Re: [PATCH bpf-next v2 8/9] selftests/bpf: Add selftests for load-acquire and store-release instructions

[Peilin Ye]

On Sat, Feb 08, 2025 at 12:20:58AM +0000, Peilin Ye wrote:
> > --- a/tools/testing/selftests/bpf/progs/arena_atomics.c
> > +++ b/tools/testing/selftests/bpf/progs/arena_atomics.c
> > @@ -6,6 +6,8 @@
> >  #include <stdbool.h>
> >  #include <stdatomic.h>
> >  #include "bpf_arena_common.h"
> > +#include "../../../include/linux/filter.h"
> > +#include "bpf_misc.h"
> >  
> >  struct {
> >  	__uint(type, BPF_MAP_TYPE_ARENA);
> > @@ -274,4 +276,90 @@ int uaf(const void *ctx)
> >  	return 0;
> >  }
> >  
> > +__u8 __arena_global load_acquire8_value = 0x12;
>    ~~~~
> 
> CI job x86_64-llvm-17 [1] failed because clang-17 crashed when compiling
> this file (arena_atomics.c):
> 
>   fatal error: error in backend: unable to write nop sequence of 1 bytes
> 
> After some digging, I believe I am hitting a known issue that Yonghong
> described in [2].

> Changing __u8 and __u16 variables to __u32 seems to resolve/work
> around the issue

Sorry, that wasn't very accurate - we need to make sure there are no
"holes" in the .addr_space.1 ELF section, e.g.:

  /* 8-byte-aligned */
  __u8 __arena_global load_acquire8_value = 0x12;
  /* 1-byte hole, causing clang-17 to crash */
  __u16 __arena_global load_acquire16_value = 0x1234;

LLVM commit f27c4903c43b ("MC: Add .data. and .rodata. prefixes to
MCContext section classification") fixed this issue.

- - -
For now, I think I should:

  1. change existing #if guards to
     "#if defined(__TARGET_ARCH_arm64) && __clang_major__ >= 18"

  2. additionally, guard "__arena_global" variable definitions behind
     "#if __clang_major >= 18" so that clang-17 doesn't try to compile
     that part (then crash)

Will fix in v3.

Thanks,
Peilin Ye