Merge commit 'ed30f24e8d07d30aa3e69d1f508f4d7bd2e8ea14' of git://git.linaro.org/landi...
[firefly-linux-kernel-4.4.55.git] / drivers / md / bcache / btree.c
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 "request.h"
27
28 #include <linux/slab.h>
29 #include <linux/bitops.h>
30 #include <linux/hash.h>
31 #include <linux/prefetch.h>
32 #include <linux/random.h>
33 #include <linux/rcupdate.h>
34 #include <trace/events/bcache.h>
35
36 /*
37  * Todo:
38  * register_bcache: Return errors out to userspace correctly
39  *
40  * Writeback: don't undirty key until after a cache flush
41  *
42  * Create an iterator for key pointers
43  *
44  * On btree write error, mark bucket such that it won't be freed from the cache
45  *
46  * Journalling:
47  *   Check for bad keys in replay
48  *   Propagate barriers
49  *   Refcount journal entries in journal_replay
50  *
51  * Garbage collection:
52  *   Finish incremental gc
53  *   Gc should free old UUIDs, data for invalid UUIDs
54  *
55  * Provide a way to list backing device UUIDs we have data cached for, and
56  * probably how long it's been since we've seen them, and a way to invalidate
57  * dirty data for devices that will never be attached again
58  *
59  * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
60  * that based on that and how much dirty data we have we can keep writeback
61  * from being starved
62  *
63  * Add a tracepoint or somesuch to watch for writeback starvation
64  *
65  * When btree depth > 1 and splitting an interior node, we have to make sure
66  * alloc_bucket() cannot fail. This should be true but is not completely
67  * obvious.
68  *
69  * Make sure all allocations get charged to the root cgroup
70  *
71  * Plugging?
72  *
73  * If data write is less than hard sector size of ssd, round up offset in open
74  * bucket to the next whole sector
75  *
76  * Also lookup by cgroup in get_open_bucket()
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 static const char * const op_types[] = {
91         "insert", "replace"
92 };
93
94 static const char *op_type(struct btree_op *op)
95 {
96         return op_types[op->type];
97 }
98
99 #define MAX_NEED_GC             64
100 #define MAX_SAVE_PRIO           72
101
102 #define PTR_DIRTY_BIT           (((uint64_t) 1 << 36))
103
104 #define PTR_HASH(c, k)                                                  \
105         (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
106
107 struct workqueue_struct *bch_gc_wq;
108 static struct workqueue_struct *btree_io_wq;
109
110 void bch_btree_op_init_stack(struct btree_op *op)
111 {
112         memset(op, 0, sizeof(struct btree_op));
113         closure_init_stack(&op->cl);
114         op->lock = -1;
115         bch_keylist_init(&op->keys);
116 }
117
118 /* Btree key manipulation */
119
120 static void bkey_put(struct cache_set *c, struct bkey *k, int level)
121 {
122         if ((level && KEY_OFFSET(k)) || !level)
123                 __bkey_put(c, k);
124 }
125
126 /* Btree IO */
127
128 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
129 {
130         uint64_t crc = b->key.ptr[0];
131         void *data = (void *) i + 8, *end = end(i);
132
133         crc = bch_crc64_update(crc, data, end - data);
134         return crc ^ 0xffffffffffffffffULL;
135 }
136
137 static void btree_bio_endio(struct bio *bio, int error)
138 {
139         struct closure *cl = bio->bi_private;
140         struct btree *b = container_of(cl, struct btree, io.cl);
141
142         if (error)
143                 set_btree_node_io_error(b);
144
145         bch_bbio_count_io_errors(b->c, bio, error, (bio->bi_rw & WRITE)
146                                  ? "writing btree" : "reading btree");
147         closure_put(cl);
148 }
149
150 static void btree_bio_init(struct btree *b)
151 {
152         BUG_ON(b->bio);
153         b->bio = bch_bbio_alloc(b->c);
154
155         b->bio->bi_end_io       = btree_bio_endio;
156         b->bio->bi_private      = &b->io.cl;
157 }
158
159 void bch_btree_read_done(struct closure *cl)
160 {
161         struct btree *b = container_of(cl, struct btree, io.cl);
162         struct bset *i = b->sets[0].data;
163         struct btree_iter *iter = b->c->fill_iter;
164         const char *err = "bad btree header";
165         BUG_ON(b->nsets || b->written);
166
167         bch_bbio_free(b->bio, b->c);
168         b->bio = NULL;
169
170         mutex_lock(&b->c->fill_lock);
171         iter->used = 0;
172
173         if (btree_node_io_error(b) ||
174             !i->seq)
175                 goto err;
176
177         for (;
178              b->written < btree_blocks(b) && i->seq == b->sets[0].data->seq;
179              i = write_block(b)) {
180                 err = "unsupported bset version";
181                 if (i->version > BCACHE_BSET_VERSION)
182                         goto err;
183
184                 err = "bad btree header";
185                 if (b->written + set_blocks(i, b->c) > btree_blocks(b))
186                         goto err;
187
188                 err = "bad magic";
189                 if (i->magic != bset_magic(b->c))
190                         goto err;
191
192                 err = "bad checksum";
193                 switch (i->version) {
194                 case 0:
195                         if (i->csum != csum_set(i))
196                                 goto err;
197                         break;
198                 case BCACHE_BSET_VERSION:
199                         if (i->csum != btree_csum_set(b, i))
200                                 goto err;
201                         break;
202                 }
203
204                 err = "empty set";
205                 if (i != b->sets[0].data && !i->keys)
206                         goto err;
207
208                 bch_btree_iter_push(iter, i->start, end(i));
209
210                 b->written += set_blocks(i, b->c);
211         }
212
213         err = "corrupted btree";
214         for (i = write_block(b);
215              index(i, b) < btree_blocks(b);
216              i = ((void *) i) + block_bytes(b->c))
217                 if (i->seq == b->sets[0].data->seq)
218                         goto err;
219
220         bch_btree_sort_and_fix_extents(b, iter);
221
222         i = b->sets[0].data;
223         err = "short btree key";
224         if (b->sets[0].size &&
225             bkey_cmp(&b->key, &b->sets[0].end) < 0)
226                 goto err;
227
228         if (b->written < btree_blocks(b))
229                 bch_bset_init_next(b);
230 out:
231
232         mutex_unlock(&b->c->fill_lock);
233
234         spin_lock(&b->c->btree_read_time_lock);
235         bch_time_stats_update(&b->c->btree_read_time, b->io_start_time);
236         spin_unlock(&b->c->btree_read_time_lock);
237
238         smp_wmb(); /* read_done is our write lock */
239         set_btree_node_read_done(b);
240
241         closure_return(cl);
242 err:
243         set_btree_node_io_error(b);
244         bch_cache_set_error(b->c, "%s at bucket %zu, block %zu, %u keys",
245                             err, PTR_BUCKET_NR(b->c, &b->key, 0),
246                             index(i, b), i->keys);
247         goto out;
248 }
249
250 void bch_btree_read(struct btree *b)
251 {
252         BUG_ON(b->nsets || b->written);
253
254         if (!closure_trylock(&b->io.cl, &b->c->cl))
255                 BUG();
256
257         b->io_start_time = local_clock();
258
259         btree_bio_init(b);
260         b->bio->bi_rw   = REQ_META|READ_SYNC;
261         b->bio->bi_size = KEY_SIZE(&b->key) << 9;
262
263         bch_bio_map(b->bio, b->sets[0].data);
264
265         pr_debug("%s", pbtree(b));
266         trace_bcache_btree_read(b->bio);
267         bch_submit_bbio(b->bio, b->c, &b->key, 0);
268
269         continue_at(&b->io.cl, bch_btree_read_done, system_wq);
270 }
271
272 static void btree_complete_write(struct btree *b, struct btree_write *w)
273 {
274         if (w->prio_blocked &&
275             !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
276                 wake_up(&b->c->alloc_wait);
277
278         if (w->journal) {
279                 atomic_dec_bug(w->journal);
280                 __closure_wake_up(&b->c->journal.wait);
281         }
282
283         if (w->owner)
284                 closure_put(w->owner);
285
286         w->prio_blocked = 0;
287         w->journal      = NULL;
288         w->owner        = NULL;
289 }
290
291 static void __btree_write_done(struct closure *cl)
292 {
293         struct btree *b = container_of(cl, struct btree, io.cl);
294         struct btree_write *w = btree_prev_write(b);
295
296         bch_bbio_free(b->bio, b->c);
297         b->bio = NULL;
298         btree_complete_write(b, w);
299
300         if (btree_node_dirty(b))
301                 queue_delayed_work(btree_io_wq, &b->work,
302                                    msecs_to_jiffies(30000));
303
304         closure_return(cl);
305 }
306
307 static void btree_write_done(struct closure *cl)
308 {
309         struct btree *b = container_of(cl, struct btree, io.cl);
310         struct bio_vec *bv;
311         int n;
312
313         __bio_for_each_segment(bv, b->bio, n, 0)
314                 __free_page(bv->bv_page);
315
316         __btree_write_done(cl);
317 }
318
319 static void do_btree_write(struct btree *b)
320 {
321         struct closure *cl = &b->io.cl;
322         struct bset *i = b->sets[b->nsets].data;
323         BKEY_PADDED(key) k;
324
325         i->version      = BCACHE_BSET_VERSION;
326         i->csum         = btree_csum_set(b, i);
327
328         btree_bio_init(b);
329         b->bio->bi_rw   = REQ_META|WRITE_SYNC;
330         b->bio->bi_size = set_blocks(i, b->c) * block_bytes(b->c);
331         bch_bio_map(b->bio, i);
332
333         bkey_copy(&k.key, &b->key);
334         SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i));
335
336         if (!bch_bio_alloc_pages(b->bio, GFP_NOIO)) {
337                 int j;
338                 struct bio_vec *bv;
339                 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
340
341                 bio_for_each_segment(bv, b->bio, j)
342                         memcpy(page_address(bv->bv_page),
343                                base + j * PAGE_SIZE, PAGE_SIZE);
344
345                 trace_bcache_btree_write(b->bio);
346                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
347
348                 continue_at(cl, btree_write_done, NULL);
349         } else {
350                 b->bio->bi_vcnt = 0;
351                 bch_bio_map(b->bio, i);
352
353                 trace_bcache_btree_write(b->bio);
354                 bch_submit_bbio(b->bio, b->c, &k.key, 0);
355
356                 closure_sync(cl);
357                 __btree_write_done(cl);
358         }
359 }
360
361 static void __btree_write(struct btree *b)
362 {
363         struct bset *i = b->sets[b->nsets].data;
364
365         BUG_ON(current->bio_list);
366
367         closure_lock(&b->io, &b->c->cl);
368         cancel_delayed_work(&b->work);
369
370         clear_bit(BTREE_NODE_dirty,      &b->flags);
371         change_bit(BTREE_NODE_write_idx, &b->flags);
372
373         bch_check_key_order(b, i);
374         BUG_ON(b->written && !i->keys);
375
376         do_btree_write(b);
377
378         pr_debug("%s block %i keys %i", pbtree(b), b->written, i->keys);
379
380         b->written += set_blocks(i, b->c);
381         atomic_long_add(set_blocks(i, b->c) * b->c->sb.block_size,
382                         &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
383
384         bch_btree_sort_lazy(b);
385
386         if (b->written < btree_blocks(b))
387                 bch_bset_init_next(b);
388 }
389
390 static void btree_write_work(struct work_struct *w)
391 {
392         struct btree *b = container_of(to_delayed_work(w), struct btree, work);
393
394         down_write(&b->lock);
395
396         if (btree_node_dirty(b))
397                 __btree_write(b);
398         up_write(&b->lock);
399 }
400
401 void bch_btree_write(struct btree *b, bool now, struct btree_op *op)
402 {
403         struct bset *i = b->sets[b->nsets].data;
404         struct btree_write *w = btree_current_write(b);
405
406         BUG_ON(b->written &&
407                (b->written >= btree_blocks(b) ||
408                 i->seq != b->sets[0].data->seq ||
409                 !i->keys));
410
411         if (!btree_node_dirty(b)) {
412                 set_btree_node_dirty(b);
413                 queue_delayed_work(btree_io_wq, &b->work,
414                                    msecs_to_jiffies(30000));
415         }
416
417         w->prio_blocked += b->prio_blocked;
418         b->prio_blocked = 0;
419
420         if (op && op->journal && !b->level) {
421                 if (w->journal &&
422                     journal_pin_cmp(b->c, w, op)) {
423                         atomic_dec_bug(w->journal);
424                         w->journal = NULL;
425                 }
426
427                 if (!w->journal) {
428                         w->journal = op->journal;
429                         atomic_inc(w->journal);
430                 }
431         }
432
433         if (current->bio_list)
434                 return;
435
436         /* Force write if set is too big */
437         if (now ||
438             b->level ||
439             set_bytes(i) > PAGE_SIZE - 48) {
440                 if (op && now) {
441                         /* Must wait on multiple writes */
442                         BUG_ON(w->owner);
443                         w->owner = &op->cl;
444                         closure_get(&op->cl);
445                 }
446
447                 __btree_write(b);
448         }
449         BUG_ON(!b->written);
450 }
451
452 /*
453  * Btree in memory cache - allocation/freeing
454  * mca -> memory cache
455  */
456
457 static void mca_reinit(struct btree *b)
458 {
459         unsigned i;
460
461         b->flags        = 0;
462         b->written      = 0;
463         b->nsets        = 0;
464
465         for (i = 0; i < MAX_BSETS; i++)
466                 b->sets[i].size = 0;
467         /*
468          * Second loop starts at 1 because b->sets[0]->data is the memory we
469          * allocated
470          */
471         for (i = 1; i < MAX_BSETS; i++)
472                 b->sets[i].data = NULL;
473 }
474
475 #define mca_reserve(c)  (((c->root && c->root->level)           \
476                           ? c->root->level : 1) * 8 + 16)
477 #define mca_can_free(c)                                         \
478         max_t(int, 0, c->bucket_cache_used - mca_reserve(c))
479
480 static void mca_data_free(struct btree *b)
481 {
482         struct bset_tree *t = b->sets;
483         BUG_ON(!closure_is_unlocked(&b->io.cl));
484
485         if (bset_prev_bytes(b) < PAGE_SIZE)
486                 kfree(t->prev);
487         else
488                 free_pages((unsigned long) t->prev,
489                            get_order(bset_prev_bytes(b)));
490
491         if (bset_tree_bytes(b) < PAGE_SIZE)
492                 kfree(t->tree);
493         else
494                 free_pages((unsigned long) t->tree,
495                            get_order(bset_tree_bytes(b)));
496
497         free_pages((unsigned long) t->data, b->page_order);
498
499         t->prev = NULL;
500         t->tree = NULL;
501         t->data = NULL;
502         list_move(&b->list, &b->c->btree_cache_freed);
503         b->c->bucket_cache_used--;
504 }
505
506 static void mca_bucket_free(struct btree *b)
507 {
508         BUG_ON(btree_node_dirty(b));
509
510         b->key.ptr[0] = 0;
511         hlist_del_init_rcu(&b->hash);
512         list_move(&b->list, &b->c->btree_cache_freeable);
513 }
514
515 static unsigned btree_order(struct bkey *k)
516 {
517         return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
518 }
519
520 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
521 {
522         struct bset_tree *t = b->sets;
523         BUG_ON(t->data);
524
525         b->page_order = max_t(unsigned,
526                               ilog2(b->c->btree_pages),
527                               btree_order(k));
528
529         t->data = (void *) __get_free_pages(gfp, b->page_order);
530         if (!t->data)
531                 goto err;
532
533         t->tree = bset_tree_bytes(b) < PAGE_SIZE
534                 ? kmalloc(bset_tree_bytes(b), gfp)
535                 : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b)));
536         if (!t->tree)
537                 goto err;
538
539         t->prev = bset_prev_bytes(b) < PAGE_SIZE
540                 ? kmalloc(bset_prev_bytes(b), gfp)
541                 : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b)));
542         if (!t->prev)
543                 goto err;
544
545         list_move(&b->list, &b->c->btree_cache);
546         b->c->bucket_cache_used++;
547         return;
548 err:
549         mca_data_free(b);
550 }
551
552 static struct btree *mca_bucket_alloc(struct cache_set *c,
553                                       struct bkey *k, gfp_t gfp)
554 {
555         struct btree *b = kzalloc(sizeof(struct btree), gfp);
556         if (!b)
557                 return NULL;
558
559         init_rwsem(&b->lock);
560         lockdep_set_novalidate_class(&b->lock);
561         INIT_LIST_HEAD(&b->list);
562         INIT_DELAYED_WORK(&b->work, btree_write_work);
563         b->c = c;
564         closure_init_unlocked(&b->io);
565
566         mca_data_alloc(b, k, gfp);
567         return b;
568 }
569
570 static int mca_reap(struct btree *b, struct closure *cl, unsigned min_order)
571 {
572         lockdep_assert_held(&b->c->bucket_lock);
573
574         if (!down_write_trylock(&b->lock))
575                 return -ENOMEM;
576
577         if (b->page_order < min_order) {
578                 rw_unlock(true, b);
579                 return -ENOMEM;
580         }
581
582         BUG_ON(btree_node_dirty(b) && !b->sets[0].data);
583
584         if (cl && btree_node_dirty(b))
585                 bch_btree_write(b, true, NULL);
586
587         if (cl)
588                 closure_wait_event_async(&b->io.wait, cl,
589                          atomic_read(&b->io.cl.remaining) == -1);
590
591         if (btree_node_dirty(b) ||
592             !closure_is_unlocked(&b->io.cl) ||
593             work_pending(&b->work.work)) {
594                 rw_unlock(true, b);
595                 return -EAGAIN;
596         }
597
598         return 0;
599 }
600
601 static int bch_mca_shrink(struct shrinker *shrink, struct shrink_control *sc)
602 {
603         struct cache_set *c = container_of(shrink, struct cache_set, shrink);
604         struct btree *b, *t;
605         unsigned long i, nr = sc->nr_to_scan;
606
607         if (c->shrinker_disabled)
608                 return 0;
609
610         if (c->try_harder)
611                 return 0;
612
613         /*
614          * If nr == 0, we're supposed to return the number of items we have
615          * cached. Not allowed to return -1.
616          */
617         if (!nr)
618                 return mca_can_free(c) * c->btree_pages;
619
620         /* Return -1 if we can't do anything right now */
621         if (sc->gfp_mask & __GFP_WAIT)
622                 mutex_lock(&c->bucket_lock);
623         else if (!mutex_trylock(&c->bucket_lock))
624                 return -1;
625
626         nr /= c->btree_pages;
627         nr = min_t(unsigned long, nr, mca_can_free(c));
628
629         i = 0;
630         list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
631                 if (!nr)
632                         break;
633
634                 if (++i > 3 &&
635                     !mca_reap(b, NULL, 0)) {
636                         mca_data_free(b);
637                         rw_unlock(true, b);
638                         --nr;
639                 }
640         }
641
642         /*
643          * Can happen right when we first start up, before we've read in any
644          * btree nodes
645          */
646         if (list_empty(&c->btree_cache))
647                 goto out;
648
649         for (i = 0; nr && i < c->bucket_cache_used; i++) {
650                 b = list_first_entry(&c->btree_cache, struct btree, list);
651                 list_rotate_left(&c->btree_cache);
652
653                 if (!b->accessed &&
654                     !mca_reap(b, NULL, 0)) {
655                         mca_bucket_free(b);
656                         mca_data_free(b);
657                         rw_unlock(true, b);
658                         --nr;
659                 } else
660                         b->accessed = 0;
661         }
662 out:
663         nr = mca_can_free(c) * c->btree_pages;
664         mutex_unlock(&c->bucket_lock);
665         return nr;
666 }
667
668 void bch_btree_cache_free(struct cache_set *c)
669 {
670         struct btree *b;
671         struct closure cl;
672         closure_init_stack(&cl);
673
674         if (c->shrink.list.next)
675                 unregister_shrinker(&c->shrink);
676
677         mutex_lock(&c->bucket_lock);
678
679 #ifdef CONFIG_BCACHE_DEBUG
680         if (c->verify_data)
681                 list_move(&c->verify_data->list, &c->btree_cache);
682 #endif
683
684         list_splice(&c->btree_cache_freeable,
685                     &c->btree_cache);
686
687         while (!list_empty(&c->btree_cache)) {
688                 b = list_first_entry(&c->btree_cache, struct btree, list);
689
690                 if (btree_node_dirty(b))
691                         btree_complete_write(b, btree_current_write(b));
692                 clear_bit(BTREE_NODE_dirty, &b->flags);
693
694                 mca_data_free(b);
695         }
696
697         while (!list_empty(&c->btree_cache_freed)) {
698                 b = list_first_entry(&c->btree_cache_freed,
699                                      struct btree, list);
700                 list_del(&b->list);
701                 cancel_delayed_work_sync(&b->work);
702                 kfree(b);
703         }
704
705         mutex_unlock(&c->bucket_lock);
706 }
707
708 int bch_btree_cache_alloc(struct cache_set *c)
709 {
710         unsigned i;
711
712         /* XXX: doesn't check for errors */
713
714         closure_init_unlocked(&c->gc);
715
716         for (i = 0; i < mca_reserve(c); i++)
717                 mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
718
719         list_splice_init(&c->btree_cache,
720                          &c->btree_cache_freeable);
721
722 #ifdef CONFIG_BCACHE_DEBUG
723         mutex_init(&c->verify_lock);
724
725         c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
726
727         if (c->verify_data &&
728             c->verify_data->sets[0].data)
729                 list_del_init(&c->verify_data->list);
730         else
731                 c->verify_data = NULL;
732 #endif
733
734         c->shrink.shrink = bch_mca_shrink;
735         c->shrink.seeks = 4;
736         c->shrink.batch = c->btree_pages * 2;
737         register_shrinker(&c->shrink);
738
739         return 0;
740 }
741
742 /* Btree in memory cache - hash table */
743
744 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
745 {
746         return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
747 }
748
749 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
750 {
751         struct btree *b;
752
753         rcu_read_lock();
754         hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
755                 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
756                         goto out;
757         b = NULL;
758 out:
759         rcu_read_unlock();
760         return b;
761 }
762
763 static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k,
764                                      int level, struct closure *cl)
765 {
766         int ret = -ENOMEM;
767         struct btree *i;
768
769         if (!cl)
770                 return ERR_PTR(-ENOMEM);
771
772         /*
773          * Trying to free up some memory - i.e. reuse some btree nodes - may
774          * require initiating IO to flush the dirty part of the node. If we're
775          * running under generic_make_request(), that IO will never finish and
776          * we would deadlock. Returning -EAGAIN causes the cache lookup code to
777          * punt to workqueue and retry.
778          */
779         if (current->bio_list)
780                 return ERR_PTR(-EAGAIN);
781
782         if (c->try_harder && c->try_harder != cl) {
783                 closure_wait_event_async(&c->try_wait, cl, !c->try_harder);
784                 return ERR_PTR(-EAGAIN);
785         }
786
787         /* XXX: tracepoint */
788         c->try_harder = cl;
789         c->try_harder_start = local_clock();
790 retry:
791         list_for_each_entry_reverse(i, &c->btree_cache, list) {
792                 int r = mca_reap(i, cl, btree_order(k));
793                 if (!r)
794                         return i;
795                 if (r != -ENOMEM)
796                         ret = r;
797         }
798
799         if (ret == -EAGAIN &&
800             closure_blocking(cl)) {
801                 mutex_unlock(&c->bucket_lock);
802                 closure_sync(cl);
803                 mutex_lock(&c->bucket_lock);
804                 goto retry;
805         }
806
807         return ERR_PTR(ret);
808 }
809
810 /*
811  * We can only have one thread cannibalizing other cached btree nodes at a time,
812  * or we'll deadlock. We use an open coded mutex to ensure that, which a
813  * cannibalize_bucket() will take. This means every time we unlock the root of
814  * the btree, we need to release this lock if we have it held.
815  */
816 void bch_cannibalize_unlock(struct cache_set *c, struct closure *cl)
817 {
818         if (c->try_harder == cl) {
819                 bch_time_stats_update(&c->try_harder_time, c->try_harder_start);
820                 c->try_harder = NULL;
821                 __closure_wake_up(&c->try_wait);
822         }
823 }
824
825 static struct btree *mca_alloc(struct cache_set *c, struct bkey *k,
826                                int level, struct closure *cl)
827 {
828         struct btree *b;
829
830         lockdep_assert_held(&c->bucket_lock);
831
832         if (mca_find(c, k))
833                 return NULL;
834
835         /* btree_free() doesn't free memory; it sticks the node on the end of
836          * the list. Check if there's any freed nodes there:
837          */
838         list_for_each_entry(b, &c->btree_cache_freeable, list)
839                 if (!mca_reap(b, NULL, btree_order(k)))
840                         goto out;
841
842         /* We never free struct btree itself, just the memory that holds the on
843          * disk node. Check the freed list before allocating a new one:
844          */
845         list_for_each_entry(b, &c->btree_cache_freed, list)
846                 if (!mca_reap(b, NULL, 0)) {
847                         mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
848                         if (!b->sets[0].data)
849                                 goto err;
850                         else
851                                 goto out;
852                 }
853
854         b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
855         if (!b)
856                 goto err;
857
858         BUG_ON(!down_write_trylock(&b->lock));
859         if (!b->sets->data)
860                 goto err;
861 out:
862         BUG_ON(!closure_is_unlocked(&b->io.cl));
863
864         bkey_copy(&b->key, k);
865         list_move(&b->list, &c->btree_cache);
866         hlist_del_init_rcu(&b->hash);
867         hlist_add_head_rcu(&b->hash, mca_hash(c, k));
868
869         lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
870         b->level        = level;
871
872         mca_reinit(b);
873
874         return b;
875 err:
876         if (b)
877                 rw_unlock(true, b);
878
879         b = mca_cannibalize(c, k, level, cl);
880         if (!IS_ERR(b))
881                 goto out;
882
883         return b;
884 }
885
886 /**
887  * bch_btree_node_get - find a btree node in the cache and lock it, reading it
888  * in from disk if necessary.
889  *
890  * If IO is necessary, it uses the closure embedded in struct btree_op to wait;
891  * if that closure is in non blocking mode, will return -EAGAIN.
892  *
893  * The btree node will have either a read or a write lock held, depending on
894  * level and op->lock.
895  */
896 struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k,
897                                  int level, struct btree_op *op)
898 {
899         int i = 0;
900         bool write = level <= op->lock;
901         struct btree *b;
902
903         BUG_ON(level < 0);
904 retry:
905         b = mca_find(c, k);
906
907         if (!b) {
908                 mutex_lock(&c->bucket_lock);
909                 b = mca_alloc(c, k, level, &op->cl);
910                 mutex_unlock(&c->bucket_lock);
911
912                 if (!b)
913                         goto retry;
914                 if (IS_ERR(b))
915                         return b;
916
917                 bch_btree_read(b);
918
919                 if (!write)
920                         downgrade_write(&b->lock);
921         } else {
922                 rw_lock(write, b, level);
923                 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
924                         rw_unlock(write, b);
925                         goto retry;
926                 }
927                 BUG_ON(b->level != level);
928         }
929
930         b->accessed = 1;
931
932         for (; i <= b->nsets && b->sets[i].size; i++) {
933                 prefetch(b->sets[i].tree);
934                 prefetch(b->sets[i].data);
935         }
936
937         for (; i <= b->nsets; i++)
938                 prefetch(b->sets[i].data);
939
940         if (!closure_wait_event(&b->io.wait, &op->cl,
941                                 btree_node_read_done(b))) {
942                 rw_unlock(write, b);
943                 b = ERR_PTR(-EAGAIN);
944         } else if (btree_node_io_error(b)) {
945                 rw_unlock(write, b);
946                 b = ERR_PTR(-EIO);
947         } else
948                 BUG_ON(!b->written);
949
950         return b;
951 }
952
953 static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level)
954 {
955         struct btree *b;
956
957         mutex_lock(&c->bucket_lock);
958         b = mca_alloc(c, k, level, NULL);
959         mutex_unlock(&c->bucket_lock);
960
961         if (!IS_ERR_OR_NULL(b)) {
962                 bch_btree_read(b);
963                 rw_unlock(true, b);
964         }
965 }
966
967 /* Btree alloc */
968
969 static void btree_node_free(struct btree *b, struct btree_op *op)
970 {
971         unsigned i;
972
973         /*
974          * The BUG_ON() in btree_node_get() implies that we must have a write
975          * lock on parent to free or even invalidate a node
976          */
977         BUG_ON(op->lock <= b->level);
978         BUG_ON(b == b->c->root);
979         pr_debug("bucket %s", pbtree(b));
980
981         if (btree_node_dirty(b))
982                 btree_complete_write(b, btree_current_write(b));
983         clear_bit(BTREE_NODE_dirty, &b->flags);
984
985         if (b->prio_blocked &&
986             !atomic_sub_return(b->prio_blocked, &b->c->prio_blocked))
987                 wake_up(&b->c->alloc_wait);
988
989         b->prio_blocked = 0;
990
991         cancel_delayed_work(&b->work);
992
993         mutex_lock(&b->c->bucket_lock);
994
995         for (i = 0; i < KEY_PTRS(&b->key); i++) {
996                 BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin));
997
998                 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
999                             PTR_BUCKET(b->c, &b->key, i));
1000         }
1001
1002         bch_bucket_free(b->c, &b->key);
1003         mca_bucket_free(b);
1004         mutex_unlock(&b->c->bucket_lock);
1005 }
1006
1007 struct btree *bch_btree_node_alloc(struct cache_set *c, int level,
1008                                    struct closure *cl)
1009 {
1010         BKEY_PADDED(key) k;
1011         struct btree *b = ERR_PTR(-EAGAIN);
1012
1013         mutex_lock(&c->bucket_lock);
1014 retry:
1015         if (__bch_bucket_alloc_set(c, WATERMARK_METADATA, &k.key, 1, cl))
1016                 goto err;
1017
1018         SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1019
1020         b = mca_alloc(c, &k.key, level, cl);
1021         if (IS_ERR(b))
1022                 goto err_free;
1023
1024         if (!b) {
1025                 cache_bug(c,
1026                         "Tried to allocate bucket that was in btree cache");
1027                 __bkey_put(c, &k.key);
1028                 goto retry;
1029         }
1030
1031         set_btree_node_read_done(b);
1032         b->accessed = 1;
1033         bch_bset_init_next(b);
1034
1035         mutex_unlock(&c->bucket_lock);
1036         return b;
1037 err_free:
1038         bch_bucket_free(c, &k.key);
1039         __bkey_put(c, &k.key);
1040 err:
1041         mutex_unlock(&c->bucket_lock);
1042         return b;
1043 }
1044
1045 static struct btree *btree_node_alloc_replacement(struct btree *b,
1046                                                   struct closure *cl)
1047 {
1048         struct btree *n = bch_btree_node_alloc(b->c, b->level, cl);
1049         if (!IS_ERR_OR_NULL(n))
1050                 bch_btree_sort_into(b, n);
1051
1052         return n;
1053 }
1054
1055 /* Garbage collection */
1056
1057 uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k)
1058 {
1059         uint8_t stale = 0;
1060         unsigned i;
1061         struct bucket *g;
1062
1063         /*
1064          * ptr_invalid() can't return true for the keys that mark btree nodes as
1065          * freed, but since ptr_bad() returns true we'll never actually use them
1066          * for anything and thus we don't want mark their pointers here
1067          */
1068         if (!bkey_cmp(k, &ZERO_KEY))
1069                 return stale;
1070
1071         for (i = 0; i < KEY_PTRS(k); i++) {
1072                 if (!ptr_available(c, k, i))
1073                         continue;
1074
1075                 g = PTR_BUCKET(c, k, i);
1076
1077                 if (gen_after(g->gc_gen, PTR_GEN(k, i)))
1078                         g->gc_gen = PTR_GEN(k, i);
1079
1080                 if (ptr_stale(c, k, i)) {
1081                         stale = max(stale, ptr_stale(c, k, i));
1082                         continue;
1083                 }
1084
1085                 cache_bug_on(GC_MARK(g) &&
1086                              (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1087                              c, "inconsistent ptrs: mark = %llu, level = %i",
1088                              GC_MARK(g), level);
1089
1090                 if (level)
1091                         SET_GC_MARK(g, GC_MARK_METADATA);
1092                 else if (KEY_DIRTY(k))
1093                         SET_GC_MARK(g, GC_MARK_DIRTY);
1094
1095                 /* guard against overflow */
1096                 SET_GC_SECTORS_USED(g, min_t(unsigned,
1097                                              GC_SECTORS_USED(g) + KEY_SIZE(k),
1098                                              (1 << 14) - 1));
1099
1100                 BUG_ON(!GC_SECTORS_USED(g));
1101         }
1102
1103         return stale;
1104 }
1105
1106 #define btree_mark_key(b, k)    __bch_btree_mark_key(b->c, b->level, k)
1107
1108 static int btree_gc_mark_node(struct btree *b, unsigned *keys,
1109                               struct gc_stat *gc)
1110 {
1111         uint8_t stale = 0;
1112         unsigned last_dev = -1;
1113         struct bcache_device *d = NULL;
1114         struct bkey *k;
1115         struct btree_iter iter;
1116         struct bset_tree *t;
1117
1118         gc->nodes++;
1119
1120         for_each_key_filter(b, k, &iter, bch_ptr_invalid) {
1121                 if (last_dev != KEY_INODE(k)) {
1122                         last_dev = KEY_INODE(k);
1123
1124                         d = KEY_INODE(k) < b->c->nr_uuids
1125                                 ? b->c->devices[last_dev]
1126                                 : NULL;
1127                 }
1128
1129                 stale = max(stale, btree_mark_key(b, k));
1130
1131                 if (bch_ptr_bad(b, k))
1132                         continue;
1133
1134                 *keys += bkey_u64s(k);
1135
1136                 gc->key_bytes += bkey_u64s(k);
1137                 gc->nkeys++;
1138
1139                 gc->data += KEY_SIZE(k);
1140                 if (KEY_DIRTY(k)) {
1141                         gc->dirty += KEY_SIZE(k);
1142                         if (d)
1143                                 d->sectors_dirty_gc += KEY_SIZE(k);
1144                 }
1145         }
1146
1147         for (t = b->sets; t <= &b->sets[b->nsets]; t++)
1148                 btree_bug_on(t->size &&
1149                              bset_written(b, t) &&
1150                              bkey_cmp(&b->key, &t->end) < 0,
1151                              b, "found short btree key in gc");
1152
1153         return stale;
1154 }
1155
1156 static struct btree *btree_gc_alloc(struct btree *b, struct bkey *k,
1157                                     struct btree_op *op)
1158 {
1159         /*
1160          * We block priorities from being written for the duration of garbage
1161          * collection, so we can't sleep in btree_alloc() ->
1162          * bch_bucket_alloc_set(), or we'd risk deadlock - so we don't pass it
1163          * our closure.
1164          */
1165         struct btree *n = btree_node_alloc_replacement(b, NULL);
1166
1167         if (!IS_ERR_OR_NULL(n)) {
1168                 swap(b, n);
1169
1170                 memcpy(k->ptr, b->key.ptr,
1171                        sizeof(uint64_t) * KEY_PTRS(&b->key));
1172
1173                 __bkey_put(b->c, &b->key);
1174                 atomic_inc(&b->c->prio_blocked);
1175                 b->prio_blocked++;
1176
1177                 btree_node_free(n, op);
1178                 up_write(&n->lock);
1179         }
1180
1181         return b;
1182 }
1183
1184 /*
1185  * Leaving this at 2 until we've got incremental garbage collection done; it
1186  * could be higher (and has been tested with 4) except that garbage collection
1187  * could take much longer, adversely affecting latency.
1188  */
1189 #define GC_MERGE_NODES  2U
1190
1191 struct gc_merge_info {
1192         struct btree    *b;
1193         struct bkey     *k;
1194         unsigned        keys;
1195 };
1196
1197 static void btree_gc_coalesce(struct btree *b, struct btree_op *op,
1198                               struct gc_stat *gc, struct gc_merge_info *r)
1199 {
1200         unsigned nodes = 0, keys = 0, blocks;
1201         int i;
1202
1203         while (nodes < GC_MERGE_NODES && r[nodes].b)
1204                 keys += r[nodes++].keys;
1205
1206         blocks = btree_default_blocks(b->c) * 2 / 3;
1207
1208         if (nodes < 2 ||
1209             __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1))
1210                 return;
1211
1212         for (i = nodes - 1; i >= 0; --i) {
1213                 if (r[i].b->written)
1214                         r[i].b = btree_gc_alloc(r[i].b, r[i].k, op);
1215
1216                 if (r[i].b->written)
1217                         return;
1218         }
1219
1220         for (i = nodes - 1; i > 0; --i) {
1221                 struct bset *n1 = r[i].b->sets->data;
1222                 struct bset *n2 = r[i - 1].b->sets->data;
1223                 struct bkey *k, *last = NULL;
1224
1225                 keys = 0;
1226
1227                 if (i == 1) {
1228                         /*
1229                          * Last node we're not getting rid of - we're getting
1230                          * rid of the node at r[0]. Have to try and fit all of
1231                          * the remaining keys into this node; we can't ensure
1232                          * they will always fit due to rounding and variable
1233                          * length keys (shouldn't be possible in practice,
1234                          * though)
1235                          */
1236                         if (__set_blocks(n1, n1->keys + r->keys,
1237                                          b->c) > btree_blocks(r[i].b))
1238                                 return;
1239
1240                         keys = n2->keys;
1241                         last = &r->b->key;
1242                 } else
1243                         for (k = n2->start;
1244                              k < end(n2);
1245                              k = bkey_next(k)) {
1246                                 if (__set_blocks(n1, n1->keys + keys +
1247                                                  bkey_u64s(k), b->c) > blocks)
1248                                         break;
1249
1250                                 last = k;
1251                                 keys += bkey_u64s(k);
1252                         }
1253
1254                 BUG_ON(__set_blocks(n1, n1->keys + keys,
1255                                     b->c) > btree_blocks(r[i].b));
1256
1257                 if (last) {
1258                         bkey_copy_key(&r[i].b->key, last);
1259                         bkey_copy_key(r[i].k, last);
1260                 }
1261
1262                 memcpy(end(n1),
1263                        n2->start,
1264                        (void *) node(n2, keys) - (void *) n2->start);
1265
1266                 n1->keys += keys;
1267
1268                 memmove(n2->start,
1269                         node(n2, keys),
1270                         (void *) end(n2) - (void *) node(n2, keys));
1271
1272                 n2->keys -= keys;
1273
1274                 r[i].keys       = n1->keys;
1275                 r[i - 1].keys   = n2->keys;
1276         }
1277
1278         btree_node_free(r->b, op);
1279         up_write(&r->b->lock);
1280
1281         pr_debug("coalesced %u nodes", nodes);
1282
1283         gc->nodes--;
1284         nodes--;
1285
1286         memmove(&r[0], &r[1], sizeof(struct gc_merge_info) * nodes);
1287         memset(&r[nodes], 0, sizeof(struct gc_merge_info));
1288 }
1289
1290 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1291                             struct closure *writes, struct gc_stat *gc)
1292 {
1293         void write(struct btree *r)
1294         {
1295                 if (!r->written)
1296                         bch_btree_write(r, true, op);
1297                 else if (btree_node_dirty(r)) {
1298                         BUG_ON(btree_current_write(r)->owner);
1299                         btree_current_write(r)->owner = writes;
1300                         closure_get(writes);
1301
1302                         bch_btree_write(r, true, NULL);
1303                 }
1304
1305                 up_write(&r->lock);
1306         }
1307
1308         int ret = 0, stale;
1309         unsigned i;
1310         struct gc_merge_info r[GC_MERGE_NODES];
1311
1312         memset(r, 0, sizeof(r));
1313
1314         while ((r->k = bch_next_recurse_key(b, &b->c->gc_done))) {
1315                 r->b = bch_btree_node_get(b->c, r->k, b->level - 1, op);
1316
1317                 if (IS_ERR(r->b)) {
1318                         ret = PTR_ERR(r->b);
1319                         break;
1320                 }
1321
1322                 r->keys = 0;
1323                 stale = btree_gc_mark_node(r->b, &r->keys, gc);
1324
1325                 if (!b->written &&
1326                     (r->b->level || stale > 10 ||
1327                      b->c->gc_always_rewrite))
1328                         r->b = btree_gc_alloc(r->b, r->k, op);
1329
1330                 if (r->b->level)
1331                         ret = btree_gc_recurse(r->b, op, writes, gc);
1332
1333                 if (ret) {
1334                         write(r->b);
1335                         break;
1336                 }
1337
1338                 bkey_copy_key(&b->c->gc_done, r->k);
1339
1340                 if (!b->written)
1341                         btree_gc_coalesce(b, op, gc, r);
1342
1343                 if (r[GC_MERGE_NODES - 1].b)
1344                         write(r[GC_MERGE_NODES - 1].b);
1345
1346                 memmove(&r[1], &r[0],
1347                         sizeof(struct gc_merge_info) * (GC_MERGE_NODES - 1));
1348
1349                 /* When we've got incremental GC working, we'll want to do
1350                  * if (should_resched())
1351                  *      return -EAGAIN;
1352                  */
1353                 cond_resched();
1354 #if 0
1355                 if (need_resched()) {
1356                         ret = -EAGAIN;
1357                         break;
1358                 }
1359 #endif
1360         }
1361
1362         for (i = 1; i < GC_MERGE_NODES && r[i].b; i++)
1363                 write(r[i].b);
1364
1365         /* Might have freed some children, must remove their keys */
1366         if (!b->written)
1367                 bch_btree_sort(b);
1368
1369         return ret;
1370 }
1371
1372 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1373                              struct closure *writes, struct gc_stat *gc)
1374 {
1375         struct btree *n = NULL;
1376         unsigned keys = 0;
1377         int ret = 0, stale = btree_gc_mark_node(b, &keys, gc);
1378
1379         if (b->level || stale > 10)
1380                 n = btree_node_alloc_replacement(b, NULL);
1381
1382         if (!IS_ERR_OR_NULL(n))
1383                 swap(b, n);
1384
1385         if (b->level)
1386                 ret = btree_gc_recurse(b, op, writes, gc);
1387
1388         if (!b->written || btree_node_dirty(b)) {
1389                 atomic_inc(&b->c->prio_blocked);
1390                 b->prio_blocked++;
1391                 bch_btree_write(b, true, n ? op : NULL);
1392         }
1393
1394         if (!IS_ERR_OR_NULL(n)) {
1395                 closure_sync(&op->cl);
1396                 bch_btree_set_root(b);
1397                 btree_node_free(n, op);
1398                 rw_unlock(true, b);
1399         }
1400
1401         return ret;
1402 }
1403
1404 static void btree_gc_start(struct cache_set *c)
1405 {
1406         struct cache *ca;
1407         struct bucket *b;
1408         struct bcache_device **d;
1409         unsigned i;
1410
1411         if (!c->gc_mark_valid)
1412                 return;
1413
1414         mutex_lock(&c->bucket_lock);
1415
1416         c->gc_mark_valid = 0;
1417         c->gc_done = ZERO_KEY;
1418
1419         for_each_cache(ca, c, i)
1420                 for_each_bucket(b, ca) {
1421                         b->gc_gen = b->gen;
1422                         if (!atomic_read(&b->pin))
1423                                 SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
1424                 }
1425
1426         for (d = c->devices;
1427              d < c->devices + c->nr_uuids;
1428              d++)
1429                 if (*d)
1430                         (*d)->sectors_dirty_gc = 0;
1431
1432         mutex_unlock(&c->bucket_lock);
1433 }
1434
1435 size_t bch_btree_gc_finish(struct cache_set *c)
1436 {
1437         size_t available = 0;
1438         struct bucket *b;
1439         struct cache *ca;
1440         struct bcache_device **d;
1441         unsigned i;
1442
1443         mutex_lock(&c->bucket_lock);
1444
1445         set_gc_sectors(c);
1446         c->gc_mark_valid = 1;
1447         c->need_gc      = 0;
1448
1449         if (c->root)
1450                 for (i = 0; i < KEY_PTRS(&c->root->key); i++)
1451                         SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i),
1452                                     GC_MARK_METADATA);
1453
1454         for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1455                 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1456                             GC_MARK_METADATA);
1457
1458         for_each_cache(ca, c, i) {
1459                 uint64_t *i;
1460
1461                 ca->invalidate_needs_gc = 0;
1462
1463                 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1464                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1465
1466                 for (i = ca->prio_buckets;
1467                      i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1468                         SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1469
1470                 for_each_bucket(b, ca) {
1471                         b->last_gc      = b->gc_gen;
1472                         c->need_gc      = max(c->need_gc, bucket_gc_gen(b));
1473
1474                         if (!atomic_read(&b->pin) &&
1475                             GC_MARK(b) == GC_MARK_RECLAIMABLE) {
1476                                 available++;
1477                                 if (!GC_SECTORS_USED(b))
1478                                         bch_bucket_add_unused(ca, b);
1479                         }
1480                 }
1481         }
1482
1483         for (d = c->devices;
1484              d < c->devices + c->nr_uuids;
1485              d++)
1486                 if (*d) {
1487                         unsigned long last =
1488                                 atomic_long_read(&((*d)->sectors_dirty));
1489                         long difference = (*d)->sectors_dirty_gc - last;
1490
1491                         pr_debug("sectors dirty off by %li", difference);
1492
1493                         (*d)->sectors_dirty_last += difference;
1494
1495                         atomic_long_set(&((*d)->sectors_dirty),
1496                                         (*d)->sectors_dirty_gc);
1497                 }
1498
1499         mutex_unlock(&c->bucket_lock);
1500         return available;
1501 }
1502
1503 static void bch_btree_gc(struct closure *cl)
1504 {
1505         struct cache_set *c = container_of(cl, struct cache_set, gc.cl);
1506         int ret;
1507         unsigned long available;
1508         struct gc_stat stats;
1509         struct closure writes;
1510         struct btree_op op;
1511
1512         uint64_t start_time = local_clock();
1513         trace_bcache_gc_start(c->sb.set_uuid);
1514         blktrace_msg_all(c, "Starting gc");
1515
1516         memset(&stats, 0, sizeof(struct gc_stat));
1517         closure_init_stack(&writes);
1518         bch_btree_op_init_stack(&op);
1519         op.lock = SHRT_MAX;
1520
1521         btree_gc_start(c);
1522
1523         ret = btree_root(gc_root, c, &op, &writes, &stats);
1524         closure_sync(&op.cl);
1525         closure_sync(&writes);
1526
1527         if (ret) {
1528                 blktrace_msg_all(c, "Stopped gc");
1529                 pr_warn("gc failed!");
1530
1531                 continue_at(cl, bch_btree_gc, bch_gc_wq);
1532         }
1533
1534         /* Possibly wait for new UUIDs or whatever to hit disk */
1535         bch_journal_meta(c, &op.cl);
1536         closure_sync(&op.cl);
1537
1538         available = bch_btree_gc_finish(c);
1539
1540         bch_time_stats_update(&c->btree_gc_time, start_time);
1541
1542         stats.key_bytes *= sizeof(uint64_t);
1543         stats.dirty     <<= 9;
1544         stats.data      <<= 9;
1545         stats.in_use    = (c->nbuckets - available) * 100 / c->nbuckets;
1546         memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1547         blktrace_msg_all(c, "Finished gc");
1548
1549         trace_bcache_gc_end(c->sb.set_uuid);
1550         wake_up(&c->alloc_wait);
1551
1552         continue_at(cl, bch_moving_gc, bch_gc_wq);
1553 }
1554
1555 void bch_queue_gc(struct cache_set *c)
1556 {
1557         closure_trylock_call(&c->gc.cl, bch_btree_gc, bch_gc_wq, &c->cl);
1558 }
1559
1560 /* Initial partial gc */
1561
1562 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op,
1563                                    unsigned long **seen)
1564 {
1565         int ret;
1566         unsigned i;
1567         struct bkey *k;
1568         struct bucket *g;
1569         struct btree_iter iter;
1570
1571         for_each_key_filter(b, k, &iter, bch_ptr_invalid) {
1572                 for (i = 0; i < KEY_PTRS(k); i++) {
1573                         if (!ptr_available(b->c, k, i))
1574                                 continue;
1575
1576                         g = PTR_BUCKET(b->c, k, i);
1577
1578                         if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i),
1579                                                 seen[PTR_DEV(k, i)]) ||
1580                             !ptr_stale(b->c, k, i)) {
1581                                 g->gen = PTR_GEN(k, i);
1582
1583                                 if (b->level)
1584                                         g->prio = BTREE_PRIO;
1585                                 else if (g->prio == BTREE_PRIO)
1586                                         g->prio = INITIAL_PRIO;
1587                         }
1588                 }
1589
1590                 btree_mark_key(b, k);
1591         }
1592
1593         if (b->level) {
1594                 k = bch_next_recurse_key(b, &ZERO_KEY);
1595
1596                 while (k) {
1597                         struct bkey *p = bch_next_recurse_key(b, k);
1598                         if (p)
1599                                 btree_node_prefetch(b->c, p, b->level - 1);
1600
1601                         ret = btree(check_recurse, k, b, op, seen);
1602                         if (ret)
1603                                 return ret;
1604
1605                         k = p;
1606                 }
1607         }
1608
1609         return 0;
1610 }
1611
1612 int bch_btree_check(struct cache_set *c, struct btree_op *op)
1613 {
1614         int ret = -ENOMEM;
1615         unsigned i;
1616         unsigned long *seen[MAX_CACHES_PER_SET];
1617
1618         memset(seen, 0, sizeof(seen));
1619
1620         for (i = 0; c->cache[i]; i++) {
1621                 size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8);
1622                 seen[i] = kmalloc(n, GFP_KERNEL);
1623                 if (!seen[i])
1624                         goto err;
1625
1626                 /* Disables the seen array until prio_read() uses it too */
1627                 memset(seen[i], 0xFF, n);
1628         }
1629
1630         ret = btree_root(check_recurse, c, op, seen);
1631 err:
1632         for (i = 0; i < MAX_CACHES_PER_SET; i++)
1633                 kfree(seen[i]);
1634         return ret;
1635 }
1636
1637 /* Btree insertion */
1638
1639 static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert)
1640 {
1641         struct bset *i = b->sets[b->nsets].data;
1642
1643         memmove((uint64_t *) where + bkey_u64s(insert),
1644                 where,
1645                 (void *) end(i) - (void *) where);
1646
1647         i->keys += bkey_u64s(insert);
1648         bkey_copy(where, insert);
1649         bch_bset_fix_lookup_table(b, where);
1650 }
1651
1652 static bool fix_overlapping_extents(struct btree *b,
1653                                     struct bkey *insert,
1654                                     struct btree_iter *iter,
1655                                     struct btree_op *op)
1656 {
1657         void subtract_dirty(struct bkey *k, int sectors)
1658         {
1659                 struct bcache_device *d = b->c->devices[KEY_INODE(k)];
1660
1661                 if (KEY_DIRTY(k) && d)
1662                         atomic_long_sub(sectors, &d->sectors_dirty);
1663         }
1664
1665         unsigned old_size, sectors_found = 0;
1666
1667         while (1) {
1668                 struct bkey *k = bch_btree_iter_next(iter);
1669                 if (!k ||
1670                     bkey_cmp(&START_KEY(k), insert) >= 0)
1671                         break;
1672
1673                 if (bkey_cmp(k, &START_KEY(insert)) <= 0)
1674                         continue;
1675
1676                 old_size = KEY_SIZE(k);
1677
1678                 /*
1679                  * We might overlap with 0 size extents; we can't skip these
1680                  * because if they're in the set we're inserting to we have to
1681                  * adjust them so they don't overlap with the key we're
1682                  * inserting. But we don't want to check them for BTREE_REPLACE
1683                  * operations.
1684                  */
1685
1686                 if (op->type == BTREE_REPLACE &&
1687                     KEY_SIZE(k)) {
1688                         /*
1689                          * k might have been split since we inserted/found the
1690                          * key we're replacing
1691                          */
1692                         unsigned i;
1693                         uint64_t offset = KEY_START(k) -
1694                                 KEY_START(&op->replace);
1695
1696                         /* But it must be a subset of the replace key */
1697                         if (KEY_START(k) < KEY_START(&op->replace) ||
1698                             KEY_OFFSET(k) > KEY_OFFSET(&op->replace))
1699                                 goto check_failed;
1700
1701                         /* We didn't find a key that we were supposed to */
1702                         if (KEY_START(k) > KEY_START(insert) + sectors_found)
1703                                 goto check_failed;
1704
1705                         if (KEY_PTRS(&op->replace) != KEY_PTRS(k))
1706                                 goto check_failed;
1707
1708                         /* skip past gen */
1709                         offset <<= 8;
1710
1711                         BUG_ON(!KEY_PTRS(&op->replace));
1712
1713                         for (i = 0; i < KEY_PTRS(&op->replace); i++)
1714                                 if (k->ptr[i] != op->replace.ptr[i] + offset)
1715                                         goto check_failed;
1716
1717                         sectors_found = KEY_OFFSET(k) - KEY_START(insert);
1718                 }
1719
1720                 if (bkey_cmp(insert, k) < 0 &&
1721                     bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) {
1722                         /*
1723                          * We overlapped in the middle of an existing key: that
1724                          * means we have to split the old key. But we have to do
1725                          * slightly different things depending on whether the
1726                          * old key has been written out yet.
1727                          */
1728
1729                         struct bkey *top;
1730
1731                         subtract_dirty(k, KEY_SIZE(insert));
1732
1733                         if (bkey_written(b, k)) {
1734                                 /*
1735                                  * We insert a new key to cover the top of the
1736                                  * old key, and the old key is modified in place
1737                                  * to represent the bottom split.
1738                                  *
1739                                  * It's completely arbitrary whether the new key
1740                                  * is the top or the bottom, but it has to match
1741                                  * up with what btree_sort_fixup() does - it
1742                                  * doesn't check for this kind of overlap, it
1743                                  * depends on us inserting a new key for the top
1744                                  * here.
1745                                  */
1746                                 top = bch_bset_search(b, &b->sets[b->nsets],
1747                                                       insert);
1748                                 shift_keys(b, top, k);
1749                         } else {
1750                                 BKEY_PADDED(key) temp;
1751                                 bkey_copy(&temp.key, k);
1752                                 shift_keys(b, k, &temp.key);
1753                                 top = bkey_next(k);
1754                         }
1755
1756                         bch_cut_front(insert, top);
1757                         bch_cut_back(&START_KEY(insert), k);
1758                         bch_bset_fix_invalidated_key(b, k);
1759                         return false;
1760                 }
1761
1762                 if (bkey_cmp(insert, k) < 0) {
1763                         bch_cut_front(insert, k);
1764                 } else {
1765                         if (bkey_written(b, k) &&
1766                             bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) {
1767                                 /*
1768                                  * Completely overwrote, so we don't have to
1769                                  * invalidate the binary search tree
1770                                  */
1771                                 bch_cut_front(k, k);
1772                         } else {
1773                                 __bch_cut_back(&START_KEY(insert), k);
1774                                 bch_bset_fix_invalidated_key(b, k);
1775                         }
1776                 }
1777
1778                 subtract_dirty(k, old_size - KEY_SIZE(k));
1779         }
1780
1781 check_failed:
1782         if (op->type == BTREE_REPLACE) {
1783                 if (!sectors_found) {
1784                         op->insert_collision = true;
1785                         return true;
1786                 } else if (sectors_found < KEY_SIZE(insert)) {
1787                         SET_KEY_OFFSET(insert, KEY_OFFSET(insert) -
1788                                        (KEY_SIZE(insert) - sectors_found));
1789                         SET_KEY_SIZE(insert, sectors_found);
1790                 }
1791         }
1792
1793         return false;
1794 }
1795
1796 static bool btree_insert_key(struct btree *b, struct btree_op *op,
1797                              struct bkey *k)
1798 {
1799         struct bset *i = b->sets[b->nsets].data;
1800         struct bkey *m, *prev;
1801         const char *status = "insert";
1802
1803         BUG_ON(bkey_cmp(k, &b->key) > 0);
1804         BUG_ON(b->level && !KEY_PTRS(k));
1805         BUG_ON(!b->level && !KEY_OFFSET(k));
1806
1807         if (!b->level) {
1808                 struct btree_iter iter;
1809                 struct bkey search = KEY(KEY_INODE(k), KEY_START(k), 0);
1810
1811                 /*
1812                  * bset_search() returns the first key that is strictly greater
1813                  * than the search key - but for back merging, we want to find
1814                  * the first key that is greater than or equal to KEY_START(k) -
1815                  * unless KEY_START(k) is 0.
1816                  */
1817                 if (KEY_OFFSET(&search))
1818                         SET_KEY_OFFSET(&search, KEY_OFFSET(&search) - 1);
1819
1820                 prev = NULL;
1821                 m = bch_btree_iter_init(b, &iter, &search);
1822
1823                 if (fix_overlapping_extents(b, k, &iter, op))
1824                         return false;
1825
1826                 while (m != end(i) &&
1827                        bkey_cmp(k, &START_KEY(m)) > 0)
1828                         prev = m, m = bkey_next(m);
1829
1830                 if (key_merging_disabled(b->c))
1831                         goto insert;
1832
1833                 /* prev is in the tree, if we merge we're done */
1834                 status = "back merging";
1835                 if (prev &&
1836                     bch_bkey_try_merge(b, prev, k))
1837                         goto merged;
1838
1839                 status = "overwrote front";
1840                 if (m != end(i) &&
1841                     KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m))
1842                         goto copy;
1843
1844                 status = "front merge";
1845                 if (m != end(i) &&
1846                     bch_bkey_try_merge(b, k, m))
1847                         goto copy;
1848         } else
1849                 m = bch_bset_search(b, &b->sets[b->nsets], k);
1850
1851 insert: shift_keys(b, m, k);
1852 copy:   bkey_copy(m, k);
1853 merged:
1854         bch_check_keys(b, "%s for %s at %s: %s", status,
1855                        op_type(op), pbtree(b), pkey(k));
1856         bch_check_key_order_msg(b, i, "%s for %s at %s: %s", status,
1857                                 op_type(op), pbtree(b), pkey(k));
1858
1859         if (b->level && !KEY_OFFSET(k))
1860                 b->prio_blocked++;
1861
1862         pr_debug("%s for %s at %s: %s", status,
1863                  op_type(op), pbtree(b), pkey(k));
1864
1865         return true;
1866 }
1867
1868 bool bch_btree_insert_keys(struct btree *b, struct btree_op *op)
1869 {
1870         bool ret = false;
1871         struct bkey *k;
1872         unsigned oldsize = bch_count_data(b);
1873
1874         while ((k = bch_keylist_pop(&op->keys))) {
1875                 bkey_put(b->c, k, b->level);
1876                 ret |= btree_insert_key(b, op, k);
1877         }
1878
1879         BUG_ON(bch_count_data(b) < oldsize);
1880         return ret;
1881 }
1882
1883 bool bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
1884                                    struct bio *bio)
1885 {
1886         bool ret = false;
1887         uint64_t btree_ptr = b->key.ptr[0];
1888         unsigned long seq = b->seq;
1889         BKEY_PADDED(k) tmp;
1890
1891         rw_unlock(false, b);
1892         rw_lock(true, b, b->level);
1893
1894         if (b->key.ptr[0] != btree_ptr ||
1895             b->seq != seq + 1 ||
1896             should_split(b))
1897                 goto out;
1898
1899         op->replace = KEY(op->inode, bio_end(bio), bio_sectors(bio));
1900
1901         SET_KEY_PTRS(&op->replace, 1);
1902         get_random_bytes(&op->replace.ptr[0], sizeof(uint64_t));
1903
1904         SET_PTR_DEV(&op->replace, 0, PTR_CHECK_DEV);
1905
1906         bkey_copy(&tmp.k, &op->replace);
1907
1908         BUG_ON(op->type != BTREE_INSERT);
1909         BUG_ON(!btree_insert_key(b, op, &tmp.k));
1910         bch_btree_write(b, false, NULL);
1911         ret = true;
1912 out:
1913         downgrade_write(&b->lock);
1914         return ret;
1915 }
1916
1917 static int btree_split(struct btree *b, struct btree_op *op)
1918 {
1919         bool split, root = b == b->c->root;
1920         struct btree *n1, *n2 = NULL, *n3 = NULL;
1921         uint64_t start_time = local_clock();
1922
1923         if (b->level)
1924                 set_closure_blocking(&op->cl);
1925
1926         n1 = btree_node_alloc_replacement(b, &op->cl);
1927         if (IS_ERR(n1))
1928                 goto err;
1929
1930         split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5;
1931
1932         pr_debug("%ssplitting at %s keys %i", split ? "" : "not ",
1933                  pbtree(b), n1->sets[0].data->keys);
1934
1935         if (split) {
1936                 unsigned keys = 0;
1937
1938                 n2 = bch_btree_node_alloc(b->c, b->level, &op->cl);
1939                 if (IS_ERR(n2))
1940                         goto err_free1;
1941
1942                 if (root) {
1943                         n3 = bch_btree_node_alloc(b->c, b->level + 1, &op->cl);
1944                         if (IS_ERR(n3))
1945                                 goto err_free2;
1946                 }
1947
1948                 bch_btree_insert_keys(n1, op);
1949
1950                 /* Has to be a linear search because we don't have an auxiliary
1951                  * search tree yet
1952                  */
1953
1954                 while (keys < (n1->sets[0].data->keys * 3) / 5)
1955                         keys += bkey_u64s(node(n1->sets[0].data, keys));
1956
1957                 bkey_copy_key(&n1->key, node(n1->sets[0].data, keys));
1958                 keys += bkey_u64s(node(n1->sets[0].data, keys));
1959
1960                 n2->sets[0].data->keys = n1->sets[0].data->keys - keys;
1961                 n1->sets[0].data->keys = keys;
1962
1963                 memcpy(n2->sets[0].data->start,
1964                        end(n1->sets[0].data),
1965                        n2->sets[0].data->keys * sizeof(uint64_t));
1966
1967                 bkey_copy_key(&n2->key, &b->key);
1968
1969                 bch_keylist_add(&op->keys, &n2->key);
1970                 bch_btree_write(n2, true, op);
1971                 rw_unlock(true, n2);
1972         } else
1973                 bch_btree_insert_keys(n1, op);
1974
1975         bch_keylist_add(&op->keys, &n1->key);
1976         bch_btree_write(n1, true, op);
1977
1978         if (n3) {
1979                 bkey_copy_key(&n3->key, &MAX_KEY);
1980                 bch_btree_insert_keys(n3, op);
1981                 bch_btree_write(n3, true, op);
1982
1983                 closure_sync(&op->cl);
1984                 bch_btree_set_root(n3);
1985                 rw_unlock(true, n3);
1986         } else if (root) {
1987                 op->keys.top = op->keys.bottom;
1988                 closure_sync(&op->cl);
1989                 bch_btree_set_root(n1);
1990         } else {
1991                 unsigned i;
1992
1993                 bkey_copy(op->keys.top, &b->key);
1994                 bkey_copy_key(op->keys.top, &ZERO_KEY);
1995
1996                 for (i = 0; i < KEY_PTRS(&b->key); i++) {
1997                         uint8_t g = PTR_BUCKET(b->c, &b->key, i)->gen + 1;
1998
1999                         SET_PTR_GEN(op->keys.top, i, g);
2000                 }
2001
2002                 bch_keylist_push(&op->keys);
2003                 closure_sync(&op->cl);
2004                 atomic_inc(&b->c->prio_blocked);
2005         }
2006
2007         rw_unlock(true, n1);
2008         btree_node_free(b, op);
2009
2010         bch_time_stats_update(&b->c->btree_split_time, start_time);
2011
2012         return 0;
2013 err_free2:
2014         __bkey_put(n2->c, &n2->key);
2015         btree_node_free(n2, op);
2016         rw_unlock(true, n2);
2017 err_free1:
2018         __bkey_put(n1->c, &n1->key);
2019         btree_node_free(n1, op);
2020         rw_unlock(true, n1);
2021 err:
2022         if (n3 == ERR_PTR(-EAGAIN) ||
2023             n2 == ERR_PTR(-EAGAIN) ||
2024             n1 == ERR_PTR(-EAGAIN))
2025                 return -EAGAIN;
2026
2027         pr_warn("couldn't split");
2028         return -ENOMEM;
2029 }
2030
2031 static int bch_btree_insert_recurse(struct btree *b, struct btree_op *op,
2032                                     struct keylist *stack_keys)
2033 {
2034         if (b->level) {
2035                 int ret;
2036                 struct bkey *insert = op->keys.bottom;
2037                 struct bkey *k = bch_next_recurse_key(b, &START_KEY(insert));
2038
2039                 if (!k) {
2040                         btree_bug(b, "no key to recurse on at level %i/%i",
2041                                   b->level, b->c->root->level);
2042
2043                         op->keys.top = op->keys.bottom;
2044                         return -EIO;
2045                 }
2046
2047                 if (bkey_cmp(insert, k) > 0) {
2048                         unsigned i;
2049
2050                         if (op->type == BTREE_REPLACE) {
2051                                 __bkey_put(b->c, insert);
2052                                 op->keys.top = op->keys.bottom;
2053                                 op->insert_collision = true;
2054                                 return 0;
2055                         }
2056
2057                         for (i = 0; i < KEY_PTRS(insert); i++)
2058                                 atomic_inc(&PTR_BUCKET(b->c, insert, i)->pin);
2059
2060                         bkey_copy(stack_keys->top, insert);
2061
2062                         bch_cut_back(k, insert);
2063                         bch_cut_front(k, stack_keys->top);
2064
2065                         bch_keylist_push(stack_keys);
2066                 }
2067
2068                 ret = btree(insert_recurse, k, b, op, stack_keys);
2069                 if (ret)
2070                         return ret;
2071         }
2072
2073         if (!bch_keylist_empty(&op->keys)) {
2074                 if (should_split(b)) {
2075                         if (op->lock <= b->c->root->level) {
2076                                 BUG_ON(b->level);
2077                                 op->lock = b->c->root->level + 1;
2078                                 return -EINTR;
2079                         }
2080                         return btree_split(b, op);
2081                 }
2082
2083                 BUG_ON(write_block(b) != b->sets[b->nsets].data);
2084
2085                 if (bch_btree_insert_keys(b, op))
2086                         bch_btree_write(b, false, op);
2087         }
2088
2089         return 0;
2090 }
2091
2092 int bch_btree_insert(struct btree_op *op, struct cache_set *c)
2093 {
2094         int ret = 0;
2095         struct keylist stack_keys;
2096
2097         /*
2098          * Don't want to block with the btree locked unless we have to,
2099          * otherwise we get deadlocks with try_harder and between split/gc
2100          */
2101         clear_closure_blocking(&op->cl);
2102
2103         BUG_ON(bch_keylist_empty(&op->keys));
2104         bch_keylist_copy(&stack_keys, &op->keys);
2105         bch_keylist_init(&op->keys);
2106
2107         while (!bch_keylist_empty(&stack_keys) ||
2108                !bch_keylist_empty(&op->keys)) {
2109                 if (bch_keylist_empty(&op->keys)) {
2110                         bch_keylist_add(&op->keys,
2111                                         bch_keylist_pop(&stack_keys));
2112                         op->lock = 0;
2113                 }
2114
2115                 ret = btree_root(insert_recurse, c, op, &stack_keys);
2116
2117                 if (ret == -EAGAIN) {
2118                         ret = 0;
2119                         closure_sync(&op->cl);
2120                 } else if (ret) {
2121                         struct bkey *k;
2122
2123                         pr_err("error %i trying to insert key for %s",
2124                                ret, op_type(op));
2125
2126                         while ((k = bch_keylist_pop(&stack_keys) ?:
2127                                     bch_keylist_pop(&op->keys)))
2128                                 bkey_put(c, k, 0);
2129                 }
2130         }
2131
2132         bch_keylist_free(&stack_keys);
2133
2134         if (op->journal)
2135                 atomic_dec_bug(op->journal);
2136         op->journal = NULL;
2137         return ret;
2138 }
2139
2140 void bch_btree_set_root(struct btree *b)
2141 {
2142         unsigned i;
2143
2144         BUG_ON(!b->written);
2145
2146         for (i = 0; i < KEY_PTRS(&b->key); i++)
2147                 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2148
2149         mutex_lock(&b->c->bucket_lock);
2150         list_del_init(&b->list);
2151         mutex_unlock(&b->c->bucket_lock);
2152
2153         b->c->root = b;
2154         __bkey_put(b->c, &b->key);
2155
2156         bch_journal_meta(b->c, NULL);
2157         pr_debug("%s for %pf", pbtree(b), __builtin_return_address(0));
2158 }
2159
2160 /* Cache lookup */
2161
2162 static int submit_partial_cache_miss(struct btree *b, struct btree_op *op,
2163                                      struct bkey *k)
2164 {
2165         struct search *s = container_of(op, struct search, op);
2166         struct bio *bio = &s->bio.bio;
2167         int ret = 0;
2168
2169         while (!ret &&
2170                !op->lookup_done) {
2171                 unsigned sectors = INT_MAX;
2172
2173                 if (KEY_INODE(k) == op->inode) {
2174                         if (KEY_START(k) <= bio->bi_sector)
2175                                 break;
2176
2177                         sectors = min_t(uint64_t, sectors,
2178                                         KEY_START(k) - bio->bi_sector);
2179                 }
2180
2181                 ret = s->d->cache_miss(b, s, bio, sectors);
2182         }
2183
2184         return ret;
2185 }
2186
2187 /*
2188  * Read from a single key, handling the initial cache miss if the key starts in
2189  * the middle of the bio
2190  */
2191 static int submit_partial_cache_hit(struct btree *b, struct btree_op *op,
2192                                     struct bkey *k)
2193 {
2194         struct search *s = container_of(op, struct search, op);
2195         struct bio *bio = &s->bio.bio;
2196         unsigned ptr;
2197         struct bio *n;
2198
2199         int ret = submit_partial_cache_miss(b, op, k);
2200         if (ret || op->lookup_done)
2201                 return ret;
2202
2203         /* XXX: figure out best pointer - for multiple cache devices */
2204         ptr = 0;
2205
2206         PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
2207
2208         while (!op->lookup_done &&
2209                KEY_INODE(k) == op->inode &&
2210                bio->bi_sector < KEY_OFFSET(k)) {
2211                 struct bkey *bio_key;
2212                 sector_t sector = PTR_OFFSET(k, ptr) +
2213                         (bio->bi_sector - KEY_START(k));
2214                 unsigned sectors = min_t(uint64_t, INT_MAX,
2215                                          KEY_OFFSET(k) - bio->bi_sector);
2216
2217                 n = bch_bio_split(bio, sectors, GFP_NOIO, s->d->bio_split);
2218                 if (!n)
2219                         return -EAGAIN;
2220
2221                 if (n == bio)
2222                         op->lookup_done = true;
2223
2224                 bio_key = &container_of(n, struct bbio, bio)->key;
2225
2226                 /*
2227                  * The bucket we're reading from might be reused while our bio
2228                  * is in flight, and we could then end up reading the wrong
2229                  * data.
2230                  *
2231                  * We guard against this by checking (in cache_read_endio()) if
2232                  * the pointer is stale again; if so, we treat it as an error
2233                  * and reread from the backing device (but we don't pass that
2234                  * error up anywhere).
2235                  */
2236
2237                 bch_bkey_copy_single_ptr(bio_key, k, ptr);
2238                 SET_PTR_OFFSET(bio_key, 0, sector);
2239
2240                 n->bi_end_io    = bch_cache_read_endio;
2241                 n->bi_private   = &s->cl;
2242
2243                 trace_bcache_cache_hit(n);
2244                 __bch_submit_bbio(n, b->c);
2245         }
2246
2247         return 0;
2248 }
2249
2250 int bch_btree_search_recurse(struct btree *b, struct btree_op *op)
2251 {
2252         struct search *s = container_of(op, struct search, op);
2253         struct bio *bio = &s->bio.bio;
2254
2255         int ret = 0;
2256         struct bkey *k;
2257         struct btree_iter iter;
2258         bch_btree_iter_init(b, &iter, &KEY(op->inode, bio->bi_sector, 0));
2259
2260         pr_debug("at %s searching for %u:%llu", pbtree(b), op->inode,
2261                  (uint64_t) bio->bi_sector);
2262
2263         do {
2264                 k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad);
2265                 if (!k) {
2266                         /*
2267                          * b->key would be exactly what we want, except that
2268                          * pointers to btree nodes have nonzero size - we
2269                          * wouldn't go far enough
2270                          */
2271
2272                         ret = submit_partial_cache_miss(b, op,
2273                                         &KEY(KEY_INODE(&b->key),
2274                                              KEY_OFFSET(&b->key), 0));
2275                         break;
2276                 }
2277
2278                 ret = b->level
2279                         ? btree(search_recurse, k, b, op)
2280                         : submit_partial_cache_hit(b, op, k);
2281         } while (!ret &&
2282                  !op->lookup_done);
2283
2284         return ret;
2285 }
2286
2287 /* Keybuf code */
2288
2289 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2290 {
2291         /* Overlapping keys compare equal */
2292         if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2293                 return -1;
2294         if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2295                 return 1;
2296         return 0;
2297 }
2298
2299 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2300                                             struct keybuf_key *r)
2301 {
2302         return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2303 }
2304
2305 static int bch_btree_refill_keybuf(struct btree *b, struct btree_op *op,
2306                                    struct keybuf *buf, struct bkey *end)
2307 {
2308         struct btree_iter iter;
2309         bch_btree_iter_init(b, &iter, &buf->last_scanned);
2310
2311         while (!array_freelist_empty(&buf->freelist)) {
2312                 struct bkey *k = bch_btree_iter_next_filter(&iter, b,
2313                                                             bch_ptr_bad);
2314
2315                 if (!b->level) {
2316                         if (!k) {
2317                                 buf->last_scanned = b->key;
2318                                 break;
2319                         }
2320
2321                         buf->last_scanned = *k;
2322                         if (bkey_cmp(&buf->last_scanned, end) >= 0)
2323                                 break;
2324
2325                         if (buf->key_predicate(buf, k)) {
2326                                 struct keybuf_key *w;
2327
2328                                 pr_debug("%s", pkey(k));
2329
2330                                 spin_lock(&buf->lock);
2331
2332                                 w = array_alloc(&buf->freelist);
2333
2334                                 w->private = NULL;
2335                                 bkey_copy(&w->key, k);
2336
2337                                 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2338                                         array_free(&buf->freelist, w);
2339
2340                                 spin_unlock(&buf->lock);
2341                         }
2342                 } else {
2343                         if (!k)
2344                                 break;
2345
2346                         btree(refill_keybuf, k, b, op, buf, end);
2347                         /*
2348                          * Might get an error here, but can't really do anything
2349                          * and it'll get logged elsewhere. Just read what we
2350                          * can.
2351                          */
2352
2353                         if (bkey_cmp(&buf->last_scanned, end) >= 0)
2354                                 break;
2355
2356                         cond_resched();
2357                 }
2358         }
2359
2360         return 0;
2361 }
2362
2363 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2364                           struct bkey *end)
2365 {
2366         struct bkey start = buf->last_scanned;
2367         struct btree_op op;
2368         bch_btree_op_init_stack(&op);
2369
2370         cond_resched();
2371
2372         btree_root(refill_keybuf, c, &op, buf, end);
2373         closure_sync(&op.cl);
2374
2375         pr_debug("found %s keys from %llu:%llu to %llu:%llu",
2376                  RB_EMPTY_ROOT(&buf->keys) ? "no" :
2377                  array_freelist_empty(&buf->freelist) ? "some" : "a few",
2378                  KEY_INODE(&start), KEY_OFFSET(&start),
2379                  KEY_INODE(&buf->last_scanned), KEY_OFFSET(&buf->last_scanned));
2380
2381         spin_lock(&buf->lock);
2382
2383         if (!RB_EMPTY_ROOT(&buf->keys)) {
2384                 struct keybuf_key *w;
2385                 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2386                 buf->start      = START_KEY(&w->key);
2387
2388                 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2389                 buf->end        = w->key;
2390         } else {
2391                 buf->start      = MAX_KEY;
2392                 buf->end        = MAX_KEY;
2393         }
2394
2395         spin_unlock(&buf->lock);
2396 }
2397
2398 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2399 {
2400         rb_erase(&w->node, &buf->keys);
2401         array_free(&buf->freelist, w);
2402 }
2403
2404 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2405 {
2406         spin_lock(&buf->lock);
2407         __bch_keybuf_del(buf, w);
2408         spin_unlock(&buf->lock);
2409 }
2410
2411 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2412                                   struct bkey *end)
2413 {
2414         bool ret = false;
2415         struct keybuf_key *p, *w, s;
2416         s.key = *start;
2417
2418         if (bkey_cmp(end, &buf->start) <= 0 ||
2419             bkey_cmp(start, &buf->end) >= 0)
2420                 return false;
2421
2422         spin_lock(&buf->lock);
2423         w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2424
2425         while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2426                 p = w;
2427                 w = RB_NEXT(w, node);
2428
2429                 if (p->private)
2430                         ret = true;
2431                 else
2432                         __bch_keybuf_del(buf, p);
2433         }
2434
2435         spin_unlock(&buf->lock);
2436         return ret;
2437 }
2438
2439 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2440 {
2441         struct keybuf_key *w;
2442         spin_lock(&buf->lock);
2443
2444         w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2445
2446         while (w && w->private)
2447                 w = RB_NEXT(w, node);
2448
2449         if (w)
2450                 w->private = ERR_PTR(-EINTR);
2451
2452         spin_unlock(&buf->lock);
2453         return w;
2454 }
2455
2456 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2457                                              struct keybuf *buf,
2458                                              struct bkey *end)
2459 {
2460         struct keybuf_key *ret;
2461
2462         while (1) {
2463                 ret = bch_keybuf_next(buf);
2464                 if (ret)
2465                         break;
2466
2467                 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2468                         pr_debug("scan finished");
2469                         break;
2470                 }
2471
2472                 bch_refill_keybuf(c, buf, end);
2473         }
2474
2475         return ret;
2476 }
2477
2478 void bch_keybuf_init(struct keybuf *buf, keybuf_pred_fn *fn)
2479 {
2480         buf->key_predicate      = fn;
2481         buf->last_scanned       = MAX_KEY;
2482         buf->keys               = RB_ROOT;
2483
2484         spin_lock_init(&buf->lock);
2485         array_allocator_init(&buf->freelist);
2486 }
2487
2488 void bch_btree_exit(void)
2489 {
2490         if (btree_io_wq)
2491                 destroy_workqueue(btree_io_wq);
2492         if (bch_gc_wq)
2493                 destroy_workqueue(bch_gc_wq);
2494 }
2495
2496 int __init bch_btree_init(void)
2497 {
2498         if (!(bch_gc_wq = create_singlethread_workqueue("bch_btree_gc")) ||
2499             !(btree_io_wq = create_singlethread_workqueue("bch_btree_io")))
2500                 return -ENOMEM;
2501
2502         return 0;
2503 }