On Fri, Aug 09, 2019 at 11:49:49AM +0200, Thomas Gleixner wrote: > On Thu, 8 Aug 2019, Fenghua Yu wrote: > > On Thu, Aug 08, 2019 at 10:32:38PM +0200, Thomas Gleixner wrote: > > > Valdis, > > > > > > On Thu, 8 Aug 2019, Valdis Klētnieks wrote: > > > > On Thu, 08 Aug 2019 22:04:03 +0200, Thomas Gleixner said: It isn't > > > > clear that whatever is doing the device_initcall()s will be able to > > > > do any reasonable recovery if we return an error, so any error > > > > recovery is going to have to happen before the function returns. It > > > > might make sense to do an 'if (ret) return;' before going further in > > > > the function, but given the comment a few lines further down about > > > > ignoring errors, it was apparently considered more important to > > > > struggle through and register stuff in sysfs even if umwait was > > > > broken.... > > > > > > I missed that when going through it. > > > > > > The right thing to do is to have a cpu_offline callback which clears > > > the umwait MSR. That covers also the failure in the cpu hotplug setup. > > > Then handling an error return makes sense and keeps everything in a > > > workable shape. > > > > When cpu is offline, the MSR won't be used. We don't need to clear it, > > right? > > Groan. If soemthing goes wrong when registering the hotplug callback, what > undoes the MSR setup which might have happened and what takes care of it on > cpus coming online later? Exactly nothing. Then you have a non-consistent > behaviour. > > Make stuff symmmetric and correct and not optimized for the sunshine case.
I see. Just want to make sure I understand it correctly: sysfs_create_group() should not be called if cpuhp_setup_state() has error. Otherwise, the sysadmin can change the MSR through the sysfs interface. After that, a CPU is online and its MSR is not updated because cpu_online is not installed. Then this online CPU has different MSR value from the other CPUs. Is that right? Thanks. -Fenghua
