xref: /openbmc/linux/drivers/md/raid5.c (revision 4a075bd4)
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 
58 #include <trace/events/block.h>
59 #include <linux/list_sort.h>
60 
61 #include "md.h"
62 #include "raid5.h"
63 #include "raid0.h"
64 #include "md-bitmap.h"
65 #include "raid5-log.h"
66 
67 #define UNSUPPORTED_MDDEV_FLAGS	(1L << MD_FAILFAST_SUPPORTED)
68 
69 #define cpu_to_group(cpu) cpu_to_node(cpu)
70 #define ANY_GROUP NUMA_NO_NODE
71 
72 static bool devices_handle_discard_safely = false;
73 module_param(devices_handle_discard_safely, bool, 0644);
74 MODULE_PARM_DESC(devices_handle_discard_safely,
75 		 "Set to Y if all devices in each array reliably return zeroes on reads from discarded regions");
76 static struct workqueue_struct *raid5_wq;
77 
78 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
79 {
80 	int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
81 	return &conf->stripe_hashtbl[hash];
82 }
83 
84 static inline int stripe_hash_locks_hash(sector_t sect)
85 {
86 	return (sect >> STRIPE_SHIFT) & STRIPE_HASH_LOCKS_MASK;
87 }
88 
89 static inline void lock_device_hash_lock(struct r5conf *conf, int hash)
90 {
91 	spin_lock_irq(conf->hash_locks + hash);
92 	spin_lock(&conf->device_lock);
93 }
94 
95 static inline void unlock_device_hash_lock(struct r5conf *conf, int hash)
96 {
97 	spin_unlock(&conf->device_lock);
98 	spin_unlock_irq(conf->hash_locks + hash);
99 }
100 
101 static inline void lock_all_device_hash_locks_irq(struct r5conf *conf)
102 {
103 	int i;
104 	spin_lock_irq(conf->hash_locks);
105 	for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
106 		spin_lock_nest_lock(conf->hash_locks + i, conf->hash_locks);
107 	spin_lock(&conf->device_lock);
108 }
109 
110 static inline void unlock_all_device_hash_locks_irq(struct r5conf *conf)
111 {
112 	int i;
113 	spin_unlock(&conf->device_lock);
114 	for (i = NR_STRIPE_HASH_LOCKS - 1; i; i--)
115 		spin_unlock(conf->hash_locks + i);
116 	spin_unlock_irq(conf->hash_locks);
117 }
118 
119 /* Find first data disk in a raid6 stripe */
120 static inline int raid6_d0(struct stripe_head *sh)
121 {
122 	if (sh->ddf_layout)
123 		/* ddf always start from first device */
124 		return 0;
125 	/* md starts just after Q block */
126 	if (sh->qd_idx == sh->disks - 1)
127 		return 0;
128 	else
129 		return sh->qd_idx + 1;
130 }
131 static inline int raid6_next_disk(int disk, int raid_disks)
132 {
133 	disk++;
134 	return (disk < raid_disks) ? disk : 0;
135 }
136 
137 /* When walking through the disks in a raid5, starting at raid6_d0,
138  * We need to map each disk to a 'slot', where the data disks are slot
139  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
140  * is raid_disks-1.  This help does that mapping.
141  */
142 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
143 			     int *count, int syndrome_disks)
144 {
145 	int slot = *count;
146 
147 	if (sh->ddf_layout)
148 		(*count)++;
149 	if (idx == sh->pd_idx)
150 		return syndrome_disks;
151 	if (idx == sh->qd_idx)
152 		return syndrome_disks + 1;
153 	if (!sh->ddf_layout)
154 		(*count)++;
155 	return slot;
156 }
157 
158 static void print_raid5_conf (struct r5conf *conf);
159 
160 static int stripe_operations_active(struct stripe_head *sh)
161 {
162 	return sh->check_state || sh->reconstruct_state ||
163 	       test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
164 	       test_bit(STRIPE_COMPUTE_RUN, &sh->state);
165 }
166 
167 static bool stripe_is_lowprio(struct stripe_head *sh)
168 {
169 	return (test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) ||
170 		test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) &&
171 	       !test_bit(STRIPE_R5C_CACHING, &sh->state);
172 }
173 
174 static void raid5_wakeup_stripe_thread(struct stripe_head *sh)
175 {
176 	struct r5conf *conf = sh->raid_conf;
177 	struct r5worker_group *group;
178 	int thread_cnt;
179 	int i, cpu = sh->cpu;
180 
181 	if (!cpu_online(cpu)) {
182 		cpu = cpumask_any(cpu_online_mask);
183 		sh->cpu = cpu;
184 	}
185 
186 	if (list_empty(&sh->lru)) {
187 		struct r5worker_group *group;
188 		group = conf->worker_groups + cpu_to_group(cpu);
189 		if (stripe_is_lowprio(sh))
190 			list_add_tail(&sh->lru, &group->loprio_list);
191 		else
192 			list_add_tail(&sh->lru, &group->handle_list);
193 		group->stripes_cnt++;
194 		sh->group = group;
195 	}
196 
197 	if (conf->worker_cnt_per_group == 0) {
198 		md_wakeup_thread(conf->mddev->thread);
199 		return;
200 	}
201 
202 	group = conf->worker_groups + cpu_to_group(sh->cpu);
203 
204 	group->workers[0].working = true;
205 	/* at least one worker should run to avoid race */
206 	queue_work_on(sh->cpu, raid5_wq, &group->workers[0].work);
207 
208 	thread_cnt = group->stripes_cnt / MAX_STRIPE_BATCH - 1;
209 	/* wakeup more workers */
210 	for (i = 1; i < conf->worker_cnt_per_group && thread_cnt > 0; i++) {
211 		if (group->workers[i].working == false) {
212 			group->workers[i].working = true;
213 			queue_work_on(sh->cpu, raid5_wq,
214 				      &group->workers[i].work);
215 			thread_cnt--;
216 		}
217 	}
218 }
219 
220 static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
221 			      struct list_head *temp_inactive_list)
222 {
223 	int i;
224 	int injournal = 0;	/* number of date pages with R5_InJournal */
225 
226 	BUG_ON(!list_empty(&sh->lru));
227 	BUG_ON(atomic_read(&conf->active_stripes)==0);
228 
229 	if (r5c_is_writeback(conf->log))
230 		for (i = sh->disks; i--; )
231 			if (test_bit(R5_InJournal, &sh->dev[i].flags))
232 				injournal++;
233 	/*
234 	 * In the following cases, the stripe cannot be released to cached
235 	 * lists. Therefore, we make the stripe write out and set
236 	 * STRIPE_HANDLE:
237 	 *   1. when quiesce in r5c write back;
238 	 *   2. when resync is requested fot the stripe.
239 	 */
240 	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) ||
241 	    (conf->quiesce && r5c_is_writeback(conf->log) &&
242 	     !test_bit(STRIPE_HANDLE, &sh->state) && injournal != 0)) {
243 		if (test_bit(STRIPE_R5C_CACHING, &sh->state))
244 			r5c_make_stripe_write_out(sh);
245 		set_bit(STRIPE_HANDLE, &sh->state);
246 	}
247 
248 	if (test_bit(STRIPE_HANDLE, &sh->state)) {
249 		if (test_bit(STRIPE_DELAYED, &sh->state) &&
250 		    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
251 			list_add_tail(&sh->lru, &conf->delayed_list);
252 		else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
253 			   sh->bm_seq - conf->seq_write > 0)
254 			list_add_tail(&sh->lru, &conf->bitmap_list);
255 		else {
256 			clear_bit(STRIPE_DELAYED, &sh->state);
257 			clear_bit(STRIPE_BIT_DELAY, &sh->state);
258 			if (conf->worker_cnt_per_group == 0) {
259 				if (stripe_is_lowprio(sh))
260 					list_add_tail(&sh->lru,
261 							&conf->loprio_list);
262 				else
263 					list_add_tail(&sh->lru,
264 							&conf->handle_list);
265 			} else {
266 				raid5_wakeup_stripe_thread(sh);
267 				return;
268 			}
269 		}
270 		md_wakeup_thread(conf->mddev->thread);
271 	} else {
272 		BUG_ON(stripe_operations_active(sh));
273 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
274 			if (atomic_dec_return(&conf->preread_active_stripes)
275 			    < IO_THRESHOLD)
276 				md_wakeup_thread(conf->mddev->thread);
277 		atomic_dec(&conf->active_stripes);
278 		if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
279 			if (!r5c_is_writeback(conf->log))
280 				list_add_tail(&sh->lru, temp_inactive_list);
281 			else {
282 				WARN_ON(test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags));
283 				if (injournal == 0)
284 					list_add_tail(&sh->lru, temp_inactive_list);
285 				else if (injournal == conf->raid_disks - conf->max_degraded) {
286 					/* full stripe */
287 					if (!test_and_set_bit(STRIPE_R5C_FULL_STRIPE, &sh->state))
288 						atomic_inc(&conf->r5c_cached_full_stripes);
289 					if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
290 						atomic_dec(&conf->r5c_cached_partial_stripes);
291 					list_add_tail(&sh->lru, &conf->r5c_full_stripe_list);
292 					r5c_check_cached_full_stripe(conf);
293 				} else
294 					/*
295 					 * STRIPE_R5C_PARTIAL_STRIPE is set in
296 					 * r5c_try_caching_write(). No need to
297 					 * set it again.
298 					 */
299 					list_add_tail(&sh->lru, &conf->r5c_partial_stripe_list);
300 			}
301 		}
302 	}
303 }
304 
305 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh,
306 			     struct list_head *temp_inactive_list)
307 {
308 	if (atomic_dec_and_test(&sh->count))
309 		do_release_stripe(conf, sh, temp_inactive_list);
310 }
311 
312 /*
313  * @hash could be NR_STRIPE_HASH_LOCKS, then we have a list of inactive_list
314  *
315  * Be careful: Only one task can add/delete stripes from temp_inactive_list at
316  * given time. Adding stripes only takes device lock, while deleting stripes
317  * only takes hash lock.
318  */
319 static void release_inactive_stripe_list(struct r5conf *conf,
320 					 struct list_head *temp_inactive_list,
321 					 int hash)
322 {
323 	int size;
324 	bool do_wakeup = false;
325 	unsigned long flags;
326 
327 	if (hash == NR_STRIPE_HASH_LOCKS) {
328 		size = NR_STRIPE_HASH_LOCKS;
329 		hash = NR_STRIPE_HASH_LOCKS - 1;
330 	} else
331 		size = 1;
332 	while (size) {
333 		struct list_head *list = &temp_inactive_list[size - 1];
334 
335 		/*
336 		 * We don't hold any lock here yet, raid5_get_active_stripe() might
337 		 * remove stripes from the list
338 		 */
339 		if (!list_empty_careful(list)) {
340 			spin_lock_irqsave(conf->hash_locks + hash, flags);
341 			if (list_empty(conf->inactive_list + hash) &&
342 			    !list_empty(list))
343 				atomic_dec(&conf->empty_inactive_list_nr);
344 			list_splice_tail_init(list, conf->inactive_list + hash);
345 			do_wakeup = true;
346 			spin_unlock_irqrestore(conf->hash_locks + hash, flags);
347 		}
348 		size--;
349 		hash--;
350 	}
351 
352 	if (do_wakeup) {
353 		wake_up(&conf->wait_for_stripe);
354 		if (atomic_read(&conf->active_stripes) == 0)
355 			wake_up(&conf->wait_for_quiescent);
356 		if (conf->retry_read_aligned)
357 			md_wakeup_thread(conf->mddev->thread);
358 	}
359 }
360 
361 /* should hold conf->device_lock already */
362 static int release_stripe_list(struct r5conf *conf,
363 			       struct list_head *temp_inactive_list)
364 {
365 	struct stripe_head *sh, *t;
366 	int count = 0;
367 	struct llist_node *head;
368 
369 	head = llist_del_all(&conf->released_stripes);
370 	head = llist_reverse_order(head);
371 	llist_for_each_entry_safe(sh, t, head, release_list) {
372 		int hash;
373 
374 		/* sh could be readded after STRIPE_ON_RELEASE_LIST is cleard */
375 		smp_mb();
376 		clear_bit(STRIPE_ON_RELEASE_LIST, &sh->state);
377 		/*
378 		 * Don't worry the bit is set here, because if the bit is set
379 		 * again, the count is always > 1. This is true for
380 		 * STRIPE_ON_UNPLUG_LIST bit too.
381 		 */
382 		hash = sh->hash_lock_index;
383 		__release_stripe(conf, sh, &temp_inactive_list[hash]);
384 		count++;
385 	}
386 
387 	return count;
388 }
389 
390 void raid5_release_stripe(struct stripe_head *sh)
391 {
392 	struct r5conf *conf = sh->raid_conf;
393 	unsigned long flags;
394 	struct list_head list;
395 	int hash;
396 	bool wakeup;
397 
398 	/* Avoid release_list until the last reference.
399 	 */
400 	if (atomic_add_unless(&sh->count, -1, 1))
401 		return;
402 
403 	if (unlikely(!conf->mddev->thread) ||
404 		test_and_set_bit(STRIPE_ON_RELEASE_LIST, &sh->state))
405 		goto slow_path;
406 	wakeup = llist_add(&sh->release_list, &conf->released_stripes);
407 	if (wakeup)
408 		md_wakeup_thread(conf->mddev->thread);
409 	return;
410 slow_path:
411 	/* we are ok here if STRIPE_ON_RELEASE_LIST is set or not */
412 	if (atomic_dec_and_lock_irqsave(&sh->count, &conf->device_lock, flags)) {
413 		INIT_LIST_HEAD(&list);
414 		hash = sh->hash_lock_index;
415 		do_release_stripe(conf, sh, &list);
416 		spin_unlock_irqrestore(&conf->device_lock, flags);
417 		release_inactive_stripe_list(conf, &list, hash);
418 	}
419 }
420 
421 static inline void remove_hash(struct stripe_head *sh)
422 {
423 	pr_debug("remove_hash(), stripe %llu\n",
424 		(unsigned long long)sh->sector);
425 
426 	hlist_del_init(&sh->hash);
427 }
428 
429 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
430 {
431 	struct hlist_head *hp = stripe_hash(conf, sh->sector);
432 
433 	pr_debug("insert_hash(), stripe %llu\n",
434 		(unsigned long long)sh->sector);
435 
436 	hlist_add_head(&sh->hash, hp);
437 }
438 
439 /* find an idle stripe, make sure it is unhashed, and return it. */
440 static struct stripe_head *get_free_stripe(struct r5conf *conf, int hash)
441 {
442 	struct stripe_head *sh = NULL;
443 	struct list_head *first;
444 
445 	if (list_empty(conf->inactive_list + hash))
446 		goto out;
447 	first = (conf->inactive_list + hash)->next;
448 	sh = list_entry(first, struct stripe_head, lru);
449 	list_del_init(first);
450 	remove_hash(sh);
451 	atomic_inc(&conf->active_stripes);
452 	BUG_ON(hash != sh->hash_lock_index);
453 	if (list_empty(conf->inactive_list + hash))
454 		atomic_inc(&conf->empty_inactive_list_nr);
455 out:
456 	return sh;
457 }
458 
459 static void shrink_buffers(struct stripe_head *sh)
460 {
461 	struct page *p;
462 	int i;
463 	int num = sh->raid_conf->pool_size;
464 
465 	for (i = 0; i < num ; i++) {
466 		WARN_ON(sh->dev[i].page != sh->dev[i].orig_page);
467 		p = sh->dev[i].page;
468 		if (!p)
469 			continue;
470 		sh->dev[i].page = NULL;
471 		put_page(p);
472 	}
473 }
474 
475 static int grow_buffers(struct stripe_head *sh, gfp_t gfp)
476 {
477 	int i;
478 	int num = sh->raid_conf->pool_size;
479 
480 	for (i = 0; i < num; i++) {
481 		struct page *page;
482 
483 		if (!(page = alloc_page(gfp))) {
484 			return 1;
485 		}
486 		sh->dev[i].page = page;
487 		sh->dev[i].orig_page = page;
488 	}
489 
490 	return 0;
491 }
492 
493 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
494 			    struct stripe_head *sh);
495 
496 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
497 {
498 	struct r5conf *conf = sh->raid_conf;
499 	int i, seq;
500 
501 	BUG_ON(atomic_read(&sh->count) != 0);
502 	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
503 	BUG_ON(stripe_operations_active(sh));
504 	BUG_ON(sh->batch_head);
505 
506 	pr_debug("init_stripe called, stripe %llu\n",
507 		(unsigned long long)sector);
508 retry:
509 	seq = read_seqcount_begin(&conf->gen_lock);
510 	sh->generation = conf->generation - previous;
511 	sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
512 	sh->sector = sector;
513 	stripe_set_idx(sector, conf, previous, sh);
514 	sh->state = 0;
515 
516 	for (i = sh->disks; i--; ) {
517 		struct r5dev *dev = &sh->dev[i];
518 
519 		if (dev->toread || dev->read || dev->towrite || dev->written ||
520 		    test_bit(R5_LOCKED, &dev->flags)) {
521 			pr_err("sector=%llx i=%d %p %p %p %p %d\n",
522 			       (unsigned long long)sh->sector, i, dev->toread,
523 			       dev->read, dev->towrite, dev->written,
524 			       test_bit(R5_LOCKED, &dev->flags));
525 			WARN_ON(1);
526 		}
527 		dev->flags = 0;
528 		dev->sector = raid5_compute_blocknr(sh, i, previous);
529 	}
530 	if (read_seqcount_retry(&conf->gen_lock, seq))
531 		goto retry;
532 	sh->overwrite_disks = 0;
533 	insert_hash(conf, sh);
534 	sh->cpu = smp_processor_id();
535 	set_bit(STRIPE_BATCH_READY, &sh->state);
536 }
537 
538 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
539 					 short generation)
540 {
541 	struct stripe_head *sh;
542 
543 	pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
544 	hlist_for_each_entry(sh, stripe_hash(conf, sector), hash)
545 		if (sh->sector == sector && sh->generation == generation)
546 			return sh;
547 	pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
548 	return NULL;
549 }
550 
551 /*
552  * Need to check if array has failed when deciding whether to:
553  *  - start an array
554  *  - remove non-faulty devices
555  *  - add a spare
556  *  - allow a reshape
557  * This determination is simple when no reshape is happening.
558  * However if there is a reshape, we need to carefully check
559  * both the before and after sections.
560  * This is because some failed devices may only affect one
561  * of the two sections, and some non-in_sync devices may
562  * be insync in the section most affected by failed devices.
563  */
564 int raid5_calc_degraded(struct r5conf *conf)
565 {
566 	int degraded, degraded2;
567 	int i;
568 
569 	rcu_read_lock();
570 	degraded = 0;
571 	for (i = 0; i < conf->previous_raid_disks; i++) {
572 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
573 		if (rdev && test_bit(Faulty, &rdev->flags))
574 			rdev = rcu_dereference(conf->disks[i].replacement);
575 		if (!rdev || test_bit(Faulty, &rdev->flags))
576 			degraded++;
577 		else if (test_bit(In_sync, &rdev->flags))
578 			;
579 		else
580 			/* not in-sync or faulty.
581 			 * If the reshape increases the number of devices,
582 			 * this is being recovered by the reshape, so
583 			 * this 'previous' section is not in_sync.
584 			 * If the number of devices is being reduced however,
585 			 * the device can only be part of the array if
586 			 * we are reverting a reshape, so this section will
587 			 * be in-sync.
588 			 */
589 			if (conf->raid_disks >= conf->previous_raid_disks)
590 				degraded++;
591 	}
592 	rcu_read_unlock();
593 	if (conf->raid_disks == conf->previous_raid_disks)
594 		return degraded;
595 	rcu_read_lock();
596 	degraded2 = 0;
597 	for (i = 0; i < conf->raid_disks; i++) {
598 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
599 		if (rdev && test_bit(Faulty, &rdev->flags))
600 			rdev = rcu_dereference(conf->disks[i].replacement);
601 		if (!rdev || test_bit(Faulty, &rdev->flags))
602 			degraded2++;
603 		else if (test_bit(In_sync, &rdev->flags))
604 			;
605 		else
606 			/* not in-sync or faulty.
607 			 * If reshape increases the number of devices, this
608 			 * section has already been recovered, else it
609 			 * almost certainly hasn't.
610 			 */
611 			if (conf->raid_disks <= conf->previous_raid_disks)
612 				degraded2++;
613 	}
614 	rcu_read_unlock();
615 	if (degraded2 > degraded)
616 		return degraded2;
617 	return degraded;
618 }
619 
620 static int has_failed(struct r5conf *conf)
621 {
622 	int degraded;
623 
624 	if (conf->mddev->reshape_position == MaxSector)
625 		return conf->mddev->degraded > conf->max_degraded;
626 
627 	degraded = raid5_calc_degraded(conf);
628 	if (degraded > conf->max_degraded)
629 		return 1;
630 	return 0;
631 }
632 
633 struct stripe_head *
634 raid5_get_active_stripe(struct r5conf *conf, sector_t sector,
635 			int previous, int noblock, int noquiesce)
636 {
637 	struct stripe_head *sh;
638 	int hash = stripe_hash_locks_hash(sector);
639 	int inc_empty_inactive_list_flag;
640 
641 	pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
642 
643 	spin_lock_irq(conf->hash_locks + hash);
644 
645 	do {
646 		wait_event_lock_irq(conf->wait_for_quiescent,
647 				    conf->quiesce == 0 || noquiesce,
648 				    *(conf->hash_locks + hash));
649 		sh = __find_stripe(conf, sector, conf->generation - previous);
650 		if (!sh) {
651 			if (!test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state)) {
652 				sh = get_free_stripe(conf, hash);
653 				if (!sh && !test_bit(R5_DID_ALLOC,
654 						     &conf->cache_state))
655 					set_bit(R5_ALLOC_MORE,
656 						&conf->cache_state);
657 			}
658 			if (noblock && sh == NULL)
659 				break;
660 
661 			r5c_check_stripe_cache_usage(conf);
662 			if (!sh) {
663 				set_bit(R5_INACTIVE_BLOCKED,
664 					&conf->cache_state);
665 				r5l_wake_reclaim(conf->log, 0);
666 				wait_event_lock_irq(
667 					conf->wait_for_stripe,
668 					!list_empty(conf->inactive_list + hash) &&
669 					(atomic_read(&conf->active_stripes)
670 					 < (conf->max_nr_stripes * 3 / 4)
671 					 || !test_bit(R5_INACTIVE_BLOCKED,
672 						      &conf->cache_state)),
673 					*(conf->hash_locks + hash));
674 				clear_bit(R5_INACTIVE_BLOCKED,
675 					  &conf->cache_state);
676 			} else {
677 				init_stripe(sh, sector, previous);
678 				atomic_inc(&sh->count);
679 			}
680 		} else if (!atomic_inc_not_zero(&sh->count)) {
681 			spin_lock(&conf->device_lock);
682 			if (!atomic_read(&sh->count)) {
683 				if (!test_bit(STRIPE_HANDLE, &sh->state))
684 					atomic_inc(&conf->active_stripes);
685 				BUG_ON(list_empty(&sh->lru) &&
686 				       !test_bit(STRIPE_EXPANDING, &sh->state));
687 				inc_empty_inactive_list_flag = 0;
688 				if (!list_empty(conf->inactive_list + hash))
689 					inc_empty_inactive_list_flag = 1;
690 				list_del_init(&sh->lru);
691 				if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
692 					atomic_inc(&conf->empty_inactive_list_nr);
693 				if (sh->group) {
694 					sh->group->stripes_cnt--;
695 					sh->group = NULL;
696 				}
697 			}
698 			atomic_inc(&sh->count);
699 			spin_unlock(&conf->device_lock);
700 		}
701 	} while (sh == NULL);
702 
703 	spin_unlock_irq(conf->hash_locks + hash);
704 	return sh;
705 }
706 
707 static bool is_full_stripe_write(struct stripe_head *sh)
708 {
709 	BUG_ON(sh->overwrite_disks > (sh->disks - sh->raid_conf->max_degraded));
710 	return sh->overwrite_disks == (sh->disks - sh->raid_conf->max_degraded);
711 }
712 
713 static void lock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
714 		__acquires(&sh1->stripe_lock)
715 		__acquires(&sh2->stripe_lock)
716 {
717 	if (sh1 > sh2) {
718 		spin_lock_irq(&sh2->stripe_lock);
719 		spin_lock_nested(&sh1->stripe_lock, 1);
720 	} else {
721 		spin_lock_irq(&sh1->stripe_lock);
722 		spin_lock_nested(&sh2->stripe_lock, 1);
723 	}
724 }
725 
726 static void unlock_two_stripes(struct stripe_head *sh1, struct stripe_head *sh2)
727 		__releases(&sh1->stripe_lock)
728 		__releases(&sh2->stripe_lock)
729 {
730 	spin_unlock(&sh1->stripe_lock);
731 	spin_unlock_irq(&sh2->stripe_lock);
732 }
733 
734 /* Only freshly new full stripe normal write stripe can be added to a batch list */
735 static bool stripe_can_batch(struct stripe_head *sh)
736 {
737 	struct r5conf *conf = sh->raid_conf;
738 
739 	if (raid5_has_log(conf) || raid5_has_ppl(conf))
740 		return false;
741 	return test_bit(STRIPE_BATCH_READY, &sh->state) &&
742 		!test_bit(STRIPE_BITMAP_PENDING, &sh->state) &&
743 		is_full_stripe_write(sh);
744 }
745 
746 /* we only do back search */
747 static void stripe_add_to_batch_list(struct r5conf *conf, struct stripe_head *sh)
748 {
749 	struct stripe_head *head;
750 	sector_t head_sector, tmp_sec;
751 	int hash;
752 	int dd_idx;
753 	int inc_empty_inactive_list_flag;
754 
755 	/* Don't cross chunks, so stripe pd_idx/qd_idx is the same */
756 	tmp_sec = sh->sector;
757 	if (!sector_div(tmp_sec, conf->chunk_sectors))
758 		return;
759 	head_sector = sh->sector - STRIPE_SECTORS;
760 
761 	hash = stripe_hash_locks_hash(head_sector);
762 	spin_lock_irq(conf->hash_locks + hash);
763 	head = __find_stripe(conf, head_sector, conf->generation);
764 	if (head && !atomic_inc_not_zero(&head->count)) {
765 		spin_lock(&conf->device_lock);
766 		if (!atomic_read(&head->count)) {
767 			if (!test_bit(STRIPE_HANDLE, &head->state))
768 				atomic_inc(&conf->active_stripes);
769 			BUG_ON(list_empty(&head->lru) &&
770 			       !test_bit(STRIPE_EXPANDING, &head->state));
771 			inc_empty_inactive_list_flag = 0;
772 			if (!list_empty(conf->inactive_list + hash))
773 				inc_empty_inactive_list_flag = 1;
774 			list_del_init(&head->lru);
775 			if (list_empty(conf->inactive_list + hash) && inc_empty_inactive_list_flag)
776 				atomic_inc(&conf->empty_inactive_list_nr);
777 			if (head->group) {
778 				head->group->stripes_cnt--;
779 				head->group = NULL;
780 			}
781 		}
782 		atomic_inc(&head->count);
783 		spin_unlock(&conf->device_lock);
784 	}
785 	spin_unlock_irq(conf->hash_locks + hash);
786 
787 	if (!head)
788 		return;
789 	if (!stripe_can_batch(head))
790 		goto out;
791 
792 	lock_two_stripes(head, sh);
793 	/* clear_batch_ready clear the flag */
794 	if (!stripe_can_batch(head) || !stripe_can_batch(sh))
795 		goto unlock_out;
796 
797 	if (sh->batch_head)
798 		goto unlock_out;
799 
800 	dd_idx = 0;
801 	while (dd_idx == sh->pd_idx || dd_idx == sh->qd_idx)
802 		dd_idx++;
803 	if (head->dev[dd_idx].towrite->bi_opf != sh->dev[dd_idx].towrite->bi_opf ||
804 	    bio_op(head->dev[dd_idx].towrite) != bio_op(sh->dev[dd_idx].towrite))
805 		goto unlock_out;
806 
807 	if (head->batch_head) {
808 		spin_lock(&head->batch_head->batch_lock);
809 		/* This batch list is already running */
810 		if (!stripe_can_batch(head)) {
811 			spin_unlock(&head->batch_head->batch_lock);
812 			goto unlock_out;
813 		}
814 		/*
815 		 * We must assign batch_head of this stripe within the
816 		 * batch_lock, otherwise clear_batch_ready of batch head
817 		 * stripe could clear BATCH_READY bit of this stripe and
818 		 * this stripe->batch_head doesn't get assigned, which
819 		 * could confuse clear_batch_ready for this stripe
820 		 */
821 		sh->batch_head = head->batch_head;
822 
823 		/*
824 		 * at this point, head's BATCH_READY could be cleared, but we
825 		 * can still add the stripe to batch list
826 		 */
827 		list_add(&sh->batch_list, &head->batch_list);
828 		spin_unlock(&head->batch_head->batch_lock);
829 	} else {
830 		head->batch_head = head;
831 		sh->batch_head = head->batch_head;
832 		spin_lock(&head->batch_lock);
833 		list_add_tail(&sh->batch_list, &head->batch_list);
834 		spin_unlock(&head->batch_lock);
835 	}
836 
837 	if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
838 		if (atomic_dec_return(&conf->preread_active_stripes)
839 		    < IO_THRESHOLD)
840 			md_wakeup_thread(conf->mddev->thread);
841 
842 	if (test_and_clear_bit(STRIPE_BIT_DELAY, &sh->state)) {
843 		int seq = sh->bm_seq;
844 		if (test_bit(STRIPE_BIT_DELAY, &sh->batch_head->state) &&
845 		    sh->batch_head->bm_seq > seq)
846 			seq = sh->batch_head->bm_seq;
847 		set_bit(STRIPE_BIT_DELAY, &sh->batch_head->state);
848 		sh->batch_head->bm_seq = seq;
849 	}
850 
851 	atomic_inc(&sh->count);
852 unlock_out:
853 	unlock_two_stripes(head, sh);
854 out:
855 	raid5_release_stripe(head);
856 }
857 
858 /* Determine if 'data_offset' or 'new_data_offset' should be used
859  * in this stripe_head.
860  */
861 static int use_new_offset(struct r5conf *conf, struct stripe_head *sh)
862 {
863 	sector_t progress = conf->reshape_progress;
864 	/* Need a memory barrier to make sure we see the value
865 	 * of conf->generation, or ->data_offset that was set before
866 	 * reshape_progress was updated.
867 	 */
868 	smp_rmb();
869 	if (progress == MaxSector)
870 		return 0;
871 	if (sh->generation == conf->generation - 1)
872 		return 0;
873 	/* We are in a reshape, and this is a new-generation stripe,
874 	 * so use new_data_offset.
875 	 */
876 	return 1;
877 }
878 
879 static void dispatch_bio_list(struct bio_list *tmp)
880 {
881 	struct bio *bio;
882 
883 	while ((bio = bio_list_pop(tmp)))
884 		generic_make_request(bio);
885 }
886 
887 static int cmp_stripe(void *priv, struct list_head *a, struct list_head *b)
888 {
889 	const struct r5pending_data *da = list_entry(a,
890 				struct r5pending_data, sibling);
891 	const struct r5pending_data *db = list_entry(b,
892 				struct r5pending_data, sibling);
893 	if (da->sector > db->sector)
894 		return 1;
895 	if (da->sector < db->sector)
896 		return -1;
897 	return 0;
898 }
899 
900 static void dispatch_defer_bios(struct r5conf *conf, int target,
901 				struct bio_list *list)
902 {
903 	struct r5pending_data *data;
904 	struct list_head *first, *next = NULL;
905 	int cnt = 0;
906 
907 	if (conf->pending_data_cnt == 0)
908 		return;
909 
910 	list_sort(NULL, &conf->pending_list, cmp_stripe);
911 
912 	first = conf->pending_list.next;
913 
914 	/* temporarily move the head */
915 	if (conf->next_pending_data)
916 		list_move_tail(&conf->pending_list,
917 				&conf->next_pending_data->sibling);
918 
919 	while (!list_empty(&conf->pending_list)) {
920 		data = list_first_entry(&conf->pending_list,
921 			struct r5pending_data, sibling);
922 		if (&data->sibling == first)
923 			first = data->sibling.next;
924 		next = data->sibling.next;
925 
926 		bio_list_merge(list, &data->bios);
927 		list_move(&data->sibling, &conf->free_list);
928 		cnt++;
929 		if (cnt >= target)
930 			break;
931 	}
932 	conf->pending_data_cnt -= cnt;
933 	BUG_ON(conf->pending_data_cnt < 0 || cnt < target);
934 
935 	if (next != &conf->pending_list)
936 		conf->next_pending_data = list_entry(next,
937 				struct r5pending_data, sibling);
938 	else
939 		conf->next_pending_data = NULL;
940 	/* list isn't empty */
941 	if (first != &conf->pending_list)
942 		list_move_tail(&conf->pending_list, first);
943 }
944 
945 static void flush_deferred_bios(struct r5conf *conf)
946 {
947 	struct bio_list tmp = BIO_EMPTY_LIST;
948 
949 	if (conf->pending_data_cnt == 0)
950 		return;
951 
952 	spin_lock(&conf->pending_bios_lock);
953 	dispatch_defer_bios(conf, conf->pending_data_cnt, &tmp);
954 	BUG_ON(conf->pending_data_cnt != 0);
955 	spin_unlock(&conf->pending_bios_lock);
956 
957 	dispatch_bio_list(&tmp);
958 }
959 
960 static void defer_issue_bios(struct r5conf *conf, sector_t sector,
961 				struct bio_list *bios)
962 {
963 	struct bio_list tmp = BIO_EMPTY_LIST;
964 	struct r5pending_data *ent;
965 
966 	spin_lock(&conf->pending_bios_lock);
967 	ent = list_first_entry(&conf->free_list, struct r5pending_data,
968 							sibling);
969 	list_move_tail(&ent->sibling, &conf->pending_list);
970 	ent->sector = sector;
971 	bio_list_init(&ent->bios);
972 	bio_list_merge(&ent->bios, bios);
973 	conf->pending_data_cnt++;
974 	if (conf->pending_data_cnt >= PENDING_IO_MAX)
975 		dispatch_defer_bios(conf, PENDING_IO_ONE_FLUSH, &tmp);
976 
977 	spin_unlock(&conf->pending_bios_lock);
978 
979 	dispatch_bio_list(&tmp);
980 }
981 
982 static void
983 raid5_end_read_request(struct bio *bi);
984 static void
985 raid5_end_write_request(struct bio *bi);
986 
987 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
988 {
989 	struct r5conf *conf = sh->raid_conf;
990 	int i, disks = sh->disks;
991 	struct stripe_head *head_sh = sh;
992 	struct bio_list pending_bios = BIO_EMPTY_LIST;
993 	bool should_defer;
994 
995 	might_sleep();
996 
997 	if (log_stripe(sh, s) == 0)
998 		return;
999 
1000 	should_defer = conf->batch_bio_dispatch && conf->group_cnt;
1001 
1002 	for (i = disks; i--; ) {
1003 		int op, op_flags = 0;
1004 		int replace_only = 0;
1005 		struct bio *bi, *rbi;
1006 		struct md_rdev *rdev, *rrdev = NULL;
1007 
1008 		sh = head_sh;
1009 		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
1010 			op = REQ_OP_WRITE;
1011 			if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
1012 				op_flags = REQ_FUA;
1013 			if (test_bit(R5_Discard, &sh->dev[i].flags))
1014 				op = REQ_OP_DISCARD;
1015 		} else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
1016 			op = REQ_OP_READ;
1017 		else if (test_and_clear_bit(R5_WantReplace,
1018 					    &sh->dev[i].flags)) {
1019 			op = REQ_OP_WRITE;
1020 			replace_only = 1;
1021 		} else
1022 			continue;
1023 		if (test_and_clear_bit(R5_SyncIO, &sh->dev[i].flags))
1024 			op_flags |= REQ_SYNC;
1025 
1026 again:
1027 		bi = &sh->dev[i].req;
1028 		rbi = &sh->dev[i].rreq; /* For writing to replacement */
1029 
1030 		rcu_read_lock();
1031 		rrdev = rcu_dereference(conf->disks[i].replacement);
1032 		smp_mb(); /* Ensure that if rrdev is NULL, rdev won't be */
1033 		rdev = rcu_dereference(conf->disks[i].rdev);
1034 		if (!rdev) {
1035 			rdev = rrdev;
1036 			rrdev = NULL;
1037 		}
1038 		if (op_is_write(op)) {
1039 			if (replace_only)
1040 				rdev = NULL;
1041 			if (rdev == rrdev)
1042 				/* We raced and saw duplicates */
1043 				rrdev = NULL;
1044 		} else {
1045 			if (test_bit(R5_ReadRepl, &head_sh->dev[i].flags) && rrdev)
1046 				rdev = rrdev;
1047 			rrdev = NULL;
1048 		}
1049 
1050 		if (rdev && test_bit(Faulty, &rdev->flags))
1051 			rdev = NULL;
1052 		if (rdev)
1053 			atomic_inc(&rdev->nr_pending);
1054 		if (rrdev && test_bit(Faulty, &rrdev->flags))
1055 			rrdev = NULL;
1056 		if (rrdev)
1057 			atomic_inc(&rrdev->nr_pending);
1058 		rcu_read_unlock();
1059 
1060 		/* We have already checked bad blocks for reads.  Now
1061 		 * need to check for writes.  We never accept write errors
1062 		 * on the replacement, so we don't to check rrdev.
1063 		 */
1064 		while (op_is_write(op) && rdev &&
1065 		       test_bit(WriteErrorSeen, &rdev->flags)) {
1066 			sector_t first_bad;
1067 			int bad_sectors;
1068 			int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
1069 					      &first_bad, &bad_sectors);
1070 			if (!bad)
1071 				break;
1072 
1073 			if (bad < 0) {
1074 				set_bit(BlockedBadBlocks, &rdev->flags);
1075 				if (!conf->mddev->external &&
1076 				    conf->mddev->sb_flags) {
1077 					/* It is very unlikely, but we might
1078 					 * still need to write out the
1079 					 * bad block log - better give it
1080 					 * a chance*/
1081 					md_check_recovery(conf->mddev);
1082 				}
1083 				/*
1084 				 * Because md_wait_for_blocked_rdev
1085 				 * will dec nr_pending, we must
1086 				 * increment it first.
1087 				 */
1088 				atomic_inc(&rdev->nr_pending);
1089 				md_wait_for_blocked_rdev(rdev, conf->mddev);
1090 			} else {
1091 				/* Acknowledged bad block - skip the write */
1092 				rdev_dec_pending(rdev, conf->mddev);
1093 				rdev = NULL;
1094 			}
1095 		}
1096 
1097 		if (rdev) {
1098 			if (s->syncing || s->expanding || s->expanded
1099 			    || s->replacing)
1100 				md_sync_acct(rdev->bdev, STRIPE_SECTORS);
1101 
1102 			set_bit(STRIPE_IO_STARTED, &sh->state);
1103 
1104 			bio_set_dev(bi, rdev->bdev);
1105 			bio_set_op_attrs(bi, op, op_flags);
1106 			bi->bi_end_io = op_is_write(op)
1107 				? raid5_end_write_request
1108 				: raid5_end_read_request;
1109 			bi->bi_private = sh;
1110 
1111 			pr_debug("%s: for %llu schedule op %d on disc %d\n",
1112 				__func__, (unsigned long long)sh->sector,
1113 				bi->bi_opf, i);
1114 			atomic_inc(&sh->count);
1115 			if (sh != head_sh)
1116 				atomic_inc(&head_sh->count);
1117 			if (use_new_offset(conf, sh))
1118 				bi->bi_iter.bi_sector = (sh->sector
1119 						 + rdev->new_data_offset);
1120 			else
1121 				bi->bi_iter.bi_sector = (sh->sector
1122 						 + rdev->data_offset);
1123 			if (test_bit(R5_ReadNoMerge, &head_sh->dev[i].flags))
1124 				bi->bi_opf |= REQ_NOMERGE;
1125 
1126 			if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1127 				WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1128 
1129 			if (!op_is_write(op) &&
1130 			    test_bit(R5_InJournal, &sh->dev[i].flags))
1131 				/*
1132 				 * issuing read for a page in journal, this
1133 				 * must be preparing for prexor in rmw; read
1134 				 * the data into orig_page
1135 				 */
1136 				sh->dev[i].vec.bv_page = sh->dev[i].orig_page;
1137 			else
1138 				sh->dev[i].vec.bv_page = sh->dev[i].page;
1139 			bi->bi_vcnt = 1;
1140 			bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1141 			bi->bi_io_vec[0].bv_offset = 0;
1142 			bi->bi_iter.bi_size = STRIPE_SIZE;
1143 			bi->bi_write_hint = sh->dev[i].write_hint;
1144 			if (!rrdev)
1145 				sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1146 			/*
1147 			 * If this is discard request, set bi_vcnt 0. We don't
1148 			 * want to confuse SCSI because SCSI will replace payload
1149 			 */
1150 			if (op == REQ_OP_DISCARD)
1151 				bi->bi_vcnt = 0;
1152 			if (rrdev)
1153 				set_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags);
1154 
1155 			if (conf->mddev->gendisk)
1156 				trace_block_bio_remap(bi->bi_disk->queue,
1157 						      bi, disk_devt(conf->mddev->gendisk),
1158 						      sh->dev[i].sector);
1159 			if (should_defer && op_is_write(op))
1160 				bio_list_add(&pending_bios, bi);
1161 			else
1162 				generic_make_request(bi);
1163 		}
1164 		if (rrdev) {
1165 			if (s->syncing || s->expanding || s->expanded
1166 			    || s->replacing)
1167 				md_sync_acct(rrdev->bdev, STRIPE_SECTORS);
1168 
1169 			set_bit(STRIPE_IO_STARTED, &sh->state);
1170 
1171 			bio_set_dev(rbi, rrdev->bdev);
1172 			bio_set_op_attrs(rbi, op, op_flags);
1173 			BUG_ON(!op_is_write(op));
1174 			rbi->bi_end_io = raid5_end_write_request;
1175 			rbi->bi_private = sh;
1176 
1177 			pr_debug("%s: for %llu schedule op %d on "
1178 				 "replacement disc %d\n",
1179 				__func__, (unsigned long long)sh->sector,
1180 				rbi->bi_opf, i);
1181 			atomic_inc(&sh->count);
1182 			if (sh != head_sh)
1183 				atomic_inc(&head_sh->count);
1184 			if (use_new_offset(conf, sh))
1185 				rbi->bi_iter.bi_sector = (sh->sector
1186 						  + rrdev->new_data_offset);
1187 			else
1188 				rbi->bi_iter.bi_sector = (sh->sector
1189 						  + rrdev->data_offset);
1190 			if (test_bit(R5_SkipCopy, &sh->dev[i].flags))
1191 				WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
1192 			sh->dev[i].rvec.bv_page = sh->dev[i].page;
1193 			rbi->bi_vcnt = 1;
1194 			rbi->bi_io_vec[0].bv_len = STRIPE_SIZE;
1195 			rbi->bi_io_vec[0].bv_offset = 0;
1196 			rbi->bi_iter.bi_size = STRIPE_SIZE;
1197 			rbi->bi_write_hint = sh->dev[i].write_hint;
1198 			sh->dev[i].write_hint = RWF_WRITE_LIFE_NOT_SET;
1199 			/*
1200 			 * If this is discard request, set bi_vcnt 0. We don't
1201 			 * want to confuse SCSI because SCSI will replace payload
1202 			 */
1203 			if (op == REQ_OP_DISCARD)
1204 				rbi->bi_vcnt = 0;
1205 			if (conf->mddev->gendisk)
1206 				trace_block_bio_remap(rbi->bi_disk->queue,
1207 						      rbi, disk_devt(conf->mddev->gendisk),
1208 						      sh->dev[i].sector);
1209 			if (should_defer && op_is_write(op))
1210 				bio_list_add(&pending_bios, rbi);
1211 			else
1212 				generic_make_request(rbi);
1213 		}
1214 		if (!rdev && !rrdev) {
1215 			if (op_is_write(op))
1216 				set_bit(STRIPE_DEGRADED, &sh->state);
1217 			pr_debug("skip op %d on disc %d for sector %llu\n",
1218 				bi->bi_opf, i, (unsigned long long)sh->sector);
1219 			clear_bit(R5_LOCKED, &sh->dev[i].flags);
1220 			set_bit(STRIPE_HANDLE, &sh->state);
1221 		}
1222 
1223 		if (!head_sh->batch_head)
1224 			continue;
1225 		sh = list_first_entry(&sh->batch_list, struct stripe_head,
1226 				      batch_list);
1227 		if (sh != head_sh)
1228 			goto again;
1229 	}
1230 
1231 	if (should_defer && !bio_list_empty(&pending_bios))
1232 		defer_issue_bios(conf, head_sh->sector, &pending_bios);
1233 }
1234 
1235 static struct dma_async_tx_descriptor *
1236 async_copy_data(int frombio, struct bio *bio, struct page **page,
1237 	sector_t sector, struct dma_async_tx_descriptor *tx,
1238 	struct stripe_head *sh, int no_skipcopy)
1239 {
1240 	struct bio_vec bvl;
1241 	struct bvec_iter iter;
1242 	struct page *bio_page;
1243 	int page_offset;
1244 	struct async_submit_ctl submit;
1245 	enum async_tx_flags flags = 0;
1246 
1247 	if (bio->bi_iter.bi_sector >= sector)
1248 		page_offset = (signed)(bio->bi_iter.bi_sector - sector) * 512;
1249 	else
1250 		page_offset = (signed)(sector - bio->bi_iter.bi_sector) * -512;
1251 
1252 	if (frombio)
1253 		flags |= ASYNC_TX_FENCE;
1254 	init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
1255 
1256 	bio_for_each_segment(bvl, bio, iter) {
1257 		int len = bvl.bv_len;
1258 		int clen;
1259 		int b_offset = 0;
1260 
1261 		if (page_offset < 0) {
1262 			b_offset = -page_offset;
1263 			page_offset += b_offset;
1264 			len -= b_offset;
1265 		}
1266 
1267 		if (len > 0 && page_offset + len > STRIPE_SIZE)
1268 			clen = STRIPE_SIZE - page_offset;
1269 		else
1270 			clen = len;
1271 
1272 		if (clen > 0) {
1273 			b_offset += bvl.bv_offset;
1274 			bio_page = bvl.bv_page;
1275 			if (frombio) {
1276 				if (sh->raid_conf->skip_copy &&
1277 				    b_offset == 0 && page_offset == 0 &&
1278 				    clen == STRIPE_SIZE &&
1279 				    !no_skipcopy)
1280 					*page = bio_page;
1281 				else
1282 					tx = async_memcpy(*page, bio_page, page_offset,
1283 						  b_offset, clen, &submit);
1284 			} else
1285 				tx = async_memcpy(bio_page, *page, b_offset,
1286 						  page_offset, clen, &submit);
1287 		}
1288 		/* chain the operations */
1289 		submit.depend_tx = tx;
1290 
1291 		if (clen < len) /* hit end of page */
1292 			break;
1293 		page_offset +=  len;
1294 	}
1295 
1296 	return tx;
1297 }
1298 
1299 static void ops_complete_biofill(void *stripe_head_ref)
1300 {
1301 	struct stripe_head *sh = stripe_head_ref;
1302 	int i;
1303 
1304 	pr_debug("%s: stripe %llu\n", __func__,
1305 		(unsigned long long)sh->sector);
1306 
1307 	/* clear completed biofills */
1308 	for (i = sh->disks; i--; ) {
1309 		struct r5dev *dev = &sh->dev[i];
1310 
1311 		/* acknowledge completion of a biofill operation */
1312 		/* and check if we need to reply to a read request,
1313 		 * new R5_Wantfill requests are held off until
1314 		 * !STRIPE_BIOFILL_RUN
1315 		 */
1316 		if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
1317 			struct bio *rbi, *rbi2;
1318 
1319 			BUG_ON(!dev->read);
1320 			rbi = dev->read;
1321 			dev->read = NULL;
1322 			while (rbi && rbi->bi_iter.bi_sector <
1323 				dev->sector + STRIPE_SECTORS) {
1324 				rbi2 = r5_next_bio(rbi, dev->sector);
1325 				bio_endio(rbi);
1326 				rbi = rbi2;
1327 			}
1328 		}
1329 	}
1330 	clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
1331 
1332 	set_bit(STRIPE_HANDLE, &sh->state);
1333 	raid5_release_stripe(sh);
1334 }
1335 
1336 static void ops_run_biofill(struct stripe_head *sh)
1337 {
1338 	struct dma_async_tx_descriptor *tx = NULL;
1339 	struct async_submit_ctl submit;
1340 	int i;
1341 
1342 	BUG_ON(sh->batch_head);
1343 	pr_debug("%s: stripe %llu\n", __func__,
1344 		(unsigned long long)sh->sector);
1345 
1346 	for (i = sh->disks; i--; ) {
1347 		struct r5dev *dev = &sh->dev[i];
1348 		if (test_bit(R5_Wantfill, &dev->flags)) {
1349 			struct bio *rbi;
1350 			spin_lock_irq(&sh->stripe_lock);
1351 			dev->read = rbi = dev->toread;
1352 			dev->toread = NULL;
1353 			spin_unlock_irq(&sh->stripe_lock);
1354 			while (rbi && rbi->bi_iter.bi_sector <
1355 				dev->sector + STRIPE_SECTORS) {
1356 				tx = async_copy_data(0, rbi, &dev->page,
1357 						     dev->sector, tx, sh, 0);
1358 				rbi = r5_next_bio(rbi, dev->sector);
1359 			}
1360 		}
1361 	}
1362 
1363 	atomic_inc(&sh->count);
1364 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
1365 	async_trigger_callback(&submit);
1366 }
1367 
1368 static void mark_target_uptodate(struct stripe_head *sh, int target)
1369 {
1370 	struct r5dev *tgt;
1371 
1372 	if (target < 0)
1373 		return;
1374 
1375 	tgt = &sh->dev[target];
1376 	set_bit(R5_UPTODATE, &tgt->flags);
1377 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1378 	clear_bit(R5_Wantcompute, &tgt->flags);
1379 }
1380 
1381 static void ops_complete_compute(void *stripe_head_ref)
1382 {
1383 	struct stripe_head *sh = stripe_head_ref;
1384 
1385 	pr_debug("%s: stripe %llu\n", __func__,
1386 		(unsigned long long)sh->sector);
1387 
1388 	/* mark the computed target(s) as uptodate */
1389 	mark_target_uptodate(sh, sh->ops.target);
1390 	mark_target_uptodate(sh, sh->ops.target2);
1391 
1392 	clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
1393 	if (sh->check_state == check_state_compute_run)
1394 		sh->check_state = check_state_compute_result;
1395 	set_bit(STRIPE_HANDLE, &sh->state);
1396 	raid5_release_stripe(sh);
1397 }
1398 
1399 /* return a pointer to the address conversion region of the scribble buffer */
1400 static struct page **to_addr_page(struct raid5_percpu *percpu, int i)
1401 {
1402 	return percpu->scribble + i * percpu->scribble_obj_size;
1403 }
1404 
1405 /* return a pointer to the address conversion region of the scribble buffer */
1406 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
1407 				 struct raid5_percpu *percpu, int i)
1408 {
1409 	return (void *) (to_addr_page(percpu, i) + sh->disks + 2);
1410 }
1411 
1412 static struct dma_async_tx_descriptor *
1413 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
1414 {
1415 	int disks = sh->disks;
1416 	struct page **xor_srcs = to_addr_page(percpu, 0);
1417 	int target = sh->ops.target;
1418 	struct r5dev *tgt = &sh->dev[target];
1419 	struct page *xor_dest = tgt->page;
1420 	int count = 0;
1421 	struct dma_async_tx_descriptor *tx;
1422 	struct async_submit_ctl submit;
1423 	int i;
1424 
1425 	BUG_ON(sh->batch_head);
1426 
1427 	pr_debug("%s: stripe %llu block: %d\n",
1428 		__func__, (unsigned long long)sh->sector, target);
1429 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1430 
1431 	for (i = disks; i--; )
1432 		if (i != target)
1433 			xor_srcs[count++] = sh->dev[i].page;
1434 
1435 	atomic_inc(&sh->count);
1436 
1437 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
1438 			  ops_complete_compute, sh, to_addr_conv(sh, percpu, 0));
1439 	if (unlikely(count == 1))
1440 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1441 	else
1442 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1443 
1444 	return tx;
1445 }
1446 
1447 /* set_syndrome_sources - populate source buffers for gen_syndrome
1448  * @srcs - (struct page *) array of size sh->disks
1449  * @sh - stripe_head to parse
1450  *
1451  * Populates srcs in proper layout order for the stripe and returns the
1452  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
1453  * destination buffer is recorded in srcs[count] and the Q destination
1454  * is recorded in srcs[count+1]].
1455  */
1456 static int set_syndrome_sources(struct page **srcs,
1457 				struct stripe_head *sh,
1458 				int srctype)
1459 {
1460 	int disks = sh->disks;
1461 	int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
1462 	int d0_idx = raid6_d0(sh);
1463 	int count;
1464 	int i;
1465 
1466 	for (i = 0; i < disks; i++)
1467 		srcs[i] = NULL;
1468 
1469 	count = 0;
1470 	i = d0_idx;
1471 	do {
1472 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1473 		struct r5dev *dev = &sh->dev[i];
1474 
1475 		if (i == sh->qd_idx || i == sh->pd_idx ||
1476 		    (srctype == SYNDROME_SRC_ALL) ||
1477 		    (srctype == SYNDROME_SRC_WANT_DRAIN &&
1478 		     (test_bit(R5_Wantdrain, &dev->flags) ||
1479 		      test_bit(R5_InJournal, &dev->flags))) ||
1480 		    (srctype == SYNDROME_SRC_WRITTEN &&
1481 		     (dev->written ||
1482 		      test_bit(R5_InJournal, &dev->flags)))) {
1483 			if (test_bit(R5_InJournal, &dev->flags))
1484 				srcs[slot] = sh->dev[i].orig_page;
1485 			else
1486 				srcs[slot] = sh->dev[i].page;
1487 		}
1488 		i = raid6_next_disk(i, disks);
1489 	} while (i != d0_idx);
1490 
1491 	return syndrome_disks;
1492 }
1493 
1494 static struct dma_async_tx_descriptor *
1495 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
1496 {
1497 	int disks = sh->disks;
1498 	struct page **blocks = to_addr_page(percpu, 0);
1499 	int target;
1500 	int qd_idx = sh->qd_idx;
1501 	struct dma_async_tx_descriptor *tx;
1502 	struct async_submit_ctl submit;
1503 	struct r5dev *tgt;
1504 	struct page *dest;
1505 	int i;
1506 	int count;
1507 
1508 	BUG_ON(sh->batch_head);
1509 	if (sh->ops.target < 0)
1510 		target = sh->ops.target2;
1511 	else if (sh->ops.target2 < 0)
1512 		target = sh->ops.target;
1513 	else
1514 		/* we should only have one valid target */
1515 		BUG();
1516 	BUG_ON(target < 0);
1517 	pr_debug("%s: stripe %llu block: %d\n",
1518 		__func__, (unsigned long long)sh->sector, target);
1519 
1520 	tgt = &sh->dev[target];
1521 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1522 	dest = tgt->page;
1523 
1524 	atomic_inc(&sh->count);
1525 
1526 	if (target == qd_idx) {
1527 		count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1528 		blocks[count] = NULL; /* regenerating p is not necessary */
1529 		BUG_ON(blocks[count+1] != dest); /* q should already be set */
1530 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1531 				  ops_complete_compute, sh,
1532 				  to_addr_conv(sh, percpu, 0));
1533 		tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
1534 	} else {
1535 		/* Compute any data- or p-drive using XOR */
1536 		count = 0;
1537 		for (i = disks; i-- ; ) {
1538 			if (i == target || i == qd_idx)
1539 				continue;
1540 			blocks[count++] = sh->dev[i].page;
1541 		}
1542 
1543 		init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1544 				  NULL, ops_complete_compute, sh,
1545 				  to_addr_conv(sh, percpu, 0));
1546 		tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
1547 	}
1548 
1549 	return tx;
1550 }
1551 
1552 static struct dma_async_tx_descriptor *
1553 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
1554 {
1555 	int i, count, disks = sh->disks;
1556 	int syndrome_disks = sh->ddf_layout ? disks : disks-2;
1557 	int d0_idx = raid6_d0(sh);
1558 	int faila = -1, failb = -1;
1559 	int target = sh->ops.target;
1560 	int target2 = sh->ops.target2;
1561 	struct r5dev *tgt = &sh->dev[target];
1562 	struct r5dev *tgt2 = &sh->dev[target2];
1563 	struct dma_async_tx_descriptor *tx;
1564 	struct page **blocks = to_addr_page(percpu, 0);
1565 	struct async_submit_ctl submit;
1566 
1567 	BUG_ON(sh->batch_head);
1568 	pr_debug("%s: stripe %llu block1: %d block2: %d\n",
1569 		 __func__, (unsigned long long)sh->sector, target, target2);
1570 	BUG_ON(target < 0 || target2 < 0);
1571 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
1572 	BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
1573 
1574 	/* we need to open-code set_syndrome_sources to handle the
1575 	 * slot number conversion for 'faila' and 'failb'
1576 	 */
1577 	for (i = 0; i < disks ; i++)
1578 		blocks[i] = NULL;
1579 	count = 0;
1580 	i = d0_idx;
1581 	do {
1582 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
1583 
1584 		blocks[slot] = sh->dev[i].page;
1585 
1586 		if (i == target)
1587 			faila = slot;
1588 		if (i == target2)
1589 			failb = slot;
1590 		i = raid6_next_disk(i, disks);
1591 	} while (i != d0_idx);
1592 
1593 	BUG_ON(faila == failb);
1594 	if (failb < faila)
1595 		swap(faila, failb);
1596 	pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
1597 		 __func__, (unsigned long long)sh->sector, faila, failb);
1598 
1599 	atomic_inc(&sh->count);
1600 
1601 	if (failb == syndrome_disks+1) {
1602 		/* Q disk is one of the missing disks */
1603 		if (faila == syndrome_disks) {
1604 			/* Missing P+Q, just recompute */
1605 			init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1606 					  ops_complete_compute, sh,
1607 					  to_addr_conv(sh, percpu, 0));
1608 			return async_gen_syndrome(blocks, 0, syndrome_disks+2,
1609 						  STRIPE_SIZE, &submit);
1610 		} else {
1611 			struct page *dest;
1612 			int data_target;
1613 			int qd_idx = sh->qd_idx;
1614 
1615 			/* Missing D+Q: recompute D from P, then recompute Q */
1616 			if (target == qd_idx)
1617 				data_target = target2;
1618 			else
1619 				data_target = target;
1620 
1621 			count = 0;
1622 			for (i = disks; i-- ; ) {
1623 				if (i == data_target || i == qd_idx)
1624 					continue;
1625 				blocks[count++] = sh->dev[i].page;
1626 			}
1627 			dest = sh->dev[data_target].page;
1628 			init_async_submit(&submit,
1629 					  ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
1630 					  NULL, NULL, NULL,
1631 					  to_addr_conv(sh, percpu, 0));
1632 			tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
1633 				       &submit);
1634 
1635 			count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_ALL);
1636 			init_async_submit(&submit, ASYNC_TX_FENCE, tx,
1637 					  ops_complete_compute, sh,
1638 					  to_addr_conv(sh, percpu, 0));
1639 			return async_gen_syndrome(blocks, 0, count+2,
1640 						  STRIPE_SIZE, &submit);
1641 		}
1642 	} else {
1643 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
1644 				  ops_complete_compute, sh,
1645 				  to_addr_conv(sh, percpu, 0));
1646 		if (failb == syndrome_disks) {
1647 			/* We're missing D+P. */
1648 			return async_raid6_datap_recov(syndrome_disks+2,
1649 						       STRIPE_SIZE, faila,
1650 						       blocks, &submit);
1651 		} else {
1652 			/* We're missing D+D. */
1653 			return async_raid6_2data_recov(syndrome_disks+2,
1654 						       STRIPE_SIZE, faila, failb,
1655 						       blocks, &submit);
1656 		}
1657 	}
1658 }
1659 
1660 static void ops_complete_prexor(void *stripe_head_ref)
1661 {
1662 	struct stripe_head *sh = stripe_head_ref;
1663 
1664 	pr_debug("%s: stripe %llu\n", __func__,
1665 		(unsigned long long)sh->sector);
1666 
1667 	if (r5c_is_writeback(sh->raid_conf->log))
1668 		/*
1669 		 * raid5-cache write back uses orig_page during prexor.
1670 		 * After prexor, it is time to free orig_page
1671 		 */
1672 		r5c_release_extra_page(sh);
1673 }
1674 
1675 static struct dma_async_tx_descriptor *
1676 ops_run_prexor5(struct stripe_head *sh, struct raid5_percpu *percpu,
1677 		struct dma_async_tx_descriptor *tx)
1678 {
1679 	int disks = sh->disks;
1680 	struct page **xor_srcs = to_addr_page(percpu, 0);
1681 	int count = 0, pd_idx = sh->pd_idx, i;
1682 	struct async_submit_ctl submit;
1683 
1684 	/* existing parity data subtracted */
1685 	struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1686 
1687 	BUG_ON(sh->batch_head);
1688 	pr_debug("%s: stripe %llu\n", __func__,
1689 		(unsigned long long)sh->sector);
1690 
1691 	for (i = disks; i--; ) {
1692 		struct r5dev *dev = &sh->dev[i];
1693 		/* Only process blocks that are known to be uptodate */
1694 		if (test_bit(R5_InJournal, &dev->flags))
1695 			xor_srcs[count++] = dev->orig_page;
1696 		else if (test_bit(R5_Wantdrain, &dev->flags))
1697 			xor_srcs[count++] = dev->page;
1698 	}
1699 
1700 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1701 			  ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1702 	tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1703 
1704 	return tx;
1705 }
1706 
1707 static struct dma_async_tx_descriptor *
1708 ops_run_prexor6(struct stripe_head *sh, struct raid5_percpu *percpu,
1709 		struct dma_async_tx_descriptor *tx)
1710 {
1711 	struct page **blocks = to_addr_page(percpu, 0);
1712 	int count;
1713 	struct async_submit_ctl submit;
1714 
1715 	pr_debug("%s: stripe %llu\n", __func__,
1716 		(unsigned long long)sh->sector);
1717 
1718 	count = set_syndrome_sources(blocks, sh, SYNDROME_SRC_WANT_DRAIN);
1719 
1720 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_PQ_XOR_DST, tx,
1721 			  ops_complete_prexor, sh, to_addr_conv(sh, percpu, 0));
1722 	tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1723 
1724 	return tx;
1725 }
1726 
1727 static struct dma_async_tx_descriptor *
1728 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1729 {
1730 	struct r5conf *conf = sh->raid_conf;
1731 	int disks = sh->disks;
1732 	int i;
1733 	struct stripe_head *head_sh = sh;
1734 
1735 	pr_debug("%s: stripe %llu\n", __func__,
1736 		(unsigned long long)sh->sector);
1737 
1738 	for (i = disks; i--; ) {
1739 		struct r5dev *dev;
1740 		struct bio *chosen;
1741 
1742 		sh = head_sh;
1743 		if (test_and_clear_bit(R5_Wantdrain, &head_sh->dev[i].flags)) {
1744 			struct bio *wbi;
1745 
1746 again:
1747 			dev = &sh->dev[i];
1748 			/*
1749 			 * clear R5_InJournal, so when rewriting a page in
1750 			 * journal, it is not skipped by r5l_log_stripe()
1751 			 */
1752 			clear_bit(R5_InJournal, &dev->flags);
1753 			spin_lock_irq(&sh->stripe_lock);
1754 			chosen = dev->towrite;
1755 			dev->towrite = NULL;
1756 			sh->overwrite_disks = 0;
1757 			BUG_ON(dev->written);
1758 			wbi = dev->written = chosen;
1759 			spin_unlock_irq(&sh->stripe_lock);
1760 			WARN_ON(dev->page != dev->orig_page);
1761 
1762 			while (wbi && wbi->bi_iter.bi_sector <
1763 				dev->sector + STRIPE_SECTORS) {
1764 				if (wbi->bi_opf & REQ_FUA)
1765 					set_bit(R5_WantFUA, &dev->flags);
1766 				if (wbi->bi_opf & REQ_SYNC)
1767 					set_bit(R5_SyncIO, &dev->flags);
1768 				if (bio_op(wbi) == REQ_OP_DISCARD)
1769 					set_bit(R5_Discard, &dev->flags);
1770 				else {
1771 					tx = async_copy_data(1, wbi, &dev->page,
1772 							     dev->sector, tx, sh,
1773 							     r5c_is_writeback(conf->log));
1774 					if (dev->page != dev->orig_page &&
1775 					    !r5c_is_writeback(conf->log)) {
1776 						set_bit(R5_SkipCopy, &dev->flags);
1777 						clear_bit(R5_UPTODATE, &dev->flags);
1778 						clear_bit(R5_OVERWRITE, &dev->flags);
1779 					}
1780 				}
1781 				wbi = r5_next_bio(wbi, dev->sector);
1782 			}
1783 
1784 			if (head_sh->batch_head) {
1785 				sh = list_first_entry(&sh->batch_list,
1786 						      struct stripe_head,
1787 						      batch_list);
1788 				if (sh == head_sh)
1789 					continue;
1790 				goto again;
1791 			}
1792 		}
1793 	}
1794 
1795 	return tx;
1796 }
1797 
1798 static void ops_complete_reconstruct(void *stripe_head_ref)
1799 {
1800 	struct stripe_head *sh = stripe_head_ref;
1801 	int disks = sh->disks;
1802 	int pd_idx = sh->pd_idx;
1803 	int qd_idx = sh->qd_idx;
1804 	int i;
1805 	bool fua = false, sync = false, discard = false;
1806 
1807 	pr_debug("%s: stripe %llu\n", __func__,
1808 		(unsigned long long)sh->sector);
1809 
1810 	for (i = disks; i--; ) {
1811 		fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1812 		sync |= test_bit(R5_SyncIO, &sh->dev[i].flags);
1813 		discard |= test_bit(R5_Discard, &sh->dev[i].flags);
1814 	}
1815 
1816 	for (i = disks; i--; ) {
1817 		struct r5dev *dev = &sh->dev[i];
1818 
1819 		if (dev->written || i == pd_idx || i == qd_idx) {
1820 			if (!discard && !test_bit(R5_SkipCopy, &dev->flags)) {
1821 				set_bit(R5_UPTODATE, &dev->flags);
1822 				if (test_bit(STRIPE_EXPAND_READY, &sh->state))
1823 					set_bit(R5_Expanded, &dev->flags);
1824 			}
1825 			if (fua)
1826 				set_bit(R5_WantFUA, &dev->flags);
1827 			if (sync)
1828 				set_bit(R5_SyncIO, &dev->flags);
1829 		}
1830 	}
1831 
1832 	if (sh->reconstruct_state == reconstruct_state_drain_run)
1833 		sh->reconstruct_state = reconstruct_state_drain_result;
1834 	else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1835 		sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1836 	else {
1837 		BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1838 		sh->reconstruct_state = reconstruct_state_result;
1839 	}
1840 
1841 	set_bit(STRIPE_HANDLE, &sh->state);
1842 	raid5_release_stripe(sh);
1843 }
1844 
1845 static void
1846 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1847 		     struct dma_async_tx_descriptor *tx)
1848 {
1849 	int disks = sh->disks;
1850 	struct page **xor_srcs;
1851 	struct async_submit_ctl submit;
1852 	int count, pd_idx = sh->pd_idx, i;
1853 	struct page *xor_dest;
1854 	int prexor = 0;
1855 	unsigned long flags;
1856 	int j = 0;
1857 	struct stripe_head *head_sh = sh;
1858 	int last_stripe;
1859 
1860 	pr_debug("%s: stripe %llu\n", __func__,
1861 		(unsigned long long)sh->sector);
1862 
1863 	for (i = 0; i < sh->disks; i++) {
1864 		if (pd_idx == i)
1865 			continue;
1866 		if (!test_bit(R5_Discard, &sh->dev[i].flags))
1867 			break;
1868 	}
1869 	if (i >= sh->disks) {
1870 		atomic_inc(&sh->count);
1871 		set_bit(R5_Discard, &sh->dev[pd_idx].flags);
1872 		ops_complete_reconstruct(sh);
1873 		return;
1874 	}
1875 again:
1876 	count = 0;
1877 	xor_srcs = to_addr_page(percpu, j);
1878 	/* check if prexor is active which means only process blocks
1879 	 * that are part of a read-modify-write (written)
1880 	 */
1881 	if (head_sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1882 		prexor = 1;
1883 		xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1884 		for (i = disks; i--; ) {
1885 			struct r5dev *dev = &sh->dev[i];
1886 			if (head_sh->dev[i].written ||
1887 			    test_bit(R5_InJournal, &head_sh->dev[i].flags))
1888 				xor_srcs[count++] = dev->page;
1889 		}
1890 	} else {
1891 		xor_dest = sh->dev[pd_idx].page;
1892 		for (i = disks; i--; ) {
1893 			struct r5dev *dev = &sh->dev[i];
1894 			if (i != pd_idx)
1895 				xor_srcs[count++] = dev->page;
1896 		}
1897 	}
1898 
1899 	/* 1/ if we prexor'd then the dest is reused as a source
1900 	 * 2/ if we did not prexor then we are redoing the parity
1901 	 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1902 	 * for the synchronous xor case
1903 	 */
1904 	last_stripe = !head_sh->batch_head ||
1905 		list_first_entry(&sh->batch_list,
1906 				 struct stripe_head, batch_list) == head_sh;
1907 	if (last_stripe) {
1908 		flags = ASYNC_TX_ACK |
1909 			(prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1910 
1911 		atomic_inc(&head_sh->count);
1912 		init_async_submit(&submit, flags, tx, ops_complete_reconstruct, head_sh,
1913 				  to_addr_conv(sh, percpu, j));
1914 	} else {
1915 		flags = prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST;
1916 		init_async_submit(&submit, flags, tx, NULL, NULL,
1917 				  to_addr_conv(sh, percpu, j));
1918 	}
1919 
1920 	if (unlikely(count == 1))
1921 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1922 	else
1923 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1924 	if (!last_stripe) {
1925 		j++;
1926 		sh = list_first_entry(&sh->batch_list, struct stripe_head,
1927 				      batch_list);
1928 		goto again;
1929 	}
1930 }
1931 
1932 static void
1933 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1934 		     struct dma_async_tx_descriptor *tx)
1935 {
1936 	struct async_submit_ctl submit;
1937 	struct page **blocks;
1938 	int count, i, j = 0;
1939 	struct stripe_head *head_sh = sh;
1940 	int last_stripe;
1941 	int synflags;
1942 	unsigned long txflags;
1943 
1944 	pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1945 
1946 	for (i = 0; i < sh->disks; i++) {
1947 		if (sh->pd_idx == i || sh->qd_idx == i)
1948 			continue;
1949 		if (!test_bit(R5_Discard, &sh->dev[i].flags))
1950 			break;
1951 	}
1952 	if (i >= sh->disks) {
1953 		atomic_inc(&sh->count);
1954 		set_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
1955 		set_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
1956 		ops_complete_reconstruct(sh);
1957 		return;
1958 	}
1959 
1960 again:
1961 	blocks = to_addr_page(percpu, j);
1962 
1963 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1964 		synflags = SYNDROME_SRC_WRITTEN;
1965 		txflags = ASYNC_TX_ACK | ASYNC_TX_PQ_XOR_DST;
1966 	} else {
1967 		synflags = SYNDROME_SRC_ALL;
1968 		txflags = ASYNC_TX_ACK;
1969 	}
1970 
1971 	count = set_syndrome_sources(blocks, sh, synflags);
1972 	last_stripe = !head_sh->batch_head ||
1973 		list_first_entry(&sh->batch_list,
1974 				 struct stripe_head, batch_list) == head_sh;
1975 
1976 	if (last_stripe) {
1977 		atomic_inc(&head_sh->count);
1978 		init_async_submit(&submit, txflags, tx, ops_complete_reconstruct,
1979 				  head_sh, to_addr_conv(sh, percpu, j));
1980 	} else
1981 		init_async_submit(&submit, 0, tx, NULL, NULL,
1982 				  to_addr_conv(sh, percpu, j));
1983 	tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1984 	if (!last_stripe) {
1985 		j++;
1986 		sh = list_first_entry(&sh->batch_list, struct stripe_head,
1987 				      batch_list);
1988 		goto again;
1989 	}
1990 }
1991 
1992 static void ops_complete_check(void *stripe_head_ref)
1993 {
1994 	struct stripe_head *sh = stripe_head_ref;
1995 
1996 	pr_debug("%s: stripe %llu\n", __func__,
1997 		(unsigned long long)sh->sector);
1998 
1999 	sh->check_state = check_state_check_result;
2000 	set_bit(STRIPE_HANDLE, &sh->state);
2001 	raid5_release_stripe(sh);
2002 }
2003 
2004 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
2005 {
2006 	int disks = sh->disks;
2007 	int pd_idx = sh->pd_idx;
2008 	int qd_idx = sh->qd_idx;
2009 	struct page *xor_dest;
2010 	struct page **xor_srcs = to_addr_page(percpu, 0);
2011 	struct dma_async_tx_descriptor *tx;
2012 	struct async_submit_ctl submit;
2013 	int count;
2014 	int i;
2015 
2016 	pr_debug("%s: stripe %llu\n", __func__,
2017 		(unsigned long long)sh->sector);
2018 
2019 	BUG_ON(sh->batch_head);
2020 	count = 0;
2021 	xor_dest = sh->dev[pd_idx].page;
2022 	xor_srcs[count++] = xor_dest;
2023 	for (i = disks; i--; ) {
2024 		if (i == pd_idx || i == qd_idx)
2025 			continue;
2026 		xor_srcs[count++] = sh->dev[i].page;
2027 	}
2028 
2029 	init_async_submit(&submit, 0, NULL, NULL, NULL,
2030 			  to_addr_conv(sh, percpu, 0));
2031 	tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
2032 			   &sh->ops.zero_sum_result, &submit);
2033 
2034 	atomic_inc(&sh->count);
2035 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
2036 	tx = async_trigger_callback(&submit);
2037 }
2038 
2039 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
2040 {
2041 	struct page **srcs = to_addr_page(percpu, 0);
2042 	struct async_submit_ctl submit;
2043 	int count;
2044 
2045 	pr_debug("%s: stripe %llu checkp: %d\n", __func__,
2046 		(unsigned long long)sh->sector, checkp);
2047 
2048 	BUG_ON(sh->batch_head);
2049 	count = set_syndrome_sources(srcs, sh, SYNDROME_SRC_ALL);
2050 	if (!checkp)
2051 		srcs[count] = NULL;
2052 
2053 	atomic_inc(&sh->count);
2054 	init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
2055 			  sh, to_addr_conv(sh, percpu, 0));
2056 	async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
2057 			   &sh->ops.zero_sum_result, percpu->spare_page, &submit);
2058 }
2059 
2060 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
2061 {
2062 	int overlap_clear = 0, i, disks = sh->disks;
2063 	struct dma_async_tx_descriptor *tx = NULL;
2064 	struct r5conf *conf = sh->raid_conf;
2065 	int level = conf->level;
2066 	struct raid5_percpu *percpu;
2067 	unsigned long cpu;
2068 
2069 	cpu = get_cpu();
2070 	percpu = per_cpu_ptr(conf->percpu, cpu);
2071 	if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
2072 		ops_run_biofill(sh);
2073 		overlap_clear++;
2074 	}
2075 
2076 	if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
2077 		if (level < 6)
2078 			tx = ops_run_compute5(sh, percpu);
2079 		else {
2080 			if (sh->ops.target2 < 0 || sh->ops.target < 0)
2081 				tx = ops_run_compute6_1(sh, percpu);
2082 			else
2083 				tx = ops_run_compute6_2(sh, percpu);
2084 		}
2085 		/* terminate the chain if reconstruct is not set to be run */
2086 		if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
2087 			async_tx_ack(tx);
2088 	}
2089 
2090 	if (test_bit(STRIPE_OP_PREXOR, &ops_request)) {
2091 		if (level < 6)
2092 			tx = ops_run_prexor5(sh, percpu, tx);
2093 		else
2094 			tx = ops_run_prexor6(sh, percpu, tx);
2095 	}
2096 
2097 	if (test_bit(STRIPE_OP_PARTIAL_PARITY, &ops_request))
2098 		tx = ops_run_partial_parity(sh, percpu, tx);
2099 
2100 	if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
2101 		tx = ops_run_biodrain(sh, tx);
2102 		overlap_clear++;
2103 	}
2104 
2105 	if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
2106 		if (level < 6)
2107 			ops_run_reconstruct5(sh, percpu, tx);
2108 		else
2109 			ops_run_reconstruct6(sh, percpu, tx);
2110 	}
2111 
2112 	if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
2113 		if (sh->check_state == check_state_run)
2114 			ops_run_check_p(sh, percpu);
2115 		else if (sh->check_state == check_state_run_q)
2116 			ops_run_check_pq(sh, percpu, 0);
2117 		else if (sh->check_state == check_state_run_pq)
2118 			ops_run_check_pq(sh, percpu, 1);
2119 		else
2120 			BUG();
2121 	}
2122 
2123 	if (overlap_clear && !sh->batch_head)
2124 		for (i = disks; i--; ) {
2125 			struct r5dev *dev = &sh->dev[i];
2126 			if (test_and_clear_bit(R5_Overlap, &dev->flags))
2127 				wake_up(&sh->raid_conf->wait_for_overlap);
2128 		}
2129 	put_cpu();
2130 }
2131 
2132 static void free_stripe(struct kmem_cache *sc, struct stripe_head *sh)
2133 {
2134 	if (sh->ppl_page)
2135 		__free_page(sh->ppl_page);
2136 	kmem_cache_free(sc, sh);
2137 }
2138 
2139 static struct stripe_head *alloc_stripe(struct kmem_cache *sc, gfp_t gfp,
2140 	int disks, struct r5conf *conf)
2141 {
2142 	struct stripe_head *sh;
2143 	int i;
2144 
2145 	sh = kmem_cache_zalloc(sc, gfp);
2146 	if (sh) {
2147 		spin_lock_init(&sh->stripe_lock);
2148 		spin_lock_init(&sh->batch_lock);
2149 		INIT_LIST_HEAD(&sh->batch_list);
2150 		INIT_LIST_HEAD(&sh->lru);
2151 		INIT_LIST_HEAD(&sh->r5c);
2152 		INIT_LIST_HEAD(&sh->log_list);
2153 		atomic_set(&sh->count, 1);
2154 		sh->raid_conf = conf;
2155 		sh->log_start = MaxSector;
2156 		for (i = 0; i < disks; i++) {
2157 			struct r5dev *dev = &sh->dev[i];
2158 
2159 			bio_init(&dev->req, &dev->vec, 1);
2160 			bio_init(&dev->rreq, &dev->rvec, 1);
2161 		}
2162 
2163 		if (raid5_has_ppl(conf)) {
2164 			sh->ppl_page = alloc_page(gfp);
2165 			if (!sh->ppl_page) {
2166 				free_stripe(sc, sh);
2167 				sh = NULL;
2168 			}
2169 		}
2170 	}
2171 	return sh;
2172 }
2173 static int grow_one_stripe(struct r5conf *conf, gfp_t gfp)
2174 {
2175 	struct stripe_head *sh;
2176 
2177 	sh = alloc_stripe(conf->slab_cache, gfp, conf->pool_size, conf);
2178 	if (!sh)
2179 		return 0;
2180 
2181 	if (grow_buffers(sh, gfp)) {
2182 		shrink_buffers(sh);
2183 		free_stripe(conf->slab_cache, sh);
2184 		return 0;
2185 	}
2186 	sh->hash_lock_index =
2187 		conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS;
2188 	/* we just created an active stripe so... */
2189 	atomic_inc(&conf->active_stripes);
2190 
2191 	raid5_release_stripe(sh);
2192 	conf->max_nr_stripes++;
2193 	return 1;
2194 }
2195 
2196 static int grow_stripes(struct r5conf *conf, int num)
2197 {
2198 	struct kmem_cache *sc;
2199 	size_t namelen = sizeof(conf->cache_name[0]);
2200 	int devs = max(conf->raid_disks, conf->previous_raid_disks);
2201 
2202 	if (conf->mddev->gendisk)
2203 		snprintf(conf->cache_name[0], namelen,
2204 			"raid%d-%s", conf->level, mdname(conf->mddev));
2205 	else
2206 		snprintf(conf->cache_name[0], namelen,
2207 			"raid%d-%p", conf->level, conf->mddev);
2208 	snprintf(conf->cache_name[1], namelen, "%.27s-alt", conf->cache_name[0]);
2209 
2210 	conf->active_name = 0;
2211 	sc = kmem_cache_create(conf->cache_name[conf->active_name],
2212 			       sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
2213 			       0, 0, NULL);
2214 	if (!sc)
2215 		return 1;
2216 	conf->slab_cache = sc;
2217 	conf->pool_size = devs;
2218 	while (num--)
2219 		if (!grow_one_stripe(conf, GFP_KERNEL))
2220 			return 1;
2221 
2222 	return 0;
2223 }
2224 
2225 /**
2226  * scribble_len - return the required size of the scribble region
2227  * @num - total number of disks in the array
2228  *
2229  * The size must be enough to contain:
2230  * 1/ a struct page pointer for each device in the array +2
2231  * 2/ room to convert each entry in (1) to its corresponding dma
2232  *    (dma_map_page()) or page (page_address()) address.
2233  *
2234  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
2235  * calculate over all devices (not just the data blocks), using zeros in place
2236  * of the P and Q blocks.
2237  */
2238 static int scribble_alloc(struct raid5_percpu *percpu,
2239 			  int num, int cnt, gfp_t flags)
2240 {
2241 	size_t obj_size =
2242 		sizeof(struct page *) * (num+2) +
2243 		sizeof(addr_conv_t) * (num+2);
2244 	void *scribble;
2245 
2246 	scribble = kvmalloc_array(cnt, obj_size, flags);
2247 	if (!scribble)
2248 		return -ENOMEM;
2249 
2250 	kvfree(percpu->scribble);
2251 
2252 	percpu->scribble = scribble;
2253 	percpu->scribble_obj_size = obj_size;
2254 	return 0;
2255 }
2256 
2257 static int resize_chunks(struct r5conf *conf, int new_disks, int new_sectors)
2258 {
2259 	unsigned long cpu;
2260 	int err = 0;
2261 
2262 	/*
2263 	 * Never shrink. And mddev_suspend() could deadlock if this is called
2264 	 * from raid5d. In that case, scribble_disks and scribble_sectors
2265 	 * should equal to new_disks and new_sectors
2266 	 */
2267 	if (conf->scribble_disks >= new_disks &&
2268 	    conf->scribble_sectors >= new_sectors)
2269 		return 0;
2270 	mddev_suspend(conf->mddev);
2271 	get_online_cpus();
2272 
2273 	for_each_present_cpu(cpu) {
2274 		struct raid5_percpu *percpu;
2275 
2276 		percpu = per_cpu_ptr(conf->percpu, cpu);
2277 		err = scribble_alloc(percpu, new_disks,
2278 				     new_sectors / STRIPE_SECTORS,
2279 				     GFP_NOIO);
2280 		if (err)
2281 			break;
2282 	}
2283 
2284 	put_online_cpus();
2285 	mddev_resume(conf->mddev);
2286 	if (!err) {
2287 		conf->scribble_disks = new_disks;
2288 		conf->scribble_sectors = new_sectors;
2289 	}
2290 	return err;
2291 }
2292 
2293 static int resize_stripes(struct r5conf *conf, int newsize)
2294 {
2295 	/* Make all the stripes able to hold 'newsize' devices.
2296 	 * New slots in each stripe get 'page' set to a new page.
2297 	 *
2298 	 * This happens in stages:
2299 	 * 1/ create a new kmem_cache and allocate the required number of
2300 	 *    stripe_heads.
2301 	 * 2/ gather all the old stripe_heads and transfer the pages across
2302 	 *    to the new stripe_heads.  This will have the side effect of
2303 	 *    freezing the array as once all stripe_heads have been collected,
2304 	 *    no IO will be possible.  Old stripe heads are freed once their
2305 	 *    pages have been transferred over, and the old kmem_cache is
2306 	 *    freed when all stripes are done.
2307 	 * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
2308 	 *    we simple return a failure status - no need to clean anything up.
2309 	 * 4/ allocate new pages for the new slots in the new stripe_heads.
2310 	 *    If this fails, we don't bother trying the shrink the
2311 	 *    stripe_heads down again, we just leave them as they are.
2312 	 *    As each stripe_head is processed the new one is released into
2313 	 *    active service.
2314 	 *
2315 	 * Once step2 is started, we cannot afford to wait for a write,
2316 	 * so we use GFP_NOIO allocations.
2317 	 */
2318 	struct stripe_head *osh, *nsh;
2319 	LIST_HEAD(newstripes);
2320 	struct disk_info *ndisks;
2321 	int err = 0;
2322 	struct kmem_cache *sc;
2323 	int i;
2324 	int hash, cnt;
2325 
2326 	md_allow_write(conf->mddev);
2327 
2328 	/* Step 1 */
2329 	sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
2330 			       sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
2331 			       0, 0, NULL);
2332 	if (!sc)
2333 		return -ENOMEM;
2334 
2335 	/* Need to ensure auto-resizing doesn't interfere */
2336 	mutex_lock(&conf->cache_size_mutex);
2337 
2338 	for (i = conf->max_nr_stripes; i; i--) {
2339 		nsh = alloc_stripe(sc, GFP_KERNEL, newsize, conf);
2340 		if (!nsh)
2341 			break;
2342 
2343 		list_add(&nsh->lru, &newstripes);
2344 	}
2345 	if (i) {
2346 		/* didn't get enough, give up */
2347 		while (!list_empty(&newstripes)) {
2348 			nsh = list_entry(newstripes.next, struct stripe_head, lru);
2349 			list_del(&nsh->lru);
2350 			free_stripe(sc, nsh);
2351 		}
2352 		kmem_cache_destroy(sc);
2353 		mutex_unlock(&conf->cache_size_mutex);
2354 		return -ENOMEM;
2355 	}
2356 	/* Step 2 - Must use GFP_NOIO now.
2357 	 * OK, we have enough stripes, start collecting inactive
2358 	 * stripes and copying them over
2359 	 */
2360 	hash = 0;
2361 	cnt = 0;
2362 	list_for_each_entry(nsh, &newstripes, lru) {
2363 		lock_device_hash_lock(conf, hash);
2364 		wait_event_cmd(conf->wait_for_stripe,
2365 				    !list_empty(conf->inactive_list + hash),
2366 				    unlock_device_hash_lock(conf, hash),
2367 				    lock_device_hash_lock(conf, hash));
2368 		osh = get_free_stripe(conf, hash);
2369 		unlock_device_hash_lock(conf, hash);
2370 
2371 		for(i=0; i<conf->pool_size; i++) {
2372 			nsh->dev[i].page = osh->dev[i].page;
2373 			nsh->dev[i].orig_page = osh->dev[i].page;
2374 		}
2375 		nsh->hash_lock_index = hash;
2376 		free_stripe(conf->slab_cache, osh);
2377 		cnt++;
2378 		if (cnt >= conf->max_nr_stripes / NR_STRIPE_HASH_LOCKS +
2379 		    !!((conf->max_nr_stripes % NR_STRIPE_HASH_LOCKS) > hash)) {
2380 			hash++;
2381 			cnt = 0;
2382 		}
2383 	}
2384 	kmem_cache_destroy(conf->slab_cache);
2385 
2386 	/* Step 3.
2387 	 * At this point, we are holding all the stripes so the array
2388 	 * is completely stalled, so now is a good time to resize
2389 	 * conf->disks and the scribble region
2390 	 */
2391 	ndisks = kcalloc(newsize, sizeof(struct disk_info), GFP_NOIO);
2392 	if (ndisks) {
2393 		for (i = 0; i < conf->pool_size; i++)
2394 			ndisks[i] = conf->disks[i];
2395 
2396 		for (i = conf->pool_size; i < newsize; i++) {
2397 			ndisks[i].extra_page = alloc_page(GFP_NOIO);
2398 			if (!ndisks[i].extra_page)
2399 				err = -ENOMEM;
2400 		}
2401 
2402 		if (err) {
2403 			for (i = conf->pool_size; i < newsize; i++)
2404 				if (ndisks[i].extra_page)
2405 					put_page(ndisks[i].extra_page);
2406 			kfree(ndisks);
2407 		} else {
2408 			kfree(conf->disks);
2409 			conf->disks = ndisks;
2410 		}
2411 	} else
2412 		err = -ENOMEM;
2413 
2414 	mutex_unlock(&conf->cache_size_mutex);
2415 
2416 	conf->slab_cache = sc;
2417 	conf->active_name = 1-conf->active_name;
2418 
2419 	/* Step 4, return new stripes to service */
2420 	while(!list_empty(&newstripes)) {
2421 		nsh = list_entry(newstripes.next, struct stripe_head, lru);
2422 		list_del_init(&nsh->lru);
2423 
2424 		for (i=conf->raid_disks; i < newsize; i++)
2425 			if (nsh->dev[i].page == NULL) {
2426 				struct page *p = alloc_page(GFP_NOIO);
2427 				nsh->dev[i].page = p;
2428 				nsh->dev[i].orig_page = p;
2429 				if (!p)
2430 					err = -ENOMEM;
2431 			}
2432 		raid5_release_stripe(nsh);
2433 	}
2434 	/* critical section pass, GFP_NOIO no longer needed */
2435 
2436 	if (!err)
2437 		conf->pool_size = newsize;
2438 	return err;
2439 }
2440 
2441 static int drop_one_stripe(struct r5conf *conf)
2442 {
2443 	struct stripe_head *sh;
2444 	int hash = (conf->max_nr_stripes - 1) & STRIPE_HASH_LOCKS_MASK;
2445 
2446 	spin_lock_irq(conf->hash_locks + hash);
2447 	sh = get_free_stripe(conf, hash);
2448 	spin_unlock_irq(conf->hash_locks + hash);
2449 	if (!sh)
2450 		return 0;
2451 	BUG_ON(atomic_read(&sh->count));
2452 	shrink_buffers(sh);
2453 	free_stripe(conf->slab_cache, sh);
2454 	atomic_dec(&conf->active_stripes);
2455 	conf->max_nr_stripes--;
2456 	return 1;
2457 }
2458 
2459 static void shrink_stripes(struct r5conf *conf)
2460 {
2461 	while (conf->max_nr_stripes &&
2462 	       drop_one_stripe(conf))
2463 		;
2464 
2465 	kmem_cache_destroy(conf->slab_cache);
2466 	conf->slab_cache = NULL;
2467 }
2468 
2469 static void raid5_end_read_request(struct bio * bi)
2470 {
2471 	struct stripe_head *sh = bi->bi_private;
2472 	struct r5conf *conf = sh->raid_conf;
2473 	int disks = sh->disks, i;
2474 	char b[BDEVNAME_SIZE];
2475 	struct md_rdev *rdev = NULL;
2476 	sector_t s;
2477 
2478 	for (i=0 ; i<disks; i++)
2479 		if (bi == &sh->dev[i].req)
2480 			break;
2481 
2482 	pr_debug("end_read_request %llu/%d, count: %d, error %d.\n",
2483 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
2484 		bi->bi_status);
2485 	if (i == disks) {
2486 		bio_reset(bi);
2487 		BUG();
2488 		return;
2489 	}
2490 	if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2491 		/* If replacement finished while this request was outstanding,
2492 		 * 'replacement' might be NULL already.
2493 		 * In that case it moved down to 'rdev'.
2494 		 * rdev is not removed until all requests are finished.
2495 		 */
2496 		rdev = conf->disks[i].replacement;
2497 	if (!rdev)
2498 		rdev = conf->disks[i].rdev;
2499 
2500 	if (use_new_offset(conf, sh))
2501 		s = sh->sector + rdev->new_data_offset;
2502 	else
2503 		s = sh->sector + rdev->data_offset;
2504 	if (!bi->bi_status) {
2505 		set_bit(R5_UPTODATE, &sh->dev[i].flags);
2506 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2507 			/* Note that this cannot happen on a
2508 			 * replacement device.  We just fail those on
2509 			 * any error
2510 			 */
2511 			pr_info_ratelimited(
2512 				"md/raid:%s: read error corrected (%lu sectors at %llu on %s)\n",
2513 				mdname(conf->mddev), STRIPE_SECTORS,
2514 				(unsigned long long)s,
2515 				bdevname(rdev->bdev, b));
2516 			atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
2517 			clear_bit(R5_ReadError, &sh->dev[i].flags);
2518 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
2519 		} else if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2520 			clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2521 
2522 		if (test_bit(R5_InJournal, &sh->dev[i].flags))
2523 			/*
2524 			 * end read for a page in journal, this
2525 			 * must be preparing for prexor in rmw
2526 			 */
2527 			set_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2528 
2529 		if (atomic_read(&rdev->read_errors))
2530 			atomic_set(&rdev->read_errors, 0);
2531 	} else {
2532 		const char *bdn = bdevname(rdev->bdev, b);
2533 		int retry = 0;
2534 		int set_bad = 0;
2535 
2536 		clear_bit(R5_UPTODATE, &sh->dev[i].flags);
2537 		atomic_inc(&rdev->read_errors);
2538 		if (test_bit(R5_ReadRepl, &sh->dev[i].flags))
2539 			pr_warn_ratelimited(
2540 				"md/raid:%s: read error on replacement device (sector %llu on %s).\n",
2541 				mdname(conf->mddev),
2542 				(unsigned long long)s,
2543 				bdn);
2544 		else if (conf->mddev->degraded >= conf->max_degraded) {
2545 			set_bad = 1;
2546 			pr_warn_ratelimited(
2547 				"md/raid:%s: read error not correctable (sector %llu on %s).\n",
2548 				mdname(conf->mddev),
2549 				(unsigned long long)s,
2550 				bdn);
2551 		} else if (test_bit(R5_ReWrite, &sh->dev[i].flags)) {
2552 			/* Oh, no!!! */
2553 			set_bad = 1;
2554 			pr_warn_ratelimited(
2555 				"md/raid:%s: read error NOT corrected!! (sector %llu on %s).\n",
2556 				mdname(conf->mddev),
2557 				(unsigned long long)s,
2558 				bdn);
2559 		} else if (atomic_read(&rdev->read_errors)
2560 			 > conf->max_nr_stripes)
2561 			pr_warn("md/raid:%s: Too many read errors, failing device %s.\n",
2562 			       mdname(conf->mddev), bdn);
2563 		else
2564 			retry = 1;
2565 		if (set_bad && test_bit(In_sync, &rdev->flags)
2566 		    && !test_bit(R5_ReadNoMerge, &sh->dev[i].flags))
2567 			retry = 1;
2568 		if (retry)
2569 			if (test_bit(R5_ReadNoMerge, &sh->dev[i].flags)) {
2570 				set_bit(R5_ReadError, &sh->dev[i].flags);
2571 				clear_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2572 			} else
2573 				set_bit(R5_ReadNoMerge, &sh->dev[i].flags);
2574 		else {
2575 			clear_bit(R5_ReadError, &sh->dev[i].flags);
2576 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
2577 			if (!(set_bad
2578 			      && test_bit(In_sync, &rdev->flags)
2579 			      && rdev_set_badblocks(
2580 				      rdev, sh->sector, STRIPE_SECTORS, 0)))
2581 				md_error(conf->mddev, rdev);
2582 		}
2583 	}
2584 	rdev_dec_pending(rdev, conf->mddev);
2585 	bio_reset(bi);
2586 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
2587 	set_bit(STRIPE_HANDLE, &sh->state);
2588 	raid5_release_stripe(sh);
2589 }
2590 
2591 static void raid5_end_write_request(struct bio *bi)
2592 {
2593 	struct stripe_head *sh = bi->bi_private;
2594 	struct r5conf *conf = sh->raid_conf;
2595 	int disks = sh->disks, i;
2596 	struct md_rdev *uninitialized_var(rdev);
2597 	sector_t first_bad;
2598 	int bad_sectors;
2599 	int replacement = 0;
2600 
2601 	for (i = 0 ; i < disks; i++) {
2602 		if (bi == &sh->dev[i].req) {
2603 			rdev = conf->disks[i].rdev;
2604 			break;
2605 		}
2606 		if (bi == &sh->dev[i].rreq) {
2607 			rdev = conf->disks[i].replacement;
2608 			if (rdev)
2609 				replacement = 1;
2610 			else
2611 				/* rdev was removed and 'replacement'
2612 				 * replaced it.  rdev is not removed
2613 				 * until all requests are finished.
2614 				 */
2615 				rdev = conf->disks[i].rdev;
2616 			break;
2617 		}
2618 	}
2619 	pr_debug("end_write_request %llu/%d, count %d, error: %d.\n",
2620 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
2621 		bi->bi_status);
2622 	if (i == disks) {
2623 		bio_reset(bi);
2624 		BUG();
2625 		return;
2626 	}
2627 
2628 	if (replacement) {
2629 		if (bi->bi_status)
2630 			md_error(conf->mddev, rdev);
2631 		else if (is_badblock(rdev, sh->sector,
2632 				     STRIPE_SECTORS,
2633 				     &first_bad, &bad_sectors))
2634 			set_bit(R5_MadeGoodRepl, &sh->dev[i].flags);
2635 	} else {
2636 		if (bi->bi_status) {
2637 			set_bit(STRIPE_DEGRADED, &sh->state);
2638 			set_bit(WriteErrorSeen, &rdev->flags);
2639 			set_bit(R5_WriteError, &sh->dev[i].flags);
2640 			if (!test_and_set_bit(WantReplacement, &rdev->flags))
2641 				set_bit(MD_RECOVERY_NEEDED,
2642 					&rdev->mddev->recovery);
2643 		} else if (is_badblock(rdev, sh->sector,
2644 				       STRIPE_SECTORS,
2645 				       &first_bad, &bad_sectors)) {
2646 			set_bit(R5_MadeGood, &sh->dev[i].flags);
2647 			if (test_bit(R5_ReadError, &sh->dev[i].flags))
2648 				/* That was a successful write so make
2649 				 * sure it looks like we already did
2650 				 * a re-write.
2651 				 */
2652 				set_bit(R5_ReWrite, &sh->dev[i].flags);
2653 		}
2654 	}
2655 	rdev_dec_pending(rdev, conf->mddev);
2656 
2657 	if (sh->batch_head && bi->bi_status && !replacement)
2658 		set_bit(STRIPE_BATCH_ERR, &sh->batch_head->state);
2659 
2660 	bio_reset(bi);
2661 	if (!test_and_clear_bit(R5_DOUBLE_LOCKED, &sh->dev[i].flags))
2662 		clear_bit(R5_LOCKED, &sh->dev[i].flags);
2663 	set_bit(STRIPE_HANDLE, &sh->state);
2664 	raid5_release_stripe(sh);
2665 
2666 	if (sh->batch_head && sh != sh->batch_head)
2667 		raid5_release_stripe(sh->batch_head);
2668 }
2669 
2670 static void raid5_error(struct mddev *mddev, struct md_rdev *rdev)
2671 {
2672 	char b[BDEVNAME_SIZE];
2673 	struct r5conf *conf = mddev->private;
2674 	unsigned long flags;
2675 	pr_debug("raid456: error called\n");
2676 
2677 	spin_lock_irqsave(&conf->device_lock, flags);
2678 
2679 	if (test_bit(In_sync, &rdev->flags) &&
2680 	    mddev->degraded == conf->max_degraded) {
2681 		/*
2682 		 * Don't allow to achieve failed state
2683 		 * Don't try to recover this device
2684 		 */
2685 		conf->recovery_disabled = mddev->recovery_disabled;
2686 		spin_unlock_irqrestore(&conf->device_lock, flags);
2687 		return;
2688 	}
2689 
2690 	set_bit(Faulty, &rdev->flags);
2691 	clear_bit(In_sync, &rdev->flags);
2692 	mddev->degraded = raid5_calc_degraded(conf);
2693 	spin_unlock_irqrestore(&conf->device_lock, flags);
2694 	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2695 
2696 	set_bit(Blocked, &rdev->flags);
2697 	set_mask_bits(&mddev->sb_flags, 0,
2698 		      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2699 	pr_crit("md/raid:%s: Disk failure on %s, disabling device.\n"
2700 		"md/raid:%s: Operation continuing on %d devices.\n",
2701 		mdname(mddev),
2702 		bdevname(rdev->bdev, b),
2703 		mdname(mddev),
2704 		conf->raid_disks - mddev->degraded);
2705 	r5c_update_on_rdev_error(mddev, rdev);
2706 }
2707 
2708 /*
2709  * Input: a 'big' sector number,
2710  * Output: index of the data and parity disk, and the sector # in them.
2711  */
2712 sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
2713 			      int previous, int *dd_idx,
2714 			      struct stripe_head *sh)
2715 {
2716 	sector_t stripe, stripe2;
2717 	sector_t chunk_number;
2718 	unsigned int chunk_offset;
2719 	int pd_idx, qd_idx;
2720 	int ddf_layout = 0;
2721 	sector_t new_sector;
2722 	int algorithm = previous ? conf->prev_algo
2723 				 : conf->algorithm;
2724 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2725 					 : conf->chunk_sectors;
2726 	int raid_disks = previous ? conf->previous_raid_disks
2727 				  : conf->raid_disks;
2728 	int data_disks = raid_disks - conf->max_degraded;
2729 
2730 	/* First compute the information on this sector */
2731 
2732 	/*
2733 	 * Compute the chunk number and the sector offset inside the chunk
2734 	 */
2735 	chunk_offset = sector_div(r_sector, sectors_per_chunk);
2736 	chunk_number = r_sector;
2737 
2738 	/*
2739 	 * Compute the stripe number
2740 	 */
2741 	stripe = chunk_number;
2742 	*dd_idx = sector_div(stripe, data_disks);
2743 	stripe2 = stripe;
2744 	/*
2745 	 * Select the parity disk based on the user selected algorithm.
2746 	 */
2747 	pd_idx = qd_idx = -1;
2748 	switch(conf->level) {
2749 	case 4:
2750 		pd_idx = data_disks;
2751 		break;
2752 	case 5:
2753 		switch (algorithm) {
2754 		case ALGORITHM_LEFT_ASYMMETRIC:
2755 			pd_idx = data_disks - sector_div(stripe2, raid_disks);
2756 			if (*dd_idx >= pd_idx)
2757 				(*dd_idx)++;
2758 			break;
2759 		case ALGORITHM_RIGHT_ASYMMETRIC:
2760 			pd_idx = sector_div(stripe2, raid_disks);
2761 			if (*dd_idx >= pd_idx)
2762 				(*dd_idx)++;
2763 			break;
2764 		case ALGORITHM_LEFT_SYMMETRIC:
2765 			pd_idx = data_disks - sector_div(stripe2, raid_disks);
2766 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2767 			break;
2768 		case ALGORITHM_RIGHT_SYMMETRIC:
2769 			pd_idx = sector_div(stripe2, raid_disks);
2770 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2771 			break;
2772 		case ALGORITHM_PARITY_0:
2773 			pd_idx = 0;
2774 			(*dd_idx)++;
2775 			break;
2776 		case ALGORITHM_PARITY_N:
2777 			pd_idx = data_disks;
2778 			break;
2779 		default:
2780 			BUG();
2781 		}
2782 		break;
2783 	case 6:
2784 
2785 		switch (algorithm) {
2786 		case ALGORITHM_LEFT_ASYMMETRIC:
2787 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2788 			qd_idx = pd_idx + 1;
2789 			if (pd_idx == raid_disks-1) {
2790 				(*dd_idx)++;	/* Q D D D P */
2791 				qd_idx = 0;
2792 			} else if (*dd_idx >= pd_idx)
2793 				(*dd_idx) += 2; /* D D P Q D */
2794 			break;
2795 		case ALGORITHM_RIGHT_ASYMMETRIC:
2796 			pd_idx = sector_div(stripe2, raid_disks);
2797 			qd_idx = pd_idx + 1;
2798 			if (pd_idx == raid_disks-1) {
2799 				(*dd_idx)++;	/* Q D D D P */
2800 				qd_idx = 0;
2801 			} else if (*dd_idx >= pd_idx)
2802 				(*dd_idx) += 2; /* D D P Q D */
2803 			break;
2804 		case ALGORITHM_LEFT_SYMMETRIC:
2805 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2806 			qd_idx = (pd_idx + 1) % raid_disks;
2807 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2808 			break;
2809 		case ALGORITHM_RIGHT_SYMMETRIC:
2810 			pd_idx = sector_div(stripe2, raid_disks);
2811 			qd_idx = (pd_idx + 1) % raid_disks;
2812 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
2813 			break;
2814 
2815 		case ALGORITHM_PARITY_0:
2816 			pd_idx = 0;
2817 			qd_idx = 1;
2818 			(*dd_idx) += 2;
2819 			break;
2820 		case ALGORITHM_PARITY_N:
2821 			pd_idx = data_disks;
2822 			qd_idx = data_disks + 1;
2823 			break;
2824 
2825 		case ALGORITHM_ROTATING_ZERO_RESTART:
2826 			/* Exactly the same as RIGHT_ASYMMETRIC, but or
2827 			 * of blocks for computing Q is different.
2828 			 */
2829 			pd_idx = sector_div(stripe2, raid_disks);
2830 			qd_idx = pd_idx + 1;
2831 			if (pd_idx == raid_disks-1) {
2832 				(*dd_idx)++;	/* Q D D D P */
2833 				qd_idx = 0;
2834 			} else if (*dd_idx >= pd_idx)
2835 				(*dd_idx) += 2; /* D D P Q D */
2836 			ddf_layout = 1;
2837 			break;
2838 
2839 		case ALGORITHM_ROTATING_N_RESTART:
2840 			/* Same a left_asymmetric, by first stripe is
2841 			 * D D D P Q  rather than
2842 			 * Q D D D P
2843 			 */
2844 			stripe2 += 1;
2845 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2846 			qd_idx = pd_idx + 1;
2847 			if (pd_idx == raid_disks-1) {
2848 				(*dd_idx)++;	/* Q D D D P */
2849 				qd_idx = 0;
2850 			} else if (*dd_idx >= pd_idx)
2851 				(*dd_idx) += 2; /* D D P Q D */
2852 			ddf_layout = 1;
2853 			break;
2854 
2855 		case ALGORITHM_ROTATING_N_CONTINUE:
2856 			/* Same as left_symmetric but Q is before P */
2857 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
2858 			qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
2859 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
2860 			ddf_layout = 1;
2861 			break;
2862 
2863 		case ALGORITHM_LEFT_ASYMMETRIC_6:
2864 			/* RAID5 left_asymmetric, with Q on last device */
2865 			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2866 			if (*dd_idx >= pd_idx)
2867 				(*dd_idx)++;
2868 			qd_idx = raid_disks - 1;
2869 			break;
2870 
2871 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
2872 			pd_idx = sector_div(stripe2, raid_disks-1);
2873 			if (*dd_idx >= pd_idx)
2874 				(*dd_idx)++;
2875 			qd_idx = raid_disks - 1;
2876 			break;
2877 
2878 		case ALGORITHM_LEFT_SYMMETRIC_6:
2879 			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
2880 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2881 			qd_idx = raid_disks - 1;
2882 			break;
2883 
2884 		case ALGORITHM_RIGHT_SYMMETRIC_6:
2885 			pd_idx = sector_div(stripe2, raid_disks-1);
2886 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
2887 			qd_idx = raid_disks - 1;
2888 			break;
2889 
2890 		case ALGORITHM_PARITY_0_6:
2891 			pd_idx = 0;
2892 			(*dd_idx)++;
2893 			qd_idx = raid_disks - 1;
2894 			break;
2895 
2896 		default:
2897 			BUG();
2898 		}
2899 		break;
2900 	}
2901 
2902 	if (sh) {
2903 		sh->pd_idx = pd_idx;
2904 		sh->qd_idx = qd_idx;
2905 		sh->ddf_layout = ddf_layout;
2906 	}
2907 	/*
2908 	 * Finally, compute the new sector number
2909 	 */
2910 	new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
2911 	return new_sector;
2912 }
2913 
2914 sector_t raid5_compute_blocknr(struct stripe_head *sh, int i, int previous)
2915 {
2916 	struct r5conf *conf = sh->raid_conf;
2917 	int raid_disks = sh->disks;
2918 	int data_disks = raid_disks - conf->max_degraded;
2919 	sector_t new_sector = sh->sector, check;
2920 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
2921 					 : conf->chunk_sectors;
2922 	int algorithm = previous ? conf->prev_algo
2923 				 : conf->algorithm;
2924 	sector_t stripe;
2925 	int chunk_offset;
2926 	sector_t chunk_number;
2927 	int dummy1, dd_idx = i;
2928 	sector_t r_sector;
2929 	struct stripe_head sh2;
2930 
2931 	chunk_offset = sector_div(new_sector, sectors_per_chunk);
2932 	stripe = new_sector;
2933 
2934 	if (i == sh->pd_idx)
2935 		return 0;
2936 	switch(conf->level) {
2937 	case 4: break;
2938 	case 5:
2939 		switch (algorithm) {
2940 		case ALGORITHM_LEFT_ASYMMETRIC:
2941 		case ALGORITHM_RIGHT_ASYMMETRIC:
2942 			if (i > sh->pd_idx)
2943 				i--;
2944 			break;
2945 		case ALGORITHM_LEFT_SYMMETRIC:
2946 		case ALGORITHM_RIGHT_SYMMETRIC:
2947 			if (i < sh->pd_idx)
2948 				i += raid_disks;
2949 			i -= (sh->pd_idx + 1);
2950 			break;
2951 		case ALGORITHM_PARITY_0:
2952 			i -= 1;
2953 			break;
2954 		case ALGORITHM_PARITY_N:
2955 			break;
2956 		default:
2957 			BUG();
2958 		}
2959 		break;
2960 	case 6:
2961 		if (i == sh->qd_idx)
2962 			return 0; /* It is the Q disk */
2963 		switch (algorithm) {
2964 		case ALGORITHM_LEFT_ASYMMETRIC:
2965 		case ALGORITHM_RIGHT_ASYMMETRIC:
2966 		case ALGORITHM_ROTATING_ZERO_RESTART:
2967 		case ALGORITHM_ROTATING_N_RESTART:
2968 			if (sh->pd_idx == raid_disks-1)
2969 				i--;	/* Q D D D P */
2970 			else if (i > sh->pd_idx)
2971 				i -= 2; /* D D P Q D */
2972 			break;
2973 		case ALGORITHM_LEFT_SYMMETRIC:
2974 		case ALGORITHM_RIGHT_SYMMETRIC:
2975 			if (sh->pd_idx == raid_disks-1)
2976 				i--; /* Q D D D P */
2977 			else {
2978 				/* D D P Q D */
2979 				if (i < sh->pd_idx)
2980 					i += raid_disks;
2981 				i -= (sh->pd_idx + 2);
2982 			}
2983 			break;
2984 		case ALGORITHM_PARITY_0:
2985 			i -= 2;
2986 			break;
2987 		case ALGORITHM_PARITY_N:
2988 			break;
2989 		case ALGORITHM_ROTATING_N_CONTINUE:
2990 			/* Like left_symmetric, but P is before Q */
2991 			if (sh->pd_idx == 0)
2992 				i--;	/* P D D D Q */
2993 			else {
2994 				/* D D Q P D */
2995 				if (i < sh->pd_idx)
2996 					i += raid_disks;
2997 				i -= (sh->pd_idx + 1);
2998 			}
2999 			break;
3000 		case ALGORITHM_LEFT_ASYMMETRIC_6:
3001 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
3002 			if (i > sh->pd_idx)
3003 				i--;
3004 			break;
3005 		case ALGORITHM_LEFT_SYMMETRIC_6:
3006 		case ALGORITHM_RIGHT_SYMMETRIC_6:
3007 			if (i < sh->pd_idx)
3008 				i += data_disks + 1;
3009 			i -= (sh->pd_idx + 1);
3010 			break;
3011 		case ALGORITHM_PARITY_0_6:
3012 			i -= 1;
3013 			break;
3014 		default:
3015 			BUG();
3016 		}
3017 		break;
3018 	}
3019 
3020 	chunk_number = stripe * data_disks + i;
3021 	r_sector = chunk_number * sectors_per_chunk + chunk_offset;
3022 
3023 	check = raid5_compute_sector(conf, r_sector,
3024 				     previous, &dummy1, &sh2);
3025 	if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
3026 		|| sh2.qd_idx != sh->qd_idx) {
3027 		pr_warn("md/raid:%s: compute_blocknr: map not correct\n",
3028 			mdname(conf->mddev));
3029 		return 0;
3030 	}
3031 	return r_sector;
3032 }
3033 
3034 /*
3035  * There are cases where we want handle_stripe_dirtying() and
3036  * schedule_reconstruction() to delay towrite to some dev of a stripe.
3037  *
3038  * This function checks whether we want to delay the towrite. Specifically,
3039  * we delay the towrite when:
3040  *
3041  *   1. degraded stripe has a non-overwrite to the missing dev, AND this
3042  *      stripe has data in journal (for other devices).
3043  *
3044  *      In this case, when reading data for the non-overwrite dev, it is
3045  *      necessary to handle complex rmw of write back cache (prexor with
3046  *      orig_page, and xor with page). To keep read path simple, we would
3047  *      like to flush data in journal to RAID disks first, so complex rmw
3048  *      is handled in the write patch (handle_stripe_dirtying).
3049  *
3050  *   2. when journal space is critical (R5C_LOG_CRITICAL=1)
3051  *
3052  *      It is important to be able to flush all stripes in raid5-cache.
3053  *      Therefore, we need reserve some space on the journal device for
3054  *      these flushes. If flush operation includes pending writes to the
3055  *      stripe, we need to reserve (conf->raid_disk + 1) pages per stripe
3056  *      for the flush out. If we exclude these pending writes from flush
3057  *      operation, we only need (conf->max_degraded + 1) pages per stripe.
3058  *      Therefore, excluding pending writes in these cases enables more
3059  *      efficient use of the journal device.
3060  *
3061  *      Note: To make sure the stripe makes progress, we only delay
3062  *      towrite for stripes with data already in journal (injournal > 0).
3063  *      When LOG_CRITICAL, stripes with injournal == 0 will be sent to
3064  *      no_space_stripes list.
3065  *
3066  *   3. during journal failure
3067  *      In journal failure, we try to flush all cached data to raid disks
3068  *      based on data in stripe cache. The array is read-only to upper
3069  *      layers, so we would skip all pending writes.
3070  *
3071  */
3072 static inline bool delay_towrite(struct r5conf *conf,
3073 				 struct r5dev *dev,
3074 				 struct stripe_head_state *s)
3075 {
3076 	/* case 1 above */
3077 	if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3078 	    !test_bit(R5_Insync, &dev->flags) && s->injournal)
3079 		return true;
3080 	/* case 2 above */
3081 	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
3082 	    s->injournal > 0)
3083 		return true;
3084 	/* case 3 above */
3085 	if (s->log_failed && s->injournal)
3086 		return true;
3087 	return false;
3088 }
3089 
3090 static void
3091 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
3092 			 int rcw, int expand)
3093 {
3094 	int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx, disks = sh->disks;
3095 	struct r5conf *conf = sh->raid_conf;
3096 	int level = conf->level;
3097 
3098 	if (rcw) {
3099 		/*
3100 		 * In some cases, handle_stripe_dirtying initially decided to
3101 		 * run rmw and allocates extra page for prexor. However, rcw is
3102 		 * cheaper later on. We need to free the extra page now,
3103 		 * because we won't be able to do that in ops_complete_prexor().
3104 		 */
3105 		r5c_release_extra_page(sh);
3106 
3107 		for (i = disks; i--; ) {
3108 			struct r5dev *dev = &sh->dev[i];
3109 
3110 			if (dev->towrite && !delay_towrite(conf, dev, s)) {
3111 				set_bit(R5_LOCKED, &dev->flags);
3112 				set_bit(R5_Wantdrain, &dev->flags);
3113 				if (!expand)
3114 					clear_bit(R5_UPTODATE, &dev->flags);
3115 				s->locked++;
3116 			} else if (test_bit(R5_InJournal, &dev->flags)) {
3117 				set_bit(R5_LOCKED, &dev->flags);
3118 				s->locked++;
3119 			}
3120 		}
3121 		/* if we are not expanding this is a proper write request, and
3122 		 * there will be bios with new data to be drained into the
3123 		 * stripe cache
3124 		 */
3125 		if (!expand) {
3126 			if (!s->locked)
3127 				/* False alarm, nothing to do */
3128 				return;
3129 			sh->reconstruct_state = reconstruct_state_drain_run;
3130 			set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3131 		} else
3132 			sh->reconstruct_state = reconstruct_state_run;
3133 
3134 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3135 
3136 		if (s->locked + conf->max_degraded == disks)
3137 			if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
3138 				atomic_inc(&conf->pending_full_writes);
3139 	} else {
3140 		BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
3141 			test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
3142 		BUG_ON(level == 6 &&
3143 			(!(test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags) ||
3144 			   test_bit(R5_Wantcompute, &sh->dev[qd_idx].flags))));
3145 
3146 		for (i = disks; i--; ) {
3147 			struct r5dev *dev = &sh->dev[i];
3148 			if (i == pd_idx || i == qd_idx)
3149 				continue;
3150 
3151 			if (dev->towrite &&
3152 			    (test_bit(R5_UPTODATE, &dev->flags) ||
3153 			     test_bit(R5_Wantcompute, &dev->flags))) {
3154 				set_bit(R5_Wantdrain, &dev->flags);
3155 				set_bit(R5_LOCKED, &dev->flags);
3156 				clear_bit(R5_UPTODATE, &dev->flags);
3157 				s->locked++;
3158 			} else if (test_bit(R5_InJournal, &dev->flags)) {
3159 				set_bit(R5_LOCKED, &dev->flags);
3160 				s->locked++;
3161 			}
3162 		}
3163 		if (!s->locked)
3164 			/* False alarm - nothing to do */
3165 			return;
3166 		sh->reconstruct_state = reconstruct_state_prexor_drain_run;
3167 		set_bit(STRIPE_OP_PREXOR, &s->ops_request);
3168 		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
3169 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
3170 	}
3171 
3172 	/* keep the parity disk(s) locked while asynchronous operations
3173 	 * are in flight
3174 	 */
3175 	set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
3176 	clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
3177 	s->locked++;
3178 
3179 	if (level == 6) {
3180 		int qd_idx = sh->qd_idx;
3181 		struct r5dev *dev = &sh->dev[qd_idx];
3182 
3183 		set_bit(R5_LOCKED, &dev->flags);
3184 		clear_bit(R5_UPTODATE, &dev->flags);
3185 		s->locked++;
3186 	}
3187 
3188 	if (raid5_has_ppl(sh->raid_conf) && sh->ppl_page &&
3189 	    test_bit(STRIPE_OP_BIODRAIN, &s->ops_request) &&
3190 	    !test_bit(STRIPE_FULL_WRITE, &sh->state) &&
3191 	    test_bit(R5_Insync, &sh->dev[pd_idx].flags))
3192 		set_bit(STRIPE_OP_PARTIAL_PARITY, &s->ops_request);
3193 
3194 	pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
3195 		__func__, (unsigned long long)sh->sector,
3196 		s->locked, s->ops_request);
3197 }
3198 
3199 /*
3200  * Each stripe/dev can have one or more bion attached.
3201  * toread/towrite point to the first in a chain.
3202  * The bi_next chain must be in order.
3203  */
3204 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx,
3205 			  int forwrite, int previous)
3206 {
3207 	struct bio **bip;
3208 	struct r5conf *conf = sh->raid_conf;
3209 	int firstwrite=0;
3210 
3211 	pr_debug("adding bi b#%llu to stripe s#%llu\n",
3212 		(unsigned long long)bi->bi_iter.bi_sector,
3213 		(unsigned long long)sh->sector);
3214 
3215 	spin_lock_irq(&sh->stripe_lock);
3216 	sh->dev[dd_idx].write_hint = bi->bi_write_hint;
3217 	/* Don't allow new IO added to stripes in batch list */
3218 	if (sh->batch_head)
3219 		goto overlap;
3220 	if (forwrite) {
3221 		bip = &sh->dev[dd_idx].towrite;
3222 		if (*bip == NULL)
3223 			firstwrite = 1;
3224 	} else
3225 		bip = &sh->dev[dd_idx].toread;
3226 	while (*bip && (*bip)->bi_iter.bi_sector < bi->bi_iter.bi_sector) {
3227 		if (bio_end_sector(*bip) > bi->bi_iter.bi_sector)
3228 			goto overlap;
3229 		bip = & (*bip)->bi_next;
3230 	}
3231 	if (*bip && (*bip)->bi_iter.bi_sector < bio_end_sector(bi))
3232 		goto overlap;
3233 
3234 	if (forwrite && raid5_has_ppl(conf)) {
3235 		/*
3236 		 * With PPL only writes to consecutive data chunks within a
3237 		 * stripe are allowed because for a single stripe_head we can
3238 		 * only have one PPL entry at a time, which describes one data
3239 		 * range. Not really an overlap, but wait_for_overlap can be
3240 		 * used to handle this.
3241 		 */
3242 		sector_t sector;
3243 		sector_t first = 0;
3244 		sector_t last = 0;
3245 		int count = 0;
3246 		int i;
3247 
3248 		for (i = 0; i < sh->disks; i++) {
3249 			if (i != sh->pd_idx &&
3250 			    (i == dd_idx || sh->dev[i].towrite)) {
3251 				sector = sh->dev[i].sector;
3252 				if (count == 0 || sector < first)
3253 					first = sector;
3254 				if (sector > last)
3255 					last = sector;
3256 				count++;
3257 			}
3258 		}
3259 
3260 		if (first + conf->chunk_sectors * (count - 1) != last)
3261 			goto overlap;
3262 	}
3263 
3264 	if (!forwrite || previous)
3265 		clear_bit(STRIPE_BATCH_READY, &sh->state);
3266 
3267 	BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
3268 	if (*bip)
3269 		bi->bi_next = *bip;
3270 	*bip = bi;
3271 	bio_inc_remaining(bi);
3272 	md_write_inc(conf->mddev, bi);
3273 
3274 	if (forwrite) {
3275 		/* check if page is covered */
3276 		sector_t sector = sh->dev[dd_idx].sector;
3277 		for (bi=sh->dev[dd_idx].towrite;
3278 		     sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
3279 			     bi && bi->bi_iter.bi_sector <= sector;
3280 		     bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
3281 			if (bio_end_sector(bi) >= sector)
3282 				sector = bio_end_sector(bi);
3283 		}
3284 		if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
3285 			if (!test_and_set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags))
3286 				sh->overwrite_disks++;
3287 	}
3288 
3289 	pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
3290 		(unsigned long long)(*bip)->bi_iter.bi_sector,
3291 		(unsigned long long)sh->sector, dd_idx);
3292 
3293 	if (conf->mddev->bitmap && firstwrite) {
3294 		/* Cannot hold spinlock over bitmap_startwrite,
3295 		 * but must ensure this isn't added to a batch until
3296 		 * we have added to the bitmap and set bm_seq.
3297 		 * So set STRIPE_BITMAP_PENDING to prevent
3298 		 * batching.
3299 		 * If multiple add_stripe_bio() calls race here they
3300 		 * much all set STRIPE_BITMAP_PENDING.  So only the first one
3301 		 * to complete "bitmap_startwrite" gets to set
3302 		 * STRIPE_BIT_DELAY.  This is important as once a stripe
3303 		 * is added to a batch, STRIPE_BIT_DELAY cannot be changed
3304 		 * any more.
3305 		 */
3306 		set_bit(STRIPE_BITMAP_PENDING, &sh->state);
3307 		spin_unlock_irq(&sh->stripe_lock);
3308 		md_bitmap_startwrite(conf->mddev->bitmap, sh->sector,
3309 				     STRIPE_SECTORS, 0);
3310 		spin_lock_irq(&sh->stripe_lock);
3311 		clear_bit(STRIPE_BITMAP_PENDING, &sh->state);
3312 		if (!sh->batch_head) {
3313 			sh->bm_seq = conf->seq_flush+1;
3314 			set_bit(STRIPE_BIT_DELAY, &sh->state);
3315 		}
3316 	}
3317 	spin_unlock_irq(&sh->stripe_lock);
3318 
3319 	if (stripe_can_batch(sh))
3320 		stripe_add_to_batch_list(conf, sh);
3321 	return 1;
3322 
3323  overlap:
3324 	set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
3325 	spin_unlock_irq(&sh->stripe_lock);
3326 	return 0;
3327 }
3328 
3329 static void end_reshape(struct r5conf *conf);
3330 
3331 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
3332 			    struct stripe_head *sh)
3333 {
3334 	int sectors_per_chunk =
3335 		previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
3336 	int dd_idx;
3337 	int chunk_offset = sector_div(stripe, sectors_per_chunk);
3338 	int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
3339 
3340 	raid5_compute_sector(conf,
3341 			     stripe * (disks - conf->max_degraded)
3342 			     *sectors_per_chunk + chunk_offset,
3343 			     previous,
3344 			     &dd_idx, sh);
3345 }
3346 
3347 static void
3348 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
3349 		     struct stripe_head_state *s, int disks)
3350 {
3351 	int i;
3352 	BUG_ON(sh->batch_head);
3353 	for (i = disks; i--; ) {
3354 		struct bio *bi;
3355 		int bitmap_end = 0;
3356 
3357 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
3358 			struct md_rdev *rdev;
3359 			rcu_read_lock();
3360 			rdev = rcu_dereference(conf->disks[i].rdev);
3361 			if (rdev && test_bit(In_sync, &rdev->flags) &&
3362 			    !test_bit(Faulty, &rdev->flags))
3363 				atomic_inc(&rdev->nr_pending);
3364 			else
3365 				rdev = NULL;
3366 			rcu_read_unlock();
3367 			if (rdev) {
3368 				if (!rdev_set_badblocks(
3369 					    rdev,
3370 					    sh->sector,
3371 					    STRIPE_SECTORS, 0))
3372 					md_error(conf->mddev, rdev);
3373 				rdev_dec_pending(rdev, conf->mddev);
3374 			}
3375 		}
3376 		spin_lock_irq(&sh->stripe_lock);
3377 		/* fail all writes first */
3378 		bi = sh->dev[i].towrite;
3379 		sh->dev[i].towrite = NULL;
3380 		sh->overwrite_disks = 0;
3381 		spin_unlock_irq(&sh->stripe_lock);
3382 		if (bi)
3383 			bitmap_end = 1;
3384 
3385 		log_stripe_write_finished(sh);
3386 
3387 		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3388 			wake_up(&conf->wait_for_overlap);
3389 
3390 		while (bi && bi->bi_iter.bi_sector <
3391 			sh->dev[i].sector + STRIPE_SECTORS) {
3392 			struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
3393 
3394 			md_write_end(conf->mddev);
3395 			bio_io_error(bi);
3396 			bi = nextbi;
3397 		}
3398 		if (bitmap_end)
3399 			md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3400 					   STRIPE_SECTORS, 0, 0);
3401 		bitmap_end = 0;
3402 		/* and fail all 'written' */
3403 		bi = sh->dev[i].written;
3404 		sh->dev[i].written = NULL;
3405 		if (test_and_clear_bit(R5_SkipCopy, &sh->dev[i].flags)) {
3406 			WARN_ON(test_bit(R5_UPTODATE, &sh->dev[i].flags));
3407 			sh->dev[i].page = sh->dev[i].orig_page;
3408 		}
3409 
3410 		if (bi) bitmap_end = 1;
3411 		while (bi && bi->bi_iter.bi_sector <
3412 		       sh->dev[i].sector + STRIPE_SECTORS) {
3413 			struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
3414 
3415 			md_write_end(conf->mddev);
3416 			bio_io_error(bi);
3417 			bi = bi2;
3418 		}
3419 
3420 		/* fail any reads if this device is non-operational and
3421 		 * the data has not reached the cache yet.
3422 		 */
3423 		if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
3424 		    s->failed > conf->max_degraded &&
3425 		    (!test_bit(R5_Insync, &sh->dev[i].flags) ||
3426 		      test_bit(R5_ReadError, &sh->dev[i].flags))) {
3427 			spin_lock_irq(&sh->stripe_lock);
3428 			bi = sh->dev[i].toread;
3429 			sh->dev[i].toread = NULL;
3430 			spin_unlock_irq(&sh->stripe_lock);
3431 			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
3432 				wake_up(&conf->wait_for_overlap);
3433 			if (bi)
3434 				s->to_read--;
3435 			while (bi && bi->bi_iter.bi_sector <
3436 			       sh->dev[i].sector + STRIPE_SECTORS) {
3437 				struct bio *nextbi =
3438 					r5_next_bio(bi, sh->dev[i].sector);
3439 
3440 				bio_io_error(bi);
3441 				bi = nextbi;
3442 			}
3443 		}
3444 		if (bitmap_end)
3445 			md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3446 					   STRIPE_SECTORS, 0, 0);
3447 		/* If we were in the middle of a write the parity block might
3448 		 * still be locked - so just clear all R5_LOCKED flags
3449 		 */
3450 		clear_bit(R5_LOCKED, &sh->dev[i].flags);
3451 	}
3452 	s->to_write = 0;
3453 	s->written = 0;
3454 
3455 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3456 		if (atomic_dec_and_test(&conf->pending_full_writes))
3457 			md_wakeup_thread(conf->mddev->thread);
3458 }
3459 
3460 static void
3461 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
3462 		   struct stripe_head_state *s)
3463 {
3464 	int abort = 0;
3465 	int i;
3466 
3467 	BUG_ON(sh->batch_head);
3468 	clear_bit(STRIPE_SYNCING, &sh->state);
3469 	if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
3470 		wake_up(&conf->wait_for_overlap);
3471 	s->syncing = 0;
3472 	s->replacing = 0;
3473 	/* There is nothing more to do for sync/check/repair.
3474 	 * Don't even need to abort as that is handled elsewhere
3475 	 * if needed, and not always wanted e.g. if there is a known
3476 	 * bad block here.
3477 	 * For recover/replace we need to record a bad block on all
3478 	 * non-sync devices, or abort the recovery
3479 	 */
3480 	if (test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery)) {
3481 		/* During recovery devices cannot be removed, so
3482 		 * locking and refcounting of rdevs is not needed
3483 		 */
3484 		rcu_read_lock();
3485 		for (i = 0; i < conf->raid_disks; i++) {
3486 			struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
3487 			if (rdev
3488 			    && !test_bit(Faulty, &rdev->flags)
3489 			    && !test_bit(In_sync, &rdev->flags)
3490 			    && !rdev_set_badblocks(rdev, sh->sector,
3491 						   STRIPE_SECTORS, 0))
3492 				abort = 1;
3493 			rdev = rcu_dereference(conf->disks[i].replacement);
3494 			if (rdev
3495 			    && !test_bit(Faulty, &rdev->flags)
3496 			    && !test_bit(In_sync, &rdev->flags)
3497 			    && !rdev_set_badblocks(rdev, sh->sector,
3498 						   STRIPE_SECTORS, 0))
3499 				abort = 1;
3500 		}
3501 		rcu_read_unlock();
3502 		if (abort)
3503 			conf->recovery_disabled =
3504 				conf->mddev->recovery_disabled;
3505 	}
3506 	md_done_sync(conf->mddev, STRIPE_SECTORS, !abort);
3507 }
3508 
3509 static int want_replace(struct stripe_head *sh, int disk_idx)
3510 {
3511 	struct md_rdev *rdev;
3512 	int rv = 0;
3513 
3514 	rcu_read_lock();
3515 	rdev = rcu_dereference(sh->raid_conf->disks[disk_idx].replacement);
3516 	if (rdev
3517 	    && !test_bit(Faulty, &rdev->flags)
3518 	    && !test_bit(In_sync, &rdev->flags)
3519 	    && (rdev->recovery_offset <= sh->sector
3520 		|| rdev->mddev->recovery_cp <= sh->sector))
3521 		rv = 1;
3522 	rcu_read_unlock();
3523 	return rv;
3524 }
3525 
3526 static int need_this_block(struct stripe_head *sh, struct stripe_head_state *s,
3527 			   int disk_idx, int disks)
3528 {
3529 	struct r5dev *dev = &sh->dev[disk_idx];
3530 	struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
3531 				  &sh->dev[s->failed_num[1]] };
3532 	int i;
3533 
3534 
3535 	if (test_bit(R5_LOCKED, &dev->flags) ||
3536 	    test_bit(R5_UPTODATE, &dev->flags))
3537 		/* No point reading this as we already have it or have
3538 		 * decided to get it.
3539 		 */
3540 		return 0;
3541 
3542 	if (dev->toread ||
3543 	    (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)))
3544 		/* We need this block to directly satisfy a request */
3545 		return 1;
3546 
3547 	if (s->syncing || s->expanding ||
3548 	    (s->replacing && want_replace(sh, disk_idx)))
3549 		/* When syncing, or expanding we read everything.
3550 		 * When replacing, we need the replaced block.
3551 		 */
3552 		return 1;
3553 
3554 	if ((s->failed >= 1 && fdev[0]->toread) ||
3555 	    (s->failed >= 2 && fdev[1]->toread))
3556 		/* If we want to read from a failed device, then
3557 		 * we need to actually read every other device.
3558 		 */
3559 		return 1;
3560 
3561 	/* Sometimes neither read-modify-write nor reconstruct-write
3562 	 * cycles can work.  In those cases we read every block we
3563 	 * can.  Then the parity-update is certain to have enough to
3564 	 * work with.
3565 	 * This can only be a problem when we need to write something,
3566 	 * and some device has failed.  If either of those tests
3567 	 * fail we need look no further.
3568 	 */
3569 	if (!s->failed || !s->to_write)
3570 		return 0;
3571 
3572 	if (test_bit(R5_Insync, &dev->flags) &&
3573 	    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3574 		/* Pre-reads at not permitted until after short delay
3575 		 * to gather multiple requests.  However if this
3576 		 * device is no Insync, the block could only be computed
3577 		 * and there is no need to delay that.
3578 		 */
3579 		return 0;
3580 
3581 	for (i = 0; i < s->failed && i < 2; i++) {
3582 		if (fdev[i]->towrite &&
3583 		    !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3584 		    !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3585 			/* If we have a partial write to a failed
3586 			 * device, then we will need to reconstruct
3587 			 * the content of that device, so all other
3588 			 * devices must be read.
3589 			 */
3590 			return 1;
3591 	}
3592 
3593 	/* If we are forced to do a reconstruct-write, either because
3594 	 * the current RAID6 implementation only supports that, or
3595 	 * because parity cannot be trusted and we are currently
3596 	 * recovering it, there is extra need to be careful.
3597 	 * If one of the devices that we would need to read, because
3598 	 * it is not being overwritten (and maybe not written at all)
3599 	 * is missing/faulty, then we need to read everything we can.
3600 	 */
3601 	if (sh->raid_conf->level != 6 &&
3602 	    sh->sector < sh->raid_conf->mddev->recovery_cp)
3603 		/* reconstruct-write isn't being forced */
3604 		return 0;
3605 	for (i = 0; i < s->failed && i < 2; i++) {
3606 		if (s->failed_num[i] != sh->pd_idx &&
3607 		    s->failed_num[i] != sh->qd_idx &&
3608 		    !test_bit(R5_UPTODATE, &fdev[i]->flags) &&
3609 		    !test_bit(R5_OVERWRITE, &fdev[i]->flags))
3610 			return 1;
3611 	}
3612 
3613 	return 0;
3614 }
3615 
3616 /* fetch_block - checks the given member device to see if its data needs
3617  * to be read or computed to satisfy a request.
3618  *
3619  * Returns 1 when no more member devices need to be checked, otherwise returns
3620  * 0 to tell the loop in handle_stripe_fill to continue
3621  */
3622 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
3623 		       int disk_idx, int disks)
3624 {
3625 	struct r5dev *dev = &sh->dev[disk_idx];
3626 
3627 	/* is the data in this block needed, and can we get it? */
3628 	if (need_this_block(sh, s, disk_idx, disks)) {
3629 		/* we would like to get this block, possibly by computing it,
3630 		 * otherwise read it if the backing disk is insync
3631 		 */
3632 		BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
3633 		BUG_ON(test_bit(R5_Wantread, &dev->flags));
3634 		BUG_ON(sh->batch_head);
3635 
3636 		/*
3637 		 * In the raid6 case if the only non-uptodate disk is P
3638 		 * then we already trusted P to compute the other failed
3639 		 * drives. It is safe to compute rather than re-read P.
3640 		 * In other cases we only compute blocks from failed
3641 		 * devices, otherwise check/repair might fail to detect
3642 		 * a real inconsistency.
3643 		 */
3644 
3645 		if ((s->uptodate == disks - 1) &&
3646 		    ((sh->qd_idx >= 0 && sh->pd_idx == disk_idx) ||
3647 		    (s->failed && (disk_idx == s->failed_num[0] ||
3648 				   disk_idx == s->failed_num[1])))) {
3649 			/* have disk failed, and we're requested to fetch it;
3650 			 * do compute it
3651 			 */
3652 			pr_debug("Computing stripe %llu block %d\n",
3653 			       (unsigned long long)sh->sector, disk_idx);
3654 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3655 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3656 			set_bit(R5_Wantcompute, &dev->flags);
3657 			sh->ops.target = disk_idx;
3658 			sh->ops.target2 = -1; /* no 2nd target */
3659 			s->req_compute = 1;
3660 			/* Careful: from this point on 'uptodate' is in the eye
3661 			 * of raid_run_ops which services 'compute' operations
3662 			 * before writes. R5_Wantcompute flags a block that will
3663 			 * be R5_UPTODATE by the time it is needed for a
3664 			 * subsequent operation.
3665 			 */
3666 			s->uptodate++;
3667 			return 1;
3668 		} else if (s->uptodate == disks-2 && s->failed >= 2) {
3669 			/* Computing 2-failure is *very* expensive; only
3670 			 * do it if failed >= 2
3671 			 */
3672 			int other;
3673 			for (other = disks; other--; ) {
3674 				if (other == disk_idx)
3675 					continue;
3676 				if (!test_bit(R5_UPTODATE,
3677 				      &sh->dev[other].flags))
3678 					break;
3679 			}
3680 			BUG_ON(other < 0);
3681 			pr_debug("Computing stripe %llu blocks %d,%d\n",
3682 			       (unsigned long long)sh->sector,
3683 			       disk_idx, other);
3684 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
3685 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
3686 			set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
3687 			set_bit(R5_Wantcompute, &sh->dev[other].flags);
3688 			sh->ops.target = disk_idx;
3689 			sh->ops.target2 = other;
3690 			s->uptodate += 2;
3691 			s->req_compute = 1;
3692 			return 1;
3693 		} else if (test_bit(R5_Insync, &dev->flags)) {
3694 			set_bit(R5_LOCKED, &dev->flags);
3695 			set_bit(R5_Wantread, &dev->flags);
3696 			s->locked++;
3697 			pr_debug("Reading block %d (sync=%d)\n",
3698 				disk_idx, s->syncing);
3699 		}
3700 	}
3701 
3702 	return 0;
3703 }
3704 
3705 /**
3706  * handle_stripe_fill - read or compute data to satisfy pending requests.
3707  */
3708 static void handle_stripe_fill(struct stripe_head *sh,
3709 			       struct stripe_head_state *s,
3710 			       int disks)
3711 {
3712 	int i;
3713 
3714 	/* look for blocks to read/compute, skip this if a compute
3715 	 * is already in flight, or if the stripe contents are in the
3716 	 * midst of changing due to a write
3717 	 */
3718 	if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
3719 	    !sh->reconstruct_state) {
3720 
3721 		/*
3722 		 * For degraded stripe with data in journal, do not handle
3723 		 * read requests yet, instead, flush the stripe to raid
3724 		 * disks first, this avoids handling complex rmw of write
3725 		 * back cache (prexor with orig_page, and then xor with
3726 		 * page) in the read path
3727 		 */
3728 		if (s->injournal && s->failed) {
3729 			if (test_bit(STRIPE_R5C_CACHING, &sh->state))
3730 				r5c_make_stripe_write_out(sh);
3731 			goto out;
3732 		}
3733 
3734 		for (i = disks; i--; )
3735 			if (fetch_block(sh, s, i, disks))
3736 				break;
3737 	}
3738 out:
3739 	set_bit(STRIPE_HANDLE, &sh->state);
3740 }
3741 
3742 static void break_stripe_batch_list(struct stripe_head *head_sh,
3743 				    unsigned long handle_flags);
3744 /* handle_stripe_clean_event
3745  * any written block on an uptodate or failed drive can be returned.
3746  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
3747  * never LOCKED, so we don't need to test 'failed' directly.
3748  */
3749 static void handle_stripe_clean_event(struct r5conf *conf,
3750 	struct stripe_head *sh, int disks)
3751 {
3752 	int i;
3753 	struct r5dev *dev;
3754 	int discard_pending = 0;
3755 	struct stripe_head *head_sh = sh;
3756 	bool do_endio = false;
3757 
3758 	for (i = disks; i--; )
3759 		if (sh->dev[i].written) {
3760 			dev = &sh->dev[i];
3761 			if (!test_bit(R5_LOCKED, &dev->flags) &&
3762 			    (test_bit(R5_UPTODATE, &dev->flags) ||
3763 			     test_bit(R5_Discard, &dev->flags) ||
3764 			     test_bit(R5_SkipCopy, &dev->flags))) {
3765 				/* We can return any write requests */
3766 				struct bio *wbi, *wbi2;
3767 				pr_debug("Return write for disc %d\n", i);
3768 				if (test_and_clear_bit(R5_Discard, &dev->flags))
3769 					clear_bit(R5_UPTODATE, &dev->flags);
3770 				if (test_and_clear_bit(R5_SkipCopy, &dev->flags)) {
3771 					WARN_ON(test_bit(R5_UPTODATE, &dev->flags));
3772 				}
3773 				do_endio = true;
3774 
3775 returnbi:
3776 				dev->page = dev->orig_page;
3777 				wbi = dev->written;
3778 				dev->written = NULL;
3779 				while (wbi && wbi->bi_iter.bi_sector <
3780 					dev->sector + STRIPE_SECTORS) {
3781 					wbi2 = r5_next_bio(wbi, dev->sector);
3782 					md_write_end(conf->mddev);
3783 					bio_endio(wbi);
3784 					wbi = wbi2;
3785 				}
3786 				md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
3787 						   STRIPE_SECTORS,
3788 						   !test_bit(STRIPE_DEGRADED, &sh->state),
3789 						   0);
3790 				if (head_sh->batch_head) {
3791 					sh = list_first_entry(&sh->batch_list,
3792 							      struct stripe_head,
3793 							      batch_list);
3794 					if (sh != head_sh) {
3795 						dev = &sh->dev[i];
3796 						goto returnbi;
3797 					}
3798 				}
3799 				sh = head_sh;
3800 				dev = &sh->dev[i];
3801 			} else if (test_bit(R5_Discard, &dev->flags))
3802 				discard_pending = 1;
3803 		}
3804 
3805 	log_stripe_write_finished(sh);
3806 
3807 	if (!discard_pending &&
3808 	    test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags)) {
3809 		int hash;
3810 		clear_bit(R5_Discard, &sh->dev[sh->pd_idx].flags);
3811 		clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
3812 		if (sh->qd_idx >= 0) {
3813 			clear_bit(R5_Discard, &sh->dev[sh->qd_idx].flags);
3814 			clear_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags);
3815 		}
3816 		/* now that discard is done we can proceed with any sync */
3817 		clear_bit(STRIPE_DISCARD, &sh->state);
3818 		/*
3819 		 * SCSI discard will change some bio fields and the stripe has
3820 		 * no updated data, so remove it from hash list and the stripe
3821 		 * will be reinitialized
3822 		 */
3823 unhash:
3824 		hash = sh->hash_lock_index;
3825 		spin_lock_irq(conf->hash_locks + hash);
3826 		remove_hash(sh);
3827 		spin_unlock_irq(conf->hash_locks + hash);
3828 		if (head_sh->batch_head) {
3829 			sh = list_first_entry(&sh->batch_list,
3830 					      struct stripe_head, batch_list);
3831 			if (sh != head_sh)
3832 					goto unhash;
3833 		}
3834 		sh = head_sh;
3835 
3836 		if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
3837 			set_bit(STRIPE_HANDLE, &sh->state);
3838 
3839 	}
3840 
3841 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
3842 		if (atomic_dec_and_test(&conf->pending_full_writes))
3843 			md_wakeup_thread(conf->mddev->thread);
3844 
3845 	if (head_sh->batch_head && do_endio)
3846 		break_stripe_batch_list(head_sh, STRIPE_EXPAND_SYNC_FLAGS);
3847 }
3848 
3849 /*
3850  * For RMW in write back cache, we need extra page in prexor to store the
3851  * old data. This page is stored in dev->orig_page.
3852  *
3853  * This function checks whether we have data for prexor. The exact logic
3854  * is:
3855  *       R5_UPTODATE && (!R5_InJournal || R5_OrigPageUPTDODATE)
3856  */
3857 static inline bool uptodate_for_rmw(struct r5dev *dev)
3858 {
3859 	return (test_bit(R5_UPTODATE, &dev->flags)) &&
3860 		(!test_bit(R5_InJournal, &dev->flags) ||
3861 		 test_bit(R5_OrigPageUPTDODATE, &dev->flags));
3862 }
3863 
3864 static int handle_stripe_dirtying(struct r5conf *conf,
3865 				  struct stripe_head *sh,
3866 				  struct stripe_head_state *s,
3867 				  int disks)
3868 {
3869 	int rmw = 0, rcw = 0, i;
3870 	sector_t recovery_cp = conf->mddev->recovery_cp;
3871 
3872 	/* Check whether resync is now happening or should start.
3873 	 * If yes, then the array is dirty (after unclean shutdown or
3874 	 * initial creation), so parity in some stripes might be inconsistent.
3875 	 * In this case, we need to always do reconstruct-write, to ensure
3876 	 * that in case of drive failure or read-error correction, we
3877 	 * generate correct data from the parity.
3878 	 */
3879 	if (conf->rmw_level == PARITY_DISABLE_RMW ||
3880 	    (recovery_cp < MaxSector && sh->sector >= recovery_cp &&
3881 	     s->failed == 0)) {
3882 		/* Calculate the real rcw later - for now make it
3883 		 * look like rcw is cheaper
3884 		 */
3885 		rcw = 1; rmw = 2;
3886 		pr_debug("force RCW rmw_level=%u, recovery_cp=%llu sh->sector=%llu\n",
3887 			 conf->rmw_level, (unsigned long long)recovery_cp,
3888 			 (unsigned long long)sh->sector);
3889 	} else for (i = disks; i--; ) {
3890 		/* would I have to read this buffer for read_modify_write */
3891 		struct r5dev *dev = &sh->dev[i];
3892 		if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3893 		     i == sh->pd_idx || i == sh->qd_idx ||
3894 		     test_bit(R5_InJournal, &dev->flags)) &&
3895 		    !test_bit(R5_LOCKED, &dev->flags) &&
3896 		    !(uptodate_for_rmw(dev) ||
3897 		      test_bit(R5_Wantcompute, &dev->flags))) {
3898 			if (test_bit(R5_Insync, &dev->flags))
3899 				rmw++;
3900 			else
3901 				rmw += 2*disks;  /* cannot read it */
3902 		}
3903 		/* Would I have to read this buffer for reconstruct_write */
3904 		if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3905 		    i != sh->pd_idx && i != sh->qd_idx &&
3906 		    !test_bit(R5_LOCKED, &dev->flags) &&
3907 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
3908 		      test_bit(R5_Wantcompute, &dev->flags))) {
3909 			if (test_bit(R5_Insync, &dev->flags))
3910 				rcw++;
3911 			else
3912 				rcw += 2*disks;
3913 		}
3914 	}
3915 
3916 	pr_debug("for sector %llu state 0x%lx, rmw=%d rcw=%d\n",
3917 		 (unsigned long long)sh->sector, sh->state, rmw, rcw);
3918 	set_bit(STRIPE_HANDLE, &sh->state);
3919 	if ((rmw < rcw || (rmw == rcw && conf->rmw_level == PARITY_PREFER_RMW)) && rmw > 0) {
3920 		/* prefer read-modify-write, but need to get some data */
3921 		if (conf->mddev->queue)
3922 			blk_add_trace_msg(conf->mddev->queue,
3923 					  "raid5 rmw %llu %d",
3924 					  (unsigned long long)sh->sector, rmw);
3925 		for (i = disks; i--; ) {
3926 			struct r5dev *dev = &sh->dev[i];
3927 			if (test_bit(R5_InJournal, &dev->flags) &&
3928 			    dev->page == dev->orig_page &&
3929 			    !test_bit(R5_LOCKED, &sh->dev[sh->pd_idx].flags)) {
3930 				/* alloc page for prexor */
3931 				struct page *p = alloc_page(GFP_NOIO);
3932 
3933 				if (p) {
3934 					dev->orig_page = p;
3935 					continue;
3936 				}
3937 
3938 				/*
3939 				 * alloc_page() failed, try use
3940 				 * disk_info->extra_page
3941 				 */
3942 				if (!test_and_set_bit(R5C_EXTRA_PAGE_IN_USE,
3943 						      &conf->cache_state)) {
3944 					r5c_use_extra_page(sh);
3945 					break;
3946 				}
3947 
3948 				/* extra_page in use, add to delayed_list */
3949 				set_bit(STRIPE_DELAYED, &sh->state);
3950 				s->waiting_extra_page = 1;
3951 				return -EAGAIN;
3952 			}
3953 		}
3954 
3955 		for (i = disks; i--; ) {
3956 			struct r5dev *dev = &sh->dev[i];
3957 			if (((dev->towrite && !delay_towrite(conf, dev, s)) ||
3958 			     i == sh->pd_idx || i == sh->qd_idx ||
3959 			     test_bit(R5_InJournal, &dev->flags)) &&
3960 			    !test_bit(R5_LOCKED, &dev->flags) &&
3961 			    !(uptodate_for_rmw(dev) ||
3962 			      test_bit(R5_Wantcompute, &dev->flags)) &&
3963 			    test_bit(R5_Insync, &dev->flags)) {
3964 				if (test_bit(STRIPE_PREREAD_ACTIVE,
3965 					     &sh->state)) {
3966 					pr_debug("Read_old block %d for r-m-w\n",
3967 						 i);
3968 					set_bit(R5_LOCKED, &dev->flags);
3969 					set_bit(R5_Wantread, &dev->flags);
3970 					s->locked++;
3971 				} else {
3972 					set_bit(STRIPE_DELAYED, &sh->state);
3973 					set_bit(STRIPE_HANDLE, &sh->state);
3974 				}
3975 			}
3976 		}
3977 	}
3978 	if ((rcw < rmw || (rcw == rmw && conf->rmw_level != PARITY_PREFER_RMW)) && rcw > 0) {
3979 		/* want reconstruct write, but need to get some data */
3980 		int qread =0;
3981 		rcw = 0;
3982 		for (i = disks; i--; ) {
3983 			struct r5dev *dev = &sh->dev[i];
3984 			if (!test_bit(R5_OVERWRITE, &dev->flags) &&
3985 			    i != sh->pd_idx && i != sh->qd_idx &&
3986 			    !test_bit(R5_LOCKED, &dev->flags) &&
3987 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
3988 			      test_bit(R5_Wantcompute, &dev->flags))) {
3989 				rcw++;
3990 				if (test_bit(R5_Insync, &dev->flags) &&
3991 				    test_bit(STRIPE_PREREAD_ACTIVE,
3992 					     &sh->state)) {
3993 					pr_debug("Read_old block "
3994 						"%d for Reconstruct\n", i);
3995 					set_bit(R5_LOCKED, &dev->flags);
3996 					set_bit(R5_Wantread, &dev->flags);
3997 					s->locked++;
3998 					qread++;
3999 				} else {
4000 					set_bit(STRIPE_DELAYED, &sh->state);
4001 					set_bit(STRIPE_HANDLE, &sh->state);
4002 				}
4003 			}
4004 		}
4005 		if (rcw && conf->mddev->queue)
4006 			blk_add_trace_msg(conf->mddev->queue, "raid5 rcw %llu %d %d %d",
4007 					  (unsigned long long)sh->sector,
4008 					  rcw, qread, test_bit(STRIPE_DELAYED, &sh->state));
4009 	}
4010 
4011 	if (rcw > disks && rmw > disks &&
4012 	    !test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4013 		set_bit(STRIPE_DELAYED, &sh->state);
4014 
4015 	/* now if nothing is locked, and if we have enough data,
4016 	 * we can start a write request
4017 	 */
4018 	/* since handle_stripe can be called at any time we need to handle the
4019 	 * case where a compute block operation has been submitted and then a
4020 	 * subsequent call wants to start a write request.  raid_run_ops only
4021 	 * handles the case where compute block and reconstruct are requested
4022 	 * simultaneously.  If this is not the case then new writes need to be
4023 	 * held off until the compute completes.
4024 	 */
4025 	if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
4026 	    (s->locked == 0 && (rcw == 0 || rmw == 0) &&
4027 	     !test_bit(STRIPE_BIT_DELAY, &sh->state)))
4028 		schedule_reconstruction(sh, s, rcw == 0, 0);
4029 	return 0;
4030 }
4031 
4032 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
4033 				struct stripe_head_state *s, int disks)
4034 {
4035 	struct r5dev *dev = NULL;
4036 
4037 	BUG_ON(sh->batch_head);
4038 	set_bit(STRIPE_HANDLE, &sh->state);
4039 
4040 	switch (sh->check_state) {
4041 	case check_state_idle:
4042 		/* start a new check operation if there are no failures */
4043 		if (s->failed == 0) {
4044 			BUG_ON(s->uptodate != disks);
4045 			sh->check_state = check_state_run;
4046 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
4047 			clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
4048 			s->uptodate--;
4049 			break;
4050 		}
4051 		dev = &sh->dev[s->failed_num[0]];
4052 		/* fall through */
4053 	case check_state_compute_result:
4054 		sh->check_state = check_state_idle;
4055 		if (!dev)
4056 			dev = &sh->dev[sh->pd_idx];
4057 
4058 		/* check that a write has not made the stripe insync */
4059 		if (test_bit(STRIPE_INSYNC, &sh->state))
4060 			break;
4061 
4062 		/* either failed parity check, or recovery is happening */
4063 		BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
4064 		BUG_ON(s->uptodate != disks);
4065 
4066 		set_bit(R5_LOCKED, &dev->flags);
4067 		s->locked++;
4068 		set_bit(R5_Wantwrite, &dev->flags);
4069 
4070 		clear_bit(STRIPE_DEGRADED, &sh->state);
4071 		set_bit(STRIPE_INSYNC, &sh->state);
4072 		break;
4073 	case check_state_run:
4074 		break; /* we will be called again upon completion */
4075 	case check_state_check_result:
4076 		sh->check_state = check_state_idle;
4077 
4078 		/* if a failure occurred during the check operation, leave
4079 		 * STRIPE_INSYNC not set and let the stripe be handled again
4080 		 */
4081 		if (s->failed)
4082 			break;
4083 
4084 		/* handle a successful check operation, if parity is correct
4085 		 * we are done.  Otherwise update the mismatch count and repair
4086 		 * parity if !MD_RECOVERY_CHECK
4087 		 */
4088 		if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
4089 			/* parity is correct (on disc,
4090 			 * not in buffer any more)
4091 			 */
4092 			set_bit(STRIPE_INSYNC, &sh->state);
4093 		else {
4094 			atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4095 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4096 				/* don't try to repair!! */
4097 				set_bit(STRIPE_INSYNC, &sh->state);
4098 				pr_warn_ratelimited("%s: mismatch sector in range "
4099 						    "%llu-%llu\n", mdname(conf->mddev),
4100 						    (unsigned long long) sh->sector,
4101 						    (unsigned long long) sh->sector +
4102 						    STRIPE_SECTORS);
4103 			} else {
4104 				sh->check_state = check_state_compute_run;
4105 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4106 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4107 				set_bit(R5_Wantcompute,
4108 					&sh->dev[sh->pd_idx].flags);
4109 				sh->ops.target = sh->pd_idx;
4110 				sh->ops.target2 = -1;
4111 				s->uptodate++;
4112 			}
4113 		}
4114 		break;
4115 	case check_state_compute_run:
4116 		break;
4117 	default:
4118 		pr_err("%s: unknown check_state: %d sector: %llu\n",
4119 		       __func__, sh->check_state,
4120 		       (unsigned long long) sh->sector);
4121 		BUG();
4122 	}
4123 }
4124 
4125 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
4126 				  struct stripe_head_state *s,
4127 				  int disks)
4128 {
4129 	int pd_idx = sh->pd_idx;
4130 	int qd_idx = sh->qd_idx;
4131 	struct r5dev *dev;
4132 
4133 	BUG_ON(sh->batch_head);
4134 	set_bit(STRIPE_HANDLE, &sh->state);
4135 
4136 	BUG_ON(s->failed > 2);
4137 
4138 	/* Want to check and possibly repair P and Q.
4139 	 * However there could be one 'failed' device, in which
4140 	 * case we can only check one of them, possibly using the
4141 	 * other to generate missing data
4142 	 */
4143 
4144 	switch (sh->check_state) {
4145 	case check_state_idle:
4146 		/* start a new check operation if there are < 2 failures */
4147 		if (s->failed == s->q_failed) {
4148 			/* The only possible failed device holds Q, so it
4149 			 * makes sense to check P (If anything else were failed,
4150 			 * we would have used P to recreate it).
4151 			 */
4152 			sh->check_state = check_state_run;
4153 		}
4154 		if (!s->q_failed && s->failed < 2) {
4155 			/* Q is not failed, and we didn't use it to generate
4156 			 * anything, so it makes sense to check it
4157 			 */
4158 			if (sh->check_state == check_state_run)
4159 				sh->check_state = check_state_run_pq;
4160 			else
4161 				sh->check_state = check_state_run_q;
4162 		}
4163 
4164 		/* discard potentially stale zero_sum_result */
4165 		sh->ops.zero_sum_result = 0;
4166 
4167 		if (sh->check_state == check_state_run) {
4168 			/* async_xor_zero_sum destroys the contents of P */
4169 			clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
4170 			s->uptodate--;
4171 		}
4172 		if (sh->check_state >= check_state_run &&
4173 		    sh->check_state <= check_state_run_pq) {
4174 			/* async_syndrome_zero_sum preserves P and Q, so
4175 			 * no need to mark them !uptodate here
4176 			 */
4177 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
4178 			break;
4179 		}
4180 
4181 		/* we have 2-disk failure */
4182 		BUG_ON(s->failed != 2);
4183 		/* fall through */
4184 	case check_state_compute_result:
4185 		sh->check_state = check_state_idle;
4186 
4187 		/* check that a write has not made the stripe insync */
4188 		if (test_bit(STRIPE_INSYNC, &sh->state))
4189 			break;
4190 
4191 		/* now write out any block on a failed drive,
4192 		 * or P or Q if they were recomputed
4193 		 */
4194 		dev = NULL;
4195 		if (s->failed == 2) {
4196 			dev = &sh->dev[s->failed_num[1]];
4197 			s->locked++;
4198 			set_bit(R5_LOCKED, &dev->flags);
4199 			set_bit(R5_Wantwrite, &dev->flags);
4200 		}
4201 		if (s->failed >= 1) {
4202 			dev = &sh->dev[s->failed_num[0]];
4203 			s->locked++;
4204 			set_bit(R5_LOCKED, &dev->flags);
4205 			set_bit(R5_Wantwrite, &dev->flags);
4206 		}
4207 		if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4208 			dev = &sh->dev[pd_idx];
4209 			s->locked++;
4210 			set_bit(R5_LOCKED, &dev->flags);
4211 			set_bit(R5_Wantwrite, &dev->flags);
4212 		}
4213 		if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4214 			dev = &sh->dev[qd_idx];
4215 			s->locked++;
4216 			set_bit(R5_LOCKED, &dev->flags);
4217 			set_bit(R5_Wantwrite, &dev->flags);
4218 		}
4219 		if (WARN_ONCE(dev && !test_bit(R5_UPTODATE, &dev->flags),
4220 			      "%s: disk%td not up to date\n",
4221 			      mdname(conf->mddev),
4222 			      dev - (struct r5dev *) &sh->dev)) {
4223 			clear_bit(R5_LOCKED, &dev->flags);
4224 			clear_bit(R5_Wantwrite, &dev->flags);
4225 			s->locked--;
4226 		}
4227 		clear_bit(STRIPE_DEGRADED, &sh->state);
4228 
4229 		set_bit(STRIPE_INSYNC, &sh->state);
4230 		break;
4231 	case check_state_run:
4232 	case check_state_run_q:
4233 	case check_state_run_pq:
4234 		break; /* we will be called again upon completion */
4235 	case check_state_check_result:
4236 		sh->check_state = check_state_idle;
4237 
4238 		/* handle a successful check operation, if parity is correct
4239 		 * we are done.  Otherwise update the mismatch count and repair
4240 		 * parity if !MD_RECOVERY_CHECK
4241 		 */
4242 		if (sh->ops.zero_sum_result == 0) {
4243 			/* both parities are correct */
4244 			if (!s->failed)
4245 				set_bit(STRIPE_INSYNC, &sh->state);
4246 			else {
4247 				/* in contrast to the raid5 case we can validate
4248 				 * parity, but still have a failure to write
4249 				 * back
4250 				 */
4251 				sh->check_state = check_state_compute_result;
4252 				/* Returning at this point means that we may go
4253 				 * off and bring p and/or q uptodate again so
4254 				 * we make sure to check zero_sum_result again
4255 				 * to verify if p or q need writeback
4256 				 */
4257 			}
4258 		} else {
4259 			atomic64_add(STRIPE_SECTORS, &conf->mddev->resync_mismatches);
4260 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery)) {
4261 				/* don't try to repair!! */
4262 				set_bit(STRIPE_INSYNC, &sh->state);
4263 				pr_warn_ratelimited("%s: mismatch sector in range "
4264 						    "%llu-%llu\n", mdname(conf->mddev),
4265 						    (unsigned long long) sh->sector,
4266 						    (unsigned long long) sh->sector +
4267 						    STRIPE_SECTORS);
4268 			} else {
4269 				int *target = &sh->ops.target;
4270 
4271 				sh->ops.target = -1;
4272 				sh->ops.target2 = -1;
4273 				sh->check_state = check_state_compute_run;
4274 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
4275 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
4276 				if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
4277 					set_bit(R5_Wantcompute,
4278 						&sh->dev[pd_idx].flags);
4279 					*target = pd_idx;
4280 					target = &sh->ops.target2;
4281 					s->uptodate++;
4282 				}
4283 				if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
4284 					set_bit(R5_Wantcompute,
4285 						&sh->dev[qd_idx].flags);
4286 					*target = qd_idx;
4287 					s->uptodate++;
4288 				}
4289 			}
4290 		}
4291 		break;
4292 	case check_state_compute_run:
4293 		break;
4294 	default:
4295 		pr_warn("%s: unknown check_state: %d sector: %llu\n",
4296 			__func__, sh->check_state,
4297 			(unsigned long long) sh->sector);
4298 		BUG();
4299 	}
4300 }
4301 
4302 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
4303 {
4304 	int i;
4305 
4306 	/* We have read all the blocks in this stripe and now we need to
4307 	 * copy some of them into a target stripe for expand.
4308 	 */
4309 	struct dma_async_tx_descriptor *tx = NULL;
4310 	BUG_ON(sh->batch_head);
4311 	clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4312 	for (i = 0; i < sh->disks; i++)
4313 		if (i != sh->pd_idx && i != sh->qd_idx) {
4314 			int dd_idx, j;
4315 			struct stripe_head *sh2;
4316 			struct async_submit_ctl submit;
4317 
4318 			sector_t bn = raid5_compute_blocknr(sh, i, 1);
4319 			sector_t s = raid5_compute_sector(conf, bn, 0,
4320 							  &dd_idx, NULL);
4321 			sh2 = raid5_get_active_stripe(conf, s, 0, 1, 1);
4322 			if (sh2 == NULL)
4323 				/* so far only the early blocks of this stripe
4324 				 * have been requested.  When later blocks
4325 				 * get requested, we will try again
4326 				 */
4327 				continue;
4328 			if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
4329 			   test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
4330 				/* must have already done this block */
4331 				raid5_release_stripe(sh2);
4332 				continue;
4333 			}
4334 
4335 			/* place all the copies on one channel */
4336 			init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
4337 			tx = async_memcpy(sh2->dev[dd_idx].page,
4338 					  sh->dev[i].page, 0, 0, STRIPE_SIZE,
4339 					  &submit);
4340 
4341 			set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
4342 			set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
4343 			for (j = 0; j < conf->raid_disks; j++)
4344 				if (j != sh2->pd_idx &&
4345 				    j != sh2->qd_idx &&
4346 				    !test_bit(R5_Expanded, &sh2->dev[j].flags))
4347 					break;
4348 			if (j == conf->raid_disks) {
4349 				set_bit(STRIPE_EXPAND_READY, &sh2->state);
4350 				set_bit(STRIPE_HANDLE, &sh2->state);
4351 			}
4352 			raid5_release_stripe(sh2);
4353 
4354 		}
4355 	/* done submitting copies, wait for them to complete */
4356 	async_tx_quiesce(&tx);
4357 }
4358 
4359 /*
4360  * handle_stripe - do things to a stripe.
4361  *
4362  * We lock the stripe by setting STRIPE_ACTIVE and then examine the
4363  * state of various bits to see what needs to be done.
4364  * Possible results:
4365  *    return some read requests which now have data
4366  *    return some write requests which are safely on storage
4367  *    schedule a read on some buffers
4368  *    schedule a write of some buffers
4369  *    return confirmation of parity correctness
4370  *
4371  */
4372 
4373 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
4374 {
4375 	struct r5conf *conf = sh->raid_conf;
4376 	int disks = sh->disks;
4377 	struct r5dev *dev;
4378 	int i;
4379 	int do_recovery = 0;
4380 
4381 	memset(s, 0, sizeof(*s));
4382 
4383 	s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state) && !sh->batch_head;
4384 	s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state) && !sh->batch_head;
4385 	s->failed_num[0] = -1;
4386 	s->failed_num[1] = -1;
4387 	s->log_failed = r5l_log_disk_error(conf);
4388 
4389 	/* Now to look around and see what can be done */
4390 	rcu_read_lock();
4391 	for (i=disks; i--; ) {
4392 		struct md_rdev *rdev;
4393 		sector_t first_bad;
4394 		int bad_sectors;
4395 		int is_bad = 0;
4396 
4397 		dev = &sh->dev[i];
4398 
4399 		pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
4400 			 i, dev->flags,
4401 			 dev->toread, dev->towrite, dev->written);
4402 		/* maybe we can reply to a read
4403 		 *
4404 		 * new wantfill requests are only permitted while
4405 		 * ops_complete_biofill is guaranteed to be inactive
4406 		 */
4407 		if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
4408 		    !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
4409 			set_bit(R5_Wantfill, &dev->flags);
4410 
4411 		/* now count some things */
4412 		if (test_bit(R5_LOCKED, &dev->flags))
4413 			s->locked++;
4414 		if (test_bit(R5_UPTODATE, &dev->flags))
4415 			s->uptodate++;
4416 		if (test_bit(R5_Wantcompute, &dev->flags)) {
4417 			s->compute++;
4418 			BUG_ON(s->compute > 2);
4419 		}
4420 
4421 		if (test_bit(R5_Wantfill, &dev->flags))
4422 			s->to_fill++;
4423 		else if (dev->toread)
4424 			s->to_read++;
4425 		if (dev->towrite) {
4426 			s->to_write++;
4427 			if (!test_bit(R5_OVERWRITE, &dev->flags))
4428 				s->non_overwrite++;
4429 		}
4430 		if (dev->written)
4431 			s->written++;
4432 		/* Prefer to use the replacement for reads, but only
4433 		 * if it is recovered enough and has no bad blocks.
4434 		 */
4435 		rdev = rcu_dereference(conf->disks[i].replacement);
4436 		if (rdev && !test_bit(Faulty, &rdev->flags) &&
4437 		    rdev->recovery_offset >= sh->sector + STRIPE_SECTORS &&
4438 		    !is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4439 				 &first_bad, &bad_sectors))
4440 			set_bit(R5_ReadRepl, &dev->flags);
4441 		else {
4442 			if (rdev && !test_bit(Faulty, &rdev->flags))
4443 				set_bit(R5_NeedReplace, &dev->flags);
4444 			else
4445 				clear_bit(R5_NeedReplace, &dev->flags);
4446 			rdev = rcu_dereference(conf->disks[i].rdev);
4447 			clear_bit(R5_ReadRepl, &dev->flags);
4448 		}
4449 		if (rdev && test_bit(Faulty, &rdev->flags))
4450 			rdev = NULL;
4451 		if (rdev) {
4452 			is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
4453 					     &first_bad, &bad_sectors);
4454 			if (s->blocked_rdev == NULL
4455 			    && (test_bit(Blocked, &rdev->flags)
4456 				|| is_bad < 0)) {
4457 				if (is_bad < 0)
4458 					set_bit(BlockedBadBlocks,
4459 						&rdev->flags);
4460 				s->blocked_rdev = rdev;
4461 				atomic_inc(&rdev->nr_pending);
4462 			}
4463 		}
4464 		clear_bit(R5_Insync, &dev->flags);
4465 		if (!rdev)
4466 			/* Not in-sync */;
4467 		else if (is_bad) {
4468 			/* also not in-sync */
4469 			if (!test_bit(WriteErrorSeen, &rdev->flags) &&
4470 			    test_bit(R5_UPTODATE, &dev->flags)) {
4471 				/* treat as in-sync, but with a read error
4472 				 * which we can now try to correct
4473 				 */
4474 				set_bit(R5_Insync, &dev->flags);
4475 				set_bit(R5_ReadError, &dev->flags);
4476 			}
4477 		} else if (test_bit(In_sync, &rdev->flags))
4478 			set_bit(R5_Insync, &dev->flags);
4479 		else if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
4480 			/* in sync if before recovery_offset */
4481 			set_bit(R5_Insync, &dev->flags);
4482 		else if (test_bit(R5_UPTODATE, &dev->flags) &&
4483 			 test_bit(R5_Expanded, &dev->flags))
4484 			/* If we've reshaped into here, we assume it is Insync.
4485 			 * We will shortly update recovery_offset to make
4486 			 * it official.
4487 			 */
4488 			set_bit(R5_Insync, &dev->flags);
4489 
4490 		if (test_bit(R5_WriteError, &dev->flags)) {
4491 			/* This flag does not apply to '.replacement'
4492 			 * only to .rdev, so make sure to check that*/
4493 			struct md_rdev *rdev2 = rcu_dereference(
4494 				conf->disks[i].rdev);
4495 			if (rdev2 == rdev)
4496 				clear_bit(R5_Insync, &dev->flags);
4497 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4498 				s->handle_bad_blocks = 1;
4499 				atomic_inc(&rdev2->nr_pending);
4500 			} else
4501 				clear_bit(R5_WriteError, &dev->flags);
4502 		}
4503 		if (test_bit(R5_MadeGood, &dev->flags)) {
4504 			/* This flag does not apply to '.replacement'
4505 			 * only to .rdev, so make sure to check that*/
4506 			struct md_rdev *rdev2 = rcu_dereference(
4507 				conf->disks[i].rdev);
4508 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4509 				s->handle_bad_blocks = 1;
4510 				atomic_inc(&rdev2->nr_pending);
4511 			} else
4512 				clear_bit(R5_MadeGood, &dev->flags);
4513 		}
4514 		if (test_bit(R5_MadeGoodRepl, &dev->flags)) {
4515 			struct md_rdev *rdev2 = rcu_dereference(
4516 				conf->disks[i].replacement);
4517 			if (rdev2 && !test_bit(Faulty, &rdev2->flags)) {
4518 				s->handle_bad_blocks = 1;
4519 				atomic_inc(&rdev2->nr_pending);
4520 			} else
4521 				clear_bit(R5_MadeGoodRepl, &dev->flags);
4522 		}
4523 		if (!test_bit(R5_Insync, &dev->flags)) {
4524 			/* The ReadError flag will just be confusing now */
4525 			clear_bit(R5_ReadError, &dev->flags);
4526 			clear_bit(R5_ReWrite, &dev->flags);
4527 		}
4528 		if (test_bit(R5_ReadError, &dev->flags))
4529 			clear_bit(R5_Insync, &dev->flags);
4530 		if (!test_bit(R5_Insync, &dev->flags)) {
4531 			if (s->failed < 2)
4532 				s->failed_num[s->failed] = i;
4533 			s->failed++;
4534 			if (rdev && !test_bit(Faulty, &rdev->flags))
4535 				do_recovery = 1;
4536 			else if (!rdev) {
4537 				rdev = rcu_dereference(
4538 				    conf->disks[i].replacement);
4539 				if (rdev && !test_bit(Faulty, &rdev->flags))
4540 					do_recovery = 1;
4541 			}
4542 		}
4543 
4544 		if (test_bit(R5_InJournal, &dev->flags))
4545 			s->injournal++;
4546 		if (test_bit(R5_InJournal, &dev->flags) && dev->written)
4547 			s->just_cached++;
4548 	}
4549 	if (test_bit(STRIPE_SYNCING, &sh->state)) {
4550 		/* If there is a failed device being replaced,
4551 		 *     we must be recovering.
4552 		 * else if we are after recovery_cp, we must be syncing
4553 		 * else if MD_RECOVERY_REQUESTED is set, we also are syncing.
4554 		 * else we can only be replacing
4555 		 * sync and recovery both need to read all devices, and so
4556 		 * use the same flag.
4557 		 */
4558 		if (do_recovery ||
4559 		    sh->sector >= conf->mddev->recovery_cp ||
4560 		    test_bit(MD_RECOVERY_REQUESTED, &(conf->mddev->recovery)))
4561 			s->syncing = 1;
4562 		else
4563 			s->replacing = 1;
4564 	}
4565 	rcu_read_unlock();
4566 }
4567 
4568 static int clear_batch_ready(struct stripe_head *sh)
4569 {
4570 	/* Return '1' if this is a member of batch, or
4571 	 * '0' if it is a lone stripe or a head which can now be
4572 	 * handled.
4573 	 */
4574 	struct stripe_head *tmp;
4575 	if (!test_and_clear_bit(STRIPE_BATCH_READY, &sh->state))
4576 		return (sh->batch_head && sh->batch_head != sh);
4577 	spin_lock(&sh->stripe_lock);
4578 	if (!sh->batch_head) {
4579 		spin_unlock(&sh->stripe_lock);
4580 		return 0;
4581 	}
4582 
4583 	/*
4584 	 * this stripe could be added to a batch list before we check
4585 	 * BATCH_READY, skips it
4586 	 */
4587 	if (sh->batch_head != sh) {
4588 		spin_unlock(&sh->stripe_lock);
4589 		return 1;
4590 	}
4591 	spin_lock(&sh->batch_lock);
4592 	list_for_each_entry(tmp, &sh->batch_list, batch_list)
4593 		clear_bit(STRIPE_BATCH_READY, &tmp->state);
4594 	spin_unlock(&sh->batch_lock);
4595 	spin_unlock(&sh->stripe_lock);
4596 
4597 	/*
4598 	 * BATCH_READY is cleared, no new stripes can be added.
4599 	 * batch_list can be accessed without lock
4600 	 */
4601 	return 0;
4602 }
4603 
4604 static void break_stripe_batch_list(struct stripe_head *head_sh,
4605 				    unsigned long handle_flags)
4606 {
4607 	struct stripe_head *sh, *next;
4608 	int i;
4609 	int do_wakeup = 0;
4610 
4611 	list_for_each_entry_safe(sh, next, &head_sh->batch_list, batch_list) {
4612 
4613 		list_del_init(&sh->batch_list);
4614 
4615 		WARN_ONCE(sh->state & ((1 << STRIPE_ACTIVE) |
4616 					  (1 << STRIPE_SYNCING) |
4617 					  (1 << STRIPE_REPLACED) |
4618 					  (1 << STRIPE_DELAYED) |
4619 					  (1 << STRIPE_BIT_DELAY) |
4620 					  (1 << STRIPE_FULL_WRITE) |
4621 					  (1 << STRIPE_BIOFILL_RUN) |
4622 					  (1 << STRIPE_COMPUTE_RUN)  |
4623 					  (1 << STRIPE_OPS_REQ_PENDING) |
4624 					  (1 << STRIPE_DISCARD) |
4625 					  (1 << STRIPE_BATCH_READY) |
4626 					  (1 << STRIPE_BATCH_ERR) |
4627 					  (1 << STRIPE_BITMAP_PENDING)),
4628 			"stripe state: %lx\n", sh->state);
4629 		WARN_ONCE(head_sh->state & ((1 << STRIPE_DISCARD) |
4630 					      (1 << STRIPE_REPLACED)),
4631 			"head stripe state: %lx\n", head_sh->state);
4632 
4633 		set_mask_bits(&sh->state, ~(STRIPE_EXPAND_SYNC_FLAGS |
4634 					    (1 << STRIPE_PREREAD_ACTIVE) |
4635 					    (1 << STRIPE_DEGRADED) |
4636 					    (1 << STRIPE_ON_UNPLUG_LIST)),
4637 			      head_sh->state & (1 << STRIPE_INSYNC));
4638 
4639 		sh->check_state = head_sh->check_state;
4640 		sh->reconstruct_state = head_sh->reconstruct_state;
4641 		spin_lock_irq(&sh->stripe_lock);
4642 		sh->batch_head = NULL;
4643 		spin_unlock_irq(&sh->stripe_lock);
4644 		for (i = 0; i < sh->disks; i++) {
4645 			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
4646 				do_wakeup = 1;
4647 			sh->dev[i].flags = head_sh->dev[i].flags &
4648 				(~((1 << R5_WriteError) | (1 << R5_Overlap)));
4649 		}
4650 		if (handle_flags == 0 ||
4651 		    sh->state & handle_flags)
4652 			set_bit(STRIPE_HANDLE, &sh->state);
4653 		raid5_release_stripe(sh);
4654 	}
4655 	spin_lock_irq(&head_sh->stripe_lock);
4656 	head_sh->batch_head = NULL;
4657 	spin_unlock_irq(&head_sh->stripe_lock);
4658 	for (i = 0; i < head_sh->disks; i++)
4659 		if (test_and_clear_bit(R5_Overlap, &head_sh->dev[i].flags))
4660 			do_wakeup = 1;
4661 	if (head_sh->state & handle_flags)
4662 		set_bit(STRIPE_HANDLE, &head_sh->state);
4663 
4664 	if (do_wakeup)
4665 		wake_up(&head_sh->raid_conf->wait_for_overlap);
4666 }
4667 
4668 static void handle_stripe(struct stripe_head *sh)
4669 {
4670 	struct stripe_head_state s;
4671 	struct r5conf *conf = sh->raid_conf;
4672 	int i;
4673 	int prexor;
4674 	int disks = sh->disks;
4675 	struct r5dev *pdev, *qdev;
4676 
4677 	clear_bit(STRIPE_HANDLE, &sh->state);
4678 	if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
4679 		/* already being handled, ensure it gets handled
4680 		 * again when current action finishes */
4681 		set_bit(STRIPE_HANDLE, &sh->state);
4682 		return;
4683 	}
4684 
4685 	if (clear_batch_ready(sh) ) {
4686 		clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
4687 		return;
4688 	}
4689 
4690 	if (test_and_clear_bit(STRIPE_BATCH_ERR, &sh->state))
4691 		break_stripe_batch_list(sh, 0);
4692 
4693 	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state) && !sh->batch_head) {
4694 		spin_lock(&sh->stripe_lock);
4695 		/*
4696 		 * Cannot process 'sync' concurrently with 'discard'.
4697 		 * Flush data in r5cache before 'sync'.
4698 		 */
4699 		if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
4700 		    !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state) &&
4701 		    !test_bit(STRIPE_DISCARD, &sh->state) &&
4702 		    test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
4703 			set_bit(STRIPE_SYNCING, &sh->state);
4704 			clear_bit(STRIPE_INSYNC, &sh->state);
4705 			clear_bit(STRIPE_REPLACED, &sh->state);
4706 		}
4707 		spin_unlock(&sh->stripe_lock);
4708 	}
4709 	clear_bit(STRIPE_DELAYED, &sh->state);
4710 
4711 	pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
4712 		"pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
4713 	       (unsigned long long)sh->sector, sh->state,
4714 	       atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
4715 	       sh->check_state, sh->reconstruct_state);
4716 
4717 	analyse_stripe(sh, &s);
4718 
4719 	if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
4720 		goto finish;
4721 
4722 	if (s.handle_bad_blocks ||
4723 	    test_bit(MD_SB_CHANGE_PENDING, &conf->mddev->sb_flags)) {
4724 		set_bit(STRIPE_HANDLE, &sh->state);
4725 		goto finish;
4726 	}
4727 
4728 	if (unlikely(s.blocked_rdev)) {
4729 		if (s.syncing || s.expanding || s.expanded ||
4730 		    s.replacing || s.to_write || s.written) {
4731 			set_bit(STRIPE_HANDLE, &sh->state);
4732 			goto finish;
4733 		}
4734 		/* There is nothing for the blocked_rdev to block */
4735 		rdev_dec_pending(s.blocked_rdev, conf->mddev);
4736 		s.blocked_rdev = NULL;
4737 	}
4738 
4739 	if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
4740 		set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
4741 		set_bit(STRIPE_BIOFILL_RUN, &sh->state);
4742 	}
4743 
4744 	pr_debug("locked=%d uptodate=%d to_read=%d"
4745 	       " to_write=%d failed=%d failed_num=%d,%d\n",
4746 	       s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
4747 	       s.failed_num[0], s.failed_num[1]);
4748 	/*
4749 	 * check if the array has lost more than max_degraded devices and,
4750 	 * if so, some requests might need to be failed.
4751 	 *
4752 	 * When journal device failed (log_failed), we will only process
4753 	 * the stripe if there is data need write to raid disks
4754 	 */
4755 	if (s.failed > conf->max_degraded ||
4756 	    (s.log_failed && s.injournal == 0)) {
4757 		sh->check_state = 0;
4758 		sh->reconstruct_state = 0;
4759 		break_stripe_batch_list(sh, 0);
4760 		if (s.to_read+s.to_write+s.written)
4761 			handle_failed_stripe(conf, sh, &s, disks);
4762 		if (s.syncing + s.replacing)
4763 			handle_failed_sync(conf, sh, &s);
4764 	}
4765 
4766 	/* Now we check to see if any write operations have recently
4767 	 * completed
4768 	 */
4769 	prexor = 0;
4770 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
4771 		prexor = 1;
4772 	if (sh->reconstruct_state == reconstruct_state_drain_result ||
4773 	    sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
4774 		sh->reconstruct_state = reconstruct_state_idle;
4775 
4776 		/* All the 'written' buffers and the parity block are ready to
4777 		 * be written back to disk
4778 		 */
4779 		BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags) &&
4780 		       !test_bit(R5_Discard, &sh->dev[sh->pd_idx].flags));
4781 		BUG_ON(sh->qd_idx >= 0 &&
4782 		       !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags) &&
4783 		       !test_bit(R5_Discard, &sh->dev[sh->qd_idx].flags));
4784 		for (i = disks; i--; ) {
4785 			struct r5dev *dev = &sh->dev[i];
4786 			if (test_bit(R5_LOCKED, &dev->flags) &&
4787 				(i == sh->pd_idx || i == sh->qd_idx ||
4788 				 dev->written || test_bit(R5_InJournal,
4789 							  &dev->flags))) {
4790 				pr_debug("Writing block %d\n", i);
4791 				set_bit(R5_Wantwrite, &dev->flags);
4792 				if (prexor)
4793 					continue;
4794 				if (s.failed > 1)
4795 					continue;
4796 				if (!test_bit(R5_Insync, &dev->flags) ||
4797 				    ((i == sh->pd_idx || i == sh->qd_idx)  &&
4798 				     s.failed == 0))
4799 					set_bit(STRIPE_INSYNC, &sh->state);
4800 			}
4801 		}
4802 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
4803 			s.dec_preread_active = 1;
4804 	}
4805 
4806 	/*
4807 	 * might be able to return some write requests if the parity blocks
4808 	 * are safe, or on a failed drive
4809 	 */
4810 	pdev = &sh->dev[sh->pd_idx];
4811 	s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
4812 		|| (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
4813 	qdev = &sh->dev[sh->qd_idx];
4814 	s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
4815 		|| (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
4816 		|| conf->level < 6;
4817 
4818 	if (s.written &&
4819 	    (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
4820 			     && !test_bit(R5_LOCKED, &pdev->flags)
4821 			     && (test_bit(R5_UPTODATE, &pdev->flags) ||
4822 				 test_bit(R5_Discard, &pdev->flags))))) &&
4823 	    (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
4824 			     && !test_bit(R5_LOCKED, &qdev->flags)
4825 			     && (test_bit(R5_UPTODATE, &qdev->flags) ||
4826 				 test_bit(R5_Discard, &qdev->flags))))))
4827 		handle_stripe_clean_event(conf, sh, disks);
4828 
4829 	if (s.just_cached)
4830 		r5c_handle_cached_data_endio(conf, sh, disks);
4831 	log_stripe_write_finished(sh);
4832 
4833 	/* Now we might consider reading some blocks, either to check/generate
4834 	 * parity, or to satisfy requests
4835 	 * or to load a block that is being partially written.
4836 	 */
4837 	if (s.to_read || s.non_overwrite
4838 	    || (conf->level == 6 && s.to_write && s.failed)
4839 	    || (s.syncing && (s.uptodate + s.compute < disks))
4840 	    || s.replacing
4841 	    || s.expanding)
4842 		handle_stripe_fill(sh, &s, disks);
4843 
4844 	/*
4845 	 * When the stripe finishes full journal write cycle (write to journal
4846 	 * and raid disk), this is the clean up procedure so it is ready for
4847 	 * next operation.
4848 	 */
4849 	r5c_finish_stripe_write_out(conf, sh, &s);
4850 
4851 	/*
4852 	 * Now to consider new write requests, cache write back and what else,
4853 	 * if anything should be read.  We do not handle new writes when:
4854 	 * 1/ A 'write' operation (copy+xor) is already in flight.
4855 	 * 2/ A 'check' operation is in flight, as it may clobber the parity
4856 	 *    block.
4857 	 * 3/ A r5c cache log write is in flight.
4858 	 */
4859 
4860 	if (!sh->reconstruct_state && !sh->check_state && !sh->log_io) {
4861 		if (!r5c_is_writeback(conf->log)) {
4862 			if (s.to_write)
4863 				handle_stripe_dirtying(conf, sh, &s, disks);
4864 		} else { /* write back cache */
4865 			int ret = 0;
4866 
4867 			/* First, try handle writes in caching phase */
4868 			if (s.to_write)
4869 				ret = r5c_try_caching_write(conf, sh, &s,
4870 							    disks);
4871 			/*
4872 			 * If caching phase failed: ret == -EAGAIN
4873 			 *    OR
4874 			 * stripe under reclaim: !caching && injournal
4875 			 *
4876 			 * fall back to handle_stripe_dirtying()
4877 			 */
4878 			if (ret == -EAGAIN ||
4879 			    /* stripe under reclaim: !caching && injournal */
4880 			    (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
4881 			     s.injournal > 0)) {
4882 				ret = handle_stripe_dirtying(conf, sh, &s,
4883 							     disks);
4884 				if (ret == -EAGAIN)
4885 					goto finish;
4886 			}
4887 		}
4888 	}
4889 
4890 	/* maybe we need to check and possibly fix the parity for this stripe
4891 	 * Any reads will already have been scheduled, so we just see if enough
4892 	 * data is available.  The parity check is held off while parity
4893 	 * dependent operations are in flight.
4894 	 */
4895 	if (sh->check_state ||
4896 	    (s.syncing && s.locked == 0 &&
4897 	     !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4898 	     !test_bit(STRIPE_INSYNC, &sh->state))) {
4899 		if (conf->level == 6)
4900 			handle_parity_checks6(conf, sh, &s, disks);
4901 		else
4902 			handle_parity_checks5(conf, sh, &s, disks);
4903 	}
4904 
4905 	if ((s.replacing || s.syncing) && s.locked == 0
4906 	    && !test_bit(STRIPE_COMPUTE_RUN, &sh->state)
4907 	    && !test_bit(STRIPE_REPLACED, &sh->state)) {
4908 		/* Write out to replacement devices where possible */
4909 		for (i = 0; i < conf->raid_disks; i++)
4910 			if (test_bit(R5_NeedReplace, &sh->dev[i].flags)) {
4911 				WARN_ON(!test_bit(R5_UPTODATE, &sh->dev[i].flags));
4912 				set_bit(R5_WantReplace, &sh->dev[i].flags);
4913 				set_bit(R5_LOCKED, &sh->dev[i].flags);
4914 				s.locked++;
4915 			}
4916 		if (s.replacing)
4917 			set_bit(STRIPE_INSYNC, &sh->state);
4918 		set_bit(STRIPE_REPLACED, &sh->state);
4919 	}
4920 	if ((s.syncing || s.replacing) && s.locked == 0 &&
4921 	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
4922 	    test_bit(STRIPE_INSYNC, &sh->state)) {
4923 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4924 		clear_bit(STRIPE_SYNCING, &sh->state);
4925 		if (test_and_clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags))
4926 			wake_up(&conf->wait_for_overlap);
4927 	}
4928 
4929 	/* If the failed drives are just a ReadError, then we might need
4930 	 * to progress the repair/check process
4931 	 */
4932 	if (s.failed <= conf->max_degraded && !conf->mddev->ro)
4933 		for (i = 0; i < s.failed; i++) {
4934 			struct r5dev *dev = &sh->dev[s.failed_num[i]];
4935 			if (test_bit(R5_ReadError, &dev->flags)
4936 			    && !test_bit(R5_LOCKED, &dev->flags)
4937 			    && test_bit(R5_UPTODATE, &dev->flags)
4938 				) {
4939 				if (!test_bit(R5_ReWrite, &dev->flags)) {
4940 					set_bit(R5_Wantwrite, &dev->flags);
4941 					set_bit(R5_ReWrite, &dev->flags);
4942 					set_bit(R5_LOCKED, &dev->flags);
4943 					s.locked++;
4944 				} else {
4945 					/* let's read it back */
4946 					set_bit(R5_Wantread, &dev->flags);
4947 					set_bit(R5_LOCKED, &dev->flags);
4948 					s.locked++;
4949 				}
4950 			}
4951 		}
4952 
4953 	/* Finish reconstruct operations initiated by the expansion process */
4954 	if (sh->reconstruct_state == reconstruct_state_result) {
4955 		struct stripe_head *sh_src
4956 			= raid5_get_active_stripe(conf, sh->sector, 1, 1, 1);
4957 		if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
4958 			/* sh cannot be written until sh_src has been read.
4959 			 * so arrange for sh to be delayed a little
4960 			 */
4961 			set_bit(STRIPE_DELAYED, &sh->state);
4962 			set_bit(STRIPE_HANDLE, &sh->state);
4963 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
4964 					      &sh_src->state))
4965 				atomic_inc(&conf->preread_active_stripes);
4966 			raid5_release_stripe(sh_src);
4967 			goto finish;
4968 		}
4969 		if (sh_src)
4970 			raid5_release_stripe(sh_src);
4971 
4972 		sh->reconstruct_state = reconstruct_state_idle;
4973 		clear_bit(STRIPE_EXPANDING, &sh->state);
4974 		for (i = conf->raid_disks; i--; ) {
4975 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
4976 			set_bit(R5_LOCKED, &sh->dev[i].flags);
4977 			s.locked++;
4978 		}
4979 	}
4980 
4981 	if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
4982 	    !sh->reconstruct_state) {
4983 		/* Need to write out all blocks after computing parity */
4984 		sh->disks = conf->raid_disks;
4985 		stripe_set_idx(sh->sector, conf, 0, sh);
4986 		schedule_reconstruction(sh, &s, 1, 1);
4987 	} else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
4988 		clear_bit(STRIPE_EXPAND_READY, &sh->state);
4989 		atomic_dec(&conf->reshape_stripes);
4990 		wake_up(&conf->wait_for_overlap);
4991 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
4992 	}
4993 
4994 	if (s.expanding && s.locked == 0 &&
4995 	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
4996 		handle_stripe_expansion(conf, sh);
4997 
4998 finish:
4999 	/* wait for this device to become unblocked */
5000 	if (unlikely(s.blocked_rdev)) {
5001 		if (conf->mddev->external)
5002 			md_wait_for_blocked_rdev(s.blocked_rdev,
5003 						 conf->mddev);
5004 		else
5005 			/* Internal metadata will immediately
5006 			 * be written by raid5d, so we don't
5007 			 * need to wait here.
5008 			 */
5009 			rdev_dec_pending(s.blocked_rdev,
5010 					 conf->mddev);
5011 	}
5012 
5013 	if (s.handle_bad_blocks)
5014 		for (i = disks; i--; ) {
5015 			struct md_rdev *rdev;
5016 			struct r5dev *dev = &sh->dev[i];
5017 			if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
5018 				/* We own a safe reference to the rdev */
5019 				rdev = conf->disks[i].rdev;
5020 				if (!rdev_set_badblocks(rdev, sh->sector,
5021 							STRIPE_SECTORS, 0))
5022 					md_error(conf->mddev, rdev);
5023 				rdev_dec_pending(rdev, conf->mddev);
5024 			}
5025 			if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
5026 				rdev = conf->disks[i].rdev;
5027 				rdev_clear_badblocks(rdev, sh->sector,
5028 						     STRIPE_SECTORS, 0);
5029 				rdev_dec_pending(rdev, conf->mddev);
5030 			}
5031 			if (test_and_clear_bit(R5_MadeGoodRepl, &dev->flags)) {
5032 				rdev = conf->disks[i].replacement;
5033 				if (!rdev)
5034 					/* rdev have been moved down */
5035 					rdev = conf->disks[i].rdev;
5036 				rdev_clear_badblocks(rdev, sh->sector,
5037 						     STRIPE_SECTORS, 0);
5038 				rdev_dec_pending(rdev, conf->mddev);
5039 			}
5040 		}
5041 
5042 	if (s.ops_request)
5043 		raid_run_ops(sh, s.ops_request);
5044 
5045 	ops_run_io(sh, &s);
5046 
5047 	if (s.dec_preread_active) {
5048 		/* We delay this until after ops_run_io so that if make_request
5049 		 * is waiting on a flush, it won't continue until the writes
5050 		 * have actually been submitted.
5051 		 */
5052 		atomic_dec(&conf->preread_active_stripes);
5053 		if (atomic_read(&conf->preread_active_stripes) <
5054 		    IO_THRESHOLD)
5055 			md_wakeup_thread(conf->mddev->thread);
5056 	}
5057 
5058 	clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
5059 }
5060 
5061 static void raid5_activate_delayed(struct r5conf *conf)
5062 {
5063 	if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
5064 		while (!list_empty(&conf->delayed_list)) {
5065 			struct list_head *l = conf->delayed_list.next;
5066 			struct stripe_head *sh;
5067 			sh = list_entry(l, struct stripe_head, lru);
5068 			list_del_init(l);
5069 			clear_bit(STRIPE_DELAYED, &sh->state);
5070 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5071 				atomic_inc(&conf->preread_active_stripes);
5072 			list_add_tail(&sh->lru, &conf->hold_list);
5073 			raid5_wakeup_stripe_thread(sh);
5074 		}
5075 	}
5076 }
5077 
5078 static void activate_bit_delay(struct r5conf *conf,
5079 	struct list_head *temp_inactive_list)
5080 {
5081 	/* device_lock is held */
5082 	struct list_head head;
5083 	list_add(&head, &conf->bitmap_list);
5084 	list_del_init(&conf->bitmap_list);
5085 	while (!list_empty(&head)) {
5086 		struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
5087 		int hash;
5088 		list_del_init(&sh->lru);
5089 		atomic_inc(&sh->count);
5090 		hash = sh->hash_lock_index;
5091 		__release_stripe(conf, sh, &temp_inactive_list[hash]);
5092 	}
5093 }
5094 
5095 static int raid5_congested(struct mddev *mddev, int bits)
5096 {
5097 	struct r5conf *conf = mddev->private;
5098 
5099 	/* No difference between reads and writes.  Just check
5100 	 * how busy the stripe_cache is
5101 	 */
5102 
5103 	if (test_bit(R5_INACTIVE_BLOCKED, &conf->cache_state))
5104 		return 1;
5105 
5106 	/* Also checks whether there is pressure on r5cache log space */
5107 	if (test_bit(R5C_LOG_TIGHT, &conf->cache_state))
5108 		return 1;
5109 	if (conf->quiesce)
5110 		return 1;
5111 	if (atomic_read(&conf->empty_inactive_list_nr))
5112 		return 1;
5113 
5114 	return 0;
5115 }
5116 
5117 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
5118 {
5119 	struct r5conf *conf = mddev->private;
5120 	sector_t sector = bio->bi_iter.bi_sector;
5121 	unsigned int chunk_sectors;
5122 	unsigned int bio_sectors = bio_sectors(bio);
5123 
5124 	WARN_ON_ONCE(bio->bi_partno);
5125 
5126 	chunk_sectors = min(conf->chunk_sectors, conf->prev_chunk_sectors);
5127 	return  chunk_sectors >=
5128 		((sector & (chunk_sectors - 1)) + bio_sectors);
5129 }
5130 
5131 /*
5132  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
5133  *  later sampled by raid5d.
5134  */
5135 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
5136 {
5137 	unsigned long flags;
5138 
5139 	spin_lock_irqsave(&conf->device_lock, flags);
5140 
5141 	bi->bi_next = conf->retry_read_aligned_list;
5142 	conf->retry_read_aligned_list = bi;
5143 
5144 	spin_unlock_irqrestore(&conf->device_lock, flags);
5145 	md_wakeup_thread(conf->mddev->thread);
5146 }
5147 
5148 static struct bio *remove_bio_from_retry(struct r5conf *conf,
5149 					 unsigned int *offset)
5150 {
5151 	struct bio *bi;
5152 
5153 	bi = conf->retry_read_aligned;
5154 	if (bi) {
5155 		*offset = conf->retry_read_offset;
5156 		conf->retry_read_aligned = NULL;
5157 		return bi;
5158 	}
5159 	bi = conf->retry_read_aligned_list;
5160 	if(bi) {
5161 		conf->retry_read_aligned_list = bi->bi_next;
5162 		bi->bi_next = NULL;
5163 		*offset = 0;
5164 	}
5165 
5166 	return bi;
5167 }
5168 
5169 /*
5170  *  The "raid5_align_endio" should check if the read succeeded and if it
5171  *  did, call bio_endio on the original bio (having bio_put the new bio
5172  *  first).
5173  *  If the read failed..
5174  */
5175 static void raid5_align_endio(struct bio *bi)
5176 {
5177 	struct bio* raid_bi  = bi->bi_private;
5178 	struct mddev *mddev;
5179 	struct r5conf *conf;
5180 	struct md_rdev *rdev;
5181 	blk_status_t error = bi->bi_status;
5182 
5183 	bio_put(bi);
5184 
5185 	rdev = (void*)raid_bi->bi_next;
5186 	raid_bi->bi_next = NULL;
5187 	mddev = rdev->mddev;
5188 	conf = mddev->private;
5189 
5190 	rdev_dec_pending(rdev, conf->mddev);
5191 
5192 	if (!error) {
5193 		bio_endio(raid_bi);
5194 		if (atomic_dec_and_test(&conf->active_aligned_reads))
5195 			wake_up(&conf->wait_for_quiescent);
5196 		return;
5197 	}
5198 
5199 	pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
5200 
5201 	add_bio_to_retry(raid_bi, conf);
5202 }
5203 
5204 static int raid5_read_one_chunk(struct mddev *mddev, struct bio *raid_bio)
5205 {
5206 	struct r5conf *conf = mddev->private;
5207 	int dd_idx;
5208 	struct bio* align_bi;
5209 	struct md_rdev *rdev;
5210 	sector_t end_sector;
5211 
5212 	if (!in_chunk_boundary(mddev, raid_bio)) {
5213 		pr_debug("%s: non aligned\n", __func__);
5214 		return 0;
5215 	}
5216 	/*
5217 	 * use bio_clone_fast to make a copy of the bio
5218 	 */
5219 	align_bi = bio_clone_fast(raid_bio, GFP_NOIO, &mddev->bio_set);
5220 	if (!align_bi)
5221 		return 0;
5222 	/*
5223 	 *   set bi_end_io to a new function, and set bi_private to the
5224 	 *     original bio.
5225 	 */
5226 	align_bi->bi_end_io  = raid5_align_endio;
5227 	align_bi->bi_private = raid_bio;
5228 	/*
5229 	 *	compute position
5230 	 */
5231 	align_bi->bi_iter.bi_sector =
5232 		raid5_compute_sector(conf, raid_bio->bi_iter.bi_sector,
5233 				     0, &dd_idx, NULL);
5234 
5235 	end_sector = bio_end_sector(align_bi);
5236 	rcu_read_lock();
5237 	rdev = rcu_dereference(conf->disks[dd_idx].replacement);
5238 	if (!rdev || test_bit(Faulty, &rdev->flags) ||
5239 	    rdev->recovery_offset < end_sector) {
5240 		rdev = rcu_dereference(conf->disks[dd_idx].rdev);
5241 		if (rdev &&
5242 		    (test_bit(Faulty, &rdev->flags) ||
5243 		    !(test_bit(In_sync, &rdev->flags) ||
5244 		      rdev->recovery_offset >= end_sector)))
5245 			rdev = NULL;
5246 	}
5247 
5248 	if (r5c_big_stripe_cached(conf, align_bi->bi_iter.bi_sector)) {
5249 		rcu_read_unlock();
5250 		bio_put(align_bi);
5251 		return 0;
5252 	}
5253 
5254 	if (rdev) {
5255 		sector_t first_bad;
5256 		int bad_sectors;
5257 
5258 		atomic_inc(&rdev->nr_pending);
5259 		rcu_read_unlock();
5260 		raid_bio->bi_next = (void*)rdev;
5261 		bio_set_dev(align_bi, rdev->bdev);
5262 		bio_clear_flag(align_bi, BIO_SEG_VALID);
5263 
5264 		if (is_badblock(rdev, align_bi->bi_iter.bi_sector,
5265 				bio_sectors(align_bi),
5266 				&first_bad, &bad_sectors)) {
5267 			bio_put(align_bi);
5268 			rdev_dec_pending(rdev, mddev);
5269 			return 0;
5270 		}
5271 
5272 		/* No reshape active, so we can trust rdev->data_offset */
5273 		align_bi->bi_iter.bi_sector += rdev->data_offset;
5274 
5275 		spin_lock_irq(&conf->device_lock);
5276 		wait_event_lock_irq(conf->wait_for_quiescent,
5277 				    conf->quiesce == 0,
5278 				    conf->device_lock);
5279 		atomic_inc(&conf->active_aligned_reads);
5280 		spin_unlock_irq(&conf->device_lock);
5281 
5282 		if (mddev->gendisk)
5283 			trace_block_bio_remap(align_bi->bi_disk->queue,
5284 					      align_bi, disk_devt(mddev->gendisk),
5285 					      raid_bio->bi_iter.bi_sector);
5286 		generic_make_request(align_bi);
5287 		return 1;
5288 	} else {
5289 		rcu_read_unlock();
5290 		bio_put(align_bi);
5291 		return 0;
5292 	}
5293 }
5294 
5295 static struct bio *chunk_aligned_read(struct mddev *mddev, struct bio *raid_bio)
5296 {
5297 	struct bio *split;
5298 	sector_t sector = raid_bio->bi_iter.bi_sector;
5299 	unsigned chunk_sects = mddev->chunk_sectors;
5300 	unsigned sectors = chunk_sects - (sector & (chunk_sects-1));
5301 
5302 	if (sectors < bio_sectors(raid_bio)) {
5303 		struct r5conf *conf = mddev->private;
5304 		split = bio_split(raid_bio, sectors, GFP_NOIO, &conf->bio_split);
5305 		bio_chain(split, raid_bio);
5306 		generic_make_request(raid_bio);
5307 		raid_bio = split;
5308 	}
5309 
5310 	if (!raid5_read_one_chunk(mddev, raid_bio))
5311 		return raid_bio;
5312 
5313 	return NULL;
5314 }
5315 
5316 /* __get_priority_stripe - get the next stripe to process
5317  *
5318  * Full stripe writes are allowed to pass preread active stripes up until
5319  * the bypass_threshold is exceeded.  In general the bypass_count
5320  * increments when the handle_list is handled before the hold_list; however, it
5321  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
5322  * stripe with in flight i/o.  The bypass_count will be reset when the
5323  * head of the hold_list has changed, i.e. the head was promoted to the
5324  * handle_list.
5325  */
5326 static struct stripe_head *__get_priority_stripe(struct r5conf *conf, int group)
5327 {
5328 	struct stripe_head *sh, *tmp;
5329 	struct list_head *handle_list = NULL;
5330 	struct r5worker_group *wg;
5331 	bool second_try = !r5c_is_writeback(conf->log) &&
5332 		!r5l_log_disk_error(conf);
5333 	bool try_loprio = test_bit(R5C_LOG_TIGHT, &conf->cache_state) ||
5334 		r5l_log_disk_error(conf);
5335 
5336 again:
5337 	wg = NULL;
5338 	sh = NULL;
5339 	if (conf->worker_cnt_per_group == 0) {
5340 		handle_list = try_loprio ? &conf->loprio_list :
5341 					&conf->handle_list;
5342 	} else if (group != ANY_GROUP) {
5343 		handle_list = try_loprio ? &conf->worker_groups[group].loprio_list :
5344 				&conf->worker_groups[group].handle_list;
5345 		wg = &conf->worker_groups[group];
5346 	} else {
5347 		int i;
5348 		for (i = 0; i < conf->group_cnt; i++) {
5349 			handle_list = try_loprio ? &conf->worker_groups[i].loprio_list :
5350 				&conf->worker_groups[i].handle_list;
5351 			wg = &conf->worker_groups[i];
5352 			if (!list_empty(handle_list))
5353 				break;
5354 		}
5355 	}
5356 
5357 	pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
5358 		  __func__,
5359 		  list_empty(handle_list) ? "empty" : "busy",
5360 		  list_empty(&conf->hold_list) ? "empty" : "busy",
5361 		  atomic_read(&conf->pending_full_writes), conf->bypass_count);
5362 
5363 	if (!list_empty(handle_list)) {
5364 		sh = list_entry(handle_list->next, typeof(*sh), lru);
5365 
5366 		if (list_empty(&conf->hold_list))
5367 			conf->bypass_count = 0;
5368 		else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
5369 			if (conf->hold_list.next == conf->last_hold)
5370 				conf->bypass_count++;
5371 			else {
5372 				conf->last_hold = conf->hold_list.next;
5373 				conf->bypass_count -= conf->bypass_threshold;
5374 				if (conf->bypass_count < 0)
5375 					conf->bypass_count = 0;
5376 			}
5377 		}
5378 	} else if (!list_empty(&conf->hold_list) &&
5379 		   ((conf->bypass_threshold &&
5380 		     conf->bypass_count > conf->bypass_threshold) ||
5381 		    atomic_read(&conf->pending_full_writes) == 0)) {
5382 
5383 		list_for_each_entry(tmp, &conf->hold_list,  lru) {
5384 			if (conf->worker_cnt_per_group == 0 ||
5385 			    group == ANY_GROUP ||
5386 			    !cpu_online(tmp->cpu) ||
5387 			    cpu_to_group(tmp->cpu) == group) {
5388 				sh = tmp;
5389 				break;
5390 			}
5391 		}
5392 
5393 		if (sh) {
5394 			conf->bypass_count -= conf->bypass_threshold;
5395 			if (conf->bypass_count < 0)
5396 				conf->bypass_count = 0;
5397 		}
5398 		wg = NULL;
5399 	}
5400 
5401 	if (!sh) {
5402 		if (second_try)
5403 			return NULL;
5404 		second_try = true;
5405 		try_loprio = !try_loprio;
5406 		goto again;
5407 	}
5408 
5409 	if (wg) {
5410 		wg->stripes_cnt--;
5411 		sh->group = NULL;
5412 	}
5413 	list_del_init(&sh->lru);
5414 	BUG_ON(atomic_inc_return(&sh->count) != 1);
5415 	return sh;
5416 }
5417 
5418 struct raid5_plug_cb {
5419 	struct blk_plug_cb	cb;
5420 	struct list_head	list;
5421 	struct list_head	temp_inactive_list[NR_STRIPE_HASH_LOCKS];
5422 };
5423 
5424 static void raid5_unplug(struct blk_plug_cb *blk_cb, bool from_schedule)
5425 {
5426 	struct raid5_plug_cb *cb = container_of(
5427 		blk_cb, struct raid5_plug_cb, cb);
5428 	struct stripe_head *sh;
5429 	struct mddev *mddev = cb->cb.data;
5430 	struct r5conf *conf = mddev->private;
5431 	int cnt = 0;
5432 	int hash;
5433 
5434 	if (cb->list.next && !list_empty(&cb->list)) {
5435 		spin_lock_irq(&conf->device_lock);
5436 		while (!list_empty(&cb->list)) {
5437 			sh = list_first_entry(&cb->list, struct stripe_head, lru);
5438 			list_del_init(&sh->lru);
5439 			/*
5440 			 * avoid race release_stripe_plug() sees
5441 			 * STRIPE_ON_UNPLUG_LIST clear but the stripe
5442 			 * is still in our list
5443 			 */
5444 			smp_mb__before_atomic();
5445 			clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
5446 			/*
5447 			 * STRIPE_ON_RELEASE_LIST could be set here. In that
5448 			 * case, the count is always > 1 here
5449 			 */
5450 			hash = sh->hash_lock_index;
5451 			__release_stripe(conf, sh, &cb->temp_inactive_list[hash]);
5452 			cnt++;
5453 		}
5454 		spin_unlock_irq(&conf->device_lock);
5455 	}
5456 	release_inactive_stripe_list(conf, cb->temp_inactive_list,
5457 				     NR_STRIPE_HASH_LOCKS);
5458 	if (mddev->queue)
5459 		trace_block_unplug(mddev->queue, cnt, !from_schedule);
5460 	kfree(cb);
5461 }
5462 
5463 static void release_stripe_plug(struct mddev *mddev,
5464 				struct stripe_head *sh)
5465 {
5466 	struct blk_plug_cb *blk_cb = blk_check_plugged(
5467 		raid5_unplug, mddev,
5468 		sizeof(struct raid5_plug_cb));
5469 	struct raid5_plug_cb *cb;
5470 
5471 	if (!blk_cb) {
5472 		raid5_release_stripe(sh);
5473 		return;
5474 	}
5475 
5476 	cb = container_of(blk_cb, struct raid5_plug_cb, cb);
5477 
5478 	if (cb->list.next == NULL) {
5479 		int i;
5480 		INIT_LIST_HEAD(&cb->list);
5481 		for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
5482 			INIT_LIST_HEAD(cb->temp_inactive_list + i);
5483 	}
5484 
5485 	if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
5486 		list_add_tail(&sh->lru, &cb->list);
5487 	else
5488 		raid5_release_stripe(sh);
5489 }
5490 
5491 static void make_discard_request(struct mddev *mddev, struct bio *bi)
5492 {
5493 	struct r5conf *conf = mddev->private;
5494 	sector_t logical_sector, last_sector;
5495 	struct stripe_head *sh;
5496 	int stripe_sectors;
5497 
5498 	if (mddev->reshape_position != MaxSector)
5499 		/* Skip discard while reshape is happening */
5500 		return;
5501 
5502 	logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5503 	last_sector = bi->bi_iter.bi_sector + (bi->bi_iter.bi_size>>9);
5504 
5505 	bi->bi_next = NULL;
5506 
5507 	stripe_sectors = conf->chunk_sectors *
5508 		(conf->raid_disks - conf->max_degraded);
5509 	logical_sector = DIV_ROUND_UP_SECTOR_T(logical_sector,
5510 					       stripe_sectors);
5511 	sector_div(last_sector, stripe_sectors);
5512 
5513 	logical_sector *= conf->chunk_sectors;
5514 	last_sector *= conf->chunk_sectors;
5515 
5516 	for (; logical_sector < last_sector;
5517 	     logical_sector += STRIPE_SECTORS) {
5518 		DEFINE_WAIT(w);
5519 		int d;
5520 	again:
5521 		sh = raid5_get_active_stripe(conf, logical_sector, 0, 0, 0);
5522 		prepare_to_wait(&conf->wait_for_overlap, &w,
5523 				TASK_UNINTERRUPTIBLE);
5524 		set_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5525 		if (test_bit(STRIPE_SYNCING, &sh->state)) {
5526 			raid5_release_stripe(sh);
5527 			schedule();
5528 			goto again;
5529 		}
5530 		clear_bit(R5_Overlap, &sh->dev[sh->pd_idx].flags);
5531 		spin_lock_irq(&sh->stripe_lock);
5532 		for (d = 0; d < conf->raid_disks; d++) {
5533 			if (d == sh->pd_idx || d == sh->qd_idx)
5534 				continue;
5535 			if (sh->dev[d].towrite || sh->dev[d].toread) {
5536 				set_bit(R5_Overlap, &sh->dev[d].flags);
5537 				spin_unlock_irq(&sh->stripe_lock);
5538 				raid5_release_stripe(sh);
5539 				schedule();
5540 				goto again;
5541 			}
5542 		}
5543 		set_bit(STRIPE_DISCARD, &sh->state);
5544 		finish_wait(&conf->wait_for_overlap, &w);
5545 		sh->overwrite_disks = 0;
5546 		for (d = 0; d < conf->raid_disks; d++) {
5547 			if (d == sh->pd_idx || d == sh->qd_idx)
5548 				continue;
5549 			sh->dev[d].towrite = bi;
5550 			set_bit(R5_OVERWRITE, &sh->dev[d].flags);
5551 			bio_inc_remaining(bi);
5552 			md_write_inc(mddev, bi);
5553 			sh->overwrite_disks++;
5554 		}
5555 		spin_unlock_irq(&sh->stripe_lock);
5556 		if (conf->mddev->bitmap) {
5557 			for (d = 0;
5558 			     d < conf->raid_disks - conf->max_degraded;
5559 			     d++)
5560 				md_bitmap_startwrite(mddev->bitmap,
5561 						     sh->sector,
5562 						     STRIPE_SECTORS,
5563 						     0);
5564 			sh->bm_seq = conf->seq_flush + 1;
5565 			set_bit(STRIPE_BIT_DELAY, &sh->state);
5566 		}
5567 
5568 		set_bit(STRIPE_HANDLE, &sh->state);
5569 		clear_bit(STRIPE_DELAYED, &sh->state);
5570 		if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5571 			atomic_inc(&conf->preread_active_stripes);
5572 		release_stripe_plug(mddev, sh);
5573 	}
5574 
5575 	bio_endio(bi);
5576 }
5577 
5578 static bool raid5_make_request(struct mddev *mddev, struct bio * bi)
5579 {
5580 	struct r5conf *conf = mddev->private;
5581 	int dd_idx;
5582 	sector_t new_sector;
5583 	sector_t logical_sector, last_sector;
5584 	struct stripe_head *sh;
5585 	const int rw = bio_data_dir(bi);
5586 	DEFINE_WAIT(w);
5587 	bool do_prepare;
5588 	bool do_flush = false;
5589 
5590 	if (unlikely(bi->bi_opf & REQ_PREFLUSH)) {
5591 		int ret = log_handle_flush_request(conf, bi);
5592 
5593 		if (ret == 0)
5594 			return true;
5595 		if (ret == -ENODEV) {
5596 			md_flush_request(mddev, bi);
5597 			return true;
5598 		}
5599 		/* ret == -EAGAIN, fallback */
5600 		/*
5601 		 * if r5l_handle_flush_request() didn't clear REQ_PREFLUSH,
5602 		 * we need to flush journal device
5603 		 */
5604 		do_flush = bi->bi_opf & REQ_PREFLUSH;
5605 	}
5606 
5607 	if (!md_write_start(mddev, bi))
5608 		return false;
5609 	/*
5610 	 * If array is degraded, better not do chunk aligned read because
5611 	 * later we might have to read it again in order to reconstruct
5612 	 * data on failed drives.
5613 	 */
5614 	if (rw == READ && mddev->degraded == 0 &&
5615 	    mddev->reshape_position == MaxSector) {
5616 		bi = chunk_aligned_read(mddev, bi);
5617 		if (!bi)
5618 			return true;
5619 	}
5620 
5621 	if (unlikely(bio_op(bi) == REQ_OP_DISCARD)) {
5622 		make_discard_request(mddev, bi);
5623 		md_write_end(mddev);
5624 		return true;
5625 	}
5626 
5627 	logical_sector = bi->bi_iter.bi_sector & ~((sector_t)STRIPE_SECTORS-1);
5628 	last_sector = bio_end_sector(bi);
5629 	bi->bi_next = NULL;
5630 
5631 	prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
5632 	for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
5633 		int previous;
5634 		int seq;
5635 
5636 		do_prepare = false;
5637 	retry:
5638 		seq = read_seqcount_begin(&conf->gen_lock);
5639 		previous = 0;
5640 		if (do_prepare)
5641 			prepare_to_wait(&conf->wait_for_overlap, &w,
5642 				TASK_UNINTERRUPTIBLE);
5643 		if (unlikely(conf->reshape_progress != MaxSector)) {
5644 			/* spinlock is needed as reshape_progress may be
5645 			 * 64bit on a 32bit platform, and so it might be
5646 			 * possible to see a half-updated value
5647 			 * Of course reshape_progress could change after
5648 			 * the lock is dropped, so once we get a reference
5649 			 * to the stripe that we think it is, we will have
5650 			 * to check again.
5651 			 */
5652 			spin_lock_irq(&conf->device_lock);
5653 			if (mddev->reshape_backwards
5654 			    ? logical_sector < conf->reshape_progress
5655 			    : logical_sector >= conf->reshape_progress) {
5656 				previous = 1;
5657 			} else {
5658 				if (mddev->reshape_backwards
5659 				    ? logical_sector < conf->reshape_safe
5660 				    : logical_sector >= conf->reshape_safe) {
5661 					spin_unlock_irq(&conf->device_lock);
5662 					schedule();
5663 					do_prepare = true;
5664 					goto retry;
5665 				}
5666 			}
5667 			spin_unlock_irq(&conf->device_lock);
5668 		}
5669 
5670 		new_sector = raid5_compute_sector(conf, logical_sector,
5671 						  previous,
5672 						  &dd_idx, NULL);
5673 		pr_debug("raid456: raid5_make_request, sector %llu logical %llu\n",
5674 			(unsigned long long)new_sector,
5675 			(unsigned long long)logical_sector);
5676 
5677 		sh = raid5_get_active_stripe(conf, new_sector, previous,
5678 				       (bi->bi_opf & REQ_RAHEAD), 0);
5679 		if (sh) {
5680 			if (unlikely(previous)) {
5681 				/* expansion might have moved on while waiting for a
5682 				 * stripe, so we must do the range check again.
5683 				 * Expansion could still move past after this
5684 				 * test, but as we are holding a reference to
5685 				 * 'sh', we know that if that happens,
5686 				 *  STRIPE_EXPANDING will get set and the expansion
5687 				 * won't proceed until we finish with the stripe.
5688 				 */
5689 				int must_retry = 0;
5690 				spin_lock_irq(&conf->device_lock);
5691 				if (mddev->reshape_backwards
5692 				    ? logical_sector >= conf->reshape_progress
5693 				    : logical_sector < conf->reshape_progress)
5694 					/* mismatch, need to try again */
5695 					must_retry = 1;
5696 				spin_unlock_irq(&conf->device_lock);
5697 				if (must_retry) {
5698 					raid5_release_stripe(sh);
5699 					schedule();
5700 					do_prepare = true;
5701 					goto retry;
5702 				}
5703 			}
5704 			if (read_seqcount_retry(&conf->gen_lock, seq)) {
5705 				/* Might have got the wrong stripe_head
5706 				 * by accident
5707 				 */
5708 				raid5_release_stripe(sh);
5709 				goto retry;
5710 			}
5711 
5712 			if (test_bit(STRIPE_EXPANDING, &sh->state) ||
5713 			    !add_stripe_bio(sh, bi, dd_idx, rw, previous)) {
5714 				/* Stripe is busy expanding or
5715 				 * add failed due to overlap.  Flush everything
5716 				 * and wait a while
5717 				 */
5718 				md_wakeup_thread(mddev->thread);
5719 				raid5_release_stripe(sh);
5720 				schedule();
5721 				do_prepare = true;
5722 				goto retry;
5723 			}
5724 			if (do_flush) {
5725 				set_bit(STRIPE_R5C_PREFLUSH, &sh->state);
5726 				/* we only need flush for one stripe */
5727 				do_flush = false;
5728 			}
5729 
5730 			set_bit(STRIPE_HANDLE, &sh->state);
5731 			clear_bit(STRIPE_DELAYED, &sh->state);
5732 			if ((!sh->batch_head || sh == sh->batch_head) &&
5733 			    (bi->bi_opf & REQ_SYNC) &&
5734 			    !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
5735 				atomic_inc(&conf->preread_active_stripes);
5736 			release_stripe_plug(mddev, sh);
5737 		} else {
5738 			/* cannot get stripe for read-ahead, just give-up */
5739 			bi->bi_status = BLK_STS_IOERR;
5740 			break;
5741 		}
5742 	}
5743 	finish_wait(&conf->wait_for_overlap, &w);
5744 
5745 	if (rw == WRITE)
5746 		md_write_end(mddev);
5747 	bio_endio(bi);
5748 	return true;
5749 }
5750 
5751 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
5752 
5753 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
5754 {
5755 	/* reshaping is quite different to recovery/resync so it is
5756 	 * handled quite separately ... here.
5757 	 *
5758 	 * On each call to sync_request, we gather one chunk worth of
5759 	 * destination stripes and flag them as expanding.
5760 	 * Then we find all the source stripes and request reads.
5761 	 * As the reads complete, handle_stripe will copy the data
5762 	 * into the destination stripe and release that stripe.
5763 	 */
5764 	struct r5conf *conf = mddev->private;
5765 	struct stripe_head *sh;
5766 	struct md_rdev *rdev;
5767 	sector_t first_sector, last_sector;
5768 	int raid_disks = conf->previous_raid_disks;
5769 	int data_disks = raid_disks - conf->max_degraded;
5770 	int new_data_disks = conf->raid_disks - conf->max_degraded;
5771 	int i;
5772 	int dd_idx;
5773 	sector_t writepos, readpos, safepos;
5774 	sector_t stripe_addr;
5775 	int reshape_sectors;
5776 	struct list_head stripes;
5777 	sector_t retn;
5778 
5779 	if (sector_nr == 0) {
5780 		/* If restarting in the middle, skip the initial sectors */
5781 		if (mddev->reshape_backwards &&
5782 		    conf->reshape_progress < raid5_size(mddev, 0, 0)) {
5783 			sector_nr = raid5_size(mddev, 0, 0)
5784 				- conf->reshape_progress;
5785 		} else if (mddev->reshape_backwards &&
5786 			   conf->reshape_progress == MaxSector) {
5787 			/* shouldn't happen, but just in case, finish up.*/
5788 			sector_nr = MaxSector;
5789 		} else if (!mddev->reshape_backwards &&
5790 			   conf->reshape_progress > 0)
5791 			sector_nr = conf->reshape_progress;
5792 		sector_div(sector_nr, new_data_disks);
5793 		if (sector_nr) {
5794 			mddev->curr_resync_completed = sector_nr;
5795 			sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5796 			*skipped = 1;
5797 			retn = sector_nr;
5798 			goto finish;
5799 		}
5800 	}
5801 
5802 	/* We need to process a full chunk at a time.
5803 	 * If old and new chunk sizes differ, we need to process the
5804 	 * largest of these
5805 	 */
5806 
5807 	reshape_sectors = max(conf->chunk_sectors, conf->prev_chunk_sectors);
5808 
5809 	/* We update the metadata at least every 10 seconds, or when
5810 	 * the data about to be copied would over-write the source of
5811 	 * the data at the front of the range.  i.e. one new_stripe
5812 	 * along from reshape_progress new_maps to after where
5813 	 * reshape_safe old_maps to
5814 	 */
5815 	writepos = conf->reshape_progress;
5816 	sector_div(writepos, new_data_disks);
5817 	readpos = conf->reshape_progress;
5818 	sector_div(readpos, data_disks);
5819 	safepos = conf->reshape_safe;
5820 	sector_div(safepos, data_disks);
5821 	if (mddev->reshape_backwards) {
5822 		BUG_ON(writepos < reshape_sectors);
5823 		writepos -= reshape_sectors;
5824 		readpos += reshape_sectors;
5825 		safepos += reshape_sectors;
5826 	} else {
5827 		writepos += reshape_sectors;
5828 		/* readpos and safepos are worst-case calculations.
5829 		 * A negative number is overly pessimistic, and causes
5830 		 * obvious problems for unsigned storage.  So clip to 0.
5831 		 */
5832 		readpos -= min_t(sector_t, reshape_sectors, readpos);
5833 		safepos -= min_t(sector_t, reshape_sectors, safepos);
5834 	}
5835 
5836 	/* Having calculated the 'writepos' possibly use it
5837 	 * to set 'stripe_addr' which is where we will write to.
5838 	 */
5839 	if (mddev->reshape_backwards) {
5840 		BUG_ON(conf->reshape_progress == 0);
5841 		stripe_addr = writepos;
5842 		BUG_ON((mddev->dev_sectors &
5843 			~((sector_t)reshape_sectors - 1))
5844 		       - reshape_sectors - stripe_addr
5845 		       != sector_nr);
5846 	} else {
5847 		BUG_ON(writepos != sector_nr + reshape_sectors);
5848 		stripe_addr = sector_nr;
5849 	}
5850 
5851 	/* 'writepos' is the most advanced device address we might write.
5852 	 * 'readpos' is the least advanced device address we might read.
5853 	 * 'safepos' is the least address recorded in the metadata as having
5854 	 *     been reshaped.
5855 	 * If there is a min_offset_diff, these are adjusted either by
5856 	 * increasing the safepos/readpos if diff is negative, or
5857 	 * increasing writepos if diff is positive.
5858 	 * If 'readpos' is then behind 'writepos', there is no way that we can
5859 	 * ensure safety in the face of a crash - that must be done by userspace
5860 	 * making a backup of the data.  So in that case there is no particular
5861 	 * rush to update metadata.
5862 	 * Otherwise if 'safepos' is behind 'writepos', then we really need to
5863 	 * update the metadata to advance 'safepos' to match 'readpos' so that
5864 	 * we can be safe in the event of a crash.
5865 	 * So we insist on updating metadata if safepos is behind writepos and
5866 	 * readpos is beyond writepos.
5867 	 * In any case, update the metadata every 10 seconds.
5868 	 * Maybe that number should be configurable, but I'm not sure it is
5869 	 * worth it.... maybe it could be a multiple of safemode_delay???
5870 	 */
5871 	if (conf->min_offset_diff < 0) {
5872 		safepos += -conf->min_offset_diff;
5873 		readpos += -conf->min_offset_diff;
5874 	} else
5875 		writepos += conf->min_offset_diff;
5876 
5877 	if ((mddev->reshape_backwards
5878 	     ? (safepos > writepos && readpos < writepos)
5879 	     : (safepos < writepos && readpos > writepos)) ||
5880 	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
5881 		/* Cannot proceed until we've updated the superblock... */
5882 		wait_event(conf->wait_for_overlap,
5883 			   atomic_read(&conf->reshape_stripes)==0
5884 			   || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5885 		if (atomic_read(&conf->reshape_stripes) != 0)
5886 			return 0;
5887 		mddev->reshape_position = conf->reshape_progress;
5888 		mddev->curr_resync_completed = sector_nr;
5889 		if (!mddev->reshape_backwards)
5890 			/* Can update recovery_offset */
5891 			rdev_for_each(rdev, mddev)
5892 				if (rdev->raid_disk >= 0 &&
5893 				    !test_bit(Journal, &rdev->flags) &&
5894 				    !test_bit(In_sync, &rdev->flags) &&
5895 				    rdev->recovery_offset < sector_nr)
5896 					rdev->recovery_offset = sector_nr;
5897 
5898 		conf->reshape_checkpoint = jiffies;
5899 		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
5900 		md_wakeup_thread(mddev->thread);
5901 		wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
5902 			   test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5903 		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5904 			return 0;
5905 		spin_lock_irq(&conf->device_lock);
5906 		conf->reshape_safe = mddev->reshape_position;
5907 		spin_unlock_irq(&conf->device_lock);
5908 		wake_up(&conf->wait_for_overlap);
5909 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
5910 	}
5911 
5912 	INIT_LIST_HEAD(&stripes);
5913 	for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
5914 		int j;
5915 		int skipped_disk = 0;
5916 		sh = raid5_get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
5917 		set_bit(STRIPE_EXPANDING, &sh->state);
5918 		atomic_inc(&conf->reshape_stripes);
5919 		/* If any of this stripe is beyond the end of the old
5920 		 * array, then we need to zero those blocks
5921 		 */
5922 		for (j=sh->disks; j--;) {
5923 			sector_t s;
5924 			if (j == sh->pd_idx)
5925 				continue;
5926 			if (conf->level == 6 &&
5927 			    j == sh->qd_idx)
5928 				continue;
5929 			s = raid5_compute_blocknr(sh, j, 0);
5930 			if (s < raid5_size(mddev, 0, 0)) {
5931 				skipped_disk = 1;
5932 				continue;
5933 			}
5934 			memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
5935 			set_bit(R5_Expanded, &sh->dev[j].flags);
5936 			set_bit(R5_UPTODATE, &sh->dev[j].flags);
5937 		}
5938 		if (!skipped_disk) {
5939 			set_bit(STRIPE_EXPAND_READY, &sh->state);
5940 			set_bit(STRIPE_HANDLE, &sh->state);
5941 		}
5942 		list_add(&sh->lru, &stripes);
5943 	}
5944 	spin_lock_irq(&conf->device_lock);
5945 	if (mddev->reshape_backwards)
5946 		conf->reshape_progress -= reshape_sectors * new_data_disks;
5947 	else
5948 		conf->reshape_progress += reshape_sectors * new_data_disks;
5949 	spin_unlock_irq(&conf->device_lock);
5950 	/* Ok, those stripe are ready. We can start scheduling
5951 	 * reads on the source stripes.
5952 	 * The source stripes are determined by mapping the first and last
5953 	 * block on the destination stripes.
5954 	 */
5955 	first_sector =
5956 		raid5_compute_sector(conf, stripe_addr*(new_data_disks),
5957 				     1, &dd_idx, NULL);
5958 	last_sector =
5959 		raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
5960 					    * new_data_disks - 1),
5961 				     1, &dd_idx, NULL);
5962 	if (last_sector >= mddev->dev_sectors)
5963 		last_sector = mddev->dev_sectors - 1;
5964 	while (first_sector <= last_sector) {
5965 		sh = raid5_get_active_stripe(conf, first_sector, 1, 0, 1);
5966 		set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
5967 		set_bit(STRIPE_HANDLE, &sh->state);
5968 		raid5_release_stripe(sh);
5969 		first_sector += STRIPE_SECTORS;
5970 	}
5971 	/* Now that the sources are clearly marked, we can release
5972 	 * the destination stripes
5973 	 */
5974 	while (!list_empty(&stripes)) {
5975 		sh = list_entry(stripes.next, struct stripe_head, lru);
5976 		list_del_init(&sh->lru);
5977 		raid5_release_stripe(sh);
5978 	}
5979 	/* If this takes us to the resync_max point where we have to pause,
5980 	 * then we need to write out the superblock.
5981 	 */
5982 	sector_nr += reshape_sectors;
5983 	retn = reshape_sectors;
5984 finish:
5985 	if (mddev->curr_resync_completed > mddev->resync_max ||
5986 	    (sector_nr - mddev->curr_resync_completed) * 2
5987 	    >= mddev->resync_max - mddev->curr_resync_completed) {
5988 		/* Cannot proceed until we've updated the superblock... */
5989 		wait_event(conf->wait_for_overlap,
5990 			   atomic_read(&conf->reshape_stripes) == 0
5991 			   || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
5992 		if (atomic_read(&conf->reshape_stripes) != 0)
5993 			goto ret;
5994 		mddev->reshape_position = conf->reshape_progress;
5995 		mddev->curr_resync_completed = sector_nr;
5996 		if (!mddev->reshape_backwards)
5997 			/* Can update recovery_offset */
5998 			rdev_for_each(rdev, mddev)
5999 				if (rdev->raid_disk >= 0 &&
6000 				    !test_bit(Journal, &rdev->flags) &&
6001 				    !test_bit(In_sync, &rdev->flags) &&
6002 				    rdev->recovery_offset < sector_nr)
6003 					rdev->recovery_offset = sector_nr;
6004 		conf->reshape_checkpoint = jiffies;
6005 		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
6006 		md_wakeup_thread(mddev->thread);
6007 		wait_event(mddev->sb_wait,
6008 			   !test_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags)
6009 			   || test_bit(MD_RECOVERY_INTR, &mddev->recovery));
6010 		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
6011 			goto ret;
6012 		spin_lock_irq(&conf->device_lock);
6013 		conf->reshape_safe = mddev->reshape_position;
6014 		spin_unlock_irq(&conf->device_lock);
6015 		wake_up(&conf->wait_for_overlap);
6016 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
6017 	}
6018 ret:
6019 	return retn;
6020 }
6021 
6022 static inline sector_t raid5_sync_request(struct mddev *mddev, sector_t sector_nr,
6023 					  int *skipped)
6024 {
6025 	struct r5conf *conf = mddev->private;
6026 	struct stripe_head *sh;
6027 	sector_t max_sector = mddev->dev_sectors;
6028 	sector_t sync_blocks;
6029 	int still_degraded = 0;
6030 	int i;
6031 
6032 	if (sector_nr >= max_sector) {
6033 		/* just being told to finish up .. nothing much to do */
6034 
6035 		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
6036 			end_reshape(conf);
6037 			return 0;
6038 		}
6039 
6040 		if (mddev->curr_resync < max_sector) /* aborted */
6041 			md_bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
6042 					   &sync_blocks, 1);
6043 		else /* completed sync */
6044 			conf->fullsync = 0;
6045 		md_bitmap_close_sync(mddev->bitmap);
6046 
6047 		return 0;
6048 	}
6049 
6050 	/* Allow raid5_quiesce to complete */
6051 	wait_event(conf->wait_for_overlap, conf->quiesce != 2);
6052 
6053 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
6054 		return reshape_request(mddev, sector_nr, skipped);
6055 
6056 	/* No need to check resync_max as we never do more than one
6057 	 * stripe, and as resync_max will always be on a chunk boundary,
6058 	 * if the check in md_do_sync didn't fire, there is no chance
6059 	 * of overstepping resync_max here
6060 	 */
6061 
6062 	/* if there is too many failed drives and we are trying
6063 	 * to resync, then assert that we are finished, because there is
6064 	 * nothing we can do.
6065 	 */
6066 	if (mddev->degraded >= conf->max_degraded &&
6067 	    test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
6068 		sector_t rv = mddev->dev_sectors - sector_nr;
6069 		*skipped = 1;
6070 		return rv;
6071 	}
6072 	if (!test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
6073 	    !conf->fullsync &&
6074 	    !md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
6075 	    sync_blocks >= STRIPE_SECTORS) {
6076 		/* we can skip this block, and probably more */
6077 		sync_blocks /= STRIPE_SECTORS;
6078 		*skipped = 1;
6079 		return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
6080 	}
6081 
6082 	md_bitmap_cond_end_sync(mddev->bitmap, sector_nr, false);
6083 
6084 	sh = raid5_get_active_stripe(conf, sector_nr, 0, 1, 0);
6085 	if (sh == NULL) {
6086 		sh = raid5_get_active_stripe(conf, sector_nr, 0, 0, 0);
6087 		/* make sure we don't swamp the stripe cache if someone else
6088 		 * is trying to get access
6089 		 */
6090 		schedule_timeout_uninterruptible(1);
6091 	}
6092 	/* Need to check if array will still be degraded after recovery/resync
6093 	 * Note in case of > 1 drive failures it's possible we're rebuilding
6094 	 * one drive while leaving another faulty drive in array.
6095 	 */
6096 	rcu_read_lock();
6097 	for (i = 0; i < conf->raid_disks; i++) {
6098 		struct md_rdev *rdev = READ_ONCE(conf->disks[i].rdev);
6099 
6100 		if (rdev == NULL || test_bit(Faulty, &rdev->flags))
6101 			still_degraded = 1;
6102 	}
6103 	rcu_read_unlock();
6104 
6105 	md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
6106 
6107 	set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
6108 	set_bit(STRIPE_HANDLE, &sh->state);
6109 
6110 	raid5_release_stripe(sh);
6111 
6112 	return STRIPE_SECTORS;
6113 }
6114 
6115 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio,
6116 			       unsigned int offset)
6117 {
6118 	/* We may not be able to submit a whole bio at once as there
6119 	 * may not be enough stripe_heads available.
6120 	 * We cannot pre-allocate enough stripe_heads as we may need
6121 	 * more than exist in the cache (if we allow ever large chunks).
6122 	 * So we do one stripe head at a time and record in
6123 	 * ->bi_hw_segments how many have been done.
6124 	 *
6125 	 * We *know* that this entire raid_bio is in one chunk, so
6126 	 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
6127 	 */
6128 	struct stripe_head *sh;
6129 	int dd_idx;
6130 	sector_t sector, logical_sector, last_sector;
6131 	int scnt = 0;
6132 	int handled = 0;
6133 
6134 	logical_sector = raid_bio->bi_iter.bi_sector &
6135 		~((sector_t)STRIPE_SECTORS-1);
6136 	sector = raid5_compute_sector(conf, logical_sector,
6137 				      0, &dd_idx, NULL);
6138 	last_sector = bio_end_sector(raid_bio);
6139 
6140 	for (; logical_sector < last_sector;
6141 	     logical_sector += STRIPE_SECTORS,
6142 		     sector += STRIPE_SECTORS,
6143 		     scnt++) {
6144 
6145 		if (scnt < offset)
6146 			/* already done this stripe */
6147 			continue;
6148 
6149 		sh = raid5_get_active_stripe(conf, sector, 0, 1, 1);
6150 
6151 		if (!sh) {
6152 			/* failed to get a stripe - must wait */
6153 			conf->retry_read_aligned = raid_bio;
6154 			conf->retry_read_offset = scnt;
6155 			return handled;
6156 		}
6157 
6158 		if (!add_stripe_bio(sh, raid_bio, dd_idx, 0, 0)) {
6159 			raid5_release_stripe(sh);
6160 			conf->retry_read_aligned = raid_bio;
6161 			conf->retry_read_offset = scnt;
6162 			return handled;
6163 		}
6164 
6165 		set_bit(R5_ReadNoMerge, &sh->dev[dd_idx].flags);
6166 		handle_stripe(sh);
6167 		raid5_release_stripe(sh);
6168 		handled++;
6169 	}
6170 
6171 	bio_endio(raid_bio);
6172 
6173 	if (atomic_dec_and_test(&conf->active_aligned_reads))
6174 		wake_up(&conf->wait_for_quiescent);
6175 	return handled;
6176 }
6177 
6178 static int handle_active_stripes(struct r5conf *conf, int group,
6179 				 struct r5worker *worker,
6180 				 struct list_head *temp_inactive_list)
6181 		__releases(&conf->device_lock)
6182 		__acquires(&conf->device_lock)
6183 {
6184 	struct stripe_head *batch[MAX_STRIPE_BATCH], *sh;
6185 	int i, batch_size = 0, hash;
6186 	bool release_inactive = false;
6187 
6188 	while (batch_size < MAX_STRIPE_BATCH &&
6189 			(sh = __get_priority_stripe(conf, group)) != NULL)
6190 		batch[batch_size++] = sh;
6191 
6192 	if (batch_size == 0) {
6193 		for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
6194 			if (!list_empty(temp_inactive_list + i))
6195 				break;
6196 		if (i == NR_STRIPE_HASH_LOCKS) {
6197 			spin_unlock_irq(&conf->device_lock);
6198 			log_flush_stripe_to_raid(conf);
6199 			spin_lock_irq(&conf->device_lock);
6200 			return batch_size;
6201 		}
6202 		release_inactive = true;
6203 	}
6204 	spin_unlock_irq(&conf->device_lock);
6205 
6206 	release_inactive_stripe_list(conf, temp_inactive_list,
6207 				     NR_STRIPE_HASH_LOCKS);
6208 
6209 	r5l_flush_stripe_to_raid(conf->log);
6210 	if (release_inactive) {
6211 		spin_lock_irq(&conf->device_lock);
6212 		return 0;
6213 	}
6214 
6215 	for (i = 0; i < batch_size; i++)
6216 		handle_stripe(batch[i]);
6217 	log_write_stripe_run(conf);
6218 
6219 	cond_resched();
6220 
6221 	spin_lock_irq(&conf->device_lock);
6222 	for (i = 0; i < batch_size; i++) {
6223 		hash = batch[i]->hash_lock_index;
6224 		__release_stripe(conf, batch[i], &temp_inactive_list[hash]);
6225 	}
6226 	return batch_size;
6227 }
6228 
6229 static void raid5_do_work(struct work_struct *work)
6230 {
6231 	struct r5worker *worker = container_of(work, struct r5worker, work);
6232 	struct r5worker_group *group = worker->group;
6233 	struct r5conf *conf = group->conf;
6234 	struct mddev *mddev = conf->mddev;
6235 	int group_id = group - conf->worker_groups;
6236 	int handled;
6237 	struct blk_plug plug;
6238 
6239 	pr_debug("+++ raid5worker active\n");
6240 
6241 	blk_start_plug(&plug);
6242 	handled = 0;
6243 	spin_lock_irq(&conf->device_lock);
6244 	while (1) {
6245 		int batch_size, released;
6246 
6247 		released = release_stripe_list(conf, worker->temp_inactive_list);
6248 
6249 		batch_size = handle_active_stripes(conf, group_id, worker,
6250 						   worker->temp_inactive_list);
6251 		worker->working = false;
6252 		if (!batch_size && !released)
6253 			break;
6254 		handled += batch_size;
6255 		wait_event_lock_irq(mddev->sb_wait,
6256 			!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags),
6257 			conf->device_lock);
6258 	}
6259 	pr_debug("%d stripes handled\n", handled);
6260 
6261 	spin_unlock_irq(&conf->device_lock);
6262 
6263 	flush_deferred_bios(conf);
6264 
6265 	r5l_flush_stripe_to_raid(conf->log);
6266 
6267 	async_tx_issue_pending_all();
6268 	blk_finish_plug(&plug);
6269 
6270 	pr_debug("--- raid5worker inactive\n");
6271 }
6272 
6273 /*
6274  * This is our raid5 kernel thread.
6275  *
6276  * We scan the hash table for stripes which can be handled now.
6277  * During the scan, completed stripes are saved for us by the interrupt
6278  * handler, so that they will not have to wait for our next wakeup.
6279  */
6280 static void raid5d(struct md_thread *thread)
6281 {
6282 	struct mddev *mddev = thread->mddev;
6283 	struct r5conf *conf = mddev->private;
6284 	int handled;
6285 	struct blk_plug plug;
6286 
6287 	pr_debug("+++ raid5d active\n");
6288 
6289 	md_check_recovery(mddev);
6290 
6291 	blk_start_plug(&plug);
6292 	handled = 0;
6293 	spin_lock_irq(&conf->device_lock);
6294 	while (1) {
6295 		struct bio *bio;
6296 		int batch_size, released;
6297 		unsigned int offset;
6298 
6299 		released = release_stripe_list(conf, conf->temp_inactive_list);
6300 		if (released)
6301 			clear_bit(R5_DID_ALLOC, &conf->cache_state);
6302 
6303 		if (
6304 		    !list_empty(&conf->bitmap_list)) {
6305 			/* Now is a good time to flush some bitmap updates */
6306 			conf->seq_flush++;
6307 			spin_unlock_irq(&conf->device_lock);
6308 			md_bitmap_unplug(mddev->bitmap);
6309 			spin_lock_irq(&conf->device_lock);
6310 			conf->seq_write = conf->seq_flush;
6311 			activate_bit_delay(conf, conf->temp_inactive_list);
6312 		}
6313 		raid5_activate_delayed(conf);
6314 
6315 		while ((bio = remove_bio_from_retry(conf, &offset))) {
6316 			int ok;
6317 			spin_unlock_irq(&conf->device_lock);
6318 			ok = retry_aligned_read(conf, bio, offset);
6319 			spin_lock_irq(&conf->device_lock);
6320 			if (!ok)
6321 				break;
6322 			handled++;
6323 		}
6324 
6325 		batch_size = handle_active_stripes(conf, ANY_GROUP, NULL,
6326 						   conf->temp_inactive_list);
6327 		if (!batch_size && !released)
6328 			break;
6329 		handled += batch_size;
6330 
6331 		if (mddev->sb_flags & ~(1 << MD_SB_CHANGE_PENDING)) {
6332 			spin_unlock_irq(&conf->device_lock);
6333 			md_check_recovery(mddev);
6334 			spin_lock_irq(&conf->device_lock);
6335 		}
6336 	}
6337 	pr_debug("%d stripes handled\n", handled);
6338 
6339 	spin_unlock_irq(&conf->device_lock);
6340 	if (test_and_clear_bit(R5_ALLOC_MORE, &conf->cache_state) &&
6341 	    mutex_trylock(&conf->cache_size_mutex)) {
6342 		grow_one_stripe(conf, __GFP_NOWARN);
6343 		/* Set flag even if allocation failed.  This helps
6344 		 * slow down allocation requests when mem is short
6345 		 */
6346 		set_bit(R5_DID_ALLOC, &conf->cache_state);
6347 		mutex_unlock(&conf->cache_size_mutex);
6348 	}
6349 
6350 	flush_deferred_bios(conf);
6351 
6352 	r5l_flush_stripe_to_raid(conf->log);
6353 
6354 	async_tx_issue_pending_all();
6355 	blk_finish_plug(&plug);
6356 
6357 	pr_debug("--- raid5d inactive\n");
6358 }
6359 
6360 static ssize_t
6361 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
6362 {
6363 	struct r5conf *conf;
6364 	int ret = 0;
6365 	spin_lock(&mddev->lock);
6366 	conf = mddev->private;
6367 	if (conf)
6368 		ret = sprintf(page, "%d\n", conf->min_nr_stripes);
6369 	spin_unlock(&mddev->lock);
6370 	return ret;
6371 }
6372 
6373 int
6374 raid5_set_cache_size(struct mddev *mddev, int size)
6375 {
6376 	int result = 0;
6377 	struct r5conf *conf = mddev->private;
6378 
6379 	if (size <= 16 || size > 32768)
6380 		return -EINVAL;
6381 
6382 	conf->min_nr_stripes = size;
6383 	mutex_lock(&conf->cache_size_mutex);
6384 	while (size < conf->max_nr_stripes &&
6385 	       drop_one_stripe(conf))
6386 		;
6387 	mutex_unlock(&conf->cache_size_mutex);
6388 
6389 	md_allow_write(mddev);
6390 
6391 	mutex_lock(&conf->cache_size_mutex);
6392 	while (size > conf->max_nr_stripes)
6393 		if (!grow_one_stripe(conf, GFP_KERNEL)) {
6394 			conf->min_nr_stripes = conf->max_nr_stripes;
6395 			result = -ENOMEM;
6396 			break;
6397 		}
6398 	mutex_unlock(&conf->cache_size_mutex);
6399 
6400 	return result;
6401 }
6402 EXPORT_SYMBOL(raid5_set_cache_size);
6403 
6404 static ssize_t
6405 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
6406 {
6407 	struct r5conf *conf;
6408 	unsigned long new;
6409 	int err;
6410 
6411 	if (len >= PAGE_SIZE)
6412 		return -EINVAL;
6413 	if (kstrtoul(page, 10, &new))
6414 		return -EINVAL;
6415 	err = mddev_lock(mddev);
6416 	if (err)
6417 		return err;
6418 	conf = mddev->private;
6419 	if (!conf)
6420 		err = -ENODEV;
6421 	else
6422 		err = raid5_set_cache_size(mddev, new);
6423 	mddev_unlock(mddev);
6424 
6425 	return err ?: len;
6426 }
6427 
6428 static struct md_sysfs_entry
6429 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
6430 				raid5_show_stripe_cache_size,
6431 				raid5_store_stripe_cache_size);
6432 
6433 static ssize_t
6434 raid5_show_rmw_level(struct mddev  *mddev, char *page)
6435 {
6436 	struct r5conf *conf = mddev->private;
6437 	if (conf)
6438 		return sprintf(page, "%d\n", conf->rmw_level);
6439 	else
6440 		return 0;
6441 }
6442 
6443 static ssize_t
6444 raid5_store_rmw_level(struct mddev  *mddev, const char *page, size_t len)
6445 {
6446 	struct r5conf *conf = mddev->private;
6447 	unsigned long new;
6448 
6449 	if (!conf)
6450 		return -ENODEV;
6451 
6452 	if (len >= PAGE_SIZE)
6453 		return -EINVAL;
6454 
6455 	if (kstrtoul(page, 10, &new))
6456 		return -EINVAL;
6457 
6458 	if (new != PARITY_DISABLE_RMW && !raid6_call.xor_syndrome)
6459 		return -EINVAL;
6460 
6461 	if (new != PARITY_DISABLE_RMW &&
6462 	    new != PARITY_ENABLE_RMW &&
6463 	    new != PARITY_PREFER_RMW)
6464 		return -EINVAL;
6465 
6466 	conf->rmw_level = new;
6467 	return len;
6468 }
6469 
6470 static struct md_sysfs_entry
6471 raid5_rmw_level = __ATTR(rmw_level, S_IRUGO | S_IWUSR,
6472 			 raid5_show_rmw_level,
6473 			 raid5_store_rmw_level);
6474 
6475 
6476 static ssize_t
6477 raid5_show_preread_threshold(struct mddev *mddev, char *page)
6478 {
6479 	struct r5conf *conf;
6480 	int ret = 0;
6481 	spin_lock(&mddev->lock);
6482 	conf = mddev->private;
6483 	if (conf)
6484 		ret = sprintf(page, "%d\n", conf->bypass_threshold);
6485 	spin_unlock(&mddev->lock);
6486 	return ret;
6487 }
6488 
6489 static ssize_t
6490 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
6491 {
6492 	struct r5conf *conf;
6493 	unsigned long new;
6494 	int err;
6495 
6496 	if (len >= PAGE_SIZE)
6497 		return -EINVAL;
6498 	if (kstrtoul(page, 10, &new))
6499 		return -EINVAL;
6500 
6501 	err = mddev_lock(mddev);
6502 	if (err)
6503 		return err;
6504 	conf = mddev->private;
6505 	if (!conf)
6506 		err = -ENODEV;
6507 	else if (new > conf->min_nr_stripes)
6508 		err = -EINVAL;
6509 	else
6510 		conf->bypass_threshold = new;
6511 	mddev_unlock(mddev);
6512 	return err ?: len;
6513 }
6514 
6515 static struct md_sysfs_entry
6516 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
6517 					S_IRUGO | S_IWUSR,
6518 					raid5_show_preread_threshold,
6519 					raid5_store_preread_threshold);
6520 
6521 static ssize_t
6522 raid5_show_skip_copy(struct mddev *mddev, char *page)
6523 {
6524 	struct r5conf *conf;
6525 	int ret = 0;
6526 	spin_lock(&mddev->lock);
6527 	conf = mddev->private;
6528 	if (conf)
6529 		ret = sprintf(page, "%d\n", conf->skip_copy);
6530 	spin_unlock(&mddev->lock);
6531 	return ret;
6532 }
6533 
6534 static ssize_t
6535 raid5_store_skip_copy(struct mddev *mddev, const char *page, size_t len)
6536 {
6537 	struct r5conf *conf;
6538 	unsigned long new;
6539 	int err;
6540 
6541 	if (len >= PAGE_SIZE)
6542 		return -EINVAL;
6543 	if (kstrtoul(page, 10, &new))
6544 		return -EINVAL;
6545 	new = !!new;
6546 
6547 	err = mddev_lock(mddev);
6548 	if (err)
6549 		return err;
6550 	conf = mddev->private;
6551 	if (!conf)
6552 		err = -ENODEV;
6553 	else if (new != conf->skip_copy) {
6554 		mddev_suspend(mddev);
6555 		conf->skip_copy = new;
6556 		if (new)
6557 			mddev->queue->backing_dev_info->capabilities |=
6558 				BDI_CAP_STABLE_WRITES;
6559 		else
6560 			mddev->queue->backing_dev_info->capabilities &=
6561 				~BDI_CAP_STABLE_WRITES;
6562 		mddev_resume(mddev);
6563 	}
6564 	mddev_unlock(mddev);
6565 	return err ?: len;
6566 }
6567 
6568 static struct md_sysfs_entry
6569 raid5_skip_copy = __ATTR(skip_copy, S_IRUGO | S_IWUSR,
6570 					raid5_show_skip_copy,
6571 					raid5_store_skip_copy);
6572 
6573 static ssize_t
6574 stripe_cache_active_show(struct mddev *mddev, char *page)
6575 {
6576 	struct r5conf *conf = mddev->private;
6577 	if (conf)
6578 		return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
6579 	else
6580 		return 0;
6581 }
6582 
6583 static struct md_sysfs_entry
6584 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
6585 
6586 static ssize_t
6587 raid5_show_group_thread_cnt(struct mddev *mddev, char *page)
6588 {
6589 	struct r5conf *conf;
6590 	int ret = 0;
6591 	spin_lock(&mddev->lock);
6592 	conf = mddev->private;
6593 	if (conf)
6594 		ret = sprintf(page, "%d\n", conf->worker_cnt_per_group);
6595 	spin_unlock(&mddev->lock);
6596 	return ret;
6597 }
6598 
6599 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6600 			       int *group_cnt,
6601 			       int *worker_cnt_per_group,
6602 			       struct r5worker_group **worker_groups);
6603 static ssize_t
6604 raid5_store_group_thread_cnt(struct mddev *mddev, const char *page, size_t len)
6605 {
6606 	struct r5conf *conf;
6607 	unsigned int new;
6608 	int err;
6609 	struct r5worker_group *new_groups, *old_groups;
6610 	int group_cnt, worker_cnt_per_group;
6611 
6612 	if (len >= PAGE_SIZE)
6613 		return -EINVAL;
6614 	if (kstrtouint(page, 10, &new))
6615 		return -EINVAL;
6616 	/* 8192 should be big enough */
6617 	if (new > 8192)
6618 		return -EINVAL;
6619 
6620 	err = mddev_lock(mddev);
6621 	if (err)
6622 		return err;
6623 	conf = mddev->private;
6624 	if (!conf)
6625 		err = -ENODEV;
6626 	else if (new != conf->worker_cnt_per_group) {
6627 		mddev_suspend(mddev);
6628 
6629 		old_groups = conf->worker_groups;
6630 		if (old_groups)
6631 			flush_workqueue(raid5_wq);
6632 
6633 		err = alloc_thread_groups(conf, new,
6634 					  &group_cnt, &worker_cnt_per_group,
6635 					  &new_groups);
6636 		if (!err) {
6637 			spin_lock_irq(&conf->device_lock);
6638 			conf->group_cnt = group_cnt;
6639 			conf->worker_cnt_per_group = worker_cnt_per_group;
6640 			conf->worker_groups = new_groups;
6641 			spin_unlock_irq(&conf->device_lock);
6642 
6643 			if (old_groups)
6644 				kfree(old_groups[0].workers);
6645 			kfree(old_groups);
6646 		}
6647 		mddev_resume(mddev);
6648 	}
6649 	mddev_unlock(mddev);
6650 
6651 	return err ?: len;
6652 }
6653 
6654 static struct md_sysfs_entry
6655 raid5_group_thread_cnt = __ATTR(group_thread_cnt, S_IRUGO | S_IWUSR,
6656 				raid5_show_group_thread_cnt,
6657 				raid5_store_group_thread_cnt);
6658 
6659 static struct attribute *raid5_attrs[] =  {
6660 	&raid5_stripecache_size.attr,
6661 	&raid5_stripecache_active.attr,
6662 	&raid5_preread_bypass_threshold.attr,
6663 	&raid5_group_thread_cnt.attr,
6664 	&raid5_skip_copy.attr,
6665 	&raid5_rmw_level.attr,
6666 	&r5c_journal_mode.attr,
6667 	&ppl_write_hint.attr,
6668 	NULL,
6669 };
6670 static struct attribute_group raid5_attrs_group = {
6671 	.name = NULL,
6672 	.attrs = raid5_attrs,
6673 };
6674 
6675 static int alloc_thread_groups(struct r5conf *conf, int cnt,
6676 			       int *group_cnt,
6677 			       int *worker_cnt_per_group,
6678 			       struct r5worker_group **worker_groups)
6679 {
6680 	int i, j, k;
6681 	ssize_t size;
6682 	struct r5worker *workers;
6683 
6684 	*worker_cnt_per_group = cnt;
6685 	if (cnt == 0) {
6686 		*group_cnt = 0;
6687 		*worker_groups = NULL;
6688 		return 0;
6689 	}
6690 	*group_cnt = num_possible_nodes();
6691 	size = sizeof(struct r5worker) * cnt;
6692 	workers = kcalloc(size, *group_cnt, GFP_NOIO);
6693 	*worker_groups = kcalloc(*group_cnt, sizeof(struct r5worker_group),
6694 				 GFP_NOIO);
6695 	if (!*worker_groups || !workers) {
6696 		kfree(workers);
6697 		kfree(*worker_groups);
6698 		return -ENOMEM;
6699 	}
6700 
6701 	for (i = 0; i < *group_cnt; i++) {
6702 		struct r5worker_group *group;
6703 
6704 		group = &(*worker_groups)[i];
6705 		INIT_LIST_HEAD(&group->handle_list);
6706 		INIT_LIST_HEAD(&group->loprio_list);
6707 		group->conf = conf;
6708 		group->workers = workers + i * cnt;
6709 
6710 		for (j = 0; j < cnt; j++) {
6711 			struct r5worker *worker = group->workers + j;
6712 			worker->group = group;
6713 			INIT_WORK(&worker->work, raid5_do_work);
6714 
6715 			for (k = 0; k < NR_STRIPE_HASH_LOCKS; k++)
6716 				INIT_LIST_HEAD(worker->temp_inactive_list + k);
6717 		}
6718 	}
6719 
6720 	return 0;
6721 }
6722 
6723 static void free_thread_groups(struct r5conf *conf)
6724 {
6725 	if (conf->worker_groups)
6726 		kfree(conf->worker_groups[0].workers);
6727 	kfree(conf->worker_groups);
6728 	conf->worker_groups = NULL;
6729 }
6730 
6731 static sector_t
6732 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
6733 {
6734 	struct r5conf *conf = mddev->private;
6735 
6736 	if (!sectors)
6737 		sectors = mddev->dev_sectors;
6738 	if (!raid_disks)
6739 		/* size is defined by the smallest of previous and new size */
6740 		raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
6741 
6742 	sectors &= ~((sector_t)conf->chunk_sectors - 1);
6743 	sectors &= ~((sector_t)conf->prev_chunk_sectors - 1);
6744 	return sectors * (raid_disks - conf->max_degraded);
6745 }
6746 
6747 static void free_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6748 {
6749 	safe_put_page(percpu->spare_page);
6750 	percpu->spare_page = NULL;
6751 	kvfree(percpu->scribble);
6752 	percpu->scribble = NULL;
6753 }
6754 
6755 static int alloc_scratch_buffer(struct r5conf *conf, struct raid5_percpu *percpu)
6756 {
6757 	if (conf->level == 6 && !percpu->spare_page) {
6758 		percpu->spare_page = alloc_page(GFP_KERNEL);
6759 		if (!percpu->spare_page)
6760 			return -ENOMEM;
6761 	}
6762 
6763 	if (scribble_alloc(percpu,
6764 			   max(conf->raid_disks,
6765 			       conf->previous_raid_disks),
6766 			   max(conf->chunk_sectors,
6767 			       conf->prev_chunk_sectors)
6768 			   / STRIPE_SECTORS,
6769 			   GFP_KERNEL)) {
6770 		free_scratch_buffer(conf, percpu);
6771 		return -ENOMEM;
6772 	}
6773 
6774 	return 0;
6775 }
6776 
6777 static int raid456_cpu_dead(unsigned int cpu, struct hlist_node *node)
6778 {
6779 	struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6780 
6781 	free_scratch_buffer(conf, per_cpu_ptr(conf->percpu, cpu));
6782 	return 0;
6783 }
6784 
6785 static void raid5_free_percpu(struct r5conf *conf)
6786 {
6787 	if (!conf->percpu)
6788 		return;
6789 
6790 	cpuhp_state_remove_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6791 	free_percpu(conf->percpu);
6792 }
6793 
6794 static void free_conf(struct r5conf *conf)
6795 {
6796 	int i;
6797 
6798 	log_exit(conf);
6799 
6800 	unregister_shrinker(&conf->shrinker);
6801 	free_thread_groups(conf);
6802 	shrink_stripes(conf);
6803 	raid5_free_percpu(conf);
6804 	for (i = 0; i < conf->pool_size; i++)
6805 		if (conf->disks[i].extra_page)
6806 			put_page(conf->disks[i].extra_page);
6807 	kfree(conf->disks);
6808 	bioset_exit(&conf->bio_split);
6809 	kfree(conf->stripe_hashtbl);
6810 	kfree(conf->pending_data);
6811 	kfree(conf);
6812 }
6813 
6814 static int raid456_cpu_up_prepare(unsigned int cpu, struct hlist_node *node)
6815 {
6816 	struct r5conf *conf = hlist_entry_safe(node, struct r5conf, node);
6817 	struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
6818 
6819 	if (alloc_scratch_buffer(conf, percpu)) {
6820 		pr_warn("%s: failed memory allocation for cpu%u\n",
6821 			__func__, cpu);
6822 		return -ENOMEM;
6823 	}
6824 	return 0;
6825 }
6826 
6827 static int raid5_alloc_percpu(struct r5conf *conf)
6828 {
6829 	int err = 0;
6830 
6831 	conf->percpu = alloc_percpu(struct raid5_percpu);
6832 	if (!conf->percpu)
6833 		return -ENOMEM;
6834 
6835 	err = cpuhp_state_add_instance(CPUHP_MD_RAID5_PREPARE, &conf->node);
6836 	if (!err) {
6837 		conf->scribble_disks = max(conf->raid_disks,
6838 			conf->previous_raid_disks);
6839 		conf->scribble_sectors = max(conf->chunk_sectors,
6840 			conf->prev_chunk_sectors);
6841 	}
6842 	return err;
6843 }
6844 
6845 static unsigned long raid5_cache_scan(struct shrinker *shrink,
6846 				      struct shrink_control *sc)
6847 {
6848 	struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6849 	unsigned long ret = SHRINK_STOP;
6850 
6851 	if (mutex_trylock(&conf->cache_size_mutex)) {
6852 		ret= 0;
6853 		while (ret < sc->nr_to_scan &&
6854 		       conf->max_nr_stripes > conf->min_nr_stripes) {
6855 			if (drop_one_stripe(conf) == 0) {
6856 				ret = SHRINK_STOP;
6857 				break;
6858 			}
6859 			ret++;
6860 		}
6861 		mutex_unlock(&conf->cache_size_mutex);
6862 	}
6863 	return ret;
6864 }
6865 
6866 static unsigned long raid5_cache_count(struct shrinker *shrink,
6867 				       struct shrink_control *sc)
6868 {
6869 	struct r5conf *conf = container_of(shrink, struct r5conf, shrinker);
6870 
6871 	if (conf->max_nr_stripes < conf->min_nr_stripes)
6872 		/* unlikely, but not impossible */
6873 		return 0;
6874 	return conf->max_nr_stripes - conf->min_nr_stripes;
6875 }
6876 
6877 static struct r5conf *setup_conf(struct mddev *mddev)
6878 {
6879 	struct r5conf *conf;
6880 	int raid_disk, memory, max_disks;
6881 	struct md_rdev *rdev;
6882 	struct disk_info *disk;
6883 	char pers_name[6];
6884 	int i;
6885 	int group_cnt, worker_cnt_per_group;
6886 	struct r5worker_group *new_group;
6887 	int ret;
6888 
6889 	if (mddev->new_level != 5
6890 	    && mddev->new_level != 4
6891 	    && mddev->new_level != 6) {
6892 		pr_warn("md/raid:%s: raid level not set to 4/5/6 (%d)\n",
6893 			mdname(mddev), mddev->new_level);
6894 		return ERR_PTR(-EIO);
6895 	}
6896 	if ((mddev->new_level == 5
6897 	     && !algorithm_valid_raid5(mddev->new_layout)) ||
6898 	    (mddev->new_level == 6
6899 	     && !algorithm_valid_raid6(mddev->new_layout))) {
6900 		pr_warn("md/raid:%s: layout %d not supported\n",
6901 			mdname(mddev), mddev->new_layout);
6902 		return ERR_PTR(-EIO);
6903 	}
6904 	if (mddev->new_level == 6 && mddev->raid_disks < 4) {
6905 		pr_warn("md/raid:%s: not enough configured devices (%d, minimum 4)\n",
6906 			mdname(mddev), mddev->raid_disks);
6907 		return ERR_PTR(-EINVAL);
6908 	}
6909 
6910 	if (!mddev->new_chunk_sectors ||
6911 	    (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
6912 	    !is_power_of_2(mddev->new_chunk_sectors)) {
6913 		pr_warn("md/raid:%s: invalid chunk size %d\n",
6914 			mdname(mddev), mddev->new_chunk_sectors << 9);
6915 		return ERR_PTR(-EINVAL);
6916 	}
6917 
6918 	conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
6919 	if (conf == NULL)
6920 		goto abort;
6921 	INIT_LIST_HEAD(&conf->free_list);
6922 	INIT_LIST_HEAD(&conf->pending_list);
6923 	conf->pending_data = kcalloc(PENDING_IO_MAX,
6924 				     sizeof(struct r5pending_data),
6925 				     GFP_KERNEL);
6926 	if (!conf->pending_data)
6927 		goto abort;
6928 	for (i = 0; i < PENDING_IO_MAX; i++)
6929 		list_add(&conf->pending_data[i].sibling, &conf->free_list);
6930 	/* Don't enable multi-threading by default*/
6931 	if (!alloc_thread_groups(conf, 0, &group_cnt, &worker_cnt_per_group,
6932 				 &new_group)) {
6933 		conf->group_cnt = group_cnt;
6934 		conf->worker_cnt_per_group = worker_cnt_per_group;
6935 		conf->worker_groups = new_group;
6936 	} else
6937 		goto abort;
6938 	spin_lock_init(&conf->device_lock);
6939 	seqcount_init(&conf->gen_lock);
6940 	mutex_init(&conf->cache_size_mutex);
6941 	init_waitqueue_head(&conf->wait_for_quiescent);
6942 	init_waitqueue_head(&conf->wait_for_stripe);
6943 	init_waitqueue_head(&conf->wait_for_overlap);
6944 	INIT_LIST_HEAD(&conf->handle_list);
6945 	INIT_LIST_HEAD(&conf->loprio_list);
6946 	INIT_LIST_HEAD(&conf->hold_list);
6947 	INIT_LIST_HEAD(&conf->delayed_list);
6948 	INIT_LIST_HEAD(&conf->bitmap_list);
6949 	init_llist_head(&conf->released_stripes);
6950 	atomic_set(&conf->active_stripes, 0);
6951 	atomic_set(&conf->preread_active_stripes, 0);
6952 	atomic_set(&conf->active_aligned_reads, 0);
6953 	spin_lock_init(&conf->pending_bios_lock);
6954 	conf->batch_bio_dispatch = true;
6955 	rdev_for_each(rdev, mddev) {
6956 		if (test_bit(Journal, &rdev->flags))
6957 			continue;
6958 		if (blk_queue_nonrot(bdev_get_queue(rdev->bdev))) {
6959 			conf->batch_bio_dispatch = false;
6960 			break;
6961 		}
6962 	}
6963 
6964 	conf->bypass_threshold = BYPASS_THRESHOLD;
6965 	conf->recovery_disabled = mddev->recovery_disabled - 1;
6966 
6967 	conf->raid_disks = mddev->raid_disks;
6968 	if (mddev->reshape_position == MaxSector)
6969 		conf->previous_raid_disks = mddev->raid_disks;
6970 	else
6971 		conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
6972 	max_disks = max(conf->raid_disks, conf->previous_raid_disks);
6973 
6974 	conf->disks = kcalloc(max_disks, sizeof(struct disk_info),
6975 			      GFP_KERNEL);
6976 
6977 	if (!conf->disks)
6978 		goto abort;
6979 
6980 	for (i = 0; i < max_disks; i++) {
6981 		conf->disks[i].extra_page = alloc_page(GFP_KERNEL);
6982 		if (!conf->disks[i].extra_page)
6983 			goto abort;
6984 	}
6985 
6986 	ret = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
6987 	if (ret)
6988 		goto abort;
6989 	conf->mddev = mddev;
6990 
6991 	if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
6992 		goto abort;
6993 
6994 	/* We init hash_locks[0] separately to that it can be used
6995 	 * as the reference lock in the spin_lock_nest_lock() call
6996 	 * in lock_all_device_hash_locks_irq in order to convince
6997 	 * lockdep that we know what we are doing.
6998 	 */
6999 	spin_lock_init(conf->hash_locks);
7000 	for (i = 1; i < NR_STRIPE_HASH_LOCKS; i++)
7001 		spin_lock_init(conf->hash_locks + i);
7002 
7003 	for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7004 		INIT_LIST_HEAD(conf->inactive_list + i);
7005 
7006 	for (i = 0; i < NR_STRIPE_HASH_LOCKS; i++)
7007 		INIT_LIST_HEAD(conf->temp_inactive_list + i);
7008 
7009 	atomic_set(&conf->r5c_cached_full_stripes, 0);
7010 	INIT_LIST_HEAD(&conf->r5c_full_stripe_list);
7011 	atomic_set(&conf->r5c_cached_partial_stripes, 0);
7012 	INIT_LIST_HEAD(&conf->r5c_partial_stripe_list);
7013 	atomic_set(&conf->r5c_flushing_full_stripes, 0);
7014 	atomic_set(&conf->r5c_flushing_partial_stripes, 0);
7015 
7016 	conf->level = mddev->new_level;
7017 	conf->chunk_sectors = mddev->new_chunk_sectors;
7018 	if (raid5_alloc_percpu(conf) != 0)
7019 		goto abort;
7020 
7021 	pr_debug("raid456: run(%s) called.\n", mdname(mddev));
7022 
7023 	rdev_for_each(rdev, mddev) {
7024 		raid_disk = rdev->raid_disk;
7025 		if (raid_disk >= max_disks
7026 		    || raid_disk < 0 || test_bit(Journal, &rdev->flags))
7027 			continue;
7028 		disk = conf->disks + raid_disk;
7029 
7030 		if (test_bit(Replacement, &rdev->flags)) {
7031 			if (disk->replacement)
7032 				goto abort;
7033 			disk->replacement = rdev;
7034 		} else {
7035 			if (disk->rdev)
7036 				goto abort;
7037 			disk->rdev = rdev;
7038 		}
7039 
7040 		if (test_bit(In_sync, &rdev->flags)) {
7041 			char b[BDEVNAME_SIZE];
7042 			pr_info("md/raid:%s: device %s operational as raid disk %d\n",
7043 				mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
7044 		} else if (rdev->saved_raid_disk != raid_disk)
7045 			/* Cannot rely on bitmap to complete recovery */
7046 			conf->fullsync = 1;
7047 	}
7048 
7049 	conf->level = mddev->new_level;
7050 	if (conf->level == 6) {
7051 		conf->max_degraded = 2;
7052 		if (raid6_call.xor_syndrome)
7053 			conf->rmw_level = PARITY_ENABLE_RMW;
7054 		else
7055 			conf->rmw_level = PARITY_DISABLE_RMW;
7056 	} else {
7057 		conf->max_degraded = 1;
7058 		conf->rmw_level = PARITY_ENABLE_RMW;
7059 	}
7060 	conf->algorithm = mddev->new_layout;
7061 	conf->reshape_progress = mddev->reshape_position;
7062 	if (conf->reshape_progress != MaxSector) {
7063 		conf->prev_chunk_sectors = mddev->chunk_sectors;
7064 		conf->prev_algo = mddev->layout;
7065 	} else {
7066 		conf->prev_chunk_sectors = conf->chunk_sectors;
7067 		conf->prev_algo = conf->algorithm;
7068 	}
7069 
7070 	conf->min_nr_stripes = NR_STRIPES;
7071 	if (mddev->reshape_position != MaxSector) {
7072 		int stripes = max_t(int,
7073 			((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4,
7074 			((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4);
7075 		conf->min_nr_stripes = max(NR_STRIPES, stripes);
7076 		if (conf->min_nr_stripes != NR_STRIPES)
7077 			pr_info("md/raid:%s: force stripe size %d for reshape\n",
7078 				mdname(mddev), conf->min_nr_stripes);
7079 	}
7080 	memory = conf->min_nr_stripes * (sizeof(struct stripe_head) +
7081 		 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
7082 	atomic_set(&conf->empty_inactive_list_nr, NR_STRIPE_HASH_LOCKS);
7083 	if (grow_stripes(conf, conf->min_nr_stripes)) {
7084 		pr_warn("md/raid:%s: couldn't allocate %dkB for buffers\n",
7085 			mdname(mddev), memory);
7086 		goto abort;
7087 	} else
7088 		pr_debug("md/raid:%s: allocated %dkB\n", mdname(mddev), memory);
7089 	/*
7090 	 * Losing a stripe head costs more than the time to refill it,
7091 	 * it reduces the queue depth and so can hurt throughput.
7092 	 * So set it rather large, scaled by number of devices.
7093 	 */
7094 	conf->shrinker.seeks = DEFAULT_SEEKS * conf->raid_disks * 4;
7095 	conf->shrinker.scan_objects = raid5_cache_scan;
7096 	conf->shrinker.count_objects = raid5_cache_count;
7097 	conf->shrinker.batch = 128;
7098 	conf->shrinker.flags = 0;
7099 	if (register_shrinker(&conf->shrinker)) {
7100 		pr_warn("md/raid:%s: couldn't register shrinker.\n",
7101 			mdname(mddev));
7102 		goto abort;
7103 	}
7104 
7105 	sprintf(pers_name, "raid%d", mddev->new_level);
7106 	conf->thread = md_register_thread(raid5d, mddev, pers_name);
7107 	if (!conf->thread) {
7108 		pr_warn("md/raid:%s: couldn't allocate thread.\n",
7109 			mdname(mddev));
7110 		goto abort;
7111 	}
7112 
7113 	return conf;
7114 
7115  abort:
7116 	if (conf) {
7117 		free_conf(conf);
7118 		return ERR_PTR(-EIO);
7119 	} else
7120 		return ERR_PTR(-ENOMEM);
7121 }
7122 
7123 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
7124 {
7125 	switch (algo) {
7126 	case ALGORITHM_PARITY_0:
7127 		if (raid_disk < max_degraded)
7128 			return 1;
7129 		break;
7130 	case ALGORITHM_PARITY_N:
7131 		if (raid_disk >= raid_disks - max_degraded)
7132 			return 1;
7133 		break;
7134 	case ALGORITHM_PARITY_0_6:
7135 		if (raid_disk == 0 ||
7136 		    raid_disk == raid_disks - 1)
7137 			return 1;
7138 		break;
7139 	case ALGORITHM_LEFT_ASYMMETRIC_6:
7140 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
7141 	case ALGORITHM_LEFT_SYMMETRIC_6:
7142 	case ALGORITHM_RIGHT_SYMMETRIC_6:
7143 		if (raid_disk == raid_disks - 1)
7144 			return 1;
7145 	}
7146 	return 0;
7147 }
7148 
7149 static int raid5_run(struct mddev *mddev)
7150 {
7151 	struct r5conf *conf;
7152 	int working_disks = 0;
7153 	int dirty_parity_disks = 0;
7154 	struct md_rdev *rdev;
7155 	struct md_rdev *journal_dev = NULL;
7156 	sector_t reshape_offset = 0;
7157 	int i;
7158 	long long min_offset_diff = 0;
7159 	int first = 1;
7160 
7161 	if (mddev_init_writes_pending(mddev) < 0)
7162 		return -ENOMEM;
7163 
7164 	if (mddev->recovery_cp != MaxSector)
7165 		pr_notice("md/raid:%s: not clean -- starting background reconstruction\n",
7166 			  mdname(mddev));
7167 
7168 	rdev_for_each(rdev, mddev) {
7169 		long long diff;
7170 
7171 		if (test_bit(Journal, &rdev->flags)) {
7172 			journal_dev = rdev;
7173 			continue;
7174 		}
7175 		if (rdev->raid_disk < 0)
7176 			continue;
7177 		diff = (rdev->new_data_offset - rdev->data_offset);
7178 		if (first) {
7179 			min_offset_diff = diff;
7180 			first = 0;
7181 		} else if (mddev->reshape_backwards &&
7182 			 diff < min_offset_diff)
7183 			min_offset_diff = diff;
7184 		else if (!mddev->reshape_backwards &&
7185 			 diff > min_offset_diff)
7186 			min_offset_diff = diff;
7187 	}
7188 
7189 	if ((test_bit(MD_HAS_JOURNAL, &mddev->flags) || journal_dev) &&
7190 	    (mddev->bitmap_info.offset || mddev->bitmap_info.file)) {
7191 		pr_notice("md/raid:%s: array cannot have both journal and bitmap\n",
7192 			  mdname(mddev));
7193 		return -EINVAL;
7194 	}
7195 
7196 	if (mddev->reshape_position != MaxSector) {
7197 		/* Check that we can continue the reshape.
7198 		 * Difficulties arise if the stripe we would write to
7199 		 * next is at or after the stripe we would read from next.
7200 		 * For a reshape that changes the number of devices, this
7201 		 * is only possible for a very short time, and mdadm makes
7202 		 * sure that time appears to have past before assembling
7203 		 * the array.  So we fail if that time hasn't passed.
7204 		 * For a reshape that keeps the number of devices the same
7205 		 * mdadm must be monitoring the reshape can keeping the
7206 		 * critical areas read-only and backed up.  It will start
7207 		 * the array in read-only mode, so we check for that.
7208 		 */
7209 		sector_t here_new, here_old;
7210 		int old_disks;
7211 		int max_degraded = (mddev->level == 6 ? 2 : 1);
7212 		int chunk_sectors;
7213 		int new_data_disks;
7214 
7215 		if (journal_dev) {
7216 			pr_warn("md/raid:%s: don't support reshape with journal - aborting.\n",
7217 				mdname(mddev));
7218 			return -EINVAL;
7219 		}
7220 
7221 		if (mddev->new_level != mddev->level) {
7222 			pr_warn("md/raid:%s: unsupported reshape required - aborting.\n",
7223 				mdname(mddev));
7224 			return -EINVAL;
7225 		}
7226 		old_disks = mddev->raid_disks - mddev->delta_disks;
7227 		/* reshape_position must be on a new-stripe boundary, and one
7228 		 * further up in new geometry must map after here in old
7229 		 * geometry.
7230 		 * If the chunk sizes are different, then as we perform reshape
7231 		 * in units of the largest of the two, reshape_position needs
7232 		 * be a multiple of the largest chunk size times new data disks.
7233 		 */
7234 		here_new = mddev->reshape_position;
7235 		chunk_sectors = max(mddev->chunk_sectors, mddev->new_chunk_sectors);
7236 		new_data_disks = mddev->raid_disks - max_degraded;
7237 		if (sector_div(here_new, chunk_sectors * new_data_disks)) {
7238 			pr_warn("md/raid:%s: reshape_position not on a stripe boundary\n",
7239 				mdname(mddev));
7240 			return -EINVAL;
7241 		}
7242 		reshape_offset = here_new * chunk_sectors;
7243 		/* here_new is the stripe we will write to */
7244 		here_old = mddev->reshape_position;
7245 		sector_div(here_old, chunk_sectors * (old_disks-max_degraded));
7246 		/* here_old is the first stripe that we might need to read
7247 		 * from */
7248 		if (mddev->delta_disks == 0) {
7249 			/* We cannot be sure it is safe to start an in-place
7250 			 * reshape.  It is only safe if user-space is monitoring
7251 			 * and taking constant backups.
7252 			 * mdadm always starts a situation like this in
7253 			 * readonly mode so it can take control before
7254 			 * allowing any writes.  So just check for that.
7255 			 */
7256 			if (abs(min_offset_diff) >= mddev->chunk_sectors &&
7257 			    abs(min_offset_diff) >= mddev->new_chunk_sectors)
7258 				/* not really in-place - so OK */;
7259 			else if (mddev->ro == 0) {
7260 				pr_warn("md/raid:%s: in-place reshape must be started in read-only mode - aborting\n",
7261 					mdname(mddev));
7262 				return -EINVAL;
7263 			}
7264 		} else if (mddev->reshape_backwards
7265 		    ? (here_new * chunk_sectors + min_offset_diff <=
7266 		       here_old * chunk_sectors)
7267 		    : (here_new * chunk_sectors >=
7268 		       here_old * chunk_sectors + (-min_offset_diff))) {
7269 			/* Reading from the same stripe as writing to - bad */
7270 			pr_warn("md/raid:%s: reshape_position too early for auto-recovery - aborting.\n",
7271 				mdname(mddev));
7272 			return -EINVAL;
7273 		}
7274 		pr_debug("md/raid:%s: reshape will continue\n", mdname(mddev));
7275 		/* OK, we should be able to continue; */
7276 	} else {
7277 		BUG_ON(mddev->level != mddev->new_level);
7278 		BUG_ON(mddev->layout != mddev->new_layout);
7279 		BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
7280 		BUG_ON(mddev->delta_disks != 0);
7281 	}
7282 
7283 	if (test_bit(MD_HAS_JOURNAL, &mddev->flags) &&
7284 	    test_bit(MD_HAS_PPL, &mddev->flags)) {
7285 		pr_warn("md/raid:%s: using journal device and PPL not allowed - disabling PPL\n",
7286 			mdname(mddev));
7287 		clear_bit(MD_HAS_PPL, &mddev->flags);
7288 		clear_bit(MD_HAS_MULTIPLE_PPLS, &mddev->flags);
7289 	}
7290 
7291 	if (mddev->private == NULL)
7292 		conf = setup_conf(mddev);
7293 	else
7294 		conf = mddev->private;
7295 
7296 	if (IS_ERR(conf))
7297 		return PTR_ERR(conf);
7298 
7299 	if (test_bit(MD_HAS_JOURNAL, &mddev->flags)) {
7300 		if (!journal_dev) {
7301 			pr_warn("md/raid:%s: journal disk is missing, force array readonly\n",
7302 				mdname(mddev));
7303 			mddev->ro = 1;
7304 			set_disk_ro(mddev->gendisk, 1);
7305 		} else if (mddev->recovery_cp == MaxSector)
7306 			set_bit(MD_JOURNAL_CLEAN, &mddev->flags);
7307 	}
7308 
7309 	conf->min_offset_diff = min_offset_diff;
7310 	mddev->thread = conf->thread;
7311 	conf->thread = NULL;
7312 	mddev->private = conf;
7313 
7314 	for (i = 0; i < conf->raid_disks && conf->previous_raid_disks;
7315 	     i++) {
7316 		rdev = conf->disks[i].rdev;
7317 		if (!rdev && conf->disks[i].replacement) {
7318 			/* The replacement is all we have yet */
7319 			rdev = conf->disks[i].replacement;
7320 			conf->disks[i].replacement = NULL;
7321 			clear_bit(Replacement, &rdev->flags);
7322 			conf->disks[i].rdev = rdev;
7323 		}
7324 		if (!rdev)
7325 			continue;
7326 		if (conf->disks[i].replacement &&
7327 		    conf->reshape_progress != MaxSector) {
7328 			/* replacements and reshape simply do not mix. */
7329 			pr_warn("md: cannot handle concurrent replacement and reshape.\n");
7330 			goto abort;
7331 		}
7332 		if (test_bit(In_sync, &rdev->flags)) {
7333 			working_disks++;
7334 			continue;
7335 		}
7336 		/* This disc is not fully in-sync.  However if it
7337 		 * just stored parity (beyond the recovery_offset),
7338 		 * when we don't need to be concerned about the
7339 		 * array being dirty.
7340 		 * When reshape goes 'backwards', we never have
7341 		 * partially completed devices, so we only need
7342 		 * to worry about reshape going forwards.
7343 		 */
7344 		/* Hack because v0.91 doesn't store recovery_offset properly. */
7345 		if (mddev->major_version == 0 &&
7346 		    mddev->minor_version > 90)
7347 			rdev->recovery_offset = reshape_offset;
7348 
7349 		if (rdev->recovery_offset < reshape_offset) {
7350 			/* We need to check old and new layout */
7351 			if (!only_parity(rdev->raid_disk,
7352 					 conf->algorithm,
7353 					 conf->raid_disks,
7354 					 conf->max_degraded))
7355 				continue;
7356 		}
7357 		if (!only_parity(rdev->raid_disk,
7358 				 conf->prev_algo,
7359 				 conf->previous_raid_disks,
7360 				 conf->max_degraded))
7361 			continue;
7362 		dirty_parity_disks++;
7363 	}
7364 
7365 	/*
7366 	 * 0 for a fully functional array, 1 or 2 for a degraded array.
7367 	 */
7368 	mddev->degraded = raid5_calc_degraded(conf);
7369 
7370 	if (has_failed(conf)) {
7371 		pr_crit("md/raid:%s: not enough operational devices (%d/%d failed)\n",
7372 			mdname(mddev), mddev->degraded, conf->raid_disks);
7373 		goto abort;
7374 	}
7375 
7376 	/* device size must be a multiple of chunk size */
7377 	mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
7378 	mddev->resync_max_sectors = mddev->dev_sectors;
7379 
7380 	if (mddev->degraded > dirty_parity_disks &&
7381 	    mddev->recovery_cp != MaxSector) {
7382 		if (test_bit(MD_HAS_PPL, &mddev->flags))
7383 			pr_crit("md/raid:%s: starting dirty degraded array with PPL.\n",
7384 				mdname(mddev));
7385 		else if (mddev->ok_start_degraded)
7386 			pr_crit("md/raid:%s: starting dirty degraded array - data corruption possible.\n",
7387 				mdname(mddev));
7388 		else {
7389 			pr_crit("md/raid:%s: cannot start dirty degraded array.\n",
7390 				mdname(mddev));
7391 			goto abort;
7392 		}
7393 	}
7394 
7395 	pr_info("md/raid:%s: raid level %d active with %d out of %d devices, algorithm %d\n",
7396 		mdname(mddev), conf->level,
7397 		mddev->raid_disks-mddev->degraded, mddev->raid_disks,
7398 		mddev->new_layout);
7399 
7400 	print_raid5_conf(conf);
7401 
7402 	if (conf->reshape_progress != MaxSector) {
7403 		conf->reshape_safe = conf->reshape_progress;
7404 		atomic_set(&conf->reshape_stripes, 0);
7405 		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7406 		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7407 		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7408 		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7409 		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7410 							"reshape");
7411 		if (!mddev->sync_thread)
7412 			goto abort;
7413 	}
7414 
7415 	/* Ok, everything is just fine now */
7416 	if (mddev->to_remove == &raid5_attrs_group)
7417 		mddev->to_remove = NULL;
7418 	else if (mddev->kobj.sd &&
7419 	    sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
7420 		pr_warn("raid5: failed to create sysfs attributes for %s\n",
7421 			mdname(mddev));
7422 	md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
7423 
7424 	if (mddev->queue) {
7425 		int chunk_size;
7426 		/* read-ahead size must cover two whole stripes, which
7427 		 * is 2 * (datadisks) * chunksize where 'n' is the
7428 		 * number of raid devices
7429 		 */
7430 		int data_disks = conf->previous_raid_disks - conf->max_degraded;
7431 		int stripe = data_disks *
7432 			((mddev->chunk_sectors << 9) / PAGE_SIZE);
7433 		if (mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
7434 			mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
7435 
7436 		chunk_size = mddev->chunk_sectors << 9;
7437 		blk_queue_io_min(mddev->queue, chunk_size);
7438 		blk_queue_io_opt(mddev->queue, chunk_size *
7439 				 (conf->raid_disks - conf->max_degraded));
7440 		mddev->queue->limits.raid_partial_stripes_expensive = 1;
7441 		/*
7442 		 * We can only discard a whole stripe. It doesn't make sense to
7443 		 * discard data disk but write parity disk
7444 		 */
7445 		stripe = stripe * PAGE_SIZE;
7446 		/* Round up to power of 2, as discard handling
7447 		 * currently assumes that */
7448 		while ((stripe-1) & stripe)
7449 			stripe = (stripe | (stripe-1)) + 1;
7450 		mddev->queue->limits.discard_alignment = stripe;
7451 		mddev->queue->limits.discard_granularity = stripe;
7452 
7453 		blk_queue_max_write_same_sectors(mddev->queue, 0);
7454 		blk_queue_max_write_zeroes_sectors(mddev->queue, 0);
7455 
7456 		rdev_for_each(rdev, mddev) {
7457 			disk_stack_limits(mddev->gendisk, rdev->bdev,
7458 					  rdev->data_offset << 9);
7459 			disk_stack_limits(mddev->gendisk, rdev->bdev,
7460 					  rdev->new_data_offset << 9);
7461 		}
7462 
7463 		/*
7464 		 * zeroing is required, otherwise data
7465 		 * could be lost. Consider a scenario: discard a stripe
7466 		 * (the stripe could be inconsistent if
7467 		 * discard_zeroes_data is 0); write one disk of the
7468 		 * stripe (the stripe could be inconsistent again
7469 		 * depending on which disks are used to calculate
7470 		 * parity); the disk is broken; The stripe data of this
7471 		 * disk is lost.
7472 		 *
7473 		 * We only allow DISCARD if the sysadmin has confirmed that
7474 		 * only safe devices are in use by setting a module parameter.
7475 		 * A better idea might be to turn DISCARD into WRITE_ZEROES
7476 		 * requests, as that is required to be safe.
7477 		 */
7478 		if (devices_handle_discard_safely &&
7479 		    mddev->queue->limits.max_discard_sectors >= (stripe >> 9) &&
7480 		    mddev->queue->limits.discard_granularity >= stripe)
7481 			blk_queue_flag_set(QUEUE_FLAG_DISCARD,
7482 						mddev->queue);
7483 		else
7484 			blk_queue_flag_clear(QUEUE_FLAG_DISCARD,
7485 						mddev->queue);
7486 
7487 		blk_queue_max_hw_sectors(mddev->queue, UINT_MAX);
7488 	}
7489 
7490 	if (log_init(conf, journal_dev, raid5_has_ppl(conf)))
7491 		goto abort;
7492 
7493 	return 0;
7494 abort:
7495 	md_unregister_thread(&mddev->thread);
7496 	print_raid5_conf(conf);
7497 	free_conf(conf);
7498 	mddev->private = NULL;
7499 	pr_warn("md/raid:%s: failed to run raid set.\n", mdname(mddev));
7500 	return -EIO;
7501 }
7502 
7503 static void raid5_free(struct mddev *mddev, void *priv)
7504 {
7505 	struct r5conf *conf = priv;
7506 
7507 	free_conf(conf);
7508 	mddev->to_remove = &raid5_attrs_group;
7509 }
7510 
7511 static void raid5_status(struct seq_file *seq, struct mddev *mddev)
7512 {
7513 	struct r5conf *conf = mddev->private;
7514 	int i;
7515 
7516 	seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
7517 		conf->chunk_sectors / 2, mddev->layout);
7518 	seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
7519 	rcu_read_lock();
7520 	for (i = 0; i < conf->raid_disks; i++) {
7521 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
7522 		seq_printf (seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
7523 	}
7524 	rcu_read_unlock();
7525 	seq_printf (seq, "]");
7526 }
7527 
7528 static void print_raid5_conf (struct r5conf *conf)
7529 {
7530 	int i;
7531 	struct disk_info *tmp;
7532 
7533 	pr_debug("RAID conf printout:\n");
7534 	if (!conf) {
7535 		pr_debug("(conf==NULL)\n");
7536 		return;
7537 	}
7538 	pr_debug(" --- level:%d rd:%d wd:%d\n", conf->level,
7539 	       conf->raid_disks,
7540 	       conf->raid_disks - conf->mddev->degraded);
7541 
7542 	for (i = 0; i < conf->raid_disks; i++) {
7543 		char b[BDEVNAME_SIZE];
7544 		tmp = conf->disks + i;
7545 		if (tmp->rdev)
7546 			pr_debug(" disk %d, o:%d, dev:%s\n",
7547 			       i, !test_bit(Faulty, &tmp->rdev->flags),
7548 			       bdevname(tmp->rdev->bdev, b));
7549 	}
7550 }
7551 
7552 static int raid5_spare_active(struct mddev *mddev)
7553 {
7554 	int i;
7555 	struct r5conf *conf = mddev->private;
7556 	struct disk_info *tmp;
7557 	int count = 0;
7558 	unsigned long flags;
7559 
7560 	for (i = 0; i < conf->raid_disks; i++) {
7561 		tmp = conf->disks + i;
7562 		if (tmp->replacement
7563 		    && tmp->replacement->recovery_offset == MaxSector
7564 		    && !test_bit(Faulty, &tmp->replacement->flags)
7565 		    && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
7566 			/* Replacement has just become active. */
7567 			if (!tmp->rdev
7568 			    || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
7569 				count++;
7570 			if (tmp->rdev) {
7571 				/* Replaced device not technically faulty,
7572 				 * but we need to be sure it gets removed
7573 				 * and never re-added.
7574 				 */
7575 				set_bit(Faulty, &tmp->rdev->flags);
7576 				sysfs_notify_dirent_safe(
7577 					tmp->rdev->sysfs_state);
7578 			}
7579 			sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
7580 		} else if (tmp->rdev
7581 		    && tmp->rdev->recovery_offset == MaxSector
7582 		    && !test_bit(Faulty, &tmp->rdev->flags)
7583 		    && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
7584 			count++;
7585 			sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
7586 		}
7587 	}
7588 	spin_lock_irqsave(&conf->device_lock, flags);
7589 	mddev->degraded = raid5_calc_degraded(conf);
7590 	spin_unlock_irqrestore(&conf->device_lock, flags);
7591 	print_raid5_conf(conf);
7592 	return count;
7593 }
7594 
7595 static int raid5_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
7596 {
7597 	struct r5conf *conf = mddev->private;
7598 	int err = 0;
7599 	int number = rdev->raid_disk;
7600 	struct md_rdev **rdevp;
7601 	struct disk_info *p = conf->disks + number;
7602 
7603 	print_raid5_conf(conf);
7604 	if (test_bit(Journal, &rdev->flags) && conf->log) {
7605 		/*
7606 		 * we can't wait pending write here, as this is called in
7607 		 * raid5d, wait will deadlock.
7608 		 * neilb: there is no locking about new writes here,
7609 		 * so this cannot be safe.
7610 		 */
7611 		if (atomic_read(&conf->active_stripes) ||
7612 		    atomic_read(&conf->r5c_cached_full_stripes) ||
7613 		    atomic_read(&conf->r5c_cached_partial_stripes)) {
7614 			return -EBUSY;
7615 		}
7616 		log_exit(conf);
7617 		return 0;
7618 	}
7619 	if (rdev == p->rdev)
7620 		rdevp = &p->rdev;
7621 	else if (rdev == p->replacement)
7622 		rdevp = &p->replacement;
7623 	else
7624 		return 0;
7625 
7626 	if (number >= conf->raid_disks &&
7627 	    conf->reshape_progress == MaxSector)
7628 		clear_bit(In_sync, &rdev->flags);
7629 
7630 	if (test_bit(In_sync, &rdev->flags) ||
7631 	    atomic_read(&rdev->nr_pending)) {
7632 		err = -EBUSY;
7633 		goto abort;
7634 	}
7635 	/* Only remove non-faulty devices if recovery
7636 	 * isn't possible.
7637 	 */
7638 	if (!test_bit(Faulty, &rdev->flags) &&
7639 	    mddev->recovery_disabled != conf->recovery_disabled &&
7640 	    !has_failed(conf) &&
7641 	    (!p->replacement || p->replacement == rdev) &&
7642 	    number < conf->raid_disks) {
7643 		err = -EBUSY;
7644 		goto abort;
7645 	}
7646 	*rdevp = NULL;
7647 	if (!test_bit(RemoveSynchronized, &rdev->flags)) {
7648 		synchronize_rcu();
7649 		if (atomic_read(&rdev->nr_pending)) {
7650 			/* lost the race, try later */
7651 			err = -EBUSY;
7652 			*rdevp = rdev;
7653 		}
7654 	}
7655 	if (!err) {
7656 		err = log_modify(conf, rdev, false);
7657 		if (err)
7658 			goto abort;
7659 	}
7660 	if (p->replacement) {
7661 		/* We must have just cleared 'rdev' */
7662 		p->rdev = p->replacement;
7663 		clear_bit(Replacement, &p->replacement->flags);
7664 		smp_mb(); /* Make sure other CPUs may see both as identical
7665 			   * but will never see neither - if they are careful
7666 			   */
7667 		p->replacement = NULL;
7668 
7669 		if (!err)
7670 			err = log_modify(conf, p->rdev, true);
7671 	}
7672 
7673 	clear_bit(WantReplacement, &rdev->flags);
7674 abort:
7675 
7676 	print_raid5_conf(conf);
7677 	return err;
7678 }
7679 
7680 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
7681 {
7682 	struct r5conf *conf = mddev->private;
7683 	int err = -EEXIST;
7684 	int disk;
7685 	struct disk_info *p;
7686 	int first = 0;
7687 	int last = conf->raid_disks - 1;
7688 
7689 	if (test_bit(Journal, &rdev->flags)) {
7690 		if (conf->log)
7691 			return -EBUSY;
7692 
7693 		rdev->raid_disk = 0;
7694 		/*
7695 		 * The array is in readonly mode if journal is missing, so no
7696 		 * write requests running. We should be safe
7697 		 */
7698 		log_init(conf, rdev, false);
7699 		return 0;
7700 	}
7701 	if (mddev->recovery_disabled == conf->recovery_disabled)
7702 		return -EBUSY;
7703 
7704 	if (rdev->saved_raid_disk < 0 && has_failed(conf))
7705 		/* no point adding a device */
7706 		return -EINVAL;
7707 
7708 	if (rdev->raid_disk >= 0)
7709 		first = last = rdev->raid_disk;
7710 
7711 	/*
7712 	 * find the disk ... but prefer rdev->saved_raid_disk
7713 	 * if possible.
7714 	 */
7715 	if (rdev->saved_raid_disk >= 0 &&
7716 	    rdev->saved_raid_disk >= first &&
7717 	    conf->disks[rdev->saved_raid_disk].rdev == NULL)
7718 		first = rdev->saved_raid_disk;
7719 
7720 	for (disk = first; disk <= last; disk++) {
7721 		p = conf->disks + disk;
7722 		if (p->rdev == NULL) {
7723 			clear_bit(In_sync, &rdev->flags);
7724 			rdev->raid_disk = disk;
7725 			if (rdev->saved_raid_disk != disk)
7726 				conf->fullsync = 1;
7727 			rcu_assign_pointer(p->rdev, rdev);
7728 
7729 			err = log_modify(conf, rdev, true);
7730 
7731 			goto out;
7732 		}
7733 	}
7734 	for (disk = first; disk <= last; disk++) {
7735 		p = conf->disks + disk;
7736 		if (test_bit(WantReplacement, &p->rdev->flags) &&
7737 		    p->replacement == NULL) {
7738 			clear_bit(In_sync, &rdev->flags);
7739 			set_bit(Replacement, &rdev->flags);
7740 			rdev->raid_disk = disk;
7741 			err = 0;
7742 			conf->fullsync = 1;
7743 			rcu_assign_pointer(p->replacement, rdev);
7744 			break;
7745 		}
7746 	}
7747 out:
7748 	print_raid5_conf(conf);
7749 	return err;
7750 }
7751 
7752 static int raid5_resize(struct mddev *mddev, sector_t sectors)
7753 {
7754 	/* no resync is happening, and there is enough space
7755 	 * on all devices, so we can resize.
7756 	 * We need to make sure resync covers any new space.
7757 	 * If the array is shrinking we should possibly wait until
7758 	 * any io in the removed space completes, but it hardly seems
7759 	 * worth it.
7760 	 */
7761 	sector_t newsize;
7762 	struct r5conf *conf = mddev->private;
7763 
7764 	if (raid5_has_log(conf) || raid5_has_ppl(conf))
7765 		return -EINVAL;
7766 	sectors &= ~((sector_t)conf->chunk_sectors - 1);
7767 	newsize = raid5_size(mddev, sectors, mddev->raid_disks);
7768 	if (mddev->external_size &&
7769 	    mddev->array_sectors > newsize)
7770 		return -EINVAL;
7771 	if (mddev->bitmap) {
7772 		int ret = md_bitmap_resize(mddev->bitmap, sectors, 0, 0);
7773 		if (ret)
7774 			return ret;
7775 	}
7776 	md_set_array_sectors(mddev, newsize);
7777 	if (sectors > mddev->dev_sectors &&
7778 	    mddev->recovery_cp > mddev->dev_sectors) {
7779 		mddev->recovery_cp = mddev->dev_sectors;
7780 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
7781 	}
7782 	mddev->dev_sectors = sectors;
7783 	mddev->resync_max_sectors = sectors;
7784 	return 0;
7785 }
7786 
7787 static int check_stripe_cache(struct mddev *mddev)
7788 {
7789 	/* Can only proceed if there are plenty of stripe_heads.
7790 	 * We need a minimum of one full stripe,, and for sensible progress
7791 	 * it is best to have about 4 times that.
7792 	 * If we require 4 times, then the default 256 4K stripe_heads will
7793 	 * allow for chunk sizes up to 256K, which is probably OK.
7794 	 * If the chunk size is greater, user-space should request more
7795 	 * stripe_heads first.
7796 	 */
7797 	struct r5conf *conf = mddev->private;
7798 	if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
7799 	    > conf->min_nr_stripes ||
7800 	    ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
7801 	    > conf->min_nr_stripes) {
7802 		pr_warn("md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
7803 			mdname(mddev),
7804 			((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
7805 			 / STRIPE_SIZE)*4);
7806 		return 0;
7807 	}
7808 	return 1;
7809 }
7810 
7811 static int check_reshape(struct mddev *mddev)
7812 {
7813 	struct r5conf *conf = mddev->private;
7814 
7815 	if (raid5_has_log(conf) || raid5_has_ppl(conf))
7816 		return -EINVAL;
7817 	if (mddev->delta_disks == 0 &&
7818 	    mddev->new_layout == mddev->layout &&
7819 	    mddev->new_chunk_sectors == mddev->chunk_sectors)
7820 		return 0; /* nothing to do */
7821 	if (has_failed(conf))
7822 		return -EINVAL;
7823 	if (mddev->delta_disks < 0 && mddev->reshape_position == MaxSector) {
7824 		/* We might be able to shrink, but the devices must
7825 		 * be made bigger first.
7826 		 * For raid6, 4 is the minimum size.
7827 		 * Otherwise 2 is the minimum
7828 		 */
7829 		int min = 2;
7830 		if (mddev->level == 6)
7831 			min = 4;
7832 		if (mddev->raid_disks + mddev->delta_disks < min)
7833 			return -EINVAL;
7834 	}
7835 
7836 	if (!check_stripe_cache(mddev))
7837 		return -ENOSPC;
7838 
7839 	if (mddev->new_chunk_sectors > mddev->chunk_sectors ||
7840 	    mddev->delta_disks > 0)
7841 		if (resize_chunks(conf,
7842 				  conf->previous_raid_disks
7843 				  + max(0, mddev->delta_disks),
7844 				  max(mddev->new_chunk_sectors,
7845 				      mddev->chunk_sectors)
7846 			    ) < 0)
7847 			return -ENOMEM;
7848 
7849 	if (conf->previous_raid_disks + mddev->delta_disks <= conf->pool_size)
7850 		return 0; /* never bother to shrink */
7851 	return resize_stripes(conf, (conf->previous_raid_disks
7852 				     + mddev->delta_disks));
7853 }
7854 
7855 static int raid5_start_reshape(struct mddev *mddev)
7856 {
7857 	struct r5conf *conf = mddev->private;
7858 	struct md_rdev *rdev;
7859 	int spares = 0;
7860 	unsigned long flags;
7861 
7862 	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
7863 		return -EBUSY;
7864 
7865 	if (!check_stripe_cache(mddev))
7866 		return -ENOSPC;
7867 
7868 	if (has_failed(conf))
7869 		return -EINVAL;
7870 
7871 	rdev_for_each(rdev, mddev) {
7872 		if (!test_bit(In_sync, &rdev->flags)
7873 		    && !test_bit(Faulty, &rdev->flags))
7874 			spares++;
7875 	}
7876 
7877 	if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
7878 		/* Not enough devices even to make a degraded array
7879 		 * of that size
7880 		 */
7881 		return -EINVAL;
7882 
7883 	/* Refuse to reduce size of the array.  Any reductions in
7884 	 * array size must be through explicit setting of array_size
7885 	 * attribute.
7886 	 */
7887 	if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
7888 	    < mddev->array_sectors) {
7889 		pr_warn("md/raid:%s: array size must be reduced before number of disks\n",
7890 			mdname(mddev));
7891 		return -EINVAL;
7892 	}
7893 
7894 	atomic_set(&conf->reshape_stripes, 0);
7895 	spin_lock_irq(&conf->device_lock);
7896 	write_seqcount_begin(&conf->gen_lock);
7897 	conf->previous_raid_disks = conf->raid_disks;
7898 	conf->raid_disks += mddev->delta_disks;
7899 	conf->prev_chunk_sectors = conf->chunk_sectors;
7900 	conf->chunk_sectors = mddev->new_chunk_sectors;
7901 	conf->prev_algo = conf->algorithm;
7902 	conf->algorithm = mddev->new_layout;
7903 	conf->generation++;
7904 	/* Code that selects data_offset needs to see the generation update
7905 	 * if reshape_progress has been set - so a memory barrier needed.
7906 	 */
7907 	smp_mb();
7908 	if (mddev->reshape_backwards)
7909 		conf->reshape_progress = raid5_size(mddev, 0, 0);
7910 	else
7911 		conf->reshape_progress = 0;
7912 	conf->reshape_safe = conf->reshape_progress;
7913 	write_seqcount_end(&conf->gen_lock);
7914 	spin_unlock_irq(&conf->device_lock);
7915 
7916 	/* Now make sure any requests that proceeded on the assumption
7917 	 * the reshape wasn't running - like Discard or Read - have
7918 	 * completed.
7919 	 */
7920 	mddev_suspend(mddev);
7921 	mddev_resume(mddev);
7922 
7923 	/* Add some new drives, as many as will fit.
7924 	 * We know there are enough to make the newly sized array work.
7925 	 * Don't add devices if we are reducing the number of
7926 	 * devices in the array.  This is because it is not possible
7927 	 * to correctly record the "partially reconstructed" state of
7928 	 * such devices during the reshape and confusion could result.
7929 	 */
7930 	if (mddev->delta_disks >= 0) {
7931 		rdev_for_each(rdev, mddev)
7932 			if (rdev->raid_disk < 0 &&
7933 			    !test_bit(Faulty, &rdev->flags)) {
7934 				if (raid5_add_disk(mddev, rdev) == 0) {
7935 					if (rdev->raid_disk
7936 					    >= conf->previous_raid_disks)
7937 						set_bit(In_sync, &rdev->flags);
7938 					else
7939 						rdev->recovery_offset = 0;
7940 
7941 					if (sysfs_link_rdev(mddev, rdev))
7942 						/* Failure here is OK */;
7943 				}
7944 			} else if (rdev->raid_disk >= conf->previous_raid_disks
7945 				   && !test_bit(Faulty, &rdev->flags)) {
7946 				/* This is a spare that was manually added */
7947 				set_bit(In_sync, &rdev->flags);
7948 			}
7949 
7950 		/* When a reshape changes the number of devices,
7951 		 * ->degraded is measured against the larger of the
7952 		 * pre and post number of devices.
7953 		 */
7954 		spin_lock_irqsave(&conf->device_lock, flags);
7955 		mddev->degraded = raid5_calc_degraded(conf);
7956 		spin_unlock_irqrestore(&conf->device_lock, flags);
7957 	}
7958 	mddev->raid_disks = conf->raid_disks;
7959 	mddev->reshape_position = conf->reshape_progress;
7960 	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
7961 
7962 	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
7963 	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
7964 	clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
7965 	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
7966 	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
7967 	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
7968 						"reshape");
7969 	if (!mddev->sync_thread) {
7970 		mddev->recovery = 0;
7971 		spin_lock_irq(&conf->device_lock);
7972 		write_seqcount_begin(&conf->gen_lock);
7973 		mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
7974 		mddev->new_chunk_sectors =
7975 			conf->chunk_sectors = conf->prev_chunk_sectors;
7976 		mddev->new_layout = conf->algorithm = conf->prev_algo;
7977 		rdev_for_each(rdev, mddev)
7978 			rdev->new_data_offset = rdev->data_offset;
7979 		smp_wmb();
7980 		conf->generation --;
7981 		conf->reshape_progress = MaxSector;
7982 		mddev->reshape_position = MaxSector;
7983 		write_seqcount_end(&conf->gen_lock);
7984 		spin_unlock_irq(&conf->device_lock);
7985 		return -EAGAIN;
7986 	}
7987 	conf->reshape_checkpoint = jiffies;
7988 	md_wakeup_thread(mddev->sync_thread);
7989 	md_new_event(mddev);
7990 	return 0;
7991 }
7992 
7993 /* This is called from the reshape thread and should make any
7994  * changes needed in 'conf'
7995  */
7996 static void end_reshape(struct r5conf *conf)
7997 {
7998 
7999 	if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
8000 		struct md_rdev *rdev;
8001 
8002 		spin_lock_irq(&conf->device_lock);
8003 		conf->previous_raid_disks = conf->raid_disks;
8004 		md_finish_reshape(conf->mddev);
8005 		smp_wmb();
8006 		conf->reshape_progress = MaxSector;
8007 		conf->mddev->reshape_position = MaxSector;
8008 		rdev_for_each(rdev, conf->mddev)
8009 			if (rdev->raid_disk >= 0 &&
8010 			    !test_bit(Journal, &rdev->flags) &&
8011 			    !test_bit(In_sync, &rdev->flags))
8012 				rdev->recovery_offset = MaxSector;
8013 		spin_unlock_irq(&conf->device_lock);
8014 		wake_up(&conf->wait_for_overlap);
8015 
8016 		/* read-ahead size must cover two whole stripes, which is
8017 		 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
8018 		 */
8019 		if (conf->mddev->queue) {
8020 			int data_disks = conf->raid_disks - conf->max_degraded;
8021 			int stripe = data_disks * ((conf->chunk_sectors << 9)
8022 						   / PAGE_SIZE);
8023 			if (conf->mddev->queue->backing_dev_info->ra_pages < 2 * stripe)
8024 				conf->mddev->queue->backing_dev_info->ra_pages = 2 * stripe;
8025 		}
8026 	}
8027 }
8028 
8029 /* This is called from the raid5d thread with mddev_lock held.
8030  * It makes config changes to the device.
8031  */
8032 static void raid5_finish_reshape(struct mddev *mddev)
8033 {
8034 	struct r5conf *conf = mddev->private;
8035 
8036 	if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
8037 
8038 		if (mddev->delta_disks <= 0) {
8039 			int d;
8040 			spin_lock_irq(&conf->device_lock);
8041 			mddev->degraded = raid5_calc_degraded(conf);
8042 			spin_unlock_irq(&conf->device_lock);
8043 			for (d = conf->raid_disks ;
8044 			     d < conf->raid_disks - mddev->delta_disks;
8045 			     d++) {
8046 				struct md_rdev *rdev = conf->disks[d].rdev;
8047 				if (rdev)
8048 					clear_bit(In_sync, &rdev->flags);
8049 				rdev = conf->disks[d].replacement;
8050 				if (rdev)
8051 					clear_bit(In_sync, &rdev->flags);
8052 			}
8053 		}
8054 		mddev->layout = conf->algorithm;
8055 		mddev->chunk_sectors = conf->chunk_sectors;
8056 		mddev->reshape_position = MaxSector;
8057 		mddev->delta_disks = 0;
8058 		mddev->reshape_backwards = 0;
8059 	}
8060 }
8061 
8062 static void raid5_quiesce(struct mddev *mddev, int quiesce)
8063 {
8064 	struct r5conf *conf = mddev->private;
8065 
8066 	if (quiesce) {
8067 		/* stop all writes */
8068 		lock_all_device_hash_locks_irq(conf);
8069 		/* '2' tells resync/reshape to pause so that all
8070 		 * active stripes can drain
8071 		 */
8072 		r5c_flush_cache(conf, INT_MAX);
8073 		conf->quiesce = 2;
8074 		wait_event_cmd(conf->wait_for_quiescent,
8075 				    atomic_read(&conf->active_stripes) == 0 &&
8076 				    atomic_read(&conf->active_aligned_reads) == 0,
8077 				    unlock_all_device_hash_locks_irq(conf),
8078 				    lock_all_device_hash_locks_irq(conf));
8079 		conf->quiesce = 1;
8080 		unlock_all_device_hash_locks_irq(conf);
8081 		/* allow reshape to continue */
8082 		wake_up(&conf->wait_for_overlap);
8083 	} else {
8084 		/* re-enable writes */
8085 		lock_all_device_hash_locks_irq(conf);
8086 		conf->quiesce = 0;
8087 		wake_up(&conf->wait_for_quiescent);
8088 		wake_up(&conf->wait_for_overlap);
8089 		unlock_all_device_hash_locks_irq(conf);
8090 	}
8091 	log_quiesce(conf, quiesce);
8092 }
8093 
8094 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
8095 {
8096 	struct r0conf *raid0_conf = mddev->private;
8097 	sector_t sectors;
8098 
8099 	/* for raid0 takeover only one zone is supported */
8100 	if (raid0_conf->nr_strip_zones > 1) {
8101 		pr_warn("md/raid:%s: cannot takeover raid0 with more than one zone.\n",
8102 			mdname(mddev));
8103 		return ERR_PTR(-EINVAL);
8104 	}
8105 
8106 	sectors = raid0_conf->strip_zone[0].zone_end;
8107 	sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
8108 	mddev->dev_sectors = sectors;
8109 	mddev->new_level = level;
8110 	mddev->new_layout = ALGORITHM_PARITY_N;
8111 	mddev->new_chunk_sectors = mddev->chunk_sectors;
8112 	mddev->raid_disks += 1;
8113 	mddev->delta_disks = 1;
8114 	/* make sure it will be not marked as dirty */
8115 	mddev->recovery_cp = MaxSector;
8116 
8117 	return setup_conf(mddev);
8118 }
8119 
8120 static void *raid5_takeover_raid1(struct mddev *mddev)
8121 {
8122 	int chunksect;
8123 	void *ret;
8124 
8125 	if (mddev->raid_disks != 2 ||
8126 	    mddev->degraded > 1)
8127 		return ERR_PTR(-EINVAL);
8128 
8129 	/* Should check if there are write-behind devices? */
8130 
8131 	chunksect = 64*2; /* 64K by default */
8132 
8133 	/* The array must be an exact multiple of chunksize */
8134 	while (chunksect && (mddev->array_sectors & (chunksect-1)))
8135 		chunksect >>= 1;
8136 
8137 	if ((chunksect<<9) < STRIPE_SIZE)
8138 		/* array size does not allow a suitable chunk size */
8139 		return ERR_PTR(-EINVAL);
8140 
8141 	mddev->new_level = 5;
8142 	mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
8143 	mddev->new_chunk_sectors = chunksect;
8144 
8145 	ret = setup_conf(mddev);
8146 	if (!IS_ERR(ret))
8147 		mddev_clear_unsupported_flags(mddev,
8148 			UNSUPPORTED_MDDEV_FLAGS);
8149 	return ret;
8150 }
8151 
8152 static void *raid5_takeover_raid6(struct mddev *mddev)
8153 {
8154 	int new_layout;
8155 
8156 	switch (mddev->layout) {
8157 	case ALGORITHM_LEFT_ASYMMETRIC_6:
8158 		new_layout = ALGORITHM_LEFT_ASYMMETRIC;
8159 		break;
8160 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
8161 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
8162 		break;
8163 	case ALGORITHM_LEFT_SYMMETRIC_6:
8164 		new_layout = ALGORITHM_LEFT_SYMMETRIC;
8165 		break;
8166 	case ALGORITHM_RIGHT_SYMMETRIC_6:
8167 		new_layout = ALGORITHM_RIGHT_SYMMETRIC;
8168 		break;
8169 	case ALGORITHM_PARITY_0_6:
8170 		new_layout = ALGORITHM_PARITY_0;
8171 		break;
8172 	case ALGORITHM_PARITY_N:
8173 		new_layout = ALGORITHM_PARITY_N;
8174 		break;
8175 	default:
8176 		return ERR_PTR(-EINVAL);
8177 	}
8178 	mddev->new_level = 5;
8179 	mddev->new_layout = new_layout;
8180 	mddev->delta_disks = -1;
8181 	mddev->raid_disks -= 1;
8182 	return setup_conf(mddev);
8183 }
8184 
8185 static int raid5_check_reshape(struct mddev *mddev)
8186 {
8187 	/* For a 2-drive array, the layout and chunk size can be changed
8188 	 * immediately as not restriping is needed.
8189 	 * For larger arrays we record the new value - after validation
8190 	 * to be used by a reshape pass.
8191 	 */
8192 	struct r5conf *conf = mddev->private;
8193 	int new_chunk = mddev->new_chunk_sectors;
8194 
8195 	if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
8196 		return -EINVAL;
8197 	if (new_chunk > 0) {
8198 		if (!is_power_of_2(new_chunk))
8199 			return -EINVAL;
8200 		if (new_chunk < (PAGE_SIZE>>9))
8201 			return -EINVAL;
8202 		if (mddev->array_sectors & (new_chunk-1))
8203 			/* not factor of array size */
8204 			return -EINVAL;
8205 	}
8206 
8207 	/* They look valid */
8208 
8209 	if (mddev->raid_disks == 2) {
8210 		/* can make the change immediately */
8211 		if (mddev->new_layout >= 0) {
8212 			conf->algorithm = mddev->new_layout;
8213 			mddev->layout = mddev->new_layout;
8214 		}
8215 		if (new_chunk > 0) {
8216 			conf->chunk_sectors = new_chunk ;
8217 			mddev->chunk_sectors = new_chunk;
8218 		}
8219 		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
8220 		md_wakeup_thread(mddev->thread);
8221 	}
8222 	return check_reshape(mddev);
8223 }
8224 
8225 static int raid6_check_reshape(struct mddev *mddev)
8226 {
8227 	int new_chunk = mddev->new_chunk_sectors;
8228 
8229 	if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
8230 		return -EINVAL;
8231 	if (new_chunk > 0) {
8232 		if (!is_power_of_2(new_chunk))
8233 			return -EINVAL;
8234 		if (new_chunk < (PAGE_SIZE >> 9))
8235 			return -EINVAL;
8236 		if (mddev->array_sectors & (new_chunk-1))
8237 			/* not factor of array size */
8238 			return -EINVAL;
8239 	}
8240 
8241 	/* They look valid */
8242 	return check_reshape(mddev);
8243 }
8244 
8245 static void *raid5_takeover(struct mddev *mddev)
8246 {
8247 	/* raid5 can take over:
8248 	 *  raid0 - if there is only one strip zone - make it a raid4 layout
8249 	 *  raid1 - if there are two drives.  We need to know the chunk size
8250 	 *  raid4 - trivial - just use a raid4 layout.
8251 	 *  raid6 - Providing it is a *_6 layout
8252 	 */
8253 	if (mddev->level == 0)
8254 		return raid45_takeover_raid0(mddev, 5);
8255 	if (mddev->level == 1)
8256 		return raid5_takeover_raid1(mddev);
8257 	if (mddev->level == 4) {
8258 		mddev->new_layout = ALGORITHM_PARITY_N;
8259 		mddev->new_level = 5;
8260 		return setup_conf(mddev);
8261 	}
8262 	if (mddev->level == 6)
8263 		return raid5_takeover_raid6(mddev);
8264 
8265 	return ERR_PTR(-EINVAL);
8266 }
8267 
8268 static void *raid4_takeover(struct mddev *mddev)
8269 {
8270 	/* raid4 can take over:
8271 	 *  raid0 - if there is only one strip zone
8272 	 *  raid5 - if layout is right
8273 	 */
8274 	if (mddev->level == 0)
8275 		return raid45_takeover_raid0(mddev, 4);
8276 	if (mddev->level == 5 &&
8277 	    mddev->layout == ALGORITHM_PARITY_N) {
8278 		mddev->new_layout = 0;
8279 		mddev->new_level = 4;
8280 		return setup_conf(mddev);
8281 	}
8282 	return ERR_PTR(-EINVAL);
8283 }
8284 
8285 static struct md_personality raid5_personality;
8286 
8287 static void *raid6_takeover(struct mddev *mddev)
8288 {
8289 	/* Currently can only take over a raid5.  We map the
8290 	 * personality to an equivalent raid6 personality
8291 	 * with the Q block at the end.
8292 	 */
8293 	int new_layout;
8294 
8295 	if (mddev->pers != &raid5_personality)
8296 		return ERR_PTR(-EINVAL);
8297 	if (mddev->degraded > 1)
8298 		return ERR_PTR(-EINVAL);
8299 	if (mddev->raid_disks > 253)
8300 		return ERR_PTR(-EINVAL);
8301 	if (mddev->raid_disks < 3)
8302 		return ERR_PTR(-EINVAL);
8303 
8304 	switch (mddev->layout) {
8305 	case ALGORITHM_LEFT_ASYMMETRIC:
8306 		new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
8307 		break;
8308 	case ALGORITHM_RIGHT_ASYMMETRIC:
8309 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
8310 		break;
8311 	case ALGORITHM_LEFT_SYMMETRIC:
8312 		new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
8313 		break;
8314 	case ALGORITHM_RIGHT_SYMMETRIC:
8315 		new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
8316 		break;
8317 	case ALGORITHM_PARITY_0:
8318 		new_layout = ALGORITHM_PARITY_0_6;
8319 		break;
8320 	case ALGORITHM_PARITY_N:
8321 		new_layout = ALGORITHM_PARITY_N;
8322 		break;
8323 	default:
8324 		return ERR_PTR(-EINVAL);
8325 	}
8326 	mddev->new_level = 6;
8327 	mddev->new_layout = new_layout;
8328 	mddev->delta_disks = 1;
8329 	mddev->raid_disks += 1;
8330 	return setup_conf(mddev);
8331 }
8332 
8333 static int raid5_change_consistency_policy(struct mddev *mddev, const char *buf)
8334 {
8335 	struct r5conf *conf;
8336 	int err;
8337 
8338 	err = mddev_lock(mddev);
8339 	if (err)
8340 		return err;
8341 	conf = mddev->private;
8342 	if (!conf) {
8343 		mddev_unlock(mddev);
8344 		return -ENODEV;
8345 	}
8346 
8347 	if (strncmp(buf, "ppl", 3) == 0) {
8348 		/* ppl only works with RAID 5 */
8349 		if (!raid5_has_ppl(conf) && conf->level == 5) {
8350 			err = log_init(conf, NULL, true);
8351 			if (!err) {
8352 				err = resize_stripes(conf, conf->pool_size);
8353 				if (err)
8354 					log_exit(conf);
8355 			}
8356 		} else
8357 			err = -EINVAL;
8358 	} else if (strncmp(buf, "resync", 6) == 0) {
8359 		if (raid5_has_ppl(conf)) {
8360 			mddev_suspend(mddev);
8361 			log_exit(conf);
8362 			mddev_resume(mddev);
8363 			err = resize_stripes(conf, conf->pool_size);
8364 		} else if (test_bit(MD_HAS_JOURNAL, &conf->mddev->flags) &&
8365 			   r5l_log_disk_error(conf)) {
8366 			bool journal_dev_exists = false;
8367 			struct md_rdev *rdev;
8368 
8369 			rdev_for_each(rdev, mddev)
8370 				if (test_bit(Journal, &rdev->flags)) {
8371 					journal_dev_exists = true;
8372 					break;
8373 				}
8374 
8375 			if (!journal_dev_exists) {
8376 				mddev_suspend(mddev);
8377 				clear_bit(MD_HAS_JOURNAL, &mddev->flags);
8378 				mddev_resume(mddev);
8379 			} else  /* need remove journal device first */
8380 				err = -EBUSY;
8381 		} else
8382 			err = -EINVAL;
8383 	} else {
8384 		err = -EINVAL;
8385 	}
8386 
8387 	if (!err)
8388 		md_update_sb(mddev, 1);
8389 
8390 	mddev_unlock(mddev);
8391 
8392 	return err;
8393 }
8394 
8395 static int raid5_start(struct mddev *mddev)
8396 {
8397 	struct r5conf *conf = mddev->private;
8398 
8399 	return r5l_start(conf->log);
8400 }
8401 
8402 static struct md_personality raid6_personality =
8403 {
8404 	.name		= "raid6",
8405 	.level		= 6,
8406 	.owner		= THIS_MODULE,
8407 	.make_request	= raid5_make_request,
8408 	.run		= raid5_run,
8409 	.start		= raid5_start,
8410 	.free		= raid5_free,
8411 	.status		= raid5_status,
8412 	.error_handler	= raid5_error,
8413 	.hot_add_disk	= raid5_add_disk,
8414 	.hot_remove_disk= raid5_remove_disk,
8415 	.spare_active	= raid5_spare_active,
8416 	.sync_request	= raid5_sync_request,
8417 	.resize		= raid5_resize,
8418 	.size		= raid5_size,
8419 	.check_reshape	= raid6_check_reshape,
8420 	.start_reshape  = raid5_start_reshape,
8421 	.finish_reshape = raid5_finish_reshape,
8422 	.quiesce	= raid5_quiesce,
8423 	.takeover	= raid6_takeover,
8424 	.congested	= raid5_congested,
8425 	.change_consistency_policy = raid5_change_consistency_policy,
8426 };
8427 static struct md_personality raid5_personality =
8428 {
8429 	.name		= "raid5",
8430 	.level		= 5,
8431 	.owner		= THIS_MODULE,
8432 	.make_request	= raid5_make_request,
8433 	.run		= raid5_run,
8434 	.start		= raid5_start,
8435 	.free		= raid5_free,
8436 	.status		= raid5_status,
8437 	.error_handler	= raid5_error,
8438 	.hot_add_disk	= raid5_add_disk,
8439 	.hot_remove_disk= raid5_remove_disk,
8440 	.spare_active	= raid5_spare_active,
8441 	.sync_request	= raid5_sync_request,
8442 	.resize		= raid5_resize,
8443 	.size		= raid5_size,
8444 	.check_reshape	= raid5_check_reshape,
8445 	.start_reshape  = raid5_start_reshape,
8446 	.finish_reshape = raid5_finish_reshape,
8447 	.quiesce	= raid5_quiesce,
8448 	.takeover	= raid5_takeover,
8449 	.congested	= raid5_congested,
8450 	.change_consistency_policy = raid5_change_consistency_policy,
8451 };
8452 
8453 static struct md_personality raid4_personality =
8454 {
8455 	.name		= "raid4",
8456 	.level		= 4,
8457 	.owner		= THIS_MODULE,
8458 	.make_request	= raid5_make_request,
8459 	.run		= raid5_run,
8460 	.start		= raid5_start,
8461 	.free		= raid5_free,
8462 	.status		= raid5_status,
8463 	.error_handler	= raid5_error,
8464 	.hot_add_disk	= raid5_add_disk,
8465 	.hot_remove_disk= raid5_remove_disk,
8466 	.spare_active	= raid5_spare_active,
8467 	.sync_request	= raid5_sync_request,
8468 	.resize		= raid5_resize,
8469 	.size		= raid5_size,
8470 	.check_reshape	= raid5_check_reshape,
8471 	.start_reshape  = raid5_start_reshape,
8472 	.finish_reshape = raid5_finish_reshape,
8473 	.quiesce	= raid5_quiesce,
8474 	.takeover	= raid4_takeover,
8475 	.congested	= raid5_congested,
8476 	.change_consistency_policy = raid5_change_consistency_policy,
8477 };
8478 
8479 static int __init raid5_init(void)
8480 {
8481 	int ret;
8482 
8483 	raid5_wq = alloc_workqueue("raid5wq",
8484 		WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_CPU_INTENSIVE|WQ_SYSFS, 0);
8485 	if (!raid5_wq)
8486 		return -ENOMEM;
8487 
8488 	ret = cpuhp_setup_state_multi(CPUHP_MD_RAID5_PREPARE,
8489 				      "md/raid5:prepare",
8490 				      raid456_cpu_up_prepare,
8491 				      raid456_cpu_dead);
8492 	if (ret) {
8493 		destroy_workqueue(raid5_wq);
8494 		return ret;
8495 	}
8496 	register_md_personality(&raid6_personality);
8497 	register_md_personality(&raid5_personality);
8498 	register_md_personality(&raid4_personality);
8499 	return 0;
8500 }
8501 
8502 static void raid5_exit(void)
8503 {
8504 	unregister_md_personality(&raid6_personality);
8505 	unregister_md_personality(&raid5_personality);
8506 	unregister_md_personality(&raid4_personality);
8507 	cpuhp_remove_multi_state(CPUHP_MD_RAID5_PREPARE);
8508 	destroy_workqueue(raid5_wq);
8509 }
8510 
8511 module_init(raid5_init);
8512 module_exit(raid5_exit);
8513 MODULE_LICENSE("GPL");
8514 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
8515 MODULE_ALIAS("md-personality-4"); /* RAID5 */
8516 MODULE_ALIAS("md-raid5");
8517 MODULE_ALIAS("md-raid4");
8518 MODULE_ALIAS("md-level-5");
8519 MODULE_ALIAS("md-level-4");
8520 MODULE_ALIAS("md-personality-8"); /* RAID6 */
8521 MODULE_ALIAS("md-raid6");
8522 MODULE_ALIAS("md-level-6");
8523 
8524 /* This used to be two separate modules, they were: */
8525 MODULE_ALIAS("raid5");
8526 MODULE_ALIAS("raid6");
8527