Merge branch 'mlxsw-Improve-IPv6-route-insertion-rate'

Ido Schimmel says:

====================
mlxsw: Improve IPv6 route insertion rate

Unlike IPv4, an IPv6 multipath route in the kernel is composed from
multiple sibling routes, each representing a single nexthop.

Therefore, an addition of a multipath route with N nexthops translates
to N in-kernel notifications. This is inefficient for device drivers
that need to program the route to the underlying device. Each time a new
nexthop is appended, a new nexthop group needs to be constructed and the
old one deleted.

This patchset improves the situation by sending a single notification
for a multipath route addition / deletion instead of one per-nexthop.
When adding thousands of multipath routes with 16 nexthops, I measured
an improvement of about x10 in the insertion rate.

Patches #1-#3 add a flag that indicates that in-kernel notifications
need to be suppressed and extend the IPv6 FIB notification info with
information about the number of sibling routes that are being notified.

Patches #4-#5 adjust the two current listeners to these notifications to
ignore notifications about IPv6 multipath routes.

Patches #6-#7 adds add / delete notifications for IPv6 multipath routes.

Patches #8-#14 do the same for mlxsw.

Patch #15 finally removes the limitations added in patches #4-#5 and
stops the kernel from sending a notification for each added / deleted
nexthop.

Patch #16 adds test cases.

v2 (David Ahern):
* Remove patch adjusting netdevsim to consume resources for each
  fib6_info. Instead, consume one resource for the entire multipath
  route
* Remove 'multipath_rt' usage in patch #10
* Remove 'multipath_rt' from 'struct fib6_entry_notifier_info' in patch
  #15. The member is only removed in this patch to prevent drivers from
  processing multipath routes twice during the series
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

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

@@ -5278,17 +5278,21 @@ mlxsw_sp_nexthop6_group_update(struct mlxsw_sp *mlxsw_sp,
 static int
 mlxsw_sp_fib6_entry_nexthop_add(struct mlxsw_sp *mlxsw_sp,
 				struct mlxsw_sp_fib6_entry *fib6_entry,
-				struct fib6_info *rt)
+				struct fib6_info **rt_arr, unsigned int nrt6)
 {
 	struct mlxsw_sp_rt6 *mlxsw_sp_rt6;
-	int err;
+	int err, i;
 
-	mlxsw_sp_rt6 = mlxsw_sp_rt6_create(rt);
-	if (IS_ERR(mlxsw_sp_rt6))
-		return PTR_ERR(mlxsw_sp_rt6);
+	for (i = 0; i < nrt6; i++) {
+		mlxsw_sp_rt6 = mlxsw_sp_rt6_create(rt_arr[i]);
+		if (IS_ERR(mlxsw_sp_rt6)) {
+			err = PTR_ERR(mlxsw_sp_rt6);
+			goto err_rt6_create;
+		}
 
-	list_add_tail(&mlxsw_sp_rt6->list, &fib6_entry->rt6_list);
-	fib6_entry->nrt6++;
+		list_add_tail(&mlxsw_sp_rt6->list, &fib6_entry->rt6_list);
+		fib6_entry->nrt6++;
+	}
 
 	err = mlxsw_sp_nexthop6_group_update(mlxsw_sp, fib6_entry);
 	if (err)
@@ -5297,27 +5301,38 @@ mlxsw_sp_fib6_entry_nexthop_add(struct mlxsw_sp *mlxsw_sp,
 	return 0;
 
 err_nexthop6_group_update:
-	fib6_entry->nrt6--;
-	list_del(&mlxsw_sp_rt6->list);
-	mlxsw_sp_rt6_destroy(mlxsw_sp_rt6);
+	i = nrt6;
+err_rt6_create:
+	for (i--; i >= 0; i--) {
+		fib6_entry->nrt6--;
+		mlxsw_sp_rt6 = list_last_entry(&fib6_entry->rt6_list,
+					       struct mlxsw_sp_rt6, list);
+		list_del(&mlxsw_sp_rt6->list);
+		mlxsw_sp_rt6_destroy(mlxsw_sp_rt6);
+	}
 	return err;
 }
 
 static void
 mlxsw_sp_fib6_entry_nexthop_del(struct mlxsw_sp *mlxsw_sp,
 				struct mlxsw_sp_fib6_entry *fib6_entry,
-				struct fib6_info *rt)
+				struct fib6_info **rt_arr, unsigned int nrt6)
 {
 	struct mlxsw_sp_rt6 *mlxsw_sp_rt6;
+	int i;
 
-	mlxsw_sp_rt6 = mlxsw_sp_fib6_entry_rt_find(fib6_entry, rt);
-	if (WARN_ON(!mlxsw_sp_rt6))
-		return;
+	for (i = 0; i < nrt6; i++) {
+		mlxsw_sp_rt6 = mlxsw_sp_fib6_entry_rt_find(fib6_entry,
+							   rt_arr[i]);
+		if (WARN_ON_ONCE(!mlxsw_sp_rt6))
+			continue;
+
+		fib6_entry->nrt6--;
+		list_del(&mlxsw_sp_rt6->list);
+		mlxsw_sp_rt6_destroy(mlxsw_sp_rt6);
+	}
 
-	fib6_entry->nrt6--;
-	list_del(&mlxsw_sp_rt6->list);
 	mlxsw_sp_nexthop6_group_update(mlxsw_sp, fib6_entry);
-	mlxsw_sp_rt6_destroy(mlxsw_sp_rt6);
 }
 
 static void mlxsw_sp_fib6_entry_type_set(struct mlxsw_sp *mlxsw_sp,
@@ -5358,29 +5373,32 @@ mlxsw_sp_fib6_entry_rt_destroy_all(struct mlxsw_sp_fib6_entry *fib6_entry)
 static struct mlxsw_sp_fib6_entry *
 mlxsw_sp_fib6_entry_create(struct mlxsw_sp *mlxsw_sp,
 			   struct mlxsw_sp_fib_node *fib_node,
-			   struct fib6_info *rt)
+			   struct fib6_info **rt_arr, unsigned int nrt6)
 {
 	struct mlxsw_sp_fib6_entry *fib6_entry;
 	struct mlxsw_sp_fib_entry *fib_entry;
 	struct mlxsw_sp_rt6 *mlxsw_sp_rt6;
-	int err;
+	int err, i;
 
 	fib6_entry = kzalloc(sizeof(*fib6_entry), GFP_KERNEL);
 	if (!fib6_entry)
 		return ERR_PTR(-ENOMEM);
 	fib_entry = &fib6_entry->common;
 
-	mlxsw_sp_rt6 = mlxsw_sp_rt6_create(rt);
-	if (IS_ERR(mlxsw_sp_rt6)) {
-		err = PTR_ERR(mlxsw_sp_rt6);
-		goto err_rt6_create;
+	INIT_LIST_HEAD(&fib6_entry->rt6_list);
+
+	for (i = 0; i < nrt6; i++) {
+		mlxsw_sp_rt6 = mlxsw_sp_rt6_create(rt_arr[i]);
+		if (IS_ERR(mlxsw_sp_rt6)) {
+			err = PTR_ERR(mlxsw_sp_rt6);
+			goto err_rt6_create;
+		}
+		list_add_tail(&mlxsw_sp_rt6->list, &fib6_entry->rt6_list);
+		fib6_entry->nrt6++;
 	}
 
-	mlxsw_sp_fib6_entry_type_set(mlxsw_sp, fib_entry, mlxsw_sp_rt6->rt);
+	mlxsw_sp_fib6_entry_type_set(mlxsw_sp, fib_entry, rt_arr[0]);
 
-	INIT_LIST_HEAD(&fib6_entry->rt6_list);
-	list_add_tail(&mlxsw_sp_rt6->list, &fib6_entry->rt6_list);
-	fib6_entry->nrt6 = 1;
 	err = mlxsw_sp_nexthop6_group_get(mlxsw_sp, fib6_entry);
 	if (err)
 		goto err_nexthop6_group_get;
@@ -5390,9 +5408,15 @@ mlxsw_sp_fib6_entry_create(struct mlxsw_sp *mlxsw_sp,
 	return fib6_entry;
 
 err_nexthop6_group_get:
-	list_del(&mlxsw_sp_rt6->list);
-	mlxsw_sp_rt6_destroy(mlxsw_sp_rt6);
+	i = nrt6;
 err_rt6_create:
+	for (i--; i >= 0; i--) {
+		fib6_entry->nrt6--;
+		mlxsw_sp_rt6 = list_last_entry(&fib6_entry->rt6_list,
+					       struct mlxsw_sp_rt6, list);
+		list_del(&mlxsw_sp_rt6->list);
+		mlxsw_sp_rt6_destroy(mlxsw_sp_rt6);
+	}
 	kfree(fib6_entry);
 	return ERR_PTR(err);
 }
@@ -5435,16 +5459,16 @@ mlxsw_sp_fib6_node_entry_find(const struct mlxsw_sp_fib_node *fib_node,
 
 static int
 mlxsw_sp_fib6_node_list_insert(struct mlxsw_sp_fib6_entry *new6_entry,
-			       bool replace)
+			       bool *p_replace)
 {
 	struct mlxsw_sp_fib_node *fib_node = new6_entry->common.fib_node;
 	struct fib6_info *nrt = mlxsw_sp_fib6_entry_rt(new6_entry);
 	struct mlxsw_sp_fib6_entry *fib6_entry;
 
-	fib6_entry = mlxsw_sp_fib6_node_entry_find(fib_node, nrt, replace);
+	fib6_entry = mlxsw_sp_fib6_node_entry_find(fib_node, nrt, *p_replace);
 
-	if (replace && WARN_ON(!fib6_entry))
-		return -EINVAL;
+	if (*p_replace && !fib6_entry)
+		*p_replace = false;
 
 	if (fib6_entry) {
 		list_add_tail(&new6_entry->common.list,
@@ -5479,11 +5503,11 @@ mlxsw_sp_fib6_node_list_remove(struct mlxsw_sp_fib6_entry *fib6_entry)
 
 static int mlxsw_sp_fib6_node_entry_link(struct mlxsw_sp *mlxsw_sp,
 					 struct mlxsw_sp_fib6_entry *fib6_entry,
-					 bool replace)
+					 bool *p_replace)
 {
 	int err;
 
-	err = mlxsw_sp_fib6_node_list_insert(fib6_entry, replace);
+	err = mlxsw_sp_fib6_node_list_insert(fib6_entry, p_replace);
 	if (err)
 		return err;
 
@@ -5556,10 +5580,12 @@ static void mlxsw_sp_fib6_entry_replace(struct mlxsw_sp *mlxsw_sp,
 }
 
 static int mlxsw_sp_router_fib6_add(struct mlxsw_sp *mlxsw_sp,
-				    struct fib6_info *rt, bool replace)
+				    struct fib6_info **rt_arr,
+				    unsigned int nrt6, bool replace)
 {
 	struct mlxsw_sp_fib6_entry *fib6_entry;
 	struct mlxsw_sp_fib_node *fib_node;
+	struct fib6_info *rt = rt_arr[0];
 	int err;
 
 	if (mlxsw_sp->router->aborted)
@@ -5584,19 +5610,21 @@ static int mlxsw_sp_router_fib6_add(struct mlxsw_sp *mlxsw_sp,
 	 */
 	fib6_entry = mlxsw_sp_fib6_node_mp_entry_find(fib_node, rt, replace);
 	if (fib6_entry) {
-		err = mlxsw_sp_fib6_entry_nexthop_add(mlxsw_sp, fib6_entry, rt);
+		err = mlxsw_sp_fib6_entry_nexthop_add(mlxsw_sp, fib6_entry,
+						      rt_arr, nrt6);
 		if (err)
 			goto err_fib6_entry_nexthop_add;
 		return 0;
 	}
 
-	fib6_entry = mlxsw_sp_fib6_entry_create(mlxsw_sp, fib_node, rt);
+	fib6_entry = mlxsw_sp_fib6_entry_create(mlxsw_sp, fib_node, rt_arr,
+						nrt6);
 	if (IS_ERR(fib6_entry)) {
 		err = PTR_ERR(fib6_entry);
 		goto err_fib6_entry_create;
 	}
 
-	err = mlxsw_sp_fib6_node_entry_link(mlxsw_sp, fib6_entry, replace);
+	err = mlxsw_sp_fib6_node_entry_link(mlxsw_sp, fib6_entry, &replace);
 	if (err)
 		goto err_fib6_node_entry_link;
 
@@ -5613,10 +5641,12 @@ static int mlxsw_sp_router_fib6_add(struct mlxsw_sp *mlxsw_sp,
 }
 
 static void mlxsw_sp_router_fib6_del(struct mlxsw_sp *mlxsw_sp,
-				     struct fib6_info *rt)
+				     struct fib6_info **rt_arr,
+				     unsigned int nrt6)
 {
 	struct mlxsw_sp_fib6_entry *fib6_entry;
 	struct mlxsw_sp_fib_node *fib_node;
+	struct fib6_info *rt = rt_arr[0];
 
 	if (mlxsw_sp->router->aborted)
 		return;
@@ -5628,11 +5658,12 @@ static void mlxsw_sp_router_fib6_del(struct mlxsw_sp *mlxsw_sp,
 	if (WARN_ON(!fib6_entry))
 		return;
 
-	/* If route is part of a multipath entry, but not the last one
-	 * removed, then only reduce its nexthop group.
+	/* If not all the nexthops are deleted, then only reduce the nexthop
+	 * group.
 	 */
-	if (!list_is_singular(&fib6_entry->rt6_list)) {
-		mlxsw_sp_fib6_entry_nexthop_del(mlxsw_sp, fib6_entry, rt);
+	if (nrt6 != fib6_entry->nrt6) {
+		mlxsw_sp_fib6_entry_nexthop_del(mlxsw_sp, fib6_entry, rt_arr,
+						nrt6);
 		return;
 	}
 
@@ -5893,10 +5924,15 @@ static void mlxsw_sp_router_fib_abort(struct mlxsw_sp *mlxsw_sp)
 		dev_warn(mlxsw_sp->bus_info->dev, "Failed to set abort trap.\n");
 }
 
+struct mlxsw_sp_fib6_event_work {
+	struct fib6_info **rt_arr;
+	unsigned int nrt6;
+};
+
 struct mlxsw_sp_fib_event_work {
 	struct work_struct work;
 	union {
-		struct fib6_entry_notifier_info fen6_info;
+		struct mlxsw_sp_fib6_event_work fib6_work;
 		struct fib_entry_notifier_info fen_info;
 		struct fib_rule_notifier_info fr_info;
 		struct fib_nh_notifier_info fnh_info;
@@ -5907,6 +5943,54 @@ struct mlxsw_sp_fib_event_work {
 	unsigned long event;
 };
 
+static int
+mlxsw_sp_router_fib6_work_init(struct mlxsw_sp_fib6_event_work *fib6_work,
+			       struct fib6_entry_notifier_info *fen6_info)
+{
+	struct fib6_info *rt = fen6_info->rt;
+	struct fib6_info **rt_arr;
+	struct fib6_info *iter;
+	unsigned int nrt6;
+	int i = 0;
+
+	nrt6 = fen6_info->nsiblings + 1;
+
+	rt_arr = kcalloc(nrt6, sizeof(struct fib6_info *), GFP_ATOMIC);
+	if (!rt_arr)
+		return -ENOMEM;
+
+	fib6_work->rt_arr = rt_arr;
+	fib6_work->nrt6 = nrt6;
+
+	rt_arr[0] = rt;
+	fib6_info_hold(rt);
+
+	if (!fen6_info->nsiblings)
+		return 0;
+
+	list_for_each_entry(iter, &rt->fib6_siblings, fib6_siblings) {
+		if (i == fen6_info->nsiblings)
+			break;
+
+		rt_arr[i + 1] = iter;
+		fib6_info_hold(iter);
+		i++;
+	}
+	WARN_ON_ONCE(i != fen6_info->nsiblings);
+
+	return 0;
+}
+
+static void
+mlxsw_sp_router_fib6_work_fini(struct mlxsw_sp_fib6_event_work *fib6_work)
+{
+	int i;
+
+	for (i = 0; i < fib6_work->nrt6; i++)
+		mlxsw_sp_rt6_release(fib6_work->rt_arr[i]);
+	kfree(fib6_work->rt_arr);
+}
+
 static void mlxsw_sp_router_fib4_event_work(struct work_struct *work)
 {
 	struct mlxsw_sp_fib_event_work *fib_work =
@@ -5965,18 +6049,21 @@ static void mlxsw_sp_router_fib6_event_work(struct work_struct *work)
 
 	switch (fib_work->event) {
 	case FIB_EVENT_ENTRY_REPLACE: /* fall through */
-	case FIB_EVENT_ENTRY_APPEND: /* fall through */
 	case FIB_EVENT_ENTRY_ADD:
 		replace = fib_work->event == FIB_EVENT_ENTRY_REPLACE;
 		err = mlxsw_sp_router_fib6_add(mlxsw_sp,
-					       fib_work->fen6_info.rt, replace);
+					       fib_work->fib6_work.rt_arr,
+					       fib_work->fib6_work.nrt6,
+					       replace);
 		if (err)
 			mlxsw_sp_router_fib_abort(mlxsw_sp);
-		mlxsw_sp_rt6_release(fib_work->fen6_info.rt);
+		mlxsw_sp_router_fib6_work_fini(&fib_work->fib6_work);
 		break;
 	case FIB_EVENT_ENTRY_DEL:
-		mlxsw_sp_router_fib6_del(mlxsw_sp, fib_work->fen6_info.rt);
-		mlxsw_sp_rt6_release(fib_work->fen6_info.rt);
+		mlxsw_sp_router_fib6_del(mlxsw_sp,
+					 fib_work->fib6_work.rt_arr,
+					 fib_work->fib6_work.nrt6);
+		mlxsw_sp_router_fib6_work_fini(&fib_work->fib6_work);
 		break;
 	case FIB_EVENT_RULE_ADD:
 		/* if we get here, a rule was added that we do not support.
@@ -6065,22 +6152,26 @@ static void mlxsw_sp_router_fib4_event(struct mlxsw_sp_fib_event_work *fib_work,
 	}
 }
 
-static void mlxsw_sp_router_fib6_event(struct mlxsw_sp_fib_event_work *fib_work,
-				       struct fib_notifier_info *info)
+static int mlxsw_sp_router_fib6_event(struct mlxsw_sp_fib_event_work *fib_work,
+				      struct fib_notifier_info *info)
 {
 	struct fib6_entry_notifier_info *fen6_info;
+	int err;
 
 	switch (fib_work->event) {
 	case FIB_EVENT_ENTRY_REPLACE: /* fall through */
-	case FIB_EVENT_ENTRY_APPEND: /* fall through */
 	case FIB_EVENT_ENTRY_ADD: /* fall through */
 	case FIB_EVENT_ENTRY_DEL:
 		fen6_info = container_of(info, struct fib6_entry_notifier_info,
 					 info);
-		fib_work->fen6_info = *fen6_info;
-		fib6_info_hold(fib_work->fen6_info.rt);
+		err = mlxsw_sp_router_fib6_work_init(&fib_work->fib6_work,
+						     fen6_info);
+		if (err)
+			return err;
 		break;
 	}
+
+	return 0;
 }
 
 static void
@@ -6221,7 +6312,9 @@ static int mlxsw_sp_router_fib_event(struct notifier_block *nb,
 		break;
 	case AF_INET6:
 		INIT_WORK(&fib_work->work, mlxsw_sp_router_fib6_event_work);
-		mlxsw_sp_router_fib6_event(fib_work, info);
+		err = mlxsw_sp_router_fib6_event(fib_work, info);
+		if (err)
+			goto err_fib_event;
 		break;
 	case RTNL_FAMILY_IP6MR:
 	case RTNL_FAMILY_IPMR:
@@ -6233,6 +6326,10 @@ static int mlxsw_sp_router_fib_event(struct notifier_block *nb,
 	mlxsw_core_schedule_work(&fib_work->work);
 
 	return NOTIFY_DONE;
+
+err_fib_event:
+	kfree(fib_work);
+	return NOTIFY_BAD;
 }
 
 struct mlxsw_sp_rif *

include/net/ip6_fib.h

@@ -377,6 +377,7 @@ typedef struct rt6_info *(*pol_lookup_t)(struct net *,
 struct fib6_entry_notifier_info {
 	struct fib_notifier_info info; /* must be first */
 	struct fib6_info *rt;
+	unsigned int nsiblings;
 };
 
 /*
@@ -450,6 +451,11 @@ int call_fib6_entry_notifiers(struct net *net,
 			      enum fib_event_type event_type,
 			      struct fib6_info *rt,
 			      struct netlink_ext_ack *extack);
+int call_fib6_multipath_entry_notifiers(struct net *net,
+					enum fib_event_type event_type,
+					struct fib6_info *rt,
+					unsigned int nsiblings,
+					struct netlink_ext_ack *extack);
 void fib6_rt_update(struct net *net, struct fib6_info *rt,
 		    struct nl_info *info);
 void inet6_rt_notify(int event, struct fib6_info *rt, struct nl_info *info,

include/net/netlink.h

@@ -378,13 +378,17 @@ struct nla_policy {
 /**
  * struct nl_info - netlink source information
  * @nlh: Netlink message header of original request
+ * @nl_net: Network namespace
  * @portid: Netlink PORTID of requesting application
+ * @skip_notify: Skip netlink notifications to user space
+ * @skip_notify_kernel: Skip selected in-kernel notifications
  */
 struct nl_info {
 	struct nlmsghdr		*nlh;
 	struct net		*nl_net;
 	u32			portid;
-	bool			skip_notify;
+	u8			skip_notify:1,
+				skip_notify_kernel:1;
 };
 
 /**

net/ipv6/ip6_fib.c

@@ -381,6 +381,22 @@ int call_fib6_entry_notifiers(struct net *net,
 	return call_fib6_notifiers(net, event_type, &info.info);
 }
 
+int call_fib6_multipath_entry_notifiers(struct net *net,
+					enum fib_event_type event_type,
+					struct fib6_info *rt,
+					unsigned int nsiblings,
+					struct netlink_ext_ack *extack)
+{
+	struct fib6_entry_notifier_info info = {
+		.info.extack = extack,
+		.rt = rt,
+		.nsiblings = nsiblings,
+	};
+
+	rt->fib6_table->fib_seq++;
+	return call_fib6_notifiers(net, event_type, &info.info);
+}
+
 struct fib6_dump_arg {
 	struct net *net;
 	struct notifier_block *nb;
@@ -1123,11 +1139,13 @@ static int fib6_add_rt2node(struct fib6_node *fn, struct fib6_info *rt,
 add:
 		nlflags |= NLM_F_CREATE;
 
-		err = call_fib6_entry_notifiers(info->nl_net,
-						FIB_EVENT_ENTRY_ADD,
-						rt, extack);
-		if (err)
-			return err;
+		if (!info->skip_notify_kernel) {
+			err = call_fib6_entry_notifiers(info->nl_net,
+							FIB_EVENT_ENTRY_ADD,
+							rt, extack);
+			if (err)
+				return err;
+		}
 
 		rcu_assign_pointer(rt->fib6_next, iter);
 		fib6_info_hold(rt);
@@ -1152,11 +1170,13 @@ static int fib6_add_rt2node(struct fib6_node *fn, struct fib6_info *rt,
 			return -ENOENT;
 		}
 
-		err = call_fib6_entry_notifiers(info->nl_net,
-						FIB_EVENT_ENTRY_REPLACE,
-						rt, extack);
-		if (err)
-			return err;
+		if (!info->skip_notify_kernel) {
+			err = call_fib6_entry_notifiers(info->nl_net,
+							FIB_EVENT_ENTRY_REPLACE,
+							rt, extack);
+			if (err)
+				return err;
+		}
 
 		fib6_info_hold(rt);
 		rcu_assign_pointer(rt->fib6_node, fn);
@@ -1839,9 +1859,11 @@ static void fib6_del_route(struct fib6_table *table, struct fib6_node *fn,
 
 	fib6_purge_rt(rt, fn, net);
 
-	call_fib6_entry_notifiers(net, FIB_EVENT_ENTRY_DEL, rt, NULL);
+	if (!info->skip_notify_kernel)
+		call_fib6_entry_notifiers(net, FIB_EVENT_ENTRY_DEL, rt, NULL);
 	if (!info->skip_notify)
 		inet6_rt_notify(RTM_DELROUTE, rt, info, 0);
+
 	fib6_info_release(rt);
 }
 

net/ipv6/route.c

@@ -3718,6 +3718,12 @@ static int __ip6_del_rt_siblings(struct fib6_info *rt, struct fib6_config *cfg)
 				info->skip_notify = 1;
 		}
 
+		info->skip_notify_kernel = 1;
+		call_fib6_multipath_entry_notifiers(net,
+						    FIB_EVENT_ENTRY_DEL,
+						    rt,
+						    rt->fib6_nsiblings,
+						    NULL);
 		list_for_each_entry_safe(sibling, next_sibling,
 					 &rt->fib6_siblings,
 					 fib6_siblings) {
@@ -4965,6 +4971,7 @@ static int ip6_route_multipath_add(struct fib6_config *cfg,
 {
 	struct fib6_info *rt_notif = NULL, *rt_last = NULL;
 	struct nl_info *info = &cfg->fc_nlinfo;
+	enum fib_event_type event_type;
 	struct fib6_config r_cfg;
 	struct rtnexthop *rtnh;
 	struct fib6_info *rt;
@@ -5042,6 +5049,11 @@ static int ip6_route_multipath_add(struct fib6_config *cfg,
 	 */
 	info->skip_notify = 1;
 
+	/* For add and replace, send one notification with all nexthops. For
+	 * append, send one notification with all appended nexthops.
+	 */
+	info->skip_notify_kernel = 1;
+
 	err_nh = NULL;
 	list_for_each_entry(nh, &rt6_nh_list, next) {
 		err = __ip6_ins_rt(nh->fib6_info, info, extack);
@@ -5078,6 +5090,15 @@ static int ip6_route_multipath_add(struct fib6_config *cfg,
 		nhn++;
 	}
 
+	event_type = replace ? FIB_EVENT_ENTRY_REPLACE : FIB_EVENT_ENTRY_ADD;
+	err = call_fib6_multipath_entry_notifiers(info->nl_net, event_type,
+						  rt_notif, nhn - 1, extack);
+	if (err) {
+		/* Delete all the siblings that were just added */
+		err_nh = NULL;
+		goto add_errout;
+	}
+
 	/* success ... tell user about new route */
 	ip6_route_mpath_notify(rt_notif, rt_last, info, nlflags);
 	goto cleanup;

tools/testing/selftests/drivers/net/mlxsw/fib_offload.sh

+#!/bin/bash
+# SPDX-License-Identifier: GPL-2.0
+#
+# Test unicast FIB offload indication.
+
+lib_dir=$(dirname $0)/../../../net/forwarding
+
+ALL_TESTS="
+	ipv6_route_add
+	ipv6_route_replace
+	ipv6_route_nexthop_group_share
+	ipv6_route_rate
+"
+NUM_NETIFS=4
+source $lib_dir/lib.sh
+source $lib_dir/devlink_lib.sh
+
+tor1_create()
+{
+	simple_if_init $tor1_p1 2001:db8:1::2/128 2001:db8:1::3/128
+}
+
+tor1_destroy()
+{
+	simple_if_fini $tor1_p1 2001:db8:1::2/128 2001:db8:1::3/128
+}
+
+tor2_create()
+{
+	simple_if_init $tor2_p1 2001:db8:2::2/128 2001:db8:2::3/128
+}
+
+tor2_destroy()
+{
+	simple_if_fini $tor2_p1 2001:db8:2::2/128 2001:db8:2::3/128
+}
+
+spine_create()
+{
+	ip link set dev $spine_p1 up
+	ip link set dev $spine_p2 up
+
+	__addr_add_del $spine_p1 add 2001:db8:1::1/64
+	__addr_add_del $spine_p2 add 2001:db8:2::1/64
+}
+
+spine_destroy()
+{
+	__addr_add_del $spine_p2 del 2001:db8:2::1/64
+	__addr_add_del $spine_p1 del 2001:db8:1::1/64
+
+	ip link set dev $spine_p2 down
+	ip link set dev $spine_p1 down
+}
+
+ipv6_offload_check()
+{
+	local pfx="$1"; shift
+	local expected_num=$1; shift
+	local num
+
+	# Try to avoid races with route offload
+	sleep .1
+
+	num=$(ip -6 route show match ${pfx} | grep "offload" | wc -l)
+
+	if [ $num -eq $expected_num ]; then
+		return 0
+	fi
+
+	return 1
+}
+
+ipv6_route_add_prefix()
+{
+	RET=0
+
+	# Add a prefix route and check that it is offloaded.
+	ip -6 route add 2001:db8:3::/64 dev $spine_p1 metric 100
+	ipv6_offload_check "2001:db8:3::/64 dev $spine_p1 metric 100" 1
+	check_err $? "prefix route not offloaded"
+
+	# Append an identical prefix route with an higher metric and check that
+	# offload indication did not change.
+	ip -6 route append 2001:db8:3::/64 dev $spine_p1 metric 200
+	ipv6_offload_check "2001:db8:3::/64 dev $spine_p1 metric 100" 1
+	check_err $? "lowest metric not offloaded after append"
+	ipv6_offload_check "2001:db8:3::/64 dev $spine_p1 metric 200" 0
+	check_err $? "highest metric offloaded when should not"
+
+	# Prepend an identical prefix route with lower metric and check that
+	# it is offloaded and the others are not.
+	ip -6 route append 2001:db8:3::/64 dev $spine_p1 metric 10
+	ipv6_offload_check "2001:db8:3::/64 dev $spine_p1 metric 10" 1
+	check_err $? "lowest metric not offloaded after prepend"
+	ipv6_offload_check "2001:db8:3::/64 dev $spine_p1 metric 100" 0
+	check_err $? "mid metric offloaded when should not"
+	ipv6_offload_check "2001:db8:3::/64 dev $spine_p1 metric 200" 0
+	check_err $? "highest metric offloaded when should not"
+
+	# Delete the routes and add the same route with a different nexthop
+	# device. Check that it is offloaded.
+	ip -6 route flush 2001:db8:3::/64 dev $spine_p1
+	ip -6 route add 2001:db8:3::/64 dev $spine_p2
+	ipv6_offload_check "2001:db8:3::/64 dev $spine_p2" 1
+
+	log_test "IPv6 prefix route add"
+
+	ip -6 route flush 2001:db8:3::/64
+}
+
+ipv6_route_add_mpath()
+{
+	RET=0
+
+	# Add a multipath route and check that it is offloaded.
+	ip -6 route add 2001:db8:3::/64 metric 100 \
+		nexthop via 2001:db8:1::2 dev $spine_p1 \
+		nexthop via 2001:db8:2::2 dev $spine_p2
+	ipv6_offload_check "2001:db8:3::/64 metric 100" 2
+	check_err $? "multipath route not offloaded when should"
+
+	# Append another nexthop and check that it is offloaded as well.
+	ip -6 route append 2001:db8:3::/64 metric 100 \
+		nexthop via 2001:db8:1::3 dev $spine_p1
+	ipv6_offload_check "2001:db8:3::/64 metric 100" 3
+	check_err $? "appended nexthop not offloaded when should"
+
+	# Mimic route replace by removing the route and adding it back with
+	# only two nexthops.
+	ip -6 route del 2001:db8:3::/64
+	ip -6 route add 2001:db8:3::/64 metric 100 \
+		nexthop via 2001:db8:1::2 dev $spine_p1 \
+		nexthop via 2001:db8:2::2 dev $spine_p2
+	ipv6_offload_check "2001:db8:3::/64 metric 100" 2
+	check_err $? "multipath route not offloaded after delete & add"
+
+	# Append a nexthop with an higher metric and check that the offload
+	# indication did not change.
+	ip -6 route append 2001:db8:3::/64 metric 200 \
+		nexthop via 2001:db8:1::3 dev $spine_p1
+	ipv6_offload_check "2001:db8:3::/64 metric 100" 2
+	check_err $? "lowest metric not offloaded after append"
+	ipv6_offload_check "2001:db8:3::/64 metric 200" 0
+	check_err $? "highest metric offloaded when should not"
+
+	# Prepend a nexthop with a lower metric and check that it is offloaded
+	# and the others are not.
+	ip -6 route append 2001:db8:3::/64 metric 10 \
+		nexthop via 2001:db8:1::3 dev $spine_p1
+	ipv6_offload_check "2001:db8:3::/64 metric 10" 1
+	check_err $? "lowest metric not offloaded after prepend"
+	ipv6_offload_check "2001:db8:3::/64 metric 100" 0
+	check_err $? "mid metric offloaded when should not"
+	ipv6_offload_check "2001:db8:3::/64 metric 200" 0
+	check_err $? "highest metric offloaded when should not"
+
+	log_test "IPv6 multipath route add"
+
+	ip -6 route flush 2001:db8:3::/64
+}
+
+ipv6_route_add()
+{
+	ipv6_route_add_prefix
+	ipv6_route_add_mpath
+}
+
+ipv6_route_replace()
+{
+	RET=0
+
+	# Replace prefix route with prefix route.
+	ip -6 route add 2001:db8:3::/64 metric 100 dev $spine_p1
+	ipv6_offload_check "2001:db8:3::/64 metric 100" 1
+	check_err $? "prefix route not offloaded when should"
+	ip -6 route replace 2001:db8:3::/64 metric 100 dev $spine_p2
+	ipv6_offload_check "2001:db8:3::/64 metric 100" 1
+	check_err $? "prefix route not offloaded after replace"
+
+	# Replace prefix route with multipath route.
+	ip -6 route replace 2001:db8:3::/64 metric 100 \
+		nexthop via 2001:db8:1::2 dev $spine_p1 \
+		nexthop via 2001:db8:2::2 dev $spine_p2
+	ipv6_offload_check "2001:db8:3::/64 metric 100" 2
+	check_err $? "multipath route not offloaded after replace"
+
+	# Replace multipath route with prefix route. A prefix route cannot
+	# replace a multipath route, so it is appended.
+	ip -6 route replace 2001:db8:3::/64 metric 100 dev $spine_p1
+	ipv6_offload_check "2001:db8:3::/64 metric 100 dev $spine_p1" 0
+	check_err $? "prefix route offloaded after 'replacing' multipath route"
+	ipv6_offload_check "2001:db8:3::/64 metric 100" 2
+	check_err $? "multipath route not offloaded after being 'replaced' by prefix route"
+
+	# Replace multipath route with multipath route.
+	ip -6 route replace 2001:db8:3::/64 metric 100 \
+		nexthop via 2001:db8:1::3 dev $spine_p1 \
+		nexthop via 2001:db8:2::3 dev $spine_p2
+	ipv6_offload_check "2001:db8:3::/64 metric 100" 2
+	check_err $? "multipath route not offloaded after replacing multipath route"
+
+	# Replace a non-existing multipath route with a multipath route and
+	# check that it is appended and not offloaded.
+	ip -6 route replace 2001:db8:3::/64 metric 200 \
+		nexthop via 2001:db8:1::3 dev $spine_p1 \
+		nexthop via 2001:db8:2::3 dev $spine_p2
+	ipv6_offload_check "2001:db8:3::/64 metric 100" 2
+	check_err $? "multipath route not offloaded after non-existing route was 'replaced'"
+	ipv6_offload_check "2001:db8:3::/64 metric 200" 0
+	check_err $? "multipath route offloaded after 'replacing' non-existing route"
+
+	log_test "IPv6 route replace"
+
+	ip -6 route flush 2001:db8:3::/64
+}
+
+ipv6_route_nexthop_group_share()
+{
+	RET=0
+
+	# The driver consolidates identical nexthop groups in order to reduce
+	# the resource usage in its adjacency table. Check that the deletion
+	# of one multipath route using the group does not affect the other.
+	ip -6 route add 2001:db8:3::/64 \
+		nexthop via 2001:db8:1::2 dev $spine_p1 \
+		nexthop via 2001:db8:2::2 dev $spine_p2
+	ip -6 route add 2001:db8:4::/64 \
+		nexthop via 2001:db8:1::2 dev $spine_p1 \
+		nexthop via 2001:db8:2::2 dev $spine_p2
+	ipv6_offload_check "2001:db8:3::/64" 2
+	check_err $? "multipath route not offloaded when should"
+	ipv6_offload_check "2001:db8:4::/64" 2
+	check_err $? "multipath route not offloaded when should"
+	ip -6 route del 2001:db8:3::/64
+	ipv6_offload_check "2001:db8:4::/64" 2
+	check_err $? "multipath route not offloaded after deletion of route sharing the nexthop group"
+
+	# Check that after unsharing a nexthop group the routes are still
+	# marked as offloaded.
+	ip -6 route add 2001:db8:3::/64 \
+		nexthop via 2001:db8:1::2 dev $spine_p1 \
+		nexthop via 2001:db8:2::2 dev $spine_p2
+	ip -6 route del 2001:db8:4::/64 \
+		nexthop via 2001:db8:1::2 dev $spine_p1
+	ipv6_offload_check "2001:db8:4::/64" 1
+	check_err $? "singlepath route not offloaded after unsharing the nexthop group"
+	ipv6_offload_check "2001:db8:3::/64" 2
+	check_err $? "multipath route not offloaded after unsharing the nexthop group"
+
+	log_test "IPv6 nexthop group sharing"
+
+	ip -6 route flush 2001:db8:3::/64
+	ip -6 route flush 2001:db8:4::/64
+}
+
+ipv6_route_rate()
+{
+	local batch_dir=$(mktemp -d)
+	local num_rts=$((40 * 1024))
+	local num_nhs=16
+	local total
+	local start
+	local diff
+	local end
+	local nhs
+	local i
+
+	RET=0
+
+	# Prepare 40K /64 multipath routes with 16 nexthops each and check how
+	# long it takes to add them. A limit of 60 seconds is set. It is much
+	# higher than insertion should take and meant to flag a serious
+	# regression.
+	total=$((nums_nhs * num_rts))
+
+	for i in $(seq 1 $num_nhs); do
+		ip -6 address add 2001:db8:1::10:$i/128 dev $tor1_p1
+		nexthops+=" nexthop via 2001:db8:1::10:$i dev $spine_p1"
+	done
+
+	for i in $(seq 1 $num_rts); do
+		echo "route add 2001:db8:8:$(printf "%x" $i)::/64$nexthops" \
+			>> $batch_dir/add.batch
+		echo "route del 2001:db8:8:$(printf "%x" $i)::/64$nexthops" \
+			>> $batch_dir/del.batch
+	done
+
+	start=$(date +%s.%N)
+
+	ip -batch $batch_dir/add.batch
+	count=$(ip -6 route show | grep offload | wc -l)
+	while [ $count -lt $total ]; do
+		sleep .01
+		count=$(ip -6 route show | grep offload | wc -l)
+	done
+
+	end=$(date +%s.%N)
+
+	diff=$(echo "$end - $start" | bc -l)
+	test "$(echo "$diff > 60" | bc -l)" -eq 0
+	check_err $? "route insertion took too long"
+	log_info "inserted $num_rts routes in $diff seconds"
+
+	log_test "IPv6 routes insertion rate"
+
+	ip -batch $batch_dir/del.batch
+	for i in $(seq 1 $num_nhs); do
+		ip -6 address del 2001:db8:1::10:$i/128 dev $tor1_p1
+	done
+	rm -rf $batch_dir
+}
+
+setup_prepare()
+{
+	spine_p1=${NETIFS[p1]}
+	tor1_p1=${NETIFS[p2]}
+
+	spine_p2=${NETIFS[p3]}
+	tor2_p1=${NETIFS[p4]}
+
+	vrf_prepare
+	forwarding_enable
+
+	tor1_create
+	tor2_create
+	spine_create
+}
+
+cleanup()
+{
+	pre_cleanup
+
+	spine_destroy
+	tor2_destroy
+	tor1_destroy
+
+	forwarding_restore
+	vrf_cleanup
+}
+
+trap cleanup EXIT
+
+setup_prepare
+setup_wait
+
+tests_run
+
+exit $EXIT_STATUS