On Thu, Oct 12, 2017 at 05:05:07PM +0100, Dave P Martin wrote:
> On Wed, Oct 11, 2017 at 05:15:58PM +0100, Catalin Marinas wrote:
> > On Tue, Oct 10, 2017 at 07:38:28PM +0100, Dave P Martin wrote:
> > > diff --git a/arch/arm64/include/asm/processor.h b/arch/arm64/include/asm/processor.h
> > > index 29adab8..4831d28 100644
> > > --- a/arch/arm64/include/asm/processor.h
> > > +++ b/arch/arm64/include/asm/processor.h
> > > @@ -39,6 +47,8 @@
> > > #define FPEXC_IDF (1 << 7)
> > >
> > > /*
> > > + * (Note: in this discussion, statements about FPSIMD apply equally to SVE.)
> > > + *
> > > * In order to reduce the number of times the FPSIMD state is needlessly saved
> > > * and restored, we need to keep track of two things:
> > > * (a) for each task, we need to remember which CPU was the last one to have
> > > @@ -99,6 +109,287 @@
> > > */
> > > static DEFINE_PER_CPU(struct fpsimd_state *, fpsimd_last_state);
> > >
> > > +static void sve_free(struct task_struct *task)
> > > +{
> > > + kfree(task->thread.sve_state);
> > > + task->thread.sve_state = NULL;
> > > +}
> >
> > I think we need a WARN_ON if TIF_SVE is still set here (and the callers
> > making sure it is cleared). I haven't checked the code paths via
> > fpsimd_release_thread() but wondering what happens if we get an
> > interrupt between freeing the state and making the pointer NULL, with
> > some context switching in a preemptible kernel.
>
> Having a WARN_ON() here may be a decent way to sanity-check that we
> don't ever have sve_state NULL with TIF_SVE set. This is a lot more
> economical than putting a WARN_ON() at each dereference of sve_state
> (of which there are quite a few). sve_free() is also a slow path.
>
> Currently, there are two callsites: sve_set_vector_length(), where we
> test_and_clear_tsk_thread_flags(task, TIF_SVE) before calling sve_free();
> and fpsimd_release_thread() where we "don't care" because the thread
> is dying.
>
> Looking more closely though, is the release_thread() path preemptible?
> I can't see anything in the scheduler core to ensure this, nor any
> general reason why it should be needed.
>
> In which case preemption during thread exit after sve_free() could
> result in a NULL deference in fpsimd_thread_switch().
>
>
> So, I think my favoured approach is:
>
> sve_release_thread()
> {
> local_bh_disable();
> fpsimd_flush_task_state(current);
> clear_thread_flag(TIF_SVE);
> local_bh_enable();
>
> sve_free();
> }
>
> The local_bh stuff is cumbersome here, and could be replaced with
> barrier()s to force the order of fpsimd_flusk_task_state() versus
> clearing TIF_SVE. Or should the barrier really be in
> fpsimd_flush_task_state()? Disabling softirqs avoids the need to answer
> such questions...
>
>
> Then:
>
> sve_free(task)
> {
> WARN_ON(test_thread_flag(TIF_SVE));
>
> barrier();
> kfree(task->thread.sve_state);
> tash->thread.sve_state = NULL;
> }
>
> I'm assuming here that kfree() can't be called safely from atomic
> context, but this is unclear. I would expect to be able to free
> GFP_ATOMIC memory from atomic context (though sve_statue is GFP_KERNEL,
> so dunno).
The kfree should be fine.
Alternative proposal: free the SVE state in arch_release_task_struct().
This is called via the RCU mechanism and the task is no longer current,
so no preemption issues.
> > Alternatively, always clear TIF_SVE here before freeing (also wondering
> > whether we should make sve_state NULL before the actual freeing but I
> > think TIF_SVE clearing should suffice).
>
> Could do. I feel that the current placement of the TIF_SVE clearing in
> sve_set_vector_length() feels "more natural", but this is a pretty
> flimsy argument. How strongly do you feel about this?
I agree with you, keep the TIF_SVE clearing in sve_set_vector_length().
--
Catalin
