xref: /linux/drivers/md/raid5.c (revision f7511d5f66f01fc451747b24e79f3ada7a3af9af)
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/module.h>
47 #include <linux/slab.h>
48 #include <linux/highmem.h>
49 #include <linux/bitops.h>
50 #include <linux/kthread.h>
51 #include <asm/atomic.h>
52 #include "raid6.h"
53 
54 #include <linux/raid/bitmap.h>
55 #include <linux/async_tx.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 #if !RAID6_USE_EMPTY_ZERO_PAGE
98 /* In .bss so it's zeroed */
99 const char raid6_empty_zero_page[PAGE_SIZE] __attribute__((aligned(256)));
100 #endif
101 
102 static inline int raid6_next_disk(int disk, int raid_disks)
103 {
104 	disk++;
105 	return (disk < raid_disks) ? disk : 0;
106 }
107 
108 static void return_io(struct bio *return_bi)
109 {
110 	struct bio *bi = return_bi;
111 	while (bi) {
112 
113 		return_bi = bi->bi_next;
114 		bi->bi_next = NULL;
115 		bi->bi_size = 0;
116 		bi->bi_end_io(bi,
117 			      test_bit(BIO_UPTODATE, &bi->bi_flags)
118 			        ? 0 : -EIO);
119 		bi = return_bi;
120 	}
121 }
122 
123 static void print_raid5_conf (raid5_conf_t *conf);
124 
125 static void __release_stripe(raid5_conf_t *conf, struct stripe_head *sh)
126 {
127 	if (atomic_dec_and_test(&sh->count)) {
128 		BUG_ON(!list_empty(&sh->lru));
129 		BUG_ON(atomic_read(&conf->active_stripes)==0);
130 		if (test_bit(STRIPE_HANDLE, &sh->state)) {
131 			if (test_bit(STRIPE_DELAYED, &sh->state)) {
132 				list_add_tail(&sh->lru, &conf->delayed_list);
133 				blk_plug_device(conf->mddev->queue);
134 			} else if (test_bit(STRIPE_BIT_DELAY, &sh->state) &&
135 				   sh->bm_seq - conf->seq_write > 0) {
136 				list_add_tail(&sh->lru, &conf->bitmap_list);
137 				blk_plug_device(conf->mddev->queue);
138 			} else {
139 				clear_bit(STRIPE_BIT_DELAY, &sh->state);
140 				list_add_tail(&sh->lru, &conf->handle_list);
141 			}
142 			md_wakeup_thread(conf->mddev->thread);
143 		} else {
144 			BUG_ON(sh->ops.pending);
145 			if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
146 				atomic_dec(&conf->preread_active_stripes);
147 				if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD)
148 					md_wakeup_thread(conf->mddev->thread);
149 			}
150 			atomic_dec(&conf->active_stripes);
151 			if (!test_bit(STRIPE_EXPANDING, &sh->state)) {
152 				list_add_tail(&sh->lru, &conf->inactive_list);
153 				wake_up(&conf->wait_for_stripe);
154 				if (conf->retry_read_aligned)
155 					md_wakeup_thread(conf->mddev->thread);
156 			}
157 		}
158 	}
159 }
160 static void release_stripe(struct stripe_head *sh)
161 {
162 	raid5_conf_t *conf = sh->raid_conf;
163 	unsigned long flags;
164 
165 	spin_lock_irqsave(&conf->device_lock, flags);
166 	__release_stripe(conf, sh);
167 	spin_unlock_irqrestore(&conf->device_lock, flags);
168 }
169 
170 static inline void remove_hash(struct stripe_head *sh)
171 {
172 	pr_debug("remove_hash(), stripe %llu\n",
173 		(unsigned long long)sh->sector);
174 
175 	hlist_del_init(&sh->hash);
176 }
177 
178 static inline void insert_hash(raid5_conf_t *conf, struct stripe_head *sh)
179 {
180 	struct hlist_head *hp = stripe_hash(conf, sh->sector);
181 
182 	pr_debug("insert_hash(), stripe %llu\n",
183 		(unsigned long long)sh->sector);
184 
185 	CHECK_DEVLOCK();
186 	hlist_add_head(&sh->hash, hp);
187 }
188 
189 
190 /* find an idle stripe, make sure it is unhashed, and return it. */
191 static struct stripe_head *get_free_stripe(raid5_conf_t *conf)
192 {
193 	struct stripe_head *sh = NULL;
194 	struct list_head *first;
195 
196 	CHECK_DEVLOCK();
197 	if (list_empty(&conf->inactive_list))
198 		goto out;
199 	first = conf->inactive_list.next;
200 	sh = list_entry(first, struct stripe_head, lru);
201 	list_del_init(first);
202 	remove_hash(sh);
203 	atomic_inc(&conf->active_stripes);
204 out:
205 	return sh;
206 }
207 
208 static void shrink_buffers(struct stripe_head *sh, int num)
209 {
210 	struct page *p;
211 	int i;
212 
213 	for (i=0; i<num ; i++) {
214 		p = sh->dev[i].page;
215 		if (!p)
216 			continue;
217 		sh->dev[i].page = NULL;
218 		put_page(p);
219 	}
220 }
221 
222 static int grow_buffers(struct stripe_head *sh, int num)
223 {
224 	int i;
225 
226 	for (i=0; i<num; i++) {
227 		struct page *page;
228 
229 		if (!(page = alloc_page(GFP_KERNEL))) {
230 			return 1;
231 		}
232 		sh->dev[i].page = page;
233 	}
234 	return 0;
235 }
236 
237 static void raid5_build_block (struct stripe_head *sh, int i);
238 
239 static void init_stripe(struct stripe_head *sh, sector_t sector, int pd_idx, int disks)
240 {
241 	raid5_conf_t *conf = sh->raid_conf;
242 	int i;
243 
244 	BUG_ON(atomic_read(&sh->count) != 0);
245 	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
246 	BUG_ON(sh->ops.pending || sh->ops.ack || sh->ops.complete);
247 
248 	CHECK_DEVLOCK();
249 	pr_debug("init_stripe called, stripe %llu\n",
250 		(unsigned long long)sh->sector);
251 
252 	remove_hash(sh);
253 
254 	sh->sector = sector;
255 	sh->pd_idx = pd_idx;
256 	sh->state = 0;
257 
258 	sh->disks = disks;
259 
260 	for (i = sh->disks; i--; ) {
261 		struct r5dev *dev = &sh->dev[i];
262 
263 		if (dev->toread || dev->read || dev->towrite || dev->written ||
264 		    test_bit(R5_LOCKED, &dev->flags)) {
265 			printk(KERN_ERR "sector=%llx i=%d %p %p %p %p %d\n",
266 			       (unsigned long long)sh->sector, i, dev->toread,
267 			       dev->read, dev->towrite, dev->written,
268 			       test_bit(R5_LOCKED, &dev->flags));
269 			BUG();
270 		}
271 		dev->flags = 0;
272 		raid5_build_block(sh, i);
273 	}
274 	insert_hash(conf, sh);
275 }
276 
277 static struct stripe_head *__find_stripe(raid5_conf_t *conf, sector_t sector, int disks)
278 {
279 	struct stripe_head *sh;
280 	struct hlist_node *hn;
281 
282 	CHECK_DEVLOCK();
283 	pr_debug("__find_stripe, sector %llu\n", (unsigned long long)sector);
284 	hlist_for_each_entry(sh, hn, stripe_hash(conf, sector), hash)
285 		if (sh->sector == sector && sh->disks == disks)
286 			return sh;
287 	pr_debug("__stripe %llu not in cache\n", (unsigned long long)sector);
288 	return NULL;
289 }
290 
291 static void unplug_slaves(mddev_t *mddev);
292 static void raid5_unplug_device(struct request_queue *q);
293 
294 static struct stripe_head *get_active_stripe(raid5_conf_t *conf, sector_t sector, int disks,
295 					     int pd_idx, int noblock)
296 {
297 	struct stripe_head *sh;
298 
299 	pr_debug("get_stripe, sector %llu\n", (unsigned long long)sector);
300 
301 	spin_lock_irq(&conf->device_lock);
302 
303 	do {
304 		wait_event_lock_irq(conf->wait_for_stripe,
305 				    conf->quiesce == 0,
306 				    conf->device_lock, /* nothing */);
307 		sh = __find_stripe(conf, sector, disks);
308 		if (!sh) {
309 			if (!conf->inactive_blocked)
310 				sh = get_free_stripe(conf);
311 			if (noblock && sh == NULL)
312 				break;
313 			if (!sh) {
314 				conf->inactive_blocked = 1;
315 				wait_event_lock_irq(conf->wait_for_stripe,
316 						    !list_empty(&conf->inactive_list) &&
317 						    (atomic_read(&conf->active_stripes)
318 						     < (conf->max_nr_stripes *3/4)
319 						     || !conf->inactive_blocked),
320 						    conf->device_lock,
321 						    raid5_unplug_device(conf->mddev->queue)
322 					);
323 				conf->inactive_blocked = 0;
324 			} else
325 				init_stripe(sh, sector, pd_idx, disks);
326 		} else {
327 			if (atomic_read(&sh->count)) {
328 			  BUG_ON(!list_empty(&sh->lru));
329 			} else {
330 				if (!test_bit(STRIPE_HANDLE, &sh->state))
331 					atomic_inc(&conf->active_stripes);
332 				if (list_empty(&sh->lru) &&
333 				    !test_bit(STRIPE_EXPANDING, &sh->state))
334 					BUG();
335 				list_del_init(&sh->lru);
336 			}
337 		}
338 	} while (sh == NULL);
339 
340 	if (sh)
341 		atomic_inc(&sh->count);
342 
343 	spin_unlock_irq(&conf->device_lock);
344 	return sh;
345 }
346 
347 /* test_and_ack_op() ensures that we only dequeue an operation once */
348 #define test_and_ack_op(op, pend) \
349 do {							\
350 	if (test_bit(op, &sh->ops.pending) &&		\
351 		!test_bit(op, &sh->ops.complete)) {	\
352 		if (test_and_set_bit(op, &sh->ops.ack)) \
353 			clear_bit(op, &pend);		\
354 		else					\
355 			ack++;				\
356 	} else						\
357 		clear_bit(op, &pend);			\
358 } while (0)
359 
360 /* find new work to run, do not resubmit work that is already
361  * in flight
362  */
363 static unsigned long get_stripe_work(struct stripe_head *sh)
364 {
365 	unsigned long pending;
366 	int ack = 0;
367 
368 	pending = sh->ops.pending;
369 
370 	test_and_ack_op(STRIPE_OP_BIOFILL, pending);
371 	test_and_ack_op(STRIPE_OP_COMPUTE_BLK, pending);
372 	test_and_ack_op(STRIPE_OP_PREXOR, pending);
373 	test_and_ack_op(STRIPE_OP_BIODRAIN, pending);
374 	test_and_ack_op(STRIPE_OP_POSTXOR, pending);
375 	test_and_ack_op(STRIPE_OP_CHECK, pending);
376 	if (test_and_clear_bit(STRIPE_OP_IO, &sh->ops.pending))
377 		ack++;
378 
379 	sh->ops.count -= ack;
380 	if (unlikely(sh->ops.count < 0)) {
381 		printk(KERN_ERR "pending: %#lx ops.pending: %#lx ops.ack: %#lx "
382 			"ops.complete: %#lx\n", pending, sh->ops.pending,
383 			sh->ops.ack, sh->ops.complete);
384 		BUG();
385 	}
386 
387 	return pending;
388 }
389 
390 static void
391 raid5_end_read_request(struct bio *bi, int error);
392 static void
393 raid5_end_write_request(struct bio *bi, int error);
394 
395 static void ops_run_io(struct stripe_head *sh)
396 {
397 	raid5_conf_t *conf = sh->raid_conf;
398 	int i, disks = sh->disks;
399 
400 	might_sleep();
401 
402 	set_bit(STRIPE_IO_STARTED, &sh->state);
403 	for (i = disks; i--; ) {
404 		int rw;
405 		struct bio *bi;
406 		mdk_rdev_t *rdev;
407 		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
408 			rw = WRITE;
409 		else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
410 			rw = READ;
411 		else
412 			continue;
413 
414 		bi = &sh->dev[i].req;
415 
416 		bi->bi_rw = rw;
417 		if (rw == WRITE)
418 			bi->bi_end_io = raid5_end_write_request;
419 		else
420 			bi->bi_end_io = raid5_end_read_request;
421 
422 		rcu_read_lock();
423 		rdev = rcu_dereference(conf->disks[i].rdev);
424 		if (rdev && test_bit(Faulty, &rdev->flags))
425 			rdev = NULL;
426 		if (rdev)
427 			atomic_inc(&rdev->nr_pending);
428 		rcu_read_unlock();
429 
430 		if (rdev) {
431 			if (test_bit(STRIPE_SYNCING, &sh->state) ||
432 				test_bit(STRIPE_EXPAND_SOURCE, &sh->state) ||
433 				test_bit(STRIPE_EXPAND_READY, &sh->state))
434 				md_sync_acct(rdev->bdev, STRIPE_SECTORS);
435 
436 			bi->bi_bdev = rdev->bdev;
437 			pr_debug("%s: for %llu schedule op %ld on disc %d\n",
438 				__func__, (unsigned long long)sh->sector,
439 				bi->bi_rw, i);
440 			atomic_inc(&sh->count);
441 			bi->bi_sector = sh->sector + rdev->data_offset;
442 			bi->bi_flags = 1 << BIO_UPTODATE;
443 			bi->bi_vcnt = 1;
444 			bi->bi_max_vecs = 1;
445 			bi->bi_idx = 0;
446 			bi->bi_io_vec = &sh->dev[i].vec;
447 			bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
448 			bi->bi_io_vec[0].bv_offset = 0;
449 			bi->bi_size = STRIPE_SIZE;
450 			bi->bi_next = NULL;
451 			if (rw == WRITE &&
452 			    test_bit(R5_ReWrite, &sh->dev[i].flags))
453 				atomic_add(STRIPE_SECTORS,
454 					&rdev->corrected_errors);
455 			generic_make_request(bi);
456 		} else {
457 			if (rw == WRITE)
458 				set_bit(STRIPE_DEGRADED, &sh->state);
459 			pr_debug("skip op %ld on disc %d for sector %llu\n",
460 				bi->bi_rw, i, (unsigned long long)sh->sector);
461 			clear_bit(R5_LOCKED, &sh->dev[i].flags);
462 			set_bit(STRIPE_HANDLE, &sh->state);
463 		}
464 	}
465 }
466 
467 static struct dma_async_tx_descriptor *
468 async_copy_data(int frombio, struct bio *bio, struct page *page,
469 	sector_t sector, struct dma_async_tx_descriptor *tx)
470 {
471 	struct bio_vec *bvl;
472 	struct page *bio_page;
473 	int i;
474 	int page_offset;
475 
476 	if (bio->bi_sector >= sector)
477 		page_offset = (signed)(bio->bi_sector - sector) * 512;
478 	else
479 		page_offset = (signed)(sector - bio->bi_sector) * -512;
480 	bio_for_each_segment(bvl, bio, i) {
481 		int len = bio_iovec_idx(bio, i)->bv_len;
482 		int clen;
483 		int b_offset = 0;
484 
485 		if (page_offset < 0) {
486 			b_offset = -page_offset;
487 			page_offset += b_offset;
488 			len -= b_offset;
489 		}
490 
491 		if (len > 0 && page_offset + len > STRIPE_SIZE)
492 			clen = STRIPE_SIZE - page_offset;
493 		else
494 			clen = len;
495 
496 		if (clen > 0) {
497 			b_offset += bio_iovec_idx(bio, i)->bv_offset;
498 			bio_page = bio_iovec_idx(bio, i)->bv_page;
499 			if (frombio)
500 				tx = async_memcpy(page, bio_page, page_offset,
501 					b_offset, clen,
502 					ASYNC_TX_DEP_ACK,
503 					tx, NULL, NULL);
504 			else
505 				tx = async_memcpy(bio_page, page, b_offset,
506 					page_offset, clen,
507 					ASYNC_TX_DEP_ACK,
508 					tx, NULL, NULL);
509 		}
510 		if (clen < len) /* hit end of page */
511 			break;
512 		page_offset +=  len;
513 	}
514 
515 	return tx;
516 }
517 
518 static void ops_complete_biofill(void *stripe_head_ref)
519 {
520 	struct stripe_head *sh = stripe_head_ref;
521 	struct bio *return_bi = NULL;
522 	raid5_conf_t *conf = sh->raid_conf;
523 	int i;
524 
525 	pr_debug("%s: stripe %llu\n", __func__,
526 		(unsigned long long)sh->sector);
527 
528 	/* clear completed biofills */
529 	for (i = sh->disks; i--; ) {
530 		struct r5dev *dev = &sh->dev[i];
531 
532 		/* acknowledge completion of a biofill operation */
533 		/* and check if we need to reply to a read request,
534 		 * new R5_Wantfill requests are held off until
535 		 * !test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending)
536 		 */
537 		if (test_and_clear_bit(R5_Wantfill, &dev->flags)) {
538 			struct bio *rbi, *rbi2;
539 
540 			/* The access to dev->read is outside of the
541 			 * spin_lock_irq(&conf->device_lock), but is protected
542 			 * by the STRIPE_OP_BIOFILL pending bit
543 			 */
544 			BUG_ON(!dev->read);
545 			rbi = dev->read;
546 			dev->read = NULL;
547 			while (rbi && rbi->bi_sector <
548 				dev->sector + STRIPE_SECTORS) {
549 				rbi2 = r5_next_bio(rbi, dev->sector);
550 				spin_lock_irq(&conf->device_lock);
551 				if (--rbi->bi_phys_segments == 0) {
552 					rbi->bi_next = return_bi;
553 					return_bi = rbi;
554 				}
555 				spin_unlock_irq(&conf->device_lock);
556 				rbi = rbi2;
557 			}
558 		}
559 	}
560 	set_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
561 
562 	return_io(return_bi);
563 
564 	set_bit(STRIPE_HANDLE, &sh->state);
565 	release_stripe(sh);
566 }
567 
568 static void ops_run_biofill(struct stripe_head *sh)
569 {
570 	struct dma_async_tx_descriptor *tx = NULL;
571 	raid5_conf_t *conf = sh->raid_conf;
572 	int i;
573 
574 	pr_debug("%s: stripe %llu\n", __func__,
575 		(unsigned long long)sh->sector);
576 
577 	for (i = sh->disks; i--; ) {
578 		struct r5dev *dev = &sh->dev[i];
579 		if (test_bit(R5_Wantfill, &dev->flags)) {
580 			struct bio *rbi;
581 			spin_lock_irq(&conf->device_lock);
582 			dev->read = rbi = dev->toread;
583 			dev->toread = NULL;
584 			spin_unlock_irq(&conf->device_lock);
585 			while (rbi && rbi->bi_sector <
586 				dev->sector + STRIPE_SECTORS) {
587 				tx = async_copy_data(0, rbi, dev->page,
588 					dev->sector, tx);
589 				rbi = r5_next_bio(rbi, dev->sector);
590 			}
591 		}
592 	}
593 
594 	atomic_inc(&sh->count);
595 	async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
596 		ops_complete_biofill, sh);
597 }
598 
599 static void ops_complete_compute5(void *stripe_head_ref)
600 {
601 	struct stripe_head *sh = stripe_head_ref;
602 	int target = sh->ops.target;
603 	struct r5dev *tgt = &sh->dev[target];
604 
605 	pr_debug("%s: stripe %llu\n", __func__,
606 		(unsigned long long)sh->sector);
607 
608 	set_bit(R5_UPTODATE, &tgt->flags);
609 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
610 	clear_bit(R5_Wantcompute, &tgt->flags);
611 	set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
612 	set_bit(STRIPE_HANDLE, &sh->state);
613 	release_stripe(sh);
614 }
615 
616 static struct dma_async_tx_descriptor *
617 ops_run_compute5(struct stripe_head *sh, unsigned long pending)
618 {
619 	/* kernel stack size limits the total number of disks */
620 	int disks = sh->disks;
621 	struct page *xor_srcs[disks];
622 	int target = sh->ops.target;
623 	struct r5dev *tgt = &sh->dev[target];
624 	struct page *xor_dest = tgt->page;
625 	int count = 0;
626 	struct dma_async_tx_descriptor *tx;
627 	int i;
628 
629 	pr_debug("%s: stripe %llu block: %d\n",
630 		__func__, (unsigned long long)sh->sector, target);
631 	BUG_ON(!test_bit(R5_Wantcompute, &tgt->flags));
632 
633 	for (i = disks; i--; )
634 		if (i != target)
635 			xor_srcs[count++] = sh->dev[i].page;
636 
637 	atomic_inc(&sh->count);
638 
639 	if (unlikely(count == 1))
640 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
641 			0, NULL, ops_complete_compute5, sh);
642 	else
643 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
644 			ASYNC_TX_XOR_ZERO_DST, NULL,
645 			ops_complete_compute5, sh);
646 
647 	/* ack now if postxor is not set to be run */
648 	if (tx && !test_bit(STRIPE_OP_POSTXOR, &pending))
649 		async_tx_ack(tx);
650 
651 	return tx;
652 }
653 
654 static void ops_complete_prexor(void *stripe_head_ref)
655 {
656 	struct stripe_head *sh = stripe_head_ref;
657 
658 	pr_debug("%s: stripe %llu\n", __func__,
659 		(unsigned long long)sh->sector);
660 
661 	set_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
662 }
663 
664 static struct dma_async_tx_descriptor *
665 ops_run_prexor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx)
666 {
667 	/* kernel stack size limits the total number of disks */
668 	int disks = sh->disks;
669 	struct page *xor_srcs[disks];
670 	int count = 0, pd_idx = sh->pd_idx, i;
671 
672 	/* existing parity data subtracted */
673 	struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
674 
675 	pr_debug("%s: stripe %llu\n", __func__,
676 		(unsigned long long)sh->sector);
677 
678 	for (i = disks; i--; ) {
679 		struct r5dev *dev = &sh->dev[i];
680 		/* Only process blocks that are known to be uptodate */
681 		if (dev->towrite && test_bit(R5_Wantprexor, &dev->flags))
682 			xor_srcs[count++] = dev->page;
683 	}
684 
685 	tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
686 		ASYNC_TX_DEP_ACK | ASYNC_TX_XOR_DROP_DST, tx,
687 		ops_complete_prexor, sh);
688 
689 	return tx;
690 }
691 
692 static struct dma_async_tx_descriptor *
693 ops_run_biodrain(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
694 		 unsigned long pending)
695 {
696 	int disks = sh->disks;
697 	int pd_idx = sh->pd_idx, i;
698 
699 	/* check if prexor is active which means only process blocks
700 	 * that are part of a read-modify-write (Wantprexor)
701 	 */
702 	int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
703 
704 	pr_debug("%s: stripe %llu\n", __func__,
705 		(unsigned long long)sh->sector);
706 
707 	for (i = disks; i--; ) {
708 		struct r5dev *dev = &sh->dev[i];
709 		struct bio *chosen;
710 		int towrite;
711 
712 		towrite = 0;
713 		if (prexor) { /* rmw */
714 			if (dev->towrite &&
715 			    test_bit(R5_Wantprexor, &dev->flags))
716 				towrite = 1;
717 		} else { /* rcw */
718 			if (i != pd_idx && dev->towrite &&
719 				test_bit(R5_LOCKED, &dev->flags))
720 				towrite = 1;
721 		}
722 
723 		if (towrite) {
724 			struct bio *wbi;
725 
726 			spin_lock(&sh->lock);
727 			chosen = dev->towrite;
728 			dev->towrite = NULL;
729 			BUG_ON(dev->written);
730 			wbi = dev->written = chosen;
731 			spin_unlock(&sh->lock);
732 
733 			while (wbi && wbi->bi_sector <
734 				dev->sector + STRIPE_SECTORS) {
735 				tx = async_copy_data(1, wbi, dev->page,
736 					dev->sector, tx);
737 				wbi = r5_next_bio(wbi, dev->sector);
738 			}
739 		}
740 	}
741 
742 	return tx;
743 }
744 
745 static void ops_complete_postxor(void *stripe_head_ref)
746 {
747 	struct stripe_head *sh = stripe_head_ref;
748 
749 	pr_debug("%s: stripe %llu\n", __func__,
750 		(unsigned long long)sh->sector);
751 
752 	set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
753 	set_bit(STRIPE_HANDLE, &sh->state);
754 	release_stripe(sh);
755 }
756 
757 static void ops_complete_write(void *stripe_head_ref)
758 {
759 	struct stripe_head *sh = stripe_head_ref;
760 	int disks = sh->disks, i, pd_idx = sh->pd_idx;
761 
762 	pr_debug("%s: stripe %llu\n", __func__,
763 		(unsigned long long)sh->sector);
764 
765 	for (i = disks; i--; ) {
766 		struct r5dev *dev = &sh->dev[i];
767 		if (dev->written || i == pd_idx)
768 			set_bit(R5_UPTODATE, &dev->flags);
769 	}
770 
771 	set_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
772 	set_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
773 
774 	set_bit(STRIPE_HANDLE, &sh->state);
775 	release_stripe(sh);
776 }
777 
778 static void
779 ops_run_postxor(struct stripe_head *sh, struct dma_async_tx_descriptor *tx,
780 		unsigned long pending)
781 {
782 	/* kernel stack size limits the total number of disks */
783 	int disks = sh->disks;
784 	struct page *xor_srcs[disks];
785 
786 	int count = 0, pd_idx = sh->pd_idx, i;
787 	struct page *xor_dest;
788 	int prexor = test_bit(STRIPE_OP_PREXOR, &pending);
789 	unsigned long flags;
790 	dma_async_tx_callback callback;
791 
792 	pr_debug("%s: stripe %llu\n", __func__,
793 		(unsigned long long)sh->sector);
794 
795 	/* check if prexor is active which means only process blocks
796 	 * that are part of a read-modify-write (written)
797 	 */
798 	if (prexor) {
799 		xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
800 		for (i = disks; i--; ) {
801 			struct r5dev *dev = &sh->dev[i];
802 			if (dev->written)
803 				xor_srcs[count++] = dev->page;
804 		}
805 	} else {
806 		xor_dest = sh->dev[pd_idx].page;
807 		for (i = disks; i--; ) {
808 			struct r5dev *dev = &sh->dev[i];
809 			if (i != pd_idx)
810 				xor_srcs[count++] = dev->page;
811 		}
812 	}
813 
814 	/* check whether this postxor is part of a write */
815 	callback = test_bit(STRIPE_OP_BIODRAIN, &pending) ?
816 		ops_complete_write : ops_complete_postxor;
817 
818 	/* 1/ if we prexor'd then the dest is reused as a source
819 	 * 2/ if we did not prexor then we are redoing the parity
820 	 * set ASYNC_TX_XOR_DROP_DST and ASYNC_TX_XOR_ZERO_DST
821 	 * for the synchronous xor case
822 	 */
823 	flags = ASYNC_TX_DEP_ACK | ASYNC_TX_ACK |
824 		(prexor ? ASYNC_TX_XOR_DROP_DST : ASYNC_TX_XOR_ZERO_DST);
825 
826 	atomic_inc(&sh->count);
827 
828 	if (unlikely(count == 1)) {
829 		flags &= ~(ASYNC_TX_XOR_DROP_DST | ASYNC_TX_XOR_ZERO_DST);
830 		tx = async_memcpy(xor_dest, xor_srcs[0], 0, 0, STRIPE_SIZE,
831 			flags, tx, callback, sh);
832 	} else
833 		tx = async_xor(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
834 			flags, tx, callback, sh);
835 }
836 
837 static void ops_complete_check(void *stripe_head_ref)
838 {
839 	struct stripe_head *sh = stripe_head_ref;
840 	int pd_idx = sh->pd_idx;
841 
842 	pr_debug("%s: stripe %llu\n", __func__,
843 		(unsigned long long)sh->sector);
844 
845 	if (test_and_clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending) &&
846 		sh->ops.zero_sum_result == 0)
847 		set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
848 
849 	set_bit(STRIPE_OP_CHECK, &sh->ops.complete);
850 	set_bit(STRIPE_HANDLE, &sh->state);
851 	release_stripe(sh);
852 }
853 
854 static void ops_run_check(struct stripe_head *sh)
855 {
856 	/* kernel stack size limits the total number of disks */
857 	int disks = sh->disks;
858 	struct page *xor_srcs[disks];
859 	struct dma_async_tx_descriptor *tx;
860 
861 	int count = 0, pd_idx = sh->pd_idx, i;
862 	struct page *xor_dest = xor_srcs[count++] = sh->dev[pd_idx].page;
863 
864 	pr_debug("%s: stripe %llu\n", __func__,
865 		(unsigned long long)sh->sector);
866 
867 	for (i = disks; i--; ) {
868 		struct r5dev *dev = &sh->dev[i];
869 		if (i != pd_idx)
870 			xor_srcs[count++] = dev->page;
871 	}
872 
873 	tx = async_xor_zero_sum(xor_dest, xor_srcs, 0, count, STRIPE_SIZE,
874 		&sh->ops.zero_sum_result, 0, NULL, NULL, NULL);
875 
876 	if (tx)
877 		set_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
878 	else
879 		clear_bit(STRIPE_OP_MOD_DMA_CHECK, &sh->ops.pending);
880 
881 	atomic_inc(&sh->count);
882 	tx = async_trigger_callback(ASYNC_TX_DEP_ACK | ASYNC_TX_ACK, tx,
883 		ops_complete_check, sh);
884 }
885 
886 static void raid5_run_ops(struct stripe_head *sh, unsigned long pending)
887 {
888 	int overlap_clear = 0, i, disks = sh->disks;
889 	struct dma_async_tx_descriptor *tx = NULL;
890 
891 	if (test_bit(STRIPE_OP_BIOFILL, &pending)) {
892 		ops_run_biofill(sh);
893 		overlap_clear++;
894 	}
895 
896 	if (test_bit(STRIPE_OP_COMPUTE_BLK, &pending))
897 		tx = ops_run_compute5(sh, pending);
898 
899 	if (test_bit(STRIPE_OP_PREXOR, &pending))
900 		tx = ops_run_prexor(sh, tx);
901 
902 	if (test_bit(STRIPE_OP_BIODRAIN, &pending)) {
903 		tx = ops_run_biodrain(sh, tx, pending);
904 		overlap_clear++;
905 	}
906 
907 	if (test_bit(STRIPE_OP_POSTXOR, &pending))
908 		ops_run_postxor(sh, tx, pending);
909 
910 	if (test_bit(STRIPE_OP_CHECK, &pending))
911 		ops_run_check(sh);
912 
913 	if (test_bit(STRIPE_OP_IO, &pending))
914 		ops_run_io(sh);
915 
916 	if (overlap_clear)
917 		for (i = disks; i--; ) {
918 			struct r5dev *dev = &sh->dev[i];
919 			if (test_and_clear_bit(R5_Overlap, &dev->flags))
920 				wake_up(&sh->raid_conf->wait_for_overlap);
921 		}
922 }
923 
924 static int grow_one_stripe(raid5_conf_t *conf)
925 {
926 	struct stripe_head *sh;
927 	sh = kmem_cache_alloc(conf->slab_cache, GFP_KERNEL);
928 	if (!sh)
929 		return 0;
930 	memset(sh, 0, sizeof(*sh) + (conf->raid_disks-1)*sizeof(struct r5dev));
931 	sh->raid_conf = conf;
932 	spin_lock_init(&sh->lock);
933 
934 	if (grow_buffers(sh, conf->raid_disks)) {
935 		shrink_buffers(sh, conf->raid_disks);
936 		kmem_cache_free(conf->slab_cache, sh);
937 		return 0;
938 	}
939 	sh->disks = conf->raid_disks;
940 	/* we just created an active stripe so... */
941 	atomic_set(&sh->count, 1);
942 	atomic_inc(&conf->active_stripes);
943 	INIT_LIST_HEAD(&sh->lru);
944 	release_stripe(sh);
945 	return 1;
946 }
947 
948 static int grow_stripes(raid5_conf_t *conf, int num)
949 {
950 	struct kmem_cache *sc;
951 	int devs = conf->raid_disks;
952 
953 	sprintf(conf->cache_name[0], "raid5-%s", mdname(conf->mddev));
954 	sprintf(conf->cache_name[1], "raid5-%s-alt", mdname(conf->mddev));
955 	conf->active_name = 0;
956 	sc = kmem_cache_create(conf->cache_name[conf->active_name],
957 			       sizeof(struct stripe_head)+(devs-1)*sizeof(struct r5dev),
958 			       0, 0, NULL);
959 	if (!sc)
960 		return 1;
961 	conf->slab_cache = sc;
962 	conf->pool_size = devs;
963 	while (num--)
964 		if (!grow_one_stripe(conf))
965 			return 1;
966 	return 0;
967 }
968 
969 #ifdef CONFIG_MD_RAID5_RESHAPE
970 static int resize_stripes(raid5_conf_t *conf, int newsize)
971 {
972 	/* Make all the stripes able to hold 'newsize' devices.
973 	 * New slots in each stripe get 'page' set to a new page.
974 	 *
975 	 * This happens in stages:
976 	 * 1/ create a new kmem_cache and allocate the required number of
977 	 *    stripe_heads.
978 	 * 2/ gather all the old stripe_heads and tranfer the pages across
979 	 *    to the new stripe_heads.  This will have the side effect of
980 	 *    freezing the array as once all stripe_heads have been collected,
981 	 *    no IO will be possible.  Old stripe heads are freed once their
982 	 *    pages have been transferred over, and the old kmem_cache is
983 	 *    freed when all stripes are done.
984 	 * 3/ reallocate conf->disks to be suitable bigger.  If this fails,
985 	 *    we simple return a failre status - no need to clean anything up.
986 	 * 4/ allocate new pages for the new slots in the new stripe_heads.
987 	 *    If this fails, we don't bother trying the shrink the
988 	 *    stripe_heads down again, we just leave them as they are.
989 	 *    As each stripe_head is processed the new one is released into
990 	 *    active service.
991 	 *
992 	 * Once step2 is started, we cannot afford to wait for a write,
993 	 * so we use GFP_NOIO allocations.
994 	 */
995 	struct stripe_head *osh, *nsh;
996 	LIST_HEAD(newstripes);
997 	struct disk_info *ndisks;
998 	int err = 0;
999 	struct kmem_cache *sc;
1000 	int i;
1001 
1002 	if (newsize <= conf->pool_size)
1003 		return 0; /* never bother to shrink */
1004 
1005 	md_allow_write(conf->mddev);
1006 
1007 	/* Step 1 */
1008 	sc = kmem_cache_create(conf->cache_name[1-conf->active_name],
1009 			       sizeof(struct stripe_head)+(newsize-1)*sizeof(struct r5dev),
1010 			       0, 0, NULL);
1011 	if (!sc)
1012 		return -ENOMEM;
1013 
1014 	for (i = conf->max_nr_stripes; i; i--) {
1015 		nsh = kmem_cache_alloc(sc, GFP_KERNEL);
1016 		if (!nsh)
1017 			break;
1018 
1019 		memset(nsh, 0, sizeof(*nsh) + (newsize-1)*sizeof(struct r5dev));
1020 
1021 		nsh->raid_conf = conf;
1022 		spin_lock_init(&nsh->lock);
1023 
1024 		list_add(&nsh->lru, &newstripes);
1025 	}
1026 	if (i) {
1027 		/* didn't get enough, give up */
1028 		while (!list_empty(&newstripes)) {
1029 			nsh = list_entry(newstripes.next, struct stripe_head, lru);
1030 			list_del(&nsh->lru);
1031 			kmem_cache_free(sc, nsh);
1032 		}
1033 		kmem_cache_destroy(sc);
1034 		return -ENOMEM;
1035 	}
1036 	/* Step 2 - Must use GFP_NOIO now.
1037 	 * OK, we have enough stripes, start collecting inactive
1038 	 * stripes and copying them over
1039 	 */
1040 	list_for_each_entry(nsh, &newstripes, lru) {
1041 		spin_lock_irq(&conf->device_lock);
1042 		wait_event_lock_irq(conf->wait_for_stripe,
1043 				    !list_empty(&conf->inactive_list),
1044 				    conf->device_lock,
1045 				    unplug_slaves(conf->mddev)
1046 			);
1047 		osh = get_free_stripe(conf);
1048 		spin_unlock_irq(&conf->device_lock);
1049 		atomic_set(&nsh->count, 1);
1050 		for(i=0; i<conf->pool_size; i++)
1051 			nsh->dev[i].page = osh->dev[i].page;
1052 		for( ; i<newsize; i++)
1053 			nsh->dev[i].page = NULL;
1054 		kmem_cache_free(conf->slab_cache, osh);
1055 	}
1056 	kmem_cache_destroy(conf->slab_cache);
1057 
1058 	/* Step 3.
1059 	 * At this point, we are holding all the stripes so the array
1060 	 * is completely stalled, so now is a good time to resize
1061 	 * conf->disks.
1062 	 */
1063 	ndisks = kzalloc(newsize * sizeof(struct disk_info), GFP_NOIO);
1064 	if (ndisks) {
1065 		for (i=0; i<conf->raid_disks; i++)
1066 			ndisks[i] = conf->disks[i];
1067 		kfree(conf->disks);
1068 		conf->disks = ndisks;
1069 	} else
1070 		err = -ENOMEM;
1071 
1072 	/* Step 4, return new stripes to service */
1073 	while(!list_empty(&newstripes)) {
1074 		nsh = list_entry(newstripes.next, struct stripe_head, lru);
1075 		list_del_init(&nsh->lru);
1076 		for (i=conf->raid_disks; i < newsize; i++)
1077 			if (nsh->dev[i].page == NULL) {
1078 				struct page *p = alloc_page(GFP_NOIO);
1079 				nsh->dev[i].page = p;
1080 				if (!p)
1081 					err = -ENOMEM;
1082 			}
1083 		release_stripe(nsh);
1084 	}
1085 	/* critical section pass, GFP_NOIO no longer needed */
1086 
1087 	conf->slab_cache = sc;
1088 	conf->active_name = 1-conf->active_name;
1089 	conf->pool_size = newsize;
1090 	return err;
1091 }
1092 #endif
1093 
1094 static int drop_one_stripe(raid5_conf_t *conf)
1095 {
1096 	struct stripe_head *sh;
1097 
1098 	spin_lock_irq(&conf->device_lock);
1099 	sh = get_free_stripe(conf);
1100 	spin_unlock_irq(&conf->device_lock);
1101 	if (!sh)
1102 		return 0;
1103 	BUG_ON(atomic_read(&sh->count));
1104 	shrink_buffers(sh, conf->pool_size);
1105 	kmem_cache_free(conf->slab_cache, sh);
1106 	atomic_dec(&conf->active_stripes);
1107 	return 1;
1108 }
1109 
1110 static void shrink_stripes(raid5_conf_t *conf)
1111 {
1112 	while (drop_one_stripe(conf))
1113 		;
1114 
1115 	if (conf->slab_cache)
1116 		kmem_cache_destroy(conf->slab_cache);
1117 	conf->slab_cache = NULL;
1118 }
1119 
1120 static void raid5_end_read_request(struct bio * bi, int error)
1121 {
1122  	struct stripe_head *sh = bi->bi_private;
1123 	raid5_conf_t *conf = sh->raid_conf;
1124 	int disks = sh->disks, i;
1125 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1126 	char b[BDEVNAME_SIZE];
1127 	mdk_rdev_t *rdev;
1128 
1129 
1130 	for (i=0 ; i<disks; i++)
1131 		if (bi == &sh->dev[i].req)
1132 			break;
1133 
1134 	pr_debug("end_read_request %llu/%d, count: %d, uptodate %d.\n",
1135 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
1136 		uptodate);
1137 	if (i == disks) {
1138 		BUG();
1139 		return;
1140 	}
1141 
1142 	if (uptodate) {
1143 		set_bit(R5_UPTODATE, &sh->dev[i].flags);
1144 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1145 			rdev = conf->disks[i].rdev;
1146 			printk(KERN_INFO "raid5:%s: read error corrected (%lu sectors at %llu on %s)\n",
1147 			       mdname(conf->mddev), STRIPE_SECTORS,
1148 			       (unsigned long long)(sh->sector + rdev->data_offset),
1149 			       bdevname(rdev->bdev, b));
1150 			clear_bit(R5_ReadError, &sh->dev[i].flags);
1151 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
1152 		}
1153 		if (atomic_read(&conf->disks[i].rdev->read_errors))
1154 			atomic_set(&conf->disks[i].rdev->read_errors, 0);
1155 	} else {
1156 		const char *bdn = bdevname(conf->disks[i].rdev->bdev, b);
1157 		int retry = 0;
1158 		rdev = conf->disks[i].rdev;
1159 
1160 		clear_bit(R5_UPTODATE, &sh->dev[i].flags);
1161 		atomic_inc(&rdev->read_errors);
1162 		if (conf->mddev->degraded)
1163 			printk(KERN_WARNING "raid5:%s: read error not correctable (sector %llu on %s).\n",
1164 			       mdname(conf->mddev),
1165 			       (unsigned long long)(sh->sector + rdev->data_offset),
1166 			       bdn);
1167 		else if (test_bit(R5_ReWrite, &sh->dev[i].flags))
1168 			/* Oh, no!!! */
1169 			printk(KERN_WARNING "raid5:%s: read error NOT corrected!! (sector %llu on %s).\n",
1170 			       mdname(conf->mddev),
1171 			       (unsigned long long)(sh->sector + rdev->data_offset),
1172 			       bdn);
1173 		else if (atomic_read(&rdev->read_errors)
1174 			 > conf->max_nr_stripes)
1175 			printk(KERN_WARNING
1176 			       "raid5:%s: Too many read errors, failing device %s.\n",
1177 			       mdname(conf->mddev), bdn);
1178 		else
1179 			retry = 1;
1180 		if (retry)
1181 			set_bit(R5_ReadError, &sh->dev[i].flags);
1182 		else {
1183 			clear_bit(R5_ReadError, &sh->dev[i].flags);
1184 			clear_bit(R5_ReWrite, &sh->dev[i].flags);
1185 			md_error(conf->mddev, rdev);
1186 		}
1187 	}
1188 	rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1189 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
1190 	set_bit(STRIPE_HANDLE, &sh->state);
1191 	release_stripe(sh);
1192 }
1193 
1194 static void raid5_end_write_request (struct bio *bi, int error)
1195 {
1196  	struct stripe_head *sh = bi->bi_private;
1197 	raid5_conf_t *conf = sh->raid_conf;
1198 	int disks = sh->disks, i;
1199 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
1200 
1201 	for (i=0 ; i<disks; i++)
1202 		if (bi == &sh->dev[i].req)
1203 			break;
1204 
1205 	pr_debug("end_write_request %llu/%d, count %d, uptodate: %d.\n",
1206 		(unsigned long long)sh->sector, i, atomic_read(&sh->count),
1207 		uptodate);
1208 	if (i == disks) {
1209 		BUG();
1210 		return;
1211 	}
1212 
1213 	if (!uptodate)
1214 		md_error(conf->mddev, conf->disks[i].rdev);
1215 
1216 	rdev_dec_pending(conf->disks[i].rdev, conf->mddev);
1217 
1218 	clear_bit(R5_LOCKED, &sh->dev[i].flags);
1219 	set_bit(STRIPE_HANDLE, &sh->state);
1220 	release_stripe(sh);
1221 }
1222 
1223 
1224 static sector_t compute_blocknr(struct stripe_head *sh, int i);
1225 
1226 static void raid5_build_block (struct stripe_head *sh, int i)
1227 {
1228 	struct r5dev *dev = &sh->dev[i];
1229 
1230 	bio_init(&dev->req);
1231 	dev->req.bi_io_vec = &dev->vec;
1232 	dev->req.bi_vcnt++;
1233 	dev->req.bi_max_vecs++;
1234 	dev->vec.bv_page = dev->page;
1235 	dev->vec.bv_len = STRIPE_SIZE;
1236 	dev->vec.bv_offset = 0;
1237 
1238 	dev->req.bi_sector = sh->sector;
1239 	dev->req.bi_private = sh;
1240 
1241 	dev->flags = 0;
1242 	dev->sector = compute_blocknr(sh, i);
1243 }
1244 
1245 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1246 {
1247 	char b[BDEVNAME_SIZE];
1248 	raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
1249 	pr_debug("raid5: error called\n");
1250 
1251 	if (!test_bit(Faulty, &rdev->flags)) {
1252 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
1253 		if (test_and_clear_bit(In_sync, &rdev->flags)) {
1254 			unsigned long flags;
1255 			spin_lock_irqsave(&conf->device_lock, flags);
1256 			mddev->degraded++;
1257 			spin_unlock_irqrestore(&conf->device_lock, flags);
1258 			/*
1259 			 * if recovery was running, make sure it aborts.
1260 			 */
1261 			set_bit(MD_RECOVERY_ERR, &mddev->recovery);
1262 		}
1263 		set_bit(Faulty, &rdev->flags);
1264 		printk (KERN_ALERT
1265 			"raid5: Disk failure on %s, disabling device.\n"
1266 			"raid5: Operation continuing on %d devices.\n",
1267 			bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1268 	}
1269 }
1270 
1271 /*
1272  * Input: a 'big' sector number,
1273  * Output: index of the data and parity disk, and the sector # in them.
1274  */
1275 static sector_t raid5_compute_sector(sector_t r_sector, unsigned int raid_disks,
1276 			unsigned int data_disks, unsigned int * dd_idx,
1277 			unsigned int * pd_idx, raid5_conf_t *conf)
1278 {
1279 	long stripe;
1280 	unsigned long chunk_number;
1281 	unsigned int chunk_offset;
1282 	sector_t new_sector;
1283 	int sectors_per_chunk = conf->chunk_size >> 9;
1284 
1285 	/* First compute the information on this sector */
1286 
1287 	/*
1288 	 * Compute the chunk number and the sector offset inside the chunk
1289 	 */
1290 	chunk_offset = sector_div(r_sector, sectors_per_chunk);
1291 	chunk_number = r_sector;
1292 	BUG_ON(r_sector != chunk_number);
1293 
1294 	/*
1295 	 * Compute the stripe number
1296 	 */
1297 	stripe = chunk_number / data_disks;
1298 
1299 	/*
1300 	 * Compute the data disk and parity disk indexes inside the stripe
1301 	 */
1302 	*dd_idx = chunk_number % data_disks;
1303 
1304 	/*
1305 	 * Select the parity disk based on the user selected algorithm.
1306 	 */
1307 	switch(conf->level) {
1308 	case 4:
1309 		*pd_idx = data_disks;
1310 		break;
1311 	case 5:
1312 		switch (conf->algorithm) {
1313 		case ALGORITHM_LEFT_ASYMMETRIC:
1314 			*pd_idx = data_disks - stripe % raid_disks;
1315 			if (*dd_idx >= *pd_idx)
1316 				(*dd_idx)++;
1317 			break;
1318 		case ALGORITHM_RIGHT_ASYMMETRIC:
1319 			*pd_idx = stripe % raid_disks;
1320 			if (*dd_idx >= *pd_idx)
1321 				(*dd_idx)++;
1322 			break;
1323 		case ALGORITHM_LEFT_SYMMETRIC:
1324 			*pd_idx = data_disks - stripe % raid_disks;
1325 			*dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1326 			break;
1327 		case ALGORITHM_RIGHT_SYMMETRIC:
1328 			*pd_idx = stripe % raid_disks;
1329 			*dd_idx = (*pd_idx + 1 + *dd_idx) % raid_disks;
1330 			break;
1331 		default:
1332 			printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1333 				conf->algorithm);
1334 		}
1335 		break;
1336 	case 6:
1337 
1338 		/**** FIX THIS ****/
1339 		switch (conf->algorithm) {
1340 		case ALGORITHM_LEFT_ASYMMETRIC:
1341 			*pd_idx = raid_disks - 1 - (stripe % raid_disks);
1342 			if (*pd_idx == raid_disks-1)
1343 				(*dd_idx)++; 	/* Q D D D P */
1344 			else if (*dd_idx >= *pd_idx)
1345 				(*dd_idx) += 2; /* D D P Q D */
1346 			break;
1347 		case ALGORITHM_RIGHT_ASYMMETRIC:
1348 			*pd_idx = stripe % raid_disks;
1349 			if (*pd_idx == raid_disks-1)
1350 				(*dd_idx)++; 	/* Q D D D P */
1351 			else if (*dd_idx >= *pd_idx)
1352 				(*dd_idx) += 2; /* D D P Q D */
1353 			break;
1354 		case ALGORITHM_LEFT_SYMMETRIC:
1355 			*pd_idx = raid_disks - 1 - (stripe % raid_disks);
1356 			*dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1357 			break;
1358 		case ALGORITHM_RIGHT_SYMMETRIC:
1359 			*pd_idx = stripe % raid_disks;
1360 			*dd_idx = (*pd_idx + 2 + *dd_idx) % raid_disks;
1361 			break;
1362 		default:
1363 			printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1364 				conf->algorithm);
1365 		}
1366 		break;
1367 	}
1368 
1369 	/*
1370 	 * Finally, compute the new sector number
1371 	 */
1372 	new_sector = (sector_t)stripe * sectors_per_chunk + chunk_offset;
1373 	return new_sector;
1374 }
1375 
1376 
1377 static sector_t compute_blocknr(struct stripe_head *sh, int i)
1378 {
1379 	raid5_conf_t *conf = sh->raid_conf;
1380 	int raid_disks = sh->disks;
1381 	int data_disks = raid_disks - conf->max_degraded;
1382 	sector_t new_sector = sh->sector, check;
1383 	int sectors_per_chunk = conf->chunk_size >> 9;
1384 	sector_t stripe;
1385 	int chunk_offset;
1386 	int chunk_number, dummy1, dummy2, dd_idx = i;
1387 	sector_t r_sector;
1388 
1389 
1390 	chunk_offset = sector_div(new_sector, sectors_per_chunk);
1391 	stripe = new_sector;
1392 	BUG_ON(new_sector != stripe);
1393 
1394 	if (i == sh->pd_idx)
1395 		return 0;
1396 	switch(conf->level) {
1397 	case 4: break;
1398 	case 5:
1399 		switch (conf->algorithm) {
1400 		case ALGORITHM_LEFT_ASYMMETRIC:
1401 		case ALGORITHM_RIGHT_ASYMMETRIC:
1402 			if (i > sh->pd_idx)
1403 				i--;
1404 			break;
1405 		case ALGORITHM_LEFT_SYMMETRIC:
1406 		case ALGORITHM_RIGHT_SYMMETRIC:
1407 			if (i < sh->pd_idx)
1408 				i += raid_disks;
1409 			i -= (sh->pd_idx + 1);
1410 			break;
1411 		default:
1412 			printk(KERN_ERR "raid5: unsupported algorithm %d\n",
1413 			       conf->algorithm);
1414 		}
1415 		break;
1416 	case 6:
1417 		if (i == raid6_next_disk(sh->pd_idx, raid_disks))
1418 			return 0; /* It is the Q disk */
1419 		switch (conf->algorithm) {
1420 		case ALGORITHM_LEFT_ASYMMETRIC:
1421 		case ALGORITHM_RIGHT_ASYMMETRIC:
1422 		  	if (sh->pd_idx == raid_disks-1)
1423 				i--; 	/* Q D D D P */
1424 			else if (i > sh->pd_idx)
1425 				i -= 2; /* D D P Q D */
1426 			break;
1427 		case ALGORITHM_LEFT_SYMMETRIC:
1428 		case ALGORITHM_RIGHT_SYMMETRIC:
1429 			if (sh->pd_idx == raid_disks-1)
1430 				i--; /* Q D D D P */
1431 			else {
1432 				/* D D P Q D */
1433 				if (i < sh->pd_idx)
1434 					i += raid_disks;
1435 				i -= (sh->pd_idx + 2);
1436 			}
1437 			break;
1438 		default:
1439 			printk (KERN_CRIT "raid6: unsupported algorithm %d\n",
1440 				conf->algorithm);
1441 		}
1442 		break;
1443 	}
1444 
1445 	chunk_number = stripe * data_disks + i;
1446 	r_sector = (sector_t)chunk_number * sectors_per_chunk + chunk_offset;
1447 
1448 	check = raid5_compute_sector (r_sector, raid_disks, data_disks, &dummy1, &dummy2, conf);
1449 	if (check != sh->sector || dummy1 != dd_idx || dummy2 != sh->pd_idx) {
1450 		printk(KERN_ERR "compute_blocknr: map not correct\n");
1451 		return 0;
1452 	}
1453 	return r_sector;
1454 }
1455 
1456 
1457 
1458 /*
1459  * Copy data between a page in the stripe cache, and one or more bion
1460  * The page could align with the middle of the bio, or there could be
1461  * several bion, each with several bio_vecs, which cover part of the page
1462  * Multiple bion are linked together on bi_next.  There may be extras
1463  * at the end of this list.  We ignore them.
1464  */
1465 static void copy_data(int frombio, struct bio *bio,
1466 		     struct page *page,
1467 		     sector_t sector)
1468 {
1469 	char *pa = page_address(page);
1470 	struct bio_vec *bvl;
1471 	int i;
1472 	int page_offset;
1473 
1474 	if (bio->bi_sector >= sector)
1475 		page_offset = (signed)(bio->bi_sector - sector) * 512;
1476 	else
1477 		page_offset = (signed)(sector - bio->bi_sector) * -512;
1478 	bio_for_each_segment(bvl, bio, i) {
1479 		int len = bio_iovec_idx(bio,i)->bv_len;
1480 		int clen;
1481 		int b_offset = 0;
1482 
1483 		if (page_offset < 0) {
1484 			b_offset = -page_offset;
1485 			page_offset += b_offset;
1486 			len -= b_offset;
1487 		}
1488 
1489 		if (len > 0 && page_offset + len > STRIPE_SIZE)
1490 			clen = STRIPE_SIZE - page_offset;
1491 		else clen = len;
1492 
1493 		if (clen > 0) {
1494 			char *ba = __bio_kmap_atomic(bio, i, KM_USER0);
1495 			if (frombio)
1496 				memcpy(pa+page_offset, ba+b_offset, clen);
1497 			else
1498 				memcpy(ba+b_offset, pa+page_offset, clen);
1499 			__bio_kunmap_atomic(ba, KM_USER0);
1500 		}
1501 		if (clen < len) /* hit end of page */
1502 			break;
1503 		page_offset +=  len;
1504 	}
1505 }
1506 
1507 #define check_xor()	do {						  \
1508 				if (count == MAX_XOR_BLOCKS) {		  \
1509 				xor_blocks(count, STRIPE_SIZE, dest, ptr);\
1510 				count = 0;				  \
1511 			   }						  \
1512 			} while(0)
1513 
1514 static void compute_parity6(struct stripe_head *sh, int method)
1515 {
1516 	raid6_conf_t *conf = sh->raid_conf;
1517 	int i, pd_idx = sh->pd_idx, qd_idx, d0_idx, disks = sh->disks, count;
1518 	struct bio *chosen;
1519 	/**** FIX THIS: This could be very bad if disks is close to 256 ****/
1520 	void *ptrs[disks];
1521 
1522 	qd_idx = raid6_next_disk(pd_idx, disks);
1523 	d0_idx = raid6_next_disk(qd_idx, disks);
1524 
1525 	pr_debug("compute_parity, stripe %llu, method %d\n",
1526 		(unsigned long long)sh->sector, method);
1527 
1528 	switch(method) {
1529 	case READ_MODIFY_WRITE:
1530 		BUG();		/* READ_MODIFY_WRITE N/A for RAID-6 */
1531 	case RECONSTRUCT_WRITE:
1532 		for (i= disks; i-- ;)
1533 			if ( i != pd_idx && i != qd_idx && sh->dev[i].towrite ) {
1534 				chosen = sh->dev[i].towrite;
1535 				sh->dev[i].towrite = NULL;
1536 
1537 				if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1538 					wake_up(&conf->wait_for_overlap);
1539 
1540 				BUG_ON(sh->dev[i].written);
1541 				sh->dev[i].written = chosen;
1542 			}
1543 		break;
1544 	case CHECK_PARITY:
1545 		BUG();		/* Not implemented yet */
1546 	}
1547 
1548 	for (i = disks; i--;)
1549 		if (sh->dev[i].written) {
1550 			sector_t sector = sh->dev[i].sector;
1551 			struct bio *wbi = sh->dev[i].written;
1552 			while (wbi && wbi->bi_sector < sector + STRIPE_SECTORS) {
1553 				copy_data(1, wbi, sh->dev[i].page, sector);
1554 				wbi = r5_next_bio(wbi, sector);
1555 			}
1556 
1557 			set_bit(R5_LOCKED, &sh->dev[i].flags);
1558 			set_bit(R5_UPTODATE, &sh->dev[i].flags);
1559 		}
1560 
1561 //	switch(method) {
1562 //	case RECONSTRUCT_WRITE:
1563 //	case CHECK_PARITY:
1564 //	case UPDATE_PARITY:
1565 		/* Note that unlike RAID-5, the ordering of the disks matters greatly. */
1566 		/* FIX: Is this ordering of drives even remotely optimal? */
1567 		count = 0;
1568 		i = d0_idx;
1569 		do {
1570 			ptrs[count++] = page_address(sh->dev[i].page);
1571 			if (count <= disks-2 && !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1572 				printk("block %d/%d not uptodate on parity calc\n", i,count);
1573 			i = raid6_next_disk(i, disks);
1574 		} while ( i != d0_idx );
1575 //		break;
1576 //	}
1577 
1578 	raid6_call.gen_syndrome(disks, STRIPE_SIZE, ptrs);
1579 
1580 	switch(method) {
1581 	case RECONSTRUCT_WRITE:
1582 		set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1583 		set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1584 		set_bit(R5_LOCKED,   &sh->dev[pd_idx].flags);
1585 		set_bit(R5_LOCKED,   &sh->dev[qd_idx].flags);
1586 		break;
1587 	case UPDATE_PARITY:
1588 		set_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1589 		set_bit(R5_UPTODATE, &sh->dev[qd_idx].flags);
1590 		break;
1591 	}
1592 }
1593 
1594 
1595 /* Compute one missing block */
1596 static void compute_block_1(struct stripe_head *sh, int dd_idx, int nozero)
1597 {
1598 	int i, count, disks = sh->disks;
1599 	void *ptr[MAX_XOR_BLOCKS], *dest, *p;
1600 	int pd_idx = sh->pd_idx;
1601 	int qd_idx = raid6_next_disk(pd_idx, disks);
1602 
1603 	pr_debug("compute_block_1, stripe %llu, idx %d\n",
1604 		(unsigned long long)sh->sector, dd_idx);
1605 
1606 	if ( dd_idx == qd_idx ) {
1607 		/* We're actually computing the Q drive */
1608 		compute_parity6(sh, UPDATE_PARITY);
1609 	} else {
1610 		dest = page_address(sh->dev[dd_idx].page);
1611 		if (!nozero) memset(dest, 0, STRIPE_SIZE);
1612 		count = 0;
1613 		for (i = disks ; i--; ) {
1614 			if (i == dd_idx || i == qd_idx)
1615 				continue;
1616 			p = page_address(sh->dev[i].page);
1617 			if (test_bit(R5_UPTODATE, &sh->dev[i].flags))
1618 				ptr[count++] = p;
1619 			else
1620 				printk("compute_block() %d, stripe %llu, %d"
1621 				       " not present\n", dd_idx,
1622 				       (unsigned long long)sh->sector, i);
1623 
1624 			check_xor();
1625 		}
1626 		if (count)
1627 			xor_blocks(count, STRIPE_SIZE, dest, ptr);
1628 		if (!nozero) set_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1629 		else clear_bit(R5_UPTODATE, &sh->dev[dd_idx].flags);
1630 	}
1631 }
1632 
1633 /* Compute two missing blocks */
1634 static void compute_block_2(struct stripe_head *sh, int dd_idx1, int dd_idx2)
1635 {
1636 	int i, count, disks = sh->disks;
1637 	int pd_idx = sh->pd_idx;
1638 	int qd_idx = raid6_next_disk(pd_idx, disks);
1639 	int d0_idx = raid6_next_disk(qd_idx, disks);
1640 	int faila, failb;
1641 
1642 	/* faila and failb are disk numbers relative to d0_idx */
1643 	/* pd_idx become disks-2 and qd_idx become disks-1 */
1644 	faila = (dd_idx1 < d0_idx) ? dd_idx1+(disks-d0_idx) : dd_idx1-d0_idx;
1645 	failb = (dd_idx2 < d0_idx) ? dd_idx2+(disks-d0_idx) : dd_idx2-d0_idx;
1646 
1647 	BUG_ON(faila == failb);
1648 	if ( failb < faila ) { int tmp = faila; faila = failb; failb = tmp; }
1649 
1650 	pr_debug("compute_block_2, stripe %llu, idx %d,%d (%d,%d)\n",
1651 	       (unsigned long long)sh->sector, dd_idx1, dd_idx2, faila, failb);
1652 
1653 	if ( failb == disks-1 ) {
1654 		/* Q disk is one of the missing disks */
1655 		if ( faila == disks-2 ) {
1656 			/* Missing P+Q, just recompute */
1657 			compute_parity6(sh, UPDATE_PARITY);
1658 			return;
1659 		} else {
1660 			/* We're missing D+Q; recompute D from P */
1661 			compute_block_1(sh, (dd_idx1 == qd_idx) ? dd_idx2 : dd_idx1, 0);
1662 			compute_parity6(sh, UPDATE_PARITY); /* Is this necessary? */
1663 			return;
1664 		}
1665 	}
1666 
1667 	/* We're missing D+P or D+D; build pointer table */
1668 	{
1669 		/**** FIX THIS: This could be very bad if disks is close to 256 ****/
1670 		void *ptrs[disks];
1671 
1672 		count = 0;
1673 		i = d0_idx;
1674 		do {
1675 			ptrs[count++] = page_address(sh->dev[i].page);
1676 			i = raid6_next_disk(i, disks);
1677 			if (i != dd_idx1 && i != dd_idx2 &&
1678 			    !test_bit(R5_UPTODATE, &sh->dev[i].flags))
1679 				printk("compute_2 with missing block %d/%d\n", count, i);
1680 		} while ( i != d0_idx );
1681 
1682 		if ( failb == disks-2 ) {
1683 			/* We're missing D+P. */
1684 			raid6_datap_recov(disks, STRIPE_SIZE, faila, ptrs);
1685 		} else {
1686 			/* We're missing D+D. */
1687 			raid6_2data_recov(disks, STRIPE_SIZE, faila, failb, ptrs);
1688 		}
1689 
1690 		/* Both the above update both missing blocks */
1691 		set_bit(R5_UPTODATE, &sh->dev[dd_idx1].flags);
1692 		set_bit(R5_UPTODATE, &sh->dev[dd_idx2].flags);
1693 	}
1694 }
1695 
1696 static int
1697 handle_write_operations5(struct stripe_head *sh, int rcw, int expand)
1698 {
1699 	int i, pd_idx = sh->pd_idx, disks = sh->disks;
1700 	int locked = 0;
1701 
1702 	if (rcw) {
1703 		/* if we are not expanding this is a proper write request, and
1704 		 * there will be bios with new data to be drained into the
1705 		 * stripe cache
1706 		 */
1707 		if (!expand) {
1708 			set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1709 			sh->ops.count++;
1710 		}
1711 
1712 		set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1713 		sh->ops.count++;
1714 
1715 		for (i = disks; i--; ) {
1716 			struct r5dev *dev = &sh->dev[i];
1717 
1718 			if (dev->towrite) {
1719 				set_bit(R5_LOCKED, &dev->flags);
1720 				if (!expand)
1721 					clear_bit(R5_UPTODATE, &dev->flags);
1722 				locked++;
1723 			}
1724 		}
1725 		if (locked + 1 == disks)
1726 			if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
1727 				atomic_inc(&sh->raid_conf->pending_full_writes);
1728 	} else {
1729 		BUG_ON(!(test_bit(R5_UPTODATE, &sh->dev[pd_idx].flags) ||
1730 			test_bit(R5_Wantcompute, &sh->dev[pd_idx].flags)));
1731 
1732 		set_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
1733 		set_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
1734 		set_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
1735 
1736 		sh->ops.count += 3;
1737 
1738 		for (i = disks; i--; ) {
1739 			struct r5dev *dev = &sh->dev[i];
1740 			if (i == pd_idx)
1741 				continue;
1742 
1743 			/* For a read-modify write there may be blocks that are
1744 			 * locked for reading while others are ready to be
1745 			 * written so we distinguish these blocks by the
1746 			 * R5_Wantprexor bit
1747 			 */
1748 			if (dev->towrite &&
1749 			    (test_bit(R5_UPTODATE, &dev->flags) ||
1750 			    test_bit(R5_Wantcompute, &dev->flags))) {
1751 				set_bit(R5_Wantprexor, &dev->flags);
1752 				set_bit(R5_LOCKED, &dev->flags);
1753 				clear_bit(R5_UPTODATE, &dev->flags);
1754 				locked++;
1755 			}
1756 		}
1757 	}
1758 
1759 	/* keep the parity disk locked while asynchronous operations
1760 	 * are in flight
1761 	 */
1762 	set_bit(R5_LOCKED, &sh->dev[pd_idx].flags);
1763 	clear_bit(R5_UPTODATE, &sh->dev[pd_idx].flags);
1764 	locked++;
1765 
1766 	pr_debug("%s: stripe %llu locked: %d pending: %lx\n",
1767 		__func__, (unsigned long long)sh->sector,
1768 		locked, sh->ops.pending);
1769 
1770 	return locked;
1771 }
1772 
1773 /*
1774  * Each stripe/dev can have one or more bion attached.
1775  * toread/towrite point to the first in a chain.
1776  * The bi_next chain must be in order.
1777  */
1778 static int add_stripe_bio(struct stripe_head *sh, struct bio *bi, int dd_idx, int forwrite)
1779 {
1780 	struct bio **bip;
1781 	raid5_conf_t *conf = sh->raid_conf;
1782 	int firstwrite=0;
1783 
1784 	pr_debug("adding bh b#%llu to stripe s#%llu\n",
1785 		(unsigned long long)bi->bi_sector,
1786 		(unsigned long long)sh->sector);
1787 
1788 
1789 	spin_lock(&sh->lock);
1790 	spin_lock_irq(&conf->device_lock);
1791 	if (forwrite) {
1792 		bip = &sh->dev[dd_idx].towrite;
1793 		if (*bip == NULL && sh->dev[dd_idx].written == NULL)
1794 			firstwrite = 1;
1795 	} else
1796 		bip = &sh->dev[dd_idx].toread;
1797 	while (*bip && (*bip)->bi_sector < bi->bi_sector) {
1798 		if ((*bip)->bi_sector + ((*bip)->bi_size >> 9) > bi->bi_sector)
1799 			goto overlap;
1800 		bip = & (*bip)->bi_next;
1801 	}
1802 	if (*bip && (*bip)->bi_sector < bi->bi_sector + ((bi->bi_size)>>9))
1803 		goto overlap;
1804 
1805 	BUG_ON(*bip && bi->bi_next && (*bip) != bi->bi_next);
1806 	if (*bip)
1807 		bi->bi_next = *bip;
1808 	*bip = bi;
1809 	bi->bi_phys_segments ++;
1810 	spin_unlock_irq(&conf->device_lock);
1811 	spin_unlock(&sh->lock);
1812 
1813 	pr_debug("added bi b#%llu to stripe s#%llu, disk %d.\n",
1814 		(unsigned long long)bi->bi_sector,
1815 		(unsigned long long)sh->sector, dd_idx);
1816 
1817 	if (conf->mddev->bitmap && firstwrite) {
1818 		bitmap_startwrite(conf->mddev->bitmap, sh->sector,
1819 				  STRIPE_SECTORS, 0);
1820 		sh->bm_seq = conf->seq_flush+1;
1821 		set_bit(STRIPE_BIT_DELAY, &sh->state);
1822 	}
1823 
1824 	if (forwrite) {
1825 		/* check if page is covered */
1826 		sector_t sector = sh->dev[dd_idx].sector;
1827 		for (bi=sh->dev[dd_idx].towrite;
1828 		     sector < sh->dev[dd_idx].sector + STRIPE_SECTORS &&
1829 			     bi && bi->bi_sector <= sector;
1830 		     bi = r5_next_bio(bi, sh->dev[dd_idx].sector)) {
1831 			if (bi->bi_sector + (bi->bi_size>>9) >= sector)
1832 				sector = bi->bi_sector + (bi->bi_size>>9);
1833 		}
1834 		if (sector >= sh->dev[dd_idx].sector + STRIPE_SECTORS)
1835 			set_bit(R5_OVERWRITE, &sh->dev[dd_idx].flags);
1836 	}
1837 	return 1;
1838 
1839  overlap:
1840 	set_bit(R5_Overlap, &sh->dev[dd_idx].flags);
1841 	spin_unlock_irq(&conf->device_lock);
1842 	spin_unlock(&sh->lock);
1843 	return 0;
1844 }
1845 
1846 static void end_reshape(raid5_conf_t *conf);
1847 
1848 static int page_is_zero(struct page *p)
1849 {
1850 	char *a = page_address(p);
1851 	return ((*(u32*)a) == 0 &&
1852 		memcmp(a, a+4, STRIPE_SIZE-4)==0);
1853 }
1854 
1855 static int stripe_to_pdidx(sector_t stripe, raid5_conf_t *conf, int disks)
1856 {
1857 	int sectors_per_chunk = conf->chunk_size >> 9;
1858 	int pd_idx, dd_idx;
1859 	int chunk_offset = sector_div(stripe, sectors_per_chunk);
1860 
1861 	raid5_compute_sector(stripe * (disks - conf->max_degraded)
1862 			     *sectors_per_chunk + chunk_offset,
1863 			     disks, disks - conf->max_degraded,
1864 			     &dd_idx, &pd_idx, conf);
1865 	return pd_idx;
1866 }
1867 
1868 static void
1869 handle_requests_to_failed_array(raid5_conf_t *conf, struct stripe_head *sh,
1870 				struct stripe_head_state *s, int disks,
1871 				struct bio **return_bi)
1872 {
1873 	int i;
1874 	for (i = disks; i--; ) {
1875 		struct bio *bi;
1876 		int bitmap_end = 0;
1877 
1878 		if (test_bit(R5_ReadError, &sh->dev[i].flags)) {
1879 			mdk_rdev_t *rdev;
1880 			rcu_read_lock();
1881 			rdev = rcu_dereference(conf->disks[i].rdev);
1882 			if (rdev && test_bit(In_sync, &rdev->flags))
1883 				/* multiple read failures in one stripe */
1884 				md_error(conf->mddev, rdev);
1885 			rcu_read_unlock();
1886 		}
1887 		spin_lock_irq(&conf->device_lock);
1888 		/* fail all writes first */
1889 		bi = sh->dev[i].towrite;
1890 		sh->dev[i].towrite = NULL;
1891 		if (bi) {
1892 			s->to_write--;
1893 			bitmap_end = 1;
1894 		}
1895 
1896 		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1897 			wake_up(&conf->wait_for_overlap);
1898 
1899 		while (bi && bi->bi_sector <
1900 			sh->dev[i].sector + STRIPE_SECTORS) {
1901 			struct bio *nextbi = r5_next_bio(bi, sh->dev[i].sector);
1902 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
1903 			if (--bi->bi_phys_segments == 0) {
1904 				md_write_end(conf->mddev);
1905 				bi->bi_next = *return_bi;
1906 				*return_bi = bi;
1907 			}
1908 			bi = nextbi;
1909 		}
1910 		/* and fail all 'written' */
1911 		bi = sh->dev[i].written;
1912 		sh->dev[i].written = NULL;
1913 		if (bi) bitmap_end = 1;
1914 		while (bi && bi->bi_sector <
1915 		       sh->dev[i].sector + STRIPE_SECTORS) {
1916 			struct bio *bi2 = r5_next_bio(bi, sh->dev[i].sector);
1917 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
1918 			if (--bi->bi_phys_segments == 0) {
1919 				md_write_end(conf->mddev);
1920 				bi->bi_next = *return_bi;
1921 				*return_bi = bi;
1922 			}
1923 			bi = bi2;
1924 		}
1925 
1926 		/* fail any reads if this device is non-operational and
1927 		 * the data has not reached the cache yet.
1928 		 */
1929 		if (!test_bit(R5_Wantfill, &sh->dev[i].flags) &&
1930 		    (!test_bit(R5_Insync, &sh->dev[i].flags) ||
1931 		      test_bit(R5_ReadError, &sh->dev[i].flags))) {
1932 			bi = sh->dev[i].toread;
1933 			sh->dev[i].toread = NULL;
1934 			if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
1935 				wake_up(&conf->wait_for_overlap);
1936 			if (bi) s->to_read--;
1937 			while (bi && bi->bi_sector <
1938 			       sh->dev[i].sector + STRIPE_SECTORS) {
1939 				struct bio *nextbi =
1940 					r5_next_bio(bi, sh->dev[i].sector);
1941 				clear_bit(BIO_UPTODATE, &bi->bi_flags);
1942 				if (--bi->bi_phys_segments == 0) {
1943 					bi->bi_next = *return_bi;
1944 					*return_bi = bi;
1945 				}
1946 				bi = nextbi;
1947 			}
1948 		}
1949 		spin_unlock_irq(&conf->device_lock);
1950 		if (bitmap_end)
1951 			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
1952 					STRIPE_SECTORS, 0, 0);
1953 	}
1954 
1955 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
1956 		if (atomic_dec_and_test(&conf->pending_full_writes))
1957 			md_wakeup_thread(conf->mddev->thread);
1958 }
1959 
1960 /* __handle_issuing_new_read_requests5 - returns 0 if there are no more disks
1961  * to process
1962  */
1963 static int __handle_issuing_new_read_requests5(struct stripe_head *sh,
1964 			struct stripe_head_state *s, int disk_idx, int disks)
1965 {
1966 	struct r5dev *dev = &sh->dev[disk_idx];
1967 	struct r5dev *failed_dev = &sh->dev[s->failed_num];
1968 
1969 	/* don't schedule compute operations or reads on the parity block while
1970 	 * a check is in flight
1971 	 */
1972 	if ((disk_idx == sh->pd_idx) &&
1973 	     test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
1974 		return ~0;
1975 
1976 	/* is the data in this block needed, and can we get it? */
1977 	if (!test_bit(R5_LOCKED, &dev->flags) &&
1978 	    !test_bit(R5_UPTODATE, &dev->flags) && (dev->toread ||
1979 	    (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags)) ||
1980 	     s->syncing || s->expanding || (s->failed &&
1981 	     (failed_dev->toread || (failed_dev->towrite &&
1982 	     !test_bit(R5_OVERWRITE, &failed_dev->flags)
1983 	     ))))) {
1984 		/* 1/ We would like to get this block, possibly by computing it,
1985 		 * but we might not be able to.
1986 		 *
1987 		 * 2/ Since parity check operations potentially make the parity
1988 		 * block !uptodate it will need to be refreshed before any
1989 		 * compute operations on data disks are scheduled.
1990 		 *
1991 		 * 3/ We hold off parity block re-reads until check operations
1992 		 * have quiesced.
1993 		 */
1994 		if ((s->uptodate == disks - 1) &&
1995 		    !test_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
1996 			set_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
1997 			set_bit(R5_Wantcompute, &dev->flags);
1998 			sh->ops.target = disk_idx;
1999 			s->req_compute = 1;
2000 			sh->ops.count++;
2001 			/* Careful: from this point on 'uptodate' is in the eye
2002 			 * of raid5_run_ops which services 'compute' operations
2003 			 * before writes. R5_Wantcompute flags a block that will
2004 			 * be R5_UPTODATE by the time it is needed for a
2005 			 * subsequent operation.
2006 			 */
2007 			s->uptodate++;
2008 			return 0; /* uptodate + compute == disks */
2009 		} else if ((s->uptodate < disks - 1) &&
2010 			test_bit(R5_Insync, &dev->flags)) {
2011 			/* Note: we hold off compute operations while checks are
2012 			 * in flight, but we still prefer 'compute' over 'read'
2013 			 * hence we only read if (uptodate < * disks-1)
2014 			 */
2015 			set_bit(R5_LOCKED, &dev->flags);
2016 			set_bit(R5_Wantread, &dev->flags);
2017 			if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2018 				sh->ops.count++;
2019 			s->locked++;
2020 			pr_debug("Reading block %d (sync=%d)\n", disk_idx,
2021 				s->syncing);
2022 		}
2023 	}
2024 
2025 	return ~0;
2026 }
2027 
2028 static void handle_issuing_new_read_requests5(struct stripe_head *sh,
2029 			struct stripe_head_state *s, int disks)
2030 {
2031 	int i;
2032 
2033 	/* Clear completed compute operations.  Parity recovery
2034 	 * (STRIPE_OP_MOD_REPAIR_PD) implies a write-back which is handled
2035 	 * later on in this routine
2036 	 */
2037 	if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2038 		!test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2039 		clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2040 		clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2041 		clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2042 	}
2043 
2044 	/* look for blocks to read/compute, skip this if a compute
2045 	 * is already in flight, or if the stripe contents are in the
2046 	 * midst of changing due to a write
2047 	 */
2048 	if (!test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2049 		!test_bit(STRIPE_OP_PREXOR, &sh->ops.pending) &&
2050 		!test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2051 		for (i = disks; i--; )
2052 			if (__handle_issuing_new_read_requests5(
2053 				sh, s, i, disks) == 0)
2054 				break;
2055 	}
2056 	set_bit(STRIPE_HANDLE, &sh->state);
2057 }
2058 
2059 static void handle_issuing_new_read_requests6(struct stripe_head *sh,
2060 			struct stripe_head_state *s, struct r6_state *r6s,
2061 			int disks)
2062 {
2063 	int i;
2064 	for (i = disks; i--; ) {
2065 		struct r5dev *dev = &sh->dev[i];
2066 		if (!test_bit(R5_LOCKED, &dev->flags) &&
2067 		    !test_bit(R5_UPTODATE, &dev->flags) &&
2068 		    (dev->toread || (dev->towrite &&
2069 		     !test_bit(R5_OVERWRITE, &dev->flags)) ||
2070 		     s->syncing || s->expanding ||
2071 		     (s->failed >= 1 &&
2072 		      (sh->dev[r6s->failed_num[0]].toread ||
2073 		       s->to_write)) ||
2074 		     (s->failed >= 2 &&
2075 		      (sh->dev[r6s->failed_num[1]].toread ||
2076 		       s->to_write)))) {
2077 			/* we would like to get this block, possibly
2078 			 * by computing it, but we might not be able to
2079 			 */
2080 			if (s->uptodate == disks-1) {
2081 				pr_debug("Computing stripe %llu block %d\n",
2082 				       (unsigned long long)sh->sector, i);
2083 				compute_block_1(sh, i, 0);
2084 				s->uptodate++;
2085 			} else if ( s->uptodate == disks-2 && s->failed >= 2 ) {
2086 				/* Computing 2-failure is *very* expensive; only
2087 				 * do it if failed >= 2
2088 				 */
2089 				int other;
2090 				for (other = disks; other--; ) {
2091 					if (other == i)
2092 						continue;
2093 					if (!test_bit(R5_UPTODATE,
2094 					      &sh->dev[other].flags))
2095 						break;
2096 				}
2097 				BUG_ON(other < 0);
2098 				pr_debug("Computing stripe %llu blocks %d,%d\n",
2099 				       (unsigned long long)sh->sector,
2100 				       i, other);
2101 				compute_block_2(sh, i, other);
2102 				s->uptodate += 2;
2103 			} else if (test_bit(R5_Insync, &dev->flags)) {
2104 				set_bit(R5_LOCKED, &dev->flags);
2105 				set_bit(R5_Wantread, &dev->flags);
2106 				s->locked++;
2107 				pr_debug("Reading block %d (sync=%d)\n",
2108 					i, s->syncing);
2109 			}
2110 		}
2111 	}
2112 	set_bit(STRIPE_HANDLE, &sh->state);
2113 }
2114 
2115 
2116 /* handle_completed_write_requests
2117  * any written block on an uptodate or failed drive can be returned.
2118  * Note that if we 'wrote' to a failed drive, it will be UPTODATE, but
2119  * never LOCKED, so we don't need to test 'failed' directly.
2120  */
2121 static void handle_completed_write_requests(raid5_conf_t *conf,
2122 	struct stripe_head *sh, int disks, struct bio **return_bi)
2123 {
2124 	int i;
2125 	struct r5dev *dev;
2126 
2127 	for (i = disks; i--; )
2128 		if (sh->dev[i].written) {
2129 			dev = &sh->dev[i];
2130 			if (!test_bit(R5_LOCKED, &dev->flags) &&
2131 				test_bit(R5_UPTODATE, &dev->flags)) {
2132 				/* We can return any write requests */
2133 				struct bio *wbi, *wbi2;
2134 				int bitmap_end = 0;
2135 				pr_debug("Return write for disc %d\n", i);
2136 				spin_lock_irq(&conf->device_lock);
2137 				wbi = dev->written;
2138 				dev->written = NULL;
2139 				while (wbi && wbi->bi_sector <
2140 					dev->sector + STRIPE_SECTORS) {
2141 					wbi2 = r5_next_bio(wbi, dev->sector);
2142 					if (--wbi->bi_phys_segments == 0) {
2143 						md_write_end(conf->mddev);
2144 						wbi->bi_next = *return_bi;
2145 						*return_bi = wbi;
2146 					}
2147 					wbi = wbi2;
2148 				}
2149 				if (dev->towrite == NULL)
2150 					bitmap_end = 1;
2151 				spin_unlock_irq(&conf->device_lock);
2152 				if (bitmap_end)
2153 					bitmap_endwrite(conf->mddev->bitmap,
2154 							sh->sector,
2155 							STRIPE_SECTORS,
2156 					 !test_bit(STRIPE_DEGRADED, &sh->state),
2157 							0);
2158 			}
2159 		}
2160 
2161 	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2162 		if (atomic_dec_and_test(&conf->pending_full_writes))
2163 			md_wakeup_thread(conf->mddev->thread);
2164 }
2165 
2166 static void handle_issuing_new_write_requests5(raid5_conf_t *conf,
2167 		struct stripe_head *sh,	struct stripe_head_state *s, int disks)
2168 {
2169 	int rmw = 0, rcw = 0, i;
2170 	for (i = disks; i--; ) {
2171 		/* would I have to read this buffer for read_modify_write */
2172 		struct r5dev *dev = &sh->dev[i];
2173 		if ((dev->towrite || i == sh->pd_idx) &&
2174 		    !test_bit(R5_LOCKED, &dev->flags) &&
2175 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2176 		      test_bit(R5_Wantcompute, &dev->flags))) {
2177 			if (test_bit(R5_Insync, &dev->flags))
2178 				rmw++;
2179 			else
2180 				rmw += 2*disks;  /* cannot read it */
2181 		}
2182 		/* Would I have to read this buffer for reconstruct_write */
2183 		if (!test_bit(R5_OVERWRITE, &dev->flags) && i != sh->pd_idx &&
2184 		    !test_bit(R5_LOCKED, &dev->flags) &&
2185 		    !(test_bit(R5_UPTODATE, &dev->flags) ||
2186 		    test_bit(R5_Wantcompute, &dev->flags))) {
2187 			if (test_bit(R5_Insync, &dev->flags)) rcw++;
2188 			else
2189 				rcw += 2*disks;
2190 		}
2191 	}
2192 	pr_debug("for sector %llu, rmw=%d rcw=%d\n",
2193 		(unsigned long long)sh->sector, rmw, rcw);
2194 	set_bit(STRIPE_HANDLE, &sh->state);
2195 	if (rmw < rcw && rmw > 0)
2196 		/* prefer read-modify-write, but need to get some data */
2197 		for (i = disks; i--; ) {
2198 			struct r5dev *dev = &sh->dev[i];
2199 			if ((dev->towrite || i == sh->pd_idx) &&
2200 			    !test_bit(R5_LOCKED, &dev->flags) &&
2201 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
2202 			    test_bit(R5_Wantcompute, &dev->flags)) &&
2203 			    test_bit(R5_Insync, &dev->flags)) {
2204 				if (
2205 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2206 					pr_debug("Read_old block "
2207 						"%d for r-m-w\n", i);
2208 					set_bit(R5_LOCKED, &dev->flags);
2209 					set_bit(R5_Wantread, &dev->flags);
2210 					if (!test_and_set_bit(
2211 						STRIPE_OP_IO, &sh->ops.pending))
2212 						sh->ops.count++;
2213 					s->locked++;
2214 				} else {
2215 					set_bit(STRIPE_DELAYED, &sh->state);
2216 					set_bit(STRIPE_HANDLE, &sh->state);
2217 				}
2218 			}
2219 		}
2220 	if (rcw <= rmw && rcw > 0)
2221 		/* want reconstruct write, but need to get some data */
2222 		for (i = disks; i--; ) {
2223 			struct r5dev *dev = &sh->dev[i];
2224 			if (!test_bit(R5_OVERWRITE, &dev->flags) &&
2225 			    i != sh->pd_idx &&
2226 			    !test_bit(R5_LOCKED, &dev->flags) &&
2227 			    !(test_bit(R5_UPTODATE, &dev->flags) ||
2228 			    test_bit(R5_Wantcompute, &dev->flags)) &&
2229 			    test_bit(R5_Insync, &dev->flags)) {
2230 				if (
2231 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2232 					pr_debug("Read_old block "
2233 						"%d for Reconstruct\n", i);
2234 					set_bit(R5_LOCKED, &dev->flags);
2235 					set_bit(R5_Wantread, &dev->flags);
2236 					if (!test_and_set_bit(
2237 						STRIPE_OP_IO, &sh->ops.pending))
2238 						sh->ops.count++;
2239 					s->locked++;
2240 				} else {
2241 					set_bit(STRIPE_DELAYED, &sh->state);
2242 					set_bit(STRIPE_HANDLE, &sh->state);
2243 				}
2244 			}
2245 		}
2246 	/* now if nothing is locked, and if we have enough data,
2247 	 * we can start a write request
2248 	 */
2249 	/* since handle_stripe can be called at any time we need to handle the
2250 	 * case where a compute block operation has been submitted and then a
2251 	 * subsequent call wants to start a write request.  raid5_run_ops only
2252 	 * handles the case where compute block and postxor are requested
2253 	 * simultaneously.  If this is not the case then new writes need to be
2254 	 * held off until the compute completes.
2255 	 */
2256 	if ((s->req_compute ||
2257 	    !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) &&
2258 		(s->locked == 0 && (rcw == 0 || rmw == 0) &&
2259 		!test_bit(STRIPE_BIT_DELAY, &sh->state)))
2260 		s->locked += handle_write_operations5(sh, rcw == 0, 0);
2261 }
2262 
2263 static void handle_issuing_new_write_requests6(raid5_conf_t *conf,
2264 		struct stripe_head *sh,	struct stripe_head_state *s,
2265 		struct r6_state *r6s, int disks)
2266 {
2267 	int rcw = 0, must_compute = 0, pd_idx = sh->pd_idx, i;
2268 	int qd_idx = r6s->qd_idx;
2269 	for (i = disks; i--; ) {
2270 		struct r5dev *dev = &sh->dev[i];
2271 		/* Would I have to read this buffer for reconstruct_write */
2272 		if (!test_bit(R5_OVERWRITE, &dev->flags)
2273 		    && i != pd_idx && i != qd_idx
2274 		    && (!test_bit(R5_LOCKED, &dev->flags)
2275 			    ) &&
2276 		    !test_bit(R5_UPTODATE, &dev->flags)) {
2277 			if (test_bit(R5_Insync, &dev->flags)) rcw++;
2278 			else {
2279 				pr_debug("raid6: must_compute: "
2280 					"disk %d flags=%#lx\n", i, dev->flags);
2281 				must_compute++;
2282 			}
2283 		}
2284 	}
2285 	pr_debug("for sector %llu, rcw=%d, must_compute=%d\n",
2286 	       (unsigned long long)sh->sector, rcw, must_compute);
2287 	set_bit(STRIPE_HANDLE, &sh->state);
2288 
2289 	if (rcw > 0)
2290 		/* want reconstruct write, but need to get some data */
2291 		for (i = disks; i--; ) {
2292 			struct r5dev *dev = &sh->dev[i];
2293 			if (!test_bit(R5_OVERWRITE, &dev->flags)
2294 			    && !(s->failed == 0 && (i == pd_idx || i == qd_idx))
2295 			    && !test_bit(R5_LOCKED, &dev->flags) &&
2296 			    !test_bit(R5_UPTODATE, &dev->flags) &&
2297 			    test_bit(R5_Insync, &dev->flags)) {
2298 				if (
2299 				  test_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2300 					pr_debug("Read_old stripe %llu "
2301 						"block %d for Reconstruct\n",
2302 					     (unsigned long long)sh->sector, i);
2303 					set_bit(R5_LOCKED, &dev->flags);
2304 					set_bit(R5_Wantread, &dev->flags);
2305 					s->locked++;
2306 				} else {
2307 					pr_debug("Request delayed stripe %llu "
2308 						"block %d for Reconstruct\n",
2309 					     (unsigned long long)sh->sector, i);
2310 					set_bit(STRIPE_DELAYED, &sh->state);
2311 					set_bit(STRIPE_HANDLE, &sh->state);
2312 				}
2313 			}
2314 		}
2315 	/* now if nothing is locked, and if we have enough data, we can start a
2316 	 * write request
2317 	 */
2318 	if (s->locked == 0 && rcw == 0 &&
2319 	    !test_bit(STRIPE_BIT_DELAY, &sh->state)) {
2320 		if (must_compute > 0) {
2321 			/* We have failed blocks and need to compute them */
2322 			switch (s->failed) {
2323 			case 0:
2324 				BUG();
2325 			case 1:
2326 				compute_block_1(sh, r6s->failed_num[0], 0);
2327 				break;
2328 			case 2:
2329 				compute_block_2(sh, r6s->failed_num[0],
2330 						r6s->failed_num[1]);
2331 				break;
2332 			default: /* This request should have been failed? */
2333 				BUG();
2334 			}
2335 		}
2336 
2337 		pr_debug("Computing parity for stripe %llu\n",
2338 			(unsigned long long)sh->sector);
2339 		compute_parity6(sh, RECONSTRUCT_WRITE);
2340 		/* now every locked buffer is ready to be written */
2341 		for (i = disks; i--; )
2342 			if (test_bit(R5_LOCKED, &sh->dev[i].flags)) {
2343 				pr_debug("Writing stripe %llu block %d\n",
2344 				       (unsigned long long)sh->sector, i);
2345 				s->locked++;
2346 				set_bit(R5_Wantwrite, &sh->dev[i].flags);
2347 			}
2348 		if (s->locked == disks)
2349 			if (!test_and_set_bit(STRIPE_FULL_WRITE, &sh->state))
2350 				atomic_inc(&conf->pending_full_writes);
2351 		/* after a RECONSTRUCT_WRITE, the stripe MUST be in-sync */
2352 		set_bit(STRIPE_INSYNC, &sh->state);
2353 
2354 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2355 			atomic_dec(&conf->preread_active_stripes);
2356 			if (atomic_read(&conf->preread_active_stripes) <
2357 			    IO_THRESHOLD)
2358 				md_wakeup_thread(conf->mddev->thread);
2359 		}
2360 	}
2361 }
2362 
2363 static void handle_parity_checks5(raid5_conf_t *conf, struct stripe_head *sh,
2364 				struct stripe_head_state *s, int disks)
2365 {
2366 	int canceled_check = 0;
2367 
2368 	set_bit(STRIPE_HANDLE, &sh->state);
2369 
2370 	/* complete a check operation */
2371 	if (test_and_clear_bit(STRIPE_OP_CHECK, &sh->ops.complete)) {
2372 	    clear_bit(STRIPE_OP_CHECK, &sh->ops.ack);
2373 	    clear_bit(STRIPE_OP_CHECK, &sh->ops.pending);
2374 		if (s->failed == 0) {
2375 			if (sh->ops.zero_sum_result == 0)
2376 				/* parity is correct (on disc,
2377 				 * not in buffer any more)
2378 				 */
2379 				set_bit(STRIPE_INSYNC, &sh->state);
2380 			else {
2381 				conf->mddev->resync_mismatches +=
2382 					STRIPE_SECTORS;
2383 				if (test_bit(
2384 				     MD_RECOVERY_CHECK, &conf->mddev->recovery))
2385 					/* don't try to repair!! */
2386 					set_bit(STRIPE_INSYNC, &sh->state);
2387 				else {
2388 					set_bit(STRIPE_OP_COMPUTE_BLK,
2389 						&sh->ops.pending);
2390 					set_bit(STRIPE_OP_MOD_REPAIR_PD,
2391 						&sh->ops.pending);
2392 					set_bit(R5_Wantcompute,
2393 						&sh->dev[sh->pd_idx].flags);
2394 					sh->ops.target = sh->pd_idx;
2395 					sh->ops.count++;
2396 					s->uptodate++;
2397 				}
2398 			}
2399 		} else
2400 			canceled_check = 1; /* STRIPE_INSYNC is not set */
2401 	}
2402 
2403 	/* check if we can clear a parity disk reconstruct */
2404 	if (test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete) &&
2405 		test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2406 
2407 		clear_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending);
2408 		clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.complete);
2409 		clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.ack);
2410 		clear_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending);
2411 	}
2412 
2413 	/* start a new check operation if there are no failures, the stripe is
2414 	 * not insync, and a repair is not in flight
2415 	 */
2416 	if (s->failed == 0 &&
2417 	    !test_bit(STRIPE_INSYNC, &sh->state) &&
2418 	    !test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending)) {
2419 		if (!test_and_set_bit(STRIPE_OP_CHECK, &sh->ops.pending)) {
2420 			BUG_ON(s->uptodate != disks);
2421 			clear_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags);
2422 			sh->ops.count++;
2423 			s->uptodate--;
2424 		}
2425 	}
2426 
2427 	/* Wait for check parity and compute block operations to complete
2428 	 * before write-back.  If a failure occurred while the check operation
2429 	 * was in flight we need to cycle this stripe through handle_stripe
2430 	 * since the parity block may not be uptodate
2431 	 */
2432 	if (!canceled_check && !test_bit(STRIPE_INSYNC, &sh->state) &&
2433 	    !test_bit(STRIPE_OP_CHECK, &sh->ops.pending) &&
2434 	    !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending)) {
2435 		struct r5dev *dev;
2436 		/* either failed parity check, or recovery is happening */
2437 		if (s->failed == 0)
2438 			s->failed_num = sh->pd_idx;
2439 		dev = &sh->dev[s->failed_num];
2440 		BUG_ON(!test_bit(R5_UPTODATE, &dev->flags));
2441 		BUG_ON(s->uptodate != disks);
2442 
2443 		set_bit(R5_LOCKED, &dev->flags);
2444 		set_bit(R5_Wantwrite, &dev->flags);
2445 		if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2446 			sh->ops.count++;
2447 
2448 		clear_bit(STRIPE_DEGRADED, &sh->state);
2449 		s->locked++;
2450 		set_bit(STRIPE_INSYNC, &sh->state);
2451 	}
2452 }
2453 
2454 
2455 static void handle_parity_checks6(raid5_conf_t *conf, struct stripe_head *sh,
2456 				struct stripe_head_state *s,
2457 				struct r6_state *r6s, struct page *tmp_page,
2458 				int disks)
2459 {
2460 	int update_p = 0, update_q = 0;
2461 	struct r5dev *dev;
2462 	int pd_idx = sh->pd_idx;
2463 	int qd_idx = r6s->qd_idx;
2464 
2465 	set_bit(STRIPE_HANDLE, &sh->state);
2466 
2467 	BUG_ON(s->failed > 2);
2468 	BUG_ON(s->uptodate < disks);
2469 	/* Want to check and possibly repair P and Q.
2470 	 * However there could be one 'failed' device, in which
2471 	 * case we can only check one of them, possibly using the
2472 	 * other to generate missing data
2473 	 */
2474 
2475 	/* If !tmp_page, we cannot do the calculations,
2476 	 * but as we have set STRIPE_HANDLE, we will soon be called
2477 	 * by stripe_handle with a tmp_page - just wait until then.
2478 	 */
2479 	if (tmp_page) {
2480 		if (s->failed == r6s->q_failed) {
2481 			/* The only possible failed device holds 'Q', so it
2482 			 * makes sense to check P (If anything else were failed,
2483 			 * we would have used P to recreate it).
2484 			 */
2485 			compute_block_1(sh, pd_idx, 1);
2486 			if (!page_is_zero(sh->dev[pd_idx].page)) {
2487 				compute_block_1(sh, pd_idx, 0);
2488 				update_p = 1;
2489 			}
2490 		}
2491 		if (!r6s->q_failed && s->failed < 2) {
2492 			/* q is not failed, and we didn't use it to generate
2493 			 * anything, so it makes sense to check it
2494 			 */
2495 			memcpy(page_address(tmp_page),
2496 			       page_address(sh->dev[qd_idx].page),
2497 			       STRIPE_SIZE);
2498 			compute_parity6(sh, UPDATE_PARITY);
2499 			if (memcmp(page_address(tmp_page),
2500 				   page_address(sh->dev[qd_idx].page),
2501 				   STRIPE_SIZE) != 0) {
2502 				clear_bit(STRIPE_INSYNC, &sh->state);
2503 				update_q = 1;
2504 			}
2505 		}
2506 		if (update_p || update_q) {
2507 			conf->mddev->resync_mismatches += STRIPE_SECTORS;
2508 			if (test_bit(MD_RECOVERY_CHECK, &conf->mddev->recovery))
2509 				/* don't try to repair!! */
2510 				update_p = update_q = 0;
2511 		}
2512 
2513 		/* now write out any block on a failed drive,
2514 		 * or P or Q if they need it
2515 		 */
2516 
2517 		if (s->failed == 2) {
2518 			dev = &sh->dev[r6s->failed_num[1]];
2519 			s->locked++;
2520 			set_bit(R5_LOCKED, &dev->flags);
2521 			set_bit(R5_Wantwrite, &dev->flags);
2522 		}
2523 		if (s->failed >= 1) {
2524 			dev = &sh->dev[r6s->failed_num[0]];
2525 			s->locked++;
2526 			set_bit(R5_LOCKED, &dev->flags);
2527 			set_bit(R5_Wantwrite, &dev->flags);
2528 		}
2529 
2530 		if (update_p) {
2531 			dev = &sh->dev[pd_idx];
2532 			s->locked++;
2533 			set_bit(R5_LOCKED, &dev->flags);
2534 			set_bit(R5_Wantwrite, &dev->flags);
2535 		}
2536 		if (update_q) {
2537 			dev = &sh->dev[qd_idx];
2538 			s->locked++;
2539 			set_bit(R5_LOCKED, &dev->flags);
2540 			set_bit(R5_Wantwrite, &dev->flags);
2541 		}
2542 		clear_bit(STRIPE_DEGRADED, &sh->state);
2543 
2544 		set_bit(STRIPE_INSYNC, &sh->state);
2545 	}
2546 }
2547 
2548 static void handle_stripe_expansion(raid5_conf_t *conf, struct stripe_head *sh,
2549 				struct r6_state *r6s)
2550 {
2551 	int i;
2552 
2553 	/* We have read all the blocks in this stripe and now we need to
2554 	 * copy some of them into a target stripe for expand.
2555 	 */
2556 	struct dma_async_tx_descriptor *tx = NULL;
2557 	clear_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2558 	for (i = 0; i < sh->disks; i++)
2559 		if (i != sh->pd_idx && (!r6s || i != r6s->qd_idx)) {
2560 			int dd_idx, pd_idx, j;
2561 			struct stripe_head *sh2;
2562 
2563 			sector_t bn = compute_blocknr(sh, i);
2564 			sector_t s = raid5_compute_sector(bn, conf->raid_disks,
2565 						conf->raid_disks -
2566 						conf->max_degraded, &dd_idx,
2567 						&pd_idx, conf);
2568 			sh2 = get_active_stripe(conf, s, conf->raid_disks,
2569 						pd_idx, 1);
2570 			if (sh2 == NULL)
2571 				/* so far only the early blocks of this stripe
2572 				 * have been requested.  When later blocks
2573 				 * get requested, we will try again
2574 				 */
2575 				continue;
2576 			if (!test_bit(STRIPE_EXPANDING, &sh2->state) ||
2577 			   test_bit(R5_Expanded, &sh2->dev[dd_idx].flags)) {
2578 				/* must have already done this block */
2579 				release_stripe(sh2);
2580 				continue;
2581 			}
2582 
2583 			/* place all the copies on one channel */
2584 			tx = async_memcpy(sh2->dev[dd_idx].page,
2585 				sh->dev[i].page, 0, 0, STRIPE_SIZE,
2586 				ASYNC_TX_DEP_ACK, tx, NULL, NULL);
2587 
2588 			set_bit(R5_Expanded, &sh2->dev[dd_idx].flags);
2589 			set_bit(R5_UPTODATE, &sh2->dev[dd_idx].flags);
2590 			for (j = 0; j < conf->raid_disks; j++)
2591 				if (j != sh2->pd_idx &&
2592 				    (!r6s || j != raid6_next_disk(sh2->pd_idx,
2593 								 sh2->disks)) &&
2594 				    !test_bit(R5_Expanded, &sh2->dev[j].flags))
2595 					break;
2596 			if (j == conf->raid_disks) {
2597 				set_bit(STRIPE_EXPAND_READY, &sh2->state);
2598 				set_bit(STRIPE_HANDLE, &sh2->state);
2599 			}
2600 			release_stripe(sh2);
2601 
2602 		}
2603 	/* done submitting copies, wait for them to complete */
2604 	if (tx) {
2605 		async_tx_ack(tx);
2606 		dma_wait_for_async_tx(tx);
2607 	}
2608 }
2609 
2610 
2611 /*
2612  * handle_stripe - do things to a stripe.
2613  *
2614  * We lock the stripe and then examine the state of various bits
2615  * to see what needs to be done.
2616  * Possible results:
2617  *    return some read request which now have data
2618  *    return some write requests which are safely on disc
2619  *    schedule a read on some buffers
2620  *    schedule a write of some buffers
2621  *    return confirmation of parity correctness
2622  *
2623  * buffers are taken off read_list or write_list, and bh_cache buffers
2624  * get BH_Lock set before the stripe lock is released.
2625  *
2626  */
2627 
2628 static void handle_stripe5(struct stripe_head *sh)
2629 {
2630 	raid5_conf_t *conf = sh->raid_conf;
2631 	int disks = sh->disks, i;
2632 	struct bio *return_bi = NULL;
2633 	struct stripe_head_state s;
2634 	struct r5dev *dev;
2635 	unsigned long pending = 0;
2636 	mdk_rdev_t *blocked_rdev = NULL;
2637 
2638 	memset(&s, 0, sizeof(s));
2639 	pr_debug("handling stripe %llu, state=%#lx cnt=%d, pd_idx=%d "
2640 		"ops=%lx:%lx:%lx\n", (unsigned long long)sh->sector, sh->state,
2641 		atomic_read(&sh->count), sh->pd_idx,
2642 		sh->ops.pending, sh->ops.ack, sh->ops.complete);
2643 
2644 	spin_lock(&sh->lock);
2645 	clear_bit(STRIPE_HANDLE, &sh->state);
2646 	clear_bit(STRIPE_DELAYED, &sh->state);
2647 
2648 	s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2649 	s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2650 	s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2651 	/* Now to look around and see what can be done */
2652 
2653 	/* clean-up completed biofill operations */
2654 	if (test_bit(STRIPE_OP_BIOFILL, &sh->ops.complete)) {
2655 		clear_bit(STRIPE_OP_BIOFILL, &sh->ops.pending);
2656 		clear_bit(STRIPE_OP_BIOFILL, &sh->ops.ack);
2657 		clear_bit(STRIPE_OP_BIOFILL, &sh->ops.complete);
2658 	}
2659 
2660 	rcu_read_lock();
2661 	for (i=disks; i--; ) {
2662 		mdk_rdev_t *rdev;
2663 		struct r5dev *dev = &sh->dev[i];
2664 		clear_bit(R5_Insync, &dev->flags);
2665 
2666 		pr_debug("check %d: state 0x%lx toread %p read %p write %p "
2667 			"written %p\n",	i, dev->flags, dev->toread, dev->read,
2668 			dev->towrite, dev->written);
2669 
2670 		/* maybe we can request a biofill operation
2671 		 *
2672 		 * new wantfill requests are only permitted while
2673 		 * STRIPE_OP_BIOFILL is clear
2674 		 */
2675 		if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread &&
2676 			!test_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2677 			set_bit(R5_Wantfill, &dev->flags);
2678 
2679 		/* now count some things */
2680 		if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2681 		if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2682 		if (test_bit(R5_Wantcompute, &dev->flags)) s.compute++;
2683 
2684 		if (test_bit(R5_Wantfill, &dev->flags))
2685 			s.to_fill++;
2686 		else if (dev->toread)
2687 			s.to_read++;
2688 		if (dev->towrite) {
2689 			s.to_write++;
2690 			if (!test_bit(R5_OVERWRITE, &dev->flags))
2691 				s.non_overwrite++;
2692 		}
2693 		if (dev->written)
2694 			s.written++;
2695 		rdev = rcu_dereference(conf->disks[i].rdev);
2696 		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2697 			blocked_rdev = rdev;
2698 			atomic_inc(&rdev->nr_pending);
2699 			break;
2700 		}
2701 		if (!rdev || !test_bit(In_sync, &rdev->flags)) {
2702 			/* The ReadError flag will just be confusing now */
2703 			clear_bit(R5_ReadError, &dev->flags);
2704 			clear_bit(R5_ReWrite, &dev->flags);
2705 		}
2706 		if (!rdev || !test_bit(In_sync, &rdev->flags)
2707 		    || test_bit(R5_ReadError, &dev->flags)) {
2708 			s.failed++;
2709 			s.failed_num = i;
2710 		} else
2711 			set_bit(R5_Insync, &dev->flags);
2712 	}
2713 	rcu_read_unlock();
2714 
2715 	if (unlikely(blocked_rdev)) {
2716 		set_bit(STRIPE_HANDLE, &sh->state);
2717 		goto unlock;
2718 	}
2719 
2720 	if (s.to_fill && !test_and_set_bit(STRIPE_OP_BIOFILL, &sh->ops.pending))
2721 		sh->ops.count++;
2722 
2723 	pr_debug("locked=%d uptodate=%d to_read=%d"
2724 		" to_write=%d failed=%d failed_num=%d\n",
2725 		s.locked, s.uptodate, s.to_read, s.to_write,
2726 		s.failed, s.failed_num);
2727 	/* check if the array has lost two devices and, if so, some requests might
2728 	 * need to be failed
2729 	 */
2730 	if (s.failed > 1 && s.to_read+s.to_write+s.written)
2731 		handle_requests_to_failed_array(conf, sh, &s, disks,
2732 						&return_bi);
2733 	if (s.failed > 1 && s.syncing) {
2734 		md_done_sync(conf->mddev, STRIPE_SECTORS,0);
2735 		clear_bit(STRIPE_SYNCING, &sh->state);
2736 		s.syncing = 0;
2737 	}
2738 
2739 	/* might be able to return some write requests if the parity block
2740 	 * is safe, or on a failed drive
2741 	 */
2742 	dev = &sh->dev[sh->pd_idx];
2743 	if ( s.written &&
2744 	     ((test_bit(R5_Insync, &dev->flags) &&
2745 	       !test_bit(R5_LOCKED, &dev->flags) &&
2746 	       test_bit(R5_UPTODATE, &dev->flags)) ||
2747 	       (s.failed == 1 && s.failed_num == sh->pd_idx)))
2748 		handle_completed_write_requests(conf, sh, disks, &return_bi);
2749 
2750 	/* Now we might consider reading some blocks, either to check/generate
2751 	 * parity, or to satisfy requests
2752 	 * or to load a block that is being partially written.
2753 	 */
2754 	if (s.to_read || s.non_overwrite ||
2755 	    (s.syncing && (s.uptodate + s.compute < disks)) || s.expanding ||
2756 	    test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2757 		handle_issuing_new_read_requests5(sh, &s, disks);
2758 
2759 	/* Now we check to see if any write operations have recently
2760 	 * completed
2761 	 */
2762 
2763 	/* leave prexor set until postxor is done, allows us to distinguish
2764 	 * a rmw from a rcw during biodrain
2765 	 */
2766 	if (test_bit(STRIPE_OP_PREXOR, &sh->ops.complete) &&
2767 		test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2768 
2769 		clear_bit(STRIPE_OP_PREXOR, &sh->ops.complete);
2770 		clear_bit(STRIPE_OP_PREXOR, &sh->ops.ack);
2771 		clear_bit(STRIPE_OP_PREXOR, &sh->ops.pending);
2772 
2773 		for (i = disks; i--; )
2774 			clear_bit(R5_Wantprexor, &sh->dev[i].flags);
2775 	}
2776 
2777 	/* if only POSTXOR is set then this is an 'expand' postxor */
2778 	if (test_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete) &&
2779 		test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete)) {
2780 
2781 		clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.complete);
2782 		clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.ack);
2783 		clear_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending);
2784 
2785 		clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2786 		clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2787 		clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2788 
2789 		/* All the 'written' buffers and the parity block are ready to
2790 		 * be written back to disk
2791 		 */
2792 		BUG_ON(!test_bit(R5_UPTODATE, &sh->dev[sh->pd_idx].flags));
2793 		for (i = disks; i--; ) {
2794 			dev = &sh->dev[i];
2795 			if (test_bit(R5_LOCKED, &dev->flags) &&
2796 				(i == sh->pd_idx || dev->written)) {
2797 				pr_debug("Writing block %d\n", i);
2798 				set_bit(R5_Wantwrite, &dev->flags);
2799 				if (!test_and_set_bit(
2800 				    STRIPE_OP_IO, &sh->ops.pending))
2801 					sh->ops.count++;
2802 				if (!test_bit(R5_Insync, &dev->flags) ||
2803 				    (i == sh->pd_idx && s.failed == 0))
2804 					set_bit(STRIPE_INSYNC, &sh->state);
2805 			}
2806 		}
2807 		if (test_and_clear_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) {
2808 			atomic_dec(&conf->preread_active_stripes);
2809 			if (atomic_read(&conf->preread_active_stripes) <
2810 				IO_THRESHOLD)
2811 				md_wakeup_thread(conf->mddev->thread);
2812 		}
2813 	}
2814 
2815 	/* Now to consider new write requests and what else, if anything
2816 	 * should be read.  We do not handle new writes when:
2817 	 * 1/ A 'write' operation (copy+xor) is already in flight.
2818 	 * 2/ A 'check' operation is in flight, as it may clobber the parity
2819 	 *    block.
2820 	 */
2821 	if (s.to_write && !test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending) &&
2822 			  !test_bit(STRIPE_OP_CHECK, &sh->ops.pending))
2823 		handle_issuing_new_write_requests5(conf, sh, &s, disks);
2824 
2825 	/* maybe we need to check and possibly fix the parity for this stripe
2826 	 * Any reads will already have been scheduled, so we just see if enough
2827 	 * data is available.  The parity check is held off while parity
2828 	 * dependent operations are in flight.
2829 	 */
2830 	if ((s.syncing && s.locked == 0 &&
2831 	     !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending) &&
2832 	     !test_bit(STRIPE_INSYNC, &sh->state)) ||
2833 	      test_bit(STRIPE_OP_CHECK, &sh->ops.pending) ||
2834 	      test_bit(STRIPE_OP_MOD_REPAIR_PD, &sh->ops.pending))
2835 		handle_parity_checks5(conf, sh, &s, disks);
2836 
2837 	if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
2838 		md_done_sync(conf->mddev, STRIPE_SECTORS,1);
2839 		clear_bit(STRIPE_SYNCING, &sh->state);
2840 	}
2841 
2842 	/* If the failed drive is just a ReadError, then we might need to progress
2843 	 * the repair/check process
2844 	 */
2845 	if (s.failed == 1 && !conf->mddev->ro &&
2846 	    test_bit(R5_ReadError, &sh->dev[s.failed_num].flags)
2847 	    && !test_bit(R5_LOCKED, &sh->dev[s.failed_num].flags)
2848 	    && test_bit(R5_UPTODATE, &sh->dev[s.failed_num].flags)
2849 		) {
2850 		dev = &sh->dev[s.failed_num];
2851 		if (!test_bit(R5_ReWrite, &dev->flags)) {
2852 			set_bit(R5_Wantwrite, &dev->flags);
2853 			if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2854 				sh->ops.count++;
2855 			set_bit(R5_ReWrite, &dev->flags);
2856 			set_bit(R5_LOCKED, &dev->flags);
2857 			s.locked++;
2858 		} else {
2859 			/* let's read it back */
2860 			set_bit(R5_Wantread, &dev->flags);
2861 			if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2862 				sh->ops.count++;
2863 			set_bit(R5_LOCKED, &dev->flags);
2864 			s.locked++;
2865 		}
2866 	}
2867 
2868 	/* Finish postxor operations initiated by the expansion
2869 	 * process
2870 	 */
2871 	if (test_bit(STRIPE_OP_POSTXOR, &sh->ops.complete) &&
2872 		!test_bit(STRIPE_OP_BIODRAIN, &sh->ops.pending)) {
2873 
2874 		clear_bit(STRIPE_EXPANDING, &sh->state);
2875 
2876 		clear_bit(STRIPE_OP_POSTXOR, &sh->ops.pending);
2877 		clear_bit(STRIPE_OP_POSTXOR, &sh->ops.ack);
2878 		clear_bit(STRIPE_OP_POSTXOR, &sh->ops.complete);
2879 
2880 		for (i = conf->raid_disks; i--; ) {
2881 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
2882 			if (!test_and_set_bit(STRIPE_OP_IO, &sh->ops.pending))
2883 				sh->ops.count++;
2884 		}
2885 	}
2886 
2887 	if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state) &&
2888 		!test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2889 		/* Need to write out all blocks after computing parity */
2890 		sh->disks = conf->raid_disks;
2891 		sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
2892 			conf->raid_disks);
2893 		s.locked += handle_write_operations5(sh, 1, 1);
2894 	} else if (s.expanded &&
2895 		!test_bit(STRIPE_OP_POSTXOR, &sh->ops.pending)) {
2896 		clear_bit(STRIPE_EXPAND_READY, &sh->state);
2897 		atomic_dec(&conf->reshape_stripes);
2898 		wake_up(&conf->wait_for_overlap);
2899 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
2900 	}
2901 
2902 	if (s.expanding && s.locked == 0 &&
2903 	    !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
2904 		handle_stripe_expansion(conf, sh, NULL);
2905 
2906 	if (sh->ops.count)
2907 		pending = get_stripe_work(sh);
2908 
2909  unlock:
2910 	spin_unlock(&sh->lock);
2911 
2912 	/* wait for this device to become unblocked */
2913 	if (unlikely(blocked_rdev))
2914 		md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
2915 
2916 	if (pending)
2917 		raid5_run_ops(sh, pending);
2918 
2919 	return_io(return_bi);
2920 
2921 }
2922 
2923 static void handle_stripe6(struct stripe_head *sh, struct page *tmp_page)
2924 {
2925 	raid6_conf_t *conf = sh->raid_conf;
2926 	int disks = sh->disks;
2927 	struct bio *return_bi = NULL;
2928 	int i, pd_idx = sh->pd_idx;
2929 	struct stripe_head_state s;
2930 	struct r6_state r6s;
2931 	struct r5dev *dev, *pdev, *qdev;
2932 	mdk_rdev_t *blocked_rdev = NULL;
2933 
2934 	r6s.qd_idx = raid6_next_disk(pd_idx, disks);
2935 	pr_debug("handling stripe %llu, state=%#lx cnt=%d, "
2936 		"pd_idx=%d, qd_idx=%d\n",
2937 	       (unsigned long long)sh->sector, sh->state,
2938 	       atomic_read(&sh->count), pd_idx, r6s.qd_idx);
2939 	memset(&s, 0, sizeof(s));
2940 
2941 	spin_lock(&sh->lock);
2942 	clear_bit(STRIPE_HANDLE, &sh->state);
2943 	clear_bit(STRIPE_DELAYED, &sh->state);
2944 
2945 	s.syncing = test_bit(STRIPE_SYNCING, &sh->state);
2946 	s.expanding = test_bit(STRIPE_EXPAND_SOURCE, &sh->state);
2947 	s.expanded = test_bit(STRIPE_EXPAND_READY, &sh->state);
2948 	/* Now to look around and see what can be done */
2949 
2950 	rcu_read_lock();
2951 	for (i=disks; i--; ) {
2952 		mdk_rdev_t *rdev;
2953 		dev = &sh->dev[i];
2954 		clear_bit(R5_Insync, &dev->flags);
2955 
2956 		pr_debug("check %d: state 0x%lx read %p write %p written %p\n",
2957 			i, dev->flags, dev->toread, dev->towrite, dev->written);
2958 		/* maybe we can reply to a read */
2959 		if (test_bit(R5_UPTODATE, &dev->flags) && dev->toread) {
2960 			struct bio *rbi, *rbi2;
2961 			pr_debug("Return read for disc %d\n", i);
2962 			spin_lock_irq(&conf->device_lock);
2963 			rbi = dev->toread;
2964 			dev->toread = NULL;
2965 			if (test_and_clear_bit(R5_Overlap, &dev->flags))
2966 				wake_up(&conf->wait_for_overlap);
2967 			spin_unlock_irq(&conf->device_lock);
2968 			while (rbi && rbi->bi_sector < dev->sector + STRIPE_SECTORS) {
2969 				copy_data(0, rbi, dev->page, dev->sector);
2970 				rbi2 = r5_next_bio(rbi, dev->sector);
2971 				spin_lock_irq(&conf->device_lock);
2972 				if (--rbi->bi_phys_segments == 0) {
2973 					rbi->bi_next = return_bi;
2974 					return_bi = rbi;
2975 				}
2976 				spin_unlock_irq(&conf->device_lock);
2977 				rbi = rbi2;
2978 			}
2979 		}
2980 
2981 		/* now count some things */
2982 		if (test_bit(R5_LOCKED, &dev->flags)) s.locked++;
2983 		if (test_bit(R5_UPTODATE, &dev->flags)) s.uptodate++;
2984 
2985 
2986 		if (dev->toread)
2987 			s.to_read++;
2988 		if (dev->towrite) {
2989 			s.to_write++;
2990 			if (!test_bit(R5_OVERWRITE, &dev->flags))
2991 				s.non_overwrite++;
2992 		}
2993 		if (dev->written)
2994 			s.written++;
2995 		rdev = rcu_dereference(conf->disks[i].rdev);
2996 		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
2997 			blocked_rdev = rdev;
2998 			atomic_inc(&rdev->nr_pending);
2999 			break;
3000 		}
3001 		if (!rdev || !test_bit(In_sync, &rdev->flags)) {
3002 			/* The ReadError flag will just be confusing now */
3003 			clear_bit(R5_ReadError, &dev->flags);
3004 			clear_bit(R5_ReWrite, &dev->flags);
3005 		}
3006 		if (!rdev || !test_bit(In_sync, &rdev->flags)
3007 		    || test_bit(R5_ReadError, &dev->flags)) {
3008 			if (s.failed < 2)
3009 				r6s.failed_num[s.failed] = i;
3010 			s.failed++;
3011 		} else
3012 			set_bit(R5_Insync, &dev->flags);
3013 	}
3014 	rcu_read_unlock();
3015 
3016 	if (unlikely(blocked_rdev)) {
3017 		set_bit(STRIPE_HANDLE, &sh->state);
3018 		goto unlock;
3019 	}
3020 	pr_debug("locked=%d uptodate=%d to_read=%d"
3021 	       " to_write=%d failed=%d failed_num=%d,%d\n",
3022 	       s.locked, s.uptodate, s.to_read, s.to_write, s.failed,
3023 	       r6s.failed_num[0], r6s.failed_num[1]);
3024 	/* check if the array has lost >2 devices and, if so, some requests
3025 	 * might need to be failed
3026 	 */
3027 	if (s.failed > 2 && s.to_read+s.to_write+s.written)
3028 		handle_requests_to_failed_array(conf, sh, &s, disks,
3029 						&return_bi);
3030 	if (s.failed > 2 && s.syncing) {
3031 		md_done_sync(conf->mddev, STRIPE_SECTORS,0);
3032 		clear_bit(STRIPE_SYNCING, &sh->state);
3033 		s.syncing = 0;
3034 	}
3035 
3036 	/*
3037 	 * might be able to return some write requests if the parity blocks
3038 	 * are safe, or on a failed drive
3039 	 */
3040 	pdev = &sh->dev[pd_idx];
3041 	r6s.p_failed = (s.failed >= 1 && r6s.failed_num[0] == pd_idx)
3042 		|| (s.failed >= 2 && r6s.failed_num[1] == pd_idx);
3043 	qdev = &sh->dev[r6s.qd_idx];
3044 	r6s.q_failed = (s.failed >= 1 && r6s.failed_num[0] == r6s.qd_idx)
3045 		|| (s.failed >= 2 && r6s.failed_num[1] == r6s.qd_idx);
3046 
3047 	if ( s.written &&
3048 	     ( r6s.p_failed || ((test_bit(R5_Insync, &pdev->flags)
3049 			     && !test_bit(R5_LOCKED, &pdev->flags)
3050 			     && test_bit(R5_UPTODATE, &pdev->flags)))) &&
3051 	     ( r6s.q_failed || ((test_bit(R5_Insync, &qdev->flags)
3052 			     && !test_bit(R5_LOCKED, &qdev->flags)
3053 			     && test_bit(R5_UPTODATE, &qdev->flags)))))
3054 		handle_completed_write_requests(conf, sh, disks, &return_bi);
3055 
3056 	/* Now we might consider reading some blocks, either to check/generate
3057 	 * parity, or to satisfy requests
3058 	 * or to load a block that is being partially written.
3059 	 */
3060 	if (s.to_read || s.non_overwrite || (s.to_write && s.failed) ||
3061 	    (s.syncing && (s.uptodate < disks)) || s.expanding)
3062 		handle_issuing_new_read_requests6(sh, &s, &r6s, disks);
3063 
3064 	/* now to consider writing and what else, if anything should be read */
3065 	if (s.to_write)
3066 		handle_issuing_new_write_requests6(conf, sh, &s, &r6s, disks);
3067 
3068 	/* maybe we need to check and possibly fix the parity for this stripe
3069 	 * Any reads will already have been scheduled, so we just see if enough
3070 	 * data is available
3071 	 */
3072 	if (s.syncing && s.locked == 0 && !test_bit(STRIPE_INSYNC, &sh->state))
3073 		handle_parity_checks6(conf, sh, &s, &r6s, tmp_page, disks);
3074 
3075 	if (s.syncing && s.locked == 0 && test_bit(STRIPE_INSYNC, &sh->state)) {
3076 		md_done_sync(conf->mddev, STRIPE_SECTORS,1);
3077 		clear_bit(STRIPE_SYNCING, &sh->state);
3078 	}
3079 
3080 	/* If the failed drives are just a ReadError, then we might need
3081 	 * to progress the repair/check process
3082 	 */
3083 	if (s.failed <= 2 && !conf->mddev->ro)
3084 		for (i = 0; i < s.failed; i++) {
3085 			dev = &sh->dev[r6s.failed_num[i]];
3086 			if (test_bit(R5_ReadError, &dev->flags)
3087 			    && !test_bit(R5_LOCKED, &dev->flags)
3088 			    && test_bit(R5_UPTODATE, &dev->flags)
3089 				) {
3090 				if (!test_bit(R5_ReWrite, &dev->flags)) {
3091 					set_bit(R5_Wantwrite, &dev->flags);
3092 					set_bit(R5_ReWrite, &dev->flags);
3093 					set_bit(R5_LOCKED, &dev->flags);
3094 				} else {
3095 					/* let's read it back */
3096 					set_bit(R5_Wantread, &dev->flags);
3097 					set_bit(R5_LOCKED, &dev->flags);
3098 				}
3099 			}
3100 		}
3101 
3102 	if (s.expanded && test_bit(STRIPE_EXPANDING, &sh->state)) {
3103 		/* Need to write out all blocks after computing P&Q */
3104 		sh->disks = conf->raid_disks;
3105 		sh->pd_idx = stripe_to_pdidx(sh->sector, conf,
3106 					     conf->raid_disks);
3107 		compute_parity6(sh, RECONSTRUCT_WRITE);
3108 		for (i = conf->raid_disks ; i-- ;  ) {
3109 			set_bit(R5_LOCKED, &sh->dev[i].flags);
3110 			s.locked++;
3111 			set_bit(R5_Wantwrite, &sh->dev[i].flags);
3112 		}
3113 		clear_bit(STRIPE_EXPANDING, &sh->state);
3114 	} else if (s.expanded) {
3115 		clear_bit(STRIPE_EXPAND_READY, &sh->state);
3116 		atomic_dec(&conf->reshape_stripes);
3117 		wake_up(&conf->wait_for_overlap);
3118 		md_done_sync(conf->mddev, STRIPE_SECTORS, 1);
3119 	}
3120 
3121 	if (s.expanding && s.locked == 0 &&
3122 	    !test_bit(STRIPE_OP_COMPUTE_BLK, &sh->ops.pending))
3123 		handle_stripe_expansion(conf, sh, &r6s);
3124 
3125  unlock:
3126 	spin_unlock(&sh->lock);
3127 
3128 	/* wait for this device to become unblocked */
3129 	if (unlikely(blocked_rdev))
3130 		md_wait_for_blocked_rdev(blocked_rdev, conf->mddev);
3131 
3132 	return_io(return_bi);
3133 
3134 	for (i=disks; i-- ;) {
3135 		int rw;
3136 		struct bio *bi;
3137 		mdk_rdev_t *rdev;
3138 		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags))
3139 			rw = WRITE;
3140 		else if (test_and_clear_bit(R5_Wantread, &sh->dev[i].flags))
3141 			rw = READ;
3142 		else
3143 			continue;
3144 
3145 		set_bit(STRIPE_IO_STARTED, &sh->state);
3146 
3147 		bi = &sh->dev[i].req;
3148 
3149 		bi->bi_rw = rw;
3150 		if (rw == WRITE)
3151 			bi->bi_end_io = raid5_end_write_request;
3152 		else
3153 			bi->bi_end_io = raid5_end_read_request;
3154 
3155 		rcu_read_lock();
3156 		rdev = rcu_dereference(conf->disks[i].rdev);
3157 		if (rdev && test_bit(Faulty, &rdev->flags))
3158 			rdev = NULL;
3159 		if (rdev)
3160 			atomic_inc(&rdev->nr_pending);
3161 		rcu_read_unlock();
3162 
3163 		if (rdev) {
3164 			if (s.syncing || s.expanding || s.expanded)
3165 				md_sync_acct(rdev->bdev, STRIPE_SECTORS);
3166 
3167 			bi->bi_bdev = rdev->bdev;
3168 			pr_debug("for %llu schedule op %ld on disc %d\n",
3169 				(unsigned long long)sh->sector, bi->bi_rw, i);
3170 			atomic_inc(&sh->count);
3171 			bi->bi_sector = sh->sector + rdev->data_offset;
3172 			bi->bi_flags = 1 << BIO_UPTODATE;
3173 			bi->bi_vcnt = 1;
3174 			bi->bi_max_vecs = 1;
3175 			bi->bi_idx = 0;
3176 			bi->bi_io_vec = &sh->dev[i].vec;
3177 			bi->bi_io_vec[0].bv_len = STRIPE_SIZE;
3178 			bi->bi_io_vec[0].bv_offset = 0;
3179 			bi->bi_size = STRIPE_SIZE;
3180 			bi->bi_next = NULL;
3181 			if (rw == WRITE &&
3182 			    test_bit(R5_ReWrite, &sh->dev[i].flags))
3183 				atomic_add(STRIPE_SECTORS, &rdev->corrected_errors);
3184 			generic_make_request(bi);
3185 		} else {
3186 			if (rw == WRITE)
3187 				set_bit(STRIPE_DEGRADED, &sh->state);
3188 			pr_debug("skip op %ld on disc %d for sector %llu\n",
3189 				bi->bi_rw, i, (unsigned long long)sh->sector);
3190 			clear_bit(R5_LOCKED, &sh->dev[i].flags);
3191 			set_bit(STRIPE_HANDLE, &sh->state);
3192 		}
3193 	}
3194 }
3195 
3196 static void handle_stripe(struct stripe_head *sh, struct page *tmp_page)
3197 {
3198 	if (sh->raid_conf->level == 6)
3199 		handle_stripe6(sh, tmp_page);
3200 	else
3201 		handle_stripe5(sh);
3202 }
3203 
3204 
3205 
3206 static void raid5_activate_delayed(raid5_conf_t *conf)
3207 {
3208 	if (atomic_read(&conf->preread_active_stripes) < IO_THRESHOLD) {
3209 		while (!list_empty(&conf->delayed_list)) {
3210 			struct list_head *l = conf->delayed_list.next;
3211 			struct stripe_head *sh;
3212 			sh = list_entry(l, struct stripe_head, lru);
3213 			list_del_init(l);
3214 			clear_bit(STRIPE_DELAYED, &sh->state);
3215 			if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
3216 				atomic_inc(&conf->preread_active_stripes);
3217 			list_add_tail(&sh->lru, &conf->hold_list);
3218 		}
3219 	} else
3220 		blk_plug_device(conf->mddev->queue);
3221 }
3222 
3223 static void activate_bit_delay(raid5_conf_t *conf)
3224 {
3225 	/* device_lock is held */
3226 	struct list_head head;
3227 	list_add(&head, &conf->bitmap_list);
3228 	list_del_init(&conf->bitmap_list);
3229 	while (!list_empty(&head)) {
3230 		struct stripe_head *sh = list_entry(head.next, struct stripe_head, lru);
3231 		list_del_init(&sh->lru);
3232 		atomic_inc(&sh->count);
3233 		__release_stripe(conf, sh);
3234 	}
3235 }
3236 
3237 static void unplug_slaves(mddev_t *mddev)
3238 {
3239 	raid5_conf_t *conf = mddev_to_conf(mddev);
3240 	int i;
3241 
3242 	rcu_read_lock();
3243 	for (i=0; i<mddev->raid_disks; i++) {
3244 		mdk_rdev_t *rdev = rcu_dereference(conf->disks[i].rdev);
3245 		if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
3246 			struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
3247 
3248 			atomic_inc(&rdev->nr_pending);
3249 			rcu_read_unlock();
3250 
3251 			blk_unplug(r_queue);
3252 
3253 			rdev_dec_pending(rdev, mddev);
3254 			rcu_read_lock();
3255 		}
3256 	}
3257 	rcu_read_unlock();
3258 }
3259 
3260 static void raid5_unplug_device(struct request_queue *q)
3261 {
3262 	mddev_t *mddev = q->queuedata;
3263 	raid5_conf_t *conf = mddev_to_conf(mddev);
3264 	unsigned long flags;
3265 
3266 	spin_lock_irqsave(&conf->device_lock, flags);
3267 
3268 	if (blk_remove_plug(q)) {
3269 		conf->seq_flush++;
3270 		raid5_activate_delayed(conf);
3271 	}
3272 	md_wakeup_thread(mddev->thread);
3273 
3274 	spin_unlock_irqrestore(&conf->device_lock, flags);
3275 
3276 	unplug_slaves(mddev);
3277 }
3278 
3279 static int raid5_congested(void *data, int bits)
3280 {
3281 	mddev_t *mddev = data;
3282 	raid5_conf_t *conf = mddev_to_conf(mddev);
3283 
3284 	/* No difference between reads and writes.  Just check
3285 	 * how busy the stripe_cache is
3286 	 */
3287 	if (conf->inactive_blocked)
3288 		return 1;
3289 	if (conf->quiesce)
3290 		return 1;
3291 	if (list_empty_careful(&conf->inactive_list))
3292 		return 1;
3293 
3294 	return 0;
3295 }
3296 
3297 /* We want read requests to align with chunks where possible,
3298  * but write requests don't need to.
3299  */
3300 static int raid5_mergeable_bvec(struct request_queue *q, struct bio *bio, struct bio_vec *biovec)
3301 {
3302 	mddev_t *mddev = q->queuedata;
3303 	sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3304 	int max;
3305 	unsigned int chunk_sectors = mddev->chunk_size >> 9;
3306 	unsigned int bio_sectors = bio->bi_size >> 9;
3307 
3308 	if (bio_data_dir(bio) == WRITE)
3309 		return biovec->bv_len; /* always allow writes to be mergeable */
3310 
3311 	max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
3312 	if (max < 0) max = 0;
3313 	if (max <= biovec->bv_len && bio_sectors == 0)
3314 		return biovec->bv_len;
3315 	else
3316 		return max;
3317 }
3318 
3319 
3320 static int in_chunk_boundary(mddev_t *mddev, struct bio *bio)
3321 {
3322 	sector_t sector = bio->bi_sector + get_start_sect(bio->bi_bdev);
3323 	unsigned int chunk_sectors = mddev->chunk_size >> 9;
3324 	unsigned int bio_sectors = bio->bi_size >> 9;
3325 
3326 	return  chunk_sectors >=
3327 		((sector & (chunk_sectors - 1)) + bio_sectors);
3328 }
3329 
3330 /*
3331  *  add bio to the retry LIFO  ( in O(1) ... we are in interrupt )
3332  *  later sampled by raid5d.
3333  */
3334 static void add_bio_to_retry(struct bio *bi,raid5_conf_t *conf)
3335 {
3336 	unsigned long flags;
3337 
3338 	spin_lock_irqsave(&conf->device_lock, flags);
3339 
3340 	bi->bi_next = conf->retry_read_aligned_list;
3341 	conf->retry_read_aligned_list = bi;
3342 
3343 	spin_unlock_irqrestore(&conf->device_lock, flags);
3344 	md_wakeup_thread(conf->mddev->thread);
3345 }
3346 
3347 
3348 static struct bio *remove_bio_from_retry(raid5_conf_t *conf)
3349 {
3350 	struct bio *bi;
3351 
3352 	bi = conf->retry_read_aligned;
3353 	if (bi) {
3354 		conf->retry_read_aligned = NULL;
3355 		return bi;
3356 	}
3357 	bi = conf->retry_read_aligned_list;
3358 	if(bi) {
3359 		conf->retry_read_aligned_list = bi->bi_next;
3360 		bi->bi_next = NULL;
3361 		bi->bi_phys_segments = 1; /* biased count of active stripes */
3362 		bi->bi_hw_segments = 0; /* count of processed stripes */
3363 	}
3364 
3365 	return bi;
3366 }
3367 
3368 
3369 /*
3370  *  The "raid5_align_endio" should check if the read succeeded and if it
3371  *  did, call bio_endio on the original bio (having bio_put the new bio
3372  *  first).
3373  *  If the read failed..
3374  */
3375 static void raid5_align_endio(struct bio *bi, int error)
3376 {
3377 	struct bio* raid_bi  = bi->bi_private;
3378 	mddev_t *mddev;
3379 	raid5_conf_t *conf;
3380 	int uptodate = test_bit(BIO_UPTODATE, &bi->bi_flags);
3381 	mdk_rdev_t *rdev;
3382 
3383 	bio_put(bi);
3384 
3385 	mddev = raid_bi->bi_bdev->bd_disk->queue->queuedata;
3386 	conf = mddev_to_conf(mddev);
3387 	rdev = (void*)raid_bi->bi_next;
3388 	raid_bi->bi_next = NULL;
3389 
3390 	rdev_dec_pending(rdev, conf->mddev);
3391 
3392 	if (!error && uptodate) {
3393 		bio_endio(raid_bi, 0);
3394 		if (atomic_dec_and_test(&conf->active_aligned_reads))
3395 			wake_up(&conf->wait_for_stripe);
3396 		return;
3397 	}
3398 
3399 
3400 	pr_debug("raid5_align_endio : io error...handing IO for a retry\n");
3401 
3402 	add_bio_to_retry(raid_bi, conf);
3403 }
3404 
3405 static int bio_fits_rdev(struct bio *bi)
3406 {
3407 	struct request_queue *q = bdev_get_queue(bi->bi_bdev);
3408 
3409 	if ((bi->bi_size>>9) > q->max_sectors)
3410 		return 0;
3411 	blk_recount_segments(q, bi);
3412 	if (bi->bi_phys_segments > q->max_phys_segments ||
3413 	    bi->bi_hw_segments > q->max_hw_segments)
3414 		return 0;
3415 
3416 	if (q->merge_bvec_fn)
3417 		/* it's too hard to apply the merge_bvec_fn at this stage,
3418 		 * just just give up
3419 		 */
3420 		return 0;
3421 
3422 	return 1;
3423 }
3424 
3425 
3426 static int chunk_aligned_read(struct request_queue *q, struct bio * raid_bio)
3427 {
3428 	mddev_t *mddev = q->queuedata;
3429 	raid5_conf_t *conf = mddev_to_conf(mddev);
3430 	const unsigned int raid_disks = conf->raid_disks;
3431 	const unsigned int data_disks = raid_disks - conf->max_degraded;
3432 	unsigned int dd_idx, pd_idx;
3433 	struct bio* align_bi;
3434 	mdk_rdev_t *rdev;
3435 
3436 	if (!in_chunk_boundary(mddev, raid_bio)) {
3437 		pr_debug("chunk_aligned_read : non aligned\n");
3438 		return 0;
3439 	}
3440 	/*
3441  	 * use bio_clone to make a copy of the bio
3442 	 */
3443 	align_bi = bio_clone(raid_bio, GFP_NOIO);
3444 	if (!align_bi)
3445 		return 0;
3446 	/*
3447 	 *   set bi_end_io to a new function, and set bi_private to the
3448 	 *     original bio.
3449 	 */
3450 	align_bi->bi_end_io  = raid5_align_endio;
3451 	align_bi->bi_private = raid_bio;
3452 	/*
3453 	 *	compute position
3454 	 */
3455 	align_bi->bi_sector =  raid5_compute_sector(raid_bio->bi_sector,
3456 					raid_disks,
3457 					data_disks,
3458 					&dd_idx,
3459 					&pd_idx,
3460 					conf);
3461 
3462 	rcu_read_lock();
3463 	rdev = rcu_dereference(conf->disks[dd_idx].rdev);
3464 	if (rdev && test_bit(In_sync, &rdev->flags)) {
3465 		atomic_inc(&rdev->nr_pending);
3466 		rcu_read_unlock();
3467 		raid_bio->bi_next = (void*)rdev;
3468 		align_bi->bi_bdev =  rdev->bdev;
3469 		align_bi->bi_flags &= ~(1 << BIO_SEG_VALID);
3470 		align_bi->bi_sector += rdev->data_offset;
3471 
3472 		if (!bio_fits_rdev(align_bi)) {
3473 			/* too big in some way */
3474 			bio_put(align_bi);
3475 			rdev_dec_pending(rdev, mddev);
3476 			return 0;
3477 		}
3478 
3479 		spin_lock_irq(&conf->device_lock);
3480 		wait_event_lock_irq(conf->wait_for_stripe,
3481 				    conf->quiesce == 0,
3482 				    conf->device_lock, /* nothing */);
3483 		atomic_inc(&conf->active_aligned_reads);
3484 		spin_unlock_irq(&conf->device_lock);
3485 
3486 		generic_make_request(align_bi);
3487 		return 1;
3488 	} else {
3489 		rcu_read_unlock();
3490 		bio_put(align_bi);
3491 		return 0;
3492 	}
3493 }
3494 
3495 /* __get_priority_stripe - get the next stripe to process
3496  *
3497  * Full stripe writes are allowed to pass preread active stripes up until
3498  * the bypass_threshold is exceeded.  In general the bypass_count
3499  * increments when the handle_list is handled before the hold_list; however, it
3500  * will not be incremented when STRIPE_IO_STARTED is sampled set signifying a
3501  * stripe with in flight i/o.  The bypass_count will be reset when the
3502  * head of the hold_list has changed, i.e. the head was promoted to the
3503  * handle_list.
3504  */
3505 static struct stripe_head *__get_priority_stripe(raid5_conf_t *conf)
3506 {
3507 	struct stripe_head *sh;
3508 
3509 	pr_debug("%s: handle: %s hold: %s full_writes: %d bypass_count: %d\n",
3510 		  __func__,
3511 		  list_empty(&conf->handle_list) ? "empty" : "busy",
3512 		  list_empty(&conf->hold_list) ? "empty" : "busy",
3513 		  atomic_read(&conf->pending_full_writes), conf->bypass_count);
3514 
3515 	if (!list_empty(&conf->handle_list)) {
3516 		sh = list_entry(conf->handle_list.next, typeof(*sh), lru);
3517 
3518 		if (list_empty(&conf->hold_list))
3519 			conf->bypass_count = 0;
3520 		else if (!test_bit(STRIPE_IO_STARTED, &sh->state)) {
3521 			if (conf->hold_list.next == conf->last_hold)
3522 				conf->bypass_count++;
3523 			else {
3524 				conf->last_hold = conf->hold_list.next;
3525 				conf->bypass_count -= conf->bypass_threshold;
3526 				if (conf->bypass_count < 0)
3527 					conf->bypass_count = 0;
3528 			}
3529 		}
3530 	} else if (!list_empty(&conf->hold_list) &&
3531 		   ((conf->bypass_threshold &&
3532 		     conf->bypass_count > conf->bypass_threshold) ||
3533 		    atomic_read(&conf->pending_full_writes) == 0)) {
3534 		sh = list_entry(conf->hold_list.next,
3535 				typeof(*sh), lru);
3536 		conf->bypass_count -= conf->bypass_threshold;
3537 		if (conf->bypass_count < 0)
3538 			conf->bypass_count = 0;
3539 	} else
3540 		return NULL;
3541 
3542 	list_del_init(&sh->lru);
3543 	atomic_inc(&sh->count);
3544 	BUG_ON(atomic_read(&sh->count) != 1);
3545 	return sh;
3546 }
3547 
3548 static int make_request(struct request_queue *q, struct bio * bi)
3549 {
3550 	mddev_t *mddev = q->queuedata;
3551 	raid5_conf_t *conf = mddev_to_conf(mddev);
3552 	unsigned int dd_idx, pd_idx;
3553 	sector_t new_sector;
3554 	sector_t logical_sector, last_sector;
3555 	struct stripe_head *sh;
3556 	const int rw = bio_data_dir(bi);
3557 	int remaining;
3558 
3559 	if (unlikely(bio_barrier(bi))) {
3560 		bio_endio(bi, -EOPNOTSUPP);
3561 		return 0;
3562 	}
3563 
3564 	md_write_start(mddev, bi);
3565 
3566 	disk_stat_inc(mddev->gendisk, ios[rw]);
3567 	disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bi));
3568 
3569 	if (rw == READ &&
3570 	     mddev->reshape_position == MaxSector &&
3571 	     chunk_aligned_read(q,bi))
3572             	return 0;
3573 
3574 	logical_sector = bi->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3575 	last_sector = bi->bi_sector + (bi->bi_size>>9);
3576 	bi->bi_next = NULL;
3577 	bi->bi_phys_segments = 1;	/* over-loaded to count active stripes */
3578 
3579 	for (;logical_sector < last_sector; logical_sector += STRIPE_SECTORS) {
3580 		DEFINE_WAIT(w);
3581 		int disks, data_disks;
3582 
3583 	retry:
3584 		prepare_to_wait(&conf->wait_for_overlap, &w, TASK_UNINTERRUPTIBLE);
3585 		if (likely(conf->expand_progress == MaxSector))
3586 			disks = conf->raid_disks;
3587 		else {
3588 			/* spinlock is needed as expand_progress may be
3589 			 * 64bit on a 32bit platform, and so it might be
3590 			 * possible to see a half-updated value
3591 			 * Ofcourse expand_progress could change after
3592 			 * the lock is dropped, so once we get a reference
3593 			 * to the stripe that we think it is, we will have
3594 			 * to check again.
3595 			 */
3596 			spin_lock_irq(&conf->device_lock);
3597 			disks = conf->raid_disks;
3598 			if (logical_sector >= conf->expand_progress)
3599 				disks = conf->previous_raid_disks;
3600 			else {
3601 				if (logical_sector >= conf->expand_lo) {
3602 					spin_unlock_irq(&conf->device_lock);
3603 					schedule();
3604 					goto retry;
3605 				}
3606 			}
3607 			spin_unlock_irq(&conf->device_lock);
3608 		}
3609 		data_disks = disks - conf->max_degraded;
3610 
3611  		new_sector = raid5_compute_sector(logical_sector, disks, data_disks,
3612 						  &dd_idx, &pd_idx, conf);
3613 		pr_debug("raid5: make_request, sector %llu logical %llu\n",
3614 			(unsigned long long)new_sector,
3615 			(unsigned long long)logical_sector);
3616 
3617 		sh = get_active_stripe(conf, new_sector, disks, pd_idx, (bi->bi_rw&RWA_MASK));
3618 		if (sh) {
3619 			if (unlikely(conf->expand_progress != MaxSector)) {
3620 				/* expansion might have moved on while waiting for a
3621 				 * stripe, so we must do the range check again.
3622 				 * Expansion could still move past after this
3623 				 * test, but as we are holding a reference to
3624 				 * 'sh', we know that if that happens,
3625 				 *  STRIPE_EXPANDING will get set and the expansion
3626 				 * won't proceed until we finish with the stripe.
3627 				 */
3628 				int must_retry = 0;
3629 				spin_lock_irq(&conf->device_lock);
3630 				if (logical_sector <  conf->expand_progress &&
3631 				    disks == conf->previous_raid_disks)
3632 					/* mismatch, need to try again */
3633 					must_retry = 1;
3634 				spin_unlock_irq(&conf->device_lock);
3635 				if (must_retry) {
3636 					release_stripe(sh);
3637 					goto retry;
3638 				}
3639 			}
3640 			/* FIXME what if we get a false positive because these
3641 			 * are being updated.
3642 			 */
3643 			if (logical_sector >= mddev->suspend_lo &&
3644 			    logical_sector < mddev->suspend_hi) {
3645 				release_stripe(sh);
3646 				schedule();
3647 				goto retry;
3648 			}
3649 
3650 			if (test_bit(STRIPE_EXPANDING, &sh->state) ||
3651 			    !add_stripe_bio(sh, bi, dd_idx, (bi->bi_rw&RW_MASK))) {
3652 				/* Stripe is busy expanding or
3653 				 * add failed due to overlap.  Flush everything
3654 				 * and wait a while
3655 				 */
3656 				raid5_unplug_device(mddev->queue);
3657 				release_stripe(sh);
3658 				schedule();
3659 				goto retry;
3660 			}
3661 			finish_wait(&conf->wait_for_overlap, &w);
3662 			set_bit(STRIPE_HANDLE, &sh->state);
3663 			clear_bit(STRIPE_DELAYED, &sh->state);
3664 			release_stripe(sh);
3665 		} else {
3666 			/* cannot get stripe for read-ahead, just give-up */
3667 			clear_bit(BIO_UPTODATE, &bi->bi_flags);
3668 			finish_wait(&conf->wait_for_overlap, &w);
3669 			break;
3670 		}
3671 
3672 	}
3673 	spin_lock_irq(&conf->device_lock);
3674 	remaining = --bi->bi_phys_segments;
3675 	spin_unlock_irq(&conf->device_lock);
3676 	if (remaining == 0) {
3677 
3678 		if ( rw == WRITE )
3679 			md_write_end(mddev);
3680 
3681 		bi->bi_end_io(bi,
3682 			      test_bit(BIO_UPTODATE, &bi->bi_flags)
3683 			        ? 0 : -EIO);
3684 	}
3685 	return 0;
3686 }
3687 
3688 static sector_t reshape_request(mddev_t *mddev, sector_t sector_nr, int *skipped)
3689 {
3690 	/* reshaping is quite different to recovery/resync so it is
3691 	 * handled quite separately ... here.
3692 	 *
3693 	 * On each call to sync_request, we gather one chunk worth of
3694 	 * destination stripes and flag them as expanding.
3695 	 * Then we find all the source stripes and request reads.
3696 	 * As the reads complete, handle_stripe will copy the data
3697 	 * into the destination stripe and release that stripe.
3698 	 */
3699 	raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3700 	struct stripe_head *sh;
3701 	int pd_idx;
3702 	sector_t first_sector, last_sector;
3703 	int raid_disks = conf->previous_raid_disks;
3704 	int data_disks = raid_disks - conf->max_degraded;
3705 	int new_data_disks = conf->raid_disks - conf->max_degraded;
3706 	int i;
3707 	int dd_idx;
3708 	sector_t writepos, safepos, gap;
3709 
3710 	if (sector_nr == 0 &&
3711 	    conf->expand_progress != 0) {
3712 		/* restarting in the middle, skip the initial sectors */
3713 		sector_nr = conf->expand_progress;
3714 		sector_div(sector_nr, new_data_disks);
3715 		*skipped = 1;
3716 		return sector_nr;
3717 	}
3718 
3719 	/* we update the metadata when there is more than 3Meg
3720 	 * in the block range (that is rather arbitrary, should
3721 	 * probably be time based) or when the data about to be
3722 	 * copied would over-write the source of the data at
3723 	 * the front of the range.
3724 	 * i.e. one new_stripe forward from expand_progress new_maps
3725 	 * to after where expand_lo old_maps to
3726 	 */
3727 	writepos = conf->expand_progress +
3728 		conf->chunk_size/512*(new_data_disks);
3729 	sector_div(writepos, new_data_disks);
3730 	safepos = conf->expand_lo;
3731 	sector_div(safepos, data_disks);
3732 	gap = conf->expand_progress - conf->expand_lo;
3733 
3734 	if (writepos >= safepos ||
3735 	    gap > (new_data_disks)*3000*2 /*3Meg*/) {
3736 		/* Cannot proceed until we've updated the superblock... */
3737 		wait_event(conf->wait_for_overlap,
3738 			   atomic_read(&conf->reshape_stripes)==0);
3739 		mddev->reshape_position = conf->expand_progress;
3740 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
3741 		md_wakeup_thread(mddev->thread);
3742 		wait_event(mddev->sb_wait, mddev->flags == 0 ||
3743 			   kthread_should_stop());
3744 		spin_lock_irq(&conf->device_lock);
3745 		conf->expand_lo = mddev->reshape_position;
3746 		spin_unlock_irq(&conf->device_lock);
3747 		wake_up(&conf->wait_for_overlap);
3748 	}
3749 
3750 	for (i=0; i < conf->chunk_size/512; i+= STRIPE_SECTORS) {
3751 		int j;
3752 		int skipped = 0;
3753 		pd_idx = stripe_to_pdidx(sector_nr+i, conf, conf->raid_disks);
3754 		sh = get_active_stripe(conf, sector_nr+i,
3755 				       conf->raid_disks, pd_idx, 0);
3756 		set_bit(STRIPE_EXPANDING, &sh->state);
3757 		atomic_inc(&conf->reshape_stripes);
3758 		/* If any of this stripe is beyond the end of the old
3759 		 * array, then we need to zero those blocks
3760 		 */
3761 		for (j=sh->disks; j--;) {
3762 			sector_t s;
3763 			if (j == sh->pd_idx)
3764 				continue;
3765 			if (conf->level == 6 &&
3766 			    j == raid6_next_disk(sh->pd_idx, sh->disks))
3767 				continue;
3768 			s = compute_blocknr(sh, j);
3769 			if (s < (mddev->array_size<<1)) {
3770 				skipped = 1;
3771 				continue;
3772 			}
3773 			memset(page_address(sh->dev[j].page), 0, STRIPE_SIZE);
3774 			set_bit(R5_Expanded, &sh->dev[j].flags);
3775 			set_bit(R5_UPTODATE, &sh->dev[j].flags);
3776 		}
3777 		if (!skipped) {
3778 			set_bit(STRIPE_EXPAND_READY, &sh->state);
3779 			set_bit(STRIPE_HANDLE, &sh->state);
3780 		}
3781 		release_stripe(sh);
3782 	}
3783 	spin_lock_irq(&conf->device_lock);
3784 	conf->expand_progress = (sector_nr + i) * new_data_disks;
3785 	spin_unlock_irq(&conf->device_lock);
3786 	/* Ok, those stripe are ready. We can start scheduling
3787 	 * reads on the source stripes.
3788 	 * The source stripes are determined by mapping the first and last
3789 	 * block on the destination stripes.
3790 	 */
3791 	first_sector =
3792 		raid5_compute_sector(sector_nr*(new_data_disks),
3793 				     raid_disks, data_disks,
3794 				     &dd_idx, &pd_idx, conf);
3795 	last_sector =
3796 		raid5_compute_sector((sector_nr+conf->chunk_size/512)
3797 				     *(new_data_disks) -1,
3798 				     raid_disks, data_disks,
3799 				     &dd_idx, &pd_idx, conf);
3800 	if (last_sector >= (mddev->size<<1))
3801 		last_sector = (mddev->size<<1)-1;
3802 	while (first_sector <= last_sector) {
3803 		pd_idx = stripe_to_pdidx(first_sector, conf,
3804 					 conf->previous_raid_disks);
3805 		sh = get_active_stripe(conf, first_sector,
3806 				       conf->previous_raid_disks, pd_idx, 0);
3807 		set_bit(STRIPE_EXPAND_SOURCE, &sh->state);
3808 		set_bit(STRIPE_HANDLE, &sh->state);
3809 		release_stripe(sh);
3810 		first_sector += STRIPE_SECTORS;
3811 	}
3812 	/* If this takes us to the resync_max point where we have to pause,
3813 	 * then we need to write out the superblock.
3814 	 */
3815 	sector_nr += conf->chunk_size>>9;
3816 	if (sector_nr >= mddev->resync_max) {
3817 		/* Cannot proceed until we've updated the superblock... */
3818 		wait_event(conf->wait_for_overlap,
3819 			   atomic_read(&conf->reshape_stripes) == 0);
3820 		mddev->reshape_position = conf->expand_progress;
3821 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
3822 		md_wakeup_thread(mddev->thread);
3823 		wait_event(mddev->sb_wait,
3824 			   !test_bit(MD_CHANGE_DEVS, &mddev->flags)
3825 			   || kthread_should_stop());
3826 		spin_lock_irq(&conf->device_lock);
3827 		conf->expand_lo = mddev->reshape_position;
3828 		spin_unlock_irq(&conf->device_lock);
3829 		wake_up(&conf->wait_for_overlap);
3830 	}
3831 	return conf->chunk_size>>9;
3832 }
3833 
3834 /* FIXME go_faster isn't used */
3835 static inline sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
3836 {
3837 	raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
3838 	struct stripe_head *sh;
3839 	int pd_idx;
3840 	int raid_disks = conf->raid_disks;
3841 	sector_t max_sector = mddev->size << 1;
3842 	int sync_blocks;
3843 	int still_degraded = 0;
3844 	int i;
3845 
3846 	if (sector_nr >= max_sector) {
3847 		/* just being told to finish up .. nothing much to do */
3848 		unplug_slaves(mddev);
3849 		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3850 			end_reshape(conf);
3851 			return 0;
3852 		}
3853 
3854 		if (mddev->curr_resync < max_sector) /* aborted */
3855 			bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
3856 					&sync_blocks, 1);
3857 		else /* completed sync */
3858 			conf->fullsync = 0;
3859 		bitmap_close_sync(mddev->bitmap);
3860 
3861 		return 0;
3862 	}
3863 
3864 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3865 		return reshape_request(mddev, sector_nr, skipped);
3866 
3867 	/* No need to check resync_max as we never do more than one
3868 	 * stripe, and as resync_max will always be on a chunk boundary,
3869 	 * if the check in md_do_sync didn't fire, there is no chance
3870 	 * of overstepping resync_max here
3871 	 */
3872 
3873 	/* if there is too many failed drives and we are trying
3874 	 * to resync, then assert that we are finished, because there is
3875 	 * nothing we can do.
3876 	 */
3877 	if (mddev->degraded >= conf->max_degraded &&
3878 	    test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3879 		sector_t rv = (mddev->size << 1) - sector_nr;
3880 		*skipped = 1;
3881 		return rv;
3882 	}
3883 	if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
3884 	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3885 	    !conf->fullsync && sync_blocks >= STRIPE_SECTORS) {
3886 		/* we can skip this block, and probably more */
3887 		sync_blocks /= STRIPE_SECTORS;
3888 		*skipped = 1;
3889 		return sync_blocks * STRIPE_SECTORS; /* keep things rounded to whole stripes */
3890 	}
3891 
3892 
3893 	bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3894 
3895 	pd_idx = stripe_to_pdidx(sector_nr, conf, raid_disks);
3896 	sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 1);
3897 	if (sh == NULL) {
3898 		sh = get_active_stripe(conf, sector_nr, raid_disks, pd_idx, 0);
3899 		/* make sure we don't swamp the stripe cache if someone else
3900 		 * is trying to get access
3901 		 */
3902 		schedule_timeout_uninterruptible(1);
3903 	}
3904 	/* Need to check if array will still be degraded after recovery/resync
3905 	 * We don't need to check the 'failed' flag as when that gets set,
3906 	 * recovery aborts.
3907 	 */
3908 	for (i=0; i<mddev->raid_disks; i++)
3909 		if (conf->disks[i].rdev == NULL)
3910 			still_degraded = 1;
3911 
3912 	bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, still_degraded);
3913 
3914 	spin_lock(&sh->lock);
3915 	set_bit(STRIPE_SYNCING, &sh->state);
3916 	clear_bit(STRIPE_INSYNC, &sh->state);
3917 	spin_unlock(&sh->lock);
3918 
3919 	handle_stripe(sh, NULL);
3920 	release_stripe(sh);
3921 
3922 	return STRIPE_SECTORS;
3923 }
3924 
3925 static int  retry_aligned_read(raid5_conf_t *conf, struct bio *raid_bio)
3926 {
3927 	/* We may not be able to submit a whole bio at once as there
3928 	 * may not be enough stripe_heads available.
3929 	 * We cannot pre-allocate enough stripe_heads as we may need
3930 	 * more than exist in the cache (if we allow ever large chunks).
3931 	 * So we do one stripe head at a time and record in
3932 	 * ->bi_hw_segments how many have been done.
3933 	 *
3934 	 * We *know* that this entire raid_bio is in one chunk, so
3935 	 * it will be only one 'dd_idx' and only need one call to raid5_compute_sector.
3936 	 */
3937 	struct stripe_head *sh;
3938 	int dd_idx, pd_idx;
3939 	sector_t sector, logical_sector, last_sector;
3940 	int scnt = 0;
3941 	int remaining;
3942 	int handled = 0;
3943 
3944 	logical_sector = raid_bio->bi_sector & ~((sector_t)STRIPE_SECTORS-1);
3945 	sector = raid5_compute_sector(	logical_sector,
3946 					conf->raid_disks,
3947 					conf->raid_disks - conf->max_degraded,
3948 					&dd_idx,
3949 					&pd_idx,
3950 					conf);
3951 	last_sector = raid_bio->bi_sector + (raid_bio->bi_size>>9);
3952 
3953 	for (; logical_sector < last_sector;
3954 	     logical_sector += STRIPE_SECTORS,
3955 		     sector += STRIPE_SECTORS,
3956 		     scnt++) {
3957 
3958 		if (scnt < raid_bio->bi_hw_segments)
3959 			/* already done this stripe */
3960 			continue;
3961 
3962 		sh = get_active_stripe(conf, sector, conf->raid_disks, pd_idx, 1);
3963 
3964 		if (!sh) {
3965 			/* failed to get a stripe - must wait */
3966 			raid_bio->bi_hw_segments = scnt;
3967 			conf->retry_read_aligned = raid_bio;
3968 			return handled;
3969 		}
3970 
3971 		set_bit(R5_ReadError, &sh->dev[dd_idx].flags);
3972 		if (!add_stripe_bio(sh, raid_bio, dd_idx, 0)) {
3973 			release_stripe(sh);
3974 			raid_bio->bi_hw_segments = scnt;
3975 			conf->retry_read_aligned = raid_bio;
3976 			return handled;
3977 		}
3978 
3979 		handle_stripe(sh, NULL);
3980 		release_stripe(sh);
3981 		handled++;
3982 	}
3983 	spin_lock_irq(&conf->device_lock);
3984 	remaining = --raid_bio->bi_phys_segments;
3985 	spin_unlock_irq(&conf->device_lock);
3986 	if (remaining == 0) {
3987 
3988 		raid_bio->bi_end_io(raid_bio,
3989 			      test_bit(BIO_UPTODATE, &raid_bio->bi_flags)
3990 			        ? 0 : -EIO);
3991 	}
3992 	if (atomic_dec_and_test(&conf->active_aligned_reads))
3993 		wake_up(&conf->wait_for_stripe);
3994 	return handled;
3995 }
3996 
3997 
3998 
3999 /*
4000  * This is our raid5 kernel thread.
4001  *
4002  * We scan the hash table for stripes which can be handled now.
4003  * During the scan, completed stripes are saved for us by the interrupt
4004  * handler, so that they will not have to wait for our next wakeup.
4005  */
4006 static void raid5d(mddev_t *mddev)
4007 {
4008 	struct stripe_head *sh;
4009 	raid5_conf_t *conf = mddev_to_conf(mddev);
4010 	int handled;
4011 
4012 	pr_debug("+++ raid5d active\n");
4013 
4014 	md_check_recovery(mddev);
4015 
4016 	handled = 0;
4017 	spin_lock_irq(&conf->device_lock);
4018 	while (1) {
4019 		struct bio *bio;
4020 
4021 		if (conf->seq_flush != conf->seq_write) {
4022 			int seq = conf->seq_flush;
4023 			spin_unlock_irq(&conf->device_lock);
4024 			bitmap_unplug(mddev->bitmap);
4025 			spin_lock_irq(&conf->device_lock);
4026 			conf->seq_write = seq;
4027 			activate_bit_delay(conf);
4028 		}
4029 
4030 		while ((bio = remove_bio_from_retry(conf))) {
4031 			int ok;
4032 			spin_unlock_irq(&conf->device_lock);
4033 			ok = retry_aligned_read(conf, bio);
4034 			spin_lock_irq(&conf->device_lock);
4035 			if (!ok)
4036 				break;
4037 			handled++;
4038 		}
4039 
4040 		sh = __get_priority_stripe(conf);
4041 
4042 		if (!sh) {
4043 			async_tx_issue_pending_all();
4044 			break;
4045 		}
4046 		spin_unlock_irq(&conf->device_lock);
4047 
4048 		handled++;
4049 		handle_stripe(sh, conf->spare_page);
4050 		release_stripe(sh);
4051 
4052 		spin_lock_irq(&conf->device_lock);
4053 	}
4054 	pr_debug("%d stripes handled\n", handled);
4055 
4056 	spin_unlock_irq(&conf->device_lock);
4057 
4058 	unplug_slaves(mddev);
4059 
4060 	pr_debug("--- raid5d inactive\n");
4061 }
4062 
4063 static ssize_t
4064 raid5_show_stripe_cache_size(mddev_t *mddev, char *page)
4065 {
4066 	raid5_conf_t *conf = mddev_to_conf(mddev);
4067 	if (conf)
4068 		return sprintf(page, "%d\n", conf->max_nr_stripes);
4069 	else
4070 		return 0;
4071 }
4072 
4073 static ssize_t
4074 raid5_store_stripe_cache_size(mddev_t *mddev, const char *page, size_t len)
4075 {
4076 	raid5_conf_t *conf = mddev_to_conf(mddev);
4077 	unsigned long new;
4078 	if (len >= PAGE_SIZE)
4079 		return -EINVAL;
4080 	if (!conf)
4081 		return -ENODEV;
4082 
4083 	if (strict_strtoul(page, 10, &new))
4084 		return -EINVAL;
4085 	if (new <= 16 || new > 32768)
4086 		return -EINVAL;
4087 	while (new < conf->max_nr_stripes) {
4088 		if (drop_one_stripe(conf))
4089 			conf->max_nr_stripes--;
4090 		else
4091 			break;
4092 	}
4093 	md_allow_write(mddev);
4094 	while (new > conf->max_nr_stripes) {
4095 		if (grow_one_stripe(conf))
4096 			conf->max_nr_stripes++;
4097 		else break;
4098 	}
4099 	return len;
4100 }
4101 
4102 static struct md_sysfs_entry
4103 raid5_stripecache_size = __ATTR(stripe_cache_size, S_IRUGO | S_IWUSR,
4104 				raid5_show_stripe_cache_size,
4105 				raid5_store_stripe_cache_size);
4106 
4107 static ssize_t
4108 raid5_show_preread_threshold(mddev_t *mddev, char *page)
4109 {
4110 	raid5_conf_t *conf = mddev_to_conf(mddev);
4111 	if (conf)
4112 		return sprintf(page, "%d\n", conf->bypass_threshold);
4113 	else
4114 		return 0;
4115 }
4116 
4117 static ssize_t
4118 raid5_store_preread_threshold(mddev_t *mddev, const char *page, size_t len)
4119 {
4120 	raid5_conf_t *conf = mddev_to_conf(mddev);
4121 	unsigned long new;
4122 	if (len >= PAGE_SIZE)
4123 		return -EINVAL;
4124 	if (!conf)
4125 		return -ENODEV;
4126 
4127 	if (strict_strtoul(page, 10, &new))
4128 		return -EINVAL;
4129 	if (new > conf->max_nr_stripes)
4130 		return -EINVAL;
4131 	conf->bypass_threshold = new;
4132 	return len;
4133 }
4134 
4135 static struct md_sysfs_entry
4136 raid5_preread_bypass_threshold = __ATTR(preread_bypass_threshold,
4137 					S_IRUGO | S_IWUSR,
4138 					raid5_show_preread_threshold,
4139 					raid5_store_preread_threshold);
4140 
4141 static ssize_t
4142 stripe_cache_active_show(mddev_t *mddev, char *page)
4143 {
4144 	raid5_conf_t *conf = mddev_to_conf(mddev);
4145 	if (conf)
4146 		return sprintf(page, "%d\n", atomic_read(&conf->active_stripes));
4147 	else
4148 		return 0;
4149 }
4150 
4151 static struct md_sysfs_entry
4152 raid5_stripecache_active = __ATTR_RO(stripe_cache_active);
4153 
4154 static struct attribute *raid5_attrs[] =  {
4155 	&raid5_stripecache_size.attr,
4156 	&raid5_stripecache_active.attr,
4157 	&raid5_preread_bypass_threshold.attr,
4158 	NULL,
4159 };
4160 static struct attribute_group raid5_attrs_group = {
4161 	.name = NULL,
4162 	.attrs = raid5_attrs,
4163 };
4164 
4165 static int run(mddev_t *mddev)
4166 {
4167 	raid5_conf_t *conf;
4168 	int raid_disk, memory;
4169 	mdk_rdev_t *rdev;
4170 	struct disk_info *disk;
4171 	struct list_head *tmp;
4172 	int working_disks = 0;
4173 
4174 	if (mddev->level != 5 && mddev->level != 4 && mddev->level != 6) {
4175 		printk(KERN_ERR "raid5: %s: raid level not set to 4/5/6 (%d)\n",
4176 		       mdname(mddev), mddev->level);
4177 		return -EIO;
4178 	}
4179 
4180 	if (mddev->reshape_position != MaxSector) {
4181 		/* Check that we can continue the reshape.
4182 		 * Currently only disks can change, it must
4183 		 * increase, and we must be past the point where
4184 		 * a stripe over-writes itself
4185 		 */
4186 		sector_t here_new, here_old;
4187 		int old_disks;
4188 		int max_degraded = (mddev->level == 5 ? 1 : 2);
4189 
4190 		if (mddev->new_level != mddev->level ||
4191 		    mddev->new_layout != mddev->layout ||
4192 		    mddev->new_chunk != mddev->chunk_size) {
4193 			printk(KERN_ERR "raid5: %s: unsupported reshape "
4194 			       "required - aborting.\n",
4195 			       mdname(mddev));
4196 			return -EINVAL;
4197 		}
4198 		if (mddev->delta_disks <= 0) {
4199 			printk(KERN_ERR "raid5: %s: unsupported reshape "
4200 			       "(reduce disks) required - aborting.\n",
4201 			       mdname(mddev));
4202 			return -EINVAL;
4203 		}
4204 		old_disks = mddev->raid_disks - mddev->delta_disks;
4205 		/* reshape_position must be on a new-stripe boundary, and one
4206 		 * further up in new geometry must map after here in old
4207 		 * geometry.
4208 		 */
4209 		here_new = mddev->reshape_position;
4210 		if (sector_div(here_new, (mddev->chunk_size>>9)*
4211 			       (mddev->raid_disks - max_degraded))) {
4212 			printk(KERN_ERR "raid5: reshape_position not "
4213 			       "on a stripe boundary\n");
4214 			return -EINVAL;
4215 		}
4216 		/* here_new is the stripe we will write to */
4217 		here_old = mddev->reshape_position;
4218 		sector_div(here_old, (mddev->chunk_size>>9)*
4219 			   (old_disks-max_degraded));
4220 		/* here_old is the first stripe that we might need to read
4221 		 * from */
4222 		if (here_new >= here_old) {
4223 			/* Reading from the same stripe as writing to - bad */
4224 			printk(KERN_ERR "raid5: reshape_position too early for "
4225 			       "auto-recovery - aborting.\n");
4226 			return -EINVAL;
4227 		}
4228 		printk(KERN_INFO "raid5: reshape will continue\n");
4229 		/* OK, we should be able to continue; */
4230 	}
4231 
4232 
4233 	mddev->private = kzalloc(sizeof (raid5_conf_t), GFP_KERNEL);
4234 	if ((conf = mddev->private) == NULL)
4235 		goto abort;
4236 	if (mddev->reshape_position == MaxSector) {
4237 		conf->previous_raid_disks = conf->raid_disks = mddev->raid_disks;
4238 	} else {
4239 		conf->raid_disks = mddev->raid_disks;
4240 		conf->previous_raid_disks = mddev->raid_disks - mddev->delta_disks;
4241 	}
4242 
4243 	conf->disks = kzalloc(conf->raid_disks * sizeof(struct disk_info),
4244 			      GFP_KERNEL);
4245 	if (!conf->disks)
4246 		goto abort;
4247 
4248 	conf->mddev = mddev;
4249 
4250 	if ((conf->stripe_hashtbl = kzalloc(PAGE_SIZE, GFP_KERNEL)) == NULL)
4251 		goto abort;
4252 
4253 	if (mddev->level == 6) {
4254 		conf->spare_page = alloc_page(GFP_KERNEL);
4255 		if (!conf->spare_page)
4256 			goto abort;
4257 	}
4258 	spin_lock_init(&conf->device_lock);
4259 	init_waitqueue_head(&conf->wait_for_stripe);
4260 	init_waitqueue_head(&conf->wait_for_overlap);
4261 	INIT_LIST_HEAD(&conf->handle_list);
4262 	INIT_LIST_HEAD(&conf->hold_list);
4263 	INIT_LIST_HEAD(&conf->delayed_list);
4264 	INIT_LIST_HEAD(&conf->bitmap_list);
4265 	INIT_LIST_HEAD(&conf->inactive_list);
4266 	atomic_set(&conf->active_stripes, 0);
4267 	atomic_set(&conf->preread_active_stripes, 0);
4268 	atomic_set(&conf->active_aligned_reads, 0);
4269 	conf->bypass_threshold = BYPASS_THRESHOLD;
4270 
4271 	pr_debug("raid5: run(%s) called.\n", mdname(mddev));
4272 
4273 	rdev_for_each(rdev, tmp, mddev) {
4274 		raid_disk = rdev->raid_disk;
4275 		if (raid_disk >= conf->raid_disks
4276 		    || raid_disk < 0)
4277 			continue;
4278 		disk = conf->disks + raid_disk;
4279 
4280 		disk->rdev = rdev;
4281 
4282 		if (test_bit(In_sync, &rdev->flags)) {
4283 			char b[BDEVNAME_SIZE];
4284 			printk(KERN_INFO "raid5: device %s operational as raid"
4285 				" disk %d\n", bdevname(rdev->bdev,b),
4286 				raid_disk);
4287 			working_disks++;
4288 		}
4289 	}
4290 
4291 	/*
4292 	 * 0 for a fully functional array, 1 or 2 for a degraded array.
4293 	 */
4294 	mddev->degraded = conf->raid_disks - working_disks;
4295 	conf->mddev = mddev;
4296 	conf->chunk_size = mddev->chunk_size;
4297 	conf->level = mddev->level;
4298 	if (conf->level == 6)
4299 		conf->max_degraded = 2;
4300 	else
4301 		conf->max_degraded = 1;
4302 	conf->algorithm = mddev->layout;
4303 	conf->max_nr_stripes = NR_STRIPES;
4304 	conf->expand_progress = mddev->reshape_position;
4305 
4306 	/* device size must be a multiple of chunk size */
4307 	mddev->size &= ~(mddev->chunk_size/1024 -1);
4308 	mddev->resync_max_sectors = mddev->size << 1;
4309 
4310 	if (conf->level == 6 && conf->raid_disks < 4) {
4311 		printk(KERN_ERR "raid6: not enough configured devices for %s (%d, minimum 4)\n",
4312 		       mdname(mddev), conf->raid_disks);
4313 		goto abort;
4314 	}
4315 	if (!conf->chunk_size || conf->chunk_size % 4) {
4316 		printk(KERN_ERR "raid5: invalid chunk size %d for %s\n",
4317 			conf->chunk_size, mdname(mddev));
4318 		goto abort;
4319 	}
4320 	if (conf->algorithm > ALGORITHM_RIGHT_SYMMETRIC) {
4321 		printk(KERN_ERR
4322 			"raid5: unsupported parity algorithm %d for %s\n",
4323 			conf->algorithm, mdname(mddev));
4324 		goto abort;
4325 	}
4326 	if (mddev->degraded > conf->max_degraded) {
4327 		printk(KERN_ERR "raid5: not enough operational devices for %s"
4328 			" (%d/%d failed)\n",
4329 			mdname(mddev), mddev->degraded, conf->raid_disks);
4330 		goto abort;
4331 	}
4332 
4333 	if (mddev->degraded > 0 &&
4334 	    mddev->recovery_cp != MaxSector) {
4335 		if (mddev->ok_start_degraded)
4336 			printk(KERN_WARNING
4337 			       "raid5: starting dirty degraded array: %s"
4338 			       "- data corruption possible.\n",
4339 			       mdname(mddev));
4340 		else {
4341 			printk(KERN_ERR
4342 			       "raid5: cannot start dirty degraded array for %s\n",
4343 			       mdname(mddev));
4344 			goto abort;
4345 		}
4346 	}
4347 
4348 	{
4349 		mddev->thread = md_register_thread(raid5d, mddev, "%s_raid5");
4350 		if (!mddev->thread) {
4351 			printk(KERN_ERR
4352 				"raid5: couldn't allocate thread for %s\n",
4353 				mdname(mddev));
4354 			goto abort;
4355 		}
4356 	}
4357 	memory = conf->max_nr_stripes * (sizeof(struct stripe_head) +
4358 		 conf->raid_disks * ((sizeof(struct bio) + PAGE_SIZE))) / 1024;
4359 	if (grow_stripes(conf, conf->max_nr_stripes)) {
4360 		printk(KERN_ERR
4361 			"raid5: couldn't allocate %dkB for buffers\n", memory);
4362 		shrink_stripes(conf);
4363 		md_unregister_thread(mddev->thread);
4364 		goto abort;
4365 	} else
4366 		printk(KERN_INFO "raid5: allocated %dkB for %s\n",
4367 			memory, mdname(mddev));
4368 
4369 	if (mddev->degraded == 0)
4370 		printk("raid5: raid level %d set %s active with %d out of %d"
4371 			" devices, algorithm %d\n", conf->level, mdname(mddev),
4372 			mddev->raid_disks-mddev->degraded, mddev->raid_disks,
4373 			conf->algorithm);
4374 	else
4375 		printk(KERN_ALERT "raid5: raid level %d set %s active with %d"
4376 			" out of %d devices, algorithm %d\n", conf->level,
4377 			mdname(mddev), mddev->raid_disks - mddev->degraded,
4378 			mddev->raid_disks, conf->algorithm);
4379 
4380 	print_raid5_conf(conf);
4381 
4382 	if (conf->expand_progress != MaxSector) {
4383 		printk("...ok start reshape thread\n");
4384 		conf->expand_lo = conf->expand_progress;
4385 		atomic_set(&conf->reshape_stripes, 0);
4386 		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4387 		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4388 		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4389 		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4390 		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4391 							"%s_reshape");
4392 	}
4393 
4394 	/* read-ahead size must cover two whole stripes, which is
4395 	 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4396 	 */
4397 	{
4398 		int data_disks = conf->previous_raid_disks - conf->max_degraded;
4399 		int stripe = data_disks *
4400 			(mddev->chunk_size / PAGE_SIZE);
4401 		if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4402 			mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4403 	}
4404 
4405 	/* Ok, everything is just fine now */
4406 	if (sysfs_create_group(&mddev->kobj, &raid5_attrs_group))
4407 		printk(KERN_WARNING
4408 		       "raid5: failed to create sysfs attributes for %s\n",
4409 		       mdname(mddev));
4410 
4411 	mddev->queue->unplug_fn = raid5_unplug_device;
4412 	mddev->queue->backing_dev_info.congested_data = mddev;
4413 	mddev->queue->backing_dev_info.congested_fn = raid5_congested;
4414 
4415 	mddev->array_size =  mddev->size * (conf->previous_raid_disks -
4416 					    conf->max_degraded);
4417 
4418 	blk_queue_merge_bvec(mddev->queue, raid5_mergeable_bvec);
4419 
4420 	return 0;
4421 abort:
4422 	if (conf) {
4423 		print_raid5_conf(conf);
4424 		safe_put_page(conf->spare_page);
4425 		kfree(conf->disks);
4426 		kfree(conf->stripe_hashtbl);
4427 		kfree(conf);
4428 	}
4429 	mddev->private = NULL;
4430 	printk(KERN_ALERT "raid5: failed to run raid set %s\n", mdname(mddev));
4431 	return -EIO;
4432 }
4433 
4434 
4435 
4436 static int stop(mddev_t *mddev)
4437 {
4438 	raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4439 
4440 	md_unregister_thread(mddev->thread);
4441 	mddev->thread = NULL;
4442 	shrink_stripes(conf);
4443 	kfree(conf->stripe_hashtbl);
4444 	mddev->queue->backing_dev_info.congested_fn = NULL;
4445 	blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
4446 	sysfs_remove_group(&mddev->kobj, &raid5_attrs_group);
4447 	kfree(conf->disks);
4448 	kfree(conf);
4449 	mddev->private = NULL;
4450 	return 0;
4451 }
4452 
4453 #ifdef DEBUG
4454 static void print_sh (struct seq_file *seq, struct stripe_head *sh)
4455 {
4456 	int i;
4457 
4458 	seq_printf(seq, "sh %llu, pd_idx %d, state %ld.\n",
4459 		   (unsigned long long)sh->sector, sh->pd_idx, sh->state);
4460 	seq_printf(seq, "sh %llu,  count %d.\n",
4461 		   (unsigned long long)sh->sector, atomic_read(&sh->count));
4462 	seq_printf(seq, "sh %llu, ", (unsigned long long)sh->sector);
4463 	for (i = 0; i < sh->disks; i++) {
4464 		seq_printf(seq, "(cache%d: %p %ld) ",
4465 			   i, sh->dev[i].page, sh->dev[i].flags);
4466 	}
4467 	seq_printf(seq, "\n");
4468 }
4469 
4470 static void printall (struct seq_file *seq, raid5_conf_t *conf)
4471 {
4472 	struct stripe_head *sh;
4473 	struct hlist_node *hn;
4474 	int i;
4475 
4476 	spin_lock_irq(&conf->device_lock);
4477 	for (i = 0; i < NR_HASH; i++) {
4478 		hlist_for_each_entry(sh, hn, &conf->stripe_hashtbl[i], hash) {
4479 			if (sh->raid_conf != conf)
4480 				continue;
4481 			print_sh(seq, sh);
4482 		}
4483 	}
4484 	spin_unlock_irq(&conf->device_lock);
4485 }
4486 #endif
4487 
4488 static void status (struct seq_file *seq, mddev_t *mddev)
4489 {
4490 	raid5_conf_t *conf = (raid5_conf_t *) mddev->private;
4491 	int i;
4492 
4493 	seq_printf (seq, " level %d, %dk chunk, algorithm %d", mddev->level, mddev->chunk_size >> 10, mddev->layout);
4494 	seq_printf (seq, " [%d/%d] [", conf->raid_disks, conf->raid_disks - mddev->degraded);
4495 	for (i = 0; i < conf->raid_disks; i++)
4496 		seq_printf (seq, "%s",
4497 			       conf->disks[i].rdev &&
4498 			       test_bit(In_sync, &conf->disks[i].rdev->flags) ? "U" : "_");
4499 	seq_printf (seq, "]");
4500 #ifdef DEBUG
4501 	seq_printf (seq, "\n");
4502 	printall(seq, conf);
4503 #endif
4504 }
4505 
4506 static void print_raid5_conf (raid5_conf_t *conf)
4507 {
4508 	int i;
4509 	struct disk_info *tmp;
4510 
4511 	printk("RAID5 conf printout:\n");
4512 	if (!conf) {
4513 		printk("(conf==NULL)\n");
4514 		return;
4515 	}
4516 	printk(" --- rd:%d wd:%d\n", conf->raid_disks,
4517 		 conf->raid_disks - conf->mddev->degraded);
4518 
4519 	for (i = 0; i < conf->raid_disks; i++) {
4520 		char b[BDEVNAME_SIZE];
4521 		tmp = conf->disks + i;
4522 		if (tmp->rdev)
4523 		printk(" disk %d, o:%d, dev:%s\n",
4524 			i, !test_bit(Faulty, &tmp->rdev->flags),
4525 			bdevname(tmp->rdev->bdev,b));
4526 	}
4527 }
4528 
4529 static int raid5_spare_active(mddev_t *mddev)
4530 {
4531 	int i;
4532 	raid5_conf_t *conf = mddev->private;
4533 	struct disk_info *tmp;
4534 
4535 	for (i = 0; i < conf->raid_disks; i++) {
4536 		tmp = conf->disks + i;
4537 		if (tmp->rdev
4538 		    && !test_bit(Faulty, &tmp->rdev->flags)
4539 		    && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
4540 			unsigned long flags;
4541 			spin_lock_irqsave(&conf->device_lock, flags);
4542 			mddev->degraded--;
4543 			spin_unlock_irqrestore(&conf->device_lock, flags);
4544 		}
4545 	}
4546 	print_raid5_conf(conf);
4547 	return 0;
4548 }
4549 
4550 static int raid5_remove_disk(mddev_t *mddev, int number)
4551 {
4552 	raid5_conf_t *conf = mddev->private;
4553 	int err = 0;
4554 	mdk_rdev_t *rdev;
4555 	struct disk_info *p = conf->disks + number;
4556 
4557 	print_raid5_conf(conf);
4558 	rdev = p->rdev;
4559 	if (rdev) {
4560 		if (test_bit(In_sync, &rdev->flags) ||
4561 		    atomic_read(&rdev->nr_pending)) {
4562 			err = -EBUSY;
4563 			goto abort;
4564 		}
4565 		p->rdev = NULL;
4566 		synchronize_rcu();
4567 		if (atomic_read(&rdev->nr_pending)) {
4568 			/* lost the race, try later */
4569 			err = -EBUSY;
4570 			p->rdev = rdev;
4571 		}
4572 	}
4573 abort:
4574 
4575 	print_raid5_conf(conf);
4576 	return err;
4577 }
4578 
4579 static int raid5_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
4580 {
4581 	raid5_conf_t *conf = mddev->private;
4582 	int found = 0;
4583 	int disk;
4584 	struct disk_info *p;
4585 
4586 	if (mddev->degraded > conf->max_degraded)
4587 		/* no point adding a device */
4588 		return 0;
4589 
4590 	/*
4591 	 * find the disk ... but prefer rdev->saved_raid_disk
4592 	 * if possible.
4593 	 */
4594 	if (rdev->saved_raid_disk >= 0 &&
4595 	    conf->disks[rdev->saved_raid_disk].rdev == NULL)
4596 		disk = rdev->saved_raid_disk;
4597 	else
4598 		disk = 0;
4599 	for ( ; disk < conf->raid_disks; disk++)
4600 		if ((p=conf->disks + disk)->rdev == NULL) {
4601 			clear_bit(In_sync, &rdev->flags);
4602 			rdev->raid_disk = disk;
4603 			found = 1;
4604 			if (rdev->saved_raid_disk != disk)
4605 				conf->fullsync = 1;
4606 			rcu_assign_pointer(p->rdev, rdev);
4607 			break;
4608 		}
4609 	print_raid5_conf(conf);
4610 	return found;
4611 }
4612 
4613 static int raid5_resize(mddev_t *mddev, sector_t sectors)
4614 {
4615 	/* no resync is happening, and there is enough space
4616 	 * on all devices, so we can resize.
4617 	 * We need to make sure resync covers any new space.
4618 	 * If the array is shrinking we should possibly wait until
4619 	 * any io in the removed space completes, but it hardly seems
4620 	 * worth it.
4621 	 */
4622 	raid5_conf_t *conf = mddev_to_conf(mddev);
4623 
4624 	sectors &= ~((sector_t)mddev->chunk_size/512 - 1);
4625 	mddev->array_size = (sectors * (mddev->raid_disks-conf->max_degraded))>>1;
4626 	set_capacity(mddev->gendisk, mddev->array_size << 1);
4627 	mddev->changed = 1;
4628 	if (sectors/2  > mddev->size && mddev->recovery_cp == MaxSector) {
4629 		mddev->recovery_cp = mddev->size << 1;
4630 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4631 	}
4632 	mddev->size = sectors /2;
4633 	mddev->resync_max_sectors = sectors;
4634 	return 0;
4635 }
4636 
4637 #ifdef CONFIG_MD_RAID5_RESHAPE
4638 static int raid5_check_reshape(mddev_t *mddev)
4639 {
4640 	raid5_conf_t *conf = mddev_to_conf(mddev);
4641 	int err;
4642 
4643 	if (mddev->delta_disks < 0 ||
4644 	    mddev->new_level != mddev->level)
4645 		return -EINVAL; /* Cannot shrink array or change level yet */
4646 	if (mddev->delta_disks == 0)
4647 		return 0; /* nothing to do */
4648 
4649 	/* Can only proceed if there are plenty of stripe_heads.
4650 	 * We need a minimum of one full stripe,, and for sensible progress
4651 	 * it is best to have about 4 times that.
4652 	 * If we require 4 times, then the default 256 4K stripe_heads will
4653 	 * allow for chunk sizes up to 256K, which is probably OK.
4654 	 * If the chunk size is greater, user-space should request more
4655 	 * stripe_heads first.
4656 	 */
4657 	if ((mddev->chunk_size / STRIPE_SIZE) * 4 > conf->max_nr_stripes ||
4658 	    (mddev->new_chunk / STRIPE_SIZE) * 4 > conf->max_nr_stripes) {
4659 		printk(KERN_WARNING "raid5: reshape: not enough stripes.  Needed %lu\n",
4660 		       (mddev->chunk_size / STRIPE_SIZE)*4);
4661 		return -ENOSPC;
4662 	}
4663 
4664 	err = resize_stripes(conf, conf->raid_disks + mddev->delta_disks);
4665 	if (err)
4666 		return err;
4667 
4668 	if (mddev->degraded > conf->max_degraded)
4669 		return -EINVAL;
4670 	/* looks like we might be able to manage this */
4671 	return 0;
4672 }
4673 
4674 static int raid5_start_reshape(mddev_t *mddev)
4675 {
4676 	raid5_conf_t *conf = mddev_to_conf(mddev);
4677 	mdk_rdev_t *rdev;
4678 	struct list_head *rtmp;
4679 	int spares = 0;
4680 	int added_devices = 0;
4681 	unsigned long flags;
4682 
4683 	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4684 		return -EBUSY;
4685 
4686 	rdev_for_each(rdev, rtmp, mddev)
4687 		if (rdev->raid_disk < 0 &&
4688 		    !test_bit(Faulty, &rdev->flags))
4689 			spares++;
4690 
4691 	if (spares - mddev->degraded < mddev->delta_disks - conf->max_degraded)
4692 		/* Not enough devices even to make a degraded array
4693 		 * of that size
4694 		 */
4695 		return -EINVAL;
4696 
4697 	atomic_set(&conf->reshape_stripes, 0);
4698 	spin_lock_irq(&conf->device_lock);
4699 	conf->previous_raid_disks = conf->raid_disks;
4700 	conf->raid_disks += mddev->delta_disks;
4701 	conf->expand_progress = 0;
4702 	conf->expand_lo = 0;
4703 	spin_unlock_irq(&conf->device_lock);
4704 
4705 	/* Add some new drives, as many as will fit.
4706 	 * We know there are enough to make the newly sized array work.
4707 	 */
4708 	rdev_for_each(rdev, rtmp, mddev)
4709 		if (rdev->raid_disk < 0 &&
4710 		    !test_bit(Faulty, &rdev->flags)) {
4711 			if (raid5_add_disk(mddev, rdev)) {
4712 				char nm[20];
4713 				set_bit(In_sync, &rdev->flags);
4714 				added_devices++;
4715 				rdev->recovery_offset = 0;
4716 				sprintf(nm, "rd%d", rdev->raid_disk);
4717 				if (sysfs_create_link(&mddev->kobj,
4718 						      &rdev->kobj, nm))
4719 					printk(KERN_WARNING
4720 					       "raid5: failed to create "
4721 					       " link %s for %s\n",
4722 					       nm, mdname(mddev));
4723 			} else
4724 				break;
4725 		}
4726 
4727 	spin_lock_irqsave(&conf->device_lock, flags);
4728 	mddev->degraded = (conf->raid_disks - conf->previous_raid_disks) - added_devices;
4729 	spin_unlock_irqrestore(&conf->device_lock, flags);
4730 	mddev->raid_disks = conf->raid_disks;
4731 	mddev->reshape_position = 0;
4732 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
4733 
4734 	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4735 	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4736 	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4737 	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4738 	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4739 						"%s_reshape");
4740 	if (!mddev->sync_thread) {
4741 		mddev->recovery = 0;
4742 		spin_lock_irq(&conf->device_lock);
4743 		mddev->raid_disks = conf->raid_disks = conf->previous_raid_disks;
4744 		conf->expand_progress = MaxSector;
4745 		spin_unlock_irq(&conf->device_lock);
4746 		return -EAGAIN;
4747 	}
4748 	md_wakeup_thread(mddev->sync_thread);
4749 	md_new_event(mddev);
4750 	return 0;
4751 }
4752 #endif
4753 
4754 static void end_reshape(raid5_conf_t *conf)
4755 {
4756 	struct block_device *bdev;
4757 
4758 	if (!test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery)) {
4759 		conf->mddev->array_size = conf->mddev->size *
4760 			(conf->raid_disks - conf->max_degraded);
4761 		set_capacity(conf->mddev->gendisk, conf->mddev->array_size << 1);
4762 		conf->mddev->changed = 1;
4763 
4764 		bdev = bdget_disk(conf->mddev->gendisk, 0);
4765 		if (bdev) {
4766 			mutex_lock(&bdev->bd_inode->i_mutex);
4767 			i_size_write(bdev->bd_inode, (loff_t)conf->mddev->array_size << 10);
4768 			mutex_unlock(&bdev->bd_inode->i_mutex);
4769 			bdput(bdev);
4770 		}
4771 		spin_lock_irq(&conf->device_lock);
4772 		conf->expand_progress = MaxSector;
4773 		spin_unlock_irq(&conf->device_lock);
4774 		conf->mddev->reshape_position = MaxSector;
4775 
4776 		/* read-ahead size must cover two whole stripes, which is
4777 		 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4778 		 */
4779 		{
4780 			int data_disks = conf->previous_raid_disks - conf->max_degraded;
4781 			int stripe = data_disks *
4782 				(conf->mddev->chunk_size / PAGE_SIZE);
4783 			if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4784 				conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4785 		}
4786 	}
4787 }
4788 
4789 static void raid5_quiesce(mddev_t *mddev, int state)
4790 {
4791 	raid5_conf_t *conf = mddev_to_conf(mddev);
4792 
4793 	switch(state) {
4794 	case 2: /* resume for a suspend */
4795 		wake_up(&conf->wait_for_overlap);
4796 		break;
4797 
4798 	case 1: /* stop all writes */
4799 		spin_lock_irq(&conf->device_lock);
4800 		conf->quiesce = 1;
4801 		wait_event_lock_irq(conf->wait_for_stripe,
4802 				    atomic_read(&conf->active_stripes) == 0 &&
4803 				    atomic_read(&conf->active_aligned_reads) == 0,
4804 				    conf->device_lock, /* nothing */);
4805 		spin_unlock_irq(&conf->device_lock);
4806 		break;
4807 
4808 	case 0: /* re-enable writes */
4809 		spin_lock_irq(&conf->device_lock);
4810 		conf->quiesce = 0;
4811 		wake_up(&conf->wait_for_stripe);
4812 		wake_up(&conf->wait_for_overlap);
4813 		spin_unlock_irq(&conf->device_lock);
4814 		break;
4815 	}
4816 }
4817 
4818 static struct mdk_personality raid6_personality =
4819 {
4820 	.name		= "raid6",
4821 	.level		= 6,
4822 	.owner		= THIS_MODULE,
4823 	.make_request	= make_request,
4824 	.run		= run,
4825 	.stop		= stop,
4826 	.status		= status,
4827 	.error_handler	= error,
4828 	.hot_add_disk	= raid5_add_disk,
4829 	.hot_remove_disk= raid5_remove_disk,
4830 	.spare_active	= raid5_spare_active,
4831 	.sync_request	= sync_request,
4832 	.resize		= raid5_resize,
4833 #ifdef CONFIG_MD_RAID5_RESHAPE
4834 	.check_reshape	= raid5_check_reshape,
4835 	.start_reshape  = raid5_start_reshape,
4836 #endif
4837 	.quiesce	= raid5_quiesce,
4838 };
4839 static struct mdk_personality raid5_personality =
4840 {
4841 	.name		= "raid5",
4842 	.level		= 5,
4843 	.owner		= THIS_MODULE,
4844 	.make_request	= make_request,
4845 	.run		= run,
4846 	.stop		= stop,
4847 	.status		= status,
4848 	.error_handler	= error,
4849 	.hot_add_disk	= raid5_add_disk,
4850 	.hot_remove_disk= raid5_remove_disk,
4851 	.spare_active	= raid5_spare_active,
4852 	.sync_request	= sync_request,
4853 	.resize		= raid5_resize,
4854 #ifdef CONFIG_MD_RAID5_RESHAPE
4855 	.check_reshape	= raid5_check_reshape,
4856 	.start_reshape  = raid5_start_reshape,
4857 #endif
4858 	.quiesce	= raid5_quiesce,
4859 };
4860 
4861 static struct mdk_personality raid4_personality =
4862 {
4863 	.name		= "raid4",
4864 	.level		= 4,
4865 	.owner		= THIS_MODULE,
4866 	.make_request	= make_request,
4867 	.run		= run,
4868 	.stop		= stop,
4869 	.status		= status,
4870 	.error_handler	= error,
4871 	.hot_add_disk	= raid5_add_disk,
4872 	.hot_remove_disk= raid5_remove_disk,
4873 	.spare_active	= raid5_spare_active,
4874 	.sync_request	= sync_request,
4875 	.resize		= raid5_resize,
4876 #ifdef CONFIG_MD_RAID5_RESHAPE
4877 	.check_reshape	= raid5_check_reshape,
4878 	.start_reshape  = raid5_start_reshape,
4879 #endif
4880 	.quiesce	= raid5_quiesce,
4881 };
4882 
4883 static int __init raid5_init(void)
4884 {
4885 	int e;
4886 
4887 	e = raid6_select_algo();
4888 	if ( e )
4889 		return e;
4890 	register_md_personality(&raid6_personality);
4891 	register_md_personality(&raid5_personality);
4892 	register_md_personality(&raid4_personality);
4893 	return 0;
4894 }
4895 
4896 static void raid5_exit(void)
4897 {
4898 	unregister_md_personality(&raid6_personality);
4899 	unregister_md_personality(&raid5_personality);
4900 	unregister_md_personality(&raid4_personality);
4901 }
4902 
4903 module_init(raid5_init);
4904 module_exit(raid5_exit);
4905 MODULE_LICENSE("GPL");
4906 MODULE_ALIAS("md-personality-4"); /* RAID5 */
4907 MODULE_ALIAS("md-raid5");
4908 MODULE_ALIAS("md-raid4");
4909 MODULE_ALIAS("md-level-5");
4910 MODULE_ALIAS("md-level-4");
4911 MODULE_ALIAS("md-personality-8"); /* RAID6 */
4912 MODULE_ALIAS("md-raid6");
4913 MODULE_ALIAS("md-level-6");
4914 
4915 /* This used to be two separate modules, they were: */
4916 MODULE_ALIAS("raid5");
4917 MODULE_ALIAS("raid6");
4918