projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
block: raid fixups for removal of bi_hw_segments
[deliverable/linux.git] / 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 Østergaard" <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 "dm-bio-list.h"
35 #include <linux/raid/raid1.h>
36 #include <linux/raid/bitmap.h>
37
38 #define DEBUG 0
39 #if DEBUG
40 #define PRINTK(x...) printk(x)
41 #else
42 #define PRINTK(x...)
43 #endif
44
45 /*
46 * Number of guaranteed r1bios in case of extreme VM load:
47 */
48 #define NR_RAID1_BIOS 256
49
50
51 static void unplug_slaves(mddev_t *mddev);
52
53 static void allow_barrier(conf_t *conf);
54 static void lower_barrier(conf_t *conf);
55
56 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
57 {
58 struct pool_info *pi = data;
59 r1bio_t *r1_bio;
60 int size = offsetof(r1bio_t, bios[pi->raid_disks]);
61
62 /* allocate a r1bio with room for raid_disks entries in the bios array */
63 r1_bio = kzalloc(size, gfp_flags);
64 if (!r1_bio)
65 unplug_slaves(pi->mddev);
66
67 return r1_bio;
68 }
69
70 static void r1bio_pool_free(void *r1_bio, void *data)
71 {
72 kfree(r1_bio);
73 }
74
75 #define RESYNC_BLOCK_SIZE (64*1024)
76 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
77 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
78 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
79 #define RESYNC_WINDOW (2048*1024)
80
81 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
82 {
83 struct pool_info *pi = data;
84 struct page *page;
85 r1bio_t *r1_bio;
86 struct bio *bio;
87 int i, j;
88
89 r1_bio = r1bio_pool_alloc(gfp_flags, pi);
90 if (!r1_bio) {
91 unplug_slaves(pi->mddev);
92 return NULL;
93 }
94
95 /*
96 * Allocate bios : 1 for reading, n-1 for writing
97 */
98 for (j = pi->raid_disks ; j-- ; ) {
99 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
100 if (!bio)
101 goto out_free_bio;
102 r1_bio->bios[j] = bio;
103 }
104 /*
105 * Allocate RESYNC_PAGES data pages and attach them to
106 * the first bio.
107 * If this is a user-requested check/repair, allocate
108 * RESYNC_PAGES for each bio.
109 */
110 if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
111 j = pi->raid_disks;
112 else
113 j = 1;
114 while(j--) {
115 bio = r1_bio->bios[j];
116 for (i = 0; i < RESYNC_PAGES; i++) {
117 page = alloc_page(gfp_flags);
118 if (unlikely(!page))
119 goto out_free_pages;
120
121 bio->bi_io_vec[i].bv_page = page;
122 }
123 }
124 /* If not user-requests, copy the page pointers to all bios */
125 if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
126 for (i=0; i<RESYNC_PAGES ; i++)
127 for (j=1; j<pi->raid_disks; j++)
128 r1_bio->bios[j]->bi_io_vec[i].bv_page =
129 r1_bio->bios[0]->bi_io_vec[i].bv_page;
130 }
131
132 r1_bio->master_bio = NULL;
133
134 return r1_bio;
135
136 out_free_pages:
137 for (i=0; i < RESYNC_PAGES ; i++)
138 for (j=0 ; j < pi->raid_disks; j++)
139 safe_put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
140 j = -1;
141 out_free_bio:
142 while ( ++j < pi->raid_disks )
143 bio_put(r1_bio->bios[j]);
144 r1bio_pool_free(r1_bio, data);
145 return NULL;
146 }
147
148 static void r1buf_pool_free(void *__r1_bio, void *data)
149 {
150 struct pool_info *pi = data;
151 int i,j;
152 r1bio_t *r1bio = __r1_bio;
153
154 for (i = 0; i < RESYNC_PAGES; i++)
155 for (j = pi->raid_disks; j-- ;) {
156 if (j == 0 ||
157 r1bio->bios[j]->bi_io_vec[i].bv_page !=
158 r1bio->bios[0]->bi_io_vec[i].bv_page)
159 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
160 }
161 for (i=0 ; i < pi->raid_disks; i++)
162 bio_put(r1bio->bios[i]);
163
164 r1bio_pool_free(r1bio, data);
165 }
166
167 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
168 {
169 int i;
170
171 for (i = 0; i < conf->raid_disks; i++) {
172 struct bio **bio = r1_bio->bios + i;
173 if (*bio && *bio != IO_BLOCKED)
174 bio_put(*bio);
175 *bio = NULL;
176 }
177 }
178
179 static void free_r1bio(r1bio_t *r1_bio)
180 {
181 conf_t *conf = mddev_to_conf(r1_bio->mddev);
182
183 /*
184 * Wake up any possible resync thread that waits for the device
185 * to go idle.
186 */
187 allow_barrier(conf);
188
189 put_all_bios(conf, r1_bio);
190 mempool_free(r1_bio, conf->r1bio_pool);
191 }
192
193 static void put_buf(r1bio_t *r1_bio)
194 {
195 conf_t *conf = mddev_to_conf(r1_bio->mddev);
196 int i;
197
198 for (i=0; i<conf->raid_disks; i++) {
199 struct bio *bio = r1_bio->bios[i];
200 if (bio->bi_end_io)
201 rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
202 }
203
204 mempool_free(r1_bio, conf->r1buf_pool);
205
206 lower_barrier(conf);
207 }
208
209 static void reschedule_retry(r1bio_t *r1_bio)
210 {
211 unsigned long flags;
212 mddev_t *mddev = r1_bio->mddev;
213 conf_t *conf = mddev_to_conf(mddev);
214
215 spin_lock_irqsave(&conf->device_lock, flags);
216 list_add(&r1_bio->retry_list, &conf->retry_list);
217 conf->nr_queued ++;
218 spin_unlock_irqrestore(&conf->device_lock, flags);
219
220 wake_up(&conf->wait_barrier);
221 md_wakeup_thread(mddev->thread);
222 }
223
224 /*
225 * raid_end_bio_io() is called when we have finished servicing a mirrored
226 * operation and are ready to return a success/failure code to the buffer
227 * cache layer.
228 */
229 static void raid_end_bio_io(r1bio_t *r1_bio)
230 {
231 struct bio *bio = r1_bio->master_bio;
232
233 /* if nobody has done the final endio yet, do it now */
234 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
235 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
236 (bio_data_dir(bio) == WRITE) ? "write" : "read",
237 (unsigned long long) bio->bi_sector,
238 (unsigned long long) bio->bi_sector +
239 (bio->bi_size >> 9) - 1);
240
241 bio_endio(bio,
242 test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
243 }
244 free_r1bio(r1_bio);
245 }
246
247 /*
248 * Update disk head position estimator based on IRQ completion info.
249 */
250 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
251 {
252 conf_t *conf = mddev_to_conf(r1_bio->mddev);
253
254 conf->mirrors[disk].head_position =
255 r1_bio->sector + (r1_bio->sectors);
256 }
257
258 static void raid1_end_read_request(struct bio *bio, int error)
259 {
260 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
261 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
262 int mirror;
263 conf_t *conf = mddev_to_conf(r1_bio->mddev);
264
265 mirror = r1_bio->read_disk;
266 /*
267 * this branch is our 'one mirror IO has finished' event handler:
268 */
269 update_head_pos(mirror, r1_bio);
270
271 if (uptodate)
272 set_bit(R1BIO_Uptodate, &r1_bio->state);
273 else {
274 /* If all other devices have failed, we want to return
275 * the error upwards rather than fail the last device.
276 * Here we redefine "uptodate" to mean "Don't want to retry"
277 */
278 unsigned long flags;
279 spin_lock_irqsave(&conf->device_lock, flags);
280 if (r1_bio->mddev->degraded == conf->raid_disks ||
281 (r1_bio->mddev->degraded == conf->raid_disks-1 &&
282 !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
283 uptodate = 1;
284 spin_unlock_irqrestore(&conf->device_lock, flags);
285 }
286
287 if (uptodate)
288 raid_end_bio_io(r1_bio);
289 else {
290 /*
291 * oops, read error:
292 */
293 char b[BDEVNAME_SIZE];
294 if (printk_ratelimit())
295 printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
296 bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
297 reschedule_retry(r1_bio);
298 }
299
300 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
301 }
302
303 static void raid1_end_write_request(struct bio *bio, int error)
304 {
305 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
306 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
307 int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
308 conf_t *conf = mddev_to_conf(r1_bio->mddev);
309 struct bio *to_put = NULL;
310
311
312 for (mirror = 0; mirror < conf->raid_disks; mirror++)
313 if (r1_bio->bios[mirror] == bio)
314 break;
315
316 if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
317 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
318 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
319 r1_bio->mddev->barriers_work = 0;
320 /* Don't rdev_dec_pending in this branch - keep it for the retry */
321 } else {
322 /*
323 * this branch is our 'one mirror IO has finished' event handler:
324 */
325 r1_bio->bios[mirror] = NULL;
326 to_put = bio;
327 if (!uptodate) {
328 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
329 /* an I/O failed, we can't clear the bitmap */
330 set_bit(R1BIO_Degraded, &r1_bio->state);
331 } else
332 /*
333 * Set R1BIO_Uptodate in our master bio, so that
334 * we will return a good error code for to the higher
335 * levels even if IO on some other mirrored buffer fails.
336 *
337 * The 'master' represents the composite IO operation to
338 * user-side. So if something waits for IO, then it will
339 * wait for the 'master' bio.
340 */
341 set_bit(R1BIO_Uptodate, &r1_bio->state);
342
343 update_head_pos(mirror, r1_bio);
344
345 if (behind) {
346 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
347 atomic_dec(&r1_bio->behind_remaining);
348
349 /* In behind mode, we ACK the master bio once the I/O has safely
350 * reached all non-writemostly disks. Setting the Returned bit
351 * ensures that this gets done only once -- we don't ever want to
352 * return -EIO here, instead we'll wait */
353
354 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
355 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
356 /* Maybe we can return now */
357 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
358 struct bio *mbio = r1_bio->master_bio;
359 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
360 (unsigned long long) mbio->bi_sector,
361 (unsigned long long) mbio->bi_sector +
362 (mbio->bi_size >> 9) - 1);
363 bio_endio(mbio, 0);
364 }
365 }
366 }
367 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
368 }
369 /*
370 *
371 * Let's see if all mirrored write operations have finished
372 * already.
373 */
374 if (atomic_dec_and_test(&r1_bio->remaining)) {
375 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
376 reschedule_retry(r1_bio);
377 else {
378 /* it really is the end of this request */
379 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
380 /* free extra copy of the data pages */
381 int i = bio->bi_vcnt;
382 while (i--)
383 safe_put_page(bio->bi_io_vec[i].bv_page);
384 }
385 /* clear the bitmap if all writes complete successfully */
386 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
387 r1_bio->sectors,
388 !test_bit(R1BIO_Degraded, &r1_bio->state),
389 behind);
390 md_write_end(r1_bio->mddev);
391 raid_end_bio_io(r1_bio);
392 }
393 }
394
395 if (to_put)
396 bio_put(to_put);
397 }
398
399
400 /*
401 * This routine returns the disk from which the requested read should
402 * be done. There is a per-array 'next expected sequential IO' sector
403 * number - if this matches on the next IO then we use the last disk.
404 * There is also a per-disk 'last know head position' sector that is
405 * maintained from IRQ contexts, both the normal and the resync IO
406 * completion handlers update this position correctly. If there is no
407 * perfect sequential match then we pick the disk whose head is closest.
408 *
409 * If there are 2 mirrors in the same 2 devices, performance degrades
410 * because position is mirror, not device based.
411 *
412 * The rdev for the device selected will have nr_pending incremented.
413 */
414 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
415 {
416 const unsigned long this_sector = r1_bio->sector;
417 int new_disk = conf->last_used, disk = new_disk;
418 int wonly_disk = -1;
419 const int sectors = r1_bio->sectors;
420 sector_t new_distance, current_distance;
421 mdk_rdev_t *rdev;
422
423 rcu_read_lock();
424 /*
425 * Check if we can balance. We can balance on the whole
426 * device if no resync is going on, or below the resync window.
427 * We take the first readable disk when above the resync window.
428 */
429 retry:
430 if (conf->mddev->recovery_cp < MaxSector &&
431 (this_sector + sectors >= conf->next_resync)) {
432 /* Choose the first operation device, for consistancy */
433 new_disk = 0;
434
435 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
436 r1_bio->bios[new_disk] == IO_BLOCKED ||
437 !rdev || !test_bit(In_sync, &rdev->flags)
438 || test_bit(WriteMostly, &rdev->flags);
439 rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
440
441 if (rdev && test_bit(In_sync, &rdev->flags) &&
442 r1_bio->bios[new_disk] != IO_BLOCKED)
443 wonly_disk = new_disk;
444
445 if (new_disk == conf->raid_disks - 1) {
446 new_disk = wonly_disk;
447 break;
448 }
449 }
450 goto rb_out;
451 }
452
453
454 /* make sure the disk is operational */
455 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
456 r1_bio->bios[new_disk] == IO_BLOCKED ||
457 !rdev || !test_bit(In_sync, &rdev->flags) ||
458 test_bit(WriteMostly, &rdev->flags);
459 rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
460
461 if (rdev && test_bit(In_sync, &rdev->flags) &&
462 r1_bio->bios[new_disk] != IO_BLOCKED)
463 wonly_disk = new_disk;
464
465 if (new_disk <= 0)
466 new_disk = conf->raid_disks;
467 new_disk--;
468 if (new_disk == disk) {
469 new_disk = wonly_disk;
470 break;
471 }
472 }
473
474 if (new_disk < 0)
475 goto rb_out;
476
477 disk = new_disk;
478 /* now disk == new_disk == starting point for search */
479
480 /*
481 * Don't change to another disk for sequential reads:
482 */
483 if (conf->next_seq_sect == this_sector)
484 goto rb_out;
485 if (this_sector == conf->mirrors[new_disk].head_position)
486 goto rb_out;
487
488 current_distance = abs(this_sector - conf->mirrors[disk].head_position);
489
490 /* Find the disk whose head is closest */
491
492 do {
493 if (disk <= 0)
494 disk = conf->raid_disks;
495 disk--;
496
497 rdev = rcu_dereference(conf->mirrors[disk].rdev);
498
499 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
500 !test_bit(In_sync, &rdev->flags) ||
501 test_bit(WriteMostly, &rdev->flags))
502 continue;
503
504 if (!atomic_read(&rdev->nr_pending)) {
505 new_disk = disk;
506 break;
507 }
508 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
509 if (new_distance < current_distance) {
510 current_distance = new_distance;
511 new_disk = disk;
512 }
513 } while (disk != conf->last_used);
514
515 rb_out:
516
517
518 if (new_disk >= 0) {
519 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
520 if (!rdev)
521 goto retry;
522 atomic_inc(&rdev->nr_pending);
523 if (!test_bit(In_sync, &rdev->flags)) {
524 /* cannot risk returning a device that failed
525 * before we inc'ed nr_pending
526 */
527 rdev_dec_pending(rdev, conf->mddev);
528 goto retry;
529 }
530 conf->next_seq_sect = this_sector + sectors;
531 conf->last_used = new_disk;
532 }
533 rcu_read_unlock();
534
535 return new_disk;
536 }
537
538 static void unplug_slaves(mddev_t *mddev)
539 {
540 conf_t *conf = mddev_to_conf(mddev);
541 int i;
542
543 rcu_read_lock();
544 for (i=0; i<mddev->raid_disks; i++) {
545 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
546 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
547 struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
548
549 atomic_inc(&rdev->nr_pending);
550 rcu_read_unlock();
551
552 blk_unplug(r_queue);
553
554 rdev_dec_pending(rdev, mddev);
555 rcu_read_lock();
556 }
557 }
558 rcu_read_unlock();
559 }
560
561 static void raid1_unplug(struct request_queue *q)
562 {
563 mddev_t *mddev = q->queuedata;
564
565 unplug_slaves(mddev);
566 md_wakeup_thread(mddev->thread);
567 }
568
569 static int raid1_congested(void *data, int bits)
570 {
571 mddev_t *mddev = data;
572 conf_t *conf = mddev_to_conf(mddev);
573 int i, ret = 0;
574
575 rcu_read_lock();
576 for (i = 0; i < mddev->raid_disks; i++) {
577 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
578 if (rdev && !test_bit(Faulty, &rdev->flags)) {
579 struct request_queue *q = bdev_get_queue(rdev->bdev);
580
581 /* Note the '|| 1' - when read_balance prefers
582 * non-congested targets, it can be removed
583 */
584 if ((bits & (1<<BDI_write_congested)) || 1)
585 ret |= bdi_congested(&q->backing_dev_info, bits);
586 else
587 ret &= bdi_congested(&q->backing_dev_info, bits);
588 }
589 }
590 rcu_read_unlock();
591 return ret;
592 }
593
594
595 static int flush_pending_writes(conf_t *conf)
596 {
597 /* Any writes that have been queued but are awaiting
598 * bitmap updates get flushed here.
599 * We return 1 if any requests were actually submitted.
600 */
601 int rv = 0;
602
603 spin_lock_irq(&conf->device_lock);
604
605 if (conf->pending_bio_list.head) {
606 struct bio *bio;
607 bio = bio_list_get(&conf->pending_bio_list);
608 blk_remove_plug(conf->mddev->queue);
609 spin_unlock_irq(&conf->device_lock);
610 /* flush any pending bitmap writes to
611 * disk before proceeding w/ I/O */
612 bitmap_unplug(conf->mddev->bitmap);
613
614 while (bio) { /* submit pending writes */
615 struct bio *next = bio->bi_next;
616 bio->bi_next = NULL;
617 generic_make_request(bio);
618 bio = next;
619 }
620 rv = 1;
621 } else
622 spin_unlock_irq(&conf->device_lock);
623 return rv;
624 }
625
626 /* Barriers....
627 * Sometimes we need to suspend IO while we do something else,
628 * either some resync/recovery, or reconfigure the array.
629 * To do this we raise a 'barrier'.
630 * The 'barrier' is a counter that can be raised multiple times
631 * to count how many activities are happening which preclude
632 * normal IO.
633 * We can only raise the barrier if there is no pending IO.
634 * i.e. if nr_pending == 0.
635 * We choose only to raise the barrier if no-one is waiting for the
636 * barrier to go down. This means that as soon as an IO request
637 * is ready, no other operations which require a barrier will start
638 * until the IO request has had a chance.
639 *
640 * So: regular IO calls 'wait_barrier'. When that returns there
641 * is no backgroup IO happening, It must arrange to call
642 * allow_barrier when it has finished its IO.
643 * backgroup IO calls must call raise_barrier. Once that returns
644 * there is no normal IO happeing. It must arrange to call
645 * lower_barrier when the particular background IO completes.
646 */
647 #define RESYNC_DEPTH 32
648
649 static void raise_barrier(conf_t *conf)
650 {
651 spin_lock_irq(&conf->resync_lock);
652
653 /* Wait until no block IO is waiting */
654 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
655 conf->resync_lock,
656 raid1_unplug(conf->mddev->queue));
657
658 /* block any new IO from starting */
659 conf->barrier++;
660
661 /* No wait for all pending IO to complete */
662 wait_event_lock_irq(conf->wait_barrier,
663 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
664 conf->resync_lock,
665 raid1_unplug(conf->mddev->queue));
666
667 spin_unlock_irq(&conf->resync_lock);
668 }
669
670 static void lower_barrier(conf_t *conf)
671 {
672 unsigned long flags;
673 spin_lock_irqsave(&conf->resync_lock, flags);
674 conf->barrier--;
675 spin_unlock_irqrestore(&conf->resync_lock, flags);
676 wake_up(&conf->wait_barrier);
677 }
678
679 static void wait_barrier(conf_t *conf)
680 {
681 spin_lock_irq(&conf->resync_lock);
682 if (conf->barrier) {
683 conf->nr_waiting++;
684 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
685 conf->resync_lock,
686 raid1_unplug(conf->mddev->queue));
687 conf->nr_waiting--;
688 }
689 conf->nr_pending++;
690 spin_unlock_irq(&conf->resync_lock);
691 }
692
693 static void allow_barrier(conf_t *conf)
694 {
695 unsigned long flags;
696 spin_lock_irqsave(&conf->resync_lock, flags);
697 conf->nr_pending--;
698 spin_unlock_irqrestore(&conf->resync_lock, flags);
699 wake_up(&conf->wait_barrier);
700 }
701
702 static void freeze_array(conf_t *conf)
703 {
704 /* stop syncio and normal IO and wait for everything to
705 * go quite.
706 * We increment barrier and nr_waiting, and then
707 * wait until nr_pending match nr_queued+1
708 * This is called in the context of one normal IO request
709 * that has failed. Thus any sync request that might be pending
710 * will be blocked by nr_pending, and we need to wait for
711 * pending IO requests to complete or be queued for re-try.
712 * Thus the number queued (nr_queued) plus this request (1)
713 * must match the number of pending IOs (nr_pending) before
714 * we continue.
715 */
716 spin_lock_irq(&conf->resync_lock);
717 conf->barrier++;
718 conf->nr_waiting++;
719 wait_event_lock_irq(conf->wait_barrier,
720 conf->nr_pending == conf->nr_queued+1,
721 conf->resync_lock,
722 ({ flush_pending_writes(conf);
723 raid1_unplug(conf->mddev->queue); }));
724 spin_unlock_irq(&conf->resync_lock);
725 }
726 static void unfreeze_array(conf_t *conf)
727 {
728 /* reverse the effect of the freeze */
729 spin_lock_irq(&conf->resync_lock);
730 conf->barrier--;
731 conf->nr_waiting--;
732 wake_up(&conf->wait_barrier);
733 spin_unlock_irq(&conf->resync_lock);
734 }
735
736
737 /* duplicate the data pages for behind I/O */
738 static struct page **alloc_behind_pages(struct bio *bio)
739 {
740 int i;
741 struct bio_vec *bvec;
742 struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
743 GFP_NOIO);
744 if (unlikely(!pages))
745 goto do_sync_io;
746
747 bio_for_each_segment(bvec, bio, i) {
748 pages[i] = alloc_page(GFP_NOIO);
749 if (unlikely(!pages[i]))
750 goto do_sync_io;
751 memcpy(kmap(pages[i]) + bvec->bv_offset,
752 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
753 kunmap(pages[i]);
754 kunmap(bvec->bv_page);
755 }
756
757 return pages;
758
759 do_sync_io:
760 if (pages)
761 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
762 put_page(pages[i]);
763 kfree(pages);
764 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
765 return NULL;
766 }
767
768 static int make_request(struct request_queue *q, struct bio * bio)
769 {
770 mddev_t *mddev = q->queuedata;
771 conf_t *conf = mddev_to_conf(mddev);
772 mirror_info_t *mirror;
773 r1bio_t *r1_bio;
774 struct bio *read_bio;
775 int i, targets = 0, disks;
776 struct bitmap *bitmap;
777 unsigned long flags;
778 struct bio_list bl;
779 struct page **behind_pages = NULL;
780 const int rw = bio_data_dir(bio);
781 const int do_sync = bio_sync(bio);
782 int do_barriers;
783 mdk_rdev_t *blocked_rdev;
784
785 /*
786 * Register the new request and wait if the reconstruction
787 * thread has put up a bar for new requests.
788 * Continue immediately if no resync is active currently.
789 * We test barriers_work *after* md_write_start as md_write_start
790 * may cause the first superblock write, and that will check out
791 * if barriers work.
792 */
793
794 md_write_start(mddev, bio); /* wait on superblock update early */
795
796 if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
797 if (rw == WRITE)
798 md_write_end(mddev);
799 bio_endio(bio, -EOPNOTSUPP);
800 return 0;
801 }
802
803 wait_barrier(conf);
804
805 bitmap = mddev->bitmap;
806
807 disk_stat_inc(mddev->gendisk, ios[rw]);
808 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
809
810 /*
811 * make_request() can abort the operation when READA is being
812 * used and no empty request is available.
813 *
814 */
815 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
816
817 r1_bio->master_bio = bio;
818 r1_bio->sectors = bio->bi_size >> 9;
819 r1_bio->state = 0;
820 r1_bio->mddev = mddev;
821 r1_bio->sector = bio->bi_sector;
822
823 if (rw == READ) {
824 /*
825 * read balancing logic:
826 */
827 int rdisk = read_balance(conf, r1_bio);
828
829 if (rdisk < 0) {
830 /* couldn't find anywhere to read from */
831 raid_end_bio_io(r1_bio);
832 return 0;
833 }
834 mirror = conf->mirrors + rdisk;
835
836 r1_bio->read_disk = rdisk;
837
838 read_bio = bio_clone(bio, GFP_NOIO);
839
840 r1_bio->bios[rdisk] = read_bio;
841
842 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
843 read_bio->bi_bdev = mirror->rdev->bdev;
844 read_bio->bi_end_io = raid1_end_read_request;
845 read_bio->bi_rw = READ | do_sync;
846 read_bio->bi_private = r1_bio;
847
848 generic_make_request(read_bio);
849 return 0;
850 }
851
852 /*
853 * WRITE:
854 */
855 /* first select target devices under spinlock and
856 * inc refcount on their rdev. Record them by setting
857 * bios[x] to bio
858 */
859 disks = conf->raid_disks;
860 #if 0
861 { static int first=1;
862 if (first) printk("First Write sector %llu disks %d\n",
863 (unsigned long long)r1_bio->sector, disks);
864 first = 0;
865 }
866 #endif
867 retry_write:
868 blocked_rdev = NULL;
869 rcu_read_lock();
870 for (i = 0; i < disks; i++) {
871 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
872 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
873 atomic_inc(&rdev->nr_pending);
874 blocked_rdev = rdev;
875 break;
876 }
877 if (rdev && !test_bit(Faulty, &rdev->flags)) {
878 atomic_inc(&rdev->nr_pending);
879 if (test_bit(Faulty, &rdev->flags)) {
880 rdev_dec_pending(rdev, mddev);
881 r1_bio->bios[i] = NULL;
882 } else
883 r1_bio->bios[i] = bio;
884 targets++;
885 } else
886 r1_bio->bios[i] = NULL;
887 }
888 rcu_read_unlock();
889
890 if (unlikely(blocked_rdev)) {
891 /* Wait for this device to become unblocked */
892 int j;
893
894 for (j = 0; j < i; j++)
895 if (r1_bio->bios[j])
896 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
897
898 allow_barrier(conf);
899 md_wait_for_blocked_rdev(blocked_rdev, mddev);
900 wait_barrier(conf);
901 goto retry_write;
902 }
903
904 BUG_ON(targets == 0); /* we never fail the last device */
905
906 if (targets < conf->raid_disks) {
907 /* array is degraded, we will not clear the bitmap
908 * on I/O completion (see raid1_end_write_request) */
909 set_bit(R1BIO_Degraded, &r1_bio->state);
910 }
911
912 /* do behind I/O ? */
913 if (bitmap &&
914 atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
915 (behind_pages = alloc_behind_pages(bio)) != NULL)
916 set_bit(R1BIO_BehindIO, &r1_bio->state);
917
918 atomic_set(&r1_bio->remaining, 0);
919 atomic_set(&r1_bio->behind_remaining, 0);
920
921 do_barriers = bio_barrier(bio);
922 if (do_barriers)
923 set_bit(R1BIO_Barrier, &r1_bio->state);
924
925 bio_list_init(&bl);
926 for (i = 0; i < disks; i++) {
927 struct bio *mbio;
928 if (!r1_bio->bios[i])
929 continue;
930
931 mbio = bio_clone(bio, GFP_NOIO);
932 r1_bio->bios[i] = mbio;
933
934 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
935 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
936 mbio->bi_end_io = raid1_end_write_request;
937 mbio->bi_rw = WRITE | do_barriers | do_sync;
938 mbio->bi_private = r1_bio;
939
940 if (behind_pages) {
941 struct bio_vec *bvec;
942 int j;
943
944 /* Yes, I really want the '__' version so that
945 * we clear any unused pointer in the io_vec, rather
946 * than leave them unchanged. This is important
947 * because when we come to free the pages, we won't
948 * know the originial bi_idx, so we just free
949 * them all
950 */
951 __bio_for_each_segment(bvec, mbio, j, 0)
952 bvec->bv_page = behind_pages[j];
953 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
954 atomic_inc(&r1_bio->behind_remaining);
955 }
956
957 atomic_inc(&r1_bio->remaining);
958
959 bio_list_add(&bl, mbio);
960 }
961 kfree(behind_pages); /* the behind pages are attached to the bios now */
962
963 bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
964 test_bit(R1BIO_BehindIO, &r1_bio->state));
965 spin_lock_irqsave(&conf->device_lock, flags);
966 bio_list_merge(&conf->pending_bio_list, &bl);
967 bio_list_init(&bl);
968
969 blk_plug_device(mddev->queue);
970 spin_unlock_irqrestore(&conf->device_lock, flags);
971
972 /* In case raid1d snuck into freeze_array */
973 wake_up(&conf->wait_barrier);
974
975 if (do_sync)
976 md_wakeup_thread(mddev->thread);
977 #if 0
978 while ((bio = bio_list_pop(&bl)) != NULL)
979 generic_make_request(bio);
980 #endif
981
982 return 0;
983 }
984
985 static void status(struct seq_file *seq, mddev_t *mddev)
986 {
987 conf_t *conf = mddev_to_conf(mddev);
988 int i;
989
990 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
991 conf->raid_disks - mddev->degraded);
992 rcu_read_lock();
993 for (i = 0; i < conf->raid_disks; i++) {
994 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
995 seq_printf(seq, "%s",
996 rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
997 }
998 rcu_read_unlock();
999 seq_printf(seq, "]");
1000 }
1001
1002
1003 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1004 {
1005 char b[BDEVNAME_SIZE];
1006 conf_t *conf = mddev_to_conf(mddev);
1007
1008 /*
1009 * If it is not operational, then we have already marked it as dead
1010 * else if it is the last working disks, ignore the error, let the
1011 * next level up know.
1012 * else mark the drive as failed
1013 */
1014 if (test_bit(In_sync, &rdev->flags)
1015 && (conf->raid_disks - mddev->degraded) == 1)
1016 /*
1017 * Don't fail the drive, act as though we were just a
1018 * normal single drive
1019 */
1020 return;
1021 if (test_and_clear_bit(In_sync, &rdev->flags)) {
1022 unsigned long flags;
1023 spin_lock_irqsave(&conf->device_lock, flags);
1024 mddev->degraded++;
1025 set_bit(Faulty, &rdev->flags);
1026 spin_unlock_irqrestore(&conf->device_lock, flags);
1027 /*
1028 * if recovery is running, make sure it aborts.
1029 */
1030 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1031 } else
1032 set_bit(Faulty, &rdev->flags);
1033 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1034 printk(KERN_ALERT "raid1: Disk failure on %s, disabling device.\n"
1035 "raid1: Operation continuing on %d devices.\n",
1036 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1037 }
1038
1039 static void print_conf(conf_t *conf)
1040 {
1041 int i;
1042
1043 printk("RAID1 conf printout:\n");
1044 if (!conf) {
1045 printk("(!conf)\n");
1046 return;
1047 }
1048 printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1049 conf->raid_disks);
1050
1051 rcu_read_lock();
1052 for (i = 0; i < conf->raid_disks; i++) {
1053 char b[BDEVNAME_SIZE];
1054 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1055 if (rdev)
1056 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1057 i, !test_bit(In_sync, &rdev->flags),
1058 !test_bit(Faulty, &rdev->flags),
1059 bdevname(rdev->bdev,b));
1060 }
1061 rcu_read_unlock();
1062 }
1063
1064 static void close_sync(conf_t *conf)
1065 {
1066 wait_barrier(conf);
1067 allow_barrier(conf);
1068
1069 mempool_destroy(conf->r1buf_pool);
1070 conf->r1buf_pool = NULL;
1071 }
1072
1073 static int raid1_spare_active(mddev_t *mddev)
1074 {
1075 int i;
1076 conf_t *conf = mddev->private;
1077
1078 /*
1079 * Find all failed disks within the RAID1 configuration
1080 * and mark them readable.
1081 * Called under mddev lock, so rcu protection not needed.
1082 */
1083 for (i = 0; i < conf->raid_disks; i++) {
1084 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1085 if (rdev
1086 && !test_bit(Faulty, &rdev->flags)
1087 && !test_and_set_bit(In_sync, &rdev->flags)) {
1088 unsigned long flags;
1089 spin_lock_irqsave(&conf->device_lock, flags);
1090 mddev->degraded--;
1091 spin_unlock_irqrestore(&conf->device_lock, flags);
1092 }
1093 }
1094
1095 print_conf(conf);
1096 return 0;
1097 }
1098
1099
1100 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1101 {
1102 conf_t *conf = mddev->private;
1103 int err = -EEXIST;
1104 int mirror = 0;
1105 mirror_info_t *p;
1106 int first = 0;
1107 int last = mddev->raid_disks - 1;
1108
1109 if (rdev->raid_disk >= 0)
1110 first = last = rdev->raid_disk;
1111
1112 for (mirror = first; mirror <= last; mirror++)
1113 if ( !(p=conf->mirrors+mirror)->rdev) {
1114
1115 blk_queue_stack_limits(mddev->queue,
1116 rdev->bdev->bd_disk->queue);
1117 /* as we don't honour merge_bvec_fn, we must never risk
1118 * violating it, so limit ->max_sector to one PAGE, as
1119 * a one page request is never in violation.
1120 */
1121 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1122 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1123 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1124
1125 p->head_position = 0;
1126 rdev->raid_disk = mirror;
1127 err = 0;
1128 /* As all devices are equivalent, we don't need a full recovery
1129 * if this was recently any drive of the array
1130 */
1131 if (rdev->saved_raid_disk < 0)
1132 conf->fullsync = 1;
1133 rcu_assign_pointer(p->rdev, rdev);
1134 break;
1135 }
1136
1137 print_conf(conf);
1138 return err;
1139 }
1140
1141 static int raid1_remove_disk(mddev_t *mddev, int number)
1142 {
1143 conf_t *conf = mddev->private;
1144 int err = 0;
1145 mdk_rdev_t *rdev;
1146 mirror_info_t *p = conf->mirrors+ number;
1147
1148 print_conf(conf);
1149 rdev = p->rdev;
1150 if (rdev) {
1151 if (test_bit(In_sync, &rdev->flags) ||
1152 atomic_read(&rdev->nr_pending)) {
1153 err = -EBUSY;
1154 goto abort;
1155 }
1156 /* Only remove non-faulty devices is recovery
1157 * is not possible.
1158 */
1159 if (!test_bit(Faulty, &rdev->flags) &&
1160 mddev->degraded < conf->raid_disks) {
1161 err = -EBUSY;
1162 goto abort;
1163 }
1164 p->rdev = NULL;
1165 synchronize_rcu();
1166 if (atomic_read(&rdev->nr_pending)) {
1167 /* lost the race, try later */
1168 err = -EBUSY;
1169 p->rdev = rdev;
1170 }
1171 }
1172 abort:
1173
1174 print_conf(conf);
1175 return err;
1176 }
1177
1178
1179 static void end_sync_read(struct bio *bio, int error)
1180 {
1181 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1182 int i;
1183
1184 for (i=r1_bio->mddev->raid_disks; i--; )
1185 if (r1_bio->bios[i] == bio)
1186 break;
1187 BUG_ON(i < 0);
1188 update_head_pos(i, r1_bio);
1189 /*
1190 * we have read a block, now it needs to be re-written,
1191 * or re-read if the read failed.
1192 * We don't do much here, just schedule handling by raid1d
1193 */
1194 if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1195 set_bit(R1BIO_Uptodate, &r1_bio->state);
1196
1197 if (atomic_dec_and_test(&r1_bio->remaining))
1198 reschedule_retry(r1_bio);
1199 }
1200
1201 static void end_sync_write(struct bio *bio, int error)
1202 {
1203 int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1204 r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1205 mddev_t *mddev = r1_bio->mddev;
1206 conf_t *conf = mddev_to_conf(mddev);
1207 int i;
1208 int mirror=0;
1209
1210 for (i = 0; i < conf->raid_disks; i++)
1211 if (r1_bio->bios[i] == bio) {
1212 mirror = i;
1213 break;
1214 }
1215 if (!uptodate) {
1216 int sync_blocks = 0;
1217 sector_t s = r1_bio->sector;
1218 long sectors_to_go = r1_bio->sectors;
1219 /* make sure these bits doesn't get cleared. */
1220 do {
1221 bitmap_end_sync(mddev->bitmap, s,
1222 &sync_blocks, 1);
1223 s += sync_blocks;
1224 sectors_to_go -= sync_blocks;
1225 } while (sectors_to_go > 0);
1226 md_error(mddev, conf->mirrors[mirror].rdev);
1227 }
1228
1229 update_head_pos(mirror, r1_bio);
1230
1231 if (atomic_dec_and_test(&r1_bio->remaining)) {
1232 md_done_sync(mddev, r1_bio->sectors, uptodate);
1233 put_buf(r1_bio);
1234 }
1235 }
1236
1237 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1238 {
1239 conf_t *conf = mddev_to_conf(mddev);
1240 int i;
1241 int disks = conf->raid_disks;
1242 struct bio *bio, *wbio;
1243
1244 bio = r1_bio->bios[r1_bio->read_disk];
1245
1246
1247 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1248 /* We have read all readable devices. If we haven't
1249 * got the block, then there is no hope left.
1250 * If we have, then we want to do a comparison
1251 * and skip the write if everything is the same.
1252 * If any blocks failed to read, then we need to
1253 * attempt an over-write
1254 */
1255 int primary;
1256 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1257 for (i=0; i<mddev->raid_disks; i++)
1258 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1259 md_error(mddev, conf->mirrors[i].rdev);
1260
1261 md_done_sync(mddev, r1_bio->sectors, 1);
1262 put_buf(r1_bio);
1263 return;
1264 }
1265 for (primary=0; primary<mddev->raid_disks; primary++)
1266 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1267 test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1268 r1_bio->bios[primary]->bi_end_io = NULL;
1269 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1270 break;
1271 }
1272 r1_bio->read_disk = primary;
1273 for (i=0; i<mddev->raid_disks; i++)
1274 if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1275 int j;
1276 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1277 struct bio *pbio = r1_bio->bios[primary];
1278 struct bio *sbio = r1_bio->bios[i];
1279
1280 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1281 for (j = vcnt; j-- ; ) {
1282 struct page *p, *s;
1283 p = pbio->bi_io_vec[j].bv_page;
1284 s = sbio->bi_io_vec[j].bv_page;
1285 if (memcmp(page_address(p),
1286 page_address(s),
1287 PAGE_SIZE))
1288 break;
1289 }
1290 } else
1291 j = 0;
1292 if (j >= 0)
1293 mddev->resync_mismatches += r1_bio->sectors;
1294 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1295 && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1296 sbio->bi_end_io = NULL;
1297 rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1298 } else {
1299 /* fixup the bio for reuse */
1300 int size;
1301 sbio->bi_vcnt = vcnt;
1302 sbio->bi_size = r1_bio->sectors << 9;
1303 sbio->bi_idx = 0;
1304 sbio->bi_phys_segments = 0;
1305 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1306 sbio->bi_flags |= 1 << BIO_UPTODATE;
1307 sbio->bi_next = NULL;
1308 sbio->bi_sector = r1_bio->sector +
1309 conf->mirrors[i].rdev->data_offset;
1310 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1311 size = sbio->bi_size;
1312 for (j = 0; j < vcnt ; j++) {
1313 struct bio_vec *bi;
1314 bi = &sbio->bi_io_vec[j];
1315 bi->bv_offset = 0;
1316 if (size > PAGE_SIZE)
1317 bi->bv_len = PAGE_SIZE;
1318 else
1319 bi->bv_len = size;
1320 size -= PAGE_SIZE;
1321 memcpy(page_address(bi->bv_page),
1322 page_address(pbio->bi_io_vec[j].bv_page),
1323 PAGE_SIZE);
1324 }
1325
1326 }
1327 }
1328 }
1329 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1330 /* ouch - failed to read all of that.
1331 * Try some synchronous reads of other devices to get
1332 * good data, much like with normal read errors. Only
1333 * read into the pages we already have so we don't
1334 * need to re-issue the read request.
1335 * We don't need to freeze the array, because being in an
1336 * active sync request, there is no normal IO, and
1337 * no overlapping syncs.
1338 */
1339 sector_t sect = r1_bio->sector;
1340 int sectors = r1_bio->sectors;
1341 int idx = 0;
1342
1343 while(sectors) {
1344 int s = sectors;
1345 int d = r1_bio->read_disk;
1346 int success = 0;
1347 mdk_rdev_t *rdev;
1348
1349 if (s > (PAGE_SIZE>>9))
1350 s = PAGE_SIZE >> 9;
1351 do {
1352 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1353 /* No rcu protection needed here devices
1354 * can only be removed when no resync is
1355 * active, and resync is currently active
1356 */
1357 rdev = conf->mirrors[d].rdev;
1358 if (sync_page_io(rdev->bdev,
1359 sect + rdev->data_offset,
1360 s<<9,
1361 bio->bi_io_vec[idx].bv_page,
1362 READ)) {
1363 success = 1;
1364 break;
1365 }
1366 }
1367 d++;
1368 if (d == conf->raid_disks)
1369 d = 0;
1370 } while (!success && d != r1_bio->read_disk);
1371
1372 if (success) {
1373 int start = d;
1374 /* write it back and re-read */
1375 set_bit(R1BIO_Uptodate, &r1_bio->state);
1376 while (d != r1_bio->read_disk) {
1377 if (d == 0)
1378 d = conf->raid_disks;
1379 d--;
1380 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1381 continue;
1382 rdev = conf->mirrors[d].rdev;
1383 atomic_add(s, &rdev->corrected_errors);
1384 if (sync_page_io(rdev->bdev,
1385 sect + rdev->data_offset,
1386 s<<9,
1387 bio->bi_io_vec[idx].bv_page,
1388 WRITE) == 0)
1389 md_error(mddev, rdev);
1390 }
1391 d = start;
1392 while (d != r1_bio->read_disk) {
1393 if (d == 0)
1394 d = conf->raid_disks;
1395 d--;
1396 if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1397 continue;
1398 rdev = conf->mirrors[d].rdev;
1399 if (sync_page_io(rdev->bdev,
1400 sect + rdev->data_offset,
1401 s<<9,
1402 bio->bi_io_vec[idx].bv_page,
1403 READ) == 0)
1404 md_error(mddev, rdev);
1405 }
1406 } else {
1407 char b[BDEVNAME_SIZE];
1408 /* Cannot read from anywhere, array is toast */
1409 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1410 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1411 " for block %llu\n",
1412 bdevname(bio->bi_bdev,b),
1413 (unsigned long long)r1_bio->sector);
1414 md_done_sync(mddev, r1_bio->sectors, 0);
1415 put_buf(r1_bio);
1416 return;
1417 }
1418 sectors -= s;
1419 sect += s;
1420 idx ++;
1421 }
1422 }
1423
1424 /*
1425 * schedule writes
1426 */
1427 atomic_set(&r1_bio->remaining, 1);
1428 for (i = 0; i < disks ; i++) {
1429 wbio = r1_bio->bios[i];
1430 if (wbio->bi_end_io == NULL ||
1431 (wbio->bi_end_io == end_sync_read &&
1432 (i == r1_bio->read_disk ||
1433 !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1434 continue;
1435
1436 wbio->bi_rw = WRITE;
1437 wbio->bi_end_io = end_sync_write;
1438 atomic_inc(&r1_bio->remaining);
1439 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1440
1441 generic_make_request(wbio);
1442 }
1443
1444 if (atomic_dec_and_test(&r1_bio->remaining)) {
1445 /* if we're here, all write(s) have completed, so clean up */
1446 md_done_sync(mddev, r1_bio->sectors, 1);
1447 put_buf(r1_bio);
1448 }
1449 }
1450
1451 /*
1452 * This is a kernel thread which:
1453 *
1454 * 1. Retries failed read operations on working mirrors.
1455 * 2. Updates the raid superblock when problems encounter.
1456 * 3. Performs writes following reads for array syncronising.
1457 */
1458
1459 static void fix_read_error(conf_t *conf, int read_disk,
1460 sector_t sect, int sectors)
1461 {
1462 mddev_t *mddev = conf->mddev;
1463 while(sectors) {
1464 int s = sectors;
1465 int d = read_disk;
1466 int success = 0;
1467 int start;
1468 mdk_rdev_t *rdev;
1469
1470 if (s > (PAGE_SIZE>>9))
1471 s = PAGE_SIZE >> 9;
1472
1473 do {
1474 /* Note: no rcu protection needed here
1475 * as this is synchronous in the raid1d thread
1476 * which is the thread that might remove
1477 * a device. If raid1d ever becomes multi-threaded....
1478 */
1479 rdev = conf->mirrors[d].rdev;
1480 if (rdev &&
1481 test_bit(In_sync, &rdev->flags) &&
1482 sync_page_io(rdev->bdev,
1483 sect + rdev->data_offset,
1484 s<<9,
1485 conf->tmppage, READ))
1486 success = 1;
1487 else {
1488 d++;
1489 if (d == conf->raid_disks)
1490 d = 0;
1491 }
1492 } while (!success && d != read_disk);
1493
1494 if (!success) {
1495 /* Cannot read from anywhere -- bye bye array */
1496 md_error(mddev, conf->mirrors[read_disk].rdev);
1497 break;
1498 }
1499 /* write it back and re-read */
1500 start = d;
1501 while (d != read_disk) {
1502 if (d==0)
1503 d = conf->raid_disks;
1504 d--;
1505 rdev = conf->mirrors[d].rdev;
1506 if (rdev &&
1507 test_bit(In_sync, &rdev->flags)) {
1508 if (sync_page_io(rdev->bdev,
1509 sect + rdev->data_offset,
1510 s<<9, conf->tmppage, WRITE)
1511 == 0)
1512 /* Well, this device is dead */
1513 md_error(mddev, rdev);
1514 }
1515 }
1516 d = start;
1517 while (d != read_disk) {
1518 char b[BDEVNAME_SIZE];
1519 if (d==0)
1520 d = conf->raid_disks;
1521 d--;
1522 rdev = conf->mirrors[d].rdev;
1523 if (rdev &&
1524 test_bit(In_sync, &rdev->flags)) {
1525 if (sync_page_io(rdev->bdev,
1526 sect + rdev->data_offset,
1527 s<<9, conf->tmppage, READ)
1528 == 0)
1529 /* Well, this device is dead */
1530 md_error(mddev, rdev);
1531 else {
1532 atomic_add(s, &rdev->corrected_errors);
1533 printk(KERN_INFO
1534 "raid1:%s: read error corrected "
1535 "(%d sectors at %llu on %s)\n",
1536 mdname(mddev), s,
1537 (unsigned long long)(sect +
1538 rdev->data_offset),
1539 bdevname(rdev->bdev, b));
1540 }
1541 }
1542 }
1543 sectors -= s;
1544 sect += s;
1545 }
1546 }
1547
1548 static void raid1d(mddev_t *mddev)
1549 {
1550 r1bio_t *r1_bio;
1551 struct bio *bio;
1552 unsigned long flags;
1553 conf_t *conf = mddev_to_conf(mddev);
1554 struct list_head *head = &conf->retry_list;
1555 int unplug=0;
1556 mdk_rdev_t *rdev;
1557
1558 md_check_recovery(mddev);
1559
1560 for (;;) {
1561 char b[BDEVNAME_SIZE];
1562
1563 unplug += flush_pending_writes(conf);
1564
1565 spin_lock_irqsave(&conf->device_lock, flags);
1566 if (list_empty(head)) {
1567 spin_unlock_irqrestore(&conf->device_lock, flags);
1568 break;
1569 }
1570 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1571 list_del(head->prev);
1572 conf->nr_queued--;
1573 spin_unlock_irqrestore(&conf->device_lock, flags);
1574
1575 mddev = r1_bio->mddev;
1576 conf = mddev_to_conf(mddev);
1577 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1578 sync_request_write(mddev, r1_bio);
1579 unplug = 1;
1580 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1581 /* some requests in the r1bio were BIO_RW_BARRIER
1582 * requests which failed with -EOPNOTSUPP. Hohumm..
1583 * Better resubmit without the barrier.
1584 * We know which devices to resubmit for, because
1585 * all others have had their bios[] entry cleared.
1586 * We already have a nr_pending reference on these rdevs.
1587 */
1588 int i;
1589 const int do_sync = bio_sync(r1_bio->master_bio);
1590 clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1591 clear_bit(R1BIO_Barrier, &r1_bio->state);
1592 for (i=0; i < conf->raid_disks; i++)
1593 if (r1_bio->bios[i])
1594 atomic_inc(&r1_bio->remaining);
1595 for (i=0; i < conf->raid_disks; i++)
1596 if (r1_bio->bios[i]) {
1597 struct bio_vec *bvec;
1598 int j;
1599
1600 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1601 /* copy pages from the failed bio, as
1602 * this might be a write-behind device */
1603 __bio_for_each_segment(bvec, bio, j, 0)
1604 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1605 bio_put(r1_bio->bios[i]);
1606 bio->bi_sector = r1_bio->sector +
1607 conf->mirrors[i].rdev->data_offset;
1608 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1609 bio->bi_end_io = raid1_end_write_request;
1610 bio->bi_rw = WRITE | do_sync;
1611 bio->bi_private = r1_bio;
1612 r1_bio->bios[i] = bio;
1613 generic_make_request(bio);
1614 }
1615 } else {
1616 int disk;
1617
1618 /* we got a read error. Maybe the drive is bad. Maybe just
1619 * the block and we can fix it.
1620 * We freeze all other IO, and try reading the block from
1621 * other devices. When we find one, we re-write
1622 * and check it that fixes the read error.
1623 * This is all done synchronously while the array is
1624 * frozen
1625 */
1626 if (mddev->ro == 0) {
1627 freeze_array(conf);
1628 fix_read_error(conf, r1_bio->read_disk,
1629 r1_bio->sector,
1630 r1_bio->sectors);
1631 unfreeze_array(conf);
1632 }
1633
1634 bio = r1_bio->bios[r1_bio->read_disk];
1635 if ((disk=read_balance(conf, r1_bio)) == -1) {
1636 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1637 " read error for block %llu\n",
1638 bdevname(bio->bi_bdev,b),
1639 (unsigned long long)r1_bio->sector);
1640 raid_end_bio_io(r1_bio);
1641 } else {
1642 const int do_sync = bio_sync(r1_bio->master_bio);
1643 r1_bio->bios[r1_bio->read_disk] =
1644 mddev->ro ? IO_BLOCKED : NULL;
1645 r1_bio->read_disk = disk;
1646 bio_put(bio);
1647 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1648 r1_bio->bios[r1_bio->read_disk] = bio;
1649 rdev = conf->mirrors[disk].rdev;
1650 if (printk_ratelimit())
1651 printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1652 " another mirror\n",
1653 bdevname(rdev->bdev,b),
1654 (unsigned long long)r1_bio->sector);
1655 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1656 bio->bi_bdev = rdev->bdev;
1657 bio->bi_end_io = raid1_end_read_request;
1658 bio->bi_rw = READ | do_sync;
1659 bio->bi_private = r1_bio;
1660 unplug = 1;
1661 generic_make_request(bio);
1662 }
1663 }
1664 }
1665 if (unplug)
1666 unplug_slaves(mddev);
1667 }
1668
1669
1670 static int init_resync(conf_t *conf)
1671 {
1672 int buffs;
1673
1674 buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1675 BUG_ON(conf->r1buf_pool);
1676 conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1677 conf->poolinfo);
1678 if (!conf->r1buf_pool)
1679 return -ENOMEM;
1680 conf->next_resync = 0;
1681 return 0;
1682 }
1683
1684 /*
1685 * perform a "sync" on one "block"
1686 *
1687 * We need to make sure that no normal I/O request - particularly write
1688 * requests - conflict with active sync requests.
1689 *
1690 * This is achieved by tracking pending requests and a 'barrier' concept
1691 * that can be installed to exclude normal IO requests.
1692 */
1693
1694 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1695 {
1696 conf_t *conf = mddev_to_conf(mddev);
1697 r1bio_t *r1_bio;
1698 struct bio *bio;
1699 sector_t max_sector, nr_sectors;
1700 int disk = -1;
1701 int i;
1702 int wonly = -1;
1703 int write_targets = 0, read_targets = 0;
1704 int sync_blocks;
1705 int still_degraded = 0;
1706
1707 if (!conf->r1buf_pool)
1708 {
1709 /*
1710 printk("sync start - bitmap %p\n", mddev->bitmap);
1711 */
1712 if (init_resync(conf))
1713 return 0;
1714 }
1715
1716 max_sector = mddev->size << 1;
1717 if (sector_nr >= max_sector) {
1718 /* If we aborted, we need to abort the
1719 * sync on the 'current' bitmap chunk (there will
1720 * only be one in raid1 resync.
1721 * We can find the current addess in mddev->curr_resync
1722 */
1723 if (mddev->curr_resync < max_sector) /* aborted */
1724 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1725 &sync_blocks, 1);
1726 else /* completed sync */
1727 conf->fullsync = 0;
1728
1729 bitmap_close_sync(mddev->bitmap);
1730 close_sync(conf);
1731 return 0;
1732 }
1733
1734 if (mddev->bitmap == NULL &&
1735 mddev->recovery_cp == MaxSector &&
1736 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1737 conf->fullsync == 0) {
1738 *skipped = 1;
1739 return max_sector - sector_nr;
1740 }
1741 /* before building a request, check if we can skip these blocks..
1742 * This call the bitmap_start_sync doesn't actually record anything
1743 */
1744 if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1745 !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1746 /* We can skip this block, and probably several more */
1747 *skipped = 1;
1748 return sync_blocks;
1749 }
1750 /*
1751 * If there is non-resync activity waiting for a turn,
1752 * and resync is going fast enough,
1753 * then let it though before starting on this new sync request.
1754 */
1755 if (!go_faster && conf->nr_waiting)
1756 msleep_interruptible(1000);
1757
1758 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1759 raise_barrier(conf);
1760
1761 conf->next_resync = sector_nr;
1762
1763 r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1764 rcu_read_lock();
1765 /*
1766 * If we get a correctably read error during resync or recovery,
1767 * we might want to read from a different device. So we
1768 * flag all drives that could conceivably be read from for READ,
1769 * and any others (which will be non-In_sync devices) for WRITE.
1770 * If a read fails, we try reading from something else for which READ
1771 * is OK.
1772 */
1773
1774 r1_bio->mddev = mddev;
1775 r1_bio->sector = sector_nr;
1776 r1_bio->state = 0;
1777 set_bit(R1BIO_IsSync, &r1_bio->state);
1778
1779 for (i=0; i < conf->raid_disks; i++) {
1780 mdk_rdev_t *rdev;
1781 bio = r1_bio->bios[i];
1782
1783 /* take from bio_init */
1784 bio->bi_next = NULL;
1785 bio->bi_flags |= 1 << BIO_UPTODATE;
1786 bio->bi_rw = READ;
1787 bio->bi_vcnt = 0;
1788 bio->bi_idx = 0;
1789 bio->bi_phys_segments = 0;
1790 bio->bi_size = 0;
1791 bio->bi_end_io = NULL;
1792 bio->bi_private = NULL;
1793
1794 rdev = rcu_dereference(conf->mirrors[i].rdev);
1795 if (rdev == NULL ||
1796 test_bit(Faulty, &rdev->flags)) {
1797 still_degraded = 1;
1798 continue;
1799 } else if (!test_bit(In_sync, &rdev->flags)) {
1800 bio->bi_rw = WRITE;
1801 bio->bi_end_io = end_sync_write;
1802 write_targets ++;
1803 } else {
1804 /* may need to read from here */
1805 bio->bi_rw = READ;
1806 bio->bi_end_io = end_sync_read;
1807 if (test_bit(WriteMostly, &rdev->flags)) {
1808 if (wonly < 0)
1809 wonly = i;
1810 } else {
1811 if (disk < 0)
1812 disk = i;
1813 }
1814 read_targets++;
1815 }
1816 atomic_inc(&rdev->nr_pending);
1817 bio->bi_sector = sector_nr + rdev->data_offset;
1818 bio->bi_bdev = rdev->bdev;
1819 bio->bi_private = r1_bio;
1820 }
1821 rcu_read_unlock();
1822 if (disk < 0)
1823 disk = wonly;
1824 r1_bio->read_disk = disk;
1825
1826 if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1827 /* extra read targets are also write targets */
1828 write_targets += read_targets-1;
1829
1830 if (write_targets == 0 || read_targets == 0) {
1831 /* There is nowhere to write, so all non-sync
1832 * drives must be failed - so we are finished
1833 */
1834 sector_t rv = max_sector - sector_nr;
1835 *skipped = 1;
1836 put_buf(r1_bio);
1837 return rv;
1838 }
1839
1840 if (max_sector > mddev->resync_max)
1841 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1842 nr_sectors = 0;
1843 sync_blocks = 0;
1844 do {
1845 struct page *page;
1846 int len = PAGE_SIZE;
1847 if (sector_nr + (len>>9) > max_sector)
1848 len = (max_sector - sector_nr) << 9;
1849 if (len == 0)
1850 break;
1851 if (sync_blocks == 0) {
1852 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1853 &sync_blocks, still_degraded) &&
1854 !conf->fullsync &&
1855 !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1856 break;
1857 BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1858 if (len > (sync_blocks<<9))
1859 len = sync_blocks<<9;
1860 }
1861
1862 for (i=0 ; i < conf->raid_disks; i++) {
1863 bio = r1_bio->bios[i];
1864 if (bio->bi_end_io) {
1865 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1866 if (bio_add_page(bio, page, len, 0) == 0) {
1867 /* stop here */
1868 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1869 while (i > 0) {
1870 i--;
1871 bio = r1_bio->bios[i];
1872 if (bio->bi_end_io==NULL)
1873 continue;
1874 /* remove last page from this bio */
1875 bio->bi_vcnt--;
1876 bio->bi_size -= len;
1877 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1878 }
1879 goto bio_full;
1880 }
1881 }
1882 }
1883 nr_sectors += len>>9;
1884 sector_nr += len>>9;
1885 sync_blocks -= (len>>9);
1886 } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1887 bio_full:
1888 r1_bio->sectors = nr_sectors;
1889
1890 /* For a user-requested sync, we read all readable devices and do a
1891 * compare
1892 */
1893 if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1894 atomic_set(&r1_bio->remaining, read_targets);
1895 for (i=0; i<conf->raid_disks; i++) {
1896 bio = r1_bio->bios[i];
1897 if (bio->bi_end_io == end_sync_read) {
1898 md_sync_acct(bio->bi_bdev, nr_sectors);
1899 generic_make_request(bio);
1900 }
1901 }
1902 } else {
1903 atomic_set(&r1_bio->remaining, 1);
1904 bio = r1_bio->bios[r1_bio->read_disk];
1905 md_sync_acct(bio->bi_bdev, nr_sectors);
1906 generic_make_request(bio);
1907
1908 }
1909 return nr_sectors;
1910 }
1911
1912 static int run(mddev_t *mddev)
1913 {
1914 conf_t *conf;
1915 int i, j, disk_idx;
1916 mirror_info_t *disk;
1917 mdk_rdev_t *rdev;
1918 struct list_head *tmp;
1919
1920 if (mddev->level != 1) {
1921 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1922 mdname(mddev), mddev->level);
1923 goto out;
1924 }
1925 if (mddev->reshape_position != MaxSector) {
1926 printk("raid1: %s: reshape_position set but not supported\n",
1927 mdname(mddev));
1928 goto out;
1929 }
1930 /*
1931 * copy the already verified devices into our private RAID1
1932 * bookkeeping area. [whatever we allocate in run(),
1933 * should be freed in stop()]
1934 */
1935 conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1936 mddev->private = conf;
1937 if (!conf)
1938 goto out_no_mem;
1939
1940 conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1941 GFP_KERNEL);
1942 if (!conf->mirrors)
1943 goto out_no_mem;
1944
1945 conf->tmppage = alloc_page(GFP_KERNEL);
1946 if (!conf->tmppage)
1947 goto out_no_mem;
1948
1949 conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1950 if (!conf->poolinfo)
1951 goto out_no_mem;
1952 conf->poolinfo->mddev = mddev;
1953 conf->poolinfo->raid_disks = mddev->raid_disks;
1954 conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1955 r1bio_pool_free,
1956 conf->poolinfo);
1957 if (!conf->r1bio_pool)
1958 goto out_no_mem;
1959
1960 spin_lock_init(&conf->device_lock);
1961 mddev->queue->queue_lock = &conf->device_lock;
1962
1963 rdev_for_each(rdev, tmp, mddev) {
1964 disk_idx = rdev->raid_disk;
1965 if (disk_idx >= mddev->raid_disks
1966 || disk_idx < 0)
1967 continue;
1968 disk = conf->mirrors + disk_idx;
1969
1970 disk->rdev = rdev;
1971
1972 blk_queue_stack_limits(mddev->queue,
1973 rdev->bdev->bd_disk->queue);
1974 /* as we don't honour merge_bvec_fn, we must never risk
1975 * violating it, so limit ->max_sector to one PAGE, as
1976 * a one page request is never in violation.
1977 */
1978 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1979 mddev->queue->max_sectors > (PAGE_SIZE>>9))
1980 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1981
1982 disk->head_position = 0;
1983 }
1984 conf->raid_disks = mddev->raid_disks;
1985 conf->mddev = mddev;
1986 INIT_LIST_HEAD(&conf->retry_list);
1987
1988 spin_lock_init(&conf->resync_lock);
1989 init_waitqueue_head(&conf->wait_barrier);
1990
1991 bio_list_init(&conf->pending_bio_list);
1992 bio_list_init(&conf->flushing_bio_list);
1993
1994
1995 mddev->degraded = 0;
1996 for (i = 0; i < conf->raid_disks; i++) {
1997
1998 disk = conf->mirrors + i;
1999
2000 if (!disk->rdev ||
2001 !test_bit(In_sync, &disk->rdev->flags)) {
2002 disk->head_position = 0;
2003 mddev->degraded++;
2004 if (disk->rdev)
2005 conf->fullsync = 1;
2006 }
2007 }
2008 if (mddev->degraded == conf->raid_disks) {
2009 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
2010 mdname(mddev));
2011 goto out_free_conf;
2012 }
2013 if (conf->raid_disks - mddev->degraded == 1)
2014 mddev->recovery_cp = MaxSector;
2015
2016 /*
2017 * find the first working one and use it as a starting point
2018 * to read balancing.
2019 */
2020 for (j = 0; j < conf->raid_disks &&
2021 (!conf->mirrors[j].rdev ||
2022 !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
2023 /* nothing */;
2024 conf->last_used = j;
2025
2026
2027 mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
2028 if (!mddev->thread) {
2029 printk(KERN_ERR
2030 "raid1: couldn't allocate thread for %s\n",
2031 mdname(mddev));
2032 goto out_free_conf;
2033 }
2034
2035 printk(KERN_INFO
2036 "raid1: raid set %s active with %d out of %d mirrors\n",
2037 mdname(mddev), mddev->raid_disks - mddev->degraded,
2038 mddev->raid_disks);
2039 /*
2040 * Ok, everything is just fine now
2041 */
2042 mddev->array_sectors = mddev->size * 2;
2043
2044 mddev->queue->unplug_fn = raid1_unplug;
2045 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2046 mddev->queue->backing_dev_info.congested_data = mddev;
2047
2048 return 0;
2049
2050 out_no_mem:
2051 printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
2052 mdname(mddev));
2053
2054 out_free_conf:
2055 if (conf) {
2056 if (conf->r1bio_pool)
2057 mempool_destroy(conf->r1bio_pool);
2058 kfree(conf->mirrors);
2059 safe_put_page(conf->tmppage);
2060 kfree(conf->poolinfo);
2061 kfree(conf);
2062 mddev->private = NULL;
2063 }
2064 out:
2065 return -EIO;
2066 }
2067
2068 static int stop(mddev_t *mddev)
2069 {
2070 conf_t *conf = mddev_to_conf(mddev);
2071 struct bitmap *bitmap = mddev->bitmap;
2072 int behind_wait = 0;
2073
2074 /* wait for behind writes to complete */
2075 while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2076 behind_wait++;
2077 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2078 set_current_state(TASK_UNINTERRUPTIBLE);
2079 schedule_timeout(HZ); /* wait a second */
2080 /* need to kick something here to make sure I/O goes? */
2081 }
2082
2083 md_unregister_thread(mddev->thread);
2084 mddev->thread = NULL;
2085 blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2086 if (conf->r1bio_pool)
2087 mempool_destroy(conf->r1bio_pool);
2088 kfree(conf->mirrors);
2089 kfree(conf->poolinfo);
2090 kfree(conf);
2091 mddev->private = NULL;
2092 return 0;
2093 }
2094
2095 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2096 {
2097 /* no resync is happening, and there is enough space
2098 * on all devices, so we can resize.
2099 * We need to make sure resync covers any new space.
2100 * If the array is shrinking we should possibly wait until
2101 * any io in the removed space completes, but it hardly seems
2102 * worth it.
2103 */
2104 mddev->array_sectors = sectors;
2105 set_capacity(mddev->gendisk, mddev->array_sectors);
2106 mddev->changed = 1;
2107 if (mddev->array_sectors / 2 > mddev->size &&
2108 mddev->recovery_cp == MaxSector) {
2109 mddev->recovery_cp = mddev->size << 1;
2110 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2111 }
2112 mddev->size = mddev->array_sectors / 2;
2113 mddev->resync_max_sectors = sectors;
2114 return 0;
2115 }
2116
2117 static int raid1_reshape(mddev_t *mddev)
2118 {
2119 /* We need to:
2120 * 1/ resize the r1bio_pool
2121 * 2/ resize conf->mirrors
2122 *
2123 * We allocate a new r1bio_pool if we can.
2124 * Then raise a device barrier and wait until all IO stops.
2125 * Then resize conf->mirrors and swap in the new r1bio pool.
2126 *
2127 * At the same time, we "pack" the devices so that all the missing
2128 * devices have the higher raid_disk numbers.
2129 */
2130 mempool_t *newpool, *oldpool;
2131 struct pool_info *newpoolinfo;
2132 mirror_info_t *newmirrors;
2133 conf_t *conf = mddev_to_conf(mddev);
2134 int cnt, raid_disks;
2135 unsigned long flags;
2136 int d, d2, err;
2137
2138 /* Cannot change chunk_size, layout, or level */
2139 if (mddev->chunk_size != mddev->new_chunk ||
2140 mddev->layout != mddev->new_layout ||
2141 mddev->level != mddev->new_level) {
2142 mddev->new_chunk = mddev->chunk_size;
2143 mddev->new_layout = mddev->layout;
2144 mddev->new_level = mddev->level;
2145 return -EINVAL;
2146 }
2147
2148 err = md_allow_write(mddev);
2149 if (err)
2150 return err;
2151
2152 raid_disks = mddev->raid_disks + mddev->delta_disks;
2153
2154 if (raid_disks < conf->raid_disks) {
2155 cnt=0;
2156 for (d= 0; d < conf->raid_disks; d++)
2157 if (conf->mirrors[d].rdev)
2158 cnt++;
2159 if (cnt > raid_disks)
2160 return -EBUSY;
2161 }
2162
2163 newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2164 if (!newpoolinfo)
2165 return -ENOMEM;
2166 newpoolinfo->mddev = mddev;
2167 newpoolinfo->raid_disks = raid_disks;
2168
2169 newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2170 r1bio_pool_free, newpoolinfo);
2171 if (!newpool) {
2172 kfree(newpoolinfo);
2173 return -ENOMEM;
2174 }
2175 newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2176 if (!newmirrors) {
2177 kfree(newpoolinfo);
2178 mempool_destroy(newpool);
2179 return -ENOMEM;
2180 }
2181
2182 raise_barrier(conf);
2183
2184 /* ok, everything is stopped */
2185 oldpool = conf->r1bio_pool;
2186 conf->r1bio_pool = newpool;
2187
2188 for (d = d2 = 0; d < conf->raid_disks; d++) {
2189 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2190 if (rdev && rdev->raid_disk != d2) {
2191 char nm[20];
2192 sprintf(nm, "rd%d", rdev->raid_disk);
2193 sysfs_remove_link(&mddev->kobj, nm);
2194 rdev->raid_disk = d2;
2195 sprintf(nm, "rd%d", rdev->raid_disk);
2196 sysfs_remove_link(&mddev->kobj, nm);
2197 if (sysfs_create_link(&mddev->kobj,
2198 &rdev->kobj, nm))
2199 printk(KERN_WARNING
2200 "md/raid1: cannot register "
2201 "%s for %s\n",
2202 nm, mdname(mddev));
2203 }
2204 if (rdev)
2205 newmirrors[d2++].rdev = rdev;
2206 }
2207 kfree(conf->mirrors);
2208 conf->mirrors = newmirrors;
2209 kfree(conf->poolinfo);
2210 conf->poolinfo = newpoolinfo;
2211
2212 spin_lock_irqsave(&conf->device_lock, flags);
2213 mddev->degraded += (raid_disks - conf->raid_disks);
2214 spin_unlock_irqrestore(&conf->device_lock, flags);
2215 conf->raid_disks = mddev->raid_disks = raid_disks;
2216 mddev->delta_disks = 0;
2217
2218 conf->last_used = 0; /* just make sure it is in-range */
2219 lower_barrier(conf);
2220
2221 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2222 md_wakeup_thread(mddev->thread);
2223
2224 mempool_destroy(oldpool);
2225 return 0;
2226 }
2227
2228 static void raid1_quiesce(mddev_t *mddev, int state)
2229 {
2230 conf_t *conf = mddev_to_conf(mddev);
2231
2232 switch(state) {
2233 case 1:
2234 raise_barrier(conf);
2235 break;
2236 case 0:
2237 lower_barrier(conf);
2238 break;
2239 }
2240 }
2241
2242
2243 static struct mdk_personality raid1_personality =
2244 {
2245 .name = "raid1",
2246 .level = 1,
2247 .owner = THIS_MODULE,
2248 .make_request = make_request,
2249 .run = run,
2250 .stop = stop,
2251 .status = status,
2252 .error_handler = error,
2253 .hot_add_disk = raid1_add_disk,
2254 .hot_remove_disk= raid1_remove_disk,
2255 .spare_active = raid1_spare_active,
2256 .sync_request = sync_request,
2257 .resize = raid1_resize,
2258 .check_reshape = raid1_reshape,
2259 .quiesce = raid1_quiesce,
2260 };
2261
2262 static int __init raid_init(void)
2263 {
2264 return register_md_personality(&raid1_personality);
2265 }
2266
2267 static void raid_exit(void)
2268 {
2269 unregister_md_personality(&raid1_personality);
2270 }
2271
2272 module_init(raid_init);
2273 module_exit(raid_exit);
2274 MODULE_LICENSE("GPL");
2275 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2276 MODULE_ALIAS("md-raid1");
2277 MODULE_ALIAS("md-level-1");
Linux kernel - Deliverables
This page took 0.133853 seconds and 5 git commands to generate.