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
47 struct reada_extctl {
48 struct list_head list;
49 struct reada_control *rc;
50 u64 generation;
51 };
52
53 struct reada_extent {
54 u64 logical;
55 u64 owner_root;
56 struct btrfs_key top;
57 struct list_head extctl;
58 int refcnt;
59 spinlock_t lock;
60 struct reada_zone *zones[BTRFS_MAX_MIRRORS];
61 int nzones;
62 int scheduled;
63 int level;
64 };
65
66 struct reada_zone {
67 u64 start;
68 u64 end;
69 u64 elems;
70 struct list_head list;
71 spinlock_t lock;
72 int locked;
73 struct btrfs_device *device;
74 struct btrfs_device *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
75 * self */
76 int ndevs;
77 struct kref refcnt;
78 };
79
80 struct reada_machine_work {
81 struct btrfs_work work;
82 struct btrfs_fs_info *fs_info;
83 };
84
85 static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
86 static void reada_control_release(struct kref *kref);
87 static void reada_zone_release(struct kref *kref);
88 static void reada_start_machine(struct btrfs_fs_info *fs_info);
89 static void __reada_start_machine(struct btrfs_fs_info *fs_info);
90
91 static int reada_add_block(struct reada_control *rc, u64 logical,
92 struct btrfs_key *top, u64 owner_root,
93 u64 generation, int level);
94
95 /* recurses */
96 /* in case of err, eb might be NULL */
__readahead_hook(struct btrfs_fs_info * fs_info,struct reada_extent * re,struct extent_buffer * eb,int err)97 static void __readahead_hook(struct btrfs_fs_info *fs_info,
98 struct reada_extent *re, struct extent_buffer *eb,
99 int err)
100 {
101 int nritems;
102 int i;
103 u64 bytenr;
104 u64 generation;
105 struct list_head list;
106
107 spin_lock(&re->lock);
108 /*
109 * just take the full list from the extent. afterwards we
110 * don't need the lock anymore
111 */
112 list_replace_init(&re->extctl, &list);
113 re->scheduled = 0;
114 spin_unlock(&re->lock);
115
116 /*
117 * this is the error case, the extent buffer has not been
118 * read correctly. We won't access anything from it and
119 * just cleanup our data structures. Effectively this will
120 * cut the branch below this node from read ahead.
121 */
122 if (err)
123 goto cleanup;
124
125 /*
126 * FIXME: currently we just set nritems to 0 if this is a leaf,
127 * effectively ignoring the content. In a next step we could
128 * trigger more readahead depending from the content, e.g.
129 * fetch the checksums for the extents in the leaf.
130 */
131 if (!btrfs_header_level(eb))
132 goto cleanup;
133
134 nritems = btrfs_header_nritems(eb);
135 generation = btrfs_header_generation(eb);
136 for (i = 0; i < nritems; i++) {
137 struct reada_extctl *rec;
138 u64 n_gen;
139 struct btrfs_key key;
140 struct btrfs_key next_key;
141
142 btrfs_node_key_to_cpu(eb, &key, i);
143 if (i + 1 < nritems)
144 btrfs_node_key_to_cpu(eb, &next_key, i + 1);
145 else
146 next_key = re->top;
147 bytenr = btrfs_node_blockptr(eb, i);
148 n_gen = btrfs_node_ptr_generation(eb, i);
149
150 list_for_each_entry(rec, &list, list) {
151 struct reada_control *rc = rec->rc;
152
153 /*
154 * if the generation doesn't match, just ignore this
155 * extctl. This will probably cut off a branch from
156 * prefetch. Alternatively one could start a new (sub-)
157 * prefetch for this branch, starting again from root.
158 * FIXME: move the generation check out of this loop
159 */
160 #ifdef DEBUG
161 if (rec->generation != generation) {
162 btrfs_debug(fs_info,
163 "generation mismatch for (%llu,%d,%llu) %llu != %llu",
164 key.objectid, key.type, key.offset,
165 rec->generation, generation);
166 }
167 #endif
168 if (rec->generation == generation &&
169 btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
170 btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
171 reada_add_block(rc, bytenr, &next_key,
172 btrfs_header_owner(eb), n_gen,
173 btrfs_header_level(eb) - 1);
174 }
175 }
176
177 cleanup:
178 /*
179 * free extctl records
180 */
181 while (!list_empty(&list)) {
182 struct reada_control *rc;
183 struct reada_extctl *rec;
184
185 rec = list_first_entry(&list, struct reada_extctl, list);
186 list_del(&rec->list);
187 rc = rec->rc;
188 kfree(rec);
189
190 kref_get(&rc->refcnt);
191 if (atomic_dec_and_test(&rc->elems)) {
192 kref_put(&rc->refcnt, reada_control_release);
193 wake_up(&rc->wait);
194 }
195 kref_put(&rc->refcnt, reada_control_release);
196
197 reada_extent_put(fs_info, re); /* one ref for each entry */
198 }
199
200 return;
201 }
202
btree_readahead_hook(struct extent_buffer * eb,int err)203 int btree_readahead_hook(struct extent_buffer *eb, int err)
204 {
205 struct btrfs_fs_info *fs_info = eb->fs_info;
206 int ret = 0;
207 struct reada_extent *re;
208
209 /* find extent */
210 spin_lock(&fs_info->reada_lock);
211 re = radix_tree_lookup(&fs_info->reada_tree,
212 eb->start >> fs_info->sectorsize_bits);
213 if (re)
214 re->refcnt++;
215 spin_unlock(&fs_info->reada_lock);
216 if (!re) {
217 ret = -1;
218 goto start_machine;
219 }
220
221 __readahead_hook(fs_info, re, eb, err);
222 reada_extent_put(fs_info, re); /* our ref */
223
224 start_machine:
225 reada_start_machine(fs_info);
226 return ret;
227 }
228
reada_find_zone(struct btrfs_device * dev,u64 logical,struct btrfs_io_context * bioc)229 static struct reada_zone *reada_find_zone(struct btrfs_device *dev, u64 logical,
230 struct btrfs_io_context *bioc)
231 {
232 struct btrfs_fs_info *fs_info = dev->fs_info;
233 int ret;
234 struct reada_zone *zone;
235 struct btrfs_block_group *cache = NULL;
236 u64 start;
237 u64 end;
238 int i;
239
240 zone = NULL;
241 spin_lock(&fs_info->reada_lock);
242 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
243 logical >> fs_info->sectorsize_bits, 1);
244 if (ret == 1 && logical >= zone->start && logical <= zone->end) {
245 kref_get(&zone->refcnt);
246 spin_unlock(&fs_info->reada_lock);
247 return zone;
248 }
249
250 spin_unlock(&fs_info->reada_lock);
251
252 cache = btrfs_lookup_block_group(fs_info, logical);
253 if (!cache)
254 return NULL;
255
256 start = cache->start;
257 end = start + cache->length - 1;
258 btrfs_put_block_group(cache);
259
260 zone = kzalloc(sizeof(*zone), GFP_KERNEL);
261 if (!zone)
262 return NULL;
263
264 ret = radix_tree_preload(GFP_KERNEL);
265 if (ret) {
266 kfree(zone);
267 return NULL;
268 }
269
270 zone->start = start;
271 zone->end = end;
272 INIT_LIST_HEAD(&zone->list);
273 spin_lock_init(&zone->lock);
274 zone->locked = 0;
275 kref_init(&zone->refcnt);
276 zone->elems = 0;
277 zone->device = dev; /* our device always sits at index 0 */
278 for (i = 0; i < bioc->num_stripes; ++i) {
279 /* bounds have already been checked */
280 zone->devs[i] = bioc->stripes[i].dev;
281 }
282 zone->ndevs = bioc->num_stripes;
283
284 spin_lock(&fs_info->reada_lock);
285 ret = radix_tree_insert(&dev->reada_zones,
286 (unsigned long)(zone->end >> fs_info->sectorsize_bits),
287 zone);
288
289 if (ret == -EEXIST) {
290 kfree(zone);
291 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
292 logical >> fs_info->sectorsize_bits, 1);
293 if (ret == 1 && logical >= zone->start && logical <= zone->end)
294 kref_get(&zone->refcnt);
295 else
296 zone = NULL;
297 }
298 spin_unlock(&fs_info->reada_lock);
299 radix_tree_preload_end();
300
301 return zone;
302 }
303
reada_find_extent(struct btrfs_fs_info * fs_info,u64 logical,struct btrfs_key * top,u64 owner_root,int level)304 static struct reada_extent *reada_find_extent(struct btrfs_fs_info *fs_info,
305 u64 logical,
306 struct btrfs_key *top,
307 u64 owner_root, int level)
308 {
309 int ret;
310 struct reada_extent *re = NULL;
311 struct reada_extent *re_exist = NULL;
312 struct btrfs_io_context *bioc = NULL;
313 struct btrfs_device *dev;
314 struct btrfs_device *prev_dev;
315 u64 length;
316 int real_stripes;
317 int nzones = 0;
318 unsigned long index = logical >> fs_info->sectorsize_bits;
319 int dev_replace_is_ongoing;
320 int have_zone = 0;
321
322 spin_lock(&fs_info->reada_lock);
323 re = radix_tree_lookup(&fs_info->reada_tree, index);
324 if (re)
325 re->refcnt++;
326 spin_unlock(&fs_info->reada_lock);
327
328 if (re)
329 return re;
330
331 re = kzalloc(sizeof(*re), GFP_KERNEL);
332 if (!re)
333 return NULL;
334
335 re->logical = logical;
336 re->top = *top;
337 INIT_LIST_HEAD(&re->extctl);
338 spin_lock_init(&re->lock);
339 re->refcnt = 1;
340 re->owner_root = owner_root;
341 re->level = level;
342
343 /*
344 * map block
345 */
346 length = fs_info->nodesize;
347 ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical,
348 &length, &bioc, 0);
349 if (ret || !bioc || length < fs_info->nodesize)
350 goto error;
351
352 if (bioc->num_stripes > BTRFS_MAX_MIRRORS) {
353 btrfs_err(fs_info,
354 "readahead: more than %d copies not supported",
355 BTRFS_MAX_MIRRORS);
356 goto error;
357 }
358
359 real_stripes = bioc->num_stripes - bioc->num_tgtdevs;
360 for (nzones = 0; nzones < real_stripes; ++nzones) {
361 struct reada_zone *zone;
362
363 dev = bioc->stripes[nzones].dev;
364
365 /* cannot read ahead on missing device. */
366 if (!dev->bdev)
367 continue;
368
369 zone = reada_find_zone(dev, logical, bioc);
370 if (!zone)
371 continue;
372
373 re->zones[re->nzones++] = zone;
374 spin_lock(&zone->lock);
375 if (!zone->elems)
376 kref_get(&zone->refcnt);
377 ++zone->elems;
378 spin_unlock(&zone->lock);
379 spin_lock(&fs_info->reada_lock);
380 kref_put(&zone->refcnt, reada_zone_release);
381 spin_unlock(&fs_info->reada_lock);
382 }
383 if (re->nzones == 0) {
384 /* not a single zone found, error and out */
385 goto error;
386 }
387
388 /* Insert extent in reada tree + all per-device trees, all or nothing */
389 down_read(&fs_info->dev_replace.rwsem);
390 ret = radix_tree_preload(GFP_KERNEL);
391 if (ret) {
392 up_read(&fs_info->dev_replace.rwsem);
393 goto error;
394 }
395
396 spin_lock(&fs_info->reada_lock);
397 ret = radix_tree_insert(&fs_info->reada_tree, index, re);
398 if (ret == -EEXIST) {
399 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
400 re_exist->refcnt++;
401 spin_unlock(&fs_info->reada_lock);
402 radix_tree_preload_end();
403 up_read(&fs_info->dev_replace.rwsem);
404 goto error;
405 }
406 if (ret) {
407 spin_unlock(&fs_info->reada_lock);
408 radix_tree_preload_end();
409 up_read(&fs_info->dev_replace.rwsem);
410 goto error;
411 }
412 radix_tree_preload_end();
413 prev_dev = NULL;
414 dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
415 &fs_info->dev_replace);
416 for (nzones = 0; nzones < re->nzones; ++nzones) {
417 dev = re->zones[nzones]->device;
418
419 if (dev == prev_dev) {
420 /*
421 * in case of DUP, just add the first zone. As both
422 * are on the same device, there's nothing to gain
423 * from adding both.
424 * Also, it wouldn't work, as the tree is per device
425 * and adding would fail with EEXIST
426 */
427 continue;
428 }
429 if (!dev->bdev)
430 continue;
431
432 if (test_bit(BTRFS_DEV_STATE_NO_READA, &dev->dev_state))
433 continue;
434
435 if (dev_replace_is_ongoing &&
436 dev == fs_info->dev_replace.tgtdev) {
437 /*
438 * as this device is selected for reading only as
439 * a last resort, skip it for read ahead.
440 */
441 continue;
442 }
443 prev_dev = dev;
444 ret = radix_tree_insert(&dev->reada_extents, index, re);
445 if (ret) {
446 while (--nzones >= 0) {
447 dev = re->zones[nzones]->device;
448 BUG_ON(dev == NULL);
449 /* ignore whether the entry was inserted */
450 radix_tree_delete(&dev->reada_extents, index);
451 }
452 radix_tree_delete(&fs_info->reada_tree, index);
453 spin_unlock(&fs_info->reada_lock);
454 up_read(&fs_info->dev_replace.rwsem);
455 goto error;
456 }
457 have_zone = 1;
458 }
459 if (!have_zone)
460 radix_tree_delete(&fs_info->reada_tree, index);
461 spin_unlock(&fs_info->reada_lock);
462 up_read(&fs_info->dev_replace.rwsem);
463
464 if (!have_zone)
465 goto error;
466
467 btrfs_put_bioc(bioc);
468 return re;
469
470 error:
471 for (nzones = 0; nzones < re->nzones; ++nzones) {
472 struct reada_zone *zone;
473
474 zone = re->zones[nzones];
475 kref_get(&zone->refcnt);
476 spin_lock(&zone->lock);
477 --zone->elems;
478 if (zone->elems == 0) {
479 /*
480 * no fs_info->reada_lock needed, as this can't be
481 * the last ref
482 */
483 kref_put(&zone->refcnt, reada_zone_release);
484 }
485 spin_unlock(&zone->lock);
486
487 spin_lock(&fs_info->reada_lock);
488 kref_put(&zone->refcnt, reada_zone_release);
489 spin_unlock(&fs_info->reada_lock);
490 }
491 btrfs_put_bioc(bioc);
492 kfree(re);
493 return re_exist;
494 }
495
reada_extent_put(struct btrfs_fs_info * fs_info,struct reada_extent * re)496 static void reada_extent_put(struct btrfs_fs_info *fs_info,
497 struct reada_extent *re)
498 {
499 int i;
500 unsigned long index = re->logical >> fs_info->sectorsize_bits;
501
502 spin_lock(&fs_info->reada_lock);
503 if (--re->refcnt) {
504 spin_unlock(&fs_info->reada_lock);
505 return;
506 }
507
508 radix_tree_delete(&fs_info->reada_tree, index);
509 for (i = 0; i < re->nzones; ++i) {
510 struct reada_zone *zone = re->zones[i];
511
512 radix_tree_delete(&zone->device->reada_extents, index);
513 }
514
515 spin_unlock(&fs_info->reada_lock);
516
517 for (i = 0; i < re->nzones; ++i) {
518 struct reada_zone *zone = re->zones[i];
519
520 kref_get(&zone->refcnt);
521 spin_lock(&zone->lock);
522 --zone->elems;
523 if (zone->elems == 0) {
524 /* no fs_info->reada_lock needed, as this can't be
525 * the last ref */
526 kref_put(&zone->refcnt, reada_zone_release);
527 }
528 spin_unlock(&zone->lock);
529
530 spin_lock(&fs_info->reada_lock);
531 kref_put(&zone->refcnt, reada_zone_release);
532 spin_unlock(&fs_info->reada_lock);
533 }
534
535 kfree(re);
536 }
537
reada_zone_release(struct kref * kref)538 static void reada_zone_release(struct kref *kref)
539 {
540 struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
541 struct btrfs_fs_info *fs_info = zone->device->fs_info;
542
543 lockdep_assert_held(&fs_info->reada_lock);
544
545 radix_tree_delete(&zone->device->reada_zones,
546 zone->end >> fs_info->sectorsize_bits);
547
548 kfree(zone);
549 }
550
reada_control_release(struct kref * kref)551 static void reada_control_release(struct kref *kref)
552 {
553 struct reada_control *rc = container_of(kref, struct reada_control,
554 refcnt);
555
556 kfree(rc);
557 }
558
reada_add_block(struct reada_control * rc,u64 logical,struct btrfs_key * top,u64 owner_root,u64 generation,int level)559 static int reada_add_block(struct reada_control *rc, u64 logical,
560 struct btrfs_key *top, u64 owner_root,
561 u64 generation, int level)
562 {
563 struct btrfs_fs_info *fs_info = rc->fs_info;
564 struct reada_extent *re;
565 struct reada_extctl *rec;
566
567 /* takes one ref */
568 re = reada_find_extent(fs_info, logical, top, owner_root, level);
569 if (!re)
570 return -1;
571
572 rec = kzalloc(sizeof(*rec), GFP_KERNEL);
573 if (!rec) {
574 reada_extent_put(fs_info, re);
575 return -ENOMEM;
576 }
577
578 rec->rc = rc;
579 rec->generation = generation;
580 atomic_inc(&rc->elems);
581
582 spin_lock(&re->lock);
583 list_add_tail(&rec->list, &re->extctl);
584 spin_unlock(&re->lock);
585
586 /* leave the ref on the extent */
587
588 return 0;
589 }
590
591 /*
592 * called with fs_info->reada_lock held
593 */
reada_peer_zones_set_lock(struct reada_zone * zone,int lock)594 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
595 {
596 int i;
597 unsigned long index = zone->end >> zone->device->fs_info->sectorsize_bits;
598
599 for (i = 0; i < zone->ndevs; ++i) {
600 struct reada_zone *peer;
601 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
602 if (peer && peer->device != zone->device)
603 peer->locked = lock;
604 }
605 }
606
607 /*
608 * called with fs_info->reada_lock held
609 */
reada_pick_zone(struct btrfs_device * dev)610 static int reada_pick_zone(struct btrfs_device *dev)
611 {
612 struct reada_zone *top_zone = NULL;
613 struct reada_zone *top_locked_zone = NULL;
614 u64 top_elems = 0;
615 u64 top_locked_elems = 0;
616 unsigned long index = 0;
617 int ret;
618
619 if (dev->reada_curr_zone) {
620 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
621 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
622 dev->reada_curr_zone = NULL;
623 }
624 /* pick the zone with the most elements */
625 while (1) {
626 struct reada_zone *zone;
627
628 ret = radix_tree_gang_lookup(&dev->reada_zones,
629 (void **)&zone, index, 1);
630 if (ret == 0)
631 break;
632 index = (zone->end >> dev->fs_info->sectorsize_bits) + 1;
633 if (zone->locked) {
634 if (zone->elems > top_locked_elems) {
635 top_locked_elems = zone->elems;
636 top_locked_zone = zone;
637 }
638 } else {
639 if (zone->elems > top_elems) {
640 top_elems = zone->elems;
641 top_zone = zone;
642 }
643 }
644 }
645 if (top_zone)
646 dev->reada_curr_zone = top_zone;
647 else if (top_locked_zone)
648 dev->reada_curr_zone = top_locked_zone;
649 else
650 return 0;
651
652 dev->reada_next = dev->reada_curr_zone->start;
653 kref_get(&dev->reada_curr_zone->refcnt);
654 reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
655
656 return 1;
657 }
658
reada_tree_block_flagged(struct btrfs_fs_info * fs_info,u64 bytenr,u64 owner_root,int level,int mirror_num,struct extent_buffer ** eb)659 static int reada_tree_block_flagged(struct btrfs_fs_info *fs_info, u64 bytenr,
660 u64 owner_root, int level, int mirror_num,
661 struct extent_buffer **eb)
662 {
663 struct extent_buffer *buf = NULL;
664 int ret;
665
666 buf = btrfs_find_create_tree_block(fs_info, bytenr, owner_root, level);
667 if (IS_ERR(buf))
668 return 0;
669
670 set_bit(EXTENT_BUFFER_READAHEAD, &buf->bflags);
671
672 ret = read_extent_buffer_pages(buf, WAIT_PAGE_LOCK, mirror_num);
673 if (ret) {
674 free_extent_buffer_stale(buf);
675 return ret;
676 }
677
678 if (test_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags)) {
679 free_extent_buffer_stale(buf);
680 return -EIO;
681 } else if (extent_buffer_uptodate(buf)) {
682 *eb = buf;
683 } else {
684 free_extent_buffer(buf);
685 }
686 return 0;
687 }
688
reada_start_machine_dev(struct btrfs_device * dev)689 static int reada_start_machine_dev(struct btrfs_device *dev)
690 {
691 struct btrfs_fs_info *fs_info = dev->fs_info;
692 struct reada_extent *re = NULL;
693 int mirror_num = 0;
694 struct extent_buffer *eb = NULL;
695 u64 logical;
696 int ret;
697 int i;
698
699 spin_lock(&fs_info->reada_lock);
700 if (dev->reada_curr_zone == NULL) {
701 ret = reada_pick_zone(dev);
702 if (!ret) {
703 spin_unlock(&fs_info->reada_lock);
704 return 0;
705 }
706 }
707 /*
708 * FIXME currently we issue the reads one extent at a time. If we have
709 * a contiguous block of extents, we could also coagulate them or use
710 * plugging to speed things up
711 */
712 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
713 dev->reada_next >> fs_info->sectorsize_bits, 1);
714 if (ret == 0 || re->logical > dev->reada_curr_zone->end) {
715 ret = reada_pick_zone(dev);
716 if (!ret) {
717 spin_unlock(&fs_info->reada_lock);
718 return 0;
719 }
720 re = NULL;
721 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
722 dev->reada_next >> fs_info->sectorsize_bits, 1);
723 }
724 if (ret == 0) {
725 spin_unlock(&fs_info->reada_lock);
726 return 0;
727 }
728 dev->reada_next = re->logical + fs_info->nodesize;
729 re->refcnt++;
730
731 spin_unlock(&fs_info->reada_lock);
732
733 spin_lock(&re->lock);
734 if (re->scheduled || list_empty(&re->extctl)) {
735 spin_unlock(&re->lock);
736 reada_extent_put(fs_info, re);
737 return 0;
738 }
739 re->scheduled = 1;
740 spin_unlock(&re->lock);
741
742 /*
743 * find mirror num
744 */
745 for (i = 0; i < re->nzones; ++i) {
746 if (re->zones[i]->device == dev) {
747 mirror_num = i + 1;
748 break;
749 }
750 }
751 logical = re->logical;
752
753 atomic_inc(&dev->reada_in_flight);
754 ret = reada_tree_block_flagged(fs_info, logical, re->owner_root,
755 re->level, mirror_num, &eb);
756 if (ret)
757 __readahead_hook(fs_info, re, NULL, ret);
758 else if (eb)
759 __readahead_hook(fs_info, re, eb, ret);
760
761 if (eb)
762 free_extent_buffer(eb);
763
764 atomic_dec(&dev->reada_in_flight);
765 reada_extent_put(fs_info, re);
766
767 return 1;
768
769 }
770
reada_start_machine_worker(struct btrfs_work * work)771 static void reada_start_machine_worker(struct btrfs_work *work)
772 {
773 struct reada_machine_work *rmw;
774 int old_ioprio;
775
776 rmw = container_of(work, struct reada_machine_work, work);
777
778 old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
779 task_nice_ioprio(current));
780 set_task_ioprio(current, BTRFS_IOPRIO_READA);
781 __reada_start_machine(rmw->fs_info);
782 set_task_ioprio(current, old_ioprio);
783
784 atomic_dec(&rmw->fs_info->reada_works_cnt);
785
786 kfree(rmw);
787 }
788
789 /* Try to start up to 10k READA requests for a group of devices */
reada_start_for_fsdevs(struct btrfs_fs_devices * fs_devices)790 static int reada_start_for_fsdevs(struct btrfs_fs_devices *fs_devices)
791 {
792 u64 enqueued;
793 u64 total = 0;
794 struct btrfs_device *device;
795
796 do {
797 enqueued = 0;
798 list_for_each_entry(device, &fs_devices->devices, dev_list) {
799 if (atomic_read(&device->reada_in_flight) <
800 MAX_IN_FLIGHT)
801 enqueued += reada_start_machine_dev(device);
802 }
803 total += enqueued;
804 } while (enqueued && total < 10000);
805
806 return total;
807 }
808
__reada_start_machine(struct btrfs_fs_info * fs_info)809 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
810 {
811 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices, *seed_devs;
812 int i;
813 u64 enqueued = 0;
814
815 mutex_lock(&fs_devices->device_list_mutex);
816
817 enqueued += reada_start_for_fsdevs(fs_devices);
818 list_for_each_entry(seed_devs, &fs_devices->seed_list, seed_list)
819 enqueued += reada_start_for_fsdevs(seed_devs);
820
821 mutex_unlock(&fs_devices->device_list_mutex);
822 if (enqueued == 0)
823 return;
824
825 /*
826 * If everything is already in the cache, this is effectively single
827 * threaded. To a) not hold the caller for too long and b) to utilize
828 * more cores, we broke the loop above after 10000 iterations and now
829 * enqueue to workers to finish it. This will distribute the load to
830 * the cores.
831 */
832 for (i = 0; i < 2; ++i) {
833 reada_start_machine(fs_info);
834 if (atomic_read(&fs_info->reada_works_cnt) >
835 BTRFS_MAX_MIRRORS * 2)
836 break;
837 }
838 }
839
reada_start_machine(struct btrfs_fs_info * fs_info)840 static void reada_start_machine(struct btrfs_fs_info *fs_info)
841 {
842 struct reada_machine_work *rmw;
843
844 rmw = kzalloc(sizeof(*rmw), GFP_KERNEL);
845 if (!rmw) {
846 /* FIXME we cannot handle this properly right now */
847 BUG();
848 }
849 btrfs_init_work(&rmw->work, reada_start_machine_worker, NULL, NULL);
850 rmw->fs_info = fs_info;
851
852 btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
853 atomic_inc(&fs_info->reada_works_cnt);
854 }
855
856 #ifdef DEBUG
dump_devs(struct btrfs_fs_info * fs_info,int all)857 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
858 {
859 struct btrfs_device *device;
860 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
861 unsigned long index;
862 int ret;
863 int i;
864 int j;
865 int cnt;
866
867 spin_lock(&fs_info->reada_lock);
868 list_for_each_entry(device, &fs_devices->devices, dev_list) {
869 btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid,
870 atomic_read(&device->reada_in_flight));
871 index = 0;
872 while (1) {
873 struct reada_zone *zone;
874 ret = radix_tree_gang_lookup(&device->reada_zones,
875 (void **)&zone, index, 1);
876 if (ret == 0)
877 break;
878 pr_debug(" zone %llu-%llu elems %llu locked %d devs",
879 zone->start, zone->end, zone->elems,
880 zone->locked);
881 for (j = 0; j < zone->ndevs; ++j) {
882 pr_cont(" %lld",
883 zone->devs[j]->devid);
884 }
885 if (device->reada_curr_zone == zone)
886 pr_cont(" curr off %llu",
887 device->reada_next - zone->start);
888 pr_cont("\n");
889 index = (zone->end >> fs_info->sectorsize_bits) + 1;
890 }
891 cnt = 0;
892 index = 0;
893 while (all) {
894 struct reada_extent *re = NULL;
895
896 ret = radix_tree_gang_lookup(&device->reada_extents,
897 (void **)&re, index, 1);
898 if (ret == 0)
899 break;
900 pr_debug(" re: logical %llu size %u empty %d scheduled %d",
901 re->logical, fs_info->nodesize,
902 list_empty(&re->extctl), re->scheduled);
903
904 for (i = 0; i < re->nzones; ++i) {
905 pr_cont(" zone %llu-%llu devs",
906 re->zones[i]->start,
907 re->zones[i]->end);
908 for (j = 0; j < re->zones[i]->ndevs; ++j) {
909 pr_cont(" %lld",
910 re->zones[i]->devs[j]->devid);
911 }
912 }
913 pr_cont("\n");
914 index = (re->logical >> fs_info->sectorsize_bits) + 1;
915 if (++cnt > 15)
916 break;
917 }
918 }
919
920 index = 0;
921 cnt = 0;
922 while (all) {
923 struct reada_extent *re = NULL;
924
925 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
926 index, 1);
927 if (ret == 0)
928 break;
929 if (!re->scheduled) {
930 index = (re->logical >> fs_info->sectorsize_bits) + 1;
931 continue;
932 }
933 pr_debug("re: logical %llu size %u list empty %d scheduled %d",
934 re->logical, fs_info->nodesize,
935 list_empty(&re->extctl), re->scheduled);
936 for (i = 0; i < re->nzones; ++i) {
937 pr_cont(" zone %llu-%llu devs",
938 re->zones[i]->start,
939 re->zones[i]->end);
940 for (j = 0; j < re->zones[i]->ndevs; ++j) {
941 pr_cont(" %lld",
942 re->zones[i]->devs[j]->devid);
943 }
944 }
945 pr_cont("\n");
946 index = (re->logical >> fs_info->sectorsize_bits) + 1;
947 }
948 spin_unlock(&fs_info->reada_lock);
949 }
950 #endif
951
952 /*
953 * interface
954 */
btrfs_reada_add(struct btrfs_root * root,struct btrfs_key * key_start,struct btrfs_key * key_end)955 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
956 struct btrfs_key *key_start, struct btrfs_key *key_end)
957 {
958 struct reada_control *rc;
959 u64 start;
960 u64 generation;
961 int ret;
962 int level;
963 struct extent_buffer *node;
964 static struct btrfs_key max_key = {
965 .objectid = (u64)-1,
966 .type = (u8)-1,
967 .offset = (u64)-1
968 };
969
970 rc = kzalloc(sizeof(*rc), GFP_KERNEL);
971 if (!rc)
972 return ERR_PTR(-ENOMEM);
973
974 rc->fs_info = root->fs_info;
975 rc->key_start = *key_start;
976 rc->key_end = *key_end;
977 atomic_set(&rc->elems, 0);
978 init_waitqueue_head(&rc->wait);
979 kref_init(&rc->refcnt);
980 kref_get(&rc->refcnt); /* one ref for having elements */
981
982 node = btrfs_root_node(root);
983 start = node->start;
984 generation = btrfs_header_generation(node);
985 level = btrfs_header_level(node);
986 free_extent_buffer(node);
987
988 ret = reada_add_block(rc, start, &max_key, root->root_key.objectid,
989 generation, level);
990 if (ret) {
991 kfree(rc);
992 return ERR_PTR(ret);
993 }
994
995 reada_start_machine(root->fs_info);
996
997 return rc;
998 }
999
1000 #ifdef DEBUG
btrfs_reada_wait(void * handle)1001 int btrfs_reada_wait(void *handle)
1002 {
1003 struct reada_control *rc = handle;
1004 struct btrfs_fs_info *fs_info = rc->fs_info;
1005
1006 while (atomic_read(&rc->elems)) {
1007 if (!atomic_read(&fs_info->reada_works_cnt))
1008 reada_start_machine(fs_info);
1009 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
1010 5 * HZ);
1011 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
1012 }
1013
1014 dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
1015
1016 kref_put(&rc->refcnt, reada_control_release);
1017
1018 return 0;
1019 }
1020 #else
btrfs_reada_wait(void * handle)1021 int btrfs_reada_wait(void *handle)
1022 {
1023 struct reada_control *rc = handle;
1024 struct btrfs_fs_info *fs_info = rc->fs_info;
1025
1026 while (atomic_read(&rc->elems)) {
1027 if (!atomic_read(&fs_info->reada_works_cnt))
1028 reada_start_machine(fs_info);
1029 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
1030 (HZ + 9) / 10);
1031 }
1032
1033 kref_put(&rc->refcnt, reada_control_release);
1034
1035 return 0;
1036 }
1037 #endif
1038
btrfs_reada_detach(void * handle)1039 void btrfs_reada_detach(void *handle)
1040 {
1041 struct reada_control *rc = handle;
1042
1043 kref_put(&rc->refcnt, reada_control_release);
1044 }
1045
1046 /*
1047 * Before removing a device (device replace or device remove ioctls), call this
1048 * function to wait for all existing readahead requests on the device and to
1049 * make sure no one queues more readahead requests for the device.
1050 *
1051 * Must be called without holding neither the device list mutex nor the device
1052 * replace semaphore, otherwise it will deadlock.
1053 */
btrfs_reada_remove_dev(struct btrfs_device * dev)1054 void btrfs_reada_remove_dev(struct btrfs_device *dev)
1055 {
1056 struct btrfs_fs_info *fs_info = dev->fs_info;
1057
1058 /* Serialize with readahead extent creation at reada_find_extent(). */
1059 spin_lock(&fs_info->reada_lock);
1060 set_bit(BTRFS_DEV_STATE_NO_READA, &dev->dev_state);
1061 spin_unlock(&fs_info->reada_lock);
1062
1063 /*
1064 * There might be readahead requests added to the radix trees which
1065 * were not yet added to the readahead work queue. We need to start
1066 * them and wait for their completion, otherwise we can end up with
1067 * use-after-free problems when dropping the last reference on the
1068 * readahead extents and their zones, as they need to access the
1069 * device structure.
1070 */
1071 reada_start_machine(fs_info);
1072 btrfs_flush_workqueue(fs_info->readahead_workers);
1073 }
1074
1075 /*
1076 * If when removing a device (device replace or device remove ioctls) an error
1077 * happens after calling btrfs_reada_remove_dev(), call this to undo what that
1078 * function did. This is safe to call even if btrfs_reada_remove_dev() was not
1079 * called before.
1080 */
btrfs_reada_undo_remove_dev(struct btrfs_device * dev)1081 void btrfs_reada_undo_remove_dev(struct btrfs_device *dev)
1082 {
1083 spin_lock(&dev->fs_info->reada_lock);
1084 clear_bit(BTRFS_DEV_STATE_NO_READA, &dev->dev_state);
1085 spin_unlock(&dev->fs_info->reada_lock);
1086 }
1087