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, bio->bi_size,
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 int raid1_end_read_request(struct bio *bio, unsigned int bytes_done, 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 if (bio->bi_size)
266 return 1;
267
268 mirror = r1_bio->read_disk;
269 /*
270 * this branch is our 'one mirror IO has finished' event handler:
271 */
272 update_head_pos(mirror, r1_bio);
273
274 if (uptodate || conf->working_disks <= 1) {
275 /*
276 * Set R1BIO_Uptodate in our master bio, so that
277 * we will return a good error code for to the higher
278 * levels even if IO on some other mirrored buffer fails.
279 *
280 * The 'master' represents the composite IO operation to
281 * user-side. So if something waits for IO, then it will
282 * wait for the 'master' bio.
283 */
284 if (uptodate)
285 set_bit(R1BIO_Uptodate, &r1_bio->state);
286
287 raid_end_bio_io(r1_bio);
288 } else {
289 /*
290 * oops, read error:
291 */
292 char b[BDEVNAME_SIZE];
293 if (printk_ratelimit())
294 printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
295 bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
296 reschedule_retry(r1_bio);
297 }
298
299 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
300 return 0;
301 }
302
303 static int raid1_end_write_request(struct bio *bio, unsigned int bytes_done, 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 if (bio->bi_size)
312 return 1;
313
314 for (mirror = 0; mirror < conf->raid_disks; mirror++)
315 if (r1_bio->bios[mirror] == bio)
316 break;
317
318 if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
319 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
320 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
321 r1_bio->mddev->barriers_work = 0;
322 /* Don't rdev_dec_pending in this branch - keep it for the retry */
323 } else {
324 /*
325 * this branch is our 'one mirror IO has finished' event handler:
326 */
327 r1_bio->bios[mirror] = NULL;
328 to_put = bio;
329 if (!uptodate) {
330 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
331 /* an I/O failed, we can't clear the bitmap */
332 set_bit(R1BIO_Degraded, &r1_bio->state);
333 } else
334 /*
335 * Set R1BIO_Uptodate in our master bio, so that
336 * we will return a good error code for to the higher
337 * levels even if IO on some other mirrored buffer fails.
338 *
339 * The 'master' represents the composite IO operation to
340 * user-side. So if something waits for IO, then it will
341 * wait for the 'master' bio.
342 */
343 set_bit(R1BIO_Uptodate, &r1_bio->state);
344
345 update_head_pos(mirror, r1_bio);
346
347 if (behind) {
348 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
349 atomic_dec(&r1_bio->behind_remaining);
350
351 /* In behind mode, we ACK the master bio once the I/O has safely
352 * reached all non-writemostly disks. Setting the Returned bit
353 * ensures that this gets done only once -- we don't ever want to
354 * return -EIO here, instead we'll wait */
355
356 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
357 test_bit(R1BIO_Uptodate, &r1_bio->state)) {
358 /* Maybe we can return now */
359 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
360 struct bio *mbio = r1_bio->master_bio;
361 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
362 (unsigned long long) mbio->bi_sector,
363 (unsigned long long) mbio->bi_sector +
364 (mbio->bi_size >> 9) - 1);
365 bio_endio(mbio, mbio->bi_size, 0);
366 }
367 }
368 }
369 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
370 }
371 /*
372 *
373 * Let's see if all mirrored write operations have finished
374 * already.
375 */
376 if (atomic_dec_and_test(&r1_bio->remaining)) {
377 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
378 reschedule_retry(r1_bio);
379 goto out;
380 }
381 /* it really is the end of this request */
382 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
383 /* free extra copy of the data pages */
384 int i = bio->bi_vcnt;
385 while (i--)
386 safe_put_page(bio->bi_io_vec[i].bv_page);
387 }
388 /* clear the bitmap if all writes complete successfully */
389 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
390 r1_bio->sectors,
391 !test_bit(R1BIO_Degraded, &r1_bio->state),
392 behind);
393 md_write_end(r1_bio->mddev);
394 raid_end_bio_io(r1_bio);
395 }
396 out:
397 if (to_put)
398 bio_put(to_put);
399
400 return 0;
401 }
402
403
404 /*
405 * This routine returns the disk from which the requested read should
406 * be done. There is a per-array 'next expected sequential IO' sector
407 * number - if this matches on the next IO then we use the last disk.
408 * There is also a per-disk 'last know head position' sector that is
409 * maintained from IRQ contexts, both the normal and the resync IO
410 * completion handlers update this position correctly. If there is no
411 * perfect sequential match then we pick the disk whose head is closest.
412 *
413 * If there are 2 mirrors in the same 2 devices, performance degrades
414 * because position is mirror, not device based.
415 *
416 * The rdev for the device selected will have nr_pending incremented.
417 */
418 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
419 {
420 const unsigned long this_sector = r1_bio->sector;
421 int new_disk = conf->last_used, disk = new_disk;
422 int wonly_disk = -1;
423 const int sectors = r1_bio->sectors;
424 sector_t new_distance, current_distance;
425 mdk_rdev_t *rdev;
426
427 rcu_read_lock();
428 /*
429 * Check if we can balance. We can balance on the whole
430 * device if no resync is going on, or below the resync window.
431 * We take the first readable disk when above the resync window.
432 */
433 retry:
434 if (conf->mddev->recovery_cp < MaxSector &&
435 (this_sector + sectors >= conf->next_resync)) {
436 /* Choose the first operation device, for consistancy */
437 new_disk = 0;
438
439 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
440 r1_bio->bios[new_disk] == IO_BLOCKED ||
441 !rdev || !test_bit(In_sync, &rdev->flags)
442 || test_bit(WriteMostly, &rdev->flags);
443 rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
444
445 if (rdev && test_bit(In_sync, &rdev->flags) &&
446 r1_bio->bios[new_disk] != IO_BLOCKED)
447 wonly_disk = new_disk;
448
449 if (new_disk == conf->raid_disks - 1) {
450 new_disk = wonly_disk;
451 break;
452 }
453 }
454 goto rb_out;
455 }
456
457
458 /* make sure the disk is operational */
459 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
460 r1_bio->bios[new_disk] == IO_BLOCKED ||
461 !rdev || !test_bit(In_sync, &rdev->flags) ||
462 test_bit(WriteMostly, &rdev->flags);
463 rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
464
465 if (rdev && test_bit(In_sync, &rdev->flags) &&
466 r1_bio->bios[new_disk] != IO_BLOCKED)
467 wonly_disk = new_disk;
468
469 if (new_disk <= 0)
470 new_disk = conf->raid_disks;
471 new_disk--;
472 if (new_disk == disk) {
473 new_disk = wonly_disk;
474 break;
475 }
476 }
477
478 if (new_disk < 0)
479 goto rb_out;
480
481 disk = new_disk;
482 /* now disk == new_disk == starting point for search */
483
484 /*
485 * Don't change to another disk for sequential reads:
486 */
487 if (conf->next_seq_sect == this_sector)
488 goto rb_out;
489 if (this_sector == conf->mirrors[new_disk].head_position)
490 goto rb_out;
491
492 current_distance = abs(this_sector - conf->mirrors[disk].head_position);
493
494 /* Find the disk whose head is closest */
495
496 do {
497 if (disk <= 0)
498 disk = conf->raid_disks;
499 disk--;
500
501 rdev = rcu_dereference(conf->mirrors[disk].rdev);
502
503 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
504 !test_bit(In_sync, &rdev->flags) ||
505 test_bit(WriteMostly, &rdev->flags))
506 continue;
507
508 if (!atomic_read(&rdev->nr_pending)) {
509 new_disk = disk;
510 break;
511 }
512 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
513 if (new_distance < current_distance) {
514 current_distance = new_distance;
515 new_disk = disk;
516 }
517 } while (disk != conf->last_used);
518
519 rb_out:
520
521
522 if (new_disk >= 0) {
523 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
524 if (!rdev)
525 goto retry;
526 atomic_inc(&rdev->nr_pending);
527 if (!test_bit(In_sync, &rdev->flags)) {
528 /* cannot risk returning a device that failed
529 * before we inc'ed nr_pending
530 */
531 rdev_dec_pending(rdev, conf->mddev);
532 goto retry;
533 }
534 conf->next_seq_sect = this_sector + sectors;
535 conf->last_used = new_disk;
536 }
537 rcu_read_unlock();
538
539 return new_disk;
540 }
541
542 static void unplug_slaves(mddev_t *mddev)
543 {
544 conf_t *conf = mddev_to_conf(mddev);
545 int i;
546
547 rcu_read_lock();
548 for (i=0; i<mddev->raid_disks; i++) {
549 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
550 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
551 request_queue_t *r_queue = bdev_get_queue(rdev->bdev);
552
553 atomic_inc(&rdev->nr_pending);
554 rcu_read_unlock();
555
556 if (r_queue->unplug_fn)
557 r_queue->unplug_fn(r_queue);
558
559 rdev_dec_pending(rdev, mddev);
560 rcu_read_lock();
561 }
562 }
563 rcu_read_unlock();
564 }
565
566 static void raid1_unplug(request_queue_t *q)
567 {
568 mddev_t *mddev = q->queuedata;
569
570 unplug_slaves(mddev);
571 md_wakeup_thread(mddev->thread);
572 }
573
574 static int raid1_issue_flush(request_queue_t *q, struct gendisk *disk,
575 sector_t *error_sector)
576 {
577 mddev_t *mddev = q->queuedata;
578 conf_t *conf = mddev_to_conf(mddev);
579 int i, ret = 0;
580
581 rcu_read_lock();
582 for (i=0; i<mddev->raid_disks && ret == 0; i++) {
583 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
584 if (rdev && !test_bit(Faulty, &rdev->flags)) {
585 struct block_device *bdev = rdev->bdev;
586 request_queue_t *r_queue = bdev_get_queue(bdev);
587
588 if (!r_queue->issue_flush_fn)
589 ret = -EOPNOTSUPP;
590 else {
591 atomic_inc(&rdev->nr_pending);
592 rcu_read_unlock();
593 ret = r_queue->issue_flush_fn(r_queue, bdev->bd_disk,
594 error_sector);
595 rdev_dec_pending(rdev, mddev);
596 rcu_read_lock();
597 }
598 }
599 }
600 rcu_read_unlock();
601 return ret;
602 }
603
604 /* Barriers....
605 * Sometimes we need to suspend IO while we do something else,
606 * either some resync/recovery, or reconfigure the array.
607 * To do this we raise a 'barrier'.
608 * The 'barrier' is a counter that can be raised multiple times
609 * to count how many activities are happening which preclude
610 * normal IO.
611 * We can only raise the barrier if there is no pending IO.
612 * i.e. if nr_pending == 0.
613 * We choose only to raise the barrier if no-one is waiting for the
614 * barrier to go down. This means that as soon as an IO request
615 * is ready, no other operations which require a barrier will start
616 * until the IO request has had a chance.
617 *
618 * So: regular IO calls 'wait_barrier'. When that returns there
619 * is no backgroup IO happening, It must arrange to call
620 * allow_barrier when it has finished its IO.
621 * backgroup IO calls must call raise_barrier. Once that returns
622 * there is no normal IO happeing. It must arrange to call
623 * lower_barrier when the particular background IO completes.
624 */
625 #define RESYNC_DEPTH 32
626
627 static void raise_barrier(conf_t *conf)
628 {
629 spin_lock_irq(&conf->resync_lock);
630
631 /* Wait until no block IO is waiting */
632 wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
633 conf->resync_lock,
634 raid1_unplug(conf->mddev->queue));
635
636 /* block any new IO from starting */
637 conf->barrier++;
638
639 /* No wait for all pending IO to complete */
640 wait_event_lock_irq(conf->wait_barrier,
641 !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
642 conf->resync_lock,
643 raid1_unplug(conf->mddev->queue));
644
645 spin_unlock_irq(&conf->resync_lock);
646 }
647
648 static void lower_barrier(conf_t *conf)
649 {
650 unsigned long flags;
651 spin_lock_irqsave(&conf->resync_lock, flags);
652 conf->barrier--;
653 spin_unlock_irqrestore(&conf->resync_lock, flags);
654 wake_up(&conf->wait_barrier);
655 }
656
657 static void wait_barrier(conf_t *conf)
658 {
659 spin_lock_irq(&conf->resync_lock);
660 if (conf->barrier) {
661 conf->nr_waiting++;
662 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
663 conf->resync_lock,
664 raid1_unplug(conf->mddev->queue));
665 conf->nr_waiting--;
666 }
667 conf->nr_pending++;
668 spin_unlock_irq(&conf->resync_lock);
669 }
670
671 static void allow_barrier(conf_t *conf)
672 {
673 unsigned long flags;
674 spin_lock_irqsave(&conf->resync_lock, flags);
675 conf->nr_pending--;
676 spin_unlock_irqrestore(&conf->resync_lock, flags);
677 wake_up(&conf->wait_barrier);
678 }
679
680 static void freeze_array(conf_t *conf)
681 {
682 /* stop syncio and normal IO and wait for everything to
683 * go quite.
684 * We increment barrier and nr_waiting, and then
685 * wait until barrier+nr_pending match nr_queued+2
686 */
687 spin_lock_irq(&conf->resync_lock);
688 conf->barrier++;
689 conf->nr_waiting++;
690 wait_event_lock_irq(conf->wait_barrier,
691 conf->barrier+conf->nr_pending == conf->nr_queued+2,
692 conf->resync_lock,
693 raid1_unplug(conf->mddev->queue));
694 spin_unlock_irq(&conf->resync_lock);
695 }
696 static void unfreeze_array(conf_t *conf)
697 {
698 /* reverse the effect of the freeze */
699 spin_lock_irq(&conf->resync_lock);
700 conf->barrier--;
701 conf->nr_waiting--;
702 wake_up(&conf->wait_barrier);
703 spin_unlock_irq(&conf->resync_lock);
704 }
705
706
707 /* duplicate the data pages for behind I/O */
708 static struct page **alloc_behind_pages(struct bio *bio)
709 {
710 int i;
711 struct bio_vec *bvec;
712 struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
713 GFP_NOIO);
714 if (unlikely(!pages))
715 goto do_sync_io;
716
717 bio_for_each_segment(bvec, bio, i) {
718 pages[i] = alloc_page(GFP_NOIO);
719 if (unlikely(!pages[i]))
720 goto do_sync_io;
721 memcpy(kmap(pages[i]) + bvec->bv_offset,
722 kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
723 kunmap(pages[i]);
724 kunmap(bvec->bv_page);
725 }
726
727 return pages;
728
729 do_sync_io:
730 if (pages)
731 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
732 put_page(pages[i]);
733 kfree(pages);
734 PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
735 return NULL;
736 }
737
738 static int make_request(request_queue_t *q, struct bio * bio)
739 {
740 mddev_t *mddev = q->queuedata;
741 conf_t *conf = mddev_to_conf(mddev);
742 mirror_info_t *mirror;
743 r1bio_t *r1_bio;
744 struct bio *read_bio;
745 int i, targets = 0, disks;
746 mdk_rdev_t *rdev;
747 struct bitmap *bitmap = mddev->bitmap;
748 unsigned long flags;
749 struct bio_list bl;
750 struct page **behind_pages = NULL;
751 const int rw = bio_data_dir(bio);
752 int do_barriers;
753
754 /*
755 * Register the new request and wait if the reconstruction
756 * thread has put up a bar for new requests.
757 * Continue immediately if no resync is active currently.
758 * We test barriers_work *after* md_write_start as md_write_start
759 * may cause the first superblock write, and that will check out
760 * if barriers work.
761 */
762
763 md_write_start(mddev, bio); /* wait on superblock update early */
764
765 if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
766 if (rw == WRITE)
767 md_write_end(mddev);
768 bio_endio(bio, bio->bi_size, -EOPNOTSUPP);
769 return 0;
770 }
771
772 wait_barrier(conf);
773
774 disk_stat_inc(mddev->gendisk, ios[rw]);
775 disk_stat_add(mddev->gendisk, sectors[rw], bio_sectors(bio));
776
777 /*
778 * make_request() can abort the operation when READA is being
779 * used and no empty request is available.
780 *
781 */
782 r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
783
784 r1_bio->master_bio = bio;
785 r1_bio->sectors = bio->bi_size >> 9;
786 r1_bio->state = 0;
787 r1_bio->mddev = mddev;
788 r1_bio->sector = bio->bi_sector;
789
790 if (rw == READ) {
791 /*
792 * read balancing logic:
793 */
794 int rdisk = read_balance(conf, r1_bio);
795
796 if (rdisk < 0) {
797 /* couldn't find anywhere to read from */
798 raid_end_bio_io(r1_bio);
799 return 0;
800 }
801 mirror = conf->mirrors + rdisk;
802
803 r1_bio->read_disk = rdisk;
804
805 read_bio = bio_clone(bio, GFP_NOIO);
806
807 r1_bio->bios[rdisk] = read_bio;
808
809 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
810 read_bio->bi_bdev = mirror->rdev->bdev;
811 read_bio->bi_end_io = raid1_end_read_request;
812 read_bio->bi_rw = READ;
813 read_bio->bi_private = r1_bio;
814
815 generic_make_request(read_bio);
816 return 0;
817 }
818
819 /*
820 * WRITE:
821 */
822 /* first select target devices under spinlock and
823 * inc refcount on their rdev. Record them by setting
824 * bios[x] to bio
825 */
826 disks = conf->raid_disks;
827 #if 0
828 { static int first=1;
829 if (first) printk("First Write sector %llu disks %d\n",
830 (unsigned long long)r1_bio->sector, disks);
831 first = 0;
832 }
833 #endif
834 rcu_read_lock();
835 for (i = 0; i < disks; i++) {
836 if ((rdev=rcu_dereference(conf->mirrors[i].rdev)) != NULL &&
837 !test_bit(Faulty, &rdev->flags)) {
838 atomic_inc(&rdev->nr_pending);
839 if (test_bit(Faulty, &rdev->flags)) {
840 rdev_dec_pending(rdev, mddev);
841 r1_bio->bios[i] = NULL;
842 } else
843 r1_bio->bios[i] = bio;
844 targets++;
845 } else
846 r1_bio->bios[i] = NULL;
847 }
848 rcu_read_unlock();
849
850 BUG_ON(targets == 0); /* we never fail the last device */
851
852 if (targets < conf->raid_disks) {
853 /* array is degraded, we will not clear the bitmap
854 * on I/O completion (see raid1_end_write_request) */
855 set_bit(R1BIO_Degraded, &r1_bio->state);
856 }
857
858 /* do behind I/O ? */
859 if (bitmap &&
860 atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
861 (behind_pages = alloc_behind_pages(bio)) != NULL)
862 set_bit(R1BIO_BehindIO, &r1_bio->state);
863
864 atomic_set(&r1_bio->remaining, 0);
865 atomic_set(&r1_bio->behind_remaining, 0);
866
867 do_barriers = bio_barrier(bio);
868 if (do_barriers)
869 set_bit(R1BIO_Barrier, &r1_bio->state);
870
871 bio_list_init(&bl);
872 for (i = 0; i < disks; i++) {
873 struct bio *mbio;
874 if (!r1_bio->bios[i])
875 continue;
876
877 mbio = bio_clone(bio, GFP_NOIO);
878 r1_bio->bios[i] = mbio;
879
880 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
881 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
882 mbio->bi_end_io = raid1_end_write_request;
883 mbio->bi_rw = WRITE | do_barriers;
884 mbio->bi_private = r1_bio;
885
886 if (behind_pages) {
887 struct bio_vec *bvec;
888 int j;
889
890 /* Yes, I really want the '__' version so that
891 * we clear any unused pointer in the io_vec, rather
892 * than leave them unchanged. This is important
893 * because when we come to free the pages, we won't
894 * know the originial bi_idx, so we just free
895 * them all
896 */
897 __bio_for_each_segment(bvec, mbio, j, 0)
898 bvec->bv_page = behind_pages[j];
899 if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
900 atomic_inc(&r1_bio->behind_remaining);
901 }
902
903 atomic_inc(&r1_bio->remaining);
904
905 bio_list_add(&bl, mbio);
906 }
907 kfree(behind_pages); /* the behind pages are attached to the bios now */
908
909 bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
910 test_bit(R1BIO_BehindIO, &r1_bio->state));
911 spin_lock_irqsave(&conf->device_lock, flags);
912 bio_list_merge(&conf->pending_bio_list, &bl);
913 bio_list_init(&bl);
914
915 blk_plug_device(mddev->queue);
916 spin_unlock_irqrestore(&conf->device_lock, flags);
917
918 #if 0
919 while ((bio = bio_list_pop(&bl)) != NULL)
920 generic_make_request(bio);
921 #endif
922
923 return 0;
924 }
925
926 static void status(struct seq_file *seq, mddev_t *mddev)
927 {
928 conf_t *conf = mddev_to_conf(mddev);
929 int i;
930
931 seq_printf(seq, " [%d/%d] [", conf->raid_disks,
932 conf->working_disks);
933 for (i = 0; i < conf->raid_disks; i++)
934 seq_printf(seq, "%s",
935 conf->mirrors[i].rdev &&
936 test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
937 seq_printf(seq, "]");
938 }
939
940
941 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
942 {
943 char b[BDEVNAME_SIZE];
944 conf_t *conf = mddev_to_conf(mddev);
945
946 /*
947 * If it is not operational, then we have already marked it as dead
948 * else if it is the last working disks, ignore the error, let the
949 * next level up know.
950 * else mark the drive as failed
951 */
952 if (test_bit(In_sync, &rdev->flags)
953 && conf->working_disks == 1)
954 /*
955 * Don't fail the drive, act as though we were just a
956 * normal single drive
957 */
958 return;
959 if (test_bit(In_sync, &rdev->flags)) {
960 mddev->degraded++;
961 conf->working_disks--;
962 /*
963 * if recovery is running, make sure it aborts.
964 */
965 set_bit(MD_RECOVERY_ERR, &mddev->recovery);
966 }
967 clear_bit(In_sync, &rdev->flags);
968 set_bit(Faulty, &rdev->flags);
969 mddev->sb_dirty = 1;
970 printk(KERN_ALERT "raid1: Disk failure on %s, disabling device. \n"
971 " Operation continuing on %d devices\n",
972 bdevname(rdev->bdev,b), conf->working_disks);
973 }
974
975 static void print_conf(conf_t *conf)
976 {
977 int i;
978 mirror_info_t *tmp;
979
980 printk("RAID1 conf printout:\n");
981 if (!conf) {
982 printk("(!conf)\n");
983 return;
984 }
985 printk(" --- wd:%d rd:%d\n", conf->working_disks,
986 conf->raid_disks);
987
988 for (i = 0; i < conf->raid_disks; i++) {
989 char b[BDEVNAME_SIZE];
990 tmp = conf->mirrors + i;
991 if (tmp->rdev)
992 printk(" disk %d, wo:%d, o:%d, dev:%s\n",
993 i, !test_bit(In_sync, &tmp->rdev->flags), !test_bit(Faulty, &tmp->rdev->flags),
994 bdevname(tmp->rdev->bdev,b));
995 }
996 }
997
998 static void close_sync(conf_t *conf)
999 {
1000 wait_barrier(conf);