Re: [PATCH 25/31] Ignore failure to read PC when resuming

[Andrew Burgess <aburgess@redhat.com>]

Pedro Alves <pedro@palves.net> writes:

> If GDB sets a GDB_THREAD_OPTION_EXIT option on a thread, and the
> thread exits, the server reports the corresponding thread exit event,
> and forgets about the thread, i.e., removes the exited thread from its
> thread list.
>
> On the GDB side, GDB set the GDB_THREAD_OPTION_EXIT option on a
> thread, GDB delays deleting the thread from its thread list until it
> sees the corresponding thread exit event, as that event needs special
> handling in infrun.
>
> When a thread disappears from the target, but it still exists on GDB's
> thread list, in all-stop RSP mode, it can happen that GDB ends up
> trying to resume such an already-exited-thread that GDB doesn't yet
> know is gone.  When that happens, against GDBserver, typically the
> ongoing execution command fails with this error:

I'm slightly confused here.  If GDB doesn't know the thread has exited
doesn't that mean the server hasn't yet reported the exit, and so should
be holding onto the thread?

I wanted to investigate this a bit more to try and understand more about
what's going on, but I couldn't find a test that was triggering the code
added in this patch.  Do you know if there's a test I can run to see
this issue?

>
>  ...
>  PC register is not available
>  (gdb)
>
> At the remote protocol level, we may see e.g., this:
>
>       [remote] Packet received: w0;p97479.978d2
>     [remote] wait: exit
>     [infrun] print_target_wait_results: target_wait (-1.0.0 [process -1], status) =
>     [infrun] print_target_wait_results:   619641.620754.0 [Thread 619641.620754],
>     [infrun] print_target_wait_results:   status->kind = THREAD_EXITED, exit_status = 0
>     [infrun] handle_inferior_event: status->kind = THREAD_EXITED, exit_status = 0
>     [infrun] context_switch: Switching context from 0.0.0 to 619641.620754.0
>     [infrun] clear_proceed_status_thread: 619641.620754.0
>
> GDB saw an exit event for thread 619641.620754.  After processing it,
> infrun decides to re-resume the target again.  To do that, infrun
> picks some other thread that isn't exited yet from GDB's perspective,
> switches to it, and calls keep_going.  Below, infrun happens to pick
> thread p97479.97479, the leader, which also exited, but GDB doesn't
> know yet:
>
> ...
>     [remote] Sending packet: $Hgp97479.97479#75
>     [remote] Packet received: OK
>     [remote] Sending packet: $g#67
>     [remote] Packet received: xxxxxxxxxxxxxxxxx (...snip...) [1120 bytes omitted]
>     [infrun] reset: reason=handling event
>     [infrun] maybe_set_commit_resumed_all_targets: not requesting commit-resumed for target remote, no resumed threads
>   [infrun] fetch_inferior_event: exit
>   PC register is not available
>   (gdb)
>
> The Linux backends, both in GDB and in GDBserver, already silently
> ignore failures to resume, with the understanding that we'll see an
> exit event soon.  Core of GDB doesn't do that yet, though.
>
> This patch is a small step in that direction.  It swallows the error
> when thrown from within resume_1.  There are likely are spots where we
> will need similar treatment, but we can tackle them as we find them.
>
> After this patch, we'll see something like this instead:
>
>     [infrun] resume_1: step=0, signal=GDB_SIGNAL_0, trap_expected=0, current thread [640478.640478.0] at 0x0
>     [infrun] do_target_resume: resume_ptid=640478.0.0, step=0, sig=GDB_SIGNAL_0
>     [remote] Sending packet: $vCont;c:p9c5de.-1#78

I'm confuse by this example.  I would have expected it to start off with
the same intro as the above, that is, send the '$g#67' packet, get back
the xxxx...etc... but then do things differently.

>     [infrun] prepare_to_wait: prepare_to_wait
>     [infrun] reset: reason=handling event
>     [infrun] maybe_set_commit_resumed_all_targets: enabling commit-resumed for target remote
>     [infrun] maybe_call_commit_resumed_all_targets: calling commit_resumed for target remote
>   [infrun] fetch_inferior_event: exit
>   [infrun] fetch_inferior_event: enter
>     [infrun] scoped_disable_commit_resumed: reason=handling event
>     [infrun] random_pending_event_thread: None found.
>     [remote] wait: enter
>       [remote] Packet received: W0;process:9c5de
>     [remote] wait: exit
>     [infrun] print_target_wait_results: target_wait (-1.0.0 [process -1], status) =
>     [infrun] print_target_wait_results:   640478.0.0 [process 640478],
>     [infrun] print_target_wait_results:   status->kind = EXITED, exit_status = 0
>     [infrun] handle_inferior_event: status->kind = EXITED, exit_status = 0
>   [Inferior 1 (process 640478) exited normally]
>     [infrun] stop_waiting: stop_waiting
>     [infrun] reset: reason=handling event
>   (gdb) [infrun] fetch_inferior_event: exit
>
> Change-Id: I7f1c7610923435c4e98e70acc5ebe5ebbac581e2
> ---
>  gdb/infrun.c | 23 ++++++++++++++++++++++-
>  1 file changed, 22 insertions(+), 1 deletion(-)
>
> diff --git a/gdb/infrun.c b/gdb/infrun.c
> index 09391d85256..21e5aa0f50e 100644
> --- a/gdb/infrun.c
> +++ b/gdb/infrun.c
> @@ -2595,7 +2595,28 @@ resume_1 (enum gdb_signal sig)
>        step = false;
>      }
>  
> -  CORE_ADDR pc = regcache_read_pc (regcache);
> +  CORE_ADDR pc = 0;

I don't think we should be picking some arbitrary $pc value (0 in this
case) and just using that as a default, instead, I think it would be
better to change the type of pc to gdb::optional<CORE_ADDR>, and then
update the rest of this function to only do the $pc relevant parts if we
have a $pc value.

> +  try
> +    {
> +      pc = regcache_read_pc (regcache);
> +    }
> +  catch (const gdb_exception_error &err)
> +    {
> +      /* Swallow errors as it may be that the current thread exited
> +	 and we've haven't seen its exit status yet.  Let the
> +	 resumption continue and we'll collect the exit event
> +	 shortly.  */
> +      if (err.error == TARGET_CLOSE_ERROR)
> +	throw;
> +
> +      if (debug_infrun)
> +	{
> +	  string_file buf;
> +	  exception_print (&buf, err);
> +	  infrun_debug_printf ("resume: swallowing error: %s",
> +			       buf.string ().c_str ());
> +	}

I guess this is the best we can probably do without changing the remote
protocol. My worry would be that there could be other reasons that the
read of $pc fails, which we are now just ignoring.  It looks like you
already ran into one such case with TARGET_CLOSE_ERROR, but maybe
there's others?

It almost feels like the ideal solution would invert the logic, so we
could write:

  catch (const gdb_exception_error &err)
    {
      /* I just invent a new error type here...  */
      if (err.err != INFERIOR_EXITED_ERROR)
        throw;

      // ... etc ...
    }

To use something like this we could have the H packet send back
something other then "OK" when GDB asks to switch to a thread that has
already exited, maybe send back the stop reply could be made to work?

I say all that really just to check if you agree or not.  I think for
now I'd be happy to go with what you present here, I think the gains
this series brings to GDB are worth some rough edges that we might want
to address in the future.

Would love to hear your thoughts,

Thanks,
Andrew

> +    }
>  
>    infrun_debug_printf ("step=%d, signal=%s, trap_expected=%d, "
>  		       "current thread [%s] at %s",
> -- 
> 2.36.0