x86/nmi/64: Switch stacks on userspace NMI entry

Returning to userspace is tricky: IRET can fail, and ESPFIX can
rearrange the stack prior to IRET.

The NMI nesting fixup relies on a precise stack layout and
atomic IRET.  Rather than trying to teach the NMI nesting fixup
to handle ESPFIX and failed IRET, punt: run NMIs that came from
user mode on the normal kernel stack.

This will make some nested NMIs visible to C code, but the C
code is okay with that.

As a side effect, this should speed up perf: it eliminates an
RDMSR when NMIs come from user mode.

Signed-off-by: Andy Lutomirski <luto@kernel.org>
Reviewed-by: Steven Rostedt <rostedt@goodmis.org>
Reviewed-by: Borislav Petkov <bp@suse.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: stable@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>

arch/x86/entry/entry_64.S

@@ -1250,18 +1250,72 @@ ENTRY(nmi)
 	 * a nested NMI that updated the copy interrupt stack frame, a
 	 * jump will be made to the repeat_nmi code that will handle the second
 	 * NMI.
+	 *
+	 * However, espfix prevents us from directly returning to userspace
+	 * with a single IRET instruction.  Similarly, IRET to user mode
+	 * can fault.  We therefore handle NMIs from user space like
+	 * other IST entries.
 	 */
 
 	/* Use %rdx as our temp variable throughout */
 	pushq	%rdx
 
-	/*
-	 * If %cs was not the kernel segment, then the NMI triggered in user
-	 * space, which means it is definitely not nested.
-	 */
-	cmpl	$__KERNEL_CS, 16(%rsp)
-	jne	first_nmi
+	testb	$3, CS-RIP+8(%rsp)
+	jz	.Lnmi_from_kernel
 
+	/*
+	 * NMI from user mode.  We need to run on the thread stack, but we
+	 * can't go through the normal entry paths: NMIs are masked, and
+	 * we don't want to enable interrupts, because then we'll end
+	 * up in an awkward situation in which IRQs are on but NMIs
+	 * are off.
+	 */
+
+	SWAPGS
+	cld
+	movq	%rsp, %rdx
+	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
+	pushq	5*8(%rdx)	/* pt_regs->ss */
+	pushq	4*8(%rdx)	/* pt_regs->rsp */
+	pushq	3*8(%rdx)	/* pt_regs->flags */
+	pushq	2*8(%rdx)	/* pt_regs->cs */
+	pushq	1*8(%rdx)	/* pt_regs->rip */
+	pushq   $-1		/* pt_regs->orig_ax */
+	pushq   %rdi		/* pt_regs->di */
+	pushq   %rsi		/* pt_regs->si */
+	pushq   (%rdx)		/* pt_regs->dx */
+	pushq   %rcx		/* pt_regs->cx */
+	pushq   %rax		/* pt_regs->ax */
+	pushq   %r8		/* pt_regs->r8 */
+	pushq   %r9		/* pt_regs->r9 */
+	pushq   %r10		/* pt_regs->r10 */
+	pushq   %r11		/* pt_regs->r11 */
+	pushq	%rbx		/* pt_regs->rbx */
+	pushq	%rbp		/* pt_regs->rbp */
+	pushq	%r12		/* pt_regs->r12 */
+	pushq	%r13		/* pt_regs->r13 */
+	pushq	%r14		/* pt_regs->r14 */
+	pushq	%r15		/* pt_regs->r15 */
+
+	/*
+	 * At this point we no longer need to worry about stack damage
+	 * due to nesting -- we're on the normal thread stack and we're
+	 * done with the NMI stack.
+	 */
+
+	movq	%rsp, %rdi
+	movq	$-1, %rsi
+	call	do_nmi
+
+	/*
+	 * Return back to user mode.  We must *not* do the normal exit
+	 * work, because we don't want to enable interrupts.  Fortunately,
+	 * do_nmi doesn't modify pt_regs.
+	 */
+	SWAPGS
+	jmp	restore_c_regs_and_iret
+
+.Lnmi_from_kernel:
 	/*
 	 * Check the special variable on the stack to see if NMIs are
 	 * executing.