On Tue, Jul 06, 2021 at 12:18:18PM +0800, Huacai Chen wrote:
> +#define __smp_load_acquire(p) \
> +({ \
> + union { typeof(*p) __val; char __c[1]; } __u; \
> + unsigned long __tmp = 0; \
> + compiletime_assert_atomic_type(*p); \
> + switch (sizeof(*p)) { \
> + case 1: \
> + *(__u8 *)__u.__c = *(volatile __u8 *)p; \
> + __smp_mb(); \
> + break; \
> + case 2: \
> + *(__u16 *)__u.__c = *(volatile __u16 *)p; \
> + __smp_mb(); \
> + break; \
> + case 4: \
> + __asm__ __volatile__( \
> + "amor.w %[val], %[tmp], %[mem] \n" \
> + : [val] "=&r" (*(__u32 *)__u.__c) \
> + : [mem] "ZB" (*(u32 *) p), [tmp] "r" (__tmp) \
> + : "memory"); \
> + break; \
> + case 8: \
> + __asm__ __volatile__( \
> + "amor.d %[val], %[tmp], %[mem] \n" \
> + : [val] "=&r" (*(__u64 *)__u.__c) \
> + : [mem] "ZB" (*(u64 *) p), [tmp] "r" (__tmp) \
> + : "memory"); \
> + break; \
> + default: \
> + barrier(); \
> + __builtin_memcpy((void *)__u.__c, (const void *)p, sizeof(*p)); \
> + __smp_mb(); \
smp_load_acquire() is explicitly not defined on longer than machine word
sizes.
> + } \
> + __u.__val; \
> +})
By using that cute AMO-fetch-or, this LOAD turns into a LOAD-STORE
cycle. Which means you cannot use it on RO memory -- also cache fail.
Surely just the volatile load and smp_mb() is faster and saner.
