Difference between revisions of "cpp/atomic/memory order"

Latest revision as of 13:26, 4 September 2024

std::memory_order specifies how memory accesses, including regular, non-atomic memory accesses, are to be ordered around an atomic operation. Absent any constraints on a multi-core system, when multiple threads simultaneously read and write to several variables, one thread can observe the values change in an order different from the order another thread wrote them. Indeed, the apparent order of changes can even differ among multiple reader threads. Some similar effects can occur even on uniprocessor systems due to compiler transformations allowed by the memory model.

The default behavior of all atomic operations in the library provides for sequentially consistent ordering (see discussion below). That default can hurt performance, but the library's atomic operations can be given an additional std::memory_order argument to specify the exact constraints, beyond atomicity, that the compiler and processor must enforce for that operation.

[edit] Constants

[edit] Formal description

Inter-thread synchronization and memory ordering determine how evaluations and side effects of expressions are ordered between different threads of execution. They are defined in the following terms:

[edit] Sequenced-before

Within the same thread, evaluation A may be sequenced-before evaluation B, as described in evaluation order.

[edit] Carries dependency

Within the same thread, evaluation A that is sequenced-before evaluation B may also carry a dependency into B (that is, B depends on A), if any of the following is true:

[edit] Modification order

All modifications to any particular atomic variable occur in a total order that is specific to this one atomic variable.

The following four requirements are guaranteed for all atomic operations:

[edit] Release sequence

After a release operation A is performed on an atomic object M, the longest continuous subsequence of the modification order of M that consists of:

Is known as release sequence headed by A.

[edit] Synchronizes with

If an atomic store in thread A is a release operation, an atomic load in thread B from the same variable is an acquire operation, and the load in thread B reads a value written by the store in thread A, then the store in thread A synchronizes-with the load in thread B.

Also, some library calls may be defined to synchronize-with other library calls on other threads.

[edit] Dependency-ordered before

Between threads, evaluation A is dependency-ordered before evaluation B if any of the following is true:

[edit] Inter-thread happens-before

Between threads, evaluation A inter-thread happens before evaluation B if any of the following is true:

[edit] Happens-before

Regardless of threads, evaluation A happens-before evaluation B if any of the following is true:

The implementation is required to ensure that the happens-before relation is acyclic, by introducing additional synchronization if necessary (it can only be necessary if a consume operation is involved, see Batty et al).

If one evaluation modifies a memory location, and the other reads or modifies the same memory location, and if at least one of the evaluations is not an atomic operation, the behavior of the program is undefined (the program has a data race) unless there exists a happens-before relationship between these two evaluations.

[edit] Strongly happens-before

Regardless of threads, evaluation A strongly happens-before evaluation B if any of the following is true:

[edit] Visible side-effects

The side-effect A on a scalar M (a write) is visible with respect to value computation B on M (a read) if both of the following are true:

If side-effect A is visible with respect to the value computation B, then the longest contiguous subset of the side-effects to M, in modification order, where B does not happen-before it is known as the visible sequence of side-effects (the value of M, determined by B, will be the value stored by one of these side effects).

Note: inter-thread synchronization boils down to preventing data races (by establishing happens-before relationships) and defining which side effects become visible under what conditions.

[edit] Consume operation

Atomic load with memory_order_consume or stronger is a consume operation. Note that std::atomic_thread_fence imposes stronger synchronization requirements than a consume operation.

[edit] Acquire operation

Atomic load with memory_order_acquire or stronger is an acquire operation. The lock() operation on a Mutex is also an acquire operation. Note that std::atomic_thread_fence imposes stronger synchronization requirements than an acquire operation.

[edit] Release operation

Atomic store with memory_order_release or stronger is a release operation. The unlock() operation on a Mutex is also a release operation. Note that std::atomic_thread_fence imposes stronger synchronization requirements than a release operation.

[edit] Explanation

[edit] Relaxed ordering

Atomic operations tagged memory_order_relaxed are not synchronization operations; they do not impose an order among concurrent memory accesses. They only guarantee atomicity and modification order consistency.

For example, with x and y initially zero,

// Thread 1:
r1 = y.load(std::memory_order_relaxed); // A
x.store(r1, std::memory_order_relaxed); // B
// Thread 2:
r2 = x.load(std::memory_order_relaxed); // C 
y.store(42, std::memory_order_relaxed); // D

is allowed to produce r1 == r2 == 42 because, although A is sequenced-before B within thread 1 and C is sequenced before D within thread 2, nothing prevents D from appearing before A in the modification order of y, and B from appearing before C in the modification order of x. The side-effect of D on y could be visible to the load A in thread 1 while the side effect of B on x could be visible to the load C in thread 2. In particular, this may occur if D is completed before C in thread 2, either due to compiler reordering or at runtime.

// Thread 1:
r1 = y.load(std::memory_order_relaxed);
if (r1 == 42)
    x.store(r1, std::memory_order_relaxed);
// Thread 2:
r2 = x.load(std::memory_order_relaxed);
if (r2 == 42)
    y.store(42, std::memory_order_relaxed);

Typical use for relaxed memory ordering is incrementing counters, such as the reference counters of std::shared_ptr, since this only requires atomicity, but not ordering or synchronization (note that decrementing the std::shared_ptr counters requires acquire-release synchronization with the destructor).

#include <atomic>
#include <iostream>
#include <thread>
#include <vector>
 
std::atomic<int> cnt = {0};
 
void f()
{
    for (int n = 0; n < 1000; ++n)
        cnt.fetch_add(1, std::memory_order_relaxed);
}
 
int main()
{
    std::vector<std::thread> v;
    for (int n = 0; n < 10; ++n)
        v.emplace_back(f);
    for (auto& t : v)
        t.join();
    std::cout << "Final counter value is " << cnt << '\n';
}

Final counter value is 10000

[edit] Release-Acquire ordering

If an atomic store in thread A is tagged memory_order_release, an atomic load in thread B from the same variable is tagged memory_order_acquire, and the load in thread B reads a value written by the store in thread A, then the store in thread A synchronizes-with the load in thread B.

All memory writes (including non-atomic and relaxed atomic) that happened-before the atomic store from the point of view of thread A, become visible side-effects in thread B. That is, once the atomic load is completed, thread B is guaranteed to see everything thread A wrote to memory. This promise only holds if B actually returns the value that A stored, or a value from later in the release sequence.

The synchronization is established only between the threads releasing and acquiring the same atomic variable. Other threads can see different order of memory accesses than either or both of the synchronized threads.

On strongly-ordered systems — x86, SPARC TSO, IBM mainframe, etc. — release-acquire ordering is automatic for the majority of operations. No additional CPU instructions are issued for this synchronization mode; only certain compiler optimizations are affected (e.g., the compiler is prohibited from moving non-atomic stores past the atomic store-release or performing non-atomic loads earlier than the atomic load-acquire). On weakly-ordered systems (ARM, Itanium, PowerPC), special CPU load or memory fence instructions are used.

Mutual exclusion locks, such as std::mutex or atomic spinlock, are an example of release-acquire synchronization: when the lock is released by thread A and acquired by thread B, everything that took place in the critical section (before the release) in the context of thread A has to be visible to thread B (after the acquire) which is executing the same critical section.

#include <atomic>
#include <cassert>
#include <string>
#include <thread>
 
std::atomic<std::string*> ptr;
int data;
 
void producer()
{
    std::string* p = new std::string("Hello");
    data = 42;
    ptr.store(p, std::memory_order_release);
}
 
void consumer()
{
    std::string* p2;
    while (!(p2 = ptr.load(std::memory_order_acquire)))
        ;
    assert(*p2 == "Hello"); // never fires
    assert(data == 42); // never fires
}
 
int main()
{
    std::thread t1(producer);
    std::thread t2(consumer);
    t1.join(); t2.join();
}

#include <atomic>
#include <cassert>
#include <thread>
#include <vector>
 
std::vector<int> data;
std::atomic<int> flag = {0};
 
void thread_1()
{
    data.push_back(42);
    flag.store(1, std::memory_order_release);
}
 
void thread_2()
{
    int expected = 1;
    // memory_order_relaxed is okay because this is an RMW,
    // and RMWs (with any ordering) following a release form a release sequence
    while (!flag.compare_exchange_strong(expected, 2, std::memory_order_relaxed))
    {
        expected = 1;
    }
}
 
void thread_3()
{
    while (flag.load(std::memory_order_acquire) < 2)
        ;
    // if we read the value 2 from the atomic flag, we see 42 in the vector
    assert(data.at(0) == 42); // will never fire
}
 
int main()
{
    std::thread a(thread_1);
    std::thread b(thread_2);
    std::thread c(thread_3);
    a.join(); b.join(); c.join();
}

[edit] Release-Consume ordering

If an atomic store in thread A is tagged memory_order_release, an atomic load in thread B from the same variable is tagged memory_order_consume, and the load in thread B reads a value written by the store in thread A, then the store in thread A is dependency-ordered before the load in thread B.

All memory writes (non-atomic and relaxed atomic) that happened-before the atomic store from the point of view of thread A, become visible side-effects within those operations in thread B into which the load operation carries dependency, that is, once the atomic load is completed, those operators and functions in thread B that use the value obtained from the load are guaranteed to see what thread A wrote to memory.

The synchronization is established only between the threads releasing and consuming the same atomic variable. Other threads can see different order of memory accesses than either or both of the synchronized threads.

On all mainstream CPUs other than DEC Alpha, dependency ordering is automatic, no additional CPU instructions are issued for this synchronization mode, only certain compiler optimizations are affected (e.g. the compiler is prohibited from performing speculative loads on the objects that are involved in the dependency chain).

Typical use cases for this ordering involve read access to rarely written concurrent data structures (routing tables, configuration, security policies, firewall rules, etc) and publisher-subscriber situations with pointer-mediated publication, that is, when the producer publishes a pointer through which the consumer can access information: there is no need to make everything else the producer wrote to memory visible to the consumer (which may be an expensive operation on weakly-ordered architectures). An example of such scenario is rcu_dereference.

See also std::kill_dependency and [[carries_dependency]] for fine-grained dependency chain control.

Note that currently (2/2015) no known production compilers track dependency chains: consume operations are lifted to acquire operations.

#include <atomic>
#include <cassert>
#include <string>
#include <thread>
 
std::atomic<std::string*> ptr;
int data;
 
void producer()
{
    std::string* p = new std::string("Hello");
    data = 42;
    ptr.store(p, std::memory_order_release);
}
 
void consumer()
{
    std::string* p2;
    while (!(p2 = ptr.load(std::memory_order_consume)))
        ;
    assert(*p2 == "Hello"); // never fires: *p2 carries dependency from ptr
    assert(data == 42); // may or may not fire: data does not carry dependency from ptr
}
 
int main()
{
    std::thread t1(producer);
    std::thread t2(consumer);
    t1.join(); t2.join();
}

[edit] Sequentially-consistent ordering

Atomic operations tagged memory_order_seq_cst not only order memory the same way as release/acquire ordering (everything that happened-before a store in one thread becomes a visible side effect in the thread that did a load), but also establish a single total modification order of all atomic operations that are so tagged.

// Thread 1:
x.store(1, std::memory_order_seq_cst); // A
y.store(1, std::memory_order_release); // B
// Thread 2:
r1 = y.fetch_add(1, std::memory_order_seq_cst); // C
r2 = y.load(std::memory_order_relaxed); // D
// Thread 3:
y.store(3, std::memory_order_seq_cst); // E
r3 = x.load(std::memory_order_seq_cst); // F

Sequential ordering may be necessary for multiple producer-multiple consumer situations where all consumers must observe the actions of all producers occurring in the same order.

Total sequential ordering requires a full memory fence CPU instruction on all multi-core systems. This may become a performance bottleneck since it forces the affected memory accesses to propagate to every core.

#include <atomic>
#include <cassert>
#include <thread>
 
std::atomic<bool> x = {false};
std::atomic<bool> y = {false};
std::atomic<int> z = {0};
 
void write_x()
{
    x.store(true, std::memory_order_seq_cst);
}
 
void write_y()
{
    y.store(true, std::memory_order_seq_cst);
}
 
void read_x_then_y()
{
    while (!x.load(std::memory_order_seq_cst))
        ;
    if (y.load(std::memory_order_seq_cst))
        ++z;
}
 
void read_y_then_x()
{
    while (!y.load(std::memory_order_seq_cst))
        ;
    if (x.load(std::memory_order_seq_cst))
        ++z;
}
 
int main()
{
    std::thread a(write_x);
    std::thread b(write_y);
    std::thread c(read_x_then_y);
    std::thread d(read_y_then_x);
    a.join(); b.join(); c.join(); d.join();
    assert(z.load() != 0); // will never happen
}

[edit] Relationship with volatile

Within a thread of execution, accesses (reads and writes) through volatile glvalues cannot be reordered past observable side-effects (including other volatile accesses) that are sequenced-before or sequenced-after within the same thread, but this order is not guaranteed to be observed by another thread, since volatile access does not establish inter-thread synchronization.

In addition, volatile accesses are not atomic (concurrent read and write is a data race) and do not order memory (non-volatile memory accesses may be freely reordered around the volatile access).

One notable exception is Visual Studio, where, with default settings, every volatile write has release semantics and every volatile read has acquire semantics (Microsoft Docs), and thus volatiles may be used for inter-thread synchronization. Standard volatile semantics are not applicable to multi-threaded programming, although they are sufficient for e.g. communication with a std::signal handler that runs in the same thread when applied to sig_atomic_t variables.

[edit] See also

[edit] External links