powerpc/64: Fix msr_check_and_set/clear MSR[EE] race

irq soft-masking means that when Linux irqs are disabled, the MSR[EE]
value can change from 1 to 0 asynchronously: if a masked interrupt of
the PACA_IRQ_MUST_HARD_MASK variety fires while irqs are disabled,
the masked handler will return with MSR[EE]=0.

This means a sequence like mtmsr(mfmsr() | MSR_FP) is racy if it can
be called with local irqs disabled, unless a hard_irq_disable has been
done.

Reported-by: Sachin Sant <sachinp@linux.ibm.com>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221004051157.308999-2-npiggin@gmail.com

1 files changed, 24 insertions, 0 deletions

diff --git a/arch/powerpc/include/asm/hw_irq.h b/arch/powerpc/include/asm/hw_irq.h
index e8de249339d8..77fa88c2aed0 100644
--- a/arch/powerpc/include/asm/hw_irq.h
+++ b/arch/powerpc/include/asm/hw_irq.h
@@ -471,6 +471,30 @@ static inline void irq_soft_mask_regs_set_state(struct pt_regs *regs, unsigned l
 }
 #endif /* CONFIG_PPC64 */
 
+static inline unsigned long mtmsr_isync_irqsafe(unsigned long msr)
+{
+#ifdef CONFIG_PPC64
+	if (arch_irqs_disabled()) {
+		/*
+		 * With soft-masking, MSR[EE] can change from 1 to 0
+		 * asynchronously when irqs are disabled, and we don't want to
+		 * set MSR[EE] back to 1 here if that has happened. A race-free
+		 * way to do this is ensure EE is already 0. Another way it
+		 * could be done is with a RESTART_TABLE handler, but that's
+		 * probably overkill here.
+		 */
+		msr &= ~MSR_EE;
+		mtmsr_isync(msr);
+		irq_soft_mask_set(IRQS_ALL_DISABLED);
+		local_paca->irq_happened |= PACA_IRQ_HARD_DIS;
+	} else
+#endif
+		mtmsr_isync(msr);
+
+	return msr;
+}
+
+
 #define ARCH_IRQ_INIT_FLAGS	IRQ_NOREQUEST
 
 #endif  /* __ASSEMBLY__ */