arm64: prefetch: don't provide spin_lock_prefetch with LSE

The LSE atomics rely on us not dirtying data at L1 if we can avoid it,
otherwise many of the potential scalability benefits are lost.

This patch replaces spin_lock_prefetch with a nop when the LSE atomics
are in use, so that users don't shoot themselves in the foot by causing
needless coherence traffic at L1.
Signed-off-by: NWill Deacon <will.deacon@arm.com>
Tested-by: NAndrew Pinski <apinski@cavium.com>
Signed-off-by: NCatalin Marinas <catalin.marinas@arm.com>

arch/arm64/include/asm/processor.h

@@ -29,6 +29,7 @@
 
 #include <linux/string.h>
 
+#include <asm/alternative.h>
 #include <asm/fpsimd.h>
 #include <asm/hw_breakpoint.h>
 #include <asm/pgtable-hwdef.h>
@@ -177,9 +178,11 @@ static inline void prefetchw(const void *ptr)
 }
 
 #define ARCH_HAS_SPINLOCK_PREFETCH
-static inline void spin_lock_prefetch(const void *x)
+static inline void spin_lock_prefetch(const void *ptr)
 {
-	prefetchw(x);
+	asm volatile(ARM64_LSE_ATOMIC_INSN(
+		     "prfm pstl1strm, %a0",
+		     "nop") : : "p" (ptr));
 }
 
 #define HAVE_ARCH_PICK_MMAP_LAYOUT