On 22.07.2019 14:06, Andrew Cooper wrote: > Does reverting back to credit1 make the issue go away? I've never > encountered this on any smt=0 test, but I also don't use credit2 at all.
I'll try to remember trying this out the next time I see it. I can't see a connection to the used scheduler though, when the message comes out of __cpu_die(). There must be an excessive delay for the dying CPU to finally call cpu_exit_clear(). I wonder if the CPU might e.g. be scrubbing memory at that point. But that shouldn't happen this early. > The sibling threads shouldn't be inserted into the scheduler in the > first place, and I thought we took deliberate steps to prevent that from > occurring. I don't think we did, but I agree this may be worthwhile to do if it wouldn't result in adding ugly special cases somewhere. Jan _______________________________________________ Xen-devel mailing list Xen-devel@lists.xenproject.org https://lists.xenproject.org/mailman/listinfo/xen-devel
