rocksdb/memtable/write_buffer_manager.cc
Akanksha Mahajan 596e9008e4 Stall writes in WriteBufferManager when memory_usage exceeds buffer_size (#7898)
Summary:
When WriteBufferManager is shared across DBs and column families
to maintain memory usage under a limit, OOMs have been observed when flush cannot
finish but writes continuously insert to memtables.
In order to avoid OOMs, when memory usage goes beyond buffer_limit_ and DBs tries to write,
this change will stall incoming writers until flush is completed and memory_usage
drops.

Design: Stall condition: When total memory usage exceeds WriteBufferManager::buffer_size_
(memory_usage() >= buffer_size_) WriterBufferManager::ShouldStall() returns true.

DBImpl first block incoming/future writers by calling write_thread_.BeginWriteStall()
(which adds dummy stall object to the writer's queue).
Then DB is blocked on a state State::Blocked (current write doesn't go
through). WBStallInterface object maintained by every DB instance is added to the queue of
WriteBufferManager.

If multiple DBs tries to write during this stall, they will also be
blocked when check WriteBufferManager::ShouldStall() returns true.

End Stall condition: When flush is finished and memory usage goes down, stall will end only if memory
waiting to be flushed is less than buffer_size/2. This lower limit will give time for flush
to complete and avoid continous stalling if memory usage remains close to buffer_size.

WriterBufferManager::EndWriteStall() is called,
which removes all instances from its queue and signal them to continue.
Their state is changed to State::Running and they are unblocked. DBImpl
then signal all incoming writers of that DB to continue by calling
write_thread_.EndWriteStall() (which removes dummy stall object from the
queue).

DB instance creates WBMStallInterface which is an interface to block and
signal DBs during stall.
When DB needs to be blocked or signalled by WriteBufferManager,
state_for_wbm_ state is changed accordingly (RUNNING or BLOCKED).

Pull Request resolved: https://github.com/facebook/rocksdb/pull/7898

Test Plan: Added a new test db/db_write_buffer_manager_test.cc

Reviewed By: anand1976

Differential Revision: D26093227

Pulled By: akankshamahajan15

fbshipit-source-id: 2bbd982a3fb7033f6de6153aa92a221249861aae
2021-04-21 13:54:02 -07:00

228 lines
7.7 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.
#include "rocksdb/write_buffer_manager.h"
#include "db/db_impl/db_impl.h"
#include "util/coding.h"
namespace ROCKSDB_NAMESPACE {
#ifndef ROCKSDB_LITE
namespace {
const size_t kSizeDummyEntry = 256 * 1024;
// The key will be longer than keys for blocks in SST files so they won't
// conflict.
const size_t kCacheKeyPrefix = kMaxVarint64Length * 4 + 1;
} // namespace
struct WriteBufferManager::CacheRep {
std::shared_ptr<Cache> cache_;
std::mutex cache_mutex_;
std::atomic<size_t> cache_allocated_size_;
// The non-prefix part will be updated according to the ID to use.
char cache_key_[kCacheKeyPrefix + kMaxVarint64Length];
uint64_t next_cache_key_id_ = 0;
std::vector<Cache::Handle*> dummy_handles_;
explicit CacheRep(std::shared_ptr<Cache> cache)
: cache_(cache), cache_allocated_size_(0) {
memset(cache_key_, 0, kCacheKeyPrefix);
size_t pointer_size = sizeof(const void*);
assert(pointer_size <= kCacheKeyPrefix);
memcpy(cache_key_, static_cast<const void*>(this), pointer_size);
}
Slice GetNextCacheKey() {
memset(cache_key_ + kCacheKeyPrefix, 0, kMaxVarint64Length);
char* end =
EncodeVarint64(cache_key_ + kCacheKeyPrefix, next_cache_key_id_++);
return Slice(cache_key_, static_cast<size_t>(end - cache_key_));
}
};
#else
struct WriteBufferManager::CacheRep {};
#endif // ROCKSDB_LITE
WriteBufferManager::WriteBufferManager(size_t _buffer_size,
std::shared_ptr<Cache> cache,
bool allow_stall)
: buffer_size_(_buffer_size),
mutable_limit_(buffer_size_ * 7 / 8),
memory_used_(0),
memory_active_(0),
dummy_size_(0),
cache_rep_(nullptr),
allow_stall_(allow_stall),
stall_active_(false) {
#ifndef ROCKSDB_LITE
if (cache) {
// Construct the cache key using the pointer to this.
cache_rep_.reset(new CacheRep(cache));
}
#else
(void)cache;
#endif // ROCKSDB_LITE
}
WriteBufferManager::~WriteBufferManager() {
#ifndef ROCKSDB_LITE
if (cache_rep_) {
for (auto* handle : cache_rep_->dummy_handles_) {
if (handle != nullptr) {
cache_rep_->cache_->Release(handle, true);
}
}
}
#endif // ROCKSDB_LITE
}
void WriteBufferManager::ReserveMem(size_t mem) {
if (cache_rep_ != nullptr) {
ReserveMemWithCache(mem);
} else if (enabled()) {
memory_used_.fetch_add(mem, std::memory_order_relaxed);
}
if (enabled()) {
memory_active_.fetch_add(mem, std::memory_order_relaxed);
}
}
// Should only be called from write thread
void WriteBufferManager::ReserveMemWithCache(size_t mem) {
#ifndef ROCKSDB_LITE
assert(cache_rep_ != nullptr);
// Use a mutex to protect various data structures. Can be optimized to a
// lock-free solution if it ends up with a performance bottleneck.
std::lock_guard<std::mutex> lock(cache_rep_->cache_mutex_);
size_t new_mem_used = memory_used_.load(std::memory_order_relaxed) + mem;
memory_used_.store(new_mem_used, std::memory_order_relaxed);
while (new_mem_used > cache_rep_->cache_allocated_size_) {
// Expand size by at least 256KB.
// Add a dummy record to the cache
Cache::Handle* handle = nullptr;
Status s =
cache_rep_->cache_->Insert(cache_rep_->GetNextCacheKey(), nullptr,
kSizeDummyEntry, nullptr, &handle);
s.PermitUncheckedError(); // TODO: What to do on error?
// We keep the handle even if insertion fails and a null handle is
// returned, so that when memory shrinks, we don't release extra
// entries from cache.
// Ideallly we should prevent this allocation from happening if
// this insertion fails. However, the callers to this code path
// are not able to handle failures properly. We'll need to improve
// it in the future.
cache_rep_->dummy_handles_.push_back(handle);
cache_rep_->cache_allocated_size_ += kSizeDummyEntry;
dummy_size_.fetch_add(kSizeDummyEntry, std::memory_order_relaxed);
}
#else
(void)mem;
#endif // ROCKSDB_LITE
}
void WriteBufferManager::ScheduleFreeMem(size_t mem) {
if (enabled()) {
memory_active_.fetch_sub(mem, std::memory_order_relaxed);
}
}
void WriteBufferManager::FreeMem(size_t mem) {
if (cache_rep_ != nullptr) {
FreeMemWithCache(mem);
} else if (enabled()) {
memory_used_.fetch_sub(mem, std::memory_order_relaxed);
}
// Check if stall is active and can be ended.
if (allow_stall_) {
EndWriteStall();
}
}
void WriteBufferManager::FreeMemWithCache(size_t mem) {
#ifndef ROCKSDB_LITE
assert(cache_rep_ != nullptr);
// Use a mutex to protect various data structures. Can be optimized to a
// lock-free solution if it ends up with a performance bottleneck.
std::lock_guard<std::mutex> lock(cache_rep_->cache_mutex_);
size_t new_mem_used = memory_used_.load(std::memory_order_relaxed) - mem;
memory_used_.store(new_mem_used, std::memory_order_relaxed);
// Gradually shrink memory costed in the block cache if the actual
// usage is less than 3/4 of what we reserve from the block cache.
// We do this because:
// 1. we don't pay the cost of the block cache immediately a memtable is
// freed, as block cache insert is expensive;
// 2. eventually, if we walk away from a temporary memtable size increase,
// we make sure shrink the memory costed in block cache over time.
// In this way, we only shrink costed memory showly even there is enough
// margin.
if (new_mem_used < cache_rep_->cache_allocated_size_ / 4 * 3 &&
cache_rep_->cache_allocated_size_ - kSizeDummyEntry > new_mem_used) {
assert(!cache_rep_->dummy_handles_.empty());
auto* handle = cache_rep_->dummy_handles_.back();
// If insert failed, handle is null so we should not release.
if (handle != nullptr) {
cache_rep_->cache_->Release(handle, true);
}
cache_rep_->dummy_handles_.pop_back();
cache_rep_->cache_allocated_size_ -= kSizeDummyEntry;
dummy_size_.fetch_sub(kSizeDummyEntry, std::memory_order_relaxed);
}
#else
(void)mem;
#endif // ROCKSDB_LITE
}
void WriteBufferManager::BeginWriteStall(StallInterface* wbm_stall) {
assert(wbm_stall != nullptr);
if (wbm_stall) {
std::unique_lock<std::mutex> lock(mu_);
queue_.push_back(wbm_stall);
}
// In case thread enqueue itself and memory got freed in parallel, end the
// stall.
if (!ShouldStall()) {
EndWriteStall();
}
}
// Called when memory is freed in FreeMem.
void WriteBufferManager::EndWriteStall() {
if (enabled() && !IsStallThresholdExceeded()) {
{
std::unique_lock<std::mutex> lock(mu_);
stall_active_.store(false, std::memory_order_relaxed);
if (queue_.empty()) {
return;
}
}
// Get the instances from the list and call WBMStallInterface::Signal to
// change the state to running and unblock the DB instances.
// Check ShouldStall() incase stall got active by other DBs.
while (!ShouldStall() && !queue_.empty()) {
std::unique_lock<std::mutex> lock(mu_);
StallInterface* wbm_stall = queue_.front();
queue_.pop_front();
wbm_stall->Signal();
}
}
}
void WriteBufferManager::RemoveDBFromQueue(StallInterface* wbm_stall) {
assert(wbm_stall != nullptr);
if (enabled() && allow_stall_) {
std::unique_lock<std::mutex> lock(mu_);
queue_.remove(wbm_stall);
wbm_stall->Signal();
}
}
} // namespace ROCKSDB_NAMESPACE