* Peter Zijlstra peterz@infradead.org wrote:
On Fri, 2013-02-22 at 14:54 +0100, Ingo Molnar wrote:
- Peter Zijlstra peterz@infradead.org wrote:
On Fri, 2013-02-22 at 13:50 +0100, Frederic Weisbecker wrote:
atomic64_read() and atomic64_set() are supposed to take care of that, without even the need for _inc() or _add() parts that use LOCK.
Are you sure? Generally atomic*_set() is not actually an atomic operation.
as per Documentation/atomic_ops.h:
I think the interesting part is:
"The setting is atomic in that the return values of the atomic operations by all threads are guaranteed to be correct reflecting either the value that has been set with this operation or set with another operation. A proper implicit or explicit memory barrier is needed before the value set with the operation is guaranteed to be readable with atomic_read from another thread."
Which does give us the wanted guarantee, however:
I checked arch/x86/include/asm/atomic64_32.h and we use cmpxchg8b for everything from _set() to _read(), which translates into 'horridly stupendifyingly slow' for a number of machines, but coherent.
That's a valid concern - and cmpxchg8b is the only 64-bit op available on most 32-bit x86 CPUs which does not involve the FPU.
Wondering how significant this range of x86 problem boxes will be by the time any of these changes reaches upstream and distros - and how much 'horridly stupendifyingly slow' is in terms of cycles expended.
Thanks,
Ingo