[PATCH bpf-next v1 4/7] selftests/bpf: make arena_base() opaque to GCC's -Warray-bounds

[PATCH bpf-next v1 4/7] selftests/bpf: make arena_base() opaque to GCC's -Warray-bounds

[Vineet Gupta]

arena_base() casts the map definition to the kernel's struct bpf_arena
and reads user_vm_start out of it.  In the object file that definition
is only the 32-byte BTF placeholder the loader rewrites, so the access
is past the end of the object and GCC says so:

  libarena/include/bpf_arena_common.h:52:69: error: array subscript 'struct bpf_arena[0]' is partly outside array bounds of 'struct <anonymous>[1]' [-Werror=array-bounds=]
     52 | #define arena_base(map) ((void __arena *)((struct bpf_arena *)(map))->user_vm_start)
  progs/arena_race.c:58:30: note: in expansion of macro 'arena_base'

GCC is right; clang has the same warning but does not apply it to a
member access through a cast.  With BPF_CFLAGS carrying -Werror this is
a build failure, today, for every prog that calls arena_base() and is
compiled by a GCC with arena support: verifier_arena.c,
verifier_arena_large.c, verifier_arena_globals1.c,
verifier_arena_globals2.c and arena_race.c.

Make the pointer opaque instead, which is what progs/stream.c already
does for its own open-coded copy of the same access.  The bounds
analysis then has nothing to reason about, and every caller is covered
without a per-file opt-out that has to be maintained -- arena_race.c
was added without one and broke the bpf_gcc build.

The barrier is behind #ifdef __clang__ so clang keeps the plain macro.
It does not need it, and the barrier is not free: it blocks reuse of
the map address, so each call site reloads it.  Across the five progs
and 20 call sites that is +35 instructions under GCC
(2528 -> 2563), about 1.4% on those objects and nothing elsewhere.
Clang output is unchanged, bit for bit.

Signed-off-by: Vineet Gupta <vineet.gupta@linux.dev>
---
 .../bpf/libarena/include/bpf_arena_common.h      | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/tools/testing/selftests/bpf/libarena/include/bpf_arena_common.h b/tools/testing/selftests/bpf/libarena/include/bpf_arena_common.h
index 82aafe879fae..4f2c6050215f 100644
--- a/tools/testing/selftests/bpf/libarena/include/bpf_arena_common.h
+++ b/tools/testing/selftests/bpf/libarena/include/bpf_arena_common.h
@@ -49,7 +49,23 @@ void __arena* bpf_arena_alloc_pages(void *map, void __arena *addr, __u32 page_cn
 int bpf_arena_reserve_pages(void *map, void __arena *addr, __u32 page_cnt) __ksym __weak;
 void bpf_arena_free_pages(void *map, void __arena *ptr, __u32 page_cnt) __ksym __weak;
 
+/*
+ * GCC's -Warray-bounds fires on the member access: the map definition the
+ * object file carries is smaller than struct bpf_arena.  Make the pointer
+ * opaque to it, as the open-coded form in progs/stream.c already does.
+ * clang does not warn and does not need the barrier, which costs a reload
+ * of the map address at every call site.
+ */
+#ifdef __clang__
 #define arena_base(map) ((void __arena *)((struct bpf_arena *)(map))->user_vm_start)
+#else
+#define arena_base(map)							\
+({									\
+	struct bpf_arena *__ma = (void *)(map);				\
+	barrier_var(__ma);						\
+	(void __arena *)__ma->user_vm_start;				\
+})
+#endif
 
 #else /* when compiled as user space code */
 
-- 
2.53.0-Meta