xref: /linux/drivers/md/raid10.c (revision 9b960d8cd6f712cb2c03e2bdd4d5ca058238037f)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * raid10.c : Multiple Devices driver for Linux
4  *
5  * Copyright (C) 2000-2004 Neil Brown
6  *
7  * RAID-10 support for md.
8  *
9  * Base on code in raid1.c.  See raid1.c for further copyright information.
10  */
11 
12 #include <linux/slab.h>
13 #include <linux/delay.h>
14 #include <linux/blkdev.h>
15 #include <linux/module.h>
16 #include <linux/seq_file.h>
17 #include <linux/ratelimit.h>
18 #include <linux/kthread.h>
19 #include <linux/raid/md_p.h>
20 #include <trace/events/block.h>
21 #include "md.h"
22 
23 #define RAID_1_10_NAME "raid10"
24 #include "raid10.h"
25 #include "raid0.h"
26 #include "md-bitmap.h"
27 #include "md-cluster.h"
28 
29 /*
30  * RAID10 provides a combination of RAID0 and RAID1 functionality.
31  * The layout of data is defined by
32  *    chunk_size
33  *    raid_disks
34  *    near_copies (stored in low byte of layout)
35  *    far_copies (stored in second byte of layout)
36  *    far_offset (stored in bit 16 of layout )
37  *    use_far_sets (stored in bit 17 of layout )
38  *    use_far_sets_bugfixed (stored in bit 18 of layout )
39  *
40  * The data to be stored is divided into chunks using chunksize.  Each device
41  * is divided into far_copies sections.   In each section, chunks are laid out
42  * in a style similar to raid0, but near_copies copies of each chunk is stored
43  * (each on a different drive).  The starting device for each section is offset
44  * near_copies from the starting device of the previous section.  Thus there
45  * are (near_copies * far_copies) of each chunk, and each is on a different
46  * drive.  near_copies and far_copies must be at least one, and their product
47  * is at most raid_disks.
48  *
49  * If far_offset is true, then the far_copies are handled a bit differently.
50  * The copies are still in different stripes, but instead of being very far
51  * apart on disk, there are adjacent stripes.
52  *
53  * The far and offset algorithms are handled slightly differently if
54  * 'use_far_sets' is true.  In this case, the array's devices are grouped into
55  * sets that are (near_copies * far_copies) in size.  The far copied stripes
56  * are still shifted by 'near_copies' devices, but this shifting stays confined
57  * to the set rather than the entire array.  This is done to improve the number
58  * of device combinations that can fail without causing the array to fail.
59  * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
60  * on a device):
61  *    A B C D    A B C D E
62  *      ...         ...
63  *    D A B C    E A B C D
64  * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
65  *    [A B] [C D]    [A B] [C D E]
66  *    |...| |...|    |...| | ... |
67  *    [B A] [D C]    [B A] [E C D]
68  */
69 
70 static void allow_barrier(struct r10conf *conf);
71 static void lower_barrier(struct r10conf *conf);
72 static int _enough(struct r10conf *conf, int previous, int ignore);
73 static int enough(struct r10conf *conf, int ignore);
74 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
75 				int *skipped);
76 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
77 static void end_reshape_write(struct bio *bio);
78 static void end_reshape(struct r10conf *conf);
79 
80 #include "raid1-10.c"
81 
82 #define NULL_CMD
83 #define cmd_before(conf, cmd) \
84 	do { \
85 		write_sequnlock_irq(&(conf)->resync_lock); \
86 		cmd; \
87 	} while (0)
88 #define cmd_after(conf) write_seqlock_irq(&(conf)->resync_lock)
89 
90 #define wait_event_barrier_cmd(conf, cond, cmd) \
91 	wait_event_cmd((conf)->wait_barrier, cond, cmd_before(conf, cmd), \
92 		       cmd_after(conf))
93 
94 #define wait_event_barrier(conf, cond) \
95 	wait_event_barrier_cmd(conf, cond, NULL_CMD)
96 
97 /*
98  * for resync bio, r10bio pointer can be retrieved from the per-bio
99  * 'struct resync_pages'.
100  */
get_resync_r10bio(struct bio * bio)101 static inline struct r10bio *get_resync_r10bio(struct bio *bio)
102 {
103 	return get_resync_pages(bio)->raid_bio;
104 }
105 
r10bio_pool_alloc(gfp_t gfp_flags,void * data)106 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
107 {
108 	struct r10conf *conf = data;
109 	int size = offsetof(struct r10bio, devs[conf->geo.raid_disks]);
110 
111 	/* allocate a r10bio with room for raid_disks entries in the
112 	 * bios array */
113 	return kzalloc(size, gfp_flags);
114 }
115 
116 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
117 /* amount of memory to reserve for resync requests */
118 #define RESYNC_WINDOW (1024*1024)
119 /* maximum number of concurrent requests, memory permitting */
120 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
121 #define CLUSTER_RESYNC_WINDOW (32 * RESYNC_WINDOW)
122 #define CLUSTER_RESYNC_WINDOW_SECTORS (CLUSTER_RESYNC_WINDOW >> 9)
123 
124 /*
125  * When performing a resync, we need to read and compare, so
126  * we need as many pages are there are copies.
127  * When performing a recovery, we need 2 bios, one for read,
128  * one for write (we recover only one drive per r10buf)
129  *
130  */
r10buf_pool_alloc(gfp_t gfp_flags,void * data)131 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
132 {
133 	struct r10conf *conf = data;
134 	struct r10bio *r10_bio;
135 	struct bio *bio;
136 	int j;
137 	int nalloc, nalloc_rp;
138 	struct resync_pages *rps;
139 
140 	r10_bio = r10bio_pool_alloc(gfp_flags, conf);
141 	if (!r10_bio)
142 		return NULL;
143 
144 	if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
145 	    test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
146 		nalloc = conf->copies; /* resync */
147 	else
148 		nalloc = 2; /* recovery */
149 
150 	/* allocate once for all bios */
151 	if (!conf->have_replacement)
152 		nalloc_rp = nalloc;
153 	else
154 		nalloc_rp = nalloc * 2;
155 	rps = kmalloc_array(nalloc_rp, sizeof(struct resync_pages), gfp_flags);
156 	if (!rps)
157 		goto out_free_r10bio;
158 
159 	/*
160 	 * Allocate bios.
161 	 */
162 	for (j = nalloc ; j-- ; ) {
163 		bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
164 		if (!bio)
165 			goto out_free_bio;
166 		bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
167 		r10_bio->devs[j].bio = bio;
168 		if (!conf->have_replacement)
169 			continue;
170 		bio = bio_kmalloc(RESYNC_PAGES, gfp_flags);
171 		if (!bio)
172 			goto out_free_bio;
173 		bio_init(bio, NULL, bio->bi_inline_vecs, RESYNC_PAGES, 0);
174 		r10_bio->devs[j].repl_bio = bio;
175 	}
176 	/*
177 	 * Allocate RESYNC_PAGES data pages and attach them
178 	 * where needed.
179 	 */
180 	for (j = 0; j < nalloc; j++) {
181 		struct bio *rbio = r10_bio->devs[j].repl_bio;
182 		struct resync_pages *rp, *rp_repl;
183 
184 		rp = &rps[j];
185 		if (rbio)
186 			rp_repl = &rps[nalloc + j];
187 
188 		bio = r10_bio->devs[j].bio;
189 
190 		if (!j || test_bit(MD_RECOVERY_SYNC,
191 				   &conf->mddev->recovery)) {
192 			if (resync_alloc_pages(rp, gfp_flags))
193 				goto out_free_pages;
194 		} else {
195 			memcpy(rp, &rps[0], sizeof(*rp));
196 			resync_get_all_pages(rp);
197 		}
198 
199 		rp->raid_bio = r10_bio;
200 		bio->bi_private = rp;
201 		if (rbio) {
202 			memcpy(rp_repl, rp, sizeof(*rp));
203 			rbio->bi_private = rp_repl;
204 		}
205 	}
206 
207 	return r10_bio;
208 
209 out_free_pages:
210 	while (--j >= 0)
211 		resync_free_pages(&rps[j]);
212 
213 	j = 0;
214 out_free_bio:
215 	for ( ; j < nalloc; j++) {
216 		if (r10_bio->devs[j].bio)
217 			bio_uninit(r10_bio->devs[j].bio);
218 		kfree(r10_bio->devs[j].bio);
219 		if (r10_bio->devs[j].repl_bio)
220 			bio_uninit(r10_bio->devs[j].repl_bio);
221 		kfree(r10_bio->devs[j].repl_bio);
222 	}
223 	kfree(rps);
224 out_free_r10bio:
225 	rbio_pool_free(r10_bio, conf);
226 	return NULL;
227 }
228 
r10buf_pool_free(void * __r10_bio,void * data)229 static void r10buf_pool_free(void *__r10_bio, void *data)
230 {
231 	struct r10conf *conf = data;
232 	struct r10bio *r10bio = __r10_bio;
233 	int j;
234 	struct resync_pages *rp = NULL;
235 
236 	for (j = conf->copies; j--; ) {
237 		struct bio *bio = r10bio->devs[j].bio;
238 
239 		if (bio) {
240 			rp = get_resync_pages(bio);
241 			resync_free_pages(rp);
242 			bio_uninit(bio);
243 			kfree(bio);
244 		}
245 
246 		bio = r10bio->devs[j].repl_bio;
247 		if (bio) {
248 			bio_uninit(bio);
249 			kfree(bio);
250 		}
251 	}
252 
253 	/* resync pages array stored in the 1st bio's .bi_private */
254 	kfree(rp);
255 
256 	rbio_pool_free(r10bio, conf);
257 }
258 
put_all_bios(struct r10conf * conf,struct r10bio * r10_bio)259 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
260 {
261 	int i;
262 
263 	for (i = 0; i < conf->geo.raid_disks; i++) {
264 		struct bio **bio = & r10_bio->devs[i].bio;
265 		if (!BIO_SPECIAL(*bio))
266 			bio_put(*bio);
267 		*bio = NULL;
268 		bio = &r10_bio->devs[i].repl_bio;
269 		if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
270 			bio_put(*bio);
271 		*bio = NULL;
272 	}
273 }
274 
free_r10bio(struct r10bio * r10_bio)275 static void free_r10bio(struct r10bio *r10_bio)
276 {
277 	struct r10conf *conf = r10_bio->mddev->private;
278 
279 	put_all_bios(conf, r10_bio);
280 	mempool_free(r10_bio, &conf->r10bio_pool);
281 }
282 
put_buf(struct r10bio * r10_bio)283 static void put_buf(struct r10bio *r10_bio)
284 {
285 	struct r10conf *conf = r10_bio->mddev->private;
286 
287 	mempool_free(r10_bio, &conf->r10buf_pool);
288 
289 	lower_barrier(conf);
290 }
291 
wake_up_barrier(struct r10conf * conf)292 static void wake_up_barrier(struct r10conf *conf)
293 {
294 	if (wq_has_sleeper(&conf->wait_barrier))
295 		wake_up(&conf->wait_barrier);
296 }
297 
reschedule_retry(struct r10bio * r10_bio)298 static void reschedule_retry(struct r10bio *r10_bio)
299 {
300 	unsigned long flags;
301 	struct mddev *mddev = r10_bio->mddev;
302 	struct r10conf *conf = mddev->private;
303 
304 	spin_lock_irqsave(&conf->device_lock, flags);
305 	list_add(&r10_bio->retry_list, &conf->retry_list);
306 	conf->nr_queued ++;
307 	spin_unlock_irqrestore(&conf->device_lock, flags);
308 
309 	/* wake up frozen array... */
310 	wake_up(&conf->wait_barrier);
311 
312 	md_wakeup_thread(mddev->thread);
313 }
314 
315 /*
316  * raid_end_bio_io() is called when we have finished servicing a mirrored
317  * operation and are ready to return a success/failure code to the buffer
318  * cache layer.
319  */
raid_end_bio_io(struct r10bio * r10_bio)320 static void raid_end_bio_io(struct r10bio *r10_bio)
321 {
322 	struct bio *bio = r10_bio->master_bio;
323 	struct r10conf *conf = r10_bio->mddev->private;
324 
325 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
326 		bio->bi_status = BLK_STS_IOERR;
327 
328 	bio_endio(bio);
329 	/*
330 	 * Wake up any possible resync thread that waits for the device
331 	 * to go idle.
332 	 */
333 	allow_barrier(conf);
334 
335 	free_r10bio(r10_bio);
336 }
337 
338 /*
339  * Update disk head position estimator based on IRQ completion info.
340  */
update_head_pos(int slot,struct r10bio * r10_bio)341 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
342 {
343 	struct r10conf *conf = r10_bio->mddev->private;
344 
345 	conf->mirrors[r10_bio->devs[slot].devnum].head_position =
346 		r10_bio->devs[slot].addr + (r10_bio->sectors);
347 }
348 
349 /*
350  * Find the disk number which triggered given bio
351  */
find_bio_disk(struct r10conf * conf,struct r10bio * r10_bio,struct bio * bio,int * slotp,int * replp)352 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
353 			 struct bio *bio, int *slotp, int *replp)
354 {
355 	int slot;
356 	int repl = 0;
357 
358 	for (slot = 0; slot < conf->geo.raid_disks; slot++) {
359 		if (r10_bio->devs[slot].bio == bio)
360 			break;
361 		if (r10_bio->devs[slot].repl_bio == bio) {
362 			repl = 1;
363 			break;
364 		}
365 	}
366 
367 	update_head_pos(slot, r10_bio);
368 
369 	if (slotp)
370 		*slotp = slot;
371 	if (replp)
372 		*replp = repl;
373 	return r10_bio->devs[slot].devnum;
374 }
375 
raid10_end_read_request(struct bio * bio)376 static void raid10_end_read_request(struct bio *bio)
377 {
378 	int uptodate = !bio->bi_status;
379 	struct r10bio *r10_bio = bio->bi_private;
380 	int slot;
381 	struct md_rdev *rdev;
382 	struct r10conf *conf = r10_bio->mddev->private;
383 
384 	slot = r10_bio->read_slot;
385 	rdev = r10_bio->devs[slot].rdev;
386 	/*
387 	 * this branch is our 'one mirror IO has finished' event handler:
388 	 */
389 	update_head_pos(slot, r10_bio);
390 
391 	if (uptodate) {
392 		/*
393 		 * Set R10BIO_Uptodate in our master bio, so that
394 		 * we will return a good error code to the higher
395 		 * levels even if IO on some other mirrored buffer fails.
396 		 *
397 		 * The 'master' represents the composite IO operation to
398 		 * user-side. So if something waits for IO, then it will
399 		 * wait for the 'master' bio.
400 		 */
401 		set_bit(R10BIO_Uptodate, &r10_bio->state);
402 	} else {
403 		/* If all other devices that store this block have
404 		 * failed, we want to return the error upwards rather
405 		 * than fail the last device.  Here we redefine
406 		 * "uptodate" to mean "Don't want to retry"
407 		 */
408 		if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
409 			     rdev->raid_disk))
410 			uptodate = 1;
411 	}
412 	if (uptodate) {
413 		raid_end_bio_io(r10_bio);
414 		rdev_dec_pending(rdev, conf->mddev);
415 	} else {
416 		/*
417 		 * oops, read error - keep the refcount on the rdev
418 		 */
419 		pr_err_ratelimited("md/raid10:%s: %pg: rescheduling sector %llu\n",
420 				   mdname(conf->mddev),
421 				   rdev->bdev,
422 				   (unsigned long long)r10_bio->sector);
423 		set_bit(R10BIO_ReadError, &r10_bio->state);
424 		reschedule_retry(r10_bio);
425 	}
426 }
427 
close_write(struct r10bio * r10_bio)428 static void close_write(struct r10bio *r10_bio)
429 {
430 	struct mddev *mddev = r10_bio->mddev;
431 
432 	md_write_end(mddev);
433 }
434 
one_write_done(struct r10bio * r10_bio)435 static void one_write_done(struct r10bio *r10_bio)
436 {
437 	if (atomic_dec_and_test(&r10_bio->remaining)) {
438 		if (test_bit(R10BIO_WriteError, &r10_bio->state))
439 			reschedule_retry(r10_bio);
440 		else {
441 			close_write(r10_bio);
442 			if (test_bit(R10BIO_MadeGood, &r10_bio->state))
443 				reschedule_retry(r10_bio);
444 			else
445 				raid_end_bio_io(r10_bio);
446 		}
447 	}
448 }
449 
raid10_end_write_request(struct bio * bio)450 static void raid10_end_write_request(struct bio *bio)
451 {
452 	struct r10bio *r10_bio = bio->bi_private;
453 	int dev;
454 	int dec_rdev = 1;
455 	struct r10conf *conf = r10_bio->mddev->private;
456 	int slot, repl;
457 	struct md_rdev *rdev = NULL;
458 	struct bio *to_put = NULL;
459 	bool discard_error;
460 
461 	discard_error = bio->bi_status && bio_op(bio) == REQ_OP_DISCARD;
462 
463 	dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
464 
465 	if (repl)
466 		rdev = conf->mirrors[dev].replacement;
467 	if (!rdev) {
468 		smp_rmb();
469 		repl = 0;
470 		rdev = conf->mirrors[dev].rdev;
471 	}
472 	/*
473 	 * this branch is our 'one mirror IO has finished' event handler:
474 	 */
475 	if (bio->bi_status && !discard_error) {
476 		if (repl)
477 			/* Never record new bad blocks to replacement,
478 			 * just fail it.
479 			 */
480 			md_error(rdev->mddev, rdev);
481 		else {
482 			set_bit(WriteErrorSeen,	&rdev->flags);
483 			if (!test_and_set_bit(WantReplacement, &rdev->flags))
484 				set_bit(MD_RECOVERY_NEEDED,
485 					&rdev->mddev->recovery);
486 
487 			dec_rdev = 0;
488 			if (test_bit(FailFast, &rdev->flags) &&
489 			    (bio->bi_opf & MD_FAILFAST)) {
490 				md_error(rdev->mddev, rdev);
491 			}
492 
493 			/*
494 			 * When the device is faulty, it is not necessary to
495 			 * handle write error.
496 			 */
497 			if (!test_bit(Faulty, &rdev->flags))
498 				set_bit(R10BIO_WriteError, &r10_bio->state);
499 			else {
500 				/* Fail the request */
501 				r10_bio->devs[slot].bio = NULL;
502 				to_put = bio;
503 				dec_rdev = 1;
504 			}
505 		}
506 	} else {
507 		/*
508 		 * Set R10BIO_Uptodate in our master bio, so that
509 		 * we will return a good error code for to the higher
510 		 * levels even if IO on some other mirrored buffer fails.
511 		 *
512 		 * The 'master' represents the composite IO operation to
513 		 * user-side. So if something waits for IO, then it will
514 		 * wait for the 'master' bio.
515 		 *
516 		 * Do not set R10BIO_Uptodate if the current device is
517 		 * rebuilding or Faulty. This is because we cannot use
518 		 * such device for properly reading the data back (we could
519 		 * potentially use it, if the current write would have felt
520 		 * before rdev->recovery_offset, but for simplicity we don't
521 		 * check this here.
522 		 */
523 		if (test_bit(In_sync, &rdev->flags) &&
524 		    !test_bit(Faulty, &rdev->flags))
525 			set_bit(R10BIO_Uptodate, &r10_bio->state);
526 
527 		/* Maybe we can clear some bad blocks. */
528 		if (rdev_has_badblock(rdev, r10_bio->devs[slot].addr,
529 				      r10_bio->sectors) &&
530 		    !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 
__raid10_find_phys(struct geom * geo,struct r10bio * r10bio)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 
raid10_find_phys(struct r10conf * conf,struct r10bio * r10bio)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 
raid10_find_virt(struct r10conf * conf,sector_t sector,int dev)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  */
read_balance(struct r10conf * conf,struct r10bio * r10_bio,int * max_sectors)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_dist_rdev, *best_pending_rdev, *rdev = NULL;
730 	int do_balance;
731 	int best_dist_slot, best_pending_slot;
732 	bool has_nonrot_disk = false;
733 	unsigned int min_pending;
734 	struct geom *geo = &conf->geo;
735 
736 	raid10_find_phys(conf, r10_bio);
737 	best_dist_slot = -1;
738 	min_pending = UINT_MAX;
739 	best_dist_rdev = NULL;
740 	best_pending_rdev = NULL;
741 	best_dist = MaxSector;
742 	best_good_sectors = 0;
743 	do_balance = 1;
744 	clear_bit(R10BIO_FailFast, &r10_bio->state);
745 
746 	if (raid1_should_read_first(conf->mddev, this_sector, sectors))
747 		do_balance = 0;
748 
749 	for (slot = 0; slot < conf->copies ; slot++) {
750 		sector_t first_bad;
751 		sector_t bad_sectors;
752 		sector_t dev_sector;
753 		unsigned int pending;
754 		bool nonrot;
755 
756 		if (r10_bio->devs[slot].bio == IO_BLOCKED)
757 			continue;
758 		disk = r10_bio->devs[slot].devnum;
759 		rdev = conf->mirrors[disk].replacement;
760 		if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
761 		    r10_bio->devs[slot].addr + sectors >
762 		    rdev->recovery_offset)
763 			rdev = conf->mirrors[disk].rdev;
764 		if (rdev == NULL ||
765 		    test_bit(Faulty, &rdev->flags))
766 			continue;
767 		if (!test_bit(In_sync, &rdev->flags) &&
768 		    r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
769 			continue;
770 
771 		dev_sector = r10_bio->devs[slot].addr;
772 		if (is_badblock(rdev, dev_sector, sectors,
773 				&first_bad, &bad_sectors)) {
774 			if (best_dist < MaxSector)
775 				/* Already have a better slot */
776 				continue;
777 			if (first_bad <= dev_sector) {
778 				/* Cannot read here.  If this is the
779 				 * 'primary' device, then we must not read
780 				 * beyond 'bad_sectors' from another device.
781 				 */
782 				bad_sectors -= (dev_sector - first_bad);
783 				if (!do_balance && sectors > bad_sectors)
784 					sectors = bad_sectors;
785 				if (best_good_sectors > sectors)
786 					best_good_sectors = sectors;
787 			} else {
788 				sector_t good_sectors =
789 					first_bad - dev_sector;
790 				if (good_sectors > best_good_sectors) {
791 					best_good_sectors = good_sectors;
792 					best_dist_slot = slot;
793 					best_dist_rdev = rdev;
794 				}
795 				if (!do_balance)
796 					/* Must read from here */
797 					break;
798 			}
799 			continue;
800 		} else
801 			best_good_sectors = sectors;
802 
803 		if (!do_balance)
804 			break;
805 
806 		nonrot = bdev_nonrot(rdev->bdev);
807 		has_nonrot_disk |= nonrot;
808 		pending = atomic_read(&rdev->nr_pending);
809 		if (min_pending > pending && nonrot) {
810 			min_pending = pending;
811 			best_pending_slot = slot;
812 			best_pending_rdev = rdev;
813 		}
814 
815 		if (best_dist_slot >= 0)
816 			/* At least 2 disks to choose from so failfast is OK */
817 			set_bit(R10BIO_FailFast, &r10_bio->state);
818 		/* This optimisation is debatable, and completely destroys
819 		 * sequential read speed for 'far copies' arrays.  So only
820 		 * keep it for 'near' arrays, and review those later.
821 		 */
822 		if (geo->near_copies > 1 && !pending)
823 			new_distance = 0;
824 
825 		/* for far > 1 always use the lowest address */
826 		else if (geo->far_copies > 1)
827 			new_distance = r10_bio->devs[slot].addr;
828 		else
829 			new_distance = abs(r10_bio->devs[slot].addr -
830 					   conf->mirrors[disk].head_position);
831 
832 		if (new_distance < best_dist) {
833 			best_dist = new_distance;
834 			best_dist_slot = slot;
835 			best_dist_rdev = rdev;
836 		}
837 	}
838 	if (slot >= conf->copies) {
839 		if (has_nonrot_disk) {
840 			slot = best_pending_slot;
841 			rdev = best_pending_rdev;
842 		} else {
843 			slot = best_dist_slot;
844 			rdev = best_dist_rdev;
845 		}
846 	}
847 
848 	if (slot >= 0) {
849 		atomic_inc(&rdev->nr_pending);
850 		r10_bio->read_slot = slot;
851 	} else
852 		rdev = NULL;
853 	*max_sectors = best_good_sectors;
854 
855 	return rdev;
856 }
857 
flush_pending_writes(struct r10conf * conf)858 static void flush_pending_writes(struct r10conf *conf)
859 {
860 	/* Any writes that have been queued but are awaiting
861 	 * bitmap updates get flushed here.
862 	 */
863 	spin_lock_irq(&conf->device_lock);
864 
865 	if (conf->pending_bio_list.head) {
866 		struct blk_plug plug;
867 		struct bio *bio;
868 
869 		bio = bio_list_get(&conf->pending_bio_list);
870 		spin_unlock_irq(&conf->device_lock);
871 
872 		/*
873 		 * As this is called in a wait_event() loop (see freeze_array),
874 		 * current->state might be TASK_UNINTERRUPTIBLE which will
875 		 * cause a warning when we prepare to wait again.  As it is
876 		 * rare that this path is taken, it is perfectly safe to force
877 		 * us to go around the wait_event() loop again, so the warning
878 		 * is a false-positive. Silence the warning by resetting
879 		 * thread state
880 		 */
881 		__set_current_state(TASK_RUNNING);
882 
883 		blk_start_plug(&plug);
884 		raid1_prepare_flush_writes(conf->mddev);
885 		wake_up(&conf->wait_barrier);
886 
887 		while (bio) { /* submit pending writes */
888 			struct bio *next = bio->bi_next;
889 
890 			raid1_submit_write(bio);
891 			bio = next;
892 			cond_resched();
893 		}
894 		blk_finish_plug(&plug);
895 	} else
896 		spin_unlock_irq(&conf->device_lock);
897 }
898 
899 /* Barriers....
900  * Sometimes we need to suspend IO while we do something else,
901  * either some resync/recovery, or reconfigure the array.
902  * To do this we raise a 'barrier'.
903  * The 'barrier' is a counter that can be raised multiple times
904  * to count how many activities are happening which preclude
905  * normal IO.
906  * We can only raise the barrier if there is no pending IO.
907  * i.e. if nr_pending == 0.
908  * We choose only to raise the barrier if no-one is waiting for the
909  * barrier to go down.  This means that as soon as an IO request
910  * is ready, no other operations which require a barrier will start
911  * until the IO request has had a chance.
912  *
913  * So: regular IO calls 'wait_barrier'.  When that returns there
914  *    is no backgroup IO happening,  It must arrange to call
915  *    allow_barrier when it has finished its IO.
916  * backgroup IO calls must call raise_barrier.  Once that returns
917  *    there is no normal IO happeing.  It must arrange to call
918  *    lower_barrier when the particular background IO completes.
919  */
920 
raise_barrier(struct r10conf * conf,int force)921 static void raise_barrier(struct r10conf *conf, int force)
922 {
923 	write_seqlock_irq(&conf->resync_lock);
924 
925 	if (WARN_ON_ONCE(force && !conf->barrier))
926 		force = false;
927 
928 	/* Wait until no block IO is waiting (unless 'force') */
929 	wait_event_barrier(conf, force || !conf->nr_waiting);
930 
931 	/* block any new IO from starting */
932 	WRITE_ONCE(conf->barrier, conf->barrier + 1);
933 
934 	/* Now wait for all pending IO to complete */
935 	wait_event_barrier(conf, !atomic_read(&conf->nr_pending) &&
936 				 conf->barrier < RESYNC_DEPTH);
937 
938 	write_sequnlock_irq(&conf->resync_lock);
939 }
940 
lower_barrier(struct r10conf * conf)941 static void lower_barrier(struct r10conf *conf)
942 {
943 	unsigned long flags;
944 
945 	write_seqlock_irqsave(&conf->resync_lock, flags);
946 	WRITE_ONCE(conf->barrier, conf->barrier - 1);
947 	write_sequnlock_irqrestore(&conf->resync_lock, flags);
948 	wake_up(&conf->wait_barrier);
949 }
950 
stop_waiting_barrier(struct r10conf * conf)951 static bool stop_waiting_barrier(struct r10conf *conf)
952 {
953 	struct bio_list *bio_list = current->bio_list;
954 	struct md_thread *thread;
955 
956 	/* barrier is dropped */
957 	if (!conf->barrier)
958 		return true;
959 
960 	/*
961 	 * If there are already pending requests (preventing the barrier from
962 	 * rising completely), and the pre-process bio queue isn't empty, then
963 	 * don't wait, as we need to empty that queue to get the nr_pending
964 	 * count down.
965 	 */
966 	if (atomic_read(&conf->nr_pending) && bio_list &&
967 	    (!bio_list_empty(&bio_list[0]) || !bio_list_empty(&bio_list[1])))
968 		return true;
969 
970 	/* daemon thread must exist while handling io */
971 	thread = rcu_dereference_protected(conf->mddev->thread, true);
972 	/*
973 	 * move on if io is issued from raid10d(), nr_pending is not released
974 	 * from original io(see handle_read_error()). All raise barrier is
975 	 * blocked until this io is done.
976 	 */
977 	if (thread->tsk == current) {
978 		WARN_ON_ONCE(atomic_read(&conf->nr_pending) == 0);
979 		return true;
980 	}
981 
982 	return false;
983 }
984 
wait_barrier_nolock(struct r10conf * conf)985 static bool wait_barrier_nolock(struct r10conf *conf)
986 {
987 	unsigned int seq = read_seqbegin(&conf->resync_lock);
988 
989 	if (READ_ONCE(conf->barrier))
990 		return false;
991 
992 	atomic_inc(&conf->nr_pending);
993 	if (!read_seqretry(&conf->resync_lock, seq))
994 		return true;
995 
996 	if (atomic_dec_and_test(&conf->nr_pending))
997 		wake_up_barrier(conf);
998 
999 	return false;
1000 }
1001 
wait_barrier(struct r10conf * conf,bool nowait)1002 static bool wait_barrier(struct r10conf *conf, bool nowait)
1003 {
1004 	bool ret = true;
1005 
1006 	if (wait_barrier_nolock(conf))
1007 		return true;
1008 
1009 	write_seqlock_irq(&conf->resync_lock);
1010 	if (conf->barrier) {
1011 		/* Return false when nowait flag is set */
1012 		if (nowait) {
1013 			ret = false;
1014 		} else {
1015 			conf->nr_waiting++;
1016 			mddev_add_trace_msg(conf->mddev, "raid10 wait barrier");
1017 			wait_event_barrier(conf, stop_waiting_barrier(conf));
1018 			conf->nr_waiting--;
1019 		}
1020 		if (!conf->nr_waiting)
1021 			wake_up(&conf->wait_barrier);
1022 	}
1023 	/* Only increment nr_pending when we wait */
1024 	if (ret)
1025 		atomic_inc(&conf->nr_pending);
1026 	write_sequnlock_irq(&conf->resync_lock);
1027 	return ret;
1028 }
1029 
allow_barrier(struct r10conf * conf)1030 static void allow_barrier(struct r10conf *conf)
1031 {
1032 	if ((atomic_dec_and_test(&conf->nr_pending)) ||
1033 			(conf->array_freeze_pending))
1034 		wake_up_barrier(conf);
1035 }
1036 
freeze_array(struct r10conf * conf,int extra)1037 static void freeze_array(struct r10conf *conf, int extra)
1038 {
1039 	/* stop syncio and normal IO and wait for everything to
1040 	 * go quiet.
1041 	 * We increment barrier and nr_waiting, and then
1042 	 * wait until nr_pending match nr_queued+extra
1043 	 * This is called in the context of one normal IO request
1044 	 * that has failed. Thus any sync request that might be pending
1045 	 * will be blocked by nr_pending, and we need to wait for
1046 	 * pending IO requests to complete or be queued for re-try.
1047 	 * Thus the number queued (nr_queued) plus this request (extra)
1048 	 * must match the number of pending IOs (nr_pending) before
1049 	 * we continue.
1050 	 */
1051 	write_seqlock_irq(&conf->resync_lock);
1052 	conf->array_freeze_pending++;
1053 	WRITE_ONCE(conf->barrier, conf->barrier + 1);
1054 	conf->nr_waiting++;
1055 	wait_event_barrier_cmd(conf, atomic_read(&conf->nr_pending) ==
1056 			conf->nr_queued + extra, flush_pending_writes(conf));
1057 	conf->array_freeze_pending--;
1058 	write_sequnlock_irq(&conf->resync_lock);
1059 }
1060 
unfreeze_array(struct r10conf * conf)1061 static void unfreeze_array(struct r10conf *conf)
1062 {
1063 	/* reverse the effect of the freeze */
1064 	write_seqlock_irq(&conf->resync_lock);
1065 	WRITE_ONCE(conf->barrier, conf->barrier - 1);
1066 	conf->nr_waiting--;
1067 	wake_up(&conf->wait_barrier);
1068 	write_sequnlock_irq(&conf->resync_lock);
1069 }
1070 
choose_data_offset(struct r10bio * r10_bio,struct md_rdev * rdev)1071 static sector_t choose_data_offset(struct r10bio *r10_bio,
1072 				   struct md_rdev *rdev)
1073 {
1074 	if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1075 	    test_bit(R10BIO_Previous, &r10_bio->state))
1076 		return rdev->data_offset;
1077 	else
1078 		return rdev->new_data_offset;
1079 }
1080 
raid10_unplug(struct blk_plug_cb * cb,bool from_schedule)1081 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1082 {
1083 	struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb, cb);
1084 	struct mddev *mddev = plug->cb.data;
1085 	struct r10conf *conf = mddev->private;
1086 	struct bio *bio;
1087 
1088 	if (from_schedule) {
1089 		spin_lock_irq(&conf->device_lock);
1090 		bio_list_merge(&conf->pending_bio_list, &plug->pending);
1091 		spin_unlock_irq(&conf->device_lock);
1092 		wake_up_barrier(conf);
1093 		md_wakeup_thread(mddev->thread);
1094 		kfree(plug);
1095 		return;
1096 	}
1097 
1098 	/* we aren't scheduling, so we can do the write-out directly. */
1099 	bio = bio_list_get(&plug->pending);
1100 	raid1_prepare_flush_writes(mddev);
1101 	wake_up_barrier(conf);
1102 
1103 	while (bio) { /* submit pending writes */
1104 		struct bio *next = bio->bi_next;
1105 
1106 		raid1_submit_write(bio);
1107 		bio = next;
1108 		cond_resched();
1109 	}
1110 	kfree(plug);
1111 }
1112 
1113 /*
1114  * 1. Register the new request and wait if the reconstruction thread has put
1115  * up a bar for new requests. Continue immediately if no resync is active
1116  * currently.
1117  * 2. If IO spans the reshape position.  Need to wait for reshape to pass.
1118  */
regular_request_wait(struct mddev * mddev,struct r10conf * conf,struct bio * bio,sector_t sectors)1119 static bool regular_request_wait(struct mddev *mddev, struct r10conf *conf,
1120 				 struct bio *bio, sector_t sectors)
1121 {
1122 	/* Bail out if REQ_NOWAIT is set for the bio */
1123 	if (!wait_barrier(conf, bio->bi_opf & REQ_NOWAIT)) {
1124 		bio_wouldblock_error(bio);
1125 		return false;
1126 	}
1127 	while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1128 	    bio->bi_iter.bi_sector < conf->reshape_progress &&
1129 	    bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1130 		allow_barrier(conf);
1131 		if (bio->bi_opf & REQ_NOWAIT) {
1132 			bio_wouldblock_error(bio);
1133 			return false;
1134 		}
1135 		mddev_add_trace_msg(conf->mddev, "raid10 wait reshape");
1136 		wait_event(conf->wait_barrier,
1137 			   conf->reshape_progress <= bio->bi_iter.bi_sector ||
1138 			   conf->reshape_progress >= bio->bi_iter.bi_sector +
1139 			   sectors);
1140 		wait_barrier(conf, false);
1141 	}
1142 	return true;
1143 }
1144 
raid10_read_request(struct mddev * mddev,struct bio * bio,struct r10bio * r10_bio,bool io_accounting)1145 static void raid10_read_request(struct mddev *mddev, struct bio *bio,
1146 				struct r10bio *r10_bio, bool io_accounting)
1147 {
1148 	struct r10conf *conf = mddev->private;
1149 	struct bio *read_bio;
1150 	int max_sectors;
1151 	struct md_rdev *rdev;
1152 	char b[BDEVNAME_SIZE];
1153 	int slot = r10_bio->read_slot;
1154 	struct md_rdev *err_rdev = NULL;
1155 	gfp_t gfp = GFP_NOIO;
1156 	int error;
1157 
1158 	if (slot >= 0 && r10_bio->devs[slot].rdev) {
1159 		/*
1160 		 * This is an error retry, but we cannot
1161 		 * safely dereference the rdev in the r10_bio,
1162 		 * we must use the one in conf.
1163 		 * If it has already been disconnected (unlikely)
1164 		 * we lose the device name in error messages.
1165 		 */
1166 		int disk;
1167 		/*
1168 		 * As we are blocking raid10, it is a little safer to
1169 		 * use __GFP_HIGH.
1170 		 */
1171 		gfp = GFP_NOIO | __GFP_HIGH;
1172 
1173 		disk = r10_bio->devs[slot].devnum;
1174 		err_rdev = conf->mirrors[disk].rdev;
1175 		if (err_rdev)
1176 			snprintf(b, sizeof(b), "%pg", err_rdev->bdev);
1177 		else {
1178 			strcpy(b, "???");
1179 			/* This never gets dereferenced */
1180 			err_rdev = r10_bio->devs[slot].rdev;
1181 		}
1182 	}
1183 
1184 	if (!regular_request_wait(mddev, conf, bio, r10_bio->sectors))
1185 		return;
1186 	rdev = read_balance(conf, r10_bio, &max_sectors);
1187 	if (!rdev) {
1188 		if (err_rdev) {
1189 			pr_crit_ratelimited("md/raid10:%s: %s: unrecoverable I/O read error for block %llu\n",
1190 					    mdname(mddev), b,
1191 					    (unsigned long long)r10_bio->sector);
1192 		}
1193 		raid_end_bio_io(r10_bio);
1194 		return;
1195 	}
1196 	if (err_rdev)
1197 		pr_err_ratelimited("md/raid10:%s: %pg: redirecting sector %llu to another mirror\n",
1198 				   mdname(mddev),
1199 				   rdev->bdev,
1200 				   (unsigned long long)r10_bio->sector);
1201 	if (max_sectors < bio_sectors(bio)) {
1202 		struct bio *split = bio_split(bio, max_sectors,
1203 					      gfp, &conf->bio_split);
1204 		if (IS_ERR(split)) {
1205 			error = PTR_ERR(split);
1206 			goto err_handle;
1207 		}
1208 		bio_chain(split, bio);
1209 		allow_barrier(conf);
1210 		submit_bio_noacct(bio);
1211 		wait_barrier(conf, false);
1212 		bio = split;
1213 		r10_bio->master_bio = bio;
1214 		r10_bio->sectors = max_sectors;
1215 	}
1216 	slot = r10_bio->read_slot;
1217 
1218 	if (io_accounting) {
1219 		md_account_bio(mddev, &bio);
1220 		r10_bio->master_bio = bio;
1221 	}
1222 	read_bio = bio_alloc_clone(rdev->bdev, bio, gfp, &mddev->bio_set);
1223 
1224 	r10_bio->devs[slot].bio = read_bio;
1225 	r10_bio->devs[slot].rdev = rdev;
1226 
1227 	read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1228 		choose_data_offset(r10_bio, rdev);
1229 	read_bio->bi_end_io = raid10_end_read_request;
1230 	if (test_bit(FailFast, &rdev->flags) &&
1231 	    test_bit(R10BIO_FailFast, &r10_bio->state))
1232 	        read_bio->bi_opf |= MD_FAILFAST;
1233 	read_bio->bi_private = r10_bio;
1234 	mddev_trace_remap(mddev, read_bio, r10_bio->sector);
1235 	submit_bio_noacct(read_bio);
1236 	return;
1237 err_handle:
1238 	atomic_dec(&rdev->nr_pending);
1239 	bio->bi_status = errno_to_blk_status(error);
1240 	set_bit(R10BIO_Uptodate, &r10_bio->state);
1241 	raid_end_bio_io(r10_bio);
1242 }
1243 
raid10_write_one_disk(struct mddev * mddev,struct r10bio * r10_bio,struct bio * bio,bool replacement,int n_copy)1244 static void raid10_write_one_disk(struct mddev *mddev, struct r10bio *r10_bio,
1245 				  struct bio *bio, bool replacement,
1246 				  int n_copy)
1247 {
1248 	unsigned long flags;
1249 	struct r10conf *conf = mddev->private;
1250 	struct md_rdev *rdev;
1251 	int devnum = r10_bio->devs[n_copy].devnum;
1252 	struct bio *mbio;
1253 
1254 	rdev = replacement ? conf->mirrors[devnum].replacement :
1255 			     conf->mirrors[devnum].rdev;
1256 
1257 	mbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO, &mddev->bio_set);
1258 	if (replacement)
1259 		r10_bio->devs[n_copy].repl_bio = mbio;
1260 	else
1261 		r10_bio->devs[n_copy].bio = mbio;
1262 
1263 	mbio->bi_iter.bi_sector	= (r10_bio->devs[n_copy].addr +
1264 				   choose_data_offset(r10_bio, rdev));
1265 	mbio->bi_end_io	= raid10_end_write_request;
1266 	if (!replacement && test_bit(FailFast,
1267 				     &conf->mirrors[devnum].rdev->flags)
1268 			 && enough(conf, devnum))
1269 		mbio->bi_opf |= MD_FAILFAST;
1270 	mbio->bi_private = r10_bio;
1271 	mddev_trace_remap(mddev, mbio, r10_bio->sector);
1272 	/* flush_pending_writes() needs access to the rdev so...*/
1273 	mbio->bi_bdev = (void *)rdev;
1274 
1275 	atomic_inc(&r10_bio->remaining);
1276 
1277 	if (!raid1_add_bio_to_plug(mddev, mbio, raid10_unplug, conf->copies)) {
1278 		spin_lock_irqsave(&conf->device_lock, flags);
1279 		bio_list_add(&conf->pending_bio_list, mbio);
1280 		spin_unlock_irqrestore(&conf->device_lock, flags);
1281 		md_wakeup_thread(mddev->thread);
1282 	}
1283 }
1284 
wait_blocked_dev(struct mddev * mddev,struct r10bio * r10_bio)1285 static void wait_blocked_dev(struct mddev *mddev, struct r10bio *r10_bio)
1286 {
1287 	struct r10conf *conf = mddev->private;
1288 	struct md_rdev *blocked_rdev;
1289 	int i;
1290 
1291 retry_wait:
1292 	blocked_rdev = NULL;
1293 	for (i = 0; i < conf->copies; i++) {
1294 		struct md_rdev *rdev, *rrdev;
1295 
1296 		rdev = conf->mirrors[i].rdev;
1297 		if (rdev) {
1298 			sector_t dev_sector = r10_bio->devs[i].addr;
1299 
1300 			/*
1301 			 * Discard request doesn't care the write result
1302 			 * so it doesn't need to wait blocked disk here.
1303 			 */
1304 			if (test_bit(WriteErrorSeen, &rdev->flags) &&
1305 			    r10_bio->sectors &&
1306 			    rdev_has_badblock(rdev, dev_sector,
1307 					      r10_bio->sectors) < 0)
1308 				/*
1309 				 * Mustn't write here until the bad
1310 				 * block is acknowledged
1311 				 */
1312 				set_bit(BlockedBadBlocks, &rdev->flags);
1313 
1314 			if (rdev_blocked(rdev)) {
1315 				blocked_rdev = rdev;
1316 				atomic_inc(&rdev->nr_pending);
1317 				break;
1318 			}
1319 		}
1320 
1321 		rrdev = conf->mirrors[i].replacement;
1322 		if (rrdev && rdev_blocked(rrdev)) {
1323 			atomic_inc(&rrdev->nr_pending);
1324 			blocked_rdev = rrdev;
1325 			break;
1326 		}
1327 	}
1328 
1329 	if (unlikely(blocked_rdev)) {
1330 		/* Have to wait for this device to get unblocked, then retry */
1331 		allow_barrier(conf);
1332 		mddev_add_trace_msg(conf->mddev,
1333 			"raid10 %s wait rdev %d blocked",
1334 			__func__, blocked_rdev->raid_disk);
1335 		md_wait_for_blocked_rdev(blocked_rdev, mddev);
1336 		wait_barrier(conf, false);
1337 		goto retry_wait;
1338 	}
1339 }
1340 
raid10_write_request(struct mddev * mddev,struct bio * bio,struct r10bio * r10_bio)1341 static void raid10_write_request(struct mddev *mddev, struct bio *bio,
1342 				 struct r10bio *r10_bio)
1343 {
1344 	struct r10conf *conf = mddev->private;
1345 	int i, k;
1346 	sector_t sectors;
1347 	int max_sectors;
1348 	int error;
1349 
1350 	if ((mddev_is_clustered(mddev) &&
1351 	     mddev->cluster_ops->area_resyncing(mddev, WRITE,
1352 						bio->bi_iter.bi_sector,
1353 						bio_end_sector(bio)))) {
1354 		DEFINE_WAIT(w);
1355 		/* Bail out if REQ_NOWAIT is set for the bio */
1356 		if (bio->bi_opf & REQ_NOWAIT) {
1357 			bio_wouldblock_error(bio);
1358 			return;
1359 		}
1360 		for (;;) {
1361 			prepare_to_wait(&conf->wait_barrier,
1362 					&w, TASK_IDLE);
1363 			if (!mddev->cluster_ops->area_resyncing(mddev, WRITE,
1364 				 bio->bi_iter.bi_sector, bio_end_sector(bio)))
1365 				break;
1366 			schedule();
1367 		}
1368 		finish_wait(&conf->wait_barrier, &w);
1369 	}
1370 
1371 	sectors = r10_bio->sectors;
1372 	if (!regular_request_wait(mddev, conf, bio, sectors))
1373 		return;
1374 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1375 	    (mddev->reshape_backwards
1376 	     ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1377 		bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1378 	     : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1379 		bio->bi_iter.bi_sector < conf->reshape_progress))) {
1380 		/* Need to update reshape_position in metadata */
1381 		mddev->reshape_position = conf->reshape_progress;
1382 		set_mask_bits(&mddev->sb_flags, 0,
1383 			      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1384 		md_wakeup_thread(mddev->thread);
1385 		if (bio->bi_opf & REQ_NOWAIT) {
1386 			allow_barrier(conf);
1387 			bio_wouldblock_error(bio);
1388 			return;
1389 		}
1390 		mddev_add_trace_msg(conf->mddev,
1391 			"raid10 wait reshape metadata");
1392 		wait_event(mddev->sb_wait,
1393 			   !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
1394 
1395 		conf->reshape_safe = mddev->reshape_position;
1396 	}
1397 
1398 	/* first select target devices under rcu_lock and
1399 	 * inc refcount on their rdev.  Record them by setting
1400 	 * bios[x] to bio
1401 	 * If there are known/acknowledged bad blocks on any device
1402 	 * on which we have seen a write error, we want to avoid
1403 	 * writing to those blocks.  This potentially requires several
1404 	 * writes to write around the bad blocks.  Each set of writes
1405 	 * gets its own r10_bio with a set of bios attached.
1406 	 */
1407 
1408 	r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1409 	raid10_find_phys(conf, r10_bio);
1410 
1411 	wait_blocked_dev(mddev, r10_bio);
1412 
1413 	max_sectors = r10_bio->sectors;
1414 
1415 	for (i = 0;  i < conf->copies; i++) {
1416 		int d = r10_bio->devs[i].devnum;
1417 		struct md_rdev *rdev, *rrdev;
1418 
1419 		rdev = conf->mirrors[d].rdev;
1420 		rrdev = conf->mirrors[d].replacement;
1421 		if (rdev && (test_bit(Faulty, &rdev->flags)))
1422 			rdev = NULL;
1423 		if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1424 			rrdev = NULL;
1425 
1426 		r10_bio->devs[i].bio = NULL;
1427 		r10_bio->devs[i].repl_bio = NULL;
1428 
1429 		if (!rdev && !rrdev)
1430 			continue;
1431 		if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1432 			sector_t first_bad;
1433 			sector_t dev_sector = r10_bio->devs[i].addr;
1434 			sector_t bad_sectors;
1435 			int is_bad;
1436 
1437 			is_bad = is_badblock(rdev, dev_sector, max_sectors,
1438 					     &first_bad, &bad_sectors);
1439 			if (is_bad && first_bad <= dev_sector) {
1440 				/* Cannot write here at all */
1441 				bad_sectors -= (dev_sector - first_bad);
1442 				if (bad_sectors < max_sectors)
1443 					/* Mustn't write more than bad_sectors
1444 					 * to other devices yet
1445 					 */
1446 					max_sectors = bad_sectors;
1447 				continue;
1448 			}
1449 			if (is_bad) {
1450 				int good_sectors;
1451 
1452 				/*
1453 				 * We cannot atomically write this, so just
1454 				 * error in that case. It could be possible to
1455 				 * atomically write other mirrors, but the
1456 				 * complexity of supporting that is not worth
1457 				 * the benefit.
1458 				 */
1459 				if (bio->bi_opf & REQ_ATOMIC) {
1460 					error = -EIO;
1461 					goto err_handle;
1462 				}
1463 
1464 				good_sectors = first_bad - dev_sector;
1465 				if (good_sectors < max_sectors)
1466 					max_sectors = good_sectors;
1467 			}
1468 		}
1469 		if (rdev) {
1470 			r10_bio->devs[i].bio = bio;
1471 			atomic_inc(&rdev->nr_pending);
1472 		}
1473 		if (rrdev) {
1474 			r10_bio->devs[i].repl_bio = bio;
1475 			atomic_inc(&rrdev->nr_pending);
1476 		}
1477 	}
1478 
1479 	if (max_sectors < r10_bio->sectors)
1480 		r10_bio->sectors = max_sectors;
1481 
1482 	if (r10_bio->sectors < bio_sectors(bio)) {
1483 		struct bio *split = bio_split(bio, r10_bio->sectors,
1484 					      GFP_NOIO, &conf->bio_split);
1485 		if (IS_ERR(split)) {
1486 			error = PTR_ERR(split);
1487 			goto err_handle;
1488 		}
1489 		bio_chain(split, bio);
1490 		allow_barrier(conf);
1491 		submit_bio_noacct(bio);
1492 		wait_barrier(conf, false);
1493 		bio = split;
1494 		r10_bio->master_bio = bio;
1495 	}
1496 
1497 	md_account_bio(mddev, &bio);
1498 	r10_bio->master_bio = bio;
1499 	atomic_set(&r10_bio->remaining, 1);
1500 
1501 	for (i = 0; i < conf->copies; i++) {
1502 		if (r10_bio->devs[i].bio)
1503 			raid10_write_one_disk(mddev, r10_bio, bio, false, i);
1504 		if (r10_bio->devs[i].repl_bio)
1505 			raid10_write_one_disk(mddev, r10_bio, bio, true, i);
1506 	}
1507 	one_write_done(r10_bio);
1508 	return;
1509 err_handle:
1510 	for (k = 0;  k < i; k++) {
1511 		int d = r10_bio->devs[k].devnum;
1512 		struct md_rdev *rdev = conf->mirrors[d].rdev;
1513 		struct md_rdev *rrdev = conf->mirrors[d].replacement;
1514 
1515 		if (r10_bio->devs[k].bio) {
1516 			rdev_dec_pending(rdev, mddev);
1517 			r10_bio->devs[k].bio = NULL;
1518 		}
1519 		if (r10_bio->devs[k].repl_bio) {
1520 			rdev_dec_pending(rrdev, mddev);
1521 			r10_bio->devs[k].repl_bio = NULL;
1522 		}
1523 	}
1524 
1525 	bio->bi_status = errno_to_blk_status(error);
1526 	set_bit(R10BIO_Uptodate, &r10_bio->state);
1527 	raid_end_bio_io(r10_bio);
1528 }
1529 
__make_request(struct mddev * mddev,struct bio * bio,int sectors)1530 static void __make_request(struct mddev *mddev, struct bio *bio, int sectors)
1531 {
1532 	struct r10conf *conf = mddev->private;
1533 	struct r10bio *r10_bio;
1534 
1535 	r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1536 
1537 	r10_bio->master_bio = bio;
1538 	r10_bio->sectors = sectors;
1539 
1540 	r10_bio->mddev = mddev;
1541 	r10_bio->sector = bio->bi_iter.bi_sector;
1542 	r10_bio->state = 0;
1543 	r10_bio->read_slot = -1;
1544 	memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) *
1545 			conf->geo.raid_disks);
1546 
1547 	if (bio_data_dir(bio) == READ)
1548 		raid10_read_request(mddev, bio, r10_bio, true);
1549 	else
1550 		raid10_write_request(mddev, bio, r10_bio);
1551 }
1552 
raid_end_discard_bio(struct r10bio * r10bio)1553 static void raid_end_discard_bio(struct r10bio *r10bio)
1554 {
1555 	struct r10conf *conf = r10bio->mddev->private;
1556 	struct r10bio *first_r10bio;
1557 
1558 	while (atomic_dec_and_test(&r10bio->remaining)) {
1559 
1560 		allow_barrier(conf);
1561 
1562 		if (!test_bit(R10BIO_Discard, &r10bio->state)) {
1563 			first_r10bio = (struct r10bio *)r10bio->master_bio;
1564 			free_r10bio(r10bio);
1565 			r10bio = first_r10bio;
1566 		} else {
1567 			md_write_end(r10bio->mddev);
1568 			bio_endio(r10bio->master_bio);
1569 			free_r10bio(r10bio);
1570 			break;
1571 		}
1572 	}
1573 }
1574 
raid10_end_discard_request(struct bio * bio)1575 static void raid10_end_discard_request(struct bio *bio)
1576 {
1577 	struct r10bio *r10_bio = bio->bi_private;
1578 	struct r10conf *conf = r10_bio->mddev->private;
1579 	struct md_rdev *rdev = NULL;
1580 	int dev;
1581 	int slot, repl;
1582 
1583 	/*
1584 	 * We don't care the return value of discard bio
1585 	 */
1586 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
1587 		set_bit(R10BIO_Uptodate, &r10_bio->state);
1588 
1589 	dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1590 	rdev = repl ? conf->mirrors[dev].replacement :
1591 		      conf->mirrors[dev].rdev;
1592 
1593 	raid_end_discard_bio(r10_bio);
1594 	rdev_dec_pending(rdev, conf->mddev);
1595 }
1596 
1597 /*
1598  * There are some limitations to handle discard bio
1599  * 1st, the discard size is bigger than stripe_size*2.
1600  * 2st, if the discard bio spans reshape progress, we use the old way to
1601  * handle discard bio
1602  */
raid10_handle_discard(struct mddev * mddev,struct bio * bio)1603 static int raid10_handle_discard(struct mddev *mddev, struct bio *bio)
1604 {
1605 	struct r10conf *conf = mddev->private;
1606 	struct geom *geo = &conf->geo;
1607 	int far_copies = geo->far_copies;
1608 	bool first_copy = true;
1609 	struct r10bio *r10_bio, *first_r10bio;
1610 	struct bio *split;
1611 	int disk;
1612 	sector_t chunk;
1613 	unsigned int stripe_size;
1614 	unsigned int stripe_data_disks;
1615 	sector_t split_size;
1616 	sector_t bio_start, bio_end;
1617 	sector_t first_stripe_index, last_stripe_index;
1618 	sector_t start_disk_offset;
1619 	unsigned int start_disk_index;
1620 	sector_t end_disk_offset;
1621 	unsigned int end_disk_index;
1622 	unsigned int remainder;
1623 
1624 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1625 		return -EAGAIN;
1626 
1627 	if (!wait_barrier(conf, bio->bi_opf & REQ_NOWAIT)) {
1628 		bio_wouldblock_error(bio);
1629 		return 0;
1630 	}
1631 
1632 	/*
1633 	 * Check reshape again to avoid reshape happens after checking
1634 	 * MD_RECOVERY_RESHAPE and before wait_barrier
1635 	 */
1636 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
1637 		goto out;
1638 
1639 	if (geo->near_copies)
1640 		stripe_data_disks = geo->raid_disks / geo->near_copies +
1641 					geo->raid_disks % geo->near_copies;
1642 	else
1643 		stripe_data_disks = geo->raid_disks;
1644 
1645 	stripe_size = stripe_data_disks << geo->chunk_shift;
1646 
1647 	bio_start = bio->bi_iter.bi_sector;
1648 	bio_end = bio_end_sector(bio);
1649 
1650 	/*
1651 	 * Maybe one discard bio is smaller than strip size or across one
1652 	 * stripe and discard region is larger than one stripe size. For far
1653 	 * offset layout, if the discard region is not aligned with stripe
1654 	 * size, there is hole when we submit discard bio to member disk.
1655 	 * For simplicity, we only handle discard bio which discard region
1656 	 * is bigger than stripe_size * 2
1657 	 */
1658 	if (bio_sectors(bio) < stripe_size*2)
1659 		goto out;
1660 
1661 	/*
1662 	 * Keep bio aligned with strip size.
1663 	 */
1664 	div_u64_rem(bio_start, stripe_size, &remainder);
1665 	if (remainder) {
1666 		split_size = stripe_size - remainder;
1667 		split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1668 		if (IS_ERR(split)) {
1669 			bio->bi_status = errno_to_blk_status(PTR_ERR(split));
1670 			bio_endio(bio);
1671 			return 0;
1672 		}
1673 		bio_chain(split, bio);
1674 		allow_barrier(conf);
1675 		/* Resend the fist split part */
1676 		submit_bio_noacct(split);
1677 		wait_barrier(conf, false);
1678 	}
1679 	div_u64_rem(bio_end, stripe_size, &remainder);
1680 	if (remainder) {
1681 		split_size = bio_sectors(bio) - remainder;
1682 		split = bio_split(bio, split_size, GFP_NOIO, &conf->bio_split);
1683 		if (IS_ERR(split)) {
1684 			bio->bi_status = errno_to_blk_status(PTR_ERR(split));
1685 			bio_endio(bio);
1686 			return 0;
1687 		}
1688 		bio_chain(split, bio);
1689 		allow_barrier(conf);
1690 		/* Resend the second split part */
1691 		submit_bio_noacct(bio);
1692 		bio = split;
1693 		wait_barrier(conf, false);
1694 	}
1695 
1696 	bio_start = bio->bi_iter.bi_sector;
1697 	bio_end = bio_end_sector(bio);
1698 
1699 	/*
1700 	 * Raid10 uses chunk as the unit to store data. It's similar like raid0.
1701 	 * One stripe contains the chunks from all member disk (one chunk from
1702 	 * one disk at the same HBA address). For layout detail, see 'man md 4'
1703 	 */
1704 	chunk = bio_start >> geo->chunk_shift;
1705 	chunk *= geo->near_copies;
1706 	first_stripe_index = chunk;
1707 	start_disk_index = sector_div(first_stripe_index, geo->raid_disks);
1708 	if (geo->far_offset)
1709 		first_stripe_index *= geo->far_copies;
1710 	start_disk_offset = (bio_start & geo->chunk_mask) +
1711 				(first_stripe_index << geo->chunk_shift);
1712 
1713 	chunk = bio_end >> geo->chunk_shift;
1714 	chunk *= geo->near_copies;
1715 	last_stripe_index = chunk;
1716 	end_disk_index = sector_div(last_stripe_index, geo->raid_disks);
1717 	if (geo->far_offset)
1718 		last_stripe_index *= geo->far_copies;
1719 	end_disk_offset = (bio_end & geo->chunk_mask) +
1720 				(last_stripe_index << geo->chunk_shift);
1721 
1722 retry_discard:
1723 	r10_bio = mempool_alloc(&conf->r10bio_pool, GFP_NOIO);
1724 	r10_bio->mddev = mddev;
1725 	r10_bio->state = 0;
1726 	r10_bio->sectors = 0;
1727 	memset(r10_bio->devs, 0, sizeof(r10_bio->devs[0]) * geo->raid_disks);
1728 	wait_blocked_dev(mddev, r10_bio);
1729 
1730 	/*
1731 	 * For far layout it needs more than one r10bio to cover all regions.
1732 	 * Inspired by raid10_sync_request, we can use the first r10bio->master_bio
1733 	 * to record the discard bio. Other r10bio->master_bio record the first
1734 	 * r10bio. The first r10bio only release after all other r10bios finish.
1735 	 * The discard bio returns only first r10bio finishes
1736 	 */
1737 	if (first_copy) {
1738 		r10_bio->master_bio = bio;
1739 		set_bit(R10BIO_Discard, &r10_bio->state);
1740 		first_copy = false;
1741 		first_r10bio = r10_bio;
1742 	} else
1743 		r10_bio->master_bio = (struct bio *)first_r10bio;
1744 
1745 	/*
1746 	 * first select target devices under rcu_lock and
1747 	 * inc refcount on their rdev.  Record them by setting
1748 	 * bios[x] to bio
1749 	 */
1750 	for (disk = 0; disk < geo->raid_disks; disk++) {
1751 		struct md_rdev *rdev, *rrdev;
1752 
1753 		rdev = conf->mirrors[disk].rdev;
1754 		rrdev = conf->mirrors[disk].replacement;
1755 		r10_bio->devs[disk].bio = NULL;
1756 		r10_bio->devs[disk].repl_bio = NULL;
1757 
1758 		if (rdev && (test_bit(Faulty, &rdev->flags)))
1759 			rdev = NULL;
1760 		if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1761 			rrdev = NULL;
1762 		if (!rdev && !rrdev)
1763 			continue;
1764 
1765 		if (rdev) {
1766 			r10_bio->devs[disk].bio = bio;
1767 			atomic_inc(&rdev->nr_pending);
1768 		}
1769 		if (rrdev) {
1770 			r10_bio->devs[disk].repl_bio = bio;
1771 			atomic_inc(&rrdev->nr_pending);
1772 		}
1773 	}
1774 
1775 	atomic_set(&r10_bio->remaining, 1);
1776 	for (disk = 0; disk < geo->raid_disks; disk++) {
1777 		sector_t dev_start, dev_end;
1778 		struct bio *mbio, *rbio = NULL;
1779 
1780 		/*
1781 		 * Now start to calculate the start and end address for each disk.
1782 		 * The space between dev_start and dev_end is the discard region.
1783 		 *
1784 		 * For dev_start, it needs to consider three conditions:
1785 		 * 1st, the disk is before start_disk, you can imagine the disk in
1786 		 * the next stripe. So the dev_start is the start address of next
1787 		 * stripe.
1788 		 * 2st, the disk is after start_disk, it means the disk is at the
1789 		 * same stripe of first disk
1790 		 * 3st, the first disk itself, we can use start_disk_offset directly
1791 		 */
1792 		if (disk < start_disk_index)
1793 			dev_start = (first_stripe_index + 1) * mddev->chunk_sectors;
1794 		else if (disk > start_disk_index)
1795 			dev_start = first_stripe_index * mddev->chunk_sectors;
1796 		else
1797 			dev_start = start_disk_offset;
1798 
1799 		if (disk < end_disk_index)
1800 			dev_end = (last_stripe_index + 1) * mddev->chunk_sectors;
1801 		else if (disk > end_disk_index)
1802 			dev_end = last_stripe_index * mddev->chunk_sectors;
1803 		else
1804 			dev_end = end_disk_offset;
1805 
1806 		/*
1807 		 * It only handles discard bio which size is >= stripe size, so
1808 		 * dev_end > dev_start all the time.
1809 		 * It doesn't need to use rcu lock to get rdev here. We already
1810 		 * add rdev->nr_pending in the first loop.
1811 		 */
1812 		if (r10_bio->devs[disk].bio) {
1813 			struct md_rdev *rdev = conf->mirrors[disk].rdev;
1814 			mbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1815 					       &mddev->bio_set);
1816 			mbio->bi_end_io = raid10_end_discard_request;
1817 			mbio->bi_private = r10_bio;
1818 			r10_bio->devs[disk].bio = mbio;
1819 			r10_bio->devs[disk].devnum = disk;
1820 			atomic_inc(&r10_bio->remaining);
1821 			md_submit_discard_bio(mddev, rdev, mbio,
1822 					dev_start + choose_data_offset(r10_bio, rdev),
1823 					dev_end - dev_start);
1824 			bio_endio(mbio);
1825 		}
1826 		if (r10_bio->devs[disk].repl_bio) {
1827 			struct md_rdev *rrdev = conf->mirrors[disk].replacement;
1828 			rbio = bio_alloc_clone(bio->bi_bdev, bio, GFP_NOIO,
1829 					       &mddev->bio_set);
1830 			rbio->bi_end_io = raid10_end_discard_request;
1831 			rbio->bi_private = r10_bio;
1832 			r10_bio->devs[disk].repl_bio = rbio;
1833 			r10_bio->devs[disk].devnum = disk;
1834 			atomic_inc(&r10_bio->remaining);
1835 			md_submit_discard_bio(mddev, rrdev, rbio,
1836 					dev_start + choose_data_offset(r10_bio, rrdev),
1837 					dev_end - dev_start);
1838 			bio_endio(rbio);
1839 		}
1840 	}
1841 
1842 	if (!geo->far_offset && --far_copies) {
1843 		first_stripe_index += geo->stride >> geo->chunk_shift;
1844 		start_disk_offset += geo->stride;
1845 		last_stripe_index += geo->stride >> geo->chunk_shift;
1846 		end_disk_offset += geo->stride;
1847 		atomic_inc(&first_r10bio->remaining);
1848 		raid_end_discard_bio(r10_bio);
1849 		wait_barrier(conf, false);
1850 		goto retry_discard;
1851 	}
1852 
1853 	raid_end_discard_bio(r10_bio);
1854 
1855 	return 0;
1856 out:
1857 	allow_barrier(conf);
1858 	return -EAGAIN;
1859 }
1860 
raid10_make_request(struct mddev * mddev,struct bio * bio)1861 static bool raid10_make_request(struct mddev *mddev, struct bio *bio)
1862 {
1863 	struct r10conf *conf = mddev->private;
1864 	sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1865 	int chunk_sects = chunk_mask + 1;
1866 	int sectors = bio_sectors(bio);
1867 
1868 	if (unlikely(bio->bi_opf & REQ_PREFLUSH)
1869 	    && md_flush_request(mddev, bio))
1870 		return true;
1871 
1872 	md_write_start(mddev, bio);
1873 
1874 	if (unlikely(bio_op(bio) == REQ_OP_DISCARD))
1875 		if (!raid10_handle_discard(mddev, bio))
1876 			return true;
1877 
1878 	/*
1879 	 * If this request crosses a chunk boundary, we need to split
1880 	 * it.
1881 	 */
1882 	if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1883 		     sectors > chunk_sects
1884 		     && (conf->geo.near_copies < conf->geo.raid_disks
1885 			 || conf->prev.near_copies <
1886 			 conf->prev.raid_disks)))
1887 		sectors = chunk_sects -
1888 			(bio->bi_iter.bi_sector &
1889 			 (chunk_sects - 1));
1890 	__make_request(mddev, bio, sectors);
1891 
1892 	/* In case raid10d snuck in to freeze_array */
1893 	wake_up_barrier(conf);
1894 	return true;
1895 }
1896 
raid10_status(struct seq_file * seq,struct mddev * mddev)1897 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1898 {
1899 	struct r10conf *conf = mddev->private;
1900 	int i;
1901 
1902 	lockdep_assert_held(&mddev->lock);
1903 
1904 	if (conf->geo.near_copies < conf->geo.raid_disks)
1905 		seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1906 	if (conf->geo.near_copies > 1)
1907 		seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1908 	if (conf->geo.far_copies > 1) {
1909 		if (conf->geo.far_offset)
1910 			seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1911 		else
1912 			seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1913 		if (conf->geo.far_set_size != conf->geo.raid_disks)
1914 			seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1915 	}
1916 	seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1917 					conf->geo.raid_disks - mddev->degraded);
1918 	for (i = 0; i < conf->geo.raid_disks; i++) {
1919 		struct md_rdev *rdev = READ_ONCE(conf->mirrors[i].rdev);
1920 
1921 		seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1922 	}
1923 	seq_printf(seq, "]");
1924 }
1925 
1926 /* check if there are enough drives for
1927  * every block to appear on atleast one.
1928  * Don't consider the device numbered 'ignore'
1929  * as we might be about to remove it.
1930  */
_enough(struct r10conf * conf,int previous,int ignore)1931 static int _enough(struct r10conf *conf, int previous, int ignore)
1932 {
1933 	int first = 0;
1934 	int has_enough = 0;
1935 	int disks, ncopies;
1936 	if (previous) {
1937 		disks = conf->prev.raid_disks;
1938 		ncopies = conf->prev.near_copies;
1939 	} else {
1940 		disks = conf->geo.raid_disks;
1941 		ncopies = conf->geo.near_copies;
1942 	}
1943 
1944 	do {
1945 		int n = conf->copies;
1946 		int cnt = 0;
1947 		int this = first;
1948 		while (n--) {
1949 			struct md_rdev *rdev;
1950 			if (this != ignore &&
1951 			    (rdev = conf->mirrors[this].rdev) &&
1952 			    test_bit(In_sync, &rdev->flags))
1953 				cnt++;
1954 			this = (this+1) % disks;
1955 		}
1956 		if (cnt == 0)
1957 			goto out;
1958 		first = (first + ncopies) % disks;
1959 	} while (first != 0);
1960 	has_enough = 1;
1961 out:
1962 	return has_enough;
1963 }
1964 
enough(struct r10conf * conf,int ignore)1965 static int enough(struct r10conf *conf, int ignore)
1966 {
1967 	/* when calling 'enough', both 'prev' and 'geo' must
1968 	 * be stable.
1969 	 * This is ensured if ->reconfig_mutex or ->device_lock
1970 	 * is held.
1971 	 */
1972 	return _enough(conf, 0, ignore) &&
1973 		_enough(conf, 1, ignore);
1974 }
1975 
1976 /**
1977  * raid10_error() - RAID10 error handler.
1978  * @mddev: affected md device.
1979  * @rdev: member device to fail.
1980  *
1981  * The routine acknowledges &rdev failure and determines new @mddev state.
1982  * If it failed, then:
1983  *	- &MD_BROKEN flag is set in &mddev->flags.
1984  * Otherwise, it must be degraded:
1985  *	- recovery is interrupted.
1986  *	- &mddev->degraded is bumped.
1987  *
1988  * @rdev is marked as &Faulty excluding case when array is failed and
1989  * &mddev->fail_last_dev is off.
1990  */
raid10_error(struct mddev * mddev,struct md_rdev * rdev)1991 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1992 {
1993 	struct r10conf *conf = mddev->private;
1994 	unsigned long flags;
1995 
1996 	spin_lock_irqsave(&conf->device_lock, flags);
1997 
1998 	if (test_bit(In_sync, &rdev->flags) && !enough(conf, rdev->raid_disk)) {
1999 		set_bit(MD_BROKEN, &mddev->flags);
2000 
2001 		if (!mddev->fail_last_dev) {
2002 			spin_unlock_irqrestore(&conf->device_lock, flags);
2003 			return;
2004 		}
2005 	}
2006 	if (test_and_clear_bit(In_sync, &rdev->flags))
2007 		mddev->degraded++;
2008 
2009 	set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2010 	set_bit(Blocked, &rdev->flags);
2011 	set_bit(Faulty, &rdev->flags);
2012 	set_mask_bits(&mddev->sb_flags, 0,
2013 		      BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
2014 	spin_unlock_irqrestore(&conf->device_lock, flags);
2015 	pr_crit("md/raid10:%s: Disk failure on %pg, disabling device.\n"
2016 		"md/raid10:%s: Operation continuing on %d devices.\n",
2017 		mdname(mddev), rdev->bdev,
2018 		mdname(mddev), conf->geo.raid_disks - mddev->degraded);
2019 }
2020 
print_conf(struct r10conf * conf)2021 static void print_conf(struct r10conf *conf)
2022 {
2023 	int i;
2024 	struct md_rdev *rdev;
2025 
2026 	pr_debug("RAID10 conf printout:\n");
2027 	if (!conf) {
2028 		pr_debug("(!conf)\n");
2029 		return;
2030 	}
2031 	pr_debug(" --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
2032 		 conf->geo.raid_disks);
2033 
2034 	lockdep_assert_held(&conf->mddev->reconfig_mutex);
2035 	for (i = 0; i < conf->geo.raid_disks; i++) {
2036 		rdev = conf->mirrors[i].rdev;
2037 		if (rdev)
2038 			pr_debug(" disk %d, wo:%d, o:%d, dev:%pg\n",
2039 				 i, !test_bit(In_sync, &rdev->flags),
2040 				 !test_bit(Faulty, &rdev->flags),
2041 				 rdev->bdev);
2042 	}
2043 }
2044 
close_sync(struct r10conf * conf)2045 static void close_sync(struct r10conf *conf)
2046 {
2047 	wait_barrier(conf, false);
2048 	allow_barrier(conf);
2049 
2050 	mempool_exit(&conf->r10buf_pool);
2051 }
2052 
raid10_spare_active(struct mddev * mddev)2053 static int raid10_spare_active(struct mddev *mddev)
2054 {
2055 	int i;
2056 	struct r10conf *conf = mddev->private;
2057 	struct raid10_info *tmp;
2058 	int count = 0;
2059 	unsigned long flags;
2060 
2061 	/*
2062 	 * Find all non-in_sync disks within the RAID10 configuration
2063 	 * and mark them in_sync
2064 	 */
2065 	for (i = 0; i < conf->geo.raid_disks; i++) {
2066 		tmp = conf->mirrors + i;
2067 		if (tmp->replacement
2068 		    && tmp->replacement->recovery_offset == MaxSector
2069 		    && !test_bit(Faulty, &tmp->replacement->flags)
2070 		    && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
2071 			/* Replacement has just become active */
2072 			if (!tmp->rdev
2073 			    || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
2074 				count++;
2075 			if (tmp->rdev) {
2076 				/* Replaced device not technically faulty,
2077 				 * but we need to be sure it gets removed
2078 				 * and never re-added.
2079 				 */
2080 				set_bit(Faulty, &tmp->rdev->flags);
2081 				sysfs_notify_dirent_safe(
2082 					tmp->rdev->sysfs_state);
2083 			}
2084 			sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
2085 		} else if (tmp->rdev
2086 			   && tmp->rdev->recovery_offset == MaxSector
2087 			   && !test_bit(Faulty, &tmp->rdev->flags)
2088 			   && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
2089 			count++;
2090 			sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
2091 		}
2092 	}
2093 	spin_lock_irqsave(&conf->device_lock, flags);
2094 	mddev->degraded -= count;
2095 	spin_unlock_irqrestore(&conf->device_lock, flags);
2096 
2097 	print_conf(conf);
2098 	return count;
2099 }
2100 
raid10_add_disk(struct mddev * mddev,struct md_rdev * rdev)2101 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
2102 {
2103 	struct r10conf *conf = mddev->private;
2104 	int err = -EEXIST;
2105 	int mirror, repl_slot = -1;
2106 	int first = 0;
2107 	int last = conf->geo.raid_disks - 1;
2108 	struct raid10_info *p;
2109 
2110 	if (mddev->recovery_cp < MaxSector)
2111 		/* only hot-add to in-sync arrays, as recovery is
2112 		 * very different from resync
2113 		 */
2114 		return -EBUSY;
2115 	if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
2116 		return -EINVAL;
2117 
2118 	if (rdev->raid_disk >= 0)
2119 		first = last = rdev->raid_disk;
2120 
2121 	if (rdev->saved_raid_disk >= first &&
2122 	    rdev->saved_raid_disk < conf->geo.raid_disks &&
2123 	    conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
2124 		mirror = rdev->saved_raid_disk;
2125 	else
2126 		mirror = first;
2127 	for ( ; mirror <= last ; mirror++) {
2128 		p = &conf->mirrors[mirror];
2129 		if (p->recovery_disabled == mddev->recovery_disabled)
2130 			continue;
2131 		if (p->rdev) {
2132 			if (test_bit(WantReplacement, &p->rdev->flags) &&
2133 			    p->replacement == NULL && repl_slot < 0)
2134 				repl_slot = mirror;
2135 			continue;
2136 		}
2137 
2138 		err = mddev_stack_new_rdev(mddev, rdev);
2139 		if (err)
2140 			return err;
2141 		p->head_position = 0;
2142 		p->recovery_disabled = mddev->recovery_disabled - 1;
2143 		rdev->raid_disk = mirror;
2144 		err = 0;
2145 		if (rdev->saved_raid_disk != mirror)
2146 			conf->fullsync = 1;
2147 		WRITE_ONCE(p->rdev, rdev);
2148 		break;
2149 	}
2150 
2151 	if (err && repl_slot >= 0) {
2152 		p = &conf->mirrors[repl_slot];
2153 		clear_bit(In_sync, &rdev->flags);
2154 		set_bit(Replacement, &rdev->flags);
2155 		rdev->raid_disk = repl_slot;
2156 		err = mddev_stack_new_rdev(mddev, rdev);
2157 		if (err)
2158 			return err;
2159 		conf->fullsync = 1;
2160 		WRITE_ONCE(p->replacement, rdev);
2161 	}
2162 
2163 	print_conf(conf);
2164 	return err;
2165 }
2166 
raid10_remove_disk(struct mddev * mddev,struct md_rdev * rdev)2167 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
2168 {
2169 	struct r10conf *conf = mddev->private;
2170 	int err = 0;
2171 	int number = rdev->raid_disk;
2172 	struct md_rdev **rdevp;
2173 	struct raid10_info *p;
2174 
2175 	print_conf(conf);
2176 	if (unlikely(number >= mddev->raid_disks))
2177 		return 0;
2178 	p = conf->mirrors + number;
2179 	if (rdev == p->rdev)
2180 		rdevp = &p->rdev;
2181 	else if (rdev == p->replacement)
2182 		rdevp = &p->replacement;
2183 	else
2184 		return 0;
2185 
2186 	if (test_bit(In_sync, &rdev->flags) ||
2187 	    atomic_read(&rdev->nr_pending)) {
2188 		err = -EBUSY;
2189 		goto abort;
2190 	}
2191 	/* Only remove non-faulty devices if recovery
2192 	 * is not possible.
2193 	 */
2194 	if (!test_bit(Faulty, &rdev->flags) &&
2195 	    mddev->recovery_disabled != p->recovery_disabled &&
2196 	    (!p->replacement || p->replacement == rdev) &&
2197 	    number < conf->geo.raid_disks &&
2198 	    enough(conf, -1)) {
2199 		err = -EBUSY;
2200 		goto abort;
2201 	}
2202 	WRITE_ONCE(*rdevp, NULL);
2203 	if (p->replacement) {
2204 		/* We must have just cleared 'rdev' */
2205 		WRITE_ONCE(p->rdev, p->replacement);
2206 		clear_bit(Replacement, &p->replacement->flags);
2207 		WRITE_ONCE(p->replacement, NULL);
2208 	}
2209 
2210 	clear_bit(WantReplacement, &rdev->flags);
2211 	err = md_integrity_register(mddev);
2212 
2213 abort:
2214 
2215 	print_conf(conf);
2216 	return err;
2217 }
2218 
__end_sync_read(struct r10bio * r10_bio,struct bio * bio,int d)2219 static void __end_sync_read(struct r10bio *r10_bio, struct bio *bio, int d)
2220 {
2221 	struct r10conf *conf = r10_bio->mddev->private;
2222 
2223 	if (!bio->bi_status)
2224 		set_bit(R10BIO_Uptodate, &r10_bio->state);
2225 	else
2226 		/* The write handler will notice the lack of
2227 		 * R10BIO_Uptodate and record any errors etc
2228 		 */
2229 		atomic_add(r10_bio->sectors,
2230 			   &conf->mirrors[d].rdev->corrected_errors);
2231 
2232 	/* for reconstruct, we always reschedule after a read.
2233 	 * for resync, only after all reads
2234 	 */
2235 	rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
2236 	if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
2237 	    atomic_dec_and_test(&r10_bio->remaining)) {
2238 		/* we have read all the blocks,
2239 		 * do the comparison in process context in raid10d
2240 		 */
2241 		reschedule_retry(r10_bio);
2242 	}
2243 }
2244 
end_sync_read(struct bio * bio)2245 static void end_sync_read(struct bio *bio)
2246 {
2247 	struct r10bio *r10_bio = get_resync_r10bio(bio);
2248 	struct r10conf *conf = r10_bio->mddev->private;
2249 	int d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
2250 
2251 	__end_sync_read(r10_bio, bio, d);
2252 }
2253 
end_reshape_read(struct bio * bio)2254 static void end_reshape_read(struct bio *bio)
2255 {
2256 	/* reshape read bio isn't allocated from r10buf_pool */
2257 	struct r10bio *r10_bio = bio->bi_private;
2258 
2259 	__end_sync_read(r10_bio, bio, r10_bio->read_slot);
2260 }
2261 
end_sync_request(struct r10bio * r10_bio)2262 static void end_sync_request(struct r10bio *r10_bio)
2263 {
2264 	struct mddev *mddev = r10_bio->mddev;
2265 
2266 	while (atomic_dec_and_test(&r10_bio->remaining)) {
2267 		if (r10_bio->master_bio == NULL) {
2268 			/* the primary of several recovery bios */
2269 			sector_t s = r10_bio->sectors;
2270 			if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2271 			    test_bit(R10BIO_WriteError, &r10_bio->state))
2272 				reschedule_retry(r10_bio);
2273 			else
2274 				put_buf(r10_bio);
2275 			md_done_sync(mddev, s, 1);
2276 			break;
2277 		} else {
2278 			struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
2279 			if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2280 			    test_bit(R10BIO_WriteError, &r10_bio->state))
2281 				reschedule_retry(r10_bio);
2282 			else
2283 				put_buf(r10_bio);
2284 			r10_bio = r10_bio2;
2285 		}
2286 	}
2287 }
2288 
end_sync_write(struct bio * bio)2289 static void end_sync_write(struct bio *bio)
2290 {
2291 	struct r10bio *r10_bio = get_resync_r10bio(bio);
2292 	struct mddev *mddev = r10_bio->mddev;
2293 	struct r10conf *conf = mddev->private;
2294 	int d;
2295 	int slot;
2296 	int repl;
2297 	struct md_rdev *rdev = NULL;
2298 
2299 	d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
2300 	if (repl)
2301 		rdev = conf->mirrors[d].replacement;
2302 	else
2303 		rdev = conf->mirrors[d].rdev;
2304 
2305 	if (bio->bi_status) {
2306 		if (repl)
2307 			md_error(mddev, rdev);
2308 		else {
2309 			set_bit(WriteErrorSeen, &rdev->flags);
2310 			if (!test_and_set_bit(WantReplacement, &rdev->flags))
2311 				set_bit(MD_RECOVERY_NEEDED,
2312 					&rdev->mddev->recovery);
2313 			set_bit(R10BIO_WriteError, &r10_bio->state);
2314 		}
2315 	} else if (rdev_has_badblock(rdev, r10_bio->devs[slot].addr,
2316 				     r10_bio->sectors)) {
2317 		set_bit(R10BIO_MadeGood, &r10_bio->state);
2318 	}
2319 
2320 	rdev_dec_pending(rdev, mddev);
2321 
2322 	end_sync_request(r10_bio);
2323 }
2324 
2325 /*
2326  * Note: sync and recover and handled very differently for raid10
2327  * This code is for resync.
2328  * For resync, we read through virtual addresses and read all blocks.
2329  * If there is any error, we schedule a write.  The lowest numbered
2330  * drive is authoritative.
2331  * However requests come for physical address, so we need to map.
2332  * For every physical address there are raid_disks/copies virtual addresses,
2333  * which is always are least one, but is not necessarly an integer.
2334  * This means that a physical address can span multiple chunks, so we may
2335  * have to submit multiple io requests for a single sync request.
2336  */
2337 /*
2338  * We check if all blocks are in-sync and only write to blocks that
2339  * aren't in sync
2340  */
sync_request_write(struct mddev * mddev,struct r10bio * r10_bio)2341 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2342 {
2343 	struct r10conf *conf = mddev->private;
2344 	int i, first;
2345 	struct bio *tbio, *fbio;
2346 	int vcnt;
2347 	struct page **tpages, **fpages;
2348 
2349 	atomic_set(&r10_bio->remaining, 1);
2350 
2351 	/* find the first device with a block */
2352 	for (i=0; i<conf->copies; i++)
2353 		if (!r10_bio->devs[i].bio->bi_status)
2354 			break;
2355 
2356 	if (i == conf->copies)
2357 		goto done;
2358 
2359 	first = i;
2360 	fbio = r10_bio->devs[i].bio;
2361 	fbio->bi_iter.bi_size = r10_bio->sectors << 9;
2362 	fbio->bi_iter.bi_idx = 0;
2363 	fpages = get_resync_pages(fbio)->pages;
2364 
2365 	vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2366 	/* now find blocks with errors */
2367 	for (i=0 ; i < conf->copies ; i++) {
2368 		int  j, d;
2369 		struct md_rdev *rdev;
2370 		struct resync_pages *rp;
2371 
2372 		tbio = r10_bio->devs[i].bio;
2373 
2374 		if (tbio->bi_end_io != end_sync_read)
2375 			continue;
2376 		if (i == first)
2377 			continue;
2378 
2379 		tpages = get_resync_pages(tbio)->pages;
2380 		d = r10_bio->devs[i].devnum;
2381 		rdev = conf->mirrors[d].rdev;
2382 		if (!r10_bio->devs[i].bio->bi_status) {
2383 			/* We know that the bi_io_vec layout is the same for
2384 			 * both 'first' and 'i', so we just compare them.
2385 			 * All vec entries are PAGE_SIZE;
2386 			 */
2387 			int sectors = r10_bio->sectors;
2388 			for (j = 0; j < vcnt; j++) {
2389 				int len = PAGE_SIZE;
2390 				if (sectors < (len / 512))
2391 					len = sectors * 512;
2392 				if (memcmp(page_address(fpages[j]),
2393 					   page_address(tpages[j]),
2394 					   len))
2395 					break;
2396 				sectors -= len/512;
2397 			}
2398 			if (j == vcnt)
2399 				continue;
2400 			atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2401 			if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2402 				/* Don't fix anything. */
2403 				continue;
2404 		} else if (test_bit(FailFast, &rdev->flags)) {
2405 			/* Just give up on this device */
2406 			md_error(rdev->mddev, rdev);
2407 			continue;
2408 		}
2409 		/* Ok, we need to write this bio, either to correct an
2410 		 * inconsistency or to correct an unreadable block.
2411 		 * First we need to fixup bv_offset, bv_len and
2412 		 * bi_vecs, as the read request might have corrupted these
2413 		 */
2414 		rp = get_resync_pages(tbio);
2415 		bio_reset(tbio, conf->mirrors[d].rdev->bdev, REQ_OP_WRITE);
2416 
2417 		md_bio_reset_resync_pages(tbio, rp, fbio->bi_iter.bi_size);
2418 
2419 		rp->raid_bio = r10_bio;
2420 		tbio->bi_private = rp;
2421 		tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
2422 		tbio->bi_end_io = end_sync_write;
2423 
2424 		bio_copy_data(tbio, fbio);
2425 
2426 		atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2427 		atomic_inc(&r10_bio->remaining);
2428 		md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2429 
2430 		if (test_bit(FailFast, &conf->mirrors[d].rdev->flags))
2431 			tbio->bi_opf |= MD_FAILFAST;
2432 		tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2433 		submit_bio_noacct(tbio);
2434 	}
2435 
2436 	/* Now write out to any replacement devices
2437 	 * that are active
2438 	 */
2439 	for (i = 0; i < conf->copies; i++) {
2440 		int d;
2441 
2442 		tbio = r10_bio->devs[i].repl_bio;
2443 		if (!tbio || !tbio->bi_end_io)
2444 			continue;
2445 		if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2446 		    && r10_bio->devs[i].bio != fbio)
2447 			bio_copy_data(tbio, fbio);
2448 		d = r10_bio->devs[i].devnum;
2449 		atomic_inc(&r10_bio->remaining);
2450 		md_sync_acct(conf->mirrors[d].replacement->bdev,
2451 			     bio_sectors(tbio));
2452 		submit_bio_noacct(tbio);
2453 	}
2454 
2455 done:
2456 	if (atomic_dec_and_test(&r10_bio->remaining)) {
2457 		md_done_sync(mddev, r10_bio->sectors, 1);
2458 		put_buf(r10_bio);
2459 	}
2460 }
2461 
2462 /*
2463  * Now for the recovery code.
2464  * Recovery happens across physical sectors.
2465  * We recover all non-is_sync drives by finding the virtual address of
2466  * each, and then choose a working drive that also has that virt address.
2467  * There is a separate r10_bio for each non-in_sync drive.
2468  * Only the first two slots are in use. The first for reading,
2469  * The second for writing.
2470  *
2471  */
fix_recovery_read_error(struct r10bio * r10_bio)2472 static void fix_recovery_read_error(struct r10bio *r10_bio)
2473 {
2474 	/* We got a read error during recovery.
2475 	 * We repeat the read in smaller page-sized sections.
2476 	 * If a read succeeds, write it to the new device or record
2477 	 * a bad block if we cannot.
2478 	 * If a read fails, record a bad block on both old and
2479 	 * new devices.
2480 	 */
2481 	struct mddev *mddev = r10_bio->mddev;
2482 	struct r10conf *conf = mddev->private;
2483 	struct bio *bio = r10_bio->devs[0].bio;
2484 	sector_t sect = 0;
2485 	int sectors = r10_bio->sectors;
2486 	int idx = 0;
2487 	int dr = r10_bio->devs[0].devnum;
2488 	int dw = r10_bio->devs[1].devnum;
2489 	struct page **pages = get_resync_pages(bio)->pages;
2490 
2491 	while (sectors) {
2492 		int s = sectors;
2493 		struct md_rdev *rdev;
2494 		sector_t addr;
2495 		int ok;
2496 
2497 		if (s > (PAGE_SIZE>>9))
2498 			s = PAGE_SIZE >> 9;
2499 
2500 		rdev = conf->mirrors[dr].rdev;
2501 		addr = r10_bio->devs[0].addr + sect;
2502 		ok = sync_page_io(rdev,
2503 				  addr,
2504 				  s << 9,
2505 				  pages[idx],
2506 				  REQ_OP_READ, false);
2507 		if (ok) {
2508 			rdev = conf->mirrors[dw].rdev;
2509 			addr = r10_bio->devs[1].addr + sect;
2510 			ok = sync_page_io(rdev,
2511 					  addr,
2512 					  s << 9,
2513 					  pages[idx],
2514 					  REQ_OP_WRITE, false);
2515 			if (!ok) {
2516 				set_bit(WriteErrorSeen, &rdev->flags);
2517 				if (!test_and_set_bit(WantReplacement,
2518 						      &rdev->flags))
2519 					set_bit(MD_RECOVERY_NEEDED,
2520 						&rdev->mddev->recovery);
2521 			}
2522 		}
2523 		if (!ok) {
2524 			/* We don't worry if we cannot set a bad block -
2525 			 * it really is bad so there is no loss in not
2526 			 * recording it yet
2527 			 */
2528 			rdev_set_badblocks(rdev, addr, s, 0);
2529 
2530 			if (rdev != conf->mirrors[dw].rdev) {
2531 				/* need bad block on destination too */
2532 				struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2533 				addr = r10_bio->devs[1].addr + sect;
2534 				ok = rdev_set_badblocks(rdev2, addr, s, 0);
2535 				if (!ok) {
2536 					/* just abort the recovery */
2537 					pr_notice("md/raid10:%s: recovery aborted due to read error\n",
2538 						  mdname(mddev));
2539 
2540 					conf->mirrors[dw].recovery_disabled
2541 						= mddev->recovery_disabled;
2542 					set_bit(MD_RECOVERY_INTR,
2543 						&mddev->recovery);
2544 					break;
2545 				}
2546 			}
2547 		}
2548 
2549 		sectors -= s;
2550 		sect += s;
2551 		idx++;
2552 	}
2553 }
2554 
recovery_request_write(struct mddev * mddev,struct r10bio * r10_bio)2555 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2556 {
2557 	struct r10conf *conf = mddev->private;
2558 	int d;
2559 	struct bio *wbio = r10_bio->devs[1].bio;
2560 	struct bio *wbio2 = r10_bio->devs[1].repl_bio;
2561 
2562 	/* Need to test wbio2->bi_end_io before we call
2563 	 * submit_bio_noacct as if the former is NULL,
2564 	 * the latter is free to free wbio2.
2565 	 */
2566 	if (wbio2 && !wbio2->bi_end_io)
2567 		wbio2 = NULL;
2568 
2569 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2570 		fix_recovery_read_error(r10_bio);
2571 		if (wbio->bi_end_io)
2572 			end_sync_request(r10_bio);
2573 		if (wbio2)
2574 			end_sync_request(r10_bio);
2575 		return;
2576 	}
2577 
2578 	/*
2579 	 * share the pages with the first bio
2580 	 * and submit the write request
2581 	 */
2582 	d = r10_bio->devs[1].devnum;
2583 	if (wbio->bi_end_io) {
2584 		atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2585 		md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2586 		submit_bio_noacct(wbio);
2587 	}
2588 	if (wbio2) {
2589 		atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2590 		md_sync_acct(conf->mirrors[d].replacement->bdev,
2591 			     bio_sectors(wbio2));
2592 		submit_bio_noacct(wbio2);
2593 	}
2594 }
2595 
r10_sync_page_io(struct md_rdev * rdev,sector_t sector,int sectors,struct page * page,enum req_op op)2596 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2597 			    int sectors, struct page *page, enum req_op op)
2598 {
2599 	if (rdev_has_badblock(rdev, sector, sectors) &&
2600 	    (op == REQ_OP_READ || test_bit(WriteErrorSeen, &rdev->flags)))
2601 		return -1;
2602 	if (sync_page_io(rdev, sector, sectors << 9, page, op, false))
2603 		/* success */
2604 		return 1;
2605 	if (op == REQ_OP_WRITE) {
2606 		set_bit(WriteErrorSeen, &rdev->flags);
2607 		if (!test_and_set_bit(WantReplacement, &rdev->flags))
2608 			set_bit(MD_RECOVERY_NEEDED,
2609 				&rdev->mddev->recovery);
2610 	}
2611 	/* need to record an error - either for the block or the device */
2612 	if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2613 		md_error(rdev->mddev, rdev);
2614 	return 0;
2615 }
2616 
2617 /*
2618  * This is a kernel thread which:
2619  *
2620  *	1.	Retries failed read operations on working mirrors.
2621  *	2.	Updates the raid superblock when problems encounter.
2622  *	3.	Performs writes following reads for array synchronising.
2623  */
2624 
fix_read_error(struct r10conf * conf,struct mddev * mddev,struct r10bio * r10_bio)2625 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2626 {
2627 	int sect = 0; /* Offset from r10_bio->sector */
2628 	int sectors = r10_bio->sectors, slot = r10_bio->read_slot;
2629 	struct md_rdev *rdev;
2630 	int d = r10_bio->devs[slot].devnum;
2631 
2632 	/* still own a reference to this rdev, so it cannot
2633 	 * have been cleared recently.
2634 	 */
2635 	rdev = conf->mirrors[d].rdev;
2636 
2637 	if (test_bit(Faulty, &rdev->flags))
2638 		/* drive has already been failed, just ignore any
2639 		   more fix_read_error() attempts */
2640 		return;
2641 
2642 	if (exceed_read_errors(mddev, rdev)) {
2643 		r10_bio->devs[slot].bio = IO_BLOCKED;
2644 		return;
2645 	}
2646 
2647 	while(sectors) {
2648 		int s = sectors;
2649 		int sl = slot;
2650 		int success = 0;
2651 		int start;
2652 
2653 		if (s > (PAGE_SIZE>>9))
2654 			s = PAGE_SIZE >> 9;
2655 
2656 		do {
2657 			d = r10_bio->devs[sl].devnum;
2658 			rdev = conf->mirrors[d].rdev;
2659 			if (rdev &&
2660 			    test_bit(In_sync, &rdev->flags) &&
2661 			    !test_bit(Faulty, &rdev->flags) &&
2662 			    rdev_has_badblock(rdev,
2663 					      r10_bio->devs[sl].addr + sect,
2664 					      s) == 0) {
2665 				atomic_inc(&rdev->nr_pending);
2666 				success = sync_page_io(rdev,
2667 						       r10_bio->devs[sl].addr +
2668 						       sect,
2669 						       s<<9,
2670 						       conf->tmppage,
2671 						       REQ_OP_READ, false);
2672 				rdev_dec_pending(rdev, mddev);
2673 				if (success)
2674 					break;
2675 			}
2676 			sl++;
2677 			if (sl == conf->copies)
2678 				sl = 0;
2679 		} while (sl != slot);
2680 
2681 		if (!success) {
2682 			/* Cannot read from anywhere, just mark the block
2683 			 * as bad on the first device to discourage future
2684 			 * reads.
2685 			 */
2686 			int dn = r10_bio->devs[slot].devnum;
2687 			rdev = conf->mirrors[dn].rdev;
2688 
2689 			if (!rdev_set_badblocks(
2690 				    rdev,
2691 				    r10_bio->devs[slot].addr
2692 				    + sect,
2693 				    s, 0)) {
2694 				md_error(mddev, rdev);
2695 				r10_bio->devs[slot].bio
2696 					= IO_BLOCKED;
2697 			}
2698 			break;
2699 		}
2700 
2701 		start = sl;
2702 		/* write it back and re-read */
2703 		while (sl != slot) {
2704 			if (sl==0)
2705 				sl = conf->copies;
2706 			sl--;
2707 			d = r10_bio->devs[sl].devnum;
2708 			rdev = conf->mirrors[d].rdev;
2709 			if (!rdev ||
2710 			    test_bit(Faulty, &rdev->flags) ||
2711 			    !test_bit(In_sync, &rdev->flags))
2712 				continue;
2713 
2714 			atomic_inc(&rdev->nr_pending);
2715 			if (r10_sync_page_io(rdev,
2716 					     r10_bio->devs[sl].addr +
2717 					     sect,
2718 					     s, conf->tmppage, REQ_OP_WRITE)
2719 			    == 0) {
2720 				/* Well, this device is dead */
2721 				pr_notice("md/raid10:%s: read correction write failed (%d sectors at %llu on %pg)\n",
2722 					  mdname(mddev), s,
2723 					  (unsigned long long)(
2724 						  sect +
2725 						  choose_data_offset(r10_bio,
2726 								     rdev)),
2727 					  rdev->bdev);
2728 				pr_notice("md/raid10:%s: %pg: failing drive\n",
2729 					  mdname(mddev),
2730 					  rdev->bdev);
2731 			}
2732 			rdev_dec_pending(rdev, mddev);
2733 		}
2734 		sl = start;
2735 		while (sl != slot) {
2736 			if (sl==0)
2737 				sl = conf->copies;
2738 			sl--;
2739 			d = r10_bio->devs[sl].devnum;
2740 			rdev = conf->mirrors[d].rdev;
2741 			if (!rdev ||
2742 			    test_bit(Faulty, &rdev->flags) ||
2743 			    !test_bit(In_sync, &rdev->flags))
2744 				continue;
2745 
2746 			atomic_inc(&rdev->nr_pending);
2747 			switch (r10_sync_page_io(rdev,
2748 					     r10_bio->devs[sl].addr +
2749 					     sect,
2750 					     s, conf->tmppage, REQ_OP_READ)) {
2751 			case 0:
2752 				/* Well, this device is dead */
2753 				pr_notice("md/raid10:%s: unable to read back corrected sectors (%d sectors at %llu on %pg)\n",
2754 				       mdname(mddev), s,
2755 				       (unsigned long long)(
2756 					       sect +
2757 					       choose_data_offset(r10_bio, rdev)),
2758 				       rdev->bdev);
2759 				pr_notice("md/raid10:%s: %pg: failing drive\n",
2760 				       mdname(mddev),
2761 				       rdev->bdev);
2762 				break;
2763 			case 1:
2764 				pr_info("md/raid10:%s: read error corrected (%d sectors at %llu on %pg)\n",
2765 				       mdname(mddev), s,
2766 				       (unsigned long long)(
2767 					       sect +
2768 					       choose_data_offset(r10_bio, rdev)),
2769 				       rdev->bdev);
2770 				atomic_add(s, &rdev->corrected_errors);
2771 			}
2772 
2773 			rdev_dec_pending(rdev, mddev);
2774 		}
2775 
2776 		sectors -= s;
2777 		sect += s;
2778 	}
2779 }
2780 
narrow_write_error(struct r10bio * r10_bio,int i)2781 static bool narrow_write_error(struct r10bio *r10_bio, int i)
2782 {
2783 	struct bio *bio = r10_bio->master_bio;
2784 	struct mddev *mddev = r10_bio->mddev;
2785 	struct r10conf *conf = mddev->private;
2786 	struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2787 	/* bio has the data to be written to slot 'i' where
2788 	 * we just recently had a write error.
2789 	 * We repeatedly clone the bio and trim down to one block,
2790 	 * then try the write.  Where the write fails we record
2791 	 * a bad block.
2792 	 * It is conceivable that the bio doesn't exactly align with
2793 	 * blocks.  We must handle this.
2794 	 *
2795 	 * We currently own a reference to the rdev.
2796 	 */
2797 
2798 	int block_sectors;
2799 	sector_t sector;
2800 	int sectors;
2801 	int sect_to_write = r10_bio->sectors;
2802 	bool ok = true;
2803 
2804 	if (rdev->badblocks.shift < 0)
2805 		return false;
2806 
2807 	block_sectors = roundup(1 << rdev->badblocks.shift,
2808 				bdev_logical_block_size(rdev->bdev) >> 9);
2809 	sector = r10_bio->sector;
2810 	sectors = ((r10_bio->sector + block_sectors)
2811 		   & ~(sector_t)(block_sectors - 1))
2812 		- sector;
2813 
2814 	while (sect_to_write) {
2815 		struct bio *wbio;
2816 		sector_t wsector;
2817 		if (sectors > sect_to_write)
2818 			sectors = sect_to_write;
2819 		/* Write at 'sector' for 'sectors' */
2820 		wbio = bio_alloc_clone(rdev->bdev, bio, GFP_NOIO,
2821 				       &mddev->bio_set);
2822 		bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2823 		wsector = r10_bio->devs[i].addr + (sector - r10_bio->sector);
2824 		wbio->bi_iter.bi_sector = wsector +
2825 				   choose_data_offset(r10_bio, rdev);
2826 		wbio->bi_opf = REQ_OP_WRITE;
2827 
2828 		if (submit_bio_wait(wbio) < 0)
2829 			/* Failure! */
2830 			ok = rdev_set_badblocks(rdev, wsector,
2831 						sectors, 0)
2832 				&& ok;
2833 
2834 		bio_put(wbio);
2835 		sect_to_write -= sectors;
2836 		sector += sectors;
2837 		sectors = block_sectors;
2838 	}
2839 	return ok;
2840 }
2841 
handle_read_error(struct mddev * mddev,struct r10bio * r10_bio)2842 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2843 {
2844 	int slot = r10_bio->read_slot;
2845 	struct bio *bio;
2846 	struct r10conf *conf = mddev->private;
2847 	struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2848 
2849 	/* we got a read error. Maybe the drive is bad.  Maybe just
2850 	 * the block and we can fix it.
2851 	 * We freeze all other IO, and try reading the block from
2852 	 * other devices.  When we find one, we re-write
2853 	 * and check it that fixes the read error.
2854 	 * This is all done synchronously while the array is
2855 	 * frozen.
2856 	 */
2857 	bio = r10_bio->devs[slot].bio;
2858 	bio_put(bio);
2859 	r10_bio->devs[slot].bio = NULL;
2860 
2861 	if (mddev->ro)
2862 		r10_bio->devs[slot].bio = IO_BLOCKED;
2863 	else if (!test_bit(FailFast, &rdev->flags)) {
2864 		freeze_array(conf, 1);
2865 		fix_read_error(conf, mddev, r10_bio);
2866 		unfreeze_array(conf);
2867 	} else
2868 		md_error(mddev, rdev);
2869 
2870 	rdev_dec_pending(rdev, mddev);
2871 	r10_bio->state = 0;
2872 	raid10_read_request(mddev, r10_bio->master_bio, r10_bio, false);
2873 	/*
2874 	 * allow_barrier after re-submit to ensure no sync io
2875 	 * can be issued while regular io pending.
2876 	 */
2877 	allow_barrier(conf);
2878 }
2879 
handle_write_completed(struct r10conf * conf,struct r10bio * r10_bio)2880 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2881 {
2882 	/* Some sort of write request has finished and it
2883 	 * succeeded in writing where we thought there was a
2884 	 * bad block.  So forget the bad block.
2885 	 * Or possibly if failed and we need to record
2886 	 * a bad block.
2887 	 */
2888 	int m;
2889 	struct md_rdev *rdev;
2890 
2891 	if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2892 	    test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2893 		for (m = 0; m < conf->copies; m++) {
2894 			int dev = r10_bio->devs[m].devnum;
2895 			rdev = conf->mirrors[dev].rdev;
2896 			if (r10_bio->devs[m].bio == NULL ||
2897 				r10_bio->devs[m].bio->bi_end_io == NULL)
2898 				continue;
2899 			if (!r10_bio->devs[m].bio->bi_status) {
2900 				rdev_clear_badblocks(
2901 					rdev,
2902 					r10_bio->devs[m].addr,
2903 					r10_bio->sectors, 0);
2904 			} else {
2905 				if (!rdev_set_badblocks(
2906 					    rdev,
2907 					    r10_bio->devs[m].addr,
2908 					    r10_bio->sectors, 0))
2909 					md_error(conf->mddev, rdev);
2910 			}
2911 			rdev = conf->mirrors[dev].replacement;
2912 			if (r10_bio->devs[m].repl_bio == NULL ||
2913 				r10_bio->devs[m].repl_bio->bi_end_io == NULL)
2914 				continue;
2915 
2916 			if (!r10_bio->devs[m].repl_bio->bi_status) {
2917 				rdev_clear_badblocks(
2918 					rdev,
2919 					r10_bio->devs[m].addr,
2920 					r10_bio->sectors, 0);
2921 			} else {
2922 				if (!rdev_set_badblocks(
2923 					    rdev,
2924 					    r10_bio->devs[m].addr,
2925 					    r10_bio->sectors, 0))
2926 					md_error(conf->mddev, rdev);
2927 			}
2928 		}
2929 		put_buf(r10_bio);
2930 	} else {
2931 		bool fail = false;
2932 		for (m = 0; m < conf->copies; m++) {
2933 			int dev = r10_bio->devs[m].devnum;
2934 			struct bio *bio = r10_bio->devs[m].bio;
2935 			rdev = conf->mirrors[dev].rdev;
2936 			if (bio == IO_MADE_GOOD) {
2937 				rdev_clear_badblocks(
2938 					rdev,
2939 					r10_bio->devs[m].addr,
2940 					r10_bio->sectors, 0);
2941 				rdev_dec_pending(rdev, conf->mddev);
2942 			} else if (bio != NULL && bio->bi_status) {
2943 				fail = true;
2944 				if (!narrow_write_error(r10_bio, m))
2945 					md_error(conf->mddev, rdev);
2946 				rdev_dec_pending(rdev, conf->mddev);
2947 			}
2948 			bio = r10_bio->devs[m].repl_bio;
2949 			rdev = conf->mirrors[dev].replacement;
2950 			if (rdev && bio == IO_MADE_GOOD) {
2951 				rdev_clear_badblocks(
2952 					rdev,
2953 					r10_bio->devs[m].addr,
2954 					r10_bio->sectors, 0);
2955 				rdev_dec_pending(rdev, conf->mddev);
2956 			}
2957 		}
2958 		if (fail) {
2959 			spin_lock_irq(&conf->device_lock);
2960 			list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2961 			conf->nr_queued++;
2962 			spin_unlock_irq(&conf->device_lock);
2963 			/*
2964 			 * In case freeze_array() is waiting for condition
2965 			 * nr_pending == nr_queued + extra to be true.
2966 			 */
2967 			wake_up(&conf->wait_barrier);
2968 			md_wakeup_thread(conf->mddev->thread);
2969 		} else {
2970 			if (test_bit(R10BIO_WriteError,
2971 				     &r10_bio->state))
2972 				close_write(r10_bio);
2973 			raid_end_bio_io(r10_bio);
2974 		}
2975 	}
2976 }
2977 
raid10d(struct md_thread * thread)2978 static void raid10d(struct md_thread *thread)
2979 {
2980 	struct mddev *mddev = thread->mddev;
2981 	struct r10bio *r10_bio;
2982 	unsigned long flags;
2983 	struct r10conf *conf = mddev->private;
2984 	struct list_head *head = &conf->retry_list;
2985 	struct blk_plug plug;
2986 
2987 	md_check_recovery(mddev);
2988 
2989 	if (!list_empty_careful(&conf->bio_end_io_list) &&
2990 	    !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2991 		LIST_HEAD(tmp);
2992 		spin_lock_irqsave(&conf->device_lock, flags);
2993 		if (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2994 			while (!list_empty(&conf->bio_end_io_list)) {
2995 				list_move(conf->bio_end_io_list.prev, &tmp);
2996 				conf->nr_queued--;
2997 			}
2998 		}
2999 		spin_unlock_irqrestore(&conf->device_lock, flags);
3000 		while (!list_empty(&tmp)) {
3001 			r10_bio = list_first_entry(&tmp, struct r10bio,
3002 						   retry_list);
3003 			list_del(&r10_bio->retry_list);
3004 
3005 			if (test_bit(R10BIO_WriteError,
3006 				     &r10_bio->state))
3007 				close_write(r10_bio);
3008 			raid_end_bio_io(r10_bio);
3009 		}
3010 	}
3011 
3012 	blk_start_plug(&plug);
3013 	for (;;) {
3014 
3015 		flush_pending_writes(conf);
3016 
3017 		spin_lock_irqsave(&conf->device_lock, flags);
3018 		if (list_empty(head)) {
3019 			spin_unlock_irqrestore(&conf->device_lock, flags);
3020 			break;
3021 		}
3022 		r10_bio = list_entry(head->prev, struct r10bio, retry_list);
3023 		list_del(head->prev);
3024 		conf->nr_queued--;
3025 		spin_unlock_irqrestore(&conf->device_lock, flags);
3026 
3027 		mddev = r10_bio->mddev;
3028 		conf = mddev->private;
3029 		if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
3030 		    test_bit(R10BIO_WriteError, &r10_bio->state))
3031 			handle_write_completed(conf, r10_bio);
3032 		else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
3033 			reshape_request_write(mddev, r10_bio);
3034 		else if (test_bit(R10BIO_IsSync, &r10_bio->state))
3035 			sync_request_write(mddev, r10_bio);
3036 		else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
3037 			recovery_request_write(mddev, r10_bio);
3038 		else if (test_bit(R10BIO_ReadError, &r10_bio->state))
3039 			handle_read_error(mddev, r10_bio);
3040 		else
3041 			WARN_ON_ONCE(1);
3042 
3043 		cond_resched();
3044 		if (mddev->sb_flags & ~(1<<MD_SB_CHANGE_PENDING))
3045 			md_check_recovery(mddev);
3046 	}
3047 	blk_finish_plug(&plug);
3048 }
3049 
init_resync(struct r10conf * conf)3050 static int init_resync(struct r10conf *conf)
3051 {
3052 	int ret, buffs, i;
3053 
3054 	buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
3055 	BUG_ON(mempool_initialized(&conf->r10buf_pool));
3056 	conf->have_replacement = 0;
3057 	for (i = 0; i < conf->geo.raid_disks; i++)
3058 		if (conf->mirrors[i].replacement)
3059 			conf->have_replacement = 1;
3060 	ret = mempool_init(&conf->r10buf_pool, buffs,
3061 			   r10buf_pool_alloc, r10buf_pool_free, conf);
3062 	if (ret)
3063 		return ret;
3064 	conf->next_resync = 0;
3065 	return 0;
3066 }
3067 
raid10_alloc_init_r10buf(struct r10conf * conf)3068 static struct r10bio *raid10_alloc_init_r10buf(struct r10conf *conf)
3069 {
3070 	struct r10bio *r10bio = mempool_alloc(&conf->r10buf_pool, GFP_NOIO);
3071 	struct rsync_pages *rp;
3072 	struct bio *bio;
3073 	int nalloc;
3074 	int i;
3075 
3076 	if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
3077 	    test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
3078 		nalloc = conf->copies; /* resync */
3079 	else
3080 		nalloc = 2; /* recovery */
3081 
3082 	for (i = 0; i < nalloc; i++) {
3083 		bio = r10bio->devs[i].bio;
3084 		rp = bio->bi_private;
3085 		bio_reset(bio, NULL, 0);
3086 		bio->bi_private = rp;
3087 		bio = r10bio->devs[i].repl_bio;
3088 		if (bio) {
3089 			rp = bio->bi_private;
3090 			bio_reset(bio, NULL, 0);
3091 			bio->bi_private = rp;
3092 		}
3093 	}
3094 	return r10bio;
3095 }
3096 
3097 /*
3098  * Set cluster_sync_high since we need other nodes to add the
3099  * range [cluster_sync_low, cluster_sync_high] to suspend list.
3100  */
raid10_set_cluster_sync_high(struct r10conf * conf)3101 static void raid10_set_cluster_sync_high(struct r10conf *conf)
3102 {
3103 	sector_t window_size;
3104 	int extra_chunk, chunks;
3105 
3106 	/*
3107 	 * First, here we define "stripe" as a unit which across
3108 	 * all member devices one time, so we get chunks by use
3109 	 * raid_disks / near_copies. Otherwise, if near_copies is
3110 	 * close to raid_disks, then resync window could increases
3111 	 * linearly with the increase of raid_disks, which means
3112 	 * we will suspend a really large IO window while it is not
3113 	 * necessary. If raid_disks is not divisible by near_copies,
3114 	 * an extra chunk is needed to ensure the whole "stripe" is
3115 	 * covered.
3116 	 */
3117 
3118 	chunks = conf->geo.raid_disks / conf->geo.near_copies;
3119 	if (conf->geo.raid_disks % conf->geo.near_copies == 0)
3120 		extra_chunk = 0;
3121 	else
3122 		extra_chunk = 1;
3123 	window_size = (chunks + extra_chunk) * conf->mddev->chunk_sectors;
3124 
3125 	/*
3126 	 * At least use a 32M window to align with raid1's resync window
3127 	 */
3128 	window_size = (CLUSTER_RESYNC_WINDOW_SECTORS > window_size) ?
3129 			CLUSTER_RESYNC_WINDOW_SECTORS : window_size;
3130 
3131 	conf->cluster_sync_high = conf->cluster_sync_low + window_size;
3132 }
3133 
3134 /*
3135  * perform a "sync" on one "block"
3136  *
3137  * We need to make sure that no normal I/O request - particularly write
3138  * requests - conflict with active sync requests.
3139  *
3140  * This is achieved by tracking pending requests and a 'barrier' concept
3141  * that can be installed to exclude normal IO requests.
3142  *
3143  * Resync and recovery are handled very differently.
3144  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
3145  *
3146  * For resync, we iterate over virtual addresses, read all copies,
3147  * and update if there are differences.  If only one copy is live,
3148  * skip it.
3149  * For recovery, we iterate over physical addresses, read a good
3150  * value for each non-in_sync drive, and over-write.
3151  *
3152  * So, for recovery we may have several outstanding complex requests for a
3153  * given address, one for each out-of-sync device.  We model this by allocating
3154  * a number of r10_bio structures, one for each out-of-sync device.
3155  * As we setup these structures, we collect all bio's together into a list
3156  * which we then process collectively to add pages, and then process again
3157  * to pass to submit_bio_noacct.
3158  *
3159  * The r10_bio structures are linked using a borrowed master_bio pointer.
3160  * This link is counted in ->remaining.  When the r10_bio that points to NULL
3161  * has its remaining count decremented to 0, the whole complex operation
3162  * is complete.
3163  *
3164  */
3165 
raid10_sync_request(struct mddev * mddev,sector_t sector_nr,sector_t max_sector,int * skipped)3166 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
3167 				    sector_t max_sector, int *skipped)
3168 {
3169 	struct r10conf *conf = mddev->private;
3170 	struct r10bio *r10_bio;
3171 	struct bio *biolist = NULL, *bio;
3172 	sector_t nr_sectors;
3173 	int i;
3174 	int max_sync;
3175 	sector_t sync_blocks;
3176 	sector_t sectors_skipped = 0;
3177 	int chunks_skipped = 0;
3178 	sector_t chunk_mask = conf->geo.chunk_mask;
3179 	int page_idx = 0;
3180 	int error_disk = -1;
3181 
3182 	/*
3183 	 * Allow skipping a full rebuild for incremental assembly
3184 	 * of a clean array, like RAID1 does.
3185 	 */
3186 	if (mddev->bitmap == NULL &&
3187 	    mddev->recovery_cp == MaxSector &&
3188 	    mddev->reshape_position == MaxSector &&
3189 	    !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
3190 	    !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
3191 	    !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
3192 	    conf->fullsync == 0) {
3193 		*skipped = 1;
3194 		return mddev->dev_sectors - sector_nr;
3195 	}
3196 
3197 	if (!mempool_initialized(&conf->r10buf_pool))
3198 		if (init_resync(conf))
3199 			return 0;
3200 
3201  skipped:
3202 	if (sector_nr >= max_sector) {
3203 		conf->cluster_sync_low = 0;
3204 		conf->cluster_sync_high = 0;
3205 
3206 		/* If we aborted, we need to abort the
3207 		 * sync on the 'current' bitmap chucks (there can
3208 		 * be several when recovering multiple devices).
3209 		 * as we may have started syncing it but not finished.
3210 		 * We can find the current address in
3211 		 * mddev->curr_resync, but for recovery,
3212 		 * we need to convert that to several
3213 		 * virtual addresses.
3214 		 */
3215 		if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
3216 			end_reshape(conf);
3217 			close_sync(conf);
3218 			return 0;
3219 		}
3220 
3221 		if (mddev->curr_resync < max_sector) { /* aborted */
3222 			if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
3223 				mddev->bitmap_ops->end_sync(mddev,
3224 							    mddev->curr_resync,
3225 							    &sync_blocks);
3226 			else for (i = 0; i < conf->geo.raid_disks; i++) {
3227 				sector_t sect =
3228 					raid10_find_virt(conf, mddev->curr_resync, i);
3229 
3230 				mddev->bitmap_ops->end_sync(mddev, sect,
3231 							    &sync_blocks);
3232 			}
3233 		} else {
3234 			/* completed sync */
3235 			if ((!mddev->bitmap || conf->fullsync)
3236 			    && conf->have_replacement
3237 			    && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3238 				/* Completed a full sync so the replacements
3239 				 * are now fully recovered.
3240 				 */
3241 				for (i = 0; i < conf->geo.raid_disks; i++) {
3242 					struct md_rdev *rdev =
3243 						conf->mirrors[i].replacement;
3244 
3245 					if (rdev)
3246 						rdev->recovery_offset = MaxSector;
3247 				}
3248 			}
3249 			conf->fullsync = 0;
3250 		}
3251 		mddev->bitmap_ops->close_sync(mddev);
3252 		close_sync(conf);
3253 		*skipped = 1;
3254 		return sectors_skipped;
3255 	}
3256 
3257 	if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
3258 		return reshape_request(mddev, sector_nr, skipped);
3259 
3260 	if (chunks_skipped >= conf->geo.raid_disks) {
3261 		pr_err("md/raid10:%s: %s fails\n", mdname(mddev),
3262 			test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ?  "resync" : "recovery");
3263 		if (error_disk >= 0 &&
3264 		    !test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3265 			/*
3266 			 * recovery fails, set mirrors.recovery_disabled,
3267 			 * device shouldn't be added to there.
3268 			 */
3269 			conf->mirrors[error_disk].recovery_disabled =
3270 						mddev->recovery_disabled;
3271 			return 0;
3272 		}
3273 		/*
3274 		 * if there has been nothing to do on any drive,
3275 		 * then there is nothing to do at all.
3276 		 */
3277 		*skipped = 1;
3278 		return (max_sector - sector_nr) + sectors_skipped;
3279 	}
3280 
3281 	if (max_sector > mddev->resync_max)
3282 		max_sector = mddev->resync_max; /* Don't do IO beyond here */
3283 
3284 	/* make sure whole request will fit in a chunk - if chunks
3285 	 * are meaningful
3286 	 */
3287 	if (conf->geo.near_copies < conf->geo.raid_disks &&
3288 	    max_sector > (sector_nr | chunk_mask))
3289 		max_sector = (sector_nr | chunk_mask) + 1;
3290 
3291 	/*
3292 	 * If there is non-resync activity waiting for a turn, then let it
3293 	 * though before starting on this new sync request.
3294 	 */
3295 	if (conf->nr_waiting)
3296 		schedule_timeout_uninterruptible(1);
3297 
3298 	/* Again, very different code for resync and recovery.
3299 	 * Both must result in an r10bio with a list of bios that
3300 	 * have bi_end_io, bi_sector, bi_bdev set,
3301 	 * and bi_private set to the r10bio.
3302 	 * For recovery, we may actually create several r10bios
3303 	 * with 2 bios in each, that correspond to the bios in the main one.
3304 	 * In this case, the subordinate r10bios link back through a
3305 	 * borrowed master_bio pointer, and the counter in the master
3306 	 * includes a ref from each subordinate.
3307 	 */
3308 	/* First, we decide what to do and set ->bi_end_io
3309 	 * To end_sync_read if we want to read, and
3310 	 * end_sync_write if we will want to write.
3311 	 */
3312 
3313 	max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3314 	if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3315 		/* recovery... the complicated one */
3316 		int j;
3317 		r10_bio = NULL;
3318 
3319 		for (i = 0 ; i < conf->geo.raid_disks; i++) {
3320 			bool still_degraded;
3321 			struct r10bio *rb2;
3322 			sector_t sect;
3323 			bool must_sync;
3324 			int any_working;
3325 			struct raid10_info *mirror = &conf->mirrors[i];
3326 			struct md_rdev *mrdev, *mreplace;
3327 
3328 			mrdev = mirror->rdev;
3329 			mreplace = mirror->replacement;
3330 
3331 			if (mrdev && (test_bit(Faulty, &mrdev->flags) ||
3332 			    test_bit(In_sync, &mrdev->flags)))
3333 				mrdev = NULL;
3334 			if (mreplace && test_bit(Faulty, &mreplace->flags))
3335 				mreplace = NULL;
3336 
3337 			if (!mrdev && !mreplace)
3338 				continue;
3339 
3340 			still_degraded = false;
3341 			/* want to reconstruct this device */
3342 			rb2 = r10_bio;
3343 			sect = raid10_find_virt(conf, sector_nr, i);
3344 			if (sect >= mddev->resync_max_sectors)
3345 				/* last stripe is not complete - don't
3346 				 * try to recover this sector.
3347 				 */
3348 				continue;
3349 			/* Unless we are doing a full sync, or a replacement
3350 			 * we only need to recover the block if it is set in
3351 			 * the bitmap
3352 			 */
3353 			must_sync = mddev->bitmap_ops->start_sync(mddev, sect,
3354 								  &sync_blocks,
3355 								  true);
3356 			if (sync_blocks < max_sync)
3357 				max_sync = sync_blocks;
3358 			if (!must_sync &&
3359 			    mreplace == NULL &&
3360 			    !conf->fullsync) {
3361 				/* yep, skip the sync_blocks here, but don't assume
3362 				 * that there will never be anything to do here
3363 				 */
3364 				chunks_skipped = -1;
3365 				continue;
3366 			}
3367 			if (mrdev)
3368 				atomic_inc(&mrdev->nr_pending);
3369 			if (mreplace)
3370 				atomic_inc(&mreplace->nr_pending);
3371 
3372 			r10_bio = raid10_alloc_init_r10buf(conf);
3373 			r10_bio->state = 0;
3374 			raise_barrier(conf, rb2 != NULL);
3375 			atomic_set(&r10_bio->remaining, 0);
3376 
3377 			r10_bio->master_bio = (struct bio*)rb2;
3378 			if (rb2)
3379 				atomic_inc(&rb2->remaining);
3380 			r10_bio->mddev = mddev;
3381 			set_bit(R10BIO_IsRecover, &r10_bio->state);
3382 			r10_bio->sector = sect;
3383 
3384 			raid10_find_phys(conf, r10_bio);
3385 
3386 			/* Need to check if the array will still be
3387 			 * degraded
3388 			 */
3389 			for (j = 0; j < conf->geo.raid_disks; j++) {
3390 				struct md_rdev *rdev = conf->mirrors[j].rdev;
3391 
3392 				if (rdev == NULL || test_bit(Faulty, &rdev->flags)) {
3393 					still_degraded = false;
3394 					break;
3395 				}
3396 			}
3397 
3398 			must_sync = mddev->bitmap_ops->start_sync(mddev, sect,
3399 						&sync_blocks, still_degraded);
3400 
3401 			any_working = 0;
3402 			for (j=0; j<conf->copies;j++) {
3403 				int k;
3404 				int d = r10_bio->devs[j].devnum;
3405 				sector_t from_addr, to_addr;
3406 				struct md_rdev *rdev = conf->mirrors[d].rdev;
3407 				sector_t sector, first_bad;
3408 				sector_t bad_sectors;
3409 				if (!rdev ||
3410 				    !test_bit(In_sync, &rdev->flags))
3411 					continue;
3412 				/* This is where we read from */
3413 				any_working = 1;
3414 				sector = r10_bio->devs[j].addr;
3415 
3416 				if (is_badblock(rdev, sector, max_sync,
3417 						&first_bad, &bad_sectors)) {
3418 					if (first_bad > sector)
3419 						max_sync = first_bad - sector;
3420 					else {
3421 						bad_sectors -= (sector
3422 								- first_bad);
3423 						if (max_sync > bad_sectors)
3424 							max_sync = bad_sectors;
3425 						continue;
3426 					}
3427 				}
3428 				bio = r10_bio->devs[0].bio;
3429 				bio->bi_next = biolist;
3430 				biolist = bio;
3431 				bio->bi_end_io = end_sync_read;
3432 				bio->bi_opf = REQ_OP_READ;
3433 				if (test_bit(FailFast, &rdev->flags))
3434 					bio->bi_opf |= MD_FAILFAST;
3435 				from_addr = r10_bio->devs[j].addr;
3436 				bio->bi_iter.bi_sector = from_addr +
3437 					rdev->data_offset;
3438 				bio_set_dev(bio, rdev->bdev);
3439 				atomic_inc(&rdev->nr_pending);
3440 				/* and we write to 'i' (if not in_sync) */
3441 
3442 				for (k=0; k<conf->copies; k++)
3443 					if (r10_bio->devs[k].devnum == i)
3444 						break;
3445 				BUG_ON(k == conf->copies);
3446 				to_addr = r10_bio->devs[k].addr;
3447 				r10_bio->devs[0].devnum = d;
3448 				r10_bio->devs[0].addr = from_addr;
3449 				r10_bio->devs[1].devnum = i;
3450 				r10_bio->devs[1].addr = to_addr;
3451 
3452 				if (mrdev) {
3453 					bio = r10_bio->devs[1].bio;
3454 					bio->bi_next = biolist;
3455 					biolist = bio;
3456 					bio->bi_end_io = end_sync_write;
3457 					bio->bi_opf = REQ_OP_WRITE;
3458 					bio->bi_iter.bi_sector = to_addr
3459 						+ mrdev->data_offset;
3460 					bio_set_dev(bio, mrdev->bdev);
3461 					atomic_inc(&r10_bio->remaining);
3462 				} else
3463 					r10_bio->devs[1].bio->bi_end_io = NULL;
3464 
3465 				/* and maybe write to replacement */
3466 				bio = r10_bio->devs[1].repl_bio;
3467 				if (bio)
3468 					bio->bi_end_io = NULL;
3469 				/* Note: if replace is not NULL, then bio
3470 				 * cannot be NULL as r10buf_pool_alloc will
3471 				 * have allocated it.
3472 				 */
3473 				if (!mreplace)
3474 					break;
3475 				bio->bi_next = biolist;
3476 				biolist = bio;
3477 				bio->bi_end_io = end_sync_write;
3478 				bio->bi_opf = REQ_OP_WRITE;
3479 				bio->bi_iter.bi_sector = to_addr +
3480 					mreplace->data_offset;
3481 				bio_set_dev(bio, mreplace->bdev);
3482 				atomic_inc(&r10_bio->remaining);
3483 				break;
3484 			}
3485 			if (j == conf->copies) {
3486 				/* Cannot recover, so abort the recovery or
3487 				 * record a bad block */
3488 				if (any_working) {
3489 					/* problem is that there are bad blocks
3490 					 * on other device(s)
3491 					 */
3492 					int k;
3493 					for (k = 0; k < conf->copies; k++)
3494 						if (r10_bio->devs[k].devnum == i)
3495 							break;
3496 					if (mrdev && !test_bit(In_sync,
3497 						      &mrdev->flags)
3498 					    && !rdev_set_badblocks(
3499 						    mrdev,
3500 						    r10_bio->devs[k].addr,
3501 						    max_sync, 0))
3502 						any_working = 0;
3503 					if (mreplace &&
3504 					    !rdev_set_badblocks(
3505 						    mreplace,
3506 						    r10_bio->devs[k].addr,
3507 						    max_sync, 0))
3508 						any_working = 0;
3509 				}
3510 				if (!any_working)  {
3511 					if (!test_and_set_bit(MD_RECOVERY_INTR,
3512 							      &mddev->recovery))
3513 						pr_warn("md/raid10:%s: insufficient working devices for recovery.\n",
3514 						       mdname(mddev));
3515 					mirror->recovery_disabled
3516 						= mddev->recovery_disabled;
3517 				} else {
3518 					error_disk = i;
3519 				}
3520 				put_buf(r10_bio);
3521 				if (rb2)
3522 					atomic_dec(&rb2->remaining);
3523 				r10_bio = rb2;
3524 				if (mrdev)
3525 					rdev_dec_pending(mrdev, mddev);
3526 				if (mreplace)
3527 					rdev_dec_pending(mreplace, mddev);
3528 				break;
3529 			}
3530 			if (mrdev)
3531 				rdev_dec_pending(mrdev, mddev);
3532 			if (mreplace)
3533 				rdev_dec_pending(mreplace, mddev);
3534 			if (r10_bio->devs[0].bio->bi_opf & MD_FAILFAST) {
3535 				/* Only want this if there is elsewhere to
3536 				 * read from. 'j' is currently the first
3537 				 * readable copy.
3538 				 */
3539 				int targets = 1;
3540 				for (; j < conf->copies; j++) {
3541 					int d = r10_bio->devs[j].devnum;
3542 					if (conf->mirrors[d].rdev &&
3543 					    test_bit(In_sync,
3544 						      &conf->mirrors[d].rdev->flags))
3545 						targets++;
3546 				}
3547 				if (targets == 1)
3548 					r10_bio->devs[0].bio->bi_opf
3549 						&= ~MD_FAILFAST;
3550 			}
3551 		}
3552 		if (biolist == NULL) {
3553 			while (r10_bio) {
3554 				struct r10bio *rb2 = r10_bio;
3555 				r10_bio = (struct r10bio*) rb2->master_bio;
3556 				rb2->master_bio = NULL;
3557 				put_buf(rb2);
3558 			}
3559 			goto giveup;
3560 		}
3561 	} else {
3562 		/* resync. Schedule a read for every block at this virt offset */
3563 		int count = 0;
3564 
3565 		/*
3566 		 * Since curr_resync_completed could probably not update in
3567 		 * time, and we will set cluster_sync_low based on it.
3568 		 * Let's check against "sector_nr + 2 * RESYNC_SECTORS" for
3569 		 * safety reason, which ensures curr_resync_completed is
3570 		 * updated in bitmap_cond_end_sync.
3571 		 */
3572 		mddev->bitmap_ops->cond_end_sync(mddev, sector_nr,
3573 					mddev_is_clustered(mddev) &&
3574 					(sector_nr + 2 * RESYNC_SECTORS > conf->cluster_sync_high));
3575 
3576 		if (!mddev->bitmap_ops->start_sync(mddev, sector_nr,
3577 						   &sync_blocks,
3578 						   mddev->degraded) &&
3579 		    !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3580 						 &mddev->recovery)) {
3581 			/* We can skip this block */
3582 			*skipped = 1;
3583 			return sync_blocks + sectors_skipped;
3584 		}
3585 		if (sync_blocks < max_sync)
3586 			max_sync = sync_blocks;
3587 		r10_bio = raid10_alloc_init_r10buf(conf);
3588 		r10_bio->state = 0;
3589 
3590 		r10_bio->mddev = mddev;
3591 		atomic_set(&r10_bio->remaining, 0);
3592 		raise_barrier(conf, 0);
3593 		conf->next_resync = sector_nr;
3594 
3595 		r10_bio->master_bio = NULL;
3596 		r10_bio->sector = sector_nr;
3597 		set_bit(R10BIO_IsSync, &r10_bio->state);
3598 		raid10_find_phys(conf, r10_bio);
3599 		r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3600 
3601 		for (i = 0; i < conf->copies; i++) {
3602 			int d = r10_bio->devs[i].devnum;
3603 			sector_t first_bad, sector;
3604 			sector_t bad_sectors;
3605 			struct md_rdev *rdev;
3606 
3607 			if (r10_bio->devs[i].repl_bio)
3608 				r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3609 
3610 			bio = r10_bio->devs[i].bio;
3611 			bio->bi_status = BLK_STS_IOERR;
3612 			rdev = conf->mirrors[d].rdev;
3613 			if (rdev == NULL || test_bit(Faulty, &rdev->flags))
3614 				continue;
3615 
3616 			sector = r10_bio->devs[i].addr;
3617 			if (is_badblock(rdev, sector, max_sync,
3618 					&first_bad, &bad_sectors)) {
3619 				if (first_bad > sector)
3620 					max_sync = first_bad - sector;
3621 				else {
3622 					bad_sectors -= (sector - first_bad);
3623 					if (max_sync > bad_sectors)
3624 						max_sync = bad_sectors;
3625 					continue;
3626 				}
3627 			}
3628 			atomic_inc(&rdev->nr_pending);
3629 			atomic_inc(&r10_bio->remaining);
3630 			bio->bi_next = biolist;
3631 			biolist = bio;
3632 			bio->bi_end_io = end_sync_read;
3633 			bio->bi_opf = REQ_OP_READ;
3634 			if (test_bit(FailFast, &rdev->flags))
3635 				bio->bi_opf |= MD_FAILFAST;
3636 			bio->bi_iter.bi_sector = sector + rdev->data_offset;
3637 			bio_set_dev(bio, rdev->bdev);
3638 			count++;
3639 
3640 			rdev = conf->mirrors[d].replacement;
3641 			if (rdev == NULL || test_bit(Faulty, &rdev->flags))
3642 				continue;
3643 
3644 			atomic_inc(&rdev->nr_pending);
3645 
3646 			/* Need to set up for writing to the replacement */
3647 			bio = r10_bio->devs[i].repl_bio;
3648 			bio->bi_status = BLK_STS_IOERR;
3649 
3650 			sector = r10_bio->devs[i].addr;
3651 			bio->bi_next = biolist;
3652 			biolist = bio;
3653 			bio->bi_end_io = end_sync_write;
3654 			bio->bi_opf = REQ_OP_WRITE;
3655 			if (test_bit(FailFast, &rdev->flags))
3656 				bio->bi_opf |= MD_FAILFAST;
3657 			bio->bi_iter.bi_sector = sector + rdev->data_offset;
3658 			bio_set_dev(bio, rdev->bdev);
3659 			count++;
3660 		}
3661 
3662 		if (count < 2) {
3663 			for (i=0; i<conf->copies; i++) {
3664 				int d = r10_bio->devs[i].devnum;
3665 				if (r10_bio->devs[i].bio->bi_end_io)
3666 					rdev_dec_pending(conf->mirrors[d].rdev,
3667 							 mddev);
3668 				if (r10_bio->devs[i].repl_bio &&
3669 				    r10_bio->devs[i].repl_bio->bi_end_io)
3670 					rdev_dec_pending(
3671 						conf->mirrors[d].replacement,
3672 						mddev);
3673 			}
3674 			put_buf(r10_bio);
3675 			biolist = NULL;
3676 			goto giveup;
3677 		}
3678 	}
3679 
3680 	nr_sectors = 0;
3681 	if (sector_nr + max_sync < max_sector)
3682 		max_sector = sector_nr + max_sync;
3683 	do {
3684 		struct page *page;
3685 		int len = PAGE_SIZE;
3686 		if (sector_nr + (len>>9) > max_sector)
3687 			len = (max_sector - sector_nr) << 9;
3688 		if (len == 0)
3689 			break;
3690 		for (bio= biolist ; bio ; bio=bio->bi_next) {
3691 			struct resync_pages *rp = get_resync_pages(bio);
3692 			page = resync_fetch_page(rp, page_idx);
3693 			if (WARN_ON(!bio_add_page(bio, page, len, 0))) {
3694 				bio->bi_status = BLK_STS_RESOURCE;
3695 				bio_endio(bio);
3696 				goto giveup;
3697 			}
3698 		}
3699 		nr_sectors += len>>9;
3700 		sector_nr += len>>9;
3701 	} while (++page_idx < RESYNC_PAGES);
3702 	r10_bio->sectors = nr_sectors;
3703 
3704 	if (mddev_is_clustered(mddev) &&
3705 	    test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3706 		/* It is resync not recovery */
3707 		if (conf->cluster_sync_high < sector_nr + nr_sectors) {
3708 			conf->cluster_sync_low = mddev->curr_resync_completed;
3709 			raid10_set_cluster_sync_high(conf);
3710 			/* Send resync message */
3711 			mddev->cluster_ops->resync_info_update(mddev,
3712 						conf->cluster_sync_low,
3713 						conf->cluster_sync_high);
3714 		}
3715 	} else if (mddev_is_clustered(mddev)) {
3716 		/* This is recovery not resync */
3717 		sector_t sect_va1, sect_va2;
3718 		bool broadcast_msg = false;
3719 
3720 		for (i = 0; i < conf->geo.raid_disks; i++) {
3721 			/*
3722 			 * sector_nr is a device address for recovery, so we
3723 			 * need translate it to array address before compare
3724 			 * with cluster_sync_high.
3725 			 */
3726 			sect_va1 = raid10_find_virt(conf, sector_nr, i);
3727 
3728 			if (conf->cluster_sync_high < sect_va1 + nr_sectors) {
3729 				broadcast_msg = true;
3730 				/*
3731 				 * curr_resync_completed is similar as
3732 				 * sector_nr, so make the translation too.
3733 				 */
3734 				sect_va2 = raid10_find_virt(conf,
3735 					mddev->curr_resync_completed, i);
3736 
3737 				if (conf->cluster_sync_low == 0 ||
3738 				    conf->cluster_sync_low > sect_va2)
3739 					conf->cluster_sync_low = sect_va2;
3740 			}
3741 		}
3742 		if (broadcast_msg) {
3743 			raid10_set_cluster_sync_high(conf);
3744 			mddev->cluster_ops->resync_info_update(mddev,
3745 						conf->cluster_sync_low,
3746 						conf->cluster_sync_high);
3747 		}
3748 	}
3749 
3750 	while (biolist) {
3751 		bio = biolist;
3752 		biolist = biolist->bi_next;
3753 
3754 		bio->bi_next = NULL;
3755 		r10_bio = get_resync_r10bio(bio);
3756 		r10_bio->sectors = nr_sectors;
3757 
3758 		if (bio->bi_end_io == end_sync_read) {
3759 			md_sync_acct_bio(bio, nr_sectors);
3760 			bio->bi_status = 0;
3761 			submit_bio_noacct(bio);
3762 		}
3763 	}
3764 
3765 	if (sectors_skipped)
3766 		/* pretend they weren't skipped, it makes
3767 		 * no important difference in this case
3768 		 */
3769 		md_done_sync(mddev, sectors_skipped, 1);
3770 
3771 	return sectors_skipped + nr_sectors;
3772  giveup:
3773 	/* There is nowhere to write, so all non-sync
3774 	 * drives must be failed or in resync, all drives
3775 	 * have a bad block, so try the next chunk...
3776 	 */
3777 	if (sector_nr + max_sync < max_sector)
3778 		max_sector = sector_nr + max_sync;
3779 
3780 	sectors_skipped += (max_sector - sector_nr);
3781 	chunks_skipped ++;
3782 	sector_nr = max_sector;
3783 	goto skipped;
3784 }
3785 
3786 static sector_t
raid10_size(struct mddev * mddev,sector_t sectors,int raid_disks)3787 raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3788 {
3789 	sector_t size;
3790 	struct r10conf *conf = mddev->private;
3791 
3792 	if (!raid_disks)
3793 		raid_disks = min(conf->geo.raid_disks,
3794 				 conf->prev.raid_disks);
3795 	if (!sectors)
3796 		sectors = conf->dev_sectors;
3797 
3798 	size = sectors >> conf->geo.chunk_shift;
3799 	sector_div(size, conf->geo.far_copies);
3800 	size = size * raid_disks;
3801 	sector_div(size, conf->geo.near_copies);
3802 
3803 	return size << conf->geo.chunk_shift;
3804 }
3805 
calc_sectors(struct r10conf * conf,sector_t size)3806 static void calc_sectors(struct r10conf *conf, sector_t size)
3807 {
3808 	/* Calculate the number of sectors-per-device that will
3809 	 * actually be used, and set conf->dev_sectors and
3810 	 * conf->stride
3811 	 */
3812 
3813 	size = size >> conf->geo.chunk_shift;
3814 	sector_div(size, conf->geo.far_copies);
3815 	size = size * conf->geo.raid_disks;
3816 	sector_div(size, conf->geo.near_copies);
3817 	/* 'size' is now the number of chunks in the array */
3818 	/* calculate "used chunks per device" */
3819 	size = size * conf->copies;
3820 
3821 	/* We need to round up when dividing by raid_disks to
3822 	 * get the stride size.
3823 	 */
3824 	size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3825 
3826 	conf->dev_sectors = size << conf->geo.chunk_shift;
3827 
3828 	if (conf->geo.far_offset)
3829 		conf->geo.stride = 1 << conf->geo.chunk_shift;
3830 	else {
3831 		sector_div(size, conf->geo.far_copies);
3832 		conf->geo.stride = size << conf->geo.chunk_shift;
3833 	}
3834 }
3835 
3836 enum geo_type {geo_new, geo_old, geo_start};
setup_geo(struct geom * geo,struct mddev * mddev,enum geo_type new)3837 static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3838 {
3839 	int nc, fc, fo;
3840 	int layout, chunk, disks;
3841 	switch (new) {
3842 	case geo_old:
3843 		layout = mddev->layout;
3844 		chunk = mddev->chunk_sectors;
3845 		disks = mddev->raid_disks - mddev->delta_disks;
3846 		break;
3847 	case geo_new:
3848 		layout = mddev->new_layout;
3849 		chunk = mddev->new_chunk_sectors;
3850 		disks = mddev->raid_disks;
3851 		break;
3852 	default: /* avoid 'may be unused' warnings */
3853 	case geo_start: /* new when starting reshape - raid_disks not
3854 			 * updated yet. */
3855 		layout = mddev->new_layout;
3856 		chunk = mddev->new_chunk_sectors;
3857 		disks = mddev->raid_disks + mddev->delta_disks;
3858 		break;
3859 	}
3860 	if (layout >> 19)
3861 		return -1;
3862 	if (chunk < (PAGE_SIZE >> 9) ||
3863 	    !is_power_of_2(chunk))
3864 		return -2;
3865 	nc = layout & 255;
3866 	fc = (layout >> 8) & 255;
3867 	fo = layout & (1<<16);
3868 	geo->raid_disks = disks;
3869 	geo->near_copies = nc;
3870 	geo->far_copies = fc;
3871 	geo->far_offset = fo;
3872 	switch (layout >> 17) {
3873 	case 0:	/* original layout.  simple but not always optimal */
3874 		geo->far_set_size = disks;
3875 		break;
3876 	case 1: /* "improved" layout which was buggy.  Hopefully no-one is
3877 		 * actually using this, but leave code here just in case.*/
3878 		geo->far_set_size = disks/fc;
3879 		WARN(geo->far_set_size < fc,
3880 		     "This RAID10 layout does not provide data safety - please backup and create new array\n");
3881 		break;
3882 	case 2: /* "improved" layout fixed to match documentation */
3883 		geo->far_set_size = fc * nc;
3884 		break;
3885 	default: /* Not a valid layout */
3886 		return -1;
3887 	}
3888 	geo->chunk_mask = chunk - 1;
3889 	geo->chunk_shift = ffz(~chunk);
3890 	return nc*fc;
3891 }
3892 
raid10_free_conf(struct r10conf * conf)3893 static void raid10_free_conf(struct r10conf *conf)
3894 {
3895 	if (!conf)
3896 		return;
3897 
3898 	mempool_exit(&conf->r10bio_pool);
3899 	kfree(conf->mirrors);
3900 	kfree(conf->mirrors_old);
3901 	kfree(conf->mirrors_new);
3902 	safe_put_page(conf->tmppage);
3903 	bioset_exit(&conf->bio_split);
3904 	kfree(conf);
3905 }
3906 
setup_conf(struct mddev * mddev)3907 static struct r10conf *setup_conf(struct mddev *mddev)
3908 {
3909 	struct r10conf *conf = NULL;
3910 	int err = -EINVAL;
3911 	struct geom geo;
3912 	int copies;
3913 
3914 	copies = setup_geo(&geo, mddev, geo_new);
3915 
3916 	if (copies == -2) {
3917 		pr_warn("md/raid10:%s: chunk size must be at least PAGE_SIZE(%ld) and be a power of 2.\n",
3918 			mdname(mddev), PAGE_SIZE);
3919 		goto out;
3920 	}
3921 
3922 	if (copies < 2 || copies > mddev->raid_disks) {
3923 		pr_warn("md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3924 			mdname(mddev), mddev->new_layout);
3925 		goto out;
3926 	}
3927 
3928 	err = -ENOMEM;
3929 	conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3930 	if (!conf)
3931 		goto out;
3932 
3933 	/* FIXME calc properly */
3934 	conf->mirrors = kcalloc(mddev->raid_disks + max(0, -mddev->delta_disks),
3935 				sizeof(struct raid10_info),
3936 				GFP_KERNEL);
3937 	if (!conf->mirrors)
3938 		goto out;
3939 
3940 	conf->tmppage = alloc_page(GFP_KERNEL);
3941 	if (!conf->tmppage)
3942 		goto out;
3943 
3944 	conf->geo = geo;
3945 	conf->copies = copies;
3946 	err = mempool_init(&conf->r10bio_pool, NR_RAID_BIOS, r10bio_pool_alloc,
3947 			   rbio_pool_free, conf);
3948 	if (err)
3949 		goto out;
3950 
3951 	err = bioset_init(&conf->bio_split, BIO_POOL_SIZE, 0, 0);
3952 	if (err)
3953 		goto out;
3954 
3955 	calc_sectors(conf, mddev->dev_sectors);
3956 	if (mddev->reshape_position == MaxSector) {
3957 		conf->prev = conf->geo;
3958 		conf->reshape_progress = MaxSector;
3959 	} else {
3960 		if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3961 			err = -EINVAL;
3962 			goto out;
3963 		}
3964 		conf->reshape_progress = mddev->reshape_position;
3965 		if (conf->prev.far_offset)
3966 			conf->prev.stride = 1 << conf->prev.chunk_shift;
3967 		else
3968 			/* far_copies must be 1 */
3969 			conf->prev.stride = conf->dev_sectors;
3970 	}
3971 	conf->reshape_safe = conf->reshape_progress;
3972 	spin_lock_init(&conf->device_lock);
3973 	INIT_LIST_HEAD(&conf->retry_list);
3974 	INIT_LIST_HEAD(&conf->bio_end_io_list);
3975 
3976 	seqlock_init(&conf->resync_lock);
3977 	init_waitqueue_head(&conf->wait_barrier);
3978 	atomic_set(&conf->nr_pending, 0);
3979 
3980 	err = -ENOMEM;
3981 	rcu_assign_pointer(conf->thread,
3982 			   md_register_thread(raid10d, mddev, "raid10"));
3983 	if (!conf->thread)
3984 		goto out;
3985 
3986 	conf->mddev = mddev;
3987 	return conf;
3988 
3989  out:
3990 	raid10_free_conf(conf);
3991 	return ERR_PTR(err);
3992 }
3993 
raid10_nr_stripes(struct r10conf * conf)3994 static unsigned int raid10_nr_stripes(struct r10conf *conf)
3995 {
3996 	unsigned int raid_disks = conf->geo.raid_disks;
3997 
3998 	if (conf->geo.raid_disks % conf->geo.near_copies)
3999 		return raid_disks;
4000 	return raid_disks / conf->geo.near_copies;
4001 }
4002 
raid10_set_queue_limits(struct mddev * mddev)4003 static int raid10_set_queue_limits(struct mddev *mddev)
4004 {
4005 	struct r10conf *conf = mddev->private;
4006 	struct queue_limits lim;
4007 	int err;
4008 
4009 	md_init_stacking_limits(&lim);
4010 	lim.max_write_zeroes_sectors = 0;
4011 	lim.io_min = mddev->chunk_sectors << 9;
4012 	lim.io_opt = lim.io_min * raid10_nr_stripes(conf);
4013 	lim.features |= BLK_FEAT_ATOMIC_WRITES;
4014 	err = mddev_stack_rdev_limits(mddev, &lim, MDDEV_STACK_INTEGRITY);
4015 	if (err)
4016 		return err;
4017 	return queue_limits_set(mddev->gendisk->queue, &lim);
4018 }
4019 
raid10_run(struct mddev * mddev)4020 static int raid10_run(struct mddev *mddev)
4021 {
4022 	struct r10conf *conf;
4023 	int i, disk_idx;
4024 	struct raid10_info *disk;
4025 	struct md_rdev *rdev;
4026 	sector_t size;
4027 	sector_t min_offset_diff = 0;
4028 	int first = 1;
4029 	int ret = -EIO;
4030 
4031 	if (mddev->private == NULL) {
4032 		conf = setup_conf(mddev);
4033 		if (IS_ERR(conf))
4034 			return PTR_ERR(conf);
4035 		mddev->private = conf;
4036 	}
4037 	conf = mddev->private;
4038 	if (!conf)
4039 		goto out;
4040 
4041 	rcu_assign_pointer(mddev->thread, conf->thread);
4042 	rcu_assign_pointer(conf->thread, NULL);
4043 
4044 	if (mddev_is_clustered(conf->mddev)) {
4045 		int fc, fo;
4046 
4047 		fc = (mddev->layout >> 8) & 255;
4048 		fo = mddev->layout & (1<<16);
4049 		if (fc > 1 || fo > 0) {
4050 			pr_err("only near layout is supported by clustered"
4051 				" raid10\n");
4052 			goto out_free_conf;
4053 		}
4054 	}
4055 
4056 	rdev_for_each(rdev, mddev) {
4057 		long long diff;
4058 
4059 		disk_idx = rdev->raid_disk;
4060 		if (disk_idx < 0)
4061 			continue;
4062 		if (disk_idx >= conf->geo.raid_disks &&
4063 		    disk_idx >= conf->prev.raid_disks)
4064 			continue;
4065 		disk = conf->mirrors + disk_idx;
4066 
4067 		if (test_bit(Replacement, &rdev->flags)) {
4068 			if (disk->replacement)
4069 				goto out_free_conf;
4070 			disk->replacement = rdev;
4071 		} else {
4072 			if (disk->rdev)
4073 				goto out_free_conf;
4074 			disk->rdev = rdev;
4075 		}
4076 		diff = (rdev->new_data_offset - rdev->data_offset);
4077 		if (!mddev->reshape_backwards)
4078 			diff = -diff;
4079 		if (diff < 0)
4080 			diff = 0;
4081 		if (first || diff < min_offset_diff)
4082 			min_offset_diff = diff;
4083 
4084 		disk->head_position = 0;
4085 		first = 0;
4086 	}
4087 
4088 	if (!mddev_is_dm(conf->mddev)) {
4089 		int err = raid10_set_queue_limits(mddev);
4090 
4091 		if (err) {
4092 			ret = err;
4093 			goto out_free_conf;
4094 		}
4095 	}
4096 
4097 	/* need to check that every block has at least one working mirror */
4098 	if (!enough(conf, -1)) {
4099 		pr_err("md/raid10:%s: not enough operational mirrors.\n",
4100 		       mdname(mddev));
4101 		goto out_free_conf;
4102 	}
4103 
4104 	if (conf->reshape_progress != MaxSector) {
4105 		/* must ensure that shape change is supported */
4106 		if (conf->geo.far_copies != 1 &&
4107 		    conf->geo.far_offset == 0)
4108 			goto out_free_conf;
4109 		if (conf->prev.far_copies != 1 &&
4110 		    conf->prev.far_offset == 0)
4111 			goto out_free_conf;
4112 	}
4113 
4114 	mddev->degraded = 0;
4115 	for (i = 0;
4116 	     i < conf->geo.raid_disks
4117 		     || i < conf->prev.raid_disks;
4118 	     i++) {
4119 
4120 		disk = conf->mirrors + i;
4121 
4122 		if (!disk->rdev && disk->replacement) {
4123 			/* The replacement is all we have - use it */
4124 			disk->rdev = disk->replacement;
4125 			disk->replacement = NULL;
4126 			clear_bit(Replacement, &disk->rdev->flags);
4127 		}
4128 
4129 		if (!disk->rdev ||
4130 		    !test_bit(In_sync, &disk->rdev->flags)) {
4131 			disk->head_position = 0;
4132 			mddev->degraded++;
4133 			if (disk->rdev &&
4134 			    disk->rdev->saved_raid_disk < 0)
4135 				conf->fullsync = 1;
4136 		}
4137 
4138 		if (disk->replacement &&
4139 		    !test_bit(In_sync, &disk->replacement->flags) &&
4140 		    disk->replacement->saved_raid_disk < 0) {
4141 			conf->fullsync = 1;
4142 		}
4143 
4144 		disk->recovery_disabled = mddev->recovery_disabled - 1;
4145 	}
4146 
4147 	if (mddev->recovery_cp != MaxSector)
4148 		pr_notice("md/raid10:%s: not clean -- starting background reconstruction\n",
4149 			  mdname(mddev));
4150 	pr_info("md/raid10:%s: active with %d out of %d devices\n",
4151 		mdname(mddev), conf->geo.raid_disks - mddev->degraded,
4152 		conf->geo.raid_disks);
4153 	/*
4154 	 * Ok, everything is just fine now
4155 	 */
4156 	mddev->dev_sectors = conf->dev_sectors;
4157 	size = raid10_size(mddev, 0, 0);
4158 	md_set_array_sectors(mddev, size);
4159 	mddev->resync_max_sectors = size;
4160 	set_bit(MD_FAILFAST_SUPPORTED, &mddev->flags);
4161 
4162 	if (md_integrity_register(mddev))
4163 		goto out_free_conf;
4164 
4165 	if (conf->reshape_progress != MaxSector) {
4166 		unsigned long before_length, after_length;
4167 
4168 		before_length = ((1 << conf->prev.chunk_shift) *
4169 				 conf->prev.far_copies);
4170 		after_length = ((1 << conf->geo.chunk_shift) *
4171 				conf->geo.far_copies);
4172 
4173 		if (max(before_length, after_length) > min_offset_diff) {
4174 			/* This cannot work */
4175 			pr_warn("md/raid10: offset difference not enough to continue reshape\n");
4176 			goto out_free_conf;
4177 		}
4178 		conf->offset_diff = min_offset_diff;
4179 
4180 		clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4181 		clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4182 		set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4183 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4184 	}
4185 
4186 	return 0;
4187 
4188 out_free_conf:
4189 	md_unregister_thread(mddev, &mddev->thread);
4190 	raid10_free_conf(conf);
4191 	mddev->private = NULL;
4192 out:
4193 	return ret;
4194 }
4195 
raid10_free(struct mddev * mddev,void * priv)4196 static void raid10_free(struct mddev *mddev, void *priv)
4197 {
4198 	raid10_free_conf(priv);
4199 }
4200 
raid10_quiesce(struct mddev * mddev,int quiesce)4201 static void raid10_quiesce(struct mddev *mddev, int quiesce)
4202 {
4203 	struct r10conf *conf = mddev->private;
4204 
4205 	if (quiesce)
4206 		raise_barrier(conf, 0);
4207 	else
4208 		lower_barrier(conf);
4209 }
4210 
raid10_resize(struct mddev * mddev,sector_t sectors)4211 static int raid10_resize(struct mddev *mddev, sector_t sectors)
4212 {
4213 	/* Resize of 'far' arrays is not supported.
4214 	 * For 'near' and 'offset' arrays we can set the
4215 	 * number of sectors used to be an appropriate multiple
4216 	 * of the chunk size.
4217 	 * For 'offset', this is far_copies*chunksize.
4218 	 * For 'near' the multiplier is the LCM of
4219 	 * near_copies and raid_disks.
4220 	 * So if far_copies > 1 && !far_offset, fail.
4221 	 * Else find LCM(raid_disks, near_copy)*far_copies and
4222 	 * multiply by chunk_size.  Then round to this number.
4223 	 * This is mostly done by raid10_size()
4224 	 */
4225 	struct r10conf *conf = mddev->private;
4226 	sector_t oldsize, size;
4227 	int ret;
4228 
4229 	if (mddev->reshape_position != MaxSector)
4230 		return -EBUSY;
4231 
4232 	if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
4233 		return -EINVAL;
4234 
4235 	oldsize = raid10_size(mddev, 0, 0);
4236 	size = raid10_size(mddev, sectors, 0);
4237 	if (mddev->external_size &&
4238 	    mddev->array_sectors > size)
4239 		return -EINVAL;
4240 
4241 	ret = mddev->bitmap_ops->resize(mddev, size, 0, false);
4242 	if (ret)
4243 		return ret;
4244 
4245 	md_set_array_sectors(mddev, size);
4246 	if (sectors > mddev->dev_sectors &&
4247 	    mddev->recovery_cp > oldsize) {
4248 		mddev->recovery_cp = oldsize;
4249 		set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4250 	}
4251 	calc_sectors(conf, sectors);
4252 	mddev->dev_sectors = conf->dev_sectors;
4253 	mddev->resync_max_sectors = size;
4254 	return 0;
4255 }
4256 
raid10_takeover_raid0(struct mddev * mddev,sector_t size,int devs)4257 static void *raid10_takeover_raid0(struct mddev *mddev, sector_t size, int devs)
4258 {
4259 	struct md_rdev *rdev;
4260 	struct r10conf *conf;
4261 
4262 	if (mddev->degraded > 0) {
4263 		pr_warn("md/raid10:%s: Error: degraded raid0!\n",
4264 			mdname(mddev));
4265 		return ERR_PTR(-EINVAL);
4266 	}
4267 	sector_div(size, devs);
4268 
4269 	/* Set new parameters */
4270 	mddev->new_level = 10;
4271 	/* new layout: far_copies = 1, near_copies = 2 */
4272 	mddev->new_layout = (1<<8) + 2;
4273 	mddev->new_chunk_sectors = mddev->chunk_sectors;
4274 	mddev->delta_disks = mddev->raid_disks;
4275 	mddev->raid_disks *= 2;
4276 	/* make sure it will be not marked as dirty */
4277 	mddev->recovery_cp = MaxSector;
4278 	mddev->dev_sectors = size;
4279 
4280 	conf = setup_conf(mddev);
4281 	if (!IS_ERR(conf)) {
4282 		rdev_for_each(rdev, mddev)
4283 			if (rdev->raid_disk >= 0) {
4284 				rdev->new_raid_disk = rdev->raid_disk * 2;
4285 				rdev->sectors = size;
4286 			}
4287 	}
4288 
4289 	return conf;
4290 }
4291 
raid10_takeover(struct mddev * mddev)4292 static void *raid10_takeover(struct mddev *mddev)
4293 {
4294 	struct r0conf *raid0_conf;
4295 
4296 	/* raid10 can take over:
4297 	 *  raid0 - providing it has only two drives
4298 	 */
4299 	if (mddev->level == 0) {
4300 		/* for raid0 takeover only one zone is supported */
4301 		raid0_conf = mddev->private;
4302 		if (raid0_conf->nr_strip_zones > 1) {
4303 			pr_warn("md/raid10:%s: cannot takeover raid 0 with more than one zone.\n",
4304 				mdname(mddev));
4305 			return ERR_PTR(-EINVAL);
4306 		}
4307 		return raid10_takeover_raid0(mddev,
4308 			raid0_conf->strip_zone->zone_end,
4309 			raid0_conf->strip_zone->nb_dev);
4310 	}
4311 	return ERR_PTR(-EINVAL);
4312 }
4313 
raid10_check_reshape(struct mddev * mddev)4314 static int raid10_check_reshape(struct mddev *mddev)
4315 {
4316 	/* Called when there is a request to change
4317 	 * - layout (to ->new_layout)
4318 	 * - chunk size (to ->new_chunk_sectors)
4319 	 * - raid_disks (by delta_disks)
4320 	 * or when trying to restart a reshape that was ongoing.
4321 	 *
4322 	 * We need to validate the request and possibly allocate
4323 	 * space if that might be an issue later.
4324 	 *
4325 	 * Currently we reject any reshape of a 'far' mode array,
4326 	 * allow chunk size to change if new is generally acceptable,
4327 	 * allow raid_disks to increase, and allow
4328 	 * a switch between 'near' mode and 'offset' mode.
4329 	 */
4330 	struct r10conf *conf = mddev->private;
4331 	struct geom geo;
4332 
4333 	if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
4334 		return -EINVAL;
4335 
4336 	if (setup_geo(&geo, mddev, geo_start) != conf->copies)
4337 		/* mustn't change number of copies */
4338 		return -EINVAL;
4339 	if (geo.far_copies > 1 && !geo.far_offset)
4340 		/* Cannot switch to 'far' mode */
4341 		return -EINVAL;
4342 
4343 	if (mddev->array_sectors & geo.chunk_mask)
4344 			/* not factor of array size */
4345 			return -EINVAL;
4346 
4347 	if (!enough(conf, -1))
4348 		return -EINVAL;
4349 
4350 	kfree(conf->mirrors_new);
4351 	conf->mirrors_new = NULL;
4352 	if (mddev->delta_disks > 0) {
4353 		/* allocate new 'mirrors' list */
4354 		conf->mirrors_new =
4355 			kcalloc(mddev->raid_disks + mddev->delta_disks,
4356 				sizeof(struct raid10_info),
4357 				GFP_KERNEL);
4358 		if (!conf->mirrors_new)
4359 			return -ENOMEM;
4360 	}
4361 	return 0;
4362 }
4363 
4364 /*
4365  * Need to check if array has failed when deciding whether to:
4366  *  - start an array
4367  *  - remove non-faulty devices
4368  *  - add a spare
4369  *  - allow a reshape
4370  * This determination is simple when no reshape is happening.
4371  * However if there is a reshape, we need to carefully check
4372  * both the before and after sections.
4373  * This is because some failed devices may only affect one
4374  * of the two sections, and some non-in_sync devices may
4375  * be insync in the section most affected by failed devices.
4376  */
calc_degraded(struct r10conf * conf)4377 static int calc_degraded(struct r10conf *conf)
4378 {
4379 	int degraded, degraded2;
4380 	int i;
4381 
4382 	degraded = 0;
4383 	/* 'prev' section first */
4384 	for (i = 0; i < conf->prev.raid_disks; i++) {
4385 		struct md_rdev *rdev = conf->mirrors[i].rdev;
4386 
4387 		if (!rdev || test_bit(Faulty, &rdev->flags))
4388 			degraded++;
4389 		else if (!test_bit(In_sync, &rdev->flags))
4390 			/* When we can reduce the number of devices in
4391 			 * an array, this might not contribute to
4392 			 * 'degraded'.  It does now.
4393 			 */
4394 			degraded++;
4395 	}
4396 	if (conf->geo.raid_disks == conf->prev.raid_disks)
4397 		return degraded;
4398 	degraded2 = 0;
4399 	for (i = 0; i < conf->geo.raid_disks; i++) {
4400 		struct md_rdev *rdev = conf->mirrors[i].rdev;
4401 
4402 		if (!rdev || test_bit(Faulty, &rdev->flags))
4403 			degraded2++;
4404 		else if (!test_bit(In_sync, &rdev->flags)) {
4405 			/* If reshape is increasing the number of devices,
4406 			 * this section has already been recovered, so
4407 			 * it doesn't contribute to degraded.
4408 			 * else it does.
4409 			 */
4410 			if (conf->geo.raid_disks <= conf->prev.raid_disks)
4411 				degraded2++;
4412 		}
4413 	}
4414 	if (degraded2 > degraded)
4415 		return degraded2;
4416 	return degraded;
4417 }
4418 
raid10_start_reshape(struct mddev * mddev)4419 static int raid10_start_reshape(struct mddev *mddev)
4420 {
4421 	/* A 'reshape' has been requested. This commits
4422 	 * the various 'new' fields and sets MD_RECOVER_RESHAPE
4423 	 * This also checks if there are enough spares and adds them
4424 	 * to the array.
4425 	 * We currently require enough spares to make the final
4426 	 * array non-degraded.  We also require that the difference
4427 	 * between old and new data_offset - on each device - is
4428 	 * enough that we never risk over-writing.
4429 	 */
4430 
4431 	unsigned long before_length, after_length;
4432 	sector_t min_offset_diff = 0;
4433 	int first = 1;
4434 	struct geom new;
4435 	struct r10conf *conf = mddev->private;
4436 	struct md_rdev *rdev;
4437 	int spares = 0;
4438 	int ret;
4439 
4440 	if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4441 		return -EBUSY;
4442 
4443 	if (setup_geo(&new, mddev, geo_start) != conf->copies)
4444 		return -EINVAL;
4445 
4446 	before_length = ((1 << conf->prev.chunk_shift) *
4447 			 conf->prev.far_copies);
4448 	after_length = ((1 << conf->geo.chunk_shift) *
4449 			conf->geo.far_copies);
4450 
4451 	rdev_for_each(rdev, mddev) {
4452 		if (!test_bit(In_sync, &rdev->flags)
4453 		    && !test_bit(Faulty, &rdev->flags))
4454 			spares++;
4455 		if (rdev->raid_disk >= 0) {
4456 			long long diff = (rdev->new_data_offset
4457 					  - rdev->data_offset);
4458 			if (!mddev->reshape_backwards)
4459 				diff = -diff;
4460 			if (diff < 0)
4461 				diff = 0;
4462 			if (first || diff < min_offset_diff)
4463 				min_offset_diff = diff;
4464 			first = 0;
4465 		}
4466 	}
4467 
4468 	if (max(before_length, after_length) > min_offset_diff)
4469 		return -EINVAL;
4470 
4471 	if (spares < mddev->delta_disks)
4472 		return -EINVAL;
4473 
4474 	conf->offset_diff = min_offset_diff;
4475 	spin_lock_irq(&conf->device_lock);
4476 	if (conf->mirrors_new) {
4477 		memcpy(conf->mirrors_new, conf->mirrors,
4478 		       sizeof(struct raid10_info)*conf->prev.raid_disks);
4479 		smp_mb();
4480 		kfree(conf->mirrors_old);
4481 		conf->mirrors_old = conf->mirrors;
4482 		conf->mirrors = conf->mirrors_new;
4483 		conf->mirrors_new = NULL;
4484 	}
4485 	setup_geo(&conf->geo, mddev, geo_start);
4486 	smp_mb();
4487 	if (mddev->reshape_backwards) {
4488 		sector_t size = raid10_size(mddev, 0, 0);
4489 		if (size < mddev->array_sectors) {
4490 			spin_unlock_irq(&conf->device_lock);
4491 			pr_warn("md/raid10:%s: array size must be reduce before number of disks\n",
4492 				mdname(mddev));
4493 			return -EINVAL;
4494 		}
4495 		mddev->resync_max_sectors = size;
4496 		conf->reshape_progress = size;
4497 	} else
4498 		conf->reshape_progress = 0;
4499 	conf->reshape_safe = conf->reshape_progress;
4500 	spin_unlock_irq(&conf->device_lock);
4501 
4502 	if (mddev->delta_disks && mddev->bitmap) {
4503 		struct mdp_superblock_1 *sb = NULL;
4504 		sector_t oldsize, newsize;
4505 
4506 		oldsize = raid10_size(mddev, 0, 0);
4507 		newsize = raid10_size(mddev, 0, conf->geo.raid_disks);
4508 
4509 		if (!mddev_is_clustered(mddev)) {
4510 			ret = mddev->bitmap_ops->resize(mddev, newsize, 0, false);
4511 			if (ret)
4512 				goto abort;
4513 			else
4514 				goto out;
4515 		}
4516 
4517 		rdev_for_each(rdev, mddev) {
4518 			if (rdev->raid_disk > -1 &&
4519 			    !test_bit(Faulty, &rdev->flags))
4520 				sb = page_address(rdev->sb_page);
4521 		}
4522 
4523 		/*
4524 		 * some node is already performing reshape, and no need to
4525 		 * call bitmap_ops->resize again since it should be called when
4526 		 * receiving BITMAP_RESIZE msg
4527 		 */
4528 		if ((sb && (le32_to_cpu(sb->feature_map) &
4529 			    MD_FEATURE_RESHAPE_ACTIVE)) || (oldsize == newsize))
4530 			goto out;
4531 
4532 		ret = mddev->bitmap_ops->resize(mddev, newsize, 0, false);
4533 		if (ret)
4534 			goto abort;
4535 
4536 		ret = mddev->cluster_ops->resize_bitmaps(mddev, newsize, oldsize);
4537 		if (ret) {
4538 			mddev->bitmap_ops->resize(mddev, oldsize, 0, false);
4539 			goto abort;
4540 		}
4541 	}
4542 out:
4543 	if (mddev->delta_disks > 0) {
4544 		rdev_for_each(rdev, mddev)
4545 			if (rdev->raid_disk < 0 &&
4546 			    !test_bit(Faulty, &rdev->flags)) {
4547 				if (raid10_add_disk(mddev, rdev) == 0) {
4548 					if (rdev->raid_disk >=
4549 					    conf->prev.raid_disks)
4550 						set_bit(In_sync, &rdev->flags);
4551 					else
4552 						rdev->recovery_offset = 0;
4553 
4554 					/* Failure here is OK */
4555 					sysfs_link_rdev(mddev, rdev);
4556 				}
4557 			} else if (rdev->raid_disk >= conf->prev.raid_disks
4558 				   && !test_bit(Faulty, &rdev->flags)) {
4559 				/* This is a spare that was manually added */
4560 				set_bit(In_sync, &rdev->flags);
4561 			}
4562 	}
4563 	/* When a reshape changes the number of devices,
4564 	 * ->degraded is measured against the larger of the
4565 	 * pre and  post numbers.
4566 	 */
4567 	spin_lock_irq(&conf->device_lock);
4568 	mddev->degraded = calc_degraded(conf);
4569 	spin_unlock_irq(&conf->device_lock);
4570 	mddev->raid_disks = conf->geo.raid_disks;
4571 	mddev->reshape_position = conf->reshape_progress;
4572 	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4573 
4574 	clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4575 	clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4576 	clear_bit(MD_RECOVERY_DONE, &mddev->recovery);
4577 	set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4578 	set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4579 	conf->reshape_checkpoint = jiffies;
4580 	md_new_event();
4581 	return 0;
4582 
4583 abort:
4584 	mddev->recovery = 0;
4585 	spin_lock_irq(&conf->device_lock);
4586 	conf->geo = conf->prev;
4587 	mddev->raid_disks = conf->geo.raid_disks;
4588 	rdev_for_each(rdev, mddev)
4589 		rdev->new_data_offset = rdev->data_offset;
4590 	smp_wmb();
4591 	conf->reshape_progress = MaxSector;
4592 	conf->reshape_safe = MaxSector;
4593 	mddev->reshape_position = MaxSector;
4594 	spin_unlock_irq(&conf->device_lock);
4595 	return ret;
4596 }
4597 
4598 /* Calculate the last device-address that could contain
4599  * any block from the chunk that includes the array-address 's'
4600  * and report the next address.
4601  * i.e. the address returned will be chunk-aligned and after
4602  * any data that is in the chunk containing 's'.
4603  */
last_dev_address(sector_t s,struct geom * geo)4604 static sector_t last_dev_address(sector_t s, struct geom *geo)
4605 {
4606 	s = (s | geo->chunk_mask) + 1;
4607 	s >>= geo->chunk_shift;
4608 	s *= geo->near_copies;
4609 	s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4610 	s *= geo->far_copies;
4611 	s <<= geo->chunk_shift;
4612 	return s;
4613 }
4614 
4615 /* Calculate the first device-address that could contain
4616  * any block from the chunk that includes the array-address 's'.
4617  * This too will be the start of a chunk
4618  */
first_dev_address(sector_t s,struct geom * geo)4619 static sector_t first_dev_address(sector_t s, struct geom *geo)
4620 {
4621 	s >>= geo->chunk_shift;
4622 	s *= geo->near_copies;
4623 	sector_div(s, geo->raid_disks);
4624 	s *= geo->far_copies;
4625 	s <<= geo->chunk_shift;
4626 	return s;
4627 }
4628 
reshape_request(struct mddev * mddev,sector_t sector_nr,int * skipped)4629 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4630 				int *skipped)
4631 {
4632 	/* We simply copy at most one chunk (smallest of old and new)
4633 	 * at a time, possibly less if that exceeds RESYNC_PAGES,
4634 	 * or we hit a bad block or something.
4635 	 * This might mean we pause for normal IO in the middle of
4636 	 * a chunk, but that is not a problem as mddev->reshape_position
4637 	 * can record any location.
4638 	 *
4639 	 * If we will want to write to a location that isn't
4640 	 * yet recorded as 'safe' (i.e. in metadata on disk) then
4641 	 * we need to flush all reshape requests and update the metadata.
4642 	 *
4643 	 * When reshaping forwards (e.g. to more devices), we interpret
4644 	 * 'safe' as the earliest block which might not have been copied
4645 	 * down yet.  We divide this by previous stripe size and multiply
4646 	 * by previous stripe length to get lowest device offset that we
4647 	 * cannot write to yet.
4648 	 * We interpret 'sector_nr' as an address that we want to write to.
4649 	 * From this we use last_device_address() to find where we might
4650 	 * write to, and first_device_address on the  'safe' position.
4651 	 * If this 'next' write position is after the 'safe' position,
4652 	 * we must update the metadata to increase the 'safe' position.
4653 	 *
4654 	 * When reshaping backwards, we round in the opposite direction
4655 	 * and perform the reverse test:  next write position must not be
4656 	 * less than current safe position.
4657 	 *
4658 	 * In all this the minimum difference in data offsets
4659 	 * (conf->offset_diff - always positive) allows a bit of slack,
4660 	 * so next can be after 'safe', but not by more than offset_diff
4661 	 *
4662 	 * We need to prepare all the bios here before we start any IO
4663 	 * to ensure the size we choose is acceptable to all devices.
4664 	 * The means one for each copy for write-out and an extra one for
4665 	 * read-in.
4666 	 * We store the read-in bio in ->master_bio and the others in
4667 	 * ->devs[x].bio and ->devs[x].repl_bio.
4668 	 */
4669 	struct r10conf *conf = mddev->private;
4670 	struct r10bio *r10_bio;
4671 	sector_t next, safe, last;
4672 	int max_sectors;
4673 	int nr_sectors;
4674 	int s;
4675 	struct md_rdev *rdev;
4676 	int need_flush = 0;
4677 	struct bio *blist;
4678 	struct bio *bio, *read_bio;
4679 	int sectors_done = 0;
4680 	struct page **pages;
4681 
4682 	if (sector_nr == 0) {
4683 		/* If restarting in the middle, skip the initial sectors */
4684 		if (mddev->reshape_backwards &&
4685 		    conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4686 			sector_nr = (raid10_size(mddev, 0, 0)
4687 				     - conf->reshape_progress);
4688 		} else if (!mddev->reshape_backwards &&
4689 			   conf->reshape_progress > 0)
4690 			sector_nr = conf->reshape_progress;
4691 		if (sector_nr) {
4692 			mddev->curr_resync_completed = sector_nr;
4693 			sysfs_notify_dirent_safe(mddev->sysfs_completed);
4694 			*skipped = 1;
4695 			return sector_nr;
4696 		}
4697 	}
4698 
4699 	/* We don't use sector_nr to track where we are up to
4700 	 * as that doesn't work well for ->reshape_backwards.
4701 	 * So just use ->reshape_progress.
4702 	 */
4703 	if (mddev->reshape_backwards) {
4704 		/* 'next' is the earliest device address that we might
4705 		 * write to for this chunk in the new layout
4706 		 */
4707 		next = first_dev_address(conf->reshape_progress - 1,
4708 					 &conf->geo);
4709 
4710 		/* 'safe' is the last device address that we might read from
4711 		 * in the old layout after a restart
4712 		 */
4713 		safe = last_dev_address(conf->reshape_safe - 1,
4714 					&conf->prev);
4715 
4716 		if (next + conf->offset_diff < safe)
4717 			need_flush = 1;
4718 
4719 		last = conf->reshape_progress - 1;
4720 		sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4721 					       & conf->prev.chunk_mask);
4722 		if (sector_nr + RESYNC_SECTORS < last)
4723 			sector_nr = last + 1 - RESYNC_SECTORS;
4724 	} else {
4725 		/* 'next' is after the last device address that we
4726 		 * might write to for this chunk in the new layout
4727 		 */
4728 		next = last_dev_address(conf->reshape_progress, &conf->geo);
4729 
4730 		/* 'safe' is the earliest device address that we might
4731 		 * read from in the old layout after a restart
4732 		 */
4733 		safe = first_dev_address(conf->reshape_safe, &conf->prev);
4734 
4735 		/* Need to update metadata if 'next' might be beyond 'safe'
4736 		 * as that would possibly corrupt data
4737 		 */
4738 		if (next > safe + conf->offset_diff)
4739 			need_flush = 1;
4740 
4741 		sector_nr = conf->reshape_progress;
4742 		last  = sector_nr | (conf->geo.chunk_mask
4743 				     & conf->prev.chunk_mask);
4744 
4745 		if (sector_nr + RESYNC_SECTORS <= last)
4746 			last = sector_nr + RESYNC_SECTORS - 1;
4747 	}
4748 
4749 	if (need_flush ||
4750 	    time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4751 		/* Need to update reshape_position in metadata */
4752 		wait_barrier(conf, false);
4753 		mddev->reshape_position = conf->reshape_progress;
4754 		if (mddev->reshape_backwards)
4755 			mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4756 				- conf->reshape_progress;
4757 		else
4758 			mddev->curr_resync_completed = conf->reshape_progress;
4759 		conf->reshape_checkpoint = jiffies;
4760 		set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
4761 		md_wakeup_thread(mddev->thread);
4762 		wait_event(mddev->sb_wait, mddev->sb_flags == 0 ||
4763 			   test_bit(MD_RECOVERY_INTR, &mddev->recovery));
4764 		if (test_bit(MD_RECOVERY_INTR, &mddev->recovery)) {
4765 			allow_barrier(conf);
4766 			return sectors_done;
4767 		}
4768 		conf->reshape_safe = mddev->reshape_position;
4769 		allow_barrier(conf);
4770 	}
4771 
4772 	raise_barrier(conf, 0);
4773 read_more:
4774 	/* Now schedule reads for blocks from sector_nr to last */
4775 	r10_bio = raid10_alloc_init_r10buf(conf);
4776 	r10_bio->state = 0;
4777 	raise_barrier(conf, 1);
4778 	atomic_set(&r10_bio->remaining, 0);
4779 	r10_bio->mddev = mddev;
4780 	r10_bio->sector = sector_nr;
4781 	set_bit(R10BIO_IsReshape, &r10_bio->state);
4782 	r10_bio->sectors = last - sector_nr + 1;
4783 	rdev = read_balance(conf, r10_bio, &max_sectors);
4784 	BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4785 
4786 	if (!rdev) {
4787 		/* Cannot read from here, so need to record bad blocks
4788 		 * on all the target devices.
4789 		 */
4790 		// FIXME
4791 		mempool_free(r10_bio, &conf->r10buf_pool);
4792 		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4793 		return sectors_done;
4794 	}
4795 
4796 	read_bio = bio_alloc_bioset(rdev->bdev, RESYNC_PAGES, REQ_OP_READ,
4797 				    GFP_KERNEL, &mddev->bio_set);
4798 	read_bio->bi_iter.bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4799 			       + rdev->data_offset);
4800 	read_bio->bi_private = r10_bio;
4801 	read_bio->bi_end_io = end_reshape_read;
4802 	r10_bio->master_bio = read_bio;
4803 	r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4804 
4805 	/*
4806 	 * Broadcast RESYNC message to other nodes, so all nodes would not
4807 	 * write to the region to avoid conflict.
4808 	*/
4809 	if (mddev_is_clustered(mddev) && conf->cluster_sync_high <= sector_nr) {
4810 		struct mdp_superblock_1 *sb = NULL;
4811 		int sb_reshape_pos = 0;
4812 
4813 		conf->cluster_sync_low = sector_nr;
4814 		conf->cluster_sync_high = sector_nr + CLUSTER_RESYNC_WINDOW_SECTORS;
4815 		sb = page_address(rdev->sb_page);
4816 		if (sb) {
4817 			sb_reshape_pos = le64_to_cpu(sb->reshape_position);
4818 			/*
4819 			 * Set cluster_sync_low again if next address for array
4820 			 * reshape is less than cluster_sync_low. Since we can't
4821 			 * update cluster_sync_low until it has finished reshape.
4822 			 */
4823 			if (sb_reshape_pos < conf->cluster_sync_low)
4824 				conf->cluster_sync_low = sb_reshape_pos;
4825 		}
4826 
4827 		mddev->cluster_ops->resync_info_update(mddev, conf->cluster_sync_low,
4828 							  conf->cluster_sync_high);
4829 	}
4830 
4831 	/* Now find the locations in the new layout */
4832 	__raid10_find_phys(&conf->geo, r10_bio);
4833 
4834 	blist = read_bio;
4835 	read_bio->bi_next = NULL;
4836 
4837 	for (s = 0; s < conf->copies*2; s++) {
4838 		struct bio *b;
4839 		int d = r10_bio->devs[s/2].devnum;
4840 		struct md_rdev *rdev2;
4841 		if (s&1) {
4842 			rdev2 = conf->mirrors[d].replacement;
4843 			b = r10_bio->devs[s/2].repl_bio;
4844 		} else {
4845 			rdev2 = conf->mirrors[d].rdev;
4846 			b = r10_bio->devs[s/2].bio;
4847 		}
4848 		if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4849 			continue;
4850 
4851 		bio_set_dev(b, rdev2->bdev);
4852 		b->bi_iter.bi_sector = r10_bio->devs[s/2].addr +
4853 			rdev2->new_data_offset;
4854 		b->bi_end_io = end_reshape_write;
4855 		b->bi_opf = REQ_OP_WRITE;
4856 		b->bi_next = blist;
4857 		blist = b;
4858 	}
4859 
4860 	/* Now add as many pages as possible to all of these bios. */
4861 
4862 	nr_sectors = 0;
4863 	pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4864 	for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4865 		struct page *page = pages[s / (PAGE_SIZE >> 9)];
4866 		int len = (max_sectors - s) << 9;
4867 		if (len > PAGE_SIZE)
4868 			len = PAGE_SIZE;
4869 		for (bio = blist; bio ; bio = bio->bi_next) {
4870 			if (WARN_ON(!bio_add_page(bio, page, len, 0))) {
4871 				bio->bi_status = BLK_STS_RESOURCE;
4872 				bio_endio(bio);
4873 				return sectors_done;
4874 			}
4875 		}
4876 		sector_nr += len >> 9;
4877 		nr_sectors += len >> 9;
4878 	}
4879 	r10_bio->sectors = nr_sectors;
4880 
4881 	/* Now submit the read */
4882 	md_sync_acct_bio(read_bio, r10_bio->sectors);
4883 	atomic_inc(&r10_bio->remaining);
4884 	read_bio->bi_next = NULL;
4885 	submit_bio_noacct(read_bio);
4886 	sectors_done += nr_sectors;
4887 	if (sector_nr <= last)
4888 		goto read_more;
4889 
4890 	lower_barrier(conf);
4891 
4892 	/* Now that we have done the whole section we can
4893 	 * update reshape_progress
4894 	 */
4895 	if (mddev->reshape_backwards)
4896 		conf->reshape_progress -= sectors_done;
4897 	else
4898 		conf->reshape_progress += sectors_done;
4899 
4900 	return sectors_done;
4901 }
4902 
4903 static void end_reshape_request(struct r10bio *r10_bio);
4904 static int handle_reshape_read_error(struct mddev *mddev,
4905 				     struct r10bio *r10_bio);
reshape_request_write(struct mddev * mddev,struct r10bio * r10_bio)4906 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4907 {
4908 	/* Reshape read completed.  Hopefully we have a block
4909 	 * to write out.
4910 	 * If we got a read error then we do sync 1-page reads from
4911 	 * elsewhere until we find the data - or give up.
4912 	 */
4913 	struct r10conf *conf = mddev->private;
4914 	int s;
4915 
4916 	if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4917 		if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4918 			/* Reshape has been aborted */
4919 			md_done_sync(mddev, r10_bio->sectors, 0);
4920 			return;
4921 		}
4922 
4923 	/* We definitely have the data in the pages, schedule the
4924 	 * writes.
4925 	 */
4926 	atomic_set(&r10_bio->remaining, 1);
4927 	for (s = 0; s < conf->copies*2; s++) {
4928 		struct bio *b;
4929 		int d = r10_bio->devs[s/2].devnum;
4930 		struct md_rdev *rdev;
4931 		if (s&1) {
4932 			rdev = conf->mirrors[d].replacement;
4933 			b = r10_bio->devs[s/2].repl_bio;
4934 		} else {
4935 			rdev = conf->mirrors[d].rdev;
4936 			b = r10_bio->devs[s/2].bio;
4937 		}
4938 		if (!rdev || test_bit(Faulty, &rdev->flags))
4939 			continue;
4940 
4941 		atomic_inc(&rdev->nr_pending);
4942 		md_sync_acct_bio(b, r10_bio->sectors);
4943 		atomic_inc(&r10_bio->remaining);
4944 		b->bi_next = NULL;
4945 		submit_bio_noacct(b);
4946 	}
4947 	end_reshape_request(r10_bio);
4948 }
4949 
end_reshape(struct r10conf * conf)4950 static void end_reshape(struct r10conf *conf)
4951 {
4952 	if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4953 		return;
4954 
4955 	spin_lock_irq(&conf->device_lock);
4956 	conf->prev = conf->geo;
4957 	md_finish_reshape(conf->mddev);
4958 	smp_wmb();
4959 	conf->reshape_progress = MaxSector;
4960 	conf->reshape_safe = MaxSector;
4961 	spin_unlock_irq(&conf->device_lock);
4962 
4963 	mddev_update_io_opt(conf->mddev, raid10_nr_stripes(conf));
4964 	conf->fullsync = 0;
4965 }
4966 
raid10_update_reshape_pos(struct mddev * mddev)4967 static void raid10_update_reshape_pos(struct mddev *mddev)
4968 {
4969 	struct r10conf *conf = mddev->private;
4970 	sector_t lo, hi;
4971 
4972 	mddev->cluster_ops->resync_info_get(mddev, &lo, &hi);
4973 	if (((mddev->reshape_position <= hi) && (mddev->reshape_position >= lo))
4974 	    || mddev->reshape_position == MaxSector)
4975 		conf->reshape_progress = mddev->reshape_position;
4976 	else
4977 		WARN_ON_ONCE(1);
4978 }
4979 
handle_reshape_read_error(struct mddev * mddev,struct r10bio * r10_bio)4980 static int handle_reshape_read_error(struct mddev *mddev,
4981 				     struct r10bio *r10_bio)
4982 {
4983 	/* Use sync reads to get the blocks from somewhere else */
4984 	int sectors = r10_bio->sectors;
4985 	struct r10conf *conf = mddev->private;
4986 	struct r10bio *r10b;
4987 	int slot = 0;
4988 	int idx = 0;
4989 	struct page **pages;
4990 
4991 	r10b = kmalloc(struct_size(r10b, devs, conf->copies), GFP_NOIO);
4992 	if (!r10b) {
4993 		set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4994 		return -ENOMEM;
4995 	}
4996 
4997 	/* reshape IOs share pages from .devs[0].bio */
4998 	pages = get_resync_pages(r10_bio->devs[0].bio)->pages;
4999 
5000 	r10b->sector = r10_bio->sector;
5001 	__raid10_find_phys(&conf->prev, r10b);
5002 
5003 	while (sectors) {
5004 		int s = sectors;
5005 		int success = 0;
5006 		int first_slot = slot;
5007 
5008 		if (s > (PAGE_SIZE >> 9))
5009 			s = PAGE_SIZE >> 9;
5010 
5011 		while (!success) {
5012 			int d = r10b->devs[slot].devnum;
5013 			struct md_rdev *rdev = conf->mirrors[d].rdev;
5014 			sector_t addr;
5015 			if (rdev == NULL ||
5016 			    test_bit(Faulty, &rdev->flags) ||
5017 			    !test_bit(In_sync, &rdev->flags))
5018 				goto failed;
5019 
5020 			addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
5021 			atomic_inc(&rdev->nr_pending);
5022 			success = sync_page_io(rdev,
5023 					       addr,
5024 					       s << 9,
5025 					       pages[idx],
5026 					       REQ_OP_READ, false);
5027 			rdev_dec_pending(rdev, mddev);
5028 			if (success)
5029 				break;
5030 		failed:
5031 			slot++;
5032 			if (slot >= conf->copies)
5033 				slot = 0;
5034 			if (slot == first_slot)
5035 				break;
5036 		}
5037 		if (!success) {
5038 			/* couldn't read this block, must give up */
5039 			set_bit(MD_RECOVERY_INTR,
5040 				&mddev->recovery);
5041 			kfree(r10b);
5042 			return -EIO;
5043 		}
5044 		sectors -= s;
5045 		idx++;
5046 	}
5047 	kfree(r10b);
5048 	return 0;
5049 }
5050 
end_reshape_write(struct bio * bio)5051 static void end_reshape_write(struct bio *bio)
5052 {
5053 	struct r10bio *r10_bio = get_resync_r10bio(bio);
5054 	struct mddev *mddev = r10_bio->mddev;
5055 	struct r10conf *conf = mddev->private;
5056 	int d;
5057 	int slot;
5058 	int repl;
5059 	struct md_rdev *rdev = NULL;
5060 
5061 	d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
5062 	rdev = repl ? conf->mirrors[d].replacement :
5063 		      conf->mirrors[d].rdev;
5064 
5065 	if (bio->bi_status) {
5066 		/* FIXME should record badblock */
5067 		md_error(mddev, rdev);
5068 	}
5069 
5070 	rdev_dec_pending(rdev, mddev);
5071 	end_reshape_request(r10_bio);
5072 }
5073 
end_reshape_request(struct r10bio * r10_bio)5074 static void end_reshape_request(struct r10bio *r10_bio)
5075 {
5076 	if (!atomic_dec_and_test(&r10_bio->remaining))
5077 		return;
5078 	md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
5079 	bio_put(r10_bio->master_bio);
5080 	put_buf(r10_bio);
5081 }
5082 
raid10_finish_reshape(struct mddev * mddev)5083 static void raid10_finish_reshape(struct mddev *mddev)
5084 {
5085 	struct r10conf *conf = mddev->private;
5086 
5087 	if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
5088 		return;
5089 
5090 	if (mddev->delta_disks > 0) {
5091 		if (mddev->recovery_cp > mddev->resync_max_sectors) {
5092 			mddev->recovery_cp = mddev->resync_max_sectors;
5093 			set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
5094 		}
5095 		mddev->resync_max_sectors = mddev->array_sectors;
5096 	} else {
5097 		int d;
5098 		for (d = conf->geo.raid_disks ;
5099 		     d < conf->geo.raid_disks - mddev->delta_disks;
5100 		     d++) {
5101 			struct md_rdev *rdev = conf->mirrors[d].rdev;
5102 			if (rdev)
5103 				clear_bit(In_sync, &rdev->flags);
5104 			rdev = conf->mirrors[d].replacement;
5105 			if (rdev)
5106 				clear_bit(In_sync, &rdev->flags);
5107 		}
5108 	}
5109 	mddev->layout = mddev->new_layout;
5110 	mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
5111 	mddev->reshape_position = MaxSector;
5112 	mddev->delta_disks = 0;
5113 	mddev->reshape_backwards = 0;
5114 }
5115 
5116 static struct md_personality raid10_personality =
5117 {
5118 	.head = {
5119 		.type	= MD_PERSONALITY,
5120 		.id	= ID_RAID10,
5121 		.name	= "raid10",
5122 		.owner	= THIS_MODULE,
5123 	},
5124 
5125 	.make_request	= raid10_make_request,
5126 	.run		= raid10_run,
5127 	.free		= raid10_free,
5128 	.status		= raid10_status,
5129 	.error_handler	= raid10_error,
5130 	.hot_add_disk	= raid10_add_disk,
5131 	.hot_remove_disk= raid10_remove_disk,
5132 	.spare_active	= raid10_spare_active,
5133 	.sync_request	= raid10_sync_request,
5134 	.quiesce	= raid10_quiesce,
5135 	.size		= raid10_size,
5136 	.resize		= raid10_resize,
5137 	.takeover	= raid10_takeover,
5138 	.check_reshape	= raid10_check_reshape,
5139 	.start_reshape	= raid10_start_reshape,
5140 	.finish_reshape	= raid10_finish_reshape,
5141 	.update_reshape_pos = raid10_update_reshape_pos,
5142 };
5143 
raid10_init(void)5144 static int __init raid10_init(void)
5145 {
5146 	return register_md_submodule(&raid10_personality.head);
5147 }
5148 
raid10_exit(void)5149 static void __exit raid10_exit(void)
5150 {
5151 	unregister_md_submodule(&raid10_personality.head);
5152 }
5153 
5154 module_init(raid10_init);
5155 module_exit(raid10_exit);
5156 MODULE_LICENSE("GPL");
5157 MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
5158 MODULE_ALIAS("md-personality-9"); /* RAID10 */
5159 MODULE_ALIAS("md-raid10");
5160 MODULE_ALIAS("md-level-10");
5161