xref: /openbmc/linux/drivers/md/raid10.c (revision ae3473231e77a3f1909d48cd144cebe5e1d049b3)
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 <trace/events/block.h>
29 #include "md.h"
30 #include "raid10.h"
31 #include "raid0.h"
32 #include "bitmap.h"
33 
34 /*
35  * RAID10 provides a combination of RAID0 and RAID1 functionality.
36  * The layout of data is defined by
37  *    chunk_size
38  *    raid_disks
39  *    near_copies (stored in low byte of layout)
40  *    far_copies (stored in second byte of layout)
41  *    far_offset (stored in bit 16 of layout )
42  *    use_far_sets (stored in bit 17 of layout )
43  *    use_far_sets_bugfixed (stored in bit 18 of layout )
44  *
45  * The data to be stored is divided into chunks using chunksize.  Each device
46  * is divided into far_copies sections.   In each section, chunks are laid out
47  * in a style similar to raid0, but near_copies copies of each chunk is stored
48  * (each on a different drive).  The starting device for each section is offset
49  * near_copies from the starting device of the previous section.  Thus there
50  * are (near_copies * far_copies) of each chunk, and each is on a different
51  * drive.  near_copies and far_copies must be at least one, and their product
52  * is at most raid_disks.
53  *
54  * If far_offset is true, then the far_copies are handled a bit differently.
55  * The copies are still in different stripes, but instead of being very far
56  * apart on disk, there are adjacent stripes.
57  *
58  * The far and offset algorithms are handled slightly differently if
59  * 'use_far_sets' is true.  In this case, the array's devices are grouped into
60  * sets that are (near_copies * far_copies) in size.  The far copied stripes
61  * are still shifted by 'near_copies' devices, but this shifting stays confined
62  * to the set rather than the entire array.  This is done to improve the number
63  * of device combinations that can fail without causing the array to fail.
64  * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
65  * on a device):
66  *    A B C D    A B C D E
67  *      ...         ...
68  *    D A B C    E A B C D
69  * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
70  *    [A B] [C D]    [A B] [C D E]
71  *    |...| |...|    |...| | ... |
72  *    [B A] [D C]    [B A] [E C D]
73  */
74 
75 /*
76  * Number of guaranteed r10bios in case of extreme VM load:
77  */
78 #define	NR_RAID10_BIOS 256
79 
80 /* when we get a read error on a read-only array, we redirect to another
81  * device without failing the first device, or trying to over-write to
82  * correct the read error.  To keep track of bad blocks on a per-bio
83  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
84  */
85 #define IO_BLOCKED ((struct bio *)1)
86 /* When we successfully write to a known bad-block, we need to remove the
87  * bad-block marking which must be done from process context.  So we record
88  * the success by setting devs[n].bio to IO_MADE_GOOD
89  */
90 #define IO_MADE_GOOD ((struct bio *)2)
91 
92 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
93 
94 /* When there are this many requests queued to be written by
95  * the raid10 thread, we become 'congested' to provide back-pressure
96  * for writeback.
97  */
98 static int max_queued_requests = 1024;
99 
100 static void allow_barrier(struct r10conf *conf);
101 static void lower_barrier(struct r10conf *conf);
102 static int _enough(struct r10conf *conf, int previous, int ignore);
103 static int enough(struct r10conf *conf, int ignore);
104 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
105 				int *skipped);
106 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
107 static void end_reshape_write(struct bio *bio);
108 static void end_reshape(struct r10conf *conf);
109 
110 #define raid10_log(md, fmt, args...)				\
111 	do { if ((md)->queue) blk_add_trace_msg((md)->queue, "raid10 " fmt, ##args); } while (0)
112 
113 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
114 {
115 	struct r10conf *conf = data;
116 	int size = offsetof(struct r10bio, devs[conf->copies]);
117 
118 	/* allocate a r10bio with room for raid_disks entries in the
119 	 * bios array */
120 	return kzalloc(size, gfp_flags);
121 }
122 
123 static void r10bio_pool_free(void *r10_bio, void *data)
124 {
125 	kfree(r10_bio);
126 }
127 
128 /* Maximum size of each resync request */
129 #define RESYNC_BLOCK_SIZE (64*1024)
130 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
131 /* amount of memory to reserve for resync requests */
132 #define RESYNC_WINDOW (1024*1024)
133 /* maximum number of concurrent requests, memory permitting */
134 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
135 
136 /*
137  * When performing a resync, we need to read and compare, so
138  * we need as many pages are there are copies.
139  * When performing a recovery, we need 2 bios, one for read,
140  * one for write (we recover only one drive per r10buf)
141  *
142  */
143 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
144 {
145 	struct r10conf *conf = data;
146 	struct page *page;
147 	struct r10bio *r10_bio;
148 	struct bio *bio;
149 	int i, j;
150 	int nalloc;
151 
152 	r10_bio = r10bio_pool_alloc(gfp_flags, conf);
153 	if (!r10_bio)
154 		return NULL;
155 
156 	if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
157 	    test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
158 		nalloc = conf->copies; /* resync */
159 	else
160 		nalloc = 2; /* recovery */
161 
162 	/*
163 	 * Allocate bios.
164 	 */
165 	for (j = nalloc ; j-- ; ) {
166 		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
167 		if (!bio)
168 			goto out_free_bio;
169 		r10_bio->devs[j].bio = bio;
170 		if (!conf->have_replacement)
171 			continue;
172 		bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
173 		if (!bio)
174 			goto out_free_bio;
175 		r10_bio->devs[j].repl_bio = bio;
176 	}
177 	/*
178 	 * Allocate RESYNC_PAGES data pages and attach them
179 	 * where needed.
180 	 */
181 	for (j = 0 ; j < nalloc; j++) {
182 		struct bio *rbio = r10_bio->devs[j].repl_bio;
183 		bio = r10_bio->devs[j].bio;
184 		for (i = 0; i < RESYNC_PAGES; i++) {
185 			if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
186 					       &conf->mddev->recovery)) {
187 				/* we can share bv_page's during recovery
188 				 * and reshape */
189 				struct bio *rbio = r10_bio->devs[0].bio;
190 				page = rbio->bi_io_vec[i].bv_page;
191 				get_page(page);
192 			} else
193 				page = alloc_page(gfp_flags);
194 			if (unlikely(!page))
195 				goto out_free_pages;
196 
197 			bio->bi_io_vec[i].bv_page = page;
198 			if (rbio)
199 				rbio->bi_io_vec[i].bv_page = page;
200 		}
201 	}
202 
203 	return r10_bio;
204 
205 out_free_pages:
206 	for ( ; i > 0 ; i--)
207 		safe_put_page(bio->bi_io_vec[i-1].bv_page);
208 	while (j--)
209 		for (i = 0; i < RESYNC_PAGES ; i++)
210 			safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
211 	j = 0;
212 out_free_bio:
213 	for ( ; j < nalloc; j++) {
214 		if (r10_bio->devs[j].bio)
215 			bio_put(r10_bio->devs[j].bio);
216 		if (r10_bio->devs[j].repl_bio)
217 			bio_put(r10_bio->devs[j].repl_bio);
218 	}
219 	r10bio_pool_free(r10_bio, conf);
220 	return NULL;
221 }
222 
223 static void r10buf_pool_free(void *__r10_bio, void *data)
224 {
225 	int i;
226 	struct r10conf *conf = data;
227 	struct r10bio *r10bio = __r10_bio;
228 	int j;
229 
230 	for (j=0; j < conf->copies; j++) {
231 		struct bio *bio = r10bio->devs[j].bio;
232 		if (bio) {
233 			for (i = 0; i < RESYNC_PAGES; i++) {
234 				safe_put_page(bio->bi_io_vec[i].bv_page);
235 				bio->bi_io_vec[i].bv_page = NULL;
236 			}
237 			bio_put(bio);
238 		}
239 		bio = r10bio->devs[j].repl_bio;
240 		if (bio)
241 			bio_put(bio);
242 	}
243 	r10bio_pool_free(r10bio, conf);
244 }
245 
246 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
247 {
248 	int i;
249 
250 	for (i = 0; i < conf->copies; i++) {
251 		struct bio **bio = & r10_bio->devs[i].bio;
252 		if (!BIO_SPECIAL(*bio))
253 			bio_put(*bio);
254 		*bio = NULL;
255 		bio = &r10_bio->devs[i].repl_bio;
256 		if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
257 			bio_put(*bio);
258 		*bio = NULL;
259 	}
260 }
261 
262 static void free_r10bio(struct r10bio *r10_bio)
263 {
264 	struct r10conf *conf = r10_bio->mddev->private;
265 
266 	put_all_bios(conf, r10_bio);
267 	mempool_free(r10_bio, conf->r10bio_pool);
268 }
269 
270 static void put_buf(struct r10bio *r10_bio)
271 {
272 	struct r10conf *conf = r10_bio->mddev->private;
273 
274 	mempool_free(r10_bio, conf->r10buf_pool);
275 
276 	lower_barrier(conf);
277 }
278 
279 static void reschedule_retry(struct r10bio *r10_bio)
280 {
281 	unsigned long flags;
282 	struct mddev *mddev = r10_bio->mddev;
283 	struct r10conf *conf = mddev->private;
284 
285 	spin_lock_irqsave(&conf->device_lock, flags);
286 	list_add(&r10_bio->retry_list, &conf->retry_list);
287 	conf->nr_queued ++;
288 	spin_unlock_irqrestore(&conf->device_lock, flags);
289 
290 	/* wake up frozen array... */
291 	wake_up(&conf->wait_barrier);
292 
293 	md_wakeup_thread(mddev->thread);
294 }
295 
296 /*
297  * raid_end_bio_io() is called when we have finished servicing a mirrored
298  * operation and are ready to return a success/failure code to the buffer
299  * cache layer.
300  */
301 static void raid_end_bio_io(struct r10bio *r10_bio)
302 {
303 	struct bio *bio = r10_bio->master_bio;
304 	int done;
305 	struct r10conf *conf = r10_bio->mddev->private;
306 
307 	if (bio->bi_phys_segments) {
308 		unsigned long flags;
309 		spin_lock_irqsave(&conf->device_lock, flags);
310 		bio->bi_phys_segments--;
311 		done = (bio->bi_phys_segments == 0);
312 		spin_unlock_irqrestore(&conf->device_lock, flags);
313 	} else
314 		done = 1;
315 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
316 		bio->bi_error = -EIO;
317 	if (done) {
318 		bio_endio(bio);
319 		/*
320 		 * Wake up any possible resync thread that waits for the device
321 		 * to go idle.
322 		 */
323 		allow_barrier(conf);
324 	}
325 	free_r10bio(r10_bio);
326 }
327 
328 /*
329  * Update disk head position estimator based on IRQ completion info.
330  */
331 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
332 {
333 	struct r10conf *conf = r10_bio->mddev->private;
334 
335 	conf->mirrors[r10_bio->devs[slot].devnum].head_position =
336 		r10_bio->devs[slot].addr + (r10_bio->sectors);
337 }
338 
339 /*
340  * Find the disk number which triggered given bio
341  */
342 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
343 			 struct bio *bio, int *slotp, int *replp)
344 {
345 	int slot;
346 	int repl = 0;
347 
348 	for (slot = 0; slot < conf->copies; slot++) {
349 		if (r10_bio->devs[slot].bio == bio)
350 			break;
351 		if (r10_bio->devs[slot].repl_bio == bio) {
352 			repl = 1;
353 			break;
354 		}
355 	}
356 
357 	BUG_ON(slot == conf->copies);
358 	update_head_pos(slot, r10_bio);
359 
360 	if (slotp)
361 		*slotp = slot;
362 	if (replp)
363 		*replp = repl;
364 	return r10_bio->devs[slot].devnum;
365 }
366 
367 static void raid10_end_read_request(struct bio *bio)
368 {
369 	int uptodate = !bio->bi_error;
370 	struct r10bio *r10_bio = bio->bi_private;
371 	int slot, dev;
372 	struct md_rdev *rdev;
373 	struct r10conf *conf = r10_bio->mddev->private;
374 
375 	slot = r10_bio->read_slot;
376 	dev = r10_bio->devs[slot].devnum;
377 	rdev = r10_bio->devs[slot].rdev;
378 	/*
379 	 * this branch is our 'one mirror IO has finished' event handler:
380 	 */
381 	update_head_pos(slot, r10_bio);
382 
383 	if (uptodate) {
384 		/*
385 		 * Set R10BIO_Uptodate in our master bio, so that
386 		 * we will return a good error code to the higher
387 		 * levels even if IO on some other mirrored buffer fails.
388 		 *
389 		 * The 'master' represents the composite IO operation to
390 		 * user-side. So if something waits for IO, then it will
391 		 * wait for the 'master' bio.
392 		 */
393 		set_bit(R10BIO_Uptodate, &r10_bio->state);
394 	} else {
395 		/* If all other devices that store this block have
396 		 * failed, we want to return the error upwards rather
397 		 * than fail the last device.  Here we redefine
398 		 * "uptodate" to mean "Don't want to retry"
399 		 */
400 		if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
401 			     rdev->raid_disk))
402 			uptodate = 1;
403 	}
404 	if (uptodate) {
405 		raid_end_bio_io(r10_bio);
406 		rdev_dec_pending(rdev, conf->mddev);
407 	} else {
408 		/*
409 		 * oops, read error - keep the refcount on the rdev
410 		 */
411 		char b[BDEVNAME_SIZE];
412 		pr_err_ratelimited("md/raid10:%s: %s: rescheduling sector %llu\n",
413 				   mdname(conf->mddev),
414 				   bdevname(rdev->bdev, b),
415 				   (unsigned long long)r10_bio->sector);
416 		set_bit(R10BIO_ReadError, &r10_bio->state);
417 		reschedule_retry(r10_bio);
418 	}
419 }
420 
421 static void close_write(struct r10bio *r10_bio)
422 {
423 	/* clear the bitmap if all writes complete successfully */
424 	bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
425 			r10_bio->sectors,
426 			!test_bit(R10BIO_Degraded, &r10_bio->state),
427 			0);
428 	md_write_end(r10_bio->mddev);
429 }
430 
431 static void one_write_done(struct r10bio *r10_bio)
432 {
433 	if (atomic_dec_and_test(&r10_bio->remaining)) {
434 		if (test_bit(R10BIO_WriteError, &r10_bio->state))
435 			reschedule_retry(r10_bio);
436 		else {
437 			close_write(r10_bio);
438 			if (test_bit(R10BIO_MadeGood, &r10_bio->state))
439 				reschedule_retry(r10_bio);
440 			else
441 				raid_end_bio_io(r10_bio);
442 		}
443 	}
444 }
445 
446 static void raid10_end_write_request(struct bio *bio)
447 {
448 	struct r10bio *r10_bio = bio->bi_private;
449 	int dev;
450 	int dec_rdev = 1;
451 	struct r10conf *conf = r10_bio->mddev->private;
452 	int slot, repl;
453 	struct md_rdev *rdev = NULL;
454 	struct bio *to_put = NULL;
455 	bool discard_error;
456 
457 	discard_error = bio->bi_error && bio_op(bio) == REQ_OP_DISCARD;
458 
459 	dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
460 
461 	if (repl)
462 		rdev = conf->mirrors[dev].replacement;
463 	if (!rdev) {
464 		smp_rmb();
465 		repl = 0;
466 		rdev = conf->mirrors[dev].rdev;
467 	}
468 	/*
469 	 * this branch is our 'one mirror IO has finished' event handler:
470 	 */
471 	if (bio->bi_error && !discard_error) {
472 		if (repl)
473 			/* Never record new bad blocks to replacement,
474 			 * just fail it.
475 			 */
476 			md_error(rdev->mddev, rdev);
477 		else {
478 			set_bit(WriteErrorSeen,	&rdev->flags);
479 			if (!test_and_set_bit(WantReplacement, &rdev->flags))
480 				set_bit(MD_RECOVERY_NEEDED,
481 					&rdev->mddev->recovery);
482 
483 			dec_rdev = 0;
484 			if (test_bit(FailFast, &rdev->flags) &&
485 			    (bio->bi_opf & MD_FAILFAST)) {
486 				md_error(rdev->mddev, rdev);
487 				if (!test_bit(Faulty, &rdev->flags))
488 					/* This is the only remaining device,
489 					 * We need to retry the write without
490 					 * FailFast
491 					 */
492 					set_bit(R10BIO_WriteError, &r10_bio->state);
493 				else {
494 					r10_bio->devs[slot].bio = NULL;
495 					to_put = bio;
496 					dec_rdev = 1;
497 				}
498 			} else
499 				set_bit(R10BIO_WriteError, &r10_bio->state);
500 		}
501 	} else {
502 		/*
503 		 * Set R10BIO_Uptodate in our master bio, so that
504 		 * we will return a good error code for to the higher
505 		 * levels even if IO on some other mirrored buffer fails.
506 		 *
507 		 * The 'master' represents the composite IO operation to
508 		 * user-side. So if something waits for IO, then it will
509 		 * wait for the 'master' bio.
510 		 */
511 		sector_t first_bad;
512 		int bad_sectors;
513 
514 		/*
515 		 * Do not set R10BIO_Uptodate if the current device is
516 		 * rebuilding or Faulty. This is because we cannot use
517 		 * such device for properly reading the data back (we could
518 		 * potentially use it, if the current write would have felt
519 		 * before rdev->recovery_offset, but for simplicity we don't
520 		 * check this here.
521 		 */
522 		if (test_bit(In_sync, &rdev->flags) &&
523 		    !test_bit(Faulty, &rdev->flags))
524 			set_bit(R10BIO_Uptodate, &r10_bio->state);
525 
526 		/* Maybe we can clear some bad blocks. */
527 		if (is_badblock(rdev,
528 				r10_bio->devs[slot].addr,
529 				r10_bio->sectors,
530 				&first_bad, &bad_sectors) && !discard_error) {
531 			bio_put(bio);
532 			if (repl)
533 				r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
534 			else
535 				r10_bio->devs[slot].bio = IO_MADE_GOOD;
536 			dec_rdev = 0;
537 			set_bit(R10BIO_MadeGood, &r10_bio->state);
538 		}
539 	}
540 
541 	/*
542 	 *
543 	 * Let's see if all mirrored write operations have finished
544 	 * already.
545 	 */
546 	one_write_done(r10_bio);
547 	if (dec_rdev)
548 		rdev_dec_pending(rdev, conf->mddev);
549 	if (to_put)
550 		bio_put(to_put);
551 }
552 
553 /*
554  * RAID10 layout manager
555  * As well as the chunksize and raid_disks count, there are two
556  * parameters: near_copies and far_copies.
557  * near_copies * far_copies must be <= raid_disks.
558  * Normally one of these will be 1.
559  * If both are 1, we get raid0.
560  * If near_copies == raid_disks, we get raid1.
561  *
562  * Chunks are laid out in raid0 style with near_copies copies of the
563  * first chunk, followed by near_copies copies of the next chunk and
564  * so on.
565  * If far_copies > 1, then after 1/far_copies of the array has been assigned
566  * as described above, we start again with a device offset of near_copies.
567  * So we effectively have another copy of the whole array further down all
568  * the drives, but with blocks on different drives.
569  * With this layout, and block is never stored twice on the one device.
570  *
571  * raid10_find_phys finds the sector offset of a given virtual sector
572  * on each device that it is on.
573  *
574  * raid10_find_virt does the reverse mapping, from a device and a
575  * sector offset to a virtual address
576  */
577 
578 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
579 {
580 	int n,f;
581 	sector_t sector;
582 	sector_t chunk;
583 	sector_t stripe;
584 	int dev;
585 	int slot = 0;
586 	int last_far_set_start, last_far_set_size;
587 
588 	last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
589 	last_far_set_start *= geo->far_set_size;
590 
591 	last_far_set_size = geo->far_set_size;
592 	last_far_set_size += (geo->raid_disks % geo->far_set_size);
593 
594 	/* now calculate first sector/dev */
595 	chunk = r10bio->sector >> geo->chunk_shift;
596 	sector = r10bio->sector & geo->chunk_mask;
597 
598 	chunk *= geo->near_copies;
599 	stripe = chunk;
600 	dev = sector_div(stripe, geo->raid_disks);
601 	if (geo->far_offset)
602 		stripe *= geo->far_copies;
603 
604 	sector += stripe << geo->chunk_shift;
605 
606 	/* and calculate all the others */
607 	for (n = 0; n < geo->near_copies; n++) {
608 		int d = dev;
609 		int set;
610 		sector_t s = sector;
611 		r10bio->devs[slot].devnum = d;
612 		r10bio->devs[slot].addr = s;
613 		slot++;
614 
615 		for (f = 1; f < geo->far_copies; f++) {
616 			set = d / geo->far_set_size;
617 			d += geo->near_copies;
618 
619 			if ((geo->raid_disks % geo->far_set_size) &&
620 			    (d > last_far_set_start)) {
621 				d -= last_far_set_start;
622 				d %= last_far_set_size;
623 				d += last_far_set_start;
624 			} else {
625 				d %= geo->far_set_size;
626 				d += geo->far_set_size * set;
627 			}
628 			s += geo->stride;
629 			r10bio->devs[slot].devnum = d;
630 			r10bio->devs[slot].addr = s;
631 			slot++;
632 		}
633 		dev++;
634 		if (dev >= geo->raid_disks) {
635 			dev = 0;
636 			sector += (geo->chunk_mask + 1);
637 		}
638 	}
639 }
640 
641 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
642 {
643 	struct geom *geo = &conf->geo;
644 
645 	if (conf->reshape_progress != MaxSector &&
646 	    ((r10bio->sector >= conf->reshape_progress) !=
647 	     conf->mddev->reshape_backwards)) {
648 		set_bit(R10BIO_Previous, &r10bio->state);
649 		geo = &conf->prev;
650 	} else
651 		clear_bit(R10BIO_Previous, &r10bio->state);
652 
653 	__raid10_find_phys(geo, r10bio);
654 }
655 
656 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
657 {
658 	sector_t offset, chunk, vchunk;
659 	/* Never use conf->prev as this is only called during resync
660 	 * or recovery, so reshape isn't happening
661 	 */
662 	struct geom *geo = &conf->geo;
663 	int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
664 	int far_set_size = geo->far_set_size;
665 	int last_far_set_start;
666 
667 	if (geo->raid_disks % geo->far_set_size) {
668 		last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
669 		last_far_set_start *= geo->far_set_size;
670 
671 		if (dev >= last_far_set_start) {
672 			far_set_size = geo->far_set_size;
673 			far_set_size += (geo->raid_disks % geo->far_set_size);
674 			far_set_start = last_far_set_start;
675 		}
676 	}
677 
678 	offset = sector & geo->chunk_mask;
679 	if (geo->far_offset) {
680 		int fc;
681 		chunk = sector >> geo->chunk_shift;
682 		fc = sector_div(chunk, geo->far_copies);
683 		dev -= fc * geo->near_copies;
684 		if (dev < far_set_start)
685 			dev += far_set_size;
686 	} else {
687 		while (sector >= geo->stride) {
688 			sector -= geo->stride;
689 			if (dev < (geo->near_copies + far_set_start))
690 				dev += far_set_size - geo->near_copies;
691 			else
692 				dev -= geo->near_copies;
693 		}
694 		chunk = sector >> geo->chunk_shift;
695 	}
696 	vchunk = chunk * geo->raid_disks + dev;
697 	sector_div(vchunk, geo->near_copies);
698 	return (vchunk << geo->chunk_shift) + offset;
699 }
700 
701 /*
702  * This routine returns the disk from which the requested read should
703  * be done. There is a per-array 'next expected sequential IO' sector
704  * number - if this matches on the next IO then we use the last disk.
705  * There is also a per-disk 'last know head position' sector that is
706  * maintained from IRQ contexts, both the normal and the resync IO
707  * completion handlers update this position correctly. If there is no
708  * perfect sequential match then we pick the disk whose head is closest.
709  *
710  * If there are 2 mirrors in the same 2 devices, performance degrades
711  * because position is mirror, not device based.
712  *
713  * The rdev for the device selected will have nr_pending incremented.
714  */
715 
716 /*
717  * FIXME: possibly should rethink readbalancing and do it differently
718  * depending on near_copies / far_copies geometry.
719  */
720 static struct md_rdev *read_balance(struct r10conf *conf,
721 				    struct r10bio *r10_bio,
722 				    int *max_sectors)
723 {
724 	const sector_t this_sector = r10_bio->sector;
725 	int disk, slot;
726 	int sectors = r10_bio->sectors;
727 	int best_good_sectors;
728 	sector_t new_distance, best_dist;
729 	struct md_rdev *best_rdev, *rdev = NULL;
730 	int do_balance;
731 	int best_slot;
732 	struct geom *geo = &conf->geo;
733 
734 	raid10_find_phys(conf, r10_bio);
735 	rcu_read_lock();
736 	sectors = r10_bio->sectors;
737 	best_slot = -1;
738 	best_rdev = NULL;
739 	best_dist = MaxSector;
740 	best_good_sectors = 0;
741 	do_balance = 1;
742 	clear_bit(R10BIO_FailFast, &r10_bio->state);
743 	/*
744 	 * Check if we can balance. We can balance on the whole
745 	 * device if no resync is going on (recovery is ok), or below
746 	 * the resync window. We take the first readable disk when
747 	 * above the resync window.
748 	 */
749 	if (conf->mddev->recovery_cp < MaxSector
750 	    && (this_sector + sectors >= conf->next_resync))
751 		do_balance = 0;
752 
753 	for (slot = 0; slot < conf->copies ; slot++) {
754 		sector_t first_bad;
755 		int bad_sectors;
756 		sector_t dev_sector;
757 
758 		if (r10_bio->devs[slot].bio == IO_BLOCKED)
759 			continue;
760 		disk = r10_bio->devs[slot].devnum;
761 		rdev = rcu_dereference(conf->mirrors[disk].replacement);
762 		if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
763 		    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
764 			rdev = rcu_dereference(conf->mirrors[disk].rdev);
765 		if (rdev == NULL ||
766 		    test_bit(Faulty, &rdev->flags))
767 			continue;
768 		if (!test_bit(In_sync, &rdev->flags) &&
769 		    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
770 			continue;
771 
772 		dev_sector = r10_bio->devs[slot].addr;
773 		if (is_badblock(rdev, dev_sector, sectors,
774 				&first_bad, &bad_sectors)) {
775 			if (best_dist < MaxSector)
776 				/* Already have a better slot */
777 				continue;
778 			if (first_bad <= dev_sector) {
779 				/* Cannot read here.  If this is the
780 				 * 'primary' device, then we must not read
781 				 * beyond 'bad_sectors' from another device.
782 				 */
783 				bad_sectors -= (dev_sector - first_bad);
784 				if (!do_balance && sectors > bad_sectors)
785 					sectors = bad_sectors;
786 				if (best_good_sectors > sectors)
787 					best_good_sectors = sectors;
788 			} else {
789 				sector_t good_sectors =
790 					first_bad - dev_sector;
791 				if (good_sectors > best_good_sectors) {
792 					best_good_sectors = good_sectors;
793 					best_slot = slot;
794 					best_rdev = rdev;
795 				}
796 				if (!do_balance)
797 					/* Must read from here */
798 					break;
799 			}
800 			continue;
801 		} else
802 			best_good_sectors = sectors;
803 
804 		if (!do_balance)
805 			break;
806 
807 		if (best_slot >= 0)
808 			/* At least 2 disks to choose from so failfast is OK */
809 			set_bit(R10BIO_FailFast, &r10_bio->state);
810 		/* This optimisation is debatable, and completely destroys
811 		 * sequential read speed for 'far copies' arrays.  So only
812 		 * keep it for 'near' arrays, and review those later.
813 		 */
814 		if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
815 			new_distance = 0;
816 
817 		/* for far > 1 always use the lowest address */
818 		else if (geo->far_copies > 1)
819 			new_distance = r10_bio->devs[slot].addr;
820 		else
821 			new_distance = abs(r10_bio->devs[slot].addr -
822 					   conf->mirrors[disk].head_position);
823 		if (new_distance < best_dist) {
824 			best_dist = new_distance;
825 			best_slot = slot;
826 			best_rdev = rdev;
827 		}
828 	}
829 	if (slot >= conf->copies) {
830 		slot = best_slot;
831 		rdev = best_rdev;
832 	}
833 
834 	if (slot >= 0) {
835 		atomic_inc(&rdev->nr_pending);
836 		r10_bio->read_slot = slot;
837 	} else
838 		rdev = NULL;
839 	rcu_read_unlock();
840 	*max_sectors = best_good_sectors;
841 
842 	return rdev;
843 }
844 
845 static int raid10_congested(struct mddev *mddev, int bits)
846 {
847 	struct r10conf *conf = mddev->private;
848 	int i, ret = 0;
849 
850 	if ((bits & (1 << WB_async_congested)) &&
851 	    conf->pending_count >= max_queued_requests)
852 		return 1;
853 
854 	rcu_read_lock();
855 	for (i = 0;
856 	     (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
857 		     && ret == 0;
858 	     i++) {
859 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
860 		if (rdev && !test_bit(Faulty, &rdev->flags)) {
861 			struct request_queue *q = bdev_get_queue(rdev->bdev);
862 
863 			ret |= bdi_congested(&q->backing_dev_info, bits);
864 		}
865 	}
866 	rcu_read_unlock();
867 	return ret;
868 }
869 
870 static void flush_pending_writes(struct r10conf *conf)
871 {
872 	/* Any writes that have been queued but are awaiting
873 	 * bitmap updates get flushed here.
874 	 */
875 	spin_lock_irq(&conf->device_lock);
876 
877 	if (conf->pending_bio_list.head) {
878 		struct bio *bio;
879 		bio = bio_list_get(&conf->pending_bio_list);
880 		conf->pending_count = 0;
881 		spin_unlock_irq(&conf->device_lock);
882 		/* flush any pending bitmap writes to disk
883 		 * before proceeding w/ I/O */
884 		bitmap_unplug(conf->mddev->bitmap);
885 		wake_up(&conf->wait_barrier);
886 
887 		while (bio) { /* submit pending writes */
888 			struct bio *next = bio->bi_next;
889 			struct md_rdev *rdev = (void*)bio->bi_bdev;
890 			bio->bi_next = NULL;
891 			bio->bi_bdev = rdev->bdev;
892 			if (test_bit(Faulty, &rdev->flags)) {
893 				bio->bi_error = -EIO;
894 				bio_endio(bio);
895 			} else if (unlikely((bio_op(bio) ==  REQ_OP_DISCARD) &&
896 					    !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
897 				/* Just ignore it */
898 				bio_endio(bio);
899 			else
900 				generic_make_request(bio);
901 			bio = next;
902 		}
903 	} else
904 		spin_unlock_irq(&conf->device_lock);
905 }
906 
907 /* Barriers....
908  * Sometimes we need to suspend IO while we do something else,
909  * either some resync/recovery, or reconfigure the array.
910  * To do this we raise a 'barrier'.
911  * The 'barrier' is a counter that can be raised multiple times
912  * to count how many activities are happening which preclude
913  * normal IO.
914  * We can only raise the barrier if there is no pending IO.
915  * i.e. if nr_pending == 0.
916  * We choose only to raise the barrier if no-one is waiting for the
917  * barrier to go down.  This means that as soon as an IO request
918  * is ready, no other operations which require a barrier will start
919  * until the IO request has had a chance.
920  *
921  * So: regular IO calls 'wait_barrier'.  When that returns there
922  *    is no backgroup IO happening,  It must arrange to call
923  *    allow_barrier when it has finished its IO.
924  * backgroup IO calls must call raise_barrier.  Once that returns
925  *    there is no normal IO happeing.  It must arrange to call
926  *    lower_barrier when the particular background IO completes.
927  */
928 
929 static void raise_barrier(struct r10conf *conf, int force)
930 {
931 	BUG_ON(force && !conf->barrier);
932 	spin_lock_irq(&conf->resync_lock);
933 
934 	/* Wait until no block IO is waiting (unless 'force') */
935 	wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
936 			    conf->resync_lock);
937 
938 	/* block any new IO from starting */
939 	conf->barrier++;
940 
941 	/* Now wait for all pending IO to complete */
942 	wait_event_lock_irq(conf->wait_barrier,
943 			    !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
944 			    conf->resync_lock);
945 
946 	spin_unlock_irq(&conf->resync_lock);
947 }
948 
949 static void lower_barrier(struct r10conf *conf)
950 {
951 	unsigned long flags;
952 	spin_lock_irqsave(&conf->resync_lock, flags);
953 	conf->barrier--;
954 	spin_unlock_irqrestore(&conf->resync_lock, flags);
955 	wake_up(&conf->wait_barrier);
956 }
957 
958 static void wait_barrier(struct r10conf *conf)
959 {
960 	spin_lock_irq(&conf->resync_lock);
961 	if (conf->barrier) {
962 		conf->nr_waiting++;
963 		/* Wait for the barrier to drop.
964 		 * However if there are already pending
965 		 * requests (preventing the barrier from
966 		 * rising completely), and the
967 		 * pre-process bio queue isn't empty,
968 		 * then don't wait, as we need to empty
969 		 * that queue to get the nr_pending
970 		 * count down.
971 		 */
972 		raid10_log(conf->mddev, "wait barrier");
973 		wait_event_lock_irq(conf->wait_barrier,
974 				    !conf->barrier ||
975 				    (atomic_read(&conf->nr_pending) &&
976 				     current->bio_list &&
977 				     !bio_list_empty(current->bio_list)),
978 				    conf->resync_lock);
979 		conf->nr_waiting--;
980 		if (!conf->nr_waiting)
981 			wake_up(&conf->wait_barrier);
982 	}
983 	atomic_inc(&conf->nr_pending);
984 	spin_unlock_irq(&conf->resync_lock);
985 }
986 
987 static void allow_barrier(struct r10conf *conf)
988 {
989 	if ((atomic_dec_and_test(&conf->nr_pending)) ||
990 			(conf->array_freeze_pending))
991 		wake_up(&conf->wait_barrier);
992 }
993 
994 static void freeze_array(struct r10conf *conf, int extra)
995 {
996 	/* stop syncio and normal IO and wait for everything to
997 	 * go quiet.
998 	 * We increment barrier and nr_waiting, and then
999 	 * wait until nr_pending match nr_queued+extra
1000 	 * This is called in the context of one normal IO request
1001 	 * that has failed. Thus any sync request that might be pending
1002 	 * will be blocked by nr_pending, and we need to wait for
1003 	 * pending IO requests to complete or be queued for re-try.
1004 	 * Thus the number queued (nr_queued) plus this request (extra)
1005 	 * must match the number of pending IOs (nr_pending) before
1006 	 * we continue.
1007 	 */
1008 	spin_lock_irq(&conf->resync_lock);
1009 	conf->array_freeze_pending++;
1010 	conf->barrier++;
1011 	conf->nr_waiting++;
1012 	wait_event_lock_irq_cmd(conf->wait_barrier,
1013 				atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
1014 				conf->resync_lock,
1015 				flush_pending_writes(conf));
1016 
1017 	conf->array_freeze_pending--;
1018 	spin_unlock_irq(&conf->resync_lock);
1019 }
1020 
1021 static void unfreeze_array(struct r10conf *conf)
1022 {
1023 	/* reverse the effect of the freeze */
1024 	spin_lock_irq(&conf->resync_lock);
1025 	conf->barrier--;
1026 	conf->nr_waiting--;
1027 	wake_up(&conf->wait_barrier);
1028 	spin_unlock_irq(&conf->resync_lock);
1029 }
1030 
1031 static sector_t choose_data_offset(struct r10bio *r10_bio,
1032 				   struct md_rdev *rdev)
1033 {
1034 	if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1035 	    test_bit(R10BIO_Previous, &r10_bio->state))
1036 		return rdev->data_offset;
1037 	else
1038 		return rdev->new_data_offset;
1039 }
1040 
1041 struct raid10_plug_cb {
1042 	struct blk_plug_cb	cb;
1043 	struct bio_list		pending;
1044 	int			pending_cnt;
1045 };
1046 
1047 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1048 {
1049 	struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1050 						   cb);
1051 	struct mddev *mddev = plug->cb.data;
1052 	struct r10conf *conf = mddev->private;
1053 	struct bio *bio;
1054 
1055 	if (from_schedule || current->bio_list) {
1056 		spin_lock_irq(&conf->device_lock);
1057 		bio_list_merge(&conf->pending_bio_list, &plug->pending);
1058 		conf->pending_count += plug->pending_cnt;
1059 		spin_unlock_irq(&conf->device_lock);
1060 		wake_up(&conf->wait_barrier);
1061 		md_wakeup_thread(mddev->thread);
1062 		kfree(plug);
1063 		return;
1064 	}
1065 
1066 	/* we aren't scheduling, so we can do the write-out directly. */
1067 	bio = bio_list_get(&plug->pending);
1068 	bitmap_unplug(mddev->bitmap);
1069 	wake_up(&conf->wait_barrier);
1070 
1071 	while (bio) { /* submit pending writes */
1072 		struct bio *next = bio->bi_next;
1073 		struct md_rdev *rdev = (void*)bio->bi_bdev;
1074 		bio->bi_next = NULL;
1075 		bio->bi_bdev = rdev->bdev;
1076 		if (test_bit(Faulty, &rdev->flags)) {
1077 			bio->bi_error = -EIO;
1078 			bio_endio(bio);
1079 		} else if (unlikely((bio_op(bio) ==  REQ_OP_DISCARD) &&
1080 				    !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1081 			/* Just ignore it */
1082 			bio_endio(bio);
1083 		else
1084 			generic_make_request(bio);
1085 		bio = next;
1086 	}
1087 	kfree(plug);
1088 }
1089 
1090 static void __make_request(struct mddev *mddev, struct bio *bio)
1091 {
1092 	struct r10conf *conf = mddev->private;
1093 	struct r10bio *r10_bio;
1094 	struct bio *read_bio;
1095 	int i;
1096 	const int op = bio_op(bio);
1097 	const int rw = bio_data_dir(bio);
1098 	const unsigned long do_sync = (bio->bi_opf & REQ_SYNC);
1099 	const unsigned long do_fua = (bio->bi_opf & REQ_FUA);
1100 	unsigned long flags;
1101 	struct md_rdev *blocked_rdev;
1102 	struct blk_plug_cb *cb;
1103 	struct raid10_plug_cb *plug = NULL;
1104 	int sectors_handled;
1105 	int max_sectors;
1106 	int sectors;
1107 
1108 	md_write_start(mddev, bio);
1109 
1110 	/*
1111 	 * Register the new request and wait if the reconstruction
1112 	 * thread has put up a bar for new requests.
1113 	 * Continue immediately if no resync is active currently.
1114 	 */
1115 	wait_barrier(conf);
1116 
1117 	sectors = bio_sectors(bio);
1118 	while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1119 	    bio->bi_iter.bi_sector < conf->reshape_progress &&
1120 	    bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1121 		/* IO spans the reshape position.  Need to wait for
1122 		 * reshape to pass
1123 		 */
1124 		raid10_log(conf->mddev, "wait reshape");
1125 		allow_barrier(conf);
1126 		wait_event(conf->wait_barrier,
1127 			   conf->reshape_progress <= bio->bi_iter.bi_sector ||
1128 			   conf->reshape_progress >= bio->bi_iter.bi_sector +
1129 			   sectors);
1130 		wait_barrier(conf);
1131 	}
1132 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1133 	    bio_data_dir(bio) == WRITE &&
1134 	    (mddev->reshape_backwards
1135 	     ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1136 		bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1137 	     : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1138 		bio->bi_iter.bi_sector < conf->reshape_progress))) {
1139 		/* Need to update reshape_position in metadata */
1140 		mddev->reshape_position = conf->reshape_progress;
1141 		set_mask_bits(&mddev->sb_flags, 0,
1142 			      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1143 		md_wakeup_thread(mddev->thread);
1144 		raid10_log(conf->mddev, "wait reshape metadata");
1145 		wait_event(mddev->sb_wait,
1146 			   !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1147 
1148 		conf->reshape_safe = mddev->reshape_position;
1149 	}
1150 
1151 	r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1152 
1153 	r10_bio->master_bio = bio;
1154 	r10_bio->sectors = sectors;
1155 
1156 	r10_bio->mddev = mddev;
1157 	r10_bio->sector = bio->bi_iter.bi_sector;
1158 	r10_bio->state = 0;
1159 
1160 	/* We might need to issue multiple reads to different
1161 	 * devices if there are bad blocks around, so we keep
1162 	 * track of the number of reads in bio->bi_phys_segments.
1163 	 * If this is 0, there is only one r10_bio and no locking
1164 	 * will be needed when the request completes.  If it is
1165 	 * non-zero, then it is the number of not-completed requests.
1166 	 */
1167 	bio->bi_phys_segments = 0;
1168 	bio_clear_flag(bio, BIO_SEG_VALID);
1169 
1170 	if (rw == READ) {
1171 		/*
1172 		 * read balancing logic:
1173 		 */
1174 		struct md_rdev *rdev;
1175 		int slot;
1176 
1177 read_again:
1178 		rdev = read_balance(conf, r10_bio, &max_sectors);
1179 		if (!rdev) {
1180 			raid_end_bio_io(r10_bio);
1181 			return;
1182 		}
1183 		slot = r10_bio->read_slot;
1184 
1185 		read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1186 		bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1187 			 max_sectors);
1188 
1189 		r10_bio->devs[slot].bio = read_bio;
1190 		r10_bio->devs[slot].rdev = rdev;
1191 
1192 		read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1193 			choose_data_offset(r10_bio, rdev);
1194 		read_bio->bi_bdev = rdev->bdev;
1195 		read_bio->bi_end_io = raid10_end_read_request;
1196 		bio_set_op_attrs(read_bio, op, do_sync);
1197 		if (test_bit(FailFast, &rdev->flags) &&
1198 		    test_bit(R10BIO_FailFast, &r10_bio->state))
1199 			read_bio->bi_opf |= MD_FAILFAST;
1200 		read_bio->bi_private = r10_bio;
1201 
1202 		if (mddev->gendisk)
1203 			trace_block_bio_remap(bdev_get_queue(read_bio->bi_bdev),
1204 					      read_bio, disk_devt(mddev->gendisk),
1205 					      r10_bio->sector);
1206 		if (max_sectors < r10_bio->sectors) {
1207 			/* Could not read all from this device, so we will
1208 			 * need another r10_bio.
1209 			 */
1210 			sectors_handled = (r10_bio->sector + max_sectors
1211 					   - bio->bi_iter.bi_sector);
1212 			r10_bio->sectors = max_sectors;
1213 			spin_lock_irq(&conf->device_lock);
1214 			if (bio->bi_phys_segments == 0)
1215 				bio->bi_phys_segments = 2;
1216 			else
1217 				bio->bi_phys_segments++;
1218 			spin_unlock_irq(&conf->device_lock);
1219 			/* Cannot call generic_make_request directly
1220 			 * as that will be queued in __generic_make_request
1221 			 * and subsequent mempool_alloc might block
1222 			 * waiting for it.  so hand bio over to raid10d.
1223 			 */
1224 			reschedule_retry(r10_bio);
1225 
1226 			r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1227 
1228 			r10_bio->master_bio = bio;
1229 			r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1230 			r10_bio->state = 0;
1231 			r10_bio->mddev = mddev;
1232 			r10_bio->sector = bio->bi_iter.bi_sector +
1233 				sectors_handled;
1234 			goto read_again;
1235 		} else
1236 			generic_make_request(read_bio);
1237 		return;
1238 	}
1239 
1240 	/*
1241 	 * WRITE:
1242 	 */
1243 	if (conf->pending_count >= max_queued_requests) {
1244 		md_wakeup_thread(mddev->thread);
1245 		raid10_log(mddev, "wait queued");
1246 		wait_event(conf->wait_barrier,
1247 			   conf->pending_count < max_queued_requests);
1248 	}
1249 	/* first select target devices under rcu_lock and
1250 	 * inc refcount on their rdev.  Record them by setting
1251 	 * bios[x] to bio
1252 	 * If there are known/acknowledged bad blocks on any device
1253 	 * on which we have seen a write error, we want to avoid
1254 	 * writing to those blocks.  This potentially requires several
1255 	 * writes to write around the bad blocks.  Each set of writes
1256 	 * gets its own r10_bio with a set of bios attached.  The number
1257 	 * of r10_bios is recored in bio->bi_phys_segments just as with
1258 	 * the read case.
1259 	 */
1260 
1261 	r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1262 	raid10_find_phys(conf, r10_bio);
1263 retry_write:
1264 	blocked_rdev = NULL;
1265 	rcu_read_lock();
1266 	max_sectors = r10_bio->sectors;
1267 
1268 	for (i = 0;  i < conf->copies; i++) {
1269 		int d = r10_bio->devs[i].devnum;
1270 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1271 		struct md_rdev *rrdev = rcu_dereference(
1272 			conf->mirrors[d].replacement);
1273 		if (rdev == rrdev)
1274 			rrdev = NULL;
1275 		if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1276 			atomic_inc(&rdev->nr_pending);
1277 			blocked_rdev = rdev;
1278 			break;
1279 		}
1280 		if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1281 			atomic_inc(&rrdev->nr_pending);
1282 			blocked_rdev = rrdev;
1283 			break;
1284 		}
1285 		if (rdev && (test_bit(Faulty, &rdev->flags)))
1286 			rdev = NULL;
1287 		if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1288 			rrdev = NULL;
1289 
1290 		r10_bio->devs[i].bio = NULL;
1291 		r10_bio->devs[i].repl_bio = NULL;
1292 
1293 		if (!rdev && !rrdev) {
1294 			set_bit(R10BIO_Degraded, &r10_bio->state);
1295 			continue;
1296 		}
1297 		if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1298 			sector_t first_bad;
1299 			sector_t dev_sector = r10_bio->devs[i].addr;
1300 			int bad_sectors;
1301 			int is_bad;
1302 
1303 			is_bad = is_badblock(rdev, dev_sector,
1304 					     max_sectors,
1305 					     &first_bad, &bad_sectors);
1306 			if (is_bad < 0) {
1307 				/* Mustn't write here until the bad block
1308 				 * is acknowledged
1309 				 */
1310 				atomic_inc(&rdev->nr_pending);
1311 				set_bit(BlockedBadBlocks, &rdev->flags);
1312 				blocked_rdev = rdev;
1313 				break;
1314 			}
1315 			if (is_bad && first_bad <= dev_sector) {
1316 				/* Cannot write here at all */
1317 				bad_sectors -= (dev_sector - first_bad);
1318 				if (bad_sectors < max_sectors)
1319 					/* Mustn't write more than bad_sectors
1320 					 * to other devices yet
1321 					 */
1322 					max_sectors = bad_sectors;
1323 				/* We don't set R10BIO_Degraded as that
1324 				 * only applies if the disk is missing,
1325 				 * so it might be re-added, and we want to
1326 				 * know to recover this chunk.
1327 				 * In this case the device is here, and the
1328 				 * fact that this chunk is not in-sync is
1329 				 * recorded in the bad block log.
1330 				 */
1331 				continue;
1332 			}
1333 			if (is_bad) {
1334 				int good_sectors = first_bad - dev_sector;
1335 				if (good_sectors < max_sectors)
1336 					max_sectors = good_sectors;
1337 			}
1338 		}
1339 		if (rdev) {
1340 			r10_bio->devs[i].bio = bio;
1341 			atomic_inc(&rdev->nr_pending);
1342 		}
1343 		if (rrdev) {
1344 			r10_bio->devs[i].repl_bio = bio;
1345 			atomic_inc(&rrdev->nr_pending);
1346 		}
1347 	}
1348 	rcu_read_unlock();
1349 
1350 	if (unlikely(blocked_rdev)) {
1351 		/* Have to wait for this device to get unblocked, then retry */
1352 		int j;
1353 		int d;
1354 
1355 		for (j = 0; j < i; j++) {
1356 			if (r10_bio->devs[j].bio) {
1357 				d = r10_bio->devs[j].devnum;
1358 				rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1359 			}
1360 			if (r10_bio->devs[j].repl_bio) {
1361 				struct md_rdev *rdev;
1362 				d = r10_bio->devs[j].devnum;
1363 				rdev = conf->mirrors[d].replacement;
1364 				if (!rdev) {
1365 					/* Race with remove_disk */
1366 					smp_mb();
1367 					rdev = conf->mirrors[d].rdev;
1368 				}
1369 				rdev_dec_pending(rdev, mddev);
1370 			}
1371 		}
1372 		allow_barrier(conf);
1373 		raid10_log(conf->mddev, "wait rdev %d blocked", blocked_rdev->raid_disk);
1374 		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1375 		wait_barrier(conf);
1376 		goto retry_write;
1377 	}
1378 
1379 	if (max_sectors < r10_bio->sectors) {
1380 		/* We are splitting this into multiple parts, so
1381 		 * we need to prepare for allocating another r10_bio.
1382 		 */
1383 		r10_bio->sectors = max_sectors;
1384 		spin_lock_irq(&conf->device_lock);
1385 		if (bio->bi_phys_segments == 0)
1386 			bio->bi_phys_segments = 2;
1387 		else
1388 			bio->bi_phys_segments++;
1389 		spin_unlock_irq(&conf->device_lock);
1390 	}
1391 	sectors_handled = r10_bio->sector + max_sectors -
1392 		bio->bi_iter.bi_sector;
1393 
1394 	atomic_set(&r10_bio->remaining, 1);
1395 	bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1396 
1397 	for (i = 0; i < conf->copies; i++) {
1398 		struct bio *mbio;
1399 		int d = r10_bio->devs[i].devnum;
1400 		if (r10_bio->devs[i].bio) {
1401 			struct md_rdev *rdev = conf->mirrors[d].rdev;
1402 			mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1403 			bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1404 				 max_sectors);
1405 			r10_bio->devs[i].bio = mbio;
1406 
1407 			mbio->bi_iter.bi_sector	= (r10_bio->devs[i].addr+
1408 					   choose_data_offset(r10_bio,
1409 							      rdev));
1410 			mbio->bi_bdev = rdev->bdev;
1411 			mbio->bi_end_io	= raid10_end_write_request;
1412 			bio_set_op_attrs(mbio, op, do_sync | do_fua);
1413 			if (test_bit(FailFast, &conf->mirrors[d].rdev->flags) &&
1414 			    enough(conf, d))
1415 				mbio->bi_opf |= MD_FAILFAST;
1416 			mbio->bi_private = r10_bio;
1417 
1418 			if (conf->mddev->gendisk)
1419 				trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1420 						      mbio, disk_devt(conf->mddev->gendisk),
1421 						      r10_bio->sector);
1422 			/* flush_pending_writes() needs access to the rdev so...*/
1423 			mbio->bi_bdev = (void*)rdev;
1424 
1425 			atomic_inc(&r10_bio->remaining);
1426 
1427 			cb = blk_check_plugged(raid10_unplug, mddev,
1428 					       sizeof(*plug));
1429 			if (cb)
1430 				plug = container_of(cb, struct raid10_plug_cb,
1431 						    cb);
1432 			else
1433 				plug = NULL;
1434 			spin_lock_irqsave(&conf->device_lock, flags);
1435 			if (plug) {
1436 				bio_list_add(&plug->pending, mbio);
1437 				plug->pending_cnt++;
1438 			} else {
1439 				bio_list_add(&conf->pending_bio_list, mbio);
1440 				conf->pending_count++;
1441 			}
1442 			spin_unlock_irqrestore(&conf->device_lock, flags);
1443 			if (!plug)
1444 				md_wakeup_thread(mddev->thread);
1445 		}
1446 
1447 		if (r10_bio->devs[i].repl_bio) {
1448 			struct md_rdev *rdev = conf->mirrors[d].replacement;
1449 			if (rdev == NULL) {
1450 				/* Replacement just got moved to main 'rdev' */
1451 				smp_mb();
1452 				rdev = conf->mirrors[d].rdev;
1453 			}
1454 			mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1455 			bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1456 				 max_sectors);
1457 			r10_bio->devs[i].repl_bio = mbio;
1458 
1459 			mbio->bi_iter.bi_sector	= (r10_bio->devs[i].addr +
1460 					   choose_data_offset(
1461 						   r10_bio, rdev));
1462 			mbio->bi_bdev = rdev->bdev;
1463 			mbio->bi_end_io	= raid10_end_write_request;
1464 			bio_set_op_attrs(mbio, op, do_sync | do_fua);
1465 			mbio->bi_private = r10_bio;
1466 
1467 			if (conf->mddev->gendisk)
1468 				trace_block_bio_remap(bdev_get_queue(mbio->bi_bdev),
1469 						      mbio, disk_devt(conf->mddev->gendisk),
1470 						      r10_bio->sector);
1471 			/* flush_pending_writes() needs access to the rdev so...*/
1472 			mbio->bi_bdev = (void*)rdev;
1473 
1474 			atomic_inc(&r10_bio->remaining);
1475 			spin_lock_irqsave(&conf->device_lock, flags);
1476 			bio_list_add(&conf->pending_bio_list, mbio);
1477 			conf->pending_count++;
1478 			spin_unlock_irqrestore(&conf->device_lock, flags);
1479 			if (!mddev_check_plugged(mddev))
1480 				md_wakeup_thread(mddev->thread);
1481 		}
1482 	}
1483 
1484 	/* Don't remove the bias on 'remaining' (one_write_done) until
1485 	 * after checking if we need to go around again.
1486 	 */
1487 
1488 	if (sectors_handled < bio_sectors(bio)) {
1489 		one_write_done(r10_bio);
1490 		/* We need another r10_bio.  It has already been counted
1491 		 * in bio->bi_phys_segments.
1492 		 */
1493 		r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1494 
1495 		r10_bio->master_bio = bio;
1496 		r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1497 
1498 		r10_bio->mddev = mddev;
1499 		r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1500 		r10_bio->state = 0;
1501 		goto retry_write;
1502 	}
1503 	one_write_done(r10_bio);
1504 }
1505 
1506 static void raid10_make_request(struct mddev *mddev, struct bio *bio)
1507 {
1508 	struct r10conf *conf = mddev->private;
1509 	sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1510 	int chunk_sects = chunk_mask + 1;
1511 
1512 	struct bio *split;
1513 
1514 	if (unlikely(bio->bi_opf & REQ_PREFLUSH)) {
1515 		md_flush_request(mddev, bio);
1516 		return;
1517 	}
1518 
1519 	do {
1520 
1521 		/*
1522 		 * If this request crosses a chunk boundary, we need to split
1523 		 * it.
1524 		 */
1525 		if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1526 			     bio_sectors(bio) > chunk_sects
1527 			     && (conf->geo.near_copies < conf->geo.raid_disks
1528 				 || conf->prev.near_copies <
1529 				 conf->prev.raid_disks))) {
1530 			split = bio_split(bio, chunk_sects -
1531 					  (bio->bi_iter.bi_sector &
1532 					   (chunk_sects - 1)),
1533 					  GFP_NOIO, fs_bio_set);
1534 			bio_chain(split, bio);
1535 		} else {
1536 			split = bio;
1537 		}
1538 
1539 		__make_request(mddev, split);
1540 	} while (split != bio);
1541 
1542 	/* In case raid10d snuck in to freeze_array */
1543 	wake_up(&conf->wait_barrier);
1544 }
1545 
1546 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1547 {
1548 	struct r10conf *conf = mddev->private;
1549 	int i;
1550 
1551 	if (conf->geo.near_copies < conf->geo.raid_disks)
1552 		seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1553 	if (conf->geo.near_copies > 1)
1554 		seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1555 	if (conf->geo.far_copies > 1) {
1556 		if (conf->geo.far_offset)
1557 			seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1558 		else
1559 			seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1560 		if (conf->geo.far_set_size != conf->geo.raid_disks)
1561 			seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1562 	}
1563 	seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1564 					conf->geo.raid_disks - mddev->degraded);
1565 	rcu_read_lock();
1566 	for (i = 0; i < conf->geo.raid_disks; i++) {
1567 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1568 		seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1569 	}
1570 	rcu_read_unlock();
1571 	seq_printf(seq, "]");
1572 }
1573 
1574 /* check if there are enough drives for
1575  * every block to appear on atleast one.
1576  * Don't consider the device numbered 'ignore'
1577  * as we might be about to remove it.
1578  */
1579 static int _enough(struct r10conf *conf, int previous, int ignore)
1580 {
1581 	int first = 0;
1582 	int has_enough = 0;
1583 	int disks, ncopies;
1584 	if (previous) {
1585 		disks = conf->prev.raid_disks;
1586 		ncopies = conf->prev.near_copies;
1587 	} else {
1588 		disks = conf->geo.raid_disks;
1589 		ncopies = conf->geo.near_copies;
1590 	}
1591 
1592 	rcu_read_lock();
1593 	do {
1594 		int n = conf->copies;
1595 		int cnt = 0;
1596 		int this = first;
1597 		while (n--) {
1598 			struct md_rdev *rdev;
1599 			if (this != ignore &&
1600 			    (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1601 			    test_bit(In_sync, &rdev->flags))
1602 				cnt++;
1603 			this = (this+1) % disks;
1604 		}
1605 		if (cnt == 0)
1606 			goto out;
1607 		first = (first + ncopies) % disks;
1608 	} while (first != 0);
1609 	has_enough = 1;
1610 out:
1611 	rcu_read_unlock();
1612 	return has_enough;
1613 }
1614 
1615 static int enough(struct r10conf *conf, int ignore)
1616 {
1617 	/* when calling 'enough', both 'prev' and 'geo' must
1618 	 * be stable.
1619 	 * This is ensured if ->reconfig_mutex or ->device_lock
1620 	 * is held.
1621 	 */
1622 	return _enough(conf, 0, ignore) &&
1623 		_enough(conf, 1, ignore);
1624 }
1625 
1626 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1627 {
1628 	char b[BDEVNAME_SIZE];
1629 	struct r10conf *conf = mddev->private;
1630 	unsigned long flags;
1631 
1632 	/*
1633 	 * If it is not operational, then we have already marked it as dead
1634 	 * else if it is the last working disks, ignore the error, let the
1635 	 * next level up know.
1636 	 * else mark the drive as failed
1637 	 */
1638 	spin_lock_irqsave(&conf->device_lock, flags);
1639 	if (test_bit(In_sync, &rdev->flags)
1640 	    && !enough(conf, rdev->raid_disk)) {
1641 		/*
1642 		 * Don't fail the drive, just return an IO error.
1643 		 */
1644 		spin_unlock_irqrestore(&conf->device_lock, flags);
1645 		return;
1646 	}
1647 	if (test_and_clear_bit(In_sync, &rdev->flags))
1648 		mddev->degraded++;
1649 	/*
1650 	 * If recovery is running, make sure it aborts.
1651 	 */
1652 	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1653 	set_bit(Blocked, &rdev->flags);
1654 	set_bit(Faulty, &rdev->flags);
1655 	set_mask_bits(&mddev->sb_flags, 0,
1656 		      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1657 	spin_unlock_irqrestore(&conf->device_lock, flags);
1658 	pr_crit("md/raid10:%s: Disk failure on %s, disabling device.\n"
1659 		"md/raid10:%s: Operation continuing on %d devices.\n",
1660 		mdname(mddev), bdevname(rdev->bdev, b),
1661 		mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1662 }
1663 
1664 static void print_conf(struct r10conf *conf)
1665 {
1666 	int i;
1667 	struct md_rdev *rdev;
1668 
1669 	pr_debug("RAID10 conf printout:\n");
1670 	if (!conf) {
1671 		pr_debug("(!conf)\n");
1672 		return;
1673 	}
1674 	pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1675 		 conf->geo.raid_disks);
1676 
1677 	/* This is only called with ->reconfix_mutex held, so
1678 	 * rcu protection of rdev is not needed */
1679 	for (i = 0; i < conf->geo.raid_disks; i++) {
1680 		char b[BDEVNAME_SIZE];
1681 		rdev = conf->mirrors[i].rdev;
1682 		if (rdev)
1683 			pr_debug(" disk %d, wo:%d, o:%d, dev:%s\n",
1684 				 i, !test_bit(In_sync, &rdev->flags),
1685 				 !test_bit(Faulty, &rdev->flags),
1686 				 bdevname(rdev->bdev,b));
1687 	}
1688 }
1689 
1690 static void close_sync(struct r10conf *conf)
1691 {
1692 	wait_barrier(conf);
1693 	allow_barrier(conf);
1694 
1695 	mempool_destroy(conf->r10buf_pool);
1696 	conf->r10buf_pool = NULL;
1697 }
1698 
1699 static int raid10_spare_active(struct mddev *mddev)
1700 {
1701 	int i;
1702 	struct r10conf *conf = mddev->private;
1703 	struct raid10_info *tmp;
1704 	int count = 0;
1705 	unsigned long flags;
1706 
1707 	/*
1708 	 * Find all non-in_sync disks within the RAID10 configuration
1709 	 * and mark them in_sync
1710 	 */
1711 	for (i = 0; i < conf->geo.raid_disks; i++) {
1712 		tmp = conf->mirrors + i;
1713 		if (tmp->replacement
1714 		    && tmp->replacement->recovery_offset == MaxSector
1715 		    && !test_bit(Faulty, &tmp->replacement->flags)
1716 		    && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1717 			/* Replacement has just become active */
1718 			if (!tmp->rdev
1719 			    || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1720 				count++;
1721 			if (tmp->rdev) {
1722 				/* Replaced device not technically faulty,
1723 				 * but we need to be sure it gets removed
1724 				 * and never re-added.
1725 				 */
1726 				set_bit(Faulty, &tmp->rdev->flags);
1727 				sysfs_notify_dirent_safe(
1728 					tmp->rdev->sysfs_state);
1729 			}
1730 			sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1731 		} else if (tmp->rdev
1732 			   && tmp->rdev->recovery_offset == MaxSector
1733 			   && !test_bit(Faulty, &tmp->rdev->flags)
1734 			   && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1735 			count++;
1736 			sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1737 		}
1738 	}
1739 	spin_lock_irqsave(&conf->device_lock, flags);
1740 	mddev->degraded -= count;
1741 	spin_unlock_irqrestore(&conf->device_lock, flags);
1742 
1743 	print_conf(conf);
1744 	return count;
1745 }
1746 
1747 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1748 {
1749 	struct r10conf *conf = mddev->private;
1750 	int err = -EEXIST;
1751 	int mirror;
1752 	int first = 0;
1753 	int last = conf->geo.raid_disks - 1;
1754 
1755 	if (mddev->recovery_cp < MaxSector)
1756 		/* only hot-add to in-sync arrays, as recovery is
1757 		 * very different from resync
1758 		 */
1759 		return -EBUSY;
1760 	if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1761 		return -EINVAL;
1762 
1763 	if (md_integrity_add_rdev(rdev, mddev))
1764 		return -ENXIO;
1765 
1766 	if (rdev->raid_disk >= 0)
1767 		first = last = rdev->raid_disk;
1768 
1769 	if (rdev->saved_raid_disk >= first &&
1770 	    conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1771 		mirror = rdev->saved_raid_disk;
1772 	else
1773 		mirror = first;
1774 	for ( ; mirror <= last ; mirror++) {
1775 		struct raid10_info *p = &conf->mirrors[mirror];
1776 		if (p->recovery_disabled == mddev->recovery_disabled)
1777 			continue;
1778 		if (p->rdev) {
1779 			if (!test_bit(WantReplacement, &p->rdev->flags) ||
1780 			    p->replacement != NULL)
1781 				continue;
1782 			clear_bit(In_sync, &rdev->flags);
1783 			set_bit(Replacement, &rdev->flags);
1784 			rdev->raid_disk = mirror;
1785 			err = 0;
1786 			if (mddev->gendisk)
1787 				disk_stack_limits(mddev->gendisk, rdev->bdev,
1788 						  rdev->data_offset << 9);
1789 			conf->fullsync = 1;
1790 			rcu_assign_pointer(p->replacement, rdev);
1791 			break;
1792 		}
1793 
1794 		if (mddev->gendisk)
1795 			disk_stack_limits(mddev->gendisk, rdev->bdev,
1796 					  rdev->data_offset << 9);
1797 
1798 		p->head_position = 0;
1799 		p->recovery_disabled = mddev->recovery_disabled - 1;
1800 		rdev->raid_disk = mirror;
1801 		err = 0;
1802 		if (rdev->saved_raid_disk != mirror)
1803 			conf->fullsync = 1;
1804 		rcu_assign_pointer(p->rdev, rdev);
1805 		break;
1806 	}
1807 	if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1808 		queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1809 
1810 	print_conf(conf);
1811 	return err;
1812 }
1813 
1814 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1815 {
1816 	struct r10conf *conf = mddev->private;
1817 	int err = 0;
1818 	int number = rdev->raid_disk;
1819 	struct md_rdev **rdevp;
1820 	struct raid10_info *p = conf->mirrors + number;
1821 
1822 	print_conf(conf);
1823 	if (rdev == p->rdev)
1824 		rdevp = &p->rdev;
1825 	else if (rdev == p->replacement)
1826 		rdevp = &p->replacement;
1827 	else
1828 		return 0;
1829 
1830 	if (test_bit(In_sync, &rdev->flags) ||
1831 	    atomic_read(&rdev->nr_pending)) {
1832 		err = -EBUSY;
1833 		goto abort;
1834 	}
1835 	/* Only remove non-faulty devices if recovery
1836 	 * is not possible.
1837 	 */
1838 	if (!test_bit(Faulty, &rdev->flags) &&
1839 	    mddev->recovery_disabled != p->recovery_disabled &&
1840 	    (!p->replacement || p->replacement == rdev) &&
1841 	    number < conf->geo.raid_disks &&
1842 	    enough(conf, -1)) {
1843 		err = -EBUSY;
1844 		goto abort;
1845 	}
1846 	*rdevp = NULL;
1847 	if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1848 		synchronize_rcu();
1849 		if (atomic_read(&rdev->nr_pending)) {
1850 			/* lost the race, try later */
1851 			err = -EBUSY;
1852 			*rdevp = rdev;
1853 			goto abort;
1854 		}
1855 	}
1856 	if (p->replacement) {
1857 		/* We must have just cleared 'rdev' */
1858 		p->rdev = p->replacement;
1859 		clear_bit(Replacement, &p->replacement->flags);
1860 		smp_mb(); /* Make sure other CPUs may see both as identical
1861 			   * but will never see neither -- if they are careful.
1862 			   */
1863 		p->replacement = NULL;
1864 		clear_bit(WantReplacement, &rdev->flags);
1865 	} else
1866 		/* We might have just remove the Replacement as faulty
1867 		 * Clear the flag just in case
1868 		 */
1869 		clear_bit(WantReplacement, &rdev->flags);
1870 
1871 	err = md_integrity_register(mddev);
1872 
1873 abort:
1874 
1875 	print_conf(conf);
1876 	return err;
1877 }
1878 
1879 static void end_sync_read(struct bio *bio)
1880 {
1881 	struct r10bio *r10_bio = bio->bi_private;
1882 	struct r10conf *conf = r10_bio->mddev->private;
1883 	int d;
1884 
1885 	if (bio == r10_bio->master_bio) {
1886 		/* this is a reshape read */
1887 		d = r10_bio->read_slot; /* really the read dev */
1888 	} else
1889 		d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1890 
1891 	if (!bio->bi_error)
1892 		set_bit(R10BIO_Uptodate, &r10_bio->state);
1893 	else
1894 		/* The write handler will notice the lack of
1895 		 * R10BIO_Uptodate and record any errors etc
1896 		 */
1897 		atomic_add(r10_bio->sectors,
1898 			   &conf->mirrors[d].rdev->corrected_errors);
1899 
1900 	/* for reconstruct, we always reschedule after a read.
1901 	 * for resync, only after all reads
1902 	 */
1903 	rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1904 	if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1905 	    atomic_dec_and_test(&r10_bio->remaining)) {
1906 		/* we have read all the blocks,
1907 		 * do the comparison in process context in raid10d
1908 		 */
1909 		reschedule_retry(r10_bio);
1910 	}
1911 }
1912 
1913 static void end_sync_request(struct r10bio *r10_bio)
1914 {
1915 	struct mddev *mddev = r10_bio->mddev;
1916 
1917 	while (atomic_dec_and_test(&r10_bio->remaining)) {
1918 		if (r10_bio->master_bio == NULL) {
1919 			/* the primary of several recovery bios */
1920 			sector_t s = r10_bio->sectors;
1921 			if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1922 			    test_bit(R10BIO_WriteError, &r10_bio->state))
1923 				reschedule_retry(r10_bio);
1924 			else
1925 				put_buf(r10_bio);
1926 			md_done_sync(mddev, s, 1);
1927 			break;
1928 		} else {
1929 			struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1930 			if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1931 			    test_bit(R10BIO_WriteError, &r10_bio->state))
1932 				reschedule_retry(r10_bio);
1933 			else
1934 				put_buf(r10_bio);
1935 			r10_bio = r10_bio2;
1936 		}
1937 	}
1938 }
1939 
1940 static void end_sync_write(struct bio *bio)
1941 {
1942 	struct r10bio *r10_bio = bio->bi_private;
1943 	struct mddev *mddev = r10_bio->mddev;
1944 	struct r10conf *conf = mddev->private;
1945 	int d;
1946 	sector_t first_bad;
1947 	int bad_sectors;
1948 	int slot;
1949 	int repl;
1950 	struct md_rdev *rdev = NULL;
1951 
1952 	d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1953 	if (repl)
1954 		rdev = conf->mirrors[d].replacement;
1955 	else
1956 		rdev = conf->mirrors[d].rdev;
1957 
1958 	if (bio->bi_error) {
1959 		if (repl)
1960 			md_error(mddev, rdev);
1961 		else {
1962 			set_bit(WriteErrorSeen, &rdev->flags);
1963 			if (!test_and_set_bit(WantReplacement, &rdev->flags))
1964 				set_bit(MD_RECOVERY_NEEDED,
1965 					&rdev->mddev->recovery);
1966 			set_bit(R10BIO_WriteError, &r10_bio->state);
1967 		}
1968 	} else if (is_badblock(rdev,
1969 			     r10_bio->devs[slot].addr,
1970 			     r10_bio->sectors,
1971 			     &first_bad, &bad_sectors))
1972 		set_bit(R10BIO_MadeGood, &r10_bio->state);
1973 
1974 	rdev_dec_pending(rdev, mddev);
1975 
1976 	end_sync_request(r10_bio);
1977 }
1978 
1979 /*
1980  * Note: sync and recover and handled very differently for raid10
1981  * This code is for resync.
1982  * For resync, we read through virtual addresses and read all blocks.
1983  * If there is any error, we schedule a write.  The lowest numbered
1984  * drive is authoritative.
1985  * However requests come for physical address, so we need to map.
1986  * For every physical address there are raid_disks/copies virtual addresses,
1987  * which is always are least one, but is not necessarly an integer.
1988  * This means that a physical address can span multiple chunks, so we may
1989  * have to submit multiple io requests for a single sync request.
1990  */
1991 /*
1992  * We check if all blocks are in-sync and only write to blocks that
1993  * aren't in sync
1994  */
1995 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1996 {
1997 	struct r10conf *conf = mddev->private;
1998 	int i, first;
1999 	struct bio *tbio, *fbio;
2000 	int vcnt;
2001 
2002 	atomic_set(&r10_bio->remaining, 1);
2003 
2004 	/* find the first device with a block */
2005 	for (i=0; i<conf->copies; i++)
2006 		if (!r10_bio->devs[i].bio->bi_error)
2007 			break;
2008 
2009 	if (i == conf->copies)
2010 		goto done;
2011 
2012 	first = i;
2013 	fbio = r10_bio->devs[i].bio;
2014 	fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2015 	fbio->bi_iter.bi_idx = 0;
2016 
2017 	vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2018 	/* now find blocks with errors */
2019 	for (i=0 ; i < conf->copies ; i++) {
2020 		int  j, d;
2021 		struct md_rdev *rdev;
2022 
2023 		tbio = r10_bio->devs[i].bio;
2024 
2025 		if (tbio->bi_end_io != end_sync_read)
2026 			continue;
2027 		if (i == first)
2028 			continue;
2029 		d = r10_bio->devs[i].devnum;
2030 		rdev = conf->mirrors[d].rdev;
2031 		if (!r10_bio->devs[i].bio->bi_error) {
2032 			/* We know that the bi_io_vec layout is the same for
2033 			 * both 'first' and 'i', so we just compare them.
2034 			 * All vec entries are PAGE_SIZE;
2035 			 */
2036 			int sectors = r10_bio->sectors;
2037 			for (j = 0; j < vcnt; j++) {
2038 				int len = PAGE_SIZE;
2039 				if (sectors < (len / 512))
2040 					len = sectors * 512;
2041 				if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2042 					   page_address(tbio->bi_io_vec[j].bv_page),
2043 					   len))
2044 					break;
2045 				sectors -= len/512;
2046 			}
2047 			if (j == vcnt)
2048 				continue;
2049 			atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2050 			if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2051 				/* Don't fix anything. */
2052 				continue;
2053 		} else if (test_bit(FailFast, &rdev->flags)) {
2054 			/* Just give up on this device */
2055 			md_error(rdev->mddev, rdev);
2056 			continue;
2057 		}
2058 		/* Ok, we need to write this bio, either to correct an
2059 		 * inconsistency or to correct an unreadable block.
2060 		 * First we need to fixup bv_offset, bv_len and
2061 		 * bi_vecs, as the read request might have corrupted these
2062 		 */
2063 		bio_reset(tbio);
2064 
2065 		tbio->bi_vcnt = vcnt;
2066 		tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
2067 		tbio->bi_private = r10_bio;
2068 		tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2069 		tbio->bi_end_io = end_sync_write;
2070 		bio_set_op_attrs(tbio, REQ_OP_WRITE, 0);
2071 
2072 		bio_copy_data(tbio, fbio);
2073 
2074 		atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2075 		atomic_inc(&r10_bio->remaining);
2076 		md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2077 
2078 		if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2079 			tbio->bi_opf |= MD_FAILFAST;
2080 		tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2081 		tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2082 		generic_make_request(tbio);
2083 	}
2084 
2085 	/* Now write out to any replacement devices
2086 	 * that are active
2087 	 */
2088 	for (i = 0; i < conf->copies; i++) {
2089 		int d;
2090 
2091 		tbio = r10_bio->devs[i].repl_bio;
2092 		if (!tbio || !tbio->bi_end_io)
2093 			continue;
2094 		if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2095 		    && r10_bio->devs[i].bio != fbio)
2096 			bio_copy_data(tbio, fbio);
2097 		d = r10_bio->devs[i].devnum;
2098 		atomic_inc(&r10_bio->remaining);
2099 		md_sync_acct(conf->mirrors[d].replacement->bdev,
2100 			     bio_sectors(tbio));
2101 		generic_make_request(tbio);
2102 	}
2103 
2104 done:
2105 	if (atomic_dec_and_test(&r10_bio->remaining)) {
2106 		md_done_sync(mddev, r10_bio->sectors, 1);
2107 		put_buf(r10_bio);
2108 	}
2109 }
2110 
2111 /*
2112  * Now for the recovery code.
2113  * Recovery happens across physical sectors.
2114  * We recover all non-is_sync drives by finding the virtual address of
2115  * each, and then choose a working drive that also has that virt address.
2116  * There is a separate r10_bio for each non-in_sync drive.
2117  * Only the first two slots are in use. The first for reading,
2118  * The second for writing.
2119  *
2120  */
2121 static void fix_recovery_read_error(struct r10bio *r10_bio)
2122 {
2123 	/* We got a read error during recovery.
2124 	 * We repeat the read in smaller page-sized sections.
2125 	 * If a read succeeds, write it to the new device or record
2126 	 * a bad block if we cannot.
2127 	 * If a read fails, record a bad block on both old and
2128 	 * new devices.
2129 	 */
2130 	struct mddev *mddev = r10_bio->mddev;
2131 	struct r10conf *conf = mddev->private;
2132 	struct bio *bio = r10_bio->devs[0].bio;
2133 	sector_t sect = 0;
2134 	int sectors = r10_bio->sectors;
2135 	int idx = 0;
2136 	int dr = r10_bio->devs[0].devnum;
2137 	int dw = r10_bio->devs[1].devnum;
2138 
2139 	while (sectors) {
2140 		int s = sectors;
2141 		struct md_rdev *rdev;
2142 		sector_t addr;
2143 		int ok;
2144 
2145 		if (s > (PAGE_SIZE>>9))
2146 			s = PAGE_SIZE >> 9;
2147 
2148 		rdev = conf->mirrors[dr].rdev;
2149 		addr = r10_bio->devs[0].addr + sect,
2150 		ok = sync_page_io(rdev,
2151 				  addr,
2152 				  s << 9,
2153 				  bio->bi_io_vec[idx].bv_page,
2154 				  REQ_OP_READ, 0, false);
2155 		if (ok) {
2156 			rdev = conf->mirrors[dw].rdev;
2157 			addr = r10_bio->devs[1].addr + sect;
2158 			ok = sync_page_io(rdev,
2159 					  addr,
2160 					  s << 9,
2161 					  bio->bi_io_vec[idx].bv_page,
2162 					  REQ_OP_WRITE, 0, false);
2163 			if (!ok) {
2164 				set_bit(WriteErrorSeen, &rdev->flags);
2165 				if (!test_and_set_bit(WantReplacement,
2166 						      &rdev->flags))
2167 					set_bit(MD_RECOVERY_NEEDED,
2168 						&rdev->mddev->recovery);
2169 			}
2170 		}
2171 		if (!ok) {
2172 			/* We don't worry if we cannot set a bad block -
2173 			 * it really is bad so there is no loss in not
2174 			 * recording it yet
2175 			 */
2176 			rdev_set_badblocks(rdev, addr, s, 0);
2177 
2178 			if (rdev != conf->mirrors[dw].rdev) {
2179 				/* need bad block on destination too */
2180 				struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2181 				addr = r10_bio->devs[1].addr + sect;
2182 				ok = rdev_set_badblocks(rdev2, addr, s, 0);
2183 				if (!ok) {
2184 					/* just abort the recovery */
2185 					pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2186 						  mdname(mddev));
2187 
2188 					conf->mirrors[dw].recovery_disabled
2189 						= mddev->recovery_disabled;
2190 					set_bit(MD_RECOVERY_INTR,
2191 						&mddev->recovery);
2192 					break;
2193 				}
2194 			}
2195 		}
2196 
2197 		sectors -= s;
2198 		sect += s;
2199 		idx++;
2200 	}
2201 }
2202 
2203 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2204 {
2205 	struct r10conf *conf = mddev->private;
2206 	int d;
2207 	struct bio *wbio, *wbio2;
2208 
2209 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2210 		fix_recovery_read_error(r10_bio);
2211 		end_sync_request(r10_bio);
2212 		return;
2213 	}
2214 
2215 	/*
2216 	 * share the pages with the first bio
2217 	 * and submit the write request
2218 	 */
2219 	d = r10_bio->devs[1].devnum;
2220 	wbio = r10_bio->devs[1].bio;
2221 	wbio2 = r10_bio->devs[1].repl_bio;
2222 	/* Need to test wbio2->bi_end_io before we call
2223 	 * generic_make_request as if the former is NULL,
2224 	 * the latter is free to free wbio2.
2225 	 */
2226 	if (wbio2 && !wbio2->bi_end_io)
2227 		wbio2 = NULL;
2228 	if (wbio->bi_end_io) {
2229 		atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2230 		md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2231 		generic_make_request(wbio);
2232 	}
2233 	if (wbio2) {
2234 		atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2235 		md_sync_acct(conf->mirrors[d].replacement->bdev,
2236 			     bio_sectors(wbio2));
2237 		generic_make_request(wbio2);
2238 	}
2239 }
2240 
2241 /*
2242  * Used by fix_read_error() to decay the per rdev read_errors.
2243  * We halve the read error count for every hour that has elapsed
2244  * since the last recorded read error.
2245  *
2246  */
2247 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2248 {
2249 	long cur_time_mon;
2250 	unsigned long hours_since_last;
2251 	unsigned int read_errors = atomic_read(&rdev->read_errors);
2252 
2253 	cur_time_mon = ktime_get_seconds();
2254 
2255 	if (rdev->last_read_error == 0) {
2256 		/* first time we've seen a read error */
2257 		rdev->last_read_error = cur_time_mon;
2258 		return;
2259 	}
2260 
2261 	hours_since_last = (long)(cur_time_mon -
2262 			    rdev->last_read_error) / 3600;
2263 
2264 	rdev->last_read_error = cur_time_mon;
2265 
2266 	/*
2267 	 * if hours_since_last is > the number of bits in read_errors
2268 	 * just set read errors to 0. We do this to avoid
2269 	 * overflowing the shift of read_errors by hours_since_last.
2270 	 */
2271 	if (hours_since_last >= 8 * sizeof(read_errors))
2272 		atomic_set(&rdev->read_errors, 0);
2273 	else
2274 		atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2275 }
2276 
2277 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2278 			    int sectors, struct page *page, int rw)
2279 {
2280 	sector_t first_bad;
2281 	int bad_sectors;
2282 
2283 	if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2284 	    && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2285 		return -1;
2286 	if (sync_page_io(rdev, sector, sectors << 9, page, rw, 0, false))
2287 		/* success */
2288 		return 1;
2289 	if (rw == WRITE) {
2290 		set_bit(WriteErrorSeen, &rdev->flags);
2291 		if (!test_and_set_bit(WantReplacement, &rdev->flags))
2292 			set_bit(MD_RECOVERY_NEEDED,
2293 				&rdev->mddev->recovery);
2294 	}
2295 	/* need to record an error - either for the block or the device */
2296 	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2297 		md_error(rdev->mddev, rdev);
2298 	return 0;
2299 }
2300 
2301 /*
2302  * This is a kernel thread which:
2303  *
2304  *	1.	Retries failed read operations on working mirrors.
2305  *	2.	Updates the raid superblock when problems encounter.
2306  *	3.	Performs writes following reads for array synchronising.
2307  */
2308 
2309 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2310 {
2311 	int sect = 0; /* Offset from r10_bio->sector */
2312 	int sectors = r10_bio->sectors;
2313 	struct md_rdev*rdev;
2314 	int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2315 	int d = r10_bio->devs[r10_bio->read_slot].devnum;
2316 
2317 	/* still own a reference to this rdev, so it cannot
2318 	 * have been cleared recently.
2319 	 */
2320 	rdev = conf->mirrors[d].rdev;
2321 
2322 	if (test_bit(Faulty, &rdev->flags))
2323 		/* drive has already been failed, just ignore any
2324 		   more fix_read_error() attempts */
2325 		return;
2326 
2327 	check_decay_read_errors(mddev, rdev);
2328 	atomic_inc(&rdev->read_errors);
2329 	if (atomic_read(&rdev->read_errors) > max_read_errors) {
2330 		char b[BDEVNAME_SIZE];
2331 		bdevname(rdev->bdev, b);
2332 
2333 		pr_notice("md/raid10:%s: %s: Raid device exceeded read_error threshold [cur %d:max %d]\n",
2334 			  mdname(mddev), b,
2335 			  atomic_read(&rdev->read_errors), max_read_errors);
2336 		pr_notice("md/raid10:%s: %s: Failing raid device\n",
2337 			  mdname(mddev), b);
2338 		md_error(mddev, rdev);
2339 		r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2340 		return;
2341 	}
2342 
2343 	while(sectors) {
2344 		int s = sectors;
2345 		int sl = r10_bio->read_slot;
2346 		int success = 0;
2347 		int start;
2348 
2349 		if (s > (PAGE_SIZE>>9))
2350 			s = PAGE_SIZE >> 9;
2351 
2352 		rcu_read_lock();
2353 		do {
2354 			sector_t first_bad;
2355 			int bad_sectors;
2356 
2357 			d = r10_bio->devs[sl].devnum;
2358 			rdev = rcu_dereference(conf->mirrors[d].rdev);
2359 			if (rdev &&
2360 			    test_bit(In_sync, &rdev->flags) &&
2361 			    !test_bit(Faulty, &rdev->flags) &&
2362 			    is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2363 					&first_bad, &bad_sectors) == 0) {
2364 				atomic_inc(&rdev->nr_pending);
2365 				rcu_read_unlock();
2366 				success = sync_page_io(rdev,
2367 						       r10_bio->devs[sl].addr +
2368 						       sect,
2369 						       s<<9,
2370 						       conf->tmppage,
2371 						       REQ_OP_READ, 0, false);
2372 				rdev_dec_pending(rdev, mddev);
2373 				rcu_read_lock();
2374 				if (success)
2375 					break;
2376 			}
2377 			sl++;
2378 			if (sl == conf->copies)
2379 				sl = 0;
2380 		} while (!success && sl != r10_bio->read_slot);
2381 		rcu_read_unlock();
2382 
2383 		if (!success) {
2384 			/* Cannot read from anywhere, just mark the block
2385 			 * as bad on the first device to discourage future
2386 			 * reads.
2387 			 */
2388 			int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2389 			rdev = conf->mirrors[dn].rdev;
2390 
2391 			if (!rdev_set_badblocks(
2392 				    rdev,
2393 				    r10_bio->devs[r10_bio->read_slot].addr
2394 				    + sect,
2395 				    s, 0)) {
2396 				md_error(mddev, rdev);
2397 				r10_bio->devs[r10_bio->read_slot].bio
2398 					= IO_BLOCKED;
2399 			}
2400 			break;
2401 		}
2402 
2403 		start = sl;
2404 		/* write it back and re-read */
2405 		rcu_read_lock();
2406 		while (sl != r10_bio->read_slot) {
2407 			char b[BDEVNAME_SIZE];
2408 
2409 			if (sl==0)
2410 				sl = conf->copies;
2411 			sl--;
2412 			d = r10_bio->devs[sl].devnum;
2413 			rdev = rcu_dereference(conf->mirrors[d].rdev);
2414 			if (!rdev ||
2415 			    test_bit(Faulty, &rdev->flags) ||
2416 			    !test_bit(In_sync, &rdev->flags))
2417 				continue;
2418 
2419 			atomic_inc(&rdev->nr_pending);
2420 			rcu_read_unlock();
2421 			if (r10_sync_page_io(rdev,
2422 					     r10_bio->devs[sl].addr +
2423 					     sect,
2424 					     s, conf->tmppage, WRITE)
2425 			    == 0) {
2426 				/* Well, this device is dead */
2427 				pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %s)\n",
2428 					  mdname(mddev), s,
2429 					  (unsigned long long)(
2430 						  sect +
2431 						  choose_data_offset(r10_bio,
2432 								     rdev)),
2433 					  bdevname(rdev->bdev, b));
2434 				pr_notice("md/raid10:%s: %s: failing drive\n",
2435 					  mdname(mddev),
2436 					  bdevname(rdev->bdev, b));
2437 			}
2438 			rdev_dec_pending(rdev, mddev);
2439 			rcu_read_lock();
2440 		}
2441 		sl = start;
2442 		while (sl != r10_bio->read_slot) {
2443 			char b[BDEVNAME_SIZE];
2444 
2445 			if (sl==0)
2446 				sl = conf->copies;
2447 			sl--;
2448 			d = r10_bio->devs[sl].devnum;
2449 			rdev = rcu_dereference(conf->mirrors[d].rdev);
2450 			if (!rdev ||
2451 			    test_bit(Faulty, &rdev->flags) ||
2452 			    !test_bit(In_sync, &rdev->flags))
2453 				continue;
2454 
2455 			atomic_inc(&rdev->nr_pending);
2456 			rcu_read_unlock();
2457 			switch (r10_sync_page_io(rdev,
2458 					     r10_bio->devs[sl].addr +
2459 					     sect,
2460 					     s, conf->tmppage,
2461 						 READ)) {
2462 			case 0:
2463 				/* Well, this device is dead */
2464 				pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %s)\n",
2465 				       mdname(mddev), s,
2466 				       (unsigned long long)(
2467 					       sect +
2468 					       choose_data_offset(r10_bio, rdev)),
2469 				       bdevname(rdev->bdev, b));
2470 				pr_notice("md/raid10:%s: %s: failing drive\n",
2471 				       mdname(mddev),
2472 				       bdevname(rdev->bdev, b));
2473 				break;
2474 			case 1:
2475 				pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %s)\n",
2476 				       mdname(mddev), s,
2477 				       (unsigned long long)(
2478 					       sect +
2479 					       choose_data_offset(r10_bio, rdev)),
2480 				       bdevname(rdev->bdev, b));
2481 				atomic_add(s, &rdev->corrected_errors);
2482 			}
2483 
2484 			rdev_dec_pending(rdev, mddev);
2485 			rcu_read_lock();
2486 		}
2487 		rcu_read_unlock();
2488 
2489 		sectors -= s;
2490 		sect += s;
2491 	}
2492 }
2493 
2494 static int narrow_write_error(struct r10bio *r10_bio, int i)
2495 {
2496 	struct bio *bio = r10_bio->master_bio;
2497 	struct mddev *mddev = r10_bio->mddev;
2498 	struct r10conf *conf = mddev->private;
2499 	struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2500 	/* bio has the data to be written to slot 'i' where
2501 	 * we just recently had a write error.
2502 	 * We repeatedly clone the bio and trim down to one block,
2503 	 * then try the write.  Where the write fails we record
2504 	 * a bad block.
2505 	 * It is conceivable that the bio doesn't exactly align with
2506 	 * blocks.  We must handle this.
2507 	 *
2508 	 * We currently own a reference to the rdev.
2509 	 */
2510 
2511 	int block_sectors;
2512 	sector_t sector;
2513 	int sectors;
2514 	int sect_to_write = r10_bio->sectors;
2515 	int ok = 1;
2516 
2517 	if (rdev->badblocks.shift < 0)
2518 		return 0;
2519 
2520 	block_sectors = roundup(1 << rdev->badblocks.shift,
2521 				bdev_logical_block_size(rdev->bdev) >> 9);
2522 	sector = r10_bio->sector;
2523 	sectors = ((r10_bio->sector + block_sectors)
2524 		   & ~(sector_t)(block_sectors - 1))
2525 		- sector;
2526 
2527 	while (sect_to_write) {
2528 		struct bio *wbio;
2529 		sector_t wsector;
2530 		if (sectors > sect_to_write)
2531 			sectors = sect_to_write;
2532 		/* Write at 'sector' for 'sectors' */
2533 		wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2534 		bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2535 		wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2536 		wbio->bi_iter.bi_sector = wsector +
2537 				   choose_data_offset(r10_bio, rdev);
2538 		wbio->bi_bdev = rdev->bdev;
2539 		bio_set_op_attrs(wbio, REQ_OP_WRITE, 0);
2540 
2541 		if (submit_bio_wait(wbio) < 0)
2542 			/* Failure! */
2543 			ok = rdev_set_badblocks(rdev, wsector,
2544 						sectors, 0)
2545 				&& ok;
2546 
2547 		bio_put(wbio);
2548 		sect_to_write -= sectors;
2549 		sector += sectors;
2550 		sectors = block_sectors;
2551 	}
2552 	return ok;
2553 }
2554 
2555 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2556 {
2557 	int slot = r10_bio->read_slot;
2558 	struct bio *bio;
2559 	struct r10conf *conf = mddev->private;
2560 	struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2561 	char b[BDEVNAME_SIZE];
2562 	unsigned long do_sync;
2563 	int max_sectors;
2564 	dev_t bio_dev;
2565 	sector_t bio_last_sector;
2566 
2567 	/* we got a read error. Maybe the drive is bad.  Maybe just
2568 	 * the block and we can fix it.
2569 	 * We freeze all other IO, and try reading the block from
2570 	 * other devices.  When we find one, we re-write
2571 	 * and check it that fixes the read error.
2572 	 * This is all done synchronously while the array is
2573 	 * frozen.
2574 	 */
2575 	bio = r10_bio->devs[slot].bio;
2576 	bdevname(bio->bi_bdev, b);
2577 	bio_dev = bio->bi_bdev->bd_dev;
2578 	bio_last_sector = r10_bio->devs[slot].addr + rdev->data_offset + r10_bio->sectors;
2579 	bio_put(bio);
2580 	r10_bio->devs[slot].bio = NULL;
2581 
2582 	if (mddev->ro)
2583 		r10_bio->devs[slot].bio = IO_BLOCKED;
2584 	else if (!test_bit(FailFast, &rdev->flags)) {
2585 		freeze_array(conf, 1);
2586 		fix_read_error(conf, mddev, r10_bio);
2587 		unfreeze_array(conf);
2588 	} else
2589 		md_error(mddev, rdev);
2590 
2591 	rdev_dec_pending(rdev, mddev);
2592 
2593 read_more:
2594 	rdev = read_balance(conf, r10_bio, &max_sectors);
2595 	if (rdev == NULL) {
2596 		pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
2597 				    mdname(mddev), b,
2598 				    (unsigned long long)r10_bio->sector);
2599 		raid_end_bio_io(r10_bio);
2600 		return;
2601 	}
2602 
2603 	do_sync = (r10_bio->master_bio->bi_opf & REQ_SYNC);
2604 	slot = r10_bio->read_slot;
2605 	pr_err_ratelimited("md/raid10:%s: %s: redirecting sector %llu to another mirror\n",
2606 			   mdname(mddev),
2607 			   bdevname(rdev->bdev, b),
2608 			   (unsigned long long)r10_bio->sector);
2609 	bio = bio_clone_mddev(r10_bio->master_bio,
2610 			      GFP_NOIO, mddev);
2611 	bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2612 	r10_bio->devs[slot].bio = bio;
2613 	r10_bio->devs[slot].rdev = rdev;
2614 	bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2615 		+ choose_data_offset(r10_bio, rdev);
2616 	bio->bi_bdev = rdev->bdev;
2617 	bio_set_op_attrs(bio, REQ_OP_READ, do_sync);
2618 	if (test_bit(FailFast, &rdev->flags) &&
2619 	    test_bit(R10BIO_FailFast, &r10_bio->state))
2620 		bio->bi_opf |= MD_FAILFAST;
2621 	bio->bi_private = r10_bio;
2622 	bio->bi_end_io = raid10_end_read_request;
2623 	trace_block_bio_remap(bdev_get_queue(bio->bi_bdev),
2624 			      bio, bio_dev,
2625 			      bio_last_sector - r10_bio->sectors);
2626 
2627 	if (max_sectors < r10_bio->sectors) {
2628 		/* Drat - have to split this up more */
2629 		struct bio *mbio = r10_bio->master_bio;
2630 		int sectors_handled =
2631 			r10_bio->sector + max_sectors
2632 			- mbio->bi_iter.bi_sector;
2633 		r10_bio->sectors = max_sectors;
2634 		spin_lock_irq(&conf->device_lock);
2635 		if (mbio->bi_phys_segments == 0)
2636 			mbio->bi_phys_segments = 2;
2637 		else
2638 			mbio->bi_phys_segments++;
2639 		spin_unlock_irq(&conf->device_lock);
2640 		generic_make_request(bio);
2641 
2642 		r10_bio = mempool_alloc(conf->r10bio_pool,
2643 					GFP_NOIO);
2644 		r10_bio->master_bio = mbio;
2645 		r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2646 		r10_bio->state = 0;
2647 		set_bit(R10BIO_ReadError,
2648 			&r10_bio->state);
2649 		r10_bio->mddev = mddev;
2650 		r10_bio->sector = mbio->bi_iter.bi_sector
2651 			+ sectors_handled;
2652 
2653 		goto read_more;
2654 	} else
2655 		generic_make_request(bio);
2656 }
2657 
2658 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2659 {
2660 	/* Some sort of write request has finished and it
2661 	 * succeeded in writing where we thought there was a
2662 	 * bad block.  So forget the bad block.
2663 	 * Or possibly if failed and we need to record
2664 	 * a bad block.
2665 	 */
2666 	int m;
2667 	struct md_rdev *rdev;
2668 
2669 	if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2670 	    test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2671 		for (m = 0; m < conf->copies; m++) {
2672 			int dev = r10_bio->devs[m].devnum;
2673 			rdev = conf->mirrors[dev].rdev;
2674 			if (r10_bio->devs[m].bio == NULL)
2675 				continue;
2676 			if (!r10_bio->devs[m].bio->bi_error) {
2677 				rdev_clear_badblocks(
2678 					rdev,
2679 					r10_bio->devs[m].addr,
2680 					r10_bio->sectors, 0);
2681 			} else {
2682 				if (!rdev_set_badblocks(
2683 					    rdev,
2684 					    r10_bio->devs[m].addr,
2685 					    r10_bio->sectors, 0))
2686 					md_error(conf->mddev, rdev);
2687 			}
2688 			rdev = conf->mirrors[dev].replacement;
2689 			if (r10_bio->devs[m].repl_bio == NULL)
2690 				continue;
2691 
2692 			if (!r10_bio->devs[m].repl_bio->bi_error) {
2693 				rdev_clear_badblocks(
2694 					rdev,
2695 					r10_bio->devs[m].addr,
2696 					r10_bio->sectors, 0);
2697 			} else {
2698 				if (!rdev_set_badblocks(
2699 					    rdev,
2700 					    r10_bio->devs[m].addr,
2701 					    r10_bio->sectors, 0))
2702 					md_error(conf->mddev, rdev);
2703 			}
2704 		}
2705 		put_buf(r10_bio);
2706 	} else {
2707 		bool fail = false;
2708 		for (m = 0; m < conf->copies; m++) {
2709 			int dev = r10_bio->devs[m].devnum;
2710 			struct bio *bio = r10_bio->devs[m].bio;
2711 			rdev = conf->mirrors[dev].rdev;
2712 			if (bio == IO_MADE_GOOD) {
2713 				rdev_clear_badblocks(
2714 					rdev,
2715 					r10_bio->devs[m].addr,
2716 					r10_bio->sectors, 0);
2717 				rdev_dec_pending(rdev, conf->mddev);
2718 			} else if (bio != NULL && bio->bi_error) {
2719 				fail = true;
2720 				if (!narrow_write_error(r10_bio, m)) {
2721 					md_error(conf->mddev, rdev);
2722 					set_bit(R10BIO_Degraded,
2723 						&r10_bio->state);
2724 				}
2725 				rdev_dec_pending(rdev, conf->mddev);
2726 			}
2727 			bio = r10_bio->devs[m].repl_bio;
2728 			rdev = conf->mirrors[dev].replacement;
2729 			if (rdev && bio == IO_MADE_GOOD) {
2730 				rdev_clear_badblocks(
2731 					rdev,
2732 					r10_bio->devs[m].addr,
2733 					r10_bio->sectors, 0);
2734 				rdev_dec_pending(rdev, conf->mddev);
2735 			}
2736 		}
2737 		if (fail) {
2738 			spin_lock_irq(&conf->device_lock);
2739 			list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2740 			conf->nr_queued++;
2741 			spin_unlock_irq(&conf->device_lock);
2742 			md_wakeup_thread(conf->mddev->thread);
2743 		} else {
2744 			if (test_bit(R10BIO_WriteError,
2745 				     &r10_bio->state))
2746 				close_write(r10_bio);
2747 			raid_end_bio_io(r10_bio);
2748 		}
2749 	}
2750 }
2751 
2752 static void raid10d(struct md_thread *thread)
2753 {
2754 	struct mddev *mddev = thread->mddev;
2755 	struct r10bio *r10_bio;
2756 	unsigned long flags;
2757 	struct r10conf *conf = mddev->private;
2758 	struct list_head *head = &conf->retry_list;
2759 	struct blk_plug plug;
2760 
2761 	md_check_recovery(mddev);
2762 
2763 	if (!list_empty_careful(&conf->bio_end_io_list) &&
2764 	    !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2765 		LIST_HEAD(tmp);
2766 		spin_lock_irqsave(&conf->device_lock, flags);
2767 		if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2768 			while (!list_empty(&conf->bio_end_io_list)) {
2769 				list_move(conf->bio_end_io_list.prev, &tmp);
2770 				conf->nr_queued--;
2771 			}
2772 		}
2773 		spin_unlock_irqrestore(&conf->device_lock, flags);
2774 		while (!list_empty(&tmp)) {
2775 			r10_bio = list_first_entry(&tmp, struct r10bio,
2776 						   retry_list);
2777 			list_del(&r10_bio->retry_list);
2778 			if (mddev->degraded)
2779 				set_bit(R10BIO_Degraded, &r10_bio->state);
2780 
2781 			if (test_bit(R10BIO_WriteError,
2782 				     &r10_bio->state))
2783 				close_write(r10_bio);
2784 			raid_end_bio_io(r10_bio);
2785 		}
2786 	}
2787 
2788 	blk_start_plug(&plug);
2789 	for (;;) {
2790 
2791 		flush_pending_writes(conf);
2792 
2793 		spin_lock_irqsave(&conf->device_lock, flags);
2794 		if (list_empty(head)) {
2795 			spin_unlock_irqrestore(&conf->device_lock, flags);
2796 			break;
2797 		}
2798 		r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2799 		list_del(head->prev);
2800 		conf->nr_queued--;
2801 		spin_unlock_irqrestore(&conf->device_lock, flags);
2802 
2803 		mddev = r10_bio->mddev;
2804 		conf = mddev->private;
2805 		if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2806 		    test_bit(R10BIO_WriteError, &r10_bio->state))
2807 			handle_write_completed(conf, r10_bio);
2808 		else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2809 			reshape_request_write(mddev, r10_bio);
2810 		else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2811 			sync_request_write(mddev, r10_bio);
2812 		else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2813 			recovery_request_write(mddev, r10_bio);
2814 		else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2815 			handle_read_error(mddev, r10_bio);
2816 		else {
2817 			/* just a partial read to be scheduled from a
2818 			 * separate context
2819 			 */
2820 			int slot = r10_bio->read_slot;
2821 			generic_make_request(r10_bio->devs[slot].bio);
2822 		}
2823 
2824 		cond_resched();
2825 		if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
2826 			md_check_recovery(mddev);
2827 	}
2828 	blk_finish_plug(&plug);
2829 }
2830 
2831 static int init_resync(struct r10conf *conf)
2832 {
2833 	int buffs;
2834 	int i;
2835 
2836 	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2837 	BUG_ON(conf->r10buf_pool);
2838 	conf->have_replacement = 0;
2839 	for (i = 0; i < conf->geo.raid_disks; i++)
2840 		if (conf->mirrors[i].replacement)
2841 			conf->have_replacement = 1;
2842 	conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2843 	if (!conf->r10buf_pool)
2844 		return -ENOMEM;
2845 	conf->next_resync = 0;
2846 	return 0;
2847 }
2848 
2849 /*
2850  * perform a "sync" on one "block"
2851  *
2852  * We need to make sure that no normal I/O request - particularly write
2853  * requests - conflict with active sync requests.
2854  *
2855  * This is achieved by tracking pending requests and a 'barrier' concept
2856  * that can be installed to exclude normal IO requests.
2857  *
2858  * Resync and recovery are handled very differently.
2859  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2860  *
2861  * For resync, we iterate over virtual addresses, read all copies,
2862  * and update if there are differences.  If only one copy is live,
2863  * skip it.
2864  * For recovery, we iterate over physical addresses, read a good
2865  * value for each non-in_sync drive, and over-write.
2866  *
2867  * So, for recovery we may have several outstanding complex requests for a
2868  * given address, one for each out-of-sync device.  We model this by allocating
2869  * a number of r10_bio structures, one for each out-of-sync device.
2870  * As we setup these structures, we collect all bio's together into a list
2871  * which we then process collectively to add pages, and then process again
2872  * to pass to generic_make_request.
2873  *
2874  * The r10_bio structures are linked using a borrowed master_bio pointer.
2875  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2876  * has its remaining count decremented to 0, the whole complex operation
2877  * is complete.
2878  *
2879  */
2880 
2881 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2882 			     int *skipped)
2883 {
2884 	struct r10conf *conf = mddev->private;
2885 	struct r10bio *r10_bio;
2886 	struct bio *biolist = NULL, *bio;
2887 	sector_t max_sector, nr_sectors;
2888 	int i;
2889 	int max_sync;
2890 	sector_t sync_blocks;
2891 	sector_t sectors_skipped = 0;
2892 	int chunks_skipped = 0;
2893 	sector_t chunk_mask = conf->geo.chunk_mask;
2894 
2895 	if (!conf->r10buf_pool)
2896 		if (init_resync(conf))
2897 			return 0;
2898 
2899 	/*
2900 	 * Allow skipping a full rebuild for incremental assembly
2901 	 * of a clean array, like RAID1 does.
2902 	 */
2903 	if (mddev->bitmap == NULL &&
2904 	    mddev->recovery_cp == MaxSector &&
2905 	    mddev->reshape_position == MaxSector &&
2906 	    !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2907 	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2908 	    !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2909 	    conf->fullsync == 0) {
2910 		*skipped = 1;
2911 		return mddev->dev_sectors - sector_nr;
2912 	}
2913 
2914  skipped:
2915 	max_sector = mddev->dev_sectors;
2916 	if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2917 	    test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2918 		max_sector = mddev->resync_max_sectors;
2919 	if (sector_nr >= max_sector) {
2920 		/* If we aborted, we need to abort the
2921 		 * sync on the 'current' bitmap chucks (there can
2922 		 * be several when recovering multiple devices).
2923 		 * as we may have started syncing it but not finished.
2924 		 * We can find the current address in
2925 		 * mddev->curr_resync, but for recovery,
2926 		 * we need to convert that to several
2927 		 * virtual addresses.
2928 		 */
2929 		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2930 			end_reshape(conf);
2931 			close_sync(conf);
2932 			return 0;
2933 		}
2934 
2935 		if (mddev->curr_resync < max_sector) { /* aborted */
2936 			if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2937 				bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2938 						&sync_blocks, 1);
2939 			else for (i = 0; i < conf->geo.raid_disks; i++) {
2940 				sector_t sect =
2941 					raid10_find_virt(conf, mddev->curr_resync, i);
2942 				bitmap_end_sync(mddev->bitmap, sect,
2943 						&sync_blocks, 1);
2944 			}
2945 		} else {
2946 			/* completed sync */
2947 			if ((!mddev->bitmap || conf->fullsync)
2948 			    && conf->have_replacement
2949 			    && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2950 				/* Completed a full sync so the replacements
2951 				 * are now fully recovered.
2952 				 */
2953 				rcu_read_lock();
2954 				for (i = 0; i < conf->geo.raid_disks; i++) {
2955 					struct md_rdev *rdev =
2956 						rcu_dereference(conf->mirrors[i].replacement);
2957 					if (rdev)
2958 						rdev->recovery_offset = MaxSector;
2959 				}
2960 				rcu_read_unlock();
2961 			}
2962 			conf->fullsync = 0;
2963 		}
2964 		bitmap_close_sync(mddev->bitmap);
2965 		close_sync(conf);
2966 		*skipped = 1;
2967 		return sectors_skipped;
2968 	}
2969 
2970 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2971 		return reshape_request(mddev, sector_nr, skipped);
2972 
2973 	if (chunks_skipped >= conf->geo.raid_disks) {
2974 		/* if there has been nothing to do on any drive,
2975 		 * then there is nothing to do at all..
2976 		 */
2977 		*skipped = 1;
2978 		return (max_sector - sector_nr) + sectors_skipped;
2979 	}
2980 
2981 	if (max_sector > mddev->resync_max)
2982 		max_sector = mddev->resync_max; /* Don't do IO beyond here */
2983 
2984 	/* make sure whole request will fit in a chunk - if chunks
2985 	 * are meaningful
2986 	 */
2987 	if (conf->geo.near_copies < conf->geo.raid_disks &&
2988 	    max_sector > (sector_nr | chunk_mask))
2989 		max_sector = (sector_nr | chunk_mask) + 1;
2990 
2991 	/*
2992 	 * If there is non-resync activity waiting for a turn, then let it
2993 	 * though before starting on this new sync request.
2994 	 */
2995 	if (conf->nr_waiting)
2996 		schedule_timeout_uninterruptible(1);
2997 
2998 	/* Again, very different code for resync and recovery.
2999 	 * Both must result in an r10bio with a list of bios that
3000 	 * have bi_end_io, bi_sector, bi_bdev set,
3001 	 * and bi_private set to the r10bio.
3002 	 * For recovery, we may actually create several r10bios
3003 	 * with 2 bios in each, that correspond to the bios in the main one.
3004 	 * In this case, the subordinate r10bios link back through a
3005 	 * borrowed master_bio pointer, and the counter in the master
3006 	 * includes a ref from each subordinate.
3007 	 */
3008 	/* First, we decide what to do and set ->bi_end_io
3009 	 * To end_sync_read if we want to read, and
3010 	 * end_sync_write if we will want to write.
3011 	 */
3012 
3013 	max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3014 	if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3015 		/* recovery... the complicated one */
3016 		int j;
3017 		r10_bio = NULL;
3018 
3019 		for (i = 0 ; i < conf->geo.raid_disks; i++) {
3020 			int still_degraded;
3021 			struct r10bio *rb2;
3022 			sector_t sect;
3023 			int must_sync;
3024 			int any_working;
3025 			struct raid10_info *mirror = &conf->mirrors[i];
3026 			struct md_rdev *mrdev, *mreplace;
3027 
3028 			rcu_read_lock();
3029 			mrdev = rcu_dereference(mirror->rdev);
3030 			mreplace = rcu_dereference(mirror->replacement);
3031 
3032 			if ((mrdev == NULL ||
3033 			     test_bit(Faulty, &mrdev->flags) ||
3034 			     test_bit(In_sync, &mrdev->flags)) &&
3035 			    (mreplace == NULL ||
3036 			     test_bit(Faulty, &mreplace->flags))) {
3037 				rcu_read_unlock();
3038 				continue;
3039 			}
3040 
3041 			still_degraded = 0;
3042 			/* want to reconstruct this device */
3043 			rb2 = r10_bio;
3044 			sect = raid10_find_virt(conf, sector_nr, i);
3045 			if (sect >= mddev->resync_max_sectors) {
3046 				/* last stripe is not complete - don't
3047 				 * try to recover this sector.
3048 				 */
3049 				rcu_read_unlock();
3050 				continue;
3051 			}
3052 			if (mreplace && test_bit(Faulty, &mreplace->flags))
3053 				mreplace = NULL;
3054 			/* Unless we are doing a full sync, or a replacement
3055 			 * we only need to recover the block if it is set in
3056 			 * the bitmap
3057 			 */
3058 			must_sync = bitmap_start_sync(mddev->bitmap, sect,
3059 						      &sync_blocks, 1);
3060 			if (sync_blocks < max_sync)
3061 				max_sync = sync_blocks;
3062 			if (!must_sync &&
3063 			    mreplace == NULL &&
3064 			    !conf->fullsync) {
3065 				/* yep, skip the sync_blocks here, but don't assume
3066 				 * that there will never be anything to do here
3067 				 */
3068 				chunks_skipped = -1;
3069 				rcu_read_unlock();
3070 				continue;
3071 			}
3072 			atomic_inc(&mrdev->nr_pending);
3073 			if (mreplace)
3074 				atomic_inc(&mreplace->nr_pending);
3075 			rcu_read_unlock();
3076 
3077 			r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3078 			r10_bio->state = 0;
3079 			raise_barrier(conf, rb2 != NULL);
3080 			atomic_set(&r10_bio->remaining, 0);
3081 
3082 			r10_bio->master_bio = (struct bio*)rb2;
3083 			if (rb2)
3084 				atomic_inc(&rb2->remaining);
3085 			r10_bio->mddev = mddev;
3086 			set_bit(R10BIO_IsRecover, &r10_bio->state);
3087 			r10_bio->sector = sect;
3088 
3089 			raid10_find_phys(conf, r10_bio);
3090 
3091 			/* Need to check if the array will still be
3092 			 * degraded
3093 			 */
3094 			rcu_read_lock();
3095 			for (j = 0; j < conf->geo.raid_disks; j++) {
3096 				struct md_rdev *rdev = rcu_dereference(
3097 					conf->mirrors[j].rdev);
3098 				if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3099 					still_degraded = 1;
3100 					break;
3101 				}
3102 			}
3103 
3104 			must_sync = bitmap_start_sync(mddev->bitmap, sect,
3105 						      &sync_blocks, still_degraded);
3106 
3107 			any_working = 0;
3108 			for (j=0; j<conf->copies;j++) {
3109 				int k;
3110 				int d = r10_bio->devs[j].devnum;
3111 				sector_t from_addr, to_addr;
3112 				struct md_rdev *rdev =
3113 					rcu_dereference(conf->mirrors[d].rdev);
3114 				sector_t sector, first_bad;
3115 				int bad_sectors;
3116 				if (!rdev ||
3117 				    !test_bit(In_sync, &rdev->flags))
3118 					continue;
3119 				/* This is where we read from */
3120 				any_working = 1;
3121 				sector = r10_bio->devs[j].addr;
3122 
3123 				if (is_badblock(rdev, sector, max_sync,
3124 						&first_bad, &bad_sectors)) {
3125 					if (first_bad > sector)
3126 						max_sync = first_bad - sector;
3127 					else {
3128 						bad_sectors -= (sector
3129 								- first_bad);
3130 						if (max_sync > bad_sectors)
3131 							max_sync = bad_sectors;
3132 						continue;
3133 					}
3134 				}
3135 				bio = r10_bio->devs[0].bio;
3136 				bio_reset(bio);
3137 				bio->bi_next = biolist;
3138 				biolist = bio;
3139 				bio->bi_private = r10_bio;
3140 				bio->bi_end_io = end_sync_read;
3141 				bio_set_op_attrs(bio, REQ_OP_READ, 0);
3142 				if (test_bit(FailFast, &rdev->flags))
3143 					bio->bi_opf |= MD_FAILFAST;
3144 				from_addr = r10_bio->devs[j].addr;
3145 				bio->bi_iter.bi_sector = from_addr +
3146 					rdev->data_offset;
3147 				bio->bi_bdev = rdev->bdev;
3148 				atomic_inc(&rdev->nr_pending);
3149 				/* and we write to 'i' (if not in_sync) */
3150 
3151 				for (k=0; k<conf->copies; k++)
3152 					if (r10_bio->devs[k].devnum == i)
3153 						break;
3154 				BUG_ON(k == conf->copies);
3155 				to_addr = r10_bio->devs[k].addr;
3156 				r10_bio->devs[0].devnum = d;
3157 				r10_bio->devs[0].addr = from_addr;
3158 				r10_bio->devs[1].devnum = i;
3159 				r10_bio->devs[1].addr = to_addr;
3160 
3161 				if (!test_bit(In_sync, &mrdev->flags)) {
3162 					bio = r10_bio->devs[1].bio;
3163 					bio_reset(bio);
3164 					bio->bi_next = biolist;
3165 					biolist = bio;
3166 					bio->bi_private = r10_bio;
3167 					bio->bi_end_io = end_sync_write;
3168 					bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3169 					bio->bi_iter.bi_sector = to_addr
3170 						+ mrdev->data_offset;
3171 					bio->bi_bdev = mrdev->bdev;
3172 					atomic_inc(&r10_bio->remaining);
3173 				} else
3174 					r10_bio->devs[1].bio->bi_end_io = NULL;
3175 
3176 				/* and maybe write to replacement */
3177 				bio = r10_bio->devs[1].repl_bio;
3178 				if (bio)
3179 					bio->bi_end_io = NULL;
3180 				/* Note: if mreplace != NULL, then bio
3181 				 * cannot be NULL as r10buf_pool_alloc will
3182 				 * have allocated it.
3183 				 * So the second test here is pointless.
3184 				 * But it keeps semantic-checkers happy, and
3185 				 * this comment keeps human reviewers
3186 				 * happy.
3187 				 */
3188 				if (mreplace == NULL || bio == NULL ||
3189 				    test_bit(Faulty, &mreplace->flags))
3190 					break;
3191 				bio_reset(bio);
3192 				bio->bi_next = biolist;
3193 				biolist = bio;
3194 				bio->bi_private = r10_bio;
3195 				bio->bi_end_io = end_sync_write;
3196 				bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3197 				bio->bi_iter.bi_sector = to_addr +
3198 					mreplace->data_offset;
3199 				bio->bi_bdev = mreplace->bdev;
3200 				atomic_inc(&r10_bio->remaining);
3201 				break;
3202 			}
3203 			rcu_read_unlock();
3204 			if (j == conf->copies) {
3205 				/* Cannot recover, so abort the recovery or
3206 				 * record a bad block */
3207 				if (any_working) {
3208 					/* problem is that there are bad blocks
3209 					 * on other device(s)
3210 					 */
3211 					int k;
3212 					for (k = 0; k < conf->copies; k++)
3213 						if (r10_bio->devs[k].devnum == i)
3214 							break;
3215 					if (!test_bit(In_sync,
3216 						      &mrdev->flags)
3217 					    && !rdev_set_badblocks(
3218 						    mrdev,
3219 						    r10_bio->devs[k].addr,
3220 						    max_sync, 0))
3221 						any_working = 0;
3222 					if (mreplace &&
3223 					    !rdev_set_badblocks(
3224 						    mreplace,
3225 						    r10_bio->devs[k].addr,
3226 						    max_sync, 0))
3227 						any_working = 0;
3228 				}
3229 				if (!any_working)  {
3230 					if (!test_and_set_bit(MD_RECOVERY_INTR,
3231 							      &mddev->recovery))
3232 						pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3233 						       mdname(mddev));
3234 					mirror->recovery_disabled
3235 						= mddev->recovery_disabled;
3236 				}
3237 				put_buf(r10_bio);
3238 				if (rb2)
3239 					atomic_dec(&rb2->remaining);
3240 				r10_bio = rb2;
3241 				rdev_dec_pending(mrdev, mddev);
3242 				if (mreplace)
3243 					rdev_dec_pending(mreplace, mddev);
3244 				break;
3245 			}
3246 			rdev_dec_pending(mrdev, mddev);
3247 			if (mreplace)
3248 				rdev_dec_pending(mreplace, mddev);
3249 			if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3250 				/* Only want this if there is elsewhere to
3251 				 * read from. 'j' is currently the first
3252 				 * readable copy.
3253 				 */
3254 				int targets = 1;
3255 				for (; j < conf->copies; j++) {
3256 					int d = r10_bio->devs[j].devnum;
3257 					if (conf->mirrors[d].rdev &&
3258 					    test_bit(In_sync,
3259 						      &conf->mirrors[d].rdev->flags))
3260 						targets++;
3261 				}
3262 				if (targets == 1)
3263 					r10_bio->devs[0].bio->bi_opf
3264 						&= ~MD_FAILFAST;
3265 			}
3266 		}
3267 		if (biolist == NULL) {
3268 			while (r10_bio) {
3269 				struct r10bio *rb2 = r10_bio;
3270 				r10_bio = (struct r10bio*) rb2->master_bio;
3271 				rb2->master_bio = NULL;
3272 				put_buf(rb2);
3273 			}
3274 			goto giveup;
3275 		}
3276 	} else {
3277 		/* resync. Schedule a read for every block at this virt offset */
3278 		int count = 0;
3279 
3280 		bitmap_cond_end_sync(mddev->bitmap, sector_nr, 0);
3281 
3282 		if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3283 				       &sync_blocks, mddev->degraded) &&
3284 		    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3285 						 &mddev->recovery)) {
3286 			/* We can skip this block */
3287 			*skipped = 1;
3288 			return sync_blocks + sectors_skipped;
3289 		}
3290 		if (sync_blocks < max_sync)
3291 			max_sync = sync_blocks;
3292 		r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3293 		r10_bio->state = 0;
3294 
3295 		r10_bio->mddev = mddev;
3296 		atomic_set(&r10_bio->remaining, 0);
3297 		raise_barrier(conf, 0);
3298 		conf->next_resync = sector_nr;
3299 
3300 		r10_bio->master_bio = NULL;
3301 		r10_bio->sector = sector_nr;
3302 		set_bit(R10BIO_IsSync, &r10_bio->state);
3303 		raid10_find_phys(conf, r10_bio);
3304 		r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3305 
3306 		for (i = 0; i < conf->copies; i++) {
3307 			int d = r10_bio->devs[i].devnum;
3308 			sector_t first_bad, sector;
3309 			int bad_sectors;
3310 			struct md_rdev *rdev;
3311 
3312 			if (r10_bio->devs[i].repl_bio)
3313 				r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3314 
3315 			bio = r10_bio->devs[i].bio;
3316 			bio_reset(bio);
3317 			bio->bi_error = -EIO;
3318 			rcu_read_lock();
3319 			rdev = rcu_dereference(conf->mirrors[d].rdev);
3320 			if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3321 				rcu_read_unlock();
3322 				continue;
3323 			}
3324 			sector = r10_bio->devs[i].addr;
3325 			if (is_badblock(rdev, sector, max_sync,
3326 					&first_bad, &bad_sectors)) {
3327 				if (first_bad > sector)
3328 					max_sync = first_bad - sector;
3329 				else {
3330 					bad_sectors -= (sector - first_bad);
3331 					if (max_sync > bad_sectors)
3332 						max_sync = bad_sectors;
3333 					rcu_read_unlock();
3334 					continue;
3335 				}
3336 			}
3337 			atomic_inc(&rdev->nr_pending);
3338 			atomic_inc(&r10_bio->remaining);
3339 			bio->bi_next = biolist;
3340 			biolist = bio;
3341 			bio->bi_private = r10_bio;
3342 			bio->bi_end_io = end_sync_read;
3343 			bio_set_op_attrs(bio, REQ_OP_READ, 0);
3344 			if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
3345 				bio->bi_opf |= MD_FAILFAST;
3346 			bio->bi_iter.bi_sector = sector + rdev->data_offset;
3347 			bio->bi_bdev = rdev->bdev;
3348 			count++;
3349 
3350 			rdev = rcu_dereference(conf->mirrors[d].replacement);
3351 			if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3352 				rcu_read_unlock();
3353 				continue;
3354 			}
3355 			atomic_inc(&rdev->nr_pending);
3356 			rcu_read_unlock();
3357 
3358 			/* Need to set up for writing to the replacement */
3359 			bio = r10_bio->devs[i].repl_bio;
3360 			bio_reset(bio);
3361 			bio->bi_error = -EIO;
3362 
3363 			sector = r10_bio->devs[i].addr;
3364 			bio->bi_next = biolist;
3365 			biolist = bio;
3366 			bio->bi_private = r10_bio;
3367 			bio->bi_end_io = end_sync_write;
3368 			bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
3369 			if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
3370 				bio->bi_opf |= MD_FAILFAST;
3371 			bio->bi_iter.bi_sector = sector + rdev->data_offset;
3372 			bio->bi_bdev = rdev->bdev;
3373 			count++;
3374 		}
3375 
3376 		if (count < 2) {
3377 			for (i=0; i<conf->copies; i++) {
3378 				int d = r10_bio->devs[i].devnum;
3379 				if (r10_bio->devs[i].bio->bi_end_io)
3380 					rdev_dec_pending(conf->mirrors[d].rdev,
3381 							 mddev);
3382 				if (r10_bio->devs[i].repl_bio &&
3383 				    r10_bio->devs[i].repl_bio->bi_end_io)
3384 					rdev_dec_pending(
3385 						conf->mirrors[d].replacement,
3386 						mddev);
3387 			}
3388 			put_buf(r10_bio);
3389 			biolist = NULL;
3390 			goto giveup;
3391 		}
3392 	}
3393 
3394 	nr_sectors = 0;
3395 	if (sector_nr + max_sync < max_sector)
3396 		max_sector = sector_nr + max_sync;
3397 	do {
3398 		struct page *page;
3399 		int len = PAGE_SIZE;
3400 		if (sector_nr + (len>>9) > max_sector)
3401 			len = (max_sector - sector_nr) << 9;
3402 		if (len == 0)
3403 			break;
3404 		for (bio= biolist ; bio ; bio=bio->bi_next) {
3405 			struct bio *bio2;
3406 			page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3407 			if (bio_add_page(bio, page, len, 0))
3408 				continue;
3409 
3410 			/* stop here */
3411 			bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3412 			for (bio2 = biolist;
3413 			     bio2 && bio2 != bio;
3414 			     bio2 = bio2->bi_next) {
3415 				/* remove last page from this bio */
3416 				bio2->bi_vcnt--;
3417 				bio2->bi_iter.bi_size -= len;
3418 				bio_clear_flag(bio2, BIO_SEG_VALID);
3419 			}
3420 			goto bio_full;
3421 		}
3422 		nr_sectors += len>>9;
3423 		sector_nr += len>>9;
3424 	} while (biolist->bi_vcnt < RESYNC_PAGES);
3425  bio_full:
3426 	r10_bio->sectors = nr_sectors;
3427 
3428 	while (biolist) {
3429 		bio = biolist;
3430 		biolist = biolist->bi_next;
3431 
3432 		bio->bi_next = NULL;
3433 		r10_bio = bio->bi_private;
3434 		r10_bio->sectors = nr_sectors;
3435 
3436 		if (bio->bi_end_io == end_sync_read) {
3437 			md_sync_acct(bio->bi_bdev, nr_sectors);
3438 			bio->bi_error = 0;
3439 			generic_make_request(bio);
3440 		}
3441 	}
3442 
3443 	if (sectors_skipped)
3444 		/* pretend they weren't skipped, it makes
3445 		 * no important difference in this case
3446 		 */
3447 		md_done_sync(mddev, sectors_skipped, 1);
3448 
3449 	return sectors_skipped + nr_sectors;
3450  giveup:
3451 	/* There is nowhere to write, so all non-sync
3452 	 * drives must be failed or in resync, all drives
3453 	 * have a bad block, so try the next chunk...
3454 	 */
3455 	if (sector_nr + max_sync < max_sector)
3456 		max_sector = sector_nr + max_sync;
3457 
3458 	sectors_skipped += (max_sector - sector_nr);
3459 	chunks_skipped ++;
3460 	sector_nr = max_sector;
3461 	goto skipped;
3462 }
3463 
3464 static sector_t
3465 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3466 {
3467 	sector_t size;
3468 	struct r10conf *conf = mddev->private;
3469 
3470 	if (!raid_disks)
3471 		raid_disks = min(conf->geo.raid_disks,
3472 				 conf->prev.raid_disks);
3473 	if (!sectors)
3474 		sectors = conf->dev_sectors;
3475 
3476 	size = sectors >> conf->geo.chunk_shift;
3477 	sector_div(size, conf->geo.far_copies);
3478 	size = size * raid_disks;
3479 	sector_div(size, conf->geo.near_copies);
3480 
3481 	return size << conf->geo.chunk_shift;
3482 }
3483 
3484 static void calc_sectors(struct r10conf *conf, sector_t size)
3485 {
3486 	/* Calculate the number of sectors-per-device that will
3487 	 * actually be used, and set conf->dev_sectors and
3488 	 * conf->stride
3489 	 */
3490 
3491 	size = size >> conf->geo.chunk_shift;
3492 	sector_div(size, conf->geo.far_copies);
3493 	size = size * conf->geo.raid_disks;
3494 	sector_div(size, conf->geo.near_copies);
3495 	/* 'size' is now the number of chunks in the array */
3496 	/* calculate "used chunks per device" */
3497 	size = size * conf->copies;
3498 
3499 	/* We need to round up when dividing by raid_disks to
3500 	 * get the stride size.
3501 	 */
3502 	size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3503 
3504 	conf->dev_sectors = size << conf->geo.chunk_shift;
3505 
3506 	if (conf->geo.far_offset)
3507 		conf->geo.stride = 1 << conf->geo.chunk_shift;
3508 	else {
3509 		sector_div(size, conf->geo.far_copies);
3510 		conf->geo.stride = size << conf->geo.chunk_shift;
3511 	}
3512 }
3513 
3514 enum geo_type {geo_new, geo_old, geo_start};
3515 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3516 {
3517 	int nc, fc, fo;
3518 	int layout, chunk, disks;
3519 	switch (new) {
3520 	case geo_old:
3521 		layout = mddev->layout;
3522 		chunk = mddev->chunk_sectors;
3523 		disks = mddev->raid_disks - mddev->delta_disks;
3524 		break;
3525 	case geo_new:
3526 		layout = mddev->new_layout;
3527 		chunk = mddev->new_chunk_sectors;
3528 		disks = mddev->raid_disks;
3529 		break;
3530 	default: /* avoid 'may be unused' warnings */
3531 	case geo_start: /* new when starting reshape - raid_disks not
3532 			 * updated yet. */
3533 		layout = mddev->new_layout;
3534 		chunk = mddev->new_chunk_sectors;
3535 		disks = mddev->raid_disks + mddev->delta_disks;
3536 		break;
3537 	}
3538 	if (layout >> 19)
3539 		return -1;
3540 	if (chunk < (PAGE_SIZE >> 9) ||
3541 	    !is_power_of_2(chunk))
3542 		return -2;
3543 	nc = layout & 255;
3544 	fc = (layout >> 8) & 255;
3545 	fo = layout & (1<<16);
3546 	geo->raid_disks = disks;
3547 	geo->near_copies = nc;
3548 	geo->far_copies = fc;
3549 	geo->far_offset = fo;
3550 	switch (layout >> 17) {
3551 	case 0:	/* original layout.  simple but not always optimal */
3552 		geo->far_set_size = disks;
3553 		break;
3554 	case 1: /* "improved" layout which was buggy.  Hopefully no-one is
3555 		 * actually using this, but leave code here just in case.*/
3556 		geo->far_set_size = disks/fc;
3557 		WARN(geo->far_set_size < fc,
3558 		     "This RAID10 layout does not provide data safety - please backup and create new array\n");
3559 		break;
3560 	case 2: /* "improved" layout fixed to match documentation */
3561 		geo->far_set_size = fc * nc;
3562 		break;
3563 	default: /* Not a valid layout */
3564 		return -1;
3565 	}
3566 	geo->chunk_mask = chunk - 1;
3567 	geo->chunk_shift = ffz(~chunk);
3568 	return nc*fc;
3569 }
3570 
3571 static struct r10conf *setup_conf(struct mddev *mddev)
3572 {
3573 	struct r10conf *conf = NULL;
3574 	int err = -EINVAL;
3575 	struct geom geo;
3576 	int copies;
3577 
3578 	copies = setup_geo(&geo, mddev, geo_new);
3579 
3580 	if (copies == -2) {
3581 		pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3582 			mdname(mddev), PAGE_SIZE);
3583 		goto out;
3584 	}
3585 
3586 	if (copies < 2 || copies > mddev->raid_disks) {
3587 		pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3588 			mdname(mddev), mddev->new_layout);
3589 		goto out;
3590 	}
3591 
3592 	err = -ENOMEM;
3593 	conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3594 	if (!conf)
3595 		goto out;
3596 
3597 	/* FIXME calc properly */
3598 	conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3599 							    max(0,-mddev->delta_disks)),
3600 				GFP_KERNEL);
3601 	if (!conf->mirrors)
3602 		goto out;
3603 
3604 	conf->tmppage = alloc_page(GFP_KERNEL);
3605 	if (!conf->tmppage)
3606 		goto out;
3607 
3608 	conf->geo = geo;
3609 	conf->copies = copies;
3610 	conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3611 					   r10bio_pool_free, conf);
3612 	if (!conf->r10bio_pool)
3613 		goto out;
3614 
3615 	calc_sectors(conf, mddev->dev_sectors);
3616 	if (mddev->reshape_position == MaxSector) {
3617 		conf->prev = conf->geo;
3618 		conf->reshape_progress = MaxSector;
3619 	} else {
3620 		if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3621 			err = -EINVAL;
3622 			goto out;
3623 		}
3624 		conf->reshape_progress = mddev->reshape_position;
3625 		if (conf->prev.far_offset)
3626 			conf->prev.stride = 1 << conf->prev.chunk_shift;
3627 		else
3628 			/* far_copies must be 1 */
3629 			conf->prev.stride = conf->dev_sectors;
3630 	}
3631 	conf->reshape_safe = conf->reshape_progress;
3632 	spin_lock_init(&conf->device_lock);
3633 	INIT_LIST_HEAD(&conf->retry_list);
3634 	INIT_LIST_HEAD(&conf->bio_end_io_list);
3635 
3636 	spin_lock_init(&conf->resync_lock);
3637 	init_waitqueue_head(&conf->wait_barrier);
3638 	atomic_set(&conf->nr_pending, 0);
3639 
3640 	conf->thread = md_register_thread(raid10d, mddev, "raid10");
3641 	if (!conf->thread)
3642 		goto out;
3643 
3644 	conf->mddev = mddev;
3645 	return conf;
3646 
3647  out:
3648 	if (conf) {
3649 		mempool_destroy(conf->r10bio_pool);
3650 		kfree(conf->mirrors);
3651 		safe_put_page(conf->tmppage);
3652 		kfree(conf);
3653 	}
3654 	return ERR_PTR(err);
3655 }
3656 
3657 static int raid10_run(struct mddev *mddev)
3658 {
3659 	struct r10conf *conf;
3660 	int i, disk_idx, chunk_size;
3661 	struct raid10_info *disk;
3662 	struct md_rdev *rdev;
3663 	sector_t size;
3664 	sector_t min_offset_diff = 0;
3665 	int first = 1;
3666 	bool discard_supported = false;
3667 
3668 	if (mddev->private == NULL) {
3669 		conf = setup_conf(mddev);
3670 		if (IS_ERR(conf))
3671 			return PTR_ERR(conf);
3672 		mddev->private = conf;
3673 	}
3674 	conf = mddev->private;
3675 	if (!conf)
3676 		goto out;
3677 
3678 	mddev->thread = conf->thread;
3679 	conf->thread = NULL;
3680 
3681 	chunk_size = mddev->chunk_sectors << 9;
3682 	if (mddev->queue) {
3683 		blk_queue_max_discard_sectors(mddev->queue,
3684 					      mddev->chunk_sectors);
3685 		blk_queue_max_write_same_sectors(mddev->queue, 0);
3686 		blk_queue_io_min(mddev->queue, chunk_size);
3687 		if (conf->geo.raid_disks % conf->geo.near_copies)
3688 			blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3689 		else
3690 			blk_queue_io_opt(mddev->queue, chunk_size *
3691 					 (conf->geo.raid_disks / conf->geo.near_copies));
3692 	}
3693 
3694 	rdev_for_each(rdev, mddev) {
3695 		long long diff;
3696 		struct request_queue *q;
3697 
3698 		disk_idx = rdev->raid_disk;
3699 		if (disk_idx < 0)
3700 			continue;
3701 		if (disk_idx >= conf->geo.raid_disks &&
3702 		    disk_idx >= conf->prev.raid_disks)
3703 			continue;
3704 		disk = conf->mirrors + disk_idx;
3705 
3706 		if (test_bit(Replacement, &rdev->flags)) {
3707 			if (disk->replacement)
3708 				goto out_free_conf;
3709 			disk->replacement = rdev;
3710 		} else {
3711 			if (disk->rdev)
3712 				goto out_free_conf;
3713 			disk->rdev = rdev;
3714 		}
3715 		q = bdev_get_queue(rdev->bdev);
3716 		diff = (rdev->new_data_offset - rdev->data_offset);
3717 		if (!mddev->reshape_backwards)
3718 			diff = -diff;
3719 		if (diff < 0)
3720 			diff = 0;
3721 		if (first || diff < min_offset_diff)
3722 			min_offset_diff = diff;
3723 
3724 		if (mddev->gendisk)
3725 			disk_stack_limits(mddev->gendisk, rdev->bdev,
3726 					  rdev->data_offset << 9);
3727 
3728 		disk->head_position = 0;
3729 
3730 		if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3731 			discard_supported = true;
3732 	}
3733 
3734 	if (mddev->queue) {
3735 		if (discard_supported)
3736 			queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3737 						mddev->queue);
3738 		else
3739 			queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3740 						  mddev->queue);
3741 	}
3742 	/* need to check that every block has at least one working mirror */
3743 	if (!enough(conf, -1)) {
3744 		pr_err("md/raid10:%s: not enough operational mirrors.\n",
3745 		       mdname(mddev));
3746 		goto out_free_conf;
3747 	}
3748 
3749 	if (conf->reshape_progress != MaxSector) {
3750 		/* must ensure that shape change is supported */
3751 		if (conf->geo.far_copies != 1 &&
3752 		    conf->geo.far_offset == 0)
3753 			goto out_free_conf;
3754 		if (conf->prev.far_copies != 1 &&
3755 		    conf->prev.far_offset == 0)
3756 			goto out_free_conf;
3757 	}
3758 
3759 	mddev->degraded = 0;
3760 	for (i = 0;
3761 	     i < conf->geo.raid_disks
3762 		     || i < conf->prev.raid_disks;
3763 	     i++) {
3764 
3765 		disk = conf->mirrors + i;
3766 
3767 		if (!disk->rdev && disk->replacement) {
3768 			/* The replacement is all we have - use it */
3769 			disk->rdev = disk->replacement;
3770 			disk->replacement = NULL;
3771 			clear_bit(Replacement, &disk->rdev->flags);
3772 		}
3773 
3774 		if (!disk->rdev ||
3775 		    !test_bit(In_sync, &disk->rdev->flags)) {
3776 			disk->head_position = 0;
3777 			mddev->degraded++;
3778 			if (disk->rdev &&
3779 			    disk->rdev->saved_raid_disk < 0)
3780 				conf->fullsync = 1;
3781 		}
3782 		disk->recovery_disabled = mddev->recovery_disabled - 1;
3783 	}
3784 
3785 	if (mddev->recovery_cp != MaxSector)
3786 		pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
3787 			  mdname(mddev));
3788 	pr_info("md/raid10:%s: active with %d out of %d devices\n",
3789 		mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3790 		conf->geo.raid_disks);
3791 	/*
3792 	 * Ok, everything is just fine now
3793 	 */
3794 	mddev->dev_sectors = conf->dev_sectors;
3795 	size = raid10_size(mddev, 0, 0);
3796 	md_set_array_sectors(mddev, size);
3797 	mddev->resync_max_sectors = size;
3798 	set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
3799 
3800 	if (mddev->queue) {
3801 		int stripe = conf->geo.raid_disks *
3802 			((mddev->chunk_sectors << 9) / PAGE_SIZE);
3803 
3804 		/* Calculate max read-ahead size.
3805 		 * We need to readahead at least twice a whole stripe....
3806 		 * maybe...
3807 		 */
3808 		stripe /= conf->geo.near_copies;
3809 		if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3810 			mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3811 	}
3812 
3813 	if (md_integrity_register(mddev))
3814 		goto out_free_conf;
3815 
3816 	if (conf->reshape_progress != MaxSector) {
3817 		unsigned long before_length, after_length;
3818 
3819 		before_length = ((1 << conf->prev.chunk_shift) *
3820 				 conf->prev.far_copies);
3821 		after_length = ((1 << conf->geo.chunk_shift) *
3822 				conf->geo.far_copies);
3823 
3824 		if (max(before_length, after_length) > min_offset_diff) {
3825 			/* This cannot work */
3826 			pr_warn("md/raid10: offset difference not enough to continue reshape\n");
3827 			goto out_free_conf;
3828 		}
3829 		conf->offset_diff = min_offset_diff;
3830 
3831 		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3832 		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3833 		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3834 		set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3835 		mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3836 							"reshape");
3837 	}
3838 
3839 	return 0;
3840 
3841 out_free_conf:
3842 	md_unregister_thread(&mddev->thread);
3843 	mempool_destroy(conf->r10bio_pool);
3844 	safe_put_page(conf->tmppage);
3845 	kfree(conf->mirrors);
3846 	kfree(conf);
3847 	mddev->private = NULL;
3848 out:
3849 	return -EIO;
3850 }
3851 
3852 static void raid10_free(struct mddev *mddev, void *priv)
3853 {
3854 	struct r10conf *conf = priv;
3855 
3856 	mempool_destroy(conf->r10bio_pool);
3857 	safe_put_page(conf->tmppage);
3858 	kfree(conf->mirrors);
3859 	kfree(conf->mirrors_old);
3860 	kfree(conf->mirrors_new);
3861 	kfree(conf);
3862 }
3863 
3864 static void raid10_quiesce(struct mddev *mddev, int state)
3865 {
3866 	struct r10conf *conf = mddev->private;
3867 
3868 	switch(state) {
3869 	case 1:
3870 		raise_barrier(conf, 0);
3871 		break;
3872 	case 0:
3873 		lower_barrier(conf);
3874 		break;
3875 	}
3876 }
3877 
3878 static int raid10_resize(struct mddev *mddev, sector_t sectors)
3879 {
3880 	/* Resize of 'far' arrays is not supported.
3881 	 * For 'near' and 'offset' arrays we can set the
3882 	 * number of sectors used to be an appropriate multiple
3883 	 * of the chunk size.
3884 	 * For 'offset', this is far_copies*chunksize.
3885 	 * For 'near' the multiplier is the LCM of
3886 	 * near_copies and raid_disks.
3887 	 * So if far_copies > 1 && !far_offset, fail.
3888 	 * Else find LCM(raid_disks, near_copy)*far_copies and
3889 	 * multiply by chunk_size.  Then round to this number.
3890 	 * This is mostly done by raid10_size()
3891 	 */
3892 	struct r10conf *conf = mddev->private;
3893 	sector_t oldsize, size;
3894 
3895 	if (mddev->reshape_position != MaxSector)
3896 		return -EBUSY;
3897 
3898 	if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3899 		return -EINVAL;
3900 
3901 	oldsize = raid10_size(mddev, 0, 0);
3902 	size = raid10_size(mddev, sectors, 0);
3903 	if (mddev->external_size &&
3904 	    mddev->array_sectors > size)
3905 		return -EINVAL;
3906 	if (mddev->bitmap) {
3907 		int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3908 		if (ret)
3909 			return ret;
3910 	}
3911 	md_set_array_sectors(mddev, size);
3912 	if (mddev->queue) {
3913 		set_capacity(mddev->gendisk, mddev->array_sectors);
3914 		revalidate_disk(mddev->gendisk);
3915 	}
3916 	if (sectors > mddev->dev_sectors &&
3917 	    mddev->recovery_cp > oldsize) {
3918 		mddev->recovery_cp = oldsize;
3919 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3920 	}
3921 	calc_sectors(conf, sectors);
3922 	mddev->dev_sectors = conf->dev_sectors;
3923 	mddev->resync_max_sectors = size;
3924 	return 0;
3925 }
3926 
3927 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
3928 {
3929 	struct md_rdev *rdev;
3930 	struct r10conf *conf;
3931 
3932 	if (mddev->degraded > 0) {
3933 		pr_warn("md/raid10:%s: Error: degraded raid0!\n",
3934 			mdname(mddev));
3935 		return ERR_PTR(-EINVAL);
3936 	}
3937 	sector_div(size, devs);
3938 
3939 	/* Set new parameters */
3940 	mddev->new_level = 10;
3941 	/* new layout: far_copies = 1, near_copies = 2 */
3942 	mddev->new_layout = (1<<8) + 2;
3943 	mddev->new_chunk_sectors = mddev->chunk_sectors;
3944 	mddev->delta_disks = mddev->raid_disks;
3945 	mddev->raid_disks *= 2;
3946 	/* make sure it will be not marked as dirty */
3947 	mddev->recovery_cp = MaxSector;
3948 	mddev->dev_sectors = size;
3949 
3950 	conf = setup_conf(mddev);
3951 	if (!IS_ERR(conf)) {
3952 		rdev_for_each(rdev, mddev)
3953 			if (rdev->raid_disk >= 0) {
3954 				rdev->new_raid_disk = rdev->raid_disk * 2;
3955 				rdev->sectors = size;
3956 			}
3957 		conf->barrier = 1;
3958 	}
3959 
3960 	return conf;
3961 }
3962 
3963 static void *raid10_takeover(struct mddev *mddev)
3964 {
3965 	struct r0conf *raid0_conf;
3966 
3967 	/* raid10 can take over:
3968 	 *  raid0 - providing it has only two drives
3969 	 */
3970 	if (mddev->level == 0) {
3971 		/* for raid0 takeover only one zone is supported */
3972 		raid0_conf = mddev->private;
3973 		if (raid0_conf->nr_strip_zones > 1) {
3974 			pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
3975 				mdname(mddev));
3976 			return ERR_PTR(-EINVAL);
3977 		}
3978 		return raid10_takeover_raid0(mddev,
3979 			raid0_conf->strip_zone->zone_end,
3980 			raid0_conf->strip_zone->nb_dev);
3981 	}
3982 	return ERR_PTR(-EINVAL);
3983 }
3984 
3985 static int raid10_check_reshape(struct mddev *mddev)
3986 {
3987 	/* Called when there is a request to change
3988 	 * - layout (to ->new_layout)
3989 	 * - chunk size (to ->new_chunk_sectors)
3990 	 * - raid_disks (by delta_disks)
3991 	 * or when trying to restart a reshape that was ongoing.
3992 	 *
3993 	 * We need to validate the request and possibly allocate
3994 	 * space if that might be an issue later.
3995 	 *
3996 	 * Currently we reject any reshape of a 'far' mode array,
3997 	 * allow chunk size to change if new is generally acceptable,
3998 	 * allow raid_disks to increase, and allow
3999 	 * a switch between 'near' mode and 'offset' mode.
4000 	 */
4001 	struct r10conf *conf = mddev->private;
4002 	struct geom geo;
4003 
4004 	if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4005 		return -EINVAL;
4006 
4007 	if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4008 		/* mustn't change number of copies */
4009 		return -EINVAL;
4010 	if (geo.far_copies > 1 && !geo.far_offset)
4011 		/* Cannot switch to 'far' mode */
4012 		return -EINVAL;
4013 
4014 	if (mddev->array_sectors & geo.chunk_mask)
4015 			/* not factor of array size */
4016 			return -EINVAL;
4017 
4018 	if (!enough(conf, -1))
4019 		return -EINVAL;
4020 
4021 	kfree(conf->mirrors_new);
4022 	conf->mirrors_new = NULL;
4023 	if (mddev->delta_disks > 0) {
4024 		/* allocate new 'mirrors' list */
4025 		conf->mirrors_new = kzalloc(
4026 			sizeof(struct raid10_info)
4027 			*(mddev->raid_disks +
4028 			  mddev->delta_disks),
4029 			GFP_KERNEL);
4030 		if (!conf->mirrors_new)
4031 			return -ENOMEM;
4032 	}
4033 	return 0;
4034 }
4035 
4036 /*
4037  * Need to check if array has failed when deciding whether to:
4038  *  - start an array
4039  *  - remove non-faulty devices
4040  *  - add a spare
4041  *  - allow a reshape
4042  * This determination is simple when no reshape is happening.
4043  * However if there is a reshape, we need to carefully check
4044  * both the before and after sections.
4045  * This is because some failed devices may only affect one
4046  * of the two sections, and some non-in_sync devices may
4047  * be insync in the section most affected by failed devices.
4048  */
4049 static int calc_degraded(struct r10conf *conf)
4050 {
4051 	int degraded, degraded2;
4052 	int i;
4053 
4054 	rcu_read_lock();
4055 	degraded = 0;
4056 	/* 'prev' section first */
4057 	for (i = 0; i < conf->prev.raid_disks; i++) {
4058 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4059 		if (!rdev || test_bit(Faulty, &rdev->flags))
4060 			degraded++;
4061 		else if (!test_bit(In_sync, &rdev->flags))
4062 			/* When we can reduce the number of devices in
4063 			 * an array, this might not contribute to
4064 			 * 'degraded'.  It does now.
4065 			 */
4066 			degraded++;
4067 	}
4068 	rcu_read_unlock();
4069 	if (conf->geo.raid_disks == conf->prev.raid_disks)
4070 		return degraded;
4071 	rcu_read_lock();
4072 	degraded2 = 0;
4073 	for (i = 0; i < conf->geo.raid_disks; i++) {
4074 		struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4075 		if (!rdev || test_bit(Faulty, &rdev->flags))
4076 			degraded2++;
4077 		else if (!test_bit(In_sync, &rdev->flags)) {
4078 			/* If reshape is increasing the number of devices,
4079 			 * this section has already been recovered, so
4080 			 * it doesn't contribute to degraded.
4081 			 * else it does.
4082 			 */
4083 			if (conf->geo.raid_disks <= conf->prev.raid_disks)
4084 				degraded2++;
4085 		}
4086 	}
4087 	rcu_read_unlock();
4088 	if (degraded2 > degraded)
4089 		return degraded2;
4090 	return degraded;
4091 }
4092 
4093 static int raid10_start_reshape(struct mddev *mddev)
4094 {
4095 	/* A 'reshape' has been requested. This commits
4096 	 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4097 	 * This also checks if there are enough spares and adds them
4098 	 * to the array.
4099 	 * We currently require enough spares to make the final
4100 	 * array non-degraded.  We also require that the difference
4101 	 * between old and new data_offset - on each device - is
4102 	 * enough that we never risk over-writing.
4103 	 */
4104 
4105 	unsigned long before_length, after_length;
4106 	sector_t min_offset_diff = 0;
4107 	int first = 1;
4108 	struct geom new;
4109 	struct r10conf *conf = mddev->private;
4110 	struct md_rdev *rdev;
4111 	int spares = 0;
4112 	int ret;
4113 
4114 	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4115 		return -EBUSY;
4116 
4117 	if (setup_geo(&new, mddev, geo_start) != conf->copies)
4118 		return -EINVAL;
4119 
4120 	before_length = ((1 << conf->prev.chunk_shift) *
4121 			 conf->prev.far_copies);
4122 	after_length = ((1 << conf->geo.chunk_shift) *
4123 			conf->geo.far_copies);
4124 
4125 	rdev_for_each(rdev, mddev) {
4126 		if (!test_bit(In_sync, &rdev->flags)
4127 		    && !test_bit(Faulty, &rdev->flags))
4128 			spares++;
4129 		if (rdev->raid_disk >= 0) {
4130 			long long diff = (rdev->new_data_offset
4131 					  - rdev->data_offset);
4132 			if (!mddev->reshape_backwards)
4133 				diff = -diff;
4134 			if (diff < 0)
4135 				diff = 0;
4136 			if (first || diff < min_offset_diff)
4137 				min_offset_diff = diff;
4138 		}
4139 	}
4140 
4141 	if (max(before_length, after_length) > min_offset_diff)
4142 		return -EINVAL;
4143 
4144 	if (spares < mddev->delta_disks)
4145 		return -EINVAL;
4146 
4147 	conf->offset_diff = min_offset_diff;
4148 	spin_lock_irq(&conf->device_lock);
4149 	if (conf->mirrors_new) {
4150 		memcpy(conf->mirrors_new, conf->mirrors,
4151 		       sizeof(struct raid10_info)*conf->prev.raid_disks);
4152 		smp_mb();
4153 		kfree(conf->mirrors_old);
4154 		conf->mirrors_old = conf->mirrors;
4155 		conf->mirrors = conf->mirrors_new;
4156 		conf->mirrors_new = NULL;
4157 	}
4158 	setup_geo(&conf->geo, mddev, geo_start);
4159 	smp_mb();
4160 	if (mddev->reshape_backwards) {
4161 		sector_t size = raid10_size(mddev, 0, 0);
4162 		if (size < mddev->array_sectors) {
4163 			spin_unlock_irq(&conf->device_lock);
4164 			pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4165 				mdname(mddev));
4166 			return -EINVAL;
4167 		}
4168 		mddev->resync_max_sectors = size;
4169 		conf->reshape_progress = size;
4170 	} else
4171 		conf->reshape_progress = 0;
4172 	conf->reshape_safe = conf->reshape_progress;
4173 	spin_unlock_irq(&conf->device_lock);
4174 
4175 	if (mddev->delta_disks && mddev->bitmap) {
4176 		ret = bitmap_resize(mddev->bitmap,
4177 				    raid10_size(mddev, 0,
4178 						conf->geo.raid_disks),
4179 				    0, 0);
4180 		if (ret)
4181 			goto abort;
4182 	}
4183 	if (mddev->delta_disks > 0) {
4184 		rdev_for_each(rdev, mddev)
4185 			if (rdev->raid_disk < 0 &&
4186 			    !test_bit(Faulty, &rdev->flags)) {
4187 				if (raid10_add_disk(mddev, rdev) == 0) {
4188 					if (rdev->raid_disk >=
4189 					    conf->prev.raid_disks)
4190 						set_bit(In_sync, &rdev->flags);
4191 					else
4192 						rdev->recovery_offset = 0;
4193 
4194 					if (sysfs_link_rdev(mddev, rdev))
4195 						/* Failure here  is OK */;
4196 				}
4197 			} else if (rdev->raid_disk >= conf->prev.raid_disks
4198 				   && !test_bit(Faulty, &rdev->flags)) {
4199 				/* This is a spare that was manually added */
4200 				set_bit(In_sync, &rdev->flags);
4201 			}
4202 	}
4203 	/* When a reshape changes the number of devices,
4204 	 * ->degraded is measured against the larger of the
4205 	 * pre and  post numbers.
4206 	 */
4207 	spin_lock_irq(&conf->device_lock);
4208 	mddev->degraded = calc_degraded(conf);
4209 	spin_unlock_irq(&conf->device_lock);
4210 	mddev->raid_disks = conf->geo.raid_disks;
4211 	mddev->reshape_position = conf->reshape_progress;
4212 	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4213 
4214 	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4215 	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4216 	clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4217 	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4218 	set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4219 
4220 	mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4221 						"reshape");
4222 	if (!mddev->sync_thread) {
4223 		ret = -EAGAIN;
4224 		goto abort;
4225 	}
4226 	conf->reshape_checkpoint = jiffies;
4227 	md_wakeup_thread(mddev->sync_thread);
4228 	md_new_event(mddev);
4229 	return 0;
4230 
4231 abort:
4232 	mddev->recovery = 0;
4233 	spin_lock_irq(&conf->device_lock);
4234 	conf->geo = conf->prev;
4235 	mddev->raid_disks = conf->geo.raid_disks;
4236 	rdev_for_each(rdev, mddev)
4237 		rdev->new_data_offset = rdev->data_offset;
4238 	smp_wmb();
4239 	conf->reshape_progress = MaxSector;
4240 	conf->reshape_safe = MaxSector;
4241 	mddev->reshape_position = MaxSector;
4242 	spin_unlock_irq(&conf->device_lock);
4243 	return ret;
4244 }
4245 
4246 /* Calculate the last device-address that could contain
4247  * any block from the chunk that includes the array-address 's'
4248  * and report the next address.
4249  * i.e. the address returned will be chunk-aligned and after
4250  * any data that is in the chunk containing 's'.
4251  */
4252 static sector_t last_dev_address(sector_t s, struct geom *geo)
4253 {
4254 	s = (s | geo->chunk_mask) + 1;
4255 	s >>= geo->chunk_shift;
4256 	s *= geo->near_copies;
4257 	s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4258 	s *= geo->far_copies;
4259 	s <<= geo->chunk_shift;
4260 	return s;
4261 }
4262 
4263 /* Calculate the first device-address that could contain
4264  * any block from the chunk that includes the array-address 's'.
4265  * This too will be the start of a chunk
4266  */
4267 static sector_t first_dev_address(sector_t s, struct geom *geo)
4268 {
4269 	s >>= geo->chunk_shift;
4270 	s *= geo->near_copies;
4271 	sector_div(s, geo->raid_disks);
4272 	s *= geo->far_copies;
4273 	s <<= geo->chunk_shift;
4274 	return s;
4275 }
4276 
4277 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4278 				int *skipped)
4279 {
4280 	/* We simply copy at most one chunk (smallest of old and new)
4281 	 * at a time, possibly less if that exceeds RESYNC_PAGES,
4282 	 * or we hit a bad block or something.
4283 	 * This might mean we pause for normal IO in the middle of
4284 	 * a chunk, but that is not a problem as mddev->reshape_position
4285 	 * can record any location.
4286 	 *
4287 	 * If we will want to write to a location that isn't
4288 	 * yet recorded as 'safe' (i.e. in metadata on disk) then
4289 	 * we need to flush all reshape requests and update the metadata.
4290 	 *
4291 	 * When reshaping forwards (e.g. to more devices), we interpret
4292 	 * 'safe' as the earliest block which might not have been copied
4293 	 * down yet.  We divide this by previous stripe size and multiply
4294 	 * by previous stripe length to get lowest device offset that we
4295 	 * cannot write to yet.
4296 	 * We interpret 'sector_nr' as an address that we want to write to.
4297 	 * From this we use last_device_address() to find where we might
4298 	 * write to, and first_device_address on the  'safe' position.
4299 	 * If this 'next' write position is after the 'safe' position,
4300 	 * we must update the metadata to increase the 'safe' position.
4301 	 *
4302 	 * When reshaping backwards, we round in the opposite direction
4303 	 * and perform the reverse test:  next write position must not be
4304 	 * less than current safe position.
4305 	 *
4306 	 * In all this the minimum difference in data offsets
4307 	 * (conf->offset_diff - always positive) allows a bit of slack,
4308 	 * so next can be after 'safe', but not by more than offset_diff
4309 	 *
4310 	 * We need to prepare all the bios here before we start any IO
4311 	 * to ensure the size we choose is acceptable to all devices.
4312 	 * The means one for each copy for write-out and an extra one for
4313 	 * read-in.
4314 	 * We store the read-in bio in ->master_bio and the others in
4315 	 * ->devs[x].bio and ->devs[x].repl_bio.
4316 	 */
4317 	struct r10conf *conf = mddev->private;
4318 	struct r10bio *r10_bio;
4319 	sector_t next, safe, last;
4320 	int max_sectors;
4321 	int nr_sectors;
4322 	int s;
4323 	struct md_rdev *rdev;
4324 	int need_flush = 0;
4325 	struct bio *blist;
4326 	struct bio *bio, *read_bio;
4327 	int sectors_done = 0;
4328 
4329 	if (sector_nr == 0) {
4330 		/* If restarting in the middle, skip the initial sectors */
4331 		if (mddev->reshape_backwards &&
4332 		    conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4333 			sector_nr = (raid10_size(mddev, 0, 0)
4334 				     - conf->reshape_progress);
4335 		} else if (!mddev->reshape_backwards &&
4336 			   conf->reshape_progress > 0)
4337 			sector_nr = conf->reshape_progress;
4338 		if (sector_nr) {
4339 			mddev->curr_resync_completed = sector_nr;
4340 			sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4341 			*skipped = 1;
4342 			return sector_nr;
4343 		}
4344 	}
4345 
4346 	/* We don't use sector_nr to track where we are up to
4347 	 * as that doesn't work well for ->reshape_backwards.
4348 	 * So just use ->reshape_progress.
4349 	 */
4350 	if (mddev->reshape_backwards) {
4351 		/* 'next' is the earliest device address that we might
4352 		 * write to for this chunk in the new layout
4353 		 */
4354 		next = first_dev_address(conf->reshape_progress - 1,
4355 					 &conf->geo);
4356 
4357 		/* 'safe' is the last device address that we might read from
4358 		 * in the old layout after a restart
4359 		 */
4360 		safe = last_dev_address(conf->reshape_safe - 1,
4361 					&conf->prev);
4362 
4363 		if (next + conf->offset_diff < safe)
4364 			need_flush = 1;
4365 
4366 		last = conf->reshape_progress - 1;
4367 		sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4368 					       & conf->prev.chunk_mask);
4369 		if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4370 			sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4371 	} else {
4372 		/* 'next' is after the last device address that we
4373 		 * might write to for this chunk in the new layout
4374 		 */
4375 		next = last_dev_address(conf->reshape_progress, &conf->geo);
4376 
4377 		/* 'safe' is the earliest device address that we might
4378 		 * read from in the old layout after a restart
4379 		 */
4380 		safe = first_dev_address(conf->reshape_safe, &conf->prev);
4381 
4382 		/* Need to update metadata if 'next' might be beyond 'safe'
4383 		 * as that would possibly corrupt data
4384 		 */
4385 		if (next > safe + conf->offset_diff)
4386 			need_flush = 1;
4387 
4388 		sector_nr = conf->reshape_progress;
4389 		last  = sector_nr | (conf->geo.chunk_mask
4390 				     & conf->prev.chunk_mask);
4391 
4392 		if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4393 			last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4394 	}
4395 
4396 	if (need_flush ||
4397 	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4398 		/* Need to update reshape_position in metadata */
4399 		wait_barrier(conf);
4400 		mddev->reshape_position = conf->reshape_progress;
4401 		if (mddev->reshape_backwards)
4402 			mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4403 				- conf->reshape_progress;
4404 		else
4405 			mddev->curr_resync_completed = conf->reshape_progress;
4406 		conf->reshape_checkpoint = jiffies;
4407 		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4408 		md_wakeup_thread(mddev->thread);
4409 		wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4410 			   test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4411 		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4412 			allow_barrier(conf);
4413 			return sectors_done;
4414 		}
4415 		conf->reshape_safe = mddev->reshape_position;
4416 		allow_barrier(conf);
4417 	}
4418 
4419 read_more:
4420 	/* Now schedule reads for blocks from sector_nr to last */
4421 	r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4422 	r10_bio->state = 0;
4423 	raise_barrier(conf, sectors_done != 0);
4424 	atomic_set(&r10_bio->remaining, 0);
4425 	r10_bio->mddev = mddev;
4426 	r10_bio->sector = sector_nr;
4427 	set_bit(R10BIO_IsReshape, &r10_bio->state);
4428 	r10_bio->sectors = last - sector_nr + 1;
4429 	rdev = read_balance(conf, r10_bio, &max_sectors);
4430 	BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4431 
4432 	if (!rdev) {
4433 		/* Cannot read from here, so need to record bad blocks
4434 		 * on all the target devices.
4435 		 */
4436 		// FIXME
4437 		mempool_free(r10_bio, conf->r10buf_pool);
4438 		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4439 		return sectors_done;
4440 	}
4441 
4442 	read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4443 
4444 	read_bio->bi_bdev = rdev->bdev;
4445 	read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4446 			       + rdev->data_offset);
4447 	read_bio->bi_private = r10_bio;
4448 	read_bio->bi_end_io = end_sync_read;
4449 	bio_set_op_attrs(read_bio, REQ_OP_READ, 0);
4450 	read_bio->bi_flags &= (~0UL << BIO_RESET_BITS);
4451 	read_bio->bi_error = 0;
4452 	read_bio->bi_vcnt = 0;
4453 	read_bio->bi_iter.bi_size = 0;
4454 	r10_bio->master_bio = read_bio;
4455 	r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4456 
4457 	/* Now find the locations in the new layout */
4458 	__raid10_find_phys(&conf->geo, r10_bio);
4459 
4460 	blist = read_bio;
4461 	read_bio->bi_next = NULL;
4462 
4463 	rcu_read_lock();
4464 	for (s = 0; s < conf->copies*2; s++) {
4465 		struct bio *b;
4466 		int d = r10_bio->devs[s/2].devnum;
4467 		struct md_rdev *rdev2;
4468 		if (s&1) {
4469 			rdev2 = rcu_dereference(conf->mirrors[d].replacement);
4470 			b = r10_bio->devs[s/2].repl_bio;
4471 		} else {
4472 			rdev2 = rcu_dereference(conf->mirrors[d].rdev);
4473 			b = r10_bio->devs[s/2].bio;
4474 		}
4475 		if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4476 			continue;
4477 
4478 		bio_reset(b);
4479 		b->bi_bdev = rdev2->bdev;
4480 		b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4481 			rdev2->new_data_offset;
4482 		b->bi_private = r10_bio;
4483 		b->bi_end_io = end_reshape_write;
4484 		bio_set_op_attrs(b, REQ_OP_WRITE, 0);
4485 		b->bi_next = blist;
4486 		blist = b;
4487 	}
4488 
4489 	/* Now add as many pages as possible to all of these bios. */
4490 
4491 	nr_sectors = 0;
4492 	for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4493 		struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4494 		int len = (max_sectors - s) << 9;
4495 		if (len > PAGE_SIZE)
4496 			len = PAGE_SIZE;
4497 		for (bio = blist; bio ; bio = bio->bi_next) {
4498 			struct bio *bio2;
4499 			if (bio_add_page(bio, page, len, 0))
4500 				continue;
4501 
4502 			/* Didn't fit, must stop */
4503 			for (bio2 = blist;
4504 			     bio2 && bio2 != bio;
4505 			     bio2 = bio2->bi_next) {
4506 				/* Remove last page from this bio */
4507 				bio2->bi_vcnt--;
4508 				bio2->bi_iter.bi_size -= len;
4509 				bio_clear_flag(bio2, BIO_SEG_VALID);
4510 			}
4511 			goto bio_full;
4512 		}
4513 		sector_nr += len >> 9;
4514 		nr_sectors += len >> 9;
4515 	}
4516 bio_full:
4517 	rcu_read_unlock();
4518 	r10_bio->sectors = nr_sectors;
4519 
4520 	/* Now submit the read */
4521 	md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4522 	atomic_inc(&r10_bio->remaining);
4523 	read_bio->bi_next = NULL;
4524 	generic_make_request(read_bio);
4525 	sector_nr += nr_sectors;
4526 	sectors_done += nr_sectors;
4527 	if (sector_nr <= last)
4528 		goto read_more;
4529 
4530 	/* Now that we have done the whole section we can
4531 	 * update reshape_progress
4532 	 */
4533 	if (mddev->reshape_backwards)
4534 		conf->reshape_progress -= sectors_done;
4535 	else
4536 		conf->reshape_progress += sectors_done;
4537 
4538 	return sectors_done;
4539 }
4540 
4541 static void end_reshape_request(struct r10bio *r10_bio);
4542 static int handle_reshape_read_error(struct mddev *mddev,
4543 				     struct r10bio *r10_bio);
4544 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4545 {
4546 	/* Reshape read completed.  Hopefully we have a block
4547 	 * to write out.
4548 	 * If we got a read error then we do sync 1-page reads from
4549 	 * elsewhere until we find the data - or give up.
4550 	 */
4551 	struct r10conf *conf = mddev->private;
4552 	int s;
4553 
4554 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4555 		if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4556 			/* Reshape has been aborted */
4557 			md_done_sync(mddev, r10_bio->sectors, 0);
4558 			return;
4559 		}
4560 
4561 	/* We definitely have the data in the pages, schedule the
4562 	 * writes.
4563 	 */
4564 	atomic_set(&r10_bio->remaining, 1);
4565 	for (s = 0; s < conf->copies*2; s++) {
4566 		struct bio *b;
4567 		int d = r10_bio->devs[s/2].devnum;
4568 		struct md_rdev *rdev;
4569 		rcu_read_lock();
4570 		if (s&1) {
4571 			rdev = rcu_dereference(conf->mirrors[d].replacement);
4572 			b = r10_bio->devs[s/2].repl_bio;
4573 		} else {
4574 			rdev = rcu_dereference(conf->mirrors[d].rdev);
4575 			b = r10_bio->devs[s/2].bio;
4576 		}
4577 		if (!rdev || test_bit(Faulty, &rdev->flags)) {
4578 			rcu_read_unlock();
4579 			continue;
4580 		}
4581 		atomic_inc(&rdev->nr_pending);
4582 		rcu_read_unlock();
4583 		md_sync_acct(b->bi_bdev, r10_bio->sectors);
4584 		atomic_inc(&r10_bio->remaining);
4585 		b->bi_next = NULL;
4586 		generic_make_request(b);
4587 	}
4588 	end_reshape_request(r10_bio);
4589 }
4590 
4591 static void end_reshape(struct r10conf *conf)
4592 {
4593 	if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4594 		return;
4595 
4596 	spin_lock_irq(&conf->device_lock);
4597 	conf->prev = conf->geo;
4598 	md_finish_reshape(conf->mddev);
4599 	smp_wmb();
4600 	conf->reshape_progress = MaxSector;
4601 	conf->reshape_safe = MaxSector;
4602 	spin_unlock_irq(&conf->device_lock);
4603 
4604 	/* read-ahead size must cover two whole stripes, which is
4605 	 * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4606 	 */
4607 	if (conf->mddev->queue) {
4608 		int stripe = conf->geo.raid_disks *
4609 			((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4610 		stripe /= conf->geo.near_copies;
4611 		if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4612 			conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4613 	}
4614 	conf->fullsync = 0;
4615 }
4616 
4617 static int handle_reshape_read_error(struct mddev *mddev,
4618 				     struct r10bio *r10_bio)
4619 {
4620 	/* Use sync reads to get the blocks from somewhere else */
4621 	int sectors = r10_bio->sectors;
4622 	struct r10conf *conf = mddev->private;
4623 	struct {
4624 		struct r10bio r10_bio;
4625 		struct r10dev devs[conf->copies];
4626 	} on_stack;
4627 	struct r10bio *r10b = &on_stack.r10_bio;
4628 	int slot = 0;
4629 	int idx = 0;
4630 	struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4631 
4632 	r10b->sector = r10_bio->sector;
4633 	__raid10_find_phys(&conf->prev, r10b);
4634 
4635 	while (sectors) {
4636 		int s = sectors;
4637 		int success = 0;
4638 		int first_slot = slot;
4639 
4640 		if (s > (PAGE_SIZE >> 9))
4641 			s = PAGE_SIZE >> 9;
4642 
4643 		rcu_read_lock();
4644 		while (!success) {
4645 			int d = r10b->devs[slot].devnum;
4646 			struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4647 			sector_t addr;
4648 			if (rdev == NULL ||
4649 			    test_bit(Faulty, &rdev->flags) ||
4650 			    !test_bit(In_sync, &rdev->flags))
4651 				goto failed;
4652 
4653 			addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4654 			atomic_inc(&rdev->nr_pending);
4655 			rcu_read_unlock();
4656 			success = sync_page_io(rdev,
4657 					       addr,
4658 					       s << 9,
4659 					       bvec[idx].bv_page,
4660 					       REQ_OP_READ, 0, false);
4661 			rdev_dec_pending(rdev, mddev);
4662 			rcu_read_lock();
4663 			if (success)
4664 				break;
4665 		failed:
4666 			slot++;
4667 			if (slot >= conf->copies)
4668 				slot = 0;
4669 			if (slot == first_slot)
4670 				break;
4671 		}
4672 		rcu_read_unlock();
4673 		if (!success) {
4674 			/* couldn't read this block, must give up */
4675 			set_bit(MD_RECOVERY_INTR,
4676 				&mddev->recovery);
4677 			return -EIO;
4678 		}
4679 		sectors -= s;
4680 		idx++;
4681 	}
4682 	return 0;
4683 }
4684 
4685 static void end_reshape_write(struct bio *bio)
4686 {
4687 	struct r10bio *r10_bio = bio->bi_private;
4688 	struct mddev *mddev = r10_bio->mddev;
4689 	struct r10conf *conf = mddev->private;
4690 	int d;
4691 	int slot;
4692 	int repl;
4693 	struct md_rdev *rdev = NULL;
4694 
4695 	d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4696 	if (repl)
4697 		rdev = conf->mirrors[d].replacement;
4698 	if (!rdev) {
4699 		smp_mb();
4700 		rdev = conf->mirrors[d].rdev;
4701 	}
4702 
4703 	if (bio->bi_error) {
4704 		/* FIXME should record badblock */
4705 		md_error(mddev, rdev);
4706 	}
4707 
4708 	rdev_dec_pending(rdev, mddev);
4709 	end_reshape_request(r10_bio);
4710 }
4711 
4712 static void end_reshape_request(struct r10bio *r10_bio)
4713 {
4714 	if (!atomic_dec_and_test(&r10_bio->remaining))
4715 		return;
4716 	md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4717 	bio_put(r10_bio->master_bio);
4718 	put_buf(r10_bio);
4719 }
4720 
4721 static void raid10_finish_reshape(struct mddev *mddev)
4722 {
4723 	struct r10conf *conf = mddev->private;
4724 
4725 	if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4726 		return;
4727 
4728 	if (mddev->delta_disks > 0) {
4729 		sector_t size = raid10_size(mddev, 0, 0);
4730 		md_set_array_sectors(mddev, size);
4731 		if (mddev->recovery_cp > mddev->resync_max_sectors) {
4732 			mddev->recovery_cp = mddev->resync_max_sectors;
4733 			set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4734 		}
4735 		mddev->resync_max_sectors = size;
4736 		if (mddev->queue) {
4737 			set_capacity(mddev->gendisk, mddev->array_sectors);
4738 			revalidate_disk(mddev->gendisk);
4739 		}
4740 	} else {
4741 		int d;
4742 		rcu_read_lock();
4743 		for (d = conf->geo.raid_disks ;
4744 		     d < conf->geo.raid_disks - mddev->delta_disks;
4745 		     d++) {
4746 			struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
4747 			if (rdev)
4748 				clear_bit(In_sync, &rdev->flags);
4749 			rdev = rcu_dereference(conf->mirrors[d].replacement);
4750 			if (rdev)
4751 				clear_bit(In_sync, &rdev->flags);
4752 		}
4753 		rcu_read_unlock();
4754 	}
4755 	mddev->layout = mddev->new_layout;
4756 	mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4757 	mddev->reshape_position = MaxSector;
4758 	mddev->delta_disks = 0;
4759 	mddev->reshape_backwards = 0;
4760 }
4761 
4762 static struct md_personality raid10_personality =
4763 {
4764 	.name		= "raid10",
4765 	.level		= 10,
4766 	.owner		= THIS_MODULE,
4767 	.make_request	= raid10_make_request,
4768 	.run		= raid10_run,
4769 	.free		= raid10_free,
4770 	.status		= raid10_status,
4771 	.error_handler	= raid10_error,
4772 	.hot_add_disk	= raid10_add_disk,
4773 	.hot_remove_disk= raid10_remove_disk,
4774 	.spare_active	= raid10_spare_active,
4775 	.sync_request	= raid10_sync_request,
4776 	.quiesce	= raid10_quiesce,
4777 	.size		= raid10_size,
4778 	.resize		= raid10_resize,
4779 	.takeover	= raid10_takeover,
4780 	.check_reshape	= raid10_check_reshape,
4781 	.start_reshape	= raid10_start_reshape,
4782 	.finish_reshape	= raid10_finish_reshape,
4783 	.congested	= raid10_congested,
4784 };
4785 
4786 static int __init raid_init(void)
4787 {
4788 	return register_md_personality(&raid10_personality);
4789 }
4790 
4791 static void raid_exit(void)
4792 {
4793 	unregister_md_personality(&raid10_personality);
4794 }
4795 
4796 module_init(raid_init);
4797 module_exit(raid_exit);
4798 MODULE_LICENSE("GPL");
4799 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4800 MODULE_ALIAS("md-personality-9"); /* RAID10 */
4801 MODULE_ALIAS("md-raid10");
4802 MODULE_ALIAS("md-level-10");
4803 
4804 module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
4805