This is the mail archive of the
gdb-patches@sourceware.org
mailing list for the GDB project.
Re: [PATCH] Readline: Cleanup some warnings
- From: Eli Zaretskii <eliz at gnu dot org>
- To: Pedro Alves <palves at redhat dot com>
- Cc: tom at tromey dot com, gdb-patches at sourceware dot org
- Date: Tue, 19 Mar 2019 22:14:36 +0200
- Subject: Re: [PATCH] Readline: Cleanup some warnings
- References: <20190130085716.75179-1-alan.hayward@arm.com> <20190131075907.GA313@adacore.com> <F71F4EDA-CBDF-4B05-B9C2-588D02471EB2@arm.com> <3463805B-A8BF-4C20-ACE3-C21AE3F7DB62@arm.com> <20190201080533.GA31043@adacore.com> <877eejvfoq.fsf@tromey.com> <1549047248.2630.7.camel@skynet.be> <310315f8-62ab-2eff-042f-9f2ae9de07da@redhat.com> <87wokxtnlt.fsf@tromey.com> <83h8c1wdr5.fsf@gnu.org> <87imwex333.fsf@tromey.com> <711b6636-b02c-edb2-308d-5fddbf4c33a9@redhat.com>
> Cc: gdb-patches@sourceware.org
> From: Pedro Alves <palves@redhat.com>
> Date: Tue, 19 Mar 2019 19:02:43 +0000
>
> > https://sourceware.org/ml/gdb-patches/2008-02/msg00423.html
Caveat: I didn't yet read that thread myself.
> Hmmm,
>
> Daniel wrote:
>
> > GDB has several SIGINT handlers which call longjmp. This is
> > problematic for at least two reasons. One is that we could be in the
> > middle of something unwise to longjmp out of, for instance malloc. In
> > practice, this never happens because we're usually waiting for I/O
> > when one of the relevant handlers is invoked, but there are a number
> > of places where it could definitely happen.
>
> That was indeed true back then, but since then, immediate_quit
> was completely eliminated, and we no longer longjmp from signal
> handlers anymore, since:
> https://sourceware.org/ml/gdb-patches/2016-03/msg00351.html
>
> Daniel wrote:
>
> > My goals in fixing this were to hide the Windows ugliness, and to fit
> > in nicely with GDB's asynchronous event loop. Since we do not return
> > to the primary event loop during target actions (for the current,
> > non-async GDB), I couldn't rely on the event loop entirely. But I
> > could use the same token mechanism and thus share the bodies of
> > handlers for async mode with the Windows case.
> >
> > The new interface is gdb_call_async_signal_handler. SIGINT handlers,
>
> This interface he mentioned, gdb_call_async_signal_handler, was
> eliminated by that series too:
>
> https://sourceware.org/ml/gdb-patches/2016-03/msg00347.html
>
> So all that's left is that little readline hack, it seems:
>
> /* With multi-threaded SIGINT handling, there is a race between the
> readline signal handler and GDB. It may still be in
> rl_prep_terminal in another thread. Do not return until it is
> done; we can check the state here because we never longjmp from
> signal handlers on Windows. */
> while (RL_ISSTATE (RL_STATE_SIGHANDLER))
> Sleep (1);
>
> (Curiously, that bit only appeared in a later version of Dan's patch,
> here: https://sourceware.org/ml/gdb-patches/2008-03/msg00034.html)
>
> I'm not seeing why we'd still need that bit, but then again,
> I'm not seeing why it was needed in the first place.
> The signal handler could run concurrently with gdb at any other
> point in the gdb code, not just here, so at any point we
> call into readline, we can be running readline code in parallel
> with a signal handler touching readline's state. It sounds like
> that should be a readline problem to worry about.
>
> That could be related to the fact that readline's signal handler
> overrides gdb's, does its thing, and then calls gdb's signal
> handler manually? If the WaitForSingleObject call had already
> woken up, then gdb's signal handler has already run and SetEvent
> on sigint_event. Then the code would go and run the deferred
> signal handler. In the remote case, that handler would
> issue prompt "Give up (and stop debugging it)? (y or n)" prompt,
> and if that is running in parallel with readline's signal
> handler still calling rl_prep_terminal, bad things would happen.
Not sure if the above refers to what I wanted to say, but: as I'm sure
you know, SIGINT handlers on Windows run in a separate thread, created
by the OS, so a Readline SIGINT handler could ruin in parallel both
with Readline's other code and in parallel with GDB's code, depending
on when exactly did the user type Ctrl-C. In a few cases where it was
important to emulate Posix behavior in order not to step on the troes
of the mainline code, I needed to stop the main thread while the
SIGINT handler was running. Could it be that the code we are
discussing does something similar?
> But again, why isn't that a readline problem, instead of
> a gdb problem?
I agree: the right solution would be for the Readline's SIGINT handler
to stop the main thread (e.g., by using SuspendThread).
> I'm still puzzled on why this isn't a readline issue. Shouldn't
> readline's Windows signal handler be synchronizing with mainline
> code such that if a signal handler is running, mainline calls into
> readline would block?
Yes, I think so.
> I think there must be something else to this.
Maybe. I will try to read that discussion soon.
Thanks.