Re: [linux-next: Tree for Jun 1] __khugepaged_exit rwsem_down_write_failed lockup

From: Vlastimil Babka
Date: Wed Jun 08 2016 - 04:19:19 EST


On 06/03/2016 05:10 PM, Andrea Arcangeli wrote:
Hello Michal,

CC'ed Hugh,

On Fri, Jun 03, 2016 at 04:46:00PM +0200, Michal Hocko wrote:
What do you think about the external dependencies mentioned above. Do
you think this is a sufficient argument wrt. occasional higher
latencies?

It's a tradeoff and both latencies would be short and uncommon so it's
hard to tell.

Shouldn't it be possible to do a mmput() before the hugepage allocation, and then again mmget_not_zero()? That way it's no longer a tradeoff?

There's also mmput_async for paths that may care about mmput
latencies. Exit itself cannot use it, it's mostly for people taking
the mm_users pin that may not want to wait for mmput to run. It also
shouldn't happen that often, it's a slow path.

The whole model inherited from KSM is to deliberately depend only on
the mmap_sem + test_exit + mm_count, and never on mm_users, which to
me in principle doesn't sound bad. I consider KSM version a
"finegrined" implementation but I never thought it would be a problem
to wait a bit in exit() in case the slow path hits. I thought it was
more of a problem if exit() runs, the parent then start a new task but
the memory wasn't freed yet.

So I would suggest Hugh to share his view on the down_write/up_write
that may temporarily block mmput (until the next test_exit bailout
point) vs higher latency in reaching exit_mmap for a real exit(2) that
would happen with the proposed change.

Thanks!
Andrea

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>