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 
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 
57 #define MAX_MIRRORS 2
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 	struct kref		refcnt;
73 	spinlock_t		lock;
74 	struct reada_zone	*zones[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[MAX_MIRRORS]; /* full list, incl self */
88 	int			ndevs;
89 	struct kref		refcnt;
90 };
91 
92 struct reada_machine_work {
93 	struct btrfs_work	work;
94 	struct btrfs_fs_info	*fs_info;
95 };
96 
97 static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *);
98 static void reada_control_release(struct kref *kref);
99 static void reada_zone_release(struct kref *kref);
100 static void reada_start_machine(struct btrfs_fs_info *fs_info);
101 static void __reada_start_machine(struct btrfs_fs_info *fs_info);
102 
103 static 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 */
__readahead_hook(struct btrfs_root * root,struct extent_buffer * eb,u64 start,int err)108 static 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)
129 		kref_get(&re->refcnt);
130 	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  */
btree_readahead_hook(struct btrfs_root * root,struct extent_buffer * eb,u64 start,int err)236 int 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 
reada_find_zone(struct btrfs_fs_info * fs_info,struct btrfs_device * dev,u64 logical,struct btrfs_bio * bbio)248 static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info,
249 					  struct btrfs_device *dev, u64 logical,
250 					  struct btrfs_bio *bbio)
251 {
252 	int ret;
253 	int looped = 0;
254 	struct reada_zone *zone;
255 	struct btrfs_block_group_cache *cache = NULL;
256 	u64 start;
257 	u64 end;
258 	int i;
259 
260 again:
261 	zone = NULL;
262 	spin_lock(&fs_info->reada_lock);
263 	ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone,
264 				     logical >> PAGE_CACHE_SHIFT, 1);
265 	if (ret == 1)
266 		kref_get(&zone->refcnt);
267 	spin_unlock(&fs_info->reada_lock);
268 
269 	if (ret == 1) {
270 		if (logical >= zone->start && logical < zone->end)
271 			return zone;
272 		spin_lock(&fs_info->reada_lock);
273 		kref_put(&zone->refcnt, reada_zone_release);
274 		spin_unlock(&fs_info->reada_lock);
275 	}
276 
277 	if (looped)
278 		return NULL;
279 
280 	cache = btrfs_lookup_block_group(fs_info, logical);
281 	if (!cache)
282 		return NULL;
283 
284 	start = cache->key.objectid;
285 	end = start + cache->key.offset - 1;
286 	btrfs_put_block_group(cache);
287 
288 	zone = kzalloc(sizeof(*zone), GFP_NOFS);
289 	if (!zone)
290 		return NULL;
291 
292 	zone->start = start;
293 	zone->end = end;
294 	INIT_LIST_HEAD(&zone->list);
295 	spin_lock_init(&zone->lock);
296 	zone->locked = 0;
297 	kref_init(&zone->refcnt);
298 	zone->elems = 0;
299 	zone->device = dev; /* our device always sits at index 0 */
300 	for (i = 0; i < bbio->num_stripes; ++i) {
301 		/* bounds have already been checked */
302 		zone->devs[i] = bbio->stripes[i].dev;
303 	}
304 	zone->ndevs = bbio->num_stripes;
305 
306 	spin_lock(&fs_info->reada_lock);
307 	ret = radix_tree_insert(&dev->reada_zones,
308 				(unsigned long)(zone->end >> PAGE_CACHE_SHIFT),
309 				zone);
310 	spin_unlock(&fs_info->reada_lock);
311 
312 	if (ret) {
313 		kfree(zone);
314 		looped = 1;
315 		goto again;
316 	}
317 
318 	return zone;
319 }
320 
reada_find_extent(struct btrfs_root * root,u64 logical,struct btrfs_key * top,int level)321 static struct reada_extent *reada_find_extent(struct btrfs_root *root,
322 					      u64 logical,
323 					      struct btrfs_key *top, int level)
324 {
325 	int ret;
326 	int looped = 0;
327 	struct reada_extent *re = NULL;
328 	struct btrfs_fs_info *fs_info = root->fs_info;
329 	struct btrfs_mapping_tree *map_tree = &fs_info->mapping_tree;
330 	struct btrfs_bio *bbio = NULL;
331 	struct btrfs_device *dev;
332 	u32 blocksize;
333 	u64 length;
334 	int nzones = 0;
335 	int i;
336 	unsigned long index = logical >> PAGE_CACHE_SHIFT;
337 
338 again:
339 	spin_lock(&fs_info->reada_lock);
340 	re = radix_tree_lookup(&fs_info->reada_tree, index);
341 	if (re)
342 		kref_get(&re->refcnt);
343 	spin_unlock(&fs_info->reada_lock);
344 
345 	if (re || looped)
346 		return re;
347 
348 	re = kzalloc(sizeof(*re), GFP_NOFS);
349 	if (!re)
350 		return NULL;
351 
352 	blocksize = btrfs_level_size(root, level);
353 	re->logical = logical;
354 	re->blocksize = blocksize;
355 	re->top = *top;
356 	INIT_LIST_HEAD(&re->extctl);
357 	spin_lock_init(&re->lock);
358 	kref_init(&re->refcnt);
359 
360 	/*
361 	 * map block
362 	 */
363 	length = blocksize;
364 	ret = btrfs_map_block(map_tree, REQ_WRITE, logical, &length, &bbio, 0);
365 	if (ret || !bbio || length < blocksize)
366 		goto error;
367 
368 	if (bbio->num_stripes > MAX_MIRRORS) {
369 		printk(KERN_ERR "btrfs readahead: more than %d copies not "
370 				"supported", MAX_MIRRORS);
371 		goto error;
372 	}
373 
374 	for (nzones = 0; nzones < bbio->num_stripes; ++nzones) {
375 		struct reada_zone *zone;
376 
377 		dev = bbio->stripes[nzones].dev;
378 		zone = reada_find_zone(fs_info, dev, logical, bbio);
379 		if (!zone)
380 			break;
381 
382 		re->zones[nzones] = zone;
383 		spin_lock(&zone->lock);
384 		if (!zone->elems)
385 			kref_get(&zone->refcnt);
386 		++zone->elems;
387 		spin_unlock(&zone->lock);
388 		spin_lock(&fs_info->reada_lock);
389 		kref_put(&zone->refcnt, reada_zone_release);
390 		spin_unlock(&fs_info->reada_lock);
391 	}
392 	re->nzones = nzones;
393 	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 	spin_lock(&fs_info->reada_lock);
400 	ret = radix_tree_insert(&fs_info->reada_tree, index, re);
401 	if (ret) {
402 		spin_unlock(&fs_info->reada_lock);
403 		if (ret != -ENOMEM) {
404 			/* someone inserted the extent in the meantime */
405 			looped = 1;
406 		}
407 		goto error;
408 	}
409 	for (i = 0; i < nzones; ++i) {
410 		dev = bbio->stripes[i].dev;
411 		ret = radix_tree_insert(&dev->reada_extents, index, re);
412 		if (ret) {
413 			while (--i >= 0) {
414 				dev = bbio->stripes[i].dev;
415 				BUG_ON(dev == NULL);
416 				radix_tree_delete(&dev->reada_extents, index);
417 			}
418 			BUG_ON(fs_info == NULL);
419 			radix_tree_delete(&fs_info->reada_tree, index);
420 			spin_unlock(&fs_info->reada_lock);
421 			goto error;
422 		}
423 	}
424 	spin_unlock(&fs_info->reada_lock);
425 
426 	kfree(bbio);
427 	return re;
428 
429 error:
430 	while (nzones) {
431 		struct reada_zone *zone;
432 
433 		--nzones;
434 		zone = re->zones[nzones];
435 		kref_get(&zone->refcnt);
436 		spin_lock(&zone->lock);
437 		--zone->elems;
438 		if (zone->elems == 0) {
439 			/*
440 			 * no fs_info->reada_lock needed, as this can't be
441 			 * the last ref
442 			 */
443 			kref_put(&zone->refcnt, reada_zone_release);
444 		}
445 		spin_unlock(&zone->lock);
446 
447 		spin_lock(&fs_info->reada_lock);
448 		kref_put(&zone->refcnt, reada_zone_release);
449 		spin_unlock(&fs_info->reada_lock);
450 	}
451 	kfree(bbio);
452 	kfree(re);
453 	if (looped)
454 		goto again;
455 	return NULL;
456 }
457 
reada_kref_dummy(struct kref * kr)458 static void reada_kref_dummy(struct kref *kr)
459 {
460 }
461 
reada_extent_put(struct btrfs_fs_info * fs_info,struct reada_extent * re)462 static void reada_extent_put(struct btrfs_fs_info *fs_info,
463 			     struct reada_extent *re)
464 {
465 	int i;
466 	unsigned long index = re->logical >> PAGE_CACHE_SHIFT;
467 
468 	spin_lock(&fs_info->reada_lock);
469 	if (!kref_put(&re->refcnt, reada_kref_dummy)) {
470 		spin_unlock(&fs_info->reada_lock);
471 		return;
472 	}
473 
474 	radix_tree_delete(&fs_info->reada_tree, index);
475 	for (i = 0; i < re->nzones; ++i) {
476 		struct reada_zone *zone = re->zones[i];
477 
478 		radix_tree_delete(&zone->device->reada_extents, index);
479 	}
480 
481 	spin_unlock(&fs_info->reada_lock);
482 
483 	for (i = 0; i < re->nzones; ++i) {
484 		struct reada_zone *zone = re->zones[i];
485 
486 		kref_get(&zone->refcnt);
487 		spin_lock(&zone->lock);
488 		--zone->elems;
489 		if (zone->elems == 0) {
490 			/* no fs_info->reada_lock needed, as this can't be
491 			 * the last ref */
492 			kref_put(&zone->refcnt, reada_zone_release);
493 		}
494 		spin_unlock(&zone->lock);
495 
496 		spin_lock(&fs_info->reada_lock);
497 		kref_put(&zone->refcnt, reada_zone_release);
498 		spin_unlock(&fs_info->reada_lock);
499 	}
500 	if (re->scheduled_for)
501 		atomic_dec(&re->scheduled_for->reada_in_flight);
502 
503 	kfree(re);
504 }
505 
reada_zone_release(struct kref * kref)506 static void reada_zone_release(struct kref *kref)
507 {
508 	struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt);
509 
510 	radix_tree_delete(&zone->device->reada_zones,
511 			  zone->end >> PAGE_CACHE_SHIFT);
512 
513 	kfree(zone);
514 }
515 
reada_control_release(struct kref * kref)516 static void reada_control_release(struct kref *kref)
517 {
518 	struct reada_control *rc = container_of(kref, struct reada_control,
519 						refcnt);
520 
521 	kfree(rc);
522 }
523 
reada_add_block(struct reada_control * rc,u64 logical,struct btrfs_key * top,int level,u64 generation)524 static int reada_add_block(struct reada_control *rc, u64 logical,
525 			   struct btrfs_key *top, int level, u64 generation)
526 {
527 	struct btrfs_root *root = rc->root;
528 	struct reada_extent *re;
529 	struct reada_extctl *rec;
530 
531 	re = reada_find_extent(root, logical, top, level); /* takes one ref */
532 	if (!re)
533 		return -1;
534 
535 	rec = kzalloc(sizeof(*rec), GFP_NOFS);
536 	if (!rec) {
537 		reada_extent_put(root->fs_info, re);
538 		return -1;
539 	}
540 
541 	rec->rc = rc;
542 	rec->generation = generation;
543 	atomic_inc(&rc->elems);
544 
545 	spin_lock(&re->lock);
546 	list_add_tail(&rec->list, &re->extctl);
547 	spin_unlock(&re->lock);
548 
549 	/* leave the ref on the extent */
550 
551 	return 0;
552 }
553 
554 /*
555  * called with fs_info->reada_lock held
556  */
reada_peer_zones_set_lock(struct reada_zone * zone,int lock)557 static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock)
558 {
559 	int i;
560 	unsigned long index = zone->end >> PAGE_CACHE_SHIFT;
561 
562 	for (i = 0; i < zone->ndevs; ++i) {
563 		struct reada_zone *peer;
564 		peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index);
565 		if (peer && peer->device != zone->device)
566 			peer->locked = lock;
567 	}
568 }
569 
570 /*
571  * called with fs_info->reada_lock held
572  */
reada_pick_zone(struct btrfs_device * dev)573 static int reada_pick_zone(struct btrfs_device *dev)
574 {
575 	struct reada_zone *top_zone = NULL;
576 	struct reada_zone *top_locked_zone = NULL;
577 	u64 top_elems = 0;
578 	u64 top_locked_elems = 0;
579 	unsigned long index = 0;
580 	int ret;
581 
582 	if (dev->reada_curr_zone) {
583 		reada_peer_zones_set_lock(dev->reada_curr_zone, 0);
584 		kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release);
585 		dev->reada_curr_zone = NULL;
586 	}
587 	/* pick the zone with the most elements */
588 	while (1) {
589 		struct reada_zone *zone;
590 
591 		ret = radix_tree_gang_lookup(&dev->reada_zones,
592 					     (void **)&zone, index, 1);
593 		if (ret == 0)
594 			break;
595 		index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
596 		if (zone->locked) {
597 			if (zone->elems > top_locked_elems) {
598 				top_locked_elems = zone->elems;
599 				top_locked_zone = zone;
600 			}
601 		} else {
602 			if (zone->elems > top_elems) {
603 				top_elems = zone->elems;
604 				top_zone = zone;
605 			}
606 		}
607 	}
608 	if (top_zone)
609 		dev->reada_curr_zone = top_zone;
610 	else if (top_locked_zone)
611 		dev->reada_curr_zone = top_locked_zone;
612 	else
613 		return 0;
614 
615 	dev->reada_next = dev->reada_curr_zone->start;
616 	kref_get(&dev->reada_curr_zone->refcnt);
617 	reada_peer_zones_set_lock(dev->reada_curr_zone, 1);
618 
619 	return 1;
620 }
621 
reada_start_machine_dev(struct btrfs_fs_info * fs_info,struct btrfs_device * dev)622 static int reada_start_machine_dev(struct btrfs_fs_info *fs_info,
623 				   struct btrfs_device *dev)
624 {
625 	struct reada_extent *re = NULL;
626 	int mirror_num = 0;
627 	struct extent_buffer *eb = NULL;
628 	u64 logical;
629 	u32 blocksize;
630 	int ret;
631 	int i;
632 	int need_kick = 0;
633 
634 	spin_lock(&fs_info->reada_lock);
635 	if (dev->reada_curr_zone == NULL) {
636 		ret = reada_pick_zone(dev);
637 		if (!ret) {
638 			spin_unlock(&fs_info->reada_lock);
639 			return 0;
640 		}
641 	}
642 	/*
643 	 * FIXME currently we issue the reads one extent at a time. If we have
644 	 * a contiguous block of extents, we could also coagulate them or use
645 	 * plugging to speed things up
646 	 */
647 	ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
648 				     dev->reada_next >> PAGE_CACHE_SHIFT, 1);
649 	if (ret == 0 || re->logical >= dev->reada_curr_zone->end) {
650 		ret = reada_pick_zone(dev);
651 		if (!ret) {
652 			spin_unlock(&fs_info->reada_lock);
653 			return 0;
654 		}
655 		re = NULL;
656 		ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re,
657 					dev->reada_next >> PAGE_CACHE_SHIFT, 1);
658 	}
659 	if (ret == 0) {
660 		spin_unlock(&fs_info->reada_lock);
661 		return 0;
662 	}
663 	dev->reada_next = re->logical + re->blocksize;
664 	kref_get(&re->refcnt);
665 
666 	spin_unlock(&fs_info->reada_lock);
667 
668 	/*
669 	 * find mirror num
670 	 */
671 	for (i = 0; i < re->nzones; ++i) {
672 		if (re->zones[i]->device == dev) {
673 			mirror_num = i + 1;
674 			break;
675 		}
676 	}
677 	logical = re->logical;
678 	blocksize = re->blocksize;
679 
680 	spin_lock(&re->lock);
681 	if (re->scheduled_for == NULL) {
682 		re->scheduled_for = dev;
683 		need_kick = 1;
684 	}
685 	spin_unlock(&re->lock);
686 
687 	reada_extent_put(fs_info, re);
688 
689 	if (!need_kick)
690 		return 0;
691 
692 	atomic_inc(&dev->reada_in_flight);
693 	ret = reada_tree_block_flagged(fs_info->extent_root, logical, blocksize,
694 			 mirror_num, &eb);
695 	if (ret)
696 		__readahead_hook(fs_info->extent_root, NULL, logical, ret);
697 	else if (eb)
698 		__readahead_hook(fs_info->extent_root, eb, eb->start, ret);
699 
700 	if (eb)
701 		free_extent_buffer(eb);
702 
703 	return 1;
704 
705 }
706 
reada_start_machine_worker(struct btrfs_work * work)707 static void reada_start_machine_worker(struct btrfs_work *work)
708 {
709 	struct reada_machine_work *rmw;
710 	struct btrfs_fs_info *fs_info;
711 
712 	rmw = container_of(work, struct reada_machine_work, work);
713 	fs_info = rmw->fs_info;
714 
715 	kfree(rmw);
716 
717 	__reada_start_machine(fs_info);
718 }
719 
__reada_start_machine(struct btrfs_fs_info * fs_info)720 static void __reada_start_machine(struct btrfs_fs_info *fs_info)
721 {
722 	struct btrfs_device *device;
723 	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
724 	u64 enqueued;
725 	u64 total = 0;
726 	int i;
727 
728 	do {
729 		enqueued = 0;
730 		list_for_each_entry(device, &fs_devices->devices, dev_list) {
731 			if (atomic_read(&device->reada_in_flight) <
732 			    MAX_IN_FLIGHT)
733 				enqueued += reada_start_machine_dev(fs_info,
734 								    device);
735 		}
736 		total += enqueued;
737 	} while (enqueued && total < 10000);
738 
739 	if (enqueued == 0)
740 		return;
741 
742 	/*
743 	 * If everything is already in the cache, this is effectively single
744 	 * threaded. To a) not hold the caller for too long and b) to utilize
745 	 * more cores, we broke the loop above after 10000 iterations and now
746 	 * enqueue to workers to finish it. This will distribute the load to
747 	 * the cores.
748 	 */
749 	for (i = 0; i < 2; ++i)
750 		reada_start_machine(fs_info);
751 }
752 
reada_start_machine(struct btrfs_fs_info * fs_info)753 static void reada_start_machine(struct btrfs_fs_info *fs_info)
754 {
755 	struct reada_machine_work *rmw;
756 
757 	rmw = kzalloc(sizeof(*rmw), GFP_NOFS);
758 	if (!rmw) {
759 		/* FIXME we cannot handle this properly right now */
760 		BUG();
761 	}
762 	rmw->work.func = reada_start_machine_worker;
763 	rmw->fs_info = fs_info;
764 
765 	btrfs_queue_worker(&fs_info->readahead_workers, &rmw->work);
766 }
767 
768 #ifdef DEBUG
dump_devs(struct btrfs_fs_info * fs_info,int all)769 static void dump_devs(struct btrfs_fs_info *fs_info, int all)
770 {
771 	struct btrfs_device *device;
772 	struct btrfs_fs_devices *fs_devices = fs_info->fs_devices;
773 	unsigned long index;
774 	int ret;
775 	int i;
776 	int j;
777 	int cnt;
778 
779 	spin_lock(&fs_info->reada_lock);
780 	list_for_each_entry(device, &fs_devices->devices, dev_list) {
781 		printk(KERN_DEBUG "dev %lld has %d in flight\n", device->devid,
782 			atomic_read(&device->reada_in_flight));
783 		index = 0;
784 		while (1) {
785 			struct reada_zone *zone;
786 			ret = radix_tree_gang_lookup(&device->reada_zones,
787 						     (void **)&zone, index, 1);
788 			if (ret == 0)
789 				break;
790 			printk(KERN_DEBUG "  zone %llu-%llu elems %llu locked "
791 				"%d devs", zone->start, zone->end, zone->elems,
792 				zone->locked);
793 			for (j = 0; j < zone->ndevs; ++j) {
794 				printk(KERN_CONT " %lld",
795 					zone->devs[j]->devid);
796 			}
797 			if (device->reada_curr_zone == zone)
798 				printk(KERN_CONT " curr off %llu",
799 					device->reada_next - zone->start);
800 			printk(KERN_CONT "\n");
801 			index = (zone->end >> PAGE_CACHE_SHIFT) + 1;
802 		}
803 		cnt = 0;
804 		index = 0;
805 		while (all) {
806 			struct reada_extent *re = NULL;
807 
808 			ret = radix_tree_gang_lookup(&device->reada_extents,
809 						     (void **)&re, index, 1);
810 			if (ret == 0)
811 				break;
812 			printk(KERN_DEBUG
813 				"  re: logical %llu size %u empty %d for %lld",
814 				re->logical, re->blocksize,
815 				list_empty(&re->extctl), re->scheduled_for ?
816 				re->scheduled_for->devid : -1);
817 
818 			for (i = 0; i < re->nzones; ++i) {
819 				printk(KERN_CONT " zone %llu-%llu devs",
820 					re->zones[i]->start,
821 					re->zones[i]->end);
822 				for (j = 0; j < re->zones[i]->ndevs; ++j) {
823 					printk(KERN_CONT " %lld",
824 						re->zones[i]->devs[j]->devid);
825 				}
826 			}
827 			printk(KERN_CONT "\n");
828 			index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
829 			if (++cnt > 15)
830 				break;
831 		}
832 	}
833 
834 	index = 0;
835 	cnt = 0;
836 	while (all) {
837 		struct reada_extent *re = NULL;
838 
839 		ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re,
840 					     index, 1);
841 		if (ret == 0)
842 			break;
843 		if (!re->scheduled_for) {
844 			index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
845 			continue;
846 		}
847 		printk(KERN_DEBUG
848 			"re: logical %llu size %u list empty %d for %lld",
849 			re->logical, re->blocksize, list_empty(&re->extctl),
850 			re->scheduled_for ? re->scheduled_for->devid : -1);
851 		for (i = 0; i < re->nzones; ++i) {
852 			printk(KERN_CONT " zone %llu-%llu devs",
853 				re->zones[i]->start,
854 				re->zones[i]->end);
855 			for (i = 0; i < re->nzones; ++i) {
856 				printk(KERN_CONT " zone %llu-%llu devs",
857 					re->zones[i]->start,
858 					re->zones[i]->end);
859 				for (j = 0; j < re->zones[i]->ndevs; ++j) {
860 					printk(KERN_CONT " %lld",
861 						re->zones[i]->devs[j]->devid);
862 				}
863 			}
864 		}
865 		printk(KERN_CONT "\n");
866 		index = (re->logical >> PAGE_CACHE_SHIFT) + 1;
867 	}
868 	spin_unlock(&fs_info->reada_lock);
869 }
870 #endif
871 
872 /*
873  * interface
874  */
btrfs_reada_add(struct btrfs_root * root,struct btrfs_key * key_start,struct btrfs_key * key_end)875 struct reada_control *btrfs_reada_add(struct btrfs_root *root,
876 			struct btrfs_key *key_start, struct btrfs_key *key_end)
877 {
878 	struct reada_control *rc;
879 	u64 start;
880 	u64 generation;
881 	int level;
882 	struct extent_buffer *node;
883 	static struct btrfs_key max_key = {
884 		.objectid = (u64)-1,
885 		.type = (u8)-1,
886 		.offset = (u64)-1
887 	};
888 
889 	rc = kzalloc(sizeof(*rc), GFP_NOFS);
890 	if (!rc)
891 		return ERR_PTR(-ENOMEM);
892 
893 	rc->root = root;
894 	rc->key_start = *key_start;
895 	rc->key_end = *key_end;
896 	atomic_set(&rc->elems, 0);
897 	init_waitqueue_head(&rc->wait);
898 	kref_init(&rc->refcnt);
899 	kref_get(&rc->refcnt); /* one ref for having elements */
900 
901 	node = btrfs_root_node(root);
902 	start = node->start;
903 	level = btrfs_header_level(node);
904 	generation = btrfs_header_generation(node);
905 	free_extent_buffer(node);
906 
907 	reada_add_block(rc, start, &max_key, level, generation);
908 
909 	reada_start_machine(root->fs_info);
910 
911 	return rc;
912 }
913 
914 #ifdef DEBUG
btrfs_reada_wait(void * handle)915 int btrfs_reada_wait(void *handle)
916 {
917 	struct reada_control *rc = handle;
918 
919 	while (atomic_read(&rc->elems)) {
920 		wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0,
921 				   5 * HZ);
922 		dump_devs(rc->root->fs_info, rc->elems < 10 ? 1 : 0);
923 	}
924 
925 	dump_devs(rc->root->fs_info, rc->elems < 10 ? 1 : 0);
926 
927 	kref_put(&rc->refcnt, reada_control_release);
928 
929 	return 0;
930 }
931 #else
btrfs_reada_wait(void * handle)932 int btrfs_reada_wait(void *handle)
933 {
934 	struct reada_control *rc = handle;
935 
936 	while (atomic_read(&rc->elems)) {
937 		wait_event(rc->wait, atomic_read(&rc->elems) == 0);
938 	}
939 
940 	kref_put(&rc->refcnt, reada_control_release);
941 
942 	return 0;
943 }
944 #endif
945 
btrfs_reada_detach(void * handle)946 void btrfs_reada_detach(void *handle)
947 {
948 	struct reada_control *rc = handle;
949 
950 	kref_put(&rc->refcnt, reada_control_release);
951 }
952