xref: /linux/drivers/md/bcache/btree.c (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
1 /*
2  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
3  *
4  * Uses a block device as cache for other block devices; optimized for SSDs.
5  * All allocation is done in buckets, which should match the erase block size
6  * of the device.
7  *
8  * Buckets containing cached data are kept on a heap sorted by priority;
9  * bucket priority is increased on cache hit, and periodically all the buckets
10  * on the heap have their priority scaled down. This currently is just used as
11  * an LRU but in the future should allow for more intelligent heuristics.
12  *
13  * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
14  * counter. Garbage collection is used to remove stale pointers.
15  *
16  * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
17  * as keys are inserted we only sort the pages that have not yet been written.
18  * When garbage collection is run, we resort the entire node.
19  *
20  * All configuration is done via sysfs; see Documentation/bcache.txt.
21  */
22 
23 #include "bcache.h"
24 #include "btree.h"
25 #include "debug.h"
26 #include "extents.h"
27 
28 #include <linux/slab.h>
29 #include <linux/bitops.h>
30 #include <linux/hash.h>
31 #include <linux/kthread.h>
32 #include <linux/prefetch.h>
33 #include <linux/random.h>
34 #include <linux/rcupdate.h>
35 #include <linux/sched/clock.h>
36 #include <linux/rculist.h>
37 
38 #include <trace/events/bcache.h>
39 
40 /*
41  * Todo:
42  * register_bcache: Return errors out to userspace correctly
43  *
44  * Writeback: don't undirty key until after a cache flush
45  *
46  * Create an iterator for key pointers
47  *
48  * On btree write error, mark bucket such that it won't be freed from the cache
49  *
50  * Journalling:
51  *   Check for bad keys in replay
52  *   Propagate barriers
53  *   Refcount journal entries in journal_replay
54  *
55  * Garbage collection:
56  *   Finish incremental gc
57  *   Gc should free old UUIDs, data for invalid UUIDs
58  *
59  * Provide a way to list backing device UUIDs we have data cached for, and
60  * probably how long it's been since we've seen them, and a way to invalidate
61  * dirty data for devices that will never be attached again
62  *
63  * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
64  * that based on that and how much dirty data we have we can keep writeback
65  * from being starved
66  *
67  * Add a tracepoint or somesuch to watch for writeback starvation
68  *
69  * When btree depth > 1 and splitting an interior node, we have to make sure
70  * alloc_bucket() cannot fail. This should be true but is not completely
71  * obvious.
72  *
73  * Plugging?
74  *
75  * If data write is less than hard sector size of ssd, round up offset in open
76  * bucket to the next whole sector
77  *
78  * Superblock needs to be fleshed out for multiple cache devices
79  *
80  * Add a sysfs tunable for the number of writeback IOs in flight
81  *
82  * Add a sysfs tunable for the number of open data buckets
83  *
84  * IO tracking: Can we track when one process is doing io on behalf of another?
85  * IO tracking: Don't use just an average, weigh more recent stuff higher
86  *
87  * Test module load/unload
88  */
89 
90 #define MAX_NEED_GC		64
91 #define MAX_SAVE_PRIO		72
92 
93 #define PTR_DIRTY_BIT		(((uint64_t) 1 << 36))
94 
95 #define PTR_HASH(c, k)							\
96 	(((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
97 
98 #define insert_lock(s, b)	((b)->level <= (s)->lock)
99 
100 /*
101  * These macros are for recursing down the btree - they handle the details of
102  * locking and looking up nodes in the cache for you. They're best treated as
103  * mere syntax when reading code that uses them.
104  *
105  * op->lock determines whether we take a read or a write lock at a given depth.
106  * If you've got a read lock and find that you need a write lock (i.e. you're
107  * going to have to split), set op->lock and return -EINTR; btree_root() will
108  * call you again and you'll have the correct lock.
109  */
110 
111 /**
112  * btree - recurse down the btree on a specified key
113  * @fn:		function to call, which will be passed the child node
114  * @key:	key to recurse on
115  * @b:		parent btree node
116  * @op:		pointer to struct btree_op
117  */
118 #define btree(fn, key, b, op, ...)					\
119 ({									\
120 	int _r, l = (b)->level - 1;					\
121 	bool _w = l <= (op)->lock;					\
122 	struct btree *_child = bch_btree_node_get((b)->c, op, key, l,	\
123 						  _w, b);		\
124 	if (!IS_ERR(_child)) {						\
125 		_r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__);	\
126 		rw_unlock(_w, _child);					\
127 	} else								\
128 		_r = PTR_ERR(_child);					\
129 	_r;								\
130 })
131 
132 /**
133  * btree_root - call a function on the root of the btree
134  * @fn:		function to call, which will be passed the child node
135  * @c:		cache set
136  * @op:		pointer to struct btree_op
137  */
138 #define btree_root(fn, c, op, ...)					\
139 ({									\
140 	int _r = -EINTR;						\
141 	do {								\
142 		struct btree *_b = (c)->root;				\
143 		bool _w = insert_lock(op, _b);				\
144 		rw_lock(_w, _b, _b->level);				\
145 		if (_b == (c)->root &&					\
146 		    _w == insert_lock(op, _b)) {			\
147 			_r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__);	\
148 		}							\
149 		rw_unlock(_w, _b);					\
150 		bch_cannibalize_unlock(c);				\
151 		if (_r == -EINTR)					\
152 			schedule();					\
153 	} while (_r == -EINTR);						\
154 									\
155 	finish_wait(&(c)->btree_cache_wait, &(op)->wait);		\
156 	_r;								\
157 })
158 
159 static inline struct bset *write_block(struct btree *b)
160 {
161 	return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
162 }
163 
164 static void bch_btree_init_next(struct btree *b)
165 {
166 	/* If not a leaf node, always sort */
167 	if (b->level && b->keys.nsets)
168 		bch_btree_sort(&b->keys, &b->c->sort);
169 	else
170 		bch_btree_sort_lazy(&b->keys, &b->c->sort);
171 
172 	if (b->written < btree_blocks(b))
173 		bch_bset_init_next(&b->keys, write_block(b),
174 				   bset_magic(&b->c->sb));
175 
176 }
177 
178 /* Btree key manipulation */
179 
180 void bkey_put(struct cache_set *c, struct bkey *k)
181 {
182 	unsigned i;
183 
184 	for (i = 0; i < KEY_PTRS(k); i++)
185 		if (ptr_available(c, k, i))
186 			atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
187 }
188 
189 /* Btree IO */
190 
191 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
192 {
193 	uint64_t crc = b->key.ptr[0];
194 	void *data = (void *) i + 8, *end = bset_bkey_last(i);
195 
196 	crc = bch_crc64_update(crc, data, end - data);
197 	return crc ^ 0xffffffffffffffffULL;
198 }
199 
200 void bch_btree_node_read_done(struct btree *b)
201 {
202 	const char *err = "bad btree header";
203 	struct bset *i = btree_bset_first(b);
204 	struct btree_iter *iter;
205 
206 	iter = mempool_alloc(b->c->fill_iter, GFP_NOIO);
207 	iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
208 	iter->used = 0;
209 
210 #ifdef CONFIG_BCACHE_DEBUG
211 	iter->b = &b->keys;
212 #endif
213 
214 	if (!i->seq)
215 		goto err;
216 
217 	for (;
218 	     b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
219 	     i = write_block(b)) {
220 		err = "unsupported bset version";
221 		if (i->version > BCACHE_BSET_VERSION)
222 			goto err;
223 
224 		err = "bad btree header";
225 		if (b->written + set_blocks(i, block_bytes(b->c)) >
226 		    btree_blocks(b))
227 			goto err;
228 
229 		err = "bad magic";
230 		if (i->magic != bset_magic(&b->c->sb))
231 			goto err;
232 
233 		err = "bad checksum";
234 		switch (i->version) {
235 		case 0:
236 			if (i->csum != csum_set(i))
237 				goto err;
238 			break;
239 		case BCACHE_BSET_VERSION:
240 			if (i->csum != btree_csum_set(b, i))
241 				goto err;
242 			break;
243 		}
244 
245 		err = "empty set";
246 		if (i != b->keys.set[0].data && !i->keys)
247 			goto err;
248 
249 		bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
250 
251 		b->written += set_blocks(i, block_bytes(b->c));
252 	}
253 
254 	err = "corrupted btree";
255 	for (i = write_block(b);
256 	     bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
257 	     i = ((void *) i) + block_bytes(b->c))
258 		if (i->seq == b->keys.set[0].data->seq)
259 			goto err;
260 
261 	bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
262 
263 	i = b->keys.set[0].data;
264 	err = "short btree key";
265 	if (b->keys.set[0].size &&
266 	    bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
267 		goto err;
268 
269 	if (b->written < btree_blocks(b))
270 		bch_bset_init_next(&b->keys, write_block(b),
271 				   bset_magic(&b->c->sb));
272 out:
273 	mempool_free(iter, b->c->fill_iter);
274 	return;
275 err:
276 	set_btree_node_io_error(b);
277 	bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
278 			    err, PTR_BUCKET_NR(b->c, &b->key, 0),
279 			    bset_block_offset(b, i), i->keys);
280 	goto out;
281 }
282 
283 static void btree_node_read_endio(struct bio *bio)
284 {
285 	struct closure *cl = bio->bi_private;
286 	closure_put(cl);
287 }
288 
289 static void bch_btree_node_read(struct btree *b)
290 {
291 	uint64_t start_time = local_clock();
292 	struct closure cl;
293 	struct bio *bio;
294 
295 	trace_bcache_btree_read(b);
296 
297 	closure_init_stack(&cl);
298 
299 	bio = bch_bbio_alloc(b->c);
300 	bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
301 	bio->bi_end_io	= btree_node_read_endio;
302 	bio->bi_private	= &cl;
303 	bio->bi_opf = REQ_OP_READ | REQ_META;
304 
305 	bch_bio_map(bio, b->keys.set[0].data);
306 
307 	bch_submit_bbio(bio, b->c, &b->key, 0);
308 	closure_sync(&cl);
309 
310 	if (bio->bi_error)
311 		set_btree_node_io_error(b);
312 
313 	bch_bbio_free(bio, b->c);
314 
315 	if (btree_node_io_error(b))
316 		goto err;
317 
318 	bch_btree_node_read_done(b);
319 	bch_time_stats_update(&b->c->btree_read_time, start_time);
320 
321 	return;
322 err:
323 	bch_cache_set_error(b->c, "io error reading bucket %zu",
324 			    PTR_BUCKET_NR(b->c, &b->key, 0));
325 }
326 
327 static void btree_complete_write(struct btree *b, struct btree_write *w)
328 {
329 	if (w->prio_blocked &&
330 	    !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
331 		wake_up_allocators(b->c);
332 
333 	if (w->journal) {
334 		atomic_dec_bug(w->journal);
335 		__closure_wake_up(&b->c->journal.wait);
336 	}
337 
338 	w->prio_blocked	= 0;
339 	w->journal	= NULL;
340 }
341 
342 static void btree_node_write_unlock(struct closure *cl)
343 {
344 	struct btree *b = container_of(cl, struct btree, io);
345 
346 	up(&b->io_mutex);
347 }
348 
349 static void __btree_node_write_done(struct closure *cl)
350 {
351 	struct btree *b = container_of(cl, struct btree, io);
352 	struct btree_write *w = btree_prev_write(b);
353 
354 	bch_bbio_free(b->bio, b->c);
355 	b->bio = NULL;
356 	btree_complete_write(b, w);
357 
358 	if (btree_node_dirty(b))
359 		schedule_delayed_work(&b->work, 30 * HZ);
360 
361 	closure_return_with_destructor(cl, btree_node_write_unlock);
362 }
363 
364 static void btree_node_write_done(struct closure *cl)
365 {
366 	struct btree *b = container_of(cl, struct btree, io);
367 
368 	bio_free_pages(b->bio);
369 	__btree_node_write_done(cl);
370 }
371 
372 static void btree_node_write_endio(struct bio *bio)
373 {
374 	struct closure *cl = bio->bi_private;
375 	struct btree *b = container_of(cl, struct btree, io);
376 
377 	if (bio->bi_error)
378 		set_btree_node_io_error(b);
379 
380 	bch_bbio_count_io_errors(b->c, bio, bio->bi_error, "writing btree");
381 	closure_put(cl);
382 }
383 
384 static void do_btree_node_write(struct btree *b)
385 {
386 	struct closure *cl = &b->io;
387 	struct bset *i = btree_bset_last(b);
388 	BKEY_PADDED(key) k;
389 
390 	i->version	= BCACHE_BSET_VERSION;
391 	i->csum		= btree_csum_set(b, i);
392 
393 	BUG_ON(b->bio);
394 	b->bio = bch_bbio_alloc(b->c);
395 
396 	b->bio->bi_end_io	= btree_node_write_endio;
397 	b->bio->bi_private	= cl;
398 	b->bio->bi_iter.bi_size	= roundup(set_bytes(i), block_bytes(b->c));
399 	b->bio->bi_opf		= REQ_OP_WRITE | REQ_META | REQ_FUA;
400 	bch_bio_map(b->bio, i);
401 
402 	/*
403 	 * If we're appending to a leaf node, we don't technically need FUA -
404 	 * this write just needs to be persisted before the next journal write,
405 	 * which will be marked FLUSH|FUA.
406 	 *
407 	 * Similarly if we're writing a new btree root - the pointer is going to
408 	 * be in the next journal entry.
409 	 *
410 	 * But if we're writing a new btree node (that isn't a root) or
411 	 * appending to a non leaf btree node, we need either FUA or a flush
412 	 * when we write the parent with the new pointer. FUA is cheaper than a
413 	 * flush, and writes appending to leaf nodes aren't blocking anything so
414 	 * just make all btree node writes FUA to keep things sane.
415 	 */
416 
417 	bkey_copy(&k.key, &b->key);
418 	SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
419 		       bset_sector_offset(&b->keys, i));
420 
421 	if (!bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
422 		int j;
423 		struct bio_vec *bv;
424 		void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
425 
426 		bio_for_each_segment_all(bv, b->bio, j)
427 			memcpy(page_address(bv->bv_page),
428 			       base + j * PAGE_SIZE, PAGE_SIZE);
429 
430 		bch_submit_bbio(b->bio, b->c, &k.key, 0);
431 
432 		continue_at(cl, btree_node_write_done, NULL);
433 	} else {
434 		b->bio->bi_vcnt = 0;
435 		bch_bio_map(b->bio, i);
436 
437 		bch_submit_bbio(b->bio, b->c, &k.key, 0);
438 
439 		closure_sync(cl);
440 		continue_at_nobarrier(cl, __btree_node_write_done, NULL);
441 	}
442 }
443 
444 void __bch_btree_node_write(struct btree *b, struct closure *parent)
445 {
446 	struct bset *i = btree_bset_last(b);
447 
448 	lockdep_assert_held(&b->write_lock);
449 
450 	trace_bcache_btree_write(b);
451 
452 	BUG_ON(current->bio_list);
453 	BUG_ON(b->written >= btree_blocks(b));
454 	BUG_ON(b->written && !i->keys);
455 	BUG_ON(btree_bset_first(b)->seq != i->seq);
456 	bch_check_keys(&b->keys, "writing");
457 
458 	cancel_delayed_work(&b->work);
459 
460 	/* If caller isn't waiting for write, parent refcount is cache set */
461 	down(&b->io_mutex);
462 	closure_init(&b->io, parent ?: &b->c->cl);
463 
464 	clear_bit(BTREE_NODE_dirty,	 &b->flags);
465 	change_bit(BTREE_NODE_write_idx, &b->flags);
466 
467 	do_btree_node_write(b);
468 
469 	atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
470 			&PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
471 
472 	b->written += set_blocks(i, block_bytes(b->c));
473 }
474 
475 void bch_btree_node_write(struct btree *b, struct closure *parent)
476 {
477 	unsigned nsets = b->keys.nsets;
478 
479 	lockdep_assert_held(&b->lock);
480 
481 	__bch_btree_node_write(b, parent);
482 
483 	/*
484 	 * do verify if there was more than one set initially (i.e. we did a
485 	 * sort) and we sorted down to a single set:
486 	 */
487 	if (nsets && !b->keys.nsets)
488 		bch_btree_verify(b);
489 
490 	bch_btree_init_next(b);
491 }
492 
493 static void bch_btree_node_write_sync(struct btree *b)
494 {
495 	struct closure cl;
496 
497 	closure_init_stack(&cl);
498 
499 	mutex_lock(&b->write_lock);
500 	bch_btree_node_write(b, &cl);
501 	mutex_unlock(&b->write_lock);
502 
503 	closure_sync(&cl);
504 }
505 
506 static void btree_node_write_work(struct work_struct *w)
507 {
508 	struct btree *b = container_of(to_delayed_work(w), struct btree, work);
509 
510 	mutex_lock(&b->write_lock);
511 	if (btree_node_dirty(b))
512 		__bch_btree_node_write(b, NULL);
513 	mutex_unlock(&b->write_lock);
514 }
515 
516 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
517 {
518 	struct bset *i = btree_bset_last(b);
519 	struct btree_write *w = btree_current_write(b);
520 
521 	lockdep_assert_held(&b->write_lock);
522 
523 	BUG_ON(!b->written);
524 	BUG_ON(!i->keys);
525 
526 	if (!btree_node_dirty(b))
527 		schedule_delayed_work(&b->work, 30 * HZ);
528 
529 	set_btree_node_dirty(b);
530 
531 	if (journal_ref) {
532 		if (w->journal &&
533 		    journal_pin_cmp(b->c, w->journal, journal_ref)) {
534 			atomic_dec_bug(w->journal);
535 			w->journal = NULL;
536 		}
537 
538 		if (!w->journal) {
539 			w->journal = journal_ref;
540 			atomic_inc(w->journal);
541 		}
542 	}
543 
544 	/* Force write if set is too big */
545 	if (set_bytes(i) > PAGE_SIZE - 48 &&
546 	    !current->bio_list)
547 		bch_btree_node_write(b, NULL);
548 }
549 
550 /*
551  * Btree in memory cache - allocation/freeing
552  * mca -> memory cache
553  */
554 
555 #define mca_reserve(c)	(((c->root && c->root->level)		\
556 			  ? c->root->level : 1) * 8 + 16)
557 #define mca_can_free(c)						\
558 	max_t(int, 0, c->btree_cache_used - mca_reserve(c))
559 
560 static void mca_data_free(struct btree *b)
561 {
562 	BUG_ON(b->io_mutex.count != 1);
563 
564 	bch_btree_keys_free(&b->keys);
565 
566 	b->c->btree_cache_used--;
567 	list_move(&b->list, &b->c->btree_cache_freed);
568 }
569 
570 static void mca_bucket_free(struct btree *b)
571 {
572 	BUG_ON(btree_node_dirty(b));
573 
574 	b->key.ptr[0] = 0;
575 	hlist_del_init_rcu(&b->hash);
576 	list_move(&b->list, &b->c->btree_cache_freeable);
577 }
578 
579 static unsigned btree_order(struct bkey *k)
580 {
581 	return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
582 }
583 
584 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
585 {
586 	if (!bch_btree_keys_alloc(&b->keys,
587 				  max_t(unsigned,
588 					ilog2(b->c->btree_pages),
589 					btree_order(k)),
590 				  gfp)) {
591 		b->c->btree_cache_used++;
592 		list_move(&b->list, &b->c->btree_cache);
593 	} else {
594 		list_move(&b->list, &b->c->btree_cache_freed);
595 	}
596 }
597 
598 static struct btree *mca_bucket_alloc(struct cache_set *c,
599 				      struct bkey *k, gfp_t gfp)
600 {
601 	struct btree *b = kzalloc(sizeof(struct btree), gfp);
602 	if (!b)
603 		return NULL;
604 
605 	init_rwsem(&b->lock);
606 	lockdep_set_novalidate_class(&b->lock);
607 	mutex_init(&b->write_lock);
608 	lockdep_set_novalidate_class(&b->write_lock);
609 	INIT_LIST_HEAD(&b->list);
610 	INIT_DELAYED_WORK(&b->work, btree_node_write_work);
611 	b->c = c;
612 	sema_init(&b->io_mutex, 1);
613 
614 	mca_data_alloc(b, k, gfp);
615 	return b;
616 }
617 
618 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
619 {
620 	struct closure cl;
621 
622 	closure_init_stack(&cl);
623 	lockdep_assert_held(&b->c->bucket_lock);
624 
625 	if (!down_write_trylock(&b->lock))
626 		return -ENOMEM;
627 
628 	BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
629 
630 	if (b->keys.page_order < min_order)
631 		goto out_unlock;
632 
633 	if (!flush) {
634 		if (btree_node_dirty(b))
635 			goto out_unlock;
636 
637 		if (down_trylock(&b->io_mutex))
638 			goto out_unlock;
639 		up(&b->io_mutex);
640 	}
641 
642 	mutex_lock(&b->write_lock);
643 	if (btree_node_dirty(b))
644 		__bch_btree_node_write(b, &cl);
645 	mutex_unlock(&b->write_lock);
646 
647 	closure_sync(&cl);
648 
649 	/* wait for any in flight btree write */
650 	down(&b->io_mutex);
651 	up(&b->io_mutex);
652 
653 	return 0;
654 out_unlock:
655 	rw_unlock(true, b);
656 	return -ENOMEM;
657 }
658 
659 static unsigned long bch_mca_scan(struct shrinker *shrink,
660 				  struct shrink_control *sc)
661 {
662 	struct cache_set *c = container_of(shrink, struct cache_set, shrink);
663 	struct btree *b, *t;
664 	unsigned long i, nr = sc->nr_to_scan;
665 	unsigned long freed = 0;
666 
667 	if (c->shrinker_disabled)
668 		return SHRINK_STOP;
669 
670 	if (c->btree_cache_alloc_lock)
671 		return SHRINK_STOP;
672 
673 	/* Return -1 if we can't do anything right now */
674 	if (sc->gfp_mask & __GFP_IO)
675 		mutex_lock(&c->bucket_lock);
676 	else if (!mutex_trylock(&c->bucket_lock))
677 		return -1;
678 
679 	/*
680 	 * It's _really_ critical that we don't free too many btree nodes - we
681 	 * have to always leave ourselves a reserve. The reserve is how we
682 	 * guarantee that allocating memory for a new btree node can always
683 	 * succeed, so that inserting keys into the btree can always succeed and
684 	 * IO can always make forward progress:
685 	 */
686 	nr /= c->btree_pages;
687 	nr = min_t(unsigned long, nr, mca_can_free(c));
688 
689 	i = 0;
690 	list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
691 		if (freed >= nr)
692 			break;
693 
694 		if (++i > 3 &&
695 		    !mca_reap(b, 0, false)) {
696 			mca_data_free(b);
697 			rw_unlock(true, b);
698 			freed++;
699 		}
700 	}
701 
702 	for (i = 0; (nr--) && i < c->btree_cache_used; i++) {
703 		if (list_empty(&c->btree_cache))
704 			goto out;
705 
706 		b = list_first_entry(&c->btree_cache, struct btree, list);
707 		list_rotate_left(&c->btree_cache);
708 
709 		if (!b->accessed &&
710 		    !mca_reap(b, 0, false)) {
711 			mca_bucket_free(b);
712 			mca_data_free(b);
713 			rw_unlock(true, b);
714 			freed++;
715 		} else
716 			b->accessed = 0;
717 	}
718 out:
719 	mutex_unlock(&c->bucket_lock);
720 	return freed;
721 }
722 
723 static unsigned long bch_mca_count(struct shrinker *shrink,
724 				   struct shrink_control *sc)
725 {
726 	struct cache_set *c = container_of(shrink, struct cache_set, shrink);
727 
728 	if (c->shrinker_disabled)
729 		return 0;
730 
731 	if (c->btree_cache_alloc_lock)
732 		return 0;
733 
734 	return mca_can_free(c) * c->btree_pages;
735 }
736 
737 void bch_btree_cache_free(struct cache_set *c)
738 {
739 	struct btree *b;
740 	struct closure cl;
741 	closure_init_stack(&cl);
742 
743 	if (c->shrink.list.next)
744 		unregister_shrinker(&c->shrink);
745 
746 	mutex_lock(&c->bucket_lock);
747 
748 #ifdef CONFIG_BCACHE_DEBUG
749 	if (c->verify_data)
750 		list_move(&c->verify_data->list, &c->btree_cache);
751 
752 	free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
753 #endif
754 
755 	list_splice(&c->btree_cache_freeable,
756 		    &c->btree_cache);
757 
758 	while (!list_empty(&c->btree_cache)) {
759 		b = list_first_entry(&c->btree_cache, struct btree, list);
760 
761 		if (btree_node_dirty(b))
762 			btree_complete_write(b, btree_current_write(b));
763 		clear_bit(BTREE_NODE_dirty, &b->flags);
764 
765 		mca_data_free(b);
766 	}
767 
768 	while (!list_empty(&c->btree_cache_freed)) {
769 		b = list_first_entry(&c->btree_cache_freed,
770 				     struct btree, list);
771 		list_del(&b->list);
772 		cancel_delayed_work_sync(&b->work);
773 		kfree(b);
774 	}
775 
776 	mutex_unlock(&c->bucket_lock);
777 }
778 
779 int bch_btree_cache_alloc(struct cache_set *c)
780 {
781 	unsigned i;
782 
783 	for (i = 0; i < mca_reserve(c); i++)
784 		if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
785 			return -ENOMEM;
786 
787 	list_splice_init(&c->btree_cache,
788 			 &c->btree_cache_freeable);
789 
790 #ifdef CONFIG_BCACHE_DEBUG
791 	mutex_init(&c->verify_lock);
792 
793 	c->verify_ondisk = (void *)
794 		__get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
795 
796 	c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
797 
798 	if (c->verify_data &&
799 	    c->verify_data->keys.set->data)
800 		list_del_init(&c->verify_data->list);
801 	else
802 		c->verify_data = NULL;
803 #endif
804 
805 	c->shrink.count_objects = bch_mca_count;
806 	c->shrink.scan_objects = bch_mca_scan;
807 	c->shrink.seeks = 4;
808 	c->shrink.batch = c->btree_pages * 2;
809 	register_shrinker(&c->shrink);
810 
811 	return 0;
812 }
813 
814 /* Btree in memory cache - hash table */
815 
816 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
817 {
818 	return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
819 }
820 
821 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
822 {
823 	struct btree *b;
824 
825 	rcu_read_lock();
826 	hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
827 		if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
828 			goto out;
829 	b = NULL;
830 out:
831 	rcu_read_unlock();
832 	return b;
833 }
834 
835 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
836 {
837 	struct task_struct *old;
838 
839 	old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
840 	if (old && old != current) {
841 		if (op)
842 			prepare_to_wait(&c->btree_cache_wait, &op->wait,
843 					TASK_UNINTERRUPTIBLE);
844 		return -EINTR;
845 	}
846 
847 	return 0;
848 }
849 
850 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
851 				     struct bkey *k)
852 {
853 	struct btree *b;
854 
855 	trace_bcache_btree_cache_cannibalize(c);
856 
857 	if (mca_cannibalize_lock(c, op))
858 		return ERR_PTR(-EINTR);
859 
860 	list_for_each_entry_reverse(b, &c->btree_cache, list)
861 		if (!mca_reap(b, btree_order(k), false))
862 			return b;
863 
864 	list_for_each_entry_reverse(b, &c->btree_cache, list)
865 		if (!mca_reap(b, btree_order(k), true))
866 			return b;
867 
868 	WARN(1, "btree cache cannibalize failed\n");
869 	return ERR_PTR(-ENOMEM);
870 }
871 
872 /*
873  * We can only have one thread cannibalizing other cached btree nodes at a time,
874  * or we'll deadlock. We use an open coded mutex to ensure that, which a
875  * cannibalize_bucket() will take. This means every time we unlock the root of
876  * the btree, we need to release this lock if we have it held.
877  */
878 static void bch_cannibalize_unlock(struct cache_set *c)
879 {
880 	if (c->btree_cache_alloc_lock == current) {
881 		c->btree_cache_alloc_lock = NULL;
882 		wake_up(&c->btree_cache_wait);
883 	}
884 }
885 
886 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
887 			       struct bkey *k, int level)
888 {
889 	struct btree *b;
890 
891 	BUG_ON(current->bio_list);
892 
893 	lockdep_assert_held(&c->bucket_lock);
894 
895 	if (mca_find(c, k))
896 		return NULL;
897 
898 	/* btree_free() doesn't free memory; it sticks the node on the end of
899 	 * the list. Check if there's any freed nodes there:
900 	 */
901 	list_for_each_entry(b, &c->btree_cache_freeable, list)
902 		if (!mca_reap(b, btree_order(k), false))
903 			goto out;
904 
905 	/* We never free struct btree itself, just the memory that holds the on
906 	 * disk node. Check the freed list before allocating a new one:
907 	 */
908 	list_for_each_entry(b, &c->btree_cache_freed, list)
909 		if (!mca_reap(b, 0, false)) {
910 			mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
911 			if (!b->keys.set[0].data)
912 				goto err;
913 			else
914 				goto out;
915 		}
916 
917 	b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
918 	if (!b)
919 		goto err;
920 
921 	BUG_ON(!down_write_trylock(&b->lock));
922 	if (!b->keys.set->data)
923 		goto err;
924 out:
925 	BUG_ON(b->io_mutex.count != 1);
926 
927 	bkey_copy(&b->key, k);
928 	list_move(&b->list, &c->btree_cache);
929 	hlist_del_init_rcu(&b->hash);
930 	hlist_add_head_rcu(&b->hash, mca_hash(c, k));
931 
932 	lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
933 	b->parent	= (void *) ~0UL;
934 	b->flags	= 0;
935 	b->written	= 0;
936 	b->level	= level;
937 
938 	if (!b->level)
939 		bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
940 				    &b->c->expensive_debug_checks);
941 	else
942 		bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
943 				    &b->c->expensive_debug_checks);
944 
945 	return b;
946 err:
947 	if (b)
948 		rw_unlock(true, b);
949 
950 	b = mca_cannibalize(c, op, k);
951 	if (!IS_ERR(b))
952 		goto out;
953 
954 	return b;
955 }
956 
957 /**
958  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
959  * in from disk if necessary.
960  *
961  * If IO is necessary and running under generic_make_request, returns -EAGAIN.
962  *
963  * The btree node will have either a read or a write lock held, depending on
964  * level and op->lock.
965  */
966 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
967 				 struct bkey *k, int level, bool write,
968 				 struct btree *parent)
969 {
970 	int i = 0;
971 	struct btree *b;
972 
973 	BUG_ON(level < 0);
974 retry:
975 	b = mca_find(c, k);
976 
977 	if (!b) {
978 		if (current->bio_list)
979 			return ERR_PTR(-EAGAIN);
980 
981 		mutex_lock(&c->bucket_lock);
982 		b = mca_alloc(c, op, k, level);
983 		mutex_unlock(&c->bucket_lock);
984 
985 		if (!b)
986 			goto retry;
987 		if (IS_ERR(b))
988 			return b;
989 
990 		bch_btree_node_read(b);
991 
992 		if (!write)
993 			downgrade_write(&b->lock);
994 	} else {
995 		rw_lock(write, b, level);
996 		if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
997 			rw_unlock(write, b);
998 			goto retry;
999 		}
1000 		BUG_ON(b->level != level);
1001 	}
1002 
1003 	b->parent = parent;
1004 	b->accessed = 1;
1005 
1006 	for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1007 		prefetch(b->keys.set[i].tree);
1008 		prefetch(b->keys.set[i].data);
1009 	}
1010 
1011 	for (; i <= b->keys.nsets; i++)
1012 		prefetch(b->keys.set[i].data);
1013 
1014 	if (btree_node_io_error(b)) {
1015 		rw_unlock(write, b);
1016 		return ERR_PTR(-EIO);
1017 	}
1018 
1019 	BUG_ON(!b->written);
1020 
1021 	return b;
1022 }
1023 
1024 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1025 {
1026 	struct btree *b;
1027 
1028 	mutex_lock(&parent->c->bucket_lock);
1029 	b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1030 	mutex_unlock(&parent->c->bucket_lock);
1031 
1032 	if (!IS_ERR_OR_NULL(b)) {
1033 		b->parent = parent;
1034 		bch_btree_node_read(b);
1035 		rw_unlock(true, b);
1036 	}
1037 }
1038 
1039 /* Btree alloc */
1040 
1041 static void btree_node_free(struct btree *b)
1042 {
1043 	trace_bcache_btree_node_free(b);
1044 
1045 	BUG_ON(b == b->c->root);
1046 
1047 	mutex_lock(&b->write_lock);
1048 
1049 	if (btree_node_dirty(b))
1050 		btree_complete_write(b, btree_current_write(b));
1051 	clear_bit(BTREE_NODE_dirty, &b->flags);
1052 
1053 	mutex_unlock(&b->write_lock);
1054 
1055 	cancel_delayed_work(&b->work);
1056 
1057 	mutex_lock(&b->c->bucket_lock);
1058 	bch_bucket_free(b->c, &b->key);
1059 	mca_bucket_free(b);
1060 	mutex_unlock(&b->c->bucket_lock);
1061 }
1062 
1063 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1064 				     int level, bool wait,
1065 				     struct btree *parent)
1066 {
1067 	BKEY_PADDED(key) k;
1068 	struct btree *b = ERR_PTR(-EAGAIN);
1069 
1070 	mutex_lock(&c->bucket_lock);
1071 retry:
1072 	if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1073 		goto err;
1074 
1075 	bkey_put(c, &k.key);
1076 	SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1077 
1078 	b = mca_alloc(c, op, &k.key, level);
1079 	if (IS_ERR(b))
1080 		goto err_free;
1081 
1082 	if (!b) {
1083 		cache_bug(c,
1084 			"Tried to allocate bucket that was in btree cache");
1085 		goto retry;
1086 	}
1087 
1088 	b->accessed = 1;
1089 	b->parent = parent;
1090 	bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1091 
1092 	mutex_unlock(&c->bucket_lock);
1093 
1094 	trace_bcache_btree_node_alloc(b);
1095 	return b;
1096 err_free:
1097 	bch_bucket_free(c, &k.key);
1098 err:
1099 	mutex_unlock(&c->bucket_lock);
1100 
1101 	trace_bcache_btree_node_alloc_fail(c);
1102 	return b;
1103 }
1104 
1105 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1106 					  struct btree_op *op, int level,
1107 					  struct btree *parent)
1108 {
1109 	return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1110 }
1111 
1112 static struct btree *btree_node_alloc_replacement(struct btree *b,
1113 						  struct btree_op *op)
1114 {
1115 	struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1116 	if (!IS_ERR_OR_NULL(n)) {
1117 		mutex_lock(&n->write_lock);
1118 		bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1119 		bkey_copy_key(&n->key, &b->key);
1120 		mutex_unlock(&n->write_lock);
1121 	}
1122 
1123 	return n;
1124 }
1125 
1126 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1127 {
1128 	unsigned i;
1129 
1130 	mutex_lock(&b->c->bucket_lock);
1131 
1132 	atomic_inc(&b->c->prio_blocked);
1133 
1134 	bkey_copy(k, &b->key);
1135 	bkey_copy_key(k, &ZERO_KEY);
1136 
1137 	for (i = 0; i < KEY_PTRS(k); i++)
1138 		SET_PTR_GEN(k, i,
1139 			    bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1140 					PTR_BUCKET(b->c, &b->key, i)));
1141 
1142 	mutex_unlock(&b->c->bucket_lock);
1143 }
1144 
1145 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1146 {
1147 	struct cache_set *c = b->c;
1148 	struct cache *ca;
1149 	unsigned i, reserve = (c->root->level - b->level) * 2 + 1;
1150 
1151 	mutex_lock(&c->bucket_lock);
1152 
1153 	for_each_cache(ca, c, i)
1154 		if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1155 			if (op)
1156 				prepare_to_wait(&c->btree_cache_wait, &op->wait,
1157 						TASK_UNINTERRUPTIBLE);
1158 			mutex_unlock(&c->bucket_lock);
1159 			return -EINTR;
1160 		}
1161 
1162 	mutex_unlock(&c->bucket_lock);
1163 
1164 	return mca_cannibalize_lock(b->c, op);
1165 }
1166 
1167 /* Garbage collection */
1168 
1169 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1170 				    struct bkey *k)
1171 {
1172 	uint8_t stale = 0;
1173 	unsigned i;
1174 	struct bucket *g;
1175 
1176 	/*
1177 	 * ptr_invalid() can't return true for the keys that mark btree nodes as
1178 	 * freed, but since ptr_bad() returns true we'll never actually use them
1179 	 * for anything and thus we don't want mark their pointers here
1180 	 */
1181 	if (!bkey_cmp(k, &ZERO_KEY))
1182 		return stale;
1183 
1184 	for (i = 0; i < KEY_PTRS(k); i++) {
1185 		if (!ptr_available(c, k, i))
1186 			continue;
1187 
1188 		g = PTR_BUCKET(c, k, i);
1189 
1190 		if (gen_after(g->last_gc, PTR_GEN(k, i)))
1191 			g->last_gc = PTR_GEN(k, i);
1192 
1193 		if (ptr_stale(c, k, i)) {
1194 			stale = max(stale, ptr_stale(c, k, i));
1195 			continue;
1196 		}
1197 
1198 		cache_bug_on(GC_MARK(g) &&
1199 			     (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1200 			     c, "inconsistent ptrs: mark = %llu, level = %i",
1201 			     GC_MARK(g), level);
1202 
1203 		if (level)
1204 			SET_GC_MARK(g, GC_MARK_METADATA);
1205 		else if (KEY_DIRTY(k))
1206 			SET_GC_MARK(g, GC_MARK_DIRTY);
1207 		else if (!GC_MARK(g))
1208 			SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1209 
1210 		/* guard against overflow */
1211 		SET_GC_SECTORS_USED(g, min_t(unsigned,
1212 					     GC_SECTORS_USED(g) + KEY_SIZE(k),
1213 					     MAX_GC_SECTORS_USED));
1214 
1215 		BUG_ON(!GC_SECTORS_USED(g));
1216 	}
1217 
1218 	return stale;
1219 }
1220 
1221 #define btree_mark_key(b, k)	__bch_btree_mark_key(b->c, b->level, k)
1222 
1223 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1224 {
1225 	unsigned i;
1226 
1227 	for (i = 0; i < KEY_PTRS(k); i++)
1228 		if (ptr_available(c, k, i) &&
1229 		    !ptr_stale(c, k, i)) {
1230 			struct bucket *b = PTR_BUCKET(c, k, i);
1231 
1232 			b->gen = PTR_GEN(k, i);
1233 
1234 			if (level && bkey_cmp(k, &ZERO_KEY))
1235 				b->prio = BTREE_PRIO;
1236 			else if (!level && b->prio == BTREE_PRIO)
1237 				b->prio = INITIAL_PRIO;
1238 		}
1239 
1240 	__bch_btree_mark_key(c, level, k);
1241 }
1242 
1243 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1244 {
1245 	uint8_t stale = 0;
1246 	unsigned keys = 0, good_keys = 0;
1247 	struct bkey *k;
1248 	struct btree_iter iter;
1249 	struct bset_tree *t;
1250 
1251 	gc->nodes++;
1252 
1253 	for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1254 		stale = max(stale, btree_mark_key(b, k));
1255 		keys++;
1256 
1257 		if (bch_ptr_bad(&b->keys, k))
1258 			continue;
1259 
1260 		gc->key_bytes += bkey_u64s(k);
1261 		gc->nkeys++;
1262 		good_keys++;
1263 
1264 		gc->data += KEY_SIZE(k);
1265 	}
1266 
1267 	for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1268 		btree_bug_on(t->size &&
1269 			     bset_written(&b->keys, t) &&
1270 			     bkey_cmp(&b->key, &t->end) < 0,
1271 			     b, "found short btree key in gc");
1272 
1273 	if (b->c->gc_always_rewrite)
1274 		return true;
1275 
1276 	if (stale > 10)
1277 		return true;
1278 
1279 	if ((keys - good_keys) * 2 > keys)
1280 		return true;
1281 
1282 	return false;
1283 }
1284 
1285 #define GC_MERGE_NODES	4U
1286 
1287 struct gc_merge_info {
1288 	struct btree	*b;
1289 	unsigned	keys;
1290 };
1291 
1292 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1293 				 struct keylist *, atomic_t *, struct bkey *);
1294 
1295 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1296 			     struct gc_stat *gc, struct gc_merge_info *r)
1297 {
1298 	unsigned i, nodes = 0, keys = 0, blocks;
1299 	struct btree *new_nodes[GC_MERGE_NODES];
1300 	struct keylist keylist;
1301 	struct closure cl;
1302 	struct bkey *k;
1303 
1304 	bch_keylist_init(&keylist);
1305 
1306 	if (btree_check_reserve(b, NULL))
1307 		return 0;
1308 
1309 	memset(new_nodes, 0, sizeof(new_nodes));
1310 	closure_init_stack(&cl);
1311 
1312 	while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1313 		keys += r[nodes++].keys;
1314 
1315 	blocks = btree_default_blocks(b->c) * 2 / 3;
1316 
1317 	if (nodes < 2 ||
1318 	    __set_blocks(b->keys.set[0].data, keys,
1319 			 block_bytes(b->c)) > blocks * (nodes - 1))
1320 		return 0;
1321 
1322 	for (i = 0; i < nodes; i++) {
1323 		new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1324 		if (IS_ERR_OR_NULL(new_nodes[i]))
1325 			goto out_nocoalesce;
1326 	}
1327 
1328 	/*
1329 	 * We have to check the reserve here, after we've allocated our new
1330 	 * nodes, to make sure the insert below will succeed - we also check
1331 	 * before as an optimization to potentially avoid a bunch of expensive
1332 	 * allocs/sorts
1333 	 */
1334 	if (btree_check_reserve(b, NULL))
1335 		goto out_nocoalesce;
1336 
1337 	for (i = 0; i < nodes; i++)
1338 		mutex_lock(&new_nodes[i]->write_lock);
1339 
1340 	for (i = nodes - 1; i > 0; --i) {
1341 		struct bset *n1 = btree_bset_first(new_nodes[i]);
1342 		struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1343 		struct bkey *k, *last = NULL;
1344 
1345 		keys = 0;
1346 
1347 		if (i > 1) {
1348 			for (k = n2->start;
1349 			     k < bset_bkey_last(n2);
1350 			     k = bkey_next(k)) {
1351 				if (__set_blocks(n1, n1->keys + keys +
1352 						 bkey_u64s(k),
1353 						 block_bytes(b->c)) > blocks)
1354 					break;
1355 
1356 				last = k;
1357 				keys += bkey_u64s(k);
1358 			}
1359 		} else {
1360 			/*
1361 			 * Last node we're not getting rid of - we're getting
1362 			 * rid of the node at r[0]. Have to try and fit all of
1363 			 * the remaining keys into this node; we can't ensure
1364 			 * they will always fit due to rounding and variable
1365 			 * length keys (shouldn't be possible in practice,
1366 			 * though)
1367 			 */
1368 			if (__set_blocks(n1, n1->keys + n2->keys,
1369 					 block_bytes(b->c)) >
1370 			    btree_blocks(new_nodes[i]))
1371 				goto out_nocoalesce;
1372 
1373 			keys = n2->keys;
1374 			/* Take the key of the node we're getting rid of */
1375 			last = &r->b->key;
1376 		}
1377 
1378 		BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1379 		       btree_blocks(new_nodes[i]));
1380 
1381 		if (last)
1382 			bkey_copy_key(&new_nodes[i]->key, last);
1383 
1384 		memcpy(bset_bkey_last(n1),
1385 		       n2->start,
1386 		       (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1387 
1388 		n1->keys += keys;
1389 		r[i].keys = n1->keys;
1390 
1391 		memmove(n2->start,
1392 			bset_bkey_idx(n2, keys),
1393 			(void *) bset_bkey_last(n2) -
1394 			(void *) bset_bkey_idx(n2, keys));
1395 
1396 		n2->keys -= keys;
1397 
1398 		if (__bch_keylist_realloc(&keylist,
1399 					  bkey_u64s(&new_nodes[i]->key)))
1400 			goto out_nocoalesce;
1401 
1402 		bch_btree_node_write(new_nodes[i], &cl);
1403 		bch_keylist_add(&keylist, &new_nodes[i]->key);
1404 	}
1405 
1406 	for (i = 0; i < nodes; i++)
1407 		mutex_unlock(&new_nodes[i]->write_lock);
1408 
1409 	closure_sync(&cl);
1410 
1411 	/* We emptied out this node */
1412 	BUG_ON(btree_bset_first(new_nodes[0])->keys);
1413 	btree_node_free(new_nodes[0]);
1414 	rw_unlock(true, new_nodes[0]);
1415 	new_nodes[0] = NULL;
1416 
1417 	for (i = 0; i < nodes; i++) {
1418 		if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1419 			goto out_nocoalesce;
1420 
1421 		make_btree_freeing_key(r[i].b, keylist.top);
1422 		bch_keylist_push(&keylist);
1423 	}
1424 
1425 	bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1426 	BUG_ON(!bch_keylist_empty(&keylist));
1427 
1428 	for (i = 0; i < nodes; i++) {
1429 		btree_node_free(r[i].b);
1430 		rw_unlock(true, r[i].b);
1431 
1432 		r[i].b = new_nodes[i];
1433 	}
1434 
1435 	memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1436 	r[nodes - 1].b = ERR_PTR(-EINTR);
1437 
1438 	trace_bcache_btree_gc_coalesce(nodes);
1439 	gc->nodes--;
1440 
1441 	bch_keylist_free(&keylist);
1442 
1443 	/* Invalidated our iterator */
1444 	return -EINTR;
1445 
1446 out_nocoalesce:
1447 	closure_sync(&cl);
1448 	bch_keylist_free(&keylist);
1449 
1450 	while ((k = bch_keylist_pop(&keylist)))
1451 		if (!bkey_cmp(k, &ZERO_KEY))
1452 			atomic_dec(&b->c->prio_blocked);
1453 
1454 	for (i = 0; i < nodes; i++)
1455 		if (!IS_ERR_OR_NULL(new_nodes[i])) {
1456 			btree_node_free(new_nodes[i]);
1457 			rw_unlock(true, new_nodes[i]);
1458 		}
1459 	return 0;
1460 }
1461 
1462 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1463 				 struct btree *replace)
1464 {
1465 	struct keylist keys;
1466 	struct btree *n;
1467 
1468 	if (btree_check_reserve(b, NULL))
1469 		return 0;
1470 
1471 	n = btree_node_alloc_replacement(replace, NULL);
1472 
1473 	/* recheck reserve after allocating replacement node */
1474 	if (btree_check_reserve(b, NULL)) {
1475 		btree_node_free(n);
1476 		rw_unlock(true, n);
1477 		return 0;
1478 	}
1479 
1480 	bch_btree_node_write_sync(n);
1481 
1482 	bch_keylist_init(&keys);
1483 	bch_keylist_add(&keys, &n->key);
1484 
1485 	make_btree_freeing_key(replace, keys.top);
1486 	bch_keylist_push(&keys);
1487 
1488 	bch_btree_insert_node(b, op, &keys, NULL, NULL);
1489 	BUG_ON(!bch_keylist_empty(&keys));
1490 
1491 	btree_node_free(replace);
1492 	rw_unlock(true, n);
1493 
1494 	/* Invalidated our iterator */
1495 	return -EINTR;
1496 }
1497 
1498 static unsigned btree_gc_count_keys(struct btree *b)
1499 {
1500 	struct bkey *k;
1501 	struct btree_iter iter;
1502 	unsigned ret = 0;
1503 
1504 	for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1505 		ret += bkey_u64s(k);
1506 
1507 	return ret;
1508 }
1509 
1510 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1511 			    struct closure *writes, struct gc_stat *gc)
1512 {
1513 	int ret = 0;
1514 	bool should_rewrite;
1515 	struct bkey *k;
1516 	struct btree_iter iter;
1517 	struct gc_merge_info r[GC_MERGE_NODES];
1518 	struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1519 
1520 	bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1521 
1522 	for (i = r; i < r + ARRAY_SIZE(r); i++)
1523 		i->b = ERR_PTR(-EINTR);
1524 
1525 	while (1) {
1526 		k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1527 		if (k) {
1528 			r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1529 						  true, b);
1530 			if (IS_ERR(r->b)) {
1531 				ret = PTR_ERR(r->b);
1532 				break;
1533 			}
1534 
1535 			r->keys = btree_gc_count_keys(r->b);
1536 
1537 			ret = btree_gc_coalesce(b, op, gc, r);
1538 			if (ret)
1539 				break;
1540 		}
1541 
1542 		if (!last->b)
1543 			break;
1544 
1545 		if (!IS_ERR(last->b)) {
1546 			should_rewrite = btree_gc_mark_node(last->b, gc);
1547 			if (should_rewrite) {
1548 				ret = btree_gc_rewrite_node(b, op, last->b);
1549 				if (ret)
1550 					break;
1551 			}
1552 
1553 			if (last->b->level) {
1554 				ret = btree_gc_recurse(last->b, op, writes, gc);
1555 				if (ret)
1556 					break;
1557 			}
1558 
1559 			bkey_copy_key(&b->c->gc_done, &last->b->key);
1560 
1561 			/*
1562 			 * Must flush leaf nodes before gc ends, since replace
1563 			 * operations aren't journalled
1564 			 */
1565 			mutex_lock(&last->b->write_lock);
1566 			if (btree_node_dirty(last->b))
1567 				bch_btree_node_write(last->b, writes);
1568 			mutex_unlock(&last->b->write_lock);
1569 			rw_unlock(true, last->b);
1570 		}
1571 
1572 		memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1573 		r->b = NULL;
1574 
1575 		if (need_resched()) {
1576 			ret = -EAGAIN;
1577 			break;
1578 		}
1579 	}
1580 
1581 	for (i = r; i < r + ARRAY_SIZE(r); i++)
1582 		if (!IS_ERR_OR_NULL(i->b)) {
1583 			mutex_lock(&i->b->write_lock);
1584 			if (btree_node_dirty(i->b))
1585 				bch_btree_node_write(i->b, writes);
1586 			mutex_unlock(&i->b->write_lock);
1587 			rw_unlock(true, i->b);
1588 		}
1589 
1590 	return ret;
1591 }
1592 
1593 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1594 			     struct closure *writes, struct gc_stat *gc)
1595 {
1596 	struct btree *n = NULL;
1597 	int ret = 0;
1598 	bool should_rewrite;
1599 
1600 	should_rewrite = btree_gc_mark_node(b, gc);
1601 	if (should_rewrite) {
1602 		n = btree_node_alloc_replacement(b, NULL);
1603 
1604 		if (!IS_ERR_OR_NULL(n)) {
1605 			bch_btree_node_write_sync(n);
1606 
1607 			bch_btree_set_root(n);
1608 			btree_node_free(b);
1609 			rw_unlock(true, n);
1610 
1611 			return -EINTR;
1612 		}
1613 	}
1614 
1615 	__bch_btree_mark_key(b->c, b->level + 1, &b->key);
1616 
1617 	if (b->level) {
1618 		ret = btree_gc_recurse(b, op, writes, gc);
1619 		if (ret)
1620 			return ret;
1621 	}
1622 
1623 	bkey_copy_key(&b->c->gc_done, &b->key);
1624 
1625 	return ret;
1626 }
1627 
1628 static void btree_gc_start(struct cache_set *c)
1629 {
1630 	struct cache *ca;
1631 	struct bucket *b;
1632 	unsigned i;
1633 
1634 	if (!c->gc_mark_valid)
1635 		return;
1636 
1637 	mutex_lock(&c->bucket_lock);
1638 
1639 	c->gc_mark_valid = 0;
1640 	c->gc_done = ZERO_KEY;
1641 
1642 	for_each_cache(ca, c, i)
1643 		for_each_bucket(b, ca) {
1644 			b->last_gc = b->gen;
1645 			if (!atomic_read(&b->pin)) {
1646 				SET_GC_MARK(b, 0);
1647 				SET_GC_SECTORS_USED(b, 0);
1648 			}
1649 		}
1650 
1651 	mutex_unlock(&c->bucket_lock);
1652 }
1653 
1654 static size_t bch_btree_gc_finish(struct cache_set *c)
1655 {
1656 	size_t available = 0;
1657 	struct bucket *b;
1658 	struct cache *ca;
1659 	unsigned i;
1660 
1661 	mutex_lock(&c->bucket_lock);
1662 
1663 	set_gc_sectors(c);
1664 	c->gc_mark_valid = 1;
1665 	c->need_gc	= 0;
1666 
1667 	for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1668 		SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1669 			    GC_MARK_METADATA);
1670 
1671 	/* don't reclaim buckets to which writeback keys point */
1672 	rcu_read_lock();
1673 	for (i = 0; i < c->nr_uuids; i++) {
1674 		struct bcache_device *d = c->devices[i];
1675 		struct cached_dev *dc;
1676 		struct keybuf_key *w, *n;
1677 		unsigned j;
1678 
1679 		if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1680 			continue;
1681 		dc = container_of(d, struct cached_dev, disk);
1682 
1683 		spin_lock(&dc->writeback_keys.lock);
1684 		rbtree_postorder_for_each_entry_safe(w, n,
1685 					&dc->writeback_keys.keys, node)
1686 			for (j = 0; j < KEY_PTRS(&w->key); j++)
1687 				SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1688 					    GC_MARK_DIRTY);
1689 		spin_unlock(&dc->writeback_keys.lock);
1690 	}
1691 	rcu_read_unlock();
1692 
1693 	for_each_cache(ca, c, i) {
1694 		uint64_t *i;
1695 
1696 		ca->invalidate_needs_gc = 0;
1697 
1698 		for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1699 			SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1700 
1701 		for (i = ca->prio_buckets;
1702 		     i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1703 			SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1704 
1705 		for_each_bucket(b, ca) {
1706 			c->need_gc	= max(c->need_gc, bucket_gc_gen(b));
1707 
1708 			if (atomic_read(&b->pin))
1709 				continue;
1710 
1711 			BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1712 
1713 			if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1714 				available++;
1715 		}
1716 	}
1717 
1718 	mutex_unlock(&c->bucket_lock);
1719 	return available;
1720 }
1721 
1722 static void bch_btree_gc(struct cache_set *c)
1723 {
1724 	int ret;
1725 	unsigned long available;
1726 	struct gc_stat stats;
1727 	struct closure writes;
1728 	struct btree_op op;
1729 	uint64_t start_time = local_clock();
1730 
1731 	trace_bcache_gc_start(c);
1732 
1733 	memset(&stats, 0, sizeof(struct gc_stat));
1734 	closure_init_stack(&writes);
1735 	bch_btree_op_init(&op, SHRT_MAX);
1736 
1737 	btree_gc_start(c);
1738 
1739 	do {
1740 		ret = btree_root(gc_root, c, &op, &writes, &stats);
1741 		closure_sync(&writes);
1742 		cond_resched();
1743 
1744 		if (ret && ret != -EAGAIN)
1745 			pr_warn("gc failed!");
1746 	} while (ret);
1747 
1748 	available = bch_btree_gc_finish(c);
1749 	wake_up_allocators(c);
1750 
1751 	bch_time_stats_update(&c->btree_gc_time, start_time);
1752 
1753 	stats.key_bytes *= sizeof(uint64_t);
1754 	stats.data	<<= 9;
1755 	stats.in_use	= (c->nbuckets - available) * 100 / c->nbuckets;
1756 	memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1757 
1758 	trace_bcache_gc_end(c);
1759 
1760 	bch_moving_gc(c);
1761 }
1762 
1763 static bool gc_should_run(struct cache_set *c)
1764 {
1765 	struct cache *ca;
1766 	unsigned i;
1767 
1768 	for_each_cache(ca, c, i)
1769 		if (ca->invalidate_needs_gc)
1770 			return true;
1771 
1772 	if (atomic_read(&c->sectors_to_gc) < 0)
1773 		return true;
1774 
1775 	return false;
1776 }
1777 
1778 static int bch_gc_thread(void *arg)
1779 {
1780 	struct cache_set *c = arg;
1781 
1782 	while (1) {
1783 		wait_event_interruptible(c->gc_wait,
1784 			   kthread_should_stop() || gc_should_run(c));
1785 
1786 		if (kthread_should_stop())
1787 			break;
1788 
1789 		set_gc_sectors(c);
1790 		bch_btree_gc(c);
1791 	}
1792 
1793 	return 0;
1794 }
1795 
1796 int bch_gc_thread_start(struct cache_set *c)
1797 {
1798 	c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1799 	if (IS_ERR(c->gc_thread))
1800 		return PTR_ERR(c->gc_thread);
1801 
1802 	return 0;
1803 }
1804 
1805 /* Initial partial gc */
1806 
1807 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1808 {
1809 	int ret = 0;
1810 	struct bkey *k, *p = NULL;
1811 	struct btree_iter iter;
1812 
1813 	for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1814 		bch_initial_mark_key(b->c, b->level, k);
1815 
1816 	bch_initial_mark_key(b->c, b->level + 1, &b->key);
1817 
1818 	if (b->level) {
1819 		bch_btree_iter_init(&b->keys, &iter, NULL);
1820 
1821 		do {
1822 			k = bch_btree_iter_next_filter(&iter, &b->keys,
1823 						       bch_ptr_bad);
1824 			if (k)
1825 				btree_node_prefetch(b, k);
1826 
1827 			if (p)
1828 				ret = btree(check_recurse, p, b, op);
1829 
1830 			p = k;
1831 		} while (p && !ret);
1832 	}
1833 
1834 	return ret;
1835 }
1836 
1837 int bch_btree_check(struct cache_set *c)
1838 {
1839 	struct btree_op op;
1840 
1841 	bch_btree_op_init(&op, SHRT_MAX);
1842 
1843 	return btree_root(check_recurse, c, &op);
1844 }
1845 
1846 void bch_initial_gc_finish(struct cache_set *c)
1847 {
1848 	struct cache *ca;
1849 	struct bucket *b;
1850 	unsigned i;
1851 
1852 	bch_btree_gc_finish(c);
1853 
1854 	mutex_lock(&c->bucket_lock);
1855 
1856 	/*
1857 	 * We need to put some unused buckets directly on the prio freelist in
1858 	 * order to get the allocator thread started - it needs freed buckets in
1859 	 * order to rewrite the prios and gens, and it needs to rewrite prios
1860 	 * and gens in order to free buckets.
1861 	 *
1862 	 * This is only safe for buckets that have no live data in them, which
1863 	 * there should always be some of.
1864 	 */
1865 	for_each_cache(ca, c, i) {
1866 		for_each_bucket(b, ca) {
1867 			if (fifo_full(&ca->free[RESERVE_PRIO]))
1868 				break;
1869 
1870 			if (bch_can_invalidate_bucket(ca, b) &&
1871 			    !GC_MARK(b)) {
1872 				__bch_invalidate_one_bucket(ca, b);
1873 				fifo_push(&ca->free[RESERVE_PRIO],
1874 					  b - ca->buckets);
1875 			}
1876 		}
1877 	}
1878 
1879 	mutex_unlock(&c->bucket_lock);
1880 }
1881 
1882 /* Btree insertion */
1883 
1884 static bool btree_insert_key(struct btree *b, struct bkey *k,
1885 			     struct bkey *replace_key)
1886 {
1887 	unsigned status;
1888 
1889 	BUG_ON(bkey_cmp(k, &b->key) > 0);
1890 
1891 	status = bch_btree_insert_key(&b->keys, k, replace_key);
1892 	if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1893 		bch_check_keys(&b->keys, "%u for %s", status,
1894 			       replace_key ? "replace" : "insert");
1895 
1896 		trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1897 					      status);
1898 		return true;
1899 	} else
1900 		return false;
1901 }
1902 
1903 static size_t insert_u64s_remaining(struct btree *b)
1904 {
1905 	long ret = bch_btree_keys_u64s_remaining(&b->keys);
1906 
1907 	/*
1908 	 * Might land in the middle of an existing extent and have to split it
1909 	 */
1910 	if (b->keys.ops->is_extents)
1911 		ret -= KEY_MAX_U64S;
1912 
1913 	return max(ret, 0L);
1914 }
1915 
1916 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1917 				  struct keylist *insert_keys,
1918 				  struct bkey *replace_key)
1919 {
1920 	bool ret = false;
1921 	int oldsize = bch_count_data(&b->keys);
1922 
1923 	while (!bch_keylist_empty(insert_keys)) {
1924 		struct bkey *k = insert_keys->keys;
1925 
1926 		if (bkey_u64s(k) > insert_u64s_remaining(b))
1927 			break;
1928 
1929 		if (bkey_cmp(k, &b->key) <= 0) {
1930 			if (!b->level)
1931 				bkey_put(b->c, k);
1932 
1933 			ret |= btree_insert_key(b, k, replace_key);
1934 			bch_keylist_pop_front(insert_keys);
1935 		} else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1936 			BKEY_PADDED(key) temp;
1937 			bkey_copy(&temp.key, insert_keys->keys);
1938 
1939 			bch_cut_back(&b->key, &temp.key);
1940 			bch_cut_front(&b->key, insert_keys->keys);
1941 
1942 			ret |= btree_insert_key(b, &temp.key, replace_key);
1943 			break;
1944 		} else {
1945 			break;
1946 		}
1947 	}
1948 
1949 	if (!ret)
1950 		op->insert_collision = true;
1951 
1952 	BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1953 
1954 	BUG_ON(bch_count_data(&b->keys) < oldsize);
1955 	return ret;
1956 }
1957 
1958 static int btree_split(struct btree *b, struct btree_op *op,
1959 		       struct keylist *insert_keys,
1960 		       struct bkey *replace_key)
1961 {
1962 	bool split;
1963 	struct btree *n1, *n2 = NULL, *n3 = NULL;
1964 	uint64_t start_time = local_clock();
1965 	struct closure cl;
1966 	struct keylist parent_keys;
1967 
1968 	closure_init_stack(&cl);
1969 	bch_keylist_init(&parent_keys);
1970 
1971 	if (btree_check_reserve(b, op)) {
1972 		if (!b->level)
1973 			return -EINTR;
1974 		else
1975 			WARN(1, "insufficient reserve for split\n");
1976 	}
1977 
1978 	n1 = btree_node_alloc_replacement(b, op);
1979 	if (IS_ERR(n1))
1980 		goto err;
1981 
1982 	split = set_blocks(btree_bset_first(n1),
1983 			   block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
1984 
1985 	if (split) {
1986 		unsigned keys = 0;
1987 
1988 		trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
1989 
1990 		n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1991 		if (IS_ERR(n2))
1992 			goto err_free1;
1993 
1994 		if (!b->parent) {
1995 			n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
1996 			if (IS_ERR(n3))
1997 				goto err_free2;
1998 		}
1999 
2000 		mutex_lock(&n1->write_lock);
2001 		mutex_lock(&n2->write_lock);
2002 
2003 		bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2004 
2005 		/*
2006 		 * Has to be a linear search because we don't have an auxiliary
2007 		 * search tree yet
2008 		 */
2009 
2010 		while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2011 			keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2012 							keys));
2013 
2014 		bkey_copy_key(&n1->key,
2015 			      bset_bkey_idx(btree_bset_first(n1), keys));
2016 		keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2017 
2018 		btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2019 		btree_bset_first(n1)->keys = keys;
2020 
2021 		memcpy(btree_bset_first(n2)->start,
2022 		       bset_bkey_last(btree_bset_first(n1)),
2023 		       btree_bset_first(n2)->keys * sizeof(uint64_t));
2024 
2025 		bkey_copy_key(&n2->key, &b->key);
2026 
2027 		bch_keylist_add(&parent_keys, &n2->key);
2028 		bch_btree_node_write(n2, &cl);
2029 		mutex_unlock(&n2->write_lock);
2030 		rw_unlock(true, n2);
2031 	} else {
2032 		trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2033 
2034 		mutex_lock(&n1->write_lock);
2035 		bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2036 	}
2037 
2038 	bch_keylist_add(&parent_keys, &n1->key);
2039 	bch_btree_node_write(n1, &cl);
2040 	mutex_unlock(&n1->write_lock);
2041 
2042 	if (n3) {
2043 		/* Depth increases, make a new root */
2044 		mutex_lock(&n3->write_lock);
2045 		bkey_copy_key(&n3->key, &MAX_KEY);
2046 		bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2047 		bch_btree_node_write(n3, &cl);
2048 		mutex_unlock(&n3->write_lock);
2049 
2050 		closure_sync(&cl);
2051 		bch_btree_set_root(n3);
2052 		rw_unlock(true, n3);
2053 	} else if (!b->parent) {
2054 		/* Root filled up but didn't need to be split */
2055 		closure_sync(&cl);
2056 		bch_btree_set_root(n1);
2057 	} else {
2058 		/* Split a non root node */
2059 		closure_sync(&cl);
2060 		make_btree_freeing_key(b, parent_keys.top);
2061 		bch_keylist_push(&parent_keys);
2062 
2063 		bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2064 		BUG_ON(!bch_keylist_empty(&parent_keys));
2065 	}
2066 
2067 	btree_node_free(b);
2068 	rw_unlock(true, n1);
2069 
2070 	bch_time_stats_update(&b->c->btree_split_time, start_time);
2071 
2072 	return 0;
2073 err_free2:
2074 	bkey_put(b->c, &n2->key);
2075 	btree_node_free(n2);
2076 	rw_unlock(true, n2);
2077 err_free1:
2078 	bkey_put(b->c, &n1->key);
2079 	btree_node_free(n1);
2080 	rw_unlock(true, n1);
2081 err:
2082 	WARN(1, "bcache: btree split failed (level %u)", b->level);
2083 
2084 	if (n3 == ERR_PTR(-EAGAIN) ||
2085 	    n2 == ERR_PTR(-EAGAIN) ||
2086 	    n1 == ERR_PTR(-EAGAIN))
2087 		return -EAGAIN;
2088 
2089 	return -ENOMEM;
2090 }
2091 
2092 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2093 				 struct keylist *insert_keys,
2094 				 atomic_t *journal_ref,
2095 				 struct bkey *replace_key)
2096 {
2097 	struct closure cl;
2098 
2099 	BUG_ON(b->level && replace_key);
2100 
2101 	closure_init_stack(&cl);
2102 
2103 	mutex_lock(&b->write_lock);
2104 
2105 	if (write_block(b) != btree_bset_last(b) &&
2106 	    b->keys.last_set_unwritten)
2107 		bch_btree_init_next(b); /* just wrote a set */
2108 
2109 	if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2110 		mutex_unlock(&b->write_lock);
2111 		goto split;
2112 	}
2113 
2114 	BUG_ON(write_block(b) != btree_bset_last(b));
2115 
2116 	if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2117 		if (!b->level)
2118 			bch_btree_leaf_dirty(b, journal_ref);
2119 		else
2120 			bch_btree_node_write(b, &cl);
2121 	}
2122 
2123 	mutex_unlock(&b->write_lock);
2124 
2125 	/* wait for btree node write if necessary, after unlock */
2126 	closure_sync(&cl);
2127 
2128 	return 0;
2129 split:
2130 	if (current->bio_list) {
2131 		op->lock = b->c->root->level + 1;
2132 		return -EAGAIN;
2133 	} else if (op->lock <= b->c->root->level) {
2134 		op->lock = b->c->root->level + 1;
2135 		return -EINTR;
2136 	} else {
2137 		/* Invalidated all iterators */
2138 		int ret = btree_split(b, op, insert_keys, replace_key);
2139 
2140 		if (bch_keylist_empty(insert_keys))
2141 			return 0;
2142 		else if (!ret)
2143 			return -EINTR;
2144 		return ret;
2145 	}
2146 }
2147 
2148 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2149 			       struct bkey *check_key)
2150 {
2151 	int ret = -EINTR;
2152 	uint64_t btree_ptr = b->key.ptr[0];
2153 	unsigned long seq = b->seq;
2154 	struct keylist insert;
2155 	bool upgrade = op->lock == -1;
2156 
2157 	bch_keylist_init(&insert);
2158 
2159 	if (upgrade) {
2160 		rw_unlock(false, b);
2161 		rw_lock(true, b, b->level);
2162 
2163 		if (b->key.ptr[0] != btree_ptr ||
2164                    b->seq != seq + 1) {
2165                        op->lock = b->level;
2166 			goto out;
2167                }
2168 	}
2169 
2170 	SET_KEY_PTRS(check_key, 1);
2171 	get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2172 
2173 	SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2174 
2175 	bch_keylist_add(&insert, check_key);
2176 
2177 	ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2178 
2179 	BUG_ON(!ret && !bch_keylist_empty(&insert));
2180 out:
2181 	if (upgrade)
2182 		downgrade_write(&b->lock);
2183 	return ret;
2184 }
2185 
2186 struct btree_insert_op {
2187 	struct btree_op	op;
2188 	struct keylist	*keys;
2189 	atomic_t	*journal_ref;
2190 	struct bkey	*replace_key;
2191 };
2192 
2193 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2194 {
2195 	struct btree_insert_op *op = container_of(b_op,
2196 					struct btree_insert_op, op);
2197 
2198 	int ret = bch_btree_insert_node(b, &op->op, op->keys,
2199 					op->journal_ref, op->replace_key);
2200 	if (ret && !bch_keylist_empty(op->keys))
2201 		return ret;
2202 	else
2203 		return MAP_DONE;
2204 }
2205 
2206 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2207 		     atomic_t *journal_ref, struct bkey *replace_key)
2208 {
2209 	struct btree_insert_op op;
2210 	int ret = 0;
2211 
2212 	BUG_ON(current->bio_list);
2213 	BUG_ON(bch_keylist_empty(keys));
2214 
2215 	bch_btree_op_init(&op.op, 0);
2216 	op.keys		= keys;
2217 	op.journal_ref	= journal_ref;
2218 	op.replace_key	= replace_key;
2219 
2220 	while (!ret && !bch_keylist_empty(keys)) {
2221 		op.op.lock = 0;
2222 		ret = bch_btree_map_leaf_nodes(&op.op, c,
2223 					       &START_KEY(keys->keys),
2224 					       btree_insert_fn);
2225 	}
2226 
2227 	if (ret) {
2228 		struct bkey *k;
2229 
2230 		pr_err("error %i", ret);
2231 
2232 		while ((k = bch_keylist_pop(keys)))
2233 			bkey_put(c, k);
2234 	} else if (op.op.insert_collision)
2235 		ret = -ESRCH;
2236 
2237 	return ret;
2238 }
2239 
2240 void bch_btree_set_root(struct btree *b)
2241 {
2242 	unsigned i;
2243 	struct closure cl;
2244 
2245 	closure_init_stack(&cl);
2246 
2247 	trace_bcache_btree_set_root(b);
2248 
2249 	BUG_ON(!b->written);
2250 
2251 	for (i = 0; i < KEY_PTRS(&b->key); i++)
2252 		BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2253 
2254 	mutex_lock(&b->c->bucket_lock);
2255 	list_del_init(&b->list);
2256 	mutex_unlock(&b->c->bucket_lock);
2257 
2258 	b->c->root = b;
2259 
2260 	bch_journal_meta(b->c, &cl);
2261 	closure_sync(&cl);
2262 }
2263 
2264 /* Map across nodes or keys */
2265 
2266 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2267 				       struct bkey *from,
2268 				       btree_map_nodes_fn *fn, int flags)
2269 {
2270 	int ret = MAP_CONTINUE;
2271 
2272 	if (b->level) {
2273 		struct bkey *k;
2274 		struct btree_iter iter;
2275 
2276 		bch_btree_iter_init(&b->keys, &iter, from);
2277 
2278 		while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2279 						       bch_ptr_bad))) {
2280 			ret = btree(map_nodes_recurse, k, b,
2281 				    op, from, fn, flags);
2282 			from = NULL;
2283 
2284 			if (ret != MAP_CONTINUE)
2285 				return ret;
2286 		}
2287 	}
2288 
2289 	if (!b->level || flags == MAP_ALL_NODES)
2290 		ret = fn(op, b);
2291 
2292 	return ret;
2293 }
2294 
2295 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2296 			  struct bkey *from, btree_map_nodes_fn *fn, int flags)
2297 {
2298 	return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2299 }
2300 
2301 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2302 				      struct bkey *from, btree_map_keys_fn *fn,
2303 				      int flags)
2304 {
2305 	int ret = MAP_CONTINUE;
2306 	struct bkey *k;
2307 	struct btree_iter iter;
2308 
2309 	bch_btree_iter_init(&b->keys, &iter, from);
2310 
2311 	while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2312 		ret = !b->level
2313 			? fn(op, b, k)
2314 			: btree(map_keys_recurse, k, b, op, from, fn, flags);
2315 		from = NULL;
2316 
2317 		if (ret != MAP_CONTINUE)
2318 			return ret;
2319 	}
2320 
2321 	if (!b->level && (flags & MAP_END_KEY))
2322 		ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2323 				     KEY_OFFSET(&b->key), 0));
2324 
2325 	return ret;
2326 }
2327 
2328 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2329 		       struct bkey *from, btree_map_keys_fn *fn, int flags)
2330 {
2331 	return btree_root(map_keys_recurse, c, op, from, fn, flags);
2332 }
2333 
2334 /* Keybuf code */
2335 
2336 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2337 {
2338 	/* Overlapping keys compare equal */
2339 	if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2340 		return -1;
2341 	if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2342 		return 1;
2343 	return 0;
2344 }
2345 
2346 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2347 					    struct keybuf_key *r)
2348 {
2349 	return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2350 }
2351 
2352 struct refill {
2353 	struct btree_op	op;
2354 	unsigned	nr_found;
2355 	struct keybuf	*buf;
2356 	struct bkey	*end;
2357 	keybuf_pred_fn	*pred;
2358 };
2359 
2360 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2361 			    struct bkey *k)
2362 {
2363 	struct refill *refill = container_of(op, struct refill, op);
2364 	struct keybuf *buf = refill->buf;
2365 	int ret = MAP_CONTINUE;
2366 
2367 	if (bkey_cmp(k, refill->end) >= 0) {
2368 		ret = MAP_DONE;
2369 		goto out;
2370 	}
2371 
2372 	if (!KEY_SIZE(k)) /* end key */
2373 		goto out;
2374 
2375 	if (refill->pred(buf, k)) {
2376 		struct keybuf_key *w;
2377 
2378 		spin_lock(&buf->lock);
2379 
2380 		w = array_alloc(&buf->freelist);
2381 		if (!w) {
2382 			spin_unlock(&buf->lock);
2383 			return MAP_DONE;
2384 		}
2385 
2386 		w->private = NULL;
2387 		bkey_copy(&w->key, k);
2388 
2389 		if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2390 			array_free(&buf->freelist, w);
2391 		else
2392 			refill->nr_found++;
2393 
2394 		if (array_freelist_empty(&buf->freelist))
2395 			ret = MAP_DONE;
2396 
2397 		spin_unlock(&buf->lock);
2398 	}
2399 out:
2400 	buf->last_scanned = *k;
2401 	return ret;
2402 }
2403 
2404 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2405 		       struct bkey *end, keybuf_pred_fn *pred)
2406 {
2407 	struct bkey start = buf->last_scanned;
2408 	struct refill refill;
2409 
2410 	cond_resched();
2411 
2412 	bch_btree_op_init(&refill.op, -1);
2413 	refill.nr_found	= 0;
2414 	refill.buf	= buf;
2415 	refill.end	= end;
2416 	refill.pred	= pred;
2417 
2418 	bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2419 			   refill_keybuf_fn, MAP_END_KEY);
2420 
2421 	trace_bcache_keyscan(refill.nr_found,
2422 			     KEY_INODE(&start), KEY_OFFSET(&start),
2423 			     KEY_INODE(&buf->last_scanned),
2424 			     KEY_OFFSET(&buf->last_scanned));
2425 
2426 	spin_lock(&buf->lock);
2427 
2428 	if (!RB_EMPTY_ROOT(&buf->keys)) {
2429 		struct keybuf_key *w;
2430 		w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2431 		buf->start	= START_KEY(&w->key);
2432 
2433 		w = RB_LAST(&buf->keys, struct keybuf_key, node);
2434 		buf->end	= w->key;
2435 	} else {
2436 		buf->start	= MAX_KEY;
2437 		buf->end	= MAX_KEY;
2438 	}
2439 
2440 	spin_unlock(&buf->lock);
2441 }
2442 
2443 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2444 {
2445 	rb_erase(&w->node, &buf->keys);
2446 	array_free(&buf->freelist, w);
2447 }
2448 
2449 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2450 {
2451 	spin_lock(&buf->lock);
2452 	__bch_keybuf_del(buf, w);
2453 	spin_unlock(&buf->lock);
2454 }
2455 
2456 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2457 				  struct bkey *end)
2458 {
2459 	bool ret = false;
2460 	struct keybuf_key *p, *w, s;
2461 	s.key = *start;
2462 
2463 	if (bkey_cmp(end, &buf->start) <= 0 ||
2464 	    bkey_cmp(start, &buf->end) >= 0)
2465 		return false;
2466 
2467 	spin_lock(&buf->lock);
2468 	w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2469 
2470 	while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2471 		p = w;
2472 		w = RB_NEXT(w, node);
2473 
2474 		if (p->private)
2475 			ret = true;
2476 		else
2477 			__bch_keybuf_del(buf, p);
2478 	}
2479 
2480 	spin_unlock(&buf->lock);
2481 	return ret;
2482 }
2483 
2484 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2485 {
2486 	struct keybuf_key *w;
2487 	spin_lock(&buf->lock);
2488 
2489 	w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2490 
2491 	while (w && w->private)
2492 		w = RB_NEXT(w, node);
2493 
2494 	if (w)
2495 		w->private = ERR_PTR(-EINTR);
2496 
2497 	spin_unlock(&buf->lock);
2498 	return w;
2499 }
2500 
2501 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2502 					  struct keybuf *buf,
2503 					  struct bkey *end,
2504 					  keybuf_pred_fn *pred)
2505 {
2506 	struct keybuf_key *ret;
2507 
2508 	while (1) {
2509 		ret = bch_keybuf_next(buf);
2510 		if (ret)
2511 			break;
2512 
2513 		if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2514 			pr_debug("scan finished");
2515 			break;
2516 		}
2517 
2518 		bch_refill_keybuf(c, buf, end, pred);
2519 	}
2520 
2521 	return ret;
2522 }
2523 
2524 void bch_keybuf_init(struct keybuf *buf)
2525 {
2526 	buf->last_scanned	= MAX_KEY;
2527 	buf->keys		= RB_ROOT;
2528 
2529 	spin_lock_init(&buf->lock);
2530 	array_allocator_init(&buf->freelist);
2531 }
2532