Re: gdb test failure debug status update

[Philippe Gerum <rpm@xenomai.org>]

Chen, Hongzhan via Xenomai <xenomai@xenomai.org> writes:

> According to my validation, gdb test fail on dovetail 5.10 branch but pass on v5.9-evl4 tag with same for-upstream/dovetail
> xenomai code base.
>
> After further debug , the issue is more clear for me. Gdb test failure because low priority thread smokey  userspace  is still
> executed after "cobalt_shadow_relaxed: state=0x4488c0 info=0x200"  like log [1] on dovetail-5.10 branch.
> The weird thing is that its following first ftrace log  happen at 62235.848583 after cobalt_shadow_relaxed in log [1].
> It is almost 3ms happened after cobalt_shadow_relaxed. The low priority smoke thread user space is executed during this
> 3ms period so that test fail.
>
> But in success case with v5.9-evl4 like in  log [2], the time interval between cobalt_shadow_relaxed and the following first ftrace log
> is only about 1us. It seems that low priority smokey userspace do not have chance to execute in this 1us because gdb test is successful.
>
> My question is why there is even no interrupt happened during that about 3ms period in failure case?  Tick seems in abnormal behavior.
> Please comment if you have any ideas to further debug it.
>
> PS: All my tests run on same up Xtream board.

<snip>

Let's put aside the tick issue for now, there may be a valid reason for
this delay with dynticks enabled.

The issue at stake may be related to the way a return to kernel space is
forced on a @user task (Dovetail has an integrated service for
triggering this called dovetail_request_ucall()).

The logic for doing so is as follows:

1. @user hits a breakpoint, which is an exception Dovetail-wise

2. @user gets XNDBGSTOP set into its flags because Cobalt notices it is
being debugged via a breakpoint trap, then relaxed as a result of taking
a exception in general, so that we may traverse the common trap handling code
safely.

3. since XNDBGSTOP is a blocking bit Cobalt-wise, it should prevent
@user from being picked for scheduling by the real-time core, next time
a Cobalt considers rescheduling that is. However, since @user is
currently relaxed, it can still run under the supervision of the common
Linux scheduler. This is what the log[1] show.

4. the common/in-band kernel code stops @user due to the ptrace stop
condition caused by the breakpoint, waiting for a continuation event to
happen.

Therefore, upon PTRACE_CONT (i.e. gdb continue), we need to force @user
to call back into kernel context (handle_ptrace_cont ->
dovetail_request_ucall), then ask for a switch to primary mode from
there, which should eventually happen when @user is about to leave the
kernel (on x86, this now happens from a generic kernel entry/exit code
in kernel/entry/*). As a result, handle_taskexit_event() runs, figures
out that @user is pending a switch to primary mode. As it switches to
primary mode, @user would be blocked by Cobalt from running further,
because XNDBGSTOP is set into its internal state.

So, I would check a few things for starters:

- is dovetail_request_ucall() working properly.
- is XNCONTHI properly set into the local Cobalt flags of @user when
  handle_user_return() is entered.
- is this path taken as expected once dovetail_request_ucall() has run
  for @user:
  exit_to_user_mode_prepare -> do_retuser -> inband_retuser_notify
  (kernel/entry/common.c)?
  
It may be a good idea to enable all cobalt tracepoints, add one to
handle_user_return() too.

-- 
Philippe.