[Bug 280846] Low memory freezes / OOM: a thread waited too long to allocate a page

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=280846

--- Comment #66 from Henrich Hartzer <henrichhartzer@tuta.io> ---
I'm glad that you've been able to reproduce this. It's nice to be able to see
the usage in top as well.

I'd like to share a few thoughts/anecdotes on this.

When I killed Telegram Desktop my laundry dropped about 400MB.

In Firefox I cleared cache and it cleared a couple hundred MB from laundry. I
need to test that some more to get better figures.

I'm wondering if this is Firefox specific, or if there's some shared component
that is causing this. This may be far fetched but I am wondering if there is a
Linux, maybe glibc behavior where you can mark a page as no longer needed but
useful if it can fit, that isn't translating to FreeBSD. I could see some
page/cache data that would be nice to hold on to if possible, but if there's
any memory contention could be readily be jettisoned. Perhaps those are our
laundry pages? Just speculation from someone who doesn't know a lot about how
allocation works.

After running on UFS for a while, I feel like the OOM behavior is much more
predictable and faster than with ZFS. Under ZFS I was getting hard locks for a
while. With UFS, it seems like it'll usually resolve itself (killing something)
in around 10 seconds). But with ZFS it could lock for minutes with this OOM
condition. I'm not sure why that is, but it's nice that when I run into the bug
it resolves pretty quickly if I'm not able to "catch" it in time.

-- 
You are receiving this mail because:
You are the assignee for the bug.