[PATCH net-next v3 0/2] Provide direct access to 1588 one step register

[PATCH net-next v3 0/2] Provide direct access to 1588 one step register

[Radu Bulie]

DPAA2 MAC supports 1588 one step timestamping.
If this option is enabled then for each transmitted PTP event packet,
the 1588 SINGLE_STEP register is accessed to modify the following fields:

-offset of the correction field inside the PTP packet
-UDP checksum update bit,  in case the PTP event packet has
 UDP encapsulation

These values can change any time, because there may be multiple
PTP clients connected, that receive various 1588 frame types:
- L2 only frame
- UDP / Ipv4
- UDP / Ipv6
- other

The current implementation uses dpni_set_single_step_cfg to update the
SINLGE_STEP register.
Using an MC command  on the Tx datapath for each transmitted 1588 message
introduces high delays, leading to low throughput and consequently to a
small number of supported PTP clients. Besides these, the nanosecond
correction field from the PTP packet will contain the high delay from the
driver which together with the originTimestamp will render timestamp
values that are unacceptable in a GM clock implementation.

This patch series replaces the dpni_set_single_step_cfg function call from
the Tx datapath for 1588 messages (when one step timestamping is enabled) 
with a callback that either implements direct access to the SINGLE_STEP
register, eliminating the overhead caused by the MC command that will need
to be dispatched by the MC firmware through the MC command portal
interface or falls back to the dpni_set_single_step_cfg in case the MC
version does not have support for returning the single step register
base address.

In other words all the delay introduced by dpni_set_single_step_cfg
function will be eliminated (if MC version has support for returning the
base address of the single step register), improving the egress driver
performance for PTP packets when single step timestamping is enabled.

The first patch adds a new attribute that contains the base address of
the SINGLE_STEP register. It will be used to directly update the register
on the Tx datapath.

The second patch updates the driver such that the SINGLE_STEP
register is either accessed directly if MC version >= 10.32 or is
accessed through dpni_set_single_step_cfg command when 1588 messages
are transmitted.

Changes in v2:
 - move global function pointer into the driver's private structure in 2/2
 - move repetitive code outside the body of the callback functions  in 2/2
 - update function dpaa2_ptp_onestep_reg_update_method  and remove goto 
   statement from non error path in 2/2	

Changes in v3:
 - remove static storage class specifier from within the structure in 2/2
 
Radu Bulie (2):
  dpaa2-eth: Update dpni_get_single_step_cfg command
  dpaa2-eth: Provide direct access to 1588 one step register

 .../net/ethernet/freescale/dpaa2/dpaa2-eth.c  | 96 +++++++++++++++++--
 .../net/ethernet/freescale/dpaa2/dpaa2-eth.h  | 12 ++-
 .../net/ethernet/freescale/dpaa2/dpni-cmd.h   |  6 +-
 drivers/net/ethernet/freescale/dpaa2/dpni.c   |  2 +
 drivers/net/ethernet/freescale/dpaa2/dpni.h   |  6 ++
 5 files changed, 110 insertions(+), 12 deletions(-)

-- 
2.17.1

[PATCH net-next v3 1/2] dpaa2-eth: Update dpni_get_single_step_cfg command

[Radu Bulie]

dpni_get_single_step_cfg is an MC firmware command used for
retrieving the contents of SINGLE_STEP 1588 register available
in a DPMAC.

This patch adds a new version of this command that returns as an extra
argument the physical base address of the aforementioned register.
The address will be used to directly modify the contents of the
SINGLE_STEP register instead of invoking the MC command
dpni_set_single_step_cgf. The former approach introduced huge delays on
the TX datapath when one step PTP events were transmitted. This led to low
throughput and high latencies observed in the PTP correction field.

Signed-off-by: Radu Bulie <radu-andrei.bulie@nxp.com>
---
Changes in v2:
 - none
Changes in v3:
 - none

 drivers/net/ethernet/freescale/dpaa2/dpni-cmd.h | 6 +++++-
 drivers/net/ethernet/freescale/dpaa2/dpni.c     | 2 ++
 drivers/net/ethernet/freescale/dpaa2/dpni.h     | 6 ++++++
 3 files changed, 13 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/freescale/dpaa2/dpni-cmd.h b/drivers/net/ethernet/freescale/dpaa2/dpni-cmd.h
index 9f80bdfeedec..828f538097af 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpni-cmd.h
+++ b/drivers/net/ethernet/freescale/dpaa2/dpni-cmd.h
@@ -98,7 +98,7 @@
 #define DPNI_CMDID_GET_LINK_CFG				DPNI_CMD(0x278)
 
 #define DPNI_CMDID_SET_SINGLE_STEP_CFG			DPNI_CMD(0x279)
-#define DPNI_CMDID_GET_SINGLE_STEP_CFG			DPNI_CMD(0x27a)
+#define DPNI_CMDID_GET_SINGLE_STEP_CFG			DPNI_CMD_V2(0x27a)
 
 /* Macros for accessing command fields smaller than 1byte */
 #define DPNI_MASK(field)	\
@@ -658,12 +658,16 @@ struct dpni_cmd_single_step_cfg {
 	__le16 flags;
 	__le16 offset;
 	__le32 peer_delay;
+	__le32 ptp_onestep_reg_base;
+	__le32 pad0;
 };
 
 struct dpni_rsp_single_step_cfg {
 	__le16 flags;
 	__le16 offset;
 	__le32 peer_delay;
+	__le32 ptp_onestep_reg_base;
+	__le32 pad0;
 };
 
 struct dpni_cmd_enable_vlan_filter {
diff --git a/drivers/net/ethernet/freescale/dpaa2/dpni.c b/drivers/net/ethernet/freescale/dpaa2/dpni.c
index d6afada99fb6..6c3b36f20fb8 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpni.c
+++ b/drivers/net/ethernet/freescale/dpaa2/dpni.c
@@ -2136,6 +2136,8 @@ int dpni_get_single_step_cfg(struct fsl_mc_io *mc_io,
 	ptp_cfg->ch_update = dpni_get_field(le16_to_cpu(rsp_params->flags),
 					    PTP_CH_UPDATE) ? 1 : 0;
 	ptp_cfg->peer_delay = le32_to_cpu(rsp_params->peer_delay);
+	ptp_cfg->ptp_onestep_reg_base =
+				  le32_to_cpu(rsp_params->ptp_onestep_reg_base);
 
 	return err;
 }
diff --git a/drivers/net/ethernet/freescale/dpaa2/dpni.h b/drivers/net/ethernet/freescale/dpaa2/dpni.h
index 7de0562bbf59..6fffd519aa00 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpni.h
+++ b/drivers/net/ethernet/freescale/dpaa2/dpni.h
@@ -1074,12 +1074,18 @@ int dpni_set_tx_shaping(struct fsl_mc_io *mc_io,
  * @peer_delay:	For peer-to-peer transparent clocks add this value to the
  *		correction field in addition to the transient time update.
  *		The value expresses nanoseconds.
+ * @ptp_onestep_reg_base: 1588 SINGLE_STEP register base address. This address
+ *			  is used to update directly the register contents.
+ *			  User has to create an address mapping for it.
+ *
+ *
  */
 struct dpni_single_step_cfg {
 	u8	en;
 	u8	ch_update;
 	u16	offset;
 	u32	peer_delay;
+	u32	ptp_onestep_reg_base;
 };
 
 int dpni_set_single_step_cfg(struct fsl_mc_io *mc_io,
-- 
2.17.1

[PATCH net-next v3 2/2] dpaa2-eth: Update SINGLE_STEP register access

[Radu Bulie]

DPAA2 MAC supports 1588 one step timestamping.
If this option is enabled then for each transmitted PTP event packet,
the 1588 SINGLE_STEP register is accessed to modify the following fields:

-offset of the correction field inside the PTP packet
-UDP checksum update bit,  in case the PTP event packet has
 UDP encapsulation

These values can change any time, because there may be multiple
PTP clients connected, that receive various 1588 frame types:
- L2 only frame
- UDP / Ipv4
- UDP / Ipv6
- other

The current implementation uses dpni_set_single_step_cfg to update the
SINLGE_STEP register.
Using an MC command  on the Tx datapath for each transmitted 1588 message
introduces high delays, leading to low throughput and consequently to a
small number of supported PTP clients. Besides these, the nanosecond
correction field from the PTP packet will contain the high delay from the
driver which together with the originTimestamp will render timestamp
values that are unacceptable in a GM clock implementation.

This patch updates the Tx datapath for 1588 messages when single step
timestamp is enabled and provides direct access to SINGLE_STEP register,
eliminating the  overhead caused by the dpni_set_single_step_cfg
MC command. MC version >= 10.32 implements this functionality.
If the MC version does not have support for returning the
single step register base address, the driver will use
dpni_set_single_step_cfg command for updates operations.

All the delay introduced by dpni_set_single_step_cfg
function will be eliminated (if MC version has support for returning the
base address of the single step register), improving the egress driver
performance for PTP packets when single step timestamping is enabled.

Before these changes the maximum throughput for 1588 messages with
single step hardware timestamp enabled was around 2000pps.
After the updates the throughput increased up to 32.82 Mbps / 46631.02 pps.

Signed-off-by: Radu Bulie <radu-andrei.bulie@nxp.com>
---
Changes in v2:
 - move global function pointer into the driver's private structure in 2/2
 - move repetitive code outside the body of the callback functions  in 2/2
 - update function dpaa2_ptp_onestep_reg_update_method  and remove goto
   statement from non error path in 2/2
Changes in v3:
 - remove static storage class specifier from within the structure in 2/2

 .../net/ethernet/freescale/dpaa2/dpaa2-eth.c  | 89 +++++++++++++++++--
 .../net/ethernet/freescale/dpaa2/dpaa2-eth.h  | 14 ++-
 2 files changed, 93 insertions(+), 10 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
index c4a48e6f1758..aab11d5da062 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
@@ -35,6 +35,75 @@ MODULE_DESCRIPTION("Freescale DPAA2 Ethernet Driver");
 struct ptp_qoriq *dpaa2_ptp;
 EXPORT_SYMBOL(dpaa2_ptp);
 
+static void dpaa2_eth_detect_features(struct dpaa2_eth_priv *priv)
+{
+	priv->features = 0;
+
+	if (dpaa2_eth_cmp_dpni_ver(priv, DPNI_PTP_ONESTEP_VER_MAJOR,
+				   DPNI_PTP_ONESTEP_VER_MINOR) >= 0)
+		priv->features |= DPAA2_ETH_FEATURE_ONESTEP_CFG_DIRECT;
+}
+
+static void dpaa2_update_ptp_onestep_indirect(struct dpaa2_eth_priv *priv,
+					      u32 offset, u8 udp)
+{
+	struct dpni_single_step_cfg cfg;
+
+	cfg.en = 1;
+	cfg.ch_update = udp;
+	cfg.offset = offset;
+	cfg.peer_delay = 0;
+
+	if (dpni_set_single_step_cfg(priv->mc_io, 0, priv->mc_token, &cfg))
+		WARN_ONCE(1, "Failed to set single step register");
+}
+
+static void dpaa2_update_ptp_onestep_direct(struct dpaa2_eth_priv *priv,
+					    u32 offset, u8 udp)
+{
+	u32 val = 0;
+
+	val = DPAA2_PTP_SINGLE_STEP_ENABLE |
+	       DPAA2_PTP_SINGLE_CORRECTION_OFF(offset);
+
+	if (udp)
+		val |= DPAA2_PTP_SINGLE_STEP_CH;
+
+	if (priv->onestep_reg_base)
+		writel(val, priv->onestep_reg_base);
+}
+
+static void dpaa2_ptp_onestep_reg_update_method(struct dpaa2_eth_priv *priv)
+{
+	struct device *dev = priv->net_dev->dev.parent;
+	struct dpni_single_step_cfg ptp_cfg;
+
+	priv->dpaa2_set_onestep_params_cb = dpaa2_update_ptp_onestep_indirect;
+
+	if (!(priv->features & DPAA2_ETH_FEATURE_ONESTEP_CFG_DIRECT))
+		return;
+
+	if (dpni_get_single_step_cfg(priv->mc_io, 0,
+				     priv->mc_token, &ptp_cfg)) {
+		dev_err(dev, "dpni_get_single_step_cfg cannot retrieve onestep reg, falling back to indirect update\n");
+		return;
+	}
+
+	if (!ptp_cfg.ptp_onestep_reg_base) {
+		dev_err(dev, "1588 onestep reg not available, falling back to indirect update\n");
+		return;
+	}
+
+	priv->onestep_reg_base = ioremap(ptp_cfg.ptp_onestep_reg_base,
+					 sizeof(u32));
+	if (!priv->onestep_reg_base) {
+		dev_err(dev, "1588 onestep reg cannot be mapped, falling back to indirect update\n");
+		return;
+	}
+
+	priv->dpaa2_set_onestep_params_cb = dpaa2_update_ptp_onestep_direct;
+}
+
 static void *dpaa2_iova_to_virt(struct iommu_domain *domain,
 				dma_addr_t iova_addr)
 {
@@ -696,7 +765,6 @@ static void dpaa2_eth_enable_tx_tstamp(struct dpaa2_eth_priv *priv,
 				       struct sk_buff *skb)
 {
 	struct ptp_tstamp origin_timestamp;
-	struct dpni_single_step_cfg cfg;
 	u8 msgtype, twostep, udp;
 	struct dpaa2_faead *faead;
 	struct dpaa2_fas *fas;
@@ -750,14 +818,12 @@ static void dpaa2_eth_enable_tx_tstamp(struct dpaa2_eth_priv *priv,
 			htonl(origin_timestamp.sec_lsb);
 		*(__be32 *)(data + offset2 + 6) = htonl(origin_timestamp.nsec);
 
-		cfg.en = 1;
-		cfg.ch_update = udp;
-		cfg.offset = offset1;
-		cfg.peer_delay = 0;
+		if (priv->ptp_correction_off == offset1)
+			return;
+
+		priv->dpaa2_set_onestep_params_cb(priv, offset1, udp);
+		priv->ptp_correction_off = offset1;
 
-		if (dpni_set_single_step_cfg(priv->mc_io, 0, priv->mc_token,
-					     &cfg))
-			WARN_ONCE(1, "Failed to set single step register");
 	}
 }
 
@@ -2407,6 +2473,9 @@ static int dpaa2_eth_ts_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
 		config.rx_filter = HWTSTAMP_FILTER_ALL;
 	}
 
+	if (priv->tx_tstamp_type == HWTSTAMP_TX_ONESTEP_SYNC)
+		dpaa2_ptp_onestep_reg_update_method(priv);
+
 	return copy_to_user(rq->ifr_data, &config, sizeof(config)) ?
 			-EFAULT : 0;
 }
@@ -4300,6 +4369,8 @@ static int dpaa2_eth_netdev_init(struct net_device *net_dev)
 		return err;
 	}
 
+	dpaa2_eth_detect_features(priv);
+
 	/* Capabilities listing */
 	supported |= IFF_LIVE_ADDR_CHANGE;
 
@@ -4758,6 +4829,8 @@ static int dpaa2_eth_remove(struct fsl_mc_device *ls_dev)
 	dpaa2_eth_free_dpbp(priv);
 	dpaa2_eth_free_dpio(priv);
 	dpaa2_eth_free_dpni(priv);
+	if (priv->onestep_reg_base)
+		iounmap(priv->onestep_reg_base);
 
 	fsl_mc_portal_free(priv->mc_io);
 
diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
index b79831cd1a94..447718483ef4 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
@@ -526,12 +526,15 @@ struct dpaa2_eth_priv {
 	u8 num_channels;
 	struct dpaa2_eth_channel *channel[DPAA2_ETH_MAX_DPCONS];
 	struct dpaa2_eth_sgt_cache __percpu *sgt_cache;
-
+	unsigned long features;
 	struct dpni_attr dpni_attrs;
 	u16 dpni_ver_major;
 	u16 dpni_ver_minor;
 	u16 tx_data_offset;
-
+	void __iomem *onestep_reg_base;
+	u8 ptp_correction_off;
+	void (*dpaa2_set_onestep_params_cb)(struct dpaa2_eth_priv *priv,
+					    u32 offset, u8 udp);
 	struct fsl_mc_device *dpbp_dev;
 	u16 rx_buf_size;
 	u16 bpid;
@@ -673,6 +676,13 @@ enum dpaa2_eth_rx_dist {
 #define DPAA2_ETH_DIST_L4DST		BIT(8)
 #define DPAA2_ETH_DIST_ALL		(~0ULL)
 
+#define DPNI_PTP_ONESTEP_VER_MAJOR 8
+#define DPNI_PTP_ONESTEP_VER_MINOR 2
+#define DPAA2_ETH_FEATURE_ONESTEP_CFG_DIRECT BIT(0)
+#define DPAA2_PTP_SINGLE_STEP_ENABLE	BIT(31)
+#define DPAA2_PTP_SINGLE_STEP_CH	BIT(7)
+#define DPAA2_PTP_SINGLE_CORRECTION_OFF(v) ((v) << 8)
+
 #define DPNI_PAUSE_VER_MAJOR		7
 #define DPNI_PAUSE_VER_MINOR		13
 #define dpaa2_eth_has_pause_support(priv)			\
-- 
2.17.1

Re: [PATCH net-next v3 0/2] Provide direct access to 1588 one step register

[patchwork-bot+netdevbpf]

Hello:

This series was applied to netdev/net-next.git (master)
by David S. Miller <davem@davemloft.net>:

On Fri, 18 Feb 2022 22:21:59 +0200 you wrote:
> DPAA2 MAC supports 1588 one step timestamping.
> If this option is enabled then for each transmitted PTP event packet,
> the 1588 SINGLE_STEP register is accessed to modify the following fields:
> 
> -offset of the correction field inside the PTP packet
> -UDP checksum update bit,  in case the PTP event packet has
>  UDP encapsulation
> 
> [...]

Here is the summary with links:
  - [net-next,v3,1/2] dpaa2-eth: Update dpni_get_single_step_cfg command
    https://git.kernel.org/netdev/net-next/c/9572594ecf02
  - [net-next,v3,2/2] dpaa2-eth: Update SINGLE_STEP register access
    https://git.kernel.org/netdev/net-next/c/c4680c978567

You are awesome, thank you!
-- 
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html