| Message ID | 20240830023901.204746-1-quic_bqiang@quicinc.com (mailing list archive) |
|---|---|
| State | Deferred |
| Delegated to: | Kalle Valo |
| Headers | show |
| Series | wifi: ath12k: avoid deadlock during regulatory update in ath12k_regd_update() | expand |
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>
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 --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; }