xref: /openbmc/linux/drivers/md/raid10.c (revision 82003e04)
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  *    use_far_sets_bugfixed (stored in bit 18 of layout )
43  *
44  * The data to be stored is divided into chunks using chunksize.  Each device
45  * is divided into far_copies sections.   In each section, chunks are laid out
46  * in a style similar to raid0, but near_copies copies of each chunk is stored
47  * (each on a different drive).  The starting device for each section is offset
48  * near_copies from the starting device of the previous section.  Thus there
49  * are (near_copies * far_copies) of each chunk, and each is on a different
50  * drive.  near_copies and far_copies must be at least one, and their product
51  * is at most raid_disks.
52  *
53  * If far_offset is true, then the far_copies are handled a bit differently.
54  * The copies are still in different stripes, but instead of being very far
55  * apart on disk, there are adjacent stripes.
56  *
57  * The far and offset algorithms are handled slightly differently if
58  * 'use_far_sets' is true.  In this case, the array's devices are grouped into
59  * sets that are (near_copies * far_copies) in size.  The far copied stripes
60  * are still shifted by 'near_copies' devices, but this shifting stays confined
61  * to the set rather than the entire array.  This is done to improve the number
62  * of device combinations that can fail without causing the array to fail.
63  * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
64  * on a device):
65  *    A B C D    A B C D E
66  *      ...         ...
67  *    D A B C    E A B C D
68  * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
69  *    [A B] [C D]    [A B] [C D E]
70  *    |...| |...|    |...| | ... |
71  *    [B A] [D C]    [B A] [E C D]
72  */
73 
74 /*
75  * Number of guaranteed r10bios in case of extreme VM load:
76  */
77 #define	NR_RAID10_BIOS 256
78 
79 /* when we get a read error on a read-only array, we redirect to another
80  * device without failing the first device, or trying to over-write to
81  * correct the read error.  To keep track of bad blocks on a per-bio
82  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
83  */
84 #define IO_BLOCKED ((struct bio *)1)
85 /* When we successfully write to a known bad-block, we need to remove the
86  * bad-block marking which must be done from process context.  So we record
87  * the success by setting devs[n].bio to IO_MADE_GOOD
88  */
89 #define IO_MADE_GOOD ((struct bio *)2)
90 
91 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
92 
93 /* When there are this many requests queued to be written by
94  * the raid10 thread, we become 'congested' to provide back-pressure
95  * for writeback.
96  */
97 static int max_queued_requests = 1024;
98 
99 static void allow_barrier(struct r10conf *conf);
100 static void lower_barrier(struct r10conf *conf);
101 static int _enough(struct r10conf *conf, int previous, int ignore);
102 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
103 				int *skipped);
104 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
105 static void end_reshape_write(struct bio *bio);
106 static void end_reshape(struct r10conf *conf);
107 
108 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
109 {
110 	struct r10conf *conf = data;
111 	int size = offsetof(struct r10bio, devs[conf->copies]);
112 
113 	/* allocate a r10bio with room for raid_disks entries in the
114 	 * bios array */
115 	return kzalloc(size, gfp_flags);
116 }
117 
118 static void r10bio_pool_free(void *r10_bio, void *data)
119 {
120 	kfree(r10_bio);
121 }
122 
123 /* Maximum size of each resync request */
124 #define RESYNC_BLOCK_SIZE (64*1024)
125 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
126 /* amount of memory to reserve for resync requests */
127 #define RESYNC_WINDOW (1024*1024)
128 /* maximum number of concurrent requests, memory permitting */
129 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
130 
131 /*
132  * When performing a resync, we need to read and compare, so
133  * we need as many pages are there are copies.
134  * When performing a recovery, we need 2 bios, one for read,
135  * one for write (we recover only one drive per r10buf)
136  *
137  */
138 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
139 {
140 	struct r10conf *conf = data;
141 	struct page *page;
142 	struct r10bio *r10_bio;
143 	struct bio *bio;
144 	int i, j;
145 	int nalloc;
146 
147 	r10_bio = r10bio_pool_alloc(gfp_flags, conf);
148 	if (!r10_bio)
149 		return NULL;
150 
151 	if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
152 	    test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
153 		nalloc = conf->copies; /* resync */
154 	else
155 		nalloc = 2; /* recovery */
156 
157 	/*
158 	 * Allocate bios.
159 	 */
160 	for (j = nalloc ; j-- ; ) {
161 		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
162 		if (!bio)
163 			goto out_free_bio;
164 		r10_bio->devs[j].bio = bio;
165 		if (!conf->have_replacement)
166 			continue;
167 		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
168 		if (!bio)
169 			goto out_free_bio;
170 		r10_bio->devs[j].repl_bio = bio;
171 	}
172 	/*
173 	 * Allocate RESYNC_PAGES data pages and attach them
174 	 * where needed.
175 	 */
176 	for (j = 0 ; j < nalloc; j++) {
177 		struct bio *rbio = r10_bio->devs[j].repl_bio;
178 		bio = r10_bio->devs[j].bio;
179 		for (i = 0; i < RESYNC_PAGES; i++) {
180 			if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
181 					       &conf->mddev->recovery)) {
182 				/* we can share bv_page's during recovery
183 				 * and reshape */
184 				struct bio *rbio = r10_bio->devs[0].bio;
185 				page = rbio->bi_io_vec[i].bv_page;
186 				get_page(page);
187 			} else
188 				page = alloc_page(gfp_flags);
189 			if (unlikely(!page))
190 				goto out_free_pages;
191 
192 			bio->bi_io_vec[i].bv_page = page;
193 			if (rbio)
194 				rbio->bi_io_vec[i].bv_page = page;
195 		}
196 	}
197 
198 	return r10_bio;
199 
200 out_free_pages:
201 	for ( ; i > 0 ; i--)
202 		safe_put_page(bio->bi_io_vec[i-1].bv_page);
203 	while (j--)
204 		for (i = 0; i < RESYNC_PAGES ; i++)
205 			safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
206 	j = 0;
207 out_free_bio:
208 	for ( ; j < nalloc; j++) {
209 		if (r10_bio->devs[j].bio)
210 			bio_put(r10_bio->devs[j].bio);
211 		if (r10_bio->devs[j].repl_bio)
212 			bio_put(r10_bio->devs[j].repl_bio);
213 	}
214 	r10bio_pool_free(r10_bio, conf);
215 	return NULL;
216 }
217 
218 static void r10buf_pool_free(void *__r10_bio, void *data)
219 {
220 	int i;
221 	struct r10conf *conf = data;
222 	struct r10bio *r10bio = __r10_bio;
223 	int j;
224 
225 	for (j=0; j < conf->copies; j++) {
226 		struct bio *bio = r10bio->devs[j].bio;
227 		if (bio) {
228 			for (i = 0; i < RESYNC_PAGES; i++) {
229 				safe_put_page(bio->bi_io_vec[i].bv_page);
230 				bio->bi_io_vec[i].bv_page = NULL;
231 			}
232 			bio_put(bio);
233 		}
234 		bio = r10bio->devs[j].repl_bio;
235 		if (bio)
236 			bio_put(bio);
237 	}
238 	r10bio_pool_free(r10bio, conf);
239 }
240 
241 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
242 {
243 	int i;
244 
245 	for (i = 0; i < conf->copies; i++) {
246 		struct bio **bio = & r10_bio->devs[i].bio;
247 		if (!BIO_SPECIAL(*bio))
248 			bio_put(*bio);
249 		*bio = NULL;
250 		bio = &r10_bio->devs[i].repl_bio;
251 		if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
252 			bio_put(*bio);
253 		*bio = NULL;
254 	}
255 }
256 
257 static void free_r10bio(struct r10bio *r10_bio)
258 {
259 	struct r10conf *conf = r10_bio->mddev->private;
260 
261 	put_all_bios(conf, r10_bio);
262 	mempool_free(r10_bio, conf->r10bio_pool);
263 }
264 
265 static void put_buf(struct r10bio *r10_bio)
266 {
267 	struct r10conf *conf = r10_bio->mddev->private;
268 
269 	mempool_free(r10_bio, conf->r10buf_pool);
270 
271 	lower_barrier(conf);
272 }
273 
274 static void reschedule_retry(struct r10bio *r10_bio)
275 {
276 	unsigned long flags;
277 	struct mddev *mddev = r10_bio->mddev;
278 	struct r10conf *conf = mddev->private;
279 
280 	spin_lock_irqsave(&conf->device_lock, flags);
281 	list_add(&r10_bio->retry_list, &conf->retry_list);
282 	conf->nr_queued ++;
283 	spin_unlock_irqrestore(&conf->device_lock, flags);
284 
285 	/* wake up frozen array... */
286 	wake_up(&conf->wait_barrier);
287 
288 	md_wakeup_thread(mddev->thread);
289 }
290 
291 /*
292  * raid_end_bio_io() is called when we have finished servicing a mirrored
293  * operation and are ready to return a success/failure code to the buffer
294  * cache layer.
295  */
296 static void raid_end_bio_io(struct r10bio *r10_bio)
297 {
298 	struct bio *bio = r10_bio->master_bio;
299 	int done;
300 	struct r10conf *conf = r10_bio->mddev->private;
301 
302 	if (bio->bi_phys_segments) {
303 		unsigned long flags;
304 		spin_lock_irqsave(&conf->device_lock, flags);
305 		bio->bi_phys_segments--;
306 		done = (bio->bi_phys_segments == 0);
307 		spin_unlock_irqrestore(&conf->device_lock, flags);
308 	} else
309 		done = 1;
310 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
311 		bio->bi_error = -EIO;
312 	if (done) {
313 		bio_endio(bio);
314 		/*
315 		 * Wake up any possible resync thread that waits for the device
316 		 * to go idle.
317 		 */
318 		allow_barrier(conf);
319 	}
320 	free_r10bio(r10_bio);
321 }
322 
323 /*
324  * Update disk head position estimator based on IRQ completion info.
325  */
326 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
327 {
328 	struct r10conf *conf = r10_bio->mddev->private;
329 
330 	conf->mirrors[r10_bio->devs[slot].devnum].head_position =
331 		r10_bio->devs[slot].addr + (r10_bio->sectors);
332 }
333 
334 /*
335  * Find the disk number which triggered given bio
336  */
337 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
338 			 struct bio *bio, int *slotp, int *replp)
339 {
340 	int slot;
341 	int repl = 0;
342 
343 	for (slot = 0; slot < conf->copies; slot++) {
344 		if (r10_bio->devs[slot].bio == bio)
345 			break;
346 		if (r10_bio->devs[slot].repl_bio == bio) {
347 			repl = 1;
348 			break;
349 		}
350 	}
351 
352 	BUG_ON(slot == conf->copies);
353 	update_head_pos(slot, r10_bio);
354 
355 	if (slotp)
356 		*slotp = slot;
357 	if (replp)
358 		*replp = repl;
359 	return r10_bio->devs[slot].devnum;
360 }
361 
362 static void raid10_end_read_request(struct bio *bio)
363 {
364 	int uptodate = !bio->bi_error;
365 	struct r10bio *r10_bio = bio->bi_private;
366 	int slot, dev;
367 	struct md_rdev *rdev;
368 	struct r10conf *conf = r10_bio->mddev->private;
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 		if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
396 			     rdev->raid_disk))
397 			uptodate = 1;
398 	}
399 	if (uptodate) {
400 		raid_end_bio_io(r10_bio);
401 		rdev_dec_pending(rdev, conf->mddev);
402 	} else {
403 		/*
404 		 * oops, read error - keep the refcount on the rdev
405 		 */
406 		char b[BDEVNAME_SIZE];
407 		printk_ratelimited(KERN_ERR
408 				   "md/raid10:%s: %s: rescheduling sector %llu\n",
409 				   mdname(conf->mddev),
410 				   bdevname(rdev->bdev, b),
411 				   (unsigned long long)r10_bio->sector);
412 		set_bit(R10BIO_ReadError, &r10_bio->state);
413 		reschedule_retry(r10_bio);
414 	}
415 }
416 
417 static void close_write(struct r10bio *r10_bio)
418 {
419 	/* clear the bitmap if all writes complete successfully */
420 	bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
421 			r10_bio->sectors,
422 			!test_bit(R10BIO_Degraded, &r10_bio->state),
423 			0);
424 	md_write_end(r10_bio->mddev);
425 }
426 
427 static void one_write_done(struct r10bio *r10_bio)
428 {
429 	if (atomic_dec_and_test(&r10_bio->remaining)) {
430 		if (test_bit(R10BIO_WriteError, &r10_bio->state))
431 			reschedule_retry(r10_bio);
432 		else {
433 			close_write(r10_bio);
434 			if (test_bit(R10BIO_MadeGood, &r10_bio->state))
435 				reschedule_retry(r10_bio);
436 			else
437 				raid_end_bio_io(r10_bio);
438 		}
439 	}
440 }
441 
442 static void raid10_end_write_request(struct bio *bio)
443 {
444 	struct r10bio *r10_bio = bio->bi_private;
445 	int dev;
446 	int dec_rdev = 1;
447 	struct r10conf *conf = r10_bio->mddev->private;
448 	int slot, repl;
449 	struct md_rdev *rdev = NULL;
450 	bool discard_error;
451 
452 	discard_error = bio->bi_error && bio_op(bio) == REQ_OP_DISCARD;
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 (bio->bi_error && !discard_error) {
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 		/*
494 		 * Do not set R10BIO_Uptodate if the current device is
495 		 * rebuilding or Faulty. This is because we cannot use
496 		 * such device for properly reading the data back (we could
497 		 * potentially use it, if the current write would have felt
498 		 * before rdev->recovery_offset, but for simplicity we don't
499 		 * check this here.
500 		 */
501 		if (test_bit(In_sync, &rdev->flags) &&
502 		    !test_bit(Faulty, &rdev->flags))
503 			set_bit(R10BIO_Uptodate, &r10_bio->state);
504 
505 		/* Maybe we can clear some bad blocks. */
506 		if (is_badblock(rdev,
507 				r10_bio->devs[slot].addr,
508 				r10_bio->sectors,
509 				&first_bad, &bad_sectors) && !discard_error) {
510 			bio_put(bio);
511 			if (repl)
512 				r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
513 			else
514 				r10_bio->devs[slot].bio = IO_MADE_GOOD;
515 			dec_rdev = 0;
516 			set_bit(R10BIO_MadeGood, &r10_bio->state);
517 		}
518 	}
519 
520 	/*
521 	 *
522 	 * Let's see if all mirrored write operations have finished
523 	 * already.
524 	 */
525 	one_write_done(r10_bio);
526 	if (dec_rdev)
527 		rdev_dec_pending(rdev, conf->mddev);
528 }
529 
530 /*
531  * RAID10 layout manager
532  * As well as the chunksize and raid_disks count, there are two
533  * parameters: near_copies and far_copies.
534  * near_copies * far_copies must be <= raid_disks.
535  * Normally one of these will be 1.
536  * If both are 1, we get raid0.
537  * If near_copies == raid_disks, we get raid1.
538  *
539  * Chunks are laid out in raid0 style with near_copies copies of the
540  * first chunk, followed by near_copies copies of the next chunk and
541  * so on.
542  * If far_copies > 1, then after 1/far_copies of the array has been assigned
543  * as described above, we start again with a device offset of near_copies.
544  * So we effectively have another copy of the whole array further down all
545  * the drives, but with blocks on different drives.
546  * With this layout, and block is never stored twice on the one device.
547  *
548  * raid10_find_phys finds the sector offset of a given virtual sector
549  * on each device that it is on.
550  *
551  * raid10_find_virt does the reverse mapping, from a device and a
552  * sector offset to a virtual address
553  */
554 
555 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
556 {
557 	int n,f;
558 	sector_t sector;
559 	sector_t chunk;
560 	sector_t stripe;
561 	int dev;
562 	int slot = 0;
563 	int last_far_set_start, last_far_set_size;
564 
565 	last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
566 	last_far_set_start *= geo->far_set_size;
567 
568 	last_far_set_size = geo->far_set_size;
569 	last_far_set_size += (geo->raid_disks % geo->far_set_size);
570 
571 	/* now calculate first sector/dev */
572 	chunk = r10bio->sector >> geo->chunk_shift;
573 	sector = r10bio->sector & geo->chunk_mask;
574 
575 	chunk *= geo->near_copies;
576 	stripe = chunk;
577 	dev = sector_div(stripe, geo->raid_disks);
578 	if (geo->far_offset)
579 		stripe *= geo->far_copies;
580 
581 	sector += stripe << geo->chunk_shift;
582 
583 	/* and calculate all the others */
584 	for (n = 0; n < geo->near_copies; n++) {
585 		int d = dev;
586 		int set;
587 		sector_t s = sector;
588 		r10bio->devs[slot].devnum = d;
589 		r10bio->devs[slot].addr = s;
590 		slot++;
591 
592 		for (f = 1; f < geo->far_copies; f++) {
593 			set = d / geo->far_set_size;
594 			d += geo->near_copies;
595 
596 			if ((geo->raid_disks % geo->far_set_size) &&
597 			    (d > last_far_set_start)) {
598 				d -= last_far_set_start;
599 				d %= last_far_set_size;
600 				d += last_far_set_start;
601 			} else {
602 				d %= geo->far_set_size;
603 				d += geo->far_set_size * set;
604 			}
605 			s += geo->stride;
606 			r10bio->devs[slot].devnum = d;
607 			r10bio->devs[slot].addr = s;
608 			slot++;
609 		}
610 		dev++;
611 		if (dev >= geo->raid_disks) {
612 			dev = 0;
613 			sector += (geo->chunk_mask + 1);
614 		}
615 	}
616 }
617 
618 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
619 {
620 	struct geom *geo = &conf->geo;
621 
622 	if (conf->reshape_progress != MaxSector &&
623 	    ((r10bio->sector >= conf->reshape_progress) !=
624 	     conf->mddev->reshape_backwards)) {
625 		set_bit(R10BIO_Previous, &r10bio->state);
626 		geo = &conf->prev;
627 	} else
628 		clear_bit(R10BIO_Previous, &r10bio->state);
629 
630 	__raid10_find_phys(geo, r10bio);
631 }
632 
633 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
634 {
635 	sector_t offset, chunk, vchunk;
636 	/* Never use conf->prev as this is only called during resync
637 	 * or recovery, so reshape isn't happening
638 	 */
639 	struct geom *geo = &conf->geo;
640 	int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
641 	int far_set_size = geo->far_set_size;
642 	int last_far_set_start;
643 
644 	if (geo->raid_disks % geo->far_set_size) {
645 		last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
646 		last_far_set_start *= geo->far_set_size;
647 
648 		if (dev >= last_far_set_start) {
649 			far_set_size = geo->far_set_size;
650 			far_set_size += (geo->raid_disks % geo->far_set_size);
651 			far_set_start = last_far_set_start;
652 		}
653 	}
654 
655 	offset = sector & geo->chunk_mask;
656 	if (geo->far_offset) {
657 		int fc;
658 		chunk = sector >> geo->chunk_shift;
659 		fc = sector_div(chunk, geo->far_copies);
660 		dev -= fc * geo->near_copies;
661 		if (dev < far_set_start)
662 			dev += far_set_size;
663 	} else {
664 		while (sector >= geo->stride) {
665 			sector -= geo->stride;
666 			if (dev < (geo->near_copies + far_set_start))
667 				dev += far_set_size - geo->near_copies;
668 			else
669 				dev -= geo->near_copies;
670 		}
671 		chunk = sector >> geo->chunk_shift;
672 	}
673 	vchunk = chunk * geo->raid_disks + dev;
674 	sector_div(vchunk, geo->near_copies);
675 	return (vchunk << geo->chunk_shift) + offset;
676 }
677 
678 /*
679  * This routine returns the disk from which the requested read should
680  * be done. There is a per-array 'next expected sequential IO' sector
681  * number - if this matches on the next IO then we use the last disk.
682  * There is also a per-disk 'last know head position' sector that is
683  * maintained from IRQ contexts, both the normal and the resync IO
684  * completion handlers update this position correctly. If there is no
685  * perfect sequential match then we pick the disk whose head is closest.
686  *
687  * If there are 2 mirrors in the same 2 devices, performance degrades
688  * because position is mirror, not device based.
689  *
690  * The rdev for the device selected will have nr_pending incremented.
691  */
692 
693 /*
694  * FIXME: possibly should rethink readbalancing and do it differently
695  * depending on near_copies / far_copies geometry.
696  */
697 static struct md_rdev *read_balance(struct r10conf *conf,
698 				    struct r10bio *r10_bio,
699 				    int *max_sectors)
700 {
701 	const sector_t this_sector = r10_bio->sector;
702 	int disk, slot;
703 	int sectors = r10_bio->sectors;
704 	int best_good_sectors;
705 	sector_t new_distance, best_dist;
706 	struct md_rdev *best_rdev, *rdev = NULL;
707 	int do_balance;
708 	int best_slot;
709 	struct geom *geo = &conf->geo;
710 
711 	raid10_find_phys(conf, r10_bio);
712 	rcu_read_lock();
713 	sectors = r10_bio->sectors;
714 	best_slot = -1;
715 	best_rdev = NULL;
716 	best_dist = MaxSector;
717 	best_good_sectors = 0;
718 	do_balance = 1;
719 	/*
720 	 * Check if we can balance. We can balance on the whole
721 	 * device if no resync is going on (recovery is ok), or below
722 	 * the resync window. We take the first readable disk when
723 	 * above the resync window.
724 	 */
725 	if (conf->mddev->recovery_cp < MaxSector
726 	    && (this_sector + sectors >= conf->next_resync))
727 		do_balance = 0;
728 
729 	for (slot = 0; slot < conf->copies ; slot++) {
730 		sector_t first_bad;
731 		int bad_sectors;
732 		sector_t dev_sector;
733 
734 		if (r10_bio->devs[slot].bio == IO_BLOCKED)
735 			continue;
736 		disk = r10_bio->devs[slot].devnum;
737 		rdev = rcu_dereference(conf->mirrors[disk].replacement);
738 		if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
739 		    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
740 			rdev = rcu_dereference(conf->mirrors[disk].rdev);
741 		if (rdev == NULL ||
742 		    test_bit(Faulty, &rdev->flags))
743 			continue;
744 		if (!test_bit(In_sync, &rdev->flags) &&
745 		    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
746 			continue;
747 
748 		dev_sector = r10_bio->devs[slot].addr;
749 		if (is_badblock(rdev, dev_sector, sectors,
750 				&first_bad, &bad_sectors)) {
751 			if (best_dist < MaxSector)
752 				/* Already have a better slot */
753 				continue;
754 			if (first_bad <= dev_sector) {
755 				/* Cannot read here.  If this is the
756 				 * 'primary' device, then we must not read
757 				 * beyond 'bad_sectors' from another device.
758 				 */
759 				bad_sectors -= (dev_sector - first_bad);
760 				if (!do_balance && sectors > bad_sectors)
761 					sectors = bad_sectors;
762 				if (best_good_sectors > sectors)
763 					best_good_sectors = sectors;
764 			} else {
765 				sector_t good_sectors =
766 					first_bad - dev_sector;
767 				if (good_sectors > best_good_sectors) {
768 					best_good_sectors = good_sectors;
769 					best_slot = slot;
770 					best_rdev = rdev;
771 				}
772 				if (!do_balance)
773 					/* Must read from here */
774 					break;
775 			}
776 			continue;
777 		} else
778 			best_good_sectors = sectors;
779 
780 		if (!do_balance)
781 			break;
782 
783 		/* This optimisation is debatable, and completely destroys
784 		 * sequential read speed for 'far copies' arrays.  So only
785 		 * keep it for 'near' arrays, and review those later.
786 		 */
787 		if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
788 			break;
789 
790 		/* for far > 1 always use the lowest address */
791 		if (geo->far_copies > 1)
792 			new_distance = r10_bio->devs[slot].addr;
793 		else
794 			new_distance = abs(r10_bio->devs[slot].addr -
795 					   conf->mirrors[disk].head_position);
796 		if (new_distance < best_dist) {
797 			best_dist = new_distance;
798 			best_slot = slot;
799 			best_rdev = rdev;
800 		}
801 	}
802 	if (slot >= conf->copies) {
803 		slot = best_slot;
804 		rdev = best_rdev;
805 	}
806 
807 	if (slot >= 0) {
808 		atomic_inc(&rdev->nr_pending);
809 		r10_bio->read_slot = slot;
810 	} else
811 		rdev = NULL;
812 	rcu_read_unlock();
813 	*max_sectors = best_good_sectors;
814 
815 	return rdev;
816 }
817 
818 static int raid10_congested(struct mddev *mddev, int bits)
819 {
820 	struct r10conf *conf = mddev->private;
821 	int i, ret = 0;
822 
823 	if ((bits & (1 << WB_async_congested)) &&
824 	    conf->pending_count >= max_queued_requests)
825 		return 1;
826 
827 	rcu_read_lock();
828 	for (i = 0;
829 	     (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
830 		     && ret == 0;
831 	     i++) {
832 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
833 		if (rdev && !test_bit(Faulty, &rdev->flags)) {
834 			struct request_queue *q = bdev_get_queue(rdev->bdev);
835 
836 			ret |= bdi_congested(&q->backing_dev_info, bits);
837 		}
838 	}
839 	rcu_read_unlock();
840 	return ret;
841 }
842 
843 static void flush_pending_writes(struct r10conf *conf)
844 {
845 	/* Any writes that have been queued but are awaiting
846 	 * bitmap updates get flushed here.
847 	 */
848 	spin_lock_irq(&conf->device_lock);
849 
850 	if (conf->pending_bio_list.head) {
851 		struct bio *bio;
852 		bio = bio_list_get(&conf->pending_bio_list);
853 		conf->pending_count = 0;
854 		spin_unlock_irq(&conf->device_lock);
855 		/* flush any pending bitmap writes to disk
856 		 * before proceeding w/ I/O */
857 		bitmap_unplug(conf->mddev->bitmap);
858 		wake_up(&conf->wait_barrier);
859 
860 		while (bio) { /* submit pending writes */
861 			struct bio *next = bio->bi_next;
862 			bio->bi_next = NULL;
863 			if (unlikely((bio_op(bio) ==  REQ_OP_DISCARD) &&
864 			    !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
865 				/* Just ignore it */
866 				bio_endio(bio);
867 			else
868 				generic_make_request(bio);
869 			bio = next;
870 		}
871 	} else
872 		spin_unlock_irq(&conf->device_lock);
873 }
874 
875 /* Barriers....
876  * Sometimes we need to suspend IO while we do something else,
877  * either some resync/recovery, or reconfigure the array.
878  * To do this we raise a 'barrier'.
879  * The 'barrier' is a counter that can be raised multiple times
880  * to count how many activities are happening which preclude
881  * normal IO.
882  * We can only raise the barrier if there is no pending IO.
883  * i.e. if nr_pending == 0.
884  * We choose only to raise the barrier if no-one is waiting for the
885  * barrier to go down.  This means that as soon as an IO request
886  * is ready, no other operations which require a barrier will start
887  * until the IO request has had a chance.
888  *
889  * So: regular IO calls 'wait_barrier'.  When that returns there
890  *    is no backgroup IO happening,  It must arrange to call
891  *    allow_barrier when it has finished its IO.
892  * backgroup IO calls must call raise_barrier.  Once that returns
893  *    there is no normal IO happeing.  It must arrange to call
894  *    lower_barrier when the particular background IO completes.
895  */
896 
897 static void raise_barrier(struct r10conf *conf, int force)
898 {
899 	BUG_ON(force && !conf->barrier);
900 	spin_lock_irq(&conf->resync_lock);
901 
902 	/* Wait until no block IO is waiting (unless 'force') */
903 	wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
904 			    conf->resync_lock);
905 
906 	/* block any new IO from starting */
907 	conf->barrier++;
908 
909 	/* Now wait for all pending IO to complete */
910 	wait_event_lock_irq(conf->wait_barrier,
911 			    !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
912 			    conf->resync_lock);
913 
914 	spin_unlock_irq(&conf->resync_lock);
915 }
916 
917 static void lower_barrier(struct r10conf *conf)
918 {
919 	unsigned long flags;
920 	spin_lock_irqsave(&conf->resync_lock, flags);
921 	conf->barrier--;
922 	spin_unlock_irqrestore(&conf->resync_lock, flags);
923 	wake_up(&conf->wait_barrier);
924 }
925 
926 static void wait_barrier(struct r10conf *conf)
927 {
928 	spin_lock_irq(&conf->resync_lock);
929 	if (conf->barrier) {
930 		conf->nr_waiting++;
931 		/* Wait for the barrier to drop.
932 		 * However if there are already pending
933 		 * requests (preventing the barrier from
934 		 * rising completely), and the
935 		 * pre-process bio queue isn't empty,
936 		 * then don't wait, as we need to empty
937 		 * that queue to get the nr_pending
938 		 * count down.
939 		 */
940 		wait_event_lock_irq(conf->wait_barrier,
941 				    !conf->barrier ||
942 				    (atomic_read(&conf->nr_pending) &&
943 				     current->bio_list &&
944 				     !bio_list_empty(current->bio_list)),
945 				    conf->resync_lock);
946 		conf->nr_waiting--;
947 		if (!conf->nr_waiting)
948 			wake_up(&conf->wait_barrier);
949 	}
950 	atomic_inc(&conf->nr_pending);
951 	spin_unlock_irq(&conf->resync_lock);
952 }
953 
954 static void allow_barrier(struct r10conf *conf)
955 {
956 	if ((atomic_dec_and_test(&conf->nr_pending)) ||
957 			(conf->array_freeze_pending))
958 		wake_up(&conf->wait_barrier);
959 }
960 
961 static void freeze_array(struct r10conf *conf, int extra)
962 {
963 	/* stop syncio and normal IO and wait for everything to
964 	 * go quiet.
965 	 * We increment barrier and nr_waiting, and then
966 	 * wait until nr_pending match nr_queued+extra
967 	 * This is called in the context of one normal IO request
968 	 * that has failed. Thus any sync request that might be pending
969 	 * will be blocked by nr_pending, and we need to wait for
970 	 * pending IO requests to complete or be queued for re-try.
971 	 * Thus the number queued (nr_queued) plus this request (extra)
972 	 * must match the number of pending IOs (nr_pending) before
973 	 * we continue.
974 	 */
975 	spin_lock_irq(&conf->resync_lock);
976 	conf->array_freeze_pending++;
977 	conf->barrier++;
978 	conf->nr_waiting++;
979 	wait_event_lock_irq_cmd(conf->wait_barrier,
980 				atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
981 				conf->resync_lock,
982 				flush_pending_writes(conf));
983 
984 	conf->array_freeze_pending--;
985 	spin_unlock_irq(&conf->resync_lock);
986 }
987 
988 static void unfreeze_array(struct r10conf *conf)
989 {
990 	/* reverse the effect of the freeze */
991 	spin_lock_irq(&conf->resync_lock);
992 	conf->barrier--;
993 	conf->nr_waiting--;
994 	wake_up(&conf->wait_barrier);
995 	spin_unlock_irq(&conf->resync_lock);
996 }
997 
998 static sector_t choose_data_offset(struct r10bio *r10_bio,
999 				   struct md_rdev *rdev)
1000 {
1001 	if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1002 	    test_bit(R10BIO_Previous, &r10_bio->state))
1003 		return rdev->data_offset;
1004 	else
1005 		return rdev->new_data_offset;
1006 }
1007 
1008 struct raid10_plug_cb {
1009 	struct blk_plug_cb	cb;
1010 	struct bio_list		pending;
1011 	int			pending_cnt;
1012 };
1013 
1014 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1015 {
1016 	struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1017 						   cb);
1018 	struct mddev *mddev = plug->cb.data;
1019 	struct r10conf *conf = mddev->private;
1020 	struct bio *bio;
1021 
1022 	if (from_schedule || current->bio_list) {
1023 		spin_lock_irq(&conf->device_lock);
1024 		bio_list_merge(&conf->pending_bio_list, &plug->pending);
1025 		conf->pending_count += plug->pending_cnt;
1026 		spin_unlock_irq(&conf->device_lock);
1027 		wake_up(&conf->wait_barrier);
1028 		md_wakeup_thread(mddev->thread);
1029 		kfree(plug);
1030 		return;
1031 	}
1032 
1033 	/* we aren't scheduling, so we can do the write-out directly. */
1034 	bio = bio_list_get(&plug->pending);
1035 	bitmap_unplug(mddev->bitmap);
1036 	wake_up(&conf->wait_barrier);
1037 
1038 	while (bio) { /* submit pending writes */
1039 		struct bio *next = bio->bi_next;
1040 		bio->bi_next = NULL;
1041 		if (unlikely((bio_op(bio) ==  REQ_OP_DISCARD) &&
1042 		    !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1043 			/* Just ignore it */
1044 			bio_endio(bio);
1045 		else
1046 			generic_make_request(bio);
1047 		bio = next;
1048 	}
1049 	kfree(plug);
1050 }
1051 
1052 static void __make_request(struct mddev *mddev, struct bio *bio)
1053 {
1054 	struct r10conf *conf = mddev->private;
1055 	struct r10bio *r10_bio;
1056 	struct bio *read_bio;
1057 	int i;
1058 	const int op = bio_op(bio);
1059 	const int rw = bio_data_dir(bio);
1060 	const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1061 	const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1062 	unsigned long flags;
1063 	struct md_rdev *blocked_rdev;
1064 	struct blk_plug_cb *cb;
1065 	struct raid10_plug_cb *plug = NULL;
1066 	int sectors_handled;
1067 	int max_sectors;
1068 	int sectors;
1069 
1070 	md_write_start(mddev, bio);
1071 
1072 	/*
1073 	 * Register the new request and wait if the reconstruction
1074 	 * thread has put up a bar for new requests.
1075 	 * Continue immediately if no resync is active currently.
1076 	 */
1077 	wait_barrier(conf);
1078 
1079 	sectors = bio_sectors(bio);
1080 	while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1081 	    bio->bi_iter.bi_sector < conf->reshape_progress &&
1082 	    bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1083 		/* IO spans the reshape position.  Need to wait for
1084 		 * reshape to pass
1085 		 */
1086 		allow_barrier(conf);
1087 		wait_event(conf->wait_barrier,
1088 			   conf->reshape_progress <= bio->bi_iter.bi_sector ||
1089 			   conf->reshape_progress >= bio->bi_iter.bi_sector +
1090 			   sectors);
1091 		wait_barrier(conf);
1092 	}
1093 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1094 	    bio_data_dir(bio) == WRITE &&
1095 	    (mddev->reshape_backwards
1096 	     ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1097 		bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1098 	     : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1099 		bio->bi_iter.bi_sector < conf->reshape_progress))) {
1100 		/* Need to update reshape_position in metadata */
1101 		mddev->reshape_position = conf->reshape_progress;
1102 		set_mask_bits(&mddev->flags, 0,
1103 			      BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1104 		md_wakeup_thread(mddev->thread);
1105 		wait_event(mddev->sb_wait,
1106 			   !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1107 
1108 		conf->reshape_safe = mddev->reshape_position;
1109 	}
1110 
1111 	r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1112 
1113 	r10_bio->master_bio = bio;
1114 	r10_bio->sectors = sectors;
1115 
1116 	r10_bio->mddev = mddev;
1117 	r10_bio->sector = bio->bi_iter.bi_sector;
1118 	r10_bio->state = 0;
1119 
1120 	/* We might need to issue multiple reads to different
1121 	 * devices if there are bad blocks around, so we keep
1122 	 * track of the number of reads in bio->bi_phys_segments.
1123 	 * If this is 0, there is only one r10_bio and no locking
1124 	 * will be needed when the request completes.  If it is
1125 	 * non-zero, then it is the number of not-completed requests.
1126 	 */
1127 	bio->bi_phys_segments = 0;
1128 	bio_clear_flag(bio, BIO_SEG_VALID);
1129 
1130 	if (rw == READ) {
1131 		/*
1132 		 * read balancing logic:
1133 		 */
1134 		struct md_rdev *rdev;
1135 		int slot;
1136 
1137 read_again:
1138 		rdev = read_balance(conf, r10_bio, &max_sectors);
1139 		if (!rdev) {
1140 			raid_end_bio_io(r10_bio);
1141 			return;
1142 		}
1143 		slot = r10_bio->read_slot;
1144 
1145 		read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1146 		bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1147 			 max_sectors);
1148 
1149 		r10_bio->devs[slot].bio = read_bio;
1150 		r10_bio->devs[slot].rdev = rdev;
1151 
1152 		read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1153 			choose_data_offset(r10_bio, rdev);
1154 		read_bio->bi_bdev = rdev->bdev;
1155 		read_bio->bi_end_io = raid10_end_read_request;
1156 		bio_set_op_attrs(read_bio, op, do_sync);
1157 		read_bio->bi_private = r10_bio;
1158 
1159 		if (max_sectors < r10_bio->sectors) {
1160 			/* Could not read all from this device, so we will
1161 			 * need another r10_bio.
1162 			 */
1163 			sectors_handled = (r10_bio->sector + max_sectors
1164 					   - bio->bi_iter.bi_sector);
1165 			r10_bio->sectors = max_sectors;
1166 			spin_lock_irq(&conf->device_lock);
1167 			if (bio->bi_phys_segments == 0)
1168 				bio->bi_phys_segments = 2;
1169 			else
1170 				bio->bi_phys_segments++;
1171 			spin_unlock_irq(&conf->device_lock);
1172 			/* Cannot call generic_make_request directly
1173 			 * as that will be queued in __generic_make_request
1174 			 * and subsequent mempool_alloc might block
1175 			 * waiting for it.  so hand bio over to raid10d.
1176 			 */
1177 			reschedule_retry(r10_bio);
1178 
1179 			r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1180 
1181 			r10_bio->master_bio = bio;
1182 			r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1183 			r10_bio->state = 0;
1184 			r10_bio->mddev = mddev;
1185 			r10_bio->sector = bio->bi_iter.bi_sector +
1186 				sectors_handled;
1187 			goto read_again;
1188 		} else
1189 			generic_make_request(read_bio);
1190 		return;
1191 	}
1192 
1193 	/*
1194 	 * WRITE:
1195 	 */
1196 	if (conf->pending_count >= max_queued_requests) {
1197 		md_wakeup_thread(mddev->thread);
1198 		wait_event(conf->wait_barrier,
1199 			   conf->pending_count < max_queued_requests);
1200 	}
1201 	/* first select target devices under rcu_lock and
1202 	 * inc refcount on their rdev.  Record them by setting
1203 	 * bios[x] to bio
1204 	 * If there are known/acknowledged bad blocks on any device
1205 	 * on which we have seen a write error, we want to avoid
1206 	 * writing to those blocks.  This potentially requires several
1207 	 * writes to write around the bad blocks.  Each set of writes
1208 	 * gets its own r10_bio with a set of bios attached.  The number
1209 	 * of r10_bios is recored in bio->bi_phys_segments just as with
1210 	 * the read case.
1211 	 */
1212 
1213 	r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1214 	raid10_find_phys(conf, r10_bio);
1215 retry_write:
1216 	blocked_rdev = NULL;
1217 	rcu_read_lock();
1218 	max_sectors = r10_bio->sectors;
1219 
1220 	for (i = 0;  i < conf->copies; i++) {
1221 		int d = r10_bio->devs[i].devnum;
1222 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1223 		struct md_rdev *rrdev = rcu_dereference(
1224 			conf->mirrors[d].replacement);
1225 		if (rdev == rrdev)
1226 			rrdev = NULL;
1227 		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1228 			atomic_inc(&rdev->nr_pending);
1229 			blocked_rdev = rdev;
1230 			break;
1231 		}
1232 		if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1233 			atomic_inc(&rrdev->nr_pending);
1234 			blocked_rdev = rrdev;
1235 			break;
1236 		}
1237 		if (rdev && (test_bit(Faulty, &rdev->flags)))
1238 			rdev = NULL;
1239 		if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1240 			rrdev = NULL;
1241 
1242 		r10_bio->devs[i].bio = NULL;
1243 		r10_bio->devs[i].repl_bio = NULL;
1244 
1245 		if (!rdev && !rrdev) {
1246 			set_bit(R10BIO_Degraded, &r10_bio->state);
1247 			continue;
1248 		}
1249 		if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1250 			sector_t first_bad;
1251 			sector_t dev_sector = r10_bio->devs[i].addr;
1252 			int bad_sectors;
1253 			int is_bad;
1254 
1255 			is_bad = is_badblock(rdev, dev_sector,
1256 					     max_sectors,
1257 					     &first_bad, &bad_sectors);
1258 			if (is_bad < 0) {
1259 				/* Mustn't write here until the bad block
1260 				 * is acknowledged
1261 				 */
1262 				atomic_inc(&rdev->nr_pending);
1263 				set_bit(BlockedBadBlocks, &rdev->flags);
1264 				blocked_rdev = rdev;
1265 				break;
1266 			}
1267 			if (is_bad && first_bad <= dev_sector) {
1268 				/* Cannot write here at all */
1269 				bad_sectors -= (dev_sector - first_bad);
1270 				if (bad_sectors < max_sectors)
1271 					/* Mustn't write more than bad_sectors
1272 					 * to other devices yet
1273 					 */
1274 					max_sectors = bad_sectors;
1275 				/* We don't set R10BIO_Degraded as that
1276 				 * only applies if the disk is missing,
1277 				 * so it might be re-added, and we want to
1278 				 * know to recover this chunk.
1279 				 * In this case the device is here, and the
1280 				 * fact that this chunk is not in-sync is
1281 				 * recorded in the bad block log.
1282 				 */
1283 				continue;
1284 			}
1285 			if (is_bad) {
1286 				int good_sectors = first_bad - dev_sector;
1287 				if (good_sectors < max_sectors)
1288 					max_sectors = good_sectors;
1289 			}
1290 		}
1291 		if (rdev) {
1292 			r10_bio->devs[i].bio = bio;
1293 			atomic_inc(&rdev->nr_pending);
1294 		}
1295 		if (rrdev) {
1296 			r10_bio->devs[i].repl_bio = bio;
1297 			atomic_inc(&rrdev->nr_pending);
1298 		}
1299 	}
1300 	rcu_read_unlock();
1301 
1302 	if (unlikely(blocked_rdev)) {
1303 		/* Have to wait for this device to get unblocked, then retry */
1304 		int j;
1305 		int d;
1306 
1307 		for (j = 0; j < i; j++) {
1308 			if (r10_bio->devs[j].bio) {
1309 				d = r10_bio->devs[j].devnum;
1310 				rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1311 			}
1312 			if (r10_bio->devs[j].repl_bio) {
1313 				struct md_rdev *rdev;
1314 				d = r10_bio->devs[j].devnum;
1315 				rdev = conf->mirrors[d].replacement;
1316 				if (!rdev) {
1317 					/* Race with remove_disk */
1318 					smp_mb();
1319 					rdev = conf->mirrors[d].rdev;
1320 				}
1321 				rdev_dec_pending(rdev, mddev);
1322 			}
1323 		}
1324 		allow_barrier(conf);
1325 		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1326 		wait_barrier(conf);
1327 		goto retry_write;
1328 	}
1329 
1330 	if (max_sectors < r10_bio->sectors) {
1331 		/* We are splitting this into multiple parts, so
1332 		 * we need to prepare for allocating another r10_bio.
1333 		 */
1334 		r10_bio->sectors = max_sectors;
1335 		spin_lock_irq(&conf->device_lock);
1336 		if (bio->bi_phys_segments == 0)
1337 			bio->bi_phys_segments = 2;
1338 		else
1339 			bio->bi_phys_segments++;
1340 		spin_unlock_irq(&conf->device_lock);
1341 	}
1342 	sectors_handled = r10_bio->sector + max_sectors -
1343 		bio->bi_iter.bi_sector;
1344 
1345 	atomic_set(&r10_bio->remaining, 1);
1346 	bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1347 
1348 	for (i = 0; i < conf->copies; i++) {
1349 		struct bio *mbio;
1350 		int d = r10_bio->devs[i].devnum;
1351 		if (r10_bio->devs[i].bio) {
1352 			struct md_rdev *rdev = conf->mirrors[d].rdev;
1353 			mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1354 			bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1355 				 max_sectors);
1356 			r10_bio->devs[i].bio = mbio;
1357 
1358 			mbio->bi_iter.bi_sector	= (r10_bio->devs[i].addr+
1359 					   choose_data_offset(r10_bio,
1360 							      rdev));
1361 			mbio->bi_bdev = rdev->bdev;
1362 			mbio->bi_end_io	= raid10_end_write_request;
1363 			bio_set_op_attrs(mbio, op, do_sync | do_fua);
1364 			mbio->bi_private = r10_bio;
1365 
1366 			atomic_inc(&r10_bio->remaining);
1367 
1368 			cb = blk_check_plugged(raid10_unplug, mddev,
1369 					       sizeof(*plug));
1370 			if (cb)
1371 				plug = container_of(cb, struct raid10_plug_cb,
1372 						    cb);
1373 			else
1374 				plug = NULL;
1375 			spin_lock_irqsave(&conf->device_lock, flags);
1376 			if (plug) {
1377 				bio_list_add(&plug->pending, mbio);
1378 				plug->pending_cnt++;
1379 			} else {
1380 				bio_list_add(&conf->pending_bio_list, mbio);
1381 				conf->pending_count++;
1382 			}
1383 			spin_unlock_irqrestore(&conf->device_lock, flags);
1384 			if (!plug)
1385 				md_wakeup_thread(mddev->thread);
1386 		}
1387 
1388 		if (r10_bio->devs[i].repl_bio) {
1389 			struct md_rdev *rdev = conf->mirrors[d].replacement;
1390 			if (rdev == NULL) {
1391 				/* Replacement just got moved to main 'rdev' */
1392 				smp_mb();
1393 				rdev = conf->mirrors[d].rdev;
1394 			}
1395 			mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1396 			bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1397 				 max_sectors);
1398 			r10_bio->devs[i].repl_bio = mbio;
1399 
1400 			mbio->bi_iter.bi_sector	= (r10_bio->devs[i].addr +
1401 					   choose_data_offset(
1402 						   r10_bio, rdev));
1403 			mbio->bi_bdev = rdev->bdev;
1404 			mbio->bi_end_io	= raid10_end_write_request;
1405 			bio_set_op_attrs(mbio, op, do_sync | do_fua);
1406 			mbio->bi_private = r10_bio;
1407 
1408 			atomic_inc(&r10_bio->remaining);
1409 			spin_lock_irqsave(&conf->device_lock, flags);
1410 			bio_list_add(&conf->pending_bio_list, mbio);
1411 			conf->pending_count++;
1412 			spin_unlock_irqrestore(&conf->device_lock, flags);
1413 			if (!mddev_check_plugged(mddev))
1414 				md_wakeup_thread(mddev->thread);
1415 		}
1416 	}
1417 
1418 	/* Don't remove the bias on 'remaining' (one_write_done) until
1419 	 * after checking if we need to go around again.
1420 	 */
1421 
1422 	if (sectors_handled < bio_sectors(bio)) {
1423 		one_write_done(r10_bio);
1424 		/* We need another r10_bio.  It has already been counted
1425 		 * in bio->bi_phys_segments.
1426 		 */
1427 		r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1428 
1429 		r10_bio->master_bio = bio;
1430 		r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1431 
1432 		r10_bio->mddev = mddev;
1433 		r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1434 		r10_bio->state = 0;
1435 		goto retry_write;
1436 	}
1437 	one_write_done(r10_bio);
1438 }
1439 
1440 static void raid10_make_request(struct mddev *mddev, struct bio *bio)
1441 {
1442 	struct r10conf *conf = mddev->private;
1443 	sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1444 	int chunk_sects = chunk_mask + 1;
1445 
1446 	struct bio *split;
1447 
1448 	if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1449 		md_flush_request(mddev, bio);
1450 		return;
1451 	}
1452 
1453 	do {
1454 
1455 		/*
1456 		 * If this request crosses a chunk boundary, we need to split
1457 		 * it.
1458 		 */
1459 		if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1460 			     bio_sectors(bio) > chunk_sects
1461 			     && (conf->geo.near_copies < conf->geo.raid_disks
1462 				 || conf->prev.near_copies <
1463 				 conf->prev.raid_disks))) {
1464 			split = bio_split(bio, chunk_sects -
1465 					  (bio->bi_iter.bi_sector &
1466 					   (chunk_sects - 1)),
1467 					  GFP_NOIO, fs_bio_set);
1468 			bio_chain(split, bio);
1469 		} else {
1470 			split = bio;
1471 		}
1472 
1473 		__make_request(mddev, split);
1474 	} while (split != bio);
1475 
1476 	/* In case raid10d snuck in to freeze_array */
1477 	wake_up(&conf->wait_barrier);
1478 }
1479 
1480 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1481 {
1482 	struct r10conf *conf = mddev->private;
1483 	int i;
1484 
1485 	if (conf->geo.near_copies < conf->geo.raid_disks)
1486 		seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1487 	if (conf->geo.near_copies > 1)
1488 		seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1489 	if (conf->geo.far_copies > 1) {
1490 		if (conf->geo.far_offset)
1491 			seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1492 		else
1493 			seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1494 		if (conf->geo.far_set_size != conf->geo.raid_disks)
1495 			seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1496 	}
1497 	seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1498 					conf->geo.raid_disks - mddev->degraded);
1499 	rcu_read_lock();
1500 	for (i = 0; i < conf->geo.raid_disks; i++) {
1501 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1502 		seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1503 	}
1504 	rcu_read_unlock();
1505 	seq_printf(seq, "]");
1506 }
1507 
1508 /* check if there are enough drives for
1509  * every block to appear on atleast one.
1510  * Don't consider the device numbered 'ignore'
1511  * as we might be about to remove it.
1512  */
1513 static int _enough(struct r10conf *conf, int previous, int ignore)
1514 {
1515 	int first = 0;
1516 	int has_enough = 0;
1517 	int disks, ncopies;
1518 	if (previous) {
1519 		disks = conf->prev.raid_disks;
1520 		ncopies = conf->prev.near_copies;
1521 	} else {
1522 		disks = conf->geo.raid_disks;
1523 		ncopies = conf->geo.near_copies;
1524 	}
1525 
1526 	rcu_read_lock();
1527 	do {
1528 		int n = conf->copies;
1529 		int cnt = 0;
1530 		int this = first;
1531 		while (n--) {
1532 			struct md_rdev *rdev;
1533 			if (this != ignore &&
1534 			    (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1535 			    test_bit(In_sync, &rdev->flags))
1536 				cnt++;
1537 			this = (this+1) % disks;
1538 		}
1539 		if (cnt == 0)
1540 			goto out;
1541 		first = (first + ncopies) % disks;
1542 	} while (first != 0);
1543 	has_enough = 1;
1544 out:
1545 	rcu_read_unlock();
1546 	return has_enough;
1547 }
1548 
1549 static int enough(struct r10conf *conf, int ignore)
1550 {
1551 	/* when calling 'enough', both 'prev' and 'geo' must
1552 	 * be stable.
1553 	 * This is ensured if ->reconfig_mutex or ->device_lock
1554 	 * is held.
1555 	 */
1556 	return _enough(conf, 0, ignore) &&
1557 		_enough(conf, 1, ignore);
1558 }
1559 
1560 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1561 {
1562 	char b[BDEVNAME_SIZE];
1563 	struct r10conf *conf = mddev->private;
1564 	unsigned long flags;
1565 
1566 	/*
1567 	 * If it is not operational, then we have already marked it as dead
1568 	 * else if it is the last working disks, ignore the error, let the
1569 	 * next level up know.
1570 	 * else mark the drive as failed
1571 	 */
1572 	spin_lock_irqsave(&conf->device_lock, flags);
1573 	if (test_bit(In_sync, &rdev->flags)
1574 	    && !enough(conf, rdev->raid_disk)) {
1575 		/*
1576 		 * Don't fail the drive, just return an IO error.
1577 		 */
1578 		spin_unlock_irqrestore(&conf->device_lock, flags);
1579 		return;
1580 	}
1581 	if (test_and_clear_bit(In_sync, &rdev->flags))
1582 		mddev->degraded++;
1583 	/*
1584 	 * If recovery is running, make sure it aborts.
1585 	 */
1586 	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1587 	set_bit(Blocked, &rdev->flags);
1588 	set_bit(Faulty, &rdev->flags);
1589 	set_mask_bits(&mddev->flags, 0,
1590 		      BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1591 	spin_unlock_irqrestore(&conf->device_lock, flags);
1592 	printk(KERN_ALERT
1593 	       "md/raid10:%s: Disk failure on %s, disabling device.\n"
1594 	       "md/raid10:%s: Operation continuing on %d devices.\n",
1595 	       mdname(mddev), bdevname(rdev->bdev, b),
1596 	       mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1597 }
1598 
1599 static void print_conf(struct r10conf *conf)
1600 {
1601 	int i;
1602 	struct md_rdev *rdev;
1603 
1604 	printk(KERN_DEBUG "RAID10 conf printout:\n");
1605 	if (!conf) {
1606 		printk(KERN_DEBUG "(!conf)\n");
1607 		return;
1608 	}
1609 	printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1610 		conf->geo.raid_disks);
1611 
1612 	/* This is only called with ->reconfix_mutex held, so
1613 	 * rcu protection of rdev is not needed */
1614 	for (i = 0; i < conf->geo.raid_disks; i++) {
1615 		char b[BDEVNAME_SIZE];
1616 		rdev = conf->mirrors[i].rdev;
1617 		if (rdev)
1618 			printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1619 				i, !test_bit(In_sync, &rdev->flags),
1620 			        !test_bit(Faulty, &rdev->flags),
1621 				bdevname(rdev->bdev,b));
1622 	}
1623 }
1624 
1625 static void close_sync(struct r10conf *conf)
1626 {
1627 	wait_barrier(conf);
1628 	allow_barrier(conf);
1629 
1630 	mempool_destroy(conf->r10buf_pool);
1631 	conf->r10buf_pool = NULL;
1632 }
1633 
1634 static int raid10_spare_active(struct mddev *mddev)
1635 {
1636 	int i;
1637 	struct r10conf *conf = mddev->private;
1638 	struct raid10_info *tmp;
1639 	int count = 0;
1640 	unsigned long flags;
1641 
1642 	/*
1643 	 * Find all non-in_sync disks within the RAID10 configuration
1644 	 * and mark them in_sync
1645 	 */
1646 	for (i = 0; i < conf->geo.raid_disks; i++) {
1647 		tmp = conf->mirrors + i;
1648 		if (tmp->replacement
1649 		    && tmp->replacement->recovery_offset == MaxSector
1650 		    && !test_bit(Faulty, &tmp->replacement->flags)
1651 		    && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1652 			/* Replacement has just become active */
1653 			if (!tmp->rdev
1654 			    || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1655 				count++;
1656 			if (tmp->rdev) {
1657 				/* Replaced device not technically faulty,
1658 				 * but we need to be sure it gets removed
1659 				 * and never re-added.
1660 				 */
1661 				set_bit(Faulty, &tmp->rdev->flags);
1662 				sysfs_notify_dirent_safe(
1663 					tmp->rdev->sysfs_state);
1664 			}
1665 			sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1666 		} else if (tmp->rdev
1667 			   && tmp->rdev->recovery_offset == MaxSector
1668 			   && !test_bit(Faulty, &tmp->rdev->flags)
1669 			   && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1670 			count++;
1671 			sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1672 		}
1673 	}
1674 	spin_lock_irqsave(&conf->device_lock, flags);
1675 	mddev->degraded -= count;
1676 	spin_unlock_irqrestore(&conf->device_lock, flags);
1677 
1678 	print_conf(conf);
1679 	return count;
1680 }
1681 
1682 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1683 {
1684 	struct r10conf *conf = mddev->private;
1685 	int err = -EEXIST;
1686 	int mirror;
1687 	int first = 0;
1688 	int last = conf->geo.raid_disks - 1;
1689 
1690 	if (mddev->recovery_cp < MaxSector)
1691 		/* only hot-add to in-sync arrays, as recovery is
1692 		 * very different from resync
1693 		 */
1694 		return -EBUSY;
1695 	if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1696 		return -EINVAL;
1697 
1698 	if (md_integrity_add_rdev(rdev, mddev))
1699 		return -ENXIO;
1700 
1701 	if (rdev->raid_disk >= 0)
1702 		first = last = rdev->raid_disk;
1703 
1704 	if (rdev->saved_raid_disk >= first &&
1705 	    conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1706 		mirror = rdev->saved_raid_disk;
1707 	else
1708 		mirror = first;
1709 	for ( ; mirror <= last ; mirror++) {
1710 		struct raid10_info *p = &conf->mirrors[mirror];
1711 		if (p->recovery_disabled == mddev->recovery_disabled)
1712 			continue;
1713 		if (p->rdev) {
1714 			if (!test_bit(WantReplacement, &p->rdev->flags) ||
1715 			    p->replacement != NULL)
1716 				continue;
1717 			clear_bit(In_sync, &rdev->flags);
1718 			set_bit(Replacement, &rdev->flags);
1719 			rdev->raid_disk = mirror;
1720 			err = 0;
1721 			if (mddev->gendisk)
1722 				disk_stack_limits(mddev->gendisk, rdev->bdev,
1723 						  rdev->data_offset << 9);
1724 			conf->fullsync = 1;
1725 			rcu_assign_pointer(p->replacement, rdev);
1726 			break;
1727 		}
1728 
1729 		if (mddev->gendisk)
1730 			disk_stack_limits(mddev->gendisk, rdev->bdev,
1731 					  rdev->data_offset << 9);
1732 
1733 		p->head_position = 0;
1734 		p->recovery_disabled = mddev->recovery_disabled - 1;
1735 		rdev->raid_disk = mirror;
1736 		err = 0;
1737 		if (rdev->saved_raid_disk != mirror)
1738 			conf->fullsync = 1;
1739 		rcu_assign_pointer(p->rdev, rdev);
1740 		break;
1741 	}
1742 	if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1743 		queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1744 
1745 	print_conf(conf);
1746 	return err;
1747 }
1748 
1749 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1750 {
1751 	struct r10conf *conf = mddev->private;
1752 	int err = 0;
1753 	int number = rdev->raid_disk;
1754 	struct md_rdev **rdevp;
1755 	struct raid10_info *p = conf->mirrors + number;
1756 
1757 	print_conf(conf);
1758 	if (rdev == p->rdev)
1759 		rdevp = &p->rdev;
1760 	else if (rdev == p->replacement)
1761 		rdevp = &p->replacement;
1762 	else
1763 		return 0;
1764 
1765 	if (test_bit(In_sync, &rdev->flags) ||
1766 	    atomic_read(&rdev->nr_pending)) {
1767 		err = -EBUSY;
1768 		goto abort;
1769 	}
1770 	/* Only remove non-faulty devices if recovery
1771 	 * is not possible.
1772 	 */
1773 	if (!test_bit(Faulty, &rdev->flags) &&
1774 	    mddev->recovery_disabled != p->recovery_disabled &&
1775 	    (!p->replacement || p->replacement == rdev) &&
1776 	    number < conf->geo.raid_disks &&
1777 	    enough(conf, -1)) {
1778 		err = -EBUSY;
1779 		goto abort;
1780 	}
1781 	*rdevp = NULL;
1782 	if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1783 		synchronize_rcu();
1784 		if (atomic_read(&rdev->nr_pending)) {
1785 			/* lost the race, try later */
1786 			err = -EBUSY;
1787 			*rdevp = rdev;
1788 			goto abort;
1789 		}
1790 	}
1791 	if (p->replacement) {
1792 		/* We must have just cleared 'rdev' */
1793 		p->rdev = p->replacement;
1794 		clear_bit(Replacement, &p->replacement->flags);
1795 		smp_mb(); /* Make sure other CPUs may see both as identical
1796 			   * but will never see neither -- if they are careful.
1797 			   */
1798 		p->replacement = NULL;
1799 		clear_bit(WantReplacement, &rdev->flags);
1800 	} else
1801 		/* We might have just remove the Replacement as faulty
1802 		 * Clear the flag just in case
1803 		 */
1804 		clear_bit(WantReplacement, &rdev->flags);
1805 
1806 	err = md_integrity_register(mddev);
1807 
1808 abort:
1809 
1810 	print_conf(conf);
1811 	return err;
1812 }
1813 
1814 static void end_sync_read(struct bio *bio)
1815 {
1816 	struct r10bio *r10_bio = bio->bi_private;
1817 	struct r10conf *conf = r10_bio->mddev->private;
1818 	int d;
1819 
1820 	if (bio == r10_bio->master_bio) {
1821 		/* this is a reshape read */
1822 		d = r10_bio->read_slot; /* really the read dev */
1823 	} else
1824 		d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1825 
1826 	if (!bio->bi_error)
1827 		set_bit(R10BIO_Uptodate, &r10_bio->state);
1828 	else
1829 		/* The write handler will notice the lack of
1830 		 * R10BIO_Uptodate and record any errors etc
1831 		 */
1832 		atomic_add(r10_bio->sectors,
1833 			   &conf->mirrors[d].rdev->corrected_errors);
1834 
1835 	/* for reconstruct, we always reschedule after a read.
1836 	 * for resync, only after all reads
1837 	 */
1838 	rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1839 	if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1840 	    atomic_dec_and_test(&r10_bio->remaining)) {
1841 		/* we have read all the blocks,
1842 		 * do the comparison in process context in raid10d
1843 		 */
1844 		reschedule_retry(r10_bio);
1845 	}
1846 }
1847 
1848 static void end_sync_request(struct r10bio *r10_bio)
1849 {
1850 	struct mddev *mddev = r10_bio->mddev;
1851 
1852 	while (atomic_dec_and_test(&r10_bio->remaining)) {
1853 		if (r10_bio->master_bio == NULL) {
1854 			/* the primary of several recovery bios */
1855 			sector_t s = r10_bio->sectors;
1856 			if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1857 			    test_bit(R10BIO_WriteError, &r10_bio->state))
1858 				reschedule_retry(r10_bio);
1859 			else
1860 				put_buf(r10_bio);
1861 			md_done_sync(mddev, s, 1);
1862 			break;
1863 		} else {
1864 			struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1865 			if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1866 			    test_bit(R10BIO_WriteError, &r10_bio->state))
1867 				reschedule_retry(r10_bio);
1868 			else
1869 				put_buf(r10_bio);
1870 			r10_bio = r10_bio2;
1871 		}
1872 	}
1873 }
1874 
1875 static void end_sync_write(struct bio *bio)
1876 {
1877 	struct r10bio *r10_bio = bio->bi_private;
1878 	struct mddev *mddev = r10_bio->mddev;
1879 	struct r10conf *conf = mddev->private;
1880 	int d;
1881 	sector_t first_bad;
1882 	int bad_sectors;
1883 	int slot;
1884 	int repl;
1885 	struct md_rdev *rdev = NULL;
1886 
1887 	d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1888 	if (repl)
1889 		rdev = conf->mirrors[d].replacement;
1890 	else
1891 		rdev = conf->mirrors[d].rdev;
1892 
1893 	if (bio->bi_error) {
1894 		if (repl)
1895 			md_error(mddev, rdev);
1896 		else {
1897 			set_bit(WriteErrorSeen, &rdev->flags);
1898 			if (!test_and_set_bit(WantReplacement, &rdev->flags))
1899 				set_bit(MD_RECOVERY_NEEDED,
1900 					&rdev->mddev->recovery);
1901 			set_bit(R10BIO_WriteError, &r10_bio->state);
1902 		}
1903 	} else if (is_badblock(rdev,
1904 			     r10_bio->devs[slot].addr,
1905 			     r10_bio->sectors,
1906 			     &first_bad, &bad_sectors))
1907 		set_bit(R10BIO_MadeGood, &r10_bio->state);
1908 
1909 	rdev_dec_pending(rdev, mddev);
1910 
1911 	end_sync_request(r10_bio);
1912 }
1913 
1914 /*
1915  * Note: sync and recover and handled very differently for raid10
1916  * This code is for resync.
1917  * For resync, we read through virtual addresses and read all blocks.
1918  * If there is any error, we schedule a write.  The lowest numbered
1919  * drive is authoritative.
1920  * However requests come for physical address, so we need to map.
1921  * For every physical address there are raid_disks/copies virtual addresses,
1922  * which is always are least one, but is not necessarly an integer.
1923  * This means that a physical address can span multiple chunks, so we may
1924  * have to submit multiple io requests for a single sync request.
1925  */
1926 /*
1927  * We check if all blocks are in-sync and only write to blocks that
1928  * aren't in sync
1929  */
1930 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1931 {
1932 	struct r10conf *conf = mddev->private;
1933 	int i, first;
1934 	struct bio *tbio, *fbio;
1935 	int vcnt;
1936 
1937 	atomic_set(&r10_bio->remaining, 1);
1938 
1939 	/* find the first device with a block */
1940 	for (i=0; i<conf->copies; i++)
1941 		if (!r10_bio->devs[i].bio->bi_error)
1942 			break;
1943 
1944 	if (i == conf->copies)
1945 		goto done;
1946 
1947 	first = i;
1948 	fbio = r10_bio->devs[i].bio;
1949 	fbio->bi_iter.bi_size = r10_bio->sectors << 9;
1950 	fbio->bi_iter.bi_idx = 0;
1951 
1952 	vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1953 	/* now find blocks with errors */
1954 	for (i=0 ; i < conf->copies ; i++) {
1955 		int  j, d;
1956 
1957 		tbio = r10_bio->devs[i].bio;
1958 
1959 		if (tbio->bi_end_io != end_sync_read)
1960 			continue;
1961 		if (i == first)
1962 			continue;
1963 		if (!r10_bio->devs[i].bio->bi_error) {
1964 			/* We know that the bi_io_vec layout is the same for
1965 			 * both 'first' and 'i', so we just compare them.
1966 			 * All vec entries are PAGE_SIZE;
1967 			 */
1968 			int sectors = r10_bio->sectors;
1969 			for (j = 0; j < vcnt; j++) {
1970 				int len = PAGE_SIZE;
1971 				if (sectors < (len / 512))
1972 					len = sectors * 512;
1973 				if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1974 					   page_address(tbio->bi_io_vec[j].bv_page),
1975 					   len))
1976 					break;
1977 				sectors -= len/512;
1978 			}
1979 			if (j == vcnt)
1980 				continue;
1981 			atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
1982 			if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1983 				/* Don't fix anything. */
1984 				continue;
1985 		}
1986 		/* Ok, we need to write this bio, either to correct an
1987 		 * inconsistency or to correct an unreadable block.
1988 		 * First we need to fixup bv_offset, bv_len and
1989 		 * bi_vecs, as the read request might have corrupted these
1990 		 */
1991 		bio_reset(tbio);
1992 
1993 		tbio->bi_vcnt = vcnt;
1994 		tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
1995 		tbio->bi_private = r10_bio;
1996 		tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
1997 		tbio->bi_end_io = end_sync_write;
1998 		bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
1999 
2000 		bio_copy_data(tbio, fbio);
2001 
2002 		d = r10_bio->devs[i].devnum;
2003 		atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2004 		atomic_inc(&r10_bio->remaining);
2005 		md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2006 
2007 		tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2008 		tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2009 		generic_make_request(tbio);
2010 	}
2011 
2012 	/* Now write out to any replacement devices
2013 	 * that are active
2014 	 */
2015 	for (i = 0; i < conf->copies; i++) {
2016 		int d;
2017 
2018 		tbio = r10_bio->devs[i].repl_bio;
2019 		if (!tbio || !tbio->bi_end_io)
2020 			continue;
2021 		if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2022 		    && r10_bio->devs[i].bio != fbio)
2023 			bio_copy_data(tbio, fbio);
2024 		d = r10_bio->devs[i].devnum;
2025 		atomic_inc(&r10_bio->remaining);
2026 		md_sync_acct(conf->mirrors[d].replacement->bdev,
2027 			     bio_sectors(tbio));
2028 		generic_make_request(tbio);
2029 	}
2030 
2031 done:
2032 	if (atomic_dec_and_test(&r10_bio->remaining)) {
2033 		md_done_sync(mddev, r10_bio->sectors, 1);
2034 		put_buf(r10_bio);
2035 	}
2036 }
2037 
2038 /*
2039  * Now for the recovery code.
2040  * Recovery happens across physical sectors.
2041  * We recover all non-is_sync drives by finding the virtual address of
2042  * each, and then choose a working drive that also has that virt address.
2043  * There is a separate r10_bio for each non-in_sync drive.
2044  * Only the first two slots are in use. The first for reading,
2045  * The second for writing.
2046  *
2047  */
2048 static void fix_recovery_read_error(struct r10bio *r10_bio)
2049 {
2050 	/* We got a read error during recovery.
2051 	 * We repeat the read in smaller page-sized sections.
2052 	 * If a read succeeds, write it to the new device or record
2053 	 * a bad block if we cannot.
2054 	 * If a read fails, record a bad block on both old and
2055 	 * new devices.
2056 	 */
2057 	struct mddev *mddev = r10_bio->mddev;
2058 	struct r10conf *conf = mddev->private;
2059 	struct bio *bio = r10_bio->devs[0].bio;
2060 	sector_t sect = 0;
2061 	int sectors = r10_bio->sectors;
2062 	int idx = 0;
2063 	int dr = r10_bio->devs[0].devnum;
2064 	int dw = r10_bio->devs[1].devnum;
2065 
2066 	while (sectors) {
2067 		int s = sectors;
2068 		struct md_rdev *rdev;
2069 		sector_t addr;
2070 		int ok;
2071 
2072 		if (s > (PAGE_SIZE>>9))
2073 			s = PAGE_SIZE >> 9;
2074 
2075 		rdev = conf->mirrors[dr].rdev;
2076 		addr = r10_bio->devs[0].addr + sect,
2077 		ok = sync_page_io(rdev,
2078 				  addr,
2079 				  s << 9,
2080 				  bio->bi_io_vec[idx].bv_page,
2081 				  REQ_OP_READ, 0, false);
2082 		if (ok) {
2083 			rdev = conf->mirrors[dw].rdev;
2084 			addr = r10_bio->devs[1].addr + sect;
2085 			ok = sync_page_io(rdev,
2086 					  addr,
2087 					  s << 9,
2088 					  bio->bi_io_vec[idx].bv_page,
2089 					  REQ_OP_WRITE, 0, false);
2090 			if (!ok) {
2091 				set_bit(WriteErrorSeen, &rdev->flags);
2092 				if (!test_and_set_bit(WantReplacement,
2093 						      &rdev->flags))
2094 					set_bit(MD_RECOVERY_NEEDED,
2095 						&rdev->mddev->recovery);
2096 			}
2097 		}
2098 		if (!ok) {
2099 			/* We don't worry if we cannot set a bad block -
2100 			 * it really is bad so there is no loss in not
2101 			 * recording it yet
2102 			 */
2103 			rdev_set_badblocks(rdev, addr, s, 0);
2104 
2105 			if (rdev != conf->mirrors[dw].rdev) {
2106 				/* need bad block on destination too */
2107 				struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2108 				addr = r10_bio->devs[1].addr + sect;
2109 				ok = rdev_set_badblocks(rdev2, addr, s, 0);
2110 				if (!ok) {
2111 					/* just abort the recovery */
2112 					printk(KERN_NOTICE
2113 					       "md/raid10:%s: recovery aborted"
2114 					       " due to read error\n",
2115 					       mdname(mddev));
2116 
2117 					conf->mirrors[dw].recovery_disabled
2118 						= mddev->recovery_disabled;
2119 					set_bit(MD_RECOVERY_INTR,
2120 						&mddev->recovery);
2121 					break;
2122 				}
2123 			}
2124 		}
2125 
2126 		sectors -= s;
2127 		sect += s;
2128 		idx++;
2129 	}
2130 }
2131 
2132 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2133 {
2134 	struct r10conf *conf = mddev->private;
2135 	int d;
2136 	struct bio *wbio, *wbio2;
2137 
2138 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2139 		fix_recovery_read_error(r10_bio);
2140 		end_sync_request(r10_bio);
2141 		return;
2142 	}
2143 
2144 	/*
2145 	 * share the pages with the first bio
2146 	 * and submit the write request
2147 	 */
2148 	d = r10_bio->devs[1].devnum;
2149 	wbio = r10_bio->devs[1].bio;
2150 	wbio2 = r10_bio->devs[1].repl_bio;
2151 	/* Need to test wbio2->bi_end_io before we call
2152 	 * generic_make_request as if the former is NULL,
2153 	 * the latter is free to free wbio2.
2154 	 */
2155 	if (wbio2 && !wbio2->bi_end_io)
2156 		wbio2 = NULL;
2157 	if (wbio->bi_end_io) {
2158 		atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2159 		md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2160 		generic_make_request(wbio);
2161 	}
2162 	if (wbio2) {
2163 		atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2164 		md_sync_acct(conf->mirrors[d].replacement->bdev,
2165 			     bio_sectors(wbio2));
2166 		generic_make_request(wbio2);
2167 	}
2168 }
2169 
2170 /*
2171  * Used by fix_read_error() to decay the per rdev read_errors.
2172  * We halve the read error count for every hour that has elapsed
2173  * since the last recorded read error.
2174  *
2175  */
2176 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2177 {
2178 	long cur_time_mon;
2179 	unsigned long hours_since_last;
2180 	unsigned int read_errors = atomic_read(&rdev->read_errors);
2181 
2182 	cur_time_mon = ktime_get_seconds();
2183 
2184 	if (rdev->last_read_error == 0) {
2185 		/* first time we've seen a read error */
2186 		rdev->last_read_error = cur_time_mon;
2187 		return;
2188 	}
2189 
2190 	hours_since_last = (long)(cur_time_mon -
2191 			    rdev->last_read_error) / 3600;
2192 
2193 	rdev->last_read_error = cur_time_mon;
2194 
2195 	/*
2196 	 * if hours_since_last is > the number of bits in read_errors
2197 	 * just set read errors to 0. We do this to avoid
2198 	 * overflowing the shift of read_errors by hours_since_last.
2199 	 */
2200 	if (hours_since_last >= 8 * sizeof(read_errors))
2201 		atomic_set(&rdev->read_errors, 0);
2202 	else
2203 		atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2204 }
2205 
2206 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2207 			    int sectors, struct page *page, int rw)
2208 {
2209 	sector_t first_bad;
2210 	int bad_sectors;
2211 
2212 	if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2213 	    && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2214 		return -1;
2215 	if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2216 		/* success */
2217 		return 1;
2218 	if (rw == WRITE) {
2219 		set_bit(WriteErrorSeen, &rdev->flags);
2220 		if (!test_and_set_bit(WantReplacement, &rdev->flags))
2221 			set_bit(MD_RECOVERY_NEEDED,
2222 				&rdev->mddev->recovery);
2223 	}
2224 	/* need to record an error - either for the block or the device */
2225 	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2226 		md_error(rdev->mddev, rdev);
2227 	return 0;
2228 }
2229 
2230 /*
2231  * This is a kernel thread which:
2232  *
2233  *	1.	Retries failed read operations on working mirrors.
2234  *	2.	Updates the raid superblock when problems encounter.
2235  *	3.	Performs writes following reads for array synchronising.
2236  */
2237 
2238 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2239 {
2240 	int sect = 0; /* Offset from r10_bio->sector */
2241 	int sectors = r10_bio->sectors;
2242 	struct md_rdev*rdev;
2243 	int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2244 	int d = r10_bio->devs[r10_bio->read_slot].devnum;
2245 
2246 	/* still own a reference to this rdev, so it cannot
2247 	 * have been cleared recently.
2248 	 */
2249 	rdev = conf->mirrors[d].rdev;
2250 
2251 	if (test_bit(Faulty, &rdev->flags))
2252 		/* drive has already been failed, just ignore any
2253 		   more fix_read_error() attempts */
2254 		return;
2255 
2256 	check_decay_read_errors(mddev, rdev);
2257 	atomic_inc(&rdev->read_errors);
2258 	if (atomic_read(&rdev->read_errors) > max_read_errors) {
2259 		char b[BDEVNAME_SIZE];
2260 		bdevname(rdev->bdev, b);
2261 
2262 		printk(KERN_NOTICE
2263 		       "md/raid10:%s: %s: Raid device exceeded "
2264 		       "read_error threshold [cur %d:max %d]\n",
2265 		       mdname(mddev), b,
2266 		       atomic_read(&rdev->read_errors), max_read_errors);
2267 		printk(KERN_NOTICE
2268 		       "md/raid10:%s: %s: Failing raid device\n",
2269 		       mdname(mddev), b);
2270 		md_error(mddev, rdev);
2271 		r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2272 		return;
2273 	}
2274 
2275 	while(sectors) {
2276 		int s = sectors;
2277 		int sl = r10_bio->read_slot;
2278 		int success = 0;
2279 		int start;
2280 
2281 		if (s > (PAGE_SIZE>>9))
2282 			s = PAGE_SIZE >> 9;
2283 
2284 		rcu_read_lock();
2285 		do {
2286 			sector_t first_bad;
2287 			int bad_sectors;
2288 
2289 			d = r10_bio->devs[sl].devnum;
2290 			rdev = rcu_dereference(conf->mirrors[d].rdev);
2291 			if (rdev &&
2292 			    test_bit(In_sync, &rdev->flags) &&
2293 			    !test_bit(Faulty, &rdev->flags) &&
2294 			    is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2295 					&first_bad, &bad_sectors) == 0) {
2296 				atomic_inc(&rdev->nr_pending);
2297 				rcu_read_unlock();
2298 				success = sync_page_io(rdev,
2299 						       r10_bio->devs[sl].addr +
2300 						       sect,
2301 						       s<<9,
2302 						       conf->tmppage,
2303 						       REQ_OP_READ, 0, false);
2304 				rdev_dec_pending(rdev, mddev);
2305 				rcu_read_lock();
2306 				if (success)
2307 					break;
2308 			}
2309 			sl++;
2310 			if (sl == conf->copies)
2311 				sl = 0;
2312 		} while (!success && sl != r10_bio->read_slot);
2313 		rcu_read_unlock();
2314 
2315 		if (!success) {
2316 			/* Cannot read from anywhere, just mark the block
2317 			 * as bad on the first device to discourage future
2318 			 * reads.
2319 			 */
2320 			int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2321 			rdev = conf->mirrors[dn].rdev;
2322 
2323 			if (!rdev_set_badblocks(
2324 				    rdev,
2325 				    r10_bio->devs[r10_bio->read_slot].addr
2326 				    + sect,
2327 				    s, 0)) {
2328 				md_error(mddev, rdev);
2329 				r10_bio->devs[r10_bio->read_slot].bio
2330 					= IO_BLOCKED;
2331 			}
2332 			break;
2333 		}
2334 
2335 		start = sl;
2336 		/* write it back and re-read */
2337 		rcu_read_lock();
2338 		while (sl != r10_bio->read_slot) {
2339 			char b[BDEVNAME_SIZE];
2340 
2341 			if (sl==0)
2342 				sl = conf->copies;
2343 			sl--;
2344 			d = r10_bio->devs[sl].devnum;
2345 			rdev = rcu_dereference(conf->mirrors[d].rdev);
2346 			if (!rdev ||
2347 			    test_bit(Faulty, &rdev->flags) ||
2348 			    !test_bit(In_sync, &rdev->flags))
2349 				continue;
2350 
2351 			atomic_inc(&rdev->nr_pending);
2352 			rcu_read_unlock();
2353 			if (r10_sync_page_io(rdev,
2354 					     r10_bio->devs[sl].addr +
2355 					     sect,
2356 					     s, conf->tmppage, WRITE)
2357 			    == 0) {
2358 				/* Well, this device is dead */
2359 				printk(KERN_NOTICE
2360 				       "md/raid10:%s: read correction "
2361 				       "write failed"
2362 				       " (%d sectors at %llu on %s)\n",
2363 				       mdname(mddev), s,
2364 				       (unsigned long long)(
2365 					       sect +
2366 					       choose_data_offset(r10_bio,
2367 								  rdev)),
2368 				       bdevname(rdev->bdev, b));
2369 				printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2370 				       "drive\n",
2371 				       mdname(mddev),
2372 				       bdevname(rdev->bdev, b));
2373 			}
2374 			rdev_dec_pending(rdev, mddev);
2375 			rcu_read_lock();
2376 		}
2377 		sl = start;
2378 		while (sl != r10_bio->read_slot) {
2379 			char b[BDEVNAME_SIZE];
2380 
2381 			if (sl==0)
2382 				sl = conf->copies;
2383 			sl--;
2384 			d = r10_bio->devs[sl].devnum;
2385 			rdev = rcu_dereference(conf->mirrors[d].rdev);
2386 			if (!rdev ||
2387 			    test_bit(Faulty, &rdev->flags) ||
2388 			    !test_bit(In_sync, &rdev->flags))
2389 				continue;
2390 
2391 			atomic_inc(&rdev->nr_pending);
2392 			rcu_read_unlock();
2393 			switch (r10_sync_page_io(rdev,
2394 					     r10_bio->devs[sl].addr +
2395 					     sect,
2396 					     s, conf->tmppage,
2397 						 READ)) {
2398 			case 0:
2399 				/* Well, this device is dead */
2400 				printk(KERN_NOTICE
2401 				       "md/raid10:%s: unable to read back "
2402 				       "corrected sectors"
2403 				       " (%d sectors at %llu on %s)\n",
2404 				       mdname(mddev), s,
2405 				       (unsigned long long)(
2406 					       sect +
2407 					       choose_data_offset(r10_bio, rdev)),
2408 				       bdevname(rdev->bdev, b));
2409 				printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2410 				       "drive\n",
2411 				       mdname(mddev),
2412 				       bdevname(rdev->bdev, b));
2413 				break;
2414 			case 1:
2415 				printk(KERN_INFO
2416 				       "md/raid10:%s: read error corrected"
2417 				       " (%d sectors at %llu on %s)\n",
2418 				       mdname(mddev), s,
2419 				       (unsigned long long)(
2420 					       sect +
2421 					       choose_data_offset(r10_bio, rdev)),
2422 				       bdevname(rdev->bdev, b));
2423 				atomic_add(s, &rdev->corrected_errors);
2424 			}
2425 
2426 			rdev_dec_pending(rdev, mddev);
2427 			rcu_read_lock();
2428 		}
2429 		rcu_read_unlock();
2430 
2431 		sectors -= s;
2432 		sect += s;
2433 	}
2434 }
2435 
2436 static int narrow_write_error(struct r10bio *r10_bio, int i)
2437 {
2438 	struct bio *bio = r10_bio->master_bio;
2439 	struct mddev *mddev = r10_bio->mddev;
2440 	struct r10conf *conf = mddev->private;
2441 	struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2442 	/* bio has the data to be written to slot 'i' where
2443 	 * we just recently had a write error.
2444 	 * We repeatedly clone the bio and trim down to one block,
2445 	 * then try the write.  Where the write fails we record
2446 	 * a bad block.
2447 	 * It is conceivable that the bio doesn't exactly align with
2448 	 * blocks.  We must handle this.
2449 	 *
2450 	 * We currently own a reference to the rdev.
2451 	 */
2452 
2453 	int block_sectors;
2454 	sector_t sector;
2455 	int sectors;
2456 	int sect_to_write = r10_bio->sectors;
2457 	int ok = 1;
2458 
2459 	if (rdev->badblocks.shift < 0)
2460 		return 0;
2461 
2462 	block_sectors = roundup(1 << rdev->badblocks.shift,
2463 				bdev_logical_block_size(rdev->bdev) >> 9);
2464 	sector = r10_bio->sector;
2465 	sectors = ((r10_bio->sector + block_sectors)
2466 		   & ~(sector_t)(block_sectors - 1))
2467 		- sector;
2468 
2469 	while (sect_to_write) {
2470 		struct bio *wbio;
2471 		sector_t wsector;
2472 		if (sectors > sect_to_write)
2473 			sectors = sect_to_write;
2474 		/* Write at 'sector' for 'sectors' */
2475 		wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2476 		bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2477 		wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2478 		wbio->bi_iter.bi_sector = wsector +
2479 				   choose_data_offset(r10_bio, rdev);
2480 		wbio->bi_bdev = rdev->bdev;
2481 		bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2482 
2483 		if (submit_bio_wait(wbio) < 0)
2484 			/* Failure! */
2485 			ok = rdev_set_badblocks(rdev, wsector,
2486 						sectors, 0)
2487 				&& ok;
2488 
2489 		bio_put(wbio);
2490 		sect_to_write -= sectors;
2491 		sector += sectors;
2492 		sectors = block_sectors;
2493 	}
2494 	return ok;
2495 }
2496 
2497 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2498 {
2499 	int slot = r10_bio->read_slot;
2500 	struct bio *bio;
2501 	struct r10conf *conf = mddev->private;
2502 	struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2503 	char b[BDEVNAME_SIZE];
2504 	unsigned long do_sync;
2505 	int max_sectors;
2506 
2507 	/* we got a read error. Maybe the drive is bad.  Maybe just
2508 	 * the block and we can fix it.
2509 	 * We freeze all other IO, and try reading the block from
2510 	 * other devices.  When we find one, we re-write
2511 	 * and check it that fixes the read error.
2512 	 * This is all done synchronously while the array is
2513 	 * frozen.
2514 	 */
2515 	bio = r10_bio->devs[slot].bio;
2516 	bdevname(bio->bi_bdev, b);
2517 	bio_put(bio);
2518 	r10_bio->devs[slot].bio = NULL;
2519 
2520 	if (mddev->ro == 0) {
2521 		freeze_array(conf, 1);
2522 		fix_read_error(conf, mddev, r10_bio);
2523 		unfreeze_array(conf);
2524 	} else
2525 		r10_bio->devs[slot].bio = IO_BLOCKED;
2526 
2527 	rdev_dec_pending(rdev, mddev);
2528 
2529 read_more:
2530 	rdev = read_balance(conf, r10_bio, &max_sectors);
2531 	if (rdev == NULL) {
2532 		printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2533 		       " read error for block %llu\n",
2534 		       mdname(mddev), b,
2535 		       (unsigned long long)r10_bio->sector);
2536 		raid_end_bio_io(r10_bio);
2537 		return;
2538 	}
2539 
2540 	do_sync = (r10_bio->master_bio->bi_opf & REQ_SYNC);
2541 	slot = r10_bio->read_slot;
2542 	printk_ratelimited(
2543 		KERN_ERR
2544 		"md/raid10:%s: %s: redirecting "
2545 		"sector %llu to another mirror\n",
2546 		mdname(mddev),
2547 		bdevname(rdev->bdev, b),
2548 		(unsigned long long)r10_bio->sector);
2549 	bio = bio_clone_mddev(r10_bio->master_bio,
2550 			      GFP_NOIO, mddev);
2551 	bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2552 	r10_bio->devs[slot].bio = bio;
2553 	r10_bio->devs[slot].rdev = rdev;
2554 	bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2555 		+ choose_data_offset(r10_bio, rdev);
2556 	bio->bi_bdev = rdev->bdev;
2557 	bio_set_op_attrs(bio, REQ_OP_READ, do_sync);
2558 	bio->bi_private = r10_bio;
2559 	bio->bi_end_io = raid10_end_read_request;
2560 	if (max_sectors < r10_bio->sectors) {
2561 		/* Drat - have to split this up more */
2562 		struct bio *mbio = r10_bio->master_bio;
2563 		int sectors_handled =
2564 			r10_bio->sector + max_sectors
2565 			- mbio->bi_iter.bi_sector;
2566 		r10_bio->sectors = max_sectors;
2567 		spin_lock_irq(&conf->device_lock);
2568 		if (mbio->bi_phys_segments == 0)
2569 			mbio->bi_phys_segments = 2;
2570 		else
2571 			mbio->bi_phys_segments++;
2572 		spin_unlock_irq(&conf->device_lock);
2573 		generic_make_request(bio);
2574 
2575 		r10_bio = mempool_alloc(conf->r10bio_pool,
2576 					GFP_NOIO);
2577 		r10_bio->master_bio = mbio;
2578 		r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2579 		r10_bio->state = 0;
2580 		set_bit(R10BIO_ReadError,
2581 			&r10_bio->state);
2582 		r10_bio->mddev = mddev;
2583 		r10_bio->sector = mbio->bi_iter.bi_sector
2584 			+ sectors_handled;
2585 
2586 		goto read_more;
2587 	} else
2588 		generic_make_request(bio);
2589 }
2590 
2591 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2592 {
2593 	/* Some sort of write request has finished and it
2594 	 * succeeded in writing where we thought there was a
2595 	 * bad block.  So forget the bad block.
2596 	 * Or possibly if failed and we need to record
2597 	 * a bad block.
2598 	 */
2599 	int m;
2600 	struct md_rdev *rdev;
2601 
2602 	if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2603 	    test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2604 		for (m = 0; m < conf->copies; m++) {
2605 			int dev = r10_bio->devs[m].devnum;
2606 			rdev = conf->mirrors[dev].rdev;
2607 			if (r10_bio->devs[m].bio == NULL)
2608 				continue;
2609 			if (!r10_bio->devs[m].bio->bi_error) {
2610 				rdev_clear_badblocks(
2611 					rdev,
2612 					r10_bio->devs[m].addr,
2613 					r10_bio->sectors, 0);
2614 			} else {
2615 				if (!rdev_set_badblocks(
2616 					    rdev,
2617 					    r10_bio->devs[m].addr,
2618 					    r10_bio->sectors, 0))
2619 					md_error(conf->mddev, rdev);
2620 			}
2621 			rdev = conf->mirrors[dev].replacement;
2622 			if (r10_bio->devs[m].repl_bio == NULL)
2623 				continue;
2624 
2625 			if (!r10_bio->devs[m].repl_bio->bi_error) {
2626 				rdev_clear_badblocks(
2627 					rdev,
2628 					r10_bio->devs[m].addr,
2629 					r10_bio->sectors, 0);
2630 			} else {
2631 				if (!rdev_set_badblocks(
2632 					    rdev,
2633 					    r10_bio->devs[m].addr,
2634 					    r10_bio->sectors, 0))
2635 					md_error(conf->mddev, rdev);
2636 			}
2637 		}
2638 		put_buf(r10_bio);
2639 	} else {
2640 		bool fail = false;
2641 		for (m = 0; m < conf->copies; m++) {
2642 			int dev = r10_bio->devs[m].devnum;
2643 			struct bio *bio = r10_bio->devs[m].bio;
2644 			rdev = conf->mirrors[dev].rdev;
2645 			if (bio == IO_MADE_GOOD) {
2646 				rdev_clear_badblocks(
2647 					rdev,
2648 					r10_bio->devs[m].addr,
2649 					r10_bio->sectors, 0);
2650 				rdev_dec_pending(rdev, conf->mddev);
2651 			} else if (bio != NULL && bio->bi_error) {
2652 				fail = true;
2653 				if (!narrow_write_error(r10_bio, m)) {
2654 					md_error(conf->mddev, rdev);
2655 					set_bit(R10BIO_Degraded,
2656 						&r10_bio->state);
2657 				}
2658 				rdev_dec_pending(rdev, conf->mddev);
2659 			}
2660 			bio = r10_bio->devs[m].repl_bio;
2661 			rdev = conf->mirrors[dev].replacement;
2662 			if (rdev && bio == IO_MADE_GOOD) {
2663 				rdev_clear_badblocks(
2664 					rdev,
2665 					r10_bio->devs[m].addr,
2666 					r10_bio->sectors, 0);
2667 				rdev_dec_pending(rdev, conf->mddev);
2668 			}
2669 		}
2670 		if (fail) {
2671 			spin_lock_irq(&conf->device_lock);
2672 			list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2673 			conf->nr_queued++;
2674 			spin_unlock_irq(&conf->device_lock);
2675 			md_wakeup_thread(conf->mddev->thread);
2676 		} else {
2677 			if (test_bit(R10BIO_WriteError,
2678 				     &r10_bio->state))
2679 				close_write(r10_bio);
2680 			raid_end_bio_io(r10_bio);
2681 		}
2682 	}
2683 }
2684 
2685 static void raid10d(struct md_thread *thread)
2686 {
2687 	struct mddev *mddev = thread->mddev;
2688 	struct r10bio *r10_bio;
2689 	unsigned long flags;
2690 	struct r10conf *conf = mddev->private;
2691 	struct list_head *head = &conf->retry_list;
2692 	struct blk_plug plug;
2693 
2694 	md_check_recovery(mddev);
2695 
2696 	if (!list_empty_careful(&conf->bio_end_io_list) &&
2697 	    !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2698 		LIST_HEAD(tmp);
2699 		spin_lock_irqsave(&conf->device_lock, flags);
2700 		if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2701 			while (!list_empty(&conf->bio_end_io_list)) {
2702 				list_move(conf->bio_end_io_list.prev, &tmp);
2703 				conf->nr_queued--;
2704 			}
2705 		}
2706 		spin_unlock_irqrestore(&conf->device_lock, flags);
2707 		while (!list_empty(&tmp)) {
2708 			r10_bio = list_first_entry(&tmp, struct r10bio,
2709 						   retry_list);
2710 			list_del(&r10_bio->retry_list);
2711 			if (mddev->degraded)
2712 				set_bit(R10BIO_Degraded, &r10_bio->state);
2713 
2714 			if (test_bit(R10BIO_WriteError,
2715 				     &r10_bio->state))
2716 				close_write(r10_bio);
2717 			raid_end_bio_io(r10_bio);
2718 		}
2719 	}
2720 
2721 	blk_start_plug(&plug);
2722 	for (;;) {
2723 
2724 		flush_pending_writes(conf);
2725 
2726 		spin_lock_irqsave(&conf->device_lock, flags);
2727 		if (list_empty(head)) {
2728 			spin_unlock_irqrestore(&conf->device_lock, flags);
2729 			break;
2730 		}
2731 		r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2732 		list_del(head->prev);
2733 		conf->nr_queued--;
2734 		spin_unlock_irqrestore(&conf->device_lock, flags);
2735 
2736 		mddev = r10_bio->mddev;
2737 		conf = mddev->private;
2738 		if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2739 		    test_bit(R10BIO_WriteError, &r10_bio->state))
2740 			handle_write_completed(conf, r10_bio);
2741 		else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2742 			reshape_request_write(mddev, r10_bio);
2743 		else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2744 			sync_request_write(mddev, r10_bio);
2745 		else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2746 			recovery_request_write(mddev, r10_bio);
2747 		else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2748 			handle_read_error(mddev, r10_bio);
2749 		else {
2750 			/* just a partial read to be scheduled from a
2751 			 * separate context
2752 			 */
2753 			int slot = r10_bio->read_slot;
2754 			generic_make_request(r10_bio->devs[slot].bio);
2755 		}
2756 
2757 		cond_resched();
2758 		if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2759 			md_check_recovery(mddev);
2760 	}
2761 	blk_finish_plug(&plug);
2762 }
2763 
2764 static int init_resync(struct r10conf *conf)
2765 {
2766 	int buffs;
2767 	int i;
2768 
2769 	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2770 	BUG_ON(conf->r10buf_pool);
2771 	conf->have_replacement = 0;
2772 	for (i = 0; i < conf->geo.raid_disks; i++)
2773 		if (conf->mirrors[i].replacement)
2774 			conf->have_replacement = 1;
2775 	conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2776 	if (!conf->r10buf_pool)
2777 		return -ENOMEM;
2778 	conf->next_resync = 0;
2779 	return 0;
2780 }
2781 
2782 /*
2783  * perform a "sync" on one "block"
2784  *
2785  * We need to make sure that no normal I/O request - particularly write
2786  * requests - conflict with active sync requests.
2787  *
2788  * This is achieved by tracking pending requests and a 'barrier' concept
2789  * that can be installed to exclude normal IO requests.
2790  *
2791  * Resync and recovery are handled very differently.
2792  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2793  *
2794  * For resync, we iterate over virtual addresses, read all copies,
2795  * and update if there are differences.  If only one copy is live,
2796  * skip it.
2797  * For recovery, we iterate over physical addresses, read a good
2798  * value for each non-in_sync drive, and over-write.
2799  *
2800  * So, for recovery we may have several outstanding complex requests for a
2801  * given address, one for each out-of-sync device.  We model this by allocating
2802  * a number of r10_bio structures, one for each out-of-sync device.
2803  * As we setup these structures, we collect all bio's together into a list
2804  * which we then process collectively to add pages, and then process again
2805  * to pass to generic_make_request.
2806  *
2807  * The r10_bio structures are linked using a borrowed master_bio pointer.
2808  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2809  * has its remaining count decremented to 0, the whole complex operation
2810  * is complete.
2811  *
2812  */
2813 
2814 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2815 			     int *skipped)
2816 {
2817 	struct r10conf *conf = mddev->private;
2818 	struct r10bio *r10_bio;
2819 	struct bio *biolist = NULL, *bio;
2820 	sector_t max_sector, nr_sectors;
2821 	int i;
2822 	int max_sync;
2823 	sector_t sync_blocks;
2824 	sector_t sectors_skipped = 0;
2825 	int chunks_skipped = 0;
2826 	sector_t chunk_mask = conf->geo.chunk_mask;
2827 
2828 	if (!conf->r10buf_pool)
2829 		if (init_resync(conf))
2830 			return 0;
2831 
2832 	/*
2833 	 * Allow skipping a full rebuild for incremental assembly
2834 	 * of a clean array, like RAID1 does.
2835 	 */
2836 	if (mddev->bitmap == NULL &&
2837 	    mddev->recovery_cp == MaxSector &&
2838 	    mddev->reshape_position == MaxSector &&
2839 	    !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2840 	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2841 	    !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2842 	    conf->fullsync == 0) {
2843 		*skipped = 1;
2844 		return mddev->dev_sectors - sector_nr;
2845 	}
2846 
2847  skipped:
2848 	max_sector = mddev->dev_sectors;
2849 	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2850 	    test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2851 		max_sector = mddev->resync_max_sectors;
2852 	if (sector_nr >= max_sector) {
2853 		/* If we aborted, we need to abort the
2854 		 * sync on the 'current' bitmap chucks (there can
2855 		 * be several when recovering multiple devices).
2856 		 * as we may have started syncing it but not finished.
2857 		 * We can find the current address in
2858 		 * mddev->curr_resync, but for recovery,
2859 		 * we need to convert that to several
2860 		 * virtual addresses.
2861 		 */
2862 		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2863 			end_reshape(conf);
2864 			close_sync(conf);
2865 			return 0;
2866 		}
2867 
2868 		if (mddev->curr_resync < max_sector) { /* aborted */
2869 			if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2870 				bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2871 						&sync_blocks, 1);
2872 			else for (i = 0; i < conf->geo.raid_disks; i++) {
2873 				sector_t sect =
2874 					raid10_find_virt(conf, mddev->curr_resync, i);
2875 				bitmap_end_sync(mddev->bitmap, sect,
2876 						&sync_blocks, 1);
2877 			}
2878 		} else {
2879 			/* completed sync */
2880 			if ((!mddev->bitmap || conf->fullsync)
2881 			    && conf->have_replacement
2882 			    && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2883 				/* Completed a full sync so the replacements
2884 				 * are now fully recovered.
2885 				 */
2886 				rcu_read_lock();
2887 				for (i = 0; i < conf->geo.raid_disks; i++) {
2888 					struct md_rdev *rdev =
2889 						rcu_dereference(conf->mirrors[i].replacement);
2890 					if (rdev)
2891 						rdev->recovery_offset = MaxSector;
2892 				}
2893 				rcu_read_unlock();
2894 			}
2895 			conf->fullsync = 0;
2896 		}
2897 		bitmap_close_sync(mddev->bitmap);
2898 		close_sync(conf);
2899 		*skipped = 1;
2900 		return sectors_skipped;
2901 	}
2902 
2903 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2904 		return reshape_request(mddev, sector_nr, skipped);
2905 
2906 	if (chunks_skipped >= conf->geo.raid_disks) {
2907 		/* if there has been nothing to do on any drive,
2908 		 * then there is nothing to do at all..
2909 		 */
2910 		*skipped = 1;
2911 		return (max_sector - sector_nr) + sectors_skipped;
2912 	}
2913 
2914 	if (max_sector > mddev->resync_max)
2915 		max_sector = mddev->resync_max; /* Don't do IO beyond here */
2916 
2917 	/* make sure whole request will fit in a chunk - if chunks
2918 	 * are meaningful
2919 	 */
2920 	if (conf->geo.near_copies < conf->geo.raid_disks &&
2921 	    max_sector > (sector_nr | chunk_mask))
2922 		max_sector = (sector_nr | chunk_mask) + 1;
2923 
2924 	/*
2925 	 * If there is non-resync activity waiting for a turn, then let it
2926 	 * though before starting on this new sync request.
2927 	 */
2928 	if (conf->nr_waiting)
2929 		schedule_timeout_uninterruptible(1);
2930 
2931 	/* Again, very different code for resync and recovery.
2932 	 * Both must result in an r10bio with a list of bios that
2933 	 * have bi_end_io, bi_sector, bi_bdev set,
2934 	 * and bi_private set to the r10bio.
2935 	 * For recovery, we may actually create several r10bios
2936 	 * with 2 bios in each, that correspond to the bios in the main one.
2937 	 * In this case, the subordinate r10bios link back through a
2938 	 * borrowed master_bio pointer, and the counter in the master
2939 	 * includes a ref from each subordinate.
2940 	 */
2941 	/* First, we decide what to do and set ->bi_end_io
2942 	 * To end_sync_read if we want to read, and
2943 	 * end_sync_write if we will want to write.
2944 	 */
2945 
2946 	max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2947 	if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2948 		/* recovery... the complicated one */
2949 		int j;
2950 		r10_bio = NULL;
2951 
2952 		for (i = 0 ; i < conf->geo.raid_disks; i++) {
2953 			int still_degraded;
2954 			struct r10bio *rb2;
2955 			sector_t sect;
2956 			int must_sync;
2957 			int any_working;
2958 			struct raid10_info *mirror = &conf->mirrors[i];
2959 			struct md_rdev *mrdev, *mreplace;
2960 
2961 			rcu_read_lock();
2962 			mrdev = rcu_dereference(mirror->rdev);
2963 			mreplace = rcu_dereference(mirror->replacement);
2964 
2965 			if ((mrdev == NULL ||
2966 			     test_bit(Faulty, &mrdev->flags) ||
2967 			     test_bit(In_sync, &mrdev->flags)) &&
2968 			    (mreplace == NULL ||
2969 			     test_bit(Faulty, &mreplace->flags))) {
2970 				rcu_read_unlock();
2971 				continue;
2972 			}
2973 
2974 			still_degraded = 0;
2975 			/* want to reconstruct this device */
2976 			rb2 = r10_bio;
2977 			sect = raid10_find_virt(conf, sector_nr, i);
2978 			if (sect >= mddev->resync_max_sectors) {
2979 				/* last stripe is not complete - don't
2980 				 * try to recover this sector.
2981 				 */
2982 				rcu_read_unlock();
2983 				continue;
2984 			}
2985 			if (mreplace && test_bit(Faulty, &mreplace->flags))
2986 				mreplace = NULL;
2987 			/* Unless we are doing a full sync, or a replacement
2988 			 * we only need to recover the block if it is set in
2989 			 * the bitmap
2990 			 */
2991 			must_sync = bitmap_start_sync(mddev->bitmap, sect,
2992 						      &sync_blocks, 1);
2993 			if (sync_blocks < max_sync)
2994 				max_sync = sync_blocks;
2995 			if (!must_sync &&
2996 			    mreplace == NULL &&
2997 			    !conf->fullsync) {
2998 				/* yep, skip the sync_blocks here, but don't assume
2999 				 * that there will never be anything to do here
3000 				 */
3001 				chunks_skipped = -1;
3002 				rcu_read_unlock();
3003 				continue;
3004 			}
3005 			atomic_inc(&mrdev->nr_pending);
3006 			if (mreplace)
3007 				atomic_inc(&mreplace->nr_pending);
3008 			rcu_read_unlock();
3009 
3010 			r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3011 			r10_bio->state = 0;
3012 			raise_barrier(conf, rb2 != NULL);
3013 			atomic_set(&r10_bio->remaining, 0);
3014 
3015 			r10_bio->master_bio = (struct bio*)rb2;
3016 			if (rb2)
3017 				atomic_inc(&rb2->remaining);
3018 			r10_bio->mddev = mddev;
3019 			set_bit(R10BIO_IsRecover, &r10_bio->state);
3020 			r10_bio->sector = sect;
3021 
3022 			raid10_find_phys(conf, r10_bio);
3023 
3024 			/* Need to check if the array will still be
3025 			 * degraded
3026 			 */
3027 			rcu_read_lock();
3028 			for (j = 0; j < conf->geo.raid_disks; j++) {
3029 				struct md_rdev *rdev = rcu_dereference(
3030 					conf->mirrors[j].rdev);
3031 				if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3032 					still_degraded = 1;
3033 					break;
3034 				}
3035 			}
3036 
3037 			must_sync = bitmap_start_sync(mddev->bitmap, sect,
3038 						      &sync_blocks, still_degraded);
3039 
3040 			any_working = 0;
3041 			for (j=0; j<conf->copies;j++) {
3042 				int k;
3043 				int d = r10_bio->devs[j].devnum;
3044 				sector_t from_addr, to_addr;
3045 				struct md_rdev *rdev =
3046 					rcu_dereference(conf->mirrors[d].rdev);
3047 				sector_t sector, first_bad;
3048 				int bad_sectors;
3049 				if (!rdev ||
3050 				    !test_bit(In_sync, &rdev->flags))
3051 					continue;
3052 				/* This is where we read from */
3053 				any_working = 1;
3054 				sector = r10_bio->devs[j].addr;
3055 
3056 				if (is_badblock(rdev, sector, max_sync,
3057 						&first_bad, &bad_sectors)) {
3058 					if (first_bad > sector)
3059 						max_sync = first_bad - sector;
3060 					else {
3061 						bad_sectors -= (sector
3062 								- first_bad);
3063 						if (max_sync > bad_sectors)
3064 							max_sync = bad_sectors;
3065 						continue;
3066 					}
3067 				}
3068 				bio = r10_bio->devs[0].bio;
3069 				bio_reset(bio);
3070 				bio->bi_next = biolist;
3071 				biolist = bio;
3072 				bio->bi_private = r10_bio;
3073 				bio->bi_end_io = end_sync_read;
3074 				bio_set_op_attrs(bio, REQ_OP_READ, 0);
3075 				from_addr = r10_bio->devs[j].addr;
3076 				bio->bi_iter.bi_sector = from_addr +
3077 					rdev->data_offset;
3078 				bio->bi_bdev = rdev->bdev;
3079 				atomic_inc(&rdev->nr_pending);
3080 				/* and we write to 'i' (if not in_sync) */
3081 
3082 				for (k=0; k<conf->copies; k++)
3083 					if (r10_bio->devs[k].devnum == i)
3084 						break;
3085 				BUG_ON(k == conf->copies);
3086 				to_addr = r10_bio->devs[k].addr;
3087 				r10_bio->devs[0].devnum = d;
3088 				r10_bio->devs[0].addr = from_addr;
3089 				r10_bio->devs[1].devnum = i;
3090 				r10_bio->devs[1].addr = to_addr;
3091 
3092 				if (!test_bit(In_sync, &mrdev->flags)) {
3093 					bio = r10_bio->devs[1].bio;
3094 					bio_reset(bio);
3095 					bio->bi_next = biolist;
3096 					biolist = bio;
3097 					bio->bi_private = r10_bio;
3098 					bio->bi_end_io = end_sync_write;
3099 					bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3100 					bio->bi_iter.bi_sector = to_addr
3101 						+ mrdev->data_offset;
3102 					bio->bi_bdev = mrdev->bdev;
3103 					atomic_inc(&r10_bio->remaining);
3104 				} else
3105 					r10_bio->devs[1].bio->bi_end_io = NULL;
3106 
3107 				/* and maybe write to replacement */
3108 				bio = r10_bio->devs[1].repl_bio;
3109 				if (bio)
3110 					bio->bi_end_io = NULL;
3111 				/* Note: if mreplace != NULL, then bio
3112 				 * cannot be NULL as r10buf_pool_alloc will
3113 				 * have allocated it.
3114 				 * So the second test here is pointless.
3115 				 * But it keeps semantic-checkers happy, and
3116 				 * this comment keeps human reviewers
3117 				 * happy.
3118 				 */
3119 				if (mreplace == NULL || bio == NULL ||
3120 				    test_bit(Faulty, &mreplace->flags))
3121 					break;
3122 				bio_reset(bio);
3123 				bio->bi_next = biolist;
3124 				biolist = bio;
3125 				bio->bi_private = r10_bio;
3126 				bio->bi_end_io = end_sync_write;
3127 				bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3128 				bio->bi_iter.bi_sector = to_addr +
3129 					mreplace->data_offset;
3130 				bio->bi_bdev = mreplace->bdev;
3131 				atomic_inc(&r10_bio->remaining);
3132 				break;
3133 			}
3134 			rcu_read_unlock();
3135 			if (j == conf->copies) {
3136 				/* Cannot recover, so abort the recovery or
3137 				 * record a bad block */
3138 				if (any_working) {
3139 					/* problem is that there are bad blocks
3140 					 * on other device(s)
3141 					 */
3142 					int k;
3143 					for (k = 0; k < conf->copies; k++)
3144 						if (r10_bio->devs[k].devnum == i)
3145 							break;
3146 					if (!test_bit(In_sync,
3147 						      &mrdev->flags)
3148 					    && !rdev_set_badblocks(
3149 						    mrdev,
3150 						    r10_bio->devs[k].addr,
3151 						    max_sync, 0))
3152 						any_working = 0;
3153 					if (mreplace &&
3154 					    !rdev_set_badblocks(
3155 						    mreplace,
3156 						    r10_bio->devs[k].addr,
3157 						    max_sync, 0))
3158 						any_working = 0;
3159 				}
3160 				if (!any_working)  {
3161 					if (!test_and_set_bit(MD_RECOVERY_INTR,
3162 							      &mddev->recovery))
3163 						printk(KERN_INFO "md/raid10:%s: insufficient "
3164 						       "working devices for recovery.\n",
3165 						       mdname(mddev));
3166 					mirror->recovery_disabled
3167 						= mddev->recovery_disabled;
3168 				}
3169 				put_buf(r10_bio);
3170 				if (rb2)
3171 					atomic_dec(&rb2->remaining);
3172 				r10_bio = rb2;
3173 				rdev_dec_pending(mrdev, mddev);
3174 				if (mreplace)
3175 					rdev_dec_pending(mreplace, mddev);
3176 				break;
3177 			}
3178 			rdev_dec_pending(mrdev, mddev);
3179 			if (mreplace)
3180 				rdev_dec_pending(mreplace, mddev);
3181 		}
3182 		if (biolist == NULL) {
3183 			while (r10_bio) {
3184 				struct r10bio *rb2 = r10_bio;
3185 				r10_bio = (struct r10bio*) rb2->master_bio;
3186 				rb2->master_bio = NULL;
3187 				put_buf(rb2);
3188 			}
3189 			goto giveup;
3190 		}
3191 	} else {
3192 		/* resync. Schedule a read for every block at this virt offset */
3193 		int count = 0;
3194 
3195 		bitmap_cond_end_sync(mddev->bitmap, sector_nr, 0);
3196 
3197 		if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3198 				       &sync_blocks, mddev->degraded) &&
3199 		    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3200 						 &mddev->recovery)) {
3201 			/* We can skip this block */
3202 			*skipped = 1;
3203 			return sync_blocks + sectors_skipped;
3204 		}
3205 		if (sync_blocks < max_sync)
3206 			max_sync = sync_blocks;
3207 		r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3208 		r10_bio->state = 0;
3209 
3210 		r10_bio->mddev = mddev;
3211 		atomic_set(&r10_bio->remaining, 0);
3212 		raise_barrier(conf, 0);
3213 		conf->next_resync = sector_nr;
3214 
3215 		r10_bio->master_bio = NULL;
3216 		r10_bio->sector = sector_nr;
3217 		set_bit(R10BIO_IsSync, &r10_bio->state);
3218 		raid10_find_phys(conf, r10_bio);
3219 		r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3220 
3221 		for (i = 0; i < conf->copies; i++) {
3222 			int d = r10_bio->devs[i].devnum;
3223 			sector_t first_bad, sector;
3224 			int bad_sectors;
3225 			struct md_rdev *rdev;
3226 
3227 			if (r10_bio->devs[i].repl_bio)
3228 				r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3229 
3230 			bio = r10_bio->devs[i].bio;
3231 			bio_reset(bio);
3232 			bio->bi_error = -EIO;
3233 			rcu_read_lock();
3234 			rdev = rcu_dereference(conf->mirrors[d].rdev);
3235 			if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3236 				rcu_read_unlock();
3237 				continue;
3238 			}
3239 			sector = r10_bio->devs[i].addr;
3240 			if (is_badblock(rdev, sector, max_sync,
3241 					&first_bad, &bad_sectors)) {
3242 				if (first_bad > sector)
3243 					max_sync = first_bad - sector;
3244 				else {
3245 					bad_sectors -= (sector - first_bad);
3246 					if (max_sync > bad_sectors)
3247 						max_sync = bad_sectors;
3248 					rcu_read_unlock();
3249 					continue;
3250 				}
3251 			}
3252 			atomic_inc(&rdev->nr_pending);
3253 			atomic_inc(&r10_bio->remaining);
3254 			bio->bi_next = biolist;
3255 			biolist = bio;
3256 			bio->bi_private = r10_bio;
3257 			bio->bi_end_io = end_sync_read;
3258 			bio_set_op_attrs(bio, REQ_OP_READ, 0);
3259 			bio->bi_iter.bi_sector = sector + rdev->data_offset;
3260 			bio->bi_bdev = rdev->bdev;
3261 			count++;
3262 
3263 			rdev = rcu_dereference(conf->mirrors[d].replacement);
3264 			if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3265 				rcu_read_unlock();
3266 				continue;
3267 			}
3268 			atomic_inc(&rdev->nr_pending);
3269 			rcu_read_unlock();
3270 
3271 			/* Need to set up for writing to the replacement */
3272 			bio = r10_bio->devs[i].repl_bio;
3273 			bio_reset(bio);
3274 			bio->bi_error = -EIO;
3275 
3276 			sector = r10_bio->devs[i].addr;
3277 			bio->bi_next = biolist;
3278 			biolist = bio;
3279 			bio->bi_private = r10_bio;
3280 			bio->bi_end_io = end_sync_write;
3281 			bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3282 			bio->bi_iter.bi_sector = sector + rdev->data_offset;
3283 			bio->bi_bdev = rdev->bdev;
3284 			count++;
3285 		}
3286 
3287 		if (count < 2) {
3288 			for (i=0; i<conf->copies; i++) {
3289 				int d = r10_bio->devs[i].devnum;
3290 				if (r10_bio->devs[i].bio->bi_end_io)
3291 					rdev_dec_pending(conf->mirrors[d].rdev,
3292 							 mddev);
3293 				if (r10_bio->devs[i].repl_bio &&
3294 				    r10_bio->devs[i].repl_bio->bi_end_io)
3295 					rdev_dec_pending(
3296 						conf->mirrors[d].replacement,
3297 						mddev);
3298 			}
3299 			put_buf(r10_bio);
3300 			biolist = NULL;
3301 			goto giveup;
3302 		}
3303 	}
3304 
3305 	nr_sectors = 0;
3306 	if (sector_nr + max_sync < max_sector)
3307 		max_sector = sector_nr + max_sync;
3308 	do {
3309 		struct page *page;
3310 		int len = PAGE_SIZE;
3311 		if (sector_nr + (len>>9) > max_sector)
3312 			len = (max_sector - sector_nr) << 9;
3313 		if (len == 0)
3314 			break;
3315 		for (bio= biolist ; bio ; bio=bio->bi_next) {
3316 			struct bio *bio2;
3317 			page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3318 			if (bio_add_page(bio, page, len, 0))
3319 				continue;
3320 
3321 			/* stop here */
3322 			bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3323 			for (bio2 = biolist;
3324 			     bio2 && bio2 != bio;
3325 			     bio2 = bio2->bi_next) {
3326 				/* remove last page from this bio */
3327 				bio2->bi_vcnt--;
3328 				bio2->bi_iter.bi_size -= len;
3329 				bio_clear_flag(bio2, BIO_SEG_VALID);
3330 			}
3331 			goto bio_full;
3332 		}
3333 		nr_sectors += len>>9;
3334 		sector_nr += len>>9;
3335 	} while (biolist->bi_vcnt < RESYNC_PAGES);
3336  bio_full:
3337 	r10_bio->sectors = nr_sectors;
3338 
3339 	while (biolist) {
3340 		bio = biolist;
3341 		biolist = biolist->bi_next;
3342 
3343 		bio->bi_next = NULL;
3344 		r10_bio = bio->bi_private;
3345 		r10_bio->sectors = nr_sectors;
3346 
3347 		if (bio->bi_end_io == end_sync_read) {
3348 			md_sync_acct(bio->bi_bdev, nr_sectors);
3349 			bio->bi_error = 0;
3350 			generic_make_request(bio);
3351 		}
3352 	}
3353 
3354 	if (sectors_skipped)
3355 		/* pretend they weren't skipped, it makes
3356 		 * no important difference in this case
3357 		 */
3358 		md_done_sync(mddev, sectors_skipped, 1);
3359 
3360 	return sectors_skipped + nr_sectors;
3361  giveup:
3362 	/* There is nowhere to write, so all non-sync
3363 	 * drives must be failed or in resync, all drives
3364 	 * have a bad block, so try the next chunk...
3365 	 */
3366 	if (sector_nr + max_sync < max_sector)
3367 		max_sector = sector_nr + max_sync;
3368 
3369 	sectors_skipped += (max_sector - sector_nr);
3370 	chunks_skipped ++;
3371 	sector_nr = max_sector;
3372 	goto skipped;
3373 }
3374 
3375 static sector_t
3376 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3377 {
3378 	sector_t size;
3379 	struct r10conf *conf = mddev->private;
3380 
3381 	if (!raid_disks)
3382 		raid_disks = min(conf->geo.raid_disks,
3383 				 conf->prev.raid_disks);
3384 	if (!sectors)
3385 		sectors = conf->dev_sectors;
3386 
3387 	size = sectors >> conf->geo.chunk_shift;
3388 	sector_div(size, conf->geo.far_copies);
3389 	size = size * raid_disks;
3390 	sector_div(size, conf->geo.near_copies);
3391 
3392 	return size << conf->geo.chunk_shift;
3393 }
3394 
3395 static void calc_sectors(struct r10conf *conf, sector_t size)
3396 {
3397 	/* Calculate the number of sectors-per-device that will
3398 	 * actually be used, and set conf->dev_sectors and
3399 	 * conf->stride
3400 	 */
3401 
3402 	size = size >> conf->geo.chunk_shift;
3403 	sector_div(size, conf->geo.far_copies);
3404 	size = size * conf->geo.raid_disks;
3405 	sector_div(size, conf->geo.near_copies);
3406 	/* 'size' is now the number of chunks in the array */
3407 	/* calculate "used chunks per device" */
3408 	size = size * conf->copies;
3409 
3410 	/* We need to round up when dividing by raid_disks to
3411 	 * get the stride size.
3412 	 */
3413 	size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3414 
3415 	conf->dev_sectors = size << conf->geo.chunk_shift;
3416 
3417 	if (conf->geo.far_offset)
3418 		conf->geo.stride = 1 << conf->geo.chunk_shift;
3419 	else {
3420 		sector_div(size, conf->geo.far_copies);
3421 		conf->geo.stride = size << conf->geo.chunk_shift;
3422 	}
3423 }
3424 
3425 enum geo_type {geo_new, geo_old, geo_start};
3426 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3427 {
3428 	int nc, fc, fo;
3429 	int layout, chunk, disks;
3430 	switch (new) {
3431 	case geo_old:
3432 		layout = mddev->layout;
3433 		chunk = mddev->chunk_sectors;
3434 		disks = mddev->raid_disks - mddev->delta_disks;
3435 		break;
3436 	case geo_new:
3437 		layout = mddev->new_layout;
3438 		chunk = mddev->new_chunk_sectors;
3439 		disks = mddev->raid_disks;
3440 		break;
3441 	default: /* avoid 'may be unused' warnings */
3442 	case geo_start: /* new when starting reshape - raid_disks not
3443 			 * updated yet. */
3444 		layout = mddev->new_layout;
3445 		chunk = mddev->new_chunk_sectors;
3446 		disks = mddev->raid_disks + mddev->delta_disks;
3447 		break;
3448 	}
3449 	if (layout >> 19)
3450 		return -1;
3451 	if (chunk < (PAGE_SIZE >> 9) ||
3452 	    !is_power_of_2(chunk))
3453 		return -2;
3454 	nc = layout & 255;
3455 	fc = (layout >> 8) & 255;
3456 	fo = layout & (1<<16);
3457 	geo->raid_disks = disks;
3458 	geo->near_copies = nc;
3459 	geo->far_copies = fc;
3460 	geo->far_offset = fo;
3461 	switch (layout >> 17) {
3462 	case 0:	/* original layout.  simple but not always optimal */
3463 		geo->far_set_size = disks;
3464 		break;
3465 	case 1: /* "improved" layout which was buggy.  Hopefully no-one is
3466 		 * actually using this, but leave code here just in case.*/
3467 		geo->far_set_size = disks/fc;
3468 		WARN(geo->far_set_size < fc,
3469 		     "This RAID10 layout does not provide data safety - please backup and create new array\n");
3470 		break;
3471 	case 2: /* "improved" layout fixed to match documentation */
3472 		geo->far_set_size = fc * nc;
3473 		break;
3474 	default: /* Not a valid layout */
3475 		return -1;
3476 	}
3477 	geo->chunk_mask = chunk - 1;
3478 	geo->chunk_shift = ffz(~chunk);
3479 	return nc*fc;
3480 }
3481 
3482 static struct r10conf *setup_conf(struct mddev *mddev)
3483 {
3484 	struct r10conf *conf = NULL;
3485 	int err = -EINVAL;
3486 	struct geom geo;
3487 	int copies;
3488 
3489 	copies = setup_geo(&geo, mddev, geo_new);
3490 
3491 	if (copies == -2) {
3492 		printk(KERN_ERR "md/raid10:%s: chunk size must be "
3493 		       "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3494 		       mdname(mddev), PAGE_SIZE);
3495 		goto out;
3496 	}
3497 
3498 	if (copies < 2 || copies > mddev->raid_disks) {
3499 		printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3500 		       mdname(mddev), mddev->new_layout);
3501 		goto out;
3502 	}
3503 
3504 	err = -ENOMEM;
3505 	conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3506 	if (!conf)
3507 		goto out;
3508 
3509 	/* FIXME calc properly */
3510 	conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3511 							    max(0,-mddev->delta_disks)),
3512 				GFP_KERNEL);
3513 	if (!conf->mirrors)
3514 		goto out;
3515 
3516 	conf->tmppage = alloc_page(GFP_KERNEL);
3517 	if (!conf->tmppage)
3518 		goto out;
3519 
3520 	conf->geo = geo;
3521 	conf->copies = copies;
3522 	conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3523 					   r10bio_pool_free, conf);
3524 	if (!conf->r10bio_pool)
3525 		goto out;
3526 
3527 	calc_sectors(conf, mddev->dev_sectors);
3528 	if (mddev->reshape_position == MaxSector) {
3529 		conf->prev = conf->geo;
3530 		conf->reshape_progress = MaxSector;
3531 	} else {
3532 		if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3533 			err = -EINVAL;
3534 			goto out;
3535 		}
3536 		conf->reshape_progress = mddev->reshape_position;
3537 		if (conf->prev.far_offset)
3538 			conf->prev.stride = 1 << conf->prev.chunk_shift;
3539 		else
3540 			/* far_copies must be 1 */
3541 			conf->prev.stride = conf->dev_sectors;
3542 	}
3543 	conf->reshape_safe = conf->reshape_progress;
3544 	spin_lock_init(&conf->device_lock);
3545 	INIT_LIST_HEAD(&conf->retry_list);
3546 	INIT_LIST_HEAD(&conf->bio_end_io_list);
3547 
3548 	spin_lock_init(&conf->resync_lock);
3549 	init_waitqueue_head(&conf->wait_barrier);
3550 	atomic_set(&conf->nr_pending, 0);
3551 
3552 	conf->thread = md_register_thread(raid10d, mddev, "raid10");
3553 	if (!conf->thread)
3554 		goto out;
3555 
3556 	conf->mddev = mddev;
3557 	return conf;
3558 
3559  out:
3560 	if (err == -ENOMEM)
3561 		printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3562 		       mdname(mddev));
3563 	if (conf) {
3564 		mempool_destroy(conf->r10bio_pool);
3565 		kfree(conf->mirrors);
3566 		safe_put_page(conf->tmppage);
3567 		kfree(conf);
3568 	}
3569 	return ERR_PTR(err);
3570 }
3571 
3572 static int raid10_run(struct mddev *mddev)
3573 {
3574 	struct r10conf *conf;
3575 	int i, disk_idx, chunk_size;
3576 	struct raid10_info *disk;
3577 	struct md_rdev *rdev;
3578 	sector_t size;
3579 	sector_t min_offset_diff = 0;
3580 	int first = 1;
3581 	bool discard_supported = false;
3582 
3583 	if (mddev->private == NULL) {
3584 		conf = setup_conf(mddev);
3585 		if (IS_ERR(conf))
3586 			return PTR_ERR(conf);
3587 		mddev->private = conf;
3588 	}
3589 	conf = mddev->private;
3590 	if (!conf)
3591 		goto out;
3592 
3593 	mddev->thread = conf->thread;
3594 	conf->thread = NULL;
3595 
3596 	chunk_size = mddev->chunk_sectors << 9;
3597 	if (mddev->queue) {
3598 		blk_queue_max_discard_sectors(mddev->queue,
3599 					      mddev->chunk_sectors);
3600 		blk_queue_max_write_same_sectors(mddev->queue, 0);
3601 		blk_queue_io_min(mddev->queue, chunk_size);
3602 		if (conf->geo.raid_disks % conf->geo.near_copies)
3603 			blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3604 		else
3605 			blk_queue_io_opt(mddev->queue, chunk_size *
3606 					 (conf->geo.raid_disks / conf->geo.near_copies));
3607 	}
3608 
3609 	rdev_for_each(rdev, mddev) {
3610 		long long diff;
3611 		struct request_queue *q;
3612 
3613 		disk_idx = rdev->raid_disk;
3614 		if (disk_idx < 0)
3615 			continue;
3616 		if (disk_idx >= conf->geo.raid_disks &&
3617 		    disk_idx >= conf->prev.raid_disks)
3618 			continue;
3619 		disk = conf->mirrors + disk_idx;
3620 
3621 		if (test_bit(Replacement, &rdev->flags)) {
3622 			if (disk->replacement)
3623 				goto out_free_conf;
3624 			disk->replacement = rdev;
3625 		} else {
3626 			if (disk->rdev)
3627 				goto out_free_conf;
3628 			disk->rdev = rdev;
3629 		}
3630 		q = bdev_get_queue(rdev->bdev);
3631 		diff = (rdev->new_data_offset - rdev->data_offset);
3632 		if (!mddev->reshape_backwards)
3633 			diff = -diff;
3634 		if (diff < 0)
3635 			diff = 0;
3636 		if (first || diff < min_offset_diff)
3637 			min_offset_diff = diff;
3638 
3639 		if (mddev->gendisk)
3640 			disk_stack_limits(mddev->gendisk, rdev->bdev,
3641 					  rdev->data_offset << 9);
3642 
3643 		disk->head_position = 0;
3644 
3645 		if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3646 			discard_supported = true;
3647 	}
3648 
3649 	if (mddev->queue) {
3650 		if (discard_supported)
3651 			queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3652 						mddev->queue);
3653 		else
3654 			queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3655 						  mddev->queue);
3656 	}
3657 	/* need to check that every block has at least one working mirror */
3658 	if (!enough(conf, -1)) {
3659 		printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3660 		       mdname(mddev));
3661 		goto out_free_conf;
3662 	}
3663 
3664 	if (conf->reshape_progress != MaxSector) {
3665 		/* must ensure that shape change is supported */
3666 		if (conf->geo.far_copies != 1 &&
3667 		    conf->geo.far_offset == 0)
3668 			goto out_free_conf;
3669 		if (conf->prev.far_copies != 1 &&
3670 		    conf->prev.far_offset == 0)
3671 			goto out_free_conf;
3672 	}
3673 
3674 	mddev->degraded = 0;
3675 	for (i = 0;
3676 	     i < conf->geo.raid_disks
3677 		     || i < conf->prev.raid_disks;
3678 	     i++) {
3679 
3680 		disk = conf->mirrors + i;
3681 
3682 		if (!disk->rdev && disk->replacement) {
3683 			/* The replacement is all we have - use it */
3684 			disk->rdev = disk->replacement;
3685 			disk->replacement = NULL;
3686 			clear_bit(Replacement, &disk->rdev->flags);
3687 		}
3688 
3689 		if (!disk->rdev ||
3690 		    !test_bit(In_sync, &disk->rdev->flags)) {
3691 			disk->head_position = 0;
3692 			mddev->degraded++;
3693 			if (disk->rdev &&
3694 			    disk->rdev->saved_raid_disk < 0)
3695 				conf->fullsync = 1;
3696 		}
3697 		disk->recovery_disabled = mddev->recovery_disabled - 1;
3698 	}
3699 
3700 	if (mddev->recovery_cp != MaxSector)
3701 		printk(KERN_NOTICE "md/raid10:%s: not clean"
3702 		       " -- starting background reconstruction\n",
3703 		       mdname(mddev));
3704 	printk(KERN_INFO
3705 		"md/raid10:%s: active with %d out of %d devices\n",
3706 		mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3707 		conf->geo.raid_disks);
3708 	/*
3709 	 * Ok, everything is just fine now
3710 	 */
3711 	mddev->dev_sectors = conf->dev_sectors;
3712 	size = raid10_size(mddev, 0, 0);
3713 	md_set_array_sectors(mddev, size);
3714 	mddev->resync_max_sectors = size;
3715 
3716 	if (mddev->queue) {
3717 		int stripe = conf->geo.raid_disks *
3718 			((mddev->chunk_sectors << 9) / PAGE_SIZE);
3719 
3720 		/* Calculate max read-ahead size.
3721 		 * We need to readahead at least twice a whole stripe....
3722 		 * maybe...
3723 		 */
3724 		stripe /= conf->geo.near_copies;
3725 		if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3726 			mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3727 	}
3728 
3729 	if (md_integrity_register(mddev))
3730 		goto out_free_conf;
3731 
3732 	if (conf->reshape_progress != MaxSector) {
3733 		unsigned long before_length, after_length;
3734 
3735 		before_length = ((1 << conf->prev.chunk_shift) *
3736 				 conf->prev.far_copies);
3737 		after_length = ((1 << conf->geo.chunk_shift) *
3738 				conf->geo.far_copies);
3739 
3740 		if (max(before_length, after_length) > min_offset_diff) {
3741 			/* This cannot work */
3742 			printk("md/raid10: offset difference not enough to continue reshape\n");
3743 			goto out_free_conf;
3744 		}
3745 		conf->offset_diff = min_offset_diff;
3746 
3747 		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3748 		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3749 		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3750 		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3751 		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3752 							"reshape");
3753 	}
3754 
3755 	return 0;
3756 
3757 out_free_conf:
3758 	md_unregister_thread(&mddev->thread);
3759 	mempool_destroy(conf->r10bio_pool);
3760 	safe_put_page(conf->tmppage);
3761 	kfree(conf->mirrors);
3762 	kfree(conf);
3763 	mddev->private = NULL;
3764 out:
3765 	return -EIO;
3766 }
3767 
3768 static void raid10_free(struct mddev *mddev, void *priv)
3769 {
3770 	struct r10conf *conf = priv;
3771 
3772 	mempool_destroy(conf->r10bio_pool);
3773 	safe_put_page(conf->tmppage);
3774 	kfree(conf->mirrors);
3775 	kfree(conf->mirrors_old);
3776 	kfree(conf->mirrors_new);
3777 	kfree(conf);
3778 }
3779 
3780 static void raid10_quiesce(struct mddev *mddev, int state)
3781 {
3782 	struct r10conf *conf = mddev->private;
3783 
3784 	switch(state) {
3785 	case 1:
3786 		raise_barrier(conf, 0);
3787 		break;
3788 	case 0:
3789 		lower_barrier(conf);
3790 		break;
3791 	}
3792 }
3793 
3794 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3795 {
3796 	/* Resize of 'far' arrays is not supported.
3797 	 * For 'near' and 'offset' arrays we can set the
3798 	 * number of sectors used to be an appropriate multiple
3799 	 * of the chunk size.
3800 	 * For 'offset', this is far_copies*chunksize.
3801 	 * For 'near' the multiplier is the LCM of
3802 	 * near_copies and raid_disks.
3803 	 * So if far_copies > 1 && !far_offset, fail.
3804 	 * Else find LCM(raid_disks, near_copy)*far_copies and
3805 	 * multiply by chunk_size.  Then round to this number.
3806 	 * This is mostly done by raid10_size()
3807 	 */
3808 	struct r10conf *conf = mddev->private;
3809 	sector_t oldsize, size;
3810 
3811 	if (mddev->reshape_position != MaxSector)
3812 		return -EBUSY;
3813 
3814 	if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3815 		return -EINVAL;
3816 
3817 	oldsize = raid10_size(mddev, 0, 0);
3818 	size = raid10_size(mddev, sectors, 0);
3819 	if (mddev->external_size &&
3820 	    mddev->array_sectors > size)
3821 		return -EINVAL;
3822 	if (mddev->bitmap) {
3823 		int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3824 		if (ret)
3825 			return ret;
3826 	}
3827 	md_set_array_sectors(mddev, size);
3828 	if (mddev->queue) {
3829 		set_capacity(mddev->gendisk, mddev->array_sectors);
3830 		revalidate_disk(mddev->gendisk);
3831 	}
3832 	if (sectors > mddev->dev_sectors &&
3833 	    mddev->recovery_cp > oldsize) {
3834 		mddev->recovery_cp = oldsize;
3835 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3836 	}
3837 	calc_sectors(conf, sectors);
3838 	mddev->dev_sectors = conf->dev_sectors;
3839 	mddev->resync_max_sectors = size;
3840 	return 0;
3841 }
3842 
3843 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3844 {
3845 	struct md_rdev *rdev;
3846 	struct r10conf *conf;
3847 
3848 	if (mddev->degraded > 0) {
3849 		printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3850 		       mdname(mddev));
3851 		return ERR_PTR(-EINVAL);
3852 	}
3853 	sector_div(size, devs);
3854 
3855 	/* Set new parameters */
3856 	mddev->new_level = 10;
3857 	/* new layout: far_copies = 1, near_copies = 2 */
3858 	mddev->new_layout = (1<<8) + 2;
3859 	mddev->new_chunk_sectors = mddev->chunk_sectors;
3860 	mddev->delta_disks = mddev->raid_disks;
3861 	mddev->raid_disks *= 2;
3862 	/* make sure it will be not marked as dirty */
3863 	mddev->recovery_cp = MaxSector;
3864 	mddev->dev_sectors = size;
3865 
3866 	conf = setup_conf(mddev);
3867 	if (!IS_ERR(conf)) {
3868 		rdev_for_each(rdev, mddev)
3869 			if (rdev->raid_disk >= 0) {
3870 				rdev->new_raid_disk = rdev->raid_disk * 2;
3871 				rdev->sectors = size;
3872 			}
3873 		conf->barrier = 1;
3874 	}
3875 
3876 	return conf;
3877 }
3878 
3879 static void *raid10_takeover(struct mddev *mddev)
3880 {
3881 	struct r0conf *raid0_conf;
3882 
3883 	/* raid10 can take over:
3884 	 *  raid0 - providing it has only two drives
3885 	 */
3886 	if (mddev->level == 0) {
3887 		/* for raid0 takeover only one zone is supported */
3888 		raid0_conf = mddev->private;
3889 		if (raid0_conf->nr_strip_zones > 1) {
3890 			printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3891 			       " with more than one zone.\n",
3892 			       mdname(mddev));
3893 			return ERR_PTR(-EINVAL);
3894 		}
3895 		return raid10_takeover_raid0(mddev,
3896 			raid0_conf->strip_zone->zone_end,
3897 			raid0_conf->strip_zone->nb_dev);
3898 	}
3899 	return ERR_PTR(-EINVAL);
3900 }
3901 
3902 static int raid10_check_reshape(struct mddev *mddev)
3903 {
3904 	/* Called when there is a request to change
3905 	 * - layout (to ->new_layout)
3906 	 * - chunk size (to ->new_chunk_sectors)
3907 	 * - raid_disks (by delta_disks)
3908 	 * or when trying to restart a reshape that was ongoing.
3909 	 *
3910 	 * We need to validate the request and possibly allocate
3911 	 * space if that might be an issue later.
3912 	 *
3913 	 * Currently we reject any reshape of a 'far' mode array,
3914 	 * allow chunk size to change if new is generally acceptable,
3915 	 * allow raid_disks to increase, and allow
3916 	 * a switch between 'near' mode and 'offset' mode.
3917 	 */
3918 	struct r10conf *conf = mddev->private;
3919 	struct geom geo;
3920 
3921 	if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3922 		return -EINVAL;
3923 
3924 	if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3925 		/* mustn't change number of copies */
3926 		return -EINVAL;
3927 	if (geo.far_copies > 1 && !geo.far_offset)
3928 		/* Cannot switch to 'far' mode */
3929 		return -EINVAL;
3930 
3931 	if (mddev->array_sectors & geo.chunk_mask)
3932 			/* not factor of array size */
3933 			return -EINVAL;
3934 
3935 	if (!enough(conf, -1))
3936 		return -EINVAL;
3937 
3938 	kfree(conf->mirrors_new);
3939 	conf->mirrors_new = NULL;
3940 	if (mddev->delta_disks > 0) {
3941 		/* allocate new 'mirrors' list */
3942 		conf->mirrors_new = kzalloc(
3943 			sizeof(struct raid10_info)
3944 			*(mddev->raid_disks +
3945 			  mddev->delta_disks),
3946 			GFP_KERNEL);
3947 		if (!conf->mirrors_new)
3948 			return -ENOMEM;
3949 	}
3950 	return 0;
3951 }
3952 
3953 /*
3954  * Need to check if array has failed when deciding whether to:
3955  *  - start an array
3956  *  - remove non-faulty devices
3957  *  - add a spare
3958  *  - allow a reshape
3959  * This determination is simple when no reshape is happening.
3960  * However if there is a reshape, we need to carefully check
3961  * both the before and after sections.
3962  * This is because some failed devices may only affect one
3963  * of the two sections, and some non-in_sync devices may
3964  * be insync in the section most affected by failed devices.
3965  */
3966 static int calc_degraded(struct r10conf *conf)
3967 {
3968 	int degraded, degraded2;
3969 	int i;
3970 
3971 	rcu_read_lock();
3972 	degraded = 0;
3973 	/* 'prev' section first */
3974 	for (i = 0; i < conf->prev.raid_disks; i++) {
3975 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3976 		if (!rdev || test_bit(Faulty, &rdev->flags))
3977 			degraded++;
3978 		else if (!test_bit(In_sync, &rdev->flags))
3979 			/* When we can reduce the number of devices in
3980 			 * an array, this might not contribute to
3981 			 * 'degraded'.  It does now.
3982 			 */
3983 			degraded++;
3984 	}
3985 	rcu_read_unlock();
3986 	if (conf->geo.raid_disks == conf->prev.raid_disks)
3987 		return degraded;
3988 	rcu_read_lock();
3989 	degraded2 = 0;
3990 	for (i = 0; i < conf->geo.raid_disks; i++) {
3991 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
3992 		if (!rdev || test_bit(Faulty, &rdev->flags))
3993 			degraded2++;
3994 		else if (!test_bit(In_sync, &rdev->flags)) {
3995 			/* If reshape is increasing the number of devices,
3996 			 * this section has already been recovered, so
3997 			 * it doesn't contribute to degraded.
3998 			 * else it does.
3999 			 */
4000 			if (conf->geo.raid_disks <= conf->prev.raid_disks)
4001 				degraded2++;
4002 		}
4003 	}
4004 	rcu_read_unlock();
4005 	if (degraded2 > degraded)
4006 		return degraded2;
4007 	return degraded;
4008 }
4009 
4010 static int raid10_start_reshape(struct mddev *mddev)
4011 {
4012 	/* A 'reshape' has been requested. This commits
4013 	 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4014 	 * This also checks if there are enough spares and adds them
4015 	 * to the array.
4016 	 * We currently require enough spares to make the final
4017 	 * array non-degraded.  We also require that the difference
4018 	 * between old and new data_offset - on each device - is
4019 	 * enough that we never risk over-writing.
4020 	 */
4021 
4022 	unsigned long before_length, after_length;
4023 	sector_t min_offset_diff = 0;
4024 	int first = 1;
4025 	struct geom new;
4026 	struct r10conf *conf = mddev->private;
4027 	struct md_rdev *rdev;
4028 	int spares = 0;
4029 	int ret;
4030 
4031 	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4032 		return -EBUSY;
4033 
4034 	if (setup_geo(&new, mddev, geo_start) != conf->copies)
4035 		return -EINVAL;
4036 
4037 	before_length = ((1 << conf->prev.chunk_shift) *
4038 			 conf->prev.far_copies);
4039 	after_length = ((1 << conf->geo.chunk_shift) *
4040 			conf->geo.far_copies);
4041 
4042 	rdev_for_each(rdev, mddev) {
4043 		if (!test_bit(In_sync, &rdev->flags)
4044 		    && !test_bit(Faulty, &rdev->flags))
4045 			spares++;
4046 		if (rdev->raid_disk >= 0) {
4047 			long long diff = (rdev->new_data_offset
4048 					  - rdev->data_offset);
4049 			if (!mddev->reshape_backwards)
4050 				diff = -diff;
4051 			if (diff < 0)
4052 				diff = 0;
4053 			if (first || diff < min_offset_diff)
4054 				min_offset_diff = diff;
4055 		}
4056 	}
4057 
4058 	if (max(before_length, after_length) > min_offset_diff)
4059 		return -EINVAL;
4060 
4061 	if (spares < mddev->delta_disks)
4062 		return -EINVAL;
4063 
4064 	conf->offset_diff = min_offset_diff;
4065 	spin_lock_irq(&conf->device_lock);
4066 	if (conf->mirrors_new) {
4067 		memcpy(conf->mirrors_new, conf->mirrors,
4068 		       sizeof(struct raid10_info)*conf->prev.raid_disks);
4069 		smp_mb();
4070 		kfree(conf->mirrors_old);
4071 		conf->mirrors_old = conf->mirrors;
4072 		conf->mirrors = conf->mirrors_new;
4073 		conf->mirrors_new = NULL;
4074 	}
4075 	setup_geo(&conf->geo, mddev, geo_start);
4076 	smp_mb();
4077 	if (mddev->reshape_backwards) {
4078 		sector_t size = raid10_size(mddev, 0, 0);
4079 		if (size < mddev->array_sectors) {
4080 			spin_unlock_irq(&conf->device_lock);
4081 			printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4082 			       mdname(mddev));
4083 			return -EINVAL;
4084 		}
4085 		mddev->resync_max_sectors = size;
4086 		conf->reshape_progress = size;
4087 	} else
4088 		conf->reshape_progress = 0;
4089 	conf->reshape_safe = conf->reshape_progress;
4090 	spin_unlock_irq(&conf->device_lock);
4091 
4092 	if (mddev->delta_disks && mddev->bitmap) {
4093 		ret = bitmap_resize(mddev->bitmap,
4094 				    raid10_size(mddev, 0,
4095 						conf->geo.raid_disks),
4096 				    0, 0);
4097 		if (ret)
4098 			goto abort;
4099 	}
4100 	if (mddev->delta_disks > 0) {
4101 		rdev_for_each(rdev, mddev)
4102 			if (rdev->raid_disk < 0 &&
4103 			    !test_bit(Faulty, &rdev->flags)) {
4104 				if (raid10_add_disk(mddev, rdev) == 0) {
4105 					if (rdev->raid_disk >=
4106 					    conf->prev.raid_disks)
4107 						set_bit(In_sync, &rdev->flags);
4108 					else
4109 						rdev->recovery_offset = 0;
4110 
4111 					if (sysfs_link_rdev(mddev, rdev))
4112 						/* Failure here  is OK */;
4113 				}
4114 			} else if (rdev->raid_disk >= conf->prev.raid_disks
4115 				   && !test_bit(Faulty, &rdev->flags)) {
4116 				/* This is a spare that was manually added */
4117 				set_bit(In_sync, &rdev->flags);
4118 			}
4119 	}
4120 	/* When a reshape changes the number of devices,
4121 	 * ->degraded is measured against the larger of the
4122 	 * pre and  post numbers.
4123 	 */
4124 	spin_lock_irq(&conf->device_lock);
4125 	mddev->degraded = calc_degraded(conf);
4126 	spin_unlock_irq(&conf->device_lock);
4127 	mddev->raid_disks = conf->geo.raid_disks;
4128 	mddev->reshape_position = conf->reshape_progress;
4129 	set_bit(MD_CHANGE_DEVS, &mddev->flags);
4130 
4131 	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4132 	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4133 	clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4134 	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4135 	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4136 
4137 	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4138 						"reshape");
4139 	if (!mddev->sync_thread) {
4140 		ret = -EAGAIN;
4141 		goto abort;
4142 	}
4143 	conf->reshape_checkpoint = jiffies;
4144 	md_wakeup_thread(mddev->sync_thread);
4145 	md_new_event(mddev);
4146 	return 0;
4147 
4148 abort:
4149 	mddev->recovery = 0;
4150 	spin_lock_irq(&conf->device_lock);
4151 	conf->geo = conf->prev;
4152 	mddev->raid_disks = conf->geo.raid_disks;
4153 	rdev_for_each(rdev, mddev)
4154 		rdev->new_data_offset = rdev->data_offset;
4155 	smp_wmb();
4156 	conf->reshape_progress = MaxSector;
4157 	conf->reshape_safe = MaxSector;
4158 	mddev->reshape_position = MaxSector;
4159 	spin_unlock_irq(&conf->device_lock);
4160 	return ret;
4161 }
4162 
4163 /* Calculate the last device-address that could contain
4164  * any block from the chunk that includes the array-address 's'
4165  * and report the next address.
4166  * i.e. the address returned will be chunk-aligned and after
4167  * any data that is in the chunk containing 's'.
4168  */
4169 static sector_t last_dev_address(sector_t s, struct geom *geo)
4170 {
4171 	s = (s | geo->chunk_mask) + 1;
4172 	s >>= geo->chunk_shift;
4173 	s *= geo->near_copies;
4174 	s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4175 	s *= geo->far_copies;
4176 	s <<= geo->chunk_shift;
4177 	return s;
4178 }
4179 
4180 /* Calculate the first device-address that could contain
4181  * any block from the chunk that includes the array-address 's'.
4182  * This too will be the start of a chunk
4183  */
4184 static sector_t first_dev_address(sector_t s, struct geom *geo)
4185 {
4186 	s >>= geo->chunk_shift;
4187 	s *= geo->near_copies;
4188 	sector_div(s, geo->raid_disks);
4189 	s *= geo->far_copies;
4190 	s <<= geo->chunk_shift;
4191 	return s;
4192 }
4193 
4194 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4195 				int *skipped)
4196 {
4197 	/* We simply copy at most one chunk (smallest of old and new)
4198 	 * at a time, possibly less if that exceeds RESYNC_PAGES,
4199 	 * or we hit a bad block or something.
4200 	 * This might mean we pause for normal IO in the middle of
4201 	 * a chunk, but that is not a problem as mddev->reshape_position
4202 	 * can record any location.
4203 	 *
4204 	 * If we will want to write to a location that isn't
4205 	 * yet recorded as 'safe' (i.e. in metadata on disk) then
4206 	 * we need to flush all reshape requests and update the metadata.
4207 	 *
4208 	 * When reshaping forwards (e.g. to more devices), we interpret
4209 	 * 'safe' as the earliest block which might not have been copied
4210 	 * down yet.  We divide this by previous stripe size and multiply
4211 	 * by previous stripe length to get lowest device offset that we
4212 	 * cannot write to yet.
4213 	 * We interpret 'sector_nr' as an address that we want to write to.
4214 	 * From this we use last_device_address() to find where we might
4215 	 * write to, and first_device_address on the  'safe' position.
4216 	 * If this 'next' write position is after the 'safe' position,
4217 	 * we must update the metadata to increase the 'safe' position.
4218 	 *
4219 	 * When reshaping backwards, we round in the opposite direction
4220 	 * and perform the reverse test:  next write position must not be
4221 	 * less than current safe position.
4222 	 *
4223 	 * In all this the minimum difference in data offsets
4224 	 * (conf->offset_diff - always positive) allows a bit of slack,
4225 	 * so next can be after 'safe', but not by more than offset_diff
4226 	 *
4227 	 * We need to prepare all the bios here before we start any IO
4228 	 * to ensure the size we choose is acceptable to all devices.
4229 	 * The means one for each copy for write-out and an extra one for
4230 	 * read-in.
4231 	 * We store the read-in bio in ->master_bio and the others in
4232 	 * ->devs[x].bio and ->devs[x].repl_bio.
4233 	 */
4234 	struct r10conf *conf = mddev->private;
4235 	struct r10bio *r10_bio;
4236 	sector_t next, safe, last;
4237 	int max_sectors;
4238 	int nr_sectors;
4239 	int s;
4240 	struct md_rdev *rdev;
4241 	int need_flush = 0;
4242 	struct bio *blist;
4243 	struct bio *bio, *read_bio;
4244 	int sectors_done = 0;
4245 
4246 	if (sector_nr == 0) {
4247 		/* If restarting in the middle, skip the initial sectors */
4248 		if (mddev->reshape_backwards &&
4249 		    conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4250 			sector_nr = (raid10_size(mddev, 0, 0)
4251 				     - conf->reshape_progress);
4252 		} else if (!mddev->reshape_backwards &&
4253 			   conf->reshape_progress > 0)
4254 			sector_nr = conf->reshape_progress;
4255 		if (sector_nr) {
4256 			mddev->curr_resync_completed = sector_nr;
4257 			sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4258 			*skipped = 1;
4259 			return sector_nr;
4260 		}
4261 	}
4262 
4263 	/* We don't use sector_nr to track where we are up to
4264 	 * as that doesn't work well for ->reshape_backwards.
4265 	 * So just use ->reshape_progress.
4266 	 */
4267 	if (mddev->reshape_backwards) {
4268 		/* 'next' is the earliest device address that we might
4269 		 * write to for this chunk in the new layout
4270 		 */
4271 		next = first_dev_address(conf->reshape_progress - 1,
4272 					 &conf->geo);
4273 
4274 		/* 'safe' is the last device address that we might read from
4275 		 * in the old layout after a restart
4276 		 */
4277 		safe = last_dev_address(conf->reshape_safe - 1,
4278 					&conf->prev);
4279 
4280 		if (next + conf->offset_diff < safe)
4281 			need_flush = 1;
4282 
4283 		last = conf->reshape_progress - 1;
4284 		sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4285 					       & conf->prev.chunk_mask);
4286 		if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4287 			sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4288 	} else {
4289 		/* 'next' is after the last device address that we
4290 		 * might write to for this chunk in the new layout
4291 		 */
4292 		next = last_dev_address(conf->reshape_progress, &conf->geo);
4293 
4294 		/* 'safe' is the earliest device address that we might
4295 		 * read from in the old layout after a restart
4296 		 */
4297 		safe = first_dev_address(conf->reshape_safe, &conf->prev);
4298 
4299 		/* Need to update metadata if 'next' might be beyond 'safe'
4300 		 * as that would possibly corrupt data
4301 		 */
4302 		if (next > safe + conf->offset_diff)
4303 			need_flush = 1;
4304 
4305 		sector_nr = conf->reshape_progress;
4306 		last  = sector_nr | (conf->geo.chunk_mask
4307 				     & conf->prev.chunk_mask);
4308 
4309 		if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4310 			last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4311 	}
4312 
4313 	if (need_flush ||
4314 	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4315 		/* Need to update reshape_position in metadata */
4316 		wait_barrier(conf);
4317 		mddev->reshape_position = conf->reshape_progress;
4318 		if (mddev->reshape_backwards)
4319 			mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4320 				- conf->reshape_progress;
4321 		else
4322 			mddev->curr_resync_completed = conf->reshape_progress;
4323 		conf->reshape_checkpoint = jiffies;
4324 		set_bit(MD_CHANGE_DEVS, &mddev->flags);
4325 		md_wakeup_thread(mddev->thread);
4326 		wait_event(mddev->sb_wait, mddev->flags == 0 ||
4327 			   test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4328 		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4329 			allow_barrier(conf);
4330 			return sectors_done;
4331 		}
4332 		conf->reshape_safe = mddev->reshape_position;
4333 		allow_barrier(conf);
4334 	}
4335 
4336 read_more:
4337 	/* Now schedule reads for blocks from sector_nr to last */
4338 	r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4339 	r10_bio->state = 0;
4340 	raise_barrier(conf, sectors_done != 0);
4341 	atomic_set(&r10_bio->remaining, 0);
4342 	r10_bio->mddev = mddev;
4343 	r10_bio->sector = sector_nr;
4344 	set_bit(R10BIO_IsReshape, &r10_bio->state);
4345 	r10_bio->sectors = last - sector_nr + 1;
4346 	rdev = read_balance(conf, r10_bio, &max_sectors);
4347 	BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4348 
4349 	if (!rdev) {
4350 		/* Cannot read from here, so need to record bad blocks
4351 		 * on all the target devices.
4352 		 */
4353 		// FIXME
4354 		mempool_free(r10_bio, conf->r10buf_pool);
4355 		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4356 		return sectors_done;
4357 	}
4358 
4359 	read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4360 
4361 	read_bio->bi_bdev = rdev->bdev;
4362 	read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4363 			       + rdev->data_offset);
4364 	read_bio->bi_private = r10_bio;
4365 	read_bio->bi_end_io = end_sync_read;
4366 	bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4367 	read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4368 	read_bio->bi_error = 0;
4369 	read_bio->bi_vcnt = 0;
4370 	read_bio->bi_iter.bi_size = 0;
4371 	r10_bio->master_bio = read_bio;
4372 	r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4373 
4374 	/* Now find the locations in the new layout */
4375 	__raid10_find_phys(&conf->geo, r10_bio);
4376 
4377 	blist = read_bio;
4378 	read_bio->bi_next = NULL;
4379 
4380 	rcu_read_lock();
4381 	for (s = 0; s < conf->copies*2; s++) {
4382 		struct bio *b;
4383 		int d = r10_bio->devs[s/2].devnum;
4384 		struct md_rdev *rdev2;
4385 		if (s&1) {
4386 			rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4387 			b = r10_bio->devs[s/2].repl_bio;
4388 		} else {
4389 			rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4390 			b = r10_bio->devs[s/2].bio;
4391 		}
4392 		if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4393 			continue;
4394 
4395 		bio_reset(b);
4396 		b->bi_bdev = rdev2->bdev;
4397 		b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4398 			rdev2->new_data_offset;
4399 		b->bi_private = r10_bio;
4400 		b->bi_end_io = end_reshape_write;
4401 		bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4402 		b->bi_next = blist;
4403 		blist = b;
4404 	}
4405 
4406 	/* Now add as many pages as possible to all of these bios. */
4407 
4408 	nr_sectors = 0;
4409 	for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4410 		struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4411 		int len = (max_sectors - s) << 9;
4412 		if (len > PAGE_SIZE)
4413 			len = PAGE_SIZE;
4414 		for (bio = blist; bio ; bio = bio->bi_next) {
4415 			struct bio *bio2;
4416 			if (bio_add_page(bio, page, len, 0))
4417 				continue;
4418 
4419 			/* Didn't fit, must stop */
4420 			for (bio2 = blist;
4421 			     bio2 && bio2 != bio;
4422 			     bio2 = bio2->bi_next) {
4423 				/* Remove last page from this bio */
4424 				bio2->bi_vcnt--;
4425 				bio2->bi_iter.bi_size -= len;
4426 				bio_clear_flag(bio2, BIO_SEG_VALID);
4427 			}
4428 			goto bio_full;
4429 		}
4430 		sector_nr += len >> 9;
4431 		nr_sectors += len >> 9;
4432 	}
4433 bio_full:
4434 	rcu_read_unlock();
4435 	r10_bio->sectors = nr_sectors;
4436 
4437 	/* Now submit the read */
4438 	md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4439 	atomic_inc(&r10_bio->remaining);
4440 	read_bio->bi_next = NULL;
4441 	generic_make_request(read_bio);
4442 	sector_nr += nr_sectors;
4443 	sectors_done += nr_sectors;
4444 	if (sector_nr <= last)
4445 		goto read_more;
4446 
4447 	/* Now that we have done the whole section we can
4448 	 * update reshape_progress
4449 	 */
4450 	if (mddev->reshape_backwards)
4451 		conf->reshape_progress -= sectors_done;
4452 	else
4453 		conf->reshape_progress += sectors_done;
4454 
4455 	return sectors_done;
4456 }
4457 
4458 static void end_reshape_request(struct r10bio *r10_bio);
4459 static int handle_reshape_read_error(struct mddev *mddev,
4460 				     struct r10bio *r10_bio);
4461 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4462 {
4463 	/* Reshape read completed.  Hopefully we have a block
4464 	 * to write out.
4465 	 * If we got a read error then we do sync 1-page reads from
4466 	 * elsewhere until we find the data - or give up.
4467 	 */
4468 	struct r10conf *conf = mddev->private;
4469 	int s;
4470 
4471 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4472 		if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4473 			/* Reshape has been aborted */
4474 			md_done_sync(mddev, r10_bio->sectors, 0);
4475 			return;
4476 		}
4477 
4478 	/* We definitely have the data in the pages, schedule the
4479 	 * writes.
4480 	 */
4481 	atomic_set(&r10_bio->remaining, 1);
4482 	for (s = 0; s < conf->copies*2; s++) {
4483 		struct bio *b;
4484 		int d = r10_bio->devs[s/2].devnum;
4485 		struct md_rdev *rdev;
4486 		rcu_read_lock();
4487 		if (s&1) {
4488 			rdev = rcu_dereference(conf->mirrors[d].replacement);
4489 			b = r10_bio->devs[s/2].repl_bio;
4490 		} else {
4491 			rdev = rcu_dereference(conf->mirrors[d].rdev);
4492 			b = r10_bio->devs[s/2].bio;
4493 		}
4494 		if (!rdev || test_bit(Faulty, &rdev->flags)) {
4495 			rcu_read_unlock();
4496 			continue;
4497 		}
4498 		atomic_inc(&rdev->nr_pending);
4499 		rcu_read_unlock();
4500 		md_sync_acct(b->bi_bdev, r10_bio->sectors);
4501 		atomic_inc(&r10_bio->remaining);
4502 		b->bi_next = NULL;
4503 		generic_make_request(b);
4504 	}
4505 	end_reshape_request(r10_bio);
4506 }
4507 
4508 static void end_reshape(struct r10conf *conf)
4509 {
4510 	if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4511 		return;
4512 
4513 	spin_lock_irq(&conf->device_lock);
4514 	conf->prev = conf->geo;
4515 	md_finish_reshape(conf->mddev);
4516 	smp_wmb();
4517 	conf->reshape_progress = MaxSector;
4518 	conf->reshape_safe = MaxSector;
4519 	spin_unlock_irq(&conf->device_lock);
4520 
4521 	/* read-ahead size must cover two whole stripes, which is
4522 	 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4523 	 */
4524 	if (conf->mddev->queue) {
4525 		int stripe = conf->geo.raid_disks *
4526 			((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4527 		stripe /= conf->geo.near_copies;
4528 		if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4529 			conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4530 	}
4531 	conf->fullsync = 0;
4532 }
4533 
4534 static int handle_reshape_read_error(struct mddev *mddev,
4535 				     struct r10bio *r10_bio)
4536 {
4537 	/* Use sync reads to get the blocks from somewhere else */
4538 	int sectors = r10_bio->sectors;
4539 	struct r10conf *conf = mddev->private;
4540 	struct {
4541 		struct r10bio r10_bio;
4542 		struct r10dev devs[conf->copies];
4543 	} on_stack;
4544 	struct r10bio *r10b = &on_stack.r10_bio;
4545 	int slot = 0;
4546 	int idx = 0;
4547 	struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4548 
4549 	r10b->sector = r10_bio->sector;
4550 	__raid10_find_phys(&conf->prev, r10b);
4551 
4552 	while (sectors) {
4553 		int s = sectors;
4554 		int success = 0;
4555 		int first_slot = slot;
4556 
4557 		if (s > (PAGE_SIZE >> 9))
4558 			s = PAGE_SIZE >> 9;
4559 
4560 		rcu_read_lock();
4561 		while (!success) {
4562 			int d = r10b->devs[slot].devnum;
4563 			struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4564 			sector_t addr;
4565 			if (rdev == NULL ||
4566 			    test_bit(Faulty, &rdev->flags) ||
4567 			    !test_bit(In_sync, &rdev->flags))
4568 				goto failed;
4569 
4570 			addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4571 			atomic_inc(&rdev->nr_pending);
4572 			rcu_read_unlock();
4573 			success = sync_page_io(rdev,
4574 					       addr,
4575 					       s << 9,
4576 					       bvec[idx].bv_page,
4577 					       REQ_OP_READ, 0, false);
4578 			rdev_dec_pending(rdev, mddev);
4579 			rcu_read_lock();
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 		rcu_read_unlock();
4590 		if (!success) {
4591 			/* couldn't read this block, must give up */
4592 			set_bit(MD_RECOVERY_INTR,
4593 				&mddev->recovery);
4594 			return -EIO;
4595 		}
4596 		sectors -= s;
4597 		idx++;
4598 	}
4599 	return 0;
4600 }
4601 
4602 static void end_reshape_write(struct bio *bio)
4603 {
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 (bio->bi_error) {
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 		if (mddev->queue) {
4654 			set_capacity(mddev->gendisk, mddev->array_sectors);
4655 			revalidate_disk(mddev->gendisk);
4656 		}
4657 	} else {
4658 		int d;
4659 		rcu_read_lock();
4660 		for (d = conf->geo.raid_disks ;
4661 		     d < conf->geo.raid_disks - mddev->delta_disks;
4662 		     d++) {
4663 			struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4664 			if (rdev)
4665 				clear_bit(In_sync, &rdev->flags);
4666 			rdev = rcu_dereference(conf->mirrors[d].replacement);
4667 			if (rdev)
4668 				clear_bit(In_sync, &rdev->flags);
4669 		}
4670 		rcu_read_unlock();
4671 	}
4672 	mddev->layout = mddev->new_layout;
4673 	mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4674 	mddev->reshape_position = MaxSector;
4675 	mddev->delta_disks = 0;
4676 	mddev->reshape_backwards = 0;
4677 }
4678 
4679 static struct md_personality raid10_personality =
4680 {
4681 	.name		= "raid10",
4682 	.level		= 10,
4683 	.owner		= THIS_MODULE,
4684 	.make_request	= raid10_make_request,
4685 	.run		= raid10_run,
4686 	.free		= raid10_free,
4687 	.status		= raid10_status,
4688 	.error_handler	= raid10_error,
4689 	.hot_add_disk	= raid10_add_disk,
4690 	.hot_remove_disk= raid10_remove_disk,
4691 	.spare_active	= raid10_spare_active,
4692 	.sync_request	= raid10_sync_request,
4693 	.quiesce	= raid10_quiesce,
4694 	.size		= raid10_size,
4695 	.resize		= raid10_resize,
4696 	.takeover	= raid10_takeover,
4697 	.check_reshape	= raid10_check_reshape,
4698 	.start_reshape	= raid10_start_reshape,
4699 	.finish_reshape	= raid10_finish_reshape,
4700 	.congested	= raid10_congested,
4701 };
4702 
4703 static int __init raid_init(void)
4704 {
4705 	return register_md_personality(&raid10_personality);
4706 }
4707 
4708 static void raid_exit(void)
4709 {
4710 	unregister_md_personality(&raid10_personality);
4711 }
4712 
4713 module_init(raid_init);
4714 module_exit(raid_exit);
4715 MODULE_LICENSE("GPL");
4716 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4717 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4718 MODULE_ALIAS("md-raid10");
4719 MODULE_ALIAS("md-level-10");
4720 
4721 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
4722