Re: [PATCH bpf-next 2/3] libbpf: Support configurable pinning of maps from BTF annotations - Toke Høiland-Jørgensen

From: "Toke Høiland-Jørgensen" <toke@redhat.com>
To: Andrii Nakryiko <andrii.nakryiko@gmail.com>
Cc: Daniel Borkmann <daniel@iogearbox.net>,
	Alexei Starovoitov <ast@kernel.org>,
	Martin KaFai Lau <kafai@fb.com>, Song Liu <songliubraving@fb.com>,
	Yonghong Song <yhs@fb.com>,
	Jesper Dangaard Brouer <brouer@redhat.com>,
	David Miller <davem@davemloft.net>,
	Networking <netdev@vger.kernel.org>, bpf <bpf@vger.kernel.org>
Subject: Re: [PATCH bpf-next 2/3] libbpf: Support configurable pinning of maps from BTF annotations
Date: Wed, 23 Oct 2019 10:53:58 +0200	[thread overview]
Message-ID: <87r233n8pl.fsf@toke.dk> (raw)
In-Reply-To: <CAEf4BzY-buKFadzzAKpCdjAZ+1_UwSpQobdRH7yQn_fFXQYX0w@mail.gmail.com>

Andrii Nakryiko <andrii.nakryiko@gmail.com> writes:

>> > 4. Once pinned, map knows its pinned path, just use that, I don't see
>> > any reasonable use case where you'd want to override path just for
>> > unpinning.
>>
>> Well, unpinning may need to re-construct the pin path. E.g.,
>> applications that exit after loading and are re-run after unloading,
>> such as iproute2, probably want to be able to unpin maps. Unfortunately
>> I don't think there is a way to get the pin path(s) of an object from
>> the kernel, though, is there? That would be kinda neat for implementing
>> something like `ip link set dev eth0 xdp off unpin`.
>
> Hm... It seems to me that if application exits and another instance
> starts, it should generate pin path using the same logic, then check
> if map is already pinned. Then based on settings, either reuse or
> unpin first. Either way, pin_path needs to be calculated from map
> attributes, not "guessed" by application.

Yeah, ideally. However, the bpf object file may not be available (it's
not for iproute2, for instance). I'm not sure there's really anything we
*can* do about that, though, other than have the application guess.
Unless we add more state to the kernel.

Would it make sense to store the fact that a map was auto-pinned as a
flag in the kernel map info? That way, an application could read that
flag along with the name and go looking in /sys/fs/bpf.

Hmm, but I guess it could do that anyway; so maybe what we need is just
a "try to find all pinned maps of this program" function? That could
then to something like:

- Get the maps IDs and names of all maps attached to that program from
  the kernel.

- Look for each map name in /sys/fs/bpf

- If a pinned map with the same name exists, compare the IDs, and unlink
  if they match

I don't suppose it'll be possible to do all that in a race-free manner,
but that would go for any use of unlink() unless I'm missing something?

-Toke