Re: Let's talk about the elephant in the room - the Linux kernel's inability to gracefully handle low memory pressure

[Johannes Weiner <hannes@cmpxchg.org>]

On Tue, Aug 06, 2019 at 09:27:05AM -0700, Suren Baghdasaryan wrote:
> On Tue, Aug 6, 2019 at 7:36 AM Michal Hocko <mhocko@kernel.org> wrote:
> >
> > On Tue 06-08-19 10:27:28, Johannes Weiner wrote:
> > > On Tue, Aug 06, 2019 at 11:36:48AM +0200, Vlastimil Babka wrote:
> > > > On 8/6/19 3:08 AM, Suren Baghdasaryan wrote:
> > > > >> @@ -1280,3 +1285,50 @@ static int __init psi_proc_init(void)
> > > > >>         return 0;
> > > > >>  }
> > > > >>  module_init(psi_proc_init);
> > > > >> +
> > > > >> +#define OOM_PRESSURE_LEVEL     80
> > > > >> +#define OOM_PRESSURE_PERIOD    (10 * NSEC_PER_SEC)
> > > > >
> > > > > 80% of the last 10 seconds spent in full stall would definitely be a
> > > > > problem. If the system was already low on memory (which it probably
> > > > > is, or we would not be reclaiming so hard and registering such a big
> > > > > stall) then oom-killer would probably kill something before 8 seconds
> > > > > are passed.
> > > >
> > > > If oom killer can act faster, than great! On small embedded systems you probably
> > > > don't enable PSI anyway?
> 
> We use PSI triggers with 1 sec tracking window. PSI averages are less
> useful on such systems because in 10 secs (which is the shortest PSI
> averaging window) memory conditions can change drastically.
> 
> > > > > If my line of thinking is correct, then do we really
> > > > > benefit from such additional protection mechanism? I might be wrong
> > > > > here because my experience is limited to embedded systems with
> > > > > relatively small amounts of memory.
> > > >
> > > > Well, Artem in his original mail describes a minutes long stall. Things are
> > > > really different on a fast desktop/laptop with SSD. I have experienced this as
> > > > well, ending up performing manual OOM by alt-sysrq-f (then I put more RAM than
> > > > 8GB in the laptop). IMHO the default limit should be set so that the user
> > > > doesn't do that manual OOM (or hard reboot) before the mechanism kicks in. 10
> > > > seconds should be fine.
> > >
> > > That's exactly what I have experienced in the past, and this was also
> > > the consistent story in the bug reports we have had.
> > >
> > > I suspect it requires a certain combination of RAM size, CPU speed,
> > > and IO capacity: the OOM killer kicks in when reclaim fails, which
> > > happens when all scanned LRU pages were locked and under IO. So IO
> > > needs to be slow enough, or RAM small enough, that the CPU can scan
> > > all LRU pages while they are temporarily unreclaimable (page lock).
> > >
> > > It may well be that on phones the RAM is small enough relative to CPU
> > > size.
> > >
> > > But on desktops/servers, we frequently see that there is a wider
> > > window of memory consumption in which reclaim efficiency doesn't drop
> > > low enough for the OOM killer to kick in. In the time it takes the CPU
> > > to scan through RAM, enough pages will have *just* finished reading
> > > for reclaim to free them again and continue to make "progress".
> > >
> > > We do know that the OOM killer might not kick in for at least 20-25
> > > minutes while the system is entirely unresponsive. People usually
> > > don't wait this long before forcibly rebooting. In a managed fleet,
> > > ssh heartbeat tests eventually fail and force a reboot.
> 
> Got it. Thanks for the explanation.
> 
> > > I'm not sure 10s is the perfect value here, but I do think the kernel
> > > should try to get out of such a state, where interacting with the
> > > system is impossible, within a reasonable amount of time.
> > >
> > > It could be a little too short for non-interactive number-crunching
> > > systems...
> >
> > Would it be possible to have a module with tunning knobs as parameters
> > and hook into the PSI infrastructure? People can play with the setting
> > to their need, we wouldn't really have think about the user visible API
> > for the tuning and this could be easily adopted as an opt-in mechanism
> > without a risk of regressions.

It's relatively easy to trigger a livelock that disables the entire
system for good, as a regular user. It's a little weird to make the
bug fix for that an opt-in with an extensive configuration interface.

This isn't like the hung task watch dog, where it's likely some kind
of kernel issue, right? This can happen on any current kernel.

What I would like to have is a way of self-recovery from a livelock. I
don't mind making it opt-out in case we make mistakes, but the kernel
should provide minimal self-protection out of the box, IMO.

> PSI averages stalls over 10, 60 and 300 seconds, so implementing 3
> corresponding thresholds would be easy. The patch Johannes posted can
> be extended to support 3 thresholds instead of 1. I can take a stab at
> it if Johannes is busy.
> If we want more flexibility we could use PSI triggers with
> configurable tracking window but that's more complex and probably not
> worth it.

This goes into quality-of-service for workloads territory again. I'm
not quite convinced yet we want to go there.