Root/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 *
43 * The data to be stored is divided into chunks using chunksize. Each device
44 * is divided into far_copies sections. In each section, chunks are laid out
45 * in a style similar to raid0, but near_copies copies of each chunk is stored
46 * (each on a different drive). The starting device for each section is offset
47 * near_copies from the starting device of the previous section. Thus there
48 * are (near_copies * far_copies) of each chunk, and each is on a different
49 * drive. near_copies and far_copies must be at least one, and their product
50 * is at most raid_disks.
51 *
52 * If far_offset is true, then the far_copies are handled a bit differently.
53 * The copies are still in different stripes, but instead of being very far
54 * apart on disk, there are adjacent stripes.
55 *
56 * The far and offset algorithms are handled slightly differently if
57 * 'use_far_sets' is true. In this case, the array's devices are grouped into
58 * sets that are (near_copies * far_copies) in size. The far copied stripes
59 * are still shifted by 'near_copies' devices, but this shifting stays confined
60 * to the set rather than the entire array. This is done to improve the number
61 * of device combinations that can fail without causing the array to fail.
62 * Example 'far' algorithm w/o 'use_far_sets' (each letter represents a chunk
63 * on a device):
64 * A B C D A B C D E
65 * ... ...
66 * D A B C E A B C D
67 * Example 'far' algorithm w/ 'use_far_sets' enabled (sets illustrated w/ []'s):
68 * [A B] [C D] [A B] [C D E]
69 * |...| |...| |...| | ... |
70 * [B A] [D C] [B A] [E C D]
71 */
72
73/*
74 * Number of guaranteed r10bios in case of extreme VM load:
75 */
76#define NR_RAID10_BIOS 256
77
78/* when we get a read error on a read-only array, we redirect to another
79 * device without failing the first device, or trying to over-write to
80 * correct the read error. To keep track of bad blocks on a per-bio
81 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
82 */
83#define IO_BLOCKED ((struct bio *)1)
84/* When we successfully write to a known bad-block, we need to remove the
85 * bad-block marking which must be done from process context. So we record
86 * the success by setting devs[n].bio to IO_MADE_GOOD
87 */
88#define IO_MADE_GOOD ((struct bio *)2)
89
90#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
91
92/* When there are this many requests queued to be written by
93 * the raid10 thread, we become 'congested' to provide back-pressure
94 * for writeback.
95 */
96static int max_queued_requests = 1024;
97
98static void allow_barrier(struct r10conf *conf);
99static void lower_barrier(struct r10conf *conf);
100static int enough(struct r10conf *conf, int ignore);
101static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
102                int *skipped);
103static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio);
104static void end_reshape_write(struct bio *bio, int error);
105static void end_reshape(struct r10conf *conf);
106
107static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
108{
109    struct r10conf *conf = data;
110    int size = offsetof(struct r10bio, devs[conf->copies]);
111
112    /* allocate a r10bio with room for raid_disks entries in the
113     * bios array */
114    return kzalloc(size, gfp_flags);
115}
116
117static void r10bio_pool_free(void *r10_bio, void *data)
118{
119    kfree(r10_bio);
120}
121
122/* Maximum size of each resync request */
123#define RESYNC_BLOCK_SIZE (64*1024)
124#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
125/* amount of memory to reserve for resync requests */
126#define RESYNC_WINDOW (1024*1024)
127/* maximum number of concurrent requests, memory permitting */
128#define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
129
130/*
131 * When performing a resync, we need to read and compare, so
132 * we need as many pages are there are copies.
133 * When performing a recovery, we need 2 bios, one for read,
134 * one for write (we recover only one drive per r10buf)
135 *
136 */
137static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
138{
139    struct r10conf *conf = data;
140    struct page *page;
141    struct r10bio *r10_bio;
142    struct bio *bio;
143    int i, j;
144    int nalloc;
145
146    r10_bio = r10bio_pool_alloc(gfp_flags, conf);
147    if (!r10_bio)
148        return NULL;
149
150    if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery) ||
151        test_bit(MD_RECOVERY_RESHAPE, &conf->mddev->recovery))
152        nalloc = conf->copies; /* resync */
153    else
154        nalloc = 2; /* recovery */
155
156    /*
157     * Allocate bios.
158     */
159    for (j = nalloc ; j-- ; ) {
160        bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
161        if (!bio)
162            goto out_free_bio;
163        r10_bio->devs[j].bio = bio;
164        if (!conf->have_replacement)
165            continue;
166        bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
167        if (!bio)
168            goto out_free_bio;
169        r10_bio->devs[j].repl_bio = bio;
170    }
171    /*
172     * Allocate RESYNC_PAGES data pages and attach them
173     * where needed.
174     */
175    for (j = 0 ; j < nalloc; j++) {
176        struct bio *rbio = r10_bio->devs[j].repl_bio;
177        bio = r10_bio->devs[j].bio;
178        for (i = 0; i < RESYNC_PAGES; i++) {
179            if (j > 0 && !test_bit(MD_RECOVERY_SYNC,
180                           &conf->mddev->recovery)) {
181                /* we can share bv_page's during recovery
182                 * and reshape */
183                struct bio *rbio = r10_bio->devs[0].bio;
184                page = rbio->bi_io_vec[i].bv_page;
185                get_page(page);
186            } else
187                page = alloc_page(gfp_flags);
188            if (unlikely(!page))
189                goto out_free_pages;
190
191            bio->bi_io_vec[i].bv_page = page;
192            if (rbio)
193                rbio->bi_io_vec[i].bv_page = page;
194        }
195    }
196
197    return r10_bio;
198
199out_free_pages:
200    for ( ; i > 0 ; i--)
201        safe_put_page(bio->bi_io_vec[i-1].bv_page);
202    while (j--)
203        for (i = 0; i < RESYNC_PAGES ; i++)
204            safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
205    j = 0;
206out_free_bio:
207    for ( ; j < nalloc; j++) {
208        if (r10_bio->devs[j].bio)
209            bio_put(r10_bio->devs[j].bio);
210        if (r10_bio->devs[j].repl_bio)
211            bio_put(r10_bio->devs[j].repl_bio);
212    }
213    r10bio_pool_free(r10_bio, conf);
214    return NULL;
215}
216
217static void r10buf_pool_free(void *__r10_bio, void *data)
218{
219    int i;
220    struct r10conf *conf = data;
221    struct r10bio *r10bio = __r10_bio;
222    int j;
223
224    for (j=0; j < conf->copies; j++) {
225        struct bio *bio = r10bio->devs[j].bio;
226        if (bio) {
227            for (i = 0; i < RESYNC_PAGES; i++) {
228                safe_put_page(bio->bi_io_vec[i].bv_page);
229                bio->bi_io_vec[i].bv_page = NULL;
230            }
231            bio_put(bio);
232        }
233        bio = r10bio->devs[j].repl_bio;
234        if (bio)
235            bio_put(bio);
236    }
237    r10bio_pool_free(r10bio, conf);
238}
239
240static void put_all_bios(struct r10conf *conf, struct r10bio *r10_bio)
241{
242    int i;
243
244    for (i = 0; i < conf->copies; i++) {
245        struct bio **bio = & r10_bio->devs[i].bio;
246        if (!BIO_SPECIAL(*bio))
247            bio_put(*bio);
248        *bio = NULL;
249        bio = &r10_bio->devs[i].repl_bio;
250        if (r10_bio->read_slot < 0 && !BIO_SPECIAL(*bio))
251            bio_put(*bio);
252        *bio = NULL;
253    }
254}
255
256static void free_r10bio(struct r10bio *r10_bio)
257{
258    struct r10conf *conf = r10_bio->mddev->private;
259
260    put_all_bios(conf, r10_bio);
261    mempool_free(r10_bio, conf->r10bio_pool);
262}
263
264static void put_buf(struct r10bio *r10_bio)
265{
266    struct r10conf *conf = r10_bio->mddev->private;
267
268    mempool_free(r10_bio, conf->r10buf_pool);
269
270    lower_barrier(conf);
271}
272
273static void reschedule_retry(struct r10bio *r10_bio)
274{
275    unsigned long flags;
276    struct mddev *mddev = r10_bio->mddev;
277    struct r10conf *conf = mddev->private;
278
279    spin_lock_irqsave(&conf->device_lock, flags);
280    list_add(&r10_bio->retry_list, &conf->retry_list);
281    conf->nr_queued ++;
282    spin_unlock_irqrestore(&conf->device_lock, flags);
283
284    /* wake up frozen array... */
285    wake_up(&conf->wait_barrier);
286
287    md_wakeup_thread(mddev->thread);
288}
289
290/*
291 * raid_end_bio_io() is called when we have finished servicing a mirrored
292 * operation and are ready to return a success/failure code to the buffer
293 * cache layer.
294 */
295static void raid_end_bio_io(struct r10bio *r10_bio)
296{
297    struct bio *bio = r10_bio->master_bio;
298    int done;
299    struct r10conf *conf = r10_bio->mddev->private;
300
301    if (bio->bi_phys_segments) {
302        unsigned long flags;
303        spin_lock_irqsave(&conf->device_lock, flags);
304        bio->bi_phys_segments--;
305        done = (bio->bi_phys_segments == 0);
306        spin_unlock_irqrestore(&conf->device_lock, flags);
307    } else
308        done = 1;
309    if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
310        clear_bit(BIO_UPTODATE, &bio->bi_flags);
311    if (done) {
312        bio_endio(bio, 0);
313        /*
314         * Wake up any possible resync thread that waits for the device
315         * to go idle.
316         */
317        allow_barrier(conf);
318    }
319    free_r10bio(r10_bio);
320}
321
322/*
323 * Update disk head position estimator based on IRQ completion info.
324 */
325static inline void update_head_pos(int slot, struct r10bio *r10_bio)
326{
327    struct r10conf *conf = r10_bio->mddev->private;
328
329    conf->mirrors[r10_bio->devs[slot].devnum].head_position =
330        r10_bio->devs[slot].addr + (r10_bio->sectors);
331}
332
333/*
334 * Find the disk number which triggered given bio
335 */
336static int find_bio_disk(struct r10conf *conf, struct r10bio *r10_bio,
337             struct bio *bio, int *slotp, int *replp)
338{
339    int slot;
340    int repl = 0;
341
342    for (slot = 0; slot < conf->copies; slot++) {
343        if (r10_bio->devs[slot].bio == bio)
344            break;
345        if (r10_bio->devs[slot].repl_bio == bio) {
346            repl = 1;
347            break;
348        }
349    }
350
351    BUG_ON(slot == conf->copies);
352    update_head_pos(slot, r10_bio);
353
354    if (slotp)
355        *slotp = slot;
356    if (replp)
357        *replp = repl;
358    return r10_bio->devs[slot].devnum;
359}
360
361static void raid10_end_read_request(struct bio *bio, int error)
362{
363    int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
364    struct r10bio *r10_bio = bio->bi_private;
365    int slot, dev;
366    struct md_rdev *rdev;
367    struct r10conf *conf = r10_bio->mddev->private;
368
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        unsigned long flags;
396        spin_lock_irqsave(&conf->device_lock, flags);
397        if (!enough(conf, rdev->raid_disk))
398            uptodate = 1;
399        spin_unlock_irqrestore(&conf->device_lock, flags);
400    }
401    if (uptodate) {
402        raid_end_bio_io(r10_bio);
403        rdev_dec_pending(rdev, conf->mddev);
404    } else {
405        /*
406         * oops, read error - keep the refcount on the rdev
407         */
408        char b[BDEVNAME_SIZE];
409        printk_ratelimited(KERN_ERR
410                   "md/raid10:%s: %s: rescheduling sector %llu\n",
411                   mdname(conf->mddev),
412                   bdevname(rdev->bdev, b),
413                   (unsigned long long)r10_bio->sector);
414        set_bit(R10BIO_ReadError, &r10_bio->state);
415        reschedule_retry(r10_bio);
416    }
417}
418
419static void close_write(struct r10bio *r10_bio)
420{
421    /* clear the bitmap if all writes complete successfully */
422    bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
423            r10_bio->sectors,
424            !test_bit(R10BIO_Degraded, &r10_bio->state),
425            0);
426    md_write_end(r10_bio->mddev);
427}
428
429static void one_write_done(struct r10bio *r10_bio)
430{
431    if (atomic_dec_and_test(&r10_bio->remaining)) {
432        if (test_bit(R10BIO_WriteError, &r10_bio->state))
433            reschedule_retry(r10_bio);
434        else {
435            close_write(r10_bio);
436            if (test_bit(R10BIO_MadeGood, &r10_bio->state))
437                reschedule_retry(r10_bio);
438            else
439                raid_end_bio_io(r10_bio);
440        }
441    }
442}
443
444static void raid10_end_write_request(struct bio *bio, int error)
445{
446    int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
447    struct r10bio *r10_bio = bio->bi_private;
448    int dev;
449    int dec_rdev = 1;
450    struct r10conf *conf = r10_bio->mddev->private;
451    int slot, repl;
452    struct md_rdev *rdev = NULL;
453
454    dev = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
455
456    if (repl)
457        rdev = conf->mirrors[dev].replacement;
458    if (!rdev) {
459        smp_rmb();
460        repl = 0;
461        rdev = conf->mirrors[dev].rdev;
462    }
463    /*
464     * this branch is our 'one mirror IO has finished' event handler:
465     */
466    if (!uptodate) {
467        if (repl)
468            /* Never record new bad blocks to replacement,
469             * just fail it.
470             */
471            md_error(rdev->mddev, rdev);
472        else {
473            set_bit(WriteErrorSeen, &rdev->flags);
474            if (!test_and_set_bit(WantReplacement, &rdev->flags))
475                set_bit(MD_RECOVERY_NEEDED,
476                    &rdev->mddev->recovery);
477            set_bit(R10BIO_WriteError, &r10_bio->state);
478            dec_rdev = 0;
479        }
480    } else {
481        /*
482         * Set R10BIO_Uptodate in our master bio, so that
483         * we will return a good error code for to the higher
484         * levels even if IO on some other mirrored buffer fails.
485         *
486         * The 'master' represents the composite IO operation to
487         * user-side. So if something waits for IO, then it will
488         * wait for the 'master' bio.
489         */
490        sector_t first_bad;
491        int bad_sectors;
492
493        set_bit(R10BIO_Uptodate, &r10_bio->state);
494
495        /* Maybe we can clear some bad blocks. */
496        if (is_badblock(rdev,
497                r10_bio->devs[slot].addr,
498                r10_bio->sectors,
499                &first_bad, &bad_sectors)) {
500            bio_put(bio);
501            if (repl)
502                r10_bio->devs[slot].repl_bio = IO_MADE_GOOD;
503            else
504                r10_bio->devs[slot].bio = IO_MADE_GOOD;
505            dec_rdev = 0;
506            set_bit(R10BIO_MadeGood, &r10_bio->state);
507        }
508    }
509
510    /*
511     *
512     * Let's see if all mirrored write operations have finished
513     * already.
514     */
515    one_write_done(r10_bio);
516    if (dec_rdev)
517        rdev_dec_pending(rdev, conf->mddev);
518}
519
520/*
521 * RAID10 layout manager
522 * As well as the chunksize and raid_disks count, there are two
523 * parameters: near_copies and far_copies.
524 * near_copies * far_copies must be <= raid_disks.
525 * Normally one of these will be 1.
526 * If both are 1, we get raid0.
527 * If near_copies == raid_disks, we get raid1.
528 *
529 * Chunks are laid out in raid0 style with near_copies copies of the
530 * first chunk, followed by near_copies copies of the next chunk and
531 * so on.
532 * If far_copies > 1, then after 1/far_copies of the array has been assigned
533 * as described above, we start again with a device offset of near_copies.
534 * So we effectively have another copy of the whole array further down all
535 * the drives, but with blocks on different drives.
536 * With this layout, and block is never stored twice on the one device.
537 *
538 * raid10_find_phys finds the sector offset of a given virtual sector
539 * on each device that it is on.
540 *
541 * raid10_find_virt does the reverse mapping, from a device and a
542 * sector offset to a virtual address
543 */
544
545static void __raid10_find_phys(struct geom *geo, struct r10bio *r10bio)
546{
547    int n,f;
548    sector_t sector;
549    sector_t chunk;
550    sector_t stripe;
551    int dev;
552    int slot = 0;
553    int last_far_set_start, last_far_set_size;
554
555    last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
556    last_far_set_start *= geo->far_set_size;
557
558    last_far_set_size = geo->far_set_size;
559    last_far_set_size += (geo->raid_disks % geo->far_set_size);
560
561    /* now calculate first sector/dev */
562    chunk = r10bio->sector >> geo->chunk_shift;
563    sector = r10bio->sector & geo->chunk_mask;
564
565    chunk *= geo->near_copies;
566    stripe = chunk;
567    dev = sector_div(stripe, geo->raid_disks);
568    if (geo->far_offset)
569        stripe *= geo->far_copies;
570
571    sector += stripe << geo->chunk_shift;
572
573    /* and calculate all the others */
574    for (n = 0; n < geo->near_copies; n++) {
575        int d = dev;
576        int set;
577        sector_t s = sector;
578        r10bio->devs[slot].devnum = d;
579        r10bio->devs[slot].addr = s;
580        slot++;
581
582        for (f = 1; f < geo->far_copies; f++) {
583            set = d / geo->far_set_size;
584            d += geo->near_copies;
585
586            if ((geo->raid_disks % geo->far_set_size) &&
587                (d > last_far_set_start)) {
588                d -= last_far_set_start;
589                d %= last_far_set_size;
590                d += last_far_set_start;
591            } else {
592                d %= geo->far_set_size;
593                d += geo->far_set_size * set;
594            }
595            s += geo->stride;
596            r10bio->devs[slot].devnum = d;
597            r10bio->devs[slot].addr = s;
598            slot++;
599        }
600        dev++;
601        if (dev >= geo->raid_disks) {
602            dev = 0;
603            sector += (geo->chunk_mask + 1);
604        }
605    }
606}
607
608static void raid10_find_phys(struct r10conf *conf, struct r10bio *r10bio)
609{
610    struct geom *geo = &conf->geo;
611
612    if (conf->reshape_progress != MaxSector &&
613        ((r10bio->sector >= conf->reshape_progress) !=
614         conf->mddev->reshape_backwards)) {
615        set_bit(R10BIO_Previous, &r10bio->state);
616        geo = &conf->prev;
617    } else
618        clear_bit(R10BIO_Previous, &r10bio->state);
619
620    __raid10_find_phys(geo, r10bio);
621}
622
623static sector_t raid10_find_virt(struct r10conf *conf, sector_t sector, int dev)
624{
625    sector_t offset, chunk, vchunk;
626    /* Never use conf->prev as this is only called during resync
627     * or recovery, so reshape isn't happening
628     */
629    struct geom *geo = &conf->geo;
630    int far_set_start = (dev / geo->far_set_size) * geo->far_set_size;
631    int far_set_size = geo->far_set_size;
632    int last_far_set_start;
633
634    if (geo->raid_disks % geo->far_set_size) {
635        last_far_set_start = (geo->raid_disks / geo->far_set_size) - 1;
636        last_far_set_start *= geo->far_set_size;
637
638        if (dev >= last_far_set_start) {
639            far_set_size = geo->far_set_size;
640            far_set_size += (geo->raid_disks % geo->far_set_size);
641            far_set_start = last_far_set_start;
642        }
643    }
644
645    offset = sector & geo->chunk_mask;
646    if (geo->far_offset) {
647        int fc;
648        chunk = sector >> geo->chunk_shift;
649        fc = sector_div(chunk, geo->far_copies);
650        dev -= fc * geo->near_copies;
651        if (dev < far_set_start)
652            dev += far_set_size;
653    } else {
654        while (sector >= geo->stride) {
655            sector -= geo->stride;
656            if (dev < (geo->near_copies + far_set_start))
657                dev += far_set_size - geo->near_copies;
658            else
659                dev -= geo->near_copies;
660        }
661        chunk = sector >> geo->chunk_shift;
662    }
663    vchunk = chunk * geo->raid_disks + dev;
664    sector_div(vchunk, geo->near_copies);
665    return (vchunk << geo->chunk_shift) + offset;
666}
667
668/**
669 * raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
670 * @q: request queue
671 * @bvm: properties of new bio
672 * @biovec: the request that could be merged to it.
673 *
674 * Return amount of bytes we can accept at this offset
675 * This requires checking for end-of-chunk if near_copies != raid_disks,
676 * and for subordinate merge_bvec_fns if merge_check_needed.
677 */
678static int raid10_mergeable_bvec(struct request_queue *q,
679                 struct bvec_merge_data *bvm,
680                 struct bio_vec *biovec)
681{
682    struct mddev *mddev = q->queuedata;
683    struct r10conf *conf = mddev->private;
684    sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
685    int max;
686    unsigned int chunk_sectors;
687    unsigned int bio_sectors = bvm->bi_size >> 9;
688    struct geom *geo = &conf->geo;
689
690    chunk_sectors = (conf->geo.chunk_mask & conf->prev.chunk_mask) + 1;
691    if (conf->reshape_progress != MaxSector &&
692        ((sector >= conf->reshape_progress) !=
693         conf->mddev->reshape_backwards))
694        geo = &conf->prev;
695
696    if (geo->near_copies < geo->raid_disks) {
697        max = (chunk_sectors - ((sector & (chunk_sectors - 1))
698                    + bio_sectors)) << 9;
699        if (max < 0)
700            /* bio_add cannot handle a negative return */
701            max = 0;
702        if (max <= biovec->bv_len && bio_sectors == 0)
703            return biovec->bv_len;
704    } else
705        max = biovec->bv_len;
706
707    if (mddev->merge_check_needed) {
708        struct {
709            struct r10bio r10_bio;
710            struct r10dev devs[conf->copies];
711        } on_stack;
712        struct r10bio *r10_bio = &on_stack.r10_bio;
713        int s;
714        if (conf->reshape_progress != MaxSector) {
715            /* Cannot give any guidance during reshape */
716            if (max <= biovec->bv_len && bio_sectors == 0)
717                return biovec->bv_len;
718            return 0;
719        }
720        r10_bio->sector = sector;
721        raid10_find_phys(conf, r10_bio);
722        rcu_read_lock();
723        for (s = 0; s < conf->copies; s++) {
724            int disk = r10_bio->devs[s].devnum;
725            struct md_rdev *rdev = rcu_dereference(
726                conf->mirrors[disk].rdev);
727            if (rdev && !test_bit(Faulty, &rdev->flags)) {
728                struct request_queue *q =
729                    bdev_get_queue(rdev->bdev);
730                if (q->merge_bvec_fn) {
731                    bvm->bi_sector = r10_bio->devs[s].addr
732                        + rdev->data_offset;
733                    bvm->bi_bdev = rdev->bdev;
734                    max = min(max, q->merge_bvec_fn(
735                              q, bvm, biovec));
736                }
737            }
738            rdev = rcu_dereference(conf->mirrors[disk].replacement);
739            if (rdev && !test_bit(Faulty, &rdev->flags)) {
740                struct request_queue *q =
741                    bdev_get_queue(rdev->bdev);
742                if (q->merge_bvec_fn) {
743                    bvm->bi_sector = r10_bio->devs[s].addr
744                        + rdev->data_offset;
745                    bvm->bi_bdev = rdev->bdev;
746                    max = min(max, q->merge_bvec_fn(
747                              q, bvm, biovec));
748                }
749            }
750        }
751        rcu_read_unlock();
752    }
753    return max;
754}
755
756/*
757 * This routine returns the disk from which the requested read should
758 * be done. There is a per-array 'next expected sequential IO' sector
759 * number - if this matches on the next IO then we use the last disk.
760 * There is also a per-disk 'last know head position' sector that is
761 * maintained from IRQ contexts, both the normal and the resync IO
762 * completion handlers update this position correctly. If there is no
763 * perfect sequential match then we pick the disk whose head is closest.
764 *
765 * If there are 2 mirrors in the same 2 devices, performance degrades
766 * because position is mirror, not device based.
767 *
768 * The rdev for the device selected will have nr_pending incremented.
769 */
770
771/*
772 * FIXME: possibly should rethink readbalancing and do it differently
773 * depending on near_copies / far_copies geometry.
774 */
775static struct md_rdev *read_balance(struct r10conf *conf,
776                    struct r10bio *r10_bio,
777                    int *max_sectors)
778{
779    const sector_t this_sector = r10_bio->sector;
780    int disk, slot;
781    int sectors = r10_bio->sectors;
782    int best_good_sectors;
783    sector_t new_distance, best_dist;
784    struct md_rdev *best_rdev, *rdev = NULL;
785    int do_balance;
786    int best_slot;
787    struct geom *geo = &conf->geo;
788
789    raid10_find_phys(conf, r10_bio);
790    rcu_read_lock();
791retry:
792    sectors = r10_bio->sectors;
793    best_slot = -1;
794    best_rdev = NULL;
795    best_dist = MaxSector;
796    best_good_sectors = 0;
797    do_balance = 1;
798    /*
799     * Check if we can balance. We can balance on the whole
800     * device if no resync is going on (recovery is ok), or below
801     * the resync window. We take the first readable disk when
802     * above the resync window.
803     */
804    if (conf->mddev->recovery_cp < MaxSector
805        && (this_sector + sectors >= conf->next_resync))
806        do_balance = 0;
807
808    for (slot = 0; slot < conf->copies ; slot++) {
809        sector_t first_bad;
810        int bad_sectors;
811        sector_t dev_sector;
812
813        if (r10_bio->devs[slot].bio == IO_BLOCKED)
814            continue;
815        disk = r10_bio->devs[slot].devnum;
816        rdev = rcu_dereference(conf->mirrors[disk].replacement);
817        if (rdev == NULL || test_bit(Faulty, &rdev->flags) ||
818            test_bit(Unmerged, &rdev->flags) ||
819            r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
820            rdev = rcu_dereference(conf->mirrors[disk].rdev);
821        if (rdev == NULL ||
822            test_bit(Faulty, &rdev->flags) ||
823            test_bit(Unmerged, &rdev->flags))
824            continue;
825        if (!test_bit(In_sync, &rdev->flags) &&
826            r10_bio->devs[slot].addr + sectors > rdev->recovery_offset)
827            continue;
828
829        dev_sector = r10_bio->devs[slot].addr;
830        if (is_badblock(rdev, dev_sector, sectors,
831                &first_bad, &bad_sectors)) {
832            if (best_dist < MaxSector)
833                /* Already have a better slot */
834                continue;
835            if (first_bad <= dev_sector) {
836                /* Cannot read here. If this is the
837                 * 'primary' device, then we must not read
838                 * beyond 'bad_sectors' from another device.
839                 */
840                bad_sectors -= (dev_sector - first_bad);
841                if (!do_balance && sectors > bad_sectors)
842                    sectors = bad_sectors;
843                if (best_good_sectors > sectors)
844                    best_good_sectors = sectors;
845            } else {
846                sector_t good_sectors =
847                    first_bad - dev_sector;
848                if (good_sectors > best_good_sectors) {
849                    best_good_sectors = good_sectors;
850                    best_slot = slot;
851                    best_rdev = rdev;
852                }
853                if (!do_balance)
854                    /* Must read from here */
855                    break;
856            }
857            continue;
858        } else
859            best_good_sectors = sectors;
860
861        if (!do_balance)
862            break;
863
864        /* This optimisation is debatable, and completely destroys
865         * sequential read speed for 'far copies' arrays. So only
866         * keep it for 'near' arrays, and review those later.
867         */
868        if (geo->near_copies > 1 && !atomic_read(&rdev->nr_pending))
869            break;
870
871        /* for far > 1 always use the lowest address */
872        if (geo->far_copies > 1)
873            new_distance = r10_bio->devs[slot].addr;
874        else
875            new_distance = abs(r10_bio->devs[slot].addr -
876                       conf->mirrors[disk].head_position);
877        if (new_distance < best_dist) {
878            best_dist = new_distance;
879            best_slot = slot;
880            best_rdev = rdev;
881        }
882    }
883    if (slot >= conf->copies) {
884        slot = best_slot;
885        rdev = best_rdev;
886    }
887
888    if (slot >= 0) {
889        atomic_inc(&rdev->nr_pending);
890        if (test_bit(Faulty, &rdev->flags)) {
891            /* Cannot risk returning a device that failed
892             * before we inc'ed nr_pending
893             */
894            rdev_dec_pending(rdev, conf->mddev);
895            goto retry;
896        }
897        r10_bio->read_slot = slot;
898    } else
899        rdev = NULL;
900    rcu_read_unlock();
901    *max_sectors = best_good_sectors;
902
903    return rdev;
904}
905
906int md_raid10_congested(struct mddev *mddev, int bits)
907{
908    struct r10conf *conf = mddev->private;
909    int i, ret = 0;
910
911    if ((bits & (1 << BDI_async_congested)) &&
912        conf->pending_count >= max_queued_requests)
913        return 1;
914
915    rcu_read_lock();
916    for (i = 0;
917         (i < conf->geo.raid_disks || i < conf->prev.raid_disks)
918             && ret == 0;
919         i++) {
920        struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
921        if (rdev && !test_bit(Faulty, &rdev->flags)) {
922            struct request_queue *q = bdev_get_queue(rdev->bdev);
923
924            ret |= bdi_congested(&q->backing_dev_info, bits);
925        }
926    }
927    rcu_read_unlock();
928    return ret;
929}
930EXPORT_SYMBOL_GPL(md_raid10_congested);
931
932static int raid10_congested(void *data, int bits)
933{
934    struct mddev *mddev = data;
935
936    return mddev_congested(mddev, bits) ||
937        md_raid10_congested(mddev, bits);
938}
939
940static void flush_pending_writes(struct r10conf *conf)
941{
942    /* Any writes that have been queued but are awaiting
943     * bitmap updates get flushed here.
944     */
945    spin_lock_irq(&conf->device_lock);
946
947    if (conf->pending_bio_list.head) {
948        struct bio *bio;
949        bio = bio_list_get(&conf->pending_bio_list);
950        conf->pending_count = 0;
951        spin_unlock_irq(&conf->device_lock);
952        /* flush any pending bitmap writes to disk
953         * before proceeding w/ I/O */
954        bitmap_unplug(conf->mddev->bitmap);
955        wake_up(&conf->wait_barrier);
956
957        while (bio) { /* submit pending writes */
958            struct bio *next = bio->bi_next;
959            bio->bi_next = NULL;
960            if (unlikely((bio->bi_rw & REQ_DISCARD) &&
961                !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
962                /* Just ignore it */
963                bio_endio(bio, 0);
964            else
965                generic_make_request(bio);
966            bio = next;
967        }
968    } else
969        spin_unlock_irq(&conf->device_lock);
970}
971
972/* Barriers....
973 * Sometimes we need to suspend IO while we do something else,
974 * either some resync/recovery, or reconfigure the array.
975 * To do this we raise a 'barrier'.
976 * The 'barrier' is a counter that can be raised multiple times
977 * to count how many activities are happening which preclude
978 * normal IO.
979 * We can only raise the barrier if there is no pending IO.
980 * i.e. if nr_pending == 0.
981 * We choose only to raise the barrier if no-one is waiting for the
982 * barrier to go down. This means that as soon as an IO request
983 * is ready, no other operations which require a barrier will start
984 * until the IO request has had a chance.
985 *
986 * So: regular IO calls 'wait_barrier'. When that returns there
987 * is no backgroup IO happening, It must arrange to call
988 * allow_barrier when it has finished its IO.
989 * backgroup IO calls must call raise_barrier. Once that returns
990 * there is no normal IO happeing. It must arrange to call
991 * lower_barrier when the particular background IO completes.
992 */
993
994static void raise_barrier(struct r10conf *conf, int force)
995{
996    BUG_ON(force && !conf->barrier);
997    spin_lock_irq(&conf->resync_lock);
998
999    /* Wait until no block IO is waiting (unless 'force') */
1000    wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
1001                conf->resync_lock);
1002
1003    /* block any new IO from starting */
1004    conf->barrier++;
1005
1006    /* Now wait for all pending IO to complete */
1007    wait_event_lock_irq(conf->wait_barrier,
1008                !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
1009                conf->resync_lock);
1010
1011    spin_unlock_irq(&conf->resync_lock);
1012}
1013
1014static void lower_barrier(struct r10conf *conf)
1015{
1016    unsigned long flags;
1017    spin_lock_irqsave(&conf->resync_lock, flags);
1018    conf->barrier--;
1019    spin_unlock_irqrestore(&conf->resync_lock, flags);
1020    wake_up(&conf->wait_barrier);
1021}
1022
1023static void wait_barrier(struct r10conf *conf)
1024{
1025    spin_lock_irq(&conf->resync_lock);
1026    if (conf->barrier) {
1027        conf->nr_waiting++;
1028        /* Wait for the barrier to drop.
1029         * However if there are already pending
1030         * requests (preventing the barrier from
1031         * rising completely), and the
1032         * pre-process bio queue isn't empty,
1033         * then don't wait, as we need to empty
1034         * that queue to get the nr_pending
1035         * count down.
1036         */
1037        wait_event_lock_irq(conf->wait_barrier,
1038                    !conf->barrier ||
1039                    (conf->nr_pending &&
1040                     current->bio_list &&
1041                     !bio_list_empty(current->bio_list)),
1042                    conf->resync_lock);
1043        conf->nr_waiting--;
1044    }
1045    conf->nr_pending++;
1046    spin_unlock_irq(&conf->resync_lock);
1047}
1048
1049static void allow_barrier(struct r10conf *conf)
1050{
1051    unsigned long flags;
1052    spin_lock_irqsave(&conf->resync_lock, flags);
1053    conf->nr_pending--;
1054    spin_unlock_irqrestore(&conf->resync_lock, flags);
1055    wake_up(&conf->wait_barrier);
1056}
1057
1058static void freeze_array(struct r10conf *conf)
1059{
1060    /* stop syncio and normal IO and wait for everything to
1061     * go quiet.
1062     * We increment barrier and nr_waiting, and then
1063     * wait until nr_pending match nr_queued+1
1064     * This is called in the context of one normal IO request
1065     * that has failed. Thus any sync request that might be pending
1066     * will be blocked by nr_pending, and we need to wait for
1067     * pending IO requests to complete or be queued for re-try.
1068     * Thus the number queued (nr_queued) plus this request (1)
1069     * must match the number of pending IOs (nr_pending) before
1070     * we continue.
1071     */
1072    spin_lock_irq(&conf->resync_lock);
1073    conf->barrier++;
1074    conf->nr_waiting++;
1075    wait_event_lock_irq_cmd(conf->wait_barrier,
1076                conf->nr_pending == conf->nr_queued+1,
1077                conf->resync_lock,
1078                flush_pending_writes(conf));
1079
1080    spin_unlock_irq(&conf->resync_lock);
1081}
1082
1083static void unfreeze_array(struct r10conf *conf)
1084{
1085    /* reverse the effect of the freeze */
1086    spin_lock_irq(&conf->resync_lock);
1087    conf->barrier--;
1088    conf->nr_waiting--;
1089    wake_up(&conf->wait_barrier);
1090    spin_unlock_irq(&conf->resync_lock);
1091}
1092
1093static sector_t choose_data_offset(struct r10bio *r10_bio,
1094                   struct md_rdev *rdev)
1095{
1096    if (!test_bit(MD_RECOVERY_RESHAPE, &rdev->mddev->recovery) ||
1097        test_bit(R10BIO_Previous, &r10_bio->state))
1098        return rdev->data_offset;
1099    else
1100        return rdev->new_data_offset;
1101}
1102
1103struct raid10_plug_cb {
1104    struct blk_plug_cb cb;
1105    struct bio_list pending;
1106    int pending_cnt;
1107};
1108
1109static void raid10_unplug(struct blk_plug_cb *cb, bool from_schedule)
1110{
1111    struct raid10_plug_cb *plug = container_of(cb, struct raid10_plug_cb,
1112                           cb);
1113    struct mddev *mddev = plug->cb.data;
1114    struct r10conf *conf = mddev->private;
1115    struct bio *bio;
1116
1117    if (from_schedule || current->bio_list) {
1118        spin_lock_irq(&conf->device_lock);
1119        bio_list_merge(&conf->pending_bio_list, &plug->pending);
1120        conf->pending_count += plug->pending_cnt;
1121        spin_unlock_irq(&conf->device_lock);
1122        wake_up(&conf->wait_barrier);
1123        md_wakeup_thread(mddev->thread);
1124        kfree(plug);
1125        return;
1126    }
1127
1128    /* we aren't scheduling, so we can do the write-out directly. */
1129    bio = bio_list_get(&plug->pending);
1130    bitmap_unplug(mddev->bitmap);
1131    wake_up(&conf->wait_barrier);
1132
1133    while (bio) { /* submit pending writes */
1134        struct bio *next = bio->bi_next;
1135        bio->bi_next = NULL;
1136        generic_make_request(bio);
1137        bio = next;
1138    }
1139    kfree(plug);
1140}
1141
1142static void make_request(struct mddev *mddev, struct bio * bio)
1143{
1144    struct r10conf *conf = mddev->private;
1145    struct r10bio *r10_bio;
1146    struct bio *read_bio;
1147    int i;
1148    sector_t chunk_mask = (conf->geo.chunk_mask & conf->prev.chunk_mask);
1149    int chunk_sects = chunk_mask + 1;
1150    const int rw = bio_data_dir(bio);
1151    const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1152    const unsigned long do_fua = (bio->bi_rw & REQ_FUA);
1153    const unsigned long do_discard = (bio->bi_rw
1154                      & (REQ_DISCARD | REQ_SECURE));
1155    const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1156    unsigned long flags;
1157    struct md_rdev *blocked_rdev;
1158    struct blk_plug_cb *cb;
1159    struct raid10_plug_cb *plug = NULL;
1160    int sectors_handled;
1161    int max_sectors;
1162    int sectors;
1163
1164    if (unlikely(bio->bi_rw & REQ_FLUSH)) {
1165        md_flush_request(mddev, bio);
1166        return;
1167    }
1168
1169    /* If this request crosses a chunk boundary, we need to
1170     * split it. This will only happen for 1 PAGE (or less) requests.
1171     */
1172    if (unlikely((bio->bi_sector & chunk_mask) + (bio->bi_size >> 9)
1173             > chunk_sects
1174             && (conf->geo.near_copies < conf->geo.raid_disks
1175             || conf->prev.near_copies < conf->prev.raid_disks))) {
1176        struct bio_pair *bp;
1177        /* Sanity check -- queue functions should prevent this happening */
1178        if ((bio->bi_vcnt != 1 && bio->bi_vcnt != 0) ||
1179            bio->bi_idx != 0)
1180            goto bad_map;
1181        /* This is a one page bio that upper layers
1182         * refuse to split for us, so we need to split it.
1183         */
1184        bp = bio_split(bio,
1185                   chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
1186
1187        /* Each of these 'make_request' calls will call 'wait_barrier'.
1188         * If the first succeeds but the second blocks due to the resync
1189         * thread raising the barrier, we will deadlock because the
1190         * IO to the underlying device will be queued in generic_make_request
1191         * and will never complete, so will never reduce nr_pending.
1192         * So increment nr_waiting here so no new raise_barriers will
1193         * succeed, and so the second wait_barrier cannot block.
1194         */
1195        spin_lock_irq(&conf->resync_lock);
1196        conf->nr_waiting++;
1197        spin_unlock_irq(&conf->resync_lock);
1198
1199        make_request(mddev, &bp->bio1);
1200        make_request(mddev, &bp->bio2);
1201
1202        spin_lock_irq(&conf->resync_lock);
1203        conf->nr_waiting--;
1204        wake_up(&conf->wait_barrier);
1205        spin_unlock_irq(&conf->resync_lock);
1206
1207        bio_pair_release(bp);
1208        return;
1209    bad_map:
1210        printk("md/raid10:%s: make_request bug: can't convert block across chunks"
1211               " or bigger than %dk %llu %d\n", mdname(mddev), chunk_sects/2,
1212               (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
1213
1214        bio_io_error(bio);
1215        return;
1216    }
1217
1218    md_write_start(mddev, bio);
1219
1220    /*
1221     * Register the new request and wait if the reconstruction
1222     * thread has put up a bar for new requests.
1223     * Continue immediately if no resync is active currently.
1224     */
1225    wait_barrier(conf);
1226
1227    sectors = bio->bi_size >> 9;
1228    while (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1229        bio->bi_sector < conf->reshape_progress &&
1230        bio->bi_sector + sectors > conf->reshape_progress) {
1231        /* IO spans the reshape position. Need to wait for
1232         * reshape to pass
1233         */
1234        allow_barrier(conf);
1235        wait_event(conf->wait_barrier,
1236               conf->reshape_progress <= bio->bi_sector ||
1237               conf->reshape_progress >= bio->bi_sector + sectors);
1238        wait_barrier(conf);
1239    }
1240    if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery) &&
1241        bio_data_dir(bio) == WRITE &&
1242        (mddev->reshape_backwards
1243         ? (bio->bi_sector < conf->reshape_safe &&
1244        bio->bi_sector + sectors > conf->reshape_progress)
1245         : (bio->bi_sector + sectors > conf->reshape_safe &&
1246        bio->bi_sector < conf->reshape_progress))) {
1247        /* Need to update reshape_position in metadata */
1248        mddev->reshape_position = conf->reshape_progress;
1249        set_bit(MD_CHANGE_DEVS, &mddev->flags);
1250        set_bit(MD_CHANGE_PENDING, &mddev->flags);
1251        md_wakeup_thread(mddev->thread);
1252        wait_event(mddev->sb_wait,
1253               !test_bit(MD_CHANGE_PENDING, &mddev->flags));
1254
1255        conf->reshape_safe = mddev->reshape_position;
1256    }
1257
1258    r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1259
1260    r10_bio->master_bio = bio;
1261    r10_bio->sectors = sectors;
1262
1263    r10_bio->mddev = mddev;
1264    r10_bio->sector = bio->bi_sector;
1265    r10_bio->state = 0;
1266
1267    /* We might need to issue multiple reads to different
1268     * devices if there are bad blocks around, so we keep
1269     * track of the number of reads in bio->bi_phys_segments.
1270     * If this is 0, there is only one r10_bio and no locking
1271     * will be needed when the request completes. If it is
1272     * non-zero, then it is the number of not-completed requests.
1273     */
1274    bio->bi_phys_segments = 0;
1275    clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1276
1277    if (rw == READ) {
1278        /*
1279         * read balancing logic:
1280         */
1281        struct md_rdev *rdev;
1282        int slot;
1283
1284read_again:
1285        rdev = read_balance(conf, r10_bio, &max_sectors);
1286        if (!rdev) {
1287            raid_end_bio_io(r10_bio);
1288            return;
1289        }
1290        slot = r10_bio->read_slot;
1291
1292        read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1293        md_trim_bio(read_bio, r10_bio->sector - bio->bi_sector,
1294                max_sectors);
1295
1296        r10_bio->devs[slot].bio = read_bio;
1297        r10_bio->devs[slot].rdev = rdev;
1298
1299        read_bio->bi_sector = r10_bio->devs[slot].addr +
1300            choose_data_offset(r10_bio, rdev);
1301        read_bio->bi_bdev = rdev->bdev;
1302        read_bio->bi_end_io = raid10_end_read_request;
1303        read_bio->bi_rw = READ | do_sync;
1304        read_bio->bi_private = r10_bio;
1305
1306        if (max_sectors < r10_bio->sectors) {
1307            /* Could not read all from this device, so we will
1308             * need another r10_bio.
1309             */
1310            sectors_handled = (r10_bio->sectors + max_sectors
1311                       - bio->bi_sector);
1312            r10_bio->sectors = max_sectors;
1313            spin_lock_irq(&conf->device_lock);
1314            if (bio->bi_phys_segments == 0)
1315                bio->bi_phys_segments = 2;
1316            else
1317                bio->bi_phys_segments++;
1318            spin_unlock(&conf->device_lock);
1319            /* Cannot call generic_make_request directly
1320             * as that will be queued in __generic_make_request
1321             * and subsequent mempool_alloc might block
1322             * waiting for it. so hand bio over to raid10d.
1323             */
1324            reschedule_retry(r10_bio);
1325
1326            r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1327
1328            r10_bio->master_bio = bio;
1329            r10_bio->sectors = ((bio->bi_size >> 9)
1330                        - sectors_handled);
1331            r10_bio->state = 0;
1332            r10_bio->mddev = mddev;
1333            r10_bio->sector = bio->bi_sector + sectors_handled;
1334            goto read_again;
1335        } else
1336            generic_make_request(read_bio);
1337        return;
1338    }
1339
1340    /*
1341     * WRITE:
1342     */
1343    if (conf->pending_count >= max_queued_requests) {
1344        md_wakeup_thread(mddev->thread);
1345        wait_event(conf->wait_barrier,
1346               conf->pending_count < max_queued_requests);
1347    }
1348    /* first select target devices under rcu_lock and
1349     * inc refcount on their rdev. Record them by setting
1350     * bios[x] to bio
1351     * If there are known/acknowledged bad blocks on any device
1352     * on which we have seen a write error, we want to avoid
1353     * writing to those blocks. This potentially requires several
1354     * writes to write around the bad blocks. Each set of writes
1355     * gets its own r10_bio with a set of bios attached. The number
1356     * of r10_bios is recored in bio->bi_phys_segments just as with
1357     * the read case.
1358     */
1359
1360    r10_bio->read_slot = -1; /* make sure repl_bio gets freed */
1361    raid10_find_phys(conf, r10_bio);
1362retry_write:
1363    blocked_rdev = NULL;
1364    rcu_read_lock();
1365    max_sectors = r10_bio->sectors;
1366
1367    for (i = 0; i < conf->copies; i++) {
1368        int d = r10_bio->devs[i].devnum;
1369        struct md_rdev *rdev = rcu_dereference(conf->mirrors[d].rdev);
1370        struct md_rdev *rrdev = rcu_dereference(
1371            conf->mirrors[d].replacement);
1372        if (rdev == rrdev)
1373            rrdev = NULL;
1374        if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1375            atomic_inc(&rdev->nr_pending);
1376            blocked_rdev = rdev;
1377            break;
1378        }
1379        if (rrdev && unlikely(test_bit(Blocked, &rrdev->flags))) {
1380            atomic_inc(&rrdev->nr_pending);
1381            blocked_rdev = rrdev;
1382            break;
1383        }
1384        if (rdev && (test_bit(Faulty, &rdev->flags)
1385                 || test_bit(Unmerged, &rdev->flags)))
1386            rdev = NULL;
1387        if (rrdev && (test_bit(Faulty, &rrdev->flags)
1388                  || test_bit(Unmerged, &rrdev->flags)))
1389            rrdev = NULL;
1390
1391        r10_bio->devs[i].bio = NULL;
1392        r10_bio->devs[i].repl_bio = NULL;
1393
1394        if (!rdev && !rrdev) {
1395            set_bit(R10BIO_Degraded, &r10_bio->state);
1396            continue;
1397        }
1398        if (rdev && test_bit(WriteErrorSeen, &rdev->flags)) {
1399            sector_t first_bad;
1400            sector_t dev_sector = r10_bio->devs[i].addr;
1401            int bad_sectors;
1402            int is_bad;
1403
1404            is_bad = is_badblock(rdev, dev_sector,
1405                         max_sectors,
1406                         &first_bad, &bad_sectors);
1407            if (is_bad < 0) {
1408                /* Mustn't write here until the bad block
1409                 * is acknowledged
1410                 */
1411                atomic_inc(&rdev->nr_pending);
1412                set_bit(BlockedBadBlocks, &rdev->flags);
1413                blocked_rdev = rdev;
1414                break;
1415            }
1416            if (is_bad && first_bad <= dev_sector) {
1417                /* Cannot write here at all */
1418                bad_sectors -= (dev_sector - first_bad);
1419                if (bad_sectors < max_sectors)
1420                    /* Mustn't write more than bad_sectors
1421                     * to other devices yet
1422                     */
1423                    max_sectors = bad_sectors;
1424                /* We don't set R10BIO_Degraded as that
1425                 * only applies if the disk is missing,
1426                 * so it might be re-added, and we want to
1427                 * know to recover this chunk.
1428                 * In this case the device is here, and the
1429                 * fact that this chunk is not in-sync is
1430                 * recorded in the bad block log.
1431                 */
1432                continue;
1433            }
1434            if (is_bad) {
1435                int good_sectors = first_bad - dev_sector;
1436                if (good_sectors < max_sectors)
1437                    max_sectors = good_sectors;
1438            }
1439        }
1440        if (rdev) {
1441            r10_bio->devs[i].bio = bio;
1442            atomic_inc(&rdev->nr_pending);
1443        }
1444        if (rrdev) {
1445            r10_bio->devs[i].repl_bio = bio;
1446            atomic_inc(&rrdev->nr_pending);
1447        }
1448    }
1449    rcu_read_unlock();
1450
1451    if (unlikely(blocked_rdev)) {
1452        /* Have to wait for this device to get unblocked, then retry */
1453        int j;
1454        int d;
1455
1456        for (j = 0; j < i; j++) {
1457            if (r10_bio->devs[j].bio) {
1458                d = r10_bio->devs[j].devnum;
1459                rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1460            }
1461            if (r10_bio->devs[j].repl_bio) {
1462                struct md_rdev *rdev;
1463                d = r10_bio->devs[j].devnum;
1464                rdev = conf->mirrors[d].replacement;
1465                if (!rdev) {
1466                    /* Race with remove_disk */
1467                    smp_mb();
1468                    rdev = conf->mirrors[d].rdev;
1469                }
1470                rdev_dec_pending(rdev, mddev);
1471            }
1472        }
1473        allow_barrier(conf);
1474        md_wait_for_blocked_rdev(blocked_rdev, mddev);
1475        wait_barrier(conf);
1476        goto retry_write;
1477    }
1478
1479    if (max_sectors < r10_bio->sectors) {
1480        /* We are splitting this into multiple parts, so
1481         * we need to prepare for allocating another r10_bio.
1482         */
1483        r10_bio->sectors = max_sectors;
1484        spin_lock_irq(&conf->device_lock);
1485        if (bio->bi_phys_segments == 0)
1486            bio->bi_phys_segments = 2;
1487        else
1488            bio->bi_phys_segments++;
1489        spin_unlock_irq(&conf->device_lock);
1490    }
1491    sectors_handled = r10_bio->sector + max_sectors - bio->bi_sector;
1492
1493    atomic_set(&r10_bio->remaining, 1);
1494    bitmap_startwrite(mddev->bitmap, r10_bio->sector, r10_bio->sectors, 0);
1495
1496    for (i = 0; i < conf->copies; i++) {
1497        struct bio *mbio;
1498        int d = r10_bio->devs[i].devnum;
1499        if (r10_bio->devs[i].bio) {
1500            struct md_rdev *rdev = conf->mirrors[d].rdev;
1501            mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1502            md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1503                    max_sectors);
1504            r10_bio->devs[i].bio = mbio;
1505
1506            mbio->bi_sector = (r10_bio->devs[i].addr+
1507                       choose_data_offset(r10_bio,
1508                                  rdev));
1509            mbio->bi_bdev = rdev->bdev;
1510            mbio->bi_end_io = raid10_end_write_request;
1511            mbio->bi_rw =
1512                WRITE | do_sync | do_fua | do_discard | do_same;
1513            mbio->bi_private = r10_bio;
1514
1515            atomic_inc(&r10_bio->remaining);
1516
1517            cb = blk_check_plugged(raid10_unplug, mddev,
1518                           sizeof(*plug));
1519            if (cb)
1520                plug = container_of(cb, struct raid10_plug_cb,
1521                            cb);
1522            else
1523                plug = NULL;
1524            spin_lock_irqsave(&conf->device_lock, flags);
1525            if (plug) {
1526                bio_list_add(&plug->pending, mbio);
1527                plug->pending_cnt++;
1528            } else {
1529                bio_list_add(&conf->pending_bio_list, mbio);
1530                conf->pending_count++;
1531            }
1532            spin_unlock_irqrestore(&conf->device_lock, flags);
1533            if (!plug)
1534                md_wakeup_thread(mddev->thread);
1535        }
1536
1537        if (r10_bio->devs[i].repl_bio) {
1538            struct md_rdev *rdev = conf->mirrors[d].replacement;
1539            if (rdev == NULL) {
1540                /* Replacement just got moved to main 'rdev' */
1541                smp_mb();
1542                rdev = conf->mirrors[d].rdev;
1543            }
1544            mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1545            md_trim_bio(mbio, r10_bio->sector - bio->bi_sector,
1546                    max_sectors);
1547            r10_bio->devs[i].repl_bio = mbio;
1548
1549            mbio->bi_sector = (r10_bio->devs[i].addr +
1550                       choose_data_offset(
1551                           r10_bio, rdev));
1552            mbio->bi_bdev = rdev->bdev;
1553            mbio->bi_end_io = raid10_end_write_request;
1554            mbio->bi_rw =
1555                WRITE | do_sync | do_fua | do_discard | do_same;
1556            mbio->bi_private = r10_bio;
1557
1558            atomic_inc(&r10_bio->remaining);
1559            spin_lock_irqsave(&conf->device_lock, flags);
1560            bio_list_add(&conf->pending_bio_list, mbio);
1561            conf->pending_count++;
1562            spin_unlock_irqrestore(&conf->device_lock, flags);
1563            if (!mddev_check_plugged(mddev))
1564                md_wakeup_thread(mddev->thread);
1565        }
1566    }
1567
1568    /* Don't remove the bias on 'remaining' (one_write_done) until
1569     * after checking if we need to go around again.
1570     */
1571
1572    if (sectors_handled < (bio->bi_size >> 9)) {
1573        one_write_done(r10_bio);
1574        /* We need another r10_bio. It has already been counted
1575         * in bio->bi_phys_segments.
1576         */
1577        r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
1578
1579        r10_bio->master_bio = bio;
1580        r10_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1581
1582        r10_bio->mddev = mddev;
1583        r10_bio->sector = bio->bi_sector + sectors_handled;
1584        r10_bio->state = 0;
1585        goto retry_write;
1586    }
1587    one_write_done(r10_bio);
1588
1589    /* In case raid10d snuck in to freeze_array */
1590    wake_up(&conf->wait_barrier);
1591}
1592
1593static void status(struct seq_file *seq, struct mddev *mddev)
1594{
1595    struct r10conf *conf = mddev->private;
1596    int i;
1597
1598    if (conf->geo.near_copies < conf->geo.raid_disks)
1599        seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
1600    if (conf->geo.near_copies > 1)
1601        seq_printf(seq, " %d near-copies", conf->geo.near_copies);
1602    if (conf->geo.far_copies > 1) {
1603        if (conf->geo.far_offset)
1604            seq_printf(seq, " %d offset-copies", conf->geo.far_copies);
1605        else
1606            seq_printf(seq, " %d far-copies", conf->geo.far_copies);
1607    }
1608    seq_printf(seq, " [%d/%d] [", conf->geo.raid_disks,
1609                    conf->geo.raid_disks - mddev->degraded);
1610    for (i = 0; i < conf->geo.raid_disks; i++)
1611        seq_printf(seq, "%s",
1612                  conf->mirrors[i].rdev &&
1613                  test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1614    seq_printf(seq, "]");
1615}
1616
1617/* check if there are enough drives for
1618 * every block to appear on atleast one.
1619 * Don't consider the device numbered 'ignore'
1620 * as we might be about to remove it.
1621 */
1622static int _enough(struct r10conf *conf, struct geom *geo, int ignore)
1623{
1624    int first = 0;
1625
1626    do {
1627        int n = conf->copies;
1628        int cnt = 0;
1629        int this = first;
1630        while (n--) {
1631            if (conf->mirrors[this].rdev &&
1632                this != ignore)
1633                cnt++;
1634            this = (this+1) % geo->raid_disks;
1635        }
1636        if (cnt == 0)
1637            return 0;
1638        first = (first + geo->near_copies) % geo->raid_disks;
1639    } while (first != 0);
1640    return 1;
1641}
1642
1643static int enough(struct r10conf *conf, int ignore)
1644{
1645    return _enough(conf, &conf->geo, ignore) &&
1646        _enough(conf, &conf->prev, ignore);
1647}
1648
1649static void error(struct mddev *mddev, struct md_rdev *rdev)
1650{
1651    char b[BDEVNAME_SIZE];
1652    struct r10conf *conf = mddev->private;
1653
1654    /*
1655     * If it is not operational, then we have already marked it as dead
1656     * else if it is the last working disks, ignore the error, let the
1657     * next level up know.
1658     * else mark the drive as failed
1659     */
1660    if (test_bit(In_sync, &rdev->flags)
1661        && !enough(conf, rdev->raid_disk))
1662        /*
1663         * Don't fail the drive, just return an IO error.
1664         */
1665        return;
1666    if (test_and_clear_bit(In_sync, &rdev->flags)) {
1667        unsigned long flags;
1668        spin_lock_irqsave(&conf->device_lock, flags);
1669        mddev->degraded++;
1670        spin_unlock_irqrestore(&conf->device_lock, flags);
1671        /*
1672         * if recovery is running, make sure it aborts.
1673         */
1674        set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1675    }
1676    set_bit(Blocked, &rdev->flags);
1677    set_bit(Faulty, &rdev->flags);
1678    set_bit(MD_CHANGE_DEVS, &mddev->flags);
1679    printk(KERN_ALERT
1680           "md/raid10:%s: Disk failure on %s, disabling device.\n"
1681           "md/raid10:%s: Operation continuing on %d devices.\n",
1682           mdname(mddev), bdevname(rdev->bdev, b),
1683           mdname(mddev), conf->geo.raid_disks - mddev->degraded);
1684}
1685
1686static void print_conf(struct r10conf *conf)
1687{
1688    int i;
1689    struct raid10_info *tmp;
1690
1691    printk(KERN_DEBUG "RAID10 conf printout:\n");
1692    if (!conf) {
1693        printk(KERN_DEBUG "(!conf)\n");
1694        return;
1695    }
1696    printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->geo.raid_disks - conf->mddev->degraded,
1697        conf->geo.raid_disks);
1698
1699    for (i = 0; i < conf->geo.raid_disks; i++) {
1700        char b[BDEVNAME_SIZE];
1701        tmp = conf->mirrors + i;
1702        if (tmp->rdev)
1703            printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1704                i, !test_bit(In_sync, &tmp->rdev->flags),
1705                    !test_bit(Faulty, &tmp->rdev->flags),
1706                bdevname(tmp->rdev->bdev,b));
1707    }
1708}
1709
1710static void close_sync(struct r10conf *conf)
1711{
1712    wait_barrier(conf);
1713    allow_barrier(conf);
1714
1715    mempool_destroy(conf->r10buf_pool);
1716    conf->r10buf_pool = NULL;
1717}
1718
1719static int raid10_spare_active(struct mddev *mddev)
1720{
1721    int i;
1722    struct r10conf *conf = mddev->private;
1723    struct raid10_info *tmp;
1724    int count = 0;
1725    unsigned long flags;
1726
1727    /*
1728     * Find all non-in_sync disks within the RAID10 configuration
1729     * and mark them in_sync
1730     */
1731    for (i = 0; i < conf->geo.raid_disks; i++) {
1732        tmp = conf->mirrors + i;
1733        if (tmp->replacement
1734            && tmp->replacement->recovery_offset == MaxSector
1735            && !test_bit(Faulty, &tmp->replacement->flags)
1736            && !test_and_set_bit(In_sync, &tmp->replacement->flags)) {
1737            /* Replacement has just become active */
1738            if (!tmp->rdev
1739                || !test_and_clear_bit(In_sync, &tmp->rdev->flags))
1740                count++;
1741            if (tmp->rdev) {
1742                /* Replaced device not technically faulty,
1743                 * but we need to be sure it gets removed
1744                 * and never re-added.
1745                 */
1746                set_bit(Faulty, &tmp->rdev->flags);
1747                sysfs_notify_dirent_safe(
1748                    tmp->rdev->sysfs_state);
1749            }
1750            sysfs_notify_dirent_safe(tmp->replacement->sysfs_state);
1751        } else if (tmp->rdev
1752               && !test_bit(Faulty, &tmp->rdev->flags)
1753               && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1754            count++;
1755            sysfs_notify_dirent_safe(tmp->rdev->sysfs_state);
1756        }
1757    }
1758    spin_lock_irqsave(&conf->device_lock, flags);
1759    mddev->degraded -= count;
1760    spin_unlock_irqrestore(&conf->device_lock, flags);
1761
1762    print_conf(conf);
1763    return count;
1764}
1765
1766
1767static int raid10_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1768{
1769    struct r10conf *conf = mddev->private;
1770    int err = -EEXIST;
1771    int mirror;
1772    int first = 0;
1773    int last = conf->geo.raid_disks - 1;
1774    struct request_queue *q = bdev_get_queue(rdev->bdev);
1775
1776    if (mddev->recovery_cp < MaxSector)
1777        /* only hot-add to in-sync arrays, as recovery is
1778         * very different from resync
1779         */
1780        return -EBUSY;
1781    if (rdev->saved_raid_disk < 0 && !_enough(conf, &conf->prev, -1))
1782        return -EINVAL;
1783
1784    if (rdev->raid_disk >= 0)
1785        first = last = rdev->raid_disk;
1786
1787    if (q->merge_bvec_fn) {
1788        set_bit(Unmerged, &rdev->flags);
1789        mddev->merge_check_needed = 1;
1790    }
1791
1792    if (rdev->saved_raid_disk >= first &&
1793        conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1794        mirror = rdev->saved_raid_disk;
1795    else
1796        mirror = first;
1797    for ( ; mirror <= last ; mirror++) {
1798        struct raid10_info *p = &conf->mirrors[mirror];
1799        if (p->recovery_disabled == mddev->recovery_disabled)
1800            continue;
1801        if (p->rdev) {
1802            if (!test_bit(WantReplacement, &p->rdev->flags) ||
1803                p->replacement != NULL)
1804                continue;
1805            clear_bit(In_sync, &rdev->flags);
1806            set_bit(Replacement, &rdev->flags);
1807            rdev->raid_disk = mirror;
1808            err = 0;
1809            disk_stack_limits(mddev->gendisk, rdev->bdev,
1810                      rdev->data_offset << 9);
1811            conf->fullsync = 1;
1812            rcu_assign_pointer(p->replacement, rdev);
1813            break;
1814        }
1815
1816        disk_stack_limits(mddev->gendisk, rdev->bdev,
1817                  rdev->data_offset << 9);
1818
1819        p->head_position = 0;
1820        p->recovery_disabled = mddev->recovery_disabled - 1;
1821        rdev->raid_disk = mirror;
1822        err = 0;
1823        if (rdev->saved_raid_disk != mirror)
1824            conf->fullsync = 1;
1825        rcu_assign_pointer(p->rdev, rdev);
1826        break;
1827    }
1828    if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1829        /* Some requests might not have seen this new
1830         * merge_bvec_fn. We must wait for them to complete
1831         * before merging the device fully.
1832         * First we make sure any code which has tested
1833         * our function has submitted the request, then
1834         * we wait for all outstanding requests to complete.
1835         */
1836        synchronize_sched();
1837        raise_barrier(conf, 0);
1838        lower_barrier(conf);
1839        clear_bit(Unmerged, &rdev->flags);
1840    }
1841    md_integrity_add_rdev(rdev, mddev);
1842    if (mddev->queue && blk_queue_discard(bdev_get_queue(rdev->bdev)))
1843        queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1844
1845    print_conf(conf);
1846    return err;
1847}
1848
1849static int raid10_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1850{
1851    struct r10conf *conf = mddev->private;
1852    int err = 0;
1853    int number = rdev->raid_disk;
1854    struct md_rdev **rdevp;
1855    struct raid10_info *p = conf->mirrors + number;
1856
1857    print_conf(conf);
1858    if (rdev == p->rdev)
1859        rdevp = &p->rdev;
1860    else if (rdev == p->replacement)
1861        rdevp = &p->replacement;
1862    else
1863        return 0;
1864
1865    if (test_bit(In_sync, &rdev->flags) ||
1866        atomic_read(&rdev->nr_pending)) {
1867        err = -EBUSY;
1868        goto abort;
1869    }
1870    /* Only remove faulty devices if recovery
1871     * is not possible.
1872     */
1873    if (!test_bit(Faulty, &rdev->flags) &&
1874        mddev->recovery_disabled != p->recovery_disabled &&
1875        (!p->replacement || p->replacement == rdev) &&
1876        number < conf->geo.raid_disks &&
1877        enough(conf, -1)) {
1878        err = -EBUSY;
1879        goto abort;
1880    }
1881    *rdevp = NULL;
1882    synchronize_rcu();
1883    if (atomic_read(&rdev->nr_pending)) {
1884        /* lost the race, try later */
1885        err = -EBUSY;
1886        *rdevp = rdev;
1887        goto abort;
1888    } else if (p->replacement) {
1889        /* We must have just cleared 'rdev' */
1890        p->rdev = p->replacement;
1891        clear_bit(Replacement, &p->replacement->flags);
1892        smp_mb(); /* Make sure other CPUs may see both as identical
1893               * but will never see neither -- if they are careful.
1894               */
1895        p->replacement = NULL;
1896        clear_bit(WantReplacement, &rdev->flags);
1897    } else
1898        /* We might have just remove the Replacement as faulty
1899         * Clear the flag just in case
1900         */
1901        clear_bit(WantReplacement, &rdev->flags);
1902
1903    err = md_integrity_register(mddev);
1904
1905abort:
1906
1907    print_conf(conf);
1908    return err;
1909}
1910
1911
1912static void end_sync_read(struct bio *bio, int error)
1913{
1914    struct r10bio *r10_bio = bio->bi_private;
1915    struct r10conf *conf = r10_bio->mddev->private;
1916    int d;
1917
1918    if (bio == r10_bio->master_bio) {
1919        /* this is a reshape read */
1920        d = r10_bio->read_slot; /* really the read dev */
1921    } else
1922        d = find_bio_disk(conf, r10_bio, bio, NULL, NULL);
1923
1924    if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1925        set_bit(R10BIO_Uptodate, &r10_bio->state);
1926    else
1927        /* The write handler will notice the lack of
1928         * R10BIO_Uptodate and record any errors etc
1929         */
1930        atomic_add(r10_bio->sectors,
1931               &conf->mirrors[d].rdev->corrected_errors);
1932
1933    /* for reconstruct, we always reschedule after a read.
1934     * for resync, only after all reads
1935     */
1936    rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1937    if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1938        atomic_dec_and_test(&r10_bio->remaining)) {
1939        /* we have read all the blocks,
1940         * do the comparison in process context in raid10d
1941         */
1942        reschedule_retry(r10_bio);
1943    }
1944}
1945
1946static void end_sync_request(struct r10bio *r10_bio)
1947{
1948    struct mddev *mddev = r10_bio->mddev;
1949
1950    while (atomic_dec_and_test(&r10_bio->remaining)) {
1951        if (r10_bio->master_bio == NULL) {
1952            /* the primary of several recovery bios */
1953            sector_t s = r10_bio->sectors;
1954            if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1955                test_bit(R10BIO_WriteError, &r10_bio->state))
1956                reschedule_retry(r10_bio);
1957            else
1958                put_buf(r10_bio);
1959            md_done_sync(mddev, s, 1);
1960            break;
1961        } else {
1962            struct r10bio *r10_bio2 = (struct r10bio *)r10_bio->master_bio;
1963            if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
1964                test_bit(R10BIO_WriteError, &r10_bio->state))
1965                reschedule_retry(r10_bio);
1966            else
1967                put_buf(r10_bio);
1968            r10_bio = r10_bio2;
1969        }
1970    }
1971}
1972
1973static void end_sync_write(struct bio *bio, int error)
1974{
1975    int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1976    struct r10bio *r10_bio = bio->bi_private;
1977    struct mddev *mddev = r10_bio->mddev;
1978    struct r10conf *conf = mddev->private;
1979    int d;
1980    sector_t first_bad;
1981    int bad_sectors;
1982    int slot;
1983    int repl;
1984    struct md_rdev *rdev = NULL;
1985
1986    d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
1987    if (repl)
1988        rdev = conf->mirrors[d].replacement;
1989    else
1990        rdev = conf->mirrors[d].rdev;
1991
1992    if (!uptodate) {
1993        if (repl)
1994            md_error(mddev, rdev);
1995        else {
1996            set_bit(WriteErrorSeen, &rdev->flags);
1997            if (!test_and_set_bit(WantReplacement, &rdev->flags))
1998                set_bit(MD_RECOVERY_NEEDED,
1999                    &rdev->mddev->recovery);
2000            set_bit(R10BIO_WriteError, &r10_bio->state);
2001        }
2002    } else if (is_badblock(rdev,
2003                 r10_bio->devs[slot].addr,
2004                 r10_bio->sectors,
2005                 &first_bad, &bad_sectors))
2006        set_bit(R10BIO_MadeGood, &r10_bio->state);
2007
2008    rdev_dec_pending(rdev, mddev);
2009
2010    end_sync_request(r10_bio);
2011}
2012
2013/*
2014 * Note: sync and recover and handled very differently for raid10
2015 * This code is for resync.
2016 * For resync, we read through virtual addresses and read all blocks.
2017 * If there is any error, we schedule a write. The lowest numbered
2018 * drive is authoritative.
2019 * However requests come for physical address, so we need to map.
2020 * For every physical address there are raid_disks/copies virtual addresses,
2021 * which is always are least one, but is not necessarly an integer.
2022 * This means that a physical address can span multiple chunks, so we may
2023 * have to submit multiple io requests for a single sync request.
2024 */
2025/*
2026 * We check if all blocks are in-sync and only write to blocks that
2027 * aren't in sync
2028 */
2029static void sync_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2030{
2031    struct r10conf *conf = mddev->private;
2032    int i, first;
2033    struct bio *tbio, *fbio;
2034    int vcnt;
2035
2036    atomic_set(&r10_bio->remaining, 1);
2037
2038    /* find the first device with a block */
2039    for (i=0; i<conf->copies; i++)
2040        if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
2041            break;
2042
2043    if (i == conf->copies)
2044        goto done;
2045
2046    first = i;
2047    fbio = r10_bio->devs[i].bio;
2048
2049    vcnt = (r10_bio->sectors + (PAGE_SIZE >> 9) - 1) >> (PAGE_SHIFT - 9);
2050    /* now find blocks with errors */
2051    for (i=0 ; i < conf->copies ; i++) {
2052        int j, d;
2053
2054        tbio = r10_bio->devs[i].bio;
2055
2056        if (tbio->bi_end_io != end_sync_read)
2057            continue;
2058        if (i == first)
2059            continue;
2060        if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
2061            /* We know that the bi_io_vec layout is the same for
2062             * both 'first' and 'i', so we just compare them.
2063             * All vec entries are PAGE_SIZE;
2064             */
2065            for (j = 0; j < vcnt; j++)
2066                if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
2067                       page_address(tbio->bi_io_vec[j].bv_page),
2068                       fbio->bi_io_vec[j].bv_len))
2069                    break;
2070            if (j == vcnt)
2071                continue;
2072            atomic64_add(r10_bio->sectors, &mddev->resync_mismatches);
2073            if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
2074                /* Don't fix anything. */
2075                continue;
2076        }
2077        /* Ok, we need to write this bio, either to correct an
2078         * inconsistency or to correct an unreadable block.
2079         * First we need to fixup bv_offset, bv_len and
2080         * bi_vecs, as the read request might have corrupted these
2081         */
2082        tbio->bi_vcnt = vcnt;
2083        tbio->bi_size = r10_bio->sectors << 9;
2084        tbio->bi_idx = 0;
2085        tbio->bi_phys_segments = 0;
2086        tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
2087        tbio->bi_flags |= 1 << BIO_UPTODATE;
2088        tbio->bi_next = NULL;
2089        tbio->bi_rw = WRITE;
2090        tbio->bi_private = r10_bio;
2091        tbio->bi_sector = r10_bio->devs[i].addr;
2092
2093        for (j=0; j < vcnt ; j++) {
2094            tbio->bi_io_vec[j].bv_offset = 0;
2095            tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
2096
2097            memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2098                   page_address(fbio->bi_io_vec[j].bv_page),
2099                   PAGE_SIZE);
2100        }
2101        tbio->bi_end_io = end_sync_write;
2102
2103        d = r10_bio->devs[i].devnum;
2104        atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2105        atomic_inc(&r10_bio->remaining);
2106        md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
2107
2108        tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
2109        tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
2110        generic_make_request(tbio);
2111    }
2112
2113    /* Now write out to any replacement devices
2114     * that are active
2115     */
2116    for (i = 0; i < conf->copies; i++) {
2117        int j, d;
2118
2119        tbio = r10_bio->devs[i].repl_bio;
2120        if (!tbio || !tbio->bi_end_io)
2121            continue;
2122        if (r10_bio->devs[i].bio->bi_end_io != end_sync_write
2123            && r10_bio->devs[i].bio != fbio)
2124            for (j = 0; j < vcnt; j++)
2125                memcpy(page_address(tbio->bi_io_vec[j].bv_page),
2126                       page_address(fbio->bi_io_vec[j].bv_page),
2127                       PAGE_SIZE);
2128        d = r10_bio->devs[i].devnum;
2129        atomic_inc(&r10_bio->remaining);
2130        md_sync_acct(conf->mirrors[d].replacement->bdev,
2131                 tbio->bi_size >> 9);
2132        generic_make_request(tbio);
2133    }
2134
2135done:
2136    if (atomic_dec_and_test(&r10_bio->remaining)) {
2137        md_done_sync(mddev, r10_bio->sectors, 1);
2138        put_buf(r10_bio);
2139    }
2140}
2141
2142/*
2143 * Now for the recovery code.
2144 * Recovery happens across physical sectors.
2145 * We recover all non-is_sync drives by finding the virtual address of
2146 * each, and then choose a working drive that also has that virt address.
2147 * There is a separate r10_bio for each non-in_sync drive.
2148 * Only the first two slots are in use. The first for reading,
2149 * The second for writing.
2150 *
2151 */
2152static void fix_recovery_read_error(struct r10bio *r10_bio)
2153{
2154    /* We got a read error during recovery.
2155     * We repeat the read in smaller page-sized sections.
2156     * If a read succeeds, write it to the new device or record
2157     * a bad block if we cannot.
2158     * If a read fails, record a bad block on both old and
2159     * new devices.
2160     */
2161    struct mddev *mddev = r10_bio->mddev;
2162    struct r10conf *conf = mddev->private;
2163    struct bio *bio = r10_bio->devs[0].bio;
2164    sector_t sect = 0;
2165    int sectors = r10_bio->sectors;
2166    int idx = 0;
2167    int dr = r10_bio->devs[0].devnum;
2168    int dw = r10_bio->devs[1].devnum;
2169
2170    while (sectors) {
2171        int s = sectors;
2172        struct md_rdev *rdev;
2173        sector_t addr;
2174        int ok;
2175
2176        if (s > (PAGE_SIZE>>9))
2177            s = PAGE_SIZE >> 9;
2178
2179        rdev = conf->mirrors[dr].rdev;
2180        addr = r10_bio->devs[0].addr + sect,
2181        ok = sync_page_io(rdev,
2182                  addr,
2183                  s << 9,
2184                  bio->bi_io_vec[idx].bv_page,
2185                  READ, false);
2186        if (ok) {
2187            rdev = conf->mirrors[dw].rdev;
2188            addr = r10_bio->devs[1].addr + sect;
2189            ok = sync_page_io(rdev,
2190                      addr,
2191                      s << 9,
2192                      bio->bi_io_vec[idx].bv_page,
2193                      WRITE, false);
2194            if (!ok) {
2195                set_bit(WriteErrorSeen, &rdev->flags);
2196                if (!test_and_set_bit(WantReplacement,
2197                              &rdev->flags))
2198                    set_bit(MD_RECOVERY_NEEDED,
2199                        &rdev->mddev->recovery);
2200            }
2201        }
2202        if (!ok) {
2203            /* We don't worry if we cannot set a bad block -
2204             * it really is bad so there is no loss in not
2205             * recording it yet
2206             */
2207            rdev_set_badblocks(rdev, addr, s, 0);
2208
2209            if (rdev != conf->mirrors[dw].rdev) {
2210                /* need bad block on destination too */
2211                struct md_rdev *rdev2 = conf->mirrors[dw].rdev;
2212                addr = r10_bio->devs[1].addr + sect;
2213                ok = rdev_set_badblocks(rdev2, addr, s, 0);
2214                if (!ok) {
2215                    /* just abort the recovery */
2216                    printk(KERN_NOTICE
2217                           "md/raid10:%s: recovery aborted"
2218                           " due to read error\n",
2219                           mdname(mddev));
2220
2221                    conf->mirrors[dw].recovery_disabled
2222                        = mddev->recovery_disabled;
2223                    set_bit(MD_RECOVERY_INTR,
2224                        &mddev->recovery);
2225                    break;
2226                }
2227            }
2228        }
2229
2230        sectors -= s;
2231        sect += s;
2232        idx++;
2233    }
2234}
2235
2236static void recovery_request_write(struct mddev *mddev, struct r10bio *r10_bio)
2237{
2238    struct r10conf *conf = mddev->private;
2239    int d;
2240    struct bio *wbio, *wbio2;
2241
2242    if (!test_bit(R10BIO_Uptodate, &r10_bio->state)) {
2243        fix_recovery_read_error(r10_bio);
2244        end_sync_request(r10_bio);
2245        return;
2246    }
2247
2248    /*
2249     * share the pages with the first bio
2250     * and submit the write request
2251     */
2252    d = r10_bio->devs[1].devnum;
2253    wbio = r10_bio->devs[1].bio;
2254    wbio2 = r10_bio->devs[1].repl_bio;
2255    if (wbio->bi_end_io) {
2256        atomic_inc(&conf->mirrors[d].rdev->nr_pending);
2257        md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
2258        generic_make_request(wbio);
2259    }
2260    if (wbio2 && wbio2->bi_end_io) {
2261        atomic_inc(&conf->mirrors[d].replacement->nr_pending);
2262        md_sync_acct(conf->mirrors[d].replacement->bdev,
2263                 wbio2->bi_size >> 9);
2264        generic_make_request(wbio2);
2265    }
2266}
2267
2268
2269/*
2270 * Used by fix_read_error() to decay the per rdev read_errors.
2271 * We halve the read error count for every hour that has elapsed
2272 * since the last recorded read error.
2273 *
2274 */
2275static void check_decay_read_errors(struct mddev *mddev, struct md_rdev *rdev)
2276{
2277    struct timespec cur_time_mon;
2278    unsigned long hours_since_last;
2279    unsigned int read_errors = atomic_read(&rdev->read_errors);
2280
2281    ktime_get_ts(&cur_time_mon);
2282
2283    if (rdev->last_read_error.tv_sec == 0 &&
2284        rdev->last_read_error.tv_nsec == 0) {
2285        /* first time we've seen a read error */
2286        rdev->last_read_error = cur_time_mon;
2287        return;
2288    }
2289
2290    hours_since_last = (cur_time_mon.tv_sec -
2291                rdev->last_read_error.tv_sec) / 3600;
2292
2293    rdev->last_read_error = cur_time_mon;
2294
2295    /*
2296     * if hours_since_last is > the number of bits in read_errors
2297     * just set read errors to 0. We do this to avoid
2298     * overflowing the shift of read_errors by hours_since_last.
2299     */
2300    if (hours_since_last >= 8 * sizeof(read_errors))
2301        atomic_set(&rdev->read_errors, 0);
2302    else
2303        atomic_set(&rdev->read_errors, read_errors >> hours_since_last);
2304}
2305
2306static int r10_sync_page_io(struct md_rdev *rdev, sector_t sector,
2307                int sectors, struct page *page, int rw)
2308{
2309    sector_t first_bad;
2310    int bad_sectors;
2311
2312    if (is_badblock(rdev, sector, sectors, &first_bad, &bad_sectors)
2313        && (rw == READ || test_bit(WriteErrorSeen, &rdev->flags)))
2314        return -1;
2315    if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
2316        /* success */
2317        return 1;
2318    if (rw == WRITE) {
2319        set_bit(WriteErrorSeen, &rdev->flags);
2320        if (!test_and_set_bit(WantReplacement, &rdev->flags))
2321            set_bit(MD_RECOVERY_NEEDED,
2322                &rdev->mddev->recovery);
2323    }
2324    /* need to record an error - either for the block or the device */
2325    if (!rdev_set_badblocks(rdev, sector, sectors, 0))
2326        md_error(rdev->mddev, rdev);
2327    return 0;
2328}
2329
2330/*
2331 * This is a kernel thread which:
2332 *
2333 * 1. Retries failed read operations on working mirrors.
2334 * 2. Updates the raid superblock when problems encounter.
2335 * 3. Performs writes following reads for array synchronising.
2336 */
2337
2338static void fix_read_error(struct r10conf *conf, struct mddev *mddev, struct r10bio *r10_bio)
2339{
2340    int sect = 0; /* Offset from r10_bio->sector */
2341    int sectors = r10_bio->sectors;
2342    struct md_rdev*rdev;
2343    int max_read_errors = atomic_read(&mddev->max_corr_read_errors);
2344    int d = r10_bio->devs[r10_bio->read_slot].devnum;
2345
2346    /* still own a reference to this rdev, so it cannot
2347     * have been cleared recently.
2348     */
2349    rdev = conf->mirrors[d].rdev;
2350
2351    if (test_bit(Faulty, &rdev->flags))
2352        /* drive has already been failed, just ignore any
2353           more fix_read_error() attempts */
2354        return;
2355
2356    check_decay_read_errors(mddev, rdev);
2357    atomic_inc(&rdev->read_errors);
2358    if (atomic_read(&rdev->read_errors) > max_read_errors) {
2359        char b[BDEVNAME_SIZE];
2360        bdevname(rdev->bdev, b);
2361
2362        printk(KERN_NOTICE
2363               "md/raid10:%s: %s: Raid device exceeded "
2364               "read_error threshold [cur %d:max %d]\n",
2365               mdname(mddev), b,
2366               atomic_read(&rdev->read_errors), max_read_errors);
2367        printk(KERN_NOTICE
2368               "md/raid10:%s: %s: Failing raid device\n",
2369               mdname(mddev), b);
2370        md_error(mddev, conf->mirrors[d].rdev);
2371        r10_bio->devs[r10_bio->read_slot].bio = IO_BLOCKED;
2372        return;
2373    }
2374
2375    while(sectors) {
2376        int s = sectors;
2377        int sl = r10_bio->read_slot;
2378        int success = 0;
2379        int start;
2380
2381        if (s > (PAGE_SIZE>>9))
2382            s = PAGE_SIZE >> 9;
2383
2384        rcu_read_lock();
2385        do {
2386            sector_t first_bad;
2387            int bad_sectors;
2388
2389            d = r10_bio->devs[sl].devnum;
2390            rdev = rcu_dereference(conf->mirrors[d].rdev);
2391            if (rdev &&
2392                !test_bit(Unmerged, &rdev->flags) &&
2393                test_bit(In_sync, &rdev->flags) &&
2394                is_badblock(rdev, r10_bio->devs[sl].addr + sect, s,
2395                    &first_bad, &bad_sectors) == 0) {
2396                atomic_inc(&rdev->nr_pending);
2397                rcu_read_unlock();
2398                success = sync_page_io(rdev,
2399                               r10_bio->devs[sl].addr +
2400                               sect,
2401                               s<<9,
2402                               conf->tmppage, READ, false);
2403                rdev_dec_pending(rdev, mddev);
2404                rcu_read_lock();
2405                if (success)
2406                    break;
2407            }
2408            sl++;
2409            if (sl == conf->copies)
2410                sl = 0;
2411        } while (!success && sl != r10_bio->read_slot);
2412        rcu_read_unlock();
2413
2414        if (!success) {
2415            /* Cannot read from anywhere, just mark the block
2416             * as bad on the first device to discourage future
2417             * reads.
2418             */
2419            int dn = r10_bio->devs[r10_bio->read_slot].devnum;
2420            rdev = conf->mirrors[dn].rdev;
2421
2422            if (!rdev_set_badblocks(
2423                    rdev,
2424                    r10_bio->devs[r10_bio->read_slot].addr
2425                    + sect,
2426                    s, 0)) {
2427                md_error(mddev, rdev);
2428                r10_bio->devs[r10_bio->read_slot].bio
2429                    = IO_BLOCKED;
2430            }
2431            break;
2432        }
2433
2434        start = sl;
2435        /* write it back and re-read */
2436        rcu_read_lock();
2437        while (sl != r10_bio->read_slot) {
2438            char b[BDEVNAME_SIZE];
2439
2440            if (sl==0)
2441                sl = conf->copies;
2442            sl--;
2443            d = r10_bio->devs[sl].devnum;
2444            rdev = rcu_dereference(conf->mirrors[d].rdev);
2445            if (!rdev ||
2446                test_bit(Unmerged, &rdev->flags) ||
2447                !test_bit(In_sync, &rdev->flags))
2448                continue;
2449
2450            atomic_inc(&rdev->nr_pending);
2451            rcu_read_unlock();
2452            if (r10_sync_page_io(rdev,
2453                         r10_bio->devs[sl].addr +
2454                         sect,
2455                         s, conf->tmppage, WRITE)
2456                == 0) {
2457                /* Well, this device is dead */
2458                printk(KERN_NOTICE
2459                       "md/raid10:%s: read correction "
2460                       "write failed"
2461                       " (%d sectors at %llu on %s)\n",
2462                       mdname(mddev), s,
2463                       (unsigned long long)(
2464                           sect +
2465                           choose_data_offset(r10_bio,
2466                                  rdev)),
2467                       bdevname(rdev->bdev, b));
2468                printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2469                       "drive\n",
2470                       mdname(mddev),
2471                       bdevname(rdev->bdev, b));
2472            }
2473            rdev_dec_pending(rdev, mddev);
2474            rcu_read_lock();
2475        }
2476        sl = start;
2477        while (sl != r10_bio->read_slot) {
2478            char b[BDEVNAME_SIZE];
2479
2480            if (sl==0)
2481                sl = conf->copies;
2482            sl--;
2483            d = r10_bio->devs[sl].devnum;
2484            rdev = rcu_dereference(conf->mirrors[d].rdev);
2485            if (!rdev ||
2486                !test_bit(In_sync, &rdev->flags))
2487                continue;
2488
2489            atomic_inc(&rdev->nr_pending);
2490            rcu_read_unlock();
2491            switch (r10_sync_page_io(rdev,
2492                         r10_bio->devs[sl].addr +
2493                         sect,
2494                         s, conf->tmppage,
2495                         READ)) {
2496            case 0:
2497                /* Well, this device is dead */
2498                printk(KERN_NOTICE
2499                       "md/raid10:%s: unable to read back "
2500                       "corrected sectors"
2501                       " (%d sectors at %llu on %s)\n",
2502                       mdname(mddev), s,
2503                       (unsigned long long)(
2504                           sect +
2505                           choose_data_offset(r10_bio, rdev)),
2506                       bdevname(rdev->bdev, b));
2507                printk(KERN_NOTICE "md/raid10:%s: %s: failing "
2508                       "drive\n",
2509                       mdname(mddev),
2510                       bdevname(rdev->bdev, b));
2511                break;
2512            case 1:
2513                printk(KERN_INFO
2514                       "md/raid10:%s: read error corrected"
2515                       " (%d sectors at %llu on %s)\n",
2516                       mdname(mddev), s,
2517                       (unsigned long long)(
2518                           sect +
2519                           choose_data_offset(r10_bio, rdev)),
2520                       bdevname(rdev->bdev, b));
2521                atomic_add(s, &rdev->corrected_errors);
2522            }
2523
2524            rdev_dec_pending(rdev, mddev);
2525            rcu_read_lock();
2526        }
2527        rcu_read_unlock();
2528
2529        sectors -= s;
2530        sect += s;
2531    }
2532}
2533
2534static void bi_complete(struct bio *bio, int error)
2535{
2536    complete((struct completion *)bio->bi_private);
2537}
2538
2539static int submit_bio_wait(int rw, struct bio *bio)
2540{
2541    struct completion event;
2542    rw |= REQ_SYNC;
2543
2544    init_completion(&event);
2545    bio->bi_private = &event;
2546    bio->bi_end_io = bi_complete;
2547    submit_bio(rw, bio);
2548    wait_for_completion(&event);
2549
2550    return test_bit(BIO_UPTODATE, &bio->bi_flags);
2551}
2552
2553static int narrow_write_error(struct r10bio *r10_bio, int i)
2554{
2555    struct bio *bio = r10_bio->master_bio;
2556    struct mddev *mddev = r10_bio->mddev;
2557    struct r10conf *conf = mddev->private;
2558    struct md_rdev *rdev = conf->mirrors[r10_bio->devs[i].devnum].rdev;
2559    /* bio has the data to be written to slot 'i' where
2560     * we just recently had a write error.
2561     * We repeatedly clone the bio and trim down to one block,
2562     * then try the write. Where the write fails we record
2563     * a bad block.
2564     * It is conceivable that the bio doesn't exactly align with
2565     * blocks. We must handle this.
2566     *
2567     * We currently own a reference to the rdev.
2568     */
2569
2570    int block_sectors;
2571    sector_t sector;
2572    int sectors;
2573    int sect_to_write = r10_bio->sectors;
2574    int ok = 1;
2575
2576    if (rdev->badblocks.shift < 0)
2577        return 0;
2578
2579    block_sectors = 1 << rdev->badblocks.shift;
2580    sector = r10_bio->sector;
2581    sectors = ((r10_bio->sector + block_sectors)
2582           & ~(sector_t)(block_sectors - 1))
2583        - sector;
2584
2585    while (sect_to_write) {
2586        struct bio *wbio;
2587        if (sectors > sect_to_write)
2588            sectors = sect_to_write;
2589        /* Write at 'sector' for 'sectors' */
2590        wbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
2591        md_trim_bio(wbio, sector - bio->bi_sector, sectors);
2592        wbio->bi_sector = (r10_bio->devs[i].addr+
2593                   choose_data_offset(r10_bio, rdev) +
2594                   (sector - r10_bio->sector));
2595        wbio->bi_bdev = rdev->bdev;
2596        if (submit_bio_wait(WRITE, wbio) == 0)
2597            /* Failure! */
2598            ok = rdev_set_badblocks(rdev, sector,
2599                        sectors, 0)
2600                && ok;
2601
2602        bio_put(wbio);
2603        sect_to_write -= sectors;
2604        sector += sectors;
2605        sectors = block_sectors;
2606    }
2607    return ok;
2608}
2609
2610static void handle_read_error(struct mddev *mddev, struct r10bio *r10_bio)
2611{
2612    int slot = r10_bio->read_slot;
2613    struct bio *bio;
2614    struct r10conf *conf = mddev->private;
2615    struct md_rdev *rdev = r10_bio->devs[slot].rdev;
2616    char b[BDEVNAME_SIZE];
2617    unsigned long do_sync;
2618    int max_sectors;
2619
2620    /* we got a read error. Maybe the drive is bad. Maybe just
2621     * the block and we can fix it.
2622     * We freeze all other IO, and try reading the block from
2623     * other devices. When we find one, we re-write
2624     * and check it that fixes the read error.
2625     * This is all done synchronously while the array is
2626     * frozen.
2627     */
2628    bio = r10_bio->devs[slot].bio;
2629    bdevname(bio->bi_bdev, b);
2630    bio_put(bio);
2631    r10_bio->devs[slot].bio = NULL;
2632
2633    if (mddev->ro == 0) {
2634        freeze_array(conf);
2635        fix_read_error(conf, mddev, r10_bio);
2636        unfreeze_array(conf);
2637    } else
2638        r10_bio->devs[slot].bio = IO_BLOCKED;
2639
2640    rdev_dec_pending(rdev, mddev);
2641
2642read_more:
2643    rdev = read_balance(conf, r10_bio, &max_sectors);
2644    if (rdev == NULL) {
2645        printk(KERN_ALERT "md/raid10:%s: %s: unrecoverable I/O"
2646               " read error for block %llu\n",
2647               mdname(mddev), b,
2648               (unsigned long long)r10_bio->sector);
2649        raid_end_bio_io(r10_bio);
2650        return;
2651    }
2652
2653    do_sync = (r10_bio->master_bio->bi_rw & REQ_SYNC);
2654    slot = r10_bio->read_slot;
2655    printk_ratelimited(
2656        KERN_ERR
2657        "md/raid10:%s: %s: redirecting "
2658        "sector %llu to another mirror\n",
2659        mdname(mddev),
2660        bdevname(rdev->bdev, b),
2661        (unsigned long long)r10_bio->sector);
2662    bio = bio_clone_mddev(r10_bio->master_bio,
2663                  GFP_NOIO, mddev);
2664    md_trim_bio(bio,
2665            r10_bio->sector - bio->bi_sector,
2666            max_sectors);
2667    r10_bio->devs[slot].bio = bio;
2668    r10_bio->devs[slot].rdev = rdev;
2669    bio->bi_sector = r10_bio->devs[slot].addr
2670        + choose_data_offset(r10_bio, rdev);
2671    bio->bi_bdev = rdev->bdev;
2672    bio->bi_rw = READ | do_sync;
2673    bio->bi_private = r10_bio;
2674    bio->bi_end_io = raid10_end_read_request;
2675    if (max_sectors < r10_bio->sectors) {
2676        /* Drat - have to split this up more */
2677        struct bio *mbio = r10_bio->master_bio;
2678        int sectors_handled =
2679            r10_bio->sector + max_sectors
2680            - mbio->bi_sector;
2681        r10_bio->sectors = max_sectors;
2682        spin_lock_irq(&conf->device_lock);
2683        if (mbio->bi_phys_segments == 0)
2684            mbio->bi_phys_segments = 2;
2685        else
2686            mbio->bi_phys_segments++;
2687        spin_unlock_irq(&conf->device_lock);
2688        generic_make_request(bio);
2689
2690        r10_bio = mempool_alloc(conf->r10bio_pool,
2691                    GFP_NOIO);
2692        r10_bio->master_bio = mbio;
2693        r10_bio->sectors = (mbio->bi_size >> 9)
2694            - sectors_handled;
2695        r10_bio->state = 0;
2696        set_bit(R10BIO_ReadError,
2697            &r10_bio->state);
2698        r10_bio->mddev = mddev;
2699        r10_bio->sector = mbio->bi_sector
2700            + sectors_handled;
2701
2702        goto read_more;
2703    } else
2704        generic_make_request(bio);
2705}
2706
2707static void handle_write_completed(struct r10conf *conf, struct r10bio *r10_bio)
2708{
2709    /* Some sort of write request has finished and it
2710     * succeeded in writing where we thought there was a
2711     * bad block. So forget the bad block.
2712     * Or possibly if failed and we need to record
2713     * a bad block.
2714     */
2715    int m;
2716    struct md_rdev *rdev;
2717
2718    if (test_bit(R10BIO_IsSync, &r10_bio->state) ||
2719        test_bit(R10BIO_IsRecover, &r10_bio->state)) {
2720        for (m = 0; m < conf->copies; m++) {
2721            int dev = r10_bio->devs[m].devnum;
2722            rdev = conf->mirrors[dev].rdev;
2723            if (r10_bio->devs[m].bio == NULL)
2724                continue;
2725            if (test_bit(BIO_UPTODATE,
2726                     &r10_bio->devs[m].bio->bi_flags)) {
2727                rdev_clear_badblocks(
2728                    rdev,
2729                    r10_bio->devs[m].addr,
2730                    r10_bio->sectors, 0);
2731            } else {
2732                if (!rdev_set_badblocks(
2733                        rdev,
2734                        r10_bio->devs[m].addr,
2735                        r10_bio->sectors, 0))
2736                    md_error(conf->mddev, rdev);
2737            }
2738            rdev = conf->mirrors[dev].replacement;
2739            if (r10_bio->devs[m].repl_bio == NULL)
2740                continue;
2741            if (test_bit(BIO_UPTODATE,
2742                     &r10_bio->devs[m].repl_bio->bi_flags)) {
2743                rdev_clear_badblocks(
2744                    rdev,
2745                    r10_bio->devs[m].addr,
2746                    r10_bio->sectors, 0);
2747            } else {
2748                if (!rdev_set_badblocks(
2749                        rdev,
2750                        r10_bio->devs[m].addr,
2751                        r10_bio->sectors, 0))
2752                    md_error(conf->mddev, rdev);
2753            }
2754        }
2755        put_buf(r10_bio);
2756    } else {
2757        for (m = 0; m < conf->copies; m++) {
2758            int dev = r10_bio->devs[m].devnum;
2759            struct bio *bio = r10_bio->devs[m].bio;
2760            rdev = conf->mirrors[dev].rdev;
2761            if (bio == IO_MADE_GOOD) {
2762                rdev_clear_badblocks(
2763                    rdev,
2764                    r10_bio->devs[m].addr,
2765                    r10_bio->sectors, 0);
2766                rdev_dec_pending(rdev, conf->mddev);
2767            } else if (bio != NULL &&
2768                   !test_bit(BIO_UPTODATE, &bio->bi_flags)) {
2769                if (!narrow_write_error(r10_bio, m)) {
2770                    md_error(conf->mddev, rdev);
2771                    set_bit(R10BIO_Degraded,
2772                        &r10_bio->state);
2773                }
2774                rdev_dec_pending(rdev, conf->mddev);
2775            }
2776            bio = r10_bio->devs[m].repl_bio;
2777            rdev = conf->mirrors[dev].replacement;
2778            if (rdev && bio == IO_MADE_GOOD) {
2779                rdev_clear_badblocks(
2780                    rdev,
2781                    r10_bio->devs[m].addr,
2782                    r10_bio->sectors, 0);
2783                rdev_dec_pending(rdev, conf->mddev);
2784            }
2785        }
2786        if (test_bit(R10BIO_WriteError,
2787                 &r10_bio->state))
2788            close_write(r10_bio);
2789        raid_end_bio_io(r10_bio);
2790    }
2791}
2792
2793static void raid10d(struct md_thread *thread)
2794{
2795    struct mddev *mddev = thread->mddev;
2796    struct r10bio *r10_bio;
2797    unsigned long flags;
2798    struct r10conf *conf = mddev->private;
2799    struct list_head *head = &conf->retry_list;
2800    struct blk_plug plug;
2801
2802    md_check_recovery(mddev);
2803
2804    blk_start_plug(&plug);
2805    for (;;) {
2806
2807        flush_pending_writes(conf);
2808
2809        spin_lock_irqsave(&conf->device_lock, flags);
2810        if (list_empty(head)) {
2811            spin_unlock_irqrestore(&conf->device_lock, flags);
2812            break;
2813        }
2814        r10_bio = list_entry(head->prev, struct r10bio, retry_list);
2815        list_del(head->prev);
2816        conf->nr_queued--;
2817        spin_unlock_irqrestore(&conf->device_lock, flags);
2818
2819        mddev = r10_bio->mddev;
2820        conf = mddev->private;
2821        if (test_bit(R10BIO_MadeGood, &r10_bio->state) ||
2822            test_bit(R10BIO_WriteError, &r10_bio->state))
2823            handle_write_completed(conf, r10_bio);
2824        else if (test_bit(R10BIO_IsReshape, &r10_bio->state))
2825            reshape_request_write(mddev, r10_bio);
2826        else if (test_bit(R10BIO_IsSync, &r10_bio->state))
2827            sync_request_write(mddev, r10_bio);
2828        else if (test_bit(R10BIO_IsRecover, &r10_bio->state))
2829            recovery_request_write(mddev, r10_bio);
2830        else if (test_bit(R10BIO_ReadError, &r10_bio->state))
2831            handle_read_error(mddev, r10_bio);
2832        else {
2833            /* just a partial read to be scheduled from a
2834             * separate context
2835             */
2836            int slot = r10_bio->read_slot;
2837            generic_make_request(r10_bio->devs[slot].bio);
2838        }
2839
2840        cond_resched();
2841        if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2842            md_check_recovery(mddev);
2843    }
2844    blk_finish_plug(&plug);
2845}
2846
2847
2848static int init_resync(struct r10conf *conf)
2849{
2850    int buffs;
2851    int i;
2852
2853    buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2854    BUG_ON(conf->r10buf_pool);
2855    conf->have_replacement = 0;
2856    for (i = 0; i < conf->geo.raid_disks; i++)
2857        if (conf->mirrors[i].replacement)
2858            conf->have_replacement = 1;
2859    conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
2860    if (!conf->r10buf_pool)
2861        return -ENOMEM;
2862    conf->next_resync = 0;
2863    return 0;
2864}
2865
2866/*
2867 * perform a "sync" on one "block"
2868 *
2869 * We need to make sure that no normal I/O request - particularly write
2870 * requests - conflict with active sync requests.
2871 *
2872 * This is achieved by tracking pending requests and a 'barrier' concept
2873 * that can be installed to exclude normal IO requests.
2874 *
2875 * Resync and recovery are handled very differently.
2876 * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
2877 *
2878 * For resync, we iterate over virtual addresses, read all copies,
2879 * and update if there are differences. If only one copy is live,
2880 * skip it.
2881 * For recovery, we iterate over physical addresses, read a good
2882 * value for each non-in_sync drive, and over-write.
2883 *
2884 * So, for recovery we may have several outstanding complex requests for a
2885 * given address, one for each out-of-sync device. We model this by allocating
2886 * a number of r10_bio structures, one for each out-of-sync device.
2887 * As we setup these structures, we collect all bio's together into a list
2888 * which we then process collectively to add pages, and then process again
2889 * to pass to generic_make_request.
2890 *
2891 * The r10_bio structures are linked using a borrowed master_bio pointer.
2892 * This link is counted in ->remaining. When the r10_bio that points to NULL
2893 * has its remaining count decremented to 0, the whole complex operation
2894 * is complete.
2895 *
2896 */
2897
2898static sector_t sync_request(struct mddev *mddev, sector_t sector_nr,
2899                 int *skipped, int go_faster)
2900{
2901    struct r10conf *conf = mddev->private;
2902    struct r10bio *r10_bio;
2903    struct bio *biolist = NULL, *bio;
2904    sector_t max_sector, nr_sectors;
2905    int i;
2906    int max_sync;
2907    sector_t sync_blocks;
2908    sector_t sectors_skipped = 0;
2909    int chunks_skipped = 0;
2910    sector_t chunk_mask = conf->geo.chunk_mask;
2911
2912    if (!conf->r10buf_pool)
2913        if (init_resync(conf))
2914            return 0;
2915
2916 skipped:
2917    max_sector = mddev->dev_sectors;
2918    if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) ||
2919        test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2920        max_sector = mddev->resync_max_sectors;
2921    if (sector_nr >= max_sector) {
2922        /* If we aborted, we need to abort the
2923         * sync on the 'current' bitmap chucks (there can
2924         * be several when recovering multiple devices).
2925         * as we may have started syncing it but not finished.
2926         * We can find the current address in
2927         * mddev->curr_resync, but for recovery,
2928         * we need to convert that to several
2929         * virtual addresses.
2930         */
2931        if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery)) {
2932            end_reshape(conf);
2933            return 0;
2934        }
2935
2936        if (mddev->curr_resync < max_sector) { /* aborted */
2937            if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
2938                bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2939                        &sync_blocks, 1);
2940            else for (i = 0; i < conf->geo.raid_disks; i++) {
2941                sector_t sect =
2942                    raid10_find_virt(conf, mddev->curr_resync, i);
2943                bitmap_end_sync(mddev->bitmap, sect,
2944                        &sync_blocks, 1);
2945            }
2946        } else {
2947            /* completed sync */
2948            if ((!mddev->bitmap || conf->fullsync)
2949                && conf->have_replacement
2950                && test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
2951                /* Completed a full sync so the replacements
2952                 * are now fully recovered.
2953                 */
2954                for (i = 0; i < conf->geo.raid_disks; i++)
2955                    if (conf->mirrors[i].replacement)
2956                        conf->mirrors[i].replacement
2957                            ->recovery_offset
2958                            = MaxSector;
2959            }
2960            conf->fullsync = 0;
2961        }
2962        bitmap_close_sync(mddev->bitmap);
2963        close_sync(conf);
2964        *skipped = 1;
2965        return sectors_skipped;
2966    }
2967
2968    if (test_bit(MD_RECOVERY_RESHAPE, &mddev->recovery))
2969        return reshape_request(mddev, sector_nr, skipped);
2970
2971    if (chunks_skipped >= conf->geo.raid_disks) {
2972        /* if there has been nothing to do on any drive,
2973         * then there is nothing to do at all..
2974         */
2975        *skipped = 1;
2976        return (max_sector - sector_nr) + sectors_skipped;
2977    }
2978
2979    if (max_sector > mddev->resync_max)
2980        max_sector = mddev->resync_max; /* Don't do IO beyond here */
2981
2982    /* make sure whole request will fit in a chunk - if chunks
2983     * are meaningful
2984     */
2985    if (conf->geo.near_copies < conf->geo.raid_disks &&
2986        max_sector > (sector_nr | chunk_mask))
2987        max_sector = (sector_nr | chunk_mask) + 1;
2988    /*
2989     * If there is non-resync activity waiting for us then
2990     * put in a delay to throttle resync.
2991     */
2992    if (!go_faster && conf->nr_waiting)
2993        msleep_interruptible(1000);
2994
2995    /* Again, very different code for resync and recovery.
2996     * Both must result in an r10bio with a list of bios that
2997     * have bi_end_io, bi_sector, bi_bdev set,
2998     * and bi_private set to the r10bio.
2999     * For recovery, we may actually create several r10bios
3000     * with 2 bios in each, that correspond to the bios in the main one.
3001     * In this case, the subordinate r10bios link back through a
3002     * borrowed master_bio pointer, and the counter in the master
3003     * includes a ref from each subordinate.
3004     */
3005    /* First, we decide what to do and set ->bi_end_io
3006     * To end_sync_read if we want to read, and
3007     * end_sync_write if we will want to write.
3008     */
3009
3010    max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
3011    if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
3012        /* recovery... the complicated one */
3013        int j;
3014        r10_bio = NULL;
3015
3016        for (i = 0 ; i < conf->geo.raid_disks; i++) {
3017            int still_degraded;
3018            struct r10bio *rb2;
3019            sector_t sect;
3020            int must_sync;
3021            int any_working;
3022            struct raid10_info *mirror = &conf->mirrors[i];
3023
3024            if ((mirror->rdev == NULL ||
3025                 test_bit(In_sync, &mirror->rdev->flags))
3026                &&
3027                (mirror->replacement == NULL ||
3028                 test_bit(Faulty,
3029                      &mirror->replacement->flags)))
3030                continue;
3031
3032            still_degraded = 0;
3033            /* want to reconstruct this device */
3034            rb2 = r10_bio;
3035            sect = raid10_find_virt(conf, sector_nr, i);
3036            if (sect >= mddev->resync_max_sectors) {
3037                /* last stripe is not complete - don't
3038                 * try to recover this sector.
3039                 */
3040                continue;
3041            }
3042            /* Unless we are doing a full sync, or a replacement
3043             * we only need to recover the block if it is set in
3044             * the bitmap
3045             */
3046            must_sync = bitmap_start_sync(mddev->bitmap, sect,
3047                              &sync_blocks, 1);
3048            if (sync_blocks < max_sync)
3049                max_sync = sync_blocks;
3050            if (!must_sync &&
3051                mirror->replacement == NULL &&
3052                !conf->fullsync) {
3053                /* yep, skip the sync_blocks here, but don't assume
3054                 * that there will never be anything to do here
3055                 */
3056                chunks_skipped = -1;
3057                continue;
3058            }
3059
3060            r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3061            raise_barrier(conf, rb2 != NULL);
3062            atomic_set(&r10_bio->remaining, 0);
3063
3064            r10_bio->master_bio = (struct bio*)rb2;
3065            if (rb2)
3066                atomic_inc(&rb2->remaining);
3067            r10_bio->mddev = mddev;
3068            set_bit(R10BIO_IsRecover, &r10_bio->state);
3069            r10_bio->sector = sect;
3070
3071            raid10_find_phys(conf, r10_bio);
3072
3073            /* Need to check if the array will still be
3074             * degraded
3075             */
3076            for (j = 0; j < conf->geo.raid_disks; j++)
3077                if (conf->mirrors[j].rdev == NULL ||
3078                    test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
3079                    still_degraded = 1;
3080                    break;
3081                }
3082
3083            must_sync = bitmap_start_sync(mddev->bitmap, sect,
3084                              &sync_blocks, still_degraded);
3085
3086            any_working = 0;
3087            for (j=0; j<conf->copies;j++) {
3088                int k;
3089                int d = r10_bio->devs[j].devnum;
3090                sector_t from_addr, to_addr;
3091                struct md_rdev *rdev;
3092                sector_t sector, first_bad;
3093                int bad_sectors;
3094                if (!conf->mirrors[d].rdev ||
3095                    !test_bit(In_sync, &conf->mirrors[d].rdev->flags))
3096                    continue;
3097                /* This is where we read from */
3098                any_working = 1;
3099                rdev = conf->mirrors[d].rdev;
3100                sector = r10_bio->devs[j].addr;
3101
3102                if (is_badblock(rdev, sector, max_sync,
3103                        &first_bad, &bad_sectors)) {
3104                    if (first_bad > sector)
3105                        max_sync = first_bad - sector;
3106                    else {
3107                        bad_sectors -= (sector
3108                                - first_bad);
3109                        if (max_sync > bad_sectors)
3110                            max_sync = bad_sectors;
3111                        continue;
3112                    }
3113                }
3114                bio = r10_bio->devs[0].bio;
3115                bio->bi_next = biolist;
3116                biolist = bio;
3117                bio->bi_private = r10_bio;
3118                bio->bi_end_io = end_sync_read;
3119                bio->bi_rw = READ;
3120                from_addr = r10_bio->devs[j].addr;
3121                bio->bi_sector = from_addr + rdev->data_offset;
3122                bio->bi_bdev = rdev->bdev;
3123                atomic_inc(&rdev->nr_pending);
3124                /* and we write to 'i' (if not in_sync) */
3125
3126                for (k=0; k<conf->copies; k++)
3127                    if (r10_bio->devs[k].devnum == i)
3128                        break;
3129                BUG_ON(k == conf->copies);
3130                to_addr = r10_bio->devs[k].addr;
3131                r10_bio->devs[0].devnum = d;
3132                r10_bio->devs[0].addr = from_addr;
3133                r10_bio->devs[1].devnum = i;
3134                r10_bio->devs[1].addr = to_addr;
3135
3136                rdev = mirror->rdev;
3137                if (!test_bit(In_sync, &rdev->flags)) {
3138                    bio = r10_bio->devs[1].bio;
3139                    bio->bi_next = biolist;
3140                    biolist = bio;
3141                    bio->bi_private = r10_bio;
3142                    bio->bi_end_io = end_sync_write;
3143                    bio->bi_rw = WRITE;
3144                    bio->bi_sector = to_addr
3145                        + rdev->data_offset;
3146                    bio->bi_bdev = rdev->bdev;
3147                    atomic_inc(&r10_bio->remaining);
3148                } else
3149                    r10_bio->devs[1].bio->bi_end_io = NULL;
3150
3151                /* and maybe write to replacement */
3152                bio = r10_bio->devs[1].repl_bio;
3153                if (bio)
3154                    bio->bi_end_io = NULL;
3155                rdev = mirror->replacement;
3156                /* Note: if rdev != NULL, then bio
3157                 * cannot be NULL as r10buf_pool_alloc will
3158                 * have allocated it.
3159                 * So the second test here is pointless.
3160                 * But it keeps semantic-checkers happy, and
3161                 * this comment keeps human reviewers
3162                 * happy.
3163                 */
3164                if (rdev == NULL || bio == NULL ||
3165                    test_bit(Faulty, &rdev->flags))
3166                    break;
3167                bio->bi_next = biolist;
3168                biolist = bio;
3169                bio->bi_private = r10_bio;
3170                bio->bi_end_io = end_sync_write;
3171                bio->bi_rw = WRITE;
3172                bio->bi_sector = to_addr + rdev->data_offset;
3173                bio->bi_bdev = rdev->bdev;
3174                atomic_inc(&r10_bio->remaining);
3175                break;
3176            }
3177            if (j == conf->copies) {
3178                /* Cannot recover, so abort the recovery or
3179                 * record a bad block */
3180                put_buf(r10_bio);
3181                if (rb2)
3182                    atomic_dec(&rb2->remaining);
3183                r10_bio = rb2;
3184                if (any_working) {
3185                    /* problem is that there are bad blocks
3186                     * on other device(s)
3187                     */
3188                    int k;
3189                    for (k = 0; k < conf->copies; k++)
3190                        if (r10_bio->devs[k].devnum == i)
3191                            break;
3192                    if (!test_bit(In_sync,
3193                              &mirror->rdev->flags)
3194                        && !rdev_set_badblocks(
3195                            mirror->rdev,
3196                            r10_bio->devs[k].addr,
3197                            max_sync, 0))
3198                        any_working = 0;
3199                    if (mirror->replacement &&
3200                        !rdev_set_badblocks(
3201                            mirror->replacement,
3202                            r10_bio->devs[k].addr,
3203                            max_sync, 0))
3204                        any_working = 0;
3205                }
3206                if (!any_working) {
3207                    if (!test_and_set_bit(MD_RECOVERY_INTR,
3208                                  &mddev->recovery))
3209                        printk(KERN_INFO "md/raid10:%s: insufficient "
3210                               "working devices for recovery.\n",
3211                               mdname(mddev));
3212                    mirror->recovery_disabled
3213                        = mddev->recovery_disabled;
3214                }
3215                break;
3216            }
3217        }
3218        if (biolist == NULL) {
3219            while (r10_bio) {
3220                struct r10bio *rb2 = r10_bio;
3221                r10_bio = (struct r10bio*) rb2->master_bio;
3222                rb2->master_bio = NULL;
3223                put_buf(rb2);
3224            }
3225            goto giveup;
3226        }
3227    } else {
3228        /* resync. Schedule a read for every block at this virt offset */
3229        int count = 0;
3230
3231        bitmap_cond_end_sync(mddev->bitmap, sector_nr);
3232
3233        if (!bitmap_start_sync(mddev->bitmap, sector_nr,
3234                       &sync_blocks, mddev->degraded) &&
3235            !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED,
3236                         &mddev->recovery)) {
3237            /* We can skip this block */
3238            *skipped = 1;
3239            return sync_blocks + sectors_skipped;
3240        }
3241        if (sync_blocks < max_sync)
3242            max_sync = sync_blocks;
3243        r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
3244
3245        r10_bio->mddev = mddev;
3246        atomic_set(&r10_bio->remaining, 0);
3247        raise_barrier(conf, 0);
3248        conf->next_resync = sector_nr;
3249
3250        r10_bio->master_bio = NULL;
3251        r10_bio->sector = sector_nr;
3252        set_bit(R10BIO_IsSync, &r10_bio->state);
3253        raid10_find_phys(conf, r10_bio);
3254        r10_bio->sectors = (sector_nr | chunk_mask) - sector_nr + 1;
3255
3256        for (i = 0; i < conf->copies; i++) {
3257            int d = r10_bio->devs[i].devnum;
3258            sector_t first_bad, sector;
3259            int bad_sectors;
3260
3261            if (r10_bio->devs[i].repl_bio)
3262                r10_bio->devs[i].repl_bio->bi_end_io = NULL;
3263
3264            bio = r10_bio->devs[i].bio;
3265            bio->bi_end_io = NULL;
3266            clear_bit(BIO_UPTODATE, &bio->bi_flags);
3267            if (conf->mirrors[d].rdev == NULL ||
3268                test_bit(Faulty, &conf->mirrors[d].rdev->flags))
3269                continue;
3270            sector = r10_bio->devs[i].addr;
3271            if (is_badblock(conf->mirrors[d].rdev,
3272                    sector, max_sync,
3273                    &first_bad, &bad_sectors)) {
3274                if (first_bad > sector)
3275                    max_sync = first_bad - sector;
3276                else {
3277                    bad_sectors -= (sector - first_bad);
3278                    if (max_sync > bad_sectors)
3279                        max_sync = bad_sectors;
3280                    continue;
3281                }
3282            }
3283            atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3284            atomic_inc(&r10_bio->remaining);
3285            bio->bi_next = biolist;
3286            biolist = bio;
3287            bio->bi_private = r10_bio;
3288            bio->bi_end_io = end_sync_read;
3289            bio->bi_rw = READ;
3290            bio->bi_sector = sector +
3291                conf->mirrors[d].rdev->data_offset;
3292            bio->bi_bdev = conf->mirrors[d].rdev->bdev;
3293            count++;
3294
3295            if (conf->mirrors[d].replacement == NULL ||
3296                test_bit(Faulty,
3297                     &conf->mirrors[d].replacement->flags))
3298                continue;
3299
3300            /* Need to set up for writing to the replacement */
3301            bio = r10_bio->devs[i].repl_bio;
3302            clear_bit(BIO_UPTODATE, &bio->bi_flags);
3303
3304            sector = r10_bio->devs[i].addr;
3305            atomic_inc(&conf->mirrors[d].rdev->nr_pending);
3306            bio->bi_next = biolist;
3307            biolist = bio;
3308            bio->bi_private = r10_bio;
3309            bio->bi_end_io = end_sync_write;
3310            bio->bi_rw = WRITE;
3311            bio->bi_sector = sector +
3312                conf->mirrors[d].replacement->data_offset;
3313            bio->bi_bdev = conf->mirrors[d].replacement->bdev;
3314            count++;
3315        }
3316
3317        if (count < 2) {
3318            for (i=0; i<conf->copies; i++) {
3319                int d = r10_bio->devs[i].devnum;
3320                if (r10_bio->devs[i].bio->bi_end_io)
3321                    rdev_dec_pending(conf->mirrors[d].rdev,
3322                             mddev);
3323                if (r10_bio->devs[i].repl_bio &&
3324                    r10_bio->devs[i].repl_bio->bi_end_io)
3325                    rdev_dec_pending(
3326                        conf->mirrors[d].replacement,
3327                        mddev);
3328            }
3329            put_buf(r10_bio);
3330            biolist = NULL;
3331            goto giveup;
3332        }
3333    }
3334
3335    for (bio = biolist; bio ; bio=bio->bi_next) {
3336
3337        bio->bi_flags &= ~(BIO_POOL_MASK - 1);
3338        if (bio->bi_end_io)
3339            bio->bi_flags |= 1 << BIO_UPTODATE;
3340        bio->bi_vcnt = 0;
3341        bio->bi_idx = 0;
3342        bio->bi_phys_segments = 0;
3343        bio->bi_size = 0;
3344    }
3345
3346    nr_sectors = 0;
3347    if (sector_nr + max_sync < max_sector)
3348        max_sector = sector_nr + max_sync;
3349    do {
3350        struct page *page;
3351        int len = PAGE_SIZE;
3352        if (sector_nr + (len>>9) > max_sector)
3353            len = (max_sector - sector_nr) << 9;
3354        if (len == 0)
3355            break;
3356        for (bio= biolist ; bio ; bio=bio->bi_next) {
3357            struct bio *bio2;
3358            page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
3359            if (bio_add_page(bio, page, len, 0))
3360                continue;
3361
3362            /* stop here */
3363            bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
3364            for (bio2 = biolist;
3365                 bio2 && bio2 != bio;
3366                 bio2 = bio2->bi_next) {
3367                /* remove last page from this bio */
3368                bio2->bi_vcnt--;
3369                bio2->bi_size -= len;
3370                bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
3371            }
3372            goto bio_full;
3373        }
3374        nr_sectors += len>>9;
3375        sector_nr += len>>9;
3376    } while (biolist->bi_vcnt < RESYNC_PAGES);
3377 bio_full:
3378    r10_bio->sectors = nr_sectors;
3379
3380    while (biolist) {
3381        bio = biolist;
3382        biolist = biolist->bi_next;
3383
3384        bio->bi_next = NULL;
3385        r10_bio = bio->bi_private;
3386        r10_bio->sectors = nr_sectors;
3387
3388        if (bio->bi_end_io == end_sync_read) {
3389            md_sync_acct(bio->bi_bdev, nr_sectors);
3390            generic_make_request(bio);
3391        }
3392    }
3393
3394    if (sectors_skipped)
3395        /* pretend they weren't skipped, it makes
3396         * no important difference in this case
3397         */
3398        md_done_sync(mddev, sectors_skipped, 1);
3399
3400    return sectors_skipped + nr_sectors;
3401 giveup:
3402    /* There is nowhere to write, so all non-sync
3403     * drives must be failed or in resync, all drives
3404     * have a bad block, so try the next chunk...
3405     */
3406    if (sector_nr + max_sync < max_sector)
3407        max_sector = sector_nr + max_sync;
3408
3409    sectors_skipped += (max_sector - sector_nr);
3410    chunks_skipped ++;
3411    sector_nr = max_sector;
3412    goto skipped;
3413}
3414
3415static sector_t
3416raid10_size(struct mddev *mddev, sector_t sectors, int raid_disks)
3417{
3418    sector_t size;
3419    struct r10conf *conf = mddev->private;
3420
3421    if (!raid_disks)
3422        raid_disks = min(conf->geo.raid_disks,
3423                 conf->prev.raid_disks);
3424    if (!sectors)
3425        sectors = conf->dev_sectors;
3426
3427    size = sectors >> conf->geo.chunk_shift;
3428    sector_div(size, conf->geo.far_copies);
3429    size = size * raid_disks;
3430    sector_div(size, conf->geo.near_copies);
3431
3432    return size << conf->geo.chunk_shift;
3433}
3434
3435static void calc_sectors(struct r10conf *conf, sector_t size)
3436{
3437    /* Calculate the number of sectors-per-device that will
3438     * actually be used, and set conf->dev_sectors and
3439     * conf->stride
3440     */
3441
3442    size = size >> conf->geo.chunk_shift;
3443    sector_div(size, conf->geo.far_copies);
3444    size = size * conf->geo.raid_disks;
3445    sector_div(size, conf->geo.near_copies);
3446    /* 'size' is now the number of chunks in the array */
3447    /* calculate "used chunks per device" */
3448    size = size * conf->copies;
3449
3450    /* We need to round up when dividing by raid_disks to
3451     * get the stride size.
3452     */
3453    size = DIV_ROUND_UP_SECTOR_T(size, conf->geo.raid_disks);
3454
3455    conf->dev_sectors = size << conf->geo.chunk_shift;
3456
3457    if (conf->geo.far_offset)
3458        conf->geo.stride = 1 << conf->geo.chunk_shift;
3459    else {
3460        sector_div(size, conf->geo.far_copies);
3461        conf->geo.stride = size << conf->geo.chunk_shift;
3462    }
3463}
3464
3465enum geo_type {geo_new, geo_old, geo_start};
3466static int setup_geo(struct geom *geo, struct mddev *mddev, enum geo_type new)
3467{
3468    int nc, fc, fo;
3469    int layout, chunk, disks;
3470    switch (new) {
3471    case geo_old:
3472        layout = mddev->layout;
3473        chunk = mddev->chunk_sectors;
3474        disks = mddev->raid_disks - mddev->delta_disks;
3475        break;
3476    case geo_new:
3477        layout = mddev->new_layout;
3478        chunk = mddev->new_chunk_sectors;
3479        disks = mddev->raid_disks;
3480        break;
3481    default: /* avoid 'may be unused' warnings */
3482    case geo_start: /* new when starting reshape - raid_disks not
3483             * updated yet. */
3484        layout = mddev->new_layout;
3485        chunk = mddev->new_chunk_sectors;
3486        disks = mddev->raid_disks + mddev->delta_disks;
3487        break;
3488    }
3489    if (layout >> 18)
3490        return -1;
3491    if (chunk < (PAGE_SIZE >> 9) ||
3492        !is_power_of_2(chunk))
3493        return -2;
3494    nc = layout & 255;
3495    fc = (layout >> 8) & 255;
3496    fo = layout & (1<<16);
3497    geo->raid_disks = disks;
3498    geo->near_copies = nc;
3499    geo->far_copies = fc;
3500    geo->far_offset = fo;
3501    geo->far_set_size = (layout & (1<<17)) ? disks / fc : disks;
3502    geo->chunk_mask = chunk - 1;
3503    geo->chunk_shift = ffz(~chunk);
3504    return nc*fc;
3505}
3506
3507static struct r10conf *setup_conf(struct mddev *mddev)
3508{
3509    struct r10conf *conf = NULL;
3510    int err = -EINVAL;
3511    struct geom geo;
3512    int copies;
3513
3514    copies = setup_geo(&geo, mddev, geo_new);
3515
3516    if (copies == -2) {
3517        printk(KERN_ERR "md/raid10:%s: chunk size must be "
3518               "at least PAGE_SIZE(%ld) and be a power of 2.\n",
3519               mdname(mddev), PAGE_SIZE);
3520        goto out;
3521    }
3522
3523    if (copies < 2 || copies > mddev->raid_disks) {
3524        printk(KERN_ERR "md/raid10:%s: unsupported raid10 layout: 0x%8x\n",
3525               mdname(mddev), mddev->new_layout);
3526        goto out;
3527    }
3528
3529    err = -ENOMEM;
3530    conf = kzalloc(sizeof(struct r10conf), GFP_KERNEL);
3531    if (!conf)
3532        goto out;
3533
3534    /* FIXME calc properly */
3535    conf->mirrors = kzalloc(sizeof(struct raid10_info)*(mddev->raid_disks +
3536                                max(0,mddev->delta_disks)),
3537                GFP_KERNEL);
3538    if (!conf->mirrors)
3539        goto out;
3540
3541    conf->tmppage = alloc_page(GFP_KERNEL);
3542    if (!conf->tmppage)
3543        goto out;
3544
3545    conf->geo = geo;
3546    conf->copies = copies;
3547    conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
3548                       r10bio_pool_free, conf);
3549    if (!conf->r10bio_pool)
3550        goto out;
3551
3552    calc_sectors(conf, mddev->dev_sectors);
3553    if (mddev->reshape_position == MaxSector) {
3554        conf->prev = conf->geo;
3555        conf->reshape_progress = MaxSector;
3556    } else {
3557        if (setup_geo(&conf->prev, mddev, geo_old) != conf->copies) {
3558            err = -EINVAL;
3559            goto out;
3560        }
3561        conf->reshape_progress = mddev->reshape_position;
3562        if (conf->prev.far_offset)
3563            conf->prev.stride = 1 << conf->prev.chunk_shift;
3564        else
3565            /* far_copies must be 1 */
3566            conf->prev.stride = conf->dev_sectors;
3567    }
3568    spin_lock_init(&conf->device_lock);
3569    INIT_LIST_HEAD(&conf->retry_list);
3570
3571    spin_lock_init(&conf->resync_lock);
3572    init_waitqueue_head(&conf->wait_barrier);
3573
3574    conf->thread = md_register_thread(raid10d, mddev, "raid10");
3575    if (!conf->thread)
3576        goto out;
3577
3578    conf->mddev = mddev;
3579    return conf;
3580
3581 out:
3582    if (err == -ENOMEM)
3583        printk(KERN_ERR "md/raid10:%s: couldn't allocate memory.\n",
3584               mdname(mddev));
3585    if (conf) {
3586        if (conf->r10bio_pool)
3587            mempool_destroy(conf->r10bio_pool);
3588        kfree(conf->mirrors);
3589        safe_put_page(conf->tmppage);
3590        kfree(conf);
3591    }
3592    return ERR_PTR(err);
3593}
3594
3595static int run(struct mddev *mddev)
3596{
3597    struct r10conf *conf;
3598    int i, disk_idx, chunk_size;
3599    struct raid10_info *disk;
3600    struct md_rdev *rdev;
3601    sector_t size;
3602    sector_t min_offset_diff = 0;
3603    int first = 1;
3604    bool discard_supported = false;
3605
3606    if (mddev->private == NULL) {
3607        conf = setup_conf(mddev);
3608        if (IS_ERR(conf))
3609            return PTR_ERR(conf);
3610        mddev->private = conf;
3611    }
3612    conf = mddev->private;
3613    if (!conf)
3614        goto out;
3615
3616    mddev->thread = conf->thread;
3617    conf->thread = NULL;
3618
3619    chunk_size = mddev->chunk_sectors << 9;
3620    if (mddev->queue) {
3621        blk_queue_max_discard_sectors(mddev->queue,
3622                          mddev->chunk_sectors);
3623        blk_queue_max_write_same_sectors(mddev->queue,
3624                         mddev->chunk_sectors);
3625        blk_queue_io_min(mddev->queue, chunk_size);
3626        if (conf->geo.raid_disks % conf->geo.near_copies)
3627            blk_queue_io_opt(mddev->queue, chunk_size * conf->geo.raid_disks);
3628        else
3629            blk_queue_io_opt(mddev->queue, chunk_size *
3630                     (conf->geo.raid_disks / conf->geo.near_copies));
3631    }
3632
3633    rdev_for_each(rdev, mddev) {
3634        long long diff;
3635        struct request_queue *q;
3636
3637        disk_idx = rdev->raid_disk;
3638        if (disk_idx < 0)
3639            continue;
3640        if (disk_idx >= conf->geo.raid_disks &&
3641            disk_idx >= conf->prev.raid_disks)
3642            continue;
3643        disk = conf->mirrors + disk_idx;
3644
3645        if (test_bit(Replacement, &rdev->flags)) {
3646            if (disk->replacement)
3647                goto out_free_conf;
3648            disk->replacement = rdev;
3649        } else {
3650            if (disk->rdev)
3651                goto out_free_conf;
3652            disk->rdev = rdev;
3653        }
3654        q = bdev_get_queue(rdev->bdev);
3655        if (q->merge_bvec_fn)
3656            mddev->merge_check_needed = 1;
3657        diff = (rdev->new_data_offset - rdev->data_offset);
3658        if (!mddev->reshape_backwards)
3659            diff = -diff;
3660        if (diff < 0)
3661            diff = 0;
3662        if (first || diff < min_offset_diff)
3663            min_offset_diff = diff;
3664
3665        if (mddev->gendisk)
3666            disk_stack_limits(mddev->gendisk, rdev->bdev,
3667                      rdev->data_offset << 9);
3668
3669        disk->head_position = 0;
3670
3671        if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
3672            discard_supported = true;
3673    }
3674
3675    if (mddev->queue) {
3676        if (discard_supported)
3677            queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
3678                        mddev->queue);
3679        else
3680            queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
3681                          mddev->queue);
3682    }
3683    /* need to check that every block has at least one working mirror */
3684    if (!enough(conf, -1)) {
3685        printk(KERN_ERR "md/raid10:%s: not enough operational mirrors.\n",
3686               mdname(mddev));
3687        goto out_free_conf;
3688    }
3689
3690    if (conf->reshape_progress != MaxSector) {
3691        /* must ensure that shape change is supported */
3692        if (conf->geo.far_copies != 1 &&
3693            conf->geo.far_offset == 0)
3694            goto out_free_conf;
3695        if (conf->prev.far_copies != 1 &&
3696            conf->geo.far_offset == 0)
3697            goto out_free_conf;
3698    }
3699
3700    mddev->degraded = 0;
3701    for (i = 0;
3702         i < conf->geo.raid_disks
3703             || i < conf->prev.raid_disks;
3704         i++) {
3705
3706        disk = conf->mirrors + i;
3707
3708        if (!disk->rdev && disk->replacement) {
3709            /* The replacement is all we have - use it */
3710            disk->rdev = disk->replacement;
3711            disk->replacement = NULL;
3712            clear_bit(Replacement, &disk->rdev->flags);
3713        }
3714
3715        if (!disk->rdev ||
3716            !test_bit(In_sync, &disk->rdev->flags)) {
3717            disk->head_position = 0;
3718            mddev->degraded++;
3719            if (disk->rdev)
3720                conf->fullsync = 1;
3721        }
3722        disk->recovery_disabled = mddev->recovery_disabled - 1;
3723    }
3724
3725    if (mddev->recovery_cp != MaxSector)
3726        printk(KERN_NOTICE "md/raid10:%s: not clean"
3727               " -- starting background reconstruction\n",
3728               mdname(mddev));
3729    printk(KERN_INFO
3730        "md/raid10:%s: active with %d out of %d devices\n",
3731        mdname(mddev), conf->geo.raid_disks - mddev->degraded,
3732        conf->geo.raid_disks);
3733    /*
3734     * Ok, everything is just fine now
3735     */
3736    mddev->dev_sectors = conf->dev_sectors;
3737    size = raid10_size(mddev, 0, 0);
3738    md_set_array_sectors(mddev, size);
3739    mddev->resync_max_sectors = size;
3740
3741    if (mddev->queue) {
3742        int stripe = conf->geo.raid_disks *
3743            ((mddev->chunk_sectors << 9) / PAGE_SIZE);
3744        mddev->queue->backing_dev_info.congested_fn = raid10_congested;
3745        mddev->queue->backing_dev_info.congested_data = mddev;
3746
3747        /* Calculate max read-ahead size.
3748         * We need to readahead at least twice a whole stripe....
3749         * maybe...
3750         */
3751        stripe /= conf->geo.near_copies;
3752        if (mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
3753            mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
3754        blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
3755    }
3756
3757
3758    if (md_integrity_register(mddev))
3759        goto out_free_conf;
3760
3761    if (conf->reshape_progress != MaxSector) {
3762        unsigned long before_length, after_length;
3763
3764        before_length = ((1 << conf->prev.chunk_shift) *
3765                 conf->prev.far_copies);
3766        after_length = ((1 << conf->geo.chunk_shift) *
3767                conf->geo.far_copies);
3768
3769        if (max(before_length, after_length) > min_offset_diff) {
3770            /* This cannot work */
3771            printk("md/raid10: offset difference not enough to continue reshape\n");
3772            goto out_free_conf;
3773        }
3774        conf->offset_diff = min_offset_diff;
3775
3776        conf->reshape_safe = conf->reshape_progress;
3777        clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
3778        clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
3779        set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
3780        set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
3781        mddev->sync_thread = md_register_thread(md_do_sync, mddev,
3782                            "reshape");
3783    }
3784
3785    return 0;
3786
3787out_free_conf:
3788    md_unregister_thread(&mddev->thread);
3789    if (conf->r10bio_pool)
3790        mempool_destroy(conf->r10bio_pool);
3791    safe_put_page(conf->tmppage);
3792    kfree(conf->mirrors);
3793    kfree(conf);
3794    mddev->private = NULL;
3795out:
3796    return -EIO;
3797}
3798
3799static int stop(struct mddev *mddev)
3800{
3801    struct r10conf *conf = mddev->private;
3802
3803    raise_barrier(conf, 0);
3804    lower_barrier(conf);
3805
3806    md_unregister_thread(&mddev->thread);
3807    if (mddev->queue)
3808        /* the unplug fn references 'conf'*/
3809        blk_sync_queue(mddev->queue);
3810
3811    if (conf->r10bio_pool)
3812        mempool_destroy(conf->r10bio_pool);
3813    kfree(conf->mirrors);
3814    kfree(conf);
3815    mddev->private = NULL;
3816    return 0;
3817}
3818
3819static void raid10_quiesce(struct mddev *mddev, int state)
3820{
3821    struct r10conf *conf = mddev->private;
3822
3823    switch(state) {
3824    case 1:
3825        raise_barrier(conf, 0);
3826        break;
3827    case 0:
3828        lower_barrier(conf);
3829        break;
3830    }
3831}
3832
3833static int raid10_resize(struct mddev *mddev, sector_t sectors)
3834{
3835    /* Resize of 'far' arrays is not supported.
3836     * For 'near' and 'offset' arrays we can set the
3837     * number of sectors used to be an appropriate multiple
3838     * of the chunk size.
3839     * For 'offset', this is far_copies*chunksize.
3840     * For 'near' the multiplier is the LCM of
3841     * near_copies and raid_disks.
3842     * So if far_copies > 1 && !far_offset, fail.
3843     * Else find LCM(raid_disks, near_copy)*far_copies and
3844     * multiply by chunk_size. Then round to this number.
3845     * This is mostly done by raid10_size()
3846     */
3847    struct r10conf *conf = mddev->private;
3848    sector_t oldsize, size;
3849
3850    if (mddev->reshape_position != MaxSector)
3851        return -EBUSY;
3852
3853    if (conf->geo.far_copies > 1 && !conf->geo.far_offset)
3854        return -EINVAL;
3855
3856    oldsize = raid10_size(mddev, 0, 0);
3857    size = raid10_size(mddev, sectors, 0);
3858    if (mddev->external_size &&
3859        mddev->array_sectors > size)
3860        return -EINVAL;
3861    if (mddev->bitmap) {
3862        int ret = bitmap_resize(mddev->bitmap, size, 0, 0);
3863        if (ret)
3864            return ret;
3865    }
3866    md_set_array_sectors(mddev, size);
3867    set_capacity(mddev->gendisk, mddev->array_sectors);
3868    revalidate_disk(mddev->gendisk);
3869    if (sectors > mddev->dev_sectors &&
3870        mddev->recovery_cp > oldsize) {
3871        mddev->recovery_cp = oldsize;
3872        set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3873    }
3874    calc_sectors(conf, sectors);
3875    mddev->dev_sectors = conf->dev_sectors;
3876    mddev->resync_max_sectors = size;
3877    return 0;
3878}
3879
3880static void *raid10_takeover_raid0(struct mddev *mddev)
3881{
3882    struct md_rdev *rdev;
3883    struct r10conf *conf;
3884
3885    if (mddev->degraded > 0) {
3886        printk(KERN_ERR "md/raid10:%s: Error: degraded raid0!\n",
3887               mdname(mddev));
3888        return ERR_PTR(-EINVAL);
3889    }
3890
3891    /* Set new parameters */
3892    mddev->new_level = 10;
3893    /* new layout: far_copies = 1, near_copies = 2 */
3894    mddev->new_layout = (1<<8) + 2;
3895    mddev->new_chunk_sectors = mddev->chunk_sectors;
3896    mddev->delta_disks = mddev->raid_disks;
3897    mddev->raid_disks *= 2;
3898    /* make sure it will be not marked as dirty */
3899    mddev->recovery_cp = MaxSector;
3900
3901    conf = setup_conf(mddev);
3902    if (!IS_ERR(conf)) {
3903        rdev_for_each(rdev, mddev)
3904            if (rdev->raid_disk >= 0)
3905                rdev->new_raid_disk = rdev->raid_disk * 2;
3906        conf->barrier = 1;
3907    }
3908
3909    return conf;
3910}
3911
3912static void *raid10_takeover(struct mddev *mddev)
3913{
3914    struct r0conf *raid0_conf;
3915
3916    /* raid10 can take over:
3917     * raid0 - providing it has only two drives
3918     */
3919    if (mddev->level == 0) {
3920        /* for raid0 takeover only one zone is supported */
3921        raid0_conf = mddev->private;
3922        if (raid0_conf->nr_strip_zones > 1) {
3923            printk(KERN_ERR "md/raid10:%s: cannot takeover raid 0"
3924                   " with more than one zone.\n",
3925                   mdname(mddev));
3926            return ERR_PTR(-EINVAL);
3927        }
3928        return raid10_takeover_raid0(mddev);
3929    }
3930    return ERR_PTR(-EINVAL);
3931}
3932
3933static int raid10_check_reshape(struct mddev *mddev)
3934{
3935    /* Called when there is a request to change
3936     * - layout (to ->new_layout)
3937     * - chunk size (to ->new_chunk_sectors)
3938     * - raid_disks (by delta_disks)
3939     * or when trying to restart a reshape that was ongoing.
3940     *
3941     * We need to validate the request and possibly allocate
3942     * space if that might be an issue later.
3943     *
3944     * Currently we reject any reshape of a 'far' mode array,
3945     * allow chunk size to change if new is generally acceptable,
3946     * allow raid_disks to increase, and allow
3947     * a switch between 'near' mode and 'offset' mode.
3948     */
3949    struct r10conf *conf = mddev->private;
3950    struct geom geo;
3951
3952    if (conf->geo.far_copies != 1 && !conf->geo.far_offset)
3953        return -EINVAL;
3954
3955    if (setup_geo(&geo, mddev, geo_start) != conf->copies)
3956        /* mustn't change number of copies */
3957        return -EINVAL;
3958    if (geo.far_copies > 1 && !geo.far_offset)
3959        /* Cannot switch to 'far' mode */
3960        return -EINVAL;
3961
3962    if (mddev->array_sectors & geo.chunk_mask)
3963            /* not factor of array size */
3964            return -EINVAL;
3965
3966    if (!enough(conf, -1))
3967        return -EINVAL;
3968
3969    kfree(conf->mirrors_new);
3970    conf->mirrors_new = NULL;
3971    if (mddev->delta_disks > 0) {
3972        /* allocate new 'mirrors' list */
3973        conf->mirrors_new = kzalloc(
3974            sizeof(struct raid10_info)
3975            *(mddev->raid_disks +
3976              mddev->delta_disks),
3977            GFP_KERNEL);
3978        if (!conf->mirrors_new)
3979            return -ENOMEM;
3980    }
3981    return 0;
3982}
3983
3984/*
3985 * Need to check if array has failed when deciding whether to:
3986 * - start an array
3987 * - remove non-faulty devices
3988 * - add a spare
3989 * - allow a reshape
3990 * This determination is simple when no reshape is happening.
3991 * However if there is a reshape, we need to carefully check
3992 * both the before and after sections.
3993 * This is because some failed devices may only affect one
3994 * of the two sections, and some non-in_sync devices may
3995 * be insync in the section most affected by failed devices.
3996 */
3997static int calc_degraded(struct r10conf *conf)
3998{
3999    int degraded, degraded2;
4000    int i;
4001
4002    rcu_read_lock();
4003    degraded = 0;
4004    /* 'prev' section first */
4005    for (i = 0; i < conf->prev.raid_disks; i++) {
4006        struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4007        if (!rdev || test_bit(Faulty, &rdev->flags))
4008            degraded++;
4009        else if (!test_bit(In_sync, &rdev->flags))
4010            /* When we can reduce the number of devices in
4011             * an array, this might not contribute to
4012             * 'degraded'. It does now.
4013             */
4014            degraded++;
4015    }
4016    rcu_read_unlock();
4017    if (conf->geo.raid_disks == conf->prev.raid_disks)
4018        return degraded;
4019    rcu_read_lock();
4020    degraded2 = 0;
4021    for (i = 0; i < conf->geo.raid_disks; i++) {
4022        struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
4023        if (!rdev || test_bit(Faulty, &rdev->flags))
4024            degraded2++;
4025        else if (!test_bit(In_sync, &rdev->flags)) {
4026            /* If reshape is increasing the number of devices,
4027             * this section has already been recovered, so
4028             * it doesn't contribute to degraded.
4029             * else it does.
4030             */
4031            if (conf->geo.raid_disks <= conf->prev.raid_disks)
4032                degraded2++;
4033        }
4034    }
4035    rcu_read_unlock();
4036    if (degraded2 > degraded)
4037        return degraded2;
4038    return degraded;
4039}
4040
4041static int raid10_start_reshape(struct mddev *mddev)
4042{
4043    /* A 'reshape' has been requested. This commits
4044     * the various 'new' fields and sets MD_RECOVER_RESHAPE
4045     * This also checks if there are enough spares and adds them
4046     * to the array.
4047     * We currently require enough spares to make the final
4048     * array non-degraded. We also require that the difference
4049     * between old and new data_offset - on each device - is
4050     * enough that we never risk over-writing.
4051     */
4052
4053    unsigned long before_length, after_length;
4054    sector_t min_offset_diff = 0;
4055    int first = 1;
4056    struct geom new;
4057    struct r10conf *conf = mddev->private;
4058    struct md_rdev *rdev;
4059    int spares = 0;
4060    int ret;
4061
4062    if (test_bit(MD_RECOVERY_RUNNING, &mddev->recovery))
4063        return -EBUSY;
4064
4065    if (setup_geo(&new, mddev, geo_start) != conf->copies)
4066        return -EINVAL;
4067
4068    before_length = ((1 << conf->prev.chunk_shift) *
4069             conf->prev.far_copies);
4070    after_length = ((1 << conf->geo.chunk_shift) *
4071            conf->geo.far_copies);
4072
4073    rdev_for_each(rdev, mddev) {
4074        if (!test_bit(In_sync, &rdev->flags)
4075            && !test_bit(Faulty, &rdev->flags))
4076            spares++;
4077        if (rdev->raid_disk >= 0) {
4078            long long diff = (rdev->new_data_offset
4079                      - rdev->data_offset);
4080            if (!mddev->reshape_backwards)
4081                diff = -diff;
4082            if (diff < 0)
4083                diff = 0;
4084            if (first || diff < min_offset_diff)
4085                min_offset_diff = diff;
4086        }
4087    }
4088
4089    if (max(before_length, after_length) > min_offset_diff)
4090        return -EINVAL;
4091
4092    if (spares < mddev->delta_disks)
4093        return -EINVAL;
4094
4095    conf->offset_diff = min_offset_diff;
4096    spin_lock_irq(&conf->device_lock);
4097    if (conf->mirrors_new) {
4098        memcpy(conf->mirrors_new, conf->mirrors,
4099               sizeof(struct raid10_info)*conf->prev.raid_disks);
4100        smp_mb();
4101        kfree(conf->mirrors_old); /* FIXME and elsewhere */
4102        conf->mirrors_old = conf->mirrors;
4103        conf->mirrors = conf->mirrors_new;
4104        conf->mirrors_new = NULL;
4105    }
4106    setup_geo(&conf->geo, mddev, geo_start);
4107    smp_mb();
4108    if (mddev->reshape_backwards) {
4109        sector_t size = raid10_size(mddev, 0, 0);
4110        if (size < mddev->array_sectors) {
4111            spin_unlock_irq(&conf->device_lock);
4112            printk(KERN_ERR "md/raid10:%s: array size must be reduce before number of disks\n",
4113                   mdname(mddev));
4114            return -EINVAL;
4115        }
4116        mddev->resync_max_sectors = size;
4117        conf->reshape_progress = size;
4118    } else
4119        conf->reshape_progress = 0;
4120    spin_unlock_irq(&conf->device_lock);
4121
4122    if (mddev->delta_disks && mddev->bitmap) {
4123        ret = bitmap_resize(mddev->bitmap,
4124                    raid10_size(mddev, 0,
4125                        conf->geo.raid_disks),
4126                    0, 0);
4127        if (ret)
4128            goto abort;
4129    }
4130    if (mddev->delta_disks > 0) {
4131        rdev_for_each(rdev, mddev)
4132            if (rdev->raid_disk < 0 &&
4133                !test_bit(Faulty, &rdev->flags)) {
4134                if (raid10_add_disk(mddev, rdev) == 0) {
4135                    if (rdev->raid_disk >=
4136                        conf->prev.raid_disks)
4137                        set_bit(In_sync, &rdev->flags);
4138                    else
4139                        rdev->recovery_offset = 0;
4140
4141                    if (sysfs_link_rdev(mddev, rdev))
4142                        /* Failure here is OK */;
4143                }
4144            } else if (rdev->raid_disk >= conf->prev.raid_disks
4145                   && !test_bit(Faulty, &rdev->flags)) {
4146                /* This is a spare that was manually added */
4147                set_bit(In_sync, &rdev->flags);
4148            }
4149    }
4150    /* When a reshape changes the number of devices,
4151     * ->degraded is measured against the larger of the
4152     * pre and post numbers.
4153     */
4154    spin_lock_irq(&conf->device_lock);
4155    mddev->degraded = calc_degraded(conf);
4156    spin_unlock_irq(&conf->device_lock);
4157    mddev->raid_disks = conf->geo.raid_disks;
4158    mddev->reshape_position = conf->reshape_progress;
4159    set_bit(MD_CHANGE_DEVS, &mddev->flags);
4160
4161    clear_bit(MD_RECOVERY_SYNC, &mddev->recovery);
4162    clear_bit(MD_RECOVERY_CHECK, &mddev->recovery);
4163    set_bit(MD_RECOVERY_RESHAPE, &mddev->recovery);
4164    set_bit(MD_RECOVERY_RUNNING, &mddev->recovery);
4165
4166    mddev->sync_thread = md_register_thread(md_do_sync, mddev,
4167                        "reshape");
4168    if (!mddev->sync_thread) {
4169        ret = -EAGAIN;
4170        goto abort;
4171    }
4172    conf->reshape_checkpoint = jiffies;
4173    md_wakeup_thread(mddev->sync_thread);
4174    md_new_event(mddev);
4175    return 0;
4176
4177abort:
4178    mddev->recovery = 0;
4179    spin_lock_irq(&conf->device_lock);
4180    conf->geo = conf->prev;
4181    mddev->raid_disks = conf->geo.raid_disks;
4182    rdev_for_each(rdev, mddev)
4183        rdev->new_data_offset = rdev->data_offset;
4184    smp_wmb();
4185    conf->reshape_progress = MaxSector;
4186    mddev->reshape_position = MaxSector;
4187    spin_unlock_irq(&conf->device_lock);
4188    return ret;
4189}
4190
4191/* Calculate the last device-address that could contain
4192 * any block from the chunk that includes the array-address 's'
4193 * and report the next address.
4194 * i.e. the address returned will be chunk-aligned and after
4195 * any data that is in the chunk containing 's'.
4196 */
4197static sector_t last_dev_address(sector_t s, struct geom *geo)
4198{
4199    s = (s | geo->chunk_mask) + 1;
4200    s >>= geo->chunk_shift;
4201    s *= geo->near_copies;
4202    s = DIV_ROUND_UP_SECTOR_T(s, geo->raid_disks);
4203    s *= geo->far_copies;
4204    s <<= geo->chunk_shift;
4205    return s;
4206}
4207
4208/* Calculate the first device-address that could contain
4209 * any block from the chunk that includes the array-address 's'.
4210 * This too will be the start of a chunk
4211 */
4212static sector_t first_dev_address(sector_t s, struct geom *geo)
4213{
4214    s >>= geo->chunk_shift;
4215    s *= geo->near_copies;
4216    sector_div(s, geo->raid_disks);
4217    s *= geo->far_copies;
4218    s <<= geo->chunk_shift;
4219    return s;
4220}
4221
4222static sector_t reshape_request(struct mddev *mddev, sector_t sector_nr,
4223                int *skipped)
4224{
4225    /* We simply copy at most one chunk (smallest of old and new)
4226     * at a time, possibly less if that exceeds RESYNC_PAGES,
4227     * or we hit a bad block or something.
4228     * This might mean we pause for normal IO in the middle of
4229     * a chunk, but that is not a problem was mddev->reshape_position
4230     * can record any location.
4231     *
4232     * If we will want to write to a location that isn't
4233     * yet recorded as 'safe' (i.e. in metadata on disk) then
4234     * we need to flush all reshape requests and update the metadata.
4235     *
4236     * When reshaping forwards (e.g. to more devices), we interpret
4237     * 'safe' as the earliest block which might not have been copied
4238     * down yet. We divide this by previous stripe size and multiply
4239     * by previous stripe length to get lowest device offset that we
4240     * cannot write to yet.
4241     * We interpret 'sector_nr' as an address that we want to write to.
4242     * From this we use last_device_address() to find where we might
4243     * write to, and first_device_address on the 'safe' position.
4244     * If this 'next' write position is after the 'safe' position,
4245     * we must update the metadata to increase the 'safe' position.
4246     *
4247     * When reshaping backwards, we round in the opposite direction
4248     * and perform the reverse test: next write position must not be
4249     * less than current safe position.
4250     *
4251     * In all this the minimum difference in data offsets
4252     * (conf->offset_diff - always positive) allows a bit of slack,
4253     * so next can be after 'safe', but not by more than offset_disk
4254     *
4255     * We need to prepare all the bios here before we start any IO
4256     * to ensure the size we choose is acceptable to all devices.
4257     * The means one for each copy for write-out and an extra one for
4258     * read-in.
4259     * We store the read-in bio in ->master_bio and the others in
4260     * ->devs[x].bio and ->devs[x].repl_bio.
4261     */
4262    struct r10conf *conf = mddev->private;
4263    struct r10bio *r10_bio;
4264    sector_t next, safe, last;
4265    int max_sectors;
4266    int nr_sectors;
4267    int s;
4268    struct md_rdev *rdev;
4269    int need_flush = 0;
4270    struct bio *blist;
4271    struct bio *bio, *read_bio;
4272    int sectors_done = 0;
4273
4274    if (sector_nr == 0) {
4275        /* If restarting in the middle, skip the initial sectors */
4276        if (mddev->reshape_backwards &&
4277            conf->reshape_progress < raid10_size(mddev, 0, 0)) {
4278            sector_nr = (raid10_size(mddev, 0, 0)
4279                     - conf->reshape_progress);
4280        } else if (!mddev->reshape_backwards &&
4281               conf->reshape_progress > 0)
4282            sector_nr = conf->reshape_progress;
4283        if (sector_nr) {
4284            mddev->curr_resync_completed = sector_nr;
4285            sysfs_notify(&mddev->kobj, NULL, "sync_completed");
4286            *skipped = 1;
4287            return sector_nr;
4288        }
4289    }
4290
4291    /* We don't use sector_nr to track where we are up to
4292     * as that doesn't work well for ->reshape_backwards.
4293     * So just use ->reshape_progress.
4294     */
4295    if (mddev->reshape_backwards) {
4296        /* 'next' is the earliest device address that we might
4297         * write to for this chunk in the new layout
4298         */
4299        next = first_dev_address(conf->reshape_progress - 1,
4300                     &conf->geo);
4301
4302        /* 'safe' is the last device address that we might read from
4303         * in the old layout after a restart
4304         */
4305        safe = last_dev_address(conf->reshape_safe - 1,
4306                    &conf->prev);
4307
4308        if (next + conf->offset_diff < safe)
4309            need_flush = 1;
4310
4311        last = conf->reshape_progress - 1;
4312        sector_nr = last & ~(sector_t)(conf->geo.chunk_mask
4313                           & conf->prev.chunk_mask);
4314        if (sector_nr + RESYNC_BLOCK_SIZE/512 < last)
4315            sector_nr = last + 1 - RESYNC_BLOCK_SIZE/512;
4316    } else {
4317        /* 'next' is after the last device address that we
4318         * might write to for this chunk in the new layout
4319         */
4320        next = last_dev_address(conf->reshape_progress, &conf->geo);
4321
4322        /* 'safe' is the earliest device address that we might
4323         * read from in the old layout after a restart
4324         */
4325        safe = first_dev_address(conf->reshape_safe, &conf->prev);
4326
4327        /* Need to update metadata if 'next' might be beyond 'safe'
4328         * as that would possibly corrupt data
4329         */
4330        if (next > safe + conf->offset_diff)
4331            need_flush = 1;
4332
4333        sector_nr = conf->reshape_progress;
4334        last = sector_nr | (conf->geo.chunk_mask
4335                     & conf->prev.chunk_mask);
4336
4337        if (sector_nr + RESYNC_BLOCK_SIZE/512 <= last)
4338            last = sector_nr + RESYNC_BLOCK_SIZE/512 - 1;
4339    }
4340
4341    if (need_flush ||
4342        time_after(jiffies, conf->reshape_checkpoint + 10*HZ)) {
4343        /* Need to update reshape_position in metadata */
4344        wait_barrier(conf);
4345        mddev->reshape_position = conf->reshape_progress;
4346        if (mddev->reshape_backwards)
4347            mddev->curr_resync_completed = raid10_size(mddev, 0, 0)
4348                - conf->reshape_progress;
4349        else
4350            mddev->curr_resync_completed = conf->reshape_progress;
4351        conf->reshape_checkpoint = jiffies;
4352        set_bit(MD_CHANGE_DEVS, &mddev->flags);
4353        md_wakeup_thread(mddev->thread);
4354        wait_event(mddev->sb_wait, mddev->flags == 0 ||
4355               kthread_should_stop());
4356        conf->reshape_safe = mddev->reshape_position;
4357        allow_barrier(conf);
4358    }
4359
4360read_more:
4361    /* Now schedule reads for blocks from sector_nr to last */
4362    r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
4363    raise_barrier(conf, sectors_done != 0);
4364    atomic_set(&r10_bio->remaining, 0);
4365    r10_bio->mddev = mddev;
4366    r10_bio->sector = sector_nr;
4367    set_bit(R10BIO_IsReshape, &r10_bio->state);
4368    r10_bio->sectors = last - sector_nr + 1;
4369    rdev = read_balance(conf, r10_bio, &max_sectors);
4370    BUG_ON(!test_bit(R10BIO_Previous, &r10_bio->state));
4371
4372    if (!rdev) {
4373        /* Cannot read from here, so need to record bad blocks
4374         * on all the target devices.
4375         */
4376        // FIXME
4377        set_bit(MD_RECOVERY_INTR, &mddev->recovery);
4378        return sectors_done;
4379    }
4380
4381    read_bio = bio_alloc_mddev(GFP_KERNEL, RESYNC_PAGES, mddev);
4382
4383    read_bio->bi_bdev = rdev->bdev;
4384    read_bio->bi_sector = (r10_bio->devs[r10_bio->read_slot].addr
4385                   + rdev->data_offset);
4386    read_bio->bi_private = r10_bio;
4387    read_bio->bi_end_io = end_sync_read;
4388    read_bio->bi_rw = READ;
4389    read_bio->bi_flags &= ~(BIO_POOL_MASK - 1);
4390    read_bio->bi_flags |= 1 << BIO_UPTODATE;
4391    read_bio->bi_vcnt = 0;
4392    read_bio->bi_idx = 0;
4393    read_bio->bi_size = 0;
4394    r10_bio->master_bio = read_bio;
4395    r10_bio->read_slot = r10_bio->devs[r10_bio->read_slot].devnum;
4396
4397    /* Now find the locations in the new layout */
4398    __raid10_find_phys(&conf->geo, r10_bio);
4399
4400    blist = read_bio;
4401    read_bio->bi_next = NULL;
4402
4403    for (s = 0; s < conf->copies*2; s++) {
4404        struct bio *b;
4405        int d = r10_bio->devs[s/2].devnum;
4406        struct md_rdev *rdev2;
4407        if (s&1) {
4408            rdev2 = conf->mirrors[d].replacement;
4409            b = r10_bio->devs[s/2].repl_bio;
4410        } else {
4411            rdev2 = conf->mirrors[d].rdev;
4412            b = r10_bio->devs[s/2].bio;
4413        }
4414        if (!rdev2 || test_bit(Faulty, &rdev2->flags))
4415            continue;
4416        b->bi_bdev = rdev2->bdev;
4417        b->bi_sector = r10_bio->devs[s/2].addr + rdev2->new_data_offset;
4418        b->bi_private = r10_bio;
4419        b->bi_end_io = end_reshape_write;
4420        b->bi_rw = WRITE;
4421        b->bi_flags &= ~(BIO_POOL_MASK - 1);
4422        b->bi_flags |= 1 << BIO_UPTODATE;
4423        b->bi_next = blist;
4424        b->bi_vcnt = 0;
4425        b->bi_idx = 0;
4426        b->bi_size = 0;
4427        blist = b;
4428    }
4429
4430    /* Now add as many pages as possible to all of these bios. */
4431
4432    nr_sectors = 0;
4433    for (s = 0 ; s < max_sectors; s += PAGE_SIZE >> 9) {
4434        struct page *page = r10_bio->devs[0].bio->bi_io_vec[s/(PAGE_SIZE>>9)].bv_page;
4435        int len = (max_sectors - s) << 9;
4436        if (len > PAGE_SIZE)
4437            len = PAGE_SIZE;
4438        for (bio = blist; bio ; bio = bio->bi_next) {
4439            struct bio *bio2;
4440            if (bio_add_page(bio, page, len, 0))
4441                continue;
4442
4443            /* Didn't fit, must stop */
4444            for (bio2 = blist;
4445                 bio2 && bio2 != bio;
4446                 bio2 = bio2->bi_next) {
4447                /* Remove last page from this bio */
4448                bio2->bi_vcnt--;
4449                bio2->bi_size -= len;
4450                bio2->bi_flags &= ~(1<<BIO_SEG_VALID);
4451            }
4452            goto bio_full;
4453        }
4454        sector_nr += len >> 9;
4455        nr_sectors += len >> 9;
4456    }
4457bio_full:
4458    r10_bio->sectors = nr_sectors;
4459
4460    /* Now submit the read */
4461    md_sync_acct(read_bio->bi_bdev, r10_bio->sectors);
4462    atomic_inc(&r10_bio->remaining);
4463    read_bio->bi_next = NULL;
4464    generic_make_request(read_bio);
4465    sector_nr += nr_sectors;
4466    sectors_done += nr_sectors;
4467    if (sector_nr <= last)
4468        goto read_more;
4469
4470    /* Now that we have done the whole section we can
4471     * update reshape_progress
4472     */
4473    if (mddev->reshape_backwards)
4474        conf->reshape_progress -= sectors_done;
4475    else
4476        conf->reshape_progress += sectors_done;
4477
4478    return sectors_done;
4479}
4480
4481static void end_reshape_request(struct r10bio *r10_bio);
4482static int handle_reshape_read_error(struct mddev *mddev,
4483                     struct r10bio *r10_bio);
4484static void reshape_request_write(struct mddev *mddev, struct r10bio *r10_bio)
4485{
4486    /* Reshape read completed. Hopefully we have a block
4487     * to write out.
4488     * If we got a read error then we do sync 1-page reads from
4489     * elsewhere until we find the data - or give up.
4490     */
4491    struct r10conf *conf = mddev->private;
4492    int s;
4493
4494    if (!test_bit(R10BIO_Uptodate, &r10_bio->state))
4495        if (handle_reshape_read_error(mddev, r10_bio) < 0) {
4496            /* Reshape has been aborted */
4497            md_done_sync(mddev, r10_bio->sectors, 0);
4498            return;
4499        }
4500
4501    /* We definitely have the data in the pages, schedule the
4502     * writes.
4503     */
4504    atomic_set(&r10_bio->remaining, 1);
4505    for (s = 0; s < conf->copies*2; s++) {
4506        struct bio *b;
4507        int d = r10_bio->devs[s/2].devnum;
4508        struct md_rdev *rdev;
4509        if (s&1) {
4510            rdev = conf->mirrors[d].replacement;
4511            b = r10_bio->devs[s/2].repl_bio;
4512        } else {
4513            rdev = conf->mirrors[d].rdev;
4514            b = r10_bio->devs[s/2].bio;
4515        }
4516        if (!rdev || test_bit(Faulty, &rdev->flags))
4517            continue;
4518        atomic_inc(&rdev->nr_pending);
4519        md_sync_acct(b->bi_bdev, r10_bio->sectors);
4520        atomic_inc(&r10_bio->remaining);
4521        b->bi_next = NULL;
4522        generic_make_request(b);
4523    }
4524    end_reshape_request(r10_bio);
4525}
4526
4527static void end_reshape(struct r10conf *conf)
4528{
4529    if (test_bit(MD_RECOVERY_INTR, &conf->mddev->recovery))
4530        return;
4531
4532    spin_lock_irq(&conf->device_lock);
4533    conf->prev = conf->geo;
4534    md_finish_reshape(conf->mddev);
4535    smp_wmb();
4536    conf->reshape_progress = MaxSector;
4537    spin_unlock_irq(&conf->device_lock);
4538
4539    /* read-ahead size must cover two whole stripes, which is
4540     * 2 * (datadisks) * chunksize where 'n' is the number of raid devices
4541     */
4542    if (conf->mddev->queue) {
4543        int stripe = conf->geo.raid_disks *
4544            ((conf->mddev->chunk_sectors << 9) / PAGE_SIZE);
4545        stripe /= conf->geo.near_copies;
4546        if (conf->mddev->queue->backing_dev_info.ra_pages < 2 * stripe)
4547            conf->mddev->queue->backing_dev_info.ra_pages = 2 * stripe;
4548    }
4549    conf->fullsync = 0;
4550}
4551
4552
4553static int handle_reshape_read_error(struct mddev *mddev,
4554                     struct r10bio *r10_bio)
4555{
4556    /* Use sync reads to get the blocks from somewhere else */
4557    int sectors = r10_bio->sectors;
4558    struct r10conf *conf = mddev->private;
4559    struct {
4560        struct r10bio r10_bio;
4561        struct r10dev devs[conf->copies];
4562    } on_stack;
4563    struct r10bio *r10b = &on_stack.r10_bio;
4564    int slot = 0;
4565    int idx = 0;
4566    struct bio_vec *bvec = r10_bio->master_bio->bi_io_vec;
4567
4568    r10b->sector = r10_bio->sector;
4569    __raid10_find_phys(&conf->prev, r10b);
4570
4571    while (sectors) {
4572        int s = sectors;
4573        int success = 0;
4574        int first_slot = slot;
4575
4576        if (s > (PAGE_SIZE >> 9))
4577            s = PAGE_SIZE >> 9;
4578
4579        while (!success) {
4580            int d = r10b->devs[slot].devnum;
4581            struct md_rdev *rdev = conf->mirrors[d].rdev;
4582            sector_t addr;
4583            if (rdev == NULL ||
4584                test_bit(Faulty, &rdev->flags) ||
4585                !test_bit(In_sync, &rdev->flags))
4586                goto failed;
4587
4588            addr = r10b->devs[slot].addr + idx * PAGE_SIZE;
4589            success = sync_page_io(rdev,
4590                           addr,
4591                           s << 9,
4592                           bvec[idx].bv_page,
4593                           READ, false);
4594            if (success)
4595                break;
4596        failed:
4597            slot++;
4598            if (slot >= conf->copies)
4599                slot = 0;
4600            if (slot == first_slot)
4601                break;
4602        }
4603        if (!success) {
4604            /* couldn't read this block, must give up */
4605            set_bit(MD_RECOVERY_INTR,
4606                &mddev->recovery);
4607            return -EIO;
4608        }
4609        sectors -= s;
4610        idx++;
4611    }
4612    return 0;
4613}
4614
4615static void end_reshape_write(struct bio *bio, int error)
4616{
4617    int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
4618    struct r10bio *r10_bio = bio->bi_private;
4619    struct mddev *mddev = r10_bio->mddev;
4620    struct r10conf *conf = mddev->private;
4621    int d;
4622    int slot;
4623    int repl;
4624    struct md_rdev *rdev = NULL;
4625
4626    d = find_bio_disk(conf, r10_bio, bio, &slot, &repl);
4627    if (repl)
4628        rdev = conf->mirrors[d].replacement;
4629    if (!rdev) {
4630        smp_mb();
4631        rdev = conf->mirrors[d].rdev;
4632    }
4633
4634    if (!uptodate) {
4635        /* FIXME should record badblock */
4636        md_error(mddev, rdev);
4637    }
4638
4639    rdev_dec_pending(rdev, mddev);
4640    end_reshape_request(r10_bio);
4641}
4642
4643static void end_reshape_request(struct r10bio *r10_bio)
4644{
4645    if (!atomic_dec_and_test(&r10_bio->remaining))
4646        return;
4647    md_done_sync(r10_bio->mddev, r10_bio->sectors, 1);
4648    bio_put(r10_bio->master_bio);
4649    put_buf(r10_bio);
4650}
4651
4652static void raid10_finish_reshape(struct mddev *mddev)
4653{
4654    struct r10conf *conf = mddev->private;
4655
4656    if (test_bit(MD_RECOVERY_INTR, &mddev->recovery))
4657        return;
4658
4659    if (mddev->delta_disks > 0) {
4660        sector_t size = raid10_size(mddev, 0, 0);
4661        md_set_array_sectors(mddev, size);
4662        if (mddev->recovery_cp > mddev->resync_max_sectors) {
4663            mddev->recovery_cp = mddev->resync_max_sectors;
4664            set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
4665        }
4666        mddev->resync_max_sectors = size;
4667        set_capacity(mddev->gendisk, mddev->array_sectors);
4668        revalidate_disk(mddev->gendisk);
4669    } else {
4670        int d;
4671        for (d = conf->geo.raid_disks ;
4672             d < conf->geo.raid_disks - mddev->delta_disks;
4673             d++) {
4674            struct md_rdev *rdev = conf->mirrors[d].rdev;
4675            if (rdev)
4676                clear_bit(In_sync, &rdev->flags);
4677            rdev = conf->mirrors[d].replacement;
4678            if (rdev)
4679                clear_bit(In_sync, &rdev->flags);
4680        }
4681    }
4682    mddev->layout = mddev->new_layout;
4683    mddev->chunk_sectors = 1 << conf->geo.chunk_shift;
4684    mddev->reshape_position = MaxSector;
4685    mddev->delta_disks = 0;
4686    mddev->reshape_backwards = 0;
4687}
4688
4689static struct md_personality raid10_personality =
4690{
4691    .name = "raid10",
4692    .level = 10,
4693    .owner = THIS_MODULE,
4694    .make_request = make_request,
4695    .run = run,
4696    .stop = stop,
4697    .status = status,
4698    .error_handler = error,
4699    .hot_add_disk = raid10_add_disk,
4700    .hot_remove_disk= raid10_remove_disk,
4701    .spare_active = raid10_spare_active,
4702    .sync_request = sync_request,
4703    .quiesce = raid10_quiesce,
4704    .size = raid10_size,
4705    .resize = raid10_resize,
4706    .takeover = raid10_takeover,
4707    .check_reshape = raid10_check_reshape,
4708    .start_reshape = raid10_start_reshape,
4709    .finish_reshape = raid10_finish_reshape,
4710};
4711
4712static int __init raid_init(void)
4713{
4714    return register_md_personality(&raid10_personality);
4715}
4716
4717static void raid_exit(void)
4718{
4719    unregister_md_personality(&raid10_personality);
4720}
4721
4722module_init(raid_init);
4723module_exit(raid_exit);
4724MODULE_LICENSE("GPL");
4725MODULE_DESCRIPTION("RAID10 (striped mirror) personality for MD");
4726MODULE_ALIAS("md-personality-9"); /* RAID10 */
4727MODULE_ALIAS("md-raid10");
4728MODULE_ALIAS("md-level-10");
4729
4730module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
4731

Archive Download this file



interactive