Merge branch 'mlxsw-mq-red-offload'

Jiri Pirko says:

====================
mlxsw: Offload multi-queue RED support

Nogah says:

Support a two level hierarchy of offloaded qdiscs in mlxsw, with sch_prio
being the root qdisc and sch_red as the children.

                +----------+
                | sch_prio |
                +----+-----+
                     |
                     |
    +----------------------------------+
    |                |                 |
    |                |                 |
    |                |                 |
+---v---+       +----v---+       +-----v--+
|sch_red|       |sch_red |       |sch_red |
+-------+       +--------+       +--------+

When setting sch_prio as the root qdisc on a physical port, mlxsw will
offload it. When adding it with sch_red as a child qdisc, it will offload
it as well.
Relocating child qdisc or connecting them to more then one child will
result in unoffloading them. Relocating child qdisc more then once is
highly unrecommended and might cause a miss match between the kernel
configuration and the offloaded one. The offloaded configuration will be
aligned with the one shown in the show command.
Changing the priomap parameter of sch_prio might cause a band that its
configuration was changed and it has offloaded sch_red set on it, to lose
some stats data as if sch_red was unoffloaded and offloaded again. However,
it won't affect the data on this band that will have sch_red continuously.

Patch 1 adds support for setting RED as the child of root qdisc.
Patches 2-4 add support for RED bstasts for offloaded child qdiscs.
Patches 5-6 handle backlog related changes for offloaded child qdiscs.
Patches 7-8 update PRIO in mlxsw to be able to have RED as child on its
bands.
Patch 9 adds offload handles for PRIO graft operations. In mlxsw it will
cause the driver to stop offloading the child in question.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/ethernet/mellanox/mlxsw/spectrum.c

@@ -1040,6 +1040,16 @@ mlxsw_sp_port_get_hw_xstats(struct net_device *dev,
 		xstats->tail_drop[i] =
 			mlxsw_reg_ppcnt_tc_no_buffer_discard_uc_get(ppcnt_pl);
 	}
+
+	for (i = 0; i < IEEE_8021QAZ_MAX_TCS; i++) {
+		err = mlxsw_sp_port_get_stats_raw(dev, MLXSW_REG_PPCNT_PRIO_CNT,
+						  i, ppcnt_pl);
+		if (err)
+			continue;
+
+		xstats->tx_packets[i] = mlxsw_reg_ppcnt_tx_frames_get(ppcnt_pl);
+		xstats->tx_bytes[i] = mlxsw_reg_ppcnt_tx_octets_get(ppcnt_pl);
+	}
 }
 
 static void update_stats_cache(struct work_struct *work)

drivers/net/ethernet/mellanox/mlxsw/spectrum.h

@@ -210,6 +210,8 @@ struct mlxsw_sp_port_xstats {
 	u64 wred_drop[TC_MAX_QUEUE];
 	u64 tail_drop[TC_MAX_QUEUE];
 	u64 backlog[TC_MAX_QUEUE];
+	u64 tx_bytes[IEEE_8021QAZ_MAX_TCS];
+	u64 tx_packets[IEEE_8021QAZ_MAX_TCS];
 };
 
 struct mlxsw_sp_port {
@@ -247,6 +249,7 @@ struct mlxsw_sp_port {
 	struct mlxsw_sp_port_sample *sample;
 	struct list_head vlans_list;
 	struct mlxsw_sp_qdisc *root_qdisc;
+	struct mlxsw_sp_qdisc *tclass_qdiscs;
 	unsigned acl_rule_count;
 	struct mlxsw_sp_acl_block *ing_acl_block;
 	struct mlxsw_sp_acl_block *eg_acl_block;

drivers/net/ethernet/mellanox/mlxsw/spectrum_qdisc.c

@@ -42,6 +42,8 @@
 #include "reg.h"
 
 #define MLXSW_SP_PRIO_BAND_TO_TCLASS(band) (IEEE_8021QAZ_MAX_TCS - band - 1)
+#define MLXSW_SP_PRIO_CHILD_TO_TCLASS(child) \
+	MLXSW_SP_PRIO_BAND_TO_TCLASS((child - 1))
 
 enum mlxsw_sp_qdisc_type {
 	MLXSW_SP_QDISC_NO_QDISC,
@@ -76,6 +78,7 @@ struct mlxsw_sp_qdisc_ops {
 struct mlxsw_sp_qdisc {
 	u32 handle;
 	u8 tclass_num;
+	u8 prio_bitmap;
 	union {
 		struct red_stats red;
 	} xstats_base;
@@ -99,6 +102,44 @@ mlxsw_sp_qdisc_compare(struct mlxsw_sp_qdisc *mlxsw_sp_qdisc, u32 handle,
 	       mlxsw_sp_qdisc->handle == handle;
 }
 
+static struct mlxsw_sp_qdisc *
+mlxsw_sp_qdisc_find(struct mlxsw_sp_port *mlxsw_sp_port, u32 parent,
+		    bool root_only)
+{
+	int tclass, child_index;
+
+	if (parent == TC_H_ROOT)
+		return mlxsw_sp_port->root_qdisc;
+
+	if (root_only || !mlxsw_sp_port->root_qdisc ||
+	    !mlxsw_sp_port->root_qdisc->ops ||
+	    TC_H_MAJ(parent) != mlxsw_sp_port->root_qdisc->handle ||
+	    TC_H_MIN(parent) > IEEE_8021QAZ_MAX_TCS)
+		return NULL;
+
+	child_index = TC_H_MIN(parent);
+	tclass = MLXSW_SP_PRIO_CHILD_TO_TCLASS(child_index);
+	return &mlxsw_sp_port->tclass_qdiscs[tclass];
+}
+
+static struct mlxsw_sp_qdisc *
+mlxsw_sp_qdisc_find_by_handle(struct mlxsw_sp_port *mlxsw_sp_port, u32 handle)
+{
+	int i;
+
+	if (mlxsw_sp_port->root_qdisc->handle == handle)
+		return mlxsw_sp_port->root_qdisc;
+
+	if (mlxsw_sp_port->root_qdisc->handle == TC_H_UNSPEC)
+		return NULL;
+
+	for (i = 0; i < IEEE_8021QAZ_MAX_TCS; i++)
+		if (mlxsw_sp_port->tclass_qdiscs[i].handle == handle)
+			return &mlxsw_sp_port->tclass_qdiscs[i];
+
+	return NULL;
+}
+
 static int
 mlxsw_sp_qdisc_destroy(struct mlxsw_sp_port *mlxsw_sp_port,
 		       struct mlxsw_sp_qdisc *mlxsw_sp_qdisc)
@@ -185,6 +226,23 @@ mlxsw_sp_qdisc_get_xstats(struct mlxsw_sp_port *mlxsw_sp_port,
 	return -EOPNOTSUPP;
 }
 
+static void
+mlxsw_sp_qdisc_bstats_per_priority_get(struct mlxsw_sp_port_xstats *xstats,
+				       u8 prio_bitmap, u64 *tx_packets,
+				       u64 *tx_bytes)
+{
+	int i;
+
+	*tx_packets = 0;
+	*tx_bytes = 0;
+	for (i = 0; i < IEEE_8021QAZ_MAX_TCS; i++) {
+		if (prio_bitmap & BIT(i)) {
+			*tx_packets += xstats->tx_packets[i];
+			*tx_bytes += xstats->tx_bytes[i];
+		}
+	}
+}
+
 static int
 mlxsw_sp_tclass_congestion_enable(struct mlxsw_sp_port *mlxsw_sp_port,
 				  int tclass_num, u32 min, u32 max,
@@ -230,17 +288,16 @@ mlxsw_sp_setup_tc_qdisc_red_clean_stats(struct mlxsw_sp_port *mlxsw_sp_port,
 	u8 tclass_num = mlxsw_sp_qdisc->tclass_num;
 	struct mlxsw_sp_qdisc_stats *stats_base;
 	struct mlxsw_sp_port_xstats *xstats;
-	struct rtnl_link_stats64 *stats;
 	struct red_stats *red_base;
 
 	xstats = &mlxsw_sp_port->periodic_hw_stats.xstats;
-	stats = &mlxsw_sp_port->periodic_hw_stats.stats;
 	stats_base = &mlxsw_sp_qdisc->stats_base;
 	red_base = &mlxsw_sp_qdisc->xstats_base.red;
 
-	stats_base->tx_packets = stats->tx_packets;
-	stats_base->tx_bytes = stats->tx_bytes;
-
+	mlxsw_sp_qdisc_bstats_per_priority_get(xstats,
+					       mlxsw_sp_qdisc->prio_bitmap,
+					       &stats_base->tx_packets,
+					       &stats_base->tx_bytes);
 	red_base->prob_mark = xstats->ecn;
 	red_base->prob_drop = xstats->wred_drop[tclass_num];
 	red_base->pdrop = xstats->tail_drop[tclass_num];
@@ -255,6 +312,12 @@ static int
 mlxsw_sp_qdisc_red_destroy(struct mlxsw_sp_port *mlxsw_sp_port,
 			   struct mlxsw_sp_qdisc *mlxsw_sp_qdisc)
 {
+	struct mlxsw_sp_qdisc *root_qdisc = mlxsw_sp_port->root_qdisc;
+
+	if (root_qdisc != mlxsw_sp_qdisc)
+		root_qdisc->stats_base.backlog -=
+					mlxsw_sp_qdisc->stats_base.backlog;
+
 	return mlxsw_sp_tclass_congestion_disable(mlxsw_sp_port,
 						  mlxsw_sp_qdisc->tclass_num);
 }
@@ -319,6 +382,7 @@ mlxsw_sp_qdisc_red_unoffload(struct mlxsw_sp_port *mlxsw_sp_port,
 	backlog = mlxsw_sp_cells_bytes(mlxsw_sp_port->mlxsw_sp,
 				       mlxsw_sp_qdisc->stats_base.backlog);
 	p->qstats->backlog -= backlog;
+	mlxsw_sp_qdisc->stats_base.backlog = 0;
 }
 
 static int
@@ -357,14 +421,16 @@ mlxsw_sp_qdisc_get_red_stats(struct mlxsw_sp_port *mlxsw_sp_port,
 	u8 tclass_num = mlxsw_sp_qdisc->tclass_num;
 	struct mlxsw_sp_qdisc_stats *stats_base;
 	struct mlxsw_sp_port_xstats *xstats;
-	struct rtnl_link_stats64 *stats;
 
 	xstats = &mlxsw_sp_port->periodic_hw_stats.xstats;
-	stats = &mlxsw_sp_port->periodic_hw_stats.stats;
 	stats_base = &mlxsw_sp_qdisc->stats_base;
 
-	tx_bytes = stats->tx_bytes - stats_base->tx_bytes;
-	tx_packets = stats->tx_packets - stats_base->tx_packets;
+	mlxsw_sp_qdisc_bstats_per_priority_get(xstats,
+					       mlxsw_sp_qdisc->prio_bitmap,
+					       &tx_packets, &tx_bytes);
+	tx_bytes = tx_bytes - stats_base->tx_bytes;
+	tx_packets = tx_packets - stats_base->tx_packets;
+
 	overlimits = xstats->wred_drop[tclass_num] + xstats->ecn -
 		     stats_base->overlimits;
 	drops = xstats->wred_drop[tclass_num] + xstats->tail_drop[tclass_num] -
@@ -406,11 +472,10 @@ int mlxsw_sp_setup_tc_red(struct mlxsw_sp_port *mlxsw_sp_port,
 {
 	struct mlxsw_sp_qdisc *mlxsw_sp_qdisc;
 
-	if (p->parent != TC_H_ROOT)
+	mlxsw_sp_qdisc = mlxsw_sp_qdisc_find(mlxsw_sp_port, p->parent, false);
+	if (!mlxsw_sp_qdisc)
 		return -EOPNOTSUPP;
 
-	mlxsw_sp_qdisc = mlxsw_sp_port->root_qdisc;
-
 	if (p->command == TC_RED_REPLACE)
 		return mlxsw_sp_qdisc_replace(mlxsw_sp_port, p->handle,
 					      mlxsw_sp_qdisc,
@@ -441,9 +506,13 @@ mlxsw_sp_qdisc_prio_destroy(struct mlxsw_sp_port *mlxsw_sp_port,
 {
 	int i;
 
-	for (i = 0; i < IEEE_8021QAZ_MAX_TCS; i++)
+	for (i = 0; i < IEEE_8021QAZ_MAX_TCS; i++) {
 		mlxsw_sp_port_prio_tc_set(mlxsw_sp_port, i,
 					  MLXSW_SP_PORT_DEFAULT_TCLASS);
+		mlxsw_sp_qdisc_destroy(mlxsw_sp_port,
+				       &mlxsw_sp_port->tclass_qdiscs[i]);
+		mlxsw_sp_port->tclass_qdiscs[i].prio_bitmap = 0;
+	}
 
 	return 0;
 }
@@ -467,16 +536,41 @@ mlxsw_sp_qdisc_prio_replace(struct mlxsw_sp_port *mlxsw_sp_port,
 			    void *params)
 {
 	struct tc_prio_qopt_offload_params *p = params;
-	int tclass, i;
+	struct mlxsw_sp_qdisc *child_qdisc;
+	int tclass, i, band, backlog;
+	u8 old_priomap;
 	int err;
 
+	for (band = 0; band < p->bands; band++) {
+		tclass = MLXSW_SP_PRIO_BAND_TO_TCLASS(band);
+		child_qdisc = &mlxsw_sp_port->tclass_qdiscs[tclass];
+		old_priomap = child_qdisc->prio_bitmap;
+		child_qdisc->prio_bitmap = 0;
 		for (i = 0; i < IEEE_8021QAZ_MAX_TCS; i++) {
-		tclass = MLXSW_SP_PRIO_BAND_TO_TCLASS(p->priomap[i]);
-		err = mlxsw_sp_port_prio_tc_set(mlxsw_sp_port, i, tclass);
+			if (p->priomap[i] == band) {
+				child_qdisc->prio_bitmap |= BIT(i);
+				if (BIT(i) & old_priomap)
+					continue;
+				err = mlxsw_sp_port_prio_tc_set(mlxsw_sp_port,
+								i, tclass);
 				if (err)
 					return err;
 			}
-
+		}
+		if (old_priomap != child_qdisc->prio_bitmap &&
+		    child_qdisc->ops && child_qdisc->ops->clean_stats) {
+			backlog = child_qdisc->stats_base.backlog;
+			child_qdisc->ops->clean_stats(mlxsw_sp_port,
+						      child_qdisc);
+			child_qdisc->stats_base.backlog = backlog;
+		}
+	}
+	for (; band < IEEE_8021QAZ_MAX_TCS; band++) {
+		tclass = MLXSW_SP_PRIO_BAND_TO_TCLASS(band);
+		child_qdisc = &mlxsw_sp_port->tclass_qdiscs[tclass];
+		child_qdisc->prio_bitmap = 0;
+		mlxsw_sp_qdisc_destroy(mlxsw_sp_port, child_qdisc);
+	}
 	return 0;
 }
 
@@ -513,6 +607,7 @@ mlxsw_sp_qdisc_get_prio_stats(struct mlxsw_sp_port *mlxsw_sp_port,
 
 	for (i = 0; i < IEEE_8021QAZ_MAX_TCS; i++) {
 		drops += xstats->tail_drop[i];
+		drops += xstats->wred_drop[i];
 		backlog += xstats->backlog[i];
 	}
 	drops = drops - stats_base->drops;
@@ -548,8 +643,10 @@ mlxsw_sp_setup_tc_qdisc_prio_clean_stats(struct mlxsw_sp_port *mlxsw_sp_port,
 	stats_base->tx_bytes = stats->tx_bytes;
 
 	stats_base->drops = 0;
-	for (i = 0; i < IEEE_8021QAZ_MAX_TCS; i++)
+	for (i = 0; i < IEEE_8021QAZ_MAX_TCS; i++) {
 		stats_base->drops += xstats->tail_drop[i];
+		stats_base->drops += xstats->wred_drop[i];
+	}
 
 	mlxsw_sp_qdisc->stats_base.backlog = 0;
 }
@@ -564,15 +661,48 @@ static struct mlxsw_sp_qdisc_ops mlxsw_sp_qdisc_ops_prio = {
 	.clean_stats = mlxsw_sp_setup_tc_qdisc_prio_clean_stats,
 };
 
+/* Grafting is not supported in mlxsw. It will result in un-offloading of the
+ * grafted qdisc as well as the qdisc in the qdisc new location.
+ * (However, if the graft is to the location where the qdisc is already at, it
+ * will be ignored completely and won't cause un-offloading).
+ */
+static int
+mlxsw_sp_qdisc_prio_graft(struct mlxsw_sp_port *mlxsw_sp_port,
+			  struct mlxsw_sp_qdisc *mlxsw_sp_qdisc,
+			  struct tc_prio_qopt_offload_graft_params *p)
+{
+	int tclass_num = MLXSW_SP_PRIO_BAND_TO_TCLASS(p->band);
+	struct mlxsw_sp_qdisc *old_qdisc;
+
+	/* Check if the grafted qdisc is already in its "new" location. If so -
+	 * nothing needs to be done.
+	 */
+	if (p->band < IEEE_8021QAZ_MAX_TCS &&
+	    mlxsw_sp_port->tclass_qdiscs[tclass_num].handle == p->child_handle)
+		return 0;
+
+	/* See if the grafted qdisc is already offloaded on any tclass. If so,
+	 * unoffload it.
+	 */
+	old_qdisc = mlxsw_sp_qdisc_find_by_handle(mlxsw_sp_port,
+						  p->child_handle);
+	if (old_qdisc)
+		mlxsw_sp_qdisc_destroy(mlxsw_sp_port, old_qdisc);
+
+	mlxsw_sp_qdisc_destroy(mlxsw_sp_port,
+			       &mlxsw_sp_port->tclass_qdiscs[tclass_num]);
+	return -EOPNOTSUPP;
+}
+
 int mlxsw_sp_setup_tc_prio(struct mlxsw_sp_port *mlxsw_sp_port,
 			   struct tc_prio_qopt_offload *p)
 {
 	struct mlxsw_sp_qdisc *mlxsw_sp_qdisc;
 
-	if (p->parent != TC_H_ROOT)
+	mlxsw_sp_qdisc = mlxsw_sp_qdisc_find(mlxsw_sp_port, p->parent, true);
+	if (!mlxsw_sp_qdisc)
 		return -EOPNOTSUPP;
 
-	mlxsw_sp_qdisc = mlxsw_sp_port->root_qdisc;
 	if (p->command == TC_PRIO_REPLACE)
 		return mlxsw_sp_qdisc_replace(mlxsw_sp_port, p->handle,
 					      mlxsw_sp_qdisc,
@@ -589,6 +719,9 @@ int mlxsw_sp_setup_tc_prio(struct mlxsw_sp_port *mlxsw_sp_port,
 	case TC_PRIO_STATS:
 		return mlxsw_sp_qdisc_get_stats(mlxsw_sp_port, mlxsw_sp_qdisc,
 						&p->stats);
+	case TC_PRIO_GRAFT:
+		return mlxsw_sp_qdisc_prio_graft(mlxsw_sp_port, mlxsw_sp_qdisc,
+						 &p->graft_params);
 	default:
 		return -EOPNOTSUPP;
 	}
@@ -596,17 +729,36 @@ int mlxsw_sp_setup_tc_prio(struct mlxsw_sp_port *mlxsw_sp_port,
 
 int mlxsw_sp_tc_qdisc_init(struct mlxsw_sp_port *mlxsw_sp_port)
 {
-	mlxsw_sp_port->root_qdisc = kzalloc(sizeof(*mlxsw_sp_port->root_qdisc),
-					    GFP_KERNEL);
-	if (!mlxsw_sp_port->root_qdisc)
-		return -ENOMEM;
+	struct mlxsw_sp_qdisc *mlxsw_sp_qdisc;
+	int i;
 
+	mlxsw_sp_qdisc = kzalloc(sizeof(*mlxsw_sp_qdisc), GFP_KERNEL);
+	if (!mlxsw_sp_qdisc)
+		goto err_root_qdisc_init;
+
+	mlxsw_sp_port->root_qdisc = mlxsw_sp_qdisc;
+	mlxsw_sp_port->root_qdisc->prio_bitmap = 0xff;
 	mlxsw_sp_port->root_qdisc->tclass_num = MLXSW_SP_PORT_DEFAULT_TCLASS;
 
+	mlxsw_sp_qdisc = kzalloc(sizeof(*mlxsw_sp_qdisc) * IEEE_8021QAZ_MAX_TCS,
+				 GFP_KERNEL);
+	if (!mlxsw_sp_qdisc)
+		goto err_tclass_qdiscs_init;
+
+	mlxsw_sp_port->tclass_qdiscs = mlxsw_sp_qdisc;
+	for (i = 0; i < IEEE_8021QAZ_MAX_TCS; i++)
+		mlxsw_sp_port->tclass_qdiscs[i].tclass_num = i;
+
 	return 0;
+
+err_tclass_qdiscs_init:
+	kfree(mlxsw_sp_port->root_qdisc);
+err_root_qdisc_init:
+	return -ENOMEM;
 }
 
 void mlxsw_sp_tc_qdisc_fini(struct mlxsw_sp_port *mlxsw_sp_port)
 {
+	kfree(mlxsw_sp_port->tclass_qdiscs);
 	kfree(mlxsw_sp_port->root_qdisc);
 }

include/net/pkt_cls.h

@@ -806,6 +806,7 @@ enum tc_prio_command {
 	TC_PRIO_REPLACE,
 	TC_PRIO_DESTROY,
 	TC_PRIO_STATS,
+	TC_PRIO_GRAFT,
 };
 
 struct tc_prio_qopt_offload_params {
@@ -818,6 +819,11 @@ struct tc_prio_qopt_offload_params {
 	struct gnet_stats_queue *qstats;
 };
 
+struct tc_prio_qopt_offload_graft_params {
+	u8 band;
+	u32 child_handle;
+};
+
 struct tc_prio_qopt_offload {
 	enum tc_prio_command command;
 	u32 handle;
@@ -825,6 +831,8 @@ struct tc_prio_qopt_offload {
 	union {
 		struct tc_prio_qopt_offload_params replace_params;
 		struct tc_qopt_offload_stats stats;
+		struct tc_prio_qopt_offload_graft_params graft_params;
 	};
 };
+
 #endif

net/sched/sch_api.c

@@ -739,6 +739,7 @@ static u32 qdisc_alloc_handle(struct net_device *dev)
 void qdisc_tree_reduce_backlog(struct Qdisc *sch, unsigned int n,
 			       unsigned int len)
 {
+	bool qdisc_is_offloaded = sch->flags & TCQ_F_OFFLOADED;
 	const struct Qdisc_class_ops *cops;
 	unsigned long cl;
 	u32 parentid;
@@ -760,8 +761,12 @@ void qdisc_tree_reduce_backlog(struct Qdisc *sch, unsigned int n,
 		 * If child was empty even before update then backlog
 		 * counter is screwed and we skip notification because
 		 * parent class is already passive.
+		 *
+		 * If the original child was offloaded then it is allowed
+		 * to be seem as empty, so the parent is notified anyway.
 		 */
-		notify = !sch->q.qlen && !WARN_ON_ONCE(!n);
+		notify = !sch->q.qlen && !WARN_ON_ONCE(!n &&
+						       !qdisc_is_offloaded);
 		/* TODO: perform the search on a per txq basis */
 		sch = qdisc_lookup(qdisc_dev(sch), TC_H_MAJ(parentid));
 		if (sch == NULL) {

net/sched/sch_prio.c

@@ -142,9 +142,8 @@ prio_reset(struct Qdisc *sch)
 	sch->q.qlen = 0;
 }
 
-static int prio_offload(struct Qdisc *sch, bool enable)
+static int prio_offload(struct Qdisc *sch, struct tc_prio_qopt *qopt)
 {
-	struct prio_sched_data *q = qdisc_priv(sch);
 	struct net_device *dev = qdisc_dev(sch);
 	struct tc_prio_qopt_offload opt = {
 		.handle = sch->handle,
@@ -154,10 +153,10 @@ static int prio_offload(struct Qdisc *sch, bool enable)
 	if (!tc_can_offload(dev) || !dev->netdev_ops->ndo_setup_tc)
 		return -EOPNOTSUPP;
 
-	if (enable) {
+	if (qopt) {
 		opt.command = TC_PRIO_REPLACE;
-		opt.replace_params.bands = q->bands;
-		memcpy(&opt.replace_params.priomap, q->prio2band,
+		opt.replace_params.bands = qopt->bands;
+		memcpy(&opt.replace_params.priomap, qopt->priomap,
 		       TC_PRIO_MAX + 1);
 		opt.replace_params.qstats = &sch->qstats;
 	} else {
@@ -174,7 +173,7 @@ prio_destroy(struct Qdisc *sch)
 	struct prio_sched_data *q = qdisc_priv(sch);
 
 	tcf_block_put(q->block);
-	prio_offload(sch, false);
+	prio_offload(sch, NULL);
 	for (prio = 0; prio < q->bands; prio++)
 		qdisc_destroy(q->queues[prio]);
 }
@@ -211,6 +210,7 @@ static int prio_tune(struct Qdisc *sch, struct nlattr *opt,
 		}
 	}
 
+	prio_offload(sch, qopt);
 	sch_tree_lock(sch);
 	q->bands = qopt->bands;
 	memcpy(q->prio2band, qopt->priomap, TC_PRIO_MAX+1);
@@ -230,7 +230,6 @@ static int prio_tune(struct Qdisc *sch, struct nlattr *opt,
 	}
 
 	sch_tree_unlock(sch);
-	prio_offload(sch, true);
 	return 0;
 }
 
@@ -309,12 +308,44 @@ static int prio_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
 		      struct Qdisc **old, struct netlink_ext_ack *extack)
 {
 	struct prio_sched_data *q = qdisc_priv(sch);
+	struct tc_prio_qopt_offload graft_offload;
+	struct net_device *dev = qdisc_dev(sch);
 	unsigned long band = arg - 1;
+	bool any_qdisc_is_offloaded;
+	int err;
 
 	if (new == NULL)
 		new = &noop_qdisc;
 
 	*old = qdisc_replace(sch, new, &q->queues[band]);
+
+	if (!tc_can_offload(dev))
+		return 0;
+
+	graft_offload.handle = sch->handle;
+	graft_offload.parent = sch->parent;
+	graft_offload.graft_params.band = band;
+	graft_offload.graft_params.child_handle = new->handle;
+	graft_offload.command = TC_PRIO_GRAFT;
+
+	err = dev->netdev_ops->ndo_setup_tc(dev, TC_SETUP_QDISC_PRIO,
+					    &graft_offload);
+
+	/* Don't report error if the graft is part of destroy operation. */
+	if (err && new != &noop_qdisc) {
+		/* Don't report error if the parent, the old child and the new
+		 * one are not offloaded.
+		 */
+		any_qdisc_is_offloaded = sch->flags & TCQ_F_OFFLOADED;
+		any_qdisc_is_offloaded |= new->flags & TCQ_F_OFFLOADED;
+		if (*old)
+			any_qdisc_is_offloaded |= (*old)->flags &
+						   TCQ_F_OFFLOADED;
+
+		if (any_qdisc_is_offloaded)
+			NL_SET_ERR_MSG(extack, "Offloading graft operation failed.");
+	}
+
 	return 0;
 }