Re: [PATCH v2 04/13] linux-nat: Don't enable async mode at the end of linux_nat_target::resume.

[Andrew Burgess <aburgess@redhat.com>]

* John Baldwin <jhb@FreeBSD.org> [2021-12-02 17:21:27 -0800]:

> On 12/2/21 3:19 AM, Andrew Burgess wrote:
> > * John Baldwin <jhb@FreeBSD.org> [2021-12-01 12:03:41 -0800]:
> > 
> > > On 12/1/21 4:02 AM, Andrew Burgess wrote:
> > > > * John Baldwin <jhb@FreeBSD.org> [2021-11-24 08:03:36 -0800]:
> > > > 
> > > > > On 11/24/21 5:24 AM, Andrew Burgess wrote:
> > > > > > * John Baldwin <jhb@FreeBSD.org> [2021-08-03 11:49:51 -0700]:
> > > > > > 
> > > > > > > Commit 5b6d1e4fa4f ("Multi-target support") enabled async mode in the caller
> > > > > > > (do_target_resume) after target::resume returns making this call
> > > > > > > redundant.
> > > > > > 
> > > > > > The only thing I spotted where this might result in a change of
> > > > > > behaviour is in record-full.c in record_full_wait_1, there's this
> > > > > > line:
> > > > > > 
> > > > > >      ops->beneath ()->resume (ptid, step, GDB_SIGNAL_0);
> > > > > > 
> > > > > > Doesn't this mean that if the record target is sitting on top of a
> > > > > > Linux target, then me could potentially see a change of behaviour
> > > > > > here?
> > > > > > 
> > > > > > I know almost nothing about the record target, so I can't really offer
> > > > > > any insights into whether the situation about could actually happen or
> > > > > > not.
> > > > > > 
> > > > > > Maybe somebody else will have some thoughts, but at a minimum, it is
> > > > > > probably worth mentioning that there might be a change for that
> > > > > > case.... Unless you know for sure that there isn't.
> > > > > 
> > > > > Hmm, I was not aware of that case.  This is a change Pedro had suggested
> > > > > and it didn't show any regressions in the test suite on a Linux/x86-64
> > > > > VM, but it very well might be that the test suite doesn't cover
> > > > > record?
> > > > 
> > > > I suspect it's not well tested.
> > > > 
> > > > > 
> > > > > Hmm, looking at record_full_target::resume, it enables async right after
> > > > > calling the beneath method and it also does so before returning, so I
> > > > > suspect it's explicit call is safe to elide for the same reason as this
> > > > > patch.  That is, in the code today, linux_nat_target::resume enables async
> > > > > mode, then would return to remote_full_target::resume which would enable
> > > > > async mode (without any other statements in between), and then return
> > > > > to do_target_resume which would enable async mode a third time (again without
> > > > > any other code in between aside from any RAII destructors when returning
> > > > > from the target methods).  Given, that I think this should be safe, and
> > > > > I can add a patch to this series to remove the call from
> > > > > remote_full_target::resume.
> > > > 
> > > > That's all true, but I wasn't commenting on
> > > > record_full_target::resume, rather record_full_wait_1, which is called
> > > > from record_full_base_target::wait.
> > > > 
> > > > I'd certainly be happy for your change to go in.  Having looked again,
> > > > I don't think there's going to be any problems, looking at other
> > > > targets, it doesn't seem like having to (re-)enable async mode is
> > > > something we'd normally expect to do on the ::wait path...
> > > > 
> > > > I only pointed this out as it seems to be a case that doesn't align
> > > > with the claims in your commit message...
> > > 
> > > Oh, humm, I see (and indeed, I had misread your earlier message).  It is
> > > true that the commit log isn't accurate (and at the least that would
> > > need to be amended).  One thing I don't really understand are the cases
> > > when async is disabled such that resume needs to re-enable it vs are
> > > the cases in resume just trying to catch an edge case with attach or
> > > create_inferior or the like where async mode isn't enabled by some other
> > > hook.
> > > 
> > > Ah, it seems that in wait_one if a target returns WAITKIND_NO_RESUMED,
> > > we disable async on that target, and then the next time we want to
> > > resume the target we re-enable async after target_resume returns.  The
> > > other places we disable async are either temporary places around
> > > readline, etc. or when when an inferior exits (INF_EXEC_COMPLETE)
> > > passed to inferior_event_handle).
> > > 
> > > Given that, I do think the code in record_full_wait_1 is ok:
> > > 
> > > 1) Any time ::wait() is invoked, it looks like target_async(1) should
> > >     have already been called by the preceding do_target_resume().
> > > 
> > > 2) The loop in record_full_wait_1() doesn't return to the wait_one()
> > >     loop where target_async(0) is invoked, so async for the underlying
> > >     target should stay enabled for the duration of the ::wait() call.
> > > 
> > > 3) All the places that can invoke INF_EXEC_COMPLETE look to be in the
> > >     gdb core itself which is "above" a target ::wait() method on the
> > >     stack.
> > 
> > Thanks for the detailed analysis.  I agree that, to the best of my
> > understanding, your change is fine.  I think it would be worth adding
> > this analysis to the commit message, then if there ever is a bug in
> > the future, we can understand what we thought was going on :)
> 
> I will include it, and I will merge the commit I had made to remove the
> explicit target_async(1) calls in the record targets into this commit
> so that the detailed commit log covers all of them.
> 
> I noticed one additional call that could perhaps be elided which is the
> call to target_async(1) near the end of remote::resume in the remote
> target.  Unlike the Linux native and record targets, this call is not
> at the very end of the function, but I don't think it will break anything
> to run the relevant code before enabling async mode rather than after:
> 
>   /* We are about to start executing the inferior, let's register it
>      with the event loop.  NOTE: this is the one place where all the
>      execution commands end up.  We could alternatively do this in each
>      of the execution commands in infcmd.c.  */
>   /* FIXME: ezannoni 1999-09-28: We may need to move this out of here
>      into infcmd.c in order to allow inferior function calls to work
>      NOT asynchronously.  */
>   if (target_can_async_p ())
>     target_async (1);
> 
>   /* We've just told the target to resume.  The remote server will
>      wait for the inferior to stop, and then send a stop reply.  In
>      the mean time, we can't start another command/query ourselves
>      because the stub wouldn't be ready to process it.  This applies
>      only to the base all-stop protocol, however.  In non-stop (which
>      only supports vCont), the stub replies with an "OK", and is
>      immediate able to process further serial input.  */
>   if (!target_is_non_stop_p ())
>     rs->waiting_for_stop_reply = 1;
> 
> This would also in theory remove the FIXME comment, though I don't know
> that this change is fully realizing the goal of that FIXME (it's also
> over 20 years old though).  If you think it makes sense to elide the
> call in the remote target I can fold this into this commit as well.
> I can also just leave the remote target as-is.

I agree that this call can be safely removed.

Thanks,
Andrew