On Tue, Aug 13, 2019 at 07:53:49PM -0700, Paul E. McKenney wrote: > On Wed, Aug 14, 2019 at 09:11:03AM +0900, Byungchul Park wrote: > > On Tue, Aug 13, 2019 at 08:41:45AM -0700, Paul E. McKenney wrote: > > > On Tue, Aug 13, 2019 at 02:29:54PM +0900, Byungchul Park wrote: > > > > On Mon, Aug 12, 2019 at 09:12:34AM -0400, Joel Fernandes wrote: > > > > > On Mon, Aug 12, 2019 at 07:10:52PM +0900, Byungchul Park wrote: > > > > > > On Sun, Aug 11, 2019 at 04:49:39PM -0700, Paul E. McKenney wrote: > > > > > > > Maybe. Note well that I said "potential issue". When I checked a few > > > > > > > years ago, none of the uses of rcu_barrier() cared about kfree_rcu(). > > > > > > > They cared instead about call_rcu() callbacks that accessed code or data > > > > > > > that was going to disappear soon, for example, due to module unload or > > > > > > > filesystem unmount. > > > > > > > > > > > > > > So it -might- be that rcu_barrier() can stay as it is, but with changes > > > > > > > as needed to documentation. > > > > > > > > > > Right, we should update the docs. Byungchul, do you mind sending a patch that > > > > > documents the rcu_barrier() behavior? > > > > > > > > Are you trying to give me the chance? I feel thankful. It doens't matter > > > > to try it at the moment though, I can't follow-up until September. I'd > > > > better do that in Septamber or give it up this time. > > > > > > Which reminds me... I recall your asking if the kfree_rcu() patch > > > might be sensitive to the exact hardware, but I cannot locate that > > > email right off-hand. This is an excellent question! When faced with > > > floods of kfree_rcu() calls, I would expect some hardware, compiler, > > > and kernel-configuration sensitivity. Which is why it will likely be > > > > Yes. > > > > > necessary to do a few more improvements over time -- for but one example, > > > accumulating callbacks into vectors in order to reduce the number of > > > kfree()-time cache misses. > > > > Yes. That would be a pretty good way to mitigate the problem. I hope > > the simple way we've done works well enough so it would never happen > > though. > > > > Or I would check the condition of all system resourses e.g. CPU and > > memory and control the bandwith of them, of course only if that actually > > happens. > > > > Thanks a lot for sharing your opinion on it! > > Didn't you say earlier that you were getting OOM on your system even > with the patches? Or did I miss the resolution of that issue? I said I saw OOM with a *larger* value of KFREE_DRAIN_JIFFIES. It was fine with the patch itself. Anyway I'm sorry I expressed it in a confusing way. Thanks, Byungchul > > Thanx, Paul > > > Thanks, > > Byungchul > > > > > Thanx, Paul > > > > > > > Thanks, > > > > Byungchul > > > > > > > > > > > It also -might- be, maybe now or maybe some time in the future, that > > > > > > > there will need to be a kfree_rcu_barrier() or some such. But if so, > > > > > > > let's not create it until it is needed. For one thing, it is reasonably > > > > > > > likely that something other than a kfree_rcu_barrier() would really > > > > > > > be what was needed. After all, the main point would be to make sure > > > > > > > that the old memory really was freed before allocating new memory. > > > > > > > > > > > > Now I fully understand what you meant thanks to you. Thank you for > > > > > > explaining it in detail. > > > > > > > > > > > > > But if the system had ample memory, why wait? In that case you don't > > > > > > > really need to wait for all the old memory to be freed, but rather for > > > > > > > sufficient memory to be available for allocation. > > > > > > > > > > > > Agree. Totally make sense. > > > > > > > > > > Agreed, all makes sense. > > > > > > > > > > thanks, > > > > > > > > > > - Joel > > > > > > > > > > [snip] > > > > > >