blob: 00cde40db57269bb173104ab3632b8a99e95c4f6 [file] [log] [blame]
Kent Overstreetcafe5632013-03-23 16:11:31 -07001/*
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"
Kent Overstreet65d45232013-12-20 17:22:05 -080026#include "extents.h"
Kent Overstreetcafe5632013-03-23 16:11:31 -070027
28#include <linux/slab.h>
29#include <linux/bitops.h>
Kent Overstreet72a44512013-10-24 17:19:26 -070030#include <linux/freezer.h>
Kent Overstreetcafe5632013-03-23 16:11:31 -070031#include <linux/hash.h>
Kent Overstreet72a44512013-10-24 17:19:26 -070032#include <linux/kthread.h>
Geert Uytterhoevencd953ed2013-03-27 18:56:28 +010033#include <linux/prefetch.h>
Kent Overstreetcafe5632013-03-23 16:11:31 -070034#include <linux/random.h>
35#include <linux/rcupdate.h>
36#include <trace/events/bcache.h>
37
38/*
39 * Todo:
40 * register_bcache: Return errors out to userspace correctly
41 *
42 * Writeback: don't undirty key until after a cache flush
43 *
44 * Create an iterator for key pointers
45 *
46 * On btree write error, mark bucket such that it won't be freed from the cache
47 *
48 * Journalling:
49 * Check for bad keys in replay
50 * Propagate barriers
51 * Refcount journal entries in journal_replay
52 *
53 * Garbage collection:
54 * Finish incremental gc
55 * Gc should free old UUIDs, data for invalid UUIDs
56 *
57 * Provide a way to list backing device UUIDs we have data cached for, and
58 * probably how long it's been since we've seen them, and a way to invalidate
59 * dirty data for devices that will never be attached again
60 *
61 * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
62 * that based on that and how much dirty data we have we can keep writeback
63 * from being starved
64 *
65 * Add a tracepoint or somesuch to watch for writeback starvation
66 *
67 * When btree depth > 1 and splitting an interior node, we have to make sure
68 * alloc_bucket() cannot fail. This should be true but is not completely
69 * obvious.
70 *
Kent Overstreetcafe5632013-03-23 16:11:31 -070071 * 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 *
Kent Overstreetcafe5632013-03-23 16:11:31 -070076 * Superblock needs to be fleshed out for multiple cache devices
77 *
78 * Add a sysfs tunable for the number of writeback IOs in flight
79 *
80 * Add a sysfs tunable for the number of open data buckets
81 *
82 * IO tracking: Can we track when one process is doing io on behalf of another?
83 * IO tracking: Don't use just an average, weigh more recent stuff higher
84 *
85 * Test module load/unload
86 */
87
Kent Overstreetcafe5632013-03-23 16:11:31 -070088#define MAX_NEED_GC 64
89#define MAX_SAVE_PRIO 72
90
91#define PTR_DIRTY_BIT (((uint64_t) 1 << 36))
92
93#define PTR_HASH(c, k) \
94 (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
95
Kent Overstreetdf8e8972013-07-24 17:37:59 -070096#define insert_lock(s, b) ((b)->level <= (s)->lock)
97
98/*
99 * These macros are for recursing down the btree - they handle the details of
100 * locking and looking up nodes in the cache for you. They're best treated as
101 * mere syntax when reading code that uses them.
102 *
103 * op->lock determines whether we take a read or a write lock at a given depth.
104 * If you've got a read lock and find that you need a write lock (i.e. you're
105 * going to have to split), set op->lock and return -EINTR; btree_root() will
106 * call you again and you'll have the correct lock.
107 */
108
109/**
110 * btree - recurse down the btree on a specified key
111 * @fn: function to call, which will be passed the child node
112 * @key: key to recurse on
113 * @b: parent btree node
114 * @op: pointer to struct btree_op
115 */
116#define btree(fn, key, b, op, ...) \
117({ \
118 int _r, l = (b)->level - 1; \
119 bool _w = l <= (op)->lock; \
Slava Pestov2452cc82014-07-12 00:22:53 -0700120 struct btree *_child = bch_btree_node_get((b)->c, op, key, l, \
121 _w, b); \
Kent Overstreetdf8e8972013-07-24 17:37:59 -0700122 if (!IS_ERR(_child)) { \
Kent Overstreetdf8e8972013-07-24 17:37:59 -0700123 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__); \
124 rw_unlock(_w, _child); \
125 } else \
126 _r = PTR_ERR(_child); \
127 _r; \
128})
129
130/**
131 * btree_root - call a function on the root of the btree
132 * @fn: function to call, which will be passed the child node
133 * @c: cache set
134 * @op: pointer to struct btree_op
135 */
136#define btree_root(fn, c, op, ...) \
137({ \
138 int _r = -EINTR; \
139 do { \
140 struct btree *_b = (c)->root; \
141 bool _w = insert_lock(op, _b); \
142 rw_lock(_w, _b, _b->level); \
143 if (_b == (c)->root && \
144 _w == insert_lock(op, _b)) { \
Kent Overstreetdf8e8972013-07-24 17:37:59 -0700145 _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__); \
146 } \
147 rw_unlock(_w, _b); \
Kent Overstreet0a63b662014-03-17 17:15:53 -0700148 bch_cannibalize_unlock(c); \
Kent Overstreet78365412013-12-17 01:29:34 -0800149 if (_r == -EINTR) \
150 schedule(); \
Kent Overstreetdf8e8972013-07-24 17:37:59 -0700151 } while (_r == -EINTR); \
152 \
Kent Overstreet0a63b662014-03-17 17:15:53 -0700153 finish_wait(&(c)->btree_cache_wait, &(op)->wait); \
Kent Overstreetdf8e8972013-07-24 17:37:59 -0700154 _r; \
155})
156
Kent Overstreeta85e9682013-12-20 17:28:16 -0800157static inline struct bset *write_block(struct btree *b)
158{
159 return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
160}
161
Kent Overstreet2a285682014-03-04 16:42:42 -0800162static void bch_btree_init_next(struct btree *b)
163{
164 /* If not a leaf node, always sort */
165 if (b->level && b->keys.nsets)
166 bch_btree_sort(&b->keys, &b->c->sort);
167 else
168 bch_btree_sort_lazy(&b->keys, &b->c->sort);
169
170 if (b->written < btree_blocks(b))
171 bch_bset_init_next(&b->keys, write_block(b),
172 bset_magic(&b->c->sb));
173
174}
175
Kent Overstreetcafe5632013-03-23 16:11:31 -0700176/* Btree key manipulation */
177
Kent Overstreet3a3b6a42013-07-24 16:46:42 -0700178void bkey_put(struct cache_set *c, struct bkey *k)
Kent Overstreete7c590e2013-09-10 18:39:16 -0700179{
180 unsigned i;
181
182 for (i = 0; i < KEY_PTRS(k); i++)
183 if (ptr_available(c, k, i))
184 atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
185}
186
Kent Overstreetcafe5632013-03-23 16:11:31 -0700187/* Btree IO */
188
189static uint64_t btree_csum_set(struct btree *b, struct bset *i)
190{
191 uint64_t crc = b->key.ptr[0];
Kent Overstreetfafff812013-12-17 21:56:21 -0800192 void *data = (void *) i + 8, *end = bset_bkey_last(i);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700193
Kent Overstreet169ef1c2013-03-28 12:50:55 -0600194 crc = bch_crc64_update(crc, data, end - data);
Kent Overstreetc19ed232013-03-26 13:49:02 -0700195 return crc ^ 0xffffffffffffffffULL;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700196}
197
Kent Overstreet78b77bf2013-12-17 22:49:08 -0800198void bch_btree_node_read_done(struct btree *b)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700199{
Kent Overstreetcafe5632013-03-23 16:11:31 -0700200 const char *err = "bad btree header";
Kent Overstreetee811282013-12-17 23:49:49 -0800201 struct bset *i = btree_bset_first(b);
Kent Overstreet57943512013-04-25 13:58:35 -0700202 struct btree_iter *iter;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700203
Kent Overstreetbcf090e2014-05-19 08:57:55 -0700204 iter = mempool_alloc(b->c->fill_iter, GFP_NOIO);
Kent Overstreet57943512013-04-25 13:58:35 -0700205 iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700206 iter->used = 0;
207
Kent Overstreet280481d2013-10-24 16:36:03 -0700208#ifdef CONFIG_BCACHE_DEBUG
Kent Overstreetc052dd92013-11-11 17:35:24 -0800209 iter->b = &b->keys;
Kent Overstreet280481d2013-10-24 16:36:03 -0700210#endif
211
Kent Overstreet57943512013-04-25 13:58:35 -0700212 if (!i->seq)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700213 goto err;
214
215 for (;
Kent Overstreeta85e9682013-12-20 17:28:16 -0800216 b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700217 i = write_block(b)) {
218 err = "unsupported bset version";
219 if (i->version > BCACHE_BSET_VERSION)
220 goto err;
221
222 err = "bad btree header";
Kent Overstreetee811282013-12-17 23:49:49 -0800223 if (b->written + set_blocks(i, block_bytes(b->c)) >
224 btree_blocks(b))
Kent Overstreetcafe5632013-03-23 16:11:31 -0700225 goto err;
226
227 err = "bad magic";
Kent Overstreet81ab4192013-10-31 15:46:42 -0700228 if (i->magic != bset_magic(&b->c->sb))
Kent Overstreetcafe5632013-03-23 16:11:31 -0700229 goto err;
230
231 err = "bad checksum";
232 switch (i->version) {
233 case 0:
234 if (i->csum != csum_set(i))
235 goto err;
236 break;
237 case BCACHE_BSET_VERSION:
238 if (i->csum != btree_csum_set(b, i))
239 goto err;
240 break;
241 }
242
243 err = "empty set";
Kent Overstreeta85e9682013-12-20 17:28:16 -0800244 if (i != b->keys.set[0].data && !i->keys)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700245 goto err;
246
Kent Overstreetfafff812013-12-17 21:56:21 -0800247 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700248
Kent Overstreetee811282013-12-17 23:49:49 -0800249 b->written += set_blocks(i, block_bytes(b->c));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700250 }
251
252 err = "corrupted btree";
253 for (i = write_block(b);
Kent Overstreeta85e9682013-12-20 17:28:16 -0800254 bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700255 i = ((void *) i) + block_bytes(b->c))
Kent Overstreeta85e9682013-12-20 17:28:16 -0800256 if (i->seq == b->keys.set[0].data->seq)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700257 goto err;
258
Kent Overstreeta85e9682013-12-20 17:28:16 -0800259 bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700260
Kent Overstreeta85e9682013-12-20 17:28:16 -0800261 i = b->keys.set[0].data;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700262 err = "short btree key";
Kent Overstreeta85e9682013-12-20 17:28:16 -0800263 if (b->keys.set[0].size &&
264 bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700265 goto err;
266
267 if (b->written < btree_blocks(b))
Kent Overstreeta85e9682013-12-20 17:28:16 -0800268 bch_bset_init_next(&b->keys, write_block(b),
269 bset_magic(&b->c->sb));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700270out:
Kent Overstreet57943512013-04-25 13:58:35 -0700271 mempool_free(iter, b->c->fill_iter);
272 return;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700273err:
274 set_btree_node_io_error(b);
Kent Overstreet88b9f8c2013-12-17 21:46:35 -0800275 bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
Kent Overstreetcafe5632013-03-23 16:11:31 -0700276 err, PTR_BUCKET_NR(b->c, &b->key, 0),
Kent Overstreet88b9f8c2013-12-17 21:46:35 -0800277 bset_block_offset(b, i), i->keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700278 goto out;
279}
280
Kent Overstreet57943512013-04-25 13:58:35 -0700281static void btree_node_read_endio(struct bio *bio, int error)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700282{
Kent Overstreet57943512013-04-25 13:58:35 -0700283 struct closure *cl = bio->bi_private;
284 closure_put(cl);
285}
Kent Overstreetcafe5632013-03-23 16:11:31 -0700286
Kent Overstreet78b77bf2013-12-17 22:49:08 -0800287static void bch_btree_node_read(struct btree *b)
Kent Overstreet57943512013-04-25 13:58:35 -0700288{
289 uint64_t start_time = local_clock();
290 struct closure cl;
291 struct bio *bio;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700292
Kent Overstreetc37511b2013-04-26 15:39:55 -0700293 trace_bcache_btree_read(b);
294
Kent Overstreet57943512013-04-25 13:58:35 -0700295 closure_init_stack(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700296
Kent Overstreet57943512013-04-25 13:58:35 -0700297 bio = bch_bbio_alloc(b->c);
298 bio->bi_rw = REQ_META|READ_SYNC;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700299 bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
Kent Overstreet57943512013-04-25 13:58:35 -0700300 bio->bi_end_io = btree_node_read_endio;
301 bio->bi_private = &cl;
302
Kent Overstreeta85e9682013-12-20 17:28:16 -0800303 bch_bio_map(bio, b->keys.set[0].data);
Kent Overstreet57943512013-04-25 13:58:35 -0700304
Kent Overstreet57943512013-04-25 13:58:35 -0700305 bch_submit_bbio(bio, b->c, &b->key, 0);
306 closure_sync(&cl);
307
308 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
309 set_btree_node_io_error(b);
310
311 bch_bbio_free(bio, b->c);
312
313 if (btree_node_io_error(b))
314 goto err;
315
316 bch_btree_node_read_done(b);
Kent Overstreet57943512013-04-25 13:58:35 -0700317 bch_time_stats_update(&b->c->btree_read_time, start_time);
Kent Overstreet57943512013-04-25 13:58:35 -0700318
319 return;
320err:
Geert Uytterhoeven61cbd252013-09-23 23:17:30 -0700321 bch_cache_set_error(b->c, "io error reading bucket %zu",
Kent Overstreet57943512013-04-25 13:58:35 -0700322 PTR_BUCKET_NR(b->c, &b->key, 0));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700323}
324
325static void btree_complete_write(struct btree *b, struct btree_write *w)
326{
327 if (w->prio_blocked &&
328 !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
Kent Overstreet119ba0f2013-04-24 19:01:12 -0700329 wake_up_allocators(b->c);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700330
331 if (w->journal) {
332 atomic_dec_bug(w->journal);
333 __closure_wake_up(&b->c->journal.wait);
334 }
335
Kent Overstreetcafe5632013-03-23 16:11:31 -0700336 w->prio_blocked = 0;
337 w->journal = NULL;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700338}
339
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800340static void btree_node_write_unlock(struct closure *cl)
341{
342 struct btree *b = container_of(cl, struct btree, io);
343
344 up(&b->io_mutex);
345}
346
Kent Overstreet57943512013-04-25 13:58:35 -0700347static void __btree_node_write_done(struct closure *cl)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700348{
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800349 struct btree *b = container_of(cl, struct btree, io);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700350 struct btree_write *w = btree_prev_write(b);
351
352 bch_bbio_free(b->bio, b->c);
353 b->bio = NULL;
354 btree_complete_write(b, w);
355
356 if (btree_node_dirty(b))
Kent Overstreet56b30772014-01-23 01:44:55 -0800357 schedule_delayed_work(&b->work, 30 * HZ);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700358
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800359 closure_return_with_destructor(cl, btree_node_write_unlock);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700360}
361
Kent Overstreet57943512013-04-25 13:58:35 -0700362static void btree_node_write_done(struct closure *cl)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700363{
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800364 struct btree *b = container_of(cl, struct btree, io);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700365 struct bio_vec *bv;
366 int n;
367
Kent Overstreet79886132013-11-23 17:19:00 -0800368 bio_for_each_segment_all(bv, b->bio, n)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700369 __free_page(bv->bv_page);
370
Kent Overstreet57943512013-04-25 13:58:35 -0700371 __btree_node_write_done(cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700372}
373
Kent Overstreet57943512013-04-25 13:58:35 -0700374static void btree_node_write_endio(struct bio *bio, int error)
375{
376 struct closure *cl = bio->bi_private;
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800377 struct btree *b = container_of(cl, struct btree, io);
Kent Overstreet57943512013-04-25 13:58:35 -0700378
379 if (error)
380 set_btree_node_io_error(b);
381
382 bch_bbio_count_io_errors(b->c, bio, error, "writing btree");
383 closure_put(cl);
384}
385
386static void do_btree_node_write(struct btree *b)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700387{
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800388 struct closure *cl = &b->io;
Kent Overstreetee811282013-12-17 23:49:49 -0800389 struct bset *i = btree_bset_last(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700390 BKEY_PADDED(key) k;
391
392 i->version = BCACHE_BSET_VERSION;
393 i->csum = btree_csum_set(b, i);
394
Kent Overstreet57943512013-04-25 13:58:35 -0700395 BUG_ON(b->bio);
396 b->bio = bch_bbio_alloc(b->c);
397
398 b->bio->bi_end_io = btree_node_write_endio;
Kent Overstreetfaadf0c2013-11-01 18:03:08 -0700399 b->bio->bi_private = cl;
Kent Overstreete49c7c32013-06-26 17:25:38 -0700400 b->bio->bi_rw = REQ_META|WRITE_SYNC|REQ_FUA;
Kent Overstreetee811282013-12-17 23:49:49 -0800401 b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
Kent Overstreet169ef1c2013-03-28 12:50:55 -0600402 bch_bio_map(b->bio, i);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700403
Kent Overstreete49c7c32013-06-26 17:25:38 -0700404 /*
405 * If we're appending to a leaf node, we don't technically need FUA -
406 * this write just needs to be persisted before the next journal write,
407 * which will be marked FLUSH|FUA.
408 *
409 * Similarly if we're writing a new btree root - the pointer is going to
410 * be in the next journal entry.
411 *
412 * But if we're writing a new btree node (that isn't a root) or
413 * appending to a non leaf btree node, we need either FUA or a flush
414 * when we write the parent with the new pointer. FUA is cheaper than a
415 * flush, and writes appending to leaf nodes aren't blocking anything so
416 * just make all btree node writes FUA to keep things sane.
417 */
418
Kent Overstreetcafe5632013-03-23 16:11:31 -0700419 bkey_copy(&k.key, &b->key);
Kent Overstreetee811282013-12-17 23:49:49 -0800420 SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
Kent Overstreeta85e9682013-12-20 17:28:16 -0800421 bset_sector_offset(&b->keys, i));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700422
Kent Overstreet501d52a2014-05-19 08:55:40 -0700423 if (!bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700424 int j;
425 struct bio_vec *bv;
426 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
427
Kent Overstreet79886132013-11-23 17:19:00 -0800428 bio_for_each_segment_all(bv, b->bio, j)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700429 memcpy(page_address(bv->bv_page),
430 base + j * PAGE_SIZE, PAGE_SIZE);
431
Kent Overstreetcafe5632013-03-23 16:11:31 -0700432 bch_submit_bbio(b->bio, b->c, &k.key, 0);
433
Kent Overstreet57943512013-04-25 13:58:35 -0700434 continue_at(cl, btree_node_write_done, NULL);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700435 } else {
436 b->bio->bi_vcnt = 0;
Kent Overstreet169ef1c2013-03-28 12:50:55 -0600437 bch_bio_map(b->bio, i);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700438
Kent Overstreetcafe5632013-03-23 16:11:31 -0700439 bch_submit_bbio(b->bio, b->c, &k.key, 0);
440
441 closure_sync(cl);
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800442 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700443 }
444}
445
Kent Overstreet2a285682014-03-04 16:42:42 -0800446void __bch_btree_node_write(struct btree *b, struct closure *parent)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700447{
Kent Overstreetee811282013-12-17 23:49:49 -0800448 struct bset *i = btree_bset_last(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700449
Kent Overstreet2a285682014-03-04 16:42:42 -0800450 lockdep_assert_held(&b->write_lock);
451
Kent Overstreetc37511b2013-04-26 15:39:55 -0700452 trace_bcache_btree_write(b);
453
Kent Overstreetcafe5632013-03-23 16:11:31 -0700454 BUG_ON(current->bio_list);
Kent Overstreet57943512013-04-25 13:58:35 -0700455 BUG_ON(b->written >= btree_blocks(b));
456 BUG_ON(b->written && !i->keys);
Kent Overstreetee811282013-12-17 23:49:49 -0800457 BUG_ON(btree_bset_first(b)->seq != i->seq);
Kent Overstreetdc9d98d2013-12-17 23:47:33 -0800458 bch_check_keys(&b->keys, "writing");
Kent Overstreetcafe5632013-03-23 16:11:31 -0700459
Kent Overstreetcafe5632013-03-23 16:11:31 -0700460 cancel_delayed_work(&b->work);
461
Kent Overstreet57943512013-04-25 13:58:35 -0700462 /* If caller isn't waiting for write, parent refcount is cache set */
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800463 down(&b->io_mutex);
464 closure_init(&b->io, parent ?: &b->c->cl);
Kent Overstreet57943512013-04-25 13:58:35 -0700465
Kent Overstreetcafe5632013-03-23 16:11:31 -0700466 clear_bit(BTREE_NODE_dirty, &b->flags);
467 change_bit(BTREE_NODE_write_idx, &b->flags);
468
Kent Overstreet57943512013-04-25 13:58:35 -0700469 do_btree_node_write(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700470
Kent Overstreetee811282013-12-17 23:49:49 -0800471 atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
Kent Overstreetcafe5632013-03-23 16:11:31 -0700472 &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
473
Kent Overstreeta85e9682013-12-20 17:28:16 -0800474 b->written += set_blocks(i, block_bytes(b->c));
Kent Overstreet2a285682014-03-04 16:42:42 -0800475}
Kent Overstreeta85e9682013-12-20 17:28:16 -0800476
Kent Overstreet2a285682014-03-04 16:42:42 -0800477void bch_btree_node_write(struct btree *b, struct closure *parent)
478{
479 unsigned nsets = b->keys.nsets;
480
481 lockdep_assert_held(&b->lock);
482
483 __bch_btree_node_write(b, parent);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700484
Kent Overstreet78b77bf2013-12-17 22:49:08 -0800485 /*
486 * do verify if there was more than one set initially (i.e. we did a
487 * sort) and we sorted down to a single set:
488 */
Kent Overstreet2a285682014-03-04 16:42:42 -0800489 if (nsets && !b->keys.nsets)
Kent Overstreet78b77bf2013-12-17 22:49:08 -0800490 bch_btree_verify(b);
491
Kent Overstreet2a285682014-03-04 16:42:42 -0800492 bch_btree_init_next(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700493}
494
Kent Overstreetf269af52013-07-23 20:48:29 -0700495static void bch_btree_node_write_sync(struct btree *b)
496{
497 struct closure cl;
498
499 closure_init_stack(&cl);
Kent Overstreet2a285682014-03-04 16:42:42 -0800500
501 mutex_lock(&b->write_lock);
Kent Overstreetf269af52013-07-23 20:48:29 -0700502 bch_btree_node_write(b, &cl);
Kent Overstreet2a285682014-03-04 16:42:42 -0800503 mutex_unlock(&b->write_lock);
504
Kent Overstreetf269af52013-07-23 20:48:29 -0700505 closure_sync(&cl);
506}
507
Kent Overstreet57943512013-04-25 13:58:35 -0700508static void btree_node_write_work(struct work_struct *w)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700509{
510 struct btree *b = container_of(to_delayed_work(w), struct btree, work);
511
Kent Overstreet2a285682014-03-04 16:42:42 -0800512 mutex_lock(&b->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700513 if (btree_node_dirty(b))
Kent Overstreet2a285682014-03-04 16:42:42 -0800514 __bch_btree_node_write(b, NULL);
515 mutex_unlock(&b->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700516}
517
Kent Overstreetc18536a2013-07-24 17:44:17 -0700518static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700519{
Kent Overstreetee811282013-12-17 23:49:49 -0800520 struct bset *i = btree_bset_last(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700521 struct btree_write *w = btree_current_write(b);
522
Kent Overstreet2a285682014-03-04 16:42:42 -0800523 lockdep_assert_held(&b->write_lock);
524
Kent Overstreet57943512013-04-25 13:58:35 -0700525 BUG_ON(!b->written);
526 BUG_ON(!i->keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700527
Kent Overstreet57943512013-04-25 13:58:35 -0700528 if (!btree_node_dirty(b))
Kent Overstreet56b30772014-01-23 01:44:55 -0800529 schedule_delayed_work(&b->work, 30 * HZ);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700530
Kent Overstreet57943512013-04-25 13:58:35 -0700531 set_btree_node_dirty(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700532
Kent Overstreetc18536a2013-07-24 17:44:17 -0700533 if (journal_ref) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700534 if (w->journal &&
Kent Overstreetc18536a2013-07-24 17:44:17 -0700535 journal_pin_cmp(b->c, w->journal, journal_ref)) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700536 atomic_dec_bug(w->journal);
537 w->journal = NULL;
538 }
539
540 if (!w->journal) {
Kent Overstreetc18536a2013-07-24 17:44:17 -0700541 w->journal = journal_ref;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700542 atomic_inc(w->journal);
543 }
544 }
545
Kent Overstreetcafe5632013-03-23 16:11:31 -0700546 /* Force write if set is too big */
Kent Overstreet57943512013-04-25 13:58:35 -0700547 if (set_bytes(i) > PAGE_SIZE - 48 &&
548 !current->bio_list)
549 bch_btree_node_write(b, NULL);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700550}
551
552/*
553 * Btree in memory cache - allocation/freeing
554 * mca -> memory cache
555 */
556
Kent Overstreetcafe5632013-03-23 16:11:31 -0700557#define mca_reserve(c) (((c->root && c->root->level) \
558 ? c->root->level : 1) * 8 + 16)
559#define mca_can_free(c) \
Kent Overstreet0a63b662014-03-17 17:15:53 -0700560 max_t(int, 0, c->btree_cache_used - mca_reserve(c))
Kent Overstreetcafe5632013-03-23 16:11:31 -0700561
562static void mca_data_free(struct btree *b)
563{
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800564 BUG_ON(b->io_mutex.count != 1);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700565
Kent Overstreeta85e9682013-12-20 17:28:16 -0800566 bch_btree_keys_free(&b->keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700567
Kent Overstreet0a63b662014-03-17 17:15:53 -0700568 b->c->btree_cache_used--;
Kent Overstreetee811282013-12-17 23:49:49 -0800569 list_move(&b->list, &b->c->btree_cache_freed);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700570}
571
572static void mca_bucket_free(struct btree *b)
573{
574 BUG_ON(btree_node_dirty(b));
575
576 b->key.ptr[0] = 0;
577 hlist_del_init_rcu(&b->hash);
578 list_move(&b->list, &b->c->btree_cache_freeable);
579}
580
581static unsigned btree_order(struct bkey *k)
582{
583 return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
584}
585
586static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
587{
Kent Overstreeta85e9682013-12-20 17:28:16 -0800588 if (!bch_btree_keys_alloc(&b->keys,
Kent Overstreetee811282013-12-17 23:49:49 -0800589 max_t(unsigned,
590 ilog2(b->c->btree_pages),
591 btree_order(k)),
592 gfp)) {
Kent Overstreet0a63b662014-03-17 17:15:53 -0700593 b->c->btree_cache_used++;
Kent Overstreetee811282013-12-17 23:49:49 -0800594 list_move(&b->list, &b->c->btree_cache);
595 } else {
596 list_move(&b->list, &b->c->btree_cache_freed);
597 }
Kent Overstreetcafe5632013-03-23 16:11:31 -0700598}
599
600static struct btree *mca_bucket_alloc(struct cache_set *c,
601 struct bkey *k, gfp_t gfp)
602{
603 struct btree *b = kzalloc(sizeof(struct btree), gfp);
604 if (!b)
605 return NULL;
606
607 init_rwsem(&b->lock);
608 lockdep_set_novalidate_class(&b->lock);
Kent Overstreet2a285682014-03-04 16:42:42 -0800609 mutex_init(&b->write_lock);
610 lockdep_set_novalidate_class(&b->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700611 INIT_LIST_HEAD(&b->list);
Kent Overstreet57943512013-04-25 13:58:35 -0700612 INIT_DELAYED_WORK(&b->work, btree_node_write_work);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700613 b->c = c;
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800614 sema_init(&b->io_mutex, 1);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700615
616 mca_data_alloc(b, k, gfp);
617 return b;
618}
619
Kent Overstreete8e1d462013-07-24 17:27:07 -0700620static int mca_reap(struct btree *b, unsigned min_order, bool flush)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700621{
Kent Overstreete8e1d462013-07-24 17:27:07 -0700622 struct closure cl;
623
624 closure_init_stack(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700625 lockdep_assert_held(&b->c->bucket_lock);
626
627 if (!down_write_trylock(&b->lock))
628 return -ENOMEM;
629
Kent Overstreeta85e9682013-12-20 17:28:16 -0800630 BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
Kent Overstreete8e1d462013-07-24 17:27:07 -0700631
Kent Overstreeta85e9682013-12-20 17:28:16 -0800632 if (b->keys.page_order < min_order)
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800633 goto out_unlock;
634
635 if (!flush) {
636 if (btree_node_dirty(b))
637 goto out_unlock;
638
639 if (down_trylock(&b->io_mutex))
640 goto out_unlock;
641 up(&b->io_mutex);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700642 }
643
Kent Overstreet2a285682014-03-04 16:42:42 -0800644 mutex_lock(&b->write_lock);
Kent Overstreetf269af52013-07-23 20:48:29 -0700645 if (btree_node_dirty(b))
Kent Overstreet2a285682014-03-04 16:42:42 -0800646 __bch_btree_node_write(b, &cl);
647 mutex_unlock(&b->write_lock);
648
649 closure_sync(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700650
Kent Overstreete8e1d462013-07-24 17:27:07 -0700651 /* wait for any in flight btree write */
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800652 down(&b->io_mutex);
653 up(&b->io_mutex);
Kent Overstreete8e1d462013-07-24 17:27:07 -0700654
Kent Overstreetcafe5632013-03-23 16:11:31 -0700655 return 0;
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800656out_unlock:
657 rw_unlock(true, b);
658 return -ENOMEM;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700659}
660
Dave Chinner7dc19d52013-08-28 10:18:11 +1000661static unsigned long bch_mca_scan(struct shrinker *shrink,
662 struct shrink_control *sc)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700663{
664 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
665 struct btree *b, *t;
666 unsigned long i, nr = sc->nr_to_scan;
Dave Chinner7dc19d52013-08-28 10:18:11 +1000667 unsigned long freed = 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700668
669 if (c->shrinker_disabled)
Dave Chinner7dc19d52013-08-28 10:18:11 +1000670 return SHRINK_STOP;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700671
Kent Overstreet0a63b662014-03-17 17:15:53 -0700672 if (c->btree_cache_alloc_lock)
Dave Chinner7dc19d52013-08-28 10:18:11 +1000673 return SHRINK_STOP;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700674
675 /* Return -1 if we can't do anything right now */
Kent Overstreeta698e082013-09-23 23:17:34 -0700676 if (sc->gfp_mask & __GFP_IO)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700677 mutex_lock(&c->bucket_lock);
678 else if (!mutex_trylock(&c->bucket_lock))
679 return -1;
680
Kent Overstreet36c9ea92013-06-03 13:04:56 -0700681 /*
682 * It's _really_ critical that we don't free too many btree nodes - we
683 * have to always leave ourselves a reserve. The reserve is how we
684 * guarantee that allocating memory for a new btree node can always
685 * succeed, so that inserting keys into the btree can always succeed and
686 * IO can always make forward progress:
687 */
Kent Overstreetcafe5632013-03-23 16:11:31 -0700688 nr /= c->btree_pages;
689 nr = min_t(unsigned long, nr, mca_can_free(c));
690
691 i = 0;
692 list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
Dave Chinner7dc19d52013-08-28 10:18:11 +1000693 if (freed >= nr)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700694 break;
695
696 if (++i > 3 &&
Kent Overstreete8e1d462013-07-24 17:27:07 -0700697 !mca_reap(b, 0, false)) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700698 mca_data_free(b);
699 rw_unlock(true, b);
Dave Chinner7dc19d52013-08-28 10:18:11 +1000700 freed++;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700701 }
702 }
703
Kent Overstreet0a63b662014-03-17 17:15:53 -0700704 for (i = 0; (nr--) && i < c->btree_cache_used; i++) {
Kent Overstreetb0f32a52013-12-10 13:24:26 -0800705 if (list_empty(&c->btree_cache))
706 goto out;
707
Kent Overstreetcafe5632013-03-23 16:11:31 -0700708 b = list_first_entry(&c->btree_cache, struct btree, list);
709 list_rotate_left(&c->btree_cache);
710
711 if (!b->accessed &&
Kent Overstreete8e1d462013-07-24 17:27:07 -0700712 !mca_reap(b, 0, false)) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700713 mca_bucket_free(b);
714 mca_data_free(b);
715 rw_unlock(true, b);
Dave Chinner7dc19d52013-08-28 10:18:11 +1000716 freed++;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700717 } else
718 b->accessed = 0;
719 }
720out:
Kent Overstreetcafe5632013-03-23 16:11:31 -0700721 mutex_unlock(&c->bucket_lock);
Dave Chinner7dc19d52013-08-28 10:18:11 +1000722 return freed;
723}
724
725static unsigned long bch_mca_count(struct shrinker *shrink,
726 struct shrink_control *sc)
727{
728 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
729
730 if (c->shrinker_disabled)
731 return 0;
732
Kent Overstreet0a63b662014-03-17 17:15:53 -0700733 if (c->btree_cache_alloc_lock)
Dave Chinner7dc19d52013-08-28 10:18:11 +1000734 return 0;
735
736 return mca_can_free(c) * c->btree_pages;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700737}
738
739void bch_btree_cache_free(struct cache_set *c)
740{
741 struct btree *b;
742 struct closure cl;
743 closure_init_stack(&cl);
744
745 if (c->shrink.list.next)
746 unregister_shrinker(&c->shrink);
747
748 mutex_lock(&c->bucket_lock);
749
750#ifdef CONFIG_BCACHE_DEBUG
751 if (c->verify_data)
752 list_move(&c->verify_data->list, &c->btree_cache);
Kent Overstreet78b77bf2013-12-17 22:49:08 -0800753
754 free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700755#endif
756
757 list_splice(&c->btree_cache_freeable,
758 &c->btree_cache);
759
760 while (!list_empty(&c->btree_cache)) {
761 b = list_first_entry(&c->btree_cache, struct btree, list);
762
763 if (btree_node_dirty(b))
764 btree_complete_write(b, btree_current_write(b));
765 clear_bit(BTREE_NODE_dirty, &b->flags);
766
767 mca_data_free(b);
768 }
769
770 while (!list_empty(&c->btree_cache_freed)) {
771 b = list_first_entry(&c->btree_cache_freed,
772 struct btree, list);
773 list_del(&b->list);
774 cancel_delayed_work_sync(&b->work);
775 kfree(b);
776 }
777
778 mutex_unlock(&c->bucket_lock);
779}
780
781int bch_btree_cache_alloc(struct cache_set *c)
782{
783 unsigned i;
784
Kent Overstreetcafe5632013-03-23 16:11:31 -0700785 for (i = 0; i < mca_reserve(c); i++)
Kent Overstreet72a44512013-10-24 17:19:26 -0700786 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
787 return -ENOMEM;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700788
789 list_splice_init(&c->btree_cache,
790 &c->btree_cache_freeable);
791
792#ifdef CONFIG_BCACHE_DEBUG
793 mutex_init(&c->verify_lock);
794
Kent Overstreet78b77bf2013-12-17 22:49:08 -0800795 c->verify_ondisk = (void *)
796 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
797
Kent Overstreetcafe5632013-03-23 16:11:31 -0700798 c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
799
800 if (c->verify_data &&
Kent Overstreeta85e9682013-12-20 17:28:16 -0800801 c->verify_data->keys.set->data)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700802 list_del_init(&c->verify_data->list);
803 else
804 c->verify_data = NULL;
805#endif
806
Dave Chinner7dc19d52013-08-28 10:18:11 +1000807 c->shrink.count_objects = bch_mca_count;
808 c->shrink.scan_objects = bch_mca_scan;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700809 c->shrink.seeks = 4;
810 c->shrink.batch = c->btree_pages * 2;
811 register_shrinker(&c->shrink);
812
813 return 0;
814}
815
816/* Btree in memory cache - hash table */
817
818static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
819{
820 return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
821}
822
823static struct btree *mca_find(struct cache_set *c, struct bkey *k)
824{
825 struct btree *b;
826
827 rcu_read_lock();
828 hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
829 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
830 goto out;
831 b = NULL;
832out:
833 rcu_read_unlock();
834 return b;
835}
836
Kent Overstreet0a63b662014-03-17 17:15:53 -0700837static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
838{
839 struct task_struct *old;
840
841 old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
842 if (old && old != current) {
843 if (op)
844 prepare_to_wait(&c->btree_cache_wait, &op->wait,
845 TASK_UNINTERRUPTIBLE);
846 return -EINTR;
847 }
848
849 return 0;
850}
851
852static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
853 struct bkey *k)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700854{
Kent Overstreete8e1d462013-07-24 17:27:07 -0700855 struct btree *b;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700856
Kent Overstreetc37511b2013-04-26 15:39:55 -0700857 trace_bcache_btree_cache_cannibalize(c);
858
Kent Overstreet0a63b662014-03-17 17:15:53 -0700859 if (mca_cannibalize_lock(c, op))
860 return ERR_PTR(-EINTR);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700861
Kent Overstreete8e1d462013-07-24 17:27:07 -0700862 list_for_each_entry_reverse(b, &c->btree_cache, list)
863 if (!mca_reap(b, btree_order(k), false))
864 return b;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700865
Kent Overstreete8e1d462013-07-24 17:27:07 -0700866 list_for_each_entry_reverse(b, &c->btree_cache, list)
867 if (!mca_reap(b, btree_order(k), true))
868 return b;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700869
Kent Overstreet0a63b662014-03-17 17:15:53 -0700870 WARN(1, "btree cache cannibalize failed\n");
Kent Overstreete8e1d462013-07-24 17:27:07 -0700871 return ERR_PTR(-ENOMEM);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700872}
873
874/*
875 * We can only have one thread cannibalizing other cached btree nodes at a time,
876 * or we'll deadlock. We use an open coded mutex to ensure that, which a
877 * cannibalize_bucket() will take. This means every time we unlock the root of
878 * the btree, we need to release this lock if we have it held.
879 */
Kent Overstreetdf8e8972013-07-24 17:37:59 -0700880static void bch_cannibalize_unlock(struct cache_set *c)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700881{
Kent Overstreet0a63b662014-03-17 17:15:53 -0700882 if (c->btree_cache_alloc_lock == current) {
883 c->btree_cache_alloc_lock = NULL;
884 wake_up(&c->btree_cache_wait);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700885 }
886}
887
Kent Overstreet0a63b662014-03-17 17:15:53 -0700888static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
889 struct bkey *k, int level)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700890{
891 struct btree *b;
892
Kent Overstreete8e1d462013-07-24 17:27:07 -0700893 BUG_ON(current->bio_list);
894
Kent Overstreetcafe5632013-03-23 16:11:31 -0700895 lockdep_assert_held(&c->bucket_lock);
896
897 if (mca_find(c, k))
898 return NULL;
899
900 /* btree_free() doesn't free memory; it sticks the node on the end of
901 * the list. Check if there's any freed nodes there:
902 */
903 list_for_each_entry(b, &c->btree_cache_freeable, list)
Kent Overstreete8e1d462013-07-24 17:27:07 -0700904 if (!mca_reap(b, btree_order(k), false))
Kent Overstreetcafe5632013-03-23 16:11:31 -0700905 goto out;
906
907 /* We never free struct btree itself, just the memory that holds the on
908 * disk node. Check the freed list before allocating a new one:
909 */
910 list_for_each_entry(b, &c->btree_cache_freed, list)
Kent Overstreete8e1d462013-07-24 17:27:07 -0700911 if (!mca_reap(b, 0, false)) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700912 mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
Kent Overstreeta85e9682013-12-20 17:28:16 -0800913 if (!b->keys.set[0].data)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700914 goto err;
915 else
916 goto out;
917 }
918
919 b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
920 if (!b)
921 goto err;
922
923 BUG_ON(!down_write_trylock(&b->lock));
Kent Overstreeta85e9682013-12-20 17:28:16 -0800924 if (!b->keys.set->data)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700925 goto err;
926out:
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800927 BUG_ON(b->io_mutex.count != 1);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700928
929 bkey_copy(&b->key, k);
930 list_move(&b->list, &c->btree_cache);
931 hlist_del_init_rcu(&b->hash);
932 hlist_add_head_rcu(&b->hash, mca_hash(c, k));
933
934 lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
Kent Overstreetd6fd3b12013-07-24 17:20:19 -0700935 b->parent = (void *) ~0UL;
Kent Overstreeta85e9682013-12-20 17:28:16 -0800936 b->flags = 0;
937 b->written = 0;
938 b->level = level;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700939
Kent Overstreet65d45232013-12-20 17:22:05 -0800940 if (!b->level)
Kent Overstreeta85e9682013-12-20 17:28:16 -0800941 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
942 &b->c->expensive_debug_checks);
Kent Overstreet65d45232013-12-20 17:22:05 -0800943 else
Kent Overstreeta85e9682013-12-20 17:28:16 -0800944 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
945 &b->c->expensive_debug_checks);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700946
947 return b;
948err:
949 if (b)
950 rw_unlock(true, b);
951
Kent Overstreet0a63b662014-03-17 17:15:53 -0700952 b = mca_cannibalize(c, op, k);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700953 if (!IS_ERR(b))
954 goto out;
955
956 return b;
957}
958
959/**
960 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
961 * in from disk if necessary.
962 *
Kent Overstreetb54d6932013-07-24 18:04:18 -0700963 * If IO is necessary and running under generic_make_request, returns -EAGAIN.
Kent Overstreetcafe5632013-03-23 16:11:31 -0700964 *
965 * The btree node will have either a read or a write lock held, depending on
966 * level and op->lock.
967 */
Kent Overstreet0a63b662014-03-17 17:15:53 -0700968struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
Slava Pestov2452cc82014-07-12 00:22:53 -0700969 struct bkey *k, int level, bool write,
970 struct btree *parent)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700971{
972 int i = 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700973 struct btree *b;
974
975 BUG_ON(level < 0);
976retry:
977 b = mca_find(c, k);
978
979 if (!b) {
Kent Overstreet57943512013-04-25 13:58:35 -0700980 if (current->bio_list)
981 return ERR_PTR(-EAGAIN);
982
Kent Overstreetcafe5632013-03-23 16:11:31 -0700983 mutex_lock(&c->bucket_lock);
Kent Overstreet0a63b662014-03-17 17:15:53 -0700984 b = mca_alloc(c, op, k, level);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700985 mutex_unlock(&c->bucket_lock);
986
987 if (!b)
988 goto retry;
989 if (IS_ERR(b))
990 return b;
991
Kent Overstreet57943512013-04-25 13:58:35 -0700992 bch_btree_node_read(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700993
994 if (!write)
995 downgrade_write(&b->lock);
996 } else {
997 rw_lock(write, b, level);
998 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
999 rw_unlock(write, b);
1000 goto retry;
1001 }
1002 BUG_ON(b->level != level);
1003 }
1004
Slava Pestov2452cc82014-07-12 00:22:53 -07001005 b->parent = parent;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001006 b->accessed = 1;
1007
Kent Overstreeta85e9682013-12-20 17:28:16 -08001008 for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1009 prefetch(b->keys.set[i].tree);
1010 prefetch(b->keys.set[i].data);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001011 }
1012
Kent Overstreeta85e9682013-12-20 17:28:16 -08001013 for (; i <= b->keys.nsets; i++)
1014 prefetch(b->keys.set[i].data);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001015
Kent Overstreet57943512013-04-25 13:58:35 -07001016 if (btree_node_io_error(b)) {
Kent Overstreetcafe5632013-03-23 16:11:31 -07001017 rw_unlock(write, b);
Kent Overstreet57943512013-04-25 13:58:35 -07001018 return ERR_PTR(-EIO);
1019 }
1020
1021 BUG_ON(!b->written);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001022
1023 return b;
1024}
1025
Slava Pestov2452cc82014-07-12 00:22:53 -07001026static void btree_node_prefetch(struct btree *parent, struct bkey *k)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001027{
1028 struct btree *b;
1029
Slava Pestov2452cc82014-07-12 00:22:53 -07001030 mutex_lock(&parent->c->bucket_lock);
1031 b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1032 mutex_unlock(&parent->c->bucket_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001033
1034 if (!IS_ERR_OR_NULL(b)) {
Slava Pestov2452cc82014-07-12 00:22:53 -07001035 b->parent = parent;
Kent Overstreet57943512013-04-25 13:58:35 -07001036 bch_btree_node_read(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001037 rw_unlock(true, b);
1038 }
1039}
1040
1041/* Btree alloc */
1042
Kent Overstreete8e1d462013-07-24 17:27:07 -07001043static void btree_node_free(struct btree *b)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001044{
Kent Overstreetc37511b2013-04-26 15:39:55 -07001045 trace_bcache_btree_node_free(b);
1046
Kent Overstreetcafe5632013-03-23 16:11:31 -07001047 BUG_ON(b == b->c->root);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001048
Kent Overstreet2a285682014-03-04 16:42:42 -08001049 mutex_lock(&b->write_lock);
1050
Kent Overstreetcafe5632013-03-23 16:11:31 -07001051 if (btree_node_dirty(b))
1052 btree_complete_write(b, btree_current_write(b));
1053 clear_bit(BTREE_NODE_dirty, &b->flags);
1054
Kent Overstreet2a285682014-03-04 16:42:42 -08001055 mutex_unlock(&b->write_lock);
1056
Kent Overstreetcafe5632013-03-23 16:11:31 -07001057 cancel_delayed_work(&b->work);
1058
1059 mutex_lock(&b->c->bucket_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001060 bch_bucket_free(b->c, &b->key);
1061 mca_bucket_free(b);
1062 mutex_unlock(&b->c->bucket_lock);
1063}
1064
Slava Pestovc5aa4a32014-04-21 18:23:12 -07001065struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
Slava Pestov2452cc82014-07-12 00:22:53 -07001066 int level, bool wait,
1067 struct btree *parent)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001068{
1069 BKEY_PADDED(key) k;
1070 struct btree *b = ERR_PTR(-EAGAIN);
1071
1072 mutex_lock(&c->bucket_lock);
1073retry:
Slava Pestovc5aa4a32014-04-21 18:23:12 -07001074 if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
Kent Overstreetcafe5632013-03-23 16:11:31 -07001075 goto err;
1076
Kent Overstreet3a3b6a42013-07-24 16:46:42 -07001077 bkey_put(c, &k.key);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001078 SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1079
Kent Overstreet0a63b662014-03-17 17:15:53 -07001080 b = mca_alloc(c, op, &k.key, level);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001081 if (IS_ERR(b))
1082 goto err_free;
1083
1084 if (!b) {
Kent Overstreetb1a67b02013-03-25 11:46:44 -07001085 cache_bug(c,
1086 "Tried to allocate bucket that was in btree cache");
Kent Overstreetcafe5632013-03-23 16:11:31 -07001087 goto retry;
1088 }
1089
Kent Overstreetcafe5632013-03-23 16:11:31 -07001090 b->accessed = 1;
Slava Pestov2452cc82014-07-12 00:22:53 -07001091 b->parent = parent;
Kent Overstreeta85e9682013-12-20 17:28:16 -08001092 bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
Kent Overstreetcafe5632013-03-23 16:11:31 -07001093
1094 mutex_unlock(&c->bucket_lock);
Kent Overstreetc37511b2013-04-26 15:39:55 -07001095
1096 trace_bcache_btree_node_alloc(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001097 return b;
1098err_free:
1099 bch_bucket_free(c, &k.key);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001100err:
1101 mutex_unlock(&c->bucket_lock);
Kent Overstreetc37511b2013-04-26 15:39:55 -07001102
Slava Pestov913dc332014-05-23 11:18:35 -07001103 trace_bcache_btree_node_alloc_fail(c);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001104 return b;
1105}
1106
Slava Pestovc5aa4a32014-04-21 18:23:12 -07001107static struct btree *bch_btree_node_alloc(struct cache_set *c,
Slava Pestov2452cc82014-07-12 00:22:53 -07001108 struct btree_op *op, int level,
1109 struct btree *parent)
Slava Pestovc5aa4a32014-04-21 18:23:12 -07001110{
Slava Pestov2452cc82014-07-12 00:22:53 -07001111 return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
Slava Pestovc5aa4a32014-04-21 18:23:12 -07001112}
1113
Kent Overstreet0a63b662014-03-17 17:15:53 -07001114static struct btree *btree_node_alloc_replacement(struct btree *b,
1115 struct btree_op *op)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001116{
Slava Pestov2452cc82014-07-12 00:22:53 -07001117 struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
Kent Overstreet67539e82013-09-10 22:53:34 -07001118 if (!IS_ERR_OR_NULL(n)) {
Kent Overstreet2a285682014-03-04 16:42:42 -08001119 mutex_lock(&n->write_lock);
Kent Overstreet89ebb4a2013-11-11 18:38:51 -08001120 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
Kent Overstreet67539e82013-09-10 22:53:34 -07001121 bkey_copy_key(&n->key, &b->key);
Kent Overstreet2a285682014-03-04 16:42:42 -08001122 mutex_unlock(&n->write_lock);
Kent Overstreet67539e82013-09-10 22:53:34 -07001123 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001124
1125 return n;
1126}
1127
Kent Overstreet8835c122013-07-24 23:18:05 -07001128static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1129{
1130 unsigned i;
1131
Kent Overstreet05335cf2014-03-17 18:22:34 -07001132 mutex_lock(&b->c->bucket_lock);
1133
1134 atomic_inc(&b->c->prio_blocked);
1135
Kent Overstreet8835c122013-07-24 23:18:05 -07001136 bkey_copy(k, &b->key);
1137 bkey_copy_key(k, &ZERO_KEY);
1138
Kent Overstreet05335cf2014-03-17 18:22:34 -07001139 for (i = 0; i < KEY_PTRS(k); i++)
1140 SET_PTR_GEN(k, i,
1141 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1142 PTR_BUCKET(b->c, &b->key, i)));
Kent Overstreet8835c122013-07-24 23:18:05 -07001143
Kent Overstreet05335cf2014-03-17 18:22:34 -07001144 mutex_unlock(&b->c->bucket_lock);
Kent Overstreet8835c122013-07-24 23:18:05 -07001145}
1146
Kent Overstreet78365412013-12-17 01:29:34 -08001147static int btree_check_reserve(struct btree *b, struct btree_op *op)
1148{
1149 struct cache_set *c = b->c;
1150 struct cache *ca;
Kent Overstreet0a63b662014-03-17 17:15:53 -07001151 unsigned i, reserve = (c->root->level - b->level) * 2 + 1;
Kent Overstreet78365412013-12-17 01:29:34 -08001152
1153 mutex_lock(&c->bucket_lock);
1154
1155 for_each_cache(ca, c, i)
1156 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1157 if (op)
Kent Overstreet0a63b662014-03-17 17:15:53 -07001158 prepare_to_wait(&c->btree_cache_wait, &op->wait,
Kent Overstreet78365412013-12-17 01:29:34 -08001159 TASK_UNINTERRUPTIBLE);
Kent Overstreet0a63b662014-03-17 17:15:53 -07001160 mutex_unlock(&c->bucket_lock);
1161 return -EINTR;
Kent Overstreet78365412013-12-17 01:29:34 -08001162 }
1163
1164 mutex_unlock(&c->bucket_lock);
Kent Overstreet0a63b662014-03-17 17:15:53 -07001165
1166 return mca_cannibalize_lock(b->c, op);
Kent Overstreet78365412013-12-17 01:29:34 -08001167}
1168
Kent Overstreetcafe5632013-03-23 16:11:31 -07001169/* Garbage collection */
1170
Kent Overstreet487dded2014-03-17 15:13:26 -07001171static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1172 struct bkey *k)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001173{
1174 uint8_t stale = 0;
1175 unsigned i;
1176 struct bucket *g;
1177
1178 /*
1179 * ptr_invalid() can't return true for the keys that mark btree nodes as
1180 * freed, but since ptr_bad() returns true we'll never actually use them
1181 * for anything and thus we don't want mark their pointers here
1182 */
1183 if (!bkey_cmp(k, &ZERO_KEY))
1184 return stale;
1185
1186 for (i = 0; i < KEY_PTRS(k); i++) {
1187 if (!ptr_available(c, k, i))
1188 continue;
1189
1190 g = PTR_BUCKET(c, k, i);
1191
Kent Overstreet3a2fd9d2014-02-27 17:51:12 -08001192 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1193 g->last_gc = PTR_GEN(k, i);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001194
1195 if (ptr_stale(c, k, i)) {
1196 stale = max(stale, ptr_stale(c, k, i));
1197 continue;
1198 }
1199
1200 cache_bug_on(GC_MARK(g) &&
1201 (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1202 c, "inconsistent ptrs: mark = %llu, level = %i",
1203 GC_MARK(g), level);
1204
1205 if (level)
1206 SET_GC_MARK(g, GC_MARK_METADATA);
1207 else if (KEY_DIRTY(k))
1208 SET_GC_MARK(g, GC_MARK_DIRTY);
Kent Overstreet4fe6a812014-03-13 13:46:29 -07001209 else if (!GC_MARK(g))
1210 SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001211
1212 /* guard against overflow */
1213 SET_GC_SECTORS_USED(g, min_t(unsigned,
1214 GC_SECTORS_USED(g) + KEY_SIZE(k),
Darrick J. Wong94717442014-01-28 16:57:39 -08001215 MAX_GC_SECTORS_USED));
Kent Overstreetcafe5632013-03-23 16:11:31 -07001216
1217 BUG_ON(!GC_SECTORS_USED(g));
1218 }
1219
1220 return stale;
1221}
1222
1223#define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k)
1224
Kent Overstreet487dded2014-03-17 15:13:26 -07001225void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1226{
1227 unsigned i;
1228
1229 for (i = 0; i < KEY_PTRS(k); i++)
1230 if (ptr_available(c, k, i) &&
1231 !ptr_stale(c, k, i)) {
1232 struct bucket *b = PTR_BUCKET(c, k, i);
1233
1234 b->gen = PTR_GEN(k, i);
1235
1236 if (level && bkey_cmp(k, &ZERO_KEY))
1237 b->prio = BTREE_PRIO;
1238 else if (!level && b->prio == BTREE_PRIO)
1239 b->prio = INITIAL_PRIO;
1240 }
1241
1242 __bch_btree_mark_key(c, level, k);
1243}
1244
Kent Overstreeta1f03582013-09-10 19:07:00 -07001245static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001246{
1247 uint8_t stale = 0;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001248 unsigned keys = 0, good_keys = 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001249 struct bkey *k;
1250 struct btree_iter iter;
1251 struct bset_tree *t;
1252
1253 gc->nodes++;
1254
Kent Overstreetc052dd92013-11-11 17:35:24 -08001255 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
Kent Overstreetcafe5632013-03-23 16:11:31 -07001256 stale = max(stale, btree_mark_key(b, k));
Kent Overstreeta1f03582013-09-10 19:07:00 -07001257 keys++;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001258
Kent Overstreeta85e9682013-12-20 17:28:16 -08001259 if (bch_ptr_bad(&b->keys, k))
Kent Overstreetcafe5632013-03-23 16:11:31 -07001260 continue;
1261
Kent Overstreetcafe5632013-03-23 16:11:31 -07001262 gc->key_bytes += bkey_u64s(k);
1263 gc->nkeys++;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001264 good_keys++;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001265
1266 gc->data += KEY_SIZE(k);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001267 }
1268
Kent Overstreeta85e9682013-12-20 17:28:16 -08001269 for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001270 btree_bug_on(t->size &&
Kent Overstreeta85e9682013-12-20 17:28:16 -08001271 bset_written(&b->keys, t) &&
Kent Overstreetcafe5632013-03-23 16:11:31 -07001272 bkey_cmp(&b->key, &t->end) < 0,
1273 b, "found short btree key in gc");
1274
Kent Overstreeta1f03582013-09-10 19:07:00 -07001275 if (b->c->gc_always_rewrite)
1276 return true;
1277
1278 if (stale > 10)
1279 return true;
1280
1281 if ((keys - good_keys) * 2 > keys)
1282 return true;
1283
1284 return false;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001285}
1286
Kent Overstreeta1f03582013-09-10 19:07:00 -07001287#define GC_MERGE_NODES 4U
Kent Overstreetcafe5632013-03-23 16:11:31 -07001288
1289struct gc_merge_info {
1290 struct btree *b;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001291 unsigned keys;
1292};
1293
Kent Overstreeta1f03582013-09-10 19:07:00 -07001294static int bch_btree_insert_node(struct btree *, struct btree_op *,
1295 struct keylist *, atomic_t *, struct bkey *);
Kent Overstreetb54d6932013-07-24 18:04:18 -07001296
Kent Overstreeta1f03582013-09-10 19:07:00 -07001297static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
Kent Overstreet0a63b662014-03-17 17:15:53 -07001298 struct gc_stat *gc, struct gc_merge_info *r)
Kent Overstreeta1f03582013-09-10 19:07:00 -07001299{
1300 unsigned i, nodes = 0, keys = 0, blocks;
1301 struct btree *new_nodes[GC_MERGE_NODES];
Kent Overstreet0a63b662014-03-17 17:15:53 -07001302 struct keylist keylist;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001303 struct closure cl;
1304 struct bkey *k;
1305
Kent Overstreet0a63b662014-03-17 17:15:53 -07001306 bch_keylist_init(&keylist);
1307
1308 if (btree_check_reserve(b, NULL))
1309 return 0;
1310
Kent Overstreeta1f03582013-09-10 19:07:00 -07001311 memset(new_nodes, 0, sizeof(new_nodes));
Kent Overstreetb54d6932013-07-24 18:04:18 -07001312 closure_init_stack(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001313
Kent Overstreeta1f03582013-09-10 19:07:00 -07001314 while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
Kent Overstreetcafe5632013-03-23 16:11:31 -07001315 keys += r[nodes++].keys;
1316
1317 blocks = btree_default_blocks(b->c) * 2 / 3;
1318
1319 if (nodes < 2 ||
Kent Overstreeta85e9682013-12-20 17:28:16 -08001320 __set_blocks(b->keys.set[0].data, keys,
Kent Overstreetee811282013-12-17 23:49:49 -08001321 block_bytes(b->c)) > blocks * (nodes - 1))
Kent Overstreeta1f03582013-09-10 19:07:00 -07001322 return 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001323
Kent Overstreeta1f03582013-09-10 19:07:00 -07001324 for (i = 0; i < nodes; i++) {
Kent Overstreet0a63b662014-03-17 17:15:53 -07001325 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001326 if (IS_ERR_OR_NULL(new_nodes[i]))
1327 goto out_nocoalesce;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001328 }
1329
Kent Overstreet0a63b662014-03-17 17:15:53 -07001330 /*
1331 * We have to check the reserve here, after we've allocated our new
1332 * nodes, to make sure the insert below will succeed - we also check
1333 * before as an optimization to potentially avoid a bunch of expensive
1334 * allocs/sorts
1335 */
1336 if (btree_check_reserve(b, NULL))
1337 goto out_nocoalesce;
1338
Kent Overstreet2a285682014-03-04 16:42:42 -08001339 for (i = 0; i < nodes; i++)
1340 mutex_lock(&new_nodes[i]->write_lock);
1341
Kent Overstreetcafe5632013-03-23 16:11:31 -07001342 for (i = nodes - 1; i > 0; --i) {
Kent Overstreetee811282013-12-17 23:49:49 -08001343 struct bset *n1 = btree_bset_first(new_nodes[i]);
1344 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001345 struct bkey *k, *last = NULL;
1346
1347 keys = 0;
1348
Kent Overstreeta1f03582013-09-10 19:07:00 -07001349 if (i > 1) {
Kent Overstreetcafe5632013-03-23 16:11:31 -07001350 for (k = n2->start;
Kent Overstreetfafff812013-12-17 21:56:21 -08001351 k < bset_bkey_last(n2);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001352 k = bkey_next(k)) {
1353 if (__set_blocks(n1, n1->keys + keys +
Kent Overstreetee811282013-12-17 23:49:49 -08001354 bkey_u64s(k),
1355 block_bytes(b->c)) > blocks)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001356 break;
1357
1358 last = k;
1359 keys += bkey_u64s(k);
1360 }
Kent Overstreeta1f03582013-09-10 19:07:00 -07001361 } else {
1362 /*
1363 * Last node we're not getting rid of - we're getting
1364 * rid of the node at r[0]. Have to try and fit all of
1365 * the remaining keys into this node; we can't ensure
1366 * they will always fit due to rounding and variable
1367 * length keys (shouldn't be possible in practice,
1368 * though)
1369 */
1370 if (__set_blocks(n1, n1->keys + n2->keys,
Kent Overstreetee811282013-12-17 23:49:49 -08001371 block_bytes(b->c)) >
1372 btree_blocks(new_nodes[i]))
Kent Overstreeta1f03582013-09-10 19:07:00 -07001373 goto out_nocoalesce;
1374
1375 keys = n2->keys;
1376 /* Take the key of the node we're getting rid of */
1377 last = &r->b->key;
1378 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001379
Kent Overstreetee811282013-12-17 23:49:49 -08001380 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1381 btree_blocks(new_nodes[i]));
Kent Overstreetcafe5632013-03-23 16:11:31 -07001382
Kent Overstreeta1f03582013-09-10 19:07:00 -07001383 if (last)
1384 bkey_copy_key(&new_nodes[i]->key, last);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001385
Kent Overstreetfafff812013-12-17 21:56:21 -08001386 memcpy(bset_bkey_last(n1),
Kent Overstreetcafe5632013-03-23 16:11:31 -07001387 n2->start,
Kent Overstreetfafff812013-12-17 21:56:21 -08001388 (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001389
1390 n1->keys += keys;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001391 r[i].keys = n1->keys;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001392
1393 memmove(n2->start,
Kent Overstreetfafff812013-12-17 21:56:21 -08001394 bset_bkey_idx(n2, keys),
1395 (void *) bset_bkey_last(n2) -
1396 (void *) bset_bkey_idx(n2, keys));
Kent Overstreetcafe5632013-03-23 16:11:31 -07001397
1398 n2->keys -= keys;
1399
Kent Overstreet0a63b662014-03-17 17:15:53 -07001400 if (__bch_keylist_realloc(&keylist,
Kent Overstreet085d2a32013-11-11 18:20:51 -08001401 bkey_u64s(&new_nodes[i]->key)))
Kent Overstreeta1f03582013-09-10 19:07:00 -07001402 goto out_nocoalesce;
1403
1404 bch_btree_node_write(new_nodes[i], &cl);
Kent Overstreet0a63b662014-03-17 17:15:53 -07001405 bch_keylist_add(&keylist, &new_nodes[i]->key);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001406 }
1407
Kent Overstreet2a285682014-03-04 16:42:42 -08001408 for (i = 0; i < nodes; i++)
1409 mutex_unlock(&new_nodes[i]->write_lock);
1410
Kent Overstreet05335cf2014-03-17 18:22:34 -07001411 closure_sync(&cl);
1412
1413 /* We emptied out this node */
1414 BUG_ON(btree_bset_first(new_nodes[0])->keys);
1415 btree_node_free(new_nodes[0]);
1416 rw_unlock(true, new_nodes[0]);
Slava Pestov400ffaa2014-07-12 21:53:11 -07001417 new_nodes[0] = NULL;
Kent Overstreet05335cf2014-03-17 18:22:34 -07001418
Kent Overstreeta1f03582013-09-10 19:07:00 -07001419 for (i = 0; i < nodes; i++) {
Kent Overstreet0a63b662014-03-17 17:15:53 -07001420 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
Kent Overstreeta1f03582013-09-10 19:07:00 -07001421 goto out_nocoalesce;
1422
Kent Overstreet0a63b662014-03-17 17:15:53 -07001423 make_btree_freeing_key(r[i].b, keylist.top);
1424 bch_keylist_push(&keylist);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001425 }
1426
Kent Overstreet0a63b662014-03-17 17:15:53 -07001427 bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1428 BUG_ON(!bch_keylist_empty(&keylist));
Kent Overstreeta1f03582013-09-10 19:07:00 -07001429
1430 for (i = 0; i < nodes; i++) {
1431 btree_node_free(r[i].b);
1432 rw_unlock(true, r[i].b);
1433
1434 r[i].b = new_nodes[i];
1435 }
1436
Kent Overstreeta1f03582013-09-10 19:07:00 -07001437 memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1438 r[nodes - 1].b = ERR_PTR(-EINTR);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001439
Kent Overstreetc37511b2013-04-26 15:39:55 -07001440 trace_bcache_btree_gc_coalesce(nodes);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001441 gc->nodes--;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001442
Kent Overstreet0a63b662014-03-17 17:15:53 -07001443 bch_keylist_free(&keylist);
1444
Kent Overstreeta1f03582013-09-10 19:07:00 -07001445 /* Invalidated our iterator */
1446 return -EINTR;
1447
1448out_nocoalesce:
1449 closure_sync(&cl);
Kent Overstreet0a63b662014-03-17 17:15:53 -07001450 bch_keylist_free(&keylist);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001451
Kent Overstreet0a63b662014-03-17 17:15:53 -07001452 while ((k = bch_keylist_pop(&keylist)))
Kent Overstreeta1f03582013-09-10 19:07:00 -07001453 if (!bkey_cmp(k, &ZERO_KEY))
1454 atomic_dec(&b->c->prio_blocked);
1455
1456 for (i = 0; i < nodes; i++)
1457 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1458 btree_node_free(new_nodes[i]);
1459 rw_unlock(true, new_nodes[i]);
1460 }
1461 return 0;
1462}
1463
Kent Overstreet0a63b662014-03-17 17:15:53 -07001464static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1465 struct btree *replace)
1466{
1467 struct keylist keys;
1468 struct btree *n;
1469
1470 if (btree_check_reserve(b, NULL))
1471 return 0;
1472
1473 n = btree_node_alloc_replacement(replace, NULL);
1474
1475 /* recheck reserve after allocating replacement node */
1476 if (btree_check_reserve(b, NULL)) {
1477 btree_node_free(n);
1478 rw_unlock(true, n);
1479 return 0;
1480 }
1481
1482 bch_btree_node_write_sync(n);
1483
1484 bch_keylist_init(&keys);
1485 bch_keylist_add(&keys, &n->key);
1486
1487 make_btree_freeing_key(replace, keys.top);
1488 bch_keylist_push(&keys);
1489
1490 bch_btree_insert_node(b, op, &keys, NULL, NULL);
1491 BUG_ON(!bch_keylist_empty(&keys));
1492
1493 btree_node_free(replace);
1494 rw_unlock(true, n);
1495
1496 /* Invalidated our iterator */
1497 return -EINTR;
1498}
1499
Kent Overstreeta1f03582013-09-10 19:07:00 -07001500static unsigned btree_gc_count_keys(struct btree *b)
1501{
1502 struct bkey *k;
1503 struct btree_iter iter;
1504 unsigned ret = 0;
1505
Kent Overstreetc052dd92013-11-11 17:35:24 -08001506 for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
Kent Overstreeta1f03582013-09-10 19:07:00 -07001507 ret += bkey_u64s(k);
1508
1509 return ret;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001510}
1511
1512static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1513 struct closure *writes, struct gc_stat *gc)
1514{
Kent Overstreeta1f03582013-09-10 19:07:00 -07001515 int ret = 0;
1516 bool should_rewrite;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001517 struct bkey *k;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001518 struct btree_iter iter;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001519 struct gc_merge_info r[GC_MERGE_NODES];
Kent Overstreet2a285682014-03-04 16:42:42 -08001520 struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001521
Kent Overstreetc052dd92013-11-11 17:35:24 -08001522 bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001523
Kent Overstreet2a285682014-03-04 16:42:42 -08001524 for (i = r; i < r + ARRAY_SIZE(r); i++)
1525 i->b = ERR_PTR(-EINTR);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001526
Kent Overstreeta1f03582013-09-10 19:07:00 -07001527 while (1) {
Kent Overstreeta85e9682013-12-20 17:28:16 -08001528 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001529 if (k) {
Kent Overstreet0a63b662014-03-17 17:15:53 -07001530 r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
Slava Pestov2452cc82014-07-12 00:22:53 -07001531 true, b);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001532 if (IS_ERR(r->b)) {
1533 ret = PTR_ERR(r->b);
1534 break;
1535 }
1536
1537 r->keys = btree_gc_count_keys(r->b);
1538
Kent Overstreet0a63b662014-03-17 17:15:53 -07001539 ret = btree_gc_coalesce(b, op, gc, r);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001540 if (ret)
1541 break;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001542 }
1543
Kent Overstreeta1f03582013-09-10 19:07:00 -07001544 if (!last->b)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001545 break;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001546
1547 if (!IS_ERR(last->b)) {
1548 should_rewrite = btree_gc_mark_node(last->b, gc);
Kent Overstreet0a63b662014-03-17 17:15:53 -07001549 if (should_rewrite) {
1550 ret = btree_gc_rewrite_node(b, op, last->b);
1551 if (ret)
Kent Overstreeta1f03582013-09-10 19:07:00 -07001552 break;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001553 }
1554
1555 if (last->b->level) {
1556 ret = btree_gc_recurse(last->b, op, writes, gc);
1557 if (ret)
1558 break;
1559 }
1560
1561 bkey_copy_key(&b->c->gc_done, &last->b->key);
1562
1563 /*
1564 * Must flush leaf nodes before gc ends, since replace
1565 * operations aren't journalled
1566 */
Kent Overstreet2a285682014-03-04 16:42:42 -08001567 mutex_lock(&last->b->write_lock);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001568 if (btree_node_dirty(last->b))
1569 bch_btree_node_write(last->b, writes);
Kent Overstreet2a285682014-03-04 16:42:42 -08001570 mutex_unlock(&last->b->write_lock);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001571 rw_unlock(true, last->b);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001572 }
1573
Kent Overstreeta1f03582013-09-10 19:07:00 -07001574 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1575 r->b = NULL;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001576
Kent Overstreetcafe5632013-03-23 16:11:31 -07001577 if (need_resched()) {
1578 ret = -EAGAIN;
1579 break;
1580 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001581 }
1582
Kent Overstreet2a285682014-03-04 16:42:42 -08001583 for (i = r; i < r + ARRAY_SIZE(r); i++)
1584 if (!IS_ERR_OR_NULL(i->b)) {
1585 mutex_lock(&i->b->write_lock);
1586 if (btree_node_dirty(i->b))
1587 bch_btree_node_write(i->b, writes);
1588 mutex_unlock(&i->b->write_lock);
1589 rw_unlock(true, i->b);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001590 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001591
Kent Overstreetcafe5632013-03-23 16:11:31 -07001592 return ret;
1593}
1594
1595static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1596 struct closure *writes, struct gc_stat *gc)
1597{
1598 struct btree *n = NULL;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001599 int ret = 0;
1600 bool should_rewrite;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001601
Kent Overstreeta1f03582013-09-10 19:07:00 -07001602 should_rewrite = btree_gc_mark_node(b, gc);
1603 if (should_rewrite) {
Kent Overstreet0a63b662014-03-17 17:15:53 -07001604 n = btree_node_alloc_replacement(b, NULL);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001605
Kent Overstreeta1f03582013-09-10 19:07:00 -07001606 if (!IS_ERR_OR_NULL(n)) {
1607 bch_btree_node_write_sync(n);
Kent Overstreet2a285682014-03-04 16:42:42 -08001608
Kent Overstreeta1f03582013-09-10 19:07:00 -07001609 bch_btree_set_root(n);
1610 btree_node_free(b);
1611 rw_unlock(true, n);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001612
Kent Overstreeta1f03582013-09-10 19:07:00 -07001613 return -EINTR;
1614 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001615 }
1616
Kent Overstreet487dded2014-03-17 15:13:26 -07001617 __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1618
Kent Overstreeta1f03582013-09-10 19:07:00 -07001619 if (b->level) {
1620 ret = btree_gc_recurse(b, op, writes, gc);
1621 if (ret)
1622 return ret;
1623 }
1624
1625 bkey_copy_key(&b->c->gc_done, &b->key);
1626
Kent Overstreetcafe5632013-03-23 16:11:31 -07001627 return ret;
1628}
1629
1630static void btree_gc_start(struct cache_set *c)
1631{
1632 struct cache *ca;
1633 struct bucket *b;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001634 unsigned i;
1635
1636 if (!c->gc_mark_valid)
1637 return;
1638
1639 mutex_lock(&c->bucket_lock);
1640
1641 c->gc_mark_valid = 0;
1642 c->gc_done = ZERO_KEY;
1643
1644 for_each_cache(ca, c, i)
1645 for_each_bucket(b, ca) {
Kent Overstreet3a2fd9d2014-02-27 17:51:12 -08001646 b->last_gc = b->gen;
Kent Overstreet29ebf462013-07-11 19:43:21 -07001647 if (!atomic_read(&b->pin)) {
Kent Overstreet4fe6a812014-03-13 13:46:29 -07001648 SET_GC_MARK(b, 0);
Kent Overstreet29ebf462013-07-11 19:43:21 -07001649 SET_GC_SECTORS_USED(b, 0);
1650 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001651 }
1652
Kent Overstreetcafe5632013-03-23 16:11:31 -07001653 mutex_unlock(&c->bucket_lock);
1654}
1655
Kent Overstreet2531d9ee2014-03-17 16:55:55 -07001656static size_t bch_btree_gc_finish(struct cache_set *c)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001657{
1658 size_t available = 0;
1659 struct bucket *b;
1660 struct cache *ca;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001661 unsigned i;
1662
1663 mutex_lock(&c->bucket_lock);
1664
1665 set_gc_sectors(c);
1666 c->gc_mark_valid = 1;
1667 c->need_gc = 0;
1668
Kent Overstreetcafe5632013-03-23 16:11:31 -07001669 for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1670 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1671 GC_MARK_METADATA);
1672
Nicholas Swensonbf0a6282013-11-26 19:14:23 -08001673 /* don't reclaim buckets to which writeback keys point */
1674 rcu_read_lock();
1675 for (i = 0; i < c->nr_uuids; i++) {
1676 struct bcache_device *d = c->devices[i];
1677 struct cached_dev *dc;
1678 struct keybuf_key *w, *n;
1679 unsigned j;
1680
1681 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1682 continue;
1683 dc = container_of(d, struct cached_dev, disk);
1684
1685 spin_lock(&dc->writeback_keys.lock);
1686 rbtree_postorder_for_each_entry_safe(w, n,
1687 &dc->writeback_keys.keys, node)
1688 for (j = 0; j < KEY_PTRS(&w->key); j++)
1689 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1690 GC_MARK_DIRTY);
1691 spin_unlock(&dc->writeback_keys.lock);
1692 }
1693 rcu_read_unlock();
1694
Kent Overstreetcafe5632013-03-23 16:11:31 -07001695 for_each_cache(ca, c, i) {
1696 uint64_t *i;
1697
1698 ca->invalidate_needs_gc = 0;
1699
1700 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1701 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1702
1703 for (i = ca->prio_buckets;
1704 i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1705 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1706
1707 for_each_bucket(b, ca) {
Kent Overstreetcafe5632013-03-23 16:11:31 -07001708 c->need_gc = max(c->need_gc, bucket_gc_gen(b));
1709
Kent Overstreet4fe6a812014-03-13 13:46:29 -07001710 if (atomic_read(&b->pin))
1711 continue;
1712
1713 BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1714
1715 if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001716 available++;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001717 }
1718 }
1719
Kent Overstreetcafe5632013-03-23 16:11:31 -07001720 mutex_unlock(&c->bucket_lock);
1721 return available;
1722}
1723
Kent Overstreet72a44512013-10-24 17:19:26 -07001724static void bch_btree_gc(struct cache_set *c)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001725{
Kent Overstreetcafe5632013-03-23 16:11:31 -07001726 int ret;
1727 unsigned long available;
1728 struct gc_stat stats;
1729 struct closure writes;
1730 struct btree_op op;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001731 uint64_t start_time = local_clock();
Kent Overstreet57943512013-04-25 13:58:35 -07001732
Kent Overstreetc37511b2013-04-26 15:39:55 -07001733 trace_bcache_gc_start(c);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001734
1735 memset(&stats, 0, sizeof(struct gc_stat));
1736 closure_init_stack(&writes);
Kent Overstreetb54d6932013-07-24 18:04:18 -07001737 bch_btree_op_init(&op, SHRT_MAX);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001738
1739 btree_gc_start(c);
1740
Kent Overstreeta1f03582013-09-10 19:07:00 -07001741 do {
1742 ret = btree_root(gc_root, c, &op, &writes, &stats);
1743 closure_sync(&writes);
Kent Overstreet57943512013-04-25 13:58:35 -07001744
Kent Overstreeta1f03582013-09-10 19:07:00 -07001745 if (ret && ret != -EAGAIN)
1746 pr_warn("gc failed!");
1747 } while (ret);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001748
1749 available = bch_btree_gc_finish(c);
Kent Overstreet57943512013-04-25 13:58:35 -07001750 wake_up_allocators(c);
1751
Kent Overstreet169ef1c2013-03-28 12:50:55 -06001752 bch_time_stats_update(&c->btree_gc_time, start_time);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001753
1754 stats.key_bytes *= sizeof(uint64_t);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001755 stats.data <<= 9;
1756 stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets;
1757 memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
Kent Overstreetcafe5632013-03-23 16:11:31 -07001758
Kent Overstreetc37511b2013-04-26 15:39:55 -07001759 trace_bcache_gc_end(c);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001760
Kent Overstreet72a44512013-10-24 17:19:26 -07001761 bch_moving_gc(c);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001762}
1763
Kent Overstreet72a44512013-10-24 17:19:26 -07001764static int bch_gc_thread(void *arg)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001765{
Kent Overstreet72a44512013-10-24 17:19:26 -07001766 struct cache_set *c = arg;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001767 struct cache *ca;
1768 unsigned i;
Kent Overstreet72a44512013-10-24 17:19:26 -07001769
1770 while (1) {
Kent Overstreeta1f03582013-09-10 19:07:00 -07001771again:
Kent Overstreet72a44512013-10-24 17:19:26 -07001772 bch_btree_gc(c);
1773
1774 set_current_state(TASK_INTERRUPTIBLE);
1775 if (kthread_should_stop())
1776 break;
1777
Kent Overstreeta1f03582013-09-10 19:07:00 -07001778 mutex_lock(&c->bucket_lock);
1779
1780 for_each_cache(ca, c, i)
1781 if (ca->invalidate_needs_gc) {
1782 mutex_unlock(&c->bucket_lock);
1783 set_current_state(TASK_RUNNING);
1784 goto again;
1785 }
1786
1787 mutex_unlock(&c->bucket_lock);
1788
Kent Overstreet72a44512013-10-24 17:19:26 -07001789 try_to_freeze();
1790 schedule();
1791 }
1792
1793 return 0;
1794}
1795
1796int bch_gc_thread_start(struct cache_set *c)
1797{
1798 c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc");
1799 if (IS_ERR(c->gc_thread))
1800 return PTR_ERR(c->gc_thread);
1801
1802 set_task_state(c->gc_thread, TASK_INTERRUPTIBLE);
1803 return 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001804}
1805
1806/* Initial partial gc */
1807
Kent Overstreet487dded2014-03-17 15:13:26 -07001808static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001809{
Kent Overstreet50310162013-09-10 17:18:59 -07001810 int ret = 0;
Kent Overstreet50310162013-09-10 17:18:59 -07001811 struct bkey *k, *p = NULL;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001812 struct btree_iter iter;
1813
Kent Overstreet487dded2014-03-17 15:13:26 -07001814 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1815 bch_initial_mark_key(b->c, b->level, k);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001816
Kent Overstreet487dded2014-03-17 15:13:26 -07001817 bch_initial_mark_key(b->c, b->level + 1, &b->key);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001818
1819 if (b->level) {
Kent Overstreetc052dd92013-11-11 17:35:24 -08001820 bch_btree_iter_init(&b->keys, &iter, NULL);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001821
Kent Overstreet50310162013-09-10 17:18:59 -07001822 do {
Kent Overstreeta85e9682013-12-20 17:28:16 -08001823 k = bch_btree_iter_next_filter(&iter, &b->keys,
1824 bch_ptr_bad);
Kent Overstreet50310162013-09-10 17:18:59 -07001825 if (k)
Slava Pestov2452cc82014-07-12 00:22:53 -07001826 btree_node_prefetch(b, k);
Kent Overstreet50310162013-09-10 17:18:59 -07001827
Kent Overstreetcafe5632013-03-23 16:11:31 -07001828 if (p)
Kent Overstreet487dded2014-03-17 15:13:26 -07001829 ret = btree(check_recurse, p, b, op);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001830
Kent Overstreet50310162013-09-10 17:18:59 -07001831 p = k;
1832 } while (p && !ret);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001833 }
1834
Kent Overstreet487dded2014-03-17 15:13:26 -07001835 return ret;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001836}
1837
Kent Overstreetc18536a2013-07-24 17:44:17 -07001838int bch_btree_check(struct cache_set *c)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001839{
Kent Overstreetc18536a2013-07-24 17:44:17 -07001840 struct btree_op op;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001841
Kent Overstreetb54d6932013-07-24 18:04:18 -07001842 bch_btree_op_init(&op, SHRT_MAX);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001843
Kent Overstreet487dded2014-03-17 15:13:26 -07001844 return btree_root(check_recurse, c, &op);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001845}
1846
Kent Overstreet2531d9ee2014-03-17 16:55:55 -07001847void bch_initial_gc_finish(struct cache_set *c)
1848{
1849 struct cache *ca;
1850 struct bucket *b;
1851 unsigned i;
1852
1853 bch_btree_gc_finish(c);
1854
1855 mutex_lock(&c->bucket_lock);
1856
1857 /*
1858 * We need to put some unused buckets directly on the prio freelist in
1859 * order to get the allocator thread started - it needs freed buckets in
1860 * order to rewrite the prios and gens, and it needs to rewrite prios
1861 * and gens in order to free buckets.
1862 *
1863 * This is only safe for buckets that have no live data in them, which
1864 * there should always be some of.
1865 */
1866 for_each_cache(ca, c, i) {
1867 for_each_bucket(b, ca) {
1868 if (fifo_full(&ca->free[RESERVE_PRIO]))
1869 break;
1870
1871 if (bch_can_invalidate_bucket(ca, b) &&
1872 !GC_MARK(b)) {
1873 __bch_invalidate_one_bucket(ca, b);
1874 fifo_push(&ca->free[RESERVE_PRIO],
1875 b - ca->buckets);
1876 }
1877 }
1878 }
1879
1880 mutex_unlock(&c->bucket_lock);
1881}
1882
Kent Overstreetcafe5632013-03-23 16:11:31 -07001883/* Btree insertion */
1884
Kent Overstreet829a60b2013-11-11 17:02:31 -08001885static bool btree_insert_key(struct btree *b, struct bkey *k,
1886 struct bkey *replace_key)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001887{
Kent Overstreet829a60b2013-11-11 17:02:31 -08001888 unsigned status;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001889
1890 BUG_ON(bkey_cmp(k, &b->key) > 0);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001891
Kent Overstreet829a60b2013-11-11 17:02:31 -08001892 status = bch_btree_insert_key(&b->keys, k, replace_key);
1893 if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1894 bch_check_keys(&b->keys, "%u for %s", status,
1895 replace_key ? "replace" : "insert");
Kent Overstreetcafe5632013-03-23 16:11:31 -07001896
Kent Overstreet829a60b2013-11-11 17:02:31 -08001897 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1898 status);
1899 return true;
1900 } else
1901 return false;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001902}
1903
Kent Overstreet59158fd2013-11-11 19:03:54 -08001904static size_t insert_u64s_remaining(struct btree *b)
1905{
Kent Overstreet35723242014-01-10 18:53:02 -08001906 long ret = bch_btree_keys_u64s_remaining(&b->keys);
Kent Overstreet59158fd2013-11-11 19:03:54 -08001907
1908 /*
1909 * Might land in the middle of an existing extent and have to split it
1910 */
1911 if (b->keys.ops->is_extents)
1912 ret -= KEY_MAX_U64S;
1913
1914 return max(ret, 0L);
1915}
1916
Kent Overstreet26c949f2013-09-10 18:41:15 -07001917static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
Kent Overstreet1b207d82013-09-10 18:52:54 -07001918 struct keylist *insert_keys,
1919 struct bkey *replace_key)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001920{
1921 bool ret = false;
Kent Overstreetdc9d98d2013-12-17 23:47:33 -08001922 int oldsize = bch_count_data(&b->keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001923
Kent Overstreet26c949f2013-09-10 18:41:15 -07001924 while (!bch_keylist_empty(insert_keys)) {
Kent Overstreetc2f95ae2013-07-24 17:24:25 -07001925 struct bkey *k = insert_keys->keys;
Kent Overstreet26c949f2013-09-10 18:41:15 -07001926
Kent Overstreet59158fd2013-11-11 19:03:54 -08001927 if (bkey_u64s(k) > insert_u64s_remaining(b))
Kent Overstreet403b6cd2013-07-24 17:22:44 -07001928 break;
1929
1930 if (bkey_cmp(k, &b->key) <= 0) {
Kent Overstreet3a3b6a42013-07-24 16:46:42 -07001931 if (!b->level)
1932 bkey_put(b->c, k);
Kent Overstreet26c949f2013-09-10 18:41:15 -07001933
Kent Overstreet829a60b2013-11-11 17:02:31 -08001934 ret |= btree_insert_key(b, k, replace_key);
Kent Overstreet26c949f2013-09-10 18:41:15 -07001935 bch_keylist_pop_front(insert_keys);
1936 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
Kent Overstreet26c949f2013-09-10 18:41:15 -07001937 BKEY_PADDED(key) temp;
Kent Overstreetc2f95ae2013-07-24 17:24:25 -07001938 bkey_copy(&temp.key, insert_keys->keys);
Kent Overstreet26c949f2013-09-10 18:41:15 -07001939
1940 bch_cut_back(&b->key, &temp.key);
Kent Overstreetc2f95ae2013-07-24 17:24:25 -07001941 bch_cut_front(&b->key, insert_keys->keys);
Kent Overstreet26c949f2013-09-10 18:41:15 -07001942
Kent Overstreet829a60b2013-11-11 17:02:31 -08001943 ret |= btree_insert_key(b, &temp.key, replace_key);
Kent Overstreet26c949f2013-09-10 18:41:15 -07001944 break;
1945 } else {
1946 break;
1947 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001948 }
1949
Kent Overstreet829a60b2013-11-11 17:02:31 -08001950 if (!ret)
1951 op->insert_collision = true;
1952
Kent Overstreet403b6cd2013-07-24 17:22:44 -07001953 BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1954
Kent Overstreetdc9d98d2013-12-17 23:47:33 -08001955 BUG_ON(bch_count_data(&b->keys) < oldsize);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001956 return ret;
1957}
1958
Kent Overstreet26c949f2013-09-10 18:41:15 -07001959static int btree_split(struct btree *b, struct btree_op *op,
1960 struct keylist *insert_keys,
Kent Overstreet1b207d82013-09-10 18:52:54 -07001961 struct bkey *replace_key)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001962{
Kent Overstreetd6fd3b12013-07-24 17:20:19 -07001963 bool split;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001964 struct btree *n1, *n2 = NULL, *n3 = NULL;
1965 uint64_t start_time = local_clock();
Kent Overstreetb54d6932013-07-24 18:04:18 -07001966 struct closure cl;
Kent Overstreet17e21a92013-07-26 12:32:38 -07001967 struct keylist parent_keys;
Kent Overstreetb54d6932013-07-24 18:04:18 -07001968
1969 closure_init_stack(&cl);
Kent Overstreet17e21a92013-07-26 12:32:38 -07001970 bch_keylist_init(&parent_keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001971
Kent Overstreet0a63b662014-03-17 17:15:53 -07001972 if (btree_check_reserve(b, op)) {
1973 if (!b->level)
1974 return -EINTR;
1975 else
1976 WARN(1, "insufficient reserve for split\n");
1977 }
Kent Overstreet78365412013-12-17 01:29:34 -08001978
Kent Overstreet0a63b662014-03-17 17:15:53 -07001979 n1 = btree_node_alloc_replacement(b, op);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001980 if (IS_ERR(n1))
1981 goto err;
1982
Kent Overstreetee811282013-12-17 23:49:49 -08001983 split = set_blocks(btree_bset_first(n1),
1984 block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001985
Kent Overstreetcafe5632013-03-23 16:11:31 -07001986 if (split) {
1987 unsigned keys = 0;
1988
Kent Overstreetee811282013-12-17 23:49:49 -08001989 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
Kent Overstreetc37511b2013-04-26 15:39:55 -07001990
Slava Pestov2452cc82014-07-12 00:22:53 -07001991 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001992 if (IS_ERR(n2))
1993 goto err_free1;
1994
Kent Overstreetd6fd3b12013-07-24 17:20:19 -07001995 if (!b->parent) {
Slava Pestov2452cc82014-07-12 00:22:53 -07001996 n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001997 if (IS_ERR(n3))
1998 goto err_free2;
1999 }
2000
Kent Overstreet2a285682014-03-04 16:42:42 -08002001 mutex_lock(&n1->write_lock);
2002 mutex_lock(&n2->write_lock);
2003
Kent Overstreet1b207d82013-09-10 18:52:54 -07002004 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002005
Kent Overstreetd6fd3b12013-07-24 17:20:19 -07002006 /*
2007 * Has to be a linear search because we don't have an auxiliary
Kent Overstreetcafe5632013-03-23 16:11:31 -07002008 * search tree yet
2009 */
2010
Kent Overstreetee811282013-12-17 23:49:49 -08002011 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2012 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
Kent Overstreetfafff812013-12-17 21:56:21 -08002013 keys));
Kent Overstreetcafe5632013-03-23 16:11:31 -07002014
Kent Overstreetfafff812013-12-17 21:56:21 -08002015 bkey_copy_key(&n1->key,
Kent Overstreetee811282013-12-17 23:49:49 -08002016 bset_bkey_idx(btree_bset_first(n1), keys));
2017 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
Kent Overstreetcafe5632013-03-23 16:11:31 -07002018
Kent Overstreetee811282013-12-17 23:49:49 -08002019 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2020 btree_bset_first(n1)->keys = keys;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002021
Kent Overstreetee811282013-12-17 23:49:49 -08002022 memcpy(btree_bset_first(n2)->start,
2023 bset_bkey_last(btree_bset_first(n1)),
2024 btree_bset_first(n2)->keys * sizeof(uint64_t));
Kent Overstreetcafe5632013-03-23 16:11:31 -07002025
2026 bkey_copy_key(&n2->key, &b->key);
2027
Kent Overstreet17e21a92013-07-26 12:32:38 -07002028 bch_keylist_add(&parent_keys, &n2->key);
Kent Overstreetb54d6932013-07-24 18:04:18 -07002029 bch_btree_node_write(n2, &cl);
Kent Overstreet2a285682014-03-04 16:42:42 -08002030 mutex_unlock(&n2->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002031 rw_unlock(true, n2);
Kent Overstreetc37511b2013-04-26 15:39:55 -07002032 } else {
Kent Overstreetee811282013-12-17 23:49:49 -08002033 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
Kent Overstreetc37511b2013-04-26 15:39:55 -07002034
Kent Overstreet2a285682014-03-04 16:42:42 -08002035 mutex_lock(&n1->write_lock);
Kent Overstreet1b207d82013-09-10 18:52:54 -07002036 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
Kent Overstreetc37511b2013-04-26 15:39:55 -07002037 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07002038
Kent Overstreet17e21a92013-07-26 12:32:38 -07002039 bch_keylist_add(&parent_keys, &n1->key);
Kent Overstreetb54d6932013-07-24 18:04:18 -07002040 bch_btree_node_write(n1, &cl);
Kent Overstreet2a285682014-03-04 16:42:42 -08002041 mutex_unlock(&n1->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002042
2043 if (n3) {
Kent Overstreetd6fd3b12013-07-24 17:20:19 -07002044 /* Depth increases, make a new root */
Kent Overstreet2a285682014-03-04 16:42:42 -08002045 mutex_lock(&n3->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002046 bkey_copy_key(&n3->key, &MAX_KEY);
Kent Overstreet17e21a92013-07-26 12:32:38 -07002047 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
Kent Overstreetb54d6932013-07-24 18:04:18 -07002048 bch_btree_node_write(n3, &cl);
Kent Overstreet2a285682014-03-04 16:42:42 -08002049 mutex_unlock(&n3->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002050
Kent Overstreetb54d6932013-07-24 18:04:18 -07002051 closure_sync(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002052 bch_btree_set_root(n3);
2053 rw_unlock(true, n3);
Kent Overstreetd6fd3b12013-07-24 17:20:19 -07002054 } else if (!b->parent) {
2055 /* Root filled up but didn't need to be split */
Kent Overstreetb54d6932013-07-24 18:04:18 -07002056 closure_sync(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002057 bch_btree_set_root(n1);
2058 } else {
Kent Overstreet17e21a92013-07-26 12:32:38 -07002059 /* Split a non root node */
Kent Overstreetb54d6932013-07-24 18:04:18 -07002060 closure_sync(&cl);
Kent Overstreet17e21a92013-07-26 12:32:38 -07002061 make_btree_freeing_key(b, parent_keys.top);
2062 bch_keylist_push(&parent_keys);
2063
Kent Overstreet17e21a92013-07-26 12:32:38 -07002064 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2065 BUG_ON(!bch_keylist_empty(&parent_keys));
Kent Overstreetcafe5632013-03-23 16:11:31 -07002066 }
2067
Kent Overstreet05335cf2014-03-17 18:22:34 -07002068 btree_node_free(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002069 rw_unlock(true, n1);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002070
Kent Overstreet169ef1c2013-03-28 12:50:55 -06002071 bch_time_stats_update(&b->c->btree_split_time, start_time);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002072
2073 return 0;
2074err_free2:
Kent Overstreet5f5837d2013-12-16 16:38:49 -08002075 bkey_put(b->c, &n2->key);
Kent Overstreete8e1d462013-07-24 17:27:07 -07002076 btree_node_free(n2);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002077 rw_unlock(true, n2);
2078err_free1:
Kent Overstreet5f5837d2013-12-16 16:38:49 -08002079 bkey_put(b->c, &n1->key);
Kent Overstreete8e1d462013-07-24 17:27:07 -07002080 btree_node_free(n1);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002081 rw_unlock(true, n1);
2082err:
Kent Overstreet0a63b662014-03-17 17:15:53 -07002083 WARN(1, "bcache: btree split failed (level %u)", b->level);
Kent Overstreet5f5837d2013-12-16 16:38:49 -08002084
Kent Overstreetcafe5632013-03-23 16:11:31 -07002085 if (n3 == ERR_PTR(-EAGAIN) ||
2086 n2 == ERR_PTR(-EAGAIN) ||
2087 n1 == ERR_PTR(-EAGAIN))
2088 return -EAGAIN;
2089
Kent Overstreetcafe5632013-03-23 16:11:31 -07002090 return -ENOMEM;
2091}
2092
Kent Overstreet26c949f2013-09-10 18:41:15 -07002093static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
Kent Overstreetc18536a2013-07-24 17:44:17 -07002094 struct keylist *insert_keys,
Kent Overstreet1b207d82013-09-10 18:52:54 -07002095 atomic_t *journal_ref,
2096 struct bkey *replace_key)
Kent Overstreet26c949f2013-09-10 18:41:15 -07002097{
Kent Overstreet2a285682014-03-04 16:42:42 -08002098 struct closure cl;
2099
Kent Overstreet17e21a92013-07-26 12:32:38 -07002100 BUG_ON(b->level && replace_key);
Kent Overstreet26c949f2013-09-10 18:41:15 -07002101
Kent Overstreet2a285682014-03-04 16:42:42 -08002102 closure_init_stack(&cl);
2103
2104 mutex_lock(&b->write_lock);
2105
2106 if (write_block(b) != btree_bset_last(b) &&
2107 b->keys.last_set_unwritten)
2108 bch_btree_init_next(b); /* just wrote a set */
2109
Kent Overstreet59158fd2013-11-11 19:03:54 -08002110 if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
Kent Overstreet2a285682014-03-04 16:42:42 -08002111 mutex_unlock(&b->write_lock);
2112 goto split;
2113 }
Kent Overstreet3b3e9e52013-12-07 03:57:58 -08002114
Kent Overstreet2a285682014-03-04 16:42:42 -08002115 BUG_ON(write_block(b) != btree_bset_last(b));
2116
2117 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2118 if (!b->level)
2119 bch_btree_leaf_dirty(b, journal_ref);
2120 else
2121 bch_btree_node_write(b, &cl);
2122 }
2123
2124 mutex_unlock(&b->write_lock);
2125
2126 /* wait for btree node write if necessary, after unlock */
2127 closure_sync(&cl);
2128
2129 return 0;
2130split:
2131 if (current->bio_list) {
2132 op->lock = b->c->root->level + 1;
2133 return -EAGAIN;
2134 } else if (op->lock <= b->c->root->level) {
2135 op->lock = b->c->root->level + 1;
2136 return -EINTR;
Kent Overstreet17e21a92013-07-26 12:32:38 -07002137 } else {
Kent Overstreet2a285682014-03-04 16:42:42 -08002138 /* Invalidated all iterators */
2139 int ret = btree_split(b, op, insert_keys, replace_key);
Kent Overstreet26c949f2013-09-10 18:41:15 -07002140
Kent Overstreet2a285682014-03-04 16:42:42 -08002141 if (bch_keylist_empty(insert_keys))
2142 return 0;
2143 else if (!ret)
2144 return -EINTR;
2145 return ret;
Kent Overstreet17e21a92013-07-26 12:32:38 -07002146 }
Kent Overstreet26c949f2013-09-10 18:41:15 -07002147}
2148
Kent Overstreete7c590e2013-09-10 18:39:16 -07002149int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2150 struct bkey *check_key)
2151{
2152 int ret = -EINTR;
2153 uint64_t btree_ptr = b->key.ptr[0];
2154 unsigned long seq = b->seq;
2155 struct keylist insert;
2156 bool upgrade = op->lock == -1;
2157
2158 bch_keylist_init(&insert);
2159
2160 if (upgrade) {
2161 rw_unlock(false, b);
2162 rw_lock(true, b, b->level);
2163
2164 if (b->key.ptr[0] != btree_ptr ||
2165 b->seq != seq + 1)
2166 goto out;
2167 }
2168
2169 SET_KEY_PTRS(check_key, 1);
2170 get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2171
2172 SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2173
2174 bch_keylist_add(&insert, check_key);
2175
Kent Overstreet1b207d82013-09-10 18:52:54 -07002176 ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
Kent Overstreete7c590e2013-09-10 18:39:16 -07002177
2178 BUG_ON(!ret && !bch_keylist_empty(&insert));
2179out:
2180 if (upgrade)
2181 downgrade_write(&b->lock);
2182 return ret;
2183}
2184
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002185struct btree_insert_op {
2186 struct btree_op op;
2187 struct keylist *keys;
2188 atomic_t *journal_ref;
2189 struct bkey *replace_key;
2190};
2191
Wei Yongjun08239ca2013-11-28 10:31:35 +08002192static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
Kent Overstreetcafe5632013-03-23 16:11:31 -07002193{
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002194 struct btree_insert_op *op = container_of(b_op,
2195 struct btree_insert_op, op);
Kent Overstreet403b6cd2013-07-24 17:22:44 -07002196
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002197 int ret = bch_btree_insert_node(b, &op->op, op->keys,
2198 op->journal_ref, op->replace_key);
2199 if (ret && !bch_keylist_empty(op->keys))
2200 return ret;
2201 else
2202 return MAP_DONE;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002203}
2204
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002205int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2206 atomic_t *journal_ref, struct bkey *replace_key)
Kent Overstreetcafe5632013-03-23 16:11:31 -07002207{
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002208 struct btree_insert_op op;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002209 int ret = 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002210
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002211 BUG_ON(current->bio_list);
Kent Overstreet4f3d4012013-09-10 18:46:36 -07002212 BUG_ON(bch_keylist_empty(keys));
Kent Overstreetcafe5632013-03-23 16:11:31 -07002213
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002214 bch_btree_op_init(&op.op, 0);
2215 op.keys = keys;
2216 op.journal_ref = journal_ref;
2217 op.replace_key = replace_key;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002218
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002219 while (!ret && !bch_keylist_empty(keys)) {
2220 op.op.lock = 0;
2221 ret = bch_btree_map_leaf_nodes(&op.op, c,
2222 &START_KEY(keys->keys),
2223 btree_insert_fn);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002224 }
2225
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002226 if (ret) {
2227 struct bkey *k;
2228
2229 pr_err("error %i", ret);
2230
2231 while ((k = bch_keylist_pop(keys)))
Kent Overstreet3a3b6a42013-07-24 16:46:42 -07002232 bkey_put(c, k);
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002233 } else if (op.op.insert_collision)
2234 ret = -ESRCH;
Kent Overstreet6054c6d2013-07-24 18:06:22 -07002235
Kent Overstreetcafe5632013-03-23 16:11:31 -07002236 return ret;
2237}
2238
2239void bch_btree_set_root(struct btree *b)
2240{
2241 unsigned i;
Kent Overstreete49c7c32013-06-26 17:25:38 -07002242 struct closure cl;
2243
2244 closure_init_stack(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002245
Kent Overstreetc37511b2013-04-26 15:39:55 -07002246 trace_bcache_btree_set_root(b);
2247
Kent Overstreetcafe5632013-03-23 16:11:31 -07002248 BUG_ON(!b->written);
2249
2250 for (i = 0; i < KEY_PTRS(&b->key); i++)
2251 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2252
2253 mutex_lock(&b->c->bucket_lock);
2254 list_del_init(&b->list);
2255 mutex_unlock(&b->c->bucket_lock);
2256
2257 b->c->root = b;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002258
Kent Overstreete49c7c32013-06-26 17:25:38 -07002259 bch_journal_meta(b->c, &cl);
2260 closure_sync(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002261}
2262
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002263/* Map across nodes or keys */
2264
2265static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2266 struct bkey *from,
2267 btree_map_nodes_fn *fn, int flags)
2268{
2269 int ret = MAP_CONTINUE;
2270
2271 if (b->level) {
2272 struct bkey *k;
2273 struct btree_iter iter;
2274
Kent Overstreetc052dd92013-11-11 17:35:24 -08002275 bch_btree_iter_init(&b->keys, &iter, from);
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002276
Kent Overstreeta85e9682013-12-20 17:28:16 -08002277 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002278 bch_ptr_bad))) {
2279 ret = btree(map_nodes_recurse, k, b,
2280 op, from, fn, flags);
2281 from = NULL;
2282
2283 if (ret != MAP_CONTINUE)
2284 return ret;
2285 }
2286 }
2287
2288 if (!b->level || flags == MAP_ALL_NODES)
2289 ret = fn(op, b);
2290
2291 return ret;
2292}
2293
2294int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2295 struct bkey *from, btree_map_nodes_fn *fn, int flags)
2296{
Kent Overstreetb54d6932013-07-24 18:04:18 -07002297 return btree_root(map_nodes_recurse, c, op, from, fn, flags);
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002298}
2299
2300static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2301 struct bkey *from, btree_map_keys_fn *fn,
2302 int flags)
2303{
2304 int ret = MAP_CONTINUE;
2305 struct bkey *k;
2306 struct btree_iter iter;
2307
Kent Overstreetc052dd92013-11-11 17:35:24 -08002308 bch_btree_iter_init(&b->keys, &iter, from);
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002309
Kent Overstreeta85e9682013-12-20 17:28:16 -08002310 while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002311 ret = !b->level
2312 ? fn(op, b, k)
2313 : btree(map_keys_recurse, k, b, op, from, fn, flags);
2314 from = NULL;
2315
2316 if (ret != MAP_CONTINUE)
2317 return ret;
2318 }
2319
2320 if (!b->level && (flags & MAP_END_KEY))
2321 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2322 KEY_OFFSET(&b->key), 0));
2323
2324 return ret;
2325}
2326
2327int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2328 struct bkey *from, btree_map_keys_fn *fn, int flags)
2329{
Kent Overstreetb54d6932013-07-24 18:04:18 -07002330 return btree_root(map_keys_recurse, c, op, from, fn, flags);
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002331}
2332
Kent Overstreetcafe5632013-03-23 16:11:31 -07002333/* Keybuf code */
2334
2335static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2336{
2337 /* Overlapping keys compare equal */
2338 if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2339 return -1;
2340 if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2341 return 1;
2342 return 0;
2343}
2344
2345static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2346 struct keybuf_key *r)
2347{
2348 return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2349}
2350
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002351struct refill {
2352 struct btree_op op;
Kent Overstreet48a915a2013-10-31 15:43:22 -07002353 unsigned nr_found;
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002354 struct keybuf *buf;
2355 struct bkey *end;
2356 keybuf_pred_fn *pred;
2357};
2358
2359static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2360 struct bkey *k)
Kent Overstreetcafe5632013-03-23 16:11:31 -07002361{
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002362 struct refill *refill = container_of(op, struct refill, op);
2363 struct keybuf *buf = refill->buf;
2364 int ret = MAP_CONTINUE;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002365
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002366 if (bkey_cmp(k, refill->end) >= 0) {
2367 ret = MAP_DONE;
2368 goto out;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002369 }
2370
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002371 if (!KEY_SIZE(k)) /* end key */
2372 goto out;
2373
2374 if (refill->pred(buf, k)) {
2375 struct keybuf_key *w;
2376
2377 spin_lock(&buf->lock);
2378
2379 w = array_alloc(&buf->freelist);
2380 if (!w) {
2381 spin_unlock(&buf->lock);
2382 return MAP_DONE;
2383 }
2384
2385 w->private = NULL;
2386 bkey_copy(&w->key, k);
2387
2388 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2389 array_free(&buf->freelist, w);
Kent Overstreet48a915a2013-10-31 15:43:22 -07002390 else
2391 refill->nr_found++;
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002392
2393 if (array_freelist_empty(&buf->freelist))
2394 ret = MAP_DONE;
2395
2396 spin_unlock(&buf->lock);
2397 }
2398out:
2399 buf->last_scanned = *k;
2400 return ret;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002401}
2402
2403void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
Kent Overstreet72c27062013-06-05 06:24:39 -07002404 struct bkey *end, keybuf_pred_fn *pred)
Kent Overstreetcafe5632013-03-23 16:11:31 -07002405{
2406 struct bkey start = buf->last_scanned;
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002407 struct refill refill;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002408
2409 cond_resched();
2410
Kent Overstreetb54d6932013-07-24 18:04:18 -07002411 bch_btree_op_init(&refill.op, -1);
Kent Overstreet48a915a2013-10-31 15:43:22 -07002412 refill.nr_found = 0;
2413 refill.buf = buf;
2414 refill.end = end;
2415 refill.pred = pred;
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002416
2417 bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2418 refill_keybuf_fn, MAP_END_KEY);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002419
Kent Overstreet48a915a2013-10-31 15:43:22 -07002420 trace_bcache_keyscan(refill.nr_found,
2421 KEY_INODE(&start), KEY_OFFSET(&start),
2422 KEY_INODE(&buf->last_scanned),
2423 KEY_OFFSET(&buf->last_scanned));
Kent Overstreetcafe5632013-03-23 16:11:31 -07002424
2425 spin_lock(&buf->lock);
2426
2427 if (!RB_EMPTY_ROOT(&buf->keys)) {
2428 struct keybuf_key *w;
2429 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2430 buf->start = START_KEY(&w->key);
2431
2432 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2433 buf->end = w->key;
2434 } else {
2435 buf->start = MAX_KEY;
2436 buf->end = MAX_KEY;
2437 }
2438
2439 spin_unlock(&buf->lock);
2440}
2441
2442static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2443{
2444 rb_erase(&w->node, &buf->keys);
2445 array_free(&buf->freelist, w);
2446}
2447
2448void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2449{
2450 spin_lock(&buf->lock);
2451 __bch_keybuf_del(buf, w);
2452 spin_unlock(&buf->lock);
2453}
2454
2455bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2456 struct bkey *end)
2457{
2458 bool ret = false;
2459 struct keybuf_key *p, *w, s;
2460 s.key = *start;
2461
2462 if (bkey_cmp(end, &buf->start) <= 0 ||
2463 bkey_cmp(start, &buf->end) >= 0)
2464 return false;
2465
2466 spin_lock(&buf->lock);
2467 w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2468
2469 while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2470 p = w;
2471 w = RB_NEXT(w, node);
2472
2473 if (p->private)
2474 ret = true;
2475 else
2476 __bch_keybuf_del(buf, p);
2477 }
2478
2479 spin_unlock(&buf->lock);
2480 return ret;
2481}
2482
2483struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2484{
2485 struct keybuf_key *w;
2486 spin_lock(&buf->lock);
2487
2488 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2489
2490 while (w && w->private)
2491 w = RB_NEXT(w, node);
2492
2493 if (w)
2494 w->private = ERR_PTR(-EINTR);
2495
2496 spin_unlock(&buf->lock);
2497 return w;
2498}
2499
2500struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002501 struct keybuf *buf,
2502 struct bkey *end,
2503 keybuf_pred_fn *pred)
Kent Overstreetcafe5632013-03-23 16:11:31 -07002504{
2505 struct keybuf_key *ret;
2506
2507 while (1) {
2508 ret = bch_keybuf_next(buf);
2509 if (ret)
2510 break;
2511
2512 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2513 pr_debug("scan finished");
2514 break;
2515 }
2516
Kent Overstreet72c27062013-06-05 06:24:39 -07002517 bch_refill_keybuf(c, buf, end, pred);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002518 }
2519
2520 return ret;
2521}
2522
Kent Overstreet72c27062013-06-05 06:24:39 -07002523void bch_keybuf_init(struct keybuf *buf)
Kent Overstreetcafe5632013-03-23 16:11:31 -07002524{
Kent Overstreetcafe5632013-03-23 16:11:31 -07002525 buf->last_scanned = MAX_KEY;
2526 buf->keys = RB_ROOT;
2527
2528 spin_lock_init(&buf->lock);
2529 array_allocator_init(&buf->freelist);
2530}