Messages in this thread | | | Date | Tue, 14 Jul 2015 14:31:03 -0700 | From | "Paul E. McKenney" <> | Subject | Re: mx6: suspicious RCU usage |
| |
On Tue, Jul 14, 2015 at 09:49:09PM +0100, Russell King - ARM Linux wrote: > On Tue, Jul 14, 2015 at 09:26:00AM -0700, Paul E. McKenney wrote: > > On Tue, Jul 14, 2015 at 01:06:32PM -0300, Fabio Estevam wrote: > > > Hi Paul, > > > > > > On Tue, Jul 14, 2015 at 1:00 PM, Paul E. McKenney > > > <paulmck@linux.vnet.ibm.com> wrote: > > > > > > > Does this patch help? > > > > > > > > https://lkml.org/lkml/2015/5/12/885 > > > > > > I am using an ARM 32-bit machine, so I used this one instead: > > > https://lkml.org/lkml/2015/5/12/892 > > > > > > , and it fixes the problem. Thanks! > > > > > > Feel free to add: > > > > > > Tested-by: Fabio Estevam <fabio.estevam@freescale.com> > > > > Glad it helped! > > > > Russell, did you want me to push this, or would you rather take it? > > If it's the one I'm thinking of (using the generic code) because it > doesn't actually solve the problem we have. It may shut up the RCU > warning, but it doesn't solve the underlying problem - one which is > caused by the use of atomics (which use the exclusive instructions) > vs cache line migration between CPUs vs speculative prefetching... > > It's possible right now that _dirty_ cache lines can be migrated to > the dying CPU, which are then lost on power down - and if we disable > the caches on the dying CPU, we then can't use exclusives, so atomics > (and all of the other normal kernel synchronisation mechanisms) are > out of the question.
OK, that could explain any number of failures.
I will leave this issue in your hands, then.
Thanx, Paul
| |