On Fri, Sep 3, 2010 at 4:20 PM, Heikki Linnakangas
<heikki.linnakangas@enterprisedb.com> wrote:
> Maybe that's ok, if I'm reading the deadlock checker code correctly, it also
> calls semop() to increment the another process' semaphore, and the deadlock
> checker can be invoked from a signal handler while in semop() to wait on our
> process' semaphore. BTW, sem_post(), which is the Posix function for
> incrementing a semaphore, is listed as a safe function to call in a signal
> handler. But it's certainly fishy.
Color me confused; I may need to backpedal rapidly here. I had
thought that what Tom was complaining about was the fact that the
signal handler was taking LWLocks, which I would have thought to be
totally unsafe. But it seems the deadlock detector does the same
thing, more or less because the signal handlers are set up so that
they don't do anything unless we're within a limited section of code
where nothing too interesting can happen. I'm not too sure why we
think that it's safe to invoke the deadlock detector that way, but
it's also not too clear to me why this case is any worse.
It furthermore appears that Simon's reply to Tom's complaint about
this function was: "This was modelled very closely on
handle_sig_alarm() and was reviewed by other hackers. I'm not great on
that, as you know, so if you can explain what it is I can't do, and
how that differs from handle_sig_alarm running the deadlock detector
in the same way, then I'll work on it some more."
I guess I need the same explanation.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company