[PATCH net-next 1/2] net: extend ndo_get_tstamp for other timestamp types

[PATCH net-next 1/2] net: extend ndo_get_tstamp for other timestamp types

[Kevin Yang]

Network device hardware timestamps (hwtstamps) and the system's
clock (ktime) often originate from different clock domains.
This makes it hard to directly calculate the duration between
a hardware-timestamped event and a system-time event by simple
subtraction.

This patch extends ndo_get_tstamp to allow a netdev to provide
a hwtstamp into the system's CLOCK_REALTIME domain. This allows a
driver to either perform a conversion by estimating or, if the
clocks are kept synchronized, return the original timestamp directly.
Other clock domains, e.g. CLOCK_MONOTONIC_RAW can also be added when
a use surfaces.

This is useful for features that need to measure the delay between
a packet's hardware arrival/departure and a later software event.
For example, the TCP stack can use this to measure precise
packet receive delays, which is a requirement for the upcoming
TCP Swift [1] congestion control algorithm.

[1] Kumar, Gautam, et al. "Swift: Delay is simple and effective
for congestion control in the datacenter." Proceedings of the
Annual conference of the ACM Special Interest Group on Data
Communication on the applications, technologies, architectures,
and protocols for computer communication. 2020.

Signed-off-by: Kevin Yang <yyd@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
---
 drivers/net/ethernet/engleder/tsnep_main.c |  8 +++++---
 drivers/net/ethernet/intel/igc/igc_main.c  |  8 +++++---
 include/linux/netdevice.h                  | 21 ++++++++++++++-------
 3 files changed, 24 insertions(+), 13 deletions(-)

diff --git a/drivers/net/ethernet/engleder/tsnep_main.c b/drivers/net/ethernet/engleder/tsnep_main.c
index b118407c30e87..7ae697fe51cf6 100644
--- a/drivers/net/ethernet/engleder/tsnep_main.c
+++ b/drivers/net/ethernet/engleder/tsnep_main.c
@@ -2275,15 +2275,17 @@ static int tsnep_netdev_set_features(struct net_device *netdev,
 
 static ktime_t tsnep_netdev_get_tstamp(struct net_device *netdev,
 				       const struct skb_shared_hwtstamps *hwtstamps,
-				       bool cycles)
+				       enum netdev_tstamp_type type)
 {
 	struct tsnep_rx_inline *rx_inline = hwtstamps->netdev_data;
 	u64 timestamp;
 
-	if (cycles)
+	if (type == NETDEV_TSTAMP_CYCLE)
 		timestamp = __le64_to_cpu(rx_inline->counter);
-	else
+	else if (type == NETDEV_TSTAMP_RAW)
 		timestamp = __le64_to_cpu(rx_inline->timestamp);
+	else
+		return 0;
 
 	return ns_to_ktime(timestamp);
 }
diff --git a/drivers/net/ethernet/intel/igc/igc_main.c b/drivers/net/ethernet/intel/igc/igc_main.c
index 7aafa60ba0c86..c233e78f474f1 100644
--- a/drivers/net/ethernet/intel/igc/igc_main.c
+++ b/drivers/net/ethernet/intel/igc/igc_main.c
@@ -6947,7 +6947,7 @@ int igc_xsk_wakeup(struct net_device *dev, u32 queue_id, u32 flags)
 
 static ktime_t igc_get_tstamp(struct net_device *dev,
 			      const struct skb_shared_hwtstamps *hwtstamps,
-			      bool cycles)
+			      enum netdev_tstamp_type type)
 {
 	struct igc_adapter *adapter = netdev_priv(dev);
 	struct igc_inline_rx_tstamps *tstamp;
@@ -6955,10 +6955,12 @@ static ktime_t igc_get_tstamp(struct net_device *dev,
 
 	tstamp = hwtstamps->netdev_data;
 
-	if (cycles)
+	if (type == NETDEV_TSTAMP_CYCLE)
 		timestamp = igc_ptp_rx_pktstamp(adapter, tstamp->timer1);
-	else
+	else if (type == NETDEV_TSTAMP_RAW)
 		timestamp = igc_ptp_rx_pktstamp(adapter, tstamp->timer0);
+	else
+		return 0;
 
 	return timestamp;
 }
diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index d99b0fbc1942a..1c1c7dcb8e801 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -1062,6 +1062,12 @@ struct netdev_net_notifier {
 	struct notifier_block *nb;
 };
 
+enum netdev_tstamp_type {
+	NETDEV_TSTAMP_RAW = 0,
+	NETDEV_TSTAMP_CYCLE,
+	NETDEV_TSTAMP_REALTIME,
+};
+
 /*
  * This structure defines the management hooks for network devices.
  * The following hooks can be defined; unless noted otherwise, they are
@@ -1406,11 +1412,10 @@ struct netdev_net_notifier {
  *     Get the forwarding path to reach the real device from the HW destination address
  * ktime_t (*ndo_get_tstamp)(struct net_device *dev,
  *			     const struct skb_shared_hwtstamps *hwtstamps,
- *			     bool cycles);
- *	Get hardware timestamp based on normal/adjustable time or free running
- *	cycle counter. This function is required if physical clock supports a
- *	free running cycle counter.
- *
+ *			     enum netdev_tstamp_type type);
+ *	Get hardware timestamp based on the type requested, or return 0 if the
+ *	requested type is not supported. This function is required if physical
+ *	clock supports a free running cycle counter.
  * int (*ndo_hwtstamp_get)(struct net_device *dev,
  *			   struct kernel_hwtstamp_config *kernel_config);
  *	Get the currently configured hardware timestamping parameters for the
@@ -1661,7 +1666,7 @@ struct net_device_ops {
                                                          struct net_device_path *path);
 	ktime_t			(*ndo_get_tstamp)(struct net_device *dev,
 						  const struct skb_shared_hwtstamps *hwtstamps,
-						  bool cycles);
+						  enum netdev_tstamp_type type);
 	int			(*ndo_hwtstamp_get)(struct net_device *dev,
 						    struct kernel_hwtstamp_config *kernel_config);
 	int			(*ndo_hwtstamp_set)(struct net_device *dev,
@@ -5236,9 +5241,11 @@ static inline ktime_t netdev_get_tstamp(struct net_device *dev,
 					bool cycles)
 {
 	const struct net_device_ops *ops = dev->netdev_ops;
+	enum netdev_tstamp_type type = cycles ? NETDEV_TSTAMP_CYCLE :
+						NETDEV_TSTAMP_RAW;
 
 	if (ops->ndo_get_tstamp)
-		return ops->ndo_get_tstamp(dev, hwtstamps, cycles);
+		return ops->ndo_get_tstamp(dev, hwtstamps, type);
 
 	return hwtstamps->hwtstamp;
 }
-- 
2.52.0.457.g6b5491de43-goog

[PATCH net-next 2/2] gve: implement ndo_get_tstamp

[Kevin Yang]

This patch implements ndo_get_tstamp in gve to support converting a
hwtstamp to the system's realtime clock.

The implementation does not assume the NIC clock is disciplined,
in other word, the NIC clock can be free-running. A periodic
job, embedded in gve's ptp_aux_work, updates the offset and slope
for the conversion.

Signed-off-by: Kevin Yang <yyd@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Harshitha Ramamurthy <hramamurthy@google.com>
---
 drivers/net/ethernet/google/gve/gve.h        |   8 ++
 drivers/net/ethernet/google/gve/gve_adminq.h |   4 +-
 drivers/net/ethernet/google/gve/gve_main.c   |  27 +++++
 drivers/net/ethernet/google/gve/gve_ptp.c    | 106 ++++++++++++++++++-
 4 files changed, 142 insertions(+), 3 deletions(-)

diff --git a/drivers/net/ethernet/google/gve/gve.h b/drivers/net/ethernet/google/gve/gve.h
index 970d5ca8cddee..13a4c450e7635 100644
--- a/drivers/net/ethernet/google/gve/gve.h
+++ b/drivers/net/ethernet/google/gve/gve.h
@@ -774,6 +774,13 @@ struct gve_flow_rule {
 	struct gve_flow_spec mask;
 };
 
+struct gve_tstamp_conversion {
+	u64 last_sync_ns;
+	seqlock_t lock; /* protects tc and cc */
+	struct timecounter tc;
+	struct cyclecounter cc;
+};
+
 struct gve_flow_rules_cache {
 	bool rules_cache_synced; /* False if the driver's rules_cache is outdated */
 	struct gve_adminq_queried_flow_rule *rules_cache;
@@ -925,6 +932,7 @@ struct gve_priv {
 	struct gve_nic_ts_report *nic_ts_report;
 	dma_addr_t nic_ts_report_bus;
 	u64 last_sync_nic_counter; /* Clock counter from last NIC TS report */
+	struct gve_tstamp_conversion ts_real;
 };
 
 enum gve_service_task_flags_bit {
diff --git a/drivers/net/ethernet/google/gve/gve_adminq.h b/drivers/net/ethernet/google/gve/gve_adminq.h
index 22a74b6aa17ea..812160b87b143 100644
--- a/drivers/net/ethernet/google/gve/gve_adminq.h
+++ b/drivers/net/ethernet/google/gve/gve_adminq.h
@@ -411,8 +411,8 @@ static_assert(sizeof(struct gve_adminq_report_nic_ts) == 16);
 
 struct gve_nic_ts_report {
 	__be64 nic_timestamp; /* NIC clock in nanoseconds */
-	__be64 reserved1;
-	__be64 reserved2;
+	__be64 cycle_pre;
+	__be64 cycle_post;
 	__be64 reserved3;
 	__be64 reserved4;
 };
diff --git a/drivers/net/ethernet/google/gve/gve_main.c b/drivers/net/ethernet/google/gve/gve_main.c
index 7eb64e1e4d858..2acc1a3d85838 100644
--- a/drivers/net/ethernet/google/gve/gve_main.c
+++ b/drivers/net/ethernet/google/gve/gve_main.c
@@ -2198,6 +2198,32 @@ static int gve_set_ts_config(struct net_device *dev,
 	return 0;
 }
 
+static ktime_t gve_get_tstamp(struct net_device *dev,
+			      const struct skb_shared_hwtstamps *hwtstamps,
+			      enum netdev_tstamp_type type)
+{
+	struct gve_priv *priv = netdev_priv(dev);
+	unsigned int seq;
+	u64 ns;
+
+	if (type == NETDEV_TSTAMP_RAW)
+		return hwtstamps->hwtstamp;
+
+	if (type != NETDEV_TSTAMP_REALTIME)
+		return 0;
+
+	/* Skip if never synced */
+	if (!READ_ONCE(priv->ts_real.last_sync_ns))
+		return 0;
+
+	do {
+		seq = read_seqbegin(&priv->ts_real.lock);
+		ns = timecounter_cyc2time(&priv->ts_real.tc,
+					  hwtstamps->hwtstamp);
+	} while (read_seqretry(&priv->ts_real.lock, seq));
+	return ns_to_ktime(ns);
+}
+
 static const struct net_device_ops gve_netdev_ops = {
 	.ndo_start_xmit		=	gve_start_xmit,
 	.ndo_features_check	=	gve_features_check,
@@ -2209,6 +2235,7 @@ static const struct net_device_ops gve_netdev_ops = {
 	.ndo_bpf		=	gve_xdp,
 	.ndo_xdp_xmit		=	gve_xdp_xmit,
 	.ndo_xsk_wakeup		=	gve_xsk_wakeup,
+	.ndo_get_tstamp		=	gve_get_tstamp,
 	.ndo_hwtstamp_get	=	gve_get_ts_config,
 	.ndo_hwtstamp_set	=	gve_set_ts_config,
 };
diff --git a/drivers/net/ethernet/google/gve/gve_ptp.c b/drivers/net/ethernet/google/gve/gve_ptp.c
index 073677d82ee8e..df32735fa940f 100644
--- a/drivers/net/ethernet/google/gve/gve_ptp.c
+++ b/drivers/net/ethernet/google/gve/gve_ptp.c
@@ -10,10 +10,91 @@
 /* Interval to schedule a nic timestamp calibration, 250ms. */
 #define GVE_NIC_TS_SYNC_INTERVAL_MS 250
 
+/* Scale ts_real.cc.mult by 1 << 31. Maximize mult for finer adjustment
+ * granularity, but ensure (mult * cycle) does not overflow in
+ * cyclecounter_cyc2ns.
+ */
+#define GVE_HWTS_REAL_CC_SHIFT 31
+#define GVE_HWTS_REAL_CC_NOMINAL BIT_ULL(GVE_HWTS_REAL_CC_SHIFT)
+
+/* Get the cross time stamp info */
+static int gve_get_cross_time(ktime_t *device,
+			      struct system_counterval_t *system, void *ctx)
+{
+	struct gve_priv *priv = ctx;
+
+	*device = ns_to_ktime(be64_to_cpu(priv->nic_ts_report->nic_timestamp));
+	system->cycles = be64_to_cpu(priv->nic_ts_report->cycle_pre) +
+			 (be64_to_cpu(priv->nic_ts_report->cycle_post) -
+			  be64_to_cpu(priv->nic_ts_report->cycle_pre)) / 2;
+	system->use_nsecs = false;
+	if (IS_ENABLED(CONFIG_X86))
+		system->cs_id = CSID_X86_TSC;
+	else if (IS_ENABLED(CONFIG_ARM_ARCH_TIMER))
+		system->cs_id = CSID_ARM_ARCH_COUNTER;
+	else
+		return -EOPNOTSUPP;
+
+	return 0;
+}
+
+static int gve_hwts_realtime_update(struct gve_priv *priv, u64 prev_nic)
+{
+	struct system_device_crosststamp cts = {};
+	struct system_time_snapshot history = {};
+	s64 nic_real_off_ns;
+	u64 real_ns;
+	int ret;
+
+	/* Step 1: Get the realtime of when NIC clock was read */
+	ktime_get_snapshot(&history);
+	ret = get_device_system_crosststamp(gve_get_cross_time, priv, &history,
+					    &cts);
+	if (ret) {
+		dev_err_ratelimited(&priv->pdev->dev,
+				    "%s crosststamp err %d\n", __func__, ret);
+		return ret;
+	}
+
+	real_ns = ktime_to_ns(cts.sys_realtime);
+
+	/* Step 2: Adjust NIC clock's offset */
+	/* Read-side ndo_get_tstamp can be called from TCP rx softirq */
+	write_seqlock_bh(&priv->ts_real.lock);
+	nic_real_off_ns = real_ns - timecounter_read(&priv->ts_real.tc);
+	timecounter_adjtime(&priv->ts_real.tc, nic_real_off_ns);
+
+	/* Step 3: Adjust NIC clock's ratio (when this is not the first sync).
+	 * The NIC clock's nominal tick ratio is 1 tick per nanosecond,
+	 * scaled by 1 << GVE_HWTS_REAL_CC_SHIFT. Adjust it to
+	 * (ktime - prev_ktime) / (nic - prev_nic). The ratio should not
+	 * deviate more than 1% from the nominal, otherwise it may suggest
+	 * there was a sudden change on NIC clock. In that case, reset ratio
+	 * to nominal. And since each sync only compares to the previous read,
+	 * this is a one-time error, not a persistent failure.
+	 */
+	if (prev_nic) {
+		const u64 lower = GVE_HWTS_REAL_CC_NOMINAL * 99 / 100;
+		const u64 upper = GVE_HWTS_REAL_CC_NOMINAL * 101 / 100;
+		u64 mult;
+
+		mult = mult_frac(GVE_HWTS_REAL_CC_NOMINAL,
+				 real_ns - priv->ts_real.last_sync_ns,
+				 priv->last_sync_nic_counter - prev_nic);
+		if (mult < lower || mult > upper)
+			mult = GVE_HWTS_REAL_CC_NOMINAL;
+		priv->ts_real.cc.mult = mult;
+	}
+
+	write_sequnlock_bh(&priv->ts_real.lock);
+	WRITE_ONCE(priv->ts_real.last_sync_ns, real_ns);
+	return 0;
+}
+
 /* Read the nic timestamp from hardware via the admin queue. */
 int gve_clock_nic_ts_read(struct gve_priv *priv)
 {
-	u64 nic_raw;
+	u64 nic_raw, prev_nic;
 	int err;
 
 	err = gve_adminq_report_nic_ts(priv, priv->nic_ts_report_bus);
@@ -21,7 +102,11 @@ int gve_clock_nic_ts_read(struct gve_priv *priv)
 		return err;
 
 	nic_raw = be64_to_cpu(priv->nic_ts_report->nic_timestamp);
+	prev_nic = priv->last_sync_nic_counter;
 	WRITE_ONCE(priv->last_sync_nic_counter, nic_raw);
+	err = gve_hwts_realtime_update(priv, prev_nic);
+	if (err)
+		return err;
 
 	return 0;
 }
@@ -57,6 +142,14 @@ static long gve_ptp_do_aux_work(struct ptp_clock_info *info)
 	return msecs_to_jiffies(GVE_NIC_TS_SYNC_INTERVAL_MS);
 }
 
+static u64 gve_cycles_read(struct cyclecounter *cc)
+{
+	const struct gve_priv *priv = container_of(cc, struct gve_priv,
+						   ts_real.cc);
+
+	return READ_ONCE(priv->last_sync_nic_counter);
+}
+
 static const struct ptp_clock_info gve_ptp_caps = {
 	.owner          = THIS_MODULE,
 	.name		= "gve clock",
@@ -89,6 +182,17 @@ static int gve_ptp_init(struct gve_priv *priv)
 		goto free_ptp;
 	}
 
+	priv->last_sync_nic_counter = 0;
+	priv->ts_real.last_sync_ns = 0;
+	seqlock_init(&priv->ts_real.lock);
+	memset(&priv->ts_real.cc, 0, sizeof(priv->ts_real.cc));
+	priv->ts_real.cc.mask = U32_MAX;
+	priv->ts_real.cc.shift = GVE_HWTS_REAL_CC_SHIFT;
+	priv->ts_real.cc.mult = GVE_HWTS_REAL_CC_NOMINAL;
+	priv->ts_real.cc.read = gve_cycles_read;
+	timecounter_init(&priv->ts_real.tc, &priv->ts_real.cc,
+			 ktime_get_real_ns());
+
 	ptp->priv = priv;
 	return 0;
 
-- 
2.52.0.457.g6b5491de43-goog

Re: [PATCH net-next 1/2] net: extend ndo_get_tstamp for other timestamp types

[Gerhard Engleder]

On 15.01.26 23:22, Kevin Yang wrote:
> Network device hardware timestamps (hwtstamps) and the system's
> clock (ktime) often originate from different clock domains.
> This makes it hard to directly calculate the duration between
> a hardware-timestamped event and a system-time event by simple
> subtraction.
> 
> This patch extends ndo_get_tstamp to allow a netdev to provide
> a hwtstamp into the system's CLOCK_REALTIME domain. This allows a
> driver to either perform a conversion by estimating or, if the
> clocks are kept synchronized, return the original timestamp directly.
> Other clock domains, e.g. CLOCK_MONOTONIC_RAW can also be added when
> a use surfaces.
> 
> This is useful for features that need to measure the delay between
> a packet's hardware arrival/departure and a later software event.
> For example, the TCP stack can use this to measure precise
> packet receive delays, which is a requirement for the upcoming
> TCP Swift [1] congestion control algorithm.
> 
> [1] Kumar, Gautam, et al. "Swift: Delay is simple and effective
> for congestion control in the datacenter." Proceedings of the
> Annual conference of the ACM Special Interest Group on Data
> Communication on the applications, technologies, architectures,
> and protocols for computer communication. 2020.
> 
> Signed-off-by: Kevin Yang <yyd@google.com>
> Reviewed-by: Willem de Bruijn <willemb@google.com>
> ---
>   drivers/net/ethernet/engleder/tsnep_main.c |  8 +++++---
>   drivers/net/ethernet/intel/igc/igc_main.c  |  8 +++++---
>   include/linux/netdevice.h                  | 21 ++++++++++++++-------
>   3 files changed, 24 insertions(+), 13 deletions(-)
>


...

>   
> +enum netdev_tstamp_type {
> +	NETDEV_TSTAMP_RAW = 0,
> +	NETDEV_TSTAMP_CYCLE,
> +	NETDEV_TSTAMP_REALTIME,
> +};

Would it make sense to document the types? By your change of the
ndo_get_tstamp() documentation below, the info that cycles is
about a free running cycle counter is lost.

IMO the name RAW for the timestamp of the adjustable physical clock is
not a good fit, because NETDEV_TSTAMP_CYCLE is more raw (not adjustable)
than NETDEV_TSTAMP_RAW (adjustable). Maybe NETDEV_TSTAMP_PHC instead of
NETDEV_TSTAMP_RAW?

If I understand your second patch correctly, then NETDEV_TSTAMP_REALTIME
can provide a timestamp with low accuracy, which can even jump backwards
(a later received package can get an earlier timestamp due to the simple
synchronisation in gve_hwts_realtime_update). Should the expected
quality of the timestamps be documented here?

> +
>   /*
>    * This structure defines the management hooks for network devices.
>    * The following hooks can be defined; unless noted otherwise, they are
> @@ -1406,11 +1412,10 @@ struct netdev_net_notifier {
>    *     Get the forwarding path to reach the real device from the HW destination address
>    * ktime_t (*ndo_get_tstamp)(struct net_device *dev,
>    *			     const struct skb_shared_hwtstamps *hwtstamps,
> - *			     bool cycles);
> - *	Get hardware timestamp based on normal/adjustable time or free running
> - *	cycle counter. This function is required if physical clock supports a
> - *	free running cycle counter.
> - *
> + *			     enum netdev_tstamp_type type);
> + *	Get hardware timestamp based on the type requested, or return 0 if the
> + *	requested type is not supported. This function is required if physical
> + *	clock supports a free running cycle counter.

...

Gerhard

Re: [PATCH net-next 1/2] net: extend ndo_get_tstamp for other timestamp types

[Jakub Kicinski]

On Thu, 15 Jan 2026 22:22:59 +0000 Kevin Yang wrote:
> Reviewed-by: Willem de Bruijn <willemb@google.com>

Other than expediency why implementing this in drivers makes sense?
There will be very little driver specific logic here.

After thinking about it for 1 minute my intuition would be for drivers
to just expose necessary information, in-kernel consumers should
"enable" the sync for all netdevs during init, and then convert
the timestamps into REALTIME / MONOTONIC by calling some helper?

Also AI code review says:

> +		mult = mult_frac(GVE_HWTS_REAL_CC_NOMINAL,
> +				 real_ns - priv->ts_real.last_sync_ns,
> +				 priv->last_sync_nic_counter - prev_nic);

Can this divide by zero? 
-- 
pw-bot: cr

Re: [PATCH net-next 1/2] net: extend ndo_get_tstamp for other timestamp types

[Kevin Yang]

On Fri, Jan 16, 2026 at 3:36 PM Gerhard Engleder
<gerhard@engleder-embedded.com> wrote:
>
> On 15.01.26 23:22, Kevin Yang wrote:
> > Network device hardware timestamps (hwtstamps) and the system's
> > clock (ktime) often originate from different clock domains.
> > This makes it hard to directly calculate the duration between
> > a hardware-timestamped event and a system-time event by simple
> > subtraction.
> >
> > This patch extends ndo_get_tstamp to allow a netdev to provide
> > a hwtstamp into the system's CLOCK_REALTIME domain. This allows a
> > driver to either perform a conversion by estimating or, if the
> > clocks are kept synchronized, return the original timestamp directly.
> > Other clock domains, e.g. CLOCK_MONOTONIC_RAW can also be added when
> > a use surfaces.
> >
> > This is useful for features that need to measure the delay between
> > a packet's hardware arrival/departure and a later software event.
> > For example, the TCP stack can use this to measure precise
> > packet receive delays, which is a requirement for the upcoming
> > TCP Swift [1] congestion control algorithm.
> >
> > [1] Kumar, Gautam, et al. "Swift: Delay is simple and effective
> > for congestion control in the datacenter." Proceedings of the
> > Annual conference of the ACM Special Interest Group on Data
> > Communication on the applications, technologies, architectures,
> > and protocols for computer communication. 2020.
> >
> > Signed-off-by: Kevin Yang <yyd@google.com>
> > Reviewed-by: Willem de Bruijn <willemb@google.com>
> > ---
> >   drivers/net/ethernet/engleder/tsnep_main.c |  8 +++++---
> >   drivers/net/ethernet/intel/igc/igc_main.c  |  8 +++++---
> >   include/linux/netdevice.h                  | 21 ++++++++++++++-------
> >   3 files changed, 24 insertions(+), 13 deletions(-)
> >
>
>
> ...
>
> >
> > +enum netdev_tstamp_type {
> > +     NETDEV_TSTAMP_RAW = 0,
> > +     NETDEV_TSTAMP_CYCLE,
> > +     NETDEV_TSTAMP_REALTIME,
> > +};
>
> Would it make sense to document the types? By your change of the
> ndo_get_tstamp() documentation below, the info that cycles is
> about a free running cycle counter is lost.
Good point. I will add a section in kdoc.

>
> IMO the name RAW for the timestamp of the adjustable physical clock is
> not a good fit, because NETDEV_TSTAMP_CYCLE is more raw (not adjustable)
> than NETDEV_TSTAMP_RAW (adjustable). Maybe NETDEV_TSTAMP_PHC instead of
> NETDEV_TSTAMP_RAW?
Agreed, will do in v2.

>
> If I understand your second patch correctly, then NETDEV_TSTAMP_REALTIME
> can provide a timestamp with low accuracy, which can even jump backwards
> (a later received package can get an earlier timestamp due to the simple
> synchronisation in gve_hwts_realtime_update). Should the expected
> quality of the timestamps be documented here?
Good point. I will mention this in the kdoc in v2.
>
> > +
> >   /*
> >    * This structure defines the management hooks for network devices.
> >    * The following hooks can be defined; unless noted otherwise, they are
> > @@ -1406,11 +1412,10 @@ struct netdev_net_notifier {
> >    *     Get the forwarding path to reach the real device from the HW destination address
> >    * ktime_t (*ndo_get_tstamp)(struct net_device *dev,
> >    *                       const struct skb_shared_hwtstamps *hwtstamps,
> > - *                        bool cycles);
> > - *   Get hardware timestamp based on normal/adjustable time or free running
> > - *   cycle counter. This function is required if physical clock supports a
> > - *   free running cycle counter.
> > - *
> > + *                        enum netdev_tstamp_type type);
> > + *   Get hardware timestamp based on the type requested, or return 0 if the
> > + *   requested type is not supported. This function is required if physical
> > + *   clock supports a free running cycle counter.
>
> ...
>
> Gerhard

Re: [PATCH net-next 1/2] net: extend ndo_get_tstamp for other timestamp types

[Kevin Yang]

Just to clarify your suggestion is that drivers should expose their internal
time/cyclecounter (along with necessary locking)? And then have a
core kernel helper performs the common conversion from the timestamp
to REALTIME(or MONOTONIC)?

If that is the case, looking at gve_get_tstamp in this series, there is
little code to de-duplicate: just a lock and timecounter_cyc2time.
Exposing those internals would likely be more complex.

Also some NICs/drivers might already provide skb_hwtstamps() in REALTIME.
Only the driver knows the device clock's details and whether
conversion is needed.
ndo_get_tstamp can help hide these details.

On Mon, Jan 19, 2026 at 2:57 PM Jakub Kicinski <kuba@kernel.org> wrote:
>
> On Thu, 15 Jan 2026 22:22:59 +0000 Kevin Yang wrote:
> > Reviewed-by: Willem de Bruijn <willemb@google.com>
>
> Other than expediency why implementing this in drivers makes sense?
> There will be very little driver specific logic here.
>
> After thinking about it for 1 minute my intuition would be for drivers
> to just expose necessary information, in-kernel consumers should
> "enable" the sync for all netdevs during init, and then convert
> the timestamps into REALTIME / MONOTONIC by calling some helper?
>
> Also AI code review says:
>
> > +             mult = mult_frac(GVE_HWTS_REAL_CC_NOMINAL,
> > +                              real_ns - priv->ts_real.last_sync_ns,
> > +                              priv->last_sync_nic_counter - prev_nic);
>
> Can this divide by zero?
Will fix this in v2.
> --
> pw-bot: cr