[PATCH 04/13] wifi: mt76: mt7921: add mutex protection in critical paths

[Felix Fietkau]

On 20.01.26 21:10, Zac wrote:
> From: Zac Bowling <zac at zacbowling.com>
> 
> Add proper mutex protection for mt7921 driver operations that access
> hardware state without proper synchronization. This fixes multiple race
> conditions that can cause system instability.
> 
> Fixes added:
> 
> 1. mac.c: mt7921_mac_reset_work()
>     - Wrap ieee80211_iterate_active_interfaces() with mt792x_mutex
>     - The vif_connect_iter callback accesses hw_encap state
> 
> 2. main.c: mt7921_remain_on_channel()
>     - Remove mt792x_mutex_acquire/release around mt7925_set_channel_state()
>     - The function is already called with mutex held from mac80211
>     - This was causing double-lock deadlock
> 
> 3. main.c: mt7921_cancel_remain_on_channel()
>     - Remove mt792x_mutex_acquire/release
>     - Function is called from mac80211 with mutex already held
> 
> 4. pci.c: mt7921_pci_pm_complete()
>     - Remove mt792x_mutex_acquire/release around ieee80211_iterate_active_interfaces
>     - This was causing deadlock as the vif connect iteration tries
>       to acquire the mutex again
> 
> 5. usb.c: mt7921_usb_pm_complete()
>     - Same fix as pci.c for USB driver path
Changelog should be below "---" after the commit description, so it 
doesn't get picked up.

> These changes prevent both missing mutex protection and mutex deadlocks
> in the mt7921 driver.
> 
> Fixes: 5c14a5f944b9 ("wifi: mt76: mt7921: introduce remain_on_channel support")
> Signed-off-by: Zac Bowling <zac at zacbowling.com>

> diff --git a/drivers/net/wireless/mediatek/mt76/mt7921/main.c b/drivers/net/wireless/mediatek/mt76/mt7921/main.c
> index 5fae9a6e273c..196fcb1e2e94 100644
> --- a/drivers/net/wireless/mediatek/mt76/mt7921/main.c
> +++ b/drivers/net/wireless/mediatek/mt76/mt7921/main.c
> @@ -373,6 +373,11 @@ void mt7921_roc_abort_sync(struct mt792x_dev *dev)
>   
>   	timer_delete_sync(&phy->roc_timer);
>   	cancel_work_sync(&phy->roc_work);
> +	/* Note: caller must hold mutex if ieee80211_iterate_interfaces is
> +	 * needed for ROC cleanup. Some call sites (like mt7921_mac_sta_remove)
> +	 * already hold the mutex via mt76_sta_remove(). For suspend paths,
> +	 * the mutex should be acquired before calling this function.
> +	 */
>   	if (test_and_clear_bit(MT76_STATE_ROC, &phy->mt76->state))
>   		ieee80211_iterate_interfaces(mt76_hw(dev),
>   					     IEEE80211_IFACE_ITER_RESUME_ALL,
> @@ -619,6 +624,7 @@ void mt7921_set_runtime_pm(struct mt792x_dev *dev)
>   	bool monitor = !!(hw->conf.flags & IEEE80211_CONF_MONITOR);
>   
>   	pm->enable = pm->enable_user && !monitor;
> +	/* Note: caller (debugfs) must hold mutex before calling this function */
>   	ieee80211_iterate_active_interfaces(hw,
>   					    IEEE80211_IFACE_ITER_RESUME_ALL,
>   					    mt7921_pm_interface_iter, dev);
> @@ -765,6 +771,9 @@ mt7921_regd_set_6ghz_power_type(struct ieee80211_vif *vif, bool is_add)
>   	struct mt792x_dev *dev = phy->dev;
>   	u32 valid_vif_num = 0;
>   
> +	/* Note: caller (mt7921_mac_sta_add/remove via mt76_sta_add/remove)
> +	 * already holds dev->mt76.mutex, so we must not acquire it here.
> +	 */
>   	ieee80211_iterate_active_interfaces(mt76_hw(dev),
>   					    IEEE80211_IFACE_ITER_RESUME_ALL,
>   					    mt7921_calc_vif_num, &valid_vif_num);

It looks like these comments should be replaced with 
lockdep_assert_held, so that these assumptions can be verified 
automatically instead of doing so by hand.


> diff --git a/drivers/net/wireless/mediatek/mt76/mt7921/pci.c b/drivers/net/wireless/mediatek/mt76/mt7921/pci.c
> index ec9686183251..9f76b334b93d 100644
> --- a/drivers/net/wireless/mediatek/mt76/mt7921/pci.c
> +++ b/drivers/net/wireless/mediatek/mt76/mt7921/pci.c
> @@ -426,7 +426,9 @@ static int mt7921_pci_suspend(struct device *device)
>   	cancel_delayed_work_sync(&pm->ps_work);
>   	cancel_work_sync(&pm->wake_work);
>   
> +	mt792x_mutex_acquire(dev);
>   	mt7921_roc_abort_sync(dev);
> +	mt792x_mutex_release(dev);
The next patch is removing those...

- Felix