Serialize the ath10k implementation of the wake_tx_queue ops. ath10k_mac_op_wake_tx_queue() must not run concurrent since it's using ieee80211_txq_schedule_start().
Fixes: bb2edb733586 ("ath10k: migrate to mac80211 txq scheduling") Reported-by: Felix Fietkau nbd@nbd.name Link: https://lore.kernel.org/r/519b5bb9-8899-ae7c-4eff-f3116cdfdb56@nbd.name CC: stable@vger.kernel.org Signed-off-by: Alexander Wetzel alexander@wetzel-home.de ---
The intend of this patch is to sort out an issue discovered in the discussion referred to by the Link tag.
I can't test it with real HW and thus just implemented the per-ac queue lock Felix suggested. One obvious alternative to the per-ac lock would be to bring back the txqs_lock commit bb2edb733586 ("ath10k: migrate to mac80211 txq scheduling") dropped.
Alexander
--- drivers/net/wireless/ath/ath10k/core.c | 3 +++ drivers/net/wireless/ath/ath10k/core.h | 3 +++ drivers/net/wireless/ath/ath10k/mac.c | 6 ++++-- 3 files changed, 10 insertions(+), 2 deletions(-)
diff --git a/drivers/net/wireless/ath/ath10k/core.c b/drivers/net/wireless/ath/ath10k/core.c index 5eb131ab916f..533ed7169e11 100644 --- a/drivers/net/wireless/ath/ath10k/core.c +++ b/drivers/net/wireless/ath/ath10k/core.c @@ -3643,6 +3643,9 @@ struct ath10k *ath10k_core_create(size_t priv_size, struct device *dev, mutex_init(&ar->dump_mutex); spin_lock_init(&ar->data_lock);
+ for (int ac = 0; ac < IEEE80211_NUM_ACS; ac++) + spin_lock_init(&ar->queue_lock[ac]); + INIT_LIST_HEAD(&ar->peers); init_waitqueue_head(&ar->peer_mapping_wq); init_waitqueue_head(&ar->htt.empty_tx_wq); diff --git a/drivers/net/wireless/ath/ath10k/core.h b/drivers/net/wireless/ath/ath10k/core.h index f5de8ce8fb45..4b5239de4018 100644 --- a/drivers/net/wireless/ath/ath10k/core.h +++ b/drivers/net/wireless/ath/ath10k/core.h @@ -1170,6 +1170,9 @@ struct ath10k { /* protects shared structure data */ spinlock_t data_lock;
+ /* serialize wake_tx_queue calls per ac */ + spinlock_t queue_lock[IEEE80211_NUM_ACS]; + struct list_head arvifs; struct list_head peers; struct ath10k_peer *peer_map[ATH10K_MAX_NUM_PEER_IDS]; diff --git a/drivers/net/wireless/ath/ath10k/mac.c b/drivers/net/wireless/ath/ath10k/mac.c index 7675858f069b..9c4bf2fdbc0f 100644 --- a/drivers/net/wireless/ath/ath10k/mac.c +++ b/drivers/net/wireless/ath/ath10k/mac.c @@ -4732,13 +4732,14 @@ static void ath10k_mac_op_wake_tx_queue(struct ieee80211_hw *hw, { struct ath10k *ar = hw->priv; int ret; - u8 ac; + u8 ac = txq->ac;
ath10k_htt_tx_txq_update(hw, txq); if (ar->htt.tx_q_state.mode != HTT_TX_MODE_SWITCH_PUSH) return;
- ac = txq->ac; + spin_lock_bh(&ar->queue_lock[ac]); + ieee80211_txq_schedule_start(hw, ac); txq = ieee80211_next_txq(hw, ac); if (!txq) @@ -4753,6 +4754,7 @@ static void ath10k_mac_op_wake_tx_queue(struct ieee80211_hw *hw, ath10k_htt_tx_txq_update(hw, txq); out: ieee80211_txq_schedule_end(hw, ac); + spin_unlock_bh(&ar->queue_lock[ac]); }
/* Must not be called with conf_mutex held as workers can use that also. */
Alexander Wetzel alexander@wetzel-home.de wrote:
Serialize the ath10k implementation of the wake_tx_queue ops. ath10k_mac_op_wake_tx_queue() must not run concurrent since it's using ieee80211_txq_schedule_start().
The intend of this patch is to sort out an issue discovered in the discussion referred to by the Link tag.
I can't test it with real hardware and thus just implemented the per-ac queue lock Felix suggested. One obvious alternative to the per-ac lock would be to bring back the txqs_lock commit bb2edb733586 ("ath10k: migrate to mac80211 txq scheduling") dropped.
Fixes: bb2edb733586 ("ath10k: migrate to mac80211 txq scheduling") Reported-by: Felix Fietkau nbd@nbd.name Link: https://lore.kernel.org/r/519b5bb9-8899-ae7c-4eff-f3116cdfdb56@nbd.name CC: stable@vger.kernel.org Signed-off-by: Alexander Wetzel alexander@wetzel-home.de Signed-off-by: Kalle Valo quic_kvalo@quicinc.com
Patch applied to ath-next branch of ath.git, thanks.
b719ebc37a1e wifi: ath10k: Serialize wake_tx_queue ops
linux-stable-mirror@lists.linaro.org