Skip to content

Fast, shared, upgradeable, non-recursive and non-fair mutex

License

Notifications You must be signed in to change notification settings

ssteinberg/shared_futex

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

83 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

shared_futex


Build status

Fast, shared, upgradeable, non-recursive and non-fair mutex with support for x86 hardware lock elision via Transactional Synchronization Extensions (TSX). Written in modern cross-platform c++17.
In active development.

Usage

shared_futex encourages a safe RAII interface.

// Lock futex in exclusive access mode
auto lock_guard = make_exclusive_lock(f);
// ...
// f will be unlocked as lock_guard goes out of scope

Similiarly shared_futex can be locked for shared access. While exclusive locks are mutually exclusive with all other holders, mutiple shared holders are allowed to access the guarded critical section simultaneously.

// Lock futex in shared access mode
auto lock_guard = make_shared_lock(f);
// ...

shared_futex can also be locked in upgradeable mode, which allows later to upgrade the lock to an exclusive lock. Upgradeable holders are mutually exclusive with other upgradeable and exclusive locks but not with shared holders.
This invariant allows concurrency between shared and upgradeable lockers and, more importantly, ensures that the state of a guarded resource does not change between the upgradeable acquisition and the upgrade to an exclusive lock.

// Lock futex in shared-upgradeable access mode
auto upgradeable_guard = make_upgradeable_lock(f);
// Find resource x to update...
// (only shared holders are permitted now)
auto exclusive_guard = upgrade_lock(std::move(upgradeable_guard));
// Update x

In addition, Lock acquisitions attempts can be made with a given time-out.

// Try lock until
const auto now = std::chrono::high_resolution_clock::now();
auto l = make_shared_lock(f, now + std::chrono::nanoseconds(100));
if (l) { /* lock acquired */ }
// Try lock for
auto l = make_shared_lock(f, std::chrono::nanoseconds(100));

Or in a non-blocking manner via std::try_to_lock

// Try lock
auto l = make_exclusive_lock(f, std::try_to_lock);
if (l) { /* lock acquired */ }

Lock guards can also be manipulated manually.

auto l = make_exclusive_lock(f, std::defer_lock); // Does not lock
if (l.try_lock()) {
	l.unlock();
}
if (l.try_lock_for(std::chrono::milliseconds(1))) {
	l.unlock();
}
const auto now = std::chrono::steady_clock::now();
if (l.try_lock_until(now + std::chrono::milliseconds(1))) {
	l.unlock();
}
l.lock();
l.unlock();

Lock guards meet the requirements of BasicLockable, Lockable and TimedLockable therefore can be used in conjunction with standard library utilities.
E.g. avoiding deadlocks when locking multiple mutexes can be done via the standard library's std::lock.

shared_futex f0;
shared_futex f1;
std::mutex m;

// Defer lock
auto l0 = make_exclusive_lock(f0, std::defer_lock);
auto l1 = make_shared_lock(f1, std::defer_lock);
std::unique_lock<std::mutex> l2(m, std::defer_lock);
// Lock multiple locks with deadlock avoidence
std::lock(l0, l1, l2);

Likewise lock guards are default constructible and also meet the requirements of MoveConstructible and MoveAssignable but not of CopyConstructible or CopyAssignable. Therefore lock guards can be safely moved around, swapped and empty assigned via empty brace initialization.

// Swap
auto l0 = make_exclusive_lock(f);
auto l1 = make_exclusive_lock(f, std::defer_lock);

std::swap(l0, l1);
assert(!l0.owns_lock() && l1.owns_lock());

l0.swap(l1);
assert(l0.owns_lock() && !l1.owns_lock());

// Move
l1 = std::move(l0);
assert(!l0.owns_lock() && l1.owns_lock());
		
l1 = {};	// Will unlock
assert(!l0.owns_lock() && !l1.owns_lock());

Locks can also be adopted.

// Drop lock
auto l0 = make_exclusive_lock(f);
auto&& lock_object = std::move(l0).drop();

// And adopt lock
auto l1 = make_exclusive_lock(f, std::move(lock_object));
assert(!l0.owns_lock() && l1.owns_lock());

// Dropping and not adopting the lock will trigger an assert

shared_futex default constructor is constexpr therefore a shared_futex is statically initialized making it safe to lock a mutex in the constructor of any static or global object.
See c++ initialization order of non-local variables.

Backoff policies

When an intial lock attempt fails the locking protocol keeps retrying until either a timeout has been reached or lock was successfully acquired. Between iterations a backoff policy is executed. The backoff policy can spin, yield or park the thread.

Three backoff policies are provided:

  1. exponential_backoff_policy - This is the default policy and generally provides a good balance between resource usages and latency, leading to best overall performance. The policy initally spins with linearly increasing spin count, and after a variable amount of spin iterations yields and then parks the thread.
    Each spin iteration has a randomness injected into the calculated spin count to break cross-thread symmetry.
  2. relaxed_backoff_policy - Avoids spinning. Yields for a few iterations the then parks the thread. Useful when the caller knows that there is contention, otherwise will perform poorly due to unnecessary contex switches.
  3. spinlock_backoff_policy - Never parks and employs aggressive spin tactic. Randomness is injected as well to break cross-thread symmetries.

Custom backoff-policies are simple to create. For examples see shared_futex_policies.hpp.

A policy can be selected via a template parameter

// Spin-lock
auto l = make_shared_lock<relaxed_backoff_policy>(f);

It is allowed to use distinct backoff policies with one shared_futex, e.g. a system might want to employ spin-locking to minimize latency on one thread (for a variety of reasons) while on another thread use the default backoff policy.

Specialized variants

shared_futex comes in a few variations which are suitable for different tasks.

  • shared_futex_micro - Small 32-bit variant with very low overhead. Performs very well at the typical low contention environment with small critical sections typically encountered at general client-side applications.
  • shared_futex - Large 128-byte variant. Aligned and does not share cachelines with local data. Provides additional machinery and bookkeeping to improve performance under contention at the expense of additional overhead.
  • shared_futex_pico - Tiny 8-bit variant. Slightly slower than the micro variant. Limited to 2^6 concurrent shared holders, overflows result in undefined behaviour. Useful when embedding locks into slots, nodes, etc..
  • shared_futex_concurrent - Large 1-kbyte variant that employs multiple slots on distinct cache lines to allow true concurrency under heavy shared contention. Can vastly outperform other locks under very high shared-to-exclusive workloads.
    Extra slots are only employed when concurrency is detected.
  • shared_futex_hle/shared_futex_micro_hle - Similar to shared_futex and shared_futex_micro, respectively, except also employs hardware lock elision (TSX HLE) on supported x86-64 systems. Can greatly increase performance in specific circumstances.
  • shared_futex_tsx_rtm - Experimental variant that performs hardware lock elision explcitly via transactional memory.

Condition variables

condition_variable is designed to replace std::condition_variable when working with a shared_futex.

condition_variable cv;
// Wait upon cv
auto lg = make_lock_when<shared_futex_lock_class::exclusive>(f, cv);

This will park until cv is signalled at which point the lock will be acquired in exclusive mode and a lock_guard returned.

A predicate can also be supplied:

auto lg = make_lock_when<shared_futex_lock_class::shared>(f, cv,
	[]() noexcept -> bool { ... });

This form checks the predicate, and if the predicate is unsatisfied, parks the calling thread in one atomic operation. Upon signalling the lock will acquired and the predicate checked again.

The above forms can also be used with a time-out:

auto lg = make_lock_when<shared_futex_lock_class::shared>(f, cv,
	std::chrono::seconds(10));
const auto now = std::chrono::steady_clock::now();
auto lg = make_lock_when<shared_futex_lock_class::shared>(f, cv,
	now + std::chrono::seconds(10));
auto lg = make_lock_when<shared_futex_lock_class::shared>(f, cv,
	[]() noexcept -> bool { ... },
	std::chrono::millisecond(500));
const auto now = std::chrono::steady_clock::now();
auto lg = make_lock_when<shared_futex_lock_class::shared>(f, cv,
	[]() noexcept -> bool { ... },
	now + std::chrono::millisecond(500));

The condition_variable can also be waited upon explicitly inside a critical section:

auto lg = make_exclusive_lock(f);
const auto predicate = []() noexcept -> bool { ... };
while (!pred())
	cv.wait(lg);

Which is functionally equivalent to cv.wait(lg, predicate). In addition a timeout can be provided via wait_for() and wait_until(). The cv.wait*() overloads return a parking_lot_wait_state indicating wait result.

Signalling is performed explicitly by calling a condition_variable's signal() or signal_n(), where the later variant sets an upper limit on the total signalled thread count. It is important to remember that unlike std::condition_variable, both forms will always unpark up to a single exclusive waiter or up to a single upgradeable waiter and multiple shared waiters.

auto lg = make_exclusive_lock(f);
const auto unparked_count = cv.signal(std::move(lg));
auto lg = make_exclusive_lock(f);
const auto unparked_count = cv.signal_n(3, std::move(lg));
assert(unparked_count <= 3);

Calling signal*() consumes the lock and releases it.

Unparks and contex-switched are expensive, therefore if a predicate is supplied to wait*() the condition_vartiable will check the unpark candidates' predicates before unparking. This is done while holding the lock supplied to signal*. It is important to remember that the lock classes supplied to wait*() and signal*() operations can be non-mutually exclusive which might result in a concurrent access to a predicate. It is the user responsibility to ensure the predicate's safety in such circumstances.

Type traits

A few type traits are provided in shared_futex_type_traits.hpp.

  • Checks if T is a shared_futex variant:
template <typename T> struct is_shared_futex_type;
template <typename T> inline constexpr bool is_shared_futex_type_v;
  • Checks if T is a shared_futex variant that supports parking:
template <typename T> struct is_shared_futex_parkable;
template <typename T> inline constexpr bool is_shared_futex_parkable_v;
  • Provides the lock class (see shared_futex_lock_class) of a lock_guard:
template <typename T> struct lock_class;
template <typename T> inline constexpr shared_futex_lock_class lock_class_v;

Hardware lock elision

Hardware lock elision is performed via the Transactional Synchronization Extensions (TSX) on modern x86-64 architectures. (For more information on TSX, see Intel� 64 and IA-32 Architectures Software Developer�s Manual, Volume 1.)

The shared_futex*_hle variants use the xacquire/xrelease x86 prefix hints for atomic operations to elide locks, when possible. This is backward compatible: On unsupported hardware the prefixes are harmlessly translated to REPNE/REPNZ and REP/REPE/REPZ (repsectively) which have no side effects on the used instructions.
Lock elision works by eliding the lock acquisition and optimistically executing the critical section as a memory transaction. The transaction can be aborted due to cache line conflict (e.g. another consumer touched the same cache line), TSX capacity and other reasons. In case of abort the execution is rolled back to the lock acquisition instruction and retried without lock elision.
Excessive aborts degrade performance therefore shared_futex variants that employ HLE are mostly suited for specific tasks, e.g. sparsely updating a vector.

*_rtm variants are experimental. Transactions are triggered explicitly via xbeing/xend instructions, which are not backward compatible. It is the users responsibility to ensure their target system supports TSX RTM. Transactions can be explictly aborted inside a critical section falling back to non-elided lock acquisition. See transactional_memory namespace.