On Fri, Jun 29, 2018 at 11:04:19AM +0200, Michal Hocko wrote: > On Thu 28-06-18 14:31:05, Paul E. McKenney wrote: > > On Thu, Jun 28, 2018 at 01:39:42PM +0200, Michal Hocko wrote: > > > On Wed 27-06-18 07:31:25, Paul E. McKenney wrote: > > > > On Wed, Jun 27, 2018 at 09:22:07AM +0200, Michal Hocko wrote: > > > > > On Tue 26-06-18 10:03:45, Paul E. McKenney wrote: > > > > > [...] > > > > > > 3. Something else? > > > > > > > > > > How hard it would be to use a different API than oom notifiers? E.g. a > > > > > shrinker which just kicks all the pending callbacks if the reclaim > > > > > priority reaches low values (e.g. 0)? > > > > > > > > Beats me. What is a shrinker? ;-) > > > > > > This is a generich mechanism to reclaim memory that is not on standard > > > LRU lists. Lwn.net surely has some nice coverage (e.g. > > > https://lwn.net/Articles/548092/). > > > > "In addition, there is little agreement over what a call to a shrinker > > really means or how the called subsystem should respond." ;-) > > > > Is this set up using register_shrinker() in mm/vmscan.c? I am guessing > > Yes, exactly. You are supposed to implement the two methods in struct > shrink_control > > > that the many mentions of shrinker in DRM are irrelevant. > > > > If my guess is correct, the API seems a poor fit for RCU. I can > > produce an approximate number of RCU callbacks for ->count_objects(), > > but a given callback might free a lot of memory or none at all. Plus, > > to actually have ->scan_objects() free them before returning, I would > > need to use something like rcu_barrier(), which might involve longer > > delays than desired.` > > Well, I am not yet sure how good fit this is because I still do not > understand the underlying problem your notifier is trying to solve. So I > will get back to this once that is settled. > > > > Or am I missing something here? > > > > > > More seriously, could you please point me at an exemplary shrinker > > > > use case so I can see what is involved? > > > > > > Well, I am not really sure what is the objective of the oom notifier to > > > point you to the right direction. IIUC you just want to kick callbacks > > > to be handled sooner under a heavy memory pressure, right? How is that > > > achieved? Kick a worker? > > > > That is achieved by enqueuing a non-lazy callback on each CPU's callback > > list, but only for those CPUs having non-empty lists. This causes > > CPUs with lists containing only lazy callbacks to be more aggressive, > > in particular, it prevents such CPUs from hanging out idle for seconds > > at a time while they have callbacks on their lists. > > > > The enqueuing happens via an IPI to the CPU in question. > > I am afraid this is too low level for my to understand what is going on > here. What are lazy callbacks and why do they need any specific action > when we are getting close to OOM? I mean, I do understand that we might > have many callers of call_rcu and free memory lazily. But there is quite > a long way before we start the reclaim until we reach the OOM killer path. > So why don't those callbacks get called during that time period? How are > their triggered when we are not hitting the OOM path? They surely cannot > sit there for ever, right? Can we trigger them sooner? Maybe the > shrinker is not the best fit but we have a retry feedback loop in the page > allocator, maybe we can kick this processing from there. The effect of RCU's current OOM code is to speed up callback invocation by at most a few seconds (assuming no stalled CPUs, in which case it is not possible to speed up callback invocation). Given that, I should just remove RCU's OOM code entirely? Thanx, Paul
