Re: [PATCH] monitor: flush messages on abort

[Markus Armbruster <armbru@redhat.com>]

Daniel P. Berrangé <berrange@redhat.com> writes:

> On Fri, Nov 03, 2023 at 03:51:00PM -0400, Steven Sistare wrote:
>> On 11/3/2023 1:33 PM, Daniel P. Berrangé wrote:
>> > On Fri, Nov 03, 2023 at 09:01:29AM -0700, Steve Sistare wrote:
>> >> Buffered monitor output is lost when abort() is called.  The pattern
>> >> error_report() followed by abort() occurs about 60 times, so valuable
>> >> information is being lost when the abort is called in the context of a
>> >> monitor command.
>> > 
>> > I'm curious, was there a particular abort() scenario that you hit ?
>> 
>> Yes, while tweaking the suspended state, and forgetting to add transitions:
>> 
>>         error_report("invalid runstate transition: '%s' -> '%s'",
>>         abort();
>> 
>> But I have previously hit this for other errors.

Can you provide a reproducer?

>> > For some crude statistics:
>> > 
>> >   $ for i in abort return exit goto ; do echo -n "$i: " ; git grep --after 1 error_report | grep $i | wc -l ; done
>> >   abort: 47
>> >   return: 512
>> >   exit: 458
>> >   goto: 177
>> > 
>> > to me those numbers say that calling "abort()" after error_report
>> > should be considered a bug, and we can blanket replace all the
>> > abort() calls with exit(EXIT_FAILURE), and thus avoid the need to
>> > special case flushing the monitor.
>> 
>> And presumably add an atexit handler to flush the monitor ala monitor_abort.
>> AFAICT currently no destructor is called for the monitor at exit time.
>
> The HMP monitor flushes at each newline,  and exit() will take care of
> flushing stdout, so I don't think there's anything else needed.

Correct.

>> > Also I think there's a decent case to be made for error_report()
>> > to call monitor_flush().

No, because the messages printed by error_report() all end in newline,
and printing a newline to a monitor triggers monitor_flush_locked().

>> A good start, but that would not help for monitors with skip_flush=true, which 
>> need to format the buffered string in a json response, which is the case I 
>> tripped over.
>
> 'skip_flush' is only set to 'true' when using a QMP monitor and invoking
> "hmp-monitor-command".

Correct.

> In such a case, the error message needs to be built into a JSON error
> reply and sent over the socket. Your patch doesn't help this case
> since you've just printed to stderr.  I don't think it is reasonable
> to expect QMP monitors to send replies on SIG_ABRT anyway. So I don't
> think the skip_flush=true scenario is a problem to be concerned with.
>
>> >> To fix, install a SIGABRT handler to flush the monitor buffer to stderr.
>> >>
>> >> Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
>> >> ---
>> >>  monitor/monitor.c | 38 ++++++++++++++++++++++++++++++++++++++
>> >>  1 file changed, 38 insertions(+)
>> >>
>> >> diff --git a/monitor/monitor.c b/monitor/monitor.c
>> >> index dc352f9..65dace0 100644
>> >> --- a/monitor/monitor.c
>> >> +++ b/monitor/monitor.c
>> >> @@ -701,6 +701,43 @@ void monitor_cleanup(void)
>> >>      }
>> >>  }
>> >>  
>> >> +#ifdef CONFIG_LINUX
>> >> +
>> >> +static void monitor_abort(int signal, siginfo_t *info, void *c)
>> >> +{
>> >> +    Monitor *mon = monitor_cur();
>> >> +
>> >> +    if (!mon || qemu_mutex_trylock(&mon->mon_lock)) {
>> >> +        return;
>> >> +    }
>> >> +
>> >> +    if (mon->outbuf && mon->outbuf->len) {
>> >> +        fputs("SIGABRT received: ", stderr);
>> >> +        fputs(mon->outbuf->str, stderr);
>> >> +        if (mon->outbuf->str[mon->outbuf->len - 1] != '\n') {
>> >> +            fputc('\n', stderr);
>> >> +        }
>> >> +    }
>> >> +
>> >> +    qemu_mutex_unlock(&mon->mon_lock);
>> > 
>> > The SIGABRT handling does not only fire in response to abort()
>> > calls, but also in response to bad memory scenarios, so we have
>> > to be careful what we do in signal handlers.
>> > 
>> > In particular using mutexes in signal handlers is a big red
>> > flag generally. Mutex APIs are not declare async signal
>> > safe, so this code is technically a POSIX compliance
>> > violation.

"Technically a POSIX compliance violation" sounds like something only
pedants would care about.  It's actually a recipe for deadlocks and
crashes.

>> Righto.  I would need to mask all signals in the sigaction to be on the safe(r) side.
>
> This is still doomed, because SIGABRT could fire while 'mon_lock' is
> already held, and so this code would deadlock trying to acquire the
> lock.

Yup.

There is no way to make async signal unsafe code safe.

>> > So I think we'd be safer just eliminating the explicit abort()
>> > calls and adding monitor_flush call to error_report.
>> 
>> I like adding a handler because it is future proof.  No need to play whack-a-mole when
>> developers re-introduce abort() calls in the future.  A minor benefit is I would not
>> need ack's from 50 maintainers to change 50 call sites from abort to exit.
>
> That's a bit of a crazy exaggeration. THe aborts() don't cover 50 different
> subsystems, and we don't require explicit acks from every subsystem maintainer
> for trivial cleanups like this.
>
>> A slight risk of the exit solution is that something bad happened at the call site, so 
>> qemu state can no longer be trusted.  Calling abort immediately may be safer than calling 
>> exit which will call the existing atexit handlers and could have side effects.
>
> If that was a real problem, then we already face it because we have
> ~500 places already calling exit() and only 50 calling abort().
>
>> A third option is to define qemu_abort() which flushes the monitor, and replaces all abort
>> calls.  That avoids async-signal-mutex hand wringing, but is still subject to whack-a-mole.
>> 
>> So: atexit, signal handler, or qemu_abort?  I will go with your preference.
>
> Just replace abort -> exit.

I disagree.

Use exit(1) for fatal errors.

Use abort() for programming errors (a.k.a. bugs).

> I'm not seeing a need for an atexit handler on top.

I'm not yet seeing a need for anything.  A reproducer may change that.