@nxtedition/rocksdb 5.2.21 → 5.2.26

package/deps/rocksdb/rocksdb/cache/cache_key.cc

@@ -0,0 +1,344 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. 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).
+
+#include "cache/cache_key.h"
+
+#include <algorithm>
+#include <atomic>
+
+#include "rocksdb/cache.h"
+#include "table/unique_id_impl.h"
+#include "util/hash.h"
+#include "util/math.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+// Value space plan for CacheKey:
+//
+// session_etc64_ | offset_etc64_ | Only generated by
+// ---------------+---------------+------------------------------------------
+//              0 |             0 | Reserved for "empty" CacheKey()
+//              0 |  > 0, < 1<<63 | CreateUniqueForCacheLifetime
+//              0 |      >= 1<<63 | CreateUniqueForProcessLifetime
+//            > 0 |           any | OffsetableCacheKey.WithOffset
+
+CacheKey CacheKey::CreateUniqueForCacheLifetime(Cache *cache) {
+  // +1 so that we can reserve all zeros for "unset" cache key
+  uint64_t id = cache->NewId() + 1;
+  // Ensure we don't collide with CreateUniqueForProcessLifetime
+  assert((id >> 63) == 0U);
+  return CacheKey(0, id);
+}
+
+CacheKey CacheKey::CreateUniqueForProcessLifetime() {
+  // To avoid colliding with CreateUniqueForCacheLifetime, assuming
+  // Cache::NewId counts up from zero, here we count down from UINT64_MAX.
+  // If this ever becomes a point of contention, we could sub-divide the
+  // space and use CoreLocalArray.
+  static std::atomic<uint64_t> counter{UINT64_MAX};
+  uint64_t id = counter.fetch_sub(1, std::memory_order_relaxed);
+  // Ensure we don't collide with CreateUniqueForCacheLifetime
+  assert((id >> 63) == 1U);
+  return CacheKey(0, id);
+}
+
+// Value plan for CacheKeys from OffsetableCacheKey, assuming that
+// db_session_ids are generated from a base_session_id and
+// session_id_counter (by SemiStructuredUniqueIdGen+EncodeSessionId
+// in DBImpl::GenerateDbSessionId):
+//
+// Conceptual inputs:
+//   db_id                   (unstructured, from GenerateRawUniqueId or equiv)
+//                           * could be shared between cloned DBs but rare
+//                           * could be constant, if session id suffices
+//   base_session_id         (unstructured, from GenerateRawUniqueId)
+//   session_id_counter      (structured)
+//                           * usually much smaller than 2**24
+//   file_number             (structured)
+//                           * usually smaller than 2**24
+//   offset_in_file          (structured, might skip lots of values)
+//                           * usually smaller than 2**32
+//   max_offset              determines placement of file_number to prevent
+//                           overlapping with offset
+//
+// Outputs come from bitwise-xor of the constituent pieces, low bits on left:
+//
+// |------------------------- session_etc64 -------------------------|
+// | +++++++++++++++ base_session_id (lower 64 bits) +++++++++++++++ |
+// |-----------------------------------------------------------------|
+// | session_id_counter ...|                                         |
+// |-----------------------------------------------------------------|
+// |                                               | ... file_number |
+// |                                               | overflow & meta |
+// |-----------------------------------------------------------------|
+//
+//
+// |------------------------- offset_etc64 --------------------------|
+// | hash of: ++++++++++++++++++++++++++++++++++++++++++++++++++++++ |
+// |  * base_session_id (upper ~39 bits)                             |
+// |  * db_id (~122 bits entropy)                                    |
+// |-----------------------------------------------------------------|
+// | offset_in_file ............... |                                |
+// |-----------------------------------------------------------------|
+// |                                              | file_number, 0-3 |
+// |                                              | lower bytes      |
+// |-----------------------------------------------------------------|
+//
+// Based on max_offset, a maximal number of bytes 0..3 is chosen for
+// including from lower bits of file_number in offset_etc64. The choice
+// is encoded in two bits of metadata going into session_etc64, though
+// the common case of 3 bytes is encoded as 0 so that session_etc64
+// is unmodified by file_number concerns in the common case.
+//
+// There is nothing preventing "file number overflow & meta" from meeting
+// and overlapping with session_id_counter, but reaching such a case requires
+// an intractable combination of large file offsets (thus at least some large
+// files), large file numbers (thus large number of files generated), and
+// large number of session IDs generated in a single process. A trillion each
+// (2**40) of session ids, offsets, and file numbers comes to 120 bits.
+// With two bits of metadata and byte granularity, this is on the verge of
+// overlap, but even in the overlap case, it doesn't seem likely that
+// a file from billions of files or session ids ago will still be live
+// or cached.
+//
+// In fact, if our SST files are all < 4TB (see
+// BlockBasedTable::kMaxFileSizeStandardEncoding), then SST files generated
+// in a single process are guaranteed to have unique cache keys, unless/until
+// number session ids * max file number = 2**86, e.g. 1 trillion DB::Open in
+// a single process and 64 trillion files generated. Even at that point, to
+// see a collision we would need a miraculous re-synchronization of session
+// id and file number, along with a live file or stale cache entry from
+// trillions of files ago.
+//
+// How https://github.com/pdillinger/unique_id applies here:
+// Every bit of output always includes "unstructured" uniqueness bits and
+// often combines with "structured" uniqueness bits. The "unstructured" bits
+// change infrequently: only when we cannot guarantee our state tracking for
+// "structured" uniqueness hasn't been cloned. Using a static
+// SemiStructuredUniqueIdGen for db_session_ids, this means we only get an
+// "all new" session id when a new process uses RocksDB. (Between processes,
+// we don't know if a DB or other persistent storage has been cloned. We
+// assume that if VM hot cloning is used, subsequently generated SST files
+// do not interact.) Within a process, only the session_lower of the
+// db_session_id changes incrementally ("structured" uniqueness).
+//
+// This basically means that our offsets, counters and file numbers allow us
+// to do somewhat "better than random" (birthday paradox) while in the
+// degenerate case of completely new session for each tiny file, we still
+// have strong uniqueness properties from the birthday paradox, with ~103
+// bit session IDs or up to 128 bits entropy with different DB IDs sharing a
+// cache.
+//
+// More collision probability analysis:
+// Suppose a RocksDB host generates (generously) 2 GB/s (10TB data, 17 DWPD)
+// with average process/session lifetime of (pessimistically) 4 minutes.
+// In 180 days (generous allowable data lifespan), we generate 31 million GB
+// of data, or 2^55 bytes, and 2^16 "all new" session IDs.
+//
+// First, suppose this is in a single DB (lifetime 180 days):
+// 128 bits cache key size
+// - 55 <- ideal size for byte offsets + file numbers
+// -  2 <- bits for offsets and file numbers not exactly powers of two
+// -  2 <- bits for file number encoding metadata
+// +  2 <- bits saved not using byte offsets in BlockBasedTable::GetCacheKey
+// ----
+//   71 <- bits remaining for distinguishing session IDs
+// The probability of a collision in 71 bits of session ID data is less than
+// 1 in 2**(71 - (2 * 16)), or roughly 1 in a trillion. And this assumes all
+// data from the last 180 days is in cache for potential collision, and that
+// cache keys under each session id exhaustively cover the remaining 57 bits
+// while in reality they'll only cover a small fraction of it.
+//
+// Although data could be transferred between hosts, each host has its own
+// cache and we are already assuming a high rate of "all new" session ids.
+// So this doesn't really change the collision calculation. Across a fleet
+// of 1 million, each with <1 in a trillion collision possibility,
+// fleetwide collision probability is <1 in a million.
+//
+// Now suppose we have many DBs per host, say 2**10, with same host-wide write
+// rate and process/session lifetime. File numbers will be ~10 bits smaller
+// and we will have 2**10 times as many session IDs because of simultaneous
+// lifetimes. So now collision chance is less than 1 in 2**(81 - (2 * 26)),
+// or roughly 1 in a billion.
+//
+// Suppose instead we generated random or hashed cache keys for each
+// (compressed) block. For 1KB compressed block size, that is 2^45 cache keys
+// in 180 days. Collision probability is more easily estimated at roughly
+// 1 in 2**(128 - (2 * 45)) or roughly 1 in a trillion (assuming all
+// data from the last 180 days is in cache, but NOT the other assumption
+// for the 1 in a trillion estimate above).
+//
+//
+// Collision probability estimation through simulation:
+// A tool ./cache_bench -stress_cache_key broadly simulates host-wide cache
+// activity over many months, by making some pessimistic simplifying
+// assumptions. See class StressCacheKey in cache_bench_tool.cc for details.
+// Here is some sample output with
+// `./cache_bench -stress_cache_key -sck_keep_bits=40`:
+//
+//   Total cache or DBs size: 32TiB  Writing 925.926 MiB/s or 76.2939TiB/day
+//   Multiply by 9.22337e+18 to correct for simulation losses (but still
+//   assume whole file cached)
+//
+// These come from default settings of 2.5M files per day of 32 MB each, and
+// `-sck_keep_bits=40` means that to represent a single file, we are only
+// keeping 40 bits of the 128-bit (base) cache key.  With file size of 2**25
+// contiguous keys (pessimistic), our simulation is about 2\*\*(128-40-25) or
+// about 9 billion billion times more prone to collision than reality.
+//
+// More default assumptions, relatively pessimistic:
+// * 100 DBs in same process (doesn't matter much)
+// * Re-open DB in same process (new session ID related to old session ID) on
+// average every 100 files generated
+// * Restart process (all new session IDs unrelated to old) 24 times per day
+//
+// After enough data, we get a result at the end (-sck_keep_bits=40):
+//
+//   (keep 40 bits)  17 collisions after 2 x 90 days, est 10.5882 days between
+//                   (9.76592e+19 corrected)
+//
+// If we believe the (pessimistic) simulation and the mathematical
+// extrapolation, we would need to run a billion machines all for 97 billion
+// days to expect a cache key collision. To help verify that our extrapolation
+// ("corrected") is robust, we can make our simulation more precise with
+// `-sck_keep_bits=41` and `42`, which takes more running time to get enough
+// collision data:
+//
+//   (keep 41 bits)  16 collisions after 4 x 90 days, est 22.5 days between
+//                   (1.03763e+20 corrected)
+//   (keep 42 bits)  19 collisions after 10 x 90 days, est 47.3684 days between
+//                   (1.09224e+20 corrected)
+//
+// The extrapolated prediction is very close. If anything, we might have some
+// very small losses of structured data (see class StressCacheKey in
+// cache_bench_tool.cc) leading to more accurate & more attractive prediction
+// with more bits kept.
+//
+// With the `-sck_randomize` option, we can see that typical workloads like
+// above have lower collision probability than "random" cache keys (note:
+// offsets still non-randomized) by a modest amount (roughly 20x less collision
+// prone than random), which should make us reasonably comfortable even in
+// "degenerate" cases (e.g. repeatedly launch a process to generate 1 file
+// with SstFileWriter):
+//
+//   (rand 40 bits) 197 collisions after 1 x 90 days, est 0.456853 days between
+//                  (4.21372e+18 corrected)
+//
+// We can see that with more frequent process restarts (all new session IDs),
+// we get closer to the "random" cache key performance:
+//
+//   (-sck_restarts_per_day=5000): 140 collisions after 1 x 90 days, ...
+//                  (5.92931e+18 corrected)
+//
+// Other tests have been run to validate other conditions behave as expected,
+// never behaving "worse than random" unless we start chopping off structured
+// data.
+//
+//
+// Conclusion: Even in extreme cases, rapidly burning through "all new" IDs
+// that only arise when a new process is started, the chance of any cache key
+// collisions in a giant fleet of machines is negligible. Especially when
+// processes live for hours or days, the chance of a cache key collision is
+// likely more plausibly due to bad hardware than to bad luck in random
+// session ID data. Software defects are surely more likely to cause corruption
+// than both of those.
+//
+// TODO: Nevertheless / regardless, an efficient way to detect (and thus
+// quantify) block cache corruptions, including collisions, should be added.
+OffsetableCacheKey::OffsetableCacheKey(const std::string &db_id,
+                                       const std::string &db_session_id,
+                                       uint64_t file_number,
+                                       uint64_t max_offset) {
+#ifndef NDEBUG
+  max_offset_ = max_offset;
+#endif
+  // Closely related to GetSstInternalUniqueId, but only need 128 bits and
+  // need to include an offset within the file.
+  // See also https://github.com/pdillinger/unique_id for background.
+  uint64_t session_upper = 0;  // Assignment to appease clang-analyze
+  uint64_t session_lower = 0;  // Assignment to appease clang-analyze
+  {
+    Status s = DecodeSessionId(db_session_id, &session_upper, &session_lower);
+    if (!s.ok()) {
+      // A reasonable fallback in case malformed
+      Hash2x64(db_session_id.data(), db_session_id.size(), &session_upper,
+               &session_lower);
+    }
+  }
+
+  // Hash the session upper (~39 bits entropy) and DB id (120+ bits entropy)
+  // for more global uniqueness entropy.
+  // (It is possible that many DBs descended from one common DB id are copied
+  // around and proliferate, in which case session id is critical, but it is
+  // more common for different DBs to have different DB ids.)
+  uint64_t db_hash = Hash64(db_id.data(), db_id.size(), session_upper);
+
+  // This establishes the db+session id part of the cache key.
+  //
+  // Exactly preserve (in common cases; see modifiers below) session lower to
+  // ensure that session ids generated during the same process lifetime are
+  // guaranteed unique.
+  //
+  // We put this first for CommonPrefixSlice(), so that a small-ish set of
+  // cache key prefixes to cover entries relevant to any DB.
+  session_etc64_ = session_lower;
+  // This provides extra entopy in case of different DB id or process
+  // generating a session id, but is also partly/variably obscured by
+  // file_number and offset (see below).
+  offset_etc64_ = db_hash;
+
+  // Into offset_etc64_ we are (eventually) going to pack & xor in an offset and
+  // a file_number, but we might need the file_number to overflow into
+  // session_etc64_. (There must only be one session_etc64_ value per
+  // file, and preferably shared among many files.)
+  //
+  // Figure out how many bytes of file_number we are going to be able to
+  // pack in with max_offset, though our encoding will only support packing
+  // in up to 3 bytes of file_number. (16M file numbers is enough for a new
+  // file number every second for half a year.)
+  int file_number_bytes_in_offset_etc =
+      (63 - FloorLog2(max_offset | 0x100000000U)) / 8;
+  int file_number_bits_in_offset_etc = file_number_bytes_in_offset_etc * 8;
+
+  // Assert two bits of metadata
+  assert(file_number_bytes_in_offset_etc >= 0 &&
+         file_number_bytes_in_offset_etc <= 3);
+  // Assert we couldn't have used a larger allowed number of bytes (shift
+  // would chop off bytes).
+  assert(file_number_bytes_in_offset_etc == 3 ||
+         (max_offset << (file_number_bits_in_offset_etc + 8) >>
+          (file_number_bits_in_offset_etc + 8)) != max_offset);
+
+  uint64_t mask = (uint64_t{1} << (file_number_bits_in_offset_etc)) - 1;
+  // Pack into high bits of etc so that offset can go in low bits of etc
+  // TODO: could be EndianSwapValue?
+  uint64_t offset_etc_modifier = ReverseBits(file_number & mask);
+  assert(offset_etc_modifier << file_number_bits_in_offset_etc == 0U);
+
+  // Overflow and 3 - byte count (likely both zero) go into session_id part
+  uint64_t session_etc_modifier =
+      (file_number >> file_number_bits_in_offset_etc << 2) |
+      static_cast<uint64_t>(3 - file_number_bytes_in_offset_etc);
+  // Packed into high bits to minimize interference with session id counter.
+  session_etc_modifier = ReverseBits(session_etc_modifier);
+
+  // Assert session_id part is only modified in extreme cases
+  assert(session_etc_modifier == 0 || file_number > /*3 bytes*/ 0xffffffU ||
+         max_offset > /*5 bytes*/ 0xffffffffffU);
+
+  // Xor in the modifiers
+  session_etc64_ ^= session_etc_modifier;
+  offset_etc64_ ^= offset_etc_modifier;
+
+  // Although DBImpl guarantees (in recent versions) that session_lower is not
+  // zero, that's not entirely sufficient to guarantee that session_etc64_ is
+  // not zero (so that the 0 case can be used by CacheKey::CreateUnique*)
+  if (session_etc64_ == 0U) {
+    session_etc64_ = session_upper | 1U;
+  }
+  assert(session_etc64_ != 0);
+}
+
+}  // namespace ROCKSDB_NAMESPACE

package/deps/rocksdb/rocksdb/cache/cache_key.h

@@ -0,0 +1,132 @@
+//  Copyright (c) Facebook, Inc. and its affiliates. 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).
+
+#pragma once
+
+#include <cstdint>
+
+#include "rocksdb/rocksdb_namespace.h"
+#include "rocksdb/slice.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+class Cache;
+
+// A standard holder for fixed-size block cache keys (and for related caches).
+// They are created through one of these, each using its own range of values:
+// * CacheKey::CreateUniqueForCacheLifetime
+// * CacheKey::CreateUniqueForProcessLifetime
+// * Default ctor ("empty" cache key)
+// * OffsetableCacheKey->WithOffset
+//
+// The first two use atomic counters to guarantee uniqueness over the given
+// lifetime and the last uses a form of universally unique identifier for
+// uniqueness with very high probabilty (and guaranteed for files generated
+// during a single process lifetime).
+//
+// CacheKeys are currently used by calling AsSlice() to pass as a key to
+// Cache. For performance, the keys are endianness-dependent (though otherwise
+// portable). (Persistable cache entries are not intended to cross platforms.)
+class CacheKey {
+ public:
+  // For convenience, constructs an "empty" cache key that is never returned
+  // by other means.
+  inline CacheKey() : session_etc64_(), offset_etc64_() {}
+
+  inline bool IsEmpty() const {
+    return (session_etc64_ == 0) & (offset_etc64_ == 0);
+  }
+
+  // Use this cache key as a Slice (byte order is endianness-dependent)
+  inline Slice AsSlice() const {
+    static_assert(sizeof(*this) == 16, "Standardized on 16-byte cache key");
+    assert(!IsEmpty());
+    return Slice(reinterpret_cast<const char *>(this), sizeof(*this));
+  }
+
+  // Create a CacheKey that is unique among others associated with this Cache
+  // instance. Depends on Cache::NewId. This is useful for block cache
+  // "reservations".
+  static CacheKey CreateUniqueForCacheLifetime(Cache *cache);
+
+  // Create a CacheKey that is unique among others for the lifetime of this
+  // process. This is useful for saving in a static data member so that
+  // different DB instances can agree on a cache key for shared entities,
+  // such as for CacheEntryStatsCollector.
+  static CacheKey CreateUniqueForProcessLifetime();
+
+ protected:
+  friend class OffsetableCacheKey;
+  CacheKey(uint64_t session_etc64, uint64_t offset_etc64)
+      : session_etc64_(session_etc64), offset_etc64_(offset_etc64) {}
+  uint64_t session_etc64_;
+  uint64_t offset_etc64_;
+};
+
+// A file-specific generator of cache keys, sometimes referred to as the
+// "base" cache key for a file because all the cache keys for various offsets
+// within the file are computed using simple arithmetic. The basis for the
+// general approach is dicussed here: https://github.com/pdillinger/unique_id
+// Heavily related to GetUniqueIdFromTableProperties.
+//
+// If the db_id, db_session_id, and file_number come from the file's table
+// properties, then the keys will be stable across DB::Open/Close, backup/
+// restore, import/export, etc.
+//
+// This class "is a" CacheKey only privately so that it is not misused as
+// a ready-to-use CacheKey.
+class OffsetableCacheKey : private CacheKey {
+ public:
+  // For convenience, constructs an "empty" cache key that should not be used.
+  inline OffsetableCacheKey() : CacheKey() {}
+
+  // Constructs an OffsetableCacheKey with the given information about a file.
+  // max_offset is based on file size (see WithOffset) and is required here to
+  // choose an appropriate (sub-)encoding. This constructor never generates an
+  // "empty" base key.
+  OffsetableCacheKey(const std::string &db_id, const std::string &db_session_id,
+                     uint64_t file_number, uint64_t max_offset);
+
+  inline bool IsEmpty() const {
+    bool result = session_etc64_ == 0;
+    assert(!(offset_etc64_ > 0 && result));
+    return result;
+  }
+
+  // Construct a CacheKey for an offset within a file, which must be
+  // <= max_offset provided in constructor. An offset is not necessarily a
+  // byte offset if a smaller unique identifier of keyable offsets is used.
+  //
+  // This class was designed to make this hot code extremely fast.
+  inline CacheKey WithOffset(uint64_t offset) const {
+    assert(!IsEmpty());
+    assert(offset <= max_offset_);
+    return CacheKey(session_etc64_, offset_etc64_ ^ offset);
+  }
+
+  // The "common prefix" is a shared prefix for all the returned CacheKeys,
+  // that also happens to usually be the same among many files in the same DB,
+  // so is efficient and highly accurate (not perfectly) for DB-specific cache
+  // dump selection (but not file-specific).
+  static constexpr size_t kCommonPrefixSize = 8;
+  inline Slice CommonPrefixSlice() const {
+    static_assert(sizeof(session_etc64_) == kCommonPrefixSize,
+                  "8 byte common prefix expected");
+    assert(!IsEmpty());
+    assert(&this->session_etc64_ == static_cast<const void *>(this));
+
+    return Slice(reinterpret_cast<const char *>(this), kCommonPrefixSize);
+  }
+
+  // For any max_offset <= this value, the same encoding scheme is guaranteed.
+  static constexpr uint64_t kMaxOffsetStandardEncoding = 0xffffffffffU;
+
+ private:
+#ifndef NDEBUG
+  uint64_t max_offset_ = 0;
+#endif
+};
+
+}  // namespace ROCKSDB_NAMESPACE

package/deps/rocksdb/rocksdb/cache/cache_reservation_manager.cc

@@ -0,0 +1,183 @@
+//  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.
+#include "cache/cache_reservation_manager.h"
+
+#include <cassert>
+#include <cstddef>
+#include <cstring>
+#include <memory>
+
+#include "cache/cache_entry_roles.h"
+#include "rocksdb/cache.h"
+#include "rocksdb/slice.h"
+#include "rocksdb/status.h"
+#include "table/block_based/reader_common.h"
+#include "util/coding.h"
+
+namespace ROCKSDB_NAMESPACE {
+
+template <CacheEntryRole R>
+CacheReservationManagerImpl<R>::CacheReservationHandle::CacheReservationHandle(
+    std::size_t incremental_memory_used,
+    std::shared_ptr<CacheReservationManagerImpl> cache_res_mgr)
+    : incremental_memory_used_(incremental_memory_used) {
+  assert(cache_res_mgr);
+  cache_res_mgr_ = cache_res_mgr;
+}
+
+template <CacheEntryRole R>
+CacheReservationManagerImpl<
+    R>::CacheReservationHandle::~CacheReservationHandle() {
+  Status s = cache_res_mgr_->ReleaseCacheReservation(incremental_memory_used_);
+  s.PermitUncheckedError();
+}
+
+template <CacheEntryRole R>
+CacheReservationManagerImpl<R>::CacheReservationManagerImpl(
+    std::shared_ptr<Cache> cache, bool delayed_decrease)
+    : delayed_decrease_(delayed_decrease),
+      cache_allocated_size_(0),
+      memory_used_(0) {
+  assert(cache != nullptr);
+  cache_ = cache;
+}
+
+template <CacheEntryRole R>
+CacheReservationManagerImpl<R>::~CacheReservationManagerImpl() {
+  for (auto* handle : dummy_handles_) {
+    cache_->Release(handle, true);
+  }
+}
+
+template <CacheEntryRole R>
+Status CacheReservationManagerImpl<R>::UpdateCacheReservation(
+    std::size_t new_mem_used) {
+  memory_used_ = new_mem_used;
+  std::size_t cur_cache_allocated_size =
+      cache_allocated_size_.load(std::memory_order_relaxed);
+  if (new_mem_used == cur_cache_allocated_size) {
+    return Status::OK();
+  } else if (new_mem_used > cur_cache_allocated_size) {
+    Status s = IncreaseCacheReservation(new_mem_used);
+    return s;
+  } else {
+    // In delayed decrease mode, we don't decrease cache reservation
+    // untill the memory usage is less than 3/4 of what we reserve
+    // in the cache.
+    // We do this because
+    // (1) Dummy entry insertion is expensive in block cache
+    // (2) Delayed releasing previously inserted dummy entries can save such
+    // expensive dummy entry insertion on memory increase in the near future,
+    // which is likely to happen when the memory usage is greater than or equal
+    // to 3/4 of what we reserve
+    if (delayed_decrease_ && new_mem_used >= cur_cache_allocated_size / 4 * 3) {
+      return Status::OK();
+    } else {
+      Status s = DecreaseCacheReservation(new_mem_used);
+      return s;
+    }
+  }
+}
+
+template <CacheEntryRole R>
+Status CacheReservationManagerImpl<R>::MakeCacheReservation(
+    std::size_t incremental_memory_used,
+    std::unique_ptr<CacheReservationManager::CacheReservationHandle>* handle) {
+  assert(handle);
+  Status s =
+      UpdateCacheReservation(GetTotalMemoryUsed() + incremental_memory_used);
+  (*handle).reset(new CacheReservationManagerImpl::CacheReservationHandle(
+      incremental_memory_used,
+      std::enable_shared_from_this<
+          CacheReservationManagerImpl<R>>::shared_from_this()));
+  return s;
+}
+
+template <CacheEntryRole R>
+Status CacheReservationManagerImpl<R>::ReleaseCacheReservation(
+    std::size_t incremental_memory_used) {
+  assert(GetTotalMemoryUsed() >= incremental_memory_used);
+  std::size_t updated_total_mem_used =
+      GetTotalMemoryUsed() - incremental_memory_used;
+  Status s = UpdateCacheReservation(updated_total_mem_used);
+  return s;
+}
+
+template <CacheEntryRole R>
+Status CacheReservationManagerImpl<R>::IncreaseCacheReservation(
+    std::size_t new_mem_used) {
+  Status return_status = Status::OK();
+  while (new_mem_used > cache_allocated_size_.load(std::memory_order_relaxed)) {
+    Cache::Handle* handle = nullptr;
+    return_status = cache_->Insert(GetNextCacheKey(), nullptr, kSizeDummyEntry,
+                                   GetNoopDeleterForRole<R>(), &handle);
+
+    if (return_status != Status::OK()) {
+      return return_status;
+    }
+
+    dummy_handles_.push_back(handle);
+    cache_allocated_size_ += kSizeDummyEntry;
+  }
+  return return_status;
+}
+
+template <CacheEntryRole R>
+Status CacheReservationManagerImpl<R>::DecreaseCacheReservation(
+    std::size_t new_mem_used) {
+  Status return_status = Status::OK();
+
+  // Decrease to the smallest multiple of kSizeDummyEntry that is greater than
+  // or equal to new_mem_used We do addition instead of new_mem_used <=
+  // cache_allocated_size_.load(std::memory_order_relaxed) - kSizeDummyEntry to
+  // avoid underflow of size_t when cache_allocated_size_ = 0
+  while (new_mem_used + kSizeDummyEntry <=
+         cache_allocated_size_.load(std::memory_order_relaxed)) {
+    assert(!dummy_handles_.empty());
+    auto* handle = dummy_handles_.back();
+    cache_->Release(handle, true);
+    dummy_handles_.pop_back();
+    cache_allocated_size_ -= kSizeDummyEntry;
+  }
+  return return_status;
+}
+
+template <CacheEntryRole R>
+std::size_t CacheReservationManagerImpl<R>::GetTotalReservedCacheSize() {
+  return cache_allocated_size_.load(std::memory_order_relaxed);
+}
+
+template <CacheEntryRole R>
+std::size_t CacheReservationManagerImpl<R>::GetTotalMemoryUsed() {
+  return memory_used_;
+}
+
+template <CacheEntryRole R>
+Slice CacheReservationManagerImpl<R>::GetNextCacheKey() {
+  // Calling this function will have the side-effect of changing the
+  // underlying cache_key_ that is shared among other keys generated from this
+  // fucntion. Therefore please make sure the previous keys are saved/copied
+  // before calling this function.
+  cache_key_ = CacheKey::CreateUniqueForCacheLifetime(cache_.get());
+  return cache_key_.AsSlice();
+}
+
+template <CacheEntryRole R>
+Cache::DeleterFn CacheReservationManagerImpl<R>::TEST_GetNoopDeleterForRole() {
+  return GetNoopDeleterForRole<R>();
+}
+
+template class CacheReservationManagerImpl<
+    CacheEntryRole::kBlockBasedTableReader>;
+template class CacheReservationManagerImpl<
+    CacheEntryRole::kCompressionDictionaryBuildingBuffer>;
+template class CacheReservationManagerImpl<CacheEntryRole::kFilterConstruction>;
+template class CacheReservationManagerImpl<CacheEntryRole::kMisc>;
+template class CacheReservationManagerImpl<CacheEntryRole::kWriteBuffer>;
+}  // namespace ROCKSDB_NAMESPACE