Re: [RFC PATCH] libbpf: Change hash_combine parameters from long to __u32

[Sidong Yang <sidong.yang@furiosa.ai>]

On Fri, Nov 15, 2024 at 11:57:24AM -0800, Andrii Nakryiko wrote:
> On Fri, Nov 15, 2024 at 2:51 AM Sidong Yang <sidong.yang@furiosa.ai> wrote:
> >
> > The hash_combine() could be trapped when compiled with sanitizer like "zig cc".
> > This patch changes parameters to __u32 to fix it.
> 
> Can you please elaborate? What exactly are you fixing? "Undefined"
> signed integer overflow? I can consider changing long to unsigned
> long, but I don't think we should downgrade from long all the way to
> 32-bit u32. I'd rather keep all those 64 bits for hash.

Hi, Andrii.

Actually I'm using libbpf-rs with maturin build that makes python package for
rust. It seems that it uses zig cc for cross compilation. It compiles libbpf
like this command.

CC="zig cc" make CFLAGS="-fsanitize-trap"

And hash_combine's result is like below.

0000000000063860 <hash_combine>:
   63860:       55                      push   %rbp
   63861:       48 89 e5                mov    %rsp,%rbp
   63864:       48 89 7d f8             mov    %rdi,-0x8(%rbp)
   63868:       48 89 75 f0             mov    %rsi,-0x10(%rbp)
   6386c:       b8 1f 00 00 00          mov    $0x1f,%eax
   63871:       48 0f af 45 f8          imul   -0x8(%rbp),%rax
   63876:       48 89 45 e8             mov    %rax,-0x18(%rbp)
   6387a:       0f 90 c0                seto   %al
   6387d:       34 ff                   xor    $0xff,%al
   6387f:       a8 01                   test   $0x1,%al
   63881:       0f 85 05 00 00 00       jne    6388c <hash_combine+0x2c>
-> 63887:       67 0f b9 40 0c          ud1    0xc(%eax),%eax
   6388c:       48 8b 45 e8             mov    -0x18(%rbp),%rax
   63890:       48 03 45 f0             add    -0x10(%rbp),%rax
   63894:       48 89 45 e0             mov    %rax,-0x20(%rbp)
   63898:       0f 90 c0                seto   %al
   6389b:       34 ff                   xor    $0xff,%al
   6389d:       a8 01                   test   $0x1,%al
   6389f:       0f 85 04 00 00 00       jne    638a9 <hash_combine+0x49>
   638a5:       67 0f b9 00             ud1    (%eax),%eax
   638a9:       48 8b 45 e0             mov    -0x20(%rbp),%rax
   638ad:       5d                      pop    %rbp
   638ae:       c3                      ret   
   638af:       90                      nop

When I'm using libbpf-rs, it receives SIGILL for ud1 instruction.
It seems more appropriate to use u64 instead of u32, doesn't it?
I'll work on it.

Thanks,
Sidong
> 
> pw-bot: cr
> 
> >
> > Signed-off-by: Sidong Yang <sidong.yang@furiosa.ai>
> > ---
> >  tools/lib/bpf/btf.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/tools/lib/bpf/btf.c b/tools/lib/bpf/btf.c
> > index 8befb8103e32..11ccb5aa4958 100644
> > --- a/tools/lib/bpf/btf.c
> > +++ b/tools/lib/bpf/btf.c
> > @@ -3548,7 +3548,7 @@ struct btf_dedup {
> >         struct strset *strs_set;
> >  };
> >
> > -static long hash_combine(long h, long value)
> > +static __u32 hash_combine(__u32 h, __u32 value)
> >  {
> >         return h * 31 + value;
> >  }
> > --
> > 2.42.0
> >
> >