x86/pae: use 64 bit atomic xchg function in native_ptep_get_and_clear

Using only 32-bit writes for the pte will result in an intermediate
L1TF vulnerable PTE. When running as a Xen PV guest this will at once
switch the guest to shadow mode resulting in a loss of performance.

Use arch_atomic64_xchg() instead which will perform the requested
operation atomically with all 64 bits.

Some performance considerations according to:

https://software.intel.com/sites/default/files/managed/ad/dc/Intel-Xeon-Scalable-Processor-throughput-latency.pdf

The main number should be the latency, as there is no tight loop around
native_ptep_get_and_clear().

"lock cmpxchg8b" has a latency of 20 cycles, while "lock xchg" (with a
memory operand) isn't mentioned in that document. "lock xadd" (with xadd
having 3 cycles less latency than xchg) has a latency of 11, so we can
assume a latency of 14 for "lock xchg".

Signed-off-by: Juergen Gross <[email protected]>
Reviewed-by: Thomas Gleixner <[email protected]>
Reviewed-by: Jan Beulich <[email protected]>
Tested-by: Jason Andryuk <[email protected]>
Signed-off-by: Boris Ostrovsky <[email protected]>

Author: jgross1

arch/x86/include/asm/pgtable-3level.h

@@ -2,6 +2,8 @@
 #ifndef _ASM_X86_PGTABLE_3LEVEL_H
 #define _ASM_X86_PGTABLE_3LEVEL_H
 
+#include <asm/atomic64_32.h>
+
 /*
  * Intel Physical Address Extension (PAE) Mode - three-level page
  * tables on PPro+ CPUs.
@@ -150,10 +152,7 @@ static inline pte_t native_ptep_get_and_clear(pte_t *ptep)
 {
 	pte_t res;
 
-	/* xchg acts as a barrier before the setting of the high bits */
-	res.pte_low = xchg(&ptep->pte_low, 0);
-	res.pte_high = ptep->pte_high;
-	ptep->pte_high = 0;
+	res.pte = (pteval_t)arch_atomic64_xchg((atomic64_t *)ptep, 0);
 
 	return res;
 }