xref: /linux/drivers/md/raid5.c (revision 1a5aeeecd550ee4344cfba1791f1134739b16dc6)
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->bm_write is the number of the last batch successfully written.
31  * conf->bm_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 bm_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/async.h>
51 #include <linux/seq_file.h>
52 #include <linux/cpu.h>
53 #include "md.h"
54 #include "raid5.h"
55 #include "bitmap.h"
56 
57 /*
58  * Stripe cache
59  */
60 
61 #define NR_STRIPES		256
62 #define STRIPE_SIZE		PAGE_SIZE
63 #define STRIPE_SHIFT		(PAGE_SHIFT - 9)
64 #define STRIPE_SECTORS		(STRIPE_SIZE>>9)
65 #define	IO_THRESHOLD		1
66 #define BYPASS_THRESHOLD	1
67 #define NR_HASH			(PAGE_SIZE / sizeof(struct hlist_head))
68 #define HASH_MASK		(NR_HASH - 1)
69 
70 #define stripe_hash(conf, sect)	(&((conf)->stripe_hashtbl[((sect) >> STRIPE_SHIFT) & HASH_MASK]))
71 
72 /* bio's attached to a stripe+device for I/O are linked together in bi_sector
73  * order without overlap.  There may be several bio's per stripe+device, and
74  * a bio could span several devices.
75  * When walking this list for a particular stripe+device, we must never proceed
76  * beyond a bio that extends past this device, as the next bio might no longer
77  * be valid.
78  * This macro is used to determine the 'next' bio in the list, given the sector
79  * of the current stripe+device
80  */
81 #define r5_next_bio(bio, sect) ( ( (bio)->bi_sector + ((bio)->bi_size>>9) < sect + STRIPE_SECTORS) ? (bio)->bi_next : NULL)
82 /*
83  * The following can be used to debug the driver
84  */
85 #define RAID5_PARANOIA	1
86 #if RAID5_PARANOIA && defined(CONFIG_SMP)
87 # define CHECK_DEVLOCK() assert_spin_locked(&conf->device_lock)
88 #else
89 # define CHECK_DEVLOCK()
90 #endif
91 
92 #ifdef DEBUG
93 #define inline
94 #define __inline__
95 #endif
96 
97 #define printk_rl(args...) ((void) (printk_ratelimit() && printk(args)))
98 
99 /*
100  * We maintain a biased count of active stripes in the bottom 16 bits of
101  * bi_phys_segments, and a count of processed stripes in the upper 16 bits
102  */
103 static inline int raid5_bi_phys_segments(struct bio *bio)
104 {
105 	return bio->bi_phys_segments & 0xffff;
106 }
107 
108 static inline int raid5_bi_hw_segments(struct bio *bio)
109 {
110 	return (bio->bi_phys_segments >> 16) & 0xffff;
111 }
112 
113 static inline int raid5_dec_bi_phys_segments(struct bio *bio)
114 {
115 	--bio->bi_phys_segments;
116 	return raid5_bi_phys_segments(bio);
117 }
118 
119 static inline int raid5_dec_bi_hw_segments(struct bio *bio)
120 {
121 	unsigned short val = raid5_bi_hw_segments(bio);
122 
123 	--val;
124 	bio->bi_phys_segments = (val << 16) | raid5_bi_phys_segments(bio);
125 	return val;
126 }
127 
128 static inline void raid5_set_bi_hw_segments(struct bio *bio, unsigned int cnt)
129 {
130 	bio->bi_phys_segments = raid5_bi_phys_segments(bio) || (cnt << 16);
131 }
132 
133 /* Find first data disk in a raid6 stripe */
134 static inline int raid6_d0(struct stripe_head *sh)
135 {
136 	if (sh->ddf_layout)
137 		/* ddf always start from first device */
138 		return 0;
139 	/* md starts just after Q block */
140 	if (sh->qd_idx == sh->disks - 1)
141 		return 0;
142 	else
143 		return sh->qd_idx + 1;
144 }
145 static inline int raid6_next_disk(int disk, int raid_disks)
146 {
147 	disk++;
148 	return (disk < raid_disks) ? disk : 0;
149 }
150 
151 /* When walking through the disks in a raid5, starting at raid6_d0,
152  * We need to map each disk to a 'slot', where the data disks are slot
153  * 0 .. raid_disks-3, the parity disk is raid_disks-2 and the Q disk
154  * is raid_disks-1.  This help does that mapping.
155  */
156 static int raid6_idx_to_slot(int idx, struct stripe_head *sh,
157 			     int *count, int syndrome_disks)
158 {
159 	int slot;
160 
161 	if (idx == sh->pd_idx)
162 		return syndrome_disks;
163 	if (idx == sh->qd_idx)
164 		return syndrome_disks + 1;
165 	slot = (*count)++;
166 	return slot;
167 }
168 
169 static void return_io(struct bio *return_bi)
170 {
171 	struct bio *bi = return_bi;
172 	while (bi) {
173 
174 		return_bi = bi->bi_next;
175 		bi->bi_next = NULL;
176 		bi->bi_size = 0;
177 		bio_endio(bi, 0);
178 		bi = return_bi;
179 	}
180 }
181 
182 static void print_raid5_conf (raid5_conf_t *conf);
183 
184 static int stripe_operations_active(struct stripe_head *sh)
185 {
186 	return sh->check_state || sh->reconstruct_state ||
187 	       test_bit(STRIPE_BIOFILL_RUN, &sh->state) ||
188 	       test_bit(STRIPE_COMPUTE_RUN, &sh->state);
189 }
190 
191 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
192 {
193 	if (atomic_dec_and_test(&sh->count)) {
194 		BUG_ON(!list_empty(&sh->lru));
195 		BUG_ON(atomic_read(&conf->active_stripes)==0);
196 		if (test_bit(STRIPE_HANDLE, &sh->state)) {
197 			if (test_bit(STRIPE_DELAYED, &sh->state)) {
198 				list_add_tail(&sh->lru, &conf->delayed_list);
199 				blk_plug_device(conf->mddev->queue);
200 			} else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
201 				   sh->bm_seq - conf->seq_write > 0) {
202 				list_add_tail(&sh->lru, &conf->bitmap_list);
203 				blk_plug_device(conf->mddev->queue);
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 	raid5_conf_t *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(raid5_conf_t *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 	CHECK_DEVLOCK();
253 	hlist_add_head(&sh->hash, hp);
254 }
255 
256 
257 /* find an idle stripe, make sure it is unhashed, and return it. */
258 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
259 {
260 	struct stripe_head *sh = NULL;
261 	struct list_head *first;
262 
263 	CHECK_DEVLOCK();
264 	if (list_empty(&conf->inactive_list))
265 		goto out;
266 	first = conf->inactive_list.next;
267 	sh = list_entry(first, struct stripe_head, lru);
268 	list_del_init(first);
269 	remove_hash(sh);
270 	atomic_inc(&conf->active_stripes);
271 out:
272 	return sh;
273 }
274 
275 static void shrink_buffers(struct stripe_head *sh, int num)
276 {
277 	struct page *p;
278 	int i;
279 
280 	for (i=0; i<num ; i++) {
281 		p = sh->dev[i].page;
282 		if (!p)
283 			continue;
284 		sh->dev[i].page = NULL;
285 		put_page(p);
286 	}
287 }
288 
289 static int grow_buffers(struct stripe_head *sh, int num)
290 {
291 	int i;
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, raid5_conf_t *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 	raid5_conf_t *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 	CHECK_DEVLOCK();
318 	pr_debug("init_stripe called, stripe %llu\n",
319 		(unsigned long long)sh->sector);
320 
321 	remove_hash(sh);
322 
323 	sh->generation = conf->generation - previous;
324 	sh->disks = previous ? conf->previous_raid_disks : conf->raid_disks;
325 	sh->sector = sector;
326 	stripe_set_idx(sector, conf, previous, sh);
327 	sh->state = 0;
328 
329 
330 	for (i = sh->disks; i--; ) {
331 		struct r5dev *dev = &sh->dev[i];
332 
333 		if (dev->toread || dev->read || dev->towrite || dev->written ||
334 		    test_bit(R5_LOCKED, &dev->flags)) {
335 			printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
336 			       (unsigned long long)sh->sector, i, dev->toread,
337 			       dev->read, dev->towrite, dev->written,
338 			       test_bit(R5_LOCKED, &dev->flags));
339 			BUG();
340 		}
341 		dev->flags = 0;
342 		raid5_build_block(sh, i, previous);
343 	}
344 	insert_hash(conf, sh);
345 }
346 
347 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector,
348 					 short generation)
349 {
350 	struct stripe_head *sh;
351 	struct hlist_node *hn;
352 
353 	CHECK_DEVLOCK();
354 	pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
355 	hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
356 		if (sh->sector == sector && sh->generation == generation)
357 			return sh;
358 	pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
359 	return NULL;
360 }
361 
362 static void unplug_slaves(mddev_t *mddev);
363 static void raid5_unplug_device(struct request_queue *q);
364 
365 static struct stripe_head *
366 get_active_stripe(raid5_conf_t *conf, sector_t sector,
367 		  int previous, int noblock, int noquiesce)
368 {
369 	struct stripe_head *sh;
370 
371 	pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
372 
373 	spin_lock_irq(&conf->device_lock);
374 
375 	do {
376 		wait_event_lock_irq(conf->wait_for_stripe,
377 				    conf->quiesce == 0 || noquiesce,
378 				    conf->device_lock, /* nothing */);
379 		sh = __find_stripe(conf, sector, conf->generation - previous);
380 		if (!sh) {
381 			if (!conf->inactive_blocked)
382 				sh = get_free_stripe(conf);
383 			if (noblock && sh == NULL)
384 				break;
385 			if (!sh) {
386 				conf->inactive_blocked = 1;
387 				wait_event_lock_irq(conf->wait_for_stripe,
388 						    !list_empty(&conf->inactive_list) &&
389 						    (atomic_read(&conf->active_stripes)
390 						     < (conf->max_nr_stripes *3/4)
391 						     || !conf->inactive_blocked),
392 						    conf->device_lock,
393 						    raid5_unplug_device(conf->mddev->queue)
394 					);
395 				conf->inactive_blocked = 0;
396 			} else
397 				init_stripe(sh, sector, previous);
398 		} else {
399 			if (atomic_read(&sh->count)) {
400 				BUG_ON(!list_empty(&sh->lru)
401 				    && !test_bit(STRIPE_EXPANDING, &sh->state));
402 			} else {
403 				if (!test_bit(STRIPE_HANDLE, &sh->state))
404 					atomic_inc(&conf->active_stripes);
405 				if (list_empty(&sh->lru) &&
406 				    !test_bit(STRIPE_EXPANDING, &sh->state))
407 					BUG();
408 				list_del_init(&sh->lru);
409 			}
410 		}
411 	} while (sh == NULL);
412 
413 	if (sh)
414 		atomic_inc(&sh->count);
415 
416 	spin_unlock_irq(&conf->device_lock);
417 	return sh;
418 }
419 
420 static void
421 raid5_end_read_request(struct bio *bi, int error);
422 static void
423 raid5_end_write_request(struct bio *bi, int error);
424 
425 static void ops_run_io(struct stripe_head *sh, struct stripe_head_state *s)
426 {
427 	raid5_conf_t *conf = sh->raid_conf;
428 	int i, disks = sh->disks;
429 
430 	might_sleep();
431 
432 	for (i = disks; i--; ) {
433 		int rw;
434 		struct bio *bi;
435 		mdk_rdev_t *rdev;
436 		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
437 			rw = WRITE;
438 		else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
439 			rw = READ;
440 		else
441 			continue;
442 
443 		bi = &sh->dev[i].req;
444 
445 		bi->bi_rw = rw;
446 		if (rw == WRITE)
447 			bi->bi_end_io = raid5_end_write_request;
448 		else
449 			bi->bi_end_io = raid5_end_read_request;
450 
451 		rcu_read_lock();
452 		rdev = rcu_dereference(conf->disks[i].rdev);
453 		if (rdev && test_bit(Faulty, &rdev->flags))
454 			rdev = NULL;
455 		if (rdev)
456 			atomic_inc(&rdev->nr_pending);
457 		rcu_read_unlock();
458 
459 		if (rdev) {
460 			if (s->syncing || s->expanding || s->expanded)
461 				md_sync_acct(rdev->bdev, STRIPE_SECTORS);
462 
463 			set_bit(STRIPE_IO_STARTED, &sh->state);
464 
465 			bi->bi_bdev = rdev->bdev;
466 			pr_debug("%s: for %llu schedule op %ld on disc %d\n",
467 				__func__, (unsigned long long)sh->sector,
468 				bi->bi_rw, i);
469 			atomic_inc(&sh->count);
470 			bi->bi_sector = sh->sector + rdev->data_offset;
471 			bi->bi_flags = 1 << BIO_UPTODATE;
472 			bi->bi_vcnt = 1;
473 			bi->bi_max_vecs = 1;
474 			bi->bi_idx = 0;
475 			bi->bi_io_vec = &sh->dev[i].vec;
476 			bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
477 			bi->bi_io_vec[0].bv_offset = 0;
478 			bi->bi_size = STRIPE_SIZE;
479 			bi->bi_next = NULL;
480 			if (rw == WRITE &&
481 			    test_bit(R5_ReWrite, &sh->dev[i].flags))
482 				atomic_add(STRIPE_SECTORS,
483 					&rdev->corrected_errors);
484 			generic_make_request(bi);
485 		} else {
486 			if (rw == WRITE)
487 				set_bit(STRIPE_DEGRADED, &sh->state);
488 			pr_debug("skip op %ld on disc %d for sector %llu\n",
489 				bi->bi_rw, i, (unsigned long long)sh->sector);
490 			clear_bit(R5_LOCKED, &sh->dev[i].flags);
491 			set_bit(STRIPE_HANDLE, &sh->state);
492 		}
493 	}
494 }
495 
496 static struct dma_async_tx_descriptor *
497 async_copy_data(int frombio, struct bio *bio, struct page *page,
498 	sector_t sector, struct dma_async_tx_descriptor *tx)
499 {
500 	struct bio_vec *bvl;
501 	struct page *bio_page;
502 	int i;
503 	int page_offset;
504 	struct async_submit_ctl submit;
505 	enum async_tx_flags flags = 0;
506 
507 	if (bio->bi_sector >= sector)
508 		page_offset = (signed)(bio->bi_sector - sector) * 512;
509 	else
510 		page_offset = (signed)(sector - bio->bi_sector) * -512;
511 
512 	if (frombio)
513 		flags |= ASYNC_TX_FENCE;
514 	init_async_submit(&submit, flags, tx, NULL, NULL, NULL);
515 
516 	bio_for_each_segment(bvl, bio, i) {
517 		int len = bio_iovec_idx(bio, i)->bv_len;
518 		int clen;
519 		int b_offset = 0;
520 
521 		if (page_offset < 0) {
522 			b_offset = -page_offset;
523 			page_offset += b_offset;
524 			len -= b_offset;
525 		}
526 
527 		if (len > 0 && page_offset + len > STRIPE_SIZE)
528 			clen = STRIPE_SIZE - page_offset;
529 		else
530 			clen = len;
531 
532 		if (clen > 0) {
533 			b_offset += bio_iovec_idx(bio, i)->bv_offset;
534 			bio_page = bio_iovec_idx(bio, i)->bv_page;
535 			if (frombio)
536 				tx = async_memcpy(page, bio_page, page_offset,
537 						  b_offset, clen, &submit);
538 			else
539 				tx = async_memcpy(bio_page, page, b_offset,
540 						  page_offset, clen, &submit);
541 		}
542 		/* chain the operations */
543 		submit.depend_tx = tx;
544 
545 		if (clen < len) /* hit end of page */
546 			break;
547 		page_offset +=  len;
548 	}
549 
550 	return tx;
551 }
552 
553 static void ops_complete_biofill(void *stripe_head_ref)
554 {
555 	struct stripe_head *sh = stripe_head_ref;
556 	struct bio *return_bi = NULL;
557 	raid5_conf_t *conf = sh->raid_conf;
558 	int i;
559 
560 	pr_debug("%s: stripe %llu\n", __func__,
561 		(unsigned long long)sh->sector);
562 
563 	/* clear completed biofills */
564 	spin_lock_irq(&conf->device_lock);
565 	for (i = sh->disks; i--; ) {
566 		struct r5dev *dev = &sh->dev[i];
567 
568 		/* acknowledge completion of a biofill operation */
569 		/* and check if we need to reply to a read request,
570 		 * new R5_Wantfill requests are held off until
571 		 * !STRIPE_BIOFILL_RUN
572 		 */
573 		if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
574 			struct bio *rbi, *rbi2;
575 
576 			BUG_ON(!dev->read);
577 			rbi = dev->read;
578 			dev->read = NULL;
579 			while (rbi && rbi->bi_sector <
580 				dev->sector + STRIPE_SECTORS) {
581 				rbi2 = r5_next_bio(rbi, dev->sector);
582 				if (!raid5_dec_bi_phys_segments(rbi)) {
583 					rbi->bi_next = return_bi;
584 					return_bi = rbi;
585 				}
586 				rbi = rbi2;
587 			}
588 		}
589 	}
590 	spin_unlock_irq(&conf->device_lock);
591 	clear_bit(STRIPE_BIOFILL_RUN, &sh->state);
592 
593 	return_io(return_bi);
594 
595 	set_bit(STRIPE_HANDLE, &sh->state);
596 	release_stripe(sh);
597 }
598 
599 static void ops_run_biofill(struct stripe_head *sh)
600 {
601 	struct dma_async_tx_descriptor *tx = NULL;
602 	raid5_conf_t *conf = sh->raid_conf;
603 	struct async_submit_ctl submit;
604 	int i;
605 
606 	pr_debug("%s: stripe %llu\n", __func__,
607 		(unsigned long long)sh->sector);
608 
609 	for (i = sh->disks; i--; ) {
610 		struct r5dev *dev = &sh->dev[i];
611 		if (test_bit(R5_Wantfill, &dev->flags)) {
612 			struct bio *rbi;
613 			spin_lock_irq(&conf->device_lock);
614 			dev->read = rbi = dev->toread;
615 			dev->toread = NULL;
616 			spin_unlock_irq(&conf->device_lock);
617 			while (rbi && rbi->bi_sector <
618 				dev->sector + STRIPE_SECTORS) {
619 				tx = async_copy_data(0, rbi, dev->page,
620 					dev->sector, tx);
621 				rbi = r5_next_bio(rbi, dev->sector);
622 			}
623 		}
624 	}
625 
626 	atomic_inc(&sh->count);
627 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_biofill, sh, NULL);
628 	async_trigger_callback(&submit);
629 }
630 
631 static void mark_target_uptodate(struct stripe_head *sh, int target)
632 {
633 	struct r5dev *tgt;
634 
635 	if (target < 0)
636 		return;
637 
638 	tgt = &sh->dev[target];
639 	set_bit(R5_UPTODATE, &tgt->flags);
640 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
641 	clear_bit(R5_Wantcompute, &tgt->flags);
642 }
643 
644 static void ops_complete_compute(void *stripe_head_ref)
645 {
646 	struct stripe_head *sh = stripe_head_ref;
647 
648 	pr_debug("%s: stripe %llu\n", __func__,
649 		(unsigned long long)sh->sector);
650 
651 	/* mark the computed target(s) as uptodate */
652 	mark_target_uptodate(sh, sh->ops.target);
653 	mark_target_uptodate(sh, sh->ops.target2);
654 
655 	clear_bit(STRIPE_COMPUTE_RUN, &sh->state);
656 	if (sh->check_state == check_state_compute_run)
657 		sh->check_state = check_state_compute_result;
658 	set_bit(STRIPE_HANDLE, &sh->state);
659 	release_stripe(sh);
660 }
661 
662 /* return a pointer to the address conversion region of the scribble buffer */
663 static addr_conv_t *to_addr_conv(struct stripe_head *sh,
664 				 struct raid5_percpu *percpu)
665 {
666 	return percpu->scribble + sizeof(struct page *) * (sh->disks + 2);
667 }
668 
669 static struct dma_async_tx_descriptor *
670 ops_run_compute5(struct stripe_head *sh, struct raid5_percpu *percpu)
671 {
672 	int disks = sh->disks;
673 	struct page **xor_srcs = percpu->scribble;
674 	int target = sh->ops.target;
675 	struct r5dev *tgt = &sh->dev[target];
676 	struct page *xor_dest = tgt->page;
677 	int count = 0;
678 	struct dma_async_tx_descriptor *tx;
679 	struct async_submit_ctl submit;
680 	int i;
681 
682 	pr_debug("%s: stripe %llu block: %d\n",
683 		__func__, (unsigned long long)sh->sector, target);
684 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
685 
686 	for (i = disks; i--; )
687 		if (i != target)
688 			xor_srcs[count++] = sh->dev[i].page;
689 
690 	atomic_inc(&sh->count);
691 
692 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST, NULL,
693 			  ops_complete_compute, sh, to_addr_conv(sh, percpu));
694 	if (unlikely(count == 1))
695 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
696 	else
697 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
698 
699 	return tx;
700 }
701 
702 /* set_syndrome_sources - populate source buffers for gen_syndrome
703  * @srcs - (struct page *) array of size sh->disks
704  * @sh - stripe_head to parse
705  *
706  * Populates srcs in proper layout order for the stripe and returns the
707  * 'count' of sources to be used in a call to async_gen_syndrome.  The P
708  * destination buffer is recorded in srcs[count] and the Q destination
709  * is recorded in srcs[count+1]].
710  */
711 static int set_syndrome_sources(struct page **srcs, struct stripe_head *sh)
712 {
713 	int disks = sh->disks;
714 	int syndrome_disks = sh->ddf_layout ? disks : (disks - 2);
715 	int d0_idx = raid6_d0(sh);
716 	int count;
717 	int i;
718 
719 	for (i = 0; i < disks; i++)
720 		srcs[i] = (void *)raid6_empty_zero_page;
721 
722 	count = 0;
723 	i = d0_idx;
724 	do {
725 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
726 
727 		srcs[slot] = sh->dev[i].page;
728 		i = raid6_next_disk(i, disks);
729 	} while (i != d0_idx);
730 	BUG_ON(count != syndrome_disks);
731 
732 	return count;
733 }
734 
735 static struct dma_async_tx_descriptor *
736 ops_run_compute6_1(struct stripe_head *sh, struct raid5_percpu *percpu)
737 {
738 	int disks = sh->disks;
739 	struct page **blocks = percpu->scribble;
740 	int target;
741 	int qd_idx = sh->qd_idx;
742 	struct dma_async_tx_descriptor *tx;
743 	struct async_submit_ctl submit;
744 	struct r5dev *tgt;
745 	struct page *dest;
746 	int i;
747 	int count;
748 
749 	if (sh->ops.target < 0)
750 		target = sh->ops.target2;
751 	else if (sh->ops.target2 < 0)
752 		target = sh->ops.target;
753 	else
754 		/* we should only have one valid target */
755 		BUG();
756 	BUG_ON(target < 0);
757 	pr_debug("%s: stripe %llu block: %d\n",
758 		__func__, (unsigned long long)sh->sector, target);
759 
760 	tgt = &sh->dev[target];
761 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
762 	dest = tgt->page;
763 
764 	atomic_inc(&sh->count);
765 
766 	if (target == qd_idx) {
767 		count = set_syndrome_sources(blocks, sh);
768 		blocks[count] = NULL; /* regenerating p is not necessary */
769 		BUG_ON(blocks[count+1] != dest); /* q should already be set */
770 		init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
771 				  ops_complete_compute, sh,
772 				  to_addr_conv(sh, percpu));
773 		tx = async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE, &submit);
774 	} else {
775 		/* Compute any data- or p-drive using XOR */
776 		count = 0;
777 		for (i = disks; i-- ; ) {
778 			if (i == target || i == qd_idx)
779 				continue;
780 			blocks[count++] = sh->dev[i].page;
781 		}
782 
783 		init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
784 				  NULL, ops_complete_compute, sh,
785 				  to_addr_conv(sh, percpu));
786 		tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE, &submit);
787 	}
788 
789 	return tx;
790 }
791 
792 static struct dma_async_tx_descriptor *
793 ops_run_compute6_2(struct stripe_head *sh, struct raid5_percpu *percpu)
794 {
795 	int i, count, disks = sh->disks;
796 	int syndrome_disks = sh->ddf_layout ? disks : disks-2;
797 	int d0_idx = raid6_d0(sh);
798 	int faila = -1, failb = -1;
799 	int target = sh->ops.target;
800 	int target2 = sh->ops.target2;
801 	struct r5dev *tgt = &sh->dev[target];
802 	struct r5dev *tgt2 = &sh->dev[target2];
803 	struct dma_async_tx_descriptor *tx;
804 	struct page **blocks = percpu->scribble;
805 	struct async_submit_ctl submit;
806 
807 	pr_debug("%s: stripe %llu block1: %d block2: %d\n",
808 		 __func__, (unsigned long long)sh->sector, target, target2);
809 	BUG_ON(target < 0 || target2 < 0);
810 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
811 	BUG_ON(!test_bit(R5_Wantcompute, &tgt2->flags));
812 
813 	/* we need to open-code set_syndrome_sources to handle to the
814 	 * slot number conversion for 'faila' and 'failb'
815 	 */
816 	for (i = 0; i < disks ; i++)
817 		blocks[i] = (void *)raid6_empty_zero_page;
818 	count = 0;
819 	i = d0_idx;
820 	do {
821 		int slot = raid6_idx_to_slot(i, sh, &count, syndrome_disks);
822 
823 		blocks[slot] = sh->dev[i].page;
824 
825 		if (i == target)
826 			faila = slot;
827 		if (i == target2)
828 			failb = slot;
829 		i = raid6_next_disk(i, disks);
830 	} while (i != d0_idx);
831 	BUG_ON(count != syndrome_disks);
832 
833 	BUG_ON(faila == failb);
834 	if (failb < faila)
835 		swap(faila, failb);
836 	pr_debug("%s: stripe: %llu faila: %d failb: %d\n",
837 		 __func__, (unsigned long long)sh->sector, faila, failb);
838 
839 	atomic_inc(&sh->count);
840 
841 	if (failb == syndrome_disks+1) {
842 		/* Q disk is one of the missing disks */
843 		if (faila == syndrome_disks) {
844 			/* Missing P+Q, just recompute */
845 			init_async_submit(&submit, ASYNC_TX_FENCE, NULL,
846 					  ops_complete_compute, sh,
847 					  to_addr_conv(sh, percpu));
848 			return async_gen_syndrome(blocks, 0, count+2,
849 						  STRIPE_SIZE, &submit);
850 		} else {
851 			struct page *dest;
852 			int data_target;
853 			int qd_idx = sh->qd_idx;
854 
855 			/* Missing D+Q: recompute D from P, then recompute Q */
856 			if (target == qd_idx)
857 				data_target = target2;
858 			else
859 				data_target = target;
860 
861 			count = 0;
862 			for (i = disks; i-- ; ) {
863 				if (i == data_target || i == qd_idx)
864 					continue;
865 				blocks[count++] = sh->dev[i].page;
866 			}
867 			dest = sh->dev[data_target].page;
868 			init_async_submit(&submit,
869 					  ASYNC_TX_FENCE|ASYNC_TX_XOR_ZERO_DST,
870 					  NULL, NULL, NULL,
871 					  to_addr_conv(sh, percpu));
872 			tx = async_xor(dest, blocks, 0, count, STRIPE_SIZE,
873 				       &submit);
874 
875 			count = set_syndrome_sources(blocks, sh);
876 			init_async_submit(&submit, ASYNC_TX_FENCE, tx,
877 					  ops_complete_compute, sh,
878 					  to_addr_conv(sh, percpu));
879 			return async_gen_syndrome(blocks, 0, count+2,
880 						  STRIPE_SIZE, &submit);
881 		}
882 	}
883 
884 	init_async_submit(&submit, ASYNC_TX_FENCE, NULL, ops_complete_compute,
885 			  sh, to_addr_conv(sh, percpu));
886 	if (failb == syndrome_disks) {
887 		/* We're missing D+P. */
888 		return async_raid6_datap_recov(syndrome_disks+2, STRIPE_SIZE,
889 					       faila, blocks, &submit);
890 	} else {
891 		/* We're missing D+D. */
892 		return async_raid6_2data_recov(syndrome_disks+2, STRIPE_SIZE,
893 					       faila, failb, blocks, &submit);
894 	}
895 }
896 
897 
898 static void ops_complete_prexor(void *stripe_head_ref)
899 {
900 	struct stripe_head *sh = stripe_head_ref;
901 
902 	pr_debug("%s: stripe %llu\n", __func__,
903 		(unsigned long long)sh->sector);
904 }
905 
906 static struct dma_async_tx_descriptor *
907 ops_run_prexor(struct stripe_head *sh, struct raid5_percpu *percpu,
908 	       struct dma_async_tx_descriptor *tx)
909 {
910 	int disks = sh->disks;
911 	struct page **xor_srcs = percpu->scribble;
912 	int count = 0, pd_idx = sh->pd_idx, i;
913 	struct async_submit_ctl submit;
914 
915 	/* existing parity data subtracted */
916 	struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
917 
918 	pr_debug("%s: stripe %llu\n", __func__,
919 		(unsigned long long)sh->sector);
920 
921 	for (i = disks; i--; ) {
922 		struct r5dev *dev = &sh->dev[i];
923 		/* Only process blocks that are known to be uptodate */
924 		if (test_bit(R5_Wantdrain, &dev->flags))
925 			xor_srcs[count++] = dev->page;
926 	}
927 
928 	init_async_submit(&submit, ASYNC_TX_FENCE|ASYNC_TX_XOR_DROP_DST, tx,
929 			  ops_complete_prexor, sh, to_addr_conv(sh, percpu));
930 	tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
931 
932 	return tx;
933 }
934 
935 static struct dma_async_tx_descriptor *
936 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
937 {
938 	int disks = sh->disks;
939 	int i;
940 
941 	pr_debug("%s: stripe %llu\n", __func__,
942 		(unsigned long long)sh->sector);
943 
944 	for (i = disks; i--; ) {
945 		struct r5dev *dev = &sh->dev[i];
946 		struct bio *chosen;
947 
948 		if (test_and_clear_bit(R5_Wantdrain, &dev->flags)) {
949 			struct bio *wbi;
950 
951 			spin_lock(&sh->lock);
952 			chosen = dev->towrite;
953 			dev->towrite = NULL;
954 			BUG_ON(dev->written);
955 			wbi = dev->written = chosen;
956 			spin_unlock(&sh->lock);
957 
958 			while (wbi && wbi->bi_sector <
959 				dev->sector + STRIPE_SECTORS) {
960 				tx = async_copy_data(1, wbi, dev->page,
961 					dev->sector, tx);
962 				wbi = r5_next_bio(wbi, dev->sector);
963 			}
964 		}
965 	}
966 
967 	return tx;
968 }
969 
970 static void ops_complete_reconstruct(void *stripe_head_ref)
971 {
972 	struct stripe_head *sh = stripe_head_ref;
973 	int disks = sh->disks;
974 	int pd_idx = sh->pd_idx;
975 	int qd_idx = sh->qd_idx;
976 	int i;
977 
978 	pr_debug("%s: stripe %llu\n", __func__,
979 		(unsigned long long)sh->sector);
980 
981 	for (i = disks; i--; ) {
982 		struct r5dev *dev = &sh->dev[i];
983 
984 		if (dev->written || i == pd_idx || i == qd_idx)
985 			set_bit(R5_UPTODATE, &dev->flags);
986 	}
987 
988 	if (sh->reconstruct_state == reconstruct_state_drain_run)
989 		sh->reconstruct_state = reconstruct_state_drain_result;
990 	else if (sh->reconstruct_state == reconstruct_state_prexor_drain_run)
991 		sh->reconstruct_state = reconstruct_state_prexor_drain_result;
992 	else {
993 		BUG_ON(sh->reconstruct_state != reconstruct_state_run);
994 		sh->reconstruct_state = reconstruct_state_result;
995 	}
996 
997 	set_bit(STRIPE_HANDLE, &sh->state);
998 	release_stripe(sh);
999 }
1000 
1001 static void
1002 ops_run_reconstruct5(struct stripe_head *sh, struct raid5_percpu *percpu,
1003 		     struct dma_async_tx_descriptor *tx)
1004 {
1005 	int disks = sh->disks;
1006 	struct page **xor_srcs = percpu->scribble;
1007 	struct async_submit_ctl submit;
1008 	int count = 0, pd_idx = sh->pd_idx, i;
1009 	struct page *xor_dest;
1010 	int prexor = 0;
1011 	unsigned long flags;
1012 
1013 	pr_debug("%s: stripe %llu\n", __func__,
1014 		(unsigned long long)sh->sector);
1015 
1016 	/* check if prexor is active which means only process blocks
1017 	 * that are part of a read-modify-write (written)
1018 	 */
1019 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_run) {
1020 		prexor = 1;
1021 		xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
1022 		for (i = disks; i--; ) {
1023 			struct r5dev *dev = &sh->dev[i];
1024 			if (dev->written)
1025 				xor_srcs[count++] = dev->page;
1026 		}
1027 	} else {
1028 		xor_dest = sh->dev[pd_idx].page;
1029 		for (i = disks; i--; ) {
1030 			struct r5dev *dev = &sh->dev[i];
1031 			if (i != pd_idx)
1032 				xor_srcs[count++] = dev->page;
1033 		}
1034 	}
1035 
1036 	/* 1/ if we prexor'd then the dest is reused as a source
1037 	 * 2/ if we did not prexor then we are redoing the parity
1038 	 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
1039 	 * for the synchronous xor case
1040 	 */
1041 	flags = ASYNC_TX_ACK |
1042 		(prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
1043 
1044 	atomic_inc(&sh->count);
1045 
1046 	init_async_submit(&submit, flags, tx, ops_complete_reconstruct, sh,
1047 			  to_addr_conv(sh, percpu));
1048 	if (unlikely(count == 1))
1049 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE, &submit);
1050 	else
1051 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE, &submit);
1052 }
1053 
1054 static void
1055 ops_run_reconstruct6(struct stripe_head *sh, struct raid5_percpu *percpu,
1056 		     struct dma_async_tx_descriptor *tx)
1057 {
1058 	struct async_submit_ctl submit;
1059 	struct page **blocks = percpu->scribble;
1060 	int count;
1061 
1062 	pr_debug("%s: stripe %llu\n", __func__, (unsigned long long)sh->sector);
1063 
1064 	count = set_syndrome_sources(blocks, sh);
1065 
1066 	atomic_inc(&sh->count);
1067 
1068 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_reconstruct,
1069 			  sh, to_addr_conv(sh, percpu));
1070 	async_gen_syndrome(blocks, 0, count+2, STRIPE_SIZE,  &submit);
1071 }
1072 
1073 static void ops_complete_check(void *stripe_head_ref)
1074 {
1075 	struct stripe_head *sh = stripe_head_ref;
1076 
1077 	pr_debug("%s: stripe %llu\n", __func__,
1078 		(unsigned long long)sh->sector);
1079 
1080 	sh->check_state = check_state_check_result;
1081 	set_bit(STRIPE_HANDLE, &sh->state);
1082 	release_stripe(sh);
1083 }
1084 
1085 static void ops_run_check_p(struct stripe_head *sh, struct raid5_percpu *percpu)
1086 {
1087 	int disks = sh->disks;
1088 	int pd_idx = sh->pd_idx;
1089 	int qd_idx = sh->qd_idx;
1090 	struct page *xor_dest;
1091 	struct page **xor_srcs = percpu->scribble;
1092 	struct dma_async_tx_descriptor *tx;
1093 	struct async_submit_ctl submit;
1094 	int count;
1095 	int i;
1096 
1097 	pr_debug("%s: stripe %llu\n", __func__,
1098 		(unsigned long long)sh->sector);
1099 
1100 	count = 0;
1101 	xor_dest = sh->dev[pd_idx].page;
1102 	xor_srcs[count++] = xor_dest;
1103 	for (i = disks; i--; ) {
1104 		if (i == pd_idx || i == qd_idx)
1105 			continue;
1106 		xor_srcs[count++] = sh->dev[i].page;
1107 	}
1108 
1109 	init_async_submit(&submit, 0, NULL, NULL, NULL,
1110 			  to_addr_conv(sh, percpu));
1111 	tx = async_xor_val(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
1112 			   &sh->ops.zero_sum_result, &submit);
1113 
1114 	atomic_inc(&sh->count);
1115 	init_async_submit(&submit, ASYNC_TX_ACK, tx, ops_complete_check, sh, NULL);
1116 	tx = async_trigger_callback(&submit);
1117 }
1118 
1119 static void ops_run_check_pq(struct stripe_head *sh, struct raid5_percpu *percpu, int checkp)
1120 {
1121 	struct page **srcs = percpu->scribble;
1122 	struct async_submit_ctl submit;
1123 	int count;
1124 
1125 	pr_debug("%s: stripe %llu checkp: %d\n", __func__,
1126 		(unsigned long long)sh->sector, checkp);
1127 
1128 	count = set_syndrome_sources(srcs, sh);
1129 	if (!checkp)
1130 		srcs[count] = NULL;
1131 
1132 	atomic_inc(&sh->count);
1133 	init_async_submit(&submit, ASYNC_TX_ACK, NULL, ops_complete_check,
1134 			  sh, to_addr_conv(sh, percpu));
1135 	async_syndrome_val(srcs, 0, count+2, STRIPE_SIZE,
1136 			   &sh->ops.zero_sum_result, percpu->spare_page, &submit);
1137 }
1138 
1139 static void raid_run_ops(struct stripe_head *sh, unsigned long ops_request)
1140 {
1141 	int overlap_clear = 0, i, disks = sh->disks;
1142 	struct dma_async_tx_descriptor *tx = NULL;
1143 	raid5_conf_t *conf = sh->raid_conf;
1144 	int level = conf->level;
1145 	struct raid5_percpu *percpu;
1146 	unsigned long cpu;
1147 
1148 	cpu = get_cpu();
1149 	percpu = per_cpu_ptr(conf->percpu, cpu);
1150 	if (test_bit(STRIPE_OP_BIOFILL, &ops_request)) {
1151 		ops_run_biofill(sh);
1152 		overlap_clear++;
1153 	}
1154 
1155 	if (test_bit(STRIPE_OP_COMPUTE_BLK, &ops_request)) {
1156 		if (level < 6)
1157 			tx = ops_run_compute5(sh, percpu);
1158 		else {
1159 			if (sh->ops.target2 < 0 || sh->ops.target < 0)
1160 				tx = ops_run_compute6_1(sh, percpu);
1161 			else
1162 				tx = ops_run_compute6_2(sh, percpu);
1163 		}
1164 		/* terminate the chain if reconstruct is not set to be run */
1165 		if (tx && !test_bit(STRIPE_OP_RECONSTRUCT, &ops_request))
1166 			async_tx_ack(tx);
1167 	}
1168 
1169 	if (test_bit(STRIPE_OP_PREXOR, &ops_request))
1170 		tx = ops_run_prexor(sh, percpu, tx);
1171 
1172 	if (test_bit(STRIPE_OP_BIODRAIN, &ops_request)) {
1173 		tx = ops_run_biodrain(sh, tx);
1174 		overlap_clear++;
1175 	}
1176 
1177 	if (test_bit(STRIPE_OP_RECONSTRUCT, &ops_request)) {
1178 		if (level < 6)
1179 			ops_run_reconstruct5(sh, percpu, tx);
1180 		else
1181 			ops_run_reconstruct6(sh, percpu, tx);
1182 	}
1183 
1184 	if (test_bit(STRIPE_OP_CHECK, &ops_request)) {
1185 		if (sh->check_state == check_state_run)
1186 			ops_run_check_p(sh, percpu);
1187 		else if (sh->check_state == check_state_run_q)
1188 			ops_run_check_pq(sh, percpu, 0);
1189 		else if (sh->check_state == check_state_run_pq)
1190 			ops_run_check_pq(sh, percpu, 1);
1191 		else
1192 			BUG();
1193 	}
1194 
1195 	if (overlap_clear)
1196 		for (i = disks; i--; ) {
1197 			struct r5dev *dev = &sh->dev[i];
1198 			if (test_and_clear_bit(R5_Overlap, &dev->flags))
1199 				wake_up(&sh->raid_conf->wait_for_overlap);
1200 		}
1201 	put_cpu();
1202 }
1203 
1204 static int grow_one_stripe(raid5_conf_t *conf)
1205 {
1206 	struct stripe_head *sh;
1207 	sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
1208 	if (!sh)
1209 		return 0;
1210 	memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
1211 	sh->raid_conf = conf;
1212 	spin_lock_init(&sh->lock);
1213 
1214 	if (grow_buffers(sh, conf->raid_disks)) {
1215 		shrink_buffers(sh, conf->raid_disks);
1216 		kmem_cache_free(conf->slab_cache, sh);
1217 		return 0;
1218 	}
1219 	sh->disks = conf->raid_disks;
1220 	/* we just created an active stripe so... */
1221 	atomic_set(&sh->count, 1);
1222 	atomic_inc(&conf->active_stripes);
1223 	INIT_LIST_HEAD(&sh->lru);
1224 	release_stripe(sh);
1225 	return 1;
1226 }
1227 
1228 static int grow_stripes(raid5_conf_t *conf, int num)
1229 {
1230 	struct kmem_cache *sc;
1231 	int devs = conf->raid_disks;
1232 
1233 	sprintf(conf->cache_name[0],
1234 		"raid%d-%s", conf->level, mdname(conf->mddev));
1235 	sprintf(conf->cache_name[1],
1236 		"raid%d-%s-alt", conf->level, mdname(conf->mddev));
1237 	conf->active_name = 0;
1238 	sc = kmem_cache_create(conf->cache_name[conf->active_name],
1239 			       sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
1240 			       0, 0, NULL);
1241 	if (!sc)
1242 		return 1;
1243 	conf->slab_cache = sc;
1244 	conf->pool_size = devs;
1245 	while (num--)
1246 		if (!grow_one_stripe(conf))
1247 			return 1;
1248 	return 0;
1249 }
1250 
1251 /**
1252  * scribble_len - return the required size of the scribble region
1253  * @num - total number of disks in the array
1254  *
1255  * The size must be enough to contain:
1256  * 1/ a struct page pointer for each device in the array +2
1257  * 2/ room to convert each entry in (1) to its corresponding dma
1258  *    (dma_map_page()) or page (page_address()) address.
1259  *
1260  * Note: the +2 is for the destination buffers of the ddf/raid6 case where we
1261  * calculate over all devices (not just the data blocks), using zeros in place
1262  * of the P and Q blocks.
1263  */
1264 static size_t scribble_len(int num)
1265 {
1266 	size_t len;
1267 
1268 	len = sizeof(struct page *) * (num+2) + sizeof(addr_conv_t) * (num+2);
1269 
1270 	return len;
1271 }
1272 
1273 static int resize_stripes(raid5_conf_t *conf, int newsize)
1274 {
1275 	/* Make all the stripes able to hold 'newsize' devices.
1276 	 * New slots in each stripe get 'page' set to a new page.
1277 	 *
1278 	 * This happens in stages:
1279 	 * 1/ create a new kmem_cache and allocate the required number of
1280 	 *    stripe_heads.
1281 	 * 2/ gather all the old stripe_heads and tranfer the pages across
1282 	 *    to the new stripe_heads.  This will have the side effect of
1283 	 *    freezing the array as once all stripe_heads have been collected,
1284 	 *    no IO will be possible.  Old stripe heads are freed once their
1285 	 *    pages have been transferred over, and the old kmem_cache is
1286 	 *    freed when all stripes are done.
1287 	 * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
1288 	 *    we simple return a failre status - no need to clean anything up.
1289 	 * 4/ allocate new pages for the new slots in the new stripe_heads.
1290 	 *    If this fails, we don't bother trying the shrink the
1291 	 *    stripe_heads down again, we just leave them as they are.
1292 	 *    As each stripe_head is processed the new one is released into
1293 	 *    active service.
1294 	 *
1295 	 * Once step2 is started, we cannot afford to wait for a write,
1296 	 * so we use GFP_NOIO allocations.
1297 	 */
1298 	struct stripe_head *osh, *nsh;
1299 	LIST_HEAD(newstripes);
1300 	struct disk_info *ndisks;
1301 	unsigned long cpu;
1302 	int err;
1303 	struct kmem_cache *sc;
1304 	int i;
1305 
1306 	if (newsize <= conf->pool_size)
1307 		return 0; /* never bother to shrink */
1308 
1309 	err = md_allow_write(conf->mddev);
1310 	if (err)
1311 		return err;
1312 
1313 	/* Step 1 */
1314 	sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1315 			       sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1316 			       0, 0, NULL);
1317 	if (!sc)
1318 		return -ENOMEM;
1319 
1320 	for (i = conf->max_nr_stripes; i; i--) {
1321 		nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1322 		if (!nsh)
1323 			break;
1324 
1325 		memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1326 
1327 		nsh->raid_conf = conf;
1328 		spin_lock_init(&nsh->lock);
1329 
1330 		list_add(&nsh->lru, &newstripes);
1331 	}
1332 	if (i) {
1333 		/* didn't get enough, give up */
1334 		while (!list_empty(&newstripes)) {
1335 			nsh = list_entry(newstripes.next, struct stripe_head, lru);
1336 			list_del(&nsh->lru);
1337 			kmem_cache_free(sc, nsh);
1338 		}
1339 		kmem_cache_destroy(sc);
1340 		return -ENOMEM;
1341 	}
1342 	/* Step 2 - Must use GFP_NOIO now.
1343 	 * OK, we have enough stripes, start collecting inactive
1344 	 * stripes and copying them over
1345 	 */
1346 	list_for_each_entry(nsh, &newstripes, lru) {
1347 		spin_lock_irq(&conf->device_lock);
1348 		wait_event_lock_irq(conf->wait_for_stripe,
1349 				    !list_empty(&conf->inactive_list),
1350 				    conf->device_lock,
1351 				    unplug_slaves(conf->mddev)
1352 			);
1353 		osh = get_free_stripe(conf);
1354 		spin_unlock_irq(&conf->device_lock);
1355 		atomic_set(&nsh->count, 1);
1356 		for(i=0; i<conf->pool_size; i++)
1357 			nsh->dev[i].page = osh->dev[i].page;
1358 		for( ; i<newsize; i++)
1359 			nsh->dev[i].page = NULL;
1360 		kmem_cache_free(conf->slab_cache, osh);
1361 	}
1362 	kmem_cache_destroy(conf->slab_cache);
1363 
1364 	/* Step 3.
1365 	 * At this point, we are holding all the stripes so the array
1366 	 * is completely stalled, so now is a good time to resize
1367 	 * conf->disks and the scribble region
1368 	 */
1369 	ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1370 	if (ndisks) {
1371 		for (i=0; i<conf->raid_disks; i++)
1372 			ndisks[i] = conf->disks[i];
1373 		kfree(conf->disks);
1374 		conf->disks = ndisks;
1375 	} else
1376 		err = -ENOMEM;
1377 
1378 	get_online_cpus();
1379 	conf->scribble_len = scribble_len(newsize);
1380 	for_each_present_cpu(cpu) {
1381 		struct raid5_percpu *percpu;
1382 		void *scribble;
1383 
1384 		percpu = per_cpu_ptr(conf->percpu, cpu);
1385 		scribble = kmalloc(conf->scribble_len, GFP_NOIO);
1386 
1387 		if (scribble) {
1388 			kfree(percpu->scribble);
1389 			percpu->scribble = scribble;
1390 		} else {
1391 			err = -ENOMEM;
1392 			break;
1393 		}
1394 	}
1395 	put_online_cpus();
1396 
1397 	/* Step 4, return new stripes to service */
1398 	while(!list_empty(&newstripes)) {
1399 		nsh = list_entry(newstripes.next, struct stripe_head, lru);
1400 		list_del_init(&nsh->lru);
1401 
1402 		for (i=conf->raid_disks; i < newsize; i++)
1403 			if (nsh->dev[i].page == NULL) {
1404 				struct page *p = alloc_page(GFP_NOIO);
1405 				nsh->dev[i].page = p;
1406 				if (!p)
1407 					err = -ENOMEM;
1408 			}
1409 		release_stripe(nsh);
1410 	}
1411 	/* critical section pass, GFP_NOIO no longer needed */
1412 
1413 	conf->slab_cache = sc;
1414 	conf->active_name = 1-conf->active_name;
1415 	conf->pool_size = newsize;
1416 	return err;
1417 }
1418 
1419 static int drop_one_stripe(raid5_conf_t *conf)
1420 {
1421 	struct stripe_head *sh;
1422 
1423 	spin_lock_irq(&conf->device_lock);
1424 	sh = get_free_stripe(conf);
1425 	spin_unlock_irq(&conf->device_lock);
1426 	if (!sh)
1427 		return 0;
1428 	BUG_ON(atomic_read(&sh->count));
1429 	shrink_buffers(sh, conf->pool_size);
1430 	kmem_cache_free(conf->slab_cache, sh);
1431 	atomic_dec(&conf->active_stripes);
1432 	return 1;
1433 }
1434 
1435 static void shrink_stripes(raid5_conf_t *conf)
1436 {
1437 	while (drop_one_stripe(conf))
1438 		;
1439 
1440 	if (conf->slab_cache)
1441 		kmem_cache_destroy(conf->slab_cache);
1442 	conf->slab_cache = NULL;
1443 }
1444 
1445 static void raid5_end_read_request(struct bio * bi, int error)
1446 {
1447 	struct stripe_head *sh = bi->bi_private;
1448 	raid5_conf_t *conf = sh->raid_conf;
1449 	int disks = sh->disks, i;
1450 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1451 	char b[BDEVNAME_SIZE];
1452 	mdk_rdev_t *rdev;
1453 
1454 
1455 	for (i=0 ; i<disks; i++)
1456 		if (bi == &sh->dev[i].req)
1457 			break;
1458 
1459 	pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1460 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
1461 		uptodate);
1462 	if (i == disks) {
1463 		BUG();
1464 		return;
1465 	}
1466 
1467 	if (uptodate) {
1468 		set_bit(R5_UPTODATE, &sh->dev[i].flags);
1469 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1470 			rdev = conf->disks[i].rdev;
1471 			printk_rl(KERN_INFO "raid5:%s: read error corrected"
1472 				  " (%lu sectors at %llu on %s)\n",
1473 				  mdname(conf->mddev), STRIPE_SECTORS,
1474 				  (unsigned long long)(sh->sector
1475 						       + rdev->data_offset),
1476 				  bdevname(rdev->bdev, b));
1477 			clear_bit(R5_ReadError, &sh->dev[i].flags);
1478 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
1479 		}
1480 		if (atomic_read(&conf->disks[i].rdev->read_errors))
1481 			atomic_set(&conf->disks[i].rdev->read_errors, 0);
1482 	} else {
1483 		const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1484 		int retry = 0;
1485 		rdev = conf->disks[i].rdev;
1486 
1487 		clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1488 		atomic_inc(&rdev->read_errors);
1489 		if (conf->mddev->degraded)
1490 			printk_rl(KERN_WARNING
1491 				  "raid5:%s: read error not correctable "
1492 				  "(sector %llu on %s).\n",
1493 				  mdname(conf->mddev),
1494 				  (unsigned long long)(sh->sector
1495 						       + rdev->data_offset),
1496 				  bdn);
1497 		else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1498 			/* Oh, no!!! */
1499 			printk_rl(KERN_WARNING
1500 				  "raid5:%s: read error NOT corrected!! "
1501 				  "(sector %llu on %s).\n",
1502 				  mdname(conf->mddev),
1503 				  (unsigned long long)(sh->sector
1504 						       + rdev->data_offset),
1505 				  bdn);
1506 		else if (atomic_read(&rdev->read_errors)
1507 			 > conf->max_nr_stripes)
1508 			printk(KERN_WARNING
1509 			       "raid5:%s: Too many read errors, failing device %s.\n",
1510 			       mdname(conf->mddev), bdn);
1511 		else
1512 			retry = 1;
1513 		if (retry)
1514 			set_bit(R5_ReadError, &sh->dev[i].flags);
1515 		else {
1516 			clear_bit(R5_ReadError, &sh->dev[i].flags);
1517 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
1518 			md_error(conf->mddev, rdev);
1519 		}
1520 	}
1521 	rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1522 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
1523 	set_bit(STRIPE_HANDLE, &sh->state);
1524 	release_stripe(sh);
1525 }
1526 
1527 static void raid5_end_write_request(struct bio *bi, int error)
1528 {
1529 	struct stripe_head *sh = bi->bi_private;
1530 	raid5_conf_t *conf = sh->raid_conf;
1531 	int disks = sh->disks, i;
1532 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1533 
1534 	for (i=0 ; i<disks; i++)
1535 		if (bi == &sh->dev[i].req)
1536 			break;
1537 
1538 	pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1539 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
1540 		uptodate);
1541 	if (i == disks) {
1542 		BUG();
1543 		return;
1544 	}
1545 
1546 	if (!uptodate)
1547 		md_error(conf->mddev, conf->disks[i].rdev);
1548 
1549 	rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1550 
1551 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
1552 	set_bit(STRIPE_HANDLE, &sh->state);
1553 	release_stripe(sh);
1554 }
1555 
1556 
1557 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous);
1558 
1559 static void raid5_build_block(struct stripe_head *sh, int i, int previous)
1560 {
1561 	struct r5dev *dev = &sh->dev[i];
1562 
1563 	bio_init(&dev->req);
1564 	dev->req.bi_io_vec = &dev->vec;
1565 	dev->req.bi_vcnt++;
1566 	dev->req.bi_max_vecs++;
1567 	dev->vec.bv_page = dev->page;
1568 	dev->vec.bv_len = STRIPE_SIZE;
1569 	dev->vec.bv_offset = 0;
1570 
1571 	dev->req.bi_sector = sh->sector;
1572 	dev->req.bi_private = sh;
1573 
1574 	dev->flags = 0;
1575 	dev->sector = compute_blocknr(sh, i, previous);
1576 }
1577 
1578 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1579 {
1580 	char b[BDEVNAME_SIZE];
1581 	raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1582 	pr_debug("raid5: error called\n");
1583 
1584 	if (!test_bit(Faulty, &rdev->flags)) {
1585 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
1586 		if (test_and_clear_bit(In_sync, &rdev->flags)) {
1587 			unsigned long flags;
1588 			spin_lock_irqsave(&conf->device_lock, flags);
1589 			mddev->degraded++;
1590 			spin_unlock_irqrestore(&conf->device_lock, flags);
1591 			/*
1592 			 * if recovery was running, make sure it aborts.
1593 			 */
1594 			set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1595 		}
1596 		set_bit(Faulty, &rdev->flags);
1597 		printk(KERN_ALERT
1598 		       "raid5: Disk failure on %s, disabling device.\n"
1599 		       "raid5: Operation continuing on %d devices.\n",
1600 		       bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1601 	}
1602 }
1603 
1604 /*
1605  * Input: a 'big' sector number,
1606  * Output: index of the data and parity disk, and the sector # in them.
1607  */
1608 static sector_t raid5_compute_sector(raid5_conf_t *conf, sector_t r_sector,
1609 				     int previous, int *dd_idx,
1610 				     struct stripe_head *sh)
1611 {
1612 	long stripe;
1613 	unsigned long chunk_number;
1614 	unsigned int chunk_offset;
1615 	int pd_idx, qd_idx;
1616 	int ddf_layout = 0;
1617 	sector_t new_sector;
1618 	int algorithm = previous ? conf->prev_algo
1619 				 : conf->algorithm;
1620 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1621 					 : conf->chunk_sectors;
1622 	int raid_disks = previous ? conf->previous_raid_disks
1623 				  : conf->raid_disks;
1624 	int data_disks = raid_disks - conf->max_degraded;
1625 
1626 	/* First compute the information on this sector */
1627 
1628 	/*
1629 	 * Compute the chunk number and the sector offset inside the chunk
1630 	 */
1631 	chunk_offset = sector_div(r_sector, sectors_per_chunk);
1632 	chunk_number = r_sector;
1633 	BUG_ON(r_sector != chunk_number);
1634 
1635 	/*
1636 	 * Compute the stripe number
1637 	 */
1638 	stripe = chunk_number / data_disks;
1639 
1640 	/*
1641 	 * Compute the data disk and parity disk indexes inside the stripe
1642 	 */
1643 	*dd_idx = chunk_number % data_disks;
1644 
1645 	/*
1646 	 * Select the parity disk based on the user selected algorithm.
1647 	 */
1648 	pd_idx = qd_idx = ~0;
1649 	switch(conf->level) {
1650 	case 4:
1651 		pd_idx = data_disks;
1652 		break;
1653 	case 5:
1654 		switch (algorithm) {
1655 		case ALGORITHM_LEFT_ASYMMETRIC:
1656 			pd_idx = data_disks - stripe % raid_disks;
1657 			if (*dd_idx >= pd_idx)
1658 				(*dd_idx)++;
1659 			break;
1660 		case ALGORITHM_RIGHT_ASYMMETRIC:
1661 			pd_idx = stripe % raid_disks;
1662 			if (*dd_idx >= pd_idx)
1663 				(*dd_idx)++;
1664 			break;
1665 		case ALGORITHM_LEFT_SYMMETRIC:
1666 			pd_idx = data_disks - stripe % raid_disks;
1667 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1668 			break;
1669 		case ALGORITHM_RIGHT_SYMMETRIC:
1670 			pd_idx = stripe % raid_disks;
1671 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1672 			break;
1673 		case ALGORITHM_PARITY_0:
1674 			pd_idx = 0;
1675 			(*dd_idx)++;
1676 			break;
1677 		case ALGORITHM_PARITY_N:
1678 			pd_idx = data_disks;
1679 			break;
1680 		default:
1681 			printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1682 				algorithm);
1683 			BUG();
1684 		}
1685 		break;
1686 	case 6:
1687 
1688 		switch (algorithm) {
1689 		case ALGORITHM_LEFT_ASYMMETRIC:
1690 			pd_idx = raid_disks - 1 - (stripe % raid_disks);
1691 			qd_idx = pd_idx + 1;
1692 			if (pd_idx == raid_disks-1) {
1693 				(*dd_idx)++;	/* Q D D D P */
1694 				qd_idx = 0;
1695 			} else if (*dd_idx >= pd_idx)
1696 				(*dd_idx) += 2; /* D D P Q D */
1697 			break;
1698 		case ALGORITHM_RIGHT_ASYMMETRIC:
1699 			pd_idx = stripe % raid_disks;
1700 			qd_idx = pd_idx + 1;
1701 			if (pd_idx == raid_disks-1) {
1702 				(*dd_idx)++;	/* Q D D D P */
1703 				qd_idx = 0;
1704 			} else if (*dd_idx >= pd_idx)
1705 				(*dd_idx) += 2; /* D D P Q D */
1706 			break;
1707 		case ALGORITHM_LEFT_SYMMETRIC:
1708 			pd_idx = raid_disks - 1 - (stripe % raid_disks);
1709 			qd_idx = (pd_idx + 1) % raid_disks;
1710 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1711 			break;
1712 		case ALGORITHM_RIGHT_SYMMETRIC:
1713 			pd_idx = stripe % raid_disks;
1714 			qd_idx = (pd_idx + 1) % raid_disks;
1715 			*dd_idx = (pd_idx + 2 + *dd_idx) % raid_disks;
1716 			break;
1717 
1718 		case ALGORITHM_PARITY_0:
1719 			pd_idx = 0;
1720 			qd_idx = 1;
1721 			(*dd_idx) += 2;
1722 			break;
1723 		case ALGORITHM_PARITY_N:
1724 			pd_idx = data_disks;
1725 			qd_idx = data_disks + 1;
1726 			break;
1727 
1728 		case ALGORITHM_ROTATING_ZERO_RESTART:
1729 			/* Exactly the same as RIGHT_ASYMMETRIC, but or
1730 			 * of blocks for computing Q is different.
1731 			 */
1732 			pd_idx = stripe % raid_disks;
1733 			qd_idx = pd_idx + 1;
1734 			if (pd_idx == raid_disks-1) {
1735 				(*dd_idx)++;	/* Q D D D P */
1736 				qd_idx = 0;
1737 			} else if (*dd_idx >= pd_idx)
1738 				(*dd_idx) += 2; /* D D P Q D */
1739 			ddf_layout = 1;
1740 			break;
1741 
1742 		case ALGORITHM_ROTATING_N_RESTART:
1743 			/* Same a left_asymmetric, by first stripe is
1744 			 * D D D P Q  rather than
1745 			 * Q D D D P
1746 			 */
1747 			pd_idx = raid_disks - 1 - ((stripe + 1) % raid_disks);
1748 			qd_idx = pd_idx + 1;
1749 			if (pd_idx == raid_disks-1) {
1750 				(*dd_idx)++;	/* Q D D D P */
1751 				qd_idx = 0;
1752 			} else if (*dd_idx >= pd_idx)
1753 				(*dd_idx) += 2; /* D D P Q D */
1754 			ddf_layout = 1;
1755 			break;
1756 
1757 		case ALGORITHM_ROTATING_N_CONTINUE:
1758 			/* Same as left_symmetric but Q is before P */
1759 			pd_idx = raid_disks - 1 - (stripe % raid_disks);
1760 			qd_idx = (pd_idx + raid_disks - 1) % raid_disks;
1761 			*dd_idx = (pd_idx + 1 + *dd_idx) % raid_disks;
1762 			ddf_layout = 1;
1763 			break;
1764 
1765 		case ALGORITHM_LEFT_ASYMMETRIC_6:
1766 			/* RAID5 left_asymmetric, with Q on last device */
1767 			pd_idx = data_disks - stripe % (raid_disks-1);
1768 			if (*dd_idx >= pd_idx)
1769 				(*dd_idx)++;
1770 			qd_idx = raid_disks - 1;
1771 			break;
1772 
1773 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
1774 			pd_idx = stripe % (raid_disks-1);
1775 			if (*dd_idx >= pd_idx)
1776 				(*dd_idx)++;
1777 			qd_idx = raid_disks - 1;
1778 			break;
1779 
1780 		case ALGORITHM_LEFT_SYMMETRIC_6:
1781 			pd_idx = data_disks - stripe % (raid_disks-1);
1782 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1783 			qd_idx = raid_disks - 1;
1784 			break;
1785 
1786 		case ALGORITHM_RIGHT_SYMMETRIC_6:
1787 			pd_idx = stripe % (raid_disks-1);
1788 			*dd_idx = (pd_idx + 1 + *dd_idx) % (raid_disks-1);
1789 			qd_idx = raid_disks - 1;
1790 			break;
1791 
1792 		case ALGORITHM_PARITY_0_6:
1793 			pd_idx = 0;
1794 			(*dd_idx)++;
1795 			qd_idx = raid_disks - 1;
1796 			break;
1797 
1798 
1799 		default:
1800 			printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1801 			       algorithm);
1802 			BUG();
1803 		}
1804 		break;
1805 	}
1806 
1807 	if (sh) {
1808 		sh->pd_idx = pd_idx;
1809 		sh->qd_idx = qd_idx;
1810 		sh->ddf_layout = ddf_layout;
1811 	}
1812 	/*
1813 	 * Finally, compute the new sector number
1814 	 */
1815 	new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1816 	return new_sector;
1817 }
1818 
1819 
1820 static sector_t compute_blocknr(struct stripe_head *sh, int i, int previous)
1821 {
1822 	raid5_conf_t *conf = sh->raid_conf;
1823 	int raid_disks = sh->disks;
1824 	int data_disks = raid_disks - conf->max_degraded;
1825 	sector_t new_sector = sh->sector, check;
1826 	int sectors_per_chunk = previous ? conf->prev_chunk_sectors
1827 					 : conf->chunk_sectors;
1828 	int algorithm = previous ? conf->prev_algo
1829 				 : conf->algorithm;
1830 	sector_t stripe;
1831 	int chunk_offset;
1832 	int chunk_number, dummy1, dd_idx = i;
1833 	sector_t r_sector;
1834 	struct stripe_head sh2;
1835 
1836 
1837 	chunk_offset = sector_div(new_sector, sectors_per_chunk);
1838 	stripe = new_sector;
1839 	BUG_ON(new_sector != stripe);
1840 
1841 	if (i == sh->pd_idx)
1842 		return 0;
1843 	switch(conf->level) {
1844 	case 4: break;
1845 	case 5:
1846 		switch (algorithm) {
1847 		case ALGORITHM_LEFT_ASYMMETRIC:
1848 		case ALGORITHM_RIGHT_ASYMMETRIC:
1849 			if (i > sh->pd_idx)
1850 				i--;
1851 			break;
1852 		case ALGORITHM_LEFT_SYMMETRIC:
1853 		case ALGORITHM_RIGHT_SYMMETRIC:
1854 			if (i < sh->pd_idx)
1855 				i += raid_disks;
1856 			i -= (sh->pd_idx + 1);
1857 			break;
1858 		case ALGORITHM_PARITY_0:
1859 			i -= 1;
1860 			break;
1861 		case ALGORITHM_PARITY_N:
1862 			break;
1863 		default:
1864 			printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1865 			       algorithm);
1866 			BUG();
1867 		}
1868 		break;
1869 	case 6:
1870 		if (i == sh->qd_idx)
1871 			return 0; /* It is the Q disk */
1872 		switch (algorithm) {
1873 		case ALGORITHM_LEFT_ASYMMETRIC:
1874 		case ALGORITHM_RIGHT_ASYMMETRIC:
1875 		case ALGORITHM_ROTATING_ZERO_RESTART:
1876 		case ALGORITHM_ROTATING_N_RESTART:
1877 			if (sh->pd_idx == raid_disks-1)
1878 				i--;	/* Q D D D P */
1879 			else if (i > sh->pd_idx)
1880 				i -= 2; /* D D P Q D */
1881 			break;
1882 		case ALGORITHM_LEFT_SYMMETRIC:
1883 		case ALGORITHM_RIGHT_SYMMETRIC:
1884 			if (sh->pd_idx == raid_disks-1)
1885 				i--; /* Q D D D P */
1886 			else {
1887 				/* D D P Q D */
1888 				if (i < sh->pd_idx)
1889 					i += raid_disks;
1890 				i -= (sh->pd_idx + 2);
1891 			}
1892 			break;
1893 		case ALGORITHM_PARITY_0:
1894 			i -= 2;
1895 			break;
1896 		case ALGORITHM_PARITY_N:
1897 			break;
1898 		case ALGORITHM_ROTATING_N_CONTINUE:
1899 			if (sh->pd_idx == 0)
1900 				i--;	/* P D D D Q */
1901 			else if (i > sh->pd_idx)
1902 				i -= 2; /* D D Q P D */
1903 			break;
1904 		case ALGORITHM_LEFT_ASYMMETRIC_6:
1905 		case ALGORITHM_RIGHT_ASYMMETRIC_6:
1906 			if (i > sh->pd_idx)
1907 				i--;
1908 			break;
1909 		case ALGORITHM_LEFT_SYMMETRIC_6:
1910 		case ALGORITHM_RIGHT_SYMMETRIC_6:
1911 			if (i < sh->pd_idx)
1912 				i += data_disks + 1;
1913 			i -= (sh->pd_idx + 1);
1914 			break;
1915 		case ALGORITHM_PARITY_0_6:
1916 			i -= 1;
1917 			break;
1918 		default:
1919 			printk(KERN_CRIT "raid6: unsupported algorithm %d\n",
1920 			       algorithm);
1921 			BUG();
1922 		}
1923 		break;
1924 	}
1925 
1926 	chunk_number = stripe * data_disks + i;
1927 	r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1928 
1929 	check = raid5_compute_sector(conf, r_sector,
1930 				     previous, &dummy1, &sh2);
1931 	if (check != sh->sector || dummy1 != dd_idx || sh2.pd_idx != sh->pd_idx
1932 		|| sh2.qd_idx != sh->qd_idx) {
1933 		printk(KERN_ERR "compute_blocknr: map not correct\n");
1934 		return 0;
1935 	}
1936 	return r_sector;
1937 }
1938 
1939 
1940 static void
1941 schedule_reconstruction(struct stripe_head *sh, struct stripe_head_state *s,
1942 			 int rcw, int expand)
1943 {
1944 	int i, pd_idx = sh->pd_idx, disks = sh->disks;
1945 	raid5_conf_t *conf = sh->raid_conf;
1946 	int level = conf->level;
1947 
1948 	if (rcw) {
1949 		/* if we are not expanding this is a proper write request, and
1950 		 * there will be bios with new data to be drained into the
1951 		 * stripe cache
1952 		 */
1953 		if (!expand) {
1954 			sh->reconstruct_state = reconstruct_state_drain_run;
1955 			set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1956 		} else
1957 			sh->reconstruct_state = reconstruct_state_run;
1958 
1959 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1960 
1961 		for (i = disks; i--; ) {
1962 			struct r5dev *dev = &sh->dev[i];
1963 
1964 			if (dev->towrite) {
1965 				set_bit(R5_LOCKED, &dev->flags);
1966 				set_bit(R5_Wantdrain, &dev->flags);
1967 				if (!expand)
1968 					clear_bit(R5_UPTODATE, &dev->flags);
1969 				s->locked++;
1970 			}
1971 		}
1972 		if (s->locked + conf->max_degraded == disks)
1973 			if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1974 				atomic_inc(&conf->pending_full_writes);
1975 	} else {
1976 		BUG_ON(level == 6);
1977 		BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1978 			test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1979 
1980 		sh->reconstruct_state = reconstruct_state_prexor_drain_run;
1981 		set_bit(STRIPE_OP_PREXOR, &s->ops_request);
1982 		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
1983 		set_bit(STRIPE_OP_RECONSTRUCT, &s->ops_request);
1984 
1985 		for (i = disks; i--; ) {
1986 			struct r5dev *dev = &sh->dev[i];
1987 			if (i == pd_idx)
1988 				continue;
1989 
1990 			if (dev->towrite &&
1991 			    (test_bit(R5_UPTODATE, &dev->flags) ||
1992 			     test_bit(R5_Wantcompute, &dev->flags))) {
1993 				set_bit(R5_Wantdrain, &dev->flags);
1994 				set_bit(R5_LOCKED, &dev->flags);
1995 				clear_bit(R5_UPTODATE, &dev->flags);
1996 				s->locked++;
1997 			}
1998 		}
1999 	}
2000 
2001 	/* keep the parity disk(s) locked while asynchronous operations
2002 	 * are in flight
2003 	 */
2004 	set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
2005 	clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2006 	s->locked++;
2007 
2008 	if (level == 6) {
2009 		int qd_idx = sh->qd_idx;
2010 		struct r5dev *dev = &sh->dev[qd_idx];
2011 
2012 		set_bit(R5_LOCKED, &dev->flags);
2013 		clear_bit(R5_UPTODATE, &dev->flags);
2014 		s->locked++;
2015 	}
2016 
2017 	pr_debug("%s: stripe %llu locked: %d ops_request: %lx\n",
2018 		__func__, (unsigned long long)sh->sector,
2019 		s->locked, s->ops_request);
2020 }
2021 
2022 /*
2023  * Each stripe/dev can have one or more bion attached.
2024  * toread/towrite point to the first in a chain.
2025  * The bi_next chain must be in order.
2026  */
2027 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
2028 {
2029 	struct bio **bip;
2030 	raid5_conf_t *conf = sh->raid_conf;
2031 	int firstwrite=0;
2032 
2033 	pr_debug("adding bh b#%llu to stripe s#%llu\n",
2034 		(unsigned long long)bi->bi_sector,
2035 		(unsigned long long)sh->sector);
2036 
2037 
2038 	spin_lock(&sh->lock);
2039 	spin_lock_irq(&conf->device_lock);
2040 	if (forwrite) {
2041 		bip = &sh->dev[dd_idx].towrite;
2042 		if (*bip == NULL && sh->dev[dd_idx].written == NULL)
2043 			firstwrite = 1;
2044 	} else
2045 		bip = &sh->dev[dd_idx].toread;
2046 	while (*bip && (*bip)->bi_sector < bi->bi_sector) {
2047 		if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
2048 			goto overlap;
2049 		bip = & (*bip)->bi_next;
2050 	}
2051 	if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
2052 		goto overlap;
2053 
2054 	BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
2055 	if (*bip)
2056 		bi->bi_next = *bip;
2057 	*bip = bi;
2058 	bi->bi_phys_segments++;
2059 	spin_unlock_irq(&conf->device_lock);
2060 	spin_unlock(&sh->lock);
2061 
2062 	pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
2063 		(unsigned long long)bi->bi_sector,
2064 		(unsigned long long)sh->sector, dd_idx);
2065 
2066 	if (conf->mddev->bitmap && firstwrite) {
2067 		bitmap_startwrite(conf->mddev->bitmap, sh->sector,
2068 				  STRIPE_SECTORS, 0);
2069 		sh->bm_seq = conf->seq_flush+1;
2070 		set_bit(STRIPE_BIT_DELAY, &sh->state);
2071 	}
2072 
2073 	if (forwrite) {
2074 		/* check if page is covered */
2075 		sector_t sector = sh->dev[dd_idx].sector;
2076 		for (bi=sh->dev[dd_idx].towrite;
2077 		     sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
2078 			     bi && bi->bi_sector <= sector;
2079 		     bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
2080 			if (bi->bi_sector + (bi->bi_size>>9) >= sector)
2081 				sector = bi->bi_sector + (bi->bi_size>>9);
2082 		}
2083 		if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
2084 			set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
2085 	}
2086 	return 1;
2087 
2088  overlap:
2089 	set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
2090 	spin_unlock_irq(&conf->device_lock);
2091 	spin_unlock(&sh->lock);
2092 	return 0;
2093 }
2094 
2095 static void end_reshape(raid5_conf_t *conf);
2096 
2097 static void stripe_set_idx(sector_t stripe, raid5_conf_t *conf, int previous,
2098 			    struct stripe_head *sh)
2099 {
2100 	int sectors_per_chunk =
2101 		previous ? conf->prev_chunk_sectors : conf->chunk_sectors;
2102 	int dd_idx;
2103 	int chunk_offset = sector_div(stripe, sectors_per_chunk);
2104 	int disks = previous ? conf->previous_raid_disks : conf->raid_disks;
2105 
2106 	raid5_compute_sector(conf,
2107 			     stripe * (disks - conf->max_degraded)
2108 			     *sectors_per_chunk + chunk_offset,
2109 			     previous,
2110 			     &dd_idx, sh);
2111 }
2112 
2113 static void
2114 handle_failed_stripe(raid5_conf_t *conf, struct stripe_head *sh,
2115 				struct stripe_head_state *s, int disks,
2116 				struct bio **return_bi)
2117 {
2118 	int i;
2119 	for (i = disks; i--; ) {
2120 		struct bio *bi;
2121 		int bitmap_end = 0;
2122 
2123 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
2124 			mdk_rdev_t *rdev;
2125 			rcu_read_lock();
2126 			rdev = rcu_dereference(conf->disks[i].rdev);
2127 			if (rdev && test_bit(In_sync, &rdev->flags))
2128 				/* multiple read failures in one stripe */
2129 				md_error(conf->mddev, rdev);
2130 			rcu_read_unlock();
2131 		}
2132 		spin_lock_irq(&conf->device_lock);
2133 		/* fail all writes first */
2134 		bi = sh->dev[i].towrite;
2135 		sh->dev[i].towrite = NULL;
2136 		if (bi) {
2137 			s->to_write--;
2138 			bitmap_end = 1;
2139 		}
2140 
2141 		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2142 			wake_up(&conf->wait_for_overlap);
2143 
2144 		while (bi && bi->bi_sector <
2145 			sh->dev[i].sector + STRIPE_SECTORS) {
2146 			struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
2147 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
2148 			if (!raid5_dec_bi_phys_segments(bi)) {
2149 				md_write_end(conf->mddev);
2150 				bi->bi_next = *return_bi;
2151 				*return_bi = bi;
2152 			}
2153 			bi = nextbi;
2154 		}
2155 		/* and fail all 'written' */
2156 		bi = sh->dev[i].written;
2157 		sh->dev[i].written = NULL;
2158 		if (bi) bitmap_end = 1;
2159 		while (bi && bi->bi_sector <
2160 		       sh->dev[i].sector + STRIPE_SECTORS) {
2161 			struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
2162 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
2163 			if (!raid5_dec_bi_phys_segments(bi)) {
2164 				md_write_end(conf->mddev);
2165 				bi->bi_next = *return_bi;
2166 				*return_bi = bi;
2167 			}
2168 			bi = bi2;
2169 		}
2170 
2171 		/* fail any reads if this device is non-operational and
2172 		 * the data has not reached the cache yet.
2173 		 */
2174 		if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
2175 		    (!test_bit(R5_Insync, &sh->dev[i].flags) ||
2176 		      test_bit(R5_ReadError, &sh->dev[i].flags))) {
2177 			bi = sh->dev[i].toread;
2178 			sh->dev[i].toread = NULL;
2179 			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2180 				wake_up(&conf->wait_for_overlap);
2181 			if (bi) s->to_read--;
2182 			while (bi && bi->bi_sector <
2183 			       sh->dev[i].sector + STRIPE_SECTORS) {
2184 				struct bio *nextbi =
2185 					r5_next_bio(bi, sh->dev[i].sector);
2186 				clear_bit(BIO_UPTODATE, &bi->bi_flags);
2187 				if (!raid5_dec_bi_phys_segments(bi)) {
2188 					bi->bi_next = *return_bi;
2189 					*return_bi = bi;
2190 				}
2191 				bi = nextbi;
2192 			}
2193 		}
2194 		spin_unlock_irq(&conf->device_lock);
2195 		if (bitmap_end)
2196 			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
2197 					STRIPE_SECTORS, 0, 0);
2198 	}
2199 
2200 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2201 		if (atomic_dec_and_test(&conf->pending_full_writes))
2202 			md_wakeup_thread(conf->mddev->thread);
2203 }
2204 
2205 /* fetch_block5 - checks the given member device to see if its data needs
2206  * to be read or computed to satisfy a request.
2207  *
2208  * Returns 1 when no more member devices need to be checked, otherwise returns
2209  * 0 to tell the loop in handle_stripe_fill5 to continue
2210  */
2211 static int fetch_block5(struct stripe_head *sh, struct stripe_head_state *s,
2212 			int disk_idx, int disks)
2213 {
2214 	struct r5dev *dev = &sh->dev[disk_idx];
2215 	struct r5dev *failed_dev = &sh->dev[s->failed_num];
2216 
2217 	/* is the data in this block needed, and can we get it? */
2218 	if (!test_bit(R5_LOCKED, &dev->flags) &&
2219 	    !test_bit(R5_UPTODATE, &dev->flags) &&
2220 	    (dev->toread ||
2221 	     (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2222 	     s->syncing || s->expanding ||
2223 	     (s->failed &&
2224 	      (failed_dev->toread ||
2225 	       (failed_dev->towrite &&
2226 		!test_bit(R5_OVERWRITE, &failed_dev->flags)))))) {
2227 		/* We would like to get this block, possibly by computing it,
2228 		 * otherwise read it if the backing disk is insync
2229 		 */
2230 		if ((s->uptodate == disks - 1) &&
2231 		    (s->failed && disk_idx == s->failed_num)) {
2232 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2233 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2234 			set_bit(R5_Wantcompute, &dev->flags);
2235 			sh->ops.target = disk_idx;
2236 			sh->ops.target2 = -1;
2237 			s->req_compute = 1;
2238 			/* Careful: from this point on 'uptodate' is in the eye
2239 			 * of raid_run_ops which services 'compute' operations
2240 			 * before writes. R5_Wantcompute flags a block that will
2241 			 * be R5_UPTODATE by the time it is needed for a
2242 			 * subsequent operation.
2243 			 */
2244 			s->uptodate++;
2245 			return 1; /* uptodate + compute == disks */
2246 		} else if (test_bit(R5_Insync, &dev->flags)) {
2247 			set_bit(R5_LOCKED, &dev->flags);
2248 			set_bit(R5_Wantread, &dev->flags);
2249 			s->locked++;
2250 			pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2251 				s->syncing);
2252 		}
2253 	}
2254 
2255 	return 0;
2256 }
2257 
2258 /**
2259  * handle_stripe_fill5 - read or compute data to satisfy pending requests.
2260  */
2261 static void handle_stripe_fill5(struct stripe_head *sh,
2262 			struct stripe_head_state *s, int disks)
2263 {
2264 	int i;
2265 
2266 	/* look for blocks to read/compute, skip this if a compute
2267 	 * is already in flight, or if the stripe contents are in the
2268 	 * midst of changing due to a write
2269 	 */
2270 	if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2271 	    !sh->reconstruct_state)
2272 		for (i = disks; i--; )
2273 			if (fetch_block5(sh, s, i, disks))
2274 				break;
2275 	set_bit(STRIPE_HANDLE, &sh->state);
2276 }
2277 
2278 /* fetch_block6 - checks the given member device to see if its data needs
2279  * to be read or computed to satisfy a request.
2280  *
2281  * Returns 1 when no more member devices need to be checked, otherwise returns
2282  * 0 to tell the loop in handle_stripe_fill6 to continue
2283  */
2284 static int fetch_block6(struct stripe_head *sh, struct stripe_head_state *s,
2285 			 struct r6_state *r6s, int disk_idx, int disks)
2286 {
2287 	struct r5dev *dev = &sh->dev[disk_idx];
2288 	struct r5dev *fdev[2] = { &sh->dev[r6s->failed_num[0]],
2289 				  &sh->dev[r6s->failed_num[1]] };
2290 
2291 	if (!test_bit(R5_LOCKED, &dev->flags) &&
2292 	    !test_bit(R5_UPTODATE, &dev->flags) &&
2293 	    (dev->toread ||
2294 	     (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
2295 	     s->syncing || s->expanding ||
2296 	     (s->failed >= 1 &&
2297 	      (fdev[0]->toread || s->to_write)) ||
2298 	     (s->failed >= 2 &&
2299 	      (fdev[1]->toread || s->to_write)))) {
2300 		/* we would like to get this block, possibly by computing it,
2301 		 * otherwise read it if the backing disk is insync
2302 		 */
2303 		BUG_ON(test_bit(R5_Wantcompute, &dev->flags));
2304 		BUG_ON(test_bit(R5_Wantread, &dev->flags));
2305 		if ((s->uptodate == disks - 1) &&
2306 		    (s->failed && (disk_idx == r6s->failed_num[0] ||
2307 				   disk_idx == r6s->failed_num[1]))) {
2308 			/* have disk failed, and we're requested to fetch it;
2309 			 * do compute it
2310 			 */
2311 			pr_debug("Computing stripe %llu block %d\n",
2312 			       (unsigned long long)sh->sector, disk_idx);
2313 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2314 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2315 			set_bit(R5_Wantcompute, &dev->flags);
2316 			sh->ops.target = disk_idx;
2317 			sh->ops.target2 = -1; /* no 2nd target */
2318 			s->req_compute = 1;
2319 			s->uptodate++;
2320 			return 1;
2321 		} else if (s->uptodate == disks-2 && s->failed >= 2) {
2322 			/* Computing 2-failure is *very* expensive; only
2323 			 * do it if failed >= 2
2324 			 */
2325 			int other;
2326 			for (other = disks; other--; ) {
2327 				if (other == disk_idx)
2328 					continue;
2329 				if (!test_bit(R5_UPTODATE,
2330 				      &sh->dev[other].flags))
2331 					break;
2332 			}
2333 			BUG_ON(other < 0);
2334 			pr_debug("Computing stripe %llu blocks %d,%d\n",
2335 			       (unsigned long long)sh->sector,
2336 			       disk_idx, other);
2337 			set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2338 			set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2339 			set_bit(R5_Wantcompute, &sh->dev[disk_idx].flags);
2340 			set_bit(R5_Wantcompute, &sh->dev[other].flags);
2341 			sh->ops.target = disk_idx;
2342 			sh->ops.target2 = other;
2343 			s->uptodate += 2;
2344 			s->req_compute = 1;
2345 			return 1;
2346 		} else if (test_bit(R5_Insync, &dev->flags)) {
2347 			set_bit(R5_LOCKED, &dev->flags);
2348 			set_bit(R5_Wantread, &dev->flags);
2349 			s->locked++;
2350 			pr_debug("Reading block %d (sync=%d)\n",
2351 				disk_idx, s->syncing);
2352 		}
2353 	}
2354 
2355 	return 0;
2356 }
2357 
2358 /**
2359  * handle_stripe_fill6 - read or compute data to satisfy pending requests.
2360  */
2361 static void handle_stripe_fill6(struct stripe_head *sh,
2362 			struct stripe_head_state *s, struct r6_state *r6s,
2363 			int disks)
2364 {
2365 	int i;
2366 
2367 	/* look for blocks to read/compute, skip this if a compute
2368 	 * is already in flight, or if the stripe contents are in the
2369 	 * midst of changing due to a write
2370 	 */
2371 	if (!test_bit(STRIPE_COMPUTE_RUN, &sh->state) && !sh->check_state &&
2372 	    !sh->reconstruct_state)
2373 		for (i = disks; i--; )
2374 			if (fetch_block6(sh, s, r6s, i, disks))
2375 				break;
2376 	set_bit(STRIPE_HANDLE, &sh->state);
2377 }
2378 
2379 
2380 /* handle_stripe_clean_event
2381  * any written block on an uptodate or failed drive can be returned.
2382  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2383  * never LOCKED, so we don't need to test 'failed' directly.
2384  */
2385 static void handle_stripe_clean_event(raid5_conf_t *conf,
2386 	struct stripe_head *sh, int disks, struct bio **return_bi)
2387 {
2388 	int i;
2389 	struct r5dev *dev;
2390 
2391 	for (i = disks; i--; )
2392 		if (sh->dev[i].written) {
2393 			dev = &sh->dev[i];
2394 			if (!test_bit(R5_LOCKED, &dev->flags) &&
2395 				test_bit(R5_UPTODATE, &dev->flags)) {
2396 				/* We can return any write requests */
2397 				struct bio *wbi, *wbi2;
2398 				int bitmap_end = 0;
2399 				pr_debug("Return write for disc %d\n", i);
2400 				spin_lock_irq(&conf->device_lock);
2401 				wbi = dev->written;
2402 				dev->written = NULL;
2403 				while (wbi && wbi->bi_sector <
2404 					dev->sector + STRIPE_SECTORS) {
2405 					wbi2 = r5_next_bio(wbi, dev->sector);
2406 					if (!raid5_dec_bi_phys_segments(wbi)) {
2407 						md_write_end(conf->mddev);
2408 						wbi->bi_next = *return_bi;
2409 						*return_bi = wbi;
2410 					}
2411 					wbi = wbi2;
2412 				}
2413 				if (dev->towrite == NULL)
2414 					bitmap_end = 1;
2415 				spin_unlock_irq(&conf->device_lock);
2416 				if (bitmap_end)
2417 					bitmap_endwrite(conf->mddev->bitmap,
2418 							sh->sector,
2419 							STRIPE_SECTORS,
2420 					 !test_bit(STRIPE_DEGRADED, &sh->state),
2421 							0);
2422 			}
2423 		}
2424 
2425 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2426 		if (atomic_dec_and_test(&conf->pending_full_writes))
2427 			md_wakeup_thread(conf->mddev->thread);
2428 }
2429 
2430 static void handle_stripe_dirtying5(raid5_conf_t *conf,
2431 		struct stripe_head *sh,	struct stripe_head_state *s, int disks)
2432 {
2433 	int rmw = 0, rcw = 0, i;
2434 	for (i = disks; i--; ) {
2435 		/* would I have to read this buffer for read_modify_write */
2436 		struct r5dev *dev = &sh->dev[i];
2437 		if ((dev->towrite || i == sh->pd_idx) &&
2438 		    !test_bit(R5_LOCKED, &dev->flags) &&
2439 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2440 		      test_bit(R5_Wantcompute, &dev->flags))) {
2441 			if (test_bit(R5_Insync, &dev->flags))
2442 				rmw++;
2443 			else
2444 				rmw += 2*disks;  /* cannot read it */
2445 		}
2446 		/* Would I have to read this buffer for reconstruct_write */
2447 		if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2448 		    !test_bit(R5_LOCKED, &dev->flags) &&
2449 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2450 		    test_bit(R5_Wantcompute, &dev->flags))) {
2451 			if (test_bit(R5_Insync, &dev->flags)) rcw++;
2452 			else
2453 				rcw += 2*disks;
2454 		}
2455 	}
2456 	pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2457 		(unsigned long long)sh->sector, rmw, rcw);
2458 	set_bit(STRIPE_HANDLE, &sh->state);
2459 	if (rmw < rcw && rmw > 0)
2460 		/* prefer read-modify-write, but need to get some data */
2461 		for (i = disks; i--; ) {
2462 			struct r5dev *dev = &sh->dev[i];
2463 			if ((dev->towrite || i == sh->pd_idx) &&
2464 			    !test_bit(R5_LOCKED, &dev->flags) &&
2465 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
2466 			    test_bit(R5_Wantcompute, &dev->flags)) &&
2467 			    test_bit(R5_Insync, &dev->flags)) {
2468 				if (
2469 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2470 					pr_debug("Read_old block "
2471 						"%d for r-m-w\n", i);
2472 					set_bit(R5_LOCKED, &dev->flags);
2473 					set_bit(R5_Wantread, &dev->flags);
2474 					s->locked++;
2475 				} else {
2476 					set_bit(STRIPE_DELAYED, &sh->state);
2477 					set_bit(STRIPE_HANDLE, &sh->state);
2478 				}
2479 			}
2480 		}
2481 	if (rcw <= rmw && rcw > 0)
2482 		/* want reconstruct write, but need to get some data */
2483 		for (i = disks; i--; ) {
2484 			struct r5dev *dev = &sh->dev[i];
2485 			if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2486 			    i != sh->pd_idx &&
2487 			    !test_bit(R5_LOCKED, &dev->flags) &&
2488 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
2489 			    test_bit(R5_Wantcompute, &dev->flags)) &&
2490 			    test_bit(R5_Insync, &dev->flags)) {
2491 				if (
2492 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2493 					pr_debug("Read_old block "
2494 						"%d for Reconstruct\n", i);
2495 					set_bit(R5_LOCKED, &dev->flags);
2496 					set_bit(R5_Wantread, &dev->flags);
2497 					s->locked++;
2498 				} else {
2499 					set_bit(STRIPE_DELAYED, &sh->state);
2500 					set_bit(STRIPE_HANDLE, &sh->state);
2501 				}
2502 			}
2503 		}
2504 	/* now if nothing is locked, and if we have enough data,
2505 	 * we can start a write request
2506 	 */
2507 	/* since handle_stripe can be called at any time we need to handle the
2508 	 * case where a compute block operation has been submitted and then a
2509 	 * subsequent call wants to start a write request.  raid_run_ops only
2510 	 * handles the case where compute block and reconstruct are requested
2511 	 * simultaneously.  If this is not the case then new writes need to be
2512 	 * held off until the compute completes.
2513 	 */
2514 	if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2515 	    (s->locked == 0 && (rcw == 0 || rmw == 0) &&
2516 	    !test_bit(STRIPE_BIT_DELAY, &sh->state)))
2517 		schedule_reconstruction(sh, s, rcw == 0, 0);
2518 }
2519 
2520 static void handle_stripe_dirtying6(raid5_conf_t *conf,
2521 		struct stripe_head *sh,	struct stripe_head_state *s,
2522 		struct r6_state *r6s, int disks)
2523 {
2524 	int rcw = 0, pd_idx = sh->pd_idx, i;
2525 	int qd_idx = sh->qd_idx;
2526 
2527 	set_bit(STRIPE_HANDLE, &sh->state);
2528 	for (i = disks; i--; ) {
2529 		struct r5dev *dev = &sh->dev[i];
2530 		/* check if we haven't enough data */
2531 		if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2532 		    i != pd_idx && i != qd_idx &&
2533 		    !test_bit(R5_LOCKED, &dev->flags) &&
2534 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2535 		      test_bit(R5_Wantcompute, &dev->flags))) {
2536 			rcw++;
2537 			if (!test_bit(R5_Insync, &dev->flags))
2538 				continue; /* it's a failed drive */
2539 
2540 			if (
2541 			  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2542 				pr_debug("Read_old stripe %llu "
2543 					"block %d for Reconstruct\n",
2544 				     (unsigned long long)sh->sector, i);
2545 				set_bit(R5_LOCKED, &dev->flags);
2546 				set_bit(R5_Wantread, &dev->flags);
2547 				s->locked++;
2548 			} else {
2549 				pr_debug("Request delayed stripe %llu "
2550 					"block %d for Reconstruct\n",
2551 				     (unsigned long long)sh->sector, i);
2552 				set_bit(STRIPE_DELAYED, &sh->state);
2553 				set_bit(STRIPE_HANDLE, &sh->state);
2554 			}
2555 		}
2556 	}
2557 	/* now if nothing is locked, and if we have enough data, we can start a
2558 	 * write request
2559 	 */
2560 	if ((s->req_compute || !test_bit(STRIPE_COMPUTE_RUN, &sh->state)) &&
2561 	    s->locked == 0 && rcw == 0 &&
2562 	    !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2563 		schedule_reconstruction(sh, s, 1, 0);
2564 	}
2565 }
2566 
2567 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2568 				struct stripe_head_state *s, int disks)
2569 {
2570 	struct r5dev *dev = NULL;
2571 
2572 	set_bit(STRIPE_HANDLE, &sh->state);
2573 
2574 	switch (sh->check_state) {
2575 	case check_state_idle:
2576 		/* start a new check operation if there are no failures */
2577 		if (s->failed == 0) {
2578 			BUG_ON(s->uptodate != disks);
2579 			sh->check_state = check_state_run;
2580 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
2581 			clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2582 			s->uptodate--;
2583 			break;
2584 		}
2585 		dev = &sh->dev[s->failed_num];
2586 		/* fall through */
2587 	case check_state_compute_result:
2588 		sh->check_state = check_state_idle;
2589 		if (!dev)
2590 			dev = &sh->dev[sh->pd_idx];
2591 
2592 		/* check that a write has not made the stripe insync */
2593 		if (test_bit(STRIPE_INSYNC, &sh->state))
2594 			break;
2595 
2596 		/* either failed parity check, or recovery is happening */
2597 		BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2598 		BUG_ON(s->uptodate != disks);
2599 
2600 		set_bit(R5_LOCKED, &dev->flags);
2601 		s->locked++;
2602 		set_bit(R5_Wantwrite, &dev->flags);
2603 
2604 		clear_bit(STRIPE_DEGRADED, &sh->state);
2605 		set_bit(STRIPE_INSYNC, &sh->state);
2606 		break;
2607 	case check_state_run:
2608 		break; /* we will be called again upon completion */
2609 	case check_state_check_result:
2610 		sh->check_state = check_state_idle;
2611 
2612 		/* if a failure occurred during the check operation, leave
2613 		 * STRIPE_INSYNC not set and let the stripe be handled again
2614 		 */
2615 		if (s->failed)
2616 			break;
2617 
2618 		/* handle a successful check operation, if parity is correct
2619 		 * we are done.  Otherwise update the mismatch count and repair
2620 		 * parity if !MD_RECOVERY_CHECK
2621 		 */
2622 		if ((sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) == 0)
2623 			/* parity is correct (on disc,
2624 			 * not in buffer any more)
2625 			 */
2626 			set_bit(STRIPE_INSYNC, &sh->state);
2627 		else {
2628 			conf->mddev->resync_mismatches += STRIPE_SECTORS;
2629 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2630 				/* don't try to repair!! */
2631 				set_bit(STRIPE_INSYNC, &sh->state);
2632 			else {
2633 				sh->check_state = check_state_compute_run;
2634 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2635 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2636 				set_bit(R5_Wantcompute,
2637 					&sh->dev[sh->pd_idx].flags);
2638 				sh->ops.target = sh->pd_idx;
2639 				sh->ops.target2 = -1;
2640 				s->uptodate++;
2641 			}
2642 		}
2643 		break;
2644 	case check_state_compute_run:
2645 		break;
2646 	default:
2647 		printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2648 		       __func__, sh->check_state,
2649 		       (unsigned long long) sh->sector);
2650 		BUG();
2651 	}
2652 }
2653 
2654 
2655 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2656 				  struct stripe_head_state *s,
2657 				  struct r6_state *r6s, int disks)
2658 {
2659 	int pd_idx = sh->pd_idx;
2660 	int qd_idx = sh->qd_idx;
2661 	struct r5dev *dev;
2662 
2663 	set_bit(STRIPE_HANDLE, &sh->state);
2664 
2665 	BUG_ON(s->failed > 2);
2666 
2667 	/* Want to check and possibly repair P and Q.
2668 	 * However there could be one 'failed' device, in which
2669 	 * case we can only check one of them, possibly using the
2670 	 * other to generate missing data
2671 	 */
2672 
2673 	switch (sh->check_state) {
2674 	case check_state_idle:
2675 		/* start a new check operation if there are < 2 failures */
2676 		if (s->failed == r6s->q_failed) {
2677 			/* The only possible failed device holds Q, so it
2678 			 * makes sense to check P (If anything else were failed,
2679 			 * we would have used P to recreate it).
2680 			 */
2681 			sh->check_state = check_state_run;
2682 		}
2683 		if (!r6s->q_failed && s->failed < 2) {
2684 			/* Q is not failed, and we didn't use it to generate
2685 			 * anything, so it makes sense to check it
2686 			 */
2687 			if (sh->check_state == check_state_run)
2688 				sh->check_state = check_state_run_pq;
2689 			else
2690 				sh->check_state = check_state_run_q;
2691 		}
2692 
2693 		/* discard potentially stale zero_sum_result */
2694 		sh->ops.zero_sum_result = 0;
2695 
2696 		if (sh->check_state == check_state_run) {
2697 			/* async_xor_zero_sum destroys the contents of P */
2698 			clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
2699 			s->uptodate--;
2700 		}
2701 		if (sh->check_state >= check_state_run &&
2702 		    sh->check_state <= check_state_run_pq) {
2703 			/* async_syndrome_zero_sum preserves P and Q, so
2704 			 * no need to mark them !uptodate here
2705 			 */
2706 			set_bit(STRIPE_OP_CHECK, &s->ops_request);
2707 			break;
2708 		}
2709 
2710 		/* we have 2-disk failure */
2711 		BUG_ON(s->failed != 2);
2712 		/* fall through */
2713 	case check_state_compute_result:
2714 		sh->check_state = check_state_idle;
2715 
2716 		/* check that a write has not made the stripe insync */
2717 		if (test_bit(STRIPE_INSYNC, &sh->state))
2718 			break;
2719 
2720 		/* now write out any block on a failed drive,
2721 		 * or P or Q if they were recomputed
2722 		 */
2723 		BUG_ON(s->uptodate < disks - 1); /* We don't need Q to recover */
2724 		if (s->failed == 2) {
2725 			dev = &sh->dev[r6s->failed_num[1]];
2726 			s->locked++;
2727 			set_bit(R5_LOCKED, &dev->flags);
2728 			set_bit(R5_Wantwrite, &dev->flags);
2729 		}
2730 		if (s->failed >= 1) {
2731 			dev = &sh->dev[r6s->failed_num[0]];
2732 			s->locked++;
2733 			set_bit(R5_LOCKED, &dev->flags);
2734 			set_bit(R5_Wantwrite, &dev->flags);
2735 		}
2736 		if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2737 			dev = &sh->dev[pd_idx];
2738 			s->locked++;
2739 			set_bit(R5_LOCKED, &dev->flags);
2740 			set_bit(R5_Wantwrite, &dev->flags);
2741 		}
2742 		if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2743 			dev = &sh->dev[qd_idx];
2744 			s->locked++;
2745 			set_bit(R5_LOCKED, &dev->flags);
2746 			set_bit(R5_Wantwrite, &dev->flags);
2747 		}
2748 		clear_bit(STRIPE_DEGRADED, &sh->state);
2749 
2750 		set_bit(STRIPE_INSYNC, &sh->state);
2751 		break;
2752 	case check_state_run:
2753 	case check_state_run_q:
2754 	case check_state_run_pq:
2755 		break; /* we will be called again upon completion */
2756 	case check_state_check_result:
2757 		sh->check_state = check_state_idle;
2758 
2759 		/* handle a successful check operation, if parity is correct
2760 		 * we are done.  Otherwise update the mismatch count and repair
2761 		 * parity if !MD_RECOVERY_CHECK
2762 		 */
2763 		if (sh->ops.zero_sum_result == 0) {
2764 			/* both parities are correct */
2765 			if (!s->failed)
2766 				set_bit(STRIPE_INSYNC, &sh->state);
2767 			else {
2768 				/* in contrast to the raid5 case we can validate
2769 				 * parity, but still have a failure to write
2770 				 * back
2771 				 */
2772 				sh->check_state = check_state_compute_result;
2773 				/* Returning at this point means that we may go
2774 				 * off and bring p and/or q uptodate again so
2775 				 * we make sure to check zero_sum_result again
2776 				 * to verify if p or q need writeback
2777 				 */
2778 			}
2779 		} else {
2780 			conf->mddev->resync_mismatches += STRIPE_SECTORS;
2781 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2782 				/* don't try to repair!! */
2783 				set_bit(STRIPE_INSYNC, &sh->state);
2784 			else {
2785 				int *target = &sh->ops.target;
2786 
2787 				sh->ops.target = -1;
2788 				sh->ops.target2 = -1;
2789 				sh->check_state = check_state_compute_run;
2790 				set_bit(STRIPE_COMPUTE_RUN, &sh->state);
2791 				set_bit(STRIPE_OP_COMPUTE_BLK, &s->ops_request);
2792 				if (sh->ops.zero_sum_result & SUM_CHECK_P_RESULT) {
2793 					set_bit(R5_Wantcompute,
2794 						&sh->dev[pd_idx].flags);
2795 					*target = pd_idx;
2796 					target = &sh->ops.target2;
2797 					s->uptodate++;
2798 				}
2799 				if (sh->ops.zero_sum_result & SUM_CHECK_Q_RESULT) {
2800 					set_bit(R5_Wantcompute,
2801 						&sh->dev[qd_idx].flags);
2802 					*target = qd_idx;
2803 					s->uptodate++;
2804 				}
2805 			}
2806 		}
2807 		break;
2808 	case check_state_compute_run:
2809 		break;
2810 	default:
2811 		printk(KERN_ERR "%s: unknown check_state: %d sector: %llu\n",
2812 		       __func__, sh->check_state,
2813 		       (unsigned long long) sh->sector);
2814 		BUG();
2815 	}
2816 }
2817 
2818 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2819 				struct r6_state *r6s)
2820 {
2821 	int i;
2822 
2823 	/* We have read all the blocks in this stripe and now we need to
2824 	 * copy some of them into a target stripe for expand.
2825 	 */
2826 	struct dma_async_tx_descriptor *tx = NULL;
2827 	clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2828 	for (i = 0; i < sh->disks; i++)
2829 		if (i != sh->pd_idx && i != sh->qd_idx) {
2830 			int dd_idx, j;
2831 			struct stripe_head *sh2;
2832 			struct async_submit_ctl submit;
2833 
2834 			sector_t bn = compute_blocknr(sh, i, 1);
2835 			sector_t s = raid5_compute_sector(conf, bn, 0,
2836 							  &dd_idx, NULL);
2837 			sh2 = get_active_stripe(conf, s, 0, 1, 1);
2838 			if (sh2 == NULL)
2839 				/* so far only the early blocks of this stripe
2840 				 * have been requested.  When later blocks
2841 				 * get requested, we will try again
2842 				 */
2843 				continue;
2844 			if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2845 			   test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2846 				/* must have already done this block */
2847 				release_stripe(sh2);
2848 				continue;
2849 			}
2850 
2851 			/* place all the copies on one channel */
2852 			init_async_submit(&submit, 0, tx, NULL, NULL, NULL);
2853 			tx = async_memcpy(sh2->dev[dd_idx].page,
2854 					  sh->dev[i].page, 0, 0, STRIPE_SIZE,
2855 					  &submit);
2856 
2857 			set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2858 			set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2859 			for (j = 0; j < conf->raid_disks; j++)
2860 				if (j != sh2->pd_idx &&
2861 				    (!r6s || j != sh2->qd_idx) &&
2862 				    !test_bit(R5_Expanded, &sh2->dev[j].flags))
2863 					break;
2864 			if (j == conf->raid_disks) {
2865 				set_bit(STRIPE_EXPAND_READY, &sh2->state);
2866 				set_bit(STRIPE_HANDLE, &sh2->state);
2867 			}
2868 			release_stripe(sh2);
2869 
2870 		}
2871 	/* done submitting copies, wait for them to complete */
2872 	if (tx) {
2873 		async_tx_ack(tx);
2874 		dma_wait_for_async_tx(tx);
2875 	}
2876 }
2877 
2878 
2879 /*
2880  * handle_stripe - do things to a stripe.
2881  *
2882  * We lock the stripe and then examine the state of various bits
2883  * to see what needs to be done.
2884  * Possible results:
2885  *    return some read request which now have data
2886  *    return some write requests which are safely on disc
2887  *    schedule a read on some buffers
2888  *    schedule a write of some buffers
2889  *    return confirmation of parity correctness
2890  *
2891  * buffers are taken off read_list or write_list, and bh_cache buffers
2892  * get BH_Lock set before the stripe lock is released.
2893  *
2894  */
2895 
2896 static bool handle_stripe5(struct stripe_head *sh)
2897 {
2898 	raid5_conf_t *conf = sh->raid_conf;
2899 	int disks = sh->disks, i;
2900 	struct bio *return_bi = NULL;
2901 	struct stripe_head_state s;
2902 	struct r5dev *dev;
2903 	mdk_rdev_t *blocked_rdev = NULL;
2904 	int prexor;
2905 
2906 	memset(&s, 0, sizeof(s));
2907 	pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d check:%d "
2908 		 "reconstruct:%d\n", (unsigned long long)sh->sector, sh->state,
2909 		 atomic_read(&sh->count), sh->pd_idx, sh->check_state,
2910 		 sh->reconstruct_state);
2911 
2912 	spin_lock(&sh->lock);
2913 	clear_bit(STRIPE_HANDLE, &sh->state);
2914 	clear_bit(STRIPE_DELAYED, &sh->state);
2915 
2916 	s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2917 	s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2918 	s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2919 
2920 	/* Now to look around and see what can be done */
2921 	rcu_read_lock();
2922 	for (i=disks; i--; ) {
2923 		mdk_rdev_t *rdev;
2924 		struct r5dev *dev = &sh->dev[i];
2925 		clear_bit(R5_Insync, &dev->flags);
2926 
2927 		pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2928 			"written %p\n",	i, dev->flags, dev->toread, dev->read,
2929 			dev->towrite, dev->written);
2930 
2931 		/* maybe we can request a biofill operation
2932 		 *
2933 		 * new wantfill requests are only permitted while
2934 		 * ops_complete_biofill is guaranteed to be inactive
2935 		 */
2936 		if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2937 		    !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
2938 			set_bit(R5_Wantfill, &dev->flags);
2939 
2940 		/* now count some things */
2941 		if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2942 		if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2943 		if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2944 
2945 		if (test_bit(R5_Wantfill, &dev->flags))
2946 			s.to_fill++;
2947 		else if (dev->toread)
2948 			s.to_read++;
2949 		if (dev->towrite) {
2950 			s.to_write++;
2951 			if (!test_bit(R5_OVERWRITE, &dev->flags))
2952 				s.non_overwrite++;
2953 		}
2954 		if (dev->written)
2955 			s.written++;
2956 		rdev = rcu_dereference(conf->disks[i].rdev);
2957 		if (blocked_rdev == NULL &&
2958 		    rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2959 			blocked_rdev = rdev;
2960 			atomic_inc(&rdev->nr_pending);
2961 		}
2962 		if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2963 			/* The ReadError flag will just be confusing now */
2964 			clear_bit(R5_ReadError, &dev->flags);
2965 			clear_bit(R5_ReWrite, &dev->flags);
2966 		}
2967 		if (!rdev || !test_bit(In_sync, &rdev->flags)
2968 		    || test_bit(R5_ReadError, &dev->flags)) {
2969 			s.failed++;
2970 			s.failed_num = i;
2971 		} else
2972 			set_bit(R5_Insync, &dev->flags);
2973 	}
2974 	rcu_read_unlock();
2975 
2976 	if (unlikely(blocked_rdev)) {
2977 		if (s.syncing || s.expanding || s.expanded ||
2978 		    s.to_write || s.written) {
2979 			set_bit(STRIPE_HANDLE, &sh->state);
2980 			goto unlock;
2981 		}
2982 		/* There is nothing for the blocked_rdev to block */
2983 		rdev_dec_pending(blocked_rdev, conf->mddev);
2984 		blocked_rdev = NULL;
2985 	}
2986 
2987 	if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
2988 		set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
2989 		set_bit(STRIPE_BIOFILL_RUN, &sh->state);
2990 	}
2991 
2992 	pr_debug("locked=%d uptodate=%d to_read=%d"
2993 		" to_write=%d failed=%d failed_num=%d\n",
2994 		s.locked, s.uptodate, s.to_read, s.to_write,
2995 		s.failed, s.failed_num);
2996 	/* check if the array has lost two devices and, if so, some requests might
2997 	 * need to be failed
2998 	 */
2999 	if (s.failed > 1 && s.to_read+s.to_write+s.written)
3000 		handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3001 	if (s.failed > 1 && s.syncing) {
3002 		md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3003 		clear_bit(STRIPE_SYNCING, &sh->state);
3004 		s.syncing = 0;
3005 	}
3006 
3007 	/* might be able to return some write requests if the parity block
3008 	 * is safe, or on a failed drive
3009 	 */
3010 	dev = &sh->dev[sh->pd_idx];
3011 	if ( s.written &&
3012 	     ((test_bit(R5_Insync, &dev->flags) &&
3013 	       !test_bit(R5_LOCKED, &dev->flags) &&
3014 	       test_bit(R5_UPTODATE, &dev->flags)) ||
3015 	       (s.failed == 1 && s.failed_num == sh->pd_idx)))
3016 		handle_stripe_clean_event(conf, sh, disks, &return_bi);
3017 
3018 	/* Now we might consider reading some blocks, either to check/generate
3019 	 * parity, or to satisfy requests
3020 	 * or to load a block that is being partially written.
3021 	 */
3022 	if (s.to_read || s.non_overwrite ||
3023 	    (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3024 		handle_stripe_fill5(sh, &s, disks);
3025 
3026 	/* Now we check to see if any write operations have recently
3027 	 * completed
3028 	 */
3029 	prexor = 0;
3030 	if (sh->reconstruct_state == reconstruct_state_prexor_drain_result)
3031 		prexor = 1;
3032 	if (sh->reconstruct_state == reconstruct_state_drain_result ||
3033 	    sh->reconstruct_state == reconstruct_state_prexor_drain_result) {
3034 		sh->reconstruct_state = reconstruct_state_idle;
3035 
3036 		/* All the 'written' buffers and the parity block are ready to
3037 		 * be written back to disk
3038 		 */
3039 		BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3040 		for (i = disks; i--; ) {
3041 			dev = &sh->dev[i];
3042 			if (test_bit(R5_LOCKED, &dev->flags) &&
3043 				(i == sh->pd_idx || dev->written)) {
3044 				pr_debug("Writing block %d\n", i);
3045 				set_bit(R5_Wantwrite, &dev->flags);
3046 				if (prexor)
3047 					continue;
3048 				if (!test_bit(R5_Insync, &dev->flags) ||
3049 				    (i == sh->pd_idx && s.failed == 0))
3050 					set_bit(STRIPE_INSYNC, &sh->state);
3051 			}
3052 		}
3053 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3054 			atomic_dec(&conf->preread_active_stripes);
3055 			if (atomic_read(&conf->preread_active_stripes) <
3056 				IO_THRESHOLD)
3057 				md_wakeup_thread(conf->mddev->thread);
3058 		}
3059 	}
3060 
3061 	/* Now to consider new write requests and what else, if anything
3062 	 * should be read.  We do not handle new writes when:
3063 	 * 1/ A 'write' operation (copy+xor) is already in flight.
3064 	 * 2/ A 'check' operation is in flight, as it may clobber the parity
3065 	 *    block.
3066 	 */
3067 	if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3068 		handle_stripe_dirtying5(conf, sh, &s, disks);
3069 
3070 	/* maybe we need to check and possibly fix the parity for this stripe
3071 	 * Any reads will already have been scheduled, so we just see if enough
3072 	 * data is available.  The parity check is held off while parity
3073 	 * dependent operations are in flight.
3074 	 */
3075 	if (sh->check_state ||
3076 	    (s.syncing && s.locked == 0 &&
3077 	     !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3078 	     !test_bit(STRIPE_INSYNC, &sh->state)))
3079 		handle_parity_checks5(conf, sh, &s, disks);
3080 
3081 	if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3082 		md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3083 		clear_bit(STRIPE_SYNCING, &sh->state);
3084 	}
3085 
3086 	/* If the failed drive is just a ReadError, then we might need to progress
3087 	 * the repair/check process
3088 	 */
3089 	if (s.failed == 1 && !conf->mddev->ro &&
3090 	    test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
3091 	    && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
3092 	    && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
3093 		) {
3094 		dev = &sh->dev[s.failed_num];
3095 		if (!test_bit(R5_ReWrite, &dev->flags)) {
3096 			set_bit(R5_Wantwrite, &dev->flags);
3097 			set_bit(R5_ReWrite, &dev->flags);
3098 			set_bit(R5_LOCKED, &dev->flags);
3099 			s.locked++;
3100 		} else {
3101 			/* let's read it back */
3102 			set_bit(R5_Wantread, &dev->flags);
3103 			set_bit(R5_LOCKED, &dev->flags);
3104 			s.locked++;
3105 		}
3106 	}
3107 
3108 	/* Finish reconstruct operations initiated by the expansion process */
3109 	if (sh->reconstruct_state == reconstruct_state_result) {
3110 		struct stripe_head *sh2
3111 			= get_active_stripe(conf, sh->sector, 1, 1, 1);
3112 		if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3113 			/* sh cannot be written until sh2 has been read.
3114 			 * so arrange for sh to be delayed a little
3115 			 */
3116 			set_bit(STRIPE_DELAYED, &sh->state);
3117 			set_bit(STRIPE_HANDLE, &sh->state);
3118 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3119 					      &sh2->state))
3120 				atomic_inc(&conf->preread_active_stripes);
3121 			release_stripe(sh2);
3122 			goto unlock;
3123 		}
3124 		if (sh2)
3125 			release_stripe(sh2);
3126 
3127 		sh->reconstruct_state = reconstruct_state_idle;
3128 		clear_bit(STRIPE_EXPANDING, &sh->state);
3129 		for (i = conf->raid_disks; i--; ) {
3130 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
3131 			set_bit(R5_LOCKED, &sh->dev[i].flags);
3132 			s.locked++;
3133 		}
3134 	}
3135 
3136 	if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3137 	    !sh->reconstruct_state) {
3138 		/* Need to write out all blocks after computing parity */
3139 		sh->disks = conf->raid_disks;
3140 		stripe_set_idx(sh->sector, conf, 0, sh);
3141 		schedule_reconstruction(sh, &s, 1, 1);
3142 	} else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3143 		clear_bit(STRIPE_EXPAND_READY, &sh->state);
3144 		atomic_dec(&conf->reshape_stripes);
3145 		wake_up(&conf->wait_for_overlap);
3146 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3147 	}
3148 
3149 	if (s.expanding && s.locked == 0 &&
3150 	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3151 		handle_stripe_expansion(conf, sh, NULL);
3152 
3153  unlock:
3154 	spin_unlock(&sh->lock);
3155 
3156 	/* wait for this device to become unblocked */
3157 	if (unlikely(blocked_rdev))
3158 		md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3159 
3160 	if (s.ops_request)
3161 		raid_run_ops(sh, s.ops_request);
3162 
3163 	ops_run_io(sh, &s);
3164 
3165 	return_io(return_bi);
3166 
3167 	return blocked_rdev == NULL;
3168 }
3169 
3170 static bool handle_stripe6(struct stripe_head *sh)
3171 {
3172 	raid5_conf_t *conf = sh->raid_conf;
3173 	int disks = sh->disks;
3174 	struct bio *return_bi = NULL;
3175 	int i, pd_idx = sh->pd_idx, qd_idx = sh->qd_idx;
3176 	struct stripe_head_state s;
3177 	struct r6_state r6s;
3178 	struct r5dev *dev, *pdev, *qdev;
3179 	mdk_rdev_t *blocked_rdev = NULL;
3180 
3181 	pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
3182 		"pd_idx=%d, qd_idx=%d\n, check:%d, reconstruct:%d\n",
3183 	       (unsigned long long)sh->sector, sh->state,
3184 	       atomic_read(&sh->count), pd_idx, qd_idx,
3185 	       sh->check_state, sh->reconstruct_state);
3186 	memset(&s, 0, sizeof(s));
3187 
3188 	spin_lock(&sh->lock);
3189 	clear_bit(STRIPE_HANDLE, &sh->state);
3190 	clear_bit(STRIPE_DELAYED, &sh->state);
3191 
3192 	s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
3193 	s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3194 	s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
3195 	/* Now to look around and see what can be done */
3196 
3197 	rcu_read_lock();
3198 	for (i=disks; i--; ) {
3199 		mdk_rdev_t *rdev;
3200 		dev = &sh->dev[i];
3201 		clear_bit(R5_Insync, &dev->flags);
3202 
3203 		pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
3204 			i, dev->flags, dev->toread, dev->towrite, dev->written);
3205 		/* maybe we can reply to a read
3206 		 *
3207 		 * new wantfill requests are only permitted while
3208 		 * ops_complete_biofill is guaranteed to be inactive
3209 		 */
3210 		if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
3211 		    !test_bit(STRIPE_BIOFILL_RUN, &sh->state))
3212 			set_bit(R5_Wantfill, &dev->flags);
3213 
3214 		/* now count some things */
3215 		if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
3216 		if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
3217 		if (test_bit(R5_Wantcompute, &dev->flags))
3218 			BUG_ON(++s.compute > 2);
3219 
3220 		if (test_bit(R5_Wantfill, &dev->flags)) {
3221 			s.to_fill++;
3222 		} else if (dev->toread)
3223 			s.to_read++;
3224 		if (dev->towrite) {
3225 			s.to_write++;
3226 			if (!test_bit(R5_OVERWRITE, &dev->flags))
3227 				s.non_overwrite++;
3228 		}
3229 		if (dev->written)
3230 			s.written++;
3231 		rdev = rcu_dereference(conf->disks[i].rdev);
3232 		if (blocked_rdev == NULL &&
3233 		    rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
3234 			blocked_rdev = rdev;
3235 			atomic_inc(&rdev->nr_pending);
3236 		}
3237 		if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3238 			/* The ReadError flag will just be confusing now */
3239 			clear_bit(R5_ReadError, &dev->flags);
3240 			clear_bit(R5_ReWrite, &dev->flags);
3241 		}
3242 		if (!rdev || !test_bit(In_sync, &rdev->flags)
3243 		    || test_bit(R5_ReadError, &dev->flags)) {
3244 			if (s.failed < 2)
3245 				r6s.failed_num[s.failed] = i;
3246 			s.failed++;
3247 		} else
3248 			set_bit(R5_Insync, &dev->flags);
3249 	}
3250 	rcu_read_unlock();
3251 
3252 	if (unlikely(blocked_rdev)) {
3253 		if (s.syncing || s.expanding || s.expanded ||
3254 		    s.to_write || s.written) {
3255 			set_bit(STRIPE_HANDLE, &sh->state);
3256 			goto unlock;
3257 		}
3258 		/* There is nothing for the blocked_rdev to block */
3259 		rdev_dec_pending(blocked_rdev, conf->mddev);
3260 		blocked_rdev = NULL;
3261 	}
3262 
3263 	if (s.to_fill && !test_bit(STRIPE_BIOFILL_RUN, &sh->state)) {
3264 		set_bit(STRIPE_OP_BIOFILL, &s.ops_request);
3265 		set_bit(STRIPE_BIOFILL_RUN, &sh->state);
3266 	}
3267 
3268 	pr_debug("locked=%d uptodate=%d to_read=%d"
3269 	       " to_write=%d failed=%d failed_num=%d,%d\n",
3270 	       s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3271 	       r6s.failed_num[0], r6s.failed_num[1]);
3272 	/* check if the array has lost >2 devices and, if so, some requests
3273 	 * might need to be failed
3274 	 */
3275 	if (s.failed > 2 && s.to_read+s.to_write+s.written)
3276 		handle_failed_stripe(conf, sh, &s, disks, &return_bi);
3277 	if (s.failed > 2 && s.syncing) {
3278 		md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3279 		clear_bit(STRIPE_SYNCING, &sh->state);
3280 		s.syncing = 0;
3281 	}
3282 
3283 	/*
3284 	 * might be able to return some write requests if the parity blocks
3285 	 * are safe, or on a failed drive
3286 	 */
3287 	pdev = &sh->dev[pd_idx];
3288 	r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3289 		|| (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3290 	qdev = &sh->dev[qd_idx];
3291 	r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == qd_idx)
3292 		|| (s.failed >= 2 && r6s.failed_num[1] == qd_idx);
3293 
3294 	if ( s.written &&
3295 	     ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3296 			     && !test_bit(R5_LOCKED, &pdev->flags)
3297 			     && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3298 	     ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3299 			     && !test_bit(R5_LOCKED, &qdev->flags)
3300 			     && test_bit(R5_UPTODATE, &qdev->flags)))))
3301 		handle_stripe_clean_event(conf, sh, disks, &return_bi);
3302 
3303 	/* Now we might consider reading some blocks, either to check/generate
3304 	 * parity, or to satisfy requests
3305 	 * or to load a block that is being partially written.
3306 	 */
3307 	if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3308 	    (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding)
3309 		handle_stripe_fill6(sh, &s, &r6s, disks);
3310 
3311 	/* Now we check to see if any write operations have recently
3312 	 * completed
3313 	 */
3314 	if (sh->reconstruct_state == reconstruct_state_drain_result) {
3315 		int qd_idx = sh->qd_idx;
3316 
3317 		sh->reconstruct_state = reconstruct_state_idle;
3318 		/* All the 'written' buffers and the parity blocks are ready to
3319 		 * be written back to disk
3320 		 */
3321 		BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
3322 		BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[qd_idx].flags));
3323 		for (i = disks; i--; ) {
3324 			dev = &sh->dev[i];
3325 			if (test_bit(R5_LOCKED, &dev->flags) &&
3326 			    (i == sh->pd_idx || i == qd_idx ||
3327 			     dev->written)) {
3328 				pr_debug("Writing block %d\n", i);
3329 				BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
3330 				set_bit(R5_Wantwrite, &dev->flags);
3331 				if (!test_bit(R5_Insync, &dev->flags) ||
3332 				    ((i == sh->pd_idx || i == qd_idx) &&
3333 				      s.failed == 0))
3334 					set_bit(STRIPE_INSYNC, &sh->state);
3335 			}
3336 		}
3337 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
3338 			atomic_dec(&conf->preread_active_stripes);
3339 			if (atomic_read(&conf->preread_active_stripes) <
3340 				IO_THRESHOLD)
3341 				md_wakeup_thread(conf->mddev->thread);
3342 		}
3343 	}
3344 
3345 	/* Now to consider new write requests and what else, if anything
3346 	 * should be read.  We do not handle new writes when:
3347 	 * 1/ A 'write' operation (copy+gen_syndrome) is already in flight.
3348 	 * 2/ A 'check' operation is in flight, as it may clobber the parity
3349 	 *    block.
3350 	 */
3351 	if (s.to_write && !sh->reconstruct_state && !sh->check_state)
3352 		handle_stripe_dirtying6(conf, sh, &s, &r6s, disks);
3353 
3354 	/* maybe we need to check and possibly fix the parity for this stripe
3355 	 * Any reads will already have been scheduled, so we just see if enough
3356 	 * data is available.  The parity check is held off while parity
3357 	 * dependent operations are in flight.
3358 	 */
3359 	if (sh->check_state ||
3360 	    (s.syncing && s.locked == 0 &&
3361 	     !test_bit(STRIPE_COMPUTE_RUN, &sh->state) &&
3362 	     !test_bit(STRIPE_INSYNC, &sh->state)))
3363 		handle_parity_checks6(conf, sh, &s, &r6s, disks);
3364 
3365 	if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3366 		md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3367 		clear_bit(STRIPE_SYNCING, &sh->state);
3368 	}
3369 
3370 	/* If the failed drives are just a ReadError, then we might need
3371 	 * to progress the repair/check process
3372 	 */
3373 	if (s.failed <= 2 && !conf->mddev->ro)
3374 		for (i = 0; i < s.failed; i++) {
3375 			dev = &sh->dev[r6s.failed_num[i]];
3376 			if (test_bit(R5_ReadError, &dev->flags)
3377 			    && !test_bit(R5_LOCKED, &dev->flags)
3378 			    && test_bit(R5_UPTODATE, &dev->flags)
3379 				) {
3380 				if (!test_bit(R5_ReWrite, &dev->flags)) {
3381 					set_bit(R5_Wantwrite, &dev->flags);
3382 					set_bit(R5_ReWrite, &dev->flags);
3383 					set_bit(R5_LOCKED, &dev->flags);
3384 					s.locked++;
3385 				} else {
3386 					/* let's read it back */
3387 					set_bit(R5_Wantread, &dev->flags);
3388 					set_bit(R5_LOCKED, &dev->flags);
3389 					s.locked++;
3390 				}
3391 			}
3392 		}
3393 
3394 	/* Finish reconstruct operations initiated by the expansion process */
3395 	if (sh->reconstruct_state == reconstruct_state_result) {
3396 		sh->reconstruct_state = reconstruct_state_idle;
3397 		clear_bit(STRIPE_EXPANDING, &sh->state);
3398 		for (i = conf->raid_disks; i--; ) {
3399 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
3400 			set_bit(R5_LOCKED, &sh->dev[i].flags);
3401 			s.locked++;
3402 		}
3403 	}
3404 
3405 	if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
3406 	    !sh->reconstruct_state) {
3407 		struct stripe_head *sh2
3408 			= get_active_stripe(conf, sh->sector, 1, 1, 1);
3409 		if (sh2 && test_bit(STRIPE_EXPAND_SOURCE, &sh2->state)) {
3410 			/* sh cannot be written until sh2 has been read.
3411 			 * so arrange for sh to be delayed a little
3412 			 */
3413 			set_bit(STRIPE_DELAYED, &sh->state);
3414 			set_bit(STRIPE_HANDLE, &sh->state);
3415 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE,
3416 					      &sh2->state))
3417 				atomic_inc(&conf->preread_active_stripes);
3418 			release_stripe(sh2);
3419 			goto unlock;
3420 		}
3421 		if (sh2)
3422 			release_stripe(sh2);
3423 
3424 		/* Need to write out all blocks after computing P&Q */
3425 		sh->disks = conf->raid_disks;
3426 		stripe_set_idx(sh->sector, conf, 0, sh);
3427 		schedule_reconstruction(sh, &s, 1, 1);
3428 	} else if (s.expanded && !sh->reconstruct_state && s.locked == 0) {
3429 		clear_bit(STRIPE_EXPAND_READY, &sh->state);
3430 		atomic_dec(&conf->reshape_stripes);
3431 		wake_up(&conf->wait_for_overlap);
3432 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3433 	}
3434 
3435 	if (s.expanding && s.locked == 0 &&
3436 	    !test_bit(STRIPE_COMPUTE_RUN, &sh->state))
3437 		handle_stripe_expansion(conf, sh, &r6s);
3438 
3439  unlock:
3440 	spin_unlock(&sh->lock);
3441 
3442 	/* wait for this device to become unblocked */
3443 	if (unlikely(blocked_rdev))
3444 		md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3445 
3446 	if (s.ops_request)
3447 		raid_run_ops(sh, s.ops_request);
3448 
3449 	ops_run_io(sh, &s);
3450 
3451 	return_io(return_bi);
3452 
3453 	return blocked_rdev == NULL;
3454 }
3455 
3456 /* returns true if the stripe was handled */
3457 static bool handle_stripe(struct stripe_head *sh)
3458 {
3459 	if (sh->raid_conf->level == 6)
3460 		return handle_stripe6(sh);
3461 	else
3462 		return handle_stripe5(sh);
3463 }
3464 
3465 static void raid5_activate_delayed(raid5_conf_t *conf)
3466 {
3467 	if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3468 		while (!list_empty(&conf->delayed_list)) {
3469 			struct list_head *l = conf->delayed_list.next;
3470 			struct stripe_head *sh;
3471 			sh = list_entry(l, struct stripe_head, lru);
3472 			list_del_init(l);
3473 			clear_bit(STRIPE_DELAYED, &sh->state);
3474 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3475 				atomic_inc(&conf->preread_active_stripes);
3476 			list_add_tail(&sh->lru, &conf->hold_list);
3477 		}
3478 	} else
3479 		blk_plug_device(conf->mddev->queue);
3480 }
3481 
3482 static void activate_bit_delay(raid5_conf_t *conf)
3483 {
3484 	/* device_lock is held */
3485 	struct list_head head;
3486 	list_add(&head, &conf->bitmap_list);
3487 	list_del_init(&conf->bitmap_list);
3488 	while (!list_empty(&head)) {
3489 		struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3490 		list_del_init(&sh->lru);
3491 		atomic_inc(&sh->count);
3492 		__release_stripe(conf, sh);
3493 	}
3494 }
3495 
3496 static void unplug_slaves(mddev_t *mddev)
3497 {
3498 	raid5_conf_t *conf = mddev->private;
3499 	int i;
3500 
3501 	rcu_read_lock();
3502 	for (i = 0; i < conf->raid_disks; i++) {
3503 		mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3504 		if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3505 			struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3506 
3507 			atomic_inc(&rdev->nr_pending);
3508 			rcu_read_unlock();
3509 
3510 			blk_unplug(r_queue);
3511 
3512 			rdev_dec_pending(rdev, mddev);
3513 			rcu_read_lock();
3514 		}
3515 	}
3516 	rcu_read_unlock();
3517 }
3518 
3519 static void raid5_unplug_device(struct request_queue *q)
3520 {
3521 	mddev_t *mddev = q->queuedata;
3522 	raid5_conf_t *conf = mddev->private;
3523 	unsigned long flags;
3524 
3525 	spin_lock_irqsave(&conf->device_lock, flags);
3526 
3527 	if (blk_remove_plug(q)) {
3528 		conf->seq_flush++;
3529 		raid5_activate_delayed(conf);
3530 	}
3531 	md_wakeup_thread(mddev->thread);
3532 
3533 	spin_unlock_irqrestore(&conf->device_lock, flags);
3534 
3535 	unplug_slaves(mddev);
3536 }
3537 
3538 static int raid5_congested(void *data, int bits)
3539 {
3540 	mddev_t *mddev = data;
3541 	raid5_conf_t *conf = mddev->private;
3542 
3543 	/* No difference between reads and writes.  Just check
3544 	 * how busy the stripe_cache is
3545 	 */
3546 	if (conf->inactive_blocked)
3547 		return 1;
3548 	if (conf->quiesce)
3549 		return 1;
3550 	if (list_empty_careful(&conf->inactive_list))
3551 		return 1;
3552 
3553 	return 0;
3554 }
3555 
3556 /* We want read requests to align with chunks where possible,
3557  * but write requests don't need to.
3558  */
3559 static int raid5_mergeable_bvec(struct request_queue *q,
3560 				struct bvec_merge_data *bvm,
3561 				struct bio_vec *biovec)
3562 {
3563 	mddev_t *mddev = q->queuedata;
3564 	sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
3565 	int max;
3566 	unsigned int chunk_sectors = mddev->chunk_sectors;
3567 	unsigned int bio_sectors = bvm->bi_size >> 9;
3568 
3569 	if ((bvm->bi_rw & 1) == WRITE)
3570 		return biovec->bv_len; /* always allow writes to be mergeable */
3571 
3572 	if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3573 		chunk_sectors = mddev->new_chunk_sectors;
3574 	max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3575 	if (max < 0) max = 0;
3576 	if (max <= biovec->bv_len && bio_sectors == 0)
3577 		return biovec->bv_len;
3578 	else
3579 		return max;
3580 }
3581 
3582 
3583 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3584 {
3585 	sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3586 	unsigned int chunk_sectors = mddev->chunk_sectors;
3587 	unsigned int bio_sectors = bio->bi_size >> 9;
3588 
3589 	if (mddev->new_chunk_sectors < mddev->chunk_sectors)
3590 		chunk_sectors = mddev->new_chunk_sectors;
3591 	return  chunk_sectors >=
3592 		((sector & (chunk_sectors - 1)) + bio_sectors);
3593 }
3594 
3595 /*
3596  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3597  *  later sampled by raid5d.
3598  */
3599 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3600 {
3601 	unsigned long flags;
3602 
3603 	spin_lock_irqsave(&conf->device_lock, flags);
3604 
3605 	bi->bi_next = conf->retry_read_aligned_list;
3606 	conf->retry_read_aligned_list = bi;
3607 
3608 	spin_unlock_irqrestore(&conf->device_lock, flags);
3609 	md_wakeup_thread(conf->mddev->thread);
3610 }
3611 
3612 
3613 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3614 {
3615 	struct bio *bi;
3616 
3617 	bi = conf->retry_read_aligned;
3618 	if (bi) {
3619 		conf->retry_read_aligned = NULL;
3620 		return bi;
3621 	}
3622 	bi = conf->retry_read_aligned_list;
3623 	if(bi) {
3624 		conf->retry_read_aligned_list = bi->bi_next;
3625 		bi->bi_next = NULL;
3626 		/*
3627 		 * this sets the active strip count to 1 and the processed
3628 		 * strip count to zero (upper 8 bits)
3629 		 */
3630 		bi->bi_phys_segments = 1; /* biased count of active stripes */
3631 	}
3632 
3633 	return bi;
3634 }
3635 
3636 
3637 /*
3638  *  The "raid5_align_endio" should check if the read succeeded and if it
3639  *  did, call bio_endio on the original bio (having bio_put the new bio
3640  *  first).
3641  *  If the read failed..
3642  */
3643 static void raid5_align_endio(struct bio *bi, int error)
3644 {
3645 	struct bio* raid_bi  = bi->bi_private;
3646 	mddev_t *mddev;
3647 	raid5_conf_t *conf;
3648 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3649 	mdk_rdev_t *rdev;
3650 
3651 	bio_put(bi);
3652 
3653 	mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3654 	conf = mddev->private;
3655 	rdev = (void*)raid_bi->bi_next;
3656 	raid_bi->bi_next = NULL;
3657 
3658 	rdev_dec_pending(rdev, conf->mddev);
3659 
3660 	if (!error && uptodate) {
3661 		bio_endio(raid_bi, 0);
3662 		if (atomic_dec_and_test(&conf->active_aligned_reads))
3663 			wake_up(&conf->wait_for_stripe);
3664 		return;
3665 	}
3666 
3667 
3668 	pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3669 
3670 	add_bio_to_retry(raid_bi, conf);
3671 }
3672 
3673 static int bio_fits_rdev(struct bio *bi)
3674 {
3675 	struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3676 
3677 	if ((bi->bi_size>>9) > queue_max_sectors(q))
3678 		return 0;
3679 	blk_recount_segments(q, bi);
3680 	if (bi->bi_phys_segments > queue_max_phys_segments(q))
3681 		return 0;
3682 
3683 	if (q->merge_bvec_fn)
3684 		/* it's too hard to apply the merge_bvec_fn at this stage,
3685 		 * just just give up
3686 		 */
3687 		return 0;
3688 
3689 	return 1;
3690 }
3691 
3692 
3693 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3694 {
3695 	mddev_t *mddev = q->queuedata;
3696 	raid5_conf_t *conf = mddev->private;
3697 	unsigned int dd_idx;
3698 	struct bio* align_bi;
3699 	mdk_rdev_t *rdev;
3700 
3701 	if (!in_chunk_boundary(mddev, raid_bio)) {
3702 		pr_debug("chunk_aligned_read : non aligned\n");
3703 		return 0;
3704 	}
3705 	/*
3706 	 * use bio_clone to make a copy of the bio
3707 	 */
3708 	align_bi = bio_clone(raid_bio, GFP_NOIO);
3709 	if (!align_bi)
3710 		return 0;
3711 	/*
3712 	 *   set bi_end_io to a new function, and set bi_private to the
3713 	 *     original bio.
3714 	 */
3715 	align_bi->bi_end_io  = raid5_align_endio;
3716 	align_bi->bi_private = raid_bio;
3717 	/*
3718 	 *	compute position
3719 	 */
3720 	align_bi->bi_sector =  raid5_compute_sector(conf, raid_bio->bi_sector,
3721 						    0,
3722 						    &dd_idx, NULL);
3723 
3724 	rcu_read_lock();
3725 	rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3726 	if (rdev && test_bit(In_sync, &rdev->flags)) {
3727 		atomic_inc(&rdev->nr_pending);
3728 		rcu_read_unlock();
3729 		raid_bio->bi_next = (void*)rdev;
3730 		align_bi->bi_bdev =  rdev->bdev;
3731 		align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3732 		align_bi->bi_sector += rdev->data_offset;
3733 
3734 		if (!bio_fits_rdev(align_bi)) {
3735 			/* too big in some way */
3736 			bio_put(align_bi);
3737 			rdev_dec_pending(rdev, mddev);
3738 			return 0;
3739 		}
3740 
3741 		spin_lock_irq(&conf->device_lock);
3742 		wait_event_lock_irq(conf->wait_for_stripe,
3743 				    conf->quiesce == 0,
3744 				    conf->device_lock, /* nothing */);
3745 		atomic_inc(&conf->active_aligned_reads);
3746 		spin_unlock_irq(&conf->device_lock);
3747 
3748 		generic_make_request(align_bi);
3749 		return 1;
3750 	} else {
3751 		rcu_read_unlock();
3752 		bio_put(align_bi);
3753 		return 0;
3754 	}
3755 }
3756 
3757 /* __get_priority_stripe - get the next stripe to process
3758  *
3759  * Full stripe writes are allowed to pass preread active stripes up until
3760  * the bypass_threshold is exceeded.  In general the bypass_count
3761  * increments when the handle_list is handled before the hold_list; however, it
3762  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3763  * stripe with in flight i/o.  The bypass_count will be reset when the
3764  * head of the hold_list has changed, i.e. the head was promoted to the
3765  * handle_list.
3766  */
3767 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3768 {
3769 	struct stripe_head *sh;
3770 
3771 	pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3772 		  __func__,
3773 		  list_empty(&conf->handle_list) ? "empty" : "busy",
3774 		  list_empty(&conf->hold_list) ? "empty" : "busy",
3775 		  atomic_read(&conf->pending_full_writes), conf->bypass_count);
3776 
3777 	if (!list_empty(&conf->handle_list)) {
3778 		sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3779 
3780 		if (list_empty(&conf->hold_list))
3781 			conf->bypass_count = 0;
3782 		else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3783 			if (conf->hold_list.next == conf->last_hold)
3784 				conf->bypass_count++;
3785 			else {
3786 				conf->last_hold = conf->hold_list.next;
3787 				conf->bypass_count -= conf->bypass_threshold;
3788 				if (conf->bypass_count < 0)
3789 					conf->bypass_count = 0;
3790 			}
3791 		}
3792 	} else if (!list_empty(&conf->hold_list) &&
3793 		   ((conf->bypass_threshold &&
3794 		     conf->bypass_count > conf->bypass_threshold) ||
3795 		    atomic_read(&conf->pending_full_writes) == 0)) {
3796 		sh = list_entry(conf->hold_list.next,
3797 				typeof(*sh), lru);
3798 		conf->bypass_count -= conf->bypass_threshold;
3799 		if (conf->bypass_count < 0)
3800 			conf->bypass_count = 0;
3801 	} else
3802 		return NULL;
3803 
3804 	list_del_init(&sh->lru);
3805 	atomic_inc(&sh->count);
3806 	BUG_ON(atomic_read(&sh->count) != 1);
3807 	return sh;
3808 }
3809 
3810 static int make_request(struct request_queue *q, struct bio * bi)
3811 {
3812 	mddev_t *mddev = q->queuedata;
3813 	raid5_conf_t *conf = mddev->private;
3814 	int dd_idx;
3815 	sector_t new_sector;
3816 	sector_t logical_sector, last_sector;
3817 	struct stripe_head *sh;
3818 	const int rw = bio_data_dir(bi);
3819 	int cpu, remaining;
3820 
3821 	if (unlikely(bio_barrier(bi))) {
3822 		bio_endio(bi, -EOPNOTSUPP);
3823 		return 0;
3824 	}
3825 
3826 	md_write_start(mddev, bi);
3827 
3828 	cpu = part_stat_lock();
3829 	part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
3830 	part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
3831 		      bio_sectors(bi));
3832 	part_stat_unlock();
3833 
3834 	if (rw == READ &&
3835 	     mddev->reshape_position == MaxSector &&
3836 	     chunk_aligned_read(q,bi))
3837 		return 0;
3838 
3839 	logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3840 	last_sector = bi->bi_sector + (bi->bi_size>>9);
3841 	bi->bi_next = NULL;
3842 	bi->bi_phys_segments = 1;	/* over-loaded to count active stripes */
3843 
3844 	for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3845 		DEFINE_WAIT(w);
3846 		int disks, data_disks;
3847 		int previous;
3848 
3849 	retry:
3850 		previous = 0;
3851 		disks = conf->raid_disks;
3852 		prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3853 		if (unlikely(conf->reshape_progress != MaxSector)) {
3854 			/* spinlock is needed as reshape_progress may be
3855 			 * 64bit on a 32bit platform, and so it might be
3856 			 * possible to see a half-updated value
3857 			 * Ofcourse reshape_progress could change after
3858 			 * the lock is dropped, so once we get a reference
3859 			 * to the stripe that we think it is, we will have
3860 			 * to check again.
3861 			 */
3862 			spin_lock_irq(&conf->device_lock);
3863 			if (mddev->delta_disks < 0
3864 			    ? logical_sector < conf->reshape_progress
3865 			    : logical_sector >= conf->reshape_progress) {
3866 				disks = conf->previous_raid_disks;
3867 				previous = 1;
3868 			} else {
3869 				if (mddev->delta_disks < 0
3870 				    ? logical_sector < conf->reshape_safe
3871 				    : logical_sector >= conf->reshape_safe) {
3872 					spin_unlock_irq(&conf->device_lock);
3873 					schedule();
3874 					goto retry;
3875 				}
3876 			}
3877 			spin_unlock_irq(&conf->device_lock);
3878 		}
3879 		data_disks = disks - conf->max_degraded;
3880 
3881 		new_sector = raid5_compute_sector(conf, logical_sector,
3882 						  previous,
3883 						  &dd_idx, NULL);
3884 		pr_debug("raid5: make_request, sector %llu logical %llu\n",
3885 			(unsigned long long)new_sector,
3886 			(unsigned long long)logical_sector);
3887 
3888 		sh = get_active_stripe(conf, new_sector, previous,
3889 				       (bi->bi_rw&RWA_MASK), 0);
3890 		if (sh) {
3891 			if (unlikely(previous)) {
3892 				/* expansion might have moved on while waiting for a
3893 				 * stripe, so we must do the range check again.
3894 				 * Expansion could still move past after this
3895 				 * test, but as we are holding a reference to
3896 				 * 'sh', we know that if that happens,
3897 				 *  STRIPE_EXPANDING will get set and the expansion
3898 				 * won't proceed until we finish with the stripe.
3899 				 */
3900 				int must_retry = 0;
3901 				spin_lock_irq(&conf->device_lock);
3902 				if (mddev->delta_disks < 0
3903 				    ? logical_sector >= conf->reshape_progress
3904 				    : logical_sector < conf->reshape_progress)
3905 					/* mismatch, need to try again */
3906 					must_retry = 1;
3907 				spin_unlock_irq(&conf->device_lock);
3908 				if (must_retry) {
3909 					release_stripe(sh);
3910 					schedule();
3911 					goto retry;
3912 				}
3913 			}
3914 
3915 			if (bio_data_dir(bi) == WRITE &&
3916 			    logical_sector >= mddev->suspend_lo &&
3917 			    logical_sector < mddev->suspend_hi) {
3918 				release_stripe(sh);
3919 				/* As the suspend_* range is controlled by
3920 				 * userspace, we want an interruptible
3921 				 * wait.
3922 				 */
3923 				flush_signals(current);
3924 				prepare_to_wait(&conf->wait_for_overlap,
3925 						&w, TASK_INTERRUPTIBLE);
3926 				if (logical_sector >= mddev->suspend_lo &&
3927 				    logical_sector < mddev->suspend_hi)
3928 					schedule();
3929 				goto retry;
3930 			}
3931 
3932 			if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3933 			    !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3934 				/* Stripe is busy expanding or
3935 				 * add failed due to overlap.  Flush everything
3936 				 * and wait a while
3937 				 */
3938 				raid5_unplug_device(mddev->queue);
3939 				release_stripe(sh);
3940 				schedule();
3941 				goto retry;
3942 			}
3943 			finish_wait(&conf->wait_for_overlap, &w);
3944 			set_bit(STRIPE_HANDLE, &sh->state);
3945 			clear_bit(STRIPE_DELAYED, &sh->state);
3946 			release_stripe(sh);
3947 		} else {
3948 			/* cannot get stripe for read-ahead, just give-up */
3949 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
3950 			finish_wait(&conf->wait_for_overlap, &w);
3951 			break;
3952 		}
3953 
3954 	}
3955 	spin_lock_irq(&conf->device_lock);
3956 	remaining = raid5_dec_bi_phys_segments(bi);
3957 	spin_unlock_irq(&conf->device_lock);
3958 	if (remaining == 0) {
3959 
3960 		if ( rw == WRITE )
3961 			md_write_end(mddev);
3962 
3963 		bio_endio(bi, 0);
3964 	}
3965 	return 0;
3966 }
3967 
3968 static sector_t raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks);
3969 
3970 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3971 {
3972 	/* reshaping is quite different to recovery/resync so it is
3973 	 * handled quite separately ... here.
3974 	 *
3975 	 * On each call to sync_request, we gather one chunk worth of
3976 	 * destination stripes and flag them as expanding.
3977 	 * Then we find all the source stripes and request reads.
3978 	 * As the reads complete, handle_stripe will copy the data
3979 	 * into the destination stripe and release that stripe.
3980 	 */
3981 	raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3982 	struct stripe_head *sh;
3983 	sector_t first_sector, last_sector;
3984 	int raid_disks = conf->previous_raid_disks;
3985 	int data_disks = raid_disks - conf->max_degraded;
3986 	int new_data_disks = conf->raid_disks - conf->max_degraded;
3987 	int i;
3988 	int dd_idx;
3989 	sector_t writepos, readpos, safepos;
3990 	sector_t stripe_addr;
3991 	int reshape_sectors;
3992 	struct list_head stripes;
3993 
3994 	if (sector_nr == 0) {
3995 		/* If restarting in the middle, skip the initial sectors */
3996 		if (mddev->delta_disks < 0 &&
3997 		    conf->reshape_progress < raid5_size(mddev, 0, 0)) {
3998 			sector_nr = raid5_size(mddev, 0, 0)
3999 				- conf->reshape_progress;
4000 		} else if (mddev->delta_disks >= 0 &&
4001 			   conf->reshape_progress > 0)
4002 			sector_nr = conf->reshape_progress;
4003 		sector_div(sector_nr, new_data_disks);
4004 		if (sector_nr) {
4005 			*skipped = 1;
4006 			return sector_nr;
4007 		}
4008 	}
4009 
4010 	/* We need to process a full chunk at a time.
4011 	 * If old and new chunk sizes differ, we need to process the
4012 	 * largest of these
4013 	 */
4014 	if (mddev->new_chunk_sectors > mddev->chunk_sectors)
4015 		reshape_sectors = mddev->new_chunk_sectors;
4016 	else
4017 		reshape_sectors = mddev->chunk_sectors;
4018 
4019 	/* we update the metadata when there is more than 3Meg
4020 	 * in the block range (that is rather arbitrary, should
4021 	 * probably be time based) or when the data about to be
4022 	 * copied would over-write the source of the data at
4023 	 * the front of the range.
4024 	 * i.e. one new_stripe along from reshape_progress new_maps
4025 	 * to after where reshape_safe old_maps to
4026 	 */
4027 	writepos = conf->reshape_progress;
4028 	sector_div(writepos, new_data_disks);
4029 	readpos = conf->reshape_progress;
4030 	sector_div(readpos, data_disks);
4031 	safepos = conf->reshape_safe;
4032 	sector_div(safepos, data_disks);
4033 	if (mddev->delta_disks < 0) {
4034 		writepos -= min_t(sector_t, reshape_sectors, writepos);
4035 		readpos += reshape_sectors;
4036 		safepos += reshape_sectors;
4037 	} else {
4038 		writepos += reshape_sectors;
4039 		readpos -= min_t(sector_t, reshape_sectors, readpos);
4040 		safepos -= min_t(sector_t, reshape_sectors, safepos);
4041 	}
4042 
4043 	/* 'writepos' is the most advanced device address we might write.
4044 	 * 'readpos' is the least advanced device address we might read.
4045 	 * 'safepos' is the least address recorded in the metadata as having
4046 	 *     been reshaped.
4047 	 * If 'readpos' is behind 'writepos', then there is no way that we can
4048 	 * ensure safety in the face of a crash - that must be done by userspace
4049 	 * making a backup of the data.  So in that case there is no particular
4050 	 * rush to update metadata.
4051 	 * Otherwise if 'safepos' is behind 'writepos', then we really need to
4052 	 * update the metadata to advance 'safepos' to match 'readpos' so that
4053 	 * we can be safe in the event of a crash.
4054 	 * So we insist on updating metadata if safepos is behind writepos and
4055 	 * readpos is beyond writepos.
4056 	 * In any case, update the metadata every 10 seconds.
4057 	 * Maybe that number should be configurable, but I'm not sure it is
4058 	 * worth it.... maybe it could be a multiple of safemode_delay???
4059 	 */
4060 	if ((mddev->delta_disks < 0
4061 	     ? (safepos > writepos && readpos < writepos)
4062 	     : (safepos < writepos && readpos > writepos)) ||
4063 	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4064 		/* Cannot proceed until we've updated the superblock... */
4065 		wait_event(conf->wait_for_overlap,
4066 			   atomic_read(&conf->reshape_stripes)==0);
4067 		mddev->reshape_position = conf->reshape_progress;
4068 		mddev->curr_resync_completed = mddev->curr_resync;
4069 		conf->reshape_checkpoint = jiffies;
4070 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
4071 		md_wakeup_thread(mddev->thread);
4072 		wait_event(mddev->sb_wait, mddev->flags == 0 ||
4073 			   kthread_should_stop());
4074 		spin_lock_irq(&conf->device_lock);
4075 		conf->reshape_safe = mddev->reshape_position;
4076 		spin_unlock_irq(&conf->device_lock);
4077 		wake_up(&conf->wait_for_overlap);
4078 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4079 	}
4080 
4081 	if (mddev->delta_disks < 0) {
4082 		BUG_ON(conf->reshape_progress == 0);
4083 		stripe_addr = writepos;
4084 		BUG_ON((mddev->dev_sectors &
4085 			~((sector_t)reshape_sectors - 1))
4086 		       - reshape_sectors - stripe_addr
4087 		       != sector_nr);
4088 	} else {
4089 		BUG_ON(writepos != sector_nr + reshape_sectors);
4090 		stripe_addr = sector_nr;
4091 	}
4092 	INIT_LIST_HEAD(&stripes);
4093 	for (i = 0; i < reshape_sectors; i += STRIPE_SECTORS) {
4094 		int j;
4095 		int skipped = 0;
4096 		sh = get_active_stripe(conf, stripe_addr+i, 0, 0, 1);
4097 		set_bit(STRIPE_EXPANDING, &sh->state);
4098 		atomic_inc(&conf->reshape_stripes);
4099 		/* If any of this stripe is beyond the end of the old
4100 		 * array, then we need to zero those blocks
4101 		 */
4102 		for (j=sh->disks; j--;) {
4103 			sector_t s;
4104 			if (j == sh->pd_idx)
4105 				continue;
4106 			if (conf->level == 6 &&
4107 			    j == sh->qd_idx)
4108 				continue;
4109 			s = compute_blocknr(sh, j, 0);
4110 			if (s < raid5_size(mddev, 0, 0)) {
4111 				skipped = 1;
4112 				continue;
4113 			}
4114 			memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
4115 			set_bit(R5_Expanded, &sh->dev[j].flags);
4116 			set_bit(R5_UPTODATE, &sh->dev[j].flags);
4117 		}
4118 		if (!skipped) {
4119 			set_bit(STRIPE_EXPAND_READY, &sh->state);
4120 			set_bit(STRIPE_HANDLE, &sh->state);
4121 		}
4122 		list_add(&sh->lru, &stripes);
4123 	}
4124 	spin_lock_irq(&conf->device_lock);
4125 	if (mddev->delta_disks < 0)
4126 		conf->reshape_progress -= reshape_sectors * new_data_disks;
4127 	else
4128 		conf->reshape_progress += reshape_sectors * new_data_disks;
4129 	spin_unlock_irq(&conf->device_lock);
4130 	/* Ok, those stripe are ready. We can start scheduling
4131 	 * reads on the source stripes.
4132 	 * The source stripes are determined by mapping the first and last
4133 	 * block on the destination stripes.
4134 	 */
4135 	first_sector =
4136 		raid5_compute_sector(conf, stripe_addr*(new_data_disks),
4137 				     1, &dd_idx, NULL);
4138 	last_sector =
4139 		raid5_compute_sector(conf, ((stripe_addr+reshape_sectors)
4140 					    * new_data_disks - 1),
4141 				     1, &dd_idx, NULL);
4142 	if (last_sector >= mddev->dev_sectors)
4143 		last_sector = mddev->dev_sectors - 1;
4144 	while (first_sector <= last_sector) {
4145 		sh = get_active_stripe(conf, first_sector, 1, 0, 1);
4146 		set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
4147 		set_bit(STRIPE_HANDLE, &sh->state);
4148 		release_stripe(sh);
4149 		first_sector += STRIPE_SECTORS;
4150 	}
4151 	/* Now that the sources are clearly marked, we can release
4152 	 * the destination stripes
4153 	 */
4154 	while (!list_empty(&stripes)) {
4155 		sh = list_entry(stripes.next, struct stripe_head, lru);
4156 		list_del_init(&sh->lru);
4157 		release_stripe(sh);
4158 	}
4159 	/* If this takes us to the resync_max point where we have to pause,
4160 	 * then we need to write out the superblock.
4161 	 */
4162 	sector_nr += reshape_sectors;
4163 	if ((sector_nr - mddev->curr_resync_completed) * 2
4164 	    >= mddev->resync_max - mddev->curr_resync_completed) {
4165 		/* Cannot proceed until we've updated the superblock... */
4166 		wait_event(conf->wait_for_overlap,
4167 			   atomic_read(&conf->reshape_stripes) == 0);
4168 		mddev->reshape_position = conf->reshape_progress;
4169 		mddev->curr_resync_completed = mddev->curr_resync + reshape_sectors;
4170 		conf->reshape_checkpoint = jiffies;
4171 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
4172 		md_wakeup_thread(mddev->thread);
4173 		wait_event(mddev->sb_wait,
4174 			   !test_bit(MD_CHANGE_DEVS, &mddev->flags)
4175 			   || kthread_should_stop());
4176 		spin_lock_irq(&conf->device_lock);
4177 		conf->reshape_safe = mddev->reshape_position;
4178 		spin_unlock_irq(&conf->device_lock);
4179 		wake_up(&conf->wait_for_overlap);
4180 		sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4181 	}
4182 	return reshape_sectors;
4183 }
4184 
4185 /* FIXME go_faster isn't used */
4186 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
4187 {
4188 	raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4189 	struct stripe_head *sh;
4190 	sector_t max_sector = mddev->dev_sectors;
4191 	int sync_blocks;
4192 	int still_degraded = 0;
4193 	int i;
4194 
4195 	if (sector_nr >= max_sector) {
4196 		/* just being told to finish up .. nothing much to do */
4197 		unplug_slaves(mddev);
4198 
4199 		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
4200 			end_reshape(conf);
4201 			return 0;
4202 		}
4203 
4204 		if (mddev->curr_resync < max_sector) /* aborted */
4205 			bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
4206 					&sync_blocks, 1);
4207 		else /* completed sync */
4208 			conf->fullsync = 0;
4209 		bitmap_close_sync(mddev->bitmap);
4210 
4211 		return 0;
4212 	}
4213 
4214 	/* Allow raid5_quiesce to complete */
4215 	wait_event(conf->wait_for_overlap, conf->quiesce != 2);
4216 
4217 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
4218 		return reshape_request(mddev, sector_nr, skipped);
4219 
4220 	/* No need to check resync_max as we never do more than one
4221 	 * stripe, and as resync_max will always be on a chunk boundary,
4222 	 * if the check in md_do_sync didn't fire, there is no chance
4223 	 * of overstepping resync_max here
4224 	 */
4225 
4226 	/* if there is too many failed drives and we are trying
4227 	 * to resync, then assert that we are finished, because there is
4228 	 * nothing we can do.
4229 	 */
4230 	if (mddev->degraded >= conf->max_degraded &&
4231 	    test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
4232 		sector_t rv = mddev->dev_sectors - sector_nr;
4233 		*skipped = 1;
4234 		return rv;
4235 	}
4236 	if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
4237 	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
4238 	    !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
4239 		/* we can skip this block, and probably more */
4240 		sync_blocks /= STRIPE_SECTORS;
4241 		*skipped = 1;
4242 		return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
4243 	}
4244 
4245 
4246 	bitmap_cond_end_sync(mddev->bitmap, sector_nr);
4247 
4248 	sh = get_active_stripe(conf, sector_nr, 0, 1, 0);
4249 	if (sh == NULL) {
4250 		sh = get_active_stripe(conf, sector_nr, 0, 0, 0);
4251 		/* make sure we don't swamp the stripe cache if someone else
4252 		 * is trying to get access
4253 		 */
4254 		schedule_timeout_uninterruptible(1);
4255 	}
4256 	/* Need to check if array will still be degraded after recovery/resync
4257 	 * We don't need to check the 'failed' flag as when that gets set,
4258 	 * recovery aborts.
4259 	 */
4260 	for (i = 0; i < conf->raid_disks; i++)
4261 		if (conf->disks[i].rdev == NULL)
4262 			still_degraded = 1;
4263 
4264 	bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
4265 
4266 	spin_lock(&sh->lock);
4267 	set_bit(STRIPE_SYNCING, &sh->state);
4268 	clear_bit(STRIPE_INSYNC, &sh->state);
4269 	spin_unlock(&sh->lock);
4270 
4271 	/* wait for any blocked device to be handled */
4272 	while (unlikely(!handle_stripe(sh)))
4273 		;
4274 	release_stripe(sh);
4275 
4276 	return STRIPE_SECTORS;
4277 }
4278 
4279 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
4280 {
4281 	/* We may not be able to submit a whole bio at once as there
4282 	 * may not be enough stripe_heads available.
4283 	 * We cannot pre-allocate enough stripe_heads as we may need
4284 	 * more than exist in the cache (if we allow ever large chunks).
4285 	 * So we do one stripe head at a time and record in
4286 	 * ->bi_hw_segments how many have been done.
4287 	 *
4288 	 * We *know* that this entire raid_bio is in one chunk, so
4289 	 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
4290 	 */
4291 	struct stripe_head *sh;
4292 	int dd_idx;
4293 	sector_t sector, logical_sector, last_sector;
4294 	int scnt = 0;
4295 	int remaining;
4296 	int handled = 0;
4297 
4298 	logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
4299 	sector = raid5_compute_sector(conf, logical_sector,
4300 				      0, &dd_idx, NULL);
4301 	last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
4302 
4303 	for (; logical_sector < last_sector;
4304 	     logical_sector += STRIPE_SECTORS,
4305 		     sector += STRIPE_SECTORS,
4306 		     scnt++) {
4307 
4308 		if (scnt < raid5_bi_hw_segments(raid_bio))
4309 			/* already done this stripe */
4310 			continue;
4311 
4312 		sh = get_active_stripe(conf, sector, 0, 1, 0);
4313 
4314 		if (!sh) {
4315 			/* failed to get a stripe - must wait */
4316 			raid5_set_bi_hw_segments(raid_bio, scnt);
4317 			conf->retry_read_aligned = raid_bio;
4318 			return handled;
4319 		}
4320 
4321 		set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
4322 		if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
4323 			release_stripe(sh);
4324 			raid5_set_bi_hw_segments(raid_bio, scnt);
4325 			conf->retry_read_aligned = raid_bio;
4326 			return handled;
4327 		}
4328 
4329 		handle_stripe(sh);
4330 		release_stripe(sh);
4331 		handled++;
4332 	}
4333 	spin_lock_irq(&conf->device_lock);
4334 	remaining = raid5_dec_bi_phys_segments(raid_bio);
4335 	spin_unlock_irq(&conf->device_lock);
4336 	if (remaining == 0)
4337 		bio_endio(raid_bio, 0);
4338 	if (atomic_dec_and_test(&conf->active_aligned_reads))
4339 		wake_up(&conf->wait_for_stripe);
4340 	return handled;
4341 }
4342 
4343 #ifdef CONFIG_MULTICORE_RAID456
4344 static void __process_stripe(void *param, async_cookie_t cookie)
4345 {
4346 	struct stripe_head *sh = param;
4347 
4348 	handle_stripe(sh);
4349 	release_stripe(sh);
4350 }
4351 
4352 static void process_stripe(struct stripe_head *sh, struct list_head *domain)
4353 {
4354 	async_schedule_domain(__process_stripe, sh, domain);
4355 }
4356 
4357 static void synchronize_stripe_processing(struct list_head *domain)
4358 {
4359 	async_synchronize_full_domain(domain);
4360 }
4361 #else
4362 static void process_stripe(struct stripe_head *sh, struct list_head *domain)
4363 {
4364 	handle_stripe(sh);
4365 	release_stripe(sh);
4366 	cond_resched();
4367 }
4368 
4369 static void synchronize_stripe_processing(struct list_head *domain)
4370 {
4371 }
4372 #endif
4373 
4374 
4375 /*
4376  * This is our raid5 kernel thread.
4377  *
4378  * We scan the hash table for stripes which can be handled now.
4379  * During the scan, completed stripes are saved for us by the interrupt
4380  * handler, so that they will not have to wait for our next wakeup.
4381  */
4382 static void raid5d(mddev_t *mddev)
4383 {
4384 	struct stripe_head *sh;
4385 	raid5_conf_t *conf = mddev->private;
4386 	int handled;
4387 	LIST_HEAD(raid_domain);
4388 
4389 	pr_debug("+++ raid5d active\n");
4390 
4391 	md_check_recovery(mddev);
4392 
4393 	handled = 0;
4394 	spin_lock_irq(&conf->device_lock);
4395 	while (1) {
4396 		struct bio *bio;
4397 
4398 		if (conf->seq_flush != conf->seq_write) {
4399 			int seq = conf->seq_flush;
4400 			spin_unlock_irq(&conf->device_lock);
4401 			bitmap_unplug(mddev->bitmap);
4402 			spin_lock_irq(&conf->device_lock);
4403 			conf->seq_write = seq;
4404 			activate_bit_delay(conf);
4405 		}
4406 
4407 		while ((bio = remove_bio_from_retry(conf))) {
4408 			int ok;
4409 			spin_unlock_irq(&conf->device_lock);
4410 			ok = retry_aligned_read(conf, bio);
4411 			spin_lock_irq(&conf->device_lock);
4412 			if (!ok)
4413 				break;
4414 			handled++;
4415 		}
4416 
4417 		sh = __get_priority_stripe(conf);
4418 
4419 		if (!sh)
4420 			break;
4421 		spin_unlock_irq(&conf->device_lock);
4422 
4423 		handled++;
4424 		process_stripe(sh, &raid_domain);
4425 
4426 		spin_lock_irq(&conf->device_lock);
4427 	}
4428 	pr_debug("%d stripes handled\n", handled);
4429 
4430 	spin_unlock_irq(&conf->device_lock);
4431 
4432 	synchronize_stripe_processing(&raid_domain);
4433 	async_tx_issue_pending_all();
4434 	unplug_slaves(mddev);
4435 
4436 	pr_debug("--- raid5d inactive\n");
4437 }
4438 
4439 static ssize_t
4440 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4441 {
4442 	raid5_conf_t *conf = mddev->private;
4443 	if (conf)
4444 		return sprintf(page, "%d\n", conf->max_nr_stripes);
4445 	else
4446 		return 0;
4447 }
4448 
4449 static ssize_t
4450 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4451 {
4452 	raid5_conf_t *conf = mddev->private;
4453 	unsigned long new;
4454 	int err;
4455 
4456 	if (len >= PAGE_SIZE)
4457 		return -EINVAL;
4458 	if (!conf)
4459 		return -ENODEV;
4460 
4461 	if (strict_strtoul(page, 10, &new))
4462 		return -EINVAL;
4463 	if (new <= 16 || new > 32768)
4464 		return -EINVAL;
4465 	while (new < conf->max_nr_stripes) {
4466 		if (drop_one_stripe(conf))
4467 			conf->max_nr_stripes--;
4468 		else
4469 			break;
4470 	}
4471 	err = md_allow_write(mddev);
4472 	if (err)
4473 		return err;
4474 	while (new > conf->max_nr_stripes) {
4475 		if (grow_one_stripe(conf))
4476 			conf->max_nr_stripes++;
4477 		else break;
4478 	}
4479 	return len;
4480 }
4481 
4482 static struct md_sysfs_entry
4483 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4484 				raid5_show_stripe_cache_size,
4485 				raid5_store_stripe_cache_size);
4486 
4487 static ssize_t
4488 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4489 {
4490 	raid5_conf_t *conf = mddev->private;
4491 	if (conf)
4492 		return sprintf(page, "%d\n", conf->bypass_threshold);
4493 	else
4494 		return 0;
4495 }
4496 
4497 static ssize_t
4498 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4499 {
4500 	raid5_conf_t *conf = mddev->private;
4501 	unsigned long new;
4502 	if (len >= PAGE_SIZE)
4503 		return -EINVAL;
4504 	if (!conf)
4505 		return -ENODEV;
4506 
4507 	if (strict_strtoul(page, 10, &new))
4508 		return -EINVAL;
4509 	if (new > conf->max_nr_stripes)
4510 		return -EINVAL;
4511 	conf->bypass_threshold = new;
4512 	return len;
4513 }
4514 
4515 static struct md_sysfs_entry
4516 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4517 					S_IRUGO | S_IWUSR,
4518 					raid5_show_preread_threshold,
4519 					raid5_store_preread_threshold);
4520 
4521 static ssize_t
4522 stripe_cache_active_show(mddev_t *mddev, char *page)
4523 {
4524 	raid5_conf_t *conf = mddev->private;
4525 	if (conf)
4526 		return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4527 	else
4528 		return 0;
4529 }
4530 
4531 static struct md_sysfs_entry
4532 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4533 
4534 static struct attribute *raid5_attrs[] =  {
4535 	&raid5_stripecache_size.attr,
4536 	&raid5_stripecache_active.attr,
4537 	&raid5_preread_bypass_threshold.attr,
4538 	NULL,
4539 };
4540 static struct attribute_group raid5_attrs_group = {
4541 	.name = NULL,
4542 	.attrs = raid5_attrs,
4543 };
4544 
4545 static sector_t
4546 raid5_size(mddev_t *mddev, sector_t sectors, int raid_disks)
4547 {
4548 	raid5_conf_t *conf = mddev->private;
4549 
4550 	if (!sectors)
4551 		sectors = mddev->dev_sectors;
4552 	if (!raid_disks) {
4553 		/* size is defined by the smallest of previous and new size */
4554 		if (conf->raid_disks < conf->previous_raid_disks)
4555 			raid_disks = conf->raid_disks;
4556 		else
4557 			raid_disks = conf->previous_raid_disks;
4558 	}
4559 
4560 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
4561 	sectors &= ~((sector_t)mddev->new_chunk_sectors - 1);
4562 	return sectors * (raid_disks - conf->max_degraded);
4563 }
4564 
4565 static void raid5_free_percpu(raid5_conf_t *conf)
4566 {
4567 	struct raid5_percpu *percpu;
4568 	unsigned long cpu;
4569 
4570 	if (!conf->percpu)
4571 		return;
4572 
4573 	get_online_cpus();
4574 	for_each_possible_cpu(cpu) {
4575 		percpu = per_cpu_ptr(conf->percpu, cpu);
4576 		safe_put_page(percpu->spare_page);
4577 		kfree(percpu->scribble);
4578 	}
4579 #ifdef CONFIG_HOTPLUG_CPU
4580 	unregister_cpu_notifier(&conf->cpu_notify);
4581 #endif
4582 	put_online_cpus();
4583 
4584 	free_percpu(conf->percpu);
4585 }
4586 
4587 static void free_conf(raid5_conf_t *conf)
4588 {
4589 	shrink_stripes(conf);
4590 	raid5_free_percpu(conf);
4591 	kfree(conf->disks);
4592 	kfree(conf->stripe_hashtbl);
4593 	kfree(conf);
4594 }
4595 
4596 #ifdef CONFIG_HOTPLUG_CPU
4597 static int raid456_cpu_notify(struct notifier_block *nfb, unsigned long action,
4598 			      void *hcpu)
4599 {
4600 	raid5_conf_t *conf = container_of(nfb, raid5_conf_t, cpu_notify);
4601 	long cpu = (long)hcpu;
4602 	struct raid5_percpu *percpu = per_cpu_ptr(conf->percpu, cpu);
4603 
4604 	switch (action) {
4605 	case CPU_UP_PREPARE:
4606 	case CPU_UP_PREPARE_FROZEN:
4607 		if (conf->level == 6 && !percpu->spare_page)
4608 			percpu->spare_page = alloc_page(GFP_KERNEL);
4609 		if (!percpu->scribble)
4610 			percpu->scribble = kmalloc(conf->scribble_len, GFP_KERNEL);
4611 
4612 		if (!percpu->scribble ||
4613 		    (conf->level == 6 && !percpu->spare_page)) {
4614 			safe_put_page(percpu->spare_page);
4615 			kfree(percpu->scribble);
4616 			pr_err("%s: failed memory allocation for cpu%ld\n",
4617 			       __func__, cpu);
4618 			return NOTIFY_BAD;
4619 		}
4620 		break;
4621 	case CPU_DEAD:
4622 	case CPU_DEAD_FROZEN:
4623 		safe_put_page(percpu->spare_page);
4624 		kfree(percpu->scribble);
4625 		percpu->spare_page = NULL;
4626 		percpu->scribble = NULL;
4627 		break;
4628 	default:
4629 		break;
4630 	}
4631 	return NOTIFY_OK;
4632 }
4633 #endif
4634 
4635 static int raid5_alloc_percpu(raid5_conf_t *conf)
4636 {
4637 	unsigned long cpu;
4638 	struct page *spare_page;
4639 	struct raid5_percpu *allcpus;
4640 	void *scribble;
4641 	int err;
4642 
4643 	allcpus = alloc_percpu(struct raid5_percpu);
4644 	if (!allcpus)
4645 		return -ENOMEM;
4646 	conf->percpu = allcpus;
4647 
4648 	get_online_cpus();
4649 	err = 0;
4650 	for_each_present_cpu(cpu) {
4651 		if (conf->level == 6) {
4652 			spare_page = alloc_page(GFP_KERNEL);
4653 			if (!spare_page) {
4654 				err = -ENOMEM;
4655 				break;
4656 			}
4657 			per_cpu_ptr(conf->percpu, cpu)->spare_page = spare_page;
4658 		}
4659 		scribble = kmalloc(scribble_len(conf->raid_disks), GFP_KERNEL);
4660 		if (!scribble) {
4661 			err = -ENOMEM;
4662 			break;
4663 		}
4664 		per_cpu_ptr(conf->percpu, cpu)->scribble = scribble;
4665 	}
4666 #ifdef CONFIG_HOTPLUG_CPU
4667 	conf->cpu_notify.notifier_call = raid456_cpu_notify;
4668 	conf->cpu_notify.priority = 0;
4669 	if (err == 0)
4670 		err = register_cpu_notifier(&conf->cpu_notify);
4671 #endif
4672 	put_online_cpus();
4673 
4674 	return err;
4675 }
4676 
4677 static raid5_conf_t *setup_conf(mddev_t *mddev)
4678 {
4679 	raid5_conf_t *conf;
4680 	int raid_disk, memory;
4681 	mdk_rdev_t *rdev;
4682 	struct disk_info *disk;
4683 
4684 	if (mddev->new_level != 5
4685 	    && mddev->new_level != 4
4686 	    && mddev->new_level != 6) {
4687 		printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4688 		       mdname(mddev), mddev->new_level);
4689 		return ERR_PTR(-EIO);
4690 	}
4691 	if ((mddev->new_level == 5
4692 	     && !algorithm_valid_raid5(mddev->new_layout)) ||
4693 	    (mddev->new_level == 6
4694 	     && !algorithm_valid_raid6(mddev->new_layout))) {
4695 		printk(KERN_ERR "raid5: %s: layout %d not supported\n",
4696 		       mdname(mddev), mddev->new_layout);
4697 		return ERR_PTR(-EIO);
4698 	}
4699 	if (mddev->new_level == 6 && mddev->raid_disks < 4) {
4700 		printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4701 		       mdname(mddev), mddev->raid_disks);
4702 		return ERR_PTR(-EINVAL);
4703 	}
4704 
4705 	if (!mddev->new_chunk_sectors ||
4706 	    (mddev->new_chunk_sectors << 9) % PAGE_SIZE ||
4707 	    !is_power_of_2(mddev->new_chunk_sectors)) {
4708 		printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4709 		       mddev->new_chunk_sectors << 9, mdname(mddev));
4710 		return ERR_PTR(-EINVAL);
4711 	}
4712 
4713 	conf = kzalloc(sizeof(raid5_conf_t), GFP_KERNEL);
4714 	if (conf == NULL)
4715 		goto abort;
4716 
4717 	conf->raid_disks = mddev->raid_disks;
4718 	conf->scribble_len = scribble_len(conf->raid_disks);
4719 	if (mddev->reshape_position == MaxSector)
4720 		conf->previous_raid_disks = mddev->raid_disks;
4721 	else
4722 		conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4723 
4724 	conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4725 			      GFP_KERNEL);
4726 	if (!conf->disks)
4727 		goto abort;
4728 
4729 	conf->mddev = mddev;
4730 
4731 	if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4732 		goto abort;
4733 
4734 	conf->level = mddev->new_level;
4735 	if (raid5_alloc_percpu(conf) != 0)
4736 		goto abort;
4737 
4738 	spin_lock_init(&conf->device_lock);
4739 	init_waitqueue_head(&conf->wait_for_stripe);
4740 	init_waitqueue_head(&conf->wait_for_overlap);
4741 	INIT_LIST_HEAD(&conf->handle_list);
4742 	INIT_LIST_HEAD(&conf->hold_list);
4743 	INIT_LIST_HEAD(&conf->delayed_list);
4744 	INIT_LIST_HEAD(&conf->bitmap_list);
4745 	INIT_LIST_HEAD(&conf->inactive_list);
4746 	atomic_set(&conf->active_stripes, 0);
4747 	atomic_set(&conf->preread_active_stripes, 0);
4748 	atomic_set(&conf->active_aligned_reads, 0);
4749 	conf->bypass_threshold = BYPASS_THRESHOLD;
4750 
4751 	pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4752 
4753 	list_for_each_entry(rdev, &mddev->disks, same_set) {
4754 		raid_disk = rdev->raid_disk;
4755 		if (raid_disk >= conf->raid_disks
4756 		    || raid_disk < 0)
4757 			continue;
4758 		disk = conf->disks + raid_disk;
4759 
4760 		disk->rdev = rdev;
4761 
4762 		if (test_bit(In_sync, &rdev->flags)) {
4763 			char b[BDEVNAME_SIZE];
4764 			printk(KERN_INFO "raid5: device %s operational as raid"
4765 				" disk %d\n", bdevname(rdev->bdev,b),
4766 				raid_disk);
4767 		} else
4768 			/* Cannot rely on bitmap to complete recovery */
4769 			conf->fullsync = 1;
4770 	}
4771 
4772 	conf->chunk_sectors = mddev->new_chunk_sectors;
4773 	conf->level = mddev->new_level;
4774 	if (conf->level == 6)
4775 		conf->max_degraded = 2;
4776 	else
4777 		conf->max_degraded = 1;
4778 	conf->algorithm = mddev->new_layout;
4779 	conf->max_nr_stripes = NR_STRIPES;
4780 	conf->reshape_progress = mddev->reshape_position;
4781 	if (conf->reshape_progress != MaxSector) {
4782 		conf->prev_chunk_sectors = mddev->chunk_sectors;
4783 		conf->prev_algo = mddev->layout;
4784 	}
4785 
4786 	memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4787 		 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4788 	if (grow_stripes(conf, conf->max_nr_stripes)) {
4789 		printk(KERN_ERR
4790 			"raid5: couldn't allocate %dkB for buffers\n", memory);
4791 		goto abort;
4792 	} else
4793 		printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4794 			memory, mdname(mddev));
4795 
4796 	conf->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4797 	if (!conf->thread) {
4798 		printk(KERN_ERR
4799 		       "raid5: couldn't allocate thread for %s\n",
4800 		       mdname(mddev));
4801 		goto abort;
4802 	}
4803 
4804 	return conf;
4805 
4806  abort:
4807 	if (conf) {
4808 		free_conf(conf);
4809 		return ERR_PTR(-EIO);
4810 	} else
4811 		return ERR_PTR(-ENOMEM);
4812 }
4813 
4814 static int run(mddev_t *mddev)
4815 {
4816 	raid5_conf_t *conf;
4817 	int working_disks = 0, chunk_size;
4818 	mdk_rdev_t *rdev;
4819 
4820 	if (mddev->recovery_cp != MaxSector)
4821 		printk(KERN_NOTICE "raid5: %s is not clean"
4822 		       " -- starting background reconstruction\n",
4823 		       mdname(mddev));
4824 	if (mddev->reshape_position != MaxSector) {
4825 		/* Check that we can continue the reshape.
4826 		 * Currently only disks can change, it must
4827 		 * increase, and we must be past the point where
4828 		 * a stripe over-writes itself
4829 		 */
4830 		sector_t here_new, here_old;
4831 		int old_disks;
4832 		int max_degraded = (mddev->level == 6 ? 2 : 1);
4833 
4834 		if (mddev->new_level != mddev->level) {
4835 			printk(KERN_ERR "raid5: %s: unsupported reshape "
4836 			       "required - aborting.\n",
4837 			       mdname(mddev));
4838 			return -EINVAL;
4839 		}
4840 		old_disks = mddev->raid_disks - mddev->delta_disks;
4841 		/* reshape_position must be on a new-stripe boundary, and one
4842 		 * further up in new geometry must map after here in old
4843 		 * geometry.
4844 		 */
4845 		here_new = mddev->reshape_position;
4846 		if (sector_div(here_new, mddev->new_chunk_sectors *
4847 			       (mddev->raid_disks - max_degraded))) {
4848 			printk(KERN_ERR "raid5: reshape_position not "
4849 			       "on a stripe boundary\n");
4850 			return -EINVAL;
4851 		}
4852 		/* here_new is the stripe we will write to */
4853 		here_old = mddev->reshape_position;
4854 		sector_div(here_old, mddev->chunk_sectors *
4855 			   (old_disks-max_degraded));
4856 		/* here_old is the first stripe that we might need to read
4857 		 * from */
4858 		if (mddev->delta_disks == 0) {
4859 			/* We cannot be sure it is safe to start an in-place
4860 			 * reshape.  It is only safe if user-space if monitoring
4861 			 * and taking constant backups.
4862 			 * mdadm always starts a situation like this in
4863 			 * readonly mode so it can take control before
4864 			 * allowing any writes.  So just check for that.
4865 			 */
4866 			if ((here_new * mddev->new_chunk_sectors !=
4867 			     here_old * mddev->chunk_sectors) ||
4868 			    mddev->ro == 0) {
4869 				printk(KERN_ERR "raid5: in-place reshape must be started"
4870 				       " in read-only mode - aborting\n");
4871 				return -EINVAL;
4872 			}
4873 		} else if (mddev->delta_disks < 0
4874 		    ? (here_new * mddev->new_chunk_sectors <=
4875 		       here_old * mddev->chunk_sectors)
4876 		    : (here_new * mddev->new_chunk_sectors >=
4877 		       here_old * mddev->chunk_sectors)) {
4878 			/* Reading from the same stripe as writing to - bad */
4879 			printk(KERN_ERR "raid5: reshape_position too early for "
4880 			       "auto-recovery - aborting.\n");
4881 			return -EINVAL;
4882 		}
4883 		printk(KERN_INFO "raid5: reshape will continue\n");
4884 		/* OK, we should be able to continue; */
4885 	} else {
4886 		BUG_ON(mddev->level != mddev->new_level);
4887 		BUG_ON(mddev->layout != mddev->new_layout);
4888 		BUG_ON(mddev->chunk_sectors != mddev->new_chunk_sectors);
4889 		BUG_ON(mddev->delta_disks != 0);
4890 	}
4891 
4892 	if (mddev->private == NULL)
4893 		conf = setup_conf(mddev);
4894 	else
4895 		conf = mddev->private;
4896 
4897 	if (IS_ERR(conf))
4898 		return PTR_ERR(conf);
4899 
4900 	mddev->thread = conf->thread;
4901 	conf->thread = NULL;
4902 	mddev->private = conf;
4903 
4904 	/*
4905 	 * 0 for a fully functional array, 1 or 2 for a degraded array.
4906 	 */
4907 	list_for_each_entry(rdev, &mddev->disks, same_set)
4908 		if (rdev->raid_disk >= 0 &&
4909 		    test_bit(In_sync, &rdev->flags))
4910 			working_disks++;
4911 
4912 	mddev->degraded = conf->raid_disks - working_disks;
4913 
4914 	if (mddev->degraded > conf->max_degraded) {
4915 		printk(KERN_ERR "raid5: not enough operational devices for %s"
4916 			" (%d/%d failed)\n",
4917 			mdname(mddev), mddev->degraded, conf->raid_disks);
4918 		goto abort;
4919 	}
4920 
4921 	/* device size must be a multiple of chunk size */
4922 	mddev->dev_sectors &= ~(mddev->chunk_sectors - 1);
4923 	mddev->resync_max_sectors = mddev->dev_sectors;
4924 
4925 	if (mddev->degraded > 0 &&
4926 	    mddev->recovery_cp != MaxSector) {
4927 		if (mddev->ok_start_degraded)
4928 			printk(KERN_WARNING
4929 			       "raid5: starting dirty degraded array: %s"
4930 			       "- data corruption possible.\n",
4931 			       mdname(mddev));
4932 		else {
4933 			printk(KERN_ERR
4934 			       "raid5: cannot start dirty degraded array for %s\n",
4935 			       mdname(mddev));
4936 			goto abort;
4937 		}
4938 	}
4939 
4940 	if (mddev->degraded == 0)
4941 		printk("raid5: raid level %d set %s active with %d out of %d"
4942 		       " devices, algorithm %d\n", conf->level, mdname(mddev),
4943 		       mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4944 		       mddev->new_layout);
4945 	else
4946 		printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4947 			" out of %d devices, algorithm %d\n", conf->level,
4948 			mdname(mddev), mddev->raid_disks - mddev->degraded,
4949 			mddev->raid_disks, mddev->new_layout);
4950 
4951 	print_raid5_conf(conf);
4952 
4953 	if (conf->reshape_progress != MaxSector) {
4954 		printk("...ok start reshape thread\n");
4955 		conf->reshape_safe = conf->reshape_progress;
4956 		atomic_set(&conf->reshape_stripes, 0);
4957 		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4958 		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4959 		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4960 		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4961 		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4962 							"%s_reshape");
4963 	}
4964 
4965 	/* read-ahead size must cover two whole stripes, which is
4966 	 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4967 	 */
4968 	{
4969 		int data_disks = conf->previous_raid_disks - conf->max_degraded;
4970 		int stripe = data_disks *
4971 			((mddev->chunk_sectors << 9) / PAGE_SIZE);
4972 		if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4973 			mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4974 	}
4975 
4976 	/* Ok, everything is just fine now */
4977 	if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4978 		printk(KERN_WARNING
4979 		       "raid5: failed to create sysfs attributes for %s\n",
4980 		       mdname(mddev));
4981 
4982 	mddev->queue->queue_lock = &conf->device_lock;
4983 
4984 	mddev->queue->unplug_fn = raid5_unplug_device;
4985 	mddev->queue->backing_dev_info.congested_data = mddev;
4986 	mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4987 
4988 	md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
4989 
4990 	blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4991 	chunk_size = mddev->chunk_sectors << 9;
4992 	blk_queue_io_min(mddev->queue, chunk_size);
4993 	blk_queue_io_opt(mddev->queue, chunk_size *
4994 			 (conf->raid_disks - conf->max_degraded));
4995 
4996 	list_for_each_entry(rdev, &mddev->disks, same_set)
4997 		disk_stack_limits(mddev->gendisk, rdev->bdev,
4998 				  rdev->data_offset << 9);
4999 
5000 	return 0;
5001 abort:
5002 	md_unregister_thread(mddev->thread);
5003 	mddev->thread = NULL;
5004 	if (conf) {
5005 		print_raid5_conf(conf);
5006 		free_conf(conf);
5007 	}
5008 	mddev->private = NULL;
5009 	printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
5010 	return -EIO;
5011 }
5012 
5013 
5014 
5015 static int stop(mddev_t *mddev)
5016 {
5017 	raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5018 
5019 	md_unregister_thread(mddev->thread);
5020 	mddev->thread = NULL;
5021 	mddev->queue->backing_dev_info.congested_fn = NULL;
5022 	blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
5023 	sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
5024 	free_conf(conf);
5025 	mddev->private = NULL;
5026 	return 0;
5027 }
5028 
5029 #ifdef DEBUG
5030 static void print_sh(struct seq_file *seq, struct stripe_head *sh)
5031 {
5032 	int i;
5033 
5034 	seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
5035 		   (unsigned long long)sh->sector, sh->pd_idx, sh->state);
5036 	seq_printf(seq, "sh %llu,  count %d.\n",
5037 		   (unsigned long long)sh->sector, atomic_read(&sh->count));
5038 	seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
5039 	for (i = 0; i < sh->disks; i++) {
5040 		seq_printf(seq, "(cache%d: %p %ld) ",
5041 			   i, sh->dev[i].page, sh->dev[i].flags);
5042 	}
5043 	seq_printf(seq, "\n");
5044 }
5045 
5046 static void printall(struct seq_file *seq, raid5_conf_t *conf)
5047 {
5048 	struct stripe_head *sh;
5049 	struct hlist_node *hn;
5050 	int i;
5051 
5052 	spin_lock_irq(&conf->device_lock);
5053 	for (i = 0; i < NR_HASH; i++) {
5054 		hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
5055 			if (sh->raid_conf != conf)
5056 				continue;
5057 			print_sh(seq, sh);
5058 		}
5059 	}
5060 	spin_unlock_irq(&conf->device_lock);
5061 }
5062 #endif
5063 
5064 static void status(struct seq_file *seq, mddev_t *mddev)
5065 {
5066 	raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
5067 	int i;
5068 
5069 	seq_printf(seq, " level %d, %dk chunk, algorithm %d", mddev->level,
5070 		mddev->chunk_sectors / 2, mddev->layout);
5071 	seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
5072 	for (i = 0; i < conf->raid_disks; i++)
5073 		seq_printf (seq, "%s",
5074 			       conf->disks[i].rdev &&
5075 			       test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
5076 	seq_printf (seq, "]");
5077 #ifdef DEBUG
5078 	seq_printf (seq, "\n");
5079 	printall(seq, conf);
5080 #endif
5081 }
5082 
5083 static void print_raid5_conf (raid5_conf_t *conf)
5084 {
5085 	int i;
5086 	struct disk_info *tmp;
5087 
5088 	printk("RAID5 conf printout:\n");
5089 	if (!conf) {
5090 		printk("(conf==NULL)\n");
5091 		return;
5092 	}
5093 	printk(" --- rd:%d wd:%d\n", conf->raid_disks,
5094 		 conf->raid_disks - conf->mddev->degraded);
5095 
5096 	for (i = 0; i < conf->raid_disks; i++) {
5097 		char b[BDEVNAME_SIZE];
5098 		tmp = conf->disks + i;
5099 		if (tmp->rdev)
5100 		printk(" disk %d, o:%d, dev:%s\n",
5101 			i, !test_bit(Faulty, &tmp->rdev->flags),
5102 			bdevname(tmp->rdev->bdev,b));
5103 	}
5104 }
5105 
5106 static int raid5_spare_active(mddev_t *mddev)
5107 {
5108 	int i;
5109 	raid5_conf_t *conf = mddev->private;
5110 	struct disk_info *tmp;
5111 
5112 	for (i = 0; i < conf->raid_disks; i++) {
5113 		tmp = conf->disks + i;
5114 		if (tmp->rdev
5115 		    && !test_bit(Faulty, &tmp->rdev->flags)
5116 		    && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
5117 			unsigned long flags;
5118 			spin_lock_irqsave(&conf->device_lock, flags);
5119 			mddev->degraded--;
5120 			spin_unlock_irqrestore(&conf->device_lock, flags);
5121 		}
5122 	}
5123 	print_raid5_conf(conf);
5124 	return 0;
5125 }
5126 
5127 static int raid5_remove_disk(mddev_t *mddev, int number)
5128 {
5129 	raid5_conf_t *conf = mddev->private;
5130 	int err = 0;
5131 	mdk_rdev_t *rdev;
5132 	struct disk_info *p = conf->disks + number;
5133 
5134 	print_raid5_conf(conf);
5135 	rdev = p->rdev;
5136 	if (rdev) {
5137 		if (number >= conf->raid_disks &&
5138 		    conf->reshape_progress == MaxSector)
5139 			clear_bit(In_sync, &rdev->flags);
5140 
5141 		if (test_bit(In_sync, &rdev->flags) ||
5142 		    atomic_read(&rdev->nr_pending)) {
5143 			err = -EBUSY;
5144 			goto abort;
5145 		}
5146 		/* Only remove non-faulty devices if recovery
5147 		 * isn't possible.
5148 		 */
5149 		if (!test_bit(Faulty, &rdev->flags) &&
5150 		    mddev->degraded <= conf->max_degraded &&
5151 		    number < conf->raid_disks) {
5152 			err = -EBUSY;
5153 			goto abort;
5154 		}
5155 		p->rdev = NULL;
5156 		synchronize_rcu();
5157 		if (atomic_read(&rdev->nr_pending)) {
5158 			/* lost the race, try later */
5159 			err = -EBUSY;
5160 			p->rdev = rdev;
5161 		}
5162 	}
5163 abort:
5164 
5165 	print_raid5_conf(conf);
5166 	return err;
5167 }
5168 
5169 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
5170 {
5171 	raid5_conf_t *conf = mddev->private;
5172 	int err = -EEXIST;
5173 	int disk;
5174 	struct disk_info *p;
5175 	int first = 0;
5176 	int last = conf->raid_disks - 1;
5177 
5178 	if (mddev->degraded > conf->max_degraded)
5179 		/* no point adding a device */
5180 		return -EINVAL;
5181 
5182 	if (rdev->raid_disk >= 0)
5183 		first = last = rdev->raid_disk;
5184 
5185 	/*
5186 	 * find the disk ... but prefer rdev->saved_raid_disk
5187 	 * if possible.
5188 	 */
5189 	if (rdev->saved_raid_disk >= 0 &&
5190 	    rdev->saved_raid_disk >= first &&
5191 	    conf->disks[rdev->saved_raid_disk].rdev == NULL)
5192 		disk = rdev->saved_raid_disk;
5193 	else
5194 		disk = first;
5195 	for ( ; disk <= last ; disk++)
5196 		if ((p=conf->disks + disk)->rdev == NULL) {
5197 			clear_bit(In_sync, &rdev->flags);
5198 			rdev->raid_disk = disk;
5199 			err = 0;
5200 			if (rdev->saved_raid_disk != disk)
5201 				conf->fullsync = 1;
5202 			rcu_assign_pointer(p->rdev, rdev);
5203 			break;
5204 		}
5205 	print_raid5_conf(conf);
5206 	return err;
5207 }
5208 
5209 static int raid5_resize(mddev_t *mddev, sector_t sectors)
5210 {
5211 	/* no resync is happening, and there is enough space
5212 	 * on all devices, so we can resize.
5213 	 * We need to make sure resync covers any new space.
5214 	 * If the array is shrinking we should possibly wait until
5215 	 * any io in the removed space completes, but it hardly seems
5216 	 * worth it.
5217 	 */
5218 	sectors &= ~((sector_t)mddev->chunk_sectors - 1);
5219 	md_set_array_sectors(mddev, raid5_size(mddev, sectors,
5220 					       mddev->raid_disks));
5221 	if (mddev->array_sectors >
5222 	    raid5_size(mddev, sectors, mddev->raid_disks))
5223 		return -EINVAL;
5224 	set_capacity(mddev->gendisk, mddev->array_sectors);
5225 	mddev->changed = 1;
5226 	revalidate_disk(mddev->gendisk);
5227 	if (sectors > mddev->dev_sectors && mddev->recovery_cp == MaxSector) {
5228 		mddev->recovery_cp = mddev->dev_sectors;
5229 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5230 	}
5231 	mddev->dev_sectors = sectors;
5232 	mddev->resync_max_sectors = sectors;
5233 	return 0;
5234 }
5235 
5236 static int check_stripe_cache(mddev_t *mddev)
5237 {
5238 	/* Can only proceed if there are plenty of stripe_heads.
5239 	 * We need a minimum of one full stripe,, and for sensible progress
5240 	 * it is best to have about 4 times that.
5241 	 * If we require 4 times, then the default 256 4K stripe_heads will
5242 	 * allow for chunk sizes up to 256K, which is probably OK.
5243 	 * If the chunk size is greater, user-space should request more
5244 	 * stripe_heads first.
5245 	 */
5246 	raid5_conf_t *conf = mddev->private;
5247 	if (((mddev->chunk_sectors << 9) / STRIPE_SIZE) * 4
5248 	    > conf->max_nr_stripes ||
5249 	    ((mddev->new_chunk_sectors << 9) / STRIPE_SIZE) * 4
5250 	    > conf->max_nr_stripes) {
5251 		printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
5252 		       ((max(mddev->chunk_sectors, mddev->new_chunk_sectors) << 9)
5253 			/ STRIPE_SIZE)*4);
5254 		return 0;
5255 	}
5256 	return 1;
5257 }
5258 
5259 static int check_reshape(mddev_t *mddev)
5260 {
5261 	raid5_conf_t *conf = mddev->private;
5262 
5263 	if (mddev->delta_disks == 0 &&
5264 	    mddev->new_layout == mddev->layout &&
5265 	    mddev->new_chunk_sectors == mddev->chunk_sectors)
5266 		return 0; /* nothing to do */
5267 	if (mddev->bitmap)
5268 		/* Cannot grow a bitmap yet */
5269 		return -EBUSY;
5270 	if (mddev->degraded > conf->max_degraded)
5271 		return -EINVAL;
5272 	if (mddev->delta_disks < 0) {
5273 		/* We might be able to shrink, but the devices must
5274 		 * be made bigger first.
5275 		 * For raid6, 4 is the minimum size.
5276 		 * Otherwise 2 is the minimum
5277 		 */
5278 		int min = 2;
5279 		if (mddev->level == 6)
5280 			min = 4;
5281 		if (mddev->raid_disks + mddev->delta_disks < min)
5282 			return -EINVAL;
5283 	}
5284 
5285 	if (!check_stripe_cache(mddev))
5286 		return -ENOSPC;
5287 
5288 	return resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
5289 }
5290 
5291 static int raid5_start_reshape(mddev_t *mddev)
5292 {
5293 	raid5_conf_t *conf = mddev->private;
5294 	mdk_rdev_t *rdev;
5295 	int spares = 0;
5296 	int added_devices = 0;
5297 	unsigned long flags;
5298 
5299 	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
5300 		return -EBUSY;
5301 
5302 	if (!check_stripe_cache(mddev))
5303 		return -ENOSPC;
5304 
5305 	list_for_each_entry(rdev, &mddev->disks, same_set)
5306 		if (rdev->raid_disk < 0 &&
5307 		    !test_bit(Faulty, &rdev->flags))
5308 			spares++;
5309 
5310 	if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
5311 		/* Not enough devices even to make a degraded array
5312 		 * of that size
5313 		 */
5314 		return -EINVAL;
5315 
5316 	/* Refuse to reduce size of the array.  Any reductions in
5317 	 * array size must be through explicit setting of array_size
5318 	 * attribute.
5319 	 */
5320 	if (raid5_size(mddev, 0, conf->raid_disks + mddev->delta_disks)
5321 	    < mddev->array_sectors) {
5322 		printk(KERN_ERR "md: %s: array size must be reduced "
5323 		       "before number of disks\n", mdname(mddev));
5324 		return -EINVAL;
5325 	}
5326 
5327 	atomic_set(&conf->reshape_stripes, 0);
5328 	spin_lock_irq(&conf->device_lock);
5329 	conf->previous_raid_disks = conf->raid_disks;
5330 	conf->raid_disks += mddev->delta_disks;
5331 	conf->prev_chunk_sectors = conf->chunk_sectors;
5332 	conf->chunk_sectors = mddev->new_chunk_sectors;
5333 	conf->prev_algo = conf->algorithm;
5334 	conf->algorithm = mddev->new_layout;
5335 	if (mddev->delta_disks < 0)
5336 		conf->reshape_progress = raid5_size(mddev, 0, 0);
5337 	else
5338 		conf->reshape_progress = 0;
5339 	conf->reshape_safe = conf->reshape_progress;
5340 	conf->generation++;
5341 	spin_unlock_irq(&conf->device_lock);
5342 
5343 	/* Add some new drives, as many as will fit.
5344 	 * We know there are enough to make the newly sized array work.
5345 	 */
5346 	list_for_each_entry(rdev, &mddev->disks, same_set)
5347 		if (rdev->raid_disk < 0 &&
5348 		    !test_bit(Faulty, &rdev->flags)) {
5349 			if (raid5_add_disk(mddev, rdev) == 0) {
5350 				char nm[20];
5351 				set_bit(In_sync, &rdev->flags);
5352 				added_devices++;
5353 				rdev->recovery_offset = 0;
5354 				sprintf(nm, "rd%d", rdev->raid_disk);
5355 				if (sysfs_create_link(&mddev->kobj,
5356 						      &rdev->kobj, nm))
5357 					printk(KERN_WARNING
5358 					       "raid5: failed to create "
5359 					       " link %s for %s\n",
5360 					       nm, mdname(mddev));
5361 			} else
5362 				break;
5363 		}
5364 
5365 	if (mddev->delta_disks > 0) {
5366 		spin_lock_irqsave(&conf->device_lock, flags);
5367 		mddev->degraded = (conf->raid_disks - conf->previous_raid_disks)
5368 			- added_devices;
5369 		spin_unlock_irqrestore(&conf->device_lock, flags);
5370 	}
5371 	mddev->raid_disks = conf->raid_disks;
5372 	mddev->reshape_position = conf->reshape_progress;
5373 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
5374 
5375 	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
5376 	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
5377 	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
5378 	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
5379 	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
5380 						"%s_reshape");
5381 	if (!mddev->sync_thread) {
5382 		mddev->recovery = 0;
5383 		spin_lock_irq(&conf->device_lock);
5384 		mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
5385 		conf->reshape_progress = MaxSector;
5386 		spin_unlock_irq(&conf->device_lock);
5387 		return -EAGAIN;
5388 	}
5389 	conf->reshape_checkpoint = jiffies;
5390 	md_wakeup_thread(mddev->sync_thread);
5391 	md_new_event(mddev);
5392 	return 0;
5393 }
5394 
5395 /* This is called from the reshape thread and should make any
5396  * changes needed in 'conf'
5397  */
5398 static void end_reshape(raid5_conf_t *conf)
5399 {
5400 
5401 	if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
5402 
5403 		spin_lock_irq(&conf->device_lock);
5404 		conf->previous_raid_disks = conf->raid_disks;
5405 		conf->reshape_progress = MaxSector;
5406 		spin_unlock_irq(&conf->device_lock);
5407 		wake_up(&conf->wait_for_overlap);
5408 
5409 		/* read-ahead size must cover two whole stripes, which is
5410 		 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
5411 		 */
5412 		{
5413 			int data_disks = conf->raid_disks - conf->max_degraded;
5414 			int stripe = data_disks * ((conf->chunk_sectors << 9)
5415 						   / PAGE_SIZE);
5416 			if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
5417 				conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
5418 		}
5419 	}
5420 }
5421 
5422 /* This is called from the raid5d thread with mddev_lock held.
5423  * It makes config changes to the device.
5424  */
5425 static void raid5_finish_reshape(mddev_t *mddev)
5426 {
5427 	raid5_conf_t *conf = mddev->private;
5428 
5429 	if (!test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
5430 
5431 		if (mddev->delta_disks > 0) {
5432 			md_set_array_sectors(mddev, raid5_size(mddev, 0, 0));
5433 			set_capacity(mddev->gendisk, mddev->array_sectors);
5434 			mddev->changed = 1;
5435 			revalidate_disk(mddev->gendisk);
5436 		} else {
5437 			int d;
5438 			mddev->degraded = conf->raid_disks;
5439 			for (d = 0; d < conf->raid_disks ; d++)
5440 				if (conf->disks[d].rdev &&
5441 				    test_bit(In_sync,
5442 					     &conf->disks[d].rdev->flags))
5443 					mddev->degraded--;
5444 			for (d = conf->raid_disks ;
5445 			     d < conf->raid_disks - mddev->delta_disks;
5446 			     d++) {
5447 				mdk_rdev_t *rdev = conf->disks[d].rdev;
5448 				if (rdev && raid5_remove_disk(mddev, d) == 0) {
5449 					char nm[20];
5450 					sprintf(nm, "rd%d", rdev->raid_disk);
5451 					sysfs_remove_link(&mddev->kobj, nm);
5452 					rdev->raid_disk = -1;
5453 				}
5454 			}
5455 		}
5456 		mddev->layout = conf->algorithm;
5457 		mddev->chunk_sectors = conf->chunk_sectors;
5458 		mddev->reshape_position = MaxSector;
5459 		mddev->delta_disks = 0;
5460 	}
5461 }
5462 
5463 static void raid5_quiesce(mddev_t *mddev, int state)
5464 {
5465 	raid5_conf_t *conf = mddev->private;
5466 
5467 	switch(state) {
5468 	case 2: /* resume for a suspend */
5469 		wake_up(&conf->wait_for_overlap);
5470 		break;
5471 
5472 	case 1: /* stop all writes */
5473 		spin_lock_irq(&conf->device_lock);
5474 		/* '2' tells resync/reshape to pause so that all
5475 		 * active stripes can drain
5476 		 */
5477 		conf->quiesce = 2;
5478 		wait_event_lock_irq(conf->wait_for_stripe,
5479 				    atomic_read(&conf->active_stripes) == 0 &&
5480 				    atomic_read(&conf->active_aligned_reads) == 0,
5481 				    conf->device_lock, /* nothing */);
5482 		conf->quiesce = 1;
5483 		spin_unlock_irq(&conf->device_lock);
5484 		/* allow reshape to continue */
5485 		wake_up(&conf->wait_for_overlap);
5486 		break;
5487 
5488 	case 0: /* re-enable writes */
5489 		spin_lock_irq(&conf->device_lock);
5490 		conf->quiesce = 0;
5491 		wake_up(&conf->wait_for_stripe);
5492 		wake_up(&conf->wait_for_overlap);
5493 		spin_unlock_irq(&conf->device_lock);
5494 		break;
5495 	}
5496 }
5497 
5498 
5499 static void *raid5_takeover_raid1(mddev_t *mddev)
5500 {
5501 	int chunksect;
5502 
5503 	if (mddev->raid_disks != 2 ||
5504 	    mddev->degraded > 1)
5505 		return ERR_PTR(-EINVAL);
5506 
5507 	/* Should check if there are write-behind devices? */
5508 
5509 	chunksect = 64*2; /* 64K by default */
5510 
5511 	/* The array must be an exact multiple of chunksize */
5512 	while (chunksect && (mddev->array_sectors & (chunksect-1)))
5513 		chunksect >>= 1;
5514 
5515 	if ((chunksect<<9) < STRIPE_SIZE)
5516 		/* array size does not allow a suitable chunk size */
5517 		return ERR_PTR(-EINVAL);
5518 
5519 	mddev->new_level = 5;
5520 	mddev->new_layout = ALGORITHM_LEFT_SYMMETRIC;
5521 	mddev->new_chunk_sectors = chunksect;
5522 
5523 	return setup_conf(mddev);
5524 }
5525 
5526 static void *raid5_takeover_raid6(mddev_t *mddev)
5527 {
5528 	int new_layout;
5529 
5530 	switch (mddev->layout) {
5531 	case ALGORITHM_LEFT_ASYMMETRIC_6:
5532 		new_layout = ALGORITHM_LEFT_ASYMMETRIC;
5533 		break;
5534 	case ALGORITHM_RIGHT_ASYMMETRIC_6:
5535 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC;
5536 		break;
5537 	case ALGORITHM_LEFT_SYMMETRIC_6:
5538 		new_layout = ALGORITHM_LEFT_SYMMETRIC;
5539 		break;
5540 	case ALGORITHM_RIGHT_SYMMETRIC_6:
5541 		new_layout = ALGORITHM_RIGHT_SYMMETRIC;
5542 		break;
5543 	case ALGORITHM_PARITY_0_6:
5544 		new_layout = ALGORITHM_PARITY_0;
5545 		break;
5546 	case ALGORITHM_PARITY_N:
5547 		new_layout = ALGORITHM_PARITY_N;
5548 		break;
5549 	default:
5550 		return ERR_PTR(-EINVAL);
5551 	}
5552 	mddev->new_level = 5;
5553 	mddev->new_layout = new_layout;
5554 	mddev->delta_disks = -1;
5555 	mddev->raid_disks -= 1;
5556 	return setup_conf(mddev);
5557 }
5558 
5559 
5560 static int raid5_check_reshape(mddev_t *mddev)
5561 {
5562 	/* For a 2-drive array, the layout and chunk size can be changed
5563 	 * immediately as not restriping is needed.
5564 	 * For larger arrays we record the new value - after validation
5565 	 * to be used by a reshape pass.
5566 	 */
5567 	raid5_conf_t *conf = mddev->private;
5568 	int new_chunk = mddev->new_chunk_sectors;
5569 
5570 	if (mddev->new_layout >= 0 && !algorithm_valid_raid5(mddev->new_layout))
5571 		return -EINVAL;
5572 	if (new_chunk > 0) {
5573 		if (!is_power_of_2(new_chunk))
5574 			return -EINVAL;
5575 		if (new_chunk < (PAGE_SIZE>>9))
5576 			return -EINVAL;
5577 		if (mddev->array_sectors & (new_chunk-1))
5578 			/* not factor of array size */
5579 			return -EINVAL;
5580 	}
5581 
5582 	/* They look valid */
5583 
5584 	if (mddev->raid_disks == 2) {
5585 		/* can make the change immediately */
5586 		if (mddev->new_layout >= 0) {
5587 			conf->algorithm = mddev->new_layout;
5588 			mddev->layout = mddev->new_layout;
5589 		}
5590 		if (new_chunk > 0) {
5591 			conf->chunk_sectors = new_chunk ;
5592 			mddev->chunk_sectors = new_chunk;
5593 		}
5594 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
5595 		md_wakeup_thread(mddev->thread);
5596 	}
5597 	return check_reshape(mddev);
5598 }
5599 
5600 static int raid6_check_reshape(mddev_t *mddev)
5601 {
5602 	int new_chunk = mddev->new_chunk_sectors;
5603 
5604 	if (mddev->new_layout >= 0 && !algorithm_valid_raid6(mddev->new_layout))
5605 		return -EINVAL;
5606 	if (new_chunk > 0) {
5607 		if (!is_power_of_2(new_chunk))
5608 			return -EINVAL;
5609 		if (new_chunk < (PAGE_SIZE >> 9))
5610 			return -EINVAL;
5611 		if (mddev->array_sectors & (new_chunk-1))
5612 			/* not factor of array size */
5613 			return -EINVAL;
5614 	}
5615 
5616 	/* They look valid */
5617 	return check_reshape(mddev);
5618 }
5619 
5620 static void *raid5_takeover(mddev_t *mddev)
5621 {
5622 	/* raid5 can take over:
5623 	 *  raid0 - if all devices are the same - make it a raid4 layout
5624 	 *  raid1 - if there are two drives.  We need to know the chunk size
5625 	 *  raid4 - trivial - just use a raid4 layout.
5626 	 *  raid6 - Providing it is a *_6 layout
5627 	 */
5628 
5629 	if (mddev->level == 1)
5630 		return raid5_takeover_raid1(mddev);
5631 	if (mddev->level == 4) {
5632 		mddev->new_layout = ALGORITHM_PARITY_N;
5633 		mddev->new_level = 5;
5634 		return setup_conf(mddev);
5635 	}
5636 	if (mddev->level == 6)
5637 		return raid5_takeover_raid6(mddev);
5638 
5639 	return ERR_PTR(-EINVAL);
5640 }
5641 
5642 
5643 static struct mdk_personality raid5_personality;
5644 
5645 static void *raid6_takeover(mddev_t *mddev)
5646 {
5647 	/* Currently can only take over a raid5.  We map the
5648 	 * personality to an equivalent raid6 personality
5649 	 * with the Q block at the end.
5650 	 */
5651 	int new_layout;
5652 
5653 	if (mddev->pers != &raid5_personality)
5654 		return ERR_PTR(-EINVAL);
5655 	if (mddev->degraded > 1)
5656 		return ERR_PTR(-EINVAL);
5657 	if (mddev->raid_disks > 253)
5658 		return ERR_PTR(-EINVAL);
5659 	if (mddev->raid_disks < 3)
5660 		return ERR_PTR(-EINVAL);
5661 
5662 	switch (mddev->layout) {
5663 	case ALGORITHM_LEFT_ASYMMETRIC:
5664 		new_layout = ALGORITHM_LEFT_ASYMMETRIC_6;
5665 		break;
5666 	case ALGORITHM_RIGHT_ASYMMETRIC:
5667 		new_layout = ALGORITHM_RIGHT_ASYMMETRIC_6;
5668 		break;
5669 	case ALGORITHM_LEFT_SYMMETRIC:
5670 		new_layout = ALGORITHM_LEFT_SYMMETRIC_6;
5671 		break;
5672 	case ALGORITHM_RIGHT_SYMMETRIC:
5673 		new_layout = ALGORITHM_RIGHT_SYMMETRIC_6;
5674 		break;
5675 	case ALGORITHM_PARITY_0:
5676 		new_layout = ALGORITHM_PARITY_0_6;
5677 		break;
5678 	case ALGORITHM_PARITY_N:
5679 		new_layout = ALGORITHM_PARITY_N;
5680 		break;
5681 	default:
5682 		return ERR_PTR(-EINVAL);
5683 	}
5684 	mddev->new_level = 6;
5685 	mddev->new_layout = new_layout;
5686 	mddev->delta_disks = 1;
5687 	mddev->raid_disks += 1;
5688 	return setup_conf(mddev);
5689 }
5690 
5691 
5692 static struct mdk_personality raid6_personality =
5693 {
5694 	.name		= "raid6",
5695 	.level		= 6,
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	= raid6_check_reshape,
5709 	.start_reshape  = raid5_start_reshape,
5710 	.finish_reshape = raid5_finish_reshape,
5711 	.quiesce	= raid5_quiesce,
5712 	.takeover	= raid6_takeover,
5713 };
5714 static struct mdk_personality raid5_personality =
5715 {
5716 	.name		= "raid5",
5717 	.level		= 5,
5718 	.owner		= THIS_MODULE,
5719 	.make_request	= make_request,
5720 	.run		= run,
5721 	.stop		= stop,
5722 	.status		= status,
5723 	.error_handler	= error,
5724 	.hot_add_disk	= raid5_add_disk,
5725 	.hot_remove_disk= raid5_remove_disk,
5726 	.spare_active	= raid5_spare_active,
5727 	.sync_request	= sync_request,
5728 	.resize		= raid5_resize,
5729 	.size		= raid5_size,
5730 	.check_reshape	= raid5_check_reshape,
5731 	.start_reshape  = raid5_start_reshape,
5732 	.finish_reshape = raid5_finish_reshape,
5733 	.quiesce	= raid5_quiesce,
5734 	.takeover	= raid5_takeover,
5735 };
5736 
5737 static struct mdk_personality raid4_personality =
5738 {
5739 	.name		= "raid4",
5740 	.level		= 4,
5741 	.owner		= THIS_MODULE,
5742 	.make_request	= make_request,
5743 	.run		= run,
5744 	.stop		= stop,
5745 	.status		= status,
5746 	.error_handler	= error,
5747 	.hot_add_disk	= raid5_add_disk,
5748 	.hot_remove_disk= raid5_remove_disk,
5749 	.spare_active	= raid5_spare_active,
5750 	.sync_request	= sync_request,
5751 	.resize		= raid5_resize,
5752 	.size		= raid5_size,
5753 	.check_reshape	= raid5_check_reshape,
5754 	.start_reshape  = raid5_start_reshape,
5755 	.finish_reshape = raid5_finish_reshape,
5756 	.quiesce	= raid5_quiesce,
5757 };
5758 
5759 static int __init raid5_init(void)
5760 {
5761 	register_md_personality(&raid6_personality);
5762 	register_md_personality(&raid5_personality);
5763 	register_md_personality(&raid4_personality);
5764 	return 0;
5765 }
5766 
5767 static void raid5_exit(void)
5768 {
5769 	unregister_md_personality(&raid6_personality);
5770 	unregister_md_personality(&raid5_personality);
5771 	unregister_md_personality(&raid4_personality);
5772 }
5773 
5774 module_init(raid5_init);
5775 module_exit(raid5_exit);
5776 MODULE_LICENSE("GPL");
5777 MODULE_ALIAS("md-personality-4"); /* RAID5 */
5778 MODULE_ALIAS("md-raid5");
5779 MODULE_ALIAS("md-raid4");
5780 MODULE_ALIAS("md-level-5");
5781 MODULE_ALIAS("md-level-4");
5782 MODULE_ALIAS("md-personality-8"); /* RAID6 */
5783 MODULE_ALIAS("md-raid6");
5784 MODULE_ALIAS("md-level-6");
5785 
5786 /* This used to be two separate modules, they were: */
5787 MODULE_ALIAS("raid5");
5788 MODULE_ALIAS("raid6");
5789