Hello,
this series introduces driver support for hardware encapsulation
offloading feature.
The main goal of the series is an overall improvement of system
performance. On a QCA9563+QCA9888-based access point in bridged mode,
encapsulation offloading increases TCP 16-streams DL throughput from
365 to 396 mbps (+8%) and UDP DL throughput from 436 to 483 mbps (+11%).
The series consist of four patches, the first three prepare the code,
and the last one introduces the offloading support itself. The first
patch reworks transmission status reporting to make it flexible enough
to support 802.11 and Ethernet encapsulated frames reporting. The second
patch reworks improves Ethernet encapsulated frames handling in the HTT
Tx layer. The third patch reworks the module parameter that configures
the main encapsulation format of frames that are passed from the driver
to the hardware. It makes it possible to configure more encapsulation
methods than just raw or not-raw. And, as stated before, the fourth
patch actually introduces offloading support. It changes a couple of
frame analysis places along the xmit path and starts reporting
offloading support to mac80211 via the corresponding hw attribute.
The new feature has been extensively tested with QCA9888. It works well
and introduces a measurable performance increase on platforms with not
so fast CPU (e.g. QCA9563). Tests with MT7621+QCA6174 show no
regression. I was unable to observe any measurable throughput increase
in tests with QCA6174 due to the poor perfomance the SISO client and OTA
testing. Only slight downlink rate stability increase with slight
reducing of the SIRQ load on the AP side.
Tests with QCA9880, QCA9980, QCA9984, QCA9377 revealed no regression
[1]. No specific throughput tests have been performed. In OTA tests with
a regular load, there is a slight increase in througput stability.
In tests of the QCA9377 in Sta mode, Edward Matijevic noticed a possible
regression in the Ethernet tx mode compared to the Native WiFi tx
mode:
> There might be a regression in download bandwidth under ETHERNET mode
> but I didn't do enough tests to see if was my PC or the adapter as the
> upload bandwidth was similar to other tests.
> NATIVE_WIFI performs the same as without the patch.
Credits: thanks to OpenWrt folks Oldřich Jedlička
<[email protected]> and Edward Matijevic <[email protected]> who
helped me a lot to perform extensive testing with various chips after a
RFC/RFT post in the OpenWrt development mailing list [1]. Thanks to
Zhijun You <[email protected]> who helped with the testing, pointed me
to the previous work (see below) and the missed parts in the HTT Tx
layer [1].
Prior work: similar patches were independently developed by QCA and then
submitted for inclusion by John Crispin <[email protected]> [2] and
latter by Zhijun You <[email protected]> [3]. The work was not merged
due to the lack of a part for 64-bits descriptors [4]. The idea of HTT
Tx layer change from this work was adopted for the series in patch #2.
1. https://www.mail-archive.com/[email protected]/msg62013.html
2. https://lore.kernel.org/all/[email protected]/
3. https://lore.kernel.org/all/[email protected]/
4. https://patchwork.kernel.org/project/linux-wireless/patch/[email protected]/
CC: Edward Matijevic <[email protected]>
CC: John Crispin <[email protected]>
CC: Oldřich Jedlička <[email protected]>
CC: Tom Psyborg <[email protected]>
CC: Vasanthakumar Thiagarajan <[email protected]>
CC: Zhijun You <[email protected]>
Changes since RFC:
* new Tested-on and Tested-by tags
* new patch #2 for better Ethernet encapsulation support in the HTT Tx
layer
* rebased on top of latest ath-next
Sergey Ryazanov (3):
ath10k: improve tx status reporting
ath10k: htt_tx: do not interpret Eth frames as WiFi
ath10k: turn rawmode into frame_mode
ath10k: add encapsulation offloading support
drivers/net/wireless/ath/ath10k/core.c | 11 ++--
drivers/net/wireless/ath/ath10k/core.h | 1 +
drivers/net/wireless/ath/ath10k/htt_tx.c | 61 ++++++++++++---------
drivers/net/wireless/ath/ath10k/mac.c | 67 +++++++++++++++++++-----
drivers/net/wireless/ath/ath10k/txrx.c | 12 ++++-
5 files changed, 108 insertions(+), 44 deletions(-)
--
2.35.1
Frame encapsulation from Ethernet into the IEEE 802.11 frame format
takes a considerable host CPU time on the xmit path. The firmware is
able to do this operation for us, so enable encapsulation offloading for
AP and Sta interface types to improve overall system performance.
The driver is almost ready for encapsulation offloading support. There
are only a few places where the driver assumes the frame format is IEEE
802.11 that need to be fixed.
Encapsulation offloading is currently disabled by default and the driver
utilizes mac80211 encapsulation support. To activate offloading, the
frame_mode=2 parameter should be passed during module loading.
On a QCA9563+QCA9888-based access point in bridged mode, encapsulation
offloading increases TCP 16-streams DL throughput from 365 to 396 mbps
(+8%) and UDP DL throughput from 436 to 483 mbps (+11%).
Tested-on: QCA9888 hw 2.0 10.4-3.9.0.2-00131
Tested-on: QCA6174 hw 3.2 PCI WLAN.RM.4.4.1-00157-QCARMSWPZ-1
Signed-off-by: Sergey Ryazanov <[email protected]>
Tested-by: Oldřich Jedlička <[email protected]> # TP-Link Archer C7 v4 & v5 (QCA9563 + QCA9880)
Tested-by: Edward Matijevic <[email protected]> # TP-Link Archer C2600 (IPQ8064 + QCA9980 10.4.1.00030-1)
Tested-by: Edward Matijevic <[email protected]> # QCA9377 PCI in Sta mode
Tested-by: Zhijun You <[email protected]> # NETGEAR R7800 (QCA9984 10.4-3.9.0.2-00159)
---
Changes since RFC:
* new Tested-on and Tested-by tags
* changed position in series: #3 -> #4
* rebased on top of latest ath-next
drivers/net/wireless/ath/ath10k/core.c | 2 +-
drivers/net/wireless/ath/ath10k/mac.c | 67 +++++++++++++++++++++-----
2 files changed, 55 insertions(+), 14 deletions(-)
diff --git a/drivers/net/wireless/ath/ath10k/core.c b/drivers/net/wireless/ath/ath10k/core.c
index 5654387acb4b..276954b70d63 100644
--- a/drivers/net/wireless/ath/ath10k/core.c
+++ b/drivers/net/wireless/ath/ath10k/core.c
@@ -55,7 +55,7 @@ MODULE_PARM_DESC(uart_print, "Uart target debugging");
MODULE_PARM_DESC(skip_otp, "Skip otp failure for calibration in testmode");
MODULE_PARM_DESC(cryptmode, "Crypto mode: 0-hardware, 1-software");
MODULE_PARM_DESC(frame_mode,
- "Datapath frame mode (0: raw, 1: native wifi (default))");
+ "Datapath frame mode (0: raw, 1: native wifi (default), 2: ethernet)");
MODULE_PARM_DESC(coredump_mask, "Bitfield of what to include in firmware crash file");
MODULE_PARM_DESC(fw_diag_log, "Diag based fw log debugging");
diff --git a/drivers/net/wireless/ath/ath10k/mac.c b/drivers/net/wireless/ath/ath10k/mac.c
index 3570a5895ea8..ebd5b9a8943e 100644
--- a/drivers/net/wireless/ath/ath10k/mac.c
+++ b/drivers/net/wireless/ath/ath10k/mac.c
@@ -3713,6 +3713,9 @@ ath10k_mac_tx_h_get_txmode(struct ath10k *ar,
const struct ath10k_skb_cb *skb_cb = ATH10K_SKB_CB(skb);
__le16 fc = hdr->frame_control;
+ if (IEEE80211_SKB_CB(skb)->flags & IEEE80211_TX_CTL_HW_80211_ENCAP)
+ return ATH10K_HW_TXRX_ETHERNET;
+
if (!vif || vif->type == NL80211_IFTYPE_MONITOR)
return ATH10K_HW_TXRX_RAW;
@@ -3873,6 +3876,12 @@ static void ath10k_mac_tx_h_fill_cb(struct ath10k *ar,
bool noack = false;
cb->flags = 0;
+
+ if (info->flags & IEEE80211_TX_CTL_HW_80211_ENCAP) {
+ cb->flags |= ATH10K_SKB_F_QOS; /* Assume data frames are QoS */
+ goto finish_cb_fill;
+ }
+
if (!ath10k_tx_h_use_hwcrypto(vif, skb))
cb->flags |= ATH10K_SKB_F_NO_HWCRYPT;
@@ -3911,6 +3920,7 @@ static void ath10k_mac_tx_h_fill_cb(struct ath10k *ar,
cb->flags |= ATH10K_SKB_F_RAW_TX;
}
+finish_cb_fill:
cb->vif = vif;
cb->txq = txq;
cb->airtime_est = airtime;
@@ -4034,7 +4044,11 @@ static int ath10k_mac_tx(struct ath10k *ar,
ath10k_tx_h_seq_no(vif, skb);
break;
case ATH10K_HW_TXRX_ETHERNET:
- ath10k_tx_h_8023(skb);
+ /* Convert 802.11->802.3 header only if the frame was erlier
+ * encapsulated to 802.11 by mac80211. Otherwise pass it as is.
+ */
+ if (!(info->flags & IEEE80211_TX_CTL_HW_80211_ENCAP))
+ ath10k_tx_h_8023(skb);
break;
case ATH10K_HW_TXRX_RAW:
if (!test_bit(ATH10K_FLAG_RAW_MODE, &ar->dev_flags) &&
@@ -4645,12 +4659,10 @@ static void ath10k_mac_op_tx(struct ieee80211_hw *hw,
struct ieee80211_vif *vif = info->control.vif;
struct ieee80211_sta *sta = control->sta;
struct ieee80211_txq *txq = NULL;
- struct ieee80211_hdr *hdr = (void *)skb->data;
enum ath10k_hw_txrx_mode txmode;
enum ath10k_mac_tx_path txpath;
bool is_htt;
bool is_mgmt;
- bool is_presp;
int ret;
u16 airtime;
@@ -4664,8 +4676,14 @@ static void ath10k_mac_op_tx(struct ieee80211_hw *hw,
is_mgmt = (txpath == ATH10K_MAC_TX_HTT_MGMT);
if (is_htt) {
+ bool is_presp = false;
+
spin_lock_bh(&ar->htt.tx_lock);
- is_presp = ieee80211_is_probe_resp(hdr->frame_control);
+ if (!(info->flags & IEEE80211_TX_CTL_HW_80211_ENCAP)) {
+ struct ieee80211_hdr *hdr = (void *)skb->data;
+
+ is_presp = ieee80211_is_probe_resp(hdr->frame_control);
+ }
ret = ath10k_htt_tx_inc_pending(htt);
if (ret) {
@@ -5465,6 +5483,30 @@ static int ath10k_mac_set_txbf_conf(struct ath10k_vif *arvif)
ar->wmi.vdev_param->txbf, value);
}
+static void ath10k_update_vif_offload(struct ieee80211_hw *hw,
+ struct ieee80211_vif *vif)
+{
+ struct ath10k_vif *arvif = (void *)vif->drv_priv;
+ struct ath10k *ar = hw->priv;
+ u32 vdev_param;
+ int ret;
+
+ if (ath10k_frame_mode != ATH10K_HW_TXRX_ETHERNET ||
+ ar->wmi.vdev_param->tx_encap_type == WMI_VDEV_PARAM_UNSUPPORTED ||
+ (vif->type != NL80211_IFTYPE_STATION &&
+ vif->type != NL80211_IFTYPE_AP))
+ vif->offload_flags &= ~IEEE80211_OFFLOAD_ENCAP_ENABLED;
+
+ vdev_param = ar->wmi.vdev_param->tx_encap_type;
+ ret = ath10k_wmi_vdev_set_param(ar, arvif->vdev_id, vdev_param,
+ ATH10K_HW_TXRX_NATIVE_WIFI);
+ /* 10.X firmware does not support this VDEV parameter. Do not warn */
+ if (ret && ret != -EOPNOTSUPP) {
+ ath10k_warn(ar, "failed to set vdev %i TX encapsulation: %d\n",
+ arvif->vdev_id, ret);
+ }
+}
+
/*
* TODO:
* Figure out how to handle WMI_VDEV_SUBTYPE_P2P_DEVICE,
@@ -5674,15 +5716,7 @@ static int ath10k_add_interface(struct ieee80211_hw *hw,
arvif->def_wep_key_idx = -1;
- vdev_param = ar->wmi.vdev_param->tx_encap_type;
- ret = ath10k_wmi_vdev_set_param(ar, arvif->vdev_id, vdev_param,
- ATH10K_HW_TXRX_NATIVE_WIFI);
- /* 10.X firmware does not support this VDEV parameter. Do not warn */
- if (ret && ret != -EOPNOTSUPP) {
- ath10k_warn(ar, "failed to set vdev %i TX encapsulation: %d\n",
- arvif->vdev_id, ret);
- goto err_vdev_delete;
- }
+ ath10k_update_vif_offload(hw, vif);
/* Configuring number of spatial stream for monitor interface is causing
* target assert in qca9888 and qca6174.
@@ -9375,6 +9409,7 @@ static const struct ieee80211_ops ath10k_ops = {
.stop = ath10k_stop,
.config = ath10k_config,
.add_interface = ath10k_add_interface,
+ .update_vif_offload = ath10k_update_vif_offload,
.remove_interface = ath10k_remove_interface,
.configure_filter = ath10k_configure_filter,
.bss_info_changed = ath10k_bss_info_changed,
@@ -10044,6 +10079,12 @@ int ath10k_mac_register(struct ath10k *ar)
if (test_bit(WMI_SERVICE_TDLS_UAPSD_BUFFER_STA, ar->wmi.svc_map))
ieee80211_hw_set(ar->hw, SUPPORTS_TDLS_BUFFER_STA);
+ if (ath10k_frame_mode == ATH10K_HW_TXRX_ETHERNET) {
+ if (ar->wmi.vdev_param->tx_encap_type !=
+ WMI_VDEV_PARAM_UNSUPPORTED)
+ ieee80211_hw_set(ar->hw, SUPPORTS_TX_ENCAP_OFFLOAD);
+ }
+
ar->hw->wiphy->flags |= WIPHY_FLAG_HAS_REMAIN_ON_CHANNEL;
ar->hw->wiphy->flags |= WIPHY_FLAG_HAS_CHANNEL_SWITCH;
ar->hw->wiphy->max_remain_on_channel_duration = 5000;
--
2.35.1
The xmit path for the Ethernet encapsulated frames become more or less
usable since d740d8fd2439 ("ath10k: unify tx mode and dispatch"). This
change reorganize the xmit path in a manageable way to properly support
various tx modes, but misses that the Ethernet encapsulated frame is a
special case. We do not have an IEEE 802.11 header at the begining of
them. But the HTT Tx handler still interprets first bytes of each frame
as an IEEE 802.11 Frame Control field.
Than this code was copied by e62ee5c381c5 ("ath10k: Add support for
htt_data_tx_desc_64 descriptor") and a2097d6444c3 ("ath10k: htt: High
latency TX support") to another handlers. In fact the issue in the high
latency (HL) handler was introduced by 83ac260151e7 ("ath10k: add mic
bytes for pmf management packet").
Ethernet encapsulated frame tx mode stay unused until 75d85fd9993c
("ath10k: introduce basic tdls functionality") started using it for TDLS
frames to avoid key selection issue in some firmwares.
Trying to interpret the begining of an Ethernet encapsulated frame as an
IEEE 802.11 header was not hurt us noticeably since we need to meet two
conditions: (1) xmit should be performed towards a TDLS peer, and (2)
the TDLS peer should have a specific OUI part of its MAC address. Looks
like that the rareness in TDLS communications of OUIs that can be
interpreted as an 802.11 management frame saves users from facing this
issue earlier.
Improve Ethernet tx mode support in the HTT Tx handler by avoiding
interpreting its first bytes as an IEEE 802.11 header. While at it, make
the ieee80211_hdr variable local to the code block that is guarded by
!is_eth check. In this way, we clarify in which cases a frame can be
interpreted as IEEE 802.11, and saves us from similar issues in the
future.
Credits: this change as part of xmit encapsulation offloading support
was originally made by QCA and then submitted for inclusion by John
Crispin [1]. But the whole work was not accepted due to the lack of a
part for 64-bits descriptors [2]. Zhijun You then pointed this out to me
in a reply to my initial RFC patch series. And I made this slightly
reworked version that covered all the HTT Tx handler variants.
1. https://lore.kernel.org/all/[email protected]/
2. https://patchwork.kernel.org/project/linux-wireless/patch/[email protected]/
Reported-by: Zhijun You <[email protected]>
Signed-off-by: Vasanthakumar Thiagarajan <[email protected]>
Signed-off-by: John Crispin <[email protected]>
Signed-off-by: Sergey Ryazanov <[email protected]>
---
Changes since RFC:
* new patch
drivers/net/wireless/ath/ath10k/htt_tx.c | 61 ++++++++++++++----------
1 file changed, 35 insertions(+), 26 deletions(-)
diff --git a/drivers/net/wireless/ath/ath10k/htt_tx.c b/drivers/net/wireless/ath/ath10k/htt_tx.c
index 9842a4b2f78f..a19b0795c86d 100644
--- a/drivers/net/wireless/ath/ath10k/htt_tx.c
+++ b/drivers/net/wireless/ath/ath10k/htt_tx.c
@@ -1275,7 +1275,6 @@ static int ath10k_htt_tx_hl(struct ath10k_htt *htt, enum ath10k_hw_txrx_mode txm
struct ath10k *ar = htt->ar;
int res, data_len;
struct htt_cmd_hdr *cmd_hdr;
- struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)msdu->data;
struct htt_data_tx_desc *tx_desc;
struct ath10k_skb_cb *skb_cb = ATH10K_SKB_CB(msdu);
struct sk_buff *tmp_skb;
@@ -1286,11 +1285,15 @@ static int ath10k_htt_tx_hl(struct ath10k_htt *htt, enum ath10k_hw_txrx_mode txm
u16 flags1 = 0;
u16 msdu_id = 0;
- if ((ieee80211_is_action(hdr->frame_control) ||
- ieee80211_is_deauth(hdr->frame_control) ||
- ieee80211_is_disassoc(hdr->frame_control)) &&
- ieee80211_has_protected(hdr->frame_control)) {
- skb_put(msdu, IEEE80211_CCMP_MIC_LEN);
+ if (!is_eth) {
+ struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)msdu->data;
+
+ if ((ieee80211_is_action(hdr->frame_control) ||
+ ieee80211_is_deauth(hdr->frame_control) ||
+ ieee80211_is_disassoc(hdr->frame_control)) &&
+ ieee80211_has_protected(hdr->frame_control)) {
+ skb_put(msdu, IEEE80211_CCMP_MIC_LEN);
+ }
}
data_len = msdu->len;
@@ -1387,7 +1390,6 @@ static int ath10k_htt_tx_32(struct ath10k_htt *htt,
{
struct ath10k *ar = htt->ar;
struct device *dev = ar->dev;
- struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)msdu->data;
struct ieee80211_tx_info *info = IEEE80211_SKB_CB(msdu);
struct ath10k_skb_cb *skb_cb = ATH10K_SKB_CB(msdu);
struct ath10k_hif_sg_item sg_items[2];
@@ -1419,15 +1421,19 @@ static int ath10k_htt_tx_32(struct ath10k_htt *htt,
txbuf_paddr = htt->txbuf.paddr +
(sizeof(struct ath10k_htt_txbuf_32) * msdu_id);
- if ((ieee80211_is_action(hdr->frame_control) ||
- ieee80211_is_deauth(hdr->frame_control) ||
- ieee80211_is_disassoc(hdr->frame_control)) &&
- ieee80211_has_protected(hdr->frame_control)) {
- skb_put(msdu, IEEE80211_CCMP_MIC_LEN);
- } else if (!(skb_cb->flags & ATH10K_SKB_F_NO_HWCRYPT) &&
- txmode == ATH10K_HW_TXRX_RAW &&
- ieee80211_has_protected(hdr->frame_control)) {
- skb_put(msdu, IEEE80211_CCMP_MIC_LEN);
+ if (!is_eth) {
+ struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)msdu->data;
+
+ if ((ieee80211_is_action(hdr->frame_control) ||
+ ieee80211_is_deauth(hdr->frame_control) ||
+ ieee80211_is_disassoc(hdr->frame_control)) &&
+ ieee80211_has_protected(hdr->frame_control)) {
+ skb_put(msdu, IEEE80211_CCMP_MIC_LEN);
+ } else if (!(skb_cb->flags & ATH10K_SKB_F_NO_HWCRYPT) &&
+ txmode == ATH10K_HW_TXRX_RAW &&
+ ieee80211_has_protected(hdr->frame_control)) {
+ skb_put(msdu, IEEE80211_CCMP_MIC_LEN);
+ }
}
skb_cb->paddr = dma_map_single(dev, msdu->data, msdu->len,
@@ -1589,7 +1595,6 @@ static int ath10k_htt_tx_64(struct ath10k_htt *htt,
{
struct ath10k *ar = htt->ar;
struct device *dev = ar->dev;
- struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)msdu->data;
struct ieee80211_tx_info *info = IEEE80211_SKB_CB(msdu);
struct ath10k_skb_cb *skb_cb = ATH10K_SKB_CB(msdu);
struct ath10k_hif_sg_item sg_items[2];
@@ -1621,15 +1626,19 @@ static int ath10k_htt_tx_64(struct ath10k_htt *htt,
txbuf_paddr = htt->txbuf.paddr +
(sizeof(struct ath10k_htt_txbuf_64) * msdu_id);
- if ((ieee80211_is_action(hdr->frame_control) ||
- ieee80211_is_deauth(hdr->frame_control) ||
- ieee80211_is_disassoc(hdr->frame_control)) &&
- ieee80211_has_protected(hdr->frame_control)) {
- skb_put(msdu, IEEE80211_CCMP_MIC_LEN);
- } else if (!(skb_cb->flags & ATH10K_SKB_F_NO_HWCRYPT) &&
- txmode == ATH10K_HW_TXRX_RAW &&
- ieee80211_has_protected(hdr->frame_control)) {
- skb_put(msdu, IEEE80211_CCMP_MIC_LEN);
+ if (!is_eth) {
+ struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)msdu->data;
+
+ if ((ieee80211_is_action(hdr->frame_control) ||
+ ieee80211_is_deauth(hdr->frame_control) ||
+ ieee80211_is_disassoc(hdr->frame_control)) &&
+ ieee80211_has_protected(hdr->frame_control)) {
+ skb_put(msdu, IEEE80211_CCMP_MIC_LEN);
+ } else if (!(skb_cb->flags & ATH10K_SKB_F_NO_HWCRYPT) &&
+ txmode == ATH10K_HW_TXRX_RAW &&
+ ieee80211_has_protected(hdr->frame_control)) {
+ skb_put(msdu, IEEE80211_CCMP_MIC_LEN);
+ }
}
skb_cb->paddr = dma_map_single(dev, msdu->data, msdu->len,
--
2.35.1
Hello Edward,
On Tue, May 17, 2022 at 5:37 AM Edward Matijevic <[email protected]> wrote:
> Hello,
> I confirm now my QCA9377 pci module has no regressions in either modes
> with these patches
> I was able to test it as an AP and it worked under ETHERNET mode
> without any obvious regressions but I'm unable to check maximum
> throughput
>
> My hardware info from dmesg and tested on kernel 5.17.7
> qca9377 hw1.0 target 0x05020000 chip_id 0x003820ff sub 11ad:0806
> firmware ver WLAN.TF.2.1-00021-QCARMSWP-1 api 6 features
> wowlan,ignore-otp crc32 42e41877
>
> My quoted statement of a possible regression seems to have been my old
> CPU acting up and is ok to remove
Thank you for the update.
--
Sergey
Hello,
I confirm now my QCA9377 pci module has no regressions in either modes
with these patches
I was able to test it as an AP and it worked under ETHERNET mode
without any obvious regressions but I'm unable to check maximum
throughput
My hardware info from dmesg and tested on kernel 5.17.7
qca9377 hw1.0 target 0x05020000 chip_id 0x003820ff sub 11ad:0806
firmware ver WLAN.TF.2.1-00021-QCARMSWP-1 api 6 features
wowlan,ignore-otp crc32 42e41877
My quoted statement of a possible regression seems to have been my old
CPU acting up and is ok to remove
On Sun, May 15, 2022 at 10:25 PM Sergey Ryazanov <[email protected]> wrote:
>
> Hello,
>
> this series introduces driver support for hardware encapsulation
> offloading feature.
>
> The main goal of the series is an overall improvement of system
> performance. On a QCA9563+QCA9888-based access point in bridged mode,
> encapsulation offloading increases TCP 16-streams DL throughput from
> 365 to 396 mbps (+8%) and UDP DL throughput from 436 to 483 mbps (+11%).
>
> The series consist of four patches, the first three prepare the code,
> and the last one introduces the offloading support itself. The first
> patch reworks transmission status reporting to make it flexible enough
> to support 802.11 and Ethernet encapsulated frames reporting. The second
> patch reworks improves Ethernet encapsulated frames handling in the HTT
> Tx layer. The third patch reworks the module parameter that configures
> the main encapsulation format of frames that are passed from the driver
> to the hardware. It makes it possible to configure more encapsulation
> methods than just raw or not-raw. And, as stated before, the fourth
> patch actually introduces offloading support. It changes a couple of
> frame analysis places along the xmit path and starts reporting
> offloading support to mac80211 via the corresponding hw attribute.
>
> The new feature has been extensively tested with QCA9888. It works well
> and introduces a measurable performance increase on platforms with not
> so fast CPU (e.g. QCA9563). Tests with MT7621+QCA6174 show no
> regression. I was unable to observe any measurable throughput increase
> in tests with QCA6174 due to the poor perfomance the SISO client and OTA
> testing. Only slight downlink rate stability increase with slight
> reducing of the SIRQ load on the AP side.
>
> Tests with QCA9880, QCA9980, QCA9984, QCA9377 revealed no regression
> [1]. No specific throughput tests have been performed. In OTA tests with
> a regular load, there is a slight increase in througput stability.
>
> In tests of the QCA9377 in Sta mode, Edward Matijevic noticed a possible
> regression in the Ethernet tx mode compared to the Native WiFi tx
> mode:
> > There might be a regression in download bandwidth under ETHERNET mode
> > but I didn't do enough tests to see if was my PC or the adapter as the
> > upload bandwidth was similar to other tests.
> > NATIVE_WIFI performs the same as without the patch.
>
> Credits: thanks to OpenWrt folks Oldřich Jedlička
> <[email protected]> and Edward Matijevic <[email protected]> who
> helped me a lot to perform extensive testing with various chips after a
> RFC/RFT post in the OpenWrt development mailing list [1]. Thanks to
> Zhijun You <[email protected]> who helped with the testing, pointed me
> to the previous work (see below) and the missed parts in the HTT Tx
> layer [1].
>
> Prior work: similar patches were independently developed by QCA and then
> submitted for inclusion by John Crispin <[email protected]> [2] and
> latter by Zhijun You <[email protected]> [3]. The work was not merged
> due to the lack of a part for 64-bits descriptors [4]. The idea of HTT
> Tx layer change from this work was adopted for the series in patch #2.
>
> 1. https://www.mail-archive.com/[email protected]/msg62013.html
> 2. https://lore.kernel.org/all/[email protected]/
> 3. https://lore.kernel.org/all/[email protected]/
> 4. https://patchwork.kernel.org/project/linux-wireless/patch/[email protected]/
>
> CC: Edward Matijevic <[email protected]>
> CC: John Crispin <[email protected]>
> CC: Oldřich Jedlička <[email protected]>
> CC: Tom Psyborg <[email protected]>
> CC: Vasanthakumar Thiagarajan <[email protected]>
> CC: Zhijun You <[email protected]>
>
> Changes since RFC:
> * new Tested-on and Tested-by tags
> * new patch #2 for better Ethernet encapsulation support in the HTT Tx
> layer
> * rebased on top of latest ath-next
>
> Sergey Ryazanov (3):
> ath10k: improve tx status reporting
> ath10k: htt_tx: do not interpret Eth frames as WiFi
> ath10k: turn rawmode into frame_mode
> ath10k: add encapsulation offloading support
>
> drivers/net/wireless/ath/ath10k/core.c | 11 ++--
> drivers/net/wireless/ath/ath10k/core.h | 1 +
> drivers/net/wireless/ath/ath10k/htt_tx.c | 61 ++++++++++++---------
> drivers/net/wireless/ath/ath10k/mac.c | 67 +++++++++++++++++++-----
> drivers/net/wireless/ath/ath10k/txrx.c | 12 ++++-
> 5 files changed, 108 insertions(+), 44 deletions(-)
>
> --
> 2.35.1
>