On Thu, May 27 2021 at 16:51, Dave Hansen wrote:

Andy Lutomirski recently noted a probable bug in write_pkru(), but
it was unclear if it was user-visible.  A recent bug report in
related code[1] forced me to take a look.

Basically, manipulation of XSAVE state is too unstructured.
get_xsave_addr() gives callers the impression they can read and
write XSAVE state when there are a lot of pitfalls, like updates
to xstate.features bits.

As a result, more than one call site screws up the modification
of PKRU in the XSAVE buffer.  This series fixes that problem up
and also hopefully carves out a less error-prone path that can
be reused for other XSAVE features.

This series:
 * Moves the PKRU manipulation to a more appropriate location,
   away from the page table code
 * Wraps get_xsave_addr() with more structured, less error-prone
   interfaces.
 * Conditionally hides a pkey debugfs file, eliminating the need
   for new runtime checks to work with the new interface.
 * Add a selftest to make it more likely to catch bugs like this
   in the future.  This improved selftest catches this issue on
   Intel CPUs.  Without the improvement, it only triggers on AMD.

I think all of this is fundamentaly wrong.

Contrary to FPU state, PKRU has to be updated at context switch
time. There is absolutely no point in having PKRU XSAVES managed.

It's broken in several ways. Anything which clears and loads the FPU
will load the wrong PKRU value. Go figure...

So the right thing is to disable PKRU in XCR0 and on sched out simply do

   task->thread.pkru = read_pkru();

and on sched in

   write_pkru(task->thread.pkru);

Simple, trivial and not going to be wreckaged by anything which fiddles
with xstates. We all know by now that xstates is a trainwreck and not
having stuff like that in there is making the fixes I'm doing way
simpler.

CET will have a similar issue, but we'll discuss that once we have the
existing horrors sorted.

Thanks,

        tglx