From:   Toke =?utf-8?Q?H=C3=B8iland-J=C3=B8rgensen?= <toke@redhat.com>
To:     Johannes Berg <johannes@sipsolutions.net>,
        Herbert Xu <herbert@gondor.apana.org.au>
Cc:     Arend Van Spriel <arend.vanspriel@broadcom.com>,
        Felix Fietkau <nbd@nbd.name>, linux-wireless@vger.kernel.org,
        Eric Dumazet <eric.dumazet@gmail.com>, netdev@vger.kernel.org
Subject: Re: [PATCH 5/5] mac80211: set NETIF_F_LLTX when using intermediate tx queues
In-Reply-To: <e4c1a3ca0c0c160b7c271ab83deebcf096f38497.camel@sipsolutions.net>
References: <20190316170634.13125-1-nbd@nbd.name> <20190316170634.13125-5-nbd@nbd.name> <87sgvmvg9g.fsf@toke.dk> <773e4dff-29fd-22b4-e4bc-cd5a94c66dc2@broadcom.com> <20190416074444.pdubbh6fbibdnhi7@gondor.apana.org.au> <73b18131-2777-da5c-a6ee-9d9b3e13cd06@broadcom.com> <20190416083636.5ttvezqyhzr2worg@gondor.apana.org.au> <eab39dd804405854997e50102f3d0ff925feb8d2.camel@sipsolutions.net> <87ef62uwfm.fsf@toke.dk> <95f86cf69dee05a176625925657cf0df0e97b5c9.camel@sipsolutions.net> <20190416093707.dtlwcmitzqopaeaw@gondor.apana.org.au> <e4c1a3ca0c0c160b7c271ab83deebcf096f38497.camel@sipsolutions.net>
Date:   Tue, 16 Apr 2019 11:02:50 +0100
Message-ID: <8736miuv1x.fsf@toke.dk>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 8BIT
Sender: linux-wireless-owner@vger.kernel.org
Precedence: bulk

Johannes Berg <johannes@sipsolutions.net> writes:

> On Tue, 2019-04-16 at 17:37 +0800, Herbert Xu wrote:
>> On Tue, Apr 16, 2019 at 11:33:50AM +0200, Johannes Berg wrote:
>> > On Tue, 2019-04-16 at 10:33 +0100, Toke Høiland-Jørgensen wrote:
>> > > 
>> > > > It is true because we have an entire buffering layer in mac80211 (in
>> > > > this case at least) and never push back to the stack.
>> > > 
>> > > I'm wondering if we should be?
>> > 
>> > I don't think so? We'd just buffer packets in yet another place.
>> 
>> But you do realise that you're giving up on the rich queueing
>> functionality that Linux provides (net/sched),
>
> Yes, that was a trade-off we always knew about. The model that Linux
> provides is just not suited for wifi.

As explained at great length here:
https://www.usenix.org/conference/atc17/technical-sessions/presentation/hoilan-jorgesen
(you already know that of course, Johannes)

>> not to mention
>> breaking certain applications that rely on congestion feedback?
>
> This I don't understand. The congestion feedback happens through socket
> buffer space etc. which is still there (as long as nobody sneaks in an
> skb_orphan() call)

Sure, for TCP, the TSQ mechanism should keep the upper-level queue low
as long as the SKBs are alive. But is this also the case for UDP?

-Toke