	sh = kmem_cache_zalloc(sc, gfp);
1986 	if (sh) {
1987 		spin_lock_init(&sh->stripe_lock);
1988 		spin_lock_init(&sh->batch_lock);
1989 		INIT_LIST_HEAD(&sh->batch_list);
1990 		INIT_LIST_HEAD(&sh->lru);
1991 		atomic_set(&sh->count, 1);
1992 	}
1993 	return sh;
1994 }
1995 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
1996 {
1997 	struct stripe_head *sh;
1998 
1999 	sh = alloc_stripe(conf->slab_cache, gfp);
2000 	if (!sh)
2001 		return 0;
2002 
2003 	sh->raid_conf = conf;
2004 
2005 	if (grow_buffers(sh, gfp)) {
2006 		shrink_buffers(sh);
2007 		kmem_cache_free(conf->slab_cache, sh);
2008 		return 0;
2009 	}
2010 	sh->hash_lock_index =
2011 		conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2012 	/* we just created an active stripe so... */
2013 	atomic_inc(&conf->active_stripes);
2014 
2015 	release_stripe(sh);
2016 	conf->max_nr_stripes++;
2017 	return 1;
2018 }
2019 
2020 static int grow_stripes(struct r5conf *conf, int num)
2021 {
2022 	struct kmem_cache *sc;
2023 	int devs = max(conf->raid_disks, conf->previous_raid_disks);
2024 
2025 	if (conf->mddev->gendisk)
2026 		sprintf(conf->cache_name[0],
2027 			"raid%d-%s", conf->level, mdname(conf->mddev));
2028 	else
2029 		sprintf(conf->cache_name[0],
2030 			"raid%d-%p", conf->level, conf->mddev);
2031 	sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
2032 
2033 	conf->active_name = 0;
2034 	sc = kmem_cache_create(conf->cache_name[conf->active_name],
2035 			       sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2036 			       0, 0, NULL);
2037 	if (!sc)
2038 		return 1;
2039 	conf->slab_cache = sc;
2040 	conf->pool_size = devs;
2041 	while (num--)
2042 		if (!grow_one_stripe(conf, GFP_KERNEL))
2043 			return 1;
2044 
2045 	return 0;
2046 }
2047 
2048 /**
2049  * scribble_len - return the required size of the scribble region
2050  * @num - total number of disks in the array
2051  *
2052  * The size must be enough to contain:
2053  * 1/ a struct page pointer for each device in the array +2
2054  * 2/ room to convert each entry in (1) to its corresponding dma
2055  *    (dma_map_page()) or page (page_address()) address.
2056  *
2057  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2058  * calculate over all devices (not just the data blocks), using zeros in place
2059  * of the P and Q blocks.
2060  */
2061 static struct flex_array *scribble_alloc(int num, int cnt, gfp_t flags)
2062 {
2063 	struct flex_array *ret;
2064 	size_t len;
2065 
2066 	len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
2067 	ret = flex_array_alloc(len, cnt, flags);
2068 	if (!ret)
2069 		return NULL;
2070 	/* always prealloc all elements, so no locking is required */
2071 	if (flex_array_prealloc(ret, 0, cnt, flags)) {
2072 		flex_array_free(ret);
2073 		return NULL;
2074 	}
2075 	return ret;
2076 }
2077 
2078 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2079 {
2080 	unsigned long cpu;
2081 	int err = 0;
2082 
2083 	mddev_suspend(conf->mddev);
2084 	get_online_cpus();
2085 	for_each_present_cpu(cpu) {
2086 		struct raid5_percpu *percpu;
2087 		struct flex_array *scribble;
2088 
2089 		percpu = per_cpu_ptr(conf->percpu, cpu);
2090 		scribble = scribble_alloc(new_disks,
2091 					  new_sectors / STRIPE_SECTORS,
2092 					  GFP_NOIO);
2093 
2094 		if (scribble) {
2095 			flex_array_free(percpu->scribble);
2096 			percpu->scribble = scribble;
2097 		} else {
2098 			err = -ENOMEM;
2099 			break;
2100 		}
2101 	}
2102 	put_online_cpus();
2103 	mddev_resume(conf->mddev);
2104 	return err;
2105 }
2106 
2107 static int resize_stripes(struct r5conf *conf, int newsize)
2108 {
2109 	/* Make all the stripes able to hold 'newsize' devices.
2110 	 * New slots in each stripe get 'page' set to a new page.
2111 	 *
2112 	 * This happens in stages:
2113 	 * 1/ create a new kmem_cache and allocate the required number of
2114 	 *    stripe_heads.
2115 	 * 2/ gather all the old stripe_heads and transfer the pages across
2116 	 *    to the new stripe_heads.  This will have the side effect of
2117 	 *    freezing the array as once all stripe_heads have been collected,
2118 	 *    no IO will be possible.  Old stripe heads are freed once their
2119 	 *    pages have been transferred over, and the old kmem_cache is
2120 	 *    freed when all stripes are done.
2121 	 * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2122 	 *    we simple return a failre status - no need to clean anything up.
2123 	 * 4/ allocate new pages for the new slots in the new stripe_heads.
2124 	 *    If this fails, we don't bother trying the shrink the
2125 	 *    stripe_heads down again, we just leave them as they are.
2126 	 *    As each stripe_head is processed the new one is released into
2127 	 *    active service.
2128 	 *
2129 	 * Once step2 is started, we cannot afford to wait for a write,
2130 	 * so we use GFP_NOIO allocations.
2131 	 */
2132 	struct stripe_head *osh, *nsh;
2133 	LIST_HEAD(newstripes);
2134 	struct disk_info *ndisks;
2135 	int err;
2136 	struct kmem_cache *sc;
2137 	int i;
2138 	int hash, cnt;
2139 
2140 	if (newsize <= conf->pool_size)
2141 		return 0; /* never bother to shrink */
2142 
2143 	err = md_allow_write(conf->mddev);
2144 	if (err)
2145 		return err;
2146 
2147 	/* Step 1 */
2148 	sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2149 			       sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2150 			       0, 0, NULL);
2151 	if (!sc)
2152 		return -ENOMEM;
2153 
2154 	for (i = conf->max_nr_stripes; i; i--) {
2155 		nsh = alloc_stripe(sc, GFP_KERNEL);
2156 		if (!nsh)
2157 			break;
2158 
2159 		nsh->raid_conf = conf;
2160 		list_add(&nsh->lru, &newstripes);
2161 	}
2162 	if (i) {
2163 		/* didn't get enough, give up */
2164 		while (!list_empty(&newstripes)) {
2165 			nsh = list_entry(newstripes.next, struct stripe_head, lru);
2166 			list_del(&nsh->lru);
2167 			kmem_cache_free(sc, nsh);
2168 		}
2169 		kmem_cache_destroy(sc);
2170 		return -ENOMEM;
2171 	}
2172 	/* Step 2 - Must use GFP_NOIO now.
2173 	 * OK, we have enough stripes, start collecting inactive
2174 	 * stripes and copying them over
2175 	 */
2176 	hash = 0;
2177 	cnt = 0;
2178 	list_for_each_entry(nsh, &newstripes, lru) {
2179 		lock_device_hash_lock(conf, hash);
2180 		wait_event_cmd(conf->wait_for_stripe,
2181 				    !list_empty(conf->inactive_list + hash),
2182 				    unlock_device_hash_lock(conf, hash),
2183 				    lock_device_hash_lock(conf, hash));
2184 		osh = get_free_stripe(conf, hash);
2185 		unlock_device_hash_lock(conf, hash);
2186 
2187 		for(i=0; i<conf->pool_size; i++) {
2188 			nsh->dev[i].page = osh->dev[i].page;
2189 			nsh->dev[i].orig_page = osh->dev[i].page;
2190 		}
2191 		nsh->hash_lock_index = hash;
2192 		kmem_cache_free(conf->slab_cache, osh);
2193 		cnt++;
2194 		if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2195 		    !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2196 			hash++;
2197 			cnt = 0;
2198 		}
2199 	}
2200 	kmem_cache_destroy(conf->slab_cache);
2201 
2202 	/* Step 3.
2203 	 * At this point, we are holding all the stripes so the array
2204 	 * is completely stalled, so now is a good time to resize
2205 	 * conf->disks and the scribble region
2206 	 */
2207 	ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
2208 	if (ndisks) {
2209 		for (i=0; i<conf->raid_disks; i++)
2210 			ndisks[i] = conf->disks[i];
2211 		kfree(conf->disks);
2212 		conf->disks = ndisks;
2213 	} else
2214 		err = -ENOMEM;
2215 
2216 	/* Step 4, return new stripes to service */
2217 	while(!list_empty(&newstripes)) {
2218 		nsh = list_entry(newstripes.next, struct stripe_head, lru);
2219 		list_del_init(&nsh->lru);
2220 
2221 		for (i=conf->raid_disks; i < newsize; i++)
2222 			if (nsh->dev[i].page == NULL) {
2223 				struct page *p = alloc_page(GFP_NOIO);
2224 				nsh->dev[i].page = p;
2225 				nsh->dev[i].orig_page = p;
2226 				if (!p)
2227 					err = -ENOMEM;
2228 			}
2229 		release_stripe(nsh);
2230 	}
2231 	/* critical section pass, GFP_NOIO no longer needed */
2232 
2233 	conf->slab_cache = sc;
2234 	conf->active_name = 1-conf->active_name;
2235 	if (!err)
2236 		conf->pool_size = newsize;
2237 	return err;
2238 }
2239 
2240 static int drop_one_stripe(struct r5conf *conf)
2241 {
2242 	struct stripe_head *sh;
2243 	int hash = (conf->max_nr_stripes - 1) % NR_STRIPE_HASH_LOCKS;
2244 
2245 	spin_lock_irq(conf->hash_locks + hash);
2246 	sh = get_free_stripe(conf, hash);
2247 	spin_unlock_irq(conf->hash_locks + hash);
2248 	if (!sh)
2249 		return 0;
2250 	BUG_ON(atomic_read(&sh->count));
2251 	shrink_buffers(sh);
2252 	kmem_cache_free(conf->slab_cache, sh);
2253 	atomic_dec(&conf->active_stripes);
2254 	conf->max_nr_stripes--;
2255 	return 1;
2256 }
2257 
2258 static void shrink_stripes(struct r5conf *conf)
2259 {
2260 	while (conf->max_nr_stripes &&
2261 	       drop_one_stripe(conf))
2262 		;
2263 
2264 	if (conf->slab_cache)
2265 		kmem_cache_destroy(conf->slab_cache);
2266 	conf->slab_cache = NULL;
2267 }
2268 
2269 static void raid5_end_read_request(struct bio * bi, int error)
2270 {
2271 	struct stripe_head *sh = bi->bi_private;
2272 	struct r5conf *conf = sh->raid_conf;
2273 	int disks = sh->disks, i;
2274 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2275 	char b[BDEVNAME_SIZE];
2276 	struct md_rdev *rdev = NULL;
2277 	sector_t s;
2278 
2279 	for (i=0 ; i<disks; i++)
2280 		if (bi == &sh->dev[i].req)
2281 			break;
2282 
2283 	pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
2284 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
2285 		uptodate);
2286 	if (i == disks) {
2287 		BUG();
2288 		return;
2289 	}
2290 	if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2291 		/* If replacement finished while this request was outstanding,
2292 		 * 'replacement' might be NULL already.
2293 		 * In that case it moved down to 'rdev'.
2294 		 * rdev is not removed until all requests are finished.
2295 		 */
2296 		rdev = conf->disks[i].replacement;
2297 	if (!rdev)
2298 		rdev = conf->disks[i].rdev;
2299 
2300 	if (use_new_offset(conf, sh))
2301 		s = sh->sector + rdev->new_data_offset;
2302 	else
2303 		s = sh->sector + rdev->data_offset;
2304 	if (uptodate) {
2305 		set_bit(R5_UPTODATE, &sh->dev[i].flags);
2306 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2307 			/* Note that this cannot happen on a
2308 			 * replacement device.  We just fail those on
2309 			 * any error
2310 			 */
2311 			printk_ratelimited(
2312 				KERN_INFO
2313 				"md/raid:%s: read error corrected"
2314 				" (%lu sectors at %llu on %s)\n",
2315 				mdname(conf->mddev), STRIPE_SECTORS,
2316 				(unsigned long long)s,
2317 				bdevname(rdev->bdev, b));
2318 			atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2319 			clear_bit(R5_ReadError, &sh->dev[i].flags);
2320 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
2321 		} else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2322 			clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2323 
2324 		if (atomic_read(&rdev->read_errors))
2325 			atomic_set(&rdev->read_errors, 0);
2326 	} else {
2327 		const char *bdn = bdevname(rdev->bdev, b);
2328 		int retry = 0;
2329 		int set_bad = 0;
2330 
2331 		clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2332 		atomic_inc(&rdev->read_errors);
2333 		if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2334 			printk_ratelimited(
2335 				KERN_WARNING
2336 				"md/raid:%s: read error on replacement device "
2337 				"(sector %llu on %s).\n",
2338 				mdname(conf->mddev),
2339 				(unsigned long long)s,
2340 				bdn);
2341 		else if (conf->mddev->degraded >= conf->max_degraded) {
2342 			set_bad = 1;
2343 			printk_ratelimited(
2344 				KERN_WARNING
2345 				"md/raid:%s: read error not correctable "
2346 				"(sector %llu on %s).\n",
2347 				mdname(conf->mddev),
2348 				(unsigned long long)s,
2349 				bdn);
2350 		} else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2351 			/* Oh, no!!! */
2352 			set_bad = 1;
2353 			printk_ratelimited(
2354 				KERN_WARNING
2355 				"md/raid:%s: read error NOT corrected!! "
2356 				"(sector %llu on %s).\n",
2357 				mdname(conf->mddev),
2358 				(unsigned long long)s,
2359 				bdn);
2360 		} else if (atomic_read(&rdev->read_errors)
2361 			 > conf->max_nr_stripes)
2362 			printk(KERN_WARNING
2363 			       "md/raid:%s: Too many read errors, failing device %s.\n",
2364 			       mdname(conf->mddev), bdn);
2365 		else
2366 			retry = 1;
2367 		if (set_bad && test_bit(In_sync, &rdev->flags)
2368 		    && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2369 			retry = 1;
2370 		if (retry)
2371 			if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2372 				set_bit(R5_ReadError, &sh->dev[i].flags);
2373 				clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2374 			} else
2375 				set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2376 		else {
2377 			clear_bit(R5_ReadError, &sh->dev[i].flags);
2378 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
2379 			if (!(set_bad
2380 			      && test_bit(In_sync, &rdev->flags)
2381 			      && rdev_set_badblocks(
2382 				      rdev, sh->sector, STRIPE_SECTORS, 0)))
2383 				md_error(conf->mddev, rdev);
2384 		}
2385 	}
2386 	rdev_dec_pending(rdev, conf->mddev);
2387 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
2388 	set_bit(STRIPE_HANDLE, &sh->state);
2389 	release_stripe(sh);
2390 }
2391 
2392 static void raid5_end_write_request(struct bio *bi, int error)
2393 {
2394 	struct stripe_head *sh = bi->bi_private;
2395 	struct r5conf *conf = sh->raid_conf;
2396 	int disks = sh->disks, i;
2397 	struct md_rdev *uninitialized_var(rdev);
2398 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
2399 	sector_t first_bad;
2400 	int bad_sectors;
2401 	int replacement = 0;
2402 
2403 	for (i = 0 ; i < disks; i++) {
2404 		if (bi == &sh->dev[i].req) {
2405 			rdev = conf->disks[i].rdev;
2406 			break;
2407 		}
2408 		if (bi == &sh->dev[i].rreq) {
2409 			rdev = conf->disks[i].replacement;
2410 			if (rdev)
2411 				replacement = 1;
2412 			else
2413 				/* rdev was removed and 'replacement'
2414 				 * replaced it.  rdev is not removed
2415 				 * until all requests are finished.
2416 				 */
2417 				rdev = conf->disks[i].rdev;
2418 			break;
2419 		}
2420 	}
2421 	pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
2422 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
2423 		uptodate);
2424 	if (i == disks) {
2425 		BUG();
2426 		return;
2427 	}
2428 
2429 	if (replacement) {
2430 		if (!uptodate)
2431 			md_error(conf->mddev, rdev);
2432 		else if (is_badblock(rdev, sh->sector,
2433 				     STRIPE_SECTORS,
2434 				     &first_bad, &bad_sectors))
2435 			set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2436 	} else {
2437 		if (!uptodate) {
2438 			set_bit(STRIPE_DEGRADED, &sh->state);
2439 			set_bit(WriteErrorSeen, &rdev->flags);
2440 			set_bit(R5_WriteError, &sh->dev[i].flags);
2441 			if (!test_and_set_bit(WantReplacement, &rdev->flags))
2442 				set_bit(MD_RECOVERY_NEEDED,
2443 					&rdev->mddev->recovery);
2444 		} else if (is_badblock(rdev, sh->sector,
2445 				       STRIPE_SECTORS,
2446 				       &first_bad, &bad_sectors)) {
2447 			set_bit(R5_MadeGood, &sh->dev[i].flags);
2448 			if (test_bit(R5_ReadError, &sh->dev[i].flags))
2449 				/* That was a successful write so make
2450 				 * sure it looks like we already did
2451 				 * a re-write.
2452 				 */
2453 				set_bit(R5_ReWrite, &sh->dev[i].flags);
2454 		}
2455 	}
2456 	rdev_dec_pending(rdev, conf->mddev);
2457 
2458 	if (sh->batch_head && !uptodate && !replacement)
2459 		set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2460 
2461 	if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2462 		clear_bit(R5_LOCKED, &sh->dev[i].flags);
2463 	set_bit(STRIPE_HANDLE, &sh->state);
2464 	release_stripe(sh);
2465 
2466 	if (sh->batch_head && sh != sh->batch_head)
2467 		release_stripe(sh->batch_head);
2468 }
2469 
2470 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
2471 
2472 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
2473 {
2474 	struct r5dev *dev = &sh->dev[i];
2475 
2476 	bio_init(&dev->req);
2477 	dev->req.bi_io_vec = &dev->vec;
2478 	dev->req.bi_max_vecs = 1;
2479 	dev->req.bi_private = sh;
2480 
2481 	bio_init(&dev->rreq);
2482 	dev->rreq.bi_io_vec = &dev->rvec;
2483 	dev->rreq.bi_max_vecs = 1;
2484 	dev->rreq.bi_private = sh;
2485 
2486 	dev->flags = 0;
2487 	dev->sector = compute_blocknr(sh, i, previous);
2488 }
2489 
2490 static void error(struct mddev *mddev, struct md_rdev *rdev)
2491 {
2492 	char b[BDEVNAME_SIZE];
2493 	struct r5conf *conf = mddev->private;
2494 	unsigned long flags;
2495 	pr_debug("raid456: error called\n");
2496 
2497 	spin_lock_irqsave(&conf->device_lock, flags);
2498 	clear_bit(In_sync, &rdev->flags);
2499 	mddev->degraded = calc_degraded(conf);
2500 	spin_unlock_irqrestore(&conf->device_lock, flags);
2501 	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2502 
2503 	set_bit(Blocked, &rdev->flags);
2504 	set_bit(Faulty, &rdev->flags);
2505 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
2506 	printk(KERN_ALERT
2507 	       "md/raid:%s: Disk failure on %s, disabling device.\n"
2508 	       "md/raid:%s: Operation continuing on %d devices.\n",
2509 	       mdname(mddev),
2510 	       bdevname(rdev->bdev, b),
2511 	       mdname(mddev),
2512 	       conf->raid_disks - mddev->degraded);
2513 }
2514 
2515 /*
2516  * Input: a 'big' sector number,
2517  * Output: index of the data and parity disk, and the sector # in them.
2518  */
2519 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2520 				     int previous, int *dd_idx,
2521 				     struct stripe_head *sh)
2522 {
2523 	sector_t stripe, stripe2;
2524 	sector_t chunk_number;
2525 	unsigned int chunk_offset;
2526 	int pd_idx, qd_idx;
2527 	int ddf_layout = 0;
2528 	sector_t new_sector;
2529 	int algorithm = previous ? conf->prev_algo
2530 				 : conf->algorithm;
2531 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2532 					 : conf->chunk_sectors;
2533 	int raid_disks = previous ? conf->previous_raid_disks
2534 				  : conf->raid_disks;
2535 	int data_disks = raid_disks - conf->max_degraded;
2536 
2537 	/* First compute the information on this sector */
2538 
2539 	/*
2540 	 * Compute the chunk number and the sector offset inside the chunk
2541 	 */
2542 	chunk_offset = sector_div(r_sector, sectors_per_chunk);
2543 	chunk_number = r_sector;
2544 
2545 	/*
2546 	 * Compute the stripe number
2547 	 */
2548 	stripe = chunk_number;
2549 	*dd_idx = sector_div(stripe, data_disks);
2550 	stripe2 = stripe;
2551 	/*
2552 	 * Select the parity disk based on the user selected algorithm.
2553 	 */
2554 	pd_idx = qd_idx = -1;
2555 	switch(conf->level) {
2556 	case 4:
2557 		pd_idx = data_disks;
2558 		break;
2559 	case 5:
2560 		switch (algorithm) {
2561 		case ALGORITHM_LEFT_ASYMMETRIC:
2562 			pd_idx = data_disks - sector_div(stripe2, raid_disks);
2563 			if (*dd_idx >= pd_idx)
2564 				(*dd_idx)++;
2565 			break;
2566 		case ALGORITHM_RIGHT_ASYMMETRIC:
2567 			pd_idx = sector_div(stripe2, raid_disks);
2568 			if (*dd_idx >= pd_idx)
2569 				(*dd_idx)++;
2570 			break;
2571 		case ALGORITHM_LEFT_SYMMETRIC:
2572 			pd_idx = data_disks - sector_div(stripe2, raid_disks);
2573 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2574 			break;
2575 		case ALGORITHM_RIGHT_SYMMETRIC:
2576 			pd_idx = sector_div(stripe2, raid_disks);
2577 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2578 			break;
2579 		case ALGORITHM_PARITY_0:
2580 			pd_idx = 0;
2581 			(*dd_idx)++;
2582 			break;
2583 		case ALGORITHM_PARITY_N:
2584 			pd_idx = data_disks;
2585 			break;
2586 		default:
2587 			BUG();
2588 		}
2589 		break;
2590 	case 6:
2591 
2592 		switch (algorithm) {
2593 		case ALGORITHM_LEFT_ASYMMETRIC:
2594 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2595 			qd_idx = pd_idx + 1;
2596 			if (pd_idx == raid_disks-1) {
2597 				(*dd_idx)++;	/* Q D D D P */
2598 				qd_idx = 0;
2599 			} else if (*dd_idx >= pd_idx)
2600 				(*dd_idx) += 2; /* D D P Q D */
2601 			break;
2602 		case ALGORITHM_RIGHT_ASYMMETRIC:
2603 			pd_idx = sector_div(stripe2, raid_disks);
2604 			qd_idx = pd_idx + 1;
2605 			if (pd_idx == raid_disks-1) {
2606 				(*dd_idx)++;	/* Q D D D P */
2607 				qd_idx = 0;
2608 			} else if (*dd_idx >= pd_idx)
2609 				(*dd_idx) += 2; /* D D P Q D */
2610 			break;
2611 		case ALGORITHM_LEFT_SYMMETRIC:
2612 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2613 			qd_idx = (pd_idx + 1) % raid_disks;
2614 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2615 			break;
2616 		case ALGORITHM_RIGHT_SYMMETRIC:
2617 			pd_idx = sector_div(stripe2, raid_disks);
2618 			qd_idx = (pd_idx + 1) % raid_disks;
2619 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2620 			break;
2621 
2622 		case ALGORITHM_PARITY_0:
2623 			pd_idx = 0;
2624 			qd_idx = 1;
2625 			(*dd_idx) += 2;
2626 			break;
2627 		case ALGORITHM_PARITY_N:
2628 			pd_idx = data_disks;
2629 			qd_idx = data_disks + 1;
2630 			break;
2631 
2632 		case ALGORITHM_ROTATING_ZERO_RESTART:
2633 			/* Exactly the same as RIGHT_ASYMMETRIC, but or
2634 			 * of blocks for computing Q is different.
2635 			 */
2636 			pd_idx = sector_div(stripe2, raid_disks);
2637 			qd_idx = pd_idx + 1;
2638 			if (pd_idx == raid_disks-1) {
2639 				(*dd_idx)++;	/* Q D D D P */
2640 				qd_idx = 0;
2641 			} else if (*dd_idx >= pd_idx)
2642 				(*dd_idx) += 2; /* D D P Q D */
2643 			ddf_layout = 1;
2644 			break;
2645 
2646 		case ALGORITHM_ROTATING_N_RESTART:
2647 			/* Same a left_asymmetric, by first stripe is
2648 			 * D D D P Q  rather than
2649 			 * Q D D D P
2650 			 */
2651 			stripe2 += 1;
2652 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2653 			qd_idx = pd_idx + 1;
2654 			if (pd_idx == raid_disks-1) {
2655 				(*dd_idx)++;	/* Q D D D P */
2656 				qd_idx = 0;
2657 			} else if (*dd_idx >= pd_idx)
2658 				(*dd_idx) += 2; /* D D P Q D */
2659 			ddf_layout = 1;
2660 			break;
2661 
2662 		case ALGORITHM_ROTATING_N_CONTINUE:
2663 			/* Same as left_symmetric but Q is before P */
2664 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2665 			qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2666 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2667 			ddf_layout = 1;
2668 			break;
2669 
2670 		case ALGORITHM_LEFT_ASYMMETRIC_6:
2671 			/* RAID5 left_asymmetric, with Q on last device */
2672 			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2673 			if (*dd_idx >= pd_idx)
2674 				(*dd_idx)++;
2675 			qd_idx = raid_disks - 1;
2676 			break;
2677 
2678 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
2679 			pd_idx = sector_div(stripe2, raid_disks-1);
2680 			if (*dd_idx >= pd_idx)
2681 				(*dd_idx)++;
2682 			qd_idx = raid_disks - 1;
2683 			break;
2684 
2685 		case ALGORITHM_LEFT_SYMMETRIC_6:
2686 			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2687 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2688 			qd_idx = raid_disks - 1;
2689 			break;
2690 
2691 		case ALGORITHM_RIGHT_SYMMETRIC_6:
2692 			pd_idx = sector_div(stripe2, raid_disks-1);
2693 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2694 			qd_idx = raid_disks - 1;
2695 			break;
2696 
2697 		case ALGORITHM_PARITY_0_6:
2698 			pd_idx = 0;
2699 			(*dd_idx)++;
2700 			qd_idx = raid_disks - 1;
2701 			break;
2702 
2703 		default:
2704 			BUG();
2705 		}
2706 		break;
2707 	}
2708 
2709 	if (sh) {
2710 		sh->pd_idx = pd_idx;
2711 		sh->qd_idx = qd_idx;
2712 		sh->ddf_layout = ddf_layout;
2713 	}
2714 	/*
2715 	 * Finally, compute the new sector number
2716 	 */
2717 	new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2718 	return new_sector;
2719 }
2720 
2721 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
2722 {
2723 	struct r5conf *conf = sh->raid_conf;
2724 	int raid_disks = sh->disks;
2725 	int data_disks = raid_disks - conf->max_degraded;
2726 	sector_t new_sector = sh->sector, check;
2727 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2728 					 : conf->chunk_sectors;
2729 	int algorithm = previous ? conf->prev_algo
2730 				 : conf->algorithm;
2731 	sector_t stripe;
2732 	int chunk_offset;
2733 	sector_t chunk_number;
2734 	int dummy1, dd_idx = i;
2735 	sector_t r_sector;
2736 	struct stripe_head sh2;
2737 
2738 	chunk_offset = sector_div(new_sector, sectors_per_chunk);
2739 	stripe = new_sector;
2740 
2741 	if (i == sh->pd_idx)
2742 		return 0;
2743 	switch(conf->level) {
2744 	case 4: break;
2745 	case 5:
2746 		switch (algorithm) {
2747 		case ALGORITHM_LEFT_ASYMMETRIC:
2748 		case ALGORITHM_RIGHT_ASYMMETRIC:
2749 			if (i > sh->pd_idx)
2750 				i--;
2751 			break;
2752 		case ALGORITHM_LEFT_SYMMETRIC:
2753 		case ALGORITHM_RIGHT_SYMMETRIC:
2754 			if (i < sh->pd_idx)
2755 				i += raid_disks;
2756 			i -= (sh->pd_idx + 1);
2757 			break;
2758 		case ALGORITHM_PARITY_0:
2759 			i -= 1;
2760 			break;
2761 		case ALGORITHM_PARITY_N:
2762 			break;
2763 		default:
2764 			BUG();
2765 		}
2766 		break;
2767 	case 6:
2768 		if (i == sh->qd_idx)
2769 			return 0; /* It is the Q disk */
2770 		switch (algorithm) {
2771 		case ALGORITHM_LEFT_ASYMMETRIC:
2772 		case ALGORITHM_RIGHT_ASYMMETRIC:
2773 		case ALGORITHM_ROTATING_ZERO_RESTART:
2774 		case ALGORITHM_ROTATING_N_RESTART:
2775 			if (sh->pd_idx == raid_disks-1)
2776 				i--;	/* Q D D D P */
2777 			else if (i > sh->pd_idx)
2778 				i -= 2; /* D D P Q D */
2779 			break;
2780 		case ALGORITHM_LEFT_SYMMETRIC:
2781 		case ALGORITHM_RIGHT_SYMMETRIC:
2782 			if (sh->pd_idx == raid_disks-1)
2783 				i--; /* Q D D D P */
2784 			else {
2785 				/* D D P Q D */
2786 				if (i < sh->pd_idx)
2787 					i += raid_disks;
2788 				i -= (sh->pd_idx + 2);
2789 			}
2790 			break;
2791 		case ALGORITHM_PARITY_0:
2792 			i -= 2;
2793 			break;
2794 		case ALGORITHM_PARITY_N:
2795 			break;
2796 		case ALGORITHM_ROTATING_N_CONTINUE:
2797 			/* Like left_symmetric, but P is before Q */
2798 			if (sh->pd_idx == 0)
2799 				i--;	/* P D D D Q */
2800 			else {
2801 				/* D D Q P D */
2802 				if (i < sh->pd_idx)
2803 					i += raid_disks;
2804 				i -= (sh->pd_idx + 1);
2805 			}
2806 			break;
2807 		case ALGORITHM_LEFT_ASYMMETRIC_6:
2808 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
2809 			if (i > sh->pd_idx)
2810 				i--;
2811 			break;
2812 		case ALGORITHM_LEFT_SYMMETRIC_6:
2813 		case ALGORITHM_RIGHT_SYMMETRIC_6:
2814 			if (i < sh->pd_idx)
2815 				i += data_disks + 1;
2816 			i -= (sh->pd_idx + 1);
2817 			break;
2818 		case ALGORITHM_PARITY_0_6:
2819 			i -= 1;
2820 			break;
2821 		default:
2822 			BUG();
2823 		}
2824 		break;
2825 	}
2826 
2827 	chunk_number = stripe * data_disks + i;
2828 	r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2829 
2830 	check = raid5_compute_sector(conf, r_sector,
2831 				     previous, &dummy1, &sh2);
2832 	if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2833 		|| sh2.qd_idx != sh->qd_idx) {
2834 		printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2835 		       mdname(conf->mddev));
2836 		return 0;
2837 	}
2838 	return r_sector;
2839 }
2840 
2841 static void
2842 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2843 			 int rcw, int expand)
2844 {
2845 	int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
2846 	struct r5conf *conf = sh->raid_conf;
2847 	int level = conf->level;
2848 
2849 	if (rcw) {
2850 
2851 		for (i = disks; i--; ) {
2852 			struct r5dev *dev = &sh->dev[i];
2853 
2854 			if (dev->towrite) {
2855 				set_bit(R5_LOCKED, &dev->flags);
2856 				set_bit(R5_Wantdrain, &dev->flags);
2857 				if (!expand)
2858 					clear_bit(R5_UPTODATE, &dev->flags);
2859 				s->locked++;
2860 			}
2861 		}
2862 		/* if we are not expanding this is a proper write request, and
2863 		 * there will be bios with new data to be drained into the
2864 		 * stripe cache
2865 		 */
2866 		if (!expand) {
2867 			if (!s->locked)
2868 				/* False alarm, nothing to do */
2869 				return;
2870 			sh->reconstruct_state = reconstruct_state_drain_run;
2871 			set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2872 		} else
2873 			sh->reconstruct_state = reconstruct_state_run;
2874 
2875 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2876 
2877 		if (s->locked + conf->max_degraded == disks)
2878 			if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2879 				atomic_inc(&conf->pending_full_writes);
2880 	} else {
2881 		BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2882 			test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2883 		BUG_ON(level == 6 &&
2884 			(!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
2885 			   test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
2886 
2887 		for (i = disks; i--; ) {
2888 			struct r5dev *dev = &sh->dev[i];
2889 			if (i == pd_idx || i == qd_idx)
2890 				continue;
2891 
2892 			if (dev->towrite &&
2893 			    (test_bit(R5_UPTODATE, &dev->flags) ||
2894 			     test_bit(R5_Wantcompute, &dev->flags))) {
2895 				set_bit(R5_Wantdrain, &dev->flags);
2896 				set_bit(R5_LOCKED, &dev->flags);
2897 				clear_bit(R5_UPTODATE, &dev->flags);
2898 				s->locked++;
2899 			}
2900 		}
2901 		if (!s->locked)
2902 			/* False alarm - nothing to do */
2903 			return;
2904 		sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2905 		set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2906 		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2907 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2908 	}
2909 
2910 	/* keep the parity disk(s) locked while asynchronous operations
2911 	 * are in flight
2912 	 */
2913 	set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2914 	clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2915 	s->locked++;
2916 
2917 	if (level == 6) {
2918 		int qd_idx = sh->qd_idx;
2919 		struct r5dev *dev = &sh->dev[qd_idx];
2920 
2921 		set_bit(R5_LOCKED, &dev->flags);
2922 		clear_bit(R5_UPTODATE, &dev->flags);
2923 		s->locked++;
2924 	}
2925 
2926 	pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2927 		__func__, (unsigned long long)sh->sector,
2928 		s->locked, s->ops_request);
2929 }
2930 
2931 /*
2932  * Each stripe/dev can have one or more bion attached.
2933  * toread/towrite point to the first in a chain.
2934  * The bi_next chain must be in order.
2935  */
2936 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
2937 			  int forwrite, int previous)
2938 {
2939 	struct bio **bip;
2940 	struct r5conf *conf = sh->raid_conf;
2941 	int firstwrite=0;
2942 
2943 	pr_debug("adding bi b#%llu to stripe s#%llu\n",
2944 		(unsigned long long)bi->bi_iter.bi_sector,
2945 		(unsigned long long)sh->sector);
2946 
2947 	/*
2948 	 * If several bio share a stripe. The bio bi_phys_segments acts as a
2949 	 * reference count to avoid race. The reference count should already be
2950 	 * increased before this function is called (for example, in
2951 	 * make_request()), so other bio sharing this stripe will not free the
2952 	 * stripe. If a stripe is owned by one stripe, the stripe lock will
2953 	 * protect it.
2954 	 */
2955 	spin_lock_irq(&sh->stripe_lock);
2956 	/* Don't allow new IO added to stripes in batch list */
2957 	if (sh->batch_head)
2958 		goto overlap;
2959 	if (forwrite) {
2960 		bip = &sh->dev[dd_idx].towrite;
2961 		if (*bip == NULL)
2962 			firstwrite = 1;
2963 	} else
2964 		bip = &sh->dev[dd_idx].toread;
2965 	while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
2966 		if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
2967 			goto overlap;
2968 		bip = & (*bip)->bi_next;
2969 	}
2970 	if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
2971 		goto overlap;
2972 
2973 	if (!forwrite || previous)
2974 		clear_bit(STRIPE_BATCH_READY, &sh->state);
2975 
2976 	BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2977 	if (*bip)
2978 		bi->bi_next = *bip;
2979 	*bip = bi;
2980 	raid5_inc_bi_active_stripes(bi);
2981 
2982 	if (forwrite) {
2983 		/* check if page is covered */
2984 		sector_t sector = sh->dev[dd_idx].sector;
2985 		for (bi=sh->dev[dd_idx].towrite;
2986 		     sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2987 			     bi && bi->bi_iter.bi_sector <= sector;
2988 		     bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2989 			if (bio_end_sector(bi) >= sector)
2990 				sector = bio_end_sector(bi);
2991 		}
2992 		if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2993 			if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
2994 				sh->overwrite_disks++;
2995 	}
2996 
2997 	pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2998 		(unsigned long long)(*bip)->bi_iter.bi_sector,
2999 		(unsigned long long)sh->sector, dd_idx);
3000 
3001 	if (conf->mddev->bitmap && firstwrite) {
3002 		/* Cannot hold spinlock over bitmap_startwrite,
3003 		 * but must ensure this isn't added to a batch until
3004 		 * we have added to the bitmap and set bm_seq.
3005 		 * So set STRIPE_BITMAP_PENDING to prevent
3006 		 * batching.
3007 		 * If multiple add_stripe_bio() calls race here they
3008 		 * much all set STRIPE_BITMAP_PENDING.  So only the first one
3009 		 * to complete "bitmap_startwrite" gets to set
3010 		 * STRIPE_BIT_DELAY.  This is important as once a stripe
3011 		 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3012 		 * any more.
3013 		 */
3014 		set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3015 		spin_unlock_irq(&sh->stripe_lock);
3016 		bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3017 				  STRIPE_SECTORS, 0);
3018 		spin_lock_irq(&sh->stripe_lock);
3019 		clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3020 		if (!sh->batch_head) {
3021 			sh->bm_seq = conf->seq_flush+1;
3022 			set_bit(STRIPE_BIT_DELAY, &sh->state);
3023 		}
3024 	}
3025 	spin_unlock_irq(&sh->stripe_lock);
3026 
3027 	if (stripe_can_batch(sh))
3028 		stripe_add_to_batch_list(conf, sh);
3029 	return 1;
3030 
3031  overlap:
3032 	set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3033 	spin_unlock_irq(&sh->stripe_lock);
3034 	return 0;
3035 }
3036 
3037 static void end_reshape(struct r5conf *conf);
3038 
3039 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3040 			    struct stripe_head *sh)
3041 {
3042 	int sectors_per_chunk =
3043 		previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3044 	int dd_idx;
3045 	int chunk_offset = sector_div(stripe, sectors_per_chunk);
3046 	int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3047 
3048 	raid5_compute_sector(conf,
3049 			     stripe * (disks - conf->max_degraded)
3050 			     *sectors_per_chunk + chunk_offset,
3051 			     previous,
3052 			     &dd_idx, sh);
3053 }
3054 
3055 static void
3056 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3057 				struct stripe_head_state *s, int disks,
3058 				struct bio **return_bi)
3059 {
3060 	int i;
3061 	BUG_ON(sh->batch_head);
3062 	for (i = disks; i--; ) {
3063 		struct bio *bi;
3064 		int bitmap_end = 0;
3065 
3066 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3067 			struct md_rdev *rdev;
3068 			rcu_read_lock();
3069 			rdev = rcu_dereference(conf->disks[i].rdev);
3070 			if (rdev && test_bit(In_sync, &rdev->flags))
3071 				atomic_inc(&rdev->nr_pending);
3072 			else
3073 				rdev = NULL;
3074 			rcu_read_unlock();
3075 			if (rdev) {
3076 				if (!rdev_set_badblocks(
3077 					    rdev,
3078 					    sh->sector,
3079 					    STRIPE_SECTORS, 0))
3080 					md_error(conf->mddev, rdev);
3081 				rdev_dec_pending(rdev, conf->mddev);
3082 			}
3083 		}
3084 		spin_lock_irq(&sh->stripe_lock);
3085 		/* fail all writes first */
3086 		bi = sh->dev[i].towrite;
3087 		sh->dev[i].towrite = NULL;
3088 		sh->overwrite_disks = 0;
3089 		spin_unlock_irq(&sh->stripe_lock);
3090 		if (bi)
3091 			bitmap_end = 1;
3092 
3093 		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3094 			wake_up(&conf->wait_for_overlap);
3095 
3096 		while (bi && bi->bi_iter.bi_sector <
3097 			sh->dev[i].sector + STRIPE_SECTORS) {
3098 			struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3099 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
3100 			if (!raid5_dec_bi_active_stripes(bi)) {
3101 				md_write_end(conf->mddev);
3102 				bi->bi_next = *return_bi;
3103 				*return_bi = bi;
3104 			}
3105 			bi = nextbi;
3106 		}
3107 		if (bitmap_end)
3108 			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3109 				STRIPE_SECTORS, 0, 0);
3110 		bitmap_end = 0;
3111 		/* and fail all 'written' */
3112 		bi = sh->dev[i].written;
3113 		sh->dev[i].written = NULL;
3114 		if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3115 			WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3116 			sh->dev[i].page = sh->dev[i].orig_page;
3117 		}
3118 
3119 		if (bi) bitmap_end = 1;
3120 		while (bi && bi->bi_iter.bi_sector <
3121 		       sh->dev[i].sector + STRIPE_SECTORS) {
3122 			struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3123 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
3124 			if (!raid5_dec_bi_active_stripes(bi)) {
3125 				md_write_end(conf->mddev);
3126 				bi->bi_next = *return_bi;
3127 				*return_bi = bi;
3128 			}
3129 			bi = bi2;
3130 		}
3131 
3132 		/* fail any reads if this device is non-operational and
3133 		 * the data has not reached the cache yet.
3134 		 */
3135 		if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3136 		    (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3137 		      test_bit(R5_ReadError, &sh->dev[i].flags))) {
3138 			spin_lock_irq(&sh->stripe_lock);
3139 			bi = sh->dev[i].toread;
3140 			sh->dev[i].toread = NULL;
3141 			spin_unlock_irq(&sh->stripe_lock);
3142 			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3143 				wake_up(&conf->wait_for_overlap);
3144 			while (bi && bi->bi_iter.bi_sector <
3145 			       sh->dev[i].sector + STRIPE_SECTORS) {
3146 				struct bio *nextbi =
3147 					r5_next_bio(bi, sh->dev[i].sector);
3148 				clear_bit(BIO_UPTODATE, &bi->bi_flags);
3149 				if (!raid5_dec_bi_active_stripes(bi)) {
3150 					bi->bi_next = *return_bi;
3151 					*return_bi = bi;
3152 				}
3153 				bi = nextbi;
3154 			}
3155 		}
3156 		if (bitmap_end)
3157 			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3158 					STRIPE_SECTORS, 0, 0);
3159 		/* If we were in the middle of a write the parity block might
3160 		 * still be locked - so just clear all R5_LOCKED flags
3161 		 */
3162 		clear_bit(R5_LOCKED, &sh->dev[i].flags);
3163 	}
3164 
3165 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3166 		if (atomic_dec_and_test(&conf->pending_full_writes))
3167 			md_wakeup_thread(conf->mddev->thread);
3168 }
3169 
3170 static void
3171 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3172 		   struct stripe_head_state *s)
3173 {
3174 	int abort = 0;
3175 	int i;
3176 
3177 	BUG_ON(sh->batch_head);
3178 	clear_bit(STRIPE_SYNCING, &sh->state);
3179 	if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3180 		wake_up(&conf->wait_for_overlap);
3181 	s->syncing = 0;
3182 	s->replacing = 0;
3183 	/* There is nothing more to do for sync/check/repair.
3184 	 * Don't even need to abort as that is handled elsewhere
3185 	 * if needed, and not always wanted e.g. if there is a known
3186 	 * bad block here.
3187 	 * For recover/replace we need to record a bad block on all
3188 	 * non-sync devices, or abort the recovery
3189 	 */
3190 	if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3191 		/* During recovery devices cannot be removed, so
3192 		 * locking and refcounting of rdevs is not needed
3193 		 */
3194 		for (i = 0; i < conf->raid_disks; i++) {
3195 			struct md_rdev *rdev = conf->disks[i].rdev;
3196 			if (rdev
3197 			    && !test_bit(Faulty, &rdev->flags)
3198 			    && !test_bit(In_sync, &rdev->flags)
3199 			    && !rdev_set_badblocks(rdev, sh->sector,
3200 						   STRIPE_SECTORS, 0))
3201 				abort = 1;
3202 			rdev = conf->disks[i].replacement;
3203 			if (rdev
3204 			    && !test_bit(Faulty, &rdev->flags)
3205 			    && !test_bit(In_sync, &rdev->flags)
3206 			    && !rdev_set_badblocks(rdev, sh->sector,
3207 						   STRIPE_SECTORS, 0))
3208 				abort = 1;
3209 		}
3210 		if (abort)
3211 			conf->recovery_disabled =
3212 				conf->mddev->recovery_disabled;
3213 	}
3214 	md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3215 }
3216 
3217 static int want_replace(struct stripe_head *sh, int disk_idx)
3218 {
3219 	struct md_rdev *rdev;
3220 	int rv = 0;
3221 	/* Doing recovery so rcu locking not required */
3222 	rdev = sh->raid_conf->disks[disk_idx].replacement;
3223 	if (rdev
3224 	    && !test_bit(Faulty, &rdev->flags)
3225 	    && !test_bit(In_sync, &rdev->flags)
3226 	    && (rdev->recovery_offset <= sh->sector
3227 		|| rdev->mddev->recovery_cp <= sh->sector))
3228 		rv = 1;
3229 
3230 	return rv;
3231 }
3232 
3233 /* fetch_block - checks the given member device to see if its data needs
3234  * to be read or computed to satisfy a request.
3235  *
3236  * Returns 1 when no more member devices need to be checked, otherwise returns
3237  * 0 to tell the loop in handle_stripe_fill to continue
3238  */
3239 
3240 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3241 			   int disk_idx, int disks)
3242 {
3243 	struct r5dev *dev = &sh->dev[disk_idx];
3244 	struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3245 				  &sh->dev[s->failed_num[1]] };
3246 	int i;
3247 
3248 
3249 	if (test_bit(R5_LOCKED, &dev->flags) ||
3250 	    test_bit(R5_UPTODATE, &dev->flags))
3251 		/* No point reading this as we already have it or have
3252 		 * decided to get it.
3253 		 */
3254 		return 0;
3255 
3256 	if (dev->toread ||
3257 	    (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3258 		/* We need this block to directly satisfy a request */
3259 		return 1;
3260 
3261 	if (s->syncing || s->expanding ||
3262 	    (s->replacing && want_replace(sh, disk_idx)))
3263 		/* When syncing, or expanding we read everything.
3264 		 * When replacing, we need the replaced block.
3265 		 */
3266 		return 1;
3267 
3268 	if ((s->failed >= 1 && fdev[0]->toread) ||
3269 	    (s->failed >= 2 && fdev[1]->toread))
3270 		/* If we want to read from a failed device, then
3271 		 * we need to actually read every other device.
3272 		 */
3273 		return 1;
3274 
3275 	/* Sometimes neither read-modify-write nor reconstruct-write
3276 	 * cycles can work.  In those cases we read every block we
3277 	 * can.  Then the parity-update is certain to have enough to
3278 	 * work with.
3279 	 * This can only be a problem when we need to write something,
3280 	 * and some device has failed.  If either of those tests
3281 	 * fail we need look no further.
3282 	 */
3283 	if (!s->failed || !s->to_write)
3284 		return 0;
3285 
3286 	if (test_bit(R5_Insync, &dev->flags) &&
3287 	    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3288 		/* Pre-reads at not permitted until after short delay
3289 		 * to gather multiple requests.  However if this
3290 		 * device is no Insync, the block could only be be computed
3291 		 * and there is no need to delay that.
3292 		 */
3293 		return 0;
3294 
3295 	for (i = 0; i < s->failed; i++) {
3296 		if (fdev[i]->towrite &&
3297 		    !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3298 		    !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3299 			/* If we have a partial write to a failed
3300 			 * device, then we will need to reconstruct
3301 			 * the content of that device, so all other
3302 			 * devices must be read.
3303 			 */
3304 			return 1;
3305 	}
3306 
3307 	/* If we are forced to do a reconstruct-write, either because
3308 	 * the current RAID6 implementation only supports that, or
3309 	 * or because parity cannot be trusted and we are currently
3310 	 * recovering it, there is extra need to be careful.
3311 	 * If one of the devices that we would need to read, because
3312 	 * it is not being overwritten (and maybe not written at all)
3313 	 * is missing/faulty, then we need to read everything we can.
3314 	 */
3315 	if (sh->raid_conf->level != 6 &&
3316 	    sh->sector < sh->raid_conf->mddev->recovery_cp)
3317 		/* reconstruct-write isn't being forced */
3318 		return 0;
3319 	for (i = 0; i < s->failed; i++) {
3320 		if (s->failed_num[i] != sh->pd_idx &&
3321 		    s->failed_num[i] != sh->qd_idx &&
3322 		    !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3323 		    !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3324 			return 1;
3325 	}
3326 
3327 	return 0;
3328 }
3329 
3330 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3331 		       int disk_idx, int disks)
3332 {
3333 	struct r5dev *dev = &sh->dev[disk_idx];
3334 
3335 	/* is the data in this block needed, and can we get it? */
3336 	if (need_this_block(sh, s, disk_idx, disks)) {
3337 		/* we would like to get this block, possibly by computing it,
3338 		 * otherwise read it if the backing disk is insync
3339 		 */
3340 		BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3341 		BUG_ON(test_bit(R5_Wantread, &dev->flags));
3342 		BUG_ON(sh->batch_head);
3343 		if ((s->uptodate == disks - 1) &&
3344 		    (s->failed && (disk_idx == s->failed_num[0] ||
3345 				   disk_idx == s->failed_num[1]))) {
3346 			/* have disk failed, and we're requested to fetch it;
3347 			 * do compute it
3348 			 */
3349 			pr_debug("Computing stripe %llu block %d\n",
3350 			       (unsigned long long)sh->sector, disk_idx);
3351 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3352 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3353 			set_bit(R5_Wantcompute, &dev->flags);
3354 			sh->ops.target = disk_idx;
3355 			sh->ops.target2 = -1; /* no 2nd target */
3356 			s->req_compute = 1;
3357 			/* Careful: from this point on 'uptodate' is in the eye
3358 			 * of raid_run_ops which services 'compute' operations
3359 			 * before writes. R5_Wantcompute flags a block that will
3360 			 * be R5_UPTODATE by the time it is needed for a
3361 			 * subsequent operation.
3362 			 */
3363 			s->uptodate++;
3364 			return 1;
3365 		} else if (s->uptodate == disks-2 && s->failed >= 2) {
3366 			/* Computing 2-failure is *very* expensive; only
3367 			 * do it if failed >= 2
3368 			 */
3369 			int other;
3370 			for (other = disks; other--; ) {
3371 				if (other == disk_idx)
3372 					continue;
3373 				if (!test_bit(R5_UPTODATE,
3374 				      &sh->dev[other].flags))
3375 					break;
3376 			}
3377 			BUG_ON(other < 0);
3378 			pr_debug("Computing stripe %llu blocks %d,%d\n",
3379 			       (unsigned long long)sh->sector,
3380 			       disk_idx, other);
3381 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3382 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3383 			set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3384 			set_bit(R5_Wantcompute, &sh->dev[other].flags);
3385 			sh->ops.target = disk_idx;
3386 			sh->ops.target2 = other;
3387 			s->uptodate += 2;
3388 			s->req_compute = 1;
3389 			return 1;
3390 		} else if (test_bit(R5_Insync, &dev->flags)) {
3391 			set_bit(R5_LOCKED, &dev->flags);
3392 			set_bit(R5_Wantread, &dev->flags);
3393 			s->locked++;
3394 			pr_debug("Reading block %d (sync=%d)\n",
3395 				disk_idx, s->syncing);
3396 		}
3397 	}
3398 
3399 	return 0;
3400 }
3401 
3402 /**
3403  * handle_stripe_fill - read or compute data to satisfy pending requests.
3404  */
3405 static void handle_stripe_fill(struct stripe_head *sh,
3406 			       struct stripe_head_state *s,
3407 			       int disks)
3408 {
3409 	int i;
3410 
3411 	/* look for blocks to read/compute, skip this if a compute
3412 	 * is already in flight, or if the stripe contents are in the
3413 	 * midst of changing due to a write
3414 	 */
3415 	if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3416 	    !sh->reconstruct_state)
3417 		for (i = disks; i--; )
3418 			if (fetch_block(sh, s, i, disks))
3419 				break;
3420 	set_bit(STRIPE_HANDLE, &sh->state);
3421 }
3422 
3423 static void break_stripe_batch_list(struct stripe_head *head_sh,
3424 				    unsigned long handle_flags);
3425 /* handle_stripe_clean_event
3426  * any written block on an uptodate or failed drive can be returned.
3427  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3428  * never LOCKED, so we don't need to test 'failed' directly.
3429  */
3430 static void handle_stripe_clean_event(struct r5conf *conf,
3431 	struct stripe_head *sh, int disks, struct bio **return_bi)
3432 {
3433 	int i;
3434 	struct r5dev *dev;
3435 	int discard_pending = 0;
3436 	struct stripe_head *head_sh = sh;
3437 	bool do_endio = false;
3438 
3439 	for (i = disks; i--; )
3440 		if (sh->dev[i].written) {
3441 			dev = &sh->dev[i];
3442 			if (!test_bit(R5_LOCKED, &dev->flags) &&
3443 			    (test_bit(R5_UPTODATE, &dev->flags) ||
3444 			     test_bit(R5_Discard, &dev->flags) ||
3445 			     test_bit(R5_SkipCopy, &dev->flags))) {
3446 				/* We can return any write requests */
3447 				struct bio *wbi, *wbi2;
3448 				pr_debug("Return write for disc %d\n", i);
3449 				if (test_and_clear_bit(R5_Discard, &dev->flags))
3450 					clear_bit(R5_UPTODATE, &dev->flags);
3451 				if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3452 					WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3453 				}
3454 				do_endio = true;
3455 
3456 returnbi:
3457 				dev->page = dev->orig_page;
3458 				wbi = dev->written;
3459 				dev->written = NULL;
3460 				while (wbi && wbi->bi_iter.bi_sector <
3461 					dev->sector + STRIPE_SECTORS) {
3462 					wbi2 = r5_next_bio(wbi, dev->sector);
3463 					if (!raid5_dec_bi_active_stripes(wbi)) {
3464 						md_write_end(conf->mddev);
3465 						wbi->bi_next = *return_bi;
3466 						*return_bi = wbi;
3467 					}
3468 					wbi = wbi2;
3469 				}
3470 				bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3471 						STRIPE_SECTORS,
3472 					 !test_bit(STRIPE_DEGRADED, &sh->state),
3473 						0);
3474 				if (head_sh->batch_head) {
3475 					sh = list_first_entry(&sh->batch_list,
3476 							      struct stripe_head,
3477 							      batch_list);
3478 					if (sh != head_sh) {
3479 						dev = &sh->dev[i];
3480 						goto returnbi;
3481 					}
3482 				}
3483 				sh = head_sh;
3484 				dev = &sh->dev[i];
3485 			} else if (test_bit(R5_Discard, &dev->flags))
3486 				discard_pending = 1;
3487 			WARN_ON(test_bit(R5_SkipCopy, &dev->flags));
3488 			WARN_ON(dev->page != dev->orig_page);
3489 		}
3490 	if (!discard_pending &&
3491 	    test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3492 		clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3493 		clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3494 		if (sh->qd_idx >= 0) {
3495 			clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3496 			clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3497 		}
3498 		/* now that discard is done we can proceed with any sync */
3499 		clear_bit(STRIPE_DISCARD, &sh->state);
3500 		/*
3501 		 * SCSI discard will change some bio fields and the stripe has
3502 		 * no updated data, so remove it from hash list and the stripe
3503 		 * will be reinitialized
3504 		 */
3505 		spin_lock_irq(&conf->device_lock);
3506 unhash:
3507 		remove_hash(sh);
3508 		if (head_sh->batch_head) {
3509 			sh = list_first_entry(&sh->batch_list,
3510 					      struct stripe_head, batch_list);
3511 			if (sh != head_sh)
3512 					goto unhash;
3513 		}
3514 		spin_unlock_irq(&conf->device_lock);
3515 		sh = head_sh;
3516 
3517 		if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3518 			set_bit(STRIPE_HANDLE, &sh->state);
3519 
3520 	}
3521 
3522 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3523 		if (atomic_dec_and_test(&conf->pending_full_writes))
3524 			md_wakeup_thread(conf->mddev->thread);
3525 
3526 	if (head_sh->batch_head && do_endio)
3527 		break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3528 }
3529 
3530 static void handle_stripe_dirtying(struct r5conf *conf,
3531 				   struct stripe_head *sh,
3532 				   struct stripe_head_state *s,
3533 				   int disks)
3534 {
3535 	int rmw = 0, rcw = 0, i;
3536 	sector_t recovery_cp = conf->mddev->recovery_cp;
3537 
3538 	/* Check whether resync is now happening or should start.
3539 	 * If yes, then the array is dirty (after unclean shutdown or
3540 	 * initial creation), so parity in some stripes might be inconsistent.
3541 	 * In this case, we need to always do reconstruct-write, to ensure
3542 	 * that in case of drive failure or read-error correction, we
3543 	 * generate correct data from the parity.
3544 	 */
3545 	if (conf->rmw_level == PARITY_DISABLE_RMW ||
3546 	    (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3547 	     s->failed == 0)) {
3548 		/* Calculate the real rcw later - for now make it
3549 		 * look like rcw is cheaper
3550 		 */
3551 		rcw = 1; rmw = 2;
3552 		pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3553 			 conf->rmw_level, (unsigned long long)recovery_cp,
3554 			 (unsigned long long)sh->sector);
3555 	} else for (i = disks; i--; ) {
3556 		/* would I have to read this buffer for read_modify_write */
3557 		struct r5dev *dev = &sh->dev[i];
3558 		if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3559 		    !test_bit(R5_LOCKED, &dev->flags) &&
3560 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
3561 		      test_bit(R5_Wantcompute, &dev->flags))) {
3562 			if (test_bit(R5_Insync, &dev->flags))
3563 				rmw++;
3564 			else
3565 				rmw += 2*disks;  /* cannot read it */
3566 		}
3567 		/* Would I have to read this buffer for reconstruct_write */
3568 		if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3569 		    i != sh->pd_idx && i != sh->qd_idx &&
3570 		    !test_bit(R5_LOCKED, &dev->flags) &&
3571 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
3572 		    test_bit(R5_Wantcompute, &dev->flags))) {
3573 			if (test_bit(R5_Insync, &dev->flags))
3574 				rcw++;
3575 			else
3576 				rcw += 2*disks;
3577 		}
3578 	}
3579 	pr_debug("for sector %llu, rmw=%d rcw=%d\n",
3580 		(unsigned long long)sh->sector, rmw, rcw);
3581 	set_bit(STRIPE_HANDLE, &sh->state);
3582 	if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_ENABLE_RMW)) && rmw > 0) {
3583 		/* prefer read-modify-write, but need to get some data */
3584 		if (conf->mddev->queue)
3585 			blk_add_trace_msg(conf->mddev->queue,
3586 					  "raid5 rmw %llu %d",
3587 					  (unsigned long long)sh->sector, rmw);
3588 		for (i = disks; i--; ) {
3589 			struct r5dev *dev = &sh->dev[i];
3590 			if ((dev->towrite || i == sh->pd_idx || i == sh->qd_idx) &&
3591 			    !test_bit(R5_LOCKED, &dev->flags) &&
3592 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
3593 			    test_bit(R5_Wantcompute, &dev->flags)) &&
3594 			    test_bit(R5_Insync, &dev->flags)) {
3595 				if (test_bit(STRIPE_PREREAD_ACTIVE,
3596 					     &sh->state)) {
3597 					pr_debug("Read_old block %d for r-m-w\n",
3598 						 i);
3599 					set_bit(R5_LOCKED, &dev->flags);
3600 					set_bit(R5_Wantread, &dev->flags);
3601 					s->locked++;
3602 				} else {
3603 					set_bit(STRIPE_DELAYED, &sh->state);
3604 					set_bit(STRIPE_HANDLE, &sh->state);
3605 				}
3606 			}
3607 		}
3608 	}
3609 	if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_ENABLE_RMW)) && rcw > 0) {
3610 		/* want reconstruct write, but need to get some data */
3611 		int qread =0;
3612 		rcw = 0;
3613 		for (i = disks; i--; ) {
3614 			struct r5dev *dev = &sh->dev[i];
3615 			if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3616 			    i != sh->pd_idx && i != sh->qd_idx &&
3617 			    !test_bit(R5_LOCKED, &dev->flags) &&
3618 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
3619 			      test_bit(R5_Wantcompute, &dev->flags))) {
3620 				rcw++;
3621 				if (test_bit(R5_Insync, &dev->flags) &&
3622 				    test_bit(STRIPE_PREREAD_ACTIVE,
3623 					     &sh->state)) {
3624 					pr_debug("Read_old block "
3625 						"%d for Reconstruct\n", i);
3626 					set_bit(R5_LOCKED, &dev->flags);
3627 					set_bit(R5_Wantread, &dev->flags);
3628 					s->locked++;
3629 					qread++;
3630 				} else {
3631 					set_bit(STRIPE_DELAYED, &sh->state);
3632 					set_bit(STRIPE_HANDLE, &sh->state);
3633 				}
3634 			}
3635 		}
3636 		if (rcw && conf->mddev->queue)
3637 			blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
3638 					  (unsigned long long)sh->sector,
3639 					  rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
3640 	}
3641 
3642 	if (rcw > disks && rmw > disks &&
3643 	    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3644 		set_bit(STRIPE_DELAYED, &sh->state);
3645 
3646 	/* now if nothing is locked, and if we have enough data,
3647 	 * we can start a write request
3648 	 */
3649 	/* since handle_stripe can be called at any time we need to handle the
3650 	 * case where a compute block operation has been submitted and then a
3651 	 * subsequent call wants to start a write request.  raid_run_ops only
3652 	 * handles the case where compute block and reconstruct are requested
3653 	 * simultaneously.  If this is not the case then new writes need to be
3654 	 * held off until the compute completes.
3655 	 */
3656 	if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
3657 	    (s->locked == 0 && (rcw == 0 || rmw == 0) &&
3658 	    !test_bit(STRIPE_BIT_DELAY, &sh->state)))
3659 		schedule_reconstruction(sh, s, rcw == 0, 0);
3660 }
3661 
3662 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
3663 				struct stripe_head_state *s, int disks)
3664 {
3665 	struct r5dev *dev = NULL;
3666 
3667 	BUG_ON(sh->batch_head);
3668 	set_bit(STRIPE_HANDLE, &sh->state);
3669 
3670 	switch (sh->check_state) {
3671 	case check_state_idle:
3672 		/* start a new check operation if there are no failures */
3673 		if (s->failed == 0) {
3674 			BUG_ON(s->uptodate != disks);
3675 			sh->check_state = check_state_run;
3676 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
3677 			clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3678 			s->uptodate--;
3679 			break;
3680 		}
3681 		dev = &sh->dev[s->failed_num[0]];
3682 		/* fall through */
3683 	case check_state_compute_result:
3684 		sh->check_state = check_state_idle;
3685 		if (!dev)
3686 			dev = &sh->dev[sh->pd_idx];
3687 
3688 		/* check that a write has not made the stripe insync */
3689 		if (test_bit(STRIPE_INSYNC, &sh->state))
3690 			break;
3691 
3692 		/* either failed parity check, or recovery is happening */
3693 		BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3694 		BUG_ON(s->uptodate != disks);
3695 
3696 		set_bit(R5_LOCKED, &dev->flags);
3697 		s->locked++;
3698 		set_bit(R5_Wantwrite, &dev->flags);
3699 
3700 		clear_bit(STRIPE_DEGRADED, &sh->state);
3701 		set_bit(STRIPE_INSYNC, &sh->state);
3702 		break;
3703 	case check_state_run:
3704 		break; /* we will be called again upon completion */
3705 	case check_state_check_result:
3706 		sh->check_state = check_state_idle;
3707 
3708 		/* if a failure occurred during the check operation, leave
3709 		 * STRIPE_INSYNC not set and let the stripe be handled again
3710 		 */
3711 		if (s->failed)
3712 			break;
3713 
3714 		/* handle a successful check operation, if parity is correct
3715 		 * we are done.  Otherwise update the mismatch count and repair
3716 		 * parity if !MD_RECOVERY_CHECK
3717 		 */
3718 		if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
3719 			/* parity is correct (on disc,
3720 			 * not in buffer any more)
3721 			 */
3722 			set_bit(STRIPE_INSYNC, &sh->state);
3723 		else {
3724 			atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3725 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3726 				/* don't try to repair!! */
3727 				set_bit(STRIPE_INSYNC, &sh->state);
3728 			else {
3729 				sh->check_state = check_state_compute_run;
3730 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3731 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3732 				set_bit(R5_Wantcompute,
3733 					&sh->dev[sh->pd_idx].flags);
3734 				sh->ops.target = sh->pd_idx;
3735 				sh->ops.target2 = -1;
3736 				s->uptodate++;
3737 			}
3738 		}
3739 		break;
3740 	case check_state_compute_run:
3741 		break;
3742 	default:
3743 		printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3744 		       __func__, sh->check_state,
3745 		       (unsigned long long) sh->sector);
3746 		BUG();
3747 	}
3748 }
3749 
3750 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
3751 				  struct stripe_head_state *s,
3752 				  int disks)
3753 {
3754 	int pd_idx = sh->pd_idx;
3755 	int qd_idx = sh->qd_idx;
3756 	struct r5dev *dev;
3757 
3758 	BUG_ON(sh->batch_head);
3759 	set_bit(STRIPE_HANDLE, &sh->state);
3760 
3761 	BUG_ON(s->failed > 2);
3762 
3763 	/* Want to check and possibly repair P and Q.
3764 	 * However there could be one 'failed' device, in which
3765 	 * case we can only check one of them, possibly using the
3766 	 * other to generate missing data
3767 	 */
3768 
3769 	switch (sh->check_state) {
3770 	case check_state_idle:
3771 		/* start a new check operation if there are < 2 failures */
3772 		if (s->failed == s->q_failed) {
3773 			/* The only possible failed device holds Q, so it
3774 			 * makes sense to check P (If anything else were failed,
3775 			 * we would have used P to recreate it).
3776 			 */
3777 			sh->check_state = check_state_run;
3778 		}
3779 		if (!s->q_failed && s->failed < 2) {
3780 			/* Q is not failed, and we didn't use it to generate
3781 			 * anything, so it makes sense to check it
3782 			 */
3783 			if (sh->check_state == check_state_run)
3784 				sh->check_state = check_state_run_pq;
3785 			else
3786 				sh->check_state = check_state_run_q;
3787 		}
3788 
3789 		/* discard potentially stale zero_sum_result */
3790 		sh->ops.zero_sum_result = 0;
3791 
3792 		if (sh->check_state == check_state_run) {
3793 			/* async_xor_zero_sum destroys the contents of P */
3794 			clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3795 			s->uptodate--;
3796 		}
3797 		if (sh->check_state >= check_state_run &&
3798 		    sh->check_state <= check_state_run_pq) {
3799 			/* async_syndrome_zero_sum preserves P and Q, so
3800 			 * no need to mark them !uptodate here
3801 			 */
3802 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
3803 			break;
3804 		}
3805 
3806 		/* we have 2-disk failure */
3807 		BUG_ON(s->failed != 2);
3808 		/* fall through */
3809 	case check_state_compute_result:
3810 		sh->check_state = check_state_idle;
3811 
3812 		/* check that a write has not made the stripe insync */
3813 		if (test_bit(STRIPE_INSYNC, &sh->state))
3814 			break;
3815 
3816 		/* now write out any block on a failed drive,
3817 		 * or P or Q if they were recomputed
3818 		 */
3819 		BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
3820 		if (s->failed == 2) {
3821 			dev = &sh->dev[s->failed_num[1]];
3822 			s->locked++;
3823 			set_bit(R5_LOCKED, &dev->flags);
3824 			set_bit(R5_Wantwrite, &dev->flags);
3825 		}
3826 		if (s->failed >= 1) {
3827 			dev = &sh->dev[s->failed_num[0]];
3828 			s->locked++;
3829 			set_bit(R5_LOCKED, &dev->flags);
3830 			set_bit(R5_Wantwrite, &dev->flags);
3831 		}
3832 		if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3833 			dev = &sh->dev[pd_idx];
3834 			s->locked++;
3835 			set_bit(R5_LOCKED, &dev->flags);
3836 			set_bit(R5_Wantwrite, &dev->flags);
3837 		}
3838 		if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3839 			dev = &sh->dev[qd_idx];
3840 			s->locked++;
3841 			set_bit(R5_LOCKED, &dev->flags);
3842 			set_bit(R5_Wantwrite, &dev->flags);
3843 		}
3844 		clear_bit(STRIPE_DEGRADED, &sh->state);
3845 
3846 		set_bit(STRIPE_INSYNC, &sh->state);
3847 		break;
3848 	case check_state_run:
3849 	case check_state_run_q:
3850 	case check_state_run_pq:
3851 		break; /* we will be called again upon completion */
3852 	case check_state_check_result:
3853 		sh->check_state = check_state_idle;
3854 
3855 		/* handle a successful check operation, if parity is correct
3856 		 * we are done.  Otherwise update the mismatch count and repair
3857 		 * parity if !MD_RECOVERY_CHECK
3858 		 */
3859 		if (sh->ops.zero_sum_result == 0) {
3860 			/* both parities are correct */
3861 			if (!s->failed)
3862 				set_bit(STRIPE_INSYNC, &sh->state);
3863 			else {
3864 				/* in contrast to the raid5 case we can validate
3865 				 * parity, but still have a failure to write
3866 				 * back
3867 				 */
3868 				sh->check_state = check_state_compute_result;
3869 				/* Returning at this point means that we may go
3870 				 * off and bring p and/or q uptodate again so
3871 				 * we make sure to check zero_sum_result again
3872 				 * to verify if p or q need writeback
3873 				 */
3874 			}
3875 		} else {
3876 			atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
3877 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
3878 				/* don't try to repair!! */
3879 				set_bit(STRIPE_INSYNC, &sh->state);
3880 			else {
3881 				int *target = &sh->ops.target;
3882 
3883 				sh->ops.target = -1;
3884 				sh->ops.target2 = -1;
3885 				sh->check_state = check_state_compute_run;
3886 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3887 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3888 				if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
3889 					set_bit(R5_Wantcompute,
3890 						&sh->dev[pd_idx].flags);
3891 					*target = pd_idx;
3892 					target = &sh->ops.target2;
3893 					s->uptodate++;
3894 				}
3895 				if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
3896 					set_bit(R5_Wantcompute,
3897 						&sh->dev[qd_idx].flags);
3898 					*target = qd_idx;
3899 					s->uptodate++;
3900 				}
3901 			}
3902 		}
3903 		break;
3904 	case check_state_compute_run:
3905 		break;
3906 	default:
3907 		printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
3908 		       __func__, sh->check_state,
3909 		       (unsigned long long) sh->sector);
3910 		BUG();
3911 	}
3912 }
3913 
3914 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
3915 {
3916 	int i;
3917 
3918 	/* We have read all the blocks in this stripe and now we need to
3919 	 * copy some of them into a target stripe for expand.
3920 	 */
3921 	struct dma_async_tx_descriptor *tx = NULL;
3922 	BUG_ON(sh->batch_head);
3923 	clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3924 	for (i = 0; i < sh->disks; i++)
3925 		if (i != sh->pd_idx && i != sh->qd_idx) {
3926 			int dd_idx, j;
3927 			struct stripe_head *sh2;
3928 			struct async_submit_ctl submit;
3929 
3930 			sector_t bn = compute_blocknr(sh, i, 1);
3931 			sector_t s = raid5_compute_sector(conf, bn, 0,
3932 							  &dd_idx, NULL);
3933 			sh2 = get_active_stripe(conf, s, 0, 1, 1);
3934 			if (sh2 == NULL)
3935 				/* so far only the early blocks of this stripe
3936 				 * have been requested.  When later blocks
3937 				 * get requested, we will try again
3938 				 */
3939 				continue;
3940 			if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
3941 			   test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
3942 				/* must have already done this block */
3943 				release_stripe(sh2);
3944 				continue;
3945 			}
3946 
3947 			/* place all the copies on one channel */
3948 			init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
3949 			tx = async_memcpy(sh2->dev[dd_idx].page,
3950 					  sh->dev[i].page, 0, 0, STRIPE_SIZE,
3951 					  &submit);
3952 
3953 			set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
3954 			set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
3955 			for (j = 0; j < conf->raid_disks; j++)
3956 				if (j != sh2->pd_idx &&
3957 				    j != sh2->qd_idx &&
3958 				    !test_bit(R5_Expanded, &sh2->dev[j].flags))
3959 					break;
3960 			if (j == conf->raid_disks) {
3961 				set_bit(STRIPE_EXPAND_READY, &sh2->state);
3962 				set_bit(STRIPE_HANDLE, &sh2->state);
3963 			}
3964 			release_stripe(sh2);
3965 
3966 		}
3967 	/* done submitting copies, wait for them to complete */
3968 	async_tx_quiesce(&tx);
3969 }
3970 
3971 /*
3972  * handle_stripe - do things to a stripe.
3973  *
3974  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
3975  * state of various bits to see what needs to be done.
3976  * Possible results:
3977  *    return some read requests which now have data
3978  *    return some write requests which are safely on storage
3979  *    schedule a read on some buffers
3980  *    schedule a write of some buffers
3981  *    return confirmation of parity correctness
3982  *
3983  */
3984 
3985 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
3986 {
3987 	struct r5conf *conf = sh->raid_conf;
3988 	int disks = sh->disks;
3989 	struct r5dev *dev;
3990 	int i;
3991 	int do_recovery = 0;
3992 
3993 	memset(s, 0, sizeof(*s));
3994 
3995 	s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
3996 	s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
3997 	s->failed_num[0] = -1;
3998 	s->failed_num[1] = -1;
3999 
4000 	/* Now to look around and see what can be done */
4001 	rcu_read_lock();
4002 	for (i=disks; i--; ) {
4003 		struct md_rdev *rdev;
4004 		sector_t first_bad;
4005 		int bad_sectors;
4006 		int is_bad = 0;
4007 
4008 		dev = &sh->dev[i];
4009 
4010 		pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4011 			 i, dev->flags,
4012 			 dev->toread, dev->towrite, dev->written);
4013 		/* maybe we can reply to a read
4014 		 *
4015 		 * new wantfill requests are only permitted while
4016 		 * ops_complete_biofill is guaranteed to be inactive
4017 		 */
4018 		if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4019 		    !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4020 			set_bit(R5_Wantfill, &dev->flags);
4021 
4022 		/* now count some things */
4023 		if (test_bit(R5_LOCKED, &dev->flags))
4024 			s->locked++;
4025 		if (test_bit(R5_UPTODATE, &dev->flags))
4026 			s->uptodate++;
4027 		if (test_bit(R5_Wantcompute, &dev->flags)) {
4028 			s->compute++;
4029 			BUG_ON(s->compute > 2);
4030 		}
4031 
4032 		if (test_bit(R5_Wantfill, &dev->flags))
4033 			s->to_fill++;
4034 		else if (dev->toread)
4035 			s->to_read++;
4036 		if (dev->towrite) {
4037 			s->to_write++;
4038 			if (!test_bit(R5_OVERWRITE, &dev->flags))
4039 				s->non_overwrite++;
4040 		}
4041 		if (dev->written)
4042 			s->written++;
4043 		/* Prefer to use the replacement for reads, but only
4044 		 * if it is recovered enough and has no bad blocks.
4045 		 */
4046 		rdev = rcu_dereference(conf->disks[i].replacement);
4047 		if (rdev && !test_bit(Faulty, &rdev->flags) &&
4048 		    rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4049 		    !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4050 				 &first_bad, &bad_sectors))
4051 			set_bit(R5_ReadRepl, &dev->flags);
4052 		else {
4053 			if (rdev)
4054 				set_bit(R5_NeedReplace, &dev->flags);
4055 			rdev = rcu_dereference(conf->disks[i].rdev);
4056 			clear_bit(R5_ReadRepl, &dev->flags);
4057 		}
4058 		if (rdev && test_bit(Faulty, &rdev->flags))
4059 			rdev = NULL;
4060 		if (rdev) {
4061 			is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4062 					     &first_bad, &bad_sectors);
4063 			if (s->blocked_rdev == NULL
4064 			    && (test_bit(Blocked, &rdev->flags)
4065 				|| is_bad < 0)) {
4066 				if (is_bad < 0)
4067 					set_bit(BlockedBadBlocks,
4068 						&rdev->flags);
4069 				s->blocked_rdev = rdev;
4070 				atomic_inc(&rdev->nr_pending);
4071 			}
4072 		}
4073 		clear_bit(R5_Insync, &dev->flags);
4074 		if (!rdev)
4075 			/* Not in-sync */;
4076 		else if (is_bad) {
4077 			/* also not in-sync */
4078 			if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4079 			    test_bit(R5_UPTODATE, &dev->flags)) {
4080 				/* treat as in-sync, but with a read error
4081 				 * which we can now try to correct
4082 				 */
4083 				set_bit(R5_Insync, &dev->flags);
4084 				set_bit(R5_ReadError, &dev->flags);
4085 			}
4086 		} else if (test_bit(In_sync, &rdev->flags))
4087 			set_bit(R5_Insync, &dev->flags);
4088 		else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4089 			/* in sync if before recovery_offset */
4090 			set_bit(R5_Insync, &dev->flags);
4091 		else if (test_bit(R5_UPTODATE, &dev->flags) &&
4092 			 test_bit(R5_Expanded, &dev->flags))
4093 			/* If we've reshaped into here, we assume it is Insync.
4094 			 * We will shortly update recovery_offset to make
4095 			 * it official.
4096 			 */
4097 			set_bit(R5_Insync, &dev->flags);
4098 
4099 		if (test_bit(R5_WriteError, &dev->flags)) {
4100 			/* This flag does not apply to '.replacement'
4101 			 * only to .rdev, so make sure to check that*/
4102 			struct md_rdev *rdev2 = rcu_dereference(
4103 				conf->disks[i].rdev);
4104 			if (rdev2 == rdev)
4105 				clear_bit(R5_Insync, &dev->flags);
4106 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4107 				s->handle_bad_blocks = 1;
4108 				atomic_inc(&rdev2->nr_pending);
4109 			} else
4110 				clear_bit(R5_WriteError, &dev->flags);
4111 		}
4112 		if (test_bit(R5_MadeGood, &dev->flags)) {
4113 			/* This flag does not apply to '.replacement'
4114 			 * only to .rdev, so make sure to check that*/
4115 			struct md_rdev *rdev2 = rcu_dereference(
4116 				conf->disks[i].rdev);
4117 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4118 				s->handle_bad_blocks = 1;
4119 				atomic_inc(&rdev2->nr_pending);
4120 			} else
4121 				clear_bit(R5_MadeGood, &dev->flags);
4122 		}
4123 		if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4124 			struct md_rdev *rdev2 = rcu_dereference(
4125 				conf->disks[i].replacement);
4126 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4127 				s->handle_bad_blocks = 1;
4128 				atomic_inc(&rdev2->nr_pending);
4129 			} else
4130 				clear_bit(R5_MadeGoodRepl, &dev->flags);
4131 		}
4132 		if (!test_bit(R5_Insync, &dev->flags)) {
4133 			/* The ReadError flag will just be confusing now */
4134 			clear_bit(R5_ReadError, &dev->flags);
4135 			clear_bit(R5_ReWrite, &dev->flags);
4136 		}
4137 		if (test_bit(R5_ReadError, &dev->flags))
4138 			clear_bit(R5_Insync, &dev->flags);
4139 		if (!test_bit(R5_Insync, &dev->flags)) {
4140 			if (s->failed < 2)
4141 				s->failed_num[s->failed] = i;
4142 			s->failed++;
4143 			if (rdev && !test_bit(Faulty, &rdev->flags))
4144 				do_recovery = 1;
4145 		}
4146 	}
4147 	if (test_bit(STRIPE_SYNCING, &sh->state)) {
4148 		/* If there is a failed device being replaced,
4149 		 *     we must be recovering.
4150 		 * else if we are after recovery_cp, we must be syncing
4151 		 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4152 		 * else we can only be replacing
4153 		 * sync and recovery both need to read all devices, and so
4154 		 * use the same flag.
4155 		 */
4156 		if (do_recovery ||
4157 		    sh->sector >= conf->mddev->recovery_cp ||
4158 		    test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4159 			s->syncing = 1;
4160 		else
4161 			s->replacing = 1;
4162 	}
4163 	rcu_read_unlock();
4164 }
4165 
4166 static int clear_batch_ready(struct stripe_head *sh)
4167 {
4168 	/* Return '1' if this is a member of batch, or
4169 	 * '0' if it is a lone stripe or a head which can now be
4170 	 * handled.
4171 	 */
4172 	struct stripe_head *tmp;
4173 	if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4174 		return (sh->batch_head && sh->batch_head != sh);
4175 	spin_lock(&sh->stripe_lock);
4176 	if (!sh->batch_head) {
4177 		spin_unlock(&sh->stripe_lock);
4178 		return 0;
4179 	}
4180 
4181 	/*
4182 	 * this stripe could be added to a batch list before we check
4183 	 * BATCH_READY, skips it
4184 	 */
4185 	if (sh->batch_head != sh) {
4186 		spin_unlock(&sh->stripe_lock);
4187 		return 1;
4188 	}
4189 	spin_lock(&sh->batch_lock);
4190 	list_for_each_entry(tmp, &sh->batch_list, batch_list)
4191 		clear_bit(STRIPE_BATCH_READY, &tmp->state);
4192 	spin_unlock(&sh->batch_lock);
4193 	spin_unlock(&sh->stripe_lock);
4194 
4195 	/*
4196 	 * BATCH_READY is cleared, no new stripes can be added.
4197 	 * batch_list can be accessed without lock
4198 	 */
4199 	return 0;
4200 }
4201 
4202 static void break_stripe_batch_list(struct stripe_head *head_sh,
4203 				    unsigned long handle_flags)
4204 {
4205 	struct stripe_head *sh, *next;
4206 	int i;
4207 	int do_wakeup = 0;
4208 
4209 	list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4210 
4211 		list_del_init(&sh->batch_list);
4212 
4213 		WARN_ON_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4214 					  (1 << STRIPE_SYNCING) |
4215 					  (1 << STRIPE_REPLACED) |
4216 					  (1 << STRIPE_PREREAD_ACTIVE) |
4217 					  (1 << STRIPE_DELAYED) |
4218 					  (1 << STRIPE_BIT_DELAY) |
4219 					  (1 << STRIPE_FULL_WRITE) |
4220 					  (1 << STRIPE_BIOFILL_RUN) |
4221 					  (1 << STRIPE_COMPUTE_RUN)  |
4222 					  (1 << STRIPE_OPS_REQ_PENDING) |
4223 					  (1 << STRIPE_DISCARD) |
4224 					  (1 << STRIPE_BATCH_READY) |
4225 					  (1 << STRIPE_BATCH_ERR) |
4226 					  (1 << STRIPE_BITMAP_PENDING)));
4227 		WARN_ON_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4228 					      (1 << STRIPE_REPLACED)));
4229 
4230 		set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4231 					    (1 << STRIPE_DEGRADED)),
4232 			      head_sh->state & (1 << STRIPE_INSYNC));
4233 
4234 		sh->check_state = head_sh->check_state;
4235 		sh->reconstruct_state = head_sh->reconstruct_state;
4236 		for (i = 0; i < sh->disks; i++) {
4237 			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4238 				do_wakeup = 1;
4239 			sh->dev[i].flags = head_sh->dev[i].flags &
4240 				(~((1 << R5_WriteError) | (1 << R5_Overlap)));
4241 		}
4242 		spin_lock_irq(&sh->stripe_lock);
4243 		sh->batch_head = NULL;
4244 		spin_unlock_irq(&sh->stripe_lock);
4245 		if (handle_flags == 0 ||
4246 		    sh->state & handle_flags)
4247 			set_bit(STRIPE_HANDLE, &sh->state);
4248 		release_stripe(sh);
4249 	}
4250 	spin_lock_irq(&head_sh->stripe_lock);
4251 	head_sh->batch_head = NULL;
4252 	spin_unlock_irq(&head_sh->stripe_lock);
4253 	for (i = 0; i < head_sh->disks; i++)
4254 		if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4255 			do_wakeup = 1;
4256 	if (head_sh->state & handle_flags)
4257 		set_bit(STRIPE_HANDLE, &head_sh->state);
4258 
4259 	if (do_wakeup)
4260 		wake_up(&head_sh->raid_conf->wait_for_overlap);
4261 }
4262 
4263 static void handle_stripe(struct stripe_head *sh)
4264 {
4265 	struct stripe_head_state s;
4266 	struct r5conf *conf = sh->raid_conf;
4267 	int i;
4268 	int prexor;
4269 	int disks = sh->disks;
4270 	struct r5dev *pdev, *qdev;
4271 
4272 	clear_bit(STRIPE_HANDLE, &sh->state);
4273 	if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4274 		/* already being handled, ensure it gets handled
4275 		 * again when current action finishes */
4276 		set_bit(STRIPE_HANDLE, &sh->state);
4277 		return;
4278 	}
4279 
4280 	if (clear_batch_ready(sh) ) {
4281 		clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4282 		return;
4283 	}
4284 
4285 	if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4286 		break_stripe_batch_list(sh, 0);
4287 
4288 	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4289 		spin_lock(&sh->stripe_lock);
4290 		/* Cannot process 'sync' concurrently with 'discard' */
4291 		if (!test_bit(STRIPE_DISCARD, &sh->state) &&
4292 		    test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4293 			set_bit(STRIPE_SYNCING, &sh->state);
4294 			clear_bit(STRIPE_INSYNC, &sh->state);
4295 			clear_bit(STRIPE_REPLACED, &sh->state);
4296 		}
4297 		spin_unlock(&sh->stripe_lock);
4298 	}
4299 	clear_bit(STRIPE_DELAYED, &sh->state);
4300 
4301 	pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4302 		"pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4303 	       (unsigned long long)sh->sector, sh->state,
4304 	       atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4305 	       sh->check_state, sh->reconstruct_state);
4306 
4307 	analyse_stripe(sh, &s);
4308 
4309 	if (s.handle_bad_blocks) {
4310 		set_bit(STRIPE_HANDLE, &sh->state);
4311 		goto finish;
4312 	}
4313 
4314 	if (unlikely(s.blocked_rdev)) {
4315 		if (s.syncing || s.expanding || s.expanded ||
4316 		    s.replacing || s.to_write || s.written) {
4317 			set_bit(STRIPE_HANDLE, &sh->state);
4318 			goto finish;
4319 		}
4320 		/* There is nothing for the blocked_rdev to block */
4321 		rdev_dec_pending(s.blocked_rdev, conf->mddev);
4322 		s.blocked_rdev = NULL;
4323 	}
4324 
4325 	if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4326 		set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4327 		set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4328 	}
4329 
4330 	pr_debug("locked=%d uptodate=%d to_read=%d"
4331 	       " to_write=%d failed=%d failed_num=%d,%d\n",
4332 	       s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4333 	       s.failed_num[0], s.failed_num[1]);
4334 	/* check if the array has lost more than max_degraded devices and,
4335 	 * if so, some requests might need to be failed.
4336 	 */
4337 	if (s.failed > conf->max_degraded) {
4338 		sh->check_state = 0;
4339 		sh->reconstruct_state = 0;
4340 		break_stripe_batch_list(sh, 0);
4341 		if (s.to_read+s.to_write+s.written)
4342 			handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
4343 		if (s.syncing + s.replacing)
4344 			handle_failed_sync(conf, sh, &s);
4345 	}
4346 
4347 	/* Now we check to see if any write operations have recently
4348 	 * completed
4349 	 */
4350 	prexor = 0;
4351 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4352 		prexor = 1;
4353 	if (sh->reconstruct_state == reconstruct_state_drain_result ||
4354 	    sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4355 		sh->reconstruct_state = reconstruct_state_idle;
4356 
4357 		/* All the 'written' buffers and the parity block are ready to
4358 		 * be written back to disk
4359 		 */
4360 		BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4361 		       !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4362 		BUG_ON(sh->qd_idx >= 0 &&
4363 		       !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4364 		       !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4365 		for (i = disks; i--; ) {
4366 			struct r5dev *dev = &sh->dev[i];
4367 			if (test_bit(R5_LOCKED, &dev->flags) &&
4368 				(i == sh->pd_idx || i == sh->qd_idx ||
4369 				 dev->written)) {
4370 				pr_debug("Writing block %d\n", i);
4371 				set_bit(R5_Wantwrite, &dev->flags);
4372 				if (prexor)
4373 					continue;
4374 				if (s.failed > 1)
4375 					continue;
4376 				if (!test_bit(R5_Insync, &dev->flags) ||
4377 				    ((i == sh->pd_idx || i == sh->qd_idx)  &&
4378 				     s.failed == 0))
4379 					set_bit(STRIPE_INSYNC, &sh->state);
4380 			}
4381 		}
4382 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4383 			s.dec_preread_active = 1;
4384 	}
4385 
4386 	/*
4387 	 * might be able to return some write requests if the parity blocks
4388 	 * are safe, or on a failed drive
4389 	 */
4390 	pdev = &sh->dev[sh->pd_idx];
4391 	s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4392 		|| (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4393 	qdev = &sh->dev[sh->qd_idx];
4394 	s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4395 		|| (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4396 		|| conf->level < 6;
4397 
4398 	if (s.written &&
4399 	    (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4400 			     && !test_bit(R5_LOCKED, &pdev->flags)
4401 			     && (test_bit(R5_UPTODATE, &pdev->flags) ||
4402 				 test_bit(R5_Discard, &pdev->flags))))) &&
4403 	    (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4404 			     && !test_bit(R5_LOCKED, &qdev->flags)
4405 			     && (test_bit(R5_UPTODATE, &qdev->flags) ||
4406 				 test_bit(R5_Discard, &qdev->flags))))))
4407 		handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
4408 
4409 	/* Now we might consider reading some blocks, either to check/generate
4410 	 * parity, or to satisfy requests
4411 	 * or to load a block that is being partially written.
4412 	 */
4413 	if (s.to_read || s.non_overwrite
4414 	    || (conf->level == 6 && s.to_write && s.failed)
4415 	    || (s.syncing && (s.uptodate + s.compute < disks))
4416 	    || s.replacing
4417 	    || s.expanding)
4418 		handle_stripe_fill(sh, &s, disks);
4419 
4420 	/* Now to consider new write requests and what else, if anything
4421 	 * should be read.  We do not handle new writes when:
4422 	 * 1/ A 'write' operation (copy+xor) is already in flight.
4423 	 * 2/ A 'check' operation is in flight, as it may clobber the parity
4424 	 *    block.
4425 	 */
4426 	if (s.to_write && !sh->reconstruct_state && !sh->check_state)
4427 		handle_stripe_dirtying(conf, sh, &s, disks);
4428 
4429 	/* maybe we need to check and possibly fix the parity for this stripe
4430 	 * Any reads will already have been scheduled, so we just see if enough
4431 	 * data is available.  The parity check is held off while parity
4432 	 * dependent operations are in flight.
4433 	 */
4434 	if (sh->check_state ||
4435 	    (s.syncing && s.locked == 0 &&
4436 	     !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4437 	     !test_bit(STRIPE_INSYNC, &sh->state))) {
4438 		if (conf->level == 6)
4439 			handle_parity_checks6(conf, sh, &s, disks);
4440 		else
4441 			handle_parity_checks5(conf, sh, &s, disks);
4442 	}
4443 
4444 	if ((s.replacing || s.syncing) && s.locked == 0
4445 	    && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4446 	    && !test_bit(STRIPE_REPLACED, &sh->state)) {
4447 		/* Write out to replacement devices where possible */
4448 		for (i = 0; i < conf->raid_disks; i++)
4449 			if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4450 				WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4451 				set_bit(R5_WantReplace, &sh->dev[i].flags);
4452 				set_bit(R5_LOCKED, &sh->dev[i].flags);
4453 				s.locked++;
4454 			}
4455 		if (s.replacing)
4456 			set_bit(STRIPE_INSYNC, &sh->state);
4457 		set_bit(STRIPE_REPLACED, &sh->state);
4458 	}
4459 	if ((s.syncing || s.replacing) && s.locked == 0 &&
4460 	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4461 	    test_bit(STRIPE_INSYNC, &sh->state)) {
4462 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4463 		clear_bit(STRIPE_SYNCING, &sh->state);
4464 		if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4465 			wake_up(&conf->wait_for_overlap);
4466 	}
4467 
4468 	/* If the failed drives are just a ReadError, then we might need
4469 	 * to progress the repair/check process
4470 	 */
4471 	if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4472 		for (i = 0; i < s.failed; i++) {
4473 			struct r5dev *dev = &sh->dev[s.failed_num[i]];
4474 			if (test_bit(R5_ReadError, &dev->flags)
4475 			    && !test_bit(R5_LOCKED, &dev->flags)
4476 			    && test_bit(R5_UPTODATE, &dev->flags)
4477 				) {
4478 				if (!test_bit(R5_ReWrite, &dev->flags)) {
4479 					set_bit(R5_Wantwrite, &dev->flags);
4480 					set_bit(R5_ReWrite, &dev->flags);
4481 					set_bit(R5_LOCKED, &dev->flags);
4482 					s.locked++;
4483 				} else {
4484 					/* let's read it back */
4485 					set_bit(R5_Wantread, &dev->flags);
4486 					set_bit(R5_LOCKED, &dev->flags);
4487 					s.locked++;
4488 				}
4489 			}
4490 		}
4491 
4492 	/* Finish reconstruct operations initiated by the expansion process */
4493 	if (sh->reconstruct_state == reconstruct_state_result) {
4494 		struct stripe_head *sh_src
4495 			= get_active_stripe(conf, sh->sector, 1, 1, 1);
4496 		if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4497 			/* sh cannot be written until sh_src has been read.
4498 			 * so arrange for sh to be delayed a little
4499 			 */
4500 			set_bit(STRIPE_DELAYED, &sh->state);
4501 			set_bit(STRIPE_HANDLE, &sh->state);
4502 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4503 					      &sh_src->state))
4504 				atomic_inc(&conf->preread_active_stripes);
4505 			release_stripe(sh_src);
4506 			goto finish;
4507 		}
4508 		if (sh_src)
4509 			release_stripe(sh_src);
4510 
4511 		sh->reconstruct_state = reconstruct_state_idle;
4512 		clear_bit(STRIPE_EXPANDING, &sh->state);
4513 		for (i = conf->raid_disks; i--; ) {
4514 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
4515 			set_bit(R5_LOCKED, &sh->dev[i].flags);
4516 			s.locked++;
4517 		}
4518 	}
4519 
4520 	if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4521 	    !sh->reconstruct_state) {
4522 		/* Need to write out all blocks after computing parity */
4523 		sh->disks = conf->raid_disks;
4524 		stripe_set_idx(sh->sector, conf, 0, sh);
4525 		schedule_reconstruction(sh, &s, 1, 1);
4526 	} else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4527 		clear_bit(STRIPE_EXPAND_READY, &sh->state);
4528 		atomic_dec(&conf->reshape_stripes);
4529 		wake_up(&conf->wait_for_overlap);
4530 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4531 	}
4532 
4533 	if (s.expanding && s.locked == 0 &&
4534 	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4535 		handle_stripe_expansion(conf, sh);
4536 
4537 finish:
4538 	/* wait for this device to become unblocked */
4539 	if (unlikely(s.blocked_rdev)) {
4540 		if (conf->mddev->external)
4541 			md_wait_for_blocked_rdev(s.blocked_rdev,
4542 						 conf->mddev);
4543 		else
4544 			/* Internal metadata will immediately
4545 			 * be written by raid5d, so we don't
4546 			 * need to wait here.
4547 			 */
4548 			rdev_dec_pending(s.blocked_rdev,
4549 					 conf->mddev);
4550 	}
4551 
4552 	if (s.handle_bad_blocks)
4553 		for (i = disks; i--; ) {
4554 			struct md_rdev *rdev;
4555 			struct r5dev *dev = &sh->dev[i];
4556 			if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
4557 				/* We own a safe reference to the rdev */
4558 				rdev = conf->disks[i].rdev;
4559 				if (!rdev_set_badblocks(rdev, sh->sector,
4560 							STRIPE_SECTORS, 0))
4561 					md_error(conf->mddev, rdev);
4562 				rdev_dec_pending(rdev, conf->mddev);
4563 			}
4564 			if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
4565 				rdev = conf->disks[i].rdev;
4566 				rdev_clear_badblocks(rdev, sh->sector,
4567 						     STRIPE_SECTORS, 0);
4568 				rdev_dec_pending(rdev, conf->mddev);
4569 			}
4570 			if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
4571 				rdev = conf->disks[i].replacement;
4572 				if (!rdev)
4573 					/* rdev have been moved down */
4574 					rdev = conf->disks[i].rdev;
4575 				rdev_clear_badblocks(rdev, sh->sector,
4576 						     STRIPE_SECTORS, 0);
4577 				rdev_dec_pending(rdev, conf->mddev);
4578 			}
4579 		}
4580 
4581 	if (s.ops_request)
4582 		raid_run_ops(sh, s.ops_request);
4583 
4584 	ops_run_io(sh, &s);
4585 
4586 	if (s.dec_preread_active) {
4587 		/* We delay this until after ops_run_io so that if make_request
4588 		 * is waiting on a flush, it won't continue until the writes
4589 		 * have actually been submitted.
4590 		 */
4591 		atomic_dec(&conf->preread_active_stripes);
4592 		if (atomic_read(&conf->preread_active_stripes) <
4593 		    IO_THRESHOLD)
4594 			md_wakeup_thread(conf->mddev->thread);
4595 	}
4596 
4597 	return_io(s.return_bi);
4598 
4599 	clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4600 }
4601 
4602 static void raid5_activate_delayed(struct r5conf *conf)
4603 {
4604 	if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
4605 		while (!list_empty(&conf->delayed_list)) {
4606 			struct list_head *l = conf->delayed_list.next;
4607 			struct stripe_head *sh;
4608 			sh = list_entry(l, struct stripe_head, lru);
4609 			list_del_init(l);
4610 			clear_bit(STRIPE_DELAYED, &sh->state);
4611 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4612 				atomic_inc(&conf->preread_active_stripes);
4613 			list_add_tail(&sh->lru, &conf->hold_list);
4614 			raid5_wakeup_stripe_thread(sh);
4615 		}
4616 	}
4617 }
4618 
4619 static void activate_bit_delay(struct r5conf *conf,
4620 	struct list_head *temp_inactive_list)
4621 {
4622 	/* device_lock is held */
4623 	struct list_head head;
4624 	list_add(&head, &conf->bitmap_list);
4625 	list_del_init(&conf->bitmap_list);
4626 	while (!list_empty(&head)) {
4627 		struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
4628 		int hash;
4629 		list_del_init(&sh->lru);
4630 		atomic_inc(&sh->count);
4631 		hash = sh->hash_lock_index;
4632 		__release_stripe(conf, sh, &temp_inactive_list[hash]);
4633 	}
4634 }
4635 
4636 static int raid5_congested(struct mddev *mddev, int bits)
4637 {
4638 	struct r5conf *conf = mddev->private;
4639 
4640 	/* No difference between reads and writes.  Just check
4641 	 * how busy the stripe_cache is
4642 	 */
4643 
4644 	if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
4645 		return 1;
4646 	if (conf->quiesce)
4647 		return 1;
4648 	if (atomic_read(&conf->empty_inactive_list_nr))
4649 		return 1;
4650 
4651 	return 0;
4652 }
4653 
4654 /* We want read requests to align with chunks where possible,
4655  * but write requests don't need to.
4656  */
4657 static int raid5_mergeable_bvec(struct mddev *mddev,
4658 				struct bvec_merge_data *bvm,
4659 				struct bio_vec *biovec)
4660 {
4661 	sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
4662 	int max;
4663 	unsigned int chunk_sectors = mddev->chunk_sectors;
4664 	unsigned int bio_sectors = bvm->bi_size >> 9;
4665 
4666 	/*
4667 	 * always allow writes to be mergeable, read as well if array
4668 	 * is degraded as we'll go through stripe cache anyway.
4669 	 */
4670 	if ((bvm->bi_rw & 1) == WRITE || mddev->degraded)
4671 		return biovec->bv_len;
4672 
4673 	if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4674 		chunk_sectors = mddev->new_chunk_sectors;
4675 	max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
4676 	if (max < 0) max = 0;
4677 	if (max <= biovec->bv_len && bio_sectors == 0)
4678 		return biovec->bv_len;
4679 	else
4680 		return max;
4681 }
4682 
4683 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
4684 {
4685 	sector_t sector = bio->bi_iter.bi_sector + get_start_sect(bio->bi_bdev);
4686 	unsigned int chunk_sectors = mddev->chunk_sectors;
4687 	unsigned int bio_sectors = bio_sectors(bio);
4688 
4689 	if (mddev->new_chunk_sectors < mddev->chunk_sectors)
4690 		chunk_sectors = mddev->new_chunk_sectors;
4691 	return  chunk_sectors >=
4692 		((sector & (chunk_sectors - 1)) + bio_sectors);
4693 }
4694 
4695 /*
4696  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
4697  *  later sampled by raid5d.
4698  */
4699 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
4700 {
4701 	unsigned long flags;
4702 
4703 	spin_lock_irqsave(&conf->device_lock, flags);
4704 
4705 	bi->bi_next = conf->retry_read_aligned_list;
4706 	conf->retry_read_aligned_list = bi;
4707 
4708 	spin_unlock_irqrestore(&conf->device_lock, flags);
4709 	md_wakeup_thread(conf->mddev->thread);
4710 }
4711 
4712 static struct bio *remove_bio_from_retry(struct r5conf *conf)
4713 {
4714 	struct bio *bi;
4715 
4716 	bi = conf->retry_read_aligned;
4717 	if (bi) {
4718 		conf->retry_read_aligned = NULL;
4719 		return bi;
4720 	}
4721 	bi = conf->retry_read_aligned_list;
4722 	if(bi) {
4723 		conf->retry_read_aligned_list = bi->bi_next;
4724 		bi->bi_next = NULL;
4725 		/*
4726 		 * this sets the active strip count to 1 and the processed
4727 		 * strip count to zero (upper 8 bits)
4728 		 */
4729 		raid5_set_bi_stripes(bi, 1); /* biased count of active stripes */
4730 	}
4731 
4732 	return bi;
4733 }
4734 
4735 /*
4736  *  The "raid5_align_endio" should check if the read succeeded and if it
4737  *  did, call bio_endio on the original bio (having bio_put the new bio
4738  *  first).
4739  *  If the read failed..
4740  */
4741 static void raid5_align_endio(struct bio *bi, int error)
4742 {
4743 	struct bio* raid_bi  = bi->bi_private;
4744 	struct mddev *mddev;
4745 	struct r5conf *conf;
4746 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
4747 	struct md_rdev *rdev;
4748 
4749 	bio_put(bi);
4750 
4751 	rdev = (void*)raid_bi->bi_next;
4752 	raid_bi->bi_next = NULL;
4753 	mddev = rdev->mddev;
4754 	conf = mddev->private;
4755 
4756 	rdev_dec_pending(rdev, conf->mddev);
4757 
4758 	if (!error && uptodate) {
4759 		trace_block_bio_complete(bdev_get_queue(raid_bi->bi_bdev),
4760 					 raid_bi, 0);
4761 		bio_endio(raid_bi, 0);
4762 		if (atomic_dec_and_test(&conf->active_aligned_reads))
4763 			wake_up(&conf->wait_for_stripe);
4764 		return;
4765 	}
4766 
4767 	pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
4768 
4769 	add_bio_to_retry(raid_bi, conf);
4770 }
4771 
4772 static int bio_fits_rdev(struct bio *bi)
4773 {
4774 	struct request_queue *q = bdev_get_queue(bi->bi_bdev);
4775 
4776 	if (bio_sectors(bi) > queue_max_sectors(q))
4777 		return 0;
4778 	blk_recount_segments(q, bi);
4779 	if (bi->bi_phys_segments > queue_max_segments(q))
4780 		return 0;
4781 
4782 	if (q->merge_bvec_fn)
4783 		/* it's too hard to apply the merge_bvec_fn at this stage,
4784 		 * just just give up
4785 		 */
4786 		return 0;
4787 
4788 	return 1;
4789 }
4790 
4791 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
4792 {
4793 	struct r5conf *conf = mddev->private;
4794 	int dd_idx;
4795 	struct bio* align_bi;
4796 	struct md_rdev *rdev;
4797 	sector_t end_sector;
4798 
4799 	if (!in_chunk_boundary(mddev, raid_bio)) {
4800 		pr_debug("chunk_aligned_read : non aligned\n");
4801 		return 0;
4802 	}
4803 	/*
4804 	 * use bio_clone_mddev to make a copy of the bio
4805 	 */
4806 	align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
4807 	if (!align_bi)
4808 		return 0;
4809 	/*
4810 	 *   set bi_end_io to a new function, and set bi_private to the
4811 	 *     original bio.
4812 	 */
4813 	align_bi->bi_end_io  = raid5_align_endio;
4814 	align_bi->bi_private = raid_bio;
4815 	/*
4816 	 *	compute position
4817 	 */
4818 	align_bi->bi_iter.bi_sector =
4819 		raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
4820 				     0, &dd_idx, NULL);
4821 
4822 	end_sector = bio_end_sector(align_bi);
4823 	rcu_read_lock();
4824 	rdev = rcu_dereference(conf->disks[dd_idx].replacement);
4825 	if (!rdev || test_bit(Faulty, &rdev->flags) ||
4826 	    rdev->recovery_offset < end_sector) {
4827 		rdev = rcu_dereference(conf->disks[dd_idx].rdev);
4828 		if (rdev &&
4829 		    (test_bit(Faulty, &rdev->flags) ||
4830 		    !(test_bit(In_sync, &rdev->flags) ||
4831 		      rdev->recovery_offset >= end_sector)))
4832 			rdev = NULL;
4833 	}
4834 	if (rdev) {
4835 		sector_t first_bad;
4836 		int bad_sectors;
4837 
4838 		atomic_inc(&rdev->nr_pending);
4839 		rcu_read_unlock();
4840 		raid_bio->bi_next = (void*)rdev;
4841 		align_bi->bi_bdev =  rdev->bdev;
4842 		__clear_bit(BIO_SEG_VALID, &align_bi->bi_flags);
4843 
4844 		if (!bio_fits_rdev(align_bi) ||
4845 		    is_badblock(rdev, align_bi->bi_iter.bi_sector,
4846 				bio_sectors(align_bi),
4847 				&first_bad, &bad_sectors)) {
4848 			/* too big in some way, or has a known bad block */
4849 			bio_put(align_bi);
4850 			rdev_dec_pending(rdev, mddev);
4851 			return 0;
4852 		}
4853 
4854 		/* No reshape active, so we can trust rdev->data_offset */
4855 		align_bi->bi_iter.bi_sector += rdev->data_offset;
4856 
4857 		spin_lock_irq(&conf->device_lock);
4858 		wait_event_lock_irq(conf->wait_for_stripe,
4859 				    conf->quiesce == 0,
4860 				    conf->device_lock);
4861 		atomic_inc(&conf->active_aligned_reads);
4862 		spin_unlock_irq(&conf->device_lock);
4863 
4864 		if (mddev->gendisk)
4865 			trace_block_bio_remap(bdev_get_queue(align_bi->bi_bdev),
4866 					      align_bi, disk_devt(mddev->gendisk),
4867 					      raid_bio->bi_iter.bi_sector);
4868 		generic_make_request(align_bi);
4869 		return 1;
4870 	} else {
4871 		rcu_read_unlock();
4872 		bio_put(align_bi);
4873 		return 0;
4874 	}
4875 }
4876 
4877 /* __get_priority_stripe - get the next stripe to process
4878  *
4879  * Full stripe writes are allowed to pass preread active stripes up until
4880  * the bypass_threshold is exceeded.  In general the bypass_count
4881  * increments when the handle_list is handled before the hold_list; however, it
4882  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
4883  * stripe with in flight i/o.  The bypass_count will be reset when the
4884  * head of the hold_list has changed, i.e. the head was promoted to the
4885  * handle_list.
4886  */
4887 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
4888 {
4889 	struct stripe_head *sh = NULL, *tmp;
4890 	struct list_head *handle_list = NULL;
4891 	struct r5worker_group *wg = NULL;
4892 
4893 	if (conf->worker_cnt_per_group == 0) {
4894 		handle_list = &conf->handle_list;
4895 	} else if (group != ANY_GROUP) {
4896 		handle_list = &conf->worker_groups[group].handle_list;
4897 		wg = &conf->worker_groups[group];
4898 	} else {
4899 		int i;
4900 		for (i = 0; i < conf->group_cnt; i++) {
4901 			handle_list = &conf->worker_groups[i].handle_list;
4902 			wg = &conf->worker_groups[i];
4903 			if (!list_empty(handle_list))
4904 				break;
4905 		}
4906 	}
4907 
4908 	pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
4909 		  __func__,
4910 		  list_empty(handle_list) ? "empty" : "busy",
4911 		  list_empty(&conf->hold_list) ? "empty" : "busy",
4912 		  atomic_read(&conf->pending_full_writes), conf->bypass_count);
4913 
4914 	if (!list_empty(handle_list)) {
4915 		sh = list_entry(handle_list->next, typeof(*sh), lru);
4916 
4917 		if (list_empty(&conf->hold_list))
4918 			conf->bypass_count = 0;
4919 		else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
4920 			if (conf->hold_list.next == conf->last_hold)
4921 				conf->bypass_count++;
4922 			else {
4923 				conf->last_hold = conf->hold_list.next;
4924 				conf->bypass_count -= conf->bypass_threshold;
4925 				if (conf->bypass_count < 0)
4926 					conf->bypass_count = 0;
4927 			}
4928 		}
4929 	} else if (!list_empty(&conf->hold_list) &&
4930 		   ((conf->bypass_threshold &&
4931 		     conf->bypass_count > conf->bypass_threshold) ||
4932 		    atomic_read(&conf->pending_full_writes) == 0)) {
4933 
4934 		list_for_each_entry(tmp, &conf->hold_list,  lru) {
4935 			if (conf->worker_cnt_per_group == 0 ||
4936 			    group == ANY_GROUP ||
4937 			    !cpu_online(tmp->cpu) ||
4938 			    cpu_to_group(tmp->cpu) == group) {
4939 				sh = tmp;
4940 				break;
4941 			}
4942 		}
4943 
4944 		if (sh) {
4945 			conf->bypass_count -= conf->bypass_threshold;
4946 			if (conf->bypass_count < 0)
4947 				conf->bypass_count = 0;
4948 		}
4949 		wg = NULL;
4950 	}
4951 
4952 	if (!sh)
4953 		return NULL;
4954 
4955 	if (wg) {
4956 		wg->stripes_cnt--;
4957 		sh->group = NULL;
4958 	}
4959 	list_del_init(&sh->lru);
4960 	BUG_ON(atomic_inc_return(&sh->count) != 1);
4961 	return sh;
4962 }
4963 
4964 struct raid5_plug_cb {
4965 	struct blk_plug_cb	cb;
4966 	struct list_head	list;
4967 	struct list_head	temp_inactive_list[NR_STRIPE_HASH_LOCKS];
4968 };
4969 
4970 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
4971 {
4972 	struct raid5_plug_cb *cb = container_of(
4973 		blk_cb, struct raid5_plug_cb, cb);
4974 	struct stripe_head *sh;
4975 	struct mddev *mddev = cb->cb.data;
4976 	struct r5conf *conf = mddev->private;
4977 	int cnt = 0;
4978 	int hash;
4979 
4980 	if (cb->list.next && !list_empty(&cb->list)) {
4981 		spin_lock_irq(&conf->device_lock);
4982 		while (!list_empty(&cb->list)) {
4983 			sh = list_first_entry(&cb->list, struct stripe_head, lru);
4984 			list_del_init(&sh->lru);
4985 			/*
4986 			 * avoid race release_stripe_plug() sees
4987 			 * STRIPE_ON_UNPLUG_LIST clear but the stripe
4988 			 * is still in our list
4989 			 */
4990 			smp_mb__before_atomic();
4991 			clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
4992 			/*
4993 			 * STRIPE_ON_RELEASE_LIST could be set here. In that
4994 			 * case, the count is always > 1 here
4995 			 */
4996 			hash = sh->hash_lock_index;
4997 			__release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
4998 			cnt++;
4999 		}
5000 		spin_unlock_irq(&conf->device_lock);
5001 	}
5002 	release_inactive_stripe_list(conf, cb->temp_inactive_list,
5003 				     NR_STRIPE_HASH_LOCKS);
5004 	if (mddev->queue)
5005 		trace_block_unplug(mddev->queue, cnt, !from_schedule);
5006 	kfree(cb);
5007 }
5008 
5009 static void release_stripe_plug(struct mddev *mddev,
5010 				struct stripe_head *sh)
5011 {
5012 	struct blk_plug_cb *blk_cb = blk_check_plugged(
5013 		raid5_unplug, mddev,
5014 		sizeof(struct raid5_plug_cb));
5015 	struct raid5_plug_cb *cb;
5016 
5017 	if (!blk_cb) {
5018 		release_stripe(sh);
5019 		return;
5020 	}
5021 
5022 	cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5023 
5024 	if (cb->list.next == NULL) {
5025 		int i;
5026 		INIT_LIST_HEAD(&cb->list);
5027 		for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5028 			INIT_LIST_HEAD(cb->temp_inactive_list + i);
5029 	}
5030 
5031 	if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5032 		list_add_tail(&sh->lru, &cb->list);
5033 	else
5034 		release_stripe(sh);
5035 }
5036 
5037 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5038 {
5039 	struct r5conf *conf = mddev->private;
5040 	sector_t logical_sector, last_sector;
5041 	struct stripe_head *sh;
5042 	int remaining;
5043 	int stripe_sectors;
5044 
5045 	if (mddev->reshape_position != MaxSector)
5046 		/* Skip discard while reshape is happening */
5047 		return;
5048 
5049 	logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5050 	last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5051 
5052 	bi->bi_next = NULL;
5053 	bi->bi_phys_segments = 1; /* over-loaded to count active stripes */
5054 
5055 	stripe_sectors = conf->chunk_sectors *
5056 		(conf->raid_disks - conf->max_degraded);
5057 	logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5058 					       stripe_sectors);
5059 	sector_div(last_sector, stripe_sectors);
5060 
5061 	logical_sector *= conf->chunk_sectors;
5062 	last_sector *= conf->chunk_sectors;
5063 
5064 	for (; logical_sector < last_sector;
5065 	     logical_sector += STRIPE_SECTORS) {
5066 		DEFINE_WAIT(w);
5067 		int d;
5068 	again:
5069 		sh = get_active_stripe(conf, logical_sector, 0, 0, 0);
5070 		prepare_to_wait(&conf->wait_for_overlap, &w,
5071 				TASK_UNINTERRUPTIBLE);
5072 		set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5073 		if (test_bit(STRIPE_SYNCING, &sh->state)) {
5074 			release_stripe(sh);
5075 			schedule();
5076 			goto again;
5077 		}
5078 		clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5079 		spin_lock_irq(&sh->stripe_lock);
5080 		for (d = 0; d < conf->raid_disks; d++) {
5081 			if (d == sh->pd_idx || d == sh->qd_idx)
5082 				continue;
5083 			if (sh->dev[d].towrite || sh->dev[d].toread) {
5084 				set_bit(R5_Overlap, &sh->dev[d].flags);
5085 				spin_unlock_irq(&sh->stripe_lock);
5086 				release_stripe(sh);
5087 				schedule();
5088 				goto again;
5089 			}
5090 		}
5091 		set_bit(STRIPE_DISCARD, &sh->state);
5092 		finish_wait(&conf->wait_for_overlap, &w);
5093 		sh->overwrite_disks = 0;
5094 		for (d = 0; d < conf->raid_disks; d++) {
5095 			if (d == sh->pd_idx || d == sh->qd_idx)
5096 				continue;
5097 			sh->dev[d].towrite = bi;
5098 			set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5099 			raid5_inc_bi_active_stripes(bi);
5100 			sh->overwrite_disks++;
5101 		}
5102 		spin_unlock_irq(&sh->stripe_lock);
5103 		if (conf->mddev->bitmap) {
5104 			for (d = 0;
5105 			     d < conf->raid_disks - conf->max_degraded;
5106 			     d++)
5107 				bitmap_startwrite(mddev->bitmap,
5108 						  sh->sector,
5109 						  STRIPE_SECTORS,
5110 						  0);
5111 			sh->bm_seq = conf->seq_flush + 1;
5112 			set_bit(STRIPE_BIT_DELAY, &sh->state);
5113 		}
5114 
5115 		set_bit(STRIPE_HANDLE, &sh->state);
5116 		clear_bit(STRIPE_DELAYED, &sh->state);
5117 		if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5118 			atomic_inc(&conf->preread_active_stripes);
5119 		release_stripe_plug(mddev, sh);
5120 	}
5121 
5122 	remaining = raid5_dec_bi_active_stripes(bi);
5123 	if (remaining == 0) {
5124 		md_write_end(mddev);
5125 		bio_endio(bi, 0);
5126 	}
5127 }
5128 
5129 static void make_request(struct mddev *mddev, struct bio * bi)
5130 {
5131 	struct r5conf *conf = mddev->private;
5132 	int dd_idx;
5133 	sector_t new_sector;
5134 	sector_t logical_sector, last_sector;
5135 	struct stripe_head *sh;
5136 	const int rw = bio_data_dir(bi);
5137 	int remaining;
5138 	DEFINE_WAIT(w);
5139 	bool do_prepare;
5140 
5141 	if (unlikely(bi->bi_rw & REQ_FLUSH)) {
5142 		md_flush_request(mddev, bi);
5143 		return;
5144 	}
5145 
5146 	md_write_start(mddev, bi);
5147 
5148 	/*
5149 	 * If array is degraded, better not do chunk aligned read because
5150 	 * later we might have to read it again in order to reconstruct
5151 	 * data on failed drives.
5152 	 */
5153 	if (rw == READ && mddev->degraded == 0 &&
5154 	     mddev->reshape_position == MaxSector &&
5155 	     chunk_aligned_read(mddev,bi))
5156 		return;
5157 
5158 	if (unlikely(bi->bi_rw & REQ_DISCARD)) {
5159 		make_discard_request(mddev, bi);
5160 		return;
5161 	}
5162 
5163 	logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5164 	last_sector = bio_end_sector(bi);
5165 	bi->bi_next = NULL;
5166 	bi->bi_phys_segments = 1;	/* over-loaded to count active stripes */
5167 
5168 	prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5169 	for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5170 		int previous;
5171 		int seq;
5172 
5173 		do_prepare = false;
5174 	retry:
5175 		seq = read_seqcount_begin(&conf->gen_lock);
5176 		previous = 0;
5177 		if (do_prepare)
5178 			prepare_to_wait(&conf->wait_for_overlap, &w,
5179 				TASK_UNINTERRUPTIBLE);
5180 		if (unlikely(conf->reshape_progress != MaxSector)) {
5181 			/* spinlock is needed as reshape_progress may be
5182 			 * 64bit on a 32bit platform, and so it might be
5183 			 * possible to see a half-updated value
5184 			 * Of course reshape_progress could change after
5185 			 * the lock is dropped, so once we get a reference
5186 			 * to the stripe that we think it is, we will have
5187 			 * to check again.
5188 			 */
5189 			spin_lock_irq(&conf->device_lock);
5190 			if (mddev->reshape_backwards
5191 			    ? logical_sector < conf->reshape_progress
5192 			    : logical_sector >= conf->reshape_progress) {
5193 				previous = 1;
5194 			} else {
5195 				if (mddev->reshape_backwards
5196 				    ? logical_sector < conf->reshape_safe
5197 				    : logical_sector >= conf->reshape_safe) {
5198 					spin_unlock_irq(&conf->device_lock);
5199 					schedule();
5200 					do_prepare = true;
5201 					goto retry;
5202 				}
5203 			}
5204 			spin_unlock_irq(&conf->device_lock);
5205 		}
5206 
5207 		new_sector = raid5_compute_sector(conf, logical_sector,
5208 						  previous,
5209 						  &dd_idx, NULL);
5210 		pr_debug("raid456: make_request, sector %llu logical %llu\n",
5211 			(unsigned long long)new_sector,
5212 			(unsigned long long)logical_sector);
5213 
5214 		sh = get_active_stripe(conf, new_sector, previous,
5215 				       (bi->bi_rw&RWA_MASK), 0);
5216 		if (sh) {
5217 			if (unlikely(previous)) {
5218 				/* expansion might have moved on while waiting for a
5219 				 * stripe, so we must do the range check again.
5220 				 * Expansion could still move past after this
5221 				 * test, but as we are holding a reference to
5222 				 * 'sh', we know that if that happens,
5223 				 *  STRIPE_EXPANDING will get set and the expansion
5224 				 * won't proceed until we finish with the stripe.
5225 				 */
5226 				int must_retry = 0;
5227 				spin_lock_irq(&conf->device_lock);
5228 				if (mddev->reshape_backwards
5229 				    ? logical_sector >= conf->reshape_progress
5230 				    : logical_sector < conf->reshape_progress)
5231 					/* mismatch, need to try again */
5232 					must_retry = 1;
5233 				spin_unlock_irq(&conf->device_lock);
5234 				if (must_retry) {
5235 					release_stripe(sh);
5236 					schedule();
5237 					do_prepare = true;
5238 					goto retry;
5239 				}
5240 			}
5241 			if (read_seqcount_retry(&conf->gen_lock, seq)) {
5242 				/* Might have got the wrong stripe_head
5243 				 * by accident
5244 				 */
5245 				release_stripe(sh);
5246 				goto retry;
5247 			}
5248 
5249 			if (rw == WRITE &&
5250 			    logical_sector >= mddev->suspend_lo &&
5251 			    logical_sector < mddev->suspend_hi) {
5252 				release_stripe(sh);
5253 				/* As the suspend_* range is controlled by
5254 				 * userspace, we want an interruptible
5255 				 * wait.
5256 				 */
5257 				flush_signals(current);
5258 				prepare_to_wait(&conf->wait_for_overlap,
5259 						&w, TASK_INTERRUPTIBLE);
5260 				if (logical_sector >= mddev->suspend_lo &&
5261 				    logical_sector < mddev->suspend_hi) {
5262 					schedule();
5263 					do_prepare = true;
5264 				}
5265 				goto retry;
5266 			}
5267 
5268 			if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5269 			    !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5270 				/* Stripe is busy expanding or
5271 				 * add failed due to overlap.  Flush everything
5272 				 * and wait a while
5273 				 */
5274 				md_wakeup_thread(mddev->thread);
5275 				release_stripe(sh);
5276 				schedule();
5277 				do_prepare = true;
5278 				goto retry;
5279 			}
5280 			set_bit(STRIPE_HANDLE, &sh->state);
5281 			clear_bit(STRIPE_DELAYED, &sh->state);
5282 			if ((!sh->batch_head || sh == sh->batch_head) &&
5283 			    (bi->bi_rw & REQ_SYNC) &&
5284 			    !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5285 				atomic_inc(&conf->preread_active_stripes);
5286 			release_stripe_plug(mddev, sh);
5287 		} else {
5288 			/* cannot get stripe for read-ahead, just give-up */
5289 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
5290 			break;
5291 		}
5292 	}
5293 	finish_wait(&conf->wait_for_overlap, &w);
5294 
5295 	remaining = raid5_dec_bi_active_stripes(bi);
5296 	if (remaining == 0) {
5297 
5298 		if ( rw == WRITE )
5299 			md_write_end(mddev);
5300 
5301 		trace_block_bio_complete(bdev_get_queue(bi->bi_bdev),
5302 					 bi, 0);
5303 		bio_endio(bi, 0);
5304 	}
5305 }
5306 
5307 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5308 
5309 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5310 {
5311 	/* reshaping is quite different to recovery/resync so it is
5312 	 * handled quite separately ... here.
5313 	 *
5314 	 * On each call to sync_request, we gather one chunk worth of
5315 	 * destination stripes and flag them as expanding.
5316 	 * Then we find all the source stripes and request reads.
5317 	 * As the reads complete, handle_stripe will copy the data
5318 	 * into the destination stripe and release that stripe.
5319 	 */
5320 	struct r5conf *conf = mddev->private;
5321 	struct stripe_head *sh;
5322 	sector_t first_sector, last_sector;
5323 	int raid_disks = conf->previous_raid_disks;
5324 	int data_disks = raid_disks - conf->max_degraded;
5325 	int new_data_disks = conf->raid_disks - conf->max_degraded;
5326 	int i;
5327 	int dd_idx;
5328 	sector_t writepos, readpos, safepos;
5329 	sector_t stripe_addr;
5330 	int reshape_sectors;
5331 	struct list_head stripes;
5332 
5333 	if (sector_nr == 0) {
5334 		/* If restarting in the middle, skip the initial sectors */
5335 		if (mddev->reshape_backwards &&
5336 		    conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5337 			sector_nr = raid5_size(mddev, 0, 0)
5338 				- conf->reshape_progress;
5339 		} else if (!mddev->reshape_backwards &&
5340 			   conf->reshape_progress > 0)
5341 			sector_nr = conf->reshape_progress;
5342 		sector_div(sector_nr, new_data_disks);
5343 		if (sector_nr) {
5344 			mddev->curr_resync_completed = sector_nr;
5345 			sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5346 			*skipped = 1;
5347 			return sector_nr;
5348 		}
5349 	}
5350 
5351 	/* We need to process a full chunk at a time.
5352 	 * If old and new chunk sizes differ, we need to process the
5353 	 * largest of these
5354 	 */
5355 	if (mddev->new_chunk_sectors > mddev->chunk_sectors)
5356 		reshape_sectors = mddev->new_chunk_sectors;
5357 	else
5358 		reshape_sectors = mddev->chunk_sectors;
5359 
5360 	/* We update the metadata at least every 10 seconds, or when
5361 	 * the data about to be copied would over-write the source of
5362 	 * the data at the front of the range.  i.e. one new_stripe
5363 	 * along from reshape_progress new_maps to after where
5364 	 * reshape_safe old_maps to
5365 	 */
5366 	writepos = conf->reshape_progress;
5367 	sector_div(writepos, new_data_disks);
5368 	readpos = conf->reshape_progress;
5369 	sector_div(readpos, data_disks);
5370 	safepos = conf->reshape_safe;
5371 	sector_div(safepos, data_disks);
5372 	if (mddev->reshape_backwards) {
5373 		writepos -= min_t(sector_t, reshape_sectors, writepos);
5374 		readpos += reshape_sectors;
5375 		safepos += reshape_sectors;
5376 	} else {
5377 		writepos += reshape_sectors;
5378 		readpos -= min_t(sector_t, reshape_sectors, readpos);
5379 		safepos -= min_t(sector_t, reshape_sectors, safepos);
5380 	}
5381 
5382 	/* Having calculated the 'writepos' possibly use it
5383 	 * to set 'stripe_addr' which is where we will write to.
5384 	 */
5385 	if (mddev->reshape_backwards) {
5386 		BUG_ON(conf->reshape_progress == 0);
5387 		stripe_addr = writepos;
5388 		BUG_ON((mddev->dev_sectors &
5389 			~((sector_t)reshape_sectors - 1))
5390 		       - reshape_sectors - stripe_addr
5391 		       != sector_nr);
5392 	} else {
5393 		BUG_ON(writepos != sector_nr + reshape_sectors);
5394 		stripe_addr = sector_nr;
5395 	}
5396 
5397 	/* 'writepos' is the most advanced device address we might write.
5398 	 * 'readpos' is the least advanced device address we might read.
5399 	 * 'safepos' is the least address recorded in the metadata as having
5400 	 *     been reshaped.
5401 	 * If there is a min_offset_diff, these are adjusted either by
5402 	 * increasing the safepos/readpos if diff is negative, or
5403 	 * increasing writepos if diff is positive.
5404 	 * If 'readpos' is then behind 'writepos', there is no way that we can
5405 	 * ensure safety in the face of a crash - that must be done by userspace
5406 	 * making a backup of the data.  So in that case there is no particular
5407 	 * rush to update metadata.
5408 	 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5409 	 * update the metadata to advance 'safepos' to match 'readpos' so that
5410 	 * we can be safe in the event of a crash.
5411 	 * So we insist on updating metadata if safepos is behind writepos and
5412 	 * readpos is beyond writepos.
5413 	 * In any case, update the metadata every 10 seconds.
5414 	 * Maybe that number should be configurable, but I'm not sure it is
5415 	 * worth it.... maybe it could be a multiple of safemode_delay???
5416 	 */
5417 	if (conf->min_offset_diff < 0) {
5418 		safepos += -conf->min_offset_diff;
5419 		readpos += -conf->min_offset_diff;
5420 	} else
5421 		writepos += conf->min_offset_diff;
5422 
5423 	if ((mddev->reshape_backwards
5424 	     ? (safepos > writepos && readpos < writepos)
5425 	     : (safepos < writepos && readpos > writepos)) ||
5426 	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5427 		/* Cannot proceed until we've updated the superblock... */
5428 		wait_event(conf->wait_for_overlap,
5429 			   atomic_read(&conf->reshape_stripes)==0
5430 			   || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5431 		if (atomic_read(&conf->reshape_stripes) != 0)
5432 			return 0;
5433 		mddev->reshape_position = conf->reshape_progress;
5434 		mddev->curr_resync_completed = sector_nr;
5435 		conf->reshape_checkpoint = jiffies;
5436 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
5437 		md_wakeup_thread(mddev->thread);
5438 		wait_event(mddev->sb_wait, mddev->flags == 0 ||
5439 			   test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5440 		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5441 			return 0;
5442 		spin_lock_irq(&conf->device_lock);
5443 		conf->reshape_safe = mddev->reshape_position;
5444 		spin_unlock_irq(&conf->device_lock);
5445 		wake_up(&conf->wait_for_overlap);
5446 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5447 	}
5448 
5449 	INIT_LIST_HEAD(&stripes);
5450 	for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5451 		int j;
5452 		int skipped_disk = 0;
5453 		sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5454 		set_bit(STRIPE_EXPANDING, &sh->state);
5455 		atomic_inc(&conf->reshape_stripes);
5456 		/* If any of this stripe is beyond the end of the old
5457 		 * array, then we need to zero those blocks
5458 		 */
5459 		for (j=sh->disks; j--;) {
5460 			sector_t s;
5461 			if (j == sh->pd_idx)
5462 				continue;
5463 			if (conf->level == 6 &&
5464 			    j == sh->qd_idx)
5465 				continue;
5466 			s = compute_blocknr(sh, j, 0);
5467 			if (s < raid5_size(mddev, 0, 0)) {
5468 				skipped_disk = 1;
5469 				continue;
5470 			}
5471 			memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5472 			set_bit(R5_Expanded, &sh->dev[j].flags);
5473 			set_bit(R5_UPTODATE, &sh->dev[j].flags);
5474 		}
5475 		if (!skipped_disk) {
5476 			set_bit(STRIPE_EXPAND_READY, &sh->state);
5477 			set_bit(STRIPE_HANDLE, &sh->state);
5478 		}
5479 		list_add(&sh->lru, &stripes);
5480 	}
5481 	spin_lock_irq(&conf->device_lock);
5482 	if (mddev->reshape_backwards)
5483 		conf->reshape_progress -= reshape_sectors * new_data_disks;
5484 	else
5485 		conf->reshape_progress += reshape_sectors * new_data_disks;
5486 	spin_unlock_irq(&conf->device_lock);
5487 	/* Ok, those stripe are ready. We can start scheduling
5488 	 * reads on the source stripes.
5489 	 * The source stripes are determined by mapping the first and last
5490 	 * block on the destination stripes.
5491 	 */
5492 	first_sector =
5493 		raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5494 				     1, &dd_idx, NULL);
5495 	last_sector =
5496 		raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5497 					    * new_data_disks - 1),
5498 				     1, &dd_idx, NULL);
5499 	if (last_sector >= mddev->dev_sectors)
5500 		last_sector = mddev->dev_sectors - 1;
5501 	while (first_sector <= last_sector) {
5502 		sh = get_active_stripe(conf, first_sector, 1, 0, 1);
5503 		set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5504 		set_bit(STRIPE_HANDLE, &sh->state);
5505 		release_stripe(sh);
5506 		first_sector += STRIPE_SECTORS;
5507 	}
5508 	/* Now that the sources are clearly marked, we can release
5509 	 * the destination stripes
5510 	 */
5511 	while (!list_empty(&stripes)) {
5512 		sh = list_entry(stripes.next, struct stripe_head, lru);
5513 		list_del_init(&sh->lru);
5514 		release_stripe(sh);
5515 	}
5516 	/* If this takes us to the resync_max point where we have to pause,
5517 	 * then we need to write out the superblock.
5518 	 */
5519 	sector_nr += reshape_sectors;
5520 	if ((sector_nr - mddev->curr_resync_completed) * 2
5521 	    >= mddev->resync_max - mddev->curr_resync_completed) {
5522 		/* Cannot proceed until we've updated the superblock... */
5523 		wait_event(conf->wait_for_overlap,
5524 			   atomic_read(&conf->reshape_stripes) == 0
5525 			   || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5526 		if (atomic_read(&conf->reshape_stripes) != 0)
5527 			goto ret;
5528 		mddev->reshape_position = conf->reshape_progress;
5529 		mddev->curr_resync_completed = sector_nr;
5530 		conf->reshape_checkpoint = jiffies;
5531 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
5532 		md_wakeup_thread(mddev->thread);
5533 		wait_event(mddev->sb_wait,
5534 			   !test_bit(MD_CHANGE_DEVS, &mddev->flags)
5535 			   || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5536 		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5537 			goto ret;
5538 		spin_lock_irq(&conf->device_lock);
5539 		conf->reshape_safe = mddev->reshape_position;
5540 		spin_unlock_irq(&conf->device_lock);
5541 		wake_up(&conf->wait_for_overlap);
5542 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5543 	}
5544 ret:
5545 	return reshape_sectors;
5546 }
5547 
5548 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5549 {
5550 	struct r5conf *conf = mddev->private;
5551 	struct stripe_head *sh;
5552 	sector_t max_sector = mddev->dev_sectors;
5553 	sector_t sync_blocks;
5554 	int still_degraded = 0;
5555 	int i;
5556 
5557 	if (sector_nr >= max_sector) {
5558 		/* just being told to finish up .. nothing much to do */
5559 
5560 		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
5561 			end_reshape(conf);
5562 			return 0;
5563 		}
5564 
5565 		if (mddev->curr_resync < max_sector) /* aborted */
5566 			bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
5567 					&sync_blocks, 1);
5568 		else /* completed sync */
5569 			conf->fullsync = 0;
5570 		bitmap_close_sync(mddev->bitmap);
5571 
5572 		return 0;
5573 	}
5574 
5575 	/* Allow raid5_quiesce to complete */
5576 	wait_event(conf->wait_for_overlap, conf->quiesce != 2);
5577 
5578 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
5579 		return reshape_request(mddev, sector_nr, skipped);
5580 
5581 	/* No need to check resync_max as we never do more than one
5582 	 * stripe, and as resync_max will always be on a chunk boundary,
5583 	 * if the check in md_do_sync didn't fire, there is no chance
5584 	 * of overstepping resync_max here
5585 	 */
5586 
5587 	/* if there is too many failed drives and we are trying
5588 	 * to resync, then assert that we are finished, because there is
5589 	 * nothing we can do.
5590 	 */
5591 	if (mddev->degraded >= conf->max_degraded &&
5592 	    test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
5593 		sector_t rv = mddev->dev_sectors - sector_nr;
5594 		*skipped = 1;
5595 		return rv;
5596 	}
5597 	if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
5598 	    !conf->fullsync &&
5599 	    !bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
5600 	    sync_blocks >= STRIPE_SECTORS) {
5601 		/* we can skip this block, and probably more */
5602 		sync_blocks /= STRIPE_SECTORS;
5603 		*skipped = 1;
5604 		return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
5605 	}
5606 
5607 	bitmap_cond_end_sync(mddev->bitmap, sector_nr);
5608 
5609 	sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
5610 	if (sh == NULL) {
5611 		sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
5612 		/* make sure we don't swamp the stripe cache if someone else
5613 		 * is trying to get access
5614 		 */
5615 		schedule_timeout_uninterruptible(1);
5616 	}
5617 	/* Need to check if array will still be degraded after recovery/resync
5618 	 * Note in case of > 1 drive failures it's possible we're rebuilding
5619 	 * one drive while leaving another faulty drive in array.
5620 	 */
5621 	rcu_read_lock();
5622 	for (i = 0; i < conf->raid_disks; i++) {
5623 		struct md_rdev *rdev = ACCESS_ONCE(conf->disks[i].rdev);
5624 
5625 		if (rdev == NULL || test_bit(Faulty, &rdev->flags))
5626 			still_degraded = 1;
5627 	}
5628 	rcu_read_unlock();
5629 
5630 	bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
5631 
5632 	set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
5633 	set_bit(STRIPE_HANDLE, &sh->state);
5634 
5635 	release_stripe(sh);
5636 
5637 	return STRIPE_SECTORS;
5638 }
5639 
5640 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
5641 {
5642 	/* We may not be able to submit a whole bio at once as there
5643 	 * may not be enough stripe_heads available.
5644 	 * We cannot pre-allocate enough stripe_heads as we may need
5645 	 * more than exist in the cache (if we allow ever large chunks).
5646 	 * So we do one stripe head at a time and record in
5647 	 * ->bi_hw_segments how many have been done.
5648 	 *
5649 	 * We *know* that this entire raid_bio is in one chunk, so
5650 	 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
5651 	 */
5652 	struct stripe_head *sh;
5653 	int dd_idx;
5654 	sector_t sector, logical_sector, last_sector;
5655 	int scnt = 0;
5656 	int remaining;
5657 	int handled = 0;
5658 
5659 	logical_sector = raid_bio->bi_iter.bi_sector &
5660 		~((sector_t)STRIPE_SECTORS-1);
5661 	sector = raid5_compute_sector(conf, logical_sector,
5662 				      0, &dd_idx, NULL);
5663 	last_sector = bio_end_sector(raid_bio);
5664 
5665 	for (; logical_sector < last_sector;
5666 	     logical_sector += STRIPE_SECTORS,
5667 		     sector += STRIPE_SECTORS,
5668 		     scnt++) {
5669 
5670 		if (scnt < raid5_bi_processed_stripes(raid_bio))
5671 			/* already done this stripe */
5672 			continue;
5673 
5674 		sh = get_active_stripe(conf, sector, 0, 1, 1);
5675 
5676 		if (!sh) {
5677 			/* failed to get a stripe - must wait */
5678 			raid5_set_bi_processed_stripes(raid_bio, scnt);
5679 			conf->retry_read_aligned = raid_bio;
5680 			return handled;
5681 		}
5682 
5683 		if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
5684 			release_stripe(sh);
5685 			raid5_set_bi_processed_stripes(raid_bio, scnt);
5686 			conf->retry_read_aligned = raid_bio;
5687 			return handled;
5688 		}
5689 
5690 		set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
5691 		handle_stripe(sh);
5692 		release_stripe(sh);
5693 		handled++;
5694 	}
5695 	remaining = raid5_dec_bi_active_stripes(raid_bio);
5696 	if (remaining == 0) {
5697 		trace_block_bio_complete(bdev_get_queue(raid_bio->bi_bdev),
5698 					 raid_bio, 0);
5699 		bio_endio(raid_bio, 0);
5700 	}
5701 	if (atomic_dec_and_test(&conf->active_aligned_reads))
5702 		wake_up(&conf->wait_for_stripe);
5703 	return handled;
5704 }
5705 
5706 static int handle_active_stripes(struct r5conf *conf, int group,
5707 				 struct r5worker *worker,
5708 				 struct list_head *temp_inactive_list)
5709 {
5710 	struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
5711 	int i, batch_size = 0, hash;
5712 	bool release_inactive = false;
5713 
5714 	while (batch_size < MAX_STRIPE_BATCH &&
5715 			(sh = __get_priority_stripe(conf, group)) != NULL)
5716 		batch[batch_size++] = sh;
5717 
5718 	if (batch_size == 0) {
5719 		for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5720 			if (!list_empty(temp_inactive_list + i))
5721 				break;
5722 		if (i == NR_STRIPE_HASH_LOCKS)
5723 			return batch_size;
5724 		release_inactive = true;
5725 	}
5726 	spin_unlock_irq(&conf->device_lock);
5727 
5728 	release_inactive_stripe_list(conf, temp_inactive_list,
5729 				     NR_STRIPE_HASH_LOCKS);
5730 
5731 	if (release_inactive) {
5732 		spin_lock_irq(&conf->device_lock);
5733 		return 0;
5734 	}
5735 
5736 	for (i = 0; i < batch_size; i++)
5737 		handle_stripe(batch[i]);
5738 
5739 	cond_resched();
5740 
5741 	spin_lock_irq(&conf->device_lock);
5742 	for (i = 0; i < batch_size; i++) {
5743 		hash = batch[i]->hash_lock_index;
5744 		__release_stripe(conf, batch[i], &temp_inactive_list[hash]);
5745 	}
5746 	return batch_size;
5747 }
5748 
5749 static void raid5_do_work(struct work_struct *work)
5750 {
5751 	struct r5worker *worker = container_of(work, struct r5worker, work);
5752 	struct r5worker_group *group = worker->group;
5753 	struct r5conf *conf = group->conf;
5754 	int group_id = group - conf->worker_groups;
5755 	int handled;
5756 	struct blk_plug plug;
5757 
5758 	pr_debug("+++ raid5worker active\n");
5759 
5760 	blk_start_plug(&plug);
5761 	handled = 0;
5762 	spin_lock_irq(&conf->device_lock);
5763 	while (1) {
5764 		int batch_size, released;
5765 
5766 		released = release_stripe_list(conf, worker->temp_inactive_list);
5767 
5768 		batch_size = handle_active_stripes(conf, group_id, worker,
5769 						   worker->temp_inactive_list);
5770 		worker->working = false;
5771 		if (!batch_size && !released)
5772 			break;
5773 		handled += batch_size;
5774 	}
5775 	pr_debug("%d stripes handled\n", handled);
5776 
5777 	spin_unlock_irq(&conf->device_lock);
5778 	blk_finish_plug(&plug);
5779 
5780 	pr_debug("--- raid5worker inactive\n");
5781 }
5782 
5783 /*
5784  * This is our raid5 kernel thread.
5785  *
5786  * We scan the hash table for stripes which can be handled now.
5787  * During the scan, completed stripes are saved for us by the interrupt
5788  * handler, so that they will not have to wait for our next wakeup.
5789  */
5790 static void raid5d(struct md_thread *thread)
5791 {
5792 	struct mddev *mddev = thread->mddev;
5793 	struct r5conf *conf = mddev->private;
5794 	int handled;
5795 	struct blk_plug plug;
5796 
5797 	pr_debug("+++ raid5d active\n");
5798 
5799 	md_check_recovery(mddev);
5800 
5801 	blk_start_plug(&plug);
5802 	handled = 0;
5803 	spin_lock_irq(&conf->device_lock);
5804 	while (1) {
5805 		struct bio *bio;
5806 		int batch_size, released;
5807 
5808 		released = release_stripe_list(conf, conf->temp_inactive_list);
5809 		if (released)
5810 			clear_bit(R5_DID_ALLOC, &conf->cache_state);
5811 
5812 		if (
5813 		    !list_empty(&conf->bitmap_list)) {
5814 			/* Now is a good time to flush some bitmap updates */
5815 			conf->seq_flush++;
5816 			spin_unlock_irq(&conf->device_lock);
5817 			bitmap_unplug(mddev->bitmap);
5818 			spin_lock_irq(&conf->device_lock);
5819 			conf->seq_write = conf->seq_flush;
5820 			activate_bit_delay(conf, conf->temp_inactive_list);
5821 		}
5822 		raid5_activate_delayed(conf);
5823 
5824 		while ((bio = remove_bio_from_retry(conf))) {
5825 			int ok;
5826 			spin_unlock_irq(&conf->device_lock);
5827 			ok = retry_aligned_read(conf, bio);
5828 			spin_lock_irq(&conf->device_lock);
5829 			if (!ok)
5830 				break;
5831 			handled++;
5832 		}
5833 
5834 		batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
5835 						   conf->temp_inactive_list);
5836 		if (!batch_size && !released)
5837 			break;
5838 		handled += batch_size;
5839 
5840 		if (mddev->flags & ~(1<<MD_CHANGE_PENDING)) {
5841 			spin_unlock_irq(&conf->device_lock);
5842 			md_check_recovery(mddev);
5843 			spin_lock_irq(&conf->device_lock);
5844 		}
5845 	}
5846 	pr_debug("%d stripes handled\n", handled);
5847 
5848 	spin_unlock_irq(&conf->device_lock);
5849 	if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state)) {
5850 		grow_one_stripe(conf, __GFP_NOWARN);
5851 		/* Set flag even if allocation failed.  This helps
5852 		 * slow down allocation requests when mem is short
5853 		 */
5854 		set_bit(R5_DID_ALLOC, &conf->cache_state);
5855 	}
5856 
5857 	async_tx_issue_pending_all();
5858 	blk_finish_plug(&plug);
5859 
5860 	pr_debug("--- raid5d inactive\n");
5861 }
5862 
5863 static ssize_t
5864 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
5865 {
5866 	struct r5conf *conf;
5867 	int ret = 0;
5868 	spin_lock(&mddev->lock);
5869 	conf = mddev->private;
5870 	if (conf)
5871 		ret = sprintf(page, "%d\n", conf->min_nr_stripes);
5872 	spin_unlock(&mddev->lock);
5873 	return ret;
5874 }
5875 
5876 int
5877 raid5_set_cache_size(struct mddev *mddev, int size)
5878 {
5879 	struct r5conf *conf = mddev->private;
5880 	int err;
5881 
5882 	if (size <= 16 || size > 32768)
5883 		return -EINVAL;
5884 
5885 	conf->min_nr_stripes = size;
5886 	while (size < conf->max_nr_stripes &&
5887 	       drop_one_stripe(conf))
5888 		;
5889 
5890 
5891 	err = md_allow_write(mddev);
5892 	if (err)
5893 		return err;
5894 
5895 	while (size > conf->max_nr_stripes)
5896 		if (!grow_one_stripe(conf, GFP_KERNEL))
5897 			break;
5898 
5899 	return 0;
5900 }
5901 EXPORT_SYMBOL(raid5_set_cache_size);
5902 
5903 static ssize_t
5904 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
5905 {
5906 	struct r5conf *conf;
5907 	unsigned long new;
5908 	int err;
5909 
5910 	if (len >= PAGE_SIZE)
5911 		return -EINVAL;
5912 	if (kstrtoul(page, 10, &new))
5913 		return -EINVAL;
5914 	err = mddev_lock(mddev);
5915 	if (err)
5916 		return err;
5917 	conf = mddev->private;
5918 	if (!conf)
5919 		err = -ENODEV;
5920 	else
5921 		err = raid5_set_cache_size(mddev, new);
5922 	mddev_unlock(mddev);
5923 
5924 	return err ?: len;
5925 }
5926 
5927 static struct md_sysfs_entry
5928 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
5929 				raid5_show_stripe_cache_size,
5930 				raid5_store_stripe_cache_size);
5931 
5932 static ssize_t
5933 raid5_show_rmw_level(struct mddev  *mddev, char *page)
5934 {
5935 	struct r5conf *conf = mddev->private;
5936 	if (conf)
5937 		return sprintf(page, "%d\n", conf->rmw_level);
5938 	else
5939 		return 0;
5940 }
5941 
5942 static ssize_t
5943 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
5944 {
5945 	struct r5conf *conf = mddev->private;
5946 	unsigned long new;
5947 
5948 	if (!conf)
5949 		return -ENODEV;
5950 
5951 	if (len >= PAGE_SIZE)
5952 		return -EINVAL;
5953 
5954 	if (kstrtoul(page, 10, &new))
5955 		return -EINVAL;
5956 
5957 	if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
5958 		return -EINVAL;
5959 
5960 	if (new != PARITY_DISABLE_RMW &&
5961 	    new != PARITY_ENABLE_RMW &&
5962 	    new != PARITY_PREFER_RMW)
5963 		return -EINVAL;
5964 
5965 	conf->rmw_level = new;
5966 	return len;
5967 }
5968 
5969 static struct md_sysfs_entry
5970 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
5971 			 raid5_show_rmw_level,
5972 			 raid5_store_rmw_level);
5973 
5974 
5975 static ssize_t
5976 raid5_show_preread_threshold(struct mddev *mddev, char *page)
5977 {
5978 	struct r5conf *conf;
5979 	int ret = 0;
5980 	spin_lock(&mddev->lock);
5981 	conf = mddev->private;
5982 	if (conf)
5983 		ret = sprintf(page, "%d\n", conf->bypass_threshold);
5984 	spin_unlock(&mddev->lock);
5985 	return ret;
5986 }
5987 
5988 static ssize_t
5989 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
5990 {
5991 	struct r5conf *conf;
5992 	unsigned long new;
5993 	int err;
5994 
5995 	if (len >= PAGE_SIZE)
5996 		return -EINVAL;
5997 	if (kstrtoul(page, 10, &new))
5998 		return -EINVAL;
5999 
6000 	err = mddev_lock(mddev);
6001 	if (err)
6002 		return err;
6003 	conf = mddev->private;
6004 	if (!conf)
6005 		err = -ENODEV;
6006 	else if (new > conf->min_nr_stripes)
6007 		err = -EINVAL;
6008 	else
6009 		conf->bypass_threshold = new;
6010 	mddev_unlock(mddev);
6011 	return err ?: len;
6012 }
6013 
6014 static struct md_sysfs_entry
6015 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6016 					S_IRUGO | S_IWUSR,
6017 					raid5_show_preread_threshold,
6018 					raid5_store_preread_threshold);
6019 
6020 static ssize_t
6021 raid5_show_skip_copy(struct mddev *mddev, char *page)
6022 {
6023 	struct r5conf *conf;
6024 	int ret = 0;
6025 	spin_lock(&mddev->lock);
6026 	conf = mddev->private;
6027 	if (conf)
6028 		ret = sprintf(page, "%d\n", conf->skip_copy);
6029 	spin_unlock(&mddev->lock);
6030 	return ret;
6031 }
6032 
6033 static ssize_t
6034 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6035 {
6036 	struct r5conf *conf;
6037 	unsigned long new;
6038 	int err;
6039 
6040 	if (len >= PAGE_SIZE)
6041 		return -EINVAL;
6042 	if (kstrtoul(page, 10, &new))
6043 		return -EINVAL;
6044 	new = !!new;
6045 
6046 	err = mddev_lock(mddev);
6047 	if (err)
6048 		return err;
6049 	conf = mddev->private;
6050 	if (!conf)
6051 		err = -ENODEV;
6052 	else if (new != conf->skip_copy) {
6053 		mddev_suspend(mddev);
6054 		conf->skip_copy = new;
6055 		if (new)
6056 			mddev->queue->backing_dev_info.capabilities |=
6057 				BDI_CAP_STABLE_WRITES;
6058 		else
6059 			mddev->queue->backing_dev_info.capabilities &=
6060 				~BDI_CAP_STABLE_WRITES;
6061 		mddev_resume(mddev);
6062 	}
6063 	mddev_unlock(mddev);
6064 	return err ?: len;
6065 }
6066 
6067 static struct md_sysfs_entry
6068 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6069 					raid5_show_skip_copy,
6070 					raid5_store_skip_copy);
6071 
6072 static ssize_t
6073 stripe_cache_active_show(struct mddev *mddev, char *page)
6074 {
6075 	struct r5conf *conf = mddev->private;
6076 	if (conf)
6077 		return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6078 	else
6079 		return 0;
6080 }
6081 
6082 static struct md_sysfs_entry
6083 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6084 
6085 static ssize_t
6086 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6087 {
6088 	struct r5conf *conf;
6089 	int ret = 0;
6090 	spin_lock(&mddev->lock);
6091 	conf = mddev->private;
6092 	if (conf)
6093 		ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6094 	spin_unlock(&mddev->lock);
6095 	return ret;
6096 }
6097 
6098 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6099 			       int *group_cnt,
6100 			       int *worker_cnt_per_group,
6101 			       struct r5worker_group **worker_groups);
6102 static ssize_t
6103 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6104 {
6105 	struct r5conf *conf;
6106 	unsigned long new;
6107 	int err;
6108 	struct r5worker_group *new_groups, *old_groups;
6109 	int group_cnt, worker_cnt_per_group;
6110 
6111 	if (len >= PAGE_SIZE)
6112 		return -EINVAL;
6113 	if (kstrtoul(page, 10, &new))
6114 		return -EINVAL;
6115 
6116 	err = mddev_lock(mddev);
6117 	if (err)
6118 		return err;
6119 	conf = mddev->private;
6120 	if (!conf)
6121 		err = -ENODEV;
6122 	else if (new != conf->worker_cnt_per_group) {
6123 		mddev_suspend(mddev);
6124 
6125 		old_groups = conf->worker_groups;
6126 		if (old_groups)
6127 			flush_workqueue(raid5_wq);
6128 
6129 		err = alloc_thread_groups(conf, new,
6130 					  &group_cnt, &worker_cnt_per_group,
6131 					  &new_groups);
6132 		if (!err) {
6133 			spin_lock_irq(&conf->device_lock);
6134 			conf->group_cnt = group_cnt;
6135 			conf->worker_cnt_per_group = worker_cnt_per_group;
6136 			conf->worker_groups = new_groups;
6137 			spin_unlock_irq(&conf->device_lock);
6138 
6139 			if (old_groups)
6140 				kfree(old_groups[0].workers);
6141 			kfree(old_groups);
6142 		}
6143 		mddev_resume(mddev);
6144 	}
6145 	mddev_unlock(mddev);
6146 
6147 	return err ?: len;
6148 }
6149 
6150 static struct md_sysfs_entry
6151 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6152 				raid5_show_group_thread_cnt,
6153 				raid5_store_group_thread_cnt);
6154 
6155 static struct attribute *raid5_attrs[] =  {
6156 	&raid5_stripecache_size.attr,
6157 	&raid5_stripecache_active.attr,
6158 	&raid5_preread_bypass_threshold.attr,
6159 	&raid5_group_thread_cnt.attr,
6160 	&raid5_skip_copy.attr,
6161 	&raid5_rmw_level.attr,
6162 	NULL,
6163 };
6164 static struct attribute_group raid5_attrs_group = {
6165 	.name = NULL,
6166 	.attrs = raid5_attrs,
6167 };
6168 
6169 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6170 			       int *group_cnt,
6171 			       int *worker_cnt_per_group,
6172 			       struct r5worker_group **worker_groups)
6173 {
6174 	int i, j, k;
6175 	ssize_t size;
6176 	struct r5worker *workers;
6177 
6178 	*worker_cnt_per_group = cnt;
6179 	if (cnt == 0) {
6180 		*group_cnt = 0;
6181 		*worker_groups = NULL;
6182 		return 0;
6183 	}
6184 	*group_cnt = num_possible_nodes();
6185 	size = sizeof(struct r5worker) * cnt;
6186 	workers = kzalloc(size * *group_cnt, GFP_NOIO);
6187 	*worker_groups = kzalloc(sizeof(struct r5worker_group) *
6188 				*group_cnt, GFP_NOIO);
6189 	if (!*worker_groups || !workers) {
6190 		kfree(workers);
6191 		kfree(*worker_groups);
6192 		return -ENOMEM;
6193 	}
6194 
6195 	for (i = 0; i < *group_cnt; i++) {
6196 		struct r5worker_group *group;
6197 
6198 		group = &(*worker_groups)[i];
6199 		INIT_LIST_HEAD(&group->handle_list);
6200 		group->conf = conf;
6201 		group->workers = workers + i * cnt;
6202 
6203 		for (j = 0; j < cnt; j++) {
6204 			struct r5worker *worker = group->workers + j;
6205 			worker->group = group;
6206 			INIT_WORK(&worker->work, raid5_do_work);
6207 
6208 			for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6209 				INIT_LIST_HEAD(worker->temp_inactive_list + k);
6210 		}
6211 	}
6212 
6213 	return 0;
6214 }
6215 
6216 static void free_thread_groups(struct r5conf *conf)
6217 {
6218 	if (conf->worker_groups)
6219 		kfree(conf->worker_groups[0].workers);
6220 	kfree(conf->worker_groups);
6221 	conf->worker_groups = NULL;
6222 }
6223 
6224 static sector_t
6225 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6226 {
6227 	struct r5conf *conf = mddev->private;
6228 
6229 	if (!sectors)
6230 		sectors = mddev->dev_sectors;
6231 	if (!raid_disks)
6232 		/* size is defined by the smallest of previous and new size */
6233 		raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6234 
6235 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
6236 	sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
6237 	return sectors * (raid_disks - conf->max_degraded);
6238 }
6239 
6240 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6241 {
6242 	safe_put_page(percpu->spare_page);
6243 	if (percpu->scribble)
6244 		flex_array_free(percpu->scribble);
6245 	percpu->spare_page = NULL;
6246 	percpu->scribble = NULL;
6247 }
6248 
6249 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6250 {
6251 	if (conf->level == 6 && !percpu->spare_page)
6252 		percpu->spare_page = alloc_page(GFP_KERNEL);
6253 	if (!percpu->scribble)
6254 		percpu->scribble = scribble_alloc(max(conf->raid_disks,
6255 						      conf->previous_raid_disks),
6256 						  max(conf->chunk_sectors,
6257 						      conf->prev_chunk_sectors)
6258 						   / STRIPE_SECTORS,
6259 						  GFP_KERNEL);
6260 
6261 	if (!percpu->scribble || (conf->level == 6 && !percpu->spare_page)) {
6262 		free_scratch_buffer(conf, percpu);
6263 		return -ENOMEM;
6264 	}
6265 
6266 	return 0;
6267 }
6268 
6269 static void raid5_free_percpu(struct r5conf *conf)
6270 {
6271 	unsigned long cpu;
6272 
6273 	if (!conf->percpu)
6274 		return;
6275 
6276 #ifdef CONFIG_HOTPLUG_CPU
6277 	unregister_cpu_notifier(&conf->cpu_notify);
6278 #endif
6279 
6280 	get_online_cpus();
6281 	for_each_possible_cpu(cpu)
6282 		free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6283 	put_online_cpus();
6284 
6285 	free_percpu(conf->percpu);
6286 }
6287 
6288 static void free_conf(struct r5conf *conf)
6289 {
6290 	if (conf->shrinker.seeks)
6291 		unregister_shrinker(&conf->shrinker);
6292 	free_thread_groups(conf);
6293 	shrink_stripes(conf);
6294 	raid5_free_percpu(conf);
6295 	kfree(conf->disks);
6296 	kfree(conf->stripe_hashtbl);
6297 	kfree(conf);
6298 }
6299 
6300 #ifdef CONFIG_HOTPLUG_CPU
6301 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
6302 			      void *hcpu)
6303 {
6304 	struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
6305 	long cpu = (long)hcpu;
6306 	struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6307 
6308 	switch (action) {
6309 	case CPU_UP_PREPARE:
6310 	case CPU_UP_PREPARE_FROZEN:
6311 		if (alloc_scratch_buffer(conf, percpu)) {
6312 			pr_err("%s: failed memory allocation for cpu%ld\n",
6313 			       __func__, cpu);
6314 			return notifier_from_errno(-ENOMEM);
6315 		}
6316 		break;
6317 	case CPU_DEAD:
6318 	case CPU_DEAD_FROZEN:
6319 		free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6320 		break;
6321 	default:
6322 		break;
6323 	}
6324 	return NOTIFY_OK;
6325 }
6326 #endif
6327 
6328 static int raid5_alloc_percpu(struct r5conf *conf)
6329 {
6330 	unsigned long cpu;
6331 	int err = 0;
6332 
6333 	conf->percpu = alloc_percpu(struct raid5_percpu);
6334 	if (!conf->percpu)
6335 		return -ENOMEM;
6336 
6337 #ifdef CONFIG_HOTPLUG_CPU
6338 	conf->cpu_notify.notifier_call = raid456_cpu_notify;
6339 	conf->cpu_notify.priority = 0;
6340 	err = register_cpu_notifier(&conf->cpu_notify);
6341 	if (err)
6342 		return err;
6343 #endif
6344 
6345 	get_online_cpus();
6346 	for_each_present_cpu(cpu) {
6347 		err = alloc_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6348 		if (err) {
6349 			pr_err("%s: failed memory allocation for cpu%ld\n",
6350 			       __func__, cpu);
6351 			break;
6352 		}
6353 	}
6354 	put_online_cpus();
6355 
6356 	return err;
6357 }
6358 
6359 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6360 				      struct shrink_control *sc)
6361 {
6362 	struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6363 	int ret = 0;
6364 	while (ret < sc->nr_to_scan) {
6365 		if (drop_one_stripe(conf) == 0)
6366 			return SHRINK_STOP;
6367 		ret++;
6368 	}
6369 	return ret;
6370 }
6371 
6372 static unsigned long raid5_cache_count(struct shrinker *shrink,
6373 				       struct shrink_control *sc)
6374 {
6375 	struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6376 
6377 	if (conf->max_nr_stripes < conf->min_nr_stripes)
6378 		/* unlikely, but not impossible */
6379 		return 0;
6380 	return conf->max_nr_stripes - conf->min_nr_stripes;
6381 }
6382 
6383 static struct r5conf *setup_conf(struct mddev *mddev)
6384 {
6385 	struct r5conf *conf;
6386 	int raid_disk, memory, max_disks;
6387 	struct md_rdev *rdev;
6388 	struct disk_info *disk;
6389 	char pers_name[6];
6390 	int i;
6391 	int group_cnt, worker_cnt_per_group;
6392 	struct r5worker_group *new_group;
6393 
6394 	if (mddev->new_level != 5
6395 	    && mddev->new_level != 4
6396 	    && mddev->new_level != 6) {
6397 		printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6398 		       mdname(mddev), mddev->new_level);
6399 		return ERR_PTR(-EIO);
6400 	}
6401 	if ((mddev->new_level == 5
6402 	     && !algorithm_valid_raid5(mddev->new_layout)) ||
6403 	    (mddev->new_level == 6
6404 	     && !algorithm_valid_raid6(mddev->new_layout))) {
6405 		printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
6406 		       mdname(mddev), mddev->new_layout);
6407 		return ERR_PTR(-EIO);
6408 	}
6409 	if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6410 		printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6411 		       mdname(mddev), mddev->raid_disks);
6412 		return ERR_PTR(-EINVAL);
6413 	}
6414 
6415 	if (!mddev->new_chunk_sectors ||
6416 	    (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6417 	    !is_power_of_2(mddev->new_chunk_sectors)) {
6418 		printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
6419 		       mdname(mddev), mddev->new_chunk_sectors << 9);
6420 		return ERR_PTR(-EINVAL);
6421 	}
6422 
6423 	conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6424 	if (conf == NULL)
6425 		goto abort;
6426 	/* Don't enable multi-threading by default*/
6427 	if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6428 				 &new_group)) {
6429 		conf->group_cnt = group_cnt;
6430 		conf->worker_cnt_per_group = worker_cnt_per_group;
6431 		conf->worker_groups = new_group;
6432 	} else
6433 		goto abort;
6434 	spin_lock_init(&conf->device_lock);
6435 	seqcount_init(&conf->gen_lock);
6436 	init_waitqueue_head(&conf->wait_for_stripe);
6437 	init_waitqueue_head(&conf->wait_for_overlap);
6438 	INIT_LIST_HEAD(&conf->handle_list);
6439 	INIT_LIST_HEAD(&conf->hold_list);
6440 	INIT_LIST_HEAD(&conf->delayed_list);
6441 	INIT_LIST_HEAD(&conf->bitmap_list);
6442 	init_llist_head(&conf->released_stripes);
6443 	atomic_set(&conf->active_stripes, 0);
6444 	atomic_set(&conf->preread_active_stripes, 0);
6445 	atomic_set(&conf->active_aligned_reads, 0);
6446 	conf->bypass_threshold = BYPASS_THRESHOLD;
6447 	conf->recovery_disabled = mddev->recovery_disabled - 1;
6448 
6449 	conf->raid_disks = mddev->raid_disks;
6450 	if (mddev->reshape_position == MaxSector)
6451 		conf->previous_raid_disks = mddev->raid_disks;
6452 	else
6453 		conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6454 	max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6455 
6456 	conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
6457 			      GFP_KERNEL);
6458 	if (!conf->disks)
6459 		goto abort;
6460 
6461 	conf->mddev = mddev;
6462 
6463 	if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6464 		goto abort;
6465 
6466 	/* We init hash_locks[0] separately to that it can be used
6467 	 * as the reference lock in the spin_lock_nest_lock() call
6468 	 * in lock_all_device_hash_locks_irq in order to convince
6469 	 * lockdep that we know what we are doing.
6470 	 */
6471 	spin_lock_init(conf->hash_locks);
6472 	for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
6473 		spin_lock_init(conf->hash_locks + i);
6474 
6475 	for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6476 		INIT_LIST_HEAD(conf->inactive_list + i);
6477 
6478 	for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6479 		INIT_LIST_HEAD(conf->temp_inactive_list + i);
6480 
6481 	conf->level = mddev->new_level;
6482 	conf->chunk_sectors = mddev->new_chunk_sectors;
6483 	if (raid5_alloc_percpu(conf) != 0)
6484 		goto abort;
6485 
6486 	pr_debug("raid456: run(%s) called.\n", mdname(mddev));
6487 
6488 	rdev_for_each(rdev, mddev) {
6489 		raid_disk = rdev->raid_disk;
6490 		if (raid_disk >= max_disks
6491 		    || raid_disk < 0)
6492 			continue;
6493 		disk = conf->disks + raid_disk;
6494 
6495 		if (test_bit(Replacement, &rdev->flags)) {
6496 			if (disk->replacement)
6497 				goto abort;
6498 			disk->replacement = rdev;
6499 		} else {
6500 			if (disk->rdev)
6501 				goto abort;
6502 			disk->rdev = rdev;
6503 		}
6504 
6505 		if (test_bit(In_sync, &rdev->flags)) {
6506 			char b[BDEVNAME_SIZE];
6507 			printk(KERN_INFO "md/raid:%s: device %s operational as raid"
6508 			       " disk %d\n",
6509 			       mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
6510 		} else if (rdev->saved_raid_disk != raid_disk)
6511 			/* Cannot rely on bitmap to complete recovery */
6512 			conf->fullsync = 1;
6513 	}
6514 
6515 	conf->level = mddev->new_level;
6516 	if (conf->level == 6) {
6517 		conf->max_degraded = 2;
6518 		if (raid6_call.xor_syndrome)
6519 			conf->rmw_level = PARITY_ENABLE_RMW;
6520 		else
6521 			conf->rmw_level = PARITY_DISABLE_RMW;
6522 	} else {
6523 		conf->max_degraded = 1;
6524 		conf->rmw_level = PARITY_ENABLE_RMW;
6525 	}
6526 	conf->algorithm = mddev->new_layout;
6527 	conf->reshape_progress = mddev->reshape_position;
6528 	if (conf->reshape_progress != MaxSector) {
6529 		conf->prev_chunk_sectors = mddev->chunk_sectors;
6530 		conf->prev_algo = mddev->layout;
6531 	}
6532 
6533 	conf->min_nr_stripes = NR_STRIPES;
6534 	memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
6535 		 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
6536 	atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
6537 	if (grow_stripes(conf, conf->min_nr_stripes)) {
6538 		printk(KERN_ERR
6539 		       "md/raid:%s: couldn't allocate %dkB for buffers\n",
6540 		       mdname(mddev), memory);
6541 		goto abort;
6542 	} else
6543 		printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
6544 		       mdname(mddev), memory);
6545 	/*
6546 	 * Losing a stripe head costs more than the time to refill it,
6547 	 * it reduces the queue depth and so can hurt throughput.
6548 	 * So set it rather large, scaled by number of devices.
6549 	 */
6550 	conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
6551 	conf->shrinker.scan_objects = raid5_cache_scan;
6552 	conf->shrinker.count_objects = raid5_cache_count;
6553 	conf->shrinker.batch = 128;
6554 	conf->shrinker.flags = 0;
6555 	register_shrinker(&conf->shrinker);
6556 
6557 	sprintf(pers_name, "raid%d", mddev->new_level);
6558 	conf->thread = md_register_thread(raid5d, mddev, pers_name);
6559 	if (!conf->thread) {
6560 		printk(KERN_ERR
6561 		       "md/raid:%s: couldn't allocate thread.\n",
6562 		       mdname(mddev));
6563 		goto abort;
6564 	}
6565 
6566 	return conf;
6567 
6568  abort:
6569 	if (conf) {
6570 		free_conf(conf);
6571 		return ERR_PTR(-EIO);
6572 	} else
6573 		return ERR_PTR(-ENOMEM);
6574 }
6575 
6576 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
6577 {
6578 	switch (algo) {
6579 	case ALGORITHM_PARITY_0:
6580 		if (raid_disk < max_degraded)
6581 			return 1;
6582 		break;
6583 	case ALGORITHM_PARITY_N:
6584 		if (raid_disk >= raid_disks - max_degraded)
6585 			return 1;
6586 		break;
6587 	case ALGORITHM_PARITY_0_6:
6588 		if (raid_disk == 0 ||
6589 		    raid_disk == raid_disks - 1)
6590 			return 1;
6591 		break;
6592 	case ALGORITHM_LEFT_ASYMMETRIC_6:
6593 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
6594 	case ALGORITHM_LEFT_SYMMETRIC_6:
6595 	case ALGORITHM_RIGHT_SYMMETRIC_6:
6596 		if (raid_disk == raid_disks - 1)
6597 			return 1;
6598 	}
6599 	return 0;
6600 }
6601 
6602 static int run(struct mddev *mddev)
6603 {
6604 	struct r5conf *conf;
6605 	int working_disks = 0;
6606 	int dirty_parity_disks = 0;
6607 	struct md_rdev *rdev;
6608 	sector_t reshape_offset = 0;
6609 	int i;
6610 	long long min_offset_diff = 0;
6611 	int first = 1;
6612 
6613 	if (mddev->recovery_cp != MaxSector)
6614 		printk(KERN_NOTICE "md/raid:%s: not clean"
6615 		       " -- starting background reconstruction\n",
6616 		       mdname(mddev));
6617 
6618 	rdev_for_each(rdev, mddev) {
6619 		long long diff;
6620 		if (rdev->raid_disk < 0)
6621 			continue;
6622 		diff = (rdev->new_data_offset - rdev->data_offset);
6623 		if (first) {
6624 			min_offset_diff = diff;
6625 			first = 0;
6626 		} else if (mddev->reshape_backwards &&
6627 			 diff < min_offset_diff)
6628 			min_offset_diff = diff;
6629 		else if (!mddev->reshape_backwards &&
6630 			 diff > min_offset_diff)
6631 			min_offset_diff = diff;
6632 	}
6633 
6634 	if (mddev->reshape_position != MaxSector) {
6635 		/* Check that we can continue the reshape.
6636 		 * Difficulties arise if the stripe we would write to
6637 		 * next is at or after the stripe we would read from next.
6638 		 * For a reshape that changes the number of devices, this
6639 		 * is only possible for a very short time, and mdadm makes
6640 		 * sure that time appears to have past before assembling
6641 		 * the array.  So we fail if that time hasn't passed.
6642 		 * For a reshape that keeps the number of devices the same
6643 		 * mdadm must be monitoring the reshape can keeping the
6644 		 * critical areas read-only and backed up.  It will start
6645 		 * the array in read-only mode, so we check for that.
6646 		 */
6647 		sector_t here_new, here_old;
6648 		int old_disks;
6649 		int max_degraded = (mddev->level == 6 ? 2 : 1);
6650 
6651 		if (mddev->new_level != mddev->level) {
6652 			printk(KERN_ERR "md/raid:%s: unsupported reshape "
6653 			       "required - aborting.\n",
6654 			       mdname(mddev));
6655 			return -EINVAL;
6656 		}
6657 		old_disks = mddev->raid_disks - mddev->delta_disks;
6658 		/* reshape_position must be on a new-stripe boundary, and one
6659 		 * further up in new geometry must map after here in old
6660 		 * geometry.
6661 		 */
6662 		here_new = mddev->reshape_position;
6663 		if (sector_div(here_new, mddev->new_chunk_sectors *
6664 			       (mddev->raid_disks - max_degraded))) {
6665 			printk(KERN_ERR "md/raid:%s: reshape_position not "
6666 			       "on a stripe boundary\n", mdname(mddev));
6667 			return -EINVAL;
6668 		}
6669 		reshape_offset = here_new * mddev->new_chunk_sectors;
6670 		/* here_new is the stripe we will write to */
6671 		here_old = mddev->reshape_position;
6672 		sector_div(here_old, mddev->chunk_sectors *
6673 			   (old_disks-max_degraded));
6674 		/* here_old is the first stripe that we might need to read
6675 		 * from */
6676 		if (mddev->delta_disks == 0) {
6677 			if ((here_new * mddev->new_chunk_sectors !=
6678 			     here_old * mddev->chunk_sectors)) {
6679 				printk(KERN_ERR "md/raid:%s: reshape position is"
6680 				       " confused - aborting\n", mdname(mddev));
6681 				return -EINVAL;
6682 			}
6683 			/* We cannot be sure it is safe to start an in-place
6684 			 * reshape.  It is only safe if user-space is monitoring
6685 			 * and taking constant backups.
6686 			 * mdadm always starts a situation like this in
6687 			 * readonly mode so it can take control before
6688 			 * allowing any writes.  So just check for that.
6689 			 */
6690 			if (abs(min_offset_diff) >= mddev->chunk_sectors &&
6691 			    abs(min_offset_diff) >= mddev->new_chunk_sectors)
6692 				/* not really in-place - so OK */;
6693 			else if (mddev->ro == 0) {
6694 				printk(KERN_ERR "md/raid:%s: in-place reshape "
6695 				       "must be started in read-only mode "
6696 				       "- aborting\n",
6697 				       mdname(mddev));
6698 				return -EINVAL;
6699 			}
6700 		} else if (mddev->reshape_backwards
6701 		    ? (here_new * mddev->new_chunk_sectors + min_offset_diff <=
6702 		       here_old * mddev->chunk_sectors)
6703 		    : (here_new * mddev->new_chunk_sectors >=
6704 		       here_old * mddev->chunk_sectors + (-min_offset_diff))) {
6705 			/* Reading from the same stripe as writing to - bad */
6706 			printk(KERN_ERR "md/raid:%s: reshape_position too early for "
6707 			       "auto-recovery - aborting.\n",
6708 			       mdname(mddev));
6709 			return -EINVAL;
6710 		}
6711 		printk(KERN_INFO "md/raid:%s: reshape will continue\n",
6712 		       mdname(mddev));
6713 		/* OK, we should be able to continue; */
6714 	} else {
6715 		BUG_ON(mddev->level != mddev->new_level);
6716 		BUG_ON(mddev->layout != mddev->new_layout);
6717 		BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
6718 		BUG_ON(mddev->delta_disks != 0);
6719 	}
6720 
6721 	if (mddev->private == NULL)
6722 		conf = setup_conf(mddev);
6723 	else
6724 		conf = mddev->private;
6725 
6726 	if (IS_ERR(conf))
6727 		return PTR_ERR(conf);
6728 
6729 	conf->min_offset_diff = min_offset_diff;
6730 	mddev->thread = conf->thread;
6731 	conf->thread = NULL;
6732 	mddev->private = conf;
6733 
6734 	for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
6735 	     i++) {
6736 		rdev = conf->disks[i].rdev;
6737 		if (!rdev && conf->disks[i].replacement) {
6738 			/* The replacement is all we have yet */
6739 			rdev = conf->disks[i].replacement;
6740 			conf->disks[i].replacement = NULL;
6741 			clear_bit(Replacement, &rdev->flags);
6742 			conf->disks[i].rdev = rdev;
6743 		}
6744 		if (!rdev)
6745 			continue;
6746 		if (conf->disks[i].replacement &&
6747 		    conf->reshape_progress != MaxSector) {
6748 			/* replacements and reshape simply do not mix. */
6749 			printk(KERN_ERR "md: cannot handle concurrent "
6750 			       "replacement and reshape.\n");
6751 			goto abort;
6752 		}
6753 		if (test_bit(In_sync, &rdev->flags)) {
6754 			working_disks++;
6755 			continue;
6756 		}
6757 		/* This disc is not fully in-sync.  However if it
6758 		 * just stored parity (beyond the recovery_offset),
6759 		 * when we don't need to be concerned about the
6760 		 * array being dirty.
6761 		 * When reshape goes 'backwards', we never have
6762 		 * partially completed devices, so we only need
6763 		 * to worry about reshape going forwards.
6764 		 */
6765 		/* Hack because v0.91 doesn't store recovery_offset properly. */
6766 		if (mddev->major_version == 0 &&
6767 		    mddev->minor_version > 90)
6768 			rdev->recovery_offset = reshape_offset;
6769 
6770 		if (rdev->recovery_offset < reshape_offset) {
6771 			/* We need to check old and new layout */
6772 			if (!only_parity(rdev->raid_disk,
6773 					 conf->algorithm,
6774 					 conf->raid_disks,
6775 					 conf->max_degraded))
6776 				continue;
6777 		}
6778 		if (!only_parity(rdev->raid_disk,
6779 				 conf->prev_algo,
6780 				 conf->previous_raid_disks,
6781 				 conf->max_degraded))
6782 			continue;
6783 		dirty_parity_disks++;
6784 	}
6785 
6786 	/*
6787 	 * 0 for a fully functional array, 1 or 2 for a degraded array.
6788 	 */
6789 	mddev->degraded = calc_degraded(conf);
6790 
6791 	if (has_failed(conf)) {
6792 		printk(KERN_ERR "md/raid:%s: not enough operational devices"
6793 			" (%d/%d failed)\n",
6794 			mdname(mddev), mddev->degraded, conf->raid_disks);
6795 		goto abort;
6796 	}
6797 
6798 	/* device size must be a multiple of chunk size */
6799 	mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
6800 	mddev->resync_max_sectors = mddev->dev_sectors;
6801 
6802 	if (mddev->degraded > dirty_parity_disks &&
6803 	    mddev->recovery_cp != MaxSector) {
6804 		if (mddev->ok_start_degraded)
6805 			printk(KERN_WARNING
6806 			       "md/raid:%s: starting dirty degraded array"
6807 			       " - data corruption possible.\n",
6808 			       mdname(mddev));
6809 		else {
6810 			printk(KERN_ERR
6811 			       "md/raid:%s: cannot start dirty degraded array.\n",
6812 			       mdname(mddev));
6813 			goto abort;
6814 		}
6815 	}
6816 
6817 	if (mddev->degraded == 0)
6818 		printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
6819 		       " devices, algorithm %d\n", mdname(mddev), conf->level,
6820 		       mddev->raid_disks-mddev->degraded, mddev->raid_disks,
6821 		       mddev->new_layout);
6822 	else
6823 		printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
6824 		       " out of %d devices, algorithm %d\n",
6825 		       mdname(mddev), conf->level,
6826 		       mddev->raid_disks - mddev->degraded,
6827 		       mddev->raid_disks, mddev->new_layout);
6828 
6829 	print_raid5_conf(conf);
6830 
6831 	if (conf->reshape_progress != MaxSector) {
6832 		conf->reshape_safe = conf->reshape_progress;
6833 		atomic_set(&conf->reshape_stripes, 0);
6834 		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
6835 		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
6836 		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
6837 		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
6838 		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
6839 							"reshape");
6840 	}
6841 
6842 	/* Ok, everything is just fine now */
6843 	if (mddev->to_remove == &raid5_attrs_group)
6844 		mddev->to_remove = NULL;
6845 	else if (mddev->kobj.sd &&
6846 	    sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
6847 		printk(KERN_WARNING
6848 		       "raid5: failed to create sysfs attributes for %s\n",
6849 		       mdname(mddev));
6850 	md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
6851 
6852 	if (mddev->queue) {
6853 		int chunk_size;
6854 		bool discard_supported = true;
6855 		/* read-ahead size must cover two whole stripes, which
6856 		 * is 2 * (datadisks) * chunksize where 'n' is the
6857 		 * number of raid devices
6858 		 */
6859 		int data_disks = conf->previous_raid_disks - conf->max_degraded;
6860 		int stripe = data_disks *
6861 			((mddev->chunk_sectors << 9) / PAGE_SIZE);
6862 		if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
6863 			mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
6864 
6865 		chunk_size = mddev->chunk_sectors << 9;
6866 		blk_queue_io_min(mddev->queue, chunk_size);
6867 		blk_queue_io_opt(mddev->queue, chunk_size *
6868 				 (conf->raid_disks - conf->max_degraded));
6869 		mddev->queue->limits.raid_partial_stripes_expensive = 1;
6870 		/*
6871 		 * We can only discard a whole stripe. It doesn't make sense to
6872 		 * discard data disk but write parity disk
6873 		 */
6874 		stripe = stripe * PAGE_SIZE;
6875 		/* Round up to power of 2, as discard handling
6876 		 * currently assumes that */
6877 		while ((stripe-1) & stripe)
6878 			stripe = (stripe | (stripe-1)) + 1;
6879 		mddev->queue->limits.discard_alignment = stripe;
6880 		mddev->queue->limits.discard_granularity = stripe;
6881 		/*
6882 		 * unaligned part of discard request will be ignored, so can't
6883 		 * guarantee discard_zeroes_data
6884 		 */
6885 		mddev->queue->limits.discard_zeroes_data = 0;
6886 
6887 		blk_queue_max_write_same_sectors(mddev->queue, 0);
6888 
6889 		rdev_for_each(rdev, mddev) {
6890 			disk_stack_limits(mddev->gendisk, rdev->bdev,
6891 					  rdev->data_offset << 9);
6892 			disk_stack_limits(mddev->gendisk, rdev->bdev,
6893 					  rdev->new_data_offset << 9);
6894 			/*
6895 			 * discard_zeroes_data is required, otherwise data
6896 			 * could be lost. Consider a scenario: discard a stripe
6897 			 * (the stripe could be inconsistent if
6898 			 * discard_zeroes_data is 0); write one disk of the
6899 			 * stripe (the stripe could be inconsistent again
6900 			 * depending on which disks are used to calculate
6901 			 * parity); the disk is broken; The stripe data of this
6902 			 * disk is lost.
6903 			 */
6904 			if (!blk_queue_discard(bdev_get_queue(rdev->bdev)) ||
6905 			    !bdev_get_queue(rdev->bdev)->
6906 						limits.discard_zeroes_data)
6907 				discard_supported = false;
6908 			/* Unfortunately, discard_zeroes_data is not currently
6909 			 * a guarantee - just a hint.  So we only allow DISCARD
6910 			 * if the sysadmin has confirmed that only safe devices
6911 			 * are in use by setting a module parameter.
6912 			 */
6913 			if (!devices_handle_discard_safely) {
6914 				if (discard_supported) {
6915 					pr_info("md/raid456: discard support disabled due to uncertainty.\n");
6916 					pr_info("Set raid456.devices_handle_discard_safely=Y to override.\n");
6917 				}
6918 				discard_supported = false;
6919 			}
6920 		}
6921 
6922 		if (discard_supported &&
6923 		   mddev->queue->limits.max_discard_sectors >= stripe &&
6924 		   mddev->queue->limits.discard_granularity >= stripe)
6925 			queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
6926 						mddev->queue);
6927 		else
6928 			queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
6929 						mddev->queue);
6930 	}
6931 
6932 	return 0;
6933 abort:
6934 	md_unregister_thread(&mddev->thread);
6935 	print_raid5_conf(conf);
6936 	free_conf(conf);
6937 	mddev->private = NULL;
6938 	printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
6939 	return -EIO;
6940 }
6941 
6942 static void raid5_free(struct mddev *mddev, void *priv)
6943 {
6944 	struct r5conf *conf = priv;
6945 
6946 	free_conf(conf);
6947 	mddev->to_remove = &raid5_attrs_group;
6948 }
6949 
6950 static void status(struct seq_file *seq, struct mddev *mddev)
6951 {
6952 	struct r5conf *conf = mddev->private;
6953 	int i;
6954 
6955 	seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
6956 		mddev->chunk_sectors / 2, mddev->layout);
6957 	seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
6958 	for (i = 0; i < conf->raid_disks; i++)
6959 		seq_printf (seq, "%s",
6960 			       conf->disks[i].rdev &&
6961 			       test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
6962 	seq_printf (seq, "]");
6963 }
6964 
6965 static void print_raid5_conf (struct r5conf *conf)
6966 {
6967 	int i;
6968 	struct disk_info *tmp;
6969 
6970 	printk(KERN_DEBUG "RAID conf printout:\n");
6971 	if (!conf) {
6972 		printk("(conf==NULL)\n");
6973 		return;
6974 	}
6975 	printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
6976 	       conf->raid_disks,
6977 	       conf->raid_disks - conf->mddev->degraded);
6978 
6979 	for (i = 0; i < conf->raid_disks; i++) {
6980 		char b[BDEVNAME_SIZE];
6981 		tmp = conf->disks + i;
6982 		if (tmp->rdev)
6983 			printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
6984 			       i, !test_bit(Faulty, &tmp->rdev->flags),
6985 			       bdevname(tmp->rdev->bdev, b));
6986 	}
6987 }
6988 
6989 static int raid5_spare_active(struct mddev *mddev)
6990 {
6991 	int i;
6992 	struct r5conf *conf = mddev->private;
6993 	struct disk_info *tmp;
6994 	int count = 0;
6995 	unsigned long flags;
6996 
6997 	for (i = 0; i < conf->raid_disks; i++) {
6998 		tmp = conf->disks + i;
6999 		if (tmp->replacement
7000 		    && tmp->replacement->recovery_offset == MaxSector
7001 		    && !test_bit(Faulty, &tmp->replacement->flags)
7002 		    && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7003 			/* Replacement has just become active. */
7004 			if (!tmp->rdev
7005 			    || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7006 				count++;
7007 			if (tmp->rdev) {
7008 				/* Replaced device not technically faulty,
7009 				 * but we need to be sure it gets removed
7010 				 * and never re-added.
7011 				 */
7012 				set_bit(Faulty, &tmp->rdev->flags);
7013 				sysfs_notify_dirent_safe(
7014 					tmp->rdev->sysfs_state);
7015 			}
7016 			sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7017 		} else if (tmp->rdev
7018 		    && tmp->rdev->recovery_offset == MaxSector
7019 		    && !test_bit(Faulty, &tmp->rdev->flags)
7020 		    && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7021 			count++;
7022 			sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7023 		}
7024 	}
7025 	spin_lock_irqsave(&conf->device_lock, flags);
7026 	mddev->degraded = calc_degraded(conf);
7027 	spin_unlock_irqrestore(&conf->device_lock, flags);
7028 	print_raid5_conf(conf);
7029 	return count;
7030 }
7031 
7032 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7033 {
7034 	struct r5conf *conf = mddev->private;
7035 	int err = 0;
7036 	int number = rdev->raid_disk;
7037 	struct md_rdev **rdevp;
7038 	struct disk_info *p = conf->disks + number;
7039 
7040 	print_raid5_conf(conf);
7041 	if (rdev == p->rdev)
7042 		rdevp = &p->rdev;
7043 	else if (rdev == p->replacement)
7044 		rdevp = &p->replacement;
7045 	else
7046 		return 0;
7047 
7048 	if (number >= conf->raid_disks &&
7049 	    conf->reshape_progress == MaxSector)
7050 		clear_bit(In_sync, &rdev->flags);
7051 
7052 	if (test_bit(In_sync, &rdev->flags) ||
7053 	    atomic_read(&rdev->nr_pending)) {
7054 		err = -EBUSY;
7055 		goto abort;
7056 	}
7057 	/* Only remove non-faulty devices if recovery
7058 	 * isn't possible.
7059 	 */
7060 	if (!test_bit(Faulty, &rdev->flags) &&
7061 	    mddev->recovery_disabled != conf->recovery_disabled &&
7062 	    !has_failed(conf) &&
7063 	    (!p->replacement || p->replacement == rdev) &&
7064 	    number < conf->raid_disks) {
7065 		err = -EBUSY;
7066 		goto abort;
7067 	}
7068 	*rdevp = NULL;
7069 	synchronize_rcu();
7070 	if (atomic_read(&rdev->nr_pending)) {
7071 		/* lost the race, try later */
7072 		err = -EBUSY;
7073 		*rdevp = rdev;
7074 	} else if (p->replacement) {
7075 		/* We must have just cleared 'rdev' */
7076 		p->rdev = p->replacement;
7077 		clear_bit(Replacement, &p->replacement->flags);
7078 		smp_mb(); /* Make sure other CPUs may see both as identical
7079 			   * but will never see neither - if they are careful
7080 			   */
7081 		p->replacement = NULL;
7082 		clear_bit(WantReplacement, &rdev->flags);
7083 	} else
7084 		/* We might have just removed the Replacement as faulty-
7085 		 * clear the bit just in case
7086 		 */
7087 		clear_bit(WantReplacement, &rdev->flags);
7088 abort:
7089 
7090 	print_raid5_conf(conf);
7091 	return err;
7092 }
7093 
7094 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7095 {
7096 	struct r5conf *conf = mddev->private;
7097 	int err = -EEXIST;
7098 	int disk;
7099 	struct disk_info *p;
7100 	int first = 0;
7101 	int last = conf->raid_disks - 1;
7102 
7103 	if (mddev->recovery_disabled == conf->recovery_disabled)
7104 		return -EBUSY;
7105 
7106 	if (rdev->saved_raid_disk < 0 && has_failed(conf))
7107 		/* no point adding a device */
7108 		return -EINVAL;
7109 
7110 	if (rdev->raid_disk >= 0)
7111 		first = last = rdev->raid_disk;
7112 
7113 	/*
7114 	 * find the disk ... but prefer rdev->saved_raid_disk
7115 	 * if possible.
7116 	 */
7117 	if (rdev->saved_raid_disk >= 0 &&
7118 	    rdev->saved_raid_disk >= first &&
7119 	    conf->disks[rdev->saved_raid_disk].rdev == NULL)
7120 		first = rdev->saved_raid_disk;
7121 
7122 	for (disk = first; disk <= last; disk++) {
7123 		p = conf->disks + disk;
7124 		if (p->rdev == NULL) {
7125 			clear_bit(In_sync, &rdev->flags);
7126 			rdev->raid_disk = disk;
7127 			err = 0;
7128 			if (rdev->saved_raid_disk != disk)
7129 				conf->fullsync = 1;
7130 			rcu_assign_pointer(p->rdev, rdev);
7131 			goto out;
7132 		}
7133 	}
7134 	for (disk = first; disk <= last; disk++) {
7135 		p = conf->disks + disk;
7136 		if (test_bit(WantReplacement, &p->rdev->flags) &&
7137 		    p->replacement == NULL) {
7138 			clear_bit(In_sync, &rdev->flags);
7139 			set_bit(Replacement, &rdev->flags);
7140 			rdev->raid_disk = disk;
7141 			err = 0;
7142 			conf->fullsync = 1;
7143 			rcu_assign_pointer(p->replacement, rdev);
7144 			break;
7145 		}
7146 	}
7147 out:
7148 	print_raid5_conf(conf);
7149 	return err;
7150 }
7151 
7152 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7153 {
7154 	/* no resync is happening, and there is enough space
7155 	 * on all devices, so we can resize.
7156 	 * We need to make sure resync covers any new space.
7157 	 * If the array is shrinking we should possibly wait until
7158 	 * any io in the removed space completes, but it hardly seems
7159 	 * worth it.
7160 	 */
7161 	sector_t newsize;
7162 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
7163 	newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7164 	if (mddev->external_size &&
7165 	    mddev->array_sectors > newsize)
7166 		return -EINVAL;
7167 	if (mddev->bitmap) {
7168 		int ret = bitmap_resize(mddev->bitmap, sectors, 0, 0);
7169 		if (ret)
7170 			return ret;
7171 	}
7172 	md_set_array_sectors(mddev, newsize);
7173 	set_capacity(mddev->gendisk, mddev->array_sectors);
7174 	revalidate_disk(mddev->gendisk);
7175 	if (sectors > mddev->dev_sectors &&
7176 	    mddev->recovery_cp > mddev->dev_sectors) {
7177 		mddev->recovery_cp = mddev->dev_sectors;
7178 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7179 	}
7180 	mddev->dev_sectors = sectors;
7181 	mddev->resync_max_sectors = sectors;
7182 	return 0;
7183 }
7184 
7185 static int check_stripe_cache(struct mddev *mddev)
7186 {
7187 	/* Can only proceed if there are plenty of stripe_heads.
7188 	 * We need a minimum of one full stripe,, and for sensible progress
7189 	 * it is best to have about 4 times that.
7190 	 * If we require 4 times, then the default 256 4K stripe_heads will
7191 	 * allow for chunk sizes up to 256K, which is probably OK.
7192 	 * If the chunk size is greater, user-space should request more
7193 	 * stripe_heads first.
7194 	 */
7195 	struct r5conf *conf = mddev->private;
7196 	if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7197 	    > conf->min_nr_stripes ||
7198 	    ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7199 	    > conf->min_nr_stripes) {
7200 		printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7201 		       mdname(mddev),
7202 		       ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7203 			/ STRIPE_SIZE)*4);
7204 		return 0;
7205 	}
7206 	return 1;
7207 }
7208 
7209 static int check_reshape(struct mddev *mddev)
7210 {
7211 	struct r5conf *conf = mddev->private;
7212 
7213 	if (mddev->delta_disks == 0 &&
7214 	    mddev->new_layout == mddev->layout &&
7215 	    mddev->new_chunk_sectors == mddev->chunk_sectors)
7216 		return 0; /* nothing to do */
7217 	if (has_failed(conf))
7218 		return -EINVAL;
7219 	if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7220 		/* We might be able to shrink, but the devices must
7221 		 * be made bigger first.
7222 		 * For raid6, 4 is the minimum size.
7223 		 * Otherwise 2 is the minimum
7224 		 */
7225 		int min = 2;
7226 		if (mddev->level == 6)
7227 			min = 4;
7228 		if (mddev->raid_disks + mddev->delta_disks < min)
7229 			return -EINVAL;
7230 	}
7231 
7232 	if (!check_stripe_cache(mddev))
7233 		return -ENOSPC;
7234 
7235 	if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7236 	    mddev->delta_disks > 0)
7237 		if (resize_chunks(conf,
7238 				  conf->previous_raid_disks
7239 				  + max(0, mddev->delta_disks),
7240 				  max(mddev->new_chunk_sectors,
7241 				      mddev->chunk_sectors)
7242 			    ) < 0)
7243 			return -ENOMEM;
7244 	return resize_stripes(conf, (conf->previous_raid_disks
7245 				     + mddev->delta_disks));
7246 }
7247 
7248 static int raid5_start_reshape(struct mddev *mddev)
7249 {
7250 	struct r5conf *conf = mddev->private;
7251 	struct md_rdev *rdev;
7252 	int spares = 0;
7253 	unsigned long flags;
7254 
7255 	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7256 		return -EBUSY;
7257 
7258 	if (!check_stripe_cache(mddev))
7259 		return -ENOSPC;
7260 
7261 	if (has_failed(conf))
7262 		return -EINVAL;
7263 
7264 	rdev_for_each(rdev, mddev) {
7265 		if (!test_bit(In_sync, &rdev->flags)
7266 		    && !test_bit(Faulty, &rdev->flags))
7267 			spares++;
7268 	}
7269 
7270 	if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7271 		/* Not enough devices even to make a degraded array
7272 		 * of that size
7273 		 */
7274 		return -EINVAL;
7275 
7276 	/* Refuse to reduce size of the array.  Any reductions in
7277 	 * array size must be through explicit setting of array_size
7278 	 * attribute.
7279 	 */
7280 	if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7281 	    < mddev->array_sectors) {
7282 		printk(KERN_ERR "md/raid:%s: array size must be reduced "
7283 		       "before number of disks\n", mdname(mddev));
7284 		return -EINVAL;
7285 	}
7286 
7287 	atomic_set(&conf->reshape_stripes, 0);
7288 	spin_lock_irq(&conf->device_lock);
7289 	write_seqcount_begin(&conf->gen_lock);
7290 	conf->previous_raid_disks = conf->raid_disks;
7291 	conf->raid_disks += mddev->delta_disks;
7292 	conf->prev_chunk_sectors = conf->chunk_sectors;
7293 	conf->chunk_sectors = mddev->new_chunk_sectors;
7294 	conf->prev_algo = conf->algorithm;
7295 	conf->algorithm = mddev->new_layout;
7296 	conf->generation++;
7297 	/* Code that selects data_offset needs to see the generation update
7298 	 * if reshape_progress has been set - so a memory barrier needed.
7299 	 */
7300 	smp_mb();
7301 	if (mddev->reshape_backwards)
7302 		conf->reshape_progress = raid5_size(mddev, 0, 0);
7303 	else
7304 		conf->reshape_progress = 0;
7305 	conf->reshape_safe = conf->reshape_progress;
7306 	write_seqcount_end(&conf->gen_lock);
7307 	spin_unlock_irq(&conf->device_lock);
7308 
7309 	/* Now make sure any requests that proceeded on the assumption
7310 	 * the reshape wasn't running - like Discard or Read - have
7311 	 * completed.
7312 	 */
7313 	mddev_suspend(mddev);
7314 	mddev_resume(mddev);
7315 
7316 	/* Add some new drives, as many as will fit.
7317 	 * We know there are enough to make the newly sized array work.
7318 	 * Don't add devices if we are reducing the number of
7319 	 * devices in the array.  This is because it is not possible
7320 	 * to correctly record the "partially reconstructed" state of
7321 	 * such devices during the reshape and confusion could result.
7322 	 */
7323 	if (mddev->delta_disks >= 0) {
7324 		rdev_for_each(rdev, mddev)
7325 			if (rdev->raid_disk < 0 &&
7326 			    !test_bit(Faulty, &rdev->flags)) {
7327 				if (raid5_add_disk(mddev, rdev) == 0) {
7328 					if (rdev->raid_disk
7329 					    >= conf->previous_raid_disks)
7330 						set_bit(In_sync, &rdev->flags);
7331 					else
7332 						rdev->recovery_offset = 0;
7333 
7334 					if (sysfs_link_rdev(mddev, rdev))
7335 						/* Failure here is OK */;
7336 				}
7337 			} else if (rdev->raid_disk >= conf->previous_raid_disks
7338 				   && !test_bit(Faulty, &rdev->flags)) {
7339 				/* This is a spare that was manually added */
7340 				set_bit(In_sync, &rdev->flags);
7341 			}
7342 
7343 		/* When a reshape changes the number of devices,
7344 		 * ->degraded is measured against the larger of the
7345 		 * pre and post number of devices.
7346 		 */
7347 		spin_lock_irqsave(&conf->device_lock, flags);
7348 		mddev->degraded = calc_degraded(conf);
7349 		spin_unlock_irqrestore(&conf->device_lock, flags);
7350 	}
7351 	mddev->raid_disks = conf->raid_disks;
7352 	mddev->reshape_position = conf->reshape_progress;
7353 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
7354 
7355 	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7356 	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7357 	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7358 	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7359 	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7360 						"reshape");
7361 	if (!mddev->sync_thread) {
7362 		mddev->recovery = 0;
7363 		spin_lock_irq(&conf->device_lock);
7364 		write_seqcount_begin(&conf->gen_lock);
7365 		mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7366 		mddev->new_chunk_sectors =
7367 			conf->chunk_sectors = conf->prev_chunk_sectors;
7368 		mddev->new_layout = conf->algorithm = conf->prev_algo;
7369 		rdev_for_each(rdev, mddev)
7370 			rdev->new_data_offset = rdev->data_offset;
7371 		smp_wmb();
7372 		conf->generation --;
7373 		conf->reshape_progress = MaxSector;
7374 		mddev->reshape_position = MaxSector;
7375 		write_seqcount_end(&conf->gen_lock);
7376 		spin_unlock_irq(&conf->device_lock);
7377 		return -EAGAIN;
7378 	}
7379 	conf->reshape_checkpoint = jiffies;
7380 	md_wakeup_thread(mddev->sync_thread);
7381 	md_new_event(mddev);
7382 	return 0;
7383 }
7384 
7385 /* This is called from the reshape thread and should make any
7386  * changes needed in 'conf'
7387  */
7388 static void end_reshape(struct r5conf *conf)
7389 {
7390 
7391 	if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
7392 		struct md_rdev *rdev;
7393 
7394 		spin_lock_irq(&conf->device_lock);
7395 		conf->previous_raid_disks = conf->raid_disks;
7396 		rdev_for_each(rdev, conf->mddev)
7397 			rdev->data_offset = rdev->new_data_offset;
7398 		smp_wmb();
7399 		conf->reshape_progress = MaxSector;
7400 		spin_unlock_irq(&conf->device_lock);
7401 		wake_up(&conf->wait_for_overlap);
7402 
7403 		/* read-ahead size must cover two whole stripes, which is
7404 		 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
7405 		 */
7406 		if (conf->mddev->queue) {
7407 			int data_disks = conf->raid_disks - conf->max_degraded;
7408 			int stripe = data_disks * ((conf->chunk_sectors << 9)
7409 						   / PAGE_SIZE);
7410 			if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
7411 				conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
7412 		}
7413 	}
7414 }
7415 
7416 /* This is called from the raid5d thread with mddev_lock held.
7417  * It makes config changes to the device.
7418  */
7419 static void raid5_finish_reshape(struct mddev *mddev)
7420 {
7421 	struct r5conf *conf = mddev->private;
7422 
7423 	if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
7424 
7425 		if (mddev->delta_disks > 0) {
7426 			md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7427 			set_capacity(mddev->gendisk, mddev->array_sectors);
7428 			revalidate_disk(mddev->gendisk);
7429 		} else {
7430 			int d;
7431 			spin_lock_irq(&conf->device_lock);
7432 			mddev->degraded = calc_degraded(conf);
7433 			spin_unlock_irq(&conf->device_lock);
7434 			for (d = conf->raid_disks ;
7435 			     d < conf->raid_disks - mddev->delta_disks;
7436 			     d++) {
7437 				struct md_rdev *rdev = conf->disks[d].rdev;
7438 				if (rdev)
7439 					clear_bit(In_sync, &rdev->flags);
7440 				rdev = conf->disks[d].replacement;
7441 				if (rdev)
7442 					clear_bit(In_sync, &rdev->flags);
7443 			}
7444 		}
7445 		mddev->layout = conf->algorithm;
7446 		mddev->chunk_sectors = conf->chunk_sectors;
7447 		mddev->reshape_position = MaxSector;
7448 		mddev->delta_disks = 0;
7449 		mddev->reshape_backwards = 0;
7450 	}
7451 }
7452 
7453 static void raid5_quiesce(struct mddev *mddev, int state)
7454 {
7455 	struct r5conf *conf = mddev->private;
7456 
7457 	switch(state) {
7458 	case 2: /* resume for a suspend */
7459 		wake_up(&conf->wait_for_overlap);
7460 		break;
7461 
7462 	case 1: /* stop all writes */
7463 		lock_all_device_hash_locks_irq(conf);
7464 		/* '2' tells resync/reshape to pause so that all
7465 		 * active stripes can drain
7466 		 */
7467 		conf->quiesce = 2;
7468 		wait_event_cmd(conf->wait_for_stripe,
7469 				    atomic_read(&conf->active_stripes) == 0 &&
7470 				    atomic_read(&conf->active_aligned_reads) == 0,
7471 				    unlock_all_device_hash_locks_irq(conf),
7472 				    lock_all_device_hash_locks_irq(conf));
7473 		conf->quiesce = 1;
7474 		unlock_all_device_hash_locks_irq(conf);
7475 		/* allow reshape to continue */
7476 		wake_up(&conf->wait_for_overlap);
7477 		break;
7478 
7479 	case 0: /* re-enable writes */
7480 		lock_all_device_hash_locks_irq(conf);
7481 		conf->quiesce = 0;
7482 		wake_up(&conf->wait_for_stripe);
7483 		wake_up(&conf->wait_for_overlap);
7484 		unlock_all_device_hash_locks_irq(conf);
7485 		break;
7486 	}
7487 }
7488 
7489 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
7490 {
7491 	struct r0conf *raid0_conf = mddev->private;
7492 	sector_t sectors;
7493 
7494 	/* for raid0 takeover only one zone is supported */
7495 	if (raid0_conf->nr_strip_zones > 1) {
7496 		printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
7497 		       mdname(mddev));
7498 		return ERR_PTR(-EINVAL);
7499 	}
7500 
7501 	sectors = raid0_conf->strip_zone[0].zone_end;
7502 	sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
7503 	mddev->dev_sectors = sectors;
7504 	mddev->new_level = level;
7505 	mddev->new_layout = ALGORITHM_PARITY_N;
7506 	mddev->new_chunk_sectors = mddev->chunk_sectors;
7507 	mddev->raid_disks += 1;
7508 	mddev->delta_disks = 1;
7509 	/* make sure it will be not marked as dirty */
7510 	mddev->recovery_cp = MaxSector;
7511 
7512 	return setup_conf(mddev);
7513 }
7514 
7515 static void *raid5_takeover_raid1(struct mddev *mddev)
7516 {
7517 	int chunksect;
7518 
7519 	if (mddev->raid_disks != 2 ||
7520 	    mddev->degraded > 1)
7521 		return ERR_PTR(-EINVAL);
7522 
7523 	/* Should check if there are write-behind devices? */
7524 
7525 	chunksect = 64*2; /* 64K by default */
7526 
7527 	/* The array must be an exact multiple of chunksize */
7528 	while (chunksect && (mddev->array_sectors & (chunksect-1)))
7529 		chunksect >>= 1;
7530 
7531 	if ((chunksect<<9) < STRIPE_SIZE)
7532 		/* array size does not allow a suitable chunk size */
7533 		return ERR_PTR(-EINVAL);
7534 
7535 	mddev->new_level = 5;
7536 	mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
7537 	mddev->new_chunk_sectors = chunksect;
7538 
7539 	return setup_conf(mddev);
7540 }
7541 
7542 static void *raid5_takeover_raid6(struct mddev *mddev)
7543 {
7544 	int new_layout;
7545 
7546 	switch (mddev->layout) {
7547 	case ALGORITHM_LEFT_ASYMMETRIC_6:
7548 		new_layout = ALGORITHM_LEFT_ASYMMETRIC;
7549 		break;
7550 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
7551 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
7552 		break;
7553 	case ALGORITHM_LEFT_SYMMETRIC_6:
7554 		new_layout = ALGORITHM_LEFT_SYMMETRIC;
7555 		break;
7556 	case ALGORITHM_RIGHT_SYMMETRIC_6:
7557 		new_layout = ALGORITHM_RIGHT_SYMMETRIC;
7558 		break;
7559 	case ALGORITHM_PARITY_0_6:
7560 		new_layout = ALGORITHM_PARITY_0;
7561 		break;
7562 	case ALGORITHM_PARITY_N:
7563 		new_layout = ALGORITHM_PARITY_N;
7564 		break;
7565 	default:
7566 		return ERR_PTR(-EINVAL);
7567 	}
7568 	mddev->new_level = 5;
7569 	mddev->new_layout = new_layout;
7570 	mddev->delta_disks = -1;
7571 	mddev->raid_disks -= 1;
7572 	return setup_conf(mddev);
7573 }
7574 
7575 static int raid5_check_reshape(struct mddev *mddev)
7576 {
7577 	/* For a 2-drive array, the layout and chunk size can be changed
7578 	 * immediately as not restriping is needed.
7579 	 * For larger arrays we record the new value - after validation
7580 	 * to be used by a reshape pass.
7581 	 */
7582 	struct r5conf *conf = mddev->private;
7583 	int new_chunk = mddev->new_chunk_sectors;
7584 
7585 	if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
7586 		return -EINVAL;
7587 	if (new_chunk > 0) {
7588 		if (!is_power_of_2(new_chunk))
7589 			return -EINVAL;
7590 		if (new_chunk < (PAGE_SIZE>>9))
7591 			return -EINVAL;
7592 		if (mddev->array_sectors & (new_chunk-1))
7593 			/* not factor of array size */
7594 			return -EINVAL;
7595 	}
7596 
7597 	/* They look valid */
7598 
7599 	if (mddev->raid_disks == 2) {
7600 		/* can make the change immediately */
7601 		if (mddev->new_layout >= 0) {
7602 			conf->algorithm = mddev->new_layout;
7603 			mddev->layout = mddev->new_layout;
7604 		}
7605 		if (new_chunk > 0) {
7606 			conf->chunk_sectors = new_chunk ;
7607 			mddev->chunk_sectors = new_chunk;
7608 		}
7609 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
7610 		md_wakeup_thread(mddev->thread);
7611 	}
7612 	return check_reshape(mddev);
7613 }
7614 
7615 static int raid6_check_reshape(struct mddev *mddev)
7616 {
7617 	int new_chunk = mddev->new_chunk_sectors;
7618 
7619 	if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
7620 		return -EINVAL;
7621 	if (new_chunk > 0) {
7622 		if (!is_power_of_2(new_chunk))
7623 			return -EINVAL;
7624 		if (new_chunk < (PAGE_SIZE >> 9))
7625 			return -EINVAL;
7626 		if (mddev->array_sectors & (new_chunk-1))
7627 			/* not factor of array size */
7628 			return -EINVAL;
7629 	}
7630 
7631 	/* They look valid */
7632 	return check_reshape(mddev);
7633 }
7634 
7635 static void *raid5_takeover(struct mddev *mddev)
7636 {
7637 	/* raid5 can take over:
7638 	 *  raid0 - if there is only one strip zone - make it a raid4 layout
7639 	 *  raid1 - if there are two drives.  We need to know the chunk size
7640 	 *  raid4 - trivial - just use a raid4 layout.
7641 	 *  raid6 - Providing it is a *_6 layout
7642 	 */
7643 	if (mddev->level == 0)
7644 		return raid45_takeover_raid0(mddev, 5);
7645 	if (mddev->level == 1)
7646 		return raid5_takeover_raid1(mddev);
7647 	if (mddev->level == 4) {
7648 		mddev->new_layout = ALGORITHM_PARITY_N;
7649 		mddev->new_level = 5;
7650 		return setup_conf(mddev);
7651 	}
7652 	if (mddev->level == 6)
7653 		return raid5_takeover_raid6(mddev);
7654 
7655 	return ERR_PTR(-EINVAL);
7656 }
7657 
7658 static void *raid4_takeover(struct mddev *mddev)
7659 {
7660 	/* raid4 can take over:
7661 	 *  raid0 - if there is only one strip zone
7662 	 *  raid5 - if layout is right
7663 	 */
7664 	if (mddev->level == 0)
7665 		return raid45_takeover_raid0(mddev, 4);
7666 	if (mddev->level == 5 &&
7667 	    mddev->layout == ALGORITHM_PARITY_N) {
7668 		mddev->new_layout = 0;
7669 		mddev->new_level = 4;
7670 		return setup_conf(mddev);
7671 	}
7672 	return ERR_PTR(-EINVAL);
7673 }
7674 
7675 static struct md_personality raid5_personality;
7676 
7677 static void *raid6_takeover(struct mddev *mddev)
7678 {
7679 	/* Currently can only take over a raid5.  We map the
7680 	 * personality to an equivalent raid6 personality
7681 	 * with the Q block at the end.
7682 	 */
7683 	int new_layout;
7684 
7685 	if (mddev->pers != &raid5_personality)
7686 		return ERR_PTR(-EINVAL);
7687 	if (mddev->degraded > 1)
7688 		return ERR_PTR(-EINVAL);
7689 	if (mddev->raid_disks > 253)
7690 		return ERR_PTR(-EINVAL);
7691 	if (mddev->raid_disks < 3)
7692 		return ERR_PTR(-EINVAL);
7693 
7694 	switch (mddev->layout) {
7695 	case ALGORITHM_LEFT_ASYMMETRIC:
7696 		new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
7697 		break;
7698 	case ALGORITHM_RIGHT_ASYMMETRIC:
7699 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
7700 		break;
7701 	case ALGORITHM_LEFT_SYMMETRIC:
7702 		new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
7703 		break;
7704 	case ALGORITHM_RIGHT_SYMMETRIC:
7705 		new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
7706 		break;
7707 	case ALGORITHM_PARITY_0:
7708 		new_layout = ALGORITHM_PARITY_0_6;
7709 		break;
7710 	case ALGORITHM_PARITY_N:
7711 		new_layout = ALGORITHM_PARITY_N;
7712 		break;
7713 	default:
7714 		return ERR_PTR(-EINVAL);
7715 	}
7716 	mddev->new_level = 6;
7717 	mddev->new_layout = new_layout;
7718 	mddev->delta_disks = 1;
7719 	mddev->raid_disks += 1;
7720 	return setup_conf(mddev);
7721 }
7722 
7723 static struct md_personality raid6_personality =
7724 {
7725 	.name		= "raid6",
7726 	.level		= 6,
7727 	.owner		= THIS_MODULE,
7728 	.make_request	= make_request,
7729 	.run		= run,
7730 	.free		= raid5_free,
7731 	.status		= status,
7732 	.error_handler	= error,
7733 	.hot_add_disk	= raid5_add_disk,
7734 	.hot_remove_disk= raid5_remove_disk,
7735 	.spare_active	= raid5_spare_active,
7736 	.sync_request	= sync_request,
7737 	.resize		= raid5_resize,
7738 	.size		= raid5_size,
7739 	.check_reshape	= raid6_check_reshape,
7740 	.start_reshape  = raid5_start_reshape,
7741 	.finish_reshape = raid5_finish_reshape,
7742 	.quiesce	= raid5_quiesce,
7743 	.takeover	= raid6_takeover,
7744 	.congested	= raid5_congested,
7745 	.mergeable_bvec	= raid5_mergeable_bvec,
7746 };
7747 static struct md_personality raid5_personality =
7748 {
7749 	.name		= "raid5",
7750 	.level		= 5,
7751 	.owner		= THIS_MODULE,
7752 	.make_request	= make_request,
7753 	.run		= run,
7754 	.free		= raid5_free,
7755 	.status		= status,
7756 	.error_handler	= error,
7757 	.hot_add_disk	= raid5_add_disk,
7758 	.hot_remove_disk= raid5_remove_disk,
7759 	.spare_active	= raid5_spare_active,
7760 	.sync_request	= sync_request,
7761 	.resize		= raid5_resize,
7762 	.size		= raid5_size,
7763 	.check_reshape	= raid5_check_reshape,
7764 	.start_reshape  = raid5_start_reshape,
7765 	.finish_reshape = raid5_finish_reshape,
7766 	.quiesce	= raid5_quiesce,
7767 	.takeover	= raid5_takeover,
7768 	.congested	= raid5_congested,
7769 	.mergeable_bvec	= raid5_mergeable_bvec,
7770 };
7771 
7772 static struct md_personality raid4_personality =
7773 {
7774 	.name		= "raid4",
7775 	.level		= 4,
7776 	.owner		= THIS_MODULE,
7777 	.make_request	= make_request,
7778 	.run		= run,
7779 	.free		= raid5_free,
7780 	.status		= status,
7781 	.error_handler	= error,
7782 	.hot_add_disk	= raid5_add_disk,
7783 	.hot_remove_disk= raid5_remove_disk,
7784 	.spare_active	= raid5_spare_active,
7785 	.sync_request	= sync_request,
7786 	.resize		= raid5_resize,
7787 	.size		= raid5_size,
7788 	.check_reshape	= raid5_check_reshape,
7789 	.start_reshape  = raid5_start_reshape,
7790 	.finish_reshape = raid5_finish_reshape,
7791 	.quiesce	= raid5_quiesce,
7792 	.takeover	= raid4_takeover,
7793 	.congested	= raid5_congested,
7794 	.mergeable_bvec	= raid5_mergeable_bvec,
7795 };
7796 
7797 static int __init raid5_init(void)
7798 {
7799 	raid5_wq = alloc_workqueue("raid5wq",
7800 		WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
7801 	if (!raid5_wq)
7802 		return -ENOMEM;
7803 	register_md_personality(&raid6_personality);
7804 	register_md_personality(&raid5_personality);
7805 	register_md_personality(&raid4_personality);
7806 	return 0;
7807 }
7808 
7809 static void raid5_exit(void)
7810 {
7811 	unregister_md_personality(&raid6_personality);
7812 	unregister_md_personality(&raid5_personality);
7813 	unregister_md_personality(&raid4_personality);
7814 	destroy_workqueue(raid5_wq);
7815 }
7816 
7817 module_init(raid5_init);
7818 module_exit(raid5_exit);
7819 MODULE_LICENSE("GPL");
7820 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
7821 MODULE_ALIAS("md-personality-4"); /* RAID5 */
7822 MODULE_ALIAS("md-raid5");
7823 MODULE_ALIAS("md-raid4");
7824 MODULE_ALIAS("md-level-5");
7825 MODULE_ALIAS("md-level-4");
7826 MODULE_ALIAS("md-personality-8"); /* RAID6 */
7827 MODULE_ALIAS("md-raid6");
7828 MODULE_ALIAS("md-level-6");
7829 
7830 /* This used to be two separate modules, they were: */
7831 MODULE_ALIAS("raid5");
7832 MODULE_ALIAS("raid6");
7833