Re: [PATCH bpf-next 2/3] bpf: Add size arg to build_id_parse function

[Jiri Olsa]

On Thu, Jan 14, 2021 at 07:47:20PM -0800, Alexei Starovoitov wrote:
> On Thu, Jan 14, 2021 at 3:44 PM Yonghong Song <yhs@xxxxxx> wrote:
> >
> >
> >
> > On 1/14/21 2:02 PM, Jiri Olsa wrote:
> > > On Thu, Jan 14, 2021 at 01:05:33PM -0800, Yonghong Song wrote:
> > >>
> > >>
> > >> On 1/14/21 12:01 PM, Jiri Olsa wrote:
> > >>> On Thu, Jan 14, 2021 at 10:56:33AM -0800, Yonghong Song wrote:
> > >>>>
> > >>>>
> > >>>> On 1/14/21 5:40 AM, Jiri Olsa wrote:
> > >>>>> It's possible to have other build id types (other than default SHA1).
> > >>>>> Currently there's also ld support for MD5 build id.
> > >>>>
> > >>>> Currently, bpf build_id based stackmap does not returns the size of
> > >>>> the build_id. Did you see an issue here? I guess user space can check
> > >>>> the length of non-zero bits of the build id to decide what kind of
> > >>>> type it is, right?
> > >>>
> > >>> you can have zero bytes in the build id hash, so you need to get the size
> > >>>
> > >>> I never saw MD5 being used in practise just SHA1, but we added the
> > >>> size to be complete and make sure we'll fit with build id, because
> > >>> there's only limited space in mmap2 event
> > >>
> > >> I am asking to check whether we should extend uapi struct
> > >> bpf_stack_build_id to include build_id_size as well. I guess
> > >> we can delay this until a real use case.
> > >
> > > right, we can try make some MD5 build id binaries and check if it
> > > explodes with some bcc tools, but I don't expect that.. I'll try
> > > to find some time for that
> >
> > Thanks. We may have issues on bcc side. For build_id collected in
> > kernel, bcc always generates a length-20 string. But for user
> > binaries, the build_id string length is equal to actual size of
> > the build_id. They may not match (MD5 length is 16).
> > The fix is probably to append '0's (up to length 20) for user
> > binary build_id's.
> >
> > I guess MD5 is very seldom used. I will wait if you can reproduce
> > the issue and then we might fix it.
> 
> Indeed.
> Jiri, please check whether md5 is really an issue.
> Sounds like we have to do something on the kernel side.
> Hopefully zero padding will be enough.
> I would prefer to avoid extending uapi struct to cover rare case.

build_id_parse is already doing the zero padding, so we are ok

I tried several bcc tools over perf bench with md5 buildid and
the results looked ok

jirka