Line data Source code
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.
4 :
5 : #ifndef BITCOIN_CUCKOOCACHE_H
6 : #define BITCOIN_CUCKOOCACHE_H
7 :
8 : #include <util/fastrange.h>
9 :
10 : #include <algorithm> // std::find
11 : #include <array>
12 : #include <atomic>
13 : #include <cmath>
14 : #include <cstring>
15 : #include <limits>
16 : #include <memory>
17 : #include <optional>
18 : #include <utility>
19 : #include <vector>
20 :
21 :
22 : /** High-performance cache primitives.
23 : *
24 : * Summary:
25 : *
26 : * 1. @ref bit_packed_atomic_flags is bit-packed atomic flags for garbage collection
27 : *
28 : * 2. @ref cache is a cache which is performant in memory usage and lookup speed. It
29 : * is lockfree for erase operations. Elements are lazily erased on the next insert.
30 : */
31 : namespace CuckooCache
32 : {
33 : /** @ref bit_packed_atomic_flags implements a container for garbage collection flags
34 : * that is only thread unsafe on calls to setup. This class bit-packs collection
35 : * flags for memory efficiency.
36 : *
37 : * All operations are `std::memory_order_relaxed` so external mechanisms must
38 : * ensure that writes and reads are properly synchronized.
39 : *
40 : * On setup(n), all bits up to `n` are marked as collected.
41 : *
42 : * Under the hood, because it is an 8-bit type, it makes sense to use a multiple
43 : * of 8 for setup, but it will be safe if that is not the case as well.
44 : */
45 : class bit_packed_atomic_flags
46 : {
47 : std::unique_ptr<std::atomic<uint8_t>[]> mem;
48 :
49 : public:
50 : /** No default constructor, as there must be some size. */
51 : bit_packed_atomic_flags() = delete;
52 :
53 : /**
54 : * bit_packed_atomic_flags constructor creates memory to sufficiently
55 : * keep track of garbage collection information for `size` entries.
56 : *
57 : * @param size the number of elements to allocate space for
58 : *
59 : * @post bit_set, bit_unset, and bit_is_set function properly forall x. x <
60 : * size
61 : * @post All calls to bit_is_set (without subsequent bit_unset) will return
62 : * true.
63 : */
64 6 : explicit bit_packed_atomic_flags(uint32_t size)
65 : {
66 : // pad out the size if needed
67 6 : size = (size + 7) / 8;
68 6 : mem.reset(new std::atomic<uint8_t>[size]);
69 131078 : for (uint32_t i = 0; i < size; ++i)
70 131072 : mem[i].store(0xFF);
71 6 : };
72 :
73 : /** setup marks all entries and ensures that bit_packed_atomic_flags can store
74 : * at least `b` entries.
75 : *
76 : * @param b the number of elements to allocate space for
77 : * @post bit_set, bit_unset, and bit_is_set function properly forall x. x <
78 : * b
79 : * @post All calls to bit_is_set (without subsequent bit_unset) will return
80 : * true.
81 : */
82 2 : inline void setup(uint32_t b)
83 : {
84 2 : bit_packed_atomic_flags d(b);
85 2 : std::swap(mem, d.mem);
86 2 : }
87 :
88 : /** bit_set sets an entry as discardable.
89 : *
90 : * @param s the index of the entry to bit_set
91 : * @post immediately subsequent call (assuming proper external memory
92 : * ordering) to bit_is_set(s) == true.
93 : */
94 0 : inline void bit_set(uint32_t s)
95 : {
96 0 : mem[s >> 3].fetch_or(uint8_t(1 << (s & 7)), std::memory_order_relaxed);
97 0 : }
98 :
99 : /** bit_unset marks an entry as something that should not be overwritten.
100 : *
101 : * @param s the index of the entry to bit_unset
102 : * @post immediately subsequent call (assuming proper external memory
103 : * ordering) to bit_is_set(s) == false.
104 : */
105 844 : inline void bit_unset(uint32_t s)
106 : {
107 844 : mem[s >> 3].fetch_and(uint8_t(~(1 << (s & 7))), std::memory_order_relaxed);
108 844 : }
109 :
110 : /** bit_is_set queries the table for discardability at `s`.
111 : *
112 : * @param s the index of the entry to read
113 : * @returns true if the bit at index `s` was set, false otherwise
114 : * */
115 844 : inline bool bit_is_set(uint32_t s) const
116 : {
117 844 : return (1 << (s & 7)) & mem[s >> 3].load(std::memory_order_relaxed);
118 : }
119 : };
120 :
121 : /** @ref cache implements a cache with properties similar to a cuckoo-set.
122 : *
123 : * The cache is able to hold up to `(~(uint32_t)0) - 1` elements.
124 : *
125 : * Read Operations:
126 : * - contains() for `erase=false`
127 : *
128 : * Read+Erase Operations:
129 : * - contains() for `erase=true`
130 : *
131 : * Erase Operations:
132 : * - allow_erase()
133 : *
134 : * Write Operations:
135 : * - setup()
136 : * - setup_bytes()
137 : * - insert()
138 : * - please_keep()
139 : *
140 : * Synchronization Free Operations:
141 : * - invalid()
142 : * - compute_hashes()
143 : *
144 : * User Must Guarantee:
145 : *
146 : * 1. Write requires synchronized access (e.g. a lock)
147 : * 2. Read requires no concurrent Write, synchronized with last insert.
148 : * 3. Erase requires no concurrent Write, synchronized with last insert.
149 : * 4. An Erase caller must release all memory before allowing a new Writer.
150 : *
151 : *
152 : * Note on function names:
153 : * - The name "allow_erase" is used because the real discard happens later.
154 : * - The name "please_keep" is used because elements may be erased anyways on insert.
155 : *
156 : * @tparam Element should be a movable and copyable type
157 : * @tparam Hash should be a function/callable which takes a template parameter
158 : * hash_select and an Element and extracts a hash from it. Should return
159 : * high-entropy uint32_t hashes for `Hash h; h<0>(e) ... h<7>(e)`.
160 : */
161 : template <typename Element, typename Hash>
162 : class cache
163 : {
164 : private:
165 : /** table stores all the elements */
166 : std::vector<Element> table;
167 :
168 : /** size stores the total available slots in the hash table */
169 4 : uint32_t size{0};
170 :
171 : /** The bit_packed_atomic_flags array is marked mutable because we want
172 : * garbage collection to be allowed to occur from const methods */
173 : mutable bit_packed_atomic_flags collection_flags;
174 :
175 : /** epoch_flags tracks how recently an element was inserted into
176 : * the cache. true denotes recent, false denotes not-recent. See insert()
177 : * method for full semantics.
178 : */
179 : mutable std::vector<bool> epoch_flags;
180 :
181 : /** epoch_heuristic_counter is used to determine when an epoch might be aged
182 : * & an expensive scan should be done. epoch_heuristic_counter is
183 : * decremented on insert and reset to the new number of inserts which would
184 : * cause the epoch to reach epoch_size when it reaches zero.
185 : */
186 4 : uint32_t epoch_heuristic_counter{0};
187 :
188 : /** epoch_size is set to be the number of elements supposed to be in a
189 : * epoch. When the number of non-erased elements in an epoch
190 : * exceeds epoch_size, a new epoch should be started and all
191 : * current entries demoted. epoch_size is set to be 45% of size because
192 : * we want to keep load around 90%, and we support 3 epochs at once --
193 : * one "dead" which has been erased, one "dying" which has been marked to be
194 : * erased next, and one "living" which new inserts add to.
195 : */
196 4 : uint32_t epoch_size{0};
197 :
198 : /** depth_limit determines how many elements insert should try to replace.
199 : * Should be set to log2(n).
200 : */
201 4 : uint8_t depth_limit{0};
202 :
203 : /** hash_function is a const instance of the hash function. It cannot be
204 : * static or initialized at call time as it may have internal state (such as
205 : * a nonce).
206 : */
207 : const Hash hash_function;
208 :
209 : /** compute_hashes is convenience for not having to write out this
210 : * expression everywhere we use the hash values of an Element.
211 : *
212 : * We need to map the 32-bit input hash onto a hash bucket in a range [0, size) in a
213 : * manner which preserves as much of the hash's uniformity as possible. Ideally
214 : * this would be done by bitmasking but the size is usually not a power of two.
215 : *
216 : * The naive approach would be to use a mod -- which isn't perfectly uniform but so
217 : * long as the hash is much larger than size it is not that bad. Unfortunately,
218 : * mod/division is fairly slow on ordinary microprocessors (e.g. 90-ish cycles on
219 : * haswell, ARM doesn't even have an instruction for it.); when the divisor is a
220 : * constant the compiler will do clever tricks to turn it into a multiply+add+shift,
221 : * but size is a run-time value so the compiler can't do that here.
222 : *
223 : * One option would be to implement the same trick the compiler uses and compute the
224 : * constants for exact division based on the size, as described in "{N}-bit Unsigned
225 : * Division via {N}-bit Multiply-Add" by Arch D. Robison in 2005. But that code is
226 : * somewhat complicated and the result is still slower than an even simpler option:
227 : * see the FastRange32 function in util/fastrange.h.
228 : *
229 : * The resulting non-uniformity is also more equally distributed which would be
230 : * advantageous for something like linear probing, though it shouldn't matter
231 : * one way or the other for a cuckoo table.
232 : *
233 : * The primary disadvantage of this approach is increased intermediate precision is
234 : * required but for a 32-bit random number we only need the high 32 bits of a
235 : * 32*32->64 multiply, which means the operation is reasonably fast even on a
236 : * typical 32-bit processor.
237 : *
238 : * @param e The element whose hashes will be returned
239 : * @returns Deterministic hashes derived from `e` uniformly mapped onto the range [0, size)
240 : */
241 7748 : inline std::array<uint32_t, 8> compute_hashes(const Element& e) const
242 : {
243 61984 : return {{FastRange32(hash_function.template operator()<0>(e), size),
244 7748 : FastRange32(hash_function.template operator()<1>(e), size),
245 7748 : FastRange32(hash_function.template operator()<2>(e), size),
246 7748 : FastRange32(hash_function.template operator()<3>(e), size),
247 7748 : FastRange32(hash_function.template operator()<4>(e), size),
248 7748 : FastRange32(hash_function.template operator()<5>(e), size),
249 7748 : FastRange32(hash_function.template operator()<6>(e), size),
250 7748 : FastRange32(hash_function.template operator()<7>(e), size)}};
251 : }
252 :
253 : /** invalid returns a special index that can never be inserted to
254 : * @returns the special constexpr index that can never be inserted to */
255 844 : constexpr uint32_t invalid() const
256 : {
257 844 : return ~(uint32_t)0;
258 : }
259 :
260 : /** allow_erase marks the element at index `n` as discardable. Threadsafe
261 : * without any concurrent insert.
262 : * @param n the index to allow erasure of
263 : */
264 0 : inline void allow_erase(uint32_t n) const
265 : {
266 0 : collection_flags.bit_set(n);
267 0 : }
268 :
269 : /** please_keep marks the element at index `n` as an entry that should be kept.
270 : * Threadsafe without any concurrent insert.
271 : * @param n the index to prioritize keeping
272 : */
273 844 : inline void please_keep(uint32_t n) const
274 : {
275 844 : collection_flags.bit_unset(n);
276 844 : }
277 :
278 : /** epoch_check handles the changing of epochs for elements stored in the
279 : * cache. epoch_check should be run before every insert.
280 : *
281 : * First, epoch_check decrements and checks the cheap heuristic, and then does
282 : * a more expensive scan if the cheap heuristic runs out. If the expensive
283 : * scan succeeds, the epochs are aged and old elements are allow_erased. The
284 : * cheap heuristic is reset to retrigger after the worst case growth of the
285 : * current epoch's elements would exceed the epoch_size.
286 : */
287 844 : void epoch_check()
288 : {
289 844 : if (epoch_heuristic_counter != 0) {
290 844 : --epoch_heuristic_counter;
291 844 : return;
292 : }
293 : // count the number of elements from the latest epoch which
294 : // have not been erased.
295 0 : uint32_t epoch_unused_count = 0;
296 0 : for (uint32_t i = 0; i < size; ++i)
297 0 : epoch_unused_count += epoch_flags[i] &&
298 0 : !collection_flags.bit_is_set(i);
299 : // If there are more non-deleted entries in the current epoch than the
300 : // epoch size, then allow_erase on all elements in the old epoch (marked
301 : // false) and move all elements in the current epoch to the old epoch
302 : // but do not call allow_erase on their indices.
303 0 : if (epoch_unused_count >= epoch_size) {
304 0 : for (uint32_t i = 0; i < size; ++i)
305 0 : if (epoch_flags[i])
306 0 : epoch_flags[i] = false;
307 : else
308 0 : allow_erase(i);
309 0 : epoch_heuristic_counter = epoch_size;
310 0 : } else
311 : // reset the epoch_heuristic_counter to next do a scan when worst
312 : // case behavior (no intermittent erases) would exceed epoch size,
313 : // with a reasonable minimum scan size.
314 : // Ordinarily, we would have to sanity check std::min(epoch_size,
315 : // epoch_unused_count), but we already know that `epoch_unused_count
316 : // < epoch_size` in this branch
317 0 : epoch_heuristic_counter = std::max(1u, std::max(epoch_size / 16,
318 0 : epoch_size - epoch_unused_count));
319 844 : }
320 :
321 : public:
322 : /** You must always construct a cache with some elements via a subsequent
323 : * call to setup or setup_bytes, otherwise operations may segfault.
324 : */
325 8 : cache() : table(), collection_flags(0), epoch_flags(), hash_function()
326 : {
327 4 : }
328 :
329 : /** setup initializes the container to store no more than new_size
330 : * elements and no less than 2 elements.
331 : *
332 : * setup should only be called once.
333 : *
334 : * @param new_size the desired number of elements to store
335 : * @returns the maximum number of elements storable
336 : */
337 2 : uint32_t setup(uint32_t new_size)
338 : {
339 : // depth_limit must be at least one otherwise errors can occur.
340 2 : size = std::max<uint32_t>(2, new_size);
341 2 : depth_limit = static_cast<uint8_t>(std::log2(static_cast<float>(size)));
342 2 : table.resize(size);
343 2 : collection_flags.setup(size);
344 2 : epoch_flags.resize(size);
345 : // Set to 45% as described above
346 2 : epoch_size = std::max(uint32_t{1}, (45 * size) / 100);
347 : // Initially set to wait for a whole epoch
348 2 : epoch_heuristic_counter = epoch_size;
349 2 : return size;
350 : }
351 :
352 : /** setup_bytes is a convenience function which accounts for internal memory
353 : * usage when deciding how many elements to store. It isn't perfect because
354 : * it doesn't account for any overhead (struct size, MallocUsage, collection
355 : * and epoch flags). This was done to simplify selecting a power of two
356 : * size. In the expected use case, an extra two bits per entry should be
357 : * negligible compared to the size of the elements.
358 : *
359 : * @param bytes the approximate number of bytes to use for this data
360 : * structure
361 : * @returns A pair of the maximum number of elements storable (see setup()
362 : * documentation for more detail) and the approxmiate total size of these
363 : * elements in bytes or std::nullopt if the size requested is too large.
364 : */
365 2 : std::optional<std::pair<uint32_t, size_t>> setup_bytes(size_t bytes)
366 : {
367 2 : size_t requested_num_elems = bytes / sizeof(Element);
368 2 : if (std::numeric_limits<uint32_t>::max() < requested_num_elems) {
369 0 : return std::nullopt;
370 : }
371 :
372 2 : auto num_elems = setup(bytes/sizeof(Element));
373 :
374 2 : size_t approx_size_bytes = num_elems * sizeof(Element);
375 2 : return std::make_pair(num_elems, approx_size_bytes);
376 2 : }
377 :
378 : /** insert loops at most depth_limit times trying to insert a hash
379 : * at various locations in the table via a variant of the Cuckoo Algorithm
380 : * with eight hash locations.
381 : *
382 : * It drops the last tried element if it runs out of depth before
383 : * encountering an open slot.
384 : *
385 : * Thus:
386 : *
387 : * ```
388 : * insert(x);
389 : * return contains(x, false);
390 : * ```
391 : *
392 : * is not guaranteed to return true.
393 : *
394 : * @param e the element to insert
395 : * @post one of the following: All previously inserted elements and e are
396 : * now in the table, one previously inserted element is evicted from the
397 : * table, the entry attempted to be inserted is evicted.
398 : */
399 844 : inline void insert(Element e)
400 : {
401 844 : epoch_check();
402 844 : uint32_t last_loc = invalid();
403 844 : bool last_epoch = true;
404 844 : std::array<uint32_t, 8> locs = compute_hashes(e);
405 : // Make sure we have not already inserted this element
406 : // If we have, make sure that it does not get deleted
407 7596 : for (const uint32_t loc : locs)
408 6752 : if (table[loc] == e) {
409 0 : please_keep(loc);
410 0 : epoch_flags[loc] = last_epoch;
411 0 : return;
412 : }
413 844 : for (uint8_t depth = 0; depth < depth_limit; ++depth) {
414 : // First try to insert to an empty slot, if one exists
415 844 : for (const uint32_t loc : locs) {
416 844 : if (!collection_flags.bit_is_set(loc))
417 0 : continue;
418 844 : table[loc] = std::move(e);
419 844 : please_keep(loc);
420 844 : epoch_flags[loc] = last_epoch;
421 844 : return;
422 : }
423 : /** Swap with the element at the location that was
424 : * not the last one looked at. Example:
425 : *
426 : * 1. On first iteration, last_loc == invalid(), find returns last, so
427 : * last_loc defaults to locs[0].
428 : * 2. On further iterations, where last_loc == locs[k], last_loc will
429 : * go to locs[k+1 % 8], i.e., next of the 8 indices wrapping around
430 : * to 0 if needed.
431 : *
432 : * This prevents moving the element we just put in.
433 : *
434 : * The swap is not a move -- we must switch onto the evicted element
435 : * for the next iteration.
436 : */
437 0 : last_loc = locs[(1 + (std::find(locs.begin(), locs.end(), last_loc) - locs.begin())) & 7];
438 0 : std::swap(table[last_loc], e);
439 : // Can't std::swap a std::vector<bool>::reference and a bool&.
440 0 : bool epoch = last_epoch;
441 0 : last_epoch = epoch_flags[last_loc];
442 0 : epoch_flags[last_loc] = epoch;
443 :
444 : // Recompute the locs -- unfortunately happens one too many times!
445 0 : locs = compute_hashes(e);
446 0 : }
447 844 : }
448 :
449 : /** contains iterates through the hash locations for a given element
450 : * and checks to see if it is present.
451 : *
452 : * contains does not check garbage collected state (in other words,
453 : * garbage is only collected when the space is needed), so:
454 : *
455 : * ```
456 : * insert(x);
457 : * if (contains(x, true))
458 : * return contains(x, false);
459 : * else
460 : * return true;
461 : * ```
462 : *
463 : * executed on a single thread will always return true!
464 : *
465 : * This is a great property for re-org performance for example.
466 : *
467 : * contains returns a bool set true if the element was found.
468 : *
469 : * @param e the element to check
470 : * @param erase whether to attempt setting the garbage collect flag
471 : *
472 : * @post if erase is true and the element is found, then the garbage collect
473 : * flag is set
474 : * @returns true if the element is found, false otherwise
475 : */
476 6904 : inline bool contains(const Element& e, const bool erase) const
477 : {
478 6904 : std::array<uint32_t, 8> locs = compute_hashes(e);
479 41928 : for (const uint32_t loc : locs)
480 37550 : if (table[loc] == e) {
481 2526 : if (erase)
482 0 : allow_erase(loc);
483 2526 : return true;
484 : }
485 4378 : return false;
486 6904 : }
487 : };
488 : } // namespace CuckooCache
489 :
490 : #endif // BITCOIN_CUCKOOCACHE_H
|
