Re: [PATCH bpf-next 1/6] bpf: implement BPF ring buffer and verifier support for it

From: Andrii Nakryiko <andrii.nakryiko@gmail.com>
To: Stanislav Fomichev <sdf@google.com>
Cc: Andrii Nakryiko <andriin@fb.com>, bpf <bpf@vger.kernel.org>,
	Networking <netdev@vger.kernel.org>,
	Alexei Starovoitov <ast@fb.com>,
	Daniel Borkmann <daniel@iogearbox.net>,
	Kernel Team <kernel-team@fb.com>,
	"Paul E . McKenney" <paulmck@kernel.org>,
	Jonathan Lemon <jonathan.lemon@gmail.com>
Subject: Re: [PATCH bpf-next 1/6] bpf: implement BPF ring buffer and verifier support for it
Date: Thu, 14 May 2020 14:13:00 -0700	[thread overview]
Message-ID: <CAEf4BzbvjQy+8T43e91OXDaLgWsy5_1RSr278=uAVUGOT0LgZw@mail.gmail.com> (raw)
In-Reply-To: <20200514205348.GB161830@google.com>

On Thu, May 14, 2020 at 1:53 PM <sdf@google.com> wrote:
>
> On 05/14, Andrii Nakryiko wrote:
> > On Thu, May 14, 2020 at 10:33 AM <sdf@google.com> wrote:
> > >
> > > On 05/13, Andrii Nakryiko wrote:
>
> [...]
>
> > > > + * void bpf_ringbuf_submit(void *data)
> > > > + *   Description
> > > > + *           Submit reserved ring buffer sample, pointed to by
> > *data*.
> > > > + *   Return
> > > > + *           Nothing.
> > > Even though you mention self-pacing properties, would it still
> > > make sense to add some argument to bpf_ringbuf_submit/bpf_ringbuf_output
> > > to indicate whether to wake up userspace or not? Maybe something like
> > > a threshold of number of outstanding events in the ringbuf after which
> > > we do the wakeup? The default 0/1 preserve the existing behavior.
> > >
> > > The example I can give is a control plane userspace thread that
> > > once a second aggregates the events, it doesn't care about millisecond
> > > resolution. With the current scheme, I suppose, if BPF generates events
> > > every 1ms, the userspace will be woken up 1000 times (if it can keep
> > > up). Most of the time, we don't really care and some buffering
> > > properties are desired.
>
> > perf buffer has setting like this, and believe me, it's so confusing
> > and dangerous, that I wouldn't want this to be exposed. Even though I
> > was aware of this behavior, I still had to debug and work-around this
> > lack on wakeup few times, it's really-really confusing feature.
>
> > In your case, though, why wouldn't user-space poll data just once a
> > second, if it's not interested in getting data as fast as possible?
> If I poll once per second I might lose the events if, for some reason,
> there is a spike. I really want to have something like: "wakeup
> userspace if the ringbuffer fill is over some threshold or
> the last wakeup was too long ago". We currently do this via a percpu
> cache map. IIRC, you've shared on lsfmmbpf that you do something like
> that as well.

Hm... don't remember such use case on our side. All applications I
know of use default perf_buffer settings with no sampling.

>
> So I was thinking how I can use new ringbuff to remove the unneeded
> copies and help with the reordering, but I'm a bit concerned about
> regressing on the number of wakeups.
>
> Maybe having a flag like RINGBUF_NO_WAKEUP in bpf_ringbuf_submit()
> will suffice? And if there is a helper or some way to obtain a
> number of unconsumed items, I can implement my own flushing policy.

Ok, I guess giving application control at each discard/commit makes
for ultimate flexibility. Let me add flags argument to commit/discard
and allow to specify NO_WAKEUP flag. As for count of unconsumed events
-- that would be a bit expensive to maintain. How about amount of data
that's not consumed? It's obviously going to be racy, but returning
(producer_pos - consumer_pos) should be sufficient enough for such
smart and best-effort heuristics? WDYT?