1 // Copyright (c) 2016 Jeremy Rubin
2 // Distributed under the MIT software license, see the accompanying
3 // file COPYING or http://www.opensource.org/licenses/mit-license.php.
5 #ifndef _BITCOIN_CUCKOOCACHE_H_
6 #define _BITCOIN_CUCKOOCACHE_H_
17 /** namespace CuckooCache provides high performance cache primitives
21 * 1) bit_packed_atomic_flags is bit-packed atomic flags for garbage collection
23 * 2) cache is a cache which is performant in memory usage and lookup speed. It
24 * is lockfree for erase operations. Elements are lazily erased on the next
29 /** bit_packed_atomic_flags implements a container for garbage collection flags
30 * that is only thread unsafe on calls to setup. This class bit-packs collection
31 * flags for memory efficiency.
33 * All operations are std::memory_order_relaxed so external mechanisms must
34 * ensure that writes and reads are properly synchronized.
36 * On setup(n), all bits up to n are marked as collected.
38 * Under the hood, because it is an 8-bit type, it makes sense to use a multiple
39 * of 8 for setup, but it will be safe if that is not the case as well.
42 class bit_packed_atomic_flags
44 std::unique_ptr
<std::atomic
<uint8_t>[]> mem
;
47 /** No default constructor as there must be some size */
48 bit_packed_atomic_flags() = delete;
51 * bit_packed_atomic_flags constructor creates memory to sufficiently
52 * keep track of garbage collection information for size entries.
54 * @param size the number of elements to allocate space for
56 * @post bit_set, bit_unset, and bit_is_set function properly forall x. x <
58 * @post All calls to bit_is_set (without subsequent bit_unset) will return
61 bit_packed_atomic_flags(uint32_t size
)
63 // pad out the size if needed
64 size
= (size
+ 7) / 8;
65 mem
.reset(new std::atomic
<uint8_t>[size
]);
66 for (uint32_t i
= 0; i
< size
; ++i
)
70 /** setup marks all entries and ensures that bit_packed_atomic_flags can store
71 * at least size entries
73 * @param b the number of elements to allocate space for
74 * @post bit_set, bit_unset, and bit_is_set function properly forall x. x <
76 * @post All calls to bit_is_set (without subsequent bit_unset) will return
79 inline void setup(uint32_t b
)
81 bit_packed_atomic_flags
d(b
);
82 std::swap(mem
, d
.mem
);
85 /** bit_set sets an entry as discardable.
87 * @param s the index of the entry to bit_set.
88 * @post immediately subsequent call (assuming proper external memory
89 * ordering) to bit_is_set(s) == true.
92 inline void bit_set(uint32_t s
)
94 mem
[s
>> 3].fetch_or(1 << (s
& 7), std::memory_order_relaxed
);
97 /** bit_unset marks an entry as something that should not be overwritten
99 * @param s the index of the entry to bit_unset.
100 * @post immediately subsequent call (assuming proper external memory
101 * ordering) to bit_is_set(s) == false.
103 inline void bit_unset(uint32_t s
)
105 mem
[s
>> 3].fetch_and(~(1 << (s
& 7)), std::memory_order_relaxed
);
108 /** bit_is_set queries the table for discardability at s
110 * @param s the index of the entry to read.
111 * @returns if the bit at index s was set.
113 inline bool bit_is_set(uint32_t s
) const
115 return (1 << (s
& 7)) & mem
[s
>> 3].load(std::memory_order_relaxed
);
119 /** cache implements a cache with properties similar to a cuckoo-set
121 * The cache is able to hold up to (~(uint32_t)0) - 1 elements.
124 * - contains(*, false)
126 * Read+Erase Operations:
127 * - contains(*, true)
138 * Synchronization Free Operations:
142 * User Must Guarantee:
144 * 1) Write Requires synchronized access (e.g., a lock)
145 * 2) Read Requires no concurrent Write, synchronized with the last insert.
146 * 3) Erase requires no concurrent Write, synchronized with last insert.
147 * 4) An Erase caller must release all memory before allowing a new Writer.
150 * Note on function names:
151 * - The name "allow_erase" is used because the real discard happens later.
152 * - The name "please_keep" is used because elements may be erased anyways on insert.
154 * @tparam Element should be a movable and copyable type
155 * @tparam Hash should be a function/callable which takes a template parameter
156 * hash_select and an Element and extracts a hash from it. Should return
157 * high-entropy hashes for `Hash h; h<0>(e) ... h<7>(e)`.
159 template <typename Element
, typename Hash
>
163 /** table stores all the elements */
164 std::vector
<Element
> table
;
166 /** size stores the total available slots in the hash table */
169 /** The bit_packed_atomic_flags array is marked mutable because we want
170 * garbage collection to be allowed to occur from const methods */
171 mutable bit_packed_atomic_flags collection_flags
;
173 /** epoch_flags tracks how recently an element was inserted into
174 * the cache. true denotes recent, false denotes not-recent. See insert()
175 * method for full semantics.
177 mutable std::vector
<bool> epoch_flags
;
179 /** epoch_heuristic_counter is used to determine when a epoch might be aged
180 * & an expensive scan should be done. epoch_heuristic_counter is
181 * decremented on insert and reset to the new number of inserts which would
182 * cause the epoch to reach epoch_size when it reaches zero.
184 uint32_t epoch_heuristic_counter
;
186 /** epoch_size is set to be the number of elements supposed to be in a
187 * epoch. When the number of non-erased elements in a epoch
188 * exceeds epoch_size, a new epoch should be started and all
189 * current entries demoted. epoch_size is set to be 45% of size because
190 * we want to keep load around 90%, and we support 3 epochs at once --
191 * one "dead" which has been erased, one "dying" which has been marked to be
192 * erased next, and one "living" which new inserts add to.
196 /** hash_mask should be set to appropriately mask out a hash such that every
197 * masked hash is [0,size), eg, if floor(log2(size)) == 20, then hash_mask
198 * should be (1<<20)-1
202 /** depth_limit determines how many elements insert should try to replace.
203 * Should be set to log2(n)*/
206 /** hash_function is a const instance of the hash function. It cannot be
207 * static or initialized at call time as it may have internal state (such as
210 const Hash hash_function
;
212 /** compute_hashes is convenience for not having to write out this
213 * expression everywhere we use the hash values of an Element.
215 * @param e the element whose hashes will be returned
216 * @returns std::array<uint32_t, 8> of deterministic hashes derived from e
218 inline std::array
<uint32_t, 8> compute_hashes(const Element
& e
) const
220 return {{hash_function
.template operator()<0>(e
) & hash_mask
,
221 hash_function
.template operator()<1>(e
) & hash_mask
,
222 hash_function
.template operator()<2>(e
) & hash_mask
,
223 hash_function
.template operator()<3>(e
) & hash_mask
,
224 hash_function
.template operator()<4>(e
) & hash_mask
,
225 hash_function
.template operator()<5>(e
) & hash_mask
,
226 hash_function
.template operator()<6>(e
) & hash_mask
,
227 hash_function
.template operator()<7>(e
) & hash_mask
}};
231 * @returns a constexpr index that can never be inserted to */
232 constexpr uint32_t invalid() const
237 /** allow_erase marks the element at index n as discardable. Threadsafe
238 * without any concurrent insert.
239 * @param n the index to allow erasure of
241 inline void allow_erase(uint32_t n
) const
243 collection_flags
.bit_set(n
);
246 /** please_keep marks the element at index n as an entry that should be kept.
247 * Threadsafe without any concurrent insert.
248 * @param n the index to prioritize keeping
250 inline void please_keep(uint32_t n
) const
252 collection_flags
.bit_unset(n
);
255 /** epoch_check handles the changing of epochs for elements stored in the
256 * cache. epoch_check should be run before every insert.
258 * First, epoch_check decrements and checks the cheap heuristic, and then does
259 * a more expensive scan if the cheap heuristic runs out. If the expensive
260 * scan succeeds, the epochs are aged and old elements are allow_erased. The
261 * cheap heuristic is reset to retrigger after the worst case growth of the
262 * current epoch's elements would exceed the epoch_size.
266 if (epoch_heuristic_counter
!= 0) {
267 --epoch_heuristic_counter
;
270 // count the number of elements from the latest epoch which
271 // have not been erased.
272 uint32_t epoch_unused_count
= 0;
273 for (uint32_t i
= 0; i
< size
; ++i
)
274 epoch_unused_count
+= epoch_flags
[i
] &&
275 !collection_flags
.bit_is_set(i
);
276 // If there are more non-deleted entries in the current epoch than the
277 // epoch size, then allow_erase on all elements in the old epoch (marked
278 // false) and move all elements in the current epoch to the old epoch
279 // but do not call allow_erase on their indices.
280 if (epoch_unused_count
>= epoch_size
) {
281 for (uint32_t i
= 0; i
< size
; ++i
)
283 epoch_flags
[i
] = false;
286 epoch_heuristic_counter
= epoch_size
;
288 // reset the epoch_heuristic_counter to next do a scan when worst
289 // case behavior (no intermittent erases) would exceed epoch size,
290 // with a reasonable minimum scan size.
291 // Ordinarily, we would have to sanity check std::min(epoch_size,
292 // epoch_unused_count), but we already know that `epoch_unused_count
293 // < epoch_size` in this branch
294 epoch_heuristic_counter
= std::max(1u, std::max(epoch_size
/ 16,
295 epoch_size
- epoch_unused_count
));
299 /** You must always construct a cache with some elements via a subsequent
300 * call to setup or setup_bytes, otherwise operations may segfault.
302 cache() : table(), size(), collection_flags(0), epoch_flags(),
303 epoch_heuristic_counter(), epoch_size(), depth_limit(0), hash_function()
307 /** setup initializes the container to store no more than new_size
308 * elements. setup rounds down to a power of two size.
310 * setup should only be called once.
312 * @param new_size the desired number of elements to store
313 * @returns the maximum number of elements storable
315 uint32_t setup(uint32_t new_size
)
317 // depth_limit must be at least one otherwise errors can occur.
318 depth_limit
= static_cast<uint8_t>(std::log2(static_cast<float>(std::max((uint32_t)2, new_size
))));
319 size
= 1 << depth_limit
;
322 collection_flags
.setup(size
);
323 epoch_flags
.resize(size
);
324 // Set to 45% as described above
325 epoch_size
= std::max((uint32_t)1, (45 * size
) / 100);
326 // Initially set to wait for a whole epoch
327 epoch_heuristic_counter
= epoch_size
;
331 /** setup_bytes is a convenience function which accounts for internal memory
332 * usage when deciding how many elements to store. It isn't perfect because
333 * it doesn't account for any overhead (struct size, MallocUsage, collection
334 * and epoch flags). This was done to simplify selecting a power of two
335 * size. In the expected use case, an extra two bits per entry should be
336 * negligible compared to the size of the elements.
338 * @param bytes the approximate number of bytes to use for this data
340 * @returns the maximum number of elements storable (see setup()
341 * documentation for more detail)
343 uint32_t setup_bytes(size_t bytes
)
345 return setup(bytes
/sizeof(Element
));
348 /** insert loops at most depth_limit times trying to insert a hash
349 * at various locations in the table via a variant of the Cuckoo Algorithm
350 * with eight hash locations.
352 * It drops the last tried element if it runs out of depth before
353 * encountering an open slot.
358 * return contains(x, false);
360 * is not guaranteed to return true.
362 * @param e the element to insert
363 * @post one of the following: All previously inserted elements and e are
364 * now in the table, one previously inserted element is evicted from the
365 * table, the entry attempted to be inserted is evicted.
368 inline void insert(Element e
)
371 uint32_t last_loc
= invalid();
372 bool last_epoch
= true;
373 std::array
<uint32_t, 8> locs
= compute_hashes(e
);
374 // Make sure we have not already inserted this element
375 // If we have, make sure that it does not get deleted
376 for (uint32_t loc
: locs
)
377 if (table
[loc
] == e
) {
379 epoch_flags
[loc
] = last_epoch
;
382 for (uint8_t depth
= 0; depth
< depth_limit
; ++depth
) {
383 // First try to insert to an empty slot, if one exists
384 for (uint32_t loc
: locs
) {
385 if (!collection_flags
.bit_is_set(loc
))
387 table
[loc
] = std::move(e
);
389 epoch_flags
[loc
] = last_epoch
;
392 /** Swap with the element at the location that was
393 * not the last one looked at. Example:
395 * 1) On first iteration, last_loc == invalid(), find returns last, so
396 * last_loc defaults to locs[0].
397 * 2) On further iterations, where last_loc == locs[k], last_loc will
398 * go to locs[k+1 % 8], i.e., next of the 8 indices wrapping around
401 * This prevents moving the element we just put in.
403 * The swap is not a move -- we must switch onto the evicted element
404 * for the next iteration.
406 last_loc
= locs
[(1 + (std::find(locs
.begin(), locs
.end(), last_loc
) - locs
.begin())) & 7];
407 std::swap(table
[last_loc
], e
);
408 // Can't std::swap a std::vector<bool>::reference and a bool&.
409 bool epoch
= last_epoch
;
410 last_epoch
= epoch_flags
[last_loc
];
411 epoch_flags
[last_loc
] = epoch
;
413 // Recompute the locs -- unfortunately happens one too many times!
414 locs
= compute_hashes(e
);
418 /* contains iterates through the hash locations for a given element
419 * and checks to see if it is present.
421 * contains does not check garbage collected state (in other words,
422 * garbage is only collected when the space is needed), so:
425 * if (contains(x, true))
426 * return contains(x, false);
430 * executed on a single thread will always return true!
432 * This is a great property for re-org performance for example.
434 * contains returns a bool set true if the element was found.
436 * @param e the element to check
439 * @post if erase is true and the element is found, then the garbage collect
441 * @returns true if the element is found, false otherwise
443 inline bool contains(const Element
& e
, const bool erase
) const
445 std::array
<uint32_t, 8> locs
= compute_hashes(e
);
446 for (uint32_t loc
: locs
)
447 if (table
[loc
] == e
) {
455 } // namespace CuckooCache