rocksdb/cache/clock_cache.h

1129 lines
50 KiB
C++

// Copyright (c) 2011-present, Facebook, Inc. All rights reserved.
// This source code is licensed under both the GPLv2 (found in the
// COPYING file in the root directory) and Apache 2.0 License
// (found in the LICENSE.Apache file in the root directory).
//
// Copyright (c) 2011 The LevelDB Authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file. See the AUTHORS file for names of contributors.
#pragma once
#include <array>
#include <atomic>
#include <cstddef>
#include <cstdint>
#include <memory>
#include <string>
#include "cache/cache_key.h"
#include "cache/sharded_cache.h"
#include "port/lang.h"
#include "port/malloc.h"
#include "port/mmap.h"
#include "port/port.h"
#include "rocksdb/cache.h"
#include "rocksdb/secondary_cache.h"
#include "util/autovector.h"
#include "util/math.h"
namespace ROCKSDB_NAMESPACE {
namespace clock_cache {
// Forward declaration of friend class.
template <class ClockCache>
class ClockCacheTest;
// HyperClockCache is an alternative to LRUCache specifically tailored for
// use as BlockBasedTableOptions::block_cache
//
// Benefits
// --------
// * Lock/wait free (no waits or spins) for efficiency under high concurrency
// * Fixed version (estimated_entry_charge > 0) is fully lock/wait free
// * Automatic version (estimated_entry_charge = 0) has rare waits among
// certain insertion or erase operations that involve the same very small
// set of entries.
// * Optimized for hot path reads. For concurrency control, most Lookup() and
// essentially all Release() are a single atomic add operation.
// * Eviction on insertion is fully parallel.
// * Uses a generalized + aging variant of CLOCK eviction that might outperform
// LRU in some cases. (For background, see
// https://en.wikipedia.org/wiki/Page_replacement_algorithm)
//
// Costs
// -----
// * FixedHyperClockCache (estimated_entry_charge > 0) - Hash table is not
// resizable (for lock-free efficiency) so capacity is not dynamically
// changeable. Rely on an estimated average value (block) size for
// space+time efficiency. (See estimated_entry_charge option details.)
// EXPERIMENTAL - This limitation is fixed in AutoHyperClockCache, activated
// with estimated_entry_charge == 0.
// * Insert usually does not (but might) overwrite a previous entry associated
// with a cache key. This is OK for RocksDB uses of Cache, though it does mess
// up our REDUNDANT block cache insertion statistics.
// * Only supports keys of exactly 16 bytes, which is what RocksDB uses for
// block cache (but not row cache or table cache).
// * Cache priorities are less aggressively enforced. Unlike LRUCache, enough
// transient LOW or BOTTOM priority items can evict HIGH priority entries that
// are not referenced recently (or often) enough.
// * If pinned entries leave little or nothing eligible for eviction,
// performance can degrade substantially, because of clock eviction eating
// CPU looking for evictable entries and because Release does not
// pro-actively delete unreferenced entries when the cache is over-full.
// Specifically, this makes this implementation more susceptible to the
// following combination:
// * num_shard_bits is high (e.g. 6)
// * capacity small (e.g. some MBs)
// * some large individual entries (e.g. non-partitioned filters)
// where individual entries occupy a large portion of their shard capacity.
// This should be mostly mitigated by the implementation picking a lower
// number of cache shards than LRUCache for a given capacity (when
// num_shard_bits is not overridden; see calls to GetDefaultCacheShardBits()).
// * With strict_capacity_limit=false, respecting the capacity limit is not as
// aggressive as LRUCache. The limit might be transiently exceeded by a very
// small number of entries even when not strictly necessary, and slower to
// recover after pinning forces limit to be substantially exceeded. (Even with
// strict_capacity_limit=true, RocksDB will nevertheless transiently allocate
// memory before discovering it is over the block cache capacity, so this
// should not be a detectable regression in respecting memory limits, except
// on exceptionally small caches.)
// * In some cases, erased or duplicated entries might not be freed
// immediately. They will eventually be freed by eviction from further Inserts.
// * Internal metadata can overflow if the number of simultaneous references
// to a cache handle reaches many millions.
//
// High-level eviction algorithm
// -----------------------------
// A score (or "countdown") is maintained for each entry, initially determined
// by priority. The score is incremented on each Lookup, up to a max of 3,
// though is easily returned to previous state if useful=false with Release.
// During CLOCK-style eviction iteration, entries with score > 0 are
// decremented if currently unreferenced and entries with score == 0 are
// evicted if currently unreferenced. Note that scoring might not be perfect
// because entries can be referenced transiently within the cache even when
// there are no outside references to the entry.
//
// Cache sharding like LRUCache is used to reduce contention on usage+eviction
// state, though here the performance improvement from more shards is small,
// and (as noted above) potentially detrimental if shard capacity is too close
// to largest entry size. Here cache sharding mostly only affects cache update
// (Insert / Erase) performance, not read performance.
//
// Read efficiency (hot path)
// --------------------------
// Mostly to minimize the cost of accessing metadata blocks with
// cache_index_and_filter_blocks=true, we focus on optimizing Lookup and
// Release. In terms of concurrency, at a minimum, these operations have
// to do reference counting (and Lookup has to compare full keys in a safe
// way). Can we fold in all the other metadata tracking *for free* with
// Lookup and Release doing a simple atomic fetch_add/fetch_sub? (Assume
// for the moment that Lookup succeeds on the first probe.)
//
// We have a clever way of encoding an entry's reference count and countdown
// clock so that Lookup and Release are each usually a single atomic addition.
// In a single metadata word we have both an "acquire" count, incremented by
// Lookup, and a "release" count, incremented by Release. If useful=false,
// Release can instead decrement the acquire count. Thus the current ref
// count is (acquires - releases), and the countdown clock is min(3, acquires).
// Note that only unreferenced entries (acquires == releases) are eligible
// for CLOCK manipulation and eviction. We tolerate use of more expensive
// compare_exchange operations for cache writes (insertions and erasures).
//
// In a cache receiving many reads and little or no writes, it is possible
// for the acquire and release counters to overflow. Assuming the *current*
// refcount never reaches to many millions, we only have to correct for
// overflow in both counters in Release, not in Lookup. The overflow check
// should be only 1-2 CPU cycles per Release because it is a predictable
// branch on a simple condition on data already in registers.
//
// Slot states
// -----------
// We encode a state indicator into the same metadata word with the
// acquire and release counters. This allows bigger state transitions to
// be atomic. States:
//
// * Empty - slot is not in use and unowned. All other metadata and data is
// in an undefined state.
// * Construction - slot is exclusively owned by one thread, the thread
// successfully entering this state, for populating or freeing data
// (de-construction, same state marker).
// * Shareable (group) - slot holds an entry with counted references for
// pinning and reading, including
// * Visible - slot holds an entry that can be returned by Lookup
// * Invisible - slot holds an entry that is not visible to Lookup
// (erased by user) but can be read by existing references, and ref count
// changed by Ref and Release.
//
// A special case is "standalone" entries, which are heap-allocated handles
// not in the table. They are always Invisible and freed on zero refs.
//
// State transitions:
// Empty -> Construction (in Insert): The encoding of state enables Insert to
// perform an optimistic atomic bitwise-or to take ownership if a slot is
// empty, or otherwise make no state change.
//
// Construction -> Visible (in Insert): This can be a simple assignment to the
// metadata word because the current thread has exclusive ownership and other
// metadata is meaningless.
//
// Visible -> Invisible (in Erase): This can be a bitwise-and while holding
// a shared reference, which is safe because the change is idempotent (in case
// of parallel Erase). By the way, we never go Invisible->Visible.
//
// Shareable -> Construction (in Evict part of Insert, in Erase, and in
// Release if Invisible): This is for starting to freeing/deleting an
// unreferenced entry. We have to use compare_exchange to ensure we only make
// this transition when there are zero refs.
//
// Construction -> Empty (in same places): This is for completing free/delete
// of an entry. A "release" atomic store suffices, as we have exclusive
// ownership of the slot but have to ensure none of the data member reads are
// re-ordered after committing the state transition.
//
// Insert
// ------
// If Insert were to guarantee replacing an existing entry for a key, there
// would be complications for concurrency and efficiency. First, consider how
// many probes to get to an entry. To ensure Lookup never waits and
// availability of a key is uninterrupted, we would need to use a different
// slot for a new entry for the same key. This means it is most likely in a
// later probing position than the old version, which should soon be removed.
// (Also, an entry is too big to replace atomically, even if no current refs.)
//
// However, overwrite capability is not really needed by RocksDB. Also, we
// know from our "redundant" stats that overwrites are very rare for the block
// cache, so we should not spend much to make them effective.
//
// FixedHyperClockCache: Instead we Insert as soon as we find an empty slot in
// the probing sequence without seeing an existing (visible) entry for the same
// key. This way we only insert if we can improve the probing performance, and
// we don't need to probe beyond our insert position, assuming we are willing
// to let the previous entry for the same key die of old age (eventual eviction
// from not being used). We can reach a similar state with concurrent
// insertions, where one will pass over the other while it is "under
// construction." This temporary duplication is acceptable for RocksDB block
// cache because we know redundant insertion is rare.
// AutoHyperClockCache: Similar, except we only notice and return an existing
// match if it is found in the search for a suitable empty slot (starting with
// the same slot as the head pointer), not by following the existing chain of
// entries. Insertions are always made to the head of the chain.
//
// Another problem to solve is what to return to the caller when we find an
// existing entry whose probing position we cannot improve on, or when the
// table occupancy limit has been reached. If strict_capacity_limit=false,
// we must never fail Insert, and if a Handle* is provided, we have to return
// a usable Cache handle on success. The solution to this (typically rare)
// problem is "standalone" handles, which are usable by the caller but not
// actually available for Lookup in the Cache. Standalone handles are allocated
// independently on the heap and specially marked so that they are freed on
// the heap when their last reference is released.
//
// Usage on capacity
// -----------------
// Insert takes different approaches to usage tracking depending on
// strict_capacity_limit setting. If true, we enforce a kind of strong
// consistency where compare-exchange is used to ensure the usage number never
// exceeds its limit, and provide threads with an authoritative signal on how
// much "usage" they have taken ownership of. With strict_capacity_limit=false,
// we use a kind of "eventual consistency" where all threads Inserting to the
// same cache shard might race on reserving the same space, but the
// over-commitment will be worked out in later insertions. It is kind of a
// dance because we don't want threads racing each other too much on paying
// down the over-commitment (with eviction) either.
//
// Eviction
// --------
// A key part of Insert is evicting some entries currently unreferenced to
// make room for new entries. The high-level eviction algorithm is described
// above, but the details are also interesting. A key part is parallelizing
// eviction with a single CLOCK pointer. This works by each thread working on
// eviction pre-emptively incrementing the CLOCK pointer, and then CLOCK-
// updating or evicting the incremented-over slot(s). To reduce contention at
// the cost of possibly evicting too much, each thread increments the clock
// pointer by 4, so commits to updating at least 4 slots per batch. As
// described above, a CLOCK update will decrement the "countdown" of
// unreferenced entries, or evict unreferenced entries with zero countdown.
// Referenced entries are not updated, because we (presumably) don't want
// long-referenced entries to age while referenced. Note however that we
// cannot distinguish transiently referenced entries from cache user
// references, so some CLOCK updates might be somewhat arbitrarily skipped.
// This is OK as long as it is rare enough that eviction order is still
// pretty good.
//
// There is no synchronization on the completion of the CLOCK updates, so it
// is theoretically possible for another thread to cycle back around and have
// two threads racing on CLOCK updates to the same slot. Thus, we cannot rely
// on any implied exclusivity to make the updates or eviction more efficient.
// These updates use an opportunistic compare-exchange (no loop), where a
// racing thread might cause the update to be skipped without retry, but in
// such case the update is likely not needed because the most likely update
// to an entry is that it has become referenced. (TODO: test efficiency of
// avoiding compare-exchange loop)
//
// Release
// -------
// In the common case, Release is a simple atomic increment of the release
// counter. There is a simple overflow check that only does another atomic
// update in extremely rare cases, so costs almost nothing.
//
// If the Release specifies "not useful", we can instead decrement the
// acquire counter, which returns to the same CLOCK state as before Lookup
// or Ref.
//
// Adding a check for over-full cache on every release to zero-refs would
// likely be somewhat expensive, increasing read contention on cache shard
// metadata. Instead we are less aggressive about deleting entries right
// away in those cases.
//
// However Release tries to immediately delete entries reaching zero refs
// if (a) erase_if_last_ref is set by the caller, or (b) the entry is already
// marked invisible. Both of these are checks on values already in CPU
// registers so do not increase cross-CPU contention when not applicable.
// When applicable, they use a compare-exchange loop to take exclusive
// ownership of the slot for freeing the entry. These are rare cases
// that should not usually affect performance.
//
// Erase
// -----
// Searches for an entry like Lookup but moves it to Invisible state if found.
// This state transition is with bit operations so is idempotent and safely
// done while only holding a shared "read" reference. Like Release, it makes
// a best effort to immediately release an Invisible entry that reaches zero
// refs, but there are some corner cases where it will only be freed by the
// clock eviction process.
// ----------------------------------------------------------------------- //
struct ClockHandleBasicData {
Cache::ObjectPtr value = nullptr;
const Cache::CacheItemHelper* helper = nullptr;
// A lossless, reversible hash of the fixed-size (16 byte) cache key. This
// eliminates the need to store a hash separately.
UniqueId64x2 hashed_key = kNullUniqueId64x2;
size_t total_charge = 0;
inline size_t GetTotalCharge() const { return total_charge; }
// Calls deleter (if non-null) on cache key and value
void FreeData(MemoryAllocator* allocator) const;
// Required by concept HandleImpl
const UniqueId64x2& GetHash() const { return hashed_key; }
};
struct ClockHandle : public ClockHandleBasicData {
// Constants for handling the atomic `meta` word, which tracks most of the
// state of the handle. The meta word looks like this:
// low bits high bits
// -----------------------------------------------------------------------
// | acquire counter | release counter | hit bit | state marker |
// -----------------------------------------------------------------------
// For reading or updating counters in meta word.
static constexpr uint8_t kCounterNumBits = 30;
static constexpr uint64_t kCounterMask = (uint64_t{1} << kCounterNumBits) - 1;
static constexpr uint8_t kAcquireCounterShift = 0;
static constexpr uint64_t kAcquireIncrement = uint64_t{1}
<< kAcquireCounterShift;
static constexpr uint8_t kReleaseCounterShift = kCounterNumBits;
static constexpr uint64_t kReleaseIncrement = uint64_t{1}
<< kReleaseCounterShift;
// For setting the hit bit
static constexpr uint8_t kHitBitShift = 2U * kCounterNumBits;
static constexpr uint64_t kHitBitMask = uint64_t{1} << kHitBitShift;
// For reading or updating the state marker in meta word
static constexpr uint8_t kStateShift = kHitBitShift + 1;
// Bits contribution to state marker.
// Occupied means any state other than empty
static constexpr uint8_t kStateOccupiedBit = 0b100;
// Shareable means the entry is reference counted (visible or invisible)
// (only set if also occupied)
static constexpr uint8_t kStateShareableBit = 0b010;
// Visible is only set if also shareable
static constexpr uint8_t kStateVisibleBit = 0b001;
// Complete state markers (not shifted into full word)
static constexpr uint8_t kStateEmpty = 0b000;
static constexpr uint8_t kStateConstruction = kStateOccupiedBit;
static constexpr uint8_t kStateInvisible =
kStateOccupiedBit | kStateShareableBit;
static constexpr uint8_t kStateVisible =
kStateOccupiedBit | kStateShareableBit | kStateVisibleBit;
// Constants for initializing the countdown clock. (Countdown clock is only
// in effect with zero refs, acquire counter == release counter, and in that
// case the countdown clock == both of those counters.)
static constexpr uint8_t kHighCountdown = 3;
static constexpr uint8_t kLowCountdown = 2;
static constexpr uint8_t kBottomCountdown = 1;
// During clock update, treat any countdown clock value greater than this
// value the same as this value.
static constexpr uint8_t kMaxCountdown = kHighCountdown;
// TODO: make these coundown values tuning parameters for eviction?
// See above. Mutable for read reference counting.
mutable std::atomic<uint64_t> meta{};
}; // struct ClockHandle
class BaseClockTable {
public:
BaseClockTable(CacheMetadataChargePolicy metadata_charge_policy,
MemoryAllocator* allocator,
const Cache::EvictionCallback* eviction_callback,
const uint32_t* hash_seed)
: metadata_charge_policy_(metadata_charge_policy),
allocator_(allocator),
eviction_callback_(*eviction_callback),
hash_seed_(*hash_seed) {}
template <class Table>
typename Table::HandleImpl* CreateStandalone(ClockHandleBasicData& proto,
size_t capacity,
bool strict_capacity_limit,
bool allow_uncharged);
template <class Table>
Status Insert(const ClockHandleBasicData& proto,
typename Table::HandleImpl** handle, Cache::Priority priority,
size_t capacity, bool strict_capacity_limit);
void Ref(ClockHandle& handle);
size_t GetOccupancy() const {
return occupancy_.load(std::memory_order_relaxed);
}
size_t GetUsage() const { return usage_.load(std::memory_order_relaxed); }
size_t GetStandaloneUsage() const {
return standalone_usage_.load(std::memory_order_relaxed);
}
uint32_t GetHashSeed() const { return hash_seed_; }
uint64_t GetYieldCount() const { return yield_count_.load(); }
struct EvictionData {
size_t freed_charge = 0;
size_t freed_count = 0;
};
void TrackAndReleaseEvictedEntry(ClockHandle* h, EvictionData* data);
#ifndef NDEBUG
// Acquire N references
void TEST_RefN(ClockHandle& handle, size_t n);
// Helper for TEST_ReleaseN
void TEST_ReleaseNMinus1(ClockHandle* handle, size_t n);
#endif
private: // fns
// Creates a "standalone" handle for returning from an Insert operation that
// cannot be completed by actually inserting into the table.
// Updates `standalone_usage_` but not `usage_` nor `occupancy_`.
template <class HandleImpl>
HandleImpl* StandaloneInsert(const ClockHandleBasicData& proto);
// Helper for updating `usage_` for new entry with given `total_charge`
// and evicting if needed under strict_capacity_limit=true rules. This
// means the operation might fail with Status::MemoryLimit. If
// `need_evict_for_occupancy`, then eviction of at least one entry is
// required, and the operation should fail if not possible.
// NOTE: Otherwise, occupancy_ is not managed in this function
template <class Table>
Status ChargeUsageMaybeEvictStrict(size_t total_charge, size_t capacity,
bool need_evict_for_occupancy,
typename Table::InsertState& state);
// Helper for updating `usage_` for new entry with given `total_charge`
// and evicting if needed under strict_capacity_limit=false rules. This
// means that updating `usage_` always succeeds even if forced to exceed
// capacity. If `need_evict_for_occupancy`, then eviction of at least one
// entry is required, and the operation should return false if such eviction
// is not possible. `usage_` is not updated in that case. Otherwise, returns
// true, indicating success.
// NOTE: occupancy_ is not managed in this function
template <class Table>
bool ChargeUsageMaybeEvictNonStrict(size_t total_charge, size_t capacity,
bool need_evict_for_occupancy,
typename Table::InsertState& state);
protected: // data
// We partition the following members into different cache lines
// to avoid false sharing among Lookup, Release, Erase and Insert
// operations in ClockCacheShard.
// Clock algorithm sweep pointer.
std::atomic<uint64_t> clock_pointer_{};
// Counter for number of times we yield to wait on another thread.
std::atomic<uint64_t> yield_count_{};
// TODO: is this separation needed if we don't do background evictions?
ALIGN_AS(CACHE_LINE_SIZE)
// Number of elements in the table.
std::atomic<size_t> occupancy_{};
// Memory usage by entries tracked by the cache (including standalone)
std::atomic<size_t> usage_{};
// Part of usage by standalone entries (not in table)
std::atomic<size_t> standalone_usage_{};
ALIGN_AS(CACHE_LINE_SIZE)
const CacheMetadataChargePolicy metadata_charge_policy_;
// From Cache, for deleter
MemoryAllocator* const allocator_;
// A reference to Cache::eviction_callback_
const Cache::EvictionCallback& eviction_callback_;
// A reference to ShardedCacheBase::hash_seed_
const uint32_t& hash_seed_;
};
// Hash table for cache entries with size determined at creation time.
// Uses open addressing and double hashing. Since entries cannot be moved,
// the "displacements" count ensures probing sequences find entries even when
// entries earlier in the probing sequence have been removed.
class FixedHyperClockTable : public BaseClockTable {
public:
// Target size to be exactly a common cache line size (see static_assert in
// clock_cache.cc)
struct ALIGN_AS(64U) HandleImpl : public ClockHandle {
// The number of elements that hash to this slot or a lower one, but wind
// up in this slot or a higher one.
std::atomic<uint32_t> displacements{};
// Whether this is a "deteched" handle that is independently allocated
// with `new` (so must be deleted with `delete`).
// TODO: ideally this would be packed into some other data field, such
// as upper bits of total_charge, but that incurs a measurable performance
// regression.
bool standalone = false;
inline bool IsStandalone() const { return standalone; }
inline void SetStandalone() { standalone = true; }
}; // struct HandleImpl
struct Opts {
explicit Opts(size_t _estimated_value_size)
: estimated_value_size(_estimated_value_size) {}
explicit Opts(const HyperClockCacheOptions& opts) {
assert(opts.estimated_entry_charge > 0);
estimated_value_size = opts.estimated_entry_charge;
}
size_t estimated_value_size;
};
FixedHyperClockTable(size_t capacity, bool strict_capacity_limit,
CacheMetadataChargePolicy metadata_charge_policy,
MemoryAllocator* allocator,
const Cache::EvictionCallback* eviction_callback,
const uint32_t* hash_seed, const Opts& opts);
~FixedHyperClockTable();
// For BaseClockTable::Insert
struct InsertState {};
void StartInsert(InsertState& state);
// Returns true iff there is room for the proposed number of entries.
bool GrowIfNeeded(size_t new_occupancy, InsertState& state);
HandleImpl* DoInsert(const ClockHandleBasicData& proto,
uint64_t initial_countdown, bool take_ref,
InsertState& state);
// Runs the clock eviction algorithm trying to reclaim at least
// requested_charge. Returns how much is evicted, which could be less
// if it appears impossible to evict the requested amount without blocking.
void Evict(size_t requested_charge, InsertState& state, EvictionData* data);
HandleImpl* Lookup(const UniqueId64x2& hashed_key);
bool Release(HandleImpl* handle, bool useful, bool erase_if_last_ref);
void Erase(const UniqueId64x2& hashed_key);
void EraseUnRefEntries();
size_t GetTableSize() const { return size_t{1} << length_bits_; }
size_t GetOccupancyLimit() const { return occupancy_limit_; }
const HandleImpl* HandlePtr(size_t idx) const { return &array_[idx]; }
#ifndef NDEBUG
size_t& TEST_MutableOccupancyLimit() {
return const_cast<size_t&>(occupancy_limit_);
}
// Release N references
void TEST_ReleaseN(HandleImpl* handle, size_t n);
#endif
// The load factor p is a real number in (0, 1) such that at all
// times at most a fraction p of all slots, without counting tombstones,
// are occupied by elements. This means that the probability that a random
// probe hits an occupied slot is at most p, and thus at most 1/p probes
// are required on average. For example, p = 70% implies that between 1 and 2
// probes are needed on average (bear in mind that this reasoning doesn't
// consider the effects of clustering over time, which should be negligible
// with double hashing).
// Because the size of the hash table is always rounded up to the next
// power of 2, p is really an upper bound on the actual load factor---the
// actual load factor is anywhere between p/2 and p. This is a bit wasteful,
// but bear in mind that slots only hold metadata, not actual values.
// Since space cost is dominated by the values (the LSM blocks),
// overprovisioning the table with metadata only increases the total cache
// space usage by a tiny fraction.
static constexpr double kLoadFactor = 0.7;
// The user can exceed kLoadFactor if the sizes of the inserted values don't
// match estimated_value_size, or in some rare cases with
// strict_capacity_limit == false. To avoid degenerate performance, we set a
// strict upper bound on the load factor.
static constexpr double kStrictLoadFactor = 0.84;
private: // functions
// Returns x mod 2^{length_bits_}.
inline size_t ModTableSize(uint64_t x) {
return BitwiseAnd(x, length_bits_mask_);
}
// Returns the first slot in the probe sequence with a handle e such that
// match_fn(e) is true. At every step, the function first tests whether
// match_fn(e) holds. If this is false, it evaluates abort_fn(e) to decide
// whether the search should be aborted, and if so, FindSlot immediately
// returns nullptr. For every handle e that is not a match and not aborted,
// FindSlot runs update_fn(e, is_last) where is_last is set to true iff that
// slot will be the last probed because the next would cycle back to the first
// slot probed. This function uses templates instead of std::function to
// minimize the risk of heap-allocated closures being created.
template <typename MatchFn, typename AbortFn, typename UpdateFn>
inline HandleImpl* FindSlot(const UniqueId64x2& hashed_key,
const MatchFn& match_fn, const AbortFn& abort_fn,
const UpdateFn& update_fn);
// Re-decrement all displacements in probe path starting from beginning
// until (not including) the given handle
inline void Rollback(const UniqueId64x2& hashed_key, const HandleImpl* h);
// Subtracts `total_charge` from `usage_` and 1 from `occupancy_`.
// Ideally this comes after releasing the entry itself so that we
// actually have the available occupancy/usage that is claimed.
// However, that means total_charge has to be saved from the handle
// before releasing it so that it can be provided to this function.
inline void ReclaimEntryUsage(size_t total_charge);
MemoryAllocator* GetAllocator() const { return allocator_; }
// Returns the number of bits used to hash an element in the hash
// table.
static int CalcHashBits(size_t capacity, size_t estimated_value_size,
CacheMetadataChargePolicy metadata_charge_policy);
private: // data
// Number of hash bits used for table index.
// The size of the table is 1 << length_bits_.
const int length_bits_;
// For faster computation of ModTableSize.
const size_t length_bits_mask_;
// Maximum number of elements the user can store in the table.
const size_t occupancy_limit_;
// Array of slots comprising the hash table.
const std::unique_ptr<HandleImpl[]> array_;
}; // class FixedHyperClockTable
// Hash table for cache entries that resizes automatically based on occupancy.
// However, it depends on a contiguous memory region to grow into
// incrementally, using linear hashing, so uses an anonymous mmap so that
// only the used portion of the memory region is mapped to physical memory
// (part of RSS).
//
// This table implementation uses the same "low-level protocol" for managing
// the contens of an entry slot as FixedHyperClockTable does, captured in the
// ClockHandle struct. The provides most of the essential data safety, but
// AutoHyperClockTable is another "high-level protocol" for organizing entries
// into a hash table, with automatic resizing.
//
// This implementation is not fully wait-free but we can call it "essentially
// wait-free," and here's why. First, like FixedHyperClockCache, there is no
// locking nor other forms of waiting at the cache or shard level. Also like
// FixedHCC there is essentially an entry-level read-write lock implemented
// with atomics, but our relaxed atomicity/consistency guarantees (e.g.
// duplicate inserts are possible) mean we do not need to wait for entry
// locking. Lookups, non-erasing Releases, and non-evicting non-growing Inserts
// are all fully wait-free. Of course, these waits are not dependent on any
// external factors such as I/O.
//
// For operations that remove entries from a chain or grow the table by
// splitting a chain, there is a chain-level locking mechanism that we call a
// "rewrite" lock, and the only waits are for these locks. On average, each
// chain lock is relevant to < 2 entries each. (The average would be less than
// one entry each, but we do not lock when there's no entry to remove or
// migrate.) And a given thread can only hold two such chain locks at a time,
// more typically just one. So in that sense alone, the waiting that does exist
// is very localized.
//
// If we look closer at the operations utilizing that locking mechanism, we
// can see why it's "essentially wait-free."
// * Grow operations to increase the size of the table: each operation splits
// an existing chain into two, and chains for splitting are chosen in table
// order. Grow operations are fully parallel except for the chain locking, but
// for one Grow operation to wait on another, it has to be feeding into the
// other, which means the table has doubled in size already from other Grow
// operations without the original one finishing. So Grow operations are very
// low latency (unlike LRUCache doubling the table size in one operation) and
// very parallelizeable. (We use some tricks to break up dependencies in
// updating metadata on the usable size of the table.) And obviously Grow
// operations are very rare after the initial population of the table.
// * Evict operations (part of many Inserts): clock updates and evictions
// sweep through the structure in table order, so like Grow operations,
// parallel Evict can only wait on each other if an Evict has lingered (slept)
// long enough that the clock pointer has wrapped around the entire structure.
// * Random erasures (Erase, Release with erase_if_last_ref, etc.): these
// operations are rare and not really considered performance critical.
// Currently they're mostly used for removing placeholder cache entries, e.g.
// for memory tracking, though that could use standalone entries instead to
// avoid potential contention in table operations. It's possible that future
// enhancements could pro-actively remove cache entries from obsolete files,
// but that's not yet implemented.
class AutoHyperClockTable : public BaseClockTable {
public:
// Target size to be exactly a common cache line size (see static_assert in
// clock_cache.cc)
struct ALIGN_AS(64U) HandleImpl : public ClockHandle {
// To orgainize AutoHyperClockTable entries into a hash table while
// allowing the table size to grow without existing entries being moved,
// a version of chaining is used. Rather than being heap allocated (and
// incurring overheads to ensure memory safety) entries must go into
// Handles ("slots") in the pre-allocated array. To improve CPU cache
// locality, the chain head pointers are interleved with the entries;
// specifically, a Handle contains
// * A head pointer for a chain of entries with this "home" location.
// * A ClockHandle, for an entry that may or may not be in the chain
// starting from that head (but for performance ideally is on that
// chain).
// * A next pointer for the continuation of the chain containing this
// entry.
//
// The pointers are not raw pointers, but are indices into the array,
// and are decorated in two ways to help detect and recover from
// relevant concurrent modifications during Lookup, so that Lookup is
// fully wait-free:
// * Each "with_shift" pointer contains a shift count that indicates
// how many hash bits were used in chosing the home address for the
// chain--specifically the next entry in the chain.
// * The end of a chain is given a special "end" marker and refers back
// to the head of the chain.
//
// Why do we need shift on each pointer? To make Lookup wait-free, we need
// to be able to query a chain without missing anything, and preferably
// avoid synchronously double-checking the length_info. Without the shifts,
// there is a risk that we start down a chain and while paused on an entry
// that goes to a new home, we then follow the rest of the
// partially-migrated chain to see the shared ending with the old home, but
// for a time were following the chain for the new home, missing some
// entries for the old home.
//
// Why do we need the end of the chain to loop back? If Lookup pauses
// at an "under construction" entry, and sees that "next" is null after
// waking up, we need something to tell whether the "under construction"
// entry was freed and reused for another chain. Otherwise, we could
// miss entries still on the original chain due in the presence of a
// concurrent modification. Until an entry is fully erased from a chain,
// it is normal to see "under construction" entries on the chain, and it
// is not safe to read their hashed key without either a read reference
// on the entry or a rewrite lock on the chain.
// Marker in a "with_shift" head pointer for some thread owning writes
// to the chain structure (except for inserts), but only if not an
// "end" pointer. Also called the "rewrite lock."
static constexpr uint64_t kHeadLocked = uint64_t{1} << 7;
// Marker in a "with_shift" pointer for the end of a chain. Must also
// point back to the head of the chain (with end marker removed).
// Also includes the "locked" bit so that attempting to lock an empty
// chain has no effect (not needed, as the lock is only needed for
// removals).
static constexpr uint64_t kNextEndFlags = (uint64_t{1} << 6) | kHeadLocked;
static inline bool IsEnd(uint64_t next_with_shift) {
// Assuming certain values never used, suffices to check this one bit
constexpr auto kCheckBit = kNextEndFlags ^ kHeadLocked;
return next_with_shift & kCheckBit;
}
// Bottom bits to right shift away to get an array index from a
// "with_shift" pointer.
static constexpr int kNextShift = 8;
// A bit mask for the "shift" associated with each "with_shift" pointer.
// Always bottommost bits.
static constexpr int kShiftMask = 63;
// A marker for head_next_with_shift that indicates this HandleImpl is
// heap allocated (standalone) rather than in the table.
static constexpr uint64_t kStandaloneMarker = UINT64_MAX;
// A marker for head_next_with_shift indicating the head is not yet part
// of the usable table, or for chain_next_with_shift indicating that the
// entry is not present or is not yet part of a chain (must not be
// "shareable" state).
static constexpr uint64_t kUnusedMarker = 0;
// See above. The head pointer is logically independent of the rest of
// the entry, including the chain next pointer.
std::atomic<uint64_t> head_next_with_shift{kUnusedMarker};
std::atomic<uint64_t> chain_next_with_shift{kUnusedMarker};
// For supporting CreateStandalone and some fallback cases.
inline bool IsStandalone() const {
return head_next_with_shift.load(std::memory_order_acquire) ==
kStandaloneMarker;
}
inline void SetStandalone() {
head_next_with_shift.store(kStandaloneMarker, std::memory_order_release);
}
}; // struct HandleImpl
struct Opts {
explicit Opts(size_t _min_avg_value_size)
: min_avg_value_size(_min_avg_value_size) {}
explicit Opts(const HyperClockCacheOptions& opts) {
assert(opts.estimated_entry_charge == 0);
min_avg_value_size = opts.min_avg_entry_charge;
}
size_t min_avg_value_size;
};
AutoHyperClockTable(size_t capacity, bool strict_capacity_limit,
CacheMetadataChargePolicy metadata_charge_policy,
MemoryAllocator* allocator,
const Cache::EvictionCallback* eviction_callback,
const uint32_t* hash_seed, const Opts& opts);
~AutoHyperClockTable();
// For BaseClockTable::Insert
struct InsertState {
uint64_t saved_length_info = 0;
size_t likely_empty_slot = 0;
};
void StartInsert(InsertState& state);
// Does initial check for whether there's hash table room for another
// inserted entry, possibly growing if needed. Returns true iff (after
// the call) there is room for the proposed number of entries.
bool GrowIfNeeded(size_t new_occupancy, InsertState& state);
HandleImpl* DoInsert(const ClockHandleBasicData& proto,
uint64_t initial_countdown, bool take_ref,
InsertState& state);
// Runs the clock eviction algorithm trying to reclaim at least
// requested_charge. Returns how much is evicted, which could be less
// if it appears impossible to evict the requested amount without blocking.
void Evict(size_t requested_charge, InsertState& state, EvictionData* data);
HandleImpl* Lookup(const UniqueId64x2& hashed_key);
bool Release(HandleImpl* handle, bool useful, bool erase_if_last_ref);
void Erase(const UniqueId64x2& hashed_key);
void EraseUnRefEntries();
size_t GetTableSize() const;
size_t GetOccupancyLimit() const;
const HandleImpl* HandlePtr(size_t idx) const { return &array_[idx]; }
#ifndef NDEBUG
size_t& TEST_MutableOccupancyLimit() {
return *reinterpret_cast<size_t*>(&occupancy_limit_);
}
// Release N references
void TEST_ReleaseN(HandleImpl* handle, size_t n);
#endif
// Maximum ratio of number of occupied slots to number of usable slots. The
// actual load factor should float pretty close to this number, which should
// be a nice space/time trade-off, though large swings in WriteBufferManager
// memory could lead to low (but very much safe) load factors (only after
// seeing high load factors). Linear hashing along with (modified) linear
// probing to find an available slot increases potential risks of high
// load factors, so are disallowed.
static constexpr double kMaxLoadFactor = 0.60;
private: // functions
// Returns true iff increased usable length. Due to load factor
// considerations, GrowIfNeeded might call this more than once to make room
// for one more entry.
bool Grow(InsertState& state);
// Operational details of splitting a chain into two for Grow().
void SplitForGrow(size_t grow_home, size_t old_home, int old_shift);
// Takes an "under construction" entry and ensures it is no longer connected
// to its home chain (in preparaion for completing erasure and freeing the
// slot). Note that previous operations might have already noticed it being
// "under (de)construction" and removed it from its chain.
void Remove(HandleImpl* h);
// Try to take ownership of an entry and erase+remove it from the table.
// Returns true if successful. Could fail if
// * There are other references to the entry
// * Some other thread has exclusive ownership or has freed it.
bool TryEraseHandle(HandleImpl* h, bool holding_ref, bool mark_invisible);
// Calculates the appropriate maximum table size, for creating the memory
// mapping.
static size_t CalcMaxUsableLength(
size_t capacity, size_t min_avg_value_size,
CacheMetadataChargePolicy metadata_charge_policy);
// Shared helper function that implements removing entries from a chain
// with proper handling to ensure all existing data is seen even in the
// presence of concurrent insertions, etc. (See implementation.)
template <class OpData>
void PurgeImpl(OpData* op_data, size_t home = SIZE_MAX);
// An RAII wrapper for locking a chain of entries for removals. See
// implementation.
class ChainRewriteLock;
// Helper function for PurgeImpl while holding a ChainRewriteLock. See
// implementation.
template <class OpData>
void PurgeImplLocked(OpData* op_data, ChainRewriteLock& rewrite_lock,
size_t home);
// Update length_info_ as much as possible without waiting, given a known
// usable (ready for inserts and lookups) grow_home. (Previous grow_homes
// might not be usable yet, but we can check if they are by looking at
// the corresponding old home.)
void CatchUpLengthInfoNoWait(size_t known_usable_grow_home);
private: // data
// mmaped area holding handles
const TypedMemMapping<HandleImpl> array_;
// Metadata for table size under linear hashing.
//
// Lowest 8 bits are the minimum number of lowest hash bits to use
// ("min shift"). The upper 56 bits are a threshold. If that minumum number
// of bits taken from a hash value is < this threshold, then one more bit of
// hash value is taken and used.
//
// Other mechanisms (shift amounts on pointers) ensure complete availability
// of data already in the table even if a reader only sees a completely
// out-of-date version of this value. In the worst case, it could take
// log time to find the correct chain, but normally this value enables
// readers to find the correct chain on the first try.
//
// To maximize parallelization of Grow() operations, this field is only
// updated opportunistically after Grow() operations and in DoInsert() where
// it is found to be out-of-date. See CatchUpLengthInfoNoWait().
std::atomic<uint64_t> length_info_;
// An already-computed version of the usable length times the max load
// factor. Could be slightly out of date but GrowIfNeeded()/Grow() handle
// that internally.
std::atomic<size_t> occupancy_limit_;
// The next index to use from array_ upon the next Grow(). Might be ahead of
// length_info_.
std::atomic<size_t> grow_frontier_;
// See explanation in AutoHyperClockTable::Evict
std::atomic<size_t> clock_pointer_mask_;
}; // class AutoHyperClockTable
// A single shard of sharded cache.
template <class TableT>
class ALIGN_AS(CACHE_LINE_SIZE) ClockCacheShard final : public CacheShardBase {
public:
using Table = TableT;
ClockCacheShard(size_t capacity, bool strict_capacity_limit,
CacheMetadataChargePolicy metadata_charge_policy,
MemoryAllocator* allocator,
const Cache::EvictionCallback* eviction_callback,
const uint32_t* hash_seed, const typename Table::Opts& opts);
// For CacheShard concept
using HandleImpl = typename Table::HandleImpl;
// Hash is lossless hash of 128-bit key
using HashVal = UniqueId64x2;
using HashCref = const HashVal&;
static inline uint32_t HashPieceForSharding(HashCref hash) {
return Upper32of64(hash[0]);
}
static inline HashVal ComputeHash(const Slice& key, uint32_t seed) {
assert(key.size() == kCacheKeySize);
HashVal in;
HashVal out;
// NOTE: endian dependence
// TODO: use GetUnaligned?
std::memcpy(&in, key.data(), kCacheKeySize);
BijectiveHash2x64(in[1], in[0] ^ seed, &out[1], &out[0]);
return out;
}
// For reconstructing key from hashed_key. Requires the caller to provide
// backing storage for the Slice in `unhashed`
static inline Slice ReverseHash(const UniqueId64x2& hashed,
UniqueId64x2* unhashed, uint32_t seed) {
BijectiveUnhash2x64(hashed[1], hashed[0], &(*unhashed)[1], &(*unhashed)[0]);
(*unhashed)[0] ^= seed;
// NOTE: endian dependence
return Slice(reinterpret_cast<const char*>(unhashed), kCacheKeySize);
}
// Although capacity is dynamically changeable, the number of table slots is
// not, so growing capacity substantially could lead to hitting occupancy
// limit.
void SetCapacity(size_t capacity);
void SetStrictCapacityLimit(bool strict_capacity_limit);
Status Insert(const Slice& key, const UniqueId64x2& hashed_key,
Cache::ObjectPtr value, const Cache::CacheItemHelper* helper,
size_t charge, HandleImpl** handle, Cache::Priority priority);
HandleImpl* CreateStandalone(const Slice& key, const UniqueId64x2& hashed_key,
Cache::ObjectPtr obj,
const Cache::CacheItemHelper* helper,
size_t charge, bool allow_uncharged);
HandleImpl* Lookup(const Slice& key, const UniqueId64x2& hashed_key);
bool Release(HandleImpl* handle, bool useful, bool erase_if_last_ref);
bool Release(HandleImpl* handle, bool erase_if_last_ref = false);
bool Ref(HandleImpl* handle);
void Erase(const Slice& key, const UniqueId64x2& hashed_key);
size_t GetCapacity() const;
size_t GetUsage() const;
size_t GetStandaloneUsage() const;
size_t GetPinnedUsage() const;
size_t GetOccupancyCount() const;
size_t GetOccupancyLimit() const;
size_t GetTableAddressCount() const;
void ApplyToSomeEntries(
const std::function<void(const Slice& key, Cache::ObjectPtr obj,
size_t charge,
const Cache::CacheItemHelper* helper)>& callback,
size_t average_entries_per_lock, size_t* state);
void EraseUnRefEntries();
std::string GetPrintableOptions() const { return std::string{}; }
HandleImpl* Lookup(const Slice& key, const UniqueId64x2& hashed_key,
const Cache::CacheItemHelper* /*helper*/,
Cache::CreateContext* /*create_context*/,
Cache::Priority /*priority*/, Statistics* /*stats*/) {
return Lookup(key, hashed_key);
}
Table& GetTable() { return table_; }
const Table& GetTable() const { return table_; }
#ifndef NDEBUG
size_t& TEST_MutableOccupancyLimit() {
return table_.TEST_MutableOccupancyLimit();
}
// Acquire/release N references
void TEST_RefN(HandleImpl* handle, size_t n);
void TEST_ReleaseN(HandleImpl* handle, size_t n);
#endif
private: // data
Table table_;
// Maximum total charge of all elements stored in the table.
std::atomic<size_t> capacity_;
// Whether to reject insertion if cache reaches its full capacity.
std::atomic<bool> strict_capacity_limit_;
}; // class ClockCacheShard
template <class Table>
class BaseHyperClockCache : public ShardedCache<ClockCacheShard<Table>> {
public:
using Shard = ClockCacheShard<Table>;
using Handle = Cache::Handle;
using CacheItemHelper = Cache::CacheItemHelper;
explicit BaseHyperClockCache(const HyperClockCacheOptions& opts);
Cache::ObjectPtr Value(Handle* handle) override;
size_t GetCharge(Handle* handle) const override;
const CacheItemHelper* GetCacheItemHelper(Handle* handle) const override;
void ReportProblems(
const std::shared_ptr<Logger>& /*info_log*/) const override;
};
class FixedHyperClockCache
#ifdef NDEBUG
final
#endif
: public BaseHyperClockCache<FixedHyperClockTable> {
public:
using BaseHyperClockCache::BaseHyperClockCache;
const char* Name() const override { return "FixedHyperClockCache"; }
void ReportProblems(
const std::shared_ptr<Logger>& /*info_log*/) const override;
}; // class FixedHyperClockCache
class AutoHyperClockCache
#ifdef NDEBUG
final
#endif
: public BaseHyperClockCache<AutoHyperClockTable> {
public:
using BaseHyperClockCache::BaseHyperClockCache;
const char* Name() const override { return "AutoHyperClockCache"; }
void ReportProblems(
const std::shared_ptr<Logger>& /*info_log*/) const override;
}; // class AutoHyperClockCache
} // namespace clock_cache
} // namespace ROCKSDB_NAMESPACE