This document describes the Distributed Switch Architecture (DSA) subsystem design principles, limitations, interactions with other subsystems, and how to develop drivers for this subsystem as well as a TODO for developers interested in joining the effort.

Design principles

The Distributed Switch Architecture subsystem was primarily designed to support Marvell Ethernet switches (MV88E6xxx, a.k.a. Link Street product line) using Linux, but has since evolved to support other vendors as well.

The original philosophy behind this design was to be able to use unmodified Linux tools such as bridge, iproute2, ifconfig to work transparently whether they configured/queried a switch port network device or a regular network device.

An Ethernet switch typically comprises multiple front-panel ports and one or more CPU or management ports. The DSA subsystem currently relies on the presence of a management port connected to an Ethernet controller capable of receiving Ethernet frames from the switch. This is a very common setup for all kinds of Ethernet switches found in Small Home and Office products: routers, gateways, or even top-of-rack switches. This host Ethernet controller will be later referred to as “conduit” and “cpu” in DSA terminology and code.

The D in DSA stands for Distributed, because the subsystem has been designed with the ability to configure and manage cascaded switches on top of each other using upstream and downstream Ethernet links between switches. These specific ports are referred to as “dsa” ports in DSA terminology and code. A collection of multiple switches connected to each other is called a “switch tree”.

For each front-panel port, DSA creates specialized network devices which are used as controlling and data-flowing endpoints for use by the Linux networking stack. These specialized network interfaces are referred to as “user” network interfaces in DSA terminology and code.

The ideal case for using DSA is when an Ethernet switch supports a “switch tag” which is a hardware feature making the switch insert a specific tag for each Ethernet frame it receives to/from specific ports to help the management interface figure out:

  • what port is this frame coming from

  • what was the reason why this frame got forwarded

  • how to send CPU originated traffic to specific ports

The subsystem does support switches not capable of inserting/stripping tags, but the features might be slightly limited in that case (traffic separation relies on Port-based VLAN IDs).

Note that DSA does not currently create network interfaces for the “cpu” and “dsa” ports because:

  • the “cpu” port is the Ethernet switch facing side of the management controller, and as such, would create a duplication of feature, since you would get two interfaces for the same conduit: conduit netdev, and “cpu” netdev

  • the “dsa” port(s) are just conduits between two or more switches, and as such cannot really be used as proper network interfaces either, only the downstream, or the top-most upstream interface makes sense with that model

NB: for the past 15 years, the DSA subsystem had been making use of the terms “master” (rather than “conduit”) and “slave” (rather than “user”). These terms have been removed from the DSA codebase and phased out of the uAPI.

Switch tagging protocols

DSA supports many vendor-specific tagging protocols, one software-defined tagging protocol, and a tag-less mode as well (DSA_TAG_PROTO_NONE).

The exact format of the tag protocol is vendor specific, but in general, they all contain something which:

  • identifies which port the Ethernet frame came from/should be sent to

  • provides a reason why this frame was forwarded to the management interface

All tagging protocols are in net/dsa/tag_*.c files and implement the methods of the struct dsa_device_ops structure, which are detailed below.

Tagging protocols generally fall in one of three categories:

  1. The switch-specific frame header is located before the Ethernet header, shifting to the right (from the perspective of the DSA conduit’s frame parser) the MAC DA, MAC SA, EtherType and the entire L2 payload.

  2. The switch-specific frame header is located before the EtherType, keeping the MAC DA and MAC SA in place from the DSA conduit’s perspective, but shifting the ‘real’ EtherType and L2 payload to the right.

  3. The switch-specific frame header is located at the tail of the packet, keeping all frame headers in place and not altering the view of the packet that the DSA conduit’s frame parser has.

A tagging protocol may tag all packets with switch tags of the same length, or the tag length might vary (for example packets with PTP timestamps might require an extended switch tag, or there might be one tag length on TX and a different one on RX). Either way, the tagging protocol driver must populate the struct dsa_device_ops::needed_headroom and/or struct dsa_device_ops::needed_tailroom with the length in octets of the longest switch frame header/trailer. The DSA framework will automatically adjust the MTU of the conduit interface to accommodate for this extra size in order for DSA user ports to support the standard MTU (L2 payload length) of 1500 octets. The needed_headroom and needed_tailroom properties are also used to request from the network stack, on a best-effort basis, the allocation of packets with enough extra space such that the act of pushing the switch tag on transmission of a packet does not cause it to reallocate due to lack of memory.

Even though applications are not expected to parse DSA-specific frame headers, the format on the wire of the tagging protocol represents an Application Binary Interface exposed by the kernel towards user space, for decoders such as libpcap. The tagging protocol driver must populate the proto member of struct dsa_device_ops with a value that uniquely describes the characteristics of the interaction required between the switch hardware and the data path driver: the offset of each bit field within the frame header and any stateful processing required to deal with the frames (as may be required for PTP timestamping).

From the perspective of the network stack, all switches within the same DSA switch tree use the same tagging protocol. In case of a packet transiting a fabric with more than one switch, the switch-specific frame header is inserted by the first switch in the fabric that the packet was received on. This header typically contains information regarding its type (whether it is a control frame that must be trapped to the CPU, or a data frame to be forwarded). Control frames should be decapsulated only by the software data path, whereas data frames might also be autonomously forwarded towards other user ports of other switches from the same fabric, and in this case, the outermost switch ports must decapsulate the packet.

Note that in certain cases, it might be the case that the tagging format used by a leaf switch (not connected directly to the CPU) is not the same as what the network stack sees. This can be seen with Marvell switch trees, where the CPU port can be configured to use either the DSA or the Ethertype DSA (EDSA) format, but the DSA links are configured to use the shorter (without Ethertype) DSA frame header, in order to reduce the autonomous packet forwarding overhead. It still remains the case that, if the DSA switch tree is configured for the EDSA tagging protocol, the operating system sees EDSA-tagged packets from the leaf switches that tagged them with the shorter DSA header. This can be done because the Marvell switch connected directly to the CPU is configured to perform tag translation between DSA and EDSA (which is simply the operation of adding or removing the ETH_P_EDSA EtherType and some padding octets).

It is possible to construct cascaded setups of DSA switches even if their tagging protocols are not compatible with one another. In this case, there are no DSA links in this fabric, and each switch constitutes a disjoint DSA switch tree. The DSA links are viewed as simply a pair of a DSA conduit (the out-facing port of the upstream DSA switch) and a CPU port (the in-facing port of the downstream DSA switch).

The tagging protocol of the attached DSA switch tree can be viewed through the dsa/tagging sysfs attribute of the DSA conduit:

cat /sys/class/net/eth0/dsa/tagging

If the hardware and driver are capable, the tagging protocol of the DSA switch tree can be changed at runtime. This is done by writing the new tagging protocol name to the same sysfs device attribute as above (the DSA conduit and all attached switch ports must be down while doing this).

It is desirable that all tagging protocols are testable with the dsa_loop mockup driver, which can be attached to any network interface. The goal is that any network interface should be capable of transmitting the same packet in the same way, and the tagger should decode the same received packet in the same way regardless of the driver used for the switch control path, and the driver used for the DSA conduit.

The transmission of a packet goes through the tagger’s xmit function. The passed struct sk_buff *skb has skb->data pointing at skb_mac_header(skb), i.e. at the destination MAC address, and the passed struct net_device *dev represents the virtual DSA user network interface whose hardware counterpart the packet must be steered to (i.e. swp0). The job of this method is to prepare the skb in a way that the switch will understand what egress port the packet is for (and not deliver it towards other ports). Typically this is fulfilled by pushing a frame header. Checking for insufficient size in the skb headroom or tailroom is unnecessary provided that the needed_headroom and needed_tailroom properties were filled out properly, because DSA ensures there is enough space before calling this method.

The reception of a packet goes through the tagger’s rcv function. The passed struct sk_buff *skb has skb->data pointing at skb_mac_header(skb) + ETH_ALEN octets, i.e. to where the first octet after the EtherType would have been, were this frame not tagged. The role of this method is to consume the frame header, adjust skb->data to really point at the first octet after the EtherType, and to change skb->dev to point to the virtual DSA user network interface corresponding to the physical front-facing switch port that the packet was received on.

Since tagging protocols in category 1 and 2 break software (and most often also hardware) packet dissection on the DSA conduit, features such as RPS (Receive Packet Steering) on the DSA conduit would be broken. The DSA framework deals with this by hooking into the flow dissector and shifting the offset at which the IP header is to be found in the tagged frame as seen by the DSA conduit. This behavior is automatic based on the overhead value of the tagging protocol. If not all packets are of equal size, the tagger can implement the flow_dissect method of the struct dsa_device_ops and override this default behavior by specifying the correct offset incurred by each individual RX packet. Tail taggers do not cause issues to the flow dissector.

Checksum offload should work with category 1 and 2 taggers when the DSA conduit driver declares NETIF_F_HW_CSUM in vlan_features and looks at csum_start and csum_offset. For those cases, DSA will shift the checksum start and offset by the tag size. If the DSA conduit driver still uses the legacy NETIF_F_IP_CSUM or NETIF_F_IPV6_CSUM in vlan_features, the offload might only work if the offload hardware already expects that specific tag (perhaps due to matching vendors). DSA user ports inherit those flags from the conduit, and it is up to the driver to correctly fall back to software checksum when the IP header is not where the hardware expects. If that check is ineffective, the packets might go to the network without a proper checksum (the checksum field will have the pseudo IP header sum). For category 3, when the offload hardware does not already expect the switch tag in use, the checksum must be calculated before any tag is inserted (i.e. inside the tagger). Otherwise, the DSA conduit would include the tail tag in the (software or hardware) checksum calculation. Then, when the tag gets stripped by the switch during transmission, it will leave an incorrect IP checksum in place.

Due to various reasons (most common being category 1 taggers being associated with DSA-unaware conduits, mangling what the conduit perceives as MAC DA), the tagging protocol may require the DSA conduit to operate in promiscuous mode, to receive all frames regardless of the value of the MAC DA. This can be done by setting the promisc_on_conduit property of the struct dsa_device_ops. Note that this assumes a DSA-unaware conduit driver, which is the norm.

Conduit network devices

Conduit network devices are regular, unmodified Linux network device drivers for the CPU/management Ethernet interface. Such a driver might occasionally need to know whether DSA is enabled (e.g.: to enable/disable specific offload features), but the DSA subsystem has been proven to work with industry standard drivers: e1000e, mv643xx_eth etc. without having to introduce modifications to these drivers. Such network devices are also often referred to as conduit network devices since they act as a pipe between the host processor and the hardware Ethernet switch.

Networking stack hooks

When a conduit netdev is used with DSA, a small hook is placed in the networking stack is in order to have the DSA subsystem process the Ethernet switch specific tagging protocol. DSA accomplishes this by registering a specific (and fake) Ethernet type (later becoming skb->protocol) with the networking stack, this is also known as a ptype or packet_type. A typical Ethernet Frame receive sequence looks like this:

Conduit network device (e.g.: e1000e):

  1. Receive interrupt fires:

    • receive function is invoked

    • basic packet processing is done: getting length, status etc.

    • packet is prepared to be processed by the Ethernet layer by calling eth_type_trans

  2. net/ethernet/eth.c:

    eth_type_trans(skb, dev)
            if (dev->dsa_ptr != NULL)
                    -> skb->protocol = ETH_P_XDSA
  3. drivers/net/ethernet/*:

            -> iterate over registered packet_type
                    -> invoke handler for ETH_P_XDSA, calls dsa_switch_rcv()
  4. net/dsa/dsa.c:

    -> dsa_switch_rcv()
            -> invoke switch tag specific protocol handler in 'net/dsa/tag_*.c'
  5. net/dsa/tag_*.c:

    • inspect and strip switch tag protocol to determine originating port

    • locate per-port network device

    • invoke eth_type_trans() with the DSA user network device

    • invoked netif_receive_skb()

Past this point, the DSA user network devices get delivered regular Ethernet frames that can be processed by the networking stack.

User network devices

User network devices created by DSA are stacked on top of their conduit network device, each of these network interfaces will be responsible for being a controlling and data-flowing end-point for each front-panel port of the switch. These interfaces are specialized in order to:

  • insert/remove the switch tag protocol (if it exists) when sending traffic to/from specific switch ports

  • query the switch for ethtool operations: statistics, link state, Wake-on-LAN, register dumps...

  • manage external/internal PHY: link, auto-negotiation, etc.

These user network devices have custom net_device_ops and ethtool_ops function pointers which allow DSA to introduce a level of layering between the networking stack/ethtool and the switch driver implementation.

Upon frame transmission from these user network devices, DSA will look up which switch tagging protocol is currently registered with these network devices and invoke a specific transmit routine which takes care of adding the relevant switch tag in the Ethernet frames.

These frames are then queued for transmission using the conduit network device ndo_start_xmit() function. Since they contain the appropriate switch tag, the Ethernet switch will be able to process these incoming frames from the management interface and deliver them to the physical switch port.

When using multiple CPU ports, it is possible to stack a LAG (bonding/team) device between the DSA user devices and the physical DSA conduits. The LAG device is thus also a DSA conduit, but the LAG slave devices continue to be DSA conduits as well (just with no user port assigned to them; this is needed for recovery in case the LAG DSA conduit disappears). Thus, the data path of the LAG DSA conduit is used asymmetrically. On RX, the ETH_P_XDSA handler, which calls dsa_switch_rcv(), is invoked early (on the physical DSA conduit; LAG slave). Therefore, the RX data path of the LAG DSA conduit is not used. On the other hand, TX takes place linearly: dsa_user_xmit calls dsa_enqueue_skb, which calls dev_queue_xmit towards the LAG DSA conduit. The latter calls dev_queue_xmit towards one physical DSA conduit or the other, and in both cases, the packet exits the system through a hardware path towards the switch.

Graphical representation

Summarized, this is basically how DSA looks like from a network device perspective:

             Unaware application
           opens and binds socket
                    |  ^
                    |  |
        |+------+ +------+ +------+ +------+|
        || swp0 | | swp1 | | swp2 | | swp3 ||
        |          DSA switch driver        |
                      |        ^
         Tag added by |        | Tag consumed by
        switch driver |        | switch driver
                      v        |
        | Unmodified host interface driver  | Software
        |       Host interface (eth0)       | Hardware
                      |        ^
      Tag consumed by |        | Tag added by
      switch hardware |        | switch hardware
                      v        |
        |               Switch              |
        |+------+ +------+ +------+ +------+|
        || swp0 | | swp1 | | swp2 | | swp3 ||

User MDIO bus

In order to be able to read to/from a switch PHY built into it, DSA creates an user MDIO bus which allows a specific switch driver to divert and intercept MDIO reads/writes towards specific PHY addresses. In most MDIO-connected switches, these functions would utilize direct or indirect PHY addressing mode to return standard MII registers from the switch builtin PHYs, allowing the PHY library and/or to return link status, link partner pages, auto-negotiation results, etc.

For Ethernet switches which have both external and internal MDIO buses, the user MII bus can be utilized to mux/demux MDIO reads and writes towards either internal or external MDIO devices this switch might be connected to: internal PHYs, external PHYs, or even external switches.

Data structures

DSA data structures are defined in include/net/dsa.h as well as net/dsa/dsa_priv.h:

  • dsa_chip_data: platform data configuration for a given switch device, this structure describes a switch device’s parent device, its address, as well as various properties of its ports: names/labels, and finally a routing table indication (when cascading switches)

  • dsa_platform_data: platform device configuration data which can reference a collection of dsa_chip_data structures if multiple switches are cascaded, the conduit network device this switch tree is attached to needs to be referenced

  • dsa_switch_tree: structure assigned to the conduit network device under dsa_ptr, this structure references a dsa_platform_data structure as well as the tagging protocol supported by the switch tree, and which receive/transmit function hooks should be invoked, information about the directly attached switch is also provided: CPU port. Finally, a collection of dsa_switch are referenced to address individual switches in the tree.

  • dsa_switch: structure describing a switch device in the tree, referencing a dsa_switch_tree as a backpointer, user network devices, conduit network device, and a reference to the backing``dsa_switch_ops``

  • dsa_switch_ops: structure referencing function pointers, see below for a full description.

Design limitations

Lack of CPU/DSA network devices

DSA does not currently create user network devices for the CPU or DSA ports, as described before. This might be an issue in the following cases:

  • inability to fetch switch CPU port statistics counters using ethtool, which can make it harder to debug MDIO switch connected using xMII interfaces

  • inability to configure the CPU port link parameters based on the Ethernet controller capabilities attached to it:

  • inability to configure specific VLAN IDs / trunking VLANs between switches when using a cascaded setup

Common pitfalls using DSA setups

Once a conduit network device is configured to use DSA (dev->dsa_ptr becomes non-NULL), and the switch behind it expects a tagging protocol, this network interface can only exclusively be used as a conduit interface. Sending packets directly through this interface (e.g.: opening a socket using this interface) will not make us go through the switch tagging protocol transmit function, so the Ethernet switch on the other end, expecting a tag will typically drop this frame.

Interactions with other subsystems

DSA currently leverages the following subsystems:

  • MDIO/PHY library: drivers/net/phy/phy.c, mdio_bus.c

  • Switchdev:net/switchdev/*

  • Device Tree for various of_* functions

  • Devlink: net/core/devlink.c

MDIO/PHY library

User network devices exposed by DSA may or may not be interfacing with PHY devices (struct phy_device as defined in include/linux/phy.h), but the DSA subsystem deals with all possible combinations:

  • internal PHY devices, built into the Ethernet switch hardware

  • external PHY devices, connected via an internal or external MDIO bus

  • internal PHY devices, connected via an internal MDIO bus

  • special, non-autonegotiated or non MDIO-managed PHY devices: SFPs, MoCA; a.k.a fixed PHYs

The PHY configuration is done by the dsa_user_phy_setup() function and the logic basically looks like this:

  • if Device Tree is used, the PHY device is looked up using the standard “phy-handle” property, if found, this PHY device is created and registered using of_phy_connect()

  • if Device Tree is used and the PHY device is “fixed”, that is, conforms to the definition of a non-MDIO managed PHY as defined in Documentation/devicetree/bindings/net/fixed-link.txt, the PHY is registered and connected transparently using the special fixed MDIO bus driver

  • finally, if the PHY is built into the switch, as is very common with standalone switch packages, the PHY is probed using the user MII bus created by DSA


DSA directly utilizes SWITCHDEV when interfacing with the bridge layer, and more specifically with its VLAN filtering portion when configuring VLANs on top of per-port user network devices. As of today, the only SWITCHDEV objects supported by DSA are the FDB and VLAN objects.

Device Tree

DSA features a standardized binding which is documented in Documentation/devicetree/bindings/net/dsa/dsa.txt. PHY/MDIO library helper functions such as of_get_phy_mode(), of_phy_connect() are also used to query per-port PHY specific details: interface connection, MDIO bus location, etc.

Driver development

DSA switch drivers need to implement a dsa_switch_ops structure which will contain the various members described below.

Probing, registration and device lifetime

DSA switches are regular device structures on buses (be they platform, SPI, I2C, MDIO or otherwise). The DSA framework is not involved in their probing with the device core.

Switch registration from the perspective of a driver means passing a valid struct dsa_switch pointer to dsa_register_switch(), usually from the switch driver’s probing function. The following members must be valid in the provided structure:

  • ds->dev: will be used to parse the switch’s OF node or platform data.

  • ds->num_ports: will be used to create the port list for this switch, and to validate the port indices provided in the OF node.

  • ds->ops: a pointer to the dsa_switch_ops structure holding the DSA method implementations.

  • ds->priv: backpointer to a driver-private data structure which can be retrieved in all further DSA method callbacks.

In addition, the following flags in the dsa_switch structure may optionally be configured to obtain driver-specific behavior from the DSA core. Their behavior when set is documented through comments in include/net/dsa.h.

  • ds->vlan_filtering_is_global

  • ds->needs_standalone_vlan_filtering

  • ds->configure_vlan_while_not_filtering

  • ds->untag_bridge_pvid

  • ds->assisted_learning_on_cpu_port

  • ds->mtu_enforcement_ingress

  • ds->fdb_isolation

Internally, DSA keeps an array of switch trees (group of switches) global to the kernel, and attaches a dsa_switch structure to a tree on registration. The tree ID to which the switch is attached is determined by the first u32 number of the dsa,member property of the switch’s OF node (0 if missing). The switch ID within the tree is determined by the second u32 number of the same OF property (0 if missing). Registering multiple switches with the same switch ID and tree ID is illegal and will cause an error. Using platform data, a single switch and a single switch tree is permitted.

In case of a tree with multiple switches, probing takes place asymmetrically. The first N-1 callers of dsa_register_switch() only add their ports to the port list of the tree (dst->ports), each port having a backpointer to its associated switch (dp->ds). Then, these switches exit their dsa_register_switch() call early, because dsa_tree_setup_routing_table() has determined that the tree is not yet complete (not all ports referenced by DSA links are present in the tree’s port list). The tree becomes complete when the last switch calls dsa_register_switch(), and this triggers the effective continuation of initialization (including the call to ds->ops->setup()) for all switches within that tree, all as part of the calling context of the last switch’s probe function.

The opposite of registration takes place when calling dsa_unregister_switch(), which removes a switch’s ports from the port list of the tree. The entire tree is torn down when the first switch unregisters.

It is mandatory for DSA switch drivers to implement the shutdown() callback of their respective bus, and call dsa_switch_shutdown() from it (a minimal version of the full teardown performed by dsa_unregister_switch()). The reason is that DSA keeps a reference on the conduit net device, and if the driver for the conduit device decides to unbind on shutdown, DSA’s reference will block that operation from finalizing.

Either dsa_switch_shutdown() or dsa_unregister_switch() must be called, but not both, and the device driver model permits the bus’ remove() method to be called even if shutdown() was already called. Therefore, drivers are expected to implement a mutual exclusion method between remove() and shutdown() by setting their drvdata to NULL after any of these has run, and checking whether the drvdata is NULL before proceeding to take any action.

After dsa_switch_shutdown() or dsa_unregister_switch() was called, no further callbacks via the provided dsa_switch_ops may take place, and the driver may free the data structures associated with the dsa_switch.

Switch configuration

  • get_tag_protocol: this is to indicate what kind of tagging protocol is supported, should be a valid value from the dsa_tag_protocol enum. The returned information does not have to be static; the driver is passed the CPU port number, as well as the tagging protocol of a possibly stacked upstream switch, in case there are hardware limitations in terms of supported tag formats.

  • change_tag_protocol: when the default tagging protocol has compatibility problems with the conduit or other issues, the driver may support changing it at runtime, either through a device tree property or through sysfs. In that case, further calls to get_tag_protocol should report the protocol in current use.

  • setup: setup function for the switch, this function is responsible for setting up the dsa_switch_ops private structure with all it needs: register maps, interrupts, mutexes, locks, etc. This function is also expected to properly configure the switch to separate all network interfaces from each other, that is, they should be isolated by the switch hardware itself, typically by creating a Port-based VLAN ID for each port and allowing only the CPU port and the specific port to be in the forwarding vector. Ports that are unused by the platform should be disabled. Past this function, the switch is expected to be fully configured and ready to serve any kind of request. It is recommended to issue a software reset of the switch during this setup function in order to avoid relying on what a previous software agent such as a bootloader/firmware may have previously configured. The method responsible for undoing any applicable allocations or operations done here is teardown.

  • port_setup and port_teardown: methods for initialization and destruction of per-port data structures. It is mandatory for some operations such as registering and unregistering devlink port regions to be done from these methods, otherwise they are optional. A port will be torn down only if it has been previously set up. It is possible for a port to be set up during probing only to be torn down immediately afterwards, for example in case its PHY cannot be found. In this case, probing of the DSA switch continues without that particular port.

  • port_change_conduit: method through which the affinity (association used for traffic termination purposes) between a user port and a CPU port can be changed. By default all user ports from a tree are assigned to the first available CPU port that makes sense for them (most of the times this means the user ports of a tree are all assigned to the same CPU port, except for H topologies as described in commit 2c0b03258b8b). The port argument represents the index of the user port, and the conduit argument represents the new DSA conduit net_device. The CPU port associated with the new conduit can be retrieved by looking at struct dsa_port *cpu_dp = conduit->dsa_ptr. Additionally, the conduit can also be a LAG device where all the slave devices are physical DSA conduits. LAG DSA also have a valid conduit->dsa_ptr pointer, however this is not unique, but rather a duplicate of the first physical DSA conduit’s (LAG slave) dsa_ptr. In case of a LAG DSA conduit, a further call to port_lag_join will be emitted separately for the physical CPU ports associated with the physical DSA conduits, requesting them to create a hardware LAG associated with the LAG interface.

Ethtool operations

  • get_strings: ethtool function used to query the driver’s strings, will typically return statistics strings, private flags strings, etc.

  • get_ethtool_stats: ethtool function used to query per-port statistics and return their values. DSA overlays user network devices general statistics: RX/TX counters from the network device, with switch driver specific statistics per port

  • get_sset_count: ethtool function used to query the number of statistics items

  • get_wol: ethtool function used to obtain Wake-on-LAN settings per-port, this function may for certain implementations also query the conduit network device Wake-on-LAN settings if this interface needs to participate in Wake-on-LAN

  • set_wol: ethtool function used to configure Wake-on-LAN settings per-port, direct counterpart to set_wol with similar restrictions

  • set_eee: ethtool function which is used to configure a switch port EEE (Green Ethernet) settings, can optionally invoke the PHY library to enable EEE at the PHY level if relevant. This function should enable EEE at the switch port MAC controller and data-processing logic

  • get_eee: ethtool function which is used to query a switch port EEE settings, this function should return the EEE state of the switch port MAC controller and data-processing logic as well as query the PHY for its currently configured EEE settings

  • get_eeprom_len: ethtool function returning for a given switch the EEPROM length/size in bytes

  • get_eeprom: ethtool function returning for a given switch the EEPROM contents

  • set_eeprom: ethtool function writing specified data to a given switch EEPROM

  • get_regs_len: ethtool function returning the register length for a given switch

  • get_regs: ethtool function returning the Ethernet switch internal register contents. This function might require user-land code in ethtool to pretty-print register values and registers

Power management

  • suspend: function invoked by the DSA platform device when the system goes to suspend, should quiesce all Ethernet switch activities, but keep ports participating in Wake-on-LAN active as well as additional wake-up logic if supported

  • resume: function invoked by the DSA platform device when the system resumes, should resume all Ethernet switch activities and re-configure the switch to be in a fully active state

  • port_enable: function invoked by the DSA user network device ndo_open function when a port is administratively brought up, this function should fully enable a given switch port. DSA takes care of marking the port with BR_STATE_BLOCKING if the port is a bridge member, or BR_STATE_FORWARDING if it was not, and propagating these changes down to the hardware

  • port_disable: function invoked by the DSA user network device ndo_close function when a port is administratively brought down, this function should fully disable a given switch port. DSA takes care of marking the port with BR_STATE_DISABLED and propagating changes to the hardware if this port is disabled while being a bridge member

Address databases

Switching hardware is expected to have a table for FDB entries, however not all of them are active at the same time. An address database is the subset (partition) of FDB entries that is active (can be matched by address learning on RX, or FDB lookup on TX) depending on the state of the port. An address database may occasionally be called “FID” (Filtering ID) in this document, although the underlying implementation may choose whatever is available to the hardware.

For example, all ports that belong to a VLAN-unaware bridge (which is currently VLAN-unaware) are expected to learn source addresses in the database associated by the driver with that bridge (and not with other VLAN-unaware bridges). During forwarding and FDB lookup, a packet received on a VLAN-unaware bridge port should be able to find a VLAN-unaware FDB entry having the same MAC DA as the packet, which is present on another port member of the same bridge. At the same time, the FDB lookup process must be able to not find an FDB entry having the same MAC DA as the packet, if that entry points towards a port which is a member of a different VLAN-unaware bridge (and is therefore associated with a different address database).

Similarly, each VLAN of each offloaded VLAN-aware bridge should have an associated address database, which is shared by all ports which are members of that VLAN, but not shared by ports belonging to different bridges that are members of the same VID.

In this context, a VLAN-unaware database means that all packets are expected to match on it irrespective of VLAN ID (only MAC address lookup), whereas a VLAN-aware database means that packets are supposed to match based on the VLAN ID from the classified 802.1Q header (or the pvid if untagged).

At the bridge layer, VLAN-unaware FDB entries have the special VID value of 0, whereas VLAN-aware FDB entries have non-zero VID values. Note that a VLAN-unaware bridge may have VLAN-aware (non-zero VID) FDB entries, and a VLAN-aware bridge may have VLAN-unaware FDB entries. As in hardware, the software bridge keeps separate address databases, and offloads to hardware the FDB entries belonging to these databases, through switchdev, asynchronously relative to the moment when the databases become active or inactive.

When a user port operates in standalone mode, its driver should configure it to use a separate database called a port private database. This is different from the databases described above, and should impede operation as standalone port (packet in, packet out to the CPU port) as little as possible. For example, on ingress, it should not attempt to learn the MAC SA of ingress traffic, since learning is a bridging layer service and this is a standalone port, therefore it would consume useless space. With no address learning, the port private database should be empty in a naive implementation, and in this case, all received packets should be trivially flooded to the CPU port.

DSA (cascade) and CPU ports are also called “shared” ports because they service multiple address databases, and the database that a packet should be associated to is usually embedded in the DSA tag. This means that the CPU port may simultaneously transport packets coming from a standalone port (which were classified by hardware in one address database), and from a bridge port (which were classified to a different address database).

Switch drivers which satisfy certain criteria are able to optimize the naive configuration by removing the CPU port from the flooding domain of the switch, and just program the hardware with FDB entries pointing towards the CPU port for which it is known that software is interested in those MAC addresses. Packets which do not match a known FDB entry will not be delivered to the CPU, which will save CPU cycles required for creating an skb just to drop it.

DSA is able to perform host address filtering for the following kinds of addresses:

  • Primary unicast MAC addresses of ports (dev->dev_addr). These are associated with the port private database of the respective user port, and the driver is notified to install them through port_fdb_add towards the CPU port.

  • Secondary unicast and multicast MAC addresses of ports (addresses added through dev_uc_add() and dev_mc_add()). These are also associated with the port private database of the respective user port.

  • Local/permanent bridge FDB entries (BR_FDB_LOCAL). These are the MAC addresses of the bridge ports, for which packets must be terminated locally and not forwarded. They are associated with the address database for that bridge.

  • Static bridge FDB entries installed towards foreign (non-DSA) interfaces present in the same bridge as some DSA switch ports. These are also associated with the address database for that bridge.

  • Dynamically learned FDB entries on foreign interfaces present in the same bridge as some DSA switch ports, only if ds->assisted_learning_on_cpu_port is set to true by the driver. These are associated with the address database for that bridge.

For various operations detailed below, DSA provides a dsa_db structure which can be of the following types:

  • DSA_DB_PORT: the FDB (or MDB) entry to be installed or deleted belongs to the port private database of user port db->dp.

  • DSA_DB_BRIDGE: the entry belongs to one of the address databases of bridge db->bridge. Separation between the VLAN-unaware database and the per-VID databases of this bridge is expected to be done by the driver.

  • DSA_DB_LAG: the entry belongs to the address database of LAG db->lag. Note: DSA_DB_LAG is currently unused and may be removed in the future.

The drivers which act upon the dsa_db argument in port_fdb_add, port_mdb_add etc should declare ds->fdb_isolation as true.

DSA associates each offloaded bridge and each offloaded LAG with a one-based ID (struct dsa_bridge :: num, struct dsa_lag :: id) for the purposes of refcounting addresses on shared ports. Drivers may piggyback on DSA’s numbering scheme (the ID is readable through db->bridge.num and db-> or may implement their own.

Only the drivers which declare support for FDB isolation are notified of FDB entries on the CPU port belonging to DSA_DB_PORT databases. For compatibility/legacy reasons, DSA_DB_BRIDGE addresses are notified to drivers even if they do not support FDB isolation. However, db->bridge.num and db-> are always set to 0 in that case (to denote the lack of isolation, for refcounting purposes).

Note that it is not mandatory for a switch driver to implement physically separate address databases for each standalone user port. Since FDB entries in the port private databases will always point to the CPU port, there is no risk for incorrect forwarding decisions. In this case, all standalone ports may share the same database, but the reference counting of host-filtered addresses (not deleting the FDB entry for a port’s MAC address if it’s still in use by another port) becomes the responsibility of the driver, because DSA is unaware that the port databases are in fact shared. This can be achieved by calling dsa_fdb_present_in_other_db() and dsa_mdb_present_in_other_db(). The down side is that the RX filtering lists of each user port are in fact shared, which means that user port A may accept a packet with a MAC DA it shouldn’t have, only because that MAC address was in the RX filtering list of user port B. These packets will still be dropped in software, however.

Bridge layer

Offloading the bridge forwarding plane is optional and handled by the methods below. They may be absent, return -EOPNOTSUPP, or ds->max_num_bridges may be non-zero and exceeded, and in this case, joining a bridge port is still possible, but the packet forwarding will take place in software, and the ports under a software bridge must remain configured in the same way as for standalone operation, i.e. have all bridging service functions (address learning etc) disabled, and send all received packets to the CPU port only.

Concretely, a port starts offloading the forwarding plane of a bridge once it returns success to the port_bridge_join method, and stops doing so after port_bridge_leave has been called. Offloading the bridge means autonomously learning FDB entries in accordance with the software bridge port’s state, and autonomously forwarding (or flooding) received packets without CPU intervention. This is optional even when offloading a bridge port. Tagging protocol drivers are expected to call dsa_default_offload_fwd_mark(skb) for packets which have already been autonomously forwarded in the forwarding domain of the ingress switch port. DSA, through dsa_port_devlink_setup(), considers all switch ports part of the same tree ID to be part of the same bridge forwarding domain (capable of autonomous forwarding to each other).

Offloading the TX forwarding process of a bridge is a distinct concept from simply offloading its forwarding plane, and refers to the ability of certain driver and tag protocol combinations to transmit a single skb coming from the bridge device’s transmit function to potentially multiple egress ports (and thereby avoid its cloning in software).

Packets for which the bridge requests this behavior are called data plane packets and have skb->offload_fwd_mark set to true in the tag protocol driver’s xmit function. Data plane packets are subject to FDB lookup, hardware learning on the CPU port, and do not override the port STP state. Additionally, replication of data plane packets (multicast, flooding) is handled in hardware and the bridge driver will transmit a single skb for each packet that may or may not need replication.

When the TX forwarding offload is enabled, the tag protocol driver is responsible to inject packets into the data plane of the hardware towards the correct bridging domain (FID) that the port is a part of. The port may be VLAN-unaware, and in this case the FID must be equal to the FID used by the driver for its VLAN-unaware address database associated with that bridge. Alternatively, the bridge may be VLAN-aware, and in that case, it is guaranteed that the packet is also VLAN-tagged with the VLAN ID that the bridge processed this packet in. It is the responsibility of the hardware to untag the VID on the egress-untagged ports, or keep the tag on the egress-tagged ones.

  • port_bridge_join: bridge layer function invoked when a given switch port is added to a bridge, this function should do what’s necessary at the switch level to permit the joining port to be added to the relevant logical domain for it to ingress/egress traffic with other members of the bridge. By setting the tx_fwd_offload argument to true, the TX forwarding process of this bridge is also offloaded.

  • port_bridge_leave: bridge layer function invoked when a given switch port is removed from a bridge, this function should do what’s necessary at the switch level to deny the leaving port from ingress/egress traffic from the remaining bridge members.

  • port_stp_state_set: bridge layer function invoked when a given switch port STP state is computed by the bridge layer and should be propagated to switch hardware to forward/block/learn traffic.

  • port_bridge_flags: bridge layer function invoked when a port must configure its settings for e.g. flooding of unknown traffic or source address learning. The switch driver is responsible for initial setup of the standalone ports with address learning disabled and egress flooding of all types of traffic, then the DSA core notifies of any change to the bridge port flags when the port joins and leaves a bridge. DSA does not currently manage the bridge port flags for the CPU port. The assumption is that address learning should be statically enabled (if supported by the hardware) on the CPU port, and flooding towards the CPU port should also be enabled, due to a lack of an explicit address filtering mechanism in the DSA core.

  • port_fast_age: bridge layer function invoked when flushing the dynamically learned FDB entries on the port is necessary. This is called when transitioning from an STP state where learning should take place to an STP state where it shouldn’t, or when leaving a bridge, or when address learning is turned off via port_bridge_flags.

Bridge VLAN filtering

  • port_vlan_filtering: bridge layer function invoked when the bridge gets configured for turning on or off VLAN filtering. If nothing specific needs to be done at the hardware level, this callback does not need to be implemented. When VLAN filtering is turned on, the hardware must be programmed with rejecting 802.1Q frames which have VLAN IDs outside of the programmed allowed VLAN ID map/rules. If there is no PVID programmed into the switch port, untagged frames must be rejected as well. When turned off the switch must accept any 802.1Q frames irrespective of their VLAN ID, and untagged frames are allowed.

  • port_vlan_add: bridge layer function invoked when a VLAN is configured (tagged or untagged) for the given switch port. The CPU port becomes a member of a VLAN only if a foreign bridge port is also a member of it (and forwarding needs to take place in software), or the VLAN is installed to the VLAN group of the bridge device itself, for termination purposes (bridge vlan add dev br0 vid 100 self). VLANs on shared ports are reference counted and removed when there is no user left. Drivers do not need to manually install a VLAN on the CPU port.

  • port_vlan_del: bridge layer function invoked when a VLAN is removed from the given switch port

  • port_fdb_add: bridge layer function invoked when the bridge wants to install a Forwarding Database entry, the switch hardware should be programmed with the specified address in the specified VLAN Id in the forwarding database associated with this VLAN ID.

  • port_fdb_del: bridge layer function invoked when the bridge wants to remove a Forwarding Database entry, the switch hardware should be programmed to delete the specified MAC address from the specified VLAN ID if it was mapped into this port forwarding database

  • port_fdb_dump: bridge bypass function invoked by ndo_fdb_dump on the physical DSA port interfaces. Since DSA does not attempt to keep in sync its hardware FDB entries with the software bridge, this method is implemented as a means to view the entries visible on user ports in the hardware database. The entries reported by this function have the self flag in the output of the bridge fdb show command.

  • port_mdb_add: bridge layer function invoked when the bridge wants to install a multicast database entry. The switch hardware should be programmed with the specified address in the specified VLAN ID in the forwarding database associated with this VLAN ID.

  • port_mdb_del: bridge layer function invoked when the bridge wants to remove a multicast database entry, the switch hardware should be programmed to delete the specified MAC address from the specified VLAN ID if it was mapped into this port forwarding database.

IEC 62439-2 (MRP)

The Media Redundancy Protocol is a topology management protocol optimized for fast fault recovery time for ring networks, which has some components implemented as a function of the bridge driver. MRP uses management PDUs (Test, Topology, LinkDown/Up, Option) sent at a multicast destination MAC address range of 01:15:4e:00:00:0x and with an EtherType of 0x88e3. Depending on the node’s role in the ring (MRM: Media Redundancy Manager, MRC: Media Redundancy Client, MRA: Media Redundancy Automanager), certain MRP PDUs might need to be terminated locally and others might need to be forwarded. An MRM might also benefit from offloading to hardware the creation and transmission of certain MRP PDUs (Test).

Normally an MRP instance can be created on top of any network interface, however in the case of a device with an offloaded data path such as DSA, it is necessary for the hardware, even if it is not MRP-aware, to be able to extract the MRP PDUs from the fabric before the driver can proceed with the software implementation. DSA today has no driver which is MRP-aware, therefore it only listens for the bare minimum switchdev objects required for the software assist to work properly. The operations are detailed below.

  • port_mrp_add and port_mrp_del: notifies driver when an MRP instance with a certain ring ID, priority, primary port and secondary port is created/deleted.

  • port_mrp_add_ring_role and port_mrp_del_ring_role: function invoked when an MRP instance changes ring roles between MRM or MRC. This affects which MRP PDUs should be trapped to software and which should be autonomously forwarded.

IEC 62439-3 (HSR/PRP)

The Parallel Redundancy Protocol (PRP) is a network redundancy protocol which works by duplicating and sequence numbering packets through two independent L2 networks (which are unaware of the PRP tail tags carried in the packets), and eliminating the duplicates at the receiver. The High-availability Seamless Redundancy (HSR) protocol is similar in concept, except all nodes that carry the redundant traffic are aware of the fact that it is HSR-tagged (because HSR uses a header with an EtherType of 0x892f) and are physically connected in a ring topology. Both HSR and PRP use supervision frames for monitoring the health of the network and for discovery of other nodes.

In Linux, both HSR and PRP are implemented in the hsr driver, which instantiates a virtual, stackable network interface with two member ports. The driver only implements the basic roles of DANH (Doubly Attached Node implementing HSR) and DANP (Doubly Attached Node implementing PRP); the roles of RedBox and QuadBox are not implemented (therefore, bridging a hsr network interface with a physical switch port does not produce the expected result).

A driver which is able of offloading certain functions of a DANP or DANH should declare the corresponding netdev features as indicated by the documentation at Documentation/networking/netdev-features.rst. Additionally, the following methods must be implemented:

  • port_hsr_join: function invoked when a given switch port is added to a DANP/DANH. The driver may return -EOPNOTSUPP and in this case, DSA will fall back to a software implementation where all traffic from this port is sent to the CPU.

  • port_hsr_leave: function invoked when a given switch port leaves a DANP/DANH and returns to normal operation as a standalone port.


Making SWITCHDEV and DSA converge towards an unified codebase

SWITCHDEV properly takes care of abstracting the networking stack with offload capable hardware, but does not enforce a strict switch device driver model. On the other DSA enforces a fairly strict device driver model, and deals with most of the switch specific. At some point we should envision a merger between these two subsystems and get the best of both worlds.