Andres Freund <andres@anarazel.de> writes:
> On Wednesday 20 January 2010 06:30:28 Tom Lane wrote:
>> Er ... what? I believe there are live platforms with sig_atomic_t = char.
>> If we're assuming more that's a must-fix.
> The reason I have asked is that the code is doing things like:
> [ grabbing a spinlock to read a single integer ]
Yes, I think we probably actually need that. The problem is not so
much whether the read is an atomic operation as whether you can rely
on getting an up-to-date value. On multiprocessors with weak memory
ordering you need some type of "sync" instruction to be sure you will
see a value that was recently written by another processor. Currently,
we embed such instructions in the spinlock acquire/release code.
There's been some discussion of exposing memory sync independently
of lock acquisition; perhaps that would be enough here, but I haven't
looked at the surrounding logic enough to say.
My complaint at the top was responding to the idea that someone might
be supposing the specific type sig_atomic_t was at least as wide as
int. That's a different matter altogether. We do assume in some places
that we can read or write the specific type TransactionId indivisibly,
but we don't try to declare it as sig_atomic_t.
> or similar things with LWLockAcquire in a signal handler
[ grows visibly pale ] *Please* tell me we are not trying to take
locks in a signal handler. What happens if it interrupts code that
is already holding that lock?
regards, tom lane