[EMAIL PROTECTED] wrote:
Ian Appru wrote:
While atomic increment/decrements (using cas) are suffiently well documented
for me to be pretty
confident in my implementation I would apreciate any input for an atomic
conditional increment.
I'd expect the following to work on OpenSolaris:
#include <atomic.h>
int atomic_conditional_increment(int *target)
{
for (;;) {
int old = *target;
if (old == 0)
return 0;
if (atomic_cas_uint((uint_t *) target, old, old + 1) == old)
return old + 1;
}
}
You will certainly need a membar before reading *target.
(If not, it can happen that one thread continuously sees "0" while
another CPU runs away incrementing the counter.
Can you elaborate? Which membar do you think is necessary? I don't see
any ordering requirement outside of atomic_cas_uint(). (Data cache
coherency is guaranteed by SPARC V9. membars only offer ordering and
sequencing guarantees.)
(And you function has a bug in that it will not always update a non-zero
target)
Again, can you elaborate? I think the load of a 0 target will always
result in an atomic_cas_uint(). That atomic_cas_uint() may fail to
increment *target, but we'll follow that up with another load of the
target and another atomic_cas_uint(), ad infinitum.
_______________________________________________
opensolaris-discuss mailing list
[email protected]
