Re: [RFC net-next 1/2] selftests: drv-net: add ability to schedule cleanup with defer()

[Jakub Kicinski <kuba@kernel.org>]

On Wed, 26 Jun 2024 12:18:58 +0200 Petr Machata wrote:
> Jakub Kicinski <kuba@kernel.org> writes:
> > +def ksft_flush_defer():
> > +    global KSFT_RESULT
> > +
> > +    while global_defer_queue:
> > +        entry = global_defer_queue[-1]
> > +        try:
> > +            entry.exec()  
> 
> I wonder if you added _exec() to invoke it here. Because then you could
> just do entry = global_defer_queue.pop() and entry._exec(), and in the
> except branch you would just have the test-related business, without the
> queue management.

Initially I had both _exec, and _dequeue as separate helpers, but then
_dequeue was identical to cancel, so I removed that one, but _exec
stayed.

As you point out _exec() would do nicely during "flush".. but linter was
angry at me for calling private functions. I couldn't quickly think of
a clean scheme of naming things. Or rather, I should say, I like that
the only non-private functions in class defer right now are
test-author-facing. At some point I considered renaming _exec() to
__call__() or run() but I was worried people will incorrectly
call it, instead of calling exec().

So I decided to stick to a bit of awkward handling in the internals for
the benefit of more obvious test-facing API. But no strong preference,
LMK if calling _exec() here is fine or I should rename it..

> > +        except Exception:  
> 
> I think this should be either an unqualified except: or except
> BaseException:.

SG


> >      print(
> >          f"# Totals: pass:{totals['pass']} fail:{totals['fail']} xfail:{totals['xfail']} xpass:0 skip:{totals['skip']} error:0"  
> 
> Majority of this hunk is just preparatory and should be in a patch of
> its own. Then in this patch it should just introduce the flush.

True, will split.

> > +    def cancel(self):  
> 
> This shouldn't dequeue if not self.queued.

I was wondering if we're better off throwing the exception from
remove() or silently ignoring (what is probably an error in the 
test code). I went with the former intentionally, but happy to
change.

> > +        self._queue.remove(self)
> > +        self.queued = False
> > +
> > +    def exec(self):  
> 
> This shouldn't exec if self.executed.
> 
> But I actually wonder if we need two flags at all. Whether the defer
> entry is resolved through exec(), cancel() or __exit__(), it's "done".
> It could be left in the queue, in which case the "done" flag is going to
> disable future exec requests. Or it can just be dropped from the queue
> when done, in which case we don't even need the "done" flag as such.

If you recall there's a rss_ctx test case which removes contexts out of
order. The flags are basically for that test. We run the .exec() to
remove a context, and then we can check 

	if thing.queued:
		.. code for context that's alive ..
	else:
		.. code for dead context ..