Root/drivers/md/raid1.c

1/*
2 * raid1.c : Multiple Devices driver for Linux
3 *
4 * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
5 *
6 * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
7 *
8 * RAID-1 management functions.
9 *
10 * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
11 *
12 * Fixes to reconstruction by Jakob Ƙstergaard" <jakob@ostenfeld.dk>
13 * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
14 *
15 * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
16 * bitmapped intelligence in resync:
17 *
18 * - bitmap marked during normal i/o
19 * - bitmap used to skip nondirty blocks during sync
20 *
21 * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
22 * - persistent bitmap code
23 *
24 * This program is free software; you can redistribute it and/or modify
25 * it under the terms of the GNU General Public License as published by
26 * the Free Software Foundation; either version 2, or (at your option)
27 * any later version.
28 *
29 * You should have received a copy of the GNU General Public License
30 * (for example /usr/src/linux/COPYING); if not, write to the Free
31 * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
32 */
33
34#include <linux/slab.h>
35#include <linux/delay.h>
36#include <linux/blkdev.h>
37#include <linux/module.h>
38#include <linux/seq_file.h>
39#include <linux/ratelimit.h>
40#include "md.h"
41#include "raid1.h"
42#include "bitmap.h"
43
44/*
45 * Number of guaranteed r1bios in case of extreme VM load:
46 */
47#define NR_RAID1_BIOS 256
48
49/* when we get a read error on a read-only array, we redirect to another
50 * device without failing the first device, or trying to over-write to
51 * correct the read error. To keep track of bad blocks on a per-bio
52 * level, we store IO_BLOCKED in the appropriate 'bios' pointer
53 */
54#define IO_BLOCKED ((struct bio *)1)
55/* When we successfully write to a known bad-block, we need to remove the
56 * bad-block marking which must be done from process context. So we record
57 * the success by setting devs[n].bio to IO_MADE_GOOD
58 */
59#define IO_MADE_GOOD ((struct bio *)2)
60
61#define BIO_SPECIAL(bio) ((unsigned long)bio <= 2)
62
63/* When there are this many requests queue to be written by
64 * the raid1 thread, we become 'congested' to provide back-pressure
65 * for writeback.
66 */
67static int max_queued_requests = 1024;
68
69static void allow_barrier(struct r1conf *conf);
70static void lower_barrier(struct r1conf *conf);
71
72static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
73{
74    struct pool_info *pi = data;
75    int size = offsetof(struct r1bio, bios[pi->raid_disks]);
76
77    /* allocate a r1bio with room for raid_disks entries in the bios array */
78    return kzalloc(size, gfp_flags);
79}
80
81static void r1bio_pool_free(void *r1_bio, void *data)
82{
83    kfree(r1_bio);
84}
85
86#define RESYNC_BLOCK_SIZE (64*1024)
87//#define RESYNC_BLOCK_SIZE PAGE_SIZE
88#define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
89#define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
90#define RESYNC_WINDOW (2048*1024)
91
92static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
93{
94    struct pool_info *pi = data;
95    struct page *page;
96    struct r1bio *r1_bio;
97    struct bio *bio;
98    int i, j;
99
100    r1_bio = r1bio_pool_alloc(gfp_flags, pi);
101    if (!r1_bio)
102        return NULL;
103
104    /*
105     * Allocate bios : 1 for reading, n-1 for writing
106     */
107    for (j = pi->raid_disks ; j-- ; ) {
108        bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
109        if (!bio)
110            goto out_free_bio;
111        r1_bio->bios[j] = bio;
112    }
113    /*
114     * Allocate RESYNC_PAGES data pages and attach them to
115     * the first bio.
116     * If this is a user-requested check/repair, allocate
117     * RESYNC_PAGES for each bio.
118     */
119    if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
120        j = pi->raid_disks;
121    else
122        j = 1;
123    while(j--) {
124        bio = r1_bio->bios[j];
125        for (i = 0; i < RESYNC_PAGES; i++) {
126            page = alloc_page(gfp_flags);
127            if (unlikely(!page))
128                goto out_free_pages;
129
130            bio->bi_io_vec[i].bv_page = page;
131            bio->bi_vcnt = i+1;
132        }
133    }
134    /* If not user-requests, copy the page pointers to all bios */
135    if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
136        for (i=0; i<RESYNC_PAGES ; i++)
137            for (j=1; j<pi->raid_disks; j++)
138                r1_bio->bios[j]->bi_io_vec[i].bv_page =
139                    r1_bio->bios[0]->bi_io_vec[i].bv_page;
140    }
141
142    r1_bio->master_bio = NULL;
143
144    return r1_bio;
145
146out_free_pages:
147    for (j=0 ; j < pi->raid_disks; j++)
148        for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
149            put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
150    j = -1;
151out_free_bio:
152    while (++j < pi->raid_disks)
153        bio_put(r1_bio->bios[j]);
154    r1bio_pool_free(r1_bio, data);
155    return NULL;
156}
157
158static void r1buf_pool_free(void *__r1_bio, void *data)
159{
160    struct pool_info *pi = data;
161    int i,j;
162    struct r1bio *r1bio = __r1_bio;
163
164    for (i = 0; i < RESYNC_PAGES; i++)
165        for (j = pi->raid_disks; j-- ;) {
166            if (j == 0 ||
167                r1bio->bios[j]->bi_io_vec[i].bv_page !=
168                r1bio->bios[0]->bi_io_vec[i].bv_page)
169                safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
170        }
171    for (i=0 ; i < pi->raid_disks; i++)
172        bio_put(r1bio->bios[i]);
173
174    r1bio_pool_free(r1bio, data);
175}
176
177static void put_all_bios(struct r1conf *conf, struct r1bio *r1_bio)
178{
179    int i;
180
181    for (i = 0; i < conf->raid_disks * 2; i++) {
182        struct bio **bio = r1_bio->bios + i;
183        if (!BIO_SPECIAL(*bio))
184            bio_put(*bio);
185        *bio = NULL;
186    }
187}
188
189static void free_r1bio(struct r1bio *r1_bio)
190{
191    struct r1conf *conf = r1_bio->mddev->private;
192
193    put_all_bios(conf, r1_bio);
194    mempool_free(r1_bio, conf->r1bio_pool);
195}
196
197static void put_buf(struct r1bio *r1_bio)
198{
199    struct r1conf *conf = r1_bio->mddev->private;
200    int i;
201
202    for (i = 0; i < conf->raid_disks * 2; i++) {
203        struct bio *bio = r1_bio->bios[i];
204        if (bio->bi_end_io)
205            rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
206    }
207
208    mempool_free(r1_bio, conf->r1buf_pool);
209
210    lower_barrier(conf);
211}
212
213static void reschedule_retry(struct r1bio *r1_bio)
214{
215    unsigned long flags;
216    struct mddev *mddev = r1_bio->mddev;
217    struct r1conf *conf = mddev->private;
218
219    spin_lock_irqsave(&conf->device_lock, flags);
220    list_add(&r1_bio->retry_list, &conf->retry_list);
221    conf->nr_queued ++;
222    spin_unlock_irqrestore(&conf->device_lock, flags);
223
224    wake_up(&conf->wait_barrier);
225    md_wakeup_thread(mddev->thread);
226}
227
228/*
229 * raid_end_bio_io() is called when we have finished servicing a mirrored
230 * operation and are ready to return a success/failure code to the buffer
231 * cache layer.
232 */
233static void call_bio_endio(struct r1bio *r1_bio)
234{
235    struct bio *bio = r1_bio->master_bio;
236    int done;
237    struct r1conf *conf = r1_bio->mddev->private;
238
239    if (bio->bi_phys_segments) {
240        unsigned long flags;
241        spin_lock_irqsave(&conf->device_lock, flags);
242        bio->bi_phys_segments--;
243        done = (bio->bi_phys_segments == 0);
244        spin_unlock_irqrestore(&conf->device_lock, flags);
245    } else
246        done = 1;
247
248    if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
249        clear_bit(BIO_UPTODATE, &bio->bi_flags);
250    if (done) {
251        bio_endio(bio, 0);
252        /*
253         * Wake up any possible resync thread that waits for the device
254         * to go idle.
255         */
256        allow_barrier(conf);
257    }
258}
259
260static void raid_end_bio_io(struct r1bio *r1_bio)
261{
262    struct bio *bio = r1_bio->master_bio;
263
264    /* if nobody has done the final endio yet, do it now */
265    if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
266        pr_debug("raid1: sync end %s on sectors %llu-%llu\n",
267             (bio_data_dir(bio) == WRITE) ? "write" : "read",
268             (unsigned long long) bio->bi_sector,
269             (unsigned long long) bio->bi_sector +
270             (bio->bi_size >> 9) - 1);
271
272        call_bio_endio(r1_bio);
273    }
274    free_r1bio(r1_bio);
275}
276
277/*
278 * Update disk head position estimator based on IRQ completion info.
279 */
280static inline void update_head_pos(int disk, struct r1bio *r1_bio)
281{
282    struct r1conf *conf = r1_bio->mddev->private;
283
284    conf->mirrors[disk].head_position =
285        r1_bio->sector + (r1_bio->sectors);
286}
287
288/*
289 * Find the disk number which triggered given bio
290 */
291static int find_bio_disk(struct r1bio *r1_bio, struct bio *bio)
292{
293    int mirror;
294    struct r1conf *conf = r1_bio->mddev->private;
295    int raid_disks = conf->raid_disks;
296
297    for (mirror = 0; mirror < raid_disks * 2; mirror++)
298        if (r1_bio->bios[mirror] == bio)
299            break;
300
301    BUG_ON(mirror == raid_disks * 2);
302    update_head_pos(mirror, r1_bio);
303
304    return mirror;
305}
306
307static void raid1_end_read_request(struct bio *bio, int error)
308{
309    int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
310    struct r1bio *r1_bio = bio->bi_private;
311    int mirror;
312    struct r1conf *conf = r1_bio->mddev->private;
313
314    mirror = r1_bio->read_disk;
315    /*
316     * this branch is our 'one mirror IO has finished' event handler:
317     */
318    update_head_pos(mirror, r1_bio);
319
320    if (uptodate)
321        set_bit(R1BIO_Uptodate, &r1_bio->state);
322    else {
323        /* If all other devices have failed, we want to return
324         * the error upwards rather than fail the last device.
325         * Here we redefine "uptodate" to mean "Don't want to retry"
326         */
327        unsigned long flags;
328        spin_lock_irqsave(&conf->device_lock, flags);
329        if (r1_bio->mddev->degraded == conf->raid_disks ||
330            (r1_bio->mddev->degraded == conf->raid_disks-1 &&
331             !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
332            uptodate = 1;
333        spin_unlock_irqrestore(&conf->device_lock, flags);
334    }
335
336    if (uptodate) {
337        raid_end_bio_io(r1_bio);
338        rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
339    } else {
340        /*
341         * oops, read error:
342         */
343        char b[BDEVNAME_SIZE];
344        printk_ratelimited(
345            KERN_ERR "md/raid1:%s: %s: "
346            "rescheduling sector %llu\n",
347            mdname(conf->mddev),
348            bdevname(conf->mirrors[mirror].rdev->bdev,
349                 b),
350            (unsigned long long)r1_bio->sector);
351        set_bit(R1BIO_ReadError, &r1_bio->state);
352        reschedule_retry(r1_bio);
353        /* don't drop the reference on read_disk yet */
354    }
355}
356
357static void close_write(struct r1bio *r1_bio)
358{
359    /* it really is the end of this request */
360    if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
361        /* free extra copy of the data pages */
362        int i = r1_bio->behind_page_count;
363        while (i--)
364            safe_put_page(r1_bio->behind_bvecs[i].bv_page);
365        kfree(r1_bio->behind_bvecs);
366        r1_bio->behind_bvecs = NULL;
367    }
368    /* clear the bitmap if all writes complete successfully */
369    bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
370            r1_bio->sectors,
371            !test_bit(R1BIO_Degraded, &r1_bio->state),
372            test_bit(R1BIO_BehindIO, &r1_bio->state));
373    md_write_end(r1_bio->mddev);
374}
375
376static void r1_bio_write_done(struct r1bio *r1_bio)
377{
378    if (!atomic_dec_and_test(&r1_bio->remaining))
379        return;
380
381    if (test_bit(R1BIO_WriteError, &r1_bio->state))
382        reschedule_retry(r1_bio);
383    else {
384        close_write(r1_bio);
385        if (test_bit(R1BIO_MadeGood, &r1_bio->state))
386            reschedule_retry(r1_bio);
387        else
388            raid_end_bio_io(r1_bio);
389    }
390}
391
392static void raid1_end_write_request(struct bio *bio, int error)
393{
394    int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
395    struct r1bio *r1_bio = bio->bi_private;
396    int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
397    struct r1conf *conf = r1_bio->mddev->private;
398    struct bio *to_put = NULL;
399
400    mirror = find_bio_disk(r1_bio, bio);
401
402    /*
403     * 'one mirror IO has finished' event handler:
404     */
405    if (!uptodate) {
406        set_bit(WriteErrorSeen,
407            &conf->mirrors[mirror].rdev->flags);
408        if (!test_and_set_bit(WantReplacement,
409                      &conf->mirrors[mirror].rdev->flags))
410            set_bit(MD_RECOVERY_NEEDED, &
411                conf->mddev->recovery);
412
413        set_bit(R1BIO_WriteError, &r1_bio->state);
414    } else {
415        /*
416         * Set R1BIO_Uptodate in our master bio, so that we
417         * will return a good error code for to the higher
418         * levels even if IO on some other mirrored buffer
419         * fails.
420         *
421         * The 'master' represents the composite IO operation
422         * to user-side. So if something waits for IO, then it
423         * will wait for the 'master' bio.
424         */
425        sector_t first_bad;
426        int bad_sectors;
427
428        r1_bio->bios[mirror] = NULL;
429        to_put = bio;
430        set_bit(R1BIO_Uptodate, &r1_bio->state);
431
432        /* Maybe we can clear some bad blocks. */
433        if (is_badblock(conf->mirrors[mirror].rdev,
434                r1_bio->sector, r1_bio->sectors,
435                &first_bad, &bad_sectors)) {
436            r1_bio->bios[mirror] = IO_MADE_GOOD;
437            set_bit(R1BIO_MadeGood, &r1_bio->state);
438        }
439    }
440
441    if (behind) {
442        if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
443            atomic_dec(&r1_bio->behind_remaining);
444
445        /*
446         * In behind mode, we ACK the master bio once the I/O
447         * has safely reached all non-writemostly
448         * disks. Setting the Returned bit ensures that this
449         * gets done only once -- we don't ever want to return
450         * -EIO here, instead we'll wait
451         */
452        if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
453            test_bit(R1BIO_Uptodate, &r1_bio->state)) {
454            /* Maybe we can return now */
455            if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
456                struct bio *mbio = r1_bio->master_bio;
457                pr_debug("raid1: behind end write sectors"
458                     " %llu-%llu\n",
459                     (unsigned long long) mbio->bi_sector,
460                     (unsigned long long) mbio->bi_sector +
461                     (mbio->bi_size >> 9) - 1);
462                call_bio_endio(r1_bio);
463            }
464        }
465    }
466    if (r1_bio->bios[mirror] == NULL)
467        rdev_dec_pending(conf->mirrors[mirror].rdev,
468                 conf->mddev);
469
470    /*
471     * Let's see if all mirrored write operations have finished
472     * already.
473     */
474    r1_bio_write_done(r1_bio);
475
476    if (to_put)
477        bio_put(to_put);
478}
479
480
481/*
482 * This routine returns the disk from which the requested read should
483 * be done. There is a per-array 'next expected sequential IO' sector
484 * number - if this matches on the next IO then we use the last disk.
485 * There is also a per-disk 'last know head position' sector that is
486 * maintained from IRQ contexts, both the normal and the resync IO
487 * completion handlers update this position correctly. If there is no
488 * perfect sequential match then we pick the disk whose head is closest.
489 *
490 * If there are 2 mirrors in the same 2 devices, performance degrades
491 * because position is mirror, not device based.
492 *
493 * The rdev for the device selected will have nr_pending incremented.
494 */
495static int read_balance(struct r1conf *conf, struct r1bio *r1_bio, int *max_sectors)
496{
497    const sector_t this_sector = r1_bio->sector;
498    int sectors;
499    int best_good_sectors;
500    int best_disk, best_dist_disk, best_pending_disk;
501    int has_nonrot_disk;
502    int disk;
503    sector_t best_dist;
504    unsigned int min_pending;
505    struct md_rdev *rdev;
506    int choose_first;
507    int choose_next_idle;
508
509    rcu_read_lock();
510    /*
511     * Check if we can balance. We can balance on the whole
512     * device if no resync is going on, or below the resync window.
513     * We take the first readable disk when above the resync window.
514     */
515 retry:
516    sectors = r1_bio->sectors;
517    best_disk = -1;
518    best_dist_disk = -1;
519    best_dist = MaxSector;
520    best_pending_disk = -1;
521    min_pending = UINT_MAX;
522    best_good_sectors = 0;
523    has_nonrot_disk = 0;
524    choose_next_idle = 0;
525
526    if (conf->mddev->recovery_cp < MaxSector &&
527        (this_sector + sectors >= conf->next_resync))
528        choose_first = 1;
529    else
530        choose_first = 0;
531
532    for (disk = 0 ; disk < conf->raid_disks * 2 ; disk++) {
533        sector_t dist;
534        sector_t first_bad;
535        int bad_sectors;
536        unsigned int pending;
537        bool nonrot;
538
539        rdev = rcu_dereference(conf->mirrors[disk].rdev);
540        if (r1_bio->bios[disk] == IO_BLOCKED
541            || rdev == NULL
542            || test_bit(Unmerged, &rdev->flags)
543            || test_bit(Faulty, &rdev->flags))
544            continue;
545        if (!test_bit(In_sync, &rdev->flags) &&
546            rdev->recovery_offset < this_sector + sectors)
547            continue;
548        if (test_bit(WriteMostly, &rdev->flags)) {
549            /* Don't balance among write-mostly, just
550             * use the first as a last resort */
551            if (best_disk < 0) {
552                if (is_badblock(rdev, this_sector, sectors,
553                        &first_bad, &bad_sectors)) {
554                    if (first_bad < this_sector)
555                        /* Cannot use this */
556                        continue;
557                    best_good_sectors = first_bad - this_sector;
558                } else
559                    best_good_sectors = sectors;
560                best_disk = disk;
561            }
562            continue;
563        }
564        /* This is a reasonable device to use. It might
565         * even be best.
566         */
567        if (is_badblock(rdev, this_sector, sectors,
568                &first_bad, &bad_sectors)) {
569            if (best_dist < MaxSector)
570                /* already have a better device */
571                continue;
572            if (first_bad <= this_sector) {
573                /* cannot read here. If this is the 'primary'
574                 * device, then we must not read beyond
575                 * bad_sectors from another device..
576                 */
577                bad_sectors -= (this_sector - first_bad);
578                if (choose_first && sectors > bad_sectors)
579                    sectors = bad_sectors;
580                if (best_good_sectors > sectors)
581                    best_good_sectors = sectors;
582
583            } else {
584                sector_t good_sectors = first_bad - this_sector;
585                if (good_sectors > best_good_sectors) {
586                    best_good_sectors = good_sectors;
587                    best_disk = disk;
588                }
589                if (choose_first)
590                    break;
591            }
592            continue;
593        } else
594            best_good_sectors = sectors;
595
596        nonrot = blk_queue_nonrot(bdev_get_queue(rdev->bdev));
597        has_nonrot_disk |= nonrot;
598        pending = atomic_read(&rdev->nr_pending);
599        dist = abs(this_sector - conf->mirrors[disk].head_position);
600        if (choose_first) {
601            best_disk = disk;
602            break;
603        }
604        /* Don't change to another disk for sequential reads */
605        if (conf->mirrors[disk].next_seq_sect == this_sector
606            || dist == 0) {
607            int opt_iosize = bdev_io_opt(rdev->bdev) >> 9;
608            struct raid1_info *mirror = &conf->mirrors[disk];
609
610            best_disk = disk;
611            /*
612             * If buffered sequential IO size exceeds optimal
613             * iosize, check if there is idle disk. If yes, choose
614             * the idle disk. read_balance could already choose an
615             * idle disk before noticing it's a sequential IO in
616             * this disk. This doesn't matter because this disk
617             * will idle, next time it will be utilized after the
618             * first disk has IO size exceeds optimal iosize. In
619             * this way, iosize of the first disk will be optimal
620             * iosize at least. iosize of the second disk might be
621             * small, but not a big deal since when the second disk
622             * starts IO, the first disk is likely still busy.
623             */
624            if (nonrot && opt_iosize > 0 &&
625                mirror->seq_start != MaxSector &&
626                mirror->next_seq_sect > opt_iosize &&
627                mirror->next_seq_sect - opt_iosize >=
628                mirror->seq_start) {
629                choose_next_idle = 1;
630                continue;
631            }
632            break;
633        }
634        /* If device is idle, use it */
635        if (pending == 0) {
636            best_disk = disk;
637            break;
638        }
639
640        if (choose_next_idle)
641            continue;
642
643        if (min_pending > pending) {
644            min_pending = pending;
645            best_pending_disk = disk;
646        }
647
648        if (dist < best_dist) {
649            best_dist = dist;
650            best_dist_disk = disk;
651        }
652    }
653
654    /*
655     * If all disks are rotational, choose the closest disk. If any disk is
656     * non-rotational, choose the disk with less pending request even the
657     * disk is rotational, which might/might not be optimal for raids with
658     * mixed ratation/non-rotational disks depending on workload.
659     */
660    if (best_disk == -1) {
661        if (has_nonrot_disk)
662            best_disk = best_pending_disk;
663        else
664            best_disk = best_dist_disk;
665    }
666
667    if (best_disk >= 0) {
668        rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
669        if (!rdev)
670            goto retry;
671        atomic_inc(&rdev->nr_pending);
672        if (test_bit(Faulty, &rdev->flags)) {
673            /* cannot risk returning a device that failed
674             * before we inc'ed nr_pending
675             */
676            rdev_dec_pending(rdev, conf->mddev);
677            goto retry;
678        }
679        sectors = best_good_sectors;
680
681        if (conf->mirrors[best_disk].next_seq_sect != this_sector)
682            conf->mirrors[best_disk].seq_start = this_sector;
683
684        conf->mirrors[best_disk].next_seq_sect = this_sector + sectors;
685    }
686    rcu_read_unlock();
687    *max_sectors = sectors;
688
689    return best_disk;
690}
691
692static int raid1_mergeable_bvec(struct request_queue *q,
693                struct bvec_merge_data *bvm,
694                struct bio_vec *biovec)
695{
696    struct mddev *mddev = q->queuedata;
697    struct r1conf *conf = mddev->private;
698    sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
699    int max = biovec->bv_len;
700
701    if (mddev->merge_check_needed) {
702        int disk;
703        rcu_read_lock();
704        for (disk = 0; disk < conf->raid_disks * 2; disk++) {
705            struct md_rdev *rdev = rcu_dereference(
706                conf->mirrors[disk].rdev);
707            if (rdev && !test_bit(Faulty, &rdev->flags)) {
708                struct request_queue *q =
709                    bdev_get_queue(rdev->bdev);
710                if (q->merge_bvec_fn) {
711                    bvm->bi_sector = sector +
712                        rdev->data_offset;
713                    bvm->bi_bdev = rdev->bdev;
714                    max = min(max, q->merge_bvec_fn(
715                              q, bvm, biovec));
716                }
717            }
718        }
719        rcu_read_unlock();
720    }
721    return max;
722
723}
724
725int md_raid1_congested(struct mddev *mddev, int bits)
726{
727    struct r1conf *conf = mddev->private;
728    int i, ret = 0;
729
730    if ((bits & (1 << BDI_async_congested)) &&
731        conf->pending_count >= max_queued_requests)
732        return 1;
733
734    rcu_read_lock();
735    for (i = 0; i < conf->raid_disks * 2; i++) {
736        struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
737        if (rdev && !test_bit(Faulty, &rdev->flags)) {
738            struct request_queue *q = bdev_get_queue(rdev->bdev);
739
740            BUG_ON(!q);
741
742            /* Note the '|| 1' - when read_balance prefers
743             * non-congested targets, it can be removed
744             */
745            if ((bits & (1<<BDI_async_congested)) || 1)
746                ret |= bdi_congested(&q->backing_dev_info, bits);
747            else
748                ret &= bdi_congested(&q->backing_dev_info, bits);
749        }
750    }
751    rcu_read_unlock();
752    return ret;
753}
754EXPORT_SYMBOL_GPL(md_raid1_congested);
755
756static int raid1_congested(void *data, int bits)
757{
758    struct mddev *mddev = data;
759
760    return mddev_congested(mddev, bits) ||
761        md_raid1_congested(mddev, bits);
762}
763
764static void flush_pending_writes(struct r1conf *conf)
765{
766    /* Any writes that have been queued but are awaiting
767     * bitmap updates get flushed here.
768     */
769    spin_lock_irq(&conf->device_lock);
770
771    if (conf->pending_bio_list.head) {
772        struct bio *bio;
773        bio = bio_list_get(&conf->pending_bio_list);
774        conf->pending_count = 0;
775        spin_unlock_irq(&conf->device_lock);
776        /* flush any pending bitmap writes to
777         * disk before proceeding w/ I/O */
778        bitmap_unplug(conf->mddev->bitmap);
779        wake_up(&conf->wait_barrier);
780
781        while (bio) { /* submit pending writes */
782            struct bio *next = bio->bi_next;
783            bio->bi_next = NULL;
784            if (unlikely((bio->bi_rw & REQ_DISCARD) &&
785                !blk_queue_discard(bdev_get_queue(bio->bi_bdev))))
786                /* Just ignore it */
787                bio_endio(bio, 0);
788            else
789                generic_make_request(bio);
790            bio = next;
791        }
792    } else
793        spin_unlock_irq(&conf->device_lock);
794}
795
796/* Barriers....
797 * Sometimes we need to suspend IO while we do something else,
798 * either some resync/recovery, or reconfigure the array.
799 * To do this we raise a 'barrier'.
800 * The 'barrier' is a counter that can be raised multiple times
801 * to count how many activities are happening which preclude
802 * normal IO.
803 * We can only raise the barrier if there is no pending IO.
804 * i.e. if nr_pending == 0.
805 * We choose only to raise the barrier if no-one is waiting for the
806 * barrier to go down. This means that as soon as an IO request
807 * is ready, no other operations which require a barrier will start
808 * until the IO request has had a chance.
809 *
810 * So: regular IO calls 'wait_barrier'. When that returns there
811 * is no backgroup IO happening, It must arrange to call
812 * allow_barrier when it has finished its IO.
813 * backgroup IO calls must call raise_barrier. Once that returns
814 * there is no normal IO happeing. It must arrange to call
815 * lower_barrier when the particular background IO completes.
816 */
817#define RESYNC_DEPTH 32
818
819static void raise_barrier(struct r1conf *conf)
820{
821    spin_lock_irq(&conf->resync_lock);
822
823    /* Wait until no block IO is waiting */
824    wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
825                conf->resync_lock);
826
827    /* block any new IO from starting */
828    conf->barrier++;
829
830    /* Now wait for all pending IO to complete */
831    wait_event_lock_irq(conf->wait_barrier,
832                !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
833                conf->resync_lock);
834
835    spin_unlock_irq(&conf->resync_lock);
836}
837
838static void lower_barrier(struct r1conf *conf)
839{
840    unsigned long flags;
841    BUG_ON(conf->barrier <= 0);
842    spin_lock_irqsave(&conf->resync_lock, flags);
843    conf->barrier--;
844    spin_unlock_irqrestore(&conf->resync_lock, flags);
845    wake_up(&conf->wait_barrier);
846}
847
848static void wait_barrier(struct r1conf *conf)
849{
850    spin_lock_irq(&conf->resync_lock);
851    if (conf->barrier) {
852        conf->nr_waiting++;
853        /* Wait for the barrier to drop.
854         * However if there are already pending
855         * requests (preventing the barrier from
856         * rising completely), and the
857         * pre-process bio queue isn't empty,
858         * then don't wait, as we need to empty
859         * that queue to get the nr_pending
860         * count down.
861         */
862        wait_event_lock_irq(conf->wait_barrier,
863                    !conf->barrier ||
864                    (conf->nr_pending &&
865                     current->bio_list &&
866                     !bio_list_empty(current->bio_list)),
867                    conf->resync_lock);
868        conf->nr_waiting--;
869    }
870    conf->nr_pending++;
871    spin_unlock_irq(&conf->resync_lock);
872}
873
874static void allow_barrier(struct r1conf *conf)
875{
876    unsigned long flags;
877    spin_lock_irqsave(&conf->resync_lock, flags);
878    conf->nr_pending--;
879    spin_unlock_irqrestore(&conf->resync_lock, flags);
880    wake_up(&conf->wait_barrier);
881}
882
883static void freeze_array(struct r1conf *conf)
884{
885    /* stop syncio and normal IO and wait for everything to
886     * go quite.
887     * We increment barrier and nr_waiting, and then
888     * wait until nr_pending match nr_queued+1
889     * This is called in the context of one normal IO request
890     * that has failed. Thus any sync request that might be pending
891     * will be blocked by nr_pending, and we need to wait for
892     * pending IO requests to complete or be queued for re-try.
893     * Thus the number queued (nr_queued) plus this request (1)
894     * must match the number of pending IOs (nr_pending) before
895     * we continue.
896     */
897    spin_lock_irq(&conf->resync_lock);
898    conf->barrier++;
899    conf->nr_waiting++;
900    wait_event_lock_irq_cmd(conf->wait_barrier,
901                conf->nr_pending == conf->nr_queued+1,
902                conf->resync_lock,
903                flush_pending_writes(conf));
904    spin_unlock_irq(&conf->resync_lock);
905}
906static void unfreeze_array(struct r1conf *conf)
907{
908    /* reverse the effect of the freeze */
909    spin_lock_irq(&conf->resync_lock);
910    conf->barrier--;
911    conf->nr_waiting--;
912    wake_up(&conf->wait_barrier);
913    spin_unlock_irq(&conf->resync_lock);
914}
915
916
917/* duplicate the data pages for behind I/O
918 */
919static void alloc_behind_pages(struct bio *bio, struct r1bio *r1_bio)
920{
921    int i;
922    struct bio_vec *bvec;
923    struct bio_vec *bvecs = kzalloc(bio->bi_vcnt * sizeof(struct bio_vec),
924                    GFP_NOIO);
925    if (unlikely(!bvecs))
926        return;
927
928    bio_for_each_segment(bvec, bio, i) {
929        bvecs[i] = *bvec;
930        bvecs[i].bv_page = alloc_page(GFP_NOIO);
931        if (unlikely(!bvecs[i].bv_page))
932            goto do_sync_io;
933        memcpy(kmap(bvecs[i].bv_page) + bvec->bv_offset,
934               kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
935        kunmap(bvecs[i].bv_page);
936        kunmap(bvec->bv_page);
937    }
938    r1_bio->behind_bvecs = bvecs;
939    r1_bio->behind_page_count = bio->bi_vcnt;
940    set_bit(R1BIO_BehindIO, &r1_bio->state);
941    return;
942
943do_sync_io:
944    for (i = 0; i < bio->bi_vcnt; i++)
945        if (bvecs[i].bv_page)
946            put_page(bvecs[i].bv_page);
947    kfree(bvecs);
948    pr_debug("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
949}
950
951struct raid1_plug_cb {
952    struct blk_plug_cb cb;
953    struct bio_list pending;
954    int pending_cnt;
955};
956
957static void raid1_unplug(struct blk_plug_cb *cb, bool from_schedule)
958{
959    struct raid1_plug_cb *plug = container_of(cb, struct raid1_plug_cb,
960                          cb);
961    struct mddev *mddev = plug->cb.data;
962    struct r1conf *conf = mddev->private;
963    struct bio *bio;
964
965    if (from_schedule || current->bio_list) {
966        spin_lock_irq(&conf->device_lock);
967        bio_list_merge(&conf->pending_bio_list, &plug->pending);
968        conf->pending_count += plug->pending_cnt;
969        spin_unlock_irq(&conf->device_lock);
970        wake_up(&conf->wait_barrier);
971        md_wakeup_thread(mddev->thread);
972        kfree(plug);
973        return;
974    }
975
976    /* we aren't scheduling, so we can do the write-out directly. */
977    bio = bio_list_get(&plug->pending);
978    bitmap_unplug(mddev->bitmap);
979    wake_up(&conf->wait_barrier);
980
981    while (bio) { /* submit pending writes */
982        struct bio *next = bio->bi_next;
983        bio->bi_next = NULL;
984        generic_make_request(bio);
985        bio = next;
986    }
987    kfree(plug);
988}
989
990static void make_request(struct mddev *mddev, struct bio * bio)
991{
992    struct r1conf *conf = mddev->private;
993    struct raid1_info *mirror;
994    struct r1bio *r1_bio;
995    struct bio *read_bio;
996    int i, disks;
997    struct bitmap *bitmap;
998    unsigned long flags;
999    const int rw = bio_data_dir(bio);
1000    const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
1001    const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
1002    const unsigned long do_discard = (bio->bi_rw
1003                      & (REQ_DISCARD | REQ_SECURE));
1004    const unsigned long do_same = (bio->bi_rw & REQ_WRITE_SAME);
1005    struct md_rdev *blocked_rdev;
1006    struct blk_plug_cb *cb;
1007    struct raid1_plug_cb *plug = NULL;
1008    int first_clone;
1009    int sectors_handled;
1010    int max_sectors;
1011
1012    /*
1013     * Register the new request and wait if the reconstruction
1014     * thread has put up a bar for new requests.
1015     * Continue immediately if no resync is active currently.
1016     */
1017
1018    md_write_start(mddev, bio); /* wait on superblock update early */
1019
1020    if (bio_data_dir(bio) == WRITE &&
1021        bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
1022        bio->bi_sector < mddev->suspend_hi) {
1023        /* As the suspend_* range is controlled by
1024         * userspace, we want an interruptible
1025         * wait.
1026         */
1027        DEFINE_WAIT(w);
1028        for (;;) {
1029            flush_signals(current);
1030            prepare_to_wait(&conf->wait_barrier,
1031                    &w, TASK_INTERRUPTIBLE);
1032            if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
1033                bio->bi_sector >= mddev->suspend_hi)
1034                break;
1035            schedule();
1036        }
1037        finish_wait(&conf->wait_barrier, &w);
1038    }
1039
1040    wait_barrier(conf);
1041
1042    bitmap = mddev->bitmap;
1043
1044    /*
1045     * make_request() can abort the operation when READA is being
1046     * used and no empty request is available.
1047     *
1048     */
1049    r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1050
1051    r1_bio->master_bio = bio;
1052    r1_bio->sectors = bio->bi_size >> 9;
1053    r1_bio->state = 0;
1054    r1_bio->mddev = mddev;
1055    r1_bio->sector = bio->bi_sector;
1056
1057    /* We might need to issue multiple reads to different
1058     * devices if there are bad blocks around, so we keep
1059     * track of the number of reads in bio->bi_phys_segments.
1060     * If this is 0, there is only one r1_bio and no locking
1061     * will be needed when requests complete. If it is
1062     * non-zero, then it is the number of not-completed requests.
1063     */
1064    bio->bi_phys_segments = 0;
1065    clear_bit(BIO_SEG_VALID, &bio->bi_flags);
1066
1067    if (rw == READ) {
1068        /*
1069         * read balancing logic:
1070         */
1071        int rdisk;
1072
1073read_again:
1074        rdisk = read_balance(conf, r1_bio, &max_sectors);
1075
1076        if (rdisk < 0) {
1077            /* couldn't find anywhere to read from */
1078            raid_end_bio_io(r1_bio);
1079            return;
1080        }
1081        mirror = conf->mirrors + rdisk;
1082
1083        if (test_bit(WriteMostly, &mirror->rdev->flags) &&
1084            bitmap) {
1085            /* Reading from a write-mostly device must
1086             * take care not to over-take any writes
1087             * that are 'behind'
1088             */
1089            wait_event(bitmap->behind_wait,
1090                   atomic_read(&bitmap->behind_writes) == 0);
1091        }
1092        r1_bio->read_disk = rdisk;
1093
1094        read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1095        md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
1096                max_sectors);
1097
1098        r1_bio->bios[rdisk] = read_bio;
1099
1100        read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
1101        read_bio->bi_bdev = mirror->rdev->bdev;
1102        read_bio->bi_end_io = raid1_end_read_request;
1103        read_bio->bi_rw = READ | do_sync;
1104        read_bio->bi_private = r1_bio;
1105
1106        if (max_sectors < r1_bio->sectors) {
1107            /* could not read all from this device, so we will
1108             * need another r1_bio.
1109             */
1110
1111            sectors_handled = (r1_bio->sector + max_sectors
1112                       - bio->bi_sector);
1113            r1_bio->sectors = max_sectors;
1114            spin_lock_irq(&conf->device_lock);
1115            if (bio->bi_phys_segments == 0)
1116                bio->bi_phys_segments = 2;
1117            else
1118                bio->bi_phys_segments++;
1119            spin_unlock_irq(&conf->device_lock);
1120            /* Cannot call generic_make_request directly
1121             * as that will be queued in __make_request
1122             * and subsequent mempool_alloc might block waiting
1123             * for it. So hand bio over to raid1d.
1124             */
1125            reschedule_retry(r1_bio);
1126
1127            r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1128
1129            r1_bio->master_bio = bio;
1130            r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1131            r1_bio->state = 0;
1132            r1_bio->mddev = mddev;
1133            r1_bio->sector = bio->bi_sector + sectors_handled;
1134            goto read_again;
1135        } else
1136            generic_make_request(read_bio);
1137        return;
1138    }
1139
1140    /*
1141     * WRITE:
1142     */
1143    if (conf->pending_count >= max_queued_requests) {
1144        md_wakeup_thread(mddev->thread);
1145        wait_event(conf->wait_barrier,
1146               conf->pending_count < max_queued_requests);
1147    }
1148    /* first select target devices under rcu_lock and
1149     * inc refcount on their rdev. Record them by setting
1150     * bios[x] to bio
1151     * If there are known/acknowledged bad blocks on any device on
1152     * which we have seen a write error, we want to avoid writing those
1153     * blocks.
1154     * This potentially requires several writes to write around
1155     * the bad blocks. Each set of writes gets it's own r1bio
1156     * with a set of bios attached.
1157     */
1158
1159    disks = conf->raid_disks * 2;
1160 retry_write:
1161    blocked_rdev = NULL;
1162    rcu_read_lock();
1163    max_sectors = r1_bio->sectors;
1164    for (i = 0; i < disks; i++) {
1165        struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1166        if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
1167            atomic_inc(&rdev->nr_pending);
1168            blocked_rdev = rdev;
1169            break;
1170        }
1171        r1_bio->bios[i] = NULL;
1172        if (!rdev || test_bit(Faulty, &rdev->flags)
1173            || test_bit(Unmerged, &rdev->flags)) {
1174            if (i < conf->raid_disks)
1175                set_bit(R1BIO_Degraded, &r1_bio->state);
1176            continue;
1177        }
1178
1179        atomic_inc(&rdev->nr_pending);
1180        if (test_bit(WriteErrorSeen, &rdev->flags)) {
1181            sector_t first_bad;
1182            int bad_sectors;
1183            int is_bad;
1184
1185            is_bad = is_badblock(rdev, r1_bio->sector,
1186                         max_sectors,
1187                         &first_bad, &bad_sectors);
1188            if (is_bad < 0) {
1189                /* mustn't write here until the bad block is
1190                 * acknowledged*/
1191                set_bit(BlockedBadBlocks, &rdev->flags);
1192                blocked_rdev = rdev;
1193                break;
1194            }
1195            if (is_bad && first_bad <= r1_bio->sector) {
1196                /* Cannot write here at all */
1197                bad_sectors -= (r1_bio->sector - first_bad);
1198                if (bad_sectors < max_sectors)
1199                    /* mustn't write more than bad_sectors
1200                     * to other devices yet
1201                     */
1202                    max_sectors = bad_sectors;
1203                rdev_dec_pending(rdev, mddev);
1204                /* We don't set R1BIO_Degraded as that
1205                 * only applies if the disk is
1206                 * missing, so it might be re-added,
1207                 * and we want to know to recover this
1208                 * chunk.
1209                 * In this case the device is here,
1210                 * and the fact that this chunk is not
1211                 * in-sync is recorded in the bad
1212                 * block log
1213                 */
1214                continue;
1215            }
1216            if (is_bad) {
1217                int good_sectors = first_bad - r1_bio->sector;
1218                if (good_sectors < max_sectors)
1219                    max_sectors = good_sectors;
1220            }
1221        }
1222        r1_bio->bios[i] = bio;
1223    }
1224    rcu_read_unlock();
1225
1226    if (unlikely(blocked_rdev)) {
1227        /* Wait for this device to become unblocked */
1228        int j;
1229
1230        for (j = 0; j < i; j++)
1231            if (r1_bio->bios[j])
1232                rdev_dec_pending(conf->mirrors[j].rdev, mddev);
1233        r1_bio->state = 0;
1234        allow_barrier(conf);
1235        md_wait_for_blocked_rdev(blocked_rdev, mddev);
1236        wait_barrier(conf);
1237        goto retry_write;
1238    }
1239
1240    if (max_sectors < r1_bio->sectors) {
1241        /* We are splitting this write into multiple parts, so
1242         * we need to prepare for allocating another r1_bio.
1243         */
1244        r1_bio->sectors = max_sectors;
1245        spin_lock_irq(&conf->device_lock);
1246        if (bio->bi_phys_segments == 0)
1247            bio->bi_phys_segments = 2;
1248        else
1249            bio->bi_phys_segments++;
1250        spin_unlock_irq(&conf->device_lock);
1251    }
1252    sectors_handled = r1_bio->sector + max_sectors - bio->bi_sector;
1253
1254    atomic_set(&r1_bio->remaining, 1);
1255    atomic_set(&r1_bio->behind_remaining, 0);
1256
1257    first_clone = 1;
1258    for (i = 0; i < disks; i++) {
1259        struct bio *mbio;
1260        if (!r1_bio->bios[i])
1261            continue;
1262
1263        mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
1264        md_trim_bio(mbio, r1_bio->sector - bio->bi_sector, max_sectors);
1265
1266        if (first_clone) {
1267            /* do behind I/O ?
1268             * Not if there are too many, or cannot
1269             * allocate memory, or a reader on WriteMostly
1270             * is waiting for behind writes to flush */
1271            if (bitmap &&
1272                (atomic_read(&bitmap->behind_writes)
1273                 < mddev->bitmap_info.max_write_behind) &&
1274                !waitqueue_active(&bitmap->behind_wait))
1275                alloc_behind_pages(mbio, r1_bio);
1276
1277            bitmap_startwrite(bitmap, r1_bio->sector,
1278                      r1_bio->sectors,
1279                      test_bit(R1BIO_BehindIO,
1280                           &r1_bio->state));
1281            first_clone = 0;
1282        }
1283        if (r1_bio->behind_bvecs) {
1284            struct bio_vec *bvec;
1285            int j;
1286
1287            /* Yes, I really want the '__' version so that
1288             * we clear any unused pointer in the io_vec, rather
1289             * than leave them unchanged. This is important
1290             * because when we come to free the pages, we won't
1291             * know the original bi_idx, so we just free
1292             * them all
1293             */
1294            __bio_for_each_segment(bvec, mbio, j, 0)
1295                bvec->bv_page = r1_bio->behind_bvecs[j].bv_page;
1296            if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1297                atomic_inc(&r1_bio->behind_remaining);
1298        }
1299
1300        r1_bio->bios[i] = mbio;
1301
1302        mbio->bi_sector = (r1_bio->sector +
1303                   conf->mirrors[i].rdev->data_offset);
1304        mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1305        mbio->bi_end_io = raid1_end_write_request;
1306        mbio->bi_rw =
1307            WRITE | do_flush_fua | do_sync | do_discard | do_same;
1308        mbio->bi_private = r1_bio;
1309
1310        atomic_inc(&r1_bio->remaining);
1311
1312        cb = blk_check_plugged(raid1_unplug, mddev, sizeof(*plug));
1313        if (cb)
1314            plug = container_of(cb, struct raid1_plug_cb, cb);
1315        else
1316            plug = NULL;
1317        spin_lock_irqsave(&conf->device_lock, flags);
1318        if (plug) {
1319            bio_list_add(&plug->pending, mbio);
1320            plug->pending_cnt++;
1321        } else {
1322            bio_list_add(&conf->pending_bio_list, mbio);
1323            conf->pending_count++;
1324        }
1325        spin_unlock_irqrestore(&conf->device_lock, flags);
1326        if (!plug)
1327            md_wakeup_thread(mddev->thread);
1328    }
1329    /* Mustn't call r1_bio_write_done before this next test,
1330     * as it could result in the bio being freed.
1331     */
1332    if (sectors_handled < (bio->bi_size >> 9)) {
1333        r1_bio_write_done(r1_bio);
1334        /* We need another r1_bio. It has already been counted
1335         * in bio->bi_phys_segments
1336         */
1337        r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
1338        r1_bio->master_bio = bio;
1339        r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
1340        r1_bio->state = 0;
1341        r1_bio->mddev = mddev;
1342        r1_bio->sector = bio->bi_sector + sectors_handled;
1343        goto retry_write;
1344    }
1345
1346    r1_bio_write_done(r1_bio);
1347
1348    /* In case raid1d snuck in to freeze_array */
1349    wake_up(&conf->wait_barrier);
1350}
1351
1352static void status(struct seq_file *seq, struct mddev *mddev)
1353{
1354    struct r1conf *conf = mddev->private;
1355    int i;
1356
1357    seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1358           conf->raid_disks - mddev->degraded);
1359    rcu_read_lock();
1360    for (i = 0; i < conf->raid_disks; i++) {
1361        struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1362        seq_printf(seq, "%s",
1363               rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1364    }
1365    rcu_read_unlock();
1366    seq_printf(seq, "]");
1367}
1368
1369
1370static void error(struct mddev *mddev, struct md_rdev *rdev)
1371{
1372    char b[BDEVNAME_SIZE];
1373    struct r1conf *conf = mddev->private;
1374
1375    /*
1376     * If it is not operational, then we have already marked it as dead
1377     * else if it is the last working disks, ignore the error, let the
1378     * next level up know.
1379     * else mark the drive as failed
1380     */
1381    if (test_bit(In_sync, &rdev->flags)
1382        && (conf->raid_disks - mddev->degraded) == 1) {
1383        /*
1384         * Don't fail the drive, act as though we were just a
1385         * normal single drive.
1386         * However don't try a recovery from this drive as
1387         * it is very likely to fail.
1388         */
1389        conf->recovery_disabled = mddev->recovery_disabled;
1390        return;
1391    }
1392    set_bit(Blocked, &rdev->flags);
1393    if (test_and_clear_bit(In_sync, &rdev->flags)) {
1394        unsigned long flags;
1395        spin_lock_irqsave(&conf->device_lock, flags);
1396        mddev->degraded++;
1397        set_bit(Faulty, &rdev->flags);
1398        spin_unlock_irqrestore(&conf->device_lock, flags);
1399        /*
1400         * if recovery is running, make sure it aborts.
1401         */
1402        set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1403    } else
1404        set_bit(Faulty, &rdev->flags);
1405    set_bit(MD_CHANGE_DEVS, &mddev->flags);
1406    printk(KERN_ALERT
1407           "md/raid1:%s: Disk failure on %s, disabling device.\n"
1408           "md/raid1:%s: Operation continuing on %d devices.\n",
1409           mdname(mddev), bdevname(rdev->bdev, b),
1410           mdname(mddev), conf->raid_disks - mddev->degraded);
1411}
1412
1413static void print_conf(struct r1conf *conf)
1414{
1415    int i;
1416
1417    printk(KERN_DEBUG "RAID1 conf printout:\n");
1418    if (!conf) {
1419        printk(KERN_DEBUG "(!conf)\n");
1420        return;
1421    }
1422    printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1423        conf->raid_disks);
1424
1425    rcu_read_lock();
1426    for (i = 0; i < conf->raid_disks; i++) {
1427        char b[BDEVNAME_SIZE];
1428        struct md_rdev *rdev = rcu_dereference(conf->mirrors[i].rdev);
1429        if (rdev)
1430            printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1431                   i, !test_bit(In_sync, &rdev->flags),
1432                   !test_bit(Faulty, &rdev->flags),
1433                   bdevname(rdev->bdev,b));
1434    }
1435    rcu_read_unlock();
1436}
1437
1438static void close_sync(struct r1conf *conf)
1439{
1440    wait_barrier(conf);
1441    allow_barrier(conf);
1442
1443    mempool_destroy(conf->r1buf_pool);
1444    conf->r1buf_pool = NULL;
1445}
1446
1447static int raid1_spare_active(struct mddev *mddev)
1448{
1449    int i;
1450    struct r1conf *conf = mddev->private;
1451    int count = 0;
1452    unsigned long flags;
1453
1454    /*
1455     * Find all failed disks within the RAID1 configuration
1456     * and mark them readable.
1457     * Called under mddev lock, so rcu protection not needed.
1458     */
1459    for (i = 0; i < conf->raid_disks; i++) {
1460        struct md_rdev *rdev = conf->mirrors[i].rdev;
1461        struct md_rdev *repl = conf->mirrors[conf->raid_disks + i].rdev;
1462        if (repl
1463            && repl->recovery_offset == MaxSector
1464            && !test_bit(Faulty, &repl->flags)
1465            && !test_and_set_bit(In_sync, &repl->flags)) {
1466            /* replacement has just become active */
1467            if (!rdev ||
1468                !test_and_clear_bit(In_sync, &rdev->flags))
1469                count++;
1470            if (rdev) {
1471                /* Replaced device not technically
1472                 * faulty, but we need to be sure
1473                 * it gets removed and never re-added
1474                 */
1475                set_bit(Faulty, &rdev->flags);
1476                sysfs_notify_dirent_safe(
1477                    rdev->sysfs_state);
1478            }
1479        }
1480        if (rdev
1481            && !test_bit(Faulty, &rdev->flags)
1482            && !test_and_set_bit(In_sync, &rdev->flags)) {
1483            count++;
1484            sysfs_notify_dirent_safe(rdev->sysfs_state);
1485        }
1486    }
1487    spin_lock_irqsave(&conf->device_lock, flags);
1488    mddev->degraded -= count;
1489    spin_unlock_irqrestore(&conf->device_lock, flags);
1490
1491    print_conf(conf);
1492    return count;
1493}
1494
1495
1496static int raid1_add_disk(struct mddev *mddev, struct md_rdev *rdev)
1497{
1498    struct r1conf *conf = mddev->private;
1499    int err = -EEXIST;
1500    int mirror = 0;
1501    struct raid1_info *p;
1502    int first = 0;
1503    int last = conf->raid_disks - 1;
1504    struct request_queue *q = bdev_get_queue(rdev->bdev);
1505
1506    if (mddev->recovery_disabled == conf->recovery_disabled)
1507        return -EBUSY;
1508
1509    if (rdev->raid_disk >= 0)
1510        first = last = rdev->raid_disk;
1511
1512    if (q->merge_bvec_fn) {
1513        set_bit(Unmerged, &rdev->flags);
1514        mddev->merge_check_needed = 1;
1515    }
1516
1517    for (mirror = first; mirror <= last; mirror++) {
1518        p = conf->mirrors+mirror;
1519        if (!p->rdev) {
1520
1521            disk_stack_limits(mddev->gendisk, rdev->bdev,
1522                      rdev->data_offset << 9);
1523
1524            p->head_position = 0;
1525            rdev->raid_disk = mirror;
1526            err = 0;
1527            /* As all devices are equivalent, we don't need a full recovery
1528             * if this was recently any drive of the array
1529             */
1530            if (rdev->saved_raid_disk < 0)
1531                conf->fullsync = 1;
1532            rcu_assign_pointer(p->rdev, rdev);
1533            break;
1534        }
1535        if (test_bit(WantReplacement, &p->rdev->flags) &&
1536            p[conf->raid_disks].rdev == NULL) {
1537            /* Add this device as a replacement */
1538            clear_bit(In_sync, &rdev->flags);
1539            set_bit(Replacement, &rdev->flags);
1540            rdev->raid_disk = mirror;
1541            err = 0;
1542            conf->fullsync = 1;
1543            rcu_assign_pointer(p[conf->raid_disks].rdev, rdev);
1544            break;
1545        }
1546    }
1547    if (err == 0 && test_bit(Unmerged, &rdev->flags)) {
1548        /* Some requests might not have seen this new
1549         * merge_bvec_fn. We must wait for them to complete
1550         * before merging the device fully.
1551         * First we make sure any code which has tested
1552         * our function has submitted the request, then
1553         * we wait for all outstanding requests to complete.
1554         */
1555        synchronize_sched();
1556        raise_barrier(conf);
1557        lower_barrier(conf);
1558        clear_bit(Unmerged, &rdev->flags);
1559    }
1560    md_integrity_add_rdev(rdev, mddev);
1561    if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
1562        queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, mddev->queue);
1563    print_conf(conf);
1564    return err;
1565}
1566
1567static int raid1_remove_disk(struct mddev *mddev, struct md_rdev *rdev)
1568{
1569    struct r1conf *conf = mddev->private;
1570    int err = 0;
1571    int number = rdev->raid_disk;
1572    struct raid1_info *p = conf->mirrors + number;
1573
1574    if (rdev != p->rdev)
1575        p = conf->mirrors + conf->raid_disks + number;
1576
1577    print_conf(conf);
1578    if (rdev == p->rdev) {
1579        if (test_bit(In_sync, &rdev->flags) ||
1580            atomic_read(&rdev->nr_pending)) {
1581            err = -EBUSY;
1582            goto abort;
1583        }
1584        /* Only remove non-faulty devices if recovery
1585         * is not possible.
1586         */
1587        if (!test_bit(Faulty, &rdev->flags) &&
1588            mddev->recovery_disabled != conf->recovery_disabled &&
1589            mddev->degraded < conf->raid_disks) {
1590            err = -EBUSY;
1591            goto abort;
1592        }
1593        p->rdev = NULL;
1594        synchronize_rcu();
1595        if (atomic_read(&rdev->nr_pending)) {
1596            /* lost the race, try later */
1597            err = -EBUSY;
1598            p->rdev = rdev;
1599            goto abort;
1600        } else if (conf->mirrors[conf->raid_disks + number].rdev) {
1601            /* We just removed a device that is being replaced.
1602             * Move down the replacement. We drain all IO before
1603             * doing this to avoid confusion.
1604             */
1605            struct md_rdev *repl =
1606                conf->mirrors[conf->raid_disks + number].rdev;
1607            raise_barrier(conf);
1608            clear_bit(Replacement, &repl->flags);
1609            p->rdev = repl;
1610            conf->mirrors[conf->raid_disks + number].rdev = NULL;
1611            lower_barrier(conf);
1612            clear_bit(WantReplacement, &rdev->flags);
1613        } else
1614            clear_bit(WantReplacement, &rdev->flags);
1615        err = md_integrity_register(mddev);
1616    }
1617abort:
1618
1619    print_conf(conf);
1620    return err;
1621}
1622
1623
1624static void end_sync_read(struct bio *bio, int error)
1625{
1626    struct r1bio *r1_bio = bio->bi_private;
1627
1628    update_head_pos(r1_bio->read_disk, r1_bio);
1629
1630    /*
1631     * we have read a block, now it needs to be re-written,
1632     * or re-read if the read failed.
1633     * We don't do much here, just schedule handling by raid1d
1634     */
1635    if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1636        set_bit(R1BIO_Uptodate, &r1_bio->state);
1637
1638    if (atomic_dec_and_test(&r1_bio->remaining))
1639        reschedule_retry(r1_bio);
1640}
1641
1642static void end_sync_write(struct bio *bio, int error)
1643{
1644    int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1645    struct r1bio *r1_bio = bio->bi_private;
1646    struct mddev *mddev = r1_bio->mddev;
1647    struct r1conf *conf = mddev->private;
1648    int mirror=0;
1649    sector_t first_bad;
1650    int bad_sectors;
1651
1652    mirror = find_bio_disk(r1_bio, bio);
1653
1654    if (!uptodate) {
1655        sector_t sync_blocks = 0;
1656        sector_t s = r1_bio->sector;
1657        long sectors_to_go = r1_bio->sectors;
1658        /* make sure these bits doesn't get cleared. */
1659        do {
1660            bitmap_end_sync(mddev->bitmap, s,
1661                    &sync_blocks, 1);
1662            s += sync_blocks;
1663            sectors_to_go -= sync_blocks;
1664        } while (sectors_to_go > 0);
1665        set_bit(WriteErrorSeen,
1666            &conf->mirrors[mirror].rdev->flags);
1667        if (!test_and_set_bit(WantReplacement,
1668                      &conf->mirrors[mirror].rdev->flags))
1669            set_bit(MD_RECOVERY_NEEDED, &
1670                mddev->recovery);
1671        set_bit(R1BIO_WriteError, &r1_bio->state);
1672    } else if (is_badblock(conf->mirrors[mirror].rdev,
1673                   r1_bio->sector,
1674                   r1_bio->sectors,
1675                   &first_bad, &bad_sectors) &&
1676           !is_badblock(conf->mirrors[r1_bio->read_disk].rdev,
1677                r1_bio->sector,
1678                r1_bio->sectors,
1679                &first_bad, &bad_sectors)
1680        )
1681        set_bit(R1BIO_MadeGood, &r1_bio->state);
1682
1683    if (atomic_dec_and_test(&r1_bio->remaining)) {
1684        int s = r1_bio->sectors;
1685        if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1686            test_bit(R1BIO_WriteError, &r1_bio->state))
1687            reschedule_retry(r1_bio);
1688        else {
1689            put_buf(r1_bio);
1690            md_done_sync(mddev, s, uptodate);
1691        }
1692    }
1693}
1694
1695static int r1_sync_page_io(struct md_rdev *rdev, sector_t sector,
1696                int sectors, struct page *page, int rw)
1697{
1698    if (sync_page_io(rdev, sector, sectors << 9, page, rw, false))
1699        /* success */
1700        return 1;
1701    if (rw == WRITE) {
1702        set_bit(WriteErrorSeen, &rdev->flags);
1703        if (!test_and_set_bit(WantReplacement,
1704                      &rdev->flags))
1705            set_bit(MD_RECOVERY_NEEDED, &
1706                rdev->mddev->recovery);
1707    }
1708    /* need to record an error - either for the block or the device */
1709    if (!rdev_set_badblocks(rdev, sector, sectors, 0))
1710        md_error(rdev->mddev, rdev);
1711    return 0;
1712}
1713
1714static int fix_sync_read_error(struct r1bio *r1_bio)
1715{
1716    /* Try some synchronous reads of other devices to get
1717     * good data, much like with normal read errors. Only
1718     * read into the pages we already have so we don't
1719     * need to re-issue the read request.
1720     * We don't need to freeze the array, because being in an
1721     * active sync request, there is no normal IO, and
1722     * no overlapping syncs.
1723     * We don't need to check is_badblock() again as we
1724     * made sure that anything with a bad block in range
1725     * will have bi_end_io clear.
1726     */
1727    struct mddev *mddev = r1_bio->mddev;
1728    struct r1conf *conf = mddev->private;
1729    struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1730    sector_t sect = r1_bio->sector;
1731    int sectors = r1_bio->sectors;
1732    int idx = 0;
1733
1734    while(sectors) {
1735        int s = sectors;
1736        int d = r1_bio->read_disk;
1737        int success = 0;
1738        struct md_rdev *rdev;
1739        int start;
1740
1741        if (s > (PAGE_SIZE>>9))
1742            s = PAGE_SIZE >> 9;
1743        do {
1744            if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1745                /* No rcu protection needed here devices
1746                 * can only be removed when no resync is
1747                 * active, and resync is currently active
1748                 */
1749                rdev = conf->mirrors[d].rdev;
1750                if (sync_page_io(rdev, sect, s<<9,
1751                         bio->bi_io_vec[idx].bv_page,
1752                         READ, false)) {
1753                    success = 1;
1754                    break;
1755                }
1756            }
1757            d++;
1758            if (d == conf->raid_disks * 2)
1759                d = 0;
1760        } while (!success && d != r1_bio->read_disk);
1761
1762        if (!success) {
1763            char b[BDEVNAME_SIZE];
1764            int abort = 0;
1765            /* Cannot read from anywhere, this block is lost.
1766             * Record a bad block on each device. If that doesn't
1767             * work just disable and interrupt the recovery.
1768             * Don't fail devices as that won't really help.
1769             */
1770            printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1771                   " for block %llu\n",
1772                   mdname(mddev),
1773                   bdevname(bio->bi_bdev, b),
1774                   (unsigned long long)r1_bio->sector);
1775            for (d = 0; d < conf->raid_disks * 2; d++) {
1776                rdev = conf->mirrors[d].rdev;
1777                if (!rdev || test_bit(Faulty, &rdev->flags))
1778                    continue;
1779                if (!rdev_set_badblocks(rdev, sect, s, 0))
1780                    abort = 1;
1781            }
1782            if (abort) {
1783                conf->recovery_disabled =
1784                    mddev->recovery_disabled;
1785                set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1786                md_done_sync(mddev, r1_bio->sectors, 0);
1787                put_buf(r1_bio);
1788                return 0;
1789            }
1790            /* Try next page */
1791            sectors -= s;
1792            sect += s;
1793            idx++;
1794            continue;
1795        }
1796
1797        start = d;
1798        /* write it back and re-read */
1799        while (d != r1_bio->read_disk) {
1800            if (d == 0)
1801                d = conf->raid_disks * 2;
1802            d--;
1803            if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1804                continue;
1805            rdev = conf->mirrors[d].rdev;
1806            if (r1_sync_page_io(rdev, sect, s,
1807                        bio->bi_io_vec[idx].bv_page,
1808                        WRITE) == 0) {
1809                r1_bio->bios[d]->bi_end_io = NULL;
1810                rdev_dec_pending(rdev, mddev);
1811            }
1812        }
1813        d = start;
1814        while (d != r1_bio->read_disk) {
1815            if (d == 0)
1816                d = conf->raid_disks * 2;
1817            d--;
1818            if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1819                continue;
1820            rdev = conf->mirrors[d].rdev;
1821            if (r1_sync_page_io(rdev, sect, s,
1822                        bio->bi_io_vec[idx].bv_page,
1823                        READ) != 0)
1824                atomic_add(s, &rdev->corrected_errors);
1825        }
1826        sectors -= s;
1827        sect += s;
1828        idx ++;
1829    }
1830    set_bit(R1BIO_Uptodate, &r1_bio->state);
1831    set_bit(BIO_UPTODATE, &bio->bi_flags);
1832    return 1;
1833}
1834
1835static int process_checks(struct r1bio *r1_bio)
1836{
1837    /* We have read all readable devices. If we haven't
1838     * got the block, then there is no hope left.
1839     * If we have, then we want to do a comparison
1840     * and skip the write if everything is the same.
1841     * If any blocks failed to read, then we need to
1842     * attempt an over-write
1843     */
1844    struct mddev *mddev = r1_bio->mddev;
1845    struct r1conf *conf = mddev->private;
1846    int primary;
1847    int i;
1848    int vcnt;
1849
1850    for (primary = 0; primary < conf->raid_disks * 2; primary++)
1851        if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1852            test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1853            r1_bio->bios[primary]->bi_end_io = NULL;
1854            rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1855            break;
1856        }
1857    r1_bio->read_disk = primary;
1858    vcnt = (r1_bio->sectors + PAGE_SIZE / 512 - 1) >> (PAGE_SHIFT - 9);
1859    for (i = 0; i < conf->raid_disks * 2; i++) {
1860        int j;
1861        struct bio *pbio = r1_bio->bios[primary];
1862        struct bio *sbio = r1_bio->bios[i];
1863        int size;
1864
1865        if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1866            continue;
1867
1868        if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1869            for (j = vcnt; j-- ; ) {
1870                struct page *p, *s;
1871                p = pbio->bi_io_vec[j].bv_page;
1872                s = sbio->bi_io_vec[j].bv_page;
1873                if (memcmp(page_address(p),
1874                       page_address(s),
1875                       sbio->bi_io_vec[j].bv_len))
1876                    break;
1877            }
1878        } else
1879            j = 0;
1880        if (j >= 0)
1881            atomic64_add(r1_bio->sectors, &mddev->resync_mismatches);
1882        if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1883                  && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1884            /* No need to write to this device. */
1885            sbio->bi_end_io = NULL;
1886            rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1887            continue;
1888        }
1889        /* fixup the bio for reuse */
1890        sbio->bi_vcnt = vcnt;
1891        sbio->bi_size = r1_bio->sectors << 9;
1892        sbio->bi_idx = 0;
1893        sbio->bi_phys_segments = 0;
1894        sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1895        sbio->bi_flags |= 1 << BIO_UPTODATE;
1896        sbio->bi_next = NULL;
1897        sbio->bi_sector = r1_bio->sector +
1898            conf->mirrors[i].rdev->data_offset;
1899        sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1900        size = sbio->bi_size;
1901        for (j = 0; j < vcnt ; j++) {
1902            struct bio_vec *bi;
1903            bi = &sbio->bi_io_vec[j];
1904            bi->bv_offset = 0;
1905            if (size > PAGE_SIZE)
1906                bi->bv_len = PAGE_SIZE;
1907            else
1908                bi->bv_len = size;
1909            size -= PAGE_SIZE;
1910            memcpy(page_address(bi->bv_page),
1911                   page_address(pbio->bi_io_vec[j].bv_page),
1912                   PAGE_SIZE);
1913        }
1914    }
1915    return 0;
1916}
1917
1918static void sync_request_write(struct mddev *mddev, struct r1bio *r1_bio)
1919{
1920    struct r1conf *conf = mddev->private;
1921    int i;
1922    int disks = conf->raid_disks * 2;
1923    struct bio *bio, *wbio;
1924
1925    bio = r1_bio->bios[r1_bio->read_disk];
1926
1927    if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1928        /* ouch - failed to read all of that. */
1929        if (!fix_sync_read_error(r1_bio))
1930            return;
1931
1932    if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1933        if (process_checks(r1_bio) < 0)
1934            return;
1935    /*
1936     * schedule writes
1937     */
1938    atomic_set(&r1_bio->remaining, 1);
1939    for (i = 0; i < disks ; i++) {
1940        wbio = r1_bio->bios[i];
1941        if (wbio->bi_end_io == NULL ||
1942            (wbio->bi_end_io == end_sync_read &&
1943             (i == r1_bio->read_disk ||
1944              !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1945            continue;
1946
1947        wbio->bi_rw = WRITE;
1948        wbio->bi_end_io = end_sync_write;
1949        atomic_inc(&r1_bio->remaining);
1950        md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1951
1952        generic_make_request(wbio);
1953    }
1954
1955    if (atomic_dec_and_test(&r1_bio->remaining)) {
1956        /* if we're here, all write(s) have completed, so clean up */
1957        int s = r1_bio->sectors;
1958        if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
1959            test_bit(R1BIO_WriteError, &r1_bio->state))
1960            reschedule_retry(r1_bio);
1961        else {
1962            put_buf(r1_bio);
1963            md_done_sync(mddev, s, 1);
1964        }
1965    }
1966}
1967
1968/*
1969 * This is a kernel thread which:
1970 *
1971 * 1. Retries failed read operations on working mirrors.
1972 * 2. Updates the raid superblock when problems encounter.
1973 * 3. Performs writes following reads for array synchronising.
1974 */
1975
1976static void fix_read_error(struct r1conf *conf, int read_disk,
1977               sector_t sect, int sectors)
1978{
1979    struct mddev *mddev = conf->mddev;
1980    while(sectors) {
1981        int s = sectors;
1982        int d = read_disk;
1983        int success = 0;
1984        int start;
1985        struct md_rdev *rdev;
1986
1987        if (s > (PAGE_SIZE>>9))
1988            s = PAGE_SIZE >> 9;
1989
1990        do {
1991            /* Note: no rcu protection needed here
1992             * as this is synchronous in the raid1d thread
1993             * which is the thread that might remove
1994             * a device. If raid1d ever becomes multi-threaded....
1995             */
1996            sector_t first_bad;
1997            int bad_sectors;
1998
1999            rdev = conf->mirrors[d].rdev;
2000            if (rdev &&
2001                (test_bit(In_sync, &rdev->flags) ||
2002                 (!test_bit(Faulty, &rdev->flags) &&
2003                  rdev->recovery_offset >= sect + s)) &&
2004                is_badblock(rdev, sect, s,
2005                    &first_bad, &bad_sectors) == 0 &&
2006                sync_page_io(rdev, sect, s<<9,
2007                     conf->tmppage, READ, false))
2008                success = 1;
2009            else {
2010                d++;
2011                if (d == conf->raid_disks * 2)
2012                    d = 0;
2013            }
2014        } while (!success && d != read_disk);
2015
2016        if (!success) {
2017            /* Cannot read from anywhere - mark it bad */
2018            struct md_rdev *rdev = conf->mirrors[read_disk].rdev;
2019            if (!rdev_set_badblocks(rdev, sect, s, 0))
2020                md_error(mddev, rdev);
2021            break;
2022        }
2023        /* write it back and re-read */
2024        start = d;
2025        while (d != read_disk) {
2026            if (d==0)
2027                d = conf->raid_disks * 2;
2028            d--;
2029            rdev = conf->mirrors[d].rdev;
2030            if (rdev &&
2031                test_bit(In_sync, &rdev->flags))
2032                r1_sync_page_io(rdev, sect, s,
2033                        conf->tmppage, WRITE);
2034        }
2035        d = start;
2036        while (d != read_disk) {
2037            char b[BDEVNAME_SIZE];
2038            if (d==0)
2039                d = conf->raid_disks * 2;
2040            d--;
2041            rdev = conf->mirrors[d].rdev;
2042            if (rdev &&
2043                test_bit(In_sync, &rdev->flags)) {
2044                if (r1_sync_page_io(rdev, sect, s,
2045                            conf->tmppage, READ)) {
2046                    atomic_add(s, &rdev->corrected_errors);
2047                    printk(KERN_INFO
2048                           "md/raid1:%s: read error corrected "
2049                           "(%d sectors at %llu on %s)\n",
2050                           mdname(mddev), s,
2051                           (unsigned long long)(sect +
2052                               rdev->data_offset),
2053                           bdevname(rdev->bdev, b));
2054                }
2055            }
2056        }
2057        sectors -= s;
2058        sect += s;
2059    }
2060}
2061
2062static void bi_complete(struct bio *bio, int error)
2063{
2064    complete((struct completion *)bio->bi_private);
2065}
2066
2067static int submit_bio_wait(int rw, struct bio *bio)
2068{
2069    struct completion event;
2070    rw |= REQ_SYNC;
2071
2072    init_completion(&event);
2073    bio->bi_private = &event;
2074    bio->bi_end_io = bi_complete;
2075    submit_bio(rw, bio);
2076    wait_for_completion(&event);
2077
2078    return test_bit(BIO_UPTODATE, &bio->bi_flags);
2079}
2080
2081static int narrow_write_error(struct r1bio *r1_bio, int i)
2082{
2083    struct mddev *mddev = r1_bio->mddev;
2084    struct r1conf *conf = mddev->private;
2085    struct md_rdev *rdev = conf->mirrors[i].rdev;
2086    int vcnt, idx;
2087    struct bio_vec *vec;
2088
2089    /* bio has the data to be written to device 'i' where
2090     * we just recently had a write error.
2091     * We repeatedly clone the bio and trim down to one block,
2092     * then try the write. Where the write fails we record
2093     * a bad block.
2094     * It is conceivable that the bio doesn't exactly align with
2095     * blocks. We must handle this somehow.
2096     *
2097     * We currently own a reference on the rdev.
2098     */
2099
2100    int block_sectors;
2101    sector_t sector;
2102    int sectors;
2103    int sect_to_write = r1_bio->sectors;
2104    int ok = 1;
2105
2106    if (rdev->badblocks.shift < 0)
2107        return 0;
2108
2109    block_sectors = 1 << rdev->badblocks.shift;
2110    sector = r1_bio->sector;
2111    sectors = ((sector + block_sectors)
2112           & ~(sector_t)(block_sectors - 1))
2113        - sector;
2114
2115    if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
2116        vcnt = r1_bio->behind_page_count;
2117        vec = r1_bio->behind_bvecs;
2118        idx = 0;
2119        while (vec[idx].bv_page == NULL)
2120            idx++;
2121    } else {
2122        vcnt = r1_bio->master_bio->bi_vcnt;
2123        vec = r1_bio->master_bio->bi_io_vec;
2124        idx = r1_bio->master_bio->bi_idx;
2125    }
2126    while (sect_to_write) {
2127        struct bio *wbio;
2128        if (sectors > sect_to_write)
2129            sectors = sect_to_write;
2130        /* Write at 'sector' for 'sectors'*/
2131
2132        wbio = bio_alloc_mddev(GFP_NOIO, vcnt, mddev);
2133        memcpy(wbio->bi_io_vec, vec, vcnt * sizeof(struct bio_vec));
2134        wbio->bi_sector = r1_bio->sector;
2135        wbio->bi_rw = WRITE;
2136        wbio->bi_vcnt = vcnt;
2137        wbio->bi_size = r1_bio->sectors << 9;
2138        wbio->bi_idx = idx;
2139
2140        md_trim_bio(wbio, sector - r1_bio->sector, sectors);
2141        wbio->bi_sector += rdev->data_offset;
2142        wbio->bi_bdev = rdev->bdev;
2143        if (submit_bio_wait(WRITE, wbio) == 0)
2144            /* failure! */
2145            ok = rdev_set_badblocks(rdev, sector,
2146                        sectors, 0)
2147                && ok;
2148
2149        bio_put(wbio);
2150        sect_to_write -= sectors;
2151        sector += sectors;
2152        sectors = block_sectors;
2153    }
2154    return ok;
2155}
2156
2157static void handle_sync_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2158{
2159    int m;
2160    int s = r1_bio->sectors;
2161    for (m = 0; m < conf->raid_disks * 2 ; m++) {
2162        struct md_rdev *rdev = conf->mirrors[m].rdev;
2163        struct bio *bio = r1_bio->bios[m];
2164        if (bio->bi_end_io == NULL)
2165            continue;
2166        if (test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2167            test_bit(R1BIO_MadeGood, &r1_bio->state)) {
2168            rdev_clear_badblocks(rdev, r1_bio->sector, s, 0);
2169        }
2170        if (!test_bit(BIO_UPTODATE, &bio->bi_flags) &&
2171            test_bit(R1BIO_WriteError, &r1_bio->state)) {
2172            if (!rdev_set_badblocks(rdev, r1_bio->sector, s, 0))
2173                md_error(conf->mddev, rdev);
2174        }
2175    }
2176    put_buf(r1_bio);
2177    md_done_sync(conf->mddev, s, 1);
2178}
2179
2180static void handle_write_finished(struct r1conf *conf, struct r1bio *r1_bio)
2181{
2182    int m;
2183    for (m = 0; m < conf->raid_disks * 2 ; m++)
2184        if (r1_bio->bios[m] == IO_MADE_GOOD) {
2185            struct md_rdev *rdev = conf->mirrors[m].rdev;
2186            rdev_clear_badblocks(rdev,
2187                         r1_bio->sector,
2188                         r1_bio->sectors, 0);
2189            rdev_dec_pending(rdev, conf->mddev);
2190        } else if (r1_bio->bios[m] != NULL) {
2191            /* This drive got a write error. We need to
2192             * narrow down and record precise write
2193             * errors.
2194             */
2195            if (!narrow_write_error(r1_bio, m)) {
2196                md_error(conf->mddev,
2197                     conf->mirrors[m].rdev);
2198                /* an I/O failed, we can't clear the bitmap */
2199                set_bit(R1BIO_Degraded, &r1_bio->state);
2200            }
2201            rdev_dec_pending(conf->mirrors[m].rdev,
2202                     conf->mddev);
2203        }
2204    if (test_bit(R1BIO_WriteError, &r1_bio->state))
2205        close_write(r1_bio);
2206    raid_end_bio_io(r1_bio);
2207}
2208
2209static void handle_read_error(struct r1conf *conf, struct r1bio *r1_bio)
2210{
2211    int disk;
2212    int max_sectors;
2213    struct mddev *mddev = conf->mddev;
2214    struct bio *bio;
2215    char b[BDEVNAME_SIZE];
2216    struct md_rdev *rdev;
2217
2218    clear_bit(R1BIO_ReadError, &r1_bio->state);
2219    /* we got a read error. Maybe the drive is bad. Maybe just
2220     * the block and we can fix it.
2221     * We freeze all other IO, and try reading the block from
2222     * other devices. When we find one, we re-write
2223     * and check it that fixes the read error.
2224     * This is all done synchronously while the array is
2225     * frozen
2226     */
2227    if (mddev->ro == 0) {
2228        freeze_array(conf);
2229        fix_read_error(conf, r1_bio->read_disk,
2230                   r1_bio->sector, r1_bio->sectors);
2231        unfreeze_array(conf);
2232    } else
2233        md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
2234    rdev_dec_pending(conf->mirrors[r1_bio->read_disk].rdev, conf->mddev);
2235
2236    bio = r1_bio->bios[r1_bio->read_disk];
2237    bdevname(bio->bi_bdev, b);
2238read_more:
2239    disk = read_balance(conf, r1_bio, &max_sectors);
2240    if (disk == -1) {
2241        printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
2242               " read error for block %llu\n",
2243               mdname(mddev), b, (unsigned long long)r1_bio->sector);
2244        raid_end_bio_io(r1_bio);
2245    } else {
2246        const unsigned long do_sync
2247            = r1_bio->master_bio->bi_rw & REQ_SYNC;
2248        if (bio) {
2249            r1_bio->bios[r1_bio->read_disk] =
2250                mddev->ro ? IO_BLOCKED : NULL;
2251            bio_put(bio);
2252        }
2253        r1_bio->read_disk = disk;
2254        bio = bio_clone_mddev(r1_bio->master_bio, GFP_NOIO, mddev);
2255        md_trim_bio(bio, r1_bio->sector - bio->bi_sector, max_sectors);
2256        r1_bio->bios[r1_bio->read_disk] = bio;
2257        rdev = conf->mirrors[disk].rdev;
2258        printk_ratelimited(KERN_ERR
2259                   "md/raid1:%s: redirecting sector %llu"
2260                   " to other mirror: %s\n",
2261                   mdname(mddev),
2262                   (unsigned long long)r1_bio->sector,
2263                   bdevname(rdev->bdev, b));
2264        bio->bi_sector = r1_bio->sector + rdev->data_offset;
2265        bio->bi_bdev = rdev->bdev;
2266        bio->bi_end_io = raid1_end_read_request;
2267        bio->bi_rw = READ | do_sync;
2268        bio->bi_private = r1_bio;
2269        if (max_sectors < r1_bio->sectors) {
2270            /* Drat - have to split this up more */
2271            struct bio *mbio = r1_bio->master_bio;
2272            int sectors_handled = (r1_bio->sector + max_sectors
2273                           - mbio->bi_sector);
2274            r1_bio->sectors = max_sectors;
2275            spin_lock_irq(&conf->device_lock);
2276            if (mbio->bi_phys_segments == 0)
2277                mbio->bi_phys_segments = 2;
2278            else
2279                mbio->bi_phys_segments++;
2280            spin_unlock_irq(&conf->device_lock);
2281            generic_make_request(bio);
2282            bio = NULL;
2283
2284            r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
2285
2286            r1_bio->master_bio = mbio;
2287            r1_bio->sectors = (mbio->bi_size >> 9)
2288                      - sectors_handled;
2289            r1_bio->state = 0;
2290            set_bit(R1BIO_ReadError, &r1_bio->state);
2291            r1_bio->mddev = mddev;
2292            r1_bio->sector = mbio->bi_sector + sectors_handled;
2293
2294            goto read_more;
2295        } else
2296            generic_make_request(bio);
2297    }
2298}
2299
2300static void raid1d(struct md_thread *thread)
2301{
2302    struct mddev *mddev = thread->mddev;
2303    struct r1bio *r1_bio;
2304    unsigned long flags;
2305    struct r1conf *conf = mddev->private;
2306    struct list_head *head = &conf->retry_list;
2307    struct blk_plug plug;
2308
2309    md_check_recovery(mddev);
2310
2311    blk_start_plug(&plug);
2312    for (;;) {
2313
2314        flush_pending_writes(conf);
2315
2316        spin_lock_irqsave(&conf->device_lock, flags);
2317        if (list_empty(head)) {
2318            spin_unlock_irqrestore(&conf->device_lock, flags);
2319            break;
2320        }
2321        r1_bio = list_entry(head->prev, struct r1bio, retry_list);
2322        list_del(head->prev);
2323        conf->nr_queued--;
2324        spin_unlock_irqrestore(&conf->device_lock, flags);
2325
2326        mddev = r1_bio->mddev;
2327        conf = mddev->private;
2328        if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
2329            if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2330                test_bit(R1BIO_WriteError, &r1_bio->state))
2331                handle_sync_write_finished(conf, r1_bio);
2332            else
2333                sync_request_write(mddev, r1_bio);
2334        } else if (test_bit(R1BIO_MadeGood, &r1_bio->state) ||
2335               test_bit(R1BIO_WriteError, &r1_bio->state))
2336            handle_write_finished(conf, r1_bio);
2337        else if (test_bit(R1BIO_ReadError, &r1_bio->state))
2338            handle_read_error(conf, r1_bio);
2339        else
2340            /* just a partial read to be scheduled from separate
2341             * context
2342             */
2343            generic_make_request(r1_bio->bios[r1_bio->read_disk]);
2344
2345        cond_resched();
2346        if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
2347            md_check_recovery(mddev);
2348    }
2349    blk_finish_plug(&plug);
2350}
2351
2352
2353static int init_resync(struct r1conf *conf)
2354{
2355    int buffs;
2356
2357    buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
2358    BUG_ON(conf->r1buf_pool);
2359    conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
2360                      conf->poolinfo);
2361    if (!conf->r1buf_pool)
2362        return -ENOMEM;
2363    conf->next_resync = 0;
2364    return 0;
2365}
2366
2367/*
2368 * perform a "sync" on one "block"
2369 *
2370 * We need to make sure that no normal I/O request - particularly write
2371 * requests - conflict with active sync requests.
2372 *
2373 * This is achieved by tracking pending requests and a 'barrier' concept
2374 * that can be installed to exclude normal IO requests.
2375 */
2376
2377static sector_t sync_request(struct mddev *mddev, sector_t sector_nr, int *skipped, int go_faster)
2378{
2379    struct r1conf *conf = mddev->private;
2380    struct r1bio *r1_bio;
2381    struct bio *bio;
2382    sector_t max_sector, nr_sectors;
2383    int disk = -1;
2384    int i;
2385    int wonly = -1;
2386    int write_targets = 0, read_targets = 0;
2387    sector_t sync_blocks;
2388    int still_degraded = 0;
2389    int good_sectors = RESYNC_SECTORS;
2390    int min_bad = 0; /* number of sectors that are bad in all devices */
2391
2392    if (!conf->r1buf_pool)
2393        if (init_resync(conf))
2394            return 0;
2395
2396    max_sector = mddev->dev_sectors;
2397    if (sector_nr >= max_sector) {
2398        /* If we aborted, we need to abort the
2399         * sync on the 'current' bitmap chunk (there will
2400         * only be one in raid1 resync.
2401         * We can find the current addess in mddev->curr_resync
2402         */
2403        if (mddev->curr_resync < max_sector) /* aborted */
2404            bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
2405                        &sync_blocks, 1);
2406        else /* completed sync */
2407            conf->fullsync = 0;
2408
2409        bitmap_close_sync(mddev->bitmap);
2410        close_sync(conf);
2411        return 0;
2412    }
2413
2414    if (mddev->bitmap == NULL &&
2415        mddev->recovery_cp == MaxSector &&
2416        !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
2417        conf->fullsync == 0) {
2418        *skipped = 1;
2419        return max_sector - sector_nr;
2420    }
2421    /* before building a request, check if we can skip these blocks..
2422     * This call the bitmap_start_sync doesn't actually record anything
2423     */
2424    if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
2425        !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2426        /* We can skip this block, and probably several more */
2427        *skipped = 1;
2428        return sync_blocks;
2429    }
2430    /*
2431     * If there is non-resync activity waiting for a turn,
2432     * and resync is going fast enough,
2433     * then let it though before starting on this new sync request.
2434     */
2435    if (!go_faster && conf->nr_waiting)
2436        msleep_interruptible(1000);
2437
2438    bitmap_cond_end_sync(mddev->bitmap, sector_nr);
2439    r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
2440    raise_barrier(conf);
2441
2442    conf->next_resync = sector_nr;
2443
2444    rcu_read_lock();
2445    /*
2446     * If we get a correctably read error during resync or recovery,
2447     * we might want to read from a different device. So we
2448     * flag all drives that could conceivably be read from for READ,
2449     * and any others (which will be non-In_sync devices) for WRITE.
2450     * If a read fails, we try reading from something else for which READ
2451     * is OK.
2452     */
2453
2454    r1_bio->mddev = mddev;
2455    r1_bio->sector = sector_nr;
2456    r1_bio->state = 0;
2457    set_bit(R1BIO_IsSync, &r1_bio->state);
2458
2459    for (i = 0; i < conf->raid_disks * 2; i++) {
2460        struct md_rdev *rdev;
2461        bio = r1_bio->bios[i];
2462
2463        /* take from bio_init */
2464        bio->bi_next = NULL;
2465        bio->bi_flags &= ~(BIO_POOL_MASK-1);
2466        bio->bi_flags |= 1 << BIO_UPTODATE;
2467        bio->bi_rw = READ;
2468        bio->bi_vcnt = 0;
2469        bio->bi_idx = 0;
2470        bio->bi_phys_segments = 0;
2471        bio->bi_size = 0;
2472        bio->bi_end_io = NULL;
2473        bio->bi_private = NULL;
2474
2475        rdev = rcu_dereference(conf->mirrors[i].rdev);
2476        if (rdev == NULL ||
2477            test_bit(Faulty, &rdev->flags)) {
2478            if (i < conf->raid_disks)
2479                still_degraded = 1;
2480        } else if (!test_bit(In_sync, &rdev->flags)) {
2481            bio->bi_rw = WRITE;
2482            bio->bi_end_io = end_sync_write;
2483            write_targets ++;
2484        } else {
2485            /* may need to read from here */
2486            sector_t first_bad = MaxSector;
2487            int bad_sectors;
2488
2489            if (is_badblock(rdev, sector_nr, good_sectors,
2490                    &first_bad, &bad_sectors)) {
2491                if (first_bad > sector_nr)
2492                    good_sectors = first_bad - sector_nr;
2493                else {
2494                    bad_sectors -= (sector_nr - first_bad);
2495                    if (min_bad == 0 ||
2496                        min_bad > bad_sectors)
2497                        min_bad = bad_sectors;
2498                }
2499            }
2500            if (sector_nr < first_bad) {
2501                if (test_bit(WriteMostly, &rdev->flags)) {
2502                    if (wonly < 0)
2503                        wonly = i;
2504                } else {
2505                    if (disk < 0)
2506                        disk = i;
2507                }
2508                bio->bi_rw = READ;
2509                bio->bi_end_io = end_sync_read;
2510                read_targets++;
2511            } else if (!test_bit(WriteErrorSeen, &rdev->flags) &&
2512                test_bit(MD_RECOVERY_SYNC, &mddev->recovery) &&
2513                !test_bit(MD_RECOVERY_CHECK, &mddev->recovery)) {
2514                /*
2515                 * The device is suitable for reading (InSync),
2516                 * but has bad block(s) here. Let's try to correct them,
2517                 * if we are doing resync or repair. Otherwise, leave
2518                 * this device alone for this sync request.
2519                 */
2520                bio->bi_rw = WRITE;
2521                bio->bi_end_io = end_sync_write;
2522                write_targets++;
2523            }
2524        }
2525        if (bio->bi_end_io) {
2526            atomic_inc(&rdev->nr_pending);
2527            bio->bi_sector = sector_nr + rdev->data_offset;
2528            bio->bi_bdev = rdev->bdev;
2529            bio->bi_private = r1_bio;
2530        }
2531    }
2532    rcu_read_unlock();
2533    if (disk < 0)
2534        disk = wonly;
2535    r1_bio->read_disk = disk;
2536
2537    if (read_targets == 0 && min_bad > 0) {
2538        /* These sectors are bad on all InSync devices, so we
2539         * need to mark them bad on all write targets
2540         */
2541        int ok = 1;
2542        for (i = 0 ; i < conf->raid_disks * 2 ; i++)
2543            if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
2544                struct md_rdev *rdev = conf->mirrors[i].rdev;
2545                ok = rdev_set_badblocks(rdev, sector_nr,
2546                            min_bad, 0
2547                    ) && ok;
2548            }
2549        set_bit(MD_CHANGE_DEVS, &mddev->flags);
2550        *skipped = 1;
2551        put_buf(r1_bio);
2552
2553        if (!ok) {
2554            /* Cannot record the badblocks, so need to
2555             * abort the resync.
2556             * If there are multiple read targets, could just
2557             * fail the really bad ones ???
2558             */
2559            conf->recovery_disabled = mddev->recovery_disabled;
2560            set_bit(MD_RECOVERY_INTR, &mddev->recovery);
2561            return 0;
2562        } else
2563            return min_bad;
2564
2565    }
2566    if (min_bad > 0 && min_bad < good_sectors) {
2567        /* only resync enough to reach the next bad->good
2568         * transition */
2569        good_sectors = min_bad;
2570    }
2571
2572    if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
2573        /* extra read targets are also write targets */
2574        write_targets += read_targets-1;
2575
2576    if (write_targets == 0 || read_targets == 0) {
2577        /* There is nowhere to write, so all non-sync
2578         * drives must be failed - so we are finished
2579         */
2580        sector_t rv;
2581        if (min_bad > 0)
2582            max_sector = sector_nr + min_bad;
2583        rv = max_sector - sector_nr;
2584        *skipped = 1;
2585        put_buf(r1_bio);
2586        return rv;
2587    }
2588
2589    if (max_sector > mddev->resync_max)
2590        max_sector = mddev->resync_max; /* Don't do IO beyond here */
2591    if (max_sector > sector_nr + good_sectors)
2592        max_sector = sector_nr + good_sectors;
2593    nr_sectors = 0;
2594    sync_blocks = 0;
2595    do {
2596        struct page *page;
2597        int len = PAGE_SIZE;
2598        if (sector_nr + (len>>9) > max_sector)
2599            len = (max_sector - sector_nr) << 9;
2600        if (len == 0)
2601            break;
2602        if (sync_blocks == 0) {
2603            if (!bitmap_start_sync(mddev->bitmap, sector_nr,
2604                           &sync_blocks, still_degraded) &&
2605                !conf->fullsync &&
2606                !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2607                break;
2608            BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2609            if ((len >> 9) > sync_blocks)
2610                len = sync_blocks<<9;
2611        }
2612
2613        for (i = 0 ; i < conf->raid_disks * 2; i++) {
2614            bio = r1_bio->bios[i];
2615            if (bio->bi_end_io) {
2616                page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2617                if (bio_add_page(bio, page, len, 0) == 0) {
2618                    /* stop here */
2619                    bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2620                    while (i > 0) {
2621                        i--;
2622                        bio = r1_bio->bios[i];
2623                        if (bio->bi_end_io==NULL)
2624                            continue;
2625                        /* remove last page from this bio */
2626                        bio->bi_vcnt--;
2627                        bio->bi_size -= len;
2628                        bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2629                    }
2630                    goto bio_full;
2631                }
2632            }
2633        }
2634        nr_sectors += len>>9;
2635        sector_nr += len>>9;
2636        sync_blocks -= (len>>9);
2637    } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2638 bio_full:
2639    r1_bio->sectors = nr_sectors;
2640
2641    /* For a user-requested sync, we read all readable devices and do a
2642     * compare
2643     */
2644    if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2645        atomic_set(&r1_bio->remaining, read_targets);
2646        for (i = 0; i < conf->raid_disks * 2 && read_targets; i++) {
2647            bio = r1_bio->bios[i];
2648            if (bio->bi_end_io == end_sync_read) {
2649                read_targets--;
2650                md_sync_acct(bio->bi_bdev, nr_sectors);
2651                generic_make_request(bio);
2652            }
2653        }
2654    } else {
2655        atomic_set(&r1_bio->remaining, 1);
2656        bio = r1_bio->bios[r1_bio->read_disk];
2657        md_sync_acct(bio->bi_bdev, nr_sectors);
2658        generic_make_request(bio);
2659
2660    }
2661    return nr_sectors;
2662}
2663
2664static sector_t raid1_size(struct mddev *mddev, sector_t sectors, int raid_disks)
2665{
2666    if (sectors)
2667        return sectors;
2668
2669    return mddev->dev_sectors;
2670}
2671
2672static struct r1conf *setup_conf(struct mddev *mddev)
2673{
2674    struct r1conf *conf;
2675    int i;
2676    struct raid1_info *disk;
2677    struct md_rdev *rdev;
2678    int err = -ENOMEM;
2679
2680    conf = kzalloc(sizeof(struct r1conf), GFP_KERNEL);
2681    if (!conf)
2682        goto abort;
2683
2684    conf->mirrors = kzalloc(sizeof(struct raid1_info)
2685                * mddev->raid_disks * 2,
2686                 GFP_KERNEL);
2687    if (!conf->mirrors)
2688        goto abort;
2689
2690    conf->tmppage = alloc_page(GFP_KERNEL);
2691    if (!conf->tmppage)
2692        goto abort;
2693
2694    conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2695    if (!conf->poolinfo)
2696        goto abort;
2697    conf->poolinfo->raid_disks = mddev->raid_disks * 2;
2698    conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2699                      r1bio_pool_free,
2700                      conf->poolinfo);
2701    if (!conf->r1bio_pool)
2702        goto abort;
2703
2704    conf->poolinfo->mddev = mddev;
2705
2706    err = -EINVAL;
2707    spin_lock_init(&conf->device_lock);
2708    rdev_for_each(rdev, mddev) {
2709        struct request_queue *q;
2710        int disk_idx = rdev->raid_disk;
2711        if (disk_idx >= mddev->raid_disks
2712            || disk_idx < 0)
2713            continue;
2714        if (test_bit(Replacement, &rdev->flags))
2715            disk = conf->mirrors + mddev->raid_disks + disk_idx;
2716        else
2717            disk = conf->mirrors + disk_idx;
2718
2719        if (disk->rdev)
2720            goto abort;
2721        disk->rdev = rdev;
2722        q = bdev_get_queue(rdev->bdev);
2723        if (q->merge_bvec_fn)
2724            mddev->merge_check_needed = 1;
2725
2726        disk->head_position = 0;
2727        disk->seq_start = MaxSector;
2728    }
2729    conf->raid_disks = mddev->raid_disks;
2730    conf->mddev = mddev;
2731    INIT_LIST_HEAD(&conf->retry_list);
2732
2733    spin_lock_init(&conf->resync_lock);
2734    init_waitqueue_head(&conf->wait_barrier);
2735
2736    bio_list_init(&conf->pending_bio_list);
2737    conf->pending_count = 0;
2738    conf->recovery_disabled = mddev->recovery_disabled - 1;
2739
2740    err = -EIO;
2741    for (i = 0; i < conf->raid_disks * 2; i++) {
2742
2743        disk = conf->mirrors + i;
2744
2745        if (i < conf->raid_disks &&
2746            disk[conf->raid_disks].rdev) {
2747            /* This slot has a replacement. */
2748            if (!disk->rdev) {
2749                /* No original, just make the replacement
2750                 * a recovering spare
2751                 */
2752                disk->rdev =
2753                    disk[conf->raid_disks].rdev;
2754                disk[conf->raid_disks].rdev = NULL;
2755            } else if (!test_bit(In_sync, &disk->rdev->flags))
2756                /* Original is not in_sync - bad */
2757                goto abort;
2758        }
2759
2760        if (!disk->rdev ||
2761            !test_bit(In_sync, &disk->rdev->flags)) {
2762            disk->head_position = 0;
2763            if (disk->rdev &&
2764                (disk->rdev->saved_raid_disk < 0))
2765                conf->fullsync = 1;
2766        }
2767    }
2768
2769    err = -ENOMEM;
2770    conf->thread = md_register_thread(raid1d, mddev, "raid1");
2771    if (!conf->thread) {
2772        printk(KERN_ERR
2773               "md/raid1:%s: couldn't allocate thread\n",
2774               mdname(mddev));
2775        goto abort;
2776    }
2777
2778    return conf;
2779
2780 abort:
2781    if (conf) {
2782        if (conf->r1bio_pool)
2783            mempool_destroy(conf->r1bio_pool);
2784        kfree(conf->mirrors);
2785        safe_put_page(conf->tmppage);
2786        kfree(conf->poolinfo);
2787        kfree(conf);
2788    }
2789    return ERR_PTR(err);
2790}
2791
2792static int stop(struct mddev *mddev);
2793static int run(struct mddev *mddev)
2794{
2795    struct r1conf *conf;
2796    int i;
2797    struct md_rdev *rdev;
2798    int ret;
2799    bool discard_supported = false;
2800
2801    if (mddev->level != 1) {
2802        printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2803               mdname(mddev), mddev->level);
2804        return -EIO;
2805    }
2806    if (mddev->reshape_position != MaxSector) {
2807        printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2808               mdname(mddev));
2809        return -EIO;
2810    }
2811    /*
2812     * copy the already verified devices into our private RAID1
2813     * bookkeeping area. [whatever we allocate in run(),
2814     * should be freed in stop()]
2815     */
2816    if (mddev->private == NULL)
2817        conf = setup_conf(mddev);
2818    else
2819        conf = mddev->private;
2820
2821    if (IS_ERR(conf))
2822        return PTR_ERR(conf);
2823
2824    if (mddev->queue)
2825        blk_queue_max_write_same_sectors(mddev->queue,
2826                         mddev->chunk_sectors);
2827    rdev_for_each(rdev, mddev) {
2828        if (!mddev->gendisk)
2829            continue;
2830        disk_stack_limits(mddev->gendisk, rdev->bdev,
2831                  rdev->data_offset << 9);
2832        if (blk_queue_discard(bdev_get_queue(rdev->bdev)))
2833            discard_supported = true;
2834    }
2835
2836    mddev->degraded = 0;
2837    for (i=0; i < conf->raid_disks; i++)
2838        if (conf->mirrors[i].rdev == NULL ||
2839            !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2840            test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2841            mddev->degraded++;
2842
2843    if (conf->raid_disks - mddev->degraded == 1)
2844        mddev->recovery_cp = MaxSector;
2845
2846    if (mddev->recovery_cp != MaxSector)
2847        printk(KERN_NOTICE "md/raid1:%s: not clean"
2848               " -- starting background reconstruction\n",
2849               mdname(mddev));
2850    printk(KERN_INFO
2851        "md/raid1:%s: active with %d out of %d mirrors\n",
2852        mdname(mddev), mddev->raid_disks - mddev->degraded,
2853        mddev->raid_disks);
2854
2855    /*
2856     * Ok, everything is just fine now
2857     */
2858    mddev->thread = conf->thread;
2859    conf->thread = NULL;
2860    mddev->private = conf;
2861
2862    md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2863
2864    if (mddev->queue) {
2865        mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2866        mddev->queue->backing_dev_info.congested_data = mddev;
2867        blk_queue_merge_bvec(mddev->queue, raid1_mergeable_bvec);
2868
2869        if (discard_supported)
2870            queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
2871                        mddev->queue);
2872        else
2873            queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD,
2874                          mddev->queue);
2875    }
2876
2877    ret = md_integrity_register(mddev);
2878    if (ret)
2879        stop(mddev);
2880    return ret;
2881}
2882
2883static int stop(struct mddev *mddev)
2884{
2885    struct r1conf *conf = mddev->private;
2886    struct bitmap *bitmap = mddev->bitmap;
2887
2888    /* wait for behind writes to complete */
2889    if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2890        printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2891               mdname(mddev));
2892        /* need to kick something here to make sure I/O goes? */
2893        wait_event(bitmap->behind_wait,
2894               atomic_read(&bitmap->behind_writes) == 0);
2895    }
2896
2897    raise_barrier(conf);
2898    lower_barrier(conf);
2899
2900    md_unregister_thread(&mddev->thread);
2901    if (conf->r1bio_pool)
2902        mempool_destroy(conf->r1bio_pool);
2903    kfree(conf->mirrors);
2904    kfree(conf->poolinfo);
2905    kfree(conf);
2906    mddev->private = NULL;
2907    return 0;
2908}
2909
2910static int raid1_resize(struct mddev *mddev, sector_t sectors)
2911{
2912    /* no resync is happening, and there is enough space
2913     * on all devices, so we can resize.
2914     * We need to make sure resync covers any new space.
2915     * If the array is shrinking we should possibly wait until
2916     * any io in the removed space completes, but it hardly seems
2917     * worth it.
2918     */
2919    sector_t newsize = raid1_size(mddev, sectors, 0);
2920    if (mddev->external_size &&
2921        mddev->array_sectors > newsize)
2922        return -EINVAL;
2923    if (mddev->bitmap) {
2924        int ret = bitmap_resize(mddev->bitmap, newsize, 0, 0);
2925        if (ret)
2926            return ret;
2927    }
2928    md_set_array_sectors(mddev, newsize);
2929    set_capacity(mddev->gendisk, mddev->array_sectors);
2930    revalidate_disk(mddev->gendisk);
2931    if (sectors > mddev->dev_sectors &&
2932        mddev->recovery_cp > mddev->dev_sectors) {
2933        mddev->recovery_cp = mddev->dev_sectors;
2934        set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2935    }
2936    mddev->dev_sectors = sectors;
2937    mddev->resync_max_sectors = sectors;
2938    return 0;
2939}
2940
2941static int raid1_reshape(struct mddev *mddev)
2942{
2943    /* We need to:
2944     * 1/ resize the r1bio_pool
2945     * 2/ resize conf->mirrors
2946     *
2947     * We allocate a new r1bio_pool if we can.
2948     * Then raise a device barrier and wait until all IO stops.
2949     * Then resize conf->mirrors and swap in the new r1bio pool.
2950     *
2951     * At the same time, we "pack" the devices so that all the missing
2952     * devices have the higher raid_disk numbers.
2953     */
2954    mempool_t *newpool, *oldpool;
2955    struct pool_info *newpoolinfo;
2956    struct raid1_info *newmirrors;
2957    struct r1conf *conf = mddev->private;
2958    int cnt, raid_disks;
2959    unsigned long flags;
2960    int d, d2, err;
2961
2962    /* Cannot change chunk_size, layout, or level */
2963    if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2964        mddev->layout != mddev->new_layout ||
2965        mddev->level != mddev->new_level) {
2966        mddev->new_chunk_sectors = mddev->chunk_sectors;
2967        mddev->new_layout = mddev->layout;
2968        mddev->new_level = mddev->level;
2969        return -EINVAL;
2970    }
2971
2972    err = md_allow_write(mddev);
2973    if (err)
2974        return err;
2975
2976    raid_disks = mddev->raid_disks + mddev->delta_disks;
2977
2978    if (raid_disks < conf->raid_disks) {
2979        cnt=0;
2980        for (d= 0; d < conf->raid_disks; d++)
2981            if (conf->mirrors[d].rdev)
2982                cnt++;
2983        if (cnt > raid_disks)
2984            return -EBUSY;
2985    }
2986
2987    newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2988    if (!newpoolinfo)
2989        return -ENOMEM;
2990    newpoolinfo->mddev = mddev;
2991    newpoolinfo->raid_disks = raid_disks * 2;
2992
2993    newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2994                 r1bio_pool_free, newpoolinfo);
2995    if (!newpool) {
2996        kfree(newpoolinfo);
2997        return -ENOMEM;
2998    }
2999    newmirrors = kzalloc(sizeof(struct raid1_info) * raid_disks * 2,
3000                 GFP_KERNEL);
3001    if (!newmirrors) {
3002        kfree(newpoolinfo);
3003        mempool_destroy(newpool);
3004        return -ENOMEM;
3005    }
3006
3007    raise_barrier(conf);
3008
3009    /* ok, everything is stopped */
3010    oldpool = conf->r1bio_pool;
3011    conf->r1bio_pool = newpool;
3012
3013    for (d = d2 = 0; d < conf->raid_disks; d++) {
3014        struct md_rdev *rdev = conf->mirrors[d].rdev;
3015        if (rdev && rdev->raid_disk != d2) {
3016            sysfs_unlink_rdev(mddev, rdev);
3017            rdev->raid_disk = d2;
3018            sysfs_unlink_rdev(mddev, rdev);
3019            if (sysfs_link_rdev(mddev, rdev))
3020                printk(KERN_WARNING
3021                       "md/raid1:%s: cannot register rd%d\n",
3022                       mdname(mddev), rdev->raid_disk);
3023        }
3024        if (rdev)
3025            newmirrors[d2++].rdev = rdev;
3026    }
3027    kfree(conf->mirrors);
3028    conf->mirrors = newmirrors;
3029    kfree(conf->poolinfo);
3030    conf->poolinfo = newpoolinfo;
3031
3032    spin_lock_irqsave(&conf->device_lock, flags);
3033    mddev->degraded += (raid_disks - conf->raid_disks);
3034    spin_unlock_irqrestore(&conf->device_lock, flags);
3035    conf->raid_disks = mddev->raid_disks = raid_disks;
3036    mddev->delta_disks = 0;
3037
3038    lower_barrier(conf);
3039
3040    set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
3041    md_wakeup_thread(mddev->thread);
3042
3043    mempool_destroy(oldpool);
3044    return 0;
3045}
3046
3047static void raid1_quiesce(struct mddev *mddev, int state)
3048{
3049    struct r1conf *conf = mddev->private;
3050
3051    switch(state) {
3052    case 2: /* wake for suspend */
3053        wake_up(&conf->wait_barrier);
3054        break;
3055    case 1:
3056        raise_barrier(conf);
3057        break;
3058    case 0:
3059        lower_barrier(conf);
3060        break;
3061    }
3062}
3063
3064static void *raid1_takeover(struct mddev *mddev)
3065{
3066    /* raid1 can take over:
3067     * raid5 with 2 devices, any layout or chunk size
3068     */
3069    if (mddev->level == 5 && mddev->raid_disks == 2) {
3070        struct r1conf *conf;
3071        mddev->new_level = 1;
3072        mddev->new_layout = 0;
3073        mddev->new_chunk_sectors = 0;
3074        conf = setup_conf(mddev);
3075        if (!IS_ERR(conf))
3076            conf->barrier = 1;
3077        return conf;
3078    }
3079    return ERR_PTR(-EINVAL);
3080}
3081
3082static struct md_personality raid1_personality =
3083{
3084    .name = "raid1",
3085    .level = 1,
3086    .owner = THIS_MODULE,
3087    .make_request = make_request,
3088    .run = run,
3089    .stop = stop,
3090    .status = status,
3091    .error_handler = error,
3092    .hot_add_disk = raid1_add_disk,
3093    .hot_remove_disk= raid1_remove_disk,
3094    .spare_active = raid1_spare_active,
3095    .sync_request = sync_request,
3096    .resize = raid1_resize,
3097    .size = raid1_size,
3098    .check_reshape = raid1_reshape,
3099    .quiesce = raid1_quiesce,
3100    .takeover = raid1_takeover,
3101};
3102
3103static int __init raid_init(void)
3104{
3105    return register_md_personality(&raid1_personality);
3106}
3107
3108static void raid_exit(void)
3109{
3110    unregister_md_personality(&raid1_personality);
3111}
3112
3113module_init(raid_init);
3114module_exit(raid_exit);
3115MODULE_LICENSE("GPL");
3116MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
3117MODULE_ALIAS("md-personality-3"); /* RAID1 */
3118MODULE_ALIAS("md-raid1");
3119MODULE_ALIAS("md-level-1");
3120
3121module_param(max_queued_requests, int, S_IRUGO|S_IWUSR);
3122

Archive Download this file



interactive