raid10: improve random reads performance
[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         sectors = r10_bio->sectors;
711         best_slot = -1;
712         best_rdev = NULL;
713         best_dist = MaxSector;
714         best_good_sectors = 0;
715         do_balance = 1;
716         /*
717          * Check if we can balance. We can balance on the whole
718          * device if no resync is going on (recovery is ok), or below
719          * the resync window. We take the first readable disk when
720          * above the resync window.
721          */
722         if (conf->mddev->recovery_cp < MaxSector
723             && (this_sector + sectors >= conf->next_resync))
724                 do_balance = 0;
725
726         for (slot = 0; slot < conf->copies ; slot++) {
727                 sector_t first_bad;
728                 int bad_sectors;
729                 sector_t dev_sector;
730
731                 if (r10_bio->devs[slot].bio == IO_BLOCKED)
732                         continue;
733                 disk = r10_bio->devs[slot].devnum;
734                 rdev = rcu_dereference(conf->mirrors[disk].replacement);
735                 if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
736                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
737                         rdev = rcu_dereference(conf->mirrors[disk].rdev);
738                 if (rdev == NULL ||
739                     test_bit(Faulty, &rdev->flags))
740                         continue;
741                 if (!test_bit(In_sync, &rdev->flags) &&
742                     r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
743                         continue;
744
745                 dev_sector = r10_bio->devs[slot].addr;
746                 if (is_badblock(rdev, dev_sector, sectors,
747                                 &first_bad, &bad_sectors)) {
748                         if (best_dist < MaxSector)
749                                 /* Already have a better slot */
750                                 continue;
751                         if (first_bad <= dev_sector) {
752                                 /* Cannot read here.  If this is the
753                                  * 'primary' device, then we must not read
754                                  * beyond 'bad_sectors' from another device.
755                                  */
756                                 bad_sectors -= (dev_sector - first_bad);
757                                 if (!do_balance && sectors > bad_sectors)
758                                         sectors = bad_sectors;
759                                 if (best_good_sectors > sectors)
760                                         best_good_sectors = sectors;
761                         } else {
762                                 sector_t good_sectors =
763                                         first_bad - dev_sector;
764                                 if (good_sectors > best_good_sectors) {
765                                         best_good_sectors = good_sectors;
766                                         best_slot = slot;
767                                         best_rdev = rdev;
768                                 }
769                                 if (!do_balance)
770                                         /* Must read from here */
771                                         break;
772                         }
773                         continue;
774                 } else
775                         best_good_sectors = sectors;
776
777                 if (!do_balance)
778                         break;
779
780                 /* This optimisation is debatable, and completely destroys
781                  * sequential read speed for 'far copies' arrays.  So only
782                  * keep it for 'near' arrays, and review those later.
783                  */
784                 if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
785                         break;
786
787                 /* for far > 1 always use the lowest address */
788                 if (geo->far_copies > 1)
789                         new_distance = r10_bio->devs[slot].addr;
790                 else
791                         new_distance = abs(r10_bio->devs[slot].addr -
792                                            conf->mirrors[disk].head_position);
793                 if (new_distance < best_dist) {
794                         best_dist = new_distance;
795                         best_slot = slot;
796                         best_rdev = rdev;
797                 }
798         }
799         if (slot >= conf->copies) {
800                 slot = best_slot;
801                 rdev = best_rdev;
802         }
803
804         if (slot >= 0) {
805                 atomic_inc(&rdev->nr_pending);
806                 r10_bio->read_slot = slot;
807         } else
808                 rdev = NULL;
809         rcu_read_unlock();
810         *max_sectors = best_good_sectors;
811
812         return rdev;
813 }
814
815 static int raid10_congested(struct mddev *mddev, int bits)
816 {
817         struct r10conf *conf = mddev->private;
818         int i, ret = 0;
819
820         if ((bits & (1 << WB_async_congested)) &&
821             conf->pending_count >= max_queued_requests)
822                 return 1;
823
824         rcu_read_lock();
825         for (i = 0;
826              (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
827                      && ret == 0;
828              i++) {
829                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
830                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
831                         struct request_queue *q = bdev_get_queue(rdev->bdev);
832
833                         ret |= bdi_congested(&q->backing_dev_info, bits);
834                 }
835         }
836         rcu_read_unlock();
837         return ret;
838 }
839
840 static void flush_pending_writes(struct r10conf *conf)
841 {
842         /* Any writes that have been queued but are awaiting
843          * bitmap updates get flushed here.
844          */
845         spin_lock_irq(&conf->device_lock);
846
847         if (conf->pending_bio_list.head) {
848                 struct bio *bio;
849                 bio = bio_list_get(&conf->pending_bio_list);
850                 conf->pending_count = 0;
851                 spin_unlock_irq(&conf->device_lock);
852                 /* flush any pending bitmap writes to disk
853                  * before proceeding w/ I/O */
854                 bitmap_unplug(conf->mddev->bitmap);
855                 wake_up(&conf->wait_barrier);
856
857                 while (bio) { /* submit pending writes */
858                         struct bio *next = bio->bi_next;
859                         bio->bi_next = NULL;
860                         if (unlikely((bio->bi_rw & REQ_DISCARD) &&
861                             !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
862                                 /* Just ignore it */
863                                 bio_endio(bio);
864                         else
865                                 generic_make_request(bio);
866                         bio = next;
867                 }
868         } else
869                 spin_unlock_irq(&conf->device_lock);
870 }
871
872 /* Barriers....
873  * Sometimes we need to suspend IO while we do something else,
874  * either some resync/recovery, or reconfigure the array.
875  * To do this we raise a 'barrier'.
876  * The 'barrier' is a counter that can be raised multiple times
877  * to count how many activities are happening which preclude
878  * normal IO.
879  * We can only raise the barrier if there is no pending IO.
880  * i.e. if nr_pending == 0.
881  * We choose only to raise the barrier if no-one is waiting for the
882  * barrier to go down.  This means that as soon as an IO request
883  * is ready, no other operations which require a barrier will start
884  * until the IO request has had a chance.
885  *
886  * So: regular IO calls 'wait_barrier'.  When that returns there
887  *    is no backgroup IO happening,  It must arrange to call
888  *    allow_barrier when it has finished its IO.
889  * backgroup IO calls must call raise_barrier.  Once that returns
890  *    there is no normal IO happeing.  It must arrange to call
891  *    lower_barrier when the particular background IO completes.
892  */
893
894 static void raise_barrier(struct r10conf *conf, int force)
895 {
896         BUG_ON(force && !conf->barrier);
897         spin_lock_irq(&conf->resync_lock);
898
899         /* Wait until no block IO is waiting (unless 'force') */
900         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
901                             conf->resync_lock);
902
903         /* block any new IO from starting */
904         conf->barrier++;
905
906         /* Now wait for all pending IO to complete */
907         wait_event_lock_irq(conf->wait_barrier,
908                             !atomic_read(&conf->nr_pending) && conf->barrier < RESYNC_DEPTH,
909                             conf->resync_lock);
910
911         spin_unlock_irq(&conf->resync_lock);
912 }
913
914 static void lower_barrier(struct r10conf *conf)
915 {
916         unsigned long flags;
917         spin_lock_irqsave(&conf->resync_lock, flags);
918         conf->barrier--;
919         spin_unlock_irqrestore(&conf->resync_lock, flags);
920         wake_up(&conf->wait_barrier);
921 }
922
923 static void wait_barrier(struct r10conf *conf)
924 {
925         spin_lock_irq(&conf->resync_lock);
926         if (conf->barrier) {
927                 conf->nr_waiting++;
928                 /* Wait for the barrier to drop.
929                  * However if there are already pending
930                  * requests (preventing the barrier from
931                  * rising completely), and the
932                  * pre-process bio queue isn't empty,
933                  * then don't wait, as we need to empty
934                  * that queue to get the nr_pending
935                  * count down.
936                  */
937                 wait_event_lock_irq(conf->wait_barrier,
938                                     !conf->barrier ||
939                                     (atomic_read(&conf->nr_pending) &&
940                                      current->bio_list &&
941                                      !bio_list_empty(current->bio_list)),
942                                     conf->resync_lock);
943                 conf->nr_waiting--;
944                 if (!conf->nr_waiting)
945                         wake_up(&conf->wait_barrier);
946         }
947         atomic_inc(&conf->nr_pending);
948         spin_unlock_irq(&conf->resync_lock);
949 }
950
951 static void allow_barrier(struct r10conf *conf)
952 {
953         if ((atomic_dec_and_test(&conf->nr_pending)) ||
954                         (conf->array_freeze_pending))
955                 wake_up(&conf->wait_barrier);
956 }
957
958 static void freeze_array(struct r10conf *conf, int extra)
959 {
960         /* stop syncio and normal IO and wait for everything to
961          * go quiet.
962          * We increment barrier and nr_waiting, and then
963          * wait until nr_pending match nr_queued+extra
964          * This is called in the context of one normal IO request
965          * that has failed. Thus any sync request that might be pending
966          * will be blocked by nr_pending, and we need to wait for
967          * pending IO requests to complete or be queued for re-try.
968          * Thus the number queued (nr_queued) plus this request (extra)
969          * must match the number of pending IOs (nr_pending) before
970          * we continue.
971          */
972         spin_lock_irq(&conf->resync_lock);
973         conf->array_freeze_pending++;
974         conf->barrier++;
975         conf->nr_waiting++;
976         wait_event_lock_irq_cmd(conf->wait_barrier,
977                                 atomic_read(&conf->nr_pending) == conf->nr_queued+extra,
978                                 conf->resync_lock,
979                                 flush_pending_writes(conf));
980
981         conf->array_freeze_pending--;
982         spin_unlock_irq(&conf->resync_lock);
983 }
984
985 static void unfreeze_array(struct r10conf *conf)
986 {
987         /* reverse the effect of the freeze */
988         spin_lock_irq(&conf->resync_lock);
989         conf->barrier--;
990         conf->nr_waiting--;
991         wake_up(&conf->wait_barrier);
992         spin_unlock_irq(&conf->resync_lock);
993 }
994
995 static sector_t choose_data_offset(struct r10bio *r10_bio,
996                                    struct md_rdev *rdev)
997 {
998         if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
999             test_bit(R10BIO_Previous, &r10_bio->state))
1000                 return rdev->data_offset;
1001         else
1002                 return rdev->new_data_offset;
1003 }
1004
1005 struct raid10_plug_cb {
1006         struct blk_plug_cb      cb;
1007         struct bio_list         pending;
1008         int                     pending_cnt;
1009 };
1010
1011 static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1012 {
1013         struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1014                                                    cb);
1015         struct mddev *mddev = plug->cb.data;
1016         struct r10conf *conf = mddev->private;
1017         struct bio *bio;
1018
1019         if (from_schedule || current->bio_list) {
1020                 spin_lock_irq(&conf->device_lock);
1021                 bio_list_merge(&conf->pending_bio_list, &plug->pending);
1022                 conf->pending_count += plug->pending_cnt;
1023                 spin_unlock_irq(&conf->device_lock);
1024                 wake_up(&conf->wait_barrier);
1025                 md_wakeup_thread(mddev->thread);
1026                 kfree(plug);
1027                 return;
1028         }
1029
1030         /* we aren't scheduling, so we can do the write-out directly. */
1031         bio = bio_list_get(&plug->pending);
1032         bitmap_unplug(mddev->bitmap);
1033         wake_up(&conf->wait_barrier);
1034
1035         while (bio) { /* submit pending writes */
1036                 struct bio *next = bio->bi_next;
1037                 bio->bi_next = NULL;
1038                 if (unlikely((bio->bi_rw & REQ_DISCARD) &&
1039                     !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
1040                         /* Just ignore it */
1041                         bio_endio(bio);
1042                 else
1043                         generic_make_request(bio);
1044                 bio = next;
1045         }
1046         kfree(plug);
1047 }
1048
1049 static void __make_request(struct mddev *mddev, struct bio *bio)
1050 {
1051         struct r10conf *conf = mddev->private;
1052         struct r10bio *r10_bio;
1053         struct bio *read_bio;
1054         int i;
1055         const int rw = bio_data_dir(bio);
1056         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1057         const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1058         const unsigned long do_discard = (bio->bi_rw
1059                                           & (REQ_DISCARD | REQ_SECURE));
1060         const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1061         unsigned long flags;
1062         struct md_rdev *blocked_rdev;
1063         struct blk_plug_cb *cb;
1064         struct raid10_plug_cb *plug = NULL;
1065         int sectors_handled;
1066         int max_sectors;
1067         int sectors;
1068
1069         /*
1070          * Register the new request and wait if the reconstruction
1071          * thread has put up a bar for new requests.
1072          * Continue immediately if no resync is active currently.
1073          */
1074         wait_barrier(conf);
1075
1076         sectors = bio_sectors(bio);
1077         while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1078             bio->bi_iter.bi_sector < conf->reshape_progress &&
1079             bio->bi_iter.bi_sector + sectors > conf->reshape_progress) {
1080                 /* IO spans the reshape position.  Need to wait for
1081                  * reshape to pass
1082                  */
1083                 allow_barrier(conf);
1084                 wait_event(conf->wait_barrier,
1085                            conf->reshape_progress <= bio->bi_iter.bi_sector ||
1086                            conf->reshape_progress >= bio->bi_iter.bi_sector +
1087                            sectors);
1088                 wait_barrier(conf);
1089         }
1090         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1091             bio_data_dir(bio) == WRITE &&
1092             (mddev->reshape_backwards
1093              ? (bio->bi_iter.bi_sector < conf->reshape_safe &&
1094                 bio->bi_iter.bi_sector + sectors > conf->reshape_progress)
1095              : (bio->bi_iter.bi_sector + sectors > conf->reshape_safe &&
1096                 bio->bi_iter.bi_sector < conf->reshape_progress))) {
1097                 /* Need to update reshape_position in metadata */
1098                 mddev->reshape_position = conf->reshape_progress;
1099                 set_mask_bits(&mddev->flags, 0,
1100                               BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1101                 md_wakeup_thread(mddev->thread);
1102                 wait_event(mddev->sb_wait,
1103                            !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1104
1105                 conf->reshape_safe = mddev->reshape_position;
1106         }
1107
1108         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1109
1110         r10_bio->master_bio = bio;
1111         r10_bio->sectors = sectors;
1112
1113         r10_bio->mddev = mddev;
1114         r10_bio->sector = bio->bi_iter.bi_sector;
1115         r10_bio->state = 0;
1116
1117         /* We might need to issue multiple reads to different
1118          * devices if there are bad blocks around, so we keep
1119          * track of the number of reads in bio->bi_phys_segments.
1120          * If this is 0, there is only one r10_bio and no locking
1121          * will be needed when the request completes.  If it is
1122          * non-zero, then it is the number of not-completed requests.
1123          */
1124         bio->bi_phys_segments = 0;
1125         bio_clear_flag(bio, BIO_SEG_VALID);
1126
1127         if (rw == READ) {
1128                 /*
1129                  * read balancing logic:
1130                  */
1131                 struct md_rdev *rdev;
1132                 int slot;
1133
1134 read_again:
1135                 rdev = read_balance(conf, r10_bio, &max_sectors);
1136                 if (!rdev) {
1137                         raid_end_bio_io(r10_bio);
1138                         return;
1139                 }
1140                 slot = r10_bio->read_slot;
1141
1142                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1143                 bio_trim(read_bio, r10_bio->sector - bio->bi_iter.bi_sector,
1144                          max_sectors);
1145
1146                 r10_bio->devs[slot].bio = read_bio;
1147                 r10_bio->devs[slot].rdev = rdev;
1148
1149                 read_bio->bi_iter.bi_sector = r10_bio->devs[slot].addr +
1150                         choose_data_offset(r10_bio, rdev);
1151                 read_bio->bi_bdev = rdev->bdev;
1152                 read_bio->bi_end_io = raid10_end_read_request;
1153                 read_bio->bi_rw = READ | do_sync;
1154                 read_bio->bi_private = r10_bio;
1155
1156                 if (max_sectors < r10_bio->sectors) {
1157                         /* Could not read all from this device, so we will
1158                          * need another r10_bio.
1159                          */
1160                         sectors_handled = (r10_bio->sector + max_sectors
1161                                            - bio->bi_iter.bi_sector);
1162                         r10_bio->sectors = max_sectors;
1163                         spin_lock_irq(&conf->device_lock);
1164                         if (bio->bi_phys_segments == 0)
1165                                 bio->bi_phys_segments = 2;
1166                         else
1167                                 bio->bi_phys_segments++;
1168                         spin_unlock_irq(&conf->device_lock);
1169                         /* Cannot call generic_make_request directly
1170                          * as that will be queued in __generic_make_request
1171                          * and subsequent mempool_alloc might block
1172                          * waiting for it.  so hand bio over to raid10d.
1173                          */
1174                         reschedule_retry(r10_bio);
1175
1176                         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1177
1178                         r10_bio->master_bio = bio;
1179                         r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1180                         r10_bio->state = 0;
1181                         r10_bio->mddev = mddev;
1182                         r10_bio->sector = bio->bi_iter.bi_sector +
1183                                 sectors_handled;
1184                         goto read_again;
1185                 } else
1186                         generic_make_request(read_bio);
1187                 return;
1188         }
1189
1190         /*
1191          * WRITE:
1192          */
1193         if (conf->pending_count >= max_queued_requests) {
1194                 md_wakeup_thread(mddev->thread);
1195                 wait_event(conf->wait_barrier,
1196                            conf->pending_count < max_queued_requests);
1197         }
1198         /* first select target devices under rcu_lock and
1199          * inc refcount on their rdev.  Record them by setting
1200          * bios[x] to bio
1201          * If there are known/acknowledged bad blocks on any device
1202          * on which we have seen a write error, we want to avoid
1203          * writing to those blocks.  This potentially requires several
1204          * writes to write around the bad blocks.  Each set of writes
1205          * gets its own r10_bio with a set of bios attached.  The number
1206          * of r10_bios is recored in bio->bi_phys_segments just as with
1207          * the read case.
1208          */
1209
1210         r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1211         raid10_find_phys(conf, r10_bio);
1212 retry_write:
1213         blocked_rdev = NULL;
1214         rcu_read_lock();
1215         max_sectors = r10_bio->sectors;
1216
1217         for (i = 0;  i < conf->copies; i++) {
1218                 int d = r10_bio->devs[i].devnum;
1219                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1220                 struct md_rdev *rrdev = rcu_dereference(
1221                         conf->mirrors[d].replacement);
1222                 if (rdev == rrdev)
1223                         rrdev = NULL;
1224                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1225                         atomic_inc(&rdev->nr_pending);
1226                         blocked_rdev = rdev;
1227                         break;
1228                 }
1229                 if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1230                         atomic_inc(&rrdev->nr_pending);
1231                         blocked_rdev = rrdev;
1232                         break;
1233                 }
1234                 if (rdev && (test_bit(Faulty, &rdev->flags)))
1235                         rdev = NULL;
1236                 if (rrdev && (test_bit(Faulty, &rrdev->flags)))
1237                         rrdev = NULL;
1238
1239                 r10_bio->devs[i].bio = NULL;
1240                 r10_bio->devs[i].repl_bio = NULL;
1241
1242                 if (!rdev && !rrdev) {
1243                         set_bit(R10BIO_Degraded, &r10_bio->state);
1244                         continue;
1245                 }
1246                 if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1247                         sector_t first_bad;
1248                         sector_t dev_sector = r10_bio->devs[i].addr;
1249                         int bad_sectors;
1250                         int is_bad;
1251
1252                         is_bad = is_badblock(rdev, dev_sector,
1253                                              max_sectors,
1254                                              &first_bad, &bad_sectors);
1255                         if (is_bad < 0) {
1256                                 /* Mustn't write here until the bad block
1257                                  * is acknowledged
1258                                  */
1259                                 atomic_inc(&rdev->nr_pending);
1260                                 set_bit(BlockedBadBlocks, &rdev->flags);
1261                                 blocked_rdev = rdev;
1262                                 break;
1263                         }
1264                         if (is_bad && first_bad <= dev_sector) {
1265                                 /* Cannot write here at all */
1266                                 bad_sectors -= (dev_sector - first_bad);
1267                                 if (bad_sectors < max_sectors)
1268                                         /* Mustn't write more than bad_sectors
1269                                          * to other devices yet
1270                                          */
1271                                         max_sectors = bad_sectors;
1272                                 /* We don't set R10BIO_Degraded as that
1273                                  * only applies if the disk is missing,
1274                                  * so it might be re-added, and we want to
1275                                  * know to recover this chunk.
1276                                  * In this case the device is here, and the
1277                                  * fact that this chunk is not in-sync is
1278                                  * recorded in the bad block log.
1279                                  */
1280                                 continue;
1281                         }
1282                         if (is_bad) {
1283                                 int good_sectors = first_bad - dev_sector;
1284                                 if (good_sectors < max_sectors)
1285                                         max_sectors = good_sectors;
1286                         }
1287                 }
1288                 if (rdev) {
1289                         r10_bio->devs[i].bio = bio;
1290                         atomic_inc(&rdev->nr_pending);
1291                 }
1292                 if (rrdev) {
1293                         r10_bio->devs[i].repl_bio = bio;
1294                         atomic_inc(&rrdev->nr_pending);
1295                 }
1296         }
1297         rcu_read_unlock();
1298
1299         if (unlikely(blocked_rdev)) {
1300                 /* Have to wait for this device to get unblocked, then retry */
1301                 int j;
1302                 int d;
1303
1304                 for (j = 0; j < i; j++) {
1305                         if (r10_bio->devs[j].bio) {
1306                                 d = r10_bio->devs[j].devnum;
1307                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1308                         }
1309                         if (r10_bio->devs[j].repl_bio) {
1310                                 struct md_rdev *rdev;
1311                                 d = r10_bio->devs[j].devnum;
1312                                 rdev = conf->mirrors[d].replacement;
1313                                 if (!rdev) {
1314                                         /* Race with remove_disk */
1315                                         smp_mb();
1316                                         rdev = conf->mirrors[d].rdev;
1317                                 }
1318                                 rdev_dec_pending(rdev, mddev);
1319                         }
1320                 }
1321                 allow_barrier(conf);
1322                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
1323                 wait_barrier(conf);
1324                 goto retry_write;
1325         }
1326
1327         if (max_sectors < r10_bio->sectors) {
1328                 /* We are splitting this into multiple parts, so
1329                  * we need to prepare for allocating another r10_bio.
1330                  */
1331                 r10_bio->sectors = max_sectors;
1332                 spin_lock_irq(&conf->device_lock);
1333                 if (bio->bi_phys_segments == 0)
1334                         bio->bi_phys_segments = 2;
1335                 else
1336                         bio->bi_phys_segments++;
1337                 spin_unlock_irq(&conf->device_lock);
1338         }
1339         sectors_handled = r10_bio->sector + max_sectors -
1340                 bio->bi_iter.bi_sector;
1341
1342         atomic_set(&r10_bio->remaining, 1);
1343         bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1344
1345         for (i = 0; i < conf->copies; i++) {
1346                 struct bio *mbio;
1347                 int d = r10_bio->devs[i].devnum;
1348                 if (r10_bio->devs[i].bio) {
1349                         struct md_rdev *rdev = conf->mirrors[d].rdev;
1350                         mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1351                         bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1352                                  max_sectors);
1353                         r10_bio->devs[i].bio = mbio;
1354
1355                         mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
1356                                            choose_data_offset(r10_bio,
1357                                                               rdev));
1358                         mbio->bi_bdev = rdev->bdev;
1359                         mbio->bi_end_io = raid10_end_write_request;
1360                         mbio->bi_rw =
1361                                 WRITE | do_sync | do_fua | do_discard | do_same;
1362                         mbio->bi_private = r10_bio;
1363
1364                         atomic_inc(&r10_bio->remaining);
1365
1366                         cb = blk_check_plugged(raid10_unplug, mddev,
1367                                                sizeof(*plug));
1368                         if (cb)
1369                                 plug = container_of(cb, struct raid10_plug_cb,
1370                                                     cb);
1371                         else
1372                                 plug = NULL;
1373                         spin_lock_irqsave(&conf->device_lock, flags);
1374                         if (plug) {
1375                                 bio_list_add(&plug->pending, mbio);
1376                                 plug->pending_cnt++;
1377                         } else {
1378                                 bio_list_add(&conf->pending_bio_list, mbio);
1379                                 conf->pending_count++;
1380                         }
1381                         spin_unlock_irqrestore(&conf->device_lock, flags);
1382                         if (!plug)
1383                                 md_wakeup_thread(mddev->thread);
1384                 }
1385
1386                 if (r10_bio->devs[i].repl_bio) {
1387                         struct md_rdev *rdev = conf->mirrors[d].replacement;
1388                         if (rdev == NULL) {
1389                                 /* Replacement just got moved to main 'rdev' */
1390                                 smp_mb();
1391                                 rdev = conf->mirrors[d].rdev;
1392                         }
1393                         mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1394                         bio_trim(mbio, r10_bio->sector - bio->bi_iter.bi_sector,
1395                                  max_sectors);
1396                         r10_bio->devs[i].repl_bio = mbio;
1397
1398                         mbio->bi_iter.bi_sector = (r10_bio->devs[i].addr +
1399                                            choose_data_offset(
1400                                                    r10_bio, rdev));
1401                         mbio->bi_bdev = rdev->bdev;
1402                         mbio->bi_end_io = raid10_end_write_request;
1403                         mbio->bi_rw =
1404                                 WRITE | do_sync | do_fua | do_discard | do_same;
1405                         mbio->bi_private = r10_bio;
1406
1407                         atomic_inc(&r10_bio->remaining);
1408                         spin_lock_irqsave(&conf->device_lock, flags);
1409                         bio_list_add(&conf->pending_bio_list, mbio);
1410                         conf->pending_count++;
1411                         spin_unlock_irqrestore(&conf->device_lock, flags);
1412                         if (!mddev_check_plugged(mddev))
1413                                 md_wakeup_thread(mddev->thread);
1414                 }
1415         }
1416
1417         /* Don't remove the bias on 'remaining' (one_write_done) until
1418          * after checking if we need to go around again.
1419          */
1420
1421         if (sectors_handled < bio_sectors(bio)) {
1422                 one_write_done(r10_bio);
1423                 /* We need another r10_bio.  It has already been counted
1424                  * in bio->bi_phys_segments.
1425                  */
1426                 r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1427
1428                 r10_bio->master_bio = bio;
1429                 r10_bio->sectors = bio_sectors(bio) - sectors_handled;
1430
1431                 r10_bio->mddev = mddev;
1432                 r10_bio->sector = bio->bi_iter.bi_sector + sectors_handled;
1433                 r10_bio->state = 0;
1434                 goto retry_write;
1435         }
1436         one_write_done(r10_bio);
1437 }
1438
1439 static void raid10_make_request(struct mddev *mddev, struct bio *bio)
1440 {
1441         struct r10conf *conf = mddev->private;
1442         sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1443         int chunk_sects = chunk_mask + 1;
1444
1445         struct bio *split;
1446
1447         if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1448                 md_flush_request(mddev, bio);
1449                 return;
1450         }
1451
1452         md_write_start(mddev, bio);
1453
1454         do {
1455
1456                 /*
1457                  * If this request crosses a chunk boundary, we need to split
1458                  * it.
1459                  */
1460                 if (unlikely((bio->bi_iter.bi_sector & chunk_mask) +
1461                              bio_sectors(bio) > chunk_sects
1462                              && (conf->geo.near_copies < conf->geo.raid_disks
1463                                  || conf->prev.near_copies <
1464                                  conf->prev.raid_disks))) {
1465                         split = bio_split(bio, chunk_sects -
1466                                           (bio->bi_iter.bi_sector &
1467                                            (chunk_sects - 1)),
1468                                           GFP_NOIO, fs_bio_set);
1469                         bio_chain(split, bio);
1470                 } else {
1471                         split = bio;
1472                 }
1473
1474                 __make_request(mddev, split);
1475         } while (split != bio);
1476
1477         /* In case raid10d snuck in to freeze_array */
1478         wake_up(&conf->wait_barrier);
1479 }
1480
1481 static void raid10_status(struct seq_file *seq, struct mddev *mddev)
1482 {
1483         struct r10conf *conf = mddev->private;
1484         int i;
1485
1486         if (conf->geo.near_copies < conf->geo.raid_disks)
1487                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1488         if (conf->geo.near_copies > 1)
1489                 seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1490         if (conf->geo.far_copies > 1) {
1491                 if (conf->geo.far_offset)
1492                         seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1493                 else
1494                         seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1495                 if (conf->geo.far_set_size != conf->geo.raid_disks)
1496                         seq_printf(seq, " %d devices per set", conf->geo.far_set_size);
1497         }
1498         seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1499                                         conf->geo.raid_disks - mddev->degraded);
1500         rcu_read_lock();
1501         for (i = 0; i < conf->geo.raid_disks; i++) {
1502                 struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1503                 seq_printf(seq, "%s", rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1504         }
1505         rcu_read_unlock();
1506         seq_printf(seq, "]");
1507 }
1508
1509 /* check if there are enough drives for
1510  * every block to appear on atleast one.
1511  * Don't consider the device numbered 'ignore'
1512  * as we might be about to remove it.
1513  */
1514 static int _enough(struct r10conf *conf, int previous, int ignore)
1515 {
1516         int first = 0;
1517         int has_enough = 0;
1518         int disks, ncopies;
1519         if (previous) {
1520                 disks = conf->prev.raid_disks;
1521                 ncopies = conf->prev.near_copies;
1522         } else {
1523                 disks = conf->geo.raid_disks;
1524                 ncopies = conf->geo.near_copies;
1525         }
1526
1527         rcu_read_lock();
1528         do {
1529                 int n = conf->copies;
1530                 int cnt = 0;
1531                 int this = first;
1532                 while (n--) {
1533                         struct md_rdev *rdev;
1534                         if (this != ignore &&
1535                             (rdev = rcu_dereference(conf->mirrors[this].rdev)) &&
1536                             test_bit(In_sync, &rdev->flags))
1537                                 cnt++;
1538                         this = (this+1) % disks;
1539                 }
1540                 if (cnt == 0)
1541                         goto out;
1542                 first = (first + ncopies) % disks;
1543         } while (first != 0);
1544         has_enough = 1;
1545 out:
1546         rcu_read_unlock();
1547         return has_enough;
1548 }
1549
1550 static int enough(struct r10conf *conf, int ignore)
1551 {
1552         /* when calling 'enough', both 'prev' and 'geo' must
1553          * be stable.
1554          * This is ensured if ->reconfig_mutex or ->device_lock
1555          * is held.
1556          */
1557         return _enough(conf, 0, ignore) &&
1558                 _enough(conf, 1, ignore);
1559 }
1560
1561 static void raid10_error(struct mddev *mddev, struct md_rdev *rdev)
1562 {
1563         char b[BDEVNAME_SIZE];
1564         struct r10conf *conf = mddev->private;
1565         unsigned long flags;
1566
1567         /*
1568          * If it is not operational, then we have already marked it as dead
1569          * else if it is the last working disks, ignore the error, let the
1570          * next level up know.
1571          * else mark the drive as failed
1572          */
1573         spin_lock_irqsave(&conf->device_lock, flags);
1574         if (test_bit(In_sync, &rdev->flags)
1575             && !enough(conf, rdev->raid_disk)) {
1576                 /*
1577                  * Don't fail the drive, just return an IO error.
1578                  */
1579                 spin_unlock_irqrestore(&conf->device_lock, flags);
1580                 return;
1581         }
1582         if (test_and_clear_bit(In_sync, &rdev->flags))
1583                 mddev->degraded++;
1584         /*
1585          * If recovery is running, make sure it aborts.
1586          */
1587         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1588         set_bit(Blocked, &rdev->flags);
1589         set_bit(Faulty, &rdev->flags);
1590         set_mask_bits(&mddev->flags, 0,
1591                       BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING));
1592         spin_unlock_irqrestore(&conf->device_lock, flags);
1593         printk(KERN_ALERT
1594                "md/raid10:%s: Disk failure on %s, disabling device.\n"
1595                "md/raid10:%s: Operation continuing on %d devices.\n",
1596                mdname(mddev), bdevname(rdev->bdev, b),
1597                mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1598 }
1599
1600 static void print_conf(struct r10conf *conf)
1601 {
1602         int i;
1603         struct md_rdev *rdev;
1604
1605         printk(KERN_DEBUG "RAID10 conf printout:\n");
1606         if (!conf) {
1607                 printk(KERN_DEBUG "(!conf)\n");
1608                 return;
1609         }
1610         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1611                 conf->geo.raid_disks);
1612
1613         /* This is only called with ->reconfix_mutex held, so
1614          * rcu protection of rdev is not needed */
1615         for (i = 0; i < conf->geo.raid_disks; i++) {
1616                 char b[BDEVNAME_SIZE];
1617                 rdev = conf->mirrors[i].rdev;
1618                 if (rdev)
1619                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1620                                 i, !test_bit(In_sync, &rdev->flags),
1621                                 !test_bit(Faulty, &rdev->flags),
1622                                 bdevname(rdev->bdev,b));
1623         }
1624 }
1625
1626 static void close_sync(struct r10conf *conf)
1627 {
1628         wait_barrier(conf);
1629         allow_barrier(conf);
1630
1631         mempool_destroy(conf->r10buf_pool);
1632         conf->r10buf_pool = NULL;
1633 }
1634
1635 static int raid10_spare_active(struct mddev *mddev)
1636 {
1637         int i;
1638         struct r10conf *conf = mddev->private;
1639         struct raid10_info *tmp;
1640         int count = 0;
1641         unsigned long flags;
1642
1643         /*
1644          * Find all non-in_sync disks within the RAID10 configuration
1645          * and mark them in_sync
1646          */
1647         for (i = 0; i < conf->geo.raid_disks; i++) {
1648                 tmp = conf->mirrors + i;
1649                 if (tmp->replacement
1650                     && tmp->replacement->recovery_offset == MaxSector
1651                     && !test_bit(Faulty, &tmp->replacement->flags)
1652                     && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1653                         /* Replacement has just become active */
1654                         if (!tmp->rdev
1655                             || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1656                                 count++;
1657                         if (tmp->rdev) {
1658                                 /* Replaced device not technically faulty,
1659                                  * but we need to be sure it gets removed
1660                                  * and never re-added.
1661                                  */
1662                                 set_bit(Faulty, &tmp->rdev->flags);
1663                                 sysfs_notify_dirent_safe(
1664                                         tmp->rdev->sysfs_state);
1665                         }
1666                         sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1667                 } else if (tmp->rdev
1668                            && tmp->rdev->recovery_offset == MaxSector
1669                            && !test_bit(Faulty, &tmp->rdev->flags)
1670                            && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1671                         count++;
1672                         sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1673                 }
1674         }
1675         spin_lock_irqsave(&conf->device_lock, flags);
1676         mddev->degraded -= count;
1677         spin_unlock_irqrestore(&conf->device_lock, flags);
1678
1679         print_conf(conf);
1680         return count;
1681 }
1682
1683 static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1684 {
1685         struct r10conf *conf = mddev->private;
1686         int err = -EEXIST;
1687         int mirror;
1688         int first = 0;
1689         int last = conf->geo.raid_disks - 1;
1690
1691         if (mddev->recovery_cp < MaxSector)
1692                 /* only hot-add to in-sync arrays, as recovery is
1693                  * very different from resync
1694                  */
1695                 return -EBUSY;
1696         if (rdev->saved_raid_disk < 0 && !_enough(conf, 1, -1))
1697                 return -EINVAL;
1698
1699         if (md_integrity_add_rdev(rdev, mddev))
1700                 return -ENXIO;
1701
1702         if (rdev->raid_disk >= 0)
1703                 first = last = rdev->raid_disk;
1704
1705         if (rdev->saved_raid_disk >= first &&
1706             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1707                 mirror = rdev->saved_raid_disk;
1708         else
1709                 mirror = first;
1710         for ( ; mirror <= last ; mirror++) {
1711                 struct raid10_info *p = &conf->mirrors[mirror];
1712                 if (p->recovery_disabled == mddev->recovery_disabled)
1713                         continue;
1714                 if (p->rdev) {
1715                         if (!test_bit(WantReplacement, &p->rdev->flags) ||
1716                             p->replacement != NULL)
1717                                 continue;
1718                         clear_bit(In_sync, &rdev->flags);
1719                         set_bit(Replacement, &rdev->flags);
1720                         rdev->raid_disk = mirror;
1721                         err = 0;
1722                         if (mddev->gendisk)
1723                                 disk_stack_limits(mddev->gendisk, rdev->bdev,
1724                                                   rdev->data_offset << 9);
1725                         conf->fullsync = 1;
1726                         rcu_assign_pointer(p->replacement, rdev);
1727                         break;
1728                 }
1729
1730                 if (mddev->gendisk)
1731                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1732                                           rdev->data_offset << 9);
1733
1734                 p->head_position = 0;
1735                 p->recovery_disabled = mddev->recovery_disabled - 1;
1736                 rdev->raid_disk = mirror;
1737                 err = 0;
1738                 if (rdev->saved_raid_disk != mirror)
1739                         conf->fullsync = 1;
1740                 rcu_assign_pointer(p->rdev, rdev);
1741                 break;
1742         }
1743         if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1744                 queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1745
1746         print_conf(conf);
1747         return err;
1748 }
1749
1750 static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1751 {
1752         struct r10conf *conf = mddev->private;
1753         int err = 0;
1754         int number = rdev->raid_disk;
1755         struct md_rdev **rdevp;
1756         struct raid10_info *p = conf->mirrors + number;
1757
1758         print_conf(conf);
1759         if (rdev == p->rdev)
1760                 rdevp = &p->rdev;
1761         else if (rdev == p->replacement)
1762                 rdevp = &p->replacement;
1763         else
1764                 return 0;
1765
1766         if (test_bit(In_sync, &rdev->flags) ||
1767             atomic_read(&rdev->nr_pending)) {
1768                 err = -EBUSY;
1769                 goto abort;
1770         }
1771         /* Only remove non-faulty devices if recovery
1772          * is not possible.
1773          */
1774         if (!test_bit(Faulty, &rdev->flags) &&
1775             mddev->recovery_disabled != p->recovery_disabled &&
1776             (!p->replacement || p->replacement == rdev) &&
1777             number < conf->geo.raid_disks &&
1778             enough(conf, -1)) {
1779                 err = -EBUSY;
1780                 goto abort;
1781         }
1782         *rdevp = NULL;
1783         if (!test_bit(RemoveSynchronized, &rdev->flags)) {
1784                 synchronize_rcu();
1785                 if (atomic_read(&rdev->nr_pending)) {
1786                         /* lost the race, try later */
1787                         err = -EBUSY;
1788                         *rdevp = rdev;
1789                         goto abort;
1790                 }
1791         }
1792         if (p->replacement) {
1793                 /* We must have just cleared 'rdev' */
1794                 p->rdev = p->replacement;
1795                 clear_bit(Replacement, &p->replacement->flags);
1796                 smp_mb(); /* Make sure other CPUs may see both as identical
1797                            * but will never see neither -- if they are careful.
1798                            */
1799                 p->replacement = NULL;
1800                 clear_bit(WantReplacement, &rdev->flags);
1801         } else
1802                 /* We might have just remove the Replacement as faulty
1803                  * Clear the flag just in case
1804                  */
1805                 clear_bit(WantReplacement, &rdev->flags);
1806
1807         err = md_integrity_register(mddev);
1808
1809 abort:
1810
1811         print_conf(conf);
1812         return err;
1813 }
1814
1815 static void end_sync_read(struct bio *bio)
1816 {
1817         struct r10bio *r10_bio = bio->bi_private;
1818         struct r10conf *conf = r10_bio->mddev->private;
1819         int d;
1820
1821         if (bio == r10_bio->master_bio) {
1822                 /* this is a reshape read */
1823                 d = r10_bio->read_slot; /* really the read dev */
1824         } else
1825                 d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1826
1827         if (!bio->bi_error)
1828                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1829         else
1830                 /* The write handler will notice the lack of
1831                  * R10BIO_Uptodate and record any errors etc
1832                  */
1833                 atomic_add(r10_bio->sectors,
1834                            &conf->mirrors[d].rdev->corrected_errors);
1835
1836         /* for reconstruct, we always reschedule after a read.
1837          * for resync, only after all reads
1838          */
1839         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1840         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1841             atomic_dec_and_test(&r10_bio->remaining)) {
1842                 /* we have read all the blocks,
1843                  * do the comparison in process context in raid10d
1844                  */
1845                 reschedule_retry(r10_bio);
1846         }
1847 }
1848
1849 static void end_sync_request(struct r10bio *r10_bio)
1850 {
1851         struct mddev *mddev = r10_bio->mddev;
1852
1853         while (atomic_dec_and_test(&r10_bio->remaining)) {
1854                 if (r10_bio->master_bio == NULL) {
1855                         /* the primary of several recovery bios */
1856                         sector_t s = r10_bio->sectors;
1857                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1858                             test_bit(R10BIO_WriteError, &r10_bio->state))
1859                                 reschedule_retry(r10_bio);
1860                         else
1861                                 put_buf(r10_bio);
1862                         md_done_sync(mddev, s, 1);
1863                         break;
1864                 } else {
1865                         struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1866                         if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1867                             test_bit(R10BIO_WriteError, &r10_bio->state))
1868                                 reschedule_retry(r10_bio);
1869                         else
1870                                 put_buf(r10_bio);
1871                         r10_bio = r10_bio2;
1872                 }
1873         }
1874 }
1875
1876 static void end_sync_write(struct bio *bio)
1877 {
1878         struct r10bio *r10_bio = bio->bi_private;
1879         struct mddev *mddev = r10_bio->mddev;
1880         struct r10conf *conf = mddev->private;
1881         int d;
1882         sector_t first_bad;
1883         int bad_sectors;
1884         int slot;
1885         int repl;
1886         struct md_rdev *rdev = NULL;
1887
1888         d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1889         if (repl)
1890                 rdev = conf->mirrors[d].replacement;
1891         else
1892                 rdev = conf->mirrors[d].rdev;
1893
1894         if (bio->bi_error) {
1895                 if (repl)
1896                         md_error(mddev, rdev);
1897                 else {
1898                         set_bit(WriteErrorSeen, &rdev->flags);
1899                         if (!test_and_set_bit(WantReplacement, &rdev->flags))
1900                                 set_bit(MD_RECOVERY_NEEDED,
1901                                         &rdev->mddev->recovery);
1902                         set_bit(R10BIO_WriteError, &r10_bio->state);
1903                 }
1904         } else if (is_badblock(rdev,
1905                              r10_bio->devs[slot].addr,
1906                              r10_bio->sectors,
1907                              &first_bad, &bad_sectors))
1908                 set_bit(R10BIO_MadeGood, &r10_bio->state);
1909
1910         rdev_dec_pending(rdev, mddev);
1911
1912         end_sync_request(r10_bio);
1913 }
1914
1915 /*
1916  * Note: sync and recover and handled very differently for raid10
1917  * This code is for resync.
1918  * For resync, we read through virtual addresses and read all blocks.
1919  * If there is any error, we schedule a write.  The lowest numbered
1920  * drive is authoritative.
1921  * However requests come for physical address, so we need to map.
1922  * For every physical address there are raid_disks/copies virtual addresses,
1923  * which is always are least one, but is not necessarly an integer.
1924  * This means that a physical address can span multiple chunks, so we may
1925  * have to submit multiple io requests for a single sync request.
1926  */
1927 /*
1928  * We check if all blocks are in-sync and only write to blocks that
1929  * aren't in sync
1930  */
1931 static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
1932 {
1933         struct r10conf *conf = mddev->private;
1934         int i, first;
1935         struct bio *tbio, *fbio;
1936         int vcnt;
1937
1938         atomic_set(&r10_bio->remaining, 1);
1939
1940         /* find the first device with a block */
1941         for (i=0; i<conf->copies; i++)
1942                 if (!r10_bio->devs[i].bio->bi_error)
1943                         break;
1944
1945         if (i == conf->copies)
1946                 goto done;
1947
1948         first = i;
1949         fbio = r10_bio->devs[i].bio;
1950         fbio->bi_iter.bi_size = r10_bio->sectors << 9;
1951         fbio->bi_iter.bi_idx = 0;
1952
1953         vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
1954         /* now find blocks with errors */
1955         for (i=0 ; i < conf->copies ; i++) {
1956                 int  j, d;
1957
1958                 tbio = r10_bio->devs[i].bio;
1959
1960                 if (tbio->bi_end_io != end_sync_read)
1961                         continue;
1962                 if (i == first)
1963                         continue;
1964                 if (!r10_bio->devs[i].bio->bi_error) {
1965                         /* We know that the bi_io_vec layout is the same for
1966                          * both 'first' and 'i', so we just compare them.
1967                          * All vec entries are PAGE_SIZE;
1968                          */
1969                         int sectors = r10_bio->sectors;
1970                         for (j = 0; j < vcnt; j++) {
1971                                 int len = PAGE_SIZE;
1972                                 if (sectors < (len / 512))
1973                                         len = sectors * 512;
1974                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1975                                            page_address(tbio->bi_io_vec[j].bv_page),
1976                                            len))
1977                                         break;
1978                                 sectors -= len/512;
1979                         }
1980                         if (j == vcnt)
1981                                 continue;
1982                         atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
1983                         if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1984                                 /* Don't fix anything. */
1985                                 continue;
1986                 }
1987                 /* Ok, we need to write this bio, either to correct an
1988                  * inconsistency or to correct an unreadable block.
1989                  * First we need to fixup bv_offset, bv_len and
1990                  * bi_vecs, as the read request might have corrupted these
1991                  */
1992                 bio_reset(tbio);
1993
1994                 tbio->bi_vcnt = vcnt;
1995                 tbio->bi_iter.bi_size = fbio->bi_iter.bi_size;
1996                 tbio->bi_rw = WRITE;
1997                 tbio->bi_private = r10_bio;
1998                 tbio->bi_iter.bi_sector = r10_bio->devs[i].addr;
1999                 tbio->bi_end_io = end_sync_write;
2000
2001                 bio_copy_data(tbio, fbio);
2002
2003                 d = r10_bio->devs[i].devnum;
2004                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2005                 atomic_inc(&r10_bio->remaining);
2006                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(tbio));
2007
2008                 tbio->bi_iter.bi_sector += conf->mirrors[d].rdev->data_offset;
2009                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2010                 generic_make_request(tbio);
2011         }
2012
2013         /* Now write out to any replacement devices
2014          * that are active
2015          */
2016         for (i = 0; i < conf->copies; i++) {
2017                 int d;
2018
2019                 tbio = r10_bio->devs[i].repl_bio;
2020                 if (!tbio || !tbio->bi_end_io)
2021                         continue;
2022                 if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2023                     && r10_bio->devs[i].bio != fbio)
2024                         bio_copy_data(tbio, fbio);
2025                 d = r10_bio->devs[i].devnum;
2026                 atomic_inc(&r10_bio->remaining);
2027                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2028                              bio_sectors(tbio));
2029                 generic_make_request(tbio);
2030         }
2031
2032 done:
2033         if (atomic_dec_and_test(&r10_bio->remaining)) {
2034                 md_done_sync(mddev, r10_bio->sectors, 1);
2035                 put_buf(r10_bio);
2036         }
2037 }
2038
2039 /*
2040  * Now for the recovery code.
2041  * Recovery happens across physical sectors.
2042  * We recover all non-is_sync drives by finding the virtual address of
2043  * each, and then choose a working drive that also has that virt address.
2044  * There is a separate r10_bio for each non-in_sync drive.
2045  * Only the first two slots are in use. The first for reading,
2046  * The second for writing.
2047  *
2048  */
2049 static void fix_recovery_read_error(struct r10bio *r10_bio)
2050 {
2051         /* We got a read error during recovery.
2052          * We repeat the read in smaller page-sized sections.
2053          * If a read succeeds, write it to the new device or record
2054          * a bad block if we cannot.
2055          * If a read fails, record a bad block on both old and
2056          * new devices.
2057          */
2058         struct mddev *mddev = r10_bio->mddev;
2059         struct r10conf *conf = mddev->private;
2060         struct bio *bio = r10_bio->devs[0].bio;
2061         sector_t sect = 0;
2062         int sectors = r10_bio->sectors;
2063         int idx = 0;
2064         int dr = r10_bio->devs[0].devnum;
2065         int dw = r10_bio->devs[1].devnum;
2066
2067         while (sectors) {
2068                 int s = sectors;
2069                 struct md_rdev *rdev;
2070                 sector_t addr;
2071                 int ok;
2072
2073                 if (s > (PAGE_SIZE>>9))
2074                         s = PAGE_SIZE >> 9;
2075
2076                 rdev = conf->mirrors[dr].rdev;
2077                 addr = r10_bio->devs[0].addr + sect,
2078                 ok = sync_page_io(rdev,
2079                                   addr,
2080                                   s << 9,
2081                                   bio->bi_io_vec[idx].bv_page,
2082                                   READ, false);
2083                 if (ok) {
2084                         rdev = conf->mirrors[dw].rdev;
2085                         addr = r10_bio->devs[1].addr + sect;
2086                         ok = sync_page_io(rdev,
2087                                           addr,
2088                                           s << 9,
2089                                           bio->bi_io_vec[idx].bv_page,
2090                                           WRITE, false);
2091                         if (!ok) {
2092                                 set_bit(WriteErrorSeen, &rdev->flags);
2093                                 if (!test_and_set_bit(WantReplacement,
2094                                                       &rdev->flags))
2095                                         set_bit(MD_RECOVERY_NEEDED,
2096                                                 &rdev->mddev->recovery);
2097                         }
2098                 }
2099                 if (!ok) {
2100                         /* We don't worry if we cannot set a bad block -
2101                          * it really is bad so there is no loss in not
2102                          * recording it yet
2103                          */
2104                         rdev_set_badblocks(rdev, addr, s, 0);
2105
2106                         if (rdev != conf->mirrors[dw].rdev) {
2107                                 /* need bad block on destination too */
2108                                 struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2109                                 addr = r10_bio->devs[1].addr + sect;
2110                                 ok = rdev_set_badblocks(rdev2, addr, s, 0);
2111                                 if (!ok) {
2112                                         /* just abort the recovery */
2113                                         printk(KERN_NOTICE
2114                                                "md/raid10:%s: recovery aborted"
2115                                                " due to read error\n",
2116                                                mdname(mddev));
2117
2118                                         conf->mirrors[dw].recovery_disabled
2119                                                 = mddev->recovery_disabled;
2120                                         set_bit(MD_RECOVERY_INTR,
2121                                                 &mddev->recovery);
2122                                         break;
2123                                 }
2124                         }
2125                 }
2126
2127                 sectors -= s;
2128                 sect += s;
2129                 idx++;
2130         }
2131 }
2132
2133 static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2134 {
2135         struct r10conf *conf = mddev->private;
2136         int d;
2137         struct bio *wbio, *wbio2;
2138
2139         if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2140                 fix_recovery_read_error(r10_bio);
2141                 end_sync_request(r10_bio);
2142                 return;
2143         }
2144
2145         /*
2146          * share the pages with the first bio
2147          * and submit the write request
2148          */
2149         d = r10_bio->devs[1].devnum;
2150         wbio = r10_bio->devs[1].bio;
2151         wbio2 = r10_bio->devs[1].repl_bio;
2152         /* Need to test wbio2->bi_end_io before we call
2153          * generic_make_request as if the former is NULL,
2154          * the latter is free to free wbio2.
2155          */
2156         if (wbio2 && !wbio2->bi_end_io)
2157                 wbio2 = NULL;
2158         if (wbio->bi_end_io) {
2159                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2160                 md_sync_acct(conf->mirrors[d].rdev->bdev, bio_sectors(wbio));
2161                 generic_make_request(wbio);
2162         }
2163         if (wbio2) {
2164                 atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2165                 md_sync_acct(conf->mirrors[d].replacement->bdev,
2166                              bio_sectors(wbio2));
2167                 generic_make_request(wbio2);
2168         }
2169 }
2170
2171 /*
2172  * Used by fix_read_error() to decay the per rdev read_errors.
2173  * We halve the read error count for every hour that has elapsed
2174  * since the last recorded read error.
2175  *
2176  */
2177 static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2178 {
2179         long cur_time_mon;
2180         unsigned long hours_since_last;
2181         unsigned int read_errors = atomic_read(&rdev->read_errors);
2182
2183         cur_time_mon = ktime_get_seconds();
2184
2185         if (rdev->last_read_error == 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 = (long)(cur_time_mon -
2192                             rdev->last_read_error) / 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, 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                             !test_bit(Faulty, &rdev->flags) &&
2295                             is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2296                                         &first_bad, &bad_sectors) == 0) {
2297                                 atomic_inc(&rdev->nr_pending);
2298                                 rcu_read_unlock();
2299                                 success = sync_page_io(rdev,
2300                                                        r10_bio->devs[sl].addr +
2301                                                        sect,
2302                                                        s<<9,
2303                                                        conf->tmppage, READ, false);
2304                                 rdev_dec_pending(rdev, mddev);
2305                                 rcu_read_lock();
2306                                 if (success)
2307                                         break;
2308                         }
2309                         sl++;
2310                         if (sl == conf->copies)
2311                                 sl = 0;
2312                 } while (!success && sl != r10_bio->read_slot);
2313                 rcu_read_unlock();
2314
2315                 if (!success) {
2316                         /* Cannot read from anywhere, just mark the block
2317                          * as bad on the first device to discourage future
2318                          * reads.
2319                          */
2320                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2321                         rdev = conf->mirrors[dn].rdev;
2322
2323                         if (!rdev_set_badblocks(
2324                                     rdev,
2325                                     r10_bio->devs[r10_bio->read_slot].addr
2326                                     + sect,
2327                                     s, 0)) {
2328                                 md_error(mddev, rdev);
2329                                 r10_bio->devs[r10_bio->read_slot].bio
2330                                         = IO_BLOCKED;
2331                         }
2332                         break;
2333                 }
2334
2335                 start = sl;
2336                 /* write it back and re-read */
2337                 rcu_read_lock();
2338                 while (sl != r10_bio->read_slot) {
2339                         char b[BDEVNAME_SIZE];
2340
2341                         if (sl==0)
2342                                 sl = conf->copies;
2343                         sl--;
2344                         d = r10_bio->devs[sl].devnum;
2345                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2346                         if (!rdev ||
2347                             test_bit(Faulty, &rdev->flags) ||
2348                             !test_bit(In_sync, &rdev->flags))
2349                                 continue;
2350
2351                         atomic_inc(&rdev->nr_pending);
2352                         rcu_read_unlock();
2353                         if (r10_sync_page_io(rdev,
2354                                              r10_bio->devs[sl].addr +
2355                                              sect,
2356                                              s, conf->tmppage, WRITE)
2357                             == 0) {
2358                                 /* Well, this device is dead */
2359                                 printk(KERN_NOTICE
2360                                        "md/raid10:%s: read correction "
2361                                        "write failed"
2362                                        " (%d sectors at %llu on %s)\n",
2363                                        mdname(mddev), s,
2364                                        (unsigned long long)(
2365                                                sect +
2366                                                choose_data_offset(r10_bio,
2367                                                                   rdev)),
2368                                        bdevname(rdev->bdev, b));
2369                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2370                                        "drive\n",
2371                                        mdname(mddev),
2372                                        bdevname(rdev->bdev, b));
2373                         }
2374                         rdev_dec_pending(rdev, mddev);
2375                         rcu_read_lock();
2376                 }
2377                 sl = start;
2378                 while (sl != r10_bio->read_slot) {
2379                         char b[BDEVNAME_SIZE];
2380
2381                         if (sl==0)
2382                                 sl = conf->copies;
2383                         sl--;
2384                         d = r10_bio->devs[sl].devnum;
2385                         rdev = rcu_dereference(conf->mirrors[d].rdev);
2386                         if (!rdev ||
2387                             test_bit(Faulty, &rdev->flags) ||
2388                             !test_bit(In_sync, &rdev->flags))
2389                                 continue;
2390
2391                         atomic_inc(&rdev->nr_pending);
2392                         rcu_read_unlock();
2393                         switch (r10_sync_page_io(rdev,
2394                                              r10_bio->devs[sl].addr +
2395                                              sect,
2396                                              s, conf->tmppage,
2397                                                  READ)) {
2398                         case 0:
2399                                 /* Well, this device is dead */
2400                                 printk(KERN_NOTICE
2401                                        "md/raid10:%s: unable to read back "
2402                                        "corrected sectors"
2403                                        " (%d sectors at %llu on %s)\n",
2404                                        mdname(mddev), s,
2405                                        (unsigned long long)(
2406                                                sect +
2407                                                choose_data_offset(r10_bio, rdev)),
2408                                        bdevname(rdev->bdev, b));
2409                                 printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2410                                        "drive\n",
2411                                        mdname(mddev),
2412                                        bdevname(rdev->bdev, b));
2413                                 break;
2414                         case 1:
2415                                 printk(KERN_INFO
2416                                        "md/raid10:%s: read error corrected"
2417                                        " (%d sectors at %llu on %s)\n",
2418                                        mdname(mddev), s,
2419                                        (unsigned long long)(
2420                                                sect +
2421                                                choose_data_offset(r10_bio, rdev)),
2422                                        bdevname(rdev->bdev, b));
2423                                 atomic_add(s, &rdev->corrected_errors);
2424                         }
2425
2426                         rdev_dec_pending(rdev, mddev);
2427                         rcu_read_lock();
2428                 }
2429                 rcu_read_unlock();
2430
2431                 sectors -= s;
2432                 sect += s;
2433         }
2434 }
2435
2436 static int narrow_write_error(struct r10bio *r10_bio, int i)
2437 {
2438         struct bio *bio = r10_bio->master_bio;
2439         struct mddev *mddev = r10_bio->mddev;
2440         struct r10conf *conf = mddev->private;
2441         struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2442         /* bio has the data to be written to slot 'i' where
2443          * we just recently had a write error.
2444          * We repeatedly clone the bio and trim down to one block,
2445          * then try the write.  Where the write fails we record
2446          * a bad block.
2447          * It is conceivable that the bio doesn't exactly align with
2448          * blocks.  We must handle this.
2449          *
2450          * We currently own a reference to the rdev.
2451          */
2452
2453         int block_sectors;
2454         sector_t sector;
2455         int sectors;
2456         int sect_to_write = r10_bio->sectors;
2457         int ok = 1;
2458
2459         if (rdev->badblocks.shift < 0)
2460                 return 0;
2461
2462         block_sectors = roundup(1 << rdev->badblocks.shift,
2463                                 bdev_logical_block_size(rdev->bdev) >> 9);
2464         sector = r10_bio->sector;
2465         sectors = ((r10_bio->sector + block_sectors)
2466                    & ~(sector_t)(block_sectors - 1))
2467                 - sector;
2468
2469         while (sect_to_write) {
2470                 struct bio *wbio;
2471                 if (sectors > sect_to_write)
2472                         sectors = sect_to_write;
2473                 /* Write at 'sector' for 'sectors' */
2474                 wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2475                 bio_trim(wbio, sector - bio->bi_iter.bi_sector, sectors);
2476                 wbio->bi_iter.bi_sector = (r10_bio->devs[i].addr+
2477                                    choose_data_offset(r10_bio, rdev) +
2478                                    (sector - r10_bio->sector));
2479                 wbio->bi_bdev = rdev->bdev;
2480                 if (submit_bio_wait(WRITE, wbio) < 0)
2481                         /* Failure! */
2482                         ok = rdev_set_badblocks(rdev, sector,
2483                                                 sectors, 0)
2484                                 && ok;
2485
2486                 bio_put(wbio);
2487                 sect_to_write -= sectors;
2488                 sector += sectors;
2489                 sectors = block_sectors;
2490         }
2491         return ok;
2492 }
2493
2494 static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2495 {
2496         int slot = r10_bio->read_slot;
2497         struct bio *bio;
2498         struct r10conf *conf = mddev->private;
2499         struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2500         char b[BDEVNAME_SIZE];
2501         unsigned long do_sync;
2502         int max_sectors;
2503
2504         /* we got a read error. Maybe the drive is bad.  Maybe just
2505          * the block and we can fix it.
2506          * We freeze all other IO, and try reading the block from
2507          * other devices.  When we find one, we re-write
2508          * and check it that fixes the read error.
2509          * This is all done synchronously while the array is
2510          * frozen.
2511          */
2512         bio = r10_bio->devs[slot].bio;
2513         bdevname(bio->bi_bdev, b);
2514         bio_put(bio);
2515         r10_bio->devs[slot].bio = NULL;
2516
2517         if (mddev->ro == 0) {
2518                 freeze_array(conf, 1);
2519                 fix_read_error(conf, mddev, r10_bio);
2520                 unfreeze_array(conf);
2521         } else
2522                 r10_bio->devs[slot].bio = IO_BLOCKED;
2523
2524         rdev_dec_pending(rdev, mddev);
2525
2526 read_more:
2527         rdev = read_balance(conf, r10_bio, &max_sectors);
2528         if (rdev == NULL) {
2529                 printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2530                        " read error for block %llu\n",
2531                        mdname(mddev), b,
2532                        (unsigned long long)r10_bio->sector);
2533                 raid_end_bio_io(r10_bio);
2534                 return;
2535         }
2536
2537         do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2538         slot = r10_bio->read_slot;
2539         printk_ratelimited(
2540                 KERN_ERR
2541                 "md/raid10:%s: %s: redirecting "
2542                 "sector %llu to another mirror\n",
2543                 mdname(mddev),
2544                 bdevname(rdev->bdev, b),
2545                 (unsigned long long)r10_bio->sector);
2546         bio = bio_clone_mddev(r10_bio->master_bio,
2547                               GFP_NOIO, mddev);
2548         bio_trim(bio, r10_bio->sector - bio->bi_iter.bi_sector, max_sectors);
2549         r10_bio->devs[slot].bio = bio;
2550         r10_bio->devs[slot].rdev = rdev;
2551         bio->bi_iter.bi_sector = r10_bio->devs[slot].addr
2552                 + choose_data_offset(r10_bio, rdev);
2553         bio->bi_bdev = rdev->bdev;
2554         bio->bi_rw = READ | do_sync;
2555         bio->bi_private = r10_bio;
2556         bio->bi_end_io = raid10_end_read_request;
2557         if (max_sectors < r10_bio->sectors) {
2558                 /* Drat - have to split this up more */
2559                 struct bio *mbio = r10_bio->master_bio;
2560                 int sectors_handled =
2561                         r10_bio->sector + max_sectors
2562                         - mbio->bi_iter.bi_sector;
2563                 r10_bio->sectors = max_sectors;
2564                 spin_lock_irq(&conf->device_lock);
2565                 if (mbio->bi_phys_segments == 0)
2566                         mbio->bi_phys_segments = 2;
2567                 else
2568                         mbio->bi_phys_segments++;
2569                 spin_unlock_irq(&conf->device_lock);
2570                 generic_make_request(bio);
2571
2572                 r10_bio = mempool_alloc(conf->r10bio_pool,
2573                                         GFP_NOIO);
2574                 r10_bio->master_bio = mbio;
2575                 r10_bio->sectors = bio_sectors(mbio) - sectors_handled;
2576                 r10_bio->state = 0;
2577                 set_bit(R10BIO_ReadError,
2578                         &r10_bio->state);
2579                 r10_bio->mddev = mddev;
2580                 r10_bio->sector = mbio->bi_iter.bi_sector
2581                         + sectors_handled;
2582
2583                 goto read_more;
2584         } else
2585                 generic_make_request(bio);
2586 }
2587
2588 static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2589 {
2590         /* Some sort of write request has finished and it
2591          * succeeded in writing where we thought there was a
2592          * bad block.  So forget the bad block.
2593          * Or possibly if failed and we need to record
2594          * a bad block.
2595          */
2596         int m;
2597         struct md_rdev *rdev;
2598
2599         if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2600             test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2601                 for (m = 0; m < conf->copies; m++) {
2602                         int dev = r10_bio->devs[m].devnum;
2603                         rdev = conf->mirrors[dev].rdev;
2604                         if (r10_bio->devs[m].bio == NULL)
2605                                 continue;
2606                         if (!r10_bio->devs[m].bio->bi_error) {
2607                                 rdev_clear_badblocks(
2608                                         rdev,
2609                                         r10_bio->devs[m].addr,
2610                                         r10_bio->sectors, 0);
2611                         } else {
2612                                 if (!rdev_set_badblocks(
2613                                             rdev,
2614                                             r10_bio->devs[m].addr,
2615                                             r10_bio->sectors, 0))
2616                                         md_error(conf->mddev, rdev);
2617                         }
2618                         rdev = conf->mirrors[dev].replacement;
2619                         if (r10_bio->devs[m].repl_bio == NULL)
2620                                 continue;
2621
2622                         if (!r10_bio->devs[m].repl_bio->bi_error) {
2623                                 rdev_clear_badblocks(
2624                                         rdev,
2625                                         r10_bio->devs[m].addr,
2626                                         r10_bio->sectors, 0);
2627                         } else {
2628                                 if (!rdev_set_badblocks(
2629                                             rdev,
2630                                             r10_bio->devs[m].addr,
2631                                             r10_bio->sectors, 0))
2632                                         md_error(conf->mddev, rdev);
2633                         }
2634                 }
2635                 put_buf(r10_bio);
2636         } else {
2637                 bool fail = false;
2638                 for (m = 0; m < conf->copies; m++) {
2639                         int dev = r10_bio->devs[m].devnum;
2640                         struct bio *bio = r10_bio->devs[m].bio;
2641                         rdev = conf->mirrors[dev].rdev;
2642                         if (bio == IO_MADE_GOOD) {
2643                                 rdev_clear_badblocks(
2644                                         rdev,
2645                                         r10_bio->devs[m].addr,
2646                                         r10_bio->sectors, 0);
2647                                 rdev_dec_pending(rdev, conf->mddev);
2648                         } else if (bio != NULL && bio->bi_error) {
2649                                 fail = true;
2650                                 if (!narrow_write_error(r10_bio, m)) {
2651                                         md_error(conf->mddev, rdev);
2652                                         set_bit(R10BIO_Degraded,
2653                                                 &r10_bio->state);
2654                                 }
2655                                 rdev_dec_pending(rdev, conf->mddev);
2656                         }
2657                         bio = r10_bio->devs[m].repl_bio;
2658                         rdev = conf->mirrors[dev].replacement;
2659                         if (rdev && bio == IO_MADE_GOOD) {
2660                                 rdev_clear_badblocks(
2661                                         rdev,
2662                                         r10_bio->devs[m].addr,
2663                                         r10_bio->sectors, 0);
2664                                 rdev_dec_pending(rdev, conf->mddev);
2665                         }
2666                 }
2667                 if (fail) {
2668                         spin_lock_irq(&conf->device_lock);
2669                         list_add(&r10_bio->retry_list, &conf->bio_end_io_list);
2670                         conf->nr_queued++;
2671                         spin_unlock_irq(&conf->device_lock);
2672                         md_wakeup_thread(conf->mddev->thread);
2673                 } else {
2674                         if (test_bit(R10BIO_WriteError,
2675                                      &r10_bio->state))
2676                                 close_write(r10_bio);
2677                         raid_end_bio_io(r10_bio);
2678                 }
2679         }
2680 }
2681
2682 static void raid10d(struct md_thread *thread)
2683 {
2684         struct mddev *mddev = thread->mddev;
2685         struct r10bio *r10_bio;
2686         unsigned long flags;
2687         struct r10conf *conf = mddev->private;
2688         struct list_head *head = &conf->retry_list;
2689         struct blk_plug plug;
2690
2691         md_check_recovery(mddev);
2692
2693         if (!list_empty_careful(&conf->bio_end_io_list) &&
2694             !test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2695                 LIST_HEAD(tmp);
2696                 spin_lock_irqsave(&conf->device_lock, flags);
2697                 if (!test_bit(MD_CHANGE_PENDING, &mddev->flags)) {
2698                         while (!list_empty(&conf->bio_end_io_list)) {
2699                                 list_move(conf->bio_end_io_list.prev, &tmp);
2700                                 conf->nr_queued--;
2701                         }
2702                 }
2703                 spin_unlock_irqrestore(&conf->device_lock, flags);
2704                 while (!list_empty(&tmp)) {
2705                         r10_bio = list_first_entry(&tmp, struct r10bio,
2706                                                    retry_list);
2707                         list_del(&r10_bio->retry_list);
2708                         if (mddev->degraded)
2709                                 set_bit(R10BIO_Degraded, &r10_bio->state);
2710
2711                         if (test_bit(R10BIO_WriteError,
2712                                      &r10_bio->state))
2713                                 close_write(r10_bio);
2714                         raid_end_bio_io(r10_bio);
2715                 }
2716         }
2717
2718         blk_start_plug(&plug);
2719         for (;;) {
2720
2721                 flush_pending_writes(conf);
2722
2723                 spin_lock_irqsave(&conf->device_lock, flags);
2724                 if (list_empty(head)) {
2725                         spin_unlock_irqrestore(&conf->device_lock, flags);
2726                         break;
2727                 }
2728                 r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2729                 list_del(head->prev);
2730                 conf->nr_queued--;
2731                 spin_unlock_irqrestore(&conf->device_lock, flags);
2732
2733                 mddev = r10_bio->mddev;
2734                 conf = mddev->private;
2735                 if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2736                     test_bit(R10BIO_WriteError, &r10_bio->state))
2737                         handle_write_completed(conf, r10_bio);
2738                 else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2739                         reshape_request_write(mddev, r10_bio);
2740                 else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2741                         sync_request_write(mddev, r10_bio);
2742                 else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2743                         recovery_request_write(mddev, r10_bio);
2744                 else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2745                         handle_read_error(mddev, r10_bio);
2746                 else {
2747                         /* just a partial read to be scheduled from a
2748                          * separate context
2749                          */
2750                         int slot = r10_bio->read_slot;
2751                         generic_make_request(r10_bio->devs[slot].bio);
2752                 }
2753
2754                 cond_resched();
2755                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2756                         md_check_recovery(mddev);
2757         }
2758         blk_finish_plug(&plug);
2759 }
2760
2761 static int init_resync(struct r10conf *conf)
2762 {
2763         int buffs;
2764         int i;
2765
2766         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2767         BUG_ON(conf->r10buf_pool);
2768         conf->have_replacement = 0;
2769         for (i = 0; i < conf->geo.raid_disks; i++)
2770                 if (conf->mirrors[i].replacement)
2771                         conf->have_replacement = 1;
2772         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2773         if (!conf->r10buf_pool)
2774                 return -ENOMEM;
2775         conf->next_resync = 0;
2776         return 0;
2777 }
2778
2779 /*
2780  * perform a "sync" on one "block"
2781  *
2782  * We need to make sure that no normal I/O request - particularly write
2783  * requests - conflict with active sync requests.
2784  *
2785  * This is achieved by tracking pending requests and a 'barrier' concept
2786  * that can be installed to exclude normal IO requests.
2787  *
2788  * Resync and recovery are handled very differently.
2789  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2790  *
2791  * For resync, we iterate over virtual addresses, read all copies,
2792  * and update if there are differences.  If only one copy is live,
2793  * skip it.
2794  * For recovery, we iterate over physical addresses, read a good
2795  * value for each non-in_sync drive, and over-write.
2796  *
2797  * So, for recovery we may have several outstanding complex requests for a
2798  * given address, one for each out-of-sync device.  We model this by allocating
2799  * a number of r10_bio structures, one for each out-of-sync device.
2800  * As we setup these structures, we collect all bio's together into a list
2801  * which we then process collectively to add pages, and then process again
2802  * to pass to generic_make_request.
2803  *
2804  * The r10_bio structures are linked using a borrowed master_bio pointer.
2805  * This link is counted in ->remaining.  When the r10_bio that points to NULL
2806  * has its remaining count decremented to 0, the whole complex operation
2807  * is complete.
2808  *
2809  */
2810
2811 static sector_t raid10_sync_request(struct mddev *mddev, sector_t sector_nr,
2812                              int *skipped)
2813 {
2814         struct r10conf *conf = mddev->private;
2815         struct r10bio *r10_bio;
2816         struct bio *biolist = NULL, *bio;
2817         sector_t max_sector, nr_sectors;
2818         int i;
2819         int max_sync;
2820         sector_t sync_blocks;
2821         sector_t sectors_skipped = 0;
2822         int chunks_skipped = 0;
2823         sector_t chunk_mask = conf->geo.chunk_mask;
2824
2825         if (!conf->r10buf_pool)
2826                 if (init_resync(conf))
2827                         return 0;
2828
2829         /*
2830          * Allow skipping a full rebuild for incremental assembly
2831          * of a clean array, like RAID1 does.
2832          */
2833         if (mddev->bitmap == NULL &&
2834             mddev->recovery_cp == MaxSector &&
2835             mddev->reshape_position == MaxSector &&
2836             !test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2837             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2838             !test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
2839             conf->fullsync == 0) {
2840                 *skipped = 1;
2841                 return mddev->dev_sectors - sector_nr;
2842         }
2843
2844  skipped:
2845         max_sector = mddev->dev_sectors;
2846         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2847             test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2848                 max_sector = mddev->resync_max_sectors;
2849         if (sector_nr >= max_sector) {
2850                 /* If we aborted, we need to abort the
2851                  * sync on the 'current' bitmap chucks (there can
2852                  * be several when recovering multiple devices).
2853                  * as we may have started syncing it but not finished.
2854                  * We can find the current address in
2855                  * mddev->curr_resync, but for recovery,
2856                  * we need to convert that to several
2857                  * virtual addresses.
2858                  */
2859                 if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2860                         end_reshape(conf);
2861                         close_sync(conf);
2862                         return 0;
2863                 }
2864
2865                 if (mddev->curr_resync < max_sector) { /* aborted */
2866                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2867                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2868                                                 &sync_blocks, 1);
2869                         else for (i = 0; i < conf->geo.raid_disks; i++) {
2870                                 sector_t sect =
2871                                         raid10_find_virt(conf, mddev->curr_resync, i);
2872                                 bitmap_end_sync(mddev->bitmap, sect,
2873                                                 &sync_blocks, 1);
2874                         }
2875                 } else {
2876                         /* completed sync */
2877                         if ((!mddev->bitmap || conf->fullsync)
2878                             && conf->have_replacement
2879                             && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2880                                 /* Completed a full sync so the replacements
2881                                  * are now fully recovered.
2882                                  */
2883                                 rcu_read_lock();
2884                                 for (i = 0; i < conf->geo.raid_disks; i++) {
2885                                         struct md_rdev *rdev =
2886                                                 rcu_dereference(conf->mirrors[i].replacement);
2887                                         if (rdev)
2888                                                 rdev->recovery_offset = MaxSector;
2889                                 }
2890                                 rcu_read_unlock();
2891                         }
2892                         conf->fullsync = 0;
2893                 }
2894                 bitmap_close_sync(mddev->bitmap);
2895                 close_sync(conf);
2896                 *skipped = 1;
2897                 return sectors_skipped;
2898         }
2899
2900         if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2901                 return reshape_request(mddev, sector_nr, skipped);
2902
2903         if (chunks_skipped >= conf->geo.raid_disks) {
2904                 /* if there has been nothing to do on any drive,
2905                  * then there is nothing to do at all..
2906                  */
2907                 *skipped = 1;
2908                 return (max_sector - sector_nr) + sectors_skipped;
2909         }
2910
2911         if (max_sector > mddev->resync_max)
2912                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
2913
2914         /* make sure whole request will fit in a chunk - if chunks
2915          * are meaningful
2916          */
2917         if (conf->geo.near_copies < conf->geo.raid_disks &&
2918             max_sector > (sector_nr | chunk_mask))
2919                 max_sector = (sector_nr | chunk_mask) + 1;
2920
2921         /*
2922          * If there is non-resync activity waiting for a turn, then let it
2923          * though before starting on this new sync request.
2924          */
2925         if (conf->nr_waiting)
2926                 schedule_timeout_uninterruptible(1);
2927
2928         /* Again, very different code for resync and recovery.
2929          * Both must result in an r10bio with a list of bios that
2930          * have bi_end_io, bi_sector, bi_bdev set,
2931          * and bi_private set to the r10bio.
2932          * For recovery, we may actually create several r10bios
2933          * with 2 bios in each, that correspond to the bios in the main one.
2934          * In this case, the subordinate r10bios link back through a
2935          * borrowed master_bio pointer, and the counter in the master
2936          * includes a ref from each subordinate.
2937          */
2938         /* First, we decide what to do and set ->bi_end_io
2939          * To end_sync_read if we want to read, and
2940          * end_sync_write if we will want to write.
2941          */
2942
2943         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
2944         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2945                 /* recovery... the complicated one */
2946                 int j;
2947                 r10_bio = NULL;
2948
2949                 for (i = 0 ; i < conf->geo.raid_disks; i++) {
2950                         int still_degraded;
2951                         struct r10bio *rb2;
2952                         sector_t sect;
2953                         int must_sync;
2954                         int any_working;
2955                         struct raid10_info *mirror = &conf->mirrors[i];
2956                         struct md_rdev *mrdev, *mreplace;
2957
2958                         rcu_read_lock();
2959                         mrdev = rcu_dereference(mirror->rdev);
2960                         mreplace = rcu_dereference(mirror->replacement);
2961
2962                         if ((mrdev == NULL ||
2963                              test_bit(Faulty, &mrdev->flags) ||
2964                              test_bit(In_sync, &mrdev->flags)) &&
2965                             (mreplace == NULL ||
2966                              test_bit(Faulty, &mreplace->flags))) {
2967                                 rcu_read_unlock();
2968                                 continue;
2969                         }
2970
2971                         still_degraded = 0;
2972                         /* want to reconstruct this device */
2973                         rb2 = r10_bio;
2974                         sect = raid10_find_virt(conf, sector_nr, i);
2975                         if (sect >= mddev->resync_max_sectors) {
2976                                 /* last stripe is not complete - don't
2977                                  * try to recover this sector.
2978                                  */
2979                                 rcu_read_unlock();
2980                                 continue;
2981                         }
2982                         if (mreplace && test_bit(Faulty, &mreplace->flags))
2983                                 mreplace = NULL;
2984                         /* Unless we are doing a full sync, or a replacement
2985                          * we only need to recover the block if it is set in
2986                          * the bitmap
2987                          */
2988                         must_sync = bitmap_start_sync(mddev->bitmap, sect,
2989                                                       &sync_blocks, 1);
2990                         if (sync_blocks < max_sync)
2991                                 max_sync = sync_blocks;
2992                         if (!must_sync &&
2993                             mreplace == NULL &&