Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Thu, 21 Mar 2019 13:16:06 -0700
From:   Jakub Kicinski <jakub.kicinski@netronome.com>
To:     Alban Crequy <alban@kinvolk.io>
Cc:     Alban Crequy <alban.crequy@gmail.com>,
        Alexei Starovoitov <ast@kernel.org>,
        Daniel Borkmann <daniel@iogearbox.net>,
        netdev <netdev@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Iago =?UTF-8?B?TMOzcGV6?= Galeiras <iago@kinvolk.io>
Subject: Re: [PATCH bpf-next v1 4/7] tools: bpftool: implement map exec
 command
Message-ID: <20190321131606.2acb3bb6@cakuba.netronome.com>
In-Reply-To: <CADZs7q7+SATGjCCdiDBfGbTBJh11V6s=o4RAAKhaKvFjfo7O_Q@mail.gmail.com>
References: <20190320173332.18105-1-alban@kinvolk.io>
        <20190320173332.18105-4-alban@kinvolk.io>
        <20190320142346.3b552895@cakuba.netronome.com>
        <CADZs7q7+SATGjCCdiDBfGbTBJh11V6s=o4RAAKhaKvFjfo7O_Q@mail.gmail.com>
Organization: Netronome Systems, Ltd.
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, 21 Mar 2019 10:57:32 +0100, Alban Crequy wrote:
> On Wed, Mar 20, 2019 at 10:23 PM Jakub Kicinski
> <jakub.kicinski@netronome.com> wrote:
> >
> > On Wed, 20 Mar 2019 18:33:29 +0100, Alban Crequy wrote:  
> > > From: Alban Crequy <alban@kinvolk.io>
> > >
> > > The map exec commands allows to open an existing map and pass the file
> > > descriptor to a child process. This enables applications to use an
> > > existing BPF map even when they don't support bpffs.
> > >
> > > Example of usage:
> > >     # bpftool map exec pinned /sys/fs/bpf/foo fd 99 cmd -- readlink /proc/self/fd/99
> > >     anon_inode:bpf-map  
> >
> > Would you mind telling us a little more about the use for this feature?
> > It seems fairly limited.  If it's about probing objects (finding out if
> > they are a map or a program) perhaps we can add a command just for that?  
> 
> I needed to know the name of the map too. I was preparing a demo based
> on python bcc tools (opensnoop) but with added feature that requires
> using a pinned map, created and maintained externally. At the moment,
> the python API for bcc does not support pinning or using external
> maps. Ideally, this should be added in the python API (some discussion
> on https://github.com/iovisor/bcc/issues/2223) but meanwhile, I use a
> workaround by executing bpftool from the python code.
> 
> Arguably, my use case is a temporary hack until we have better support
> in python bcc. But other tools implements similar commands to pass
> file descriptors between processes: "ip netns exec" and "tc exec bpf".
> So I think it could be useful for other scripting use cases.

The thing is the receiver of the FD has to be bpf-aware, because there
isn't really much one can do with that file descriptor, in which case
it's kind of strange that the receiver doesn't know how to open a
pinned object..

> In my demo, I used the two hacks:
> - if the pinned map fd is not given to the python script, re-execute
> itself with bpftool:
> os.execvp("bpftool", ["bpftool", "map", "exec", "pinned", pin_path,
> "fd", "90", "cmd", "--"] + sys.argv)
> - once we have the fd 90 (number specified above) of the pinned map in
> the python script, overwrite the empty fd created by bcc:
> os.dup2(90, 6)
> I call dup2() between the bpf map creation and the bpf program
> creation. To check which map fd to overwrite, I just call
> os.system("bpftool map show fd 6...").

I see, thanks for the explanation.  That does indeed seems like a hack.

> Thanks a lot for the reviews. I'll need some time to address it (maybe
> a week or 2).
> 
> > (I guess bpftool -f isn't really the cleanest way of getting at that
> > info.)
> >  
> > > Documentation and bash completion updated as well.
> > >
> > > Signed-off-by: Alban Crequy <alban@kinvolk.io>