Line data Source code
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 :
60 : struct reada_extctl {
61 : struct list_head list;
62 : struct reada_control *rc;
63 : u64 generation;
64 : };
65 :
66 : struct reada_extent {
67 : u64 logical;
68 : struct btrfs_key top;
69 : u32 blocksize;
70 : int err;
71 : struct list_head extctl;
72 : int refcnt;
73 : spinlock_t lock;
74 : struct reada_zone *zones[BTRFS_MAX_MIRRORS];
75 : int nzones;
76 : struct btrfs_device *scheduled_for;
77 : };
78 :
79 : struct reada_zone {
80 : u64 start;
81 : u64 end;
82 : u64 elems;
83 : struct list_head list;
84 : spinlock_t lock;
85 : int locked;
86 : struct btrfs_device *device;
87 : struct btrfs_device *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
88 : * self */
89 : int ndevs;
90 : struct kref refcnt;
91 : };
92 :
93 : struct reada_machine_work {
94 : struct btrfs_work work;
95 : struct btrfs_fs_info *fs_info;
96 : };
97 :
98 : static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
99 : static void reada_control_release(struct kref *kref);
100 : static void reada_zone_release(struct kref *kref);
101 : static void reada_start_machine(struct btrfs_fs_info *fs_info);
102 : static void __reada_start_machine(struct btrfs_fs_info *fs_info);
103 :
104 : static int reada_add_block(struct reada_control *rc, u64 logical,
105 : struct btrfs_key *top, int level, u64 generation);
106 :
107 : /* recurses */
108 : /* in case of err, eb might be NULL */
109 61312 : static int __readahead_hook(struct btrfs_root *root, 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 reada_extent *re;
118 15890 : struct btrfs_fs_info *fs_info = root->fs_info;
119 : struct list_head list;
120 15890 : unsigned long index = start >> PAGE_CACHE_SHIFT;
121 : struct btrfs_device *for_dev;
122 :
123 15890 : if (eb)
124 15889 : level = btrfs_header_level(eb);
125 :
126 : /* find extent */
127 : spin_lock(&fs_info->reada_lock);
128 15891 : re = radix_tree_lookup(&fs_info->reada_tree, index);
129 15891 : if (re)
130 15825 : re->refcnt++;
131 : spin_unlock(&fs_info->reada_lock);
132 :
133 15891 : if (!re)
134 : return -1;
135 :
136 : spin_lock(&re->lock);
137 : /*
138 : * just take the full list from the extent. afterwards we
139 : * don't need the lock anymore
140 : */
141 15825 : list_replace_init(&re->extctl, &list);
142 15825 : for_dev = re->scheduled_for;
143 15825 : re->scheduled_for = NULL;
144 : spin_unlock(&re->lock);
145 :
146 15855 : if (err == 0) {
147 29533 : nritems = level ? btrfs_header_nritems(eb) : 0;
148 : generation = btrfs_header_generation(eb);
149 : /*
150 : * FIXME: currently we just set nritems to 0 if this is a leaf,
151 : * effectively ignoring the content. In a next step we could
152 : * trigger more readahead depending from the content, e.g.
153 : * fetch the checksums for the extents in the leaf.
154 : */
155 : } else {
156 : /*
157 : * this is the error case, the extent buffer has not been
158 : * read correctly. We won't access anything from it and
159 : * just cleanup our data structures. Effectively this will
160 : * cut the branch below this node from read ahead.
161 : */
162 : nritems = 0;
163 : generation = 0;
164 : }
165 :
166 354306 : for (i = 0; i < nritems; i++) {
167 : struct reada_extctl *rec;
168 : u64 n_gen;
169 : struct btrfs_key key;
170 : struct btrfs_key next_key;
171 :
172 : btrfs_node_key_to_cpu(eb, &key, i);
173 342788 : if (i + 1 < nritems)
174 : btrfs_node_key_to_cpu(eb, &next_key, i + 1);
175 : else
176 13708 : next_key = re->top;
177 : bytenr = btrfs_node_blockptr(eb, i);
178 : n_gen = btrfs_node_ptr_generation(eb, i);
179 :
180 347858 : list_for_each_entry(rec, &list, list) {
181 9407 : struct reada_control *rc = rec->rc;
182 :
183 : /*
184 : * if the generation doesn't match, just ignore this
185 : * extctl. This will probably cut off a branch from
186 : * prefetch. Alternatively one could start a new (sub-)
187 : * prefetch for this branch, starting again from root.
188 : * FIXME: move the generation check out of this loop
189 : */
190 : #ifdef DEBUG
191 : if (rec->generation != generation) {
192 : btrfs_debug(root->fs_info,
193 : "generation mismatch for (%llu,%d,%llu) %llu != %llu",
194 : key.objectid, key.type, key.offset,
195 : rec->generation, generation);
196 : }
197 : #endif
198 18812 : if (rec->generation == generation &&
199 12792 : btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
200 3383 : btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
201 1762 : reada_add_block(rc, bytenr, &next_key,
202 : level - 1, n_gen);
203 : }
204 : }
205 : /*
206 : * free extctl records
207 : */
208 17813 : while (!list_empty(&list)) {
209 : struct reada_control *rc;
210 : struct reada_extctl *rec;
211 :
212 : rec = list_first_entry(&list, struct reada_extctl, list);
213 1988 : list_del(&rec->list);
214 1988 : rc = rec->rc;
215 1988 : kfree(rec);
216 :
217 1987 : kref_get(&rc->refcnt);
218 3978 : if (atomic_dec_and_test(&rc->elems)) {
219 : kref_put(&rc->refcnt, reada_control_release);
220 226 : wake_up(&rc->wait);
221 : }
222 : kref_put(&rc->refcnt, reada_control_release);
223 :
224 1988 : reada_extent_put(fs_info, re); /* one ref for each entry */
225 : }
226 15825 : reada_extent_put(fs_info, re); /* our ref */
227 15825 : if (for_dev)
228 15825 : atomic_dec(&for_dev->reada_in_flight);
229 :
230 : return 0;
231 : }
232 :
233 : /*
234 : * start is passed separately in case eb in NULL, which may be the case with
235 : * failed I/O
236 : */
237 144 : int btree_readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
238 : u64 start, int err)
239 : {
240 : int ret;
241 :
242 144 : ret = __readahead_hook(root, eb, start, err);
243 :
244 144 : reada_start_machine(root->fs_info);
245 :
246 144 : return ret;
247 : }
248 :
249 4204 : static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
250 : struct btrfs_device *dev, u64 logical,
251 : struct btrfs_bio *bbio)
252 : {
253 : int ret;
254 : struct reada_zone *zone;
255 : struct btrfs_block_group_cache *cache = NULL;
256 : u64 start;
257 : u64 end;
258 : int i;
259 :
260 4204 : zone = NULL;
261 : spin_lock(&fs_info->reada_lock);
262 4206 : ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
263 4206 : logical >> PAGE_CACHE_SHIFT, 1);
264 4206 : if (ret == 1)
265 3972 : kref_get(&zone->refcnt);
266 : spin_unlock(&fs_info->reada_lock);
267 :
268 4206 : if (ret == 1) {
269 3972 : if (logical >= zone->start && logical < zone->end)
270 : return zone;
271 : spin_lock(&fs_info->reada_lock);
272 0 : kref_put(&zone->refcnt, reada_zone_release);
273 : spin_unlock(&fs_info->reada_lock);
274 : }
275 :
276 234 : cache = btrfs_lookup_block_group(fs_info, logical);
277 234 : if (!cache)
278 : return NULL;
279 :
280 234 : start = cache->key.objectid;
281 234 : end = start + cache->key.offset - 1;
282 234 : btrfs_put_block_group(cache);
283 :
284 234 : zone = kzalloc(sizeof(*zone), GFP_NOFS);
285 234 : if (!zone)
286 : return NULL;
287 :
288 234 : zone->start = start;
289 234 : zone->end = end;
290 234 : INIT_LIST_HEAD(&zone->list);
291 234 : spin_lock_init(&zone->lock);
292 234 : zone->locked = 0;
293 : kref_init(&zone->refcnt);
294 234 : zone->elems = 0;
295 234 : zone->device = dev; /* our device always sits at index 0 */
296 675 : for (i = 0; i < bbio->num_stripes; ++i) {
297 : /* bounds have already been checked */
298 441 : zone->devs[i] = bbio->stripes[i].dev;
299 : }
300 234 : zone->ndevs = bbio->num_stripes;
301 :
302 : spin_lock(&fs_info->reada_lock);
303 234 : ret = radix_tree_insert(&dev->reada_zones,
304 234 : (unsigned long)(zone->end >> PAGE_CACHE_SHIFT),
305 : zone);
306 :
307 234 : if (ret == -EEXIST) {
308 0 : kfree(zone);
309 0 : ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
310 : logical >> PAGE_CACHE_SHIFT, 1);
311 0 : if (ret == 1)
312 0 : kref_get(&zone->refcnt);
313 : }
314 : spin_unlock(&fs_info->reada_lock);
315 :
316 234 : return zone;
317 : }
318 :
319 3975 : static struct reada_extent *reada_find_extent(struct btrfs_root *root,
320 : u64 logical,
321 : struct btrfs_key *top, int level)
322 : {
323 : int ret;
324 : struct reada_extent *re = NULL;
325 : struct reada_extent *re_exist = NULL;
326 1988 : struct btrfs_fs_info *fs_info = root->fs_info;
327 1988 : struct btrfs_bio *bbio = NULL;
328 : struct btrfs_device *dev;
329 : struct btrfs_device *prev_dev;
330 : u32 blocksize;
331 : u64 length;
332 : int nzones = 0;
333 : int i;
334 1988 : unsigned long index = logical >> PAGE_CACHE_SHIFT;
335 : int dev_replace_is_ongoing;
336 :
337 : spin_lock(&fs_info->reada_lock);
338 1989 : re = radix_tree_lookup(&fs_info->reada_tree, index);
339 1989 : if (re)
340 1 : re->refcnt++;
341 : spin_unlock(&fs_info->reada_lock);
342 :
343 1989 : if (re)
344 : return re;
345 :
346 1988 : re = kzalloc(sizeof(*re), GFP_NOFS);
347 1987 : if (!re)
348 : return NULL;
349 :
350 : blocksize = btrfs_level_size(root, level);
351 1987 : re->logical = logical;
352 1987 : re->blocksize = blocksize;
353 1987 : re->top = *top;
354 1987 : INIT_LIST_HEAD(&re->extctl);
355 1987 : spin_lock_init(&re->lock);
356 1987 : re->refcnt = 1;
357 :
358 : /*
359 : * map block
360 : */
361 1987 : length = blocksize;
362 1987 : ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical, &length,
363 : &bbio, 0);
364 1986 : if (ret || !bbio || length < blocksize)
365 : goto error;
366 :
367 1988 : if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
368 0 : btrfs_err(root->fs_info,
369 : "readahead: more than %d copies not supported",
370 : BTRFS_MAX_MIRRORS);
371 0 : goto error;
372 : }
373 :
374 4206 : for (nzones = 0; nzones < bbio->num_stripes; ++nzones) {
375 : struct reada_zone *zone;
376 :
377 4204 : dev = bbio->stripes[nzones].dev;
378 4204 : zone = reada_find_zone(fs_info, dev, logical, bbio);
379 4206 : if (!zone)
380 : break;
381 :
382 4206 : re->zones[nzones] = zone;
383 : spin_lock(&zone->lock);
384 4206 : if (!zone->elems)
385 234 : kref_get(&zone->refcnt);
386 4206 : ++zone->elems;
387 : spin_unlock(&zone->lock);
388 : spin_lock(&fs_info->reada_lock);
389 4206 : kref_put(&zone->refcnt, reada_zone_release);
390 : spin_unlock(&fs_info->reada_lock);
391 : }
392 1990 : re->nzones = nzones;
393 1990 : if (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 1988 : btrfs_dev_replace_lock(&fs_info->dev_replace);
400 : spin_lock(&fs_info->reada_lock);
401 1988 : ret = radix_tree_insert(&fs_info->reada_tree, index, re);
402 1988 : if (ret == -EEXIST) {
403 0 : re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
404 0 : BUG_ON(!re_exist);
405 0 : re_exist->refcnt++;
406 : spin_unlock(&fs_info->reada_lock);
407 0 : btrfs_dev_replace_unlock(&fs_info->dev_replace);
408 0 : goto error;
409 : }
410 1988 : if (ret) {
411 : spin_unlock(&fs_info->reada_lock);
412 0 : btrfs_dev_replace_unlock(&fs_info->dev_replace);
413 0 : goto error;
414 : }
415 : prev_dev = NULL;
416 1988 : dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing(
417 : &fs_info->dev_replace);
418 6194 : for (i = 0; i < nzones; ++i) {
419 4206 : dev = bbio->stripes[i].dev;
420 4206 : if (dev == prev_dev) {
421 : /*
422 : * in case of DUP, just add the first zone. As both
423 : * are on the same device, there's nothing to gain
424 : * from adding both.
425 : * Also, it wouldn't work, as the tree is per device
426 : * and adding would fail with EEXIST
427 : */
428 622 : continue;
429 : }
430 3584 : if (!dev->bdev) {
431 : /*
432 : * cannot read ahead on missing device, but for RAID5/6,
433 : * REQ_GET_READ_MIRRORS return 1. So don't skip missing
434 : * device for such case.
435 : */
436 0 : if (nzones > 1)
437 0 : continue;
438 : }
439 5201 : if (dev_replace_is_ongoing &&
440 1617 : dev == fs_info->dev_replace.tgtdev) {
441 : /*
442 : * as this device is selected for reading only as
443 : * a last resort, skip it for read ahead.
444 : */
445 508 : continue;
446 : }
447 : prev_dev = dev;
448 3076 : ret = radix_tree_insert(&dev->reada_extents, index, re);
449 3076 : if (ret) {
450 0 : while (--i >= 0) {
451 0 : dev = bbio->stripes[i].dev;
452 0 : BUG_ON(dev == NULL);
453 : /* ignore whether the entry was inserted */
454 0 : radix_tree_delete(&dev->reada_extents, index);
455 : }
456 0 : BUG_ON(fs_info == NULL);
457 0 : radix_tree_delete(&fs_info->reada_tree, index);
458 : spin_unlock(&fs_info->reada_lock);
459 0 : btrfs_dev_replace_unlock(&fs_info->dev_replace);
460 0 : goto error;
461 : }
462 : }
463 : spin_unlock(&fs_info->reada_lock);
464 1988 : btrfs_dev_replace_unlock(&fs_info->dev_replace);
465 :
466 1988 : kfree(bbio);
467 1988 : return re;
468 :
469 : error:
470 0 : while (nzones) {
471 : struct reada_zone *zone;
472 :
473 0 : --nzones;
474 0 : zone = re->zones[nzones];
475 0 : kref_get(&zone->refcnt);
476 : spin_lock(&zone->lock);
477 0 : --zone->elems;
478 0 : 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 0 : kfree(bbio);
492 0 : kfree(re);
493 0 : return re_exist;
494 : }
495 :
496 33872 : static void reada_extent_put(struct btrfs_fs_info *fs_info,
497 : struct reada_extent *re)
498 : {
499 : int i;
500 33872 : unsigned long index = re->logical >> PAGE_CACHE_SHIFT;
501 :
502 : spin_lock(&fs_info->reada_lock);
503 33932 : if (--re->refcnt) {
504 : spin_unlock(&fs_info->reada_lock);
505 33930 : return;
506 : }
507 :
508 1988 : radix_tree_delete(&fs_info->reada_tree, index);
509 6194 : for (i = 0; i < re->nzones; ++i) {
510 4206 : struct reada_zone *zone = re->zones[i];
511 :
512 4206 : radix_tree_delete(&zone->device->reada_extents, index);
513 : }
514 :
515 : spin_unlock(&fs_info->reada_lock);
516 :
517 6194 : for (i = 0; i < re->nzones; ++i) {
518 4206 : struct reada_zone *zone = re->zones[i];
519 :
520 4206 : kref_get(&zone->refcnt);
521 : spin_lock(&zone->lock);
522 4206 : --zone->elems;
523 4206 : 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 1988 : if (re->scheduled_for)
535 66 : atomic_dec(&re->scheduled_for->reada_in_flight);
536 :
537 1988 : kfree(re);
538 : }
539 :
540 234 : static void reada_zone_release(struct kref *kref)
541 : {
542 234 : struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
543 :
544 234 : radix_tree_delete(&zone->device->reada_zones,
545 234 : zone->end >> PAGE_CACHE_SHIFT);
546 :
547 234 : kfree(zone);
548 234 : }
549 :
550 226 : static void reada_control_release(struct kref *kref)
551 : {
552 226 : struct reada_control *rc = container_of(kref, struct reada_control,
553 : refcnt);
554 :
555 226 : kfree(rc);
556 226 : }
557 :
558 1988 : static int reada_add_block(struct reada_control *rc, u64 logical,
559 : struct btrfs_key *top, int level, u64 generation)
560 : {
561 1988 : struct btrfs_root *root = rc->root;
562 : struct reada_extent *re;
563 : struct reada_extctl *rec;
564 :
565 1988 : re = reada_find_extent(root, logical, top, level); /* takes one ref */
566 1989 : if (!re)
567 : return -1;
568 :
569 1989 : rec = kzalloc(sizeof(*rec), GFP_NOFS);
570 1989 : if (!rec) {
571 0 : reada_extent_put(root->fs_info, re);
572 0 : return -1;
573 : }
574 :
575 1989 : rec->rc = rc;
576 1989 : rec->generation = generation;
577 1989 : atomic_inc(&rc->elems);
578 :
579 : spin_lock(&re->lock);
580 1989 : list_add_tail(&rec->list, &re->extctl);
581 : spin_unlock(&re->lock);
582 :
583 : /* leave the ref on the extent */
584 :
585 1989 : return 0;
586 : }
587 :
588 : /*
589 : * called with fs_info->reada_lock held
590 : */
591 32334 : static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
592 : {
593 : int i;
594 32334 : unsigned long index = zone->end >> PAGE_CACHE_SHIFT;
595 :
596 103026 : for (i = 0; i < zone->ndevs; ++i) {
597 : struct reada_zone *peer;
598 70692 : peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
599 70692 : if (peer && peer->device != zone->device)
600 15331 : peer->locked = lock;
601 : }
602 32334 : }
603 :
604 : /*
605 : * called with fs_info->reada_lock held
606 : */
607 26952 : static int reada_pick_zone(struct btrfs_device *dev)
608 : {
609 : struct reada_zone *top_zone = NULL;
610 : struct reada_zone *top_locked_zone = NULL;
611 : u64 top_elems = 0;
612 : u64 top_locked_elems = 0;
613 : unsigned long index = 0;
614 : int ret;
615 :
616 26952 : if (dev->reada_curr_zone) {
617 16167 : reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
618 16167 : kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
619 16167 : dev->reada_curr_zone = NULL;
620 : }
621 : /* pick the zone with the most elements */
622 : while (1) {
623 : struct reada_zone *zone;
624 :
625 43138 : ret = radix_tree_gang_lookup(&dev->reada_zones,
626 : (void **)&zone, index, 1);
627 43138 : if (ret == 0)
628 : break;
629 16186 : index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
630 16186 : if (zone->locked) {
631 5330 : if (zone->elems > top_locked_elems) {
632 : top_locked_elems = zone->elems;
633 : top_locked_zone = zone;
634 : }
635 : } else {
636 10856 : if (zone->elems > top_elems) {
637 : top_elems = zone->elems;
638 : top_zone = zone;
639 : }
640 : }
641 16186 : }
642 26952 : if (top_zone)
643 10845 : dev->reada_curr_zone = top_zone;
644 16107 : else if (top_locked_zone)
645 5322 : dev->reada_curr_zone = top_locked_zone;
646 : else
647 : return 0;
648 :
649 16167 : dev->reada_next = dev->reada_curr_zone->start;
650 16167 : kref_get(&dev->reada_curr_zone->refcnt);
651 16167 : reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
652 :
653 16167 : return 1;
654 : }
655 :
656 28911 : static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
657 : struct btrfs_device *dev)
658 : {
659 28911 : struct reada_extent *re = NULL;
660 : int mirror_num = 0;
661 28911 : struct extent_buffer *eb = NULL;
662 : u64 logical;
663 : u32 blocksize;
664 : int ret;
665 : int i;
666 : int need_kick = 0;
667 :
668 : spin_lock(&fs_info->reada_lock);
669 28913 : if (dev->reada_curr_zone == NULL) {
670 10785 : ret = reada_pick_zone(dev);
671 10785 : if (!ret) {
672 : spin_unlock(&fs_info->reada_lock);
673 10566 : return 0;
674 : }
675 : }
676 : /*
677 : * FIXME currently we issue the reads one extent at a time. If we have
678 : * a contiguous block of extents, we could also coagulate them or use
679 : * plugging to speed things up
680 : */
681 18347 : ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
682 18347 : dev->reada_next >> PAGE_CACHE_SHIFT, 1);
683 18347 : if (ret == 0 || re->logical >= dev->reada_curr_zone->end) {
684 16167 : ret = reada_pick_zone(dev);
685 16167 : if (!ret) {
686 : spin_unlock(&fs_info->reada_lock);
687 219 : return 0;
688 : }
689 15948 : re = NULL;
690 15948 : ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
691 15948 : dev->reada_next >> PAGE_CACHE_SHIFT, 1);
692 : }
693 18128 : if (ret == 0) {
694 : spin_unlock(&fs_info->reada_lock);
695 2010 : return 0;
696 : }
697 16118 : dev->reada_next = re->logical + re->blocksize;
698 16118 : re->refcnt++;
699 :
700 : spin_unlock(&fs_info->reada_lock);
701 :
702 : /*
703 : * find mirror num
704 : */
705 17815 : for (i = 0; i < re->nzones; ++i) {
706 17815 : if (re->zones[i]->device == dev) {
707 16118 : mirror_num = i + 1;
708 16118 : break;
709 : }
710 : }
711 16118 : logical = re->logical;
712 16118 : blocksize = re->blocksize;
713 :
714 : spin_lock(&re->lock);
715 16118 : if (re->scheduled_for == NULL) {
716 15891 : re->scheduled_for = dev;
717 : need_kick = 1;
718 : }
719 : spin_unlock(&re->lock);
720 :
721 16118 : reada_extent_put(fs_info, re);
722 :
723 16118 : if (!need_kick)
724 : return 0;
725 :
726 15891 : atomic_inc(&dev->reada_in_flight);
727 31637 : ret = reada_tree_block_flagged(fs_info->extent_root, logical, blocksize,
728 : mirror_num, &eb);
729 15890 : if (ret)
730 0 : __readahead_hook(fs_info->extent_root, NULL, logical, ret);
731 15890 : else if (eb)
732 31492 : __readahead_hook(fs_info->extent_root, eb, eb->start, ret);
733 :
734 15891 : if (eb)
735 15747 : free_extent_buffer(eb);
736 :
737 : return 1;
738 :
739 : }
740 :
741 372 : static void reada_start_machine_worker(struct btrfs_work *work)
742 : {
743 : struct reada_machine_work *rmw;
744 : struct btrfs_fs_info *fs_info;
745 : int old_ioprio;
746 :
747 : rmw = container_of(work, struct reada_machine_work, work);
748 372 : fs_info = rmw->fs_info;
749 :
750 372 : kfree(rmw);
751 :
752 744 : old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
753 : task_nice_ioprio(current));
754 372 : set_task_ioprio(current, BTRFS_IOPRIO_READA);
755 372 : __reada_start_machine(fs_info);
756 372 : set_task_ioprio(current, old_ioprio);
757 371 : }
758 :
759 370 : static void __reada_start_machine(struct btrfs_fs_info *fs_info)
760 : {
761 : struct btrfs_device *device;
762 370 : struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
763 : u64 enqueued;
764 : u64 total = 0;
765 : int i;
766 :
767 : do {
768 : enqueued = 0;
769 43929 : list_for_each_entry(device, &fs_devices->devices, dev_list) {
770 29170 : if (atomic_read(&device->reada_in_flight) <
771 : MAX_IN_FLIGHT)
772 28911 : enqueued += reada_start_machine_dev(fs_info,
773 : device);
774 : }
775 14759 : total += enqueued;
776 14759 : } while (enqueued && total < 10000);
777 :
778 372 : if (enqueued == 0)
779 372 : return;
780 :
781 : /*
782 : * If everything is already in the cache, this is effectively single
783 : * threaded. To a) not hold the caller for too long and b) to utilize
784 : * more cores, we broke the loop above after 10000 iterations and now
785 : * enqueue to workers to finish it. This will distribute the load to
786 : * the cores.
787 : */
788 2 : for (i = 0; i < 2; ++i)
789 2 : reada_start_machine(fs_info);
790 : }
791 :
792 372 : static void reada_start_machine(struct btrfs_fs_info *fs_info)
793 : {
794 : struct reada_machine_work *rmw;
795 :
796 372 : rmw = kzalloc(sizeof(*rmw), GFP_NOFS);
797 372 : if (!rmw) {
798 : /* FIXME we cannot handle this properly right now */
799 0 : BUG();
800 : }
801 372 : btrfs_init_work(&rmw->work, btrfs_readahead_helper,
802 : reada_start_machine_worker, NULL, NULL);
803 372 : rmw->fs_info = fs_info;
804 :
805 372 : btrfs_queue_work(fs_info->readahead_workers, &rmw->work);
806 372 : }
807 :
808 : #ifdef DEBUG
809 : static 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_CACHE_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 for %lld",
854 : re->logical, re->blocksize,
855 : list_empty(&re->extctl), re->scheduled_for ?
856 : re->scheduled_for->devid : -1);
857 :
858 : for (i = 0; i < re->nzones; ++i) {
859 : printk(KERN_CONT " zone %llu-%llu devs",
860 : re->zones[i]->start,
861 : re->zones[i]->end);
862 : for (j = 0; j < re->zones[i]->ndevs; ++j) {
863 : printk(KERN_CONT " %lld",
864 : re->zones[i]->devs[j]->devid);
865 : }
866 : }
867 : printk(KERN_CONT "\n");
868 : index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
869 : if (++cnt > 15)
870 : break;
871 : }
872 : }
873 :
874 : index = 0;
875 : cnt = 0;
876 : while (all) {
877 : struct reada_extent *re = NULL;
878 :
879 : ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
880 : index, 1);
881 : if (ret == 0)
882 : break;
883 : if (!re->scheduled_for) {
884 : index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
885 : continue;
886 : }
887 : printk(KERN_DEBUG
888 : "re: logical %llu size %u list empty %d for %lld",
889 : re->logical, re->blocksize, list_empty(&re->extctl),
890 : re->scheduled_for ? re->scheduled_for->devid : -1);
891 : for (i = 0; i < re->nzones; ++i) {
892 : printk(KERN_CONT " zone %llu-%llu devs",
893 : re->zones[i]->start,
894 : re->zones[i]->end);
895 : for (i = 0; i < re->nzones; ++i) {
896 : printk(KERN_CONT " zone %llu-%llu devs",
897 : re->zones[i]->start,
898 : re->zones[i]->end);
899 : for (j = 0; j < re->zones[i]->ndevs; ++j) {
900 : printk(KERN_CONT " %lld",
901 : re->zones[i]->devs[j]->devid);
902 : }
903 : }
904 : }
905 : printk(KERN_CONT "\n");
906 : index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
907 : }
908 : spin_unlock(&fs_info->reada_lock);
909 : }
910 : #endif
911 :
912 : /*
913 : * interface
914 : */
915 226 : struct reada_control *btrfs_reada_add(struct btrfs_root *root,
916 : struct btrfs_key *key_start, struct btrfs_key *key_end)
917 : {
918 : struct reada_control *rc;
919 : u64 start;
920 : u64 generation;
921 : int level;
922 226 : struct extent_buffer *node;
923 : static struct btrfs_key max_key = {
924 : .objectid = (u64)-1,
925 : .type = (u8)-1,
926 : .offset = (u64)-1
927 : };
928 :
929 226 : rc = kzalloc(sizeof(*rc), GFP_NOFS);
930 226 : if (!rc)
931 : return ERR_PTR(-ENOMEM);
932 :
933 226 : rc->root = root;
934 226 : rc->key_start = *key_start;
935 226 : rc->key_end = *key_end;
936 : atomic_set(&rc->elems, 0);
937 226 : init_waitqueue_head(&rc->wait);
938 : kref_init(&rc->refcnt);
939 226 : kref_get(&rc->refcnt); /* one ref for having elements */
940 :
941 226 : node = btrfs_root_node(root);
942 226 : start = node->start;
943 226 : level = btrfs_header_level(node);
944 : generation = btrfs_header_generation(node);
945 226 : free_extent_buffer(node);
946 :
947 226 : if (reada_add_block(rc, start, &max_key, level, generation)) {
948 0 : kfree(rc);
949 0 : return ERR_PTR(-ENOMEM);
950 : }
951 :
952 226 : reada_start_machine(root->fs_info);
953 :
954 226 : return rc;
955 : }
956 :
957 : #ifdef DEBUG
958 : int btrfs_reada_wait(void *handle)
959 : {
960 : struct reada_control *rc = handle;
961 :
962 : while (atomic_read(&rc->elems)) {
963 : wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
964 : 5 * HZ);
965 : dump_devs(rc->root->fs_info,
966 : atomic_read(&rc->elems) < 10 ? 1 : 0);
967 : }
968 :
969 : dump_devs(rc->root->fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0);
970 :
971 : kref_put(&rc->refcnt, reada_control_release);
972 :
973 : return 0;
974 : }
975 : #else
976 226 : int btrfs_reada_wait(void *handle)
977 : {
978 : struct reada_control *rc = handle;
979 :
980 572 : while (atomic_read(&rc->elems)) {
981 480 : wait_event(rc->wait, atomic_read(&rc->elems) == 0);
982 : }
983 :
984 226 : kref_put(&rc->refcnt, reada_control_release);
985 :
986 226 : return 0;
987 : }
988 : #endif
989 :
990 0 : void btrfs_reada_detach(void *handle)
991 : {
992 : struct reada_control *rc = handle;
993 :
994 0 : kref_put(&rc->refcnt, reada_control_release);
995 0 : }
|
