pybind11/docs/advanced/deadlock.md
Thomas Köppe af67e87393
docs/advanced A document about deadlock potential with C++ statics (#5394)
* [docs/advanced] A document about deadlock potential with C++ statics

* [docs/advanced] Refer to deadlock.md from misc.rst

* [docs/advanced] Fix tables in deadlock.md

* Use :ref:`deadlock-reference-label`

* Revert "Use :ref:`deadlock-reference-label`"

This reverts commit e5734d275f.

* Add simple references to docs/advanced/deadlock.md filename. (Maybe someone can work on clickable links later.)

---------

Co-authored-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>
2024-10-08 10:51:20 -07:00

17 KiB
Raw Blame History

Double locking, deadlocking, GIL

[TOC]

Introduction

Overview

In concurrent programming with locks, deadlocks can arise when more than one mutex is locked at the same time, and careful attention has to be paid to lock ordering to avoid this. Here we will look at a common situation that occurs in native extensions for CPython written in C++.

Deadlocks

A deadlock can occur when more than one thread attempts to lock more than one mutex, and two of the threads lock two of the mutexes in different orders. For example, consider mutexes mu1 and mu2, and threads T1 and T2, executing:

T1 T2
1 mu1.lock(){.good} mu2.lock(){.good}
2 mu2.lock(){.bad} mu1.lock(){.bad}
3 /* work */ /* work */
4 mu2.unlock() mu1.unlock()
5 mu1.unlock() mu2.unlock()

Now if T1 manages to lock mu1 and T2 manages to lock mu2 (as indicated in green), then both threads will block while trying to lock the respective other mutex (as indicated in red), but they are also unable to release the mutex that they have locked (step 5).

The problem is that it is possible for one thread to attempt to lock mu1 and then mu2, and for another thread to attempt to lock mu2 and then mu1. Note that it does not matter if either mutex is unlocked at any intermediate point; what matters is only the order of any attempt to lock the mutexes. For example, the following, more complex series of operations is just as prone to deadlock:

T1 T2
1 mu1.lock(){.good} mu1.lock(){.good}
2 waiting for T2 mu2.lock(){.good}
3 waiting for T2 /* work */
3 waiting for T2 mu1.unlock()
3 mu2.lock(){.bad} /* work */
3 /* work */ mu1.lock(){.bad}
3 /* work */ /* work */
4 mu2.unlock() mu1.unlock()
5 mu1.unlock() mu2.unlock()

When the mutexes involved in a locking sequence are known at compile-time, then avoiding deadlocks is “merely” a matter of arranging the lock operations carefully so as to only occur in one single, fixed order. However, it is also possible for mutexes to only be determined at runtime. A typical example of this is a database where each row has its own mutex. An operation that modifies two rows in a single transaction (e.g. “transferring an amount from one account to another”) must lock two row mutexes, but the locking order cannot be established at compile time. In this case, a dynamic “deadlock avoidance algorithm” is needed. (In C++, std::lock provides such an algorithm. An algorithm might use a non-blocking try_lock operation on a mutex, which can either succeed or fail to lock the mutex, but returns without blocking.)

Conceptually, one could also consider it a deadlock if the same thread attempts to lock a mutex that it has already locked (e.g. when some locked operation accidentally recurses into itself): mu.lock();{.good} mu.lock();{.bad} However, this is a slightly separate issue: Typical mutexes are either of recursive or non-recursive kind. A recursive mutex allows repeated locking and requires balanced unlocking. A non-recursive mutex can be implemented more efficiently, and/but for efficiency reasons does not actually guarantee a deadlock on second lock. Instead, the API simply forbids such use, making it a precondition that the thread not already hold the mutex, with undefined behaviour on violation.

“Once” initialization

A common programming problem is to have an operation happen precisely once, even if requested concurrently. While it is clear that we need to track in some shared state somewhere whether the operation has already happened, it is worth noting that this state only ever transitions, once, from false to true. This is considerably simpler than a general shared state that can change values arbitrarily. Next, we also need a mechanism for all but one thread to block until the initialization has completed, which we can provide with a mutex. The simplest solution just always locks the mutex:

// The "once" mechanism:
constinit absl::Mutex mu(absl::kConstInit);
constinit bool init_done = false;

// The operation of interest:
void f();

void InitOnceNaive() {
  absl::MutexLock lock(&mu);
  if (!init_done) {
    f();
    init_done = true;
  }
}

This works, but the efficiency-minded reader will observe that once the operation has completed, all future lock contention on the mutex is unnecessary. This leads to the (in)famous “double-locking” algorithm, which was historically hard to write correctly. The idea is to check the boolean before locking the mutex, and avoid locking if the operation has already completed. However, accessing shared state concurrently when at least one access is a write is prone to causing a data race and needs to be done according to an appropriate concurrent programming model. In C++ we use atomic variables:

// The "once" mechanism:
constinit absl::Mutex mu(absl::kConstInit);
constinit std::atomic<bool> init_done = false;

// The operation of interest:
void f();

void InitOnceWithFastPath() {
  if (!init_done.load(std::memory_order_acquire)) {
    absl::MutexLock lock(&mu);
    if (!init_done.load(std::memory_order_relaxed)) {
      f();
      init_done.store(true, std::memory_order_release);
    }
  }
}

Checking the flag now happens without holding the mutex lock, and if the operation has already completed, we return immediately. After locking the mutex, we need to check the flag again, since multiple threads can reach this point.

Atomic details. Since the atomic flag variable is accessed concurrently, we have to think about the memory order of the accesses. There are two separate cases: The first, outer check outside the mutex lock, and the second, inner check under the lock. The outer check and the flag update form an acquire/release pair: if the load sees the value true (which must have been written by the store operation), then it also sees everything that happened before the store, namely the operation f(). By contrast, the inner check can use relaxed memory ordering, since in that case the mutex operations provide the necessary ordering: if the inner load sees the value true, it happened after the lock(), which happened after the unlock(), which happened after the store.

The C++ standard library, and Abseil, provide a ready-made solution of this algorithm called std::call_once/absl::call_once. (The interface is the same, but the Abseil implementation is possibly better.)

// The "once" mechanism:
constinit absl::once_flag init_flag;

// The operation of interest:
void f();

void InitOnceWithCallOnce() {
  absl::call_once(once_flag, f);
}

Even though conceptually this is performing the same algorithm, this implementation has some considerable advantages: The once_flag type is a small and trivial, integer-like type and is trivially destructible. Not only does it take up less space than a mutex, it also generates less code since it does not have to run a destructor, which would need to be added to the program's global destructor list.

The final clou comes with the C++ semantics of a static variable declared at block scope: According to [stmt.dcl]:

Dynamic initialization of a block variable with static storage duration or thread storage duration is performed the first time control passes through its declaration; such a variable is considered initialized upon the completion of its initialization. [...] If control enters the declaration concurrently while the variable is being initialized, the concurrent execution shall wait for completion of the initialization.

This is saying that the initialization of a local, static variable precisely has the “once” semantics that we have been discussing. We can therefore write the above example as follows:

// The operation of interest:
void f();

void InitOnceWithStatic() {
  static int unused = (f(), 0);
}

This approach is by far the simplest and easiest, but the big difference is that the mutex (or mutex-like object) in this implementation is no longer visible or in the users control. This is perfectly fine if the initializer is simple, but if the initializer itself attempts to lock any other mutex (including by initializing another static variable!), then we have no control over the lock ordering!

Finally, you may have noticed the constinits around the earlier code. Both constinit and constexpr specifiers on a declaration mean that the variable is constant-initialized, which means that no initialization is performed at runtime (the initial value is already known at compile time). This in turn means that a static variable guard mutex may not be needed, and static initialization never blocks. The difference between the two is that a constexpr-specified variable is also const, and a variable cannot be constexpr if it has a non-trivial destructor. Such a destructor also means that the guard mutex is needed after all, since the destructor must be registered to run at exit, conditionally on initialization having happened.

Python, CPython, GIL

With CPython, a Python program can call into native code. To this end, the native code registers callback functions with the Python runtime via the CPython API. In order to ensure that the internal state of the Python runtime remains consistent, there is a single, shared mutex called the “global interpreter lock”, or GIL for short. Upon entry of one of the user-provided callback functions, the GIL is locked (or “held”), so that no other mutations of the Python runtime state can occur until the native callback returns.

Many native extensions do not interact with the Python runtime for at least some part of them, and so it is common for native extensions to release the GIL, do some work, and then reacquire the GIL before returning. Similarly, when code is generally not holding the GIL but needs to interact with the runtime briefly, it will first reacquire the GIL. The GIL is reentrant, and constructions to acquire and subsequently release the GIL are common, and often don't worry about whether the GIL is already held.

If the native code is written in C++ and contains local, static variables, then we are now dealing with at least two mutexes: the static variable guard mutex, and the GIL from CPython.

A common problem in such code is an operation with “only once” semantics that also ends up requiring the GIL to be held at some point. As per the above description of “once”-style techniques, one might find a static variable:

// CPython callback, assumes that the GIL is held on entry.
PyObject* InvokeWidget(PyObject* self) {
  static PyObject* impl = CreateWidget();
  return PyObject_CallOneArg(impl, self);
}

This seems reasonable, but bear in mind that there are two mutexes (the "guard mutex" and "the GIL"), and we must think about the lock order. Otherwise, if the callback is called from multiple threads, a deadlock may ensue.

Let us consider what we can see here: On entry, the GIL is already locked, and we are locking the guard mutex. This is one lock order. Inside the initializer CreateWidget, with both mutexes already locked, the function can freely access the Python runtime.

However, it is entirely possible that CreateWidget will want to release the GIL at one point and reacquire it later:

// Assumes that the GIL is held on entry.
// Ensures that the GIL is held on exit.
PyObject* CreateWidget() {
  // ...
  Py_BEGIN_ALLOW_THREADS  // releases GIL
  // expensive work, not accessing the Python runtime
  Py_END_ALLOW_THREADS    // acquires GIL, #!
  // ...
  return result;
}

Now we have a second lock order: the guard mutex is locked, and then the GIL is locked (at #!). To see how this deadlocks, consider threads T1 and T2 both having the runtime attempt to call InvokeWidget. T1 locks the GIL and proceeds, locking the guard mutex and calling CreateWidget; T2 is blocked waiting for the GIL. Then T1 releases the GIL to do “expensive work”, and T2 awakes and locks the GIL. Now T2 is blocked trying to acquire the guard mutex, but T1 is blocked reacquiring the GIL (at #!).

In other words: if we want to support “once-called” functions that can arbitrarily release and reacquire the GIL, as is very common, then the only lock order that we can ensure is: guard mutex first, GIL second.

To implement this, we must rewrite our code. Naively, we could always release the GIL before a static variable with blocking initializer:

// CPython callback, assumes that the GIL is held on entry.
PyObject* InvokeWidget(PyObject* self) {
  Py_BEGIN_ALLOW_THREADS  // releases GIL
  static PyObject* impl = CreateWidget();
  Py_END_ALLOW_THREADS    // acquires GIL

  return PyObject_CallOneArg(impl, self);
}

But similar to the InitOnceNaive example above, this code cycles the GIL (possibly descheduling the thread) even when the static variable has already been initialized. If we want to avoid this, we need to abandon the use of a static variable, since we do not control the guard mutex well enough. Instead, we use an operation whose mutex locking is under our control, such as call_once. For example:

// CPython callback, assumes that the GIL is held on entry.
PyObject* InvokeWidget(PyObject* self) {
  static constinit PyObject* impl = nullptr;
  static constinit std::atomic<bool> init_done = false;
  static constinit absl::once_flag init_flag;

  if (!init_done.load(std::memory_order_acquire)) {
    Py_BEGIN_ALLOW_THREADS                       // releases GIL
    absl::call_once(init_flag, [&]() {
      PyGILState_STATE s = PyGILState_Ensure();  // acquires GIL
      impl = CreateWidget();
      PyGILState_Release(s);                     // releases GIL
      init_done.store(true, std::memory_order_release);
    });
    Py_END_ALLOW_THREADS                         // acquires GIL
  }

  return PyObject_CallOneArg(impl, self);
}

The lock order is now always guard mutex first, GIL second. Unfortunately we have to duplicate the “double-checked done flag”, effectively leading to triple checking, because the flag state inside the absl::once_flag is not accessible to the user. In other words, we cannot ask init_flag whether it has been used yet.

However, we can perform one last, minor optimisation: since we assume that the GIL is held on entry, and again when the initializing operation returns, the GIL actually serializes access to our done flag variable, which therefore does not need to be atomic. (The difference to the previous, atomic code may be small, depending on the architecture. For example, on x86-64, acquire/release on a bool is nearly free (demo).)

// CPython callback, assumes that the GIL is held on entry, and indeed anywhere
// directly in this function (i.e. the GIL can be released inside CreateWidget,
// but must be reaqcuired when that call returns).
PyObject* InvokeWidget(PyObject* self) {
  static constinit PyObject* impl = nullptr;
  static constinit bool init_done = false;       // guarded by GIL
  static constinit absl::once_flag init_flag;

  if (!init_done) {
    Py_BEGIN_ALLOW_THREADS                       // releases GIL
                                                 // (multiple threads may enter here)
    absl::call_once(init_flag, [&]() {
                                                 // (only one thread enters here)
      PyGILState_STATE s = PyGILState_Ensure();  // acquires GIL
      impl = CreateWidget();
      init_done = true;                          // (GIL is held)
      PyGILState_Release(s);                     // releases GIL
    });

    Py_END_ALLOW_THREADS                         // acquires GIL
  }

  return PyObject_CallOneArg(impl, self);
}

Debugging tips

  • Build with symbols.
  • Ctrl-C sends SIGINT, Ctrl-\ sends SIGQUIT. Both have their uses.
  • Useful gdb commands:
    • py-bt prints a Python backtrace if you are in a Python frame.
    • thread apply all bt 10 prints the top-10 frames for each thread. A full backtrace can be prohibitively expensive, and the top few frames are often good enough.
    • p PyGILState_Check() shows whether a thread is holding the GIL. For all threads, run thread apply all p PyGILState_Check() to find out which thread is holding the GIL.
    • The static variable guard mutex is accessed with functions like cxa_guard_acquire (though this depends on ABI details and can vary). The guard mutex itself contains information about which thread is currently holding it.