diff mbox series

wifi: ath12k: avoid deadlock during regulatory update in ath12k_regd_update()

Message ID 20240830023901.204746-1-quic_bqiang@quicinc.com
State New
Headers show
Series wifi: ath12k: avoid deadlock during regulatory update in ath12k_regd_update() | expand

Commit Message

Baochen Qiang Aug. 30, 2024, 2:39 a.m. UTC 
From: Wen Gong <quic_wgong@quicinc.com>

Running this test in a loop it is easy to reproduce an rtnl deadlock:

iw reg set FI
ifconfig wlan0 down

What happens is that thread A (workqueue) tries to update the regulatory:

    try to acquire the rtnl_lock of ar->regd_update_work

    rtnl_lock
    ath12k_regd_update [ath12k]
    ath12k_regd_update_work [ath12k]
    process_one_work
    worker_thread
    kthread
    ret_from_fork

And thread B (ifconfig) tries to stop the interface:

    try to cancel_work_sync(&ar->regd_update_work) in ath12k_mac_op_stop().
    ifconfig  3109 [003]  2414.232506: probe:

    ath12k_mac_op_stop [ath12k]
    drv_stop [mac80211]
    ieee80211_do_stop [mac80211]
    ieee80211_stop [mac80211]

The sequence of deadlock is:

1. Thread B calls rtnl_lock().

2. Thread A starts to run and calls rtnl_lock() from within
   ath12k_regd_update_work(), then enters wait state because the lock is owned by
   thread B.

3. Thread B tries to call cancel_work_sync(&ar->regd_update_work), but thread A is in
   ath12k_regd_update_work() waiting for rtnl_lock(). So cancel_work_sync()
   forever waits for ath12k_regd_update_work() to finish and we have a deadlock.

Change to use regulatory_set_wiphy_regd(), which is the asynchronous version of
regulatory_set_wiphy_regd_sync(). This way rtnl & wiphy locks are not required so can
be removed, and in the end the deadlock issue can be avoided.

But a side effect introduced by the asynchronous regd update is that, some essential
information used in ath12k_reg_update_chan_list(), which would be called later in
ath12k_regd_update(), might has not been updated by cfg80211, as a result wrong
channel parameters sent to firmware.

To handle this side effect, move ath12k_reg_update_chan_list() to ath12k_reg_notifier(),
and advertise WIPHY_FLAG_NOTIFY_REGDOM_BY_DRIVER to cfg80211. This works because,
in the process of the asynchronous regd update, after the new regd is processed,
cfg80211 will notify ath12k by calling ath12k_reg_notifier(). Since all essential
information is updated at that time, we are good to do channel list update.

Please note ath12k_reg_notifier() could also be called due to other reasons, like
core/beacon/user hints etc. For them we are not allowed to call
ath12k_reg_update_chan_list() because regd has not been updated. This is done by
verifying  the initiator.

Tested-on: WCN7850 hw2.0 PCI WLAN.HMT.1.0.c5-00481-QCAHMTSWPL_V1.0_V2.0_SILICONZ-3

Signed-off-by: Wen Gong <quic_wgong@quicinc.com>
Co-developed-by: Baochen Qiang <quic_bqiang@quicinc.com>
Signed-off-by: Baochen Qiang <quic_bqiang@quicinc.com>
---
 drivers/net/wireless/ath/ath12k/reg.c | 35 +++++++++++++++------------
 1 file changed, 20 insertions(+), 15 deletions(-)


base-commit: 8fb3b2b8d6d489416a7ff8a28cd4083340ad9e55

Comments

Jeff Johnson Sept. 4, 2024, 4:33 p.m. UTC | #1 
On 8/29/2024 7:39 PM, Baochen Qiang wrote:
> From: Wen Gong <quic_wgong@quicinc.com>
> 
> Running this test in a loop it is easy to reproduce an rtnl deadlock:
> 
> iw reg set FI
> ifconfig wlan0 down
> 
> What happens is that thread A (workqueue) tries to update the regulatory:
> 
>     try to acquire the rtnl_lock of ar->regd_update_work
> 
>     rtnl_lock
>     ath12k_regd_update [ath12k]
>     ath12k_regd_update_work [ath12k]
>     process_one_work
>     worker_thread
>     kthread
>     ret_from_fork
> 
> And thread B (ifconfig) tries to stop the interface:
> 
>     try to cancel_work_sync(&ar->regd_update_work) in ath12k_mac_op_stop().
>     ifconfig  3109 [003]  2414.232506: probe:
> 
>     ath12k_mac_op_stop [ath12k]
>     drv_stop [mac80211]
>     ieee80211_do_stop [mac80211]
>     ieee80211_stop [mac80211]
> 
> The sequence of deadlock is:
> 
> 1. Thread B calls rtnl_lock().
> 
> 2. Thread A starts to run and calls rtnl_lock() from within
>    ath12k_regd_update_work(), then enters wait state because the lock is owned by

checkpatch complains that the commit description exceeds 75 columns

at a minimum you should avoid exceeding 80 columns

Kalle, do you want to reformat when you pull into pending?
Or are you ok with the current formatting?

>    thread B.
> 
> 3. Thread B tries to call cancel_work_sync(&ar->regd_update_work), but thread A is in
>    ath12k_regd_update_work() waiting for rtnl_lock(). So cancel_work_sync()
>    forever waits for ath12k_regd_update_work() to finish and we have a deadlock.
> 
> Change to use regulatory_set_wiphy_regd(), which is the asynchronous version of
> regulatory_set_wiphy_regd_sync(). This way rtnl & wiphy locks are not required so can
> be removed, and in the end the deadlock issue can be avoided.
> 
> But a side effect introduced by the asynchronous regd update is that, some essential
> information used in ath12k_reg_update_chan_list(), which would be called later in
> ath12k_regd_update(), might has not been updated by cfg80211, as a result wrong
> channel parameters sent to firmware.
> 
> To handle this side effect, move ath12k_reg_update_chan_list() to ath12k_reg_notifier(),
> and advertise WIPHY_FLAG_NOTIFY_REGDOM_BY_DRIVER to cfg80211. This works because,
> in the process of the asynchronous regd update, after the new regd is processed,
> cfg80211 will notify ath12k by calling ath12k_reg_notifier(). Since all essential
> information is updated at that time, we are good to do channel list update.
> 
> Please note ath12k_reg_notifier() could also be called due to other reasons, like
> core/beacon/user hints etc. For them we are not allowed to call
> ath12k_reg_update_chan_list() because regd has not been updated. This is done by
> verifying  the initiator.
> 
> Tested-on: WCN7850 hw2.0 PCI WLAN.HMT.1.0.c5-00481-QCAHMTSWPL_V1.0_V2.0_SILICONZ-3
> 
> Signed-off-by: Wen Gong <quic_wgong@quicinc.com>
> Co-developed-by: Baochen Qiang <quic_bqiang@quicinc.com>
> Signed-off-by: Baochen Qiang <quic_bqiang@quicinc.com>
code change itself LGTM, so...
Acked-by: Jeff Johnson <quic_jjohnson@quicinc.com>
Kalle Valo Sept. 4, 2024, 4:47 p.m. UTC | #2 
Jeff Johnson <quic_jjohnson@quicinc.com> writes:

> On 8/29/2024 7:39 PM, Baochen Qiang wrote:
>
>> From: Wen Gong <quic_wgong@quicinc.com>
>> 
>> Running this test in a loop it is easy to reproduce an rtnl deadlock:
>> 
>> iw reg set FI
>> ifconfig wlan0 down
>> 
>> What happens is that thread A (workqueue) tries to update the regulatory:
>> 
>>     try to acquire the rtnl_lock of ar->regd_update_work
>> 
>>     rtnl_lock
>>     ath12k_regd_update [ath12k]
>>     ath12k_regd_update_work [ath12k]
>>     process_one_work
>>     worker_thread
>>     kthread
>>     ret_from_fork
>> 
>> And thread B (ifconfig) tries to stop the interface:
>> 
>>     try to cancel_work_sync(&ar->regd_update_work) in ath12k_mac_op_stop().
>>     ifconfig  3109 [003]  2414.232506: probe:
>> 
>>     ath12k_mac_op_stop [ath12k]
>>     drv_stop [mac80211]
>>     ieee80211_do_stop [mac80211]
>>     ieee80211_stop [mac80211]
>> 
>> The sequence of deadlock is:
>> 
>> 1. Thread B calls rtnl_lock().
>> 
>> 2. Thread A starts to run and calls rtnl_lock() from within
>>    ath12k_regd_update_work(), then enters wait state because the lock is owned by
>
> checkpatch complains that the commit description exceeds 75 columns
>
> at a minimum you should avoid exceeding 80 columns
>
> Kalle, do you want to reformat when you pull into pending?

Yes, I can reformat it in the pending branch. But I'm busy right now so
it might take a while.
diff mbox series

Patch

diff --git a/drivers/net/wireless/ath/ath12k/reg.c b/drivers/net/wireless/ath/ath12k/reg.c
index 439d61f284d8..ea03f3f50e50 100644
--- a/drivers/net/wireless/ath/ath12k/reg.c
+++ b/drivers/net/wireless/ath/ath12k/reg.c
@@ -55,6 +55,24 @@  ath12k_reg_notifier(struct wiphy *wiphy, struct regulatory_request *request)
 	ath12k_dbg(ar->ab, ATH12K_DBG_REG,
 		   "Regulatory Notification received for %s\n", wiphy_name(wiphy));
 
+	if (request->initiator == NL80211_REGDOM_SET_BY_DRIVER) {
+		ath12k_dbg(ar->ab, ATH12K_DBG_REG,
+			   "driver initiated regd update\n");
+		if (ah->state != ATH12K_HW_STATE_ON)
+			return;
+
+		for_each_ar(ah, ar, i) {
+			ret = ath12k_reg_update_chan_list(ar);
+			if (ret) {
+				ath12k_warn(ar->ab,
+					    "failed to update chan list for pdev %u, ret %d\n",
+					    i, ret);
+				break;
+			}
+		}
+		return;
+	}
+
 	/* Currently supporting only General User Hints. Cell base user
 	 * hints to be handled later.
 	 * Hints from other sources like Core, Beacons are not expected for
@@ -211,7 +229,6 @@  int ath12k_regd_update(struct ath12k *ar, bool init)
 	struct ieee80211_regdomain *regd, *regd_copy = NULL;
 	int ret, regd_len, pdev_id;
 	struct ath12k_base *ab;
-	int i;
 
 	ab = ar->ab;
 
@@ -275,11 +292,7 @@  int ath12k_regd_update(struct ath12k *ar, bool init)
 		goto err;
 	}
 
-	rtnl_lock();
-	wiphy_lock(hw->wiphy);
-	ret = regulatory_set_wiphy_regd_sync(hw->wiphy, regd_copy);
-	wiphy_unlock(hw->wiphy);
-	rtnl_unlock();
+	ret = regulatory_set_wiphy_regd(hw->wiphy, regd_copy);
 
 	kfree(regd_copy);
 
@@ -290,15 +303,6 @@  int ath12k_regd_update(struct ath12k *ar, bool init)
 		goto skip;
 
 	ah->regd_updated = true;
-	/* Apply the new regd to all the radios, this is expected to be received only once
-	 * since we check for ah->regd_updated and allow here only once.
-	 */
-	for_each_ar(ah, ar, i) {
-		ab = ar->ab;
-		ret = ath12k_reg_update_chan_list(ar);
-		if (ret)
-			goto err;
-	}
 skip:
 	return 0;
 err:
@@ -770,6 +774,7 @@  void ath12k_regd_update_work(struct work_struct *work)
 void ath12k_reg_init(struct ieee80211_hw *hw)
 {
 	hw->wiphy->regulatory_flags = REGULATORY_WIPHY_SELF_MANAGED;
+	hw->wiphy->flags |= WIPHY_FLAG_NOTIFY_REGDOM_BY_DRIVER;
 	hw->wiphy->reg_notifier = ath12k_reg_notifier;
 }