Messages in this thread | | | Date | Mon, 11 Aug 2008 17:42:21 -0700 | From | "H. Peter Anvin" <> | Subject | Re: Kernel oops with 2.6.26, padlock and ipsec: probably problem with fpu state changes |
| |
Herbert Xu wrote: > On Mon, Aug 11, 2008 at 01:19:01PM -0700, Suresh Siddha wrote: > . >> we can't unconditionally do clts() in the process context. We have >> to disable pre-emption to avoid interactions with context switch and >> lazy restore. So there will be RT latency issues aswell. > > Yes disabling preemption is the real killer. > > This is just a quick band-aid. Longer term we should add a task > flag that indicates the task is currently doing kernel FPU which > will tell the scheduler to clear TS the next time it's run. That > way we won't need to disable preemtion or pollute the user task's > FPU used state.
That's not sufficient, though, because you have to track all the state and how it relates to everything. You now have to track both the userspace FPU state and the potential kernel FPU state. The VIA instructions are special (in the short bus to school sense) in that they use a mechanism intended to protect specific state to protect -- exactly nothing.
-hpa
| |