Re: [PATCH rtw-next v12 0/3] wifi: rtw88: channel switch in AP mode

From: Luka Gejak

Date: Wed Nov 18 2026 - 07:48:21 EST


On November 18, 2026 12:26:21 AM GMT+01:00, Mehmet Fide
<mehmet.fide@xxxxxxxxx> wrote:
>From: Mehmet Fide <mehmet.fide@xxxxxxxxxxxxxxxxxx>
>
>An rtw88 AP cannot change its channel: the driver does not announce
>channel switch support, so hostapd's CHAN_SWITCH is refused and the
>only way is to tear the AP down. The series fixes the beacon download
>the reserved page build makes twice, counts a switch down on top of
>the firmware's beacon page, and times the countdown downloads from the
>TBTT so that every count goes out on the air.
>
>v12 applies on top of patches 1-4 of the firmware recovery fixes [2];
>patch 5 of [2] is superseded by commit 3f2238709f5d ("wifi: rtw88:
>clear RTW_FLAG_RESTARTING on stop and start failure") in rtw-next.
>
>v12 was tested with [2] and the mac80211 patch [1] on top, [1] is in
>mainline now as commit 12f97c692dcb, on a kernel with PROVE_LOCKING,
>DEBUG_ATOMIC_SLEEP, kmemleak and KFENCE, on an RTL8821CU (two
>adapters), an RTL8822BU, an RTL8822CU (USB) and an RTL8822CE (PCIe)
>with hostapd as AP, the beacons sniffed from a second machine, on
>rtw-next fb51e39dfaea:
>
>- the regression suite on the five adapters on 2.4 and 5 GHz:
> beacons, association, power save, scans, a hardware restart,
> rebinding, a firmware crash on the RTL8822CE, and 300 switches with
> counts 2, 5 and 10. The AP switched in 300 of 300 and every count
> went out in 296; the other four were on 2.4 GHz on the RTL8822BU and
> the RTL8822CU, which lost 46 % and 2.4 % of their beacons in the
> beacon tests of the same run. One missing count was the single CSA
> beacon of a count of 2, and the station lost the AP. In 30 more
> switches with a count of 2 on that adapter every count went out, and
> in 29 of 30 with the v12 of before [2]. In two more switches with a
> count of 2 the station lost the AP although the sniffers heard its
> single CSA beacon; the same happens with v11. No lockdep report, and
> no kmemleak report from the drivers
>- a test build that fails the first download of a countdown on purpose,
> with counts 2 and 5 on the RTL8821CU, the RTL8822BU and the RTL8822CE:
> the AP switched in 30 of 30 and every count went out in 25. In 15
> more runs each, the v12 of before [2] and this one lost counts in 2
> of 15 each. With v11 a count of 2 switched without any CSA beacon in
> 5 of 5
>- the same build holding the first step back by 95 ms: the AP switched
> in 30 of 30 and every count went out in 27. With v11, held back by 90
> to 100 ms, a count of 2 switched without any CSA beacon in 10 of 12
>- an IBSS left during its countdown, without [1] and with a monitor
> interface keeping the radio up, on four adapters: the countdown work
> stops and a scan works; with v11 the work ran on and every scan was
> refused
>- a station interface associated first, an AP started next to it and
> stopped during a countdown of 50 on the RTL8822CE and the RTL8822CU:
> the station stayed associated in the 5 runs in which it had
> associated
>- an off-channel period of 300 ms during a countdown of 20 on the
> RTL8821CU: the AP switched in 3 of 3. The beacons after an off-channel
> period are partly missing also without the series
>
>Every count missing in the first three items fell on a TBTT at which
>neither sniffer heard any beacon, as happens on these adapters without
>a countdown too.
>
>The RTL8822BU and the RTL8822CE ran again on rtw-next badf4c3c7679:
>
>- the regression suite on 2.4 and 5 GHz with 36 switches: every
> count went out, lockdep stayed on, no kmemleak report from the
> drivers
>- a firmware crash during a countdown of 40 on the RTL8822CE, the AP
> alone and next to an associated station interface: no lockdep or
> DEBUG_ATOMIC_SLEEP report; the AP alone ended on its new channel
>- a firmware crash with the hardware restart held back by 4 s in a test
> build, the AP and its interface stopped in between: the restart flag
> was clear afterwards and the next crash was recovered, 2 of 2
>
>In IBSS a countdown leaves more beacons off the air where the channel
>is busy; the count goes on after each of them. With 20 s with and
>without a countdown taking turns on one channel, the beacons that no
>sniffer heard are, on 5 GHz, 1.9 % and 3.5 % of the countdown of an
>RTL8821CU and the RTL8822BU against 0.3 % and 0.2 % without one, and
>none on the RTL8822C; on 2.4 GHz 6.6 to 12.1 % against 2.7 to 7.2 % on
>the five adapters. An AP shows no such difference. The cause is the
>beacon function, which rtw_fw_write_data_rsvd_page() switches off for
>the download: left on for the count steps, the difference is gone on
>5 GHz, but one RTL8821CU AP then lost its station in 6 of 110
>switches with counts 5 and 10 on 2.4 GHz, and in none of 100 without
>the change, so the series leaves the download as it is. The TSF
>stands still for the download as well, 2.8 ms on the USB chips and
>0.3 ms on the RTL8822CE, so the beacons of a countdown are about
>105 ms apart on USB.
>
>Not run for v12: Windows stations. The WoWLAN paths are built but not
>run: after a suspend the PCIe card of the test board is inaccessible and
>the USB adapters are probed again. Without the series as well, an IBSS
>interface stops beaconing after a scan. Patch 1 changed only in its
>commit message since v6.
>
>v12:
>- 2/3: a failed download of a count is repeated before the next step,
> and the switch completes only after the count of 1 reached the
> firmware (Luka). The work steps with the count that
> ieee80211_beacon_update_cntdwn() returns and no longer calls
> ieee80211_beacon_cntdwn_is_complete(), which warned when the AP
> stopped in between (Sashiko)
>- 2/3: the work never clears csa_vif, and rtw_ops_stop_ap() downloads
> the page again whenever a station interface stays (Luka)
>- 2/3: the countdown stops while beaconing is disabled, for an IBSS
> that is left or an off-channel period, and starts again when it is
> enabled with the switch still announced (Sashiko)
>- 2/3: WoWLAN suspend keeps the interface of the countdown and resume
> starts it again, without iterating the interfaces (Luka)
>- 2/3: a failed download of the replay after a hardware restart is
> repeated before the next step
>- 3/3: each step reads the TSF after its download and the next one runs
> 10 ms after the first TBTT at least 10 ms away; no step runs less than
> 5 ms after or 10 ms before a TBTT, the one that finishes the switch
> included. A late step delays the switch instead of taking a count off
> the air (Luka)
>- 3/3: the commit message says which count the first CSA beacon
> carries (Sashiko)
>- 1/3: the commit message says why there is no Fixes tag and what
> mac80211 reports for the double fetch (Sashiko)
>- rebased on patches 1-4 of [2]
>
>v11:
>- 2/3: the channel_switch_beacon op is gone, the BSS_CHANGED_BEACON
> path starts every countdown (Luka)
>- 2/3: rtw_fw_csa_start() does nothing on a device that is not
> running
>- 2/3: only the countdown work advances the count. The beacon of the
> reserved page is fetched as a template while a switch is announced,
> so the replay after a firmware recovery and the page of the WoWLAN
> firmware repeat the count that is on the air
>- 2/3: an AP that stops in its countdown downloads the reserved page
> again if a station interface stays behind; the downloads the
> countdown held back were lost for it
>- 3/3: the first download shares the branch of a late one, and the
> comment says what the branch catches (Luka)
>- the beacon interval that stays at 100 TU on the air is fixed by a
> separate patch (Luka)
>- rebased on rtw-next
>
>v10:
>- 2/3: rtw_core_stop() clears csa_vif and cancels the countdown work
> again once it has the mutex back, for a switch started while the
> mutex was dropped (Luka)
>- 3/3: the first count is downloaded by the countdown work, 10 ms
> after the next TBTT, and no longer when mac80211 changes the beacon.
> My reply on v9 said v10 only drops the back-off; that is not enough
> on the RTL8822CE, where a TBTT shortly before the download sends no
> beacon and the next one stays out too, so a count of 2 went out
> without a CSA beacon (Luka). The switch completes up to one beacon
> interval later. Ping-Ke's Acked-by and Luka's Reviewed-by are
> dropped for the changed code
>- rebased on rtw-next
>
>v9:
>- 2/3: the countdown starts again after a WoWLAN resume that worked,
> which mac80211 follows with no replay (Luka); Ping-Ke's Acked-by is
> dropped for the new code
>- 3/3: Acked-by from Ping-Ke on v8, Reviewed-by from Luka
>- an IBSS left during its countdown and joined again kept csa_active;
> that is mac80211's, fixed by a separate patch [1] (Luka)
>
>v8:
>- 2/3: Acked-by from Ping-Ke
>- 3/3: the TSF read retries once instead of looping, rtw_fw_csa_queue()
> reads the TSF once also on start, the port 0 comment is gone, empty
> lines where asked (Ping-Ke)
>
>v7:
>- new patch 3: the countdown downloads are timed from the TSF, 10 ms
> after each TBTT, so a download that runs late by less than a beacon
> interval still reaches the firmware before the next beacon; one
> later than that is followed by its successor at once. Ping-Ke's
> question on v6: with the absolute due time a late download and its
> successor could land in one beacon period and leave a count out of
> the air
>- patch 3 takes a TBTT within the margin before the start as the one
> that sent the countdown template: a TBTT between the template
> download and the first TSF read made the stations see one count
> twice
>- patch 2 keeps csa_vif until the beacon after the switch: a TIM
> update before it fetched one beacon past count 1 and hit the WARN
> in ieee80211_beacon_update_cntdwn()
>- rebased on rtw-next
>
>v6:
>- the wiphy work patch of v5 is gone: the countdown work is an
> ieee80211 delayed work and csa_vif is driver state, both under the
> driver mutex, so rtw88 takes no wiphy lock anywhere (Ping-Ke)
>- rtw_core_stop() cancels the countdown work and clears csa_vif, which
> covers the IPS path and the firmware recovery; the wiphy_lock() in
> the recovery work is gone (Ping-Ke)
>- the BSS_CHANGED_BEACON path says why it starts the countdown, the
> replay after a hardware restart, and stops it when the beacon comes
> back without a switch announced (Ping-Ke)
>- the downloads are paced from the first one with an absolute due
> time. Queued after each download, they drifted by about 10 ms per
> beacon interval on the ordered mac80211 workqueue; on a count of 50
> the AP switched half a second after its station had, and the
> station re-associated
>
>v5:
>- 2/3 is new: update_beacon_work becomes a wiphy work, so that it can
> look at the channel switch state under the wiphy lock; it checks
> RTW_FLAG_RUNNING instead of being cancelled in rtw_core_stop(), which
> runs without the wiphy lock on the IPS path (Ping-Ke); reverted in v6
>- 3/3: rtw_fw_csa_active/start/stop() assert the wiphy lock (Ping-Ke);
> reverted in v6, they assert the driver mutex
>- 3/3: hw_scan() takes the mutex with guard() (Ping-Ke)
>- 3/3: a comment on the association download names the STA+AP
> combination it skips the download for (Ping-Ke)
>- the whole set ran on a PROVE_LOCKING kernel without a lockdep report:
> switches with counts 1 to 100, hostapd stopped and restarted during
> a countdown, scans during a countdown, a firmware crash during a
> countdown, three stop/idle/start cycles, IBSS, 50 switches with a
> station attached, and on the RTL8822CU a station interface
> associating while the AP interface counted down
>
>v4:
>- 1/2: Reviewed-by from Luka
>- 2/2: the countdown work is back in the device, initialised once,
> with csa_vif naming the interface; no interface iteration and no
> lookup through the reserved page any more (Ping-Ke)
>- 2/2: the countdown is cancelled before a firmware recovery restarts
> the hardware, instead of guarding the work with flags (Ping-Ke)
>- 2/2: the page download on association skips a running countdown,
> like the TIM update and the PG backup (Luka)
>- 2/2: the scan refusal comes before the scan offload test, so that
> firmware without scan offload does not fall back to a software scan
> during a countdown (Luka)
>
>v3:
>- 1/2: the beacon out-pointer is initialised in the caller (Luka)
>- 2/2: the countdown work lives in the interface, remove_interface
> cancels it unconditionally, the queue wrapper and the comments are
> gone (Ping-Ke)
>- 2/2: the TIM update and the PG backup on set_key skip the page
> download while a switch is announced; every beacon fetch advances
> the countdown (Luka)
>- 2/2: the commit message names the refused hardware scan (Ping-Ke)
>
>v2:
>- 2/2: a hardware scan is refused while a switch is announced
>- the rtw89 mention is gone from the cover
>
>[1]
>https://lore.kernel.org/linux-wireless/20261025193615.1123718-1-mehmet.fide@xxxxxxxxx/
>[2]
>https://lore.kernel.org/linux-wireless/20261117171003.1840491-1-mehmet.fide@xxxxxxxxx/
>
>Mehmet Fide (3):
> wifi: rtw88: download the beacon the reserved page was built with
> wifi: rtw88: support channel switch in AP mode
> wifi: rtw88: time the countdown downloads from the TBTT
>
> drivers/net/wireless/realtek/rtw88/fw.c | 182 ++++++++++++++++--
> drivers/net/wireless/realtek/rtw88/fw.h | 5 +
> drivers/net/wireless/realtek/rtw88/mac80211.c | 60 +++++-
> drivers/net/wireless/realtek/rtw88/main.c | 9 +-
> drivers/net/wireless/realtek/rtw88/main.h | 5 +
> drivers/net/wireless/realtek/rtw88/reg.h | 1 +
> 6 files changed, 237 insertions(+), 25 deletions(-)
>
>
>base-commit: badf4c3c7679871a50cff1150f66e63187201757
>prerequisite-patch-id: 61a489b0da7e1d0028cc1c72c0ecd1301f3b8ec7
>prerequisite-patch-id: 5040450eb00c500bda0b95bb35bd821721f149ee
>prerequisite-patch-id: 8c5c1393ce4b56fecf3238b21bd34fcdbae8769a
>prerequisite-patch-id: 6ecc624d0c38a71c64993f725622e02ae2ad1629
>--
>2.56.0
>
>
I forgot to mention, v12 doesn't apply to the rtw-next branch, so it needs
a rebase.

Best regards,
Luka Gejak