Sep 20 – 24, 2021
US/Pacific timezone

DSA switches: domesticating a savage beast

Sep 21, 2021, 10:20 AM
Networking and BPF Summit/Virtual-Room (LPC Virtual)

Networking and BPF Summit/Virtual-Room

LPC Virtual

Networking & BPF Summit (Closed) BPF & Networking Summit


Vladimir Oltean (NXP Semiconductors)


The DSA subsystem was originally built around Marvell devices, but has since
been extended to cover a wide variety of hardware with even wider views of
their management model. This presentation discusses the changes in DSA that
took place in the last years for this wide variety of switches to offer more
services, and in a more uniform way, to the larger network stack.

Summarized, these changes are:

  • Acknowledging switches which only have DSA tags for control plane packets,
    and modifying the bridge driver to accept termination of data plane packets
    from these switches.

  • Support for unoffloaded upper interfaces.

  • Support for more cross-chip topologies than the basic daisy chain, while
    maintaining the basic principle that network interfaces under the same bridge
    can forward from one to another, and interfaces under different bridges

The data plane and the control plane

The original DSA architecture of exposing one virtual network interface for
each front-facing switch port, and not exposing virtual network interfaces for
the ports facing inwards (CPU ports, DSA/cascade ports) has remained unchanged
to this day. DSA network interfaces should not only be conduit interfaces for
retrieving ethtool statistics and registering with the PHY library, but they
should be fully capable of sending and receiving packets. This is accomplished
via the DSA tagging protocol drivers, which are hooks in the RX and TX path of
the host Ethernet controller (the DSA master) which multiplex and demultiplex
packets to/from the correct virtual switch interface based on switch-specific
metadata that is placed in the packets.

In this model, the basic function of a network switch from a hardware
designer's perspective, which is to switch packets, is an optional feature from
the Linux network stack's perspective, and was added years after the original
design had been established.

Behind the seemingly uniform implementation of DSA tagging protocols and switch
drivers, which are tightly managed by the DSA framework, lie many differences
and subtleties that make the feature set exposed by two different DSA switches
to the network stack very different.

The majority of network switches capable of management have some sort of
distinction between the data plane packets and the control packets.

At the most basic level, control packets, which must be used for link-layer
control protocols like STP and PTP, have the ability to target a specific
egress port and to override its STP state (inject into a BLOCKING port). These
packets typically bypass the forwarding layer of the switch and the frame
analysis stage of the ingress (CPU) port and are injected directly into the
egress port. The implications are that metadata such as QoS class and VLAN ID
must be specified by the operating system driver directly as part of the DSA
tag, and that hardware address learning is not performed on the CPU port.

On the opposite side of the spectrum, data plane packets do not perform STP
state override, are subject to hardware address learning on the CPU port, but
also cannot be steered towards a precise destination port, since they are also
subject to the forwarding rules of the switch.

At the extreme, there exists a DSA_TAG_PROTO_NONE tagging protocol, which
admits defeat and does not attempt to multiplex/demultiplex virtual switch
interfaces from the DSA tag, and all network I/O through such a switch takes
place through the DSA master which is seen as a switched endpoint. The network
interfaces registered for the switches are only used for control operations
(ethtool, PHY library) and are "dead" to the network stack both for control
plane and for data plane packets. These are the "unmanaged" switches.

Finally, in some switch designs, injecting a control packet is an expensive
operation which cannot be sustained at line rate, and the bulk of the traffic
(the data plane packets) should be injected, from the hardware designer's
perspective, directly through the DSA master interface, with no DSA tag.
These are the "lightly managed" switches, and their virtual DSA interfaces are
similarly "dead" to the network stack except for link-local packets.

The most basic and common approach with this type of hardware is to simply set
up a user space configuration to perform the traffic termination from the
switching domain on the DSA master itself. For some packets to target a single
switch port, the user is required to install a bridge VLAN on the switch port
which is egress-tagged towards the CPU port, then create an 8021q upper with
the same VLAN ID on top of the DSA master, and send/receive traffic through the
8021q upper of the DSA master. This approach is, however, undesirable because
bridging DSA interfaces with non-DSA (foreign) interfaces is impossible, which
is an important use case for boards with a switch and a Wi-Fi AP (home routers).
Interfaces that are DSA masters cannot be added to a bridge either.

A slightly better integrated way of achieving the same result is the relatively
new software-defined DSA tagging protocol named tag_8021q, which can bring both
the lightly managed and unmanaged switches closer to the user model exposed by
DSA switches with hardware support for a DSA tagging protocol.

The tag_8021q protocol is fundamentally still sending data plane packets from
the perspective of the hardware, so there are things it cannot accomplish, like
STP state override. Additionally, the DSA framework has traditionally not
enforced any meaningful distinction between data plane and control plane
packets, since originally, the assumption was that all packets injected by the
software network stack should be control packets.

To unify the hardware and the software notions, and to use these chips in the
way they were meant to, the network stack must be taught about data plane
packets. The tag_8021q model breaks down when DSA switch interfaces offload a
VLAN-aware bridge, which is in fact their primary use cases. This is because
the source port of the switch cannot be retrieved based on the VLAN ID by the
tagging protocol driver on RX, because the VLANs are under the control of the
bridge driver, not DSA, and there is no guarantee that a VLAN is uniquely
installed on a single switch port. So bridging with foreign interfaces becomes
equally impossible.

The decisive changes which made these switches correctly offload a VLAN-aware
bridge come in the form of not attempting to report a precise source port on RX
for data plane packets, just a plausible/imprecise one. As long as some
requirements inside the software bridge's ingress path are satisfied (valid STP
state, VLAN ID from the packet is in the port's membership list), the bridge is
happy to accept the packet as valid, and process it on behalf of the imprecise
DSA interface that was reported.

Complications arise due to the fact that the software bridge might learn the
MAC SA of these packets on a potentially wrong port, and deliver those packets
on the return path towards the wrong port. Additionally, due to bandwidth
constraints, DSA interfaces do not synchronize their hardware FDB with the FDB
of the software bridge, so the software bridge does not have an opportunity to
figure out the real source port of imprecise packets.

To give DSA the chance to right a wrong, the bridge driver was modified to
support TX forwarding offload. With this feature, the software bridge avoids
cloning an skb which needs to be flooded to multiple ports, and sends only one
copy of the packet towards a single network interface from each "hardware
domain" that the flooded packet must reach. The port driver is responsible with
looking up its hardware FDB for that packet and replicate the packet as needed.
This is a useful feature in itself, because with switches with a large port
count, multicast traffic on the bottlenecked link between the DSA master and
the CPU port is reduced, and packets are replicated inside the hardware.
But with the lightly-managed and unmanaged switches, it makes the imprecise RX
work correctly, since the TX is also imprecise. So even though the software
bridge did learn the MAC SA of the packets on the wrong source port, that
source port is in the same hardware domain with the right port, and even though
the software FDB is incorrect, the hardware FDB isn't. So DSA drivers for
lightly-managed and unmanaged switches have a chance to properly terminate
traffic on behalf of a VLAN-aware bridge, in a way that is compatible with
bridging with foreign interfaces, and with a user space interaction procedure
that is much more uniform with DSA drivers that always send and receive packets
with a DSA tag.

Unoffloaded software upper interfaces

Recently, DSA has also gained support for offloading other virtual network
interfaces than the Linux bridge. These are the hsr driver (which supports the
HSR and PRP industrial redundancy protocols) and the bonding/team drivers
(which support the link aggregation protocol).

Not all switches are capable of offloading hsr and team/bonding, and DSA's
policy is to fall back to a software implementation when hardware offload
cannot be achieved: the bandwidth to/from the CPU is often times good enough
that this is not impractical.

However, DSA's policy could not be enforced right away with the expected
results, due to two roadblocks that led to further changes in the kernel code

To not offload an upper interface means for DSA that the physical port should
behave exactly as it would if it was a standalone interface with no switching
to the others except the CPU port, and which is capable of IP termination.

But when the unoffloaded upper interface (the software LAG) is part of a
bridge, the bridge driver makes the incorrect assumption that it is capable of
hardware forwarding towards all other ports which report the same physical
switch ID. Instead, forwarding to/from a software LAG should take place in
software. This has led to a redesign of the switchdev API, in that drivers must
now explicitly mark to the bridge the network interfaces that are capable of
autonomous forwarding; the new default being that they aren't. In the new
model, even if two interfaces report the same physical switch ID, they might
yet not be part of the same hardware domain for autonomous forwarding as far as
the bridge is concerned.

The second roadblock, even after the bridge was taught to allow software
forwarding between some interfaces which have the same physical switch ID, was
FDB isolation in DSA switches. Up until this point, the vast majority of DSA
drivers, as well as the DSA core, have considered that it is enough to offload
multiple bridges by enforcing a separation between the ports of one bridge and
the ports of another at the forwarding level. This works as long as the same
MAC address (or MAC+VLAN pair, in VLAN-aware bridges) is not present in more
than one bridging domain at the same time. This is an apparently reasonable
restriction that should never be seen in real life, so no precautions have been
taken against it in drivers or the core.

The issue is that a DSA switch is still a switch, and for every packet it
forwards, regardless of whether it is received on a standalone port, a port
under a VLAN-unaware bridge or under a VLAN-aware one, it will attempt to look
up the FDB to find the destination. With unoffloaded LAGs on top of a
standalone DSA port, where forwarding between the switched domain and the
standalone port takes place in software, the expectation that a MAC address is
only present in one bridging domain is no longer true. From the perspective of
the ports under the hardware bridge, a MAC address might come from the outside
world, whereas from the perspective of the standalone ports, the same MAC
address might come from the CPU port. So without FDB isolation, the standalone
port might look up the FDB for a MAC address and see that it could forward the
packet directly to the port in the hardware bridge domain, where that packet
was learned by the bridge port, shortcircuiting the CPU. But the forwarding
isolation rules put in place will prevent this from happening, so packets will
be dropped instead of being forwarded in software.

Individual drivers have started receiving patches for FDB isolation between
standalone ports and bridged ports, but it is possible to conceive real life
situations where even FDB isolation between one bridge and another must be
maintained. Since the DSA core does not enforce FDB isolation through its API
and many drivers already have been written without it in mind, it is to be
expected that many years pass until DSA offers a uniform set of services to
upper layers in this regard.

Switch topology changes

Traditionally, the cross-chip setups supported by DSA have been daisy chains,
where all switches except the top-most one lack a dedicated CPU port, and are
simply cascaded towards an upstream switch. There are two new switch topologies
supported by DSA now.

The first is the disjoint tree topology. A DSA tree is comprised of all
switches directly connected to each other which use a compatible tagging
protocol (one switch understands the packets from the other one, and can
push/pop them as needed). Disjoint trees are used when DSA switches are
connected to each other, but their tagging protocols are not compatible.
As opposed to one switch understanding another's, tag stacking takes place, so
in software, more than one DSA tagging protocol driver needs to be invoked for
the same packet. In such a system, each switch forms its own tree. Disjoint
trees were already supported, but the new changes also permit some hardware
forwarding to take place between switches belonging to different trees. For
example, be there an embedded 5 port DSA switch that has 3 external DSA
switches connected to 3 of its ports. Each embedded DSA switch interface is a
DSA master for the external DSA switch beneath it, and there are 4 DSA disjoint
trees in this system. For a packet to be sent from external switch 1 to
external switch 2, it must be forwarded towards the CPU port. In the most basic
configuration, forwarding between the two external switches can take place in
software. However, it is desirable that the embedded DSA switch that is a
master of external switches 1 and 2 can accelerate the forwarding between the
two (because the external switches are tagger-compatible, they are just
separated by a switch which isn't tagger-compatible with them). Under some
conditions, this is possible as long as the embedded DSA switch still has some
elementary understanding of the packets, and can still forward them by MAC DA
and optionally VLAN ID, even though they are DSA-tagged. With the vast majority
of DSA tagging protocols, the MAC DA of the packets is not altered even when a
DSA tag is inserted, so the embedded DSA master can sanely forward packets
between one external switch and another. This is one of the only special cases
where DSA master interfaces can be bridged (they are part of a separate bridge
compared to the external switch ports), because in this case, the DSA masters
are part of a bridge with no software data plane, just a hardware data plane.
The second requirement is for both the embedded and the external switches to
have the same understanding of what constitutes a data plane packet, and what
constitutes a control plane packet: STP packets received by the external switch
should not be flooded by the embedded switch. Due to the same reason that the
embedded switch must still preserve an elementary understanding of the MAC DA
of packets tagged with the external switch's tagging protocol, this will also
be the case, since typical link-layer protocols have unique link-local
multicast MAC addresses.

The second is the H tree topology. In such a system, there are multiple
switches laterally interconnected through cascade ports, but to reach the CPU,
each switch has its own dedicated CPU port. It turns out that to support such a
system, there are two distinct issues.

First, with regard to RX filtering, an H tree topology is very similar in
challenges to a single switch with multiple CPU ports. Hardware address
learning on the CPU port, if at all available, is of no use and leads to
addresses bouncing and packet drops. All MAC addresses which need to be
filtered to the host need to be installed on all CPU ports as static FDB
entries. This has led to the extension of the bridge switchdev FDB notifiers to
cover FDB entries that are local to the bridge, and which should not be

Secondly, in an H topology it is actually possible to have packet loops with
the TX forwarding offload feature enabled, because TX data plane packets sent
by the stack to one switch might also be flooded through the cascade port to
the other switch, where they might be again flooded to the second switch's CPU
port, where they will be processed as RX packets. Currently, drivers which
support this topology need to be individually patched to cut RX from cascade
ports that go towards switches that have their own CPU port, because the DSA
driver API does not have the necessary insight into driver internals as to be
able to cut forwarding between two ports only in a specific direction.

Future changes

One of the most important features still absent from DSA is the support for
multiple CPU ports, the ability to dynamically change DSA masters and the
option to configure the CPU ports in a link aggregation group. However, with
many roadblocks such as basic RX filtering support now out of the way, this
functionality will arrive sooner rather than later.

There is also the emerging topic of Ethernet controllers as DSA masters that
are aware of the DSA switches beneath them, which is typical when both the
switch and the Ethernet controller are made by the same silicon vendor.
Right now DSA can freely inherit all master->vlan_features, such as TX
checksumming offloads, but this does not work for all switch and DSA master
combinations, so it must be refined and only the known-working master and
switch combinations inherit the extra features.

On the same topic of DSA-aware masters, SR-IOV capable masters are expected to
still work when attached to a DSA switch, but the network stack's model of this
use case is unclear. VFs on top of a DSA master should be treated as switched
endpoints, but the VF driver's transmit and receive procedures do not go
through the DSA tagging protocol hooks, and these packets are therefore
DSA-untagged. So hardware manufacturers have the option of inserting DSA tags
in hardware for packets sent through a VF that goes through a DSA switch. It is
unclear, however, according to which bridging domain are these VFs being
forwarded. An effort should be made to standardize the way in which the network
stack treats these interfaces. It appears reasonable that DSA switches might
have to register virtual network interfaces that are facing each VF of the
master, in order to enforce their bridging domain, but this makes the DSA
master and switch drivers closely coupled.

On the other hand, letting other code paths than the DSA tagging protocol
driver inject packets into the switch risks compromising the integrity of the
hardware, which is an issue that currently exists and needs to be addressed.

As a conclusion, taming DSA switches and making them behave completely in
accordance with the network stack's expectations proves to be a much more
ambitious challenge than initially foreseen, thus the fight continues.

I agree to abide by the anti-harassment policy I agree

Primary author

Vladimir Oltean (NXP Semiconductors)

Presentation materials

Diamond Sponsor

Platinum Sponsor

Gold Sponsors

Silver Sponsors

Speaker Gift Sponsor

T-Shirt Sponsor

Conference Services provided by