Re: [PATCH net-next v7 0/5] veth: add Byte Queue Limits (BQL) support

From 1a6d2f0a15c23bb4ebb44139193c2649de1142ad Mon Sep 17 00:00:00 2001
From: Simon Schippers <[email protected]>
Date: Thu, 6 Aug 2026 16:17:35 +0200
Subject: [PATCH net-next v8 1/6] net: dql: Add completion coalescing for
 software interfaces

Software interfaces like veth or tun have no hardware completion interrupt
and consume packets one at a time. Calling netdev_tx_completed_queue() per
packet makes DQL converge on a limit of two packets, which is a regression:
BQL sizes the limit from how much completes per call, so a caller that
reports a single object every time forces such a limit. Add an optional
coalescing window so such a queue can batch completions and report them
periodically instead.

The window is set per queue with dql_set_coal_usecs(). While it is
non-zero, dql_completed() accumulates the count in coal_pending and returns
false instead of recalculating the limit, until the window has elapsed or
the batch has grown beyond the limit. That condition is evaluated before
any field of the first cache line is read, so a call that only batches does
not touch that line at all.

A held-back batch cannot stall the queue. dql_queued() stops the queue once
num_queued exceeds adj_limit, so a stopped queue has at least limit + 1
objects in flight. Draining it makes coal_pending exceed the limit and
forces a flush on the call that completes the last one. A batch can
therefore only defer a wake-up that is not needed yet.

Coalescing is active when coal_usecs or coal_pending is set, rather than
being controlled by a separate enable flag, so no enable state can be
toggled out from under a pending batch. The second term keeps an
outstanding batch reachable after the window is set back to 0. Without it
those objects would never be folded into num_completed and the queue would
stall.

Timestamps come from local_clock(), which unlike raw sched_clock() is
bounded across CPUs. They are kept in units of 1024 ns to avoid a division,
so the window is ~2.4% longer than configured, and the u32 wraps every ~73
minutes, which unsigned subtraction handles. A CPU migration can move the
timestamp backwards within local_clock()'s drift bound and flush one batch
early. That is rare and only costs a little batching. The clock is read
only once there is something pending, so a caller that flushes a queue
with an empty batch does not pay for it.

Room for the three new fields comes from moving max_limit and min_limit
into the enqueue cache line, which had eight bytes free, and reading them
up front in dql_completed() next to num_queued and stall_thrs. The
completion path already touched that line once, so the move adds no cache
line touch. Both cache lines are now exactly full on x86_64 and
sizeof(struct dql) is unchanged.

Signed-off-by: Simon Schippers <[email protected]>
---
 include/linux/dynamic_queue_limits.h | 53 ++++++++++++++++++++---
 lib/dynamic_queue_limits.c           | 63 +++++++++++++++++++++++++---
 2 files changed, 104 insertions(+), 12 deletions(-)

diff --git a/include/linux/dynamic_queue_limits.h b/include/linux/dynamic_queue_limits.h
index 808b1a5102e7..cb7338209246 100644
--- a/include/linux/dynamic_queue_limits.h
+++ b/include/linux/dynamic_queue_limits.h
@@ -10,7 +10,8 @@
  *
  *   1) Objects are queued up to some limit specified as number of objects.
  *   2) Periodically a completion process executes which retires consumed
- *      objects.
+ *      objects, or objects are coalesced, which suits software interfaces
+ *      completing one packet at a time.
  *   3) Starvation occurs when limit has been reached, all queued data has
  *      actually been consumed, but completion processing has not yet run
  *      so queuing new data is blocked.
@@ -39,6 +40,7 @@
 #ifdef __KERNEL__
 
 #include <linux/bitops.h>
+#include <linux/types.h>
 #include <asm/bug.h>
 
 #define DQL_HIST_LEN		4
@@ -53,6 +55,9 @@ struct dql {
 	/* Stall threshold (in jiffies), defined by user */
 	unsigned short	stall_thrs;
 
+	unsigned int	max_limit;		/* Max limit */
+	unsigned int	min_limit;		/* Minimum limit */
+
 	unsigned long	history_head;		/* top 58 bits of jiffies */
 	/* stall entries, a bit per entry */
 	unsigned long	history[DQL_HIST_LEN];
@@ -69,21 +74,38 @@ struct dql {
 	unsigned int	lowest_slack;		/* Lowest slack found */
 	unsigned long	slack_start_time;	/* Time slacks seen */
 
-	/* Configuration */
-	unsigned int	max_limit;		/* Max limit */
-	unsigned int	min_limit;		/* Minimum limit */
 	unsigned int	slack_hold_time;	/* Time to measure slack */
 
 	/* Longest stall detected, reported to user */
 	unsigned short	stall_max;
+
+	/* Completion coalescing window in us, 0 disables coalescing */
+	u16		coal_usecs;
+
 	unsigned long	last_reap;		/* Last reap (in jiffies) */
 	unsigned long	stall_cnt;		/* Number of stalls */
+
+	/* Objects completed but not yet accounted for, held back by
+	 * completion coalescing.
+	 */
+	u32		coal_pending;
+	/* Last flush, local_clock() >> DQL_COAL_NS_TO_USECS_SHIFT */
+	u32		coal_last_flush;
 };
 
 /* Set some static maximums */
 #define DQL_MAX_OBJECT (UINT_MAX / 16)
 #define DQL_MAX_LIMIT ((UINT_MAX / 2) - DQL_MAX_OBJECT)
 
+/* local_clock() is in ns.  Shifting by 10 (dividing by 1024) instead of
+ * dividing by NSEC_PER_USEC (1000) is close enough for a coalescing window and
+ * avoids a division, at the cost of a window ~2.4% longer than configured.
+ */
+#define DQL_COAL_NS_TO_USECS_SHIFT	10
+
+/* Maximum coalescing window, bounded by the width of ->coal_usecs */
+#define DQL_COAL_MAX_USECS		U16_MAX
+
 /* Populate the bitmap to be processed later in dql_check_stall() */
 static inline void dql_queue_stall(struct dql *dql)
 {
@@ -149,8 +171,27 @@ static inline int dql_avail(const struct dql *dql)
 	return READ_ONCE(dql->adj_limit) - READ_ONCE(dql->num_queued);
 }
 
-/* Record number of completed objects and recalculate the limit. */
-void dql_completed(struct dql *dql, unsigned int count);
+/* Set the completion coalescing window in us, 0 disables coalescing. */
+static inline void dql_set_coal_usecs(struct dql *dql, unsigned int usecs)
+{
+	if (WARN_ON_ONCE(usecs > DQL_COAL_MAX_USECS))
+		usecs = DQL_COAL_MAX_USECS;
+
+	WRITE_ONCE(dql->coal_usecs, usecs);
+}
+
+/* Return the completion coalescing window in us, 0 if disabled. */
+static inline unsigned int dql_get_coal_usecs(const struct dql *dql)
+{
+	return READ_ONCE(dql->coal_usecs);
+}
+
+/* Record number of completed objects and recalculate the limit.
+ *
+ * Returns true if the completion was applied and the limit recalculated, false
+ * if it was only batched by completion coalescing.
+ */
+bool dql_completed(struct dql *dql, unsigned int count);
 
 /* Reset dql state */
 void dql_reset(struct dql *dql);
diff --git a/lib/dynamic_queue_limits.c b/lib/dynamic_queue_limits.c
index f97a752e900a..af4ae1d11c81 100644
--- a/lib/dynamic_queue_limits.c
+++ b/lib/dynamic_queue_limits.c
@@ -10,6 +10,7 @@
 #include <linux/dynamic_queue_limits.h>
 #include <linux/compiler.h>
 #include <linux/export.h>
+#include <linux/sched/clock.h>
 #include <trace/events/napi.h>
 
 #define POSDIFF(A, B) ((int)((A) - (B)) > 0 ? (A) - (B) : 0)
@@ -80,20 +81,65 @@ static void dql_check_stall(struct dql *dql, unsigned short stall_thrs)
 }
 
 /* Records completed count and recalculates the queue limit */
-void dql_completed(struct dql *dql, unsigned int count)
+bool dql_completed(struct dql *dql, unsigned int count)
 {
 	unsigned int inprogress, prev_inprogress, limit;
 	unsigned int ovlimit, completed, num_queued;
+	unsigned int max_limit, min_limit;
 	unsigned short stall_thrs;
 	bool all_prev_completed;
+	u16 coal_usecs;
+	bool coal;
+
+	/* Coalescing is active while a window is configured, or while a batch
+	 * is still outstanding and needs draining. Read coal_usecs once so that
+	 * the two uses below cannot disagree if it is changed concurrently.
+	 */
+	coal_usecs = READ_ONCE(dql->coal_usecs);
+	coal = coal_usecs || dql->coal_pending;
+
+	/* Without coalescing there is nothing to do unless something completed.
+	 * A coalescing queue may be called with @count == 0 to flush a batch
+	 * whose window has elapsed.
+	 */
+	if (!coal && !count)
+		return false;
+
+	if (coal) {
+		u32 now;
+
+		dql->coal_pending += count;
+		if (!dql->coal_pending)
+			return false;
+
+		/* Read the clock only once there is something to hold back, so
+		 * that a flush call on a queue with nothing pending is free.
+		 */
+		now = local_clock() >> DQL_COAL_NS_TO_USECS_SHIFT;
+
+		/* Hold coal_pending back until the window has elapsed.
+		 * coal_pending above the limit means the queue is starved: it
+		 * cannot make progress until those objects are accounted for,
+		 * so flush regardless of the window.
+		 */
+		if (now - dql->coal_last_flush < coal_usecs &&
+		    dql->coal_pending <= dql->limit)
+			return false;
+
+		count = dql->coal_pending;
+		dql->coal_pending = 0;
+		dql->coal_last_flush = now;
+	}
 
 	num_queued = READ_ONCE(dql->num_queued);
-	/* Read stall_thrs in advance since it belongs to the same (first)
-	 * cache line as ->num_queued. This way, dql_check_stall() does not
-	 * need to touch the first cache line again later, reducing the window
-	 * of possible false sharing.
+	/* Read stall_thrs, max_limit and min_limit in advance since they belong
+	 * to the same (first) cache line as ->num_queued. This way, neither
+	 * dql_check_stall() nor the clamp() below need to touch the first cache
+	 * line again later, reducing the window of possible false sharing.
 	 */
 	stall_thrs = READ_ONCE(dql->stall_thrs);
+	max_limit = READ_ONCE(dql->max_limit);
+	min_limit = READ_ONCE(dql->min_limit);
 
 	/* Can't complete more than what's in queue */
 	BUG_ON(count > num_queued - dql->num_completed);
@@ -170,7 +216,7 @@ void dql_completed(struct dql *dql, unsigned int count)
 	}
 
 	/* Enforce bounds on limit */
-	limit = clamp(limit, dql->min_limit, dql->max_limit);
+	limit = clamp(limit, min_limit, max_limit);
 
 	if (limit != dql->limit) {
 		dql->limit = limit;
@@ -184,6 +230,8 @@ void dql_completed(struct dql *dql, unsigned int count)
 	dql->prev_num_queued = num_queued;
 
 	dql_check_stall(dql, stall_thrs);
+
+	return true;
 }
 EXPORT_SYMBOL(dql_completed);
 
@@ -199,6 +247,8 @@ void dql_reset(struct dql *dql)
 	dql->prev_ovlimit = 0;
 	dql->lowest_slack = UINT_MAX;
 	dql->slack_start_time = jiffies;
+	dql->coal_pending = 0;
+	dql->coal_last_flush = local_clock() >> DQL_COAL_NS_TO_USECS_SHIFT;
 
 	dql->last_reap = jiffies;
 	dql->history_head = jiffies / BITS_PER_LONG;
@@ -212,6 +262,7 @@ void dql_init(struct dql *dql, unsigned int hold_time)
 	dql->min_limit = 0;
 	dql->slack_hold_time = hold_time;
 	dql->stall_thrs = 0;
+	dql->coal_usecs = 0;
 	dql_reset(dql);
 }
 EXPORT_SYMBOL(dql_init);

base-commit: 001b5d347d8ba39b2dccaefcc57967b18caec8fe
-- 
2.43.0

From ad97ecb1ea25a16bd61549e95dce1a86b4efcd7e Mon Sep 17 00:00:00 2001
From: Simon Schippers <[email protected]>
Date: Thu, 6 Aug 2026 16:18:52 +0200
Subject: [PATCH net-next v8 2/6] net: bql: Provide support for dql coalescing
 for software interfaces

Make the DQL coalescing window from the previous commit usable by BQL.

netdev_tx_completed_queue() now returns false early if the completion was
only batched by DQL coalescing, avoiding the memory barrier and the
wake-up. netdev_txq_completed_mb() is adjusted accordingly and issues the
smp_mb() itself if no completion was applied.

netif_set_bql_coalesce_usecs() sets the window on every TX queue of a
device, both at setup and later on a live device, and
netif_get_bql_coalesce_usecs() reads it back from queue 0 for drivers
that want to report it. Both compile away to nothing when CONFIG_BQL is
off, so a driver does not need its own guard. The window can also be read
and written per queue as byte_queue_limits/coal_usecs, where values above
DQL_COAL_MAX_USECS are rejected with -EINVAL.

Signed-off-by: Simon Schippers <[email protected]>
---
 .../ABI/testing/sysfs-class-net-queues        | 15 ++++
 include/linux/netdevice.h                     | 73 +++++++++++++++++--
 include/net/netdev_queues.h                   |  4 +-
 net/core/net-sysfs.c                          | 29 ++++++++
 4 files changed, 111 insertions(+), 10 deletions(-)

diff --git a/Documentation/ABI/testing/sysfs-class-net-queues b/Documentation/ABI/testing/sysfs-class-net-queues
index 84aa25e0d14d..4a0413d353a9 100644
--- a/Documentation/ABI/testing/sysfs-class-net-queues
+++ b/Documentation/ABI/testing/sysfs-class-net-queues
@@ -97,6 +97,21 @@ Description:
 		queued on this network device transmit queue. Default value is
 		0.
 
+What:		/sys/class/net/<iface>/queues/tx-<queue>/byte_queue_limits/coal_usecs
+Date:		August 2026
+KernelVersion:	7.4
+Contact:	[email protected]
+Description:
+		Completion coalescing window for this transmit queue, in
+		microseconds. While it is non-zero, completions reported to BQL
+		are accumulated and the queue limit is only recalculated once
+		the window has elapsed or the accumulated count would starve the
+		queue. This is meant for software interfaces without a hardware
+		completion interrupt, whose completion routine therefore runs
+		once per packet. Writing 0 disables coalescing and restores a
+		limit recalculation on every completion. Default value is 0,
+		except on devices whose driver sets a window of its own.
+
 What:		/sys/class/net/<iface>/queues/tx-<queue>/byte_queue_limits/stall_thrs
 Date:		Jan 2024
 KernelVersion:	6.9
diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index db9dce7f0aa6..36a8d4362d3e 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -3975,28 +3975,41 @@ static inline bool __netdev_sent_queue(struct net_device *dev,
  *
  *	Must be called at most once per TX completion round (and not per
  *	individual packet), so that BQL can adjust its limits appropriately.
+ *	A queue with DQL completion coalescing enabled is exempt from this,
+ *	see netif_set_bql_coalesce_usecs(): coalescing does the batching, so
+ *	such a queue may report per packet.
+ *
+ *	Returns: true if the completion was applied and the limit recalculated,
+ *	false if nothing completed or the completion was only batched by
+ *	coalescing. A true return also means the memory barrier below was
+ *	issued, which netdev_txq_completed_mb() depends on.
  */
-static inline void netdev_tx_completed_queue(struct netdev_queue *dev_queue,
+static inline bool netdev_tx_completed_queue(struct netdev_queue *dev_queue,
 					     unsigned int pkts, unsigned int bytes)
 {
 #ifdef CONFIG_BQL
-	if (unlikely(!bytes))
-		return;
-
-	dql_completed(&dev_queue->dql, bytes);
+	/* There is nothing to wake if nothing completed, or if the completion
+	 * was only batched by coalescing and the limit has not moved.
+	 */
+	if (!dql_completed(&dev_queue->dql, bytes))
+		return false;
 
 	/*
 	 * Without the memory barrier there is a small possibility that
 	 * netdev_tx_sent_queue will miss the update and cause the queue to
 	 * be stopped forever
 	 */
-	smp_mb(); /* NOTE: netdev_txq_completed_mb() assumes this exists */
+	smp_mb();
 
 	if (unlikely(dql_avail(&dev_queue->dql) < 0))
-		return;
+		return true;
 
 	if (test_and_clear_bit(__QUEUE_STATE_STACK_XOFF, &dev_queue->state))
 		netif_schedule_queue(dev_queue);
+
+	return true;
+#else
+	return false;
 #endif
 }
 
@@ -4024,6 +4037,52 @@ static inline void netdev_tx_reset_queue(struct netdev_queue *q)
 #endif
 }
 
+/**
+ * netif_set_bql_coalesce_usecs - set the BQL completion coalescing window
+ * @dev: network device
+ * @usecs: window in microseconds, 0 disables coalescing
+ *
+ * Configure DQL completion coalescing on all TX queues of @dev, so that
+ * netdev_tx_completed_queue() batches completions instead of recalculating the
+ * limit on every call. Intended for software interfaces, which have no hardware
+ * completion interrupt and therefore complete per packet.
+ *
+ * Safe to call on a live device to change the window. @usecs must not exceed
+ * DQL_COAL_MAX_USECS.
+ */
+static inline void netif_set_bql_coalesce_usecs(struct net_device *dev,
+						unsigned int usecs)
+{
+#ifdef CONFIG_BQL
+	unsigned int i;
+
+	for (i = 0; i < dev->num_tx_queues; i++)
+		dql_set_coal_usecs(&netdev_get_tx_queue(dev, i)->dql, usecs);
+#endif
+}
+
+/**
+ * netif_get_bql_coalesce_usecs - get the BQL completion coalescing window
+ * @dev: network device
+ *
+ * The window is per queue and can also be set through
+ * byte_queue_limits/coal_usecs, so this reports TX queue 0 only. Meant for
+ * drivers that set the same window on every queue with
+ * netif_set_bql_coalesce_usecs().
+ *
+ * Returns: the window in microseconds, 0 if coalescing is disabled or BQL is
+ * not built in.
+ */
+static inline unsigned int
+netif_get_bql_coalesce_usecs(const struct net_device *dev)
+{
+#ifdef CONFIG_BQL
+	return dql_get_coal_usecs(&netdev_get_tx_queue(dev, 0)->dql);
+#else
+	return 0;
+#endif
+}
+
 /**
  * netdev_tx_reset_subqueue - reset the BQL stats and state of a netdev queue
  * @dev: network device
diff --git a/include/net/netdev_queues.h b/include/net/netdev_queues.h
index 70c9fe9e83cc..c17d0a49c6db 100644
--- a/include/net/netdev_queues.h
+++ b/include/net/netdev_queues.h
@@ -279,9 +279,7 @@ static inline void
 netdev_txq_completed_mb(struct netdev_queue *dev_queue,
 			unsigned int pkts, unsigned int bytes)
 {
-	if (IS_ENABLED(CONFIG_BQL))
-		netdev_tx_completed_queue(dev_queue, pkts, bytes);
-	else if (bytes)
+	if (!netdev_tx_completed_queue(dev_queue, pkts, bytes) && bytes)
 		smp_mb();
 }
 
diff --git a/net/core/net-sysfs.c b/net/core/net-sysfs.c
index 25546deacec8..286abb88dad8 100644
--- a/net/core/net-sysfs.c
+++ b/net/core/net-sysfs.c
@@ -1667,6 +1667,34 @@ BQL_ATTR(limit, limit);
 BQL_ATTR(limit_max, max_limit);
 BQL_ATTR(limit_min, min_limit);
 
+static ssize_t bql_show_coal_usecs(struct kobject *kobj, struct attribute *attr,
+				   struct netdev_queue *queue, char *buf)
+{
+	return sysfs_emit(buf, "%u\n", dql_get_coal_usecs(&queue->dql));
+}
+
+static ssize_t bql_set_coal_usecs(struct kobject *kobj, struct attribute *attr,
+				  struct netdev_queue *queue, const char *buf,
+				  size_t len)
+{
+	unsigned int value;
+	int err;
+
+	err = kstrtouint(buf, 10, &value);
+	if (err < 0)
+		return err;
+
+	if (value > DQL_COAL_MAX_USECS)
+		return -EINVAL;
+
+	dql_set_coal_usecs(&queue->dql, value);
+
+	return len;
+}
+
+static struct netdev_queue_attribute bql_coal_usecs_attribute __ro_after_init
+	= __ATTR(coal_usecs, 0644, bql_show_coal_usecs, bql_set_coal_usecs);
+
 static struct attribute *dql_attrs[] __ro_after_init = {
 	&bql_limit_attribute.attr,
 	&bql_limit_max_attribute.attr,
@@ -1676,6 +1704,7 @@ static struct attribute *dql_attrs[] __ro_after_init = {
 	&bql_stall_thrs_attribute.attr,
 	&bql_stall_cnt_attribute.attr,
 	&bql_stall_max_attribute.attr,
+	&bql_coal_usecs_attribute.attr,
 	NULL
 };
 
-- 
2.43.0

From be9913fa1bd36bec9bf4f38daf33ab51923656ac Mon Sep 17 00:00:00 2001
From: Jesper Dangaard Brouer <[email protected]>
Date: Thu, 6 Aug 2026 16:22:00 +0200
Subject: [PATCH net-next v8 3/6] net: add dev->bql flag to allow BQL sysfs for
 IFF_NO_QUEUE devices
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Virtual devices with IFF_NO_QUEUE or lltx are excluded from BQL sysfs
by netdev_uses_bql(), since they traditionally lack real hardware
queues. However, some virtual devices like veth implement a real
ptr_ring FIFO with NAPI processing and benefit from BQL to limit
in-flight bytes and reduce latency.

Add a per-device 'bql' bitfield boolean in the priv_flags_slow section
of struct net_device. When set, it overrides the IFF_NO_QUEUE/lltx
exclusion and exposes BQL sysfs entries (/sys/class/net/<dev>/queues/
tx-<n>/byte_queue_limits/). The flag is still gated on CONFIG_BQL.

This allows drivers that use BQL despite being IFF_NO_QUEUE to opt in
to sysfs visibility for monitoring and debugging.

Signed-off-by: Jesper Dangaard Brouer <[email protected]>
Tested-by: Jonas Köppeler <[email protected]>
Signed-off-by: Simon Schippers <[email protected]>
---
 Documentation/networking/net_cachelines/net_device.rst | 1 +
 include/linux/netdevice.h                              | 2 ++
 net/core/net-sysfs.c                                   | 8 +++++++-
 3 files changed, 10 insertions(+), 1 deletion(-)

diff --git a/Documentation/networking/net_cachelines/net_device.rst b/Documentation/networking/net_cachelines/net_device.rst
index 512f6d6fa3d8..52709375ee21 100644
--- a/Documentation/networking/net_cachelines/net_device.rst
+++ b/Documentation/networking/net_cachelines/net_device.rst
@@ -168,6 +168,7 @@ unsigned_long:1                     see_all_hwtstamp_requests
 unsigned_long:1                     change_proto_down
 unsigned_long:1                     netns_immutable
 unsigned_long:1                     fcoe_mtu
+unsigned_long:1                     bql                                                                 netdev_uses_bql(net-sysfs.c)
 struct list_head                    net_notifier_list
 struct macsec_ops*                  macsec_ops
 struct udp_tunnel_nic_info*         udp_tunnel_nic_info
diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 36a8d4362d3e..cdd25832357e 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -2093,6 +2093,7 @@ enum netdev_reg_state {
  *	@change_proto_down: device supports setting carrier via IFLA_PROTO_DOWN
  *	@netns_immutable: interface can't change network namespaces
  *	@fcoe_mtu:	device supports maximum FCoE MTU, 2158 bytes
+ *	@bql:		device uses BQL (DQL sysfs) despite having IFF_NO_QUEUE
  *
  *	@net_notifier_list:	List of per-net netdev notifier block
  *				that follow this device when it is moved
@@ -2513,6 +2514,7 @@ struct net_device {
 	unsigned long		change_proto_down:1;
 	unsigned long		netns_immutable:1;
 	unsigned long		fcoe_mtu:1;
+	unsigned long		bql:1;
 
 	struct list_head	net_notifier_list;
 
diff --git a/net/core/net-sysfs.c b/net/core/net-sysfs.c
index 286abb88dad8..180065ad7ed9 100644
--- a/net/core/net-sysfs.c
+++ b/net/core/net-sysfs.c
@@ -1968,10 +1968,16 @@ static const struct kobj_type netdev_queue_ktype = {
 
 static bool netdev_uses_bql(const struct net_device *dev)
 {
+	if (!IS_ENABLED(CONFIG_BQL))
+		return false;
+
+	if (dev->bql)
+		return true;
+
 	if (dev->lltx || (dev->priv_flags & IFF_NO_QUEUE))
 		return false;
 
-	return IS_ENABLED(CONFIG_BQL);
+	return true;
 }
 
 static int netdev_queue_add_kobject(struct net_device *dev, int index)
-- 
2.43.0

From 64b7a0a2b41aa3f4d0a639dc86819a1470706118 Mon Sep 17 00:00:00 2001
From: Jesper Dangaard Brouer <[email protected]>
Date: Thu, 6 Aug 2026 16:28:28 +0200
Subject: [PATCH net-next v8 4/6] veth: implement Byte Queue Limits (BQL) for
 latency reduction
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Commit dc82a33297fc ("veth: apply qdisc backpressure on full ptr_ring to
reduce TX drops") let a qdisc see when veth's ptr_ring is full, but noted
that the 256-entry ring still sits in front of the qdisc as a dark buffer:
the qdisc cannot shape until the ring overflows. Add BQL so it gets
feedback earlier. With fq_codel under UDP load, ping RTT drops from ~6.61ms
to ~0.36ms (18x).

Charge one fixed unit per packet instead of skb->len. veth has no link
speed -- the ring drains at CPU speed and is packet-indexed. With
byte-based charging, small packets get many more entries into the ring
before the queue stops, deepening the dark buffer again: a concurrent
min-size flood degrades ping RTT by 3.7x with skb->len and not at all with
a fixed unit. As a result byte_queue_limits/limit, limit_max and inflight
report packets for veth. Nothing in the ABI communicates the unit, so it is
stated here.

Charge in veth_xdp_rx() under the ptr_ring producer_lock, once the ring is
known not to be full. The charge must precede the produce, because the peer
NAPI can complete the skb the moment it becomes visible. Holding the lock
across both avoids a pre-charge/undo pattern. veth_xmit() therefore
resolves the txq up front and reuses it on the NETDEV_TX_BUSY path. Charged
skbs are tagged with VETH_BQL_FLAG in the ptr_ring entry, because the qdisc
can be replaced while they are in flight and each skb has to carry the
decision made at enqueue time.

Program order is not enough across CPUs. The smp_wmb() in
__ptr_ring_produce() publishes the charge ahead of the entry, but
dql_completed() reads num_queued, which is not reached through the entry
and so is not dependency-ordered by the consume. veth_xdp_rcv() therefore
pairs an smp_rmb() with it before completing.

Only enable BQL when a real qdisc is attached (!qdisc_txq_has_no_queue),
since dql_queued() needs the serialization of HARD_TX_LOCK, which lltx
devices like veth do not take.

veth has no completion interrupt, so veth_xdp_rcv() completes every skb it
consumes. Against plain BQL that pays a full dql_completed() per packet and
breaks the "at most once per TX completion round" contract, so enable DQL
coalescing instead, default VETH_BQL_COAL_USECS (100 us) and tunable with
ethtool -C tx-usecs for the direction that device transmits in. With
CONFIG_BQL=n that knob reads back 0 and setting it returns -EOPNOTSUPP.
Accumulating per NAPI poll in the driver is not equivalent, because a poll
is not the periodic interval DQL asks for. Each poll therefore starts with
a zero-count completion, so a batch held back at the end of the previous
one is not left waiting on traffic that may never arrive.

BQL adds a second queue-stop mechanism (STACK_XOFF) next to the existing
ring-full one, and both must be clear for the queue to transmit. At
teardown veth_napi_del_range() drains the leftover ring entries after
synchronize_net(), once NAPI is gone and the producer has stopped charging
because it observes rq->napi == NULL, and balances the accounting by
completing the outstanding charges rather than calling
netdev_tx_reset_queue(), whose dql_reset() would race with a concurrent
producer. The peer txq is still woken to clear any DRV_XOFF a late
veth_xmit() may have set.

Signed-off-by: Jesper Dangaard Brouer <[email protected]>
Co-developed-by: Jonas Köppeler <[email protected]>
Signed-off-by: Jonas Köppeler <[email protected]>
Co-developed-by: Simon Schippers <[email protected]>
Signed-off-by: Simon Schippers <[email protected]>
---
 drivers/net/veth.c | 170 +++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 157 insertions(+), 13 deletions(-)

diff --git a/drivers/net/veth.c b/drivers/net/veth.c
index 42e4f246c91d..51ef8474f447 100644
--- a/drivers/net/veth.c
+++ b/drivers/net/veth.c
@@ -34,9 +34,13 @@
 #define DRV_VERSION	"1.0"
 
 #define VETH_XDP_FLAG		BIT(0)
+#define VETH_BQL_FLAG		BIT(1)
 #define VETH_RING_SIZE		256
 #define VETH_XDP_HEADROOM	(XDP_PACKET_HEADROOM + NET_IP_ALIGN)
 
+/* Default DQL completion coalescing window (us), tunable with ethtool -C */
+#define VETH_BQL_COAL_USECS	100
+
 #define VETH_XDP_TX_BULK_SIZE	16
 #define VETH_XDP_BATCH		16
 
@@ -262,7 +266,51 @@ static void veth_get_channels(struct net_device *dev,
 static int veth_set_channels(struct net_device *dev,
 			     struct ethtool_channels *ch);
 
+static int veth_get_coalesce(struct net_device *dev,
+			     struct ethtool_coalesce *ec,
+			     struct kernel_ethtool_coalesce *kernel_coal,
+			     struct netlink_ext_ack *extack)
+{
+	/* Read back from DQL rather than a shadow copy, so that a window set
+	 * through byte_queue_limits/coal_usecs is reported here too.
+	 */
+	ec->tx_coalesce_usecs = netif_get_bql_coalesce_usecs(dev);
+
+	return 0;
+}
+
+static int veth_set_coalesce(struct net_device *dev,
+			     struct ethtool_coalesce *ec,
+			     struct kernel_ethtool_coalesce *kernel_coal,
+			     struct netlink_ext_ack *extack)
+{
+	struct veth_priv *priv = netdev_priv(dev);
+	struct net_device *peer;
+
+	if (!IS_ENABLED(CONFIG_BQL)) {
+		NL_SET_ERR_MSG_MOD(extack, "BQL is not enabled in this kernel");
+		return -EOPNOTSUPP;
+	}
+
+	if (ec->tx_coalesce_usecs > DQL_COAL_MAX_USECS) {
+		NL_SET_ERR_MSG_MOD(extack, "tx-usecs too large");
+		return -EINVAL;
+	}
+
+	netif_set_bql_coalesce_usecs(dev, ec->tx_coalesce_usecs);
+
+	/* Mirror onto the peer to keep the pair symmetric: both directions
+	 * coalesce with the same tx-usecs. Called under RTNL.
+	 */
+	peer = rtnl_dereference(priv->peer);
+	if (peer)
+		netif_set_bql_coalesce_usecs(peer, ec->tx_coalesce_usecs);
+
+	return 0;
+}
+
 static const struct ethtool_ops veth_ethtool_ops = {
+	.supported_coalesce_params = ETHTOOL_COALESCE_TX_USECS,
 	.get_drvinfo		= veth_get_drvinfo,
 	.get_link		= ethtool_op_get_link,
 	.get_strings		= veth_get_strings,
@@ -272,6 +320,8 @@ static const struct ethtool_ops veth_ethtool_ops = {
 	.get_ts_info		= ethtool_op_get_ts_info,
 	.get_channels		= veth_get_channels,
 	.set_channels		= veth_set_channels,
+	.get_coalesce		= veth_get_coalesce,
+	.set_coalesce		= veth_set_coalesce,
 };
 
 /* general routines */
@@ -281,6 +331,21 @@ static bool veth_is_xdp_frame(void *ptr)
 	return (unsigned long)ptr & VETH_XDP_FLAG;
 }
 
+static bool veth_ptr_is_bql(void *ptr)
+{
+	return (unsigned long)ptr & VETH_BQL_FLAG;
+}
+
+static struct sk_buff *veth_ptr_to_skb(void *ptr)
+{
+	return (void *)((unsigned long)ptr & ~VETH_BQL_FLAG);
+}
+
+static void *veth_skb_to_ptr(struct sk_buff *skb, bool bql)
+{
+	return bql ? (void *)((unsigned long)skb | VETH_BQL_FLAG) : skb;
+}
+
 static struct xdp_frame *veth_ptr_to_xdp(void *ptr)
 {
 	return (void *)((unsigned long)ptr & ~VETH_XDP_FLAG);
@@ -296,7 +361,21 @@ static void veth_ptr_free(void *ptr)
 	if (veth_is_xdp_frame(ptr))
 		xdp_return_frame(veth_ptr_to_xdp(ptr));
 	else
-		kfree_skb(ptr);
+		kfree_skb(veth_ptr_to_skb(ptr));
+}
+
+static unsigned int veth_ptr_ring_drain(struct ptr_ring *ring)
+{
+	unsigned int n_bql = 0;
+	void *ptr;
+
+	while ((ptr = ptr_ring_consume(ring))) {
+		if (veth_ptr_is_bql(ptr))
+			n_bql++;
+		veth_ptr_free(ptr);
+	}
+
+	return n_bql;
 }
 
 static void __veth_xdp_flush(struct veth_rq *rq)
@@ -310,19 +389,36 @@ static void __veth_xdp_flush(struct veth_rq *rq)
 	}
 }
 
-static int veth_xdp_rx(struct veth_rq *rq, struct sk_buff *skb)
+static int veth_xdp_rx(struct veth_rq *rq, struct sk_buff *skb, bool do_bql,
+		       struct netdev_queue *txq)
 {
-	if (unlikely(ptr_ring_produce(&rq->xdp_ring, skb)))
+	struct ptr_ring *ring = &rq->xdp_ring;
+
+	spin_lock(&ring->producer_lock);
+	if (unlikely(__ptr_ring_check_produce(ring))) {
+		spin_unlock(&ring->producer_lock);
 		return NETDEV_TX_BUSY; /* signal qdisc layer */
+	}
+
+	if (do_bql)
+		netdev_tx_sent_queue(txq, 1); /* one unit per packet */
+
+	/* Its smp_wmb() orders the BQL charge above ahead of the entry, so
+	 * the peer NAPI cannot see the skb without its charge. Pairs with
+	 * the smp_rmb() in veth_xdp_rcv().
+	 */
+	__ptr_ring_produce(ring, veth_skb_to_ptr(skb, do_bql));
+	spin_unlock(&ring->producer_lock);
 
 	return NET_RX_SUCCESS; /* same as NETDEV_TX_OK */
 }
 
 static int veth_forward_skb(struct net_device *dev, struct sk_buff *skb,
-			    struct veth_rq *rq, bool xdp)
+			    struct veth_rq *rq, bool xdp, bool do_bql,
+			    struct netdev_queue *txq)
 {
 	return __dev_forward_skb(dev, skb) ?: xdp ?
-		veth_xdp_rx(rq, skb) :
+		veth_xdp_rx(rq, skb, do_bql, txq) :
 		__netif_rx(skb);
 }
 
@@ -349,10 +445,11 @@ static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct veth_priv *rcv_priv, *priv = netdev_priv(dev);
 	struct veth_rq *rq = NULL;
-	struct netdev_queue *txq;
+	struct netdev_queue *txq = NULL;
 	struct net_device *rcv;
 	int length = skb->len;
 	bool use_napi = false;
+	bool do_bql = false;
 	int ret, rxq;
 
 	rcu_read_lock();
@@ -377,7 +474,12 @@ static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
 
 	skb_tx_timestamp(skb);
 
-	ret = veth_forward_skb(rcv, skb, rq, use_napi);
+	if (rxq < dev->real_num_tx_queues) {
+		txq = netdev_get_tx_queue(dev, rxq);
+		do_bql = use_napi && !qdisc_txq_has_no_queue(txq);
+	}
+
+	ret = veth_forward_skb(rcv, skb, rq, use_napi, do_bql, txq);
 	switch (ret) {
 	case NET_RX_SUCCESS: /* same as NETDEV_TX_OK */
 		if (!use_napi)
@@ -389,9 +491,7 @@ static netdev_tx_t veth_xmit(struct sk_buff *skb, struct net_device *dev)
 		/* If a qdisc is attached to our virtual device, returning
 		 * NETDEV_TX_BUSY is allowed.
 		 */
-		txq = netdev_get_tx_queue(dev, rxq);
-
-		if (qdisc_txq_has_no_queue(txq)) {
+		if (!txq || qdisc_txq_has_no_queue(txq)) {
 			dev_kfree_skb_any(skb);
 			goto drop;
 		}
@@ -901,11 +1001,18 @@ static struct sk_buff *veth_xdp_rcv_skb(struct veth_rq *rq,
 
 static int veth_xdp_rcv(struct veth_rq *rq, int budget,
 			struct veth_xdp_tx_bq *bq,
-			struct veth_stats *stats)
+			struct veth_stats *stats,
+			struct netdev_queue *peer_txq)
 {
 	int i, done = 0, n_xdpf = 0;
 	void *xdpf[VETH_XDP_BATCH];
 
+	/* Flush completions batched during an earlier poll whose coalescing
+	 * window has elapsed in the meantime.
+	 */
+	if (peer_txq)
+		netdev_tx_completed_queue(peer_txq, 0, 0);
+
 	for (i = 0; i < budget; i++) {
 		void *ptr = __ptr_ring_consume(&rq->xdp_ring);
 
@@ -929,9 +1036,21 @@ static int veth_xdp_rcv(struct veth_rq *rq, int budget,
 			}
 		} else {
 			/* ndo_start_xmit */
-			struct sk_buff *skb = ptr;
+			bool bql_charged = veth_ptr_is_bql(ptr);
+			struct sk_buff *skb = veth_ptr_to_skb(ptr);
 
 			stats->xdp_bytes += skb->len;
+			if (peer_txq && bql_charged) {
+				/* Pairs with the smp_wmb() in
+				 * __ptr_ring_produce(). The charge is not
+				 * reached through ptr, so nothing else
+				 * orders dql_completed()'s read of
+				 * num_queued against the consume above.
+				 */
+				smp_rmb();
+				netdev_tx_completed_queue(peer_txq, 1, 1);
+			}
+
 			skb = veth_xdp_rcv_skb(rq, skb, bq, stats);
 			if (skb) {
 				if (skb_shared(skb) || skb_unclone(skb, GFP_ATOMIC))
@@ -977,7 +1096,7 @@ static int veth_poll(struct napi_struct *napi, int budget)
 		   netdev_get_tx_queue(peer_dev, queue_idx) : NULL;
 
 	xdp_set_return_frame_no_direct();
-	done = veth_xdp_rcv(rq, budget, &bq, &stats);
+	done = veth_xdp_rcv(rq, budget, &bq, &stats, peer_txq);
 
 	if (stats.xdp_redirect > 0)
 		xdp_do_flush();
@@ -1075,6 +1194,7 @@ static int __veth_napi_enable(struct net_device *dev)
 static void veth_napi_del_range(struct net_device *dev, int start, int end)
 {
 	struct veth_priv *priv = netdev_priv(dev);
+	struct net_device *peer;
 	int i;
 
 	for (i = start; i < end; i++) {
@@ -1086,11 +1206,31 @@ static void veth_napi_del_range(struct net_device *dev, int start, int end)
 	}
 	synchronize_net();
 
+	peer = rtnl_dereference(priv->peer);
+
 	for (i = start; i < end; i++) {
 		struct veth_rq *rq = &priv->rq[i];
+		struct netdev_queue *txq;
+		unsigned int n_bql;
 
 		rq->rx_notify_masked = false;
+
+		/* Drain leftover ring frames, counting BQL-charged SKBs that
+		 * were charged via netdev_tx_sent_queue() but never consumed.
+		 */
+		n_bql = veth_ptr_ring_drain(&rq->xdp_ring);
 		ptr_ring_cleanup(&rq->xdp_ring, veth_ptr_free);
+
+		if (!peer || i >= peer->num_tx_queues)
+			continue;
+
+		txq = netdev_get_tx_queue(peer, i);
+
+		if (n_bql)
+			netdev_tx_completed_queue(txq, n_bql, n_bql);
+
+		if (netif_running(peer))
+			netif_tx_wake_queue(txq);
 	}
 
 	for (i = start; i < end; i++) {
@@ -1744,6 +1884,7 @@ static void veth_setup(struct net_device *dev)
 	dev->priv_flags |= IFF_PHONY_HEADROOM;
 	dev->priv_flags |= IFF_DISABLE_NETPOLL;
 	dev->lltx = true;
+	dev->bql = true;
 
 	dev->netdev_ops = &veth_netdev_ops;
 	dev->xdp_metadata_ops = &veth_xdp_metadata_ops;
@@ -1919,6 +2060,9 @@ static int veth_newlink(struct net_device *dev,
 	veth_set_xdp_features(dev);
 	veth_set_xdp_features(peer);
 
+	netif_set_bql_coalesce_usecs(dev, VETH_BQL_COAL_USECS);
+	netif_set_bql_coalesce_usecs(peer, VETH_BQL_COAL_USECS);
+
 	return 0;
 
 err_peer_queues:
-- 
2.43.0

From ab29d87b9b357e34b935376e1fe81786672e7310 Mon Sep 17 00:00:00 2001
From: Jesper Dangaard Brouer <[email protected]>
Date: Thu, 6 Aug 2026 16:28:28 +0200
Subject: [PATCH net-next v8 5/6] net: sched: add timeout count to NETDEV
 WATCHDOG message
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add the per-queue timeout counter (trans_timeout) to the core NETDEV
WATCHDOG log message.  This makes it easy to determine how frequently
a particular queue is stalling from a single log line, without having
to search through and correlate spaced-out log entries.

Useful for production monitoring where timeouts are spaced by the
watchdog interval, making frequency hard to judge.

Suggested-by: Jakub Kicinski <[email protected]>
Signed-off-by: Jesper Dangaard Brouer <[email protected]>
Tested-by: Jonas Köppeler <[email protected]>
Signed-off-by: Simon Schippers <[email protected]>
---
 net/sched/sch_generic.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c
index ef2b4bf51564..4cbefeab726a 100644
--- a/net/sched/sch_generic.c
+++ b/net/sched/sch_generic.c
@@ -533,6 +533,7 @@ static void dev_watchdog(struct timer_list *t)
 		    netif_running(dev) &&
 		    netif_carrier_ok(dev)) {
 			unsigned int timedout_ms = 0;
+			unsigned long trans_timeout = 0;
 			unsigned int i;
 			unsigned long trans_start;
 			unsigned long oldest_start = jiffies;
@@ -553,6 +554,7 @@ static void dev_watchdog(struct timer_list *t)
 				if (time_after(jiffies, trans_start + dev->watchdog_timeo)) {
 					timedout_ms = jiffies_to_msecs(jiffies - trans_start);
 					atomic_long_inc(&txq->trans_timeout);
+					trans_timeout = atomic_long_read(&txq->trans_timeout);
 					break;
 				}
 				if (time_after(oldest_start, trans_start))
@@ -561,9 +563,9 @@ static void dev_watchdog(struct timer_list *t)
 
 			if (unlikely(timedout_ms)) {
 				trace_net_dev_xmit_timeout(dev, i);
-				netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms\n",
+				netdev_crit(dev, "NETDEV WATCHDOG: CPU: %d: transmit queue %u timed out %u ms (n:%ld)\n",
 					    raw_smp_processor_id(),
-					    i, timedout_ms);
+					    i, timedout_ms, trans_timeout);
 				netif_freeze_queues(dev);
 				dev->netdev_ops->ndo_tx_timeout(dev, i);
 				netif_unfreeze_queues(dev);
-- 
2.43.0

From f75b304361388b050425d5b3796e5583dccb0d96 Mon Sep 17 00:00:00 2001
From: Jesper Dangaard Brouer <[email protected]>
Date: Thu, 6 Aug 2026 16:34:14 +0200
Subject: [PATCH net-next v8 6/6] veth: add tx_timeout watchdog as BQL safety
 net
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

With the introduction of BQL (Byte Queue Limits) for veth, there are
now two independent mechanisms that can stop a transmit queue:

 - DRV_XOFF: set by netif_tx_stop_queue() when the ptr_ring is full
 - STACK_XOFF: set by BQL when the byte-in-flight limit is reached

If either mechanism stalls without a corresponding wake/completion,
the queue stops permanently. Enable the net device watchdog timer and
implement ndo_tx_timeout as a failsafe recovery.

The timeout handler resets BQL state (clearing STACK_XOFF) and wakes
the queue (clearing DRV_XOFF), covering both stop mechanisms. The
watchdog fires after 16 seconds, which accommodates worst-case NAPI
processing (budget=64 packets x 250ms per-packet consumer delay)
without false positives under normal backpressure.

Signed-off-by: Jesper Dangaard Brouer <[email protected]>
Tested-by: Jonas Köppeler <[email protected]>
Signed-off-by: Simon Schippers <[email protected]>
---
 drivers/net/veth.c | 25 +++++++++++++++++++++++++
 1 file changed, 25 insertions(+)

diff --git a/drivers/net/veth.c b/drivers/net/veth.c
index 51ef8474f447..2310d994da51 100644
--- a/drivers/net/veth.c
+++ b/drivers/net/veth.c
@@ -44,6 +44,13 @@
 #define VETH_XDP_TX_BULK_SIZE	16
 #define VETH_XDP_BATCH		16
 
+/* tx_timeout watchdog timeout. DRV_XOFF is only cleared at the end of a NAPI
+ * veth_poll() (netif_tx_wake_queue()), so the timeout must outlast a full
+ * worst-case poll: a 64-packet budget with a pessimistic 250 ms/pkt consumer
+ * delay => 64 * 250 ms = 16 s.
+ */
+#define VETH_WATCHDOG_TIMEOUT_MS	(64 * 250)
+
 struct veth_stats {
 	u64	rx_drops;
 	/* xdp */
@@ -1522,6 +1529,22 @@ static int veth_set_channels(struct net_device *dev,
 	goto out;
 }
 
+static void veth_tx_timeout(struct net_device *dev, unsigned int txqueue)
+{
+	struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue);
+
+	netdev_err(dev,
+		   "veth backpressure(0x%lX) stalled(n:%ld) TXQ(%u) re-enable\n",
+		   txq->state, atomic_long_read(&txq->trans_timeout), txqueue);
+
+	/* Cannot call netdev_tx_reset_queue(): dql_reset() races with
+	 * peer NAPI calling dql_completed() concurrently.
+	 * Just clear the stop bits; the qdisc will re-stop if still stuck.
+	 */
+	clear_bit(__QUEUE_STATE_STACK_XOFF, &txq->state);
+	netif_tx_wake_queue(txq);
+}
+
 static int veth_open(struct net_device *dev)
 {
 	struct veth_priv *priv = netdev_priv(dev);
@@ -1860,6 +1883,7 @@ static const struct net_device_ops veth_netdev_ops = {
 	.ndo_bpf		= veth_xdp,
 	.ndo_xdp_xmit		= veth_ndo_xdp_xmit,
 	.ndo_get_peer_dev	= veth_peer_dev,
+	.ndo_tx_timeout		= veth_tx_timeout,
 };
 
 static const struct xdp_metadata_ops veth_xdp_metadata_ops = {
@@ -1899,6 +1923,7 @@ static void veth_setup(struct net_device *dev)
 	dev->priv_destructor = veth_dev_free;
 	dev->pcpu_stat_type = NETDEV_PCPU_STAT_TSTATS;
 	dev->max_mtu = ETH_MAX_MTU;
+	dev->watchdog_timeo = msecs_to_jiffies(VETH_WATCHDOG_TIMEOUT_MS);
 
 	dev->hw_features = VETH_FEATURES;
 	dev->hw_enc_features = VETH_FEATURES;
-- 
2.43.0