Re: [PATCH 05/11] x86/fpu: set PKRU state for kernel threads

[Andy Lutomirski <luto@amacapital.net>]

On Thu, Oct 18, 2018 at 11:25 AM Sebastian Andrzej Siewior
<bigeasy@linutronix.de> wrote:
>
> On 2018-10-18 09:48:24 [-0700], Andy Lutomirski wrote:
> > > On Oct 18, 2018, at 9:26 AM, Sebastian Andrzej Siewior <bigeasy@linutronix.de> wrote:
> > >> On 2018-10-12 11:02:18 [-0700], Andy Lutomirski wrote:
> > >> On Fri, Oct 12, 2018 at 10:54 AM Dave Hansen
> > >>> So I'm kinda missing the point of the patch.
> > >>
> > >> use_mm().
> > >
> > > So. I would drop that patch from queue. Anyone feels different about it?
> > >
> >
> > I think we *do* want the patch. It’s a bugfix for use_mm users, right?
>
> This is the loophole that has been pointed out. I am not convinced what
> the correct behaviour should be here (and we have five users of that
> interface). For instance f_fs[0].  It reads data from the USB EP and
> then writes it to userland task. Due to $circumstances it happens in a
> workqueue instead of the task's context.  So it borrows the mm with
> use_mm().  The current behaviour random because the PKRU value can not
> be predicted. It may or may not work.
>
> Setting it to allow-all/none would let the operation always fail or
> succeed which might be an improvement in terms of debugging. However it
> is hard to judge what the correct behaviour should be. Should fail or
> succeed.
> But this is not the only loophole: There is ptrace interface which is
> used by gdb (just checked) and also bypasses PKRU. So…
>
> [0] drivers/usb/gadget/function/f_fs.c::ffs_user_copy_worker()
>
> Sebastian

I think we need an entirely new API:

user_mm_ctx_t ctx = user_mm_ctx_get();

...

use_user_mm_ctx(ctx);
unuse_user_mm_ctx(ctx);

...

user_mm_ctx_put(ctx);

and ctx will store a copy of mm and PKRU.