* [PATCH net-next V2 1/5] net/mlx5: E-switch, add match-metadata allocator for nested VF/SF eswitches
2026-11-20 7:59 [PATCH net-next V2 0/5] net/mlx5: LAG over VFs/SFs (virtual LAG) Tariq Toukan
@ 2026-11-20 7:59 ` Tariq Toukan
2026-11-20 7:59 ` [PATCH net-next V2 2/5] net/mlx5: E-switch, match master egress rule by vport metadata Tariq Toukan
` (3 subsequent siblings)
4 siblings, 0 replies; 6+ messages in thread
From: Tariq Toukan @ 2026-11-20 7:59 UTC (permalink / raw)
To: Andrew Lunn, David S. Miller, Eric Dumazet, Jakub Kicinski,
netdev, Paolo Abeni
Cc: Akiva Goldberger, Gal Pressman, Jason Gunthorpe, Leon Romanovsky,
linux-rdma, Mark Bloch, Moshe Shemesh, Or Har-Toov,
Saeed Mahameed, Shay Drory, Tariq Toukan
From: Shay Drory <shayd@nvidia.com>
The eswitch source-port match metadata is laid out as 4 bits of PFNUM
and 12 bits of a unique id (ESW_SOURCE_PORT_METADATA_BITS). PFNUM is the
native_port_num reported in HCA_CAP, so all VFs/SFs of the same PF share
a single PFNUM. Each VF/SF eswitch allocates its id from its own
vport_metadata_ida, restarting from the lowest value, so two VFs/SFs of
the same PF can end up with identical metadata. Once their eswitches
share an FDB (virt LAG), that metadata no longer uniquely identifies the
source port.
Add a dedicated allocator for nested (virt LAG) eswitches that replaces
the PFNUM field with the VF/SF's slot in the LAG (its xarray index in
ldev->fns). That index is unique per function within the LAG, so
combined with a per-VF/SF id it yields unique metadata regardless of
which PF the VFs/SFs belong to. The split is 8 bits for the virt LAG
index - which fits the U8_MAX cap on the LAG xarray - and the remaining
8 bits for the id; the two still sum to ESW_SOURCE_PORT_METADATA_BITS,
so nothing downstream of the match changes. Describe the nested layout
next to the reg_c_0 layout in include/linux/mlx5/eswitch.h.
The PFNUM value 0xf is reserved for internal port offload, which stamps
it into reg_c_0 and matches it with the same mask used for vport
metadata. That offload is gated on switchdev mode only, so it can be
used on a nested eswitch as well. Cap the LAG index accordingly.
Select the nested layout for a non-PF eswitch, unless it is virt LAG
capable but not a LAG member, e.g. one refused as a late joiner and left
standalone. Such a device keeps the PFNUM layout: it does not pair with
the eswitches of the LAG members, so its metadata is never compared
against theirs. PFs, which never report the virt LAG capabilities, keep
the PFNUM layout.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Moshe Shemesh <moshe@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
---
.../mellanox/mlx5/core/eswitch_offloads.c | 48 +++++++++++++++++++
.../net/ethernet/mellanox/mlx5/core/lag/lag.c | 17 +++++++
.../ethernet/mellanox/mlx5/core/mlx5_core.h | 1 +
include/linux/mlx5/eswitch.h | 4 ++
4 files changed, 70 insertions(+)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c b/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
index 1323114ce7f9..3c1355f91090 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
@@ -3650,6 +3650,48 @@ static u32 mlx5_esw_match_metadata_reserved(struct mlx5_eswitch *esw)
return MLX5_ESW_METADATA_RSVD_UPLINK;
}
+/* Nested (virt LAG) source-port metadata layout: the upper ESW_NESTED_IDX_BITS
+ * hold the VF/SF's slot in the LAG and the lower ESW_NESTED_VPORT_BITS hold a
+ * vport id unique within its eswitch. The two fields still sum to
+ * ESW_SOURCE_PORT_METADATA_BITS. The LAG slot (0-254, per the xa_alloc() limit
+ * in mlx5_ldev_add_mdev()) fits in ESW_NESTED_IDX_BITS; slots 0xf0 and up would
+ * drive the PFNUM bits to the 0xf reserved by internal port offload, so only
+ * 0-239 are used.
+ */
+#define ESW_NESTED_VPORT_BITS 8
+#define ESW_NESTED_IDX_BITS (ESW_SOURCE_PORT_METADATA_BITS - ESW_NESTED_VPORT_BITS)
+
+static u32 mlx5_nested_esw_match_metadata_alloc(struct mlx5_eswitch *esw)
+{
+ u32 vport_end_ida = (1 << ESW_NESTED_VPORT_BITS) - 1;
+ u32 max_idx = (1 << ESW_NESTED_IDX_BITS) -
+ (1 << (ESW_NESTED_IDX_BITS - ESW_PFNUM_BITS)) - 1;
+ int idx;
+ int id;
+
+ idx = mlx5_lag_get_member_idx(esw->dev);
+ if (idx < 0 || idx > max_idx)
+ return 0;
+
+ /* Use only non-zero id's (2-255). */
+ id = ida_alloc_range(&esw->offloads.vport_metadata_ida,
+ MLX5_ESW_METADATA_RSVD_UPLINK + 1,
+ vport_end_ida, GFP_KERNEL);
+ if (id < 0)
+ return 0;
+ id = (idx << ESW_NESTED_VPORT_BITS) | id;
+ return id;
+}
+
+static void mlx5_nested_esw_match_metadata_free(struct mlx5_eswitch *esw,
+ u32 metadata)
+{
+ u32 vport_bit_mask = (1 << ESW_NESTED_VPORT_BITS) - 1;
+
+ /* Metadata holds only ESW_NESTED_VPORT_BITS of actual ida id. */
+ ida_free(&esw->offloads.vport_metadata_ida, metadata & vport_bit_mask);
+}
+
u32 mlx5_esw_match_metadata_alloc(struct mlx5_eswitch *esw)
{
u32 vport_end_ida = (1 << ESW_VPORT_BITS) - 1;
@@ -3658,6 +3700,9 @@ u32 mlx5_esw_match_metadata_alloc(struct mlx5_eswitch *esw)
int pf_num;
int id;
+ if (!mlx5_core_is_pf(esw->dev) && !mlx5_virt_lag_nonmember(esw->dev))
+ return mlx5_nested_esw_match_metadata_alloc(esw);
+
/* Only 4 bits of pf_num */
pf_num = mlx5_sd_pf_num_get(esw->dev);
if (pf_num < 0 || pf_num > max_pf_num)
@@ -3678,6 +3723,9 @@ void mlx5_esw_match_metadata_free(struct mlx5_eswitch *esw, u32 metadata)
{
u32 vport_bit_mask = (1 << ESW_VPORT_BITS) - 1;
+ if (!mlx5_core_is_pf(esw->dev) && !mlx5_virt_lag_nonmember(esw->dev))
+ return mlx5_nested_esw_match_metadata_free(esw, metadata);
+
/* Metadata contains only 12 bits of actual ida id */
ida_free(&esw->offloads.vport_metadata_ida, metadata & vport_bit_mask);
}
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c b/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c
index abcab76fced0..193fbe47fef1 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c
@@ -516,6 +516,23 @@ int mlx5_lag_get_dev_seq(struct mlx5_core_dev *dev)
}
EXPORT_SYMBOL(mlx5_lag_get_dev_seq);
+int mlx5_lag_get_member_idx(struct mlx5_core_dev *dev)
+{
+ struct mlx5_lag *ldev = mlx5_lag_dev(dev);
+ struct lag_fn *fn;
+ int idx;
+
+ if (!ldev)
+ return -ENOENT;
+
+ xa_lock(&ldev->fns);
+ fn = mlx5_lag_fn_by_dev(ldev, dev);
+ idx = fn ? fn->idx : -ENOENT;
+ xa_unlock(&ldev->fns);
+
+ return idx;
+}
+
/* seq 0 = master, then all remaining devices */
static int mlx5_lag_get_dev_index_by_seq_all(struct mlx5_lag *ldev, int seq)
{
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/mlx5_core.h b/drivers/net/ethernet/mellanox/mlx5/core/mlx5_core.h
index d84d0424d04e..f5f105762d88 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/mlx5_core.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/mlx5_core.h
@@ -275,6 +275,7 @@ void mlx5_lag_add_mdev(struct mlx5_core_dev *dev);
void mlx5_lag_remove_mdev(struct mlx5_core_dev *dev);
void mlx5_lag_disable_change(struct mlx5_core_dev *dev);
void mlx5_lag_enable_change(struct mlx5_core_dev *dev);
+int mlx5_lag_get_member_idx(struct mlx5_core_dev *dev);
int mlx5_events_init(struct mlx5_core_dev *dev);
void mlx5_events_cleanup(struct mlx5_core_dev *dev);
diff --git a/include/linux/mlx5/eswitch.h b/include/linux/mlx5/eswitch.h
index 03d3620141c8..06d7d634de0d 100644
--- a/include/linux/mlx5/eswitch.h
+++ b/include/linux/mlx5/eswitch.h
@@ -95,6 +95,10 @@ bool mlx5_eswitch_vport_match_metadata_enabled(const struct mlx5_eswitch *esw);
* unique non-zero vport id (range 1-4095). The rest (lowest 16 bits) is left
* for user data objects managed by a common mapping context.
* PFNUM + VPORT comprise the SOURCE_PORT matching.
+ *
+ * A nested (virt LAG) VF/SF eswitch splits the SOURCE_PORT bits differently:
+ * the highest 8 bits are its slot in the LAG (range 0-239) and the next 8
+ * bits are a unique non-zero vport id (range 2-255).
*/
#define ESW_VPORT_BITS 12
#define ESW_PFNUM_BITS 4
--
2.44.0
^ permalink raw reply related [flat|nested] 6+ messages in thread* [PATCH net-next V2 2/5] net/mlx5: E-switch, match master egress rule by vport metadata
2026-11-20 7:59 [PATCH net-next V2 0/5] net/mlx5: LAG over VFs/SFs (virtual LAG) Tariq Toukan
2026-11-20 7:59 ` [PATCH net-next V2 1/5] net/mlx5: E-switch, add match-metadata allocator for nested VF/SF eswitches Tariq Toukan
@ 2026-11-20 7:59 ` Tariq Toukan
2026-11-20 7:59 ` [PATCH net-next V2 3/5] net/mlx5: E-switch, drop master egress rule under a SW vport LAG Tariq Toukan
` (2 subsequent siblings)
4 siblings, 0 replies; 6+ messages in thread
From: Tariq Toukan @ 2026-11-20 7:59 UTC (permalink / raw)
To: Andrew Lunn, David S. Miller, Eric Dumazet, Jakub Kicinski,
netdev, Paolo Abeni
Cc: Akiva Goldberger, Gal Pressman, Jason Gunthorpe, Leon Romanovsky,
linux-rdma, Mark Bloch, Moshe Shemesh, Or Har-Toov,
Saeed Mahameed, Shay Drory, Tariq Toukan
From: Shay Drory <shayd@nvidia.com>
The master egress-ACL bounce group and rule matched the slave uplink by
source_port + source_eswitch_owner_vhca_id. Virt LAG master cannot match
that way; it must match by vport metadata (reg_c_0).
Virt LAG only supports MPESW, where every uplink has a unique metadata.
Hence, use metadata matching when LAG is virt LAG, and keep the
source_port + source_eswitch_owner_vhca_id match otherwise.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Akiva Goldberger <agoldberger@nvidia.com>
Reviewed-by: Moshe Shemesh <moshe@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
---
.../mellanox/mlx5/core/eswitch_offloads.c | 78 +++++++++++++------
1 file changed, 55 insertions(+), 23 deletions(-)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c b/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
index 3c1355f91090..57f58e79478e 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
@@ -2999,6 +2999,26 @@ static int esw_set_slave_root_fdb(struct mlx5_core_dev *master,
return err;
}
+static void mlx5_esw_set_spec_source_port_vhca(struct mlx5_flow_spec *spec,
+ u16 vport, u16 vhca_id)
+{
+ void *misc;
+
+ spec->match_criteria_enable = MLX5_MATCH_MISC_PARAMETERS;
+
+ misc = MLX5_ADDR_OF(fte_match_param, spec->match_value,
+ misc_parameters);
+ MLX5_SET(fte_match_set_misc, misc, source_port, vport);
+ MLX5_SET(fte_match_set_misc, misc, source_eswitch_owner_vhca_id,
+ vhca_id);
+
+ misc = MLX5_ADDR_OF(fte_match_param, spec->match_criteria,
+ misc_parameters);
+ MLX5_SET_TO_ONES(fte_match_set_misc, misc, source_port);
+ MLX5_SET_TO_ONES(fte_match_set_misc, misc,
+ source_eswitch_owner_vhca_id);
+}
+
static int __esw_set_master_egress_rule(struct mlx5_core_dev *master,
struct mlx5_core_dev *slave,
struct mlx5_vport *vport,
@@ -3010,22 +3030,22 @@ static int __esw_set_master_egress_rule(struct mlx5_core_dev *master,
struct mlx5_flow_act flow_act = {};
struct mlx5_flow_spec *spec;
int err = 0;
- void *misc;
+
+ if (mlx5_virt_lag_is_supported(master) &&
+ !mlx5_eswitch_vport_match_metadata_enabled(slave->priv.eswitch))
+ return -EINVAL;
spec = kvzalloc_obj(*spec);
if (!spec)
return -ENOMEM;
- spec->match_criteria_enable = MLX5_MATCH_MISC_PARAMETERS;
- misc = MLX5_ADDR_OF(fte_match_param, spec->match_value,
- misc_parameters);
- MLX5_SET(fte_match_set_misc, misc, source_port, MLX5_VPORT_UPLINK);
- MLX5_SET(fte_match_set_misc, misc, source_eswitch_owner_vhca_id, slave_index);
-
- misc = MLX5_ADDR_OF(fte_match_param, spec->match_criteria, misc_parameters);
- MLX5_SET_TO_ONES(fte_match_set_misc, misc, source_port);
- MLX5_SET_TO_ONES(fte_match_set_misc, misc,
- source_eswitch_owner_vhca_id);
+ if (mlx5_virt_lag_is_supported(master) &&
+ mlx5_eswitch_vport_match_metadata_enabled(slave->priv.eswitch))
+ mlx5_esw_set_spec_source_port(slave->priv.eswitch,
+ MLX5_VPORT_UPLINK, spec);
+ else
+ mlx5_esw_set_spec_source_port_vhca(spec, MLX5_VPORT_UPLINK,
+ slave_index);
flow_act.action = MLX5_FLOW_CONTEXT_ACTION_FWD_DEST;
dest.type = MLX5_FLOW_DESTINATION_TYPE_VPORT;
@@ -3178,6 +3198,21 @@ static void esw_unset_slave_egress_rule(struct mlx5_core_dev *master,
esw_slave_egress_destroy_resources(slave_vport);
}
+static void mlx5_esw_set_flow_group_source_port_vhca(u32 *flow_group_in)
+{
+ void *match_criteria = MLX5_ADDR_OF(create_flow_group_in,
+ flow_group_in, match_criteria);
+
+ MLX5_SET_TO_ONES(fte_match_param, match_criteria,
+ misc_parameters.source_port);
+ MLX5_SET_TO_ONES(fte_match_param, match_criteria,
+ misc_parameters.source_eswitch_owner_vhca_id);
+ MLX5_SET(create_flow_group_in, flow_group_in,
+ match_criteria_enable, MLX5_MATCH_MISC_PARAMETERS);
+ MLX5_SET(create_flow_group_in, flow_group_in,
+ source_eswitch_owner_vhca_id_valid, 1);
+}
+
static int esw_master_egress_create_resources(struct mlx5_eswitch *esw,
struct mlx5_flow_namespace *egress_ns,
struct mlx5_vport *vport, size_t count)
@@ -3188,13 +3223,16 @@ static int esw_master_egress_create_resources(struct mlx5_eswitch *esw,
};
struct mlx5_flow_table *acl;
struct mlx5_flow_group *g;
- void *match_criteria;
u32 *flow_group_in;
int err;
if (vport->egress.acl)
return 0;
+ if (mlx5_virt_lag_is_supported(esw->dev) &&
+ !mlx5_eswitch_vport_match_metadata_enabled(esw))
+ return -EINVAL;
+
flow_group_in = kvzalloc(inlen, GFP_KERNEL);
if (!flow_group_in)
return -ENOMEM;
@@ -3208,17 +3246,11 @@ static int esw_master_egress_create_resources(struct mlx5_eswitch *esw,
goto out;
}
- match_criteria = MLX5_ADDR_OF(create_flow_group_in, flow_group_in,
- match_criteria);
- MLX5_SET_TO_ONES(fte_match_param, match_criteria,
- misc_parameters.source_port);
- MLX5_SET_TO_ONES(fte_match_param, match_criteria,
- misc_parameters.source_eswitch_owner_vhca_id);
- MLX5_SET(create_flow_group_in, flow_group_in, match_criteria_enable,
- MLX5_MATCH_MISC_PARAMETERS);
-
- MLX5_SET(create_flow_group_in, flow_group_in,
- source_eswitch_owner_vhca_id_valid, 1);
+ if (mlx5_virt_lag_is_supported(esw->dev) &&
+ mlx5_eswitch_vport_match_metadata_enabled(esw))
+ mlx5_esw_set_flow_group_source_port(esw, flow_group_in, 0);
+ else
+ mlx5_esw_set_flow_group_source_port_vhca(flow_group_in);
MLX5_SET(create_flow_group_in, flow_group_in, start_flow_index, 0);
MLX5_SET(create_flow_group_in, flow_group_in, end_flow_index, count);
--
2.44.0
^ permalink raw reply related [flat|nested] 6+ messages in thread* [PATCH net-next V2 3/5] net/mlx5: E-switch, drop master egress rule under a SW vport LAG
2026-11-20 7:59 [PATCH net-next V2 0/5] net/mlx5: LAG over VFs/SFs (virtual LAG) Tariq Toukan
2026-11-20 7:59 ` [PATCH net-next V2 1/5] net/mlx5: E-switch, add match-metadata allocator for nested VF/SF eswitches Tariq Toukan
2026-11-20 7:59 ` [PATCH net-next V2 2/5] net/mlx5: E-switch, match master egress rule by vport metadata Tariq Toukan
@ 2026-11-20 7:59 ` Tariq Toukan
2026-11-20 7:59 ` [PATCH net-next V2 4/5] net/mlx5e: Register eswitch-manager VF/SF self-port with switch-id Tariq Toukan
2026-11-20 7:59 ` [PATCH net-next V2 5/5] net/mlx5: LAG, enable LAG over functions that do not represent a physical port Tariq Toukan
4 siblings, 0 replies; 6+ messages in thread
From: Tariq Toukan @ 2026-11-20 7:59 UTC (permalink / raw)
To: Andrew Lunn, David S. Miller, Eric Dumazet, Jakub Kicinski,
netdev, Paolo Abeni
Cc: Akiva Goldberger, Gal Pressman, Jason Gunthorpe, Leon Romanovsky,
linux-rdma, Mark Bloch, Moshe Shemesh, Or Har-Toov,
Saeed Mahameed, Shay Drory, Tariq Toukan
From: Shay Drory <shayd@nvidia.com>
VPORT_LAG steers all traffic of a LAG slave to the LAG master. The
eswitch master egress rule does the opposite direction: it bounces
traffic that reached the master's manager vport into the slave eswitch.
With a FW vport LAG the FW steers that bounced packet to the LAG master,
so the two coexist.
The software vport LAG emulation cannot. It is a catch-all rule in the
slave eswitch's egress ACL that forwards to the master's manager vport,
while the master egress rule redirects the packet after the FDB - past
the point where the slave egress rule is evaluated - so the bounced
packet never reaches it.
The two therefore cannot coexist: drop the master egress rule when the
software vport LAG is installed, and restore it when it is removed. Both
rules are keyed by the peer vhca_id, so each slave swaps its own pair
independently. Restoring needs the ACL size, so pass ldev->max_fns down
to mlx5_eswitch_offloads_vport_lag_del_one(), as the single FDB path
already does.
Also, release the MPESW metadata after the shared FDB is destroyed,
mirroring the creation order. Releasing it first makes all slaves fall
back to the same default uplink metadata, so the restored master egress
rules collide.
Note that vport LAG on its own only gets traffic to the LAG master; the
distribution to the other slaves esw managers, which was managed by
master egress rules, is done by the LAG demux.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Moshe Shemesh <moshe@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
---
.../net/ethernet/mellanox/mlx5/core/eswitch.h | 3 ++-
.../mellanox/mlx5/core/eswitch_offloads.c | 24 +++++++++++++++++--
.../ethernet/mellanox/mlx5/core/lag/mpesw.c | 2 +-
.../mellanox/mlx5/core/lag/shared_fdb.c | 6 +++--
4 files changed, 29 insertions(+), 6 deletions(-)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/eswitch.h b/drivers/net/ethernet/mellanox/mlx5/core/eswitch.h
index 48eb6acc1e87..bcf053b0a38d 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/eswitch.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/eswitch.h
@@ -958,7 +958,8 @@ bool mlx5_esw_offloads_controller_valid(const struct mlx5_eswitch *esw, u32 cont
int mlx5_eswitch_offloads_vport_lag_add_one(struct mlx5_eswitch *master_esw,
struct mlx5_eswitch *slave_esw);
void mlx5_eswitch_offloads_vport_lag_del_one(struct mlx5_eswitch *master_esw,
- struct mlx5_eswitch *slave_esw);
+ struct mlx5_eswitch *slave_esw,
+ int max_slaves);
int mlx5_eswitch_offloads_single_fdb_add_one(struct mlx5_eswitch *master_esw,
struct mlx5_eswitch *slave_esw, int max_slaves);
void mlx5_eswitch_offloads_single_fdb_del_one(struct mlx5_eswitch *master_esw,
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c b/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
index 57f58e79478e..2072921c1fff 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
@@ -3380,12 +3380,32 @@ void mlx5_eswitch_offloads_single_fdb_del_one(struct mlx5_eswitch *master_esw,
int mlx5_eswitch_offloads_vport_lag_add_one(struct mlx5_eswitch *master_esw,
struct mlx5_eswitch *slave_esw)
{
- return esw_set_slave_egress_rule(master_esw->dev, slave_esw->dev);
+ int err;
+
+ err = esw_set_slave_egress_rule(master_esw->dev, slave_esw->dev);
+ if (err)
+ return err;
+
+ esw_unset_master_egress_rule(master_esw->dev, slave_esw->dev);
+
+ return 0;
}
void mlx5_eswitch_offloads_vport_lag_del_one(struct mlx5_eswitch *master_esw,
- struct mlx5_eswitch *slave_esw)
+ struct mlx5_eswitch *slave_esw,
+ int max_slaves)
{
+ int err;
+
+ if (mlx5_sd_is_primary(slave_esw->dev)) {
+ err = esw_set_master_egress_rule(master_esw->dev,
+ slave_esw->dev, max_slaves);
+ if (err)
+ esw_warn(master_esw->dev,
+ "Failed to restore master egress rule (%d)\n",
+ err);
+ }
+
esw_unset_slave_egress_rule(master_esw->dev, slave_esw->dev);
}
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/lag/mpesw.c b/drivers/net/ethernet/mellanox/mlx5/core/lag/mpesw.c
index 8a867a04916a..2d3e61541d8a 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/lag/mpesw.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/lag/mpesw.c
@@ -208,10 +208,10 @@ void mlx5_lag_disable_mpesw(struct mlx5_lag *ldev)
if (ldev->mode != MLX5_LAG_MODE_MPESW)
return;
- mlx5_mpesw_metadata_cleanup(ldev);
mlx5_lag_shared_fdb_destroy(ldev, MLX5_LAG_FILTER_ALL);
if (mlx5_lag_has_sd_group(ldev))
mlx5_mpesw_restore_sd_fdb(ldev);
+ mlx5_mpesw_metadata_cleanup(ldev);
/* Pair with the master mark taken at MPESW create for VFs. */
if (ldev->virt_lag)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/lag/shared_fdb.c b/drivers/net/ethernet/mellanox/mlx5/core/lag/shared_fdb.c
index 404278a15ef7..90209e4b99f7 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/lag/shared_fdb.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/lag/shared_fdb.c
@@ -126,7 +126,8 @@ int mlx5_lag_create_vport_lag(struct mlx5_lag *ldev, u32 group_id)
if (j == master_idx)
continue;
slave_esw = mlx5_lag_fn(ldev, j)->dev->priv.eswitch;
- mlx5_eswitch_offloads_vport_lag_del_one(master_esw, slave_esw);
+ mlx5_eswitch_offloads_vport_lag_del_one(master_esw, slave_esw,
+ ldev->max_fns);
}
return err;
}
@@ -153,7 +154,8 @@ int mlx5_lag_destroy_vport_lag(struct mlx5_lag *ldev, u32 group_id)
continue;
dev = mlx5_lag_fn(ldev, i)->dev;
mlx5_eswitch_offloads_vport_lag_del_one(master_esw,
- dev->priv.eswitch);
+ dev->priv.eswitch,
+ ldev->max_fns);
}
return 0;
}
--
2.44.0
^ permalink raw reply related [flat|nested] 6+ messages in thread* [PATCH net-next V2 4/5] net/mlx5e: Register eswitch-manager VF/SF self-port with switch-id
2026-11-20 7:59 [PATCH net-next V2 0/5] net/mlx5: LAG over VFs/SFs (virtual LAG) Tariq Toukan
` (2 preceding siblings ...)
2026-11-20 7:59 ` [PATCH net-next V2 3/5] net/mlx5: E-switch, drop master egress rule under a SW vport LAG Tariq Toukan
@ 2026-11-20 7:59 ` Tariq Toukan
2026-11-20 7:59 ` [PATCH net-next V2 5/5] net/mlx5: LAG, enable LAG over functions that do not represent a physical port Tariq Toukan
4 siblings, 0 replies; 6+ messages in thread
From: Tariq Toukan @ 2026-11-20 7:59 UTC (permalink / raw)
To: Andrew Lunn, David S. Miller, Eric Dumazet, Jakub Kicinski,
netdev, Paolo Abeni
Cc: Akiva Goldberger, Gal Pressman, Jason Gunthorpe, Leon Romanovsky,
linux-rdma, Mark Bloch, Moshe Shemesh, Or Har-Toov,
Saeed Mahameed, Shay Drory, Tariq Toukan
From: Shay Drory <shayd@nvidia.com>
A non-PF function's self-port (its own uplink netdev) was always
registered with DEVLINK_PORT_FLAVOUR_VIRTUAL, which doesn't carries
switch_id. Consequently a function that acts as an eswitch manager - a
VF (or SF) running its own nested eswitch - exposed no phys_switch_id on
its uplink netdev, so it could not be grouped with the representors it
manages.
When the function is an eswitch manager, register its self-port with the
switch_id taken from the system image guid. The uplink netdev is already
bound to this devlink port, so phys_switch_id is derived by the core,
and the managed representors - which stamp the same switch_id - group
with it.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Moshe Shemesh <moshe@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
---
drivers/net/ethernet/mellanox/mlx5/core/en/devlink.c | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/devlink.c b/drivers/net/ethernet/mellanox/mlx5/core/en/devlink.c
index 8818f65d1fbc..79adb1ecc450 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en/devlink.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/devlink.c
@@ -63,6 +63,11 @@ int mlx5e_devlink_port_register(struct mlx5e_dev *mlx5e_dev,
dl_port_index = mlx5_esw_vport_to_devlink_port_index(mdev,
MLX5_VPORT_UPLINK);
} else {
+ if (MLX5_ESWITCH_MANAGER(mdev)) {
+ mlx5e_devlink_get_port_parent_id(mdev, &ppid);
+ memcpy(attrs.switch_id.id, ppid.id, ppid.id_len);
+ attrs.switch_id.id_len = ppid.id_len;
+ }
attrs.flavour = DEVLINK_PORT_FLAVOUR_VIRTUAL;
dl_port_index = mlx5_esw_vport_to_devlink_port_index(mdev, 0);
}
--
2.44.0
^ permalink raw reply related [flat|nested] 6+ messages in thread* [PATCH net-next V2 5/5] net/mlx5: LAG, enable LAG over functions that do not represent a physical port
2026-11-20 7:59 [PATCH net-next V2 0/5] net/mlx5: LAG over VFs/SFs (virtual LAG) Tariq Toukan
` (3 preceding siblings ...)
2026-11-20 7:59 ` [PATCH net-next V2 4/5] net/mlx5e: Register eswitch-manager VF/SF self-port with switch-id Tariq Toukan
@ 2026-11-20 7:59 ` Tariq Toukan
4 siblings, 0 replies; 6+ messages in thread
From: Tariq Toukan @ 2026-11-20 7:59 UTC (permalink / raw)
To: Andrew Lunn, David S. Miller, Eric Dumazet, Jakub Kicinski,
netdev, Paolo Abeni
Cc: Akiva Goldberger, Gal Pressman, Jason Gunthorpe, Leon Romanovsky,
linux-rdma, Mark Bloch, Moshe Shemesh, Or Har-Toov,
Saeed Mahameed, Shay Drory, Tariq Toukan
From: Shay Drory <shayd@nvidia.com>
LAG membership assumes every member represents a physical port.
A nested LAG of VFs/SFs does not fit that assumption; its member
functions sit behind a multi_uplink_eswitch and each reports
num_lag_ports < 2, so they do not represent a physical port.
Detect this with mlx5_virt_lag_is_supported(): let such a function join
the LAG in mlx5_lag_add_mdev(), and skip the MLX5_LAG_XA_MARK_PORT mark
and the PORT_CHANGE event notifier in mlx5_ldev_add_mdev(), since it is
not a physical port.
These function only support MPESW, and netdevs addition to LAG is only
needed for linux bond. Hence, don't allow such a lag to add netdevs to
itself. For the same reason, don't register the LAG netdev notifier for
it. Multipath LAG is not supported either, so skip mlx5_lag_mp_init()
and do not register the FIB notifier for it.
Such a function may report num_lag_ports == 0, which ldev->ports keeps
as is. Report at least one port from mlx5_lag_get_num_ports(), so user
would avoid divide by 0.
Such a function selects the uplink through port_select_eswitch, not
through a TX port affinity. Yet once its LAG is active, RDMA sets
lag_active, and mlx5_ib_lag_should_assign_affinity() then assigns a TX
port affinity to every QP, bypassing the num_lag_ports > 1 and
lag_tx_port_affinity capability checks. Make
mlx5_ib_lag_should_assign_affinity() return false for such a function.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Moshe Shemesh <moshe@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
---
drivers/infiniband/hw/mlx5/mlx5_ib.h | 3 ++
.../net/ethernet/mellanox/mlx5/core/lag/lag.c | 32 +++++++++++++------
.../ethernet/mellanox/mlx5/core/lag/mpesw.c | 2 +-
3 files changed, 26 insertions(+), 11 deletions(-)
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 4a6ee14ffd7d..494b7036321b 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -1612,6 +1612,9 @@ static inline void mlx5r_deref_wait_odp_mkey(struct mlx5_ib_mkey *mmkey)
static inline bool mlx5_ib_lag_should_assign_affinity(struct mlx5_ib_dev *dev)
{
+ if (dev->lag_active && mlx5_virt_lag_is_supported(dev->mdev))
+ return 0;
+
/*
* If the driver is in hash mode and the port_select_flow_table_bypass cap
* is supported, it means that the driver no longer needs to assign the port
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c b/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c
index 193fbe47fef1..2cd2319f9318 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c
@@ -366,6 +366,7 @@ static struct mlx5_lag *mlx5_lag_dev_alloc(struct mlx5_core_dev *dev)
ldev->ports = MLX5_CAP_GEN(dev, num_lag_ports);
ldev->buckets = 1;
+ ldev->virt_lag = mlx5_virt_lag_is_supported(dev);
ldev->v2p_map = kcalloc(ldev->ports * MLX5_LAG_MAX_HASH_BUCKETS,
sizeof(*ldev->v2p_map), GFP_KERNEL);
if (!ldev->v2p_map)
@@ -381,7 +382,7 @@ static struct mlx5_lag *mlx5_lag_dev_alloc(struct mlx5_core_dev *dev)
INIT_DELAYED_WORK(&ldev->bond_work, mlx5_do_bond_work);
INIT_WORK(&ldev->speed_update_work, mlx5_mpesw_speed_update_work);
- if (!mlx5_sd_is_supported(dev)) {
+ if (!mlx5_sd_is_supported(dev) && !ldev->virt_lag) {
ldev->nb.notifier_call = mlx5_lag_netdev_event;
write_pnet(&ldev->net, mlx5_core_net(dev));
if (register_netdevice_notifier_net(read_pnet(&ldev->net),
@@ -392,10 +393,13 @@ static struct mlx5_lag *mlx5_lag_dev_alloc(struct mlx5_core_dev *dev)
}
ldev->mode = MLX5_LAG_MODE_NONE;
- err = mlx5_lag_mp_init(ldev);
- if (err)
- mlx5_core_err(dev, "Failed to init multipath lag err=%d\n",
- err);
+ if (!ldev->virt_lag) {
+ err = mlx5_lag_mp_init(ldev);
+ if (err)
+ mlx5_core_err(dev,
+ "Failed to init multipath lag err=%d\n",
+ err);
+ }
ldev->max_fns = ldev->ports;
return ldev;
@@ -468,7 +472,9 @@ int mlx5_lag_get_dev_index_by_seq(struct mlx5_lag *ldev, int seq)
/* Return the appropriate iterator filter for a device in LAG:
* - SD shared FDB active: iterate only the device's SD group
* - SD group exists but shared FDB not active: iterate all devices
- * - No SD: iterate ports only
+ * - Virt LAG: iterate all devices, since its members are not marked as
+ * ports, as they do not represent a physical port
+ * - Otherwise: iterate ports only
*/
static u32 mlx5_lag_get_filter(struct mlx5_lag *ldev, struct mlx5_core_dev *dev)
{
@@ -476,7 +482,7 @@ static u32 mlx5_lag_get_filter(struct mlx5_lag *ldev, struct mlx5_core_dev *dev)
if (fn && fn->sd_fdb_active)
return fn->group_id;
- if (fn && fn->group_id)
+ if (fn && (fn->group_id || ldev->virt_lag))
return MLX5_LAG_FILTER_ALL;
return MLX5_LAG_FILTER_PORTS;
}
@@ -2398,7 +2404,7 @@ int mlx5_ldev_add_mdev(struct mlx5_lag *ldev,
fn->group_id = group_id;
dev->priv.lag = ldev;
- if (group_id)
+ if (group_id || ldev->virt_lag)
return 0;
xa_set_mark(&ldev->fns, idx, MLX5_LAG_XA_MARK_PORT);
@@ -2547,7 +2553,7 @@ void mlx5_lag_add_mdev(struct mlx5_core_dev *dev)
{
int err;
- if (!mlx5_lag_is_supported(dev))
+ if (!mlx5_lag_is_supported(dev) && !mlx5_virt_lag_is_supported(dev))
return;
if (mlx5_lag_register_hca_devcom_comp(dev))
@@ -2582,6 +2588,9 @@ void mlx5_lag_remove_netdev(struct mlx5_core_dev *dev,
if (!ldev)
return;
+ if (ldev->virt_lag)
+ return;
+
mutex_lock(&ldev->lock);
mlx5_ldev_remove_netdev(ldev, netdev);
clear_bit(MLX5_LAG_FLAG_NDEVS_READY, &ldev->state_flags);
@@ -2603,6 +2612,9 @@ void mlx5_lag_add_netdev(struct mlx5_core_dev *dev,
if (!ldev)
return;
+ if (ldev->virt_lag)
+ return;
+
mutex_lock(&ldev->lock);
mlx5_ldev_add_netdev(ldev, dev, netdev);
num = mlx5_lag_num_netdevs(ldev);
@@ -2886,7 +2898,7 @@ u8 mlx5_lag_get_num_ports(struct mlx5_core_dev *dev)
if (!ldev)
return 0;
- return ldev->ports;
+ return max_t(u8, ldev->ports, 1);
}
EXPORT_SYMBOL(mlx5_lag_get_num_ports);
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/lag/mpesw.c b/drivers/net/ethernet/mellanox/mlx5/core/lag/mpesw.c
index 2d3e61541d8a..0ef2d13d0096 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/lag/mpesw.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/lag/mpesw.c
@@ -336,7 +336,7 @@ bool mlx5_lag_is_mpesw(struct mlx5_core_dev *dev)
struct mlx5_lag *ldev = mlx5_lag_dev(dev);
return ldev && ldev->mode == MLX5_LAG_MODE_MPESW &&
- __mlx5_lag_dev_is_port(ldev, dev);
+ (ldev->virt_lag || __mlx5_lag_dev_is_port(ldev, dev));
}
EXPORT_SYMBOL(mlx5_lag_is_mpesw);
--
2.44.0
^ permalink raw reply related [flat|nested] 6+ messages in thread