* [PATCH rtw-next v2 1/3] wifi: rtw88: free the reserved pages of an interface on firmware recovery
2026-11-16 7:20 [PATCH rtw-next v2 0/3] wifi: rtw88: three fixes around the firmware recovery Mehmet Fide
@ 2026-11-16 7:20 ` Mehmet Fide
2026-11-16 16:02 ` Luka Gejak
2026-11-16 7:20 ` [PATCH rtw-next v2 2/3] wifi: rtw88: report the frames purged from the TX report queue Mehmet Fide
2026-11-16 7:20 ` [PATCH rtw-next v2 3/3] wifi: rtw88: reset the stations outside of RCU on firmware recovery Mehmet Fide
2 siblings, 1 reply; 6+ messages in thread
From: Mehmet Fide @ 2026-11-16 7:20 UTC (permalink / raw)
To: Ping-Ke Shih
Cc: Luka Gejak, Zong-Zhe Yang, Tzu-En Huang, Yan-Hsuan Chuang,
Kalle Valo, Stanislaw Gruszka, Brian Norris, linux-wireless,
linux-kernel, mehmet.fide
From: Mehmet Fide <mehmet.fide@screeningeagle.com>
The recovery after a firmware crash resets the driver state of every
interface, and mac80211 then adds the interfaces again without removing
them first. rtw_ops_add_interface() starts with an empty reserved page
list, so the pages the interface had before the crash are lost: the
next page build takes them off the build list and nothing frees them.
kmemleak after a firmware crash with an AP interface:
unreferenced object 0xffff0000cbe22c60 (size 96):
comm "hostapd", pid 25545, jiffies 4296463067
backtrace (crc cbdc21c4):
kmemleak_alloc+0xcc/0xe4
__kmalloc_cache_noprof+0x234/0x3bc
rtw_add_rsvd_page+0x38/0xa8 [rtw88_core]
rtw_add_rsvd_page_bcn+0x30/0x58 [rtw88_core]
rtw_ops_add_interface+0x144/0x2cc [rtw88_core]
rtw_ops_change_interface+0x6c/0x80 [rtw88_core]
drv_change_interface+0x100/0x258 [mac80211]
Free the pages where the recovery resets the interface.
rtw_ops_set_tim() can queue the beacon update work between this reset
and the reconfiguration, for a frame buffered for a dozing station. The
work would then find no page to build and warn about an empty reserved
page. Let it do nothing while RTW_FLAG_RESTARTING is set: the
reconfiguration downloads the beacon again.
Fixes: 5c831644e1f4 ("rtw88: handle and recover when firmware crash")
Signed-off-by: Mehmet Fide <mehmet.fide@screeningeagle.com>
---
Notes:
v2: the beacon update work does nothing while RTW_FLAG_RESTARTING is set
(Luka Gejak): rtw_ops_set_tim() can queue it between the reset and the
reconfiguration, where it found no page and warned
drivers/net/wireless/realtek/rtw88/fw.c | 6 ++++--
drivers/net/wireless/realtek/rtw88/main.c | 1 +
2 files changed, 5 insertions(+), 2 deletions(-)
diff --git a/drivers/net/wireless/realtek/rtw88/fw.c b/drivers/net/wireless/realtek/rtw88/fw.c
index 945fedcd375b..c9224749826d 100644
--- a/drivers/net/wireless/realtek/rtw88/fw.c
+++ b/drivers/net/wireless/realtek/rtw88/fw.c
@@ -1800,8 +1800,10 @@ void rtw_fw_update_beacon_work(struct work_struct *work)
update_beacon_work);
mutex_lock(&rtwdev->mutex);
- rtw_fw_download_rsvd_page(rtwdev);
- rtw_send_rsvd_page_h2c(rtwdev);
+ if (!test_bit(RTW_FLAG_RESTARTING, rtwdev->flags)) {
+ rtw_fw_download_rsvd_page(rtwdev);
+ rtw_send_rsvd_page_h2c(rtwdev);
+ }
mutex_unlock(&rtwdev->mutex);
}
diff --git a/drivers/net/wireless/realtek/rtw88/main.c b/drivers/net/wireless/realtek/rtw88/main.c
index f7708ad17b6f..adc54b063382 100644
--- a/drivers/net/wireless/realtek/rtw88/main.c
+++ b/drivers/net/wireless/realtek/rtw88/main.c
@@ -627,6 +627,7 @@ static void rtw_reset_vif_iter(void *data, u8 *mac, struct ieee80211_vif *vif)
rtw_bf_disassoc(rtwdev, vif, NULL);
rtw_vif_assoc_changed(rtwvif, NULL);
rtw_txq_cleanup(rtwdev, vif->txq);
+ rtw_remove_rsvd_page(rtwdev, rtwvif);
rtw_release_macid(rtwdev, rtwvif->mac_id);
}
--
2.56.0
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: [PATCH rtw-next v2 1/3] wifi: rtw88: free the reserved pages of an interface on firmware recovery
2026-11-16 7:20 ` [PATCH rtw-next v2 1/3] wifi: rtw88: free the reserved pages of an interface on " Mehmet Fide
@ 2026-11-16 16:02 ` Luka Gejak
0 siblings, 0 replies; 6+ messages in thread
From: Luka Gejak @ 2026-11-16 16:02 UTC (permalink / raw)
To: Mehmet Fide, Ping-Ke Shih
Cc: Zong-Zhe Yang, Tzu-En Huang, Yan-Hsuan Chuang, Kalle Valo,
Stanislaw Gruszka, Brian Norris, linux-wireless, linux-kernel,
mehmet.fide
Hi Mehmet,
Thanks for the update. Three comments below.
On November 16, 2026 8:20:12 AM GMT+01:00, Mehmet Fide
<mehmet.fide@gmail.com> wrote:
> @@ -1800,8 +1800,10 @@ void rtw_fw_update_beacon_work(struct work_struct *work)
> + if (!test_bit(RTW_FLAG_RESTARTING, rtwdev->flags)) {
> + rtw_fw_download_rsvd_page(rtwdev);
> + rtw_send_rsvd_page_h2c(rtwdev);
> + }
Could a skipped TIM update be retried at reconfig_complete()? After
reconfig's last beacon download, sta_info_cleanup() can still expire a
buffered frame and queue this work. If it runs before the flag clears,
the beacon keeps the old TIM bit.
[ ... ]
> + if (!test_bit(RTW_FLAG_RESTARTING, rtwdev->flags)) {
Also, if the last interface closes before restart work takes the wiphy
lock, open_count is zero and reconfig_complete() is skipped. The flag
survives normal reopening, so the worker still skips TIM updates.
Could that exit be covered too?
[ ... ]
> @@ -627,6 +627,7 @@ static void rtw_reset_vif_iter(void *data, u8 *mac, struct ieee80211_vif *vif)
> + rtw_remove_rsvd_page(rtwdev, rtwvif);
As far as I can see, rtw_core_stop() drops the driver mutex without
holding wiphy. A concurrent beacon change can then reach:
ieee80211_change_beacon()
rtw_ops_bss_info_changed()
rtw_fw_download_rsvd_page()
The lists are already empty, and mac80211 restart has not begun yet.
This seems to leave a path to the empty reserved page warning.
Could the direct download path be covered too?
Best regards,
Luka Gejak
^ permalink raw reply [flat|nested] 6+ messages in thread
* [PATCH rtw-next v2 2/3] wifi: rtw88: report the frames purged from the TX report queue
2026-11-16 7:20 [PATCH rtw-next v2 0/3] wifi: rtw88: three fixes around the firmware recovery Mehmet Fide
2026-11-16 7:20 ` [PATCH rtw-next v2 1/3] wifi: rtw88: free the reserved pages of an interface on " Mehmet Fide
@ 2026-11-16 7:20 ` Mehmet Fide
2026-11-16 7:20 ` [PATCH rtw-next v2 3/3] wifi: rtw88: reset the stations outside of RCU on firmware recovery Mehmet Fide
2 siblings, 0 replies; 6+ messages in thread
From: Mehmet Fide @ 2026-11-16 7:20 UTC (permalink / raw)
To: Ping-Ke Shih
Cc: Luka Gejak, Zong-Zhe Yang, Tzu-En Huang, Yan-Hsuan Chuang,
Kalle Valo, Stanislaw Gruszka, Brian Norris, linux-wireless,
linux-kernel, mehmet.fide
From: Mehmet Fide <mehmet.fide@screeningeagle.com>
A frame whose sender asked for its TX status waits in the TX report
queue for the report of the firmware. When no report comes, the purge
timer frees the frames with skb_queue_purge(), and rtw_core_deinit()
does the same. mac80211 is not told: the sender gets no status, and the
copy mac80211 keeps for the status of a frame from user space stays in
its list until ieee80211_free_hw() warns about it. On an RTL8822CE with
hostapd, after a firmware crash and an unbind of the driver:
rtw88_8822ce 0000:01:00.0: failed to get tx report from firmware
[...]
Have pending ack frames!
WARNING: net/mac80211/main.c:1761 at ieee80211_free_ack_frame+0x4c/0x70 [mac80211], CPU#1: sh/24725
Call trace:
ieee80211_free_ack_frame+0x4c/0x70 [mac80211] (P)
idr_for_each+0x6c/0xf4
ieee80211_free_hw+0x30/0x8c [mac80211]
rtw_pci_remove+0xd8/0xf0 [rtw88_pci]
pci_device_remove+0x44/0xd4
Take the frames off the queue under its lock and hand them to
ieee80211_purge_tx_queue(), which reports each one as not sent.
Fixes: e3037485c68e ("rtw88: new Realtek 802.11ac driver")
Signed-off-by: Mehmet Fide <mehmet.fide@screeningeagle.com>
Reviewed-by: Luka Gejak <luka.gejak@linux.dev>
---
Notes:
v2: Reviewed-by Luka Gejak, no change
drivers/net/wireless/realtek/rtw88/main.c | 5 +----
drivers/net/wireless/realtek/rtw88/tx.c | 20 ++++++++++++++++----
drivers/net/wireless/realtek/rtw88/tx.h | 1 +
3 files changed, 18 insertions(+), 8 deletions(-)
diff --git a/drivers/net/wireless/realtek/rtw88/main.c b/drivers/net/wireless/realtek/rtw88/main.c
index adc54b063382..ce65a1a80840 100644
--- a/drivers/net/wireless/realtek/rtw88/main.c
+++ b/drivers/net/wireless/realtek/rtw88/main.c
@@ -2225,7 +2225,6 @@ void rtw_core_deinit(struct rtw_dev *rtwdev)
struct rtw_fw_state *fw = &rtwdev->fw;
struct rtw_fw_state *wow_fw = &rtwdev->wow_fw;
struct rtw_rsvd_page *rsvd_pkt, *tmp;
- unsigned long flags;
rtw_wait_firmware_completion(rtwdev);
@@ -2237,9 +2236,7 @@ void rtw_core_deinit(struct rtw_dev *rtwdev)
destroy_workqueue(rtwdev->tx_wq);
timer_delete_sync(&rtwdev->tx_report.purge_timer);
- spin_lock_irqsave(&rtwdev->tx_report.q_lock, flags);
- skb_queue_purge(&rtwdev->tx_report.queue);
- spin_unlock_irqrestore(&rtwdev->tx_report.q_lock, flags);
+ rtw_tx_report_purge(rtwdev);
skb_queue_purge(&rtwdev->coex.queue);
skb_queue_purge(&rtwdev->c2h_queue);
diff --git a/drivers/net/wireless/realtek/rtw88/tx.c b/drivers/net/wireless/realtek/rtw88/tx.c
index 701ed7c4f3f7..87d993282b1c 100644
--- a/drivers/net/wireless/realtek/rtw88/tx.c
+++ b/drivers/net/wireless/realtek/rtw88/tx.c
@@ -176,21 +176,33 @@ static void rtw_tx_report_enable(struct rtw_dev *rtwdev,
pkt_info->report = true;
}
+void rtw_tx_report_purge(struct rtw_dev *rtwdev)
+{
+ struct rtw_tx_report *tx_report = &rtwdev->tx_report;
+ struct sk_buff_head purge;
+ unsigned long flags;
+
+ __skb_queue_head_init(&purge);
+
+ spin_lock_irqsave(&tx_report->q_lock, flags);
+ skb_queue_splice_init(&tx_report->queue, &purge);
+ spin_unlock_irqrestore(&tx_report->q_lock, flags);
+
+ ieee80211_purge_tx_queue(rtwdev->hw, &purge);
+}
+
void rtw_tx_report_purge_timer(struct timer_list *t)
{
struct rtw_dev *rtwdev = timer_container_of(rtwdev, t,
tx_report.purge_timer);
struct rtw_tx_report *tx_report = &rtwdev->tx_report;
- unsigned long flags;
if (skb_queue_len(&tx_report->queue) == 0)
return;
rtw_warn(rtwdev, "failed to get tx report from firmware\n");
- spin_lock_irqsave(&tx_report->q_lock, flags);
- skb_queue_purge(&tx_report->queue);
- spin_unlock_irqrestore(&tx_report->q_lock, flags);
+ rtw_tx_report_purge(rtwdev);
}
void rtw_tx_report_enqueue(struct rtw_dev *rtwdev, struct sk_buff *skb, u8 sn)
diff --git a/drivers/net/wireless/realtek/rtw88/tx.h b/drivers/net/wireless/realtek/rtw88/tx.h
index 168f48e87853..363f63525965 100644
--- a/drivers/net/wireless/realtek/rtw88/tx.h
+++ b/drivers/net/wireless/realtek/rtw88/tx.h
@@ -98,6 +98,7 @@ void rtw_tx_pkt_info_update(struct rtw_dev *rtwdev,
struct sk_buff *skb);
void rtw_tx_fill_tx_desc(struct rtw_dev *rtwdev,
struct rtw_tx_pkt_info *pkt_info, struct sk_buff *skb);
+void rtw_tx_report_purge(struct rtw_dev *rtwdev);
void rtw_tx_report_enqueue(struct rtw_dev *rtwdev, struct sk_buff *skb, u8 sn);
void rtw_tx_report_handle(struct rtw_dev *rtwdev, struct sk_buff *skb, int src);
void rtw_tx_rsvd_page_pkt_info_update(struct rtw_dev *rtwdev,
--
2.56.0
^ permalink raw reply [flat|nested] 6+ messages in thread* [PATCH rtw-next v2 3/3] wifi: rtw88: reset the stations outside of RCU on firmware recovery
2026-11-16 7:20 [PATCH rtw-next v2 0/3] wifi: rtw88: three fixes around the firmware recovery Mehmet Fide
2026-11-16 7:20 ` [PATCH rtw-next v2 1/3] wifi: rtw88: free the reserved pages of an interface on " Mehmet Fide
2026-11-16 7:20 ` [PATCH rtw-next v2 2/3] wifi: rtw88: report the frames purged from the TX report queue Mehmet Fide
@ 2026-11-16 7:20 ` Mehmet Fide
2026-11-16 15:55 ` Luka Gejak
2 siblings, 1 reply; 6+ messages in thread
From: Mehmet Fide @ 2026-11-16 7:20 UTC (permalink / raw)
To: Ping-Ke Shih
Cc: Luka Gejak, Zong-Zhe Yang, Tzu-En Huang, Yan-Hsuan Chuang,
Kalle Valo, Stanislaw Gruszka, Brian Norris, linux-wireless,
linux-kernel, mehmet.fide
From: Mehmet Fide <mehmet.fide@screeningeagle.com>
The recovery after a firmware crash removes the driver state of every
station with rtw_sta_remove(), called from an atomic station iterator.
rtw_sta_remove() cancels the rate work of the station with
cancel_work_sync(), which can sleep. With a station associated:
BUG: sleeping function called from invalid context at kernel/workqueue.c:4722
in_atomic(): 0, irqs_disabled(): 0, non_block: 0, pid: 14, name: kworker/u16:1
RCU nest depth: 1, expected: 0
Workqueue: phy35 rtw_fw_recovery_work [rtw88_core]
Call trace:
__might_resched+0x13c/0x1f0
__might_sleep+0x48/0x7c
__cancel_work_sync+0xa8/0xb4
cancel_work_sync+0x14/0x20
rtw_sta_remove+0x34/0x10c [rtw88_core]
rtw_reset_sta_iter+0x20/0x48 [rtw88_core]
ieee80211_iterate_stations_atomic+0x80/0x144 [mac80211]
rtw_fw_recovery_work+0x1dc/0x2b8 [rtw88_core]
Use rtw_iterate_stas(), which collects the stations under RCU and calls
the iterator after it. The recovery holds the driver mutex it asks for.
Fixes: bcafcb959a57 ("wifi: rtw88: use work to update rate to avoid RCU warning")
Signed-off-by: Mehmet Fide <mehmet.fide@screeningeagle.com>
Reviewed-by: Luka Gejak <luka.gejak@linux.dev>
---
Notes:
v2: Reviewed-by Luka Gejak, no change
drivers/net/wireless/realtek/rtw88/main.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/net/wireless/realtek/rtw88/main.c b/drivers/net/wireless/realtek/rtw88/main.c
index ce65a1a80840..8e7ae747b900 100644
--- a/drivers/net/wireless/realtek/rtw88/main.c
+++ b/drivers/net/wireless/realtek/rtw88/main.c
@@ -664,7 +664,7 @@ static void __fw_recovery_work(struct rtw_dev *rtwdev)
WARN(1, "firmware crash, start reset and recover\n");
rtw_iterate_keys_atomic(rtwdev, NULL, rtw_reset_key_iter, rtwdev);
- rtw_iterate_stas_atomic(rtwdev, rtw_reset_sta_iter, rtwdev);
+ rtw_iterate_stas(rtwdev, rtw_reset_sta_iter, rtwdev);
rtw_iterate_vifs_atomic(rtwdev, rtw_reset_vif_iter, rtwdev);
bitmap_zero(rtwdev->hw_port, RTW_PORT_NUM);
rtw_enter_ips(rtwdev);
--
2.56.0
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: [PATCH rtw-next v2 3/3] wifi: rtw88: reset the stations outside of RCU on firmware recovery
2026-11-16 7:20 ` [PATCH rtw-next v2 3/3] wifi: rtw88: reset the stations outside of RCU on firmware recovery Mehmet Fide
@ 2026-11-16 15:55 ` Luka Gejak
0 siblings, 0 replies; 6+ messages in thread
From: Luka Gejak @ 2026-11-16 15:55 UTC (permalink / raw)
To: Mehmet Fide, Ping-Ke Shih
Cc: Zong-Zhe Yang, Tzu-En Huang, Yan-Hsuan Chuang, Kalle Valo,
Stanislaw Gruszka, Brian Norris, linux-wireless, linux-kernel,
mehmet.fide
Hi Mehmet,
Thanks for the update. One comment below.
On November 16, 2026 8:20:14 AM GMT+01:00, Mehmet Fide
<mehmet.fide@gmail.com> wrote:
> @@ -664,7 +664,7 @@ static void __fw_recovery_work(struct rtw_dev *rtwdev)
> WARN(1, "firmware crash, start reset and recover\n");
>
> rtw_iterate_keys_atomic(rtwdev, NULL, rtw_reset_key_iter, rtwdev);
> - rtw_iterate_stas_atomic(rtwdev, rtw_reset_sta_iter, rtwdev);
> + rtw_iterate_stas(rtwdev, rtw_reset_sta_iter, rtwdev);
> rtw_iterate_vifs_atomic(rtwdev, rtw_reset_vif_iter, rtwdev);
I followed the allocation failure case from v1 a bit further:
https://lore.kernel.org/r/699D1E7C-1793-44A7-AF57-CFE16AD036E6@linux.dev/
As far as I can see, skipping a station also skips rtw_txq_cleanup().
For a queue outside the voice access category, wake_tx_queue() queues
tx_work. If that work has not run yet, the omitted station's queue can
stay on rtwdev->txqs until reconfiguration. rtw_sta_add() then calls
rtw_txq_init(), which does this without unlinking it:
INIT_LIST_HEAD(&rtwtxq->list);
The device list still points at that node, but the node now points to
itself. __rtw_tx_work() seems to keep visiting it while holding
txq_lock. Its list_del_init() cannot repair the device list because
the node has already lost its old links.
rtw_enter_ips() does not drain rtwdev->tx_wq, and the flush in
ieee80211_restart_work() is of mac80211's separate workqueue. So I think
this can hang the TX worker, beyond the station count and MAC ID leak
I mentioned before.
Could recovery use an iteration that cannot silently skip stations?
Best regards,
Luka Gejak
^ permalink raw reply [flat|nested] 6+ messages in thread