On Tue, 5 May 2020 11:27:20 -0700 "Paul E. McKenney" <paulmck@xxxxxxxxxx> wrote: > On Tue, May 05, 2020 at 07:49:43PM +0200, SeongJae Park wrote: > > On Tue, 5 May 2020 10:23:58 -0700 "Paul E. McKenney" <paulmck@xxxxxxxxxx> wrote: > > > > > On Tue, May 05, 2020 at 09:25:06AM -0700, Eric Dumazet wrote: > > > > > > > > > > > > On 5/5/20 9:13 AM, SeongJae Park wrote: > > > > > On Tue, 5 May 2020 09:00:44 -0700 Eric Dumazet <edumazet@xxxxxxxxxx> wrote: > > > > > > > > > >> On Tue, May 5, 2020 at 8:47 AM SeongJae Park <sjpark@xxxxxxxxxx> wrote: > > > > >>> > > > > >>> On Tue, 5 May 2020 08:20:50 -0700 Eric Dumazet <eric.dumazet@xxxxxxxxx> wrote: > > > > >>> > > > > >>>> > > > > >>>> > > > > >>>> On 5/5/20 8:07 AM, SeongJae Park wrote: > > > > >>>>> On Tue, 5 May 2020 07:53:39 -0700 Eric Dumazet <edumazet@xxxxxxxxxx> wrote: > > > > >>>>> > > > > >>>> > > > > >>>>>> Why do we have 10,000,000 objects around ? Could this be because of > > > > >>>>>> some RCU problem ? > > > > >>>>> > > > > >>>>> Mainly because of a long RCU grace period, as you guess. I have no idea how > > > > >>>>> the grace period became so long in this case. > > > > >>>>> > > > > >>>>> As my test machine was a virtual machine instance, I guess RCU readers > > > > >>>>> preemption[1] like problem might affected this. > > > > >>>>> > > > > >>>>> [1] https://www.usenix.org/system/files/conference/atc17/atc17-prasad.pdf > > > > > > If this is the root cause of the problem, then it will be necessary to > > > provide a hint to the hypervisor. Or, in the near term, avoid loading > > > the hypervisor the point that vCPU preemption is so lengthy. > > > > > > RCU could also provide some sort of pre-stall-warning notification that > > > some of the CPUs aren't passing through quiescent states, which might > > > allow the guest OS's userspace to take corrective action. > > > > > > But first, what are you doing to either confirm or invalidate the > > > hypothesis that this might be due to vCPU preemption? > > > > Nothing, I was just guessing. Sorry if this made you confused. > > > > > > > > > >>>>>> Once Al patches reverted, do you have 10,000,000 sock_alloc around ? > > > > >>>>> > > > > >>>>> Yes, both the old kernel that prior to Al's patches and the recent kernel > > > > >>>>> reverting the Al's patches didn't reproduce the problem. > > > > >>>>> > > > > >>>> > > > > >>>> I repeat my question : Do you have 10,000,000 (smaller) objects kept in slab caches ? > > > > >>>> > > > > >>>> TCP sockets use the (very complex, error prone) SLAB_TYPESAFE_BY_RCU, but not the struct socket_wq > > > > >>>> object that was allocated in sock_alloc_inode() before Al patches. > > > > >>>> > > > > >>>> These objects should be visible in kmalloc-64 kmem cache. > > > > >>> > > > > >>> Not exactly the 10,000,000, as it is only the possible highest number, but I > > > > >>> was able to observe clear exponential increase of the number of the objects > > > > >>> using slabtop. Before the start of the problematic workload, the number of > > > > >>> objects of 'kmalloc-64' was 5760, but I was able to observe the number increase > > > > >>> to 1,136,576. > > > > >>> > > > > >>> OBJS ACTIVE USE OBJ SIZE SLABS OBJ/SLAB CACHE SIZE NAME > > > > >>> before: 5760 5088 88% 0.06K 90 64 360K kmalloc-64 > > > > >>> after: 1136576 1136576 100% 0.06K 17759 64 71036K kmalloc-64 > > > > >>> > > > > >> > > > > >> Great, thanks. > > > > >> > > > > >> How recent is the kernel you are running for your experiment ? > > > > > > > > > > It's based on 5.4.35. > > > > > > Is it possible to retest on v5.6? I have been adding various mechanisms > > > to make RCU keep up better with heavy callback overload. > > > > I will try soon! > > > > > > > > Also, could you please provide the .config? If either NO_HZ_FULL or > > > RCU_NOCB_CPU, please also provide the kernel boot parameters. > > > > NO_HZ_FULL is not set, but RCU_NOCB_CPU is y. > > OK, this is important information. > > > I think I should check whether it's ok to share the full config and boot > > parameters. Please wait this. > > I probably don't need the whole thing. So, if it makes it easier to > gain approval... > > The main thing I need are CONFIG_PREEMPT and the various Kconfig options > having "RCU" in their names. For example, I have no need for any of the > options pertaining to device drivers. (As far as I know at the moment, > anyway!) > > For the boot parameters, I am very interested in rcu_nocbs=. Along with > any other boot parameters whose names contain "rcu". I guess this would be ok. It uses no 'rcu_nocbs=' boot parameter. The configs you asked are as below: # CONFIG_PREEMPT is not set # # RCU Subsystem # CONFIG_TREE_RCU=y CONFIG_RCU_EXPERT=y CONFIG_SRCU=y CONFIG_TREE_SRCU=y CONFIG_RCU_STALL_COMMON=y CONFIG_RCU_NEED_SEGCBLIST=y CONFIG_RCU_FANOUT=64 CONFIG_RCU_FANOUT_LEAF=16 # CONFIG_RCU_FAST_NO_HZ is not set CONFIG_RCU_NOCB_CPU=y # end of RCU Subsystem > > If rcu_nocbs does designate have any CPUs listed, another thing to check > is where the rcuo kthreads are permitted to run. The reason that this > is important is that any CPU listed in the rcu_nocbs= boot parameter > has its RCU callbacks invoked by one of the rcuo kthreads. If you have > booted with (say) "rcu_nocbs=1,63" and then bound all of the resulting > rcuo kthreads to CPU 0, you just tied RCU's hands, making it unable to > keep up with any reasonable RCU callback load. > > This sort of configuration is permitted, but it is intended for tightly > controlled real-time or HPC systems whose configurations and workloads > avoid tossing out large numbers of callbacks. Which might not be the > case for your workload. > > > > > >> Let's make sure the bug is not in RCU. > > > > > > > > > > One thing I can currently say is that the grace period passes at last. I > > > > > modified the benchmark to repeat not 10,000 times but only 5,000 times to run > > > > > the test without OOM but easily observable memory pressure. As soon as the > > > > > benchmark finishes, the memory were freed. > > > > > > > > > > If you need more tests, please let me know. > > > > > > > > I would ask Paul opinion on this issue, because we have many objects > > > > being freed after RCU grace periods. > > > > > > As always, "It depends." > > > > > > o If the problem is a too-long RCU reader, RCU is prohibited from > > > ending the grace period. The reader duration must be shortened, > > > and until it is shortened, there is nothing RCU can do. > > > > > > o In some special cases of the above, RCU can and does help, for > > > example, by enlisting the aid of cond_resched(). So perhaps > > > there is a long in-kernel loop that needs a cond_resched(). > > > > > > And perhaps RCU can help for some types of vCPU preemption. > > > > > > o As Al suggested offline and as has been discussed in the past, > > > it would not be hard to cause RCU to burn CPU to attain faster > > > grace periods during OOM events. This could be helpful, but only > > > given that RCU readers are completing in reasonable timeframes. > > > > Totally agreed. > > > > > > If RCU subsystem can not keep-up, I guess other workloads will also suffer. > > > > > > If readers are not excessively long, RCU should be able to keep up. > > > (In the absence of misconfigurations, for example, both NO_HZ_FULL and > > > then binding all the rcuo kthreads to a single CPU on a 100-CPU system > > > or some such.) > > > > > > > Sure, we can revert patches there and there trying to work around the issue, > > > > but for objects allocated from process context, we should not have these problems. > > > > > > Agreed, let's get more info on what is happening to RCU. > > > > > > One approach is to shorten the RCU CPU stall warning timeout > > > (rcupdate.rcu_cpu_stall_timeout=10 for 10 seconds). > > > > I will also try this and let you know the results. > > Sounds good, thank you! :) Thanks, SeongJae Park
