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