selftests/bpf: Fix selftest verifier_arena_large failure

With latest llvm22, I got the following verification failure:

  ...
  ; int big_alloc2(void *ctx) @ verifier_arena_large.c:207
  0: (b4) w6 = 1                        ; R6_w=1
  ...
  ; if (err) @ verifier_arena_large.c:233
  53: (56) if w6 != 0x0 goto pc+62      ; R6=0
  54: (b7) r7 = -4                      ; R7_w=-4
  55: (18) r8 = 0x7f4000000000          ; R8_w=scalar()
  57: (bf) r9 = addr_space_cast(r8, 0, 1)       ; R8_w=scalar() R9_w=arena
  58: (b4) w6 = 5                       ; R6_w=5
  ; pg = page[i]; @ verifier_arena_large.c:238
  59: (bf) r1 = r7                      ; R1_w=-4 R7_w=-4
  60: (07) r1 += 4                      ; R1_w=0
  61: (79) r2 = *(u64 *)(r9 +0)         ; R2_w=scalar() R9_w=arena
  ; if (*pg != i) @ verifier_arena_large.c:239
  62: (bf) r3 = addr_space_cast(r2, 0, 1)       ; R2_w=scalar() R3_w=arena
  63: (71) r3 = *(u8 *)(r3 +0)          ; R3_w=scalar(smin=smin32=0,smax=umax=smax32=umax32=255,var_off=(0x0; 0xff))
  64: (5d) if r1 != r3 goto pc+51       ; R1_w=0 R3_w=0
  ; bpf_arena_free_pages(&arena, (void __arena *)pg, 2); @ verifier_arena_large.c:241
  65: (18) r1 = 0xff11000114548000      ; R1_w=map_ptr(map=arena,ks=0,vs=0)
  67: (b4) w3 = 2                       ; R3_w=2
  68: (85) call bpf_arena_free_pages#72675      ;
  69: (b7) r1 = 0                       ; R1_w=0
  ; page[i + 1] = NULL; @ verifier_arena_large.c:243
  70: (7b) *(u64 *)(r8 +8) = r1
  R8 invalid mem access 'scalar'
  processed 61 insns (limit 1000000) max_states_per_insn 0 total_states 6 peak_states 6 mark_read 2
  =============
  #489/5   verifier_arena_large/big_alloc2:FAIL

The main reason is that 'r8' in insn '70' is not an arena pointer.
Further debugging at llvm side shows that llvm commit ([1]) caused
the failure. For the original code:
  page[i] = NULL;
  page[i + 1] = NULL;
the llvm transformed it to something like below at source level:
  __builtin_memset(&page[i], 0, 16)
Such transformation prevents llvm BPFCheckAndAdjustIR pass from
generating proper addr_space_cast insns ([2]).

Adding support in llvm BPFCheckAndAdjustIR pass should work, but
not sure that such a pattern exists or not in real applications.
At the same time, simply adding a memory barrier between two 'page'
assignment can fix the issue.

  [1] llvm/llvm-project#155415
  [2] llvm/llvm-project#84410

Cc: Eduard Zingerman <[email protected]>
Signed-off-by: Yonghong Song <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
Signed-off-by: Alexei Starovoitov <[email protected]>

Author: Yonghong Song

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

@@ -240,6 +240,7 @@ int big_alloc2(void *ctx)
 			return 5;
 		bpf_arena_free_pages(&arena, (void __arena *)pg, 2);
 		page[i] = NULL;
+		barrier();
 		page[i + 1] = NULL;
 		cond_break;
 	}