folly学习笔记
1 ThreadLocal 部分
参考的代码commit at 77e08c29d9f3b60b55dd74c1ceede35a8297e586,第一个就看ThreadLocal,是因为这几天解BUG涉及到这块,所以就先写了
1.1 ThreadLocal的结构
参考folly自己对于threadlocal的描述,可以发现实际上针对每种Tag类型,只会占用一个pthreadke。参考实现能发现,folly内部实际上是一个per thread存储一个指针,这个指针指向一个ThreadEntry::elements数组
- Similar speed as using
pthread_getspecific
directly, but only consumes a singlepthread_key_t
perTag
template param.
We keep a
__thread
array of pointers to objects (ThreadEntry::elements
) where each array has an index for each unique instance of theThreadLocalPtr
object. EachThreadLocalPtr
object has a unique id that is an index into these arrays so we can fetch the correct object from thread local storage very efficiently.In order to prevent unbounded growth of the id space and thus huge
ThreadEntry::elements
arrays, for example due to continuous creation and destruction ofThreadLocalPtr
objects, we keep track of all active instances by linking them together into a list. When an instance is destroyed we remove it from the chain and insert the id intofreeIds_
for reuse. These operations require a global mutex, but only happen at construction and destruction time.accessAllThreads
also acquires this global mutex.We use a single global
pthread_key_t
perTag
to manage object destruction and memory cleanup upon thread exit because there is a finite number ofpthread_key_t
’s available per machine.
这个描述不够清楚,我再用自己的话简单描述一下,
template <class T, class Tag = void, class AccessMode = void> ThreadLocal,是ThreadLocalPtr的简单包裹,没有太多需要担心的东西
template <class T, class Tag = void, class AccessMode = void> ThreadLocalPtr,ThreadLocalPtr如果初始化的时候没有特别的初始化Tag,那么实际上都是从StaticMeta里面拿到具体的element的,简单陈述就是从static thread_local ThreadEntry* threadEntryTL的数组找到element然后拿出来
typedef threadlocal_detail::StaticMeta<Tag, AccessMode> StaticMeta,全局单例的静态元数据,里面放了一个pthread_key,用来在新线程找来的时候找到对应的thread_entry。默认的tag是void,意味着即使你存储着一堆不同类型T的对象,只要tag一致,它们都会放在一个thread_entry指向的element数组里面。
struct folly::threadlocal_detail::ThreadEntry,每个线程存储着自己的Per-thread entry,进程里面有几个StaticMeta,那么线程就会存储几个ThreadEntry。每次拿取threadentry都是从ThreadLocalPtr里面的get进去,然后就开始perthread的操作了
用一张图来表示,就是ThreadLocalPtr统一入口,如果已经分配了Per-thread entry和元素,那就直接返回并分配,没分配的话就走第一次分配的流程,也就是下面1.2节的ThreadLocal第一次执行操作时的流程
这个描述还没看明白的话,可以对着下面的操作流程看下
1.2 ThreadLocal第一次执行操作时的流程
简单来说,当一个线程第一次调用folly::ThreadLocal的时候,实际上走的基本都是slow函数,调用关系伪代码如下
olly::ThreadLocalPtr<T, Tag, AccessMode>::get() {
olly::threadlocal_detail::StaticMeta<Tag, AccessMode>::get {
if (tls entry capacity <= id) {
StaticMeta<...>::getSlowReserveAndCache {
tls entry = StaticMeta<Tag, AccessMode>::getThreadEntrySlow {
tls entry = pthread_getspecific(meta.pthreadKey_)
if (tls entry not set)
alloc and pthread_setspecific
return tls entry
}
if (tls entry capacity <= id)
alloc and return tls entry
}
}
threadEntry->elements[id]
}
}
流程图还是不够精细,让我们直接来看源码,从最外层开始,走一遍一个线程第一次调用threadlocal类型的流程。
ThreadLocal 是ThreadLocalPtr的简单包裹,可以参考下面的代码,就是做个一个构造函数初始化一下,实际上每次get的时候都是拿取的tlp_,即自己的ThreadLocalPtr指针
template <class T, class Tag, class AccessMode>
class ThreadLocalPtr;
template <class T, class Tag = void, class AccessMode = void>
class ThreadLocal {
public:
...
FOLLY_ERASE T* get() const {
auto const ptr = tlp_.get();
return FOLLY_LIKELY(!!ptr) ? ptr : makeTlp();
}
...
T* operator->() const { return get(); }
T& operator*() const { return *get(); }
...
mutable ThreadLocalPtr<T, Tag, AccessMode> tlp_;
std::function<T*()> constructor_;
};
那么ThreadLocalPtr的结果是什么,看下面实际上就包裹了一个StaticMeta::EntryID的对象,那么怎么做到的per线程?关注get函数StaticMeta::get(&id_);
template <class T, class Tag = void, class AccessMode = void>
class ThreadLocalPtr {
private:
typedef threadlocal_detail::StaticMeta<Tag, AccessMode> StaticMeta;
...
public:
constexpr ThreadLocalPtr() : id_() {}
T* get() const {
threadlocal_detail::ElementWrapper& w = StaticMeta::get(&id_);
return static_cast<T*>(w.ptr);
}
...
mutable typename StaticMeta::EntryID id_;
};
函数StaticMeta::get可以看到,有一个简单的判断,根据是否使用kUseThreadLocal,来拿取threadEntryTL,这是一个真正的static thread_local类型对象,每个线程都有独一的一份,注意,这里开始了第一次分化。如果当前的capacity小于ID,那就需要走getSlowReserveAndCache分配出来容量。有人问为什么需要id_很简单,当你编译ThreadLocalPtr的模板参数没有指定Tag时,就用的都是一套StaticMeta<void,void>::get,也就是说你存储不同类别的(T类别)的对象,最后都会被塞到一个StaticMeta<void,void>::get,拿的都是一套threadEntryNonTL,既然如此就得给这个threadEntryNonTL所保存的element数组多分配点空间和元素。
如果你指定了独特的Tag,那么就不用多分配elements了,第一次分配了拿了就行。
我们模拟的是一个新线程走这一块,那么我们可以认为threadEntryTL第一次就是个空指针,就会再往getSlowReserveAndCache里面走,如果是第二次到来且已经给新的id分配过空间了,那么可以直接就拿初始化过的threadEntryTL对应的element返回了
FOLLY_EXPORT FOLLY_ALWAYS_INLINE static ElementWrapper& get(EntryID* ent) {
// Eliminate as many branches and as much extra code as possible in the
// cached fast path, leaving only one branch here and one indirection below.
uint32_t id = ent->getOrInvalid();
static thread_local ThreadEntry* threadEntryTL{};
ThreadEntry* threadEntryNonTL{};
auto& threadEntry = kUseThreadLocal ? threadEntryTL : threadEntryNonTL;
static thread_local size_t capacityTL{};
size_t capacityNonTL{};
auto& capacity = kUseThreadLocal ? capacityTL : capacityNonTL;
if (FOLLY_UNLIKELY(capacity <= id)) {
getSlowReserveAndCache(ent, id, threadEntry, capacity);
}
return threadEntry->elements[id];
}
顺着getSlowReserveAndCache查看调用过程,可以看到当前线程ThreadEntry thread local static指针被instance的threadEntry_函数赋值,这个地方threadEntry = inst.threadEntry_()这里拿到了当前数组的threadEntry,实际上的函数是getThreadEntrySlow。Here comes the fun part,是怎么拿到perthread的threadentry呢?
首先还是找到instance,然后拿到该static StaticMeta<Tag, AccessMode>保存的pthread_key,毕竟是第一次走到这里,当前线程调用pthread_getspecific发现,呀,我这个线程还没有绑定这个key的独有空间。ok,那么我可得申请一个。
先调用getThreadEntryList拿到StaticMeta的Theandentrylist,然后再构造一个static thread local类型entry,然后取地址挂到StaticMeta的链上。挂完了在把自己的tid和tid_os挂上,最终调用pthread_setspecific绑定该thread_entry,下这样子下次如果再进入getThreadEntrySlow就不需要再申请新的threadentry了,毕竟太慢了。
getThreadEntrySlow跑完了,实际上就可以分配element了,这时候就又回到了getSlowReserveAndCache里面,过程就比较直接了,inst里面reserve足够多的ent,然后拿一下id,返回即可,我们看看对应的reserve函数
FOLLY_NOINLINE static void getSlowReserveAndCache(
EntryID* ent, uint32_t& id, ThreadEntry*& threadEntry, size_t& capacity) {
auto& inst = instance();
threadEntry = inst.threadEntry_();
if (UNLIKELY(threadEntry->getElementsCapacity() <= id)) {
inst.reserve(ent);
id = ent->getOrInvalid();
}
capacity = threadEntry->getElementsCapacity();
assert(capacity > id);
}
FOLLY_EXPORT FOLLY_NOINLINE static ThreadEntry* getThreadEntrySlow() {
auto& meta = instance();
auto key = meta.pthreadKey_;
ThreadEntry* threadEntry =
static_cast<ThreadEntry*>(pthread_getspecific(key));
if (!threadEntry) {
ThreadEntryList* threadEntryList = StaticMeta::getThreadEntryList();
if (kUseThreadLocal) {
static thread_local ThreadEntry threadEntrySingleton;
threadEntry = &threadEntrySingleton;
} else {
threadEntry = new ThreadEntry();
}
// if the ThreadEntry already exists
// but pthread_getspecific returns NULL
// do not add the same entry twice to the list
// since this would create a loop in the list
if (!threadEntry->list) {
threadEntry->list = threadEntryList;
threadEntry->listNext = threadEntryList->head;
threadEntryList->head = threadEntry;
}
threadEntry->tid() = std::this_thread::get_id();
threadEntry->tid_os = folly::getOSThreadID();
// if we're adding a thread entry
// we need to increment the list count
// even if the entry is reused
threadEntryList->count++;
threadEntry->meta = &meta;
int ret = pthread_setspecific(key, threadEntry);
checkPosixError(ret, "pthread_setspecific failed");
}
return threadEntry;
}
reserve过程实际上是从新分配element数组的个数,可以认为是扩充以后把原本的element数组的内容再原样拷贝回去,所以得加锁。这里的reallocate过程实际上是有倍数的,其会根据情况在两个增长因子计算出来的数字做分配,简单说就是要么分配(idval + 5) * kSmallGrowthFactor,要么分配(idval + 5) * kBigGrowthFactor,+5是为了避免过慢增长。
void StaticMetaBase::reserve(EntryID* id) {
auto& meta = *this;
ThreadEntry* threadEntry = (*threadEntry_)();
size_t prevCapacity = threadEntry->getElementsCapacity();
uint32_t idval = id->getOrAllocate(meta);
if (prevCapacity > idval) {
return;
}
size_t newCapacity;
ElementWrapper* reallocated = reallocate(threadEntry, idval, newCapacity);
// Success, update the entry
{
std::lock_guard<std::mutex> g(meta.lock_);
if (reallocated) {
/*
* Note: we need to hold the meta lock when copying data out of
* the old vector, because some other thread might be
* destructing a ThreadLocal and writing to the elements vector
* of this thread.
*/
if (prevCapacity != 0) {
memcpy(
reallocated,
threadEntry->elements,
sizeof(*reallocated) * prevCapacity);
}
std::swap(reallocated, threadEntry->elements);
}
for (size_t i = prevCapacity; i < newCapacity; i++) {
threadEntry->elements[i].node.initZero(threadEntry, i);
}
threadEntry->setElementsCapacity(newCapacity);
}
free(reallocated);
}
好了,threadlocalPtr部分就这个样子了
1.3 Threadlocalptr的一些问题
一些问题,
- 如果在执行getSlowReserveAndCache,上下文切换了,当前线程被丢到另外的cpu上执行怎么办?没啥问题,它是threadlocal,不是cpulocal,所以上下文都还是传递过去了。
- 为什么还需要pthread_key放在staticmeta里面,直接从静态函数进去走static pthread local,换句话说用c17以后的_thread不行吗?我理解是两点,但是感觉这解释也不太对劲
- pthread_key这个东西,每个Meta只有一个,就能够大大节省_thread成员创建时候对内存的浪费
- 在主线程退出时,能针对性的去做清理,释放掉对应的tls
结尾
唉,尴尬