7 * A bkey contains a key, a size field, a variable number of pointers, and some
10 * We use two different functions for validating bkeys, bch_ptr_invalid and
13 * bch_ptr_invalid() primarily filters out keys and pointers that would be
14 * invalid due to some sort of bug, whereas bch_ptr_bad() filters out keys and
15 * pointer that occur in normal practice but don't point to real data.
17 * The one exception to the rule that ptr_invalid() filters out invalid keys is
18 * that it also filters out keys of size 0 - these are keys that have been
19 * completely overwritten. It'd be safe to delete these in memory while leaving
20 * them on disk, just unnecessary work - so we filter them out when resorting
23 * We can't filter out stale keys when we're resorting, because garbage
24 * collection needs to find them to ensure bucket gens don't wrap around -
25 * unless we're rewriting the btree node those stale keys still exist on disk.
27 * We also implement functions here for removing some number of sectors from the
28 * front or the back of a bkey - this is mainly used for fixing overlapping
29 * extents, by removing the overlapping sectors from the older key.
33 * A bset is an array of bkeys laid out contiguously in memory in sorted order,
34 * along with a header. A btree node is made up of a number of these, written at
37 * There could be many of them on disk, but we never allow there to be more than
38 * 4 in memory - we lazily resort as needed.
40 * We implement code here for creating and maintaining auxiliary search trees
41 * (described below) for searching an individial bset, and on top of that we
42 * implement a btree iterator.
46 * Most of the code in bcache doesn't care about an individual bset - it needs
47 * to search entire btree nodes and iterate over them in sorted order.
49 * The btree iterator code serves both functions; it iterates through the keys
50 * in a btree node in sorted order, starting from either keys after a specific
51 * point (if you pass it a search key) or the start of the btree node.
53 * AUXILIARY SEARCH TREES:
55 * Since keys are variable length, we can't use a binary search on a bset - we
56 * wouldn't be able to find the start of the next key. But binary searches are
57 * slow anyways, due to terrible cache behaviour; bcache originally used binary
58 * searches and that code topped out at under 50k lookups/second.
60 * So we need to construct some sort of lookup table. Since we only insert keys
61 * into the last (unwritten) set, most of the keys within a given btree node are
62 * usually in sets that are mostly constant. We use two different types of
63 * lookup tables to take advantage of this.
65 * Both lookup tables share in common that they don't index every key in the
66 * set; they index one key every BSET_CACHELINE bytes, and then a linear search
67 * is used for the rest.
69 * For sets that have been written to disk and are no longer being inserted
70 * into, we construct a binary search tree in an array - traversing a binary
71 * search tree in an array gives excellent locality of reference and is very
72 * fast, since both children of any node are adjacent to each other in memory
73 * (and their grandchildren, and great grandchildren...) - this means
74 * prefetching can be used to great effect.
76 * It's quite useful performance wise to keep these nodes small - not just
77 * because they're more likely to be in L2, but also because we can prefetch
78 * more nodes on a single cacheline and thus prefetch more iterations in advance
79 * when traversing this tree.
81 * Nodes in the auxiliary search tree must contain both a key to compare against
82 * (we don't want to fetch the key from the set, that would defeat the purpose),
83 * and a pointer to the key. We use a few tricks to compress both of these.
85 * To compress the pointer, we take advantage of the fact that one node in the
86 * search tree corresponds to precisely BSET_CACHELINE bytes in the set. We have
87 * a function (to_inorder()) that takes the index of a node in a binary tree and
88 * returns what its index would be in an inorder traversal, so we only have to
89 * store the low bits of the offset.
91 * The key is 84 bits (KEY_DEV + key->key, the offset on the device). To
92 * compress that, we take advantage of the fact that when we're traversing the
93 * search tree at every iteration we know that both our search key and the key
94 * we're looking for lie within some range - bounded by our previous
95 * comparisons. (We special case the start of a search so that this is true even
96 * at the root of the tree).
98 * So we know the key we're looking for is between a and b, and a and b don't
99 * differ higher than bit 50, we don't need to check anything higher than bit
102 * We don't usually need the rest of the bits, either; we only need enough bits
103 * to partition the key range we're currently checking. Consider key n - the
104 * key our auxiliary search tree node corresponds to, and key p, the key
105 * immediately preceding n. The lowest bit we need to store in the auxiliary
106 * search tree is the highest bit that differs between n and p.
108 * Note that this could be bit 0 - we might sometimes need all 80 bits to do the
109 * comparison. But we'd really like our nodes in the auxiliary search tree to be
112 * The solution is to make them fixed size, and when we're constructing a node
113 * check if p and n differed in the bits we needed them to. If they don't we
114 * flag that node, and when doing lookups we fallback to comparing against the
115 * real key. As long as this doesn't happen to often (and it seems to reliably
116 * happen a bit less than 1% of the time), we win - even on failures, that key
117 * is then more likely to be in cache than if we were doing binary searches all
118 * the way, since we're touching so much less memory.
120 * The keys in the auxiliary search tree are stored in (software) floating
121 * point, with an exponent and a mantissa. The exponent needs to be big enough
122 * to address all the bits in the original key, but the number of bits in the
123 * mantissa is somewhat arbitrary; more bits just gets us fewer failures.
125 * We need 7 bits for the exponent and 3 bits for the key's offset (since keys
126 * are 8 byte aligned); using 22 bits for the mantissa means a node is 4 bytes.
127 * We need one node per 128 bytes in the btree node, which means the auxiliary
128 * search trees take up 3% as much memory as the btree itself.
130 * Constructing these auxiliary search trees is moderately expensive, and we
131 * don't want to be constantly rebuilding the search tree for the last set
132 * whenever we insert another key into it. For the unwritten set, we use a much
133 * simpler lookup table - it's just a flat array, so index i in the lookup table
134 * corresponds to the i range of BSET_CACHELINE bytes in the set. Indexing
135 * within each byte range works the same as with the auxiliary search trees.
137 * These are much easier to keep up to date when we insert a key - we do it
138 * somewhat lazily; when we shift a key up we usually just increment the pointer
139 * to it, only when it would overflow do we go to the trouble of finding the
140 * first key in that range of bytes again.
143 /* Btree key comparison/iteration */
147 struct btree_iter_set {
148 struct bkey *k, *end;
154 * We construct a binary tree in an array as if the array
155 * started at 1, so that things line up on the same cachelines
156 * better: see comments in bset.c at cacheline_to_bkey() for
160 /* size of the binary tree and prev array */
163 /* function of size - precalculated for to_inorder() */
166 /* copy of the last key in the set */
168 struct bkey_float *tree;
171 * The nodes in the bset tree point to specific keys - this
172 * array holds the sizes of the previous key.
174 * Conceptually it's a member of struct bkey_float, but we want
175 * to keep bkey_float to 4 bytes and prev isn't used in the fast
180 /* The actual btree node, with pointers to each sorted set */
184 static __always_inline int64_t bkey_cmp(const struct bkey *l,
185 const struct bkey *r)
187 return unlikely(KEY_INODE(l) != KEY_INODE(r))
188 ? (int64_t) KEY_INODE(l) - (int64_t) KEY_INODE(r)
189 : (int64_t) KEY_OFFSET(l) - (int64_t) KEY_OFFSET(r);
192 static inline size_t bkey_u64s(const struct bkey *k)
194 BUG_ON(KEY_CSUM(k) > 1);
195 return 2 + KEY_PTRS(k) + (KEY_CSUM(k) ? 1 : 0);
198 static inline size_t bkey_bytes(const struct bkey *k)
200 return bkey_u64s(k) * sizeof(uint64_t);
203 static inline void bkey_copy(struct bkey *dest, const struct bkey *src)
205 memcpy(dest, src, bkey_bytes(src));
208 static inline void bkey_copy_key(struct bkey *dest, const struct bkey *src)
213 SET_KEY_INODE(dest, KEY_INODE(src));
214 SET_KEY_OFFSET(dest, KEY_OFFSET(src));
217 static inline struct bkey *bkey_next(const struct bkey *k)
219 uint64_t *d = (void *) k;
220 return (struct bkey *) (d + bkey_u64s(k));
232 /* Enough room for btree_split's keys without realloc */
233 #define KEYLIST_INLINE 16
234 uint64_t d[KEYLIST_INLINE];
237 static inline void bch_keylist_init(struct keylist *l)
239 l->top = (void *) (l->list = l->d);
242 static inline void bch_keylist_push(struct keylist *l)
244 l->top = bkey_next(l->top);
247 static inline void bch_keylist_add(struct keylist *l, struct bkey *k)
249 bkey_copy(l->top, k);
253 static inline bool bch_keylist_empty(struct keylist *l)
255 return l->top == (void *) l->list;
258 static inline void bch_keylist_free(struct keylist *l)
264 void bch_keylist_copy(struct keylist *, struct keylist *);
265 struct bkey *bch_keylist_pop(struct keylist *);
266 int bch_keylist_realloc(struct keylist *, int, struct cache_set *);
268 void bch_bkey_copy_single_ptr(struct bkey *, const struct bkey *,
270 bool __bch_cut_front(const struct bkey *, struct bkey *);
271 bool __bch_cut_back(const struct bkey *, struct bkey *);
273 static inline bool bch_cut_front(const struct bkey *where, struct bkey *k)
275 BUG_ON(bkey_cmp(where, k) > 0);
276 return __bch_cut_front(where, k);
279 static inline bool bch_cut_back(const struct bkey *where, struct bkey *k)
281 BUG_ON(bkey_cmp(where, &START_KEY(k)) < 0);
282 return __bch_cut_back(where, k);
285 const char *bch_ptr_status(struct cache_set *, const struct bkey *);
286 bool __bch_ptr_invalid(struct cache_set *, int level, const struct bkey *);
287 bool bch_ptr_bad(struct btree *, const struct bkey *);
289 static inline uint8_t gen_after(uint8_t a, uint8_t b)
292 return r > 128U ? 0 : r;
295 static inline uint8_t ptr_stale(struct cache_set *c, const struct bkey *k,
298 return gen_after(PTR_BUCKET(c, k, i)->gen, PTR_GEN(k, i));
301 static inline bool ptr_available(struct cache_set *c, const struct bkey *k,
304 return (PTR_DEV(k, i) < MAX_CACHES_PER_SET) && PTR_CACHE(c, k, i);
308 typedef bool (*ptr_filter_fn)(struct btree *, const struct bkey *);
310 struct bkey *bch_next_recurse_key(struct btree *, struct bkey *);
311 struct bkey *bch_btree_iter_next(struct btree_iter *);
312 struct bkey *bch_btree_iter_next_filter(struct btree_iter *,
313 struct btree *, ptr_filter_fn);
315 void bch_btree_iter_push(struct btree_iter *, struct bkey *, struct bkey *);
316 struct bkey *__bch_btree_iter_init(struct btree *, struct btree_iter *,
317 struct bkey *, struct bset_tree *);
320 #define BKEY_MID_BITS 3
321 #define BKEY_EXPONENT_BITS 7
322 #define BKEY_MANTISSA_BITS 22
323 #define BKEY_MANTISSA_MASK ((1 << BKEY_MANTISSA_BITS) - 1)
326 unsigned exponent:BKEY_EXPONENT_BITS;
327 unsigned m:BKEY_MID_BITS;
328 unsigned mantissa:BKEY_MANTISSA_BITS;
332 * BSET_CACHELINE was originally intended to match the hardware cacheline size -
333 * it used to be 64, but I realized the lookup code would touch slightly less
334 * memory if it was 128.
336 * It definites the number of bytes (in struct bset) per struct bkey_float in
337 * the auxiliar search tree - when we're done searching the bset_float tree we
338 * have this many bytes left that we do a linear search over.
340 * Since (after level 5) every level of the bset_tree is on a new cacheline,
341 * we're touching one fewer cacheline in the bset tree in exchange for one more
342 * cacheline in the linear search - but the linear search might stop before it
343 * gets to the second cacheline.
346 #define BSET_CACHELINE 128
347 #define bset_tree_space(b) (btree_data_space(b) / BSET_CACHELINE)
349 #define bset_tree_bytes(b) (bset_tree_space(b) * sizeof(struct bkey_float))
350 #define bset_prev_bytes(b) (bset_tree_space(b) * sizeof(uint8_t))
352 void bch_bset_init_next(struct btree *);
354 void bch_bset_fix_invalidated_key(struct btree *, struct bkey *);
355 void bch_bset_fix_lookup_table(struct btree *, struct bkey *);
357 struct bkey *__bch_bset_search(struct btree *, struct bset_tree *,
358 const struct bkey *);
360 static inline struct bkey *bch_bset_search(struct btree *b, struct bset_tree *t,
361 const struct bkey *search)
363 return search ? __bch_bset_search(b, t, search) : t->data->start;
366 bool bch_bkey_try_merge(struct btree *, struct bkey *, struct bkey *);
367 void bch_btree_sort_lazy(struct btree *);
368 void bch_btree_sort_into(struct btree *, struct btree *);
369 void bch_btree_sort_and_fix_extents(struct btree *, struct btree_iter *);
370 void bch_btree_sort_partial(struct btree *, unsigned);
372 static inline void bch_btree_sort(struct btree *b)
374 bch_btree_sort_partial(b, 0);
377 int bch_bset_print_stats(struct cache_set *, char *);