On Fri, 14 Jun 2013, David Rientjes wrote: > Even with all of the above (which is not actually that invasive of a > patch), I still think we need memory.oom_delay_millisecs. I probably made > a mistake in describing what that is addressing if it seems like it's > trying to address any of the above. > > If a userspace oom handler fails to respond even with access to those > "memcg reserves", the kernel needs to kill within that memcg. Do we do > that above a set time period (this patch) or when the reserves are > completely exhausted? That's debatable, but if we are to allow it for > global oom conditions as well then my opinion was to make it as safe as > possible; today, we can't disable the global oom killer from userspace and > I don't think we should ever allow it to be disabled. I think we should > allow userspace a reasonable amount of time to respond and then kill if it > is exceeded. > > For the global oom case, we want to have a priority-based memcg selection. > Select the lowest priority top-level memcg and kill within it. If it has > an oom notifier, send it a signal to kill something. If it fails to > react, kill something after memory.oom_delay_millisecs has elapsed. If > there isn't a userspace oom notifier, kill something within that lowest > priority memcg. > > The bottomline with my approach is that I don't believe there is ever a > reason for an oom memcg to remain oom indefinitely. That's why I hate > memory.oom_control == 1 and I think for the global notification it would > be deemed a nonstarter since you couldn't even login to the machine. > What's the status of this patch? I'd love to be able to introduce memcg reserves so that userspace oom notifications can actually work, but we still require a failsafe in the kernel if those reserves are depleted or userspace cannot respond. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx"> email@xxxxxxxxx </a>