xref: /linux/drivers/md/raid5.c (revision f2ee442115c9b6219083c019939a9cc0c9abb2f8)
1 /*
2  * raid5.c : Multiple Devices driver for Linux
3  *	   Copyright (C) 1996, 1997 Ingo Molnar, Miguel de Icaza, Gadi Oxman
4  *	   Copyright (C) 1999, 2000 Ingo Molnar
5  *	   Copyright (C) 2002, 2003 H. Peter Anvin
6  *
7  * RAID-4/5/6 management functions.
8  * Thanks to Penguin Computing for making the RAID-6 development possible
9  * by donating a test server!
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20 
21 /*
22  * BITMAP UNPLUGGING:
23  *
24  * The sequencing for updating the bitmap reliably is a little
25  * subtle (and I got it wrong the first time) so it deserves some
26  * explanation.
27  *
28  * We group bitmap updates into batches.  Each batch has a number.
29  * We may write out several batches at once, but that isn't very important.
30  * conf->seq_write is the number of the last batch successfully written.
31  * conf->seq_flush is the number of the last batch that was closed to
32  *    new additions.
33  * When we discover that we will need to write to any block in a stripe
34  * (in add_stripe_bio) we update the in-memory bitmap and record in sh->bm_seq
35  * the number of the batch it will be in. This is seq_flush+1.
36  * When we are ready to do a write, if that batch hasn't been written yet,
37  *   we plug the array and queue the stripe for later.
38  * When an unplug happens, we increment bm_flush, thus closing the current
39  *   batch.
40  * When we notice that bm_flush > bm_write, we write out all pending updates
41  * to the bitmap, and advance bm_write to where bm_flush was.
42  * This may occasionally write a bit out twice, but is sure never to
43  * miss any bits.
44  */
45 
46 #include <linux/blkdev.h>
47 #include <linux/kthread.h>
48 #include <linux/raid/pq.h>
49 #include <linux/async_tx.h>
50 #include <linux/module.h>
51 #include <linux/async.h>
52 #include <linux/seq_file.h>
53 #include <linux/cpu.h>
54 #include <linux/slab.h>
55 #include <linux/ratelimit.h>
56 #include "md.h"
57 #include "raid5.h"
58 #include "raid0.h"
59 #include "bitmap.h"
60 
61 /*
62  * Stripe cache
63  */
64 
65 #define NR_STRIPES		256
66 #define STRIPE_SIZE		PAGE_SIZE
67 #define STRIPE_SHIFT		(PAGE_SHIFT - 9)
68 #define STRIPE_SECTORS		(STRIPE_SIZE>>9)
69 #define	IO_THRESHOLD		1
70 #define BYPASS_THRESHOLD	1
71 #define NR_HASH			(PAGE_SIZE / sizeof(struct hlist_head))
72 #define HASH_MASK		(NR_HASH - 1)
73 
74 static inline struct hlist_head *stripe_hash(struct r5conf *conf, sector_t sect)
75 {
76 	int hash = (sect >> STRIPE_SHIFT) & HASH_MASK;
77 	return &conf->stripe_hashtbl[hash];
78 }
79 
80 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
81  * order without overlap.  There may be several bio's per stripe+device, and
82  * a bio could span several devices.
83  * When walking this list for a particular stripe+device, we must never proceed
84  * beyond a bio that extends past this device, as the next bio might no longer
85  * be valid.
86  * This function is used to determine the 'next' bio in the list, given the sector
87  * of the current stripe+device
88  */
89 static inline struct bio *r5_next_bio(struct bio *bio, sector_t sector)
90 {
91 	int sectors = bio->bi_size >> 9;
92 	if (bio->bi_sector + sectors < sector + STRIPE_SECTORS)
93 		return bio->bi_next;
94 	else
95 		return NULL;
96 }
97 
98 /*
99  * We maintain a biased count of active stripes in the bottom 16 bits of
100  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
101  */
102 static inline int raid5_bi_phys_segments(struct bio *bio)
103 {
104 	return bio->bi_phys_segments & 0xffff;
105 }
106 
107 static inline int raid5_bi_hw_segments(struct bio *bio)
108 {
109 	return (bio->bi_phys_segments >> 16) & 0xffff;
110 }
111 
112 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
113 {
114 	--bio->bi_phys_segments;
115 	return raid5_bi_phys_segments(bio);
116 }
117 
118 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
119 {
120 	unsigned short val = raid5_bi_hw_segments(bio);
121 
122 	--val;
123 	bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
124 	return val;
125 }
126 
127 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
128 {
129 	bio->bi_phys_segments = raid5_bi_phys_segments(bio) | (cnt << 16);
130 }
131 
132 /* Find first data disk in a raid6 stripe */
133 static inline int raid6_d0(struct stripe_head *sh)
134 {
135 	if (sh->ddf_layout)
136 		/* ddf always start from first device */
137 		return 0;
138 	/* md starts just after Q block */
139 	if (sh->qd_idx == sh->disks - 1)
140 		return 0;
141 	else
142 		return sh->qd_idx + 1;
143 }
144 static inline int raid6_next_disk(int disk, int raid_disks)
145 {
146 	disk++;
147 	return (disk < raid_disks) ? disk : 0;
148 }
149 
150 /* When walking through the disks in a raid5, starting at raid6_d0,
151  * We need to map each disk to a 'slot', where the data disks are slot
152  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
153  * is raid_disks-1.  This help does that mapping.
154  */
155 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
156 			     int *count, int syndrome_disks)
157 {
158 	int slot = *count;
159 
160 	if (sh->ddf_layout)
161 		(*count)++;
162 	if (idx == sh->pd_idx)
163 		return syndrome_disks;
164 	if (idx == sh->qd_idx)
165 		return syndrome_disks + 1;
166 	if (!sh->ddf_layout)
167 		(*count)++;
168 	return slot;
169 }
170 
171 static void return_io(struct bio *return_bi)
172 {
173 	struct bio *bi = return_bi;
174 	while (bi) {
175 
176 		return_bi = bi->bi_next;
177 		bi->bi_next = NULL;
178 		bi->bi_size = 0;
179 		bio_endio(bi, 0);
180 		bi = return_bi;
181 	}
182 }
183 
184 static void print_raid5_conf (struct r5conf *conf);
185 
186 static int stripe_operations_active(struct stripe_head *sh)
187 {
188 	return sh->check_state || sh->reconstruct_state ||
189 	       test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
190 	       test_bit(STRIPE_COMPUTE_RUN, &sh->state);
191 }
192 
193 static void __release_stripe(struct r5conf *conf, struct stripe_head *sh)
194 {
195 	if (atomic_dec_and_test(&sh->count)) {
196 		BUG_ON(!list_empty(&sh->lru));
197 		BUG_ON(atomic_read(&conf->active_stripes)==0);
198 		if (test_bit(STRIPE_HANDLE, &sh->state)) {
199 			if (test_bit(STRIPE_DELAYED, &sh->state))
200 				list_add_tail(&sh->lru, &conf->delayed_list);
201 			else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
202 				   sh->bm_seq - conf->seq_write > 0)
203 				list_add_tail(&sh->lru, &conf->bitmap_list);
204 			else {
205 				clear_bit(STRIPE_BIT_DELAY, &sh->state);
206 				list_add_tail(&sh->lru, &conf->handle_list);
207 			}
208 			md_wakeup_thread(conf->mddev->thread);
209 		} else {
210 			BUG_ON(stripe_operations_active(sh));
211 			if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
212 				atomic_dec(&conf->preread_active_stripes);
213 				if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
214 					md_wakeup_thread(conf->mddev->thread);
215 			}
216 			atomic_dec(&conf->active_stripes);
217 			if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
218 				list_add_tail(&sh->lru, &conf->inactive_list);
219 				wake_up(&conf->wait_for_stripe);
220 				if (conf->retry_read_aligned)
221 					md_wakeup_thread(conf->mddev->thread);
222 			}
223 		}
224 	}
225 }
226 
227 static void release_stripe(struct stripe_head *sh)
228 {
229 	struct r5conf *conf = sh->raid_conf;
230 	unsigned long flags;
231 
232 	spin_lock_irqsave(&conf->device_lock, flags);
233 	__release_stripe(conf, sh);
234 	spin_unlock_irqrestore(&conf->device_lock, flags);
235 }
236 
237 static inline void remove_hash(struct stripe_head *sh)
238 {
239 	pr_debug("remove_hash(), stripe %llu\n",
240 		(unsigned long long)sh->sector);
241 
242 	hlist_del_init(&sh->hash);
243 }
244 
245 static inline void insert_hash(struct r5conf *conf, struct stripe_head *sh)
246 {
247 	struct hlist_head *hp = stripe_hash(conf, sh->sector);
248 
249 	pr_debug("insert_hash(), stripe %llu\n",
250 		(unsigned long long)sh->sector);
251 
252 	hlist_add_head(&sh->hash, hp);
253 }
254 
255 
256 /* find an idle stripe, make sure it is unhashed, and return it. */
257 static struct stripe_head *get_free_stripe(struct r5conf *conf)
258 {
259 	struct stripe_head *sh = NULL;
260 	struct list_head *first;
261 
262 	if (list_empty(&conf->inactive_list))
263 		goto out;
264 	first = conf->inactive_list.next;
265 	sh = list_entry(first, struct stripe_head, lru);
266 	list_del_init(first);
267 	remove_hash(sh);
268 	atomic_inc(&conf->active_stripes);
269 out:
270 	return sh;
271 }
272 
273 static void shrink_buffers(struct stripe_head *sh)
274 {
275 	struct page *p;
276 	int i;
277 	int num = sh->raid_conf->pool_size;
278 
279 	for (i = 0; i < num ; i++) {
280 		p = sh->dev[i].page;
281 		if (!p)
282 			continue;
283 		sh->dev[i].page = NULL;
284 		put_page(p);
285 	}
286 }
287 
288 static int grow_buffers(struct stripe_head *sh)
289 {
290 	int i;
291 	int num = sh->raid_conf->pool_size;
292 
293 	for (i = 0; i < num; i++) {
294 		struct page *page;
295 
296 		if (!(page = alloc_page(GFP_KERNEL))) {
297 			return 1;
298 		}
299 		sh->dev[i].page = page;
300 	}
301 	return 0;
302 }
303 
304 static void raid5_build_block(struct stripe_head *sh, int i, int previous);
305 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
306 			    struct stripe_head *sh);
307 
308 static void init_stripe(struct stripe_head *sh, sector_t sector, int previous)
309 {
310 	struct r5conf *conf = sh->raid_conf;
311 	int i;
312 
313 	BUG_ON(atomic_read(&sh->count) != 0);
314 	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
315 	BUG_ON(stripe_operations_active(sh));
316 
317 	pr_debug("init_stripe called, stripe %llu\n",
318 		(unsigned long long)sh->sector);
319 
320 	remove_hash(sh);
321 
322 	sh->generation = conf->generation - previous;
323 	sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
324 	sh->sector = sector;
325 	stripe_set_idx(sector, conf, previous, sh);
326 	sh->state = 0;
327 
328 
329 	for (i = sh->disks; i--; ) {
330 		struct r5dev *dev = &sh->dev[i];
331 
332 		if (dev->toread || dev->read || dev->towrite || dev->written ||
333 		    test_bit(R5_LOCKED, &dev->flags)) {
334 			printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
335 			       (unsigned long long)sh->sector, i, dev->toread,
336 			       dev->read, dev->towrite, dev->written,
337 			       test_bit(R5_LOCKED, &dev->flags));
338 			WARN_ON(1);
339 		}
340 		dev->flags = 0;
341 		raid5_build_block(sh, i, previous);
342 	}
343 	insert_hash(conf, sh);
344 }
345 
346 static struct stripe_head *__find_stripe(struct r5conf *conf, sector_t sector,
347 					 short generation)
348 {
349 	struct stripe_head *sh;
350 	struct hlist_node *hn;
351 
352 	pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
353 	hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
354 		if (sh->sector == sector && sh->generation == generation)
355 			return sh;
356 	pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
357 	return NULL;
358 }
359 
360 /*
361  * Need to check if array has failed when deciding whether to:
362  *  - start an array
363  *  - remove non-faulty devices
364  *  - add a spare
365  *  - allow a reshape
366  * This determination is simple when no reshape is happening.
367  * However if there is a reshape, we need to carefully check
368  * both the before and after sections.
369  * This is because some failed devices may only affect one
370  * of the two sections, and some non-in_sync devices may
371  * be insync in the section most affected by failed devices.
372  */
373 static int has_failed(struct r5conf *conf)
374 {
375 	int degraded;
376 	int i;
377 	if (conf->mddev->reshape_position == MaxSector)
378 		return conf->mddev->degraded > conf->max_degraded;
379 
380 	rcu_read_lock();
381 	degraded = 0;
382 	for (i = 0; i < conf->previous_raid_disks; i++) {
383 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
384 		if (!rdev || test_bit(Faulty, &rdev->flags))
385 			degraded++;
386 		else if (test_bit(In_sync, &rdev->flags))
387 			;
388 		else
389 			/* not in-sync or faulty.
390 			 * If the reshape increases the number of devices,
391 			 * this is being recovered by the reshape, so
392 			 * this 'previous' section is not in_sync.
393 			 * If the number of devices is being reduced however,
394 			 * the device can only be part of the array if
395 			 * we are reverting a reshape, so this section will
396 			 * be in-sync.
397 			 */
398 			if (conf->raid_disks >= conf->previous_raid_disks)
399 				degraded++;
400 	}
401 	rcu_read_unlock();
402 	if (degraded > conf->max_degraded)
403 		return 1;
404 	rcu_read_lock();
405 	degraded = 0;
406 	for (i = 0; i < conf->raid_disks; i++) {
407 		struct md_rdev *rdev = rcu_dereference(conf->disks[i].rdev);
408 		if (!rdev || test_bit(Faulty, &rdev->flags))
409 			degraded++;
410 		else if (test_bit(In_sync, &rdev->flags))
411 			;
412 		else
413 			/* not in-sync or faulty.
414 			 * If reshape increases the number of devices, this
415 			 * section has already been recovered, else it
416 			 * almost certainly hasn't.
417 			 */
418 			if (conf->raid_disks <= conf->previous_raid_disks)
419 				degraded++;
420 	}
421 	rcu_read_unlock();
422 	if (degraded > conf->max_degraded)
423 		return 1;
424 	return 0;
425 }
426 
427 static struct stripe_head *
428 get_active_stripe(struct r5conf *conf, sector_t sector,
429 		  int previous, int noblock, int noquiesce)
430 {
431 	struct stripe_head *sh;
432 
433 	pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
434 
435 	spin_lock_irq(&conf->device_lock);
436 
437 	do {
438 		wait_event_lock_irq(conf->wait_for_stripe,
439 				    conf->quiesce == 0 || noquiesce,
440 				    conf->device_lock, /* nothing */);
441 		sh = __find_stripe(conf, sector, conf->generation - previous);
442 		if (!sh) {
443 			if (!conf->inactive_blocked)
444 				sh = get_free_stripe(conf);
445 			if (noblock && sh == NULL)
446 				break;
447 			if (!sh) {
448 				conf->inactive_blocked = 1;
449 				wait_event_lock_irq(conf->wait_for_stripe,
450 						    !list_empty(&conf->inactive_list) &&
451 						    (atomic_read(&conf->active_stripes)
452 						     < (conf->max_nr_stripes *3/4)
453 						     || !conf->inactive_blocked),
454 						    conf->device_lock,
455 						    );
456 				conf->inactive_blocked = 0;
457 			} else
458 				init_stripe(sh, sector, previous);
459 		} else {
460 			if (atomic_read(&sh->count)) {
461 				BUG_ON(!list_empty(&sh->lru)
462 				    && !test_bit(STRIPE_EXPANDING, &sh->state));
463 			} else {
464 				if (!test_bit(STRIPE_HANDLE, &sh->state))
465 					atomic_inc(&conf->active_stripes);
466 				if (list_empty(&sh->lru) &&
467 				    !test_bit(STRIPE_EXPANDING, &sh->state))
468 					BUG();
469 				list_del_init(&sh->lru);
470 			}
471 		}
472 	} while (sh == NULL);
473 
474 	if (sh)
475 		atomic_inc(&sh->count);
476 
477 	spin_unlock_irq(&conf->device_lock);
478 	return sh;
479 }
480 
481 static void
482 raid5_end_read_request(struct bio *bi, int error);
483 static void
484 raid5_end_write_request(struct bio *bi, int error);
485 
486 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
487 {
488 	struct r5conf *conf = sh->raid_conf;
489 	int i, disks = sh->disks;
490 
491 	might_sleep();
492 
493 	for (i = disks; i--; ) {
494 		int rw;
495 		struct bio *bi;
496 		struct md_rdev *rdev;
497 		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
498 			if (test_and_clear_bit(R5_WantFUA, &sh->dev[i].flags))
499 				rw = WRITE_FUA;
500 			else
501 				rw = WRITE;
502 		} else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
503 			rw = READ;
504 		else
505 			continue;
506 
507 		bi = &sh->dev[i].req;
508 
509 		bi->bi_rw = rw;
510 		if (rw & WRITE)
511 			bi->bi_end_io = raid5_end_write_request;
512 		else
513 			bi->bi_end_io = raid5_end_read_request;
514 
515 		rcu_read_lock();
516 		rdev = rcu_dereference(conf->disks[i].rdev);
517 		if (rdev && test_bit(Faulty, &rdev->flags))
518 			rdev = NULL;
519 		if (rdev)
520 			atomic_inc(&rdev->nr_pending);
521 		rcu_read_unlock();
522 
523 		/* We have already checked bad blocks for reads.  Now
524 		 * need to check for writes.
525 		 */
526 		while ((rw & WRITE) && rdev &&
527 		       test_bit(WriteErrorSeen, &rdev->flags)) {
528 			sector_t first_bad;
529 			int bad_sectors;
530 			int bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
531 					      &first_bad, &bad_sectors);
532 			if (!bad)
533 				break;
534 
535 			if (bad < 0) {
536 				set_bit(BlockedBadBlocks, &rdev->flags);
537 				if (!conf->mddev->external &&
538 				    conf->mddev->flags) {
539 					/* It is very unlikely, but we might
540 					 * still need to write out the
541 					 * bad block log - better give it
542 					 * a chance*/
543 					md_check_recovery(conf->mddev);
544 				}
545 				md_wait_for_blocked_rdev(rdev, conf->mddev);
546 			} else {
547 				/* Acknowledged bad block - skip the write */
548 				rdev_dec_pending(rdev, conf->mddev);
549 				rdev = NULL;
550 			}
551 		}
552 
553 		if (rdev) {
554 			if (s->syncing || s->expanding || s->expanded)
555 				md_sync_acct(rdev->bdev, STRIPE_SECTORS);
556 
557 			set_bit(STRIPE_IO_STARTED, &sh->state);
558 
559 			bi->bi_bdev = rdev->bdev;
560 			pr_debug("%s: for %llu schedule op %ld on disc %d\n",
561 				__func__, (unsigned long long)sh->sector,
562 				bi->bi_rw, i);
563 			atomic_inc(&sh->count);
564 			bi->bi_sector = sh->sector + rdev->data_offset;
565 			bi->bi_flags = 1 << BIO_UPTODATE;
566 			bi->bi_vcnt = 1;
567 			bi->bi_max_vecs = 1;
568 			bi->bi_idx = 0;
569 			bi->bi_io_vec = &sh->dev[i].vec;
570 			bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
571 			bi->bi_io_vec[0].bv_offset = 0;
572 			bi->bi_size = STRIPE_SIZE;
573 			bi->bi_next = NULL;
574 			generic_make_request(bi);
575 		} else {
576 			if (rw & WRITE)
577 				set_bit(STRIPE_DEGRADED, &sh->state);
578 			pr_debug("skip op %ld on disc %d for sector %llu\n",
579 				bi->bi_rw, i, (unsigned long long)sh->sector);
580 			clear_bit(R5_LOCKED, &sh->dev[i].flags);
581 			set_bit(STRIPE_HANDLE, &sh->state);
582 		}
583 	}
584 }
585 
586 static struct dma_async_tx_descriptor *
587 async_copy_data(int frombio, struct bio *bio, struct page *page,
588 	sector_t sector, struct dma_async_tx_descriptor *tx)
589 {
590 	struct bio_vec *bvl;
591 	struct page *bio_page;
592 	int i;
593 	int page_offset;
594 	struct async_submit_ctl submit;
595 	enum async_tx_flags flags = 0;
596 
597 	if (bio->bi_sector >= sector)
598 		page_offset = (signed)(bio->bi_sector - sector) * 512;
599 	else
600 		page_offset = (signed)(sector - bio->bi_sector) * -512;
601 
602 	if (frombio)
603 		flags |= ASYNC_TX_FENCE;
604 	init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
605 
606 	bio_for_each_segment(bvl, bio, i) {
607 		int len = bvl->bv_len;
608 		int clen;
609 		int b_offset = 0;
610 
611 		if (page_offset < 0) {
612 			b_offset = -page_offset;
613 			page_offset += b_offset;
614 			len -= b_offset;
615 		}
616 
617 		if (len > 0 && page_offset + len > STRIPE_SIZE)
618 			clen = STRIPE_SIZE - page_offset;
619 		else
620 			clen = len;
621 
622 		if (clen > 0) {
623 			b_offset += bvl->bv_offset;
624 			bio_page = bvl->bv_page;
625 			if (frombio)
626 				tx = async_memcpy(page, bio_page, page_offset,
627 						  b_offset, clen, &submit);
628 			else
629 				tx = async_memcpy(bio_page, page, b_offset,
630 						  page_offset, clen, &submit);
631 		}
632 		/* chain the operations */
633 		submit.depend_tx = tx;
634 
635 		if (clen < len) /* hit end of page */
636 			break;
637 		page_offset +=  len;
638 	}
639 
640 	return tx;
641 }
642 
643 static void ops_complete_biofill(void *stripe_head_ref)
644 {
645 	struct stripe_head *sh = stripe_head_ref;
646 	struct bio *return_bi = NULL;
647 	struct r5conf *conf = sh->raid_conf;
648 	int i;
649 
650 	pr_debug("%s: stripe %llu\n", __func__,
651 		(unsigned long long)sh->sector);
652 
653 	/* clear completed biofills */
654 	spin_lock_irq(&conf->device_lock);
655 	for (i = sh->disks; i--; ) {
656 		struct r5dev *dev = &sh->dev[i];
657 
658 		/* acknowledge completion of a biofill operation */
659 		/* and check if we need to reply to a read request,
660 		 * new R5_Wantfill requests are held off until
661 		 * !STRIPE_BIOFILL_RUN
662 		 */
663 		if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
664 			struct bio *rbi, *rbi2;
665 
666 			BUG_ON(!dev->read);
667 			rbi = dev->read;
668 			dev->read = NULL;
669 			while (rbi && rbi->bi_sector <
670 				dev->sector + STRIPE_SECTORS) {
671 				rbi2 = r5_next_bio(rbi, dev->sector);
672 				if (!raid5_dec_bi_phys_segments(rbi)) {
673 					rbi->bi_next = return_bi;
674 					return_bi = rbi;
675 				}
676 				rbi = rbi2;
677 			}
678 		}
679 	}
680 	spin_unlock_irq(&conf->device_lock);
681 	clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
682 
683 	return_io(return_bi);
684 
685 	set_bit(STRIPE_HANDLE, &sh->state);
686 	release_stripe(sh);
687 }
688 
689 static void ops_run_biofill(struct stripe_head *sh)
690 {
691 	struct dma_async_tx_descriptor *tx = NULL;
692 	struct r5conf *conf = sh->raid_conf;
693 	struct async_submit_ctl submit;
694 	int i;
695 
696 	pr_debug("%s: stripe %llu\n", __func__,
697 		(unsigned long long)sh->sector);
698 
699 	for (i = sh->disks; i--; ) {
700 		struct r5dev *dev = &sh->dev[i];
701 		if (test_bit(R5_Wantfill, &dev->flags)) {
702 			struct bio *rbi;
703 			spin_lock_irq(&conf->device_lock);
704 			dev->read = rbi = dev->toread;
705 			dev->toread = NULL;
706 			spin_unlock_irq(&conf->device_lock);
707 			while (rbi && rbi->bi_sector <
708 				dev->sector + STRIPE_SECTORS) {
709 				tx = async_copy_data(0, rbi, dev->page,
710 					dev->sector, tx);
711 				rbi = r5_next_bio(rbi, dev->sector);
712 			}
713 		}
714 	}
715 
716 	atomic_inc(&sh->count);
717 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
718 	async_trigger_callback(&submit);
719 }
720 
721 static void mark_target_uptodate(struct stripe_head *sh, int target)
722 {
723 	struct r5dev *tgt;
724 
725 	if (target < 0)
726 		return;
727 
728 	tgt = &sh->dev[target];
729 	set_bit(R5_UPTODATE, &tgt->flags);
730 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
731 	clear_bit(R5_Wantcompute, &tgt->flags);
732 }
733 
734 static void ops_complete_compute(void *stripe_head_ref)
735 {
736 	struct stripe_head *sh = stripe_head_ref;
737 
738 	pr_debug("%s: stripe %llu\n", __func__,
739 		(unsigned long long)sh->sector);
740 
741 	/* mark the computed target(s) as uptodate */
742 	mark_target_uptodate(sh, sh->ops.target);
743 	mark_target_uptodate(sh, sh->ops.target2);
744 
745 	clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
746 	if (sh->check_state == check_state_compute_run)
747 		sh->check_state = check_state_compute_result;
748 	set_bit(STRIPE_HANDLE, &sh->state);
749 	release_stripe(sh);
750 }
751 
752 /* return a pointer to the address conversion region of the scribble buffer */
753 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
754 				 struct raid5_percpu *percpu)
755 {
756 	return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
757 }
758 
759 static struct dma_async_tx_descriptor *
760 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
761 {
762 	int disks = sh->disks;
763 	struct page **xor_srcs = percpu->scribble;
764 	int target = sh->ops.target;
765 	struct r5dev *tgt = &sh->dev[target];
766 	struct page *xor_dest = tgt->page;
767 	int count = 0;
768 	struct dma_async_tx_descriptor *tx;
769 	struct async_submit_ctl submit;
770 	int i;
771 
772 	pr_debug("%s: stripe %llu block: %d\n",
773 		__func__, (unsigned long long)sh->sector, target);
774 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
775 
776 	for (i = disks; i--; )
777 		if (i != target)
778 			xor_srcs[count++] = sh->dev[i].page;
779 
780 	atomic_inc(&sh->count);
781 
782 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
783 			  ops_complete_compute, sh, to_addr_conv(sh, percpu));
784 	if (unlikely(count == 1))
785 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
786 	else
787 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
788 
789 	return tx;
790 }
791 
792 /* set_syndrome_sources - populate source buffers for gen_syndrome
793  * @srcs - (struct page *) array of size sh->disks
794  * @sh - stripe_head to parse
795  *
796  * Populates srcs in proper layout order for the stripe and returns the
797  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
798  * destination buffer is recorded in srcs[count] and the Q destination
799  * is recorded in srcs[count+1]].
800  */
801 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
802 {
803 	int disks = sh->disks;
804 	int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
805 	int d0_idx = raid6_d0(sh);
806 	int count;
807 	int i;
808 
809 	for (i = 0; i < disks; i++)
810 		srcs[i] = NULL;
811 
812 	count = 0;
813 	i = d0_idx;
814 	do {
815 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
816 
817 		srcs[slot] = sh->dev[i].page;
818 		i = raid6_next_disk(i, disks);
819 	} while (i != d0_idx);
820 
821 	return syndrome_disks;
822 }
823 
824 static struct dma_async_tx_descriptor *
825 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
826 {
827 	int disks = sh->disks;
828 	struct page **blocks = percpu->scribble;
829 	int target;
830 	int qd_idx = sh->qd_idx;
831 	struct dma_async_tx_descriptor *tx;
832 	struct async_submit_ctl submit;
833 	struct r5dev *tgt;
834 	struct page *dest;
835 	int i;
836 	int count;
837 
838 	if (sh->ops.target < 0)
839 		target = sh->ops.target2;
840 	else if (sh->ops.target2 < 0)
841 		target = sh->ops.target;
842 	else
843 		/* we should only have one valid target */
844 		BUG();
845 	BUG_ON(target < 0);
846 	pr_debug("%s: stripe %llu block: %d\n",
847 		__func__, (unsigned long long)sh->sector, target);
848 
849 	tgt = &sh->dev[target];
850 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
851 	dest = tgt->page;
852 
853 	atomic_inc(&sh->count);
854 
855 	if (target == qd_idx) {
856 		count = set_syndrome_sources(blocks, sh);
857 		blocks[count] = NULL; /* regenerating p is not necessary */
858 		BUG_ON(blocks[count+1] != dest); /* q should already be set */
859 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
860 				  ops_complete_compute, sh,
861 				  to_addr_conv(sh, percpu));
862 		tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
863 	} else {
864 		/* Compute any data- or p-drive using XOR */
865 		count = 0;
866 		for (i = disks; i-- ; ) {
867 			if (i == target || i == qd_idx)
868 				continue;
869 			blocks[count++] = sh->dev[i].page;
870 		}
871 
872 		init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
873 				  NULL, ops_complete_compute, sh,
874 				  to_addr_conv(sh, percpu));
875 		tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
876 	}
877 
878 	return tx;
879 }
880 
881 static struct dma_async_tx_descriptor *
882 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
883 {
884 	int i, count, disks = sh->disks;
885 	int syndrome_disks = sh->ddf_layout ? disks : disks-2;
886 	int d0_idx = raid6_d0(sh);
887 	int faila = -1, failb = -1;
888 	int target = sh->ops.target;
889 	int target2 = sh->ops.target2;
890 	struct r5dev *tgt = &sh->dev[target];
891 	struct r5dev *tgt2 = &sh->dev[target2];
892 	struct dma_async_tx_descriptor *tx;
893 	struct page **blocks = percpu->scribble;
894 	struct async_submit_ctl submit;
895 
896 	pr_debug("%s: stripe %llu block1: %d block2: %d\n",
897 		 __func__, (unsigned long long)sh->sector, target, target2);
898 	BUG_ON(target < 0 || target2 < 0);
899 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
900 	BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
901 
902 	/* we need to open-code set_syndrome_sources to handle the
903 	 * slot number conversion for 'faila' and 'failb'
904 	 */
905 	for (i = 0; i < disks ; i++)
906 		blocks[i] = NULL;
907 	count = 0;
908 	i = d0_idx;
909 	do {
910 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
911 
912 		blocks[slot] = sh->dev[i].page;
913 
914 		if (i == target)
915 			faila = slot;
916 		if (i == target2)
917 			failb = slot;
918 		i = raid6_next_disk(i, disks);
919 	} while (i != d0_idx);
920 
921 	BUG_ON(faila == failb);
922 	if (failb < faila)
923 		swap(faila, failb);
924 	pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
925 		 __func__, (unsigned long long)sh->sector, faila, failb);
926 
927 	atomic_inc(&sh->count);
928 
929 	if (failb == syndrome_disks+1) {
930 		/* Q disk is one of the missing disks */
931 		if (faila == syndrome_disks) {
932 			/* Missing P+Q, just recompute */
933 			init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
934 					  ops_complete_compute, sh,
935 					  to_addr_conv(sh, percpu));
936 			return async_gen_syndrome(blocks, 0, syndrome_disks+2,
937 						  STRIPE_SIZE, &submit);
938 		} else {
939 			struct page *dest;
940 			int data_target;
941 			int qd_idx = sh->qd_idx;
942 
943 			/* Missing D+Q: recompute D from P, then recompute Q */
944 			if (target == qd_idx)
945 				data_target = target2;
946 			else
947 				data_target = target;
948 
949 			count = 0;
950 			for (i = disks; i-- ; ) {
951 				if (i == data_target || i == qd_idx)
952 					continue;
953 				blocks[count++] = sh->dev[i].page;
954 			}
955 			dest = sh->dev[data_target].page;
956 			init_async_submit(&submit,
957 					  ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
958 					  NULL, NULL, NULL,
959 					  to_addr_conv(sh, percpu));
960 			tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
961 				       &submit);
962 
963 			count = set_syndrome_sources(blocks, sh);
964 			init_async_submit(&submit, ASYNC_TX_FENCE, tx,
965 					  ops_complete_compute, sh,
966 					  to_addr_conv(sh, percpu));
967 			return async_gen_syndrome(blocks, 0, count+2,
968 						  STRIPE_SIZE, &submit);
969 		}
970 	} else {
971 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
972 				  ops_complete_compute, sh,
973 				  to_addr_conv(sh, percpu));
974 		if (failb == syndrome_disks) {
975 			/* We're missing D+P. */
976 			return async_raid6_datap_recov(syndrome_disks+2,
977 						       STRIPE_SIZE, faila,
978 						       blocks, &submit);
979 		} else {
980 			/* We're missing D+D. */
981 			return async_raid6_2data_recov(syndrome_disks+2,
982 						       STRIPE_SIZE, faila, failb,
983 						       blocks, &submit);
984 		}
985 	}
986 }
987 
988 
989 static void ops_complete_prexor(void *stripe_head_ref)
990 {
991 	struct stripe_head *sh = stripe_head_ref;
992 
993 	pr_debug("%s: stripe %llu\n", __func__,
994 		(unsigned long long)sh->sector);
995 }
996 
997 static struct dma_async_tx_descriptor *
998 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
999 	       struct dma_async_tx_descriptor *tx)
1000 {
1001 	int disks = sh->disks;
1002 	struct page **xor_srcs = percpu->scribble;
1003 	int count = 0, pd_idx = sh->pd_idx, i;
1004 	struct async_submit_ctl submit;
1005 
1006 	/* existing parity data subtracted */
1007 	struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1008 
1009 	pr_debug("%s: stripe %llu\n", __func__,
1010 		(unsigned long long)sh->sector);
1011 
1012 	for (i = disks; i--; ) {
1013 		struct r5dev *dev = &sh->dev[i];
1014 		/* Only process blocks that are known to be uptodate */
1015 		if (test_bit(R5_Wantdrain, &dev->flags))
1016 			xor_srcs[count++] = dev->page;
1017 	}
1018 
1019 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
1020 			  ops_complete_prexor, sh, to_addr_conv(sh, percpu));
1021 	tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1022 
1023 	return tx;
1024 }
1025 
1026 static struct dma_async_tx_descriptor *
1027 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
1028 {
1029 	int disks = sh->disks;
1030 	int i;
1031 
1032 	pr_debug("%s: stripe %llu\n", __func__,
1033 		(unsigned long long)sh->sector);
1034 
1035 	for (i = disks; i--; ) {
1036 		struct r5dev *dev = &sh->dev[i];
1037 		struct bio *chosen;
1038 
1039 		if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
1040 			struct bio *wbi;
1041 
1042 			spin_lock_irq(&sh->raid_conf->device_lock);
1043 			chosen = dev->towrite;
1044 			dev->towrite = NULL;
1045 			BUG_ON(dev->written);
1046 			wbi = dev->written = chosen;
1047 			spin_unlock_irq(&sh->raid_conf->device_lock);
1048 
1049 			while (wbi && wbi->bi_sector <
1050 				dev->sector + STRIPE_SECTORS) {
1051 				if (wbi->bi_rw & REQ_FUA)
1052 					set_bit(R5_WantFUA, &dev->flags);
1053 				tx = async_copy_data(1, wbi, dev->page,
1054 					dev->sector, tx);
1055 				wbi = r5_next_bio(wbi, dev->sector);
1056 			}
1057 		}
1058 	}
1059 
1060 	return tx;
1061 }
1062 
1063 static void ops_complete_reconstruct(void *stripe_head_ref)
1064 {
1065 	struct stripe_head *sh = stripe_head_ref;
1066 	int disks = sh->disks;
1067 	int pd_idx = sh->pd_idx;
1068 	int qd_idx = sh->qd_idx;
1069 	int i;
1070 	bool fua = false;
1071 
1072 	pr_debug("%s: stripe %llu\n", __func__,
1073 		(unsigned long long)sh->sector);
1074 
1075 	for (i = disks; i--; )
1076 		fua |= test_bit(R5_WantFUA, &sh->dev[i].flags);
1077 
1078 	for (i = disks; i--; ) {
1079 		struct r5dev *dev = &sh->dev[i];
1080 
1081 		if (dev->written || i == pd_idx || i == qd_idx) {
1082 			set_bit(R5_UPTODATE, &dev->flags);
1083 			if (fua)
1084 				set_bit(R5_WantFUA, &dev->flags);
1085 		}
1086 	}
1087 
1088 	if (sh->reconstruct_state == reconstruct_state_drain_run)
1089 		sh->reconstruct_state = reconstruct_state_drain_result;
1090 	else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
1091 		sh->reconstruct_state = reconstruct_state_prexor_drain_result;
1092 	else {
1093 		BUG_ON(sh->reconstruct_state != reconstruct_state_run);
1094 		sh->reconstruct_state = reconstruct_state_result;
1095 	}
1096 
1097 	set_bit(STRIPE_HANDLE, &sh->state);
1098 	release_stripe(sh);
1099 }
1100 
1101 static void
1102 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1103 		     struct dma_async_tx_descriptor *tx)
1104 {
1105 	int disks = sh->disks;
1106 	struct page **xor_srcs = percpu->scribble;
1107 	struct async_submit_ctl submit;
1108 	int count = 0, pd_idx = sh->pd_idx, i;
1109 	struct page *xor_dest;
1110 	int prexor = 0;
1111 	unsigned long flags;
1112 
1113 	pr_debug("%s: stripe %llu\n", __func__,
1114 		(unsigned long long)sh->sector);
1115 
1116 	/* check if prexor is active which means only process blocks
1117 	 * that are part of a read-modify-write (written)
1118 	 */
1119 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1120 		prexor = 1;
1121 		xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1122 		for (i = disks; i--; ) {
1123 			struct r5dev *dev = &sh->dev[i];
1124 			if (dev->written)
1125 				xor_srcs[count++] = dev->page;
1126 		}
1127 	} else {
1128 		xor_dest = sh->dev[pd_idx].page;
1129 		for (i = disks; i--; ) {
1130 			struct r5dev *dev = &sh->dev[i];
1131 			if (i != pd_idx)
1132 				xor_srcs[count++] = dev->page;
1133 		}
1134 	}
1135 
1136 	/* 1/ if we prexor'd then the dest is reused as a source
1137 	 * 2/ if we did not prexor then we are redoing the parity
1138 	 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1139 	 * for the synchronous xor case
1140 	 */
1141 	flags = ASYNC_TX_ACK |
1142 		(prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1143 
1144 	atomic_inc(&sh->count);
1145 
1146 	init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1147 			  to_addr_conv(sh, percpu));
1148 	if (unlikely(count == 1))
1149 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1150 	else
1151 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1152 }
1153 
1154 static void
1155 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1156 		     struct dma_async_tx_descriptor *tx)
1157 {
1158 	struct async_submit_ctl submit;
1159 	struct page **blocks = percpu->scribble;
1160 	int count;
1161 
1162 	pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1163 
1164 	count = set_syndrome_sources(blocks, sh);
1165 
1166 	atomic_inc(&sh->count);
1167 
1168 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1169 			  sh, to_addr_conv(sh, percpu));
1170 	async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1171 }
1172 
1173 static void ops_complete_check(void *stripe_head_ref)
1174 {
1175 	struct stripe_head *sh = stripe_head_ref;
1176 
1177 	pr_debug("%s: stripe %llu\n", __func__,
1178 		(unsigned long long)sh->sector);
1179 
1180 	sh->check_state = check_state_check_result;
1181 	set_bit(STRIPE_HANDLE, &sh->state);
1182 	release_stripe(sh);
1183 }
1184 
1185 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1186 {
1187 	int disks = sh->disks;
1188 	int pd_idx = sh->pd_idx;
1189 	int qd_idx = sh->qd_idx;
1190 	struct page *xor_dest;
1191 	struct page **xor_srcs = percpu->scribble;
1192 	struct dma_async_tx_descriptor *tx;
1193 	struct async_submit_ctl submit;
1194 	int count;
1195 	int i;
1196 
1197 	pr_debug("%s: stripe %llu\n", __func__,
1198 		(unsigned long long)sh->sector);
1199 
1200 	count = 0;
1201 	xor_dest = sh->dev[pd_idx].page;
1202 	xor_srcs[count++] = xor_dest;
1203 	for (i = disks; i--; ) {
1204 		if (i == pd_idx || i == qd_idx)
1205 			continue;
1206 		xor_srcs[count++] = sh->dev[i].page;
1207 	}
1208 
1209 	init_async_submit(&submit, 0, NULL, NULL, NULL,
1210 			  to_addr_conv(sh, percpu));
1211 	tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1212 			   &sh->ops.zero_sum_result, &submit);
1213 
1214 	atomic_inc(&sh->count);
1215 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1216 	tx = async_trigger_callback(&submit);
1217 }
1218 
1219 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1220 {
1221 	struct page **srcs = percpu->scribble;
1222 	struct async_submit_ctl submit;
1223 	int count;
1224 
1225 	pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1226 		(unsigned long long)sh->sector, checkp);
1227 
1228 	count = set_syndrome_sources(srcs, sh);
1229 	if (!checkp)
1230 		srcs[count] = NULL;
1231 
1232 	atomic_inc(&sh->count);
1233 	init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1234 			  sh, to_addr_conv(sh, percpu));
1235 	async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1236 			   &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1237 }
1238 
1239 static void __raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1240 {
1241 	int overlap_clear = 0, i, disks = sh->disks;
1242 	struct dma_async_tx_descriptor *tx = NULL;
1243 	struct r5conf *conf = sh->raid_conf;
1244 	int level = conf->level;
1245 	struct raid5_percpu *percpu;
1246 	unsigned long cpu;
1247 
1248 	cpu = get_cpu();
1249 	percpu = per_cpu_ptr(conf->percpu, cpu);
1250 	if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1251 		ops_run_biofill(sh);
1252 		overlap_clear++;
1253 	}
1254 
1255 	if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1256 		if (level < 6)
1257 			tx = ops_run_compute5(sh, percpu);
1258 		else {
1259 			if (sh->ops.target2 < 0 || sh->ops.target < 0)
1260 				tx = ops_run_compute6_1(sh, percpu);
1261 			else
1262 				tx = ops_run_compute6_2(sh, percpu);
1263 		}
1264 		/* terminate the chain if reconstruct is not set to be run */
1265 		if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1266 			async_tx_ack(tx);
1267 	}
1268 
1269 	if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1270 		tx = ops_run_prexor(sh, percpu, tx);
1271 
1272 	if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1273 		tx = ops_run_biodrain(sh, tx);
1274 		overlap_clear++;
1275 	}
1276 
1277 	if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1278 		if (level < 6)
1279 			ops_run_reconstruct5(sh, percpu, tx);
1280 		else
1281 			ops_run_reconstruct6(sh, percpu, tx);
1282 	}
1283 
1284 	if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1285 		if (sh->check_state == check_state_run)
1286 			ops_run_check_p(sh, percpu);
1287 		else if (sh->check_state == check_state_run_q)
1288 			ops_run_check_pq(sh, percpu, 0);
1289 		else if (sh->check_state == check_state_run_pq)
1290 			ops_run_check_pq(sh, percpu, 1);
1291 		else
1292 			BUG();
1293 	}
1294 
1295 	if (overlap_clear)
1296 		for (i = disks; i--; ) {
1297 			struct r5dev *dev = &sh->dev[i];
1298 			if (test_and_clear_bit(R5_Overlap, &dev->flags))
1299 				wake_up(&sh->raid_conf->wait_for_overlap);
1300 		}
1301 	put_cpu();
1302 }
1303 
1304 #ifdef CONFIG_MULTICORE_RAID456
1305 static void async_run_ops(void *param, async_cookie_t cookie)
1306 {
1307 	struct stripe_head *sh = param;
1308 	unsigned long ops_request = sh->ops.request;
1309 
1310 	clear_bit_unlock(STRIPE_OPS_REQ_PENDING, &sh->state);
1311 	wake_up(&sh->ops.wait_for_ops);
1312 
1313 	__raid_run_ops(sh, ops_request);
1314 	release_stripe(sh);
1315 }
1316 
1317 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1318 {
1319 	/* since handle_stripe can be called outside of raid5d context
1320 	 * we need to ensure sh->ops.request is de-staged before another
1321 	 * request arrives
1322 	 */
1323 	wait_event(sh->ops.wait_for_ops,
1324 		   !test_and_set_bit_lock(STRIPE_OPS_REQ_PENDING, &sh->state));
1325 	sh->ops.request = ops_request;
1326 
1327 	atomic_inc(&sh->count);
1328 	async_schedule(async_run_ops, sh);
1329 }
1330 #else
1331 #define raid_run_ops __raid_run_ops
1332 #endif
1333 
1334 static int grow_one_stripe(struct r5conf *conf)
1335 {
1336 	struct stripe_head *sh;
1337 	sh = kmem_cache_zalloc(conf->slab_cache, GFP_KERNEL);
1338 	if (!sh)
1339 		return 0;
1340 
1341 	sh->raid_conf = conf;
1342 	#ifdef CONFIG_MULTICORE_RAID456
1343 	init_waitqueue_head(&sh->ops.wait_for_ops);
1344 	#endif
1345 
1346 	if (grow_buffers(sh)) {
1347 		shrink_buffers(sh);
1348 		kmem_cache_free(conf->slab_cache, sh);
1349 		return 0;
1350 	}
1351 	/* we just created an active stripe so... */
1352 	atomic_set(&sh->count, 1);
1353 	atomic_inc(&conf->active_stripes);
1354 	INIT_LIST_HEAD(&sh->lru);
1355 	release_stripe(sh);
1356 	return 1;
1357 }
1358 
1359 static int grow_stripes(struct r5conf *conf, int num)
1360 {
1361 	struct kmem_cache *sc;
1362 	int devs = max(conf->raid_disks, conf->previous_raid_disks);
1363 
1364 	if (conf->mddev->gendisk)
1365 		sprintf(conf->cache_name[0],
1366 			"raid%d-%s", conf->level, mdname(conf->mddev));
1367 	else
1368 		sprintf(conf->cache_name[0],
1369 			"raid%d-%p", conf->level, conf->mddev);
1370 	sprintf(conf->cache_name[1], "%s-alt", conf->cache_name[0]);
1371 
1372 	conf->active_name = 0;
1373 	sc = kmem_cache_create(conf->cache_name[conf->active_name],
1374 			       sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1375 			       0, 0, NULL);
1376 	if (!sc)
1377 		return 1;
1378 	conf->slab_cache = sc;
1379 	conf->pool_size = devs;
1380 	while (num--)
1381 		if (!grow_one_stripe(conf))
1382 			return 1;
1383 	return 0;
1384 }
1385 
1386 /**
1387  * scribble_len - return the required size of the scribble region
1388  * @num - total number of disks in the array
1389  *
1390  * The size must be enough to contain:
1391  * 1/ a struct page pointer for each device in the array +2
1392  * 2/ room to convert each entry in (1) to its corresponding dma
1393  *    (dma_map_page()) or page (page_address()) address.
1394  *
1395  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1396  * calculate over all devices (not just the data blocks), using zeros in place
1397  * of the P and Q blocks.
1398  */
1399 static size_t scribble_len(int num)
1400 {
1401 	size_t len;
1402 
1403 	len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1404 
1405 	return len;
1406 }
1407 
1408 static int resize_stripes(struct r5conf *conf, int newsize)
1409 {
1410 	/* Make all the stripes able to hold 'newsize' devices.
1411 	 * New slots in each stripe get 'page' set to a new page.
1412 	 *
1413 	 * This happens in stages:
1414 	 * 1/ create a new kmem_cache and allocate the required number of
1415 	 *    stripe_heads.
1416 	 * 2/ gather all the old stripe_heads and tranfer the pages across
1417 	 *    to the new stripe_heads.  This will have the side effect of
1418 	 *    freezing the array as once all stripe_heads have been collected,
1419 	 *    no IO will be possible.  Old stripe heads are freed once their
1420 	 *    pages have been transferred over, and the old kmem_cache is
1421 	 *    freed when all stripes are done.
1422 	 * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1423 	 *    we simple return a failre status - no need to clean anything up.
1424 	 * 4/ allocate new pages for the new slots in the new stripe_heads.
1425 	 *    If this fails, we don't bother trying the shrink the
1426 	 *    stripe_heads down again, we just leave them as they are.
1427 	 *    As each stripe_head is processed the new one is released into
1428 	 *    active service.
1429 	 *
1430 	 * Once step2 is started, we cannot afford to wait for a write,
1431 	 * so we use GFP_NOIO allocations.
1432 	 */
1433 	struct stripe_head *osh, *nsh;
1434 	LIST_HEAD(newstripes);
1435 	struct disk_info *ndisks;
1436 	unsigned long cpu;
1437 	int err;
1438 	struct kmem_cache *sc;
1439 	int i;
1440 
1441 	if (newsize <= conf->pool_size)
1442 		return 0; /* never bother to shrink */
1443 
1444 	err = md_allow_write(conf->mddev);
1445 	if (err)
1446 		return err;
1447 
1448 	/* Step 1 */
1449 	sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1450 			       sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1451 			       0, 0, NULL);
1452 	if (!sc)
1453 		return -ENOMEM;
1454 
1455 	for (i = conf->max_nr_stripes; i; i--) {
1456 		nsh = kmem_cache_zalloc(sc, GFP_KERNEL);
1457 		if (!nsh)
1458 			break;
1459 
1460 		nsh->raid_conf = conf;
1461 		#ifdef CONFIG_MULTICORE_RAID456
1462 		init_waitqueue_head(&nsh->ops.wait_for_ops);
1463 		#endif
1464 
1465 		list_add(&nsh->lru, &newstripes);
1466 	}
1467 	if (i) {
1468 		/* didn't get enough, give up */
1469 		while (!list_empty(&newstripes)) {
1470 			nsh = list_entry(newstripes.next, struct stripe_head, lru);
1471 			list_del(&nsh->lru);
1472 			kmem_cache_free(sc, nsh);
1473 		}
1474 		kmem_cache_destroy(sc);
1475 		return -ENOMEM;
1476 	}
1477 	/* Step 2 - Must use GFP_NOIO now.
1478 	 * OK, we have enough stripes, start collecting inactive
1479 	 * stripes and copying them over
1480 	 */
1481 	list_for_each_entry(nsh, &newstripes, lru) {
1482 		spin_lock_irq(&conf->device_lock);
1483 		wait_event_lock_irq(conf->wait_for_stripe,
1484 				    !list_empty(&conf->inactive_list),
1485 				    conf->device_lock,
1486 				    );
1487 		osh = get_free_stripe(conf);
1488 		spin_unlock_irq(&conf->device_lock);
1489 		atomic_set(&nsh->count, 1);
1490 		for(i=0; i<conf->pool_size; i++)
1491 			nsh->dev[i].page = osh->dev[i].page;
1492 		for( ; i<newsize; i++)
1493 			nsh->dev[i].page = NULL;
1494 		kmem_cache_free(conf->slab_cache, osh);
1495 	}
1496 	kmem_cache_destroy(conf->slab_cache);
1497 
1498 	/* Step 3.
1499 	 * At this point, we are holding all the stripes so the array
1500 	 * is completely stalled, so now is a good time to resize
1501 	 * conf->disks and the scribble region
1502 	 */
1503 	ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1504 	if (ndisks) {
1505 		for (i=0; i<conf->raid_disks; i++)
1506 			ndisks[i] = conf->disks[i];
1507 		kfree(conf->disks);
1508 		conf->disks = ndisks;
1509 	} else
1510 		err = -ENOMEM;
1511 
1512 	get_online_cpus();
1513 	conf->scribble_len = scribble_len(newsize);
1514 	for_each_present_cpu(cpu) {
1515 		struct raid5_percpu *percpu;
1516 		void *scribble;
1517 
1518 		percpu = per_cpu_ptr(conf->percpu, cpu);
1519 		scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1520 
1521 		if (scribble) {
1522 			kfree(percpu->scribble);
1523 			percpu->scribble = scribble;
1524 		} else {
1525 			err = -ENOMEM;
1526 			break;
1527 		}
1528 	}
1529 	put_online_cpus();
1530 
1531 	/* Step 4, return new stripes to service */
1532 	while(!list_empty(&newstripes)) {
1533 		nsh = list_entry(newstripes.next, struct stripe_head, lru);
1534 		list_del_init(&nsh->lru);
1535 
1536 		for (i=conf->raid_disks; i < newsize; i++)
1537 			if (nsh->dev[i].page == NULL) {
1538 				struct page *p = alloc_page(GFP_NOIO);
1539 				nsh->dev[i].page = p;
1540 				if (!p)
1541 					err = -ENOMEM;
1542 			}
1543 		release_stripe(nsh);
1544 	}
1545 	/* critical section pass, GFP_NOIO no longer needed */
1546 
1547 	conf->slab_cache = sc;
1548 	conf->active_name = 1-conf->active_name;
1549 	conf->pool_size = newsize;
1550 	return err;
1551 }
1552 
1553 static int drop_one_stripe(struct r5conf *conf)
1554 {
1555 	struct stripe_head *sh;
1556 
1557 	spin_lock_irq(&conf->device_lock);
1558 	sh = get_free_stripe(conf);
1559 	spin_unlock_irq(&conf->device_lock);
1560 	if (!sh)
1561 		return 0;
1562 	BUG_ON(atomic_read(&sh->count));
1563 	shrink_buffers(sh);
1564 	kmem_cache_free(conf->slab_cache, sh);
1565 	atomic_dec(&conf->active_stripes);
1566 	return 1;
1567 }
1568 
1569 static void shrink_stripes(struct r5conf *conf)
1570 {
1571 	while (drop_one_stripe(conf))
1572 		;
1573 
1574 	if (conf->slab_cache)
1575 		kmem_cache_destroy(conf->slab_cache);
1576 	conf->slab_cache = NULL;
1577 }
1578 
1579 static void raid5_end_read_request(struct bio * bi, int error)
1580 {
1581 	struct stripe_head *sh = bi->bi_private;
1582 	struct r5conf *conf = sh->raid_conf;
1583 	int disks = sh->disks, i;
1584 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1585 	char b[BDEVNAME_SIZE];
1586 	struct md_rdev *rdev;
1587 
1588 
1589 	for (i=0 ; i<disks; i++)
1590 		if (bi == &sh->dev[i].req)
1591 			break;
1592 
1593 	pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1594 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
1595 		uptodate);
1596 	if (i == disks) {
1597 		BUG();
1598 		return;
1599 	}
1600 
1601 	if (uptodate) {
1602 		set_bit(R5_UPTODATE, &sh->dev[i].flags);
1603 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1604 			rdev = conf->disks[i].rdev;
1605 			printk_ratelimited(
1606 				KERN_INFO
1607 				"md/raid:%s: read error corrected"
1608 				" (%lu sectors at %llu on %s)\n",
1609 				mdname(conf->mddev), STRIPE_SECTORS,
1610 				(unsigned long long)(sh->sector
1611 						     + rdev->data_offset),
1612 				bdevname(rdev->bdev, b));
1613 			atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
1614 			clear_bit(R5_ReadError, &sh->dev[i].flags);
1615 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
1616 		}
1617 		if (atomic_read(&conf->disks[i].rdev->read_errors))
1618 			atomic_set(&conf->disks[i].rdev->read_errors, 0);
1619 	} else {
1620 		const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1621 		int retry = 0;
1622 		rdev = conf->disks[i].rdev;
1623 
1624 		clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1625 		atomic_inc(&rdev->read_errors);
1626 		if (conf->mddev->degraded >= conf->max_degraded)
1627 			printk_ratelimited(
1628 				KERN_WARNING
1629 				"md/raid:%s: read error not correctable "
1630 				"(sector %llu on %s).\n",
1631 				mdname(conf->mddev),
1632 				(unsigned long long)(sh->sector
1633 						     + rdev->data_offset),
1634 				bdn);
1635 		else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1636 			/* Oh, no!!! */
1637 			printk_ratelimited(
1638 				KERN_WARNING
1639 				"md/raid:%s: read error NOT corrected!! "
1640 				"(sector %llu on %s).\n",
1641 				mdname(conf->mddev),
1642 				(unsigned long long)(sh->sector
1643 						     + rdev->data_offset),
1644 				bdn);
1645 		else if (atomic_read(&rdev->read_errors)
1646 			 > conf->max_nr_stripes)
1647 			printk(KERN_WARNING
1648 			       "md/raid:%s: Too many read errors, failing device %s.\n",
1649 			       mdname(conf->mddev), bdn);
1650 		else
1651 			retry = 1;
1652 		if (retry)
1653 			set_bit(R5_ReadError, &sh->dev[i].flags);
1654 		else {
1655 			clear_bit(R5_ReadError, &sh->dev[i].flags);
1656 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
1657 			md_error(conf->mddev, rdev);
1658 		}
1659 	}
1660 	rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1661 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
1662 	set_bit(STRIPE_HANDLE, &sh->state);
1663 	release_stripe(sh);
1664 }
1665 
1666 static void raid5_end_write_request(struct bio *bi, int error)
1667 {
1668 	struct stripe_head *sh = bi->bi_private;
1669 	struct r5conf *conf = sh->raid_conf;
1670 	int disks = sh->disks, i;
1671 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1672 	sector_t first_bad;
1673 	int bad_sectors;
1674 
1675 	for (i=0 ; i<disks; i++)
1676 		if (bi == &sh->dev[i].req)
1677 			break;
1678 
1679 	pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1680 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
1681 		uptodate);
1682 	if (i == disks) {
1683 		BUG();
1684 		return;
1685 	}
1686 
1687 	if (!uptodate) {
1688 		set_bit(WriteErrorSeen, &conf->disks[i].rdev->flags);
1689 		set_bit(R5_WriteError, &sh->dev[i].flags);
1690 	} else if (is_badblock(conf->disks[i].rdev, sh->sector, STRIPE_SECTORS,
1691 			       &first_bad, &bad_sectors))
1692 		set_bit(R5_MadeGood, &sh->dev[i].flags);
1693 
1694 	rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1695 
1696 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
1697 	set_bit(STRIPE_HANDLE, &sh->state);
1698 	release_stripe(sh);
1699 }
1700 
1701 
1702 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1703 
1704 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1705 {
1706 	struct r5dev *dev = &sh->dev[i];
1707 
1708 	bio_init(&dev->req);
1709 	dev->req.bi_io_vec = &dev->vec;
1710 	dev->req.bi_vcnt++;
1711 	dev->req.bi_max_vecs++;
1712 	dev->vec.bv_page = dev->page;
1713 	dev->vec.bv_len = STRIPE_SIZE;
1714 	dev->vec.bv_offset = 0;
1715 
1716 	dev->req.bi_sector = sh->sector;
1717 	dev->req.bi_private = sh;
1718 
1719 	dev->flags = 0;
1720 	dev->sector = compute_blocknr(sh, i, previous);
1721 }
1722 
1723 static void error(struct mddev *mddev, struct md_rdev *rdev)
1724 {
1725 	char b[BDEVNAME_SIZE];
1726 	struct r5conf *conf = mddev->private;
1727 	pr_debug("raid456: error called\n");
1728 
1729 	if (test_and_clear_bit(In_sync, &rdev->flags)) {
1730 		unsigned long flags;
1731 		spin_lock_irqsave(&conf->device_lock, flags);
1732 		mddev->degraded++;
1733 		spin_unlock_irqrestore(&conf->device_lock, flags);
1734 		/*
1735 		 * if recovery was running, make sure it aborts.
1736 		 */
1737 		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1738 	}
1739 	set_bit(Blocked, &rdev->flags);
1740 	set_bit(Faulty, &rdev->flags);
1741 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
1742 	printk(KERN_ALERT
1743 	       "md/raid:%s: Disk failure on %s, disabling device.\n"
1744 	       "md/raid:%s: Operation continuing on %d devices.\n",
1745 	       mdname(mddev),
1746 	       bdevname(rdev->bdev, b),
1747 	       mdname(mddev),
1748 	       conf->raid_disks - mddev->degraded);
1749 }
1750 
1751 /*
1752  * Input: a 'big' sector number,
1753  * Output: index of the data and parity disk, and the sector # in them.
1754  */
1755 static sector_t raid5_compute_sector(struct r5conf *conf, sector_t r_sector,
1756 				     int previous, int *dd_idx,
1757 				     struct stripe_head *sh)
1758 {
1759 	sector_t stripe, stripe2;
1760 	sector_t chunk_number;
1761 	unsigned int chunk_offset;
1762 	int pd_idx, qd_idx;
1763 	int ddf_layout = 0;
1764 	sector_t new_sector;
1765 	int algorithm = previous ? conf->prev_algo
1766 				 : conf->algorithm;
1767 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1768 					 : conf->chunk_sectors;
1769 	int raid_disks = previous ? conf->previous_raid_disks
1770 				  : conf->raid_disks;
1771 	int data_disks = raid_disks - conf->max_degraded;
1772 
1773 	/* First compute the information on this sector */
1774 
1775 	/*
1776 	 * Compute the chunk number and the sector offset inside the chunk
1777 	 */
1778 	chunk_offset = sector_div(r_sector, sectors_per_chunk);
1779 	chunk_number = r_sector;
1780 
1781 	/*
1782 	 * Compute the stripe number
1783 	 */
1784 	stripe = chunk_number;
1785 	*dd_idx = sector_div(stripe, data_disks);
1786 	stripe2 = stripe;
1787 	/*
1788 	 * Select the parity disk based on the user selected algorithm.
1789 	 */
1790 	pd_idx = qd_idx = -1;
1791 	switch(conf->level) {
1792 	case 4:
1793 		pd_idx = data_disks;
1794 		break;
1795 	case 5:
1796 		switch (algorithm) {
1797 		case ALGORITHM_LEFT_ASYMMETRIC:
1798 			pd_idx = data_disks - sector_div(stripe2, raid_disks);
1799 			if (*dd_idx >= pd_idx)
1800 				(*dd_idx)++;
1801 			break;
1802 		case ALGORITHM_RIGHT_ASYMMETRIC:
1803 			pd_idx = sector_div(stripe2, raid_disks);
1804 			if (*dd_idx >= pd_idx)
1805 				(*dd_idx)++;
1806 			break;
1807 		case ALGORITHM_LEFT_SYMMETRIC:
1808 			pd_idx = data_disks - sector_div(stripe2, raid_disks);
1809 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1810 			break;
1811 		case ALGORITHM_RIGHT_SYMMETRIC:
1812 			pd_idx = sector_div(stripe2, raid_disks);
1813 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1814 			break;
1815 		case ALGORITHM_PARITY_0:
1816 			pd_idx = 0;
1817 			(*dd_idx)++;
1818 			break;
1819 		case ALGORITHM_PARITY_N:
1820 			pd_idx = data_disks;
1821 			break;
1822 		default:
1823 			BUG();
1824 		}
1825 		break;
1826 	case 6:
1827 
1828 		switch (algorithm) {
1829 		case ALGORITHM_LEFT_ASYMMETRIC:
1830 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1831 			qd_idx = pd_idx + 1;
1832 			if (pd_idx == raid_disks-1) {
1833 				(*dd_idx)++;	/* Q D D D P */
1834 				qd_idx = 0;
1835 			} else if (*dd_idx >= pd_idx)
1836 				(*dd_idx) += 2; /* D D P Q D */
1837 			break;
1838 		case ALGORITHM_RIGHT_ASYMMETRIC:
1839 			pd_idx = sector_div(stripe2, raid_disks);
1840 			qd_idx = pd_idx + 1;
1841 			if (pd_idx == raid_disks-1) {
1842 				(*dd_idx)++;	/* Q D D D P */
1843 				qd_idx = 0;
1844 			} else if (*dd_idx >= pd_idx)
1845 				(*dd_idx) += 2; /* D D P Q D */
1846 			break;
1847 		case ALGORITHM_LEFT_SYMMETRIC:
1848 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1849 			qd_idx = (pd_idx + 1) % raid_disks;
1850 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1851 			break;
1852 		case ALGORITHM_RIGHT_SYMMETRIC:
1853 			pd_idx = sector_div(stripe2, raid_disks);
1854 			qd_idx = (pd_idx + 1) % raid_disks;
1855 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1856 			break;
1857 
1858 		case ALGORITHM_PARITY_0:
1859 			pd_idx = 0;
1860 			qd_idx = 1;
1861 			(*dd_idx) += 2;
1862 			break;
1863 		case ALGORITHM_PARITY_N:
1864 			pd_idx = data_disks;
1865 			qd_idx = data_disks + 1;
1866 			break;
1867 
1868 		case ALGORITHM_ROTATING_ZERO_RESTART:
1869 			/* Exactly the same as RIGHT_ASYMMETRIC, but or
1870 			 * of blocks for computing Q is different.
1871 			 */
1872 			pd_idx = sector_div(stripe2, raid_disks);
1873 			qd_idx = pd_idx + 1;
1874 			if (pd_idx == raid_disks-1) {
1875 				(*dd_idx)++;	/* Q D D D P */
1876 				qd_idx = 0;
1877 			} else if (*dd_idx >= pd_idx)
1878 				(*dd_idx) += 2; /* D D P Q D */
1879 			ddf_layout = 1;
1880 			break;
1881 
1882 		case ALGORITHM_ROTATING_N_RESTART:
1883 			/* Same a left_asymmetric, by first stripe is
1884 			 * D D D P Q  rather than
1885 			 * Q D D D P
1886 			 */
1887 			stripe2 += 1;
1888 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1889 			qd_idx = pd_idx + 1;
1890 			if (pd_idx == raid_disks-1) {
1891 				(*dd_idx)++;	/* Q D D D P */
1892 				qd_idx = 0;
1893 			} else if (*dd_idx >= pd_idx)
1894 				(*dd_idx) += 2; /* D D P Q D */
1895 			ddf_layout = 1;
1896 			break;
1897 
1898 		case ALGORITHM_ROTATING_N_CONTINUE:
1899 			/* Same as left_symmetric but Q is before P */
1900 			pd_idx = raid_disks - 1 - sector_div(stripe2, raid_disks);
1901 			qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1902 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1903 			ddf_layout = 1;
1904 			break;
1905 
1906 		case ALGORITHM_LEFT_ASYMMETRIC_6:
1907 			/* RAID5 left_asymmetric, with Q on last device */
1908 			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1909 			if (*dd_idx >= pd_idx)
1910 				(*dd_idx)++;
1911 			qd_idx = raid_disks - 1;
1912 			break;
1913 
1914 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
1915 			pd_idx = sector_div(stripe2, raid_disks-1);
1916 			if (*dd_idx >= pd_idx)
1917 				(*dd_idx)++;
1918 			qd_idx = raid_disks - 1;
1919 			break;
1920 
1921 		case ALGORITHM_LEFT_SYMMETRIC_6:
1922 			pd_idx = data_disks - sector_div(stripe2, raid_disks-1);
1923 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1924 			qd_idx = raid_disks - 1;
1925 			break;
1926 
1927 		case ALGORITHM_RIGHT_SYMMETRIC_6:
1928 			pd_idx = sector_div(stripe2, raid_disks-1);
1929 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1930 			qd_idx = raid_disks - 1;
1931 			break;
1932 
1933 		case ALGORITHM_PARITY_0_6:
1934 			pd_idx = 0;
1935 			(*dd_idx)++;
1936 			qd_idx = raid_disks - 1;
1937 			break;
1938 
1939 		default:
1940 			BUG();
1941 		}
1942 		break;
1943 	}
1944 
1945 	if (sh) {
1946 		sh->pd_idx = pd_idx;
1947 		sh->qd_idx = qd_idx;
1948 		sh->ddf_layout = ddf_layout;
1949 	}
1950 	/*
1951 	 * Finally, compute the new sector number
1952 	 */
1953 	new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1954 	return new_sector;
1955 }
1956 
1957 
1958 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1959 {
1960 	struct r5conf *conf = sh->raid_conf;
1961 	int raid_disks = sh->disks;
1962 	int data_disks = raid_disks - conf->max_degraded;
1963 	sector_t new_sector = sh->sector, check;
1964 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1965 					 : conf->chunk_sectors;
1966 	int algorithm = previous ? conf->prev_algo
1967 				 : conf->algorithm;
1968 	sector_t stripe;
1969 	int chunk_offset;
1970 	sector_t chunk_number;
1971 	int dummy1, dd_idx = i;
1972 	sector_t r_sector;
1973 	struct stripe_head sh2;
1974 
1975 
1976 	chunk_offset = sector_div(new_sector, sectors_per_chunk);
1977 	stripe = new_sector;
1978 
1979 	if (i == sh->pd_idx)
1980 		return 0;
1981 	switch(conf->level) {
1982 	case 4: break;
1983 	case 5:
1984 		switch (algorithm) {
1985 		case ALGORITHM_LEFT_ASYMMETRIC:
1986 		case ALGORITHM_RIGHT_ASYMMETRIC:
1987 			if (i > sh->pd_idx)
1988 				i--;
1989 			break;
1990 		case ALGORITHM_LEFT_SYMMETRIC:
1991 		case ALGORITHM_RIGHT_SYMMETRIC:
1992 			if (i < sh->pd_idx)
1993 				i += raid_disks;
1994 			i -= (sh->pd_idx + 1);
1995 			break;
1996 		case ALGORITHM_PARITY_0:
1997 			i -= 1;
1998 			break;
1999 		case ALGORITHM_PARITY_N:
2000 			break;
2001 		default:
2002 			BUG();
2003 		}
2004 		break;
2005 	case 6:
2006 		if (i == sh->qd_idx)
2007 			return 0; /* It is the Q disk */
2008 		switch (algorithm) {
2009 		case ALGORITHM_LEFT_ASYMMETRIC:
2010 		case ALGORITHM_RIGHT_ASYMMETRIC:
2011 		case ALGORITHM_ROTATING_ZERO_RESTART:
2012 		case ALGORITHM_ROTATING_N_RESTART:
2013 			if (sh->pd_idx == raid_disks-1)
2014 				i--;	/* Q D D D P */
2015 			else if (i > sh->pd_idx)
2016 				i -= 2; /* D D P Q D */
2017 			break;
2018 		case ALGORITHM_LEFT_SYMMETRIC:
2019 		case ALGORITHM_RIGHT_SYMMETRIC:
2020 			if (sh->pd_idx == raid_disks-1)
2021 				i--; /* Q D D D P */
2022 			else {
2023 				/* D D P Q D */
2024 				if (i < sh->pd_idx)
2025 					i += raid_disks;
2026 				i -= (sh->pd_idx + 2);
2027 			}
2028 			break;
2029 		case ALGORITHM_PARITY_0:
2030 			i -= 2;
2031 			break;
2032 		case ALGORITHM_PARITY_N:
2033 			break;
2034 		case ALGORITHM_ROTATING_N_CONTINUE:
2035 			/* Like left_symmetric, but P is before Q */
2036 			if (sh->pd_idx == 0)
2037 				i--;	/* P D D D Q */
2038 			else {
2039 				/* D D Q P D */
2040 				if (i < sh->pd_idx)
2041 					i += raid_disks;
2042 				i -= (sh->pd_idx + 1);
2043 			}
2044 			break;
2045 		case ALGORITHM_LEFT_ASYMMETRIC_6:
2046 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
2047 			if (i > sh->pd_idx)
2048 				i--;
2049 			break;
2050 		case ALGORITHM_LEFT_SYMMETRIC_6:
2051 		case ALGORITHM_RIGHT_SYMMETRIC_6:
2052 			if (i < sh->pd_idx)
2053 				i += data_disks + 1;
2054 			i -= (sh->pd_idx + 1);
2055 			break;
2056 		case ALGORITHM_PARITY_0_6:
2057 			i -= 1;
2058 			break;
2059 		default:
2060 			BUG();
2061 		}
2062 		break;
2063 	}
2064 
2065 	chunk_number = stripe * data_disks + i;
2066 	r_sector = chunk_number * sectors_per_chunk + chunk_offset;
2067 
2068 	check = raid5_compute_sector(conf, r_sector,
2069 				     previous, &dummy1, &sh2);
2070 	if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
2071 		|| sh2.qd_idx != sh->qd_idx) {
2072 		printk(KERN_ERR "md/raid:%s: compute_blocknr: map not correct\n",
2073 		       mdname(conf->mddev));
2074 		return 0;
2075 	}
2076 	return r_sector;
2077 }
2078 
2079 
2080 static void
2081 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
2082 			 int rcw, int expand)
2083 {
2084 	int i, pd_idx = sh->pd_idx, disks = sh->disks;
2085 	struct r5conf *conf = sh->raid_conf;
2086 	int level = conf->level;
2087 
2088 	if (rcw) {
2089 		/* if we are not expanding this is a proper write request, and
2090 		 * there will be bios with new data to be drained into the
2091 		 * stripe cache
2092 		 */
2093 		if (!expand) {
2094 			sh->reconstruct_state = reconstruct_state_drain_run;
2095 			set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2096 		} else
2097 			sh->reconstruct_state = reconstruct_state_run;
2098 
2099 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2100 
2101 		for (i = disks; i--; ) {
2102 			struct r5dev *dev = &sh->dev[i];
2103 
2104 			if (dev->towrite) {
2105 				set_bit(R5_LOCKED, &dev->flags);
2106 				set_bit(R5_Wantdrain, &dev->flags);
2107 				if (!expand)
2108 					clear_bit(R5_UPTODATE, &dev->flags);
2109 				s->locked++;
2110 			}
2111 		}
2112 		if (s->locked + conf->max_degraded == disks)
2113 			if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2114 				atomic_inc(&conf->pending_full_writes);
2115 	} else {
2116 		BUG_ON(level == 6);
2117 		BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
2118 			test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
2119 
2120 		sh->reconstruct_state = reconstruct_state_prexor_drain_run;
2121 		set_bit(STRIPE_OP_PREXOR, &s->ops_request);
2122 		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2123 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
2124 
2125 		for (i = disks; i--; ) {
2126 			struct r5dev *dev = &sh->dev[i];
2127 			if (i == pd_idx)
2128 				continue;
2129 
2130 			if (dev->towrite &&
2131 			    (test_bit(R5_UPTODATE, &dev->flags) ||
2132 			     test_bit(R5_Wantcompute, &dev->flags))) {
2133 				set_bit(R5_Wantdrain, &dev->flags);
2134 				set_bit(R5_LOCKED, &dev->flags);
2135 				clear_bit(R5_UPTODATE, &dev->flags);
2136 				s->locked++;
2137 			}
2138 		}
2139 	}
2140 
2141 	/* keep the parity disk(s) locked while asynchronous operations
2142 	 * are in flight
2143 	 */
2144 	set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2145 	clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2146 	s->locked++;
2147 
2148 	if (level == 6) {
2149 		int qd_idx = sh->qd_idx;
2150 		struct r5dev *dev = &sh->dev[qd_idx];
2151 
2152 		set_bit(R5_LOCKED, &dev->flags);
2153 		clear_bit(R5_UPTODATE, &dev->flags);
2154 		s->locked++;
2155 	}
2156 
2157 	pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2158 		__func__, (unsigned long long)sh->sector,
2159 		s->locked, s->ops_request);
2160 }
2161 
2162 /*
2163  * Each stripe/dev can have one or more bion attached.
2164  * toread/towrite point to the first in a chain.
2165  * The bi_next chain must be in order.
2166  */
2167 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2168 {
2169 	struct bio **bip;
2170 	struct r5conf *conf = sh->raid_conf;
2171 	int firstwrite=0;
2172 
2173 	pr_debug("adding bi b#%llu to stripe s#%llu\n",
2174 		(unsigned long long)bi->bi_sector,
2175 		(unsigned long long)sh->sector);
2176 
2177 
2178 	spin_lock_irq(&conf->device_lock);
2179 	if (forwrite) {
2180 		bip = &sh->dev[dd_idx].towrite;
2181 		if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2182 			firstwrite = 1;
2183 	} else
2184 		bip = &sh->dev[dd_idx].toread;
2185 	while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2186 		if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2187 			goto overlap;
2188 		bip = & (*bip)->bi_next;
2189 	}
2190 	if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2191 		goto overlap;
2192 
2193 	BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2194 	if (*bip)
2195 		bi->bi_next = *bip;
2196 	*bip = bi;
2197 	bi->bi_phys_segments++;
2198 
2199 	if (forwrite) {
2200 		/* check if page is covered */
2201 		sector_t sector = sh->dev[dd_idx].sector;
2202 		for (bi=sh->dev[dd_idx].towrite;
2203 		     sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2204 			     bi && bi->bi_sector <= sector;
2205 		     bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2206 			if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2207 				sector = bi->bi_sector + (bi->bi_size>>9);
2208 		}
2209 		if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2210 			set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2211 	}
2212 	spin_unlock_irq(&conf->device_lock);
2213 
2214 	pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2215 		(unsigned long long)(*bip)->bi_sector,
2216 		(unsigned long long)sh->sector, dd_idx);
2217 
2218 	if (conf->mddev->bitmap && firstwrite) {
2219 		bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2220 				  STRIPE_SECTORS, 0);
2221 		sh->bm_seq = conf->seq_flush+1;
2222 		set_bit(STRIPE_BIT_DELAY, &sh->state);
2223 	}
2224 	return 1;
2225 
2226  overlap:
2227 	set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2228 	spin_unlock_irq(&conf->device_lock);
2229 	return 0;
2230 }
2231 
2232 static void end_reshape(struct r5conf *conf);
2233 
2234 static void stripe_set_idx(sector_t stripe, struct r5conf *conf, int previous,
2235 			    struct stripe_head *sh)
2236 {
2237 	int sectors_per_chunk =
2238 		previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2239 	int dd_idx;
2240 	int chunk_offset = sector_div(stripe, sectors_per_chunk);
2241 	int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2242 
2243 	raid5_compute_sector(conf,
2244 			     stripe * (disks - conf->max_degraded)
2245 			     *sectors_per_chunk + chunk_offset,
2246 			     previous,
2247 			     &dd_idx, sh);
2248 }
2249 
2250 static void
2251 handle_failed_stripe(struct r5conf *conf, struct stripe_head *sh,
2252 				struct stripe_head_state *s, int disks,
2253 				struct bio **return_bi)
2254 {
2255 	int i;
2256 	for (i = disks; i--; ) {
2257 		struct bio *bi;
2258 		int bitmap_end = 0;
2259 
2260 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2261 			struct md_rdev *rdev;
2262 			rcu_read_lock();
2263 			rdev = rcu_dereference(conf->disks[i].rdev);
2264 			if (rdev && test_bit(In_sync, &rdev->flags))
2265 				atomic_inc(&rdev->nr_pending);
2266 			else
2267 				rdev = NULL;
2268 			rcu_read_unlock();
2269 			if (rdev) {
2270 				if (!rdev_set_badblocks(
2271 					    rdev,
2272 					    sh->sector,
2273 					    STRIPE_SECTORS, 0))
2274 					md_error(conf->mddev, rdev);
2275 				rdev_dec_pending(rdev, conf->mddev);
2276 			}
2277 		}
2278 		spin_lock_irq(&conf->device_lock);
2279 		/* fail all writes first */
2280 		bi = sh->dev[i].towrite;
2281 		sh->dev[i].towrite = NULL;
2282 		if (bi) {
2283 			s->to_write--;
2284 			bitmap_end = 1;
2285 		}
2286 
2287 		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2288 			wake_up(&conf->wait_for_overlap);
2289 
2290 		while (bi && bi->bi_sector <
2291 			sh->dev[i].sector + STRIPE_SECTORS) {
2292 			struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2293 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
2294 			if (!raid5_dec_bi_phys_segments(bi)) {
2295 				md_write_end(conf->mddev);
2296 				bi->bi_next = *return_bi;
2297 				*return_bi = bi;
2298 			}
2299 			bi = nextbi;
2300 		}
2301 		/* and fail all 'written' */
2302 		bi = sh->dev[i].written;
2303 		sh->dev[i].written = NULL;
2304 		if (bi) bitmap_end = 1;
2305 		while (bi && bi->bi_sector <
2306 		       sh->dev[i].sector + STRIPE_SECTORS) {
2307 			struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2308 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
2309 			if (!raid5_dec_bi_phys_segments(bi)) {
2310 				md_write_end(conf->mddev);
2311 				bi->bi_next = *return_bi;
2312 				*return_bi = bi;
2313 			}
2314 			bi = bi2;
2315 		}
2316 
2317 		/* fail any reads if this device is non-operational and
2318 		 * the data has not reached the cache yet.
2319 		 */
2320 		if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2321 		    (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2322 		      test_bit(R5_ReadError, &sh->dev[i].flags))) {
2323 			bi = sh->dev[i].toread;
2324 			sh->dev[i].toread = NULL;
2325 			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2326 				wake_up(&conf->wait_for_overlap);
2327 			if (bi) s->to_read--;
2328 			while (bi && bi->bi_sector <
2329 			       sh->dev[i].sector + STRIPE_SECTORS) {
2330 				struct bio *nextbi =
2331 					r5_next_bio(bi, sh->dev[i].sector);
2332 				clear_bit(BIO_UPTODATE, &bi->bi_flags);
2333 				if (!raid5_dec_bi_phys_segments(bi)) {
2334 					bi->bi_next = *return_bi;
2335 					*return_bi = bi;
2336 				}
2337 				bi = nextbi;
2338 			}
2339 		}
2340 		spin_unlock_irq(&conf->device_lock);
2341 		if (bitmap_end)
2342 			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2343 					STRIPE_SECTORS, 0, 0);
2344 		/* If we were in the middle of a write the parity block might
2345 		 * still be locked - so just clear all R5_LOCKED flags
2346 		 */
2347 		clear_bit(R5_LOCKED, &sh->dev[i].flags);
2348 	}
2349 
2350 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2351 		if (atomic_dec_and_test(&conf->pending_full_writes))
2352 			md_wakeup_thread(conf->mddev->thread);
2353 }
2354 
2355 static void
2356 handle_failed_sync(struct r5conf *conf, struct stripe_head *sh,
2357 		   struct stripe_head_state *s)
2358 {
2359 	int abort = 0;
2360 	int i;
2361 
2362 	md_done_sync(conf->mddev, STRIPE_SECTORS, 0);
2363 	clear_bit(STRIPE_SYNCING, &sh->state);
2364 	s->syncing = 0;
2365 	/* There is nothing more to do for sync/check/repair.
2366 	 * For recover we need to record a bad block on all
2367 	 * non-sync devices, or abort the recovery
2368 	 */
2369 	if (!test_bit(MD_RECOVERY_RECOVER, &conf->mddev->recovery))
2370 		return;
2371 	/* During recovery devices cannot be removed, so locking and
2372 	 * refcounting of rdevs is not needed
2373 	 */
2374 	for (i = 0; i < conf->raid_disks; i++) {
2375 		struct md_rdev *rdev = conf->disks[i].rdev;
2376 		if (!rdev
2377 		    || test_bit(Faulty, &rdev->flags)
2378 		    || test_bit(In_sync, &rdev->flags))
2379 			continue;
2380 		if (!rdev_set_badblocks(rdev, sh->sector,
2381 					STRIPE_SECTORS, 0))
2382 			abort = 1;
2383 	}
2384 	if (abort) {
2385 		conf->recovery_disabled = conf->mddev->recovery_disabled;
2386 		set_bit(MD_RECOVERY_INTR, &conf->mddev->recovery);
2387 	}
2388 }
2389 
2390 /* fetch_block - checks the given member device to see if its data needs
2391  * to be read or computed to satisfy a request.
2392  *
2393  * Returns 1 when no more member devices need to be checked, otherwise returns
2394  * 0 to tell the loop in handle_stripe_fill to continue
2395  */
2396 static int fetch_block(struct stripe_head *sh, struct stripe_head_state *s,
2397 		       int disk_idx, int disks)
2398 {
2399 	struct r5dev *dev = &sh->dev[disk_idx];
2400 	struct r5dev *fdev[2] = { &sh->dev[s->failed_num[0]],
2401 				  &sh->dev[s->failed_num[1]] };
2402 
2403 	/* is the data in this block needed, and can we get it? */
2404 	if (!test_bit(R5_LOCKED, &dev->flags) &&
2405 	    !test_bit(R5_UPTODATE, &dev->flags) &&
2406 	    (dev->toread ||
2407 	     (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2408 	     s->syncing || s->expanding ||
2409 	     (s->failed >= 1 && fdev[0]->toread) ||
2410 	     (s->failed >= 2 && fdev[1]->toread) ||
2411 	     (sh->raid_conf->level <= 5 && s->failed && fdev[0]->towrite &&
2412 	      !test_bit(R5_OVERWRITE, &fdev[0]->flags)) ||
2413 	     (sh->raid_conf->level == 6 && s->failed && s->to_write))) {
2414 		/* we would like to get this block, possibly by computing it,
2415 		 * otherwise read it if the backing disk is insync
2416 		 */
2417 		BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2418 		BUG_ON(test_bit(R5_Wantread, &dev->flags));
2419 		if ((s->uptodate == disks - 1) &&
2420 		    (s->failed && (disk_idx == s->failed_num[0] ||
2421 				   disk_idx == s->failed_num[1]))) {
2422 			/* have disk failed, and we're requested to fetch it;
2423 			 * do compute it
2424 			 */
2425 			pr_debug("Computing stripe %llu block %d\n",
2426 			       (unsigned long long)sh->sector, disk_idx);
2427 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2428 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2429 			set_bit(R5_Wantcompute, &dev->flags);
2430 			sh->ops.target = disk_idx;
2431 			sh->ops.target2 = -1; /* no 2nd target */
2432 			s->req_compute = 1;
2433 			/* Careful: from this point on 'uptodate' is in the eye
2434 			 * of raid_run_ops which services 'compute' operations
2435 			 * before writes. R5_Wantcompute flags a block that will
2436 			 * be R5_UPTODATE by the time it is needed for a
2437 			 * subsequent operation.
2438 			 */
2439 			s->uptodate++;
2440 			return 1;
2441 		} else if (s->uptodate == disks-2 && s->failed >= 2) {
2442 			/* Computing 2-failure is *very* expensive; only
2443 			 * do it if failed >= 2
2444 			 */
2445 			int other;
2446 			for (other = disks; other--; ) {
2447 				if (other == disk_idx)
2448 					continue;
2449 				if (!test_bit(R5_UPTODATE,
2450 				      &sh->dev[other].flags))
2451 					break;
2452 			}
2453 			BUG_ON(other < 0);
2454 			pr_debug("Computing stripe %llu blocks %d,%d\n",
2455 			       (unsigned long long)sh->sector,
2456 			       disk_idx, other);
2457 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2458 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2459 			set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2460 			set_bit(R5_Wantcompute, &sh->dev[other].flags);
2461 			sh->ops.target = disk_idx;
2462 			sh->ops.target2 = other;
2463 			s->uptodate += 2;
2464 			s->req_compute = 1;
2465 			return 1;
2466 		} else if (test_bit(R5_Insync, &dev->flags)) {
2467 			set_bit(R5_LOCKED, &dev->flags);
2468 			set_bit(R5_Wantread, &dev->flags);
2469 			s->locked++;
2470 			pr_debug("Reading block %d (sync=%d)\n",
2471 				disk_idx, s->syncing);
2472 		}
2473 	}
2474 
2475 	return 0;
2476 }
2477 
2478 /**
2479  * handle_stripe_fill - read or compute data to satisfy pending requests.
2480  */
2481 static void handle_stripe_fill(struct stripe_head *sh,
2482 			       struct stripe_head_state *s,
2483 			       int disks)
2484 {
2485 	int i;
2486 
2487 	/* look for blocks to read/compute, skip this if a compute
2488 	 * is already in flight, or if the stripe contents are in the
2489 	 * midst of changing due to a write
2490 	 */
2491 	if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2492 	    !sh->reconstruct_state)
2493 		for (i = disks; i--; )
2494 			if (fetch_block(sh, s, i, disks))
2495 				break;
2496 	set_bit(STRIPE_HANDLE, &sh->state);
2497 }
2498 
2499 
2500 /* handle_stripe_clean_event
2501  * any written block on an uptodate or failed drive can be returned.
2502  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2503  * never LOCKED, so we don't need to test 'failed' directly.
2504  */
2505 static void handle_stripe_clean_event(struct r5conf *conf,
2506 	struct stripe_head *sh, int disks, struct bio **return_bi)
2507 {
2508 	int i;
2509 	struct r5dev *dev;
2510 
2511 	for (i = disks; i--; )
2512 		if (sh->dev[i].written) {
2513 			dev = &sh->dev[i];
2514 			if (!test_bit(R5_LOCKED, &dev->flags) &&
2515 				test_bit(R5_UPTODATE, &dev->flags)) {
2516 				/* We can return any write requests */
2517 				struct bio *wbi, *wbi2;
2518 				int bitmap_end = 0;
2519 				pr_debug("Return write for disc %d\n", i);
2520 				spin_lock_irq(&conf->device_lock);
2521 				wbi = dev->written;
2522 				dev->written = NULL;
2523 				while (wbi && wbi->bi_sector <
2524 					dev->sector + STRIPE_SECTORS) {
2525 					wbi2 = r5_next_bio(wbi, dev->sector);
2526 					if (!raid5_dec_bi_phys_segments(wbi)) {
2527 						md_write_end(conf->mddev);
2528 						wbi->bi_next = *return_bi;
2529 						*return_bi = wbi;
2530 					}
2531 					wbi = wbi2;
2532 				}
2533 				if (dev->towrite == NULL)
2534 					bitmap_end = 1;
2535 				spin_unlock_irq(&conf->device_lock);
2536 				if (bitmap_end)
2537 					bitmap_endwrite(conf->mddev->bitmap,
2538 							sh->sector,
2539 							STRIPE_SECTORS,
2540 					 !test_bit(STRIPE_DEGRADED, &sh->state),
2541 							0);
2542 			}
2543 		}
2544 
2545 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2546 		if (atomic_dec_and_test(&conf->pending_full_writes))
2547 			md_wakeup_thread(conf->mddev->thread);
2548 }
2549 
2550 static void handle_stripe_dirtying(struct r5conf *conf,
2551 				   struct stripe_head *sh,
2552 				   struct stripe_head_state *s,
2553 				   int disks)
2554 {
2555 	int rmw = 0, rcw = 0, i;
2556 	if (conf->max_degraded == 2) {
2557 		/* RAID6 requires 'rcw' in current implementation
2558 		 * Calculate the real rcw later - for now fake it
2559 		 * look like rcw is cheaper
2560 		 */
2561 		rcw = 1; rmw = 2;
2562 	} else for (i = disks; i--; ) {
2563 		/* would I have to read this buffer for read_modify_write */
2564 		struct r5dev *dev = &sh->dev[i];
2565 		if ((dev->towrite || i == sh->pd_idx) &&
2566 		    !test_bit(R5_LOCKED, &dev->flags) &&
2567 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2568 		      test_bit(R5_Wantcompute, &dev->flags))) {
2569 			if (test_bit(R5_Insync, &dev->flags))
2570 				rmw++;
2571 			else
2572 				rmw += 2*disks;  /* cannot read it */
2573 		}
2574 		/* Would I have to read this buffer for reconstruct_write */
2575 		if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2576 		    !test_bit(R5_LOCKED, &dev->flags) &&
2577 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2578 		    test_bit(R5_Wantcompute, &dev->flags))) {
2579 			if (test_bit(R5_Insync, &dev->flags)) rcw++;
2580 			else
2581 				rcw += 2*disks;
2582 		}
2583 	}
2584 	pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2585 		(unsigned long long)sh->sector, rmw, rcw);
2586 	set_bit(STRIPE_HANDLE, &sh->state);
2587 	if (rmw < rcw && rmw > 0)
2588 		/* prefer read-modify-write, but need to get some data */
2589 		for (i = disks; i--; ) {
2590 			struct r5dev *dev = &sh->dev[i];
2591 			if ((dev->towrite || i == sh->pd_idx) &&
2592 			    !test_bit(R5_LOCKED, &dev->flags) &&
2593 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
2594 			    test_bit(R5_Wantcompute, &dev->flags)) &&
2595 			    test_bit(R5_Insync, &dev->flags)) {
2596 				if (
2597 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2598 					pr_debug("Read_old block "
2599 						"%d for r-m-w\n", i);
2600 					set_bit(R5_LOCKED, &dev->flags);
2601 					set_bit(R5_Wantread, &dev->flags);
2602 					s->locked++;
2603 				} else {
2604 					set_bit(STRIPE_DELAYED, &sh->state);
2605 					set_bit(STRIPE_HANDLE, &sh->state);
2606 				}
2607 			}
2608 		}
2609 	if (rcw <= rmw && rcw > 0) {
2610 		/* want reconstruct write, but need to get some data */
2611 		rcw = 0;
2612 		for (i = disks; i--; ) {
2613 			struct r5dev *dev = &sh->dev[i];
2614 			if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2615 			    i != sh->pd_idx && i != sh->qd_idx &&
2616 			    !test_bit(R5_LOCKED, &dev->flags) &&
2617 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
2618 			      test_bit(R5_Wantcompute, &dev->flags))) {
2619 				rcw++;
2620 				if (!test_bit(R5_Insync, &dev->flags))
2621 					continue; /* it's a failed drive */
2622 				if (
2623 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2624 					pr_debug("Read_old block "
2625 						"%d for Reconstruct\n", i);
2626 					set_bit(R5_LOCKED, &dev->flags);
2627 					set_bit(R5_Wantread, &dev->flags);
2628 					s->locked++;
2629 				} else {
2630 					set_bit(STRIPE_DELAYED, &sh->state);
2631 					set_bit(STRIPE_HANDLE, &sh->state);
2632 				}
2633 			}
2634 		}
2635 	}
2636 	/* now if nothing is locked, and if we have enough data,
2637 	 * we can start a write request
2638 	 */
2639 	/* since handle_stripe can be called at any time we need to handle the
2640 	 * case where a compute block operation has been submitted and then a
2641 	 * subsequent call wants to start a write request.  raid_run_ops only
2642 	 * handles the case where compute block and reconstruct are requested
2643 	 * simultaneously.  If this is not the case then new writes need to be
2644 	 * held off until the compute completes.
2645 	 */
2646 	if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2647 	    (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2648 	    !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2649 		schedule_reconstruction(sh, s, rcw == 0, 0);
2650 }
2651 
2652 static void handle_parity_checks5(struct r5conf *conf, struct stripe_head *sh,
2653 				struct stripe_head_state *s, int disks)
2654 {
2655 	struct r5dev *dev = NULL;
2656 
2657 	set_bit(STRIPE_HANDLE, &sh->state);
2658 
2659 	switch (sh->check_state) {
2660 	case check_state_idle:
2661 		/* start a new check operation if there are no failures */
2662 		if (s->failed == 0) {
2663 			BUG_ON(s->uptodate != disks);
2664 			sh->check_state = check_state_run;
2665 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
2666 			clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2667 			s->uptodate--;
2668 			break;
2669 		}
2670 		dev = &sh->dev[s->failed_num[0]];
2671 		/* fall through */
2672 	case check_state_compute_result:
2673 		sh->check_state = check_state_idle;
2674 		if (!dev)
2675 			dev = &sh->dev[sh->pd_idx];
2676 
2677 		/* check that a write has not made the stripe insync */
2678 		if (test_bit(STRIPE_INSYNC, &sh->state))
2679 			break;
2680 
2681 		/* either failed parity check, or recovery is happening */
2682 		BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2683 		BUG_ON(s->uptodate != disks);
2684 
2685 		set_bit(R5_LOCKED, &dev->flags);
2686 		s->locked++;
2687 		set_bit(R5_Wantwrite, &dev->flags);
2688 
2689 		clear_bit(STRIPE_DEGRADED, &sh->state);
2690 		set_bit(STRIPE_INSYNC, &sh->state);
2691 		break;
2692 	case check_state_run:
2693 		break; /* we will be called again upon completion */
2694 	case check_state_check_result:
2695 		sh->check_state = check_state_idle;
2696 
2697 		/* if a failure occurred during the check operation, leave
2698 		 * STRIPE_INSYNC not set and let the stripe be handled again
2699 		 */
2700 		if (s->failed)
2701 			break;
2702 
2703 		/* handle a successful check operation, if parity is correct
2704 		 * we are done.  Otherwise update the mismatch count and repair
2705 		 * parity if !MD_RECOVERY_CHECK
2706 		 */
2707 		if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2708 			/* parity is correct (on disc,
2709 			 * not in buffer any more)
2710 			 */
2711 			set_bit(STRIPE_INSYNC, &sh->state);
2712 		else {
2713 			conf->mddev->resync_mismatches += STRIPE_SECTORS;
2714 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2715 				/* don't try to repair!! */
2716 				set_bit(STRIPE_INSYNC, &sh->state);
2717 			else {
2718 				sh->check_state = check_state_compute_run;
2719 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2720 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2721 				set_bit(R5_Wantcompute,
2722 					&sh->dev[sh->pd_idx].flags);
2723 				sh->ops.target = sh->pd_idx;
2724 				sh->ops.target2 = -1;
2725 				s->uptodate++;
2726 			}
2727 		}
2728 		break;
2729 	case check_state_compute_run:
2730 		break;
2731 	default:
2732 		printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2733 		       __func__, sh->check_state,
2734 		       (unsigned long long) sh->sector);
2735 		BUG();
2736 	}
2737 }
2738 
2739 
2740 static void handle_parity_checks6(struct r5conf *conf, struct stripe_head *sh,
2741 				  struct stripe_head_state *s,
2742 				  int disks)
2743 {
2744 	int pd_idx = sh->pd_idx;
2745 	int qd_idx = sh->qd_idx;
2746 	struct r5dev *dev;
2747 
2748 	set_bit(STRIPE_HANDLE, &sh->state);
2749 
2750 	BUG_ON(s->failed > 2);
2751 
2752 	/* Want to check and possibly repair P and Q.
2753 	 * However there could be one 'failed' device, in which
2754 	 * case we can only check one of them, possibly using the
2755 	 * other to generate missing data
2756 	 */
2757 
2758 	switch (sh->check_state) {
2759 	case check_state_idle:
2760 		/* start a new check operation if there are < 2 failures */
2761 		if (s->failed == s->q_failed) {
2762 			/* The only possible failed device holds Q, so it
2763 			 * makes sense to check P (If anything else were failed,
2764 			 * we would have used P to recreate it).
2765 			 */
2766 			sh->check_state = check_state_run;
2767 		}
2768 		if (!s->q_failed && s->failed < 2) {
2769 			/* Q is not failed, and we didn't use it to generate
2770 			 * anything, so it makes sense to check it
2771 			 */
2772 			if (sh->check_state == check_state_run)
2773 				sh->check_state = check_state_run_pq;
2774 			else
2775 				sh->check_state = check_state_run_q;
2776 		}
2777 
2778 		/* discard potentially stale zero_sum_result */
2779 		sh->ops.zero_sum_result = 0;
2780 
2781 		if (sh->check_state == check_state_run) {
2782 			/* async_xor_zero_sum destroys the contents of P */
2783 			clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2784 			s->uptodate--;
2785 		}
2786 		if (sh->check_state >= check_state_run &&
2787 		    sh->check_state <= check_state_run_pq) {
2788 			/* async_syndrome_zero_sum preserves P and Q, so
2789 			 * no need to mark them !uptodate here
2790 			 */
2791 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
2792 			break;
2793 		}
2794 
2795 		/* we have 2-disk failure */
2796 		BUG_ON(s->failed != 2);
2797 		/* fall through */
2798 	case check_state_compute_result:
2799 		sh->check_state = check_state_idle;
2800 
2801 		/* check that a write has not made the stripe insync */
2802 		if (test_bit(STRIPE_INSYNC, &sh->state))
2803 			break;
2804 
2805 		/* now write out any block on a failed drive,
2806 		 * or P or Q if they were recomputed
2807 		 */
2808 		BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2809 		if (s->failed == 2) {
2810 			dev = &sh->dev[s->failed_num[1]];
2811 			s->locked++;
2812 			set_bit(R5_LOCKED, &dev->flags);
2813 			set_bit(R5_Wantwrite, &dev->flags);
2814 		}
2815 		if (s->failed >= 1) {
2816 			dev = &sh->dev[s->failed_num[0]];
2817 			s->locked++;
2818 			set_bit(R5_LOCKED, &dev->flags);
2819 			set_bit(R5_Wantwrite, &dev->flags);
2820 		}
2821 		if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2822 			dev = &sh->dev[pd_idx];
2823 			s->locked++;
2824 			set_bit(R5_LOCKED, &dev->flags);
2825 			set_bit(R5_Wantwrite, &dev->flags);
2826 		}
2827 		if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2828 			dev = &sh->dev[qd_idx];
2829 			s->locked++;
2830 			set_bit(R5_LOCKED, &dev->flags);
2831 			set_bit(R5_Wantwrite, &dev->flags);
2832 		}
2833 		clear_bit(STRIPE_DEGRADED, &sh->state);
2834 
2835 		set_bit(STRIPE_INSYNC, &sh->state);
2836 		break;
2837 	case check_state_run:
2838 	case check_state_run_q:
2839 	case check_state_run_pq:
2840 		break; /* we will be called again upon completion */
2841 	case check_state_check_result:
2842 		sh->check_state = check_state_idle;
2843 
2844 		/* handle a successful check operation, if parity is correct
2845 		 * we are done.  Otherwise update the mismatch count and repair
2846 		 * parity if !MD_RECOVERY_CHECK
2847 		 */
2848 		if (sh->ops.zero_sum_result == 0) {
2849 			/* both parities are correct */
2850 			if (!s->failed)
2851 				set_bit(STRIPE_INSYNC, &sh->state);
2852 			else {
2853 				/* in contrast to the raid5 case we can validate
2854 				 * parity, but still have a failure to write
2855 				 * back
2856 				 */
2857 				sh->check_state = check_state_compute_result;
2858 				/* Returning at this point means that we may go
2859 				 * off and bring p and/or q uptodate again so
2860 				 * we make sure to check zero_sum_result again
2861 				 * to verify if p or q need writeback
2862 				 */
2863 			}
2864 		} else {
2865 			conf->mddev->resync_mismatches += STRIPE_SECTORS;
2866 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2867 				/* don't try to repair!! */
2868 				set_bit(STRIPE_INSYNC, &sh->state);
2869 			else {
2870 				int *target = &sh->ops.target;
2871 
2872 				sh->ops.target = -1;
2873 				sh->ops.target2 = -1;
2874 				sh->check_state = check_state_compute_run;
2875 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2876 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2877 				if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2878 					set_bit(R5_Wantcompute,
2879 						&sh->dev[pd_idx].flags);
2880 					*target = pd_idx;
2881 					target = &sh->ops.target2;
2882 					s->uptodate++;
2883 				}
2884 				if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2885 					set_bit(R5_Wantcompute,
2886 						&sh->dev[qd_idx].flags);
2887 					*target = qd_idx;
2888 					s->uptodate++;
2889 				}
2890 			}
2891 		}
2892 		break;
2893 	case check_state_compute_run:
2894 		break;
2895 	default:
2896 		printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2897 		       __func__, sh->check_state,
2898 		       (unsigned long long) sh->sector);
2899 		BUG();
2900 	}
2901 }
2902 
2903 static void handle_stripe_expansion(struct r5conf *conf, struct stripe_head *sh)
2904 {
2905 	int i;
2906 
2907 	/* We have read all the blocks in this stripe and now we need to
2908 	 * copy some of them into a target stripe for expand.
2909 	 */
2910 	struct dma_async_tx_descriptor *tx = NULL;
2911 	clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2912 	for (i = 0; i < sh->disks; i++)
2913 		if (i != sh->pd_idx && i != sh->qd_idx) {
2914 			int dd_idx, j;
2915 			struct stripe_head *sh2;
2916 			struct async_submit_ctl submit;
2917 
2918 			sector_t bn = compute_blocknr(sh, i, 1);
2919 			sector_t s = raid5_compute_sector(conf, bn, 0,
2920 							  &dd_idx, NULL);
2921 			sh2 = get_active_stripe(conf, s, 0, 1, 1);
2922 			if (sh2 == NULL)
2923 				/* so far only the early blocks of this stripe
2924 				 * have been requested.  When later blocks
2925 				 * get requested, we will try again
2926 				 */
2927 				continue;
2928 			if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2929 			   test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2930 				/* must have already done this block */
2931 				release_stripe(sh2);
2932 				continue;
2933 			}
2934 
2935 			/* place all the copies on one channel */
2936 			init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2937 			tx = async_memcpy(sh2->dev[dd_idx].page,
2938 					  sh->dev[i].page, 0, 0, STRIPE_SIZE,
2939 					  &submit);
2940 
2941 			set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2942 			set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2943 			for (j = 0; j < conf->raid_disks; j++)
2944 				if (j != sh2->pd_idx &&
2945 				    j != sh2->qd_idx &&
2946 				    !test_bit(R5_Expanded, &sh2->dev[j].flags))
2947 					break;
2948 			if (j == conf->raid_disks) {
2949 				set_bit(STRIPE_EXPAND_READY, &sh2->state);
2950 				set_bit(STRIPE_HANDLE, &sh2->state);
2951 			}
2952 			release_stripe(sh2);
2953 
2954 		}
2955 	/* done submitting copies, wait for them to complete */
2956 	if (tx) {
2957 		async_tx_ack(tx);
2958 		dma_wait_for_async_tx(tx);
2959 	}
2960 }
2961 
2962 
2963 /*
2964  * handle_stripe - do things to a stripe.
2965  *
2966  * We lock the stripe and then examine the state of various bits
2967  * to see what needs to be done.
2968  * Possible results:
2969  *    return some read request which now have data
2970  *    return some write requests which are safely on disc
2971  *    schedule a read on some buffers
2972  *    schedule a write of some buffers
2973  *    return confirmation of parity correctness
2974  *
2975  * buffers are taken off read_list or write_list, and bh_cache buffers
2976  * get BH_Lock set before the stripe lock is released.
2977  *
2978  */
2979 
2980 static void analyse_stripe(struct stripe_head *sh, struct stripe_head_state *s)
2981 {
2982 	struct r5conf *conf = sh->raid_conf;
2983 	int disks = sh->disks;
2984 	struct r5dev *dev;
2985 	int i;
2986 
2987 	memset(s, 0, sizeof(*s));
2988 
2989 	s->syncing = test_bit(STRIPE_SYNCING, &sh->state);
2990 	s->expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2991 	s->expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2992 	s->failed_num[0] = -1;
2993 	s->failed_num[1] = -1;
2994 
2995 	/* Now to look around and see what can be done */
2996 	rcu_read_lock();
2997 	spin_lock_irq(&conf->device_lock);
2998 	for (i=disks; i--; ) {
2999 		struct md_rdev *rdev;
3000 		sector_t first_bad;
3001 		int bad_sectors;
3002 		int is_bad = 0;
3003 
3004 		dev = &sh->dev[i];
3005 
3006 		pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3007 			i, dev->flags, dev->toread, dev->towrite, dev->written);
3008 		/* maybe we can reply to a read
3009 		 *
3010 		 * new wantfill requests are only permitted while
3011 		 * ops_complete_biofill is guaranteed to be inactive
3012 		 */
3013 		if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3014 		    !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3015 			set_bit(R5_Wantfill, &dev->flags);
3016 
3017 		/* now count some things */
3018 		if (test_bit(R5_LOCKED, &dev->flags))
3019 			s->locked++;
3020 		if (test_bit(R5_UPTODATE, &dev->flags))
3021 			s->uptodate++;
3022 		if (test_bit(R5_Wantcompute, &dev->flags)) {
3023 			s->compute++;
3024 			BUG_ON(s->compute > 2);
3025 		}
3026 
3027 		if (test_bit(R5_Wantfill, &dev->flags))
3028 			s->to_fill++;
3029 		else if (dev->toread)
3030 			s->to_read++;
3031 		if (dev->towrite) {
3032 			s->to_write++;
3033 			if (!test_bit(R5_OVERWRITE, &dev->flags))
3034 				s->non_overwrite++;
3035 		}
3036 		if (dev->written)
3037 			s->written++;
3038 		rdev = rcu_dereference(conf->disks[i].rdev);
3039 		if (rdev) {
3040 			is_bad = is_badblock(rdev, sh->sector, STRIPE_SECTORS,
3041 					     &first_bad, &bad_sectors);
3042 			if (s->blocked_rdev == NULL
3043 			    && (test_bit(Blocked, &rdev->flags)
3044 				|| is_bad < 0)) {
3045 				if (is_bad < 0)
3046 					set_bit(BlockedBadBlocks,
3047 						&rdev->flags);
3048 				s->blocked_rdev = rdev;
3049 				atomic_inc(&rdev->nr_pending);
3050 			}
3051 		}
3052 		clear_bit(R5_Insync, &dev->flags);
3053 		if (!rdev)
3054 			/* Not in-sync */;
3055 		else if (is_bad) {
3056 			/* also not in-sync */
3057 			if (!test_bit(WriteErrorSeen, &rdev->flags)) {
3058 				/* treat as in-sync, but with a read error
3059 				 * which we can now try to correct
3060 				 */
3061 				set_bit(R5_Insync, &dev->flags);
3062 				set_bit(R5_ReadError, &dev->flags);
3063 			}
3064 		} else if (test_bit(In_sync, &rdev->flags))
3065 			set_bit(R5_Insync, &dev->flags);
3066 		else if (!test_bit(Faulty, &rdev->flags)) {
3067 			/* in sync if before recovery_offset */
3068 			if (sh->sector + STRIPE_SECTORS <= rdev->recovery_offset)
3069 				set_bit(R5_Insync, &dev->flags);
3070 		}
3071 		if (test_bit(R5_WriteError, &dev->flags)) {
3072 			clear_bit(R5_Insync, &dev->flags);
3073 			if (!test_bit(Faulty, &rdev->flags)) {
3074 				s->handle_bad_blocks = 1;
3075 				atomic_inc(&rdev->nr_pending);
3076 			} else
3077 				clear_bit(R5_WriteError, &dev->flags);
3078 		}
3079 		if (test_bit(R5_MadeGood, &dev->flags)) {
3080 			if (!test_bit(Faulty, &rdev->flags)) {
3081 				s->handle_bad_blocks = 1;
3082 				atomic_inc(&rdev->nr_pending);
3083 			} else
3084 				clear_bit(R5_MadeGood, &dev->flags);
3085 		}
3086 		if (!test_bit(R5_Insync, &dev->flags)) {
3087 			/* The ReadError flag will just be confusing now */
3088 			clear_bit(R5_ReadError, &dev->flags);
3089 			clear_bit(R5_ReWrite, &dev->flags);
3090 		}
3091 		if (test_bit(R5_ReadError, &dev->flags))
3092 			clear_bit(R5_Insync, &dev->flags);
3093 		if (!test_bit(R5_Insync, &dev->flags)) {
3094 			if (s->failed < 2)
3095 				s->failed_num[s->failed] = i;
3096 			s->failed++;
3097 		}
3098 	}
3099 	spin_unlock_irq(&conf->device_lock);
3100 	rcu_read_unlock();
3101 }
3102 
3103 static void handle_stripe(struct stripe_head *sh)
3104 {
3105 	struct stripe_head_state s;
3106 	struct r5conf *conf = sh->raid_conf;
3107 	int i;
3108 	int prexor;
3109 	int disks = sh->disks;
3110 	struct r5dev *pdev, *qdev;
3111 
3112 	clear_bit(STRIPE_HANDLE, &sh->state);
3113 	if (test_and_set_bit_lock(STRIPE_ACTIVE, &sh->state)) {
3114 		/* already being handled, ensure it gets handled
3115 		 * again when current action finishes */
3116 		set_bit(STRIPE_HANDLE, &sh->state);
3117 		return;
3118 	}
3119 
3120 	if (test_and_clear_bit(STRIPE_SYNC_REQUESTED, &sh->state)) {
3121 		set_bit(STRIPE_SYNCING, &sh->state);
3122 		clear_bit(STRIPE_INSYNC, &sh->state);
3123 	}
3124 	clear_bit(STRIPE_DELAYED, &sh->state);
3125 
3126 	pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3127 		"pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3128 	       (unsigned long long)sh->sector, sh->state,
3129 	       atomic_read(&sh->count), sh->pd_idx, sh->qd_idx,
3130 	       sh->check_state, sh->reconstruct_state);
3131 
3132 	analyse_stripe(sh, &s);
3133 
3134 	if (s.handle_bad_blocks) {
3135 		set_bit(STRIPE_HANDLE, &sh->state);
3136 		goto finish;
3137 	}
3138 
3139 	if (unlikely(s.blocked_rdev)) {
3140 		if (s.syncing || s.expanding || s.expanded ||
3141 		    s.to_write || s.written) {
3142 			set_bit(STRIPE_HANDLE, &sh->state);
3143 			goto finish;
3144 		}
3145 		/* There is nothing for the blocked_rdev to block */
3146 		rdev_dec_pending(s.blocked_rdev, conf->mddev);
3147 		s.blocked_rdev = NULL;
3148 	}
3149 
3150 	if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3151 		set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3152 		set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3153 	}
3154 
3155 	pr_debug("locked=%d uptodate=%d to_read=%d"
3156 	       " to_write=%d failed=%d failed_num=%d,%d\n",
3157 	       s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3158 	       s.failed_num[0], s.failed_num[1]);
3159 	/* check if the array has lost more than max_degraded devices and,
3160 	 * if so, some requests might need to be failed.
3161 	 */
3162 	if (s.failed > conf->max_degraded) {
3163 		sh->check_state = 0;
3164 		sh->reconstruct_state = 0;
3165 		if (s.to_read+s.to_write+s.written)
3166 			handle_failed_stripe(conf, sh, &s, disks, &s.return_bi);
3167 		if (s.syncing)
3168 			handle_failed_sync(conf, sh, &s);
3169 	}
3170 
3171 	/*
3172 	 * might be able to return some write requests if the parity blocks
3173 	 * are safe, or on a failed drive
3174 	 */
3175 	pdev = &sh->dev[sh->pd_idx];
3176 	s.p_failed = (s.failed >= 1 && s.failed_num[0] == sh->pd_idx)
3177 		|| (s.failed >= 2 && s.failed_num[1] == sh->pd_idx);
3178 	qdev = &sh->dev[sh->qd_idx];
3179 	s.q_failed = (s.failed >= 1 && s.failed_num[0] == sh->qd_idx)
3180 		|| (s.failed >= 2 && s.failed_num[1] == sh->qd_idx)
3181 		|| conf->level < 6;
3182 
3183 	if (s.written &&
3184 	    (s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3185 			     && !test_bit(R5_LOCKED, &pdev->flags)
3186 			     && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3187 	    (s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3188 			     && !test_bit(R5_LOCKED, &qdev->flags)
3189 			     && test_bit(R5_UPTODATE, &qdev->flags)))))
3190 		handle_stripe_clean_event(conf, sh, disks, &s.return_bi);
3191 
3192 	/* Now we might consider reading some blocks, either to check/generate
3193 	 * parity, or to satisfy requests
3194 	 * or to load a block that is being partially written.
3195 	 */
3196 	if (s.to_read || s.non_overwrite
3197 	    || (conf->level == 6 && s.to_write && s.failed)
3198 	    || (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3199 		handle_stripe_fill(sh, &s, disks);
3200 
3201 	/* Now we check to see if any write operations have recently
3202 	 * completed
3203 	 */
3204 	prexor = 0;
3205 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3206 		prexor = 1;
3207 	if (sh->reconstruct_state == reconstruct_state_drain_result ||
3208 	    sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3209 		sh->reconstruct_state = reconstruct_state_idle;
3210 
3211 		/* All the 'written' buffers and the parity block are ready to
3212 		 * be written back to disk
3213 		 */
3214 		BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3215 		BUG_ON(sh->qd_idx >= 0 &&
3216 		       !test_bit(R5_UPTODATE, &sh->dev[sh->qd_idx].flags));
3217 		for (i = disks; i--; ) {
3218 			struct r5dev *dev = &sh->dev[i];
3219 			if (test_bit(R5_LOCKED, &dev->flags) &&
3220 				(i == sh->pd_idx || i == sh->qd_idx ||
3221 				 dev->written)) {
3222 				pr_debug("Writing block %d\n", i);
3223 				set_bit(R5_Wantwrite, &dev->flags);
3224 				if (prexor)
3225 					continue;
3226 				if (!test_bit(R5_Insync, &dev->flags) ||
3227 				    ((i == sh->pd_idx || i == sh->qd_idx)  &&
3228 				     s.failed == 0))
3229 					set_bit(STRIPE_INSYNC, &sh->state);
3230 			}
3231 		}
3232 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3233 			s.dec_preread_active = 1;
3234 	}
3235 
3236 	/* Now to consider new write requests and what else, if anything
3237 	 * should be read.  We do not handle new writes when:
3238 	 * 1/ A 'write' operation (copy+xor) is already in flight.
3239 	 * 2/ A 'check' operation is in flight, as it may clobber the parity
3240 	 *    block.
3241 	 */
3242 	if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3243 		handle_stripe_dirtying(conf, sh, &s, disks);
3244 
3245 	/* maybe we need to check and possibly fix the parity for this stripe
3246 	 * Any reads will already have been scheduled, so we just see if enough
3247 	 * data is available.  The parity check is held off while parity
3248 	 * dependent operations are in flight.
3249 	 */
3250 	if (sh->check_state ||
3251 	    (s.syncing && s.locked == 0 &&
3252 	     !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3253 	     !test_bit(STRIPE_INSYNC, &sh->state))) {
3254 		if (conf->level == 6)
3255 			handle_parity_checks6(conf, sh, &s, disks);
3256 		else
3257 			handle_parity_checks5(conf, sh, &s, disks);
3258 	}
3259 
3260 	if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3261 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3262 		clear_bit(STRIPE_SYNCING, &sh->state);
3263 	}
3264 
3265 	/* If the failed drives are just a ReadError, then we might need
3266 	 * to progress the repair/check process
3267 	 */
3268 	if (s.failed <= conf->max_degraded && !conf->mddev->ro)
3269 		for (i = 0; i < s.failed; i++) {
3270 			struct r5dev *dev = &sh->dev[s.failed_num[i]];
3271 			if (test_bit(R5_ReadError, &dev->flags)
3272 			    && !test_bit(R5_LOCKED, &dev->flags)
3273 			    && test_bit(R5_UPTODATE, &dev->flags)
3274 				) {
3275 				if (!test_bit(R5_ReWrite, &dev->flags)) {
3276 					set_bit(R5_Wantwrite, &dev->flags);
3277 					set_bit(R5_ReWrite, &dev->flags);
3278 					set_bit(R5_LOCKED, &dev->flags);
3279 					s.locked++;
3280 				} else {
3281 					/* let's read it back */
3282 					set_bit(R5_Wantread, &dev->flags);
3283 					set_bit(R5_LOCKED, &dev->flags);
3284 					s.locked++;
3285 				}
3286 			}
3287 		}
3288 
3289 
3290 	/* Finish reconstruct operations initiated by the expansion process */
3291 	if (sh->reconstruct_state == reconstruct_state_result) {
3292 		struct stripe_head *sh_src
3293 			= get_active_stripe(conf, sh->sector, 1, 1, 1);
3294 		if (sh_src && test_bit(STRIPE_EXPAND_SOURCE, &sh_src->state)) {
3295 			/* sh cannot be written until sh_src has been read.
3296 			 * so arrange for sh to be delayed a little
3297 			 */
3298 			set_bit(STRIPE_DELAYED, &sh->state);
3299 			set_bit(STRIPE_HANDLE, &sh->state);
3300 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3301 					      &sh_src->state))
3302 				atomic_inc(&conf->preread_active_stripes);
3303 			release_stripe(sh_src);
3304 			goto finish;
3305 		}
3306 		if (sh_src)
3307 			release_stripe(sh_src);
3308 
3309 		sh->reconstruct_state = reconstruct_state_idle;
3310 		clear_bit(STRIPE_EXPANDING, &sh->state);
3311 		for (i = conf->raid_disks; i--; ) {
3312 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
3313 			set_bit(R5_LOCKED, &sh->dev[i].flags);
3314 			s.locked++;
3315 		}
3316 	}
3317 
3318 	if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3319 	    !sh->reconstruct_state) {
3320 		/* Need to write out all blocks after computing parity */
3321 		sh->disks = conf->raid_disks;
3322 		stripe_set_idx(sh->sector, conf, 0, sh);
3323 		schedule_reconstruction(sh, &s, 1, 1);
3324 	} else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3325 		clear_bit(STRIPE_EXPAND_READY, &sh->state);
3326 		atomic_dec(&conf->reshape_stripes);
3327 		wake_up(&conf->wait_for_overlap);
3328 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3329 	}
3330 
3331 	if (s.expanding && s.locked == 0 &&
3332 	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3333 		handle_stripe_expansion(conf, sh);
3334 
3335 finish:
3336 	/* wait for this device to become unblocked */
3337 	if (conf->mddev->external && unlikely(s.blocked_rdev))
3338 		md_wait_for_blocked_rdev(s.blocked_rdev, conf->mddev);
3339 
3340 	if (s.handle_bad_blocks)
3341 		for (i = disks; i--; ) {
3342 			struct md_rdev *rdev;
3343 			struct r5dev *dev = &sh->dev[i];
3344 			if (test_and_clear_bit(R5_WriteError, &dev->flags)) {
3345 				/* We own a safe reference to the rdev */
3346 				rdev = conf->disks[i].rdev;
3347 				if (!rdev_set_badblocks(rdev, sh->sector,
3348 							STRIPE_SECTORS, 0))
3349 					md_error(conf->mddev, rdev);
3350 				rdev_dec_pending(rdev, conf->mddev);
3351 			}
3352 			if (test_and_clear_bit(R5_MadeGood, &dev->flags)) {
3353 				rdev = conf->disks[i].rdev;
3354 				rdev_clear_badblocks(rdev, sh->sector,
3355 						     STRIPE_SECTORS);
3356 				rdev_dec_pending(rdev, conf->mddev);
3357 			}
3358 		}
3359 
3360 	if (s.ops_request)
3361 		raid_run_ops(sh, s.ops_request);
3362 
3363 	ops_run_io(sh, &s);
3364 
3365 	if (s.dec_preread_active) {
3366 		/* We delay this until after ops_run_io so that if make_request
3367 		 * is waiting on a flush, it won't continue until the writes
3368 		 * have actually been submitted.
3369 		 */
3370 		atomic_dec(&conf->preread_active_stripes);
3371 		if (atomic_read(&conf->preread_active_stripes) <
3372 		    IO_THRESHOLD)
3373 			md_wakeup_thread(conf->mddev->thread);
3374 	}
3375 
3376 	return_io(s.return_bi);
3377 
3378 	clear_bit_unlock(STRIPE_ACTIVE, &sh->state);
3379 }
3380 
3381 static void raid5_activate_delayed(struct r5conf *conf)
3382 {
3383 	if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3384 		while (!list_empty(&conf->delayed_list)) {
3385 			struct list_head *l = conf->delayed_list.next;
3386 			struct stripe_head *sh;
3387 			sh = list_entry(l, struct stripe_head, lru);
3388 			list_del_init(l);
3389 			clear_bit(STRIPE_DELAYED, &sh->state);
3390 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3391 				atomic_inc(&conf->preread_active_stripes);
3392 			list_add_tail(&sh->lru, &conf->hold_list);
3393 		}
3394 	}
3395 }
3396 
3397 static void activate_bit_delay(struct r5conf *conf)
3398 {
3399 	/* device_lock is held */
3400 	struct list_head head;
3401 	list_add(&head, &conf->bitmap_list);
3402 	list_del_init(&conf->bitmap_list);
3403 	while (!list_empty(&head)) {
3404 		struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3405 		list_del_init(&sh->lru);
3406 		atomic_inc(&sh->count);
3407 		__release_stripe(conf, sh);
3408 	}
3409 }
3410 
3411 int md_raid5_congested(struct mddev *mddev, int bits)
3412 {
3413 	struct r5conf *conf = mddev->private;
3414 
3415 	/* No difference between reads and writes.  Just check
3416 	 * how busy the stripe_cache is
3417 	 */
3418 
3419 	if (conf->inactive_blocked)
3420 		return 1;
3421 	if (conf->quiesce)
3422 		return 1;
3423 	if (list_empty_careful(&conf->inactive_list))
3424 		return 1;
3425 
3426 	return 0;
3427 }
3428 EXPORT_SYMBOL_GPL(md_raid5_congested);
3429 
3430 static int raid5_congested(void *data, int bits)
3431 {
3432 	struct mddev *mddev = data;
3433 
3434 	return mddev_congested(mddev, bits) ||
3435 		md_raid5_congested(mddev, bits);
3436 }
3437 
3438 /* We want read requests to align with chunks where possible,
3439  * but write requests don't need to.
3440  */
3441 static int raid5_mergeable_bvec(struct request_queue *q,
3442 				struct bvec_merge_data *bvm,
3443 				struct bio_vec *biovec)
3444 {
3445 	struct mddev *mddev = q->queuedata;
3446 	sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3447 	int max;
3448 	unsigned int chunk_sectors = mddev->chunk_sectors;
3449 	unsigned int bio_sectors = bvm->bi_size >> 9;
3450 
3451 	if ((bvm->bi_rw & 1) == WRITE)
3452 		return biovec->bv_len; /* always allow writes to be mergeable */
3453 
3454 	if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3455 		chunk_sectors = mddev->new_chunk_sectors;
3456 	max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3457 	if (max < 0) max = 0;
3458 	if (max <= biovec->bv_len && bio_sectors == 0)
3459 		return biovec->bv_len;
3460 	else
3461 		return max;
3462 }
3463 
3464 
3465 static int in_chunk_boundary(struct mddev *mddev, struct bio *bio)
3466 {
3467 	sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3468 	unsigned int chunk_sectors = mddev->chunk_sectors;
3469 	unsigned int bio_sectors = bio->bi_size >> 9;
3470 
3471 	if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3472 		chunk_sectors = mddev->new_chunk_sectors;
3473 	return  chunk_sectors >=
3474 		((sector & (chunk_sectors - 1)) + bio_sectors);
3475 }
3476 
3477 /*
3478  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3479  *  later sampled by raid5d.
3480  */
3481 static void add_bio_to_retry(struct bio *bi,struct r5conf *conf)
3482 {
3483 	unsigned long flags;
3484 
3485 	spin_lock_irqsave(&conf->device_lock, flags);
3486 
3487 	bi->bi_next = conf->retry_read_aligned_list;
3488 	conf->retry_read_aligned_list = bi;
3489 
3490 	spin_unlock_irqrestore(&conf->device_lock, flags);
3491 	md_wakeup_thread(conf->mddev->thread);
3492 }
3493 
3494 
3495 static struct bio *remove_bio_from_retry(struct r5conf *conf)
3496 {
3497 	struct bio *bi;
3498 
3499 	bi = conf->retry_read_aligned;
3500 	if (bi) {
3501 		conf->retry_read_aligned = NULL;
3502 		return bi;
3503 	}
3504 	bi = conf->retry_read_aligned_list;
3505 	if(bi) {
3506 		conf->retry_read_aligned_list = bi->bi_next;
3507 		bi->bi_next = NULL;
3508 		/*
3509 		 * this sets the active strip count to 1 and the processed
3510 		 * strip count to zero (upper 8 bits)
3511 		 */
3512 		bi->bi_phys_segments = 1; /* biased count of active stripes */
3513 	}
3514 
3515 	return bi;
3516 }
3517 
3518 
3519 /*
3520  *  The "raid5_align_endio" should check if the read succeeded and if it
3521  *  did, call bio_endio on the original bio (having bio_put the new bio
3522  *  first).
3523  *  If the read failed..
3524  */
3525 static void raid5_align_endio(struct bio *bi, int error)
3526 {
3527 	struct bio* raid_bi  = bi->bi_private;
3528 	struct mddev *mddev;
3529 	struct r5conf *conf;
3530 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3531 	struct md_rdev *rdev;
3532 
3533 	bio_put(bi);
3534 
3535 	rdev = (void*)raid_bi->bi_next;
3536 	raid_bi->bi_next = NULL;
3537 	mddev = rdev->mddev;
3538 	conf = mddev->private;
3539 
3540 	rdev_dec_pending(rdev, conf->mddev);
3541 
3542 	if (!error && uptodate) {
3543 		bio_endio(raid_bi, 0);
3544 		if (atomic_dec_and_test(&conf->active_aligned_reads))
3545 			wake_up(&conf->wait_for_stripe);
3546 		return;
3547 	}
3548 
3549 
3550 	pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3551 
3552 	add_bio_to_retry(raid_bi, conf);
3553 }
3554 
3555 static int bio_fits_rdev(struct bio *bi)
3556 {
3557 	struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3558 
3559 	if ((bi->bi_size>>9) > queue_max_sectors(q))
3560 		return 0;
3561 	blk_recount_segments(q, bi);
3562 	if (bi->bi_phys_segments > queue_max_segments(q))
3563 		return 0;
3564 
3565 	if (q->merge_bvec_fn)
3566 		/* it's too hard to apply the merge_bvec_fn at this stage,
3567 		 * just just give up
3568 		 */
3569 		return 0;
3570 
3571 	return 1;
3572 }
3573 
3574 
3575 static int chunk_aligned_read(struct mddev *mddev, struct bio * raid_bio)
3576 {
3577 	struct r5conf *conf = mddev->private;
3578 	int dd_idx;
3579 	struct bio* align_bi;
3580 	struct md_rdev *rdev;
3581 
3582 	if (!in_chunk_boundary(mddev, raid_bio)) {
3583 		pr_debug("chunk_aligned_read : non aligned\n");
3584 		return 0;
3585 	}
3586 	/*
3587 	 * use bio_clone_mddev to make a copy of the bio
3588 	 */
3589 	align_bi = bio_clone_mddev(raid_bio, GFP_NOIO, mddev);
3590 	if (!align_bi)
3591 		return 0;
3592 	/*
3593 	 *   set bi_end_io to a new function, and set bi_private to the
3594 	 *     original bio.
3595 	 */
3596 	align_bi->bi_end_io  = raid5_align_endio;
3597 	align_bi->bi_private = raid_bio;
3598 	/*
3599 	 *	compute position
3600 	 */
3601 	align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3602 						    0,
3603 						    &dd_idx, NULL);
3604 
3605 	rcu_read_lock();
3606 	rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3607 	if (rdev && test_bit(In_sync, &rdev->flags)) {
3608 		sector_t first_bad;
3609 		int bad_sectors;
3610 
3611 		atomic_inc(&rdev->nr_pending);
3612 		rcu_read_unlock();
3613 		raid_bio->bi_next = (void*)rdev;
3614 		align_bi->bi_bdev =  rdev->bdev;
3615 		align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3616 		align_bi->bi_sector += rdev->data_offset;
3617 
3618 		if (!bio_fits_rdev(align_bi) ||
3619 		    is_badblock(rdev, align_bi->bi_sector, align_bi->bi_size>>9,
3620 				&first_bad, &bad_sectors)) {
3621 			/* too big in some way, or has a known bad block */
3622 			bio_put(align_bi);
3623 			rdev_dec_pending(rdev, mddev);
3624 			return 0;
3625 		}
3626 
3627 		spin_lock_irq(&conf->device_lock);
3628 		wait_event_lock_irq(conf->wait_for_stripe,
3629 				    conf->quiesce == 0,
3630 				    conf->device_lock, /* nothing */);
3631 		atomic_inc(&conf->active_aligned_reads);
3632 		spin_unlock_irq(&conf->device_lock);
3633 
3634 		generic_make_request(align_bi);
3635 		return 1;
3636 	} else {
3637 		rcu_read_unlock();
3638 		bio_put(align_bi);
3639 		return 0;
3640 	}
3641 }
3642 
3643 /* __get_priority_stripe - get the next stripe to process
3644  *
3645  * Full stripe writes are allowed to pass preread active stripes up until
3646  * the bypass_threshold is exceeded.  In general the bypass_count
3647  * increments when the handle_list is handled before the hold_list; however, it
3648  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3649  * stripe with in flight i/o.  The bypass_count will be reset when the
3650  * head of the hold_list has changed, i.e. the head was promoted to the
3651  * handle_list.
3652  */
3653 static struct stripe_head *__get_priority_stripe(struct r5conf *conf)
3654 {
3655 	struct stripe_head *sh;
3656 
3657 	pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3658 		  __func__,
3659 		  list_empty(&conf->handle_list) ? "empty" : "busy",
3660 		  list_empty(&conf->hold_list) ? "empty" : "busy",
3661 		  atomic_read(&conf->pending_full_writes), conf->bypass_count);
3662 
3663 	if (!list_empty(&conf->handle_list)) {
3664 		sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3665 
3666 		if (list_empty(&conf->hold_list))
3667 			conf->bypass_count = 0;
3668 		else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3669 			if (conf->hold_list.next == conf->last_hold)
3670 				conf->bypass_count++;
3671 			else {
3672 				conf->last_hold = conf->hold_list.next;
3673 				conf->bypass_count -= conf->bypass_threshold;
3674 				if (conf->bypass_count < 0)
3675 					conf->bypass_count = 0;
3676 			}
3677 		}
3678 	} else if (!list_empty(&conf->hold_list) &&
3679 		   ((conf->bypass_threshold &&
3680 		     conf->bypass_count > conf->bypass_threshold) ||
3681 		    atomic_read(&conf->pending_full_writes) == 0)) {
3682 		sh = list_entry(conf->hold_list.next,
3683 				typeof(*sh), lru);
3684 		conf->bypass_count -= conf->bypass_threshold;
3685 		if (conf->bypass_count < 0)
3686 			conf->bypass_count = 0;
3687 	} else
3688 		return NULL;
3689 
3690 	list_del_init(&sh->lru);
3691 	atomic_inc(&sh->count);
3692 	BUG_ON(atomic_read(&sh->count) != 1);
3693 	return sh;
3694 }
3695 
3696 static void make_request(struct mddev *mddev, struct bio * bi)
3697 {
3698 	struct r5conf *conf = mddev->private;
3699 	int dd_idx;
3700 	sector_t new_sector;
3701 	sector_t logical_sector, last_sector;
3702 	struct stripe_head *sh;
3703 	const int rw = bio_data_dir(bi);
3704 	int remaining;
3705 	int plugged;
3706 
3707 	if (unlikely(bi->bi_rw & REQ_FLUSH)) {
3708 		md_flush_request(mddev, bi);
3709 		return;
3710 	}
3711 
3712 	md_write_start(mddev, bi);
3713 
3714 	if (rw == READ &&
3715 	     mddev->reshape_position == MaxSector &&
3716 	     chunk_aligned_read(mddev,bi))
3717 		return;
3718 
3719 	logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3720 	last_sector = bi->bi_sector + (bi->bi_size>>9);
3721 	bi->bi_next = NULL;
3722 	bi->bi_phys_segments = 1;	/* over-loaded to count active stripes */
3723 
3724 	plugged = mddev_check_plugged(mddev);
3725 	for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3726 		DEFINE_WAIT(w);
3727 		int disks, data_disks;
3728 		int previous;
3729 
3730 	retry:
3731 		previous = 0;
3732 		disks = conf->raid_disks;
3733 		prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3734 		if (unlikely(conf->reshape_progress != MaxSector)) {
3735 			/* spinlock is needed as reshape_progress may be
3736 			 * 64bit on a 32bit platform, and so it might be
3737 			 * possible to see a half-updated value
3738 			 * Of course reshape_progress could change after
3739 			 * the lock is dropped, so once we get a reference
3740 			 * to the stripe that we think it is, we will have
3741 			 * to check again.
3742 			 */
3743 			spin_lock_irq(&conf->device_lock);
3744 			if (mddev->delta_disks < 0
3745 			    ? logical_sector < conf->reshape_progress
3746 			    : logical_sector >= conf->reshape_progress) {
3747 				disks = conf->previous_raid_disks;
3748 				previous = 1;
3749 			} else {
3750 				if (mddev->delta_disks < 0
3751 				    ? logical_sector < conf->reshape_safe
3752 				    : logical_sector >= conf->reshape_safe) {
3753 					spin_unlock_irq(&conf->device_lock);
3754 					schedule();
3755 					goto retry;
3756 				}
3757 			}
3758 			spin_unlock_irq(&conf->device_lock);
3759 		}
3760 		data_disks = disks - conf->max_degraded;
3761 
3762 		new_sector = raid5_compute_sector(conf, logical_sector,
3763 						  previous,
3764 						  &dd_idx, NULL);
3765 		pr_debug("raid456: make_request, sector %llu logical %llu\n",
3766 			(unsigned long long)new_sector,
3767 			(unsigned long long)logical_sector);
3768 
3769 		sh = get_active_stripe(conf, new_sector, previous,
3770 				       (bi->bi_rw&RWA_MASK), 0);
3771 		if (sh) {
3772 			if (unlikely(previous)) {
3773 				/* expansion might have moved on while waiting for a
3774 				 * stripe, so we must do the range check again.
3775 				 * Expansion could still move past after this
3776 				 * test, but as we are holding a reference to
3777 				 * 'sh', we know that if that happens,
3778 				 *  STRIPE_EXPANDING will get set and the expansion
3779 				 * won't proceed until we finish with the stripe.
3780 				 */
3781 				int must_retry = 0;
3782 				spin_lock_irq(&conf->device_lock);
3783 				if (mddev->delta_disks < 0
3784 				    ? logical_sector >= conf->reshape_progress
3785 				    : logical_sector < conf->reshape_progress)
3786 					/* mismatch, need to try again */
3787 					must_retry = 1;
3788 				spin_unlock_irq(&conf->device_lock);
3789 				if (must_retry) {
3790 					release_stripe(sh);
3791 					schedule();
3792 					goto retry;
3793 				}
3794 			}
3795 
3796 			if (rw == WRITE &&
3797 			    logical_sector >= mddev->suspend_lo &&
3798 			    logical_sector < mddev->suspend_hi) {
3799 				release_stripe(sh);
3800 				/* As the suspend_* range is controlled by
3801 				 * userspace, we want an interruptible
3802 				 * wait.
3803 				 */
3804 				flush_signals(current);
3805 				prepare_to_wait(&conf->wait_for_overlap,
3806 						&w, TASK_INTERRUPTIBLE);
3807 				if (logical_sector >= mddev->suspend_lo &&
3808 				    logical_sector < mddev->suspend_hi)
3809 					schedule();
3810 				goto retry;
3811 			}
3812 
3813 			if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3814 			    !add_stripe_bio(sh, bi, dd_idx, rw)) {
3815 				/* Stripe is busy expanding or
3816 				 * add failed due to overlap.  Flush everything
3817 				 * and wait a while
3818 				 */
3819 				md_wakeup_thread(mddev->thread);
3820 				release_stripe(sh);
3821 				schedule();
3822 				goto retry;
3823 			}
3824 			finish_wait(&conf->wait_for_overlap, &w);
3825 			set_bit(STRIPE_HANDLE, &sh->state);
3826 			clear_bit(STRIPE_DELAYED, &sh->state);
3827 			if ((bi->bi_rw & REQ_SYNC) &&
3828 			    !test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3829 				atomic_inc(&conf->preread_active_stripes);
3830 			release_stripe(sh);
3831 		} else {
3832 			/* cannot get stripe for read-ahead, just give-up */
3833 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
3834 			finish_wait(&conf->wait_for_overlap, &w);
3835 			break;
3836 		}
3837 
3838 	}
3839 	if (!plugged)
3840 		md_wakeup_thread(mddev->thread);
3841 
3842 	spin_lock_irq(&conf->device_lock);
3843 	remaining = raid5_dec_bi_phys_segments(bi);
3844 	spin_unlock_irq(&conf->device_lock);
3845 	if (remaining == 0) {
3846 
3847 		if ( rw == WRITE )
3848 			md_write_end(mddev);
3849 
3850 		bio_endio(bi, 0);
3851 	}
3852 }
3853 
3854 static sector_t raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks);
3855 
3856 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr, int *skipped)
3857 {
3858 	/* reshaping is quite different to recovery/resync so it is
3859 	 * handled quite separately ... here.
3860 	 *
3861 	 * On each call to sync_request, we gather one chunk worth of
3862 	 * destination stripes and flag them as expanding.
3863 	 * Then we find all the source stripes and request reads.
3864 	 * As the reads complete, handle_stripe will copy the data
3865 	 * into the destination stripe and release that stripe.
3866 	 */
3867 	struct r5conf *conf = mddev->private;
3868 	struct stripe_head *sh;
3869 	sector_t first_sector, last_sector;
3870 	int raid_disks = conf->previous_raid_disks;
3871 	int data_disks = raid_disks - conf->max_degraded;
3872 	int new_data_disks = conf->raid_disks - conf->max_degraded;
3873 	int i;
3874 	int dd_idx;
3875 	sector_t writepos, readpos, safepos;
3876 	sector_t stripe_addr;
3877 	int reshape_sectors;
3878 	struct list_head stripes;
3879 
3880 	if (sector_nr == 0) {
3881 		/* If restarting in the middle, skip the initial sectors */
3882 		if (mddev->delta_disks < 0 &&
3883 		    conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3884 			sector_nr = raid5_size(mddev, 0, 0)
3885 				- conf->reshape_progress;
3886 		} else if (mddev->delta_disks >= 0 &&
3887 			   conf->reshape_progress > 0)
3888 			sector_nr = conf->reshape_progress;
3889 		sector_div(sector_nr, new_data_disks);
3890 		if (sector_nr) {
3891 			mddev->curr_resync_completed = sector_nr;
3892 			sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3893 			*skipped = 1;
3894 			return sector_nr;
3895 		}
3896 	}
3897 
3898 	/* We need to process a full chunk at a time.
3899 	 * If old and new chunk sizes differ, we need to process the
3900 	 * largest of these
3901 	 */
3902 	if (mddev->new_chunk_sectors > mddev->chunk_sectors)
3903 		reshape_sectors = mddev->new_chunk_sectors;
3904 	else
3905 		reshape_sectors = mddev->chunk_sectors;
3906 
3907 	/* we update the metadata when there is more than 3Meg
3908 	 * in the block range (that is rather arbitrary, should
3909 	 * probably be time based) or when the data about to be
3910 	 * copied would over-write the source of the data at
3911 	 * the front of the range.
3912 	 * i.e. one new_stripe along from reshape_progress new_maps
3913 	 * to after where reshape_safe old_maps to
3914 	 */
3915 	writepos = conf->reshape_progress;
3916 	sector_div(writepos, new_data_disks);
3917 	readpos = conf->reshape_progress;
3918 	sector_div(readpos, data_disks);
3919 	safepos = conf->reshape_safe;
3920 	sector_div(safepos, data_disks);
3921 	if (mddev->delta_disks < 0) {
3922 		writepos -= min_t(sector_t, reshape_sectors, writepos);
3923 		readpos += reshape_sectors;
3924 		safepos += reshape_sectors;
3925 	} else {
3926 		writepos += reshape_sectors;
3927 		readpos -= min_t(sector_t, reshape_sectors, readpos);
3928 		safepos -= min_t(sector_t, reshape_sectors, safepos);
3929 	}
3930 
3931 	/* 'writepos' is the most advanced device address we might write.
3932 	 * 'readpos' is the least advanced device address we might read.
3933 	 * 'safepos' is the least address recorded in the metadata as having
3934 	 *     been reshaped.
3935 	 * If 'readpos' is behind 'writepos', then there is no way that we can
3936 	 * ensure safety in the face of a crash - that must be done by userspace
3937 	 * making a backup of the data.  So in that case there is no particular
3938 	 * rush to update metadata.
3939 	 * Otherwise if 'safepos' is behind 'writepos', then we really need to
3940 	 * update the metadata to advance 'safepos' to match 'readpos' so that
3941 	 * we can be safe in the event of a crash.
3942 	 * So we insist on updating metadata if safepos is behind writepos and
3943 	 * readpos is beyond writepos.
3944 	 * In any case, update the metadata every 10 seconds.
3945 	 * Maybe that number should be configurable, but I'm not sure it is
3946 	 * worth it.... maybe it could be a multiple of safemode_delay???
3947 	 */
3948 	if ((mddev->delta_disks < 0
3949 	     ? (safepos > writepos && readpos < writepos)
3950 	     : (safepos < writepos && readpos > writepos)) ||
3951 	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
3952 		/* Cannot proceed until we've updated the superblock... */
3953 		wait_event(conf->wait_for_overlap,
3954 			   atomic_read(&conf->reshape_stripes)==0);
3955 		mddev->reshape_position = conf->reshape_progress;
3956 		mddev->curr_resync_completed = sector_nr;
3957 		conf->reshape_checkpoint = jiffies;
3958 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
3959 		md_wakeup_thread(mddev->thread);
3960 		wait_event(mddev->sb_wait, mddev->flags == 0 ||
3961 			   kthread_should_stop());
3962 		spin_lock_irq(&conf->device_lock);
3963 		conf->reshape_safe = mddev->reshape_position;
3964 		spin_unlock_irq(&conf->device_lock);
3965 		wake_up(&conf->wait_for_overlap);
3966 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
3967 	}
3968 
3969 	if (mddev->delta_disks < 0) {
3970 		BUG_ON(conf->reshape_progress == 0);
3971 		stripe_addr = writepos;
3972 		BUG_ON((mddev->dev_sectors &
3973 			~((sector_t)reshape_sectors - 1))
3974 		       - reshape_sectors - stripe_addr
3975 		       != sector_nr);
3976 	} else {
3977 		BUG_ON(writepos != sector_nr + reshape_sectors);
3978 		stripe_addr = sector_nr;
3979 	}
3980 	INIT_LIST_HEAD(&stripes);
3981 	for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
3982 		int j;
3983 		int skipped_disk = 0;
3984 		sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
3985 		set_bit(STRIPE_EXPANDING, &sh->state);
3986 		atomic_inc(&conf->reshape_stripes);
3987 		/* If any of this stripe is beyond the end of the old
3988 		 * array, then we need to zero those blocks
3989 		 */
3990 		for (j=sh->disks; j--;) {
3991 			sector_t s;
3992 			if (j == sh->pd_idx)
3993 				continue;
3994 			if (conf->level == 6 &&
3995 			    j == sh->qd_idx)
3996 				continue;
3997 			s = compute_blocknr(sh, j, 0);
3998 			if (s < raid5_size(mddev, 0, 0)) {
3999 				skipped_disk = 1;
4000 				continue;
4001 			}
4002 			memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4003 			set_bit(R5_Expanded, &sh->dev[j].flags);
4004 			set_bit(R5_UPTODATE, &sh->dev[j].flags);
4005 		}
4006 		if (!skipped_disk) {
4007 			set_bit(STRIPE_EXPAND_READY, &sh->state);
4008 			set_bit(STRIPE_HANDLE, &sh->state);
4009 		}
4010 		list_add(&sh->lru, &stripes);
4011 	}
4012 	spin_lock_irq(&conf->device_lock);
4013 	if (mddev->delta_disks < 0)
4014 		conf->reshape_progress -= reshape_sectors * new_data_disks;
4015 	else
4016 		conf->reshape_progress += reshape_sectors * new_data_disks;
4017 	spin_unlock_irq(&conf->device_lock);
4018 	/* Ok, those stripe are ready. We can start scheduling
4019 	 * reads on the source stripes.
4020 	 * The source stripes are determined by mapping the first and last
4021 	 * block on the destination stripes.
4022 	 */
4023 	first_sector =
4024 		raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4025 				     1, &dd_idx, NULL);
4026 	last_sector =
4027 		raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4028 					    * new_data_disks - 1),
4029 				     1, &dd_idx, NULL);
4030 	if (last_sector >= mddev->dev_sectors)
4031 		last_sector = mddev->dev_sectors - 1;
4032 	while (first_sector <= last_sector) {
4033 		sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4034 		set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4035 		set_bit(STRIPE_HANDLE, &sh->state);
4036 		release_stripe(sh);
4037 		first_sector += STRIPE_SECTORS;
4038 	}
4039 	/* Now that the sources are clearly marked, we can release
4040 	 * the destination stripes
4041 	 */
4042 	while (!list_empty(&stripes)) {
4043 		sh = list_entry(stripes.next, struct stripe_head, lru);
4044 		list_del_init(&sh->lru);
4045 		release_stripe(sh);
4046 	}
4047 	/* If this takes us to the resync_max point where we have to pause,
4048 	 * then we need to write out the superblock.
4049 	 */
4050 	sector_nr += reshape_sectors;
4051 	if ((sector_nr - mddev->curr_resync_completed) * 2
4052 	    >= mddev->resync_max - mddev->curr_resync_completed) {
4053 		/* Cannot proceed until we've updated the superblock... */
4054 		wait_event(conf->wait_for_overlap,
4055 			   atomic_read(&conf->reshape_stripes) == 0);
4056 		mddev->reshape_position = conf->reshape_progress;
4057 		mddev->curr_resync_completed = sector_nr;
4058 		conf->reshape_checkpoint = jiffies;
4059 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
4060 		md_wakeup_thread(mddev->thread);
4061 		wait_event(mddev->sb_wait,
4062 			   !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4063 			   || kthread_should_stop());
4064 		spin_lock_irq(&conf->device_lock);
4065 		conf->reshape_safe = mddev->reshape_position;
4066 		spin_unlock_irq(&conf->device_lock);
4067 		wake_up(&conf->wait_for_overlap);
4068 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4069 	}
4070 	return reshape_sectors;
4071 }
4072 
4073 /* FIXME go_faster isn't used */
4074 static inline sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
4075 {
4076 	struct r5conf *conf = mddev->private;
4077 	struct stripe_head *sh;
4078 	sector_t max_sector = mddev->dev_sectors;
4079 	sector_t sync_blocks;
4080 	int still_degraded = 0;
4081 	int i;
4082 
4083 	if (sector_nr >= max_sector) {
4084 		/* just being told to finish up .. nothing much to do */
4085 
4086 		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4087 			end_reshape(conf);
4088 			return 0;
4089 		}
4090 
4091 		if (mddev->curr_resync < max_sector) /* aborted */
4092 			bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4093 					&sync_blocks, 1);
4094 		else /* completed sync */
4095 			conf->fullsync = 0;
4096 		bitmap_close_sync(mddev->bitmap);
4097 
4098 		return 0;
4099 	}
4100 
4101 	/* Allow raid5_quiesce to complete */
4102 	wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4103 
4104 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4105 		return reshape_request(mddev, sector_nr, skipped);
4106 
4107 	/* No need to check resync_max as we never do more than one
4108 	 * stripe, and as resync_max will always be on a chunk boundary,
4109 	 * if the check in md_do_sync didn't fire, there is no chance
4110 	 * of overstepping resync_max here
4111 	 */
4112 
4113 	/* if there is too many failed drives and we are trying
4114 	 * to resync, then assert that we are finished, because there is
4115 	 * nothing we can do.
4116 	 */
4117 	if (mddev->degraded >= conf->max_degraded &&
4118 	    test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4119 		sector_t rv = mddev->dev_sectors - sector_nr;
4120 		*skipped = 1;
4121 		return rv;
4122 	}
4123 	if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4124 	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4125 	    !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4126 		/* we can skip this block, and probably more */
4127 		sync_blocks /= STRIPE_SECTORS;
4128 		*skipped = 1;
4129 		return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4130 	}
4131 
4132 
4133 	bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4134 
4135 	sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4136 	if (sh == NULL) {
4137 		sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4138 		/* make sure we don't swamp the stripe cache if someone else
4139 		 * is trying to get access
4140 		 */
4141 		schedule_timeout_uninterruptible(1);
4142 	}
4143 	/* Need to check if array will still be degraded after recovery/resync
4144 	 * We don't need to check the 'failed' flag as when that gets set,
4145 	 * recovery aborts.
4146 	 */
4147 	for (i = 0; i < conf->raid_disks; i++)
4148 		if (conf->disks[i].rdev == NULL)
4149 			still_degraded = 1;
4150 
4151 	bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4152 
4153 	set_bit(STRIPE_SYNC_REQUESTED, &sh->state);
4154 
4155 	handle_stripe(sh);
4156 	release_stripe(sh);
4157 
4158 	return STRIPE_SECTORS;
4159 }
4160 
4161 static int  retry_aligned_read(struct r5conf *conf, struct bio *raid_bio)
4162 {
4163 	/* We may not be able to submit a whole bio at once as there
4164 	 * may not be enough stripe_heads available.
4165 	 * We cannot pre-allocate enough stripe_heads as we may need
4166 	 * more than exist in the cache (if we allow ever large chunks).
4167 	 * So we do one stripe head at a time and record in
4168 	 * ->bi_hw_segments how many have been done.
4169 	 *
4170 	 * We *know* that this entire raid_bio is in one chunk, so
4171 	 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4172 	 */
4173 	struct stripe_head *sh;
4174 	int dd_idx;
4175 	sector_t sector, logical_sector, last_sector;
4176 	int scnt = 0;
4177 	int remaining;
4178 	int handled = 0;
4179 
4180 	logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4181 	sector = raid5_compute_sector(conf, logical_sector,
4182 				      0, &dd_idx, NULL);
4183 	last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4184 
4185 	for (; logical_sector < last_sector;
4186 	     logical_sector += STRIPE_SECTORS,
4187 		     sector += STRIPE_SECTORS,
4188 		     scnt++) {
4189 
4190 		if (scnt < raid5_bi_hw_segments(raid_bio))
4191 			/* already done this stripe */
4192 			continue;
4193 
4194 		sh = get_active_stripe(conf, sector, 0, 1, 0);
4195 
4196 		if (!sh) {
4197 			/* failed to get a stripe - must wait */
4198 			raid5_set_bi_hw_segments(raid_bio, scnt);
4199 			conf->retry_read_aligned = raid_bio;
4200 			return handled;
4201 		}
4202 
4203 		set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4204 		if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4205 			release_stripe(sh);
4206 			raid5_set_bi_hw_segments(raid_bio, scnt);
4207 			conf->retry_read_aligned = raid_bio;
4208 			return handled;
4209 		}
4210 
4211 		handle_stripe(sh);
4212 		release_stripe(sh);
4213 		handled++;
4214 	}
4215 	spin_lock_irq(&conf->device_lock);
4216 	remaining = raid5_dec_bi_phys_segments(raid_bio);
4217 	spin_unlock_irq(&conf->device_lock);
4218 	if (remaining == 0)
4219 		bio_endio(raid_bio, 0);
4220 	if (atomic_dec_and_test(&conf->active_aligned_reads))
4221 		wake_up(&conf->wait_for_stripe);
4222 	return handled;
4223 }
4224 
4225 
4226 /*
4227  * This is our raid5 kernel thread.
4228  *
4229  * We scan the hash table for stripes which can be handled now.
4230  * During the scan, completed stripes are saved for us by the interrupt
4231  * handler, so that they will not have to wait for our next wakeup.
4232  */
4233 static void raid5d(struct mddev *mddev)
4234 {
4235 	struct stripe_head *sh;
4236 	struct r5conf *conf = mddev->private;
4237 	int handled;
4238 	struct blk_plug plug;
4239 
4240 	pr_debug("+++ raid5d active\n");
4241 
4242 	md_check_recovery(mddev);
4243 
4244 	blk_start_plug(&plug);
4245 	handled = 0;
4246 	spin_lock_irq(&conf->device_lock);
4247 	while (1) {
4248 		struct bio *bio;
4249 
4250 		if (atomic_read(&mddev->plug_cnt) == 0 &&
4251 		    !list_empty(&conf->bitmap_list)) {
4252 			/* Now is a good time to flush some bitmap updates */
4253 			conf->seq_flush++;
4254 			spin_unlock_irq(&conf->device_lock);
4255 			bitmap_unplug(mddev->bitmap);
4256 			spin_lock_irq(&conf->device_lock);
4257 			conf->seq_write = conf->seq_flush;
4258 			activate_bit_delay(conf);
4259 		}
4260 		if (atomic_read(&mddev->plug_cnt) == 0)
4261 			raid5_activate_delayed(conf);
4262 
4263 		while ((bio = remove_bio_from_retry(conf))) {
4264 			int ok;
4265 			spin_unlock_irq(&conf->device_lock);
4266 			ok = retry_aligned_read(conf, bio);
4267 			spin_lock_irq(&conf->device_lock);
4268 			if (!ok)
4269 				break;
4270 			handled++;
4271 		}
4272 
4273 		sh = __get_priority_stripe(conf);
4274 
4275 		if (!sh)
4276 			break;
4277 		spin_unlock_irq(&conf->device_lock);
4278 
4279 		handled++;
4280 		handle_stripe(sh);
4281 		release_stripe(sh);
4282 		cond_resched();
4283 
4284 		if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
4285 			md_check_recovery(mddev);
4286 
4287 		spin_lock_irq(&conf->device_lock);
4288 	}
4289 	pr_debug("%d stripes handled\n", handled);
4290 
4291 	spin_unlock_irq(&conf->device_lock);
4292 
4293 	async_tx_issue_pending_all();
4294 	blk_finish_plug(&plug);
4295 
4296 	pr_debug("--- raid5d inactive\n");
4297 }
4298 
4299 static ssize_t
4300 raid5_show_stripe_cache_size(struct mddev *mddev, char *page)
4301 {
4302 	struct r5conf *conf = mddev->private;
4303 	if (conf)
4304 		return sprintf(page, "%d\n", conf->max_nr_stripes);
4305 	else
4306 		return 0;
4307 }
4308 
4309 int
4310 raid5_set_cache_size(struct mddev *mddev, int size)
4311 {
4312 	struct r5conf *conf = mddev->private;
4313 	int err;
4314 
4315 	if (size <= 16 || size > 32768)
4316 		return -EINVAL;
4317 	while (size < conf->max_nr_stripes) {
4318 		if (drop_one_stripe(conf))
4319 			conf->max_nr_stripes--;
4320 		else
4321 			break;
4322 	}
4323 	err = md_allow_write(mddev);
4324 	if (err)
4325 		return err;
4326 	while (size > conf->max_nr_stripes) {
4327 		if (grow_one_stripe(conf))
4328 			conf->max_nr_stripes++;
4329 		else break;
4330 	}
4331 	return 0;
4332 }
4333 EXPORT_SYMBOL(raid5_set_cache_size);
4334 
4335 static ssize_t
4336 raid5_store_stripe_cache_size(struct mddev *mddev, const char *page, size_t len)
4337 {
4338 	struct r5conf *conf = mddev->private;
4339 	unsigned long new;
4340 	int err;
4341 
4342 	if (len >= PAGE_SIZE)
4343 		return -EINVAL;
4344 	if (!conf)
4345 		return -ENODEV;
4346 
4347 	if (strict_strtoul(page, 10, &new))
4348 		return -EINVAL;
4349 	err = raid5_set_cache_size(mddev, new);
4350 	if (err)
4351 		return err;
4352 	return len;
4353 }
4354 
4355 static struct md_sysfs_entry
4356 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4357 				raid5_show_stripe_cache_size,
4358 				raid5_store_stripe_cache_size);
4359 
4360 static ssize_t
4361 raid5_show_preread_threshold(struct mddev *mddev, char *page)
4362 {
4363 	struct r5conf *conf = mddev->private;
4364 	if (conf)
4365 		return sprintf(page, "%d\n", conf->bypass_threshold);
4366 	else
4367 		return 0;
4368 }
4369 
4370 static ssize_t
4371 raid5_store_preread_threshold(struct mddev *mddev, const char *page, size_t len)
4372 {
4373 	struct r5conf *conf = mddev->private;
4374 	unsigned long new;
4375 	if (len >= PAGE_SIZE)
4376 		return -EINVAL;
4377 	if (!conf)
4378 		return -ENODEV;
4379 
4380 	if (strict_strtoul(page, 10, &new))
4381 		return -EINVAL;
4382 	if (new > conf->max_nr_stripes)
4383 		return -EINVAL;
4384 	conf->bypass_threshold = new;
4385 	return len;
4386 }
4387 
4388 static struct md_sysfs_entry
4389 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4390 					S_IRUGO | S_IWUSR,
4391 					raid5_show_preread_threshold,
4392 					raid5_store_preread_threshold);
4393 
4394 static ssize_t
4395 stripe_cache_active_show(struct mddev *mddev, char *page)
4396 {
4397 	struct r5conf *conf = mddev->private;
4398 	if (conf)
4399 		return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4400 	else
4401 		return 0;
4402 }
4403 
4404 static struct md_sysfs_entry
4405 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4406 
4407 static struct attribute *raid5_attrs[] =  {
4408 	&raid5_stripecache_size.attr,
4409 	&raid5_stripecache_active.attr,
4410 	&raid5_preread_bypass_threshold.attr,
4411 	NULL,
4412 };
4413 static struct attribute_group raid5_attrs_group = {
4414 	.name = NULL,
4415 	.attrs = raid5_attrs,
4416 };
4417 
4418 static sector_t
4419 raid5_size(struct mddev *mddev, sector_t sectors, int raid_disks)
4420 {
4421 	struct r5conf *conf = mddev->private;
4422 
4423 	if (!sectors)
4424 		sectors = mddev->dev_sectors;
4425 	if (!raid_disks)
4426 		/* size is defined by the smallest of previous and new size */
4427 		raid_disks = min(conf->raid_disks, conf->previous_raid_disks);
4428 
4429 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4430 	sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4431 	return sectors * (raid_disks - conf->max_degraded);
4432 }
4433 
4434 static void raid5_free_percpu(struct r5conf *conf)
4435 {
4436 	struct raid5_percpu *percpu;
4437 	unsigned long cpu;
4438 
4439 	if (!conf->percpu)
4440 		return;
4441 
4442 	get_online_cpus();
4443 	for_each_possible_cpu(cpu) {
4444 		percpu = per_cpu_ptr(conf->percpu, cpu);
4445 		safe_put_page(percpu->spare_page);
4446 		kfree(percpu->scribble);
4447 	}
4448 #ifdef CONFIG_HOTPLUG_CPU
4449 	unregister_cpu_notifier(&conf->cpu_notify);
4450 #endif
4451 	put_online_cpus();
4452 
4453 	free_percpu(conf->percpu);
4454 }
4455 
4456 static void free_conf(struct r5conf *conf)
4457 {
4458 	shrink_stripes(conf);
4459 	raid5_free_percpu(conf);
4460 	kfree(conf->disks);
4461 	kfree(conf->stripe_hashtbl);
4462 	kfree(conf);
4463 }
4464 
4465 #ifdef CONFIG_HOTPLUG_CPU
4466 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4467 			      void *hcpu)
4468 {
4469 	struct r5conf *conf = container_of(nfb, struct r5conf, cpu_notify);
4470 	long cpu = (long)hcpu;
4471 	struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4472 
4473 	switch (action) {
4474 	case CPU_UP_PREPARE:
4475 	case CPU_UP_PREPARE_FROZEN:
4476 		if (conf->level == 6 && !percpu->spare_page)
4477 			percpu->spare_page = alloc_page(GFP_KERNEL);
4478 		if (!percpu->scribble)
4479 			percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4480 
4481 		if (!percpu->scribble ||
4482 		    (conf->level == 6 && !percpu->spare_page)) {
4483 			safe_put_page(percpu->spare_page);
4484 			kfree(percpu->scribble);
4485 			pr_err("%s: failed memory allocation for cpu%ld\n",
4486 			       __func__, cpu);
4487 			return notifier_from_errno(-ENOMEM);
4488 		}
4489 		break;
4490 	case CPU_DEAD:
4491 	case CPU_DEAD_FROZEN:
4492 		safe_put_page(percpu->spare_page);
4493 		kfree(percpu->scribble);
4494 		percpu->spare_page = NULL;
4495 		percpu->scribble = NULL;
4496 		break;
4497 	default:
4498 		break;
4499 	}
4500 	return NOTIFY_OK;
4501 }
4502 #endif
4503 
4504 static int raid5_alloc_percpu(struct r5conf *conf)
4505 {
4506 	unsigned long cpu;
4507 	struct page *spare_page;
4508 	struct raid5_percpu __percpu *allcpus;
4509 	void *scribble;
4510 	int err;
4511 
4512 	allcpus = alloc_percpu(struct raid5_percpu);
4513 	if (!allcpus)
4514 		return -ENOMEM;
4515 	conf->percpu = allcpus;
4516 
4517 	get_online_cpus();
4518 	err = 0;
4519 	for_each_present_cpu(cpu) {
4520 		if (conf->level == 6) {
4521 			spare_page = alloc_page(GFP_KERNEL);
4522 			if (!spare_page) {
4523 				err = -ENOMEM;
4524 				break;
4525 			}
4526 			per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4527 		}
4528 		scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4529 		if (!scribble) {
4530 			err = -ENOMEM;
4531 			break;
4532 		}
4533 		per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4534 	}
4535 #ifdef CONFIG_HOTPLUG_CPU
4536 	conf->cpu_notify.notifier_call = raid456_cpu_notify;
4537 	conf->cpu_notify.priority = 0;
4538 	if (err == 0)
4539 		err = register_cpu_notifier(&conf->cpu_notify);
4540 #endif
4541 	put_online_cpus();
4542 
4543 	return err;
4544 }
4545 
4546 static struct r5conf *setup_conf(struct mddev *mddev)
4547 {
4548 	struct r5conf *conf;
4549 	int raid_disk, memory, max_disks;
4550 	struct md_rdev *rdev;
4551 	struct disk_info *disk;
4552 
4553 	if (mddev->new_level != 5
4554 	    && mddev->new_level != 4
4555 	    && mddev->new_level != 6) {
4556 		printk(KERN_ERR "md/raid:%s: raid level not set to 4/5/6 (%d)\n",
4557 		       mdname(mddev), mddev->new_level);
4558 		return ERR_PTR(-EIO);
4559 	}
4560 	if ((mddev->new_level == 5
4561 	     && !algorithm_valid_raid5(mddev->new_layout)) ||
4562 	    (mddev->new_level == 6
4563 	     && !algorithm_valid_raid6(mddev->new_layout))) {
4564 		printk(KERN_ERR "md/raid:%s: layout %d not supported\n",
4565 		       mdname(mddev), mddev->new_layout);
4566 		return ERR_PTR(-EIO);
4567 	}
4568 	if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4569 		printk(KERN_ERR "md/raid:%s: not enough configured devices (%d, minimum 4)\n",
4570 		       mdname(mddev), mddev->raid_disks);
4571 		return ERR_PTR(-EINVAL);
4572 	}
4573 
4574 	if (!mddev->new_chunk_sectors ||
4575 	    (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4576 	    !is_power_of_2(mddev->new_chunk_sectors)) {
4577 		printk(KERN_ERR "md/raid:%s: invalid chunk size %d\n",
4578 		       mdname(mddev), mddev->new_chunk_sectors << 9);
4579 		return ERR_PTR(-EINVAL);
4580 	}
4581 
4582 	conf = kzalloc(sizeof(struct r5conf), GFP_KERNEL);
4583 	if (conf == NULL)
4584 		goto abort;
4585 	spin_lock_init(&conf->device_lock);
4586 	init_waitqueue_head(&conf->wait_for_stripe);
4587 	init_waitqueue_head(&conf->wait_for_overlap);
4588 	INIT_LIST_HEAD(&conf->handle_list);
4589 	INIT_LIST_HEAD(&conf->hold_list);
4590 	INIT_LIST_HEAD(&conf->delayed_list);
4591 	INIT_LIST_HEAD(&conf->bitmap_list);
4592 	INIT_LIST_HEAD(&conf->inactive_list);
4593 	atomic_set(&conf->active_stripes, 0);
4594 	atomic_set(&conf->preread_active_stripes, 0);
4595 	atomic_set(&conf->active_aligned_reads, 0);
4596 	conf->bypass_threshold = BYPASS_THRESHOLD;
4597 	conf->recovery_disabled = mddev->recovery_disabled - 1;
4598 
4599 	conf->raid_disks = mddev->raid_disks;
4600 	if (mddev->reshape_position == MaxSector)
4601 		conf->previous_raid_disks = mddev->raid_disks;
4602 	else
4603 		conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4604 	max_disks = max(conf->raid_disks, conf->previous_raid_disks);
4605 	conf->scribble_len = scribble_len(max_disks);
4606 
4607 	conf->disks = kzalloc(max_disks * sizeof(struct disk_info),
4608 			      GFP_KERNEL);
4609 	if (!conf->disks)
4610 		goto abort;
4611 
4612 	conf->mddev = mddev;
4613 
4614 	if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4615 		goto abort;
4616 
4617 	conf->level = mddev->new_level;
4618 	if (raid5_alloc_percpu(conf) != 0)
4619 		goto abort;
4620 
4621 	pr_debug("raid456: run(%s) called.\n", mdname(mddev));
4622 
4623 	list_for_each_entry(rdev, &mddev->disks, same_set) {
4624 		raid_disk = rdev->raid_disk;
4625 		if (raid_disk >= max_disks
4626 		    || raid_disk < 0)
4627 			continue;
4628 		disk = conf->disks + raid_disk;
4629 
4630 		disk->rdev = rdev;
4631 
4632 		if (test_bit(In_sync, &rdev->flags)) {
4633 			char b[BDEVNAME_SIZE];
4634 			printk(KERN_INFO "md/raid:%s: device %s operational as raid"
4635 			       " disk %d\n",
4636 			       mdname(mddev), bdevname(rdev->bdev, b), raid_disk);
4637 		} else if (rdev->saved_raid_disk != raid_disk)
4638 			/* Cannot rely on bitmap to complete recovery */
4639 			conf->fullsync = 1;
4640 	}
4641 
4642 	conf->chunk_sectors = mddev->new_chunk_sectors;
4643 	conf->level = mddev->new_level;
4644 	if (conf->level == 6)
4645 		conf->max_degraded = 2;
4646 	else
4647 		conf->max_degraded = 1;
4648 	conf->algorithm = mddev->new_layout;
4649 	conf->max_nr_stripes = NR_STRIPES;
4650 	conf->reshape_progress = mddev->reshape_position;
4651 	if (conf->reshape_progress != MaxSector) {
4652 		conf->prev_chunk_sectors = mddev->chunk_sectors;
4653 		conf->prev_algo = mddev->layout;
4654 	}
4655 
4656 	memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4657 		 max_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4658 	if (grow_stripes(conf, conf->max_nr_stripes)) {
4659 		printk(KERN_ERR
4660 		       "md/raid:%s: couldn't allocate %dkB for buffers\n",
4661 		       mdname(mddev), memory);
4662 		goto abort;
4663 	} else
4664 		printk(KERN_INFO "md/raid:%s: allocated %dkB\n",
4665 		       mdname(mddev), memory);
4666 
4667 	conf->thread = md_register_thread(raid5d, mddev, NULL);
4668 	if (!conf->thread) {
4669 		printk(KERN_ERR
4670 		       "md/raid:%s: couldn't allocate thread.\n",
4671 		       mdname(mddev));
4672 		goto abort;
4673 	}
4674 
4675 	return conf;
4676 
4677  abort:
4678 	if (conf) {
4679 		free_conf(conf);
4680 		return ERR_PTR(-EIO);
4681 	} else
4682 		return ERR_PTR(-ENOMEM);
4683 }
4684 
4685 
4686 static int only_parity(int raid_disk, int algo, int raid_disks, int max_degraded)
4687 {
4688 	switch (algo) {
4689 	case ALGORITHM_PARITY_0:
4690 		if (raid_disk < max_degraded)
4691 			return 1;
4692 		break;
4693 	case ALGORITHM_PARITY_N:
4694 		if (raid_disk >= raid_disks - max_degraded)
4695 			return 1;
4696 		break;
4697 	case ALGORITHM_PARITY_0_6:
4698 		if (raid_disk == 0 ||
4699 		    raid_disk == raid_disks - 1)
4700 			return 1;
4701 		break;
4702 	case ALGORITHM_LEFT_ASYMMETRIC_6:
4703 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
4704 	case ALGORITHM_LEFT_SYMMETRIC_6:
4705 	case ALGORITHM_RIGHT_SYMMETRIC_6:
4706 		if (raid_disk == raid_disks - 1)
4707 			return 1;
4708 	}
4709 	return 0;
4710 }
4711 
4712 static int run(struct mddev *mddev)
4713 {
4714 	struct r5conf *conf;
4715 	int working_disks = 0;
4716 	int dirty_parity_disks = 0;
4717 	struct md_rdev *rdev;
4718 	sector_t reshape_offset = 0;
4719 
4720 	if (mddev->recovery_cp != MaxSector)
4721 		printk(KERN_NOTICE "md/raid:%s: not clean"
4722 		       " -- starting background reconstruction\n",
4723 		       mdname(mddev));
4724 	if (mddev->reshape_position != MaxSector) {
4725 		/* Check that we can continue the reshape.
4726 		 * Currently only disks can change, it must
4727 		 * increase, and we must be past the point where
4728 		 * a stripe over-writes itself
4729 		 */
4730 		sector_t here_new, here_old;
4731 		int old_disks;
4732 		int max_degraded = (mddev->level == 6 ? 2 : 1);
4733 
4734 		if (mddev->new_level != mddev->level) {
4735 			printk(KERN_ERR "md/raid:%s: unsupported reshape "
4736 			       "required - aborting.\n",
4737 			       mdname(mddev));
4738 			return -EINVAL;
4739 		}
4740 		old_disks = mddev->raid_disks - mddev->delta_disks;
4741 		/* reshape_position must be on a new-stripe boundary, and one
4742 		 * further up in new geometry must map after here in old
4743 		 * geometry.
4744 		 */
4745 		here_new = mddev->reshape_position;
4746 		if (sector_div(here_new, mddev->new_chunk_sectors *
4747 			       (mddev->raid_disks - max_degraded))) {
4748 			printk(KERN_ERR "md/raid:%s: reshape_position not "
4749 			       "on a stripe boundary\n", mdname(mddev));
4750 			return -EINVAL;
4751 		}
4752 		reshape_offset = here_new * mddev->new_chunk_sectors;
4753 		/* here_new is the stripe we will write to */
4754 		here_old = mddev->reshape_position;
4755 		sector_div(here_old, mddev->chunk_sectors *
4756 			   (old_disks-max_degraded));
4757 		/* here_old is the first stripe that we might need to read
4758 		 * from */
4759 		if (mddev->delta_disks == 0) {
4760 			/* We cannot be sure it is safe to start an in-place
4761 			 * reshape.  It is only safe if user-space if monitoring
4762 			 * and taking constant backups.
4763 			 * mdadm always starts a situation like this in
4764 			 * readonly mode so it can take control before
4765 			 * allowing any writes.  So just check for that.
4766 			 */
4767 			if ((here_new * mddev->new_chunk_sectors !=
4768 			     here_old * mddev->chunk_sectors) ||
4769 			    mddev->ro == 0) {
4770 				printk(KERN_ERR "md/raid:%s: in-place reshape must be started"
4771 				       " in read-only mode - aborting\n",
4772 				       mdname(mddev));
4773 				return -EINVAL;
4774 			}
4775 		} else if (mddev->delta_disks < 0
4776 		    ? (here_new * mddev->new_chunk_sectors <=
4777 		       here_old * mddev->chunk_sectors)
4778 		    : (here_new * mddev->new_chunk_sectors >=
4779 		       here_old * mddev->chunk_sectors)) {
4780 			/* Reading from the same stripe as writing to - bad */
4781 			printk(KERN_ERR "md/raid:%s: reshape_position too early for "
4782 			       "auto-recovery - aborting.\n",
4783 			       mdname(mddev));
4784 			return -EINVAL;
4785 		}
4786 		printk(KERN_INFO "md/raid:%s: reshape will continue\n",
4787 		       mdname(mddev));
4788 		/* OK, we should be able to continue; */
4789 	} else {
4790 		BUG_ON(mddev->level != mddev->new_level);
4791 		BUG_ON(mddev->layout != mddev->new_layout);
4792 		BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4793 		BUG_ON(mddev->delta_disks != 0);
4794 	}
4795 
4796 	if (mddev->private == NULL)
4797 		conf = setup_conf(mddev);
4798 	else
4799 		conf = mddev->private;
4800 
4801 	if (IS_ERR(conf))
4802 		return PTR_ERR(conf);
4803 
4804 	mddev->thread = conf->thread;
4805 	conf->thread = NULL;
4806 	mddev->private = conf;
4807 
4808 	/*
4809 	 * 0 for a fully functional array, 1 or 2 for a degraded array.
4810 	 */
4811 	list_for_each_entry(rdev, &mddev->disks, same_set) {
4812 		if (rdev->raid_disk < 0)
4813 			continue;
4814 		if (test_bit(In_sync, &rdev->flags)) {
4815 			working_disks++;
4816 			continue;
4817 		}
4818 		/* This disc is not fully in-sync.  However if it
4819 		 * just stored parity (beyond the recovery_offset),
4820 		 * when we don't need to be concerned about the
4821 		 * array being dirty.
4822 		 * When reshape goes 'backwards', we never have
4823 		 * partially completed devices, so we only need
4824 		 * to worry about reshape going forwards.
4825 		 */
4826 		/* Hack because v0.91 doesn't store recovery_offset properly. */
4827 		if (mddev->major_version == 0 &&
4828 		    mddev->minor_version > 90)
4829 			rdev->recovery_offset = reshape_offset;
4830 
4831 		if (rdev->recovery_offset < reshape_offset) {
4832 			/* We need to check old and new layout */
4833 			if (!only_parity(rdev->raid_disk,
4834 					 conf->algorithm,
4835 					 conf->raid_disks,
4836 					 conf->max_degraded))
4837 				continue;
4838 		}
4839 		if (!only_parity(rdev->raid_disk,
4840 				 conf->prev_algo,
4841 				 conf->previous_raid_disks,
4842 				 conf->max_degraded))
4843 			continue;
4844 		dirty_parity_disks++;
4845 	}
4846 
4847 	mddev->degraded = (max(conf->raid_disks, conf->previous_raid_disks)
4848 			   - working_disks);
4849 
4850 	if (has_failed(conf)) {
4851 		printk(KERN_ERR "md/raid:%s: not enough operational devices"
4852 			" (%d/%d failed)\n",
4853 			mdname(mddev), mddev->degraded, conf->raid_disks);
4854 		goto abort;
4855 	}
4856 
4857 	/* device size must be a multiple of chunk size */
4858 	mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4859 	mddev->resync_max_sectors = mddev->dev_sectors;
4860 
4861 	if (mddev->degraded > dirty_parity_disks &&
4862 	    mddev->recovery_cp != MaxSector) {
4863 		if (mddev->ok_start_degraded)
4864 			printk(KERN_WARNING
4865 			       "md/raid:%s: starting dirty degraded array"
4866 			       " - data corruption possible.\n",
4867 			       mdname(mddev));
4868 		else {
4869 			printk(KERN_ERR
4870 			       "md/raid:%s: cannot start dirty degraded array.\n",
4871 			       mdname(mddev));
4872 			goto abort;
4873 		}
4874 	}
4875 
4876 	if (mddev->degraded == 0)
4877 		printk(KERN_INFO "md/raid:%s: raid level %d active with %d out of %d"
4878 		       " devices, algorithm %d\n", mdname(mddev), conf->level,
4879 		       mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4880 		       mddev->new_layout);
4881 	else
4882 		printk(KERN_ALERT "md/raid:%s: raid level %d active with %d"
4883 		       " out of %d devices, algorithm %d\n",
4884 		       mdname(mddev), conf->level,
4885 		       mddev->raid_disks - mddev->degraded,
4886 		       mddev->raid_disks, mddev->new_layout);
4887 
4888 	print_raid5_conf(conf);
4889 
4890 	if (conf->reshape_progress != MaxSector) {
4891 		conf->reshape_safe = conf->reshape_progress;
4892 		atomic_set(&conf->reshape_stripes, 0);
4893 		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4894 		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4895 		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4896 		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4897 		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4898 							"reshape");
4899 	}
4900 
4901 
4902 	/* Ok, everything is just fine now */
4903 	if (mddev->to_remove == &raid5_attrs_group)
4904 		mddev->to_remove = NULL;
4905 	else if (mddev->kobj.sd &&
4906 	    sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4907 		printk(KERN_WARNING
4908 		       "raid5: failed to create sysfs attributes for %s\n",
4909 		       mdname(mddev));
4910 	md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4911 
4912 	if (mddev->queue) {
4913 		int chunk_size;
4914 		/* read-ahead size must cover two whole stripes, which
4915 		 * is 2 * (datadisks) * chunksize where 'n' is the
4916 		 * number of raid devices
4917 		 */
4918 		int data_disks = conf->previous_raid_disks - conf->max_degraded;
4919 		int stripe = data_disks *
4920 			((mddev->chunk_sectors << 9) / PAGE_SIZE);
4921 		if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4922 			mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4923 
4924 		blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4925 
4926 		mddev->queue->backing_dev_info.congested_data = mddev;
4927 		mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4928 
4929 		chunk_size = mddev->chunk_sectors << 9;
4930 		blk_queue_io_min(mddev->queue, chunk_size);
4931 		blk_queue_io_opt(mddev->queue, chunk_size *
4932 				 (conf->raid_disks - conf->max_degraded));
4933 
4934 		list_for_each_entry(rdev, &mddev->disks, same_set)
4935 			disk_stack_limits(mddev->gendisk, rdev->bdev,
4936 					  rdev->data_offset << 9);
4937 	}
4938 
4939 	return 0;
4940 abort:
4941 	md_unregister_thread(&mddev->thread);
4942 	print_raid5_conf(conf);
4943 	free_conf(conf);
4944 	mddev->private = NULL;
4945 	printk(KERN_ALERT "md/raid:%s: failed to run raid set.\n", mdname(mddev));
4946 	return -EIO;
4947 }
4948 
4949 static int stop(struct mddev *mddev)
4950 {
4951 	struct r5conf *conf = mddev->private;
4952 
4953 	md_unregister_thread(&mddev->thread);
4954 	if (mddev->queue)
4955 		mddev->queue->backing_dev_info.congested_fn = NULL;
4956 	free_conf(conf);
4957 	mddev->private = NULL;
4958 	mddev->to_remove = &raid5_attrs_group;
4959 	return 0;
4960 }
4961 
4962 static void status(struct seq_file *seq, struct mddev *mddev)
4963 {
4964 	struct r5conf *conf = mddev->private;
4965 	int i;
4966 
4967 	seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
4968 		mddev->chunk_sectors / 2, mddev->layout);
4969 	seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4970 	for (i = 0; i < conf->raid_disks; i++)
4971 		seq_printf (seq, "%s",
4972 			       conf->disks[i].rdev &&
4973 			       test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4974 	seq_printf (seq, "]");
4975 }
4976 
4977 static void print_raid5_conf (struct r5conf *conf)
4978 {
4979 	int i;
4980 	struct disk_info *tmp;
4981 
4982 	printk(KERN_DEBUG "RAID conf printout:\n");
4983 	if (!conf) {
4984 		printk("(conf==NULL)\n");
4985 		return;
4986 	}
4987 	printk(KERN_DEBUG " --- level:%d rd:%d wd:%d\n", conf->level,
4988 	       conf->raid_disks,
4989 	       conf->raid_disks - conf->mddev->degraded);
4990 
4991 	for (i = 0; i < conf->raid_disks; i++) {
4992 		char b[BDEVNAME_SIZE];
4993 		tmp = conf->disks + i;
4994 		if (tmp->rdev)
4995 			printk(KERN_DEBUG " disk %d, o:%d, dev:%s\n",
4996 			       i, !test_bit(Faulty, &tmp->rdev->flags),
4997 			       bdevname(tmp->rdev->bdev, b));
4998 	}
4999 }
5000 
5001 static int raid5_spare_active(struct mddev *mddev)
5002 {
5003 	int i;
5004 	struct r5conf *conf = mddev->private;
5005 	struct disk_info *tmp;
5006 	int count = 0;
5007 	unsigned long flags;
5008 
5009 	for (i = 0; i < conf->raid_disks; i++) {
5010 		tmp = conf->disks + i;
5011 		if (tmp->rdev
5012 		    && tmp->rdev->recovery_offset == MaxSector
5013 		    && !test_bit(Faulty, &tmp->rdev->flags)
5014 		    && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5015 			count++;
5016 			sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
5017 		}
5018 	}
5019 	spin_lock_irqsave(&conf->device_lock, flags);
5020 	mddev->degraded -= count;
5021 	spin_unlock_irqrestore(&conf->device_lock, flags);
5022 	print_raid5_conf(conf);
5023 	return count;
5024 }
5025 
5026 static int raid5_remove_disk(struct mddev *mddev, int number)
5027 {
5028 	struct r5conf *conf = mddev->private;
5029 	int err = 0;
5030 	struct md_rdev *rdev;
5031 	struct disk_info *p = conf->disks + number;
5032 
5033 	print_raid5_conf(conf);
5034 	rdev = p->rdev;
5035 	if (rdev) {
5036 		if (number >= conf->raid_disks &&
5037 		    conf->reshape_progress == MaxSector)
5038 			clear_bit(In_sync, &rdev->flags);
5039 
5040 		if (test_bit(In_sync, &rdev->flags) ||
5041 		    atomic_read(&rdev->nr_pending)) {
5042 			err = -EBUSY;
5043 			goto abort;
5044 		}
5045 		/* Only remove non-faulty devices if recovery
5046 		 * isn't possible.
5047 		 */
5048 		if (!test_bit(Faulty, &rdev->flags) &&
5049 		    mddev->recovery_disabled != conf->recovery_disabled &&
5050 		    !has_failed(conf) &&
5051 		    number < conf->raid_disks) {
5052 			err = -EBUSY;
5053 			goto abort;
5054 		}
5055 		p->rdev = NULL;
5056 		synchronize_rcu();
5057 		if (atomic_read(&rdev->nr_pending)) {
5058 			/* lost the race, try later */
5059 			err = -EBUSY;
5060 			p->rdev = rdev;
5061 		}
5062 	}
5063 abort:
5064 
5065 	print_raid5_conf(conf);
5066 	return err;
5067 }
5068 
5069 static int raid5_add_disk(struct mddev *mddev, struct md_rdev *rdev)
5070 {
5071 	struct r5conf *conf = mddev->private;
5072 	int err = -EEXIST;
5073 	int disk;
5074 	struct disk_info *p;
5075 	int first = 0;
5076 	int last = conf->raid_disks - 1;
5077 
5078 	if (mddev->recovery_disabled == conf->recovery_disabled)
5079 		return -EBUSY;
5080 
5081 	if (has_failed(conf))
5082 		/* no point adding a device */
5083 		return -EINVAL;
5084 
5085 	if (rdev->raid_disk >= 0)
5086 		first = last = rdev->raid_disk;
5087 
5088 	/*
5089 	 * find the disk ... but prefer rdev->saved_raid_disk
5090 	 * if possible.
5091 	 */
5092 	if (rdev->saved_raid_disk >= 0 &&
5093 	    rdev->saved_raid_disk >= first &&
5094 	    conf->disks[rdev->saved_raid_disk].rdev == NULL)
5095 		disk = rdev->saved_raid_disk;
5096 	else
5097 		disk = first;
5098 	for ( ; disk <= last ; disk++)
5099 		if ((p=conf->disks + disk)->rdev == NULL) {
5100 			clear_bit(In_sync, &rdev->flags);
5101 			rdev->raid_disk = disk;
5102 			err = 0;
5103 			if (rdev->saved_raid_disk != disk)
5104 				conf->fullsync = 1;
5105 			rcu_assign_pointer(p->rdev, rdev);
5106 			break;
5107 		}
5108 	print_raid5_conf(conf);
5109 	return err;
5110 }
5111 
5112 static int raid5_resize(struct mddev *mddev, sector_t sectors)
5113 {
5114 	/* no resync is happening, and there is enough space
5115 	 * on all devices, so we can resize.
5116 	 * We need to make sure resync covers any new space.
5117 	 * If the array is shrinking we should possibly wait until
5118 	 * any io in the removed space completes, but it hardly seems
5119 	 * worth it.
5120 	 */
5121 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5122 	md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5123 					       mddev->raid_disks));
5124 	if (mddev->array_sectors >
5125 	    raid5_size(mddev, sectors, mddev->raid_disks))
5126 		return -EINVAL;
5127 	set_capacity(mddev->gendisk, mddev->array_sectors);
5128 	revalidate_disk(mddev->gendisk);
5129 	if (sectors > mddev->dev_sectors &&
5130 	    mddev->recovery_cp > mddev->dev_sectors) {
5131 		mddev->recovery_cp = mddev->dev_sectors;
5132 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5133 	}
5134 	mddev->dev_sectors = sectors;
5135 	mddev->resync_max_sectors = sectors;
5136 	return 0;
5137 }
5138 
5139 static int check_stripe_cache(struct mddev *mddev)
5140 {
5141 	/* Can only proceed if there are plenty of stripe_heads.
5142 	 * We need a minimum of one full stripe,, and for sensible progress
5143 	 * it is best to have about 4 times that.
5144 	 * If we require 4 times, then the default 256 4K stripe_heads will
5145 	 * allow for chunk sizes up to 256K, which is probably OK.
5146 	 * If the chunk size is greater, user-space should request more
5147 	 * stripe_heads first.
5148 	 */
5149 	struct r5conf *conf = mddev->private;
5150 	if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5151 	    > conf->max_nr_stripes ||
5152 	    ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5153 	    > conf->max_nr_stripes) {
5154 		printk(KERN_WARNING "md/raid:%s: reshape: not enough stripes.  Needed %lu\n",
5155 		       mdname(mddev),
5156 		       ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5157 			/ STRIPE_SIZE)*4);
5158 		return 0;
5159 	}
5160 	return 1;
5161 }
5162 
5163 static int check_reshape(struct mddev *mddev)
5164 {
5165 	struct r5conf *conf = mddev->private;
5166 
5167 	if (mddev->delta_disks == 0 &&
5168 	    mddev->new_layout == mddev->layout &&
5169 	    mddev->new_chunk_sectors == mddev->chunk_sectors)
5170 		return 0; /* nothing to do */
5171 	if (mddev->bitmap)
5172 		/* Cannot grow a bitmap yet */
5173 		return -EBUSY;
5174 	if (has_failed(conf))
5175 		return -EINVAL;
5176 	if (mddev->delta_disks < 0) {
5177 		/* We might be able to shrink, but the devices must
5178 		 * be made bigger first.
5179 		 * For raid6, 4 is the minimum size.
5180 		 * Otherwise 2 is the minimum
5181 		 */
5182 		int min = 2;
5183 		if (mddev->level == 6)
5184 			min = 4;
5185 		if (mddev->raid_disks + mddev->delta_disks < min)
5186 			return -EINVAL;
5187 	}
5188 
5189 	if (!check_stripe_cache(mddev))
5190 		return -ENOSPC;
5191 
5192 	return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5193 }
5194 
5195 static int raid5_start_reshape(struct mddev *mddev)
5196 {
5197 	struct r5conf *conf = mddev->private;
5198 	struct md_rdev *rdev;
5199 	int spares = 0;
5200 	unsigned long flags;
5201 
5202 	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5203 		return -EBUSY;
5204 
5205 	if (!check_stripe_cache(mddev))
5206 		return -ENOSPC;
5207 
5208 	list_for_each_entry(rdev, &mddev->disks, same_set)
5209 		if (!test_bit(In_sync, &rdev->flags)
5210 		    && !test_bit(Faulty, &rdev->flags))
5211 			spares++;
5212 
5213 	if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5214 		/* Not enough devices even to make a degraded array
5215 		 * of that size
5216 		 */
5217 		return -EINVAL;
5218 
5219 	/* Refuse to reduce size of the array.  Any reductions in
5220 	 * array size must be through explicit setting of array_size
5221 	 * attribute.
5222 	 */
5223 	if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5224 	    < mddev->array_sectors) {
5225 		printk(KERN_ERR "md/raid:%s: array size must be reduced "
5226 		       "before number of disks\n", mdname(mddev));
5227 		return -EINVAL;
5228 	}
5229 
5230 	atomic_set(&conf->reshape_stripes, 0);
5231 	spin_lock_irq(&conf->device_lock);
5232 	conf->previous_raid_disks = conf->raid_disks;
5233 	conf->raid_disks += mddev->delta_disks;
5234 	conf->prev_chunk_sectors = conf->chunk_sectors;
5235 	conf->chunk_sectors = mddev->new_chunk_sectors;
5236 	conf->prev_algo = conf->algorithm;
5237 	conf->algorithm = mddev->new_layout;
5238 	if (mddev->delta_disks < 0)
5239 		conf->reshape_progress = raid5_size(mddev, 0, 0);
5240 	else
5241 		conf->reshape_progress = 0;
5242 	conf->reshape_safe = conf->reshape_progress;
5243 	conf->generation++;
5244 	spin_unlock_irq(&conf->device_lock);
5245 
5246 	/* Add some new drives, as many as will fit.
5247 	 * We know there are enough to make the newly sized array work.
5248 	 * Don't add devices if we are reducing the number of
5249 	 * devices in the array.  This is because it is not possible
5250 	 * to correctly record the "partially reconstructed" state of
5251 	 * such devices during the reshape and confusion could result.
5252 	 */
5253 	if (mddev->delta_disks >= 0) {
5254 		int added_devices = 0;
5255 		list_for_each_entry(rdev, &mddev->disks, same_set)
5256 			if (rdev->raid_disk < 0 &&
5257 			    !test_bit(Faulty, &rdev->flags)) {
5258 				if (raid5_add_disk(mddev, rdev) == 0) {
5259 					if (rdev->raid_disk
5260 					    >= conf->previous_raid_disks) {
5261 						set_bit(In_sync, &rdev->flags);
5262 						added_devices++;
5263 					} else
5264 						rdev->recovery_offset = 0;
5265 
5266 					if (sysfs_link_rdev(mddev, rdev))
5267 						/* Failure here is OK */;
5268 				}
5269 			} else if (rdev->raid_disk >= conf->previous_raid_disks
5270 				   && !test_bit(Faulty, &rdev->flags)) {
5271 				/* This is a spare that was manually added */
5272 				set_bit(In_sync, &rdev->flags);
5273 				added_devices++;
5274 			}
5275 
5276 		/* When a reshape changes the number of devices,
5277 		 * ->degraded is measured against the larger of the
5278 		 * pre and post number of devices.
5279 		 */
5280 		spin_lock_irqsave(&conf->device_lock, flags);
5281 		mddev->degraded += (conf->raid_disks - conf->previous_raid_disks)
5282 			- added_devices;
5283 		spin_unlock_irqrestore(&conf->device_lock, flags);
5284 	}
5285 	mddev->raid_disks = conf->raid_disks;
5286 	mddev->reshape_position = conf->reshape_progress;
5287 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
5288 
5289 	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5290 	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5291 	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5292 	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5293 	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5294 						"reshape");
5295 	if (!mddev->sync_thread) {
5296 		mddev->recovery = 0;
5297 		spin_lock_irq(&conf->device_lock);
5298 		mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5299 		conf->reshape_progress = MaxSector;
5300 		spin_unlock_irq(&conf->device_lock);
5301 		return -EAGAIN;
5302 	}
5303 	conf->reshape_checkpoint = jiffies;
5304 	md_wakeup_thread(mddev->sync_thread);
5305 	md_new_event(mddev);
5306 	return 0;
5307 }
5308 
5309 /* This is called from the reshape thread and should make any
5310  * changes needed in 'conf'
5311  */
5312 static void end_reshape(struct r5conf *conf)
5313 {
5314 
5315 	if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5316 
5317 		spin_lock_irq(&conf->device_lock);
5318 		conf->previous_raid_disks = conf->raid_disks;
5319 		conf->reshape_progress = MaxSector;
5320 		spin_unlock_irq(&conf->device_lock);
5321 		wake_up(&conf->wait_for_overlap);
5322 
5323 		/* read-ahead size must cover two whole stripes, which is
5324 		 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5325 		 */
5326 		if (conf->mddev->queue) {
5327 			int data_disks = conf->raid_disks - conf->max_degraded;
5328 			int stripe = data_disks * ((conf->chunk_sectors << 9)
5329 						   / PAGE_SIZE);
5330 			if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5331 				conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5332 		}
5333 	}
5334 }
5335 
5336 /* This is called from the raid5d thread with mddev_lock held.
5337  * It makes config changes to the device.
5338  */
5339 static void raid5_finish_reshape(struct mddev *mddev)
5340 {
5341 	struct r5conf *conf = mddev->private;
5342 
5343 	if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5344 
5345 		if (mddev->delta_disks > 0) {
5346 			md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5347 			set_capacity(mddev->gendisk, mddev->array_sectors);
5348 			revalidate_disk(mddev->gendisk);
5349 		} else {
5350 			int d;
5351 			mddev->degraded = conf->raid_disks;
5352 			for (d = 0; d < conf->raid_disks ; d++)
5353 				if (conf->disks[d].rdev &&
5354 				    test_bit(In_sync,
5355 					     &conf->disks[d].rdev->flags))
5356 					mddev->degraded--;
5357 			for (d = conf->raid_disks ;
5358 			     d < conf->raid_disks - mddev->delta_disks;
5359 			     d++) {
5360 				struct md_rdev *rdev = conf->disks[d].rdev;
5361 				if (rdev && raid5_remove_disk(mddev, d) == 0) {
5362 					sysfs_unlink_rdev(mddev, rdev);
5363 					rdev->raid_disk = -1;
5364 				}
5365 			}
5366 		}
5367 		mddev->layout = conf->algorithm;
5368 		mddev->chunk_sectors = conf->chunk_sectors;
5369 		mddev->reshape_position = MaxSector;
5370 		mddev->delta_disks = 0;
5371 	}
5372 }
5373 
5374 static void raid5_quiesce(struct mddev *mddev, int state)
5375 {
5376 	struct r5conf *conf = mddev->private;
5377 
5378 	switch(state) {
5379 	case 2: /* resume for a suspend */
5380 		wake_up(&conf->wait_for_overlap);
5381 		break;
5382 
5383 	case 1: /* stop all writes */
5384 		spin_lock_irq(&conf->device_lock);
5385 		/* '2' tells resync/reshape to pause so that all
5386 		 * active stripes can drain
5387 		 */
5388 		conf->quiesce = 2;
5389 		wait_event_lock_irq(conf->wait_for_stripe,
5390 				    atomic_read(&conf->active_stripes) == 0 &&
5391 				    atomic_read(&conf->active_aligned_reads) == 0,
5392 				    conf->device_lock, /* nothing */);
5393 		conf->quiesce = 1;
5394 		spin_unlock_irq(&conf->device_lock);
5395 		/* allow reshape to continue */
5396 		wake_up(&conf->wait_for_overlap);
5397 		break;
5398 
5399 	case 0: /* re-enable writes */
5400 		spin_lock_irq(&conf->device_lock);
5401 		conf->quiesce = 0;
5402 		wake_up(&conf->wait_for_stripe);
5403 		wake_up(&conf->wait_for_overlap);
5404 		spin_unlock_irq(&conf->device_lock);
5405 		break;
5406 	}
5407 }
5408 
5409 
5410 static void *raid45_takeover_raid0(struct mddev *mddev, int level)
5411 {
5412 	struct r0conf *raid0_conf = mddev->private;
5413 	sector_t sectors;
5414 
5415 	/* for raid0 takeover only one zone is supported */
5416 	if (raid0_conf->nr_strip_zones > 1) {
5417 		printk(KERN_ERR "md/raid:%s: cannot takeover raid0 with more than one zone.\n",
5418 		       mdname(mddev));
5419 		return ERR_PTR(-EINVAL);
5420 	}
5421 
5422 	sectors = raid0_conf->strip_zone[0].zone_end;
5423 	sector_div(sectors, raid0_conf->strip_zone[0].nb_dev);
5424 	mddev->dev_sectors = sectors;
5425 	mddev->new_level = level;
5426 	mddev->new_layout = ALGORITHM_PARITY_N;
5427 	mddev->new_chunk_sectors = mddev->chunk_sectors;
5428 	mddev->raid_disks += 1;
5429 	mddev->delta_disks = 1;
5430 	/* make sure it will be not marked as dirty */
5431 	mddev->recovery_cp = MaxSector;
5432 
5433 	return setup_conf(mddev);
5434 }
5435 
5436 
5437 static void *raid5_takeover_raid1(struct mddev *mddev)
5438 {
5439 	int chunksect;
5440 
5441 	if (mddev->raid_disks != 2 ||
5442 	    mddev->degraded > 1)
5443 		return ERR_PTR(-EINVAL);
5444 
5445 	/* Should check if there are write-behind devices? */
5446 
5447 	chunksect = 64*2; /* 64K by default */
5448 
5449 	/* The array must be an exact multiple of chunksize */
5450 	while (chunksect && (mddev->array_sectors & (chunksect-1)))
5451 		chunksect >>= 1;
5452 
5453 	if ((chunksect<<9) < STRIPE_SIZE)
5454 		/* array size does not allow a suitable chunk size */
5455 		return ERR_PTR(-EINVAL);
5456 
5457 	mddev->new_level = 5;
5458 	mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5459 	mddev->new_chunk_sectors = chunksect;
5460 
5461 	return setup_conf(mddev);
5462 }
5463 
5464 static void *raid5_takeover_raid6(struct mddev *mddev)
5465 {
5466 	int new_layout;
5467 
5468 	switch (mddev->layout) {
5469 	case ALGORITHM_LEFT_ASYMMETRIC_6:
5470 		new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5471 		break;
5472 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
5473 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5474 		break;
5475 	case ALGORITHM_LEFT_SYMMETRIC_6:
5476 		new_layout = ALGORITHM_LEFT_SYMMETRIC;
5477 		break;
5478 	case ALGORITHM_RIGHT_SYMMETRIC_6:
5479 		new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5480 		break;
5481 	case ALGORITHM_PARITY_0_6:
5482 		new_layout = ALGORITHM_PARITY_0;
5483 		break;
5484 	case ALGORITHM_PARITY_N:
5485 		new_layout = ALGORITHM_PARITY_N;
5486 		break;
5487 	default:
5488 		return ERR_PTR(-EINVAL);
5489 	}
5490 	mddev->new_level = 5;
5491 	mddev->new_layout = new_layout;
5492 	mddev->delta_disks = -1;
5493 	mddev->raid_disks -= 1;
5494 	return setup_conf(mddev);
5495 }
5496 
5497 
5498 static int raid5_check_reshape(struct mddev *mddev)
5499 {
5500 	/* For a 2-drive array, the layout and chunk size can be changed
5501 	 * immediately as not restriping is needed.
5502 	 * For larger arrays we record the new value - after validation
5503 	 * to be used by a reshape pass.
5504 	 */
5505 	struct r5conf *conf = mddev->private;
5506 	int new_chunk = mddev->new_chunk_sectors;
5507 
5508 	if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5509 		return -EINVAL;
5510 	if (new_chunk > 0) {
5511 		if (!is_power_of_2(new_chunk))
5512 			return -EINVAL;
5513 		if (new_chunk < (PAGE_SIZE>>9))
5514 			return -EINVAL;
5515 		if (mddev->array_sectors & (new_chunk-1))
5516 			/* not factor of array size */
5517 			return -EINVAL;
5518 	}
5519 
5520 	/* They look valid */
5521 
5522 	if (mddev->raid_disks == 2) {
5523 		/* can make the change immediately */
5524 		if (mddev->new_layout >= 0) {
5525 			conf->algorithm = mddev->new_layout;
5526 			mddev->layout = mddev->new_layout;
5527 		}
5528 		if (new_chunk > 0) {
5529 			conf->chunk_sectors = new_chunk ;
5530 			mddev->chunk_sectors = new_chunk;
5531 		}
5532 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
5533 		md_wakeup_thread(mddev->thread);
5534 	}
5535 	return check_reshape(mddev);
5536 }
5537 
5538 static int raid6_check_reshape(struct mddev *mddev)
5539 {
5540 	int new_chunk = mddev->new_chunk_sectors;
5541 
5542 	if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5543 		return -EINVAL;
5544 	if (new_chunk > 0) {
5545 		if (!is_power_of_2(new_chunk))
5546 			return -EINVAL;
5547 		if (new_chunk < (PAGE_SIZE >> 9))
5548 			return -EINVAL;
5549 		if (mddev->array_sectors & (new_chunk-1))
5550 			/* not factor of array size */
5551 			return -EINVAL;
5552 	}
5553 
5554 	/* They look valid */
5555 	return check_reshape(mddev);
5556 }
5557 
5558 static void *raid5_takeover(struct mddev *mddev)
5559 {
5560 	/* raid5 can take over:
5561 	 *  raid0 - if there is only one strip zone - make it a raid4 layout
5562 	 *  raid1 - if there are two drives.  We need to know the chunk size
5563 	 *  raid4 - trivial - just use a raid4 layout.
5564 	 *  raid6 - Providing it is a *_6 layout
5565 	 */
5566 	if (mddev->level == 0)
5567 		return raid45_takeover_raid0(mddev, 5);
5568 	if (mddev->level == 1)
5569 		return raid5_takeover_raid1(mddev);
5570 	if (mddev->level == 4) {
5571 		mddev->new_layout = ALGORITHM_PARITY_N;
5572 		mddev->new_level = 5;
5573 		return setup_conf(mddev);
5574 	}
5575 	if (mddev->level == 6)
5576 		return raid5_takeover_raid6(mddev);
5577 
5578 	return ERR_PTR(-EINVAL);
5579 }
5580 
5581 static void *raid4_takeover(struct mddev *mddev)
5582 {
5583 	/* raid4 can take over:
5584 	 *  raid0 - if there is only one strip zone
5585 	 *  raid5 - if layout is right
5586 	 */
5587 	if (mddev->level == 0)
5588 		return raid45_takeover_raid0(mddev, 4);
5589 	if (mddev->level == 5 &&
5590 	    mddev->layout == ALGORITHM_PARITY_N) {
5591 		mddev->new_layout = 0;
5592 		mddev->new_level = 4;
5593 		return setup_conf(mddev);
5594 	}
5595 	return ERR_PTR(-EINVAL);
5596 }
5597 
5598 static struct md_personality raid5_personality;
5599 
5600 static void *raid6_takeover(struct mddev *mddev)
5601 {
5602 	/* Currently can only take over a raid5.  We map the
5603 	 * personality to an equivalent raid6 personality
5604 	 * with the Q block at the end.
5605 	 */
5606 	int new_layout;
5607 
5608 	if (mddev->pers != &raid5_personality)
5609 		return ERR_PTR(-EINVAL);
5610 	if (mddev->degraded > 1)
5611 		return ERR_PTR(-EINVAL);
5612 	if (mddev->raid_disks > 253)
5613 		return ERR_PTR(-EINVAL);
5614 	if (mddev->raid_disks < 3)
5615 		return ERR_PTR(-EINVAL);
5616 
5617 	switch (mddev->layout) {
5618 	case ALGORITHM_LEFT_ASYMMETRIC:
5619 		new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5620 		break;
5621 	case ALGORITHM_RIGHT_ASYMMETRIC:
5622 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5623 		break;
5624 	case ALGORITHM_LEFT_SYMMETRIC:
5625 		new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5626 		break;
5627 	case ALGORITHM_RIGHT_SYMMETRIC:
5628 		new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5629 		break;
5630 	case ALGORITHM_PARITY_0:
5631 		new_layout = ALGORITHM_PARITY_0_6;
5632 		break;
5633 	case ALGORITHM_PARITY_N:
5634 		new_layout = ALGORITHM_PARITY_N;
5635 		break;
5636 	default:
5637 		return ERR_PTR(-EINVAL);
5638 	}
5639 	mddev->new_level = 6;
5640 	mddev->new_layout = new_layout;
5641 	mddev->delta_disks = 1;
5642 	mddev->raid_disks += 1;
5643 	return setup_conf(mddev);
5644 }
5645 
5646 
5647 static struct md_personality raid6_personality =
5648 {
5649 	.name		= "raid6",
5650 	.level		= 6,
5651 	.owner		= THIS_MODULE,
5652 	.make_request	= make_request,
5653 	.run		= run,
5654 	.stop		= stop,
5655 	.status		= status,
5656 	.error_handler	= error,
5657 	.hot_add_disk	= raid5_add_disk,
5658 	.hot_remove_disk= raid5_remove_disk,
5659 	.spare_active	= raid5_spare_active,
5660 	.sync_request	= sync_request,
5661 	.resize		= raid5_resize,
5662 	.size		= raid5_size,
5663 	.check_reshape	= raid6_check_reshape,
5664 	.start_reshape  = raid5_start_reshape,
5665 	.finish_reshape = raid5_finish_reshape,
5666 	.quiesce	= raid5_quiesce,
5667 	.takeover	= raid6_takeover,
5668 };
5669 static struct md_personality raid5_personality =
5670 {
5671 	.name		= "raid5",
5672 	.level		= 5,
5673 	.owner		= THIS_MODULE,
5674 	.make_request	= make_request,
5675 	.run		= run,
5676 	.stop		= stop,
5677 	.status		= status,
5678 	.error_handler	= error,
5679 	.hot_add_disk	= raid5_add_disk,
5680 	.hot_remove_disk= raid5_remove_disk,
5681 	.spare_active	= raid5_spare_active,
5682 	.sync_request	= sync_request,
5683 	.resize		= raid5_resize,
5684 	.size		= raid5_size,
5685 	.check_reshape	= raid5_check_reshape,
5686 	.start_reshape  = raid5_start_reshape,
5687 	.finish_reshape = raid5_finish_reshape,
5688 	.quiesce	= raid5_quiesce,
5689 	.takeover	= raid5_takeover,
5690 };
5691 
5692 static struct md_personality raid4_personality =
5693 {
5694 	.name		= "raid4",
5695 	.level		= 4,
5696 	.owner		= THIS_MODULE,
5697 	.make_request	= make_request,
5698 	.run		= run,
5699 	.stop		= stop,
5700 	.status		= status,
5701 	.error_handler	= error,
5702 	.hot_add_disk	= raid5_add_disk,
5703 	.hot_remove_disk= raid5_remove_disk,
5704 	.spare_active	= raid5_spare_active,
5705 	.sync_request	= sync_request,
5706 	.resize		= raid5_resize,
5707 	.size		= raid5_size,
5708 	.check_reshape	= raid5_check_reshape,
5709 	.start_reshape  = raid5_start_reshape,
5710 	.finish_reshape = raid5_finish_reshape,
5711 	.quiesce	= raid5_quiesce,
5712 	.takeover	= raid4_takeover,
5713 };
5714 
5715 static int __init raid5_init(void)
5716 {
5717 	register_md_personality(&raid6_personality);
5718 	register_md_personality(&raid5_personality);
5719 	register_md_personality(&raid4_personality);
5720 	return 0;
5721 }
5722 
5723 static void raid5_exit(void)
5724 {
5725 	unregister_md_personality(&raid6_personality);
5726 	unregister_md_personality(&raid5_personality);
5727 	unregister_md_personality(&raid4_personality);
5728 }
5729 
5730 module_init(raid5_init);
5731 module_exit(raid5_exit);
5732 MODULE_LICENSE("GPL");
5733 MODULE_DESCRIPTION("RAID4/5/6 (striping with parity) personality for MD");
5734 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5735 MODULE_ALIAS("md-raid5");
5736 MODULE_ALIAS("md-raid4");
5737 MODULE_ALIAS("md-level-5");
5738 MODULE_ALIAS("md-level-4");
5739 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5740 MODULE_ALIAS("md-raid6");
5741 MODULE_ALIAS("md-level-6");
5742 
5743 /* This used to be two separate modules, they were: */
5744 MODULE_ALIAS("raid5");
5745 MODULE_ALIAS("raid6");
5746