Re: [PATCH v2 bpf-next 0/8] bpf: Allow bpf tcp iter to do bpf_(get|set)sockopt

From: Martin KaFai Lau <kafai@fb.com>
To: Eric Dumazet <eric.dumazet@gmail.com>
Cc: <bpf@vger.kernel.org>, Alexei Starovoitov <ast@kernel.org>,
	Daniel Borkmann <daniel@iogearbox.net>,
	Eric Dumazet <edumazet@google.com>, <kernel-team@fb.com>,
	Neal Cardwell <ncardwell@google.com>, <netdev@vger.kernel.org>,
	Yonghong Song <yhs@fb.com>, Yuchung Cheng <ycheng@google.com>
Subject: Re: [PATCH v2 bpf-next 0/8] bpf: Allow bpf tcp iter to do bpf_(get|set)sockopt
Date: Thu, 22 Jul 2021 14:01:44 -0700	[thread overview]
Message-ID: <20210722210144.qrnpycup4rmejvnx@kafai-mbp.dhcp.thefacebook.com> (raw)
In-Reply-To: <d5ffdaf5-08e5-2b28-d891-73d507bae5fa@gmail.com>

On Thu, Jul 22, 2021 at 03:25:39PM +0200, Eric Dumazet wrote:
> 
> 
> On 7/1/21 10:05 PM, Martin KaFai Lau wrote:
> > This set is to allow bpf tcp iter to call bpf_(get|set)sockopt.
> > 
> > With bpf-tcp-cc, new algo rollout happens more often.  Instead of
> > restarting the applications to pick up the new tcp-cc, this set
> > allows the bpf tcp iter to call bpf_(get|set)sockopt(TCP_CONGESTION).
> > It is not limited to TCP_CONGESTION, the bpf tcp iter can call
> > bpf_(get|set)sockopt() with other options.  The bpf tcp iter can read
> > into all the fields of a tcp_sock, so there is a lot of flexibility
> > to select the desired sk to do setsockopt(), e.g. it can test for
> > TCP_LISTEN only and leave the established connections untouched,
> > or check the addr/port, or check the current tcp-cc name, ...etc.
> > 
> > Patch 1-4 are some cleanup and prep work in the tcp and bpf seq_file.
> > 
> > Patch 5 is to have the tcp seq_file iterate on the
> > port+addr lhash2 instead of the port only listening_hash.
> > 
> > Patch 6 is to have the bpf tcp iter doing batching which
> > then allows lock_sock.  lock_sock is needed for setsockopt.
> > 
> > Patch 7 allows the bpf tcp iter to call bpf_(get|set)sockopt.
> > 
> > v2:
> > - Use __GFP_NOWARN in patch 6
> > - Add bpf_getsockopt() in patch 7 to give a symmetrical user experience.
> >   selftest in patch 8 is changed to also cover bpf_getsockopt().
> > - Remove CAP_NET_ADMIN check in patch 7. Tracing bpf prog has already
> >   required CAP_SYS_ADMIN or CAP_PERFMON.
> > - Move some def macros to bpf_tracing_net.h in patch 8
> > 
> > Martin KaFai Lau (8):
> >   tcp: seq_file: Avoid skipping sk during tcp_seek_last_pos
> >   tcp: seq_file: Refactor net and family matching
> >   bpf: tcp: seq_file: Remove bpf_seq_afinfo from tcp_iter_state
> >   tcp: seq_file: Add listening_get_first()
> >   tcp: seq_file: Replace listening_hash with lhash2
> >   bpf: tcp: bpf iter batching and lock_sock
> >   bpf: tcp: Support bpf_(get|set)sockopt in bpf tcp iter
> >   bpf: selftest: Test batching and bpf_(get|set)sockopt in bpf tcp iter
> 
> For the whole series :
> 
> Reviewed-by: Eric Dumazet <edumazet@google.com>
> 
> Sorry for the delay.
> 
> BTW, it seems weird for new BPF features to use /proc/net "legacy"
> infrastructure and update it.
bpf iter uses seq_file, so the initial bpf_iter_tcp reuses most
of the pieces from /proc/net/tcp.

This set refactored a few things such that the bpf_iter_tcp only
shares the legacy tcp_seek_last_pos(), so the dependency on
/proc/net/tcp should be less going forward.

A similar modification could also be done to bpf_iter_udp in the future.

Thanks for the review!