mujoco / data /src /engine /engine_global_table.h
introvoyz041's picture
Migrated from GitHub
2c55b92 verified
// Copyright 2024 DeepMind Technologies Limited
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#ifndef MUJOCO_SRC_ENGINE_ENGINE_GLOBAL_TABLE_H_
#define MUJOCO_SRC_ENGINE_ENGINE_GLOBAL_TABLE_H_
#include <atomic>
#include <cctype>
#include <cstdio>
#include <mutex>
#include <new>
#include <string>
#include <string_view>
#include <type_traits>
#include "engine/engine_util_errmem.h"
namespace mujoco {
static constexpr int kCacheLineBytes = 256;
static inline bool CaseInsensitiveEqual(std::string_view s1, std::string_view s2) {
if (s1.length() != s2.length()) {
return false;
}
auto len = s1.length();
for (decltype(len) i = 0; i < len; ++i) {
if (std::tolower(s1[i]) != std::tolower(s2[i])) {
return false;
}
}
return true;
}
// A table intended for use as global storage for extension objects such as plugins, implemented as
// a linked list of array "blocks". This is a compromise that maintains a good degree of memory
// locality while not invalidating existing pointers when growing the table. It is expected that for
// most users, the number of objects loaded will be small enough to fit in the initial block, and
// so the global table will behave like an array. Since pointers are never invalidated, we do not
// need to apply a read lock on the global table when resolving an element.
template<typename T>
struct alignas(kCacheLineBytes) TableBlock {
static constexpr int kBlockSize = 15;
TableBlock() : objects{}, next(nullptr) {
static_assert(
sizeof(TableBlock<T>) / kCacheLineBytes ==
sizeof(TableBlock<T>::objects) / kCacheLineBytes
+ (sizeof(TableBlock<T>::objects) % kCacheLineBytes > 0),
"TableBlock::next doesn't fit in the same cache line as the end of TableBlock::objects");
}
T objects[kBlockSize];
TableBlock<T>* next;
};
using Mutex = std::mutex;
class ReentrantWriteLock {
public:
ReentrantWriteLock(Mutex& mutex) : mutex_(mutex) {
if (LockCountOnCurrentThread() == 0) {
mutex_.lock();
}
++LockCountOnCurrentThread();
}
~ReentrantWriteLock() {
if (--LockCountOnCurrentThread() == 0) {
mutex_.unlock();
}
}
private:
Mutex& mutex_;
static int& LockCountOnCurrentThread() noexcept {
thread_local int counter = 0;
return counter;
}
};
template<typename T>
class GlobalTable {
public:
static GlobalTable<T>& GetSingleton() {
static_assert(std::is_trivially_destructible_v<GlobalTable<T>>);
static GlobalTable<T> global;
return global;
}
// Each extension object type T must implement these.
using ErrorMessage = char[512];
static const char* HumanReadableTypeName();
static std::string_view ObjectKey(const T&);
static bool ObjectEqual(const T&, const T&);
static bool CopyObject(T& dst, const T& src, ErrorMessage& err);
int count() {
return count_.load(std::memory_order_acquire);
}
ReentrantWriteLock LockExclusively() {
return ReentrantWriteLock(mutex());
}
int AppendIfUnique(const T& obj) {
ErrorMessage err = "\0";
// ========= ATTENTION! ========================================================================
// Do not handle objects with nontrivial destructors outside of this lambda.
// Do not call mju_error inside this lambda.
int slot = [&]() {
auto lock = LockExclusively();
int count = count_.load(std::memory_order_acquire);
int local_idx = 0;
TableBlock<T>* block = &first_block_;
// check if a non-identical object has already been registered
for (int i = 0; i < count; ++i, ++local_idx) {
if (local_idx == TableBlock<T>::kBlockSize) {
local_idx = 0;
block = block->next;
}
const T& existing = block->objects[local_idx];
if (CaseInsensitiveEqual(ObjectKey(obj), ObjectKey(existing))) {
if (!ObjectEqual(obj, existing)) {
std::snprintf(err, sizeof(err), "%s '%s' is already registered",
HumanReadableTypeName(), std::string(ObjectKey(obj)).c_str());
return -1;
} else {
return i;
}
}
}
// allocate a new block if the last allocated block is full
if (local_idx == TableBlock<T>::kBlockSize) {
local_idx = 0;
block->next = new(std::nothrow) TableBlock<T>;
if (!block->next) {
std::snprintf(err, sizeof(err), "failed to allocate memory for a new %s table block",
HumanReadableTypeName());
return -1;
}
block = block->next;
}
// copy the new object into the table
if (!CopyObject(block->objects[local_idx], obj, err)) {
return -1;
}
// increment the global count with a release memory barrier
count_.store(count + 1, std::memory_order_release);
return count;
}();
// ========= ATTENTION! ========================================================================
// End of safe lambda, do not handle objects with non-trivial destructors beyond this point.
// registration failed, throw an mju_error
if (slot < 0) {
err[sizeof(err) - 1] = '\0';
mju_error("%s", err);
}
return slot;
}
// look up by slot number, assuming that count() has already been called
const T* GetAtSlotUnsafe(int slot, int nslot) {
if (slot < 0 || slot >= nslot) {
return nullptr;
}
TableBlock<T>* block = &first_block_;
// iterate over blocks in the global table until the local index is less than the block size
int local_idx = slot;
while (local_idx >= TableBlock<T>::kBlockSize) {
local_idx -= TableBlock<T>::kBlockSize;
block = block->next;
if (!block) {
return nullptr;
}
}
// local_idx is now a valid index into the current block
T* obj = &(block->objects[local_idx]);
// check if obj has been initialized
if (obj && ObjectKey(*obj).empty()) {
return nullptr;
}
return obj;
}
// look up by key, assuming that count() has already been called
const T* GetByKeyUnsafe(std::string_view key, int* slot, int nslot) {
if (slot) *slot = -1;
if (key.empty()) {
return nullptr;
}
TableBlock<T>* block = &first_block_;
int found_slot = 0;
while (block) {
for (int i = 0;
i < TableBlock<T>::kBlockSize && found_slot < nslot;
++i, ++found_slot) {
const T& obj = block->objects[i];
// reached an uninitialized object, which means that iterated beyond the object count
// this should never happen if `count` was actually returned by count()
std::string_view candidate_key = ObjectKey(obj);
if (candidate_key.empty()) {
return nullptr;
}
// check if key matches the query
if (CaseInsensitiveEqual(candidate_key, key)) {
if (slot) *slot = found_slot;
return &obj;
}
}
block = block->next;
}
return nullptr;
}
const T* GetAtSlot(int slot) {
// count() uses memory_order_acquire which acts as a barrier that guarantees that all
// objects up to `count` have been completely inserted
return GetAtSlotUnsafe(slot, count());
}
const T* GetByKey(std::string_view key, int* slot) {
// count() uses memory_order_acquire which acts as a barrier that guarantees that all
// objects up to `count` have been completely inserted
return GetByKeyUnsafe(key, slot, count());
}
private:
GlobalTable() {
new(mutex_) Mutex;
}
Mutex& mutex() {
return *std::launder(reinterpret_cast<Mutex*>(&mutex_));
}
TableBlock<T> first_block_;
std::atomic_int count_;
// A mutex whose destructor is never run.
// When a C++ program terminates, the destructors for function static objects and globals will be
// executed by whichever thread started that termination but there is no guarantee that other
// threads have terminated. In other words, a static object may be accessed by another thread
// after it is deleted. We avoid destruction issues by never running the destructor.
alignas(Mutex) unsigned char mutex_[sizeof(Mutex)];
};
} // namespace mujoco
#endif // MUJOCO_SRC_ENGINE_ENGINE_GLOBAL_TABLE_H_