On Wed, Nov 24, 2021 at 11:43:05AM +0100, Vlastimil Babka wrote:

quoted

quoted

Any thoughts?  For now I can just hack around this by skipping
reclaim_throttle if cgroup_reclaim() == true, but that's probably not
the correct fix. :)

No, it wouldn't be but a possibility is throttling for only 1 jiffy if
reclaiming within a memcg and the zone is balanced overall.

The interruptible part should just be the patch below. I need to poke at
the cgroup limit part a bit

As the throttle timeout is short anyway, will the TASK_UNINTERRUPTIBLE vs
TASK_INTERRUPTIBLE make a difference for the (ability to kill? AFAIU
typically this inability to kill is because of a loop that doesn't check for
fatal_signal_pending().

Yep, and the fatal_signal_pending() is lacking within reclaim in general
but I'm undecided on how much that should change in the context of reclaim
throttling but at minimum, I don't want the signal delivery to be masked
or delayed.

-- 
Mel Gorman
SUSE Labs