On 08/25, Linus Torvalds wrote: > > On Mon, Aug 25, 2014 at 7:41 AM, Oleg Nesterov <o...@redhat.com> wrote: > > > > I think this should be safe, because this thread and/or swapper/0 can > > do nothing with with fpu->state, and they should not use fpu. > > .. but if that's the case, then what was wrong with the old code
Confused... Just in case, I think that you mean current code, and ignoring the lack of preempt_disable() around math_state_restore() it is correct. I'd like to change it only because this code is the main source of the nasty special case, used_math() and/or __thread_has_fpu(current) can be false even if use_eager_fpu(). > that > just copied the state over the unused space from the user space > buffer? But it is not unused? Although I probably misunderstood you from the very beginning. OK, what I meant that without switch_fpu_xstate(init_task.fpu.state) or another hack we can't avoid drop_fpu() which leads to this special case. Currently __copy_from_user(&xstate->xsave) copies the new registers right into this thread's fpu->state. If this thread is preempted before math_state_restore(), the context switch (__save_init_fpu) will overwrite the same buffer, the result of __copy_from_user() can be simply lost (entirely or not). With this patch we can safely do __copy_from_user(xstate), this buffer is not used until the 2nd switch_fpu_xstate(). > You can't have it both ways. Either the old code was fine (because it > doesn't use the buffer while it is in flux), or the new code is broken > (because it uses the shared buffer). Your choice.No? It uses the shared buffer, yes. But in this case (I think! please correct me!), when this thread uses the swapper's fpu->state, schedule() -> fpu_xsave() into this shared buffer should be fine because it should write the same content? Oleg. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
