raid1/raid10: slow down resync if there is non-resync activity pending
[sfrench/cifs-2.6.git] / drivers / md / raid10.c
1 /*
2  * raid10.c : Multiple Devices driver for Linux
3  *
4  * Copyright (C) 2000-2004 Neil Brown
5  *
6  * RAID-10 support for md.
7  *
8  * Base on code in raid1.c.  See raid1.c for further copyright information.
9  *
10  *
11  * This program is free software; you can redistribute it and/or modify
12  * it under the terms of the GNU General Public License as published by
13  * the Free Software Foundation; either version 2, or (at your option)
14  * any later version.
15  *
16  * You should have received a copy of the GNU General Public License
17  * (for example /usr/src/linux/COPYING); if not, write to the Free
18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
19  */
20
21 #include <linux/slab.h>
22 #include <linux/delay.h>
23 #include <linux/blkdev.h>
24 #include <linux/module.h>
25 #include <linux/seq_file.h>
26 #include <linux/ratelimit.h>
27 #include <linux/kthread.h>
28 #include "md.h"
29 #include "raid10.h"
30 #include "raid0.h"
31 #include "bitmap.h"
32
33 /*
34  * RAID10 provides a combination of RAID0 and RAID1 functionality.
35  * The layout of data is defined by
36  *    chunk_size
37  *    raid_disks
38  *    near_copies (stored in low byte of layout)
39  *    far_copies (stored in second byte of layout)
40  *    far_offset (stored in bit 16 of layout )
41  *    use_far_sets (stored in bit 17 of layout )
42  *    use_far_sets_bugfixed (stored in bit 18 of layout )
43  *
44  * The data to be stored is divided into chunks using chunksize.  Each device
45  * is divided into far_copies sections.   In each section, chunks are laid out
46  * in a style similar to raid0, but near_copies copies of each chunk is stored
47  * (each on a different drive).  The starting device for each section is offset
48  * near_copies from the starting device of the previous section.  Thus there
49  * are (near_copies * far_copies) of each chunk, and each is on a different
50  * drive.  near_copies and far_copies must be at least one, and their product
51  * is at most raid_disks.
52  *
53  * If far_offset is true, then the far_copies are handled a bit differently.
54  * The copies are still in different stripes, but instead of being very far
55  * apart on disk, there are adjacent stripes.
56  *
57  * The far and offset algorithms are handled slightly differently if
58  * 'use_far_sets' is true.  In this case, the array's devices are grouped into
59  * sets that are (near_copies * far_copies) in size.  The far copied stripes
60  * are still shifted by 'near_copies' devices, but this shifting stays confined
61  * to the set rather than the entire array.  This is done to improve the number
62  * of device combinations that can fail without causing the array to fail.
63  * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
64  * on a device):
65  *    A B C D    A B C D E
66  *      ...         ...
67  *    D A B C    E A B C D
68  * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
69  *    [A B] [C D]    [A B] [C D E]
70  *    |...| |...|    |...| | ... |
71  *    [B A] [D C]    [B A] [E C D]
72  */
73
74 /*
75  * Number of guaranteed r10bios in case of extreme VM load:
76  */
77 #define NR_RAID10_BIOS 256
78
79 /* when we get a read error on a read-only array, we redirect to another
80  * device without failing the first device, or trying to over-write to
81  * correct the read error.  To keep track of bad blocks on a per-bio
82  * level, we store IO_BLOCKED in the appropriate 'bios' pointer
83  */
84 #define IO_BLOCKED ((struct bio *)1)
85 /* When we successfully write to a known bad-block, we need to remove the
86  * bad-block marking which must be done from process context.  So we record
87  * the success by setting devs[n].bio to IO_MADE_GOOD
88  */
89 #define IO_MADE_GOOD ((struct bio *)2)
90
91 #define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
92
93 /* When there are this many requests queued to be written by
94  * the raid10 thread, we become 'congested' to provide back-pressure
95  * for writeback.
96  */
97 static int max_queued_requests = 1024;
98
99 static void allow_barrier(struct r10conf *conf);
100 static void lower_barrier(struct r10conf *conf);
101 static int _enough(struct r10conf *conf, int previous, int ignore);
102 static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
103                                 int *skipped);
104 static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
105 static void end_reshape_write(struct bio *bio);
106 static void end_reshape(struct r10conf *conf);
107
108 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
109 {
110         struct r10conf *conf = data;
111         int size = offsetof(struct r10bio, devs[conf->copies]);
112
113         /* allocate a r10bio with room for raid_disks entries in the
114          * bios array */
115         return kzalloc(size, gfp_flags);
116 }
117
118 static void r10bio_pool_free(void *r10_bio, void *data)
119 {
120         kfree(r10_bio);
121 }
122
123 /* Maximum size of each resync request */
124 #define RESYNC_BLOCK_SIZE (64*1024)
125 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
126 /* amount of memory to reserve for resync requests */
127 #define RESYNC_WINDOW (1024*1024)
128 /* maximum number of concurrent requests, memory permitting */
129 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
130
131 /*
132  * When performing a resync, we need to read and compare, so
133  * we need as many pages are there are copies.
134  * When performing a recovery, we need 2 bios, one for read,
135  * one for write (we recover only one drive per r10buf)
136  *
137  */
138 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
139 {
140         struct r10conf *conf = data;
141         struct page *page;
142         struct r10bio *r10_bio;
143         struct bio *bio;
144         int i, j;
145         int nalloc;
146
147         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
148         if (!r10_bio)
149                 return NULL;
150
151         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
152             test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
153                 nalloc = conf->copies; /* resync */
154         else
155                 nalloc = 2; /* recovery */
156
157         /*
158          * Allocate bios.
159          */
160         for (j = nalloc ; j-- ; ) {
161                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
162                 if (!bio)
163                         goto out_free_bio;
164                 r10_bio->devs[j].bio = bio;
165                 if (!conf->have_replacement)
166                         continue;
167                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
168                 if (!bio)
169                         goto out_free_bio;
170                 r10_bio->devs[j].repl_bio = bio;
171         }
172         /*
173          * Allocate RESYNC_PAGES data pages and attach them
174          * where needed.
175          */
176         for (j = 0 ; j < nalloc; j++) {
177                 struct bio *rbio = r10_bio->devs[j].repl_bio;
178                 bio = r10_bio->devs[j].bio;
179                 for (i = 0; i < RESYNC_PAGES; i++) {
180                         if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
181                                                &conf->mddev->recovery)) {
182                                 /* we can share bv_page's during recovery
183                                  * and reshape */
184                                 struct bio *rbio = r10_bio->devs[0].bio;
185                                 page = rbio->bi_io_vec[i].bv_page;
186                                 get_page(page);
187                         } else
188                                 page = alloc_page(gfp_flags);
189                         if (unlikely(!page))
190                                 goto out_free_pages;
191
192                         bio->bi_io_vec[i].bv_page = page;
193                         if (rbio)
194                                 rbio->bi_io_vec[i].bv_page = page;
195                 }
196         }
197
198         return r10_bio;
199
200 out_free_pages:
201         for ( ; i > 0 ; i--)
202                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
203         while (j--)
204                 for (i = 0; i < RESYNC_PAGES ; i++)
205                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
206         j = 0;
207 out_free_bio:
208         for ( ; j < nalloc; j++) {
209                 if (r10_bio->devs[j].bio)
210                         bio_put(r10_bio->devs[j].bio);
211                 if (r10_bio->devs[j].repl_bio)
212                         bio_put(r10_bio->devs[j].repl_bio);
213         }
214         r10bio_pool_free(r10_bio, conf);
215         return NULL;
216 }
217
218 static void r10buf_pool_free(void *__r10_bio, void *data)
219 {
220         int i;
221         struct r10conf *conf = data;
222         struct r10bio *r10bio = __r10_bio;
223         int j;
224
225         for (j=0; j < conf->copies; j++) {
226                 struct bio *bio = r10bio->devs[j].bio;
227                 if (bio) {
228                         for (i = 0; i < RESYNC_PAGES; i++) {
229                                 safe_put_page(bio->bi_io_vec[i].bv_page);
230                                 bio->bi_io_vec[i].bv_page = NULL;
231                         }
232                         bio_put(bio);
233                 }
234                 bio = r10bio->devs[j].repl_bio;
235                 if (bio)
236                         bio_put(bio);
237         }
238         r10bio_pool_free(r10bio, conf);
239 }
240
241 static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
242 {
243         int i;
244
245         for (i = 0; i < conf->copies; i++) {
246                 struct bio **bio = & r10_bio->devs[i].bio;
247                 if (!BIO_SPECIAL(*bio))
248                         bio_put(*bio);
249                 *bio = NULL;
250                 bio = &r10_bio->devs[i].repl_bio;
251                 if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
252                         bio_put(*bio);
253                 *bio = NULL;
254         }
255 }
256
257 static void free_r10bio(struct r10bio *r10_bio)
258 {
259         struct r10conf *conf = r10_bio->mddev->private;
260
261         put_all_bios(conf, r10_bio);
262         mempool_free(r10_bio, conf->r10bio_pool);
263 }
264
265 static void put_buf(struct r10bio *r10_bio)
266 {
267         struct r10conf *conf = r10_bio->mddev->private;
268
269         mempool_free(r10_bio, conf->r10buf_pool);
270
271         lower_barrier(conf);
272 }
273
274 static void reschedule_retry(struct r10bio *r10_bio)
275 {
276         unsigned long flags;
277         struct mddev *mddev = r10_bio->mddev;
278         struct r10conf *conf = mddev->private;
279
280         spin_lock_irqsave(&conf->device_lock, flags);
281         list_add(&r10_bio->retry_list, &conf->retry_list);
282         conf->nr_queued ++;
283         spin_unlock_irqrestore(&conf->device_lock, flags);
284
285         /* wake up frozen array... */
286         wake_up(&conf->wait_barrier);
287
288         md_wakeup_thread(mddev->thread);
289 }
290
291 /*
292  * raid_end_bio_io() is called when we have finished servicing a mirrored
293  * operation and are ready to return a success/failure code to the buffer
294  * cache layer.
295  */
296 static void raid_end_bio_io(struct r10bio *r10_bio)
297 {
298         struct bio *bio = r10_bio->master_bio;
299         int done;
300         struct r10conf *conf = r10_bio->mddev->private;
301
302         if (bio->bi_phys_segments) {
303                 unsigned long flags;
304                 spin_lock_irqsave(&conf->device_lock, flags);
305                 bio->bi_phys_segments--;
306                 done = (bio->bi_phys_segments == 0);
307                 spin_unlock_irqrestore(&conf->device_lock, flags);
308         } else
309                 done = 1;
310         if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
311                 bio->bi_error = -EIO;
312         if (done) {
313                 bio_endio(bio);
314                 /*
315                  * Wake up any possible resync thread that waits for the device
316                  * to go idle.
317                  */
318                 allow_barrier(conf);
319         }
320         free_r10bio(r10_bio);
321 }
322
323 /*
324  * Update disk head position estimator based on IRQ completion info.
325  */
326 static inline void update_head_pos(int slot, struct r10bio *r10_bio)
327 {
328         struct r10conf *conf = r10_bio->mddev->private;
329
330         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
331                 r10_bio->devs[slot].addr + (r10_bio->sectors);
332 }
333
334 /*
335  * Find the disk number which triggered given bio
336  */
337 static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
338                          struct bio *bio, int *slotp, int *replp)
339 {
340         int slot;
341         int repl = 0;
342
343         for (slot = 0; slot < conf->copies; slot++) {
344                 if (r10_bio->devs[slot].bio == bio)
345                         break;
346                 if (r10_bio->devs[slot].repl_bio == bio) {
347                         repl = 1;
348                         break;
349                 }
350         }
351
352         BUG_ON(slot == conf->copies);
353         update_head_pos(slot, r10_bio);
354
355         if (slotp)
356                 *slotp = slot;
357         if (replp)
358                 *replp = repl;
359         return r10_bio->devs[slot].devnum;
360 }
361
362 static void raid10_end_read_request(struct bio *bio)
363 {
364         int uptodate = !bio->bi_error;
365         struct r10bio *r10_bio = bio->bi_private;
366         int slot, dev;
367         struct md_rdev *rdev;
368         struct r10conf *conf = r10_bio->mddev->private;
369
370         slot = r10_bio->read_slot;
371         dev = r10_bio->devs[slot].devnum;
372         rdev = r10_bio->devs[slot].rdev;
373         /*
374          * this branch is our 'one mirror IO has finished' event handler:
375          */
376         update_head_pos(slot, r10_bio);
377
378         if (uptodate) {
379                 /*
380                  * Set R10BIO_Uptodate in our master bio, so that
381                  * we will return a good error code to the higher
382                  * levels even if IO on some other mirrored buffer fails.
383                  *
384                  * The 'master' represents the composite IO operation to
385                  * user-side. So if something waits for IO, then it will
386                  * wait for the 'master' bio.
387                  */
388                 set_bit(R10BIO_Uptodate, &r10_bio->state);
389         } else {
390                 /* If all other devices that store this block have
391                  * failed, we want to return the error upwards rather
392                  * than fail the last device.  Here we redefine
393                  * "uptodate" to mean "Don't want to retry"
394                  */
395                 if (!_enough(conf, test_bit(R10BIO_Previous, &r10_bio->state),
396                              rdev->raid_disk))
397                         uptodate = 1;
398         }
399         if (uptodate) {
400                 raid_end_bio_io(r10_bio);
401                 rdev_dec_pending(rdev, conf->mddev);
402         } else {
403                 /*
404                  * oops, read error - keep the refcount on the rdev
405                  */
406                 char b[BDEVNAME_SIZE];
407                 printk_ratelimited(KERN_ERR
408                                    "md/raid10:%s: %s: rescheduling sector %llu\n",
409                                    mdname(conf->mddev),
410                                    bdevname(rdev->bdev, b),
411                                    (unsigned long long)r10_bio->sector);
412                 set_bit(R10BIO_ReadError, &r10_bio->state);
413                 reschedule_retry(r10_bio);
414         }
415 }
416
417 static void close_write(struct r10bio *r10_bio)
418 {
419         /* clear the bitmap if all writes complete successfully */
420         bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
421                         r10_bio->sectors,
422                         !test_bit(R10BIO_Degraded, &r10_bio->state),
423                         0);
424         md_write_end(r10_bio->mddev);
425 }
426
427 static void one_write_done(struct r10bio *r10_bio)
428 {
429         if (atomic_dec_and_test(&r10_bio->remaining)) {
430                 if (test_bit(R10BIO_WriteError, &r10_bio->state))
431                         reschedule_retry(r10_bio);
432                 else {
433                         close_write(r10_bio);
434                         if (test_bit(R10BIO_MadeGood, &r10_bio->state))
435                                 reschedule_retry(r10_bio);
436                         else
437                                 raid_end_bio_io(r10_bio);
438                 }
439         }
440 }
441
442 static void raid10_end_write_request(struct bio *bio)
443 {
444         struct r10bio *r10_bio = bio->bi_private;
445         int dev;
446         int dec_rdev = 1;
447         struct r10conf *conf = r10_bio->mddev->private;
448         int slot, repl;
449         struct md_rdev *rdev = NULL;
450
451         dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
452
453         if (repl)
454                 rdev = conf->mirrors[dev].replacement;
455         if (!rdev) {
456                 smp_rmb();
457                 repl = 0;
458                 rdev = conf->mirrors[dev].rdev;
459         }
460         /*
461          * this branch is our 'one mirror IO has finished' event handler:
462          */
463         if (bio->bi_error) {
464                 if (repl)
465                         /* Never record new bad blocks to replacement,
466                          * just fail it.
467                          */
468                         md_error(rdev->mddev, rdev);
469                 else {
470                         set_bit(WriteErrorSeen, &rdev->flags);
471                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
472                                 set_bit(MD_RECOVERY_NEEDED,
473                                         &rdev->mddev->recovery);
474                         set_bit(R10BIO_WriteError, &r10_bio->state);
475                         dec_rdev = 0;
476                 }
477         } else {
478                 /*
479                  * Set R10BIO_Uptodate in our master bio, so that
480                  * we will return a good error code for to the higher
481                  * levels even if IO on some other mirrored buffer fails.
482                  *
483                  * The 'master' represents the composite IO operation to
484                  * user-side. So if something waits for IO, then it will
485                  * wait for the 'master' bio.
486                  */
487                 sector_t first_bad;
488                 int bad_sectors;
489
490                 /*
491                  * Do not set R10BIO_Uptodate if the current device is
492                  * rebuilding or Faulty. This is because we cannot use
493                  * such device for properly reading the data back (we could
494                  * potentially use it, if the current write would have felt
495                  * before rdev->recovery_offset, but for simplicity we don't
496                  * check this here.
497                  */
498                 if (test_bit(In_sync, &rdev->flags) &&
499                     !test_bit(Faulty, &rdev->flags))
500                         set_bit(R10BIO_Uptodate, &r10_bio->state);
501
502                 /* Maybe we can clear some bad blocks. */
503                 if (is_badblock(rdev,
504                                 r10_bio->devs[slot].addr,
505                                 r10_bio->sectors,
506                                 &first_bad, &bad_sectors)) {
507                         bio_put(bio);
508                         if (repl)
509                                 r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
510                         else
511                                 r10_bio->devs[slot].bio = IO_MADE_GOOD;
512                         dec_rdev = 0;
513                         set_bit(R10BIO_MadeGood, &r10_bio->state);
514                 }
515         }
516
517         /*
518          *
519          * Let's see if all mirrored write operations have finished
520          * already.
521          */
522         one_write_done(r10_bio);
523         if (dec_rdev)
524                 rdev_dec_pending(rdev, conf->mddev);
525 }
526
527 /*
528  * RAID10 layout manager
529  * As well as the chunksize and raid_disks count, there are two
530  * parameters: near_copies and far_copies.
531  * near_copies * far_copies must be <= raid_disks.
532  * Normally one of these will be 1.
533  * If both are 1, we get raid0.
534  * If near_copies == raid_disks, we get raid1.
535  *
536  * Chunks are laid out in raid0 style with near_copies copies of the
537  * first chunk, followed by near_copies copies of the next chunk and
538  * so on.
539  * If far_copies > 1, then after 1/far_copies of the array has been assigned
540  * as described above, we start again with a device offset of near_copies.
541  * So we effectively have another copy of the whole array further down all
542  * the drives, but with blocks on different drives.
543  * With this layout, and block is never stored twice on the one device.
544  *
545  * raid10_find_phys finds the sector offset of a given virtual sector
546  * on each device that it is on.
547  *
548  * raid10_find_virt does the reverse mapping, from a device and a
549  * sector offset to a virtual address
550  */
551
552 static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
553 {
554         int n,f;
555         sector_t sector;
556         sector_t chunk;
557         sector_t stripe;
558         int dev;
559         int slot = 0;
560         int last_far_set_start, last_far_set_size;
561
562         last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
563         last_far_set_start *= geo->far_set_size;
564
565         last_far_set_size = geo->far_set_size;
566         last_far_set_size += (geo->raid_disks % geo->far_set_size);
567
568         /* now calculate first sector/dev */
569         chunk = r10bio->sector >> geo->chunk_shift;
570         sector = r10bio->sector & geo->chunk_mask;
571
572         chunk *= geo->near_copies;
573         stripe = chunk;
574         dev = sector_div(stripe, geo->raid_disks);
575         if (geo->far_offset)
576                 stripe *= geo->far_copies;
577
578         sector += stripe << geo->chunk_shift;
579
580         /* and calculate all the others */
581         for (n = 0; n < geo->near_copies; n++) {
582                 int d = dev;
583                 int set;
584                 sector_t s = sector;
585                 r10bio->devs[slot].devnum = d;
586                 r10bio->devs[slot].addr = s;
587                 slot++;
588
589                 for (f = 1; f < geo->far_copies; f++) {
590                         set = d / geo->far_set_size;
591                         d += geo->near_copies;
592
593                         if ((geo->raid_disks % geo->far_set_size) &&
594                             (d > last_far_set_start)) {
595                                 d -= last_far_set_start;
596                                 d %= last_far_set_size;
597                                 d += last_far_set_start;
598                         } else {
599                                 d %= geo->far_set_size;
600                                 d += geo->far_set_size * set;
601                         }
602                         s += geo->stride;
603                         r10bio->devs[slot].devnum = d;
604                         r10bio->devs[slot].addr = s;
605                         slot++;
606                 }
607                 dev++;
608                 if (dev >= geo->raid_disks) {
609                         dev = 0;
610                         sector += (geo->chunk_mask + 1);
611                 }
612         }
613 }
614
615 static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
616 {
617         struct geom *geo = &conf->geo;
618
619         if (conf->reshape_progress != MaxSector &&
620             ((r10bio->sector >= conf->reshape_progress) !=
621              conf->mddev->reshape_backwards)) {
622                 set_bit(R10BIO_Previous, &r10bio->state);
623                 geo = &conf->prev;
624         } else
625                 clear_bit(R10BIO_Previous, &r10bio->state);
626
627         __raid10_find_phys(geo, r10bio);
628 }
629
630 static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
631 {
632         sector_t offset, chunk, vchunk;
633         /* Never use conf->prev as this is only called during resync
634          * or recovery, so reshape isn't happening
635          */
636         struct geom *geo = &conf->geo;
637         int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
638         int far_set_size = geo->far_set_size;
639         int last_far_set_start;
640
641         if (geo->raid_disks % geo->far_set_size) {
642                 last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
643                 last_far_set_start *= geo->far_set_size;
644
645                 if (dev >= last_far_set_start) {
646                         far_set_size = geo->far_set_size;
647                         far_set_size += (geo->raid_disks % geo->far_set_size);
648                         far_set_start = last_far_set_start;
649                 }
650         }
651
652         offset = sector & geo->chunk_mask;
653         if (geo->far_offset) {
654                 int fc;
655                 chunk = sector >> geo->chunk_shift;
656                 fc = sector_div(chunk, geo->far_copies);
657                 dev -= fc * geo->near_copies;
658                 if (dev < far_set_start)
659                         dev += far_set_size;
660         } else {
661                 while (sector >= geo->stride) {
662                         sector -= geo->stride;
663                         if (dev < (geo->near_copies + far_set_start))
664                                 dev += far_set_size - geo->near_copies;
665                         else
666                                 dev -= geo->near_copies;
667                 }
668                 chunk = sector >> geo->chunk_shift;
669         }
670         vchunk = chunk * geo->raid_disks + dev;
671         sector_div(vchunk, geo->near_copies);
672         return (vchunk << geo->chunk_shift) + offset;
673 }
674
675 /*
676  * This routine returns the disk from which the requested read should
677  * be done. There is a per-array 'next expected sequential IO' sector
678  * number - if this matches on the next IO then we use the last disk.
679  * There is also a per-disk 'last know head position' sector that is
680  * maintained from IRQ contexts, both the normal and the resync IO
681  * completion handlers update this position correctly. If there is no
682  * perfect sequential match then we pick the disk whose head is closest.
683  *
684  * If there are 2 mirrors in the same 2 devices, performance degrades
685  * because position is mirror, not device based.
686  *
687  * The rdev for the device selected will have nr_pending incremented.
688  */
689
690 /*
691  * FIXME: possibly should rethink readbalancing and do it differently
692  * depending on near_copies / far_copies geometry.
693  */
694 static struct md_rdev *read_balance(struct r10conf *conf,
695                                     struct r10bio *r10_bio,
696                                     int *max_sectors)
697 {
698         const sector_t this_sector = r10_bio->sector;
699         int disk, slot;
700         int sectors = r10_bio->sectors;
701         int best_good_sectors;
702         sector_t new_distance, best_dist;
703         struct md_rdev *best_rdev, *rdev = NULL;
704         int do_balance;
705         int best_slot;
706         struct geom *geo = &conf->geo;
707
708         raid10_find_phys(conf, r10_bio);
709         rcu_read_lock();
710 retry:
711         sectors = r10_bio->sectors;
712         best_slot = -1;
713         best_rdev = NULL;
714         best_dist = MaxSector;
715         best_good_sectors = 0;
716         do_balance = 1;
717         /*
718          * Check if we can balance. We can balance on the whole
719          * device if no resync is going on (recovery is ok), or below
720          * the resync window. We take the first readable disk when
721          * above the resync window.
722          */
723         if (conf->mddev->recovery_cp < MaxSector
724             && (this_sector + sectors >= conf->next_resync))
725                 do_balance = 0;
726
727         for (slot = 0; slot < conf->copies ; slot++) {
728                 sector_t first_bad;
729                 int bad_sectors;
730                 sector_t dev_sector;
731
732                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
733                         continue;
734                 disk = r10_bio->devs[slot].devnum;
735                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
736                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
737                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
738                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
739                 if (rdev == NULL ||
740                     test_bit(Faulty, &rdev->flags))
741                         continue;
742                 if (!test_bit(In_sync, &rdev->flags) &&
743                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
744                         continue;
745
746                 dev_sector = r10_bio->devs[slot].addr;
747                 if (is_badblock(rdev, dev_sector, sectors,
748                                 &first_bad, &bad_sectors)) {
749                         if (best_dist < MaxSector)
750                                 /* Already have a better slot */
751                                 continue;
752                         if (first_bad <= dev_sector) {
753                                 /* Cannot read here.  If this is the
754                                  * 'primary' device, then we must not read
755                                  * beyond 'bad_sectors' from another device.
756                                  */
757                                 bad_sectors -= (dev_sector - first_bad);
758                                 if (!do_balance && sectors > bad_sectors)
759                                         sectors = bad_sectors;
760                                 if (best_good_sectors > sectors)
761                                         best_good_sectors = sectors;
762                         } else {
763                                 sector_t good_sectors =
764                                         first_bad - dev_sector;
765                                 if (good_sectors > best_good_sectors) {
766                                         best_good_sectors = good_sectors;
767                                         best_slot = slot;
768                                         best_rdev = rdev;
769                                 }
770                                 if (!do_balance)
771                                         /* Must read from here */
772                                         break;
773                         }
774                         continue;
775                 } else
776                         best_good_sectors = sectors;
777
778                 if (!do_balance)
779                         break;
780
781                 /* This optimisation is debatable, and completely destroys
782                  * sequential read speed for 'far copies' arrays.  So only
783                  * keep it for 'near' arrays, and review those later.
784                  */
785                 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
786                         break;
787
788                 /* for far > 1 always use the lowest address */
789                 if (geo->far_copies > 1)
790                         new_distance = r10_bio->devs[slot].addr;
791                 else
792                         new_distance = abs(r10_bio->devs[slot].addr -
793                                            conf->mirrors[disk].head_position);
794                 if (new_distance < best_dist) {
795                         best_dist = new_distance;
796                         best_slot = slot;
797                         best_rdev = rdev;
798                 }
799         }
800         if (slot >= conf->copies) {
801                 slot = best_slot;
802                 rdev = best_rdev;
803         }
804
805         if (slot >= 0) {
806                 atomic_inc(&rdev->nr_pending);
807                 if (test_bit(Faulty, &rdev->flags)) {
808                         /* Cannot risk returning a device that failed
809                          * before we inc'ed nr_pending
810                          */
811                         rdev_dec_pending(rdev, conf->mddev);
812                         goto retry;
813                 }
814                 r10_bio->read_slot = slot;
815         } else
816                 rdev = NULL;
817         rcu_read_unlock();
818         *max_sectors = best_good_sectors;
819
820         return rdev;
821 }
822
823 static int raid10_congested(struct mddev *mddev, int bits)
824 {
825         struct r10conf *conf = mddev->private;
826         int i, ret = 0;
827
828         if ((bits & (1 << WB_async_congested)) &&
829             conf->pending_count >= max_queued_requests)
830                 return 1;
831
832         rcu_read_lock();
833         for (i = 0;
834              (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
835                      && ret == 0;
836              i++) {
837                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
838                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
839                         struct request_queue *q = bdev_get_queue(rdev->bdev);
840
841                         ret |= bdi_congested(&q->backing_dev_info, bits);
842                 }
843         }
844         rcu_read_unlock();
845         return ret;
846 }
847
848 static void flush_pending_writes(struct r10conf *conf)
849 {
850         /* Any writes that have been queued but are awaiting
851          * bitmap updates get flushed here.
852          */
853         spin_lock_irq(&conf->device_lock);
854
855         if (conf->pending_bio_list.head) {
856                 struct bio *bio;
857                 bio = bio_list_get(&conf->pending_bio_list);
858                 conf->pending_count = 0;
859                 spin_unlock_irq(&conf->device_lock);
860                 /* flush any pending bitmap writes to disk
861                  * before proceeding w/ I/O */
862                 bitmap_unplug(conf->mddev->bitmap);
863                 wake_up(&conf->wait_barrier);
864
865                 while (bio) { /* submit pending writes */
866                         struct bio *next = bio->bi_next;
867                         bio->bi_next = NULL;
868                         if (unlikely((bio->bi_rw & REQ_DISCARD) &&
869                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
870                                 /* Just ignore it */
871                                 bio_endio(bio);
872                         else
873                                 generic_make_request(bio);
874                         bio = next;
875                 }
876         } else
877                 spin_unlock_irq(&conf->device_lock);
878 }
879
880 /* Barriers....
881  * Sometimes we need to suspend IO while we do something else,
882  * either some resync/recovery, or reconfigure the array.
883  * To do this we raise a 'barrier'.
884  * The 'barrier' is a counter that can be raised multiple times
885  * to count how many activities are happening which preclude
886  * normal IO.
887  * We can only raise the barrier if there is no pending IO.
888  * i.e. if nr_pending == 0.
889  * We choose only to raise the barrier if no-one is waiting for the
890  * barrier to go down.  This means that as soon as an IO request
891  * is ready, no other operations which require a barrier will start
892  * until the IO request has had a chance.
893  *
894  * So: regular IO calls 'wait_barrier'.  When that returns there
895  *    is no backgroup IO happening,  It must arrange to call
896  *    allow_barrier when it has finished its IO.
897  * backgroup IO calls must call raise_barrier.  Once that returns
898  *    there is no normal IO happeing.  It must arrange to call
899  *    lower_barrier when the particular background IO completes.
900  */
901
902 static void raise_barrier(struct r10conf *conf, int force)
903 {
904         BUG_ON(force && !conf->barrier);
905         spin_lock_irq(&conf->resync_lock);
906
907         /* Wait until no block IO is waiting (unless 'force') */
908         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
909                             conf->resync_lock);
910
911         /* block any new IO from starting */
912         conf->barrier++;
913
914         /* Now wait for all pending IO to complete */
915         wait_event_lock_irq(conf->wait_barrier,
916                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
917                             conf->resync_lock);
918
919         spin_unlock_irq(&conf->resync_lock);
920 }
921
922 static void lower_barrier(struct r10conf *conf)
923 {
924         unsigned long flags;
925         spin_lock_irqsave(&conf->resync_lock, flags);
926         conf->barrier--;
927         spin_unlock_irqrestore(&conf->resync_lock, flags);
928         wake_up(&conf->wait_barrier);
929 }
930
931 static void wait_barrier(struct r10conf *conf)
932 {
933         spin_lock_irq(&conf->resync_lock);
934         if (conf->barrier) {
935                 conf->nr_waiting++;
936                 /* Wait for the barrier to drop.
937                  * However if there are already pending
938                  * requests (preventing the barrier from
939                  * rising completely), and the
940                  * pre-process bio queue isn't empty,
941                  * then don't wait, as we need to empty
942                  * that queue to get the nr_pending
943                  * count down.
944                  */
945                 wait_event_lock_irq(conf->wait_barrier,
946                                     !conf->barrier ||
947                                     (conf->nr_pending &&
948                                      current->bio_list &&
949                                      !bio_list_empty(current->bio_list)),
950                                     conf->resync_lock);
951                 conf->nr_waiting--;
952         }
953         conf->nr_pending++;
954         spin_unlock_irq(&conf->resync_lock);
955 }
956
957 static void allow_barrier(struct r10conf *conf)
958 {
959         unsigned long flags;
960         spin_lock_irqsave(&conf->resync_lock, flags);
961         conf->nr_pending--;
962         spin_unlock_irqrestore(&conf->resync_lock, flags);
963         wake_up(&conf->wait_barrier);
964 }
965
966 static void freeze_array(struct r10conf *conf, int extra)
967 {
968         /* stop syncio and normal IO and wait for everything to
969          * go quiet.
970          * We increment barrier and nr_waiting, and then
971          * wait until nr_pending match nr_queued+extra
972          * This is called in the context of one normal IO request
973          * that has failed. Thus any sync request that might be pending
974          * will be blocked by nr_pending, and we need to wait for
975          * pending IO requests to complete or be queued for re-try.
976          * Thus the number queued (nr_queued) plus this request (extra)
977          * must match the number of pending IOs (nr_pending) before
978          * we continue.
979          */
980         spin_lock_irq(&conf->resync_lock);
981         conf->barrier++;
982         conf->nr_waiting++;
983         wait_event_lock_irq_cmd(conf->wait_barrier,
984                                 conf->nr_pending == conf->nr_queued+extra,
985                                 conf->resync_lock,
986                                 flush_pending_writes(conf));
987
988         spin_unlock_irq(&conf->resync_lock);
989 }
990
991 static void unfreeze_array(struct r10conf *conf)
992 {
993         /* reverse the effect of the freeze */
994         spin_lock_irq(&conf->resync_lock);
995         conf->barrier--;
996         conf->nr_waiting--;
997         wake_up(&conf->wait_barrier);
998         spin_unlock_irq(&conf->resync_lock);
999 }
1000
1001 static sector_t choose_data_offset(struct r10bio *r10_bio,
1002                                    struct md_rdev *rdev)
1003 {
1004         if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1005             test_bit(R10BIO_Previous, &r10_bio->state))
1006                 return rdev->data_offset;
1007         else
1008                 return rdev->new_data_offset;
1009 }
1010
1011 struct raid10_plug_cb {
1012         struct blk_plug_cb      cb;
1013         struct bio_list         pending;
1014         int                     pending_cnt;
1015 };
1016
1017 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1018 {
1019         struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1020                                                    cb);
1021         struct mddev *mddev = plug->cb.data;
1022         struct r10conf *conf = mddev->private;
1023         struct bio *bio;
1024
1025         if (from_schedule || current->bio_list) {
1026                 spin_lock_irq(&conf->device_lock);
1027                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1028                 conf->pending_count += plug->pending_cnt;
1029                 spin_unlock_irq(&conf->device_lock);
1030                 wake_up(&conf->wait_barrier);
1031                 md_wakeup_thread(mddev->thread);
1032                 kfree(plug);
1033                 return;
1034         }
1035
1036         /* we aren't scheduling, so we can do the write-out directly. */
1037         bio = bio_list_get(&plug->pending);
1038         bitmap_unplug(mddev->bitmap);
1039         wake_up(&conf->wait_barrier);
1040
1041         while (bio) { /* submit pending writes */
1042                 struct bio *next = bio->bi_next;
1043                 bio->bi_next = NULL;
1044                 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1045                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1046                         /* Just ignore it */
1047                         bio_endio(bio);
1048                 else
1049                         generic_make_request(bio);
1050                 bio = next;
1051         }
1052         kfree(plug);
1053 }
1054
1055 static void __make_request(struct mddev *mddev, struct bio *bio)
1056 {
1057         struct r10conf *conf = mddev->private;
1058         struct r10bio *r10_bio;
1059         struct bio *read_bio;
1060         int i;
1061         const int rw = bio_data_dir(bio);
1062         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1063         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1064         const unsigned long do_discard = (bio->bi_rw
1065                                           & (REQ_DISCARD | REQ_SECURE));
1066         const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1067         unsigned long flags;
1068         struct md_rdev *blocked_rdev;
1069         struct blk_plug_cb *cb;
1070         struct raid10_plug_cb *plug = NULL;
1071         int sectors_handled;
1072         int max_sectors;
1073         int sectors;
1074
1075         /*
1076          * Register the new request and wait if the reconstruction
1077          * thread has put up a bar for new requests.
1078          * Continue immediately if no resync is active currently.
1079          */
1080         wait_barrier(conf);
1081
1082         sectors = bio_sectors(bio);
1083         while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1084             bio->bi_iter.bi_sector < conf->reshape_progress &&
1085             bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1086                 /* IO spans the reshape position.  Need to wait for
1087                  * reshape to pass
1088                  */
1089                 allow_barrier(conf);
1090                 wait_event(conf->wait_barrier,
1091                            conf->reshape_progress <= bio->bi_iter.bi_sector ||
1092                            conf->reshape_progress >= bio->bi_iter.bi_sector +
1093                            sectors);
1094                 wait_barrier(conf);
1095         }
1096         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1097             bio_data_dir(bio) == WRITE &&
1098             (mddev->reshape_backwards
1099              ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1100                 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1101              : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1102                 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1103                 /* Need to update reshape_position in metadata */
1104                 mddev->reshape_position = conf->reshape_progress;
1105                 set_mask_bits(&mddev->flags, 0,
1106                               BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1107                 md_wakeup_thread(mddev->thread);
1108                 wait_event(mddev->sb_wait,
1109                            !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1110
1111                 conf->reshape_safe = mddev->reshape_position;
1112         }
1113
1114         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1115
1116         r10_bio->master_bio = bio;
1117         r10_bio->sectors = sectors;
1118
1119         r10_bio->mddev = mddev;
1120         r10_bio->sector = bio->bi_iter.bi_sector;
1121         r10_bio->state = 0;
1122
1123         /* We might need to issue multiple reads to different
1124          * devices if there are bad blocks around, so we keep
1125          * track of the number of reads in bio->bi_phys_segments.
1126          * If this is 0, there is only one r10_bio and no locking
1127          * will be needed when the request completes.  If it is
1128          * non-zero, then it is the number of not-completed requests.
1129          */
1130         bio->bi_phys_segments = 0;
1131         bio_clear_flag(bio, BIO_SEG_VALID);
1132
1133         if (rw == READ) {
1134                 /*
1135                  * read balancing logic:
1136                  */
1137                 struct md_rdev *rdev;
1138                 int slot;
1139
1140 read_again:
1141                 rdev = read_balance(conf, r10_bio, &max_sectors);
1142                 if (!rdev) {
1143                         raid_end_bio_io(r10_bio);
1144                         return;
1145                 }
1146                 slot = r10_bio->read_slot;
1147
1148                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1149                 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1150                          max_sectors);
1151
1152                 r10_bio->devs[slot].bio = read_bio;
1153                 r10_bio->devs[slot].rdev = rdev;
1154
1155                 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1156                         choose_data_offset(r10_bio, rdev);
1157                 read_bio->bi_bdev = rdev->bdev;
1158                 read_bio->bi_end_io = raid10_end_read_request;
1159                 read_bio->bi_rw = READ | do_sync;
1160                 read_bio->bi_private = r10_bio;
1161
1162                 if (max_sectors < r10_bio->sectors) {
1163                         /* Could not read all from this device, so we will
1164                          * need another r10_bio.
1165                          */
1166                         sectors_handled = (r10_bio->sector + max_sectors
1167                                            - bio->bi_iter.bi_sector);
1168                         r10_bio->sectors = max_sectors;
1169                         spin_lock_irq(&conf->device_lock);
1170                         if (bio->bi_phys_segments == 0)
1171                                 bio->bi_phys_segments = 2;
1172                         else
1173                                 bio->bi_phys_segments++;
1174                         spin_unlock_irq(&conf->device_lock);
1175                         /* Cannot call generic_make_request directly
1176                          * as that will be queued in __generic_make_request
1177                          * and subsequent mempool_alloc might block
1178                          * waiting for it.  so hand bio over to raid10d.
1179                          */
1180                         reschedule_retry(r10_bio);
1181
1182                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1183
1184                         r10_bio->master_bio = bio;
1185                         r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1186                         r10_bio->state = 0;
1187                         r10_bio->mddev = mddev;
1188                         r10_bio->sector = bio->bi_iter.bi_sector +
1189                                 sectors_handled;
1190                         goto read_again;
1191                 } else
1192                         generic_make_request(read_bio);
1193                 return;
1194         }
1195
1196         /*
1197          * WRITE:
1198          */
1199         if (conf->pending_count >= max_queued_requests) {
1200                 md_wakeup_thread(mddev->thread);
1201                 wait_event(conf->wait_barrier,
1202                            conf->pending_count < max_queued_requests);
1203         }
1204         /* first select target devices under rcu_lock and
1205          * inc refcount on their rdev.  Record them by setting
1206          * bios[x] to bio
1207          * If there are known/acknowledged bad blocks on any device
1208          * on which we have seen a write error, we want to avoid
1209          * writing to those blocks.  This potentially requires several
1210          * writes to write around the bad blocks.  Each set of writes
1211          * gets its own r10_bio with a set of bios attached.  The number
1212          * of r10_bios is recored in bio->bi_phys_segments just as with
1213          * the read case.
1214          */
1215
1216         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1217         raid10_find_phys(conf, r10_bio);
1218 retry_write:
1219         blocked_rdev = NULL;
1220         rcu_read_lock();
1221         max_sectors = r10_bio->sectors;
1222
1223         for (i = 0;  i < conf->copies; i++) {
1224                 int d = r10_bio->devs[i].devnum;
1225                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1226                 struct md_rdev *rrdev = rcu_dereference(
1227                         conf->mirrors[d].replacement);
1228                 if (rdev == rrdev)
1229                         rrdev = NULL;
1230                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1231                         atomic_inc(&rdev->nr_pending);
1232                         blocked_rdev = rdev;
1233                         break;
1234                 }
1235                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1236                         atomic_inc(&rrdev->nr_pending);
1237                         blocked_rdev = rrdev;
1238                         break;
1239                 }
1240                 if (rdev && (test_bit(Faulty, &rdev->flags)))
1241                         rdev = NULL;
1242                 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1243                         rrdev = NULL;
1244
1245                 r10_bio->devs[i].bio = NULL;
1246                 r10_bio->devs[i].repl_bio = NULL;
1247
1248                 if (!rdev && !rrdev) {
1249                         set_bit(R10BIO_Degraded, &r10_bio->state);
1250                         continue;
1251                 }
1252                 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1253                         sector_t first_bad;
1254                         sector_t dev_sector = r10_bio->devs[i].addr;
1255                         int bad_sectors;
1256                         int is_bad;
1257
1258                         is_bad = is_badblock(rdev, dev_sector,
1259                                              max_sectors,
1260                                              &first_bad, &bad_sectors);
1261                         if (is_bad < 0) {
1262                                 /* Mustn't write here until the bad block
1263                                  * is acknowledged
1264                                  */
1265                                 atomic_inc(&rdev->nr_pending);
1266                                 set_bit(BlockedBadBlocks, &rdev->flags);
1267                                 blocked_rdev = rdev;
1268                                 break;
1269                         }
1270                         if (is_bad && first_bad <= dev_sector) {
1271                                 /* Cannot write here at all */
1272                                 bad_sectors -= (dev_sector - first_bad);
1273                                 if (bad_sectors < max_sectors)
1274                                         /* Mustn't write more than bad_sectors
1275                                          * to other devices yet
1276                                          */
1277                                         max_sectors = bad_sectors;
1278                                 /* We don't set R10BIO_Degraded as that
1279                                  * only applies if the disk is missing,
1280                                  * so it might be re-added, and we want to
1281                                  * know to recover this chunk.
1282                                  * In this case the device is here, and the
1283                                  * fact that this chunk is not in-sync is
1284                                  * recorded in the bad block log.
1285                                  */
1286                                 continue;
1287                         }
1288                         if (is_bad) {
1289                                 int good_sectors = first_bad - dev_sector;
1290                                 if (good_sectors < max_sectors)
1291                                         max_sectors = good_sectors;
1292                         }
1293                 }
1294                 if (rdev) {
1295                         r10_bio->devs[i].bio = bio;
1296                         atomic_inc(&rdev->nr_pending);
1297                 }
1298                 if (rrdev) {
1299                         r10_bio->devs[i].repl_bio = bio;
1300                         atomic_inc(&rrdev->nr_pending);
1301                 }
1302         }
1303         rcu_read_unlock();
1304
1305         if (unlikely(blocked_rdev)) {
1306                 /* Have to wait for this device to get unblocked, then retry */
1307                 int j;
1308                 int d;
1309
1310                 for (j = 0; j < i; j++) {
1311                         if (r10_bio->devs[j].bio) {
1312                                 d = r10_bio->devs[j].devnum;
1313                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1314                         }
1315                         if (r10_bio->devs[j].repl_bio) {
1316                                 struct md_rdev *rdev;
1317                                 d = r10_bio->devs[j].devnum;
1318                                 rdev = conf->mirrors[d].replacement;
1319                                 if (!rdev) {
1320                                         /* Race with remove_disk */
1321                                         smp_mb();
1322                                         rdev = conf->mirrors[d].rdev;
1323                                 }
1324                                 rdev_dec_pending(rdev, mddev);
1325                         }
1326                 }
1327                 allow_barrier(conf);
1328                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1329                 wait_barrier(conf);
1330                 goto retry_write;
1331         }
1332
1333         if (max_sectors < r10_bio->sectors) {
1334                 /* We are splitting this into multiple parts, so
1335                  * we need to prepare for allocating another r10_bio.
1336                  */
1337                 r10_bio->sectors = max_sectors;
1338                 spin_lock_irq(&conf->device_lock);
1339                 if (bio->bi_phys_segments == 0)
1340                         bio->bi_phys_segments = 2;
1341                 else
1342                         bio->bi_phys_segments++;
1343                 spin_unlock_irq(&conf->device_lock);
1344         }
1345         sectors_handled = r10_bio->sector + max_sectors -
1346                 bio->bi_iter.bi_sector;
1347
1348         atomic_set(&r10_bio->remaining, 1);
1349         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1350
1351         for (i = 0; i < conf->copies; i++) {
1352                 struct bio *mbio;
1353                 int d = r10_bio->devs[i].devnum;
1354                 if (r10_bio->devs[i].bio) {
1355                         struct md_rdev *rdev = conf->mirrors[d].rdev;
1356                         mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1357                         bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1358                                  max_sectors);
1359                         r10_bio->devs[i].bio = mbio;
1360
1361                         mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1362                                            choose_data_offset(r10_bio,
1363                                                               rdev));
1364                         mbio->bi_bdev = rdev->bdev;
1365                         mbio->bi_end_io = raid10_end_write_request;
1366                         mbio->bi_rw =
1367                                 WRITE | do_sync | do_fua | do_discard | do_same;
1368                         mbio->bi_private = r10_bio;
1369
1370                         atomic_inc(&r10_bio->remaining);
1371
1372                         cb = blk_check_plugged(raid10_unplug, mddev,
1373                                                sizeof(*plug));
1374                         if (cb)
1375                                 plug = container_of(cb, struct raid10_plug_cb,
1376                                                     cb);
1377                         else
1378                                 plug = NULL;
1379                         spin_lock_irqsave(&conf->device_lock, flags);
1380                         if (plug) {
1381                                 bio_list_add(&plug->pending, mbio);
1382                                 plug->pending_cnt++;
1383                         } else {
1384                                 bio_list_add(&conf->pending_bio_list, mbio);
1385                                 conf->pending_count++;
1386                         }
1387                         spin_unlock_irqrestore(&conf->device_lock, flags);
1388                         if (!plug)
1389                                 md_wakeup_thread(mddev->thread);
1390                 }
1391
1392                 if (r10_bio->devs[i].repl_bio) {
1393                         struct md_rdev *rdev = conf->mirrors[d].replacement;
1394                         if (rdev == NULL) {
1395                                 /* Replacement just got moved to main 'rdev' */
1396                                 smp_mb();
1397                                 rdev = conf->mirrors[d].rdev;
1398                         }
1399                         mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1400                         bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1401                                  max_sectors);
1402                         r10_bio->devs[i].repl_bio = mbio;
1403
1404                         mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1405                                            choose_data_offset(
1406                                                    r10_bio, rdev));
1407                         mbio->bi_bdev = rdev->bdev;
1408                         mbio->bi_end_io = raid10_end_write_request;
1409                         mbio->bi_rw =
1410                                 WRITE | do_sync | do_fua | do_discard | do_same;
1411                         mbio->bi_private = r10_bio;
1412
1413                         atomic_inc(&r10_bio->remaining);
1414                         spin_lock_irqsave(&conf->device_lock, flags);
1415                         bio_list_add(&conf->pending_bio_list, mbio);
1416                         conf->pending_count++;
1417                         spin_unlock_irqrestore(&conf->device_lock, flags);
1418                         if (!mddev_check_plugged(mddev))
1419                                 md_wakeup_thread(mddev->thread);
1420                 }
1421         }
1422
1423         /* Don't remove the bias on 'remaining' (one_write_done) until
1424          * after checking if we need to go around again.
1425          */
1426
1427         if (sectors_handled < bio_sectors(bio)) {
1428                 one_write_done(r10_bio);
1429                 /* We need another r10_bio.  It has already been counted
1430                  * in bio->bi_phys_segments.
1431                  */
1432                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1433
1434                 r10_bio->master_bio = bio;
1435                 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1436
1437                 r10_bio->mddev = mddev;
1438                 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1439                 r10_bio->state = 0;
1440                 goto retry_write;
1441         }
1442         one_write_done(r10_bio);
1443 }
1444
1445 static void raid10_make_request(struct mddev *mddev, struct bio *bio)
1446 {
1447         struct r10conf *conf = mddev->private;
1448         sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1449         int chunk_sects = chunk_mask + 1;
1450
1451         struct bio *split;
1452
1453         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1454                 md_flush_request(mddev, bio);
1455                 return;
1456         }
1457
1458         md_write_start(mddev, bio);
1459
1460         do {
1461
1462                 /*
1463                  * If this request crosses a chunk boundary, we need to split
1464                  * it.
1465                  */
1466                 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1467                              bio_sectors(bio) > chunk_sects
1468                              && (conf->geo.near_copies < conf->geo.raid_disks
1469                                  || conf->prev.near_copies <
1470                                  conf->prev.raid_disks))) {
1471                         split = bio_split(bio, chunk_sects -
1472                                           (bio->bi_iter.bi_sector &
1473                                            (chunk_sects - 1)),
1474                                           GFP_NOIO, fs_bio_set);
1475                         bio_chain(split, bio);
1476                 } else {
1477                         split = bio;
1478                 }
1479
1480                 __make_request(mddev, split);
1481         } while (split != bio);
1482
1483         /* In case raid10d snuck in to freeze_array */
1484         wake_up(&conf->wait_barrier);
1485 }
1486
1487 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1488 {
1489         struct r10conf *conf = mddev->private;
1490         int i;
1491
1492         if (conf->geo.near_copies < conf->geo.raid_disks)
1493                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1494         if (conf->geo.near_copies > 1)
1495                 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1496         if (conf->geo.far_copies > 1) {
1497                 if (conf->geo.far_offset)
1498                         seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1499                 else
1500                         seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1501                 if (conf->geo.far_set_size != conf->geo.raid_disks)
1502                         seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1503         }
1504         seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1505                                         conf->geo.raid_disks - mddev->degraded);
1506         for (i = 0; i < conf->geo.raid_disks; i++)
1507                 seq_printf(seq, "%s",
1508                               conf->mirrors[i].rdev &&
1509                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1510         seq_printf(seq, "]");
1511 }
1512
1513 /* check if there are enough drives for
1514  * every block to appear on atleast one.
1515  * Don't consider the device numbered 'ignore'
1516  * as we might be about to remove it.
1517  */
1518 static int _enough(struct r10conf *conf, int previous, int ignore)
1519 {
1520         int first = 0;
1521         int has_enough = 0;
1522         int disks, ncopies;
1523         if (previous) {
1524                 disks = conf->prev.raid_disks;
1525                 ncopies = conf->prev.near_copies;
1526         } else {
1527                 disks = conf->geo.raid_disks;
1528                 ncopies = conf->geo.near_copies;
1529         }
1530
1531         rcu_read_lock();
1532         do {
1533                 int n = conf->copies;
1534                 int cnt = 0;
1535                 int this = first;
1536                 while (n--) {
1537                         struct md_rdev *rdev;
1538                         if (this != ignore &&
1539                             (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1540                             test_bit(In_sync, &rdev->flags))
1541                                 cnt++;
1542                         this = (this+1) % disks;
1543                 }
1544                 if (cnt == 0)
1545                         goto out;
1546                 first = (first + ncopies) % disks;
1547         } while (first != 0);
1548         has_enough = 1;
1549 out:
1550         rcu_read_unlock();
1551         return has_enough;
1552 }
1553
1554 static int enough(struct r10conf *conf, int ignore)
1555 {
1556         /* when calling 'enough', both 'prev' and 'geo' must
1557          * be stable.
1558          * This is ensured if ->reconfig_mutex or ->device_lock
1559          * is held.
1560          */
1561         return _enough(conf, 0, ignore) &&
1562                 _enough(conf, 1, ignore);
1563 }
1564
1565 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1566 {
1567         char b[BDEVNAME_SIZE];
1568         struct r10conf *conf = mddev->private;
1569         unsigned long flags;
1570
1571         /*
1572          * If it is not operational, then we have already marked it as dead
1573          * else if it is the last working disks, ignore the error, let the
1574          * next level up know.
1575          * else mark the drive as failed
1576          */
1577         spin_lock_irqsave(&conf->device_lock, flags);
1578         if (test_bit(In_sync, &rdev->flags)
1579             && !enough(conf, rdev->raid_disk)) {
1580                 /*
1581                  * Don't fail the drive, just return an IO error.
1582                  */
1583                 spin_unlock_irqrestore(&conf->device_lock, flags);
1584                 return;
1585         }
1586         if (test_and_clear_bit(In_sync, &rdev->flags))
1587                 mddev->degraded++;
1588         /*
1589          * If recovery is running, make sure it aborts.
1590          */
1591         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1592         set_bit(Blocked, &rdev->flags);
1593         set_bit(Faulty, &rdev->flags);
1594         set_mask_bits(&mddev->flags, 0,
1595                       BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1596         spin_unlock_irqrestore(&conf->device_lock, flags);
1597         printk(KERN_ALERT
1598                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1599                "md/raid10:%s: Operation continuing on %d devices.\n",
1600                mdname(mddev), bdevname(rdev->bdev, b),
1601                mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1602 }
1603
1604 static void print_conf(struct r10conf *conf)
1605 {
1606         int i;
1607         struct raid10_info *tmp;
1608
1609         printk(KERN_DEBUG "RAID10 conf printout:\n");
1610         if (!conf) {
1611                 printk(KERN_DEBUG "(!conf)\n");
1612                 return;
1613         }
1614         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1615                 conf->geo.raid_disks);
1616
1617         for (i = 0; i < conf->geo.raid_disks; i++) {
1618                 char b[BDEVNAME_SIZE];
1619                 tmp = conf->mirrors + i;
1620                 if (tmp->rdev)
1621                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1622                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1623                                 !test_bit(Faulty, &tmp->rdev->flags),
1624                                 bdevname(tmp->rdev->bdev,b));
1625         }
1626 }
1627
1628 static void close_sync(struct r10conf *conf)
1629 {
1630         wait_barrier(conf);
1631         allow_barrier(conf);
1632
1633         mempool_destroy(conf->r10buf_pool);
1634         conf->r10buf_pool = NULL;
1635 }
1636
1637 static int raid10_spare_active(struct mddev *mddev)
1638 {
1639         int i;
1640         struct r10conf *conf = mddev->private;
1641         struct raid10_info *tmp;
1642         int count = 0;
1643         unsigned long flags;
1644
1645         /*
1646          * Find all non-in_sync disks within the RAID10 configuration
1647          * and mark them in_sync
1648          */
1649         for (i = 0; i < conf->geo.raid_disks; i++) {
1650                 tmp = conf->mirrors + i;
1651                 if (tmp->replacement
1652                     && tmp->replacement->recovery_offset == MaxSector
1653                     && !test_bit(Faulty, &tmp->replacement->flags)
1654                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1655                         /* Replacement has just become active */
1656                         if (!tmp->rdev
1657                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1658                                 count++;
1659                         if (tmp->rdev) {
1660                                 /* Replaced device not technically faulty,
1661                                  * but we need to be sure it gets removed
1662                                  * and never re-added.
1663                                  */
1664                                 set_bit(Faulty, &tmp->rdev->flags);
1665                                 sysfs_notify_dirent_safe(
1666                                         tmp->rdev->sysfs_state);
1667                         }
1668                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1669                 } else if (tmp->rdev
1670                            && tmp->rdev->recovery_offset == MaxSector
1671                            && !test_bit(Faulty, &tmp->rdev->flags)
1672                            && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1673                         count++;
1674                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1675                 }
1676         }
1677         spin_lock_irqsave(&conf->device_lock, flags);
1678         mddev->degraded -= count;
1679         spin_unlock_irqrestore(&conf->device_lock, flags);
1680
1681         print_conf(conf);
1682         return count;
1683 }
1684
1685 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1686 {
1687         struct r10conf *conf = mddev->private;
1688         int err = -EEXIST;
1689         int mirror;
1690         int first = 0;
1691         int last = conf->geo.raid_disks - 1;
1692
1693         if (mddev->recovery_cp < MaxSector)
1694                 /* only hot-add to in-sync arrays, as recovery is
1695                  * very different from resync
1696                  */
1697                 return -EBUSY;
1698         if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1699                 return -EINVAL;
1700
1701         if (md_integrity_add_rdev(rdev, mddev))
1702                 return -ENXIO;
1703
1704         if (rdev->raid_disk >= 0)
1705                 first = last = rdev->raid_disk;
1706
1707         if (rdev->saved_raid_disk >= first &&
1708             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1709                 mirror = rdev->saved_raid_disk;
1710         else
1711                 mirror = first;
1712         for ( ; mirror <= last ; mirror++) {
1713                 struct raid10_info *p = &conf->mirrors[mirror];
1714                 if (p->recovery_disabled == mddev->recovery_disabled)
1715                         continue;
1716                 if (p->rdev) {
1717                         if (!test_bit(WantReplacement, &p->rdev->flags) ||
1718                             p->replacement != NULL)
1719                                 continue;
1720                         clear_bit(In_sync, &rdev->flags);
1721                         set_bit(Replacement, &rdev->flags);
1722                         rdev->raid_disk = mirror;
1723                         err = 0;
1724                         if (mddev->gendisk)
1725                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1726                                                   rdev->data_offset << 9);
1727                         conf->fullsync = 1;
1728                         rcu_assign_pointer(p->replacement, rdev);
1729                         break;
1730                 }
1731
1732                 if (mddev->gendisk)
1733                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1734                                           rdev->data_offset << 9);
1735
1736                 p->head_position = 0;
1737                 p->recovery_disabled = mddev->recovery_disabled - 1;
1738                 rdev->raid_disk = mirror;
1739                 err = 0;
1740                 if (rdev->saved_raid_disk != mirror)
1741                         conf->fullsync = 1;
1742                 rcu_assign_pointer(p->rdev, rdev);
1743                 break;
1744         }
1745         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1746                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1747
1748         print_conf(conf);
1749         return err;
1750 }
1751
1752 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1753 {
1754         struct r10conf *conf = mddev->private;
1755         int err = 0;
1756         int number = rdev->raid_disk;
1757         struct md_rdev **rdevp;
1758         struct raid10_info *p = conf->mirrors + number;
1759
1760         print_conf(conf);
1761         if (rdev == p->rdev)
1762                 rdevp = &p->rdev;
1763         else if (rdev == p->replacement)
1764                 rdevp = &p->replacement;
1765         else
1766                 return 0;
1767
1768         if (test_bit(In_sync, &rdev->flags) ||
1769             atomic_read(&rdev->nr_pending)) {
1770                 err = -EBUSY;
1771                 goto abort;
1772         }
1773         /* Only remove faulty devices if recovery
1774          * is not possible.
1775          */
1776         if (!test_bit(Faulty, &rdev->flags) &&
1777             mddev->recovery_disabled != p->recovery_disabled &&
1778             (!p->replacement || p->replacement == rdev) &&
1779             number < conf->geo.raid_disks &&
1780             enough(conf, -1)) {
1781                 err = -EBUSY;
1782                 goto abort;
1783         }
1784         *rdevp = NULL;
1785         synchronize_rcu();
1786         if (atomic_read(&rdev->nr_pending)) {
1787                 /* lost the race, try later */
1788                 err = -EBUSY;
1789                 *rdevp = rdev;
1790                 goto abort;
1791         } else if (p->replacement) {
1792                 /* We must have just cleared 'rdev' */
1793                 p->rdev = p->replacement;
1794                 clear_bit(Replacement, &p->replacement->flags);
1795                 smp_mb(); /* Make sure other CPUs may see both as identical
1796                            * but will never see neither -- if they are careful.
1797                            */
1798                 p->replacement = NULL;
1799                 clear_bit(WantReplacement, &rdev->flags);
1800         } else
1801                 /* We might have just remove the Replacement as faulty
1802                  * Clear the flag just in case
1803                  */
1804                 clear_bit(WantReplacement, &rdev->flags);
1805
1806         err = md_integrity_register(mddev);
1807
1808 abort:
1809
1810         print_conf(conf);
1811         return err;
1812 }
1813
1814 static void end_sync_read(struct bio *bio)
1815 {
1816         struct r10bio *r10_bio = bio->bi_private;
1817         struct r10conf *conf = r10_bio->mddev->private;
1818         int d;
1819
1820         if (bio == r10_bio->master_bio) {
1821                 /* this is a reshape read */
1822                 d = r10_bio->read_slot; /* really the read dev */
1823         } else
1824                 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1825
1826         if (!bio->bi_error)
1827                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1828         else
1829                 /* The write handler will notice the lack of
1830                  * R10BIO_Uptodate and record any errors etc
1831                  */
1832                 atomic_add(r10_bio->sectors,
1833                            &conf->mirrors[d].rdev->corrected_errors);
1834
1835         /* for reconstruct, we always reschedule after a read.
1836          * for resync, only after all reads
1837          */
1838         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1839         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1840             atomic_dec_and_test(&r10_bio->remaining)) {
1841                 /* we have read all the blocks,
1842                  * do the comparison in process context in raid10d
1843                  */
1844                 reschedule_retry(r10_bio);
1845         }
1846 }
1847
1848 static void end_sync_request(struct r10bio *r10_bio)
1849 {
1850         struct mddev *mddev = r10_bio->mddev;
1851
1852         while (atomic_dec_and_test(&r10_bio->remaining)) {
1853                 if (r10_bio->master_bio == NULL) {
1854                         /* the primary of several recovery bios */
1855                         sector_t s = r10_bio->sectors;
1856                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1857                             test_bit(R10BIO_WriteError, &r10_bio->state))
1858                                 reschedule_retry(r10_bio);
1859                         else
1860                                 put_buf(r10_bio);
1861                         md_done_sync(mddev, s, 1);
1862                         break;
1863                 } else {
1864                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1865                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1866                             test_bit(R10BIO_WriteError, &r10_bio->state))
1867                                 reschedule_retry(r10_bio);
1868                         else
1869                                 put_buf(r10_bio);
1870                         r10_bio = r10_bio2;
1871                 }
1872         }
1873 }
1874
1875 static void end_sync_write(struct bio *bio)
1876 {
1877         struct r10bio *r10_bio = bio->bi_private;
1878         struct mddev *mddev = r10_bio->mddev;
1879         struct r10conf *conf = mddev->private;
1880         int d;
1881         sector_t first_bad;
1882         int bad_sectors;
1883         int slot;
1884         int repl;
1885         struct md_rdev *rdev = NULL;
1886
1887         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1888         if (repl)
1889                 rdev = conf->mirrors[d].replacement;
1890         else
1891                 rdev = conf->mirrors[d].rdev;
1892
1893         if (bio->bi_error) {
1894                 if (repl)
1895                         md_error(mddev, rdev);
1896                 else {
1897                         set_bit(WriteErrorSeen, &rdev->flags);
1898                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
1899                                 set_bit(MD_RECOVERY_NEEDED,
1900                                         &rdev->mddev->recovery);
1901                         set_bit(R10BIO_WriteError, &r10_bio->state);
1902                 }
1903         } else if (is_badblock(rdev,
1904                              r10_bio->devs[slot].addr,
1905                              r10_bio->sectors,
1906                              &first_bad, &bad_sectors))
1907                 set_bit(R10BIO_MadeGood, &r10_bio->state);
1908
1909         rdev_dec_pending(rdev, mddev);
1910
1911         end_sync_request(r10_bio);
1912 }
1913
1914 /*
1915  * Note: sync and recover and handled very differently for raid10
1916  * This code is for resync.
1917  * For resync, we read through virtual addresses and read all blocks.
1918  * If there is any error, we schedule a write.  The lowest numbered
1919  * drive is authoritative.
1920  * However requests come for physical address, so we need to map.
1921  * For every physical address there are raid_disks/copies virtual addresses,
1922  * which is always are least one, but is not necessarly an integer.
1923  * This means that a physical address can span multiple chunks, so we may
1924  * have to submit multiple io requests for a single sync request.
1925  */
1926 /*
1927  * We check if all blocks are in-sync and only write to blocks that
1928  * aren't in sync
1929  */
1930 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1931 {
1932         struct r10conf *conf = mddev->private;
1933         int i, first;
1934         struct bio *tbio, *fbio;
1935         int vcnt;
1936
1937         atomic_set(&r10_bio->remaining, 1);
1938
1939         /* find the first device with a block */
1940         for (i=0; i<conf->copies; i++)
1941                 if (!r10_bio->devs[i].bio->bi_error)
1942                         break;
1943
1944         if (i == conf->copies)
1945                 goto done;
1946
1947         first = i;
1948         fbio = r10_bio->devs[i].bio;
1949         fbio->bi_iter.bi_size = r10_bio->sectors << 9;
1950         fbio->bi_iter.bi_idx = 0;
1951
1952         vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1953         /* now find blocks with errors */
1954         for (i=0 ; i < conf->copies ; i++) {
1955                 int  j, d;
1956
1957                 tbio = r10_bio->devs[i].bio;
1958
1959                 if (tbio->bi_end_io != end_sync_read)
1960                         continue;
1961                 if (i == first)
1962                         continue;
1963                 if (!r10_bio->devs[i].bio->bi_error) {
1964                         /* We know that the bi_io_vec layout is the same for
1965                          * both 'first' and 'i', so we just compare them.
1966                          * All vec entries are PAGE_SIZE;
1967                          */
1968                         int sectors = r10_bio->sectors;
1969                         for (j = 0; j < vcnt; j++) {
1970                                 int len = PAGE_SIZE;
1971                                 if (sectors < (len / 512))
1972                                         len = sectors * 512;
1973                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1974                                            page_address(tbio->bi_io_vec[j].bv_page),
1975                                            len))
1976                                         break;
1977                                 sectors -= len/512;
1978                         }
1979                         if (j == vcnt)
1980                                 continue;
1981                         atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
1982                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1983                                 /* Don't fix anything. */
1984                                 continue;
1985                 }
1986                 /* Ok, we need to write this bio, either to correct an
1987                  * inconsistency or to correct an unreadable block.
1988                  * First we need to fixup bv_offset, bv_len and
1989                  * bi_vecs, as the read request might have corrupted these
1990                  */
1991                 bio_reset(tbio);
1992
1993                 tbio->bi_vcnt = vcnt;
1994                 tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
1995                 tbio->bi_rw = WRITE;
1996                 tbio->bi_private = r10_bio;
1997                 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
1998                 tbio->bi_end_io = end_sync_write;
1999
2000                 bio_copy_data(tbio, fbio);
2001
2002                 d = r10_bio->devs[i].devnum;
2003                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2004                 atomic_inc(&r10_bio->remaining);
2005                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2006
2007                 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2008                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2009                 generic_make_request(tbio);
2010         }
2011
2012         /* Now write out to any replacement devices
2013          * that are active
2014          */
2015         for (i = 0; i < conf->copies; i++) {
2016                 int d;
2017
2018                 tbio = r10_bio->devs[i].repl_bio;
2019                 if (!tbio || !tbio->bi_end_io)
2020                         continue;
2021                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2022                     && r10_bio->devs[i].bio != fbio)
2023                         bio_copy_data(tbio, fbio);
2024                 d = r10_bio->devs[i].devnum;
2025                 atomic_inc(&r10_bio->remaining);
2026                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2027                              bio_sectors(tbio));
2028                 generic_make_request(tbio);
2029         }
2030
2031 done:
2032         if (atomic_dec_and_test(&r10_bio->remaining)) {
2033                 md_done_sync(mddev, r10_bio->sectors, 1);
2034                 put_buf(r10_bio);
2035         }
2036 }
2037
2038 /*
2039  * Now for the recovery code.
2040  * Recovery happens across physical sectors.
2041  * We recover all non-is_sync drives by finding the virtual address of
2042  * each, and then choose a working drive that also has that virt address.
2043  * There is a separate r10_bio for each non-in_sync drive.
2044  * Only the first two slots are in use. The first for reading,
2045  * The second for writing.
2046  *
2047  */
2048 static void fix_recovery_read_error(struct r10bio *r10_bio)
2049 {
2050         /* We got a read error during recovery.
2051          * We repeat the read in smaller page-sized sections.
2052          * If a read succeeds, write it to the new device or record
2053          * a bad block if we cannot.
2054          * If a read fails, record a bad block on both old and
2055          * new devices.
2056          */
2057         struct mddev *mddev = r10_bio->mddev;
2058         struct r10conf *conf = mddev->private;
2059         struct bio *bio = r10_bio->devs[0].bio;
2060         sector_t sect = 0;
2061         int sectors = r10_bio->sectors;
2062         int idx = 0;
2063         int dr = r10_bio->devs[0].devnum;
2064         int dw = r10_bio->devs[1].devnum;
2065
2066         while (sectors) {
2067                 int s = sectors;
2068                 struct md_rdev *rdev;
2069                 sector_t addr;
2070                 int ok;
2071
2072                 if (s > (PAGE_SIZE>>9))
2073                         s = PAGE_SIZE >> 9;
2074
2075                 rdev = conf->mirrors[dr].rdev;
2076                 addr = r10_bio->devs[0].addr + sect,
2077                 ok = sync_page_io(rdev,
2078                                   addr,
2079                                   s << 9,
2080                                   bio->bi_io_vec[idx].bv_page,
2081                                   READ, false);
2082                 if (ok) {
2083                         rdev = conf->mirrors[dw].rdev;
2084                         addr = r10_bio->devs[1].addr + sect;
2085                         ok = sync_page_io(rdev,
2086                                           addr,
2087                                           s << 9,
2088                                           bio->bi_io_vec[idx].bv_page,
2089                                           WRITE, false);
2090                         if (!ok) {
2091                                 set_bit(WriteErrorSeen, &rdev->flags);
2092                                 if (!test_and_set_bit(WantReplacement,
2093                                                       &rdev->flags))
2094                                         set_bit(MD_RECOVERY_NEEDED,
2095                                                 &rdev->mddev->recovery);
2096                         }
2097                 }
2098                 if (!ok) {
2099                         /* We don't worry if we cannot set a bad block -
2100                          * it really is bad so there is no loss in not
2101                          * recording it yet
2102                          */
2103                         rdev_set_badblocks(rdev, addr, s, 0);
2104
2105                         if (rdev != conf->mirrors[dw].rdev) {
2106                                 /* need bad block on destination too */
2107                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2108                                 addr = r10_bio->devs[1].addr + sect;
2109                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2110                                 if (!ok) {
2111                                         /* just abort the recovery */
2112                                         printk(KERN_NOTICE
2113                                                "md/raid10:%s: recovery aborted"
2114                                                " due to read error\n",
2115                                                mdname(mddev));
2116
2117                                         conf->mirrors[dw].recovery_disabled
2118                                                 = mddev->recovery_disabled;
2119                                         set_bit(MD_RECOVERY_INTR,
2120                                                 &mddev->recovery);
2121                                         break;
2122                                 }
2123                         }
2124                 }
2125
2126                 sectors -= s;
2127                 sect += s;
2128                 idx++;
2129         }
2130 }
2131
2132 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2133 {
2134         struct r10conf *conf = mddev->private;
2135         int d;
2136         struct bio *wbio, *wbio2;
2137
2138         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2139                 fix_recovery_read_error(r10_bio);
2140                 end_sync_request(r10_bio);
2141                 return;
2142         }
2143
2144         /*
2145          * share the pages with the first bio
2146          * and submit the write request
2147          */
2148         d = r10_bio->devs[1].devnum;
2149         wbio = r10_bio->devs[1].bio;
2150         wbio2 = r10_bio->devs[1].repl_bio;
2151         /* Need to test wbio2->bi_end_io before we call
2152          * generic_make_request as if the former is NULL,
2153          * the latter is free to free wbio2.
2154          */
2155         if (wbio2 && !wbio2->bi_end_io)
2156                 wbio2 = NULL;
2157         if (wbio->bi_end_io) {
2158                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2159                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2160                 generic_make_request(wbio);
2161         }
2162         if (wbio2) {
2163                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2164                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2165                              bio_sectors(wbio2));
2166                 generic_make_request(wbio2);
2167         }
2168 }
2169
2170 /*
2171  * Used by fix_read_error() to decay the per rdev read_errors.
2172  * We halve the read error count for every hour that has elapsed
2173  * since the last recorded read error.
2174  *
2175  */
2176 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2177 {
2178         struct timespec cur_time_mon;
2179         unsigned long hours_since_last;
2180         unsigned int read_errors = atomic_read(&rdev->read_errors);
2181
2182         ktime_get_ts(&cur_time_mon);
2183
2184         if (rdev->last_read_error.tv_sec == 0 &&
2185             rdev->last_read_error.tv_nsec == 0) {
2186                 /* first time we've seen a read error */
2187                 rdev->last_read_error = cur_time_mon;
2188                 return;
2189         }
2190
2191         hours_since_last = (cur_time_mon.tv_sec -
2192                             rdev->last_read_error.tv_sec) / 3600;
2193
2194         rdev->last_read_error = cur_time_mon;
2195
2196         /*
2197          * if hours_since_last is > the number of bits in read_errors
2198          * just set read errors to 0. We do this to avoid
2199          * overflowing the shift of read_errors by hours_since_last.
2200          */
2201         if (hours_since_last >= 8 * sizeof(read_errors))
2202                 atomic_set(&rdev->read_errors, 0);
2203         else
2204                 atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2205 }
2206
2207 static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2208                             int sectors, struct page *page, int rw)
2209 {
2210         sector_t first_bad;
2211         int bad_sectors;
2212
2213         if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2214             && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2215                 return -1;
2216         if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2217                 /* success */
2218                 return 1;
2219         if (rw == WRITE) {
2220                 set_bit(WriteErrorSeen, &rdev->flags);
2221                 if (!test_and_set_bit(WantReplacement, &rdev->flags))
2222                         set_bit(MD_RECOVERY_NEEDED,
2223                                 &rdev->mddev->recovery);
2224         }
2225         /* need to record an error - either for the block or the device */
2226         if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2227                 md_error(rdev->mddev, rdev);
2228         return 0;
2229 }
2230
2231 /*
2232  * This is a kernel thread which:
2233  *
2234  *      1.      Retries failed read operations on working mirrors.
2235  *      2.      Updates the raid superblock when problems encounter.
2236  *      3.      Performs writes following reads for array synchronising.
2237  */
2238
2239 static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2240 {
2241         int sect = 0; /* Offset from r10_bio->sector */
2242         int sectors = r10_bio->sectors;
2243         struct md_rdev*rdev;
2244         int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2245         int d = r10_bio->devs[r10_bio->read_slot].devnum;
2246
2247         /* still own a reference to this rdev, so it cannot
2248          * have been cleared recently.
2249          */
2250         rdev = conf->mirrors[d].rdev;
2251
2252         if (test_bit(Faulty, &rdev->flags))
2253                 /* drive has already been failed, just ignore any
2254                    more fix_read_error() attempts */
2255                 return;
2256
2257         check_decay_read_errors(mddev, rdev);
2258         atomic_inc(&rdev->read_errors);
2259         if (atomic_read(&rdev->read_errors) > max_read_errors) {
2260                 char b[BDEVNAME_SIZE];
2261                 bdevname(rdev->bdev, b);
2262
2263                 printk(KERN_NOTICE
2264                        "md/raid10:%s: %s: Raid device exceeded "
2265                        "read_error threshold [cur %d:max %d]\n",
2266                        mdname(mddev), b,
2267                        atomic_read(&rdev->read_errors), max_read_errors);
2268                 printk(KERN_NOTICE
2269                        "md/raid10:%s: %s: Failing raid device\n",
2270                        mdname(mddev), b);
2271                 md_error(mddev, conf->mirrors[d].rdev);
2272                 r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2273                 return;
2274         }
2275
2276         while(sectors) {
2277                 int s = sectors;
2278                 int sl = r10_bio->read_slot;
2279                 int success = 0;
2280                 int start;
2281
2282                 if (s > (PAGE_SIZE>>9))
2283                         s = PAGE_SIZE >> 9;
2284
2285                 rcu_read_lock();
2286                 do {
2287                         sector_t first_bad;
2288                         int bad_sectors;
2289
2290                         d = r10_bio->devs[sl].devnum;
2291                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2292                         if (rdev &&
2293                             test_bit(In_sync, &rdev->flags) &&
2294                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2295                                         &first_bad, &bad_sectors) == 0) {
2296                                 atomic_inc(&rdev->nr_pending);
2297                                 rcu_read_unlock();
2298                                 success = sync_page_io(rdev,
2299                                                        r10_bio->devs[sl].addr +
2300                                                        sect,
2301                                                        s<<9,
2302                                                        conf->tmppage, READ, false);
2303                                 rdev_dec_pending(rdev, mddev);
2304                                 rcu_read_lock();
2305                                 if (success)
2306                                         break;
2307                         }
2308                         sl++;
2309                         if (sl == conf->copies)
2310                                 sl = 0;
2311                 } while (!success && sl != r10_bio->read_slot);
2312                 rcu_read_unlock();
2313
2314                 if (!success) {
2315                         /* Cannot read from anywhere, just mark the block
2316                          * as bad on the first device to discourage future
2317                          * reads.
2318                          */
2319                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2320                         rdev = conf->mirrors[dn].rdev;
2321
2322                         if (!rdev_set_badblocks(
2323                                     rdev,
2324                                     r10_bio->devs[r10_bio->read_slot].addr
2325                                     + sect,
2326                                     s, 0)) {
2327                                 md_error(mddev, rdev);
2328                                 r10_bio->devs[r10_bio->read_slot].bio
2329                                         = IO_BLOCKED;
2330                         }
2331                         break;
2332                 }
2333
2334                 start = sl;
2335                 /* write it back and re-read */
2336                 rcu_read_lock();
2337                 while (sl != r10_bio->read_slot) {
2338                         char b[BDEVNAME_SIZE];
2339
2340                         if (sl==0)
2341                                 sl = conf->copies;
2342                         sl--;
2343                         d = r10_bio->devs[sl].devnum;
2344                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2345                         if (!rdev ||
2346                             !test_bit(In_sync, &rdev->flags))
2347                                 continue;
2348
2349                         atomic_inc(&rdev->nr_pending);
2350                         rcu_read_unlock();
2351                         if (r10_sync_page_io(rdev,
2352                                              r10_bio->devs[sl].addr +
2353                                              sect,
2354                                              s, conf->tmppage, WRITE)
2355                             == 0) {
2356                                 /* Well, this device is dead */
2357                                 printk(KERN_NOTICE
2358                                        "md/raid10:%s: read correction "
2359                                        "write failed"
2360                                        " (%d sectors at %llu on %s)\n",
2361                                        mdname(mddev), s,
2362                                        (unsigned long long)(
2363                                                sect +
2364                                                choose_data_offset(r10_bio,
2365                                                                   rdev)),
2366                                        bdevname(rdev->bdev, b));
2367                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2368                                        "drive\n",
2369                                        mdname(mddev),
2370                                        bdevname(rdev->bdev, b));
2371                         }
2372                         rdev_dec_pending(rdev, mddev);
2373                         rcu_read_lock();
2374                 }
2375                 sl = start;
2376                 while (sl != r10_bio->read_slot) {
2377                         char b[BDEVNAME_SIZE];
2378
2379                         if (sl==0)
2380                                 sl = conf->copies;
2381                         sl--;
2382                         d = r10_bio->devs[sl].devnum;
2383                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2384                         if (!rdev ||
2385                             !test_bit(In_sync, &rdev->flags))
2386                                 continue;
2387
2388                         atomic_inc(&rdev->nr_pending);
2389                         rcu_read_unlock();
2390                         switch (r10_sync_page_io(rdev,
2391                                              r10_bio->devs[sl].addr +
2392                                              sect,
2393                                              s, conf->tmppage,
2394                                                  READ)) {
2395                         case 0:
2396                                 /* Well, this device is dead */
2397                                 printk(KERN_NOTICE
2398                                        "md/raid10:%s: unable to read back "
2399                                        "corrected sectors"
2400                                        " (%d sectors at %llu on %s)\n",
2401                                        mdname(mddev), s,
2402                                        (unsigned long long)(
2403                                                sect +
2404                                                choose_data_offset(r10_bio, rdev)),
2405                                        bdevname(rdev->bdev, b));
2406                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2407                                        "drive\n",
2408                                        mdname(mddev),
2409                                        bdevname(rdev->bdev, b));
2410                                 break;
2411                         case 1:
2412                                 printk(KERN_INFO
2413                                        "md/raid10:%s: read error corrected"
2414                                        " (%d sectors at %llu on %s)\n",
2415                                        mdname(mddev), s,
2416                                        (unsigned long long)(
2417                                                sect +
2418                                                choose_data_offset(r10_bio, rdev)),
2419                                        bdevname(rdev->bdev, b));
2420                                 atomic_add(s, &rdev->corrected_errors);
2421                         }
2422
2423                         rdev_dec_pending(rdev, mddev);
2424                         rcu_read_lock();
2425                 }
2426                 rcu_read_unlock();
2427
2428                 sectors -= s;
2429                 sect += s;
2430         }
2431 }
2432
2433 static int narrow_write_error(struct r10bio *r10_bio, int i)
2434 {
2435         struct bio *bio = r10_bio->master_bio;
2436         struct mddev *mddev = r10_bio->mddev;
2437         struct r10conf *conf = mddev->private;
2438         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2439         /* bio has the data to be written to slot 'i' where
2440          * we just recently had a write error.
2441          * We repeatedly clone the bio and trim down to one block,
2442          * then try the write.  Where the write fails we record
2443          * a bad block.
2444          * It is conceivable that the bio doesn't exactly align with
2445          * blocks.  We must handle this.
2446          *
2447          * We currently own a reference to the rdev.
2448          */
2449
2450         int block_sectors;
2451         sector_t sector;
2452         int sectors;
2453         int sect_to_write = r10_bio->sectors;
2454         int ok = 1;
2455
2456         if (rdev->badblocks.shift < 0)
2457                 return 0;
2458
2459         block_sectors = roundup(1 << rdev->badblocks.shift,
2460                                 bdev_logical_block_size(rdev->bdev) >> 9);
2461         sector = r10_bio->sector;
2462         sectors = ((r10_bio->sector + block_sectors)
2463                    & ~(sector_t)(block_sectors - 1))
2464                 - sector;
2465
2466         while (sect_to_write) {
2467                 struct bio *wbio;
2468                 if (sectors > sect_to_write)
2469                         sectors = sect_to_write;
2470                 /* Write at 'sector' for 'sectors' */
2471                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2472                 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2473                 wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2474                                    choose_data_offset(r10_bio, rdev) +
2475                                    (sector - r10_bio->sector));
2476                 wbio->bi_bdev = rdev->bdev;
2477                 if (submit_bio_wait(WRITE, wbio) < 0)
2478                         /* Failure! */
2479                         ok = rdev_set_badblocks(rdev, sector,
2480                                                 sectors, 0)
2481                                 && ok;
2482
2483                 bio_put(wbio);
2484                 sect_to_write -= sectors;
2485                 sector += sectors;
2486                 sectors = block_sectors;
2487         }
2488         return ok;
2489 }
2490
2491 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2492 {
2493         int slot = r10_bio->read_slot;
2494         struct bio *bio;
2495         struct r10conf *conf = mddev->private;
2496         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2497         char b[BDEVNAME_SIZE];
2498         unsigned long do_sync;
2499         int max_sectors;
2500
2501         /* we got a read error. Maybe the drive is bad.  Maybe just
2502          * the block and we can fix it.
2503          * We freeze all other IO, and try reading the block from
2504          * other devices.  When we find one, we re-write
2505          * and check it that fixes the read error.
2506          * This is all done synchronously while the array is
2507          * frozen.
2508          */
2509         bio = r10_bio->devs[slot].bio;
2510         bdevname(bio->bi_bdev, b);
2511         bio_put(bio);
2512         r10_bio->devs[slot].bio = NULL;
2513
2514         if (mddev->ro == 0) {
2515                 freeze_array(conf, 1);
2516                 fix_read_error(conf, mddev, r10_bio);
2517                 unfreeze_array(conf);
2518         } else
2519                 r10_bio->devs[slot].bio = IO_BLOCKED;
2520
2521         rdev_dec_pending(rdev, mddev);
2522
2523 read_more:
2524         rdev = read_balance(conf, r10_bio, &max_sectors);
2525         if (rdev == NULL) {
2526                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2527                        " read error for block %llu\n",
2528                        mdname(mddev), b,
2529                        (unsigned long long)r10_bio->sector);
2530                 raid_end_bio_io(r10_bio);
2531                 return;
2532         }
2533
2534         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2535         slot = r10_bio->read_slot;
2536         printk_ratelimited(
2537                 KERN_ERR
2538                 "md/raid10:%s: %s: redirecting "
2539                 "sector %llu to another mirror\n",
2540                 mdname(mddev),
2541                 bdevname(rdev->bdev, b),
2542                 (unsigned long long)r10_bio->sector);
2543         bio = bio_clone_mddev(r10_bio->master_bio,
2544                               GFP_NOIO, mddev);
2545         bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2546         r10_bio->devs[slot].bio = bio;
2547         r10_bio->devs[slot].rdev = rdev;
2548         bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2549                 + choose_data_offset(r10_bio, rdev);
2550         bio->bi_bdev = rdev->bdev;
2551         bio->bi_rw = READ | do_sync;
2552         bio->bi_private = r10_bio;
2553         bio->bi_end_io = raid10_end_read_request;
2554         if (max_sectors < r10_bio->sectors) {
2555                 /* Drat - have to split this up more */
2556                 struct bio *mbio = r10_bio->master_bio;
2557                 int sectors_handled =
2558                         r10_bio->sector + max_sectors
2559                         - mbio->bi_iter.bi_sector;
2560                 r10_bio->sectors = max_sectors;
2561                 spin_lock_irq(&conf->device_lock);
2562                 if (mbio->bi_phys_segments == 0)
2563                         mbio->bi_phys_segments = 2;
2564                 else
2565                         mbio->bi_phys_segments++;
2566                 spin_unlock_irq(&conf->device_lock);
2567                 generic_make_request(bio);
2568
2569                 r10_bio = mempool_alloc(conf->r10bio_pool,
2570                                         GFP_NOIO);
2571                 r10_bio->master_bio = mbio;
2572                 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2573                 r10_bio->state = 0;
2574                 set_bit(R10BIO_ReadError,
2575                         &r10_bio->state);
2576                 r10_bio->mddev = mddev;
2577                 r10_bio->sector = mbio->bi_iter.bi_sector
2578                         + sectors_handled;
2579
2580                 goto read_more;
2581         } else
2582                 generic_make_request(bio);
2583 }
2584
2585 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2586 {
2587         /* Some sort of write request has finished and it
2588          * succeeded in writing where we thought there was a
2589          * bad block.  So forget the bad block.
2590          * Or possibly if failed and we need to record
2591          * a bad block.
2592          */
2593         int m;
2594         struct md_rdev *rdev;
2595
2596         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2597             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2598                 for (m = 0; m < conf->copies; m++) {
2599                         int dev = r10_bio->devs[m].devnum;
2600                         rdev = conf->mirrors[dev].rdev;
2601                         if (r10_bio->devs[m].bio == NULL)
2602                                 continue;
2603                         if (!r10_bio->devs[m].bio->bi_error) {
2604                                 rdev_clear_badblocks(
2605                                         rdev,
2606                                         r10_bio->devs[m].addr,
2607                                         r10_bio->sectors, 0);
2608                         } else {
2609                                 if (!rdev_set_badblocks(
2610                                             rdev,
2611                                             r10_bio->devs[m].addr,
2612                                             r10_bio->sectors, 0))
2613                                         md_error(conf->mddev, rdev);
2614                         }
2615                         rdev = conf->mirrors[dev].replacement;
2616                         if (r10_bio->devs[m].repl_bio == NULL)
2617                                 continue;
2618
2619                         if (!r10_bio->devs[m].repl_bio->bi_error) {
2620                                 rdev_clear_badblocks(
2621                                         rdev,
2622                                         r10_bio->devs[m].addr,
2623                                         r10_bio->sectors, 0);
2624                         } else {
2625                                 if (!rdev_set_badblocks(
2626                                             rdev,
2627                                             r10_bio->devs[m].addr,
2628                                             r10_bio->sectors, 0))
2629                                         md_error(conf->mddev, rdev);
2630                         }
2631                 }
2632                 put_buf(r10_bio);
2633         } else {
2634                 bool fail = false;
2635                 for (m = 0; m < conf->copies; m++) {
2636                         int dev = r10_bio->devs[m].devnum;
2637                         struct bio *bio = r10_bio->devs[m].bio;
2638                         rdev = conf->mirrors[dev].rdev;
2639                         if (bio == IO_MADE_GOOD) {
2640                                 rdev_clear_badblocks(
2641                                         rdev,
2642                                         r10_bio->devs[m].addr,
2643                                         r10_bio->sectors, 0);
2644                                 rdev_dec_pending(rdev, conf->mddev);
2645                         } else if (bio != NULL && bio->bi_error) {
2646                                 fail = true;
2647                                 if (!narrow_write_error(r10_bio, m)) {
2648                                         md_error(conf->mddev, rdev);
2649                                         set_bit(R10BIO_Degraded,
2650                                                 &r10_bio->state);
2651                                 }
2652                                 rdev_dec_pending(rdev, conf->mddev);
2653                         }
2654                         bio = r10_bio->devs[m].repl_bio;
2655                         rdev = conf->mirrors[dev].replacement;
2656                         if (rdev && bio == IO_MADE_GOOD) {
2657                                 rdev_clear_badblocks(
2658                                         rdev,
2659                                         r10_bio->devs[m].addr,
2660                                         r10_bio->sectors, 0);
2661                                 rdev_dec_pending(rdev, conf->mddev);
2662                         }
2663                 }
2664                 if (fail) {
2665                         spin_lock_irq(&conf->device_lock);
2666                         list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2667                         conf->nr_queued++;
2668                         spin_unlock_irq(&conf->device_lock);
2669                         md_wakeup_thread(conf->mddev->thread);
2670                 } else {
2671                         if (test_bit(R10BIO_WriteError,
2672                                      &r10_bio->state))
2673                                 close_write(r10_bio);
2674                         raid_end_bio_io(r10_bio);
2675                 }
2676         }
2677 }
2678
2679 static void raid10d(struct md_thread *thread)
2680 {
2681         struct mddev *mddev = thread->mddev;
2682         struct r10bio *r10_bio;
2683         unsigned long flags;
2684         struct r10conf *conf = mddev->private;
2685         struct list_head *head = &conf->retry_list;
2686         struct blk_plug plug;
2687
2688         md_check_recovery(mddev);
2689
2690         if (!list_empty_careful(&conf->bio_end_io_list) &&
2691             !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2692                 LIST_HEAD(tmp);
2693                 spin_lock_irqsave(&conf->device_lock, flags);
2694                 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2695                         while (!list_empty(&conf->bio_end_io_list)) {
2696                                 list_move(conf->bio_end_io_list.prev, &tmp);
2697                                 conf->nr_queued--;
2698                         }
2699                 }
2700                 spin_unlock_irqrestore(&conf->device_lock, flags);
2701                 while (!list_empty(&tmp)) {
2702                         r10_bio = list_first_entry(&tmp, struct r10bio,
2703                                                    retry_list);
2704                         list_del(&r10_bio->retry_list);
2705                         if (mddev->degraded)
2706                                 set_bit(R10BIO_Degraded, &r10_bio->state);
2707
2708                         if (test_bit(R10BIO_WriteError,
2709                                      &r10_bio->state))
2710                                 close_write(r10_bio);
2711                         raid_end_bio_io(r10_bio);
2712                 }
2713         }
2714
2715         blk_start_plug(&plug);
2716         for (;;) {
2717
2718                 flush_pending_writes(conf);
2719
2720                 spin_lock_irqsave(&conf->device_lock, flags);
2721                 if (list_empty(head)) {
2722                         spin_unlock_irqrestore(&conf->device_lock, flags);
2723                         break;
2724                 }
2725                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2726                 list_del(head->prev);
2727                 conf->nr_queued--;
2728                 spin_unlock_irqrestore(&conf->device_lock, flags);
2729
2730                 mddev = r10_bio->mddev;
2731                 conf = mddev->private;
2732                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2733                     test_bit(R10BIO_WriteError, &r10_bio->state))
2734                         handle_write_completed(conf, r10_bio);
2735                 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2736                         reshape_request_write(mddev, r10_bio);
2737                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2738                         sync_request_write(mddev, r10_bio);
2739                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2740                         recovery_request_write(mddev, r10_bio);
2741                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2742                         handle_read_error(mddev, r10_bio);
2743                 else {
2744                         /* just a partial read to be scheduled from a
2745                          * separate context
2746                          */
2747                         int slot = r10_bio->read_slot;
2748                         generic_make_request(r10_bio->devs[slot].bio);
2749                 }
2750
2751                 cond_resched();
2752                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2753                         md_check_recovery(mddev);
2754         }
2755         blk_finish_plug(&plug);
2756 }
2757
2758 static int init_resync(struct r10conf *conf)
2759 {
2760         int buffs;
2761         int i;
2762
2763         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2764         BUG_ON(conf->r10buf_pool);
2765         conf->have_replacement = 0;
2766         for (i = 0; i < conf->geo.raid_disks; i++)
2767                 if (conf->mirrors[i].replacement)
2768                         conf->have_replacement = 1;
2769         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2770         if (!conf->r10buf_pool)
2771                 return -ENOMEM;
2772         conf->next_resync = 0;
2773         return 0;
2774 }
2775
2776 /*
2777  * perform a "sync" on one "block"
2778  *
2779  * We need to make sure that no normal I/O request - particularly write
2780  * requests - conflict with active sync requests.
2781  *
2782  * This is achieved by tracking pending requests and a 'barrier' concept
2783  * that can be installed to exclude normal IO requests.
2784  *
2785  * Resync and recovery are handled very differently.
2786  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2787  *
2788  * For resync, we iterate over virtual addresses, read all copies,
2789  * and update if there are differences.  If only one copy is live,
2790  * skip it.
2791  * For recovery, we iterate over physical addresses, read a good
2792  * value for each non-in_sync drive, and over-write.
2793  *
2794  * So, for recovery we may have several outstanding complex requests for a
2795  * given address, one for each out-of-sync device.  We model this by allocating
2796  * a number of r10_bio structures, one for each out-of-sync device.
2797  * As we setup these structures, we collect all bio's together into a list
2798  * which we then process collectively to add pages, and then process again
2799  * to pass to generic_make_request.
2800  *
2801  * The r10_bio structures are linked using a borrowed master_bio pointer.
2802  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2803  * has its remaining count decremented to 0, the whole complex operation
2804  * is complete.
2805  *
2806  */
2807
2808 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2809                              int *skipped)
2810 {
2811         struct r10conf *conf = mddev->private;
2812         struct r10bio *r10_bio;
2813         struct bio *biolist = NULL, *bio;
2814         sector_t max_sector, nr_sectors;
2815         int i;
2816         int max_sync;
2817         sector_t sync_blocks;
2818         sector_t sectors_skipped = 0;
2819         int chunks_skipped = 0;
2820         sector_t chunk_mask = conf->geo.chunk_mask;
2821
2822         if (!conf->r10buf_pool)
2823                 if (init_resync(conf))
2824                         return 0;
2825
2826         /*
2827          * Allow skipping a full rebuild for incremental assembly
2828          * of a clean array, like RAID1 does.
2829          */
2830         if (mddev->bitmap == NULL &&
2831             mddev->recovery_cp == MaxSector &&
2832             mddev->reshape_position == MaxSector &&
2833             !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2834             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2835             !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2836             conf->fullsync == 0) {
2837                 *skipped = 1;
2838                 return mddev->dev_sectors - sector_nr;
2839         }
2840
2841  skipped:
2842         max_sector = mddev->dev_sectors;
2843         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2844             test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2845                 max_sector = mddev->resync_max_sectors;
2846         if (sector_nr >= max_sector) {
2847                 /* If we aborted, we need to abort the
2848                  * sync on the 'current' bitmap chucks (there can
2849                  * be several when recovering multiple devices).
2850                  * as we may have started syncing it but not finished.
2851                  * We can find the current address in
2852                  * mddev->curr_resync, but for recovery,
2853                  * we need to convert that to several
2854                  * virtual addresses.
2855                  */
2856                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2857                         end_reshape(conf);
2858                         close_sync(conf);
2859                         return 0;
2860                 }
2861
2862                 if (mddev->curr_resync < max_sector) { /* aborted */
2863                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2864                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2865                                                 &sync_blocks, 1);
2866                         else for (i = 0; i < conf->geo.raid_disks; i++) {
2867                                 sector_t sect =
2868                                         raid10_find_virt(conf, mddev->curr_resync, i);
2869                                 bitmap_end_sync(mddev->bitmap, sect,
2870                                                 &sync_blocks, 1);
2871                         }
2872                 } else {
2873                         /* completed sync */
2874                         if ((!mddev->bitmap || conf->fullsync)
2875                             && conf->have_replacement
2876                             && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2877                                 /* Completed a full sync so the replacements
2878                                  * are now fully recovered.
2879                                  */
2880                                 for (i = 0; i < conf->geo.raid_disks; i++)
2881                                         if (conf->mirrors[i].replacement)
2882                                                 conf->mirrors[i].replacement
2883                                                         ->recovery_offset
2884                                                         = MaxSector;
2885                         }
2886                         conf->fullsync = 0;
2887                 }
2888                 bitmap_close_sync(mddev->bitmap);
2889                 close_sync(conf);
2890                 *skipped = 1;
2891                 return sectors_skipped;
2892         }
2893
2894         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2895                 return reshape_request(mddev, sector_nr, skipped);
2896
2897         if (chunks_skipped >= conf->geo.raid_disks) {
2898                 /* if there has been nothing to do on any drive,
2899                  * then there is nothing to do at all..
2900                  */
2901                 *skipped = 1;
2902                 return (max_sector - sector_nr) + sectors_skipped;
2903         }
2904
2905         if (max_sector > mddev->resync_max)
2906                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2907
2908         /* make sure whole request will fit in a chunk - if chunks
2909          * are meaningful
2910          */
2911         if (conf->geo.near_copies < conf->geo.raid_disks &&
2912             max_sector > (sector_nr | chunk_mask))
2913                 max_sector = (sector_nr | chunk_mask) + 1;
2914
2915         /*
2916          * If there is non-resync activity waiting for a turn, then let it
2917          * though before starting on this new sync request.
2918          */
2919         if (conf->nr_waiting)
2920                 schedule_timeout_uninterruptible(1);
2921
2922         /* Again, very different code for resync and recovery.
2923          * Both must result in an r10bio with a list of bios that
2924          * have bi_end_io, bi_sector, bi_bdev set,
2925          * and bi_private set to the r10bio.
2926          * For recovery, we may actually create several r10bios
2927          * with 2 bios in each, that correspond to the bios in the main one.
2928          * In this case, the subordinate r10bios link back through a
2929          * borrowed master_bio pointer, and the counter in the master
2930          * includes a ref from each subordinate.
2931          */
2932         /* First, we decide what to do and set ->bi_end_io
2933          * To end_sync_read if we want to read, and
2934          * end_sync_write if we will want to write.
2935          */
2936
2937         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2938         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2939                 /* recovery... the complicated one */
2940                 int j;
2941                 r10_bio = NULL;
2942
2943                 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2944                         int still_degraded;
2945                         struct r10bio *rb2;
2946                         sector_t sect;
2947                         int must_sync;
2948                         int any_working;
2949                         struct raid10_info *mirror = &conf->mirrors[i];
2950
2951                         if ((mirror->rdev == NULL ||
2952                              test_bit(In_sync, &mirror->rdev->flags))
2953                             &&
2954                             (mirror->replacement == NULL ||
2955                              test_bit(Faulty,
2956                                       &mirror->replacement->flags)))
2957                                 continue;
2958
2959                         still_degraded = 0;
2960                         /* want to reconstruct this device */
2961                         rb2 = r10_bio;
2962                         sect = raid10_find_virt(conf, sector_nr, i);
2963                         if (sect >= mddev->resync_max_sectors) {
2964                                 /* last stripe is not complete - don't
2965                                  * try to recover this sector.
2966                                  */
2967                                 continue;
2968                         }
2969                         /* Unless we are doing a full sync, or a replacement
2970                          * we only need to recover the block if it is set in
2971                          * the bitmap
2972                          */
2973                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2974                                                       &sync_blocks, 1);
2975                         if (sync_blocks < max_sync)
2976                                 max_sync = sync_blocks;
2977                         if (!must_sync &&
2978                             mirror->replacement == NULL &&
2979                             !conf->fullsync) {
2980                                 /* yep, skip the sync_blocks here, but don't assume
2981                                  * that there will never be anything to do here
2982                                  */
2983                                 chunks_skipped = -1;
2984                                 continue;
2985                         }
2986
2987                         r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
2988                         r10_bio->state = 0;
2989                         raise_barrier(conf, rb2 != NULL);
2990                         atomic_set(&r10_bio->remaining, 0);
2991
2992                         r10_bio->master_bio = (struct bio*)rb2;
2993                         if (rb2)