2016-08-17 15:33:13

by Rajkumar Manoharan

Subject: [PATCH] ath10k: improve wake_tx_queue ops performance

txqs_lock is interfering with wake_tx_queue submitting more frames.
so queues don't get filled in and don't keep firmware/hardware busy
enough. This change helps to reduce the txqs_lock contention and
wake_tx_queue() blockage to being possible in txrx_unref().

To reduce turn around time of wake_tx_queue ops and to maintain fairness
among all txqs, the callback is updated to push first txq alone from
pending list for every wake_tx_queue call. Remaining txqs will be
processed later upon tx completion.

Below improvements are observed in push-only mode and validated on
IPQ4019 platform. With this change, in AP mode ~10Mbps increase is
observed in downlink (AP -> STA) traffic and approx. 5-10% of CPU
usage is reduced.

Major improvement is observed in 1-hop Mesh mode topology in 11ACVHT80.
Compared to Infra mode, CPU overhead is higher in Mesh mode due to path
lookup and no fast-xmit support. So reducing spin lock contention is
helping in Mesh.

TOT +change
-------- --------
TCP DL 545 Mbps 595 Mbps
TCP UL 555 Mbps 585 Mbps

Signed-off-by: Rajkumar Manoharan <[email protected]>
---
drivers/net/wireless/ath/ath10k/mac.c | 18 +++++++++++++++++-
1 file changed, 17 insertions(+), 1 deletion(-)

diff --git a/drivers/net/wireless/ath/ath10k/mac.c b/drivers/net/wireless/ath/ath10k/mac.c
index c7c7ceb9be73..93465ffec2eb 100644
--- a/drivers/net/wireless/ath/ath10k/mac.c
+++ b/drivers/net/wireless/ath/ath10k/mac.c
@@ -4106,13 +4106,29 @@ static void ath10k_mac_op_wake_tx_queue(struct ieee80211_hw *hw,
{
struct ath10k *ar = hw->priv;
struct ath10k_txq *artxq = (void *)txq->drv_priv;
+ struct ieee80211_txq *f_txq;
+ struct ath10k_txq *f_artxq;
+ int ret = 0;
+ int max = 16;

spin_lock_bh(&ar->txqs_lock);
if (list_empty(&artxq->list))
list_add_tail(&artxq->list, &ar->txqs);
+
+ f_artxq = list_first_entry(&ar->txqs, struct ath10k_txq, list);
+ f_txq = container_of((void *)f_artxq, struct ieee80211_txq, drv_priv);
+ list_del_init(&f_artxq->list);
+
+ while (ath10k_mac_tx_can_push(hw, f_txq) && max--) {
+ ret = ath10k_mac_tx_push_txq(hw, f_txq);
+ if (ret)
+ break;
+ }
+ if (ret != -ENOENT)
+ list_add_tail(&f_artxq->list, &ar->txqs);
spin_unlock_bh(&ar->txqs_lock);

- ath10k_mac_tx_push_pending(ar);
+ ath10k_htt_tx_txq_update(hw, f_txq);
ath10k_htt_tx_txq_update(hw, txq);
}

--
2.9.2

2016-09-02 15:52:16

by Kalle Valo

Subject: Re: ath10k: improve wake_tx_queue ops performance

Rajkumar Manoharan <[email protected]> wrote:
> txqs_lock is interfering with wake_tx_queue submitting more frames.
> so queues don't get filled in and don't keep firmware/hardware busy
> enough. This change helps to reduce the txqs_lock contention and
> wake_tx_queue() blockage to being possible in txrx_unref().
>
> To reduce turn around time of wake_tx_queue ops and to maintain fairness
> among all txqs, the callback is updated to push first txq alone from
> pending list for every wake_tx_queue call. Remaining txqs will be
> processed later upon tx completion.
>
> Below improvements are observed in push-only mode and validated on
> IPQ4019 platform. With this change, in AP mode ~10Mbps increase is
> observed in downlink (AP -> STA) traffic and approx. 5-10% of CPU
> usage is reduced.
>
> Major improvement is observed in 1-hop Mesh mode topology in 11ACVHT80.
> Compared to Infra mode, CPU overhead is higher in Mesh mode due to path
> lookup and no fast-xmit support. So reducing spin lock contention is
> helping in Mesh.
>
> TOT +change
> -------- --------
> TCP DL 545 Mbps 595 Mbps
> TCP UL 555 Mbps 585 Mbps
>
> Signed-off-by: Rajkumar Manoharan <[email protected]>

Thanks, 1 patch applied to ath-next branch of ath.git:

83e164b7679d ath10k: improve wake_tx_queue ops performance

--
Sent by pwcli
https://patchwork.kernel.org/patch/9286083/