selftests/bpf: Fix arena_atomics failure due to llvm change

llvm change [1] made a change such that __sync_fetch_and_{and,or,xor}()
will generate atomic_fetch_*() insns even if the return value is not used.
This is a deliberate choice to make sure barrier semantics are preserved
from source code to asm insn.

But the change in [1] caused arena_atomics selftest failure.

  test_arena_atomics:PASS:arena atomics skeleton open 0 nsec
  libbpf: prog 'and': BPF program load failed: Permission denied
  libbpf: prog 'and': -- BEGIN PROG LOAD LOG --
  arg#0 reference type('UNKNOWN ') size cannot be determined: -22
  0: R1=ctx() R10=fp0
  ; if (pid != (bpf_get_current_pid_tgid() >> 32)) @ arena_atomics.c:87
  0: (18) r1 = 0xffffc90000064000       ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4)
  2: (61) r6 = *(u32 *)(r1 +0)          ; R1_w=map_value(map=arena_at.bss,ks=4,vs=4) R6_w=scalar(smin=0,smax=umax=0xffffffff,v
ar_off=(0x0; 0xffffffff))
  3: (85) call bpf_get_current_pid_tgid#14      ; R0_w=scalar()
  4: (77) r0 >>= 32                     ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff))
  5: (5d) if r0 != r6 goto pc+11        ; R0_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0xffffffff)) R6_w=scalar(smin=0,smax=umax=0xffffffff,var_off=(0x0; 0x)
  ; __sync_fetch_and_and(&and64_value, 0x011ull << 32); @ arena_atomics.c:91
  6: (18) r1 = 0x100000000060           ; R1_w=scalar()
  8: (bf) r1 = addr_space_cast(r1, 0, 1)        ; R1_w=arena
  9: (18) r2 = 0x1100000000             ; R2_w=0x1100000000
  11: (db) r2 = atomic64_fetch_and((u64 *)(r1 +0), r2)
  BPF_ATOMIC stores into R1 arena is not allowed
  processed 9 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0
  -- END PROG LOAD LOG --
  libbpf: prog 'and': failed to load: -13
  libbpf: failed to load object 'arena_atomics'
  libbpf: failed to load BPF skeleton 'arena_atomics': -13
  test_arena_atomics:FAIL:arena atomics skeleton load unexpected error: -13 (errno 13)
  openbmc#3       arena_atomics:FAIL

The reason of the failure is due to [2] where atomic{64,}_fetch_{and,or,xor}() are not
allowed by arena addresses.

Version 2 of the patch fixed the issue by using inline asm ([3]). But further discussion
suggested to find a way from source to generate locked insn which is more user
friendly. So in not-merged llvm patch ([4]), if relax memory ordering is used and
the return value is not used, locked insn could be generated.

So with llvm patch [4] to compile the bpf selftest, the following code
  __c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
is able to generate locked insn, hence fixing the selftest failure.

  [1] llvm/llvm-project#106494
  [2] d503a04 ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
  [3] https://lore.kernel.org/bpf/20240803025928.4184433-1-yonghong.song@linux.dev/
  [4] llvm/llvm-project#107343

Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
Link: https://lore.kernel.org/r/20240909223431.1666305-1-yonghong.song@linux.dev
Signed-off-by: Alexei Starovoitov <ast@kernel.org>

tools/testing/selftests/bpf/progs/arena_atomics.c

@@ -4,6 +4,7 @@
 #include <bpf/bpf_helpers.h>
 #include <bpf/bpf_tracing.h>
 #include <stdbool.h>
+#include <stdatomic.h>
 #include "bpf_arena_common.h"
 
 struct {
@@ -77,50 +78,79 @@ int sub(const void *ctx)
 	return 0;
 }
 
+#ifdef __BPF_FEATURE_ATOMIC_MEM_ORDERING
+_Atomic __u64 __arena_global and64_value = (0x110ull << 32);
+_Atomic __u32 __arena_global and32_value = 0x110;
+#else
 __u64 __arena_global and64_value = (0x110ull << 32);
 __u32 __arena_global and32_value = 0x110;
+#endif
 
 SEC("raw_tp/sys_enter")
 int and(const void *ctx)
 {
 	if (pid != (bpf_get_current_pid_tgid() >> 32))
 		return 0;
 #ifdef ENABLE_ATOMICS_TESTS
-
+#ifdef __BPF_FEATURE_ATOMIC_MEM_ORDERING
+	__c11_atomic_fetch_and(&and64_value, 0x011ull << 32, memory_order_relaxed);
+	__c11_atomic_fetch_and(&and32_value, 0x011, memory_order_relaxed);
+#else
 	__sync_fetch_and_and(&and64_value, 0x011ull << 32);
 	__sync_fetch_and_and(&and32_value, 0x011);
+#endif
 #endif
 
 	return 0;
 }
 
+#ifdef __BPF_FEATURE_ATOMIC_MEM_ORDERING
+_Atomic __u32 __arena_global or32_value = 0x110;
+_Atomic __u64 __arena_global or64_value = (0x110ull << 32);
+#else
 __u32 __arena_global or32_value = 0x110;
 __u64 __arena_global or64_value = (0x110ull << 32);
+#endif
 
 SEC("raw_tp/sys_enter")
 int or(const void *ctx)
 {
 	if (pid != (bpf_get_current_pid_tgid() >> 32))
 		return 0;
 #ifdef ENABLE_ATOMICS_TESTS
+#ifdef __BPF_FEATURE_ATOMIC_MEM_ORDERING
+	__c11_atomic_fetch_or(&or64_value, 0x011ull << 32, memory_order_relaxed);
+	__c11_atomic_fetch_or(&or32_value, 0x011, memory_order_relaxed);
+#else
 	__sync_fetch_and_or(&or64_value, 0x011ull << 32);
 	__sync_fetch_and_or(&or32_value, 0x011);
+#endif
 #endif
 
 	return 0;
 }
 
+#ifdef __BPF_FEATURE_ATOMIC_MEM_ORDERING
+_Atomic __u64 __arena_global xor64_value = (0x110ull << 32);
+_Atomic __u32 __arena_global xor32_value = 0x110;
+#else
 __u64 __arena_global xor64_value = (0x110ull << 32);
 __u32 __arena_global xor32_value = 0x110;
+#endif
 
 SEC("raw_tp/sys_enter")
 int xor(const void *ctx)
 {
 	if (pid != (bpf_get_current_pid_tgid() >> 32))
 		return 0;
 #ifdef ENABLE_ATOMICS_TESTS
+#ifdef __BPF_FEATURE_ATOMIC_MEM_ORDERING
+	__c11_atomic_fetch_xor(&xor64_value, 0x011ull << 32, memory_order_relaxed);
+	__c11_atomic_fetch_xor(&xor32_value, 0x011, memory_order_relaxed);
+#else
 	__sync_fetch_and_xor(&xor64_value, 0x011ull << 32);
 	__sync_fetch_and_xor(&xor32_value, 0x011);
+#endif
 #endif
 
 	return 0;