Drivers without native channel contexts get the emulated ops, and
ieee80211_emulate_remove_chanctx() reconfigures the hardware with no
context left, which ieee80211_calc_hw_conf_chan() turns into a channel
change to the default channel. A monitor-mode retune releases its
context before using the next one, so every retune costs the driver two
channel changes: one to the default channel and one to the channel that
was asked for. Nothing is on the air without a context and the next
context sets the channel, so drop that reconfiguration; the radar flag
is still cleared for the next configuration to carry.

On an AR9271 (ath9k_htc), where every channel change is a full reset
over USB, hopping across the 13 standard 2.4 GHz channels with an
unmodified driver showed 2006 drv_config CHANGE_CHANNEL calls for 1000
hops with the reconfiguration and 1006 without, a median hop of 123.5 ms
against 92.0 ms (p95 183 ms against 105 ms), and 99.1% against 98.3% of
injected frames delivered on a second card.

Signed-off-by: Nerijus Bendžiūnas <nerijus.bendziunas@gmail.com>
---
 net/mac80211/main.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/net/mac80211/main.c b/net/mac80211/main.c
index a59837b9f480..b3f9b3bc3cac 100644
--- a/net/mac80211/main.c
+++ b/net/mac80211/main.c
@@ -289,9 +289,14 @@ void ieee80211_emulate_remove_chanctx(struct ieee80211_hw *hw,
 {
 	struct ieee80211_local *local = hw_to_local(hw);
 
+	/*
+	 * With no context left there is nothing to configure: nothing is on
+	 * the air, and the next context sets the channel. Parking the radio
+	 * on the default channel here only to move it again when the next
+	 * context is added costs drivers without native channel contexts a
+	 * full channel change per retune.
+	 */
 	local->hw.conf.radar_enabled = false;
-
-	_ieee80211_hw_conf_chan(local, NULL);
 }
 EXPORT_SYMBOL(ieee80211_emulate_remove_chanctx);
 

base-commit: ca800a9302764c445de0da0e84d2252400a770ee
-- 
2.55.0

On Sat, 2026-08-29 at 13:15 +0300, Nerijus Bendžiūnas wrote:
> Drivers without native channel contexts get the emulated ops, and
> ieee80211_emulate_remove_chanctx() reconfigures the hardware with no
> context left, which ieee80211_calc_hw_conf_chan() turns into a channel
> change to the default channel. A monitor-mode retune releases its
> context before using the next one, so every retune costs the driver two
> channel changes: one to the default channel and one to the channel that
> was asked for. Nothing is on the air without a context and the next
> context sets the channel, so drop that reconfiguration; the radar flag
> is still cleared for the next configuration to carry.
> 
> On an AR9271 (ath9k_htc), where every channel change is a full reset
> over USB, hopping across the 13 standard 2.4 GHz channels with an
> unmodified driver showed 2006 drv_config CHANGE_CHANNEL calls for 1000
> hops with the reconfiguration and 1006 without, a median hop of 123.5 ms
> against 92.0 ms (p95 183 ms against 105 ms), and 99.1% against 98.3% of
> injected frames delivered on a second card.
> 
> Signed-off-by: Nerijus Bendžiūnas <nerijus.bendziunas@gmail.com>

You're expected to disclose LLM usage.

https://www.kernel.org/doc/html/latest/process/coding-assistants.html

And please clean up the wordering everywhere before submitting again.

johannes