blob: e538d45fa65a9c45ec502cebc7e0634d36a6a5d2 [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; \
Kent Overstreet0a63b662014-03-17 17:15:53 -0700120 struct btree *_child = bch_btree_node_get((b)->c, op, key, l, _w);\
Kent Overstreetdf8e8972013-07-24 17:37:59 -0700121 if (!IS_ERR(_child)) { \
122 _child->parent = (b); \
123 _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)) { \
145 _b->parent = NULL; \
146 _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__); \
147 } \
148 rw_unlock(_w, _b); \
Kent Overstreet0a63b662014-03-17 17:15:53 -0700149 bch_cannibalize_unlock(c); \
Kent Overstreet78365412013-12-17 01:29:34 -0800150 if (_r == -EINTR) \
151 schedule(); \
Kent Overstreetdf8e8972013-07-24 17:37:59 -0700152 } while (_r == -EINTR); \
153 \
Kent Overstreet0a63b662014-03-17 17:15:53 -0700154 finish_wait(&(c)->btree_cache_wait, &(op)->wait); \
Kent Overstreetdf8e8972013-07-24 17:37:59 -0700155 _r; \
156})
157
Kent Overstreeta85e9682013-12-20 17:28:16 -0800158static inline struct bset *write_block(struct btree *b)
159{
160 return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
161}
162
Kent Overstreet2a285682014-03-04 16:42:42 -0800163static void bch_btree_init_next(struct btree *b)
164{
165 /* If not a leaf node, always sort */
166 if (b->level && b->keys.nsets)
167 bch_btree_sort(&b->keys, &b->c->sort);
168 else
169 bch_btree_sort_lazy(&b->keys, &b->c->sort);
170
171 if (b->written < btree_blocks(b))
172 bch_bset_init_next(&b->keys, write_block(b),
173 bset_magic(&b->c->sb));
174
175}
176
Kent Overstreetcafe5632013-03-23 16:11:31 -0700177/* Btree key manipulation */
178
Kent Overstreet3a3b6a42013-07-24 16:46:42 -0700179void bkey_put(struct cache_set *c, struct bkey *k)
Kent Overstreete7c590e2013-09-10 18:39:16 -0700180{
181 unsigned i;
182
183 for (i = 0; i < KEY_PTRS(k); i++)
184 if (ptr_available(c, k, i))
185 atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
186}
187
Kent Overstreetcafe5632013-03-23 16:11:31 -0700188/* Btree IO */
189
190static uint64_t btree_csum_set(struct btree *b, struct bset *i)
191{
192 uint64_t crc = b->key.ptr[0];
Kent Overstreetfafff812013-12-17 21:56:21 -0800193 void *data = (void *) i + 8, *end = bset_bkey_last(i);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700194
Kent Overstreet169ef1c2013-03-28 12:50:55 -0600195 crc = bch_crc64_update(crc, data, end - data);
Kent Overstreetc19ed232013-03-26 13:49:02 -0700196 return crc ^ 0xffffffffffffffffULL;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700197}
198
Kent Overstreet78b77bf2013-12-17 22:49:08 -0800199void bch_btree_node_read_done(struct btree *b)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700200{
Kent Overstreetcafe5632013-03-23 16:11:31 -0700201 const char *err = "bad btree header";
Kent Overstreetee811282013-12-17 23:49:49 -0800202 struct bset *i = btree_bset_first(b);
Kent Overstreet57943512013-04-25 13:58:35 -0700203 struct btree_iter *iter;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700204
Kent Overstreetbcf090e2014-05-19 08:57:55 -0700205 iter = mempool_alloc(b->c->fill_iter, GFP_NOIO);
Kent Overstreet57943512013-04-25 13:58:35 -0700206 iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700207 iter->used = 0;
208
Kent Overstreet280481d2013-10-24 16:36:03 -0700209#ifdef CONFIG_BCACHE_DEBUG
Kent Overstreetc052dd92013-11-11 17:35:24 -0800210 iter->b = &b->keys;
Kent Overstreet280481d2013-10-24 16:36:03 -0700211#endif
212
Kent Overstreet57943512013-04-25 13:58:35 -0700213 if (!i->seq)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700214 goto err;
215
216 for (;
Kent Overstreeta85e9682013-12-20 17:28:16 -0800217 b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700218 i = write_block(b)) {
219 err = "unsupported bset version";
220 if (i->version > BCACHE_BSET_VERSION)
221 goto err;
222
223 err = "bad btree header";
Kent Overstreetee811282013-12-17 23:49:49 -0800224 if (b->written + set_blocks(i, block_bytes(b->c)) >
225 btree_blocks(b))
Kent Overstreetcafe5632013-03-23 16:11:31 -0700226 goto err;
227
228 err = "bad magic";
Kent Overstreet81ab4192013-10-31 15:46:42 -0700229 if (i->magic != bset_magic(&b->c->sb))
Kent Overstreetcafe5632013-03-23 16:11:31 -0700230 goto err;
231
232 err = "bad checksum";
233 switch (i->version) {
234 case 0:
235 if (i->csum != csum_set(i))
236 goto err;
237 break;
238 case BCACHE_BSET_VERSION:
239 if (i->csum != btree_csum_set(b, i))
240 goto err;
241 break;
242 }
243
244 err = "empty set";
Kent Overstreeta85e9682013-12-20 17:28:16 -0800245 if (i != b->keys.set[0].data && !i->keys)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700246 goto err;
247
Kent Overstreetfafff812013-12-17 21:56:21 -0800248 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700249
Kent Overstreetee811282013-12-17 23:49:49 -0800250 b->written += set_blocks(i, block_bytes(b->c));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700251 }
252
253 err = "corrupted btree";
254 for (i = write_block(b);
Kent Overstreeta85e9682013-12-20 17:28:16 -0800255 bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700256 i = ((void *) i) + block_bytes(b->c))
Kent Overstreeta85e9682013-12-20 17:28:16 -0800257 if (i->seq == b->keys.set[0].data->seq)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700258 goto err;
259
Kent Overstreeta85e9682013-12-20 17:28:16 -0800260 bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700261
Kent Overstreeta85e9682013-12-20 17:28:16 -0800262 i = b->keys.set[0].data;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700263 err = "short btree key";
Kent Overstreeta85e9682013-12-20 17:28:16 -0800264 if (b->keys.set[0].size &&
265 bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700266 goto err;
267
268 if (b->written < btree_blocks(b))
Kent Overstreeta85e9682013-12-20 17:28:16 -0800269 bch_bset_init_next(&b->keys, write_block(b),
270 bset_magic(&b->c->sb));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700271out:
Kent Overstreet57943512013-04-25 13:58:35 -0700272 mempool_free(iter, b->c->fill_iter);
273 return;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700274err:
275 set_btree_node_io_error(b);
Kent Overstreet88b9f8c2013-12-17 21:46:35 -0800276 bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
Kent Overstreetcafe5632013-03-23 16:11:31 -0700277 err, PTR_BUCKET_NR(b->c, &b->key, 0),
Kent Overstreet88b9f8c2013-12-17 21:46:35 -0800278 bset_block_offset(b, i), i->keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700279 goto out;
280}
281
Kent Overstreet57943512013-04-25 13:58:35 -0700282static void btree_node_read_endio(struct bio *bio, int error)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700283{
Kent Overstreet57943512013-04-25 13:58:35 -0700284 struct closure *cl = bio->bi_private;
285 closure_put(cl);
286}
Kent Overstreetcafe5632013-03-23 16:11:31 -0700287
Kent Overstreet78b77bf2013-12-17 22:49:08 -0800288static void bch_btree_node_read(struct btree *b)
Kent Overstreet57943512013-04-25 13:58:35 -0700289{
290 uint64_t start_time = local_clock();
291 struct closure cl;
292 struct bio *bio;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700293
Kent Overstreetc37511b2013-04-26 15:39:55 -0700294 trace_bcache_btree_read(b);
295
Kent Overstreet57943512013-04-25 13:58:35 -0700296 closure_init_stack(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700297
Kent Overstreet57943512013-04-25 13:58:35 -0700298 bio = bch_bbio_alloc(b->c);
299 bio->bi_rw = REQ_META|READ_SYNC;
Kent Overstreet4f024f32013-10-11 15:44:27 -0700300 bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
Kent Overstreet57943512013-04-25 13:58:35 -0700301 bio->bi_end_io = btree_node_read_endio;
302 bio->bi_private = &cl;
303
Kent Overstreeta85e9682013-12-20 17:28:16 -0800304 bch_bio_map(bio, b->keys.set[0].data);
Kent Overstreet57943512013-04-25 13:58:35 -0700305
Kent Overstreet57943512013-04-25 13:58:35 -0700306 bch_submit_bbio(bio, b->c, &b->key, 0);
307 closure_sync(&cl);
308
309 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
310 set_btree_node_io_error(b);
311
312 bch_bbio_free(bio, b->c);
313
314 if (btree_node_io_error(b))
315 goto err;
316
317 bch_btree_node_read_done(b);
Kent Overstreet57943512013-04-25 13:58:35 -0700318 bch_time_stats_update(&b->c->btree_read_time, start_time);
Kent Overstreet57943512013-04-25 13:58:35 -0700319
320 return;
321err:
Geert Uytterhoeven61cbd252013-09-23 23:17:30 -0700322 bch_cache_set_error(b->c, "io error reading bucket %zu",
Kent Overstreet57943512013-04-25 13:58:35 -0700323 PTR_BUCKET_NR(b->c, &b->key, 0));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700324}
325
326static void btree_complete_write(struct btree *b, struct btree_write *w)
327{
328 if (w->prio_blocked &&
329 !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
Kent Overstreet119ba0f2013-04-24 19:01:12 -0700330 wake_up_allocators(b->c);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700331
332 if (w->journal) {
333 atomic_dec_bug(w->journal);
334 __closure_wake_up(&b->c->journal.wait);
335 }
336
Kent Overstreetcafe5632013-03-23 16:11:31 -0700337 w->prio_blocked = 0;
338 w->journal = NULL;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700339}
340
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800341static void btree_node_write_unlock(struct closure *cl)
342{
343 struct btree *b = container_of(cl, struct btree, io);
344
345 up(&b->io_mutex);
346}
347
Kent Overstreet57943512013-04-25 13:58:35 -0700348static void __btree_node_write_done(struct closure *cl)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700349{
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800350 struct btree *b = container_of(cl, struct btree, io);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700351 struct btree_write *w = btree_prev_write(b);
352
353 bch_bbio_free(b->bio, b->c);
354 b->bio = NULL;
355 btree_complete_write(b, w);
356
357 if (btree_node_dirty(b))
Kent Overstreet56b30772014-01-23 01:44:55 -0800358 schedule_delayed_work(&b->work, 30 * HZ);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700359
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800360 closure_return_with_destructor(cl, btree_node_write_unlock);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700361}
362
Kent Overstreet57943512013-04-25 13:58:35 -0700363static void btree_node_write_done(struct closure *cl)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700364{
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800365 struct btree *b = container_of(cl, struct btree, io);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700366 struct bio_vec *bv;
367 int n;
368
Kent Overstreet79886132013-11-23 17:19:00 -0800369 bio_for_each_segment_all(bv, b->bio, n)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700370 __free_page(bv->bv_page);
371
Kent Overstreet57943512013-04-25 13:58:35 -0700372 __btree_node_write_done(cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700373}
374
Kent Overstreet57943512013-04-25 13:58:35 -0700375static void btree_node_write_endio(struct bio *bio, int error)
376{
377 struct closure *cl = bio->bi_private;
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800378 struct btree *b = container_of(cl, struct btree, io);
Kent Overstreet57943512013-04-25 13:58:35 -0700379
380 if (error)
381 set_btree_node_io_error(b);
382
383 bch_bbio_count_io_errors(b->c, bio, error, "writing btree");
384 closure_put(cl);
385}
386
387static void do_btree_node_write(struct btree *b)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700388{
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800389 struct closure *cl = &b->io;
Kent Overstreetee811282013-12-17 23:49:49 -0800390 struct bset *i = btree_bset_last(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700391 BKEY_PADDED(key) k;
392
393 i->version = BCACHE_BSET_VERSION;
394 i->csum = btree_csum_set(b, i);
395
Kent Overstreet57943512013-04-25 13:58:35 -0700396 BUG_ON(b->bio);
397 b->bio = bch_bbio_alloc(b->c);
398
399 b->bio->bi_end_io = btree_node_write_endio;
Kent Overstreetfaadf0c2013-11-01 18:03:08 -0700400 b->bio->bi_private = cl;
Kent Overstreete49c7c32013-06-26 17:25:38 -0700401 b->bio->bi_rw = REQ_META|WRITE_SYNC|REQ_FUA;
Kent Overstreetee811282013-12-17 23:49:49 -0800402 b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
Kent Overstreet169ef1c2013-03-28 12:50:55 -0600403 bch_bio_map(b->bio, i);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700404
Kent Overstreete49c7c32013-06-26 17:25:38 -0700405 /*
406 * If we're appending to a leaf node, we don't technically need FUA -
407 * this write just needs to be persisted before the next journal write,
408 * which will be marked FLUSH|FUA.
409 *
410 * Similarly if we're writing a new btree root - the pointer is going to
411 * be in the next journal entry.
412 *
413 * But if we're writing a new btree node (that isn't a root) or
414 * appending to a non leaf btree node, we need either FUA or a flush
415 * when we write the parent with the new pointer. FUA is cheaper than a
416 * flush, and writes appending to leaf nodes aren't blocking anything so
417 * just make all btree node writes FUA to keep things sane.
418 */
419
Kent Overstreetcafe5632013-03-23 16:11:31 -0700420 bkey_copy(&k.key, &b->key);
Kent Overstreetee811282013-12-17 23:49:49 -0800421 SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
Kent Overstreeta85e9682013-12-20 17:28:16 -0800422 bset_sector_offset(&b->keys, i));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700423
Kent Overstreet8e51e412013-06-06 18:15:57 -0700424 if (!bio_alloc_pages(b->bio, GFP_NOIO)) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700425 int j;
426 struct bio_vec *bv;
427 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
428
Kent Overstreet79886132013-11-23 17:19:00 -0800429 bio_for_each_segment_all(bv, b->bio, j)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700430 memcpy(page_address(bv->bv_page),
431 base + j * PAGE_SIZE, PAGE_SIZE);
432
Kent Overstreetcafe5632013-03-23 16:11:31 -0700433 bch_submit_bbio(b->bio, b->c, &k.key, 0);
434
Kent Overstreet57943512013-04-25 13:58:35 -0700435 continue_at(cl, btree_node_write_done, NULL);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700436 } else {
437 b->bio->bi_vcnt = 0;
Kent Overstreet169ef1c2013-03-28 12:50:55 -0600438 bch_bio_map(b->bio, i);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700439
Kent Overstreetcafe5632013-03-23 16:11:31 -0700440 bch_submit_bbio(b->bio, b->c, &k.key, 0);
441
442 closure_sync(cl);
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800443 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700444 }
445}
446
Kent Overstreet2a285682014-03-04 16:42:42 -0800447void __bch_btree_node_write(struct btree *b, struct closure *parent)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700448{
Kent Overstreetee811282013-12-17 23:49:49 -0800449 struct bset *i = btree_bset_last(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700450
Kent Overstreet2a285682014-03-04 16:42:42 -0800451 lockdep_assert_held(&b->write_lock);
452
Kent Overstreetc37511b2013-04-26 15:39:55 -0700453 trace_bcache_btree_write(b);
454
Kent Overstreetcafe5632013-03-23 16:11:31 -0700455 BUG_ON(current->bio_list);
Kent Overstreet57943512013-04-25 13:58:35 -0700456 BUG_ON(b->written >= btree_blocks(b));
457 BUG_ON(b->written && !i->keys);
Kent Overstreetee811282013-12-17 23:49:49 -0800458 BUG_ON(btree_bset_first(b)->seq != i->seq);
Kent Overstreetdc9d98d2013-12-17 23:47:33 -0800459 bch_check_keys(&b->keys, "writing");
Kent Overstreetcafe5632013-03-23 16:11:31 -0700460
Kent Overstreetcafe5632013-03-23 16:11:31 -0700461 cancel_delayed_work(&b->work);
462
Kent Overstreet57943512013-04-25 13:58:35 -0700463 /* If caller isn't waiting for write, parent refcount is cache set */
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800464 down(&b->io_mutex);
465 closure_init(&b->io, parent ?: &b->c->cl);
Kent Overstreet57943512013-04-25 13:58:35 -0700466
Kent Overstreetcafe5632013-03-23 16:11:31 -0700467 clear_bit(BTREE_NODE_dirty, &b->flags);
468 change_bit(BTREE_NODE_write_idx, &b->flags);
469
Kent Overstreet57943512013-04-25 13:58:35 -0700470 do_btree_node_write(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700471
Kent Overstreetee811282013-12-17 23:49:49 -0800472 atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
Kent Overstreetcafe5632013-03-23 16:11:31 -0700473 &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
474
Kent Overstreeta85e9682013-12-20 17:28:16 -0800475 b->written += set_blocks(i, block_bytes(b->c));
Kent Overstreet2a285682014-03-04 16:42:42 -0800476}
Kent Overstreeta85e9682013-12-20 17:28:16 -0800477
Kent Overstreet2a285682014-03-04 16:42:42 -0800478void bch_btree_node_write(struct btree *b, struct closure *parent)
479{
480 unsigned nsets = b->keys.nsets;
481
482 lockdep_assert_held(&b->lock);
483
484 __bch_btree_node_write(b, parent);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700485
Kent Overstreet78b77bf2013-12-17 22:49:08 -0800486 /*
487 * do verify if there was more than one set initially (i.e. we did a
488 * sort) and we sorted down to a single set:
489 */
Kent Overstreet2a285682014-03-04 16:42:42 -0800490 if (nsets && !b->keys.nsets)
Kent Overstreet78b77bf2013-12-17 22:49:08 -0800491 bch_btree_verify(b);
492
Kent Overstreet2a285682014-03-04 16:42:42 -0800493 bch_btree_init_next(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700494}
495
Kent Overstreetf269af52013-07-23 20:48:29 -0700496static void bch_btree_node_write_sync(struct btree *b)
497{
498 struct closure cl;
499
500 closure_init_stack(&cl);
Kent Overstreet2a285682014-03-04 16:42:42 -0800501
502 mutex_lock(&b->write_lock);
Kent Overstreetf269af52013-07-23 20:48:29 -0700503 bch_btree_node_write(b, &cl);
Kent Overstreet2a285682014-03-04 16:42:42 -0800504 mutex_unlock(&b->write_lock);
505
Kent Overstreetf269af52013-07-23 20:48:29 -0700506 closure_sync(&cl);
507}
508
Kent Overstreet57943512013-04-25 13:58:35 -0700509static void btree_node_write_work(struct work_struct *w)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700510{
511 struct btree *b = container_of(to_delayed_work(w), struct btree, work);
512
Kent Overstreet2a285682014-03-04 16:42:42 -0800513 mutex_lock(&b->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700514 if (btree_node_dirty(b))
Kent Overstreet2a285682014-03-04 16:42:42 -0800515 __bch_btree_node_write(b, NULL);
516 mutex_unlock(&b->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700517}
518
Kent Overstreetc18536a2013-07-24 17:44:17 -0700519static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700520{
Kent Overstreetee811282013-12-17 23:49:49 -0800521 struct bset *i = btree_bset_last(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700522 struct btree_write *w = btree_current_write(b);
523
Kent Overstreet2a285682014-03-04 16:42:42 -0800524 lockdep_assert_held(&b->write_lock);
525
Kent Overstreet57943512013-04-25 13:58:35 -0700526 BUG_ON(!b->written);
527 BUG_ON(!i->keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700528
Kent Overstreet57943512013-04-25 13:58:35 -0700529 if (!btree_node_dirty(b))
Kent Overstreet56b30772014-01-23 01:44:55 -0800530 schedule_delayed_work(&b->work, 30 * HZ);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700531
Kent Overstreet57943512013-04-25 13:58:35 -0700532 set_btree_node_dirty(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700533
Kent Overstreetc18536a2013-07-24 17:44:17 -0700534 if (journal_ref) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700535 if (w->journal &&
Kent Overstreetc18536a2013-07-24 17:44:17 -0700536 journal_pin_cmp(b->c, w->journal, journal_ref)) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700537 atomic_dec_bug(w->journal);
538 w->journal = NULL;
539 }
540
541 if (!w->journal) {
Kent Overstreetc18536a2013-07-24 17:44:17 -0700542 w->journal = journal_ref;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700543 atomic_inc(w->journal);
544 }
545 }
546
Kent Overstreetcafe5632013-03-23 16:11:31 -0700547 /* Force write if set is too big */
Kent Overstreet57943512013-04-25 13:58:35 -0700548 if (set_bytes(i) > PAGE_SIZE - 48 &&
549 !current->bio_list)
550 bch_btree_node_write(b, NULL);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700551}
552
553/*
554 * Btree in memory cache - allocation/freeing
555 * mca -> memory cache
556 */
557
Kent Overstreetcafe5632013-03-23 16:11:31 -0700558#define mca_reserve(c) (((c->root && c->root->level) \
559 ? c->root->level : 1) * 8 + 16)
560#define mca_can_free(c) \
Kent Overstreet0a63b662014-03-17 17:15:53 -0700561 max_t(int, 0, c->btree_cache_used - mca_reserve(c))
Kent Overstreetcafe5632013-03-23 16:11:31 -0700562
563static void mca_data_free(struct btree *b)
564{
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800565 BUG_ON(b->io_mutex.count != 1);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700566
Kent Overstreeta85e9682013-12-20 17:28:16 -0800567 bch_btree_keys_free(&b->keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700568
Kent Overstreet0a63b662014-03-17 17:15:53 -0700569 b->c->btree_cache_used--;
Kent Overstreetee811282013-12-17 23:49:49 -0800570 list_move(&b->list, &b->c->btree_cache_freed);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700571}
572
573static void mca_bucket_free(struct btree *b)
574{
575 BUG_ON(btree_node_dirty(b));
576
577 b->key.ptr[0] = 0;
578 hlist_del_init_rcu(&b->hash);
579 list_move(&b->list, &b->c->btree_cache_freeable);
580}
581
582static unsigned btree_order(struct bkey *k)
583{
584 return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
585}
586
587static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
588{
Kent Overstreeta85e9682013-12-20 17:28:16 -0800589 if (!bch_btree_keys_alloc(&b->keys,
Kent Overstreetee811282013-12-17 23:49:49 -0800590 max_t(unsigned,
591 ilog2(b->c->btree_pages),
592 btree_order(k)),
593 gfp)) {
Kent Overstreet0a63b662014-03-17 17:15:53 -0700594 b->c->btree_cache_used++;
Kent Overstreetee811282013-12-17 23:49:49 -0800595 list_move(&b->list, &b->c->btree_cache);
596 } else {
597 list_move(&b->list, &b->c->btree_cache_freed);
598 }
Kent Overstreetcafe5632013-03-23 16:11:31 -0700599}
600
601static struct btree *mca_bucket_alloc(struct cache_set *c,
602 struct bkey *k, gfp_t gfp)
603{
604 struct btree *b = kzalloc(sizeof(struct btree), gfp);
605 if (!b)
606 return NULL;
607
608 init_rwsem(&b->lock);
609 lockdep_set_novalidate_class(&b->lock);
Kent Overstreet2a285682014-03-04 16:42:42 -0800610 mutex_init(&b->write_lock);
611 lockdep_set_novalidate_class(&b->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700612 INIT_LIST_HEAD(&b->list);
Kent Overstreet57943512013-04-25 13:58:35 -0700613 INIT_DELAYED_WORK(&b->work, btree_node_write_work);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700614 b->c = c;
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800615 sema_init(&b->io_mutex, 1);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700616
617 mca_data_alloc(b, k, gfp);
618 return b;
619}
620
Kent Overstreete8e1d462013-07-24 17:27:07 -0700621static int mca_reap(struct btree *b, unsigned min_order, bool flush)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700622{
Kent Overstreete8e1d462013-07-24 17:27:07 -0700623 struct closure cl;
624
625 closure_init_stack(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700626 lockdep_assert_held(&b->c->bucket_lock);
627
628 if (!down_write_trylock(&b->lock))
629 return -ENOMEM;
630
Kent Overstreeta85e9682013-12-20 17:28:16 -0800631 BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
Kent Overstreete8e1d462013-07-24 17:27:07 -0700632
Kent Overstreeta85e9682013-12-20 17:28:16 -0800633 if (b->keys.page_order < min_order)
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800634 goto out_unlock;
635
636 if (!flush) {
637 if (btree_node_dirty(b))
638 goto out_unlock;
639
640 if (down_trylock(&b->io_mutex))
641 goto out_unlock;
642 up(&b->io_mutex);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700643 }
644
Kent Overstreet2a285682014-03-04 16:42:42 -0800645 mutex_lock(&b->write_lock);
Kent Overstreetf269af52013-07-23 20:48:29 -0700646 if (btree_node_dirty(b))
Kent Overstreet2a285682014-03-04 16:42:42 -0800647 __bch_btree_node_write(b, &cl);
648 mutex_unlock(&b->write_lock);
649
650 closure_sync(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700651
Kent Overstreete8e1d462013-07-24 17:27:07 -0700652 /* wait for any in flight btree write */
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800653 down(&b->io_mutex);
654 up(&b->io_mutex);
Kent Overstreete8e1d462013-07-24 17:27:07 -0700655
Kent Overstreetcafe5632013-03-23 16:11:31 -0700656 return 0;
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800657out_unlock:
658 rw_unlock(true, b);
659 return -ENOMEM;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700660}
661
Dave Chinner7dc19d52013-08-28 10:18:11 +1000662static unsigned long bch_mca_scan(struct shrinker *shrink,
663 struct shrink_control *sc)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700664{
665 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
666 struct btree *b, *t;
667 unsigned long i, nr = sc->nr_to_scan;
Dave Chinner7dc19d52013-08-28 10:18:11 +1000668 unsigned long freed = 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700669
670 if (c->shrinker_disabled)
Dave Chinner7dc19d52013-08-28 10:18:11 +1000671 return SHRINK_STOP;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700672
Kent Overstreet0a63b662014-03-17 17:15:53 -0700673 if (c->btree_cache_alloc_lock)
Dave Chinner7dc19d52013-08-28 10:18:11 +1000674 return SHRINK_STOP;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700675
676 /* Return -1 if we can't do anything right now */
Kent Overstreeta698e082013-09-23 23:17:34 -0700677 if (sc->gfp_mask & __GFP_IO)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700678 mutex_lock(&c->bucket_lock);
679 else if (!mutex_trylock(&c->bucket_lock))
680 return -1;
681
Kent Overstreet36c9ea92013-06-03 13:04:56 -0700682 /*
683 * It's _really_ critical that we don't free too many btree nodes - we
684 * have to always leave ourselves a reserve. The reserve is how we
685 * guarantee that allocating memory for a new btree node can always
686 * succeed, so that inserting keys into the btree can always succeed and
687 * IO can always make forward progress:
688 */
Kent Overstreetcafe5632013-03-23 16:11:31 -0700689 nr /= c->btree_pages;
690 nr = min_t(unsigned long, nr, mca_can_free(c));
691
692 i = 0;
693 list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
Dave Chinner7dc19d52013-08-28 10:18:11 +1000694 if (freed >= nr)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700695 break;
696
697 if (++i > 3 &&
Kent Overstreete8e1d462013-07-24 17:27:07 -0700698 !mca_reap(b, 0, false)) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700699 mca_data_free(b);
700 rw_unlock(true, b);
Dave Chinner7dc19d52013-08-28 10:18:11 +1000701 freed++;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700702 }
703 }
704
Kent Overstreet0a63b662014-03-17 17:15:53 -0700705 for (i = 0; (nr--) && i < c->btree_cache_used; i++) {
Kent Overstreetb0f32a52013-12-10 13:24:26 -0800706 if (list_empty(&c->btree_cache))
707 goto out;
708
Kent Overstreetcafe5632013-03-23 16:11:31 -0700709 b = list_first_entry(&c->btree_cache, struct btree, list);
710 list_rotate_left(&c->btree_cache);
711
712 if (!b->accessed &&
Kent Overstreete8e1d462013-07-24 17:27:07 -0700713 !mca_reap(b, 0, false)) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700714 mca_bucket_free(b);
715 mca_data_free(b);
716 rw_unlock(true, b);
Dave Chinner7dc19d52013-08-28 10:18:11 +1000717 freed++;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700718 } else
719 b->accessed = 0;
720 }
721out:
Kent Overstreetcafe5632013-03-23 16:11:31 -0700722 mutex_unlock(&c->bucket_lock);
Dave Chinner7dc19d52013-08-28 10:18:11 +1000723 return freed;
724}
725
726static unsigned long bch_mca_count(struct shrinker *shrink,
727 struct shrink_control *sc)
728{
729 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
730
731 if (c->shrinker_disabled)
732 return 0;
733
Kent Overstreet0a63b662014-03-17 17:15:53 -0700734 if (c->btree_cache_alloc_lock)
Dave Chinner7dc19d52013-08-28 10:18:11 +1000735 return 0;
736
737 return mca_can_free(c) * c->btree_pages;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700738}
739
740void bch_btree_cache_free(struct cache_set *c)
741{
742 struct btree *b;
743 struct closure cl;
744 closure_init_stack(&cl);
745
746 if (c->shrink.list.next)
747 unregister_shrinker(&c->shrink);
748
749 mutex_lock(&c->bucket_lock);
750
751#ifdef CONFIG_BCACHE_DEBUG
752 if (c->verify_data)
753 list_move(&c->verify_data->list, &c->btree_cache);
Kent Overstreet78b77bf2013-12-17 22:49:08 -0800754
755 free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700756#endif
757
758 list_splice(&c->btree_cache_freeable,
759 &c->btree_cache);
760
761 while (!list_empty(&c->btree_cache)) {
762 b = list_first_entry(&c->btree_cache, struct btree, list);
763
764 if (btree_node_dirty(b))
765 btree_complete_write(b, btree_current_write(b));
766 clear_bit(BTREE_NODE_dirty, &b->flags);
767
768 mca_data_free(b);
769 }
770
771 while (!list_empty(&c->btree_cache_freed)) {
772 b = list_first_entry(&c->btree_cache_freed,
773 struct btree, list);
774 list_del(&b->list);
775 cancel_delayed_work_sync(&b->work);
776 kfree(b);
777 }
778
779 mutex_unlock(&c->bucket_lock);
780}
781
782int bch_btree_cache_alloc(struct cache_set *c)
783{
784 unsigned i;
785
Kent Overstreetcafe5632013-03-23 16:11:31 -0700786 for (i = 0; i < mca_reserve(c); i++)
Kent Overstreet72a44512013-10-24 17:19:26 -0700787 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
788 return -ENOMEM;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700789
790 list_splice_init(&c->btree_cache,
791 &c->btree_cache_freeable);
792
793#ifdef CONFIG_BCACHE_DEBUG
794 mutex_init(&c->verify_lock);
795
Kent Overstreet78b77bf2013-12-17 22:49:08 -0800796 c->verify_ondisk = (void *)
797 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
798
Kent Overstreetcafe5632013-03-23 16:11:31 -0700799 c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
800
801 if (c->verify_data &&
Kent Overstreeta85e9682013-12-20 17:28:16 -0800802 c->verify_data->keys.set->data)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700803 list_del_init(&c->verify_data->list);
804 else
805 c->verify_data = NULL;
806#endif
807
Dave Chinner7dc19d52013-08-28 10:18:11 +1000808 c->shrink.count_objects = bch_mca_count;
809 c->shrink.scan_objects = bch_mca_scan;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700810 c->shrink.seeks = 4;
811 c->shrink.batch = c->btree_pages * 2;
812 register_shrinker(&c->shrink);
813
814 return 0;
815}
816
817/* Btree in memory cache - hash table */
818
819static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
820{
821 return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
822}
823
824static struct btree *mca_find(struct cache_set *c, struct bkey *k)
825{
826 struct btree *b;
827
828 rcu_read_lock();
829 hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
830 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
831 goto out;
832 b = NULL;
833out:
834 rcu_read_unlock();
835 return b;
836}
837
Kent Overstreet0a63b662014-03-17 17:15:53 -0700838static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
839{
840 struct task_struct *old;
841
842 old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
843 if (old && old != current) {
844 if (op)
845 prepare_to_wait(&c->btree_cache_wait, &op->wait,
846 TASK_UNINTERRUPTIBLE);
847 return -EINTR;
848 }
849
850 return 0;
851}
852
853static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
854 struct bkey *k)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700855{
Kent Overstreete8e1d462013-07-24 17:27:07 -0700856 struct btree *b;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700857
Kent Overstreetc37511b2013-04-26 15:39:55 -0700858 trace_bcache_btree_cache_cannibalize(c);
859
Kent Overstreet0a63b662014-03-17 17:15:53 -0700860 if (mca_cannibalize_lock(c, op))
861 return ERR_PTR(-EINTR);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700862
Kent Overstreete8e1d462013-07-24 17:27:07 -0700863 list_for_each_entry_reverse(b, &c->btree_cache, list)
864 if (!mca_reap(b, btree_order(k), false))
865 return b;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700866
Kent Overstreete8e1d462013-07-24 17:27:07 -0700867 list_for_each_entry_reverse(b, &c->btree_cache, list)
868 if (!mca_reap(b, btree_order(k), true))
869 return b;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700870
Kent Overstreet0a63b662014-03-17 17:15:53 -0700871 WARN(1, "btree cache cannibalize failed\n");
Kent Overstreete8e1d462013-07-24 17:27:07 -0700872 return ERR_PTR(-ENOMEM);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700873}
874
875/*
876 * We can only have one thread cannibalizing other cached btree nodes at a time,
877 * or we'll deadlock. We use an open coded mutex to ensure that, which a
878 * cannibalize_bucket() will take. This means every time we unlock the root of
879 * the btree, we need to release this lock if we have it held.
880 */
Kent Overstreetdf8e8972013-07-24 17:37:59 -0700881static void bch_cannibalize_unlock(struct cache_set *c)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700882{
Kent Overstreet0a63b662014-03-17 17:15:53 -0700883 if (c->btree_cache_alloc_lock == current) {
884 c->btree_cache_alloc_lock = NULL;
885 wake_up(&c->btree_cache_wait);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700886 }
887}
888
Kent Overstreet0a63b662014-03-17 17:15:53 -0700889static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
890 struct bkey *k, int level)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700891{
892 struct btree *b;
893
Kent Overstreete8e1d462013-07-24 17:27:07 -0700894 BUG_ON(current->bio_list);
895
Kent Overstreetcafe5632013-03-23 16:11:31 -0700896 lockdep_assert_held(&c->bucket_lock);
897
898 if (mca_find(c, k))
899 return NULL;
900
901 /* btree_free() doesn't free memory; it sticks the node on the end of
902 * the list. Check if there's any freed nodes there:
903 */
904 list_for_each_entry(b, &c->btree_cache_freeable, list)
Kent Overstreete8e1d462013-07-24 17:27:07 -0700905 if (!mca_reap(b, btree_order(k), false))
Kent Overstreetcafe5632013-03-23 16:11:31 -0700906 goto out;
907
908 /* We never free struct btree itself, just the memory that holds the on
909 * disk node. Check the freed list before allocating a new one:
910 */
911 list_for_each_entry(b, &c->btree_cache_freed, list)
Kent Overstreete8e1d462013-07-24 17:27:07 -0700912 if (!mca_reap(b, 0, false)) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700913 mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
Kent Overstreeta85e9682013-12-20 17:28:16 -0800914 if (!b->keys.set[0].data)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700915 goto err;
916 else
917 goto out;
918 }
919
920 b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
921 if (!b)
922 goto err;
923
924 BUG_ON(!down_write_trylock(&b->lock));
Kent Overstreeta85e9682013-12-20 17:28:16 -0800925 if (!b->keys.set->data)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700926 goto err;
927out:
Kent Overstreetcb7a5832013-12-16 15:27:25 -0800928 BUG_ON(b->io_mutex.count != 1);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700929
930 bkey_copy(&b->key, k);
931 list_move(&b->list, &c->btree_cache);
932 hlist_del_init_rcu(&b->hash);
933 hlist_add_head_rcu(&b->hash, mca_hash(c, k));
934
935 lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
Kent Overstreetd6fd3b12013-07-24 17:20:19 -0700936 b->parent = (void *) ~0UL;
Kent Overstreeta85e9682013-12-20 17:28:16 -0800937 b->flags = 0;
938 b->written = 0;
939 b->level = level;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700940
Kent Overstreet65d45232013-12-20 17:22:05 -0800941 if (!b->level)
Kent Overstreeta85e9682013-12-20 17:28:16 -0800942 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
943 &b->c->expensive_debug_checks);
Kent Overstreet65d45232013-12-20 17:22:05 -0800944 else
Kent Overstreeta85e9682013-12-20 17:28:16 -0800945 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
946 &b->c->expensive_debug_checks);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700947
948 return b;
949err:
950 if (b)
951 rw_unlock(true, b);
952
Kent Overstreet0a63b662014-03-17 17:15:53 -0700953 b = mca_cannibalize(c, op, k);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700954 if (!IS_ERR(b))
955 goto out;
956
957 return b;
958}
959
960/**
961 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
962 * in from disk if necessary.
963 *
Kent Overstreetb54d6932013-07-24 18:04:18 -0700964 * If IO is necessary and running under generic_make_request, returns -EAGAIN.
Kent Overstreetcafe5632013-03-23 16:11:31 -0700965 *
966 * The btree node will have either a read or a write lock held, depending on
967 * level and op->lock.
968 */
Kent Overstreet0a63b662014-03-17 17:15:53 -0700969struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
970 struct bkey *k, int level, bool write)
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
1005 b->accessed = 1;
1006
Kent Overstreeta85e9682013-12-20 17:28:16 -08001007 for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1008 prefetch(b->keys.set[i].tree);
1009 prefetch(b->keys.set[i].data);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001010 }
1011
Kent Overstreeta85e9682013-12-20 17:28:16 -08001012 for (; i <= b->keys.nsets; i++)
1013 prefetch(b->keys.set[i].data);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001014
Kent Overstreet57943512013-04-25 13:58:35 -07001015 if (btree_node_io_error(b)) {
Kent Overstreetcafe5632013-03-23 16:11:31 -07001016 rw_unlock(write, b);
Kent Overstreet57943512013-04-25 13:58:35 -07001017 return ERR_PTR(-EIO);
1018 }
1019
1020 BUG_ON(!b->written);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001021
1022 return b;
1023}
1024
1025static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level)
1026{
1027 struct btree *b;
1028
1029 mutex_lock(&c->bucket_lock);
Kent Overstreet0a63b662014-03-17 17:15:53 -07001030 b = mca_alloc(c, NULL, k, level);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001031 mutex_unlock(&c->bucket_lock);
1032
1033 if (!IS_ERR_OR_NULL(b)) {
Kent Overstreet57943512013-04-25 13:58:35 -07001034 bch_btree_node_read(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001035 rw_unlock(true, b);
1036 }
1037}
1038
1039/* Btree alloc */
1040
Kent Overstreete8e1d462013-07-24 17:27:07 -07001041static void btree_node_free(struct btree *b)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001042{
Kent Overstreetc37511b2013-04-26 15:39:55 -07001043 trace_bcache_btree_node_free(b);
1044
Kent Overstreetcafe5632013-03-23 16:11:31 -07001045 BUG_ON(b == b->c->root);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001046
Kent Overstreet2a285682014-03-04 16:42:42 -08001047 mutex_lock(&b->write_lock);
1048
Kent Overstreetcafe5632013-03-23 16:11:31 -07001049 if (btree_node_dirty(b))
1050 btree_complete_write(b, btree_current_write(b));
1051 clear_bit(BTREE_NODE_dirty, &b->flags);
1052
Kent Overstreet2a285682014-03-04 16:42:42 -08001053 mutex_unlock(&b->write_lock);
1054
Kent Overstreetcafe5632013-03-23 16:11:31 -07001055 cancel_delayed_work(&b->work);
1056
1057 mutex_lock(&b->c->bucket_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001058 bch_bucket_free(b->c, &b->key);
1059 mca_bucket_free(b);
1060 mutex_unlock(&b->c->bucket_lock);
1061}
1062
Slava Pestovc5aa4a32014-04-21 18:23:12 -07001063struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1064 int level, bool wait)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001065{
1066 BKEY_PADDED(key) k;
1067 struct btree *b = ERR_PTR(-EAGAIN);
1068
1069 mutex_lock(&c->bucket_lock);
1070retry:
Slava Pestovc5aa4a32014-04-21 18:23:12 -07001071 if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
Kent Overstreetcafe5632013-03-23 16:11:31 -07001072 goto err;
1073
Kent Overstreet3a3b6a42013-07-24 16:46:42 -07001074 bkey_put(c, &k.key);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001075 SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1076
Kent Overstreet0a63b662014-03-17 17:15:53 -07001077 b = mca_alloc(c, op, &k.key, level);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001078 if (IS_ERR(b))
1079 goto err_free;
1080
1081 if (!b) {
Kent Overstreetb1a67b02013-03-25 11:46:44 -07001082 cache_bug(c,
1083 "Tried to allocate bucket that was in btree cache");
Kent Overstreetcafe5632013-03-23 16:11:31 -07001084 goto retry;
1085 }
1086
Kent Overstreetcafe5632013-03-23 16:11:31 -07001087 b->accessed = 1;
Kent Overstreeta85e9682013-12-20 17:28:16 -08001088 bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
Kent Overstreetcafe5632013-03-23 16:11:31 -07001089
1090 mutex_unlock(&c->bucket_lock);
Kent Overstreetc37511b2013-04-26 15:39:55 -07001091
1092 trace_bcache_btree_node_alloc(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001093 return b;
1094err_free:
1095 bch_bucket_free(c, &k.key);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001096err:
1097 mutex_unlock(&c->bucket_lock);
Kent Overstreetc37511b2013-04-26 15:39:55 -07001098
1099 trace_bcache_btree_node_alloc_fail(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001100 return b;
1101}
1102
Slava Pestovc5aa4a32014-04-21 18:23:12 -07001103static struct btree *bch_btree_node_alloc(struct cache_set *c,
1104 struct btree_op *op, int level)
1105{
1106 return __bch_btree_node_alloc(c, op, level, op != NULL);
1107}
1108
Kent Overstreet0a63b662014-03-17 17:15:53 -07001109static struct btree *btree_node_alloc_replacement(struct btree *b,
1110 struct btree_op *op)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001111{
Kent Overstreet0a63b662014-03-17 17:15:53 -07001112 struct btree *n = bch_btree_node_alloc(b->c, op, b->level);
Kent Overstreet67539e82013-09-10 22:53:34 -07001113 if (!IS_ERR_OR_NULL(n)) {
Kent Overstreet2a285682014-03-04 16:42:42 -08001114 mutex_lock(&n->write_lock);
Kent Overstreet89ebb4a2013-11-11 18:38:51 -08001115 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
Kent Overstreet67539e82013-09-10 22:53:34 -07001116 bkey_copy_key(&n->key, &b->key);
Kent Overstreet2a285682014-03-04 16:42:42 -08001117 mutex_unlock(&n->write_lock);
Kent Overstreet67539e82013-09-10 22:53:34 -07001118 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001119
1120 return n;
1121}
1122
Kent Overstreet8835c122013-07-24 23:18:05 -07001123static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1124{
1125 unsigned i;
1126
Kent Overstreet05335cf2014-03-17 18:22:34 -07001127 mutex_lock(&b->c->bucket_lock);
1128
1129 atomic_inc(&b->c->prio_blocked);
1130
Kent Overstreet8835c122013-07-24 23:18:05 -07001131 bkey_copy(k, &b->key);
1132 bkey_copy_key(k, &ZERO_KEY);
1133
Kent Overstreet05335cf2014-03-17 18:22:34 -07001134 for (i = 0; i < KEY_PTRS(k); i++)
1135 SET_PTR_GEN(k, i,
1136 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1137 PTR_BUCKET(b->c, &b->key, i)));
Kent Overstreet8835c122013-07-24 23:18:05 -07001138
Kent Overstreet05335cf2014-03-17 18:22:34 -07001139 mutex_unlock(&b->c->bucket_lock);
Kent Overstreet8835c122013-07-24 23:18:05 -07001140}
1141
Kent Overstreet78365412013-12-17 01:29:34 -08001142static int btree_check_reserve(struct btree *b, struct btree_op *op)
1143{
1144 struct cache_set *c = b->c;
1145 struct cache *ca;
Kent Overstreet0a63b662014-03-17 17:15:53 -07001146 unsigned i, reserve = (c->root->level - b->level) * 2 + 1;
Kent Overstreet78365412013-12-17 01:29:34 -08001147
1148 mutex_lock(&c->bucket_lock);
1149
1150 for_each_cache(ca, c, i)
1151 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1152 if (op)
Kent Overstreet0a63b662014-03-17 17:15:53 -07001153 prepare_to_wait(&c->btree_cache_wait, &op->wait,
Kent Overstreet78365412013-12-17 01:29:34 -08001154 TASK_UNINTERRUPTIBLE);
Kent Overstreet0a63b662014-03-17 17:15:53 -07001155 mutex_unlock(&c->bucket_lock);
1156 return -EINTR;
Kent Overstreet78365412013-12-17 01:29:34 -08001157 }
1158
1159 mutex_unlock(&c->bucket_lock);
Kent Overstreet0a63b662014-03-17 17:15:53 -07001160
1161 return mca_cannibalize_lock(b->c, op);
Kent Overstreet78365412013-12-17 01:29:34 -08001162}
1163
Kent Overstreetcafe5632013-03-23 16:11:31 -07001164/* Garbage collection */
1165
Kent Overstreet487dded2014-03-17 15:13:26 -07001166static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1167 struct bkey *k)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001168{
1169 uint8_t stale = 0;
1170 unsigned i;
1171 struct bucket *g;
1172
1173 /*
1174 * ptr_invalid() can't return true for the keys that mark btree nodes as
1175 * freed, but since ptr_bad() returns true we'll never actually use them
1176 * for anything and thus we don't want mark their pointers here
1177 */
1178 if (!bkey_cmp(k, &ZERO_KEY))
1179 return stale;
1180
1181 for (i = 0; i < KEY_PTRS(k); i++) {
1182 if (!ptr_available(c, k, i))
1183 continue;
1184
1185 g = PTR_BUCKET(c, k, i);
1186
Kent Overstreet3a2fd9d2014-02-27 17:51:12 -08001187 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1188 g->last_gc = PTR_GEN(k, i);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001189
1190 if (ptr_stale(c, k, i)) {
1191 stale = max(stale, ptr_stale(c, k, i));
1192 continue;
1193 }
1194
1195 cache_bug_on(GC_MARK(g) &&
1196 (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1197 c, "inconsistent ptrs: mark = %llu, level = %i",
1198 GC_MARK(g), level);
1199
1200 if (level)
1201 SET_GC_MARK(g, GC_MARK_METADATA);
1202 else if (KEY_DIRTY(k))
1203 SET_GC_MARK(g, GC_MARK_DIRTY);
Kent Overstreet4fe6a812014-03-13 13:46:29 -07001204 else if (!GC_MARK(g))
1205 SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001206
1207 /* guard against overflow */
1208 SET_GC_SECTORS_USED(g, min_t(unsigned,
1209 GC_SECTORS_USED(g) + KEY_SIZE(k),
Darrick J. Wong94717442014-01-28 16:57:39 -08001210 MAX_GC_SECTORS_USED));
Kent Overstreetcafe5632013-03-23 16:11:31 -07001211
1212 BUG_ON(!GC_SECTORS_USED(g));
1213 }
1214
1215 return stale;
1216}
1217
1218#define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k)
1219
Kent Overstreet487dded2014-03-17 15:13:26 -07001220void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1221{
1222 unsigned i;
1223
1224 for (i = 0; i < KEY_PTRS(k); i++)
1225 if (ptr_available(c, k, i) &&
1226 !ptr_stale(c, k, i)) {
1227 struct bucket *b = PTR_BUCKET(c, k, i);
1228
1229 b->gen = PTR_GEN(k, i);
1230
1231 if (level && bkey_cmp(k, &ZERO_KEY))
1232 b->prio = BTREE_PRIO;
1233 else if (!level && b->prio == BTREE_PRIO)
1234 b->prio = INITIAL_PRIO;
1235 }
1236
1237 __bch_btree_mark_key(c, level, k);
1238}
1239
Kent Overstreeta1f03582013-09-10 19:07:00 -07001240static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001241{
1242 uint8_t stale = 0;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001243 unsigned keys = 0, good_keys = 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001244 struct bkey *k;
1245 struct btree_iter iter;
1246 struct bset_tree *t;
1247
1248 gc->nodes++;
1249
Kent Overstreetc052dd92013-11-11 17:35:24 -08001250 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
Kent Overstreetcafe5632013-03-23 16:11:31 -07001251 stale = max(stale, btree_mark_key(b, k));
Kent Overstreeta1f03582013-09-10 19:07:00 -07001252 keys++;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001253
Kent Overstreeta85e9682013-12-20 17:28:16 -08001254 if (bch_ptr_bad(&b->keys, k))
Kent Overstreetcafe5632013-03-23 16:11:31 -07001255 continue;
1256
Kent Overstreetcafe5632013-03-23 16:11:31 -07001257 gc->key_bytes += bkey_u64s(k);
1258 gc->nkeys++;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001259 good_keys++;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001260
1261 gc->data += KEY_SIZE(k);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001262 }
1263
Kent Overstreeta85e9682013-12-20 17:28:16 -08001264 for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001265 btree_bug_on(t->size &&
Kent Overstreeta85e9682013-12-20 17:28:16 -08001266 bset_written(&b->keys, t) &&
Kent Overstreetcafe5632013-03-23 16:11:31 -07001267 bkey_cmp(&b->key, &t->end) < 0,
1268 b, "found short btree key in gc");
1269
Kent Overstreeta1f03582013-09-10 19:07:00 -07001270 if (b->c->gc_always_rewrite)
1271 return true;
1272
1273 if (stale > 10)
1274 return true;
1275
1276 if ((keys - good_keys) * 2 > keys)
1277 return true;
1278
1279 return false;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001280}
1281
Kent Overstreeta1f03582013-09-10 19:07:00 -07001282#define GC_MERGE_NODES 4U
Kent Overstreetcafe5632013-03-23 16:11:31 -07001283
1284struct gc_merge_info {
1285 struct btree *b;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001286 unsigned keys;
1287};
1288
Kent Overstreeta1f03582013-09-10 19:07:00 -07001289static int bch_btree_insert_node(struct btree *, struct btree_op *,
1290 struct keylist *, atomic_t *, struct bkey *);
Kent Overstreetb54d6932013-07-24 18:04:18 -07001291
Kent Overstreeta1f03582013-09-10 19:07:00 -07001292static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
Kent Overstreet0a63b662014-03-17 17:15:53 -07001293 struct gc_stat *gc, struct gc_merge_info *r)
Kent Overstreeta1f03582013-09-10 19:07:00 -07001294{
1295 unsigned i, nodes = 0, keys = 0, blocks;
1296 struct btree *new_nodes[GC_MERGE_NODES];
Kent Overstreet0a63b662014-03-17 17:15:53 -07001297 struct keylist keylist;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001298 struct closure cl;
1299 struct bkey *k;
1300
Kent Overstreet0a63b662014-03-17 17:15:53 -07001301 bch_keylist_init(&keylist);
1302
1303 if (btree_check_reserve(b, NULL))
1304 return 0;
1305
Kent Overstreeta1f03582013-09-10 19:07:00 -07001306 memset(new_nodes, 0, sizeof(new_nodes));
Kent Overstreetb54d6932013-07-24 18:04:18 -07001307 closure_init_stack(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001308
Kent Overstreeta1f03582013-09-10 19:07:00 -07001309 while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
Kent Overstreetcafe5632013-03-23 16:11:31 -07001310 keys += r[nodes++].keys;
1311
1312 blocks = btree_default_blocks(b->c) * 2 / 3;
1313
1314 if (nodes < 2 ||
Kent Overstreeta85e9682013-12-20 17:28:16 -08001315 __set_blocks(b->keys.set[0].data, keys,
Kent Overstreetee811282013-12-17 23:49:49 -08001316 block_bytes(b->c)) > blocks * (nodes - 1))
Kent Overstreeta1f03582013-09-10 19:07:00 -07001317 return 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001318
Kent Overstreeta1f03582013-09-10 19:07:00 -07001319 for (i = 0; i < nodes; i++) {
Kent Overstreet0a63b662014-03-17 17:15:53 -07001320 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001321 if (IS_ERR_OR_NULL(new_nodes[i]))
1322 goto out_nocoalesce;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001323 }
1324
Kent Overstreet0a63b662014-03-17 17:15:53 -07001325 /*
1326 * We have to check the reserve here, after we've allocated our new
1327 * nodes, to make sure the insert below will succeed - we also check
1328 * before as an optimization to potentially avoid a bunch of expensive
1329 * allocs/sorts
1330 */
1331 if (btree_check_reserve(b, NULL))
1332 goto out_nocoalesce;
1333
Kent Overstreet2a285682014-03-04 16:42:42 -08001334 for (i = 0; i < nodes; i++)
1335 mutex_lock(&new_nodes[i]->write_lock);
1336
Kent Overstreetcafe5632013-03-23 16:11:31 -07001337 for (i = nodes - 1; i > 0; --i) {
Kent Overstreetee811282013-12-17 23:49:49 -08001338 struct bset *n1 = btree_bset_first(new_nodes[i]);
1339 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001340 struct bkey *k, *last = NULL;
1341
1342 keys = 0;
1343
Kent Overstreeta1f03582013-09-10 19:07:00 -07001344 if (i > 1) {
Kent Overstreetcafe5632013-03-23 16:11:31 -07001345 for (k = n2->start;
Kent Overstreetfafff812013-12-17 21:56:21 -08001346 k < bset_bkey_last(n2);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001347 k = bkey_next(k)) {
1348 if (__set_blocks(n1, n1->keys + keys +
Kent Overstreetee811282013-12-17 23:49:49 -08001349 bkey_u64s(k),
1350 block_bytes(b->c)) > blocks)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001351 break;
1352
1353 last = k;
1354 keys += bkey_u64s(k);
1355 }
Kent Overstreeta1f03582013-09-10 19:07:00 -07001356 } else {
1357 /*
1358 * Last node we're not getting rid of - we're getting
1359 * rid of the node at r[0]. Have to try and fit all of
1360 * the remaining keys into this node; we can't ensure
1361 * they will always fit due to rounding and variable
1362 * length keys (shouldn't be possible in practice,
1363 * though)
1364 */
1365 if (__set_blocks(n1, n1->keys + n2->keys,
Kent Overstreetee811282013-12-17 23:49:49 -08001366 block_bytes(b->c)) >
1367 btree_blocks(new_nodes[i]))
Kent Overstreeta1f03582013-09-10 19:07:00 -07001368 goto out_nocoalesce;
1369
1370 keys = n2->keys;
1371 /* Take the key of the node we're getting rid of */
1372 last = &r->b->key;
1373 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001374
Kent Overstreetee811282013-12-17 23:49:49 -08001375 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1376 btree_blocks(new_nodes[i]));
Kent Overstreetcafe5632013-03-23 16:11:31 -07001377
Kent Overstreeta1f03582013-09-10 19:07:00 -07001378 if (last)
1379 bkey_copy_key(&new_nodes[i]->key, last);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001380
Kent Overstreetfafff812013-12-17 21:56:21 -08001381 memcpy(bset_bkey_last(n1),
Kent Overstreetcafe5632013-03-23 16:11:31 -07001382 n2->start,
Kent Overstreetfafff812013-12-17 21:56:21 -08001383 (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001384
1385 n1->keys += keys;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001386 r[i].keys = n1->keys;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001387
1388 memmove(n2->start,
Kent Overstreetfafff812013-12-17 21:56:21 -08001389 bset_bkey_idx(n2, keys),
1390 (void *) bset_bkey_last(n2) -
1391 (void *) bset_bkey_idx(n2, keys));
Kent Overstreetcafe5632013-03-23 16:11:31 -07001392
1393 n2->keys -= keys;
1394
Kent Overstreet0a63b662014-03-17 17:15:53 -07001395 if (__bch_keylist_realloc(&keylist,
Kent Overstreet085d2a32013-11-11 18:20:51 -08001396 bkey_u64s(&new_nodes[i]->key)))
Kent Overstreeta1f03582013-09-10 19:07:00 -07001397 goto out_nocoalesce;
1398
1399 bch_btree_node_write(new_nodes[i], &cl);
Kent Overstreet0a63b662014-03-17 17:15:53 -07001400 bch_keylist_add(&keylist, &new_nodes[i]->key);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001401 }
1402
Kent Overstreet2a285682014-03-04 16:42:42 -08001403 for (i = 0; i < nodes; i++)
1404 mutex_unlock(&new_nodes[i]->write_lock);
1405
Kent Overstreet05335cf2014-03-17 18:22:34 -07001406 closure_sync(&cl);
1407
1408 /* We emptied out this node */
1409 BUG_ON(btree_bset_first(new_nodes[0])->keys);
1410 btree_node_free(new_nodes[0]);
1411 rw_unlock(true, new_nodes[0]);
1412
Kent Overstreeta1f03582013-09-10 19:07:00 -07001413 for (i = 0; i < nodes; i++) {
Kent Overstreet0a63b662014-03-17 17:15:53 -07001414 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
Kent Overstreeta1f03582013-09-10 19:07:00 -07001415 goto out_nocoalesce;
1416
Kent Overstreet0a63b662014-03-17 17:15:53 -07001417 make_btree_freeing_key(r[i].b, keylist.top);
1418 bch_keylist_push(&keylist);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001419 }
1420
Kent Overstreet0a63b662014-03-17 17:15:53 -07001421 bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1422 BUG_ON(!bch_keylist_empty(&keylist));
Kent Overstreeta1f03582013-09-10 19:07:00 -07001423
1424 for (i = 0; i < nodes; i++) {
1425 btree_node_free(r[i].b);
1426 rw_unlock(true, r[i].b);
1427
1428 r[i].b = new_nodes[i];
1429 }
1430
Kent Overstreeta1f03582013-09-10 19:07:00 -07001431 memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1432 r[nodes - 1].b = ERR_PTR(-EINTR);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001433
Kent Overstreetc37511b2013-04-26 15:39:55 -07001434 trace_bcache_btree_gc_coalesce(nodes);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001435 gc->nodes--;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001436
Kent Overstreet0a63b662014-03-17 17:15:53 -07001437 bch_keylist_free(&keylist);
1438
Kent Overstreeta1f03582013-09-10 19:07:00 -07001439 /* Invalidated our iterator */
1440 return -EINTR;
1441
1442out_nocoalesce:
1443 closure_sync(&cl);
Kent Overstreet0a63b662014-03-17 17:15:53 -07001444 bch_keylist_free(&keylist);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001445
Kent Overstreet0a63b662014-03-17 17:15:53 -07001446 while ((k = bch_keylist_pop(&keylist)))
Kent Overstreeta1f03582013-09-10 19:07:00 -07001447 if (!bkey_cmp(k, &ZERO_KEY))
1448 atomic_dec(&b->c->prio_blocked);
1449
1450 for (i = 0; i < nodes; i++)
1451 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1452 btree_node_free(new_nodes[i]);
1453 rw_unlock(true, new_nodes[i]);
1454 }
1455 return 0;
1456}
1457
Kent Overstreet0a63b662014-03-17 17:15:53 -07001458static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1459 struct btree *replace)
1460{
1461 struct keylist keys;
1462 struct btree *n;
1463
1464 if (btree_check_reserve(b, NULL))
1465 return 0;
1466
1467 n = btree_node_alloc_replacement(replace, NULL);
1468
1469 /* recheck reserve after allocating replacement node */
1470 if (btree_check_reserve(b, NULL)) {
1471 btree_node_free(n);
1472 rw_unlock(true, n);
1473 return 0;
1474 }
1475
1476 bch_btree_node_write_sync(n);
1477
1478 bch_keylist_init(&keys);
1479 bch_keylist_add(&keys, &n->key);
1480
1481 make_btree_freeing_key(replace, keys.top);
1482 bch_keylist_push(&keys);
1483
1484 bch_btree_insert_node(b, op, &keys, NULL, NULL);
1485 BUG_ON(!bch_keylist_empty(&keys));
1486
1487 btree_node_free(replace);
1488 rw_unlock(true, n);
1489
1490 /* Invalidated our iterator */
1491 return -EINTR;
1492}
1493
Kent Overstreeta1f03582013-09-10 19:07:00 -07001494static unsigned btree_gc_count_keys(struct btree *b)
1495{
1496 struct bkey *k;
1497 struct btree_iter iter;
1498 unsigned ret = 0;
1499
Kent Overstreetc052dd92013-11-11 17:35:24 -08001500 for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
Kent Overstreeta1f03582013-09-10 19:07:00 -07001501 ret += bkey_u64s(k);
1502
1503 return ret;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001504}
1505
1506static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1507 struct closure *writes, struct gc_stat *gc)
1508{
Kent Overstreeta1f03582013-09-10 19:07:00 -07001509 int ret = 0;
1510 bool should_rewrite;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001511 struct bkey *k;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001512 struct btree_iter iter;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001513 struct gc_merge_info r[GC_MERGE_NODES];
Kent Overstreet2a285682014-03-04 16:42:42 -08001514 struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001515
Kent Overstreetc052dd92013-11-11 17:35:24 -08001516 bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001517
Kent Overstreet2a285682014-03-04 16:42:42 -08001518 for (i = r; i < r + ARRAY_SIZE(r); i++)
1519 i->b = ERR_PTR(-EINTR);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001520
Kent Overstreeta1f03582013-09-10 19:07:00 -07001521 while (1) {
Kent Overstreeta85e9682013-12-20 17:28:16 -08001522 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001523 if (k) {
Kent Overstreet0a63b662014-03-17 17:15:53 -07001524 r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1525 true);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001526 if (IS_ERR(r->b)) {
1527 ret = PTR_ERR(r->b);
1528 break;
1529 }
1530
1531 r->keys = btree_gc_count_keys(r->b);
1532
Kent Overstreet0a63b662014-03-17 17:15:53 -07001533 ret = btree_gc_coalesce(b, op, gc, r);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001534 if (ret)
1535 break;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001536 }
1537
Kent Overstreeta1f03582013-09-10 19:07:00 -07001538 if (!last->b)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001539 break;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001540
1541 if (!IS_ERR(last->b)) {
1542 should_rewrite = btree_gc_mark_node(last->b, gc);
Kent Overstreet0a63b662014-03-17 17:15:53 -07001543 if (should_rewrite) {
1544 ret = btree_gc_rewrite_node(b, op, last->b);
1545 if (ret)
Kent Overstreeta1f03582013-09-10 19:07:00 -07001546 break;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001547 }
1548
1549 if (last->b->level) {
1550 ret = btree_gc_recurse(last->b, op, writes, gc);
1551 if (ret)
1552 break;
1553 }
1554
1555 bkey_copy_key(&b->c->gc_done, &last->b->key);
1556
1557 /*
1558 * Must flush leaf nodes before gc ends, since replace
1559 * operations aren't journalled
1560 */
Kent Overstreet2a285682014-03-04 16:42:42 -08001561 mutex_lock(&last->b->write_lock);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001562 if (btree_node_dirty(last->b))
1563 bch_btree_node_write(last->b, writes);
Kent Overstreet2a285682014-03-04 16:42:42 -08001564 mutex_unlock(&last->b->write_lock);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001565 rw_unlock(true, last->b);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001566 }
1567
Kent Overstreeta1f03582013-09-10 19:07:00 -07001568 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1569 r->b = NULL;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001570
Kent Overstreetcafe5632013-03-23 16:11:31 -07001571 if (need_resched()) {
1572 ret = -EAGAIN;
1573 break;
1574 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001575 }
1576
Kent Overstreet2a285682014-03-04 16:42:42 -08001577 for (i = r; i < r + ARRAY_SIZE(r); i++)
1578 if (!IS_ERR_OR_NULL(i->b)) {
1579 mutex_lock(&i->b->write_lock);
1580 if (btree_node_dirty(i->b))
1581 bch_btree_node_write(i->b, writes);
1582 mutex_unlock(&i->b->write_lock);
1583 rw_unlock(true, i->b);
Kent Overstreeta1f03582013-09-10 19:07:00 -07001584 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001585
Kent Overstreetcafe5632013-03-23 16:11:31 -07001586 return ret;
1587}
1588
1589static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1590 struct closure *writes, struct gc_stat *gc)
1591{
1592 struct btree *n = NULL;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001593 int ret = 0;
1594 bool should_rewrite;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001595
Kent Overstreeta1f03582013-09-10 19:07:00 -07001596 should_rewrite = btree_gc_mark_node(b, gc);
1597 if (should_rewrite) {
Kent Overstreet0a63b662014-03-17 17:15:53 -07001598 n = btree_node_alloc_replacement(b, NULL);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001599
Kent Overstreeta1f03582013-09-10 19:07:00 -07001600 if (!IS_ERR_OR_NULL(n)) {
1601 bch_btree_node_write_sync(n);
Kent Overstreet2a285682014-03-04 16:42:42 -08001602
Kent Overstreeta1f03582013-09-10 19:07:00 -07001603 bch_btree_set_root(n);
1604 btree_node_free(b);
1605 rw_unlock(true, n);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001606
Kent Overstreeta1f03582013-09-10 19:07:00 -07001607 return -EINTR;
1608 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001609 }
1610
Kent Overstreet487dded2014-03-17 15:13:26 -07001611 __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1612
Kent Overstreeta1f03582013-09-10 19:07:00 -07001613 if (b->level) {
1614 ret = btree_gc_recurse(b, op, writes, gc);
1615 if (ret)
1616 return ret;
1617 }
1618
1619 bkey_copy_key(&b->c->gc_done, &b->key);
1620
Kent Overstreetcafe5632013-03-23 16:11:31 -07001621 return ret;
1622}
1623
1624static void btree_gc_start(struct cache_set *c)
1625{
1626 struct cache *ca;
1627 struct bucket *b;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001628 unsigned i;
1629
1630 if (!c->gc_mark_valid)
1631 return;
1632
1633 mutex_lock(&c->bucket_lock);
1634
1635 c->gc_mark_valid = 0;
1636 c->gc_done = ZERO_KEY;
1637
1638 for_each_cache(ca, c, i)
1639 for_each_bucket(b, ca) {
Kent Overstreet3a2fd9d2014-02-27 17:51:12 -08001640 b->last_gc = b->gen;
Kent Overstreet29ebf462013-07-11 19:43:21 -07001641 if (!atomic_read(&b->pin)) {
Kent Overstreet4fe6a812014-03-13 13:46:29 -07001642 SET_GC_MARK(b, 0);
Kent Overstreet29ebf462013-07-11 19:43:21 -07001643 SET_GC_SECTORS_USED(b, 0);
1644 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001645 }
1646
Kent Overstreetcafe5632013-03-23 16:11:31 -07001647 mutex_unlock(&c->bucket_lock);
1648}
1649
Kent Overstreet2531d9ee2014-03-17 16:55:55 -07001650static size_t bch_btree_gc_finish(struct cache_set *c)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001651{
1652 size_t available = 0;
1653 struct bucket *b;
1654 struct cache *ca;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001655 unsigned i;
1656
1657 mutex_lock(&c->bucket_lock);
1658
1659 set_gc_sectors(c);
1660 c->gc_mark_valid = 1;
1661 c->need_gc = 0;
1662
Kent Overstreetcafe5632013-03-23 16:11:31 -07001663 for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1664 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1665 GC_MARK_METADATA);
1666
Nicholas Swensonbf0a6282013-11-26 19:14:23 -08001667 /* don't reclaim buckets to which writeback keys point */
1668 rcu_read_lock();
1669 for (i = 0; i < c->nr_uuids; i++) {
1670 struct bcache_device *d = c->devices[i];
1671 struct cached_dev *dc;
1672 struct keybuf_key *w, *n;
1673 unsigned j;
1674
1675 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1676 continue;
1677 dc = container_of(d, struct cached_dev, disk);
1678
1679 spin_lock(&dc->writeback_keys.lock);
1680 rbtree_postorder_for_each_entry_safe(w, n,
1681 &dc->writeback_keys.keys, node)
1682 for (j = 0; j < KEY_PTRS(&w->key); j++)
1683 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1684 GC_MARK_DIRTY);
1685 spin_unlock(&dc->writeback_keys.lock);
1686 }
1687 rcu_read_unlock();
1688
Kent Overstreetcafe5632013-03-23 16:11:31 -07001689 for_each_cache(ca, c, i) {
1690 uint64_t *i;
1691
1692 ca->invalidate_needs_gc = 0;
1693
1694 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1695 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1696
1697 for (i = ca->prio_buckets;
1698 i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1699 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1700
1701 for_each_bucket(b, ca) {
Kent Overstreetcafe5632013-03-23 16:11:31 -07001702 c->need_gc = max(c->need_gc, bucket_gc_gen(b));
1703
Kent Overstreet4fe6a812014-03-13 13:46:29 -07001704 if (atomic_read(&b->pin))
1705 continue;
1706
1707 BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1708
1709 if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001710 available++;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001711 }
1712 }
1713
Kent Overstreetcafe5632013-03-23 16:11:31 -07001714 mutex_unlock(&c->bucket_lock);
1715 return available;
1716}
1717
Kent Overstreet72a44512013-10-24 17:19:26 -07001718static void bch_btree_gc(struct cache_set *c)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001719{
Kent Overstreetcafe5632013-03-23 16:11:31 -07001720 int ret;
1721 unsigned long available;
1722 struct gc_stat stats;
1723 struct closure writes;
1724 struct btree_op op;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001725 uint64_t start_time = local_clock();
Kent Overstreet57943512013-04-25 13:58:35 -07001726
Kent Overstreetc37511b2013-04-26 15:39:55 -07001727 trace_bcache_gc_start(c);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001728
1729 memset(&stats, 0, sizeof(struct gc_stat));
1730 closure_init_stack(&writes);
Kent Overstreetb54d6932013-07-24 18:04:18 -07001731 bch_btree_op_init(&op, SHRT_MAX);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001732
1733 btree_gc_start(c);
1734
Kent Overstreeta1f03582013-09-10 19:07:00 -07001735 do {
1736 ret = btree_root(gc_root, c, &op, &writes, &stats);
1737 closure_sync(&writes);
Kent Overstreet57943512013-04-25 13:58:35 -07001738
Kent Overstreeta1f03582013-09-10 19:07:00 -07001739 if (ret && ret != -EAGAIN)
1740 pr_warn("gc failed!");
1741 } while (ret);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001742
1743 available = bch_btree_gc_finish(c);
Kent Overstreet57943512013-04-25 13:58:35 -07001744 wake_up_allocators(c);
1745
Kent Overstreet169ef1c2013-03-28 12:50:55 -06001746 bch_time_stats_update(&c->btree_gc_time, start_time);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001747
1748 stats.key_bytes *= sizeof(uint64_t);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001749 stats.data <<= 9;
1750 stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets;
1751 memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
Kent Overstreetcafe5632013-03-23 16:11:31 -07001752
Kent Overstreetc37511b2013-04-26 15:39:55 -07001753 trace_bcache_gc_end(c);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001754
Kent Overstreet72a44512013-10-24 17:19:26 -07001755 bch_moving_gc(c);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001756}
1757
Kent Overstreet72a44512013-10-24 17:19:26 -07001758static int bch_gc_thread(void *arg)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001759{
Kent Overstreet72a44512013-10-24 17:19:26 -07001760 struct cache_set *c = arg;
Kent Overstreeta1f03582013-09-10 19:07:00 -07001761 struct cache *ca;
1762 unsigned i;
Kent Overstreet72a44512013-10-24 17:19:26 -07001763
1764 while (1) {
Kent Overstreeta1f03582013-09-10 19:07:00 -07001765again:
Kent Overstreet72a44512013-10-24 17:19:26 -07001766 bch_btree_gc(c);
1767
1768 set_current_state(TASK_INTERRUPTIBLE);
1769 if (kthread_should_stop())
1770 break;
1771
Kent Overstreeta1f03582013-09-10 19:07:00 -07001772 mutex_lock(&c->bucket_lock);
1773
1774 for_each_cache(ca, c, i)
1775 if (ca->invalidate_needs_gc) {
1776 mutex_unlock(&c->bucket_lock);
1777 set_current_state(TASK_RUNNING);
1778 goto again;
1779 }
1780
1781 mutex_unlock(&c->bucket_lock);
1782
Kent Overstreet72a44512013-10-24 17:19:26 -07001783 try_to_freeze();
1784 schedule();
1785 }
1786
1787 return 0;
1788}
1789
1790int bch_gc_thread_start(struct cache_set *c)
1791{
1792 c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc");
1793 if (IS_ERR(c->gc_thread))
1794 return PTR_ERR(c->gc_thread);
1795
1796 set_task_state(c->gc_thread, TASK_INTERRUPTIBLE);
1797 return 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001798}
1799
1800/* Initial partial gc */
1801
Kent Overstreet487dded2014-03-17 15:13:26 -07001802static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001803{
Kent Overstreet50310162013-09-10 17:18:59 -07001804 int ret = 0;
Kent Overstreet50310162013-09-10 17:18:59 -07001805 struct bkey *k, *p = NULL;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001806 struct btree_iter iter;
1807
Kent Overstreet487dded2014-03-17 15:13:26 -07001808 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1809 bch_initial_mark_key(b->c, b->level, k);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001810
Kent Overstreet487dded2014-03-17 15:13:26 -07001811 bch_initial_mark_key(b->c, b->level + 1, &b->key);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001812
1813 if (b->level) {
Kent Overstreetc052dd92013-11-11 17:35:24 -08001814 bch_btree_iter_init(&b->keys, &iter, NULL);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001815
Kent Overstreet50310162013-09-10 17:18:59 -07001816 do {
Kent Overstreeta85e9682013-12-20 17:28:16 -08001817 k = bch_btree_iter_next_filter(&iter, &b->keys,
1818 bch_ptr_bad);
Kent Overstreet50310162013-09-10 17:18:59 -07001819 if (k)
1820 btree_node_prefetch(b->c, k, b->level - 1);
1821
Kent Overstreetcafe5632013-03-23 16:11:31 -07001822 if (p)
Kent Overstreet487dded2014-03-17 15:13:26 -07001823 ret = btree(check_recurse, p, b, op);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001824
Kent Overstreet50310162013-09-10 17:18:59 -07001825 p = k;
1826 } while (p && !ret);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001827 }
1828
Kent Overstreet487dded2014-03-17 15:13:26 -07001829 return ret;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001830}
1831
Kent Overstreetc18536a2013-07-24 17:44:17 -07001832int bch_btree_check(struct cache_set *c)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001833{
Kent Overstreetc18536a2013-07-24 17:44:17 -07001834 struct btree_op op;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001835
Kent Overstreetb54d6932013-07-24 18:04:18 -07001836 bch_btree_op_init(&op, SHRT_MAX);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001837
Kent Overstreet487dded2014-03-17 15:13:26 -07001838 return btree_root(check_recurse, c, &op);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001839}
1840
Kent Overstreet2531d9ee2014-03-17 16:55:55 -07001841void bch_initial_gc_finish(struct cache_set *c)
1842{
1843 struct cache *ca;
1844 struct bucket *b;
1845 unsigned i;
1846
1847 bch_btree_gc_finish(c);
1848
1849 mutex_lock(&c->bucket_lock);
1850
1851 /*
1852 * We need to put some unused buckets directly on the prio freelist in
1853 * order to get the allocator thread started - it needs freed buckets in
1854 * order to rewrite the prios and gens, and it needs to rewrite prios
1855 * and gens in order to free buckets.
1856 *
1857 * This is only safe for buckets that have no live data in them, which
1858 * there should always be some of.
1859 */
1860 for_each_cache(ca, c, i) {
1861 for_each_bucket(b, ca) {
1862 if (fifo_full(&ca->free[RESERVE_PRIO]))
1863 break;
1864
1865 if (bch_can_invalidate_bucket(ca, b) &&
1866 !GC_MARK(b)) {
1867 __bch_invalidate_one_bucket(ca, b);
1868 fifo_push(&ca->free[RESERVE_PRIO],
1869 b - ca->buckets);
1870 }
1871 }
1872 }
1873
1874 mutex_unlock(&c->bucket_lock);
1875}
1876
Kent Overstreetcafe5632013-03-23 16:11:31 -07001877/* Btree insertion */
1878
Kent Overstreet829a60b2013-11-11 17:02:31 -08001879static bool btree_insert_key(struct btree *b, struct bkey *k,
1880 struct bkey *replace_key)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001881{
Kent Overstreet829a60b2013-11-11 17:02:31 -08001882 unsigned status;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001883
1884 BUG_ON(bkey_cmp(k, &b->key) > 0);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001885
Kent Overstreet829a60b2013-11-11 17:02:31 -08001886 status = bch_btree_insert_key(&b->keys, k, replace_key);
1887 if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1888 bch_check_keys(&b->keys, "%u for %s", status,
1889 replace_key ? "replace" : "insert");
Kent Overstreetcafe5632013-03-23 16:11:31 -07001890
Kent Overstreet829a60b2013-11-11 17:02:31 -08001891 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1892 status);
1893 return true;
1894 } else
1895 return false;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001896}
1897
Kent Overstreet59158fd2013-11-11 19:03:54 -08001898static size_t insert_u64s_remaining(struct btree *b)
1899{
Kent Overstreet35723242014-01-10 18:53:02 -08001900 long ret = bch_btree_keys_u64s_remaining(&b->keys);
Kent Overstreet59158fd2013-11-11 19:03:54 -08001901
1902 /*
1903 * Might land in the middle of an existing extent and have to split it
1904 */
1905 if (b->keys.ops->is_extents)
1906 ret -= KEY_MAX_U64S;
1907
1908 return max(ret, 0L);
1909}
1910
Kent Overstreet26c949f2013-09-10 18:41:15 -07001911static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
Kent Overstreet1b207d82013-09-10 18:52:54 -07001912 struct keylist *insert_keys,
1913 struct bkey *replace_key)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001914{
1915 bool ret = false;
Kent Overstreetdc9d98d2013-12-17 23:47:33 -08001916 int oldsize = bch_count_data(&b->keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001917
Kent Overstreet26c949f2013-09-10 18:41:15 -07001918 while (!bch_keylist_empty(insert_keys)) {
Kent Overstreetc2f95ae2013-07-24 17:24:25 -07001919 struct bkey *k = insert_keys->keys;
Kent Overstreet26c949f2013-09-10 18:41:15 -07001920
Kent Overstreet59158fd2013-11-11 19:03:54 -08001921 if (bkey_u64s(k) > insert_u64s_remaining(b))
Kent Overstreet403b6cd2013-07-24 17:22:44 -07001922 break;
1923
1924 if (bkey_cmp(k, &b->key) <= 0) {
Kent Overstreet3a3b6a42013-07-24 16:46:42 -07001925 if (!b->level)
1926 bkey_put(b->c, k);
Kent Overstreet26c949f2013-09-10 18:41:15 -07001927
Kent Overstreet829a60b2013-11-11 17:02:31 -08001928 ret |= btree_insert_key(b, k, replace_key);
Kent Overstreet26c949f2013-09-10 18:41:15 -07001929 bch_keylist_pop_front(insert_keys);
1930 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
Kent Overstreet26c949f2013-09-10 18:41:15 -07001931 BKEY_PADDED(key) temp;
Kent Overstreetc2f95ae2013-07-24 17:24:25 -07001932 bkey_copy(&temp.key, insert_keys->keys);
Kent Overstreet26c949f2013-09-10 18:41:15 -07001933
1934 bch_cut_back(&b->key, &temp.key);
Kent Overstreetc2f95ae2013-07-24 17:24:25 -07001935 bch_cut_front(&b->key, insert_keys->keys);
Kent Overstreet26c949f2013-09-10 18:41:15 -07001936
Kent Overstreet829a60b2013-11-11 17:02:31 -08001937 ret |= btree_insert_key(b, &temp.key, replace_key);
Kent Overstreet26c949f2013-09-10 18:41:15 -07001938 break;
1939 } else {
1940 break;
1941 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07001942 }
1943
Kent Overstreet829a60b2013-11-11 17:02:31 -08001944 if (!ret)
1945 op->insert_collision = true;
1946
Kent Overstreet403b6cd2013-07-24 17:22:44 -07001947 BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1948
Kent Overstreetdc9d98d2013-12-17 23:47:33 -08001949 BUG_ON(bch_count_data(&b->keys) < oldsize);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001950 return ret;
1951}
1952
Kent Overstreet26c949f2013-09-10 18:41:15 -07001953static int btree_split(struct btree *b, struct btree_op *op,
1954 struct keylist *insert_keys,
Kent Overstreet1b207d82013-09-10 18:52:54 -07001955 struct bkey *replace_key)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001956{
Kent Overstreetd6fd3b12013-07-24 17:20:19 -07001957 bool split;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001958 struct btree *n1, *n2 = NULL, *n3 = NULL;
1959 uint64_t start_time = local_clock();
Kent Overstreetb54d6932013-07-24 18:04:18 -07001960 struct closure cl;
Kent Overstreet17e21a92013-07-26 12:32:38 -07001961 struct keylist parent_keys;
Kent Overstreetb54d6932013-07-24 18:04:18 -07001962
1963 closure_init_stack(&cl);
Kent Overstreet17e21a92013-07-26 12:32:38 -07001964 bch_keylist_init(&parent_keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001965
Kent Overstreet0a63b662014-03-17 17:15:53 -07001966 if (btree_check_reserve(b, op)) {
1967 if (!b->level)
1968 return -EINTR;
1969 else
1970 WARN(1, "insufficient reserve for split\n");
1971 }
Kent Overstreet78365412013-12-17 01:29:34 -08001972
Kent Overstreet0a63b662014-03-17 17:15:53 -07001973 n1 = btree_node_alloc_replacement(b, op);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001974 if (IS_ERR(n1))
1975 goto err;
1976
Kent Overstreetee811282013-12-17 23:49:49 -08001977 split = set_blocks(btree_bset_first(n1),
1978 block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001979
Kent Overstreetcafe5632013-03-23 16:11:31 -07001980 if (split) {
1981 unsigned keys = 0;
1982
Kent Overstreetee811282013-12-17 23:49:49 -08001983 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
Kent Overstreetc37511b2013-04-26 15:39:55 -07001984
Kent Overstreet0a63b662014-03-17 17:15:53 -07001985 n2 = bch_btree_node_alloc(b->c, op, b->level);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001986 if (IS_ERR(n2))
1987 goto err_free1;
1988
Kent Overstreetd6fd3b12013-07-24 17:20:19 -07001989 if (!b->parent) {
Kent Overstreet0a63b662014-03-17 17:15:53 -07001990 n3 = bch_btree_node_alloc(b->c, op, b->level + 1);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001991 if (IS_ERR(n3))
1992 goto err_free2;
1993 }
1994
Kent Overstreet2a285682014-03-04 16:42:42 -08001995 mutex_lock(&n1->write_lock);
1996 mutex_lock(&n2->write_lock);
1997
Kent Overstreet1b207d82013-09-10 18:52:54 -07001998 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001999
Kent Overstreetd6fd3b12013-07-24 17:20:19 -07002000 /*
2001 * Has to be a linear search because we don't have an auxiliary
Kent Overstreetcafe5632013-03-23 16:11:31 -07002002 * search tree yet
2003 */
2004
Kent Overstreetee811282013-12-17 23:49:49 -08002005 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2006 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
Kent Overstreetfafff812013-12-17 21:56:21 -08002007 keys));
Kent Overstreetcafe5632013-03-23 16:11:31 -07002008
Kent Overstreetfafff812013-12-17 21:56:21 -08002009 bkey_copy_key(&n1->key,
Kent Overstreetee811282013-12-17 23:49:49 -08002010 bset_bkey_idx(btree_bset_first(n1), keys));
2011 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
Kent Overstreetcafe5632013-03-23 16:11:31 -07002012
Kent Overstreetee811282013-12-17 23:49:49 -08002013 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2014 btree_bset_first(n1)->keys = keys;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002015
Kent Overstreetee811282013-12-17 23:49:49 -08002016 memcpy(btree_bset_first(n2)->start,
2017 bset_bkey_last(btree_bset_first(n1)),
2018 btree_bset_first(n2)->keys * sizeof(uint64_t));
Kent Overstreetcafe5632013-03-23 16:11:31 -07002019
2020 bkey_copy_key(&n2->key, &b->key);
2021
Kent Overstreet17e21a92013-07-26 12:32:38 -07002022 bch_keylist_add(&parent_keys, &n2->key);
Kent Overstreetb54d6932013-07-24 18:04:18 -07002023 bch_btree_node_write(n2, &cl);
Kent Overstreet2a285682014-03-04 16:42:42 -08002024 mutex_unlock(&n2->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002025 rw_unlock(true, n2);
Kent Overstreetc37511b2013-04-26 15:39:55 -07002026 } else {
Kent Overstreetee811282013-12-17 23:49:49 -08002027 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
Kent Overstreetc37511b2013-04-26 15:39:55 -07002028
Kent Overstreet2a285682014-03-04 16:42:42 -08002029 mutex_lock(&n1->write_lock);
Kent Overstreet1b207d82013-09-10 18:52:54 -07002030 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
Kent Overstreetc37511b2013-04-26 15:39:55 -07002031 }
Kent Overstreetcafe5632013-03-23 16:11:31 -07002032
Kent Overstreet17e21a92013-07-26 12:32:38 -07002033 bch_keylist_add(&parent_keys, &n1->key);
Kent Overstreetb54d6932013-07-24 18:04:18 -07002034 bch_btree_node_write(n1, &cl);
Kent Overstreet2a285682014-03-04 16:42:42 -08002035 mutex_unlock(&n1->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002036
2037 if (n3) {
Kent Overstreetd6fd3b12013-07-24 17:20:19 -07002038 /* Depth increases, make a new root */
Kent Overstreet2a285682014-03-04 16:42:42 -08002039 mutex_lock(&n3->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002040 bkey_copy_key(&n3->key, &MAX_KEY);
Kent Overstreet17e21a92013-07-26 12:32:38 -07002041 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
Kent Overstreetb54d6932013-07-24 18:04:18 -07002042 bch_btree_node_write(n3, &cl);
Kent Overstreet2a285682014-03-04 16:42:42 -08002043 mutex_unlock(&n3->write_lock);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002044
Kent Overstreetb54d6932013-07-24 18:04:18 -07002045 closure_sync(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002046 bch_btree_set_root(n3);
2047 rw_unlock(true, n3);
Kent Overstreetd6fd3b12013-07-24 17:20:19 -07002048 } else if (!b->parent) {
2049 /* Root filled up but didn't need to be split */
Kent Overstreetb54d6932013-07-24 18:04:18 -07002050 closure_sync(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002051 bch_btree_set_root(n1);
2052 } else {
Kent Overstreet17e21a92013-07-26 12:32:38 -07002053 /* Split a non root node */
Kent Overstreetb54d6932013-07-24 18:04:18 -07002054 closure_sync(&cl);
Kent Overstreet17e21a92013-07-26 12:32:38 -07002055 make_btree_freeing_key(b, parent_keys.top);
2056 bch_keylist_push(&parent_keys);
2057
Kent Overstreet17e21a92013-07-26 12:32:38 -07002058 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2059 BUG_ON(!bch_keylist_empty(&parent_keys));
Kent Overstreetcafe5632013-03-23 16:11:31 -07002060 }
2061
Kent Overstreet05335cf2014-03-17 18:22:34 -07002062 btree_node_free(b);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002063 rw_unlock(true, n1);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002064
Kent Overstreet169ef1c2013-03-28 12:50:55 -06002065 bch_time_stats_update(&b->c->btree_split_time, start_time);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002066
2067 return 0;
2068err_free2:
Kent Overstreet5f5837d2013-12-16 16:38:49 -08002069 bkey_put(b->c, &n2->key);
Kent Overstreete8e1d462013-07-24 17:27:07 -07002070 btree_node_free(n2);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002071 rw_unlock(true, n2);
2072err_free1:
Kent Overstreet5f5837d2013-12-16 16:38:49 -08002073 bkey_put(b->c, &n1->key);
Kent Overstreete8e1d462013-07-24 17:27:07 -07002074 btree_node_free(n1);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002075 rw_unlock(true, n1);
2076err:
Kent Overstreet0a63b662014-03-17 17:15:53 -07002077 WARN(1, "bcache: btree split failed (level %u)", b->level);
Kent Overstreet5f5837d2013-12-16 16:38:49 -08002078
Kent Overstreetcafe5632013-03-23 16:11:31 -07002079 if (n3 == ERR_PTR(-EAGAIN) ||
2080 n2 == ERR_PTR(-EAGAIN) ||
2081 n1 == ERR_PTR(-EAGAIN))
2082 return -EAGAIN;
2083
Kent Overstreetcafe5632013-03-23 16:11:31 -07002084 return -ENOMEM;
2085}
2086
Kent Overstreet26c949f2013-09-10 18:41:15 -07002087static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
Kent Overstreetc18536a2013-07-24 17:44:17 -07002088 struct keylist *insert_keys,
Kent Overstreet1b207d82013-09-10 18:52:54 -07002089 atomic_t *journal_ref,
2090 struct bkey *replace_key)
Kent Overstreet26c949f2013-09-10 18:41:15 -07002091{
Kent Overstreet2a285682014-03-04 16:42:42 -08002092 struct closure cl;
2093
Kent Overstreet17e21a92013-07-26 12:32:38 -07002094 BUG_ON(b->level && replace_key);
Kent Overstreet26c949f2013-09-10 18:41:15 -07002095
Kent Overstreet2a285682014-03-04 16:42:42 -08002096 closure_init_stack(&cl);
2097
2098 mutex_lock(&b->write_lock);
2099
2100 if (write_block(b) != btree_bset_last(b) &&
2101 b->keys.last_set_unwritten)
2102 bch_btree_init_next(b); /* just wrote a set */
2103
Kent Overstreet59158fd2013-11-11 19:03:54 -08002104 if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
Kent Overstreet2a285682014-03-04 16:42:42 -08002105 mutex_unlock(&b->write_lock);
2106 goto split;
2107 }
Kent Overstreet3b3e9e52013-12-07 03:57:58 -08002108
Kent Overstreet2a285682014-03-04 16:42:42 -08002109 BUG_ON(write_block(b) != btree_bset_last(b));
2110
2111 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2112 if (!b->level)
2113 bch_btree_leaf_dirty(b, journal_ref);
2114 else
2115 bch_btree_node_write(b, &cl);
2116 }
2117
2118 mutex_unlock(&b->write_lock);
2119
2120 /* wait for btree node write if necessary, after unlock */
2121 closure_sync(&cl);
2122
2123 return 0;
2124split:
2125 if (current->bio_list) {
2126 op->lock = b->c->root->level + 1;
2127 return -EAGAIN;
2128 } else if (op->lock <= b->c->root->level) {
2129 op->lock = b->c->root->level + 1;
2130 return -EINTR;
Kent Overstreet17e21a92013-07-26 12:32:38 -07002131 } else {
Kent Overstreet2a285682014-03-04 16:42:42 -08002132 /* Invalidated all iterators */
2133 int ret = btree_split(b, op, insert_keys, replace_key);
Kent Overstreet26c949f2013-09-10 18:41:15 -07002134
Kent Overstreet2a285682014-03-04 16:42:42 -08002135 if (bch_keylist_empty(insert_keys))
2136 return 0;
2137 else if (!ret)
2138 return -EINTR;
2139 return ret;
Kent Overstreet17e21a92013-07-26 12:32:38 -07002140 }
Kent Overstreet26c949f2013-09-10 18:41:15 -07002141}
2142
Kent Overstreete7c590e2013-09-10 18:39:16 -07002143int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2144 struct bkey *check_key)
2145{
2146 int ret = -EINTR;
2147 uint64_t btree_ptr = b->key.ptr[0];
2148 unsigned long seq = b->seq;
2149 struct keylist insert;
2150 bool upgrade = op->lock == -1;
2151
2152 bch_keylist_init(&insert);
2153
2154 if (upgrade) {
2155 rw_unlock(false, b);
2156 rw_lock(true, b, b->level);
2157
2158 if (b->key.ptr[0] != btree_ptr ||
2159 b->seq != seq + 1)
2160 goto out;
2161 }
2162
2163 SET_KEY_PTRS(check_key, 1);
2164 get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2165
2166 SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2167
2168 bch_keylist_add(&insert, check_key);
2169
Kent Overstreet1b207d82013-09-10 18:52:54 -07002170 ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
Kent Overstreete7c590e2013-09-10 18:39:16 -07002171
2172 BUG_ON(!ret && !bch_keylist_empty(&insert));
2173out:
2174 if (upgrade)
2175 downgrade_write(&b->lock);
2176 return ret;
2177}
2178
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002179struct btree_insert_op {
2180 struct btree_op op;
2181 struct keylist *keys;
2182 atomic_t *journal_ref;
2183 struct bkey *replace_key;
2184};
2185
Wei Yongjun08239ca2013-11-28 10:31:35 +08002186static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
Kent Overstreetcafe5632013-03-23 16:11:31 -07002187{
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002188 struct btree_insert_op *op = container_of(b_op,
2189 struct btree_insert_op, op);
Kent Overstreet403b6cd2013-07-24 17:22:44 -07002190
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002191 int ret = bch_btree_insert_node(b, &op->op, op->keys,
2192 op->journal_ref, op->replace_key);
2193 if (ret && !bch_keylist_empty(op->keys))
2194 return ret;
2195 else
2196 return MAP_DONE;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002197}
2198
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002199int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2200 atomic_t *journal_ref, struct bkey *replace_key)
Kent Overstreetcafe5632013-03-23 16:11:31 -07002201{
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002202 struct btree_insert_op op;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002203 int ret = 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002204
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002205 BUG_ON(current->bio_list);
Kent Overstreet4f3d4012013-09-10 18:46:36 -07002206 BUG_ON(bch_keylist_empty(keys));
Kent Overstreetcafe5632013-03-23 16:11:31 -07002207
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002208 bch_btree_op_init(&op.op, 0);
2209 op.keys = keys;
2210 op.journal_ref = journal_ref;
2211 op.replace_key = replace_key;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002212
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002213 while (!ret && !bch_keylist_empty(keys)) {
2214 op.op.lock = 0;
2215 ret = bch_btree_map_leaf_nodes(&op.op, c,
2216 &START_KEY(keys->keys),
2217 btree_insert_fn);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002218 }
2219
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002220 if (ret) {
2221 struct bkey *k;
2222
2223 pr_err("error %i", ret);
2224
2225 while ((k = bch_keylist_pop(keys)))
Kent Overstreet3a3b6a42013-07-24 16:46:42 -07002226 bkey_put(c, k);
Kent Overstreetcc7b8812013-07-24 18:07:22 -07002227 } else if (op.op.insert_collision)
2228 ret = -ESRCH;
Kent Overstreet6054c6d2013-07-24 18:06:22 -07002229
Kent Overstreetcafe5632013-03-23 16:11:31 -07002230 return ret;
2231}
2232
2233void bch_btree_set_root(struct btree *b)
2234{
2235 unsigned i;
Kent Overstreete49c7c32013-06-26 17:25:38 -07002236 struct closure cl;
2237
2238 closure_init_stack(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002239
Kent Overstreetc37511b2013-04-26 15:39:55 -07002240 trace_bcache_btree_set_root(b);
2241
Kent Overstreetcafe5632013-03-23 16:11:31 -07002242 BUG_ON(!b->written);
2243
2244 for (i = 0; i < KEY_PTRS(&b->key); i++)
2245 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2246
2247 mutex_lock(&b->c->bucket_lock);
2248 list_del_init(&b->list);
2249 mutex_unlock(&b->c->bucket_lock);
2250
2251 b->c->root = b;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002252
Kent Overstreete49c7c32013-06-26 17:25:38 -07002253 bch_journal_meta(b->c, &cl);
2254 closure_sync(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002255}
2256
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002257/* Map across nodes or keys */
2258
2259static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2260 struct bkey *from,
2261 btree_map_nodes_fn *fn, int flags)
2262{
2263 int ret = MAP_CONTINUE;
2264
2265 if (b->level) {
2266 struct bkey *k;
2267 struct btree_iter iter;
2268
Kent Overstreetc052dd92013-11-11 17:35:24 -08002269 bch_btree_iter_init(&b->keys, &iter, from);
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002270
Kent Overstreeta85e9682013-12-20 17:28:16 -08002271 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002272 bch_ptr_bad))) {
2273 ret = btree(map_nodes_recurse, k, b,
2274 op, from, fn, flags);
2275 from = NULL;
2276
2277 if (ret != MAP_CONTINUE)
2278 return ret;
2279 }
2280 }
2281
2282 if (!b->level || flags == MAP_ALL_NODES)
2283 ret = fn(op, b);
2284
2285 return ret;
2286}
2287
2288int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2289 struct bkey *from, btree_map_nodes_fn *fn, int flags)
2290{
Kent Overstreetb54d6932013-07-24 18:04:18 -07002291 return btree_root(map_nodes_recurse, c, op, from, fn, flags);
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002292}
2293
2294static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2295 struct bkey *from, btree_map_keys_fn *fn,
2296 int flags)
2297{
2298 int ret = MAP_CONTINUE;
2299 struct bkey *k;
2300 struct btree_iter iter;
2301
Kent Overstreetc052dd92013-11-11 17:35:24 -08002302 bch_btree_iter_init(&b->keys, &iter, from);
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002303
Kent Overstreeta85e9682013-12-20 17:28:16 -08002304 while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002305 ret = !b->level
2306 ? fn(op, b, k)
2307 : btree(map_keys_recurse, k, b, op, from, fn, flags);
2308 from = NULL;
2309
2310 if (ret != MAP_CONTINUE)
2311 return ret;
2312 }
2313
2314 if (!b->level && (flags & MAP_END_KEY))
2315 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2316 KEY_OFFSET(&b->key), 0));
2317
2318 return ret;
2319}
2320
2321int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2322 struct bkey *from, btree_map_keys_fn *fn, int flags)
2323{
Kent Overstreetb54d6932013-07-24 18:04:18 -07002324 return btree_root(map_keys_recurse, c, op, from, fn, flags);
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002325}
2326
Kent Overstreetcafe5632013-03-23 16:11:31 -07002327/* Keybuf code */
2328
2329static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2330{
2331 /* Overlapping keys compare equal */
2332 if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2333 return -1;
2334 if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2335 return 1;
2336 return 0;
2337}
2338
2339static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2340 struct keybuf_key *r)
2341{
2342 return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2343}
2344
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002345struct refill {
2346 struct btree_op op;
Kent Overstreet48a915a2013-10-31 15:43:22 -07002347 unsigned nr_found;
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002348 struct keybuf *buf;
2349 struct bkey *end;
2350 keybuf_pred_fn *pred;
2351};
2352
2353static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2354 struct bkey *k)
Kent Overstreetcafe5632013-03-23 16:11:31 -07002355{
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002356 struct refill *refill = container_of(op, struct refill, op);
2357 struct keybuf *buf = refill->buf;
2358 int ret = MAP_CONTINUE;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002359
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002360 if (bkey_cmp(k, refill->end) >= 0) {
2361 ret = MAP_DONE;
2362 goto out;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002363 }
2364
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002365 if (!KEY_SIZE(k)) /* end key */
2366 goto out;
2367
2368 if (refill->pred(buf, k)) {
2369 struct keybuf_key *w;
2370
2371 spin_lock(&buf->lock);
2372
2373 w = array_alloc(&buf->freelist);
2374 if (!w) {
2375 spin_unlock(&buf->lock);
2376 return MAP_DONE;
2377 }
2378
2379 w->private = NULL;
2380 bkey_copy(&w->key, k);
2381
2382 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2383 array_free(&buf->freelist, w);
Kent Overstreet48a915a2013-10-31 15:43:22 -07002384 else
2385 refill->nr_found++;
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002386
2387 if (array_freelist_empty(&buf->freelist))
2388 ret = MAP_DONE;
2389
2390 spin_unlock(&buf->lock);
2391 }
2392out:
2393 buf->last_scanned = *k;
2394 return ret;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002395}
2396
2397void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
Kent Overstreet72c27062013-06-05 06:24:39 -07002398 struct bkey *end, keybuf_pred_fn *pred)
Kent Overstreetcafe5632013-03-23 16:11:31 -07002399{
2400 struct bkey start = buf->last_scanned;
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002401 struct refill refill;
Kent Overstreetcafe5632013-03-23 16:11:31 -07002402
2403 cond_resched();
2404
Kent Overstreetb54d6932013-07-24 18:04:18 -07002405 bch_btree_op_init(&refill.op, -1);
Kent Overstreet48a915a2013-10-31 15:43:22 -07002406 refill.nr_found = 0;
2407 refill.buf = buf;
2408 refill.end = end;
2409 refill.pred = pred;
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002410
2411 bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2412 refill_keybuf_fn, MAP_END_KEY);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002413
Kent Overstreet48a915a2013-10-31 15:43:22 -07002414 trace_bcache_keyscan(refill.nr_found,
2415 KEY_INODE(&start), KEY_OFFSET(&start),
2416 KEY_INODE(&buf->last_scanned),
2417 KEY_OFFSET(&buf->last_scanned));
Kent Overstreetcafe5632013-03-23 16:11:31 -07002418
2419 spin_lock(&buf->lock);
2420
2421 if (!RB_EMPTY_ROOT(&buf->keys)) {
2422 struct keybuf_key *w;
2423 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2424 buf->start = START_KEY(&w->key);
2425
2426 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2427 buf->end = w->key;
2428 } else {
2429 buf->start = MAX_KEY;
2430 buf->end = MAX_KEY;
2431 }
2432
2433 spin_unlock(&buf->lock);
2434}
2435
2436static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2437{
2438 rb_erase(&w->node, &buf->keys);
2439 array_free(&buf->freelist, w);
2440}
2441
2442void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2443{
2444 spin_lock(&buf->lock);
2445 __bch_keybuf_del(buf, w);
2446 spin_unlock(&buf->lock);
2447}
2448
2449bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2450 struct bkey *end)
2451{
2452 bool ret = false;
2453 struct keybuf_key *p, *w, s;
2454 s.key = *start;
2455
2456 if (bkey_cmp(end, &buf->start) <= 0 ||
2457 bkey_cmp(start, &buf->end) >= 0)
2458 return false;
2459
2460 spin_lock(&buf->lock);
2461 w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2462
2463 while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2464 p = w;
2465 w = RB_NEXT(w, node);
2466
2467 if (p->private)
2468 ret = true;
2469 else
2470 __bch_keybuf_del(buf, p);
2471 }
2472
2473 spin_unlock(&buf->lock);
2474 return ret;
2475}
2476
2477struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2478{
2479 struct keybuf_key *w;
2480 spin_lock(&buf->lock);
2481
2482 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2483
2484 while (w && w->private)
2485 w = RB_NEXT(w, node);
2486
2487 if (w)
2488 w->private = ERR_PTR(-EINTR);
2489
2490 spin_unlock(&buf->lock);
2491 return w;
2492}
2493
2494struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
Kent Overstreet48dad8b2013-09-10 18:48:51 -07002495 struct keybuf *buf,
2496 struct bkey *end,
2497 keybuf_pred_fn *pred)
Kent Overstreetcafe5632013-03-23 16:11:31 -07002498{
2499 struct keybuf_key *ret;
2500
2501 while (1) {
2502 ret = bch_keybuf_next(buf);
2503 if (ret)
2504 break;
2505
2506 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2507 pr_debug("scan finished");
2508 break;
2509 }
2510
Kent Overstreet72c27062013-06-05 06:24:39 -07002511 bch_refill_keybuf(c, buf, end, pred);
Kent Overstreetcafe5632013-03-23 16:11:31 -07002512 }
2513
2514 return ret;
2515}
2516
Kent Overstreet72c27062013-06-05 06:24:39 -07002517void bch_keybuf_init(struct keybuf *buf)
Kent Overstreetcafe5632013-03-23 16:11:31 -07002518{
Kent Overstreetcafe5632013-03-23 16:11:31 -07002519 buf->last_scanned = MAX_KEY;
2520 buf->keys = RB_ROOT;
2521
2522 spin_lock_init(&buf->lock);
2523 array_allocator_init(&buf->freelist);
2524}