On Wed 20-04-16 06:55:42, Tetsuo Handa wrote: > Michal Hocko wrote: > > > This patch adds a timeout for handling corner cases where a TIF_MEMDIE > > > thread got stuck. Since the timeout is checked at oom_unkillable_task(), > > > oom_scan_process_thread() will not find TIF_MEMDIE thread > > > (for !oom_kill_allocating_task case) and oom_badness() will return 0 > > > (for oom_kill_allocating_task case). > > > > > > By applying this patch, the kernel will automatically press SysRq-f if > > > the OOM reaper cannot reap the victim's memory, and we will never OOM > > > livelock forever as long as the OOM killer is called. > > > > Which will not guarantee anything as already pointed out several times > > before. So I think this is not really that useful. I have said it > > earlier and will repeat it again. Any timeout based solution which > > doesn't guarantee that the system will be in a consistent state (reboot, > > panic or kill all existing tasks) after the specified timeout is > > pointless. > > Triggering the reboot/panic is the worst action. Killing all existing tasks > is the next worst action. Thus, I prefer killing tasks one by one. killing a task by task doesn't guarantee any convergence to a usable state. If somebody really cares about these highly unlikely lockups I am pretty sure he would really appreciate to have a _reliable_ and _guaranteed_ way out of that situation. Having a fuzzy mechanism to do something in a good hope of resolving that state is just unhelpful. If I was an admin and had a machine on the other side of the globe and that machine just locked up due to OOM I would pretty much wanted to force reboot as my other means of fixing that situation would be pretty much close to zero otherwise. > I'm OK with shortening the timeout like N (when waiting for the 1st victim) > + N/2 (the 2nd victim) + N/4 (the 3rd victim) + N/8 (the 4th victim) + ... > but does it worth complicating the least unlikely path? No it is not IMHO. > > I believe that the chances of the lockup are much less likely with the > > oom reaper and that we are not really urged to provide a new knob with a > > random semantic. If we really want to have a timeout based thing better > > make it behave reliably. > > The threshold which the administrator can wait for ranges. Some may want to > set few seconds because of 10 seconds /dev/watchdog timeout, others may want > to set one minute because of not using watchdog. Thus, I think we should not > hard code the timeout. I guess you missed my point here. I didn't say this should be hardcoded in any way. I am just saying that if we really want to do some timeout based decisions we should better think about the semantic and that should provide a reliable and deterministic means to resolve the problem. -- Michal Hocko SUSE Labs -- 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>