Re: [patch 7/9] x86/iopl: Restrict iopl() permission scope

From: Thomas Gleixner
Date: Sun Nov 10 2019 - 16:22:11 EST


On Sun, 10 Nov 2019, Andy Lutomirski wrote:
> On Sun, Nov 10, 2019 at 12:31 PM Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> >
> > On Sun, 10 Nov 2019, Andy Lutomirski wrote:
> > > On 11/6/19 11:35 AM, Thomas Gleixner wrote:
> > > > +
> > > > + if (IS_ENABLED(CONFIG_X86_IOPL_EMULATION)) {
> > > > + struct tss_struct *tss;
> > > > + unsigned int tss_base;
> > > > +
> > > > + /* Prevent racing against a task switch */
> > > > + preempt_disable();
> > > > + tss = this_cpu_ptr(&cpu_tss_rw);
> > > > + if (level == 3) {
> > > > + /* Grant access to all I/O ports */
> > > > + set_thread_flag(TIF_IO_BITMAP);
> > > > + tss_base = IO_BITMAP_OFFSET_VALID_ALL;
> > >
> > > Where is the actual TSS updated?
> >
> > Here. It sets the offset to the all zero bitmap. That's all it needs.
>
> Ah, I see.
>
> > > I think what you need to do is have a single function, called by
> > > exit_thread(), switch_to(), and here, that updates the TSS to match a
> > > given task's IO bitmap state. This is probably switch_to_bitmap() or
> > > similar.
> >
> > Well, no. exit_thread() and this here actually fiddle with the TIF bit
> > which is not what the switch to case does. There is some stuff which can be
> > shared.
>
> I was thinking that the code that read iopl_emul and t->io_bitmap_ptr
> and updated x86_tss.io_bitmap_base could be factored out of
> switch_to(). Suppose you call that tss_update_io_bitmap(). And you
> could have another tiny helper:
>
> static void update_tif_io_bitmap(void)
> {
> if (...->iopl_emul || ...->io_bitmap_ptr)
> set_thread_flag(TIF_IO_BITMAP);
> else
> clear_thread_flag(TIF_IO_BITMAP);
> }
>
> Then the whole iopl emulation path becomes:

Yes. I'm almost done with that already :)
> > It's in that very same patch:
> >
> > > -static void tss_update_io_bitmap(struct tss_struct *tss,
> > > - struct thread_struct *thread)
>
> But where did the line you just deleted come from? I'm obviously just
> bad at reading emails somewhere.

patch 5/9 :)

Let me finish that stuff and send out another version which has that reuse
of the switch to code in a separate patch.

There is another new detail in the series. I split out all the io bitmap
data into a separate data struct, which also allows me to stick a refcount
into it which removes the kmemdup() from the fork path.

Thanks,

tglx