[ic] Re: ALERT: bad pipe signal received for /page.html

Tue Dec 12 18:43:29 EST 2006

> >> > Hello, I've been plagued by apache2 segfaults ever since I started
> >> > using Interchange::Link years ago.  The latest Link.pm has ALERT
> >> > messages accompanying the segfaults in error_log:
> >> >
> >> > ALERT: bad pipe signal received for /page.html
> >> > [Sat Dec 09 10:27:55 2006] [notice] child pid 21337 exit signal
> >> > Segmentation fault (11)
> >> >
> >> > Does anyone have any advice on solving this?  I'm using
> >> apache-2.0.58
> >> > and mod_perl-2.0.2 in Gentoo Linux.
> >>
> >> Also, here is the portion of Link.pm that the ALERT seems to come
> >> from:
> >>
> >> # Return this message to the browser when the server is not running.
> >> # Log an error log entry if set to notify
> >>
> >> sub die_page {
> >>
> >>     my $r = shift;
> >>     my $msg;
> >>
> >>     warn "ALERT: bad pipe signal received for $ENV{SCRIPT_NAME}\n";
> >>
> >>     $r->content_type ("text/html");
> >>     $r->print (<<EOF);
> >> <HTML><HEAD><TITLE>Interrupted</TITLE></HEAD>
> >> <BODY BGCOLOR="#FFFFFF">
> >> <H3>Someone pressed stop...</H3>
> >> <P>
> >> We have aborted this request because someone terminated it.
> >> Please try again soon.
> >> </BODY></HTML>
> >> EOF
> >>
> >> }
> >>
> >> Please let me know if you have any ideas.
> >
> > The segfaults are eliminated by commenting out the $r stuff in the
> > die_page sub.  I still get the ALERTs though.  Does anyone have any
> > advice on figuring out why I'm having the bad pipe problem?  Is there
> > an easy way to add extra debugging info to the sub?
> >
> > Also, restarting IC with PERL_SIGNALS=unsafe increases the ALERTs
> > 50 fold.
>
> I've been seeing this too, on my Apache 2 and latest Link.pm. I also
> had to use PERL_SIGNALS=unsafe and so I get quite a lot of these.
>
> The visible effect on the browser is that the page or image (which
> Link.pm apparently still has some part in delivering) does not load.
> I get them myself when browsing and testing my websites, and I have
> never stopped loading a page or had any other problems on non-IC
> sites I host.
>
> I was told the problem stems from either the browser and a stop
> button or some other network fault. I may go back to Apache 1.3 to
> get around this.

I've been working on this all day and I think I may have a solution.

Incidentally, I want to mention that if I add $! to the Link.pm warn
line, I can see that there are two different types of ALERTs in
error_log:

Broken pipe
Inappropriate ioctl for device

Also incidentally, hitting F5 repeatedly always prints a few ALERTS in
error_log, but the number of ALERTs printed seems to be about 10 times
more for an http page than for the same page accessed via https.  I
checked and tested the settings in my *:80 and *:443 vhosts trying to
narrow that down, but I didn't come up with anything.

Now, as I mentioned a few posts back, commenting out the html delivery
stuff in the die_page sub of Link.pm eliminates the segfaults in
error_log but not the ALERTs.  Today I enabled the prefork mpm in
apache2 and tinkered with the Server/Child settings in interchange.cfg
and httpd.conf and that did seem to reduce the ALERTs somewhat.

After all of this, I've been browsing around my site and I haven't
seen a single image or page fail to load like it used to.  The thing
is, ALERTs still show up in error_log even for requests I'm sure I
created myself AND WERE SERVED SUCCESSFULLY.  I hypothesize that the
failed requests were because a $SIG{PIPE} was generated for whatever
reason (although not because the user clicked the stop button), the
html was delivered, it caused a segfault, and the request failed.  I
should mention though, that if the failing request syndrome is fixed,
it could also be from the apache2, httpd.conf, or interchange.cfg
modifications mentioned above.  I doubt it though.

Incidentally, I'd like to mention that the ALERTs can be generated
pretty reliably by hitting the browser's stop button in the middle of
a request, clicking on the same link again or a different link in the
middle of a request (my Mom probably still double clicks links), or
just refreshing the page in the middle of a request.

If the failed requests do stop at this point, I think the die_page
section of Link.pm is doing more harm than good.  Is there anything
wrong with not handling $SIG{PIPE} at all?

- Grant