Received-SPF: pass (google.com: domain of linux-wireless-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   Kalle Valo <kvalo@kernel.org>
To:     ath11k@lists.infradead.org
Cc:     linux-wireless@vger.kernel.org
Subject: [PATCH] wifi: ath11k: avoid deadlock during regulatory update in ath11k_regd_update()
Date:   Thu,  6 Oct 2022 18:17:47 +0300
Message-Id: <20221006151747.13757-1-kvalo@kernel.org>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

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+0x17/0x20
    ath11k_regd_update+0x15a/0x260 [ath11k]
    ath11k_regd_update_work+0x15/0x20 [ath11k]
    process_one_work+0x228/0x670
    worker_thread+0x4d/0x440
    kthread+0x16d/0x1b0
    ret_from_fork+0x22/0x30

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

    try to cancel_work_sync(&ar->regd_update_work) in ath11k_mac_op_stop().
    ifconfig  3109 [003]  2414.232506: probe:

    ath11k_mac_op_stop: (ffffffffc14187a0)
    drv_stop+0x30 ([mac80211])
    ieee80211_do_stop+0x5d2 ([mac80211])
    ieee80211_stop+0x3e ([mac80211])
    __dev_close_many+0x9e ([kernel.kallsyms])
    __dev_change_flags+0xbe ([kernel.kallsyms])
    dev_change_flags+0x23 ([kernel.kallsyms])
    devinet_ioctl+0x5e3 ([kernel.kallsyms])
    inet_ioctl+0x197 ([kernel.kallsyms])
    sock_do_ioctl+0x4d ([kernel.kallsyms])
    sock_ioctl+0x264 ([kernel.kallsyms])
    __x64_sys_ioctl+0x92 ([kernel.kallsyms])
    do_syscall_64+0x3a ([kernel.kallsyms])
    entry_SYSCALL_64_after_hwframe+0x63 ([kernel.kallsyms])
    __GI___ioctl+0x7 (/lib/x86_64-linux-gnu/libc-2.23.so)

The sequence of deadlock is:

1. Thread B calls rtnl_lock().

2. Thread A starts to run and calls rtnl_lock() from within
   ath11k_regd_update_work(), then enters wait state because the lock is owned by
   thread B.

3. Thread B continues to run and tries to call
   cancel_work_sync(&ar->regd_update_work), but thread A is in
   ath11k_regd_update_work() waiting for rtnl_lock(). So cancel_work_sync()
   forever waits for ath11k_regd_update_work() to finish and we have a deadlock.

Fix this by switching from using regulatory_set_wiphy_regd_sync() to
regulatory_set_wiphy_regd(). Now cfg80211 will schedule another workqueue which
handles the locking on it's own. So the ath11k workqueue can simply exit without
taking any locks, avoiding the deadlock.

Tested-on: WCN6855 hw2.0 PCI WLAN.HSP.1.1-03125-QCAHSPSWPL_V1_V2_SILICONZ_LITE-3

Signed-off-by: Wen Gong <quic_wgong@quicinc.com>
[kvalo: improve commit log]
Signed-off-by: Kalle Valo <quic_kvalo@quicinc.com>
---
 drivers/net/wireless/ath/ath11k/reg.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/drivers/net/wireless/ath/ath11k/reg.c b/drivers/net/wireless/ath/ath11k/reg.c
index 7ee3ff69dfc8..6fae4e61ede7 100644
--- a/drivers/net/wireless/ath/ath11k/reg.c
+++ b/drivers/net/wireless/ath/ath11k/reg.c
@@ -287,11 +287,7 @@ int ath11k_regd_update(struct ath11k *ar)
 		goto err;
 	}
 
-	rtnl_lock();
-	wiphy_lock(ar->hw->wiphy);
-	ret = regulatory_set_wiphy_regd_sync(ar->hw->wiphy, regd_copy);
-	wiphy_unlock(ar->hw->wiphy);
-	rtnl_unlock();
+	ret = regulatory_set_wiphy_regd(ar->hw->wiphy, regd_copy);
 
 	kfree(regd_copy);
 

base-commit: 023baf1318ef21442fab3842bf03883bc81223e0
-- 
2.30.2