Skip site navigation (1) Skip section navigation (2)

Restartable signals 'n all that

From: Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>
To: pgsql-hackers(at)postgreSQL(dot)org
Subject: Restartable signals 'n all that
Date: 2007-07-01 16:11:26
Message-ID: 12819.1183306286@sss.pgh.pa.us (view raw or flat)
Thread:
Lists: pgsql-hackers
While poking at the vacuum-launcher issue currently under discussion,
I got annoyed again at the behavior we've known for a long while that
on some platforms pg_usleep() won't be interrupted by signals.  (In
particular I see this on HPUX, though not on Linux or Darwin.  Anyone
know if it happens on any BSDen?)  I noticed that with the launcher set
to sleep at most one second between checks for signals, it seemed to
*always* take the full second before shutting down, which seemed awfully
unlucky.

Some more testing and man-page-reading revealed the full truth of what's
going on.  The Single Unix Spec's select(2) page says under ERRORS

    [EINTR]
    The select() function was interrupted before any of the selected events
    occurred and before the timeout interval expired. If SA_RESTART has been
    set for the interrupting signal, it is implementation-dependent whether
    select() restarts or returns with [EINTR].

Since pqsignal() sets SA_RESTART for all trapped signals except SIGALRM,
that means we are exposing ourselves to the implementation dependency.
What I furthermore realized while tracing is that "restart" means
"start counting down the full timeout interval over again".  Thus, if
we have told select() to time out after 1 second, and SIGINT arrives
after 0.9 second, we will wait a full second more before responding.

Bad as that is, it gets worse rapidly: each new signal arrival restarts
the cycle.  So a continuous flow of signals at a spacing of less than
1 second would prevent the delay from *ever* terminating.

This may be why some kernels reduce the timeout value before returning,
so that a "restart" behavior in userland behaves sanely.  But that's
not what's happening for me :-(.

To me, this calls into question whether we should try to avoid using
SA_RESTART at all.  The reason for doing it of course is to avoid
unexpected syscall EINTR failures as well as short read/short write
behaviors during disk I/O.  However, if that's the plan then what the
heck is pqsignal() doing giving an exception for SIGALRM?  As soon as
you have even one non-restartable trapped signal, it seems you need
to handle EINTR everywhere.

I looked into the CVS history and found that we inherited the SIGALRM
exception from Berkeley (in fact it's in the v4r2 sources from 1994).
Back then the system's usage of SIGALRM was pretty darn narrow --- it
was used only to trigger the deadlock checker, which means it applied
only while waiting for a lock, and the range of code in which the
interrupt could occur was just a few lines.  Now that we use SIGALRM for
statement_timeout, the interrupt can potentially happen almost anywhere
in the backend code.

So we've got two problems: SA_RESTART is preventing some EINTRs from
happening when we'd like, and yet it seems we are at risk of unwanted
EINTRs anyway.

The only really clean solution I can see is to stop using SA_RESTART
and try to make all our syscalls EINTR-proof.  But the probability
of bugs-of-omission seems just about 100%, especially in third party
backend add-ons that we don't get to review the code for.

If we do nothing, anyone using statement_timeout is at risk.  The
risk is somewhat ameliorated by the fact that occurrence of the
interrupt means transaction cancellation anyway, so an unexpected
error of some other type isn't really a fatal problem.  But it's
still a bit nervous-making.  I don't currently see a way to get
corrupted data from an EINTR (bufmgr is fairly robust about write
failures, for instance) but ...

If we decide to live with that, and fix any reported problems, then
one thing we could do to ameliorate the sleep problem is to turn
off SA_RESTART for all activity-cancelling interrupts, in particular
SIGINT/SIGTERM/SIGQUIT.  This wouldn't make it safe for bgwriter
and friends to go back to long sleep intervals, because they are
watching for other interrupts too that don't represent reasons to
cancel transactions.  But it would at least solve the problem of
slow response to shutdown requests.

Comments?  I sure hope someone has a better idea.

			regards, tom lane

Responses

pgsql-hackers by date

Next:From: Tom LaneDate: 2007-07-01 16:25:34
Subject: Re: [COMMITTERS] pgsql: Improve logging of checkpoints.
Previous:From: Tom LaneDate: 2007-07-01 14:06:42
Subject: Re: Something is fairly whacko about shutdown in CVS HEAD

Privacy Policy | About PostgreSQL
Copyright © 1996-2014 The PostgreSQL Global Development Group