[PATCH] mac80211: don't apply flow control on management frames

[PATCH] mac80211: don't apply flow control on management frames

[Johannes Berg]

From: Johannes Berg <johannes.berg@intel.com>

In some cases (depending on the driver, but it's true e.g. for
iwlwifi) we're using an internal TXQ for management packets,
mostly to simplify the code and to have a place to queue them.
However, it appears that in certain cases we can confuse the
code and management frames are dropped, which is certainly not
what we want.

Short-circuit the processing of management frames. To keep the
impact minimal, only put them on the frags queue and check the
tid == management only for doing that and to skip the airtime
fairness checks, if applicable.

Signed-off-by: Johannes Berg <johannes.berg@intel.com>
---
 net/mac80211/tx.c | 18 +++++++++++++++---
 1 file changed, 15 insertions(+), 3 deletions(-)

diff --git a/net/mac80211/tx.c b/net/mac80211/tx.c
index 5d06de61047a..b2d09acb9fb0 100644
--- a/net/mac80211/tx.c
+++ b/net/mac80211/tx.c
@@ -5,7 +5,7 @@
  * Copyright 2006-2007	Jiri Benc <jbenc@suse.cz>
  * Copyright 2007	Johannes Berg <johannes@sipsolutions.net>
  * Copyright 2013-2014  Intel Mobile Communications GmbH
- * Copyright (C) 2018-2020 Intel Corporation
+ * Copyright (C) 2018-2021 Intel Corporation
  *
  * Transmit and frame generation functions.
  */
@@ -1388,8 +1388,17 @@ static void ieee80211_txq_enqueue(struct ieee80211_local *local,
 	ieee80211_set_skb_enqueue_time(skb);
 
 	spin_lock_bh(&fq->lock);
-	fq_tin_enqueue(fq, tin, flow_idx, skb,
-		       fq_skb_free_func);
+	/*
+	 * For management frames, don't really apply codel etc.,
+	 * we don't want to apply any shaping or anything we just
+	 * want to simplify the driver API by having them on the
+	 * txqi.
+	 */
+	if (unlikely(txqi->txq.tid == IEEE80211_NUM_TIDS))
+		__skb_queue_tail(&txqi->frags, skb);
+	else
+		fq_tin_enqueue(fq, tin, flow_idx, skb,
+			       fq_skb_free_func);
 	spin_unlock_bh(&fq->lock);
 }
 
@@ -3835,6 +3844,9 @@ bool ieee80211_txq_airtime_check(struct ieee80211_hw *hw,
 	if (!txq->sta)
 		return true;
 
+	if (unlikely(txq->tid == IEEE80211_NUM_TIDS))
+		return true;
+
 	sta = container_of(txq->sta, struct sta_info, sta);
 	if (atomic_read(&sta->airtime[txq->ac].aql_tx_pending) <
 	    sta->airtime[txq->ac].aql_limit_low)
-- 
2.30.2

Re: [PATCH] mac80211: don't apply flow control on management frames

[Toke Høiland-Jørgensen]

Johannes Berg <johannes@sipsolutions.net> writes:

> From: Johannes Berg <johannes.berg@intel.com>
>
> In some cases (depending on the driver, but it's true e.g. for
> iwlwifi) we're using an internal TXQ for management packets,
> mostly to simplify the code and to have a place to queue them.
> However, it appears that in certain cases we can confuse the
> code and management frames are dropped, which is certainly not
> what we want.
>
> Short-circuit the processing of management frames. To keep the
> impact minimal, only put them on the frags queue and check the
> tid == management only for doing that and to skip the airtime
> fairness checks, if applicable.
>
> Signed-off-by: Johannes Berg <johannes.berg@intel.com>
> ---
>  net/mac80211/tx.c | 18 +++++++++++++++---
>  1 file changed, 15 insertions(+), 3 deletions(-)
>
> diff --git a/net/mac80211/tx.c b/net/mac80211/tx.c
> index 5d06de61047a..b2d09acb9fb0 100644
> --- a/net/mac80211/tx.c
> +++ b/net/mac80211/tx.c
> @@ -5,7 +5,7 @@
>   * Copyright 2006-2007	Jiri Benc <jbenc@suse.cz>
>   * Copyright 2007	Johannes Berg <johannes@sipsolutions.net>
>   * Copyright 2013-2014  Intel Mobile Communications GmbH
> - * Copyright (C) 2018-2020 Intel Corporation
> + * Copyright (C) 2018-2021 Intel Corporation
>   *
>   * Transmit and frame generation functions.
>   */
> @@ -1388,8 +1388,17 @@ static void ieee80211_txq_enqueue(struct ieee80211_local *local,
>  	ieee80211_set_skb_enqueue_time(skb);
>  
>  	spin_lock_bh(&fq->lock);
> -	fq_tin_enqueue(fq, tin, flow_idx, skb,
> -		       fq_skb_free_func);
> +	/*
> +	 * For management frames, don't really apply codel etc.,
> +	 * we don't want to apply any shaping or anything we just
> +	 * want to simplify the driver API by having them on the
> +	 * txqi.
> +	 */
> +	if (unlikely(txqi->txq.tid == IEEE80211_NUM_TIDS))
> +		__skb_queue_tail(&txqi->frags, skb);
> +	else
> +		fq_tin_enqueue(fq, tin, flow_idx, skb,
> +			       fq_skb_free_func);

One consequence of this is that we create a strict priority queue for
management frames. With all the possibilities for badness (such as the
ability of starving all other queues) that carries with it. I guess
that's probably fine for management frames, though, right? As in, there
is some other mechanism that prevents abuse of this?

-Toke

Re: [PATCH] mac80211: don't apply flow control on management frames

[Johannes Berg]

On Sat, 2021-03-20 at 01:13 +0100, Toke Høiland-Jørgensen wrote:
> 
> One consequence of this is that we create a strict priority queue for
> management frames.

Yes, for iwlwifi that's actually a change. For everyone else (not
setting BUFF_MMPDU_TXQ or STA_MMPDU_TXQ) it already is since it goes
directly to ->tx() and from there to the hardware queue(s).

> With all the possibilities for badness (such as the
> ability of starving all other queues) that carries with it. I guess
> that's probably fine for management frames, though, right? As in, there
> is some other mechanism that prevents abuse of this?

Well, there's just not that many management frames to start with? And
only wpa_supplicant (or root in general) can create them. So I don't
think we need to worry about that yet. With QoS-mgmt frames we might
eventually want to think about that, but even there I'd say we never
really want to drop them.

johannes

Re: [PATCH] mac80211: don't apply flow control on management frames

[Toke Høiland-Jørgensen]

Johannes Berg <johannes@sipsolutions.net> writes:

> On Sat, 2021-03-20 at 01:13 +0100, Toke Høiland-Jørgensen wrote:
>> 
>> One consequence of this is that we create a strict priority queue for
>> management frames.
>
> Yes, for iwlwifi that's actually a change. For everyone else (not
> setting BUFF_MMPDU_TXQ or STA_MMPDU_TXQ) it already is since it goes
> directly to ->tx() and from there to the hardware queue(s).

Ah, right, of course; so not much change at all. Cool.

>> With all the possibilities for badness (such as the
>> ability of starving all other queues) that carries with it. I guess
>> that's probably fine for management frames, though, right? As in, there
>> is some other mechanism that prevents abuse of this?
>
> Well, there's just not that many management frames to start with? And
> only wpa_supplicant (or root in general) can create them. So I don't
> think we need to worry about that yet. With QoS-mgmt frames we might
> eventually want to think about that, but even there I'd say we never
> really want to drop them.

Yup, that's what I meant with "some other mechanism to prevent abuse".
Great.

Feel free to add my:

Acked-by: Toke Høiland-Jørgensen <toke@redhat.com>

Re: [PATCH] mac80211: don't apply flow control on management frames

[Maxime Bizon]

On Fri, 2021-03-19 at 23:28 +0100, Johannes Berg wrote:

Hello Johannes,

> 
> Short-circuit the processing of management frames. To keep the
> impact minimal, only put them on the frags queue and check the
> tid == management only for doing that and to skip the airtime
> fairness checks, if applicable.

After your patch, what are the actual effects of HW_STA_MMPDU_TXQ and
HW_BUFF_MMPDU_TXQ ?

Thanks,

-- 
Maxime

Re: [PATCH] mac80211: don't apply flow control on management frames

[Johannes Berg]

Hi,

On Mon, 2021-03-22 at 10:43 +0100, Maxime Bizon wrote:
> > Short-circuit the processing of management frames. To keep the
> > impact minimal, only put them on the frags queue and check the
> > tid == management only for doing that and to skip the airtime
> > fairness checks, if applicable.
> 
> After your patch, what are the actual effects of HW_STA_MMPDU_TXQ and
> HW_BUFF_MMPDU_TXQ ?

Well, my patch doesn't change the effect of those significantly.

The idea for iwlwifi was that it doesn't actually like ->tx() to get
called, but much prefers a TXQ where the frame is, and then it can pull
it whenever it can transmit it.

This was the key requirement here, and it doesn't change: instead of
tx() getting called with the frames, the frames go to the TXQ instead
and wake_tx_queue() is called, and then the driver later pulls the
frames and pushes them to the hardware.

What does change is that management frames are no longer subject to
codel and inter-flow issues, and also note that the hash of a management
frame isn't actually well-defined.

johannes