x86/asm/entry/64/compat: Use SYSRETL to return from compat mode SYSENTER

SYSEXIT is scary on 64-bit kernels -- SYSEXIT must be invoked
with usergs and IRQs on.  That means that we rely on STI to
correctly mask interrupts for one instruction.  This is okay by
itself, but the semantics with respect to NMIs are unclear.

Avoid the whole issue by using SYSRETL instead.  For background,
Intel CPUs don't allow SYSCALL from compat mode, but they do
allow SYSRETL back to compat mode.  Go figure.

To avoid doing too much at once, this doesn't revamp the calling
convention.  We still return with EBP, EDX, and ECX on the user
stack.

Oddly this seems to be 30 cycles or so faster.  Avoiding POPFQ
and STI will account for under half of that, I think, so my best
guess is that Intel just optimizes SYSRET much better than
SYSEXIT.

Signed-off-by: Andy Lutomirski <luto@kernel.org>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: Denys Vlasenko <vda.linux@googlemail.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/57a0bf1b5230b2716a64ebe48e9bc1110f7ab433.1428019097.git.luto@kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>

arch/x86/ia32/ia32entry.S

@@ -186,28 +186,55 @@ sysenter_dispatch:
 	testl	$_TIF_ALLWORK_MASK, ASM_THREAD_INFO(TI_flags, %rsp, SIZEOF_PTREGS)
 	jnz	sysexit_audit
 sysexit_from_sys_call:
+	/*
+	 * NB: SYSEXIT is not obviously safe for 64-bit kernels -- an
+	 * NMI between STI and SYSEXIT has poorly specified behavior,
+	 * and and NMI followed by an IRQ with usergs is fatal.  So
+	 * we just pretend we're using SYSEXIT but we really use
+	 * SYSRETL instead.
+	 *
+	 * This code path is still called 'sysexit' because it pairs
+	 * with 'sysenter' and it uses the SYSENTER calling convention.
+	 */
 	andl    $~TS_COMPAT,ASM_THREAD_INFO(TI_status, %rsp, SIZEOF_PTREGS)
-	/* clear IF, that popfq doesn't enable interrupts early */
-	andl	$~0x200,EFLAGS(%rsp)
-	movl	RIP(%rsp),%edx		/* User %eip */
-	CFI_REGISTER rip,rdx
+	movl	RIP(%rsp),%ecx		/* User %eip */
+	CFI_REGISTER rip,rcx
 	RESTORE_RSI_RDI
-	/* pop everything except ss,rsp,rflags slots */
-	REMOVE_PT_GPREGS_FROM_STACK 3*8
+	xorl	%edx,%edx		/* avoid info leaks */
 	xorq	%r8,%r8
 	xorq	%r9,%r9
 	xorq	%r10,%r10
-	xorq	%r11,%r11
-	popfq_cfi
+	movl	EFLAGS(%rsp),%r11d	/* User eflags */
 	/*CFI_RESTORE rflags*/
-	popq_cfi %rcx				/* User %esp */
-	CFI_REGISTER rsp,rcx
 	TRACE_IRQS_ON
+
 	/*
-	 * 32bit SYSEXIT restores eip from edx, esp from ecx.
-	 * cs and ss are loaded from MSRs.
+	 * SYSRETL works even on Intel CPUs.  Use it in preference to SYSEXIT,
+	 * since it avoids a dicey window with interrupts enabled.
 	 */
-	ENABLE_INTERRUPTS_SYSEXIT32
+	movl	RSP(%rsp),%esp
+
+	/*
+	 * USERGS_SYSRET32 does:
+	 *  gsbase = user's gs base
+	 *  eip = ecx
+	 *  rflags = r11
+	 *  cs = __USER32_CS
+	 *  ss = __USER_DS
+	 *
+	 * The prologue set RIP(%rsp) to VDSO32_SYSENTER_RETURN, which does:
+	 *
+	 *  pop %ebp
+	 *  pop %edx
+	 *  pop %ecx
+	 *
+	 * Therefore, we invoke SYSRETL with EDX and R8-R10 zeroed to
+	 * avoid info leaks.  R11 ends up with VDSO32_SYSENTER_RETURN's
+	 * address (already known to user code), and R12-R15 are
+	 * callee-saved and therefore don't contain any interesting
+	 * kernel data.
+	 */
+	USERGS_SYSRET32
 
 	CFI_RESTORE_STATE