Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20201029160938.154084-1-irogers@google.com> <497F86C5-BD00-4C38-BD87-C6EFB92D1088@fb.com>
In-Reply-To: <497F86C5-BD00-4C38-BD87-C6EFB92D1088@fb.com>
From:   Ian Rogers <irogers@google.com>
Date:   Thu, 29 Oct 2020 12:37:20 -0700
Message-ID: <CAP-5=fUVSSWEYWWswi19nQHY-b5Vn8-oi7uvtXWnmo1usLOzNw@mail.gmail.com>
Subject: Re: [PATCH] libbpf hashmap: Fix undefined behavior in hash_bits
To:     Song Liu <songliubraving@fb.com>
Cc:     Alexei Starovoitov <ast@kernel.org>,
        Daniel Borkmann <daniel@iogearbox.net>,
        Martin Lau <kafai@fb.com>, Yonghong Song <yhs@fb.com>,
        Andrii Nakryiko <andriin@fb.com>,
        John Fastabend <john.fastabend@gmail.com>,
        KP Singh <kpsingh@chromium.org>,
        "netdev@vger.kernel.org" <netdev@vger.kernel.org>,
        "bpf@vger.kernel.org" <bpf@vger.kernel.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Thu, Oct 29, 2020 at 10:45 AM Song Liu <songliubraving@fb.com> wrote:
>
> > On Oct 29, 2020, at 9:09 AM, Ian Rogers <irogers@google.com> wrote:
> >
> > If bits is 0, the case when the map is empty, then the >> is the size of
> > the register which is undefined behavior - on x86 it is the same as a
> > shift by 0. Fix by handling the 0 case explicitly when running with
> > address sanitizer.
> >
> > A variant of this patch was posted previously as:
> > https://lore.kernel.org/lkml/20200508063954.256593-1-irogers@google.com/
> >
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
> > tools/lib/bpf/hashmap.h | 14 ++++++++++++++
> > 1 file changed, 14 insertions(+)
> >
> > diff --git a/tools/lib/bpf/hashmap.h b/tools/lib/bpf/hashmap.h
> > index d9b385fe808c..27d0556527d3 100644
> > --- a/tools/lib/bpf/hashmap.h
> > +++ b/tools/lib/bpf/hashmap.h
> > @@ -12,9 +12,23 @@
> > #include <stddef.h>
> > #include <limits.h>
> >
> > +#ifdef __has_feature
> > +#define HAVE_FEATURE(f) __has_feature(f)
> > +#else
> > +#define HAVE_FEATURE(f) 0
> > +#endif
> > +
> > static inline size_t hash_bits(size_t h, int bits)
> > {
> >       /* shuffle bits and return requested number of upper bits */
> > +#if defined(ADDRESS_SANITIZER) || HAVE_FEATURE(address_sanitizer)
>
> I am not very familiar with these features. Is address sanitizer same
> as undefined behavior sanitizer (mentioned in previous version)?

My preference would be to special case bits == 0 without the feature
guards as per the original change, this is the most correct. There is
some feature support for detecting ubsan:
https://github.com/google/sanitizers/issues/765
In my case I see this with address sanitizer and older versions of
clang don't expose ubsan as a feature.

> > +     /*
> > +      * If the requested bits == 0 avoid undefined behavior from a
> > +      * greater-than bit width shift right (aka invalid-shift-exponent).
> > +      */
> > +     if (bits == 0)
> > +             return -1;
>
> Shall we return 0 or -1 (0xffffffff) here?

The value isn't used and so doesn't matter. -1 seemed less likely to
silently succeed.

> Also, we have HASHMAP_MIN_CAP_BITS == 2. Shall we just make sure we
> never feed bits == 0 into hash_bits()?

I think that'd be a different change. I'd be happy to see it.

Thanks,
Ian

> Thanks,
> Song
>
>
> > +#endif
> > #if (__SIZEOF_SIZE_T__ == __SIZEOF_LONG_LONG__)
> >       /* LP64 case */
> >       return (h * 11400714819323198485llu) >> (__SIZEOF_LONG_LONG__ * 8 - bits);
> > --
> > 2.29.1.341.ge80a0c044ae-goog
> >
>