On Wed, Jan 12, 2022 at 01:28:58AM +0000, Sean Christopherson wrote:
> On Tue, Jan 11, 2022, Peter Zijlstra wrote:
> > On Thu, Jan 06, 2022 at 04:35:23PM +0800, kernel test robot wrote:
> > > 
> > > 
> > > Greeting,
> > > 
> > > FYI, we noticed the following commit (built with clang-14):
> > > 
> > > commit: aa93e2ad7464ffb90155a5ffdde963816f86d5dc ("x86/entry_32: Remove .fixup usage")
> > > https://git.kernel.org/cgit/linux/kernel/git/tip/tip.git x86/core
> > > 
> > > in testcase: kernel-selftests
> > > version: 
> > > with following parameters:
> > > 
> > > 	group: x86
> > > 
> > 
> > It would be very useful if this thing would also say which of the many
> > x86 selftests fails... it appears to be: ldt_gdt_32.
> > 
> > The below fixes it, but I'm still not entirely sure what the actual
> > problem is, although Andy did find a bug in that the exception handler
> > should do: *(ss:esp) = 0, adding ss-base (using insn_get_seg_base())
> > doesn't seem to cure things.
> 
> Because I was curious...
> 
> The issue is that PARANOID_EXIT_TO_KERNEL_MODE in the handle_exception_return
> path overwrites the entry stack data with the task stack data, restoring the "bad"
> segment value.

Oh gawd... that's terrible, and yes, that now makes perfect sense.

However did you find that?