michael@0: /* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
michael@0: /* vim: set ts=8 sts=2 et sw=2 tw=80: */
michael@0: /* This Source Code Form is subject to the terms of the Mozilla Public
michael@0:  * License, v. 2.0. If a copy of the MPL was not distributed with this
michael@0:  * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
michael@0: 
michael@0: /*
michael@0:  * A counting Bloom filter implementation.  This allows consumers to
michael@0:  * do fast probabilistic "is item X in set Y?" testing which will
michael@0:  * never answer "no" when the correct answer is "yes" (but might
michael@0:  * incorrectly answer "yes" when the correct answer is "no").
michael@0:  */
michael@0: 
michael@0: #ifndef mozilla_BloomFilter_h
michael@0: #define mozilla_BloomFilter_h
michael@0: 
michael@0: #include "mozilla/Assertions.h"
michael@0: #include "mozilla/Likely.h"
michael@0: 
michael@0: #include <stdint.h>
michael@0: #include <string.h>
michael@0: 
michael@0: namespace mozilla {
michael@0: 
michael@0: /*
michael@0:  * This class implements a counting Bloom filter as described at
michael@0:  * <http://en.wikipedia.org/wiki/Bloom_filter#Counting_filters>, with
michael@0:  * 8-bit counters.  This allows quick probabilistic answers to the
michael@0:  * question "is object X in set Y?" where the contents of Y might not
michael@0:  * be time-invariant.  The probabilistic nature of the test means that
michael@0:  * sometimes the answer will be "yes" when it should be "no".  If the
michael@0:  * answer is "no", then X is guaranteed not to be in Y.
michael@0:  *
michael@0:  * The filter is parametrized on KeySize, which is the size of the key
michael@0:  * generated by each of hash functions used by the filter, in bits,
michael@0:  * and the type of object T being added and removed.  T must implement
michael@0:  * a |uint32_t hash() const| method which returns a uint32_t hash key
michael@0:  * that will be used to generate the two separate hash functions for
michael@0:  * the Bloom filter.  This hash key MUST be well-distributed for good
michael@0:  * results!  KeySize is not allowed to be larger than 16.
michael@0:  *
michael@0:  * The filter uses exactly 2**KeySize bytes of memory.  From now on we
michael@0:  * will refer to the memory used by the filter as M.
michael@0:  *
michael@0:  * The expected rate of incorrect "yes" answers depends on M and on
michael@0:  * the number N of objects in set Y.  As long as N is small compared
michael@0:  * to M, the rate of such answers is expected to be approximately
michael@0:  * 4*(N/M)**2 for this filter.  In practice, if Y has a few hundred
michael@0:  * elements then using a KeySize of 12 gives a reasonably low
michael@0:  * incorrect answer rate.  A KeySize of 12 has the additional benefit
michael@0:  * of using exactly one page for the filter in typical hardware
michael@0:  * configurations.
michael@0:  */
michael@0: 
michael@0: template<unsigned KeySize, class T>
michael@0: class BloomFilter
michael@0: {
michael@0:     /*
michael@0:      * A counting Bloom filter with 8-bit counters.  For now we assume
michael@0:      * that having two hash functions is enough, but we may revisit that
michael@0:      * decision later.
michael@0:      *
michael@0:      * The filter uses an array with 2**KeySize entries.
michael@0:      *
michael@0:      * Assuming a well-distributed hash function, a Bloom filter with
michael@0:      * array size M containing N elements and
michael@0:      * using k hash function has expected false positive rate exactly
michael@0:      *
michael@0:      * $  (1 - (1 - 1/M)^{kN})^k  $
michael@0:      *
michael@0:      * because each array slot has a
michael@0:      *
michael@0:      * $  (1 - 1/M)^{kN}  $
michael@0:      *
michael@0:      * chance of being 0, and the expected false positive rate is the
michael@0:      * probability that all of the k hash functions will hit a nonzero
michael@0:      * slot.
michael@0:      *
michael@0:      * For reasonable assumptions (M large, kN large, which should both
michael@0:      * hold if we're worried about false positives) about M and kN this
michael@0:      * becomes approximately
michael@0:      *
michael@0:      * $$  (1 - \exp(-kN/M))^k   $$
michael@0:      *
michael@0:      * For our special case of k == 2, that's $(1 - \exp(-2N/M))^2$,
michael@0:      * or in other words
michael@0:      *
michael@0:      * $$    N/M = -0.5 * \ln(1 - \sqrt(r))   $$
michael@0:      *
michael@0:      * where r is the false positive rate.  This can be used to compute
michael@0:      * the desired KeySize for a given load N and false positive rate r.
michael@0:      *
michael@0:      * If N/M is assumed small, then the false positive rate can
michael@0:      * further be approximated as 4*N^2/M^2.  So increasing KeySize by
michael@0:      * 1, which doubles M, reduces the false positive rate by about a
michael@0:      * factor of 4, and a false positive rate of 1% corresponds to
michael@0:      * about M/N == 20.
michael@0:      *
michael@0:      * What this means in practice is that for a few hundred keys using a
michael@0:      * KeySize of 12 gives false positive rates on the order of 0.25-4%.
michael@0:      *
michael@0:      * Similarly, using a KeySize of 10 would lead to a 4% false
michael@0:      * positive rate for N == 100 and to quite bad false positive
michael@0:      * rates for larger N.
michael@0:      */
michael@0:   public:
michael@0:     BloomFilter() {
michael@0:         static_assert(KeySize <= keyShift, "KeySize too big");
michael@0: 
michael@0:         // Should we have a custom operator new using calloc instead and
michael@0:         // require that we're allocated via the operator?
michael@0:         clear();
michael@0:     }
michael@0: 
michael@0:     /*
michael@0:      * Clear the filter.  This should be done before reusing it, because
michael@0:      * just removing all items doesn't clear counters that hit the upper
michael@0:      * bound.
michael@0:      */
michael@0:     void clear();
michael@0: 
michael@0:     /*
michael@0:      * Add an item to the filter.
michael@0:      */
michael@0:     void add(const T* t);
michael@0: 
michael@0:     /*
michael@0:      * Remove an item from the filter.
michael@0:      */
michael@0:     void remove(const T* t);
michael@0: 
michael@0:     /*
michael@0:      * Check whether the filter might contain an item.  This can
michael@0:      * sometimes return true even if the item is not in the filter,
michael@0:      * but will never return false for items that are actually in the
michael@0:      * filter.
michael@0:      */
michael@0:     bool mightContain(const T* t) const;
michael@0: 
michael@0:     /*
michael@0:      * Methods for add/remove/contain when we already have a hash computed
michael@0:      */
michael@0:     void add(uint32_t hash);
michael@0:     void remove(uint32_t hash);
michael@0:     bool mightContain(uint32_t hash) const;
michael@0: 
michael@0:   private:
michael@0:     static const size_t arraySize = (1 << KeySize);
michael@0:     static const uint32_t keyMask = (1 << KeySize) - 1;
michael@0:     static const uint32_t keyShift = 16;
michael@0: 
michael@0:     static uint32_t hash1(uint32_t hash) { return hash & keyMask; }
michael@0:     static uint32_t hash2(uint32_t hash) { return (hash >> keyShift) & keyMask; }
michael@0: 
michael@0:     uint8_t& firstSlot(uint32_t hash) { return counters[hash1(hash)]; }
michael@0:     uint8_t& secondSlot(uint32_t hash) { return counters[hash2(hash)]; }
michael@0:     const uint8_t& firstSlot(uint32_t hash) const { return counters[hash1(hash)]; }
michael@0:     const uint8_t& secondSlot(uint32_t hash) const { return counters[hash2(hash)]; }
michael@0: 
michael@0:     static bool full(const uint8_t& slot) { return slot == UINT8_MAX; }
michael@0: 
michael@0:     uint8_t counters[arraySize];
michael@0: };
michael@0: 
michael@0: template<unsigned KeySize, class T>
michael@0: inline void
michael@0: BloomFilter<KeySize, T>::clear()
michael@0: {
michael@0:   memset(counters, 0, arraySize);
michael@0: }
michael@0: 
michael@0: template<unsigned KeySize, class T>
michael@0: inline void
michael@0: BloomFilter<KeySize, T>::add(uint32_t hash)
michael@0: {
michael@0:   uint8_t& slot1 = firstSlot(hash);
michael@0:   if (MOZ_LIKELY(!full(slot1)))
michael@0:     ++slot1;
michael@0: 
michael@0:   uint8_t& slot2 = secondSlot(hash);
michael@0:   if (MOZ_LIKELY(!full(slot2)))
michael@0:     ++slot2;
michael@0: }
michael@0: 
michael@0: template<unsigned KeySize, class T>
michael@0: MOZ_ALWAYS_INLINE void
michael@0: BloomFilter<KeySize, T>::add(const T* t)
michael@0: {
michael@0:   uint32_t hash = t->hash();
michael@0:   return add(hash);
michael@0: }
michael@0: 
michael@0: template<unsigned KeySize, class T>
michael@0: inline void
michael@0: BloomFilter<KeySize, T>::remove(uint32_t hash)
michael@0: {
michael@0:   // If the slots are full, we don't know whether we bumped them to be
michael@0:   // there when we added or not, so just leave them full.
michael@0:   uint8_t& slot1 = firstSlot(hash);
michael@0:   if (MOZ_LIKELY(!full(slot1)))
michael@0:     --slot1;
michael@0: 
michael@0:   uint8_t& slot2 = secondSlot(hash);
michael@0:   if (MOZ_LIKELY(!full(slot2)))
michael@0:     --slot2;
michael@0: }
michael@0: 
michael@0: template<unsigned KeySize, class T>
michael@0: MOZ_ALWAYS_INLINE void
michael@0: BloomFilter<KeySize, T>::remove(const T* t)
michael@0: {
michael@0:   uint32_t hash = t->hash();
michael@0:   remove(hash);
michael@0: }
michael@0: 
michael@0: template<unsigned KeySize, class T>
michael@0: MOZ_ALWAYS_INLINE bool
michael@0: BloomFilter<KeySize, T>::mightContain(uint32_t hash) const
michael@0: {
michael@0:   // Check that all the slots for this hash contain something
michael@0:   return firstSlot(hash) && secondSlot(hash);
michael@0: }
michael@0: 
michael@0: template<unsigned KeySize, class T>
michael@0: MOZ_ALWAYS_INLINE bool
michael@0: BloomFilter<KeySize, T>::mightContain(const T* t) const
michael@0: {
michael@0:   uint32_t hash = t->hash();
michael@0:   return mightContain(hash);
michael@0: }
michael@0: 
michael@0: } // namespace mozilla
michael@0: 
michael@0: #endif /* mozilla_BloomFilter_h */