On 7/10/25 10:54 AM, Puranjay Mohan wrote:
As BPF doesn't include any barrier instructions, smp_mb() is implemented by doing a dummy value returning atomic operation. Such an operation acts a full barrier as enforced by LKMM and also by the work in progress BPF memory model. If the returned value is not used, clang[1] can optimize the value returning atomic instruction in to a normal atomic instruction which provides no ordering guarantees. Mark the variable as volatile so the above optimization is never performed and smp_mb() works as expected. [1] https://godbolt.org/z/qzze7bG6z
You are using llvm19 in the above godbolt run. But from llvm20, instead of 'lock ...' insn, 'atomic_fetch_or' will be generated so barrier semantics will be preserved. Since CI is using llvm20, so we should not have any problem. But for llvm19 or lower, the patch does fix a problem for arm64 etc. So in case that maintainer agrees with this patch, my ACK is below: Acked-by: Yonghong Song <yonghong.song@xxxxxxxxx>
Fixes: 88d706ba7cc5 ("selftests/bpf: Introduce arena spin lock") Signed-off-by: Puranjay Mohan <puranjay@xxxxxxxxxx> --- tools/testing/selftests/bpf/bpf_atomic.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tools/testing/selftests/bpf/bpf_atomic.h b/tools/testing/selftests/bpf/bpf_atomic.h index a9674e544322..c550e5711967 100644 --- a/tools/testing/selftests/bpf/bpf_atomic.h +++ b/tools/testing/selftests/bpf/bpf_atomic.h @@ -61,7 +61,7 @@ extern bool CONFIG_X86_64 __kconfig __weak;#define smp_mb() \({ \ - unsigned long __val; \ + volatile unsigned long __val; \ __sync_fetch_and_add(&__val, 0); \ })
