xref: /openbmc/linux/drivers/md/raid10.c (revision f35e839a)
1 /*
2  * raid10.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 2000-2004 Neil Brown
5  *
6  * RAID-10 support for md.
7  *
8  * Base on code in raid1.c.  See raid1.c for further copyright information.
9  *
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 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
28 #include "md.h"
29 #include "raid10.h"
30 #include "raid0.h"
31 #include "bitmap.h"
32 
33 /*
34  * RAID10 provides a combination of RAID0 and RAID1 functionality.
35  * The layout of data is defined by
36  *    chunk_size
37  *    raid_disks
38  *    near_copies (stored in low byte of layout)
39  *    far_copies (stored in second byte of layout)
40  *    far_offset (stored in bit 16 of layout )
41  *    use_far_sets (stored in bit 17 of layout )
42  *
43  * The data to be stored is divided into chunks using chunksize.  Each device
44  * is divided into far_copies sections.   In each section, chunks are laid out
45  * in a style similar to raid0, but near_copies copies of each chunk is stored
46  * (each on a different drive).  The starting device for each section is offset
47  * near_copies from the starting device of the previous section.  Thus there
48  * are (near_copies * far_copies) of each chunk, and each is on a different
49  * drive.  near_copies and far_copies must be at least one, and their product
50  * is at most raid_disks.
51  *
52  * If far_offset is true, then the far_copies are handled a bit differently.
53  * The copies are still in different stripes, but instead of being very far
54  * apart on disk, there are adjacent stripes.
55  *
56  * The far and offset algorithms are handled slightly differently if
57  * 'use_far_sets' is true.  In this case, the array's devices are grouped into
58  * sets that are (near_copies * far_copies) in size.  The far copied stripes
59  * are still shifted by 'near_copies' devices, but this shifting stays confined
60  * to the set rather than the entire array.  This is done to improve the number
61  * of device combinations that can fail without causing the array to fail.
62  * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
63  * on a device):
64  *    A B C D    A B C D E
65  *      ...         ...
66  *    D A B C    E A B C D
67  * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
68  *    [A B] [C D]    [A B] [C D E]
69  *    |...| |...|    |...| | ... |
70  *    [B A] [D C]    [B A] [E C D]
71  */
72 
73 /*
74  * Number of guaranteed r10bios in case of extreme VM load:
75  */
76 #define	NR_RAID10_BIOS 256
77 
78 /* when we get a read error on a read-only array, we redirect to another
79  * device without failing the first device, or trying to over-write to
80  * correct the read error.  To keep track of bad blocks on a per-bio
81  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
82  */
83 #define IO_BLOCKED ((struct bio *)1)
84 /* When we successfully write to a known bad-block, we need to remove the
85  * bad-block marking which must be done from process context.  So we record
86  * the success by setting devs[n].bio to IO_MADE_GOOD
87  */
88 #define IO_MADE_GOOD ((struct bio *)2)
89 
90 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
91 
92 /* When there are this many requests queued to be written by
93  * the raid10 thread, we become 'congested' to provide back-pressure
94  * for writeback.
95  */
96 static int max_queued_requests = 1024;
97 
98 static void allow_barrier(struct r10conf *conf);
99 static void lower_barrier(struct r10conf *conf);
100 static int enough(struct r10conf *conf, int ignore);
101 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
102 				int *skipped);
103 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
104 static void end_reshape_write(struct bio *bio, int error);
105 static void end_reshape(struct r10conf *conf);
106 
107 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
108 {
109 	struct r10conf *conf = data;
110 	int size = offsetof(struct r10bio, devs[conf->copies]);
111 
112 	/* allocate a r10bio with room for raid_disks entries in the
113 	 * bios array */
114 	return kzalloc(size, gfp_flags);
115 }
116 
117 static void r10bio_pool_free(void *r10_bio, void *data)
118 {
119 	kfree(r10_bio);
120 }
121 
122 /* Maximum size of each resync request */
123 #define RESYNC_BLOCK_SIZE (64*1024)
124 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
125 /* amount of memory to reserve for resync requests */
126 #define RESYNC_WINDOW (1024*1024)
127 /* maximum number of concurrent requests, memory permitting */
128 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
129 
130 /*
131  * When performing a resync, we need to read and compare, so
132  * we need as many pages are there are copies.
133  * When performing a recovery, we need 2 bios, one for read,
134  * one for write (we recover only one drive per r10buf)
135  *
136  */
137 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
138 {
139 	struct r10conf *conf = data;
140 	struct page *page;
141 	struct r10bio *r10_bio;
142 	struct bio *bio;
143 	int i, j;
144 	int nalloc;
145 
146 	r10_bio = r10bio_pool_alloc(gfp_flags, conf);
147 	if (!r10_bio)
148 		return NULL;
149 
150 	if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
151 	    test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
152 		nalloc = conf->copies; /* resync */
153 	else
154 		nalloc = 2; /* recovery */
155 
156 	/*
157 	 * Allocate bios.
158 	 */
159 	for (j = nalloc ; j-- ; ) {
160 		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
161 		if (!bio)
162 			goto out_free_bio;
163 		r10_bio->devs[j].bio = bio;
164 		if (!conf->have_replacement)
165 			continue;
166 		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
167 		if (!bio)
168 			goto out_free_bio;
169 		r10_bio->devs[j].repl_bio = bio;
170 	}
171 	/*
172 	 * Allocate RESYNC_PAGES data pages and attach them
173 	 * where needed.
174 	 */
175 	for (j = 0 ; j < nalloc; j++) {
176 		struct bio *rbio = r10_bio->devs[j].repl_bio;
177 		bio = r10_bio->devs[j].bio;
178 		for (i = 0; i < RESYNC_PAGES; i++) {
179 			if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
180 					       &conf->mddev->recovery)) {
181 				/* we can share bv_page's during recovery
182 				 * and reshape */
183 				struct bio *rbio = r10_bio->devs[0].bio;
184 				page = rbio->bi_io_vec[i].bv_page;
185 				get_page(page);
186 			} else
187 				page = alloc_page(gfp_flags);
188 			if (unlikely(!page))
189 				goto out_free_pages;
190 
191 			bio->bi_io_vec[i].bv_page = page;
192 			if (rbio)
193 				rbio->bi_io_vec[i].bv_page = page;
194 		}
195 	}
196 
197 	return r10_bio;
198 
199 out_free_pages:
200 	for ( ; i > 0 ; i--)
201 		safe_put_page(bio->bi_io_vec[i-1].bv_page);
202 	while (j--)
203 		for (i = 0; i < RESYNC_PAGES ; i++)
204 			safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
205 	j = 0;
206 out_free_bio:
207 	for ( ; j < nalloc; j++) {
208 		if (r10_bio->devs[j].bio)
209 			bio_put(r10_bio->devs[j].bio);
210 		if (r10_bio->devs[j].repl_bio)
211 			bio_put(r10_bio->devs[j].repl_bio);
212 	}
213 	r10bio_pool_free(r10_bio, conf);
214 	return NULL;
215 }
216 
217 static void r10buf_pool_free(void *__r10_bio, void *data)
218 {
219 	int i;
220 	struct r10conf *conf = data;
221 	struct r10bio *r10bio = __r10_bio;
222 	int j;
223 
224 	for (j=0; j < conf->copies; j++) {
225 		struct bio *bio = r10bio->devs[j].bio;
226 		if (bio) {
227 			for (i = 0; i < RESYNC_PAGES; i++) {
228 				safe_put_page(bio->bi_io_vec[i].bv_page);
229 				bio->bi_io_vec[i].bv_page = NULL;
230 			}
231 			bio_put(bio);
232 		}
233 		bio = r10bio->devs[j].repl_bio;
234 		if (bio)
235 			bio_put(bio);
236 	}
237 	r10bio_pool_free(r10bio, conf);
238 }
239 
240 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
241 {
242 	int i;
243 
244 	for (i = 0; i < conf->copies; i++) {
245 		struct bio **bio = & r10_bio->devs[i].bio;
246 		if (!BIO_SPECIAL(*bio))
247 			bio_put(*bio);
248 		*bio = NULL;
249 		bio = &r10_bio->devs[i].repl_bio;
250 		if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
251 			bio_put(*bio);
252 		*bio = NULL;
253 	}
254 }
255 
256 static void free_r10bio(struct r10bio *r10_bio)
257 {
258 	struct r10conf *conf = r10_bio->mddev->private;
259 
260 	put_all_bios(conf, r10_bio);
261 	mempool_free(r10_bio, conf->r10bio_pool);
262 }
263 
264 static void put_buf(struct r10bio *r10_bio)
265 {
266 	struct r10conf *conf = r10_bio->mddev->private;
267 
268 	mempool_free(r10_bio, conf->r10buf_pool);
269 
270 	lower_barrier(conf);
271 }
272 
273 static void reschedule_retry(struct r10bio *r10_bio)
274 {
275 	unsigned long flags;
276 	struct mddev *mddev = r10_bio->mddev;
277 	struct r10conf *conf = mddev->private;
278 
279 	spin_lock_irqsave(&conf->device_lock, flags);
280 	list_add(&r10_bio->retry_list, &conf->retry_list);
281 	conf->nr_queued ++;
282 	spin_unlock_irqrestore(&conf->device_lock, flags);
283 
284 	/* wake up frozen array... */
285 	wake_up(&conf->wait_barrier);
286 
287 	md_wakeup_thread(mddev->thread);
288 }
289 
290 /*
291  * raid_end_bio_io() is called when we have finished servicing a mirrored
292  * operation and are ready to return a success/failure code to the buffer
293  * cache layer.
294  */
295 static void raid_end_bio_io(struct r10bio *r10_bio)
296 {
297 	struct bio *bio = r10_bio->master_bio;
298 	int done;
299 	struct r10conf *conf = r10_bio->mddev->private;
300 
301 	if (bio->bi_phys_segments) {
302 		unsigned long flags;
303 		spin_lock_irqsave(&conf->device_lock, flags);
304 		bio->bi_phys_segments--;
305 		done = (bio->bi_phys_segments == 0);
306 		spin_unlock_irqrestore(&conf->device_lock, flags);
307 	} else
308 		done = 1;
309 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
310 		clear_bit(BIO_UPTODATE, &bio->bi_flags);
311 	if (done) {
312 		bio_endio(bio, 0);
313 		/*
314 		 * Wake up any possible resync thread that waits for the device
315 		 * to go idle.
316 		 */
317 		allow_barrier(conf);
318 	}
319 	free_r10bio(r10_bio);
320 }
321 
322 /*
323  * Update disk head position estimator based on IRQ completion info.
324  */
325 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
326 {
327 	struct r10conf *conf = r10_bio->mddev->private;
328 
329 	conf->mirrors[r10_bio->devs[slot].devnum].head_position =
330 		r10_bio->devs[slot].addr + (r10_bio->sectors);
331 }
332 
333 /*
334  * Find the disk number which triggered given bio
335  */
336 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
337 			 struct bio *bio, int *slotp, int *replp)
338 {
339 	int slot;
340 	int repl = 0;
341 
342 	for (slot = 0; slot < conf->copies; slot++) {
343 		if (r10_bio->devs[slot].bio == bio)
344 			break;
345 		if (r10_bio->devs[slot].repl_bio == bio) {
346 			repl = 1;
347 			break;
348 		}
349 	}
350 
351 	BUG_ON(slot == conf->copies);
352 	update_head_pos(slot, r10_bio);
353 
354 	if (slotp)
355 		*slotp = slot;
356 	if (replp)
357 		*replp = repl;
358 	return r10_bio->devs[slot].devnum;
359 }
360 
361 static void raid10_end_read_request(struct bio *bio, int error)
362 {
363 	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
364 	struct r10bio *r10_bio = bio->bi_private;
365 	int slot, dev;
366 	struct md_rdev *rdev;
367 	struct r10conf *conf = r10_bio->mddev->private;
368 
369 
370 	slot = r10_bio->read_slot;
371 	dev = r10_bio->devs[slot].devnum;
372 	rdev = r10_bio->devs[slot].rdev;
373 	/*
374 	 * this branch is our 'one mirror IO has finished' event handler:
375 	 */
376 	update_head_pos(slot, r10_bio);
377 
378 	if (uptodate) {
379 		/*
380 		 * Set R10BIO_Uptodate in our master bio, so that
381 		 * we will return a good error code to the higher
382 		 * levels even if IO on some other mirrored buffer fails.
383 		 *
384 		 * The 'master' represents the composite IO operation to
385 		 * user-side. So if something waits for IO, then it will
386 		 * wait for the 'master' bio.
387 		 */
388 		set_bit(R10BIO_Uptodate, &r10_bio->state);
389 	} else {
390 		/* If all other devices that store this block have
391 		 * failed, we want to return the error upwards rather
392 		 * than fail the last device.  Here we redefine
393 		 * "uptodate" to mean "Don't want to retry"
394 		 */
395 		unsigned long flags;
396 		spin_lock_irqsave(&conf->device_lock, flags);
397 		if (!enough(conf, rdev->raid_disk))
398 			uptodate = 1;
399 		spin_unlock_irqrestore(&conf->device_lock, flags);
400 	}
401 	if (uptodate) {
402 		raid_end_bio_io(r10_bio);
403 		rdev_dec_pending(rdev, conf->mddev);
404 	} else {
405 		/*
406 		 * oops, read error - keep the refcount on the rdev
407 		 */
408 		char b[BDEVNAME_SIZE];
409 		printk_ratelimited(KERN_ERR
410 				   "md/raid10:%s: %s: rescheduling sector %llu\n",
411 				   mdname(conf->mddev),
412 				   bdevname(rdev->bdev, b),
413 				   (unsigned long long)r10_bio->sector);
414 		set_bit(R10BIO_ReadError, &r10_bio->state);
415 		reschedule_retry(r10_bio);
416 	}
417 }
418 
419 static void close_write(struct r10bio *r10_bio)
420 {
421 	/* clear the bitmap if all writes complete successfully */
422 	bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
423 			r10_bio->sectors,
424 			!test_bit(R10BIO_Degraded, &r10_bio->state),
425 			0);
426 	md_write_end(r10_bio->mddev);
427 }
428 
429 static void one_write_done(struct r10bio *r10_bio)
430 {
431 	if (atomic_dec_and_test(&r10_bio->remaining)) {
432 		if (test_bit(R10BIO_WriteError, &r10_bio->state))
433 			reschedule_retry(r10_bio);
434 		else {
435 			close_write(r10_bio);
436 			if (test_bit(R10BIO_MadeGood, &r10_bio->state))
437 				reschedule_retry(r10_bio);
438 			else
439 				raid_end_bio_io(r10_bio);
440 		}
441 	}
442 }
443 
444 static void raid10_end_write_request(struct bio *bio, int error)
445 {
446 	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
447 	struct r10bio *r10_bio = bio->bi_private;
448 	int dev;
449 	int dec_rdev = 1;
450 	struct r10conf *conf = r10_bio->mddev->private;
451 	int slot, repl;
452 	struct md_rdev *rdev = NULL;
453 
454 	dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
455 
456 	if (repl)
457 		rdev = conf->mirrors[dev].replacement;
458 	if (!rdev) {
459 		smp_rmb();
460 		repl = 0;
461 		rdev = conf->mirrors[dev].rdev;
462 	}
463 	/*
464 	 * this branch is our 'one mirror IO has finished' event handler:
465 	 */
466 	if (!uptodate) {
467 		if (repl)
468 			/* Never record new bad blocks to replacement,
469 			 * just fail it.
470 			 */
471 			md_error(rdev->mddev, rdev);
472 		else {
473 			set_bit(WriteErrorSeen,	&rdev->flags);
474 			if (!test_and_set_bit(WantReplacement, &rdev->flags))
475 				set_bit(MD_RECOVERY_NEEDED,
476 					&rdev->mddev->recovery);
477 			set_bit(R10BIO_WriteError, &r10_bio->state);
478 			dec_rdev = 0;
479 		}
480 	} else {
481 		/*
482 		 * Set R10BIO_Uptodate in our master bio, so that
483 		 * we will return a good error code for to the higher
484 		 * levels even if IO on some other mirrored buffer fails.
485 		 *
486 		 * The 'master' represents the composite IO operation to
487 		 * user-side. So if something waits for IO, then it will
488 		 * wait for the 'master' bio.
489 		 */
490 		sector_t first_bad;
491 		int bad_sectors;
492 
493 		set_bit(R10BIO_Uptodate, &r10_bio->state);
494 
495 		/* Maybe we can clear some bad blocks. */
496 		if (is_badblock(rdev,
497 				r10_bio->devs[slot].addr,
498 				r10_bio->sectors,
499 				&first_bad, &bad_sectors)) {
500 			bio_put(bio);
501 			if (repl)
502 				r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
503 			else
504 				r10_bio->devs[slot].bio = IO_MADE_GOOD;
505 			dec_rdev = 0;
506 			set_bit(R10BIO_MadeGood, &r10_bio->state);
507 		}
508 	}
509 
510 	/*
511 	 *
512 	 * Let's see if all mirrored write operations have finished
513 	 * already.
514 	 */
515 	one_write_done(r10_bio);
516 	if (dec_rdev)
517 		rdev_dec_pending(rdev, conf->mddev);
518 }
519 
520 /*
521  * RAID10 layout manager
522  * As well as the chunksize and raid_disks count, there are two
523  * parameters: near_copies and far_copies.
524  * near_copies * far_copies must be <= raid_disks.
525  * Normally one of these will be 1.
526  * If both are 1, we get raid0.
527  * If near_copies == raid_disks, we get raid1.
528  *
529  * Chunks are laid out in raid0 style with near_copies copies of the
530  * first chunk, followed by near_copies copies of the next chunk and
531  * so on.
532  * If far_copies > 1, then after 1/far_copies of the array has been assigned
533  * as described above, we start again with a device offset of near_copies.
534  * So we effectively have another copy of the whole array further down all
535  * the drives, but with blocks on different drives.
536  * With this layout, and block is never stored twice on the one device.
537  *
538  * raid10_find_phys finds the sector offset of a given virtual sector
539  * on each device that it is on.
540  *
541  * raid10_find_virt does the reverse mapping, from a device and a
542  * sector offset to a virtual address
543  */
544 
545 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
546 {
547 	int n,f;
548 	sector_t sector;
549 	sector_t chunk;
550 	sector_t stripe;
551 	int dev;
552 	int slot = 0;
553 	int last_far_set_start, last_far_set_size;
554 
555 	last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
556 	last_far_set_start *= geo->far_set_size;
557 
558 	last_far_set_size = geo->far_set_size;
559 	last_far_set_size += (geo->raid_disks % geo->far_set_size);
560 
561 	/* now calculate first sector/dev */
562 	chunk = r10bio->sector >> geo->chunk_shift;
563 	sector = r10bio->sector & geo->chunk_mask;
564 
565 	chunk *= geo->near_copies;
566 	stripe = chunk;
567 	dev = sector_div(stripe, geo->raid_disks);
568 	if (geo->far_offset)
569 		stripe *= geo->far_copies;
570 
571 	sector += stripe << geo->chunk_shift;
572 
573 	/* and calculate all the others */
574 	for (n = 0; n < geo->near_copies; n++) {
575 		int d = dev;
576 		int set;
577 		sector_t s = sector;
578 		r10bio->devs[slot].devnum = d;
579 		r10bio->devs[slot].addr = s;
580 		slot++;
581 
582 		for (f = 1; f < geo->far_copies; f++) {
583 			set = d / geo->far_set_size;
584 			d += geo->near_copies;
585 
586 			if ((geo->raid_disks % geo->far_set_size) &&
587 			    (d > last_far_set_start)) {
588 				d -= last_far_set_start;
589 				d %= last_far_set_size;
590 				d += last_far_set_start;
591 			} else {
592 				d %= geo->far_set_size;
593 				d += geo->far_set_size * set;
594 			}
595 			s += geo->stride;
596 			r10bio->devs[slot].devnum = d;
597 			r10bio->devs[slot].addr = s;
598 			slot++;
599 		}
600 		dev++;
601 		if (dev >= geo->raid_disks) {
602 			dev = 0;
603 			sector += (geo->chunk_mask + 1);
604 		}
605 	}
606 }
607 
608 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
609 {
610 	struct geom *geo = &conf->geo;
611 
612 	if (conf->reshape_progress != MaxSector &&
613 	    ((r10bio->sector >= conf->reshape_progress) !=
614 	     conf->mddev->reshape_backwards)) {
615 		set_bit(R10BIO_Previous, &r10bio->state);
616 		geo = &conf->prev;
617 	} else
618 		clear_bit(R10BIO_Previous, &r10bio->state);
619 
620 	__raid10_find_phys(geo, r10bio);
621 }
622 
623 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
624 {
625 	sector_t offset, chunk, vchunk;
626 	/* Never use conf->prev as this is only called during resync
627 	 * or recovery, so reshape isn't happening
628 	 */
629 	struct geom *geo = &conf->geo;
630 	int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
631 	int far_set_size = geo->far_set_size;
632 	int last_far_set_start;
633 
634 	if (geo->raid_disks % geo->far_set_size) {
635 		last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
636 		last_far_set_start *= geo->far_set_size;
637 
638 		if (dev >= last_far_set_start) {
639 			far_set_size = geo->far_set_size;
640 			far_set_size += (geo->raid_disks % geo->far_set_size);
641 			far_set_start = last_far_set_start;
642 		}
643 	}
644 
645 	offset = sector & geo->chunk_mask;
646 	if (geo->far_offset) {
647 		int fc;
648 		chunk = sector >> geo->chunk_shift;
649 		fc = sector_div(chunk, geo->far_copies);
650 		dev -= fc * geo->near_copies;
651 		if (dev < far_set_start)
652 			dev += far_set_size;
653 	} else {
654 		while (sector >= geo->stride) {
655 			sector -= geo->stride;
656 			if (dev < (geo->near_copies + far_set_start))
657 				dev += far_set_size - geo->near_copies;
658 			else
659 				dev -= geo->near_copies;
660 		}
661 		chunk = sector >> geo->chunk_shift;
662 	}
663 	vchunk = chunk * geo->raid_disks + dev;
664 	sector_div(vchunk, geo->near_copies);
665 	return (vchunk << geo->chunk_shift) + offset;
666 }
667 
668 /**
669  *	raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
670  *	@q: request queue
671  *	@bvm: properties of new bio
672  *	@biovec: the request that could be merged to it.
673  *
674  *	Return amount of bytes we can accept at this offset
675  *	This requires checking for end-of-chunk if near_copies != raid_disks,
676  *	and for subordinate merge_bvec_fns if merge_check_needed.
677  */
678 static int raid10_mergeable_bvec(struct request_queue *q,
679 				 struct bvec_merge_data *bvm,
680 				 struct bio_vec *biovec)
681 {
682 	struct mddev *mddev = q->queuedata;
683 	struct r10conf *conf = mddev->private;
684 	sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
685 	int max;
686 	unsigned int chunk_sectors;
687 	unsigned int bio_sectors = bvm->bi_size >> 9;
688 	struct geom *geo = &conf->geo;
689 
690 	chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
691 	if (conf->reshape_progress != MaxSector &&
692 	    ((sector >= conf->reshape_progress) !=
693 	     conf->mddev->reshape_backwards))
694 		geo = &conf->prev;
695 
696 	if (geo->near_copies < geo->raid_disks) {
697 		max = (chunk_sectors - ((sector & (chunk_sectors - 1))
698 					+ bio_sectors)) << 9;
699 		if (max < 0)
700 			/* bio_add cannot handle a negative return */
701 			max = 0;
702 		if (max <= biovec->bv_len && bio_sectors == 0)
703 			return biovec->bv_len;
704 	} else
705 		max = biovec->bv_len;
706 
707 	if (mddev->merge_check_needed) {
708 		struct {
709 			struct r10bio r10_bio;
710 			struct r10dev devs[conf->copies];
711 		} on_stack;
712 		struct r10bio *r10_bio = &on_stack.r10_bio;
713 		int s;
714 		if (conf->reshape_progress != MaxSector) {
715 			/* Cannot give any guidance during reshape */
716 			if (max <= biovec->bv_len && bio_sectors == 0)
717 				return biovec->bv_len;
718 			return 0;
719 		}
720 		r10_bio->sector = sector;
721 		raid10_find_phys(conf, r10_bio);
722 		rcu_read_lock();
723 		for (s = 0; s < conf->copies; s++) {
724 			int disk = r10_bio->devs[s].devnum;
725 			struct md_rdev *rdev = rcu_dereference(
726 				conf->mirrors[disk].rdev);
727 			if (rdev && !test_bit(Faulty, &rdev->flags)) {
728 				struct request_queue *q =
729 					bdev_get_queue(rdev->bdev);
730 				if (q->merge_bvec_fn) {
731 					bvm->bi_sector = r10_bio->devs[s].addr
732 						+ rdev->data_offset;
733 					bvm->bi_bdev = rdev->bdev;
734 					max = min(max, q->merge_bvec_fn(
735 							  q, bvm, biovec));
736 				}
737 			}
738 			rdev = rcu_dereference(conf->mirrors[disk].replacement);
739 			if (rdev && !test_bit(Faulty, &rdev->flags)) {
740 				struct request_queue *q =
741 					bdev_get_queue(rdev->bdev);
742 				if (q->merge_bvec_fn) {
743 					bvm->bi_sector = r10_bio->devs[s].addr
744 						+ rdev->data_offset;
745 					bvm->bi_bdev = rdev->bdev;
746 					max = min(max, q->merge_bvec_fn(
747 							  q, bvm, biovec));
748 				}
749 			}
750 		}
751 		rcu_read_unlock();
752 	}
753 	return max;
754 }
755 
756 /*
757  * This routine returns the disk from which the requested read should
758  * be done. There is a per-array 'next expected sequential IO' sector
759  * number - if this matches on the next IO then we use the last disk.
760  * There is also a per-disk 'last know head position' sector that is
761  * maintained from IRQ contexts, both the normal and the resync IO
762  * completion handlers update this position correctly. If there is no
763  * perfect sequential match then we pick the disk whose head is closest.
764  *
765  * If there are 2 mirrors in the same 2 devices, performance degrades
766  * because position is mirror, not device based.
767  *
768  * The rdev for the device selected will have nr_pending incremented.
769  */
770 
771 /*
772  * FIXME: possibly should rethink readbalancing and do it differently
773  * depending on near_copies / far_copies geometry.
774  */
775 static struct md_rdev *read_balance(struct r10conf *conf,
776 				    struct r10bio *r10_bio,
777 				    int *max_sectors)
778 {
779 	const sector_t this_sector = r10_bio->sector;
780 	int disk, slot;
781 	int sectors = r10_bio->sectors;
782 	int best_good_sectors;
783 	sector_t new_distance, best_dist;
784 	struct md_rdev *best_rdev, *rdev = NULL;
785 	int do_balance;
786 	int best_slot;
787 	struct geom *geo = &conf->geo;
788 
789 	raid10_find_phys(conf, r10_bio);
790 	rcu_read_lock();
791 retry:
792 	sectors = r10_bio->sectors;
793 	best_slot = -1;
794 	best_rdev = NULL;
795 	best_dist = MaxSector;
796 	best_good_sectors = 0;
797 	do_balance = 1;
798 	/*
799 	 * Check if we can balance. We can balance on the whole
800 	 * device if no resync is going on (recovery is ok), or below
801 	 * the resync window. We take the first readable disk when
802 	 * above the resync window.
803 	 */
804 	if (conf->mddev->recovery_cp < MaxSector
805 	    && (this_sector + sectors >= conf->next_resync))
806 		do_balance = 0;
807 
808 	for (slot = 0; slot < conf->copies ; slot++) {
809 		sector_t first_bad;
810 		int bad_sectors;
811 		sector_t dev_sector;
812 
813 		if (r10_bio->devs[slot].bio == IO_BLOCKED)
814 			continue;
815 		disk = r10_bio->devs[slot].devnum;
816 		rdev = rcu_dereference(conf->mirrors[disk].replacement);
817 		if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
818 		    test_bit(Unmerged, &rdev->flags) ||
819 		    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
820 			rdev = rcu_dereference(conf->mirrors[disk].rdev);
821 		if (rdev == NULL ||
822 		    test_bit(Faulty, &rdev->flags) ||
823 		    test_bit(Unmerged, &rdev->flags))
824 			continue;
825 		if (!test_bit(In_sync, &rdev->flags) &&
826 		    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
827 			continue;
828 
829 		dev_sector = r10_bio->devs[slot].addr;
830 		if (is_badblock(rdev, dev_sector, sectors,
831 				&first_bad, &bad_sectors)) {
832 			if (best_dist < MaxSector)
833 				/* Already have a better slot */
834 				continue;
835 			if (first_bad <= dev_sector) {
836 				/* Cannot read here.  If this is the
837 				 * 'primary' device, then we must not read
838 				 * beyond 'bad_sectors' from another device.
839 				 */
840 				bad_sectors -= (dev_sector - first_bad);
841 				if (!do_balance && sectors > bad_sectors)
842 					sectors = bad_sectors;
843 				if (best_good_sectors > sectors)
844 					best_good_sectors = sectors;
845 			} else {
846 				sector_t good_sectors =
847 					first_bad - dev_sector;
848 				if (good_sectors > best_good_sectors) {
849 					best_good_sectors = good_sectors;
850 					best_slot = slot;
851 					best_rdev = rdev;
852 				}
853 				if (!do_balance)
854 					/* Must read from here */
855 					break;
856 			}
857 			continue;
858 		} else
859 			best_good_sectors = sectors;
860 
861 		if (!do_balance)
862 			break;
863 
864 		/* This optimisation is debatable, and completely destroys
865 		 * sequential read speed for 'far copies' arrays.  So only
866 		 * keep it for 'near' arrays, and review those later.
867 		 */
868 		if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
869 			break;
870 
871 		/* for far > 1 always use the lowest address */
872 		if (geo->far_copies > 1)
873 			new_distance = r10_bio->devs[slot].addr;
874 		else
875 			new_distance = abs(r10_bio->devs[slot].addr -
876 					   conf->mirrors[disk].head_position);
877 		if (new_distance < best_dist) {
878 			best_dist = new_distance;
879 			best_slot = slot;
880 			best_rdev = rdev;
881 		}
882 	}
883 	if (slot >= conf->copies) {
884 		slot = best_slot;
885 		rdev = best_rdev;
886 	}
887 
888 	if (slot >= 0) {
889 		atomic_inc(&rdev->nr_pending);
890 		if (test_bit(Faulty, &rdev->flags)) {
891 			/* Cannot risk returning a device that failed
892 			 * before we inc'ed nr_pending
893 			 */
894 			rdev_dec_pending(rdev, conf->mddev);
895 			goto retry;
896 		}
897 		r10_bio->read_slot = slot;
898 	} else
899 		rdev = NULL;
900 	rcu_read_unlock();
901 	*max_sectors = best_good_sectors;
902 
903 	return rdev;
904 }
905 
906 int md_raid10_congested(struct mddev *mddev, int bits)
907 {
908 	struct r10conf *conf = mddev->private;
909 	int i, ret = 0;
910 
911 	if ((bits & (1 << BDI_async_congested)) &&
912 	    conf->pending_count >= max_queued_requests)
913 		return 1;
914 
915 	rcu_read_lock();
916 	for (i = 0;
917 	     (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
918 		     && ret == 0;
919 	     i++) {
920 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
921 		if (rdev && !test_bit(Faulty, &rdev->flags)) {
922 			struct request_queue *q = bdev_get_queue(rdev->bdev);
923 
924 			ret |= bdi_congested(&q->backing_dev_info, bits);
925 		}
926 	}
927 	rcu_read_unlock();
928 	return ret;
929 }
930 EXPORT_SYMBOL_GPL(md_raid10_congested);
931 
932 static int raid10_congested(void *data, int bits)
933 {
934 	struct mddev *mddev = data;
935 
936 	return mddev_congested(mddev, bits) ||
937 		md_raid10_congested(mddev, bits);
938 }
939 
940 static void flush_pending_writes(struct r10conf *conf)
941 {
942 	/* Any writes that have been queued but are awaiting
943 	 * bitmap updates get flushed here.
944 	 */
945 	spin_lock_irq(&conf->device_lock);
946 
947 	if (conf->pending_bio_list.head) {
948 		struct bio *bio;
949 		bio = bio_list_get(&conf->pending_bio_list);
950 		conf->pending_count = 0;
951 		spin_unlock_irq(&conf->device_lock);
952 		/* flush any pending bitmap writes to disk
953 		 * before proceeding w/ I/O */
954 		bitmap_unplug(conf->mddev->bitmap);
955 		wake_up(&conf->wait_barrier);
956 
957 		while (bio) { /* submit pending writes */
958 			struct bio *next = bio->bi_next;
959 			bio->bi_next = NULL;
960 			if (unlikely((bio->bi_rw & REQ_DISCARD) &&
961 			    !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
962 				/* Just ignore it */
963 				bio_endio(bio, 0);
964 			else
965 				generic_make_request(bio);
966 			bio = next;
967 		}
968 	} else
969 		spin_unlock_irq(&conf->device_lock);
970 }
971 
972 /* Barriers....
973  * Sometimes we need to suspend IO while we do something else,
974  * either some resync/recovery, or reconfigure the array.
975  * To do this we raise a 'barrier'.
976  * The 'barrier' is a counter that can be raised multiple times
977  * to count how many activities are happening which preclude
978  * normal IO.
979  * We can only raise the barrier if there is no pending IO.
980  * i.e. if nr_pending == 0.
981  * We choose only to raise the barrier if no-one is waiting for the
982  * barrier to go down.  This means that as soon as an IO request
983  * is ready, no other operations which require a barrier will start
984  * until the IO request has had a chance.
985  *
986  * So: regular IO calls 'wait_barrier'.  When that returns there
987  *    is no backgroup IO happening,  It must arrange to call
988  *    allow_barrier when it has finished its IO.
989  * backgroup IO calls must call raise_barrier.  Once that returns
990  *    there is no normal IO happeing.  It must arrange to call
991  *    lower_barrier when the particular background IO completes.
992  */
993 
994 static void raise_barrier(struct r10conf *conf, int force)
995 {
996 	BUG_ON(force && !conf->barrier);
997 	spin_lock_irq(&conf->resync_lock);
998 
999 	/* Wait until no block IO is waiting (unless 'force') */
1000 	wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
1001 			    conf->resync_lock);
1002 
1003 	/* block any new IO from starting */
1004 	conf->barrier++;
1005 
1006 	/* Now wait for all pending IO to complete */
1007 	wait_event_lock_irq(conf->wait_barrier,
1008 			    !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
1009 			    conf->resync_lock);
1010 
1011 	spin_unlock_irq(&conf->resync_lock);
1012 }
1013 
1014 static void lower_barrier(struct r10conf *conf)
1015 {
1016 	unsigned long flags;
1017 	spin_lock_irqsave(&conf->resync_lock, flags);
1018 	conf->barrier--;
1019 	spin_unlock_irqrestore(&conf->resync_lock, flags);
1020 	wake_up(&conf->wait_barrier);
1021 }
1022 
1023 static void wait_barrier(struct r10conf *conf)
1024 {
1025 	spin_lock_irq(&conf->resync_lock);
1026 	if (conf->barrier) {
1027 		conf->nr_waiting++;
1028 		/* Wait for the barrier to drop.
1029 		 * However if there are already pending
1030 		 * requests (preventing the barrier from
1031 		 * rising completely), and the
1032 		 * pre-process bio queue isn't empty,
1033 		 * then don't wait, as we need to empty
1034 		 * that queue to get the nr_pending
1035 		 * count down.
1036 		 */
1037 		wait_event_lock_irq(conf->wait_barrier,
1038 				    !conf->barrier ||
1039 				    (conf->nr_pending &&
1040 				     current->bio_list &&
1041 				     !bio_list_empty(current->bio_list)),
1042 				    conf->resync_lock);
1043 		conf->nr_waiting--;
1044 	}
1045 	conf->nr_pending++;
1046 	spin_unlock_irq(&conf->resync_lock);
1047 }
1048 
1049 static void allow_barrier(struct r10conf *conf)
1050 {
1051 	unsigned long flags;
1052 	spin_lock_irqsave(&conf->resync_lock, flags);
1053 	conf->nr_pending--;
1054 	spin_unlock_irqrestore(&conf->resync_lock, flags);
1055 	wake_up(&conf->wait_barrier);
1056 }
1057 
1058 static void freeze_array(struct r10conf *conf)
1059 {
1060 	/* stop syncio and normal IO and wait for everything to
1061 	 * go quiet.
1062 	 * We increment barrier and nr_waiting, and then
1063 	 * wait until nr_pending match nr_queued+1
1064 	 * This is called in the context of one normal IO request
1065 	 * that has failed. Thus any sync request that might be pending
1066 	 * will be blocked by nr_pending, and we need to wait for
1067 	 * pending IO requests to complete or be queued for re-try.
1068 	 * Thus the number queued (nr_queued) plus this request (1)
1069 	 * must match the number of pending IOs (nr_pending) before
1070 	 * we continue.
1071 	 */
1072 	spin_lock_irq(&conf->resync_lock);
1073 	conf->barrier++;
1074 	conf->nr_waiting++;
1075 	wait_event_lock_irq_cmd(conf->wait_barrier,
1076 				conf->nr_pending == conf->nr_queued+1,
1077 				conf->resync_lock,
1078 				flush_pending_writes(conf));
1079 
1080 	spin_unlock_irq(&conf->resync_lock);
1081 }
1082 
1083 static void unfreeze_array(struct r10conf *conf)
1084 {
1085 	/* reverse the effect of the freeze */
1086 	spin_lock_irq(&conf->resync_lock);
1087 	conf->barrier--;
1088 	conf->nr_waiting--;
1089 	wake_up(&conf->wait_barrier);
1090 	spin_unlock_irq(&conf->resync_lock);
1091 }
1092 
1093 static sector_t choose_data_offset(struct r10bio *r10_bio,
1094 				   struct md_rdev *rdev)
1095 {
1096 	if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1097 	    test_bit(R10BIO_Previous, &r10_bio->state))
1098 		return rdev->data_offset;
1099 	else
1100 		return rdev->new_data_offset;
1101 }
1102 
1103 struct raid10_plug_cb {
1104 	struct blk_plug_cb	cb;
1105 	struct bio_list		pending;
1106 	int			pending_cnt;
1107 };
1108 
1109 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1110 {
1111 	struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1112 						   cb);
1113 	struct mddev *mddev = plug->cb.data;
1114 	struct r10conf *conf = mddev->private;
1115 	struct bio *bio;
1116 
1117 	if (from_schedule || current->bio_list) {
1118 		spin_lock_irq(&conf->device_lock);
1119 		bio_list_merge(&conf->pending_bio_list, &plug->pending);
1120 		conf->pending_count += plug->pending_cnt;
1121 		spin_unlock_irq(&conf->device_lock);
1122 		wake_up(&conf->wait_barrier);
1123 		md_wakeup_thread(mddev->thread);
1124 		kfree(plug);
1125 		return;
1126 	}
1127 
1128 	/* we aren't scheduling, so we can do the write-out directly. */
1129 	bio = bio_list_get(&plug->pending);
1130 	bitmap_unplug(mddev->bitmap);
1131 	wake_up(&conf->wait_barrier);
1132 
1133 	while (bio) { /* submit pending writes */
1134 		struct bio *next = bio->bi_next;
1135 		bio->bi_next = NULL;
1136 		if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1137 		    !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1138 			/* Just ignore it */
1139 			bio_endio(bio, 0);
1140 		else
1141 			generic_make_request(bio);
1142 		bio = next;
1143 	}
1144 	kfree(plug);
1145 }
1146 
1147 static void make_request(struct mddev *mddev, struct bio * bio)
1148 {
1149 	struct r10conf *conf = mddev->private;
1150 	struct r10bio *r10_bio;
1151 	struct bio *read_bio;
1152 	int i;
1153 	sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1154 	int chunk_sects = chunk_mask + 1;
1155 	const int rw = bio_data_dir(bio);
1156 	const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1157 	const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1158 	const unsigned long do_discard = (bio->bi_rw
1159 					  & (REQ_DISCARD | REQ_SECURE));
1160 	const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1161 	unsigned long flags;
1162 	struct md_rdev *blocked_rdev;
1163 	struct blk_plug_cb *cb;
1164 	struct raid10_plug_cb *plug = NULL;
1165 	int sectors_handled;
1166 	int max_sectors;
1167 	int sectors;
1168 
1169 	if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1170 		md_flush_request(mddev, bio);
1171 		return;
1172 	}
1173 
1174 	/* If this request crosses a chunk boundary, we need to
1175 	 * split it.  This will only happen for 1 PAGE (or less) requests.
1176 	 */
1177 	if (unlikely((bio->bi_sector & chunk_mask) + bio_sectors(bio)
1178 		     > chunk_sects
1179 		     && (conf->geo.near_copies < conf->geo.raid_disks
1180 			 || conf->prev.near_copies < conf->prev.raid_disks))) {
1181 		struct bio_pair *bp;
1182 		/* Sanity check -- queue functions should prevent this happening */
1183 		if (bio_segments(bio) > 1)
1184 			goto bad_map;
1185 		/* This is a one page bio that upper layers
1186 		 * refuse to split for us, so we need to split it.
1187 		 */
1188 		bp = bio_split(bio,
1189 			       chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1190 
1191 		/* Each of these 'make_request' calls will call 'wait_barrier'.
1192 		 * If the first succeeds but the second blocks due to the resync
1193 		 * thread raising the barrier, we will deadlock because the
1194 		 * IO to the underlying device will be queued in generic_make_request
1195 		 * and will never complete, so will never reduce nr_pending.
1196 		 * So increment nr_waiting here so no new raise_barriers will
1197 		 * succeed, and so the second wait_barrier cannot block.
1198 		 */
1199 		spin_lock_irq(&conf->resync_lock);
1200 		conf->nr_waiting++;
1201 		spin_unlock_irq(&conf->resync_lock);
1202 
1203 		make_request(mddev, &bp->bio1);
1204 		make_request(mddev, &bp->bio2);
1205 
1206 		spin_lock_irq(&conf->resync_lock);
1207 		conf->nr_waiting--;
1208 		wake_up(&conf->wait_barrier);
1209 		spin_unlock_irq(&conf->resync_lock);
1210 
1211 		bio_pair_release(bp);
1212 		return;
1213 	bad_map:
1214 		printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1215 		       " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1216 		       (unsigned long long)bio->bi_sector, bio_sectors(bio) / 2);
1217 
1218 		bio_io_error(bio);
1219 		return;
1220 	}
1221 
1222 	md_write_start(mddev, bio);
1223 
1224 	/*
1225 	 * Register the new request and wait if the reconstruction
1226 	 * thread has put up a bar for new requests.
1227 	 * Continue immediately if no resync is active currently.
1228 	 */
1229 	wait_barrier(conf);
1230 
1231 	sectors = bio_sectors(bio);
1232 	while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1233 	    bio->bi_sector < conf->reshape_progress &&
1234 	    bio->bi_sector + sectors > conf->reshape_progress) {
1235 		/* IO spans the reshape position.  Need to wait for
1236 		 * reshape to pass
1237 		 */
1238 		allow_barrier(conf);
1239 		wait_event(conf->wait_barrier,
1240 			   conf->reshape_progress <= bio->bi_sector ||
1241 			   conf->reshape_progress >= bio->bi_sector + sectors);
1242 		wait_barrier(conf);
1243 	}
1244 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1245 	    bio_data_dir(bio) == WRITE &&
1246 	    (mddev->reshape_backwards
1247 	     ? (bio->bi_sector < conf->reshape_safe &&
1248 		bio->bi_sector + sectors > conf->reshape_progress)
1249 	     : (bio->bi_sector + sectors > conf->reshape_safe &&
1250 		bio->bi_sector < conf->reshape_progress))) {
1251 		/* Need to update reshape_position in metadata */
1252 		mddev->reshape_position = conf->reshape_progress;
1253 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
1254 		set_bit(MD_CHANGE_PENDING, &mddev->flags);
1255 		md_wakeup_thread(mddev->thread);
1256 		wait_event(mddev->sb_wait,
1257 			   !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1258 
1259 		conf->reshape_safe = mddev->reshape_position;
1260 	}
1261 
1262 	r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1263 
1264 	r10_bio->master_bio = bio;
1265 	r10_bio->sectors = sectors;
1266 
1267 	r10_bio->mddev = mddev;
1268 	r10_bio->sector = bio->bi_sector;
1269 	r10_bio->state = 0;
1270 
1271 	/* We might need to issue multiple reads to different
1272 	 * devices if there are bad blocks around, so we keep
1273 	 * track of the number of reads in bio->bi_phys_segments.
1274 	 * If this is 0, there is only one r10_bio and no locking
1275 	 * will be needed when the request completes.  If it is
1276 	 * non-zero, then it is the number of not-completed requests.
1277 	 */
1278 	bio->bi_phys_segments = 0;
1279 	clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1280 
1281 	if (rw == READ) {
1282 		/*
1283 		 * read balancing logic:
1284 		 */
1285 		struct md_rdev *rdev;
1286 		int slot;
1287 
1288 read_again:
1289 		rdev = read_balance(conf, r10_bio, &max_sectors);
1290 		if (!rdev) {
1291 			raid_end_bio_io(r10_bio);
1292 			return;
1293 		}
1294 		slot = r10_bio->read_slot;
1295 
1296 		read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1297 		md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1298 			    max_sectors);
1299 
1300 		r10_bio->devs[slot].bio = read_bio;
1301 		r10_bio->devs[slot].rdev = rdev;
1302 
1303 		read_bio->bi_sector = r10_bio->devs[slot].addr +
1304 			choose_data_offset(r10_bio, rdev);
1305 		read_bio->bi_bdev = rdev->bdev;
1306 		read_bio->bi_end_io = raid10_end_read_request;
1307 		read_bio->bi_rw = READ | do_sync;
1308 		read_bio->bi_private = r10_bio;
1309 
1310 		if (max_sectors < r10_bio->sectors) {
1311 			/* Could not read all from this device, so we will
1312 			 * need another r10_bio.
1313 			 */
1314 			sectors_handled = (r10_bio->sectors + max_sectors
1315 					   - bio->bi_sector);
1316 			r10_bio->sectors = max_sectors;
1317 			spin_lock_irq(&conf->device_lock);
1318 			if (bio->bi_phys_segments == 0)
1319 				bio->bi_phys_segments = 2;
1320 			else
1321 				bio->bi_phys_segments++;
1322 			spin_unlock(&conf->device_lock);
1323 			/* Cannot call generic_make_request directly
1324 			 * as that will be queued in __generic_make_request
1325 			 * and subsequent mempool_alloc might block
1326 			 * waiting for it.  so hand bio over to raid10d.
1327 			 */
1328 			reschedule_retry(r10_bio);
1329 
1330 			r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1331 
1332 			r10_bio->master_bio = bio;
1333 			r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1334 			r10_bio->state = 0;
1335 			r10_bio->mddev = mddev;
1336 			r10_bio->sector = bio->bi_sector + sectors_handled;
1337 			goto read_again;
1338 		} else
1339 			generic_make_request(read_bio);
1340 		return;
1341 	}
1342 
1343 	/*
1344 	 * WRITE:
1345 	 */
1346 	if (conf->pending_count >= max_queued_requests) {
1347 		md_wakeup_thread(mddev->thread);
1348 		wait_event(conf->wait_barrier,
1349 			   conf->pending_count < max_queued_requests);
1350 	}
1351 	/* first select target devices under rcu_lock and
1352 	 * inc refcount on their rdev.  Record them by setting
1353 	 * bios[x] to bio
1354 	 * If there are known/acknowledged bad blocks on any device
1355 	 * on which we have seen a write error, we want to avoid
1356 	 * writing to those blocks.  This potentially requires several
1357 	 * writes to write around the bad blocks.  Each set of writes
1358 	 * gets its own r10_bio with a set of bios attached.  The number
1359 	 * of r10_bios is recored in bio->bi_phys_segments just as with
1360 	 * the read case.
1361 	 */
1362 
1363 	r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1364 	raid10_find_phys(conf, r10_bio);
1365 retry_write:
1366 	blocked_rdev = NULL;
1367 	rcu_read_lock();
1368 	max_sectors = r10_bio->sectors;
1369 
1370 	for (i = 0;  i < conf->copies; i++) {
1371 		int d = r10_bio->devs[i].devnum;
1372 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1373 		struct md_rdev *rrdev = rcu_dereference(
1374 			conf->mirrors[d].replacement);
1375 		if (rdev == rrdev)
1376 			rrdev = NULL;
1377 		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1378 			atomic_inc(&rdev->nr_pending);
1379 			blocked_rdev = rdev;
1380 			break;
1381 		}
1382 		if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1383 			atomic_inc(&rrdev->nr_pending);
1384 			blocked_rdev = rrdev;
1385 			break;
1386 		}
1387 		if (rdev && (test_bit(Faulty, &rdev->flags)
1388 			     || test_bit(Unmerged, &rdev->flags)))
1389 			rdev = NULL;
1390 		if (rrdev && (test_bit(Faulty, &rrdev->flags)
1391 			      || test_bit(Unmerged, &rrdev->flags)))
1392 			rrdev = NULL;
1393 
1394 		r10_bio->devs[i].bio = NULL;
1395 		r10_bio->devs[i].repl_bio = NULL;
1396 
1397 		if (!rdev && !rrdev) {
1398 			set_bit(R10BIO_Degraded, &r10_bio->state);
1399 			continue;
1400 		}
1401 		if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1402 			sector_t first_bad;
1403 			sector_t dev_sector = r10_bio->devs[i].addr;
1404 			int bad_sectors;
1405 			int is_bad;
1406 
1407 			is_bad = is_badblock(rdev, dev_sector,
1408 					     max_sectors,
1409 					     &first_bad, &bad_sectors);
1410 			if (is_bad < 0) {
1411 				/* Mustn't write here until the bad block
1412 				 * is acknowledged
1413 				 */
1414 				atomic_inc(&rdev->nr_pending);
1415 				set_bit(BlockedBadBlocks, &rdev->flags);
1416 				blocked_rdev = rdev;
1417 				break;
1418 			}
1419 			if (is_bad && first_bad <= dev_sector) {
1420 				/* Cannot write here at all */
1421 				bad_sectors -= (dev_sector - first_bad);
1422 				if (bad_sectors < max_sectors)
1423 					/* Mustn't write more than bad_sectors
1424 					 * to other devices yet
1425 					 */
1426 					max_sectors = bad_sectors;
1427 				/* We don't set R10BIO_Degraded as that
1428 				 * only applies if the disk is missing,
1429 				 * so it might be re-added, and we want to
1430 				 * know to recover this chunk.
1431 				 * In this case the device is here, and the
1432 				 * fact that this chunk is not in-sync is
1433 				 * recorded in the bad block log.
1434 				 */
1435 				continue;
1436 			}
1437 			if (is_bad) {
1438 				int good_sectors = first_bad - dev_sector;
1439 				if (good_sectors < max_sectors)
1440 					max_sectors = good_sectors;
1441 			}
1442 		}
1443 		if (rdev) {
1444 			r10_bio->devs[i].bio = bio;
1445 			atomic_inc(&rdev->nr_pending);
1446 		}
1447 		if (rrdev) {
1448 			r10_bio->devs[i].repl_bio = bio;
1449 			atomic_inc(&rrdev->nr_pending);
1450 		}
1451 	}
1452 	rcu_read_unlock();
1453 
1454 	if (unlikely(blocked_rdev)) {
1455 		/* Have to wait for this device to get unblocked, then retry */
1456 		int j;
1457 		int d;
1458 
1459 		for (j = 0; j < i; j++) {
1460 			if (r10_bio->devs[j].bio) {
1461 				d = r10_bio->devs[j].devnum;
1462 				rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1463 			}
1464 			if (r10_bio->devs[j].repl_bio) {
1465 				struct md_rdev *rdev;
1466 				d = r10_bio->devs[j].devnum;
1467 				rdev = conf->mirrors[d].replacement;
1468 				if (!rdev) {
1469 					/* Race with remove_disk */
1470 					smp_mb();
1471 					rdev = conf->mirrors[d].rdev;
1472 				}
1473 				rdev_dec_pending(rdev, mddev);
1474 			}
1475 		}
1476 		allow_barrier(conf);
1477 		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1478 		wait_barrier(conf);
1479 		goto retry_write;
1480 	}
1481 
1482 	if (max_sectors < r10_bio->sectors) {
1483 		/* We are splitting this into multiple parts, so
1484 		 * we need to prepare for allocating another r10_bio.
1485 		 */
1486 		r10_bio->sectors = max_sectors;
1487 		spin_lock_irq(&conf->device_lock);
1488 		if (bio->bi_phys_segments == 0)
1489 			bio->bi_phys_segments = 2;
1490 		else
1491 			bio->bi_phys_segments++;
1492 		spin_unlock_irq(&conf->device_lock);
1493 	}
1494 	sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1495 
1496 	atomic_set(&r10_bio->remaining, 1);
1497 	bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1498 
1499 	for (i = 0; i < conf->copies; i++) {
1500 		struct bio *mbio;
1501 		int d = r10_bio->devs[i].devnum;
1502 		if (r10_bio->devs[i].bio) {
1503 			struct md_rdev *rdev = conf->mirrors[d].rdev;
1504 			mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1505 			md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1506 				    max_sectors);
1507 			r10_bio->devs[i].bio = mbio;
1508 
1509 			mbio->bi_sector	= (r10_bio->devs[i].addr+
1510 					   choose_data_offset(r10_bio,
1511 							      rdev));
1512 			mbio->bi_bdev = rdev->bdev;
1513 			mbio->bi_end_io	= raid10_end_write_request;
1514 			mbio->bi_rw =
1515 				WRITE | do_sync | do_fua | do_discard | do_same;
1516 			mbio->bi_private = r10_bio;
1517 
1518 			atomic_inc(&r10_bio->remaining);
1519 
1520 			cb = blk_check_plugged(raid10_unplug, mddev,
1521 					       sizeof(*plug));
1522 			if (cb)
1523 				plug = container_of(cb, struct raid10_plug_cb,
1524 						    cb);
1525 			else
1526 				plug = NULL;
1527 			spin_lock_irqsave(&conf->device_lock, flags);
1528 			if (plug) {
1529 				bio_list_add(&plug->pending, mbio);
1530 				plug->pending_cnt++;
1531 			} else {
1532 				bio_list_add(&conf->pending_bio_list, mbio);
1533 				conf->pending_count++;
1534 			}
1535 			spin_unlock_irqrestore(&conf->device_lock, flags);
1536 			if (!plug)
1537 				md_wakeup_thread(mddev->thread);
1538 		}
1539 
1540 		if (r10_bio->devs[i].repl_bio) {
1541 			struct md_rdev *rdev = conf->mirrors[d].replacement;
1542 			if (rdev == NULL) {
1543 				/* Replacement just got moved to main 'rdev' */
1544 				smp_mb();
1545 				rdev = conf->mirrors[d].rdev;
1546 			}
1547 			mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1548 			md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1549 				    max_sectors);
1550 			r10_bio->devs[i].repl_bio = mbio;
1551 
1552 			mbio->bi_sector	= (r10_bio->devs[i].addr +
1553 					   choose_data_offset(
1554 						   r10_bio, rdev));
1555 			mbio->bi_bdev = rdev->bdev;
1556 			mbio->bi_end_io	= raid10_end_write_request;
1557 			mbio->bi_rw =
1558 				WRITE | do_sync | do_fua | do_discard | do_same;
1559 			mbio->bi_private = r10_bio;
1560 
1561 			atomic_inc(&r10_bio->remaining);
1562 			spin_lock_irqsave(&conf->device_lock, flags);
1563 			bio_list_add(&conf->pending_bio_list, mbio);
1564 			conf->pending_count++;
1565 			spin_unlock_irqrestore(&conf->device_lock, flags);
1566 			if (!mddev_check_plugged(mddev))
1567 				md_wakeup_thread(mddev->thread);
1568 		}
1569 	}
1570 
1571 	/* Don't remove the bias on 'remaining' (one_write_done) until
1572 	 * after checking if we need to go around again.
1573 	 */
1574 
1575 	if (sectors_handled < bio_sectors(bio)) {
1576 		one_write_done(r10_bio);
1577 		/* We need another r10_bio.  It has already been counted
1578 		 * in bio->bi_phys_segments.
1579 		 */
1580 		r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1581 
1582 		r10_bio->master_bio = bio;
1583 		r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1584 
1585 		r10_bio->mddev = mddev;
1586 		r10_bio->sector = bio->bi_sector + sectors_handled;
1587 		r10_bio->state = 0;
1588 		goto retry_write;
1589 	}
1590 	one_write_done(r10_bio);
1591 
1592 	/* In case raid10d snuck in to freeze_array */
1593 	wake_up(&conf->wait_barrier);
1594 }
1595 
1596 static void status(struct seq_file *seq, struct mddev *mddev)
1597 {
1598 	struct r10conf *conf = mddev->private;
1599 	int i;
1600 
1601 	if (conf->geo.near_copies < conf->geo.raid_disks)
1602 		seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1603 	if (conf->geo.near_copies > 1)
1604 		seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1605 	if (conf->geo.far_copies > 1) {
1606 		if (conf->geo.far_offset)
1607 			seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1608 		else
1609 			seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1610 	}
1611 	seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1612 					conf->geo.raid_disks - mddev->degraded);
1613 	for (i = 0; i < conf->geo.raid_disks; i++)
1614 		seq_printf(seq, "%s",
1615 			      conf->mirrors[i].rdev &&
1616 			      test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1617 	seq_printf(seq, "]");
1618 }
1619 
1620 /* check if there are enough drives for
1621  * every block to appear on atleast one.
1622  * Don't consider the device numbered 'ignore'
1623  * as we might be about to remove it.
1624  */
1625 static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
1626 {
1627 	int first = 0;
1628 
1629 	do {
1630 		int n = conf->copies;
1631 		int cnt = 0;
1632 		int this = first;
1633 		while (n--) {
1634 			if (conf->mirrors[this].rdev &&
1635 			    this != ignore)
1636 				cnt++;
1637 			this = (this+1) % geo->raid_disks;
1638 		}
1639 		if (cnt == 0)
1640 			return 0;
1641 		first = (first + geo->near_copies) % geo->raid_disks;
1642 	} while (first != 0);
1643 	return 1;
1644 }
1645 
1646 static int enough(struct r10conf *conf, int ignore)
1647 {
1648 	return _enough(conf, &conf->geo, ignore) &&
1649 		_enough(conf, &conf->prev, ignore);
1650 }
1651 
1652 static void error(struct mddev *mddev, struct md_rdev *rdev)
1653 {
1654 	char b[BDEVNAME_SIZE];
1655 	struct r10conf *conf = mddev->private;
1656 
1657 	/*
1658 	 * If it is not operational, then we have already marked it as dead
1659 	 * else if it is the last working disks, ignore the error, let the
1660 	 * next level up know.
1661 	 * else mark the drive as failed
1662 	 */
1663 	if (test_bit(In_sync, &rdev->flags)
1664 	    && !enough(conf, rdev->raid_disk))
1665 		/*
1666 		 * Don't fail the drive, just return an IO error.
1667 		 */
1668 		return;
1669 	if (test_and_clear_bit(In_sync, &rdev->flags)) {
1670 		unsigned long flags;
1671 		spin_lock_irqsave(&conf->device_lock, flags);
1672 		mddev->degraded++;
1673 		spin_unlock_irqrestore(&conf->device_lock, flags);
1674 		/*
1675 		 * if recovery is running, make sure it aborts.
1676 		 */
1677 		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1678 	}
1679 	set_bit(Blocked, &rdev->flags);
1680 	set_bit(Faulty, &rdev->flags);
1681 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
1682 	printk(KERN_ALERT
1683 	       "md/raid10:%s: Disk failure on %s, disabling device.\n"
1684 	       "md/raid10:%s: Operation continuing on %d devices.\n",
1685 	       mdname(mddev), bdevname(rdev->bdev, b),
1686 	       mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1687 }
1688 
1689 static void print_conf(struct r10conf *conf)
1690 {
1691 	int i;
1692 	struct raid10_info *tmp;
1693 
1694 	printk(KERN_DEBUG "RAID10 conf printout:\n");
1695 	if (!conf) {
1696 		printk(KERN_DEBUG "(!conf)\n");
1697 		return;
1698 	}
1699 	printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1700 		conf->geo.raid_disks);
1701 
1702 	for (i = 0; i < conf->geo.raid_disks; i++) {
1703 		char b[BDEVNAME_SIZE];
1704 		tmp = conf->mirrors + i;
1705 		if (tmp->rdev)
1706 			printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1707 				i, !test_bit(In_sync, &tmp->rdev->flags),
1708 			        !test_bit(Faulty, &tmp->rdev->flags),
1709 				bdevname(tmp->rdev->bdev,b));
1710 	}
1711 }
1712 
1713 static void close_sync(struct r10conf *conf)
1714 {
1715 	wait_barrier(conf);
1716 	allow_barrier(conf);
1717 
1718 	mempool_destroy(conf->r10buf_pool);
1719 	conf->r10buf_pool = NULL;
1720 }
1721 
1722 static int raid10_spare_active(struct mddev *mddev)
1723 {
1724 	int i;
1725 	struct r10conf *conf = mddev->private;
1726 	struct raid10_info *tmp;
1727 	int count = 0;
1728 	unsigned long flags;
1729 
1730 	/*
1731 	 * Find all non-in_sync disks within the RAID10 configuration
1732 	 * and mark them in_sync
1733 	 */
1734 	for (i = 0; i < conf->geo.raid_disks; i++) {
1735 		tmp = conf->mirrors + i;
1736 		if (tmp->replacement
1737 		    && tmp->replacement->recovery_offset == MaxSector
1738 		    && !test_bit(Faulty, &tmp->replacement->flags)
1739 		    && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1740 			/* Replacement has just become active */
1741 			if (!tmp->rdev
1742 			    || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1743 				count++;
1744 			if (tmp->rdev) {
1745 				/* Replaced device not technically faulty,
1746 				 * but we need to be sure it gets removed
1747 				 * and never re-added.
1748 				 */
1749 				set_bit(Faulty, &tmp->rdev->flags);
1750 				sysfs_notify_dirent_safe(
1751 					tmp->rdev->sysfs_state);
1752 			}
1753 			sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1754 		} else if (tmp->rdev
1755 			   && !test_bit(Faulty, &tmp->rdev->flags)
1756 			   && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1757 			count++;
1758 			sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1759 		}
1760 	}
1761 	spin_lock_irqsave(&conf->device_lock, flags);
1762 	mddev->degraded -= count;
1763 	spin_unlock_irqrestore(&conf->device_lock, flags);
1764 
1765 	print_conf(conf);
1766 	return count;
1767 }
1768 
1769 
1770 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1771 {
1772 	struct r10conf *conf = mddev->private;
1773 	int err = -EEXIST;
1774 	int mirror;
1775 	int first = 0;
1776 	int last = conf->geo.raid_disks - 1;
1777 	struct request_queue *q = bdev_get_queue(rdev->bdev);
1778 
1779 	if (mddev->recovery_cp < MaxSector)
1780 		/* only hot-add to in-sync arrays, as recovery is
1781 		 * very different from resync
1782 		 */
1783 		return -EBUSY;
1784 	if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
1785 		return -EINVAL;
1786 
1787 	if (rdev->raid_disk >= 0)
1788 		first = last = rdev->raid_disk;
1789 
1790 	if (q->merge_bvec_fn) {
1791 		set_bit(Unmerged, &rdev->flags);
1792 		mddev->merge_check_needed = 1;
1793 	}
1794 
1795 	if (rdev->saved_raid_disk >= first &&
1796 	    conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1797 		mirror = rdev->saved_raid_disk;
1798 	else
1799 		mirror = first;
1800 	for ( ; mirror <= last ; mirror++) {
1801 		struct raid10_info *p = &conf->mirrors[mirror];
1802 		if (p->recovery_disabled == mddev->recovery_disabled)
1803 			continue;
1804 		if (p->rdev) {
1805 			if (!test_bit(WantReplacement, &p->rdev->flags) ||
1806 			    p->replacement != NULL)
1807 				continue;
1808 			clear_bit(In_sync, &rdev->flags);
1809 			set_bit(Replacement, &rdev->flags);
1810 			rdev->raid_disk = mirror;
1811 			err = 0;
1812 			disk_stack_limits(mddev->gendisk, rdev->bdev,
1813 					  rdev->data_offset << 9);
1814 			conf->fullsync = 1;
1815 			rcu_assign_pointer(p->replacement, rdev);
1816 			break;
1817 		}
1818 
1819 		disk_stack_limits(mddev->gendisk, rdev->bdev,
1820 				  rdev->data_offset << 9);
1821 
1822 		p->head_position = 0;
1823 		p->recovery_disabled = mddev->recovery_disabled - 1;
1824 		rdev->raid_disk = mirror;
1825 		err = 0;
1826 		if (rdev->saved_raid_disk != mirror)
1827 			conf->fullsync = 1;
1828 		rcu_assign_pointer(p->rdev, rdev);
1829 		break;
1830 	}
1831 	if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1832 		/* Some requests might not have seen this new
1833 		 * merge_bvec_fn.  We must wait for them to complete
1834 		 * before merging the device fully.
1835 		 * First we make sure any code which has tested
1836 		 * our function has submitted the request, then
1837 		 * we wait for all outstanding requests to complete.
1838 		 */
1839 		synchronize_sched();
1840 		raise_barrier(conf, 0);
1841 		lower_barrier(conf);
1842 		clear_bit(Unmerged, &rdev->flags);
1843 	}
1844 	md_integrity_add_rdev(rdev, mddev);
1845 	if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1846 		queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1847 
1848 	print_conf(conf);
1849 	return err;
1850 }
1851 
1852 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1853 {
1854 	struct r10conf *conf = mddev->private;
1855 	int err = 0;
1856 	int number = rdev->raid_disk;
1857 	struct md_rdev **rdevp;
1858 	struct raid10_info *p = conf->mirrors + number;
1859 
1860 	print_conf(conf);
1861 	if (rdev == p->rdev)
1862 		rdevp = &p->rdev;
1863 	else if (rdev == p->replacement)
1864 		rdevp = &p->replacement;
1865 	else
1866 		return 0;
1867 
1868 	if (test_bit(In_sync, &rdev->flags) ||
1869 	    atomic_read(&rdev->nr_pending)) {
1870 		err = -EBUSY;
1871 		goto abort;
1872 	}
1873 	/* Only remove faulty devices if recovery
1874 	 * is not possible.
1875 	 */
1876 	if (!test_bit(Faulty, &rdev->flags) &&
1877 	    mddev->recovery_disabled != p->recovery_disabled &&
1878 	    (!p->replacement || p->replacement == rdev) &&
1879 	    number < conf->geo.raid_disks &&
1880 	    enough(conf, -1)) {
1881 		err = -EBUSY;
1882 		goto abort;
1883 	}
1884 	*rdevp = NULL;
1885 	synchronize_rcu();
1886 	if (atomic_read(&rdev->nr_pending)) {
1887 		/* lost the race, try later */
1888 		err = -EBUSY;
1889 		*rdevp = rdev;
1890 		goto abort;
1891 	} else if (p->replacement) {
1892 		/* We must have just cleared 'rdev' */
1893 		p->rdev = p->replacement;
1894 		clear_bit(Replacement, &p->replacement->flags);
1895 		smp_mb(); /* Make sure other CPUs may see both as identical
1896 			   * but will never see neither -- if they are careful.
1897 			   */
1898 		p->replacement = NULL;
1899 		clear_bit(WantReplacement, &rdev->flags);
1900 	} else
1901 		/* We might have just remove the Replacement as faulty
1902 		 * Clear the flag just in case
1903 		 */
1904 		clear_bit(WantReplacement, &rdev->flags);
1905 
1906 	err = md_integrity_register(mddev);
1907 
1908 abort:
1909 
1910 	print_conf(conf);
1911 	return err;
1912 }
1913 
1914 
1915 static void end_sync_read(struct bio *bio, int error)
1916 {
1917 	struct r10bio *r10_bio = bio->bi_private;
1918 	struct r10conf *conf = r10_bio->mddev->private;
1919 	int d;
1920 
1921 	if (bio == r10_bio->master_bio) {
1922 		/* this is a reshape read */
1923 		d = r10_bio->read_slot; /* really the read dev */
1924 	} else
1925 		d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1926 
1927 	if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1928 		set_bit(R10BIO_Uptodate, &r10_bio->state);
1929 	else
1930 		/* The write handler will notice the lack of
1931 		 * R10BIO_Uptodate and record any errors etc
1932 		 */
1933 		atomic_add(r10_bio->sectors,
1934 			   &conf->mirrors[d].rdev->corrected_errors);
1935 
1936 	/* for reconstruct, we always reschedule after a read.
1937 	 * for resync, only after all reads
1938 	 */
1939 	rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1940 	if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1941 	    atomic_dec_and_test(&r10_bio->remaining)) {
1942 		/* we have read all the blocks,
1943 		 * do the comparison in process context in raid10d
1944 		 */
1945 		reschedule_retry(r10_bio);
1946 	}
1947 }
1948 
1949 static void end_sync_request(struct r10bio *r10_bio)
1950 {
1951 	struct mddev *mddev = r10_bio->mddev;
1952 
1953 	while (atomic_dec_and_test(&r10_bio->remaining)) {
1954 		if (r10_bio->master_bio == NULL) {
1955 			/* the primary of several recovery bios */
1956 			sector_t s = r10_bio->sectors;
1957 			if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1958 			    test_bit(R10BIO_WriteError, &r10_bio->state))
1959 				reschedule_retry(r10_bio);
1960 			else
1961 				put_buf(r10_bio);
1962 			md_done_sync(mddev, s, 1);
1963 			break;
1964 		} else {
1965 			struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1966 			if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1967 			    test_bit(R10BIO_WriteError, &r10_bio->state))
1968 				reschedule_retry(r10_bio);
1969 			else
1970 				put_buf(r10_bio);
1971 			r10_bio = r10_bio2;
1972 		}
1973 	}
1974 }
1975 
1976 static void end_sync_write(struct bio *bio, int error)
1977 {
1978 	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1979 	struct r10bio *r10_bio = bio->bi_private;
1980 	struct mddev *mddev = r10_bio->mddev;
1981 	struct r10conf *conf = mddev->private;
1982 	int d;
1983 	sector_t first_bad;
1984 	int bad_sectors;
1985 	int slot;
1986 	int repl;
1987 	struct md_rdev *rdev = NULL;
1988 
1989 	d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1990 	if (repl)
1991 		rdev = conf->mirrors[d].replacement;
1992 	else
1993 		rdev = conf->mirrors[d].rdev;
1994 
1995 	if (!uptodate) {
1996 		if (repl)
1997 			md_error(mddev, rdev);
1998 		else {
1999 			set_bit(WriteErrorSeen, &rdev->flags);
2000 			if (!test_and_set_bit(WantReplacement, &rdev->flags))
2001 				set_bit(MD_RECOVERY_NEEDED,
2002 					&rdev->mddev->recovery);
2003 			set_bit(R10BIO_WriteError, &r10_bio->state);
2004 		}
2005 	} else if (is_badblock(rdev,
2006 			     r10_bio->devs[slot].addr,
2007 			     r10_bio->sectors,
2008 			     &first_bad, &bad_sectors))
2009 		set_bit(R10BIO_MadeGood, &r10_bio->state);
2010 
2011 	rdev_dec_pending(rdev, mddev);
2012 
2013 	end_sync_request(r10_bio);
2014 }
2015 
2016 /*
2017  * Note: sync and recover and handled very differently for raid10
2018  * This code is for resync.
2019  * For resync, we read through virtual addresses and read all blocks.
2020  * If there is any error, we schedule a write.  The lowest numbered
2021  * drive is authoritative.
2022  * However requests come for physical address, so we need to map.
2023  * For every physical address there are raid_disks/copies virtual addresses,
2024  * which is always are least one, but is not necessarly an integer.
2025  * This means that a physical address can span multiple chunks, so we may
2026  * have to submit multiple io requests for a single sync request.
2027  */
2028 /*
2029  * We check if all blocks are in-sync and only write to blocks that
2030  * aren't in sync
2031  */
2032 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2033 {
2034 	struct r10conf *conf = mddev->private;
2035 	int i, first;
2036 	struct bio *tbio, *fbio;
2037 	int vcnt;
2038 
2039 	atomic_set(&r10_bio->remaining, 1);
2040 
2041 	/* find the first device with a block */
2042 	for (i=0; i<conf->copies; i++)
2043 		if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2044 			break;
2045 
2046 	if (i == conf->copies)
2047 		goto done;
2048 
2049 	first = i;
2050 	fbio = r10_bio->devs[i].bio;
2051 
2052 	vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2053 	/* now find blocks with errors */
2054 	for (i=0 ; i < conf->copies ; i++) {
2055 		int  j, d;
2056 
2057 		tbio = r10_bio->devs[i].bio;
2058 
2059 		if (tbio->bi_end_io != end_sync_read)
2060 			continue;
2061 		if (i == first)
2062 			continue;
2063 		if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2064 			/* We know that the bi_io_vec layout is the same for
2065 			 * both 'first' and 'i', so we just compare them.
2066 			 * All vec entries are PAGE_SIZE;
2067 			 */
2068 			for (j = 0; j < vcnt; j++)
2069 				if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2070 					   page_address(tbio->bi_io_vec[j].bv_page),
2071 					   fbio->bi_io_vec[j].bv_len))
2072 					break;
2073 			if (j == vcnt)
2074 				continue;
2075 			atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2076 			if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2077 				/* Don't fix anything. */
2078 				continue;
2079 		}
2080 		/* Ok, we need to write this bio, either to correct an
2081 		 * inconsistency or to correct an unreadable block.
2082 		 * First we need to fixup bv_offset, bv_len and
2083 		 * bi_vecs, as the read request might have corrupted these
2084 		 */
2085 		bio_reset(tbio);
2086 
2087 		tbio->bi_vcnt = vcnt;
2088 		tbio->bi_size = r10_bio->sectors << 9;
2089 		tbio->bi_rw = WRITE;
2090 		tbio->bi_private = r10_bio;
2091 		tbio->bi_sector = r10_bio->devs[i].addr;
2092 
2093 		for (j=0; j < vcnt ; j++) {
2094 			tbio->bi_io_vec[j].bv_offset = 0;
2095 			tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2096 
2097 			memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2098 			       page_address(fbio->bi_io_vec[j].bv_page),
2099 			       PAGE_SIZE);
2100 		}
2101 		tbio->bi_end_io = end_sync_write;
2102 
2103 		d = r10_bio->devs[i].devnum;
2104 		atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2105 		atomic_inc(&r10_bio->remaining);
2106 		md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2107 
2108 		tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
2109 		tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2110 		generic_make_request(tbio);
2111 	}
2112 
2113 	/* Now write out to any replacement devices
2114 	 * that are active
2115 	 */
2116 	for (i = 0; i < conf->copies; i++) {
2117 		int j, d;
2118 
2119 		tbio = r10_bio->devs[i].repl_bio;
2120 		if (!tbio || !tbio->bi_end_io)
2121 			continue;
2122 		if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2123 		    && r10_bio->devs[i].bio != fbio)
2124 			for (j = 0; j < vcnt; j++)
2125 				memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2126 				       page_address(fbio->bi_io_vec[j].bv_page),
2127 				       PAGE_SIZE);
2128 		d = r10_bio->devs[i].devnum;
2129 		atomic_inc(&r10_bio->remaining);
2130 		md_sync_acct(conf->mirrors[d].replacement->bdev,
2131 			     bio_sectors(tbio));
2132 		generic_make_request(tbio);
2133 	}
2134 
2135 done:
2136 	if (atomic_dec_and_test(&r10_bio->remaining)) {
2137 		md_done_sync(mddev, r10_bio->sectors, 1);
2138 		put_buf(r10_bio);
2139 	}
2140 }
2141 
2142 /*
2143  * Now for the recovery code.
2144  * Recovery happens across physical sectors.
2145  * We recover all non-is_sync drives by finding the virtual address of
2146  * each, and then choose a working drive that also has that virt address.
2147  * There is a separate r10_bio for each non-in_sync drive.
2148  * Only the first two slots are in use. The first for reading,
2149  * The second for writing.
2150  *
2151  */
2152 static void fix_recovery_read_error(struct r10bio *r10_bio)
2153 {
2154 	/* We got a read error during recovery.
2155 	 * We repeat the read in smaller page-sized sections.
2156 	 * If a read succeeds, write it to the new device or record
2157 	 * a bad block if we cannot.
2158 	 * If a read fails, record a bad block on both old and
2159 	 * new devices.
2160 	 */
2161 	struct mddev *mddev = r10_bio->mddev;
2162 	struct r10conf *conf = mddev->private;
2163 	struct bio *bio = r10_bio->devs[0].bio;
2164 	sector_t sect = 0;
2165 	int sectors = r10_bio->sectors;
2166 	int idx = 0;
2167 	int dr = r10_bio->devs[0].devnum;
2168 	int dw = r10_bio->devs[1].devnum;
2169 
2170 	while (sectors) {
2171 		int s = sectors;
2172 		struct md_rdev *rdev;
2173 		sector_t addr;
2174 		int ok;
2175 
2176 		if (s > (PAGE_SIZE>>9))
2177 			s = PAGE_SIZE >> 9;
2178 
2179 		rdev = conf->mirrors[dr].rdev;
2180 		addr = r10_bio->devs[0].addr + sect,
2181 		ok = sync_page_io(rdev,
2182 				  addr,
2183 				  s << 9,
2184 				  bio->bi_io_vec[idx].bv_page,
2185 				  READ, false);
2186 		if (ok) {
2187 			rdev = conf->mirrors[dw].rdev;
2188 			addr = r10_bio->devs[1].addr + sect;
2189 			ok = sync_page_io(rdev,
2190 					  addr,
2191 					  s << 9,
2192 					  bio->bi_io_vec[idx].bv_page,
2193 					  WRITE, false);
2194 			if (!ok) {
2195 				set_bit(WriteErrorSeen, &rdev->flags);
2196 				if (!test_and_set_bit(WantReplacement,
2197 						      &rdev->flags))
2198 					set_bit(MD_RECOVERY_NEEDED,
2199 						&rdev->mddev->recovery);
2200 			}
2201 		}
2202 		if (!ok) {
2203 			/* We don't worry if we cannot set a bad block -
2204 			 * it really is bad so there is no loss in not
2205 			 * recording it yet
2206 			 */
2207 			rdev_set_badblocks(rdev, addr, s, 0);
2208 
2209 			if (rdev != conf->mirrors[dw].rdev) {
2210 				/* need bad block on destination too */
2211 				struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2212 				addr = r10_bio->devs[1].addr + sect;
2213 				ok = rdev_set_badblocks(rdev2, addr, s, 0);
2214 				if (!ok) {
2215 					/* just abort the recovery */
2216 					printk(KERN_NOTICE
2217 					       "md/raid10:%s: recovery aborted"
2218 					       " due to read error\n",
2219 					       mdname(mddev));
2220 
2221 					conf->mirrors[dw].recovery_disabled
2222 						= mddev->recovery_disabled;
2223 					set_bit(MD_RECOVERY_INTR,
2224 						&mddev->recovery);
2225 					break;
2226 				}
2227 			}
2228 		}
2229 
2230 		sectors -= s;
2231 		sect += s;
2232 		idx++;
2233 	}
2234 }
2235 
2236 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2237 {
2238 	struct r10conf *conf = mddev->private;
2239 	int d;
2240 	struct bio *wbio, *wbio2;
2241 
2242 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2243 		fix_recovery_read_error(r10_bio);
2244 		end_sync_request(r10_bio);
2245 		return;
2246 	}
2247 
2248 	/*
2249 	 * share the pages with the first bio
2250 	 * and submit the write request
2251 	 */
2252 	d = r10_bio->devs[1].devnum;
2253 	wbio = r10_bio->devs[1].bio;
2254 	wbio2 = r10_bio->devs[1].repl_bio;
2255 	if (wbio->bi_end_io) {
2256 		atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2257 		md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2258 		generic_make_request(wbio);
2259 	}
2260 	if (wbio2 && wbio2->bi_end_io) {
2261 		atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2262 		md_sync_acct(conf->mirrors[d].replacement->bdev,
2263 			     bio_sectors(wbio2));
2264 		generic_make_request(wbio2);
2265 	}
2266 }
2267 
2268 
2269 /*
2270  * Used by fix_read_error() to decay the per rdev read_errors.
2271  * We halve the read error count for every hour that has elapsed
2272  * since the last recorded read error.
2273  *
2274  */
2275 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2276 {
2277 	struct timespec cur_time_mon;
2278 	unsigned long hours_since_last;
2279 	unsigned int read_errors = atomic_read(&rdev->read_errors);
2280 
2281 	ktime_get_ts(&cur_time_mon);
2282 
2283 	if (rdev->last_read_error.tv_sec == 0 &&
2284 	    rdev->last_read_error.tv_nsec == 0) {
2285 		/* first time we've seen a read error */
2286 		rdev->last_read_error = cur_time_mon;
2287 		return;
2288 	}
2289 
2290 	hours_since_last = (cur_time_mon.tv_sec -
2291 			    rdev->last_read_error.tv_sec) / 3600;
2292 
2293 	rdev->last_read_error = cur_time_mon;
2294 
2295 	/*
2296 	 * if hours_since_last is > the number of bits in read_errors
2297 	 * just set read errors to 0. We do this to avoid
2298 	 * overflowing the shift of read_errors by hours_since_last.
2299 	 */
2300 	if (hours_since_last >= 8 * sizeof(read_errors))
2301 		atomic_set(&rdev->read_errors, 0);
2302 	else
2303 		atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2304 }
2305 
2306 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2307 			    int sectors, struct page *page, int rw)
2308 {
2309 	sector_t first_bad;
2310 	int bad_sectors;
2311 
2312 	if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2313 	    && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2314 		return -1;
2315 	if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2316 		/* success */
2317 		return 1;
2318 	if (rw == WRITE) {
2319 		set_bit(WriteErrorSeen, &rdev->flags);
2320 		if (!test_and_set_bit(WantReplacement, &rdev->flags))
2321 			set_bit(MD_RECOVERY_NEEDED,
2322 				&rdev->mddev->recovery);
2323 	}
2324 	/* need to record an error - either for the block or the device */
2325 	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2326 		md_error(rdev->mddev, rdev);
2327 	return 0;
2328 }
2329 
2330 /*
2331  * This is a kernel thread which:
2332  *
2333  *	1.	Retries failed read operations on working mirrors.
2334  *	2.	Updates the raid superblock when problems encounter.
2335  *	3.	Performs writes following reads for array synchronising.
2336  */
2337 
2338 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2339 {
2340 	int sect = 0; /* Offset from r10_bio->sector */
2341 	int sectors = r10_bio->sectors;
2342 	struct md_rdev*rdev;
2343 	int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2344 	int d = r10_bio->devs[r10_bio->read_slot].devnum;
2345 
2346 	/* still own a reference to this rdev, so it cannot
2347 	 * have been cleared recently.
2348 	 */
2349 	rdev = conf->mirrors[d].rdev;
2350 
2351 	if (test_bit(Faulty, &rdev->flags))
2352 		/* drive has already been failed, just ignore any
2353 		   more fix_read_error() attempts */
2354 		return;
2355 
2356 	check_decay_read_errors(mddev, rdev);
2357 	atomic_inc(&rdev->read_errors);
2358 	if (atomic_read(&rdev->read_errors) > max_read_errors) {
2359 		char b[BDEVNAME_SIZE];
2360 		bdevname(rdev->bdev, b);
2361 
2362 		printk(KERN_NOTICE
2363 		       "md/raid10:%s: %s: Raid device exceeded "
2364 		       "read_error threshold [cur %d:max %d]\n",
2365 		       mdname(mddev), b,
2366 		       atomic_read(&rdev->read_errors), max_read_errors);
2367 		printk(KERN_NOTICE
2368 		       "md/raid10:%s: %s: Failing raid device\n",
2369 		       mdname(mddev), b);
2370 		md_error(mddev, conf->mirrors[d].rdev);
2371 		r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2372 		return;
2373 	}
2374 
2375 	while(sectors) {
2376 		int s = sectors;
2377 		int sl = r10_bio->read_slot;
2378 		int success = 0;
2379 		int start;
2380 
2381 		if (s > (PAGE_SIZE>>9))
2382 			s = PAGE_SIZE >> 9;
2383 
2384 		rcu_read_lock();
2385 		do {
2386 			sector_t first_bad;
2387 			int bad_sectors;
2388 
2389 			d = r10_bio->devs[sl].devnum;
2390 			rdev = rcu_dereference(conf->mirrors[d].rdev);
2391 			if (rdev &&
2392 			    !test_bit(Unmerged, &rdev->flags) &&
2393 			    test_bit(In_sync, &rdev->flags) &&
2394 			    is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2395 					&first_bad, &bad_sectors) == 0) {
2396 				atomic_inc(&rdev->nr_pending);
2397 				rcu_read_unlock();
2398 				success = sync_page_io(rdev,
2399 						       r10_bio->devs[sl].addr +
2400 						       sect,
2401 						       s<<9,
2402 						       conf->tmppage, READ, false);
2403 				rdev_dec_pending(rdev, mddev);
2404 				rcu_read_lock();
2405 				if (success)
2406 					break;
2407 			}
2408 			sl++;
2409 			if (sl == conf->copies)
2410 				sl = 0;
2411 		} while (!success && sl != r10_bio->read_slot);
2412 		rcu_read_unlock();
2413 
2414 		if (!success) {
2415 			/* Cannot read from anywhere, just mark the block
2416 			 * as bad on the first device to discourage future
2417 			 * reads.
2418 			 */
2419 			int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2420 			rdev = conf->mirrors[dn].rdev;
2421 
2422 			if (!rdev_set_badblocks(
2423 				    rdev,
2424 				    r10_bio->devs[r10_bio->read_slot].addr
2425 				    + sect,
2426 				    s, 0)) {
2427 				md_error(mddev, rdev);
2428 				r10_bio->devs[r10_bio->read_slot].bio
2429 					= IO_BLOCKED;
2430 			}
2431 			break;
2432 		}
2433 
2434 		start = sl;
2435 		/* write it back and re-read */
2436 		rcu_read_lock();
2437 		while (sl != r10_bio->read_slot) {
2438 			char b[BDEVNAME_SIZE];
2439 
2440 			if (sl==0)
2441 				sl = conf->copies;
2442 			sl--;
2443 			d = r10_bio->devs[sl].devnum;
2444 			rdev = rcu_dereference(conf->mirrors[d].rdev);
2445 			if (!rdev ||
2446 			    test_bit(Unmerged, &rdev->flags) ||
2447 			    !test_bit(In_sync, &rdev->flags))
2448 				continue;
2449 
2450 			atomic_inc(&rdev->nr_pending);
2451 			rcu_read_unlock();
2452 			if (r10_sync_page_io(rdev,
2453 					     r10_bio->devs[sl].addr +
2454 					     sect,
2455 					     s, conf->tmppage, WRITE)
2456 			    == 0) {
2457 				/* Well, this device is dead */
2458 				printk(KERN_NOTICE
2459 				       "md/raid10:%s: read correction "
2460 				       "write failed"
2461 				       " (%d sectors at %llu on %s)\n",
2462 				       mdname(mddev), s,
2463 				       (unsigned long long)(
2464 					       sect +
2465 					       choose_data_offset(r10_bio,
2466 								  rdev)),
2467 				       bdevname(rdev->bdev, b));
2468 				printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2469 				       "drive\n",
2470 				       mdname(mddev),
2471 				       bdevname(rdev->bdev, b));
2472 			}
2473 			rdev_dec_pending(rdev, mddev);
2474 			rcu_read_lock();
2475 		}
2476 		sl = start;
2477 		while (sl != r10_bio->read_slot) {
2478 			char b[BDEVNAME_SIZE];
2479 
2480 			if (sl==0)
2481 				sl = conf->copies;
2482 			sl--;
2483 			d = r10_bio->devs[sl].devnum;
2484 			rdev = rcu_dereference(conf->mirrors[d].rdev);
2485 			if (!rdev ||
2486 			    !test_bit(In_sync, &rdev->flags))
2487 				continue;
2488 
2489 			atomic_inc(&rdev->nr_pending);
2490 			rcu_read_unlock();
2491 			switch (r10_sync_page_io(rdev,
2492 					     r10_bio->devs[sl].addr +
2493 					     sect,
2494 					     s, conf->tmppage,
2495 						 READ)) {
2496 			case 0:
2497 				/* Well, this device is dead */
2498 				printk(KERN_NOTICE
2499 				       "md/raid10:%s: unable to read back "
2500 				       "corrected sectors"
2501 				       " (%d sectors at %llu on %s)\n",
2502 				       mdname(mddev), s,
2503 				       (unsigned long long)(
2504 					       sect +
2505 					       choose_data_offset(r10_bio, rdev)),
2506 				       bdevname(rdev->bdev, b));
2507 				printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2508 				       "drive\n",
2509 				       mdname(mddev),
2510 				       bdevname(rdev->bdev, b));
2511 				break;
2512 			case 1:
2513 				printk(KERN_INFO
2514 				       "md/raid10:%s: read error corrected"
2515 				       " (%d sectors at %llu on %s)\n",
2516 				       mdname(mddev), s,
2517 				       (unsigned long long)(
2518 					       sect +
2519 					       choose_data_offset(r10_bio, rdev)),
2520 				       bdevname(rdev->bdev, b));
2521 				atomic_add(s, &rdev->corrected_errors);
2522 			}
2523 
2524 			rdev_dec_pending(rdev, mddev);
2525 			rcu_read_lock();
2526 		}
2527 		rcu_read_unlock();
2528 
2529 		sectors -= s;
2530 		sect += s;
2531 	}
2532 }
2533 
2534 static int narrow_write_error(struct r10bio *r10_bio, int i)
2535 {
2536 	struct bio *bio = r10_bio->master_bio;
2537 	struct mddev *mddev = r10_bio->mddev;
2538 	struct r10conf *conf = mddev->private;
2539 	struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2540 	/* bio has the data to be written to slot 'i' where
2541 	 * we just recently had a write error.
2542 	 * We repeatedly clone the bio and trim down to one block,
2543 	 * then try the write.  Where the write fails we record
2544 	 * a bad block.
2545 	 * It is conceivable that the bio doesn't exactly align with
2546 	 * blocks.  We must handle this.
2547 	 *
2548 	 * We currently own a reference to the rdev.
2549 	 */
2550 
2551 	int block_sectors;
2552 	sector_t sector;
2553 	int sectors;
2554 	int sect_to_write = r10_bio->sectors;
2555 	int ok = 1;
2556 
2557 	if (rdev->badblocks.shift < 0)
2558 		return 0;
2559 
2560 	block_sectors = 1 << rdev->badblocks.shift;
2561 	sector = r10_bio->sector;
2562 	sectors = ((r10_bio->sector + block_sectors)
2563 		   & ~(sector_t)(block_sectors - 1))
2564 		- sector;
2565 
2566 	while (sect_to_write) {
2567 		struct bio *wbio;
2568 		if (sectors > sect_to_write)
2569 			sectors = sect_to_write;
2570 		/* Write at 'sector' for 'sectors' */
2571 		wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2572 		md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2573 		wbio->bi_sector = (r10_bio->devs[i].addr+
2574 				   choose_data_offset(r10_bio, rdev) +
2575 				   (sector - r10_bio->sector));
2576 		wbio->bi_bdev = rdev->bdev;
2577 		if (submit_bio_wait(WRITE, wbio) == 0)
2578 			/* Failure! */
2579 			ok = rdev_set_badblocks(rdev, sector,
2580 						sectors, 0)
2581 				&& ok;
2582 
2583 		bio_put(wbio);
2584 		sect_to_write -= sectors;
2585 		sector += sectors;
2586 		sectors = block_sectors;
2587 	}
2588 	return ok;
2589 }
2590 
2591 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2592 {
2593 	int slot = r10_bio->read_slot;
2594 	struct bio *bio;
2595 	struct r10conf *conf = mddev->private;
2596 	struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2597 	char b[BDEVNAME_SIZE];
2598 	unsigned long do_sync;
2599 	int max_sectors;
2600 
2601 	/* we got a read error. Maybe the drive is bad.  Maybe just
2602 	 * the block and we can fix it.
2603 	 * We freeze all other IO, and try reading the block from
2604 	 * other devices.  When we find one, we re-write
2605 	 * and check it that fixes the read error.
2606 	 * This is all done synchronously while the array is
2607 	 * frozen.
2608 	 */
2609 	bio = r10_bio->devs[slot].bio;
2610 	bdevname(bio->bi_bdev, b);
2611 	bio_put(bio);
2612 	r10_bio->devs[slot].bio = NULL;
2613 
2614 	if (mddev->ro == 0) {
2615 		freeze_array(conf);
2616 		fix_read_error(conf, mddev, r10_bio);
2617 		unfreeze_array(conf);
2618 	} else
2619 		r10_bio->devs[slot].bio = IO_BLOCKED;
2620 
2621 	rdev_dec_pending(rdev, mddev);
2622 
2623 read_more:
2624 	rdev = read_balance(conf, r10_bio, &max_sectors);
2625 	if (rdev == NULL) {
2626 		printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2627 		       " read error for block %llu\n",
2628 		       mdname(mddev), b,
2629 		       (unsigned long long)r10_bio->sector);
2630 		raid_end_bio_io(r10_bio);
2631 		return;
2632 	}
2633 
2634 	do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2635 	slot = r10_bio->read_slot;
2636 	printk_ratelimited(
2637 		KERN_ERR
2638 		"md/raid10:%s: %s: redirecting "
2639 		"sector %llu to another mirror\n",
2640 		mdname(mddev),
2641 		bdevname(rdev->bdev, b),
2642 		(unsigned long long)r10_bio->sector);
2643 	bio = bio_clone_mddev(r10_bio->master_bio,
2644 			      GFP_NOIO, mddev);
2645 	md_trim_bio(bio,
2646 		    r10_bio->sector - bio->bi_sector,
2647 		    max_sectors);
2648 	r10_bio->devs[slot].bio = bio;
2649 	r10_bio->devs[slot].rdev = rdev;
2650 	bio->bi_sector = r10_bio->devs[slot].addr
2651 		+ choose_data_offset(r10_bio, rdev);
2652 	bio->bi_bdev = rdev->bdev;
2653 	bio->bi_rw = READ | do_sync;
2654 	bio->bi_private = r10_bio;
2655 	bio->bi_end_io = raid10_end_read_request;
2656 	if (max_sectors < r10_bio->sectors) {
2657 		/* Drat - have to split this up more */
2658 		struct bio *mbio = r10_bio->master_bio;
2659 		int sectors_handled =
2660 			r10_bio->sector + max_sectors
2661 			- mbio->bi_sector;
2662 		r10_bio->sectors = max_sectors;
2663 		spin_lock_irq(&conf->device_lock);
2664 		if (mbio->bi_phys_segments == 0)
2665 			mbio->bi_phys_segments = 2;
2666 		else
2667 			mbio->bi_phys_segments++;
2668 		spin_unlock_irq(&conf->device_lock);
2669 		generic_make_request(bio);
2670 
2671 		r10_bio = mempool_alloc(conf->r10bio_pool,
2672 					GFP_NOIO);
2673 		r10_bio->master_bio = mbio;
2674 		r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2675 		r10_bio->state = 0;
2676 		set_bit(R10BIO_ReadError,
2677 			&r10_bio->state);
2678 		r10_bio->mddev = mddev;
2679 		r10_bio->sector = mbio->bi_sector
2680 			+ sectors_handled;
2681 
2682 		goto read_more;
2683 	} else
2684 		generic_make_request(bio);
2685 }
2686 
2687 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2688 {
2689 	/* Some sort of write request has finished and it
2690 	 * succeeded in writing where we thought there was a
2691 	 * bad block.  So forget the bad block.
2692 	 * Or possibly if failed and we need to record
2693 	 * a bad block.
2694 	 */
2695 	int m;
2696 	struct md_rdev *rdev;
2697 
2698 	if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2699 	    test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2700 		for (m = 0; m < conf->copies; m++) {
2701 			int dev = r10_bio->devs[m].devnum;
2702 			rdev = conf->mirrors[dev].rdev;
2703 			if (r10_bio->devs[m].bio == NULL)
2704 				continue;
2705 			if (test_bit(BIO_UPTODATE,
2706 				     &r10_bio->devs[m].bio->bi_flags)) {
2707 				rdev_clear_badblocks(
2708 					rdev,
2709 					r10_bio->devs[m].addr,
2710 					r10_bio->sectors, 0);
2711 			} else {
2712 				if (!rdev_set_badblocks(
2713 					    rdev,
2714 					    r10_bio->devs[m].addr,
2715 					    r10_bio->sectors, 0))
2716 					md_error(conf->mddev, rdev);
2717 			}
2718 			rdev = conf->mirrors[dev].replacement;
2719 			if (r10_bio->devs[m].repl_bio == NULL)
2720 				continue;
2721 			if (test_bit(BIO_UPTODATE,
2722 				     &r10_bio->devs[m].repl_bio->bi_flags)) {
2723 				rdev_clear_badblocks(
2724 					rdev,
2725 					r10_bio->devs[m].addr,
2726 					r10_bio->sectors, 0);
2727 			} else {
2728 				if (!rdev_set_badblocks(
2729 					    rdev,
2730 					    r10_bio->devs[m].addr,
2731 					    r10_bio->sectors, 0))
2732 					md_error(conf->mddev, rdev);
2733 			}
2734 		}
2735 		put_buf(r10_bio);
2736 	} else {
2737 		for (m = 0; m < conf->copies; m++) {
2738 			int dev = r10_bio->devs[m].devnum;
2739 			struct bio *bio = r10_bio->devs[m].bio;
2740 			rdev = conf->mirrors[dev].rdev;
2741 			if (bio == IO_MADE_GOOD) {
2742 				rdev_clear_badblocks(
2743 					rdev,
2744 					r10_bio->devs[m].addr,
2745 					r10_bio->sectors, 0);
2746 				rdev_dec_pending(rdev, conf->mddev);
2747 			} else if (bio != NULL &&
2748 				   !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2749 				if (!narrow_write_error(r10_bio, m)) {
2750 					md_error(conf->mddev, rdev);
2751 					set_bit(R10BIO_Degraded,
2752 						&r10_bio->state);
2753 				}
2754 				rdev_dec_pending(rdev, conf->mddev);
2755 			}
2756 			bio = r10_bio->devs[m].repl_bio;
2757 			rdev = conf->mirrors[dev].replacement;
2758 			if (rdev && bio == IO_MADE_GOOD) {
2759 				rdev_clear_badblocks(
2760 					rdev,
2761 					r10_bio->devs[m].addr,
2762 					r10_bio->sectors, 0);
2763 				rdev_dec_pending(rdev, conf->mddev);
2764 			}
2765 		}
2766 		if (test_bit(R10BIO_WriteError,
2767 			     &r10_bio->state))
2768 			close_write(r10_bio);
2769 		raid_end_bio_io(r10_bio);
2770 	}
2771 }
2772 
2773 static void raid10d(struct md_thread *thread)
2774 {
2775 	struct mddev *mddev = thread->mddev;
2776 	struct r10bio *r10_bio;
2777 	unsigned long flags;
2778 	struct r10conf *conf = mddev->private;
2779 	struct list_head *head = &conf->retry_list;
2780 	struct blk_plug plug;
2781 
2782 	md_check_recovery(mddev);
2783 
2784 	blk_start_plug(&plug);
2785 	for (;;) {
2786 
2787 		flush_pending_writes(conf);
2788 
2789 		spin_lock_irqsave(&conf->device_lock, flags);
2790 		if (list_empty(head)) {
2791 			spin_unlock_irqrestore(&conf->device_lock, flags);
2792 			break;
2793 		}
2794 		r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2795 		list_del(head->prev);
2796 		conf->nr_queued--;
2797 		spin_unlock_irqrestore(&conf->device_lock, flags);
2798 
2799 		mddev = r10_bio->mddev;
2800 		conf = mddev->private;
2801 		if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2802 		    test_bit(R10BIO_WriteError, &r10_bio->state))
2803 			handle_write_completed(conf, r10_bio);
2804 		else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2805 			reshape_request_write(mddev, r10_bio);
2806 		else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2807 			sync_request_write(mddev, r10_bio);
2808 		else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2809 			recovery_request_write(mddev, r10_bio);
2810 		else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2811 			handle_read_error(mddev, r10_bio);
2812 		else {
2813 			/* just a partial read to be scheduled from a
2814 			 * separate context
2815 			 */
2816 			int slot = r10_bio->read_slot;
2817 			generic_make_request(r10_bio->devs[slot].bio);
2818 		}
2819 
2820 		cond_resched();
2821 		if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2822 			md_check_recovery(mddev);
2823 	}
2824 	blk_finish_plug(&plug);
2825 }
2826 
2827 
2828 static int init_resync(struct r10conf *conf)
2829 {
2830 	int buffs;
2831 	int i;
2832 
2833 	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2834 	BUG_ON(conf->r10buf_pool);
2835 	conf->have_replacement = 0;
2836 	for (i = 0; i < conf->geo.raid_disks; i++)
2837 		if (conf->mirrors[i].replacement)
2838 			conf->have_replacement = 1;
2839 	conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2840 	if (!conf->r10buf_pool)
2841 		return -ENOMEM;
2842 	conf->next_resync = 0;
2843 	return 0;
2844 }
2845 
2846 /*
2847  * perform a "sync" on one "block"
2848  *
2849  * We need to make sure that no normal I/O request - particularly write
2850  * requests - conflict with active sync requests.
2851  *
2852  * This is achieved by tracking pending requests and a 'barrier' concept
2853  * that can be installed to exclude normal IO requests.
2854  *
2855  * Resync and recovery are handled very differently.
2856  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2857  *
2858  * For resync, we iterate over virtual addresses, read all copies,
2859  * and update if there are differences.  If only one copy is live,
2860  * skip it.
2861  * For recovery, we iterate over physical addresses, read a good
2862  * value for each non-in_sync drive, and over-write.
2863  *
2864  * So, for recovery we may have several outstanding complex requests for a
2865  * given address, one for each out-of-sync device.  We model this by allocating
2866  * a number of r10_bio structures, one for each out-of-sync device.
2867  * As we setup these structures, we collect all bio's together into a list
2868  * which we then process collectively to add pages, and then process again
2869  * to pass to generic_make_request.
2870  *
2871  * The r10_bio structures are linked using a borrowed master_bio pointer.
2872  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2873  * has its remaining count decremented to 0, the whole complex operation
2874  * is complete.
2875  *
2876  */
2877 
2878 static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2879 			     int *skipped, int go_faster)
2880 {
2881 	struct r10conf *conf = mddev->private;
2882 	struct r10bio *r10_bio;
2883 	struct bio *biolist = NULL, *bio;
2884 	sector_t max_sector, nr_sectors;
2885 	int i;
2886 	int max_sync;
2887 	sector_t sync_blocks;
2888 	sector_t sectors_skipped = 0;
2889 	int chunks_skipped = 0;
2890 	sector_t chunk_mask = conf->geo.chunk_mask;
2891 
2892 	if (!conf->r10buf_pool)
2893 		if (init_resync(conf))
2894 			return 0;
2895 
2896 	/*
2897 	 * Allow skipping a full rebuild for incremental assembly
2898 	 * of a clean array, like RAID1 does.
2899 	 */
2900 	if (mddev->bitmap == NULL &&
2901 	    mddev->recovery_cp == MaxSector &&
2902 	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2903 	    conf->fullsync == 0) {
2904 		*skipped = 1;
2905 		max_sector = mddev->dev_sectors;
2906 		if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2907 		    test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2908 			max_sector = mddev->resync_max_sectors;
2909 		return max_sector - sector_nr;
2910 	}
2911 
2912  skipped:
2913 	max_sector = mddev->dev_sectors;
2914 	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2915 	    test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2916 		max_sector = mddev->resync_max_sectors;
2917 	if (sector_nr >= max_sector) {
2918 		/* If we aborted, we need to abort the
2919 		 * sync on the 'current' bitmap chucks (there can
2920 		 * be several when recovering multiple devices).
2921 		 * as we may have started syncing it but not finished.
2922 		 * We can find the current address in
2923 		 * mddev->curr_resync, but for recovery,
2924 		 * we need to convert that to several
2925 		 * virtual addresses.
2926 		 */
2927 		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2928 			end_reshape(conf);
2929 			return 0;
2930 		}
2931 
2932 		if (mddev->curr_resync < max_sector) { /* aborted */
2933 			if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2934 				bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2935 						&sync_blocks, 1);
2936 			else for (i = 0; i < conf->geo.raid_disks; i++) {
2937 				sector_t sect =
2938 					raid10_find_virt(conf, mddev->curr_resync, i);
2939 				bitmap_end_sync(mddev->bitmap, sect,
2940 						&sync_blocks, 1);
2941 			}
2942 		} else {
2943 			/* completed sync */
2944 			if ((!mddev->bitmap || conf->fullsync)
2945 			    && conf->have_replacement
2946 			    && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2947 				/* Completed a full sync so the replacements
2948 				 * are now fully recovered.
2949 				 */
2950 				for (i = 0; i < conf->geo.raid_disks; i++)
2951 					if (conf->mirrors[i].replacement)
2952 						conf->mirrors[i].replacement
2953 							->recovery_offset
2954 							= MaxSector;
2955 			}
2956 			conf->fullsync = 0;
2957 		}
2958 		bitmap_close_sync(mddev->bitmap);
2959 		close_sync(conf);
2960 		*skipped = 1;
2961 		return sectors_skipped;
2962 	}
2963 
2964 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2965 		return reshape_request(mddev, sector_nr, skipped);
2966 
2967 	if (chunks_skipped >= conf->geo.raid_disks) {
2968 		/* if there has been nothing to do on any drive,
2969 		 * then there is nothing to do at all..
2970 		 */
2971 		*skipped = 1;
2972 		return (max_sector - sector_nr) + sectors_skipped;
2973 	}
2974 
2975 	if (max_sector > mddev->resync_max)
2976 		max_sector = mddev->resync_max; /* Don't do IO beyond here */
2977 
2978 	/* make sure whole request will fit in a chunk - if chunks
2979 	 * are meaningful
2980 	 */
2981 	if (conf->geo.near_copies < conf->geo.raid_disks &&
2982 	    max_sector > (sector_nr | chunk_mask))
2983 		max_sector = (sector_nr | chunk_mask) + 1;
2984 	/*
2985 	 * If there is non-resync activity waiting for us then
2986 	 * put in a delay to throttle resync.
2987 	 */
2988 	if (!go_faster && conf->nr_waiting)
2989 		msleep_interruptible(1000);
2990 
2991 	/* Again, very different code for resync and recovery.
2992 	 * Both must result in an r10bio with a list of bios that
2993 	 * have bi_end_io, bi_sector, bi_bdev set,
2994 	 * and bi_private set to the r10bio.
2995 	 * For recovery, we may actually create several r10bios
2996 	 * with 2 bios in each, that correspond to the bios in the main one.
2997 	 * In this case, the subordinate r10bios link back through a
2998 	 * borrowed master_bio pointer, and the counter in the master
2999 	 * includes a ref from each subordinate.
3000 	 */
3001 	/* First, we decide what to do and set ->bi_end_io
3002 	 * To end_sync_read if we want to read, and
3003 	 * end_sync_write if we will want to write.
3004 	 */
3005 
3006 	max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3007 	if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3008 		/* recovery... the complicated one */
3009 		int j;
3010 		r10_bio = NULL;
3011 
3012 		for (i = 0 ; i < conf->geo.raid_disks; i++) {
3013 			int still_degraded;
3014 			struct r10bio *rb2;
3015 			sector_t sect;
3016 			int must_sync;
3017 			int any_working;
3018 			struct raid10_info *mirror = &conf->mirrors[i];
3019 
3020 			if ((mirror->rdev == NULL ||
3021 			     test_bit(In_sync, &mirror->rdev->flags))
3022 			    &&
3023 			    (mirror->replacement == NULL ||
3024 			     test_bit(Faulty,
3025 				      &mirror->replacement->flags)))
3026 				continue;
3027 
3028 			still_degraded = 0;
3029 			/* want to reconstruct this device */
3030 			rb2 = r10_bio;
3031 			sect = raid10_find_virt(conf, sector_nr, i);
3032 			if (sect >= mddev->resync_max_sectors) {
3033 				/* last stripe is not complete - don't
3034 				 * try to recover this sector.
3035 				 */
3036 				continue;
3037 			}
3038 			/* Unless we are doing a full sync, or a replacement
3039 			 * we only need to recover the block if it is set in
3040 			 * the bitmap
3041 			 */
3042 			must_sync = bitmap_start_sync(mddev->bitmap, sect,
3043 						      &sync_blocks, 1);
3044 			if (sync_blocks < max_sync)
3045 				max_sync = sync_blocks;
3046 			if (!must_sync &&
3047 			    mirror->replacement == NULL &&
3048 			    !conf->fullsync) {
3049 				/* yep, skip the sync_blocks here, but don't assume
3050 				 * that there will never be anything to do here
3051 				 */
3052 				chunks_skipped = -1;
3053 				continue;
3054 			}
3055 
3056 			r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3057 			raise_barrier(conf, rb2 != NULL);
3058 			atomic_set(&r10_bio->remaining, 0);
3059 
3060 			r10_bio->master_bio = (struct bio*)rb2;
3061 			if (rb2)
3062 				atomic_inc(&rb2->remaining);
3063 			r10_bio->mddev = mddev;
3064 			set_bit(R10BIO_IsRecover, &r10_bio->state);
3065 			r10_bio->sector = sect;
3066 
3067 			raid10_find_phys(conf, r10_bio);
3068 
3069 			/* Need to check if the array will still be
3070 			 * degraded
3071 			 */
3072 			for (j = 0; j < conf->geo.raid_disks; j++)
3073 				if (conf->mirrors[j].rdev == NULL ||
3074 				    test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3075 					still_degraded = 1;
3076 					break;
3077 				}
3078 
3079 			must_sync = bitmap_start_sync(mddev->bitmap, sect,
3080 						      &sync_blocks, still_degraded);
3081 
3082 			any_working = 0;
3083 			for (j=0; j<conf->copies;j++) {
3084 				int k;
3085 				int d = r10_bio->devs[j].devnum;
3086 				sector_t from_addr, to_addr;
3087 				struct md_rdev *rdev;
3088 				sector_t sector, first_bad;
3089 				int bad_sectors;
3090 				if (!conf->mirrors[d].rdev ||
3091 				    !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3092 					continue;
3093 				/* This is where we read from */
3094 				any_working = 1;
3095 				rdev = conf->mirrors[d].rdev;
3096 				sector = r10_bio->devs[j].addr;
3097 
3098 				if (is_badblock(rdev, sector, max_sync,
3099 						&first_bad, &bad_sectors)) {
3100 					if (first_bad > sector)
3101 						max_sync = first_bad - sector;
3102 					else {
3103 						bad_sectors -= (sector
3104 								- first_bad);
3105 						if (max_sync > bad_sectors)
3106 							max_sync = bad_sectors;
3107 						continue;
3108 					}
3109 				}
3110 				bio = r10_bio->devs[0].bio;
3111 				bio_reset(bio);
3112 				bio->bi_next = biolist;
3113 				biolist = bio;
3114 				bio->bi_private = r10_bio;
3115 				bio->bi_end_io = end_sync_read;
3116 				bio->bi_rw = READ;
3117 				from_addr = r10_bio->devs[j].addr;
3118 				bio->bi_sector = from_addr + rdev->data_offset;
3119 				bio->bi_bdev = rdev->bdev;
3120 				atomic_inc(&rdev->nr_pending);
3121 				/* and we write to 'i' (if not in_sync) */
3122 
3123 				for (k=0; k<conf->copies; k++)
3124 					if (r10_bio->devs[k].devnum == i)
3125 						break;
3126 				BUG_ON(k == conf->copies);
3127 				to_addr = r10_bio->devs[k].addr;
3128 				r10_bio->devs[0].devnum = d;
3129 				r10_bio->devs[0].addr = from_addr;
3130 				r10_bio->devs[1].devnum = i;
3131 				r10_bio->devs[1].addr = to_addr;
3132 
3133 				rdev = mirror->rdev;
3134 				if (!test_bit(In_sync, &rdev->flags)) {
3135 					bio = r10_bio->devs[1].bio;
3136 					bio_reset(bio);
3137 					bio->bi_next = biolist;
3138 					biolist = bio;
3139 					bio->bi_private = r10_bio;
3140 					bio->bi_end_io = end_sync_write;
3141 					bio->bi_rw = WRITE;
3142 					bio->bi_sector = to_addr
3143 						+ rdev->data_offset;
3144 					bio->bi_bdev = rdev->bdev;
3145 					atomic_inc(&r10_bio->remaining);
3146 				} else
3147 					r10_bio->devs[1].bio->bi_end_io = NULL;
3148 
3149 				/* and maybe write to replacement */
3150 				bio = r10_bio->devs[1].repl_bio;
3151 				if (bio)
3152 					bio->bi_end_io = NULL;
3153 				rdev = mirror->replacement;
3154 				/* Note: if rdev != NULL, then bio
3155 				 * cannot be NULL as r10buf_pool_alloc will
3156 				 * have allocated it.
3157 				 * So the second test here is pointless.
3158 				 * But it keeps semantic-checkers happy, and
3159 				 * this comment keeps human reviewers
3160 				 * happy.
3161 				 */
3162 				if (rdev == NULL || bio == NULL ||
3163 				    test_bit(Faulty, &rdev->flags))
3164 					break;
3165 				bio_reset(bio);
3166 				bio->bi_next = biolist;
3167 				biolist = bio;
3168 				bio->bi_private = r10_bio;
3169 				bio->bi_end_io = end_sync_write;
3170 				bio->bi_rw = WRITE;
3171 				bio->bi_sector = to_addr + rdev->data_offset;
3172 				bio->bi_bdev = rdev->bdev;
3173 				atomic_inc(&r10_bio->remaining);
3174 				break;
3175 			}
3176 			if (j == conf->copies) {
3177 				/* Cannot recover, so abort the recovery or
3178 				 * record a bad block */
3179 				put_buf(r10_bio);
3180 				if (rb2)
3181 					atomic_dec(&rb2->remaining);
3182 				r10_bio = rb2;
3183 				if (any_working) {
3184 					/* problem is that there are bad blocks
3185 					 * on other device(s)
3186 					 */
3187 					int k;
3188 					for (k = 0; k < conf->copies; k++)
3189 						if (r10_bio->devs[k].devnum == i)
3190 							break;
3191 					if (!test_bit(In_sync,
3192 						      &mirror->rdev->flags)
3193 					    && !rdev_set_badblocks(
3194 						    mirror->rdev,
3195 						    r10_bio->devs[k].addr,
3196 						    max_sync, 0))
3197 						any_working = 0;
3198 					if (mirror->replacement &&
3199 					    !rdev_set_badblocks(
3200 						    mirror->replacement,
3201 						    r10_bio->devs[k].addr,
3202 						    max_sync, 0))
3203 						any_working = 0;
3204 				}
3205 				if (!any_working)  {
3206 					if (!test_and_set_bit(MD_RECOVERY_INTR,
3207 							      &mddev->recovery))
3208 						printk(KERN_INFO "md/raid10:%s: insufficient "
3209 						       "working devices for recovery.\n",
3210 						       mdname(mddev));
3211 					mirror->recovery_disabled
3212 						= mddev->recovery_disabled;
3213 				}
3214 				break;
3215 			}
3216 		}
3217 		if (biolist == NULL) {
3218 			while (r10_bio) {
3219 				struct r10bio *rb2 = r10_bio;
3220 				r10_bio = (struct r10bio*) rb2->master_bio;
3221 				rb2->master_bio = NULL;
3222 				put_buf(rb2);
3223 			}
3224 			goto giveup;
3225 		}
3226 	} else {
3227 		/* resync. Schedule a read for every block at this virt offset */
3228 		int count = 0;
3229 
3230 		bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3231 
3232 		if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3233 				       &sync_blocks, mddev->degraded) &&
3234 		    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3235 						 &mddev->recovery)) {
3236 			/* We can skip this block */
3237 			*skipped = 1;
3238 			return sync_blocks + sectors_skipped;
3239 		}
3240 		if (sync_blocks < max_sync)
3241 			max_sync = sync_blocks;
3242 		r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3243 
3244 		r10_bio->mddev = mddev;
3245 		atomic_set(&r10_bio->remaining, 0);
3246 		raise_barrier(conf, 0);
3247 		conf->next_resync = sector_nr;
3248 
3249 		r10_bio->master_bio = NULL;
3250 		r10_bio->sector = sector_nr;
3251 		set_bit(R10BIO_IsSync, &r10_bio->state);
3252 		raid10_find_phys(conf, r10_bio);
3253 		r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3254 
3255 		for (i = 0; i < conf->copies; i++) {
3256 			int d = r10_bio->devs[i].devnum;
3257 			sector_t first_bad, sector;
3258 			int bad_sectors;
3259 
3260 			if (r10_bio->devs[i].repl_bio)
3261 				r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3262 
3263 			bio = r10_bio->devs[i].bio;
3264 			bio_reset(bio);
3265 			clear_bit(BIO_UPTODATE, &bio->bi_flags);
3266 			if (conf->mirrors[d].rdev == NULL ||
3267 			    test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3268 				continue;
3269 			sector = r10_bio->devs[i].addr;
3270 			if (is_badblock(conf->mirrors[d].rdev,
3271 					sector, max_sync,
3272 					&first_bad, &bad_sectors)) {
3273 				if (first_bad > sector)
3274 					max_sync = first_bad - sector;
3275 				else {
3276 					bad_sectors -= (sector - first_bad);
3277 					if (max_sync > bad_sectors)
3278 						max_sync = bad_sectors;
3279 					continue;
3280 				}
3281 			}
3282 			atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3283 			atomic_inc(&r10_bio->remaining);
3284 			bio->bi_next = biolist;
3285 			biolist = bio;
3286 			bio->bi_private = r10_bio;
3287 			bio->bi_end_io = end_sync_read;
3288 			bio->bi_rw = READ;
3289 			bio->bi_sector = sector +
3290 				conf->mirrors[d].rdev->data_offset;
3291 			bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3292 			count++;
3293 
3294 			if (conf->mirrors[d].replacement == NULL ||
3295 			    test_bit(Faulty,
3296 				     &conf->mirrors[d].replacement->flags))
3297 				continue;
3298 
3299 			/* Need to set up for writing to the replacement */
3300 			bio = r10_bio->devs[i].repl_bio;
3301 			bio_reset(bio);
3302 			clear_bit(BIO_UPTODATE, &bio->bi_flags);
3303 
3304 			sector = r10_bio->devs[i].addr;
3305 			atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3306 			bio->bi_next = biolist;
3307 			biolist = bio;
3308 			bio->bi_private = r10_bio;
3309 			bio->bi_end_io = end_sync_write;
3310 			bio->bi_rw = WRITE;
3311 			bio->bi_sector = sector +
3312 				conf->mirrors[d].replacement->data_offset;
3313 			bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3314 			count++;
3315 		}
3316 
3317 		if (count < 2) {
3318 			for (i=0; i<conf->copies; i++) {
3319 				int d = r10_bio->devs[i].devnum;
3320 				if (r10_bio->devs[i].bio->bi_end_io)
3321 					rdev_dec_pending(conf->mirrors[d].rdev,
3322 							 mddev);
3323 				if (r10_bio->devs[i].repl_bio &&
3324 				    r10_bio->devs[i].repl_bio->bi_end_io)
3325 					rdev_dec_pending(
3326 						conf->mirrors[d].replacement,
3327 						mddev);
3328 			}
3329 			put_buf(r10_bio);
3330 			biolist = NULL;
3331 			goto giveup;
3332 		}
3333 	}
3334 
3335 	nr_sectors = 0;
3336 	if (sector_nr + max_sync < max_sector)
3337 		max_sector = sector_nr + max_sync;
3338 	do {
3339 		struct page *page;
3340 		int len = PAGE_SIZE;
3341 		if (sector_nr + (len>>9) > max_sector)
3342 			len = (max_sector - sector_nr) << 9;
3343 		if (len == 0)
3344 			break;
3345 		for (bio= biolist ; bio ; bio=bio->bi_next) {
3346 			struct bio *bio2;
3347 			page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3348 			if (bio_add_page(bio, page, len, 0))
3349 				continue;
3350 
3351 			/* stop here */
3352 			bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3353 			for (bio2 = biolist;
3354 			     bio2 && bio2 != bio;
3355 			     bio2 = bio2->bi_next) {
3356 				/* remove last page from this bio */
3357 				bio2->bi_vcnt--;
3358 				bio2->bi_size -= len;
3359 				bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3360 			}
3361 			goto bio_full;
3362 		}
3363 		nr_sectors += len>>9;
3364 		sector_nr += len>>9;
3365 	} while (biolist->bi_vcnt < RESYNC_PAGES);
3366  bio_full:
3367 	r10_bio->sectors = nr_sectors;
3368 
3369 	while (biolist) {
3370 		bio = biolist;
3371 		biolist = biolist->bi_next;
3372 
3373 		bio->bi_next = NULL;
3374 		r10_bio = bio->bi_private;
3375 		r10_bio->sectors = nr_sectors;
3376 
3377 		if (bio->bi_end_io == end_sync_read) {
3378 			md_sync_acct(bio->bi_bdev, nr_sectors);
3379 			generic_make_request(bio);
3380 		}
3381 	}
3382 
3383 	if (sectors_skipped)
3384 		/* pretend they weren't skipped, it makes
3385 		 * no important difference in this case
3386 		 */
3387 		md_done_sync(mddev, sectors_skipped, 1);
3388 
3389 	return sectors_skipped + nr_sectors;
3390  giveup:
3391 	/* There is nowhere to write, so all non-sync
3392 	 * drives must be failed or in resync, all drives
3393 	 * have a bad block, so try the next chunk...
3394 	 */
3395 	if (sector_nr + max_sync < max_sector)
3396 		max_sector = sector_nr + max_sync;
3397 
3398 	sectors_skipped += (max_sector - sector_nr);
3399 	chunks_skipped ++;
3400 	sector_nr = max_sector;
3401 	goto skipped;
3402 }
3403 
3404 static sector_t
3405 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3406 {
3407 	sector_t size;
3408 	struct r10conf *conf = mddev->private;
3409 
3410 	if (!raid_disks)
3411 		raid_disks = min(conf->geo.raid_disks,
3412 				 conf->prev.raid_disks);
3413 	if (!sectors)
3414 		sectors = conf->dev_sectors;
3415 
3416 	size = sectors >> conf->geo.chunk_shift;
3417 	sector_div(size, conf->geo.far_copies);
3418 	size = size * raid_disks;
3419 	sector_div(size, conf->geo.near_copies);
3420 
3421 	return size << conf->geo.chunk_shift;
3422 }
3423 
3424 static void calc_sectors(struct r10conf *conf, sector_t size)
3425 {
3426 	/* Calculate the number of sectors-per-device that will
3427 	 * actually be used, and set conf->dev_sectors and
3428 	 * conf->stride
3429 	 */
3430 
3431 	size = size >> conf->geo.chunk_shift;
3432 	sector_div(size, conf->geo.far_copies);
3433 	size = size * conf->geo.raid_disks;
3434 	sector_div(size, conf->geo.near_copies);
3435 	/* 'size' is now the number of chunks in the array */
3436 	/* calculate "used chunks per device" */
3437 	size = size * conf->copies;
3438 
3439 	/* We need to round up when dividing by raid_disks to
3440 	 * get the stride size.
3441 	 */
3442 	size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3443 
3444 	conf->dev_sectors = size << conf->geo.chunk_shift;
3445 
3446 	if (conf->geo.far_offset)
3447 		conf->geo.stride = 1 << conf->geo.chunk_shift;
3448 	else {
3449 		sector_div(size, conf->geo.far_copies);
3450 		conf->geo.stride = size << conf->geo.chunk_shift;
3451 	}
3452 }
3453 
3454 enum geo_type {geo_new, geo_old, geo_start};
3455 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3456 {
3457 	int nc, fc, fo;
3458 	int layout, chunk, disks;
3459 	switch (new) {
3460 	case geo_old:
3461 		layout = mddev->layout;
3462 		chunk = mddev->chunk_sectors;
3463 		disks = mddev->raid_disks - mddev->delta_disks;
3464 		break;
3465 	case geo_new:
3466 		layout = mddev->new_layout;
3467 		chunk = mddev->new_chunk_sectors;
3468 		disks = mddev->raid_disks;
3469 		break;
3470 	default: /* avoid 'may be unused' warnings */
3471 	case geo_start: /* new when starting reshape - raid_disks not
3472 			 * updated yet. */
3473 		layout = mddev->new_layout;
3474 		chunk = mddev->new_chunk_sectors;
3475 		disks = mddev->raid_disks + mddev->delta_disks;
3476 		break;
3477 	}
3478 	if (layout >> 18)
3479 		return -1;
3480 	if (chunk < (PAGE_SIZE >> 9) ||
3481 	    !is_power_of_2(chunk))
3482 		return -2;
3483 	nc = layout & 255;
3484 	fc = (layout >> 8) & 255;
3485 	fo = layout & (1<<16);
3486 	geo->raid_disks = disks;
3487 	geo->near_copies = nc;
3488 	geo->far_copies = fc;
3489 	geo->far_offset = fo;
3490 	geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3491 	geo->chunk_mask = chunk - 1;
3492 	geo->chunk_shift = ffz(~chunk);
3493 	return nc*fc;
3494 }
3495 
3496 static struct r10conf *setup_conf(struct mddev *mddev)
3497 {
3498 	struct r10conf *conf = NULL;
3499 	int err = -EINVAL;
3500 	struct geom geo;
3501 	int copies;
3502 
3503 	copies = setup_geo(&geo, mddev, geo_new);
3504 
3505 	if (copies == -2) {
3506 		printk(KERN_ERR "md/raid10:%s: chunk size must be "
3507 		       "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3508 		       mdname(mddev), PAGE_SIZE);
3509 		goto out;
3510 	}
3511 
3512 	if (copies < 2 || copies > mddev->raid_disks) {
3513 		printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3514 		       mdname(mddev), mddev->new_layout);
3515 		goto out;
3516 	}
3517 
3518 	err = -ENOMEM;
3519 	conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3520 	if (!conf)
3521 		goto out;
3522 
3523 	/* FIXME calc properly */
3524 	conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3525 							    max(0,mddev->delta_disks)),
3526 				GFP_KERNEL);
3527 	if (!conf->mirrors)
3528 		goto out;
3529 
3530 	conf->tmppage = alloc_page(GFP_KERNEL);
3531 	if (!conf->tmppage)
3532 		goto out;
3533 
3534 	conf->geo = geo;
3535 	conf->copies = copies;
3536 	conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3537 					   r10bio_pool_free, conf);
3538 	if (!conf->r10bio_pool)
3539 		goto out;
3540 
3541 	calc_sectors(conf, mddev->dev_sectors);
3542 	if (mddev->reshape_position == MaxSector) {
3543 		conf->prev = conf->geo;
3544 		conf->reshape_progress = MaxSector;
3545 	} else {
3546 		if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3547 			err = -EINVAL;
3548 			goto out;
3549 		}
3550 		conf->reshape_progress = mddev->reshape_position;
3551 		if (conf->prev.far_offset)
3552 			conf->prev.stride = 1 << conf->prev.chunk_shift;
3553 		else
3554 			/* far_copies must be 1 */
3555 			conf->prev.stride = conf->dev_sectors;
3556 	}
3557 	spin_lock_init(&conf->device_lock);
3558 	INIT_LIST_HEAD(&conf->retry_list);
3559 
3560 	spin_lock_init(&conf->resync_lock);
3561 	init_waitqueue_head(&conf->wait_barrier);
3562 
3563 	conf->thread = md_register_thread(raid10d, mddev, "raid10");
3564 	if (!conf->thread)
3565 		goto out;
3566 
3567 	conf->mddev = mddev;
3568 	return conf;
3569 
3570  out:
3571 	if (err == -ENOMEM)
3572 		printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3573 		       mdname(mddev));
3574 	if (conf) {
3575 		if (conf->r10bio_pool)
3576 			mempool_destroy(conf->r10bio_pool);
3577 		kfree(conf->mirrors);
3578 		safe_put_page(conf->tmppage);
3579 		kfree(conf);
3580 	}
3581 	return ERR_PTR(err);
3582 }
3583 
3584 static int run(struct mddev *mddev)
3585 {
3586 	struct r10conf *conf;
3587 	int i, disk_idx, chunk_size;
3588 	struct raid10_info *disk;
3589 	struct md_rdev *rdev;
3590 	sector_t size;
3591 	sector_t min_offset_diff = 0;
3592 	int first = 1;
3593 	bool discard_supported = false;
3594 
3595 	if (mddev->private == NULL) {
3596 		conf = setup_conf(mddev);
3597 		if (IS_ERR(conf))
3598 			return PTR_ERR(conf);
3599 		mddev->private = conf;
3600 	}
3601 	conf = mddev->private;
3602 	if (!conf)
3603 		goto out;
3604 
3605 	mddev->thread = conf->thread;
3606 	conf->thread = NULL;
3607 
3608 	chunk_size = mddev->chunk_sectors << 9;
3609 	if (mddev->queue) {
3610 		blk_queue_max_discard_sectors(mddev->queue,
3611 					      mddev->chunk_sectors);
3612 		blk_queue_max_write_same_sectors(mddev->queue,
3613 						 mddev->chunk_sectors);
3614 		blk_queue_io_min(mddev->queue, chunk_size);
3615 		if (conf->geo.raid_disks % conf->geo.near_copies)
3616 			blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3617 		else
3618 			blk_queue_io_opt(mddev->queue, chunk_size *
3619 					 (conf->geo.raid_disks / conf->geo.near_copies));
3620 	}
3621 
3622 	rdev_for_each(rdev, mddev) {
3623 		long long diff;
3624 		struct request_queue *q;
3625 
3626 		disk_idx = rdev->raid_disk;
3627 		if (disk_idx < 0)
3628 			continue;
3629 		if (disk_idx >= conf->geo.raid_disks &&
3630 		    disk_idx >= conf->prev.raid_disks)
3631 			continue;
3632 		disk = conf->mirrors + disk_idx;
3633 
3634 		if (test_bit(Replacement, &rdev->flags)) {
3635 			if (disk->replacement)
3636 				goto out_free_conf;
3637 			disk->replacement = rdev;
3638 		} else {
3639 			if (disk->rdev)
3640 				goto out_free_conf;
3641 			disk->rdev = rdev;
3642 		}
3643 		q = bdev_get_queue(rdev->bdev);
3644 		if (q->merge_bvec_fn)
3645 			mddev->merge_check_needed = 1;
3646 		diff = (rdev->new_data_offset - rdev->data_offset);
3647 		if (!mddev->reshape_backwards)
3648 			diff = -diff;
3649 		if (diff < 0)
3650 			diff = 0;
3651 		if (first || diff < min_offset_diff)
3652 			min_offset_diff = diff;
3653 
3654 		if (mddev->gendisk)
3655 			disk_stack_limits(mddev->gendisk, rdev->bdev,
3656 					  rdev->data_offset << 9);
3657 
3658 		disk->head_position = 0;
3659 
3660 		if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3661 			discard_supported = true;
3662 	}
3663 
3664 	if (mddev->queue) {
3665 		if (discard_supported)
3666 			queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3667 						mddev->queue);
3668 		else
3669 			queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3670 						  mddev->queue);
3671 	}
3672 	/* need to check that every block has at least one working mirror */
3673 	if (!enough(conf, -1)) {
3674 		printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3675 		       mdname(mddev));
3676 		goto out_free_conf;
3677 	}
3678 
3679 	if (conf->reshape_progress != MaxSector) {
3680 		/* must ensure that shape change is supported */
3681 		if (conf->geo.far_copies != 1 &&
3682 		    conf->geo.far_offset == 0)
3683 			goto out_free_conf;
3684 		if (conf->prev.far_copies != 1 &&
3685 		    conf->geo.far_offset == 0)
3686 			goto out_free_conf;
3687 	}
3688 
3689 	mddev->degraded = 0;
3690 	for (i = 0;
3691 	     i < conf->geo.raid_disks
3692 		     || i < conf->prev.raid_disks;
3693 	     i++) {
3694 
3695 		disk = conf->mirrors + i;
3696 
3697 		if (!disk->rdev && disk->replacement) {
3698 			/* The replacement is all we have - use it */
3699 			disk->rdev = disk->replacement;
3700 			disk->replacement = NULL;
3701 			clear_bit(Replacement, &disk->rdev->flags);
3702 		}
3703 
3704 		if (!disk->rdev ||
3705 		    !test_bit(In_sync, &disk->rdev->flags)) {
3706 			disk->head_position = 0;
3707 			mddev->degraded++;
3708 			if (disk->rdev)
3709 				conf->fullsync = 1;
3710 		}
3711 		disk->recovery_disabled = mddev->recovery_disabled - 1;
3712 	}
3713 
3714 	if (mddev->recovery_cp != MaxSector)
3715 		printk(KERN_NOTICE "md/raid10:%s: not clean"
3716 		       " -- starting background reconstruction\n",
3717 		       mdname(mddev));
3718 	printk(KERN_INFO
3719 		"md/raid10:%s: active with %d out of %d devices\n",
3720 		mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3721 		conf->geo.raid_disks);
3722 	/*
3723 	 * Ok, everything is just fine now
3724 	 */
3725 	mddev->dev_sectors = conf->dev_sectors;
3726 	size = raid10_size(mddev, 0, 0);
3727 	md_set_array_sectors(mddev, size);
3728 	mddev->resync_max_sectors = size;
3729 
3730 	if (mddev->queue) {
3731 		int stripe = conf->geo.raid_disks *
3732 			((mddev->chunk_sectors << 9) / PAGE_SIZE);
3733 		mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3734 		mddev->queue->backing_dev_info.congested_data = mddev;
3735 
3736 		/* Calculate max read-ahead size.
3737 		 * We need to readahead at least twice a whole stripe....
3738 		 * maybe...
3739 		 */
3740 		stripe /= conf->geo.near_copies;
3741 		if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3742 			mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3743 		blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3744 	}
3745 
3746 
3747 	if (md_integrity_register(mddev))
3748 		goto out_free_conf;
3749 
3750 	if (conf->reshape_progress != MaxSector) {
3751 		unsigned long before_length, after_length;
3752 
3753 		before_length = ((1 << conf->prev.chunk_shift) *
3754 				 conf->prev.far_copies);
3755 		after_length = ((1 << conf->geo.chunk_shift) *
3756 				conf->geo.far_copies);
3757 
3758 		if (max(before_length, after_length) > min_offset_diff) {
3759 			/* This cannot work */
3760 			printk("md/raid10: offset difference not enough to continue reshape\n");
3761 			goto out_free_conf;
3762 		}
3763 		conf->offset_diff = min_offset_diff;
3764 
3765 		conf->reshape_safe = conf->reshape_progress;
3766 		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3767 		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3768 		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3769 		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3770 		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3771 							"reshape");
3772 	}
3773 
3774 	return 0;
3775 
3776 out_free_conf:
3777 	md_unregister_thread(&mddev->thread);
3778 	if (conf->r10bio_pool)
3779 		mempool_destroy(conf->r10bio_pool);
3780 	safe_put_page(conf->tmppage);
3781 	kfree(conf->mirrors);
3782 	kfree(conf);
3783 	mddev->private = NULL;
3784 out:
3785 	return -EIO;
3786 }
3787 
3788 static int stop(struct mddev *mddev)
3789 {
3790 	struct r10conf *conf = mddev->private;
3791 
3792 	raise_barrier(conf, 0);
3793 	lower_barrier(conf);
3794 
3795 	md_unregister_thread(&mddev->thread);
3796 	if (mddev->queue)
3797 		/* the unplug fn references 'conf'*/
3798 		blk_sync_queue(mddev->queue);
3799 
3800 	if (conf->r10bio_pool)
3801 		mempool_destroy(conf->r10bio_pool);
3802 	safe_put_page(conf->tmppage);
3803 	kfree(conf->mirrors);
3804 	kfree(conf);
3805 	mddev->private = NULL;
3806 	return 0;
3807 }
3808 
3809 static void raid10_quiesce(struct mddev *mddev, int state)
3810 {
3811 	struct r10conf *conf = mddev->private;
3812 
3813 	switch(state) {
3814 	case 1:
3815 		raise_barrier(conf, 0);
3816 		break;
3817 	case 0:
3818 		lower_barrier(conf);
3819 		break;
3820 	}
3821 }
3822 
3823 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3824 {
3825 	/* Resize of 'far' arrays is not supported.
3826 	 * For 'near' and 'offset' arrays we can set the
3827 	 * number of sectors used to be an appropriate multiple
3828 	 * of the chunk size.
3829 	 * For 'offset', this is far_copies*chunksize.
3830 	 * For 'near' the multiplier is the LCM of
3831 	 * near_copies and raid_disks.
3832 	 * So if far_copies > 1 && !far_offset, fail.
3833 	 * Else find LCM(raid_disks, near_copy)*far_copies and
3834 	 * multiply by chunk_size.  Then round to this number.
3835 	 * This is mostly done by raid10_size()
3836 	 */
3837 	struct r10conf *conf = mddev->private;
3838 	sector_t oldsize, size;
3839 
3840 	if (mddev->reshape_position != MaxSector)
3841 		return -EBUSY;
3842 
3843 	if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3844 		return -EINVAL;
3845 
3846 	oldsize = raid10_size(mddev, 0, 0);
3847 	size = raid10_size(mddev, sectors, 0);
3848 	if (mddev->external_size &&
3849 	    mddev->array_sectors > size)
3850 		return -EINVAL;
3851 	if (mddev->bitmap) {
3852 		int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3853 		if (ret)
3854 			return ret;
3855 	}
3856 	md_set_array_sectors(mddev, size);
3857 	set_capacity(mddev->gendisk, mddev->array_sectors);
3858 	revalidate_disk(mddev->gendisk);
3859 	if (sectors > mddev->dev_sectors &&
3860 	    mddev->recovery_cp > oldsize) {
3861 		mddev->recovery_cp = oldsize;
3862 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3863 	}
3864 	calc_sectors(conf, sectors);
3865 	mddev->dev_sectors = conf->dev_sectors;
3866 	mddev->resync_max_sectors = size;
3867 	return 0;
3868 }
3869 
3870 static void *raid10_takeover_raid0(struct mddev *mddev)
3871 {
3872 	struct md_rdev *rdev;
3873 	struct r10conf *conf;
3874 
3875 	if (mddev->degraded > 0) {
3876 		printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3877 		       mdname(mddev));
3878 		return ERR_PTR(-EINVAL);
3879 	}
3880 
3881 	/* Set new parameters */
3882 	mddev->new_level = 10;
3883 	/* new layout: far_copies = 1, near_copies = 2 */
3884 	mddev->new_layout = (1<<8) + 2;
3885 	mddev->new_chunk_sectors = mddev->chunk_sectors;
3886 	mddev->delta_disks = mddev->raid_disks;
3887 	mddev->raid_disks *= 2;
3888 	/* make sure it will be not marked as dirty */
3889 	mddev->recovery_cp = MaxSector;
3890 
3891 	conf = setup_conf(mddev);
3892 	if (!IS_ERR(conf)) {
3893 		rdev_for_each(rdev, mddev)
3894 			if (rdev->raid_disk >= 0)
3895 				rdev->new_raid_disk = rdev->raid_disk * 2;
3896 		conf->barrier = 1;
3897 	}
3898 
3899 	return conf;
3900 }
3901 
3902 static void *raid10_takeover(struct mddev *mddev)
3903 {
3904 	struct r0conf *raid0_conf;
3905 
3906 	/* raid10 can take over:
3907 	 *  raid0 - providing it has only two drives
3908 	 */
3909 	if (mddev->level == 0) {
3910 		/* for raid0 takeover only one zone is supported */
3911 		raid0_conf = mddev->private;
3912 		if (raid0_conf->nr_strip_zones > 1) {
3913 			printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3914 			       " with more than one zone.\n",
3915 			       mdname(mddev));
3916 			return ERR_PTR(-EINVAL);
3917 		}
3918 		return raid10_takeover_raid0(mddev);
3919 	}
3920 	return ERR_PTR(-EINVAL);
3921 }
3922 
3923 static int raid10_check_reshape(struct mddev *mddev)
3924 {
3925 	/* Called when there is a request to change
3926 	 * - layout (to ->new_layout)
3927 	 * - chunk size (to ->new_chunk_sectors)
3928 	 * - raid_disks (by delta_disks)
3929 	 * or when trying to restart a reshape that was ongoing.
3930 	 *
3931 	 * We need to validate the request and possibly allocate
3932 	 * space if that might be an issue later.
3933 	 *
3934 	 * Currently we reject any reshape of a 'far' mode array,
3935 	 * allow chunk size to change if new is generally acceptable,
3936 	 * allow raid_disks to increase, and allow
3937 	 * a switch between 'near' mode and 'offset' mode.
3938 	 */
3939 	struct r10conf *conf = mddev->private;
3940 	struct geom geo;
3941 
3942 	if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3943 		return -EINVAL;
3944 
3945 	if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3946 		/* mustn't change number of copies */
3947 		return -EINVAL;
3948 	if (geo.far_copies > 1 && !geo.far_offset)
3949 		/* Cannot switch to 'far' mode */
3950 		return -EINVAL;
3951 
3952 	if (mddev->array_sectors & geo.chunk_mask)
3953 			/* not factor of array size */
3954 			return -EINVAL;
3955 
3956 	if (!enough(conf, -1))
3957 		return -EINVAL;
3958 
3959 	kfree(conf->mirrors_new);
3960 	conf->mirrors_new = NULL;
3961 	if (mddev->delta_disks > 0) {
3962 		/* allocate new 'mirrors' list */
3963 		conf->mirrors_new = kzalloc(
3964 			sizeof(struct raid10_info)
3965 			*(mddev->raid_disks +
3966 			  mddev->delta_disks),
3967 			GFP_KERNEL);
3968 		if (!conf->mirrors_new)
3969 			return -ENOMEM;
3970 	}
3971 	return 0;
3972 }
3973 
3974 /*
3975  * Need to check if array has failed when deciding whether to:
3976  *  - start an array
3977  *  - remove non-faulty devices
3978  *  - add a spare
3979  *  - allow a reshape
3980  * This determination is simple when no reshape is happening.
3981  * However if there is a reshape, we need to carefully check
3982  * both the before and after sections.
3983  * This is because some failed devices may only affect one
3984  * of the two sections, and some non-in_sync devices may
3985  * be insync in the section most affected by failed devices.
3986  */
3987 static int calc_degraded(struct r10conf *conf)
3988 {
3989 	int degraded, degraded2;
3990 	int i;
3991 
3992 	rcu_read_lock();
3993 	degraded = 0;
3994 	/* 'prev' section first */
3995 	for (i = 0; i < conf->prev.raid_disks; i++) {
3996 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3997 		if (!rdev || test_bit(Faulty, &rdev->flags))
3998 			degraded++;
3999 		else if (!test_bit(In_sync, &rdev->flags))
4000 			/* When we can reduce the number of devices in
4001 			 * an array, this might not contribute to
4002 			 * 'degraded'.  It does now.
4003 			 */
4004 			degraded++;
4005 	}
4006 	rcu_read_unlock();
4007 	if (conf->geo.raid_disks == conf->prev.raid_disks)
4008 		return degraded;
4009 	rcu_read_lock();
4010 	degraded2 = 0;
4011 	for (i = 0; i < conf->geo.raid_disks; i++) {
4012 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4013 		if (!rdev || test_bit(Faulty, &rdev->flags))
4014 			degraded2++;
4015 		else if (!test_bit(In_sync, &rdev->flags)) {
4016 			/* If reshape is increasing the number of devices,
4017 			 * this section has already been recovered, so
4018 			 * it doesn't contribute to degraded.
4019 			 * else it does.
4020 			 */
4021 			if (conf->geo.raid_disks <= conf->prev.raid_disks)
4022 				degraded2++;
4023 		}
4024 	}
4025 	rcu_read_unlock();
4026 	if (degraded2 > degraded)
4027 		return degraded2;
4028 	return degraded;
4029 }
4030 
4031 static int raid10_start_reshape(struct mddev *mddev)
4032 {
4033 	/* A 'reshape' has been requested. This commits
4034 	 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4035 	 * This also checks if there are enough spares and adds them
4036 	 * to the array.
4037 	 * We currently require enough spares to make the final
4038 	 * array non-degraded.  We also require that the difference
4039 	 * between old and new data_offset - on each device - is
4040 	 * enough that we never risk over-writing.
4041 	 */
4042 
4043 	unsigned long before_length, after_length;
4044 	sector_t min_offset_diff = 0;
4045 	int first = 1;
4046 	struct geom new;
4047 	struct r10conf *conf = mddev->private;
4048 	struct md_rdev *rdev;
4049 	int spares = 0;
4050 	int ret;
4051 
4052 	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4053 		return -EBUSY;
4054 
4055 	if (setup_geo(&new, mddev, geo_start) != conf->copies)
4056 		return -EINVAL;
4057 
4058 	before_length = ((1 << conf->prev.chunk_shift) *
4059 			 conf->prev.far_copies);
4060 	after_length = ((1 << conf->geo.chunk_shift) *
4061 			conf->geo.far_copies);
4062 
4063 	rdev_for_each(rdev, mddev) {
4064 		if (!test_bit(In_sync, &rdev->flags)
4065 		    && !test_bit(Faulty, &rdev->flags))
4066 			spares++;
4067 		if (rdev->raid_disk >= 0) {
4068 			long long diff = (rdev->new_data_offset
4069 					  - rdev->data_offset);
4070 			if (!mddev->reshape_backwards)
4071 				diff = -diff;
4072 			if (diff < 0)
4073 				diff = 0;
4074 			if (first || diff < min_offset_diff)
4075 				min_offset_diff = diff;
4076 		}
4077 	}
4078 
4079 	if (max(before_length, after_length) > min_offset_diff)
4080 		return -EINVAL;
4081 
4082 	if (spares < mddev->delta_disks)
4083 		return -EINVAL;
4084 
4085 	conf->offset_diff = min_offset_diff;
4086 	spin_lock_irq(&conf->device_lock);
4087 	if (conf->mirrors_new) {
4088 		memcpy(conf->mirrors_new, conf->mirrors,
4089 		       sizeof(struct raid10_info)*conf->prev.raid_disks);
4090 		smp_mb();
4091 		kfree(conf->mirrors_old); /* FIXME and elsewhere */
4092 		conf->mirrors_old = conf->mirrors;
4093 		conf->mirrors = conf->mirrors_new;
4094 		conf->mirrors_new = NULL;
4095 	}
4096 	setup_geo(&conf->geo, mddev, geo_start);
4097 	smp_mb();
4098 	if (mddev->reshape_backwards) {
4099 		sector_t size = raid10_size(mddev, 0, 0);
4100 		if (size < mddev->array_sectors) {
4101 			spin_unlock_irq(&conf->device_lock);
4102 			printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4103 			       mdname(mddev));
4104 			return -EINVAL;
4105 		}
4106 		mddev->resync_max_sectors = size;
4107 		conf->reshape_progress = size;
4108 	} else
4109 		conf->reshape_progress = 0;
4110 	spin_unlock_irq(&conf->device_lock);
4111 
4112 	if (mddev->delta_disks && mddev->bitmap) {
4113 		ret = bitmap_resize(mddev->bitmap,
4114 				    raid10_size(mddev, 0,
4115 						conf->geo.raid_disks),
4116 				    0, 0);
4117 		if (ret)
4118 			goto abort;
4119 	}
4120 	if (mddev->delta_disks > 0) {
4121 		rdev_for_each(rdev, mddev)
4122 			if (rdev->raid_disk < 0 &&
4123 			    !test_bit(Faulty, &rdev->flags)) {
4124 				if (raid10_add_disk(mddev, rdev) == 0) {
4125 					if (rdev->raid_disk >=
4126 					    conf->prev.raid_disks)
4127 						set_bit(In_sync, &rdev->flags);
4128 					else
4129 						rdev->recovery_offset = 0;
4130 
4131 					if (sysfs_link_rdev(mddev, rdev))
4132 						/* Failure here  is OK */;
4133 				}
4134 			} else if (rdev->raid_disk >= conf->prev.raid_disks
4135 				   && !test_bit(Faulty, &rdev->flags)) {
4136 				/* This is a spare that was manually added */
4137 				set_bit(In_sync, &rdev->flags);
4138 			}
4139 	}
4140 	/* When a reshape changes the number of devices,
4141 	 * ->degraded is measured against the larger of the
4142 	 * pre and  post numbers.
4143 	 */
4144 	spin_lock_irq(&conf->device_lock);
4145 	mddev->degraded = calc_degraded(conf);
4146 	spin_unlock_irq(&conf->device_lock);
4147 	mddev->raid_disks = conf->geo.raid_disks;
4148 	mddev->reshape_position = conf->reshape_progress;
4149 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
4150 
4151 	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4152 	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4153 	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4154 	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4155 
4156 	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4157 						"reshape");
4158 	if (!mddev->sync_thread) {
4159 		ret = -EAGAIN;
4160 		goto abort;
4161 	}
4162 	conf->reshape_checkpoint = jiffies;
4163 	md_wakeup_thread(mddev->sync_thread);
4164 	md_new_event(mddev);
4165 	return 0;
4166 
4167 abort:
4168 	mddev->recovery = 0;
4169 	spin_lock_irq(&conf->device_lock);
4170 	conf->geo = conf->prev;
4171 	mddev->raid_disks = conf->geo.raid_disks;
4172 	rdev_for_each(rdev, mddev)
4173 		rdev->new_data_offset = rdev->data_offset;
4174 	smp_wmb();
4175 	conf->reshape_progress = MaxSector;
4176 	mddev->reshape_position = MaxSector;
4177 	spin_unlock_irq(&conf->device_lock);
4178 	return ret;
4179 }
4180 
4181 /* Calculate the last device-address that could contain
4182  * any block from the chunk that includes the array-address 's'
4183  * and report the next address.
4184  * i.e. the address returned will be chunk-aligned and after
4185  * any data that is in the chunk containing 's'.
4186  */
4187 static sector_t last_dev_address(sector_t s, struct geom *geo)
4188 {
4189 	s = (s | geo->chunk_mask) + 1;
4190 	s >>= geo->chunk_shift;
4191 	s *= geo->near_copies;
4192 	s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4193 	s *= geo->far_copies;
4194 	s <<= geo->chunk_shift;
4195 	return s;
4196 }
4197 
4198 /* Calculate the first device-address that could contain
4199  * any block from the chunk that includes the array-address 's'.
4200  * This too will be the start of a chunk
4201  */
4202 static sector_t first_dev_address(sector_t s, struct geom *geo)
4203 {
4204 	s >>= geo->chunk_shift;
4205 	s *= geo->near_copies;
4206 	sector_div(s, geo->raid_disks);
4207 	s *= geo->far_copies;
4208 	s <<= geo->chunk_shift;
4209 	return s;
4210 }
4211 
4212 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4213 				int *skipped)
4214 {
4215 	/* We simply copy at most one chunk (smallest of old and new)
4216 	 * at a time, possibly less if that exceeds RESYNC_PAGES,
4217 	 * or we hit a bad block or something.
4218 	 * This might mean we pause for normal IO in the middle of
4219 	 * a chunk, but that is not a problem was mddev->reshape_position
4220 	 * can record any location.
4221 	 *
4222 	 * If we will want to write to a location that isn't
4223 	 * yet recorded as 'safe' (i.e. in metadata on disk) then
4224 	 * we need to flush all reshape requests and update the metadata.
4225 	 *
4226 	 * When reshaping forwards (e.g. to more devices), we interpret
4227 	 * 'safe' as the earliest block which might not have been copied
4228 	 * down yet.  We divide this by previous stripe size and multiply
4229 	 * by previous stripe length to get lowest device offset that we
4230 	 * cannot write to yet.
4231 	 * We interpret 'sector_nr' as an address that we want to write to.
4232 	 * From this we use last_device_address() to find where we might
4233 	 * write to, and first_device_address on the  'safe' position.
4234 	 * If this 'next' write position is after the 'safe' position,
4235 	 * we must update the metadata to increase the 'safe' position.
4236 	 *
4237 	 * When reshaping backwards, we round in the opposite direction
4238 	 * and perform the reverse test:  next write position must not be
4239 	 * less than current safe position.
4240 	 *
4241 	 * In all this the minimum difference in data offsets
4242 	 * (conf->offset_diff - always positive) allows a bit of slack,
4243 	 * so next can be after 'safe', but not by more than offset_disk
4244 	 *
4245 	 * We need to prepare all the bios here before we start any IO
4246 	 * to ensure the size we choose is acceptable to all devices.
4247 	 * The means one for each copy for write-out and an extra one for
4248 	 * read-in.
4249 	 * We store the read-in bio in ->master_bio and the others in
4250 	 * ->devs[x].bio and ->devs[x].repl_bio.
4251 	 */
4252 	struct r10conf *conf = mddev->private;
4253 	struct r10bio *r10_bio;
4254 	sector_t next, safe, last;
4255 	int max_sectors;
4256 	int nr_sectors;
4257 	int s;
4258 	struct md_rdev *rdev;
4259 	int need_flush = 0;
4260 	struct bio *blist;
4261 	struct bio *bio, *read_bio;
4262 	int sectors_done = 0;
4263 
4264 	if (sector_nr == 0) {
4265 		/* If restarting in the middle, skip the initial sectors */
4266 		if (mddev->reshape_backwards &&
4267 		    conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4268 			sector_nr = (raid10_size(mddev, 0, 0)
4269 				     - conf->reshape_progress);
4270 		} else if (!mddev->reshape_backwards &&
4271 			   conf->reshape_progress > 0)
4272 			sector_nr = conf->reshape_progress;
4273 		if (sector_nr) {
4274 			mddev->curr_resync_completed = sector_nr;
4275 			sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4276 			*skipped = 1;
4277 			return sector_nr;
4278 		}
4279 	}
4280 
4281 	/* We don't use sector_nr to track where we are up to
4282 	 * as that doesn't work well for ->reshape_backwards.
4283 	 * So just use ->reshape_progress.
4284 	 */
4285 	if (mddev->reshape_backwards) {
4286 		/* 'next' is the earliest device address that we might
4287 		 * write to for this chunk in the new layout
4288 		 */
4289 		next = first_dev_address(conf->reshape_progress - 1,
4290 					 &conf->geo);
4291 
4292 		/* 'safe' is the last device address that we might read from
4293 		 * in the old layout after a restart
4294 		 */
4295 		safe = last_dev_address(conf->reshape_safe - 1,
4296 					&conf->prev);
4297 
4298 		if (next + conf->offset_diff < safe)
4299 			need_flush = 1;
4300 
4301 		last = conf->reshape_progress - 1;
4302 		sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4303 					       & conf->prev.chunk_mask);
4304 		if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4305 			sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4306 	} else {
4307 		/* 'next' is after the last device address that we
4308 		 * might write to for this chunk in the new layout
4309 		 */
4310 		next = last_dev_address(conf->reshape_progress, &conf->geo);
4311 
4312 		/* 'safe' is the earliest device address that we might
4313 		 * read from in the old layout after a restart
4314 		 */
4315 		safe = first_dev_address(conf->reshape_safe, &conf->prev);
4316 
4317 		/* Need to update metadata if 'next' might be beyond 'safe'
4318 		 * as that would possibly corrupt data
4319 		 */
4320 		if (next > safe + conf->offset_diff)
4321 			need_flush = 1;
4322 
4323 		sector_nr = conf->reshape_progress;
4324 		last  = sector_nr | (conf->geo.chunk_mask
4325 				     & conf->prev.chunk_mask);
4326 
4327 		if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4328 			last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4329 	}
4330 
4331 	if (need_flush ||
4332 	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4333 		/* Need to update reshape_position in metadata */
4334 		wait_barrier(conf);
4335 		mddev->reshape_position = conf->reshape_progress;
4336 		if (mddev->reshape_backwards)
4337 			mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4338 				- conf->reshape_progress;
4339 		else
4340 			mddev->curr_resync_completed = conf->reshape_progress;
4341 		conf->reshape_checkpoint = jiffies;
4342 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
4343 		md_wakeup_thread(mddev->thread);
4344 		wait_event(mddev->sb_wait, mddev->flags == 0 ||
4345 			   kthread_should_stop());
4346 		conf->reshape_safe = mddev->reshape_position;
4347 		allow_barrier(conf);
4348 	}
4349 
4350 read_more:
4351 	/* Now schedule reads for blocks from sector_nr to last */
4352 	r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4353 	raise_barrier(conf, sectors_done != 0);
4354 	atomic_set(&r10_bio->remaining, 0);
4355 	r10_bio->mddev = mddev;
4356 	r10_bio->sector = sector_nr;
4357 	set_bit(R10BIO_IsReshape, &r10_bio->state);
4358 	r10_bio->sectors = last - sector_nr + 1;
4359 	rdev = read_balance(conf, r10_bio, &max_sectors);
4360 	BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4361 
4362 	if (!rdev) {
4363 		/* Cannot read from here, so need to record bad blocks
4364 		 * on all the target devices.
4365 		 */
4366 		// FIXME
4367 		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4368 		return sectors_done;
4369 	}
4370 
4371 	read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4372 
4373 	read_bio->bi_bdev = rdev->bdev;
4374 	read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4375 			       + rdev->data_offset);
4376 	read_bio->bi_private = r10_bio;
4377 	read_bio->bi_end_io = end_sync_read;
4378 	read_bio->bi_rw = READ;
4379 	read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4380 	read_bio->bi_flags |= 1 << BIO_UPTODATE;
4381 	read_bio->bi_vcnt = 0;
4382 	read_bio->bi_size = 0;
4383 	r10_bio->master_bio = read_bio;
4384 	r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4385 
4386 	/* Now find the locations in the new layout */
4387 	__raid10_find_phys(&conf->geo, r10_bio);
4388 
4389 	blist = read_bio;
4390 	read_bio->bi_next = NULL;
4391 
4392 	for (s = 0; s < conf->copies*2; s++) {
4393 		struct bio *b;
4394 		int d = r10_bio->devs[s/2].devnum;
4395 		struct md_rdev *rdev2;
4396 		if (s&1) {
4397 			rdev2 = conf->mirrors[d].replacement;
4398 			b = r10_bio->devs[s/2].repl_bio;
4399 		} else {
4400 			rdev2 = conf->mirrors[d].rdev;
4401 			b = r10_bio->devs[s/2].bio;
4402 		}
4403 		if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4404 			continue;
4405 
4406 		bio_reset(b);
4407 		b->bi_bdev = rdev2->bdev;
4408 		b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4409 		b->bi_private = r10_bio;
4410 		b->bi_end_io = end_reshape_write;
4411 		b->bi_rw = WRITE;
4412 		b->bi_next = blist;
4413 		blist = b;
4414 	}
4415 
4416 	/* Now add as many pages as possible to all of these bios. */
4417 
4418 	nr_sectors = 0;
4419 	for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4420 		struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4421 		int len = (max_sectors - s) << 9;
4422 		if (len > PAGE_SIZE)
4423 			len = PAGE_SIZE;
4424 		for (bio = blist; bio ; bio = bio->bi_next) {
4425 			struct bio *bio2;
4426 			if (bio_add_page(bio, page, len, 0))
4427 				continue;
4428 
4429 			/* Didn't fit, must stop */
4430 			for (bio2 = blist;
4431 			     bio2 && bio2 != bio;
4432 			     bio2 = bio2->bi_next) {
4433 				/* Remove last page from this bio */
4434 				bio2->bi_vcnt--;
4435 				bio2->bi_size -= len;
4436 				bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4437 			}
4438 			goto bio_full;
4439 		}
4440 		sector_nr += len >> 9;
4441 		nr_sectors += len >> 9;
4442 	}
4443 bio_full:
4444 	r10_bio->sectors = nr_sectors;
4445 
4446 	/* Now submit the read */
4447 	md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4448 	atomic_inc(&r10_bio->remaining);
4449 	read_bio->bi_next = NULL;
4450 	generic_make_request(read_bio);
4451 	sector_nr += nr_sectors;
4452 	sectors_done += nr_sectors;
4453 	if (sector_nr <= last)
4454 		goto read_more;
4455 
4456 	/* Now that we have done the whole section we can
4457 	 * update reshape_progress
4458 	 */
4459 	if (mddev->reshape_backwards)
4460 		conf->reshape_progress -= sectors_done;
4461 	else
4462 		conf->reshape_progress += sectors_done;
4463 
4464 	return sectors_done;
4465 }
4466 
4467 static void end_reshape_request(struct r10bio *r10_bio);
4468 static int handle_reshape_read_error(struct mddev *mddev,
4469 				     struct r10bio *r10_bio);
4470 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4471 {
4472 	/* Reshape read completed.  Hopefully we have a block
4473 	 * to write out.
4474 	 * If we got a read error then we do sync 1-page reads from
4475 	 * elsewhere until we find the data - or give up.
4476 	 */
4477 	struct r10conf *conf = mddev->private;
4478 	int s;
4479 
4480 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4481 		if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4482 			/* Reshape has been aborted */
4483 			md_done_sync(mddev, r10_bio->sectors, 0);
4484 			return;
4485 		}
4486 
4487 	/* We definitely have the data in the pages, schedule the
4488 	 * writes.
4489 	 */
4490 	atomic_set(&r10_bio->remaining, 1);
4491 	for (s = 0; s < conf->copies*2; s++) {
4492 		struct bio *b;
4493 		int d = r10_bio->devs[s/2].devnum;
4494 		struct md_rdev *rdev;
4495 		if (s&1) {
4496 			rdev = conf->mirrors[d].replacement;
4497 			b = r10_bio->devs[s/2].repl_bio;
4498 		} else {
4499 			rdev = conf->mirrors[d].rdev;
4500 			b = r10_bio->devs[s/2].bio;
4501 		}
4502 		if (!rdev || test_bit(Faulty, &rdev->flags))
4503 			continue;
4504 		atomic_inc(&rdev->nr_pending);
4505 		md_sync_acct(b->bi_bdev, r10_bio->sectors);
4506 		atomic_inc(&r10_bio->remaining);
4507 		b->bi_next = NULL;
4508 		generic_make_request(b);
4509 	}
4510 	end_reshape_request(r10_bio);
4511 }
4512 
4513 static void end_reshape(struct r10conf *conf)
4514 {
4515 	if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4516 		return;
4517 
4518 	spin_lock_irq(&conf->device_lock);
4519 	conf->prev = conf->geo;
4520 	md_finish_reshape(conf->mddev);
4521 	smp_wmb();
4522 	conf->reshape_progress = MaxSector;
4523 	spin_unlock_irq(&conf->device_lock);
4524 
4525 	/* read-ahead size must cover two whole stripes, which is
4526 	 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4527 	 */
4528 	if (conf->mddev->queue) {
4529 		int stripe = conf->geo.raid_disks *
4530 			((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4531 		stripe /= conf->geo.near_copies;
4532 		if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4533 			conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4534 	}
4535 	conf->fullsync = 0;
4536 }
4537 
4538 
4539 static int handle_reshape_read_error(struct mddev *mddev,
4540 				     struct r10bio *r10_bio)
4541 {
4542 	/* Use sync reads to get the blocks from somewhere else */
4543 	int sectors = r10_bio->sectors;
4544 	struct r10conf *conf = mddev->private;
4545 	struct {
4546 		struct r10bio r10_bio;
4547 		struct r10dev devs[conf->copies];
4548 	} on_stack;
4549 	struct r10bio *r10b = &on_stack.r10_bio;
4550 	int slot = 0;
4551 	int idx = 0;
4552 	struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4553 
4554 	r10b->sector = r10_bio->sector;
4555 	__raid10_find_phys(&conf->prev, r10b);
4556 
4557 	while (sectors) {
4558 		int s = sectors;
4559 		int success = 0;
4560 		int first_slot = slot;
4561 
4562 		if (s > (PAGE_SIZE >> 9))
4563 			s = PAGE_SIZE >> 9;
4564 
4565 		while (!success) {
4566 			int d = r10b->devs[slot].devnum;
4567 			struct md_rdev *rdev = conf->mirrors[d].rdev;
4568 			sector_t addr;
4569 			if (rdev == NULL ||
4570 			    test_bit(Faulty, &rdev->flags) ||
4571 			    !test_bit(In_sync, &rdev->flags))
4572 				goto failed;
4573 
4574 			addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4575 			success = sync_page_io(rdev,
4576 					       addr,
4577 					       s << 9,
4578 					       bvec[idx].bv_page,
4579 					       READ, false);
4580 			if (success)
4581 				break;
4582 		failed:
4583 			slot++;
4584 			if (slot >= conf->copies)
4585 				slot = 0;
4586 			if (slot == first_slot)
4587 				break;
4588 		}
4589 		if (!success) {
4590 			/* couldn't read this block, must give up */
4591 			set_bit(MD_RECOVERY_INTR,
4592 				&mddev->recovery);
4593 			return -EIO;
4594 		}
4595 		sectors -= s;
4596 		idx++;
4597 	}
4598 	return 0;
4599 }
4600 
4601 static void end_reshape_write(struct bio *bio, int error)
4602 {
4603 	int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4604 	struct r10bio *r10_bio = bio->bi_private;
4605 	struct mddev *mddev = r10_bio->mddev;
4606 	struct r10conf *conf = mddev->private;
4607 	int d;
4608 	int slot;
4609 	int repl;
4610 	struct md_rdev *rdev = NULL;
4611 
4612 	d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4613 	if (repl)
4614 		rdev = conf->mirrors[d].replacement;
4615 	if (!rdev) {
4616 		smp_mb();
4617 		rdev = conf->mirrors[d].rdev;
4618 	}
4619 
4620 	if (!uptodate) {
4621 		/* FIXME should record badblock */
4622 		md_error(mddev, rdev);
4623 	}
4624 
4625 	rdev_dec_pending(rdev, mddev);
4626 	end_reshape_request(r10_bio);
4627 }
4628 
4629 static void end_reshape_request(struct r10bio *r10_bio)
4630 {
4631 	if (!atomic_dec_and_test(&r10_bio->remaining))
4632 		return;
4633 	md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4634 	bio_put(r10_bio->master_bio);
4635 	put_buf(r10_bio);
4636 }
4637 
4638 static void raid10_finish_reshape(struct mddev *mddev)
4639 {
4640 	struct r10conf *conf = mddev->private;
4641 
4642 	if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4643 		return;
4644 
4645 	if (mddev->delta_disks > 0) {
4646 		sector_t size = raid10_size(mddev, 0, 0);
4647 		md_set_array_sectors(mddev, size);
4648 		if (mddev->recovery_cp > mddev->resync_max_sectors) {
4649 			mddev->recovery_cp = mddev->resync_max_sectors;
4650 			set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4651 		}
4652 		mddev->resync_max_sectors = size;
4653 		set_capacity(mddev->gendisk, mddev->array_sectors);
4654 		revalidate_disk(mddev->gendisk);
4655 	} else {
4656 		int d;
4657 		for (d = conf->geo.raid_disks ;
4658 		     d < conf->geo.raid_disks - mddev->delta_disks;
4659 		     d++) {
4660 			struct md_rdev *rdev = conf->mirrors[d].rdev;
4661 			if (rdev)
4662 				clear_bit(In_sync, &rdev->flags);
4663 			rdev = conf->mirrors[d].replacement;
4664 			if (rdev)
4665 				clear_bit(In_sync, &rdev->flags);
4666 		}
4667 	}
4668 	mddev->layout = mddev->new_layout;
4669 	mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4670 	mddev->reshape_position = MaxSector;
4671 	mddev->delta_disks = 0;
4672 	mddev->reshape_backwards = 0;
4673 }
4674 
4675 static struct md_personality raid10_personality =
4676 {
4677 	.name		= "raid10",
4678 	.level		= 10,
4679 	.owner		= THIS_MODULE,
4680 	.make_request	= make_request,
4681 	.run		= run,
4682 	.stop		= stop,
4683 	.status		= status,
4684 	.error_handler	= error,
4685 	.hot_add_disk	= raid10_add_disk,
4686 	.hot_remove_disk= raid10_remove_disk,
4687 	.spare_active	= raid10_spare_active,
4688 	.sync_request	= sync_request,
4689 	.quiesce	= raid10_quiesce,
4690 	.size		= raid10_size,
4691 	.resize		= raid10_resize,
4692 	.takeover	= raid10_takeover,
4693 	.check_reshape	= raid10_check_reshape,
4694 	.start_reshape	= raid10_start_reshape,
4695 	.finish_reshape	= raid10_finish_reshape,
4696 };
4697 
4698 static int __init raid_init(void)
4699 {
4700 	return register_md_personality(&raid10_personality);
4701 }
4702 
4703 static void raid_exit(void)
4704 {
4705 	unregister_md_personality(&raid10_personality);
4706 }
4707 
4708 module_init(raid_init);
4709 module_exit(raid_exit);
4710 MODULE_LICENSE("GPL");
4711 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4712 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4713 MODULE_ALIAS("md-raid10");
4714 MODULE_ALIAS("md-level-10");
4715 
4716 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
4717