[RFC PATCH 05/16] x86/split_lock: Use non atomic set and clear bit instructions in clear_cpufeature()

[Fenghua Yu]

x86_capability can span two cache lines depending on kernel configuration
and building environment. When #AC exception is enabled for split locked
accesses, clear_cpufeature() may generate #AC exception because of atomic
setting or clearing bits in x86_capability.

But kernel clears cpufeature only during a CPU is booting up. Therefore,
there is no racing condition when clear_cpufeature() is called and no need
to atomically clear or set bits in x86_capability.

To avoid #AC exception caused by split lock, call non atomic __set_bit()
and __clear_bit(). They are faster than atomic set_bit() and clear_bit()
as well.

Signed-off-by: Fenghua Yu <fenghua.yu@xxxxxxxxx>
---
 arch/x86/kernel/cpu/cpuid-deps.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/arch/x86/kernel/cpu/cpuid-deps.c b/arch/x86/kernel/cpu/cpuid-deps.c
index 2c0bd38a44ab..b2c2a004c769 100644
--- a/arch/x86/kernel/cpu/cpuid-deps.c
+++ b/arch/x86/kernel/cpu/cpuid-deps.c
@@ -65,15 +65,15 @@ static const struct cpuid_dep cpuid_deps[] = {
 static inline void clear_feature(struct cpuinfo_x86 *c, unsigned int feature)
 {
 	/*
-	 * Note: This could use the non atomic __*_bit() variants, but the
-	 * rest of the cpufeature code uses atomics as well, so keep it for
-	 * consistency. Cleanup all of it separately.
+	 * Because this code is only called during boot time and there
+	 * is no need to be atomic, use non atomic __*_bit() for better
+	 * performance and to avoid #AC exception for split locked access.
 	 */
 	if (!c) {
 		clear_cpu_cap(&boot_cpu_data, feature);
-		set_bit(feature, (unsigned long *)cpu_caps_cleared);
+		__set_bit(feature, (unsigned long *)cpu_caps_cleared);
 	} else {
-		clear_bit(feature, (unsigned long *)c->x86_capability);
+		__clear_bit(feature, (unsigned long *)c->x86_capability);
 	}
 }
 
-- 
2.5.0