Bitcoin Core  27.99.0
P2P Digital Currency
cuckoocache.h
Go to the documentation of this file.
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 <utility>
18 #include <vector>
19 
20 
30 namespace CuckooCache
31 {
45 {
46  std::unique_ptr<std::atomic<uint8_t>[]> mem;
47 
48 public:
51 
63  explicit bit_packed_atomic_flags(uint32_t size)
64  {
65  // pad out the size if needed
66  size = (size + 7) / 8;
67  mem.reset(new std::atomic<uint8_t>[size]);
68  for (uint32_t i = 0; i < size; ++i)
69  mem[i].store(0xFF);
70  };
71 
81  inline void setup(uint32_t b)
82  {
84  std::swap(mem, d.mem);
85  }
86 
93  inline void bit_set(uint32_t s)
94  {
95  mem[s >> 3].fetch_or(uint8_t(1 << (s & 7)), std::memory_order_relaxed);
96  }
97 
104  inline void bit_unset(uint32_t s)
105  {
106  mem[s >> 3].fetch_and(uint8_t(~(1 << (s & 7))), std::memory_order_relaxed);
107  }
108 
114  inline bool bit_is_set(uint32_t s) const
115  {
116  return (1 << (s & 7)) & mem[s >> 3].load(std::memory_order_relaxed);
117  }
118 };
119 
160 template <typename Element, typename Hash>
161 class cache
162 {
163 private:
165  std::vector<Element> table;
166 
168  uint32_t size{0};
169 
173 
178  mutable std::vector<bool> epoch_flags;
179 
186 
195  uint32_t epoch_size{0};
196 
200  uint8_t depth_limit{0};
201 
207 
240  inline std::array<uint32_t, 8> compute_hashes(const Element& e) const
241  {
242  return {{FastRange32(hash_function.template operator()<0>(e), size),
243  FastRange32(hash_function.template operator()<1>(e), size),
244  FastRange32(hash_function.template operator()<2>(e), size),
245  FastRange32(hash_function.template operator()<3>(e), size),
246  FastRange32(hash_function.template operator()<4>(e), size),
247  FastRange32(hash_function.template operator()<5>(e), size),
248  FastRange32(hash_function.template operator()<6>(e), size),
249  FastRange32(hash_function.template operator()<7>(e), size)}};
250  }
251 
254  constexpr uint32_t invalid() const
255  {
256  return ~(uint32_t)0;
257  }
258 
263  inline void allow_erase(uint32_t n) const
264  {
266  }
267 
272  inline void please_keep(uint32_t n) const
273  {
275  }
276 
286  void epoch_check()
287  {
288  if (epoch_heuristic_counter != 0) {
290  return;
291  }
292  // count the number of elements from the latest epoch which
293  // have not been erased.
294  uint32_t epoch_unused_count = 0;
295  for (uint32_t i = 0; i < size; ++i)
296  epoch_unused_count += epoch_flags[i] &&
298  // If there are more non-deleted entries in the current epoch than the
299  // epoch size, then allow_erase on all elements in the old epoch (marked
300  // false) and move all elements in the current epoch to the old epoch
301  // but do not call allow_erase on their indices.
302  if (epoch_unused_count >= epoch_size) {
303  for (uint32_t i = 0; i < size; ++i)
304  if (epoch_flags[i])
305  epoch_flags[i] = false;
306  else
307  allow_erase(i);
309  } else
310  // reset the epoch_heuristic_counter to next do a scan when worst
311  // case behavior (no intermittent erases) would exceed epoch size,
312  // with a reasonable minimum scan size.
313  // Ordinarily, we would have to sanity check std::min(epoch_size,
314  // epoch_unused_count), but we already know that `epoch_unused_count
315  // < epoch_size` in this branch
316  epoch_heuristic_counter = std::max(1u, std::max(epoch_size / 16,
317  epoch_size - epoch_unused_count));
318  }
319 
320 public:
325  {
326  }
327 
336  uint32_t setup(uint32_t new_size)
337  {
338  // depth_limit must be at least one otherwise errors can occur.
339  size = std::max<uint32_t>(2, new_size);
340  depth_limit = static_cast<uint8_t>(std::log2(static_cast<float>(size)));
341  table.resize(size);
343  epoch_flags.resize(size);
344  // Set to 45% as described above
345  epoch_size = std::max(uint32_t{1}, (45 * size) / 100);
346  // Initially set to wait for a whole epoch
348  return size;
349  }
350 
364  std::pair<uint32_t, size_t> setup_bytes(size_t bytes)
365  {
366  uint32_t requested_num_elems(std::min<size_t>(
367  bytes / sizeof(Element),
368  std::numeric_limits<uint32_t>::max()));
369 
370  auto num_elems = setup(requested_num_elems);
371 
372  size_t approx_size_bytes = num_elems * sizeof(Element);
373  return std::make_pair(num_elems, approx_size_bytes);
374  }
375 
397  inline void insert(Element e)
398  {
399  epoch_check();
400  uint32_t last_loc = invalid();
401  bool last_epoch = true;
402  std::array<uint32_t, 8> locs = compute_hashes(e);
403  // Make sure we have not already inserted this element
404  // If we have, make sure that it does not get deleted
405  for (const uint32_t loc : locs)
406  if (table[loc] == e) {
407  please_keep(loc);
408  epoch_flags[loc] = last_epoch;
409  return;
410  }
411  for (uint8_t depth = 0; depth < depth_limit; ++depth) {
412  // First try to insert to an empty slot, if one exists
413  for (const uint32_t loc : locs) {
414  if (!collection_flags.bit_is_set(loc))
415  continue;
416  table[loc] = std::move(e);
417  please_keep(loc);
418  epoch_flags[loc] = last_epoch;
419  return;
420  }
435  last_loc = locs[(1 + (std::find(locs.begin(), locs.end(), last_loc) - locs.begin())) & 7];
436  std::swap(table[last_loc], e);
437  // Can't std::swap a std::vector<bool>::reference and a bool&.
438  bool epoch = last_epoch;
439  last_epoch = epoch_flags[last_loc];
440  epoch_flags[last_loc] = epoch;
441 
442  // Recompute the locs -- unfortunately happens one too many times!
443  locs = compute_hashes(e);
444  }
445  }
446 
474  inline bool contains(const Element& e, const bool erase) const
475  {
476  std::array<uint32_t, 8> locs = compute_hashes(e);
477  for (const uint32_t loc : locs)
478  if (table[loc] == e) {
479  if (erase)
480  allow_erase(loc);
481  return true;
482  }
483  return false;
484  }
485 };
486 } // namespace CuckooCache
487 
488 #endif // BITCOIN_CUCKOOCACHE_H
bit_packed_atomic_flags implements a container for garbage collection flags that is only thread unsaf...
Definition: cuckoocache.h:45
void bit_set(uint32_t s)
bit_set sets an entry as discardable.
Definition: cuckoocache.h:93
void setup(uint32_t b)
setup marks all entries and ensures that bit_packed_atomic_flags can store at least b entries.
Definition: cuckoocache.h:81
bit_packed_atomic_flags()=delete
No default constructor, as there must be some size.
bool bit_is_set(uint32_t s) const
bit_is_set queries the table for discardability at s.
Definition: cuckoocache.h:114
void bit_unset(uint32_t s)
bit_unset marks an entry as something that should not be overwritten.
Definition: cuckoocache.h:104
bit_packed_atomic_flags(uint32_t size)
bit_packed_atomic_flags constructor creates memory to sufficiently keep track of garbage collection i...
Definition: cuckoocache.h:63
std::unique_ptr< std::atomic< uint8_t >[]> mem
Definition: cuckoocache.h:46
cache implements a cache with properties similar to a cuckoo-set.
Definition: cuckoocache.h:162
uint32_t size
size stores the total available slots in the hash table
Definition: cuckoocache.h:168
uint8_t depth_limit
depth_limit determines how many elements insert should try to replace.
Definition: cuckoocache.h:200
std::vector< Element > table
table stores all the elements
Definition: cuckoocache.h:165
uint32_t epoch_heuristic_counter
epoch_heuristic_counter is used to determine when an epoch might be aged & an expensive scan should b...
Definition: cuckoocache.h:185
uint32_t setup(uint32_t new_size)
setup initializes the container to store no more than new_size elements and no less than 2 elements.
Definition: cuckoocache.h:336
void epoch_check()
epoch_check handles the changing of epochs for elements stored in the cache.
Definition: cuckoocache.h:286
uint32_t epoch_size
epoch_size is set to be the number of elements supposed to be in a epoch.
Definition: cuckoocache.h:195
std::vector< bool > epoch_flags
epoch_flags tracks how recently an element was inserted into the cache.
Definition: cuckoocache.h:178
void insert(Element e)
insert loops at most depth_limit times trying to insert a hash at various locations in the table via ...
Definition: cuckoocache.h:397
bit_packed_atomic_flags collection_flags
The bit_packed_atomic_flags array is marked mutable because we want garbage collection to be allowed ...
Definition: cuckoocache.h:172
std::array< uint32_t, 8 > compute_hashes(const Element &e) const
compute_hashes is convenience for not having to write out this expression everywhere we use the hash ...
Definition: cuckoocache.h:240
constexpr uint32_t invalid() const
invalid returns a special index that can never be inserted to
Definition: cuckoocache.h:254
const Hash hash_function
hash_function is a const instance of the hash function.
Definition: cuckoocache.h:206
void allow_erase(uint32_t n) const
allow_erase marks the element at index n as discardable.
Definition: cuckoocache.h:263
bool contains(const Element &e, const bool erase) const
contains iterates through the hash locations for a given element and checks to see if it is present.
Definition: cuckoocache.h:474
void please_keep(uint32_t n) const
please_keep marks the element at index n as an entry that should be kept.
Definition: cuckoocache.h:272
cache()
You must always construct a cache with some elements via a subsequent call to setup or setup_bytes,...
Definition: cuckoocache.h:324
std::pair< uint32_t, size_t > setup_bytes(size_t bytes)
setup_bytes is a convenience function which accounts for internal memory usage when deciding how many...
Definition: cuckoocache.h:364
static uint32_t FastRange32(uint32_t x, uint32_t n)
Fast range reduction with 32-bit input and 32-bit range.
Definition: fastrange.h:19
uint256 Hash(const T &in1)
Compute the 256-bit hash of an object.
Definition: hash.h:75
High-performance cache primitives.
Definition: cuckoocache.h:31