The failure of Dekker-style synchronization is typically explained with reordering of instructions. I.e., if we write
atomic_int X;
atomic_int Y;
int r1, r2;
static void t1() {
X.store(1, std::memory_order_relaxed)
r1 = Y.load(std::memory_order_relaxed);
}
static void t2() {
Y.store(1, std::memory_order_relaxed)
r2 = X.load(std::memory_order_relaxed);
}
Then the loads can be reordered with the stores, leading to r1==r2==0
.
I was expecting an acquire_release fence to prevent this kind of reordering:
static void t1() {
X.store(1, std::memory_order_relaxed);
atomic_thread_fence(std::memory_order_acq_rel);
r1 = Y.load(std::memory_order_relaxed);
}
static void t2() {
Y.store(1, std::memory_order_relaxed);
atomic_thread_fence(std::memory_order_acq_rel);
r2 = X.load(std::memory_order_relaxed);
}
The load cannot be moved above the fence and the store cannot be moved below the fence, and so the bad result should be prevented.
However, experiments show r1==r2==0
can still occur. Is there a reordering-based explanation for this? Where's the flaw in my reasoning?
And finally, ifthe relaxed atomic load reads the value written by the relaxed atomic store, the C++11 standard says that the fences synchronize-witheach other, just as I’ve shown. I like C++11’s approach to portable memory fences.
The most important thing to know about acquire and release fences is that they can establish a synchronizes-withrelationship, which means that they prohibit memory reordering in a way that allows you to pass information reliably between threads.
First things first: Acquire and release fences are considered low-levellock-free operations. If you stick with higher-level, sequentially consistentatomic types, such as volatilevariables in Java 5+, or default atomics in C++11, you don’t need acquire and release fences.
On the SPARC-V9 architecture, an acquire fence can be implemented using the membar #LoadLoad | #LoadStoreinstruction, and an a release fence can be implemented as membar #LoadStore | #StoreStore.
As I understand it (mainly from reading Jeff Preshings blog), an atomic_thread_fence(std::memory_order_acq_rel)
prevents any reorderings except for StoreLoad
, i.e., it still allows to reorder a Store
with a subsequent Load
. However, this is exactly the reordering that has to be prevented in your example.
More precisely, an atomic_thread_fence(std::memory_order_acquire)
prevents the reordering of any previous Load
with any subsequent Store
and any subsequent Load
, i.e., it prevents LoadLoad
and LoadStore
reorderings across the fence.
An atomic_thread_fence(std::memory_order_release)
prevents the reordering of any subsequent Store
with any preceding Store
and any preceding Load
, i.e., it prevents LoadStore
and StoreStore
reorderings across the fence.
An atomic_thread_fence(std::memory_order_acq_rel)
then prevents the union, i.e., it prevents LoadLoad
, LoadStore
, and StoreStore
, which means that only StoreLoad
may still happen.
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With