On Wed, Mar 8, 2023 at 2:20 AM Juri Lelli <juri.lelli@xxxxxxxxxx> wrote: > > On 01/03/23 17:03, Qais Yousef wrote: > > On 03/01/23 15:26, Juri Lelli wrote: <...> > > > BTW, do you have a repro script of some sort handy I might play with? > > > > Sorry no. You'll just need to suspend to ram. I had a simple patch to measure > > the time around the call and trace_printk'ed the result. > > > > I was working on a android phone which just suspends to ram if you turn the > > screen off and disconnect the usb. > > Looks like I could come up with the following > > https://github.com/jlelli/linux.git deadline/rework-cpusets > https://github.com/jlelli/linux/tree/deadline/rework-cpusets > > which I don't think it's at a point that I feel comfortable to propose > as an RFC (not even sure if it actually makes sense), but it survived my > very light testing. > > Could you please take a look and, if it makes some sense in theory, give > it a try on your end? > Hi Juri, Thanks for coming up with the RFC. I can test your changes in the server environment. I observed the same issue on my side and I can reproduce. I sync'ed up with Qais offline earlier yesterday, and was preparing a couple of patches that optimize the cpuset.cpus writes. Tracking dl tasks in cpusets is one of them. But I am happy to take your patches and do the testing. Note that I won't be able to test the dl part of the patch, only the latency impact on rebuild_root_domains(), as we don't have dl tasks in our system. The other patch is fixing cpuset_rwsem. I see you switched it back to mutex. I did observe performance issues with cpuset_rwsem. Basically, using percpu_rwsem generates very very long latency tails for writers, but mutex doesn't. After some debugging, I found it was because percpu_rwsem requires every writer to call a synchronize_rcu() for acquiring the lock. So in my patches, I disabled the fastpath of readers for cpuset_rwsem. This has been done before[1]. But mutex also worked. Anyway, I'm happy to test your patches and ack once they are sent out. [1] https://lore.kernel.org/all/20160811165413.GA22807@xxxxxxxxxx/T/#m13ef9a728c89cccc9c02af3c96924096c76162a5 Hao
