In the native case, PER_CPU_VAR(cpu_tss_rw + TSS_sp0) is the
trampoline stack. But XEN pv doesn't use trampoline stack, so
PER_CPU_VAR(cpu_tss_rw + TSS_sp0) is also the kernel stack.

In that case, source and destination stacks are identical, which means
that reusing swapgs_restore_regs_and_return_to_usermode() in XEN pv
would cause %rsp to move up to the top of the kernel stack and leave the
IRET frame below %rsp.

This is dangerous as it can be corrupted if #NMI / #MC hit as either of
these events occurring in the middle of the stack pushing would clobber
data on the (original) stack.

And, with  XEN pv, swapgs_restore_regs_and_return_to_usermode() pushing
the IRET frame on to the original address is useless and error-prone
when there is any future attempt to modify the code.

 [ bp: Massage commit message. ]

Fixes: 7f2590a1 ("x86/entry/64: Use a per-CPU trampoline stack for IDT entries")
Signed-off-by: NLai Jiangshan <laijs@linux.alibaba.com>
Signed-off-by: NBorislav Petkov <bp@suse.de>
Reviewed-by: NBoris Ostrovsky <boris.ostrovsky@oracle.com>
Link: https://lkml.kernel.org/r/20211126101209.8613-4-jiangshanlai@gmail.com