Development discussion of WireGuard
 help / color / mirror / Atom feed
From: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
To: "Jason A. Donenfeld" <Jason@zx2c4.com>
Cc: WireGuard mailing list <wireguard@lists.zx2c4.com>,
	Netdev <netdev@vger.kernel.org>,
	"David S. Miller" <davem@davemloft.net>,
	Jakub Kicinski <kuba@kernel.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	Peter Zijlstra <peterz@infradead.org>
Subject: Re: [RFC] wiregard RX packet processing.
Date: Tue, 11 Jan 2022 16:40:31 +0100	[thread overview]
Message-ID: <Yd2k7yCGCQXfibk3@linutronix.de> (raw)
In-Reply-To: <CAHmME9rzEjKg41eq5jBtsLXF+vZSEnvdomZJ-rTzx8Q=ac1ayg@mail.gmail.com>

On 2021-12-20 18:29:49 [+0100], Jason A. Donenfeld wrote:
> Hi Sebastian,
> 
> Seems like you've identified two things, the use of need_resched, and
> potentially surrounding napi_schedule in local_bh_{disable,enable}.
> 
> Regarding need_resched, I pulled that out of other code that seemed to
> have the "same requirements", as vaguely conceived. It indeed might
> not be right. The intent is to have that worker running at maximum
> throughput for extended periods of time, but not preventing other
> threads from running elsewhere, so that, e.g., a user's machine
> doesn't have a jenky mouse when downloading a file.
>
> What are the effects of unconditionally calling cond_resched() without
> checking for if (need_resched())? Sounds like you're saying none at
> all?

I stand to be corrected but "if need_resched() cond_resched())" is not
something one should do. If you hold a lock and need to drop it first
and und you don't want to drop the lock if there is no need for
scheduling then there is cond_resched_lock() for instance. If you need
to do something more complex (say set a marker if you drop the lock)
then okay _but_ in this case you do more than just the "if …" from
above.

cond_resched() gets optimized away on a preemptible kernel. The side
effect is that you have always a branch (to cond_resched()) including a
possible RCU section (urgently needed quiescent state).

> Regarding napi_schedule, I actually wasn't aware that it's requirement
> to _only_ ever run from softirq was a strict one. When I switched to
> using napi_schedule in this way, throughput really jumped up
> significantly. Part of this indeed is from the batching, so that the
> napi callback can then handle more packets in one go later. But I
> assumed it was something inside of NAPI that was batching and
> scheduling it, rather than a mistake on my part to call this from a wq
> and not from a softirq.

There is no strict requirement to do napi_schedule() from hard-IRQ but
it makes sense actually. So napi_schedule() invokes
__raise_softirq_irqoff() which only ors a bit in the softirq state.
Nothing else. The only reason that the softirq is invoked in a
deterministic way is that irq_exit() has this "if
(local_softirq_pending()) invoke_softirq()" check before returing (to
interrupted user/ kernel code).

So if you use it in a worker (for instance) the NAPI call is delayed
until the next IRQ (due to irq_exit() part) or a random
local_bh_enable() user.

> What, then, are the effects of surrounding that in
> local_bh_{disable,enable} as you've done in the patch? You mentioned
> one aspect is that it will "invoke wg_packet_rx_poll() where you see
> only one skb." It sounds like that'd be bad for performance, though,
> given that the design of napi is really geared toward batching.

As Toke Høiland-Jørgensen wrote in the previous reply, I missed the BH
disable/ enable in ptr_ring_consume_bh(). So what happens is that
ptr_ring_consume_bh() gives you one skb, you do
wg_queue_enqueue_per_peer_rx() which raises NAPI then the following
ptr_ring_consume_bh() (that local_bh_enable() to be exact) invokes the
NAPI callback (I guess wg_packet_rx_poll() but as I wrote earlier, I
didn't figure out how the skbs move from here to the other queue for
that callback).

So there is probably no batching assuming that one skb is processed in
the NAPI callback.

> Jason

Sebastian

  parent reply	other threads:[~2022-01-11 15:40 UTC|newest]

Thread overview: 5+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-12-08 17:32 Sebastian Andrzej Siewior
2021-12-20 17:29 ` Jason A. Donenfeld
2022-01-05  0:14   ` Toke Høiland-Jørgensen
2022-01-11 15:40   ` Sebastian Andrzej Siewior [this message]
2022-02-10  0:21     ` Feature Request :: Configuration-Option "ospf = yes|no" on Multi-Peer-Interfaces markus

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=Yd2k7yCGCQXfibk3@linutronix.de \
    --to=bigeasy@linutronix.de \
    --cc=Jason@zx2c4.com \
    --cc=davem@davemloft.net \
    --cc=kuba@kernel.org \
    --cc=netdev@vger.kernel.org \
    --cc=peterz@infradead.org \
    --cc=tglx@linutronix.de \
    --cc=wireguard@lists.zx2c4.com \
    --subject='Re: [RFC] wiregard RX packet processing.' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).