Re: [PATCH] ring-buffer: Add set/clear_current_oom_origin() during allocations

[Michal Hocko <mhocko@kernel.org>]

On Wed 04-04-18 16:59:18, Joel Fernandes wrote:
> Hi Steve,
> 
> On Wed, Apr 4, 2018 at 9:18 AM, Joel Fernandes <joelaf@google.com> wrote:
> > On Wed, Apr 4, 2018 at 9:13 AM, Steven Rostedt <rostedt@goodmis.org> wrote:
> > [..]
> >>>
> >>> Also, I agree with the new patch and its nice idea to do that.
> >>
> >> Thanks, want to give it a test too?
> 
> With the latest tree and the below diff, I can still OOM-kill a victim
> process doing a large buffer_size_kb write:
> 
> I pulled your ftrace/core and added this:
> +       /*
>         i = si_mem_available();
>         if (i < nr_pages)
>                 return -ENOMEM;
> +       */
> 
> Here's a run in Qemu with 4-cores 1GB total memory:
> 
> bash-4.3# ./m -m 1M &
> [1] 1056
> bash-4.3#
> bash-4.3#
> bash-4.3#
> bash-4.3# echo 10000000 > /d/tracing/buffer_size_kb
> [   33.213988] Out of memory: Kill process 1042 (bash) score
> 1712050900 or sacrifice child
> [   33.215349] Killed process 1056 (m) total-vm:9220kB,
> anon-rss:7564kB, file-rss:4kB, shmem-rss:640kB

OK, so the reason your memory hog is triggered is that your echo is
built-in and we properly select bask as an oom_origin but then another
clever heuristic jumps in and tries to reduce the damage by sacrificing
a child process. And your memory hog runs as a child from the same bash
session.

I cannot say I would love this heuristic. In fact I would really love to
dig it deep under the ground. But this is a harder sell than it might
seem. Anyway is your testing scenario really representative enough to
care? Does the buffer_size_kb updater runs in the same process as any
large memory process?

> bash: echo: write error: Cannot allocate memory
> [1]+  Killed                  ./m -m 1M
> bash-4.3#
> --

-- 
Michal Hocko
SUSE Labs