On Mon, Jul 06, 2020 at 05:20:57PM -0400, Dave Jones wrote: > On Mon, Jul 06, 2020 at 04:59:52PM +0200, Peter Zijlstra wrote: > > On Fri, Jul 03, 2020 at 04:51:53PM -0400, Dave Jones wrote: > > > On Fri, Jul 03, 2020 at 12:40:33PM +0200, Peter Zijlstra wrote: > > > > > > looked promising the first few hours, but as soon as it hit four hours > > > of uptime, loadavg spiked and is now pinned to at least 1.00 > > > > OK, lots of cursing later, I now have the below... > > > > The TL;DR is that while schedule() doesn't change p->state once it > > starts, it does read it quite a bit, and ttwu() will actually change it > > to TASK_WAKING. So if ttwu() changes it to WAKING before schedule() > > reads it to do loadavg accounting, things go sideways. > > > > The below is extra complicated by the fact that I've had to scrounge up > > a bunch of load-store ordering without actually adding barriers. It adds > > yet another control dependency to ttwu(), so take that C standard :-) > > Man this stuff is subtle. I could've read this a hundred times and not > even come close to approaching this. > > Basically me reading scheduler code: > http://www.quickmeme.com/img/96/9642ed212bbced00885592b39880ec55218e922245e0637cf94db2e41857d558.jpg
Heh, that one made me nearly spill my tea, much funnies :-) But yes, Dave Chinner also complained about this for the previous fix. I've written this: https://lore.kernel.org/lkml/20200703133259.ge4...@hirez.programming.kicks-ass.net/ to help with that. But clearly I'll need to update that patch again after this little adventure. > > I've booted it, and build a few kernels with it and checked loadavg > > drops to 0 after each build, so from that pov all is well, but since > > I'm not confident I can reproduce the issue, I can't tell this actually > > fixes anything, except maybe phantoms of my imagination. > > Five hours in, looking good so far. I think you nailed it. \o/ hooray! Thanks for testing Dave!
