selftests/x86/sigreturn/64: Fix spurious failures on AMD CPUs

When I wrote the sigreturn test, I didn't realize that AMD's busted
IRET behavior was different from Intel's busted IRET behavior:

On AMD CPUs, the CPU leaks the high 32 bits of the kernel stack pointer
to certain userspace contexts.  Gee, thanks.  There's very little
the kernel can do about it.  Modify the test so it passes.
Signed-off-by: Andy Lutomirski <luto@kernel.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/86e7fd3564497f657de30a36da4505799eebef01.1530076529.git.luto@kernel.orgSigned-off-by: Ingo Molnar <mingo@kernel.org>

tools/testing/selftests/x86/sigreturn.c

@@ -612,19 +612,38 @@ static int test_valid_sigreturn(int cs_bits, bool use_16bit_ss, int force_ss)
 		greg_t req = requested_regs[i], res = resulting_regs[i];
 		if (i == REG_TRAPNO || i == REG_IP)
 			continue;	/* don't care */
-		if (i == REG_SP) {
-			printf("\tSP: %llx -> %llx\n", (unsigned long long)req,
-			       (unsigned long long)res);
 
+		if (i == REG_SP) {
 			/*
-			 * In many circumstances, the high 32 bits of rsp
-			 * are zeroed.  For example, we could be a real
-			 * 32-bit program, or we could hit any of a number
-			 * of poorly-documented IRET or segmented ESP
-			 * oddities.  If this happens, it's okay.
+			 * If we were using a 16-bit stack segment, then
+			 * the kernel is a bit stuck: IRET only restores
+			 * the low 16 bits of ESP/RSP if SS is 16-bit.
+			 * The kernel uses a hack to restore bits 31:16,
+			 * but that hack doesn't help with bits 63:32.
+			 * On Intel CPUs, bits 63:32 end up zeroed, and, on
+			 * AMD CPUs, they leak the high bits of the kernel
+			 * espfix64 stack pointer.  There's very little that
+			 * the kernel can do about it.
+			 *
+			 * Similarly, if we are returning to a 32-bit context,
+			 * the CPU will often lose the high 32 bits of RSP.
 			 */
-			if (res == (req & 0xFFFFFFFF))
-				continue;  /* OK; not expected to work */
+
+			if (res == req)
+				continue;
+
+			if (cs_bits != 64 && ((res ^ req) & 0xFFFFFFFF) == 0) {
+				printf("[NOTE]\tSP: %llx -> %llx\n",
+				       (unsigned long long)req,
+				       (unsigned long long)res);
+				continue;
+			}
+
+			printf("[FAIL]\tSP mismatch: requested 0x%llx; got 0x%llx\n",
+			       (unsigned long long)requested_regs[i],
+			       (unsigned long long)resulting_regs[i]);
+			nerrs++;
+			continue;
 		}
 
 		bool ignore_reg = false;
@@ -663,13 +682,6 @@ static int test_valid_sigreturn(int cs_bits, bool use_16bit_ss, int force_ss)
 		}
 
 		if (requested_regs[i] != resulting_regs[i] && !ignore_reg) {
-			/*
-			 * SP is particularly interesting here.  The
-			 * usual cause of failures is that we hit the
-			 * nasty IRET case of returning to a 16-bit SS,
-			 * in which case bits 16:31 of the *kernel*
-			 * stack pointer persist in ESP.
-			 */
 			printf("[FAIL]\tReg %d mismatch: requested 0x%llx; got 0x%llx\n",
 			       i, (unsigned long long)requested_regs[i],
 			       (unsigned long long)resulting_regs[i]);