Re: [PATCH] copy over oom_adj value at fork time

[Paul Menage <menage@google.com>]

On Tue, Jul 14, 2009 at 6:15 PM, Paul Menage<menage@google.com> wrote:
>
> But ironically, with this fix applied the main part of the original
> change (force all threads in a process to share a single oom_adj
> value) will start to break my code - it's no longer possible to have
> the regular threads in a process be oom-immune, then vfork() and set a
> non-disabled oom_adj in the child, since this will set it for the
> entire process. (Our job scheduler does something like this, in order
> to have the launcher be OOM immune and the running jobs be at various
> oom_adj levels depending on their priority).
>

How about if instead of having the oom_adj be per-mm, we kept an array
of counters in the mm, tracking how many users were at each oom_adj
level; the OOM killer could then use the level of the mm's highest
oom_adj user when deciding how to calculate the badness of a thread
using that mm.

That would preserve the previous semantics of letting a spawned child
inherit a per-thread oom_adj value, while avoiding the specific
problem of the OOM killer getting livelocked (that David's patch
originally addressed) and the more general case of the inconsistency
in determining the oom_adj level of an mm depending on which thread
you look at.

For the very common case where all users of the mm are at the same
oom_adj level, this array could be left unallocated as NULL, and the
thread's own oom_adj used.

Paul