mac80211: check ATF flag in ieee80211_next_txq()

Commit Message

Ryder Lee Jan. 6, 2021, 1:01 a.m. UTC

The selected txq should be scheduled unconditionally if
NL80211_EXT_FEATURE_AIRTIME_FAIRNESS is not set by driver.

Also put the sta to the end of the active_txqs list if
deficit is negative then move on to the next txq.

Tested-by: Sujuan Chen <[e-mail]>
Signed-off-by: Ryder Lee <[e-mail]>
---
 net/mac80211/tx.c | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

Comments

Toke Høiland-Jørgensen Jan. 6, 2021, 10:51 a.m. UTC | #1

Ryder Lee <[e-mail]> writes:

> The selected txq should be scheduled unconditionally if
> NL80211_EXT_FEATURE_AIRTIME_FAIRNESS is not set by driver.
>
> Also put the sta to the end of the active_txqs list if
> deficit is negative then move on to the next txq.

Why is this needed? If the feature is not set, no airtime should ever be
accounted to the station, and so sta->airtime[txqi->txq.ac].deficit will
always be 0 - so you're just adding another check that doesn't actually
change the behaviour, aren't you?

-Toke

Ryder Lee Jan. 6, 2021, 11:50 a.m. UTC | #2

On Wed, 2021-01-06 at 11:51 +0100, Toke Høiland-Jørgensen wrote:
> Ryder Lee <[e-mail]> writes:
> 
> > The selected txq should be scheduled unconditionally if
> > NL80211_EXT_FEATURE_AIRTIME_FAIRNESS is not set by driver.
> >
> > Also put the sta to the end of the active_txqs list if
> > deficit is negative then move on to the next txq.
> 
> Why is this needed? If the feature is not set, no airtime should ever be
> accounted to the station, and so sta->airtime[txqi->txq.ac].deficit will
> always be 0 - so you're just adding another check that doesn't actually
> change the behaviour, aren't you?
> 
> -Toke

You're right. I think the problem is we still call
ieee80211_sta_register_airtime() in driver even if feature is not set,
and that makes deficit < 0. Will fix it in driver.

Ryder

Felix Fietkau Jan. 6, 2021, 1:30 p.m. UTC | #3

On 2021-01-06 11:51, Toke Høiland-Jørgensen wrote:
> Ryder Lee <[e-mail]> writes:
> 
>> The selected txq should be scheduled unconditionally if
>> NL80211_EXT_FEATURE_AIRTIME_FAIRNESS is not set by driver.
>>
>> Also put the sta to the end of the active_txqs list if
>> deficit is negative then move on to the next txq.
> 
> Why is this needed? If the feature is not set, no airtime should ever be
> accounted to the station, and so sta->airtime[txqi->txq.ac].deficit will
> always be 0 - so you're just adding another check that doesn't actually
> change the behaviour, aren't you?
I think it might make sense to keep airtime reporting even when airtime
fairness is disabled at run time, so this patch makes sense to me.
Instead of this patch, the right place to deal with this would probably
be ieee80211_sta_register_airtime.

- Felix

Toke Høiland-Jørgensen Jan. 6, 2021, 3:41 p.m. UTC | #4

Felix Fietkau <[e-mail]> writes:

> On 2021-01-06 11:51, Toke Høiland-Jørgensen wrote:
>> Ryder Lee <[e-mail]> writes:
>> 
>>> The selected txq should be scheduled unconditionally if
>>> NL80211_EXT_FEATURE_AIRTIME_FAIRNESS is not set by driver.
>>>
>>> Also put the sta to the end of the active_txqs list if
>>> deficit is negative then move on to the next txq.
>> 
>> Why is this needed? If the feature is not set, no airtime should ever be
>> accounted to the station, and so sta->airtime[txqi->txq.ac].deficit will
>> always be 0 - so you're just adding another check that doesn't actually
>> change the behaviour, aren't you?
>
> I think it might make sense to keep airtime reporting even when airtime
> fairness is disabled at run time, so this patch makes sense to me.
> Instead of this patch, the right place to deal with this would probably
> be ieee80211_sta_register_airtime.

When the fairness mechanism is user-disabled I agree it makes sense to
still keep the accounting; and in fact that's what
ieee80211_sta_register_airtime() already does when the accounting is
turned off by way of the airtime_flags field... So don't think anything
else is needed there either?

-Toke

Ryder Lee Jan. 7, 2021, 2:11 a.m. UTC | #5

On Wed, 2021-01-06 at 16:41 +0100, Toke Høiland-Jørgensen wrote:
> Felix Fietkau <[e-mail]> writes:
> 
> > On 2021-01-06 11:51, Toke Høiland-Jørgensen wrote:
> >> Ryder Lee <[e-mail]> writes:
> >> 
> >>> The selected txq should be scheduled unconditionally if
> >>> NL80211_EXT_FEATURE_AIRTIME_FAIRNESS is not set by driver.
> >>>
> >>> Also put the sta to the end of the active_txqs list if
> >>> deficit is negative then move on to the next txq.
> >> 
> >> Why is this needed? If the feature is not set, no airtime should ever be
> >> accounted to the station, and so sta->airtime[txqi->txq.ac].deficit will
> >> always be 0 - so you're just adding another check that doesn't actually
> >> change the behaviour, aren't you?
> >
> > I think it might make sense to keep airtime reporting even when airtime
> > fairness is disabled at run time, so this patch makes sense to me.
> > Instead of this patch, the right place to deal with this would probably
> > be ieee80211_sta_register_airtime.
> 
> When the fairness mechanism is user-disabled I agree it makes sense to
> still keep the accounting; and in fact that's what
> ieee80211_sta_register_airtime() already does when the accounting is
> turned off by way of the airtime_flags field... So don't think anything
> else is needed there either?
> 
> -Toke

Not sure I get this right. Are you talking about local->airtime_flags =
AIRTIME_USE_TX | AIRTIME_USE_RX ? I think that's different and we still
need to take NL80211_EXT_FEATURE_AIRTIME_FAIRNESS into account, right?

Ryder

Toke Høiland-Jørgensen Jan. 7, 2021, 1:08 p.m. UTC | #6

Ryder Lee <[e-mail]> writes:

> On Wed, 2021-01-06 at 16:41 +0100, Toke Høiland-Jørgensen wrote:
>> Felix Fietkau <[e-mail]> writes:
>> 
>> > On 2021-01-06 11:51, Toke Høiland-Jørgensen wrote:
>> >> Ryder Lee <[e-mail]> writes:
>> >> 
>> >>> The selected txq should be scheduled unconditionally if
>> >>> NL80211_EXT_FEATURE_AIRTIME_FAIRNESS is not set by driver.
>> >>>
>> >>> Also put the sta to the end of the active_txqs list if
>> >>> deficit is negative then move on to the next txq.
>> >> 
>> >> Why is this needed? If the feature is not set, no airtime should ever be
>> >> accounted to the station, and so sta->airtime[txqi->txq.ac].deficit will
>> >> always be 0 - so you're just adding another check that doesn't actually
>> >> change the behaviour, aren't you?
>> >
>> > I think it might make sense to keep airtime reporting even when airtime
>> > fairness is disabled at run time, so this patch makes sense to me.
>> > Instead of this patch, the right place to deal with this would probably
>> > be ieee80211_sta_register_airtime.
>> 
>> When the fairness mechanism is user-disabled I agree it makes sense to
>> still keep the accounting; and in fact that's what
>> ieee80211_sta_register_airtime() already does when the accounting is
>> turned off by way of the airtime_flags field... So don't think anything
>> else is needed there either?
>> 
>> -Toke
>
> Not sure I get this right. Are you talking about local->airtime_flags =
> AIRTIME_USE_TX | AIRTIME_USE_RX ? I think that's different and we still
> need to take NL80211_EXT_FEATURE_AIRTIME_FAIRNESS into account, right?

I just meant that what Felix was asking for (a way *for the user* to
disable airtime fairness while still getting the airtime usage
accounted) is possible by setting those flags. The EXT_FEATURE flag is
meant as a way for the driver to signal to mac80211 that it supports
reporting airtime at all; so ideally it should be a flag that is only
set once.

Going back and reading your initial response it seems like you may be
toggling the flag dynamically in the driver, though? Is this accurate?
And if so, why? Is it not enough for you to fiddle with the
USE_TX/USE_RX flags? :)

-Toke

Ryder Lee Jan. 8, 2021, 2:25 a.m. UTC | #7

On Thu, 2021-01-07 at 14:08 +0100, Toke Høiland-Jørgensen wrote:
> Ryder Lee <[e-mail]> writes:
> 
> > On Wed, 2021-01-06 at 16:41 +0100, Toke Høiland-Jørgensen wrote:
> >> Felix Fietkau <[e-mail]> writes:
> >> 
> >> > On 2021-01-06 11:51, Toke Høiland-Jørgensen wrote:
> >> >> Ryder Lee <[e-mail]> writes:
> >> >> 
> >> >>> The selected txq should be scheduled unconditionally if
> >> >>> NL80211_EXT_FEATURE_AIRTIME_FAIRNESS is not set by driver.
> >> >>>
> >> >>> Also put the sta to the end of the active_txqs list if
> >> >>> deficit is negative then move on to the next txq.
> >> >> 
> >> >> Why is this needed? If the feature is not set, no airtime should ever be
> >> >> accounted to the station, and so sta->airtime[txqi->txq.ac].deficit will
> >> >> always be 0 - so you're just adding another check that doesn't actually
> >> >> change the behaviour, aren't you?
> >> >
> >> > I think it might make sense to keep airtime reporting even when airtime
> >> > fairness is disabled at run time, so this patch makes sense to me.
> >> > Instead of this patch, the right place to deal with this would probably
> >> > be ieee80211_sta_register_airtime.
> >> 
> >> When the fairness mechanism is user-disabled I agree it makes sense to
> >> still keep the accounting; and in fact that's what
> >> ieee80211_sta_register_airtime() already does when the accounting is
> >> turned off by way of the airtime_flags field... So don't think anything
> >> else is needed there either?
> >> 
> >> -Toke
> >
> > Not sure I get this right. Are you talking about local->airtime_flags =
> > AIRTIME_USE_TX | AIRTIME_USE_RX ? I think that's different and we still
> > need to take NL80211_EXT_FEATURE_AIRTIME_FAIRNESS into account, right?
> 
> I just meant that what Felix was asking for (a way *for the user* to
> disable airtime fairness while still getting the airtime usage
> accounted) is possible by setting those flags. The EXT_FEATURE flag is
> meant as a way for the driver to signal to mac80211 that it supports
> reporting airtime at all; so ideally it should be a flag that is only
> set once.
> 
> Going back and reading your initial response it seems like you may be
> toggling the flag dynamically in the driver, though? Is this accurate?
> And if so, why? Is it not enough for you to fiddle with the
> USE_TX/USE_RX flags? :)
> 
> -Toke

Gotcha. We just set it once indeed. So the way you think is disable the
EXT_FEATURE flag and clear AIRTIME_USE_TX through debugfs
(DEBUGFS_ADD_MODE(airtime_flags, 0600)) in the meantime.

Ryder

Patch

diff --git a/net/mac80211/tx.c b/net/mac80211/tx.c
index 6422da6690f7..5640c9428596 100644
--- a/net/mac80211/tx.c
+++ b/net/mac80211/tx.c
@@ -3760,14 +3760,19 @@  struct ieee80211_txq *ieee80211_next_txq(struct ieee80211_hw *hw, u8 ac)
 		struct sta_info *sta = container_of(txqi->txq.sta,
 						    struct sta_info, sta);
 		bool aql_check = ieee80211_txq_airtime_check(hw, &txqi->txq);
-		s64 deficit = sta->airtime[txqi->txq.ac].deficit;
+		s64 deficit = 0;
 
 		if (aql_check)
 			found_eligible_txq = true;
 
-		if (deficit < 0)
-			sta->airtime[txqi->txq.ac].deficit +=
-				sta->airtime_weight;
+		if (wiphy_ext_feature_isset(local->hw.wiphy,
+					    NL80211_EXT_FEATURE_AIRTIME_FAIRNESS)) {
+			deficit = sta->airtime[txqi->txq.ac].deficit;
+
+			if (deficit < 0)
+				sta->airtime[txqi->txq.ac].deficit +=
+					sta->airtime_weight;
+		}
 
 		if (deficit < 0 || !aql_check) {
 			list_move_tail(&txqi->schedule_order,