[PATCH v2] powerpc/interrupt: Use early_radix_enabled() in NMI real-mode guard

radix_enabled() uses a jump label which is only valid after
mmu_feature_keys_init() is called. Before that point, on a pSeries
hash guest, early_check_vec5() clears MMU_FTR_TYPE_RADIX in
cur_cpu_spec->mmu_features but the jump label has not yet been patched,
so radix_enabled() incorrectly returns true.

Replace radix_enabled() with early_radix_enabled() which does a plain
bitmask check against cur_cpu_spec->mmu_features and is correct at all
times, including before jump label initialization.

Add the missing #include <asm/mmu.h> since early_radix_enabled() is
declared there.

Fixes: 8d0e21012743 ("powerpc/mce: Avoid nmi_enter/exit in real mode on pseries hash")
Signed-off-by: Venkat Rao Bagalkote <redacted>
Reviewed-by: Mukesh Kumar Chaurasiya <redacted>
---
v2:
 - Added Fixes: tag referencing commit 8d0e21012743.
 - Included <asm/mmu.h> in alphabetical order.
 - Added Reviewed-by tag from Mukesh.

 arch/powerpc/include/asm/interrupt.h | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/arch/powerpc/include/asm/interrupt.h b/arch/powerpc/include/asm/interrupt.h
index 1b45a49e9bed..355f6bbf9894 100644
--- a/arch/powerpc/include/asm/interrupt.h
+++ b/arch/powerpc/include/asm/interrupt.h

@@ -70,6 +70,7 @@
 #include <linux/irq-entry-common.h>
 
 #include <asm/kprobes.h>
+#include <asm/mmu.h>
 #include <asm/runlatch.h>
 
 #ifdef CONFIG_PPC_IRQ_SOFT_MASK_DEBUG

@@ -290,7 +291,7 @@ interrupt_handler long func(struct pt_regs *regs)			\
 		state = irqentry_nmi_enter(regs);			\
 	} else if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) &&			\
 		   firmware_has_feature(FW_FEATURE_LPAR) &&		\
-		   !radix_enabled()) {					\
+		   !early_radix_enabled()) {				\
 		/* no nmi_entry for a pseries hash guest		\
 		 * taking a real mode exception */			\
 	} else if (IS_ENABLED(CONFIG_KASAN)) {				\

@@ -307,7 +308,7 @@ interrupt_handler long func(struct pt_regs *regs)			\
 		irqentry_nmi_exit(regs, state);				\
 	} else if (IS_ENABLED(CONFIG_PPC_BOOK3S_64) &&			\
 		   firmware_has_feature(FW_FEATURE_LPAR) &&		\
-		   !radix_enabled()) {					\
+		   !early_radix_enabled()) {				\
 		/* no nmi_exit for a pseries hash guest			\
 		 * taking a real mode exception */			\
 	} else if (IS_ENABLED(CONFIG_KASAN)) {				\

-- 
2.45.2

On 9/9/26 2:42 PM, Venkat Rao Bagalkote wrote:

radix_enabled() uses a jump label which is only valid after
mmu_feature_keys_init() is called. Before that point, on a pSeries
hash guest, early_check_vec5() clears MMU_FTR_TYPE_RADIX in
cur_cpu_spec->mmu_features but the jump label has not yet been patched,
so radix_enabled() incorrectly returns true.

The usage window where it goes wrong is:

early_setup:
	configure_exceptions();
	<exception can happen now, and could prepare/exit of corresponding exception>
		<those could use radix_enabled() >
	setup_feature_keys();
	<jump labels are setup and post this it is safe to use radix_enabled()>

But since common wrappers don't have the context of early or late,
using early_radix_enabled() is safer option.

Commit message could convey the same.

Replace radix_enabled() with early_radix_enabled() which does a plain
bitmask check against cur_cpu_spec->mmu_features and is correct at all
times, including before jump label initialization.

If you have data, please put the values of both early_radix_enabled and radix_enabled()
in hash/radix guests in the chaneglog to ensure replacing with early variant
is functionally safe.