Re: [PATCH] workqueue: defer to waiting stop_machine

[Andreas Mohr <andi@lisas.de>]

> Isn't the problem that the kworker wouldn't yield to the higher
> priority stopper task while a work item keeps requeueing itself if
> preemption is not enabled?  If so, isn't the correct solution just
> adding cond_resched() in the work item processing loop?  The analysis
> and solution seem to have gone a bit stray....

While I did not quite follow the very fine and detailed analysis,
I had the same feeling about it.

The previous solution seemed less preferable e.g. for two reasons,
from a modularity/dependency POV:
- required a very specific (code smell?) stop_machine handling dependency
  in work queue code (machine stop handling arguably definitely
  is a corner case, and thereby supposed to remain just that!)
- new stop_machine_pending() helper is pretty bloated,
  and called in a semi-hotpath to boot (since it's using && operators
  rather than ||, seems like it would be called pretty much every time)

Preemption checks being expected to be much more general and widespread
thus seems like a much better fit.


Or, to put it another way, could it be that that extra very specific
stop_machine check was simply added since due to missing preemption checks
we were busy-handling there and thus not getting back to standard handling areas
where some *usual*, *hotpath/mainstream* stop_machine checks would have been made?
If so, perhaps there actually are some other cases of wasteful stop_machine check
code sites in the kernel where instead we could simply have a much cheaper
reschedule done, thereby go back to hitting one central (and thus cache-hot)
code site with stop_machine check etc.?


Afraid of having stated the glaringly obvious ;),

Andreas Mohr