Re: GDB (not) handling SIGINT...?

On 11/14/2018 10:32 PM, Simon Marchi wrote:
> On 2018-11-13 17:37, Paul Smith wrote:
>> Hi all; I'm using GDB 8.1 on a modern 64bit GNU/Linux system (Ubuntu
>> 18.04LTS).
>>
>> Recently I added a call to sigtimedwait() for SIGINT into my
>> (multithreaded) program and now I'm having an issue with GDB.
>>
>> What I want is that if I attach to my program then continue, then use
>> ^C at the GDB terminal, I should get a (gdb) prompt back but I do NOT
>> want the SIGINT delivered to my program to wake up my sigtimedwait()
>> call (because it will cause my program to do various things that I
>> don't want it to do).
>>
>> I see that SIGINT is set to nopass:
>>
>>   (gdb) info signals SIGINT
>>   Signal        Stop      Print   Pass to program Description
>>   SIGINT        Yes       Yes     No              Interrupt
>>
>> but yet when I use ^C at the GDB prompt my sigtimedwait() call inside
>> my program does return with "2" (SIGINT), which I don't want.
>>
>> I guess I don't understand what the docs mean when they say that the
>> signal is not passed to the process under debug.  Note that in my case
>> I'm attaching to the program from a completely different terminal so
>> there's no issue with process groups etc. and as far as I can
>> understand it, the signal should only be delivered to GDB not my
>> process, so unless there's some weird magic at work here it must be GDB
>> forwarding that signal to my process.
>>
>> Anyone have any thoughts about this?
>>
>> Cheers!
> 
> I was able to reproduce it with, it really sounds like a bug.
> 
> If others want to try, here's my test program:
> 
> #include <stdio.h>
> #include <signal.h>
> #include <stdlib.h>
> 
> int main()
> {
>     sigset_t set;
>     sigemptyset(&set);
>     sigaddset(&set, SIGINT);
>     sigprocmask(SIG_BLOCK, &set, NULL);
>     for (int i = 0; i < 10; i++) {
>         int n = sigwaitinfo(&set, NULL);
>         printf("signal %d\n", n);
>     }
> }
> 
> 
> Run in a terminal, attach with GDB in another terminal, continue, then ctrl-C in GDB's terminal.  The first call to sigwaitinfo returns -1, I think it's expected as the syscall gets interrupted.  But the subsequent calls return 2, showing that indeed the process has received a SIGINT.
> 
> Paul, could you please file a bug?  You can re-use this test program if you want.

There are a few things going on here, none of it is really new, though:

#1 - If your program blocks SIGINT and then uses sigwait, then GDB won't
     ever intercept the SIGINT, because ptrace doesn't ever see the signal.
     This is an ancient issue.  See:
       https://sourceware.org/bugzilla/show_bug.cgi?id=9425

#2 - If you attach to a process instead of running it from GDB, then ctrl-c
     reaches gdb first, and then GDB sends/forwards a SIGINT to the child process.
     Normally this then causes ptrace to intercept the SIGINT, but, see #1 above.
     This case ("attach"), could be handled by GDB instead stopping the
     process with SIGTOP or even better, PTRACE_INTERRUPT.  Note, TBC, would
     only work with attach/a separate terminal.

#3 - If you run the process instead of attaching, in the same terminal
     as GDB, then ctrl-c reaches the inferior process first (because
     the inferior is put in the foreground), gdb would have no chance to
     do PTRACE_INTERRUPT at all.  That would be fixable (combined with 
     PTRACE_INTERRUPT) by my WIP branch here:
       https://github.com/palves/gdb/commits/palves/tty-always-separate-session
     which effectively makes the "run" case the same as the "attach" case,
     by always putting the inferior in a separate terminal session.

Thanks,
Pedro Alves