Re: [PATCH 3/3] Wait for target to stop before running agent commands

[Pedro Alves <palves at redhat dot com>]

On 09/18/2014 10:59 AM, Gary Benson wrote:
> This commit fixes a bug on where agent_run_command is called on
> GNU/Linux targets that may not have stopped yet.
> 
> gdb/ChangeLog:
> 
> 	* linux-nat.c (linux_child_static_tracepoint_markers_by_strid):
> 	Use target_stop_and_wait in place of target_stop.
> ---
>  gdb/ChangeLog   |    5 +++++
>  gdb/linux-nat.c |    2 +-
>  2 files changed, 6 insertions(+), 1 deletions(-)
> 
> diff --git a/gdb/linux-nat.c b/gdb/linux-nat.c
> index 5a82d23..f2895ff 100644
> --- a/gdb/linux-nat.c
> +++ b/gdb/linux-nat.c
> @@ -4458,7 +4458,7 @@ linux_child_static_tracepoint_markers_by_strid (struct target_ops *self,
>    ptid_t ptid = ptid_build (pid, 0, 0);
>  
>    /* Pause all */
> -  target_stop (ptid);
> +  target_stop_and_wait (ptid);

More would be needed for fix this fully.

Unlike the comment above suggests, target_stop_and_wait only
waits for one thread to stop, not all.

See the gdbserver version (run_inferior_command) -- this is also
missing uninserting breakpoints after fully pausing the target.

If we handle this internally to linux-nat.c, then similarly
to gdbserver's version, this could use
stop_callback/stop_wait_callback to leave pending any events
that pausing might trip on.  Then we'd need to be careful to
only re-resume threads that were already resumed before, and
leave those with pending events stopped.  If we want to
use existing target methods, then we need to teach gdb's
struct thread_info layer about leaving events pending, like in:

  https://sourceware.org/ml/gdb-patches/2011-12/msg00564.html

Alternative, we can would promote target_stop_and_wait to
a real target method that handles the leaving of events pending,
like gdbserver's pause_all/unpause_all methods.

Thanks,
Pedro Alves