2020-05-08 06:42:02

by Ian Rogers

Subject: [PATCH] libbpf hashmap: fix undefined behavior in hash_bits

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.

Signed-off-by: Ian Rogers <[email protected]>
---
tools/lib/bpf/hashmap.h | 2 ++
1 file changed, 2 insertions(+)

diff --git a/tools/lib/bpf/hashmap.h b/tools/lib/bpf/hashmap.h
index d5ef212a55ba..781db653d16c 100644
--- a/tools/lib/bpf/hashmap.h
+++ b/tools/lib/bpf/hashmap.h
@@ -19,6 +19,8 @@
static inline size_t hash_bits(size_t h, int bits)
{
/* shuffle bits and return requested number of upper bits */
+ if (bits == 0)
+ return 0;
return (h * 11400714819323198485llu) >> (__WORDSIZE - bits);
}

--
2.26.2.645.ge9eca65c58-goog

2020-05-08 07:14:06

by Andrii Nakryiko

Subject: Re: [PATCH] libbpf hashmap: fix undefined behavior in hash_bits

On Thu, May 7, 2020 at 11:40 PM Ian Rogers <[email protected]> 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.
>
> Signed-off-by: Ian Rogers <[email protected]>
> ---

No need. The only case when bits can be 0 is when hashmap is
completely empty (no elements have ever been added yet). In that case,
it doesn't matter what value hash_bits() returns,
hashmap__for_each_key_entry/hashmap__for_each_key_entry_safe will
behave correctly, because map->buckets will be NULL.

> tools/lib/bpf/hashmap.h | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/tools/lib/bpf/hashmap.h b/tools/lib/bpf/hashmap.h
> index d5ef212a55ba..781db653d16c 100644
> --- a/tools/lib/bpf/hashmap.h
> +++ b/tools/lib/bpf/hashmap.h
> @@ -19,6 +19,8 @@
> static inline size_t hash_bits(size_t h, int bits)
> {
> /* shuffle bits and return requested number of upper bits */
> + if (bits == 0)
> + return 0;
> return (h * 11400714819323198485llu) >> (__WORDSIZE - bits);
> }
>
> --
> 2.26.2.645.ge9eca65c58-goog
>

2020-05-08 07:24:43

by Ian Rogers

Subject: Re: [PATCH] libbpf hashmap: fix undefined behavior in hash_bits

On Fri, May 8, 2020 at 12:12 AM Andrii Nakryiko
<[email protected]> wrote:
>
> On Thu, May 7, 2020 at 11:40 PM Ian Rogers <[email protected]> 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.
> >
> > Signed-off-by: Ian Rogers <[email protected]>
> > ---
>
> No need. The only case when bits can be 0 is when hashmap is
> completely empty (no elements have ever been added yet). In that case,
> it doesn't matter what value hash_bits() returns,
> hashmap__for_each_key_entry/hashmap__for_each_key_entry_safe will
> behave correctly, because map->buckets will be NULL.

Agreed. Unfortunately the LLVM undefined behavior sanitizer (I've not
tested with GCC to the same extent) will cause an exit when it sees >>
64 regardless of whether the value is used or not. It'd be possible to
#ifdef this code on whether a sanitizer was present.

Thanks,
Ian

> > tools/lib/bpf/hashmap.h | 2 ++
> > 1 file changed, 2 insertions(+)
> >
> > diff --git a/tools/lib/bpf/hashmap.h b/tools/lib/bpf/hashmap.h
> > index d5ef212a55ba..781db653d16c 100644
> > --- a/tools/lib/bpf/hashmap.h
> > +++ b/tools/lib/bpf/hashmap.h
> > @@ -19,6 +19,8 @@
> > static inline size_t hash_bits(size_t h, int bits)
> > {
> > /* shuffle bits and return requested number of upper bits */
> > + if (bits == 0)
> > + return 0;
> > return (h * 11400714819323198485llu) >> (__WORDSIZE - bits);
> > }
> >
> > --
> > 2.26.2.645.ge9eca65c58-goog
> >

2020-05-08 18:06:46

by Andrii Nakryiko

Subject: Re: [PATCH] libbpf hashmap: fix undefined behavior in hash_bits

On Fri, May 8, 2020 at 12:21 AM Ian Rogers <[email protected]> wrote:
>
> On Fri, May 8, 2020 at 12:12 AM Andrii Nakryiko
> <[email protected]> wrote:
> >
> > On Thu, May 7, 2020 at 11:40 PM Ian Rogers <[email protected]> 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.
> > >
> > > Signed-off-by: Ian Rogers <[email protected]>
> > > ---
> >
> > No need. The only case when bits can be 0 is when hashmap is
> > completely empty (no elements have ever been added yet). In that case,
> > it doesn't matter what value hash_bits() returns,
> > hashmap__for_each_key_entry/hashmap__for_each_key_entry_safe will
> > behave correctly, because map->buckets will be NULL.
>
> Agreed. Unfortunately the LLVM undefined behavior sanitizer (I've not
> tested with GCC to the same extent) will cause an exit when it sees >>
> 64 regardless of whether the value is used or not. It'd be possible to
> #ifdef this code on whether a sanitizer was present.

Yeah, let's do that rather than slowing down hashing function.

>
> Thanks,
> Ian
>
> > > tools/lib/bpf/hashmap.h | 2 ++
> > > 1 file changed, 2 insertions(+)
> > >
> > > diff --git a/tools/lib/bpf/hashmap.h b/tools/lib/bpf/hashmap.h
> > > index d5ef212a55ba..781db653d16c 100644
> > > --- a/tools/lib/bpf/hashmap.h
> > > +++ b/tools/lib/bpf/hashmap.h
> > > @@ -19,6 +19,8 @@
> > > static inline size_t hash_bits(size_t h, int bits)
> > > {
> > > /* shuffle bits and return requested number of upper bits */
> > > + if (bits == 0)
> > > + return 0;
> > > return (h * 11400714819323198485llu) >> (__WORDSIZE - bits);
> > > }
> > >
> > > --
> > > 2.26.2.645.ge9eca65c58-goog
> > >