powerpc/64s: Fix unrecoverable MCE calling async handler from NMI

The machine check handler is not considered NMI on 64s. The early
handler is the true NMI handler, and then it schedules the
machine_check_exception handler to run when interrupts are enabled.

This works fine except the case of an unrecoverable MCE, where the true
NMI is taken when MSR[RI] is clear, it can not recover, so it calls
machine_check_exception directly so something might be done about it.

Calling an async handler from NMI context can result in irq state and
other things getting corrupted. This can also trigger the BUG at
  arch/powerpc/include/asm/interrupt.h:168
  BUG_ON(!arch_irq_disabled_regs(regs) && !(regs->msr & MSR_EE));

Fix this by making an _async version of the handler which is called
in the normal case, and a NMI version that is called for unrecoverable
interrupts.

Fixes: 2b43dd7653cc ("powerpc/64: enable MSR[EE] in irq replay pt_regs")
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Tested-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20211004145642.1331214-6-npiggin@gmail.com

1 files changed, 18 insertions, 13 deletions

diff --git a/arch/powerpc/kernel/traps.c b/arch/powerpc/kernel/traps.c
index e453b666613b..11741703d26e 100644
--- a/arch/powerpc/kernel/traps.c
+++ b/arch/powerpc/kernel/traps.c
@@ -796,24 +796,22 @@ void die_mce(const char *str, struct pt_regs *regs, long err)
 	 * do_exit() checks for in_interrupt() and panics in that case, so
 	 * exit the irq/nmi before calling die.
 	 */
-	if (IS_ENABLED(CONFIG_PPC_BOOK3S_64))
-		irq_exit();
-	else
+	if (in_nmi())
 		nmi_exit();
+	else
+		irq_exit();
 	die(str, regs, err);
 }
 
 /*
- * BOOK3S_64 does not call this handler as a non-maskable interrupt
+ * BOOK3S_64 does not usually call this handler as a non-maskable interrupt
  * (it uses its own early real-mode handler to handle the MCE proper
  * and then raises irq_work to call this handler when interrupts are
- * enabled).
+ * enabled). The only time when this is not true is if the early handler
+ * is unrecoverable, then it does call this directly to try to get a
+ * message out.
  */
-#ifdef CONFIG_PPC_BOOK3S_64
-DEFINE_INTERRUPT_HANDLER_ASYNC(machine_check_exception)
-#else
-DEFINE_INTERRUPT_HANDLER_NMI(machine_check_exception)
-#endif
+static void __machine_check_exception(struct pt_regs *regs)
 {
 	int recover = 0;
 
@@ -847,12 +845,19 @@ bail:
 	/* Must die if the interrupt is not recoverable */
 	if (regs_is_unrecoverable(regs))
 		die_mce("Unrecoverable Machine check", regs, SIGBUS);
+}
 
 #ifdef CONFIG_PPC_BOOK3S_64
-	return;
-#else
-	return 0;
+DEFINE_INTERRUPT_HANDLER_ASYNC(machine_check_exception_async)
+{
+	__machine_check_exception(regs);
+}
 #endif
+DEFINE_INTERRUPT_HANDLER_NMI(machine_check_exception)
+{
+	__machine_check_exception(regs);
+
+	return 0;
 }
 
 DEFINE_INTERRUPT_HANDLER(SMIException) /* async? */