Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (C) 2011 STRATO. All rights reserved.
4 */
5
6#include <linux/sched.h>
7#include <linux/pagemap.h>
8#include <linux/writeback.h>
9#include <linux/blkdev.h>
10#include <linux/slab.h>
11#include <linux/workqueue.h>
12#include "ctree.h"
13#include "volumes.h"
14#include "disk-io.h"
15#include "transaction.h"
16#include "dev-replace.h"
17#include "block-group.h"
18
19#undef DEBUG
20
21/*
22 * This is the implementation for the generic read ahead framework.
23 *
24 * To trigger a readahead, btrfs_reada_add must be called. It will start
25 * a read ahead for the given range [start, end) on tree root. The returned
26 * handle can either be used to wait on the readahead to finish
27 * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).
28 *
29 * The read ahead works as follows:
30 * On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
31 * reada_start_machine will then search for extents to prefetch and trigger
32 * some reads. When a read finishes for a node, all contained node/leaf
33 * pointers that lie in the given range will also be enqueued. The reads will
34 * be triggered in sequential order, thus giving a big win over a naive
35 * enumeration. It will also make use of multi-device layouts. Each disk
36 * will have its on read pointer and all disks will by utilized in parallel.
37 * Also will no two disks read both sides of a mirror simultaneously, as this
38 * would waste seeking capacity. Instead both disks will read different parts
39 * of the filesystem.
40 * Any number of readaheads can be started in parallel. The read order will be
41 * determined globally, i.e. 2 parallel readaheads will normally finish faster
42 * than the 2 started one after another.
43 */
44
45#define MAX_IN_FLIGHT 6
46
47struct reada_extctl {
48 struct list_head list;
49 struct reada_control *rc;
50 u64 generation;
51};
52
53struct reada_extent {
54 u64 logical;
55 struct btrfs_key top;
56 struct list_head extctl;
57 int refcnt;
58 spinlock_t lock;
59 struct reada_zone *zones[BTRFS_MAX_MIRRORS];
60 int nzones;
61 int scheduled;
62};
63
64struct reada_zone {
65 u64 start;
66 u64 end;
67 u64 elems;
68 struct list_head list;
69 spinlock_t lock;
70 int locked;
71 struct btrfs_device *device;
72 struct btrfs_device *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
73 * self */
74 int ndevs;
75 struct kref refcnt;
76};
77
78struct reada_machine_work {
79 struct btrfs_work work;
80 struct btrfs_fs_info *fs_info;
81};
82
83static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
84static void reada_control_release(struct kref *kref);
85static void reada_zone_release(struct kref *kref);
86static void reada_start_machine(struct btrfs_fs_info *fs_info);
87static void __reada_start_machine(struct btrfs_fs_info *fs_info);
88
89static int reada_add_block(struct reada_control *rc, u64 logical,
90 struct btrfs_key *top, u64 generation);
91
92/* recurses */
93/* in case of err, eb might be NULL */
94static void __readahead_hook(struct btrfs_fs_info *fs_info,
95 struct reada_extent *re, struct extent_buffer *eb,
96 int err)
97{
98 int nritems;
99 int i;
100 u64 bytenr;
101 u64 generation;
102 struct list_head list;
103
104 spin_lock(&re->lock);
105 /*
106 * just take the full list from the extent. afterwards we
107 * don't need the lock anymore
108 */
109 list_replace_init(&re->extctl, &list);
110 re->scheduled = 0;
111 spin_unlock(&re->lock);
112
113 /*
114 * this is the error case, the extent buffer has not been
115 * read correctly. We won't access anything from it and
116 * just cleanup our data structures. Effectively this will
117 * cut the branch below this node from read ahead.
118 */
119 if (err)
120 goto cleanup;
121
122 /*
123 * FIXME: currently we just set nritems to 0 if this is a leaf,
124 * effectively ignoring the content. In a next step we could
125 * trigger more readahead depending from the content, e.g.
126 * fetch the checksums for the extents in the leaf.
127 */
128 if (!btrfs_header_level(eb))
129 goto cleanup;
130
131 nritems = btrfs_header_nritems(eb);
132 generation = btrfs_header_generation(eb);
133 for (i = 0; i < nritems; i++) {
134 struct reada_extctl *rec;
135 u64 n_gen;
136 struct btrfs_key key;
137 struct btrfs_key next_key;
138
139 btrfs_node_key_to_cpu(eb, &key, i);
140 if (i + 1 < nritems)
141 btrfs_node_key_to_cpu(eb, &next_key, i + 1);
142 else
143 next_key = re->top;
144 bytenr = btrfs_node_blockptr(eb, i);
145 n_gen = btrfs_node_ptr_generation(eb, i);
146
147 list_for_each_entry(rec, &list, list) {
148 struct reada_control *rc = rec->rc;
149
150 /*
151 * if the generation doesn't match, just ignore this
152 * extctl. This will probably cut off a branch from
153 * prefetch. Alternatively one could start a new (sub-)
154 * prefetch for this branch, starting again from root.
155 * FIXME: move the generation check out of this loop
156 */
157#ifdef DEBUG
158 if (rec->generation != generation) {
159 btrfs_debug(fs_info,
160 "generation mismatch for (%llu,%d,%llu) %llu != %llu",
161 key.objectid, key.type, key.offset,
162 rec->generation, generation);
163 }
164#endif
165 if (rec->generation == generation &&
166 btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
167 btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
168 reada_add_block(rc, bytenr, &next_key, n_gen);
169 }
170 }
171
172cleanup:
173 /*
174 * free extctl records
175 */
176 while (!list_empty(&list)) {
177 struct reada_control *rc;
178 struct reada_extctl *rec;
179
180 rec = list_first_entry(&list, struct reada_extctl, list);
181 list_del(&rec->list);
182 rc = rec->rc;
183 kfree(rec);
184
185 kref_get(&rc->refcnt);
186 if (atomic_dec_and_test(&rc->elems)) {
187 kref_put(&rc->refcnt, reada_control_release);
188 wake_up(&rc->wait);
189 }
190 kref_put(&rc->refcnt, reada_control_release);
191
192 reada_extent_put(fs_info, re); /* one ref for each entry */
193 }
194
195 return;
196}
197
198int btree_readahead_hook(struct extent_buffer *eb, int err)
199{
200 struct btrfs_fs_info *fs_info = eb->fs_info;
201 int ret = 0;
202 struct reada_extent *re;
203
204 /* find extent */
205 spin_lock(&fs_info->reada_lock);
206 re = radix_tree_lookup(&fs_info->reada_tree,
207 eb->start >> PAGE_SHIFT);
208 if (re)
209 re->refcnt++;
210 spin_unlock(&fs_info->reada_lock);
211 if (!re) {
212 ret = -1;
213 goto start_machine;
214 }
215
216 __readahead_hook(fs_info, re, eb, err);
217 reada_extent_put(fs_info, re); /* our ref */
218
219start_machine:
220 reada_start_machine(fs_info);
221 return ret;
222}
223
224static struct reada_zone *reada_find_zone(struct btrfs_device *dev, u64 logical,
225 struct btrfs_bio *bbio)
226{
227 struct btrfs_fs_info *fs_info = dev->fs_info;
228 int ret;
229 struct reada_zone *zone;
230 struct btrfs_block_group_cache *cache = NULL;
231 u64 start;
232 u64 end;
233 int i;
234
235 zone = NULL;
236 spin_lock(&fs_info->reada_lock);
237 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
238 logical >> PAGE_SHIFT, 1);
239 if (ret == 1 && logical >= zone->start && logical <= zone->end) {
240 kref_get(&zone->refcnt);
241 spin_unlock(&fs_info->reada_lock);
242 return zone;
243 }
244
245 spin_unlock(&fs_info->reada_lock);
246
247 cache = btrfs_lookup_block_group(fs_info, logical);
248 if (!cache)
249 return NULL;
250
251 start = cache->key.objectid;
252 end = start + cache->key.offset - 1;
253 btrfs_put_block_group(cache);
254
255 zone = kzalloc(sizeof(*zone), GFP_KERNEL);
256 if (!zone)
257 return NULL;
258
259 ret = radix_tree_preload(GFP_KERNEL);
260 if (ret) {
261 kfree(zone);
262 return NULL;
263 }
264
265 zone->start = start;
266 zone->end = end;
267 INIT_LIST_HEAD(&zone->list);
268 spin_lock_init(&zone->lock);
269 zone->locked = 0;
270 kref_init(&zone->refcnt);
271 zone->elems = 0;
272 zone->device = dev; /* our device always sits at index 0 */
273 for (i = 0; i < bbio->num_stripes; ++i) {
274 /* bounds have already been checked */
275 zone->devs[i] = bbio->stripes[i].dev;
276 }
277 zone->ndevs = bbio->num_stripes;
278
279 spin_lock(&fs_info->reada_lock);
280 ret = radix_tree_insert(&dev->reada_zones,
281 (unsigned long)(zone->end >> PAGE_SHIFT),
282 zone);
283
284 if (ret == -EEXIST) {
285 kfree(zone);
286 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
287 logical >> PAGE_SHIFT, 1);
288 if (ret == 1 && logical >= zone->start && logical <= zone->end)
289 kref_get(&zone->refcnt);
290 else
291 zone = NULL;
292 }
293 spin_unlock(&fs_info->reada_lock);
294 radix_tree_preload_end();
295
296 return zone;
297}
298
299static struct reada_extent *reada_find_extent(struct btrfs_fs_info *fs_info,
300 u64 logical,
301 struct btrfs_key *top)
302{
303 int ret;
304 struct reada_extent *re = NULL;
305 struct reada_extent *re_exist = NULL;
306 struct btrfs_bio *bbio = NULL;
307 struct btrfs_device *dev;
308 struct btrfs_device *prev_dev;
309 u64 length;
310 int real_stripes;
311 int nzones = 0;
312 unsigned long index = logical >> PAGE_SHIFT;
313 int dev_replace_is_ongoing;
314 int have_zone = 0;
315
316 spin_lock(&fs_info->reada_lock);
317 re = radix_tree_lookup(&fs_info->reada_tree, index);
318 if (re)
319 re->refcnt++;
320 spin_unlock(&fs_info->reada_lock);
321
322 if (re)
323 return re;
324
325 re = kzalloc(sizeof(*re), GFP_KERNEL);
326 if (!re)
327 return NULL;
328
329 re->logical = logical;
330 re->top = *top;
331 INIT_LIST_HEAD(&re->extctl);
332 spin_lock_init(&re->lock);
333 re->refcnt = 1;
334
335 /*
336 * map block
337 */
338 length = fs_info->nodesize;
339 ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
340 &length, &bbio, 0);
341 if (ret || !bbio || length < fs_info->nodesize)
342 goto error;
343
344 if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
345 btrfs_err(fs_info,
346 "readahead: more than %d copies not supported",
347 BTRFS_MAX_MIRRORS);
348 goto error;
349 }
350
351 real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
352 for (nzones = 0; nzones < real_stripes; ++nzones) {
353 struct reada_zone *zone;
354
355 dev = bbio->stripes[nzones].dev;
356
357 /* cannot read ahead on missing device. */
358 if (!dev->bdev)
359 continue;
360
361 zone = reada_find_zone(dev, logical, bbio);
362 if (!zone)
363 continue;
364
365 re->zones[re->nzones++] = zone;
366 spin_lock(&zone->lock);
367 if (!zone->elems)
368 kref_get(&zone->refcnt);
369 ++zone->elems;
370 spin_unlock(&zone->lock);
371 spin_lock(&fs_info->reada_lock);
372 kref_put(&zone->refcnt, reada_zone_release);
373 spin_unlock(&fs_info->reada_lock);
374 }
375 if (re->nzones == 0) {
376 /* not a single zone found, error and out */
377 goto error;
378 }
379
380 /* Insert extent in reada tree + all per-device trees, all or nothing */
381 down_read(&fs_info->dev_replace.rwsem);
382 ret = radix_tree_preload(GFP_KERNEL);
383 if (ret) {
384 up_read(&fs_info->dev_replace.rwsem);
385 goto error;
386 }
387
388 spin_lock(&fs_info->reada_lock);
389 ret = radix_tree_insert(&fs_info->reada_tree, index, re);
390 if (ret == -EEXIST) {
391 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
392 re_exist->refcnt++;
393 spin_unlock(&fs_info->reada_lock);
394 radix_tree_preload_end();
395 up_read(&fs_info->dev_replace.rwsem);
396 goto error;
397 }
398 if (ret) {
399 spin_unlock(&fs_info->reada_lock);
400 radix_tree_preload_end();
401 up_read(&fs_info->dev_replace.rwsem);
402 goto error;
403 }
404 radix_tree_preload_end();
405 prev_dev = NULL;
406 dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
407 &fs_info->dev_replace);
408 for (nzones = 0; nzones < re->nzones; ++nzones) {
409 dev = re->zones[nzones]->device;
410
411 if (dev == prev_dev) {
412 /*
413 * in case of DUP, just add the first zone. As both
414 * are on the same device, there's nothing to gain
415 * from adding both.
416 * Also, it wouldn't work, as the tree is per device
417 * and adding would fail with EEXIST
418 */
419 continue;
420 }
421 if (!dev->bdev)
422 continue;
423
424 if (dev_replace_is_ongoing &&
425 dev == fs_info->dev_replace.tgtdev) {
426 /*
427 * as this device is selected for reading only as
428 * a last resort, skip it for read ahead.
429 */
430 continue;
431 }
432 prev_dev = dev;
433 ret = radix_tree_insert(&dev->reada_extents, index, re);
434 if (ret) {
435 while (--nzones >= 0) {
436 dev = re->zones[nzones]->device;
437 BUG_ON(dev == NULL);
438 /* ignore whether the entry was inserted */
439 radix_tree_delete(&dev->reada_extents, index);
440 }
441 radix_tree_delete(&fs_info->reada_tree, index);
442 spin_unlock(&fs_info->reada_lock);
443 up_read(&fs_info->dev_replace.rwsem);
444 goto error;
445 }
446 have_zone = 1;
447 }
448 spin_unlock(&fs_info->reada_lock);
449 up_read(&fs_info->dev_replace.rwsem);
450
451 if (!have_zone)
452 goto error;
453
454 btrfs_put_bbio(bbio);
455 return re;
456
457error:
458 for (nzones = 0; nzones < re->nzones; ++nzones) {
459 struct reada_zone *zone;
460
461 zone = re->zones[nzones];
462 kref_get(&zone->refcnt);
463 spin_lock(&zone->lock);
464 --zone->elems;
465 if (zone->elems == 0) {
466 /*
467 * no fs_info->reada_lock needed, as this can't be
468 * the last ref
469 */
470 kref_put(&zone->refcnt, reada_zone_release);
471 }
472 spin_unlock(&zone->lock);
473
474 spin_lock(&fs_info->reada_lock);
475 kref_put(&zone->refcnt, reada_zone_release);
476 spin_unlock(&fs_info->reada_lock);
477 }
478 btrfs_put_bbio(bbio);
479 kfree(re);
480 return re_exist;
481}
482
483static void reada_extent_put(struct btrfs_fs_info *fs_info,
484 struct reada_extent *re)
485{
486 int i;
487 unsigned long index = re->logical >> PAGE_SHIFT;
488
489 spin_lock(&fs_info->reada_lock);
490 if (--re->refcnt) {
491 spin_unlock(&fs_info->reada_lock);
492 return;
493 }
494
495 radix_tree_delete(&fs_info->reada_tree, index);
496 for (i = 0; i < re->nzones; ++i) {
497 struct reada_zone *zone = re->zones[i];
498
499 radix_tree_delete(&zone->device->reada_extents, index);
500 }
501
502 spin_unlock(&fs_info->reada_lock);
503
504 for (i = 0; i < re->nzones; ++i) {
505 struct reada_zone *zone = re->zones[i];
506
507 kref_get(&zone->refcnt);
508 spin_lock(&zone->lock);
509 --zone->elems;
510 if (zone->elems == 0) {
511 /* no fs_info->reada_lock needed, as this can't be
512 * the last ref */
513 kref_put(&zone->refcnt, reada_zone_release);
514 }
515 spin_unlock(&zone->lock);
516
517 spin_lock(&fs_info->reada_lock);
518 kref_put(&zone->refcnt, reada_zone_release);
519 spin_unlock(&fs_info->reada_lock);
520 }
521
522 kfree(re);
523}
524
525static void reada_zone_release(struct kref *kref)
526{
527 struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
528
529 radix_tree_delete(&zone->device->reada_zones,
530 zone->end >> PAGE_SHIFT);
531
532 kfree(zone);
533}
534
535static void reada_control_release(struct kref *kref)
536{
537 struct reada_control *rc = container_of(kref, struct reada_control,
538 refcnt);
539
540 kfree(rc);
541}
542
543static int reada_add_block(struct reada_control *rc, u64 logical,
544 struct btrfs_key *top, u64 generation)
545{
546 struct btrfs_fs_info *fs_info = rc->fs_info;
547 struct reada_extent *re;
548 struct reada_extctl *rec;
549
550 /* takes one ref */
551 re = reada_find_extent(fs_info, logical, top);
552 if (!re)
553 return -1;
554
555 rec = kzalloc(sizeof(*rec), GFP_KERNEL);
556 if (!rec) {
557 reada_extent_put(fs_info, re);
558 return -ENOMEM;
559 }
560
561 rec->rc = rc;
562 rec->generation = generation;
563 atomic_inc(&rc->elems);
564
565 spin_lock(&re->lock);
566 list_add_tail(&rec->list, &re->extctl);
567 spin_unlock(&re->lock);
568
569 /* leave the ref on the extent */
570
571 return 0;
572}
573
574/*
575 * called with fs_info->reada_lock held
576 */
577static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
578{
579 int i;
580 unsigned long index = zone->end >> PAGE_SHIFT;
581
582 for (i = 0; i < zone->ndevs; ++i) {
583 struct reada_zone *peer;
584 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
585 if (peer && peer->device != zone->device)
586 peer->locked = lock;
587 }
588}
589
590/*
591 * called with fs_info->reada_lock held
592 */
593static int reada_pick_zone(struct btrfs_device *dev)
594{
595 struct reada_zone *top_zone = NULL;
596 struct reada_zone *top_locked_zone = NULL;
597 u64 top_elems = 0;
598 u64 top_locked_elems = 0;
599 unsigned long index = 0;
600 int ret;
601
602 if (dev->reada_curr_zone) {
603 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
604 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
605 dev->reada_curr_zone = NULL;
606 }
607 /* pick the zone with the most elements */
608 while (1) {
609 struct reada_zone *zone;
610
611 ret = radix_tree_gang_lookup(&dev->reada_zones,
612 (void **)&zone, index, 1);
613 if (ret == 0)
614 break;
615 index = (zone->end >> PAGE_SHIFT) + 1;
616 if (zone->locked) {
617 if (zone->elems > top_locked_elems) {
618 top_locked_elems = zone->elems;
619 top_locked_zone = zone;
620 }
621 } else {
622 if (zone->elems > top_elems) {
623 top_elems = zone->elems;
624 top_zone = zone;
625 }
626 }
627 }
628 if (top_zone)
629 dev->reada_curr_zone = top_zone;
630 else if (top_locked_zone)
631 dev->reada_curr_zone = top_locked_zone;
632 else
633 return 0;
634
635 dev->reada_next = dev->reada_curr_zone->start;
636 kref_get(&dev->reada_curr_zone->refcnt);
637 reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
638
639 return 1;
640}
641
642static int reada_tree_block_flagged(struct btrfs_fs_info *fs_info, u64 bytenr,
643 int mirror_num, struct extent_buffer **eb)
644{
645 struct extent_buffer *buf = NULL;
646 int ret;
647
648 buf = btrfs_find_create_tree_block(fs_info, bytenr);
649 if (IS_ERR(buf))
650 return 0;
651
652 set_bit(EXTENT_BUFFER_READAHEAD, &buf->bflags);
653
654 ret = read_extent_buffer_pages(buf, WAIT_PAGE_LOCK, mirror_num);
655 if (ret) {
656 free_extent_buffer_stale(buf);
657 return ret;
658 }
659
660 if (test_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags)) {
661 free_extent_buffer_stale(buf);
662 return -EIO;
663 } else if (extent_buffer_uptodate(buf)) {
664 *eb = buf;
665 } else {
666 free_extent_buffer(buf);
667 }
668 return 0;
669}
670
671static int reada_start_machine_dev(struct btrfs_device *dev)
672{
673 struct btrfs_fs_info *fs_info = dev->fs_info;
674 struct reada_extent *re = NULL;
675 int mirror_num = 0;
676 struct extent_buffer *eb = NULL;
677 u64 logical;
678 int ret;
679 int i;
680
681 spin_lock(&fs_info->reada_lock);
682 if (dev->reada_curr_zone == NULL) {
683 ret = reada_pick_zone(dev);
684 if (!ret) {
685 spin_unlock(&fs_info->reada_lock);
686 return 0;
687 }
688 }
689 /*
690 * FIXME currently we issue the reads one extent at a time. If we have
691 * a contiguous block of extents, we could also coagulate them or use
692 * plugging to speed things up
693 */
694 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
695 dev->reada_next >> PAGE_SHIFT, 1);
696 if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
697 ret = reada_pick_zone(dev);
698 if (!ret) {
699 spin_unlock(&fs_info->reada_lock);
700 return 0;
701 }
702 re = NULL;
703 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
704 dev->reada_next >> PAGE_SHIFT, 1);
705 }
706 if (ret == 0) {
707 spin_unlock(&fs_info->reada_lock);
708 return 0;
709 }
710 dev->reada_next = re->logical + fs_info->nodesize;
711 re->refcnt++;
712
713 spin_unlock(&fs_info->reada_lock);
714
715 spin_lock(&re->lock);
716 if (re->scheduled || list_empty(&re->extctl)) {
717 spin_unlock(&re->lock);
718 reada_extent_put(fs_info, re);
719 return 0;
720 }
721 re->scheduled = 1;
722 spin_unlock(&re->lock);
723
724 /*
725 * find mirror num
726 */
727 for (i = 0; i < re->nzones; ++i) {
728 if (re->zones[i]->device == dev) {
729 mirror_num = i + 1;
730 break;
731 }
732 }
733 logical = re->logical;
734
735 atomic_inc(&dev->reada_in_flight);
736 ret = reada_tree_block_flagged(fs_info, logical, mirror_num, &eb);
737 if (ret)
738 __readahead_hook(fs_info, re, NULL, ret);
739 else if (eb)
740 __readahead_hook(fs_info, re, eb, ret);
741
742 if (eb)
743 free_extent_buffer(eb);
744
745 atomic_dec(&dev->reada_in_flight);
746 reada_extent_put(fs_info, re);
747
748 return 1;
749
750}
751
752static void reada_start_machine_worker(struct btrfs_work *work)
753{
754 struct reada_machine_work *rmw;
755 struct btrfs_fs_info *fs_info;
756 int old_ioprio;
757
758 rmw = container_of(work, struct reada_machine_work, work);
759 fs_info = rmw->fs_info;
760
761 kfree(rmw);
762
763 old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
764 task_nice_ioprio(current));
765 set_task_ioprio(current, BTRFS_IOPRIO_READA);
766 __reada_start_machine(fs_info);
767 set_task_ioprio(current, old_ioprio);
768
769 atomic_dec(&fs_info->reada_works_cnt);
770}
771
772static void __reada_start_machine(struct btrfs_fs_info *fs_info)
773{
774 struct btrfs_device *device;
775 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
776 u64 enqueued;
777 u64 total = 0;
778 int i;
779
780again:
781 do {
782 enqueued = 0;
783 mutex_lock(&fs_devices->device_list_mutex);
784 list_for_each_entry(device, &fs_devices->devices, dev_list) {
785 if (atomic_read(&device->reada_in_flight) <
786 MAX_IN_FLIGHT)
787 enqueued += reada_start_machine_dev(device);
788 }
789 mutex_unlock(&fs_devices->device_list_mutex);
790 total += enqueued;
791 } while (enqueued && total < 10000);
792 if (fs_devices->seed) {
793 fs_devices = fs_devices->seed;
794 goto again;
795 }
796
797 if (enqueued == 0)
798 return;
799
800 /*
801 * If everything is already in the cache, this is effectively single
802 * threaded. To a) not hold the caller for too long and b) to utilize
803 * more cores, we broke the loop above after 10000 iterations and now
804 * enqueue to workers to finish it. This will distribute the load to
805 * the cores.
806 */
807 for (i = 0; i < 2; ++i) {
808 reada_start_machine(fs_info);
809 if (atomic_read(&fs_info->reada_works_cnt) >
810 BTRFS_MAX_MIRRORS * 2)
811 break;
812 }
813}
814
815static void reada_start_machine(struct btrfs_fs_info *fs_info)
816{
817 struct reada_machine_work *rmw;
818
819 rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
820 if (!rmw) {
821 /* FIXME we cannot handle this properly right now */
822 BUG();
823 }
824 btrfs_init_work(&rmw->work, btrfs_readahead_helper,
825 reada_start_machine_worker, NULL, NULL);
826 rmw->fs_info = fs_info;
827
828 btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
829 atomic_inc(&fs_info->reada_works_cnt);
830}
831
832#ifdef DEBUG
833static void dump_devs(struct btrfs_fs_info *fs_info, int all)
834{
835 struct btrfs_device *device;
836 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
837 unsigned long index;
838 int ret;
839 int i;
840 int j;
841 int cnt;
842
843 spin_lock(&fs_info->reada_lock);
844 list_for_each_entry(device, &fs_devices->devices, dev_list) {
845 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
846 atomic_read(&device->reada_in_flight));
847 index = 0;
848 while (1) {
849 struct reada_zone *zone;
850 ret = radix_tree_gang_lookup(&device->reada_zones,
851 (void **)&zone, index, 1);
852 if (ret == 0)
853 break;
854 pr_debug(" zone %llu-%llu elems %llu locked %d devs",
855 zone->start, zone->end, zone->elems,
856 zone->locked);
857 for (j = 0; j < zone->ndevs; ++j) {
858 pr_cont(" %lld",
859 zone->devs[j]->devid);
860 }
861 if (device->reada_curr_zone == zone)
862 pr_cont(" curr off %llu",
863 device->reada_next - zone->start);
864 pr_cont("\n");
865 index = (zone->end >> PAGE_SHIFT) + 1;
866 }
867 cnt = 0;
868 index = 0;
869 while (all) {
870 struct reada_extent *re = NULL;
871
872 ret = radix_tree_gang_lookup(&device->reada_extents,
873 (void **)&re, index, 1);
874 if (ret == 0)
875 break;
876 pr_debug(" re: logical %llu size %u empty %d scheduled %d",
877 re->logical, fs_info->nodesize,
878 list_empty(&re->extctl), re->scheduled);
879
880 for (i = 0; i < re->nzones; ++i) {
881 pr_cont(" zone %llu-%llu devs",
882 re->zones[i]->start,
883 re->zones[i]->end);
884 for (j = 0; j < re->zones[i]->ndevs; ++j) {
885 pr_cont(" %lld",
886 re->zones[i]->devs[j]->devid);
887 }
888 }
889 pr_cont("\n");
890 index = (re->logical >> PAGE_SHIFT) + 1;
891 if (++cnt > 15)
892 break;
893 }
894 }
895
896 index = 0;
897 cnt = 0;
898 while (all) {
899 struct reada_extent *re = NULL;
900
901 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
902 index, 1);
903 if (ret == 0)
904 break;
905 if (!re->scheduled) {
906 index = (re->logical >> PAGE_SHIFT) + 1;
907 continue;
908 }
909 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
910 re->logical, fs_info->nodesize,
911 list_empty(&re->extctl), re->scheduled);
912 for (i = 0; i < re->nzones; ++i) {
913 pr_cont(" zone %llu-%llu devs",
914 re->zones[i]->start,
915 re->zones[i]->end);
916 for (j = 0; j < re->zones[i]->ndevs; ++j) {
917 pr_cont(" %lld",
918 re->zones[i]->devs[j]->devid);
919 }
920 }
921 pr_cont("\n");
922 index = (re->logical >> PAGE_SHIFT) + 1;
923 }
924 spin_unlock(&fs_info->reada_lock);
925}
926#endif
927
928/*
929 * interface
930 */
931struct reada_control *btrfs_reada_add(struct btrfs_root *root,
932 struct btrfs_key *key_start, struct btrfs_key *key_end)
933{
934 struct reada_control *rc;
935 u64 start;
936 u64 generation;
937 int ret;
938 struct extent_buffer *node;
939 static struct btrfs_key max_key = {
940 .objectid = (u64)-1,
941 .type = (u8)-1,
942 .offset = (u64)-1
943 };
944
945 rc = kzalloc(sizeof(*rc), GFP_KERNEL);
946 if (!rc)
947 return ERR_PTR(-ENOMEM);
948
949 rc->fs_info = root->fs_info;
950 rc->key_start = *key_start;
951 rc->key_end = *key_end;
952 atomic_set(&rc->elems, 0);
953 init_waitqueue_head(&rc->wait);
954 kref_init(&rc->refcnt);
955 kref_get(&rc->refcnt); /* one ref for having elements */
956
957 node = btrfs_root_node(root);
958 start = node->start;
959 generation = btrfs_header_generation(node);
960 free_extent_buffer(node);
961
962 ret = reada_add_block(rc, start, &max_key, generation);
963 if (ret) {
964 kfree(rc);
965 return ERR_PTR(ret);
966 }
967
968 reada_start_machine(root->fs_info);
969
970 return rc;
971}
972
973#ifdef DEBUG
974int btrfs_reada_wait(void *handle)
975{
976 struct reada_control *rc = handle;
977 struct btrfs_fs_info *fs_info = rc->fs_info;
978
979 while (atomic_read(&rc->elems)) {
980 if (!atomic_read(&fs_info->reada_works_cnt))
981 reada_start_machine(fs_info);
982 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
983 5 * HZ);
984 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
985 }
986
987 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
988
989 kref_put(&rc->refcnt, reada_control_release);
990
991 return 0;
992}
993#else
994int btrfs_reada_wait(void *handle)
995{
996 struct reada_control *rc = handle;
997 struct btrfs_fs_info *fs_info = rc->fs_info;
998
999 while (atomic_read(&rc->elems)) {
1000 if (!atomic_read(&fs_info->reada_works_cnt))
1001 reada_start_machine(fs_info);
1002 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
1003 (HZ + 9) / 10);
1004 }
1005
1006 kref_put(&rc->refcnt, reada_control_release);
1007
1008 return 0;
1009}
1010#endif
1011
1012void btrfs_reada_detach(void *handle)
1013{
1014 struct reada_control *rc = handle;
1015
1016 kref_put(&rc->refcnt, reada_control_release);
1017}
1/*
2 * Copyright (C) 2011 STRATO. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public
6 * License v2 as published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful,
9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 *
13 * You should have received a copy of the GNU General Public
14 * License along with this program; if not, write to the
15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
16 * Boston, MA 021110-1307, USA.
17 */
18
19#include <linux/sched.h>
20#include <linux/pagemap.h>
21#include <linux/writeback.h>
22#include <linux/blkdev.h>
23#include <linux/rbtree.h>
24#include <linux/slab.h>
25#include <linux/workqueue.h>
26#include "ctree.h"
27#include "volumes.h"
28#include "disk-io.h"
29#include "transaction.h"
30#include "dev-replace.h"
31
32#undef DEBUG
33
34/*
35 * This is the implementation for the generic read ahead framework.
36 *
37 * To trigger a readahead, btrfs_reada_add must be called. It will start
38 * a read ahead for the given range [start, end) on tree root. The returned
39 * handle can either be used to wait on the readahead to finish
40 * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).
41 *
42 * The read ahead works as follows:
43 * On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
44 * reada_start_machine will then search for extents to prefetch and trigger
45 * some reads. When a read finishes for a node, all contained node/leaf
46 * pointers that lie in the given range will also be enqueued. The reads will
47 * be triggered in sequential order, thus giving a big win over a naive
48 * enumeration. It will also make use of multi-device layouts. Each disk
49 * will have its on read pointer and all disks will by utilized in parallel.
50 * Also will no two disks read both sides of a mirror simultaneously, as this
51 * would waste seeking capacity. Instead both disks will read different parts
52 * of the filesystem.
53 * Any number of readaheads can be started in parallel. The read order will be
54 * determined globally, i.e. 2 parallel readaheads will normally finish faster
55 * than the 2 started one after another.
56 */
57
58#define MAX_IN_FLIGHT 6
59
60struct reada_extctl {
61 struct list_head list;
62 struct reada_control *rc;
63 u64 generation;
64};
65
66struct reada_extent {
67 u64 logical;
68 struct btrfs_key top;
69 int err;
70 struct list_head extctl;
71 int refcnt;
72 spinlock_t lock;
73 struct reada_zone *zones[BTRFS_MAX_MIRRORS];
74 int nzones;
75 int scheduled;
76};
77
78struct reada_zone {
79 u64 start;
80 u64 end;
81 u64 elems;
82 struct list_head list;
83 spinlock_t lock;
84 int locked;
85 struct btrfs_device *device;
86 struct btrfs_device *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
87 * self */
88 int ndevs;
89 struct kref refcnt;
90};
91
92struct reada_machine_work {
93 struct btrfs_work work;
94 struct btrfs_fs_info *fs_info;
95};
96
97static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
98static void reada_control_release(struct kref *kref);
99static void reada_zone_release(struct kref *kref);
100static void reada_start_machine(struct btrfs_fs_info *fs_info);
101static void __reada_start_machine(struct btrfs_fs_info *fs_info);
102
103static int reada_add_block(struct reada_control *rc, u64 logical,
104 struct btrfs_key *top, u64 generation);
105
106/* recurses */
107/* in case of err, eb might be NULL */
108static void __readahead_hook(struct btrfs_fs_info *fs_info,
109 struct reada_extent *re, struct extent_buffer *eb,
110 u64 start, int err)
111{
112 int level = 0;
113 int nritems;
114 int i;
115 u64 bytenr;
116 u64 generation;
117 struct list_head list;
118
119 if (eb)
120 level = btrfs_header_level(eb);
121
122 spin_lock(&re->lock);
123 /*
124 * just take the full list from the extent. afterwards we
125 * don't need the lock anymore
126 */
127 list_replace_init(&re->extctl, &list);
128 re->scheduled = 0;
129 spin_unlock(&re->lock);
130
131 /*
132 * this is the error case, the extent buffer has not been
133 * read correctly. We won't access anything from it and
134 * just cleanup our data structures. Effectively this will
135 * cut the branch below this node from read ahead.
136 */
137 if (err)
138 goto cleanup;
139
140 /*
141 * FIXME: currently we just set nritems to 0 if this is a leaf,
142 * effectively ignoring the content. In a next step we could
143 * trigger more readahead depending from the content, e.g.
144 * fetch the checksums for the extents in the leaf.
145 */
146 if (!level)
147 goto cleanup;
148
149 nritems = btrfs_header_nritems(eb);
150 generation = btrfs_header_generation(eb);
151 for (i = 0; i < nritems; i++) {
152 struct reada_extctl *rec;
153 u64 n_gen;
154 struct btrfs_key key;
155 struct btrfs_key next_key;
156
157 btrfs_node_key_to_cpu(eb, &key, i);
158 if (i + 1 < nritems)
159 btrfs_node_key_to_cpu(eb, &next_key, i + 1);
160 else
161 next_key = re->top;
162 bytenr = btrfs_node_blockptr(eb, i);
163 n_gen = btrfs_node_ptr_generation(eb, i);
164
165 list_for_each_entry(rec, &list, list) {
166 struct reada_control *rc = rec->rc;
167
168 /*
169 * if the generation doesn't match, just ignore this
170 * extctl. This will probably cut off a branch from
171 * prefetch. Alternatively one could start a new (sub-)
172 * prefetch for this branch, starting again from root.
173 * FIXME: move the generation check out of this loop
174 */
175#ifdef DEBUG
176 if (rec->generation != generation) {
177 btrfs_debug(fs_info,
178 "generation mismatch for (%llu,%d,%llu) %llu != %llu",
179 key.objectid, key.type, key.offset,
180 rec->generation, generation);
181 }
182#endif
183 if (rec->generation == generation &&
184 btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
185 btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
186 reada_add_block(rc, bytenr, &next_key, n_gen);
187 }
188 }
189
190cleanup:
191 /*
192 * free extctl records
193 */
194 while (!list_empty(&list)) {
195 struct reada_control *rc;
196 struct reada_extctl *rec;
197
198 rec = list_first_entry(&list, struct reada_extctl, list);
199 list_del(&rec->list);
200 rc = rec->rc;
201 kfree(rec);
202
203 kref_get(&rc->refcnt);
204 if (atomic_dec_and_test(&rc->elems)) {
205 kref_put(&rc->refcnt, reada_control_release);
206 wake_up(&rc->wait);
207 }
208 kref_put(&rc->refcnt, reada_control_release);
209
210 reada_extent_put(fs_info, re); /* one ref for each entry */
211 }
212
213 return;
214}
215
216/*
217 * start is passed separately in case eb in NULL, which may be the case with
218 * failed I/O
219 */
220int btree_readahead_hook(struct btrfs_fs_info *fs_info,
221 struct extent_buffer *eb, u64 start, int err)
222{
223 int ret = 0;
224 struct reada_extent *re;
225
226 /* find extent */
227 spin_lock(&fs_info->reada_lock);
228 re = radix_tree_lookup(&fs_info->reada_tree,
229 start >> PAGE_SHIFT);
230 if (re)
231 re->refcnt++;
232 spin_unlock(&fs_info->reada_lock);
233 if (!re) {
234 ret = -1;
235 goto start_machine;
236 }
237
238 __readahead_hook(fs_info, re, eb, start, err);
239 reada_extent_put(fs_info, re); /* our ref */
240
241start_machine:
242 reada_start_machine(fs_info);
243 return ret;
244}
245
246static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
247 struct btrfs_device *dev, u64 logical,
248 struct btrfs_bio *bbio)
249{
250 int ret;
251 struct reada_zone *zone;
252 struct btrfs_block_group_cache *cache = NULL;
253 u64 start;
254 u64 end;
255 int i;
256
257 zone = NULL;
258 spin_lock(&fs_info->reada_lock);
259 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
260 logical >> PAGE_SHIFT, 1);
261 if (ret == 1 && logical >= zone->start && logical <= zone->end) {
262 kref_get(&zone->refcnt);
263 spin_unlock(&fs_info->reada_lock);
264 return zone;
265 }
266
267 spin_unlock(&fs_info->reada_lock);
268
269 cache = btrfs_lookup_block_group(fs_info, logical);
270 if (!cache)
271 return NULL;
272
273 start = cache->key.objectid;
274 end = start + cache->key.offset - 1;
275 btrfs_put_block_group(cache);
276
277 zone = kzalloc(sizeof(*zone), GFP_KERNEL);
278 if (!zone)
279 return NULL;
280
281 zone->start = start;
282 zone->end = end;
283 INIT_LIST_HEAD(&zone->list);
284 spin_lock_init(&zone->lock);
285 zone->locked = 0;
286 kref_init(&zone->refcnt);
287 zone->elems = 0;
288 zone->device = dev; /* our device always sits at index 0 */
289 for (i = 0; i < bbio->num_stripes; ++i) {
290 /* bounds have already been checked */
291 zone->devs[i] = bbio->stripes[i].dev;
292 }
293 zone->ndevs = bbio->num_stripes;
294
295 spin_lock(&fs_info->reada_lock);
296 ret = radix_tree_insert(&dev->reada_zones,
297 (unsigned long)(zone->end >> PAGE_SHIFT),
298 zone);
299
300 if (ret == -EEXIST) {
301 kfree(zone);
302 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
303 logical >> PAGE_SHIFT, 1);
304 if (ret == 1 && logical >= zone->start && logical <= zone->end)
305 kref_get(&zone->refcnt);
306 else
307 zone = NULL;
308 }
309 spin_unlock(&fs_info->reada_lock);
310
311 return zone;
312}
313
314static struct reada_extent *reada_find_extent(struct btrfs_root *root,
315 u64 logical,
316 struct btrfs_key *top)
317{
318 int ret;
319 struct reada_extent *re = NULL;
320 struct reada_extent *re_exist = NULL;
321 struct btrfs_fs_info *fs_info = root->fs_info;
322 struct btrfs_bio *bbio = NULL;
323 struct btrfs_device *dev;
324 struct btrfs_device *prev_dev;
325 u32 blocksize;
326 u64 length;
327 int real_stripes;
328 int nzones = 0;
329 unsigned long index = logical >> PAGE_SHIFT;
330 int dev_replace_is_ongoing;
331 int have_zone = 0;
332
333 spin_lock(&fs_info->reada_lock);
334 re = radix_tree_lookup(&fs_info->reada_tree, index);
335 if (re)
336 re->refcnt++;
337 spin_unlock(&fs_info->reada_lock);
338
339 if (re)
340 return re;
341
342 re = kzalloc(sizeof(*re), GFP_KERNEL);
343 if (!re)
344 return NULL;
345
346 blocksize = root->nodesize;
347 re->logical = logical;
348 re->top = *top;
349 INIT_LIST_HEAD(&re->extctl);
350 spin_lock_init(&re->lock);
351 re->refcnt = 1;
352
353 /*
354 * map block
355 */
356 length = blocksize;
357 ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical, &length,
358 &bbio, 0);
359 if (ret || !bbio || length < blocksize)
360 goto error;
361
362 if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
363 btrfs_err(root->fs_info,
364 "readahead: more than %d copies not supported",
365 BTRFS_MAX_MIRRORS);
366 goto error;
367 }
368
369 real_stripes = bbio->num_stripes - bbio->num_tgtdevs;
370 for (nzones = 0; nzones < real_stripes; ++nzones) {
371 struct reada_zone *zone;
372
373 dev = bbio->stripes[nzones].dev;
374
375 /* cannot read ahead on missing device. */
376 if (!dev->bdev)
377 continue;
378
379 zone = reada_find_zone(fs_info, dev, logical, bbio);
380 if (!zone)
381 continue;
382
383 re->zones[re->nzones++] = zone;
384 spin_lock(&zone->lock);
385 if (!zone->elems)
386 kref_get(&zone->refcnt);
387 ++zone->elems;
388 spin_unlock(&zone->lock);
389 spin_lock(&fs_info->reada_lock);
390 kref_put(&zone->refcnt, reada_zone_release);
391 spin_unlock(&fs_info->reada_lock);
392 }
393 if (re->nzones == 0) {
394 /* not a single zone found, error and out */
395 goto error;
396 }
397
398 /* insert extent in reada_tree + all per-device trees, all or nothing */
399 btrfs_dev_replace_lock(&fs_info->dev_replace, 0);
400 spin_lock(&fs_info->reada_lock);
401 ret = radix_tree_insert(&fs_info->reada_tree, index, re);
402 if (ret == -EEXIST) {
403 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
404 BUG_ON(!re_exist);
405 re_exist->refcnt++;
406 spin_unlock(&fs_info->reada_lock);
407 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
408 goto error;
409 }
410 if (ret) {
411 spin_unlock(&fs_info->reada_lock);
412 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
413 goto error;
414 }
415 prev_dev = NULL;
416 dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
417 &fs_info->dev_replace);
418 for (nzones = 0; nzones < re->nzones; ++nzones) {
419 dev = re->zones[nzones]->device;
420
421 if (dev == prev_dev) {
422 /*
423 * in case of DUP, just add the first zone. As both
424 * are on the same device, there's nothing to gain
425 * from adding both.
426 * Also, it wouldn't work, as the tree is per device
427 * and adding would fail with EEXIST
428 */
429 continue;
430 }
431 if (!dev->bdev)
432 continue;
433
434 if (dev_replace_is_ongoing &&
435 dev == fs_info->dev_replace.tgtdev) {
436 /*
437 * as this device is selected for reading only as
438 * a last resort, skip it for read ahead.
439 */
440 continue;
441 }
442 prev_dev = dev;
443 ret = radix_tree_insert(&dev->reada_extents, index, re);
444 if (ret) {
445 while (--nzones >= 0) {
446 dev = re->zones[nzones]->device;
447 BUG_ON(dev == NULL);
448 /* ignore whether the entry was inserted */
449 radix_tree_delete(&dev->reada_extents, index);
450 }
451 BUG_ON(fs_info == NULL);
452 radix_tree_delete(&fs_info->reada_tree, index);
453 spin_unlock(&fs_info->reada_lock);
454 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
455 goto error;
456 }
457 have_zone = 1;
458 }
459 spin_unlock(&fs_info->reada_lock);
460 btrfs_dev_replace_unlock(&fs_info->dev_replace, 0);
461
462 if (!have_zone)
463 goto error;
464
465 btrfs_put_bbio(bbio);
466 return re;
467
468error:
469 for (nzones = 0; nzones < re->nzones; ++nzones) {
470 struct reada_zone *zone;
471
472 zone = re->zones[nzones];
473 kref_get(&zone->refcnt);
474 spin_lock(&zone->lock);
475 --zone->elems;
476 if (zone->elems == 0) {
477 /*
478 * no fs_info->reada_lock needed, as this can't be
479 * the last ref
480 */
481 kref_put(&zone->refcnt, reada_zone_release);
482 }
483 spin_unlock(&zone->lock);
484
485 spin_lock(&fs_info->reada_lock);
486 kref_put(&zone->refcnt, reada_zone_release);
487 spin_unlock(&fs_info->reada_lock);
488 }
489 btrfs_put_bbio(bbio);
490 kfree(re);
491 return re_exist;
492}
493
494static void reada_extent_put(struct btrfs_fs_info *fs_info,
495 struct reada_extent *re)
496{
497 int i;
498 unsigned long index = re->logical >> PAGE_SHIFT;
499
500 spin_lock(&fs_info->reada_lock);
501 if (--re->refcnt) {
502 spin_unlock(&fs_info->reada_lock);
503 return;
504 }
505
506 radix_tree_delete(&fs_info->reada_tree, index);
507 for (i = 0; i < re->nzones; ++i) {
508 struct reada_zone *zone = re->zones[i];
509
510 radix_tree_delete(&zone->device->reada_extents, index);
511 }
512
513 spin_unlock(&fs_info->reada_lock);
514
515 for (i = 0; i < re->nzones; ++i) {
516 struct reada_zone *zone = re->zones[i];
517
518 kref_get(&zone->refcnt);
519 spin_lock(&zone->lock);
520 --zone->elems;
521 if (zone->elems == 0) {
522 /* no fs_info->reada_lock needed, as this can't be
523 * the last ref */
524 kref_put(&zone->refcnt, reada_zone_release);
525 }
526 spin_unlock(&zone->lock);
527
528 spin_lock(&fs_info->reada_lock);
529 kref_put(&zone->refcnt, reada_zone_release);
530 spin_unlock(&fs_info->reada_lock);
531 }
532
533 kfree(re);
534}
535
536static void reada_zone_release(struct kref *kref)
537{
538 struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
539
540 radix_tree_delete(&zone->device->reada_zones,
541 zone->end >> PAGE_SHIFT);
542
543 kfree(zone);
544}
545
546static void reada_control_release(struct kref *kref)
547{
548 struct reada_control *rc = container_of(kref, struct reada_control,
549 refcnt);
550
551 kfree(rc);
552}
553
554static int reada_add_block(struct reada_control *rc, u64 logical,
555 struct btrfs_key *top, u64 generation)
556{
557 struct btrfs_root *root = rc->root;
558 struct reada_extent *re;
559 struct reada_extctl *rec;
560
561 re = reada_find_extent(root, logical, top); /* takes one ref */
562 if (!re)
563 return -1;
564
565 rec = kzalloc(sizeof(*rec), GFP_KERNEL);
566 if (!rec) {
567 reada_extent_put(root->fs_info, re);
568 return -ENOMEM;
569 }
570
571 rec->rc = rc;
572 rec->generation = generation;
573 atomic_inc(&rc->elems);
574
575 spin_lock(&re->lock);
576 list_add_tail(&rec->list, &re->extctl);
577 spin_unlock(&re->lock);
578
579 /* leave the ref on the extent */
580
581 return 0;
582}
583
584/*
585 * called with fs_info->reada_lock held
586 */
587static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
588{
589 int i;
590 unsigned long index = zone->end >> PAGE_SHIFT;
591
592 for (i = 0; i < zone->ndevs; ++i) {
593 struct reada_zone *peer;
594 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
595 if (peer && peer->device != zone->device)
596 peer->locked = lock;
597 }
598}
599
600/*
601 * called with fs_info->reada_lock held
602 */
603static int reada_pick_zone(struct btrfs_device *dev)
604{
605 struct reada_zone *top_zone = NULL;
606 struct reada_zone *top_locked_zone = NULL;
607 u64 top_elems = 0;
608 u64 top_locked_elems = 0;
609 unsigned long index = 0;
610 int ret;
611
612 if (dev->reada_curr_zone) {
613 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
614 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
615 dev->reada_curr_zone = NULL;
616 }
617 /* pick the zone with the most elements */
618 while (1) {
619 struct reada_zone *zone;
620
621 ret = radix_tree_gang_lookup(&dev->reada_zones,
622 (void **)&zone, index, 1);
623 if (ret == 0)
624 break;
625 index = (zone->end >> PAGE_SHIFT) + 1;
626 if (zone->locked) {
627 if (zone->elems > top_locked_elems) {
628 top_locked_elems = zone->elems;
629 top_locked_zone = zone;
630 }
631 } else {
632 if (zone->elems > top_elems) {
633 top_elems = zone->elems;
634 top_zone = zone;
635 }
636 }
637 }
638 if (top_zone)
639 dev->reada_curr_zone = top_zone;
640 else if (top_locked_zone)
641 dev->reada_curr_zone = top_locked_zone;
642 else
643 return 0;
644
645 dev->reada_next = dev->reada_curr_zone->start;
646 kref_get(&dev->reada_curr_zone->refcnt);
647 reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
648
649 return 1;
650}
651
652static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
653 struct btrfs_device *dev)
654{
655 struct reada_extent *re = NULL;
656 int mirror_num = 0;
657 struct extent_buffer *eb = NULL;
658 u64 logical;
659 int ret;
660 int i;
661
662 spin_lock(&fs_info->reada_lock);
663 if (dev->reada_curr_zone == NULL) {
664 ret = reada_pick_zone(dev);
665 if (!ret) {
666 spin_unlock(&fs_info->reada_lock);
667 return 0;
668 }
669 }
670 /*
671 * FIXME currently we issue the reads one extent at a time. If we have
672 * a contiguous block of extents, we could also coagulate them or use
673 * plugging to speed things up
674 */
675 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
676 dev->reada_next >> PAGE_SHIFT, 1);
677 if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
678 ret = reada_pick_zone(dev);
679 if (!ret) {
680 spin_unlock(&fs_info->reada_lock);
681 return 0;
682 }
683 re = NULL;
684 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
685 dev->reada_next >> PAGE_SHIFT, 1);
686 }
687 if (ret == 0) {
688 spin_unlock(&fs_info->reada_lock);
689 return 0;
690 }
691 dev->reada_next = re->logical + fs_info->tree_root->nodesize;
692 re->refcnt++;
693
694 spin_unlock(&fs_info->reada_lock);
695
696 spin_lock(&re->lock);
697 if (re->scheduled || list_empty(&re->extctl)) {
698 spin_unlock(&re->lock);
699 reada_extent_put(fs_info, re);
700 return 0;
701 }
702 re->scheduled = 1;
703 spin_unlock(&re->lock);
704
705 /*
706 * find mirror num
707 */
708 for (i = 0; i < re->nzones; ++i) {
709 if (re->zones[i]->device == dev) {
710 mirror_num = i + 1;
711 break;
712 }
713 }
714 logical = re->logical;
715
716 atomic_inc(&dev->reada_in_flight);
717 ret = reada_tree_block_flagged(fs_info->extent_root, logical,
718 mirror_num, &eb);
719 if (ret)
720 __readahead_hook(fs_info, re, NULL, logical, ret);
721 else if (eb)
722 __readahead_hook(fs_info, re, eb, eb->start, ret);
723
724 if (eb)
725 free_extent_buffer(eb);
726
727 atomic_dec(&dev->reada_in_flight);
728 reada_extent_put(fs_info, re);
729
730 return 1;
731
732}
733
734static void reada_start_machine_worker(struct btrfs_work *work)
735{
736 struct reada_machine_work *rmw;
737 struct btrfs_fs_info *fs_info;
738 int old_ioprio;
739
740 rmw = container_of(work, struct reada_machine_work, work);
741 fs_info = rmw->fs_info;
742
743 kfree(rmw);
744
745 old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
746 task_nice_ioprio(current));
747 set_task_ioprio(current, BTRFS_IOPRIO_READA);
748 __reada_start_machine(fs_info);
749 set_task_ioprio(current, old_ioprio);
750
751 atomic_dec(&fs_info->reada_works_cnt);
752}
753
754static void __reada_start_machine(struct btrfs_fs_info *fs_info)
755{
756 struct btrfs_device *device;
757 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
758 u64 enqueued;
759 u64 total = 0;
760 int i;
761
762 do {
763 enqueued = 0;
764 list_for_each_entry(device, &fs_devices->devices, dev_list) {
765 if (atomic_read(&device->reada_in_flight) <
766 MAX_IN_FLIGHT)
767 enqueued += reada_start_machine_dev(fs_info,
768 device);
769 }
770 total += enqueued;
771 } while (enqueued && total < 10000);
772
773 if (enqueued == 0)
774 return;
775
776 /*
777 * If everything is already in the cache, this is effectively single
778 * threaded. To a) not hold the caller for too long and b) to utilize
779 * more cores, we broke the loop above after 10000 iterations and now
780 * enqueue to workers to finish it. This will distribute the load to
781 * the cores.
782 */
783 for (i = 0; i < 2; ++i) {
784 reada_start_machine(fs_info);
785 if (atomic_read(&fs_info->reada_works_cnt) >
786 BTRFS_MAX_MIRRORS * 2)
787 break;
788 }
789}
790
791static void reada_start_machine(struct btrfs_fs_info *fs_info)
792{
793 struct reada_machine_work *rmw;
794
795 rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
796 if (!rmw) {
797 /* FIXME we cannot handle this properly right now */
798 BUG();
799 }
800 btrfs_init_work(&rmw->work, btrfs_readahead_helper,
801 reada_start_machine_worker, NULL, NULL);
802 rmw->fs_info = fs_info;
803
804 btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
805 atomic_inc(&fs_info->reada_works_cnt);
806}
807
808#ifdef DEBUG
809static void dump_devs(struct btrfs_fs_info *fs_info, int all)
810{
811 struct btrfs_device *device;
812 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
813 unsigned long index;
814 int ret;
815 int i;
816 int j;
817 int cnt;
818
819 spin_lock(&fs_info->reada_lock);
820 list_for_each_entry(device, &fs_devices->devices, dev_list) {
821 printk(KERN_DEBUG "dev %lld has %d in flight\n", device->devid,
822 atomic_read(&device->reada_in_flight));
823 index = 0;
824 while (1) {
825 struct reada_zone *zone;
826 ret = radix_tree_gang_lookup(&device->reada_zones,
827 (void **)&zone, index, 1);
828 if (ret == 0)
829 break;
830 printk(KERN_DEBUG " zone %llu-%llu elems %llu locked "
831 "%d devs", zone->start, zone->end, zone->elems,
832 zone->locked);
833 for (j = 0; j < zone->ndevs; ++j) {
834 printk(KERN_CONT " %lld",
835 zone->devs[j]->devid);
836 }
837 if (device->reada_curr_zone == zone)
838 printk(KERN_CONT " curr off %llu",
839 device->reada_next - zone->start);
840 printk(KERN_CONT "\n");
841 index = (zone->end >> PAGE_SHIFT) + 1;
842 }
843 cnt = 0;
844 index = 0;
845 while (all) {
846 struct reada_extent *re = NULL;
847
848 ret = radix_tree_gang_lookup(&device->reada_extents,
849 (void **)&re, index, 1);
850 if (ret == 0)
851 break;
852 printk(KERN_DEBUG
853 " re: logical %llu size %u empty %d scheduled %d",
854 re->logical, fs_info->tree_root->nodesize,
855 list_empty(&re->extctl), re->scheduled);
856
857 for (i = 0; i < re->nzones; ++i) {
858 printk(KERN_CONT " zone %llu-%llu devs",
859 re->zones[i]->start,
860 re->zones[i]->end);
861 for (j = 0; j < re->zones[i]->ndevs; ++j) {
862 printk(KERN_CONT " %lld",
863 re->zones[i]->devs[j]->devid);
864 }
865 }
866 printk(KERN_CONT "\n");
867 index = (re->logical >> PAGE_SHIFT) + 1;
868 if (++cnt > 15)
869 break;
870 }
871 }
872
873 index = 0;
874 cnt = 0;
875 while (all) {
876 struct reada_extent *re = NULL;
877
878 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
879 index, 1);
880 if (ret == 0)
881 break;
882 if (!re->scheduled) {
883 index = (re->logical >> PAGE_SHIFT) + 1;
884 continue;
885 }
886 printk(KERN_DEBUG
887 "re: logical %llu size %u list empty %d scheduled %d",
888 re->logical, fs_info->tree_root->nodesize,
889 list_empty(&re->extctl), re->scheduled);
890 for (i = 0; i < re->nzones; ++i) {
891 printk(KERN_CONT " zone %llu-%llu devs",
892 re->zones[i]->start,
893 re->zones[i]->end);
894 for (j = 0; j < re->zones[i]->ndevs; ++j) {
895 printk(KERN_CONT " %lld",
896 re->zones[i]->devs[j]->devid);
897 }
898 }
899 printk(KERN_CONT "\n");
900 index = (re->logical >> PAGE_SHIFT) + 1;
901 }
902 spin_unlock(&fs_info->reada_lock);
903}
904#endif
905
906/*
907 * interface
908 */
909struct reada_control *btrfs_reada_add(struct btrfs_root *root,
910 struct btrfs_key *key_start, struct btrfs_key *key_end)
911{
912 struct reada_control *rc;
913 u64 start;
914 u64 generation;
915 int ret;
916 struct extent_buffer *node;
917 static struct btrfs_key max_key = {
918 .objectid = (u64)-1,
919 .type = (u8)-1,
920 .offset = (u64)-1
921 };
922
923 rc = kzalloc(sizeof(*rc), GFP_KERNEL);
924 if (!rc)
925 return ERR_PTR(-ENOMEM);
926
927 rc->root = root;
928 rc->key_start = *key_start;
929 rc->key_end = *key_end;
930 atomic_set(&rc->elems, 0);
931 init_waitqueue_head(&rc->wait);
932 kref_init(&rc->refcnt);
933 kref_get(&rc->refcnt); /* one ref for having elements */
934
935 node = btrfs_root_node(root);
936 start = node->start;
937 generation = btrfs_header_generation(node);
938 free_extent_buffer(node);
939
940 ret = reada_add_block(rc, start, &max_key, generation);
941 if (ret) {
942 kfree(rc);
943 return ERR_PTR(ret);
944 }
945
946 reada_start_machine(root->fs_info);
947
948 return rc;
949}
950
951#ifdef DEBUG
952int btrfs_reada_wait(void *handle)
953{
954 struct reada_control *rc = handle;
955 struct btrfs_fs_info *fs_info = rc->root->fs_info;
956
957 while (atomic_read(&rc->elems)) {
958 if (!atomic_read(&fs_info->reada_works_cnt))
959 reada_start_machine(fs_info);
960 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
961 5 * HZ);
962 dump_devs(rc->root->fs_info,
963 atomic_read(&rc->elems) < 10 ? 1 : 0);
964 }
965
966 dump_devs(rc->root->fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
967
968 kref_put(&rc->refcnt, reada_control_release);
969
970 return 0;
971}
972#else
973int btrfs_reada_wait(void *handle)
974{
975 struct reada_control *rc = handle;
976 struct btrfs_fs_info *fs_info = rc->root->fs_info;
977
978 while (atomic_read(&rc->elems)) {
979 if (!atomic_read(&fs_info->reada_works_cnt))
980 reada_start_machine(fs_info);
981 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
982 (HZ + 9) / 10);
983 }
984
985 kref_put(&rc->refcnt, reada_control_release);
986
987 return 0;
988}
989#endif
990
991void btrfs_reada_detach(void *handle)
992{
993 struct reada_control *rc = handle;
994
995 kref_put(&rc->refcnt, reada_control_release);
996}