Received-SPF: pass (google.com: domain of linux-wireless-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <b2d6360f289ea0de4985a2909919049ff0f53e2b.camel@sipsolutions.net>
Subject: Re: [RFC] mac80211: Switch all drivers over to iTXQs
From:   Johannes Berg <johannes@sipsolutions.net>
To:     Alexander Wetzel <alexander@wetzel-home.de>
Cc:     linux-wireless@vger.kernel.org
Date:   Fri, 08 Jul 2022 23:11:56 +0200
In-Reply-To: <20220707143230.21686-1-alexander@wetzel-home.de>
References: <20220707143230.21686-1-alexander@wetzel-home.de>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
User-Agent: Evolution 3.44.3 (3.44.3-1.fc36) 
MIME-Version: 1.0
Precedence: bulk

Nice!

I see your follow-up, but a few comments anyway here on this version.

> +	/* Driver has no native support for iTXQ, handle the queues */
> +	ieee80211_txq_schedule_start(&local->hw, txq->txq.ac);
> +	while ((queue =3D ieee80211_next_txq(&local->hw, txq->txq.ac)) !=3D NUL=
L) {
> +		control.sta =3D queue->sta;
> +		while ((skb =3D ieee80211_tx_dequeue(&local->hw, queue)))
> +			drv_tx(local, &control, skb);

Yeah, I think I see how that might not work right - you don't
necessarily want to pull out everything here.

But it should be easy to break out here, I guess, and continue if the
driver restarts the corresponding queue?

> -/* Indicate which queue to use. */
>  u16 ieee80211_select_queue(struct ieee80211_sub_if_data *sdata,
>  			   struct sk_buff *skb)
>  {
> -[...]
> +	/* mac80211 only supports iTXQ nowadays. Only calculate the hash while
> +	 * the skb is still hot in the CPU cache
> +	 */
> +	skb_get_hash(skb);
> +	return 0;

Hmm. Is that really worth keeping? I'd probably rather remove the
netdev_select_queue callbacks completely.

Regardless, once this works I'm sure we can make a lot of cleanups
elsewhere, e.g. a couple of "!txq" checks in tx.c (tx_prep_agg,
amsdu_aggregate).


Eventually with this implementation I think it'd be worth doing it for
all the cases where ieee80211_get_txq() returns NULL or for where it's
skipped in ieee80211_queue_skb(), i.e.

 * monitor mode,
 * management frames,
 * after-DTIM
 * PS response
 *=C2=A0STA not uploaded

as well, and just going into the mac80211 iTXQ scheduling code you wrote
in wake_tx_queue() for those cases.

That way, we could get rid of all the special cases in
ieee80211_queue_skb() and callers that depend on the return value.

That would allow us to remove a LOT of code, all the stuff handling
local->pending (tx_pending), tx_frags, etc.

johannes