Hi,
This patchset adds tx and rx checksum offloading
to ath10k driver.
I don't have any hard numbers but I did observe a
consistently sligthly better performance on AP135
+ ath10k. Nothing ground breaking though.
Michal Kazior (2):
ath10k: implement rx checksum offloading
ath10k: implement tx checksum offloading
drivers/net/wireless/ath/ath10k/htt_rx.c | 40 ++++++++++++++++++++++++++++++
drivers/net/wireless/ath/ath10k/htt_tx.c | 2 ++
drivers/net/wireless/ath/ath10k/mac.c | 2 ++
3 files changed, 44 insertions(+)
--
1.7.9.5
HW supports L3/L4 rx checksum offloading.
This should reduce CPU load and improve
performance on slow host machines.
Signed-off-by: Michal Kazior <[email protected]>
---
drivers/net/wireless/ath/ath10k/htt_rx.c | 40 ++++++++++++++++++++++++++++++
1 file changed, 40 insertions(+)
diff --git a/drivers/net/wireless/ath/ath10k/htt_rx.c b/drivers/net/wireless/ath/ath10k/htt_rx.c
index 04f08d9..e784c40 100644
--- a/drivers/net/wireless/ath/ath10k/htt_rx.c
+++ b/drivers/net/wireless/ath/ath10k/htt_rx.c
@@ -804,6 +804,37 @@ static bool ath10k_htt_rx_has_fcs_err(struct sk_buff *skb)
return false;
}
+static int ath10k_htt_rx_get_csum_state(struct sk_buff *skb)
+{
+ struct htt_rx_desc *rxd;
+ u32 flags, info;
+ bool is_ip4, is_ip6;
+ bool is_tcp, is_udp;
+ bool ip_csum_ok, tcpudp_csum_ok;
+
+ rxd = (void *)skb->data - sizeof(*rxd);
+ flags = __le32_to_cpu(rxd->attention.flags);
+ info = __le32_to_cpu(rxd->msdu_start.info1);
+
+ is_ip4 = !!(info & RX_MSDU_START_INFO1_IPV4_PROTO);
+ is_ip6 = !!(info & RX_MSDU_START_INFO1_IPV6_PROTO);
+ is_tcp = !!(info & RX_MSDU_START_INFO1_TCP_PROTO);
+ is_udp = !!(info & RX_MSDU_START_INFO1_UDP_PROTO);
+ ip_csum_ok = !(flags & RX_ATTENTION_FLAGS_IP_CHKSUM_FAIL);
+ tcpudp_csum_ok = !(flags & RX_ATTENTION_FLAGS_TCP_UDP_CHKSUM_FAIL);
+
+ if (!is_ip4 && !is_ip6)
+ return CHECKSUM_NONE;
+ if (!is_tcp && !is_udp)
+ return CHECKSUM_NONE;
+ if (!ip_csum_ok)
+ return CHECKSUM_NONE;
+ if (!tcpudp_csum_ok)
+ return CHECKSUM_NONE;
+
+ return CHECKSUM_UNNECESSARY;
+}
+
static void ath10k_htt_rx_handler(struct ath10k_htt *htt,
struct htt_rx_indication *rx)
{
@@ -815,6 +846,7 @@ static void ath10k_htt_rx_handler(struct ath10k_htt *htt,
u8 *fw_desc;
int i, j;
int ret;
+ int ip_summed;
memset(&info, 0, sizeof(info));
@@ -889,6 +921,11 @@ static void ath10k_htt_rx_handler(struct ath10k_htt *htt,
continue;
}
+ /* The skb is not yet processed and it may be
+ * reallocated. Since the offload is in the original
+ * skb extract the checksum now and assign it later */
+ ip_summed = ath10k_htt_rx_get_csum_state(msdu_head);
+
info.skb = msdu_head;
info.fcs_err = ath10k_htt_rx_has_fcs_err(msdu_head);
info.signal = ATH10K_DEFAULT_NOISE_FLOOR;
@@ -914,6 +951,8 @@ static void ath10k_htt_rx_handler(struct ath10k_htt *htt,
if (ath10k_htt_rx_hdr_is_amsdu((void *)info.skb->data))
ath10k_dbg(ATH10K_DBG_HTT, "htt mpdu is amsdu\n");
+ info.skb->ip_summed = ip_summed;
+
ath10k_dbg_dump(ATH10K_DBG_HTT_DUMP, NULL, "htt mpdu: ",
info.skb->data, info.skb->len);
ath10k_process_rx(htt->ar, &info);
@@ -980,6 +1019,7 @@ static void ath10k_htt_rx_frag_handler(struct ath10k_htt *htt,
info.status = HTT_RX_IND_MPDU_STATUS_OK;
info.encrypt_type = MS(__le32_to_cpu(rxd->mpdu_start.info0),
RX_MPDU_START_INFO0_ENCRYPT_TYPE);
+ info.skb->ip_summed = ath10k_htt_rx_get_csum_state(info.skb);
if (tkip_mic_err) {
ath10k_warn("tkip mic error\n");
--
1.7.9.5
HW supports L3/L4 tx checksum offloading.
This should reduce CPU load and improve
performance on slow host machines.
Signed-off-by: Michal Kazior <[email protected]>
---
drivers/net/wireless/ath/ath10k/htt_tx.c | 2 ++
drivers/net/wireless/ath/ath10k/mac.c | 2 ++
2 files changed, 4 insertions(+)
diff --git a/drivers/net/wireless/ath/ath10k/htt_tx.c b/drivers/net/wireless/ath/ath10k/htt_tx.c
index dc3f3e8..656c254 100644
--- a/drivers/net/wireless/ath/ath10k/htt_tx.c
+++ b/drivers/net/wireless/ath/ath10k/htt_tx.c
@@ -465,6 +465,8 @@ int ath10k_htt_tx(struct ath10k_htt *htt, struct sk_buff *msdu)
flags1 = 0;
flags1 |= SM((u16)vdev_id, HTT_DATA_TX_DESC_FLAGS1_VDEV_ID);
flags1 |= SM((u16)tid, HTT_DATA_TX_DESC_FLAGS1_EXT_TID);
+ flags1 |= HTT_DATA_TX_DESC_FLAGS1_CKSUM_L3_OFFLOAD;
+ flags1 |= HTT_DATA_TX_DESC_FLAGS1_CKSUM_L4_OFFLOAD;
frags_paddr = ATH10K_SKB_CB(txfrag)->paddr;
diff --git a/drivers/net/wireless/ath/ath10k/mac.c b/drivers/net/wireless/ath/ath10k/mac.c
index 344ad27..5784337 100644
--- a/drivers/net/wireless/ath/ath10k/mac.c
+++ b/drivers/net/wireless/ath/ath10k/mac.c
@@ -3309,6 +3309,8 @@ int ath10k_mac_register(struct ath10k *ar)
ar->hw->wiphy->iface_combinations = &ath10k_if_comb;
ar->hw->wiphy->n_iface_combinations = 1;
+ ar->hw->netdev_features = NETIF_F_HW_CSUM;
+
ret = ath_regd_init(&ar->ath_common.regulatory, ar->hw->wiphy,
ath10k_reg_notifier);
if (ret) {
--
1.7.9.5
Michal Kazior <[email protected]> writes:
> Hi,
>
> This patchset adds tx and rx checksum offloading
> to ath10k driver.
>
> I don't have any hard numbers but I did observe a
> consistently sligthly better performance on AP135
> + ath10k. Nothing ground breaking though.
>
>
> Michal Kazior (2):
> ath10k: implement rx checksum offloading
> ath10k: implement tx checksum offloading
Applied, thanks.
--
Kalle Valo