Re: Proposal for Signal Detection Refactoring

On Tue, Sep 25, 2018 at 3:03 AM Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> Michael Paquier <michael(at)paquier(dot)xyz> writes:
> > And then within separate signal handlers things like:
> > void
> > StatementCancelHandler(SIGNAL_ARGS)
> > {
> > [...]
> > signalPendingFlags |= PENDING_INTERRUPT | PENDING_CANCEL_QUERY;
> > [...]
> > }
>
> AFAICS this still wouldn't work. The machine code is still going to
> look (on many machines) like "load from signalPendingFlags,
> OR in some bits, store to signalPendingFlags". So there's still a
> window for another signal handler to interrupt that and store some
> bits that will get lost.
>
> You could only fix that by blocking all signal handling during the
> handler, which would be expensive and rather pointless.
>
> I do not think that it's readily possible to improve on the current
> situation with one sig_atomic_t per flag.
>

After a fair bit of reading I think there are ways of doing this in C11 but
I don't think those are portable to C99.

In C99 (and, in practice C89, as the C99 committee noted there were no
known C89 implementations where reading was unsafe), reading or writing a
static sig_atomic_t inside a signal handler is safe, but a round-trip is
*not* guaranteed not to clobber. While I could be wrong, I think it is
only in C11 that you have any round-trip operations which are guaranteed
not to clobber in the language itself.

Basically we are a long way out to be able to consider these a single value
as flags.

However, what I think one could do is use a struct of volatile
sig_atomic_t members and macros for checking/setting. Simply writing a
value is safe in C89 and higher.

> regards, tom lane
>

--
Best Regards,
Chris Travers
Head of Database

Tel: +49 162 9037 210 | Skype: einhverfr | www.adjust.com
Saarbrücker Straße 37a, 10405 Berlin