[PATCH v5 3/8] arm: fixmap: implement __set_fixmap()

[Kees Cook]

On Tue, Sep 9, 2014 at 5:38 AM, Will Deacon <will.deacon at arm.com> wrote:
> On Mon, Sep 08, 2014 at 11:40:43PM +0100, Kees Cook wrote:
>> On Mon, Sep 8, 2014 at 2:55 PM, Rabin Vincent <rabin at rab.in> wrote:
>> > On Mon, Sep 08, 2014 at 12:16:34PM -0700, Kees Cook wrote:
>> >> On Thu, Sep 04, 2014 at 06:27:48PM +0100, Will Deacon wrote:
>> >> > On Thu, Sep 04, 2014 at 06:23:42PM +0100, Kees Cook wrote:
>> >> > > Ah! If this is the case, perhaps we can get away with
>> >> > > local_flush_tlb_kernel_range() then?
>> >> >
>> >> > That's a bit tricky, since you need to ensure that preemption is disabled
>> >> > until the mapping is put back like it was.
>> >>
>> >> Okay, under both real hardware with the errata, and under QEMU, things seem
>> >> to work with this change to the series. What do you think?
>> >
>> > Preemption is already disabled until the mapping is put back in this
>> > patch.c code because interrupts are disabled from before the time
>> > set_fixmap() is called until after clear_fixmap() is called.
>>
>> Should I drop the preempt_disable/enable(), and just add a comment to
>> set_fixmap()?
>>
>> > I'd guess that Will meant other (future) callers of set_fixmap() would
>> > have to ensure similar behaviour with set_fixmap() / clear_fixmap().
>> >
>> > Unless I'm missing something set/clear_fixmap() seem to be quite arch
>> > specific and only really used on x86, so we could ensure that future
>> > users on ARM perform the correct tlb flush:  the first user on ARM with
>> > a non-atomic context (or you) could implement a set_fixmap() which does
>> > the global flush and have this patch.c (and any other atomic context
>> > callers) call __set_fixmap() directly.
>> >
>> > The change to local_flush_tlb_kernel_range() in __set_fixmap() would of
>> > course be needed in that case, and IIRC that was what my original patch
>> > had (via set_top_pte()).
>>
>> Ah, so it was, yes! Will, which version of this logic would you prefer?
>
> I still don't think we're solving the general problem here -- we're actually
> just making the ftrace case work. It is perfectly possible for another CPU
> to undergo a TLB miss and refill whilst the page table is being modified by
> the CPU with preemption disabled. In this case, a local tlb flush won't
> invalidate that entry on the other core, and we have no way of knowing when
> the original permissions are actually observed across the system.

The fixmap is used by anything doing patching _except_ ftrace,
actually. It's used by jump labels, kprobes, and kgdb. This code is
the general case. Access to set_fixmap is done via the kernel patching
interface: patch_text().

Right now, the patch_text interface checks cache_ops_need_broadcast(),
and conditionally runs under stop_machine(). We could make this
unconditional, and we'll avoid any problem with TLB misses on another
CPU.

> So I think we need to figure out a way to invalidate the TLB properly. What
> do architectures that use IPIs for TLB broadcasting do (x86, some powerpc,
> mips, ...)? They must have exactly the same problem.

I don't think this should be done at the set_fixmap level, as it is
more a primitive. I think making sure patch_text() always works would
be best. What do you think of using an unconditional stop_machine()
instead?

-Kees

-- 
Kees Cook
Chrome OS Security