blob: a955669519a265bbfb5f13de4723c87932f8047e [file] [log] [blame]
Arne Jansen7414a032011-05-23 14:33:49 +02001/*
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
31#undef DEBUG
32
33/*
34 * This is the implementation for the generic read ahead framework.
35 *
36 * To trigger a readahead, btrfs_reada_add must be called. It will start
37 * a read ahead for the given range [start, end) on tree root. The returned
38 * handle can either be used to wait on the readahead to finish
39 * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach).
40 *
41 * The read ahead works as follows:
42 * On btrfs_reada_add, the root of the tree is inserted into a radix_tree.
43 * reada_start_machine will then search for extents to prefetch and trigger
44 * some reads. When a read finishes for a node, all contained node/leaf
45 * pointers that lie in the given range will also be enqueued. The reads will
46 * be triggered in sequential order, thus giving a big win over a naive
47 * enumeration. It will also make use of multi-device layouts. Each disk
48 * will have its on read pointer and all disks will by utilized in parallel.
49 * Also will no two disks read both sides of a mirror simultaneously, as this
50 * would waste seeking capacity. Instead both disks will read different parts
51 * of the filesystem.
52 * Any number of readaheads can be started in parallel. The read order will be
53 * determined globally, i.e. 2 parallel readaheads will normally finish faster
54 * than the 2 started one after another.
55 */
56
Arne Jansen7414a032011-05-23 14:33:49 +020057#define MAX_IN_FLIGHT 6
58
59struct reada_extctl {
60 struct list_head list;
61 struct reada_control *rc;
62 u64 generation;
63};
64
65struct reada_extent {
66 u64 logical;
67 struct btrfs_key top;
68 u32 blocksize;
69 int err;
70 struct list_head extctl;
Al Viro99621b42012-08-29 16:31:33 -040071 int refcnt;
Arne Jansen7414a032011-05-23 14:33:49 +020072 spinlock_t lock;
Stefan Behrens94598ba2012-03-27 14:21:26 -040073 struct reada_zone *zones[BTRFS_MAX_MIRRORS];
Arne Jansen7414a032011-05-23 14:33:49 +020074 int nzones;
75 struct btrfs_device *scheduled_for;
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;
Stefan Behrens94598ba2012-03-27 14:21:26 -040086 struct btrfs_device *devs[BTRFS_MAX_MIRRORS]; /* full list, incl
87 * self */
Arne Jansen7414a032011-05-23 14:33:49 +020088 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, int level, u64 generation);
105
106/* recurses */
107/* in case of err, eb might be NULL */
108static int __readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
109 u64 start, int err)
110{
111 int level = 0;
112 int nritems;
113 int i;
114 u64 bytenr;
115 u64 generation;
116 struct reada_extent *re;
117 struct btrfs_fs_info *fs_info = root->fs_info;
118 struct list_head list;
119 unsigned long index = start >> PAGE_CACHE_SHIFT;
120 struct btrfs_device *for_dev;
121
122 if (eb)
123 level = btrfs_header_level(eb);
124
125 /* find extent */
126 spin_lock(&fs_info->reada_lock);
127 re = radix_tree_lookup(&fs_info->reada_tree, index);
128 if (re)
Al Viro99621b42012-08-29 16:31:33 -0400129 re->refcnt++;
Arne Jansen7414a032011-05-23 14:33:49 +0200130 spin_unlock(&fs_info->reada_lock);
131
132 if (!re)
133 return -1;
134
135 spin_lock(&re->lock);
136 /*
137 * just take the full list from the extent. afterwards we
138 * don't need the lock anymore
139 */
140 list_replace_init(&re->extctl, &list);
141 for_dev = re->scheduled_for;
142 re->scheduled_for = NULL;
143 spin_unlock(&re->lock);
144
145 if (err == 0) {
146 nritems = level ? btrfs_header_nritems(eb) : 0;
147 generation = btrfs_header_generation(eb);
148 /*
149 * FIXME: currently we just set nritems to 0 if this is a leaf,
150 * effectively ignoring the content. In a next step we could
151 * trigger more readahead depending from the content, e.g.
152 * fetch the checksums for the extents in the leaf.
153 */
154 } else {
155 /*
156 * this is the error case, the extent buffer has not been
157 * read correctly. We won't access anything from it and
158 * just cleanup our data structures. Effectively this will
159 * cut the branch below this node from read ahead.
160 */
161 nritems = 0;
162 generation = 0;
163 }
164
165 for (i = 0; i < nritems; i++) {
166 struct reada_extctl *rec;
167 u64 n_gen;
168 struct btrfs_key key;
169 struct btrfs_key next_key;
170
171 btrfs_node_key_to_cpu(eb, &key, i);
172 if (i + 1 < nritems)
173 btrfs_node_key_to_cpu(eb, &next_key, i + 1);
174 else
175 next_key = re->top;
176 bytenr = btrfs_node_blockptr(eb, i);
177 n_gen = btrfs_node_ptr_generation(eb, i);
178
179 list_for_each_entry(rec, &list, list) {
180 struct reada_control *rc = rec->rc;
181
182 /*
183 * if the generation doesn't match, just ignore this
184 * extctl. This will probably cut off a branch from
185 * prefetch. Alternatively one could start a new (sub-)
186 * prefetch for this branch, starting again from root.
187 * FIXME: move the generation check out of this loop
188 */
189#ifdef DEBUG
190 if (rec->generation != generation) {
191 printk(KERN_DEBUG "generation mismatch for "
192 "(%llu,%d,%llu) %llu != %llu\n",
193 key.objectid, key.type, key.offset,
194 rec->generation, generation);
195 }
196#endif
197 if (rec->generation == generation &&
198 btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 &&
199 btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0)
200 reada_add_block(rc, bytenr, &next_key,
201 level - 1, n_gen);
202 }
203 }
204 /*
205 * free extctl records
206 */
207 while (!list_empty(&list)) {
208 struct reada_control *rc;
209 struct reada_extctl *rec;
210
211 rec = list_first_entry(&list, struct reada_extctl, list);
212 list_del(&rec->list);
213 rc = rec->rc;
214 kfree(rec);
215
216 kref_get(&rc->refcnt);
217 if (atomic_dec_and_test(&rc->elems)) {
218 kref_put(&rc->refcnt, reada_control_release);
219 wake_up(&rc->wait);
220 }
221 kref_put(&rc->refcnt, reada_control_release);
222
223 reada_extent_put(fs_info, re); /* one ref for each entry */
224 }
225 reada_extent_put(fs_info, re); /* our ref */
226 if (for_dev)
227 atomic_dec(&for_dev->reada_in_flight);
228
229 return 0;
230}
231
232/*
233 * start is passed separately in case eb in NULL, which may be the case with
234 * failed I/O
235 */
236int btree_readahead_hook(struct btrfs_root *root, struct extent_buffer *eb,
237 u64 start, int err)
238{
239 int ret;
240
241 ret = __readahead_hook(root, eb, start, err);
242
243 reada_start_machine(root->fs_info);
244
245 return ret;
246}
247
248static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
249 struct btrfs_device *dev, u64 logical,
Ilya Dryomov21ca5432011-11-04 09:41:02 -0400250 struct btrfs_bio *bbio)
Arne Jansen7414a032011-05-23 14:33:49 +0200251{
252 int ret;
Arne Jansen7414a032011-05-23 14:33:49 +0200253 struct reada_zone *zone;
254 struct btrfs_block_group_cache *cache = NULL;
255 u64 start;
256 u64 end;
257 int i;
258
Arne Jansen7414a032011-05-23 14:33:49 +0200259 zone = NULL;
260 spin_lock(&fs_info->reada_lock);
261 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
262 logical >> PAGE_CACHE_SHIFT, 1);
263 if (ret == 1)
264 kref_get(&zone->refcnt);
265 spin_unlock(&fs_info->reada_lock);
266
267 if (ret == 1) {
268 if (logical >= zone->start && logical < zone->end)
269 return zone;
270 spin_lock(&fs_info->reada_lock);
271 kref_put(&zone->refcnt, reada_zone_release);
272 spin_unlock(&fs_info->reada_lock);
273 }
274
Arne Jansen7414a032011-05-23 14:33:49 +0200275 cache = btrfs_lookup_block_group(fs_info, logical);
276 if (!cache)
277 return NULL;
278
279 start = cache->key.objectid;
280 end = start + cache->key.offset - 1;
281 btrfs_put_block_group(cache);
282
283 zone = kzalloc(sizeof(*zone), GFP_NOFS);
284 if (!zone)
285 return NULL;
286
287 zone->start = start;
288 zone->end = end;
289 INIT_LIST_HEAD(&zone->list);
290 spin_lock_init(&zone->lock);
291 zone->locked = 0;
292 kref_init(&zone->refcnt);
293 zone->elems = 0;
294 zone->device = dev; /* our device always sits at index 0 */
Ilya Dryomov21ca5432011-11-04 09:41:02 -0400295 for (i = 0; i < bbio->num_stripes; ++i) {
Arne Jansen7414a032011-05-23 14:33:49 +0200296 /* bounds have already been checked */
Ilya Dryomov21ca5432011-11-04 09:41:02 -0400297 zone->devs[i] = bbio->stripes[i].dev;
Arne Jansen7414a032011-05-23 14:33:49 +0200298 }
Ilya Dryomov21ca5432011-11-04 09:41:02 -0400299 zone->ndevs = bbio->num_stripes;
Arne Jansen7414a032011-05-23 14:33:49 +0200300
301 spin_lock(&fs_info->reada_lock);
302 ret = radix_tree_insert(&dev->reada_zones,
Chris Masona1754232012-02-28 12:42:44 -0500303 (unsigned long)(zone->end >> PAGE_CACHE_SHIFT),
Arne Jansen7414a032011-05-23 14:33:49 +0200304 zone);
Arne Jansen7414a032011-05-23 14:33:49 +0200305
Arne Jansen8c9c2bf2012-02-25 09:09:30 +0100306 if (ret == -EEXIST) {
Arne Jansen7414a032011-05-23 14:33:49 +0200307 kfree(zone);
Arne Jansen8c9c2bf2012-02-25 09:09:30 +0100308 ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
309 logical >> PAGE_CACHE_SHIFT, 1);
310 if (ret == 1)
311 kref_get(&zone->refcnt);
Arne Jansen7414a032011-05-23 14:33:49 +0200312 }
Arne Jansen8c9c2bf2012-02-25 09:09:30 +0100313 spin_unlock(&fs_info->reada_lock);
Arne Jansen7414a032011-05-23 14:33:49 +0200314
315 return zone;
316}
317
318static struct reada_extent *reada_find_extent(struct btrfs_root *root,
319 u64 logical,
320 struct btrfs_key *top, int level)
321{
322 int ret;
Arne Jansen7414a032011-05-23 14:33:49 +0200323 struct reada_extent *re = NULL;
Arne Jansen8c9c2bf2012-02-25 09:09:30 +0100324 struct reada_extent *re_exist = NULL;
Arne Jansen7414a032011-05-23 14:33:49 +0200325 struct btrfs_fs_info *fs_info = root->fs_info;
326 struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
Ilya Dryomov21ca5432011-11-04 09:41:02 -0400327 struct btrfs_bio *bbio = NULL;
Arne Jansen7414a032011-05-23 14:33:49 +0200328 struct btrfs_device *dev;
Arne Jansen207a2322012-02-25 09:09:47 +0100329 struct btrfs_device *prev_dev;
Arne Jansen7414a032011-05-23 14:33:49 +0200330 u32 blocksize;
331 u64 length;
332 int nzones = 0;
333 int i;
334 unsigned long index = logical >> PAGE_CACHE_SHIFT;
335
Arne Jansen7414a032011-05-23 14:33:49 +0200336 spin_lock(&fs_info->reada_lock);
337 re = radix_tree_lookup(&fs_info->reada_tree, index);
338 if (re)
Al Viro99621b42012-08-29 16:31:33 -0400339 re->refcnt++;
Arne Jansen7414a032011-05-23 14:33:49 +0200340 spin_unlock(&fs_info->reada_lock);
341
Arne Jansen8c9c2bf2012-02-25 09:09:30 +0100342 if (re)
Arne Jansen7414a032011-05-23 14:33:49 +0200343 return re;
344
345 re = kzalloc(sizeof(*re), GFP_NOFS);
346 if (!re)
347 return NULL;
348
349 blocksize = btrfs_level_size(root, level);
350 re->logical = logical;
351 re->blocksize = blocksize;
352 re->top = *top;
353 INIT_LIST_HEAD(&re->extctl);
354 spin_lock_init(&re->lock);
Al Viro99621b42012-08-29 16:31:33 -0400355 re->refcnt = 1;
Arne Jansen7414a032011-05-23 14:33:49 +0200356
357 /*
358 * map block
359 */
360 length = blocksize;
Ilya Dryomov21ca5432011-11-04 09:41:02 -0400361 ret = btrfs_map_block(map_tree, REQ_WRITE, logical, &length, &bbio, 0);
362 if (ret || !bbio || length < blocksize)
Arne Jansen7414a032011-05-23 14:33:49 +0200363 goto error;
364
Stefan Behrens94598ba2012-03-27 14:21:26 -0400365 if (bbio->num_stripes > BTRFS_MAX_MIRRORS) {
Arne Jansen7414a032011-05-23 14:33:49 +0200366 printk(KERN_ERR "btrfs readahead: more than %d copies not "
Stefan Behrens94598ba2012-03-27 14:21:26 -0400367 "supported", BTRFS_MAX_MIRRORS);
Arne Jansen7414a032011-05-23 14:33:49 +0200368 goto error;
369 }
370
Ilya Dryomov21ca5432011-11-04 09:41:02 -0400371 for (nzones = 0; nzones < bbio->num_stripes; ++nzones) {
Arne Jansen7414a032011-05-23 14:33:49 +0200372 struct reada_zone *zone;
373
Ilya Dryomov21ca5432011-11-04 09:41:02 -0400374 dev = bbio->stripes[nzones].dev;
375 zone = reada_find_zone(fs_info, dev, logical, bbio);
Arne Jansen7414a032011-05-23 14:33:49 +0200376 if (!zone)
377 break;
378
379 re->zones[nzones] = zone;
380 spin_lock(&zone->lock);
381 if (!zone->elems)
382 kref_get(&zone->refcnt);
383 ++zone->elems;
384 spin_unlock(&zone->lock);
385 spin_lock(&fs_info->reada_lock);
386 kref_put(&zone->refcnt, reada_zone_release);
387 spin_unlock(&fs_info->reada_lock);
388 }
389 re->nzones = nzones;
390 if (nzones == 0) {
391 /* not a single zone found, error and out */
392 goto error;
393 }
394
395 /* insert extent in reada_tree + all per-device trees, all or nothing */
396 spin_lock(&fs_info->reada_lock);
397 ret = radix_tree_insert(&fs_info->reada_tree, index, re);
Arne Jansen8c9c2bf2012-02-25 09:09:30 +0100398 if (ret == -EEXIST) {
399 re_exist = radix_tree_lookup(&fs_info->reada_tree, index);
400 BUG_ON(!re_exist);
Al Viro99621b42012-08-29 16:31:33 -0400401 re_exist->refcnt++;
Arne Jansen8c9c2bf2012-02-25 09:09:30 +0100402 spin_unlock(&fs_info->reada_lock);
403 goto error;
404 }
Arne Jansen7414a032011-05-23 14:33:49 +0200405 if (ret) {
406 spin_unlock(&fs_info->reada_lock);
Arne Jansen7414a032011-05-23 14:33:49 +0200407 goto error;
408 }
Arne Jansen207a2322012-02-25 09:09:47 +0100409 prev_dev = NULL;
Arne Jansen7414a032011-05-23 14:33:49 +0200410 for (i = 0; i < nzones; ++i) {
Ilya Dryomov21ca5432011-11-04 09:41:02 -0400411 dev = bbio->stripes[i].dev;
Arne Jansen207a2322012-02-25 09:09:47 +0100412 if (dev == prev_dev) {
413 /*
414 * in case of DUP, just add the first zone. As both
415 * are on the same device, there's nothing to gain
416 * from adding both.
417 * Also, it wouldn't work, as the tree is per device
418 * and adding would fail with EEXIST
419 */
420 continue;
421 }
422 prev_dev = dev;
Arne Jansen7414a032011-05-23 14:33:49 +0200423 ret = radix_tree_insert(&dev->reada_extents, index, re);
424 if (ret) {
425 while (--i >= 0) {
Ilya Dryomov21ca5432011-11-04 09:41:02 -0400426 dev = bbio->stripes[i].dev;
Arne Jansen7414a032011-05-23 14:33:49 +0200427 BUG_ON(dev == NULL);
428 radix_tree_delete(&dev->reada_extents, index);
429 }
430 BUG_ON(fs_info == NULL);
431 radix_tree_delete(&fs_info->reada_tree, index);
432 spin_unlock(&fs_info->reada_lock);
433 goto error;
434 }
435 }
436 spin_unlock(&fs_info->reada_lock);
437
Ilya Dryomov21ca5432011-11-04 09:41:02 -0400438 kfree(bbio);
Arne Jansen7414a032011-05-23 14:33:49 +0200439 return re;
440
441error:
442 while (nzones) {
443 struct reada_zone *zone;
444
445 --nzones;
446 zone = re->zones[nzones];
447 kref_get(&zone->refcnt);
448 spin_lock(&zone->lock);
449 --zone->elems;
450 if (zone->elems == 0) {
451 /*
452 * no fs_info->reada_lock needed, as this can't be
453 * the last ref
454 */
455 kref_put(&zone->refcnt, reada_zone_release);
456 }
457 spin_unlock(&zone->lock);
458
459 spin_lock(&fs_info->reada_lock);
460 kref_put(&zone->refcnt, reada_zone_release);
461 spin_unlock(&fs_info->reada_lock);
462 }
Ilya Dryomov21ca5432011-11-04 09:41:02 -0400463 kfree(bbio);
Arne Jansen7414a032011-05-23 14:33:49 +0200464 kfree(re);
Arne Jansen8c9c2bf2012-02-25 09:09:30 +0100465 return re_exist;
Arne Jansen7414a032011-05-23 14:33:49 +0200466}
467
Arne Jansen7414a032011-05-23 14:33:49 +0200468static void reada_extent_put(struct btrfs_fs_info *fs_info,
469 struct reada_extent *re)
470{
471 int i;
472 unsigned long index = re->logical >> PAGE_CACHE_SHIFT;
473
474 spin_lock(&fs_info->reada_lock);
Al Viro99621b42012-08-29 16:31:33 -0400475 if (--re->refcnt) {
Arne Jansen7414a032011-05-23 14:33:49 +0200476 spin_unlock(&fs_info->reada_lock);
477 return;
478 }
479
480 radix_tree_delete(&fs_info->reada_tree, index);
481 for (i = 0; i < re->nzones; ++i) {
482 struct reada_zone *zone = re->zones[i];
483
484 radix_tree_delete(&zone->device->reada_extents, index);
485 }
486
487 spin_unlock(&fs_info->reada_lock);
488
489 for (i = 0; i < re->nzones; ++i) {
490 struct reada_zone *zone = re->zones[i];
491
492 kref_get(&zone->refcnt);
493 spin_lock(&zone->lock);
494 --zone->elems;
495 if (zone->elems == 0) {
496 /* no fs_info->reada_lock needed, as this can't be
497 * the last ref */
498 kref_put(&zone->refcnt, reada_zone_release);
499 }
500 spin_unlock(&zone->lock);
501
502 spin_lock(&fs_info->reada_lock);
503 kref_put(&zone->refcnt, reada_zone_release);
504 spin_unlock(&fs_info->reada_lock);
505 }
506 if (re->scheduled_for)
507 atomic_dec(&re->scheduled_for->reada_in_flight);
508
509 kfree(re);
510}
511
512static void reada_zone_release(struct kref *kref)
513{
514 struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
515
516 radix_tree_delete(&zone->device->reada_zones,
517 zone->end >> PAGE_CACHE_SHIFT);
518
519 kfree(zone);
520}
521
522static void reada_control_release(struct kref *kref)
523{
524 struct reada_control *rc = container_of(kref, struct reada_control,
525 refcnt);
526
527 kfree(rc);
528}
529
530static int reada_add_block(struct reada_control *rc, u64 logical,
531 struct btrfs_key *top, int level, u64 generation)
532{
533 struct btrfs_root *root = rc->root;
534 struct reada_extent *re;
535 struct reada_extctl *rec;
536
537 re = reada_find_extent(root, logical, top, level); /* takes one ref */
538 if (!re)
539 return -1;
540
541 rec = kzalloc(sizeof(*rec), GFP_NOFS);
542 if (!rec) {
543 reada_extent_put(root->fs_info, re);
544 return -1;
545 }
546
547 rec->rc = rc;
548 rec->generation = generation;
549 atomic_inc(&rc->elems);
550
551 spin_lock(&re->lock);
552 list_add_tail(&rec->list, &re->extctl);
553 spin_unlock(&re->lock);
554
555 /* leave the ref on the extent */
556
557 return 0;
558}
559
560/*
561 * called with fs_info->reada_lock held
562 */
563static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
564{
565 int i;
566 unsigned long index = zone->end >> PAGE_CACHE_SHIFT;
567
568 for (i = 0; i < zone->ndevs; ++i) {
569 struct reada_zone *peer;
570 peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
571 if (peer && peer->device != zone->device)
572 peer->locked = lock;
573 }
574}
575
576/*
577 * called with fs_info->reada_lock held
578 */
579static int reada_pick_zone(struct btrfs_device *dev)
580{
581 struct reada_zone *top_zone = NULL;
582 struct reada_zone *top_locked_zone = NULL;
583 u64 top_elems = 0;
584 u64 top_locked_elems = 0;
585 unsigned long index = 0;
586 int ret;
587
588 if (dev->reada_curr_zone) {
589 reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
590 kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
591 dev->reada_curr_zone = NULL;
592 }
593 /* pick the zone with the most elements */
594 while (1) {
595 struct reada_zone *zone;
596
597 ret = radix_tree_gang_lookup(&dev->reada_zones,
598 (void **)&zone, index, 1);
599 if (ret == 0)
600 break;
601 index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
602 if (zone->locked) {
603 if (zone->elems > top_locked_elems) {
604 top_locked_elems = zone->elems;
605 top_locked_zone = zone;
606 }
607 } else {
608 if (zone->elems > top_elems) {
609 top_elems = zone->elems;
610 top_zone = zone;
611 }
612 }
613 }
614 if (top_zone)
615 dev->reada_curr_zone = top_zone;
616 else if (top_locked_zone)
617 dev->reada_curr_zone = top_locked_zone;
618 else
619 return 0;
620
621 dev->reada_next = dev->reada_curr_zone->start;
622 kref_get(&dev->reada_curr_zone->refcnt);
623 reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
624
625 return 1;
626}
627
628static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
629 struct btrfs_device *dev)
630{
631 struct reada_extent *re = NULL;
632 int mirror_num = 0;
633 struct extent_buffer *eb = NULL;
634 u64 logical;
635 u32 blocksize;
636 int ret;
637 int i;
638 int need_kick = 0;
639
640 spin_lock(&fs_info->reada_lock);
641 if (dev->reada_curr_zone == NULL) {
642 ret = reada_pick_zone(dev);
643 if (!ret) {
644 spin_unlock(&fs_info->reada_lock);
645 return 0;
646 }
647 }
648 /*
649 * FIXME currently we issue the reads one extent at a time. If we have
650 * a contiguous block of extents, we could also coagulate them or use
651 * plugging to speed things up
652 */
653 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
654 dev->reada_next >> PAGE_CACHE_SHIFT, 1);
655 if (ret == 0 || re->logical >= dev->reada_curr_zone->end) {
656 ret = reada_pick_zone(dev);
657 if (!ret) {
658 spin_unlock(&fs_info->reada_lock);
659 return 0;
660 }
661 re = NULL;
662 ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
663 dev->reada_next >> PAGE_CACHE_SHIFT, 1);
664 }
665 if (ret == 0) {
666 spin_unlock(&fs_info->reada_lock);
667 return 0;
668 }
669 dev->reada_next = re->logical + re->blocksize;
Al Viro99621b42012-08-29 16:31:33 -0400670 re->refcnt++;
Arne Jansen7414a032011-05-23 14:33:49 +0200671
672 spin_unlock(&fs_info->reada_lock);
673
674 /*
675 * find mirror num
676 */
677 for (i = 0; i < re->nzones; ++i) {
678 if (re->zones[i]->device == dev) {
679 mirror_num = i + 1;
680 break;
681 }
682 }
683 logical = re->logical;
684 blocksize = re->blocksize;
685
686 spin_lock(&re->lock);
687 if (re->scheduled_for == NULL) {
688 re->scheduled_for = dev;
689 need_kick = 1;
690 }
691 spin_unlock(&re->lock);
692
693 reada_extent_put(fs_info, re);
694
695 if (!need_kick)
696 return 0;
697
698 atomic_inc(&dev->reada_in_flight);
699 ret = reada_tree_block_flagged(fs_info->extent_root, logical, blocksize,
700 mirror_num, &eb);
701 if (ret)
702 __readahead_hook(fs_info->extent_root, NULL, logical, ret);
703 else if (eb)
704 __readahead_hook(fs_info->extent_root, eb, eb->start, ret);
705
706 if (eb)
707 free_extent_buffer(eb);
708
709 return 1;
710
711}
712
713static void reada_start_machine_worker(struct btrfs_work *work)
714{
715 struct reada_machine_work *rmw;
716 struct btrfs_fs_info *fs_info;
Stefan Behrens3d136a12012-02-03 11:20:04 +0100717 int old_ioprio;
Arne Jansen7414a032011-05-23 14:33:49 +0200718
719 rmw = container_of(work, struct reada_machine_work, work);
720 fs_info = rmw->fs_info;
721
722 kfree(rmw);
723
Stefan Behrens3d136a12012-02-03 11:20:04 +0100724 old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current),
725 task_nice_ioprio(current));
726 set_task_ioprio(current, BTRFS_IOPRIO_READA);
Arne Jansen7414a032011-05-23 14:33:49 +0200727 __reada_start_machine(fs_info);
Stefan Behrens3d136a12012-02-03 11:20:04 +0100728 set_task_ioprio(current, old_ioprio);
Arne Jansen7414a032011-05-23 14:33:49 +0200729}
730
731static void __reada_start_machine(struct btrfs_fs_info *fs_info)
732{
733 struct btrfs_device *device;
734 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
735 u64 enqueued;
736 u64 total = 0;
737 int i;
738
739 do {
740 enqueued = 0;
741 list_for_each_entry(device, &fs_devices->devices, dev_list) {
742 if (atomic_read(&device->reada_in_flight) <
743 MAX_IN_FLIGHT)
744 enqueued += reada_start_machine_dev(fs_info,
745 device);
746 }
747 total += enqueued;
748 } while (enqueued && total < 10000);
749
750 if (enqueued == 0)
751 return;
752
753 /*
754 * If everything is already in the cache, this is effectively single
755 * threaded. To a) not hold the caller for too long and b) to utilize
756 * more cores, we broke the loop above after 10000 iterations and now
757 * enqueue to workers to finish it. This will distribute the load to
758 * the cores.
759 */
760 for (i = 0; i < 2; ++i)
761 reada_start_machine(fs_info);
762}
763
764static void reada_start_machine(struct btrfs_fs_info *fs_info)
765{
766 struct reada_machine_work *rmw;
767
768 rmw = kzalloc(sizeof(*rmw), GFP_NOFS);
769 if (!rmw) {
770 /* FIXME we cannot handle this properly right now */
771 BUG();
772 }
773 rmw->work.func = reada_start_machine_worker;
774 rmw->fs_info = fs_info;
775
776 btrfs_queue_worker(&fs_info->readahead_workers, &rmw->work);
777}
778
779#ifdef DEBUG
780static void dump_devs(struct btrfs_fs_info *fs_info, int all)
781{
782 struct btrfs_device *device;
783 struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
784 unsigned long index;
785 int ret;
786 int i;
787 int j;
788 int cnt;
789
790 spin_lock(&fs_info->reada_lock);
791 list_for_each_entry(device, &fs_devices->devices, dev_list) {
792 printk(KERN_DEBUG "dev %lld has %d in flight\n", device->devid,
793 atomic_read(&device->reada_in_flight));
794 index = 0;
795 while (1) {
796 struct reada_zone *zone;
797 ret = radix_tree_gang_lookup(&device->reada_zones,
798 (void **)&zone, index, 1);
799 if (ret == 0)
800 break;
801 printk(KERN_DEBUG " zone %llu-%llu elems %llu locked "
802 "%d devs", zone->start, zone->end, zone->elems,
803 zone->locked);
804 for (j = 0; j < zone->ndevs; ++j) {
805 printk(KERN_CONT " %lld",
806 zone->devs[j]->devid);
807 }
808 if (device->reada_curr_zone == zone)
809 printk(KERN_CONT " curr off %llu",
810 device->reada_next - zone->start);
811 printk(KERN_CONT "\n");
812 index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
813 }
814 cnt = 0;
815 index = 0;
816 while (all) {
817 struct reada_extent *re = NULL;
818
819 ret = radix_tree_gang_lookup(&device->reada_extents,
820 (void **)&re, index, 1);
821 if (ret == 0)
822 break;
823 printk(KERN_DEBUG
824 " re: logical %llu size %u empty %d for %lld",
825 re->logical, re->blocksize,
826 list_empty(&re->extctl), re->scheduled_for ?
827 re->scheduled_for->devid : -1);
828
829 for (i = 0; i < re->nzones; ++i) {
830 printk(KERN_CONT " zone %llu-%llu devs",
831 re->zones[i]->start,
832 re->zones[i]->end);
833 for (j = 0; j < re->zones[i]->ndevs; ++j) {
834 printk(KERN_CONT " %lld",
835 re->zones[i]->devs[j]->devid);
836 }
837 }
838 printk(KERN_CONT "\n");
839 index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
840 if (++cnt > 15)
841 break;
842 }
843 }
844
845 index = 0;
846 cnt = 0;
847 while (all) {
848 struct reada_extent *re = NULL;
849
850 ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
851 index, 1);
852 if (ret == 0)
853 break;
854 if (!re->scheduled_for) {
855 index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
856 continue;
857 }
858 printk(KERN_DEBUG
859 "re: logical %llu size %u list empty %d for %lld",
860 re->logical, re->blocksize, list_empty(&re->extctl),
861 re->scheduled_for ? re->scheduled_for->devid : -1);
862 for (i = 0; i < re->nzones; ++i) {
863 printk(KERN_CONT " zone %llu-%llu devs",
864 re->zones[i]->start,
865 re->zones[i]->end);
866 for (i = 0; i < re->nzones; ++i) {
867 printk(KERN_CONT " zone %llu-%llu devs",
868 re->zones[i]->start,
869 re->zones[i]->end);
870 for (j = 0; j < re->zones[i]->ndevs; ++j) {
871 printk(KERN_CONT " %lld",
872 re->zones[i]->devs[j]->devid);
873 }
874 }
875 }
876 printk(KERN_CONT "\n");
877 index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
878 }
879 spin_unlock(&fs_info->reada_lock);
880}
881#endif
882
883/*
884 * interface
885 */
886struct reada_control *btrfs_reada_add(struct btrfs_root *root,
887 struct btrfs_key *key_start, struct btrfs_key *key_end)
888{
889 struct reada_control *rc;
890 u64 start;
891 u64 generation;
892 int level;
893 struct extent_buffer *node;
894 static struct btrfs_key max_key = {
895 .objectid = (u64)-1,
896 .type = (u8)-1,
897 .offset = (u64)-1
898 };
899
900 rc = kzalloc(sizeof(*rc), GFP_NOFS);
901 if (!rc)
902 return ERR_PTR(-ENOMEM);
903
904 rc->root = root;
905 rc->key_start = *key_start;
906 rc->key_end = *key_end;
907 atomic_set(&rc->elems, 0);
908 init_waitqueue_head(&rc->wait);
909 kref_init(&rc->refcnt);
910 kref_get(&rc->refcnt); /* one ref for having elements */
911
912 node = btrfs_root_node(root);
913 start = node->start;
914 level = btrfs_header_level(node);
915 generation = btrfs_header_generation(node);
916 free_extent_buffer(node);
917
918 reada_add_block(rc, start, &max_key, level, generation);
919
920 reada_start_machine(root->fs_info);
921
922 return rc;
923}
924
925#ifdef DEBUG
926int btrfs_reada_wait(void *handle)
927{
928 struct reada_control *rc = handle;
929
930 while (atomic_read(&rc->elems)) {
931 wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
932 5 * HZ);
933 dump_devs(rc->root->fs_info, rc->elems < 10 ? 1 : 0);
934 }
935
936 dump_devs(rc->root->fs_info, rc->elems < 10 ? 1 : 0);
937
938 kref_put(&rc->refcnt, reada_control_release);
939
940 return 0;
941}
942#else
943int btrfs_reada_wait(void *handle)
944{
945 struct reada_control *rc = handle;
946
947 while (atomic_read(&rc->elems)) {
948 wait_event(rc->wait, atomic_read(&rc->elems) == 0);
949 }
950
951 kref_put(&rc->refcnt, reada_control_release);
952
953 return 0;
954}
955#endif
956
957void btrfs_reada_detach(void *handle)
958{
959 struct reada_control *rc = handle;
960
961 kref_put(&rc->refcnt, reada_control_release);
962}